text
stringlengths
0
643k
meta
stringlengths
137
151
# Introduction {#Intro} The original 1970 conception of Mastermind by Mordechai Meirowitz was a sequence reconstruction game. One player (the *codemaker*) would construct a hidden sequence of four pegs, each peg being one of six colors, and the other player (the *codebreaker*) would make guesses of the same form, receiving feedback after each guess regarding how close they were to the hidden sequence. In 1963 Erdős and Rényi studied the two-color variant of this game, and after the release of Mastermind, Knuth showed that a minimax strategy guarantees guessing the hidden vector in no more than 5 turns. Many authors have since studied algorithms to minimize the number of guesses required in the worst case, and almost all of these results will be introduced and discussed at relevant points in this paper. We note here that the work of Jäger and Peczarski and Goddard deal with finding explicit optimal bounds for small numbers of colors and pegs, whereas we deal with asypmtotics when both of these quantities are large. The variants of Mastermind which we study are defined by the following parameters: 1. (\(k\)) *Size of Alphabet* 2. (\(n\)) *Length of Sequence.* The hidden vector and all guess vectors will be elements of \([k]^n\). 3. (\(\Delta\)) *Distance Function.* \(\Delta\) takes as inputs two vectors in \([k]^n\). The output may, for example, be a single integer, but this will not always be the case. Most research studies the two following distance functions: 1. *"Black-peg and white-peg."* Informally, a black peg denotes "the correct color in the correct spot," and a white peg denotes "the correct color in an incorrect spot." For two vectors \(Q_t\) and \(H\), the black-peg and white-peg distance function is the ordered pair \(\Delta(Q_t, H) := (b(Q_t, H), w(Q_t, H))\) where \[\label{blackHitsDefinition} b(Q_t, H) = \left|\{i\in [1,n] \mid q_i = h_i\}\right|,\] and \[\label{whiteHitsDefinition} w(Q_t, H) = \max_{\sigma}~b(\sigma(Q_t), H)-b(Q_t, H),\] where \(\sigma\) iterates over all permutations of \(Q_t\). This variant is the distance function used in the original game of Mastermind. 2. *"Black-peg-only."* This is simply \(\Delta(Q_t, H) := b(Q_t, H)\), where \(b\) is defined as above. 4. (\(R\)) *Repetition.* A commonly-studied variant of the game introduces the restriction that the guesses and vectors cannot have repeated components, i.e. they are vectors of the form \(v \mid i \neq j \Rightarrow v_i \neq v_j\). 5. (\(A\)) *Adaptiveness.* In the *adaptive* variant of Mastermind, the codebreaker receives \(\Delta(Q_t, H)\) after each guess \(Q_t\), and may use this information to inform the selection of \(Q_{t+1}\). In the *non-adaptive* variant, the codebreaker submits any number of queries \(Q_1, Q_2, \ldots, Q_m\) all at once (so the codebreaker chooses \(m\)). The codemaker then reports the distances \((\Delta(Q_1, H), \Delta(Q_2, H), \ldots, \Delta(Q_m, H))\), after which the codebreaker must determine \(H\) without submitting any additional guesses. The adaptive variant with repetitions allowed is the most extensively studied in the literature. Doerr, Spöhel, Thomas, and Winzen obtain both strong asymptotic lower bounds and an asymptotic improvement in algorithm performance, in part applying techniques from. We focus only on analyzing the worst-case performance of query strategies for these variants of Mastermind. That is, we always consider the number of queries necessary to guarantee identification of any hidden vector. As such, we will be only be focusing on deterministic strategies for the codebreaker. Our first main result concerns the *Permutation Variant*, the black-peg, adaptive, no-repeats variant in which \(n=k\). In this variant, the hidden sequence \(H\) and all guesses \(Q_t\) are permutations of \([n]\), hence the name. Explicit algorithms that take \(O(n \log n)\) turns to solve this variant were developed by Ko and Teng, and El Ouali and Sauerland. Ko and Teng approach the problem with an algorithm akin to binary search. El Ouali and Sauerland improve this algorithm and extend it to handle variants with \(k\ge n\), while also achieving an average factor of two reduction in the number of queries needed to identify \(H\). Via a basic information-theoretic argument, one can show that the Permutation Game satisfies \(f(n, n)\ge n-n/\log n + c\) for some constant \(c>0\). We improve this lower bound to \(f(n, n)\ge n-\log \log n\) for sufficiently large \(n\). To our knowledge, this constitutes the first improvement over the trivial information-theoretic lower bound for the Permutation Game variant of Mastermind. Our second result concerns Knuth's Minimax algorithm for adaptive variants, which was first introduced in 1976: Knuth's Minimax algorithm is empirically near-optimal for solving small games of Mastermind (\(n\) and \(k\) less than 10) in as few guesses as possible. However, it has proven difficult to analyze the asymptotic performance of the Minimax algorithm, primarily because its behavior is determined by the distribution remaining solutions after a series of guesses, which is difficult to analyze in general. To our knowledge, this is the first upper bound on the worst-case performance of the minimax algorithm, but if it performs near-optimally for large \(n\) and \(k\), we would expect this bound to be much smaller. We know, for example, that algorithms exist that use only \(O(n \log k)\) guesses when \(k\) is not too large. Our third result concerns to non-adaptive variants of Mastermind. We extend the following theorem: proving the result in the no-repeats case. For neither variant do the known upper bounds match these lower bounds when \(n > k\); in the with-repeats case a corollary of a result in gives an upper bound of \(O(k \log k)\) guesses, and for the no-repeats case no improvement over the \(nk\) bound is known. On the other hand, the authors of are able to extend a result of Chvátal to provide tight bounds for \(n \le k\) with in the with-repeats case. ## Structure of the Paper In Section 2 we prove Theorem 1, and the proof is found in 2.2. In Section 3 we prove Theorem 2, and in Section 4 we discuss Theorem 3, with the proof in 4.2. # Adaptive Variants of Mastermind ## The Permutation Game We begin with notation necessary for the proof of Theorem 1. Recall that this is the adaptive, no-repeats variant with \(n=k\). White pegs provide no information in this variant, so without loss of generality we assume black-peg responses. Throughout this section we will let \(f(n)\) be the number of guesses required by an optimal strategy for the permutation game on \([n]^n\). We will bound \(f(n)\) from below. We will use the derangement function \(D(n)\), which counts the number of permutations in \(S_n\) with no fixed points. It has the explicit form \[\label{derangements} D(n) = n!\sum_{i=0}^n \frac{(-1)^i}{i!},\] and is the nearest integer to \(n!/e\). ### Trivial Lower Bound To motivate the proof of Theorem 1, we begin with the following simple result. ### Solution subsets {#Buckets Definition} We now improve this trivial bound. Assume the codebreaker has some arbitrary fixed, deterministic guessing strategy. In \(t\) turns of the game, the codebreaker has submitted queries \(Q_1, \dots, Q_t\) and received responses \(r_1,\dots,r_t\). The set of solution vectors \(h\) satisfying \(\Delta(Q_i,h) = r_i\) for all \(i\) will be called the *remaining solution set* \(S_t\), and the codebreaker wins exactly when \(|S_t| = 1\). Now we analyze the performance of the codebreaker's strategy. On turn \(t\), the codebreaker makes the guess \(Q_t\) according to some deterministic procedure. Each of the \(n\) possible responses \(\Delta(Q_t,H)\) produces a different remaining solution set \(S_t\). Moreover, these different choices of \(S_t\) partition \(S_{t-1}\), as every vector in \(S_{t-1}\) agrees with exactly one value of \(\Delta(Q_t,H)\). We call the subset of \(S_{t-1}\) corresponding to a response \(r\) a *solution subset*, formally defined as \[B_t(r) = \left\{h\in S_{t-1}\mid \Delta(Q_t, h) = r\right\},\] where \(S_0\) is the set of all permutations of \([n]\). Following this notation, we see that \(S_t = B_{t}(\Delta(Q_t,H))\). It will be useful to know the sizes of the sets \(B(r,Q) := \{h | \Delta(h,Q) = r\}\). The number of elements in \(B(r,Q)\) is equal to the number of ways to choose \(r\) indices that are fixed points with respect to the query sequence \(Q_1\) multiplied by the number of ways to permute the remaining \(n-r\) colors without any fixed points. Hence we have \[|B(r,Q)| = \binom{n}{r}D(n-r).\] So we may define \(B(r) := B(r,Q)\) as the right-hand side is independent of \(Q\). By definition, \(B_t(r)\) for a fixed \(Q_t\) is a subset of \(B(r,Q_t)\), and so \[|B_t(r)|\le |B(r)| = \binom{n}{r}D(n-r).\] ## Proof of Theorem 1 Continuing, we make use of two technical lemmas. The first bounds the sums of sizes of the subsets defined above: This will allow us to prove a bound on the worst-case size of \(|S_t|\). Note that for a fixed strategy and hidden vector \(H\), every query \(Q_t\), response \(r_t\), and remaining solution subset \(S_t\) are completely determined. We prove Lemma 2 at the end of the section.\ ## Proof of Lemma 2 **Update:** El Ouali, Glazik, Sauerland, and Srivastav have announced an improvement of the lower bound in Theorem [\[permutationGameTheorem\]](#permutationGameTheorem){reference-type="ref" reference="permutationGameTheorem"} from \(n-\log\log(n)\) to \(n\), and an extension to \(k > n\) with a lower bound of \(k\). # Linear Algebra and the Minimax Algorithm We now turn to a general upper bound on all Mastermind variants and note its application to Knuth's minimax algorithm in particular. We will represent an arbitrary query or hidden vector as a \((0,1)\)-vector \(A \in \R^{nk}\) in the following manner: \[A_{in+j} = \begin{cases} 1 & \text{this guess/solution assigns the \(i^\text{th}\) spot the \(j^\text{th}\) color} \\ 0 & \text{otherwise}, \end{cases}\] where \(0 \leq i \leq n-1\) and \(0 \leq j \leq k-1\). As such, each set of indices \(A_{in},\dots,A_{in+k-1}\) will have exactly one 1, as the \(i^\text{th}\) position is exactly one color. With this notation, the black-peg distance becomes the dot product of the guess and the hidden vector, as there will be a contribution to the dot product exactly when both vectors have a one in the same spot, i.e. there is the same color in the same spot of both vectors. The goal of Mastermind is then to find the unique valid \((0,1)\)-vector such that its dot product with the hidden vector is \(n\). By linearity of the dot product, once the codebreaker has queried any set of queries, it is possible to deduce the black-peg response to any linear combination of those queries as the corresponding linear combination of their responses. It follows that a winning strategy is simply to query a basis for the span of the set of valid queries, which, as a subspace of \(\R^{nk}\), is of size at most \(nk\). Note that this strategy does not make use of adaptive feedback, of white-peg responses, or the condition on repeated colors, and hence applies to every variant of Mastermind studied in this paper. This logic also allows us to bound the minimax algorithm (Definition [\[minimax def\]](#minimax def){reference-type="ref" reference="minimax def"}). ## Proof of Theorem 2 The theorem can be easily reduced to the following lemma: Theorem [\[minimaxTheorem\]](#minimaxTheorem){reference-type="ref" reference="minimaxTheorem"} follows as an immediate corollary from this lemma: Since these vectors are members of \(\R^{nk}\), the minimax algorithm can make at most \(nk\) linearly independent guesses. After that, the minimax cannot make a linearly independent guess and by the above lemma, the hidden vector must be uniquely determined. # Non-Adaptive Variants This section follows closely the reasoning in as they analyze non-adaptive games. We perform an analysis of the black-peg, non-adaptive, no-repeats variant. We make use of *(Shannon) Entropy*, defined for a random variable \(X\) to be: \[H(X) := \sum_{x\in \text{Domain}(X)}\mathbb{P}[X=x]\cdot (-\log_2(\mathbb{P}[X=x])).\] Entropy is *subadditive*, that is, if \(X_1, X_2, \ldots, X_n\) are random variables and \(X = (X_1,\dots,X_n)\) is the random vector containing the \(X_i\) as entries, then \[\label{subadditivityDefinition} H(X)\le \sum_{i=1}^{n}H(X_i).\] ## Proof of Theorem [\[nonAdaptiveTheorem\]](#nonAdaptiveTheorem){reference-type="ref" reference="nonAdaptiveTheorem"} ## A Note on Non-Adaptive Variants with Repeats Consider non-adaptive variants of Mastermind in which repetitions are allowed. When \(k=n\), Theorem 13 of guarantees the existence of a set of \(O(n \log n)\) queries which uniquely identify any hidden sequence \(H\). When \(k > n\), one can simply extend \(H\) and all queries \(Q_t\) by \(k-n\) "auxiliary" positions. We fill these auxiliary positions with arbitrary colors, and adjust the codemaker's responses accordingly. Applying Theorem 13 of now guarantees the existence of a set of \(O(k \log k)\) queries which will uniquely identify any hidden sequence. # Further Work
{'timestamp': '2017-09-27T02:03:53', 'yymm': '1607', 'arxiv_id': '1607.04597', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04597'}
null
null
null
null
null
null
# Introduction {#sec:intro} A major challenge towards self-organizing networks (SON) is the joint optimization of multiple SON use cases by coordinately handling multiple configuration parameters. Widely studied SON use cases include coverage and capacity optimization (CCO), mobility load balancing (MLB) and mobility robustness optimization (MRO). However, most of these works study an isolated single use case and ignore the conflicts or interactions between the use cases . In contrast, this paper considers a joint optimization of two strongly coupled use cases: CCO and MLB. The objective is to achieve a good trade-off between coverage and capacity performance, while ensuring a load-balanced network. The SON functionalities are usually implemented at the network management layer and are designed to deal with ''long-term'' network performance. Short-term optimization of individual users is left to lower layers of the protocol stack. To capture long-term global changes in a network, we consider a cluster-based network scenario, where users served by the same base station (BS) with similar SINR distribution are adaptively grouped into clusters. Our objective is to jointly optimizing the following variables: - Cluster-based BS assignment and power allocation. - BS-based antenna tilt optimization and power allocation. The joint optimization of assignment, antenna tilts, and powers is an inherently challenging problem. The interference and the resulting performance measures depend on these variables in a complex and intertwined manner. Such a problem, to the best of the authors' knowledge, has been studied in only a few works. For example, in a problem of jointly optimizing antenna tilt and cell selection to improve the spectral and energy efficiency is stated, however, the solution derived by a structured searching algorithm may not be optimal. In this paper, we propose a robust algorithmic framework built on a utility model, which enables fast and near-optimal uplink solutions and sub-optimal downlink solutions by exploiting three properties: 1) the monotonic property and fixed point of the monotone and strictly subhomogenoues (MSS) functions [^1], 2) decoupled property of the antenna tilt and BS assignment optimization in the uplink network, and 3) uplink-downlink duality. The first property admits global optimal solution with fixed-point iteration for two specific problems: utility-constrained power minimization and power-constrained max-min utility balancing . The second and third properties enable decomposition of the high-dimensional optimization problem, such as the joint beamforming and power control proposed in . Our distinct contributions in this work can be summarized as follows:\ 1) We propose a max-min utility balancing algorithm for capacity-coverage trade-off optimization over a joint space of antenna tilts, BS assignments and powers. The utility defined as a convex combination of the average SINR and the worst-case SINR implies the balanced performance of capacity and coverage. Load balancing is improved as well due to a uniform distribution of the interference among the BSs.\ 2) The proposed utility is formulated based on the MSS functions, which allows us to find the optimal solution by applying fixed-point iterations.\ 3) Note that antenna tilts are BS-specific variables, while assignments are cluster-specific, we develop two optimization problems with the same objective functions, formulated either as a problem of per-cluster variables or as a problem of per-base variables. We propose a two-step optimization algorithm in the uplink to iteratively optimize the per BS variables (antenna tilts and BS power budgets) and the cluster-based variables (assignments and cluster power). Since both problems aim at optimizing the same objective function, the algorithm is shown to be convergent.\ 4) The decoupled property of antenna tilt and assignment in the uplink decomposes the high-dimensional optimization problem and enables more efficient optimization algorithm. We then analyze the uplink-downlink duality by using the Perron-Frobenius theory, and propose an efficient sub-optimal solution in the downlink by utilizing optimized variables in the dual uplink. # System Model {#sec:Model} We consider a multicell wireless network composed of a set of BSs \(\set{N}:=\{1,\ldots, N\}\) and a set of users \(\set{K}:=\{1,\ldots, K\}\). Using fuzzy C-means clustering algorithm, we group users with similar SINR distributions[^2] and served by the same BS into clusters. The clustering algorithm is beyond the scope of this paper. Let the set of user clusters be denoted by \(\set{C}:=\{1,\ldots,C\}\), and let \(\bm{A}\) denote a \(C\times K\) binary user/cluster assignment matrix whose columns sum to one. The BS/cluster assignment is defined by a \(N\times C\) binary matrix \(\bm{B}\) whose columns also sum to one. Throughout the paper, we assume a frequency flat channel. The average/long-term downlink path attenuation between \(N\) BSs and \(K\) users are collected in a channel gain matrix \(\bm{H}\in {\field{R}}^{N\times K}\). We introduce the cross-link gain matrix \(\bm{V}\in{\field{R}}^{K\times K}\), where the entry \(v_{lk}(\theta_j)\) is the cross-link gain between user \(l\) served by BS \(j\), and user \(k\) served by BS \(i\), i.e., between the transmitter of the link \((j, l)\) and the receiver of the link \((i, k)\). Note that \(v_{lk}(\theta_j)\) depends on the antenna downtilt \(\theta_j\). Let the BS/user assignment matrix be denoted by \(\bm{J}\) so that we have \(\bm{J}:=\bm{B}\bm{A}\in\{0,1\}^{N\times K}\), and \(\bm{V}:=\bm{J}^T\bm{H}\). We denote by \(\bm{r}:=[r_1, \ldots, r_N]^T\), \(\bm{q}:=[q_1, \ldots, q_C]^T\) and \(\bm{p}:=[p_1, \ldots, p_K]^T\)the BS transmission power budget, the cluster power allocation and the user power allocation, respectively. [^1]: Many literatures use the term *interference function* for the functions satisfy three condotions, positivity, monotonicity and scalability. Positivity is shown to be a consequence of the other two properties, and we use the term *strctly subhomogeneous* in place of scalable from a constraction mapping point of view in keeping with some related literature. [^2]: We assume the Kullback-Leibler divergence as the distance metric.
{'timestamp': '2016-07-19T02:04:46', 'yymm': '1607', 'arxiv_id': '1607.04754', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04754'}
# Introduction Cloning algorithms are numerical procedures aimed at simulating rare events efficiently, using a population dynamics scheme. In such algorithms, copies of the system are evolved in parallel and the ones showing the rare behavior of interest are multiplied iteratively [@Anderson1975_63_4; @Glasserman_1996; @2001IbaYukito; @GRASSBERGER200264; @4117599; @CappeGuillinetal; @Forwardinterfacesampling; @PhysRevLett.94.018104; @delmoral2005; @Dean2009562; @GuyaderArnaudetal; @giardina_direct_2006; @lecomte_numerical_2007; @JulienNaturephys_2007; @tailleur_simulation_2009; @giardina_simulating_2011; @1751-8121-46-25-254002_2013; @1751-8121-49-20-205002_2016] (See Fig. [\[fig:traj\]](#fig:traj){reference-type="ref" reference="fig:traj"}). One of these algorithms proposed by Giardinà *et al.* [@giardina_direct_2006; @lecomte_numerical_2007; @JulienNaturephys_2007; @tailleur_simulation_2009; @giardina_simulating_2011; @1751-8121-46-25-254002_2013; @1751-8121-49-20-205002_2016] is used to evaluate numerically the cumulant generating function (a large deviation function, LDF) of additive (or "time-extensive") observables in Markov processes . It has been applied to many physical systems, including chaotic systems, glassy dynamics and non-equilibrium lattice gas models, and it has allowed the study of novel properties, such as the behavior of breathers in the Fermi-Pasta-Ulam-Tsingou chain , dynamical phase transitions in kinetically constrained models , and an additivity principle for simple exclusion processes . While the method has been used widely, there have been fewer studies focusing on the analytical justification of the algorithm. Even though it is heuristically believed that the LDF estimator converges to the correct result as the number of copies \(N_c\) increases, there is no proof of this convergence. Related to this lack of the proof, although we use the algorithm by assuming its validity, we do not have any clue how fast the estimator converges as \(N_c \rightarrow \infty\). In order to discuss this convergence, we define two types of numerical errors. First, for a fixed finite \(N_c\), averaging over a large number of realizations, the LDF estimator converges to an incorrect value, which is different from the desired large deviation result. We call this deviation from the correct value, *systematic errors*. Compared with these errors, we also consider the fluctuations of the estimated value. More precisely, for a fixed value of \(N_c\), the results obtained in different realizations are distributed around this incorrect value. We call the errors associated to these fluctuations the *stochastic errors*. Although both errors are important in numerical simulations, the former one can lead this algorithm to produce wrong results. For example as seen in Ref. , the systematic error grows exponentially as a temperature decreases (or generically in the weak noise limit of diffusive dynamics). In order to study these errors, we employ a birth-death process  description of the population dynamics algorithm as explained below: We focus on physical systems described by a Markov dynamics  with a finite number of states \(M\), and we denote by \(i\) (\(i=0,1,\cdots M-1\)) the states of the system. This Markov process has its own stochastic dynamics, described by the transition rates \(w(i\rightarrow j)\). In population dynamics algorithms, in order to study its rare trajectories, one prepares \(N_c\) copies of the system, and simulate these copies according to *(i)* the dynamics of \(w(i\rightarrow j)\) (followed independently by all copies) and *(ii)* 'cloning' step in which the ensemble of copies is directly manipulated, *i.e.*, some copies are eliminated while some are multiplied (See Table [\[tablecorrespondence\_\]](#tablecorrespondence_){reference-type="ref" reference="tablecorrespondence_"}). Formally, the population dynamics represents, for a *single* copy of the system, a process that does not preserve probability. This fact has motivated the studies of auxiliary processes , effective processes  and driven processes  to construct modified dynamics (and their approximations ) that preserve probability. Different from these methods, in this article, we formulate explicitly the meta-dynamics of the copies themselves by using a stochastic birth-death process. The process preserves probability, and it allows us to study the numerical errors of the algorithm when evaluating LDF.
{'timestamp': '2017-05-11T02:06:08', 'yymm': '1607', 'arxiv_id': '1607.04752', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04752'}
null
null
null
null
# Introduction While recurrent neural networks (RNNs) are seeing widespread success across many tasks, the fundamental architecture presents challenges to typical training algorithms. In particular, the problem of 'vanishing/exploding gradients'  in gradient-based optimization persists, where gradients either vanish or diverge as one goes deeper into the network, resulting in slow training or numerical instability. The long short-term memory (LSTM) network  was designed to overcome this issue. Recently, the use of norm-preserving operators in the transition matrix-the matrix of weights connecting subsequent *internal* states-of the RNN have been explored . Using operators with bounded eigenvalue spectrum should, as demonstrated by  , bound the norms of the gradients in the network, assuming an appropriate non-linearity is applied. Unitary matrices satisfy this requirement and are the focus of this work. Imposing unitarity (or orthogonality) on this transition matrix is however challenging for gradient-based optimization methods, as additive updates typically do not preserve unitarity. Solutions include *re-unitarizing* after each batch, or using a *parametrization* of unitary matrices closed under addition. In this work we propose a solution in the second category, using results from the theory of Lie algebras and Lie groups to define a general parametrization of unitary matrices in terms of skew-Hermitian matrices (elements of the Lie algebra associated to the Lie group of unitary matrices). As explained in more detail below, elements of this Lie algebra can be identified with unitary matrices, while the algebra is closed under addition, forming a vector space over real numbers. While we are motivated by the issues of RNNs, and we consider an application in RNNs, our primary focus here is on a core question: how can unitary matrices be learned? *Assuming* the choice of a unitary transition matrix is an appropriate modelling choice, the gradients on this operator should ultimately guide it towards unitarity, if it is possible under the parametrization or learning scheme used. It is therefore useful to know which approach is best. We distil the problem into its simplest form (a learning task described in detail later), so that our findings cannot be confounded by other factors specific to the RNN long-term memory task, before demonstrating our parametrization in that setting. ## Related work We draw most inspiration from the recent work of  , who proposed a specific parametrization of unitary matrices and demonstrated its utility in standard long-term memory tasks for RNNs. We describe their parametrization here, as we use it later. Citing the difficulty of obtaining a general and efficient parametrization of unitary matrices, they use the fact that the unitary group is closed under matrix multiplication to form a composite operator: \[U = \mathbf{D}_3 \mathbf{R}_2 \mathcal{F}^{-1} \mathbf{D}_2 \Pi \mathbf{R}_1 \mathcal{F} \mathbf{D}_1 \label{eqn:arjovsky}\] where each component is unitary and easily parametrized: - \(\mathbf{D}\) is a diagonal matrix with entries of the form \(e^{i \alpha}\), \(\alpha \in \mathbb{R}\) - \(\mathbf{R}\) is a complex reflection operator; \(\mathbf{R} = \mathbb{I}-2\frac{\mathbf{v}\mathbf{v}^{\dagger}}{\| \mathbf{v} \|^2}\) (\(\dagger\) denotes Hermitian conjugate) - \(\mathcal{F}\) and \(\mathcal{F}^{-1}\) are the Fourier and inverse Fourier transforms (or, in practice, their discrete matrix representations) - \(\Pi\) is a fixed permutation matrix In total, this parametrization has \(7n\) real learnable parameters (\(2n\) for each reflection and \(n\) for each diagonal operator), so describes a subspace of unitary matrices (which have \(n^2\) real parameters). Nonetheless, they find that an RNN using this operator as its transition matrix outperforms LSTMs on the adding and memory tasks described first in  . This prompted us to consider other parametrizations of unitary matrices which might be more expressive or interpretable.   constrain a part of the transition matrix to be close to the identity, acting as a form of long-term memory store, while   *initialize* it to the identity, and then use ReLUs as non-linearities.   study analytic solutions to the long-term memory task, supporting observations and intuitions that orthogonal (or unitary) matrices would be appropriate as transition matrices for this task. They also study initializations to orthogonal and identity matrices, and consider experiments where an additional term in the loss function encourages an orthogonal solution to the transition matrix, without using an explicit parametrization.   study exact solutions to learning dynamics in deep networks and find that orthogonal weight initializations at each layer lead to depth-independent learning (thus escaping the vanishing/exploding gradient problem). Interestingly, they attribute this to the eigenvalue spectrum of orthogonal matrices lying on the unit circle. They compare with weights initialized to random, scaled Gaussian values, which preserve norms in expectation (over values of the random matrix) and find orthogonal matrices superior. It therefore appears that preserving norms is *not* sufficient to stabilize gradients over network depth, but that the eigenvalue spectrum must also be strictly controlled. In a related but separate vein,   penalize the difference of difference of norms between subsequent hidden states in the network. This is *not* equivalent to imposing orthogonality of the *transition* matrix, as the norm of the hidden state may be influenced by the inputs and non-linearities, and their method directly addresses this norm. The theory of Lie groups and Lie algebras has seen most application in machine learning for its use in capturing notions of *invariance*. For example,  , learn infinitesimal Lie group generators (elements of the Lie algebra) associated with affine transformations of images, corresponding to visual perceptual invariances. This is different to our setting as our generators are already known (we assume the Lie group \(U(n)\)) and wish to learn the coefficients of a given transformation relative to that basis set of generators. However, our approach could be extended to the case where the basis of \(\mathfrak{u}(n)\) is *unknown*, and must be learned. As we find later (appendix B), the choice of basis can impact performance, and so may be an important consideration.   learn commutative subgroups of \(SO(n)\) (known as toroidal subgroups), motivated by learning the irreducible representations of the symmetry group corresponding to invariant properties of images. Their choice of group parametrization is equivalent to selecting a particular basis of the corresponding Lie algebra, as they describe, but primarily exploit the algebra to understand properties of toroidal subgroups.   perform motion estimation by defining a regression function in terms of a function on the Lie algebra of affine transformations, and then learning this. This is similar to our approach in the sense that they do optimization in the Lie algebra, although as they consider two-dimensional affine transformations only, their parametrization of the Lie algebra is straight forward. Finally,   describe an online learning algorithm for orthogonal matrices--which are the real-valued equivalent to unitary matrices. They also claim that the approach is extends easily to unitary matrices. ## Structure of this paper We begin with an introduction to the relevant facts and definitions from the theory of Lie groups and Lie algebras, to properly situate this work in its mathematical context. Further exposition is beyond the scope of this paper, and we refer the interested reader to any of the comprehensive introductory texts on the matter. We explain our parametrization in detail and describe a method to calculate the derivative of the matrix exponential-a quantity otherwise computationally intractable. Then, we describe a simple but clear experiment designed to test our core question of learning unitary matrices. We compare to an approach using the parametrization of   and one using polar decomposition to 'back-project' to the closest unitary matrix. We use this experimental set-up to probe aspects of our model, studying the importance of the choice of basis (appendix B), and the impact of the restricted parameter set used by one of the alternate approaches. We additionally implement our parametrization in a recurrent neural network as a 'general unitary RNN', and evaluate its performance on standard long-memory tasks. # The Lie algebra \(\mathfrak{u}(n)\) {#section:lie} ## Basics of Lie groups and Lie algebras A Lie group is a group which is also a differentiable manifold, with elements of the group corresponding to points on the manifold. The group operations (multiplication and inversion) must be smooth maps (infinitely differentiable) back to the group. In this work we consider the group \(U(n)\): the set of \(n \times n\) unitary matrices, with matrix multiplication. These are the complex-valued analogue to orthogonal matrices, satisfying the property \[U^{\dagger} U = U U^{\dagger} = \mathbb{I} \label{eqn:unitary}\] where \(\dagger\) denotes the conjugate transpose (or Hermitian conjugate). Unitary matrices preserve matrix norms, and have eigenvalues lying on the (complex) unit circle, which is the desired property of the transition matrix in a RNN. The differentiable manifold property of Lie groups opens the door for the study of the Lie *algebra*. This object is the *tangent space* to the Lie group at the identity (the group must have an identity element). Consider a curve through the Lie group \(U(n)\)-a one-dimensional subspace parametrized by a variable \(t\), where \(U(t=0) = \mathbb{I}\) (this is a matrix \(U(t)\) in \(U(n)\) parametrised by \(t\), not a group). Consider the defining property of unitary matrices (Equation [\[eqn:unitary\]](#eqn:unitary){reference-type="ref" reference="eqn:unitary"}), and take the derivative along this curve: \[U(t)^{\dagger}U(t) = \mathbb{I} \rightarrow \dot{U}(t)^{\dagger} U(t) + U^{\dagger}(t) \dot{U}(t) = 0\] Taking \(t \rightarrow 0\), \(U(t) \rightarrow \mathbb{I}\), we have \[\dot{U}(0)^{\dagger} \mathbb{I} + \mathbb{I}^{\dagger} \dot{U}(0) = 0 \Rightarrow \dot{U}(0)^{\dagger} =-\dot{U}(0) \label{eqn:alg}\] The elements \(\dot{U}(0)\) belong to the Lie algebra. We refer to this Lie algebra as \(\mathfrak{u}(n)\), and an arbitrary element as \(L\). Then Equation [\[eqn:alg\]](#eqn:alg){reference-type="ref" reference="eqn:alg"} defines the properites of these Lie algebra elements; they are \(n \times n\) skew-Hermitian matrices: \(L^{\dagger} =-L\). As vector spaces, Lie algebras are closed under addition. In particular \(\mathfrak{u}(n)\) is a vector space over \(\mathbb{R}\), so a *real* linear combination of its elements is once again in \(\mathfrak{u}(n)\) (this is also clear from the definition of skew-Hermitian). We exploit this fact later. Lie algebras are also endowed with an operation known as the *Lie bracket*, which has many interesting properties, but is beyond the scope of this work. Lie algebras are interesting algebraic objects and have been studied deeply, but in this work we use \(\mathfrak{u}(n)\) because of the *exponential map*. Above, it was shown that elements of the algebra can be derived from the group (considering infinitesimal steps away from the identity). There is a *reverse* operation, allowing elements of the group to be recovered from the algebra: this is the *exponential map*. In the case of matrix groups, the exponential map is simply the *matrix exponential*: \[\exp(L) = \sum_{j = 0}^{\infty}\frac{L^j}{j!}\] Very simply, \(L \in \mathfrak{u}(n)\), then \(\exp(L) \in U(n)\). While this map is not in *general* surjective, it so happens that \(U(n)\) is a compact, connected group and so \(\exp\) *is* indeed surjective. That is, for any \(U \in U(n)\), there exists *some* \(L \in \mathfrak{u}(n)\) such that \(\exp(L) = U\). Notably, while orthogonal matrices also form a Lie group \(O(n)\), with associated Lie algebra \(\mathfrak{o}(n)\) consisting of skew-symmetric matrices, \(O(n)\) is *not* connected, and so the exponential map can only produce *special* orthogonal matrices-those with determinant one-\(SO(n)\) being the component of \(O(n)\) containing the identity. ## Parametrization of \(U(n)\) in terms of \(\mathfrak{u}(n)\) The dimension of \(\mathfrak{u}(n)\) as a real vector space is \(n^2\). This is readily derived from noting that an arbitrary \(n \times n\) complex matrix has \(2n^2\) free real parameters, and the requirement of \(L^{\dagger} =-L\) imposes \(n^2\) constraints. So, a set of \(n^2\) linearly-independent skew-Hermitian matrices defines a basis for the space; \(\{ T_j \}_{j = \{1, \dots, n^2 \}}\). Then any element \(L\) can be written as \[L = \sum_{j=1}^{n^2} \lambda_j T_j\] where \(\{\lambda_j\}_{j = 1, \dots, n^2}\) are \(n^2\) real numbers; the coefficients of \(L\) with respect to the basis. Using the exponential map, \[U = \exp(L) = \exp\left(\sum_{j=1}^{n^2} \lambda_j T_j \right) \label{eqn:parametrization}\] we see that these \(\{\lambda_j\}_{j = 1, \dots, n^2}\) suffice as *parameters* of \(U\) (given the basis \(T_j\)). This is the parametrization we propose. It has two attractive properties: 1. It is a fully general parametrization, as the exponential map is surjective 2. Gradient updates on \(\{\lambda_j\}_{j = 1, \dots, n^2}\) preserve unitarity automatically, as the algebra is closed under addition This parametrization means gradient steps are taken in the vector space of \(\mathfrak{u}(n)\), rather than the manifold of \(U(n)\), which *may* provide a flatter cost landscape-although confirming this intuition would require further analysis. This work is intended to explore the use of this parametrization for learning arbitrary unitary matrices. There are many possible choices of basis for \(\mathfrak{u}(n)\). We went for the following set of sparse matrices: 1. \(n\) diagonal, imaginary matrices: \(T_a\) is \(i\) on the \(a\)-th diagonal, else zero. 2. \(\frac{n(n-1)}{2}\) symmetric, imaginary matrices with two non-zero elements, e.g., for \(n=2\), \(\left(\begin{matrix} 0 & i \\ i & 0 \end{matrix}\right)\) 3. \(\frac{n(n-1)}{2}\) anti-symmetric, real matrices with two non-zero elements, e.g., for \(n=2\), \(\left(\begin{matrix} 0 & 1 \\-1 & 0 \end{matrix}\right)\) We explore the effects of choice of basis in appendix B. ## Derivatives of the matrix exponential The matrix exponential appearing in Equation [\[eqn:parametrization\]](#eqn:parametrization){reference-type="ref" reference="eqn:parametrization"} poses an issue for gradient calculations. In general, the derivative of the matrix exponential does not have a closed-form expression, so computing gradients is intractable. In early stages of this work, we used the method of finite differences to approximate gradients, which would prohibit its use in larger-scale applications (such as RNNs). In the appendix we describe an investigation into using random projections to overcome this limitation, which while promising turned out to yield minimal benefit. We therefore sought mathematical solutions to this complexity issue, which we describe here and in further detail in the appendix. Exploiting the fact that \(L\) is skew-Hermitian, we can derive an analytical expression for the derivative of \(U\) with respect to each of its parameters, negating the need for finite differences. This expression takes the form: \[\frac{\partial U}{\partial \lambda_a} = W V_a W^{\dagger} \label{eqn:twomat}\] where \(W\) is a unitary matrix of eigenvectors obtained in the eigenvalue decomposition of \(U\); \(U = W D W^{\dagger}\), (\(D\) = diag(\(d_1\), \(\dots\), \(d_{n^2}\)); \(d_i\) are the eigenvalues of \(U\)). Each \(V_a\) is a matrix defined component-wise \[\begin{aligned} i = j:& V_{ii} =(W^{\dagger} T_a W)_{ii} e^{d_i}\label{eqn:vii}\\ i \neq j:& V_{ij} = (W^{\dagger} T_a W)_{ij} \left(\frac{e^{d_i}-e^{d_j}}{d_i-d_j}\right)\label{eqn:vij} \end{aligned}\] Where \(T_a\) is the basis matrix of the Lie algebra in the \(a\)-th direction. We provide the derivation, based on work from   and   in Appendix A. We can simplify the expression \(W^{\dagger} T_a W\) for each \(T_a\), depending on the type of basis element. In these expressions, \(\mathbf{w}_a\) refers to the \(a\)-th *row* of W. 1. \(T_a\) purely imaginary; \(W^{\dagger} T_a W = i \cdot \mathrm{outer}(\mathbf{w}^*_a, \mathbf{w}_a)\) 2. \(T_a\) symmetric imaginary, nonzero in positions \((r, s)\) and \((s, r)\): \(W^{\dagger} T_{rs} W = i \cdot (\mathrm{outer}(\mathbf{w}^*_s, \mathbf{w}_r) + \mathrm{outer}(\mathbf{w}^*_r, \mathbf{w}_s))\) 3. \(T_a\) antisymmetric real, nonzero in positions \((r, s)\) and \((s, r)\): \(W^{\dagger} T_{rs} W = \mathrm{outer}(\mathbf{w}^*_r, \mathbf{w}_s)-\mathrm{outer}(\mathbf{w}^*_s, \mathbf{w}_r)\) These expressions follow from the sparsity of the basis and are derived in appendix A. Thus, we reduce the calculation of \(W^{\dagger} T_a W\) from two matrix multplications to at most two vector outer products. Overall, we have reduced the cost of calculating gradients to a single eigenvalue decomposition, and for each parameter two matrix multiplications (equation [\[eqn:twomat\]](#eqn:twomat){reference-type="ref" reference="eqn:twomat"}), one or two vector outer products, and element-wise multiplication of two matrices (equations [\[eqn:vii\]](#eqn:vii){reference-type="ref" reference="eqn:vii"}, [\[eqn:vij\]](#eqn:vij){reference-type="ref" reference="eqn:vij"}). As we see in the RNN experiments, this actually makes our approach faster than the (restricted)uRNN of  for roughly equivalent numbers of parameters. # Supervised Learning of Unitary Operators {#section:learning} We consider the supervised learning problem of learning the unitary matrix \(U\) that generated a \(\mathbf{y}\) from \(\mathbf{x}\); \(\mathbf{y} = U\mathbf{x}\), given examples of such \(\mathbf{x}\)s and \(\mathbf{y}\)s. This is the core learning problem that needs to be solved for the state-transformation matrix in RNNs. It is similar to the setting considered in   (they consider an online learning problem). We compare a number of methods for learning \(U\) at different values of \(n\). We further consider the case where we have artificially restricted the number of learnable variables in our parametrization (for the sake of comparison), and generate a pathological change of basis to demonstrate the relevance of selecting a good basis (appendix B). ## Task The experimental setup is as follows: we create a \(n \times n\) unitary matrix \(U\) (the next section describes how this is done), then sample vectors \(\mathbf{x} \in \mathbb{C}^n\) with normally-distributed coefficients. We create \(\mathbf{y}_j = U \mathbf{x}_j + \epsilon_j\) where \(\epsilon \sim \mathcal{N}(0, \sigma^2)\). The objective is to recover \(U\) from the \(\{\mathbf{x}_j, \mathbf{y}_j\}\) pairs by minimizing the squared Euclidean distance between predicted and true \(\mathbf{y}\) values; \[U = \mathop{\mathrm{argmin}}_U \frac{1}{N} \sum_j^N \| \hat{\mathbf{y_j}}-\mathbf{y_j} \|^2 = \mathop{\mathrm{argmin}}_U \frac{1}{N} \sum_j^N \| U \mathbf{x}_j-\mathbf{y_j} \|^2\] While this problem is easily solved in the batch setting using least-squares, we wish to learn \(U\) through mini-batch stochastic gradient descent, to emulate a deep learning scenario. For each experimental run (a single \(U\)), we generate one million training \(\{\mathbf{x}_j, \mathbf{y}_j\}\) pairs, divided into batches of size 20. The test and validation sets both contain \(100,000\) examples. In practice we set \(\sigma^2 = 0.01\) and use a fixed learning rate of \(0.001\). For larger dimensions, we run the model through the data for multiple epochs, shuffling and re-batching each time. All experiments were implemented in Python. The code is available here: `https://github.com/ratschlab/uRNN`. For the matrix exponential, we use the scipy builtin `expm`, which uses Pade approximation . We make use of the fact that \(iL\) is Hermitian to use `eigh` (also in scipy) to perform eigenvalue decompositions. ## Generating the ground-truth unitary matrix {#section:generating} The \(U\) we wish to recover is generated by one of three methods: 1. QR decomposition: we create a \(n \times n\) complex matrix with normally-distributed entries and then perform a QR decomposition, producing a unitary matrix \(U\) and an upper triangular matrix (which is discarded). This approach is also used to sample orthogonal matrices in  , noting a result from   demonstrating that this is equivalent to sampling from the appropriate Haar measure. 2. Lie algebra: given the standard basis of \(\mathfrak{u}(n)\), we sample \(n^2\) normally-distributed real \(\lambda_j\) to produce \(U = \exp{\left( \sum_j \lambda_j T_j\right)}\) 3. Unitary composition: we compose parametrized unitary operators as in    (Equation [\[eqn:arjovsky\]](#eqn:arjovsky){reference-type="ref" reference="eqn:arjovsky"}). The parameters are sampled as follows: angles in \(D\) come from \(\mathcal{U}(-\pi, \pi)\). The complex reflection vectors in \(\mathbf{R}\) come from \(\mathcal{U}(-s, s)\) where \(s = \sqrt{\frac{6}{2n}}\). We study the effects of this generation method on test-set loss in a later section. While we find no significant association between generation method and learning approach, in our experiments we nonetheless average over an equal number of experiments using each method, to compensate for possible unseen bias. ## Approaches We compare the following approaches for learning \(U\): 1. `projection`: \(U\) is represented as an unconstrained \(n \times n\) complex matrix, but after each gradient update we *project* it to the closest unitary matrix, using polar decomposition . This amounts to \(2n^2\) real parameters. 2. `arjovsky`: \(U\) is parametrized as in Equation [\[eqn:arjovsky\]](#eqn:arjovsky){reference-type="ref" reference="eqn:arjovsky"}, which comes to \(7n\) real parameters. 3. `lie_algebra`: (we refer to this as \(\mathfrak{u}(n)\)) \(U\) is parametrized by its \(n^2\) real coefficients \(\{ \lambda_j \}\) in the Lie algebra, as in Equation [\[eqn:parametrization\]](#eqn:parametrization){reference-type="ref" reference="eqn:parametrization"}. As baselines we use the `true` matrix \(U\), and a `random` unitary matrix \(U_R\) generated by the same method as \(U\) (in that experimental run). We also implemented the algorithm described in   and considered both unitary and orthogonal learning tasks (our parametrization contains orthogonal matrices as a special case) but found it too numerically unstable and therefore excluded it from our analyses. ## Comparison of Approaches {#section:results} Table [\[table:accuracies\]](#table:accuracies){reference-type="ref" reference="table:accuracies"} shows the test-set loss for different values of \(n\) and different approaches for learning \(U\). We performed between 6 and 18 replicates of each experiment, and show bootstrap estimates of means and standard errors over these replicates. As we can see, the learning task becomes more challenging as \(n\) increases, but our parametrization (\(\mathfrak{u}(n)\)) consistently outperforms the other approaches. ## Restricting to \(7n\) parameters {#section:restriction} As mentioned, `arjovsky` uses only \(7n\) parameters. To check if this difference accounts for the differences in loss observed in Table [\[table:accuracies\]](#table:accuracies){reference-type="ref" reference="table:accuracies"}, we ran experiments where we fixed all but \(7n\) (selected randomly) of the \(\{\lambda_j\}\) in the `lie_algebra` parametrization. The fixed parameters retained their initial values throughout the experiment. We observe that, as suspected, restricting to \(7n\) parameters results in a performance degradation equivalent to that of `arjovsky`. Table [\[table:restrict\]](#table:restrict){reference-type="ref" reference="table:restrict"} shows the results for \(n=8, 14, 20\). The fact that the restricted case is consistently within error of the `arjovsky` model supports our hypothesis that the difference in learnable parameters accounts for the difference in performance. This suggests that generalising the model of to allow for \(n^2\) parameters may result in performance similar to our approach. However, how to go about such a generalisation is unclear, as a naive approach would simply use a composition of \(n^2\) operators, and this would likely become computationally intractable. ## Method of generating \(U\) {#section:method} As described, we used three methods to generate the true \(U\). One of these produces \(U\) in the subspace available to the composition parametrization (Equation [\[eqn:arjovsky\]](#eqn:arjovsky){reference-type="ref" reference="eqn:arjovsky"}), so we were curious to see if this parametrization performed better on experiments using that method. We were also concerned that generating \(U\) using the Lie algebra parametrization might make the task too 'easy' for our approach, as its random initialization could lie close to the true solution. Figure [\[fig:boxplots\]](#fig:boxplots){reference-type="ref" reference="fig:boxplots"} shows box-plots of the distribution of test losses from these approaches for the three methods, comparing our approach (\(\mathfrak{u}(n)\)) with that of  , denoted `arjovsky`. To combine results from experiments using different values of \(n\), we first scaled test-set losses by the performance of `rand` (the random unitary matrix), so the y-axis ranges from 0 (perfect) to 1 (random performance). The dotted line denotes the average (over methods) of the test-set loss for `true`, similarly scaled. The right panel in Figure [\[fig:boxplots\]](#fig:boxplots){reference-type="ref" reference="fig:boxplots"} shows a zoomed-in version of the \(\mathfrak{u}(n)\) result where the comparison with `true` is more meaningful, and a comparison with the case where we have restricted to \(7n\) learnable parameters (see earlier). We do not observe a difference (within error) between the methods, which is consistent between \(\mathfrak{u}(n)\) and `arjovsky`. Our concern that using the Lie algebra to generate \(U\) would make the task 'too easy' for \(\mathfrak{u}(n)\) was seemingly unfounded. # Unitary Recurrent Neural Network for Long Memory Tasks To demonstrate that our approach is practical for use in deep learning, we incorporate it into a recurrent neural network to solve standard long-memory tasks. Specifically, we define a general unitary RNN with recurrence relation \[\mathbf{h}_t = f\left( \beta U \mathbf{h}_{t-1} + V \mathbf{x}_t + \mathbf{b} \right)\] where \(f\) is a nonlinearity, \(\beta\) is a free scaling factor, \(U\) is our unitary matrix parametrised as in equation [\[eqn:parametrization\]](#eqn:parametrization){reference-type="ref" reference="eqn:parametrization"}, \(\mathbf{h}_t\) is the hidden state of the RNN and \(\mathbf{x}_t\) is the input data at 'time-point' \(t\). We refer to this as a 'general unitary RNN' (guRNN), to distinguish it from the restricted uRNN of  . We use the guRNN on two tasks: the 'adding problem' and the 'memory problem', first described in . For the sake of brevity we refer to  for specific experimental details, as we use an identical experimental setup (reproduced in TensorFlow; see above github link for code). We compare our model (guRNN) with the restricted uRNN (ruRNN) parametrised as in equation [\[eqn:arjovsky\]](#eqn:arjovsky){reference-type="ref" reference="eqn:arjovsky"}, a LSTM , and the IRNN of  . Figure [\[fig:RNN\]](#fig:RNN){reference-type="ref" reference="fig:RNN"} shows the results for each task where the sequence length or the memory duration is \(T=100\). While our model guarantees unitarity of \(U\), this is *not* sufficient to prevent gradients from vanishing. Consider the norm of the gradient of the cost \(C\) with respect to the data at time \(\tau\), and use submultiplicativity of the norm to write; \[\left\| \frac{\partial C}{\partial \mathbf{x}_\tau} \right\| \leq \left\| \frac{\partial C}{\partial \mathbf{x}_T} \right\| \left( \prod_{t = \tau}^{T-1} \| f' \left( U \mathbf{h}_t + V \mathbf{x}_t + \mathbf{b}\right) \| \| U \|\right) \left\|\frac{\partial \mathbf{h}_\tau}{\mathbf{x}_\tau} \right\|\] where \(f'\) is a diagonal matrix giving the derivatives of the nonlinearity. Using a unitary matrix fixes \(\| U \| = 1\), but beyond further restrictions (on \(V\) and \(\mathbf{b}\)) does nothing to control the norm of \(f'\), which is at *most* 1 for common nonlinearities. Designing a nonlinearity to better preserve gradient norms is beyond the scope of this work, so we simply scaled \(U\) by a constant multiplicative factor \(\beta\) to counteract the tendency of the nonlinearity to shrink gradients. In Figure [\[fig:RNN\]](#fig:RNN){reference-type="ref" reference="fig:RNN"} we denote this setup by guRNN\(_\beta\). Confirming our intuition, this simple modification greatly improves performance on both tasks. Perhaps owing to our efficient gradient calculation (appendix A) and simpler recurrence relation, our model runs faster than that of  (in our implementation), by a factor of 4.8 and 2.6 in the adding and memory tasks shown in Figure [\[fig:RNN\]](#fig:RNN){reference-type="ref" reference="fig:RNN"} respectively. This amounts to the guRNN processing 61.2 and 37.0 examples per second in the two tasks, on a GeForce GTX 1080 GPU. # Discussion {#section:discussion} Drawing from the rich theory of Lie groups and Lie algebras, we have described a parametrization of unitary matrices appropriate for use in deep learning. This parametrization exploits the Lie group-Lie algebra correspondence through the exponential map to represent unitary matrices in terms of real coefficients relative to a given basis of the Lie algebra \(\mathfrak{u}(n)\). As this map from \(\mathfrak{u}(n)\) to \(U(n)\) is surjective, the parametrization can describe any unitary matrix. We have demonstrated that unitary matrices can be learned with high accuracy using simple gradient descent, and that this approach outperforms a recently-proposed parametrization (from  ) and significantly outperforms the approach of 're-unitarizing' after gradient updates. This experimental design is quite simple, designed to probe a core problem, before considering the broader setting of RNNs. Our experiments with general unitary RNNs using this parametrization showed that this approach is practical for deep learning. With a fraction of the parameters, our model outperforms LSTMs on the standard 'memory problem' and attains comparable (although inferior) performance on the adding problem . Further work is required to understand the difference in performance between our approach and the ruRNN of -perhaps the \(7n\)-dimensional subspace captured by their parametrization is serendipitously *beneficial* for these RNN tasks-although we note that the results presented here are not the fruit of exhaustive hyperparameter exploration. Of particular interest is the impressive performance of both uRNNs on the memory task, where the LSTM and IRNN appear to fail to learn. While our RNN experiments have demonstrated the utility of using a unitary operator for these tasks, we believe that the role of the nonlinearity in the vanishing and exploding gradient problem must not be discounted. We have shown that a simple scaling factor can help reduce the vanishing gradient problem induced by the choice of nonlinearity. More analysis considering the combination of nonlinearity and transition operator must be performed to better tackle this problem. The success of our parametrization for unitary operator learning suggests that the approach of performing gradient updates in the Lie algebra is particularly effective. As Lie groups describe many naturally-occuring symmetries, the Lie group-Lie algebra correspondence could be rich for further exploitation to enhance performance in tasks beyond our initial motivation of recurrent neural networks. # Appendix A: Derivation of derivative of the matrix exponential This derivation draws elements from   and  . We have \(U = \exp(L)\), and seek \(dU\). For what follows, we simply require that \(L\) be normal, so the results are more general than the unitary case. In this case, \(L\) is skew-Hermitian, which is normal and therefore diagonalisable by unitary matrices. Thus, there exist \(W \in U(n)\) and \(D = \mathrm{diag}(d_1, \dots, d_n)\) such that \(L = W D W^{\dagger}\), and therefore \[U = W \tilde{D} W^{\dagger} \label{eqn:-1}\] where \(\tilde{D} = \mathrm{diag}(e^{d_1}, \dots, e^{d_n})\). We assume we can calculate: \(dL\), \(W\), and \(D\) and seek an expression for \(dU\). Then using [\[eqn:-1\]](#eqn:-1){reference-type="ref" reference="eqn:-1"}: \[\begin{split} dU &= d(W \tilde{D} W^{\dagger}) \\ &= dW \tilde{D} W^{\dagger} + W d\tilde{D} W^{\dagger} + W \tilde{D} dW^{\dagger} \end{split}\] Pre-multiplying with \(W^{\dagger}\) and post-multiplying with \(W\): \[W^{\dagger} dU W = W^{\dagger} dW \tilde{D} + d\tilde{D} + \tilde{D} dW^{\dagger} W \label{eqn:0}\] The last term can be simplified by differentiating both sides of \(W^{\dagger} W = \mathbb{I}\) (this follows from unitarity of \(W\)); \[W^{\dagger} W + W^{\dagger} dW = 0 \Rightarrow dW^{\dagger} W =-W^{\dagger} dW \label{eqn:1}\] and substituting back into [\[eqn:0\]](#eqn:0){reference-type="ref" reference="eqn:0"} to get: \[W^{\dagger} dU W = W^{\dagger} dW \tilde{D}-\tilde{D} W^{\dagger} dW + d\tilde{D}\] We can then say that \(dU = W V W^{\dagger}\) where \[V = W^{\dagger} dW \tilde{D}-\tilde{D} W^{\dagger} dW + d\tilde{D}\] Similarly, \(dL = W A W^{\dagger}\) where (replacing \(\tilde{D}\) with \(D\)) \[A = W^{\dagger} dW D-D W^{\dagger} dW + dD \label{eqn:2}\] and also \(A = W^{\dagger} dL W\). ## Calculating \(V\) We use the convention that repeated indices denote summation over that index, unless otherwise stated. Looking at the components of \(V\); \[V_{ij} = (W^{\dagger} dW \tilde{D})_{ij}-(\tilde{D} W^{\dagger} dW)_{ij} + d\tilde{D}_{ij}\] ### Diagonal case (\(i = j\)): (no summation over \(i\)) \[V_{ii} = W^{\dagger}_{ia} dW_{ab} \tilde{D}_{bi}-\tilde{D}_{ia} W^{\dagger}_{ab} dW_{bi} + d\tilde{D}_{ii}\] Since \(\tilde{D}_{bi} = \delta_{bi} \tilde{d}_i\), the first two terms cancel: \[\begin{split} V_{ii} =& W^{\dagger}_{ia} dW_{ab} \delta_{bi} \tilde{d}_i-\delta_{ai} \tilde{d}_i W^{\dagger}_{ab} dW_{bi} + d\tilde{D}_{ii}\\ =& W^{\dagger}_{ia} dW_{ai} \tilde{d}_i-\tilde{d}_i W^{\dagger}_{ib} dW_{bi} + d\tilde{D}_{ii}\\ =& d\tilde{D}_{ii} \label{eqn:3} \end{split}\] Using [\[eqn:2\]](#eqn:2){reference-type="ref" reference="eqn:2"} we get \(A_{ii} = dD_{ii} = (W^{\dagger} dL W)_{ii}\) Recall that the diagonal elements of \(\tilde{D}\) are the exponentiated versions of the diagonal elements of \(D\), so \(\tilde{D}_{ii} = e^{d_i}\). Then \[d\tilde{D}_{ii} = d(d_i) e^{d_i} = dD_{ii} \tilde{D}_{ii}\] Inserting that into Equation [\[eqn:3\]](#eqn:3){reference-type="ref" reference="eqn:3"}: \[V_{ii} = dD_{ii} \tilde{D}_{ii} = (W^{\dagger} dL W)_{ii} \tilde{D}_{ii} = (W^{\dagger} dL W)_{ii} e^{d_i}\] This produces Equation [\[eqn:vii\]](#eqn:vii){reference-type="ref" reference="eqn:vii"} in the main paper. ### Off-diagonal case (\(i \neq j\)): (no summation over \(i, j\)) In this case, the purely diagonal part vanishes. We get: \[\begin{split} V_{ij} =& W^{\dagger}_{ia} dW_{ab} \delta_{bj} \tilde{d}_j-\delta_{ai} \tilde{d}_i W^{\dagger}_{ab} dW_{bj}\\ =& W^{\dagger}_{ia} dW_{aj} \tilde{d}_j-W^{\dagger}_{ib} dW_{bj} \tilde{d}_i \\ &= (W^{\dagger} dW)_{ij} (\tilde{d}_j-\tilde{d}_i) \end{split} \label{eqn:previj}\] Similarly, \[A_{ij} = (W^{\dagger} dW)_{ij} (d_j-d_i)\] Remembering that this is all component-wise multiplication (no summation over \(i\) and \(j\)), we can rearrange expressions to get: \[(W^{\dagger} dW)_{ij} = \frac{A_{ij}}{d_j-d_i} = \frac{(W^{\dagger} dL W)_{ij}}{d_j-d_i}\] Combining this with [\[eqn:previj\]](#eqn:previj){reference-type="ref" reference="eqn:previj"} and remembering \(\tilde{d}_a = e^{d_a}\), we have, for \(i \neq j\): \[V_{ij} = (W^{\dagger} dL W)_{ij} \left(\frac{e^{d_i}-e^{d_j}}{d_i-d_j}\right)\] This is Equation [\[eqn:vij\]](#eqn:vij){reference-type="ref" reference="eqn:vij"} in the main paper. ## Efficiently calculating \(W^{\dagger} dL W\) This section is specific to our work, as it relies on the choice of basis for \(\mathfrak{u}(n)\). In our case, \(dL\) is simple. \(L\) is a linear combination of the parameters \(\lambda_i\); \[L = \sum_i^{n^2} \lambda_i T_i\] Where \(T_i\) are the basis matrices of \(\mathfrak{u}(n)\). Then \[dL_a = \frac{\partial L}{\partial \lambda_a} = T_a\] We need \(W^{\dagger} T_a W\) for all \(a\). Since the \(T_a\)s are sparse, this is cheaper than performing \(n^2\) full matrix multiplications, as we demonstrate now. In components; \[(W^{\dagger} T_a W)_{i j} = W^{\dagger}_{i k} {T_a}_{k l} W_{l j}\] Cases: ### \(T_a\) diagonal, purely imaginary \(T_a\) is zero except for a \(i\) in the \(a\)-th position on the diagonal. \[\begin{split} (W^{\dagger} T_a W)_{i j} &= i W^{\dagger}_{i a} W_{a j} = i W^*_{a i} W_{a j}\\ \Rightarrow W^{\dagger} T_a W &= i \cdot \mathrm{outer}(\mathbf{w}^*_a, \mathbf{w}_a) \end{split}\] where \(\mathbf{w}_a\) is the \(a\)-th *row* of \(W\). ### \(T_a\) symmetric, purely imaginary \(T_{rs}\) is zero except for \(i\) in position \((r, s)\) and \((s, r)\). \[\begin{split} (W^{\dagger} T_{rs} W)_{i j} &= i W^{\dagger}_{i k} (\delta_{ks, lr} + \delta_{kr, ls}) W_{l j} \\ &= i (W^{\dagger}_{is} W_{rj} + W^{\dagger}_{ir} W_{s j}) = i ( W^*_{si} W_{rj} + W^*_{ri} W_{sj}) \\ \Rightarrow W^{\dagger} T_{rs} W &= i \cdot (\mathrm{outer}(\mathbf{w}^*_s, \mathbf{w}_r) + \mathrm{outer}(\mathbf{w}^*_r, \mathbf{w}_s)) \end{split}\] ### \(T_a\) antisymmetric, purely real \(T_{rs}\) is zero except for \(1\) in position \((r, s)\) and \(-1\) in position \((s, r)\). \[\begin{split} (W^{\dagger} T_{rs} W)_{i j} &= W^{\dagger}_{i k} (\delta_{kr, sl}-\delta_{ks, rl}) W_{l j} \\ &= W^{\dagger}_{ir} W_{sj}-W^{\dagger}_{is} W_{r j} = W^*_{ri} W_{sj}-W^*_{si} W_{rj} \\ \Rightarrow W^{\dagger} T_{rs} W &= \mathrm{outer}(\mathbf{w}^*_r, \mathbf{w}_s)-\mathrm{outer}(\mathbf{w}^*_s, \mathbf{w}_r) \end{split}\] These reproduce the expressions in the main paper. The outer product of two \(n\)-dimensional vectors is an \(O(n^2)\) operation, and so this provides a (up to) factor \(n\) speed-up on matrix multiplication. # Appendix B: Changing the basis of \(\mathfrak{u}(n)\) {#section:basis} The Lie group parametrization assumes a fixed basis of \(\mathfrak{u}(n)\). Our intuition is that this makes some regions of \(U(n)\) more 'accessible' to the optimization procedure, elements whose coefficients are small given this basis. Learning a matrix \(U\) which came from elsewhere in \(U(n)\) may therefore be more challenging. We emulated this 'change of basis' without needing to explicitly construct a new basis by generating a change of basis matrix, \(M\). That is, if \(V_j\) is the \(j\)-th element of the new basis, it is given by \[V_j = \sum_k M_{jk} T_k\] If \(\{\tilde{\lambda}\}_a\) are the coefficients of \(L\) relative to the basis \(V\), the coefficients relative to the old basis \(T\) are given by: \[\lambda_b = \sum_k \tilde{\lambda}_k M_{kb} = \tilde{\lambda}^T \cdot M\] A change of basis matrix must be full-rank. We generate one by sampling a square, \(n^2 \times n^2\) matrix from a continuous uniform distribution \(\mathcal{U}(-c, c)\) (\(c\) is a constant we vary in experiments, see Figure [\[fig:basis\]](#fig:basis){reference-type="ref" reference="fig:basis"}). This is very unlikely to be singular. We choose the \(c\) range of the distribution such that \(M\) will have 'large' values relative to the true matrix \(U\), whose parameters \(\lambda\) (relative to \(T\)) are drawn from \(\mathcal{N}(0, 0.01)\). Preliminary experiments suggested that the learning rate must be adjusted to compensate for the change of scale-evidence for this is visible in the first column of Figure [\[fig:basis\]](#fig:basis){reference-type="ref" reference="fig:basis"}, where changing the basis without changing the learning rate results in an unstable validation set trace. Poor performance resulting from an inappropriate learning rate is not our focus here, so we performed experiments for different values of the learning rate. Figure [\[fig:basis\]](#fig:basis){reference-type="ref" reference="fig:basis"} shows a grid of validation set losses as we vary the learning rate (columns) and the value of \(c\) (rows). Our intuition is that if the performance under the change of basis is *purely* driven by the difference in scale, using an appropriately-scaled learning rate should negate its affect. Each parameter \(\lambda_j\) is scaled by a variable uniformly distributed between \((-c, c)\). The expectation value of the absolute value of this quantity is \(c^2/2\), so we consider learning rates normalised by this factor. As seen in Figure [\[fig:basis\]](#fig:basis){reference-type="ref" reference="fig:basis"}, the graphs on the diagonal are *not* identical, suggesting that merely scaling the learning rate does not account for the change of learning behavior given a new basis-at least in expectation. Nonetheless, it is reassuring to observe that for all choices of \(c\) explored, there exists a learning rate which facilitates learning, even if it markedly slower than the 'ideal' case. While having a 'misspecified' basis does appear to negatively impact learning, it can be largely overcome with choice of learning rate.
{'timestamp': '2017-01-11T02:05:09', 'yymm': '1607', 'arxiv_id': '1607.04903', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04903'}
null
null
null
null
null
null
# Introduction The noise assessment in physical measurements time series is an important measure of its statistical characteristics and overall quality. Among the most effective approaches to analyzing measurement noise (scatter) is Allan variance (AVAR), which was originally introduced to estimate the frequency standards instability. Later, AVAR has proved to be a powerful statistical tool for time series analysis, particularly, for the analysis of geodetic and astronomical observations. AVAR has been used for quality assessment and improvement of the celestial reference frame (CRF) , the time series analysis of station position and baseline length , and studies on the Earth rotation and geodynamics. AVAR estimates of noise characteristics have important advantages over classical variance estimates such as standard deviation (STD) and weighted root-mean-square (WRMS) residual. The latter cannot distinguish the different significant types of noise, which is important in several astro-geodetic tasks. Another advantage of AVAR is that it is practically independent of the long-term systematic components in the investigated time series. AVAR can also be used to investigate the spectral characteristics of the signal that is actively used for analysis of astrometric and geodetic data. However, the application of original AVAR to the time series analysis of astro-geodetic measurements may not yield satisfactory results. Unlike clock comparison, geodetic and astrometric measurements mostly consist of data points with unequal uncertainties. This requires a proper weighting of the measurements during the data analysis. Moreover, one often deals with multi-dimensional quantities in geodesy and astronomy. For example, the station coordinates \(X\), \(Y\), and \(Z\) form 3D vector of a geocentric station position (although this example is more complicated because the vertical and horizontal station displacements caused by the geophysical reasons may have different statistical characteristics, including AVAR estimates, see and references therein). The coordinates of a celestial object, right ascension and declination, also form a 2D position vector. To analyze such data types, AVAR modifications were proposed in, including weighted AVAR (WAVAR), multi-dimensional AVAR (MAVAR), and weighted multi-dimensional AVAR (WMAVAR). These modifications should be distinguished from the classical modified AVAR introduced in. The rest of the paper is organized as follows. Section [2](#sect:overview){reference-type="ref" reference="sect:overview"} introduces AVAR and its modification, and gives several practical illustrations of their basic features. In Section [3](#sect:results){reference-type="ref" reference="sect:results"}, a brief overview is provided of the works that employ AVAR in geodesy and astrometry, and basic results obtained with the AVAR technique are presented. Additional details and discussion on the use of AVAR in space geodesy and astrometry can be found in. # Overview of AVAR and its modifications {#sect:overview} The classical time-domain AVAR applied to the time series \(y_i, i=1, \dots, n\) is given by \[AVAR = \frac{1}{2(n-1)}\sum_{i=1}^{n-1}(y_i-y_{i+1})^2\. \label{eq:AVAR}\] Allan deviation ADEV = \(\sqrt{\rm{AVAR}}\) is used as a noise characteristics in many data analysis applications. Both AVAR and ADEV estimates will be used throughout the paper depending on the context. In metrology, the analyst is normally interested in computing not only the parameter under investigation but also its uncertainty, as a measure of reliability of obtained result. A method of estimating the AVAR confidence interval is proposed in. The method of estimating the AVAR confidence interval proposed in was applied in for analysis of geodetic time series. The original AVAR definition supposes that all measurements (observations) have uniform precision. In most geodetic and astrometric applications, however, measurements display different level of precision. Therefore, appropriate weighting is necessary during analysis. Given the measurements \(y_1,y_2,\dots,y_n\) with the associated uncertainties \(s_1,s_2,\dots,s_n\) Then the WAVAR estimate can be defined to treat unequally weighted data : \[\begin{array}{rl} & WAVAR = \frac{\displaystyle 1}{\displaystyle 2p} \displaystyle\sum_{i=1}^{n-1}p_i(y_i-y_{i+1})^2 \,, \\[2ex] & p = \displaystyle\sum_{i=1}^{n-1}{p_i}\,, \\[4ex] & p_i = (s_i^2+s_{i+1}^2)^{-1} \,, \end{array} \label{eq:WAVAR}\] where \(p_i\) are the weights. Figure [\[fig:WAVAR_example\]](#fig:WAVAR_example){reference-type="ref" reference="fig:WAVAR_example"} illustrates the difference between AVAR and WAVAR for three time series with outliers: daily estimates of the station height and two series of celestial pole offsets (CPO). The latter are the differences in \(dX\) and \(dY\) between the observed \(X\) and \(Y\) coordinates of the celestial pole from the celestial pole position defined by the International Astronomical Union IAU 2000/2006 precession-nutation model as measured by VLBI (very long baseline interferometry). Here, we consider points with abnormally large deviations from the mean and/or with abnormally large uncertainties to be outliers. Analysts usually aim to identify and reject outliers before conducting statistical analysis; however, generally this is not a simple problem. One can see from Fig. [\[fig:WAVAR_example\]](#fig:WAVAR_example){reference-type="ref" reference="fig:WAVAR_example"} that WADEV can provide more robust estimate even in the presence of outliers that were not previously detected and eliminated. In other words, WADEV is less dependent on the quality of the outlier detection procedure used prior to data analysis. Notably, WADEV dependence on outliers is low when the outliers exhibit greater uncertainties than the uncertainties of "good" measurements. This is mostly the case in practice. Exceptions are possible, as shown in the bottom plot of Fig. [\[fig:WAVAR_example\]](#fig:WAVAR_example){reference-type="ref" reference="fig:WAVAR_example"} although they are rare. It is important to keep in mind that AVAR, as well as any other statistics has limitations depending on data under analysis. Let us consider an example of time series of Z coordinate of the station position presented in Fig. [\[fig:hflk\]](#fig:hflk){reference-type="ref" reference="fig:hflk"}, upper plot. One can see that both ADEV and WADEV give unsatisfactory results. Both estimates are affected by jumps in the time series and are too large as compared with the actual measurement noise. Moreover, WADEV gave even worse result than ADEV. The reason of such a confusing WADEV estimate is the following. As can be clearly seen in Fig. [\[fig:hflk\]](#fig:hflk){reference-type="ref" reference="fig:hflk"}, bottom plot, the measurement errors changed with time. In particular, the smallest errors are observed in the second part of the time series, where the largest jump occurred. As a result, the measurements made around the second, largest jump were entered to the WADEV computation with larger weight. For comparison, if only the part of the time series after the last jump (epoch 2007.5) is used for analysis, ADEV = WADEV = 0.2 cm. So, ADEV, as well as most of other statistics, cannot provide a satisfactory result for non-stationary time series. In particular, if the time series under investigation has heterogeneous uncertainty, the use of dynamic AVAR (DAVAR) is clearly advantageous. The DAVAR is defined as AVAR computed over a sliding window moving along the time series, and thus it allows estimating the variation of the time series noise characteristics with time. In astro-geodetic time series analysis, we often address multi-dimensional values, such as geocentric station coordinates \(X\), \(Y\), and \(Z\) (3D vector) and the position of celestial objects (right ascension and declination) in the sky (2D unit vector). Further extension of the WAVAR estimator WMAVAR can facilitate the analysis of such data. Given \(k\)-dimensional vector of measurements \(y_i=(y_i^1,y_i^2,\dots,y_i^k)\) with the \(k\)-dimensional vector of corresponding uncertainties \(s_i=(s_i^1,s_i^2,\dots,s_i^k)\), WMAVAR considers the measurements to be the points in the \(k\)-dimensional space. Then, the difference between adjacent measurements is taken as the Euclidean length between these values. Thus, the \(k\)-dimensional WMAVAR estimate can be defined as \[\begin{array}{rl} & WMAVAR = \frac{\displaystyle 1}{\displaystyle 2p} \displaystyle\sum_{i=1}^{n-1}p_i \, d_i^2\,, \\ & d_i = |y_i-y_{i+1}|\,, \\[1ex] & p = \displaystyle\sum_{i=1}^{n-1}{p_i}\,, \\[3ex] & p_i = \left( \displaystyle\sum\limits_{j=1}^k \left[(y_i^j-y_{i+1}^j)/d_i\right]^2 \left[(s_i^j)^2+(s_{i+1}^j)^2\right] \right)^{-1} \,, \end{array} \label{eq:WMAVAR}\] where \(|\dots|\) denotes the Euclidean norm, and \(p_i\) are the weights. The expression for \(p_i\) is derived from classical error propagation low and is therefore theoretically correct. However, this expression cannot be applied in practice because it generates a singular when \(d_i=0\), that is, a case in which adjacent measurements are equal. This situation is not extraordinary. To eliminate this problem, a simplified formula was proposed in : \[p_i=\left( \displaystyle\sum\limits_{j=1}^k \left[(s_i^j)^2 + (s_{i+1}^j)^2 \right] \right)^{-1} \. \label{eq:p_alt}\] Numerous tests with real data revealed no practical difference between the use of Eqs. ([\[eq:WMAVAR\]](#eq:WMAVAR){reference-type="ref" reference="eq:WMAVAR"}) and that of ([\[eq:p_alt\]](#eq:p_alt){reference-type="ref" reference="eq:p_alt"}) in computing for WMAVAR. It is easy to see that WMAVAR (WMADEV) is a universal definition that encompasses all other variants such as AVAR (ADEV), WAVAR (WADEV), and MAVAR (MADEV) as special cases. It can be also noted that for a time series whose \(s_i\) values are close to one another WMADEV is approximately \(\sqrt{k}\) times as large as the WADEV computed for one (each) vector dimension. For example, given three series of equally precise measurements of station coordinates \(X\), \(Y\), and \(Z\) having about the same WADEV estimate \(\sigma_1\), the 3D WMADEV estimate will be equal to roughly \(\sqrt{3}\sigma_1\). A main advantage of AVAR when used as a noise level estimator is its weak dependence on low-frequency variations. Figure [\[fig:cpo_fcn\]](#fig:cpo_fcn){reference-type="ref" reference="fig:cpo_fcn"} depicts the results of the CPO determination in the International VLBI Service for Geodesy and astrometry (IVS). One can see that the CPO series include quasi-harmonic signal with period about 430 days corresponding to the free Earth's core nutation (FCN) and low-frequency components caused by the errors in the precession-nutation model. In this study, CPO is modeled by the ZM2 and ZM4 models The results presented in Fig. [\[fig:cpo_fcn\]](#fig:cpo_fcn){reference-type="ref" reference="fig:cpo_fcn"} show that the WRMS estimate depends heavily on the low-frequency components of the analyzed signal. The removal of these components requires highly accurate modeling of the underlying physical processes. By contrast, AVAR is practically unaffected by the presence of trend and low-frequency harmonics. An important AVAR application is in characterizing the spectral behavior of a time series. The assessment of noise spectral characteristics may be very important in astrometric and geodetic data analyses. For example, a choice of the method for computing the station velocity uncertainty from the station position time series crucially depends on the noise spectral type. Knowledge of the spectral noise structure facilitates the construction of a proper covariance matrix of a geodetic signal. As with clock deviations, the noise processes in geodetic and astronomical time series can generally be described by a power law \[S_y(f) = S_0 \, f^\alpha \,, \label{eq:power_law}\] where \(S_0\) is a normalizing constant and \(\alpha\) is the spectral index. It was supposed above that \(\{y_i\}\) are the original measurements and that AVAR is just used as a scatter level estimate. The original AVAR definition as given in full by is \[AVAR = \sigma^2(\tau) = \frac{1}{2} \langle (\bar y_i-\bar y_{i+1})^2 \rangle \,, \label{eq:original_AVAR}\] where \(\bar y_i\) are averaged values over the sampling interval (number of points) \(\tau\). Then, the following relation can be used to identify noise spectral characteristics: \[\log(\sigma^2(\tau)) = \mu \, \log(\tau) \. \label{eq:avar_tau}\] The \(\mu\) value is connected to the spectral index by \(\alpha=-(\mu+1)\) for \(-3 < \alpha < 1\) (\(-2 < \mu < 2\)). Then, the prevailing noise type in the time series can by classified as white noise (\(\mu=-1, \alpha=0\)), flicker noise (\(\mu=0, \alpha=-1\)), random walk (\(\mu=1, \alpha=-2\)), or another noise type (see Fig. [\[fig:noise_type\]](#fig:noise_type){reference-type="ref" reference="fig:noise_type"} for illustration). It must be kept in mind, however, that the analogy of the AVAR-derived log-log slope \(\mu\) with the spectral index is only justified for stationary stochastic processes. Given real signals, noise often exhibits different spectral characteristics at various frequency bands; thus, a full noise spectrum consists of several components: \[S_y = \sum_{j=1}^{N_f}{c_j\,f_j^{\alpha_j} \.} \label{eq:complex_spectra}\] It must be noted that this noise classification method must be practiced with care. First, the analyzed time series should be sufficiently long; at least several hundreds of measurements are needed to obtain a reliable result. Moreover, trend, seasonal and other long-periodic terms (beyond the frequency band of the analyst's interest) from the analyzed signal should be removed prior to spectral analysis; otherwise, the log-log spectrum ([\[eq:avar_tau\]](#eq:avar_tau){reference-type="ref" reference="eq:avar_tau"}) can be distorted. In case of uniform noise, computing and analyzing the classical power spectra may be preferable, such as through Fourier transform. In case of mixed noise, however, AVAR analysis can provide valuable supplemental information for adjusting the analysis procedure. Indeed, the use of the AVAR algorithm in characterizing the spectral behavior of a time series can easily be generalized for weighted and multi-dimensional measurements through the application of WMAVAR and weighted average at each sampling interval \(\tau\). It can be mentioned here that the computation of weighted average uncertainty does not generate an unambiguous solution. A possible approach to solving this problem in practical applications is discussed in. Among other applications, AVAR can be used to estimate the Hurst parameter. The Hurst parameter (or Hurst exponent) \(H\) is related to such time series statistical properties as long-term memory, self-similarity (fractal dimension) and spectral index. Notably, the original AVAR is defined for evenly spaced time series, which generally does not apply to all geodetic and especially astronomical data. The simplest method of solving this problem is to form so-called normal points that average data over equally spaced intervals. Such a method was used in for radio source position time series and in for station position time series. Nonetheless, the process of averaging the original series inevitably leads to loss of information regarding high-frequency noise characteristics. However, if AVAR is merely used as a measure of random deviation level, uneven data spacing should not influence analysis results, supposing stationary noise. A special type of unevenly spaced data is the time series that consists of equally spaced measurements with gaps. In, a special technique has been developed to fill in data gaps prior to statistical analysis. Additional considerations for the application of AVAR to series with gaps are given in. # Use of AVAR in geodesy and astrometry {#sect:results} This section provides an overview of the applications of AVAR in astrometry, geodesy, and geodynamics. In these fields, AVAR is mainly applied to investigate space geodesy station displacements and the stability of VLBI-derived celestial object positions. Several examples of AVAR applications are discussed in the following subsections to effectively illustrate some features of AVAR estimates. ## Celestial reference frame The studies described in this subsection are CRF-related investigations, such as the analysis of radio source coordinates catalogs and the improvement of the International Celestial Reference Frame (ICRF) through the detection of radio sources with unstable positional behavior. Accuracy assessment of the CRF catalogs is not a trivial task because only the differences between the radio source positions in various catalogs are known (this situation is very similar to the case of clock comparison). A possible absolute method of catalog comparison is based on the comparison of the noise levels in the CPO series as computed with the compared source catalogs. In this study, the radio source position catalogues obtained in 8 IVS analysis centers were compared by means of two scatter indices. Both indices are based on the analysis of the CPO series computed with different catalogs. The first index is computed from analysis of the residuals of CPO series with respect to the IAU 2000/2006 precession-nutation model. This index occurred to be less sensitive, and, besides, it is not fully independent, since the IAU model was determined form VLBI data analysis using some CRF realizations under study. Conversely, WADEV and WMADEV indices provide an independent estimate of the quality of CRF realizations. The WMADEV occurred to be most sensitive to differences in CRF realizations, and thus can be considered as a preferred scatter index. With the use of this method, the ICRF catalog, hereafter referred to as ICRF1, was compared with the catalog RSC(PUL)07C02 derived at the Pulkovo observatory. The latter was computed as combined catalog using several individual catalogs obtained in the IVS analysis centers in 2005. It uses more observations than ICRF1 catalog computed in 1995, and a new method of determination and elimination of the systematic errors in source positions. Two CPO time series for 2002--2006 obtained with two CRF realizations were computed. The results of comparison are presented in Table [1](#tab:cpo_cat){reference-type="ref" reference="tab:cpo_cat"} and show that the noise level of the CPO series computed with the Pulkovo catalog is lower than that of the series computed with ICRF1 This finding suggests that the catalog RSC(PUL)07C02 was more accurate than ICRF1. This conclusion was later confirmed by improvement of the EOP results obtained with the new catalog. The analysis of the astrometric aspects of catalog computation and comparison is beyond the scope of this paper. It is only important here to show that estimation of the noise level in the CPO time series can serve as an external characterization of the accuracy of radio source catalogs (CRF realizations). Subsequently, a similar comparison was made in between two ICRF realizations, ICRF1 and ICRF2. The latter was computed in 2009 and using about 6.5 mln observations collected during 1979--2009 as compared with 1.6 mln observations collected in 1979--1995 used for ICRF1. As a result, several direct catalog comparisons showed much better precision and of ICRF2 with respect to ICRF1. Two CPO time series for 2002--2010 obtained with two CRF realizations were computed. The result of the CPO noise analysis made in and presented in Table [2](#tab:icrf2-icrf1){reference-type="ref" reference="tab:icrf2-icrf1"} independently confirmed that ICRF2 is more accurate than ICRF1. It should be realized that both \(dX\) and \(dY\) are merely the spherical coordinates of the celestial pole, and their statistical behaviors are closely correlated. Thus, the 2D WMADEV estimate serves as a compact and convenient tool for describing the noise component in CPO (and other similar) series. The noise characteristics of a source position time series can be used for source ranking in terms of temporal and spatial stabilities, as well as for compiling a list of sources that are not stable enough to be included in VLBI global solution and thus requires special handling. The position instabilities of 16 radio sources with long and dense observational history were investigated in detail, including the spectral characteristics of the source position time series through the AVAR analysis. The analysis results showed that the time variabilities of these objects exhibit varied spectral characteristics that may be ascribed to diversified physical processes. Several statistical techniques were applied to source position time series to assess the VLBI-derived CRF in. In particular, AVAR analysis was performed to assess the noise spectrum of 60 source position time series. Sampling interval \(\tau\) varied from the initial time span of the series to 1/4 \(\ldots\) 1/3 of the data span (six months to four years). It was found in that roughly 60% of the series predominantly exhibited white noise and 40% of the series mainly displayed flicker noise. Approximately 160 sources have sufficiently long and dense observation histories from which the AVAR can be derived for a one-year sampling interval. This information was employed to investigate source position stability further. An extended similar work was conducted to select a set of positionally stable sources through an analysis of source position time series with the use of WRMS and ADEV. This study also included an analysis of the apparent drifts in radio source positions. In this study, 707 sources were analyzed, and 199 were recommended as candidate core sources for the subsequent ICRF realization. In particular, roughly 60%, 35%, and 5% of the series exhibited white noise, flicker noise, and random work noise, respectively. The methods developed in for source stability characterization were also used to analyze recent source position time series. This study investigated the position stability of ICRF2 sources using AVAR and considered the apparent linear source motions. It was found that the selection of stable core sources depends on the configuration of the VLBI solution used to derive the position time series and may differ by up to 20% of the selected sources. Furthermore, source stability found in improved significantly with respect to that found in because of the advances in observing technology as well as the increase in observing networks, which directly affect the precision and accuracy of VLBI results. Another analysis performed in found that the log-log slope \(\mu\) computed according Eq. ([\[eq:avar_tau\]](#eq:avar_tau){reference-type="ref" reference="eq:avar_tau"}) may not remain constant with time for long radio source position time series. Subsequently, this work was extended by incorporating the most recent observations in the analysis. AVAR was also applied to the time series analysis of 15 source positions provided by 9 IVS analysis centers in the framework of ICRF2 preparation campaign. In particular, it was found that AVAR generates an inadequate estimate of noise level when jumps are present in a time series, as discussed above. In general cases, a composite index of WRMS and WADEV can be applied to enhance the robustness of the measure of source position stability, as previously discussed in. ## Geodesy Many studies have confirmed that AVAR is an effective tool for investigating the noise characteristics of station position time series, particularly those obtained with space geodetic methods, such as VLBI, GPS, satellite laser ranging (SLR), and Doppler orbitography and radiopositioning integrated by satellite (DORIS). AVAR was used to estimate the week-to-week repeatability of the station position of the European Permanent GPS Network (EPN) station position in the presence of seasonal station movements. The use of AVAR in this case is reasonable because AVAR does not require the estimation of a systematic trend and seasonal variations in station positions, which are difficult to model. In the process, the WRMS noise estimate that was previously used to compare station position time series is distorted. AVAR was also applied to the time series analysis for baseline length in relation to GPS-based deformation monitoring applications, where this approach was used to determine the temporal upper limit accuracy of each monitoring system and to help distinguish a systematic error from a genuine geophysical signal for a new measurement. The noise spectrum in DORIS ground station motion was investigated by applying AVAR to the decomposition of the 3D signal into its principal components in the time domain. The noise level was estimated, and white noise was dominant after the linear drift in the series of station coordinates was eliminated. In addition, the AVAR signature was sensitive to periodic terms. To avoid bias in the estimation of noise type, three periodic terms, namely annual, semi-annual, and 117.3-day, were filtered out prior to computation of AVAR. The annual and semiannual seasonal variations are usually observed in the station position series derived from astronomical and space geodesy observations see and literature cited therein. The 117.3-day signal is specific for DORIS observations. It is connected with the 117.3-day period in orbit node movement for the TOPEX/Poseidon satellite extensively used during many years for DORIS. Besides, the method of estimating the AVAR confidence interval proposed in was applied in to the DORIS stations position time series. This became usual practice in future works. The time series of the station coordinates obtained from various space geodetic observations were analyzed in. It was found that the station position series obtained from VLBI, SLR, and DORIS contain white noise, whereas the majority of GPS station motions exhibit flicker noise. The atmospheric loading time series (change in the station positions caused by variations of the atmospheric pressure) also contain white noise, as do the series of the transformation parameters in the midst of position time series obtained by different techniques. A similar but more detailed analysis was presented in. The noise and low-frequency components of coordinate variations were investigated with the data obtained from three permanent GPS stations located in the territory of Pulkovo Observatory. AVAR was used in this work to assess the noise in the station position time series. The results were compared with the spectral index obtained from Hurst exponent analysis. Both methods suggested that the position time series for the three Pulkovo stations are generally dominated by flicker and mixed white-flicker noises. An AVAR analysis of the noise characteristics in the time series for DORIS station position was presented in. Once the trends and seasonal components from the analyzed time series were eliminated, the results of the subsequent AVAR analysis indicated that the three solutions are dominated by white noise in all three components (north, east, and vertical). Upon confirming the dominant noise type in the position time series, an adequate method for further analysis was chosen. ## Earth's rotation and geodynamics Earth orientation parameters (EOP) define the rotation of the terrestrial reference system ITRS with respect to the celestial reference system ICRS. These parameters include polar motion (movement of the Earth's rotation axis in ITRS), precession-nutation (movement of the Earth's rotation axis in ICRS), and Universal Time, which depends on the rotational speed of the Earth around its axis. EOP are determined through the same space geodetic methods described in the previous subsection. The final EOP series for practical use are derived by combining the results obtained at different EOP analysis centers. The noise characteristics of EOP series should be investigated for different tasks, such as quality assessment of the EOP series and the weighting of these series during combination. The computation method of the combined EOP series at the Bureau International de L'Heure (BIH) involved an AVAR analysis of the input EOP series. Specifically, AVAR was used to obtain the frequency dependence of the noise power spectrum in the combined series. In particular, it was found that the yearly bias in a time series with respect to the combined series is dominated by flicker noise, which may indicate accidental local and instrumental changes. However, the individual results obtained at separate stations are not stable enough; thus, corrections for individual stations cannot be reliably applied. AVAR analysis is also a main step in the EOP combination procedure employed by the International Earth Rotation and Reference Systems Service (IERS), which succeeded BIH. AVAR is used to characterize the internal random deviations in the EOP time series at various time scales. White noise and flicker noise dominate these series Another application of the AVAR is in comparing EOP series to evaluate their internal precision. These data are used to assign proper weights to EOP series during combination. An analysis of several CPO series computed at different IVS analysis centers was performed in. The results of this study confirmed that the WRMS estimate of the noise in CPO series is heavily dependent on the model of low-frequency variations, whereas WADEV is practically independent of such a model. The use of WADEV estimates also mitigated the influence of outliers significantly. Using WADEV estimates also allowed us to severely mitigate the influence of outliers. Another result of the study is that 2D WMADEV estimates are close to the averaged WADEV estimates computed for \(dX\) and \(dY\) and multiplied by \(\sqrt{2}\), as was discussed in Section [2](#sect:overview){reference-type="ref" reference="sect:overview"}. Seasonal and non-seasonal components in the geocenter motion signal measured by satellite geodesy techniques were analyzed to determine the connection of these components with geophysical fluids. AVAR was used in this study to compute and compare the spectral characteristics of geodetic and geophysical geocenter time series over a sampling period of one month to one year. In particular, the dependence of the log-log \(\tau\)-AVAR pattern on the seasonal components in the analyzed signal was determined. Geodetic and modeled excitation functions were compared via AVAR in. In this study AVAR analysis was applied to Earth's rotation excitation functions, complementing other methods of statistical analysis. Bias and seasonal variations were removed from the signals prior to AVAR computation. Comparison of ADEV obtained for geodetic and geophysical excitations series, and their residuals for different sample interval \(\tau\) allowed to draw a conclusion on the geophysical excitation of the Earth's rotation. # Conclusions AVAR is an effective statistical tool for analyzing the time series of observational data in astronomy and geodesy, as well as of all other time series. Important independent characteristics of the noise component in the studied signal can be obtained through this tool. The main application of AVAR to time series analysis is in determining the signal scatter level and spectral analysis with the primary aim to identify the dominant noise type in a time series. This information can be applied to refined data analysis, such as computing realistic uncertainty of station velocities ) and regularization of EOP series. AVAR modifications, namely, WAVAR, MAVAR, and WMAVAR, were proposed for processing weighted and/or multi-dimensional data. These modifications serve as effective and convenient tools for data analysis in geodesy and astronomy. WMAVAR is the most general estimator and encompasses AVAR, WAVAR, and MAVAR as special cases. In particular, 2D WMAVAR can be used for complex data processing. Indeed, the AVAR modifications we proposed should not be confused with the "original" modified AVAR definition. An important advantage of AVAR and its more refine versions over WRMS in practical application is its weak sensitivity to low-frequency signal variations. By contrast, WRMS depends heavily on the model used to eliminate the systematic component of the studied signal. Our study showed that WAVAR is more robust to outliers than the classical AVAR is; however, both AVAR and WAVAR may estimate noise level erroneously when jumps occur in a time series. AVAR is also widely used to investigate the spectral characteristics of a time series and is a powerful tool for noise type identification through log-log representation. In particular, AVAR facilitates the effective analysis of signals with different types of noise at various frequency bands. AVAR is supposed to be more computationally effective than classical spectral methods, such as Fourier transform; in our opinion, however, this advantage is no longer significant at present. A detailed comparison of two methods for estimating spectral noise characteristics may be interesting and useful though. It must be noted that the AVAR method needs further investigation and development for some applications. First, many geodetic and astronomical series are unevenly spaced as was discussed in Section [2](#sect:overview){reference-type="ref" reference="sect:overview"}. Frequent examples include station position time series with gaps, radio source position time series, VLBI-derived session-wise EOP series. Another open issue regarding the application of AVAR to the analysis of geodetic and astronomical time series stems from possible correlations between the measurements that may distort statistical analysis results substantially. Finally, we can conclude that despite its limitations and some unresolved issues, AVAR nevertheless remains one of the most powerful tools for analyzing a wide range of physical measurement time series.
{'timestamp': '2016-07-19T02:03:14', 'yymm': '1607', 'arxiv_id': '1607.04712', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04712'}
# Introduction There are many ways to define the Fibonacci numbers. An equivalent approach to the standard recurrence relation, where \(F_{n+1} = F_n + F_{n-1}\) and \(F_1 = 1\) and \(F_2=2\), is that they are the unique sequence of integers such that every positive number can be written uniquely as a sum of non-adjacent terms. This expansion is called the Zeckendorf decomposition, and much is known about it. In particular, the distribution of the number of summands of \(m \in [F_n, F_{n+1})\) converges to a Gaussian as \(n\to\infty\), with mean and variance growing linearly with \(n\). Similar results hold for a large class of sequences which have a notion of legal decomposition leading to unique decomposition; see. Given a sequence \(\{H_n\}\), one can frequently prove that the mean and the variance of the number of summands of \(m \in [H_n, H_{n+1})\) grows linearly with \(n\). Explicitly, there are constants \(a, b, C\) and \(d\) such that the mean is \(an+b+o(1)\) and the variance is \(Cn + d + o(1)\). The difficulty is proving that \(a\) and \(C\) are positive, which is needed for the proofs of Gaussian behavior. Until recently, the only approaches have been technical and involved generating functions, partial fraction expansions and generalized Binet formulas applied to polynomials associated to the characteristic polynomials of the sequence, which have required a lot of work to show the leading terms are positive for such recurrences. The point of this work is to bypass these arguments through elementary counting. We concentrate on positive linear recurrence sequences (defined below) to highlight the main ideas of the method; with additional work these arguments can be extended to more general sequences (see ). In addition to the arguments below, one can also obtain similar results (though not as elementarily) through Markov chains or through an analysis of two dimensional recurrences. The following two examples clarify the above. Before we state our main result we first set some notation. As remarked above, previous work has shown that \(\E{K_n} = an + b + f(n)\) where \(a > 0\) and \(f(n) = o(1)\); this can be proved through very simple counting arguments (see ). While it is also known that \(\V{K_n} = Cn + d + o(1)\), previous approaches could not easily show \(C \neq 0\). We elementarily prove \(C > 0\) by giving a positive lower bound \(c\) for it. We sketch the proof. We can remove the second to last block of a legal decomposition to get a shorter legal decomposition, forming relations between longer legal decompositions and shorter legal decompositions. We then use strong induction and conditional probabilities to prove the theorem. # Lemmas derived from Expectation We first determine a relationship between \(K_n\) and \(Z_n\). Then, with the help of \(\E{K_n} = an + b + f(n)\), we explain how to explicitly determine the positive lower bound \(c\). For an arbitrary \(\omega \in S_t\), the second to last block has size \(Z_n = t\), and the remaining blocks form a legal decomposition in \(\Omega_{n-\ell(t)}\) with size \(K_{n-\ell(t)}(h_t(\omega))\), so \(K_n(\omega) = K_{n-\ell(t)}(h_t(\omega)) + t\). Since \(h\) is a bijection, we have the following two equations: \[\label{1} \begin{split} \E{K_n|Z_n = t} &\ = \ \E{K_{n-\ell(t)} + t}\\ &\ = \ a(n-\ell(t)) + b + f(n-\ell(t)) + t, \end{split}\] and \[\label{2} \begin{split} \E{K_n^2|Z_n = t} &\ = \ \E{(K_{n-\ell(t)} + t)^2}\\ &\ = \ \E{K_{n-\ell(t)}^2 + 2t{K_{n-\ell(t)}} + t^2}\\ &\ = \ \E{K_{n-\ell(t)}^2} + 2t\E{K_{n-\ell(t)}} + t^2\\ &\ = \ \E{K_{n-\ell(t)}^2} + 2t[a(n-\ell(t)) + b + f(n-\ell(t))] + t^2. \end{split}\] Furthermore, by \(\eqref{1}\) we have \[\label{3} \begin{split} \E{K_n} &\ = \ \sum\limits_{t = 0}^{S-1} \PP{Z_n = t}\cdot \E{K_n|Z_n = t}\\ &\ = \ \sum\limits_{t = 0}^{S-1} \PP{Z_n = t}\cdot [a(n-\ell(t)) + b + f(n-\ell(t)) + t]\\ &\ = \ an + b + \sum\limits_{t = 0} ^ {S-1} \PP{Z_n = t}\cdot [t + f(n-\ell(t))-a\ell(t)]\\ &\ = \ an + b + f(n),\\ \end{split}\] where the last equality comes from the definition of \(f(n)\). If we set \(Y_n(\omega) := Z_n(\omega) + f(n-L_n(\omega))-aL_n(\omega)\), then we have \[\label{4} \E{Y_n} \ =\ \sum\limits_{t = 0} ^ {S-1} \PP{Z_n = t}\cdot [t + f(n-\ell(t))-a\ell(t)] \ = \ f(n).\] Now that we have \(\E{Y_n}\), we use it to estimate \(\V{Y_n}\). Finally, we choose \(c\). Let \[c\ =\ \text{min}\left\{\frac{\V{K_{L+1}}}{L+1},\ \frac{\V{K_{L+2}}}{L+2},\ \dots,\ \frac{\V{K_{N}}}{N},\ \frac{a^2}{2SL}\right\},\] Where \(N\) is as determined in Lemma [\[lem:varyn\]](#lem:varyn){reference-type="ref" reference="lem:varyn"}. For all \(n > L\), \(H_{n + 1}-H_n > 1\), so there are at least two integers in \([H_n, H_{n+1})\). Since the legal decomposition of \(H_n\) has only one summand while that of \(H_n + 1\) has two summands, \(\V{K_n}\) is nonzero when \(n > L\). Hence, \(c > 0\). In the next section we show \(\V{K_n} \ge cn\) for all \(n > L\). # A lower bound for the Variance We prove Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} by strong induction. While the algebra is long, the main idea is easily stated: we condition based on how many summands are in the second to last block, which must be a type 2 block, and then use conditional probability arguments (inputting results for the mean and smaller cases) to compute the desired quantities.
{'timestamp': '2016-07-19T02:02:24', 'yymm': '1607', 'arxiv_id': '1607.04692', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04692'}
# Introduction The work presented here, roughly speaking, aims at proving a version of the folklore statement: Gravitational waves carry energy away from an isolated system. More precisely, a system with no radiation emitted to infinity must in fact be stationary. On physical grounds it is expected that truly time-periodic gravitational systems do not exist. This is due to the hyperbolic structure of the Einstein equations; any dynamical solution loses energy through outward radiation. However as remarked by Alexakis-Schlue, this is a subtle mathematical question. It is true at the linearised level when we consider, for example, the free wave equation \(\square \phi=0\) on a Minkowski background: Formally, an outgoing wave with vanishing radiation field has to be stationary; that is, if the function \(\phi\) decays faster than \(r^{-1}\) in the null-outgoing direction then it must be time-independent. This is no longer the case for a suitably perturbed wave operator \(L = \square + V\) as remarked by Alexakis, Schlue and Shao in. This problem can also be phrased as a question of uniqueness of solutions for differential operators given boundary conditions. Hörmander,, provided general conditions for the uniqueness property to hold across an hypersurface. In the context of an hyperbolic principal symbol, these conditions reduce to requiring the hypersurface to be pseudo-convex (see section [\[section:Carleman estimates\]](#section:Carleman estimates){reference-type="ref" reference="section:Carleman estimates"} for more details). Complementary to this is the work of Alinhac,, he showed that (generically) if one of Hörmander's conditions is violated then non-uniqueness of solutions across an hypersurface ensue.\ In this paper we focus on the consequences of the lack of gravitational waves and radiation. Firstly, we give a brief account of their mathematical meaning. It is easier to start at the linear level: Consider a small perturbation of the Minkowski metric in flat spacetime. The perturbation then is required to satisfy the linearised Einstein equations around the trivial solution. After imposing suitable gauge conditions, it is found that the perturbation obeys a wave equation and therefore it behaves as a wave propagating causally on the background. Moreover, if we impose vanishing boundary conditions at infinity then the leading order term can be interpreted as power radiated to infinity. These perturbations are the so-called *gravitational waves*. The space of solutions, of both linear Einstein and gauge equations, consists of the linear combination of two scalar waves, that is, gravitational waves have 2 degrees of freedom or polarisation modes.\ Now, a description of gravitational waves in the non-linear case is more delicate. To the author's knowledge there is no way of splitting the gravity field into "stationary" and "dynamic" parts; this is due to the non-linear nature of Einstein equations. Whence the lack of meaning of the expression: Here is a gravitational wave (to be attributed to the dynamical part) propagating in a background (the stationary part). Nevertheless considerable efforts have been made to understand this statement and it is possible to make sense of it *at infinity* by imposing *asymptotically flat* boundary conditions[^1]. In Section [\[section:CK analysis\]](#section:CK analysis){reference-type="ref" reference="section:CK analysis"} we review briefly the conclusions of the asymptotic analysis carried out by Christodoulou and Klainerman in order to prove the non-linear stability of Minkowski spacetime,.\ To the author's knowledge, it was Papapetrou, ,, who initiated the study of the relation radiation-stationarity for the full non-linear Einstein equations. Although his result is not conclusive, it provided strong evidence for the validity of the theorem presented here. More precisely, Papapetrou showed the incompatibility of the following two conditions: a) A spacetime is stationary below a characteristic hypersurface and non-radiative above it; b) There is a shock wave along that characteristic hypersurface or above it.\ Later on, Bičák, Sholtz and Tod,, gave a rigorous proof in the class of analytic metrics using ideas from Gibbons and Stewart,. Bičák-Scholtz-Tod used the (undesired) hypothesis of analyticity all the way up to infinity to split Einstein equations order by order at infinity and then used the non-radiative condition to conclude time-independence to all orders inductively. Then the desired stationarity conclusion in the interior follows by analytic continuation.\ The theorem was later improved to cover smooth metrics by Alexakis and Schlue in. The purpose of this paper is to show that the result holds as well when gravity is coupled to electromagnetism. Alexakis and Schlue approach relies again on first proving stationarity to all orders at infinity as in. Then, in order to extend this condition to the interior, they use unique continuation from infinity techniques based on Carleman estimates in the spirit of. For the first part of the proof, the exact null structure of the equations plays an important role in order to compute the metric, connection coefficients and curvature components to all orders at infinity with just the radiation field. For the second part a wave equation satisfied by the deformation tensor of the Weyl curvature, \(\mathcal{L}_T C\), is derived and exploited to conclude stationarity in a neighbourhood of infinity by means of energy estimates.\ We consider here the extended problem to include a Maxwell field. It is found that the results from hold as well, that is, gravity and electromagnetism cannot balance each other to produce a periodic solution. The key point is that Maxwell equations also have a special non-linear structure and the coupled Einstein-Maxwell system can be treated in a similar way as before. The two deformation tensors, \(\mathcal{L}_T C\) and \(\mathcal{L}_T F\), also obey wave equations; however the coupling terms do not decay fast enough for the Alexakis-Schlue argument to work. Hence we are forced to revise and adapt their proof at the level of Carleman estimates to conclude the vanishing of the deformation tensors a little bit into the interior of the spacetime. The main result of this paper is, We comment briefly about the assumptions. In this paper we will consider the class of asymptotically flat spacetimes admitting coordinates \((t, r, \theta^i)\) which are Minkowski to leading order and the decaying metric coefficients admit an infinite asymptotic expansion in terms of inverse powers of \(r\), an area parameter. Moreover, it is required that these expansions are well-behaved with respect to derivatives up to second order (cf. definition [\[def:AF\]](#def:AF){reference-type="ref" reference="def:AF"}). We remark that this condition of asymptotic flatness is morally equivalent to smoothness at null infinity in the conformal picture, the advantage of this formulation lies in its adaptability to more general asymptotic conditions, e.g., polyhomogeneous expansions. Indeed, the methods employed here can be generalised to cope with time-independent logarithmic singularities; and while we take into account the asymptotic behaviour obtained by Klainerman-Christodoulou up to order three, it is a definite hypothesis of Theorem [\[thm1\]](#thm1){reference-type="ref" reference="thm1"} that we work with metrics which are smooth at higher orders. For more details about the smoothness assumption we refer the reader to the introduction of Section [\[chapter:asymptotics\]](#chapter:asymptotics){reference-type="ref" reference="chapter:asymptotics"}.\ By non-radiating we mean that the Bondi mass is constant in a neighbourhood of spatial infinity. Recall the Bondi mass loss formula for future null infinity: \[\partial_u M =-\frac{1}{32\pi} \int_{S^2} (|\Xi|^2 + |\underline{A}(F)|^2) d\mu_{\breve{\gamma}},\] where \(\Xi\) and \(\underline{A}(F)\) can be regarded as the gravitational and electromagnetic power radiated to infinity per unit solid angle, respectively. Moreover, they correspond to the leading order components of the Weyl and Faraday tensors, respectively. Hence, the condition of constant Bondi mass is equivalent to requiring the vanishing of the gravitational and electromagnetic radiation fields, i.e., the leading order terms. In fact as we will see in Lemma [\[lem:decay non-radiating\]](#lem:decay non-radiating){reference-type="ref" reference="lem:decay non-radiating"}, the components of the Weyl and Faraday tensors for non-radiating spacetimes are \(\mathcal{O}(r^{-3})\) and \(\mathcal{O}(r^{-2})\), respectively. We refer the reader to Section [\[section:CK analysis\]](#section:CK analysis){reference-type="ref" reference="section:CK analysis"} for more details.\ The conclusion of Theorem [\[thm1\]](#thm1){reference-type="ref" reference="thm1"} can be seen from two point of views. Firstly, as a rigidity result for metrics with fast decaying curvature. Secondly, it is also a inheritance of symmetry result since the gravity and electromagnetic field both turned out to be time-independent. As we will see shortly, this is not the case for other matter-energy models coupled to gravity.\ We remark that the above result can be generalised to include a massless Klein-Gordon scalar field coupled with gravity and electromagnetism. A more precise statement accompanied with a sketched proof is presented in the Appendix. We stress that the fall-off conditions required for the whole system have to be assumed at this point due to the lack of a stability result for an Einstein-Klein-Gordon system. In Theorem [\[prop:Klein-Gordon\]](#prop:Klein-Gordon){reference-type="ref" reference="prop:Klein-Gordon"} we consider the fall-off obtained from the linear analysis performed in Winicour together with a smoothness assumption at higher orders as in definition [\[def:AF\]](#def:AF){reference-type="ref" reference="def:AF"}.\ Bičák-Scholtz-Tod also tackled this Einstein-Klein-Gordon problem in (a continuation of ). They again go around the fall-off condition by requiring analyticity of the fields all the way up to infinity and conclude that time-periodic asymptotically flat Einstein-massless-Klein-Gordon systems are in fact stationary.\ The results presented in this paper are local around spatial infinity and hence are applicable to any system whose matter source is spatially compact. We mention here what is known about the case when the matter content extends to infinity. The simplest model to consider is that of a massive Klein-Gordon field coupled to Einstein equations. As mentioned above the techniques used in this work carry on to the Einstein-massless-Klein-Gordon system; however as we will see, the massive case is different.\ In, Herdeiro and Rady constructed time-periodic Einstein-Klein-Gordon solutions. They start by analysing linear perturbations of the Klein-Gordon equation on a fixed Kerr background and take them as starting points to solve the full non-linear equations numerically. The underlying family of spacetimes are asymptotically flat, stationary, axisymmetric and with regular horizons. The scalar field however does not share the symmetries since the ansatz is taken to be \(\phi(r,\theta)e^{i(m \varphi-\omega t)}\). Thus, this spacetimes can be regarded as hairy black holes bifurcating off the Kerr solution. Later on, Chodosh and Shlapentokh-Rothman,, retook the problem and proved, analytically, the existence of such spacetimes in a small neighbourhood of the Kerr family. All this is in contrast with our result where the symmetry is inherited by all the fields. The crucial difference in the assumptions is the presence of a non-zero mass for the scalar field, which acts as a non-decaying potential.\ Related to the above hairy black holes is the existence of countably many time-periodic, spherically symmetric asymptotically flat *boson stars* given by Bizoń and Wasserman in. These are solutions of the Einstein-Klein-Gordon equations where the underlying spacetime is static while the complex scalar field has a positive mass and takes the form of a standing wave \(\phi(r)e^{i\omega t}\). Therefore the matter field does not inherit the time-like symmetry.\ \ *Outline of the paper.* In **Section 2** we introduce null coordinates adapted to future null infinity. We do this following the work of Christodoulou-Klainerman, Christodoulou and Klainerman-Nicolò. In addition to providing a background, these coordinates also give us a time-like quasi-symmetry which will play the role of the candidate Killing field.\ Also, a tetrad adapted to these coordinates is defined and the structure equations are written with respect to this basis. Then the gauge conditions are explicitly spell out and we briefly explain the null hierarchy of the equations. To end this section we summarise the Christodoulou and Klainerman analysis based on the initial value formulation. In particular their deduced asymptotic expansions and the concepts of mass and radiation are reviewed.\ In **Section 3** we use the tetrad formalism to achieve the first step in the proof of the main theorem. That is, it is proven that a non-radiating spacetime is stationary to all orders at infinity. We recursively compute the connection and curvature coefficients to all orders at infinity (here is where the smoothness assumption of asymptotic flatness in definition [\[def:AF\]](#def:AF){reference-type="ref" reference="def:AF"} plays an important role in the form of asymptotic expansions in inverse powers of \(r\)). The hierarchy found by BMS, also interpreted as signature levels in the language of Christodoulou-Klainerman, helps to understand the limiting structure at future null infinity of the Einstein equations. Then the aforementioned hierarchy helps us identify levels where the equations become linear for the quantities belonging to that level. Moreover, the radiation fields can be regarded as the necessary initial data to run an induction argument and find recurrence relations. In particular, in the absence of radiation fields, all the asymptotic quantities are found to be time-independent. Also, the procedure sheds light on the way the different terms in the \(r\)-expansion of the metric and connection coefficients couple to each other via the Einstein equations.\ The above results together with an analytical condition already imply the stationarity of non-radiating electrovacuum spacetimes in a neighbourhood of infinity as in. However, one of the goals of this paper is to dispense with the analyticity assumption. We are able to remain in the smooth class by using the methods explained in Section 4 and conclude the stationarity of non-radiating spacetimes in the class of smooth metrics.\ The technique to compute the asymptotic quantities to all orders at infinity is so basic that we feel compelled to give an account here of the main idea with a toy model: An out-going wave on Minkowski spacetime. The reader will find that the procedure is straightforward and the difficulty to apply it to the Einstein equations lies only in the intricacy of the equations themselves.\ \ *Toy model.* Consider the Minkowski metric in spherical coordinates: \[\eta =-dt^2 + dr^2+ r^2 \breve{\gamma},\] where \(\breve{\gamma}\) is the round metric on the sphere \(\mathbb{S}^2\). We change to out-going null coordinates which are better suited for the problem at hand. That is, let \(u=t-r\), then the metric can be written as \[\eta =-du^2-2dudr + r^2 \breve{\gamma}.\] Note that the level sets of \(u\) are null hypersurfaces ruled by \(\partial_r\). Now, let \(\phi\) be a solution of the free wave equation, \[\square_{\eta} \phi = 0.\] Assume moreover that \(\phi\) admits an expansion of the form[^2] \[\phi (u,r,\theta^2, \theta^3) = \sum_{n=1}^{\infty} \overset{(n)}{\phi} (u, \theta^2, \theta^3) \frac{1}{r^n} ,\] which is well-behaved with respect to derivatives.\ We wish to find \(\overset{(n)}{\phi}\) in terms of \(\overset{(1)}{\phi}\) and \(\overset{(n)}{\phi_0} := \lim_{u\rightarrow-\infty} \overset{(n)}{\phi}\). To do so we consider \(X = \grad \phi\) and rewrite the wave equation as the 1st order system, \[\begin{aligned} \curl X = 0, \\ \divergence X = 0. \end{aligned}\] For easy comparison with the CK notation we define the null components of \(X\) as \[x:= X_r, \quad \underline{x}:=X_u, \quad \cancel{X} := \frac1r \breve{\grad} \, \phi\] For the time being let us denote by \(\breve{\grad}\), \(\breve{\divergence}\) and \(\breve{\Delta}\) the gradient, divergence and Laplace operators on the unit round sphere, respectively. In coordinates, the previous equations read \[\begin{aligned} \partial_r \cancel{X} &= \frac1r \breve{\grad} \, x-\frac1r \cancel{X} , \quad & \partial_u \cancel{X} &= \frac1r \breve{\grad} \, \underline{x}-\frac1r \cancel{X} , \\ \partial_r \underline{x} &= \partial_u x, \quad & \breve{\curl} \, \cancel{X} &= 0 , \\ \partial_u x &=-\partial_r \underline{x} + \frac{1}{r^2}\partial_r(r^2 x) + \frac1r \breve{\divergence} \, X. \end{aligned}\] These equations (the first column) imply the following recurrence relations for the asymptotic quantities, \[\begin{aligned} -(n-1) \overset{(n)}{\cancel{X}} &= \breve{\grad} \overset{(n)}{x}, \\ -n \overset{(n)}{\underline{x}} &= \partial_u \overset{(n+1)}{x}, \\ \partial_u \overset{(n+1)}{x} &= n\overset{(n)}{\underline{x}}-(n-2) \overset{(n)}{x}-\breve{\divergence} \, \overset{(n)}{\cancel{X}}. \end{aligned}\] From these we can deduce an evolution equation for each \(\overset{(n+1)}{x}\) in terms of lower order data, \[\begin{aligned} 2\partial_u \overset{(n+1)}{x} = \frac{1}{n-1} \breve{\Delta} \overset{(n)}{x}-(n-2) \overset{(n)}{x}, \end{aligned}\] which can be solved given initial data \(\overset{(n+1)}{x_0}\) at \(u=-\infty\). Note that the asymptotic expansion for \(\phi\) implies \(\overset{(1)}{x}=0\) and \(\overset{(2)}{x}=-\overset{(1)}{\phi}\), in particular, we can compute \(\overset{(n+1)}{x}\) inductively for \(n>2\) assuming we know the radiation field \(\overset{(1)}{\phi}\).\ Therefore, we can compute \(\overset{(n)}{x}\), \(\overset{(n)}{\underline{x}}\) and \(\overset{(n)}{\cancel{X}}\) for all \(n\in \mathbb{N}\) along future null infinity in terms of the radiation field \(\overset{(1)}{\phi}\) and the "pole moments at spatial infinity" \(\overset{(n)}{x_0}\), \(n\in \mathbb{N}\). \(_{\blacksquare}\)\ This toy model reflects perfectly well the asymptotic relations when the equations can be extended smoothly up to infinity. The analysis for Maxwell and Einstein equations is completely analogous albeit more technical. The previous relations will be used in the Appendix when we discuss the extension of the main theorem to include a massless Klein-Gordon scalar field.\ In **Section 4** we complete the proof of the main theorem. We explain how the previous asymptotic values at infinity can be translated to a local result in the spacetime. That is, the condition of stationarity to all orders at infinity is extended to a neighbourhood of spatial infinity.\ The main technical tool is that of Carleman estimates. These are a priori wave estimates for functions with fast decay (faster than any polynomial) towards infinity. They can also be thought as energy estimates adapted to time-like boundaries conditions where now we aim at controlling the bulk term while making the boundary terms vanish. A key ingredient in these inequalities is a 1-parameter weight function which accounts for the exponential rate of decay. With the help of this weight, it is possible to bound the \(L^2\)-norms of a function \(\phi\) and its first derivatives in terms of \(\square \phi\), provided that the boundary of the region under consideration satisfies the so-called pseudo-convex condition.\ At this point we make a parenthesis to look for wave equations satisfied by the components of the deformation tensors \(\mathcal{L}_T C\) and \(\mathcal{L}_T F\). These are the Ionescu-Klainerman tensorial equations,. We need to revise them in order to cope with a non-vacuum spacetime. These wave equations, written in suitable coordinates, in conjunction with the Carleman estimates are used to find \(L^2\)-bounds for the components of the deformation tensors and its first derivatives. These bounds depend on an arbitrarily large parameter which can be taken to infinity to conclude the vanishing of the functions in the interior.\ Finally, an **Appendix** is included where we state a variation of Theorem [\[thm1\]](#thm1){reference-type="ref" reference="thm1"} to include a Klein-Gordon field, as well as the corresponding recurrence relations. We also include a brief review of the tetrad formalism employed in this paper, as well as the CK notation. Finally a table is included to compare the CK and NP notations. # Coordinates and gauge conditions {#section:tetrad and coordinates} Here we construct the necessary out-going null coordinates in the context of asymptotically flat spacetimes. These are given by a foliation by null hypersurfaces complemented with a transversal coordinate given by the candidate Killing field, which is roughly speaking, a time-like symmetry to first order. The existence and construction of such foliation is taken from Christodoulou.\ Let \((\Sigma_0, h,K)\) be a strongly asymptotically flat initial data set and consider an exhaustion of \(\Sigma_0\) by balls \(B_d\), where \(d\) is an asymptotically flat radial coordinate on \(\Sigma_0\). Within the corresponding Cauchy development define \(C_d^-\), resp. \(C_d^+\), to be the ingoing, resp. outgoing, null hypersurface consisting of geodesic rays emanating from \(S_d =\partial B_d\). Fix some \(S_0=\partial B_{d_0}\), thanks to Christodoulou-Klainerman, we can choose it so that the generators of \(C_0^{\pm} := C_{d_0}^{\pm}\) have no future end points. Finally let \(S_0^* = C_0^+ \cap C_{d^*}^-\) for some \(d^*\) large.\ At \(S_0^*\) we can define the future-directed null outgoing and ingoing normal vectors \(L\) and \(\underline{L}\), respectively, by the conditions: \[\langle L, \underline{L}\rangle =-2, \quad \langle L,L \rangle = \langle \underline{L}, \underline{L}\rangle = \langle L, X \rangle = \langle \underline{L}, X \rangle = 0, \quad X\in \mathcal{T}_pS_0^*.\] These conditions fix \(L, \underline{L}\) up to rescaling by a function \(a\) on \(S_0^*\): \[L \mapsto a L, \qquad \underline{L} \mapsto \frac1a \underline{L}.\] Next, \((L, \underline{L})\) are extended to \(C_{d^*}^-\) as follows: Let \(\tilde{\underline{L}}\) be the extension of \(\underline{L}\) by geodesic flow, that is, \[\nabla_{\tilde{\underline{L}}}\tilde{\underline{L}} = 0, \quad \textrm{along } C_{d^*}^-\text{, with} \quad \tilde{\underline{L}} = \underline{L} \quad \textrm{on } S_0^*.\] With the help of its affine function \(t\) (that is, \(\tilde{\underline{L}} \cdot t = 1\) on \(C_{d^*}^-\) and \(t=0\) on \(S_0^*\)) we define the *retarded time* function \(u\) on \(C_{d^*}^-\) to be the area parameter on a level set of \(t\), \[u(t) := 2 (r^*_0-r^*_t) ,\] here \(r^*_t\) is the area radius of \(S_t^*\), the level set of \(t\) on \(C_{d^*}^-\). Now, we set \(\underline{L}\) to be the velocity field of \(u\) on \(C_{d^*}^-\), that is, \(\underline{L} = \frac{dt}{du}\tilde{\underline{L}}\) along the null geodesics ruling \(C_{d^*}^-\). In particular, \(\underline{L}(u)=1\).\ Then, let \(L\) be defined on \(C_{d^*}^-\) as the null conjugate to \(\underline{L}\) with respect to level sets of \(u\), \(S_u^*\): \[\langle L, \underline{L}\rangle =-2, \quad \langle L, X \rangle = 0, \quad X\in \mathcal{T}_pS_u^*.\] Finally we extend \(u\) by choosing a solution of the eikonal equation, \[g^{\alpha \beta} \partial_{\alpha} u \partial_{\beta} u = 0\] with the already prescribed value of \(u\) along \(C_{d^*}^-\). Let \(C_u^+\) be the level sets of \(u\), we extend \(L\) by geodesics, \[\nabla_L L = 0, \quad \textrm{on the relevant open set.}\] Let \(s\) be the affine parameter: \(L\cdot s = 1\) and \(s|_{S_u^*}=r_u^*\).\ The above procedure also gives coordinates \((s,u, \theta^i)\), by choosing \((\theta^i)\) on \(S_0^*\) then \(u-\)flowing them along \(\underline{L}\) on \(C_{d^*}^-\) to obtain coordinates \((u, \theta^i)\) and then \(s-\)flowing the latter along \(L\) to cover an open set near infinity.\ It worth noticing that the above construction is engineered so that \(L\), \(\underline{L}\) and the candidate Killing field \(\partial_u\) are related to leading order in the usual way, that is, \[\begin{aligned} \partial_u = \frac12 \left( L + \underline{L}\right) + \mathcal{O}(r^{-1}). \end{aligned}\] ## Gauge Conditions It will be convenient to work with a tetrad \[\begin{aligned} \label{tetrad} \begin{aligned} e_0 & := L = \partial_s, \\ e_1 & := \underline{L} = f^0\partial_s + f^1 \partial_u + f^i\partial_{\theta^i}, \\ e_i & := {h_i}^j \partial_{\theta^j} , \end{aligned} \end{aligned}\] satisfying \[\begin{aligned} \label{metric-matrix} \eta_{\mu\nu} := \langle e_{\mu}, e_{\nu} \rangle = \left( \begin{array}{ccc} 0 &-2 & 0 \\ -2 & 0 & 0 \\ 0 & 0 & \textrm{Id}_{2}\\ \end{array}\right). \end{aligned}\] The above change of basis can be written compactly in terms of an *orthonormalisation matrix*, \[\begin{aligned} e_{\mu} = {h_{\mu}}^{a} \partial_{x^{a}}. \end{aligned}\] That is, \[\begin{aligned} {h_{\mu}}^{a} := \left( \begin{array}{cc} \begin{array}{cc} 1 & f^0 \\ 0 & f^1 \end{array} & \begin{array}{cc} 0 & 0 \\ 0 & 0 \end{array} \\ \begin{array}{cc} 0 & f^2 \\ 0 & f^3 \end{array} & \begin{array}{c} {h_{i}}^k \end{array} \end{array} \right). \end{aligned}\] The metric components are related to the orthonormalisation matrix by \(g^{ab} = {h_{\mu}}^{a} \eta^{\mu \nu} {h_{\nu}}^{b}\), where \(g^{ab}\) is the inverse matrix of \(g_{ab} := \langle \partial_{x^{a}}, \partial_{x^{b}} \rangle\). Explicitly, \[\begin{aligned} g^{ab} = \left( \begin{array}{cc} \begin{array}{cc} 0 &-\frac12 f^1 \\ -\frac12 f^1 &-f^1 f^0 + (f^2)^2 + (f^3)^2 \end{array} & \begin{array}{cc} 0 & 0 \\ f^i {h_i}^2 & f^i {h_i}^3 \end{array} \\ \begin{array}{cc} 0 \qquad \qquad \; & f^i {h_i}^2 \qquad \quad \\ 0 \qquad \qquad \; & f^i {h_i}^3 \qquad \quad \end{array} & \begin{array}{c} {h_{i}}^k \delta^{ij} {h_j}^l \end{array} \end{array} \right). \end{aligned}\] **Remark.** The function \(f^1\) is constant and we can choose it so that \(f^1=2\). This can be showed by first noting that \(\partial_s \langle \partial_u, \partial_s \rangle = \langle \nabla_{\partial_s} \partial_u, \partial_s \rangle + \langle \partial_u, \nabla_{\partial_s} \partial_s \rangle = \langle \nabla_{\partial_u} \partial_s, \partial_s \rangle = \frac12 \partial_u \langle \partial_s, \partial_s \rangle = 0\), hence \(\langle \partial_u, \partial_s \rangle\) is constant along \(s-\)lines. Then the assertion follows from \(-2 =\langle L, \underline{L} \rangle = f^1 \langle \partial_u, \partial_s \rangle\).\ We call all these choices the *null-geodesic gauge conditions*.\ The connection coefficients are the components of the derivative operator with respect to the tetrad \(e_{\mu}\): \[\begin{aligned} \omega_{\lambda \mu \nu} := \langle \nabla_{e_{\lambda}} e_{\nu}, e_{\mu} \rangle. \end{aligned}\] They satisfy \(\omega_{\lambda \mu \nu} =-\omega_{\lambda \nu \mu}\), this is a consequence of \(\eta_{\mu \nu}\) being a constant matrix. Moreover, the connection coefficients and the orthonormalisation matrix are related by the *frame equations* (the torsion-free property of the connection in this language, see Appendix for details). In the null gauge case they read \[\begin{aligned} \omega_{001} &= 0, \quad & -2\zeta_i = \omega_{10i} &= \omega_{i01} =-V_i, \\ \omega_{00i} &= 0 , \quad & -\chi_{ij}= \omega_{i0j} &= \omega_{j0i} =-\chi_{ji}, \\ -2\underline{\zeta}_i=\omega_{01i} &= \omega_{i10}=V_i , \quad & -\underline{\chi}_{ij} = \omega_{i1j} &= \omega_{j1i} =-\underline{\chi}_{ji}, \\ e_0(f^0) &= \frac12 \omega_{110} = 2 \underline{\omega}, \quad & e_i(f^0) &= \frac12 \omega_{11i} =\lambda_i, \\ e_0(f^i) &= ( {\omega_0}^j{}_1-{\omega_1}^j{}_0 ){h_j}^i ,\quad & e_i({h_j}^l)-e_j({h_i}^l) &= ({\omega_i}^k{}_j-{\omega_j}^k{}_i ){h_k}^l, \\ e_0({h_i}^j) &= ( {\omega_0}^l{}_i-{\omega_i}^l{}_0 ){h_l}^j, \quad & e_1({h_i}^j)-e_i(f^j) &= ( {\omega_1}^l{}_i-{\omega_i}^l{}_i ){h_l}^j. \end{aligned}\] We refer the reader to the Appendix for a review of the CK notation and a comparison with the NP spin coefficients and the slight variations adopted in this paper.\ The previous construction of coordinates and tetrad fixes \(e_0=L\) and \(e_1=\underline{L}\) up to rescaling \[L \mapsto a L, \qquad \underline{L} \mapsto \frac1a \underline{L} ,\] where \(a\) is a function of the sphere \(S^*_0\). In view of the corresponding scaling of the second fundamental forms, \[\chi \mapsto a \chi, \qquad \underline{\chi} \mapsto \frac1a \underline{\chi}.\] The function \(a\) can be chosen so that \[\lim_{r \rightarrow \infty} ( \textrm{tr} \chi + \textrm{tr} \underline{\chi} ) = 0 .\] In other words, it is required that the first variation of area of \(S_0^*\) along geodesics defined by the direction of \(L+\underline{L}\) tends to \(0\) at infinity. We define then \[T := \partial_u\] to be the candidate Killing vector field.\ Finally, we can also arrange for \(\omega_{0ij}=0\) as in Lemma [\[lem:comoving gauge\]](#lem:comoving gauge){reference-type="ref" reference="lem:comoving gauge"}. This is done by solving an ODE, so we still have the freedom to choose an initial condition for the system. This translates to a freedom to adapt the frame \(\{e_i\}\) at infinity. More precisely, we are still free to choose the leading coefficient in the asymptotic expansion of \({h_i}^j\). Since we are approaching the round unit-sphere at infinity, we can set \[\begin{aligned} \overset{(1)}{{h_i}^j} = \left( \begin{array}{cc} 1 & 0 \\ 0 & \frac{1}{\sin\theta} \\ \end{array}\right). \end{aligned}\] We call the choice \(\omega_{0ij}=0\) the *co-moving frame gauge*.\ In the next section we work with the above choice of gauge conditions. While the null-geodesic gauge is essential for the procedure, the co-moving frame gauge is used mostly for convenience. The arguments go through provided that we think of \(\omega_{0ij}\) and \(\overset{(1)}{{h_i}^j}\) as given data.\ Note that the frame equations imply that \(\chi_{ij}, \underline{\chi}_{ij}\) are symmetric 2-tensors, \(2\zeta_i=V_i=-2\underline{\zeta}_i\), \(\begin{align} e_0(f^0) &= 2 \underline{\omega}, \quad & e_i(f^0) &= \lambda_i, \\ e_0(f^i) &=-\zeta^k {h_k}^i ,\quad & e_i({h_j}^l)-e_j({h_i}^l) &= ({\omega_i}^k{}_j-{\omega_j}^k{}_i ){h_k}^l, \\ e_0({h_i}^j) &=-{\chi_i}^k{h_k}^j, \quad & e_1({h_i}^j)-e_i(f^j) &= ( {\omega_1}^l{}_i-{\omega_i}^l{}_i ){h_l}^j. \end{align} \subsection{Structure equations} Recall the formula for the Riemann curvature tensor components in terms of the connection coefficients: \[ R_{\rho \sigma \mu \nu} = e_{\rho} (\omega_{\sigma \mu \nu} )-e_{\sigma} (\omega_{\rho \mu \nu} )-{\omega_{\rho \mu}}^{\alpha} \omega_{\sigma \nu \alpha} + {\omega_{\sigma \mu}}^{\alpha} \omega_{\rho \nu \alpha}-{\omega_{\sigma \rho}}^{\alpha} \omega_{\alpha \mu \nu} + {\omega_{\rho \sigma}}^{\alpha} \omega_{\alpha \mu \nu} . \] The splitting into trace and traceless part is the usual \[ R_{\rho \sigma \mu \nu} = C_{\rho \sigma \mu \nu} + \frac{1}{2} ( \eta_{\rho \mu} S_{\sigma \nu}-\eta_{\sigma \mu} S_{\rho \nu} + \eta_{\sigma \nu} S_{\rho \mu}-\eta_{\rho \nu} S_{\sigma \mu} ), \] where\)C\_\(and\)S\_\(are the components of the Weyl and Schouten tensors, respectively. \\ The \emph{structure equations} are the combination of these two preceding equations, we write them schematically as \begin{align} \label{structure equations} C + \eta \wedge S = e \wedge \omega + \omega \wedge \omega. \end{align} The Einstein equations appear implicitly when we regard the Schouten tensor as the stress-energy tensor of the matter/energy content. \\ In order to understand the structure equations in the null-geodesic gauge we introduce the concept of the signature of a tensor, as discussed in p 148. We just quote here their clear exposition: \\ \emph{Given any covariant tensor \)U\( at a point of spacetime, we define a null component of it to be any tensor tangent to the sphere \)S_{u,s}\( at a point, which is derived from \)U\( by contractions with either \)e_0\( or \)e_1\( and projections to \)S_{u,s}\(. To any such component we assign a signature that it is defined as the difference between the total number of contractions with \)e_0\( and the total number of contractions with \)e_3\(. We are now ready to state the following heuristic principle. \\ \\ \textbf{Principle of Conservation of signature:} Consider an arbitrary covariant tensor \)U\( that can be expressed as a multilinear form in an arbitrary number of covariant tensors \)U_1 \ldots U_p\(, with coefficients depending only on the spacetime metric and its volume form. Then the signature of any null term of U, expressed in terms of the null components of \)U_1\ldots U_p\(, is equal to the sum of the signatures of each constituent in the decomposition.}\\ In particular the signature of a tensor does not change by lowering/raising indices. We have included in Table 1 of the Appendix the signature of the different null components relevant to us. The above principle will allow us to guess the quadratic terms appearing in the structure equations without performing the calculations. Indeed, we already know that in the null-geodesic gauge the only connection coefficient with signature equal 2 vanishes; this fact is the one responsible for the success of the hierarchy-induction argument explained in Section \ref{section:quantities} to obtain the asymptotic quantities to all orders at infinity.\\ Now we illustrate how to apply this principle to the structure equations. We will denote by\)U(l)\(a null component of\)U\(with signature\)l\(. The signature 2 case yields schematically: \begin{align*} C(2) + \eta \wedge S(2) &= e(1) \omega(1) + e(0)\omega(2) + \omega(1) \wedge \omega(1) + \omega(0) \wedge \omega(2), \\ &=e(1) \omega(1) + \omega(1) \wedge \omega(1). \end{align*} where in the second line we have used the fact that the only connection coefficient with signature 2,\)\_0\(, vanishes in the null-geodesic gauge. More generally we have the following: \\ \\ \emph{\textbf{Hierarchy of structure equations:} When working in the null-geodesic gauge there is no connection coefficient of signature \)l-2\( on the right-hand side of the structure equation of signature \)l\(, \)l=2,1,0\(.}\\ This is precisely the hierarchy found by BMS,. The exact form of the structure equations is presented in equations (\ref{hierarchy0})-(\ref{hierarchy4}). \subsection{Christodoulou-Klainerman estimates } \label{section:CK analysis} In their book The Global Nonlinear Stability of the Minkowski Space, Christodoulou and Klainerman proved that any strongly asymptotically flat initial data set that satisfies a global smallness assumption, leads to a unique globally hyperbolic and geodesically complete solution of the Einstein vacuum equations. Moreover the development is globally asymptotically flat in the sense that the curvature vanishes at infinity in all directions. \\ Here we list some conclusions of their work regarding the asymptotic behaviour of solutions suitably close to Minkowski spacetime. Of particular interest are the decay rates obtained for the connection coefficients and the null components of the Weyl tensor. It is observed that the peeling property may not hold in general. \\ While all the previous was done in the vacuum case, the extension to include an Electromagnetic field was worked out by Zipser in her PhD dissertation. The same conclusion for an Einstein-massless-Klein-Gordon system is expected to hold; however to the author's knowledge such result is still lacking in the literature. \\ Using the null out-going coordinates as constructed in the previous section, Christodoulou-Klainerman showed that, for fixed\)u\(,\)S\_s,u\(converges to the standard round sphere embedded in Minkowski, that is, let\[be the induced metric on\)S\_s,u\(, then \begin{align*} \lim_{u; r \rightarrow \infty} r^{-2} \cancel{g} &= \breve{\gamma}, \qquad & \lim_{u; r \rightarrow \infty} K[r^{-2} \cancel{g} ] &= 1, \\ \lim_{u; r \rightarrow \infty} r \tr \chi &= 2, \qquad & \lim_{u; r \rightarrow \infty} r \tr \underline{\chi} &=-2, \end{align*} here the limit\)\_u; r \(is taken along a level set of\)u\(while letting the area function\)r\(tend to infinity;\]is the standard round metric on\)\^2\(and\)K\(is the Gauss curvature.\\ Denote by\[the trace-free part of\], then to next order they obtained \begin{align*} \lim_{u; r \rightarrow \infty} r^{2} \hat{\chi} &= \Sigma, \qquad && \lim_{u; r \rightarrow \infty} r(r\tr\chi-2) = H, \\ \lim_{u; r \rightarrow \infty} r \underline{\hat{\chi}} &= \Xi, \qquad && \lim_{u; r \rightarrow \infty} r(r\tr\underline{\chi}-2) = \underline{H}, \end{align*} where\[,\]are symmetric trace-less\)u\(-dependent 2-covariant tensors defined on\)\^2\(and\)H\(,\)[H]{.ul}\(are\)u\(-independent functions on\)\^2\(of vanishing mean. Moreover, \begin{align*} |\Xi(u, \cdot)|_{\breve{\gamma}} \leq C(1 + |u|)^{-3/2}. \end{align*} Recall the null decomposition the Weyl tensor with respect to a tetrad\){e_0, e_1, e_i }\(, \begin{align} \label{null Weyl} \alpha_{ij} &= C_{i0j0}, \qquad && \underline{\alpha}_{ij} = C_{i1j1}, \\ 2 \beta_i &= C_{i010}, \qquad && 2 \underline{\beta}_i = C_{i110}, \\ 4 \rho &= C_{1010}, \qquad && \sigma = C_{0123}. \end{align} The curvature null components also decay. Explicitly, there exist a symmetric trace-less 2-tensor\)[A]{.ul}\(, a 1-form\)[B]{.ul}\(and functions\)P\(,\)Q\(defined on\)\^2\(and\)u\(-dependent satisfying \begin{align} \label{null Weyl2} \lim_{u; r \rightarrow \infty} r \underline{\alpha} &= \underline{A}, \qquad & \lim_{u; r \rightarrow \infty} r^2 \underline{\beta} &= \underline{B}, \\ \lim_{u; r \rightarrow \infty} r^{3} \rho &= P, \qquad & \lim_{u; r \rightarrow \infty} r^3 \sigma &= Q. \end{align} Moreover, these limits decay in\)u\(as follows, \begin{align} |\underline{A}(u, \cdot)| & \leq C (1+|u|)^{-5/2}, \qquad & |\underline{B} (u, \cdot)| &\leq C(1+|u|)^{-3/2}, \\ |P(u, \cdot)-\overline{P}(u)| & \leq (1+|u|)^{-1/2}, \qquad & |Q(u, \cdot)-\overline{Q}(u)| &\leq (1+|u|)^{-1/2}, \end{align} and\)\_u - (u)=0\(,\)\_u - (u)=0\(.\\ This is all consistent with the presence of peeling. In contrast, they could only deduce a weaker fall-off for the remaining curvature components, namely, \begin{align*} |\alpha|_{\cancel{g}} = \mathcal{O} (r^{-7/2}), \qquad |\beta|_{\cancel{g}} = \mathcal{O} (r^{-7/2}), \end{align*} For completeness we mention that the presence of the full peeling property would be achieved by\)\|\|\_ = (r\^-5)\(and\)\|\|\_ = (r\^-4)\(. \\ Finally, the following relations are implied by the structure equations, \begin{align*} \partial_u \Xi &=-\frac12 \underline{A}, \qquad & \breve{ \cancel{\divergence} }\ \Xi &= \underline{B}, \\ \partial_u \Sigma &=-\frac12 \Xi, \qquad & \breve{ \cancel{\divergence} }\ \Sigma &= \frac12 \breve{\cancel{\nabla}} H + Z, \\ \partial_u \underline{H} &=-\frac12|\Xi|^2, \end{align*} where\[and\]stand for the connection and divergence operators on the round sphere, respectively. \subsubsection{Mass} The \emph{Hawking mass} enclosed by a 2-sphere\)S\_s,u\(is defined to be, \begin{align*} m(s,u) = \frac{r(s,u)}{2} \left( 1 + \frac{1}{16\pi} \int_{S_{s,u}} \tr \chi \tr \underline{\chi} \right), \end{align*} Christodoulou and Klainerman also deduced from their analysis the decay rate of the Hawking mass as well as its evolution equation. Explicitly, there exists a function\)M(u)\(such that along a level set of\)u\(we have \begin{align*} m(s,u) = M(u) + \mathcal{O} (r^{-1}). \end{align*} The function\)M(u)\(is called the Bondi mass. Furthermore, it satisfies the Bondi mass formula \begin{align*} \partial_u M =-\frac{1}{32\pi} \int_{\mathbb{S}^2} |\Xi(u, \cdot)|^2 d\mu_{\breve{\gamma}}. \end{align*} \subsubsection{Gravity coupled with electromagnetism} The previous analysis was done for vacuum solutions suitably close to Minkowski spacetime. The generalisation to an Einstein-Maxwell system was carried out by Zipser in her PhD thesis. She found that the null components of the Faraday tensor decay as expected from the linear analysis (cf. ). \\ Recall the null decomposition of the Faraday tensor. \begin{align} \label{null Faraday} \alpha(F)_i &= F_{i0} , \quad& \underline{\alpha}(F)_i &= F_{i1} , \\ \rho (F) &= \frac12F_{10}, \quad& \sigma(F) &= F_{23}. \end{align} As a consequence of Zipser's analysis we have that the following limits exist, \begin{align} \label{null Faraday2} \lim_{u; r \rightarrow \infty} r \underline{\alpha}(F) &= \underline{A}(F), \\ \lim_{u; r \rightarrow \infty} r^{2} \rho(F) &= P(F) , \\ \lim_{u; r \rightarrow \infty} r^2 \sigma (F) &= Q(F). \end{align} However, full peeling is not available for the remaining component. Instead we have the weaker estimate along a level set of\)u\(, \begin{align} \label{null Faraday3} \alpha(F) = \mathcal{O}(r^{-\frac52}). \end{align} The generalisation made by Zipser has as a consequence that the Hawking mass also has a limit, that we may call Bondi mass, and this limit satisfies the following Bondi mass formula: \begin{align} \label{Bondi mass formula} \partial_u M =-\frac{1}{32\pi} \int_{\mathbb{S}^2} |\Xi(u, \cdot)|^2 + |\underline{A}(F)(u, \cdot)|^2 d\mu_{\breve{\gamma}}. \end{align} For this reason we call\)\_ij\(and\)[A]{.ul}(F)\_i\(the \emph{radiation fields}. \begin{definition} \label{non-radiating} \emph{Non-radiating spacetimes.} An asymptotically flat solution to the Einstein-Maxwell equations is called non-radiating (towards future null infinity) if the Bondi mass\)M(u)\(is constant as a function of retarded time\)u\(. \end{definition} This is strongly related to the order of decay of the curvature at infinity. Indeed we have the following lemma which is a consequence of Corollary \ref{cor:non-radiating}. \begin{lem} \label{lem:decay non-radiating} Assume we are working with the tetrad, coordinates and gauge conditions imposed in this section, then an asymptotically flat solution of the Einstein-Maxwell equations is non-radiating towards future infinity if and only if \begin{align*} C_{\rho \sigma \mu \nu} = \mathcal{O}(r^{-3}) \quad \textrm{and} \quad F_{\mu \nu} = \mathcal{O}(r^{-2}), \end{align*} \end{lem} \section{Asymptotic behaviour of the fields} \label{chapter:asymptotics} \setcounter{footnote}{1} In this section we start by presenting the main technical assumptions required to `push' the Einstein equations to infinity at all orders. Namely we estate the class of functional spaces and asymptotically flat spacetimes that we will consider. Without further comment all the spacetimes in this paper will be assumed asymptotically flat in the sense of definition \ref{def:AF} below. \\ Next, we compute asymptotic quantities at infinity. We follow Bi\v{c}\'ak-Sholtz-Tod and Alexakis-Schlue in their asymptotic analysis and extent their results to show that the radiation fields determine the evolution of the metric to all orders at infinity. Indeed, by assuming an asymptotic expansion we can take the limit for each order of the structure equations and find algebraic and\)u\(-transport relations at infinity at a given order. These relations are well-suited for an induction process provided that we know the radiation fields,\)\_ij\(and\)[A]{.ul}(F)\_i\(along null infinity and initial conditions for\) \(and\) \(at a sphere. The main results of this section state that these data determines the solution to all orders at infinity, see Proposition \ref{prop:recurrence relations} and its ensuing corollaries. \\ Before giving our definition of asymptotic flatness, we digress briefly about the different coordinate systems that we will be using. These are time-space, null out-going and double-null coordinates. \begin{itemize} \item \emph{Time-space coordinates.} These are important as they split naturally the initial data and the evolution equations. It is a definite hypothesis of this work that the asymptotic properties at spatial infinity of the Cauchy data are preserved under the evolution; this is the content of part\)i)\(in definition \ref{def:AF}, below. It would be desirable to deduced this fact\footnote{That is, that smoothness at spatial infinity is preserved for finite but arbitrarily large time. Then, the desired expansion would hold for, e.g., time-periodic spacetimes.} from the Einstein equations; this would require a `preservation of regularity estimate for all orders at spatial infinity' which appears to be missing in the literature. \\ We stress here that the CK analysis is not suited for this problem as their gauge conditions are naturally adapted to null inifnity. On the other hand, Friedrich's analysis of his conformal equations,, provides strong evidence about the preservation of regularity at spatial infinity, since the equations can be extended regularly there.\\ \item \emph{Null out-going coordinates.} Their construction is one of the crucial steps in the CK analysis. These coordinates are used to define the 15-parameter family of quasi-symmetries (corresponding to translations, Lorentz rotations, scaling and inverted translations in Minkowski spacetime) which serve to compare all geometric quantities to that of flat spacetime. Of particular importance to us is the construction of the time-like quasi-translations, which will play the role of candidate Killing field.\\ In they define\)u\(by setting initial conditions on a ``last Cauchy hypersurface'' and then extending by geodesics. Here we have followed Christodoulou and Klainerman-Nicol\`o in that initial conditions for\)u\(are set on a null hypersurface by means of the area radius parameter and then are extended by geodesics as before. With respect to this coordinate system the candidate Killing field is given by\)T=\_u\(. \item \emph{Double-null coordinates.} In order to capture the decaying properties of a function towards both past and future null infinities one is forced to use double-null coordinates. Indeed, Klainerman-Nicol\`o revised the methods of using a double-null foliation which is better suited for the analysis of the structure equations. \\ The relevance of these coordinates in this paper relies on the fact that the Carleman estimates, Theorem \ref{thm:Carleman}, are stated with respect to double-null coordinates. It is important to stress, however, that the type of metrics allowed by definition \ref{def:AF} do not satisfy a priori the decay conditions, (\ref{metric AS}), required for the Carleman estimates to hold. However, in the context of non-radiating spacetimes, it can be proved that the Cauchy development of initial data close to Kerr-Newman which happens to be non-radiating do admit double-null coordinates of the desired form (\ref{metric AS}). We will prove this in Lemma \ref{lem:doublenull}. \end{itemize} \subsection{Asymptotic conditions} Now we present the functional spaces relevant for our class of asymptotically flat spacetimes. \\ We say that a function\)f *Regularity at spatial infinity.* We say that \((\mathcal{M}, g)\) is regular at spatial infinity if there are coordinates \((t,r, \vartheta^i)\) such that for large \(r\), the metric admits an expansion of the form \[\begin{aligned} g =-dt^2 + dr^2 + r^2 \breve{\gamma}_{ij} d\vartheta^i d\vartheta^j + g^{\infty} \end{aligned}\] where \(\breve{\gamma}\) is the round metric on \(\mathbb{S}^2\) and the components \(g_{\alpha\beta}^{\infty}\) of \(g^{\infty}\) belong to the class \(\mathcal{O}^{\infty}_{2} (r^{-1})\). *Smoothness at null infinity.* \((\mathcal{M},g)\) is called *smooth at future null infinity* if, with respect to the null out-going coordinates \((u,s, \theta^2, \theta^3)\) previously defined, the vanishing of a component of a geometric quantity, \(w = w (s, u, \theta^1, \theta^2)\), in the sense that \[\label{vanish} \lim_{u; r \rightarrow \infty} r^q w = 0,\] implies \(r^q \omega = \mathcal{O}(r^{-1})\) and the existence of the limit of \(w\) to order \(q+1\), in the sense that \[\partial_s (r^{q+1} w) = \mathcal{O}(r^{-2}).\] Moreover, we require a mild angular regularity condition, namely, if \(w\) vanishes in the sense of ([\[vanish\]](#vanish){reference-type="ref" reference="vanish"}) then so do its angular derivatives: \[\lim_{u; r \rightarrow \infty} r^q \partial_{\theta^i} w = 0.\] Analogously, we require these conditions to hold at past null infinity. *Weak peeling.* We also require that the Weyl tensor decays accordingly to \[\begin{aligned} |\underline{\alpha} |_{\cancel{g}} &= \mathcal{O}(r^{-1}) \\ |\underline{\beta} |_{\cancel{g}} &= \mathcal{O}(r^{-2}) \\ |\rho|, |\sigma| &= \mathcal{O}(r^{-3}) \\ |\partial_u \beta |_{\cancel{g}} &= \mathcal{O}(r^{-4}) \\ |\partial_u \alpha|_{\cancel{g}} &= \mathcal{O}(r^{-5}) \end{aligned}\] *Compatibility of \(I^+\) and \(I^-\) at spatial infinity.* We say that the *poles are finite towards spatial infinity* if all the components of the curvature \(\kappa = \alpha, \beta, \rho, \sigma, \underline{\beta}, \underline{\alpha}\) admit an expansion for large \(r\) (near future and past null infinities \(I^+\), \(I^-\)), \[\kappa \sim \sum_{n=0}^{\infty} \overset{(n+1)}{\kappa}_{\pm} (u_{\pm}, with the property that \[ \lim_{u_{\pm} \rightarrow \mp \infty} | \overset{(n)}{\kappa_{\pm}} (u_{\pm}, \end{enumerate} \end{definition} These are all technical conditions which are sufficient to allow us to deal with the Einstein equations one order at a time at infinity. Some remarks are in order. \\ \\ \textbf{Remarks.} \begin{itemize} \item The first condition states that the initial regularity assumptions at spatial infinity are propagated through out the evolution. This property may not hold in general since logarithmic divergences are expected to appear when \(t\rightarrow \infty\). However, it should be possible to prove its validity for arbitrarily large \(t\) by using Friedrich's conformal equations,. Then condition \(i)\) would hold for time-periodic spacetimes. \item The purpose of condition \(ii)\) is to rule out \emph{logarithmic singularities}. For example, the function \(\frac{\log r}{r}\) decays at infinity but it is not \(\mathcal{O}(r^{-1})\). It is known that these kind of singularities are present in Cauchy developments of asymptotically flat initial data and actually are expected to be ubiquitous in dynamical spacetimes. However, as remarked by Christodoulou the coefficients accompanying the logarithmic singularities are time-independent for relevant systems, e.g., an \(N\)-body system with no incoming radiation. Therefore we can get rid off the logarithmic singularity by taking an extra time-derivative. In the context of non-radiating spacetimes taking an extra time-derivative does not affect the general argument thanks to the \emph{finiteness of poles condition}, condition \(iv)\) which prohibits polynomial growth. \\ Summarising, at first glance it would appear that smoothness at null infinity is too strong an assumption; however, thanks to the CK analysis, we can circumvent the problem by noting that an extra time-derivative makes the geometric quantities smooth at null infinity. The same issue arises when discussing the Peeling property. It does not hold in general for all the components of the Weyl tensor, but when the failure is understood as being time-independent, an extra time-derivative allows to state the weak peeling property as in \(iii)\) of the above definition.\\ Moreover, in the context of time-periodic spacetimes, it is easy to prove that the affine parameter \(s\) is comparable to \(r\) by appealing to a fundamental domain argument and condition \(i)\). That is, in a time-periodic spacetime the regularity assumption at null infinity follows from the regularity condition at spatial infinity.\\ At this point it is also worth stressing that the \(r\) function appearing in time-space coordinates is also compatible with the \(r\) function appearing in double-null coordinates. This is a consequence of Lemma \ref{lem:doublenull}. Therefore, when discussing asymptotic properties, we can freely use any of these \(r\) or \(s\) as parameters along the outgoing null geodesics. \item Finally, there are dynamical spacetimes satisfying the above regularity assumptions. Indeed, Cutler and Wald,, and Chru\'sciel and Delay, provided examples of electrovacuum and vacuum spacetimes, respectively, having a smooth null infinity. \end{itemize} \subsection{Quantities to all orders at infinity} \label{section:quantities} Following the the work of Bi\v{c}\'ak-Sholtz-Tod and Alexakis-Schlue we now proceed to show that the radiation fields \(\Xi_{ij}:=\overset{(1)}{\hat{\chi}_{ij}}\) and \(\underline{A}(F)_i :=\overset{(1)}{\underline{\alpha}(F)_i} \) characterise the metric at infinity to all orders (up to stationary data to be described below). To do so we rely on the hierarchy discovered by BMS which is also tied to the splitting of the Weyl tensor into its null components and can be interpreted as signature levels. \\ Formally, in Proposition \ref{prop:recurrence relations} we state recurrence relations satisfied by the different orders of the physical quantities. This is done by translating the structure equations to infinity, hence obtaining a non-linear algebraic system of equations for the asymptotic coefficients. Then the aforementioned hierarchy helps us to identify levels where the equations become linear for the quantities belonging to that level. Moreover, the structure of that hierarchy leads to identifying the radiation fields as the necessary initial data to run an induction argument. \\ Then we further specialise to the case when \(\Xi_{ij}=0\) and \(A(F)_i=0\), that is, when no radiation is escaping to infinity (Definition \ref{non-radiating}). Then by running the hierarchy-induction procedure we prove in Corollary \ref{cor:non-radiating} that \(\partial_u\) is a Killing symmetry to all orders at infinity. \\ We remark here that we will be using generically the word \emph{quantities} to refer to the components of either the orthonormalisation matrix, the connection coefficients or the curvature components. \\ Before stating the main results we set some notation. Recall that \(\overset{(n)}{f}\) denotes the best \(r^{-n}\)-approximation of \(f\) (see discussion preceding definition \ref{def:AF}). The symbol \(\{\phi_1, \ldots, \phi_n\}\) will denote any expression involving the functions \(\phi_1, \ldots, \phi_n\). The symbol \(\lfloor n \rfloor \) denotes an expression involving the connection coefficients up to order \(n\) and the orthonormalisation matrix up to order \(n-1\). That is, \] \lfloor n \rfloor = \{ \overset{(0)}{\omega_{\mu \nu \lambda}}, \ldots, \overset{(n)}{\omega_{\mu \nu \lambda}}, \overset{(0)}{{h_\mu}^{\nu}}, \ldots \overset{(n-1)}{{h_\mu}^{\nu}} \} , \[ with the convention that \(\lfloor 0 \rfloor =0\). The symbol \(Q(\phi,\ldots; \varphi, \ldots)\) stands for a quadratic expression containing terms of the form \(\phi \varphi\). We remind the reader that asymptotic flatness is to be understood as stated in definition \ref{def:AF}. \\ \begin{prop} \label{prop:recurrence relations} Let \((M, g, F)\) be an asymptotically flat solution of the Einstein-Maxwell equations. Then the asymptotic quantities satisfy the following recurrence relations along the out-going null hypersurfaces \(C_u^+\) for any \(n\in \mathbb{N}\), \begin{subequations} \label{recursion0} \begin{align} \overset{(n+1)}{\alpha_{ij}} &= (n-1) \overset{(n)}{\chi_{ij}}-2 \overset{(n)}{\alpha(F)_k} \overset{(1)}{\alpha(F)^k} \eta_{ij} + \lfloor n-1 \rfloor, \\ \overset{(n)}{{h_i}^j} &= \overset{(n)}{{\chi_i}^k} \overset{(1)}{{h_k}^j} + \lfloor n-1 \rfloor, \end{align} \end{subequations} \begin{subequations} \label{recursion1} \begin{align} \overset{(n+1)}{\beta_i}, \overset{(n)}{\rho(F)}, \overset{(n)}{\sigma(F)}, \overset{(n)}{\omega_{jji}}, \overset{(n)}{\zeta_i}, \overset{(n)}{f^i} &= \{ \overset{(n)}{\chi_{ij}}, \overset{(n)}{\alpha(F)_i}, \lfloor n-1 \rfloor \} , \quad n> 3 \end{align} \end{subequations} \begin{subequations} \label{recursion2} \begin{align} \overset{(n+1)}{\rho}, \overset{(n+1)}{\sigma}, \overset{(n)}{\underline{\alpha}(F)}, \overset{(n)}{\underline{\omega}}, \overset{(n)}{\omega_{123}}, \overset{(n)}{\underline{\chi}_{ij}}, \overset{(n)}{\lambda_i}, \overset{(n-1)}{f^0} &= \{ \overset{(n)}{\chi_{ij}}, \overset{(n)}{\alpha(F)_i}, \lfloor n-1 \rfloor \}, \quad n> 2. \end{align} \end{subequations} Moreover, \begin{subequations} \label{recursion3} \begin{align} 2\partial_u \overset{(n+1)}{\chi_{ij}} &=-n \overset{(n)}{\underline{\chi}_{ij}} + \{ \overset{(n)}{\chi_{ij}}, \overset{(n)}{\alpha(F)_i}, \lfloor n-1 \rfloor \}, \\ 2\partial_u \overset{(n+1)}{\alpha(F)_i} &= \{ \overset{(n)}{\alpha(F)_i}, \overset{(n)}{\rho(F)}, \overset{(n)}{\sigma(F)}, \lfloor n \rfloor \}, \end{align} \end{subequations} \begin{subequations} \label{recursion4} \begin{align} \overset{(n+1)}{\underline{\beta}_i} &= \partial_u \overset{(n+1)}{\zeta_i} + \lfloor n \rfloor, \\ \overset{(n+1)}{\underline{\alpha}_{ij}} &= \partial_u \overset{(n+1)}{\underline{\chi}_{ij}} + \lfloor n \rfloor. \end{align} \end{subequations} \end{prop} This proposition can be interpreted as saying that the asymptotic quantities can be computed recursively starting from the radiation fields. The cases \(n=1,2,3\) are special. Roughly speaking \(n=1\) corresponds to the choice of gauges at infinity. On the other hand, for \(n=2\) we have to specify a mass aspect function as data at the \((\rho, \sigma)\)-level; also an EM-charge aspect function at the \(\beta_i\)-level. Finally, for \(n=3\) at the \(\beta_i\)-level an angular momentum aspect vector is required. It is worth remarking that these functions are not freely specifiable but they obey constraint/evolution equations given by the Einstein and Maxwell equations of signature \(-1\) and \(-2\). We deal with these cases during the proof of Proposition \ref{prop:recurrence relations}. Here we state the main relations found there in: \begin{lem} \label{lem:evolution equations} Under the same hypothesis of Proposition \ref{prop:recurrence relations} the following relations hold, \begin{subequations} \label{mass and angular momentum} \begin{align} \partial_u \tr \overset{(2)}{\underline{\chi}} &= Q(\overset{(1)}{\underline{\chi}}; \overset{(1)}{\underline{\chi}}) + Q(\overset{(1)}{\underline{\alpha}(F)}; \overset{(1)}{\underline{\alpha}(F)}), \\ \overset{(3)}{\rho}, \overset{(3)}{\sigma}, \overset{(2)}{\underline{\alpha}(F)}, \overset{(2)}{\underline{\omega}}, \overset{(2)}{\omega_{123}}, \overset{(2)}{\hat{\underline{\chi}}_{ij}}, \overset{(2)}{\lambda_i}, \overset{(1)}{f^0} &= \{ \overset{(2)}{\chi_{ij}}, \overset{(2)}{\alpha(F)_i}, \tr \overset{(2)}{\underline{\chi}}, \lfloor 1 \rfloor \} , \\ \overset{(3)}{\omega}_{jji} &= \overset{(1)}{h^{-1}} \cdot ( \breve{e}_k \overset{(3)}{{h_i}^j} + \lfloor 2 \rfloor ), \\ \partial_u\overset{(3)}{\zeta_i} &=-\partial_u\overset{(3)}{\omega_{jji}} + Q(\overset{(1)}{{ \underline{\alpha}(F)}}; \overset{(2)}{\rho(F)}, \overset{(2)}{\sigma(F)}) + \lfloor 2 \rfloor , \\ 2 \partial_u\overset{(2)}{\rho(F)} &= \breve{\cancel{\divergence}} \, \underline{A}(F), \\ 2 \partial_u\overset{(2)}{\sigma(F)} &= \breve{\cancel{\curl}} \, \underline{A}(F). \end{align} \end{subequations} \end{lem} Here \(\breve{e}_k := \overset{(1)}{{h_k}^j} \partial_j\) is the standard orthonormal basis on the round sphere and \(\breve{\cancel{\divergence}}\) and \(\breve{\cancel{\curl}}\) are the corresponding operators. We remark that equations (\ref{mass and angular momentum}a), (\ref{mass and angular momentum}d), (\ref{mass and angular momentum}e) and (\ref{mass and angular momentum}f) can be interpreted as evolution formulas for the mass, angular momentum and EM-charges. From the point of view of a characteristic initial value formulation, these are constraint equations at null infinity. \\ In particular, combining these relations with the non-radiating condition we get stationarity to all orders at infinity: \\ \begin{cor} \label{cor:non-radiating} Let \((M, g, F)\) be an asymptotically flat electrovacuum spacetime. Assume it is non-radiating, that is, \(\Xi_{ij}=0=\underline{A}(F)_{i}\). Then all the asymptotic quantities are \(u\)-independent. \\ \end{cor} Along the same lines we can prove a related result, namely, that all the asymptotic quantities can be found recursively in terms of the radiation fields and the ``pole moments'' of the initial Cauchy data. In other words, the radiation fields determine completely the dynamics of the spacetime at infinity. \\ \begin{cor} Let \((M, g, F)\) be an asymptotically flat electrovacuum spacetime. Then all the asymptotic quantities depend solely on the radiation fields \(\Xi_{ij}\), \(\underline{A}(F)_{i}\), the initial mass and angular momentum and the pole moments of \(\chi_{ij}\) and \(\alpha(F)_i\) (The latter are the limits as \(u\rightarrow-\infty\) of the asymptotic components \(\overset{(n)}{\chi_{ij}}\) and \(\overset{(n)}{\alpha(F)_i}\), \(n \in\mathbb{N}\)). \end{cor} Before proving these results we present the structure equations below. We have grouped them in levels representing the hierarchy found first by BMS, ,. Note that (\ref{hierarchy0}b), (\ref{hierarchy1}e), (\ref{hierarchy1}f), (\ref{hierarchy2}i) and (\ref{hierarchy2}j) are frame equations while (\ref{hierarchy1}g), (\ref{hierarchy2}k) are 1st Bianchi identities. All of these are satisfied on any Lorentzian manifold. In contrast, the Einstein equations appear implicitly when we regard \(S_{\mu\nu}\) as a quadratic expression in \(F_{\mu \nu}\). Finally, (\ref{hierarchy1}c), (\ref{hierarchy1}d), (\ref{hierarchy2}g), (\ref{hierarchy2}h), (\ref{hierarchy3}c), (\ref{hierarchy3}d) are the Maxwell equations. \begin{enumerate} [a)] \item The \(\alpha_{ij}\) or \(\Psi_0\) or signature \(2\) level. This includes the 2nd fundamental form \(\chi_{ij}\) and \({h_i}^j\): \begin{subequations} \label{hierarchy0} \begin{align} \alpha_{ij} + \frac12 \eta_{ij} S_{00} &=-e_0(\chi_{ij})-\chi_{ik} {\chi^k}_j, \\ e_0({h_i}^j)& =-{\chi_i}^k {h_k}^j. \end{align} \end{subequations} \item The \(\beta_i\) or \(\Psi_1\) or signature \(1\) level. This includes the torsion \(\zeta_i\) and the coefficients of the induced connection on \(S_{s,u}\), that is, \(\omega_{iij}\), and \(f^i\): \begin{subequations} \label{hierarchy1} \begin{align} 2\beta_i + S_{0i} &=-2 e_0(\zeta_i)-4 \zeta^k \chi_{ki} , \\ \beta_i-\frac12 S_{0i} &= e_0(\omega_{jji})-\zeta_j\chi_{ii} + \chi_{jj}\zeta_i + {\chi_j}^k\omega_{kji}, \quad i\neq j, \\ e_0(\rho(F)) &=-\cancel{\divergence} \, \alpha(F)-\tr \chi \rho(F)-\zeta_i\alpha(F)^i, \\ e_0(\sigma(F)) &=-\cancel{\curl} \, \alpha(F)-\tr \chi \sigma(F) + \varepsilon^{ik}\zeta_i\alpha(F)_k, \\ e_i({h_j}^k)-e_j({h_i}^k) &= ({\omega_i}^n{}_j-{\omega_j}^n{}_i) {h_{n}}^k, \\ e_0(f^i) &=-\zeta^k {h_k}^i, \\ e_1(\omega_{jji})-e_j(\omega_{1ji}) &=-e_j(\underline{\chi}_{ji}) + e_i(\underline{\chi}_{jj}) \nonumber \\ & \qquad-(\omega\wedge\omega)_{ji1j} + (\omega \wedge \omega)_{1jji}, \quad i\neq j. \end{align} \end{subequations} with no summation on repeated \(j\)'s. \item The \((\rho, \sigma)\) or \(\Psi_2\) or signature \(0\) level. This includes \(\underline{\omega}\), \(\omega_{123}\), \(\underline{\chi}_{ij}\), \(\lambda_i\) and \(f^0\): \begin{subequations} \label{hierarchy2} \begin{align} -\rho + \frac12(S_{22} + S_{33}) &= e_2(\omega_{323})-e_3(\omega_{223}) + \frac12 \chi_{22} \underline{\chi}_{33} + \frac12 \chi_{33} \underline{\chi}_{22} \nonumber \\ & \qquad-\chi_{23} \underline{\chi}_{23}-(\omega_{232})^2-(\omega_{323})^2, \\ -\sigma &=-e_2(\zeta_3) + e_3(\zeta_2) + \omega_{323}\zeta_3-\omega_{232}\zeta_2 \nonumber \\ & \qquad -{\chi_2}^k \underline{\chi}_{3k} + {\chi_3}^k \underline{\chi}_{2k} , \\ \rho +\frac12 S_{01} &= e_0(\underline{\omega})-3\zeta^k\zeta_k, \\ -2 \sigma &= e_0(\omega_{123}) + 4 \zeta^k \omega_{k23} , \\ -4 \rho-S_{jj} + \frac12 S_{01} &=-e_0(\underline{\chi}_{jj})-2 e_j(\zeta_j) + 2 \zeta^k \omega_{jkj}-{\chi_j}^k\underline{\chi}_{kj}-2\zeta_j \zeta_j, \\ -\sigma-S_{23} &=-e_0(\underline{\chi}_{23})-2 e_2(\zeta_3) + 2 \zeta^k\omega_{2k3}-2 \zeta_2\zeta_3-{\chi_2}^k\underline{\chi}_{k3}, \\ e_0(\underline{\alpha}(F)_i) &=-e_i(\rho(F)) + {\varepsilon_i}^je_j(\sigma(F))-(F\wedge \omega)_{i} , \\ e_1(\alpha(F)_i) &= e_i(\rho(F)) + {\varepsilon_i}^k e_k(\sigma(F))-({}^* F\wedge \omega)_{i} , \\ e_0(f^0) &= \underline{\omega}, \\ e_i(f^0) &= \lambda_i , \\ e_1(\chi_{ij}) -e_i(\zeta_j) &= e_0(\underline{\chi}_{ij})-e_j(\zeta_i)-(\omega \wedge \omega)_{1i0j} + (\omega \wedge \omega)_{0j1i}. \end{align} \end{subequations} With no summation on repeated \(j\)'s. \item The \(\underline{\beta}_i\) or \(\Psi_3\) or signature \(-1\) level: \begin{subequations} \label{hierarchy3} \begin{align} 2 \underline{\beta}_i-S_{1i} &= e_0(\lambda_i) +2 e_1(\zeta_i)-(\omega\wedge \omega)_{01i1} \\ -\underline{\beta}_i-\frac12 S_{1i} &= e_1(\omega_{jji})-e_j(\omega_{1ji})-(\omega \wedge \omega)_{1jji}. \\ e_1(\rho(F)) & = \cancel{\divergence} \, \underline{\alpha}(F) + \tr \underline{\chi} \rho(F) + \zeta_i\alpha(F)^i, \\ e_1(\sigma(F)) & =-\cancel{\curl} \, \underline{\alpha}(F)-\tr \underline{\chi} \sigma(F) + \varepsilon^{ik}\zeta_i\alpha(F)_k, \end{align} \end{subequations} With no summation on repeated \(j\)'s. \item Finally, the \(\underline{\alpha}_{ij}\) or \(\Psi_4\) or signature \(-2\) level: \begin{align} \label{hierarchy4} \underline{\alpha}_{ij} + \frac12 \eta_{ij}S_{11} =-e_i(\lambda_i)-e_1(\underline{\chi}_{ij})-(\omega\wedge \omega)_{i1j1}. \end{align} \end{enumerate} Now we proceed to prove the results stated above. It is convenient to think of the structure-Einstein-Maxwell equations as a 1st order system schematically of the form: \begin{align*} \nabla h + \omega \wedge h &= 0, \\ \nabla \omega + \omega \wedge \omega &= C + \eta \wedge S, \\ S &= F^2, \\ \nabla F + \omega \wedge F &= 0. \end{align*} The procedure will be analogous to the one described in the Introduction for a wave equation. \\ \textbf{Proof of Proposition \ref{prop:recurrence relations}.} Firstly, we state the 1st order values of the relevant coefficients that we will be using to obtain the recurrence relations and also in the induction argument. These are a consequence of the CK analysis and the gauge conditions. \begin{align*} \overset{(0)}{f^0} &= 1 , \qquad f^1 \equiv 2 , \qquad \overset{(1)}{{h_i}^j} = \mathrm{diag}(1, \frac{1}{\sin \theta}), \\ \overset{(1)}{\underline{\chi}_{ij}} & = \Xi_{ij} + \eta_{ij}, \qquad \overset{(1)}{\chi_{ij}} = \eta_{ij}. \end{align*} Recall the Einstein equations: \] S_{\mu \nu } = 2 F_{\mu \sigma} {F^{\sigma}}_{\nu}-\frac12 \eta_{\mu \nu} F_{\alpha \beta} F^{\alpha \beta}. \[ They are quadratic in \(F_{\mu \nu}\), in particular then \(\overset{(n+1)}{S_{\mu \nu}} = \{ \overset{(1)}{F_{\mu \nu}}, \ldots, \overset{(n)}{F_{\mu \nu}} \}\). A more detailed null decomposition is given by \begin{align*} S_{00} &= 2 \alpha(F)_i \alpha(F)^i = \mathcal{O} (r^{-5}), \\ S_{01} &= \rho(F)^2 + \sigma(F)^2 = \mathcal{O} (r^{-4}), \\ S_{ij} &= Q (\rho(F), \sigma(F); \rho(F), \sigma(F)) + Q(\alpha(F); \underline{\alpha}(F)) = \mathcal{O}(r^{-\frac72}) , \\ S_{0i} &= Q (\alpha(F)_i; \rho(F), \sigma(F)) = \mathcal{O}(r^{-\frac92}), \\ S_{1i} &= Q (\underline{\alpha}(F)_i; \rho(F), \sigma(F)) = \mathcal{O}(r^{-3}) \\ S_{11} &= 2 \underline{\alpha}(F)_i \underline{\alpha}(F)^i = \mathcal{O}(r^{-2}) , \end{align*} where we have included the peeling behaviour of the Faraday tensor, (\ref{null Faraday2})-(\ref{null Faraday3}). Now we proceed to derive the recurrence relations. \begin{enumerate} [a)] \item It can be seen that the \((n+1)\)-order of the \(\alpha\)-level equations together with \] \overset{(n+1)}{S_{00}} = \overset{(n)}{\alpha(F)_i} \overset{(1)}{\alpha(F)^i} + \{ \overset{(1)} \alpha(F)_i, \ldots, \overset{(n-1)} \alpha(F)_i \} \[ imply the recurrence relations (\ref{recursion0}) for \(\overset{(n+1)}{\alpha_{ij}}\) and \(\overset{(n)}{{h_i}^j}\). \item Now we look at the \((n+1)\)-order of the \(\beta\)-level equations. From this level onwards the equations become more intricate. The important thing to remember is that we only need to keep track of the coefficients accompanying the variables that we want to find at a given order and level. \\ We start with the \((n+1)\)-order of the frame equation: \]e_i({h_j}^k)-e_j({h_i}^k) = ({\omega_i}^n{}_j-{\omega_j}^n{}_i) {h_{n}}^k. \[ Note that the left-hand-side does not contain terms with \(\overset{(n+1)}{{h_j}^k}\) since \(e_i = {h_i}^j\partial_j\) and \({h_i}^j\) decays. Writing down the equations explicitly for \(k=2,3\) we get the system \begin{eqnarray*} \left( \begin{array}{cc} \overset{(1)}{{h_2}^2} &-\overset{(1)}{{h_2}^3}\\ -\overset{(1)}{{h_2}^2} & \overset{(1)}{{h_2}^3} \end{array}\right) \left( \begin{array}{c} \overset{(n)}{\omega_{223}} \\ \overset{(n)}{\omega_{332}} \end{array}\right) = \breve{e}_k \overset{(n)}{{h_i}^j} + \lfloor n-1 \rfloor. \end{eqnarray*} Here we recall that \(\breve{e}_k := \overset{(1)}{{h_k}^j} \partial_j\) is the standard orthonormal basis on the unit round sphere. This system tells us that \(\overset{(n)}{\omega_{jji}} = \{ \breve{e}_k \overset{(n)}{{h_i}^j}, \lfloor n-1 \rfloor \} = \{ \breve{e}_k \overset{(n)}{\chi_{ij}}, \lfloor n-1 \rfloor \}\), where we have used the recurrence relation for \(\overset{(n)}{{h_i}^j}\). In particular for \(n=3\) we get equation (\ref{mass and angular momentum}c).\\ We are left with the variables \(\beta_i\), \(\rho(F)\), \(\sigma(F)\) and \(\zeta_i\). The \((n+1)\)-order of equations (\ref{hierarchy1}a)-(\ref{hierarchy1}d) give the following linear system valid for \(n\geq 3\), \begin{eqnarray*} \left( \begin{array}{cccc} 2 & 0 & 0 &-(2n-4) \\ 1 & 0 & 0 &-1 \\ 0 &-(n-2) & 0 & 0 \\ 0 & 0 &-(n-2) & 0 \end{array}\right) \left( \begin{array}{c} \overset{(n+1)}{\beta_i} \\ \overset{(n)}{\rho(F)} \\ \overset{(n)}{\sigma(F)} \\ \overset{(n)}{\zeta_i} \end{array}\right) = \{ \overset{(n)}{\chi_{ij}},\overset{(n)}{\alpha(F)_i}, \lfloor n-1 \rfloor \}, \end{eqnarray*} where the zeros on the first two rows correspond to the fact that \(\overset{(1)}{\alpha(F)} = 0 = \overset{(2)}{\alpha(F)} \) and the quadratic character of \(S_{0i}\) stated before. Similarly the zeros on the last column come also from \(\overset{(1)}{\alpha(F)} = 0 = \overset{(2)}{\alpha(F)}\). The above system can be solved (note that the degenerate cases correspond to \(n =2,3\)) to obtain the desired recurrence relations. \\ It is worth noticing that \(\overset{(1)}{\zeta_i} = 0\), while \(\overset{(1)}{\omega_{jji}}\) corresponds to the connection coefficients of the standard round sphere.\\ \\ \textbf{Remark.} The case \(n=2,3\) are special. For \(n=2\) the Maxwell equations degenerate and we have to specify \(\overset{(2)}{\rho(F)}\) and \(\overset{(2)}{\sigma(F)}\) as initial data, these can be regarded as electromagnetic charges. For \(n=3\), the degeneracy is telling us that we have to prescribe \(\overset{(3)}{\zeta_i}\) or \(\overset{(4)}{\beta_i}\) as initial data. However, these data are not freely specifiable, in the sense that they satisfy evolutions equations (\ref{mass and angular momentum}). We will find the same situation on the \((\rho,\sigma)\)-level where the corresponding initial data can be regarded as a mass aspect function. \item This is the 0-signature case; a signature count gives us that \(\lambda_i\) does not appear on the \((\omega \wedge \omega)\)-term. Thus, the \((n+1)\)-order of equations (\ref{hierarchy2}a)-(\ref{hierarchy2}g) give us 9 equations for 9 variables, namely, \(\overset{(n+1)}{\rho}\), \(\overset{(n+1)}{\sigma}\), \(\overset{(n)}{\alpha(F)_i}\), \(\overset{(n)}{\underline{\omega}}\), \(\overset{(n)}{\omega_{123}}\) and \(\overset{(n)}{\underline{\chi}_{ij}}\). The linear system can be solved provided it is non-degenerate, which can be checked by direct computation for \(n>2\). This gives the desired recurrence relation. Here we state the linear system obtained by considering the \((n+1)\)-order of equations (\ref{hierarchy2}a)-(\ref{hierarchy2}g): \begin{eqnarray*} \left( \begin{array}{cccccccc} 1 & 0 & \frac12 & \frac12 & 0 & 0 & 0 & * \\ 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 \\ 1 & 0 & 0 & 0 & 0 & n & 0 & * \\ 0 & 1 & 0 & 0 & 0 & 0 &-n & 0\\ 1 & 0 & n-1 & 0 & 0 & 0 & 0 & * \\ 1 & 0 & 0 & n-1 & 0 & 0 & 0 & *\\ 0 & 1 & 0 & 0 & 0 & 0 & n-1 & *\\ n-1 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\ \end{array}\right) \left( \begin{array}{c} \overset{(n+1)}{\rho} \\ \overset{(n+1)}{\sigma} \\ \overset{(n)}{\underline{\chi}_{22}} \\ \overset{(n)}{\underline{\chi}_{33}} \\ \overset{(n)}{\underline{\chi}_{23}} \\ \overset{(n)}{\underline{\omega}} \\ \overset{(n)}{\underline{\omega}_{123}} \\ \overset{(n)}{\underline{\alpha}(F)_i} \\ \end{array}\right) = \textrm{Known data}, \end{eqnarray*} where `Known data' can be described more precisely as a term of the form \] \{ \overset{(n)}{\chi_{ij}}, \overset{(n)}{\zeta_i},\overset{(n)}{\omega_{jji}}, \overset{(n)}{\alpha(F)_i}, \overset{(n)}{\rho(F)},\overset{(n)}{\sigma(F)}, \lfloor n-1 \rfloor \}.\[ Once again the case \(n=2\) is special. The 1st, 5th and 6th rows become linearly dependent. Due to this degeneracy we have to specify \(\tr \overset{(2)}{\underline{\chi}}\) along null infinity, this corresponds to the Bondi mass (up to gauge terms). Finally we remark that the Einstein equation at the \(\underline{\alpha}\)-level provides us with an evolution equation, (\ref{mass and angular momentum}a), for \(\tr \overset{(2)}{\chi}\), which corresponds to the Bondi mass formula (\ref{Bondi mass formula}). \item Finally, equations (\ref{recursion3}) are obtained by taking the \((n+1)\)-order of equations (\ref{hierarchy2}k) and (\ref{hierarchy2}h), respectively. \item Now we focus on Lemma \ref{lem:evolution equations}. The evolution equations for the EM-charges, (\ref{mass and angular momentum}e) and (\ref{mass and angular momentum}f), follow from Maxwell's equations (\ref{hierarchy3}c) and (\ref{hierarchy3}d), respectively. \item As explained before we can obtain an evolution equation for \(\overset{(3)}{\zeta_i}\) by considering the system \((\ref{hierarchy3}a)\) and \((\ref{hierarchy3}b)\). The equation obtained after getting rid of the Weyl terms is of the form: \] \partial_u\overset{(3)}{\zeta}_i =-\partial_u \overset{(3)}{\omega_{jji}} + Q(\overset{(1)}{\underline{\chi}}; \overset{(2)}{\zeta}) + Q(\overset{(1)}{\underline{\alpha}(F)};\overset{(2)}{\rho(F)}, \overset{(2)}{\sigma(F)}) + \lfloor 2 \rfloor. \[ This proves the identity (\ref{mass and angular momentum}d). \item Again a signature analysis as above helps us to find the structure of the \(\underline{\alpha}\)-equation. We have that to 2nd order the trace of (\ref{hierarchy4}) reads \] \partial_u \tr \overset{(2)}{\underline{\chi}} = Q(\overset{(1)}{\underline{\chi}}; \overset{(1)}{\underline{\chi}}) + Q(\overset{(1)}{\underline{\alpha}(F)}; \overset{(1)}{\underline{\alpha}(F)}), \[ where we have used \(\overset{(1)}{\zeta_i}=0\). The remaining system for \(\overset{(3)}{\rho}\), \(\overset{(3)}{\sigma}\), etc, is non-degenerate and can be solved. \\ This finishes the proof of Proposition \ref{prop:recurrence relations} and Lemma \ref{lem:evolution equations}. \(_{\blacksquare}\)\\ \end{enumerate} \textbf{Proof of Corollary \ref{cor:non-radiating}.} All the hard work was done in the proof of Proposition \ref{prop:recurrence relations}. Now we have to check the lower order cases. \\ Firstly, we show that the 1st order asymptotic quantities are all \(u-\)independent. Recall that \(\overset{(1)}{{h_i}^j}=\mathrm{diag(1, 1/\sin\theta)}\) and \(\overset{(1)}{\chi_{ij}}=\eta_{ij}\). Moreover, \(\overset{(1)}{\underline{\alpha}(F)_i}=0\) and \(\overset{(1)}{\underline{\chi}_{ij}}=\eta_{ij}\) by the non-radiating condition. We also know that \(\overset{(n)}{\alpha_{ij}}= \overset{(n)}{\beta_i}=0\) for \(n=0,1,2,3\); \(\overset{(n)}{\alpha(F)_i} = \overset{(n)}{\rho}= \overset{(n)}{\sigma}=0\) for \(n=0,1,2\); and \(\overset{(n)}{\rho(F)}= \overset{(n)}{\sigma(F)}=0\) for \(n=0,1\).\\ Now we focus on \(\overset{(1)}{\zeta_i}\) and look at equations (\ref{hierarchy3}), we can get rid of the Weyl term by subtracting them. To 1st order we find that \] \partial_u \overset{(1)}{\zeta_i} = \{ \overset{(1)}{S_{1i}} \}, \[ which vanishes since \(S_{1i} = \mathcal{O}(r^{-3})\) due to Einstein-Maxwell equations. Hence \(\overset{(1)}{\zeta_i}\) is also \(u\)-independent. The frame equation \(e_0(f^i) =-\zeta^k {h_k}^i\) to 2nd order gives us that \(\overset{(1)}{f^i} = \overset{(1)}{\zeta^k} \overset{(1)}{{h_i}^i}\) and thus it is also \(u\)-independent. This finishes the \(\beta\)-level.\\ Moving on to the \((\rho,\sigma)\)-level: Recall that \(\overset{(1)}{\underline{\chi}_{ij}} = \Xi_{ij} + \eta_{ij}\). By assumption \(\Xi_{ij} =0\), so \(\overset{(1)}{\underline{\chi}_{ij}}\) is \(u\)-independent. Moreover, \(\overset{(2)}{\rho} =\overset{(2)}{\sigma} = \overset{(2)}{S_{01}} =0\), then the structure equations (\ref{hierarchy2}c) and (\ref{hierarchy2}d) to 2nd order give us that \(\overset{(1)}{\underline{\omega}}\) and \(\overset{(1)}{\omega_{123}}\) depends on previous data, so they must be \(u\)-independent. Finally, to 1st order, the frame equations \( e_i(f^0) = \lambda_i\) tell us that \(\overset{(1)}{\lambda_i}=0\), since \(\overset{(0)}{f^0}=1\). Therefore all the 1st order quantities at this level are \(u\)-independent.\\ We will see that we can repeat the argument now for asymptotic quantities of order 2. Firstly, the \(u\)-derivative of the recurrence relation (\ref{recursion3}) for \(n=1\) tell us that \begin{align*} \partial_u^2\overset{(2)}{\chi_{ij}} &= \partial_u^2\overset{(2)}{\alpha(F)_i}, \\ \partial_u^2\overset{(2)}{\alpha(F)_i} &= 0, \end{align*} where we have used that the 1st order quantities are \(u\)-independent. Now we make use of the hypothesis that \(\overset{(3)}{\alpha_{ij}}\) and \(\overset{(2)}{\alpha(F)_i}\) must remain finite\footnote{At this point is not really necessary to use this hypothesis since we already know that \(\overset{(2)}{\alpha(F)_i}=0\). We have phrased it this way for the sake of generality in the induction argument.} as \(u \rightarrow-\infty\) to rule out linear growth of \(\overset{(2)}{\chi_{ij}}\) and \(\overset{(2)}{\alpha(F)_i}\). Hence, \(\overset{(2)}{\chi_{ij}}\) and \( \overset{(2)}{\alpha(F)_i}\) are \(u\)-independent. \\ Now, at the \(\beta\)-level we find that the equation (\ref{hierarchy1}a) degenerates at this order; the \(\overset{(2)}{\zeta_i}\)-terms cancel. Also \(\overset{(3)}{S_{0i}}=0\), this together with equation (\ref{hierarchy1}b) gives a relation of the form: \]\overset{(2)}{\zeta_i} = \overset{(2)}{\omega_{jji}}. \[ On the other hand we can use the frame equation (\ref{hierarchy1}e) as before to find \(\overset{(2)}{\omega_{jji}}\) in terms of \(\overset{(2)}{{h_i}^j}\) and 1st order terms. This proves that \(\overset{(2)}{\zeta_i}\) and \(\overset{(2)}{\omega_{jji}} \) are \(u\)-independent. \\ As forecast, the Maxwell equations (\ref{hierarchy1}c) and (\ref{hierarchy1}d) also degenerate. This suggests that we need to provide \(\overset{(2)}{\rho(F)}\) and \(\overset{(2)}{\sigma(F)}\) as initial data. However, equations (\ref{mass and angular momentum}e) and (\ref{mass and angular momentum}f) impose evolution equations for these. Therefore they are \(u\)-independent in the non-radiating case. This finishes the \(\beta\)-level at this order.\\ Next, as seen in part c) of the proof of Proposition \ref{prop:recurrence relations}, the equations at the \((\rho, \sigma)\)-level also degenerate. Quick inspection and the fact that \(\overset{(3)}{S_{01}}=\overset{(3)}{S_{ij}}=0\) lead to identifying \(\overset{(2)}{\tr \underline{\chi}}\) as initial data. Then all the remaining quantities can be computed in terms of this. \\ Moreover, the evolution equation for \(\overset{(2)}{\tr \underline{\chi}}\) implies that it is \(u\)-independent in the non-radiating case. Therefore all the quantities at this level are also \(u\)-independent. \\ Thus we have established \(u\)-independence of all the quantities to 1st and 2nd order. In general, we can prove that the 1st and 2nd order quantities depend solely on the radiation fields and three aspect functions: mass and EM-charges. \\ It is clear now how to repeat the procedure inductively for \(n>2\): Assume that \(\overset{(n)}{\chi_{ij}}\), \(\overset{(n)}{\alpha(F)_i}\) and \(\lfloor n-1 \rfloor\) are \(u\)-independent then run the hierarchy-recursion argument to conclude that \(\lfloor n \rfloor\) is \(u\)-independent. \\ At this point we encounter one difficulty when analysing the \(\beta\)-level to order 3. It degenerates, but we already know that we have to prescribe an angular momentum aspect vector \(\overset{(3)}{\zeta_i}\) which comes with an evolution equation (\ref{mass and angular momentum}d). We now that we can compute \(\overset{(3)}{\omega_{jji}}\) from previous data, equation (\ref{mass and angular momentum}c), hence it is \(u\)-independent in the non-radiating case. Therefore the evolution equation for \(\overset{(3)}{\zeta_i}\) implies that it is also \(u\)-independent and the induction works through. \\ Finally, the evolution equations for \(\overset{(n+1)}{\chi_{ij}}\) and \(\overset{(n+1)}{\alpha(F)}\), (\ref{recursion3}a) and (\ref{recursion3}b), ensure that they will also be \(u\)-independent (provided they have finite limits as \(u \rightarrow-\infty\)) and the inductive step is proved. \(_{\blacksquare}\) \section{Unique continuation from infinity } \setcounter{footnote}{1} In the previous section we showed that all the asymptotic quantities depend only on \(\chi_{ij}\) along the out-going null hypersurface \(C_0^+\) and the radiating fields \(\Xi_{ij}\) and \(\underline{A}(F)_i\). In particular, if \(\Xi_{ij}\) and \(\underline{A}(F)_i\) vanish then all the quantities are \(u\)-independent; that is, \(T=\partial_u\) is a Killing field to all orders at infinity. The goal now is to extend this symmetry into the spacetime. We do this in Proposition \ref{prop:unique continuation2}, which in conjuction with Corollary \ref{cor:non-radiating}, allows us to prove our main result, Theorem \ref{thm1}.\\ We start here by presenting the necessary techniques to extend the time-like symmetry into a neighbourhood of infinity. The motivation comes from successful applications of the so-called Carleman estimates to prove uniqueness of solutions in the context of hyperbolic equations with pseudo-convex boundary conditions. For example, Ionescu-Klainerman used this approach to prove local unique extension of Killing vector fields across a pseudo-convex hypersurface in. Also, in, Alexakis-Ionescu-Klainerman showed uniqueness of smooth stationary black holes for small perturbations of Kerr by proving the unique continuation property for the Simon-Mars tensor. Here we present the main results when the boundary condition imposed is that of asymptotic flatness. Alexakis-Schlue-Shao showed in that linear waves satisfy the unique continuation from infinity property provided they decay faster than any polynomial. The main technical tool is the new Carleman estimates they derived in the context of asymptotically flat spacetimes, we include them here as Theorem \ref{thm:Carleman}.\\ In the second part of this section we revise, and extend to the non-vacuum case, the tensorial equations satisfied by the deformation tensors \(\mathcal{L}_Tg\), \(\mathcal{L}_TC\) and \(\mathcal{L}_T F\) obtained by. These take the form of transport equations for \(\mathcal{L}_T g\) and its first derivatives along the outgoing null direction and wave equations for \(\mathcal{L}_TC\) and \(\mathcal{L}_T F\). The latter are a consequence of the usual wave equations satisfied by \(C\) and \(F\), which in turn are implied by the Bianchi and Maxwell equations.\\ Then, in the third section we use Cartesian coordinates to cast the equations in a suitable form (with ``fast decaying coefficients'') such that the Carleman estimates from Theorem \ref{thm:Carleman} can be applied. Finally, a standard argument is used to bound the weighted \(L^2\)-norms of the deformation tensors and hence conclude its vanishing in a neighbourhood of infinity. Special care has to be taken regarding the wave equation for \(\mathcal{L}_T F\) whose coefficients do not decay fast enough. To deal with this problem, we ``borrow'' some decay from the coupling coefficient appearing in the wave equation for \(\mathcal{L}_T C\). \subsection{Carleman estimates} \label{section:Carleman estimates} In, Alexakis and Schlue proved Theorem \ref{thm1} for a vacuum spacetime. In order to generalise the proof to include a Maxwell field we need to adapt their argument at the level of Carleman estimates. These are inequalities that can be thought of as a priori estimates for functions decaying faster than any polynomial at infinity. In conjunction with a wave equation satisfied by the function, this method can be used to prove the vanishing of the function in a neighbourhood of infinity. \\ We present here the Carleman estimates obtained by Alexakis, Schlue and Shao in for spacetimes of positive mass. To do so we introduce some notation first. \\ The class of spacetimes considered in are incomplete 4-manifolds of the form \((-\infty,0) \times (0, \infty) \times \mathbb{S}^2\) with coordinates \((u,v, y^i)\) equipped with a metric given by: \begin{align} \label{metric AS} g & = g_{uu}du^2-4 K du dv + g_{vv} dv^2 + \sum_{i.j=2}^3 \gamma_{ij} dy^i dy^j \\ \nonumber & \qquad \sum_{i=2}^3 (g_{iu} dy^i du + g_{iv} dy^i dv), \end{align} where \(m\) and \(r\) are smooth positive functions. In addition Alexakis-Schlue-Shao require the following bounds to hold: \begin{itemize} \item The metric components satisfy \begin{align*} K & = 1-\frac{2m}{r} ,\qquad & g_{uu}, g_{vv} &= \mathcal{O}_1(r^{-3}), \\ g_{iu}, g_{iv} &= \mathcal{O}_1(r^{-1}) , \qquad & \gamma_{ij} &= \breve{\gamma}_{ij} + \mathcal{O}_1(r^{-1}). \end{align*} \item The function \(m\) is uniformly bounded away from \(0\), \(m\geq m_{min} >0\), and for some \(\eta>0\), the differential of \(m\) satisfies \begin{align*} |\partial_i m| = \mathcal{O}(r(-uv)^{-\eta}), \qquad |\partial_um|, |\partial_v m| = \mathcal{O}(r^{-2}). \end{align*} \item The function \(r\) is bounded on a level set of \(v-u\), that is, there exist constants \(C_1\), \(C_2\) such that \begin{align*} v(p)-u(p) = C_1 \quad \textrm{implies} \quad 1 \ll |r(p)| < C_2. \end{align*} The differential of \(r\) satisfies the following estimate \begin{align*} \left( 1+ \frac{2m}{r}\right) dr = (1+\mathcal{O}(r^{-2})) dv-(1+\mathcal{O}(r^{-2}))du + \sum_{i=2}^3 \mathcal{O}(r^{-1})dy^i. \end{align*} Note that these two conditions imply that \(r\) and \(v-u\) are comparable, that is, there exists a constant \(C\) such that \(r \leq C |v-u|\). \item Finally they also require \begin{align*} | \square_g \left( \frac{m}{r}\right)| = \mathcal{O}((-uv)^{-1-\eta}). \end{align*} \end{itemize} \textbf{Remark.} It is not clear that an asymptotically flat spacetime in the sense of definition \ref{def:AF} satisfies the above condition when cast in double-null coordinates. While this may not be true in general, it holds for non-radiating spacetimes. We prove this in Lemma \ref{lem:doublenull}. \\ Now, an important ingredient in the deduction of the Carleman estimates is a \emph{pseudo-convex} function. Geometrically, in the Lorentzian context, these are functions whose level sets are convex with respect to null geodesics; that is, any null geodesic tangent to a level set locally remains on one side of that level set. This is equivalent to the following quantitative condition,: A function \(f\) on \((\mathcal{M},g)\) is pseudo-convex if there exists a function \(h\) such that \[ \pi := h g-\nabla^2 f\] is positive definite when restricted to the tangent space of the level sets of \(f\). Alexakis-Schlue-Shao proved that, with respect to the above coordinates, the function \[f:= \frac{1}{(-u)v},\] is pseudo-convex. Therefore we expect the unique continuation property to hold across its level sets. One technicality arises in that this pseudo-convexity degenerates towards infinity. To cope with this degeneracy, which takes the form of vanishing/blowing up weights towards infinity, Alexakis-Schlue-Shao rely on a reparametrisation of \(f\), \[F(f) := \log f-f^{2\delta} ,\] for some \(\delta>0\) to be chosen later. Then they are able to conclude the unique continuation from infinity property in a neighbourhood of the form \[\mathcal{D}_{\omega} := \{ (u,v,y^i): 0 < f(u,v) < \omega \} .\] In order to understand the geometry of pseudo-convex timelike hypersurfaces we start by analysing them in Minkoswki spacetime. Consider double null coordinates \((u,v, \theta^2, \theta^3)\). The pseudo-convex timelike hypersurfaces considered by Alexakis-Schlue-Shao are given by the positive level sets of the function \(f_{\epsilon} = \frac{1}{(-u + \epsilon)(v + \epsilon)}\), \(\epsilon>0\). The \(\epsilon\)-perturbation is necessary to accomplish the pseudo-convexity condition in the absence of a mass. See Figure [\[fig:pseudo-convex\]](#fig:pseudo-convex){reference-type="ref" reference="fig:pseudo-convex"}.\ On positive-mass spacetimes the situation is qualitatively different. Consider for example Schwarzschild spacetime in double null coordinates, \((u,v,\theta^2, \theta^3)\), recall that these are defined by \[u = \frac{t-r_* }{2}, \qquad v = \frac{t+r_* }{2},\] where \[r_*(r) = \int_{r_0}^r\left( 1-\frac{2M}{s}\right)^{-1}ds, \qquad r_0>2M.\] Due to the presence of a mass now the function \(f = \frac{1}{(-u )v}\) is pseudo-convex.\ We are now in position to state the main technical tool of this section. For convenience we introduce the weight function \(\mathcal{W}\) and associated weighted norms. For any \(\lambda>0\) and domain \(\mathcal{D}=\mathcal{D}_{\omega}\), \(\omega>0\) \[\mathcal{W} = \textrm{e}^{-\lambda F} f^{\frac12}, \quad \| \cdot \|_{\mathcal{W}} = \| \mathcal{W} \cdot \|_2, \quad \| \phi \|^2_2 = \int_{\mathcal{D}} |\phi|^2 d\mu_g.\] **Sketch of proof.** Consider \(\varphi = \textrm{e}^{-\lambda F(f)} \phi\). The idea is to obtain an energy estimate for \(\varphi\), but here we wish for the bulk terms of the integral to be positive and for the boundary terms to vanish. That is, consider the modified energy current \[J_{\beta}^{w}[\varphi] = Q_{\alpha \beta}[\varphi] \nabla^{\alpha} f + \left( \frac12 \partial_{\beta} w + \frac12 \lambda^2 (\nabla^{\alpha}f)(\nabla_{\alpha}f)(F')^2 \partial_{\beta} f \right) \varphi^2-\frac12 w\partial_{\beta}(\varphi^2),\] where \(Q_{\alpha \beta}[\varphi]\) is the standard energy-momentum tensor for \(\square \varphi=0\).\ The function \(w\) is to be chosen appropriately so that \(\divergence J_{\beta}^{w}\) produces the tensor \(\pi\), hence capturing the pseudo-convexity of \(f\). Specifically, the choice \[w = h-\frac12 \square f-\frac12,\] produces positive bulk terms that are quadratic in \(\partial_X \varphi\) for \(X\) tangent to the level sets of \(f\); in order to obtain positivity in the normal direction one relies on the choice of reparametrisation \(F(f)\).\ The above procedure ultimately results in an inequality of the form \[\begin{aligned} \int_{\mathcal{D}_{\omega}} \mathcal{W}_L |\mathcal{P} \varphi|^2 \geq & \, C \lambda \int_{\mathcal{D}_{\omega}} \left( \mathcal{W}_N |\nabla_N \varphi|^2 +\mathcal{W}_T |\nabla_T \varphi|^2 +\mathcal{W}_T \sum |\nabla_{e_i} \varphi|^2 \right) \\ & C \lambda^3 \int_{\mathcal{D}_{\omega}} \mathcal{W}_0 |\varphi|^2 + \int_{\mathcal{D}_{\omega}} \mathcal{E}, \end{aligned}\] where \(\mathcal{P} (\varphi) = \textrm{e}^{-\lambda F(f)}\square(\textrm{e}^{\lambda F(f)} \varphi)\) and \(\mathcal{W}_L\), \(\mathcal{W}_N\), \(\mathcal{W}_T\), \(\mathcal{W}_0\) are positive weights. The only non-positive term is the error \(\mathcal{E}\) which must be absorbed into the remaining positive terms. \(_{\blacksquare}\)\ Alexakis-Schlue-Shao used the above Carleman estimate to extend the vanishing of a function \(\phi\) into a neighbourhood of infinity,. The proof is based on standard arguments using cut-off functions to cope with the "inner decay" necessary for Theorem [\[thm:Carleman\]](#thm:Carleman){reference-type="ref" reference="thm:Carleman"} to hold. The wave equation is used to substitute \(\square \phi\) by lower order derivatives which can be absorbed on the left-hand side of ([\[Carleman\]](#Carleman){reference-type="ref" reference="Carleman"}) due to the decaying conditions. Finally the result follows by taking \(\lambda\) to infinity. We omit a detailed proof since we will follow the same program in the proof of Proposition [\[prop:unique continuation2\]](#prop:unique continuation2){reference-type="ref" reference="prop:unique continuation2"}.\ We will also need Carleman estimates for transport equations involving \(\mathcal{L}_T g\) and its first derivatives. These are covered by the following lemma proved in. To conclude this section we present a technical result necessary to ensure that we can change from time-space coordinates in the asymptotically flat sense of definition [\[def:AF\]](#def:AF){reference-type="ref" reference="def:AF"} to the double-null coordinates required for the Carleman estimates. **Proof.** We will prove that the asymptotic expansion ([\[Kerr-Newman\]](#Kerr-Newman){reference-type="ref" reference="Kerr-Newman"}) holds through out the development when written in \(T\)-adapted coordinates. Then a transformation to 'comoving coordinates', as in, shows that \((\mathcal{D}, g)\) is of the desired form.\ First we need to find estimates for \(T\) at \(\Sigma\) with respect to the given coordinate system. Let us write, \[T = \tau \partial_t + \rho \partial_r + \theta^i \partial_{\vartheta^i}.\] Recall that \(T = \partial_u = \frac{L + \underline{L}}{2} + \sum_{\mu=0}^3 \mathcal{O}(r^{-1})e_\mu\).\ We have abused notation by using the same symbol, \(r\), for the area function of Section [2](#section:tetrad and coordinates){reference-type="ref" reference="section:tetrad and coordinates"} and the coordinate function in ([\[Kerr-Newman\]](#Kerr-Newman){reference-type="ref" reference="Kerr-Newman"}). However, due to the asymptotic expansion ([\[Kerr-Newman\]](#Kerr-Newman){reference-type="ref" reference="Kerr-Newman"}) assumed on \(\Sigma\) we have that they agree up to \(\mathcal{O}(1)\), hence the symbol \(\mathcal{O}(r^{-k})\) is unambiguous. Moreover, the cone \(C^-_{d^*}\) was chosen so that intersects \(\Sigma\) on \(r=d^*\), this guarantee that \(L\) and \(\underline{L}\) agree to leading order with the in-going and out-going null directions defined by the level sets of \(r\). Therefore, given the expansion ([\[Kerr-Newman\]](#Kerr-Newman){reference-type="ref" reference="Kerr-Newman"}), \(T\) must agree with \(\partial_t\) to leading order, that is, \(\tau \rightarrow 1\) and \(\rho, \theta^i \rightarrow 0\) as \(r \rightarrow \infty\) along \(\Sigma\).\ Now we use the hypothesis that \(T\) is Killing to order 4 to find more refined estimates for \(\tau, \rho, \theta^i\) at \(\Sigma\). Indeed, we have, \[\begin{aligned} (\mathcal{L}_T g)_{\mu \nu} & = (\partial_{\mu} \tau) g_{t \nu} + (\partial_{\nu} \tau) g_{\mu t} + \tau(\partial_t g)_{\mu \nu} \\ & \quad + (\partial_{\mu} \rho) g_{r \nu} + (\partial_{\nu} \rho) g_{\mu r} + \rho(\partial_r g)_{\mu \nu} \\ & \quad + (\partial_{\mu} \theta^i) g_{\vartheta^i \nu} + (\partial_{\nu} \theta^i) g_{\mu \vartheta^i} + \theta^i(\partial_{\vartheta^i} g)_{\mu \nu} . \end{aligned}\] Also, ([\[Kerr-Newman\]](#Kerr-Newman){reference-type="ref" reference="Kerr-Newman"}) gives control on \(\partial g\) along \(\Sigma\): \[\begin{aligned} (\partial_t g)_{rr} &= \mathcal{O}^{\infty}_1(r^{-4}), &\quad (\partial_t g)_{rt} &= \mathcal{O}^{\infty}_1(r^{-4}), &\quad (\partial_t g)_{r \vartheta^i} &= \mathcal{O}^{\infty}_1(r^{-3}), \\ (\partial_r g)_{rr} &=-\frac{2M}{r^2} + \mathcal{O}^{\infty}_1(r^{-4}), &\quad (\partial_r g)_{rt} &= \mathcal{O}^{\infty}_1(r^{-5}), &\quad (\partial_r g)_{r \vartheta^i} &= \mathcal{O}^{\infty}_1(r^{-4}), \\ (\partial_{\vartheta^i} g)_{rr} &= \mathcal{O}^{\infty}_1(r^{-2}), &\quad (\partial_{\vartheta^i} g)_{rt} &= \mathcal{O}^{\infty}_1(r^{-4}), &\quad (\partial_{\vartheta^i} g)_{r \vartheta^j} &= \mathcal{O}^{\infty}_1(r^{-3}). \end{aligned}\] Therefore, to leading order, the \((r,{\vartheta^i})\), \((r,r)\) and \((r,t)\) components of \(\mathcal{L}_T g\) imply the following differential estimates: \[\begin{aligned} \partial_r \theta^i + \frac{1}{r^2}(\partial_{\vartheta^i} \rho) + \frac{1}{r^2}(\partial_r \tau)\mathcal{O}(r^{-1}) &= \mathcal{O}(r^{-4}), \\ 2(\partial_r \rho)\left( 1 +\frac{2M}{r} \right)-\left( \frac{2M}{r^2} + \mathcal{O}(r^{-2}) \right) \rho + \theta^i \mathcal{O}(r^{-2}) & = \mathcal{O}(r^{-4}), \\ -2(\partial_r \tau)\left( 1-\frac{2M}{r} \right) + (\partial_r \theta)\mathcal{O}(r^{-1}) + \theta^1\mathcal{O}(r^{-3}) & = \mathcal{O}(r^{-4}). \end{aligned}\] Now, the a priori estimates, \(\tau \rightarrow 1\) and \(\rho, \theta^i \rightarrow 0\), together with the above bounds show that \(\partial_r\theta^i, \partial_{r} \rho, \partial_r \tau = \mathcal{O}(r^{-2})\). Thus, \[\begin{aligned} \tau-1, \rho, \theta^i = \mathcal{O} (r^{-3}). \end{aligned}\] Finally, we define \(T\)-adapted coordinates by setting \(\bar{t}=0\) on \(\Sigma\) and \(T(\bar{t})=1\), while coordinates \((r, y^2, y^3)\) are defined to be constant along \(T\)-lines with \((r, y^2, y^3)= (r, \vartheta^2, \vartheta^3)\) on \(\Sigma\). The goal of the previous analysis was to achieve the correct asymptotic behaviour of the metric in this new coordinates, indeed, it is easy to check that \[\begin{aligned} g = &-\left( 1-\frac{2M}{r} + \frac{e^2}{r^2} + \mathcal{O}^{\infty}_2(r^{-3}) \right) d\bar{t}^2 + \mathcal{O}^{\infty}_2(r^{-4}) d\bar{t} dr \nonumber \\ & + \left( 1 + \frac{2M}{r}-\frac{a^2 + e^2}{r^2}\sin^2y^2 + \frac{2Ma^2}{r^3}\cos^2y^2 + \mathcal{O}^{\infty}_2(r^{-4}) \right) dr^2 \nonumber \\ & + \left( \breve{\gamma}_{ij} + \mathcal{O}^{\infty}_2(1) \right) dy^i dy^j + \sum_{i=2}^3 \left( \mathcal{O}^{\infty}_2(r^{-1})d\bar{t} dy^i + \mathcal{O}^{\infty}_2(r^{-3})dr dy^i \right) \end{aligned}\] In these coordinates, the metric components are easily controlled by the initial data and \(\mathcal{L}_T g\), indeed, \[\left| g_{\mu \nu} (\bar{t}, r, y^2, y^3)-g_{\mu \nu} (0,r, y^2, y^3) \right| \leq \int_0^{\bar{t}} | (\mathcal{L}_T g)_{\mu \nu} (t,r,y^2, y^3)| dt.\] Therefore on a domain of the form, \[\mathcal{D} = \{ (\bar{t}, r, y^2, y^3): r\geq R, |\bar{t}| \leq r + 2M\log r-(R+2M\log R) \} ,\] we have that \(|g_{\mu \nu} (\bar{t})-g_{\mu \nu}(0)| = \mathcal{O}(r^{-3})\).\ Then a transformation to comoving coordinates as in achieves the desired double-null form ([\[metric AS\]](#metric AS){reference-type="ref" reference="metric AS"}). \(_{\blacksquare}\) ## Proof of the main Theorem Now we proceed to the last part of the proof of Theorem [\[thm1\]](#thm1){reference-type="ref" reference="thm1"}. Namely, we show that the vector field \(T=\partial_u\) is indeed a symmetry of the gravitational and electromagnetic fields provided that it is a symmetry to all orders at infinity. More precisely we have the following: Note that conditions ([\[decay\]](#decay){reference-type="ref" reference="decay"}) corresponds precisely to the non-radiating hypothesis of Theorem [\[thm1\]](#thm1){reference-type="ref" reference="thm1"}; while conditions ([\[killing to all orders\]](#killing to all orders){reference-type="ref" reference="killing to all orders"}) are those deduced from Corollary [\[cor:non-radiating\]](#cor:non-radiating){reference-type="ref" reference="cor:non-radiating"}.\ The strategy to prove Proposition [\[prop:unique continuation2\]](#prop:unique continuation2){reference-type="ref" reference="prop:unique continuation2"} is analogous to the one used by Ionescu-Klainerman in and Alexakis-Schlue in. We start by deducing tensorial wave equations for \(\mathcal{L}_T C\) and \(\mathcal{L}_T F\). When written in Cartesian coordinates those equations fit into the framework provided by the Carleman estimates of Theorem [\[thm:Carleman\]](#thm:Carleman){reference-type="ref" reference="thm:Carleman"}. Once we get those differential estimates a standard argument to deduce unique continuation follows through.\ We emphasise here that the full system of equations, ([\[coupled system\]](#coupled system){reference-type="ref" reference="coupled system"}) below, does not fit into the Alexakis-Schlue argument since one of the coupling terms does not decay fast enough. We deal with this problem by borrowing decay from the well-behaved coefficients and by using different Carleman weights for each of \(\mathcal{L}_T C\) and \(\mathcal{L}_T F\). See equations ([\[CarlemanW3\]](#CarlemanW3){reference-type="ref" reference="CarlemanW3"}), ([\[CarlemanE\]](#CarlemanE){reference-type="ref" reference="CarlemanE"}) and ensuing discussion. ## Ionescu-Klainerman tensorial equations In this section we recall the equations satisfied by the deformation tensors deduced by Ionescu and Klainerman in. We remind the reader that their work assumes a vacuum spacetime, that is, the Riemann and Weyl tensors are equal, \(R_{abcd}=C_{abcd}\). This is in contrast with our current approach where the Riemann tensor is coupled to the Faraday tensor via the Einstein equations. Hence we need to be careful now to distinguish between \(R\) and \(C\).\ \ *Notation.* Through out this section we will denote schematically by \[U \cdot V\] any linear combination of the product of two tensors \(U\) and \(V\), and contractions thereof. For example the relation \[R_{abcd} = C_{abcd} + \frac{1}{n-1} \left( g_{ac} S_{bd}-g_{dc} S_{ba} + g_{db} S_{ca}-g_{ab} S_{cd} \right).\] will be abbreviated to \[R = C+ g \cdot S. \\ \] We proceed now to compute the Ionescu-Klainerman tensorial wave equations for \(\mathcal{L}_T C\) and \(\mathcal{L}_T F\). The idea is to use the wave equations satisfied by \(C\) and \(F\) and then commute the covariant and Lie derivatives with the help of Lemma [\[lem:commutator\]](#lem:commutator){reference-type="ref" reference="lem:commutator"}. The relevant equations are naturally coupled to \(\pi_{ab}:=\mathcal{L}_T g_{ab}\) and \(\nabla_a \pi_{bc}\), hence the necessity of finding (transport) equations for them. In order to get a closed system of equations we need to perform some algebraic tricks (related to the symmetries of the Weyl tensor) and work instead with auxiliary variables, \(B_{ab}\) and \(P_{abc}\); see Proposition [\[prop:tensorial equations\]](#prop:tensorial equations){reference-type="ref" reference="prop:tensorial equations"}. This section is entirely based on.\ We begin by noticing that \(\mathcal{L}_T C\) is not trace-less. Indeed, \[g^{ac} \mathcal{L}_T C_{abcd} = (-\mathcal{L}_T g^{ac})C_{abcd} = \pi^{ac}C_{abcd}.\] To remedy this we define the modified Lie derivative \[\begin{aligned} \hat{\mathcal{L}}_T C := \mathcal{L}_T C-B\odot C. \end{aligned}\] where \(B\) is a 2-covariant tensor and \((B\odot C)_{abcd} := {B_a}^eC_{ebcd} + {B_b}^eC_{aecd}+ {B_c}^eC_{abed}+ {B_d}^eC_{abce}\). If we take \[B = \frac12( \pi + \omega) ,\] with \(\omega\) any antisymmetric 2-form, a simple calculation leads to Next, we need to compute the commutator of the Lie and covariant derivatives. This is given by the following: **Proof.** We compute \[\begin{aligned} \mathcal{L}_T V_{a_1 \ldots a_n} &= T^c \nabla_c V_{a_1 \ldots a_n} + \sum_i (\nabla_{a_i}T^c)V_{a_1 \ldots c \ldots a_n} , \\ \nabla_b \mathcal{L}_T V_{a_1 \ldots a_n} &= (\nabla_b T^c ) \nabla_c V_{a_1 \ldots a_n} + T^c\nabla_b \nabla_c V_{a_1 \ldots a_n} \\ & \qquad + \sum_i (\nabla_b \nabla_{a_i}T^c)V_{a_1 \ldots c \ldots a_n} + (\nabla_{a_i}T^c) \nabla_b V_{a_1 \ldots c \ldots a_n}, \\ \mathcal{L}_T \nabla_b V_{a_1 \ldots a_n} &= T^c \nabla_c \nabla_b V_{a_1 \ldots a_n} + (\nabla_b T^c) \nabla_c V_{a_1 \ldots a_n} + \sum_i (\nabla_{a_i}T^c) \nabla_b V_{a_1 \ldots c \ldots a_n}. \end{aligned}\] Then \[\begin{aligned} V_{a_1 \ldots a_n} = \sum_i \left ( T^c {R_{cba_i}}^f V_{a_1 \ldots f \ldots a_n} + (\nabla_b \nabla_{a_i} T^c) V_{a_1 \ldots c \ldots a_n} \right). \end{aligned}\] And the results follows from the identity \[\begin{aligned} \nabla_b\nabla_a T_c = R_{cabd}T^d + \Pi_{abc}. \end{aligned}\] To prove this equality we just evaluate, commute derivatives and make use of the 1st Bianchi identity: \[\begin{aligned} R_{cabd}T^d + \Pi_{abc} &= R_{cabd}T^d + \frac12 \left( \nabla_a (\nabla_b T_c + \nabla_cT_b) + \nabla_b(\nabla_aT_c + \nabla_cT_a)-\nabla_c(\nabla_aT_b + \nabla_bT_a)\right), \\ &= R_{cabd}T^d + \frac12 \left( R_{acbd}T^d + R_{bcad}T^d + 2 \nabla_a\nabla_bT_c-R_{bacd}T^d \right), \\ &= R_{cabd}T^d + \frac12 \left( 2R_{acbd}T^d + 2 \nabla_a\nabla_bT_c \right), \\ &= \nabla_a\nabla_bT_c. \qquad {}_{\blacksquare} \end{aligned}\] Now we present the variables and equations which play the crucial role in the unique continuation analysis. The variables are minor modifications of \(\pi\), \(\nabla \pi\) and \(\mathcal{L}_T C\) which give us a closed system of equations. During the proof we will make use of the following identities: **Proof of Lemma.** We start by showing \(L^aL^b\pi_{ab}=0\). Indeed, \[\begin{aligned} L^aL^b\pi_{ab} &= L^aL^b (\nabla_a T_b + \nabla_b T_a ) = L^b T^a\nabla_a L_b + L^a T^b\nabla_b L_a \\ &= T^a \nabla_a (L_b L^b) = 0. \end{aligned}\] Now, by commuting derivatives we can find a transport equation for \(L^b\pi_{ab}\). \[\begin{aligned} L^c\nabla_c (L^b\pi_{ab} ) &= L^c \nabla_c (L^b (\nabla_a T_b + \nabla_b T_a ) ,\\ &= L^c L^b ( \nabla_a \nabla_c T_b + {R_{cab}}^e T_e) + L^c \nabla_c (T^b \nabla_b L_a) , \\ &= \nabla_a(L^cL^b) \nabla_c T_b + {R_{cab}}^e L^c L^b T_e) + {R_{cba}}^e L^cT^bL_e + T^c \nabla_c (L^b \nabla_b L_a),\\ &= (\nabla_aL^c) (L^b \pi_{cb}), \end{aligned}\] which is a homogeneous equation for \(L^b \pi_{ab}\). In our context, the choice of \(T\) implies that the deformation tensor \(\pi\) vanishes to first order at infinity. In particular \(L^b\pi_{ab}\) also vanishes to first order at infinity; this together with the above transport equation implies that \(L^b\pi_{ab}\equiv 0\) as desired.\ Next, a straightforward computation yields, \[\begin{aligned} 2 P_{abc}L^c &= L^c(\nabla_a \pi_{bc}-\nabla_b \pi_{ac}-\nabla_c \omega_{ab}) , \\ &=-\pi_{bc}\nabla_aL^c + \pi_{ca}\nabla_bL^c-\pi_{ac} \nabla_b L^c + \pi_{bc} \nabla_aL^c = 0, \end{aligned}\] where we have used \(L^c\pi_{ac}=0\) and the definition of \(\omega_{ab}\).\ Finally, for the last equality we deduce a transport equation for \(L^b\omega_{ab}\) using its definition, \[\begin{aligned} L^c \nabla_c(L^b\omega_{ab}) &= L^b L^c \nabla_c \omega_{ab}, \\ &= L^b (\pi_{ac} \nabla_bL^c-\pi_{bc} \nabla_aL^c) = 0, \end{aligned}\] Hence \(L^b\omega_{ab}\) is constant and we can choose initial conditions for \(\omega_{ab}\) such that \(L^b\omega_{ab}\) vanishes. \(_{\blacksquare}\)\ **Proof of Proposition [\[prop:tensorial equations\]](#prop:tensorial equations){reference-type="ref" reference="prop:tensorial equations"}.** For the transport equation for \(B_{ab}\) we compute: \[\begin{aligned} 2(L^c P_{cba}-B_{cb}\nabla_aL^c) &= L^c(\nabla_c\pi_{ba}-\nabla_b\pi_{ac}-\nabla_a\omega_{cb})-(\pi_{cb}+\omega_{cb})\nabla_aL^c \\ &= \nabla_L \pi_{ab} + \pi_{ac}\nabla_b L^c + \omega_{cb}\nabla_aL^c-(\pi_{cb}+\omega_{cb})\nabla_aL^c \\ &= 2 \nabla_L B_{ab}, \end{aligned}\] where we have used \(L^c\pi_{cb}=L^c\omega_{cb}=0\) and the transport equation defining \(\omega_{ab}\).\ Next we deduce the transport equation for \(P_{abc}\). Recall the following identity[^4] proved in for \(\tilde{P}_{abc} := \frac12 (\nabla_a\pi_{bc}-\nabla_b\pi_{ac})\), \[\nabla_d\tilde{P}_{abc}-\nabla_c\tilde{P}_{abd} = \mathcal{L}_T R_{abcd}-\frac12 {\pi_a}^f R_{fbcd}-\frac12 {\pi_b}^f R_{afcd}.\] Note that \(P_{abc} = \tilde{P}_{abc}-\frac12 \nabla_c \omega_{ab}\), hence \[\begin{aligned} L^d(\nabla_d P_{abc}-\nabla_c P_{abd}) &= L^d \left(\mathcal{L}_T R_{abcd}-\frac12 {\pi_a}^f R_{fbcd}-\frac12 {\pi_b}^f R_{afcd} + \frac12 (\nabla_c\nabla_d \omega_{ab}-\nabla_d\nabla_c \omega_{ab}) \right), \\ &= L^d \left(\mathcal{L}_T R_{abcd}-\frac12 {\pi_a}^f R_{fbcd}-\frac12 {\pi_b}^f R_{afcd}-\frac12 {\omega_a}^f R_{fbcd}-\frac12 {\omega_b}^f R_{afcd} \right), \\ &= L^d \left(\mathcal{L}_T R_{abcd}-{B_a}^f R_{fbcd}-{B_b}^f R_{afcd} \right), \end{aligned}\] the final result follows by noticing that \(L^d\nabla_c P_{abd} =-P_{abd}\nabla_cL^d\) since \(L^dP_{abd}=0\).\ Finally, we deduce the wave equation for \(W\) by commuting Lie and covariant derivatives. We are interested only in the general structure of the equations, in particular, in the coefficients accompanying \(W\) and \(\nabla W\). Hence we do not any longer keep track of the different contractions but just on the bilinear structure of products and the different terms involving our variables \(P\), \(B\) and \(W\).\ \ *Notation.* During the following computations we will substitute freely \(B\) instead of \(\pi\) since \(\pi_{ab} = B_{ab} + B_{ba}\).\ We start now with a divergence equation for \(W\). **Proof.** We will prove the schematic version: \[\divergence W = B \cdot \nabla C + P \cdot C .\] We have that \[\begin{aligned} \nabla_e W = \mathcal{L}_T (\nabla_e C) + \Pi_e \odot C-\nabla_e (B \odot C), \end{aligned}\] hence, \[\begin{aligned} \divergence W &= (\mathcal{L}_T g) \cdot \nabla C + \mathcal{L}_T \divergence C + (\Pi_e-\nabla_e B) \cdot C-B\cdot \nabla C, \\ &= B \cdot \nabla C + P \cdot C, \end{aligned}\] where we used \(\pi_{ab} = B_{ab} + B_{ba}\), \(\Pi_{aeb}-\nabla_e B_{ab} = P_{abe}\) and \(\divergence C = 0\). \(_\blacksquare\)\ We proceed similarly to obtain the wave equation for \(W\), \[\begin{aligned} \square W &= \nabla ^e( \mathcal{L}_T \nabla_e C + \Pi_e \odot C-\nabla_e(B \odot C) ) , \\ &= \mathcal{L}_T \square C + \Pi^e \cdot \nabla_e C + \nabla ^e(\Pi_e \odot C-\nabla_e(B \odot C)) , \\ &= \mathcal{L}_T \square C + \nabla B \cdot \nabla C + \nabla^e( \Pi_e-\nabla_e B) \odot C + B\cdot \square C , \\ &= \mathcal{L}_T \square C + \nabla B \cdot \nabla C + \nabla P \cdot C + B\cdot \square C, \end{aligned}\] where we have used once more \(\Pi_{aeb}-\nabla_e B_{ab} = P_{abe}\) and \(\Pi=\nabla \pi = \nabla B\) (schematically). This finishes the proof of Proposition [\[prop:tensorial equations\]](#prop:tensorial equations){reference-type="ref" reference="prop:tensorial equations"}. \(_{\blacksquare}\)\ Using the same argument as above we can prove the following general statement: This last lemma can be applied to the Faraday tensor. To conclude this section we state the full system of equations relevant for the Carleman estimates. We start by recalling the wave equations satisfied by the Weyl and Faraday tensors, as well as the Einstein equations in schematic form: \[\begin{aligned} \square C &= R \cdot C = (C + g\cdot S) \cdot C, \\ \square F &= R \cdot F = (C + g\cdot S) \cdot F, \\ S &= F \cdot F \end{aligned}\] Hence \[\begin{aligned} \mathcal{L}_T\square C & = (\mathcal{L}_T C + \pi \cdot F^2 + \mathcal{L}_T F \cdot F)\cdot C + R \cdot \mathcal{L}_T C, \\ & = (W + B\cdot C + B \cdot F^2 + E \cdot F + B \cdot F^2)\cdot C + R \cdot W + R\cdot B \cdot W. \end{aligned}\] Similarly, for the modified Lie derivative of the Faraday tensor we obtain \[\mathcal{L}_T\square E = (W + B\cdot C + B \cdot F^2 + E \cdot F + B \cdot F^2)\cdot F + R \cdot E + R\cdot B \cdot F.\] Therefore, we have proved: ## Estimates for the deformation tensors Now we use the Carleman estimates of Theorem [\[thm:Carleman\]](#thm:Carleman){reference-type="ref" reference="thm:Carleman"} applied to the wave equations for \(W\) and \(E\). We start by writing them in Cartesian coordinates. This type of coordinates are chosen due to its uniform decay in all directions. We also have to check that the fields \(W\) and \(E\) satisfy the vanishing condition ([\[vanishing condition\]](#vanishing condition){reference-type="ref" reference="vanishing condition"}). This latter condition is fulfilled on the "exterior" part thanks to the hypothesis of Proposition [\[prop:unique continuation2\]](#prop:unique continuation2){reference-type="ref" reference="prop:unique continuation2"}, that is, the decaying properties corresponds to the fields vanishing to all orders at infinity. However in order to cope with the "interior" decay also included in the vanishing condition ([\[vanishing condition\]](#vanishing condition){reference-type="ref" reference="vanishing condition"}) a cut-off function needs to be introduced. Finally, by absorbing the lower order terms into the terms already present in the Carleman estimates we can conclude the vanishing of \(W\), \(E\), \(B\) and \(P\) in a neighbourhood of spatial infinity.\ Firstly, we pass to Cartesian coordinates \((x^0, x^1, x^2, x^3)\) such that the metric takes the form \[g =-(dx^0)^2 + (dx^1)^2 + (dx^2)^2 + (dx^3)^2 + \sum_{\mu, \nu=0}^3 \mathcal{O}_2(r^{-1}) dx^{\mu}dx^{\nu} ,\] and \(\partial_{x^0}|_{x^1, x^i}\) coincides with \(\partial_u|_{s,\theta^i}\) as \(r\rightarrow \infty\), where \[r = \sqrt{(x^1)^2 + (x^2)^2 + (x^3)^2}.\] Then, in these coordinates the Christoffel symbols decay suitably fast. That is, \[\Gamma_{\mu \nu}^{\alpha} = \mathcal{O}_1(r^{-2}) \quad \textrm{and} \quad \partial_{\beta} \Gamma_{\mu \nu}^{\alpha} = \mathcal{O}(r^{-3}) .\] In what follows, for brevity, we will denote by \((V)\) the components of a tensor \(V\) with respect to these Cartesian coordinates. In particular, we have schematically \[\begin{aligned} (\nabla V) &= \partial (V) + \Gamma \cdot (V), \\ (\square V) &= \square (V) + \Gamma \cdot \partial (V) + \partial \Gamma \cdot (V). \end{aligned}\] Hence, the wave equations for the components of \(W\) and \(E\) are \[\begin{aligned} \square (W) & = [(R) + \partial \Gamma]\cdot (W) + \Gamma \cdot \partial (W) + [(F)\cdot (C)] \cdot (E) + [(R)\cdot (C) + \partial (C) \cdot \Gamma]\cdot (B) \\ & \quad + \partial (C) \cdot \partial (B) + (C) \cdot \partial (P) + (C)\cdot \Gamma \cdot (P), \\ \square (E) & = [(F) + \partial \Gamma] \cdot (W) + \Gamma \cdot \partial (E) + [(F)^2 + \partial\Gamma] \cdot (E) + [(R)\cdot (F) + \partial (F) \cdot \Gamma] \cdot (B) \\ & \quad + \partial (F) \cdot \partial (B) + (F) \cdot \partial (P) + (F) \cdot \Gamma \cdot (P), \end{aligned}\] where we have just kept the leading order terms multiplying \(W\), \(E\), etc. In view of the asymptotic behaviour assumed for the Weyl and Faraday tensors we have the following estimates \[\begin{aligned} \label{coupled system} \begin{aligned} \square (W) & = \mathcal{O}(r^{-3}) (W) + \mathcal{O}(r^{-2})\partial (W) + \mathcal{O}(r^{-5}) (E) + \mathcal{O}(r^{-6}) (B) \\ & \qquad + \mathcal{O}(r^{-4}) \partial (B) + \mathcal{O}(r^{-5}) (P) + \mathcal{O}(r^{-3}) \partial (P) , \\ \square (E) & = \mathcal{O}(r^{-2}) (W) + \mathcal{O}(r^{-3}) (E) + \mathcal{O}(r^{-2})\partial (E) + \mathcal{O}(r^{-5}) (B) \\ & \qquad + \mathcal{O}(r^{-3}) \partial (B) + \mathcal{O}(r^{-4}) (P) + \mathcal{O}(r^{-2}) \partial (P). \end{aligned} \end{aligned}\] While these are morally the reason for the unique continuation we still need to compensate for the fact that the coefficient accompanying \((W)\) in the second equation does not decay fast enough[^5]. However the coupling term in the first equation, \(\mathcal{O}(r^{-5}) (E)\), allows us to borrow some decay by modifying the Carleman weight.\ Before applying the Carleman estimates of Theorem [\[thm:Carleman\]](#thm:Carleman){reference-type="ref" reference="thm:Carleman"} we have to guarantee that all the quantities vanish to all orders in the sense of ([\[vanishing condition\]](#vanishing condition){reference-type="ref" reference="vanishing condition"}), \[\lim_{k\rightarrow \infty } \int_{\partial \mathcal{U}_k} r^N(\phi^2 + |\partial\phi|^2) = 0.\] We deal with the "interior" and "exterior" parts of the the boundary of \(\mathcal{U}_k\) differently.\ The "exterior" boundary approaches infinity as \(r\rightarrow \infty\) and so it captures the idea of \(\phi\) vanishing to all orders at infinity. By assumption the quantities \(\pi =\mathcal{L}_T g\), \(\mathcal{L}_T C\) and \(\mathcal{L}_T F\) vanish to all orders at infinity. Now we check the modified versions. Firstly, \((\omega)\) satisfies the transport equation \[\nabla_L (\omega) = \mathcal{O}(r^{-1})(\pi).\] with \((\omega)=0\) at infinity by construction. Thus, \((\omega)\) vanishes to all orders at infinity. It follows that \[\begin{aligned} (B) &= (\pi) + (\omega), \\ (P) &= \partial (\pi + \omega) + \Gamma \cdot (\pi + \omega), \\ (W) &= (\mathcal{L}_T C) + (R) \odot (B), \\ (E) &= (\mathcal{L}_T F) + (F) \odot (B), \end{aligned}\] also vanish to all orders at infinity.\ To deal with the "interior" part of \(\partial \mathcal{U}_k\) a cut-off function is used. This technique is standard for unique continuation problems.\ Let \(\chi\) be a cut-off function whose level sets coincide with those of \(F\circ f\), \[\chi =1 \, \textrm{on } \mathcal{D}_{\omega_0}, \qquad \chi=0 \, \textrm{on } \mathcal{M}\setminus \mathcal{D}_{\omega_1}, \qquad \omega_0<\omega_1<\omega.\] Then the functions \(\chi \cdot (W)\), \(\chi \cdot (E)\), etc., satisfy the vanishing condition ([\[vanishing condition\]](#vanishing condition){reference-type="ref" reference="vanishing condition"}). The price to pay is that we have introduced extra terms in the wave equations, however these are easy to treat since they are supported only in the cut-off region. Indeed, \[\begin{aligned} \square (\chi\cdot (W)) &= (\square \chi) \cdot (W) + (\partial \chi) \cdot \partial (W) + \chi \cdot \square (W), \\ &= (\square \chi) \cdot (W) + (\partial \chi) \cdot \partial (W) + (\partial \chi) \cdot \{(W), \partial(W), \ldots \} \nonumber \\ & \quad+ \{ \chi \cdot (W), \partial (\chi \cdot (W)), \ldots \}, \\ &= \nabla \chi \mathbf{M} + \{ \chi \cdot (W), \partial (\chi \cdot (W)), \ldots \}. \end{aligned}\] Hereafter we will use the symbol \(\nabla \chi \mathbf{M}\) to denote multiples of \((W)\), \((E)\), etc., which are only supported in the cut-off region \(\mathcal{D}_{\omega_1} \setminus \mathcal{D}_{\omega_0}\). Recall also that we have used the notation \(\{ (W), (E), \ldots \}\) to denote a function involving \((W)\), \((E)\), etc. Hence, after applying the Carleman estimates we can focus only on the terms supported on \(\mathcal{D}_{\omega_0}\).\ \ **Weighted Carleman estimates.**\ \ We are now ready to apply Theorem [\[thm:Carleman\]](#thm:Carleman){reference-type="ref" reference="thm:Carleman"} to the functions \(\chi \cdot (W)\), \(\chi \cdot (E)\), etc. To keep the notation simple and readable we will omit in the next argument the cut-off function and the parenthesis denoting Cartesian components. We follow the standard procedure to bound the \(L^2\)-norms of \(W\), \(B\), \(P\) and its first derivatives.\ The Carleman estimate for \(W\) combined with its wave equation reads \[\begin{aligned} \label{CarlemanW} \lambda^3 \| f^{\delta} W \|_{\mathcal{W}} + \lambda \| f^{-\frac12} \Psi^{\frac12} \partial W \|_{\mathcal{W}} & \lesssim \| f^{-1} \square W \|_{\mathcal{W}} , \nonumber \\ & \lesssim \| f^{-1} r^{-3} W \|_{\mathcal{W}} + \| f^{-1} r^{-2} \partial W \|_{\mathcal{W}} \nonumber \\ & \quad + \| f^{-1} r^{-6} B \|_{\mathcal{W}} + \| f^{-1} r^{-4} \partial B \|_{\mathcal{W}} \nonumber \\ & \quad + \| f^{-1} r^{-5} P \|_{\mathcal{W}} + \| f^{-1} r^{-3} \partial P \|_{\mathcal{W}} \nonumber \\ & \quad + \| f^{-1} r^{-5} E \|_{\mathcal{W}} + \| \nabla\chi \textbf{M} \|_{\mathcal{W}}. \end{aligned}\] The following estimates will be used throughout; they follow from the definition of \(f\) and \(\Psi\): \[\begin{aligned} \label{estimates f} f \gtrsim \frac{1}{r^2}, \qquad f\Psi \gtrsim \frac{1}{r^3}. \end{aligned}\] They imply \[\begin{aligned} f^{-\frac12} \Psi^{\frac12} & =f^{-1}(f\Psi)^{\frac12} \gtrsim f^{-1}r^{-\frac32}, \\ \label{estimates f2} f^{\delta} &\gtrsim \frac{1}{r^{2\delta}} > \frac{1}{r}, \qquad \textrm{for } 0<2\delta < 1. \end{aligned}\] These last inequalities tell us that the first two terms on the right hand side of ([\[CarlemanW\]](#CarlemanW){reference-type="ref" reference="CarlemanW"}) can be absorbed into the corresponding terms on the left hand side, since \(f^{-1}r^{-3} \lesssim f^{\delta}\) and \(f^{-1}r^{-2} \lesssim f^{-\frac12} \Psi^{\frac12}\). This will be the main trick during the proof.\ Next, we aim at controlling the \(B\) terms. The Carleman estimate from Lemma ([\[lem:CarlemanODE\]](#lem:CarlemanODE){reference-type="ref" reference="lem:CarlemanODE"}) together with the transport equation for \(B\) read \[\begin{aligned} \lambda \|\frac1r f^{-1} r^{-4} B \|_{\mathcal{W}} & \lesssim \| f^{-1} r^{-4} \nabla_L B \|_{\mathcal{W}} , \nonumber \\ & \lesssim \| f^{-1} r^{-4} P \|_{\mathcal{W}} + \| f^{-1} r^{-5} B \|_{\mathcal{W}} + \| \nabla\chi \textbf{M} \|_{\mathcal{W}}. \end{aligned}\] We add this inequality to ([\[CarlemanW\]](#CarlemanW){reference-type="ref" reference="CarlemanW"}) and observe that the terms \(\| f^{-1} r^{-5} B \|_{\mathcal{W}}\) and \(\| f^{-1} r^{-6} B \|_{\mathcal{W}}\) can be absorbed into the left hand side. We have thus obtained, \[\begin{aligned} \label{CarlemanW2} \lambda^3 \| f^{\delta} W \|_{\mathcal{W}} + \lambda \| f^{-\frac12} \Psi^{\frac12} \partial W \|_{\mathcal{W}} + \lambda\| f^{-1} r^{-5} B \|_{\mathcal{W}} & \lesssim \| f^{-1} r^{-5} E \|_{\mathcal{W}} + \| f^{-1} r^{-4} \partial B \|_{\mathcal{W}} \nonumber \\ & \quad + \| f^{-1} r^{-4} P \|_{\mathcal{W}} + \| f^{-1} r^{-3} \partial P \|_{\mathcal{W}} \nonumber \\ & \quad + \| \nabla\chi \textbf{M} \|_{\mathcal{W}}. \end{aligned}\] Now we proceed similarly to bound \(P\), \(\partial B\) and \(\partial P\). We have \[\begin{aligned} \lambda \|\frac1r & f^{-1} r^{-3} P \|_{\mathcal{W}} \lesssim \| f^{-1} r^{-3} \nabla_L P \|_{\mathcal{W}} , \nonumber \\ & \lesssim \| f^{-1} r^{-3} W \|_{\mathcal{W}} + \| f^{-1} r^{-6} B \|_{\mathcal{W}} + \| f^{-1} r^{-5} E \|_{\mathcal{W}} + \| f^{-1} r^{-4} P \|_{\mathcal{W}} + \|\nabla\chi \textbf{M} \|_{\mathcal{W}}, \\ \lambda \|\frac1r & f^{-1} r^{-3} \partial B \|_{\mathcal{W}} \lesssim \| f^{-1} r^{-3} \nabla_L \partial B \|_{\mathcal{W}} , \nonumber \\ & \lesssim \| f^{-1} r^{-5} B \|_{\mathcal{W}} + \| f^{-1} r^{-4} \partial B \|_{\mathcal{W}} + \| f^{-1} r^{-4} P \|_{\mathcal{W}} + \| f^{-1} r^{-3} \partial P \|_{\mathcal{W}} + \|\nabla\chi \textbf{M} \|_{\mathcal{W}}, \\ \lambda \|\frac1r &f^{-1} r^{-2} \partial P \|_{\mathcal{W}} \lesssim \| f^{-1} r^{-2} \nabla_L \partial P \|_{\mathcal{W}} , \nonumber \\ & \lesssim \| f^{-1} r^{-3} W \|_{\mathcal{W}} + \| f^{-1} r^{-2} \partial W \|_{\mathcal{W}} + \| f^{-1} r^{-6} B \|_{\mathcal{W}} + \| f^{-1} r^{-5} \partial B \|_{\mathcal{W}} \\ & \qquad + \| f^{-1} r^{-4} P \|_{\mathcal{W}} + \| f^{-1} r^{-5} E \|_{\mathcal{W}} + \| f^{-1} r^{-4} \partial E \|_{\mathcal{W}} + \|\nabla\chi \textbf{M} \|_{\mathcal{W}}. \end{aligned}\] We add these inequalities to ([\[CarlemanW2\]](#CarlemanW2){reference-type="ref" reference="CarlemanW2"}) and observe that the \(W\), \(B\), \(\partial B\), \(P\) and \(\partial P\) terms can be absorbed into the left hand side, thus obtaining, \[\begin{aligned} \label{CarlemanW3} \lambda^3 \| &f^{\delta} W \|_{\mathcal{W}} + \lambda \| f^{-\frac12} \Psi^{\frac12} \partial W \|_{\mathcal{W}} + \lambda\| f^{-1} r^{-5} B \|_{\mathcal{W}} + \lambda\| f^{-1} r^{-4} \partial B \|_{\mathcal{W}} \nonumber \\ & + \lambda\| f^{-1} r^{-4} P \|_{\mathcal{W}} + \lambda\| f^{-1} r^{-3} \partial P \|_{\mathcal{W}} \lesssim \| f^{-1} r^{-5} E \|_{\mathcal{W}} + \| \nabla\chi \textbf{M} \|_{\mathcal{W}}. \end{aligned}\] for sufficiently large \(\lambda\) and \(0<\delta<\frac12\). An analogous argument gives (note the different Carleman parameter), \[\begin{aligned} \label{CarlemanE} \lambda'^3 \| &f^{\delta} E \|_{\mathcal{W}'} + \lambda' \| f^{-\frac12} \Psi^{\frac12} \partial E \|_{\mathcal{W}'} + \lambda' \| f^{-1} r^{-5} B \|_{\mathcal{W}'} + \lambda' \| f^{-1} r^{-4} \partial B \|_{\mathcal{W}'} \nonumber \\ & + \lambda' \| f^{-1} r^{-3} P \|_{\mathcal{W}'} + \lambda' \| f^{-1} r^{-2} \partial P \|_{\mathcal{W}'} \lesssim \| f^{-1} r^{-2} W \|_{\mathcal{W}'} + \| \nabla\chi \textbf{M} \|_{\mathcal{W}'}. \end{aligned}\] We would like to add these last two inequalities and absorb the \(W\) term on the left hand side to obtain the desired bound. However, for \(\lambda =\lambda'\), this is not possible as \(f^{-1} r^{-2}=\mathcal{O}(1)\) does not decay fast enough. To remedy this, we make the observation that the norms depend on \(\lambda\) and by taking slightly different weights we can perform the procedure just described. More precisely, we want to find \(\lambda'\) such that \[\begin{aligned} \label{inequalities} \begin{aligned} (e^{-\lambda F}f^{\frac12})f^{-1} r^{-5} & \lesssim e^{-\lambda' F} f^{\frac12} f^{\delta}, \\ (e^{-\lambda' F}f^{\frac12})f^{-1} r^{-2} & \lesssim e^{-\lambda F} f^{\frac12} f^{\delta}. \end{aligned} \end{aligned}\] Indeed, we will show that the choice \(\lambda' := \lambda-\delta\) achieves the previous inequalities. Firstly, note that \[\begin{aligned} r^{-5} & \lesssim f^{1+2\delta}, \\ r^{-2} & \lesssim f, \end{aligned}\] these are a consequence of estimates ([\[estimates f\]](#estimates f){reference-type="ref" reference="estimates f"}) and ([\[estimates f2\]](#estimates f2){reference-type="ref" reference="estimates f2"}). They imply that \[\begin{aligned} f^{-1} r^{-5} & \lesssim e^{\delta F} f^{\delta}, \\ e^{\delta F} f^{-1} r^{-2} & \lesssim f^{\delta}, \end{aligned}\] since \(F=\mathcal{O}(\log f)\). Finally, it is easy to see that these last inequalities are equivalent to ([\[inequalities\]](#inequalities){reference-type="ref" reference="inequalities"}), with \(\lambda' := \lambda-\delta\).\ \ **Remark.** It is worth noticing that the previous argument did not make any special use of the power \(r^{-5}\) accompanying \(E\). The procedure will work for any \(r^{-q}\) with \(q>2\) by choosing \(\delta>0\) small enough.\ Now we are in position to close the argument. We add inequalities ([\[CarlemanW3\]](#CarlemanW3){reference-type="ref" reference="CarlemanW3"}) and ([\[CarlemanE\]](#CarlemanE){reference-type="ref" reference="CarlemanE"}), with \(\lambda'= \lambda-\delta\). Inequalities ([\[inequalities\]](#inequalities){reference-type="ref" reference="inequalities"}) ensure that the terms \(\| f^{-1} r^{-5} E \|_{\mathcal{W}}\) and \(\| f^{-1} r^{-2} W \|_{\mathcal{W}'}\) can be absorbed into the left hand side. Moreover, the Carleman weights \(e^{-\lambda F}\) and \(e^{-\lambda' F}\) are monotonic increasing functions, so we can substitute its minimum value on left hand side and its maximum value on the right hand side (since these terms are supported in the cut-off region). Thus, after dropping the weight factors from the inequality we obtain the desired \(L^2\)-bound, \[\begin{aligned} \lambda^3 \| f^{\frac12} f^{\delta} W \|_2 &+ \lambda \| \Psi^{\frac12} \partial W \|_2 + \lambda\| f^{\frac12} f^{-1} r^{-5} B \|_2 + \lambda\| f^{\frac12} f^{-1} r^{-4} \partial B \|_2 \\ & + \lambda\| f^{\frac12} f^{-1} r^{-4} P \|_2 + \lambda\| f^{\frac12} f^{-1} r^{-3} \partial P \|_2 \\ & + \lambda'^3 \| f^{\frac12} f^{\delta} E \|_2 + \lambda' \| \Psi^{\frac12} \partial E \|_2 + \lambda' \| f^{\frac12} f^{-1} r^{-5} B \|_2 + \lambda' \| f^{\frac12} f^{-1} r^{-4} \partial B \|_2 \\ & + \lambda' \| f^{\frac12} f^{-1} r^{-3} P \|_2 + \lambda' \| f^{\frac12} f^{-1} r^{-2} \partial P \|_2 \\ & \qquad \lesssim \| \nabla\chi \textbf{M} \|_2 + \| \nabla\chi \textbf{M} \|_2. \end{aligned}\] Finally, the left hand side can be interpreted as integrated over the smaller domain \(\mathcal{D}_{\omega_0}\) where \(\chi=1\) and by taking \(\lambda \rightarrow \infty\) we conclude that \(B \equiv 0\), \(P \equiv 0\), \(W\equiv 0\) and \(E\equiv 0\) on \(\mathcal{D}_{\omega_0}\). In particular \[\mathcal{L}_T g \equiv 0 \qquad \textrm{and} \qquad \mathcal{L}_T F \equiv 0 \qquad \textrm{on } \mathcal{D}_{\omega_0}.\] This finishes the proof of Proposition [\[prop:unique continuation2\]](#prop:unique continuation2){reference-type="ref" reference="prop:unique continuation2"} and Theorem [\[thm1\]](#thm1){reference-type="ref" reference="thm1"}. \(_{\blacksquare}\) [^1]: Alternatively, one could impose de Sitter or anti-de Sitter boundary conditions. [^2]: Such an expansion is morally equivalent to analyticity all the way to infinity; an assumption arguably incompatible with the wave equation. Here we assume it for the sake of simplicity. The role of the wave equation in this argument will be merely computational as opposed to evolutionary. An asymptotic expansion would work as well. [^3]: A nested family of subsets, with piece-wise smooth time-like boundaries, whose union is all of \(\mathcal{D}_{\omega}\). [^4]: Which is basically an antisymmetrised identity for \(\nabla_d \nabla_a (\nabla_{(b} T_{c)})\) necessary to cope with the symmetries of \(R_{abcd}\). The idea to prove it is to commute derivatives. [^5]: We need a power strictly greater than 2 in order to run the Alexakis-Schlue argument, see proof below where the corresponding term is controlled by choosing different \(\lambda\)-weights for each Carleman inequality.
{'timestamp': '2016-08-16T02:13:25', 'yymm': '1607', 'arxiv_id': '1607.04882', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04882'}
null
null
# Introduction For the first decade of microlens planet detections, beginning with OGLE-2003-BLG-235Lb, the great majority of detections required a combination of survey and followup data. This is a consequence of two effects. First, the survey coverage was generally too sparse to characterize the planetary anomalies in the detected events. Second, thanks to aggressive alert capability, pioneered by the Optical Gravitational Lensing Experiment (OGLE) Early Warning System (EWS, @ews1 [@ews2]), it became possible to organize intensive followup of planet-sensitive events--or even ongoing planetary anomalies --and so obtain sufficient time resolution to detect and characterize planets. However, as surveys have become more powerful over the past decade, they have become increasingly capable of detecting planets without followup observations. That is, making use of larger cameras, the surveys are able to monitor fairly wide areas at cadences of up to several times per hour. While still substantially lower than followup observations of the handful of events that were monitored by followup groups, this is still adequate to detect most planets (provided that the anomalies occur when the survey is observing). Very simple reasoning given below, which is supported by detailed simulations, leads one to expect that the transition from survey+followup to survey-only mode implies a corresponding transition from planets detected primarily in high-magnification events via central and resonant caustics to planets primarily detected in lower magnification events via planetary caustics. High-magnification events are intrinsically sensitive to planets because they probe the so-called "central caustic" that lies close to (or overlays) the position of the host. Planets that are separated from the hosts by substantially more (less) than the Einstein radius generate one (two) other caustics that are typically much larger than the central caustic and thus have a higher cross section for anomalous deviations from a point-lens light curve due to a random source trajectory. However, for high-magnification events, the source is by definition passing close to the host and hence close to or over the central caustic. For planet-host separations that are comparable to the Einstein radius, the two sets of caustics merge into a single (and larger) "resonant caustic", which is even more likely to generate anomalous deviations of a high-magnification event. For many years, the Microlensing Follow Up Network (\(\mu\)FUN) employed a strategy based on this high planet sensitivity of high-magnification events. They made detailed analyses of alerts of ongoing events from the OGLE and the Microlensing Observations in Astrophysics (MOA) teams to predict high-magnification events and then mobilized followup observations over the predicted peak. showed that \(\mu\)FUN was able to get substantial data over peak for about 50% of all identified events with maximum magnification \(A_{\rm max}>200\), but that its success rate dropped off dramatically at lower magnification, i.e., even for \(100<A_{\rm max}<200\). The reason for this drop off was fundamentally limited observing resources: there are twice as many events \(A_{\rm max}>100\) compared to \(A_{\rm max}>200\), and monitoring the full-width half-maximum requires twice as much observing time. Hence, observations grow quadratically with effective magnification cutoff. By contrast, because planetary caustics are typically much larger than central caustics, most planets detected in survey-only mode are expected to be from anomalies generated by the former, which occur primarily in garden-variety (rather than high-mag) events. For example, detected a large planetary caustic in OGLE-2012-BLG-0406 based purely upon OGLE data, while detected one in MOA-bin-1 based mostly on MOA data. In the latter case it would have been completely impossible to discover the planet by survey+followup mode because the "primary event" (due to the host) was so weak that it was never detected in the data. Nevertheless, there has been a steady stream of survey-only detections of planets in high-magnification events as well. The first of these was MOA-2007-BLG-192Lb, a magnification \(A_{\rm max}> 200\) event, which required a combination of MOA and OGLE data. The first planet detected by combining three surveys (MOA, OGLE, Wise), MOA-2011-BLG-322Lb, was also via a central caustic, although in this case the caustic was very large so that the magnification did not have to be extremely large \((A_{\rm max}\sim 20)\). Similarly, detected a large central caustic due to the large planet OGLE-2015-BLG-0954Lb despite modest peak magnification of the underlying event \(A_{\rm max}\sim 20\). This case was notable because high-cadence data from the Korea Microlensing Telescope Network (KMTNet) captured the caustic entrance despite the extremely short source self-crossing time, \(t_*=16\,\)min. There also exist 2 planets MOA-2008-BLG-379Lb and OGLE-2012-BLG-0724Lb that were detected by the OGLE+MOA surveys through the high-magnification channel. KMTNet is still in the process of testing its reduction pipeline. Motivated by the above experience, the KMTNet team focused its tests on high-magnification events identified as such on the OGLE web page. In addition to exposing the reduction algorithms to a wide range of brightnesses, this testing has the added advantage that there is a high probability to find planets. Here we report on the first planet found by these tests from among the first seven high-mag events that were examined: OGLE-2016-BLG-(0261,0353,0471,0528,0572,0596,0612). These events were chosen to have model point-lens magnifications \(A>20\) and modeled peak times \(2457439<t_0<2457492\). The lower limit was set by the beginning of the KMTNet observing season and the upper limit was the time of the last OGLE update when the seven events were selected. # Observations On 2016 April 8 UT 12:15 (HJD\(^\prime=\) HJD\(-2450000=7487.0\)), OGLE alerted the community to a new microlensing event OGLE-2016-BLG-0596 based on observations with the 1.4 deg\(^2\) camera on its 1.3m Warsaw Telescope at the Las Campanas Observatory in Chile using its EWS real-time event detection software. Most observations were in \(I\) band, but with some \(V\) band observations that are, in general, taken for source characterization. These \(V\)-band data are not used in the modeling. At equatorial coordinates \((17^{\rm h} 51^{\rm m} 12^{\rm s}\hskip-2pt.81,-30^\circ 50' 59''\hskip-2pt.4)\) and Galactic coordinates \((-1.01^\circ,-2.03^\circ)\), this event lies in OGLE field BLG534, with an observing cadence during the period of the anomaly of roughly 0.4 per hour[^1]. KMTNet employs three \(4.0\,\rm deg^2\) cameras mounted on 1.6m telescopes at CTIO/Chile, SAAO/South Africa, and SSO/Australia. In 2015 KMTNet had concentrated observations on 4 fields. However, in 2016, this strategy was radically modified to cover (12, 40, 80) \(\rm deg^2\) at cadences of \((4,\ \geq 1,\ \geq 0.4)\,\rm hr^{-1}\). For the three highest-cadence fields, KMTNet observations are alternately offset by about \(6^\prime\) in order to ensure coverage of events in gaps between chips. As a result, OGLE-2016-BLG-0596 lies in two slightly offset fields BLG01 and BLG41, which are each observed at a cadence 2 per hour[^2]. KMTNet observes primarily in \(I\) band, but 1/11 observations from CTIO and 1/21 observations from SAAO are in \(V\)-band. Reductions of the primary data were made using difference image analysis (DIA). However, due to issues discussed in Section 3, special variants of DIA were developed specifically for this event. See below. KMT CTIO \(V\) and \(I\) images were, in addition, reduced using DoPHOT, solely for the purpose of determining the source color. # Light Curve Variability ## Evidence of Variability The OGLE-2016-BLG-0596 light curve shows clear variability over the course of 6 seasons of OGLE-IV data prior to 2016. This variability is roughly consistent with being sinusoidal at period \(P=126.5\,\) days. See Figure [\[fig:one\]](#fig:one){reference-type="ref" reference="fig:one"}. While the semi-amplitude of the variability is only \(8.7\%\) of the baseline flux \((I_{\rm base}\sim 19)\), this semi-amplitude turns out to be roughly equal to the source flux derived from the model. ## Importance of Variability Assuming (as proves to be the case) that it is not the source itself that is variable, such low level variability cannot significantly impact characterization of the anomalous features of the lightcurve because they are at relatively high magnification and take place on much shorter timescales. However, if not properly accounted for, such variability can seriously impact the estimate of the source flux and, as a direct consequence of this, the Einstein timescale \[t_{\rm E} = {\theta_{\rm E}\over\mu}; \qquad \theta_{\rm E}^2\equiv \kappa M \pi_{\rm rel}; \qquad \kappa\equiv {4 G\over c^2\au}\simeq 8.14{{\rm mas}\over M_\odot}. \label{eqn:tedef}\] Here \(\theta_{\rm E}\) is the angular Einstein radius, \(\pi_{\rm rel} = \au(D_L^{-1}-D_S^{-1})\) is the lens-source relative parallax, and \(\mu\) is the lens-source relative proper motion in the Earth frame. Errors in these quantities would propagate into the estimates of the planet-star mass ratio \(q\), the Einstein radius \(\theta_{\rm E}\), and the proper motion \(\mu\), all of which are important for assessing the physical implications of the detection. The reason that \(t_{\rm E}\) is potentially impacted by unmodeled variability is that it is determined primarily from the wings of the light curve where the amplitude of the amplification of flux is comparable to that of the variability. Hence, it is important to track down the source of this variability and correct for it to the extent possible. See, e.g.,. ## Removal of Variability I: Variations of Neighbor In principle, tracking down such a low level of variability in such a crowded field could have been very difficult. However, in the present case, it turns out to be due to a star \(1.5^{\prime\prime}\) to the southeast, which is quite bright \(I\sim 14.5\) and shows variability with the same period and phase. Within the framework of standard DIA, it is natural that this variable should impact the microlensing light curve because the difference image contains residuals from the variable that overlap the point spread function (PSF) of the microlensed star. Hence, when the difference image is dot-multiplied by the PSF to estimate the flux, it includes a contribution from the residual flux of the variable. It is straightforward to simultaneously fit for two (or \(n\)) variables with possibly overlapping PSFs. After constructing difference images in the standard way, one simply generalizes the normal procedure by calculating the \(n\) flux-difference values \[F_i = \sum_{j=1}^n c_{ij} d_j; \qquad c\equiv b^{-1}, \label{eqn:multifit1}\] where \[b_{ij} = \sum_k {P_{i,k}P_{j,k}\over \sigma_k^2}; \qquad d_i = \sum_k {P_{i,k}f_k\over \sigma_k^2}, \label{eqn:multifit2}\] \(P_{i,k}\) is the (unit normalized: \(\sum_k P_{i,k}=1\)) amplitude of the \(i\)th PSF in the \(k\)th pixel and \((f_k,\sigma_k)\) are the value and error of the difference flux in the \(k\)th pixel. We use a variant of this formalism to reduce the OGLE data with \(n=3\) stars, including the microlensed source, the bright variable, and one other very nearby (but non-variable) blended star. The result is shown in the second panel of Figure [\[fig:one\]](#fig:one){reference-type="ref" reference="fig:one"}. First note that the fluxes have decreased by about 0.5 mag because the non-variable neighboring blend (third star in the fit) has been removed from the baseline flux. The semi-periodic variations are removed. However there remains an annual trend. ## Removal of Variability II: Annual Variations The bottom panel of Figure [\[fig:one\]](#fig:one){reference-type="ref" reference="fig:one"} shows that this annual trend is due to variations with sidereal time, almost certainly due to the impact of the bright red neighbor (even if it were constant) via differential refraction. We fit this variation to polynomials of order \(n\), but find that there is no significant improvement beyond \(n=1\), for which \(f(t) = 0.2415 + 0.1490(t-0.5)\), where \(f\) is flux in units of \(I=18\). Note that the variation from 0.3 to 0.7 (on the figure) is \(0.4\times 0.149\sim 0.06\) which is 1.5 times larger than the source flux derived below. The third panel shows the results of applying this sidereal-time correction. As expected the annual trend is gone. We apply this flux correction to all data, not just the baseline data shown in this figure. We find (as expected) that this corrects the slope of the rising part of the light curve, which indeed impacts the estimate of \(t_{\rm E}\), though by less than 10%. Note that while, in most cases, it is possible to derive an accurate estimate of the photometric error bars from those reported by the photometric pipeline, in this case we did not apply this simple prescription because the data were reduced using a special customized pipeline. ## Correction of KMTNet Data We apply the same formalism given by Equations ([\[eqn:multifit1\]](#eqn:multifit1){reference-type="ref" reference="eqn:multifit1"}) and ([\[eqn:multifit2\]](#eqn:multifit2){reference-type="ref" reference="eqn:multifit2"}) to the KMTNet data, but with only two stars, i.e., the microlensed source and the neighboring variable. Note that this difference between OGLE and KMTNet reductions plays no role in the final result because the third star incorporated into the OGLE fit is not variable, and the KMTNet flux scale is ultimately aligned to OGLE through the microlens fit. Thus, in particular, we retain the advantage of resolving out this blend, thus placing better limits on flux from the lens. We note that it is difficult to correct for the annual variation in the KMTNet data. Reliable measurement of the annual variation would require baseline data, which do not exist because the photometry system was not the same for the whole of 2015 compared to 2016. In principle, we could have applied the OGLE-based correction to KMTNet data, but this type of correction is observatory-specific and this would not have been a reliable approach and could easily cause more problem than what was being corrected for. From checking the impact of only correcting the OGLE data for variable-contamination (but not annual variation), however, we found essentially no change. The scatter (hence renormalized error bars) are very slightly smaller, but no change in parameters. The same would be case for KMTNet data. ## Guideline for Assessing the Need of Multi-star Fitting The formalism introduced in Section 3.3 can also be used to gain intuition about the impact of uncorrected variability, which can then be used to assess whether such corrections are necessary in specific cases. First note that for \(n=1\), Equations ([\[eqn:multifit1\]](#eqn:multifit1){reference-type="ref" reference="eqn:multifit1"}) and ([\[eqn:multifit2\]](#eqn:multifit2){reference-type="ref" reference="eqn:multifit2"}) reduce to the standard formula: \[F_1 = {d_1\over b_{11}}; \qquad b_{11} = \sum_k {P_{1,k}^2\over \sigma_k^2}; \qquad d_1 = \sum_k {P_{1,k}f_k\over \sigma_k^2} \label{eqn:singlefit}\] Equation ([\[eqn:singlefit\]](#eqn:singlefit){reference-type="ref" reference="eqn:singlefit"}) then allows us to express the properly corrected "true" photometry in terms of the "naive" single-source photometry that ignores neighbors. We first "infer" the value, \(d_{i,\rm inferred} = b_{ii}*F_{i,\rm naive}\), which then yields, \[F_{i,\rm true} = \sum_{j=1}^2 c_{ij}d_{j,\rm inferred}= {F_{i,\rm naive}-(b_{12}/b_{ii})F_{(3-i),\rm naive}\over 1-b_{12}^2/(b_{11}b_{22})}. \label{eqn:backout}\] Hence, the correction is governed by the ratio of the PSF overlap integral \(b_{12}\) to the integral of the PSF squared, \(b_{11}\). We can evaluate this explicitly for the special case of a Gaussian PSF and below-sky sources (\(\sigma_k =\)const) \[{b_{12}\over b_{11}} = 4^{-(\Delta \theta/{\rm FWHM})^2} \qquad (\rm below-sky\ Gaussian), \label{eqn:gaussbelow}\] where \(\Delta\theta\) is the separation between the two sources. If the two sources are reasonably well separated, \(\Delta\theta\ga\)FWHM, and (as in the present case) the target (1) is below sky while the contaminating variable (2) is well above sky, then the effect is roughly half of that given by Equation ([\[eqn:gaussbelow\]](#eqn:gaussbelow){reference-type="ref" reference="eqn:gaussbelow"}). This is because the squared PSF integral is basically unaffected while the half of the contribution to the overlap integral that is closer to the contaminant is heavily suppressed by the higher flux errors per pixel. We close by re-emphasizing that these order-or-magnitude estimates are not used in the present analysis but are intended as guidance for future cases. # Light Curve Analysis The lightcurve, presented in Figure [\[fig:two\]](#fig:two){reference-type="ref" reference="fig:two"}, has two principal features: a strong caustic entrance near peak at HJD\(^\prime = 7486.4\) and a weak caustic exit at HJD\(^\prime = 7502.6\). Apart from these caustic crossings, the morphology is that of a slightly distorted point-lens event. This morphology points to a binary lens with very unequal mass ratio \(q\ll 1\), i.e., in the brown-dwarf or planetary regime. The long duration of the caustic (16 days) then points to a resonant caustic, and so projected separation (in units of \(\theta_{\rm E}\)) of \(s\sim 1\). A thorough search of the parameter space spanning \(-1.0\leq \log s \leq 1.0\) and \(-5.0\leq \log q \leq 1.0\) leads to only one viable solution, which confirms the above naive reasoning. The uniqueness of the solution is shown in Figure [\[fig:three\]](#fig:three){reference-type="ref" reference="fig:three"}, where we present the \(\Delta\chi^2\) map of the MCMC chain in the \(\log s\)--\(\log q\) parameter space obtained from the preliminary grid search. In fact, initial modeling based on data taken up through HJD\(^\prime = 7500.8\) (so, before the caustic exit) already led to essentially this same solution (although the predicted caustic exit was 2.6 days later than the one subsequently observed). The model is described by seven parameters. These include three that are analogous to a point-lens event \((t_0,u_0,t_{\rm E})\), i.e., the time of closest approach to the center of magnification, the impact parameter normalized to \(\theta_{\rm E}\) and the Einstein crossing time; three to describe the binary companion \((s,q,\alpha)\) where \(\alpha\) is the angle of binary axis relative to the source trajectory; and \(\rho\equiv \theta_*/\theta_{\rm E}\), where \(\theta_*\) is the angular radius of the source. The best fit parameters and errors (determined from a Markov Chain) are given in Table [\[table:one\]](#table:one){reference-type="ref" reference="table:one"}. We present the model light curve superposed on the data points in Figure [\[fig:two\]](#fig:two){reference-type="ref" reference="fig:two"} and the lens geometry is shown in Figure [\[fig:four\]](#fig:four){reference-type="ref" reference="fig:four"}. We also fit the lightcurve for the microlens parallax effect, but found no improvement. We note that compared to other planetary and binary events with well-covered caustic crossings, the parameter \(\rho=(6.0\pm 0.8)\times 10^{-4}\) (and the parameter combination \(t_* = \rho t_{\rm E} = 0.049 \pm 0.007\)) have relatively large errors. These parameters are usually better measured because caustic crossings tend to be bright (since the caustic itself is a contour of formally infinite magnification), which means that the photometry over the caustic crossing is relatively precise. Since \(t_*\) depends almost entirely on the duration of this crossing, with only weak dependence on other model parameters, it can then be determined quite precisely. In the present case, however, the first crossing was entirely missed simply because it was not visible from any of the five survey telescopes currently in operation (OGLE, MOA, and three from KMTNet: CTIO, SAAO, SSO). The caustic exit was captured by KMTNet SAAO, with 12 points taken over 3.63 hours (i.e., 20 minute cadence). However, since this caustic was quite weak, peaking at \(I\sim 18.1\), the photometry has much larger errors than the SAAO photometry near peak. See upper two panels of Figure [\[fig:two\]](#fig:two){reference-type="ref" reference="fig:two"}. Adopting a more glass-half-full orientation, we should assess the prior probability that either of the two caustic crossings would have been adequately observed to measure \(\rho\). Considering the 20 days between HJD\(^\prime\) 7485 and 7505, the three KMTNet observatories each took at least two points on 13 nights, with total durations, (2.11,2.76,2.52) days for SSO, SAAO, and CTIO, i.e., a total 7.39 days. Essentially all of these 39 intervals had approximately continuous coverage. We estimate that the probability that \(\rho\) can be measured is the same as the probability that the caustic peak is covered, which may be slightly too conservative. Under this assumption, the probability that the caustics would be observed are each 37%, so that the probability that at least one would be observed is \(1-(1-0.37)^2=46\%\). Of course, since the midpoint of the two caustic crossings was 16 April, this probability is adversely affected by the shortness of the bulge observing window relative to microlensing "high season" (21 May--21 July). At that time the observing window is roughly 2.5 hours longer, and so (assuming comparable weather conditions), the probability for each crossing would be 52% and the probability for at least one would be 77%. Nevertheless, the mid-April values may be considered as a proxy for the microlensing season as a whole. # Physical Parameters We use KMTNet CTIO DoPHOT reductions to construct an instrumental color magnitude diagram (CMD) that is presented in Figure [\[fig:five\]](#fig:five){reference-type="ref" reference="fig:five"}.[^3] We find the instrumental source color from model-independent regression and the instrumental source magnitude by fitting the \(I\) band light curve to the model. We then find the offset from the clump \(\Delta[(V-I),I]=(-0.23,4.06)\pm (0.03,0.10)\), where the error in the color offset is dominated by the regression measurement while the error in the magnitude offset is dominated by fitting for the clump centroid. We then adopt \([(V-I),I]_{0,\rm clump}=(1.06,14.49)\) to obtain \([(V-I),I]_{0,s}=(0.83,18.55)\). Then using standard techniques, we convert from \(V/I\) to \(V/K\) using the \(VIK\) color-color relations and then use the color/surface-brightness relations to derive \[\eqalign{ \theta_* = 0.690\pm 0.065\,\ \mu{\rm as}; \cr \theta_{\rm E} = {\theta_*\over\rho} = 1.15\pm 0.18\ {\rm mas}; \cr \mu = {\theta_{\rm E}\over t_{\rm E}} = 5.1\pm 0.8\,\ {\rm mas}\,{\rm yr}^{-1}.\cr } \label{eqn:thetastar}\] The error in \(\theta_*\) is dominated by the uncertainties in transforming from color to surface brightness (8%), with a significant contribution from the error in \(I_s\) (5%). The fractional errors in \(\theta_{\rm E}\) and \(\mu\) are substantially larger than in \(\theta_*\) due to the relatively large error in \(\rho\). See Section 4. The relatively large value of \(\theta_{\rm E}\) almost certainly implies that the lens lies in the Galactic disk since the lens-source relative parallax is \[\pi_{\rm rel} = {\theta_{\rm E}^2\over\kappa M} = (0.16\pm 0.05 {\rm mas}) \biggl({M\over M_\odot}\biggr)^{-1}. \label{eqn:mpirel}\] That is, only if the lens were substantially heavier than \(1\,M_\odot\) could it be in the bulge (\(\pi_{\rm rel}\la 0.03\)). However, first, there are almost no such massive stars in the bulge and second, its light would then exceed the blended light \((I_b\sim 19.7)\), even allowing for the \(A_I=2.96\) extinction toward this line of sight. The only exception to this line of reasoning would be if the lens were a black hole. Although the model considering parallax effects does not have improvement compared to the non-parallax model, non-detection of \(\pi_{\rm E}\) can give constraints on the mass and distance. In Figure [\[fig:six\]](#fig:six){reference-type="ref" reference="fig:six"}, we present the \(\Delta\chi^2\) map of the MCMC chain in the \(\pi_{{\rm E},E}\)--\(\pi_{E,{\rm N}}\) parameter space obtained from the modeling considering both the lens orbital motion and the microlens parallax effect. The upper limit of the microlens parallax as measured 3\(\sigma\) level is \(\pi_{\rm E} \lesssim 0.4\). This gives the lower limits of the mass and distance of \(M\gtrsim 0.35\ M_\odot\) and \(D_{\rm L}\gtrsim 1.7\) kpc. # Resolving the Nature of the Planet The most notable characteristic of OGLE-2016-BLG-0596 is its high mass ratio \(q=0.0117\pm 0.0004\), implying that the mass is \(m_p = 12.2\,M_{\rm jup} (M/M_\odot)\). Hence, if the host is one solar mass, this planet would be just below the the deuterium-burning limit (usually regarded as the planet/brown-dwarf boundary). While the host could in principle have arbitrarily low mass (and so, by Equation ([\[eqn:mpirel\]](#eqn:mpirel){reference-type="ref" reference="eqn:mpirel"}), be arbitrarily close), distances closer than \(D_L\la 1\,{\rm kpc}\) are strongly disfavored by the relatively low proper motion, the parallax constraint, and the paucity of nearby lenses. At this limiting distance, and so \(M =\theta_{\rm E}^2/\kappa\pi_{\rm rel} \sim 0.18\,M_\odot\), the planet would still be \(m_p\sim 2\,M_{\rm jup}\), i.e., quite massive for such a low-mass host. Hence, regardless of the host mass, this is a fairly extreme system. To distinguish among these interesting possibilities will require measuring (or strongly constraining) the host mass. This can be accomplished with high resolution imaging, either using the *Hubble Space Telescope (HST)* or ground-based adaptive optics (AO) imaging on an 8m class telescope. An advantage of *HST* is that it can observe in the \(I\) band for which the source flux is directly measured from the event. Hence, the source light can be most reliably separated from the blended light in \(I\). In contrast to many previous cases, there are no \(H\) band observations during the event, so ground-based AO observations (which must be in the infrared) cannot be directly compared to an event-derived source flux. Nevertheless, it is probably possible to transform from \(V/I\) light-curve measurement to \(H_s\) source flux with a precision of 0.2 mag, using a \(VIH\) color-color diagram. For definiteness, we will assume that the lens can be reliably detected from *HST* or AO observations provided that the flux is at least half that of the source, i.e., \(I_L<22.3\). For example, if the lens were an \(M=0.5\,M_\odot\) early M dwarf (so \(D_L\sim 2.2\,{\rm kpc}\)), then it would have \(I_L \sim 18.8 + A_I\leq 21.8\) (and brighter if, as is almost certainly the case, a substantial fraction of the dust is beyond 2.2 kpc). Thus, there is a good chance that AO or *HST* observations could detect the lens, and even if this failed, the observations would strongly constrain the host to be of very low mass. # Discussion OGLE-2016-BLG-0596Lb is a very high mass-ratio \((q=0.0117)\) planet that lies projected very close to the Einstein ring \((s=1.075)\), which consequently generated a huge resonant caustic that required 16 days for the source to traverse. The underlying event was of quite high magnification \((A_{\rm point-lens}\sim 100)\), which led to pronounced features at peak. It therefore would seem to be extremely easy to discover. While the data set posted on the OGLE web site are adversely affected by the nearby variable, it is still the case that a free fit to these data leads to a solution qualitatively similar to the one presented here (except that it lacks a measurement of \(\rho\)). It is therefore striking that none of the automated programs nor active individual investigators that query this site noticed this event (or at least they did not alert the community to what they found as they do for a wide range of other events, many less interesting). This indicate the possibility that there may be many other planets "hidden in plain sight" in existing data. This is also supported by the planet discoveries MOA-2008-BLG-397Lb, OGLE-2008-BLG-355Lb, and MOA-2010-BLG-353Lb, for which the planetary signals were not noticed during the progress of events. These three characteristics, high-magnification (which is usually associated with survey+followup rather than survey-only mode), very high mass ratio, and apparent failure of both machine and by-eye recognition of the planetary perturbation, prompt us to address two questions. First, how do the real (as opposed to theoretical) planet sensitivities differ between survey-only and survey+followup modes. Second, why was this planet discovered only based on systematic analysis and what does this imply about the need for such systematic analysis of all events? ## Summary of Microlens Planet Detections in the Observational \((s,q)\) Plane Many papers contain figures that summarize microlensing planet detections in the physical plane of planet-mass versus projected separation (with the latter sometimes normalized by the snow line), e.g., Figure 1 of. And there are many studies that show plots of *planet sensitivity* in the observational \((s,q)\) plane, (e.g., @gaudi02 [@gould10]). But to our knowledge, there are no published figures (or even figures shown at conferences) showing the census of microlensing planet discoveries on this plane. Figure [\[fig:seven\]](#fig:seven){reference-type="ref" reference="fig:seven"} illustrates the position of OGLE-2016-BLG-0596 (green pentagon) among the 44 previously published planets (or, to the extent we have such knowledge, submitted for publication). Discovered bodies are defined to be "planets" if their measured or best-estimated mass \(m_p<13\,M_{\rm jup}\) and if they are known to orbit a more massive body[^4]. Planets are color-coded by discovery method: discovered by followup observations (blue), discovered (or discoverable) in survey-only observations but requiring followup for full-characterization (magenta), fully (or essentially fully) characterized by survey observations (green). The shapes of the symbols indicate the type of caustic that gave rise to the planetary perturbation: circles, squares, and triangles for planetary, central, and resonant caustics, respectively. In many cases, solutions with \((q,s)\) and \((q,1/s)\) yield almost equally good fits to the data. In these cases, the two solutions are shown as open symbols in order to diminish their individual visual "weight" relative to the filled symbols used when this degeneracy is broken. Hence, OGLE-2016-BLG-0596Lb would be a green filled triangle if it were not being singled out by making it larger pentagon. The most striking feature of this figure is that, in sharp contrast to the triangular appearance of high-magnification-event planet-sensitivity plots (e.g., @gould10) and to "double pronged" low-magnification sensitivity plots (e.g., @gaudi02), this detection plot looks basically like a cross, with a vertical band of detections near \(\log s\sim 0\) and a horizontal band near \(\log(q)\sim-2.5\). The part of this structure at high mass ratio \(\log(q)>-2\) is easily explained: companions with high mass ratio are, a priori, most likely stars or brown dwarfs (BDs) and can only be claimed as "planets" if the host mass is known to be low. This in turn usually requires a measurement of the microlens parallax, which for ground based observations is much more likely if there is a large caustic and so \(s\sim 1\). We note that there are 4 planet detections in the region \((\log(s)>+0.15, \log(q)<-3)\), while there is no detection in the opposite quadrant \((\log(s)<-0.15, \log(q)<-3)\). All the 4 planets derive from planetary caustics and 3 of them are pure survey detections: MOA-2011-BLG-028Lb[^5], OGLE-2008-BLG-092Lb, and MOA-2013-BLG-605Lb. The remaining planet, OGLE-2005-BLG-390Lb, dates from an era when followup groups intensively monitored the wings of events, primarily due to the paucity of better targets. Thus we may expect that surveys will gradually fill in this quadrant. The difference in the detection rates between the quadrants with \(\log(s)>+0.15\) and \(\log(s)<-0.15\) can be explained by the difference in the size of the planetary caustics with \(s<1\) and \(s>1\). In the case of \(s>1\), there exist a single planetary caustic. In the case of \(s<1\), on the other hand, there exist two sets of planetary caustics and each one is smaller than the planetary caustic with \(s>1\). As a result, the planetary caustic with \(s>1\) has a larger cross section and thus higher sensitivity. Furthermore, smaller caustic size of planets with \(s<1\) makes planetary signals tend to be heavily affected by finite-source effects, which diminish planetary signals, while signals of planets with \(s>1\) can survive and show up in the wings of light curves. Actually, all 4 events with planet detections via planetary-caustic perturbations are involved with large source stars, i.e. giant and subgiant stars for which finite-source effects are important. Apart from this quadrant, it is not obvious that surveys are probing a different part of parameter space from the previously dominant survey+followup mode. To further investigate this, we show in Figures [\[fig:eight\]](#fig:eight){reference-type="ref" reference="fig:eight"} and [\[fig:nine\]](#fig:nine){reference-type="ref" reference="fig:nine"} the cumulative distributions of planets by log mass ratio \(\log(q)\) and (absolute value of) log separation \(|\log(s)|\). In this case we distinguish only between events that could be fully characterized by survey observations (green) and those that required significant followup (including auto-followup by surveys). These distributions generally appear quite similar. For the mass ratio distribution, the greatest difference (0.259) is at \(\log(q)=-2.319\), which is very typical (Kolmogorov-Smirnoff (KS) probability 40%). The greatest difference for the separation distribution (0.334 at \(|\log(s)|=0.124\)) has a KS probability of 8.5%. This may be indicative of a real difference. If so, the difference would be that pure-survey is relatively more efficient at finding widely separated lenses, which was already hinted at by inspection of the \((q,s)\) scatter plot. Finally, in Figure [\[fig:ten\]](#fig:ten){reference-type="ref" reference="fig:ten"}, we show cumulative distributions by year of discovery. One might expect that with the massive ramp-up of surveys, survey-only discoveries would move strongly ahead of survey+followup. This expectation is confirmed in its sign but not its magnitude by Figure [\[fig:ten\]](#fig:ten){reference-type="ref" reference="fig:ten"}. It shows that in (2014, 2015, 2016) there have been (2,2,1) and (0,1,0) discoveries by survey-only and survey+followup, respectively. This is certainly not a complete accounting, in part because 2016 has just begun and in part because historically there has been a considerable delay in microlensing planet publications for a variety of reasons. For example, of the 28 planets discovered prior to 2012, the number with delays (publication year minus discovery year) of \((0,1,\ldots,9)\) years was \(N=(1,5,9,5,1,2,4,0,0,1)\). In the history of microlensing, there has been only one planet published during the discovery year, OGLE-2005-BLG-071Lb. Hence, we will only get a full picture of this transition after a few years. ## Challenges to the By-Eye and By-Machine discovery of OGLE-2016-BLG-0596 There are three interrelated reasons why OGLE-2016-BLG-0596 may have escaped notice as a potentially planetary event until the KMTNet data for this event were examined (for reasons unrelated to any apparent anomaly). First, it is relatively faint at peak. Second, it has a variable baseline. Third, it was not announced as a microlensing event until one day after the peak. As a general rule, high-magnification events are singled out for intensive followup observations only if they are still rising. When such intensive observations would have been conducted, they would have immediately revealed the anomalous nature of the event, probably triggering additional observations. This is how many of the planets discovered by \(\mu\)FUN were found. While \(\mu\)FUN itself is now semi-dormant, its protocols are directly relevant here because what is of interest is whether there is prima facie evidence for a population of missed planets during past years, during most of which \(\mu\)FUN was active. Now, in fact, OGLE-2016-BLG-0596 met the criteria for an OGLE alert 24 hours previously, but no alert was issued because of caution due to the variable baseline. Nevertheless, even if such an alert had been issued, it would not have triggered any followup observations because (due to the anomaly) the event would have appeared to have already peaked at that time. Finally, the variability of the baseline may have influenced modelers and followup groups to discount the evident irregularities in the light curve near peak as being due to data artifacts. This could have been exacerbated by the faintness of the event, which increases both the formal error bars and the probability of centroiding errors (hence irregular photometry) due to bright blends. Both of these effects reduce the confidence of modelers that apparent anomalies in online "quick look" photometry are due to physical effects. It is nevertheless a fact that when the original OGLE data are modeled, they show a clear signal for a massive planet or low-mass BD, which would trigger a re-reduction of the data, such as the one we report here. We therefore conclude that while OGLE-2016-BLG-0596 has some near-unique features that increased the difficulty of recognizing it as a planetary event, such recognition was clearly feasible. Hence, we do indeed regard this event as prima facie evidence for more such events in archival data, particularly OGLE-IV data 2010-2015. The OGLE project has received funding from the National Science Centre, Poland, grant MAESTRO 2014/14/A/ST9/00121 to AU. Work by C.H. was supported by Creative Research Initiative Program (2009-0081561) of National Research Foundation of Korea. The OGLE Team thanks Profs. M. Kubiak and G. Pietrzyński, former members of the OGLE team, for their contribution to the collection of the OGLE photometric data over the past years. WZ and AG were supported by NSF grant AST-1516842. Work by JCY was performed under contract with the California Institute of Technology (Caltech)/Jet Propulsion Laboratory (JPL) funded by NASA through the Sagan Fellowship Program executed by the NASA Exoplanet Science Institute. This research has made the telescopes of KMTNet operated by the Korea Astronomy and Space Science Institute (KASI). [^1]: OGLE cadences were significantly adjusted at the time of the peak and planetary anomaly of this event, due to the *Kepler* K2 Campaign 9 microlensing campaign. The five fields covering the K2 field were observed 3 times per hour, while other fields (including BLG534) were observed somewhat less frequently (very roughly 2/3) compared to their usual rates. [^2]: Like OGLE, KMTNet also adjusted its schedule for the K2 campaign, but in a different way. First, CTIO observations were not adjusted. Second, KMTNet only began "K2 mode" on 2016 April 23. This was after the event peak and caustic entrance but before the exit. Therefore, in particular, the caustic exit observations from SAAO were at the lower cadence (reduced by a factor 0.75) [^3]: Correction of the DoPHOT data for variation is not done, but this would have little effect on the result. The variable is extraordinarily red, \(\sim 1.2\) magnitude redder than the clump, whereas the source is \(\sim 0.2\) magnitude bluer than the clump. Hence, by a naive estimate, the variations would be fractionally smaller by a factor 4. The full amplitude of these variations in \(I\) band is of order the source flux, whereas the color measurement is made when the source is magnified 60 to 100 times. The color measurement is differential over short timescales of a few days, whereas the period is a large fraction of a year. Combining these very small factors, we expect the color measurement to be impacted at the level \((1/4)\times (1/80)\times (3/(126/\pi)) \sim 2\times 10^{-4}\). It is general practice to ignore such small errors, which in this case are more than hundred times smaller than the measurement error. We also note that the dependence of the color measurement on the choice of the \(V\)-band data set (OGLE or KMTNet) is small considering that the offset from the clump has an accuracy of 0.03 magnitude whereas the precision of the color measurement is 0.05 magnitude. Furthermore, the SAAO \(V\)-band data are taken for redundancy, primarily in a case there is no CTIO data due to bad weather when the event is well magnified or for very short, highly magnified events that peak of South Africa. [^4]: To facilitate comparison with future compilations, we list here the 45 planets used to construct this figure and those that follow. We compress, e.g., OGLE-2003-BLG-235Lb to OB03235 for compactness and only use "b,c" for multiple planets: OB03235, OB05071, OB05169, OB05390, MB06bin1, OB06109b, OB06109c, MB07192, MB07400, OB07349, OB07368, MB08310, MB08379, OB08092, OB08355, MB09266, MB09319, MB09387, MB10073, MB10328, MB10353, MB10477, MB11028, MB11262, MB11293, MB11322, OB110251, OB110265, OB120026b, OB120026c, OB120358, OB120406, OB120455, OB120563, OB120724, MB13220, MB13605, OB130102, OB130341, OB140124, OB141760, OB150051, OB150954, OB150966, OB160596. [^5]: We note that this event's light curve does contain some followup data, but it is not essential for characterizing the planet.
{'timestamp': '2017-02-21T02:08:56', 'yymm': '1607', 'arxiv_id': '1607.04919', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04919'}
null
null
# Introduction {#sxn:intro} Graphs, long popular in computer science and discrete mathematics, have received renewed interest recently in statistics, machine learning, data analysis, and related areas because they provide a useful way to model many types of relational data. In this way, graphs can be used to extract insight from data arising in many application domains. In biology, e.g., graphs are routinely used to generate hypotheses for experimental validation ; and in neurscience, they are used to study the networks and circuits in the brain . Given their ubiquity, graphs and graph-based data have been approached from several different perspectives. In computer science, it is common to develop algorithms, e.g., for connected component, minimum spanning tree, and maximum flow problems, to run on data that are modeled as a precisely-specified input graph. These algorithms are then characterized by the number of operations they use for the worst-possible input at any size. In statistics and machine learning, on the other hand, it is common to use graphs as models to perform inference about unseen data. In this case, one often hypothesizes an unseen graph with a particular structure, such as block structure, hierarchical structure, low-rank structure in the adjacency matrix, etc. Then one runs algorithms on the observed data in order to impute entries in the unobserved hypothesized graph. These methods may be characterized in terms of running time, but they are also characterized in terms of the amount of data needed to recover the hidden hypothesized structure. In many application areas where the end goal is to obtain some sort of domain-specific insight, e.g., such as in social network analysis, neuroscience, medical imaging, etc., one constructs graphs from primary data, and then one runs a computational procedure that does *not* come with either of these traditional types of theoretical guarantees. As an example, consider the GeneRank method , where we have a set of genes related to an experimental condition in a microarray study. This set of genes is "refined" via a locally-biased graph algorithm closely related to those we will discuss. Importantly, this operational refinement procedure does *not* come with the sort of theory traditional in statistics, machine learning, or computer science. As anther example, e.g., in social network applications, one might run a random walk process for a few steps from a starting node of interest, and if the walk "gets stuck" then one might interpret this as evidence that that region of the graph is meaningful in the application domain . These are examples of the types of heuristics commonly-used in applications. By heuristic, we mean an algorithm in the sense that it performs a sequence of well-defined steps, but one where precise theory is lacking (although usually heuristics come with strong intuitive motivation and are justified in terms of some downstream application). In particular, typically heuristics do not explicitly optimize a well-defined objective function and typically they do not come with well-defined inferential guarantees. Note that, in both of these examples, the goal is to find "local" or "small scale" structure in the data graph. Both examples also correspond to what practitioners interested in downstream applications actually do. Existing algorithmic and statistical theory, however, has challenges with these local or small-scale structures. For instance, a very "good" algorithmic runtime on a graph is traditionally one that is *linear* in the number of vertices and edges. If the output of interest is only a vanishingly small fraction of a large graph, however, then this theory may not provide strong qualitative guidance on how these locally-biased methods behave in practice. Likewise, inferential methods often assume that the structures inferred constitute a substantial fraction of the graph, and many statistical techniques have challenges differentiating very small structure from random noise. In this overview, we describe a class of graph algorithms that has proven to be very useful for identifying and interpreting small-scale local structure in large-scale data. For this class of algorithms, however, strong algorithmic and statistical theory has been developed. In particular, these graph algorithms are *locally-biased* in one of several precisely-quantified senses. We will describe what we mean by this in more detail below, but, informally, this means that the algorithms are most interested in only a small part of a large graph. As opposed to heuristic operational procedures, however, many of these algorithms do come with strong worst-case algorithmic guarantees, and many of these algorithms also do come with statistical guarantees that prove they have *implicit* regularization properties. This complementary algorithmic-statistical theory helps explain their improved performance in many practical applications. While the approach of locally-biased graph algorithms is very general, it has been developed most extensively for the fundamental problem of finding locally-biased graph partitions, i.e., clusters or communities, and so we will focus on locally-biased approaches for the graph clustering problem. Of course, this partitioning question is of interest in many more application-driven areas, where one is interested in finding useful or meaningful clusters as part of a data analysis pipeline. ## The rationale for local analysis in real-world data As a quick example of why local graph analysis is frequently used in data and data science applications, we present in Figure [\[fig:small-cluster\]](#fig:small-cluster){reference-type="ref" reference="fig:small-cluster"} the results of finding the best partition of both a random geometric graph and a more typical data graph. Standard graph partitioning algorithms must operate on, or "touch", each vertex and edge of the graph to identify these partitions. The best partition of the geometric graph is around half the data, where it is reasonable to run an algorithm that touches all the data. On the other hand, the best partition of the data graph is very small, and in this case touching the entire graph to find it can be too costly in terms of computation time. The local graph clustering techniques discussed in this paper can find this cluster touching only edges and nodes in the output cluster, greatly reducing the computation time. Far from being a pathology or a peculiarity, the finding that optimal partitions of real-world networks are often extremely imbalanced, thus leading to very small optimal clusters, is endemic to many of the graphs arising in large-scale data analysis .[^1] Let us now explain in more detail Figure [\[fig:small-cluster\]](#fig:small-cluster){reference-type="ref" reference="fig:small-cluster"}. Figure [\[fig:small-cluster-a\]](#fig:small-cluster-a){reference-type="ref" reference="fig:small-cluster-a"} shows a graph with 3000 nodes that is typical of many graphs that possess a strong underlying *geometry*, e.g., those used in computer graphics, computer vision, logistics planning, road network analysis, and so on. This particular graph is produced by generating 3000 random points in the plane and connecting all points within a small radius, such that the final graph is connected. The geometric graph can be nearly bisected by optimizing a measure known as conductance (we shall define this shortly), which is designed to balance partition size and quality. In Figure [\[fig:small-cluster-b\]](#fig:small-cluster-b){reference-type="ref" reference="fig:small-cluster-b"}, we show a more typical data graph of around 10680 nodes , where this particular data graph is based on the trust relationships in a PGP (Pretty Good Privacy) key-chain. Optimizing the same conductance objective function results in a tiny set, Figure [\[fig:small-cluster-d\]](#fig:small-cluster-d){reference-type="ref" reference="fig:small-cluster-d"}, and not the near-bisection as in the geometric case, Figure [\[fig:small-cluster-c\]](#fig:small-cluster-c){reference-type="ref" reference="fig:small-cluster-c"}. (We were able to use integer optimization techniques to directly solve the NP-hard problems at the cost of months of computation.) Many other examples of this general phenomenon can be found in prior work . ## Partitioning as a model problem The problem of finding good partitions or clusters is ubiquitous. Representative examples include biomedical applications, internet and world-wide-web, social graphs, human communication graphs, human mobility graphs, voting graphs in political science, protein interaction graphs, material science, neuroscience , collaboration graphs. All of the features of locally-biased computations are present in this model partitioning problem. For example, while some of these algorithms read the entire graph as input but are engineered to return answers that are biased toward and meaningful for a smaller part of the input graph , other algorithms can take as input a small "seed set" of nodes as well as an oracle with which to access neighbors of a node, and they return meaningful answers without even touching the entire graph . Similarly, while these algorithms are often formulated in the language of theoretical computer science as approximation algorithms, i.e., they come with running time guarantees and can be shown to *approximate* to within some quality-of-approximation factor some objective function of interest, e.g., conductance, in other cases one can prove statistical results such as that they *exactly* solve a regularized version of that objective . Importantly, this statistical regularization is implicit rather than explicit. Typically, regularization is explicitly used in statistics, when fitting models with a large number of parameters, in order to avoid overfitting to the given data. It is also used to select answers biased towards specific sets---for example, sparse solution sets by using the Lasso . In the case of locally-biased graph algorithms, one simply runs a *faster* algorithm for a problem. In some cases, the nature of the approximation that this fast algorithm makes can be related back to a regularized variant of the objective function. ## Overview In this paper, we will consider partitioning from the perspective of conductance, and we will survey a recent class of results about a common localizing construction. These methods will let us find the optimal sets in Figure [\[fig:small-cluster\]](#fig:small-cluster){reference-type="ref" reference="fig:small-cluster"} without resorting to integer optimization (but also losing the proof of optimality). Depending on the exact construction, they will also come with a variety of helpful statistical properties. We'll conclude with a variety of different perspectives on these problems and some open questions. In Section [13](#sec:prelim){reference-type="ref" reference="sec:prelim"} we describe assumptions, notation and preliminary results which we will use in this paper. In Section [10](#sec:global_cond_spcut){reference-type="ref" reference="sec:global_cond_spcut"} we discuss two global graph partitioning problems and their spectral relaxation. In Section [12](#sec:localgraphpart){reference-type="ref" reference="sec:localgraphpart"} we describe the local graph clustering application. In Section [8](#sec:empirical){reference-type="ref" reference="sec:empirical"} we provide empirical evaluations for the global and local graph clustering algorithms which are described in this paper. Finally, in Section [7](#sec:disc_conclusion){reference-type="ref" reference="sec:disc_conclusion"} we give our conclusions. # Text that has been cut # Intro Motivated by these very large-scale data analysis applications where only small meaningful clusters exist and where the graphs are so large that iterating over all edges and nodes is measured in minutes or hours, recent work has focused on *locally-biased graph algorithms*. This refers to a broad class of algorithms that are designed to take as input a large graph, e.g., a billion node social or information network, and to solve problems that are in some sense "nearby" an exogenously-specified "seed set" of nodes. While these algorithms are often very useful in practice, and while many of them have connections with heuristic procedures that practitioners often apply to their data, a distinguishing feature of them is their non-trivial algorithmic and/or statistical theory. In this overview paper, we will cover recent work of global and local graph algorithms, touching on each of these topics, reviewing complementary aspects, and focusing on optimization questions. We split this paper into three parts, optimization objectives, global and local algorithms. First we will present optimization objectives which by consensus quantify the notion of a cluster in a graph. The optimization problems that we present are NP-hard, thus cannot be optimized in polynomial time. Then we discuss state-of-the-art polynomial time relaxations that result in spectral or max-flow problems. Second, we present an overview of locally-biased graph partitioning methods in terms of *flow-based methods* and *spectral-based methods*. Finally, we illustrate their performance in practice and comment on the implicit or explicit regularization that is applied by each algorithm. # Regularization, Beyond \(\ell_1\) and \(\ell_2\) It is easy to see that the above problem is a continuous representation of \(s\)-\(t\) Min-Cut problem. The constraints \(x_s=1\) and \(x_t=0\) represent regularization such that the trivial solution of a constant vector \(x\) is obtained. At the same time these two constraints affect the separation of the graph in two parts such that the nodes \(v_s\) and \(v_t\) are in different parts of the partition. The approximation is based on the distortion error of embedding the semi-metric to the so called \(\ell_1\)-metrics set, see ([Leighton Rao]{style="color: red"}) for details. #### Theoretical issues: combining spectral and flow. [The material below is also covered in the SDP section, we need to combine them.]{style="color: red"} Although it isn't our main focus, we comment briefly on one set of well-known results and how it fits in with our discussion. A natural question in the above is whether one can combine the complementary ideas from spectral and flow to obtain improved results. The answer is yes, and this is the ARV algorithm. It (1) can be written as a tighter relaxation of the expansion/conductance problem, (2) can be written as an SDP, (3) provides improvement over spectral as well as LR, (4) leads to related optimization-based methods/approaches such as KRF and MWUM that lead to similar improvements. XXX. GIVE THE OBJECTIVE AND RELAXATION AND RESULTS AND LEAVE IT AT THAT. BTW, also, LMO showed that the practice follows the theory for graphs up to thousands to tens of thousands of nodes , but, aside from this, these methods haven't really been implemented. XXX. WE'LL PROBABLY SAY IN THE CONCLUSION THAT LOCAL ARV METHODS ARE OPEN. ## Embedding aspects {#subsection:embeddings} [I have to embed ( :-) ) this section in previous sections since most of the material below is already mentioned above.]{style="color: red"} These two approaches have an interpretation in terms of embeddings (which is of theoretical interest, but which also can be seen empirically), and this provides a way to think of the two together as complementary (in the global case, but we will do the local case below). To start, return to min-cut, and note that that objective, while tractable, often leads to trivial solutions, in that it cuts on a very small number of nodes, e.g., one or two. Thus, it is common to normalize by the volume. (Alternatively, one could have a balance constraint or use methods that explicitly or implicitly provide a bias toward large sets of nodes.) In this case, one can define the expansion of a set of nodes to be the usual expansion-SA-to-Volume measure. XXX. DEFINE. If the graph is weighted or has variable degrees, then one could define the conductance to be the usual conductance-SA-to-Vol measure. (They are the same for unweighted/degree-regular graphs, but not otherwise and the differences are important in theory/practice.) Note that the input to these objectives is just a graph \(G=(V,E,W)\). These objectives are intractable, but they capture the qualitative idea of what most people think of by clusters---relatively large, relatively a lot of stuff inside, relatively less interaction between the cluster and its complement. The point is that spectral and flow both give approximation algorithms for expansion/conductance. Spectral relaxes one way, i.e., the \(\{-1,1\}\) constraint to being \(\pm1\) on average, and this can be solved by computing eigenvectors or diffusions Flow relaxes another way, i.e, the \(\{0,1\}\) constraint to be a metric, and this can be written as the solution of an LP, which can then be solved with flow methods like Dinic or related methods. From this embedding perspective, it is clear that spectral embeds the graph not just on a like, and it then does a sweep cut, but it also embeds the graph in a complete graph, i.e., the Laplacian of a complete graph, and it is here that the duality is the tightest. It is here that the quadratic approximation result of Cheeger arises. And note that for spectral methods random walks mix quickly on constant-degree expanders and slowly on line graphs, and this can be sued to construct the Guattery-Miller cockroach graphs. XXX. SOMETHING ABOUT CONNECTIONS WITH L2. Similarly, flow methods embed you in an \(\ell_1\) space, and Bourgain's embedding lemma says that an \(O(\log n)\) distortion can be expected, and this happens on constant degree expanders. We have gone through this in some detail since many of the ideas are essential for constructing local versions of these methods. Many of these results are "obvious" to certain research areas and not even thought about by other research areas, so we went through them in the more well-developed global partitioning area, so we can draw similarities with local methods. ## Some additional observations There is, of course, the usual theory-practice gap. Theoretical work typically considers the problem of two clusters, and applied work is often more interested in more than two clusters. The latter, while also using a lot of algorithm engineering, often uses ideas that boil down to the former. Here, we mention a few things. #### Practical issues: going beyond spectral and flow. Global graph partitioning methods basically boil down to some combination of spectral and flow, also couples with ideas like multi-resolution. The canonical example of this is METIS . This, coupled with doing things "right," e.g., doing the sweep cut right with spectral of doing something analogous like Goldberg does with flow algorithms, can have a big effect on the quality of the result. We won't go into a lot of detail about this, but we expect that a lot of these ideas can also lead to improved local and # Spectral methods The material in this section is based on. Then \[vol(E(S,S^c)) = \frac{x^T L x}{(c_1-c_2)^2}\] and \[x^TDx = c_1^2 vol(S) + c_2^2 vol(S^c).\] Depending on how we define \(c_1\) and \(c_2\) we obtain different problems. For example, if \[c_1= \sqrt{\frac{vol(S^c)}{vol(S)}}, \ c_2=-\sqrt{\frac{vol(S)}{vol(S^c)}}\] then the following problem \[\begin{aligned} \label{spcutLD} \mbox{minimize} & \quad \frac{x^TLx}{x^TDx} \\\nonumber \mbox{subject to} &\quad x\in\{c_1,c_2\}^{|\mathcal{V}|} \end{aligned}\] is equivalent to the Sparsest-Cut problem [\[spcut\]](#spcut){reference-type="eqref" reference="spcut"}. If \(c_1=1\) and \(c_2=0\) then [\[spcutLD\]](#spcutLD){reference-type="eqref" reference="spcutLD"} is equivalent to the Minimum-Isoperimetry problem [\[isop\]](#isop){reference-type="eqref" reference="isop"}. It is clear that for different values of \(c_1\) and \(c_2\) we impose different regularization on the optimal partition of the graph. However for any values of \(c_1\) and \(c_2\) problem [\[spcutLD\]](#spcutLD){reference-type="eqref" reference="spcutLD"} is still NP-hard. A possible way to overcome this difficulty is to relax the solution set to \(\{u\in\mathbb{R}^{|\mathcal{V}|} \ | 1_{|\mathcal{V}|}^T D u = 0\}\), where \(1_{|\mathcal{V}|}\) is a column vector of ones of length \(|\mathcal{V}|\). This relaxation is equivalent to finding the second largest eigenvalue of the normalized graph Laplacian \(\mathcal{L}\), which justifies the name spectral relaxation. The relaxed problem is \[\begin{aligned} \label{spcutLD} \mbox{minimize} & \quad \frac{x^TLx}{x^TDx} \\\nonumber \mbox{subject to} &\quad x\in\{u\in\mathbb{R}^{|\mathcal{V}|} \ | \ 1_{|\mathcal{V}|}^T D u = 0\}, \end{aligned}\] which is a non-convex problem, but all local solutions are of the same importance since all have the same objective value. The local solutions differ only by a positive constant factor and a sign change. The reason for imposing the orthogonality constraint \(1_{|\mathcal{V}|}^T D x = 0\) is because otherwise [\[spcutLD\]](#spcutLD){reference-type="eqref" reference="spcutLD"} has a trivial solution, the eigevector \(D^{1/2} 1_{|\mathcal{V}|}\) (any other constant vector can be used instead of \(1_{|\mathcal{V}|}\)) of the zero (the smallest) eigevalue of \(\mathcal{L}\). Let us define \(\delta_S(v_i,v_j) = |x_i-x_j|^2\). Using \(\delta_S\) we can rewrite the numerator in the objective function as \[x^T L x = \sum_{i=1}^{|\mathcal{V}|}\sum_{j=1}^{|\mathcal{V}|} C_{ij}\delta_S(v_i,v_j).\] Another way of viewing the spectral relaxation is by noticing that if instead of defining \(x: \mathcal{V} \to \{c_1,c_2\}\) to be an indicator function, we allow it to be any arbitrary function \(x: \mathcal{V} \to \mathbb{R}\), then we relax the cut-metric \(\delta_S\) to be the Euclidean square distance. # Local methods ## The Local-Isoperimetry Problem Recall that we introduced the Sparsest-Cut and Minimum-Isoperimetry problems as regularizations of the Min-Cut problem that were designed to achieve some control over the size of the set identified. The Local-Isoperimetry problem is an integer problem that adds more regularization (or a constraint) to achieve a local solution . The problem depends on a graph \(G\) and a reference set \(R\) that must be contained within the solution. We also make the size of the final set a parameter \(k\). The new program is \[\begin{aligned} \hat{\phi}(\mathcal{G};R,k) \quad := \quad \mbox{minimize} & \; \tilde{\phi}(S)= \frac{vol(E(S,S^c))}{vol(S)} \\\nonumber \mbox{subject to} & \; S\subset \mathcal{V},\ vol(S) \le k \\ \label{eq:localiso} & \; R \subseteq S. \nonumber \end{aligned}\] It differs from the Minimum-Isoperimetry problem because we have replaced the volume bound with a constant \(k\) and require that the solution set \(S\) contain the reference set \(R\). The idea with this problem is that we wish to localize the solution to the Minimum-Isoperimetry problem around a subset of nodes \(R\). We'll often take \(R\) to be a single node. Note that if \(k = vol(G)\) then this problem may be solved as a parametric flow problem using techniques akin to those in . If \(k = vol(G)/2\), then we are able to use \(|V|\) runs of this algorithm to solve the Minimum-Isoperimetry problem by setting \(R\) to be each vertex. So solving this is NP-hard in general. We illustrate one counter-intuitive feature of this problem in Figure [\[fig:local-iso\]](#fig:local-iso){reference-type="ref" reference="fig:local-iso"}. First, let us note that one of the global solutions of the Minimum-Isoperimetry problem must be a connected subgraph. However, the solution set of the Local-Isoperimetry problem may be a disconnected graph. This occurs if there is another set in the graph with a sufficiently small isoperimetry value such that it is better to pay the price for disconnecting the reference graph \(R\) and returning the union of the two. # Discussion and conclusion {#sec:disc_conclusion} Although the optimization approach we have adopted is designed to highlight similarities between different variants of locally-biased graph algorithms, it is also worth emphasizing that there are a number of quite different and complementary perspectives people in different research communities have adopted thus far on these methods. For example: **Theoretical and empirical.** The theoretical implications of these locally-biased algorithms are often used to improve the performance of long-standing problems in theoretical computer science by improving runtimes, improving approximation constants, and handling special cases. Empirically, these algorithms are used to *study* real-world data and to accelerate and improve performance on discovery and prediction tasks. Due to the strong locality, the fast runtimes for theory often manifest as extremely fast algorithms in practice. Well-implemented strongly-local algorithms often have runtimes in milliseconds even on billion-edge graphs . **Algorithmic and statistical.** Some of the work is motivated by having better algorithmic results, e.g., being fast and/or being a rigorous approximation algorithm, i.e., worst-case guarantees in terms of approximating the optimal solution of a combinatorial problem, while other work has provided an interpretation in terms of statistical properties, e.g., explicitly optimizing a regularized objective or implicitly having output that are nice in some sense, i.e., well-connected output cluster. Often, locally-biased algorithms alone suffice as the result is an improvement to some downstream activity that will necessarily look at all the data anyway. **Optimization and operational.** The locally-biased methods tend to result from stating an optimization problem and solving it with some sort of black box or white box. Strongly-local algorithms often arise by studying a specific procedure on a graph and showing that it satisfies some condition, e.g., that it terminates so quickly that it cannot explore the entire graph, that it leads to a solution with certain quality-of-approximation guarantees, etc. See, for instance, the spectral algorithms and the flow-based algorithms. In light of these complementary approaches as well as the ubiquity with which graphs are used to model data, we expect that locally-biased graph algorithms and our optimization perspective on locally-biased graph algorithms will find increasing usefulness in many application areas. # Empirical Evaluation {#sec:empirical} In this section, we illustrate differences among global, weakly local, and strongly local solutions to the problems discussed in Section [12](#sec:localgraphpart){reference-type="ref" reference="sec:localgraphpart"}. Additionally, we discuss differences between spectral and flow methods (which is equivalently \(\ell_2\) vs. \(\ell_1\) metrics for node distances). To do so, we make use of the following real-world undirected and unweighted networks. Note that the small-scale vs. large-scale clustering properties of the first three networks have been characterized previously . In addition, it is known that US-Roads has a downward-sloping network community profile. #### Global, weakly local, and strongly local solutions We first demonstrate differences among global, weakly local, and strongly local algorithms. Let us start with a comparison among spectral algorithms. By comparing algorithms that use that same metric, i.e., \(\ell_2\), to measure distances among nodes we minimize factors that can affect the solution, and we focus on weak vs. strong locality. In all figures we show the solution obtained by an algorithm without applying any rounding procedure. We illustrate the importance of the nodes by colouring and size; details are explained in the captions of the figures and in the text. The layout for all graphs has been obtained using the force-directed algorithm, which is available from the graph-tool project.[^3] For US-Senate, the comparison is shown in Figure [\[Fig_senate_spectral\]](#Fig_senate_spectral){reference-type="ref" reference="Fig_senate_spectral"}. Figures [\[Fig_senate_spectral_1\]](#Fig_senate_spectral_1){reference-type="ref" reference="Fig_senate_spectral_1"} and [\[Fig_senate_spectral_2\]](#Fig_senate_spectral_2){reference-type="ref" reference="Fig_senate_spectral_2"} show the solutions of global algorithms, Spectral relaxation and MOV global (\(z=1_{|\mathcal{V}|}\) and then we orthogonalize \(z\) with respect to \(D1_{|\mathcal{V}|}\)), respectively. As expected, the US-Senate graph has two large clusters, i.e., before the year \(1913\) and after that year, that partition along the one-dimensional time axis. This global structure is nicely captured by Spectral relaxation and MOV global in Figures [\[Fig_senate_spectral_1\]](#Fig_senate_spectral_1){reference-type="ref" reference="Fig_senate_spectral_1"} and [\[Fig_senate_spectral_2\]](#Fig_senate_spectral_2){reference-type="ref" reference="Fig_senate_spectral_2"}, respectively. Given an input seed set, Figures [\[Fig_senate_spectral_3\]](#Fig_senate_spectral_3){reference-type="ref" reference="Fig_senate_spectral_3"} and [\[Fig_senate_spectral_4\]](#Fig_senate_spectral_4){reference-type="ref" reference="Fig_senate_spectral_4"} illustrate the weakly and strongly local solutions by MOV and \(\ell_1\)-regularized Page-Rank, respectively. For MOV in Figures [\[Fig_senate_spectral_3\]](#Fig_senate_spectral_3){reference-type="ref" reference="Fig_senate_spectral_3"} we set \(z_i = 1\) for all \(i\) in the input seed set and \(z_i=0\) for all \(i\) outside the input seed set. Then we orthogonalize \(z\) with respect to \(D\cdot1_{|\mathcal{V}|}\). For \(\ell_1\)-regularized Page-Rank, we only give a single node as an input seed set, i.e., \(h_i=1\) where \(i\) is the input node and \(h_i=0\) for all other nodes. Moreover, we set the locality parameter \(\epsilon\) large enough such that the solution is very sparse, i.e., strongly local. In Figures [\[Fig_senate_spectral_3\]](#Fig_senate_spectral_3){reference-type="ref" reference="Fig_senate_spectral_3"} and [\[Fig_senate_spectral_4\]](#Fig_senate_spectral_4){reference-type="ref" reference="Fig_senate_spectral_4"}, we demonstrate the input seed sets by nodes with a blue halo around them. In Figure [\[Fig_senate_spectral_3\]](#Fig_senate_spectral_3){reference-type="ref" reference="Fig_senate_spectral_3"}, the cluster which is found by MOV consists of the nodes which have large mass concentration around the input seed set, i.e., the nodes around the input seed set that have large size and are coloured with a bright red shade. MOV recovers this cluster by examining the whole graph; each node has a weight assigned to it in Figure [\[Fig_senate_spectral_3\]](#Fig_senate_spectral_3){reference-type="ref" reference="Fig_senate_spectral_3"}. On the other hand, a similar cluster is found in Figure [\[Fig_senate_spectral_4\]](#Fig_senate_spectral_4){reference-type="ref" reference="Fig_senate_spectral_4"} by using \(\ell_1\)-regularized Page-Rank without examining the whole graph. This is possible because nodes of the graph have zero weight assigned and need not be considered. *This speed and data advantage, along with the sparsity-based implicit regularization , are some of the reasons that strongly-local algorithms, such as \(\ell_1\)-PageRank, are used so often in practice .* In Figure [\[Fig_grqc_spectral\]](#Fig_grqc_spectral){reference-type="ref" reference="Fig_grqc_spectral"}, we present global, weakly local, and strongly local solutions for the less well-partitionable and thus less easily-visualizable CA-GrQc graph. As already mentioned in the description of this data-set, this graph has many small clusters with small conductance ratio and large clusters have large ratio. This is also justified in our experiment by the fact that global methods, such as the Spectral relaxation and MOV global in Figures [\[Fig_grqc_spectral_1\]](#Fig_grqc_spectral_1){reference-type="ref" reference="Fig_grqc_spectral_1"} and [\[Fig_grqc_spectral_2\]](#Fig_grqc_spectral_2){reference-type="ref" reference="Fig_grqc_spectral_2"}, respectively, recover small clusters. The two global procedures find small clusters which are presented in Figures [\[Fig_grqc_spectral_1\]](#Fig_grqc_spectral_1){reference-type="ref" reference="Fig_grqc_spectral_1"} and [\[Fig_grqc_spectral_2\]](#Fig_grqc_spectral_2){reference-type="ref" reference="Fig_grqc_spectral_2"} with red, orange and yellow colours. However, since there are many small clusters of small conductance ratio, one might want to find different clusters than the ones obtained by Spectral relaxation and MOV global. This is possible using localized procedures such as MOV and \(\ell_1\)-regularized Page-Rank. Given two different seed sets we demonstrate in Figures [\[Fig_grqc_spectral_3\]](#Fig_grqc_spectral_3){reference-type="ref" reference="Fig_grqc_spectral_3"} and [\[Fig_grqc_spectral_4\]](#Fig_grqc_spectral_4){reference-type="ref" reference="Fig_grqc_spectral_4"} that MOV successfully finds other clusters than the ones obtained by the global methods. The same is shown in Figures [\[Fig_grqc_spectral_5\]](#Fig_grqc_spectral_5){reference-type="ref" reference="Fig_grqc_spectral_5"} and [\[Fig_grqc_spectral_6\]](#Fig_grqc_spectral_6){reference-type="ref" reference="Fig_grqc_spectral_6"} for \(\ell_1\)-regularized Page-Rank. Notice that MOV assigns weights (perhaps small) to all the nodes of the graph; on the other hand, \(\ell_1\)-regularized Page-Rank, as a strongly local procedure, assigns weights only to a small number of nodes, without examining all of the graph. We now use the FB-Johns55 graph which has an expander-like behaviour at all size scales, i.e., all small and large clusters have large conductance ratio. See Figure \(6\) in for details. We present the results of this experiment in Figure [\[Fig_fb_spectral\]](#Fig_fb_spectral){reference-type="ref" reference="Fig_fb_spectral"}. Notice that in Figures [\[Fig_fb_spectral_1\]](#Fig_fb_spectral_1){reference-type="ref" reference="Fig_fb_spectral_1"} and [\[Fig_fb_spectral_2\]](#Fig_fb_spectral_2){reference-type="ref" reference="Fig_fb_spectral_2"} the global methods identify three main clusters, one small (red nodes), one medium size (orange nodes) and one large (yellow nodes). All these clusters have similar conductance ratio. In Figures [\[Fig_fb_spectral_3\]](#Fig_fb_spectral_3){reference-type="ref" reference="Fig_fb_spectral_3"} and [\[Fig_fb_spectral_4\]](#Fig_fb_spectral_4){reference-type="ref" reference="Fig_fb_spectral_4"} we show that MOV can recover the medium or the small size clusters, respectively, by giving a localized seed set. In Figures [\[Fig_fb_spectral_5\]](#Fig_fb_spectral_5){reference-type="ref" reference="Fig_fb_spectral_5"} and [\[Fig_fb_spectral_6\]](#Fig_fb_spectral_6){reference-type="ref" reference="Fig_fb_spectral_6"} we illustrate that using \(\ell_1\)-regularized Page-Rank one can find very similar clusters while exploiting the strongly local running time of the method. Let us now present the performance of flow-based algorithms on the same graphs. We begin with US-Senate in Figure [\[Fig_senate_flow\]](#Fig_senate_flow){reference-type="ref" reference="Fig_senate_flow"}. In this figure, the red nodes are part of the solution of Flow Improve or Local Flow Improve, depending on the experiment; the yellow nodes are part of the seed set only; and the orange nodes are in both the solution of the flow algorithm and the input seed set. In Figure [\[Fig_senate_flow_1\]](#Fig_senate_flow_1){reference-type="ref" reference="Fig_senate_flow_1"}, we used as an input seed set to Flow Improve the cluster obtained by applying sweep cut with respect to the conductance ratio on the Spectral relaxation solution. Figures [\[Fig_senate_flow_2\]](#Fig_senate_flow_2){reference-type="ref" reference="Fig_senate_flow_2"} and [\[Fig_senate_flow_3\]](#Fig_senate_flow_3){reference-type="ref" reference="Fig_senate_flow_3"} present a clear distinction between Flow Improve and Local Flow Improve, weakly and strongly local algorithms, respectively. For both figures, the input seed set is located at about the middle of the graph. Flow Improve as a weakly local algorithm examines the whole graph and returns a cluster which includes the period before \(1913\). Also, it includes big part of the input seed set in the cluster due to the overlapping regularization term in the denominator of its objective function. See the definition of the objective function \(\phi_R\) for Flow Improve in Section [12](#sec:localgraphpart){reference-type="ref" reference="sec:localgraphpart"}. On the other hand, in Figure [\[Fig_senate_flow_3\]](#Fig_senate_flow_3){reference-type="ref" reference="Fig_senate_flow_3"} Local Flow Improve as a strongly local algorithm does not examine the whole graph and its solution is concentrated only around the input seed set. The distinction that we discussed in the previous paragraph between Flow Improve and Local Flow Improve is easy to visualize in the relatively well-structured US-Senate, but it is not so clear in all graphs. For example, in Figure [\[Fig_cagrqc_flow\]](#Fig_cagrqc_flow){reference-type="ref" reference="Fig_cagrqc_flow"} we present the performance of these two algorithms for the CA-GrQc graph. Since this graph has only small clusters of small conductance ratio, Flow Improve and Local Flow Improve find the same clusters. This is clearly shown by comparing Figures [\[Fig_cagrqc_flow_1\]](#Fig_cagrqc_flow_1){reference-type="ref" reference="Fig_cagrqc_flow_1"} and [\[Fig_cagrqc_flow_2\]](#Fig_cagrqc_flow_2){reference-type="ref" reference="Fig_cagrqc_flow_2"} and Figures [\[Fig_cagrqc_flow_3\]](#Fig_cagrqc_flow_3){reference-type="ref" reference="Fig_cagrqc_flow_3"} and [\[Fig_cagrqc_flow_4\]](#Fig_cagrqc_flow_4){reference-type="ref" reference="Fig_cagrqc_flow_4"}. A similar performance is observed for the FB-Johns55 graph in Figure [\[Fig_fb_flow\]](#Fig_fb_flow){reference-type="ref" reference="Fig_fb_flow"}, except that the solution of Flow Improve and Local Flow Improve are not exactly the same but only very similar. #### Flow vs. spectral, or \(\ell_1\) vs. \(\ell_2\) Spectral algorithms measure distances of the nodes based on the \(\ell_2\) norm. Generally this means that the nodes of the graph are embedded on the real line. On the other hand, flow algorithms measure distances of the nodes based on the \(\ell_1\) norm. The solution to flow-based algorithms that we discussed is binary, either a node is added in the solution with weight \(1\) or it is not and it has weight \(0\). In this case, the objective function \(\|Bx\|_{1,C}\) of the flow algorithms is a locally-biased variation on \(\cut(S)\), where \(S\) is constructed based on the binary \(x\). Therefore, the flow algorithms aim to find a balance between finding good cuts and identifying the input seed set. This implies that the flow algorithms try to minimize the absolute number of edges that cross the partition, but at the same time they try to take into account the volume regularization effect of the denominator in the objective function. In this section, we will try to isolate the effect of \(\ell_1\) and \(\ell_2\) metrics in the output solution. We do this by employing MQI and spectral MQI, which are flow (i.e., \(\ell_1\)) and spectral (i.e., \(\ell_2\)) problems, respectively. The first set of results is shown in Figure [\[Fig_senate_l1vl2\]](#Fig_senate_l1vl2){reference-type="ref" reference="Fig_senate_l1vl2"}. Notice in Figure [\[Fig_senate_l1vl2_1\]](#Fig_senate_l1vl2_1){reference-type="ref" reference="Fig_senate_l1vl2_1"} and [\[Fig_senate_l1vl2_2\]](#Fig_senate_l1vl2_2){reference-type="ref" reference="Fig_senate_l1vl2_2"} that MQI and spectral MQI + sweep cut recover the large clusters, i.e., before and after the year \(1913\). There are only minor differences between the two solutions. Moreover, observe that Spectral MQI returns a solution which is not binary. This is illustrated in Figure [\[Fig_senate_l1vl2_3\]](#Fig_senate_l1vl2_3){reference-type="ref" reference="Fig_senate_l1vl2_3"}, where the weights of the nodes are real numbers. Then sweep cut is applied on the solution of spectral MQI to obtain a binary solution with small conductance ratio, i.e., Figure [\[Fig_senate_l1vl2_2\]](#Fig_senate_l1vl2_2){reference-type="ref" reference="Fig_senate_l1vl2_2"}. The previous example did not reveal any difference between MQI and spectral MQI other than the fact that spectral MQI has to be combined with the sweep cut rounding procedure to obtain a binary solution. In Figure [\[Fig_usroads_l1vl2\]](#Fig_usroads_l1vl2){reference-type="ref" reference="Fig_usroads_l1vl2"}, we present a result showing where the solutions have substantial differences. The graph that we used for this is the US-Roads, and the input seed set consists of nodes near Minneapolis together with some suburban areas around the city. Notice in Figures [\[Fig_usroads_l1vl2_1\]](#Fig_usroads_l1vl2_1){reference-type="ref" reference="Fig_usroads_l1vl2_1"} that MQI, i.e., \(\ell_1\) metric, shrinks the boundaries of the input seed set. However, MQI does not accurately recover Minneapolis. The reason is the volume regularization which is imposed by the denominator of the objective function of MQI. This regularization forces the solution to have large volume. On the other hand, spectral MQI + sweep cut in Figure [\[Fig_usroads_l1vl2_2\]](#Fig_usroads_l1vl2_2){reference-type="ref" reference="Fig_usroads_l1vl2_2"} recovers Minneapolis. The reason is that for spectral MQI the regularization effect of the denominator is unimportant since the objective function is invariant to scalar multiplications of the solution vector. It's the solution of spectral MQI, i.e., the eigenvector of smallest eigenvalue, which is presented in Figure [\[Fig_usroads_l1vl2_3\]](#Fig_usroads_l1vl2_3){reference-type="ref" reference="Fig_usroads_l1vl2_3"}, that is concentrated closely around Minneapolis. Due to this concentration of the eigenvector around Minneapolis, the sweep is successful. Briefly, spectral MQI, which is a continuous relaxation of MQI, implicitly offers an additional level of volume regularization, which turns out to be useful in this example. Finally, we present another set of results using the FB-Johns55 graph in Figure [\[Fig_fb_l1vl2\]](#Fig_fb_l1vl2){reference-type="ref" reference="Fig_fb_l1vl2"}. As we saw before, notice that for this less well-structured graph the solutions of MQI and spectral MQI + sweep cut are nearly the same. This happens because the regularization effect of the denominator of MQI and the regularization imposed by spectral MQI have nearly the same effect on this graph. This is also justified by the fact that MQI in Figure [\[Fig_fb_l1vl2_1\]](#Fig_fb_l1vl2_1){reference-type="ref" reference="Fig_fb_l1vl2_1"} and spectral MQI without sweep cut in Figure [\[Fig_fb_l1vl2_3\]](#Fig_fb_l1vl2_3){reference-type="ref" reference="Fig_fb_l1vl2_3"} recover nearly the same cluster. # Conductance partitioning # Sparsest-Cut, Minimum-Conductance, and Spectral Relaxations {#sec:global_cond_spcut} In this section, we present two ubiquitous combinatorial optimization problems: Sparsest-Cut and Minimum-Conductance. These problems are NP-hard, but they can be relaxed to tractable convex optimization problems . We discuss one of the commonly used relaxation techniques which will motivate part of our discussion for local graph clustering algorithms. Both Sparsest-Cut and Minimum-Conductance give different ways of balancing the size of a partition with its quality. Sparsest-Cut finds the partition that minimizes the ratio of the fraction of edges that are removed divided by the scaled product of volumes of the two disjoint sets of nodes defined by removing those edges. In particular, if we have a partition \((S, S^c)\), where \(S^c\) is defined in Section [13](#sec:prelim){reference-type="ref" reference="sec:prelim"}, then \(\cut(S)\) is the number of edges that are removed, and the scaled product of volumes \(\vol(S)\vol(S^c)/\vol(\mathcal{V})\) is the volume of the disjoint sets of nodes. Putting all together in an optimization problem, we obtain \[\begin{aligned} \label{spcut} \tilde{\phi}(\mathcal{G}):= \mbox{minimize} & \ \tilde{\phi}(S):= \frac{\cut(S)}{\frac{1}{\vol(\mathcal{V})}\vol(S)\vol(S^c)} \\\nonumber \mbox{subject to} &\ S\subset \mathcal{V}. \end{aligned}\] We shall use the term *expansion of a set* to refer to the ratio \(\tilde{\phi}(S)\), and the term *expansion of the graph* to refer to \(\tilde{\phi}(\mathcal{G})\), in which case this problem is known as the Sparsest-Cut problem. Another way of balancing the partition is \[\begin{aligned} \label{isop} \phi(\mathcal{G}):= \mbox{minimize} & \ \phi(S):= \frac{\cut(S)}{\min(\vol(S),\vol(S^c))} \\\nonumber \mbox{subject to} &\ S\subset \mathcal{V}. \end{aligned}\] In this case, we divide by the minimum of \(\vol(S)\) and \(\vol(S^c)\), rather than their product. We shall use the term *conductance of a set* to refer to the ratio of cut to volume \(\phi(S)\), and the term *conductance of the graph* to refer to \(\phi(\mathcal{G})\), in which case this problem is known as the Minimum-Conductance problem. The difference between the two problems [\[spcut\]](#spcut){reference-type="eqref" reference="spcut"} and [\[isop\]](#isop){reference-type="eqref" reference="isop"} is that the former regularizes based on the number of connections lost among pairs of nodes, while the latter regularizes based on the size of the small side of the partition. Optimal solutions to these problems differ by a factor of \(2\): \[\label{eq:equiv_cond_sparsecut} \frac{1}{2}\tilde{\phi}(S) \le \phi(S) \le \tilde{\phi}(S),\] leading to the two objectives \(\tilde{\phi}(S)\) and \(\phi(S)\) being almost substitutable from a theoretical computer science perspective . However, this does not mean that the actual obtained solutions by solving [\[spcut\]](#spcut){reference-type="eqref" reference="spcut"} and [\[isop\]](#isop){reference-type="eqref" reference="isop"} are similar; in general, they are not. There are three major relaxation techniques for the NP-hard problems [\[spcut\]](#spcut){reference-type="eqref" reference="spcut"} and [\[isop\]](#isop){reference-type="eqref" reference="isop"}: spectral relaxation; all pairs multi-commodity flow or linear programming (LP) relaxation; and semidefinite programming (SDP) relaxation. For detailed descriptions about the LP and SDP relaxations we refer the reader to and, respectively. We focus here on spectral relaxation since similar relaxations are widely used for the development of local clustering methods, which we discuss in subsequent sections. ## Spectral Relaxation {#subsec:spcut} Spectral graph partitioning is one of the best known relaxations of the Sparsest-Cut [\[spcut\]](#spcut){reference-type="eqref" reference="spcut"} and Minimum-Conductance [\[isop\]](#isop){reference-type="eqref" reference="isop"}. The relaxation is the same for both problems, although the diversity of derivations of spectral partitioning does not always make this connection clear. For a partition \((S, S^c)\), let's associate a vector \(x\in\{c_1,c_2\}^{|\mathcal{V}|}\) such that \(x_i=c_1\) if \(v_i\in S\) and \(x_i=c_2\) if \(v_i\in S^c\). (For simplicity, think of \(c_1 = 1\) and \(c_2 = 0\), so \(x\) is the set indicator vector.) The spectral clustering relaxation uses a continuous relaxation of the set indicator vector in problems [\[spcut\]](#spcut){reference-type="eqref" reference="spcut"} and [\[isop\]](#isop){reference-type="eqref" reference="isop"} to produce an eigenvector. The relaxed problem is \[\begin{aligned} \label{spcutLD} \lambda_2 := \mbox{minimize} & \ \frac{\|Bx\|_{2,C}^2}{2\|x\|_{2,D}^2} \\\nonumber \mbox{subject to} &\ 1_{|\mathcal{V}|}^T D x = 0\\\nonumber &\ x\in\mathbb{R}^{|\mathcal{V}|}-\{0_{|\mathcal{V}|}\}. \end{aligned}\] (The denominator \(\|x\|_{2,D}^2 = \sum_i |x_i|^2 d_i\).) To see why [\[spcutLD\]](#spcutLD){reference-type="eqref" reference="spcutLD"} is a continuous relaxation of [\[spcut\]](#spcut){reference-type="eqref" reference="spcut"} we make two observations. First, notice that for all \(x\) in \(\mathbb{R}^{|\mathcal{V}|}-\{0_{|\mathcal{V}|}\}\) such that \(1_{|\mathcal{V}|}^T D x = 0\) the denominator in [\[spcutLD\]](#spcutLD){reference-type="eqref" reference="spcutLD"} satisfies \(2\vol(\mathcal{V})\|x\|_{2,D}^2 = \sum_{i=1}^{|\mathcal{V}|}\sum_{j=1}^{|\mathcal{V}|} d_id_j|x_i-x_j|^2\). Therefore, \(\lambda_2\) in [\[spcutLD\]](#spcutLD){reference-type="eqref" reference="spcutLD"} is equivalent to \[\begin{aligned} \label{eq:1434} \lambda_2 = \mbox{minimize} & \ \frac{\|Bx\|_{2,C}^2}{\frac{1}{\vol(\mathcal{V})}\sum_{i=1}^{|\mathcal{V}|}\sum_{j=1}^{|\mathcal{V}|} d_id_j|x_i-x_j|^2} \\\nonumber \mbox{subject to} &\ 1_{|\mathcal{V}|}^T D x = 0\\\nonumber &\ x\in\mathbb{R}^{|\mathcal{V}|}-\{0_{|\mathcal{V}|}\}. \end{aligned}\] The optimal value of the right hand side in [\[eq:1434\]](#eq:1434){reference-type="eqref" reference="eq:1434"} is equivalent to the optimal value of right hand side in the following expression \[\begin{aligned} \label{eq:1435} \lambda_2 = \mbox{minimize} & \ \frac{\|Bx\|_{2,C}^2}{\frac{1}{\vol(\mathcal{V})}\sum_{i=1}^{|\mathcal{V}|}\sum_{j=1}^{|\mathcal{V}|} d_id_j|x_i-x_j|^2} \\\nonumber \mbox{subject to} & \ x\in\mathbb{R}^{|\mathcal{V}|}-\{0_{|\mathcal{V}|},1_{|\mathcal{V}|}\}. \end{aligned}\] To prove this notice that the objective function of the right hand side in [\[eq:1435\]](#eq:1435){reference-type="eqref" reference="eq:1435"} is invariant to constant shifts of \(x\), i.e., \(x\) and \(x + c 1_{|\mathcal{V}|}\) have the same objective function, where \(c\) is a constant. Therefore, if \(\tilde{x}\) is an optimal solution of the right hand side in [\[eq:1435\]](#eq:1435){reference-type="eqref" reference="eq:1435"} then \(\hat{x} = \tilde{x}-\frac{1^T_{|\mathcal{V}|}D\tilde{x}}{\vol(\mathcal{V})} 1_{|\mathcal{V}|}\) has the same optimal objective value and also \(1_{|\mathcal{V}|}^TD \hat{x} = 0\). Second, by restricting the solution in [\[eq:1435\]](#eq:1435){reference-type="eqref" reference="eq:1435"} in \(\{0,1\}^{|\mathcal{V}|}\) instead of \(\mathbb{R}^{|\mathcal{V}|}\) we get that \(\cut(S) = \|Bx\|_{2,C}^2\) and \(\sum_{i=1}^{|\mathcal{V}|}\sum_{j=1}^{|\mathcal{V}|} d_id_j|x_i-x_j|^2 = \vol(S)\vol(S^c)\). Using these two observations, it is easy to see that [\[spcutLD\]](#spcutLD){reference-type="eqref" reference="spcutLD"} is a continuous relaxation of [\[spcut\]](#spcut){reference-type="eqref" reference="spcut"}. Using [\[eq:equiv_cond_sparsecut\]](#eq:equiv_cond_sparsecut){reference-type="eqref" reference="eq:equiv_cond_sparsecut"}, it is easy to see that [\[spcutLD\]](#spcutLD){reference-type="eqref" reference="spcutLD"} is a relaxation for [\[isop\]](#isop){reference-type="eqref" reference="isop"} as well. The quality of approximation of relaxation [\[spcutLD\]](#spcutLD){reference-type="eqref" reference="spcutLD"} to Sparsest-Cut [\[spcut\]](#spcut){reference-type="eqref" reference="spcut"} is given by Cheeger's inequality \[{\lambda_2}/{\vol(\mathcal{V})}\le \tilde{\phi}(\mathcal{G}) \le {(8 \lambda_2)^{1/2}}/{\vol(\mathcal{V})}\] while the approximation guarantee for the Minimum-Conductance problem [\[isop\]](#isop){reference-type="eqref" reference="isop"} is \[{\lambda_2}/{2}\le \phi(\mathcal{G}) \le (2 \lambda_2)^{1/2},\] which can be found in. (A generalization of these bounds holds for arbitrary vectors .) Both of these approximation ratios can be realized by rounding procedures described below. Another form of relaxation is the combinatorial model relaxation, which is formulation as problem [\[spcutLD\]](#spcutLD){reference-type="eqref" reference="spcutLD"} by ignoring the orthogonality constraint and restricting \(x\in\{0,1\}^{|\mathcal{V}|}\) instead of \(x\in\mathbb{R}^{|\mathcal{V}|}\). An extensive study of the spectral and the combinatorial mode relaxation can be found in, while empirical comparisons between these relaxations are discussed in. ## Rounding {#subsec:rounding} In practice, the solution obtained by the spectral relaxation is unlikely to lie in \(\{0,1\}^{|\mathcal{V}|}\), i.e., it is unlikely to be the indicator vector of a set. Therefore, it is necessary to have an efficient post-processing procedure where the solution is rounded to a set. At the same time it is important to guarantee that the rounded solution has good worst-case guarantees in terms of the conductance or sparsest cut objective. One of the most efficient and theoretically justified rounding procedures for spectral relaxation is the *sweep cut*. The sweep cut procedure is summarized in the following steps. Notice that sweep cut can be used to obtain approximate solutions for both Sparsest-Cut and Minimum-Conductance. In fact, the proof for the upper inequalities of the approximation guarantees of spectral relaxation to Sparsest-Cut and Minimum-Conductance are obtained by using the sweep cut procedure. # Regularization, \(\ell_2\) versus \(\ell_1\) In the previous section we followed a theoretical computer science perspective. We discussed tractable combinatorial problems, i.e., min-cut and \(s\)-\(t\) min-cut, and NP-hard combinatorial problems, i.e., minimum conductance and isoperimetry. We discussed how to relax the latter NP-hard problems in order to obtain continuous versions that are "good\" approximations in terms of worst-case objective value. In this section we make a step back and study the regularization properties that are imposed by min-cut and certain spectral problems. Understanding the regularization properties that are imposed by each problem on the solution will allow us to understand the structure of the output solution in practice. In subsequent sections we will use the regularization problems by min-cut and spectral to understand what properties are imposed on the solution by current state-of-the-art local algorithms. #### Properties of \(s\)-\(t\) Min-Cut We now discuss regularization properties for the \(s\)-\(t\) min-cut problem. Similar properties hold for min-cut since this is the best \(s\)-\(t\) min-cut over all pairs of nodes in \(\mathcal{V}\). It can be shown that the following optimization problem \[\begin{aligned} \label{eq:mincutL1} \nonumber \mbox{minimize} & \quad \|Bx\|_{1,C}\\ \mbox{subject to} & \quad x_s =1, \ x_t = 0 \\\nonumber &\quad x\ge 0. \end{aligned}\] is equivalent to \(s\)-\(t\) min-cut, which we introduced in Section [\[subsec:mincutmaxflow\]](#subsec:mincutmaxflow){reference-type="ref" reference="subsec:mincutmaxflow"}. The solution to \(s\)-\(t\) min-cut is binary and lies in \([0,1]^{|\mathcal{V}|}\), see. Thus, the solution defines a partition \(S, S^c\). Since we assumed that the given graph has edges with unit-weights then the \(s\)-\(t\) min-cut objective penalizes partitions \(S,S^c\) that have many outgoing edges from \(S\) to \(S^c\). Intuitively, this implies that min-cut looks for partitions which have clear boundaries. For many applications, depending on the given graph such a property might be very important. Later on we will see that there are local algorithms which preserve this property close to a given input seed set. #### Properties of spectral Let us now replace the \(\ell_1\) norm in [\[eq:mincutL1\]](#eq:mincutL1){reference-type="eqref" reference="eq:mincutL1"} with the \(\ell_2\) norm to obtain the following problem \[\begin{aligned} \label{eq:spectralL2} \nonumber \mbox{minimize} & \quad \|Bx\|_{2,C}\\ \mbox{subject to} & \quad x_s =1, \ x_t = 0 \\\nonumber &\quad x\ge 0. \end{aligned}\] From the preliminaries we know that \(\|Bx\|_{2,C} = x^T L x\). This problem is related to spectral relaxation [\[spcutLD\]](#spcutLD){reference-type="eqref" reference="spcutLD"}. The only difference between [\[eq:spectralL2\]](#eq:spectralL2){reference-type="eqref" reference="eq:spectralL2"} and [\[spcutLD\]](#spcutLD){reference-type="eqref" reference="spcutLD"} is that we do not scale the \(\ell_2\) norm in the constraint with the diagonal matrix \(D\). This means that we do not put weights on nodes based on their degree, i.e., all nodes are of the same importance. Similarly to our discussion on spectral relaxation it can be shown that problem [\[eq:spectralL2\]](#eq:spectralL2){reference-type="eqref" reference="eq:spectralL2"} is a relaxation of the Minimum-Conductance and Minimum-Isoperimetry problems where all nodes are of the same importance. However, in this section our focus is the properties imposed by the \(\ell_2\) norm and not the relaxation properties of problem [\[eq:spectralL2\]](#eq:spectralL2){reference-type="eqref" reference="eq:spectralL2"}. The solution to [\[eq:spectralL2\]](#eq:spectralL2){reference-type="eqref" reference="eq:spectralL2"} satisfies \[L x = \lambda_2 x,\] where \(\lambda_2\) is the second smallest eigenvalue of \(L\) for a unit-length eigenvector \(x\). Since \(L\) is the Laplacian operator of the given graph, then the eigenvector which satisfies the previous equation can be seen as the solution to a diffusion process. This means that the solution to problem [\[eq:spectralL2\]](#eq:spectralL2){reference-type="eqref" reference="eq:spectralL2"} is generally "smooth\", meaning that the values of the components of \(x\) are in \(\mathbb{R}\) and they are not binary. This is an important difference between [\[eq:mincutL1\]](#eq:mincutL1){reference-type="eqref" reference="eq:mincutL1"} and [\[eq:spectralL2\]](#eq:spectralL2){reference-type="eqref" reference="eq:spectralL2"}, which is due to the geometry that \(\ell_1\) and \(\ell_2\) norms define. Intuitively, the "smoothness\" of the solution given by spectral relaxation implies that we loose the property of [\[eq:mincutL1\]](#eq:mincutL1){reference-type="eqref" reference="eq:mincutL1"}, which is to preserve boundaries. If one takes a clustering perspective, then this drawback can be fixed by combining spectral with some sweep procedure which will possibly ignore components of \(x\) with very small values, i.e., nodes outside the boundary. This two step-procedure is provably at most as good as [\[eq:mincutL1\]](#eq:mincutL1){reference-type="eqref" reference="eq:mincutL1"} if our objective is to find the best partition with the fewest outgoing edges. However, one should not conclude that using \(\ell_1\) norm than \(\ell_2\) defines a better geometry. The decision between \(\ell_1\) and \(\ell_2\) is application depended. If one wants to preserve boundaries then they should use the \(\ell_1\) norm. If one wants to perform a diffusion on a given graph, then one should use the \(\ell_2\) norm. # Locally-biased graph partitioning methods {#sec:localgraphpart} All of the algorithms described in Section [10](#sec:global_cond_spcut){reference-type="ref" reference="sec:global_cond_spcut"} are "global," in that they touch all of the nodes of the input graph at least once, and thus they have a running time that is at least linear in the size of the input graph. Informally, locally-biased graph clustering algorithms find clusters near a specified seed set of vertices, in many cases without even touching the entire input graph. In this section, we will describe several local graph clustering algorithms, each of which has somewhat different properties. To understand the seemingly-quite-different algorithms we will discuss, we will distinguish local graph clustering algorithms based on three major features. The local graph clustering algorithms that we consider in the following sections and their basic properties with respect to the above three features are given in Table [1](#table1){reference-type="ref" reference="table1"}. ## A general localizing construction We describe these locally-biased graph methods in terms of an augmented graph we call the *reference cut graph*. We should emphasize that this is a conceptual construction to highlight the similarities and differences between locally-biased algorithms in Table [1](#table1){reference-type="ref" reference="table1"}---in particular, these algorithms do *not* explicitly construct the reference cut graph. Let \(h,g\in\mathbb{R}^{|\mathcal{V}|}\), \(h,g\ge 0\), and \(g-h \ge 0\), and let \(\alpha\), \(\beta\), and \(\gamma\) be parameters specified below.. Then the reference cut graph is constructed from a simple, undirected graph \(\mathcal{G}\) as follows: Let \(H := diag(h)\), \(G := diag(g)\) and \(Z = G-H\). Then we can also view the augmented graph through its incidence matrix and weight matrix: \[\tilde{B} = \begin{bmatrix} 1_{|\mathcal{V}|} &-I_{|\mathcal{V}|} & 0 \\ 0 & B & 0 \\ 0 & I_{|\mathcal{V}|} &-1_{|\mathcal{V}|} \end{bmatrix}, \ \tilde{C}= \begin{bmatrix} \alpha H & 0 & 0 \\ 0 & \gamma C & 0 \\ 0 & 0 & \beta Z \end{bmatrix},\] respectively. The above construction might look overly complicated. However, we will see in the following subsections that it simplifies for local spectral and flow graph clustering algorithms with specific settings for \(h,g\) and \(\gamma\). ## Weakly-Local and Strongly-Local flow methods Although finding the set of minimum conductance is NP-hard in general, there are a number of cases and variations that admit polynomial time algorithms and can be solved via Max-Flow/Min-Cut or a parametric Max-Flow method. These algorithms begin with a reference set \(R\) of nodes, and they return a smaller (or not much larger) set of nodes that is a better partition in terms of the conductance ratio \(\phi\). Typically, the returned value is *optimal* for a variation on the Minimum-Conductance and/or the Sparsest-Cut objective. The methods themselves are highly flexible and apply to other variations of Minimum-Conductance and Sparsest-Cut. For the sake of simplicity, we will describe them for conductance. All of the following procedures adopt the following meta-algorithm starting with working set \(W\) initialized to an input reference set of nodes \(R\), values \(\alpha_1, \beta_1, \gamma_1\) and vectors \(h = d_R\), \(g=d\), where \(d_R\) is a vector of length \(|\mathcal{V}|\) with components equal to \(d_i\)'s for nodes \(v_i\in R\) and zeros for nodes \(v_i\in R^c\). Figure [\[fig:cut-graph\]](#fig:cut-graph){reference-type="ref" reference="fig:cut-graph"} illustrates the construction of an augmented graph based on the previous setting of \(\alpha,\beta,\gamma\) and \(h,g\). **The Local-Flow Meta-algorithm** Next, we describe several procedures that are instantiations of this basic Local-Flow Meta-algorithm. #### MQI The first algorithm we consider is the MQI procedure due to Lang and Rao . This method is designed to take the reference set \(R\) with \(\vol(R) \le \vol(G)/2\) and identify a subset of it \(S \subseteq R\). The method instantiates the Local-Flow Meta-algorithm using \(\alpha_k = \cut(W_k), \gamma_k = \vol(W_k), \beta_k = \infty\) and \(h=d_R\), \(g=d\). The idea with this method is that the reference cut graph will have an \(s,t\)-Min-Cut value strictly less than \(\alpha_k \gamma_k\) if and only if there is a strict subset \(S \subset W_k\) that has conductances less than \(\alpha_k / \gamma_k\). (See  for the proof.) If there is such a set, then the result \(W_{k+1}\) will be a set with conductance less than \(\alpha_k / \gamma_k\). Since \(\alpha_k\) and \(\gamma_k\) are picked based on the current working set \(W_k\), at each step the algorithm monotonically improves the conductance. Also, each step minimizes the objective \[\begin{aligned} \mbox{minimize} & \; \| \tilde{B}x \|_{1,\tilde{C}_k} \label{eq:mqi-l1} \\ \mbox{subject to} & \; x_i = 0 \ \forall v_i\in R^c, x_s = 1, x_t = 0. \end{aligned}\] In fact, when the algorithm terminates, that means that there is no subset of \(R\) with conductance less than \(\alpha_k / \gamma_k\). Hence, we have solved the following variation on the conductance problem \[\begin{aligned} \mbox{minimize} & \; \phi(S)= \frac{\cut(S)}{\vol(S)} \\ \mbox{subject to} & \; S \subseteq R. \end{aligned}\] The key difference from the standard conductance problem is that we have restricted ourselves to a *subset* of the reference set \(R\). Procedurally, this is guaranteed because the edges connecting \(R^c\) to \(t\) have weight infinity, so they will never be cut. Thus, operationally, MQI is always a strongly local algorithm since it only operates within the input seed set \(R\). Nodes connected to \(t\) with weight infinity can be agglomerated or merged into a mega-sink node \(T\). The resulting graph has the same size as \(R\) along with the source and sink. (This is how the MQI construction is described in the original paper.) The MQI problem can be solved using Max-Flow method on the resulting graph a logarithmic number of times. Therefore, the running time for solving MQI depends on the Max-Flow algorithm that is used. Details about running times of Max-Flow algorithms can be found in . #### Flow-Improve The Flow-Improve method due to Andersen and Lang  was inspired by MQI and designed to address the weakness that the algorithm will always find an output set within the reference set \(R\), i.e., that is a subset of \(R\). (As an illustration, see Figure [\[fig:mqi-vs-flow-improve\]](#fig:mqi-vs-flow-improve){reference-type="ref" reference="fig:mqi-vs-flow-improve"}.) Again, Flow-Improve takes as input a reference set \(R\) with volume less than half the graph. The idea behind Flow-Improve is that we want to find a set with conductance at least as good as \(R\) and that also is highly correlated with \(R\). To do this, consider the following variant of conductance \[\phi_R(S) = \frac{\cut(S)}{\vol(S \cap R)-\theta \vol(S \cap R^c)}\] where \(\theta = \vol(R)/\vol(R^c)\), and where the value is \(\infty\) if the denominator is negative. For any set \(S\), \(\phi_R(S) \ge \phi(S)\). Thus, this modified conductance score is an upper-bound on the true conductance. Again, we are able to show that the Local-Flow Meta-algorithm can solve for the exact value of \(\phi_R(S)\) in polynomial time. To do so, instantiate that algorithm with \(\alpha_k = \phi_R(W_k)\), \(\beta_k = \theta \phi_R(W_k)\), \(\gamma_k = 1\) and \(h=d_R\), \(g=d\). The value of a cut on set \(S\) in the resulting graph is \[\cut(S) + \alpha_k \vol(R)-\alpha_k [\vol(S \cap R)-\theta \vol(S \cap R^c)].\] (See  for the justification.) Andersen and Lang show that the algorithm monotonically reduces \(\phi_R(W_{k})\) at each iteration as well. Each iteration now solves the \(s,t\)-Min-Cut problem \[\begin{aligned} \mbox{minimize} & \; \| \tilde{B}x \|_{1,\tilde{C}_k} \label{eq:FlowImprove} \\ \mbox{subject to} & \; x_s = 1, x_t = 0. \nonumber \end{aligned}\] In order to match precisely their Flow-Improve procedure, we would need to modify our Meta-algorithm to check the value of \(\phi_R(W_k)\), instead of conductance (at Step [\[step:cond-or-variant\]](#step:cond-or-variant){reference-type="ref" reference="step:cond-or-variant"} of the Local-Flow Meta-algorithm above), for monotonic decrease. The authors also show that this procedure terminates in a finite number of iterations. At termination, the Flow-Improve algorithm has exactly solved \(\mbox{minimize} \; \phi_R(S), S \subseteq V\). This can be considered a locally-biased variation of the conductance objective, where we penalize departure from the reference set \(R\). Consequently, the solutions will tend to identify small conductance sets nearby \(R\). Flow-Improve is a very useful algorithm, but it has two small weaknesses. The first is that it is a weakly local algorithm. At each step, we have to solve a Min-Cut problem that is the size of the original graph. The second is that the Min-Cut problems do not have integer weights. (Note that \(\theta\) will almost never be an integer.) Most fast Max-Flow/Min-Cut procedures and implementations assume integer weights. For instance, many implementations of the push-relabel method (hipr ) only allows integer weights. Boykov and Kolmogorov's solver is a notable exception . Similarly to MQI, the running time of solving the Max-Flow/Min-Cut problem depends on the particular solver that is used. A summary of Max-Flow/Min-Cut methods can be found in. #### Local-Flow-Improve The Local-Flow-Improve algorithm due to Orecchia and Zhu  sought to address the weak-locality of the Flow-Improve method and create a strongly local flow based method. This involved two key innovations: a modification to the construction and objective that enables strong locality; and an algorithm to realize that strong locality. This Local-Flow-Improve method essentially interpolates between MQI and Flow-Improve. In one limit, it is strictly local to the reference graph and exactly reproduces the MQI output. In the other limit, it is exactly Flow-Improve. To do this, Orecchia and Zhu alter the objective function used for Flow-Improve to place an additional penalty on deviating from the set \(R\). They describe this as increasing the weight of connections \(\beta_k\) in the reference cut graph by scaling these by a value \(\kappa \ge 1\). If \(\kappa = 1\), then their construction is exactly that of Flow-Improve. If \(\kappa = \infty\), then this construction is equivalent to that of MQI. The effect of \(\kappa\) is illustrated in Figure [\[fig:local-flow-improve\]](#fig:local-flow-improve){reference-type="ref" reference="fig:local-flow-improve"}. In terms of the optimization framework, their modification corresponds to using \[\phi_R'(S;\kappa) = \frac{\cut(S)}{\vol(S \cap R)-\theta \kappa \vol(S \cap R^c)} \ \] where \(\kappa \ge 1\) and \(\theta = \vol(R)/\vol(R^c)\) as in Flow-Improve, and again the value is \(\infty\) if the denominator is negative. This result corresponds to instantiating the Local-Flow Meta-algorithm using \(\alpha_k = \phi_R'(W;\kappa)\), \(\beta_k = \phi_R'(W;\kappa) \theta \kappa\) and \(h=d_R\), \(g=d\). The second innovation is that they describe an algorithm to solve the Min-Cut problem on the reference cut graph that does not need to explore the entire graph. This second piece used a novel modification of Dinic's procedure  to compute a Max-Flow/Min-Cut that exploited the locality. We refer interested readers back to Orecchia and Zhu for details of this second somewhat complicated construction. In our recent work , however, we describe a simplified framework for the Local-Flow-Improve method that shows that the strong locality in their modification results from implicitly regularizing the Flow-Improve objective with a \(\ell_1\) norm regularizer. (This will mirror strongly local spectral results in the forthcoming spectral section.) In fact, our recent work  shows that each iteration exactly solves \[\begin{aligned} \mbox{minimize} & \; \| \tilde{B} x\|_{1,\tilde{C}_k'} + \varepsilon \| D x \|_1 \label{eq:FlowImprove-l1} \\ \mbox{subject to} & \; x_s = 1, x_t = 0, \nonumber \end{aligned}\] where \(\tilde{C}_k'\) is a small perturbation on the above definition and \(\varepsilon\) is a locality parameter. The volume of the output cluster \(S\) of the method in is bounded \(\vol(S) \le \vol(R)(1+2/\epsilon) + E(R,R^c)\), where \(\epsilon:= \vol(R)/\vol(R^c) + \delta\) and \(\delta\ge 0\) is a constant. That work also describes a simple procedure to realize the strong locality that leverages any Max-Flow/Min-Cut solver on a sequence of sub-problems whose size is bounded independent of the graph. ## Weakly-and-Strongly local spectral methods There are spectral analogues for each of the three flow-based constructions on the augmented graph. The development of these ideas occurred in parallel, largely independently, and it was not obvious until recently that the ideas were very related. Here, we make these connections explicit. Of the three flow constructions, the simplest is the MQI objective. We begin with it. #### SpectralMQI The method we call SpectralMQI was proposed as the Dirichlet partitioning problem in . Given a graph \(\mathcal{G}\) and a subset of vertices, consider finding the set \(S\) of minimal local conductance \(\phi'(S) = \cut(S)/\vol(S)\) such that \(S \subseteq R\), where, again, \(R\) is a reference set specified in the input. Note that the only difference from conductance is that we don't have the minimum in the denominator. A spectral algorithm to find an approximate minimizer of this is to solve the generalized eigenvector problem \[\begin{aligned} \lambda_R = \mbox{minimize} & \quad \frac{\|Bx\|_{2,C}^2}{\|x\|_{2,D}^2} \\ \mbox{subject to} &\quad x_{i} = 0 \ \forall v_i \in R^c. \end{aligned}\] The solution vector \(x\) and value \(\lambda_R\) are related to the smallest eigenvalue of the sub-matrix of the normalized Laplacian corresponding to the nodes in \(R\). (Note that we take the sub-matrix of the normalized Laplacian, rather than the normalized Laplacian on the sub-graph induced by \(R\).) A sweep-cut over the eigenvector \(x\) produces a set \(S\) that satisfies a Cheeger inequality with respect to the best possible solution . This definition of local conductance is also called \(\texttt{NCut'}\) by Hochbaum , who gave a polynomial time algorithm to compute it that is closely related to the MQI procedure. In this case, if \(R\) has volume less than half the graph, then this is exactly the spectral analogue of the MQI procedure and the result is a Cheeger-like bound on the overall conductance. #### MOV The Mahoney-Orecchia-Vishnoi (MOV) objective  is a spectral analogue of the FlowImprove method, with a few subtle differences. The goal is to find a small Rayleigh quotient, as in [\[spcutLD\]](#spcutLD){reference-type="eqref" reference="spcutLD"}, that is highly correlated with an input vector \(z\in\mathbb{R}^{|V|}\), where \(z\) represents the seed or local-bias. Given this, the MOV objective is \[\begin{aligned} \mbox{minimize} & \ \frac{\|Bx\|_{2,C}^2}{\|x\|_{2,D}^2} \\\nonumber \mbox{subject to} &\ 1_{|\mathcal{V}|}^T D x = 0 \\\nonumber &\ (z^T D x)^2 \ge \kappa &\ x\in\mathbb{R}^{\mathcal{V}}. \end{aligned}\] The solution of this problem represents an embedding of the nodes in \(\mathcal{V}\) which is locally biased, i.e., large values for components/nodes that are considered important and small or zero values for the rest. According to, there is a constant \(\rho\), i.e., the optimal dual variable for the locally-biased constraint, such that the solution to the MOV problem satisfies \((L+\rho D) x = \rho D z\). The null space of \(L\) is the vector \(1_{|\mathcal{V}|}\), and assuming that \(1_{|\mathcal{V}|}Dz = 0\), then the solution to the previous system is unique. One final detail is that the MOV construction fixes \(\| x \|_2 = 1\). Consequently, the MOV solution is \[\label{eq:mov} x = c (L + \rho D)^\dagger D z \quad c = (\| (L + \rho D)^\dagger D z \|_2)^{-1}.\] In the MOV paper , they show that \(\rho\) can be chosen such that \(x^T D z = \kappa\), the desired correlation strength with the input vector \(z\), through a simple bisection procedure. Solving the linear system [\[eq:mov\]](#eq:mov){reference-type="eqref" reference="eq:mov"} results in a weakly-local method that satisfies another Cheeger-like inequality. Recent extensions show that it is possible to get multiple locally-biased vectors that are akin to eigenvectors from this setup . The methodology is able to leverage the large number of Laplacian system solvers that can find an approximate solution to [\[eq:mov\]](#eq:mov){reference-type="eqref" reference="eq:mov"} in nearly linear time. The pseudo-inverse allows us to "pass through" \(\rho = 0\) and approach \(\rho =-\lambda_2\). (This system is singular at \(\rho = 0\) and \(\rho = \lambda_2\).) What is interesting is that taking the limit \(\rho \to-\lambda_2\) directly maps to the spectral relaxation [\[spcutLD\]](#spcutLD){reference-type="eqref" reference="spcutLD"}. Thus, the \(\rho\) parameter interpolates between the global spectral relaxation [\[spcutLD\]](#spcutLD){reference-type="eqref" reference="spcutLD"} and a spectral-version of the Min-Cut problem in each step of FlowImprove. Based on the reference cut graph, the MOV objective is \(\mbox{minimize} \ \|\tilde{B}\tilde{x}\|_{2,\tilde{C}}^2\), where \(\tilde{x}: = [1; x; 0]\). The reference graph cut setting is \(\gamma=1\), \(g=d\), \(h=Dz\) and \(\alpha=\beta=\rho\ge0\) controls the desired strength of the correlation to the input vector \(z\). Notice that the MOV problem is a spectral, i.e., \(\ell_2\), version of the \(s,t\)-Min-Cut problem. This observation was made first in. If \(\rho\) is extremely large, the solution to the above problem would have perfect correlation with the input vector \(h\). As \(\rho \to 0\), we decrease the effective correlation with the input vector \(h\). (These arguments are formal in .) # Preliminaries and Notation {#sec:prelim} **Graph assumptions:** We use the letter \(\mathcal{G}\) to denote a given connected graph. We assume that \(\mathcal{G}\) is undirected with no self-loops. Many of the constructions we will use operate on *weighted graphs* and so we assume that each edge may have a positive capacity. Graphs that are unweighted should have all of their capacities set to \(1\). **Nodes, edges and cuts:** Let \(\mathcal{V} = \{v_1,v_2\dots,v_n\}\) be a given set of \(|\mathcal{V}|\) nodes of graph \(\mathcal{G}\). We denote with \(e_{ij}\) an edge in the graph between nodes \(v_i\) and \(v_j\). Let \(\mathcal{E}\) be a given set of \(|\mathcal{E}|\) edges of graph \(\mathcal{G}\). A subset \(S\subset\mathcal{V}\) of nodes can be used to define a partitioning of \(\mathcal{V}\) into \(S\) and \(S^c:=\mathcal{V} \backslash S\). We define a cut as subset \(E \subset \mathcal{E}\) which partitions the graph \(\mathcal{G}\) into two sets. Given a partition \(S\subset\mathcal{V}\) and \(S^c\), then \(E(S,S^c)=\{e_{ij}\in \mathcal{E} \ | \ v_i\in S \mbox{ and } v_j\in S^c\}\) is the set of edges with one side in \(S\) and the other side in \(S^c\). If the partition is clear from the context we write the cut set as \(E\) instead of \(E(S,S^c)\). Let \(c_{ij}\) be a weight of the edge \(e_{ij}\), then we define the cut \(S\) as \[\cut(S) := \cut(E(S,S^c)) := \sum_{e_{ij}\in E(S,S^c)} c_{ij}.\] The volume of a set \(S\) is \[\vol(S) := \sum_{v_i \in S} \sum_{e_{ij} \in \mathcal{E}} c_{ij}.\] For simplicity of notation, we will drop the input \(\mathcal{G}\) in \(\mathcal{V}\) and \(\mathcal{E}\) if it is clear from the context that we are referring to a single graph \(\mathcal{G}\). **Matrix notation:** We denote with \(A\in\mathbb{R}^{|\mathcal{V}|\times |\mathcal{V}|}\) the adjacency matrix for a given graph, where \(A_{ij} = c_{ij}\) \(\forall e_{ij}\in \mathcal{E}\) and zero elsewhere. Let \(d_i\) be the degree of node \(v_i\in\mathcal{V}\), \(D\in\mathbb{R}^{|\mathcal{V}|\times |\mathcal{V}|}\) be the degree diagonal matrix \(D_{ii} = d_i\), \(L=D-A\) be the graph Laplacian, and \(\mathcal{L} = D^{-1/2}LD^{-1/2}\) be the symmetric normalized graph Laplacian. Note that the volume of a subset \(S\) is \(\vol(S) = \sum_{v_i\in S} d_{i}\). We denote with \(B\in\mathbb{R}^{|\mathcal{E}|\times |\mathcal{V}|}\) the incidence matrix of the given graph \(\mathcal{G}\). Every row of the incidence matrix corresponds to an edge \(e_{ij}\in \mathcal{E}\) in \(\mathcal{G}\). Assuming arbitrary ordering of the edges of the graph, in this paper we define the rows of the incidence matrix as \(B_{e_{ij}} = e_i-e_j\) \(\forall e_{ij}\in \mathcal{E}\), where \(e_i\in\mathbb{R}^{|\mathcal{V}|}\) is equal to one at the \(i^{th}\) position and zero elsewhere. Finally, \(C\in\mathbb{R}^{|\mathcal{E}|\times|\mathcal{E}|}\) is a diagonal matrix of the weights of each edge, i.e., \(C_{ij} = c_{ij}\) \(\forall i,j\). In this notation, the Laplacian matrix \(L = B^T C B\). **Norms:** For all \(x \in \mathbb{R}^{|\mathcal{E}|}\), we define the weighted \(\ell_1\) and \(\ell_2\) norms \(\|x\|_{1,C}: = \sum_{e_{ij}\in E} c_{ij} |x_{ij}|\) and \(\|x\|_{2,C}^2: = \sum_{e_{ij}\in E} c_{ij} |x_{ij}|^2\), respectively. Given a partition \(S,S^c\) and a vector \(x\in\{0,1\}^{|\mathcal{V}|}\) such that \(x_i=1\) if \(v_i\in S\) and \(x_i = 0\) if \(v_i\in S^c\), then \(\cut(S) = \|Bx\|_{1,C}\). Moreover, notice that \(\cut(S) = \|Bx\|_{2,C}^2 = x^T B^T C B x = x^T L x\). **Miscellaneous:** We use \([x;y;z]\) to denote a single column vector where the individual vectors \(x,y,z\) are stacked in this order. Finally, the vectors \(0_{|\mathcal{V}|}\) and \(1_{|\mathcal{V}|}\) are the all zeros and ones vectors of length \(|\mathcal{V}|\), respectively. ## Strongly-local flow and spectral methods #### \(\ell_1\)-Regularized Page-Rank The \(\ell_1\)-regularized Page-Rank problem was initially studied in and then further refined in. In the latter work, the problem is defined as \[\begin{aligned} \label{eq:l1pr} \mbox{minimize} & \quad \frac{1}{2}\|\tilde{B}\tilde{x}\|_{\tilde{C},2}^2 + \epsilon \|Dx\|_1, \end{aligned}\] where \(\tilde{x}: = [1; x; 0]\). The reference cut graph setting for [\[eq:l1pr\]](#eq:l1pr){reference-type="eqref" reference="eq:l1pr"} is \(g=d\) and \(h\ge 0\) is a vector that satisfies \(\|h\|_1=1\) and \(\|h\|_\infty \ge \epsilon\). The latter condition is to guarantee that the solution to [\[eq:l1pr\]](#eq:l1pr){reference-type="eqref" reference="eq:l1pr"} is not the zero vector. Moreover, \(\alpha = \beta\) and \(\gamma = (1-\alpha)/2\). Similarly to \(z\) for MOV, the vector \(h\) controls the input seed set and the weights of nodes in that set. The larger the weights the more the solution will be correlated with the corresponding nodes in the input seed set. The solution vector to problem [\[eq:l1pr\]](#eq:l1pr){reference-type="eqref" reference="eq:l1pr"} is component-wise non-negative and the parameter \(\alpha\) controls how much energy is concentrated close to the input seed set. Formally, based on theoretical guarantees in the vector \(h\) should be an indicator vector for a single seed node, around which there is a target cluster of nodes \(C\). The algorithm is not guaranteed to find the exact target cluster \(C\), but if \(C\) has conductance less than \(\alpha/10\) then it is guaranteed to return a cluster with conductance of \(\mathcal{O}(\sqrt{\alpha \log(\vol(C))})\). We refer the reader to for a detailed description of the theoretical graph clustering guarantees. The idea of \(\ell_1\)-regularized Page-Rank graph clustering initially appeared in in the form of implicit regularization. In particular, the authors in suggest solving a personalized Page-Rank linear system approximately. In, the authors noticed that the termination criteria in are related to the first-order optimality conditions of the above \(\ell_1\)-regularized Page-Rank problem, and they draw the connection to explicit \(\ell_1\) regularization. It is shown in that solving the \(\ell_1\)-regularized Page-Rank problem has the same Cheeger-like worst-case approximation guarantees to the Minimum-Conductance problem as the original procedure in. However, there is an important technical difference: one advantage of solving the \(\ell_1\)-regularized problem is that the locality of the solution is a property of the optimization problem as opposed to a property of an algorithm. In particular, by solving the \(\ell_1\)-regularized problem it is guaranteed to obtain the same solution regardless of the algorithm used. In comparison, applying the procedure in, where the output depends on the setting of the procedure, i.e., the strategy for choosing nodes to be updated at every iteration, leads to somewhat different solutions, depending on the specific settings chosen. Let \(x^*\) be the optimal solution of [\[eq:l1pr\]](#eq:l1pr){reference-type="eqref" reference="eq:l1pr"} and \(S^*\) be the set of nodes where \(x^*\) is non-zero. In, it is shown that many standard optimization algorithms such as iterative soft-thresholding or block iterative soft-thresholding solve [\[eq:l1pr\]](#eq:l1pr){reference-type="eqref" reference="eq:l1pr"} with running time \(\mathcal{O}(\vol(S^*)/\alpha)\), which can be independent of the volume of the whole graph \(\vol(\mathcal{V})\). This opens up the possibility of the use of these algorithms more generally. For details about the algorithms, we refer the reader to . [^1]: An important applied question has to do with the meaningfulness, usefulness, etc., of such small clusters. We do not consider those questions here, and instead we refer the interested reader to prior work . Here, we instead focus on the algorithmic and statistical properties of these locally-biased algorithms. [^2]: <http://snap.stanford.edu/data> [^3]: <https://graph-tool.skewed.de>
{'timestamp': '2016-12-06T02:08:31', 'yymm': '1607', 'arxiv_id': '1607.04940', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04940'}
null
null
# Introduction With the discovery of the Higgs Boson in 2012 by both the ATLAS  and the CMS  collaborations, our understanding of the physics of the fundamental particles up to the \(1\) TeV scale, otherwise known as the Standard Model (SM), has been experimentally confirmed. However, though we know with certainty that SM is the correct picture, we also know with an equal degree of certainty that it is not the complete one. The mathematical framework of the SM is inconsistent with that of general relativity, but the later has been experimentally verified. Other experimental discoveries such as the neutrino oscillations, along with theoretical questions such as the origin of mass, the matter antimatter asymmetry, the dark matter and dark energy, the strong CP problem etc provide ample evidence for the existence of new physics (NP) beyond the SM. Over the past three decades, flavour physics has emerged as an important testing ground for the existence of NP. For example, tensions between SM expectations and experimental results have been found in B physics for observables such as the isospin asymmetry \(A_I(B\rightarrow K\mu^+\mu^-)\) , the longitudinal polarization fraction in \(B_s\rightarrow K^*K^*\) , \(R_{D^*}^{\tau,l}=\frac{\text{BR}(B\rightarrow D^*\tau\nu_{\tau})}{\text{BR}(B\rightarrow D^*l\nu_l)}\), \(R_{D}^{\tau,l}=\frac{\text{BR}(B\rightarrow D\tau\nu_{\tau})}{\text{BR}(B\rightarrow Dl\nu_l)}\)  and \(R_K=\frac{\text{BR}(B\rightarrow K\mu^+\mu^-)}{\text{BR}(B\rightarrow Ke^+e^-)}\) . A number of theoretical works have been and are still being undertaken following these results. Top quark physics seems to be important for the search of NP in the up quark sector. However, no signal for NP has yet been detected in the top quark sector. Searches for NP have also been carried out in charm. A considerable amount of work on the \(D\rightarrow PP\) and \(D\rightarrow VP\) modes  has been undertaken over the last thirty years. In 2012, LHCb  and CDF  reported the first observation of a CP asymmetry between the \(D^0\rightarrow\pi^+\pi^-\) and the \(D^0\rightarrow K^+K^-\) modes. This was followed by a large amount of work , where the authors mostly used NP models to explain the same. The \(3.2\sigma\) hint has since then slowly disappeared. However, the \(D\rightarrow VV\) hadronic modes have received lesser attention . It is not an easy task to have a solid theoretical understanding of all the charm hadronic modes. This is due to the mass of the charm quark, which, unlike the bottom quark is not sufficiently heavy for the realization of the infinitely heavy quark limit. Hence, the well known approaches based on QCD that lead to satisfactory predictions for \(B\) decays like the heavy quark effective theory , the QCD-factorization , the perturbative QCD approach  and the soft-collinear effective theory , fail to achieve the same for \(D\) decays. The charm quark is also not light enough for the application of a chiral perturbation theory. Furthermore, in the case of hadronic vector final states, the calculation of the form factors poses greater difficulty than their \(D\rightarrow P\) counterparts. In absence of any reliable and effective theoretical models, it might be a good idea to look for possible NP in \(D\rightarrow VV\) decays in a model independent way. This is what we have tried to achieve in this paper. The method is similar to the one used in  for \(B\) decays. The idea is to define observables which can be experimentally measured, linear combinations of which have a value inconsistent with zero under the presence of new physics. Although extensively used for the \(B\rightarrow VV\) decays, such a formalism, to the best of our knowledge, has not yet been pursued for the \(D\rightarrow VV\) hadronic decays. Our paper is organised as follows. In section [\[ang_an\]](#ang_an){reference-type="ref" reference="ang_an"} we describe the formalism for defining such model independent observables in detail. The next section (section [\[pol\]](#pol){reference-type="ref" reference="pol"}) is about the connection between the polarization basis that we use for our analysis and the \(a\), \(b\), \(c\) amplitudes used in the most general covariant expression for a heavy pseudoscalar to two-body vector decay amplitude. We then discuss our observables and the effect NP has on them in section [\[results\]](#results){reference-type="ref" reference="results"}. We also extract numerical values for the amplitudes in the polarization basis for \(D^0\rightarrow \overline{K^{*0}}\rho^0\) mode, which is the best measured \(D\rightarrow VV\) mode in . Finally, we summarize and conclude in section [\[summary\]](#summary){reference-type="ref" reference="summary"}. # The time-dependent analysis[\[ang_an\]]{#ang_an label="ang_an"} Consider a \(D\rightarrow V_1V_2\) decay. Let the SM contribution to this decay be parametrized by a single decay amplitude along with a corresponding strong phase. The NP contribution will in general be characterised by a different decay amplitude with a different strong phase, along with a NP weak phase. The corresponding CP conjugate decay will then have the same components, with the sign of the NP weak phase reversed. A decay of the type \(D\rightarrow\bar{V_1}\bar{V_2}\), where the bar denotes the CP conjugate state, and it's corresponding CP conjugate decay will in general be parametrized by some different SM and NP parameters. Hence, for each of the above cases, the decay amplitude for each of the three possible helicity states may be written as: \[\begin{aligned} \label{amplitude} A_{\lambda}&\equiv& \text{Amp}(D\rightarrow V_1V_2)=a_{\lambda}e^{i\delta_{\lambda}^a}+b_{\lambda}e^{i\phi}e^{i\delta_{\lambda}^b}\nonumber\\ \bar{A}_{\bar\lambda}&\equiv& \text{Amp}(\bar{D}\rightarrow \bar{V_1}\bar{V_2})=a_{\lambda}e^{i\delta_{\lambda}^a}+b_{\lambda}e^{-i\phi}e^{i\delta_{\lambda}^b}\nonumber\\ A_{\bar\lambda}&\equiv& \text{Amp}(D\rightarrow \bar{V_1}\bar{V_2})=c_{\lambda}e^{i\delta_{\lambda}^c}+d_{\lambda}e^{i\gamma}e^{i\delta_{\lambda}^d}\nonumber\\ \bar{A}_{\lambda}&\equiv& \text{Amp}(\bar{D}\rightarrow V_1V_2)=c_{\lambda}e^{i\delta_{\lambda}^c}+d_{\lambda}e^{-i\gamma}e^{i\delta_{\lambda}^d}. \end{aligned}\] In the above, \(a_{\lambda}\), \(c_{\lambda}\) represent the SM decay amplitudes; \(\delta_{\lambda}^a\) and \(\delta_{\lambda}^c\) denote the SM strong phases; \(b_{\lambda}\), \(d_{\lambda}\) are the NP decay amplitudes; \(\delta_{\lambda}^b\) and \(\delta_{\lambda}^d\) the NP strong phases and \(\phi\), \(\gamma\) are the NP weak phases. The helicity index \(\lambda\) takes the values \({0, \parallel, \perp}\) denoting the longitudinal, transverse parallel and the transverse perpendicular helicities respectively. Using CPT invariance, the full amplitude for each of the above decays can be written as \[\begin{aligned} \label{amplitude} \mathcal{A}_{f}&\equiv&\text{Amp}\left(D(t)\rightarrow V_{1}V_{2}\right)={A_{f}}_{0}g_{0}+{A_{f}}_{\parallel}g_{\parallel}+i{A_{f}}_{\perp}g_{\perp}\nonumber\\ \mathcal{A}_{\bar f}&\equiv&\text{Amp}\left(D(t)\rightarrow\overline{V}_{1}\overline{V}_{2}\right)={A_{\bar f}}_{0}g_{0}+{A_{\bar f}}_{\parallel}g_{\parallel}+i{A_{\bar f}}_{\perp}g_{\perp}\nonumber\\ \bar{\mathcal{A}}_{f}&\equiv&\text{Amp}\left(\overline{D}(t)\rightarrow V_{1}V_{2}\right)={\overline{A}_{f}}_{0}g_{0}+{\overline{A}_{f}}_{\parallel}g_{\parallel}-i{\overline{A}_{f}}_{\perp}g_{\perp}\nonumber\\ \bar{\mathcal{A}}_{\bar f}&\equiv&\text{Amp}\left(\overline{D}(t)\rightarrow\overline{V}_{1}\overline{V}_{2}\right)={\overline{A}_{\bar f}}_{0}g_{0}+ {\overline{A}_{\bar f}}_{\parallel}g_{\parallel}-i{\overline{A}_{\bar f}}_{\perp}g_{\perp}. \end{aligned}\] where \(g_{\lambda}\)'s are basically functions of the angles describing the kinematics for the the corresponding decay. Based on the parametrization discussed above, one can now look into the time dependent decay rates of a neutral \(D\) meson going to two vector final states and perform an angular analysis of all the \(D(t)\to V_1V_2\), \(D(t)\to\overline{V}_1\overline{V}_2\) modes and their CP conjugate processes. The general expressions for the time-dependent decay rates of a neutral meson \(M^0\) going to two vector final states are given by \[\begin{aligned} \Gamma \left(M^0(t)\to V_1V_2\right)=&N_fe^{-\text{\(\Gamma \)t}}\left(\left|\mathcal{A}_f\right|^2+\left|\frac{q}{p} \bar{\mathcal{A}}_f\right|^2\right)\cosh\left(\text{y\(\Gamma \)t}\right) +\left(\left|\mathcal{A}_f\right|^2-\left|\frac{q}{p} \bar{\mathcal{A}}_f\right|^2\right)\cos\left(\text{x\(\Gamma \)t}\right)\nonumber&\\ &+2\text{Re}\left(\left(\frac{q}{p}\right)\mathcal{A}_{f}^{*}\bar{\mathcal{A}}_{f}\right)\sinh(\text{y\(\Gamma\)t}) -2\text{Im}\left(\left(\frac{q}{p}\right)\mathcal{A}_{f}^{*}\bar{\mathcal{A}}_{f}\right)\text{sin}\left(\text{x\(\Gamma\)t}\right)\nonumber&\\ \Gamma \left(M^0(t)\to \overline{V}_1 \overline{V}_2 \right)=&N_fe^{-\text{\(\Gamma \)t}}\left(\left|\mathcal{A}_{\bar f}\right|^2+\left|\frac{q}{p} \bar{\mathcal{A}}_{\bar f}\right|^2\right)\cosh\left(\text{y\(\Gamma \)t}\right) +\left(\left|\mathcal{A}_{\bar f}\right|^2-\left|\frac{q}{p} \bar{\mathcal{A}}_{\bar f}\right|^2\right)\cos\left(\text{x\(\Gamma \)t}\right)\nonumber&\\ &+2\text{Re}\left(\left(\frac{q}{p}\right)\mathcal{A}_{\bar f}^{*}\bar{\mathcal{A}}_{\bar f}\right)\sinh(\text{y\(\Gamma\)t}) -2\text{Im}\left(\left(\frac{q}{p}\right)\mathcal{A}_{\bar f}^{*}\bar{\mathcal{A}}_{\bar f}\right)\text{sin}\left(\text{x\(\Gamma\)t}\right)\nonumber&\\ \Gamma \left(\overline{M^0}(t)\to V_1V_2\right)=&N_fe^{-\text{\(\Gamma \)t}}\left(\left|\frac{p}{q} \mathcal{A}_f\right|^2+\left|\bar{\mathcal{A}}_f\right|^2\right)\cosh\left(\text{y\(\Gamma \)t}\right) -\left(\left|\frac{p}{q} \mathcal{A}_f\right|^2-\left|\bar{\mathcal{A}}_f\right|^2\right)\cos\left(\text{x\(\Gamma \)t}\right)\nonumber&\\ &+2\text{Re}\left(\left(\frac{p}{q}\right)\mathcal{A}_{f}\bar{\mathcal{A}}_{f}^{*}\right)\sinh(\text{y\(\Gamma\)t}) -2\text{Im}\left(\left(\frac{p}{q}\right)\mathcal{A}_{f}\bar{\mathcal{A}}_{f}^{*}\right)\text{sin}\left(\text{x\(\Gamma\)t}\right)\nonumber&\\ \Gamma \left(\overline{M^0}(t)\to \overline{V}_1 \overline{V}_2\right)=&N_fe^{-\text{\(\Gamma \)t}}\left(\left|\frac{p}{q} \mathcal{A}_{\bar f}\right|^2+\left|\bar{\mathcal{A}}_{\bar f}\right|^2\right)\cosh\left(\text{y\(\Gamma \)t}\right) -\left(\left|\frac{p}{q} \mathcal{A}_{\bar f}\right|^2-\left|\bar{\mathcal{A}}_{\bar f}\right|^2\right)\cos\left(\text{x\(\Gamma \)t}\right)\nonumber&\\ &+2\text{Re}\left(\left(\frac{p}{q}\right)\mathcal{A}_{\bar f}\bar{\mathcal{A}}_{\bar f}^{*}\right)\sinh(\text{y\(\Gamma\)t}) -2\text{Im}\left(\left(\frac{p}{q}\right)\mathcal{A}_{\bar f}\bar{\mathcal{A}}_{\bar f}^{*}\right)\text{sin}\left(\text{x\(\Gamma\)t}\right).& \end{aligned}\] Here \(x=\frac{\Delta m}{\Gamma}\), \(y=\frac{\Delta\Gamma}{2\Gamma}\), \(\Delta m=m_{H}-m_{L}\), \(\Delta\Gamma=\Gamma_{H}-\Gamma_{L}\). The indices H, L stand for the heavy and light mass eigenstates. \(N_{f}\) is a time independent normalization factor. Decays occurring without \(M^0-\overline{M^0}\) oscillations (pure decays) are given by the terms proportional to \(\left|\mathcal{A}_{f}\right|^{2}\) or \(\left|\bar{A}_{f}\right|^{2}\). Terms proportional to \(\left|\frac{q}{p}\bar{\mathcal{A}}_{f}\right|^{2}\) or \(\left|\frac{p}{q}\mathcal{A}_{f}\right|^{2}\) are due to decays following the \(M^0-\overline{M^0}\) oscillation (decays after mixing). BABAR, while measuring the effect of CP-violation on the time-dependent decay rates in the \(D\rightarrow K\pi\) system, parametrized the decay rates in powers of \(\Gamma t\) . They truncate the series at second order in \(\Gamma t\). However the sensitivity of the coefficient of the second order term to the decay rate is less than that of the coefficient of the first order term . We follow the same parametrization and keep the terms up to the quadratic order in \(\Gamma t\). It is clear that our observables will be linear combinations of the coefficients of the sinh, sin, cosh and cos terms. For the case of neutral charm decays, the corresponding decay rates look like \[\begin{aligned} \label{decay rate} \Gamma \left(D^0(t)\to V_{1}V_{2}\right)=&N_fe^{-\text{\(\Gamma \)t}}\bigg(2\left|\mathcal{A}_{f}\right|^{2}+\left[2yRe\left(\frac{q}{p}\mathcal{A}_{f}^{*}\bar{\mathcal{A}}_{f}\right)-2xIm\left(\frac{q}{p}\mathcal{A}_{f}^{*}\bar{\mathcal{A}}_{f}\right)\right]\Gamma t\nonumber\\ &+\left[\frac{(y^{2}-x^{2})}{2}\left|\mathcal{A}_{f}\right|^{2}+\frac{(x^{2}+y^{2})}{2}\left|\frac{q}{p}\bar{\mathcal{A}}_{f}\right|^{2}\right](\Gamma t)^{2}\bigg)\nonumber\\ \Gamma \left(D^0(t)\to \overline{V}_{1} \overline{V}_{2}\right)=&N_fe^{-\text{\(\Gamma \)t}}\bigg(2\left|\mathcal{A}_{\bar f}\right|^{2}+\left[2yRe\left(\frac{q}{p}\mathcal{A}_{\bar f}^{*}\bar{\mathcal{A}}_{\bar f}\right)-2xIm\left(\frac{q}{p}\mathcal{A}_{\bar f}^{*}\bar{\mathcal{A}}_{\bar f}\right)\right]\Gamma t\nonumber\\ &+\left[\frac{(y^{2}-x^{2})}{2}\left|\mathcal{A}_{\bar f}\right|^{2}+\frac{(x^{2}+y^{2})}{2}\left|\frac{q}{p}\bar{\mathcal{A}}_{\bar f}\right|^{2}\right](\Gamma t)^{2}\bigg)\nonumber\\ \Gamma \left(\overline{D^0}(t)\to V_{1}V_{2}\right)=&N_fe^{-\text{\(\Gamma \)t}}\bigg(2\left|\bar{\mathcal{A}}_{f}\right|^{2}+\left[2yRe\left(\frac{p}{q}\bar{\mathcal{A}}_{f}^{*}\mathcal{A}_{f}\right)-2xIm\left(\frac{p}{q}\bar{\mathcal{A}}_{f}^{*}\mathcal{A}_{f}\right)\right]\Gamma t\nonumber\\ &+\left[\frac{(y^{2}-x^{2})}{2}\left|\bar{\mathcal{A}}_{f}\right|^{2}+\frac{(x^{2}+y^{2})}{2}\left|\frac{p}{q}\mathcal{A}_{f}\right|^{2}\right](\Gamma t)^{2}\bigg)\nonumber\\ \Gamma \left(\overline{D^0}(t)\to\overline{V}_{1}\overline{V}_{2}\right)=&N_fe^{-\text{\(\Gamma \)t}}\bigg(2\left|\bar{\mathcal{A}}_{\bar f}\right|^{2}+\left[2yRe\left(\frac{p}{q}\bar{\mathcal{A}}_{\bar f}^{*}\mathcal{A}_{\bar f}\right)-2xIm\left(\frac{p}{q}\bar{\mathcal{A}}_{\bar f}^{*}\mathcal{A}_{\bar f}\right)\right]\Gamma t\nonumber\\ &+\left[\frac{(y^{2}-x^{2})}{2}\left|\bar{\mathcal{A}}_{\bar f}\right|^{2}+\frac{(x^{2}+y^{2})}{2}\left|\frac{p}{q}\mathcal{A}_{\bar f}\right|^{2}\right](\Gamma t)^{2}\bigg) \end{aligned}\] Using eqns ([\[amplitude\]](#amplitude){reference-type="ref" reference="amplitude"}) and ([\[decay rate\]](#decay rate){reference-type="ref" reference="decay rate"}) we can write the time dependent decay rates as \[\begin{aligned} \Gamma\left(D^0(t)\to V_{1}V_{2}\right)&=&e^{-\Gamma t}\sum_{\lambda\leq\lambda'}\left(X^{V_{1}V_{2}}_{\lambda\lambda'}+Y^{V_{1}V_{2}}_{\lambda\lambda'}\,\Gamma t+Z^{V_{1}V_{2}}_{\lambda\lambda'}\,(\Gamma t)^2\right)g_{\lambda}g_{\lambda'}\nonumber\\ \Gamma\left(D^0(t)\to\overline{V}_{1}\overline{V}_{2}\right)&=&e^{-\Gamma t}\sum_{\lambda\leq\lambda'}\left(X^{\overline{V}_{1}\overline{V}_{2}}_{\lambda\lambda'}+Y^{\overline{V}_{1}\overline{V}_{2}}_{\lambda\lambda'}\,\Gamma t+Z^{\overline{V}_{1}\overline{V}_{2}}_{\lambda\lambda'}\,(\Gamma t)^2\right)g_{\lambda}g_{\lambda'}\nonumber\\ \Gamma\left(\overline{D^0}(t)\to V_{1}V_{2}\right)&=&e^{-\Gamma t}\sum_{\lambda\leq\lambda'}\left(\bar{X}^{V_{1}V_{2}}_{\lambda\lambda'}+\bar{Y}^{V_{1}V_{2}}_{\lambda\lambda'}\,\Gamma t+\bar{Z}^{V_{1}V_{2}}_{\lambda\lambda'}\,(\Gamma t)^2\right)g_{\lambda}g_{\lambda'}\nonumber\\ \Gamma\left(\overline{D^0}(t)\to\overline{V}_{1}\overline{V}_{2}\right)&=&e^{-\Gamma t}\sum_{\lambda\leq\lambda'}\left(\bar{X}^{\overline{V}_{1}\overline{V}_{2}}_{\lambda\lambda'}+\bar{Y}^{\overline{V}_{1}\overline{V}_{2}}_{\lambda\lambda'}\,\Gamma t+\bar{Z}^{\overline{V}_{1}\overline{V}_{2}}_{\lambda\lambda'}\,(\Gamma t)^2\right)g_{\lambda}g_{\lambda'}. \end{aligned}\] Thus by a time dependent angular analysis of the decay modes \(D^0(t)\to V_{1}V_{2}\), \(D^0(t)\to\overline{V}_{1}\overline{V}_{2}\), \(\overline{D^0}(t)\to V_{1}V_{2}\), \(\overline{D^0}(t)\to\overline{V}_{1}\overline{V}_{2}\), one can define 72 observables. However, not all of these observables are independent. For example, it can be easily verified that the \(Z\)'s (the coefficient of the \((\Gamma t)^2\)) terms) are linear combinations of the \(X\)'s. In fact, observables extracted from the coefficients of the terms that are even powers i \(\Gamma t\) can always be written as linear combinations of the \(X\)'s, while those extracted from the coefficients of the terms that are odd powers in \(\Gamma t\) can be written as linear combinations of \(Y\)'s. Hence, in our analysis, we deal with a total of 48 observables (24 \(X\)'s and 24 \(Y\)'s). In terms of the amplitudes, these are given by: \[\begin{aligned} \label{pol_obs} X^{V_{1}V_{2}}_{\lambda\lambda}=&2(A_{f}^{*})_{\lambda}(A_{f})_{\lambda}\nonumber\\ X^{V_{1}V_{2}}_{i\perp}=&-4Im\left[(A_{f}^{*})_{i}(A_{f})_{\perp}\right]\nonumber\\ X^{V_{1}V_{2}}_{0\parallel}=&4Re\left[(A_{f})_{0}^{*}(A_{f})_{\parallel}\right]\nonumber\\ X^{\overline{V}_{1}\overline{V}_{2}}_{\lambda\lambda}=&2(A_{\bar f}^{*})_{\lambda}(A_{\bar f})_{\lambda}\nonumber\\ X^{\overline{V}_{1}\overline{V}_{2}}_{i\perp}=&-4Im\left[(A_{\bar f}^{*})_{i}(A_{\bar f})_{\perp}\right]\nonumber\\ X^{\overline{V}_{1}\overline{V}_{2}}_{0\parallel}=&4Re\left[(A_{\bar f})_{0}^{*}(A_{\bar f})_{\parallel}\right]\nonumber\\ \bar{X}^{V_{1}V_{2}}_{\lambda\lambda}=&2(\bar{A}_{f}^{*})_{\lambda}(\bar{A}_{f})_{\lambda}\nonumber\\ \bar{X}^{V_{1}V_{2}}_{i\perp}=&-4Im\left[(\bar{A}_{f}^{*})_{\perp}(\bar{A}_{f})_{i}\right]\nonumber\\ \bar{X}^{V_{1}V_{2}}_{0\parallel}=&4Re\left[(\bar{A}_{f})_{0}^{*}(\bar{A}_{f})_{\parallel}\right]\nonumber\\ \bar{X}^{\overline{V}_{1}\overline{V}_{2}}_{\lambda\lambda}=&2(\bar{A}_{\bar f}^{*})_{\lambda}(\bar{A}_{\bar f})_{\lambda}\nonumber\\ \bar{X}^{\overline{V}_{1}\overline{V}_{2}}_{i\perp}=&-4Im\left[(\bar{A}_{\bar f}^{*})_{\perp}(\bar{A}_{\bar f})_{i}\right]\nonumber\\ \bar{X}^{\overline{V}_{1}\overline{V}_{2}}_{0\parallel}=&4Re\left[(\bar{A}_{\bar f})_{0}^{*}(\bar{A}_{\bar f})_{\parallel}\right]\nonumber\\ Y^{V_{1}V_{2}}_{ii}=&2yRe\left[\frac{q}{p}(A_{f}^{*})_{i}(\bar{A}_{f})_{i}\right]-2xIm\left[\frac{q}{p}(A_{f}^{*})_{i}(\bar{A}_{f})_{i}\right]\nonumber\\ Y^{V_{1}V_{2}}_{\perp\perp}=&-2yRe\left[\frac{q}{p}(A_{f}^{*})_{\perp}(\bar{A}_{f})_{\perp}\right] +2xIm\left[\frac{q}{p}(A_{f}^{*})_{\perp}(\bar{A}_{f})_{\perp}\right]\nonumber\\ Y^{V_{1}V_{2}}_{i\perp}=&-2yRe\left[\frac{q}{p}i\left((A_{f}^{*})_{i}(\bar{A}_{f})_{\perp}+(A_{f}^{*})_{\perp}(\bar{A}_{f})_{i}\right)\right] +2xIm\left[\frac{q}{p}i\left((A_{f}^{*})_{i}(\bar{A}_{f})_{\perp}+(A_{f}^{*})_{\perp}(\bar{A}_{f})_{i}\right)\right]\nonumber\\ Y^{V_{1}V_{2}}_{0\parallel}=&2yRe\left[\frac{q}{p}\left((A_{f})_{0}^{*}(\bar{A}_{f})_{\parallel}+(A_{f})_{\parallel}^{*}(\bar{A}_{f})_{0}\right)\right]-2xIm\left[\frac{q}{p}\left((A_{f})_{0}^{*}(\bar{A}_{f})_{\parallel}+(A_{f})_{\parallel}^{*}(\bar{A}_{f})_{0}\right)\right]\nonumber\\ Y^{\overline{V}_{1}\overline{V}_{2}}_{ii}=&2yRe\left[\frac{q}{p}(A_{\bar f}^{*})_{i}(\bar{A}_{\bar f})_{i}\right]-2xIm\left[\frac{q}{p}(A_{\bar f}^{*})_{i}(\bar{A}_{\bar f})_{i}\right]\nonumber\\ Y^{\overline{V}_{1}\overline{V}_{2}}_{\perp\perp}=&-2yRe\left[\frac{q}{p}(A_{\bar f}^{*})_{\perp}(\bar{A}_{\bar f})_{\perp}\right] +2xIm\left[\frac{q}{p}(A_{\bar f}^{*})_{\perp}(\bar{A}_{\bar f})_{\perp}\right]\nonumber\\ Y^{\overline{V}_{1}\overline{V}_{2}}_{i\perp}=&-2yRe\left[\frac{q}{p}i\left((A_{\bar f}^{*})_{i}(\bar{A}_{\bar f})_{\perp}+(A_{\bar f}^{*})_{\perp}(\bar{A}_{\bar f})_{i}\right)\right] +2xIm\left[\frac{q}{p}i\left((A_{\bar f}^{*})_{i}(\bar{A}_{\bar f})_{\perp}+(A_{\bar f}^{*})_{\perp}(\bar{A}_{\bar f})_{i}\right)\right]\nonumber\\ \end{aligned}\] \[\begin{aligned} Y^{\overline{V}_{1}\overline{V}_{2}}_{0\parallel}=&2yRe\left[\frac{q}{p}\left((A_{\bar f})_{0}^{*}(\bar{A}_{\bar f})_{\parallel}+(A_{\bar f})_{\parallel}^{*}(\bar{A}_{\bar f})_{0}\right)\right]-2xIm\left[\frac{q}{p}\left((A_{\bar f})_{0}^{*}(\bar{A}_{\bar f})_{\parallel}+(A_{\bar f})_{\parallel}^{*}(\bar{A}_{\bar f})_{0}\right)\right]\nonumber\\ \bar{Y}^{V_{1}V_{2}}_{ii}=&2yRe\left[\frac{p}{q}(\bar{A}_{f}^{*})_{i}(A_{f})_{i}\right]-2xIm\left[\frac{p}{q}(\bar{A}_{f}^{*})_{i}(A_{f})_{i}\right]\nonumber\\ \bar{Y}^{V_{1}V_{2}}_{\perp\perp}=&-2yRe\left[\frac{p}{q}(\bar{A}_{f}^{*})_{i}(A_{f})_{i}\right] +2xIm\left[\frac{p}{q}(\bar{A}_{f}^{*})_{i}(A_{f})_{i}\right]\nonumber\\ \bar{Y}^{V_{1}V_{2}}_{i\perp}=&2yRe\left[\frac{p}{q}i\left((\bar{A}_{f}^{*})_{i}(A_{f})_{\perp}+(\bar{A}_{f}^{*})_{\perp}(A_{f})_{i}\right)\right]-2xIm\left[\frac{p}{q}i\left((\bar{A}_{f}^{*})_{i}(A_{f})_{\perp}+(\bar{A}_{f}^{*})_{\perp}(A_{f})_{i}\right)\right]\nonumber\\ \bar{Y}^{V_{1}V_{2}}_{0\parallel}=&2yRe\left[\frac{p}{q}\left((\bar{A}_{f})_{0}^{*}(A_{f})_{\parallel}+(\bar{A}_{f})_{\parallel}^{*}(A_{f})_{0}\right)\right]-2xIm\left[\frac{p}{q}\left((\bar{A}_{f})_{0}^{*}(A_{f})_{\parallel}+(\bar{A}_{f})_{\parallel}^{*}(A_{f})_{0}\right)\right]\nonumber\\ \bar{Y}^{\overline{V}_{1}\overline{V}_{2}}_{ii}=&2yRe\left[\frac{p}{q}(\bar{A}_{\bar f}^{*})_{i}(A_{\bar f})_{i}\right]-2xIm\left[\frac{p}{q}(\bar{A}_{\bar f}^{*})_{i}(A_{\bar f})_{i}\right]\nonumber\\ \bar{Y}^{\overline{V}_{1}\overline{V}_{2}}_{\perp\perp}=&-2yRe\left[\frac{p}{q}(\bar{A}_{\bar f}^{*})_{\perp}(A_{\bar f})_{\perp}\right] +2xIm\left[\frac{p}{q}(\bar{A}_{\bar f}^{*})_{\perp}(A_{\bar f})_{\perp}\right]\nonumber\\ \bar{Y}^{\overline{V}_{1}\overline{V}_{2}}_{i\perp}=&2yRe\left[\frac{p}{q}i\left((\bar{A}_{\bar f}^{*})_{i}(A_{\bar f})_{\perp}+(\bar{A}_{\bar f}^{*})_{\perp}(A_{\bar f})_{i}\right)\right]-2xIm\left[\frac{p}{q}i\left((\bar{A}_{\bar f}^{*})_{i}(A_{\bar f})_{\perp}+(\bar{A}_{\bar f}^{*})_{\perp}(A_{\bar f})_{i}\right)\right]\nonumber\\ \bar{Y}^{\overline{V}_{1}\overline{V}_{2}}_{0\parallel}=&2yRe\left[\frac{p}{q}\left((\bar{A}_{\bar f})_{0}^{*}(A_{\bar f})_{\parallel}+(\bar{A}_{\bar f})_{\parallel}^{*}(A_{\bar f})_{0}\right)\right]-2xIm\left[\frac{p}{q}\left((\bar{A}_{\bar f})_{0}^{*}(A_{\bar f})_{\parallel}+(\bar{A}_{\bar f})_{\parallel}^{*}(A_{\bar f})_{0}\right)\right] \end{aligned}\] where \(\frac{q}{p}=re^{i(\alpha_{SM}+\alpha_{NP})} \footnote{Within the SM, CP violation in the charm sector is known to be very small (\)\mathcal O(\lambda^5)\( in the Wolfenstein parameterization of the CKM matrix). However, no NP has yet been found in the charm sector, and hence, any NP effect, if present, has to be very tiny too. It is therefore not feasible to neglect the tiny CP violating phase in charm mixing when one tries to probe NP in the charm sector. Hence we parameterize the CP violating mixing phase in charm according to the SM (\)\alpha_{SM}\() and the NP (\)\alpha_{NP}\() counterparts.}\). These 48 observables can be written in terms of 28 independent parameters: \(r\), \(\alpha_{SM}, \alpha_{NP}\), \(\phi\), \(\chi\), three each of \(a_{\lambda}\)'s, \(b_{\lambda}\)'s, \(c_{\lambda}\)'s, \(d_{\lambda}\)'s, and eleven strong phase differences \(\delta^{ab}_{\lambda}=\delta^{a}_{\lambda}-\delta^{b}_{\lambda}\), \(\delta^{cd}_{\lambda}=\delta^{c}_{\lambda}-\delta^{d}_{\lambda}\), \(\delta^{ac}_{\lambda}=\delta^{a}_{\lambda}-\delta^{c}_{\lambda}\), \(\Delta^{a}_{\parallel}=\delta^{a}_{\perp}-\delta^{a}_{\parallel}\). The expressions of the observables in terms of the theoretical parameters can be found in appendix [\[obs\]](#obs){reference-type="ref" reference="obs"}. # The \(\textbf{a}\), \(\textbf{b}\), \(\textbf{c}\) amplitudes[\[pol\]]{#pol label="pol"} In this section, we briefly discuss the relation of the polarization basis to the \(a\), \(b\) and \(c\) amplitudes used in the most general covariant expression for a \(D\rightarrow V_1V_2\) decay. We closely follow  in the discussions of this section. Consider the decay \(D^{0}\to V_{1}V_{2}\). Angular momentum conservation (\(\vec{J}_{D}=\vec{L}_{V_{1}V_{2}}+\vec{S}_{V_{1}V_{2}}\)), dictates that \(\vec{L}_{V_{1}V_{2}}\) can be 0, 1, 2. Thus, one obtains three independent amplitudes corresponding to the three different \(\vec{L}_{V_{1}V_{2}}\) values. The most general covariant amplitude for a \(D^{0}\to V_{1}V_{2}\) decay can be written as  \[\begin{aligned} \label{covariant amplitude} M_{\lambda_{1}\lambda_{2}}=&<V_{1}(\lambda_{1})V_{2}(\lambda_{2})|H_{wk}|D>\nonumber\\ =&\epsilon_{1\mu}^{*}\epsilon_{2\nu}^{*}\bigg[ag^{\mu\nu}+\frac{b}{m_{1}m_{2}}p_{2}^{\mu}p_{1}^{\nu}+i\frac{c}{m_{1}m_{2}}\epsilon^{\mu\nu\alpha\beta}p_{1\alpha}p_{2\beta}\bigg], \end{aligned}\] Here \(\epsilon_{1}\), \(\epsilon_{2}\) represent the polarization vectors and \(m_{1}\), \(m_{2}\) the masses of the vector mesons \(V_{1}\) and \(V_{2}\) respectively. The invariant amplitudes \(a\), \(b\), \(c\) each carry the dimension of energy. The corresponding decay rate in terms of the \(a\), \(b\) and \(c\) amplitudes is given by \[\label{gamma_cov} \Gamma\left(D\to V_{1}V_{2}\right)=\frac{\left|\bf k\right|}{8\pi m_{D}^{2}}\left(2|a|^{2}+|xa+(x^{2}-1)b|^{2}+2(x^{2}-1)|c|^{2}\right],\] where \(\left|\bf k\right|\) is the decay momentum and \(x=\frac{p_{1}.p_{2}}{m_{1}m_{2}}=\frac{m_{D}^{2}-m_{1}^{2}-m_{2}^{2}}{2m_{1}m_{2}}\). The \(D^{0}\to V_{1}V_{2}\) amplitude can also be written in the linear polarization basis as, \[\begin{aligned} \label{linear polarisation amplitude} A\left(D^{0}\to V_{1}V_{2}\right)=A_{0}\epsilon_{1}^{*L}\epsilon_{2}^{*L}-A_{\parallel}\vec{\epsilon}_{1}^{*T}.\vec{\epsilon}_{2}^{*T}/\sqrt{2} -iA_{\perp}\vec{\epsilon}_{1}^{*}\times\vec{\epsilon}_{2}^{*}.\hat{p}_{2}/\sqrt{2}, \end{aligned}\] where \(\epsilon_{i}^{*L}=\vec{\epsilon}_{i}^{*}.\hat{p}_{2}\), \(\vec{\epsilon}_{i}^{*T}=\vec{\epsilon}_{i}^{*}-\epsilon_{i}^{*L}\hat{p}_{2}\), \(\hat{p}_{2}\) is a unit vector in the direction of the momentum of the meson \(V_{2}\) in the \(V_{1}\) rest frame. Comparing eqn.([\[linear polarisation amplitude\]](#linear polarisation amplitude){reference-type="ref" reference="linear polarisation amplitude"}) to eqn.([\[amplitude\]](#amplitude){reference-type="ref" reference="amplitude"}), it can easily be verified that \(g_{0}=\epsilon_{1}^{*L}\epsilon_{2}^{*L}\), \(g_{\parallel}=-\frac{1}{\sqrt{2}}\epsilon_{1}^{*T}.\epsilon_{2}^{*T}\) and \(g_{\perp}=-\frac{1}{\sqrt{2}}\epsilon_{1}^{*}\times\epsilon_{2}^{*}.\hat{p}_{2}\). The corresponding decay width in the linear polarization basis is given by \[\label{gamma_pol} \Gamma\left(D^{0}\to V_{1}V_{2}\right)=\frac{\left|\bf k\right|}{8\pi m_{D}^{2}}\left(|A_{0}|^{2}+|A_{\parallel}|^{2}+|A_{\perp}|^{2}\right).\] From eqns.([\[gamma_cov\]](#gamma_cov){reference-type="ref" reference="gamma_cov"}) and ([\[gamma_pol\]](#gamma_pol){reference-type="ref" reference="gamma_pol"}), it is evident that the two bases are related as \[\begin{aligned} A_{\parallel}=\sqrt{2}a,\,\,\,A_{0}=-ax-b(x^{2}-1),\,\,\,A_{\perp}=\sqrt{2(x^{2}-1)}c. \end{aligned}\] # Results and discussions[\[results\]]{#results label="results"} A priori, the dependence of the observables on the theoretical parameters may not be trivial as is evident from appendix [\[obs\]](#obs){reference-type="ref" reference="obs"}. However, a careful and systematic study of these observables enables one to propose multiple ways for the identification and extraction of new physics as will be shown in this section. In what follows, \(\lambda=\{0,\parallel,\perp\}\) and \(i=\{0,\parallel\}\). Let us look at the SM case first. For this case the NP decay amplitudes \(b_{\lambda}\), \(d_{\lambda}\), the NP strong phases \(\delta^{b}_{\lambda}\), \(\delta^{d}_{\lambda}\) and the NP weak phases \(\phi\) and \(\chi\) and \(\alpha_{NP}\) are all equal to 0. With these values, we get the following relations between the observables \[\begin{aligned} &X^{V_1V_2}_{\lambda\lambda}=\bar{X}^{\bar{V_1}\bar{V_2}}_{\lambda\lambda},\; X^{\bar{V_1}\bar{V_2}}_{\lambda\lambda}=\bar{X}^{V_1V_2}_{\lambda\lambda},\; X^{V_1V_2}_{0\parallel}=\bar{X}^{\bar{V_1}\bar{V_2}}_{0\parallel},\label{SM_1}\\ &X^{\bar{V_1}\bar{V_2}}_{0\parallel}=\bar{X}^{V_1V_2}_{0\parallel},\; X^{V_1V_2}_{i\perp}=-\bar{X}^{\bar{V_1}\bar{V_2}}_{i\perp},\; X^{\bar{V_1}\bar{V_2}}_{i\perp}=-\bar{X}^{V_1V_2}_{i\perp}\label{SM_2}. \end{aligned}\] It is possible to construct linear combinations of the \(Y\)'s which depend directly on the phases \(\alpha_{SM}\) and the decay phase difference \(\delta_{\lambda}^{ac}\) as follows: \[\begin{aligned} &1+\frac{Y_{\lambda \lambda }^{\bar{V_1} \bar{V_2}}+Y_{\lambda \lambda }^{V_1 V_2}}{r^2 \bar{Y}_{\lambda \lambda }^{\bar{V_1} \bar{V_2}}+r^2 \bar{Y}_{\lambda \lambda }^{V_1 V_2}}=\frac{2y\cos\alpha_{SM}}{y\cos\alpha_{SM}+x\sin\alpha_{SM}}\label{SM_3}\\ &\frac{r^2 \bar{Y}_{\lambda \lambda }^{\bar{V_1} \bar{V_2}}+r^2 \bar{Y}_{\lambda \lambda }^{V_1 V_2}+Y_{\lambda \lambda }^{\bar{V_1} \bar{V_2}}+Y_{\lambda \lambda }^{V_1 V_2}}{r^2 \bar{Y}_{\lambda \lambda }^{\bar{V_1} \bar{V_2}}-r^2 \bar{Y}_{\lambda \lambda }^{V_1 V_2}-Y_{\lambda \lambda }^{\bar{V_1} \bar{V_2}}+Y_{\lambda \lambda }^{V_1 V_2}}=\frac{x}{y}\tan\delta_{\lambda}^{ac}\label{SM_4} \end{aligned}\] Eqn.([\[SM_3\]](#SM_3){reference-type="ref" reference="SM_3"}) is of particular interest. The right hand side of the equation is known to appreciable accuracy for the SM, and so is \(r\) on the left hand side. The observables \(Y\) can be measured experimentally. Hence this relation can be used as a smoking gun signal for detecting the presence of NP. One can go further and identify whether the NP manifests itself in pure decays or as a CP violating effect in mixing or both. For instance, suppose there is no NP in pure decays but some signal for CP violation has been observed. In that case, the formalism dictates that \(b_{\lambda}=d_{\lambda}=\delta^{b}_{\lambda}=\delta^{d}_{\lambda}=\phi=\chi=0\), but \(\alpha_{NP}\neq0\). This case is very similar to the SM case discussed above. In particular the relations ([\[SM_1\]](#SM_1){reference-type="ref" reference="SM_1"}), ([\[SM_2\]](#SM_2){reference-type="ref" reference="SM_2"}) and ([\[SM_4\]](#SM_4){reference-type="ref" reference="SM_4"}) hold exactly in the same way. This is because the \(X\)'s do not depend on the (small) NP phase \(\alpha_{NP}\) at all, and in the other case the \(\alpha_{NP}\) dependence gets canceled in the left hand side of eqn.([\[SM_4\]](#SM_4){reference-type="ref" reference="SM_4"}). However eqn.([\[SM_3\]](#SM_3){reference-type="ref" reference="SM_3"}) is modified to \[1+\frac{Y_{\lambda \lambda }^{\bar{V_1} \bar{V_2}}+Y_{\lambda \lambda }^{V_1 V_2}}{r^2 \bar{Y}_{\lambda \lambda }^{\bar{V_1} \bar{V_2}}+r^2 \bar{Y}_{\lambda \lambda }^{V_1 V_2}}=\frac{2y}{y+x\tan(\alpha_{SM}+\alpha_{NP})}\label{no_decay_1}\\ \] Eqn.([\[no_decay_1\]](#no_decay_1){reference-type="ref" reference="no_decay_1"}) along with relations of the type \[Y^{V_1V_2}_{\lambda\lambda}+r^2\bar{Y}^{\bar{V_1}\bar{V_2}}_{\lambda\lambda}+Y^{\bar{V_1}\bar{V_2}}_{\lambda\lambda}+r^2\bar{Y}^{V_1V_2}_{\lambda\lambda}=4y\frac{\cos\alpha_{SM}+\alpha_{NP}}{y\cos\alpha_{SM}+\alpha_{NP}-x\sin\alpha_{SM}+\alpha_{NP}}\frac{Y^{V_1V_2}_{\lambda\lambda}+Y^{\bar{V_1}\bar{V_2}}_{\lambda\lambda}}{\sqrt{X^{V_1V_2}_{\lambda\lambda}X^{\bar{V_1}\bar{V_2}}_{\lambda\lambda}}}.\] can be simultaneously solved for the determination of \(r\) and \(\alpha_{NP}\) in this case. We next look into the case where NP is manifested only in pure decays. In this case, \(\alpha_{NP}=0\), but \(b_{\lambda}\), \(d_{\lambda}\), \(\delta^{b}_{\lambda}\), \(\delta^{d}_{\lambda}\), \(\phi\), \(\chi\) \(\neq0\). For these values of the parameters, we have the following relations among the observables \(X\) and \(Y\): \[\frac{x^2({Y^{V_1V_2}_{\lambda\lambda}}+{\bar{Y}^{V_1V_2}_{\lambda\lambda}})^2+y^2({Y^{V_1V_2}_{\lambda\lambda}}-{\bar{Y}^{V_1V_2}_{\lambda\lambda}})^2}{x^2({Y^{\bar{V_1}\bar{V_2}}_{\lambda\lambda}}+{\bar{Y}^{\bar{V_1}\bar{V_2}}_{\lambda\lambda}})^2+y^2({Y^{\bar{V_1}\bar{V_2}}_{\lambda\lambda}}-{\bar{Y}^{\bar{V_1}\bar{V_2}}_{\lambda\lambda}})^2}=\frac{\bar{X}^{V_1V_2}_{\lambda\lambda}{X^{V_1V_2}_{\lambda\lambda}}}{{\bar{X}^{\bar{V_1}\bar{V_2}}_{\lambda\lambda}}{X^{\bar{V_1}\bar{V_2}}_{\lambda\lambda}}}.\label{no_mixing_1}\] It would be appropriate here to point out relations of the type \[\begin{aligned} &{X^{V_1V_2}_{\lambda\lambda}}-{\bar{X}^{\bar{V_1}\bar{V_2}}_{\lambda\lambda}}=8\sin\phi\sin\delta^{ab}_{\lambda}a_{\lambda}b_{\lambda},\label{no_mixing_2}\\ &{X^{\bar{V_1}\bar{V_2}}_{\lambda\lambda}}-{\bar{X}^{V_1V_2}_{\lambda\lambda}}=8\sin\chi\sin\delta^{cd}_{\lambda}c_{\lambda}d_{\lambda}\label{no_mixing_3}. \end{aligned}\] Note that a non zero value of either of the above linear combinations is again a smoking gun signal for NP, since that will mean that the NP weak phases (\(\phi\), \(\chi\)) and the NP weak decay amplitudes (\(b_{\lambda}\), \(d_{\lambda}\)) \(\neq0\). Eqns.([\[no_mixing_1\]](#no_mixing_1){reference-type="ref" reference="no_mixing_1"}), ([\[no_mixing_2\]](#no_mixing_2){reference-type="ref" reference="no_mixing_2"}) and ([\[no_mixing_3\]](#no_mixing_3){reference-type="ref" reference="no_mixing_3"}) hold in the most general case also were NP is present both in pure decays and mixing. The extraction of \(\mid\frac{q}{p}\mid\) in the most general scenario can be obtained from relations of the type: \[x^2({Y^{V_1V_2}_{\lambda\lambda}}+{r^2\bar{Y}^{V_1V_2}_{\lambda\lambda}})^2+y^2({Y^{V_1V_2}_{\lambda\lambda}}-{r^2\bar{Y}^{V_1V_2}_{\lambda\lambda}})^2=4r^2x^2y^2{\bar{X}^{V_1V_2}_{\lambda\lambda}}{X^{V_1V_2}_{\lambda\lambda}}.\] Let us now discuss if these observables can be used for the extraction of all the parameters. It is appropriate at this point to look into the case of final states in \(D^0\) decays that are CP eigenstates. These include modes like \(D^0\rightarrow\phi\omega\), \(D^0\rightarrow\rho^0\phi\), \(D^0\rightarrow\rho^0\omega\), \(D^0\rightarrow\rho\rho\), \(D^0\rightarrow\omega\omega\) etc. For these modes, \(O^{V_1V_2}=O^{\bar{V_1}\bar{V_2}}\) where \(O=\{X, \bar{X}, Y, \bar{Y}\}\) for all combinations of the polarization indices. Therefore the number of observables reduces from 48 to 24. However there are 28 parameters. Discarding \(\alpha_{SM}\) (which is precisely known already) and \(r\) (since the NP effect, if present, must be very small, it is a reasonable approximation to incorporate the complete effect of NP for the ratio \(q/p\) into the phase \(\alpha_{NP}\) without changing \(r\), which, then, is again precisely known) still leaves 26 parameters to be fitted to 24 observables. Hence, the numerical estimates for all the parameters cannot be obtained for final states that are CP eigenstates without some approximations to reduce the number of parameters. However, relations like ([\[no_mixing_2\]](#no_mixing_2){reference-type="ref" reference="no_mixing_2"}) (eqn.([\[no_mixing_2\]](#no_mixing_2){reference-type="ref" reference="no_mixing_2"}) and eqn.([\[no_mixing_3\]](#no_mixing_3){reference-type="ref" reference="no_mixing_3"}) are same in this case) where linear combinations of experimentally measured observables vanish in the absence of new physics can still be devised as tests for the detection of NP effects. For a complete estimation of all the parameters, one has to look at final states which are non CP eigenstates, i.e. modes involving \(K^{*0}\) as one of the final states. At present, there is serious dearth of data in the \(D\rightarrow VV\) sector. The most well measured mode is \(D^0\rightarrow\bar{K^{*0}}\rho^0\) where there are estimates for the total branching fraction, the total transverse wave, S-wave, longitudinal S-wave, P-wave and D-wave branching fractions . The P-wave and longitudinal branching fraction are limits. This gives one six branching fractions which can be written in terms of the theoretical parameters. With the corresponding measurements of \(D^0\rightarrow K^{*0}\rho^0\), \(\bar{D^0}\rightarrow K^{*0}\rho^0\) and \(\bar{D^0}\rightarrow \bar{K^{*0}}\rho^0\) (which are well within the present experimental reach), one has 24 branching fractions written in terms of 27 parameters (discarding \(\alpha_{SM}\)). It is hence not possible to obtain numerical estimates for all the 27 theoretical parameters with the present data set. However with the measurement of another observable in the future, (say the total longitudinal wave), it is possible to extract numerical values for all the theoretical parameters from a fit to data. With the present data set, one can find numerical estimates for the absolute values of the \(A_0\), \(A_{\parallel}\) and the \(A_{\perp}\) amplitudes (and the corresponding observables \(X_{\lambda\lambda}\) from the first relation in ([\[pol_obs\]](#pol_obs){reference-type="ref" reference="pol_obs"})) and the interference phase between \(A_0\) and \(A_{\perp}\) for the \(D^0\rightarrow\bar{K^{*0}}\rho^0\) mode. We have extracted the amplitudes and the interference phase (which we denote by \(\theta\)) via a numerical fit using the package HEPfit . The values of the amplitudes and the phase are displayed in table [1](#tab1){reference-type="ref" reference="tab1"}. The corresponding values thus obtained for the branching ratios are given in table [2](#tab2){reference-type="ref" reference="tab2"}. These values are consistent with experiment within errors. # Summary and conclusions[\[summary\]]{#summary label="summary"} We present a model independent formalism for the detection of NP in \(D^0\rightarrow VV\) decays, which involves a time-dependent analysis of the neutral \(D\) mesons. This formalism has been previously used to analyze the \(B\rightarrow VV\) decays. However the form of the observables change due to the different parametrization of the time dependent \(D^0\) decays. We show that it is possible to construct combinations of these observables which are smoking gun signals for NP. We discuss the relations between the observables in the different cases of the SM, NP in decay, NP in mixing and the most general case where NP manifests itself both in mixing and decay. We also extract the absolute values for the longitudinal, the transverse parallel, the transverse perpendicular amplitudes and the interference phase between the longitudinal and the transverse perpendicular amplitudes for the \(D^0\rightarrow\overline{K^{*0}}\rho^0\) mode. We find that the interference phase is consistent with zero. We thank Nita Sinha and Rahul Sinha for fruitful discussions regarding this work.
{'timestamp': '2016-10-28T02:03:38', 'yymm': '1607', 'arxiv_id': '1607.04837', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04837'}
null
null
# Introduction The inflationary universe scenario in which the early universe undergoes a rapid expansion has been generally accepted as a solution to the horizon problem and some other related problems of the standard big-bang cosmology. The origin of the field that drives inflation is still unknown and is subject to speculations. Among many models of inflation a popular class comprise tachyon inflation models. These models are of particular interest as in these models inflation is driven by the tachyon field originating in string theory. The tachyon potential is derived from string theory and has to satisfy some definite properties to describe tachyon condensation and other requirements in string theory. However, Kofman and Linde have shown that the slow-roll conditions are not compatible with a string coupling much smaller than one, and the compactification length scale much larger than the Planck length. This leads to the density fluctuations produced during inflation being incompatible with observational constraint on the amplitude of the scalar perturbations. This criticism is based on the string theory motivated values of the parameters in the tachyon potential, i.e., the brane tension and the parameters in the four-dimensional Newton constant obtained via conventional string compactification. Of course, if one relaxes the string theory constraints on the above mentioned parameters, the effective tachyon theory will naturally lead to a type of inflation which will slightly deviate from the conventional inflation based on the canonical scalar field theory. Steer and Vernizzi have noted a deviation from the standard single field inflation in the second order consistency relations. Based on their analysis they concluded that the tachyon inflation could not be ruled out by the then available observations. It seems like the present observations could perhaps discriminate between different tachyon models and disfavor or rule out some of these models (for a recent discussion on phenomenological constraints imposed by Planck 2015, see, e.g., ref ). A simple tachyon model can be analyzed in the framework of the second Randall-Sundrum (RSII) model. The RSII model was originally proposed as a possible mechanism for localizing gravity on the 3+1 universe embedded in a 4+1 dimensional spacetime without compactification of the extra dimension. The model is a 4+1 dimensional Anti de Sitter (AdS\(_5\)) universe containing two 3-branes with opposite tensions separated in the fifth dimension: observers reside on the positive tension brane and the negative tension brane is pushed off to infinity. The Planck mass scale is determined by the curvature of the AdS spacetime rather then by the size of the fifth dimension. The fluctuation of the interbrane distance along the extra dimension implies the existence of the so called *radion*--a massless scalar field that causes a distortion of the bulk geometry. In this regard, a stabilization mechanism of the interbrane distance has been proposed by assuming the presence of scalar fields in the bulk. The stabilization mechanism is relevant for the RSI model where the interbrane distance is kept finite. In RSII model, as the negative tension brane is pushed off to infinity the radion disappears. However, it has been shown by Kim, Tupper, and Viollier that a disappearance of the radion in RSII is an artifact of linear theory and hence, when going beyond linear theory the radion remains a dynamical field in the RSII model. Moreover, owing to the radion, the distance between branes remains finite in the RSII limit of infinite coordinate bulk even though the coordinate position of the second brane is infinite. The presence of the radion may have interesting physical implications. The radion field has been proposed as an inflaton. It has been shown that the interaction of the radion with the tachyon can alter the tachyon equation of state: by averaging over large scales the effective equation of state describes the warm dark matter. One of our aims here is to answer the question whether radion could drastically change the early cosmology in the framework of the RSII braneworld model. In this paper we propose a simple tachyon condensate as a model for inflation and we analyze two effects: the coupling of the tachyon with the radion and the modification of the standard cosmology in the RSII scenario. To study these effects we will consider an additional dynamical 3-brane moving in the AdS\(_5\) background of the RSII model. The action of the 3+1 dimensional brane in the five dimensional bulk is equivalent to the Dirac-Born-Infeld (DBI) description of the Nambu-Goto 3-brane.. It is a simple matter to show that this additional 3-brane behaves effectively as a tachyon with the inverse quartic potential. The tachyon model of this kind falls into the class of the power-law tachyon potentials \(V(\theta) \propto \theta^{-n}\) with \(n>2\) which drive a dark-matter attractor : for \(\theta \rightarrow \infty\), the pressure tends to unity very quickly with the unpleasant feature of cold dark matter (CDM) domination at the end of inflation. Actually, this problem is imminent for all tachyon models with the ground state at \(\theta\rightarrow\infty\). The tachyon field rolls towards its ground state without oscillating about it and the conventional reheating mechanism does not work. Typically these scenarios suffer from a reheating problem, since gravitational particle production is not efficient compared to the standard non-gravitational particle production by an oscillating inflaton field. This will be discussed in more detail at the end of Sec. [3](#conditions){reference-type="ref" reference="conditions"}. The remainder of the paper is organized as follows. In Sec. [2](#Sec:Randal){reference-type="ref" reference="Sec:Randal"} we present the RSII model with back-reaction. The system of dynamical equations is presented in a dimensionless form suitable for the calculation of the expansion rate and slow-roll parameters. In Sec. [3](#conditions){reference-type="ref" reference="conditions"} the slow-roll approximation and initial conditions for our model are discussed in detail. Numerical results are presented and discussed in Sec. [4](#Sec:numres){reference-type="ref" reference="Sec:numres"}. In the concluding section, Sec. [5](#conclusions){reference-type="ref" reference="conclusions"}, we summarize our results and give conclusions. # Randall-Sundrum model with back-reaction {#Sec:Randal} The bulk spacetime of the extended RSII model which includes the back-reaction of the radion in Fefferman-Graham coordinates is described by the line element \[ds^2_{(5)}=G_{ab} dX^a dX^b=\frac{1}{k^2 z^2}\left[\left(1+ k^2 z^2\eta(x)\right)g^{\mu\nu}dx^\mu dx^\nu -\frac{1}{\left(1+ k^2 z^2\eta(x)\right)^2} dz^2\right] , \label{eq2001}\] where \(k=1/\ell\) is the inverse of the AdS curvature radius \(\ell\) and \(\eta (x)\) is the radion field. The observer brane is placed at \(z=\ell\) and \(g_{\mu\nu}\) is the metric on the brane. Consider a dynamical 3-brane moving in the bulk. The brane Lagrangian describes the dynamic of the tachyon field modified by the interaction with the radion. After integrating out the fifth coordinate the total effective action is given by \[S = \int d^4x \sqrt{-g} \left(-\frac{R}{16\pi G}+\frac12 g^{\mu\nu}\Phi_{,\mu}\Phi_{,\nu} \right)+ S_{\rm br} , \label{eq3015}\] where \(\Phi\) is the canonically normalized radion field related to \(\eta\) as \[\eta = \sinh^2 \left( \sqrt{\frac{4\pi G}{3}} \Phi \right).\] The brane action \(S_{\rm br}\) is derived in Ref. in terms of the induced metric or the "pull back\" of the bulk space-time metric \(G_{ab}\) to the brane, \[g^{(\rm ind)}_{\mu\nu}=G_{ab} \frac{\partial X^a}{\partial x^\mu} \frac{\partial X^b}{\partial x^\nu} \,. \label{eq0002}\] The action takes the form \[S_{\rm br} =-\sigma \int d^4x \sqrt{-\det g^{\rm ind}_{\mu \nu}} =-\int d^{4}x \sqrt{-g} \, \frac{\sigma}{k^4\Theta^4}(1+k^2\Theta^2 \eta)^2 \sqrt{1-\frac{g^{\mu\nu}\Theta_{,\mu}\Theta_{,\nu}}{(1+k^2\Theta^2\eta)^3}}, \label{eq2006}\] where \(\sigma\) denotes the brane tension and \(\Theta\) is the tachyon field. In the absence of the radion (\(\Phi = 0\)) the brane action is just the tachyon condensate with the inverse quartic potential \[S_{\rm br}^{(0)} =-\int d^{4}x \sqrt{-g} \frac{\lambda}{\Theta^4} \sqrt{1-g^{\mu\nu}\Theta_{,\mu}\Theta_{,\nu}}, \label{tachS}\] where \[\lambda = \frac{\sigma}{k^4}.\] The combined brane-radion Lagrangian reads \[{\cal{L}} = \frac12 g^{\mu\nu}\Phi_{,\mu}\Phi_{,\nu} -\frac{\lambda\psi^2}{\Theta^4}\sqrt{1-\frac{g^{\mu\nu}\Theta_{,\mu}\Theta_{,\nu}}{\psi^3}}, \label{lagrangian}\] where \[\psi=1+k^2 \Theta^2 \eta.\] Hence, in our model, the tachyon is as usual the Dirac-Born-Infeld type scalar field whereas the radion is a canonical scalar. In the following we will assume the spatially flat FRW spacetime on the observer brane with four dimensional line element in the standard form \[ds^2=g_{\mu\nu}dx^\mu dx^\nu=dt^2-a^2(t)(dr^2+r^2 d\Omega^2) \label{eq0012}\] The treatment of our system in a cosmological context is conveniently performed in the Hamiltonian formalism. For this purpose we first define the conjugate momentum fields as \[\Pi_\Phi^\mu= \frac{\partial{\cal{L}}}{\partial\Phi_{,\mu}}, \label{eq2115} \quad \Pi_\Theta^\mu= \frac{\partial{\cal{L}}}{\partial\Theta_{,\mu}}.\] In the cosmological context \(\Pi_\Phi^\mu\) and \(\Pi_\theta^\mu\) are time-like so we may also define their magnitudes as \[\Pi_\Phi=\sqrt{g_{\mu\nu}\Pi_\Phi^\mu\Pi_\Phi^\nu}, \hspace{1cm} \Pi_\theta=\sqrt{g_{\mu\nu}\Pi_\theta^\mu\Pi_\theta^\nu}. \label{eq2118}\] The Hamiltonian density may be derived from the stress tensor corresponding to the Lagrangian ([\[lagrangian\]](#lagrangian){reference-type="ref" reference="lagrangian"}) or by the Legendre transformation. Either way one finds \[{\cal{H}} = \frac{1}{2}\Pi_{\Phi}^2+\frac{\lambda\psi^2}{\Theta^4}\sqrt{1+\Pi_{\Theta}^2\Theta^8/(\lambda^2\psi)}. \label{totH}\] For later use we may also need the Hamiltonian density in terms of \(\Phi_{,\mu}\) and \(\Theta_{,\mu}\), \[{\cal{H}} = \frac12 g^{\mu\nu}\Phi_{,\mu}\Phi_{,\nu} +\frac{\lambda\psi^2}{\Theta^4}\left(1-\frac{g^{\mu\nu}\Theta_{,\mu}\Theta_{,\nu}}{\psi^3}\right)^{-1/2}. \label{eq014}\] Next, we can write Hamilton's equations in the form \[\begin{aligned} \dot{\Phi} = \frac{\partial{\cal{H}}}{\partial\Pi_\Phi},\label{eqHam1}\\ \dot{\Theta} = \frac{\partial{\cal{H}}}{\partial\Pi_\Theta},\label{eqHam2} \\ \dot{\Pi}_\Phi +3H\Pi_\Phi=-\frac{\partial{\cal{H}}}{\partial\Phi},\label{eqHam3}\\ \dot{\Pi}_\Theta + 3H\Pi_\Theta=-\frac{\partial{\cal{H}}}{\partial\Theta}. \label{eqHam4} \end{aligned}\] In the spatially flat Randall-Sundrum cosmology the Hubble expansion rate \(H\) is related to the Hamiltonian via the modified Friedmann equation \[\begin{aligned} H\equiv\frac{\dot{a}}{a}=\sqrt{\frac{8 \pi G}{3} \mathcal{H}\left(1+ \frac{2 \pi G}{3k^2} \mathcal{H}\right) }. \label{scale_a} \end{aligned}\] In addition, we will make use of the energy-momentum conservation equation \[\dot{\mathcal{H}}+3H(\mathcal{H}+\mathcal{L})=0 , \label{3200}\] which, combined with the time derivative of ([\[scale_a\]](#scale_a){reference-type="ref" reference="scale_a"}), yields the second Friedmann equation in the form \[\dot{H}=-4\pi G(\mathcal{H}+\mathcal{L})\left(1+ \frac{4 \pi G}{3k^2} \mathcal{H}\right). \label{eq3222}\] Thus, the Friedman equations are modified in the RSII cosmology. As far as we know the effects of these modifications on tachyon inflation were first studied by Bento, Bertolami and Sen. To solve the system of equations ([\[eqHam1\]](#eqHam1){reference-type="ref" reference="eqHam1"})-([\[scale_a\]](#scale_a){reference-type="ref" reference="scale_a"}), it is convenient to rescale the time as \(t=\tau/k\) and express the system in terms of dimensionless quantities. Besides, we can eliminate the coupling constant \(\lambda\) from the equations by appropriately rescaling the fields \(\Phi\) and \(\Theta\) and their conjugate fields \(\Pi_{\Phi}\) and \(\Pi_{\Theta}\). To this end we introduce the dimensionless functions \[\begin{aligned} h = H/k, \quad \phi = \Phi/(k \sqrt{\lambda}),\quad \pi_\phi = \Pi_\Phi/(k^2\sqrt{\lambda})),\qua \theta=k \Theta, \quad \pi_\theta = \Pi_{\Theta}/(k^4 \lambda), \label{eq002} \end{aligned}\] and rescale the Lagrangian and Hamiltonian to obtain the dimensionless pressure and energy density: \[\bar{p}\equiv \frac{\mathcal{L}}{k^4\lambda} =\frac{1}{2}\pi_\phi^2-\frac{\psi^2}{\theta^4} \frac{1}{\sqrt{1+\theta^8\pi_\theta^2/\psi}} , \label{eq504}\] \[\bar{\rho}\equiv \frac{\mathcal{H}}{k^4\lambda}=\frac{1}{2}\pi_\phi^2+\frac{\psi^2}{\theta^4} \sqrt{1+\theta^8\pi_\theta^2/\psi}. \label{eq505}\] Following Steer and Vernizzi we also introduce a combined dimensionless coupling \[\kappa^2=8\pi\lambda G k^2. \label{eq102}\] Then, from ([\[eqHam1\]](#eqHam1){reference-type="ref" reference="eqHam1"})-([\[scale_a\]](#scale_a){reference-type="ref" reference="scale_a"}) we obtain the following set of equations \[\label{sysRT1} \dot \phi=\pi_\phi,\] \[\label{sysRT2} \dot \theta=\frac{\theta^4\psi\pi_{\theta}} {\sqrt{1+\theta^8\pi_{\theta}^2/\psi}},\] \[\label{sysRT3} \dot \pi_\phi=-3h\pi_\phi-\frac{\psi}{2\theta^2}\,\frac{4+3\theta^8\pi_{\theta}^2/\psi} {\sqrt{1+\theta^8\pi_{\theta}^2/\psi}}\eta',\] \[\label{sysRT4} \dot \pi_{\theta}=-3h{{\pi }_{\theta}} +\frac{\psi}{\theta^5}\,\frac{4-3\theta^{10}\eta\pi_{\theta}^2/\psi} {\sqrt{1+\theta^8\pi_{\theta}^2/\psi}},\] where \[\begin{aligned} \label{h} h\equiv \frac{\dot{a}}{a}=\sqrt{\frac{\kappa^2}{3}\bar{\rho}\left(1+\frac{\kappa^2}{12}\bar{\rho} \right)}, \end{aligned}\] \[\psi=1+\theta^2 \eta,\] \[\eta = \sinh^2 \left( \sqrt{\frac{\kappa^2}{6}} \phi \right) ,\] \[\eta^{\prime}=\frac{d\eta}{d\phi} = \sqrt{\frac{\kappa^2}{6}} \sinh \left( \sqrt{\frac{2 \kappa^2}{3}} \phi \right). \label{eq012}\] In addition to Eqs. ([\[sysRT1\]](#sysRT1){reference-type="ref" reference="sysRT1"})-([\[h\]](#h){reference-type="ref" reference="h"}) we can solve in parallel the second Friedman equation \[\begin{aligned} \dot{h}=-\frac{\kappa^2}{2}(\bar{\rho}+\bar{p})\left(1+\frac{\kappa^2}{6}\bar{\rho} \right), \label{h2} \end{aligned}\] and \[\dot{N}=h \label{n}\] where \(N\) is the number of e-folds as a function of \(\tau\). In Eqs. ([\[sysRT1\]](#sysRT1){reference-type="ref" reference="sysRT1"})-([\[n\]](#n){reference-type="ref" reference="n"}) and from now on the overdot denotes a derivative with respect to \(\tau\). Obviously, the explicit dependence on \(\lambda\) in Eqs. ([\[sysRT1\]](#sysRT1){reference-type="ref" reference="sysRT1"})-([\[eq012\]](#eq012){reference-type="ref" reference="eq012"}) is eliminated and the only remaining free parameter is \(\kappa\). Equations ([\[sysRT1\]](#sysRT1){reference-type="ref" reference="sysRT1"}) and ([\[sysRT2\]](#sysRT2){reference-type="ref" reference="sysRT2"}) can be used to express the pressure and energy density in terms of \(\dot{\phi}\) and \(\dot{\theta}\). One thus finds \[\bar{p}=\frac{1}{2}\dot{\phi}^2-\frac{\psi^2}{\theta^4} \sqrt{1-\dot{\theta}^2/\psi^3} , \label{eq015}\] \[\bar{\rho}=\frac{1}{2}\dot{\phi}^2+\frac{\psi^2}{\theta^4} \frac{1}{\sqrt{1-\dot{\theta}^2/\psi^3}}. \label{eq016}\] The functional dependence of \(h\) on \(\tau\) will be used to calculate the slow-roll parameters and the number of e-folds. The slow-roll parameters are defined as \[\epsilon_{i} \equiv \frac{d\ln| \epsilon_{i-1}|}{Hdt},\qquad i \geq 1,\] where \[\epsilon_0 \equiv \frac{H_*}{H}.\] and \(H_*\) is the Hubble rate at an arbitrarily chosen time. Using the previously defined dimensionless Hubble rate \(h\), the first two parameters can be written as \[\label{eps1} \epsilon_1 =-\frac{\dot{h}}{h^2},\] \[\label{eps2} \epsilon_2 = 2\epsilon_1+\frac{\ddot{h}}{h\dot{h}}.\] The conditions for a slow-roll regime are satisfied when \(\epsilon_1 < 1\) and \(\epsilon_2 < 1\), and inflation ends when any of them exceeds unity. The effect of the radion and the tachyon can be seen if we compare the slow-roll parameters for the full model with those for the model with inverse quartic tachyon potential. The number of e-folds is defined through Eq. ([\[n\]](#n){reference-type="ref" reference="n"}) as \[N=\int_{\tau_{\rm i}}^{\tau_{\rm f}}d\tau h,\] where the subscripts \({\rm i}\) and \({\rm f}\) denote the beginning and the end of inflation, respectively. The slow-roll parameters are related to observable quantities, in particular to the tensor-to-scalar ratio \(r\) and the scalar spectral index \(n_{\rm s}\) defined by \[r=\frac{\mathcal{P}_{\rm T}}{\mathcal{P}_{{\rm S}}}, \label{eq3005}\] \[n_{\rm s}= \frac{d\ln \mathcal{P}_{{\rm S}}}{d\ln q}, \label{eq3006}\] where \(\mathcal{P}_{{\rm S}}\) and \(\mathcal{P}_{\rm T}\) are the power spectra of scalar and tensor perturbations, respectively, evaluated at the horizon, i.e., for a wave-number satisfying \(q=aH\). Calculation of the spectra proceeds by identifying the proper canonical field and imposing quantization of the quadratic action for the near free field. The procedure for a general k-inflation is described in and applied to the tachyon fluid in Refs.. It turns out that at the lowest order in \(\epsilon_1\) and \(\epsilon_2\), the power spectra may be expressed in the same way as in the standard tachyon inflation : \[\mathcal{P}_{\rm T} \simeq [1-2(1+C)\epsilon_1]\frac{16GH^2}{\pi}, \label{eq3007}\] \[\mathcal{P}_{{\rm S}} \simeq [1-2(1+C-\alpha)\epsilon_1-C \epsilon_2]\frac{GH^2}{\pi\epsilon_1 }, \label{eq3008}\] where \(C=-2+\ln 2 +\gamma\simeq-0.72\), and \(\alpha\) is a parameter related to the speed of sound expanded in \(\epsilon_1\): \[c_{\rm s}=1-2\alpha\epsilon_1 +O(\epsilon_1^2). \label{eq3009}\] The parameter \(\alpha\) will be calculated in the next section using the relation between \(\epsilon_1\) and the speed of sound in the slow-roll approximation. # Conditions for tachyon inflation {#conditions} ## Inverse quartic potential To solve the system of equations ([\[sysRT1\]](#sysRT1){reference-type="ref" reference="sysRT1"})-([\[sysRT4\]](#sysRT4){reference-type="ref" reference="sysRT4"}) we need to choose initial conditions relevant for inflation. To this end we first solve a simpler case in which the radion is absent. The model is described by the pure tachyon Lagrangian with the inverse quartic potential \[{\cal{L}} = -\frac{\lambda}{\Theta^4}\sqrt{1-g^{\mu\nu}\Theta_{,\mu}\Theta_{,\nu}} \,, \label{eq000}\] with the corresponding Hamiltonian \[{\cal{H}} =\frac{\lambda}{\Theta^4}\sqrt{1+\Pi_{\Theta}^2\Theta^8/(\lambda^2)}. \label{eq001}\] As before, we rescale the time as \(t=\tau/k\), absorb the coupling constant \(\lambda\) into the conjugate field \(\Pi_{\Theta}\) and introduce the dimensionless functions as in ([\[eq002\]](#eq002){reference-type="ref" reference="eq002"}). Then, using ([\[eqHam2\]](#eqHam2){reference-type="ref" reference="eqHam2"}), ([\[eqHam4\]](#eqHam4){reference-type="ref" reference="eqHam4"}) and ([\[eq001\]](#eq001){reference-type="ref" reference="eq001"}) we obtain the Hamilton equations in the form \[\dot \theta=\frac{\theta^4\pi_{\theta}} {\sqrt{1+\theta^8\pi_{\theta}^2}} \label{eq003}\] \[\dot \pi_\theta=-3h\pi_\theta +\frac{4}{\theta^5 \sqrt{1+\theta^8\pi_\theta^2}} \label{eq004}\] The dimensionless pressure and energy density are given by \[\bar{p}=-\frac{1}{\theta^4\sqrt{1+\theta^8\pi_{\theta}^2}}= -\frac{1}{\theta^4}\sqrt{1-\dot{\theta}^2}, \label{eq008}\] \[\bar{\rho}= \frac{1}{\theta^4}\sqrt{1+\theta^8\pi_{\theta}^2}= \frac{1}{\theta^4}\frac{1}{\sqrt{1-\dot{\theta}^2}}, \label{eq2008}\] Using this we also find a simple expression for the sound speed \[c_{\rm s}^2\equiv \left.\frac{\partial p}{\partial\rho}\right|_{\theta} =\frac{1}{1+\theta^8\pi_{\theta}^2}= 1-\dot{\theta}^2. \label{eq2009}\] The Friedmann equations are given by Eq. ([\[h\]](#h){reference-type="ref" reference="h"}) and ([\[h2\]](#h2){reference-type="ref" reference="h2"}) with ([\[eq008\]](#eq008){reference-type="ref" reference="eq008"}) and ([\[eq2008\]](#eq2008){reference-type="ref" reference="eq2008"}). Hence, as before, the coupling \(\lambda\) drops out and the only remaining parameter is \(\kappa\) defined in ([\[eq102\]](#eq102){reference-type="ref" reference="eq102"}). The system of equations ([\[eq003\]](#eq003){reference-type="ref" reference="eq003"})-([\[eq004\]](#eq004){reference-type="ref" reference="eq004"}) can be solved numerically for a chosen set of initial conditions. Once the solution for \(H\) is found the slow-roll parameters can be easily calculated. ## Slow-roll approximation and initial conditions {#initial} To find appropriate initial value of the field \(\theta\), we first consider the pure tachyon model in the slow-roll approximation. Tachyon inflation is based upon the slow evolution of \(\theta\) with the slow-roll conditions \[\dot{\theta}\simeq \theta^4\pi_\theta \ll 1, \quad \dot\pi_\theta \ll 3h \pi_\theta,\] so that in the slow-roll approximation we may neglect the factors \((1-\dot{\theta}^2)^{1/2}\). Then, during inflation we have \[h\simeq \frac{\kappa}{\sqrt3 \theta^2}\left(1+\frac{\kappa^2}{12\theta^4}\right)^{1/2}, \label{eq1007}\] \[\dot{\theta} \simeq \frac{4}{3h\theta} \simeq \frac{4\theta }{\sqrt3 \kappa} \left(1+\frac{\kappa^2}{12\theta^4}\right)^{-1/2} , \label{eq1008}\] \[\ddot{\theta} \simeq \frac{4\dot{\theta} }{\sqrt3 \kappa} \left(1+\frac{\kappa^2}{12\theta^4}\right)^{-3/2} \left(1+\frac{\kappa^2}{4\theta^4}\right), \label{eq1009}\] and using ([\[h2\]](#h2){reference-type="ref" reference="h2"}) we also find \[\dot{h}=-\frac{\kappa^2 \dot{\theta}^2}{2\theta^4}\left(1+\frac{\kappa^2}{6\theta^4}\right).\] As a consequence, the slow-roll parameters ([\[eps1\]](#eps1){reference-type="ref" reference="eps1"}) and ([\[eps2\]](#eps2){reference-type="ref" reference="eps2"}) can be approximated by \[\begin{aligned} \epsilon_1&\simeq&\frac32 \dot{\theta}^2\left(1+\frac{\kappa^2}{6\theta^4}\right) \left(1+\frac{\kappa^2}{12\theta^4}\right)^{-1} \nonumber \\ &\simeq&\frac{8\theta^2}{\kappa^2}\left(1+\frac{\kappa^2}{6\theta^4}\right) \left(1+\frac{\kappa^2}{12\theta^4}\right)^{-2}, \label{eq009} \end{aligned}\] \[\begin{aligned} \epsilon_2 &\simeq& 2\frac{\ddot{\theta}}{h\dot{\theta}}-\dot{\theta}^2\frac{\kappa^2}{4\theta^4} \left(1+\frac{\kappa^2}{6\theta^4}\right)^{-1}\left(1+\frac{\kappa^2}{12\theta^4}\right)^{-1} \nonumber \\ &\simeq& \frac{8\theta^2}{\kappa^2} \left(1+\frac{\kappa^2}{12\theta^4}\right)^{-2}\left[ 1+\frac{\kappa^2}{4\theta^4}-\frac{\kappa^2}{6\theta^4}\left(1+\frac{\kappa^2}{6\theta^4}\right)^{-1} \right]. \label{eq010} \end{aligned}\] In the slow roll regime we have \(\kappa^2/\theta^4\gg 1\), so the corrections due to the RSII modification in Eqs. ([\[eq1007\]](#eq1007){reference-type="ref" reference="eq1007"})-([\[eq010\]](#eq010){reference-type="ref" reference="eq010"}) will dominate over unity and we find \[h\simeq \frac16 \frac{\kappa^2}{\theta^4}, \quad\quad \dot{\theta}\simeq 8 \frac{\theta^3 }{ \kappa^2} , \quad\quad \ddot{\theta}\simeq 24 \frac{\theta^2}{\kappa^2}\dot{\theta}, \label{eq021}\] \[\epsilon_1\simeq 3 \dot{\theta}^2\simeq 192 \frac{\theta^6}{\kappa^4}, \quad\quad \epsilon_2\simeq 288 \frac{\theta^6}{\kappa^4} \simeq \frac32\epsilon_1. \label{eq018}\] In contrast, if we disregarded the RSII corrections we would obtain the usual slow-roll equations of tachyon inflation for the potential \(V=\lambda/ \theta^{4}\) \[h\simeq \frac{1}{\sqrt{3}} \frac{\kappa}{\theta^2}, \quad\quad \dot{\theta}\simeq \frac{4}{\sqrt{3}} \frac{\theta}{ \kappa}, \quad\quad \ddot{\theta}\simeq \frac{4}{\sqrt{3}} \frac{\dot{\theta}}{\kappa}, \label{eq0021}\] \[\epsilon_1\simeq \frac32 \dot{\theta}^2\simeq 8 \frac{\theta^2}{\kappa^2}, \quad\quad \epsilon_2\simeq \epsilon_1. \label{eq0018}\] Hence, in the slow-roll regime the tachyon inflation in the RSII modified cosmology proceeds in a quite distinct way compared with that in the standard FRW cosmology. However, close to and at the end of inflation we have \(\kappa^2/\theta_{\rm f}^4\ll 1\) and we can neglect the RSII cosmology corrections. Hence, the expressions ([\[eq0021\]](#eq0021){reference-type="ref" reference="eq0021"}) and ([\[eq0018\]](#eq0018){reference-type="ref" reference="eq0018"}) can be used at the end of inflation where we find \[\epsilon_1(\theta_{\rm f})\simeq \epsilon_2(\theta_{\rm f})\simeq \frac{8\theta_{\rm f}^2}{\kappa^2} \simeq 1, \label{eq017}\] and \[h(\theta_{\rm f})\simeq \frac{8}{\sqrt3 \kappa}. \label{eq2007}\] In the slow-roll approximation the number of e-folds is given by \[\begin{aligned} \label{inteq} N \simeq\frac{\kappa^2}{4} \int_{\theta_0}^{\theta_{\rm{f}}} \frac{d\theta}{\theta^3} \left(1+\frac{\kappa^2}{12\theta^4}\right) \simeq \frac{\kappa^2}{8\theta_0^2}\left(1+\frac{\kappa^2}{36\theta_0^4}\right)-1 \simeq \frac23 \frac{1}{\epsilon_1(\theta_0)}-1, \end{aligned}\] where we have exploited \(\kappa^2/(36\theta_0^4)\gg 1\) at the beginning and \(\kappa^2/(36\theta_{\rm f}^4)\ll 1\) at the end of inflation together with the condition ([\[eq017\]](#eq017){reference-type="ref" reference="eq017"}). For comparison, in the standard tachyon inflation described by ([\[eq0021\]](#eq0021){reference-type="ref" reference="eq0021"}) and ([\[eq0018\]](#eq0018){reference-type="ref" reference="eq0018"}) we would obtain \[N_{\rm st.tach} \simeq\frac{\kappa^2}{8\theta_0^2}-1 \simeq \frac{1}{\epsilon_1(\theta_0)}-1. \label{eq319}\] For example, the choice \(\kappa^2=5\) and \(\theta_0 = 0.25\) leads to \(N_{\rm st.tach} \simeq 9\) whereas in RSII cosmology with the same parameters one finds \(N \simeq 365\). At this point it is convenient to estimate a phenomenologically acceptable range of the couplings \(\lambda\) and \(\kappa\). Although the evolution equations do not depend on \(\lambda\), its approximate value is needed for choosing appropriate initial conditions for the radion field. The value of \(\lambda\) may be estimated using the observational constraint on the amplitude of scalar perturbations. The approximate expression \[\mathcal{P}_{\rm S}\simeq \frac{GH^2}{\pi\epsilon_1},\] which follows from ([\[eq3008\]](#eq3008){reference-type="ref" reference="eq3008"}), is to be compared with the power spectrum amplitude \(A_s\) measured by Planck : \[A_s\simeq 2.2\times 10^{-9}.\] Hence, we must make sure that the condition \[\frac{H}{M_{\rm P}} \lesssim \sqrt{\pi A_s}\simeq 8.31\times 10^{-5}\] is satisfied close to and at the end of inflation (where \(\epsilon_1 \lesssim 1\)). (See also Ref.)̇ Here we define the Planck mass as \(M_{\rm P}=G^{-1/2}\). According to Eq. ([\[eq2007\]](#eq2007){reference-type="ref" reference="eq2007"}) in the slow-roll approximation near the end of inflation we have \[\frac{H}{M_{\rm P}}\simeq \frac{8}{\sqrt{3}} \frac{k}{\kappa M_{\rm P}}= \sqrt{\frac{8 }{3\pi}}\frac{k^2}{\sqrt{\sigma}} , \label{eq03}\] yielding \[\frac{k}{\sigma^{1/4}}\lesssim 10^{-2}, \label{eq01}\] or \[\lambda\gtrsim 10^{8}. \label{eq02}\] To estimate \(\kappa\) note first that the tension of a D\(p\)-brane is given by \[\sigma =\frac{1}{(2\pi)^p \alpha^{\prime (p+1)/2} g_{\rm s}},\] where \(g_{\rm s}\) is the string coupling constant and \(1/(2\pi\alpha^{\prime})\) is the string tension. Using this for \(p=3\) from ([\[eq01\]](#eq01){reference-type="ref" reference="eq01"}) we find a constraint \[g_{\rm s}\lesssim 4\times 10^{-11}\left(\frac{M_{\rm s}}{k}\right)^4 ,\] where \(M_{\rm s}=1/\sqrt{\alpha^{\prime}}\). Using this constraint we can choose \(k\) and \(M_{\rm s}\) such that the scale hierarchy \[H < k < M_{\rm s} < M_{\rm P}\] is satisfied. With this in mind we can give an order of magnitude estimate for an acceptable range of values of our free parameter \(\kappa\). Requiring \(k\gtrsim H\), from ([\[eq03\]](#eq03){reference-type="ref" reference="eq03"}) it follows \[\kappa\gtrsim 8/\sqrt{3},\] so we can safely choose \(\kappa\) to vary in the range \(1<\kappa<20\). Equations ([\[eq018\]](#eq018){reference-type="ref" reference="eq018"}) and ([\[inteq\]](#inteq){reference-type="ref" reference="inteq"}) can be used to estimate the value of the ratio \(\kappa^2/\theta_0^3\) by fitting the model to the observational parameters \(r\) and \(n_{\rm s}\) defined by ([\[eq3005\]](#eq3005){reference-type="ref" reference="eq3005"}) and ([\[eq3006\]](#eq3006){reference-type="ref" reference="eq3006"}). First, using ([\[eq2009\]](#eq2009){reference-type="ref" reference="eq2009"}) and ([\[eq018\]](#eq018){reference-type="ref" reference="eq018"}) we can express the sound speed in terms of \(\epsilon_1\): \[c_s \simeq1-\frac16 \epsilon_1,\] and comparing this with ([\[eq3009\]](#eq3009){reference-type="ref" reference="eq3009"}) we find \(\alpha=1/12\). In contrast, in the standard tachyon inflation the value \(\alpha=1/6\) is obtained. Using ([\[eq3007\]](#eq3007){reference-type="ref" reference="eq3007"}) and ([\[eq3008\]](#eq3008){reference-type="ref" reference="eq3008"}) we find up to the second order \[r=16 \epsilon_1\left[1-\frac16 \epsilon_1+C\epsilon_2 \right], \label{eq005}\] \[n_{\rm s}=1-2 \epsilon_1-\epsilon_2-\left[2 \epsilon_1^2+\left(2C +\frac{17}{6} \right)\epsilon_1\epsilon_2 +C\epsilon_2\epsilon_3\right], \label{eq006}\] where it is understood that \(\epsilon_1\) and \(\epsilon_2\) take their values at or close to the beginning of inflation. Using these equations at linear order with \(\epsilon_2\simeq 3\epsilon_1/2\) and ([\[inteq\]](#inteq){reference-type="ref" reference="inteq"}), we find the approximate relations \[r=\frac{32}{3}\frac{1}{N+1} , \label{eq501}\] \[n_{\rm s}=1-\frac73 \frac{1}{N+1}, \label{eq502}\] \[r=\frac{32}{7}(1-n_{\rm s}). \label{eq503}\] Had we used the usual relation \(\epsilon_2\simeq \epsilon_1\) and ([\[eq319\]](#eq319){reference-type="ref" reference="eq319"}) we would have obtained the standard tachyon-inflation relations \[r=\frac{16}{N+1},\] \[n_{\rm s}=1-\frac{3}{N+1},\] \[r=\frac{16}{3} (1-n_{\rm s}). \label{eq5004}\] Comparing \(r\) and \(n_{\rm s}\) with the latest observations we can fix the parameters \(\epsilon_1\), \(\epsilon_2\), and \(N\). Then, from ([\[inteq\]](#inteq){reference-type="ref" reference="inteq"}) we can determine the ratio \(\kappa^2/\theta_0^3\) and for a chosen set of values of \(\kappa\) we find the corresponding \(\theta_0\). These values of \(\kappa\) and the corresponding initial values \(\theta_0\) as initial conditions are then used to solve the system of equation ([\[eq003\]](#eq003){reference-type="ref" reference="eq003"})-([\[eq004\]](#eq004){reference-type="ref" reference="eq004"}) for the pure tachyon and ([\[sysRT1\]](#sysRT1){reference-type="ref" reference="sysRT1"})-([\[sysRT4\]](#sysRT4){reference-type="ref" reference="sysRT4"}) for the tachyon-radion system. There is no a priori reason to restrict possible initial values of the radion field \(\Phi\) so we can choose a range of initial values based on the natural scale dictated by observations, i.e., the scale between \(H\) and \(M_{\rm P}\). Hence, a natural initial value for \(\Phi\) would be of the order of \(k\) or a few orders larger, say in the range 10 to 1000 \(k\). However, according to ([\[eq002\]](#eq002){reference-type="ref" reference="eq002"}) the dimensionless radion field \(\phi\) is rescaled with respect to \(\Phi/k\) by a factor of \(1/\sqrt{\lambda}\simeq 10^{-4}\) so we can choose the initial value \(\phi_0=\phi(0)\) in the range from 0.001 to 0.5. As we have mentioned in Introduction our tachyon model suffers from the so called reheating problem. A possible way out is provided by string theory. String theory D-branes couple to the (pull-back of) antisymmetric tensor field \(\mathcal{B}\) that combines the Kalb-Ramond and electromagnetic fields. In this way there exist a natural coupling between the tachyon and the electromagnetic field. Therefore, this interaction could serve as a possible reheating mechanism at the end of inflation. In a tachyon model based on brane-antibrane annihilation resulting in a time dependent tachyon condensate it has been shown that a coupling of massless fields to the time dependent tachyon condensate could yield a reheating efficient enough to overcome the above mention problem of a CDM dominance. Another possible way out of the reheating problem could be the so called *warm inflation*. Warm inflation is an alternative inflation scenario with no need for a reheating period. In warm inflation, dissipative effects are included during inflation, so that radiation is produced in parallel with the inflationary expansion and inflation ends when the universe heats up to become radiation dominated. This scenario has been successfully applied to tachyon inflation models and, in principle, should also work for our model. This requires further investigation which goes beyond the scope of the present paper. # Numerical results {#Sec:numres} The system of equations ([\[sysRT1\]](#sysRT1){reference-type="ref" reference="sysRT1"})-([\[sysRT4\]](#sysRT4){reference-type="ref" reference="sysRT4"}) is evolved numerically starting from \(\tau=0\) up to some large \(\tau\) of the order of 100. The initial values \(\theta_0\) and \(\phi_0\) are chosen as described in the previous section and the initial conjugate momenta are taken to be \(\pi_{\theta 0}=\pi_{\phi 0}=0\). The function \(N(\tau)\) is solved simultaneously using ([\[n\]](#n){reference-type="ref" reference="n"}) with \(N(0)=0\). The time evolution of the slow roll parameters \(\epsilon_1\) and \(\epsilon_2\) are obtained using ([\[eps1\]](#eps1){reference-type="ref" reference="eps1"}) and ([\[eps2\]](#eps2){reference-type="ref" reference="eps2"}). The inflation ends at a point \(\tau_{\rm f}\) at which \(\epsilon_1(\tau_{\rm f})=1\). The beginning of inflation at \(\tau_i\) is then found by requiring \(N(\tau_{\rm f})-N(\tau_{\rm i})=N\). The results for \(\kappa=2\), \(\theta_0=0.25\) and \(\phi_0=0.4\) are presented in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"} together with the results calculated for the system ([\[eq003\]](#eq003){reference-type="ref" reference="eq003"})-([\[eq004\]](#eq004){reference-type="ref" reference="eq004"}) within the standard FRW cosmology with the same initial \(\theta_0=0.25\) and \(\pi_{\theta 0}=0\). To calculate the quantities \(n_s\) and \(r\) we proceed as follows. For a chosen pair of \((N, \kappa)\) we first find the initial value \(\theta_0\) from Eq. ([\[inteq\]](#inteq){reference-type="ref" reference="inteq"}). Then, for a chosen set of initial values \(\phi_0\), \(\pi_{\theta 0}\), \(\pi_{\phi 0}\) (at \(\tau=0\)) we find the corresponding \(\phi_{\rm i}\), \(\pi_{\theta \rm i}\), \(\pi_{\phi \rm i}\) at \(\tau=\tau_{\rm i}\) and calculate \(h\), \(\dot{h}\) and \(\ddot{h}\) using equations ([\[sysRT1\]](#sysRT1){reference-type="ref" reference="sysRT1"})-([\[h\]](#h){reference-type="ref" reference="h"}) with ([\[eq504\]](#eq504){reference-type="ref" reference="eq504"}) and ([\[eq505\]](#eq505){reference-type="ref" reference="eq505"}). From this we find \(\epsilon_1(\theta_{\rm{i}})\) and \(\epsilon_2(\theta_{\rm{i}})\) using the defining expressions ([\[eps1\]](#eps1){reference-type="ref" reference="eps1"}) and ([\[eps2\]](#eps2){reference-type="ref" reference="eps2"}) and calculate the parameters \(r\) and \(n_{\rm s}\) from ([\[eq005\]](#eq005){reference-type="ref" reference="eq005"}) and ([\[eq006\]](#eq006){reference-type="ref" reference="eq006"}), respectively. In Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"} we present \(n_s\) and \(r\) as functions of \(\kappa\) and \(N\) for a fixed \(\phi_0=0.01\). In Fig. [\[fig3\]](#fig3){reference-type="ref" reference="fig3"} we present the \(n_s-r\) diagram with 10000 points superimposed on the observational constraints taken from the Planck Collaboration 2015. Each point in the diagram corresponds to a set \((N, \kappa, \phi_0)\) chosen randomly in the range \(60 \leq N \leq 120\), \(1\leq \kappa \leq 12\) and \(0\leq \phi_0 \leq 0.5\). To obtain a more favorable distribution of points within the 2\(\sigma\) area measured by the Planck collaboration we have used the distribution histograms of the number of e-folds \(N\) and the parameters \(\kappa\) and \(\phi_0\). In this way we have been guided to restrict \(N\) to range between 85 and 110, \(\kappa\) between 1 and 8, and \(\phi_0\) between 0 and 0.5. The outcome of these constraints is presented in Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}. In this figure one notices basically two bands: the dominant one almost parallel to the \(r\) axis with most data points concentrated close to the line corresponding to the RSII model prediction without radion and the less dens one almost parallel to the \(n_s\) axis, with \(r\) between 0.1 and 0.13. From a quite large data set of numerical results we find that smaller values of \(\kappa\) give a better agreement with observational data from the Planck mission. For example, if we fix \(\kappa=2\) and vary \(N\) in a wider range \(N\) \(60\leq N \leq 120\)) and \(\phi_0\) in a bit narrower range \(0 \leq\phi_0 \leq 0.25\)) we obtain an interesting splitting of points into three disconnected clusters (Fig. [\[fig5\]](#fig5){reference-type="ref" reference="fig5"}). This splitting is a clear manifestation of the nonlinearity of Eqs. ([\[sysRT1\]](#sysRT1){reference-type="ref" reference="sysRT1"})-([\[sysRT4\]](#sysRT4){reference-type="ref" reference="sysRT4"}). The most pronounced cluster of numerical points fits in quite well within the 2\(\sigma\) regions of constraints given by the Planck collaboration. A still better agreement between our model and the observational data constraints is obtained for a very narrow range of parameters as shown in Figs. [\[fig6\]](#fig6){reference-type="ref" reference="fig6"} and [\[fig7\]](#fig7){reference-type="ref" reference="fig7"}. The best fit is obtained for \(\kappa=1.25\), \(\phi_0=0.05\) and \(115 \leq N \leq 120\). # Summary and conclusions {#conclusions} We have investigated a model of inflation based on the dynamics of a D3-brane in the AdS\(_5\) bulk of the RSII model. The bulk metric is extended to include the back reaction of the radion excitations. We have shown that the slow-roll equations of the tachyon inflation are quite distinct to those of the standard tachyon inflation with the same potential. In particular, the departure of the sound speed from unity equals \(c_{\rm s}-1\simeq-\epsilon_1/6\) in contrast to the standard result \(c_{\rm s}-1\simeq-\epsilon_1/3\). The \(n_{\rm s}-r\) relation in our model is substantially different from the standard one and is closer to the best observational value, as shown in Fig. [\[fig6\]](#fig6){reference-type="ref" reference="fig6"}. Note that the largest concentration of numerical results, represented by the points in the plot in Fig. [\[fig6\]](#fig6){reference-type="ref" reference="fig6"}, are within 2\(\sigma\) of the *Planck* TT+lowP and *Planck* TT+lowP+lensing+BAO+JLA+H\(_0\) (red and light blue shaded regions, respectively). Clearly, the agreement with observations is not ideal and it is fair to say that the present model is disfavored but not excluded. However, one should bare in mind that the model is based on the brane dynamics which results in a definite potential with one free parameter only. In contrast, the majority of other tachyon potentials discussed in the literature are chosen arbitrarily and adjusted so that the results agree with observations. In this work we have analyzed the simplest tachyon model that stems from the dynamics of a D3-brane in an AdS\(_5\) bulk yielding basically an inverse quartic potential. In principle, the same mechanism could lead to a more general tachyon potential if the AdS\(_5\) background metric is deformed by the presence of matter in the bulk, e.g., in the form of a minimally coupled scalar field with an arbitrary self-interaction potential. This will be the subject of our investigation in the future.
{'timestamp': '2017-02-21T02:11:47', 'yymm': '1607', 'arxiv_id': '1607.04524', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04524'}
# Introduction {#sec1} Optimization problems are prevalent and have held great importance throughout history in engineering applications and scientific endeavors. For instance, many problems in the field of artificial intelligence (AI) can be viewed as optimization problems. Accordingly, generic *local* optimization methods, such as hill climbing and the gradient method, have been successfully adopted to solve AI problems since early research on the topic. On the other hand, the application of *global* optimization to AI problems has been studied much less despite its practical importance. This is mainly due to the lack of necessary computational power in the past and the absence of a practical global optimization method with a strong theoretical basis. Of these two obstacles, the former is becoming less serious today, as evidenced by a number of studies on global optimization in the past two decades. The aim of this paper is to partially address the latter obstacle. The inherent difficulty of the global optimization problem has led to two distinct research directions: development of heuristics without theoretically guaranteed performance and advancement of theoretically supported methods regardless of its difficulty. A degree of practical success has resulted from heuristic approaches such as simulated annealing, genetic algorithms (for a brief introduction on the context of AI, see ), and swarm-based optimization (for an interesting example of a recent study, see ). Although these methods are heuristics without strong theoretical supports, they became very popular partly because their optimization mechanisms aesthetically mimic nature's physical or biological optimization mechanism. On the other hand, the Lipschitzian approach to global optimization aims to accomplish the global optimization task in a theoretically supported manner. Despite its early successes in theoretical viewpoints, the early studies were based on an assumption that is impractical in most applications: the Lipschitz constant, which is the bound of the slope on the objective function, is known. The relaxation of this crucial assumption resulted in the well-known DIRECT algorithm that has worked well in practice, yet guarantees only consistency property. Recently, the Simultaneous Optimistic Optimization (SOO) algorithm achieved the guarantee of a finite-time error bound without knowledge of the Lipschitz constant. However, the practical performance of the algorithm is unclear. In this paper, we propose a generic global optimization algorithm that is aimed to achieve both a satisfactory performance in practice and a finite-loss bound as the theoretical basis without strong additional assumption
{'timestamp': '2016-07-19T02:05:57', 'yymm': '1607', 'arxiv_id': '1607.04817', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04817'}
null
null
# Introduction > *"One Picture is Worth a Thousand Words"*\ > \[An advertisement for the *San Antonio Light* (1918)\] Teaching undergraduate *Analysis of Algorithms* has been a rewarding, although a bit taxing, experience. I was often surprised to learn that many basic problems that clearly belong to its core syllabus had been left unanswered or partially answered. Also, it seemed a bit odd to me that many otherwise decent texts offered unnecessarily imprecise computations[^1] of several rather fundamental results. In this article, I pursue a seemingly marginal topic, the best-case behavior of a well-known sorting algorithm \({\tt MergeSort}\), which pursuit, however, yields some interesting findings that could hardly be characterized as *"marginal."* It turns out that-contrary to what a casual student of this subject might believe-computing the exact formula for the number of comparisons of keys that \({\tt MergeSort}\) performs on any \(n\)-element array in the best case is not a routine exercise and leads to a problem that gained some notoriety for being a hard nut to crack analytically: the *sum of digits* problem. Even more unexpectedly, a relatively straightforward[^2] formula for the said number of comparisons yields an improvement of a well-known answer to this instance of the *sum of digits* problem: > How many \(1\)s appear in binary representations of all integers between (but not including) \(0\) and \(n\)? # \({\tt MergeSort}\) and its best-case behavior {#sec:mergsor} A call to \({\tt MergeSort}\) inherits an \(n\)-element array \({\tt A}\) of integers and sorts it non-decreasingly, following the steps described below. A Java code of \({\tt Merge}\) is shown on the Figure [\[fig:Merge\]](#fig:Merge){reference-type="ref" reference="fig:Merge"}. A typical measure of the running time of \({\tt MergeSort}\) is the number of *comparisons of keys*, which for brevity I call *comps*, that it performs while sorting array \({\tt A}\). Since no comps are performed outside \({\tt Merge}\), the running time of \({\tt MergeSort}\) can be computed as the sum of numbers of comps performed by all calls to \({\tt Merge}\) during the execution of \({\tt MergeSort}\). Since the minimum number of comps performed by \({\tt Merge}\) on two list is equal to the length of the shorter list, and any increasingly sorted array on any size \(N \geq 2\) produces only best-case scenarios for all subsequent calls to \({\tt Merge}\), a rudimentary analysis of the recursion tree for \({\tt MergeSort}\) easily yields the exact formula for the minimum number of comps for the entire \({\tt MergeSort}\). The problem arises when one tries to reduce the said formula, which naturally involves long summations, to one that can be evaluated in a logarithmic time. ## Recursion tree {#subsecsec:rectre} The obvious recursion tree for \({\tt MergeSort}\) and sufficiently large \(n\) is shown on Figure [\[fig:rectre\]](#fig:rectre){reference-type="ref" reference="fig:rectre"}. A recursive application of the equality[^3] \[\label{eq:n/2=n+1/2} \lceil \frac{n}{2}\rceil = \lfloor \frac{n+1}{2}\rfloor\] allows for rewriting of that tree onto one whose first four levels are shown on Figure [\[fig:rectre1\]](#fig:rectre1){reference-type="ref" reference="fig:rectre1"}. ## Best-case and its characterization \(B(n)\) {#subsec:bestchar} The best-case arrays of sizes \(\lfloor \frac{n}{2} \rfloor\) and \(\lceil \frac{n}{2} \rceil\) for \({\tt Merge}\), where \(n \geq 2\), are those in which every element of the first array is less than all elements of the second one. In such a case, \({\tt MergeSort}\) performs \(\lfloor \frac{n}{2} \rfloor\) of comps. Thus the following recurrence relation for the least number \(B(n)\) of comparisons of keys that \({\tt MergeSort}\) performs on any \(n\)-element array is straightforward to derive from its description given by Algorithm [\[alg:mersor\]](#alg:mersor){reference-type="ref" reference="alg:mersor"}. \[\label{eq:rec_rel1} B(1) = 0,\] and, for \(n \geq 2\), \[\label{eq:rec_rel2} B(n) = \lfloor \frac{n}{2}\rfloor + B( \lfloor \frac{n}{2}\rfloor) + B( \lceil \frac{n}{2}\rceil).\] Using the equality ([\[eq:n/2=n+1/2\]](#eq:n/2=n+1/2){reference-type="ref" reference="eq:n/2=n+1/2"}), the recurrence relation ([\[eq:rec_rel2\]](#eq:rec_rel2){reference-type="ref" reference="eq:rec_rel2"}) is equivalent to: \[\label{eq:rec_rel3} B(n) = \lfloor \frac{n}{2}\rfloor + B( \lfloor \frac{n}{2}\rfloor) + B( \lfloor \frac{n+1}{2}\rfloor) .\] A graph of \(B(n)\) is shown on Figure [\[fig:Best-case_rec_solution\]](#fig:Best-case_rec_solution){reference-type="ref" reference="fig:Best-case_rec_solution"}. Unfolding the recurrence ([\[eq:rec_rel3\]](#eq:rec_rel3){reference-type="ref" reference="eq:rec_rel3"}) allows for noticing that the minimum number \(B(n)\) of comps performed by all calls to \({\tt Merge}\) is equal to the sum of all values shown at nodes highlighted yellow in the recursion tree \(T\) of Figure [\[fig:rectre1\]](#fig:rectre1){reference-type="ref" reference="fig:rectre1"}. They may be summed-up level-by-level. One can notice from Figure [\[fig:rectre1\]](#fig:rectre1){reference-type="ref" reference="fig:rectre1"} that the number of comps performed at any level \(k\) with the maximal number \(2^k\) of nodes is given by this formula: \[\label{eq:formula_level_i} \sum _{i=0} ^{2^k-1} \lfloor \frac{n+i}{2^{k+1}} \rfloor.\] What is not clear is whether all levels of the recursion tree \(T\) are maximal. Fortunately, the answer to this question does not depend on whether given instance of \({\tt MergeSort}\) is running on a best-case array or on any other case of array. It has been known form a classic analysis of the worst-case running time of \({\tt MergeSort}\) that every level of its recursion tree \(T\) that contains at least one non-leaf, or-in other words-a node that shows value \(p \geq 2\), is maximal. [\[sec:mergesort\]](#sec:mergesort){reference-type="ref" reference="sec:mergesort"} page  contains a detailed derivation of that fact. Thus all levels 0 through \(h-1\) of \(T\) are maximal. Therefore, the formula ([\[eq:formula_level_i\]](#eq:formula_level_i){reference-type="ref" reference="eq:formula_level_i"}) gives the number of comps for every level \(0 \leq k \leq h-1\). The last level \(h\) of \(T\) may be not maximal because the level \(h-1\) may contain leaves, or-in other words-nodes that show value \(p = 1\), where \(p = \lfloor \frac{n+i}{2^{h-1}} \rfloor\) for some \(0 \leq i \leq 2^{h-1}-1\), and as such do not have any children in level \(h\). However, for each such node the value of \(\lfloor \frac{p}{2} \rfloor = \lfloor \frac{n+i}{2^{h}} \rfloor\) is 0, so it can be included in summation ([\[eq:formula_level_i\]](#eq:formula_level_i){reference-type="ref" reference="eq:formula_level_i"}) without affecting its value even though the said value does not correspond to any node in level \(h\). Therefore, the formula ([\[eq:formula_level_i\]](#eq:formula_level_i){reference-type="ref" reference="eq:formula_level_i"}) gives the number of comps for level \(k = h\). Also, the depth of \(T\) is \(\lfloor \lg n \rfloor\), as the Theorem [\[thm:depthRectree\]](#thm:depthRectree){reference-type="ref" reference="thm:depthRectree"} page in [\[sec:excerpts\]](#sec:excerpts){reference-type="ref" reference="sec:excerpts"} states. Thus the minimum number of comps performed by \({\tt MergeSort}\) is given by this formula: \[\label{eq:formula1} \sum _{k=0} ^{\lfloor \lg n \rfloor} \sum _{i=0} ^{2^k-1} \lfloor \frac{n+i}{2^{k+1}} \rfloor.\] Unfortunately, the summation ([\[eq:formula1\]](#eq:formula1){reference-type="ref" reference="eq:formula1"}) contains \(n-1\) non-zero terms, so it cannot be evaluated quickly in its present form. Fortunately, its inner summation ([\[eq:formula_level_i\]](#eq:formula_level_i){reference-type="ref" reference="eq:formula_level_i"}) can be reduced to a closed-form formula. ## *Zigzag* function {#subsec:zig} In order to reduce ([\[eq:formula_level_i\]](#eq:formula_level_i){reference-type="ref" reference="eq:formula_level_i"}) to a closed form, I am going to use function *Zigzag* defined by: \[\label{eq:def_zigzag} \mbox{\textit{Zigzag}}\,(n) = \min (x-\lfloor x \rfloor, \lceil x \rceil-x).\] The following fact is instrumental for that purpose. Here is the closed-form of the summation ([\[eq:formula_level_i\]](#eq:formula_level_i){reference-type="ref" reference="eq:formula_level_i"}). The following theorem yields the formula ([\[eq:formula_Zigzag\*\]](#eq:formula_Zigzag*){reference-type="ref" reference="eq:formula_Zigzag*"}) for the minimum number \(B(n)\) of comps performed by \({\tt MergeSort}\). Formula ([\[eq:formula_Zigzag\*\]](#eq:formula_Zigzag*){reference-type="ref" reference="eq:formula_Zigzag*"}), although not quite closed-form, comprises of summation with only \(\lfloor \lg n \rfloor +1\) closed-form terms, so it may be evaluated in \(O(\log ^{c})\) time, where \(c\) is a constant. I will show in Section [3](#sec:frac){reference-type="ref" reference="sec:frac"} that ([\[eq:formula_Zigzag\*\]](#eq:formula_Zigzag*){reference-type="ref" reference="eq:formula_Zigzag*"}) does not have a closed form. Graphs of both sides of equality ([\[eq:formula_Zigzag\*\]](#eq:formula_Zigzag*){reference-type="ref" reference="eq:formula_Zigzag*"}) are shown on Figure [\[fig:Best-case_MergeSort_worksheet_best-case\]](#fig:Best-case_MergeSort_worksheet_best-case){reference-type="ref" reference="fig:Best-case_MergeSort_worksheet_best-case"}. Once can see that for natural numbers \(n\) they coincide with the solution \(B(n)\) of recurrences [\[eq:rec_rel1\]](#eq:rec_rel1){reference-type="eqref" reference="eq:rec_rel1"} and [\[eq:rec_rel2\]](#eq:rec_rel2){reference-type="eqref" reference="eq:rec_rel2"} visualized on Figure [\[fig:Best-case_rec_solution\]](#fig:Best-case_rec_solution){reference-type="ref" reference="fig:Best-case_rec_solution"}. # A fractal in \(B(n)\) {#sec:frac} A deceitfully simple expression \[\label{eq:sum_zigzag} \sum _{k=0} ^{\lfloor \lg x \rfloor} 2^{k+1} \mbox{\textit{Zigzag}}\,(\frac{x}{2^{k+1}}),\] half of which occurs in formula ([\[eq:formula_Zigzag_B\]](#eq:formula_Zigzag_B){reference-type="ref" reference="eq:formula_Zigzag_B"}) of Corollary [\[thm:formula_Zigzag_B\]](#thm:formula_Zigzag_B){reference-type="ref" reference="thm:formula_Zigzag_B"}, is a formidable adversary for those who may try to turn it into a closed form, although the time required for its evaluation for any given \(n\) is \(O (\log ^ c)\) [^6]. That does not come as a surprise, taking into account that its graph, shown on Figure [\[fig:Best-case_MergeSort_worksheet_sum_zigzag\]](#fig:Best-case_MergeSort_worksheet_sum_zigzag){reference-type="ref" reference="fig:Best-case_MergeSort_worksheet_sum_zigzag"}, bears a resemblance of fractal. This can be easily seen as soon as a sawtooth function \(2^{\lfloor \lg x \rfloor + 1}-x\) is subtracted from it, yielding the function \(F(x)\) given by \[\label{eq:def_F} F(x) = \sum _{k=0} ^{\lfloor \lg x \rfloor} 2^{k+1} \mbox{\textit{Zigzag}}\,(\frac{x}{2^{k+1}})-2^{\lfloor \lg x \rfloor + 1} + x.\] Since \(\frac{1}{2} \leq \frac{x}{2^{\lfloor \lg x \rfloor + 1}} < 1\), equality [\[eq:def_zigzag\]](#eq:def_zigzag){reference-type="eqref" reference="eq:def_zigzag"} implies \[\nonumber \mbox{\textit{Zigzag}}\,(\frac{x}{2^{\lfloor \lg x \rfloor+1}}) = 1-\frac{x}{2^{\lfloor \lg x \rfloor+1}},\] or \[\label{eq:zigprop1} 2^{\lfloor \lg x \rfloor+1} \mbox{\textit{Zigzag}}\,(\frac{x}{2^{\lfloor \lg x \rfloor+1}}) = 2^{\lfloor \lg x \rfloor+1}-x.\] The equality [\[eq:zigprop1\]](#eq:zigprop1){reference-type="eqref" reference="eq:zigprop1"} simplifies definition [\[eq:def_F\]](#eq:def_F){reference-type="eqref" reference="eq:def_F"} of function \(F\) to \[\label{eq:fractal} F(x) = \sum _{k=1} ^{\lfloor \lg x \rfloor} 2^{k} \mbox{\textit{Zigzag}}\,(\frac{x}{2^{k}}),\] visualized on Figure [\[fig:Best-case_MergeSort_worksheet_sum_zigzag_minus_chainsaw\]](#fig:Best-case_MergeSort_worksheet_sum_zigzag_minus_chainsaw){reference-type="ref" reference="fig:Best-case_MergeSort_worksheet_sum_zigzag_minus_chainsaw"}. The function \(F\) is a fractal with quasi similarity that repeats at intervals of exponentially growing length. It is a union \[\label{eq:F_union_def} F = \bigcup _{k=0} ^{\infty} f_k\] of functions \(f_k\), each having an interval \([ 2^k, 2^{k+1})\) as its domain. In other words, for every integer \(k \geq 0\), \[\label{eq:def_f_k} f_k = F \restriction [ 2^k, 2^{k+1}),\] which, of course, yields [\[eq:F_union_def\]](#eq:F_union_def){reference-type="eqref" reference="eq:F_union_def"}. Let \(\hat{f}_k\) be the normalized \(f_k\) on interval \([0,1)\), defined by: \[\label{eq:f_k_normalized} \hat{f}_k (x) = \frac{1}{2^k} f_k (2^k (x + 1)),\] and \(\tilde{f}_k\) be the periodized \(\hat{f}_k\) by composing it with a sawtooth function \(x-\lfloor x \rfloor\),[^7] defined by: \[\label{eq:f_k_periodized} \tilde{f}_k(x) = \hat{f}_k(x-\lfloor x \rfloor).\] Contracting definitions [\[eq:def_f\_k\]](#eq:def_f_k){reference-type="eqref" reference="eq:def_f_k"}, [\[eq:f_k\_normalized\]](#eq:f_k_normalized){reference-type="eqref" reference="eq:f_k_normalized"}, and [\[eq:f_k\_periodized\]](#eq:f_k_periodized){reference-type="eqref" reference="eq:f_k_periodized"}, yields \[\label{eq:f_k_simplified_a} \tilde{f}_k(x) = \frac{1}{2^k} F (2^k (x-\lfloor x \rfloor + 1)).\] One can compute[^8] from [\[eq:f_k\_simplified_a\]](#eq:f_k_simplified_a){reference-type="eqref" reference="eq:f_k_simplified_a"} the following alternative formula for \(\tilde{f}_k(x)\): \[\label{eq:f_k_simplified} \tilde{f}_k(x) = \sum _{i=0} ^{k-1} \frac{1}{2^i} \mbox{\textit{Zigzag}} \, (2^i x).\] Figure [\[fig:Best-case_MergeSort_worksheet_sum_zigzag_minus_chainsaw_progression\]](#fig:Best-case_MergeSort_worksheet_sum_zigzag_minus_chainsaw_progression){reference-type="ref" reference="fig:Best-case_MergeSort_worksheet_sum_zigzag_minus_chainsaw_progression"} shows functions \(\tilde{f}_0...,\tilde{f}_6\) drawn on the same graph. Since each function \(f_k\), and-therefore-each function \(\hat{f}_k\), and-therefore-each function \(\tilde{f}_k\), are a result of smaller and smaller triangles piled, originating in function *Zigzag* of definition ([\[eq:fractal\]](#eq:fractal){reference-type="ref" reference="eq:fractal"}) of function \(F\), on one another as shown on Figure [\[fig:Best-case_MergeSort_worksheet_sum_zigzag_minus_chainsaw_progression\]](#fig:Best-case_MergeSort_worksheet_sum_zigzag_minus_chainsaw_progression){reference-type="ref" reference="fig:Best-case_MergeSort_worksheet_sum_zigzag_minus_chainsaw_progression"}, for any integers \(0 \leq i < j\), \(\tilde{f}_i\) linearly interpolates \(\tilde{f}_j\). Because of that, each \(\tilde{f}_i\) linearly interpolates the limit \(\tilde{F}\) of all \(\tilde{f}_k\)s defined by: \[\label{eq:def_lim} \tilde{F}(x) = \lim _{k \rightarrow \infty} \tilde{f}_k (x),\] as Figure [\[fig:Takagi_function_progression_5\]](#fig:Takagi_function_progression_5){reference-type="ref" reference="fig:Takagi_function_progression_5"} illustrates. An application of [\[eq:f_k\_simplified\]](#eq:f_k_simplified){reference-type="eqref" reference="eq:f_k_simplified"} to [\[eq:def_lim\]](#eq:def_lim){reference-type="eqref" reference="eq:def_lim"} yields: \[\label{eq:def_lim2} \tilde{F}(x) = \sum _{i=0} ^{\infty} \frac{1}{2^i} \mbox{\textit{Zigzag}} \, (2^i x).\] Since for every integer \(n\) and \(i \geq k\), \(2^i \frac{n}{2^k}\) is integer, \(\mbox{\textit{Zigzag}} \, (2^i \frac{n}{2^k}) = 0\). Therefore, by virtue of [\[eq:f_k\_simplified\]](#eq:f_k_simplified){reference-type="eqref" reference="eq:f_k_simplified"} and [\[eq:def_lim2\]](#eq:def_lim2){reference-type="eqref" reference="eq:def_lim2"}, for every non-negative integer \(k\) and \(n\), \[\label{eq:Ftilde=ftilde_k} \tilde{F}(\frac{n}{2^k}) = \tilde{f}_k(\frac{n}{2^k}).\] This and [\[eq:f_k\_simplified\]](#eq:f_k_simplified){reference-type="eqref" reference="eq:f_k_simplified"} eliminate the need for infinite summation[^9] while computing \(\tilde{F}(\frac{n}{2^k})\). It can be shown that although a continuous function, \(\tilde{F}\) is nowhere-differentiable. As such, it does not have a closed-form formula as any closed-form formula on a real interval must define a function have a derivative at every point of that interval, except for a non-dense set of its points. Since \(\tilde{F}\) can be expressed in function, described by a closed-form formula, of the right-hand side of formula [\[eq:formula_Zigzag\*\]](#eq:formula_Zigzag*){reference-type="eqref" reference="eq:formula_Zigzag*"}, the latter does not have a closed-form formula, either. This way I arrived at the following conclusion. ***Note***. One can apply the reverse transformations to those used in Section [3](#sec:frac){reference-type="ref" reference="sec:frac"} on function \(\tilde{F}\) and construct a fractal function \(\breve{F}\), shown on Figure [\[fig:Best-case_MergeSort_worksheet_sum_zig_minus_chainsaw_w\_upper\]](#fig:Best-case_MergeSort_worksheet_sum_zig_minus_chainsaw_w_upper){reference-type="ref" reference="fig:Best-case_MergeSort_worksheet_sum_zig_minus_chainsaw_w_upper"}, given by the equation \[\label{eq:def_F_u} \breve{F} (x) = 2^{\lfloor \lg x \rfloor } \tilde{F}(\frac{x}{2^{\lfloor \lg x \rfloor}}),\] that for every positive integer \(n\) satisfies \[\label{eq:F_u=F} \breve{F} (n) = F(n),\] where \(F\) is given by [\[eq:fractal\]](#eq:fractal){reference-type="eqref" reference="eq:fractal"}. # Computing \(\tilde{F} (x)\) and \(B(n)\) from one another {#sec:compfromfrac} Computing values of function \(\tilde{F} (x)\) does not have to be as complex as (or more complex than) the definition [\[eq:def_lim2\]](#eq:def_lim2){reference-type="eqref" reference="eq:def_lim2"} implies. Of course, for every integer \(n\), \(\tilde{F} (n)\) \(=\) \(0\). One can apply some elementary arguments based on a structure visualized on Figure [\[fig:Takagi_function_progression_5\]](#fig:Takagi_function_progression_5){reference-type="ref" reference="fig:Takagi_function_progression_5"} to conclude that \[\tilde{F} (\frac{2}{3}) = \tilde{F} (\frac{1}{3}) = \frac{2}{3},\] (the latter being the maximum of \(\tilde{F} (x)\)) or that for every positive integer \(k\), \[\tilde{F} (\frac{1}{2^k}) = \frac{ k }{2^k}.\] It takes a bit more work to compute \[\tilde{F} (\frac{3}{2^k}) = \frac{3 k-4 }{2^k}.\] It turns out that computing values of function \(\tilde{F} (x)\) for every \(x\) that has a finite binary expansion can be done easily if an oracle for computing the values of the function \(B(n)\) defined by [\[eq:rec_rel1\]](#eq:rec_rel1){reference-type="eqref" reference="eq:rec_rel1"} and [\[eq:rec_rel2\]](#eq:rec_rel2){reference-type="eqref" reference="eq:rec_rel2"} is given[^10]. Once that is accomplished, since \(\tilde{F} (x)\) is a continuous function and the set of numbers with finite binary expansions is dense in the set \(\mathfrak{R}\) of reals, it allows for fast approximations of \(\tilde{F} (x)\) for every real \(x\). [^11] Theorem [\[thm:mainB\]](#thm:mainB){reference-type="ref" reference="thm:mainB"} allows for easy computing of \(B(n)\) if \(\tilde{F} (\frac{n}{2^k})\) is given for some \(k \geq \lg n\) using this form of [\[eq:mainB\]](#eq:mainB){reference-type="eqref" reference="eq:mainB"}: For instance, putting \(k = \lfloor \lg n \rfloor + 1\) in [\[eq:mainB2\]](#eq:mainB2){reference-type="eqref" reference="eq:mainB2"} easily yields [\[eq:formula_Zigzag_B\]](#eq:formula_Zigzag_B){reference-type="eqref" reference="eq:formula_Zigzag_B"}. For \(k = \lceil \lg n \rceil\) we obtain \[B(n) = \frac{n \lceil \lg n \rceil }{2}-2^{\lceil \lg n \rceil-1} \tilde{F} (\frac{n}{2^{\lceil \lg n \rceil}}) =\] \[by [\[eq:def_lim2\]](#eq:def_lim2){reference-type="eqref" reference="eq:def_lim2"}\] \[= \frac{n \lceil \lg n \rceil }{2}-2^{\lceil \lg n \rceil-1} \sum _{i=0} ^{\infty} \frac{1}{2^i} \mbox{\textit{Zigzag}} \, (2^i \frac{n}{2^{\lceil \lg n \rceil}}) =\] \[since for \(i \geq \lceil \lg n \rceil\), \(2^i \frac{n}{2^{\lceil \lg n \rceil}}\) is integer and \(\mbox{\textit{Zigzag}} \, (2^i \frac{n}{2^{\lceil \lg n \rceil}}) = 0\)\] \[= \frac{n \lceil \lg n \rceil }{2}-2^{\lceil \lg n \rceil-1} \sum _{i=0} ^{\lceil \lg n \rceil-1} \frac{1}{2^i} \mbox{\textit{Zigzag}} \, (2^i \frac{n}{2^{\lceil \lg n \rceil}}) =\] \[= \frac{n \lceil \lg n \rceil }{2}-\frac{1}{2} \sum _{i=0} ^{\lceil \lg n \rceil-1} 2^{\lceil \lg n \rceil-i} \mbox{\textit{Zigzag}} \, ( \frac{n}{2^{\lceil \lg n \rceil-i}}) .\] Substituting \(k\) for \(\lceil \lg n \rceil-i\) we conclude \[\label{eq:mainAlt} B(n) = \frac{n \lceil \lg n \rceil }{2}-\frac{1}{2} \sum _{k=1} ^{\lceil \lg n \rceil} 2^{k} \mbox{\textit{Zigzag}} \, ( \frac{n}{2^{k}}),\] a similar to [\[eq:formula_Zigzag_B\]](#eq:formula_Zigzag_B){reference-type="eqref" reference="eq:formula_Zigzag_B"} characterization of \(B(n)\). # Relationship between the best case and the worst case A casual student of \({\tt MergeSort}\) tends to believe that its worst-case behavior is about twice as bad as its best-case behavior. This, of course, is only approximately true. In this Section, I will derive the exact difference between \(2 B(n)\) and \(W(n)\) using function \(F\) defined by [\[eq:def_F\]](#eq:def_F){reference-type="eqref" reference="eq:def_F"} page . An exact formula for the number \(W(n)\) of comparisons of keys performed by \({\tt MergeSort}\) in the worst case is known[^13] and is given for any positive integer \(n\) by the following equality: \[\label{eq_worst} W(n) = \sum _{i = 1} ^{n} \lceil \lg i \rceil.\] From [\[eq:formula_Zigzag_B\]](#eq:formula_Zigzag_B){reference-type="eqref" reference="eq:formula_Zigzag_B"} and [\[eq:def_F\]](#eq:def_F){reference-type="eqref" reference="eq:def_F"}, one can derive \[2B(n) = n \lfloor \lg n \rfloor-2^{\lfloor \lg n \rfloor + 1} + 2n-F(n) =\] \[by \(\sum _{i = 1} ^{n} \lceil \lg i \rceil = n \lfloor \lg n \rfloor-2^{\lfloor \lg n \rfloor + 1} + n + 1\) from \] \[= \sum _{i = 1} ^{n} \lceil \lg i \rceil-1 + n-F(n) =\] \[by [\[eq_worst\]](#eq_worst){reference-type="eqref" reference="eq_worst"}\] \[= W(n)-1 + n-F(n).\] The above yield the following characterization. In particular, since for every positive integer \(n\), \[\label{eq:F_ub} 0 \leq F(n) \leq \frac{n-1}{2}\] (see Figure [\[fig:Best-case_MergeSort_worksheet_sum_zigzag_minus_chainsaw\]](#fig:Best-case_MergeSort_worksheet_sum_zigzag_minus_chainsaw){reference-type="ref" reference="fig:Best-case_MergeSort_worksheet_sum_zigzag_minus_chainsaw"} for explanation), I conclude with the following tight linear bounds on \(2B(n)-W(n)\). Obviously, \(2B(n)-W(n) = n-1\) whenever \(F(n) = 0\), that is, whenever \(n = 2^{\lfloor \lg n \rfloor}\). It can be shown that \(2B(n)-W(n) = \frac{n-1}{2}\) whenever \(n\) \(=\) \(\frac{1}{3}(2^{k+1} + (-1)^k)\) for some integer \(k \geq 0\). A graph of \(2B(n)-W(n)\) and its tight bounds are shown on Figure [\[fig:2B-W_with_bounds\]](#fig:2B-W_with_bounds){reference-type="ref" reference="fig:2B-W_with_bounds"}. # The sum of digits problem {#sec:sumdig} A known explicit formula, published in, for the total number of bits in all integers between 0 and \(n\) (not including 0 and \(n\)) is expressed in terms of function *Zigzag* (referred to as \(2 g\) in ) and is given by:[^14] ![image](Quote_from_Trollope_A){width="1\\linewidth"} \(\,\,\) Let \(g(x)\) be periodic of period 1 and defined on \([0,1]\) by ![image](Quote_from_Trollope){width="1\\linewidth"} It has been shown in that the recurrence relation for \(A(n,2)\) is the same as the recurrence relation for \(B(n)\) given by ([\[eq:rec_rel1\]](#eq:rec_rel1){reference-type="ref" reference="eq:rec_rel1"}) and ([\[eq:rec_rel2\]](#eq:rec_rel2){reference-type="ref" reference="eq:rec_rel2"}). Therefore, the formula ([\[eq:formula_Zigzag\*\]](#eq:formula_Zigzag*){reference-type="ref" reference="eq:formula_Zigzag*"}) derived in this paper is equivalent to \(A(n,2)\) given above by the considerably more complicated definition. Interestingly, the above definition can be simplified to [\[eq:formula_Zigzag\*\]](#eq:formula_Zigzag*){reference-type="eqref" reference="eq:formula_Zigzag*"} along the lines of the elementary derivation of the alternative formula [\[eq:mainAlt\]](#eq:mainAlt){reference-type="eqref" reference="eq:mainAlt"} for \(B(n)\) on page [^15]. # Proof of Theorem [\[thm:formula_level_Zigzag\]](#thm:formula_level_Zigzag){reference-type="ref" reference="thm:formula_level_Zigzag"} page , Subsection [2.3](#subsec:zig){reference-type="ref" reference="subsec:zig"} {#sec:proof_formula_level_Zigzag} In this Section, I provide an analytic proof of the experimentally-derived Theorem [\[thm:formula_level_Zigzag\]](#thm:formula_level_Zigzag){reference-type="ref" reference="thm:formula_level_Zigzag"} page , Subsection [2.3](#subsec:zig){reference-type="ref" reference="subsec:zig"} that was instrumental for the derivation of a logarithmic-length formula[^16] for \(B(n)\). The result and its proof have a flavor of Concrete Mathematics. Although they are interesting in their own right, they cannot be found in. # Proof of Theorem [\[thm:no_closed_form\]](#thm:no_closed_form){reference-type="ref" reference="thm:no_closed_form"} page , Section [3](#sec:frac){reference-type="ref" reference="sec:frac"} [\[sec:proof_no_closed_form\]]{#sec:proof_no_closed_form label="sec:proof_no_closed_form"} {#proof-of-theorem-thmno_closed_form-page-section-secfrac-secproof_no_closed_form} In this Section, I present a brief discussion/motivation of what can be generally considered a *closed-form formula* for a function from the set of real numbers into a set of real numbers. I provide an analytic proof of Theorem [\[thm:no_closed_form\]](#thm:no_closed_form){reference-type="ref" reference="thm:no_closed_form"} page , Section [3](#sec:frac){reference-type="ref" reference="sec:frac"} that implies the non-existence of closed-form formula for the minimum number \(B(n)\) of comparisons of keys by \({\tt MergeSort}\) while sorting an \(n\)-element array. I am going to use the acronym \(cf\!f\) as an abbreviation for *closed-form formula*. For reader's convenience, the Theorem [\[thm:no_closed_form\]](#thm:no_closed_form){reference-type="ref" reference="thm:no_closed_form"} is quoted below as Theorem [\[thm:no_closed_form_add\]](#thm:no_closed_form_add){reference-type="ref" reference="thm:no_closed_form_add"}. The rest of this Section constitutes the proof of Theorem [\[thm:no_closed_form_add\]](#thm:no_closed_form_add){reference-type="ref" reference="thm:no_closed_form_add"}. First, let me use an example of function \(2^x: \mathbb{R} \longrightarrow \mathbb{R}\) as an insight of what may be accepted as a \(cf\!f\) for a *continuous* function-like, say, \(\tilde{F}(x)\)-on the set \(\mathbb{R}\) of reals or on an interval thereof. One picks a dense[^17] subset \(\mathbb{Q}\) of \(\mathbb{R}\), with a collection of mappings \(\rho_x(i): \mathbb{N} \longrightarrow \mathbb{Q}\), where \(x \in \mathbb{R}\), given by \(\rho_x(i) = \frac{\lfloor i \times x \rfloor}{i}\) so that \(\lim _{i \rightarrow \infty}\rho_x(i) = x\). Since for any \(x \in \mathbb{R} \setminus \mathbb{Q}\), \(2^x\) has been defined as \[2^x = \lim _{i \rightarrow \infty} 2^{\rho_x (i)} = \lim _{i \rightarrow \infty} \sqrt[i]{2^{\lfloor i \times x \rfloor}},\] \(\lim _{i \rightarrow \infty} \sqrt[i]{2^{\lfloor i \times x \rfloor}}\) is considered a \(cf\!f\) \(\alpha: \mathbb{R} \longrightarrow \mathbb{R}\) for \(2^x\). Should the nowhere-differentiable function \(\tilde{F}\) have a \(cf\!f\), it would be differentiable everywhere except, perhaps, on a non-dense subset of \(\mathbb{R}\). The following inductive argument demonstrates that. All atomic \(cf\!f\)s are differentiable except, perhaps, on a non-dense subset of \(\mathbb{R}\). If a finite number of \(cf\!f\)s are differentiable except, perhaps, on non-dense subsets of \(\mathbb{R}\) then their composition is differentiable except, perhaps, on non-dense subsets of \(\mathbb{R}\). [^20] Thus \(\tilde{F}\) has no \(cf\!f\). The above observation, together with Lemma [\[lem:main_no_cff\]](#lem:main_no_cff){reference-type="ref" reference="lem:main_no_cff"}, complete the proof of Theorem [\[thm:no_closed_form_add\]](#thm:no_closed_form_add){reference-type="ref" reference="thm:no_closed_form_add"}. # Proof of Theorem [\[thm:mainB\]](#thm:mainB){reference-type="ref" reference="thm:mainB"} page , Section [4](#sec:compfromfrac){reference-type="ref" reference="sec:compfromfrac"} [\[sec:comment\]]{#sec:comment label="sec:comment"} {#proof-of-theorem-thmmainb-page-section-seccompfromfrac-seccomment} In this Section, I provide an analytic proof of experimentally-derived Theorem [\[thm:mainB\]](#thm:mainB){reference-type="ref" reference="thm:mainB"} page , Section [4](#sec:compfromfrac){reference-type="ref" reference="sec:compfromfrac"}. This result, re-stated by Theorem [\[thm:mainB_add\]](#thm:mainB_add){reference-type="ref" reference="thm:mainB_add"} below, allows for practically efficient computations of values of the continuous Blackmange function for reals with finite binary floating-point representations. I also provide some properties (Lammas [\[lem:100\]](#lem:100){reference-type="ref" reference="lem:100"}, [\[lem:200\]](#lem:200){reference-type="ref" reference="lem:200"}, and [\[lem:300\]](#lem:300){reference-type="ref" reference="lem:300"}) of the *Zigzag* function, given by the equality [\[eq:def_zigzag\]](#eq:def_zigzag){reference-type="eqref" reference="eq:def_zigzag"} page  and visualized on Figure [\[fig:zigzag\]](#fig:zigzag){reference-type="ref" reference="fig:zigzag"} page , that are useful for a neat derivation of a formula for the Blancmange function as (the limit of) a finite sum of some values of the *Zigzag* function. Let the function[^21] \(\tilde{F}\), visualized on Figure [\[fig:Tak\]](#fig:Tak){reference-type="ref" reference="fig:Tak"} page , be defined by [\[eq:def_lim2\]](#eq:def_lim2){reference-type="eqref" reference="eq:def_lim2"} page , and \(B(n)\), given by [\[eq:rec_rel1\]](#eq:rec_rel1){reference-type="eqref" reference="eq:rec_rel1"} and [\[eq:rec_rel3\]](#eq:rec_rel3){reference-type="eqref" reference="eq:rec_rel3"} page , be the least number of comparisons of keys that \({\tt MergeSort}\) performs while sorting an \(n\)-element array. The reminder of this Section constitutes a proof of Theorem [\[thm:mainB_add\]](#thm:mainB_add){reference-type="ref" reference="thm:mainB_add"}. *Note*. Function *Zigzag*, visualized on Figure [\[fig:zigzag\]](#fig:zigzag){reference-type="ref" reference="fig:zigzag"} page , has been defined by [\[eq:def_zigzag\]](#eq:def_zigzag){reference-type="eqref" reference="eq:def_zigzag"} page . At this point, we are ready to conclude the proof of Theorem [\[thm:mainB\]](#thm:mainB){reference-type="ref" reference="thm:mainB"}. By virtue of [\[eq:formula_Zigzag_B\]](#eq:formula_Zigzag_B){reference-type="eqref" reference="eq:formula_Zigzag_B"} page , we have: \[2 B(n) = n(\lfloor \lg n \rfloor + 1)-\sum _{k=0} ^{\lfloor \lg n \rfloor} 2^{k+1} Zigzag (\frac{n}{2^{k+1}}) =\] \[= n(\lfloor \lg n \rfloor + 1)-\sum _{i=1} ^{\lfloor \lg n \rfloor+1} 2^{i} Zigzag (\frac{n}{2^{i}}) =\] \[= n(\lfloor \lg n \rfloor + 1)-\sum _{i=1} ^{k} 2^{i} Zigzag (\frac{n}{2^{i}}) + \sum _{i=\lfloor \lg n \rfloor+2} ^{k} 2^{i} Zigzag (\frac{n}{2^{i}}) =\] \[by Lemmas [\[lem:200\]](#lem:200){reference-type="ref" reference="lem:200"} and [\[lem:300\]](#lem:300){reference-type="ref" reference="lem:300"}\] \[= n(\lfloor \lg n \rfloor + 1)-2^k \tilde{F} (\frac{n}{2^k}) + n \times k-n(\lfloor \lg n \rfloor + 1) = n \times k-2^k \tilde{F} (\frac{n}{2^k}) ,\] that is, \[2 B(n) = n \times k-2^k \tilde{F} (\frac{n}{2^k}) ,\] from which [\[eq:100x\]](#eq:100x){reference-type="eqref" reference="eq:100x"} follows. This completes the proof of Theorem [\[thm:mainB\]](#thm:mainB){reference-type="ref" reference="thm:mainB"}. *Note*. A glance at the proof of Lemma [\[lem:200\]](#lem:200){reference-type="ref" reference="lem:200"} suffices to notice that it fails if \(n > 2^k\), and so does Theorem [\[thm:mainB\]](#thm:mainB){reference-type="ref" reference="thm:mainB"}. In particular, for \(k = \lfloor \lg n \rfloor\), Lemma [\[lem:cff_F\_tilde\]](#lem:cff_F_tilde){reference-type="ref" reference="lem:cff_F_tilde"} page  yields \[\tilde{F} (\frac{n}{2^k}) = \frac{n \times k-2 B(n)}{2^k} + \frac{2n}{2^k}-2 > \frac{n \times k-2 B(n)}{2^k}\] since for \(n > 2^{\lfloor \lg n \rfloor}\), \(\frac{2n}{2^{\lfloor \lg n \rfloor}}-2 >0.\) [^1]: A notable exception in this category is . [^2]: Although not quite *closed-form*. [^3]: It can be verified separately for odd and even values of \(n\). [^4]: [\[foot:Mat\]]{#foot:Mat label="foot:Mat"} I used Wolfram Mathematica for that purpose. [^5]: Analytic proof of that fact is a straightforward exercise; see Appendix [\[sec:proof1\]](#sec:proof1){reference-type="ref" reference="sec:proof1"} page . [^6]: So, to all practical purposes, ([\[eq:formula_Zigzag_B\]](#eq:formula_Zigzag_B){reference-type="ref" reference="eq:formula_Zigzag_B"}) is a closed-form formula. [^7]: The fractional part of \(x\). [^8]: An elementary geometric argument based on the graph visualized on Figure [\[fig:Best-case_MergeSort_worksheet_sum_zigzag_minus_chainsaw_progression\]](#fig:Best-case_MergeSort_worksheet_sum_zigzag_minus_chainsaw_progression){reference-type="ref" reference="fig:Best-case_MergeSort_worksheet_sum_zigzag_minus_chainsaw_progression"} will do. [^9]: As it appears in [\[eq:def_lim2\]](#eq:def_lim2){reference-type="eqref" reference="eq:def_lim2"} [^10]: Which is not that surprising after a glance at Figure [\[fig:Best-case_MergeSort_worksheet_sum_zig_minus_chainsaw_w\_upper\]](#fig:Best-case_MergeSort_worksheet_sum_zig_minus_chainsaw_w_upper){reference-type="ref" reference="fig:Best-case_MergeSort_worksheet_sum_zig_minus_chainsaw_w_upper"}. [^11]: It helps to remember that \(\tilde{F}\) is a periodic function with \(\tilde{F} (x) = \tilde{F} (x-\lfloor x \rfloor)\). [^12]: Of course, one if free to assume that \(n\) is odd here. [^13]: See [\[eq:recmergesort900\]](#eq:recmergesort900){reference-type="eqref" reference="eq:recmergesort900"} in the [\[sec:excerpts\]](#sec:excerpts){reference-type="ref" reference="sec:excerpts"}. [^14]: The following are screen shots and an excerpt from. [^15]: Even more interestingly, if someone did bother to simplify Trollope's formula of then I am not aware of it. [^16]: \(B(n) = \frac{n}{2}(\lfloor \lg n \rfloor + 1)-\sum _{k=0} ^{\lfloor \lg n \rfloor} 2^k \mbox{\textit{Zigzag}}\,(\frac{n}{2^{k+1}})\), where \(\mbox{\textit{Zigzag}}\,(x) = \min (x-\lfloor x \rfloor, \lceil x \rceil-x)\). [^17]: In the metric topology of \(\mathbb{R}\). [^18]: Also, together with its partial sums, on Figure [\[fig:Takagi_function_progression_5\]](#fig:Takagi_function_progression_5){reference-type="ref" reference="fig:Takagi_function_progression_5"}, page . [^19]: It is a trivial exercise to show that every real number with finite binary expansion in the interval \([1,2)\) is of the form \(\frac{n}{2^{\lfloor \lg n \rfloor}}\) for some \(n \in \mathbb{N}\), and it is obvious that every real number of that form has a finite binary expansion and falls into that interval. [^20]: For instance, function \(\sqrt{x}Zigzag(\frac{1}{x})\) is differentiable on \([0,1]\), except for the non-dense set \(\{0\} \cup \{ \frac{1}{n} \mid n \in \mathbb{N} \}\). [^21]: Known as the Blancmange function. [^22]: Of course, one if free to assume that \(n\) is odd here.
{'timestamp': '2016-12-06T02:14:23', 'yymm': '1607', 'arxiv_id': '1607.04604', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04604'}
# Introduction We start with a brief recall of \(t\)-designs. Let \({\mathcal{P}}\) be a set of \(v \ge 1\) elements, and let \({\mathcal{B}}\) be a set of \(k\)-subsets of \({\mathcal{P}}\), where \(k\) is a positive integer with \(1 \leq k \leq v\). Let \(t\) be a positive integer with \(t \leq k\). The pair \({\mathbb{D}} = ({\mathcal{P}}, {\mathcal{B}})\) is called a \(t\)-\((v, k, \lambda)\) *design*, or simply *\(t\)-design*, if every \(t\)-subset of \({\mathcal{P}}\) is contained in exactly \(\lambda\) elements of \({\mathcal{B}}\). The elements of \({\mathcal{P}}\) are called points, and those of \({\mathcal{B}}\) are referred to as blocks. We usually use \(b\) to denote the number of blocks in \({\mathcal{B}}\). A \(t\)-design is called *simple* if \({\mathcal{B}}\) does not contain repeated blocks. In this paper, we consider only simple \(t\)-designs. A \(t\)-design is called *symmetric* if \(v = b\). It is clear that \(t\)-designs with \(k = t\) or \(k = v\) always exist. Such \(t\)-designs are *trivial*. In this paper, we consider only \(t\)-designs with \(v > k > t\). A \(t\)-\((v,k,\lambda)\) design is referred to as a *Steiner system* if \(t \geq 2\) and \(\lambda=1\), and is denoted by \(S(t,k, v)\). A necessary condition for the existence of a \(t\)-\((v, k, \lambda)\) design is that \[\begin{aligned} \label{eqn-tdesignnecessty} \binom{k-i}{t-i} \mbox{ divides } \lambda \binom{v-i}{t-i} \end{aligned}\] for all integer \(i\) with \(0 \leq i \leq t\). The interplay between codes and \(t\)-designs goes in two directions. In one direction, the incidence matrix of any \(t\)-design generates a linear code over any finite field \({\mathrm{GF}}(q)\). A lot of progress in this direction has been made and documented in the literature (see, for examples,, , ). In the other direction, the codewords of a fixed Hamming weight in a linear or nonlinear code may hold a \(t\)-design. Some linear and nonlinear codes were employed to construct \(t\)-designs . Binary and ternary Golay codes of certain parameters give \(4\)-designs and \(5\)-designs. However, the largest \(t\) for which an infinite family of \(t\)-designs is derived directly from codes is \(t=3\). According to the references, and, not much progress on the construction of \(t\)-designs from codes has been made so far, while many other constructions of \(t\)-designs are documented in the literature (). The objective of this paper is to construct infinite families of \(2\)-designs and \(3\)-designs from a type of binary linear codes with five weights. The obtained \(t\)-designs depend only on the weight distribution of the underlying binary codes. The total number of \(2\)-designs and \(3\)-designs presented in this paper are exponential in \(m\), where \(m \geq 5\) is an odd integer. In addition, the block size of the designs can vary in a huge range. # The classical construction of \(t\)-designs from codes Let \({\mathcal{C}}\) be a \([v, \kappa, d]\) linear code over \({\mathrm{GF}}(q)\). Let \(A_i:=A_i({\mathcal{C}})\), which denotes the number of codewords with Hamming weight \(i\) in \({\mathcal{C}}\), where \(0 \leq i \leq v\). The sequence \((A_0, A_1, \cdots, A_{v})\) is called the *weight distribution* of \({\mathcal{C}}\), and \(\sum_{i=0}^v A_iz^i\) is referred to as the *weight enumerator* of \({\mathcal{C}}\). For each \(k\) with \(A_k \neq 0\), let \({\mathcal{B}}_k\) denote the set of the supports of all codewords with Hamming weight \(k\) in \({\mathcal{C}}\), where the coordinates of a codeword are indexed by \((0,1,2, \cdots, v-1)\). Let \({\mathcal{P}}=\{0, 1, 2, \cdots, v-1\}\). The pair \(({\mathcal{P}}, {\mathcal{B}}_k)\) may be a \(t\)-\((v, k, \lambda)\) design for some positive integer \(\lambda\). The following theorems, developed by Assumus and Mattson, show that the pair \(({\mathcal{P}}, {\mathcal{B}}_k)\) defined by a linear code is a \(t\)-design under certain conditions. To construct \(t\)-designs via Theorem [\[thm-AM1\]](#thm-AM1){reference-type="ref" reference="thm-AM1"}, we will need the following lemma in subsequent sections, which is a variant of the MacWilliam Identity . Later in this paper, we will need also the following theorem. # A type of binary linear codes with five-weights and related codes {#sec-maincodes} In this section, we first introduce a type of binary linear codes \({\mathcal{C}}_m\) of length \(n=2^m-1\), which has the weight distribution of Table [1](#tab-zhou3){reference-type="ref" reference="tab-zhou3"}, and then analyze their dual codes \({\mathcal{C}}_m^\perp\), the extended codes \(\overline{{\mathcal{C}}_m^\perp}\), and the duals \(\overline{{\mathcal{C}}_m^\perp}^\perp\). Such codes will be employed to construct \(t\)-designs in Sections [4](#sec-2designs){reference-type="ref" reference="sec-2designs"} and [5](#sec-3designs){reference-type="ref" reference="sec-3designs"}. Examples of such codes will be given in Section [6](#sec-examplecodes){reference-type="ref" reference="sec-examplecodes"}. Finally, we settle the weight distribution of the code \(\overline{{\mathcal{C}}_{m}^\perp}\). # Infinite families of \(2\)-designs from \({\mathcal{C}}_{m}^\perp\) and \({\mathcal{C}}_{m}\) {#sec-2designs} # Infinite families of \(3\)-designs from \(\overline{{\mathcal{C}}_{m}^\perp}\) and \(\overline{{\mathcal{C}}_{m}^\perp}^\perp\) {#sec-3designs} # Two families of binary cyclic codes with the weight distribution of Table [1](#tab-zhou3){reference-type="ref" reference="tab-zhou3"} {#sec-examplecodes} To prove the existence of the \(2\)-designs in Section [4](#sec-2designs){reference-type="ref" reference="sec-2designs"} and the \(3\)-designs in Section [5](#sec-3designs){reference-type="ref" reference="sec-3designs"}, we present two families of binary codes of length \(2^m-1\) with the weight distribution of Table [1](#tab-zhou3){reference-type="ref" reference="tab-zhou3"}. Let \(n=q^m-1\), where \(m\) is a positive integer. Let \(\alpha\) be a generator of \({\mathrm{GF}}(q^m)^*\). For any \(i\) with \(0 \leq i \leq n-1\), let \(\mathbb{M}_i(x)\) denote the minimal polynomial of \(\beta^i\) over \({\mathrm{GF}}(q)\). For any \(2 \leq \delta \leq n\), define \[\begin{aligned} \label{eqn-BCHgeneratorPolyn} g_{(q,n,\delta,b)}(x)={\mathrm{lcm}}(\mathbb{M}_{b}(x), \mathbb{M}_{b+1}(x), \cdots, \mathbb{M}_{b+\delta-2}(x)), \end{aligned}\] where \(b\) is an integer, \({\mathrm{lcm}}\) denotes the least common multiple of these minimal polynomials, and the addition in the subscript \(b+i\) of \(\mathbb{M}_{b+i}(x)\) always means the integer addition modulo \(n\). Let \({\mathcal{C}}_{(q, n, \delta,b)}\) denote the cyclic code of length \(n\) with generator polynomial \(g_{(q, n,\delta, b)}(x)\). \({\mathcal{C}}_{(q, n, \delta, b)}\) is called a *primitive BCH code* with *designed distance* \(\delta\). When \(b=1\), the set \({\mathcal{C}}_{(q, n, \delta, b)}\) is called a *narrow-sense primitive BCH code*. Although primitive BCH codes are not good asymptotically, they are among the best linear codes when the length of the codes is not very large. So far, we have very limited knowledge of BCH codes, as the dimension and minimum distance of BCH codes are in general open, in spite of some recent progress. However, in a few cases the weight distribution of a BCH code can be settled. The following theorem introduces such a case. It is known that the dual of a BCH code may not be a BCH code. The following theorem describes a family of cyclic codes having the weight distribution of Table [1](#tab-zhou3){reference-type="ref" reference="tab-zhou3"}, which may not be BCH codes. # Summary and concluding remarks In this paper, with any binary linear code of length \(2^m-1\) and the weight distribution of Table [1](#tab-zhou3){reference-type="ref" reference="tab-zhou3"}, a huge number of infinite families of \(2\)-designs and \(3\)-designs with various block sizes were constructed. These constructions clearly show that the coding theory approach to constructing \(t\)-designs are in fact promising, and may stimulate further investigations in this direction. It is open if the codewords of a fixed weight in a family of linear codes can hold an infinite family of \(t\)-designs for some \(t \geq 4\). It is noticed that the technical details of this paper are tedious. However, one has to settle the weight distribution of a linear code and the minimum distance of its dual at the same time, if one would like to employ the Assmus-Mattson Theorem for the construction of \(t\)-designs. Note that it could be very difficult to prove that a linear code has minimum weight \(7\). This explains why the proofs of some of the theorems are messy and tedious, but necessary.
{'timestamp': '2016-07-19T02:05:49', 'yymm': '1607', 'arxiv_id': '1607.04815', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04815'}
null
null
# Introduction {#sect:intro} Waves measured by a collection of nearby sensors, called an array of receivers, carry information about their source and the medium through which they travel. We consider a typical remote sensing regime with sources of small (point-like) support, and study the inverse problem of determining them from the array measurements. When the waves travel in a known and non-scattering (e.g. homogeneous) medium, the sources can be localized with reverse time migration also known as backprojection. This estimates the source locations as the peaks of the image formed by superposing the array recordings delayed by travel times from the receivers to the imaging points. The accuracy of the estimates depends on the array aperture, the distance of the sources from the array, and the temporal support of the signals emitted by the sources. It may be improved under certain conditions by using \(l_1\) optimization, which seeks to invert the linear mapping from supposedly sparse vectors of the discretized source amplitude on some mesh, to the array measurements. The fast growing literature of imaging with \(l_1\) optimization in homogeneous media includes compressed sensing studies such as, synthetic radar imaging studies like, array imaging studies like, and the resolution study. In this paper we assume that the waves travel in heterogeneous media with fluctuations of the wave speed caused by numerous inhomogeneities. The amplitude of the fluctuations is small, meaning that a single inhomogeneity is a weak scatterer. However, there are many inhomogeneities that interact with the waves on their way from the sources to the receivers, and their scattering effect accumulates. Because in applications it is impossible to know the inhomogeneities in detail, and these cannot be estimated from the array measurements as part of the inversion, the fluctuations of the wave speed are uncertain. We model this uncertainty with a random process, and thus study inversion in random media. In this stochastic framework, the actual heterogeneous medium in which the waves propagate is one realization of the random model. The data measured at the array are uncertain and the question is how to mitigate the uncertainty to get images that are robust with respect to arbitrary medium realizations, i.e., they are statistically stable. The mitigation of uncertainty in the wave propagation model becomes important when the sources are further than a few scattering mean free paths from the array. The scattering mean free path is the length scale on which the waves randomize, meaning that their fluctuations from one medium realization to another are large in comparison with their coherent (statistical expectation) part. The random wave distortions registered at the array are very different from additive and uncorrelated noise assumed usually in inversion. They are more difficult to mitigate and lead to poor and unreliable source reconstructions by coherent methods like reverse time migration or standard \(l_1\) optimization. The Coherent Interferometric (CINT) method is designed to deal efficiently with such random distortions, as long as there is some residual coherence in the array measurements. This holds when the sources are separated from the array by distances (ranges) that are large with respect to the scattering mean free path, but do not exceed a transport mean free path, which is the distance at which the waves forget their initial direction. The transport mean free path defines the range limit of applicability of coherent inversion methods. Beyond it only incoherent methods based on transport or diffusion equations can be used. In this paper we assume a scattering regime where the CINT method is useful. It forms images by superposing cross-correlations of the array measurements, delayed by travel times between the receivers and the imaging points. As shown in, the cross-correlations must be computed locally, in appropriate time windows, and over limited receiver offsets. This introduces a smoothing in the CINT image formation, which is essential for stabilizing statistically the images, at the expense of resolution. The larger the random distortions of the array measurements, the more smoothing is needed and the worse the resolution. Thus, it is natural to ask if it is possible to improve the source localization by using the prior information that the sources have small support. We show that under generic conditions, the CINT imaging function is approximately a discrete convolution of the vector of source intensities discretized on the imaging mesh, with a blurring kernel. To reconstruct the sources we seek to undo the convolution using convex (\(l_1\)) optimization. We present an analysis of the method in a scattering regime where the random medium effects on the array measurements can be modeled by large wavefront distortions, as assumed in adaptive optics. We derive from first principles the CINT blurring kernel in this regime, and state the inversion problem as an \(l_1\) optimization. We also quantify the quality of the reconstruction with error estimates that depend on the separation of the sources, or of clusters of sources, but are independent of the source placement on or off the imaging mesh, as long as the sources are sufficiently far apart. The analysis shows that we can expect almost exact reconstructions when the sources are further apart than the CINT resolution limits. This is similar to the super-resolution results in, that show that one dimensional discrete convolutions can be undone by convex optimization, assuming that the minimum distance between the points in the support of the unknown vectors is \(2/f_c\), where \(f_c\) is the largest "frequency" in the Fourier transform of the convolution kernel. When the sources are clustered together, the \(l_1\) reconstruction is not guaranteed to be close to the true vector of source intensities in the point-wise sense. However, we show that its support is in the vicinity of the clusters, and its entries are related to the average source intensities there. The paper is organized as follows: We begin in section [2](#sect:form){reference-type="ref" reference="sect:form"} with the formulation of the inverse problem as an \(l_1\) optimization. The analysis of the method is in sections [3](#sect:anal){reference-type="ref" reference="sect:anal"}-[5](#sect:res){reference-type="ref" reference="sect:res"}, and we demonstrate its performance with numerical simulations in section [6](#sect:num){reference-type="ref" reference="sect:num"}. We end with a summary in section [7](#sect:sum){reference-type="ref" reference="sect:sum"}. # Formulation of the inverse problem {#sect:form} Consider the inversion setup illustrated in Figure [\[fig:setup\]](#fig:setup){reference-type="ref" reference="fig:setup"}, where \(N_s\) sources located at \(\vec{\bm{y}}_s\), for \(s = 1, \ldots, N_s\), emit signals \(f_s(t)\) that generate sound waves recorded at a remote array of receivers placed at \(\vec{\bm{x}}_r\), for \(r = 1, \ldots, N_r\). For simplicity we assume that the array aperture is planar and square, of side \(a\). This allows us to introduce a system of coordinates centered at the array, with range direction orthogonal to the array, and cross-range plane parallel to the array. In this system of coordinates we have \(\vec{\bm{x}}_r = (\bm{x}_r,0)\), with cross-range vectors \(\bm{x}_r = (x_{r,1},x_{r,2})\) satisfying \(|x_{r,1}|,|x_{r,2}| \le a/2\). The sources are at \(\vec{\bm{y}}_s = (\bm{y}_s,y_{s,3})\), with range coordinates \(y_{s,3}\) of order \(L\), satisfying \(L \gg a\), and two dimensional cross-range vectors \(\bm{y}_s\). In general, the signals \(f_s(t)\) emitted by the sources may be pulses, chirps or even noise-like, with Fourier transforms \[\hat f_s(\omega) = \int_{-\infty}^\infty dt \, f_s(t) e^{i \omega t} \label{eq:f2}\] supported in the frequency interval \([\omega_o-\pi B,\omega_o + \pi B]\), where \(\omega_o\) is the central frequency and \(B\) is the bandwidth. We denote the recorded waves by \(p(\vec{\bm{x}}_r,t)\), and use the linearity of the wave equation to write \[p(\vec{\bm{x}}_r,t) = \int_{\omega_o-\pi B}^{\omega_o + \pi B} \frac{d \omega}{2 \pi} \, \hat p(\vec{\bm{x}}_r, \omega) e^{-i \omega t}, \qquad \hat p(\vec{\bm{x}}_r,\omega) = \sum_{s=1}^{N_s} \hat f_s(\omega) \hat G(\vec{\bm{x}}_r,\vec{\bm{y}}_s,\omega) + \hat n(\vec{\bm{x}}_r,\omega), \label{eq:f1}\] for \(r = 1, \ldots, N_r\). Here the Green's function \(\hat G\) models the propagation of time harmonic waves in the medium, and \(\hat n\) denotes additive and uncorrelated noise. The inverse problem is to determine the sources from these array measurements. ## Imaging in homogeneous media The Green's function in media with constant speed \(c_o\) is \[\hat G_o(\vec{\bm{x}}_r,\vec{\bm{y}}_s,\omega) = \frac{\exp[i \omega \tau(\vec{\bm{x}}_r,\vec{\bm{y}}_s)]}{4 \pi |\vec{\bm{x}}_r-\vec{\bm{y}}_s|}, \label{eq:Go}\] where \[\tau(\vec{\bm{x}}_r,\vec{\bm{y}}_s) = {|\vec{\bm{x}}_r-\vec{\bm{y}}_s|}/{c_o} \label{eq:tau}\] is the travel time from the source at \(\vec{\bm{y}}_s\) to the receiver at \(\vec{\bm{x}}_r\). The measurements are of the form \[p_o(\vec{\bm{x}}_r,t) = \sum_{s=1}^{N_s} \frac{f_s\left(t- \tau(\vec{\bm{x}}_r,\vec{\bm{y}}_s)\right)}{4 \pi |\vec{\bm{x}}_r-\vec{\bm{y}}_s|} + n(\vec{\bm{x}}_r,t), \label{eq:f4}\] and in reverse time migration they are synchronized using travel time delays with respect to a presumed source location at a search point \(\vec{\bm{y}}\), and then superposed to form an image \[\begin{aligned} \mathcal{J}_o(\vec{\bm{y}}) = \sum_{r=1}^{N_r} p_o\left(\vec{\bm{x}}_r,\tau(\vec{\bm{x}}_r,\vec{\bm{y}})\right). \label{eq:f5} \end{aligned}\] The resolution limits of the imaging function [\[eq:f5\]](#eq:f5){reference-type="eqref" reference="eq:f5"} are well known. The cross-range resolution is of order \(\lambda_o L/a\), where \(\lambda_o = 2 \pi c_o/\omega_o\) is the central wavelength, and the range resolution is inverse proportional to the temporal support of \(f_s(t)\), which determines the precision of the travel time estimation. If the signals \(f_s(t)\) are pulses, their temporal support is of order \(1/B\), and the range resolution is of order \(c_o/B\). If they are chirps or other long signals that are known, then they must be compressed in time by cross-correlation (matched filtering) with the time reversed \(f_s(t)\) to achieve the \(c_o/B\) range resolution. If the signals are unknown and noise-like, then imaging must be based on cross-correlations of the array measurements, like in CINT. We refer to for the formulation of the inverse source problem as an \(l_1\) optimization, and recall from there that the resolution limits \(\lambda_o L/a\) and \(c_o/B\) also play a role in the successful recovery of the presumed sparse source support. ## Coherent interferometric imaging The imaging function \(\mathcal{J}_o(\vec{\bm{y}})\) does not work well in random media at ranges \(L\) that exceed a few scattering mean free paths. This is because the measurements have large random distortions that are very different than the additive noise \(n(\vec{\bm{x}}_r,t)\), and cannot be reduced by simply summing over the receivers as in [\[eq:f5\]](#eq:f5){reference-type="eqref" reference="eq:f5"}. To mitigate these distortions we image with the CINT function \[\begin{aligned} \mathcal{J}(\vec{\bm{y}}) = \int_{-\infty}^\infty d \omega \int_{-\infty}^\infty d \widetilde \omega \, \hat \phi\left(\frac{\widetilde \omega}{\Omega}\right) \sum_{r=1}^{N_r} \sum_{r' = 1}^{N_r} \psi\left(\frac{|\bm{x}_r-\bm{x}_{r'}|}{X(\omega)}\right) \hat p (\vec{\bm{x}}_r,\omega + \widetilde \omega/2) \overline{\hat p(\vec{\bm{x}}_{r'},\omega-\widetilde \omega/2)} \nonumber \\ \times \exp[-i (\omega + \widetilde \omega/2) \tau(\vec{\bm{x}}_r,\vec{\bm{y}}) + i (\omega- \widetilde \omega/2) \tau(\vec{\bm{x}}_{r'},\vec{\bm{y}})], \label{eq:f6} \end{aligned}\] where \(\hat \phi\) and \(\psi\) are smooth window functions of dimensionless argument and support of order one, and the domain of integration is restricted by the finite bandwidth that supports the measurements, \[\omega \pm \widetilde \omega/2 \in [\omega_o-\pi B,\omega_o + \pi B].\] As in reverse time migration, the travel times are used in [\[eq:f6\]](#eq:f6){reference-type="eqref" reference="eq:f6"} to synchronize the waves due to a presumed source at the search location \(\vec{\bm{y}}\). However, the image is formed by superposing cross-correlations of the array measurements \(p(\vec{\bm{x}}_r,t)\), instead of the measurements themselves. The cross-correlations are convolutions of \(p(\vec{\bm{x}}_r,t)\) with the time reversed \(p(\vec{\bm{x}}_{r'},t)\), for \(r, r' = 1, \ldots, N_r\). The time reversal appears as complex conjugation in the frequency domain, denoted with the bar in [\[eq:f6\]](#eq:f6){reference-type="eqref" reference="eq:f6"}. The time window \[\label{eq:defphi} \phi\left(\Omega t\right) = \frac{1}{2 \pi \Omega}\int_{-\infty}^\infty d \omega\,\hat \phi\left(\frac{\omega}{\Omega}\right)e^{-i \omega t}\] and spatial window \(\psi(|\bm{x}|/X)\) ensure that the cross-correlations are computed locally, over receiver offsets that do not exceed the distance \(X\), and over time offsets of order \(1/\Omega\). These threshold parameters account for the fact that scattering in random media decorrelates statistically the waves at frequencies separated by more than \(\Omega_d\), the decoherence frequency, and points separated by more than \(X_d\), the decoherence length. We refer to and the next section for more details. Here it suffices to recall that [\[eq:f6\]](#eq:f6){reference-type="eqref" reference="eq:f6"} is robust[^1] when \(X \lesssim X_d\) and \(\Omega \lesssim \Omega_d\), and that the image \(\mathcal{J}(\vec{\bm{y}})\) has a cross-range resolution of order \(\lambda_o L /X\) and a range resolution of order \(c_o/\Omega\). The best focus occurs at \(X \approx X_d\) and \(\Omega \approx \Omega_d\), so the decoherence parameters \(X_d\) and \(\Omega_d\) can be estimated with optimization, as explained in. To state the inverse problem as a convex optimization, we make the simplifying assumption that the sources emit the same known pulse \(f(t)\), so that \[\hat f_s(\omega) = \hat f(\omega) \rho(\vec{\bm{y}}_s), \qquad s = 1, \ldots, N_s, \label{eq:f7}\] for an unknown, complex valued amplitude \(\rho(\vec{\bm{y}})\). Using [\[eq:f7\]](#eq:f7){reference-type="eqref" reference="eq:f7"} in [\[eq:f1\]](#eq:f1){reference-type="eqref" reference="eq:f1"} and substituting in [\[eq:f6\]](#eq:f6){reference-type="eqref" reference="eq:f6"} we obtain \[\mathcal{J}(\vec{\bm{y}}) \approx \sum_{s=1}^{N_s} \sum_{s'=1}^{N_s} \rho(\vec{\bm{y}}_s) \overline{\rho(\vec{\bm{y}}_{s'})}\, \kappa(\vec{\bm{y}}, \vec{\bm{y}}_s,\vec{\bm{y}}_{s'}), \label{eq:f8}\] with kernel \[\begin{aligned} \kappa(\vec{\bm{y}},\vec{\bm{y}}_s,\vec{\bm{y}}_{s'}) =& \sum_{r=1}^{N_r} \sum_{r' = 1}^{N_r} \psi\left(\frac{|\bm{x}_r-\bm{x}_{r'}|}{X}\right) \int_{-\infty}^\infty d \omega \int_{-\infty}^\infty d \widetilde \omega \, \hat \phi\left(\frac{\widetilde \omega}{\Omega}\right) \hat f\left(\omega + {\widetilde \omega}/{2}\right) \overline{\hat f\left(\omega-{\widetilde \omega}/{2}\right)} \nonumber \\ & \hspace{-0.5in}\times \hat G\left(\vec{\bm{x}}_r,\vec{\bm{y}}_s,\omega+{\widetilde \omega}/{2}\right)\overline{\hat G(\vec{\bm{x}}_{r'},\vec{\bm{y}}_{s'},\omega-\widetilde \omega/2)}\exp[-i (\omega + \widetilde \omega/2) \tau(\vec{\bm{x}}_r,\vec{\bm{y}})] + i (\omega-\widetilde \omega/2)\tau(\vec{\bm{x}}_{r'},\vec{\bm{y}})], \label{eq:f9} \end{aligned}\] where the approximation is because we neglect the additive noise[^2]. In the analysis of the next section we take the Gaussian pulse \[\hat f(\omega) = \left(\frac{\sqrt{2 \pi}}{B}\right)^{1/2} \exp \left[ -\frac{(\omega-\omega_o)^2}{4 B^2}\right], \label{eq:f10}\] normalized by \[\|f \|_{_2} = \left[\int_{-\infty}^\infty dt \, |f(t)|^2 \right]^{1/2}= \left[\int_{-\infty}^\infty \frac{d \omega}{2 \pi}\, |\hat f(\omega)|^2\right]^{1/2} = 1.\] This choice allows us to obtain an explicit expression of the kernel [\[eq:f9\]](#eq:f9){reference-type="eqref" reference="eq:f9"}. Naturally, in practice the pulses may not be Gaussian, and the sources may emit different signals. The method described here still applies to such cases, with \(\rho(\vec{\bm{y}}_s)\) replaced in [\[eq:f8\]](#eq:f8){reference-type="eqref" reference="eq:f8"} by \(\|f_s\|_{_2}\), and the substitution \[\hat f(\omega + \widetilde \omega/2) \overline{\hat f(\omega-\widetilde \omega/2)} \leadsto \frac{\hat f_s(\omega + \widetilde \omega/2)}{\|f_s\|} \frac{\overline{\hat f_{s'}(\omega-\widetilde \omega/2)}}{\|f_{s'}\|_{_2}}\] in [\[eq:f9\]](#eq:f9){reference-type="eqref" reference="eq:f9"}. Since \(f_s\) is unknown in general, we may only estimate the kernel \(\kappa\) up to unknown, constant multiplicative factors. This still allows the determination of the location of the sources, but does not give good estimates of their intensities. ## The optimization problem {#sect:OPT} Let us consider a reconstruction mesh with \(N_z\) points denoted generically by \(\vec{\bm{z}}\), and name \(\bm{\rho}\) the column vector with entries given by the unknown \(\rho(\vec{\bm{z}})\). We sample the CINT image \(\mathcal{J}(\vec{\bm{y}})\) at \(N_y < N_z\) points, and gather these samples in the "data" vector \(\bm{d}\). It is natural to take \(N_y < N_z\), because we seek to super-resolve the CINT image, which is blurred by the kernel [\[eq:CINTK\]](#eq:CINTK){reference-type="eqref" reference="eq:CINTK"}. At first glance it appears that we may use equation [\[eq:f8\]](#eq:f8){reference-type="eqref" reference="eq:f8"} to formulate the inversion as an optimization problem for recovering the rank one matrix \(\bm{\rho} \bm{\rho}^\star\), as in. However, as shown in the next section, in strong random media where CINT is needed, the kernel \(\kappa(\vec{\bm{y}},\vec{\bm{z}},\vec{\bm{z}}')\) is large only when \(\vec{\bm{z}}\) and \(\vec{\bm{z}}'\) are nearby. In fact, for reasonable mesh sizes on which we can expect to obtain unique reconstructions, the kernel satisfies \[\kappa(\vec{\bm{y}},\vec{\bm{z}},\vec{\bm{z}}') \approx \left\{ \begin{array}{ll} \kappa(\vec{\bm{y}},\vec{\bm{z}},\vec{\bm{z}}) \quad &\mbox{if} ~ \vec{\bm{z}}'=\vec{\bm{z}}, \\ 0 & \mbox{if}~ \vec{\bm{z}} \ne \vec{\bm{z}}'. \end{array}\right. \label{eq:diagKappa}\] Thus, only the diagonal entries \(|\rho(\vec{\bm{z}})|^2\) of \(\bm{\rho} \bm{\rho}^\star\) play a role. These are the source intensities and we denote by \(\bm{u} \in \mathbb{R}^{N_z}\) the vector of unknowns formed by them. Equation [\[eq:f8\]](#eq:f8){reference-type="eqref" reference="eq:f8"} gives \[\bm{\mathcal{M}}\, \bm{u} \approx \bm{d}, \label{eq:f11}\] where \(\bm{\mathcal{M}}\) is the \(N_y \times N_z\) "measurement" matrix with entries \(\kappa(\vec{\bm{y}},\vec{\bm{z}},\vec{\bm{z}}).\) We formulate the inversion as the \(l_1\) optimization problem \[\min_{\bm{u} \in \mathbb{R}^{N_z}} \|\bm{u}\|_{_1} ~ ~ \mbox{such that} ~~ \|\bm{\mathcal{M}}\, \bm{u}-\bm{d}\|_{_2} \le ~\mbox{tolerance}, \label{eq:f12}\] where the tolerance accounts for additive noise effects and the random fluctuations of the CINT imaging function, which are small for large enough array aperture \(a\) and bandwidth \(B\), as shown in . We prove in section [4](#sect:CINT){reference-type="ref" reference="sect:CINT"} that the left hand side in [\[eq:f11\]](#eq:f11){reference-type="eqref" reference="eq:f11"} is approximately a discrete convolution. The \(l_1\) optimization is useful in this context, and recovers well sources that are well separated, as expected from the results in. We rediscover such results in this paper using a different analysis. We also consider cases where the sources are clustered together, and show that although we cannot expect good reconstructions in the point-wise sense, the \(l_1\) minimizer is supported in the vicinity of the clusters, and estimates the average source intensity there. # Setup of the analysis {#sect:anal} We introduce in this section the random wave speed model and the scaling assumptions which define the relations between the wavelength \(\lambda_o\), the typical size \(\ell\) of the inhomogeneities in the medium, the standard deviation \(\sigma\) of the fluctuations of the wave speed, the array aperture \(a\), the range \(L\), and the extent of the imaging region. The scaling allows us to describe the scattering effects of the random medium as large wavefront distortions. This is a simple wave propagation model that is convenient for analysis, and captures qualitatively all the important features of imaging with CINT. That is to say, equation [\[eq:f11\]](#eq:f11){reference-type="eqref" reference="eq:f11"} holds in general scattering regimes, and the CINT kernel \(\kappa\) has a similar form, but the mathematical expression of the decoherence length \(X_d\) and frequency \(\Omega_d\), which quantify the blurring by the kernel, are expected to change. The expressions of \(X_d\) and \(\Omega_d\) in terms of \(\lambda_o, \sigma, \ell\) and \(L\) are needed for analysis, but they are unlikely to be known in practice. This is why one should determine the decoherence parameters directly form the data or adaptively, during the CINT image formation, as in. The model of the wave speed \(c(\vec{\bm{x}})\) is \[c(\vec{\bm{x}}) = c_o \left[ 1 + \sigma \mu \left(\frac{\vec{\bm{x}}}{\ell}\right) \right]^{-1/2}, \label{eq:a1}\] where \(\mu\) is a mean zero, stationary random process of dimensionless argument. We suppose that \(\mu\) is bounded almost surely and denote by \(\mathcal{R}\) its auto-correlation, assumed isotropic and Gaussian for convenience \[\mathcal{R}(\vec{\bm{x}}) = \mathbb{E} \left[ \mu(\vec{\bm{x}} + \vec{\bm{x}}') \mu(\vec{\bm{x}})\right] = e^{- |\vec{\bm{x}}|^2/2}. \label{eq:a2}\] Then, \(\sigma \ll 1\) quantifies the small amplitude of the fluctuations of \(c(\vec{\bm{x}})\), and \(\ell\) is the correlation length, which characterizes the typical size of the inhomogeneities in the medium. We explain in section [3.1](#sect:sc1){reference-type="ref" reference="sect:sc1"} how far the waves should propagate in media modeled by [\[eq:a1\]](#eq:a1){reference-type="eqref" reference="eq:a1"}, so that the cumulative scattering effects can be described by large wavefront distortions. We are interested in imaging with finite size arrays at long distances, so the waves propagate in the range direction, within a cone (beam) of small opening angle. This is the paraxial regime defined in section [3.2](#sect:sc2){reference-type="ref" reference="sect:sc2"}. We describe in section [3.3](#sect:sc3){reference-type="ref" reference="sect:sc3"} the wave randomization quantified by the scattering mean free path, and the statistical decorrelation quantified by the decoherence length and frequency. We end the section with a summary of the scaling assumptions. ## Wave scattering regime {#sect:sc1} We use a geometrical optics (Rytov) wave propagation model that holds in high frequency regimes with separation of scales \[\label{eq:a3} \lambda_o \ll \ell \ll L,\] and standard deviation \(\sigma\) of the fluctuations satisfying \[\sigma \ll \min \left\{ \left(\frac{\ell}{L}\right)^{3/2}, \frac{\sqrt{\ell \lambda_o}}{L}\right\}. \label{eq:a4}\] It is shown in that the first bound in [\[eq:a4\]](#eq:a4){reference-type="eqref" reference="eq:a4"} ensures that the rays remain straight and the variance of the amplitude of the Green's function is negligible, so we can use the same geometrical spreading factor as in the homogeneous medium. The second bound in [\[eq:a4\]](#eq:a4){reference-type="eqref" reference="eq:a4"} ensures that only first order (in \(\sigma\)) corrections of the travel time matter, so the propagation model is \[\hat G(\vec{\bm{x}},\vec{\bm{y}},\omega) \approx \hat G_o(\vec{\bm{x}},\vec{\bm{y}},\omega) \exp \left[ i \omega \tau(\vec{\bm{x}},\vec{\bm{y}})\frac{\sigma}{2} \int_0^1 d \vartheta \, \mu \left(\frac{(1-\vartheta) \vec{\bm{y}} + \vartheta \vec{\bm{x}}}{\ell}\right) \right]. \label{eq:a6}\] Let us write the random phase correction in [\[eq:a6\]](#eq:a6){reference-type="eqref" reference="eq:a6"} as \[\omega \delta \tau(\vec{\bm{x}},\vec{\bm{y}}) = \frac{(2 \pi)^{1/4}}{2} \sigma k \sqrt{\ell |\vec{\bm{x}}-\vec{\bm{y}}|} \, \nu(\vec{\bm{x}},\vec{\bm{y}}), \label{eq:a7p}\] where \(k = \omega/c_o\) is the wavenumber and \[\nu(\vec{\bm{x}},\vec{\bm{y}}) = \frac{1}{(2 \pi)^{1/4}} \sqrt{\frac{|\vec{\bm{x}}-\vec{\bm{y}}|}{\ell}} \int_0^1 d \vartheta \, \mu \left(\frac{(1-\vartheta) \vec{\bm{y}} + \vartheta \vec{\bm{x}}}{\ell}\right) \label{eq:a7}\] is defined by the integral of the fluctuations along the straight ray between \(\vec{\bm{y}}\) and \(\vec{\bm{x}}\). It is shown in [@borcea2011enhanced Lemma 3.1] that \(\nu(\vec{\bm{x}},\vec{\bm{y}})\) converges in distribution as \(\ell/|\vec{\bm{x}}-\vec{\bm{y}}| \sim \ell/L \to 0\) to a Gaussian process. Its mean \[\mathbb{E} \left[ \nu(\vec{\bm{x}},\vec{\bm{y}})\right] = 0,\] and variance \[\mathbb{E} \left[ \nu^2(\vec{\bm{x}},\vec{\bm{y}}) \right] = \frac{|\vec{\bm{x}}-\vec{\bm{y}}|}{\ell \sqrt{2 \pi}}\int_0^1 d \vartheta \int_0^1 d \vartheta' \, \mathcal{R} \left( \frac{(\vartheta'-\vartheta) (\vec{\bm{x}}-\vec{\bm{y}})}{\ell}\right) \approx 1, \label{eq:a8}\] are calculated from definition [\[eq:a7\]](#eq:a7){reference-type="eqref" reference="eq:a7"} and the expression [\[eq:a2\]](#eq:a2){reference-type="eqref" reference="eq:a2"} of \(\mathcal{R}\). The approximation is for \(\ell/L\) small. We conclude from [\[eq:a7p\]](#eq:a7p){reference-type="eqref" reference="eq:a7p"}, [\[eq:a8\]](#eq:a8){reference-type="eqref" reference="eq:a8"} and \(k = O( k_o)\), with \(k_o = 2 \pi /\lambda_o\), that the random phase fluctuations in [\[eq:a6\]](#eq:a6){reference-type="eqref" reference="eq:a6"} have standard deviation of order \(\sigma \sqrt{\ell L}/\lambda_o\). When this is small, the random medium effects are negligible and any coherent imaging method works well. We are interested in the case of large fluctuations, so we ask that \[\label{eq:a9} \sigma \gg \frac{\lambda_o}{\sqrt{\ell L}}.\] This is consistent with [\[eq:a4\]](#eq:a4){reference-type="eqref" reference="eq:a4"} when \[\frac{\lambda_o/\sqrt{\ell L}}{(\ell/L)^{3/2}} = \frac{\lambda_o L}{\ell^2} \ll 1, \qquad \frac{\lambda_o/\sqrt{\ell L}}{\sqrt{\ell \lambda_o}/{L} } = \frac{\sqrt{\lambda_o L}}{\ell} \ll 1,\] so we tighten our assumption [\[eq:a3\]](#eq:a3){reference-type="eqref" reference="eq:a3"} on the correlation length as \[\sqrt{\lambda_o L} \ll \ell \ll L. \label{eq:a10}\] ## The paraxial regime {#sect:sc2} Suppose that the sources are contained in the search (imaging) region \[\label{eq:imreg} \mathfrak{D} = \left[-D/2,D/2\right]\times\left[-D/2,D/2\right]\left[-D_3/2,D_3/2\right],\] which is a rectangular prism of sides \(D\) in cross-range and \(D_3\) in range, as illustrated in Figure [\[fig:setup\]](#fig:setup){reference-type="ref" reference="fig:setup"}. When \(D\) and the array aperture \(a\) are small with respect to the range scale \(L\), the rays connecting the sources and the receivers are contained within a cone (beam) of small opening angle, and we can use the paraxial approximation to simplify the calculations. The paraxial regime is defined by the scaling relations \[\lambda_o \ll D \lesssim a \ll L, \qquad D_3 \ll L, \qquad \frac{a^4}{\lambda_o L^3} \ll 1, \qquad \frac{a^2 D_3}{\lambda_o L^2} \ll 1, \label{eq:a11}\] so that for \(\vec{\bm{y}} = (\bm{y},y_3)\) and \(\vec{\bm{x}} = (\bm{x},0)\) we get \[\begin{aligned} k |\vec{\bm{x}}-\vec{\bm{y}}| &= k \left( y_3 + \frac{|\bm{x}|^2}{2 L}-\frac{\bm{x} \cdot \bm{y}}{L} + \frac{|\bm{y}|^2}{2 L} \right) + O \left( \frac{a^4}{\lambda_o L^3}\right) + O\left(\frac{a^2 D_3}{\lambda_o L^2}\right) \nonumber \\ &\approx k \left( y_3 + \frac{|\bm{x}|^2}{2 L}-\frac{\bm{x} \cdot \bm{y}}{L} + \frac{|\bm{y}|^2}{2 L} \right), \label{eq:a12} \end{aligned}\] and \[\frac{1}{4 \pi |\vec{\bm{x}}-\vec{\bm{y}}|} = \frac{1}{4 \pi L} \left[ 1 + O\left(\frac{D_3}{L}\right) + O\left(\frac{a^2}{L^2} \right) \right]\approx \frac{1}{4 \pi L}. \label{eq:a13}\] These approximations are proved in appendix [\[ap:DerPar\]](#ap:DerPar){reference-type="ref" reference="ap:DerPar"}, and the deterministic factor in [\[eq:a6\]](#eq:a6){reference-type="eqref" reference="eq:a6"} becomes \[\hat G_o(\vec{\bm{x}},\vec{\bm{y}},\omega) \approx \frac{1}{4 \pi L} \exp \left[ i k \left(y_3 + \frac{|\bm{x}|^2}{2 L}-\frac{\bm{x} \cdot \bm{y}}{L} + \frac{|\bm{y}|^2}{2 L} \right) \right]. \label{eq:a14}\] ## Randomization and statistical decorrelation of the waves {#sect:sc3} Here we quantify the scattering mean free path \(\mathcal{S}\), the length scale on which the waves randomize (lose coherence), and the decoherence length \(X_d\) and frequency \(\Omega_d\), which describe the statistical decorrelation of the waves due to scattering. These important scales appear in the definition of the CINT blurring kernel defined in section [4](#sect:CINT){reference-type="ref" reference="sect:CINT"}. This result, proved in appendix [\[ap:moments\]](#ap:moments){reference-type="ref" reference="ap:moments"}, shows that the wavefront distortions due to scattering in the random medium do not average out. The expectation of \(\hat{G}\) is not the same as the Green's function in the homogeneous medium, but decays exponentially with the distance of propagation on the scale \(\mathcal{S}(\omega)\), the scattering mean free path. The scaling assumption [\[eq:a9\]](#eq:a9){reference-type="eqref" reference="eq:a9"} and definition [\[eq:prop1.2\]](#eq:prop1.2){reference-type="eqref" reference="eq:prop1.2"} give \[\frac{|\vec{\bm{x}}-\vec{\bm{y}}|}{\mathcal{S}(\omega)} = O \left(\frac{L}{\mathcal{S}(\omega)}\right) = O(\sigma^2 k^2 \ell L) \gg 1,\] which is why the expectation in [\[eq:prop1.1\]](#eq:prop1.1){reference-type="eqref" reference="eq:prop1.1"} is almost zero. The standard deviation of the fluctuations is approximately \[\mbox{std}[\hat G(\vec{\bm{x}},\vec{\bm{y}},\omega)] \approx \sqrt{\Big|\hat G_o(\vec{\bm{x}},\vec{\bm{y}},\omega)\Big|^2-\Big|\mathbb{E}\big[ \hat G(\vec{\bm{x}},\vec{\bm{y}},\omega)\big] \Big|^2} \approx \Big| \hat G_o(\vec{\bm{x}},\vec{\bm{y}},\omega)\Big|,\] where we used that \(|\hat G(\vec{\bm{x}},\vec{\bm{y}},\omega)| \approx |\hat G_o(\vec{\bm{x}},\vec{\bm{y}},\omega)|\). Thus, the random fluctuations of the waves dominate their coherent part (the expectation) at the ranges considered in our analysis, \[\frac{ \Big| \mathbb{E} \big[ \hat G(\vec{\bm{x}},\vec{\bm{y}},\omega) \big] \Big|}{\mbox{std}[\hat G(\vec{\bm{x}},\vec{\bm{y}},\omega)]} \approx e^{-\frac{|\vec{\bm{x}}-\vec{\bm{y}}|}{S(\omega)}} \approx 0,\] and the wave is randomized. Reverse time migration or standard \(l_1\) optimization methods cannot mitigate these large random distortions, as we illustrate with numerical simulations. This is why we base our inversion on the CINT method. Moment formula [\[eq:prop2.1\]](#eq:prop2.1){reference-type="eqref" reference="eq:prop2.1"} is proved in appendix [\[ap:moments\]](#ap:moments){reference-type="ref" reference="ap:moments"}, and the inequalities in [\[eq:prop2.2\]](#eq:prop2.2){reference-type="eqref" reference="eq:prop2.2"} are due to assumption [\[eq:a9\]](#eq:a9){reference-type="eqref" reference="eq:a9"}. The first exponential factor in [\[eq:prop2.1\]](#eq:prop2.1){reference-type="eqref" reference="eq:prop2.1"} accounts for the randomization due to the travel time fluctuations between the two ranges. In our scaling \(|y_3-y_3'| \lesssim D_3\), and by the last inequality in [\[eq:a11\]](#eq:a11){reference-type="eqref" reference="eq:a11"}, and the definition of the scattering mean free path, we have \[\frac{|y_3-y_3'|}{\mathcal{S}(\omega_o)} = O \left(D_3 \frac{\sigma^2\ell}{\lambda_o^2} \right) \ll O\left(\frac{\lambda_o L^2}{a^2} \frac{\sigma^2 \ell}{\lambda_o^2} \right) \ll O\left(\frac{\ell^2}{a^2} \right), \label{eq:a15}\] where we used the bound [\[eq:a4\]](#eq:a4){reference-type="eqref" reference="eq:a4"} on \(\sigma\). It is shown in that the standard deviation of the CINT image is small with respect to the expectation of its peak value (i.e., the imaging function is statistically stable) when \[a > \ell. \label{eq:a16}\] Stability is essential for imaging to succeed, so we ask that the array aperture satisfy [\[eq:a16\]](#eq:a16){reference-type="eqref" reference="eq:a16"}, and conclude from [\[eq:a15\]](#eq:a15){reference-type="eqref" reference="eq:a15"} that the second moments [\[eq:prop1.1\]](#eq:prop1.1){reference-type="eqref" reference="eq:prop1.1"} simplify as[^3] \[\begin{aligned} \mathbb{E}\left[ \hat G(\vec{\bm{x}},\vec{\bm{y}},\omega) \overline{\hat G(\vec{\bm{x}}',\vec{\bm{y}}',\omega')}\right] &\approx \hat G_o(\vec{\bm{x}},\vec{\bm{y}},\omega) \overline{\hat G_o(\vec{\bm{x}}',\vec{\bm{y}}',\omega')}e^{-\frac{(\omega-\omega')^2}{2 \Omega_d^2}-\frac{|\bm{y}-\bm{y}'|^2 + (\bm{y}-\bm{y}')\cdot(\bm{x}-\bm{x}') + |\bm{x}-\bm{x}'|^2}{2 X_d^2}}, \label{eq:a17} \end{aligned}\] with \(\hat G_o\) given in [\[eq:a14\]](#eq:a14){reference-type="eqref" reference="eq:a14"}. The exponential decay in [\[eq:a17\]](#eq:a17){reference-type="eqref" reference="eq:a17"} models the statistical decorrelation of the waves due to scattering. In our context, the spatial decorrelation, modeled by the decay in \(\bm{x}-\bm{x}'\) and \(\bm{y}-\bm{y}'\), can be explained by the fact that rays connecting sources to far apart receivers traverse through different parts of the random medium. Because \(\mu\) does not have long range correlations, the fluctuations of the travel time along such different rays are statistically uncorrelated. The waves at far apart frequencies are uncorrelated because they interact differently with the random medium. This gives the decay in \(\omega-\omega'\) in equation [\[eq:a17\]](#eq:a17){reference-type="eqref" reference="eq:a17"}. Definition [\[eq:f9\]](#eq:f9){reference-type="eqref" reference="eq:f9"} of the CINT kernel involves the superposition of \(\hat G(\vec{\bm{x}},\vec{\bm{y}},\omega) \overline{\hat G(\vec{\bm{x}}',\vec{\bm{y}}',\omega')}\) over the array elements and frequencies. If the array aperture \(a\) is large with respect to \(X_d\), the superposition stabilizes statistically because we sum many uncorrelated entries, as in the law of large numbers. This is why CINT is robust with respect to the uncertainty of the fluctuations of the wave speed, as shown in. ## Summary of the scaling assumptions {#sect:sc4} We gather here the scaling assumptions stated throughout the section, and complement them with extra assumptions on the bandwidth and the size of the imaging region. We refer to appendix [\[ap:verify\]](#ap:verify){reference-type="ref" reference="ap:verify"} for the verification of their consistency. The wavelength \(\lambda_o\) is the smallest length scale, and the range \(L\) is the largest. The assumptions [\[eq:a16\]](#eq:a16){reference-type="eqref" reference="eq:a16"} and [\[eq:a11\]](#eq:a11){reference-type="eqref" reference="eq:a11"} on the aperture are \[\ell < a \ll (\lambda_o L^3)^{1/4}. \label{eq:as1}\] The upper bound ensures that we can use the paraxial approximation and the lower bound gives \(a > \ell \gg X_d\), so that the CINT image is statistically stable. Assumption [\[eq:a10\]](#eq:a10){reference-type="eqref" reference="eq:a10"} combined with [\[eq:as1\]](#eq:as1){reference-type="eqref" reference="eq:as1"} gives that the correlation length of the wave speed fluctuations should satisfy \[\sqrt{\lambda_o L} \ll \ell \ll (\lambda_o L^3)^{1/4} \ll L. \label{eq:as2}\] The standard deviation \(\sigma\) of the fluctuations is bounded above as in [\[eq:a4\]](#eq:a4){reference-type="eqref" reference="eq:a4"}, and below as in [\[eq:a9\]](#eq:a9){reference-type="eqref" reference="eq:a9"}, \[\frac{\lambda_o}{\sqrt{\ell L}} \ll \sigma \ll \frac{\sqrt{\ell \lambda_o}}{L}. \label{eq:as3}\] The cross-range and range sizes \(D\) and \(D_3\) of the imaging region should be large with respect to the CINT resolution limits of \(\lambda_o L/X\) in cross-range and \(c_o/\Omega\) in range (see next section), so we can observe the image focus. We take the threshold parameters \[X/X_d = O(1), \qquad \Omega/\min\{\Omega_d,B\} = O(1), \label{eq:as4pp}\] and recalling the scaling assumptions [\[eq:a11\]](#eq:a11){reference-type="eqref" reference="eq:a11"} that allow us to use the paraxial approximation, we obtain \[\frac{c_o}{\Omega} \ll D_3 \ll \frac{\lambda_o L^2}{a^2} \quad \mbox{and} \quad \frac{\lambda_o L}{X} \ll D \lesssim a. \label{eq:as5}\] In general, the CINT image is statistically stable if in addition to having \(a \gg X_d\), which follows from [\[eq:prop2.2\]](#eq:prop2.2){reference-type="eqref" reference="eq:prop2.2"} and [\[eq:as1\]](#eq:as1){reference-type="eqref" reference="eq:as1"}, the bandwidth \(B\) is larger than the decoherence frequency \(\Omega_d\). However, for the propagation model [\[eq:a6\]](#eq:a6){reference-type="eqref" reference="eq:a6"} considered in this section, where the effect of the random medium consists only of wavefront distortions and no delay spread (reverberations), the bandwidth \(B\) does not play a role in the stabilization of CINT, as shown in. Thus, we study imaging in both narrowband and broadband regimes: **The narrowband regime** is defined by \(B\) satisfying \[\omega_o\left(\frac{a}{L}\right)^2 \ll B \ll \omega_o \min\left\{1, \frac{\lambda_o L}{a X_d} \right\}= \omega_o \frac{\lambda_o L}{a X_d}. \label{eq:as4}\] As verified in Appendix [\[ap:verify\]](#ap:verify){reference-type="ref" reference="ap:verify"}, \[\frac{\lambda_o L}{X_d} \ll \ell < a, \qquad \frac{\lambda_o L}{a X_d} \ll \frac{\Omega_d}{\omega_o}, \label{eq:as4p}\] so \(B \ll \Omega_d\). This choice leads to a simpler expression of the CINT blurring kernel, but since \(\Omega\) is of the order of \(B\), the range resolution is the same as in the homogeneous medium, and cannot be improved with optimization unless the sources are very far apart in range. However, the optimization can improve the cross-range focusing. **The broadband regime** is defined by \[\Omega_d \ll B \ll \omega_o, \label{eq:as4bb}\] so we may seek to improve the CINT resolution in both range and cross-range. The expression of the CINT kernel is more complicated in this case, but it simplifies slightly when \[\frac{\lambda_o}{\sqrt{\ell L}} \ll \frac{\lambda_o^{2/3} \ell^{1/6}}{L^{5/6}} \ll \sigma \ll \frac{\sqrt{\ell \lambda_o}}{L}, \qquad \frac{\ell}{a} = O(1). \label{eq:as3bb}\] We present the analysis that uses these conditions, which say that the fluctuations in the random medium are even stronger than in [\[eq:as3\]](#eq:as3){reference-type="eqref" reference="eq:as3"}, but the correlation length is not much smaller than \(a\). Extensions to larger apertures are possible, although the analysis is more complicated. # The CINT blurring kernel {#sect:CINT} Here we derive the CINT convolution model. To obtain an explicit expression of the kernel [\[eq:f9\]](#eq:f9){reference-type="eqref" reference="eq:f9"}, we use the Gaussian pulse [\[eq:f10\]](#eq:f10){reference-type="eqref" reference="eq:f10"} and the Gaussian threshold windows \[\hat \phi\left(\frac{\widetilde \omega}{\Omega}\right) = e^{-\frac{\widetilde \omega^2}{2 \Omega^2}}, \qquad \psi\left(\frac{|\widetilde {\bx}|}{X}\right) = e^{-\frac{|\widetilde {\bx}|^2}{2 \Omega^2}}, \label{eq:R1}\] with \(X\) and \(\Omega\) satisfying \[{X}/{X_d} = O(1), \qquad \Omega = \left\{ \begin{array}{ll} B \quad & \mbox{in narrowband regime} \\ \\ O(\Omega_d) & \mbox{in broadband regime}. \end{array} \right. \label{eq:R2}\] As stated previously, and shown in, \(X_d\) and \(\Omega_d\) can be estimated adaptively, by optimizing the focusing of the CINT image. This is why we can assume that \(X_d\) is known approximately. The same holds for \(\Omega_d\), if the bandwidth is big enough. The expression of the CINT kernel is simpler in the narrowband scaling [\[eq:as4\]](#eq:as4){reference-type="eqref" reference="eq:as4"}, where \(B \ll \Omega_d\), as shown in section [4.1](#sect:CINT1){reference-type="ref" reference="sect:CINT1"}, and we take \(\Omega = B\). The broadband regime is discussed in section [4.2](#sect:CINT2){reference-type="ref" reference="sect:CINT2"}. Typically, the receivers are separated by distances of order \(\lambda_o\), so that they behave collectively as an array. Since \(\lambda_o \ll a\), we have \(N_r = O\left(a^2/\lambda_o^2\right) \gg 1,\) and we can approximate the sums in [\[eq:f9\]](#eq:f9){reference-type="eqref" reference="eq:f9"} by integrals \[\sum_{r=1}^N \leadsto \frac{N_r}{a^2} \int_{\mathcal{A}} d \bm{x},\] where \(\mathcal{A}\) denotes the array aperture, the square of side \(a\). To avoid specifying the finite aperture in the integrals, and to simplify the calculations, we use a Gaussian apodization factor \[\psi_\mathcal{A}(\bm{x}) = \exp\left[-\frac{|\bm{x}|^2}{2(a/6)^2}\right], \label{eq:R3}\] which is negligible outside the disk of radius \(a/2\). ## The CINT kernel in the narrowband regime {#sect:CINT1} The calculation of the kernel [\[eq:f9\]](#eq:f9){reference-type="eqref" reference="eq:f9"} is in appendix [\[sect:apCINT\]](#sect:apCINT){reference-type="ref" reference="sect:apCINT"}, and we state the results in the next proposition. The parameter \(R\) defined in [\[eq:prop3.1\]](#eq:prop3.1){reference-type="eqref" reference="eq:prop3.1"} is the CINT cross-range resolution limit, the length scale of exponential decay of the kernel \(\kappa(\vec{\bm{y}},\vec{\bm{z}},\vec{\bm{z}}')\) with \(\overline{\bz}-\bm{y}\). Definitions [\[eq:R4\]](#eq:R4){reference-type="eqref" reference="eq:R4"}, [\[eq:prop2.2\]](#eq:prop2.2){reference-type="eqref" reference="eq:prop2.2"} and assumption [\[eq:R2\]](#eq:R2){reference-type="eqref" reference="eq:R2"} give that \[X_e = O(X_d) \ll a,\] so the resolution is worse than in homogeneous media, \[R \gg \frac{L}{k_o a}. \label{eq:R6}\] This is due to the smoothing needed to stabilize statistically the image. The goal of the convex optimization [\[eq:f12\]](#eq:f12){reference-type="eqref" reference="eq:f12"} is to overcome this blurring and localize better the sources in cross-range. The parameter \(R_3\) is the CINT range resolution limit. Because we are in the narrowband regime, we obtain from definition [\[eq:R4\]](#eq:R4){reference-type="eqref" reference="eq:R4"} and [\[eq:R2\]](#eq:R2){reference-type="eqref" reference="eq:R2"} that \(\Omega_e \approx B,\) and therefore \(R_3\) is similar to the range resolution in homogeneous media, \[R_3 = \frac{c_o}{\Omega_e} \approx \frac{c_o}{B}.\] The results obtained in for imaging with \(l_1\) optimization in homogeneous media show that it is not possible to improve the \(c_o/B\) range resolution, unless the sources are very far apart in range. We cannot expect to do better in random media, so we do not seek any super-resolution in range, in the narrowband regime. Note that by the first inequality in [\[eq:as4p\]](#eq:as4p){reference-type="eqref" reference="eq:as4p"} we have \(\widetilde{R} \ll X_d\), so the kernel decays with the offsets \(\widetilde{\bz}\) and \(\widetilde{z}\) on the length scales \(\widetilde{R}\) and \(\widetilde{R}_3\). These scales are, up to a constant of order one, the resolution limits of imaging in homogeneous media. ## The CINT kernel in the broadband regime {#sect:CINT2} The expression of the CINT kernel is stated in the next proposition, proved in appendix [\[sect:apCINT\]](#sect:apCINT){reference-type="ref" reference="sect:apCINT"}. Because \(X_e = O(X_d)\) and \(\Omega_e = \Omega_d\), we obtain from definitions [\[eq:defTheta\]](#eq:defTheta){reference-type="eqref" reference="eq:defTheta"} and [\[eq:prop2.2\]](#eq:prop2.2){reference-type="eqref" reference="eq:prop2.2"} that \[\theta = O\left(\frac{\ell}{a} \right) = O(1), \label{eq:KernBB1}\] where we used the assumption [\[eq:as3bb\]](#eq:as3bb){reference-type="eqref" reference="eq:as3bb"} on the aperture.[^4] The first term in the exponential in [\[eq:CINTKernBB\]](#eq:CINTKernBB){reference-type="eqref" reference="eq:CINTKernBB"} gives the focusing in cross-range, which is the same as in the narrowband case: \(|\overline{\bz}-\bm{y}| = O(R)\). This means that the denominators in [\[eq:CINTKernBB\]](#eq:CINTKernBB){reference-type="eqref" reference="eq:CINTKernBB"} are order one, \[1 + \frac{|\overline{\bz}-\bm{y}|^2}{2 \theta^2 R^2} = O(1).\] The second term in the exponential in [\[eq:CINTKernBB\]](#eq:CINTKernBB){reference-type="eqref" reference="eq:CINTKernBB"} gives the focusing in range. In our setting we have by the paraxial approximation that \[\overline{z}_3-y_3 + \frac{|\overline{\bz}|^2-|\bm{y}|^2}{2L} \approx |(\overline{\bz},\overline{z}_3)|-|(\bm{y},y_3)|,\] so CINT estimates the distance from the center of the array to \((\overline{\bz},\overline{z}_3)\) with resolution of order \(R_3\). Since \(\Omega_e \ll B\), this resolution is worse than in homogeneous media \[R_3 = \frac{c_o}{\Omega_e} = O\left(\frac{c_o}{B}\right) = O(\widetilde{R}_3),\] so in the broadband regime it makes sense to seek an improvement of both the range and cross-range resolution with optimization. Equations [\[eq:CINTKernBB\]](#eq:CINTKernBB){reference-type="eqref" reference="eq:CINTKernBB"} and [\[eq:KernBB\]](#eq:KernBB){reference-type="eqref" reference="eq:KernBB"} show that the kernel decays with the offset \(\widetilde{z}_3\) on the same scale \(\widetilde{R}_3\) as before. To see the decay with the offset \(\widetilde{\bz}\), we note that the last two terms in [\[eq:KernBB\]](#eq:KernBB){reference-type="eqref" reference="eq:KernBB"} satisfy \[\begin{aligned} \frac{|\widetilde{\bz}|^2}{2} \left[ \frac{1}{\gamma X_d^2 } + \frac{1}{\widetilde{R}^2} \right]-\frac{\left|\frac{(\overline{\bz}-\bm{y})}{R} \cdot\frac{\widetilde{\bz}}{\widetilde{R}}\right|^2 }{4 \theta^2\big[1 + \frac{|\overline{\bz}-\bm{y}|^2}{2 \theta^2 R^2}\big]} &= \frac{|\widetilde{\bz}|^2}{2 R^2 \big[1 + \frac{|\overline{\bz}-\bm{y}|^2}{2 \theta^2 R^2}\big]} \left[ 1 + \frac{\widetilde{R}^2 \big[1 + \frac{|\overline{\bz}-\bm{y}|^2}{2 \theta^2 R^2}\big]}{\gamma X_d^2} \right] + \frac{\frac{|\overline{\bz}-\bm{y}|^2}{R^2} \frac{|\widetilde{\bz}|^2}{\widetilde{R}^2}-\left|\frac{(\overline{\bz}-\bm{y})}{R}\cdot\frac{\widetilde{\bz}}{\widetilde {R}}\right|^2}{4 \theta^2 \big[1 + \frac{|\overline{\bz}-\bm{y}|^2}{2 \theta^2 R^2}\big]} \\ &\gtrsim \frac{|\widetilde{\bz}|^2}{2 R^2 \big[1 + \frac{|\overline{\bz}-\bm{y}|^2}{2 \theta^2 R^2}\big]}, \end{aligned}\] where we used that \(\widetilde{R} \ll X_d\), as explained in the previous section. This shows that the kernel decays with the cross-range offsets on the same scale \(\widetilde{R}\) as before. ## The approximate convolution model {#sect:CINTConv} Let us discretize the imaging region \(\mathfrak{D}\) defined in [\[eq:imreg\]](#eq:imreg){reference-type="eqref" reference="eq:imreg"} on a mesh with size \(\vec{\bm{h}} = (h,h,h_3)\). In principle, the steps \(h\) and \(h_3\) may be chosen arbitrarily small, to avoid discretization error due to sources being off the mesh. However, we know from and the analysis below and the numerical simulations that we cannot expect reconstructions at scales that are finer than the resolution limits in homogeneous media. This motivates us to formulate the inversion using the assumption that the sources are further apart than \(3 \widetilde{R}\) in cross-range and \(3 \widetilde{R}_3\) in range. This leads to a simpler optimization problem because by Propositions [\[prop.3\]](#prop.3){reference-type="ref" reference="prop.3"} and [\[prop.3bb\]](#prop.3bb){reference-type="ref" reference="prop.3bb"} we have \[|\mathfrak{M}(\vec{\bm{y}},\vec{\bm{z}},\vec{\bm{z}}')| \le \exp\left(-\frac{9}{2}\right) \ll 1, \quad \mbox{if} ~ |\bm{z}-\bm{z}'| \ge 3 \widetilde{R} ~ \mbox{or}~ |z_3-z_3'| \ge 3 \widetilde{R}_3,\] and we may work only with the diagonal part of the CINT kernel. We obtain the linear system of equations [\[eq:f11\]](#eq:f11){reference-type="eqref" reference="eq:f11"}, with vector \(\bm{u}\) of components \(|\rho(\vec{\bm{z}})|^2\) at the \(N_z\) mesh points in \(\mathfrak{D}\). The "data" vector \(\bm{d}\) consists of the samples of the CINT image at \(N_y < N_z\) equidistant points in \(\mathfrak{D}\), and in the narrowband regime the \(N_y \times N_z\) matrix \(\bm{\mathcal{M}}\) has entries \[m_{\vec{\bm{y}},\vec{\bm{z}}} = C\exp\left[-\frac{|\bm{z}-\bm{y}|^2}{2 R^2}-\frac{ (\overline{z}_3-y_3)^2}{2 R_3^2} \right], \qquad B \ll \Omega_d, \label{eq:R8}\] with constant \(C\). This depends only on \(\vec{\bm{y}}-\vec{\bm{z}}\), so we have a convolution as stated in section [2.3](#sect:OPT){reference-type="ref" reference="sect:OPT"}. In the broadband regime, the entries of \(\bm{\mathcal{M}}\) are \[m_{\vec{\bm{y}},\vec{\bm{z}}} = \frac{C}{\sqrt{1+\frac{|\bm{z}-\bm{y}|^2}{2 \theta^2 R^2}}} \exp\left[-\frac{|\bm{z}-\bm{y}|^2}{2 R^2}-\frac{ | z_3-y_3 + \frac{|\bm{z}-\bm{y}|^2}{2 L } + \frac{\bm{y} \cdot (\bm{z}-\bm{y})}{L}|^2}{2 R_3^2\big[1+\frac{|\bm{z}-\bm{y}|^2}{2 \theta^2 R^2}\big]} \right], \qquad B \gg \Omega_d, \label{eq:R8bb}\] with constant \(C\). Were it not for the last term in [\[eq:R8bb\]](#eq:R8bb){reference-type="eqref" reference="eq:R8bb"}, we would have a convolution. This term is large only at points \(\vec{\bm{y}} = (\bm{y},y_3)\) with \(\bm{y}\) near the boundary of the imaging region (\(|\bm{y}| \sim D < a\)), because by definition [\[eq:defTheta\]](#eq:defTheta){reference-type="eqref" reference="eq:defTheta"} and the assumption \(\theta = O(1)\) we get \[\left| \frac{\bm{y} \cdot (\bm{z}-\bm{y})/L}{R_3}\right| = O\left(\frac{|\bm{y}|}{L} \frac{R}{R_3}\right) = O\left(\frac{|\bm{y}|}{a \theta}\right) = O\left(\frac{|\bm{y}|}{a}\right).\] For points with \(|\bm{y}| \ll D < a\) the right hand side in [\[eq:R8bb\]](#eq:R8bb){reference-type="eqref" reference="eq:R8bb"} is approximately a function of \(\vec{\bm{y}}-\vec{\bm{z}}\), corresponding to a convolution model. # Resolution analysis {#sect:res} In this section we analyze the reconstruction of the vector \(\bm{u}\) of source intensities using the convex optimization formulation described in section [2.3](#sect:OPT){reference-type="ref" reference="sect:OPT"}. To simplify the analysis, we treat the approximation in [\[eq:f11\]](#eq:f11){reference-type="eqref" reference="eq:f11"} as an equality, and study the \(l_1\) optimization \[\min_{\bm{u} \in \mathbb{R}^{N_z}} \|\bm{u}\|_{_1} ~ ~ \mbox{such that} ~~ \bm{\mathcal{M}}\, \bm{u} = \bm{d}. \label{eq:Res1}\] This neglects additive noise and random fluctuations of the CINT function, which are small in our scaling. It also implies that the sources are on the reconstruction mesh, so that the equality constraint in [\[eq:Res1\]](#eq:Res1){reference-type="eqref" reference="eq:Res1"} holds for the true discretized source intensity. Naturally, in practice the sources may lie anywhere in \(\mathfrak{D}\), and noise and distortions due to the random medium play a role. This is why we use the more robust formulation [\[eq:f12\]](#eq:f12){reference-type="eqref" reference="eq:f12"} in the numerical simulations in section [6](#sect:num){reference-type="ref" reference="sect:num"}. We expect from the study of deconvolution using \(l_1\) optimization that the solution of [\[eq:Res1\]](#eq:Res1){reference-type="eqref" reference="eq:Res1"} should be a good approximation of the unknown vector of intensities if the sources are well separated. We show in this section that this is indeed the case. We also consider the case of clusters of nearby sources, and show that the \(l_1\) solution is useful when the clusters are well separated. The analysis is built on our recent results in. ## Definitions {#sect:T1Def} Let \(\bm{Y} = \left\{\vec{\bm{y}}_s, ~ ~ s = 1, \ldots, N_s\right\}\) be the set that supports the unknown, point-like sources in \(\mathfrak{D}\). We quantify the spatial separation between them using the following definition: For example, if the sources are all in the same cross-range plane, and the minimum distance between any two of them is \(H_{min}\), we may take \(H = H_{min}\) and \(H_3 = D_3\). We search the sources on a mesh with \(N_z\) points denoted generically by \(\vec{\bm{z}}\). The mesh discretizes \(\mathfrak{D}\), and we call it \(\mathfrak{D}_z\). For simplicity we let \(\bm{Y} \subset \mathfrak{D}_z\). To any \(\vec{\bm{z}} \in \mathfrak{D}_z\), we associate the column vector \(\bm{m}_{\vec{\bm{z}}} \in \mathbb{R}^{N_y}\) of the matrix \(\bm{\mathcal{M}}\). Its entries are given in [\[eq:R8\]](#eq:R8){reference-type="eqref" reference="eq:R8"} in the narrowband regime and by [\[eq:R8bb\]](#eq:R8bb){reference-type="eqref" reference="eq:R8bb"} in the broadband regime, for \(N_y < N_z\) points \(\vec{\bm{y}}\) at which we sample the CINT image. Note that [\[eq:Res2\]](#eq:Res2){reference-type="eqref" reference="eq:Res2"} is symmetric and non-negative, and attains its maximum at \(\vec{\bm{z}}' = \vec{\bm{z}}\), where \(\mathcal{I}_{\vec{\bm{z}},\vec{\bm{z}}} = 1\). We will show below that [\[eq:Res2\]](#eq:Res2){reference-type="eqref" reference="eq:Res2"} decreases as the points \(\vec{\bm{z}}\) and \(\vec{\bm{z}}'\) grow apart. This motivates the next definition which uses \(\mathcal{I}_{\vec{\bm{z}},\vec{\bm{z}}'}\) to measure the distance between \(\vec{\bm{z}}\) and \(\vec{\bm{z}}'\). We will show that \(\mathcal{I}_{\vec{\bm{z}},\vec{\bm{z}}'}\) decays as \(\|\vec{\bm{z}}-\vec{\bm{z}}'\|_2\) grows. Thus, we say that points \(\vec{\bm{z}}'\) outside \(\mathscr{B}_r(\vec{\bm{z}})\) have a weaker interaction with \(\vec{\bm{z}}\) than points inside \(\mathscr{B}_r(\vec{\bm{z}})\). Moreover, we may relate intuitively \(\Delta(\vec{\bm{z}},\vec{\bm{z}}')\) to the Euclidian distance \(\|\vec{\bm{z}}-\vec{\bm{z}}'\|_2\). In general more than one point may be closest to \(\vec{\bm{z}}\). If this is so, we let \(\mathscr{N}(\vec{\bm{z}})\) be any such point. ## Results The results stated here describe the relation between the reconstruction \(\bm{u}_\star\), the solution of the convex optimization problem [\[eq:Res1\]](#eq:Res1){reference-type="eqref" reference="eq:Res1"}, and the true unknown vector \(\bm{u}\) of source intensities. The next theorem shows that \(\bm{u}_\star\) is essentially supported in the set \(\bm{Y}\), when the points there are well separated. Note that the scales of separation between the sources are the resolution parameters \(R\) and \(R_3\) of CINT. The parameters \(\alpha\) and \(\alpha_3\) in the separation assumption may be any non-negative real numbers, but the statement of the theorem is useful only when the coefficient in front of \(\|\bm{u}_\star\|_1\) in [\[eq:Res6\]](#eq:Res6){reference-type="eqref" reference="eq:Res6"} is smaller than one. This happens for large enough \(\alpha\) and \(\alpha_3\). The larger the separation between the sources, the smaller the right hand side in [\[eq:Res6\]](#eq:Res6){reference-type="eqref" reference="eq:Res6"} is, and the better the concentration of the support of \(\bm{u}_\star\) near the points in \(\bm{Y}\). The next corollary gives an estimate of the error of the reconstruction. This result says that when the sources are far apart, the effective intensity vector \(\overline{\bm{u}}_\star\) is close to the true solution \(\bm{u}\). By definition, the support of \(\overline{\bm{u}}_\star\) is at the source points in \(\bm{Y}\). Its entries \(\overline{u}_\star(\vec{\bm{z}})\) at \(\vec{\bm{z}} \in \bm{Y}\) are weighted averages of the entries of \(\bm{u}_\star\) at points \(\vec{\bm{z}}' \in \mathscr{B}_r(\vec{\bm{z}})\), with weights \(\mathcal{I}_{\vec{\bm{z}},\vec{\bm{z}}'}\). When \(r\) is small, these weights are close to one, so \(\overline{u}_\star(\vec{\bm{z}})\) is approximately the sum of the entries of \(\bm{u}_\star\) supported in the ball \(\mathscr{B}_r(\vec{\bm{z}})\). Theorem [\[thm.1\]](#thm.1){reference-type="ref" reference="thm.1"} and its corollary are not useful when the sources are clustered together. The next result deals with this case, when the clusters are well separated. Equation [\[eq:Res9\]](#eq:Res9){reference-type="eqref" reference="eq:Res9"} says that we can cover the sources with disjoint balls of radius \(\epsilon\), centered at the points in \(\bm{Y}_\epsilon\). Thus, we call \(\bm{Y}_\epsilon\) the effective support of the sources, and \(\epsilon\) the radius of the clusters. The statement of the theorem is that when the clusters are well separated, and they have small radius, the \(l_1\) minimizer will be supported in their vicinity. As expected, [\[eq:Res10\]](#eq:Res10){reference-type="eqref" reference="eq:Res10"} converges to [\[eq:Res6\]](#eq:Res6){reference-type="eqref" reference="eq:Res6"} in the limit \(\epsilon \to 0\). ## Proofs {#sect:proofs} We use which state that for the decomposition of the \(l_1\) minimizer \(\bm{u}_\star\) as in Theorem [\[thm.1\]](#thm.1){reference-type="ref" reference="thm.1"}, and for the effective reconstructed source intensity vector \(\overline{\bm{u}}_\star\) defined in [\[eq:Res7\]](#eq:Res7){reference-type="eqref" reference="eq:Res7"}, we have \[\|\bm{u}_\star^{(o)}\|_1 \le \frac{2 \mathcal{I}(\bm{Y})}{r} \|\bm{u}_\star\|_1 ~ ~ \mbox{and} ~ ~ \frac{\|\bm{u}- \overline{\bm{u}}_\star\|_1}{\|\bm{u}\|_1} \le \frac{2 \mathcal{I}(\bm{Y})}{r}. \label{eq:Pf1}\] To determine the interaction coefficient \(\mathcal{I}(\bm{Y})\) of the sources, we estimate first the cross-correlations \(\mathcal{I}_{\vec{\bm{z}}, \vec{\bm{z}}'}\): The proof is in Appendix [\[ap:PfLem1\]](#ap:PfLem1){reference-type="ref" reference="ap:PfLem1"}. The next lemma, proved in sections [5.3.1](#sect:pfLemnb){reference-type="ref" reference="sect:pfLemnb"} and [5.3.2](#sect:pfLembb){reference-type="ref" reference="sect:pfLembb"}, gives the estimate of \(\mathcal{I}(\bm{Y})\), that combined with [\[eq:Pf1\]](#eq:Pf1){reference-type="eqref" reference="eq:Pf1"} proves Theorem [\[thm.1\]](#thm.1){reference-type="ref" reference="thm.1"} and Corollary [\[cor.1\]](#cor.1){reference-type="ref" reference="cor.1"}. To prove Theorem [\[thm.2\]](#thm.2){reference-type="ref" reference="thm.2"}, we use [@borcea2015resolution Theorem 4.4] which states that \[\|\bm{u}_\star^{(o)}\|_1 \le \frac{2 \mathcal{I}(\bm{Y}_\epsilon)}{r} \|\bm{u}_\star\|_1 + \frac{\|\bm{u}\|-\|\overline{\bm{u}}\|_1}{r}, \label{eq:PfT2.1}\] for \(\overline{\bm{u}}\) defined by \[\overline{\bm{u}}(\vec{\bm{z}}) = \left\{ \begin{array}{ll} \sum_{\vec{\bm{z}}' \in \mathscr{B}_\epsilon(\vec{\bm{z}})\cap \bm{Y}} u(\vec{\bm{z}}') \, \mathcal{I}_{\vec{\bm{z}},\vec{\bm{z}}'}, \quad &\mbox{if} ~ \vec{\bm{z}} \in \bm{Y}_\epsilon, \\ 0, &\mbox{otherwise}. \end{array} \right. \label{eq:PfT2.2}\] The interaction coefficient of the effective support \(\bm{Y}_\epsilon\) is as in Lemma [\[lem.2\]](#lem.2){reference-type="ref" reference="lem.2"}, so it remains to estimate the last term in [\[eq:PfT2.1\]](#eq:PfT2.1){reference-type="eqref" reference="eq:PfT2.1"}. Let us define the set \[S_{\vec{\bm{z}},\epsilon} = \left\{ \vec{\bm{z}}' \in \bm{Y} ~ ~ \mbox{s.t.} ~ ~ \vec{\bm{z}}' \in \mathscr{B}_\epsilon(\vec{\bm{z}})\right\},\] so that with definition [\[eq:PfT2.2\]](#eq:PfT2.2){reference-type="eqref" reference="eq:PfT2.2"} we can write \[\|\overline{\bm{u}}\|_1 = \sum_{\vec{\bm{z}} \in \bm{Y}_\epsilon} \left| \sum_{\vec{\bm{z}}' \in S_{\vec{\bm{z}},\epsilon}} u(\vec{\bm{z}}') \, \mathcal{I}_{\vec{\bm{z}},\vec{\bm{z}}'} \right| = \sum_{\vec{\bm{z}} \in \bm{Y}_\epsilon} \sum_{\vec{\bm{z}}' \in S_{\vec{\bm{z}},\epsilon}} |u(\vec{\bm{z}}')| \, \mathcal{I}_{\vec{\bm{z}},\vec{\bm{z}}'},\] where we used that by definition \(\mathcal{I}_{\vec{\bm{z}},\vec{\bm{z}}'} \ge 0\) and \(u(\vec{\bm{z}}') = |\rho(\vec{\bm{z}}')|^2 \ge 0\). Since the norm of the vector of the true source intensities is given by \[\|\bm{u}\|_1 = \sum_{\vec{\bm{z}} \in \bm{Y}} |u(\vec{\bm{z}})| = \sum_{\vec{\bm{z}} \in \bm{Y}_\epsilon} \sum_{\vec{\bm{z}}' \in S_{\vec{\bm{z}}\,\epsilon}} |u(\vec{\bm{z}}')|,\] we obtain that \[\|\bm{u}\|_1-\|\overline{\bm{u}}\|_1 = \sum_{\vec{\bm{z}} \in \bm{Y}_\epsilon} \sum_{\vec{\bm{z}}' \in S_{\vec{\bm{z}},\epsilon}} |u(\vec{\bm{z}}')| \left(1-\mathcal{I}_{\vec{\bm{z}},\vec{\bm{z}}'}\right) < \epsilon \sum_{\vec{\bm{z}} \in \bm{Y}_\epsilon} \sum_{\vec{\bm{z}}' \in S_{\vec{\bm{z}},\epsilon}} |u(\vec{\bm{z}}')| = \epsilon \|\bm{u}\|_1. \label{eq:PfT2.4}\] The inequality is because \[\Delta(\vec{\bm{z}},\vec{\bm{z}}') = 1-\mathcal{I}_{\vec{\bm{z}},\vec{\bm{z}}'} < \epsilon, \quad \forall \, \vec{\bm{z}}' \in \mathscr{B}_\epsilon(\vec{\bm{z}}).\] The statement of Theorem [\[thm.2\]](#thm.2){reference-type="ref" reference="thm.2"} follows by substitution of [\[eq:PfT2.4\]](#eq:PfT2.4){reference-type="eqref" reference="eq:PfT2.4"} in [\[eq:PfT2.1\]](#eq:PfT2.1){reference-type="eqref" reference="eq:PfT2.1"}, and using the estimate in Lemma [\[lem.2\]](#lem.2){reference-type="ref" reference="lem.2"}, with \(\bm{Y}\) replaced by \(\bm{Y}_\epsilon\). \(\Box\). ### Proof of Lemma [\[lem.2\]](#lem.2){reference-type="ref" reference="lem.2"} in the narrowband regime {#sect:pfLemnb} Recall Definition [\[def.4\]](#def.4){reference-type="ref" reference="def.4"} of \(\mathcal{I}(\bm{Y})\), and let \(\vec{\bm{z}} \in \mathfrak{D}_z\) be the maximizer of the sum in [\[eq:Res5\]](#eq:Res5){reference-type="eqref" reference="eq:Res5"}, so that \[\mathcal{I}(\bm{Y}) = \sum_{\vec{\bm{y}} \in \bm{Y}\setminus \mathscr{N}(\vec{\bm{z}})} \mathcal{I}_{\vec{\bm{z}},\vec{\bm{y}}}. \label{eq:Pf4}\] We denote the components of \(\vec{\bm{z}}\) by \(z_j\), for \(j=1,2,3\), and conclude from the source separation assumption in the lemma that the set \[\mathscr{S}_{\vec{\bm{z}}} = \left\{ \vec{\bm{z}}'= (z_1',z_2',z_3') \in \mathfrak{D} ~ ~ \mbox{s.t.} ~ ~ |z_j-z_j'| < H, ~j = 1, 2, ~ ~ |z_3-z_3'| < H_3\right\} \label{eq:defSz}\] contains at most one point in \(\bm{Y}\). This may be \(\mathscr{N}(\vec{\bm{z}})\), the closest point in \(\bm{Y}\) to \(\vec{\bm{z}}\) with respect to the semimetric \(\Delta\), satisfying \[\mathcal{I}_{\vec{\bm{z}},\mathscr{N}(\vec{\bm{z}})} \ge \mathcal{I}_{\vec{\bm{z}},\vec{\bm{y}}}. \qquad \forall \, \vec{\bm{y}} \in \bm{Y}. \label{eq:Pf5}\] Alternatively, \(\mathscr{S}_{\vec{\bm{z}}}\) may be empty or contain another point in \(\bm{Y}\). In either case, we obtain from equations [\[eq:Pf4\]](#eq:Pf4){reference-type="eqref" reference="eq:Pf4"} and [\[eq:Pf5\]](#eq:Pf5){reference-type="eqref" reference="eq:Pf5"} that \[\mathcal{I}(\bm{Y}) \le \sum_{\vec{\bm{y}} \in \bm{Y}\setminus \mathscr{S}_{\vec{\bm{z}}}} \mathcal{I}_{\vec{\bm{z}},\vec{\bm{y}}},\] and from the bound in Lemma [\[lem.1\]](#lem.1){reference-type="ref" reference="lem.1"}, \[\mathcal{I}(\bm{Y}) \le C \sum_{\vec{\bm{y}} \in \bm{Y}\setminus \mathscr{S}_{\vec{\bm{z}}}} \mathscr{E}_{\vec{\bm{z}},\vec{\bm{y}}}, ~ ~\mbox{for} ~ ~ \mathscr{E}_{\vec{\bm{z}},\vec{\bm{y}}} = \exp\left[-\frac{|\bm{z}-\bm{y}|^2}{8 R^2} -\frac{(z_3-y_3)^2}{R_3}\right]. \label{eq:Pf6}\] Using again the source separation assumption in the lemma, we conclude that for any \(\vec{\bm{y}} \in \bm{Y}\), we can define a set \(\mathscr{H}_{\vec{\bm{y}}}\), in the form of a rectangular prism of sides \(H/2\) in cross-range and \(H_3/2\) in range, satisfying \[\vec{\bm{y}} \in \mathscr{H}_{\vec{\bm{y}}} ~ ~ \mbox{and} ~ ~ \mathscr{H}_{\vec{\bm{y}}} \bigcap \mathscr{H}_{\vec{\bm{y}}'} = \emptyset, \qquad \forall \, \vec{\bm{y}} \ne \vec{\bm{y}}' \in \bm{Y}. \label{eq:Pf.7}\] There are many such sets, but we make our choice so that \(\vec{\bm{y}}\) is the furthermost point to \(\vec{\bm{z}}\) in \(\mathscr{H}_{\vec{\bm{y}}}\), satisfying \[\mathscr{E}_{\vec{\bm{z}},\vec{\bm{y}}} \le \mathscr{E}_{\vec{\bm{z}},\vec{\bm{z}}'}, \qquad \forall \, \vec{\bm{z}}' \in \mathscr{H}_{\vec{\bm{y}}}. \label{eq:Pf.8}\] This allows us to write \[\mathscr{E}_{\vec{\bm{z}},\vec{\bm{y}}} \le \frac{8}{H^2 H_3} \int_{\mathscr{H}_{\vec{\bm{y}}}} d \vec{\bm{z}}' \mathscr{E}_{\vec{\bm{z}},\vec{\bm{z}}'}, \qquad \forall \, \vec{\bm{y}} \in \bm{Y}, \label{eq:Pf.9}\] and obtain from [\[eq:Pf6\]](#eq:Pf6){reference-type="eqref" reference="eq:Pf6"} that \[\mathcal{I}(\bm{Y}) \le \frac{8 C^2}{H^2 H_3} \sum_{\bm{y} \in \bm{Y} \setminus \mathscr{S}_{\vec{\bm{z}}}} \int_{\mathscr{H}_{\vec{\bm{y}}}} d \vec{\bm{z}}' \mathscr{E}_{\vec{\bm{z}},\vec{\bm{z}}'} \le \frac{8 C^2}{H^2 H_3} \int_{\mathbb{R}^3 \setminus \mathscr{S}_{\vec{\bm{z}},\frac{1}{2}}} d \vec{\bm{z}}' \mathscr{E}_{\vec{\bm{z}},\vec{\bm{z}}'}, \label{eq:Pf10}\] with \(\mathscr{S}_{\vec{\bm{z}}, \frac{1}{2}}\) defined as in [\[eq:defSz\]](#eq:defSz){reference-type="eqref" reference="eq:defSz"}, with half the values of \(H\) and \(H_3\), \[\mathscr{S}_{\vec{\bm{z}},\frac{1}{2}} = \left\{ \vec{\bm{z}}'= (z_1',z_2',z_3') \in \mathfrak{D} ~ ~ \mbox{s.t.} ~ ~ |z_j-z_j'| < H/2, ~j = 1, 2, ~ ~ |z_3-z_3'| < H_3/2\right\}. \label{eq:defSzhalf}\] The last inequality in [\[eq:Pf10\]](#eq:Pf10){reference-type="eqref" reference="eq:Pf10"} is because the integrand is positive, the sets \(\mathscr{H}_{\vec{\bm{y}}}\) are disjoint, and \[\bigcup_{\vec{\bm{y}} \in \bm{Y}\setminus \mathscr{S}_{\vec{\bm{z}}}} \mathscr{H}_{\vec{\bm{y}}} \subset \mathbb{R}^3 \setminus \mathscr{S}_{\vec{\bm{z}},\frac{1}{2}}.\] We estimate the integral in [\[eq:Pf10\]](#eq:Pf10){reference-type="eqref" reference="eq:Pf10"} by decomposing the set \(\mathscr{S}_{\vec{\bm{z}},\frac{1}{2}}^c = \mathbb{R}^3\setminus \mathscr{S}_{\vec{\bm{z}},\frac{1}{2}}\) in three components denoted by \(\mathscr{C}_{\vec{\bm{z}},j}\), where \[\mathscr{C}_{\vec{\bm{z}},j} = \left\{\vec{\bm{z}}' \in \mathbb{R}^3 ~ ~ \mbox{s.t.} ~ ~ |z_j-z_j'| \ge H/2 \right\}, \quad j = 1,2,\] and \[\mathscr{C}_{\vec{\bm{z}},3} = \left\{\vec{\bm{z}}' \in \mathbb{R}^3 ~ ~ \mbox{s.t.} ~ ~ |z_j-z_j'| < H/2, ~ |z_3-z_3'| \ge H_3/2 \right\}.\] We have \[\begin{aligned} \frac{8}{H^2 H_3} \int_{\mathscr{C}_{\vec{\bm{z}},1}} d \vec{\bm{z}}' \, \mathscr{E}_{\vec{\bm{z}},\vec{\bm{z}}'} &= \frac{8}{H^2 H_3} \int_{|z_1'-z_1| \ge H/2} d z_1'\, e^{-\frac{(z_1'-z_1)^2}{4 R^2}} \int_{-\infty}^\infty d z_2'\, e^{-\frac{(z_2'-z_2)^2}{4 R^2}} \int_{-\infty}^\infty d z_3' \, e^{-\frac{(z_3'-z_3)^2}{4 R^2_3}} \nonumber \\ &=\frac{32\pi R R_3}{H^2 H_3} 2 \int_{H/2}^\infty dt \, e^{-\frac{t^2}{4 R^2}} \nonumber \\ &= \frac{64 \sqrt{\pi}}{\alpha^2 \alpha_3} \mbox{erfc} \left(\frac{\alpha}{4}\right), \label{eq:Pf11} \end{aligned}\] where we evaluated the integrals over \(z_3'\) and \(z_2'\) in the second line, and used [\[eq:asH\]](#eq:asH){reference-type="eqref" reference="eq:asH"} in the last line. The integral over \(\mathscr{C}_{\vec{\bm{z}},2}\) is the same, so it remains to estimate \[\begin{aligned} \frac{8}{H^2 H_3} \int_{\mathscr{C}_{\vec{\bm{z}},3}} d \vec{\bm{z}}' \, \mathscr{E}_{\vec{\bm{z}},\vec{\bm{z}}'} = \frac{8}{H^2 H_3} \int_{|z_1'-z_1| < H/2} d z_1'\, e^{-\frac{(z_1'-z_1)^2}{4 R^2}} \int_{|z_2'-z_2| < H/2} d z_2'\, e^{-\frac{(z_2'-z_2)^2}{4 R^2}}\nonumber \\ \times \int_{|z_3'-z_3| \ge H_3/2} d z_3' \, e^{-\frac{(z_3'-z_3)^2}{4 R^2_3}}. \label{eq:Pf12} \end{aligned}\] We bound the integrals over \(z_1'\) and \(z_2'\) by those of the real line, and rewrite the integral over \(z_3'\) in terms of the complementary error function, to obtain \[\begin{aligned} \frac{8}{H^2 H_3 } \int_{\mathscr{C}_{\vec{\bm{z}},3}} d \vec{\bm{z}}' \, \mathscr{E}_{\vec{\bm{z}},\vec{\bm{z}}'} &\le \frac{64 \sqrt{\pi}}{\alpha^2 \alpha_3} \mbox{erfc} \left(\frac{\alpha_3}{4}\right). \label{eq:Pf14} \end{aligned}\] The statement of Lemma [\[lem.2\]](#lem.2){reference-type="ref" reference="lem.2"} follows from [\[eq:Pf10\]](#eq:Pf10){reference-type="eqref" reference="eq:Pf10"}, with right hand side given by the sum of the integrals over \(\mathscr{C}_{\vec{\bm{z}},1}\) and \(\mathscr{C}_{\vec{\bm{z}},2}\) estimated in [\[eq:Pf11\]](#eq:Pf11){reference-type="eqref" reference="eq:Pf11"}, and over \(\mathscr{C}_{\vec{\bm{z}},3}\), estimated in [\[eq:Pf14\]](#eq:Pf14){reference-type="eqref" reference="eq:Pf14"}. \(\Box\) ### Proof of Lemma [\[lem.2\]](#lem.2){reference-type="ref" reference="lem.2"} in the broadband regime {#sect:pfLembb} We obtain from Lemma [\[lem.1\]](#lem.1){reference-type="ref" reference="lem.1"}, the same way as above, and with the same notation, that \[\mathcal{I}(\bm{Y}) \le C \sum_{\vec{\bm{y}} \in \bm{Y}\setminus \mathscr{S}_{\vec{\bm{z}}}} \mathscr{E}_{\vec{\bm{z}},\vec{\bm{y}}}, ~ ~\mbox{for} ~ ~ \mathscr{E}_{\vec{\bm{z}},\vec{\bm{y}}} = \exp\left[-\frac{|\bm{z}-\bm{y}|^2}{8 R^2} -\frac{\theta\left|z_3-y_3 + \frac{|\bm{z}|^2-|\bm{y}|^2}{2L}\right|}{R_3}\right]. \label{eq:Pf6bb}\] We also define as before, using the source separation assumption in the lemma, the set \(\mathscr{H}_{\vec{\bm{y}}}\), satisfying [\[eq:Pf.7\]](#eq:Pf.7){reference-type="eqref" reference="eq:Pf.7"} and [\[eq:Pf.8\]](#eq:Pf.8){reference-type="eqref" reference="eq:Pf.8"}. This leads us to the bound [\[eq:Pf10\]](#eq:Pf10){reference-type="eqref" reference="eq:Pf10"}, with the set \(\mathscr{S}_{\vec{\bm{z}},\frac{1}{2}}\) defined in [\[eq:defSzhalf\]](#eq:defSzhalf){reference-type="eqref" reference="eq:defSzhalf"}. We estimate the integral in [\[eq:Pf10\]](#eq:Pf10){reference-type="eqref" reference="eq:Pf10"} by decomposing the set \(\mathbb{R}^3\setminus \mathscr{S}_{\vec{\bm{z}},\frac{1}{2}}\) in three parts \(S_{\vec{\bm{z}},j}\), where \[S_{\vec{\bm{z}},j} = \left\{\vec{\bm{z}}' \in \mathbb{R}^3 ~ ~ \mbox{s.t.} ~ ~ |z_j-z_j'| \ge H/2 \right\}, \quad j = 1,2,\] and \[S_{\vec{\bm{z}},3} = \left\{\vec{\bm{z}}' \in \mathbb{R}^3 ~ ~ \mbox{s.t.} ~ ~ |z_j-z_j'| < H/2, ~ |z_3-z_3'| \ge H_3/2 \right\}.\] We have \[\begin{aligned} \frac{8}{H^2 H_3} \int_{S_{\vec{\bm{z}},1}} d \vec{\bm{z}}' \, \mathscr{E}_{\vec{\bm{z}},\vec{\bm{z}}'} &= \frac{8}{H^2 H_3} \int_{|z_1'-z_1| \ge H/2} d z_1'\, e^{-\frac{(z_1'-z_1)^2}{8 R^2}} \int_{-\infty}^\infty d z_2'\, e^{-\frac{(z_2'-z_2)^2}{8 R^2}} \int_{-\infty}^\infty d z_3' \, e^{-\frac{\theta}{ R_3} \left|z_3'-z_3 + \frac{|\bm{z}|^2-|\bm{z}'|^2}{2 L} \right|} \nonumber \\ &=\frac{128 \pi R^2 R_3}{H^2 H_3 \theta} \mbox{erfc}\left(\frac{H}{4 \sqrt{2} R}\right)\nonumber \\ &\le \frac{128\pi }{\alpha^2 \alpha_3 \theta}e^{-\left(\frac{h}{4 \sqrt{2} R}\right)^2} \nonumber \\ &\le \frac{16 \pi}{\alpha^3} e^{-\left(\frac{\alpha}{4 \sqrt{2}}\right)^2}, \label{eq:Pf11bb} \end{aligned}\] where we evaluated the integrals over \(z_3'\) and \(z_2'\) in the second line. The first inequality is because the complementary error function satisfies \(\mbox{erfc}(x) \le \exp(-x^2)\), and the second inequality is by the assumption on \(\alpha_3\). The integral over \(S_{\vec{\bm{z}},2}\) is the same, so it remains to estimate \[\begin{aligned} \frac{8}{H^2 H_3} \int_{S_{\vec{\bm{z}},3}} d \vec{\bm{z}}' \, \mathscr{E}_{\vec{\bm{z}},\vec{\bm{z}}'} \approx \frac{8}{H^2 H_3} \int_{|z_1'-z_1| < H/2} d z_1'\, e^{-\frac{(z_1'-z_1)^2}{8 R^2}} \int_{|z_2'-z_2| < H/2} d z_2'\, e^{-\frac{(z_2'-z_2)^2}{8 R^2}}\nonumber \\ \times \int_{|z_3'-z_3| \ge H_3/2} d z_3' \, e^{-\frac{\theta}{R_3} \left|z_3'-z_3 + \frac{|\bm{z}'|^2-|\bm{z}|^2}{2 L} \right|}. \label{eq:Pf12bb} \end{aligned}\] Because \(|z_j'-z_j| < H/2\) for \(j = 1,2\), and therefore \(|\bm{z}'-\bm{z}| < H/\sqrt{2}\), we have \[\begin{aligned} \theta \left|\frac{|\bm{z}'|^2-|\bm{z}|^2}{2 L R_3}\right| \le \frac{\theta|\bm{z}'-\bm{z}| |\bm{z}'+\bm{z}|}{2 L R_3} \le \frac{\theta H D}{2 L R_3} = \frac{\theta \alpha D R}{2 L R_3} = \frac{3 D \alpha }{a} \lesssim {3 \alpha}. \end{aligned}\] Here we used that \(|\bm{z}'+\bm{z}| \le D \sqrt{2}\) for all \(\bm{z},\bm{z}' \in [-D/2,D/2] \times [-D/2,D/2]\), and substituted \(H = R \alpha\) and definitions [\[eq:prop3.1\]](#eq:prop3.1){reference-type="eqref" reference="eq:prop3.1"} and [\[eq:defTheta\]](#eq:defTheta){reference-type="eqref" reference="eq:defTheta"}. The last inequality is by assumption [\[eq:as5\]](#eq:as5){reference-type="eqref" reference="eq:as5"}. With this bound we can estimate the integral over \(z_3'\) as follows \[\begin{aligned} \int_{|z_3'-z_3| \ge H_3/2} d z_3' \, e^{-\frac{\theta}{ R_3} \left|z_3'-z_3 + \frac{|\bm{z}'|^2-|\bm{z}|^2}{2 L} \right|} &= \frac{R_3}{\theta} \left[\int_{\frac{\theta H_3}{2R_3}}^\infty dt \, e^{-\left|-t + \frac{\theta(|\bm{z}'|^2-|\bm{z}|^2)}{2 L R_3}\right|} + \int_{\frac{\theta H_3}{2 R_3}}^\infty dt \, e^{-\left|t + \frac{\theta(|\bm{z}'|^2-|\bm{z}|^2)}{2 LR_3}\right|}\right] \nonumber \\ & \le \frac{2 R_3}{\theta} \int_{\frac{\theta H_3}{2 R_3}}^\infty dt \, e^{-(t-3 \alpha)} \nonumber \\ &= \frac{2 R_3}{\theta} e^{-\left(\frac{\theta\alpha_3}{2 }-3 \alpha \right)} \int_0^\infty dt e^{-t} \nonumber \\ &\le \frac{2 R_3}{\theta} e^{-\alpha}, \label{eq:Pf13bb} \end{aligned}\] where the last inequality is by the assumption \(H_3/R_3 = \alpha_3 \ge 8 \alpha/\theta\). Substituting in [\[eq:Pf12bb\]](#eq:Pf12bb){reference-type="eqref" reference="eq:Pf12bb"} we get \[\begin{aligned} \frac{8}{H^2 H_3 } \int_{S_{\vec{\bm{z}},3}} d \vec{\bm{z}}' \, \mathscr{E}_{\vec{\bm{z}},\vec{\bm{z}}'} &\le \frac{16 R_3}{H^2 H_3 \theta}e^{-\alpha} \int_{|z_1'-z_1| < H/2} d z_1'\, e^{-\frac{(z_1'-z_1)^2}{8 R^2}} \int_{|z_2'-z_2| < H/2} d z_2'\, e^{-\frac{(z_2'-z_2)^2}{8 R^2}} \nonumber \\ &\le \frac{128 \pi R^2 R_3}{H^2 H_3 \theta} e^{-\alpha} = \frac{128 \pi}{\alpha^2 \alpha_3 \theta} \le \frac{16 \pi}{\alpha^3} e^{-\alpha}. \label{eq:Pf14bb} \end{aligned}\] Here we bounded each Gaussian integral by \(2 \sqrt{2 \pi} R\), which is the integral over the real line, and used again the assumption on \(\alpha_3\). The statement of Lemma [\[lem.2\]](#lem.2){reference-type="ref" reference="lem.2"} follows from [\[eq:Pf10\]](#eq:Pf10){reference-type="eqref" reference="eq:Pf10"}, with right hand side given by the sum of the integrals over \(S_{\vec{\bm{z}},1}\) and \(S_{\vec{\bm{z}},2}\) estimated in [\[eq:Pf11bb\]](#eq:Pf11bb){reference-type="eqref" reference="eq:Pf11bb"}, and over \(S_{\vec{\bm{z}},3}\), estimated in [\[eq:Pf14bb\]](#eq:Pf14bb){reference-type="eqref" reference="eq:Pf14bb"}. \(\Box\) # Numerical simulations {#sect:num} We present here numerical simulations obtained with the wave propagation model described in equation [\[eq:a6\]](#eq:a6){reference-type="eqref" reference="eq:a6"}, with random travel time fluctuations computed by the line integrals in [\[eq:a7p\]](#eq:a7p){reference-type="eqref" reference="eq:a7p"}, in one realization of the random process \(\mu\). We generate it numerically using random Fourier series, for the Gaussian autocorrelation [\[eq:a2\]](#eq:a2){reference-type="eqref" reference="eq:a2"}. All the length scales are normalized by \(\ell\) in the simulations, and are chosen to satisfy marginally the assumptions in section [3.4](#sect:sc4){reference-type="ref" reference="sect:sc4"}. Specifically, we take \(\lambda_o = 1.75 \cdot 10^{-4} \,\ell\) and \(L = 800\, \ell,\) so that \[\sqrt{\lambda_o L} = 0.12 < \ell < 9.72 \, \ell = (\lambda_o L^3)^{1/4},\] and the aperture is \(a = 16 \ell\). This is slightly larger than the bound in [\[eq:as1\]](#eq:as1){reference-type="eqref" reference="eq:as1"}, but we also have the apodization [\[eq:R3\]](#eq:R3){reference-type="eqref" reference="eq:R3"}. We verify that \[\frac{\lambda_o}{\sqrt{\ell L}} = 6.17 \cdot 10^{-7} \ll \frac{\sqrt{\lambda_o \ell}}{L} = 5.22 \cdot 10^{-6} \ll \left(\frac{\ell}{L}\right)^{3/2} = 4.42 \cdot 10^{-5},\] and we take the strength of the fluctuations \(\sigma = 1.5 \cdot 10^{-6}\). With this choice we obtain \[\frac{\Omega_d}{\omega_o} = 0.083 \quad \mbox{and} \quad X_d = 0.14 \ell.\] We show results in two dimensions, for a linear array and a narrowband regime with bandwidth \(B = 0.0032 \omega_o\). Since in this regime we can only expect improvements in the cross-range localization of the sources, we focus attention at a given range, and display cross-range sections of the images. The migration and CINT images are calculated as in equations [\[eq:f5\]](#eq:f5){reference-type="eqref" reference="eq:f5"} and [\[eq:f6\]](#eq:f6){reference-type="eqref" reference="eq:f6"} from the data contaminated with \(5\%\) additive, uncorrelated, Gaussian noise. The thresholding parameters in the CINT image formation are \(\Omega = B/2\) and \(X = X_d/2\), and the sources are off the reconstruction mesh. The mesh size is \(H = \lambda_o L/(6X)\), unless stated otherwise The optimization formulation is \[\min_{\bm{u}} \|\bm{u}\|_1 \quad \mbox{such that } \quad \|\bm{\mathcal{M}}\, \bm{u}-\bm{d}\|_{2} \le \delta, \label{eq:OPTForm}\] with tolerance \(\delta = 0.05 \|\bm{d}\|_2\). The entries of matrix \(\bm{\mathcal{M}}\) are as defined in [\[eq:R8\]](#eq:R8){reference-type="eqref" reference="eq:R8"}, with constant \(C\) given in appendix [\[sect:apCINT\]](#sect:apCINT){reference-type="ref" reference="sect:apCINT"}. For comparison, we also present the results of a direct application of \(L_1\) optimization to the array data, without using the CINT image formation, as in. We call this method \"direct \(l_1\) optimization\" and refer to for details. We also refer to [\[eq:OPTForm\]](#eq:OPTForm){reference-type="eqref" reference="eq:OPTForm"} as \"\(l_1\) optimization\" and solve it with the software package. We begin with the results in Figure [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"}, for two sources that are at about \(2 \lambda_o L/X\) apart. The exact source locations are indicated on the abscissa in the plots, where the units are in \(\lambda_o L/X\). We display in the top row the CINT and migration images, and in the bottom row the solutions of the \(l_1\) optimization and the direct \(l_1\) optimization. Both migration and direct \(l_1\) optimization give spurious peaks, due to the random medium. To confirm this, we show in Figure [\[fig:2\]](#fig:2){reference-type="ref" reference="fig:2"} the results of the direct \(l_1\) optimization for the same sources in the homogeneous medium, where the reconstruction is very good. The CINT image shown in the top left plot is blurry, and it cannot distinguish the two sources. The \(l_1\) optimization improves the result, although there is a small shift in the estimate of the source locations. This shift changes from one realization to another, and it is due to the small random fluctuations of the CINT image. To illustrate the robustness of the methods to different realizations of the random medium, we display in Figure [\[fig:3\]](#fig:3){reference-type="ref" reference="fig:3"} the histograms of the number of peaks found by each method at a given cross-range location. We define filtered peaks as local maxima whose values are above \(33\%\) of the maximum of the image. The height of the histograms varies among the plots in Figure [\[fig:3\]](#fig:3){reference-type="ref" reference="fig:3"} because each method finds a different number of peaks. On the average, the migration images have \(9.5\) peaks, the direct \(l_1\) method finds \(9.65\) peaks, the CINT image has \(1.01\) peaks and the \(l_1\) optimization finds \(2.04\) peaks. While both migration and direct \(l_1\) find many spurious peaks, that are far from the source locations, the \(l_1\) optimization separates well the two sources and almost always peaks at their true locations. We display in Figure [\[fig:4\]](#fig:4){reference-type="ref" reference="fig:4"} the effect of the mesh size \(H\) on the \(l_1\) optimization results. Because we have \(5\) sources in this simulation, that are closer apart than \(\lambda_o L/X\), we do not expect a nearly exact reconstruction with the \(l_1\) optimization. Thus, we display in addition to the actual reconstructions the aggregated values recovered in the intervals of length \(r = \lambda_o L/(4X),\) centered at the true source locations. We observe that the results improve as we decrease the mesh size from \(H = \lambda_o L/X\) to \(\lambda_o L/(8X)\). This is due to the fact that the sources are off the grid, and the discretization error decreases as we reduce \(H\). The last illustration, in Figure [\[fig:5\]](#fig:5){reference-type="ref" reference="fig:5"}, shows the effect of the source separation on the quality of the reconstructions. As expected from the results in section [5](#sect:res){reference-type="ref" reference="sect:res"}, the reconstruction is better when the sources are further apart. # Summary {#sect:sum} We studied receiver array imaging of remote localized sources in random media, using convex optimization. The scattering regime is defined by precise scaling assumptions, and leads to large random wavefront distortions of the waves measured at the array. Conventional imaging methods like reverse time migration, also known as backprojection , or standard \(l_1\) optimization, cannot deal with such distortions and produce poor and unreliable results. We base our imaging on the coherent interferometric (CINT) method which mitigates random media effects like wavefront distortions at the expense of image resolution. The goal of the convex optimization is to remove the blur in the CINT images and thus improve the source localization. We show with a detailed analysis that under generic conditions the CINT imaging function is approximately a convolution of a blurring kernel with the discretized unknown source intensity on the imaging mesh. The kernel has a generic expression, that depends on the known CINT resolution limits obtained in for various wave propagation models. The optimization seeks to undo this convolution. The analysis and numerical simulations show that it gives very good estimates of the source locations when they are sufficiently far apart. This is in agreement with the results in. We also show that when the sources are clustered together, the estimates are not close to the true locations pointwise, but they are supported in their vicinity.
{'timestamp': '2016-07-18T02:09:47', 'yymm': '1607', 'arxiv_id': '1607.04625', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04625'}
null
null
# Introduction Recently, the LHCb collaboration performed the first full amplitude analysis of the decays \(B^+\to J/\psi \phi K^+\) with \(J/\psi\to\mu^+\mu^-\), \(\phi\to K^+K^-\) with a data sample of 3 fb\(^{-1}\) of \(pp\) collision data collected at \(\sqrt{s}=7\) and \(8\) TeV with the LHCb detector, confirmed the two old particles \(X(4140)\) and \(X(4274)\) in the \(J/\psi \phi\) mass spectrum with statistical significance \(8.4\sigma\) and \(6.0\sigma\), respectively, determined the quantum numbers to be \(J^{PC} =1^{++}\) with statistical significance \(5.7\sigma\) and \(5.8\sigma\), respectively. Moreover, the LHCb collaboration observed two new particles \(X(4500)\) and \(X(4700)\) in the \(J/\psi \phi\) mass spectrum with statistical significance \(6.1\sigma\) and \(5.6\sigma\), respectively, determined the quantum numbers to be \(J^{PC} =0^{++}\) with statistical significance \(4.0\sigma\) and \(4.5\sigma\), respectively. The measured masses and widths are There have been several possible assignments for the two new particles \(X(4500)\) and \(X(4700)\). In Ref., Chen et al study the newly observed \(X(4500)\) and \(X(4700)\) based on the diquark-antidiquark configuration within the framework of QCD sum rules, and interpret them as the D-wave \(cs\bar{c}\bar{s}\) tetraquark states with \(J^P =0^+\). In Ref., Liu studies the possible rescattering effects contribute to the process \(B^+\to J/\psi\phi K^+\), and observes that the \(D_{s}^{*+}D_{s}^-\) rescattering via the open-charmed meson loops and the \(\psi^\prime \phi\) rescattering via the \(\psi^\prime K_1\) loops may simulate the structures of the \(X(4140)\) and \(X(4700)\), respectively, and it is hard to attribute the \(X(4274)\) and \(X(4500)\) to the P-wave threshold rescattering effects. In Ref., Maiani, Polosa and Riquer assign the \(X(4500)\) and \(X(4700)\) to be the 2S tetraquark states based on the constituent diquark model. Also in Ref., Zhu assigns the \(X(4500)\) and \(X(4700)\) to be the 2S tetraquark states based on the constituent diquark model. In Ref., Lebed and Polosa propose that the \(X(3915)\) is the ground state \(cs\bar c \bar s\) state based on lacking of the observed \(D\bar D\) and \(D^*\bar{D}^*\) decays, and attribute the single known decay mode \(J/\psi \omega\) to the \(\omega-\phi\) mixing effect. The diquarks \(\varepsilon^{ijk} q^{T}_j C\Gamma q^{\prime}_k\) in color antitriplet have five structures in Dirac spinor space, where \(C\Gamma=C\gamma_5\), \(C\), \(C\gamma_\mu \gamma_5\), \(C\gamma_\mu\) and \(C\sigma_{\mu\nu}\) for the scalar, pseudoscalar, vector, axialvector and tensor diquarks, respectively. In Ref., we study the masses and pole residues of the \(X(3915)\), \(X(4500)\) and \(X(4700)\) in the scenario of tetraquark states with the QCD sum rules by calculating the contributions of the vacuum condensates up to dimension 10. The theoretical calculations support assigning the \(X(3915)\) and \(X(4500)\) to be the 1S and 2S \(C\gamma_\mu\otimes\gamma^\mu C\) type scalar \(cs\bar{c}\bar{s}\) tetraquark states, respectively, and assigning the \(X(4700)\) to be the 1S \(C\gamma_\mu\gamma_5\otimes\gamma_5\gamma^\mu C\) type scalar \(cs\bar{c}\bar{s}\) tetraquark state. In subsequent work, we take the \(X(4140)\) as the diquark-antidiquark type \(cs\bar{c}\bar{s}\) tetraquark state with \(J^{PC}=1^{++}\), and study the mass and pole residue with the QCD sum rules in details by constructing two types interpolating currents. The numerical results \(M_{X_{L,+}}=3.95\pm0.09\,\rm{GeV}\) and \(M_{X_{H,+}}=5.00\pm0.10\,\rm{GeV}\) disfavor assigning the \(X(4140)\) to be the \(C\gamma_5\otimes \gamma_\mu C+C\gamma_\mu\otimes \gamma_5 C\) type or \(C\otimes\gamma_5 \gamma_\mu C+C\gamma_\mu\gamma_5\otimes C\) type tetraquark state with \(J^{PC}=1^{++}\). The attractive interactions of one-gluon exchange favor formation of the diquarks in color antitriplet \(\overline{3}_{ c}\), flavor antitriplet \(\overline{3}_{ f}\) and spin singlet \(1_s\) or flavor sextet \(6_{ f}\) and spin triplet \(3_s\). The calculations based on the QCD sum rules also indicate that the favored configurations are the \(C\gamma_5\) and \(C\gamma_\mu\) diquark states, and the heavy-light \(C\gamma_5\) and \(C\gamma_\mu\) diquark states have almost degenerate masses . In Ref., we construct the \(C\gamma_5\otimes \gamma_5C\), \(C\gamma_\mu\otimes \gamma^\mu C\), \(C\gamma_\mu\gamma_5\otimes \gamma_5\gamma^\mu C\) type interpolating currents to study the scalar tetraquark states with the QCD sum rules in a systematic way, and observe that the \(C\gamma_5\otimes \gamma_5C\) type and \(C\gamma_\mu\otimes \gamma^\mu C\) type scalar \(cs\bar{c}\bar{s}\) tetraquark states have almost degenerate masses, about \(4.44\,\rm{GeV}\). The value \(4.44\,\rm{GeV}\) is not robust as the masses are extracted from the QCD spectral densities at the energy scale \(\mu=1\,\rm{GeV}\). In Refs., we explore the energy scale dependence of the masses \(M_{X/Y/Z}\) of the hidden charm (bottom) tetraquark states in details for the first time, and suggest a formula, \[\begin{aligned} \mu&=&\sqrt{M^2_{X/Y/Z}-(2{\mathbb{M}}_Q)^2} \,, \end{aligned}\] with the effective heavy quark mass \({\mathbb{M}}_Q\) to determine the energy scales of the QCD spectral densities in the QCD sum rules, which works well. Now we take a short digression to discuss the energy scale dependence of the QCD sum rules for the hidden charm or hidden bottom tetraquark states. The correlation functions \(\Pi(p)\) can be written as \[\begin{aligned} \Pi(p)&=&\int_{4m^2_Q(\mu)}^{s_0} ds \frac{\rho_{QCD}(s,\mu)}{s-p^2}+\int_{s_0}^\infty ds \frac{\rho_{QCD}(s,\mu)}{s-p^2} \,, \end{aligned}\] through dispersion relation at the QCD side, where the \(s_0\) are continuum threshold parameters. The \(\Pi(p)\) are energy scale independent, \[\begin{aligned} \frac{d}{d\mu}\Pi(p)&=&0\,, \end{aligned}\] which does not mean \[\begin{aligned} \frac{d}{d\mu}\int_{4m^2_Q(\mu)}^{s_0} ds \frac{\rho_{QCD}(s,\mu)}{s-p^2}\rightarrow 0 \,, \end{aligned}\] due to the following two reasons inherited from the QCD sum rules:\ \(\bullet\) Perturbative corrections are neglected, the higher dimensional vacuum condensates are factorized into lower dimensional ones therefore the energy scale dependence of the higher dimensional vacuum condensates is modified;\ \(\bullet\) Truncations \(s_0\) set in, the correlation between the threshold \(4m^2_Q(\mu)\) and continuum threshold \(s_0\) is unknown.\ In the QCD sum rules for the hidden charm or hidden bottom tetraquark states, the integrals \[\begin{aligned} \int_{4m^2_Q(\mu)}^{s_0} ds \,\rho_{QCD}(s,\mu)\,\exp\left(-\frac{s}{T^2} \right)\,, \end{aligned}\] are sensitive to the heavy quark masses \(m_Q(\mu)\) or the energy scales \(\mu\), where the \(T^2\) denotes the Borel parameters. Variations of the heavy quark masses or the energy scales \(\mu\) lead to changes of integral ranges \(4m_Q^2(\mu)-s_0\) of the variable \(ds\) besides the QCD spectral densities \(\rho_{QCD}(s,\mu)\), therefore changes of the Borel windows and predicted masses and pole residues. We cannot obtain energy scale independent QCD sum rules, but we have an energy scale formula to determine the energy scales consistently. According to the formula, the energy scale \(\mu=1\,\rm{GeV}\) is too low to result in robust predictions . In Ref., we use the energy scale formula in Eq.(2) to determine the energy scales of the QCD spectral densities in the QCD sum rules, and observe that the \(X(3915)\) and \(X(4500)\) can be assigned to be the 1S and 2S \(C\gamma_\mu\otimes\gamma^\mu C\) type scalar \(cs\bar{c}\bar{s}\) tetraquark states, respectively, and the \(X(4700)\) can be assigned to be the 1S \(C\gamma_\mu\gamma_5\otimes\gamma_5\gamma^\mu C\) type scalar \(cs\bar{c}\bar{s}\) tetraquark state. In Ref., we obtain the values \(M_{X(3915)}=3.91^{+0.21}_{-0.17}\,\rm{GeV}\), \(M_{X(4500)}=4.50^{+0.08}_{-0.09}\,\rm{GeV}\) and \(M_{X(4700)}=4.70^{+0.08}_{-0.09}\,\rm{GeV}\). The value \(M_{X(3915)}=3.91^{+0.21}_{-0.17}\,\rm{GeV}\) is much smaller than the value \(4.44\,\rm{GeV}\) obtained in Ref.. If the \(C\gamma_5\otimes \gamma_5C\) type and \(C\gamma_\mu\otimes \gamma^\mu C\) type tetraquark states have degenerate masses, then the \(X(3915)\) and \(X(4500)\) can have another diquark-antidiquark structure, \(C\gamma_5\otimes \gamma_5C\). In this article, we study the \(C\gamma_5\otimes \gamma_5C\) type and \(C\otimes C\) type scalar \(cs\bar{c}\bar{s}\) tetraquark states with the QCD sum rules by calculating the contributions of the vacuum condensates up to dimension 10 in a consistent way. In calculations, we use the energy scale formula to determine the optimal energy scales of the QCD spectral densities to extract to tetraquark masses to identify the \(X(3915)\), \(X(4500)\) and \(X(4700)\). The article is arranged as follows: we derive the QCD sum rules for the masses and pole residues of the 1S \(C\gamma_5\otimes \gamma_5C\) type and \(C\otimes C\) type \(cs\bar{c}\bar{s}\) tetraquark states in section 2; in section 3, we derive the QCD sum rules for the masses and pole residues of the 1S and 2S \(C\gamma_5\otimes \gamma_5C\) type \(cs\bar{c}\bar{s}\) tetraquark states; section 4 is reserved for our conclusion. # QCD sum rules for the scalar tetraquark states \(X_L\) and \(X_H\) In the following, we write down the two-point correlation functions \(\Pi_{L/H}(p)\) in the QCD sum rules, \[\begin{aligned} \Pi_{L/H}(p)&=&i\int d^4x e^{ip \cdot x} \langle0|T\left\{J_{L/H}(x)J_{L/H}^{\dagger}(0)\right\}|0\rangle \,, \end{aligned}\] where \[\begin{aligned} J_L(x)&=&\epsilon^{ijk}\epsilon^{imn}s^j(x)C\gamma_5 c^k(x) \bar{s}^m(x)\gamma_5 C \bar{c}^n(x) \,, \nonumber\\ J_H(x)&=&\epsilon^{ijk}\epsilon^{imn}s^j(x)C c^k(x) \bar{s}^m(x) C \bar{c}^n(x)\,, \end{aligned}\] where the \(i\), \(j\), \(k\), \(m\), \(n\) are color indexes, the \(C\) is the charge conjugation matrix. The scalar diquark states are more stable than the pseudoscalar diquark states, so we expect that the \(C\gamma_5\otimes \gamma_5C\) type tetraquark states have much smaller masses than the corresponding \(C\otimes C\) type tetraquark states, and add the subscripts \(L\) and \(H\) to denote the light and heavy tetraquark states, respectively. At the phenomenological side, we can insert a complete set of intermediate hadronic states with the same quantum numbers as the current operators \(J_{L/H}(x)\) into the correlation functions \(\Pi_{L/H}(p)\) to obtain the hadronic representation. After isolating the ground state contributions of the scalar \(cs\bar{c}\bar{s}\) tetraquark states \(X_{L/H}\), we get the results, \[\begin{aligned} \Pi_{L}(p)&=&\frac{\lambda_{ X_L}^2}{M_{X_L}^2-p^2} +\cdots \, \,, \nonumber\\ \Pi_{H}(p)&=&\frac{\lambda_{ X_H}^2}{M_{X_H}^2-p^2} +\cdots \, \,, \end{aligned}\] where the pole residues \(\lambda_{X_{L/H}}\) are defined by \(\langle 0|J_{L/H}(0)|X_{L/H}(p)\rangle = \lambda_{X_{L/H}}\). In the following, we briefly outline the operator product expansion for the correlation functions \(\Pi_{L/H}(p)\) in perturbative QCD. We contract the \(s\) and \(c\) quark fields in the correlation functions \(\Pi_{L/H}(p)\) with Wick theorem, and obtain the results: \[\begin{aligned} \Pi_{L}(p)&=&i\epsilon^{ijk}\epsilon^{imn}\epsilon^{i^{\prime}j^{\prime}k^{\prime}}\epsilon^{i^{\prime}m^{\prime}n^{\prime}}\int d^4x e^{ip \cdot x} \nonumber\\ &&{\rm Tr}\left[ \gamma_{5}C^{kk^{\prime}}(x)\gamma_{5} CS^{jj^{\prime}T}(x)C\right] {\rm Tr}\left[ \gamma_{5} C^{n^{\prime}n}(-x)\gamma_{5} C S^{m^{\prime}mT}(-x)C\right] \,, \nonumber\\ \Pi_{H}(p)&=&i\epsilon^{ijk}\epsilon^{imn}\epsilon^{i^{\prime}j^{\prime}k^{\prime}}\epsilon^{i^{\prime}m^{\prime}n^{\prime}}\int d^4x e^{ip \cdot x} \nonumber\\ &&{\rm Tr}\left[ C^{kk^{\prime}}(x) CS^{jj^{\prime}T}(x)C\right] {\rm Tr}\left[ C^{n^{\prime}n}(-x) C S^{m^{\prime}mT}(-x)C\right] \,, \end{aligned}\] where the \(S_{ij}(x)\) and \(C_{ij}(x)\) are the full \(s\) and \(c\) quark propagators, respectively, \[\begin{aligned} S^{ij}(x)&=& \frac{i\delta_{ij}\!\not\!{x}}{ 2\pi^2x^4}-\frac{\delta_{ij}m_s}{4\pi^2x^2}-\frac{\delta_{ij}\langle \bar{s}s\rangle}{12} +\frac{i\delta_{ij}\!\not\!{x}m_s \langle\bar{s}s\rangle}{48}-\frac{\delta_{ij}x^2\langle \bar{s}g_s\sigma Gs\rangle}{192}+\frac{i\delta_{ij}x^2\!\not\!{x} m_s\langle \bar{s}g_s\sigma Gs\rangle }{1152}\nonumber\\ &&-\frac{ig_s G^{a}_{\alpha\beta}t^a_{ij}(\!\not\!{x} \sigma^{\alpha\beta}+\sigma^{\alpha\beta} \!\not\!{x})}{32\pi^2x^2}-\frac{i\delta_{ij}x^2\!\not\!{x}g_s^2\langle \bar{s} s\rangle^2}{7776}-\frac{\delta_{ij}x^4\langle \bar{s}s \rangle\langle g_s^2 GG\rangle}{27648}-\frac{1}{8}\langle\bar{s}_j\sigma^{\mu\nu}s_i \rangle \sigma_{\mu\nu} \nonumber\\ && -\frac{1}{4}\langle\bar{s}_j\gamma^{\mu}s_i\rangle \gamma_{\mu }+\cdots \,, \end{aligned}\] \[\begin{aligned} C_{ij}(x)&=&\frac{i}{(2\pi)^4}\int d^4k e^{-ik \cdot x} \left\{ \frac{\delta_{ij}}{\!\not\!{k}-m_c}-\frac{g_sG^n_{\alpha\beta}t^n_{ij}}{4}\frac{\sigma^{\alpha\beta}(\!\not\!{k}+m_c)+(\!\not\!{k}+m_c) \sigma^{\alpha\beta}}{(k^2-m_c^2)^2}\right.\nonumber\\ &&\left. +\frac{g_s D_\alpha G^n_{\beta\lambda}t^n_{ij}(f^{\lambda\beta\alpha}+f^{\lambda\alpha\beta}) }{3(k^2-m_c^2)^4}-\frac{g_s^2 (t^at^b)_{ij} G^a_{\alpha\beta}G^b_{\mu\nu}(f^{\alpha\beta\mu\nu}+f^{\alpha\mu\beta\nu}+f^{\alpha\mu\nu\beta}) }{4(k^2-m_c^2)^5}+\cdots\right\} \,, \nonumber \\ \end{aligned}\] \[\begin{aligned} f^{\lambda\alpha\beta}&=&(\!\not\!{k}+m_c)\gamma^\lambda(\!\not\!{k}+m_c)\gamma^\alpha(\!\not\!{k}+m_c)\gamma^\beta(\!\not\!{k}+m_c)\, ,\nonumber\\ f^{\alpha\beta\mu\nu}&=&(\!\not\!{k}+m_c)\gamma^\alpha(\!\not\!{k}+m_c)\gamma^\beta(\!\not\!{k}+m_c)\gamma^\mu(\!\not\!{k}+m_c)\gamma^\nu(\!\not\!{k}+m_c)\,, \end{aligned}\] and \(t^n=\frac{\lambda^n}{2}\), the \(\lambda^n\) is the Gell-Mann matrix, \(D_\alpha=\partial_\alpha-ig_sG^n_\alpha t^n\). Then we compute the integrals both in coordinate space and in momentum space, and obtain the correlation functions \(\Pi_{L/H}(p)\), therefore the QCD spectral densities through dispersion relation. In this article, we take into account the vacuum condensates which are vacuum expectations of the operators of the orders \(\mathcal{O}( \alpha_s^{k})\) with \(k\leq 1\) consistently. For the technical details, one can consult Ref.. We neglect the radiative \(\mathcal{O}(\alpha_s)\) corrections for the perturbative contributions, it is a challenging or formidable work to calculate the radiative \(\mathcal{O}(\alpha_s)\) corrections in the QCD sum rules for hidden charm or hidden bottom tetraquark states, though the corrections may be large in the presence of two heavy quarks, just like in the QCD sum rules for the vector and axialvector \(B_c\) mesons. Once the analytical QCD spectral densities are obtained, we take the quark-hadron duality below the continuum thresholds \(s^0_{L/H}\) and perform Borel transform with respect to the variable \(P^2=-p^2\) to obtain the QCD sum rules: \[\begin{aligned} \lambda^2_{X_L}\, \exp\left(-\frac{M^2_{X_L}}{T^2}\right)= \int_{4m_c^2}^{s_L^0} ds\, \rho_L(s) \, \exp\left(-\frac{s}{T^2}\right) \,, \\ \lambda^2_{X_H}\, \exp\left(-\frac{M^2_{X_H}}{T^2}\right)= \int_{4m_c^2}^{s_H^0} ds\, \rho_H(s) \, \exp\left(-\frac{s}{T^2}\right) \,, \end{aligned}\] where \[\begin{aligned} \rho_L(s)&=&\rho_{0}(s)+\rho_{3}(s) +\rho_{4}(s)+\rho_{5}(s)+\rho_{6}(s)+\rho_{7}(s) +\rho_{8}(s)+\rho_{10}(s)\,, \nonumber\\ \rho_H(s)&=&\rho_L(s)|_{m_c\to-m_c} \,, \end{aligned}\] \[\begin{aligned} \rho_{0}(s)&=&\frac{1}{512\pi^6}\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \, yz\, (1-y-z)^3\left(s-\overline{m}_c^2\right)^2\left(7s^2-6s\overline{m}_c^2+\overline{m}_c^4 \right) \nonumber\\ &&+\frac{m_sm_c}{256\pi^6}\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \, (y+z)\, (1-y-z)^2\left(s-\overline{m}_c^2\right)^2\left(5s-2\overline{m}_c^2 \right)\,, \end{aligned}\] \[\begin{aligned} \rho_{3}(s)&=&-\frac{m_c\langle \bar{s}s\rangle}{16\pi^4}\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \, (y+z)(1-y-z)\left(s-\overline{m}_c^2\right)\left(2s-\overline{m}_c^2\right) \nonumber\\ &&+\frac{m_s\langle \bar{s}s\rangle}{16\pi^4}\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \, yz\, (1-y-z)\left(10s^2-12s\overline{m}_c^2+3\overline{m}_c^4 \right)\nonumber\\ &&-\frac{m_sm_c^2\langle \bar{s}s\rangle}{8\pi^4}\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \left( s-\overline{m}_c^2\right) \,, \end{aligned}\] \[\begin{aligned} \rho_{4}(s)&=&-\frac{m_c^2}{384\pi^4} \langle\frac{\alpha_s GG}{\pi}\rangle\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \left( \frac{z}{y^2}+\frac{y}{z^2}\right)(1-y-z)^3 \nonumber\\ &&\left\{ 2s-\overline{m}_c^2+\frac{s^2}{6}\delta\left(s-\overline{m}_c^2\right)\right\} \nonumber\\ &&+\frac{1}{512\pi^4} \langle\frac{\alpha_s GG}{\pi}\rangle\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \left( y+z\right)(1-y-z)^2 \left( 10s^2-12s\overline{m}_c^2+3\overline{m}_c^4\right) \nonumber\\ &&-\frac{m_sm_c^3}{384\pi^4} \langle\frac{\alpha_s GG}{\pi}\rangle\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \left( \frac{1}{y^3}+\frac{1}{z^3}\right)\left( y+z\right)(1-y-z)^2 \left\{1+\frac{s}{2}\delta\left(s-\overline{m}_c^2\right)\right\} \nonumber\\ &&+\frac{m_sm_c}{256\pi^4} \langle\frac{\alpha_s GG}{\pi}\rangle\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \left( \frac{z}{y^2}+\frac{y}{z^2}\right)(1-y-z)^2 \left(3s-2\overline{m}_c^2\right)\nonumber\\ &&+\frac{m_sm_c}{128\pi^4} \langle\frac{\alpha_s GG}{\pi}\rangle\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \,(1-y-z) \left(3s-2\overline{m}_c^2\right)\,, \end{aligned}\] \[\begin{aligned} \rho_{5}(s)&=&\frac{m_c\langle \bar{s}g_s\sigma Gs\rangle}{64\pi^4}\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \, (y+z) \left(3s-2\overline{m}_c^2 \right) \nonumber\\ &&-\frac{m_c\langle \bar{s}g_s\sigma Gs\rangle}{64\pi^4}\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \, \left( \frac{y}{z}+\frac{z}{y}\right) (1-y-z) \left(3s-2\overline{m}_c^2 \right) \nonumber\\ &&-\frac{m_s\langle \bar{s}g_s\sigma Gs\rangle}{16\pi^4}\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \,yz \left\{2s-\overline{m}_c^2 +\frac{s^2}{6}\delta\left(s-\overline{m}_c^2 \right)\right\} \nonumber\\ &&+\frac{m_sm_c^2\langle \bar{s}g_s\sigma Gs\rangle}{32\pi^4}\int_{y_i}^{y_f}dy \nonumber\\ &&-\frac{m_sm_c^2\langle \bar{s}g_s\sigma Gs\rangle}{64\pi^4}\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \, \left( \frac{1}{y}+\frac{1}{z}\right) \,, \end{aligned}\] \[\begin{aligned} \rho_{6}(s)&=&\frac{m_c^2\langle\bar{s}s\rangle^2}{12\pi^2}\int_{y_i}^{y_f}dy +\frac{g_s^2\langle\bar{s}s\rangle^2}{108\pi^4}\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz\, yz \left\{2s-\overline{m}_c^2 +\frac{s^2}{6}\delta\left(s-\overline{m}_c^2 \right)\right\}\nonumber\\ &&-\frac{g_s^2\langle\bar{s}s\rangle^2}{512\pi^4}\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \, (1-y-z)\left\{ 2\left(\frac{z}{y}+\frac{y}{z} \right)\left(3s-2\overline{m}_c^2 \right)+\left(\frac{z}{y^2}+\frac{y}{z^2} \right)\right.\nonumber\\ &&\left.m_c^2\left[ 2+ s\,\delta\left(s-\overline{m}_c^2 \right)\right] \right\} \nonumber\\ &&-\frac{g_s^2\langle\bar{s}s\rangle^2}{3888\pi^4}\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \, (1-y-z)\left\{ 3\left(\frac{z}{y}+\frac{y}{z} \right)\left(3s-2\overline{m}_c^2 \right)+\left(\frac{z}{y^2}+\frac{y}{z^2} \right)\right. \nonumber\\ &&\left.m_c^2\left[ 2+s\,\delta\left(s-\overline{m}_c^2\right)\right]+(y+z)\left[12\left(2s-\overline{m}_c^2\right) +2s^2\delta\left(s-\overline{m}_c^2\right)\right] \right\}\nonumber\\ &&-\frac{m_s m_c\langle\bar{s}s\rangle^2}{12\pi^2}\int_{y_i}^{y_f}dy \left\{1 +\frac{s}{2}\delta\left(s-\widetilde{m}_c^2 \right)\right\}\nonumber\\ && +\frac{m_s m_c g_s^2\langle\bar{s}s\rangle^2}{648\pi^4}\int_{y_i}^{y_f}dy \left\{1 +\frac{s}{2}\delta\left(s-\widetilde{m}_c^2 \right)\right\}\nonumber\\ &&-\frac{m_s m_c g_s^2\langle\bar{s}s\rangle^2}{216\pi^4}\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \left\{\frac{1}{y}+\frac{1}{z}+\left( \frac{1}{y^2}+\frac{1}{z^2}\right)\frac{m_c^2}{2}\delta\left(s-\overline{m}_c^2 \right)\right\}\nonumber\\ &&-\frac{m_s m_c g_s^2\langle\bar{s}s\rangle^2}{648\pi^4}\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \left( \frac{y}{z}+\frac{z}{y}\right) \left\{1+\frac{s}{2}\delta\left(s-\overline{m}_c^2 \right)\right\}\,, \end{aligned}\] \[\begin{aligned} \rho_7(s)&=&\frac{m_c^3\langle\bar{s}s\rangle}{288\pi^2 }\langle\frac{\alpha_sGG}{\pi}\rangle\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \left(\frac{1}{y^3}+\frac{1}{z^3} \right)(y+z)(1-y-z)\left(1+\frac{ s}{T^2}\right) \delta\left(s-\overline{m}_c^2\right)\nonumber\\ &&-\frac{m_c\langle\bar{s}s\rangle}{96\pi^2}\langle\frac{\alpha_sGG}{\pi}\rangle\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \left(\frac{y}{z^2}+\frac{z}{y^2}\right)(1-y-z) \left\{2+s\,\delta\left(s-\overline{m}_c^2\right) \right\}\nonumber\\ &&-\frac{m_c\langle\bar{s}s\rangle}{96\pi^2}\langle\frac{\alpha_sGG}{\pi}\rangle\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz\left\{2+ s\, \delta\left(s-\overline{m}_c^2\right) \right\} \nonumber\\ &&-\frac{m_c\langle\bar{s}s\rangle}{576\pi^2}\langle\frac{\alpha_sGG}{\pi}\rangle\int_{y_i}^{y_f}dy \left\{2+ s \, \delta \left(s-\widetilde{m}_c^2\right) \right\} \nonumber\\ &&-\frac{m_s m_c^2\langle\bar{s}s\rangle}{144\pi^2 }\langle\frac{\alpha_sGG}{\pi}\rangle\int_{0}^{1}dy \int_{0}^{1-y}dz \left(\frac{z}{y^2}+\frac{y}{z^2} \right)(1-y-z)\left(1+\frac{ s}{T^2}+\frac{ 2s^2}{T^4}\right) \delta\left(s-\overline{m}_c^2\right)\nonumber\\ &&+\frac{m_s m_c^4\langle\bar{s}s\rangle}{144\pi^2T^2 }\langle\frac{\alpha_sGG}{\pi}\rangle\int_{0}^{1}dy \int_{0}^{1-y}dz \left(\frac{1}{y^3}+\frac{1}{z^3} \right) \delta\left(s-\overline{m}_c^2\right)\nonumber\\ &&-\frac{m_s m_c^2\langle\bar{s}s\rangle}{48\pi^2 }\langle\frac{\alpha_sGG}{\pi}\rangle\int_{0}^{1}dy \int_{0}^{1-y}dz \left(\frac{1}{y^2}+\frac{1}{z^2} \right) \delta\left(s-\overline{m}_c^2\right)\nonumber\\ &&+\frac{m_s \langle\bar{s}s\rangle}{64\pi^2 }\langle\frac{\alpha_sGG}{\pi}\rangle\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \left(y+z \right) \left\{ 1+\left( \frac{2s}{3}+\frac{s^2}{6T^2}\right)\delta\left(s-\overline{m}_c^2\right)\right\}\nonumber\\ &&-\frac{m_s m_c^2\langle\bar{s}s\rangle}{288\pi^2 }\langle\frac{\alpha_sGG}{\pi}\rangle\int_{0}^{1}dy \left( 1+\frac{s}{T^2}\right)\delta\left(s-\widetilde{m}_c^2\right) \,, \end{aligned}\] \[\begin{aligned} \rho_8(s)&=&-\frac{m_c^2\langle\bar{s}s\rangle\langle\bar{s}g_s\sigma Gs\rangle}{24\pi^2}\int_0^1 dy \left(1+\frac{s}{T^2} \right)\delta\left(s-\widetilde{m}_c^2\right)\nonumber \\ &&+\frac{ \langle\bar{s}s\rangle\langle\bar{s}g_s\sigma Gs\rangle}{48\pi^2}\int_{0}^{1} dy \,s\,\delta\left(s-\widetilde{m}_c^2\right) \nonumber\\ &&+\frac{ 5m_s m_c\langle\bar{s}s\rangle\langle\bar{s}g_s\sigma Gs\rangle}{144\pi^2}\int_{0}^{1} dy\left(1+\frac{s}{T^2}+\frac{s^2}{2T^4} \right)\delta\left(s-\widetilde{m}_c^2\right) \nonumber\\ &&-\frac{ m_s m_c\langle\bar{s}s\rangle\langle\bar{s}g_s\sigma Gs\rangle}{96\pi^2}\int_{0}^{1} dy \left( \frac{1-y}{y}+\frac{y}{1-y}\right)\left(1+\frac{s}{T^2} \right)\delta\left(s-\widetilde{m}_c^2\right)\,, \end{aligned}\] \[\begin{aligned} \rho_{10}(s)&=&\frac{m_c^2\langle\bar{s}g_s\sigma Gs\rangle^2}{192\pi^2T^6}\int_0^1 dy \, s^2 \, \delta \left( s-\widetilde{m}_c^2\right) \nonumber \\ &&-\frac{m_c^4\langle\bar{s}s\rangle^2}{216T^4}\langle\frac{\alpha_sGG}{\pi}\rangle\int_0^1 dy \left\{ \frac{1}{y^3}+\frac{1}{(1-y)^3}\right\} \delta\left( s-\widetilde{m}_c^2\right)\nonumber\\ &&+\frac{m_c^2\langle\bar{s}s\rangle^2}{72T^2}\langle\frac{\alpha_sGG}{\pi}\rangle\int_0^1 dy \left\{ \frac{1}{y^2}+\frac{1}{(1-y)^2}\right\} \delta\left( s-\widetilde{m}_c^2\right)\nonumber\\ &&-\frac{\langle\bar{s}g_s\sigma Gs\rangle^2}{192 \pi^2T^4} \int_0^1 dy \,s^2 \, \delta\left( s-\widetilde{m}_c^2\right)\nonumber\\ &&+\frac{ \langle\bar{s}g_s\sigma Gs\rangle^2}{128 \pi^2T^2} \int_0^1 dy \,s\, \delta\left( s-\widetilde{m}_c^2\right)\nonumber \\ &&+\frac{m_c^2\langle\bar{s} s\rangle^2}{216 T^6}\langle\frac{\alpha_sGG}{\pi}\rangle\int_0^1 dy \, s^2 \, \delta \left( s-\widetilde{m}_c^2\right) \nonumber\\ &&-\frac{m_sm_c \langle\bar{s}g_s\sigma Gs\rangle^2}{576\pi^2T^8}\int_0^1 dy \, s^3 \, \delta \left( s-\widetilde{m}_c^2\right) \nonumber \\ &&-\frac{m_s m_c^3\langle\bar{s}s\rangle^2}{432T^4}\langle\frac{\alpha_sGG}{\pi}\rangle\int_0^1 dy \left\{ \frac{1}{y^3}+\frac{1}{(1-y)^3}\right\}\left(1-\frac{s}{T^2} \right) \delta\left( s-\widetilde{m}_c^2\right)\nonumber\\ &&-\frac{m_s m_c\langle\bar{s}s\rangle^2}{144T^4}\langle\frac{\alpha_sGG}{\pi}\rangle\int_0^1 dy \left\{ \frac{1-y}{y^2}+\frac{y}{(1-y)^2}\right\}\,s\, \delta\left( s-\widetilde{m}_c^2\right)\nonumber\\ &&+\frac{m_sm_c \langle\bar{s}g_s\sigma Gs\rangle^2}{576\pi^2T^6}\int_0^1 dy \left( \frac{1-y}{y}+\frac{y}{1-y}\right)\, s^2 \, \delta \left( s-\widetilde{m}_c^2\right) \nonumber \\ &&-\frac{m_s m_c\langle\bar{s}s\rangle^2}{864T^8}\langle\frac{\alpha_sGG}{\pi}\rangle\int_0^1 dy \,s^3\, \delta\left( s-\widetilde{m}_c^2\right) \,, \end{aligned}\] the subscripts \(0\), \(3\), \(4\), \(5\), \(6\), \(7\), \(8\), \(10\) denote the dimensions of the vacuum condensates, \(y_{f}=\frac{1+\sqrt{1-4m_c^2/s}}{2}\), \(y_{i}=\frac{1-\sqrt{1-4m_c^2/s}}{2}\), \(z_{i}=\frac{y m_c^2}{y s-m_c^2}\), \(\overline{m}_c^2=\frac{(y+z)m_c^2}{yz}\), \(\widetilde{m}_c^2=\frac{m_c^2}{y(1-y)}\), \(\int_{y_i}^{y_f}dy \to \int_{0}^{1}dy\), \(\int_{z_i}^{1-y}dz \to \int_{0}^{1-y}dz\), when the \(\delta\) functions \(\delta\left(s-\overline{m}_c^2\right)\) and \(\delta\left(s-\widetilde{m}_c^2\right)\) appear. We derive Eqs.(14-15) with respect to \(\frac{1}{T^2}\), then eliminate the pole residues \(\lambda_{X_{L/H}}\), and obtain the QCD sum rules for the masses \(M_{X_{L/H}}\) of the scalar tetraquark states, \[\begin{aligned} M^2_{X_{L/H}}&=&-\frac{\int_{4m_c^2}^{s_0} ds\,\frac{d}{d \left(1/T^2\right)}\,\rho_{L/H}(s)\,\exp\left(-\frac{s}{T^2}\right)}{\int_{4m_c^2}^{s_0} ds \, \rho_{L/H}(s)\, \exp\left(-\frac{s}{T^2}\right)}\,. \end{aligned}\] The vacuum condensates are taken to be the standard values \(\langle\bar{q}q \rangle=-(0.24\pm 0.01\, \rm{GeV})^3\), \(\langle\bar{s}s \rangle=(0.8\pm0.1)\langle\bar{q}q \rangle\), \(\langle\bar{s}g_s\sigma G s \rangle=m_0^2\langle \bar{s}s \rangle\), \(m_0^2=(0.8 \pm 0.1)\,\rm{GeV}^2\), \(\langle \frac{\alpha_s GG}{\pi}\rangle=(0.33\,\rm{GeV})^4\) at the energy scale \(\mu=1\, \rm{GeV}\). The quark condensates and mixed quark condensates evolve with the renormalization group equation, \(\langle\bar{s}s \rangle(\mu)=\langle\bar{s}s \rangle(Q)\left[\frac{\alpha_{s}(Q)}{\alpha_{s}(\mu)}\right]^{\frac{4}{9}}\), and \(\langle\bar{s}g_s \sigma Gs \rangle(\mu)=\langle\bar{s}g_s \sigma Gs \rangle(Q)\left[\frac{\alpha_{s}(Q)}{\alpha_{s}(\mu)}\right]^{\frac{2}{27}}\). We take the \(\overline{MS}\) masses \(m_{c}(m_c)=(1.275\pm0.025)\,\rm{GeV}\) and \(m_s(\mu=2\,\rm{GeV})=(0.095\pm0.005)\,\rm{GeV}\) from the Particle Data Group, and take into account the energy-scale dependence of the \(\overline{MS}\) masses from the renormalization group equation, \[\begin{aligned} m_c(\mu)&=&m_c(m_c)\left[\frac{\alpha_{s}(\mu)}{\alpha_{s}(m_c)}\right]^{\frac{12}{25}} \, ,\nonumber\\ m_s(\mu)&=&m_s({\rm 2GeV} )\left[\frac{\alpha_{s}(\mu)}{\alpha_{s}({\rm 2GeV})}\right]^{\frac{4}{9}} \, ,\nonumber\\ \alpha_s(\mu)&=&\frac{1}{b_0t}\left[1-\frac{b_1}{b_0^2}\frac{\log t}{t} +\frac{b_1^2(\log^2{t}-\log{t}-1)+b_0b_2}{b_0^4t^2}\right]\,, \end{aligned}\] where \(t=\log \frac{\mu^2}{\Lambda^2}\), \(b_0=\frac{33-2n_f}{12\pi}\), \(b_1=\frac{153-19n_f}{24\pi^2}\), \(b_2=\frac{2857-\frac{5033}{9}n_f+\frac{325}{27}n_f^2}{128\pi^3}\), \(\Lambda=213\,\rm{MeV}\), \(296\,\rm{MeV}\) and \(339\,\rm{MeV}\) for the flavors \(n_f=5\), \(4\) and \(3\), respectively . In the four-quark system \(q\bar{q}^{\prime}Q\bar{Q}\), the \(Q\)-quark serves as a static well potential and attracts the light quark \(q\) to form a heavy diquark \(\mathcal{D}\) in color antitriplet, while the \(\bar{Q}\)-quark serves as another static well potential and attracts the light antiquark \(\bar{q}^\prime\) to form a heavy antidiquark \(\mathcal{\bar{D}}\) in color triplet. Then the \(\mathcal{D}\) and \(\mathcal{\bar{D}}\) attract each other to form a compact tetraquark state, the two heavy quarks \(Q\) and \(\bar{Q}\) stabilize the tetraquark state. The tetraquark states \(\mathcal{D\bar{D}}\) are characterized by the effective heavy quark masses \({\mathbb{M}}_Q\) and the virtuality \(V=\sqrt{M^2_{X/Y/Z}-(2{\mathbb{M}}_Q)^2}\). It is natural to take the energy scale \(\mu=V\). We cannot obtain energy scale independent QCD sum rules, but we have an energy scale formula to determine the energy scales consistently. We fit the effective \(Q\)-quark mass \({\mathbb{M}}_{Q}\) to reproduce the experimental value \(M_{Z_c(3900)/Z_b(10610)}\), the empirical effective \(Q\)-quark mass \({\mathbb{M}}_Q\) is universal in the QCD sum rules for the hidden-charm or hidden bottom tetraquark states and embodies the net effects of the complex dynamics. We take the empirical energy scale formula and reproduce the experimental values of the masses of the \(X(3872)\), \(Z_c(3900)\), \(Z_c(4020)\), \(Z_c(4025)\), \(Z(4430)\), \(Y(4660)\), \(Z_b(10610)\) and \(Z_b(10650)\) in the scenario of tetraquark states . In this article, we take the updated value \({\mathbb{M}}_c=1.82\,\rm{GeV}\). We search for the Borel parameters \(T^2\) and continuum threshold parameters \(s^0_{L/H}\) according to the four criteria: \(\bf{1_\cdot}\) Pole dominance at the phenomenological side; \(\bf{2_\cdot}\) Convergence of the operator product expansion; \(\bf{3_\cdot}\) Appearance of the Borel platforms; \(\bf{4_\cdot}\) Satisfying the energy scale formula. Now we take a short digression to discuss how to choose the Borel parameters. At the phenomenological side of the QCD sum rules, we prefer smaller Borel parameters so as to depress the contributions of the higher excited states and continuum states and determine the upper bound of the Borel parameters \(T^2_{max}\). At the QCD side of the QCD sum rules, we prefer larger Borel parameters so as to warrant the convergence of the operator product expansion and determine the lower bound of the Borel parameters \(T^2_{min}\). In the QCD sum rules for the tetraquark states, the operator product expansion converges slowly, the \(T^2_{min}\) is postponed to large value, the Borel window \(T^2_{max}-T^2_{min}\) is rather small. However, the small Borel window does exist, the Borel parameter is just a free parameter, the predicted masses and pole residues should be independent on this parameter, in other words, there appears Borel platform. The resulting Borel parameters \(T^2\) and threshold parameters \(s^0_{L/H}\) are \[\begin{aligned} X_L &:& T^2 = (2.7-3.1) \mbox{ GeV}^2 \, ,\, s^0_{L} = (4.4\pm0.1 \mbox{ GeV})^2 \,, \\ X_H &:& T^2 = (5.2-5.6) \mbox{ GeV}^2 \, ,\, s^0_{H} = (6.0\pm0.1 \mbox{ GeV})^2 \,. \end{aligned}\] The pole contributions are \[\begin{aligned} X_L &:& {\rm pole} = (39-62) \% \, \,\, {\rm at}\, \, \, \mu = 1.40 \mbox{ GeV} \,, \\ X_H &:& {\rm pole} = (43-58) \% \, \,\, {\rm at}\, \, \, \mu = 4.10 \mbox{ GeV} \,, \end{aligned}\] the pole dominance condition is satisfied. The contributions come from the vacuum condensates of dimension \(i\) \(D_{i}\) are \[\begin{aligned} X_L &:& D_{0}=(37-39)\%\,, \,\, D_{3}=(65-67)\%\,, \,\, D_{4}=1\%\,, \,\,D_{5}=-(14-17)\%\,, \,\,D_{6}=(9-11)\%\, ,\nonumber\\ &&D_{7}=2\%\,, \,\, D_{8}=-(1-2)\%\,, \,\, D_{10} \ll 1 \% \, \,\, {\rm at}\, \,\, \mu = 1.40 \mbox{ GeV} \,, \\ X_H &:& D_{0}=(165-174)\%\,, \,\, D_{3}=-(69-78)\%\,, \,\, D_{4}=1\%\,, \,\,D_{5}=1\%\,, \,\,D_{6}=(2-3)\%\, ,\nonumber\\ &&D_{7}=-1\%\,, \,\, D_{8}\ll 1\%\,, \,\, D_{10} \ll 1 \% \, \,\, {\rm at}\, \,\, \mu = 4.10 \mbox{ GeV} \,, \end{aligned}\] where \(i=0,\,3,\,4,\,5,\,6,\,7,\,8,\,10\), the operator product expansion is well convergent. Now we take into account uncertainties of all the input parameters, and obtain the values of the masses and pole residues of the \(X_L\) and \(X_H\), \[\begin{aligned} M_{X_L}&=&3.89\pm 0.05\,\rm{GeV} \,, \,\,\, {\rm Experimental\,\, value} \,\,\,\,3918.4\pm 1.9\,\rm{ MeV} \, \, , \\ M_{X_H}&=&5.48\pm0.10\,\rm{GeV} \,, \,\,\, {\rm Experimental\,\, value} \,\,\,\,4704 \pm 10 ^{+14}_{-24} \mbox{ GeV}\, \, , \\ \lambda_{X_L}&=&(2.21 \pm 0.22)\times 10^{-2}\,\rm{GeV}^5 \,, \nonumber\\ \lambda_{X_H}&=&(1.98\pm0.08)\times 10^{-1}\,\rm{GeV}^5 \, , \end{aligned}\] which are also shown in Figs.1-2. In Figs.1-2, we plot the masses and pole residues with variations of the Borel parameters at much larger intervals than the Borel windows shown in Eqs.(27-28). From Eqs.(33-34) and Figs.1-2, we can see that the energy scale formula is satisfied and there appear platforms in the Borel windows, the uncertainties originate from the Borel parameters in the Borel windows are very small, \(\delta M_{X_L}/M_{X_L},\,\delta M_{X_H}/M_{X_H}\ll 1\%\), while the uncertainties originate from the Borel parameters outside of the Borel windows are rather large compared to the central values, where the criterion \(\bf{1}\) or the criterion \(\bf{2}\) is not satisfied. In the Borel windows, the four criteria are all satisfied, we expect to make reliable predictions. In calculations, we observe that the predicted masses \(M_{X_L}\) and \(M_{X_H}\) decrease monotonously with increase of the energy scales, the uncertainty \(\delta \mu=\pm0.1\,\rm{GeV}\) can lead to uncertainties \(\delta M_{X_L}=\pm20\,\rm{MeV}\) and \(\delta M_{X_H}=\pm3\,\rm{MeV}\). The present prediction \(M_{X_L}=3.89\pm 0.05\,\rm{GeV}\) is compatible with the experimental value \(M_{X(3915)}=3918.4\pm 1.9\,\rm{ MeV}\), so it is reasonable to assign the \(X_L\) to be the \(X(3915)\). The predicted mass \(M_{X_H}=5.48\pm0.10\,\rm{GeV}\) lies above the upper bound of the experimental value \(M_{X(4700)} = 4704 \pm 10 ^{+14}_{-24} \mbox{ GeV}\), it is impossible to assign the \(X(4700)\) to be the \(C-C\) type scalar tetraquark state. In Ref., we obtain the values \(M_{X(3915)}=3.91^{+0.21}_{-0.17}\,\rm{GeV}\), \(M_{X(4500)}=4.50^{+0.08}_{-0.09}\,\rm{GeV}\), \(M_{X(4700)}=4.70^{+0.08}_{-0.09}\,\rm{GeV}\), which are in excellent agreement with the experimental data, and support assigning the \(X(3915)\) and \(X(4500)\) to be the 1S and 2S \(C\gamma_\mu\otimes\gamma^\mu C\) type tetraquark states, and support assigning the \(X(4700)\) to be the 1S \(C\gamma_\mu\gamma_5\otimes\gamma_5\gamma^\mu C\) type tetraquark state. For the hidden-charm mesons, the energy gaps between the ground states and the first radial excited states are \(M_{\psi^\prime}-M_{J/\psi}=589\,\rm{MeV}\), \(M_{\eta_c^\prime}-M_{\eta_c}=656\,\rm{MeV}\), \(M_{Z(4430)}-M_{Z_c(3900)}=576\,\rm{MeV}\). In this article, we choose the continuum threshold parameters as \(\sqrt{s^0_{L}}-M_{X_L} = 0.4\sim 0.6 \mbox{ GeV}\) and \(\sqrt{s^0_{H}}-M_{X_H} = 0.4\sim 0.6 \mbox{ GeV}\), the contaminations of the radial excited states can be neglected. On the other hand, the currents \(J_{L/H}(x)\) couple potentially to the scattering states \(J/\psi\omega\), \(J/\psi \phi\), \(D_s^{*\pm} D_s^{*\mp}\), \(\cdots\), we can take into account the contributions of the intermediate meson-loops. All the renormalized self-energies contribute a finite imaginary part to modify the dispersion relation, we can take into account the finite width effects by the simple replacement of the hadronic spectral densities, \[\begin{aligned} \delta \left(s-M^2_{L/H} \right) &\to& \frac{1}{\pi}\frac{\sqrt{s}\,\Gamma_{L/H}(s)}{\left(s-M_{L/H}^2\right)^2+s\,\Gamma_{L/H}^2(s)}\,. \end{aligned}\] The widths \(\Gamma_{X(3915)} = 20 \pm 5 \mbox{ MeV}\) and \(\Gamma_{X(4700)} = 120 \pm 31 {}^{+42}_{-33} \mbox{ MeV}\) are not broad, the effects of the finite widths can be absorbed safely into the pole residues, the present predictions of the masses are reasonable. Now we can obtain the conclusion tentatively that the \(X(3915)\) maybe have both \(C\gamma_\mu\otimes\gamma^\mu C\) type and \(C\gamma_5\otimes\gamma_5 C\) type tetraquark components. In the following, we perform Fierz re-arrangement to the currents \(J_L\) and \(J_H\) both in the color and Dirac-spinor spaces to obtain the results, \[\begin{aligned} J_L&=&\frac{1}{4}\left\{\,-\bar{c} c\,\bar{s} s+\bar{c}i\gamma_5 c\,\bar{s}i\gamma_5 s-\bar{c} \gamma^\mu c\,\bar{s}\gamma_\mu s-\bar{c} \gamma^\mu\gamma_5 c\,\bar{s}\gamma_\mu\gamma_5 s+\frac{1}{2}\bar{c}\sigma_{\mu\nu} c\,\bar{s}\sigma^{\mu\nu} s\right. \nonumber\\ &&\left.+\bar{c} s\,\bar{s} c-\bar{c}i\gamma_5 s\,\bar{s}i\gamma_5 c+\bar{c} \gamma^\mu s\,\bar{s}\gamma_\mu c+\bar{c} \gamma^\mu\gamma_5 s\,\bar{s}\gamma_\mu\gamma_5 c-\frac{1}{2}\bar{c}\sigma_{\mu\nu} s\,\bar{s}\sigma^{\mu\nu} c \,\right\} \,, \end{aligned}\] \[\begin{aligned} J_H&=&\frac{1}{4}\left\{\,\bar{c} i\gamma_5c\,\bar{s} i\gamma_5 s-\bar{c}c\,\bar{s} s+\bar{c} \gamma^\mu \gamma_5c\,\bar{s}\gamma_\mu \gamma_5s +\bar{c} \gamma^\mu c\,\bar{s} \gamma_\mu s+\frac{1}{2}\bar{c}\sigma_{\mu\nu} \gamma_5c\,\bar{s}\sigma^{\mu\nu}\gamma_5 s\right. \nonumber\\ &&\left.+\bar{c} s\,\bar{s} c-\bar{c}i\gamma_5 s\,\bar{s}i \gamma_5c-\bar{c} \gamma^\mu s\,\bar{s}\gamma_\mu c-\bar{c} \gamma^\mu\gamma_5 s\,\bar{s}\gamma_\mu \gamma_5c-\frac{1}{2}\bar{c}\sigma_{\mu\nu} s\,\bar{s}\sigma^{\mu\nu} c \,\right\} \,, \end{aligned}\] the components couple to the meson pairs, for example, \(J/\psi \phi\), \(D_s\bar{D}_s\), \(D_s^*\bar{D}_s^*\). The two-body strong decays \[\begin{aligned} X_L &\to& J/\psi \phi \to J/\psi\omega \,\,\, (\phi-\omega\,\,{\rm mixing})\,, \nonumber\\ X_H&\to& J/\psi\phi \,\,, \, \, D_{s}\bar{D}_{s}\,\,, \, \, D^*_{s}\bar{D}^*_{s}\,\,, \, \, D_{s0}\bar{D}_{s0}\,\,, \, \, D_{s1}\bar{D}_{s1} \end{aligned}\] are Okubo-Zweig-Iizuka super-allowed. We can search for the \(X_L\) and \(X_H\) in those decays in the future. The diquark-antidiquark type tetraquark state can be taken as a special superposition of a series of meson-meson pairs, and embodies the net effects. The decays to its components, for example, \(J/\psi \phi\), are Okubo-Zweig-Iizuka super-allowed, but the re-arrangements in the color-space are non-trivial. # QCD sum rules for the \(X(3915)\) and \(X(4500)\) as the \(C\gamma_5 \otimes \gamma_5C\) type tetraquark states Now we tentatively assign the \(X(3915)\) and \(X(4500)\) to be the 1S and 2S \(C\gamma_5 \otimes \gamma_5C\) type tetraquark states, respectively, and study their masses and pole residues with the QCD sum rules. At the phenomenological side, we isolate the 1S and 2S scalar \(cs\bar{c}\bar{s}\) tetraquark states in the correlation function \(\Pi_{L}(p)\), and get the following result, \[\begin{aligned} \Pi_{L}(p)&=&\frac{\lambda_{ X_L{(\rm 1S)}}^2}{M_{X_L{(\rm 1S)}}^2-p^2} +\frac{\lambda_{ X_L{(\rm 2S)}}^2}{M_{X_L{(\rm 2S)}}^2-p^2} +\cdots \, \,, \end{aligned}\] where the pole residues \(\lambda_{X_L{(\rm 1S,2S)}}\) are defined by \(\langle 0|J_{L}(0)|X_L{(\rm 1S,2S)}(p)\rangle = \lambda_{X_L{(\rm 1S,2S)}}\). As the current \(J_{L}(x)\) couples potentially to the scattering states \(J/\psi\omega\), \(J/\psi \phi\), \(D_s^{*\pm} D_s^{*\mp}\), \(\cdots\), which contribute a finite imaginary part to modify the dispersion relation, we can take into account the finite width effects according to Eq.(36), and absorb the finite width effects into the pole residues. The contributions of the scattering states \(J/\psi\omega\), \(J/\psi \phi\), \(D_s^{*\pm} D_s^{*\mp}\), \(\cdots\) can be safely neglected if only the predicted masses \(M_{X_L{(\rm 1S,2S)}}\) are concerned. We take into account the contribution of the \(X(4500)\) and postpone the continuum threshold \(s_L^0\) to the large value \(s_{X(4500)}^0\) in the QCD sum rule in Eq.(14) to obtain the QCD sum rule, \[\begin{aligned} \lambda^2_{X_L{(\rm 1S)}}\, \exp\left(-\frac{M^2_{X_L{(\rm 1S)}}}{T^2}\right)+\lambda^2_{X_L{(\rm 2S)}}\, \exp\left(-\frac{M^2_{X_L{(\rm 2S)}}}{T^2}\right)= \int_{4m_c^2}^{s_{X(4500)}^0} ds\, \rho_L(s) \, \exp\left(-\frac{s}{T^2}\right) \,, \end{aligned}\] then we introduce the notations \(\tau=\frac{1}{T^2}\), \(D^n=\left(-\frac{d}{d\tau}\right)^n\), and use the subscripts \(1\) and \(2\) to denote the 1S state \(X(3915)\) and the 2S state \(X(4500)\) respectively for simplicity. The QCD sum rule can be rewritten as \[\begin{aligned} \lambda_1^2\exp\left(-\tau M_1^2 \right)+\lambda_2^2\exp\left(-\tau M_2^2 \right)&=&\Pi_{QCD}(\tau) \,, \end{aligned}\] here we add the subscript \(QCD\) to denote the QCD side of the correlation function \(\Pi_L(\tau)\). We derive the QCD sum rule in Eq.(42) with respect to \(\tau\) to obtain \[\begin{aligned} \lambda_1^2M_1^2\exp\left(-\tau M_1^2 \right)+\lambda_2^2M_2^2\exp\left(-\tau M_2^2 \right)&=&D\Pi_{QCD}(\tau) \,. \end{aligned}\] Then we have two equations, and obtain the QCD sum rules, \[\begin{aligned} \lambda_i^2\exp\left(-\tau M_i^2 \right)&=&\frac{\left(D-M_j^2\right)\Pi_{QCD}(\tau)}{M_i^2-M_j^2} \,, \end{aligned}\] where \(i \neq j\). Now we derive the QCD sum rules in Eq.(44) with respect to \(\tau\) to obtain \[\begin{aligned} M_i^2&=&\frac{\left(D^2-M_j^2D\right)\Pi_{QCD}(\tau)}{\left(D-M_j^2\right)\Pi_{QCD}(\tau)} \,, \nonumber\\ M_i^4&=&\frac{\left(D^3-M_j^2D^2\right)\Pi_{QCD}(\tau)}{\left(D-M_j^2\right)\Pi_{QCD}(\tau)}\,. \end{aligned}\] The squared masses \(M_i^2\) satisfy the following equation, \[\begin{aligned} M_i^4-b M_i^2+c&=&0\,, \end{aligned}\] where \[\begin{aligned} b&=&\frac{D^3\otimes D^0-D^2\otimes D}{D^2\otimes D^0-D\otimes D}\,, \nonumber\\ c&=&\frac{D^3\otimes D-D^2\otimes D^2}{D^2\otimes D^0-D\otimes D}\,, \nonumber\\ D^j \otimes D^k&=&D^j\Pi_{QCD}(\tau) \, D^k\Pi_{QCD}(\tau)\,, \end{aligned}\] \(i=1,2\), \(j,k=0,1,2,3\). We solve the equation in Eq.(46) and obtain the solutions \[\begin{aligned} M_1^2=\frac{b-\sqrt{b^2-4c} }{2} \,, \\ M_2^2=\frac{b+\sqrt{b^2-4c} }{2} \,. \end{aligned}\] In Ref., M. S. Maior de Sousa and R. Rodrigues da Silva study the masses and decay constants of the \(\rho({\rm 1S,2S})\), \(\psi({\rm 1S,2S})\), \(\Upsilon({\rm 1S,2S})\) using Eqs.(48-49), and observe that the ground state masses are (much) smaller than the experimental values. In Ref., we apply this approach to study the \(Z_c(3900)\) and \(Z(4430)\) as the 1S and 2S axialvector hidden-charm tetraquark states, respectively, and use the energy scale formula \(\mu=\sqrt{M^2_{X/Y/Z}-(2{\mathbb{M}}_c)^2}\) to overcome the shortcoming, and reproduce the experimental values of the masses \(M_{Z_c(3900)}\) and \(M_{Z(4430)}\). In Ref., we use the same approach to study the \(X(3915)\) and \(X(4500)\) as the 1S and 2S \(C\gamma_\mu\otimes\gamma^\mu C\) type scalar \(cs\bar{c}\bar{s}\) tetraquark states, respectively, and reproduce the experimental values of the masses \(M_{X(3915)}\) and \(M_{X(4500)}\). Now we tentatively assign the \(X(3915)\) and \(X(4500)\) to be the 1S and 2S \(C\gamma_5 \otimes \gamma_5C\) type tetraquark states, respectively, and resort to the same approach study the masses and pole residues with the QCD sum rules. Again, we search for the Borel parameter \(T^2\) and continuum threshold parameter \(s^0_{X(4500)}\) according to the four criteria: \(\bf{1_\cdot}\) Pole dominance at the phenomenological side; \(\bf{2_\cdot}\) Convergence of the operator product expansion; \(\bf{3_\cdot}\) Appearance of the Borel platforms; \(\bf{4_\cdot}\) Satisfying the energy scale formula.\ The resulting Borel parameter and threshold parameter are \[\begin{aligned} X(3890/4350) &:& T^2 = (1.8-2.2) \mbox{ GeV}^2 \, ,\, s^0_{X(4350)} = (4.9\pm0.1 \mbox{ GeV})^2 \,, \end{aligned}\] here we use the notations \(3890\) and \(4350\) in stead of \(3915\) and \(4500\) according to the masses extracted from the QCD sum rules. The pole contributions are \[\begin{aligned} X(3890)+X(4350) &:& {\rm pole} = (87-98) \% \, \,\, {\rm at}\, \,\, \mu = 1.25 \mbox{ GeV} \,, \\ X(3890)+X(4350) &:& {\rm pole} = (93-99) \% \, \,\, {\rm at}\, \,\, \mu = 2.40 \mbox{ GeV} \,, \end{aligned}\] the pole dominance condition is well satisfied. The contributions come from the vacuum condensates of dimension 10 \(D_{10}\) are \[\begin{aligned} X(3890)+X(4350) &:& D_{10} = (1-3) \% \, \,\, {\rm at}\, \,\, \mu = 1.25 \mbox{ GeV} \,, \\ X(3890)+X(4350) &:& D_{10} = (0-1) \% \, \,\, {\rm at}\, \,\, \mu = 2.40 \mbox{ GeV} \,, \end{aligned}\] the operator product expansion is well convergent. Now we take into account uncertainties of all the input parameters, and obtain the masses and pole residues of the \(X(3890)\) and \(X(4350)\), \[\begin{aligned} M_{X(3890)}&=&3.85^{+0.18}_{-0.17}\,\rm{GeV} \,, \,\,\, {\rm Experimental\,\, value} \,\,\,\,3918.4\pm 1.9\,\rm{ MeV} \, \, , \nonumber\\ M_{X(4350)}&=&4.83 \,\rm{GeV} \,, \nonumber\\ \lambda_{X(3890)}&=&1.84^{+0.76}_{-0.54}\times 10^{-2}\,\rm{GeV}^5 \,, \nonumber\\ \lambda_{X(4350)}&=&5.82\times 10^{-2}\,\rm{GeV}^5 \, , \end{aligned}\] at the energy scale \(\mu=1.25\,\rm{GeV}\), \[\begin{aligned} M_{X(3890)}&=&3.29\,\rm{GeV} \, , \nonumber\\ M_{X(4350)}&=&4.35^{+0.10}_{-0.11}\,\rm{GeV} \,, \,\,\, {\rm Experimental\,\, value} \,\,\,\,4506 \pm 11 ^{+12}_{-15}\, \rm {MeV}\, \, , \nonumber\\ \lambda_{X(3890)}&=&1.16\times 10^{-2}\,\rm{GeV}^5 \,, \nonumber\\ \lambda_{X(4350)}&=&6.01^{+0.93}_{-0.80}\times 10^{-2}\,\rm{GeV}^5 \, , \end{aligned}\] at the energy scale \(\mu=2.40\,\rm{GeV}\), where we have neglected the uncertainties out of control. Then we take the central values of the masses and pole residues as the input parameters, and obtain the pole contributions of the \(X(3890)\) and \(X(4350)\) respectively, \[\begin{aligned} {\rm pole}_{X(3890)}&=&(72-90)\% \,, \nonumber \\ {\rm pole}_{X(4350)}&=&(8-15)\% \,, \end{aligned}\] at the energy scale \(\mu=1.25\,\rm{GeV}\) and \[\begin{aligned} {\rm pole}_{X(3890)}&=&(55-76)\% \,, \nonumber \\ {\rm pole}_{X(4350)}&=&(23-38)\% \,, \end{aligned}\] at the energy scale \(\mu=2.40\,\rm{GeV}\). The pole contribution of the \(X(3890/4350)\) at \(\mu=1.25/2.40\,\rm{GeV}\) is larger than that at \(\mu=2.40/1.25\,\rm{GeV}\), we prefer to extract the mass and pole residue of the \(X(3890/4350)\) from the QCD spectral density at \(\mu=1.25/2.40\,\rm{GeV}\) and neglect the ones at \(\mu=2.40/1.25\,\rm{GeV}\). The neglected values of the masses do not satisfy the energy scale formula in Eq.(2). In this article, we choose the values \[\begin{aligned} M_{X(3890)}&=&3.85^{+0.18}_{-0.17}\,\rm{GeV} \,, \,\,\, {\rm Experimental\,\, value} \,\,\,\,3918.4\pm 1.9\,\rm{ MeV} \, \, , \nonumber\\ M_{X(4350)}&=&4.35^{+0.10}_{-0.11}\,\rm{GeV} \,, \,\,\, {\rm Experimental\,\, value} \,\,\,\,4506 \pm 11 ^{+12}_{-15}\, \rm {MeV}\, \, ,\\ \lambda_{X(3890)}&=&1.84^{+0.76}_{-0.54}\times 10^{-2}\,\rm{GeV}^5 \,, \nonumber\\ \lambda_{X(4350)}&=&6.01^{+0.93}_{-0.80}\times 10^{-2}\,\rm{GeV}^5 \, , \end{aligned}\] which are also shown in Figs.3-4. In Figs.3-4, we plot the masses and pole residues with variations of the Borel parameter \(T^2\) at much larger interval than the Borel window shown in Eq.(50). There appear platforms not very flat in the Borel window, which result in uncertainties \(\delta M_{X_L}/M_{X_L},\,\delta M_{X_H}/M_{X_H}=\pm 1\%\), while the uncertainties originate from the Borel parameter outside of the Borel window are rather large compared to the central values. The predicted masses \(M_{X(3890)}=3.85^{+0.18}_{-0.17}\,\rm{GeV}\) and \(M_{X(4350)}=4.35^{+0.10}_{-0.11}\,\rm{GeV}\) satisfy the energy scale formula. From Fig.3, we can see that the predicted mass \(M_{X(3890)}=3.85^{+0.18}_{-0.17}\,\rm{GeV}\) is compatible with the experimental value \(M_{X(3915)}=3918.4\pm 1.9\,\rm{ MeV}\), so it is reasonable to assign the \(X(3890)\) to be the \(X(3915)\). The predicted mass \(M_{X(4350)}=4.35^{+0.10}_{-0.11}\,\rm{GeV}\) lies below the lower bound of the experimental value \(M_{X(4500)} = 4506 \pm 11 ^{+12}_{-15}\, \rm {MeV}\), it is not favored to assign the \(X(4500)\) to be the 2S \(C\gamma_5\otimes \gamma_5C\) type scalar tetraquark state. Now we can draw the conclusion tentatively that the \(X(3915)\) maybe have both \(C\gamma_\mu\otimes\gamma^\mu C\) type and \(C\gamma_5\otimes\gamma_5 C\) type tetraquark components, while the \(X(4500)\) can be assigned to the 2S \(C\gamma_\mu\otimes\gamma^\mu C\) type tetraquark state. In the Borel or Laplace QCD sum rules, we introduce a new parameter \(T^2\) therefore an exponential factor \(\exp\left(-\frac{s}{T^2}\right)\) to suppress the experimentally unknown higher resonances. The Borel parameter \(T^2\) has no physical significance other than being a mathematical artifact, leads to rather narrow stability window for the hidden charm or hidden bottom tetraquark states. On the other hand, the continuum threshold \(s_0\), which has a clear physical interpretation, is often exponentially suppressed \(\exp\left(-\frac{s_0}{T^2}\right)\), or at best reduced in importance. While in the finite energy QCD sum rules or Hilbert moment QCD sum rules, the threshold \(s_0\) shows a power-like welcome feature, it is interesting to study the hidden charm or hidden bottom tetraquark states with the finite energy QCD sum rules or Hilbert moment QCD sum rules, this may be our next work. # Conclusion In this article, we study the \(C\gamma_5\otimes \gamma_5C\) type and \(C\otimes C\) type scalar \(cs\bar{c}\bar{s}\) tetraquark states with the QCD sum rules by calculating the contributions of the vacuum condensates up to dimension 10 in a consistent way. In calculations, we use the energy scale formula \(\mu=\sqrt{M^2_{X/Y/Z}-(2{\mathbb{M}}_c)^2}\) to determine the optimal energy scales of the QCD spectral densities. The ground state masses \(M_{C\gamma_5\otimes \gamma_5C}=3.89\pm 0.05\,\rm{GeV}\) and \(M_{C\otimes C}=5.48\pm0.10\,\rm{GeV}\) support assigning the \(X(3915)\) to be the 1S \(C\gamma_5\otimes \gamma_5C\) type tetraquark state with \(J^{PC}=0^{++}\), but do not support assigning the \(X(4700)\) to be the 1S \(C\otimes C\) type \(cs\bar{c}\bar{s}\) tetraquark state with \(J^{PC}=0^{++}\). Then we tentatively assign the \(X(3915)\) and \(X(4500)\) to be the \(\rm 1S\) and \(\rm 2S\) \(C\gamma_5\otimes \gamma_5C\) type scalar \(cs\bar{c}\bar{s}\) tetraquark states respectively, and obtain the \(\rm{1S}\) mass \(M_{\rm 1S}=3.85^{+0.18}_{-0.17}\,\rm{GeV}\) and \(\rm{2S}\) mass \(M_{\rm 2S}=4.35^{+0.10}_{-0.11}\,\rm{GeV}\) from the QCD sum rules, which support assigning the \(X(3915)\) to be the \(\rm{1S}\) \(C\gamma_5\otimes \gamma_5C\) type tetraquark state, but do not support assigning the \(X(4500)\) to be the \({\rm 2S}\) \(C\gamma_5\otimes \gamma_5C\) type tetraquark state. The \(X(3915)\) maybe have both \(C\gamma_\mu\otimes\gamma^\mu C\) type and \(C\gamma_5\otimes\gamma_5 C\) type tetraquark components, while the \(X(4500)\) can be assigned to the 2S \(C\gamma_\mu\otimes\gamma^\mu C\) type tetraquark state.
{'timestamp': '2016-12-15T02:06:12', 'yymm': '1607', 'arxiv_id': '1607.04840', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04840'}
# Introduction Let \({\mathcal{P}}\) be a set of \(v \ge 1\) elements, and let \({\mathcal{B}}\) be a set of \(k\)-subsets of \({\mathcal{P}}\), where \(k\) is a positive integer with \(1 \leq k \leq v\). Let \(t\) be a positive integer with \(t \leq k\). The pair \({\mathbb{D}} = ({\mathcal{P}}, {\mathcal{B}})\) is called a \(t\)-\((v, k, \lambda)\) *design*, or simply *\(t\)-design*, if every \(t\)-subset of \({\mathcal{P}}\) is contained in exactly \(\lambda\) elements of \({\mathcal{B}}\). The elements of \({\mathcal{P}}\) are called points, and those of \({\mathcal{B}}\) are referred to as blocks. We usually use \(b\) to denote the number of blocks in \({\mathcal{B}}\). A \(t\)-design is called *simple* if \({\mathcal{B}}\) does not contain repeated blocks. In this paper, we consider only simple \(t\)-designs. A \(t\)-design is called *symmetric* if \(v = b\). It is clear that \(t\)-designs with \(k = t\) or \(k = v\) always exist. Such \(t\)-designs are *trivial*. In this paper, we consider only \(t\)-designs with \(v > k > t\). A \(t\)-\((v,k,\lambda)\) design is referred to as a *Steiner system* if \(t \geq 2\) and \(\lambda=1\), and is denoted by \(S(t,k, v)\). A necessary condition for the existence of a \(t\)-\((v, k, \lambda)\) design is that \[\begin{aligned} \label{eqn-tdesignnecessty} \binom{k-i}{t-i} \mbox{ divides } \lambda \binom{v-i}{t-i} \end{aligned}\] for all integer \(i\) with \(0 \leq i \leq t\). There has been an interplay between codes and \(t\)-designs for decades. The incidence matrix of any \(t\)-design spans a linear code over any finite field \({\mathrm{GF}}(q)\). A lot of progress in this direction has been made and documented in the literature (see, for examples,, , ). On the other hand, both linear and nonlinear codes may hold \(t\)-designs. Some linear and nonlinear codes were employed to construct \(2\)-designs and \(3\)-designs. Binary and ternary Golay codes of certain parameters hold \(4\)-designs and \(5\)-designs. However, the largest \(t\) for which an infinite family of \(t\)-designs is derived directly from codes is \(t=3\). It looks that not much progress on the construction of \(t\)-designs from codes has been made so far, while many other constructions of \(t\)-designs are documented in the literature (). The main objective of this paper is to construct infinite families of \(2\)-designs and \(3\)-designs from linear codes. In addition, we determine the parameters of some known \(t\)-designs, and present many conjectured infinite families of \(2\)-designs that are based on projective ternary cyclic codes. # The classical construction of \(t\)-designs from codes and highly nonlinear functions Let \({\mathcal{C}}\) be a \([v, \kappa, d]\) linear code over \({\mathrm{GF}}(q)\). Let \(A_i:=A_i({\mathcal{C}})\), which denotes the number of codewords with Hamming weight \(i\) in \({\mathcal{C}}\), where \(0 \leq i \leq v\). The sequence \((A_0, A_1, \cdots, A_{v})\) is called the *weight distribution* of \({\mathcal{C}}\), and \(\sum_{i=0}^v A_iz^i\) is referred to as the *weight enumerator* of \({\mathcal{C}}\). For each \(k\) with \(A_k \neq 0\), let \({\mathcal{B}}_k\) denote the set of supports of all codewords of Hamming weight \(k\) in \({\mathcal{C}}\), where the coordinates of a codeword are indexed by \((0,1,2, \cdots, v-1)\). Let \({\mathcal{P}}=\{0, 1, 2, \cdots, v-1\}\). The pair \(({\mathcal{P}}, {\mathcal{B}}_k)\) may be a \(t\)-\((v, k, \lambda)\) design for some positive integer \(\lambda\). The following theorems, developed by Assumus and Mattson, show that the pair \(({\mathcal{P}}, {\mathcal{B}}_k)\) defined by a linear code is a \(t\)-design under certain conditions. The Assmus-Mattson Theorem for nonbinary codes is given as follows \[Assmus-Mattson Theorem, \] The Assmus-Mattson Theorems documented above are very powerful tools in constructing \(t\)-designs from linear codes. We will employ them heavily in this paper. It should be noted that the conditions in Theorems [\[thm-AM1\]](#thm-AM1){reference-type="ref" reference="thm-AM1"} and [\[thm-AM2\]](#thm-AM2){reference-type="ref" reference="thm-AM2"} are sufficient, but not necessary for obtaining \(t\)-designs. To construct \(t\)-designs via Theorems [\[thm-AM1\]](#thm-AM1){reference-type="ref" reference="thm-AM1"} and [\[thm-AM2\]](#thm-AM2){reference-type="ref" reference="thm-AM2"}, we will need the following lemma in subsequent sections, which is a variant of the MacWilliam Identity . A function \(f\) from \({\mathrm{GF}}(q^m)\) to itself is called *planar* or *perfect nonlinear* (PN) if \[\max_{0\ne a\in{\mathrm{GF}}(q^m)}\max_{b\in{\mathrm{GF}}(q^m)}|\{x\in{\mathrm{GF}}(q^m): f(x+a)-f(x)=b\}|=1,\] and *almost perfect nonlinear* (APN) if \[\max_{0\ne a\in{\mathrm{GF}}(q^m)}\max_{b\in{\mathrm{GF}}(q^m)}|\{x\in{\mathrm{GF}}(q^m): f(x+a)-f(x)=b\}|=2.\] Later in this paper, we will employ such functions in the constructions of linear codes and thus our constructions of \(t\)-designs. # Infinite families of \(3\)-designs from the binary RM codes {#sec-brmdesigns} It was known that Reed-Muller codes give families of \(3\)-\((2^m, k, \lambda)\) designs (, ). However, the parameters of \(k\) and \(\lambda\) may not be specifically given in the literature. The purpose of this section is to determine the parameters of some \(3\)-designs derived from binary Reed-Muller codes. We use \({\mathrm{RM}}(r, m)\) to denote the binary Reed-Muller code of length \(2^m\) and order \(r\). Note that \({\mathrm{RM}}(m-r, m)^\perp={\mathrm{RM}}(r-1, m)\), where \(2 \leq r < m\). The definition and information about binary Reed-Muller codes can be found in and. The following theorem gives parameters of all the \(3\)-designs in both \({\mathrm{RM}}(m-2, m)\) and \({\mathrm{RM}}(1, m)\). As a corollary of Theorem [\[thm-brmdesign1\]](#thm-brmdesign1){reference-type="ref" reference="thm-brmdesign1"}, we have the following, which is well known. The following theorem is also well known, and tells us that Reed-Muller codes give much more \(3\)-designs . Determining the weight distribution of \({\mathrm{RM}}(m-r, m)\) may be hard for \(3 \leq r \leq m-3\) in general. Therefore, it may be difficult to find out the parameters \((\kappa, \lambda)\) of all the \(3\)-designs. The following problem is open in general. Some progress on the open problem above was made by Kasami and Tokura and Kasami, Tokura and Azumi. Detailed information on this problem can be found in. # Designs from cyclic Hamming codes {#sec-hmdesigns} Let \(\alpha\) be a generator of \({\mathrm{GF}}(q^m)^*\). Set \(\beta=\alpha^{q-1}\). Let \(g(x)\) be the minimal polynomial of \(\beta\) over \({\mathrm{GF}}(q)\). Let \({\mathcal{C}}_{(q,m)}\) denote the cyclic code of length \(v=(q^m-1)/(q-1)\) over \({\mathrm{GF}}(q)\) with generator polynomial \(g(x)\). Then \({\mathcal{C}}_{(q,m)}\) has parameters \([(q^m-1)/(q-1), (q^m-1)/(q-1)-m, d]\), where \(d \in \{2,3\}\). When \(\gcd(q-1, m)=1\), \({\mathcal{C}}_{(q,m)}\) has minimum weight \(3\) and is equivalent to the Hamming code. A code of minimum distance \(d=2e+1\) is *perfect*, if the spheres of radius \(e\) around the codewords cover the whole space. The following theorem introduces a relation between perfect codes and \(t\)-designs and is due to Assmus and Mattson . It is known that the Hamming code over \({\mathrm{GF}}(q)\) is perfect, and the codewords of weight \(3\) hold a \(2\)-design by Theorem [\[thm-perfectcodedesign\]](#thm-perfectcodedesign){reference-type="ref" reference="thm-perfectcodedesign"}. The \(2\)-designs documented in the following theorem may be viewed as an extension of this result. Theorem [\[thm-HMdesign171\]](#thm-HMdesign171){reference-type="ref" reference="thm-HMdesign171"} tells us that for some \(k \geq 3\) with \(A_k \neq 0\), the supports of the codewords with weight \(k\) in \({\mathcal{C}}_{(q,m)}\) form \(2\)-\(((q^m-1)/(q-1), k, \lambda)\) design. However, it looks complicated to determine the parameter \(\lambda\) corresponding to this \(k \geq 4\). We draw the reader's attention to the following open problem. Notice that two binary codewords have the same support if and only if they are equal. When \(q=2\), Theorem [\[thm-HMdesign171\]](#thm-HMdesign171){reference-type="ref" reference="thm-HMdesign171"} becomes the following. Corollary [\[cor-HMdesign171\]](#cor-HMdesign171){reference-type="ref" reference="cor-HMdesign171"} says that each binary Hamming code \({\mathcal{C}}_{(2,m)}\) and its dual code give a total number \(2^m-4\) of \(2\)-designs with varying block sizes. The following are examples of the \(2\)-designs held in the binary Hamming code. # Designs from a class of binary codes with two zeros and their duals {#sec-newdesigns} In this section, we construct many infinite families of \(2\)-designs and \(3\)-designs with several classes of binary cyclic codes whose duals have two zeros. These binary codes are defined by almost perfect nonlinear (APN) functions over \({\mathrm{GF}}(2^m)\). To demonstrate the existence of the \(2\)-designs and \(3\)-designs presented in Theorems [\[thm-newdesigns1\]](#thm-newdesigns1){reference-type="ref" reference="thm-newdesigns1"} and [\[thm-newdesigns2\]](#thm-newdesigns2){reference-type="ref" reference="thm-newdesigns2"}, respectively, we describe a list of binary codes that have the weight distribution of Table [1](#tab-CG1){reference-type="ref" reference="tab-CG1"} below. Let \(\alpha\) be a generator of \({\mathrm{GF}}(2^m)^*\). Let \(g_s(x)=\mathbb{M}_1(x)\mathbb{M}_s(x)\), where \(\mathbb{M}_i(x)\) is the minimal polynomial of \(\alpha^i\) over \({\mathrm{GF}}(2)\). Let \({\mathcal{C}}_m\) denote the cyclic code of length \(v=2^m-1\) over \({\mathrm{GF}}(2)\) with generator polynomial \(g_s(x)\). It is known that \({\mathcal{C}}_m^\perp\) has dimension \(2m\) and the weight distribution of Table [1](#tab-CG1){reference-type="ref" reference="tab-CG1"} when \(m\) is odd and \(s\) takes on the following values: 1. \(s=2^h+1\), where \(\gcd(h, m)=1\) and \(h\) is a positive integer. 2. \(s=2^{2h}-2^h+1\), where \(h\) is a positive integer. 3. \(s=2^{(m-1)/2}+3\). 4. \(s=2^{(m-1)/2}+2^{(m-1)/4}-1\), where \(m \equiv 1 \pmod{4}\). 5. \(s=2^{(m-1)/2}+2^{(3m-1)/4}-1\), where \(m \equiv 3 \pmod{4}\). In all these cases, \({\mathcal{C}}_m\) has parameters \([2^m-1, 2^m-1-2m, 5]\) and is optimal. It is also known that the binary narrow-sense primitive BCH code with designed distance \(2^{m-1}-2^{(m-1)/2}\) has also the weight distribution of Table [1](#tab-CG1){reference-type="ref" reference="tab-CG1"}. These codes and their extended codes give \(2\)-designs and \(3\)-designs when they are plugged into Theorems [\[thm-newdesigns1\]](#thm-newdesigns1){reference-type="ref" reference="thm-newdesigns1"} and [\[thm-newdesigns2\]](#thm-newdesigns2){reference-type="ref" reference="thm-newdesigns2"}. It is known that \({\mathcal{C}}_m\) has parameters \([2^m-1, 2^m-1-2m, 5]\) if and only if \(x^e\) is an APN monomial over \({\mathrm{GF}}(2^m)\). However, even if \(x^e\) is APN, the dual code \({\mathcal{C}}_m^\perp\) may have many weights, and thus the code \({\mathcal{C}}_m\) and its dual \({\mathcal{C}}_m^\perp\) may not give \(2\)-designs. One of such examples is the inverse APN monomial. # Infinite families of \(2\)-designs from a type of ternary linear codes {#sec-june28} In this section, we will construct infinite families of \(2\)-designs with a type of primitive ternary cyclic codes. Theorem [\[thm-newdesigns228\]](#thm-newdesigns228){reference-type="ref" reference="thm-newdesigns228"} gives more \(2\)-designs. However, determining the corresponding value \(\lambda\) may be hard, as the number of blocks in the design may be difficult to derive from \(\overline{A}_k\) or \(\overline{A}_k^\perp\). To demonstrate the existence of the \(2\)-designs presented in Theorem [\[thm-newdesigns228\]](#thm-newdesigns228){reference-type="ref" reference="thm-newdesigns228"}, we present a list of ternary cyclic codes that have the weight distribution of Table [2](#tab-CG3){reference-type="ref" reference="tab-CG3"} below. Put \(n=3^m-1\). Let \(\alpha\) be a generator of \({\mathrm{GF}}(3^m)^*\). Let \(g_s(x)=\mathbb{M}_{n-1}(x)\mathbb{M}_{n-s}(x)\), where \(\mathbb{M}_i(x)\) is the minimal polynomial of \(\alpha^i\) over \({\mathrm{GF}}(3)\). Let \({\mathcal{C}}_m\) denote the cyclic code of length \(n=3^m-1\) over \({\mathrm{GF}}(3)\) with generator polynomial \(g_s(x)\). It is known that \({\mathcal{C}}_m^\perp\) has dimension \(2m\) and the weight distribution of Table [2](#tab-CG3){reference-type="ref" reference="tab-CG3"} when \(m\) is odd and \(s\) takes on the following values: 1. \(s=3^h+1\), \(h \geq 0\) is an integer. 2. \(s=(3^h + 1)/2\), where \(h\) is a positive integer and \(\gcd(m, h)=1\). In these two cases, \(x^s\) is a planar function on \({\mathrm{GF}}(3^m)\). Hence, these ternary codes are extremal in the sense that they are defined by planar functions whose differentiality is extremal. More classes of ternary codes such that their duals have the weight distribution of Table [2](#tab-CG3){reference-type="ref" reference="tab-CG3"} are documented in. They give also \(2\)-designs via Theorem [\[thm-newdesigns228\]](#thm-newdesigns228){reference-type="ref" reference="thm-newdesigns228"}. There are also ternary cyclic codes with three weights but different weight distributions in. They may also hold \(2\)-designs. # Conjectured infinite families of \(2\)-designs from projective cyclic codes {#sec-conjectureddesigns} Throughout this section, let \(m \geq 3\) be an odd integer, and let \(v=(3^m-1)/2\). The objective of this section is to present a number of conjectured infinite families of \(2\)-designs derived from linear projective ternary cyclic codes. [\[table4\]]{#table4 label="table4"} Below we present two examples of ternary linear codes \({\mathcal{C}}_m\) such that their duals \({\mathcal{C}}_m^\perp\) have the weight distribution of Table [\[Tab-GG2\]](#Tab-GG2){reference-type="ref" reference="Tab-GG2"}. There are a survey on Steiner systems \(S(2,4,v)\) and a book chapter on Steiner systems. It is known that a Steiner system \(S(2,4,v)\) exists if and only if \(v \equiv 1 \mbox{ or } 4 \pmod{12}\). If Conjecture [\[conj-j261\]](#conj-j261){reference-type="ref" reference="conj-j261"} is true, so is Conjecture [\[conj-j262\]](#conj-j262){reference-type="ref" reference="conj-j262"}. In this case, a coding theory construction of a Steiner system \(S(2, 4, (3^m-1)/2)\) for all odd \(m \ge 3\) is obtained. Even if some or all of the three conjectures are not true for ternary codes with the weight distribution of Table [\[Tab-GG2\]](#Tab-GG2){reference-type="ref" reference="Tab-GG2"}, these conjectures might still be valid for the two classes of ternary cyclic codes descried in Examples [\[exam-j271\]](#exam-j271){reference-type="ref" reference="exam-j271"} and [\[exam-j272\]](#exam-j272){reference-type="ref" reference="exam-j272"}. Note that Theorem [\[thm-AM2\]](#thm-AM2){reference-type="ref" reference="thm-AM2"} does not apply to the three conjectures above. We need to develop different methods for settling these conjectures. # Summary and concluding remarks In the last section of this paper, we mention some applications of \(t\)-designs and summarize the main contributions of this paper. ## Some applications of \(2\)-designs Let \({\mathcal{P}}\) be an Abelian group of order \(v\) under a binary operation denoted by \(+\). Let \({\mathcal{B}}=\{B_1, B_2, \cdots, B_b\}\), where all \(B_i\) are \(k\)-subsets of \({\mathcal{P}}\) and \(k\) is a positive integer. We define \(\Delta(B_i)\) to be the multiset \(\{ x-y: x \in B_i,\ y \in B_i\}\). If every nonzero element of \({\mathcal{P}}\) appears exactly \(\delta\) times in the multiset \(\bigcup_{i=1}^b \Delta(B_i)\), we call \({\mathcal{B}}\) a \((v, k, \delta)\) difference family in \(({\mathcal{P}}, +)\). The following theorems are straightforward and should be well known. Difference families have applications in the design and analysis of optical orthogonal codes, frequency hopping sequences, and other engineering areas. By Theorems [\[thm-june261\]](#thm-june261){reference-type="ref" reference="thm-june261"} and [\[thm-june262\]](#thm-june262){reference-type="ref" reference="thm-june262"}, \(t\)-designs with \(t \geq 2\) have also applications in these areas. In addition, \(2\)-designs give naturally linear codes. These show the importance of \(2\)-designs in applications. ## Summary It is well known that binary Reed-Muller codes hold \(3\)-designs. Hence, the only contribution of Section [3](#sec-brmdesigns){reference-type="ref" reference="sec-brmdesigns"} is the determination of the specific parameters of the \(3\)-designs held in \({\mathrm{RM}}(m-2, m)\) and its dual code, which are documented in Theorem [\[thm-brmdesign1\]](#thm-brmdesign1){reference-type="ref" reference="thm-brmdesign1"}. It has also been known for a long time that the codewords of weight \(3\) in the Hamming code hold a \(2\)-\(((q^m-1)/(q-1), 3, q-1)\) design. The contribution of Section [4](#sec-hmdesigns){reference-type="ref" reference="sec-hmdesigns"} is Theorem [\[thm-HMdesign171\]](#thm-HMdesign171){reference-type="ref" reference="thm-HMdesign171"}, which may be viewed as an extension of the known \(2\)-\(((q^m-1)/(q-1), 3, q-1)\) design held in the Hamming code, and also the parameters of the infinite families of \(2\)-designs derived from the binary Hamming codes, which are documented in Examples [\[exam-hmdesign4\]](#exam-hmdesign4){reference-type="ref" reference="exam-hmdesign4"}, [\[exam-hmdesign5\]](#exam-hmdesign5){reference-type="ref" reference="exam-hmdesign5"}, [\[exam-hmdesign6\]](#exam-hmdesign6){reference-type="ref" reference="exam-hmdesign6"}, and [\[exam-hmdesign7\]](#exam-hmdesign7){reference-type="ref" reference="exam-hmdesign7"}. A major contribution of this paper is presented in Section [5](#sec-newdesigns){reference-type="ref" reference="sec-newdesigns"}, where Theorems [\[thm-newdesigns1\]](#thm-newdesigns1){reference-type="ref" reference="thm-newdesigns1"} and [\[thm-newdesigns2\]](#thm-newdesigns2){reference-type="ref" reference="thm-newdesigns2"} document many infinite families of \(2\)-design and \(3\)-designs. The parameters of these \(2\)-designs and \(3\)-designs are given specifically. These designs are derived from binary cyclic codes that are defined by special almost perfect nonlinear functions. Another major contribution of this paper is documented in Section [6](#sec-june28){reference-type="ref" reference="sec-june28"}, where Theorem [\[thm-newdesigns228\]](#thm-newdesigns228){reference-type="ref" reference="thm-newdesigns228"} and its two corollaries describe several infinite families of \(2\)-designs. These \(2\)-designs are related to planar functions. It is noticed that the total number of \(3\)-designs presented in this paper (see Theorems [\[thm-brmdesign1\]](#thm-brmdesign1){reference-type="ref" reference="thm-brmdesign1"} and [\[thm-newdesigns2\]](#thm-newdesigns2){reference-type="ref" reference="thm-newdesigns2"}) are exponential. All of them are derived from linear codes. After comparing the list of infinite families of \(3\)-designs in with the \(3\)-designs presented in this paper, one may conclude that many, if not most, of the known infinite families of \(3\)-designs are from coding theory. Section [7](#sec-conjectureddesigns){reference-type="ref" reference="sec-conjectureddesigns"} presents many conjectured infinite families of \(2\)-designs. The reader is cordially invited to attack these conjectures and solve other open problems presented in this paper.
{'timestamp': '2016-07-19T02:05:49', 'yymm': '1607', 'arxiv_id': '1607.04813', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04813'}
# Introduction Several natural graph classes are known to have polynomially many minimal separators, w.r.t. the number \(n\) of vertices of the graph. It is the case for *chordal* graphs, which have at most \(n\) minimal separators , *weakly chordal*, *circular-arc* and *circle* graphs, which have \(\mathcal{O}(n^2)\) minimal separators . The property of having polynomially many minimal separators has been used in algorithms for decades, initially in an ad-hoc manner, i.e., algorithms were based on minimal separators but also other specific features of particular graph classes (see, e.g.,  ). Later, it was observed that minimal separators are sufficient for solving problems like [Treewidth]{.smallcaps} or [Minimum fill-in]{.smallcaps} . Both problems are related to *minimal triangulations*. Given an arbitrary graph \(G\), a minimal triangulation is a minimal chordal supergraph \(H\) of \(G\), on the same vertex set. Bouchitté and Todinca  introduced the notion of *potential maximal clique*, that is, a vertex set of \(G\) inducing a maximal clique in some minimal triangulation \(H\) of \(G\). Their algorithm for treewidth is based on dynamic programming over minimal separators and potential maximal clique s. The same authors proved that the number of potential maximal clique s is polynomially bounded in the number of minimal separators . Fomin and Villanger  found a more surprising application of minimal separators and potential maximal clique s, proving that they were sufficient for solving problems like [Maximum Independent Set]{.smallcaps}, [Maximum Induced Forest]{.smallcaps}, and more generally for finding a maximum induced subgraph \(G[F]\) of treewidth at most \(t\), where \(t\) is a constant. More formally, let \(\operatorname{poly}\) be some polynomial. We call \(\mathcal{G}_{\poly}\) the family of graphs such that \(G \in \mathcal{G}_{\operatorname{poly}}\) if and only if \(G\) has at most \(\operatorname{poly}(n)\) minimal separators. By , the problem of finding a maximum induced subgraph of treewidth at most \(t\) can be solved in polynomial time on \(\mathcal{G}_{\poly}\). The exponent of the polynomial depends on \(\operatorname{poly}\) and on \(t\). In , Fomin *et al.* further extend the technique to compute large induced subgraphs of bounded treewidth, and satisfying some CMSO property (expressible in counting monadic second-order logic). That allows to capture problems like [Longest induced path]{.smallcaps}. They also point out some limits of the approach. It is asked in  whether the techniques can be extended for solving the [Connected Vertex Cover]{.smallcaps} problem, which is equivalent to finding a maximum independent set \(F\) such that \(G-F\) is connected. More generally, their algorithm computes an induced subgraph \(G[F]\) of treewidth at most \(t\) satisfying some CMSO property, but is not able to ensure any property relating the induced subgraph to the initial graph. Here we make some progress in this direction. First, we consider the problem [Distance-\(d\) Independent Set]{.smallcaps} on \(\mathcal{G}_{\poly}\), where the goal is to find a maximum independent set \(F\) of the input graph \(G\), such that the vertices of \(F\) are at pairwise distance at least \(d\) in \(G\) (in the literature this problem is also known as \(d\)-Scattered-Set). This is equivalent to finding a maximum independent set in graph \(G^{d-1}\), the \((d-1)\)-th power of \(G\). Eto *et al.*  already studied the problem on chordal graphs, and proved that it is polynomial for every even \(d\), and NP-hard for any odd \(d\geq 3\) (it is even \(W[1]\)-hard when parameterized by the solution size). Their positive result is based on the observation that for any even \(d\), if \(G\) is chordal then so is \(G^{d-1}\). Eto *et al.*  ask if [Distance-\(d\) Independent Set]{.smallcaps} is polynomial on chordal bipartite graphs (which are *not* chordal but weakly chordal, see Section [2](#se:prelim){reference-type="ref" reference="se:prelim"}), a subclass of \(\mathcal{G}_{\poly}\). We bring a positive answer to their question for even values \(d\), by a result of combinatorial nature: for any graph \(G\) and any odd \(k\), the graph \(G^k\) has no more minimal separators than \(G\) (see Section [3](#se:pow){reference-type="ref" reference="se:pow"}). Consequently, [Distance-\(d\) Independent Set]{.smallcaps} is polynomial on \(\mathcal{G}_{\poly}\), for any even value \(d\) and any polynomial \(\operatorname{poly}\), and NP-hard for any odd \(d \geq 3\) and any \(\operatorname{poly}(n)\) asymptotically larger than \(n\). Such a dichotomy between odd and even values also appears when computing large \(d\)-clubs, that are induced subgraphs of diameter at most \(d\) , and for quite similar reasons. Second, we consider [Connected Vertex Cover]{.smallcaps}, [Connected Feedback Vertex Set]{.smallcaps} and more generally the problem of finding a maximum induced subgraph \(G[F]\) of treewidth at most \(t\), such that \(G-F\) is connected. We show (Section [4](#se:cvc){reference-type="ref" reference="se:cvc"}) that the problems are polynomially solvable for subclasses of \(\mathcal{G}_{\poly}\), like chordal and circular-arc graphs. This does not settle the complexity of these problems on \(\mathcal{G}_{\poly}\). As we shall discuss in Section [5](#se:ind){reference-type="ref" reference="se:ind"}, when restricted to bipartite graphs in \(\mathcal{G}_{\poly}\), [Connected Vertex Cover]{.smallcaps} can be reduced from [Red-Blue Dominating Set]{.smallcaps} (see ). It might be that this latter problem is NP-hard on bipartite graphs of \(\mathcal{G}_{\poly}\); that was our hope, since the very related problem [Independent Dominating Set]{.smallcaps} is NP-hard on chordal bipartite graphs , and on circle graphs . This question is still open, however we will observe that the [Red-Blue Dominating Set]{.smallcaps} is polynomial on the two natural classes of bipartite graphs with polynomially many minimal separators: chordal bipartite and circle bipartite graphs. # Preliminaries {#se:prelim} Let \(G = (V,E)\) be a graph. Let \(dist_G(u,v)\) denote the distance between vertices \(u\) and \(v\) (the minimum number of edges of a \(uv\)-path). We denote by \(N_G^k[v]\) the set of vertices at distance at most \(k\) from \(v\). Let also \(N_G^k(v) = N_G^k[v] \setminus \{v\}\), and we call these sets the *closed* and *open neighborhoods at distance* \(k\) of \(v\), respectively. Similarly, for a set of vertices \(U \subseteq V\), we call the sets \(N_G^k(U) = \cup_{u\in U} N_G^k(u) \backslash~ U\) and \(N_G^k[U] = \cup_{u\in U} N_G^k[u]\) the open and closed neighborhoods at distance \(k\) of \(U\), respectively. For \(k=1\), we simply denote by \(N_G(U)\), respectively \(N_G[U]\), the open and closed neighborhoods of \(U\); the subscript is omitted if clear from the context. A *clique* (resp. *independent set*) of \(G\) is a set of pairwise adjacent (resp. non-adjacent) vertices. A distance-\(d\) independent set is a set of vertices at pairwise distance at least \(d\). Equivalently, it is an independent set of the \((d-1)\)-th power \(G^{d-1}\) of \(G\). Graph \(G^k = (V,E^k)\) is obtained from \(G\) by adding an edge between every pair of vertices at distance at most \(k\). Given a vertex subset \(C\) of \(G\), we denote by \(G[C]\) the subgraph induced by \(C\). We say that \(C\) is a connected component of \(G\) if \(G[C]\) is connected and \(C\) is inclusion-maximal for this property. For \(S \subseteq V\), we simply denote \(G-S\) the graph \(G[V \setminus S]\). We say that \(S\) is a *\(a,b\)-minimal separator* of \(G\) if \(a\) and \(b\) are in distinct components \(C\) and \(D\) of \(G-S\), and \(N(C) = N(D) = S\). We also say that \(S\) is a minimal separator if it is an \(a,b\)-minimal separator for some pair of vertices \(a\) and \(b\). ## Graph classes A graph is *chordal* if it has no induced cycle with more than three vertices. A graph \(G\) is *weakly chordal* if \(G\) and its complement \(\overline{G}\) have no induced cycle with more than four vertices. The classes of *circle* and *circular-arc graphs* are defined by their intersection model. A graph \(G\) is a *circle graph* (resp. a *circular-arc graph*) if every vertex of the graph can be associated to a chord (resp. to an arc) of a circle such that two vertices are adjacent in \(G\) if and only if the corresponding chords (resp. arcs) intersect. We may assume w.l.o.g. that, in the intersection model, no two chords (resp. no two arcs) share an endpoint. On the circle, we add a *scanpoint* between each two consecutive endpoints of the set of chords (resp. arcs). A *scanline* is a line segment between two scanpoints. Given an intersection model of a circle (resp. circular-arc) graph \(G\), for any minimal separator \(S\) of \(G\) there is a scanline such that the vertices of \(S\) correspond exactly to the chords (resp. arcs) intersecting the scanline, see, e.g., . Chordal graphs have at most \(n\) minimal separators ; weakly chordal, circle and circular-arc graphs all have \(\mathcal{O}(n^2)\) minimal separators . ## Dynamic programming over minimal triangulations {#ss:FoVi10} Let \(G=(V,E)\) be an arbitrary graph. A chordal supergraph \(H = (V, E')\) (i.e., with \(E \subseteq E'\)), is called a *triangulation* of \(G\). If, moreover, \(E'\) is inclusion-minimal among all possible triangulations, we say that \(H\) is a *minimal triangulation* of \(G\). The *treewidth* of a chordal graph is its maximum clique size, minus one. Forests have treewidth \(1\), and graphs with no edges have treewidth \(0\). The treewidth \(\tw(G)\) of an arbitrary graph \(G\) is the minimum treewidth over all (minimal) triangulations \(H\) of \(G\). Cliques of minimal triangulations play a central role in treewidth. A *potential maximal clique* of \(G\) is a set of vertices that induces a maximal clique in some minimal triangulation \(H\) of \(G\). By , if \(\Omega\) is a potential maximal clique, then for every component \(C_i\) of \(G-\Omega\), its neighborhood \(S_i\) is a minimal separator. Moreover, the sets \(S_i\) are exactly the minimal separators of \(G\) contained in \(\Omega\). Minimal separators and potential maximal clique s have been used for computing treewidth and other parameters related to minimal triangulations, on \(\mathcal{G}_{\poly}\). Fomin and Villanger  extend the techniques to a family of problems: Clearly, [Maximum Independent Set]{.smallcaps} (which is equivalent to [Minimum Vertex Cover]{.smallcaps}) and [Maximum Induced Forest]{.smallcaps} (which is equivalent to [Minimum Feedback Vertex Set]{.smallcaps}) fit into this framework: they consist in finding maximum induced subgraphs \(G[F]\) of treewidth at most 0, respectively at most 1. The first ingredient of  is the following observation. Note that, for any clique \(\Omega\) of \(H_G\), we have that \(F \cap \Omega\) induces a clique in \(H_F\). In particular, if \(\tw(G[F]) \leq t\) and the clique size of \(H_F\) is at most \(t+1\), then every maximal clique of \(H_G\) intersects \(F\) in at most \(t+1\) vertices. The second ingredient is a dynamic programming scheme that we describe below. Let \(S\) be a minimal separator of \(G\), and \(C\) be a component of \(G-S\) such that \(N(C) = S\). The pair \((S,C)\) is called a *block*. Let \(\Omega\) be a potential maximal clique such that \(S \subset \Omega \subseteq S \cup C\). Then \((S,C,\Omega)\) is called a *good triple*. In the sequel, \(W\) denotes a set of at most \(t+1\) vertices. Observe that the two variants of compatibility differ by parameter \(\Omega\) and the last two conditions. We denote by \(\alpha(S,C,W)\) (resp. \(\beta(S,C,\Omega,W)\)) the size of a largest partial solution compatible with \((S,C,W)\) (resp. \((S,C,\Omega,W)\)). We now show how these quantities can be computed over all blocks and all good triples. The dynamic programming will proceed by increasing size over the blocks \((S,C)\), the size of the block being \(|S \cup C|\). It is based on the following equations (see  for details and proofs and Figure [\[fi:alphabeta\]](#fi:alphabeta){reference-type="ref" reference="fi:alphabeta"} for an illustration). #### Base case. It occurs for good triples \((S,C,\Omega)\) such that \(\Omega = S \cup C\). In this case, for each subset \(W\) of \(\Omega\) of size at most \(t+1\), \[\label{eq:base} \beta(S,C,\Omega,W) = |W|.\] #### Computing \(\alpha\) from \(\beta\). The following equation allows to compute the \(\alpha\) values from \(\beta\) values: \[\label{eq:afb} \alpha(S,C,W) = \max_{\Omega, W'} \beta(S,C,\Omega,W'),\] where the maximum is taken over all potential maximal clique s \(\Omega\) such that \((S,C,\Omega)\) is a good triple, and all subsets \(W'\) of \(\Omega\), of size at most \(t+1\), such that \(W = W' \cap S\). #### Computing \(\beta\) from \(\alpha\). Let \((S,C,\Omega)\) be a good triple, and fix an order \(C_1, C_2,\dots,C_p\) on the connected components of \(G[C \setminus \Omega]\). Let \(S_i = N_G(C_i)\), for all \(1 \leq i \leq p\). By , \((S_i,C_i)\) are also blocks of \(G\). A partial solution \(F\) compatible with \((S,C,\Omega,W)\) is obtained as a union of partial solutions \(F_i\) compatible with \((S_i,C_i,W \cap S_i)\), for each \(1 \leq i \leq p\), and the set \(W\). Denote by \(\gamma_i(S,C,\Omega,W)\) the size of the largest partial solution \(F\) compatible[^1] with \((S,C,\Omega,W)\), contained in \(\Omega \cup C_1 \cup \dots \cup C_i\) (hence \(F\) is not allowed to intersect the components \(C_{i+1}\) to \(C_p\)). We have the following equations. \[\label{eq:g1} \gamma_1(S,C,\Omega,W) = \alpha(S,C,\Omega,W \cap S_1) + |W|-|W \cap S_1|.\] For all \(i\), \(2 \leq i \leq p\), \[\label{eq:gi} \gamma_i(S,C,\Omega,W) = \gamma_{i-1}(S,C,\Omega,W) + \alpha(S,C,\Omega,W \cap S_i)-|W \cap S_i|.\] and finally \[\label{eq:bfg} \beta(S,C,\Omega,W) = \gamma_p(S,C,\Omega,W).\] For convenience we also consider that \(\emptyset\) is a minimal separator, and \((\emptyset,V)\) is a block. Then the size of the optimal global solution is simply \(\alpha(\emptyset,V,\emptyset)\). The algorithm can be adapted to output an optimal solution, not only its size. # Powers of graphs with polynomially many minimal separators {#se:pow} Let us prove that for any odd \(k\), graph \(G^k\) has no more minimal separators than \(G\). Recall that [Distance-\(d\) Independent set]{.smallcaps} on \(G\) is equivalent to [Maximum Independent Set]{.smallcaps} on \(G^{d-1}\). Since the latter problem is polynomial on \(\mathcal{G}_{\poly}\) by Proposition [\[pr:FoVi10\]](#pr:FoVi10){reference-type="ref" reference="pr:FoVi10"}, we deduce: We remind that for any odd value \(d\), problem [Distance-\(d\) Independent set]{.smallcaps} is NP-hard on chordal graphs , thus on \(\mathcal{G}_{\poly}\) for any polynomial \(\operatorname{poly}\) asymptotically larger than \(n\). The construction of  also shows that even powers of chordal graphs may contain exponentially many minimal separators. # On [Connected Vertex Cover]{.smallcaps} and [Connected Feedback Vertex Set]{.smallcaps} {#se:cvc} Let us consider the problem of finding a maximum induced subgraph \(G[F]\) such that \(\tw(G[F])\leq t\) and \(G-F\) is connected. One can easily observe that, for \(t=0\) (resp. \(t=1\)), this problem is equivalent to [Connected Vertex Cover]{.smallcaps} (resp. [Connected Feedback Vertex Set]{.smallcaps}), in the sense that if \(F\) is an optimal solution for the former, than \(V(G)-F\) is an optimal solution for the latter. Our goal is to enrich the dynamic programming scheme described in Subsection [2.2](#ss:FoVi10){reference-type="ref" reference="ss:FoVi10"} in order to ensure the connectivity of \(G-F\). One should think of this dynamic programming scheme of Subsection [2.2](#ss:FoVi10){reference-type="ref" reference="ss:FoVi10"} as similar to dynamic programming algorithms for bounded treewidth. The difference is that the bags (here, the potential maximal clique s) are not small but polynomially many, and we parse simultaneously through a set of decompositions. Nevertheless, we can borrow several classical ideas from treewidth-based algorithms. In general, for checking some property for the solution \(F\), we add a notion of *characteristics* of partial solutions. Then, for a characteristic \(c\), we update the Definition [\[de:partcomp\]](#de:partcomp){reference-type="ref" reference="de:partcomp"} in order to define partial solutions compatible with \((S,C,W,c)\) (resp. \((S,C,\Omega,W,c)\)), by requesting the partial solution to be compatible with characteristic \(c\). Parameter \(c\) will also appear in the updated version of Equations [\[eq:base\]](#eq:base){reference-type="ref" reference="eq:base"} to [\[eq:bfg\]](#eq:bfg){reference-type="ref" reference="eq:bfg"}. As usual in dynamic programming, the characteristics must satisfy several properties: (1) we must be able to compute the characteristic for the base case, (2) the characteristic of a partial solution \(F\) obtained from gluing smaller partial solutions \(F_i\) must only depend on the characteristics of \(F_i\), and (3) the characteristic of a global solution should indicate whether it is acceptable or not. Moreover, for a polynomial algorithm, we need the set of possible characteristics to be polynomially bounded. For checking connectivity conditions on \(G-F\), we define the characteristics of partial solutions in a natural way. Consider a block \((S,C)\) (resp. a good triple \((S,C,\Omega)\)) and a subset \(W\) of \(S\) (resp. of \(\Omega\)). Let \(F\) be a partial solution compatible with \((S,C,W)\) (resp. \((S,C,\Omega,W)\)), see Definition [\[de:partcomp\]](#de:partcomp){reference-type="ref" reference="de:partcomp"}. The *characteristic* \(c\) of \(F\) for \((S,C,W)\) (resp. for \((S,C,\Omega,W)\)) is defined as the partition induced on \(S \setminus W\) (resp. on \(\Omega \setminus W\)) by the connected components of \(G[S \cup C]-F\). More formally, let \(D_1, \dots, D_q\) denote the connected components of \(G[S \cup C]-F\), and let \(P_j = D_j \cap S\) (resp. \(P_j = D_j \cap \Omega\)), for all \(1 \leq j \leq q\). Then \(c = \{P_1,\dots, P_q\}\). We decide that if \(S \neq \emptyset\), partial solutions \(F\) having some component \(D_j\) that does not intersect \(S\) (resp. \(\Omega\)) are immediately rejected; indeed, for any extension \(F'\) of \(F\), the graph \(G-F'\) remains disconnected. Hence we may assume that all sets \(P_j\) are non-empty. We say that a partial solution \(F\) is *compatible with \((S,C,W,c)\)* (resp. *with \((S,C,\Omega,W,c)\)*) if it satisfies the conditions of Defintion [\[de:partcomp\]](#de:partcomp){reference-type="ref" reference="de:partcomp"}, and \(c\) is the characteristic of \(F\) for \((S,C,W)\) (resp. for \((S,C,\Omega,W)\)). We also define functions \(\alpha(S,C,W,c)\), \(\beta(S,C,\Omega,W,c)\) and \(\gamma_i(S,C,\Omega,W,c)\) like in Subsection [2.2](#ss:FoVi10){reference-type="ref" reference="ss:FoVi10"}, as the maximum size of partial solutions \(F\) compatible with the parameters. We can update Equations [\[eq:base\]](#eq:base){reference-type="ref" reference="eq:base"} to [\[eq:bfg\]](#eq:bfg){reference-type="ref" reference="eq:bfg"} as follows. #### Base case. For the good triples \((S,C,\Omega)\) such that \((S,C)\) is inclusion-minimal (hence \(\Omega = S \cup C\)), \[\label{eq:base2} \beta(S,C,\Omega,W,c) = |W|~\text{if \(c\) corresponds to the connected components of \(G[\Omega \setminus W]\).}\] Otherwise we set \(\beta(S,C,\Omega,W,c) =-\infty\). #### Computing \(\alpha\) from \(\beta\). \[\label{eq:afb2} \alpha(S,C,W,c) = \max_{\Omega, W',c'} \beta(S,C,\Omega,W',c'),\] where the maximum is taken over all potential maximal clique s \(\Omega\) such that \((S,C,\Omega)\) is a good triple, and all subsets \(W'\) of \(\Omega\), of size at most \(t+1\), such that \(W = W' \cap S\), and all characteristics \(c'\) such that each part of \(c\) corresponds to the intersection between \(S\) and a part of \(c'\). If \(S \neq \emptyset\) we also request that each part of \(c'\) intersects \(S\). This condition allows to reject partial solutions \(F\) for which a component of \(G[S \cup C]-F\) is strictly contained in \(C\). Indeed, such partial solutions cannot extend to global solutions \(F'\) such that \(G-F'\) is connected. For the particular case \(S = \emptyset\) (hence \(C=V\) and \(W = \emptyset\)) we only consider characteristics \(c'\) with a single part. This ensures that the global solution \(F\) satisfies that \(G-F\) is connected. If there is no such triple \((\Omega, W',c')\), then we set \(\alpha(S,C,W,c) =-\infty\) (we can assume that, when it has no parameters, function \(\max\) returns \(-\infty\)). #### Computing \(\beta\) from \(\alpha\). \[\label{eq:g12} \gamma_1(S,C,\Omega,W,c) = \max_{c'} (\alpha(S_1,C_1,\Omega,W \cap S_1, c') + |W|-|W \cap S_1|),\] over all characteristics if \(c'\) that *map correctly* on \(c\), in the following sense. Consider a characteristic \(c'\) and let \(G_{c'}[\Omega \setminus W]\) be the graph obtained from \(G[\Omega \setminus W]\) by completing each part \(D \in c'\) into a clique. We say that a characteristic \(c'\) maps correctly on \(c\) if \(c\) is the partition of \(\Omega \setminus W\) defined by the connected components of \(G_{c'}[\Omega \setminus W]\). The notion of mapping transforms the characteristic of the partial solution \(F_1\) w.r.t. \((S_1,C_1,W \cap S_1)\) into the characteristic of \(F_1 \cup W\) w.r.t. the quadruple \((S,C,\Omega,W)\). For all \(i\), \(2 \leq i \leq p\), \[\label{eq:gi2} \gamma_i(S,C,\Omega,W,c) = \max_{c_{i-1},c_i}(\gamma_{i-1}(S,C,\Omega,W,c_{i-1}) + \alpha(S,C,\Omega,W \cap S_i,c_i)-|W \cap S_i|),\] over all pairs of characteristics \(c_{i-1},c_i\) that *map correctly* on \(c\). That is, \(c\) must correspond to the connected components of \(G_{c_{i-1},c_i}[\Omega \setminus W]\), obtained from \(G[\Omega \setminus W]\) by completing each part of \(c_{i-1}\) and each part \(c_i\) of into a clique. Finally \[\label{eq:bfg2} \beta(S,C,\Omega,W,c) = \gamma_p(S,C,\Omega,W,c).\] The optimal solution size is, as before, \(\alpha(\emptyset,V,\emptyset,\{\emptyset\})\). In general, the number of characteristics may be exponential. Nevertheless, there are classes of graphs with the property that each minimal separator \(S\) and each potential maximal clique \(\Omega\) can be partitioned into at most a constant number of cliques. With this constraint, the number of characteristics is polynomial (even constant, for any given triple \((S,C,W)\) or quadruple \((S,C,\Omega,W)\)). This is the case for chordal graphs, where each minimal separator and each potential maximal clique induces a clique in \(G\). It is also the case for circular-arc graphs. Recall that each minimal separator corresponds to the set of arcs intersecting a pair of scanpoints . Moreover, by , each potential maximal clique corresponds to the set of arcs intersecting a triple of scanpoints. Since arcs intersecting a given scanpoint form a clique, we have that each minimal separator can be partitioned into two cliques, and each potential maximal clique can be partitioned into three cliques. We deduce: Note that Escoffier *et al.*  already observed that [Connected Vertex Cover]{.smallcaps} is polynomial for chordal graphs. # [Independent Dominating Set]{.smallcaps} and variants {#se:ind} The [Independent Dominating Set]{.smallcaps} problem consists in finding a *minimum* independent set \(F\) of \(G\) such that \(F\) dominates \(G\). Hence the solution \(F\) induces a graph of treewidth \(0\) and it is natural to ask if similar techniques work in this case. The fact that we have a minimization problem is not a difficulty: the general dynamic programming scheme applies in this case, and for any weighted problem with polynomially bounded weights, including negative ones . [Independent Dominating Set]{.smallcaps} is known to be NP-complete in chordal bipartite graphs  and in circle graphs . Therefore, it is NP-hard on \(\mathcal{G}_{\poly}\) for some polynomials \(\operatorname{poly}\). But, again, we can use our scheme in the case of circular-arc graphs, for this problem or any problem of the type minimum dominating induced subgraph of treewidth at most a constant \(t\). Let \((S,C)\) be a block an let \(F \subseteq S \cup C\) be a partial solution compatible with \((S,C,W)\) for some \(W \subseteq S\) of size at most \(t+1\) (in the sense of Definition [\[de:partcomp\]](#de:partcomp){reference-type="ref" reference="de:partcomp"}). The natural way for defining the characteristic of \(F\) is to specify which vertices of \(S\) are dominated by \(F\) and which are not (we already know that \(F \cap S = W\)). It is thus enough to memorize which vertices of \(S\) are dominated by \(F \cap C\). In circular-arc graphs, this information can be encoded using a polynomial number of characteristics. Indeed, a minimal separator \(S\) corresponds to arcs intersecting a scanline, between two scanpoints \(p_1\) and \(p_2\) of some intersection model of \(G\). Moreover (see ), the vertices of component \(C\) correspond to the arcs situated on one of the sides of the scanline. Let \(s^1_1,s^1_2,\dots, s^1_{l_1}\) be the arcs of the model containing scanpoint \(p_1\), ordered by increasing intersection with the side of \(p_1p_2\) corresponding to \(C\). Simply observe that if \(F \cap C\) dominates vertex \(s^1_i\), it also dominates all vertices \(s^1_j\) with \(j>i\). Therefore we only have to store the vertex \(s^1_{min_1}\) dominated by \(F \cap C\) which has a minimum intersection with the side of the scanline corresponding to component \(C\), and proceed similary for the arcs of \(S\) containing scanpoint \(p_2\). These two vertices of \(S\) will define the characteristic of \(F\), and they suffice to identify all vertices of \(S\) dominated by \(F \cap C\). These characteristics can be used to compute a minimum dominating induced subgraph of treewidth at most \(t\), for circular-arc graphs, in polynomial time. We will not show, in details, how to do it, since the technique is quite classical. Problem [Independent Dominating Set]{.smallcaps} is already known to be polynomial for this class . The algorithm of Vatshelle  is more general, based on parameters called *boolean-width* and *MIM-width*, which are small (\(\mathcal{O}(\log n)\) for the former, constant for the latter) on circular-arc graphs and also other graph classes. Another problem of similar flavor, combining domination and independence, is [Red-Blue Dominating Set]{.smallcaps}. In this problem we are given a bipartite graph \(G=(R,B,E)\) with red and blue vertices, and an integer \(k\), and the goal is to find a set of at most \(k\) blue vertices dominating all the red ones. [Red-Blue Dominating Set]{.smallcaps} can be reduced to [Connected Vertex Cover]{.smallcaps} as follows . Let \(G'\) be the graph obtained from \(G=(R,B,E)\) by adding a new vertex \(u\) adjacent to all vertices of \(B\) and then, for each \(v \in R \cup \{u\}\), a pendant vertex \(v'\) adjacent only to \(v\). Then \(G\) has a red-blue dominating set of size at most \(k\) if and only if \(G'\) has a connected vertex cover of size at most \(k+|B|+1\). Indeed any minimum connected vertex cover of \(G'\) must contain \(u\), \(R\), and a subset of \(B\) dominating \(R\). It is not hard to prove that this reduction increases the number of minimal separators by at most \(\mathcal{O}(n)\), see Appendix [\[app:red\]](#app:red){reference-type="ref" reference="app:red"}. Therefore, if [Red-Blue Dominating Set]{.smallcaps} is NP-hard on (bipartite) \(\mathcal{G}_{\poly}\) for some \(poly\), so is [Connected Vertex Cover]{.smallcaps}. There are two natural, well-studied classes of bipartite graphs with polynomial number of minimal separators, and it turns out that [Red-Blue Dominating Set]{.smallcaps} is polynomial for both. One is the class of chordal bipartite graphs (which are actually defined as the bipartite, *weakly* chordal graphs). For this class, [Red-Blue Dominating Set]{.smallcaps} is polynomial by . Reference  considers the total domination problem for the class, but the approach is based on red-blue domination. The second natural class is the class of circle bipartite graphs, i.e., bipartite graphs that are also circle graphs. They have an elegant characterization established by de Fraysseix . Let \(H = (V,E)\) be a planar multigraph, and partition its edge set into two parts \(E_R\) and \(E_B\) such that \(T=(V,E_R)\) is a spanning tree of \(H\). Let \(B(H, E_R) = (E_R, E_B, E')\) be the bipartite graph defined as follows: \(E_R\) is the set of red vertices, \(E_B\) is the set of blue vertices, and \(e_R \in E_R\) is adjacent to \(e_B \in E_B\) if the unique cycle obtained from the spanning tree \(T\) by adding \(e_B\) contains the edge \(e_R\). We say that \(B(H, E_R)\) is a fundamental graph of \(H\). By , a graph is circle bipartite if and only if it is the fundamental graph \(B(H, E_R)\) of a planar multigraph \(H\). Consider now the [Tree augmentation]{.smallcaps} problem that consists in finding, on input \(G\) and a spanning tree \(T\) of \(G\), a minimum set of edges \(D \subseteq E(G)-E(T)\) such that each edge in \(E(T)\) is contained in at least one cycle of \(G' = (V, E(T) \cup D)\). In is shown that [Tree augmentation]{.smallcaps} is polynomial when the input graph is planar. Is direct to see that a set \(S \subseteq E_B\) is a solution of the [Tree augmentation]{.smallcaps} problem on input \(H = (V, E_R \cup E_B)\) and \(T = (V, E_R)\), if and only if \(S\) is a solution of [Red-Blue Dominating Set]{.smallcaps} on input \(B(H) = (E_R, E_B, E')\). This observation, together with and, impliy that [Red-Blue Dominating Set]{.smallcaps} is polynomial in circle bipartite graphs. # Discussion We showed how the dynamic programming scheme of  can be extended for other optimization problems, on *subclasses* of \(\mathcal{G}_{\poly}\). Note that the algorithm of  allows to find in polynomial time, on \(\mathcal{G}_{\poly}\), a maximum (weight) subgraph \(G[F]\) of treewidth at most \(t\), satisfying some property expressible in CMSO. It also handles annotated versions, where the vertices/edges of \(G[F]\) must be selected from a prescribed set. We have seen that [Distance-\(d\) Independent Set]{.smallcaps} can be solved in polynomial time on \(\mathcal{G}_{\poly}\) for any even \(d\). This also holds for the more general problem of finding an induced subgraph \(G[F]\) whose components are at pairwise distance at least \(d\), and such that each component is isomorphic to a graph in a fixed family. E.g., each component could be an edge, to have a variant of [Maximum Induced Matching]{.smallcaps} where edges should be at pairwise distance at least \(d\). For this we need to solve the corresponding problem on \(G^{d-1}\), using only edges from \(G\), as in . When seeking for maximum (resp. minimum) induced subgraphs \(G[F]\) of treewidth at most \(t\) such that \(G-F\) is connected (resp. \(F\) dominates \(G\)) on particular subclasses of \(\mathcal{G}_{\poly}\), we can add any CMSO condition on \(G[F]\). It is not unlikely that the techniques can be extended to other classes than circular-arc graphs (and chordal graphs, for connectivity constraints). We also believe that the interplay between graphs of bounded MIM-width  and \(\mathcal{G}_{\poly}\) deserves to be studied. None of the classes contains the other, but several natural graph classes are in their intersection, and they are both somehow related to induced matchings. We leave as open problems the complexity of [Connected Vertex Cover]{.smallcaps} and [Connected Feedback Vertex set]{.smallcaps} in weakly chordal graphs, and on \(\mathcal{G}_{\poly}\). We have examples showing that, even for weakly chordal graphs, the natural set of characteristics that we used in Section [4](#se:cvc){reference-type="ref" reference="se:cvc"} is not polynomially bounded.
{'timestamp': '2016-07-18T02:07:51', 'yymm': '1607', 'arxiv_id': '1607.04545', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04545'}
null
null
# Introduction Let \(A\) and \(B\) be nonempty subsets of an abelian group, and define the *sumset* of \(A\) and \(B\) and the *\(h\)-fold sumset* of \(A\) as \[\begin{aligned} A+B := \{ a + b: a \in A, b \in B\} \text{\hspace{3px} and \hspace{3px}} hA := \{ a_1 + \ldots + a_h: a_i \in A\}, \end{aligned}\] respectively. When the set \(A\) is implicitly understood, we will reserve the letter \(K\) to denote the doubling constant of \(A\); that is, \(K := |A+A|/|A|\). A classical result of Plünnecke bounds the cardinality of \(hA\) in terms of \(K\) and \(|A|\). See the survey of Ruzsa for variations, generalizations, and a graph theoretic proof of Theorem [\[plunnecke\]](#plunnecke){reference-type="ref" reference="plunnecke"}; see Petridis for a new inductive proof. Given \(\lambda \in \mathbb{Z}\), define a *dilate* of \(A\) as \[\begin{aligned} \lambda \cdot A := \{ \lambda a: a \in A\}. \end{aligned}\] Suppose \(\lambda_1, \ldots, \lambda_h\) are nonzero integers. Since \(\lambda_i \cdot A \subseteq \lambda_i A\), one can apply Theorem [\[plunnecke\]](#plunnecke){reference-type="ref" reference="plunnecke"} to conclude that \[\begin{aligned} |\lambda_1 \cdot A + \ldots + \lambda_h \cdot A| \le K^{\sum_i |\lambda_i|}|A|. \end{aligned}\] Bukh significantly improved this by considering the binary expansion of \(\lambda_i\) and using Ruzsa's covering lemma and triangle inequality. If \(|\lambda_i| \le 2^r\) for all \(i\), then Theorem [\[buk08\]](#buk08){reference-type="ref" reference="buk08"} yields that \[\begin{aligned} \label{logbound} | \lambda_1 \cdot A + \ldots + \lambda_h \cdot A| \le K^{O(rh)}|A|. \end{aligned}\] In this paper we prove a bound that improves [\[logbound\]](#logbound){reference-type="eqref" reference="logbound"} when \(r\) and \(h\) are sufficiently large and comparable to each other. Throughout the paper \(ln\) stands for the natural logarithm. The proof of Theorem [\[mainthm\]](#mainthm){reference-type="ref" reference="mainthm"} relies on Theorem [\[buk08\]](#buk08){reference-type="ref" reference="buk08"} as well as a result of Tuza  on decomposing bipartite graphs into bicliques (complete bipartite subgraphs). The key idea is to connect Bukh's technique of considering the binary expansion of \(\lambda_i\) to the graph decomposition problem that allows us to efficiently group certain powers of 2. We remark here that in all of the above theorems, the condition \(|A+A| = K|A|\) can be replaced with \(|A-A| = K|A|\) with no change to the conclusion. It is likely that Theorem [\[mainthm\]](#mainthm){reference-type="ref" reference="mainthm"} is not best possible--we discuss this in the last section. # Basic Tools We need the following analogue of Ruzsa's triangle inequality, see. A useful corollary of Theorem [\[ruzsatriangle\]](#ruzsatriangle){reference-type="ref" reference="ruzsatriangle"} is as follows. We can use Corollary [\[easyobs\]](#easyobs){reference-type="ref" reference="easyobs"} to prove the following proposition that we will use in the proof of Theorem [\[mainthm\]](#mainthm){reference-type="ref" reference="mainthm"}. # Proof of Theorem [\[mainthm\]](#mainthm){reference-type="ref" reference="mainthm"} Given \(\lambda_1, \ldots, \lambda_h \in \mathbb{Z} \setminus \{0\}\), we define \[\begin{aligned} \label{rands} r := \max_i \lfloor \log_2 |\lambda_i| \rfloor \end{aligned}\] and write the binary expansion of \(\lambda_i\) as \[\begin{aligned} \label{lambda}\lambda_i = \epsilon_i \sum_{j=0}^r \lambda_{i,j} 2^j \text{, where \(\lambda_{i,j} \in \{0,1\}\) and \(\epsilon_i \in \{-1, 1\}\)}. \end{aligned}\] Bukh's proof of Theorem [\[buk08\]](#buk08){reference-type="ref" reference="buk08"} actually gives the following stronger statement. In his proof of Theorem [\[buk08\]](#buk08){reference-type="ref" reference="buk08"}, the first step is to observe that \[\lambda_1 \cdot A + \ldots + \lambda_h \cdot A \subseteq \sum_{j=0}^r (\lambda_{1,j} 2^j) \cdot A + \ldots + \sum_{j=0}^r (\lambda_{h,j} 2^j) \cdot A.\] In our proof, we also consider the binary expansion of \(\lambda_i\), but we do the above step more efficiently by first grouping together \(\lambda_i\) that have shared binary digits. In order to do this systematically, we view the problem as a graph theoretic problem and apply the following result of Tuza . Tuza stated this result for the *covers* of \(G\), where a *cover* of \(G\) is a collection of subgraphs of \(G\) such that every edge of \(G\) is contained in at least one of these subgraphs. However, the cover provided in his proof is indeed a decomposition. Furthermore, the assumption \(n\ge 10 (\ln m) ^2\) was not stated in but such kind of assumption is needed.[^1] Indeed, [\[eq:Tuza\]](#eq:Tuza){reference-type="eqref" reference="eq:Tuza"} becomes false when \(n= o(\ln m)\) because \(\sum_{i=1}^q |V(H_i)| \ge m+ n\) for any cover \(H_1, \ldots, H_q\) of \(G\) if \(G\) has no isolated vertices. Note that [\[eq:Tuza\]](#eq:Tuza){reference-type="eqref" reference="eq:Tuza"} is tight up to a constant factor. Indeed, Tuza  provided a bipartite graph \(G\) with two parts of size \(n\le m\) such that every biclique *cover* \(H_1, \ldots, H_q\) of \(G\) satisfies \(\sum_{i=1}^q |V(H_i)| \ge mn/ (e^2\ln m)\), where \(e= 2.718...\). # Concluding Remarks Instead of Theorem [\[ces\]](#ces){reference-type="ref" reference="ces"}, in an earlier version of the paper we applied a result of Chung, Erdős, and Spencer, which states that every graph on \(n\) vertices has a biclique decomposition \(H_1, \ldots, H_q\) such that \(\sum_{i=1}^q |V(H_i)| \le (1+o(1)) n^2/ (2\ln n)\). Instead of [\[eq:main\]](#eq:main){reference-type="eqref" reference="eq:main"}, we obtained that \[|\lambda_1 \cdot A + \ldots + \lambda_h \cdot A | \le K^{O((r+h)^2/\ln (r+h) ) } |A|.\] This bound is equivalent to [\[eq:main\]](#eq:main){reference-type="eqref" reference="eq:main"} when \(r=\Theta(h)\) but weaker than Bukh's bound [\[logbound\]](#logbound){reference-type="eqref" reference="logbound"} when \(r\) and \(h\) are not close to each other. Although the assumption [\[eq:mn\]](#eq:mn){reference-type="eqref" reference="eq:mn"} may not be optimal, Theorem [\[mainthm\]](#mainthm){reference-type="ref" reference="mainthm"} is not true without any assumption on \(r\) and \(h\). For example, when \(r\) is large and \(h=o(\ln r)\), [\[eq:main\]](#eq:main){reference-type="eqref" reference="eq:main"} becomes \(|\lambda_1 \cdot A + \ldots + \lambda_h \cdot A | \le K^{ o(r ) } |A|\), which is false when \(A= \{1, \dots, n\}\). If each of \(\lambda_1, \dots, \lambda_r\) has \(O(1)\) digits in its binary expansion, then Theorem [\[buk08\]](#buk08){reference-type="ref" reference="buk08"} yields that \(|\lambda_1 \cdot A + \ldots + \lambda_r \cdot A| \le K^{O(r)}|A|\). Bukh asked if this bound holds whenever \(\lambda_i \le 2^r\): In light of Question [\[bukhconj\]](#bukhconj){reference-type="ref" reference="bukhconj"}, one can view Theorem [\[mainthm\]](#mainthm){reference-type="ref" reference="mainthm"} as providing modest progress by proving a subquadratic bound of quality \(O(r^2/\ln r)\) whereas Theorem [\[buk08\]](#buk08){reference-type="ref" reference="buk08"} shows that the exponent is \(O(r^2)\). Generalized arithmetic progressions give supporting evidence for Question [\[bukhconj\]](#bukhconj){reference-type="ref" reference="bukhconj"}. A *generalized arithmetic progression* \(P\) is a set of the form \[P := \{ d + x_1 d_1 + \ldots + x_k d_k: 0 \le x_i < L_i \}.\] Moreover, \(P\) is said to be *proper* if \(|P| = L_1 \cdot \ldots \cdot L_k\). One can calculate that if \(P\) is proper, then \[|P+P| \le (2L_i-1)^k \le 2^k |P| =: K |P|.\] Additionally, one can calculate that for any \(\lambda_1, \ldots, \lambda_h \in \mathbb{Z}^+\), if \(\lambda_i \le 2^r\) then \[|\lambda_1 \cdot P + \ldots + \lambda_h \cdot P| \le (\lambda_1 + \ldots + \lambda_h)^k |P| = 2^{k \log_2(\lambda_1 + \ldots + \lambda_h)}|P| = K^{r + \log_2{h}}|P|.\] Freiman's theorem says that, roughly speaking, sets with small doubling are contained in generalized arithmetic progressions with bounded dimension. Using this line of reasoning, Schoen and Shkredov proved that \[|\lambda_1 \cdot A + \ldots + \lambda_h \cdot A | \le e^{O( \log_2^6 (2K) \log_2\log_2 (4K)) (h + \log_2 \sum_i |\lambda_i| ) }|A|.\] This naturally leads us to ask a more precise version of Question [\[bukhconj\]](#bukhconj){reference-type="ref" reference="bukhconj"}.
{'timestamp': '2017-08-29T02:02:17', 'yymm': '1607', 'arxiv_id': '1607.04888', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04888'}
# Introduction Photonic crystals are systems with a spatially periodic structure of refractive medium. Electromagnetic (EM) fields in a photonic crystal are given by the Bloch states where frequency dispersion may have an energy gap as the electronic band dispersion of solids. This gives us an opportunity to simulate microscopic quantum states of electrons by the macroscopic EM waves governed by the classical Maxwell equations. Especially, simulating topologically nontrivial phases in quantum solids has been an extensively studied topic these days. A prototypical topological phase of matter is an integer quantum Hall state, where its topological nature is encoded in a topological invariant, the Chern number, calculated using the Bloch wave functions. A physical consequence of the topological nature is quantization of the Hall conductance, since it is proportional to the Chern number. On the other hand, the quantization is also explained in terms of the chiral edge states. Generally, there is an intimate relation between bulk topology and edge states, as known as the bulk-edge correspondence. So, we can access topological properties of the bulk via edge states, or inversely, we can know about the edge/surface states by the bulk information alone. The idea of the bulk-edge correspondence is especially important in photonic systems, since there is no "Hall conductance" in photonic systems, while edge states are always physical observable. Since we have the Bloch states and the Brillouin zone in periodic photonic systems, the Chern number is well-defined as in electronic systems. Then, comparing the bulk Chern number and the edge states is a natural strategy to attack topological issues in photonic systems. In fact, quantum Hall state analogs in 2D photonic crystals have been discussed focusing on the edge states. Note that in such studies, the time reversal symmetry (TRS) breaking is mandatory to realize a quantum Hall analog. In this paper, a Gaussian type localized basis set based method is introduced for numerical evaluation of the Bloch states of the EM waves. With the localized basis set, momentum, which defines the Brillouin zone, is treated as a twisted boundary condition \(e^{i\phi_\alpha}\), \(\alpha=x,y\), and then, the basis functions naturally and strictly becomes periodic in momentum \(\phi_\alpha\). Since we will explain later, this feature is advantageous in the Chern number calculation, and this is the reason for our choice of the localized basis set. The Gaussian expansion is not the only choice for the localized basis set, i.e., for instance, the finite element method gives an alternative. However, the Gaussian basis element is convenient when derivatives of the basis functions are required, obviously due to its Gaussian nature. Such a situation really arises in 3D crystals where the decomposition of the EM field into TE or TM modes is impossible. We demonstrate the merit of the Gaussian expansion through the calculation of the "section" Chern numbre of the 3D photonic crystals. As we explained, the introduced method is applied to 3D photonic crystals. First of all, note that it is possible to define and use the Chern number to characterize the given system even in 3D cases as follows. That is, let us fix one of the three components of the momentum, say \(k_3\), and define the Chern number \(C (k_3)\) using \((k_1,k_2)\) as a periodic parameter space. Since this parameter space is a 2D section of the 3D Brillouin zone, we call \(C (k_3)\) the section Chern number. This section Chern number is well defined only when the band gap is always finite on the constant \(k_3\) plane (section). Different from the usual Chern number in 2D systems, the section Chern number can be finite even in systems having TRS. This feature enables us to realize a topologically nontrivial state in photonic crystals without magneto-optical media. Specifically, \(C (k_3)\) can take a nonzero value when the TRS *or* the spatial inversion symmetry (SIS) is broken. Since the section Chern number \(C (k_i)\) is a topological invariant, it can change only when the band gap vanishes at a certain point on the constant \(k_i\) plane. Generally, this gap closing point is a Weyl point in the band structure, having a linear dispersion around it. Intriguing topological properties are expected in the system with the Weyl points. Also, it has been already shown that the Weyl points emerge in the double gyroid photonic crystal when TRS or SIS is broken. In this paper, we consider a photonic crystal of a simple structure with Weyl points to discuss the topological nature. In this paper, we limit ourselves to the cases with TRS but without SIS. The considered system is anisotropic and the Weyl points are clearly resolved only in a constant \(k_z\) plane in the Brillouin zone. We also discuss the edge states associated with the nonzero section Chern number, which can be realized by making a wave packet. This paper is organized as follows. In Sec. II, we introduce the Gaussian expansion method for the photonic crystals, and explain the importance of the localized basis set for the calculation of the Chern number. Section III is devoted to demonstrate the validity and the usefulness of the introduced method by applying the method to simple models. In Sec. IV, a specific photonic crystal having the Weyl points in its photonic band structure is proposed, and the section Chern number for that model is evaluated. The relation between the section Chern number and the edge modes is also discussed. Summary and discussion are given in Sec. V. # Methods The Maxwell equations for the normal modes introduced as \(\boldsymbol{E} \negthinspace \left( \boldsymbol{r}, t \right) = {\mathrm{e}^{-\mathrm{i} \omega t}} \boldsymbol{E} \negthinspace \left( \boldsymbol{r} \right)\) and \(\boldsymbol{H} \negthinspace \left( \boldsymbol{r}, t \right) = {\mathrm{e}^{-\mathrm{i}\omega t}} \boldsymbol{H} \negthinspace \left( \boldsymbol{r} \right)\) are written as \[\begin{aligned} \nabla \times \boldsymbol{E} \negthinspace \left( \boldsymbol{r} \right) &= {\mathrm{i} \omega} \boldsymbol{B}(\bm{r}), & \nabla \times \boldsymbol{H} \negthinspace \left( \boldsymbol{r} \right) &= {-\mathrm{i} \omega } \boldsymbol{D}(\bm{r}), \\ \nabla \cdot \boldsymbol{D} \negthinspace \left( \boldsymbol{r} \right) &= 0, & \nabla \cdot \boldsymbol{B} \negthinspace \left( \boldsymbol{r} \right) &= 0. \label{Maxwell-eq} \end{aligned}\] Here, we assume that the permittivity and the permeability are linear and lossless, i.e., we assume \[\boldsymbol{D} \negthinspace \left( \boldsymbol{r} \right) = \varepsilon_0 \hat{\varepsilon} \negthinspace \left( \boldsymbol{r} \right) \boldsymbol{E} \negthinspace \left( \boldsymbol{r} \right), \thickspace \boldsymbol{B} \negthinspace \left( \boldsymbol{r} \right) = \mu_0 \hat{\mu} \negthinspace \left( \boldsymbol{r} \right) \boldsymbol{E} \negthinspace \left( \boldsymbol{r} \right),\] where \(\hat{\varepsilon} \negthinspace \left( \boldsymbol{r} \right)\) and \(\hat{\mu} \negthinspace \left( \boldsymbol{r} \right)\) are \(3 \times 3\) tensors satisfying the Hermitian conditions \(\hat{\varepsilon}^\dagger = \hat{\varepsilon}\) and \(\hat{\mu}^\dagger = \hat{\mu}\). With this assumption, the Maxwell equations are rewritten into Hermitian eigenequation forms as \[\begin{aligned} \nabla \times \hat{\varepsilon}^{-1}\negthinspace \left( \boldsymbol{r} \right) \nabla \times \boldsymbol{H} \negthinspace \left( \boldsymbol{r} \right) &= \left( \frac{\omega}{c} \right)^2 \hat{\mu} \negthinspace \left( \boldsymbol{r} \right) \boldsymbol{H} \negthinspace \left( \boldsymbol{r} \right), \label{eigenequation-H}\\ \nabla \times \hat{\mu}^{-1}(\bm{r}) \nabla \times \boldsymbol{E} \negthinspace \left( \boldsymbol{r} \right) &= \left( \frac{\omega}{c} \right)^2 \hat{\varepsilon} \negthinspace \left( \boldsymbol{r} \right) \boldsymbol{E} \negthinspace \left( \boldsymbol{r} \right). \label{eigenequation-E} \end{aligned}\] We only have to solve the either of these eigenequations with respect to \(\boldsymbol{E}\) or \(\boldsymbol{H}\). In a spatially periodic system, all the eigenstates are represented as Bloch states. Roughly speaking, there are two ways to expand a Bloch state. The first way is to use a basis set localized in wave space like plane waves, and the second way is to use a basis set localized in real space. For the plane wave type basis sets, a Bloch state is usually represented as \[\psi_{k} \negthinspace \left( \boldsymbol{r} \right) = \sum_G f_G {\mathrm{e}^{\mathrm{i} \left(\boldsymbol{G} + \boldsymbol{k} \right)\cdot \boldsymbol{r} }}, \label{plane-wave}\] with \(\boldsymbol{G}\) denoting reciprocal lattice vectors. In this form, each basis element is not periodic for a parameter change \(\boldsymbol{k} \rightarrow \boldsymbol{k} + \boldsymbol{G}\), since \(\mathrm{e}^{\mathrm{i} \boldsymbol{k} \cdot \boldsymbol{r} }\) and \(\mathrm{e}^{\mathrm{i} \left( \boldsymbol{G} + \boldsymbol{k}\right) \cdot \boldsymbol{r} }\) are not identical. On the other hand, for spatially localized basis sets, the matrix eigenequation can be naturally made into periodic in parameter \(k\). A Bloch state is represented by localized basis set as \[\psi_k \negthinspace \left( \boldsymbol{r} \right) = \frac{1}{\sqrt{N}} \sum_m {\mathrm{e}^{\mathrm{i} \boldsymbol{k} \cdot \boldsymbol{R}_m }} w_m \negthinspace \left(\boldsymbol{r} \right), \label{localize-wavenumber}\] with \(\boldsymbol{R}_m\) and \(N\) representing a lattice vector in real space and the number of unit cells of the system under consideration respectively, and \(w_m \negthinspace \left( \boldsymbol{r} \right) = w_0 \negthinspace \left( \boldsymbol{r}-\boldsymbol{R}_m \right)\). In this form, a parameter \(k\) only appears as \(\boldsymbol{e}^{\mathrm{i} \boldsymbol{k} \cdot \boldsymbol{R}_m }\) and is introduced in the eigenequation as a boundary condition. This periodicity is crucial in the numerical evaluation of the Chern number. Representative localized functions are the maximally localized Wannier functions, however, it is known that the so-called *composite bands*, which compose the maximally localized Wannier functions, must be distinguished from the bands below and above by gaps with vanishing Chern number. However, photonic bands are generally tangling in higher frequencies and it is not common for photonic band structures to have an isolated set of bands enclosed by gaps. Then, we use a Gaussian basis set as a typical localized basis set in the following. Explicitly, the Gaussian basis element is written as \[\boldsymbol{g}_{i,k}^{\lambda} \negthinspace \left( \boldsymbol{r} \right) = \frac{1}{\sqrt{N}} \sum_{m,\lambda} {\mathrm{e}^{\mathrm{i} \boldsymbol{k} \cdot \boldsymbol{R}_m}} \boldsymbol{f}^{\lambda} \negthinspace \left( \boldsymbol{r}_{im} \right) \mathrm{e}^{- r_{im}^2 / \alpha^2}, \label{gaussian-wavenumber}\] with \[\bm{f}^{ \lambda} (\bm{r}_{im})= \begin{pmatrix} f_x^{ \lambda} \negthinspace \left(x_{im},y_{im},z_{im} \right)\\ f_y^{ \lambda} \negthinspace \left(x_{im},y_{im},z_{im} \right)\\ f_z^{ \lambda} \negthinspace \left(x_{im},y_{im},z_{im} \right) \end{pmatrix}.\] Here, suffices \(i\) and \(\lambda\) specify the grid position within the unit cell and the polarization degrees of freedom respectively, and \(\boldsymbol{r}_{im} = \boldsymbol{r}-\left( \boldsymbol{R}_m + \boldsymbol{r}_i \right)\). The \(f_i^{ \lambda}\negthinspace \left( x, y, z \right)\) is a polynomial of \(\left(x, y, z\right)\). Using this basis set for the eigenfunction expansion in Eq. [\[eigenequation-H\]](#eigenequation-H){reference-type="eqref" reference="eigenequation-H"}, we obtain a matrix eigenequation \[\sum_{j,\lambda'} \Theta_{i\lambda,j \lambda'} \negthinspace \left( \boldsymbol{k} \right) c_{j}^{\lambda'} = \left( \frac{\omega}{c} \right)^2 \sum_{j,\lambda'} O_{i\lambda,j\lambda'} \negthinspace \left( \boldsymbol{k} \right) c_{j}^{\lambda'}. \label{eigenequation-matrix}\] Here, \(\Theta_{i\lambda,j\lambda'} \negthinspace \left(\boldsymbol{k}\right)\) is the matrix element of the operator in the left hand side of Eq. [\[eigenequation-H\]](#eigenequation-H){reference-type="eqref" reference="eigenequation-H"}, i.e., \(\hat{\Theta} = \nabla \times \hat{\varepsilon}^{-1} \negthinspace \left( \boldsymbol{r} \right) \nabla \times\), which is explicitly written as \[\begin{aligned} \Theta_{i\lambda,j\lambda'} \negthinspace \left( \boldsymbol{k} \right) &&= \left< \boldsymbol{g}_{ik}^{ \lambda} \right| \nabla \times \hat{\varepsilon}^{-1} \negthinspace \left( \boldsymbol{r} \right) \nabla \times \left| \boldsymbol{g}_{jk}^{\lambda'} \right> \notag \\ &&= \sum_{m} {\mathrm{e}^{-\mathrm{i} \boldsymbol{k} \cdot \boldsymbol{R}_m }} \int \mathrm{d}^3 r \left( \nabla \times \boldsymbol{f}^{\lambda} \negthinspace \left( \boldsymbol{r}_{im} \right) \mathrm{e}^{-r_{im}^2/\alpha^2 }\right)^\ast \notag \\ && \qquad \qquad \cdot \hat{\varepsilon}^{-1} \negthinspace \left( \boldsymbol{r} \right) \left( \nabla \times \boldsymbol{f}^{\lambda} \negthinspace \left( \boldsymbol{r}_{j0} \right) \mathrm{e}^{-r_{j0}^2 / \alpha^2 }\right). \label{matrix-element-theta} \end{aligned}\] Here we used the Hermiticity of the rotation, \(\int \mathrm{d}^3 r \thickspace \boldsymbol{g}^\ast \negthinspace \left( \boldsymbol{r}\right) \cdot \left\{ \nabla \times \boldsymbol{f} \negthinspace \left( \boldsymbol{r}\right) \right\} = \int \mathrm{d}^3 r \thickspace \boldsymbol{f} \negthinspace \left( \boldsymbol{r}\right) \cdot \left\{ \nabla \times \boldsymbol{g} \negthinspace \left( \boldsymbol{r}\right) \right\}^\ast\) under the boundary condition, \(\displaystyle \lim_{r \rightarrow \infty } \boldsymbol{g}_{i,k}^{ \lambda} \negthinspace \left( \boldsymbol{r} \right) = 0\). On the other hand, \(O_{i\lambda,j\lambda'} \negthinspace \left( \boldsymbol{k} \right)\) is the overlap, which is obtained as, \[\begin{aligned} O_{i\lambda,j\lambda'} \negthinspace \left( \boldsymbol{k} \right) &&= \left< \boldsymbol{g}_{ik}^{\lambda} \right| \hat{\mu} \negthinspace \left( \boldsymbol{r} \right) \left| \boldsymbol{g}_{jk}^{\lambda'} \right> \notag \\ &&= \sum_{m} {\mathrm{e}^{-\mathrm{i} \boldsymbol{k} \cdot \boldsymbol{R}_m } } \int \mathrm{d}^3 r \left( \boldsymbol{f}^{\lambda} \negthinspace \left( \boldsymbol{r}_{im} \right) \mathrm{e}^{- r_{im}^2 / \alpha^2 } \right)^\ast \notag \\ && \qquad \qquad \cdot \hat{\mu} \negthinspace \left( \boldsymbol{r} \right) \left( \boldsymbol{f}^{\lambda'} \negthinspace \left( \boldsymbol{r}_{j0} \right) \mathrm{e}^{- r_{j0}^2 / \alpha^2} \right). \label{matrix-element-overlap} \end{aligned}\] The integration is performed over the infinite spatial region. In many cases, a photonic crystal is constituted by repetition of structures having different permittivity or permeability from the uniform media. Then, it is convenient to separate the inverse of permittivity \(\hat{\varepsilon}^{-1} \negthinspace \left( \boldsymbol{r} \right)\) as \[\hat{\varepsilon}^{-1} \negthinspace \left( \boldsymbol{r} \right) = \hat{\varepsilon}_c^{-1} + \left( \hat{\varepsilon}^{-1} \negthinspace \left( \boldsymbol{r} \right)-\hat{\varepsilon}_c^{-1} \right) \label{permittivity-separate}.\] The first term of the right hand side is constant corresponding to the uniform background media. The second term is nonzero only in the region where the structures exist. For the constant term, the integrals of the matrix elements are the standard Gaussian integral, which can be analytically evaluated. For the second term, the numerical integration is easy since the integrand is finite only in the limited region. The same argument also applies to the inverse of the permeability \(\hat{\mu}^{-1} \negthinspace \left(\boldsymbol{r} \right)\). For a uniform system all the matrix elements are given without numerical integration. ## Gaussian basis set (2D case) First, we consider the 2D case. Here, 2D means that \(\hat{\varepsilon} \negthinspace \left( \boldsymbol{r} \right)\) and \(\hat{\mu} \negthinspace \left( \boldsymbol{r} \right)\) have no dependence on \(z\). In such a system, a mirror plane perpendicular to the \(z\) axis exists and the solutions of Eqs. [\[eigenequation-H\]](#eigenequation-H){reference-type="eqref" reference="eigenequation-H"} and [\[eigenequation-E\]](#eigenequation-E){reference-type="eqref" reference="eigenequation-E"} are separated into TE modes (\(H_z\) polarization) or TM modes (\(E_z\) polarization). As a result, we only have to consider the \(z\) component of either field \(\boldsymbol{H}\) or \(\boldsymbol{E}\), which indicates that the equation to be solved becomes a scalar equation. For the scalar eigenequation, we use a basis element \[g_{ik} \negthinspace \left( \boldsymbol{r} \right) = \frac{1}{\sqrt{N}} \sum_m {\mathrm{e}^{\mathrm{i} \boldsymbol{k} \cdot \boldsymbol{R}_m }} \mathrm{e}^{-r_{im}^2 / \alpha^2 }. \label{2d-gaussian-base}\] Each basis function is cylindrically symmetric and there are finite overlaps between the basis functions. Then, these basis elements are put on the grid dividing the unit cell regularly (Fig.[\[gaussian-picture\]](#gaussian-picture){reference-type="ref" reference="gaussian-picture"}). Later, we will see that the simple configuration of the basis elements gives sufficient accuracy for our purpose, and there arises no need for further attention on the distribution of the grids, which is often important in the standard finite element methods. For this basis set, the integral of the overlap \(O_{ij} \negthinspace \left( \boldsymbol{k} \right)\) for the constant term (when \(\thinspace \hat{\mu}_c = \hat{1}\)) is obtained as \[O_{ij}^\text{empty} \negthinspace \left( \boldsymbol{k} \right) = \sum_{m} \frac{\pi \alpha^2 }{2} {\mathrm{e}^{-\mathrm{i} \boldsymbol{k} \cdot \boldsymbol{R}_m}} \mathrm{e}^{-\frac{{r_m'}^2 }{2 \alpha^2} }, \label{2d-overlap-value}\] where \(\boldsymbol{r}_m' = \boldsymbol{r}_i + \boldsymbol{R}_m-\boldsymbol{r}_j\). On the other hand, the integral of the matrix element \(\Theta_{ij} \negthinspace \left( \boldsymbol{k} \right)\) for the constant term (when \(\hat{\varepsilon}_c^{-1} = \hat{1}\)) is \[\Theta_{ij}^\text{empty} \negthinspace \left( \boldsymbol{k} \right) = \sum_{m} \frac{\pi \alpha^2}{2} {\mathrm{e}^{-\mathrm{i} \boldsymbol{k} \cdot \boldsymbol{R}_m}} \mathrm{e}^{-\frac{{r_m'}^2}{2 \alpha^2} } \frac{1}{\alpha^2} \left( 2-\frac{{r_m'}^2}{\alpha^2} \right).\label{2d-matrix-value}\] For the empty lattice, all the matrix elements of Eq. [\[eigenequation-matrix\]](#eigenequation-matrix){reference-type="eqref" reference="eigenequation-matrix"} are given by using Eqs. [\[2d-overlap-value\]](#2d-overlap-value){reference-type="eqref" reference="2d-overlap-value"} and [\[2d-matrix-value\]](#2d-matrix-value){reference-type="eqref" reference="2d-matrix-value"}. When some structures are introduced, the additional work to obtain matrix elements is the numerical integration within the structures. ## Gaussian basis set (3D case) For a 3D system, it is generally impossible to decompose the EM fields into TE or TM modes, and we have to handle all three components of the vector eigenequation. A naive thought suggests us to use the scalar Gaussian functions \[Eq. [\[2d-gaussian-base\]](#2d-gaussian-base){reference-type="eqref" reference="2d-gaussian-base"}\] for each component of the vector as \[\boldsymbol{\psi}_k \negthinspace \left( \boldsymbol{r} \right) = \sum_{ \lambda = 1}^3 \boldsymbol{e}_{ \lambda} \sum_i c_{ik}^{\lambda} g_{ik} \negthinspace \left( \boldsymbol{r} \right). \label{extensionAtFirst}\] Here, \(\boldsymbol{e}_{ \lambda}\) represents unit vectors in three orthogonal directions. This naively introduced basis set, however, has a deficiency in the following sense. Firstly, note that the divergence of the Eq. [\[eigenequation-H\]](#eigenequation-H){reference-type="eqref" reference="eigenequation-H"} vanishes because of \(\nabla \cdot \left( \nabla \times \boldsymbol{f} \negthinspace \left( \boldsymbol{r} \right) \right) = 0\) for any vector field \(\boldsymbol{f} \negthinspace \left( \boldsymbol{r} \right)\). Thus, when the permeability is constant and isotropic, the solution of Eq. [\[eigenequation-H\]](#eigenequation-H){reference-type="eqref" reference="eigenequation-H"} satisfies the constraint, \[\nabla \cdot \boldsymbol{H} \negthinspace \left( \boldsymbol{r} \right) = 0 \label{constraint}.\] On the other hand, the divergence of Eq. [\[extensionAtFirst\]](#extensionAtFirst){reference-type="eqref" reference="extensionAtFirst"} is \[\nabla \cdot \sum_{i, \lambda} \boldsymbol{e}_{ \lambda} c_{ik}^{\lambda} g_{ik} \negthinspace \left( \boldsymbol{r} \right) = \sum_{i,\lambda} \left(-2 \alpha x_{\lambda} \right) c_{ik}^{\lambda} g_{ik} \negthinspace \left( \boldsymbol{r} \right),\] where the right-hand side does not vanish except in the trivial case with \(c_{ik}^{ \lambda} = 0\) for all combinations of \(i\) and \({\lambda}\). Namely, expansion Eq. [\[extensionAtFirst\]](#extensionAtFirst){reference-type="eqref" reference="extensionAtFirst"} does not satisfy constraint Eq.[\[constraint\]](#constraint){reference-type="eqref" reference="constraint"}, and as a consequence, the spectrum of the eigenvalues includes unphysical spurious values. For each basis to satisfy constraint Eq. [\[constraint\]](#constraint){reference-type="eqref" reference="constraint"}, we modify each basis by taking its rotation as \[\begin{split} \boldsymbol{g}_{ik}^{\lambda} \negthinspace \left( \boldsymbol{r}\right) &= \nabla \times \boldsymbol{e}_{\lambda} g_{ik} \negthinspace \left( \boldsymbol{r} \right) \\ &= \sum_{ \mu, \nu = 1}^3 \mathrm{\varepsilon}_{\mu \lambda \nu} \left(-2 \alpha x_{\mu} \right) g_{ik} \negthinspace \left( \boldsymbol{r} \right) \boldsymbol{e}_{\nu}, \label{the-base} \end{split}\] where \(\varepsilon_{\lambda \mu \nu}\) is an anti-symmetric symbol \(\left(\varepsilon_{123} = 1 \right)\). This is also obtained by choosing three independent basis elements of the lowest order from Eq. [\[gaussian-wavenumber\]](#gaussian-wavenumber){reference-type="eqref" reference="gaussian-wavenumber"} under the divergence-free condition. The modified basis element Eq. [\[the-base\]](#the-base){reference-type="eqref" reference="the-base"} has no divergence, and satisfies the constraint \[Eq. [\[constraint\]](#constraint){reference-type="eqref" reference="constraint"}\]. We expand the solution using these basis elements as \[\boldsymbol{\psi}_k \negthinspace \left( \boldsymbol{r} \right) = \sum_{{\lambda} = 1}^3 \sum_i c_{ik}^{\lambda} \boldsymbol{g}_{ik}^{\lambda} \negthinspace \left( \boldsymbol{r} \right). \label{expansion-vec}\] Here again, we write down the integral values for this basis set. The integrals for the constant term of the overlaps \(O_{iz,jz} \left( \boldsymbol{k} \right)\) and \(O_{ix,jz} \left( \boldsymbol{k} \right)\) are \[\begin{aligned} O_{iz,jz}^\text{empty} \negthickspace \left( \boldsymbol{k} \right) && = \sum_{m} \sqrt{\left(\frac{\pi \alpha^2}{2} \right)^3} {\mathrm{e}^{-\mathrm{i} \boldsymbol{k} \cdot \boldsymbol{R}_m }} \mathrm{e}^{-\frac{{r_m'}^2}{2 \alpha^2} } \notag \\ && \qquad \qquad \qquad \times \frac{1}{\alpha^2} \left\{ 2-\frac{1}{\alpha^2} \left( {x_m'}^2 + {y_m'}^2 \right) \right\} \label{3d-overlap-value1} \quad \\ O_{ix,jz}^\text{empty} \negthickspace \left( \boldsymbol{k} \right) && = \sum_m \sqrt{\left(\frac{\pi \alpha^2}{2}\right)^3} {\mathrm{e}^{-\mathrm{i} \boldsymbol{k} \cdot \boldsymbol{R}_m}} \mathrm{e}^{-\frac{{r_m'}^2}{2 \alpha^2} } \frac{1}{\alpha^4} x_m' z_m' \label{3d-overlap-value2} \end{aligned}\] respectively, while the integrals for the matrix elements \(\Theta_{iz,jz} \left(\boldsymbol{k}\right)\) and \(\Theta_{ix,jz} \left(\boldsymbol{k}\right)\) are \[\begin{aligned} && \Theta_{iz,jz}^\text{empty} \negthickspace \left(\boldsymbol{k}\right) = \sum_m \sqrt{\left(\frac{\pi \alpha^2}{2}\right)^3} {\mathrm{e}^{-\mathrm{i} \boldsymbol{k} \cdot \boldsymbol{R}_m }} \mathrm{e}^{-\frac{{r_m'}^2}{2 \alpha^2}} \notag \\ && \qquad \times \frac{1}{\alpha^4} \Big[ \frac{1}{\alpha^4} \left( {x_m'}^2 + {y_m'}^2 \right)^2-\frac{9}{\alpha^2} \left( {x_m'}^2 + {y_m'}^2 \right)\notag \\ && \qquad \qquad + \frac{{z_m'}^2}{\alpha^2} \Big\{ \frac{1}{\alpha^2} \left( {x_m'}^2 + {y_m'}^2 \right)-2 \Big\} + 10\Big] \label{3d-matrix-value1} \\ && \Theta_{ix,jz}^\text{empty} \negthickspace \left(\boldsymbol{k}\right) = \sum_m \sqrt{\left(\frac{\pi \alpha^2}{2}\right)^3} {\mathrm{e}^{-\mathrm{i} \boldsymbol{k} \cdot \boldsymbol{R}_m } } \mathrm{e}^{-\frac{{r_m'}^2 }{2 \alpha^2} } \notag \\ && \qquad \qquad \qquad \qquad \times \frac{{x_m'} {z_m'}}{\alpha^6} \left( 7-\frac{{r_m'}^2}{\alpha^2} \right) \label{3d-matrix-value2} \end{aligned}\] respectively. The integrals of the other components are obtained by permutations of the indices \(x\), \(y\), and \(z\). In addition, the integrals are invariant against the simultaneous exchange of the position of the localization center and the direction of the polarization \(\left( i \lambda' \leftrightarrow j \lambda \right)\). Thus all matrix components for an isotropic bulk dielectric are given by Eqs. [\[3d-overlap-value2\]](#3d-overlap-value2){reference-type="eqref" reference="3d-overlap-value2"}-[\[3d-matrix-value2\]](#3d-matrix-value2){reference-type="eqref" reference="3d-matrix-value2"}. The grid points for these basis elements can be chosen to divide the unit cell of the Bravais lattice with a regular spacing like Fig. [\[gaussian-picture\]](#gaussian-picture){reference-type="ref" reference="gaussian-picture"}. Here, we make some comments on the case of spatially varying \(\hat{\mu}(\bm{r})\). In such a case, we have \(\nabla\cdot\bm{H}(\bm{r})\neq{0}\), but the expansion (Eq. [\[extensionAtFirst\]](#extensionAtFirst){reference-type="eqref" reference="extensionAtFirst"}) still produces a spurious spectrum. However, the basis set (Eq. [\[the-base\]](#the-base){reference-type="eqref" reference="the-base"}) is also insufficient because it has no longitudinal component of \(\boldsymbol{H} \negthinspace \left( \boldsymbol{r} \right)\), which should arise in this case. ## Chern number calculation The most sophisticated method for the calculation of the Chern number seems to be the one devised by Fukui, Hatsugai and Suzuki. Here, let us review the method briefly. In this method, the Chern number is derived on a discretized mesh of the Brillouin zone. An \(U \negthinspace \left( 1 \right)\) link variable is defined as \[U_\mu \negthinspace \left( \boldsymbol{k}_l \right) \equiv \frac{ \langle n \negthinspace \left( \boldsymbol{k}_l \right) | n \negthinspace \left( \boldsymbol{k}_l + \boldsymbol{\mathrm{e}}_\mu \right) \rangle} {\left| \langle n \negthinspace \left( \boldsymbol{k}_l \right) | n \negthinspace \left( \boldsymbol{k}_l + \boldsymbol{\mathrm{e}}_\mu \right) \rangle \right|} \label{link-variable},\] where \(\left| n \negthinspace \left( \boldsymbol{k}_l \right) \right>\) is the \(n\)th eigenvector obtained by diagonalizing an eigenequation, \(\boldsymbol{k}_l\) represents a lattice point in the discretized, Brillouin zone and \(\boldsymbol{\mathrm{e}}_\mu\) represents one lattice displacement in the direction \(\mu\left( = 1, 2\right)\). The \(\boldsymbol{k}_l\) is invariant under the displacement of periodic length \(\boldsymbol{k}_l + N_\mu \boldsymbol{\mathrm{e}}_\mu = \boldsymbol{k}_l\). From the link variable, a lattice field strength is taken as \[\begin{aligned} \tilde{F}_{12} \negthinspace \left( \boldsymbol{k}_l \right) &&\equiv \ln U_1 \negthickspace \left( \boldsymbol{k}_l \right) U_2 \negthickspace \left( \boldsymbol{k}_l \negthickspace + \negthickspace \boldsymbol{\mathrm{e}}_1 \right) U_1^{\scriptscriptstyle-1} \negthickspace \left( \boldsymbol{k}_l \negthickspace + \negthickspace \boldsymbol{\mathrm{e}}_2 \right) U_2^{\scriptscriptstyle-1} \negthickspace \left( \boldsymbol{k}_l \right) \notag \\ &&-\pi < \frac{1}{\mathrm{i}} \tilde{F}_{12} \negthinspace \left( \boldsymbol{k}_l \right) \le \pi. \label{field-strength} \end{aligned}\] Here the field strength is defined as the principal branch of the logarithm. From this field strength, the Chern number is computed as \[\tilde{c} = \frac{1}{2 \pi \mathrm{i} } \sum_l \tilde{F}_{12} \negthinspace \left(\boldsymbol{k}_l \right),\] where the sum is taken over all the lattice points in Brillouin zone. It is straightforward to extend this method to multiband cases. Equation [\[field-strength\]](#field-strength){reference-type="eqref" reference="field-strength"} naturally reflects gauge invariance. The result is always given as an integer because of the periodicity of the parameters (\(| n\negthinspace \left(\boldsymbol{k}_l + N_\mu \boldsymbol{\mathrm{e}}_\mu \right) \rangle = | n\negthinspace \left(\boldsymbol{k}_l \right) \rangle\)) and it can be seen by fixing the gauge over the entire lattice points. A gauge potential is defined as \[\widetilde{A}_\mu \negthinspace \left(\boldsymbol{k}_l\right) = \ln U_\mu \negthinspace \left( \boldsymbol{k}_l\right), \quad-\pi < \frac{1}{\mathrm{i}} \widetilde{A}_\mu \negthinspace \left( \boldsymbol{k}_l \right) \le \pi,\] which is periodic on the lattice \(\tilde{A}_\mu \negthinspace \left( \boldsymbol{k}_l + N_\mu \boldsymbol{\mathrm{e}}_\mu \right) = \tilde{A}_\mu \negthinspace \left( \boldsymbol{k}_l \right)\). The field strength (Eq. [\[field-strength\]](#field-strength){reference-type="eqref" reference="field-strength"}) is related to this gauge potential by \[\tilde{F}_{12} \negthinspace \left(\boldsymbol{k}_l \right) = \triangle_1 \widetilde{A}_2-\triangle_2 \widetilde{A}_1 + 2 \pi \mathrm{i} n_{12} \negthinspace \left( \boldsymbol{k}_l \right), \label{field-strength-gauge-potential}\] where \(\triangle_\mu\) represents a forward difference operation \(\triangle_\mu \widetilde{\boldsymbol{A}} \negthinspace \left( \boldsymbol{k}_l \right) = \widetilde{\boldsymbol{A}} \negthinspace \left( \boldsymbol{k}_l + \boldsymbol{e}_\mu \right)-\widetilde{\boldsymbol{A}} \negthinspace \left( \boldsymbol{k}_l \right)\) and \(n_{12} \negthinspace \left( \boldsymbol{k}_l \right)\) is an integer-valued field that makes \(-\pi < \tilde{F}_{12}/\mathrm{i} \leq \pi\). Taking the sum of Eq. [\[field-strength-gauge-potential\]](#field-strength-gauge-potential){reference-type="eqref" reference="field-strength-gauge-potential"}, the first and second terms of the right hand side cancel between adjacent links. Then the total sum of those terms vanishes because of the periodicity, and the result is given by \[\tilde{c} = \sum_l n_{12} \negthinspace \left( \boldsymbol{k}_l \right).\] For the Gaussian(spatially localized) basis sets, \(\left| n \negthinspace \left(\boldsymbol{k} \right) \right>\) is given by the eigenvector of the matrix eigenequation (Eq. [\[eigenequation-matrix\]](#eigenequation-matrix){reference-type="eqref" reference="eigenequation-matrix"}). Since Eq. [\[eigenequation-matrix\]](#eigenequation-matrix){reference-type="eqref" reference="eigenequation-matrix"} is invariant under the displacement \(\boldsymbol{k} \rightarrow \boldsymbol{k} + \boldsymbol{G}\), it is guaranteed that the computed Chern number is an integer. # Test for the method ## Empty lattice In this subsection, we give appropriate settings for the Gaussian basis sets to produce sufficiently accurate results for empty lattices in 2D and 3D cases. Further, using the determined setting, we calculate band structures of typical photonic crystals of both 2D and 3D as examples. Let us begin with the 2D empty lattice, in which the equation to be solved becomes a scalar equation. The results for a square lattice are shown in Fig. [\[empty-lattice\]](#empty-lattice){reference-type="ref" reference="empty-lattice"}(a) for two different settings represented as solid and broken lines. For both of the settings, we use \(64\) basis elements on regularly aligned grid points, i.e., eight basis elements per each direction. The broken line is obtained by setting the localization factor in the \(\mathrm{e}^{-r^2 / \alpha^2}\) to \(\alpha = 0.87 d\) (\(d\) being the spacing between two gird points) and the solid line is obtained by setting it to \(\alpha = d\). A good convergence for these settings requires matrix elements up to the fifth-nearest-neighbor pairs of the grid points. The necessary furthest pairs for matrix elements are determined by the overlap length of the matrix \(\Theta\) since the convergence of the matrix elements of \(\Theta\) is slower than those of \(O\). Within the shown frequency scale, the solid and broken lines overlap well and the solid lines match perfectly with the analytically derived dispersion, namely, \(\omega = c k\) folded in the first Brillouin zone. However, at the \(\Gamma\)-point, the lowest band does not converge to \(\omega \rightarrow 0\), in \(k\rightarrow 0\) limit. This mismatch comes from the fact that the Gaussian basis elements should have sufficiently large overlaps between them in order to compose a spatially homogeneous state, which is expected to be realized in the \(\omega\rightarrow 0\) and \(k\rightarrow 0\) limit. On the other hand, the accuracy for higher frequency modes is improved by increasing the number of basis elements since higher modes vary more rapidly in space. The failure in the \(\omega\rightarrow{0}\) limit is usually irrelevant in the discussion of topological states, since we usually focus on photonic gaps at finite \(\omega\) to see topological phenomena. For the 3D case, the vector equations should be treated and the basis set \(\left\{ | \boldsymbol{\tilde{g}}_{ik}^{\lambda} \rangle \right\}\) is used. Also in this case, the exact band structure of the empty lattice is well restored by using \(1,536\) basis elements, i.e., 512 grid points with eight grid points per each direction, and three basis elements \(\left( \lambda = 1, 2, 3\right)\) on each grid point. The results are shown in Fig. [\[empty-lattice\]](#empty-lattice){reference-type="ref" reference="empty-lattice"}(b) for two different settings. The broken line is obtained by setting the localization factor in the \(| \boldsymbol{\tilde{g}}_{ik}^{\lambda} \rangle\)\[Eq. [\[the-base\]](#the-base){reference-type="eqref" reference="the-base"}\] to \(\alpha = d\) and the solid line is obtained by setting it to \(\alpha = 1.07 d\). However, since the \(|\boldsymbol{\tilde{g}}_{ik}^{\lambda} \rangle\) is the Gaussian function multiplied by the first-order polynomial, it extends in a broader region than the genuine Gaussian function does. The difference of the first eigenvalues between \(\alpha = d\) and \(\alpha = 1.07 d\) near \(\Gamma\)-point is more prominent than the 2D case. We have confirmed that the results are well converged and match perfectly to the dispersion, \(\omega = c k\), folded in the first Brillouin zone in the shown frequency scale when we set \(\alpha\) more than \(1.07 d\), which takes the matrix elements up to the \(8\)th nearest neighbor points. Yet the lowest eigenmode at \(\Gamma\) point is never restored with the basis set \(\{|\bm{\tilde{g}}_{ik}^{ \lambda}\rangle\}\), due to the difficulty in representing a uniform field by \(|\bm{\tilde{g}}_{ik}^{\lambda}\rangle\), since \(|\bm{\tilde{g}}_{ij}^{ \lambda}\rangle\) is an odd function with respect to its center. ## The case of spatially modulated permittivity To begin with, we choose an array of dielectric columns arranged on a square lattice as a sample system for the 2D Gaussian basis set. We assume the radius of the column to be \(0.378 a\) with \(a\) representing the lattice spacing, and the relative permittivity to be \(\varepsilon = 8.9\) within the columns (\(\varepsilon = 1.0\) out of the columns). The calculated dispersions for TE modes (solid lines) and TM modes (dashed lines) are shown in Fig. [\[square-columns\]](#square-columns){reference-type="ref" reference="square-columns"}. The calculation is done with the same setting as the empty lattice \[the solid line in Fig. [\[empty-lattice\]](#empty-lattice){reference-type="ref" reference="empty-lattice"}(a)\]. The numerical integration between Gaussian basis elements converges by fining the integral mesh up to \(\alpha / 5\). The precision of the results increases with the number of basis elements in one direction \(n_x\). In this system, the most of eigenvalues converge in the shown scale before \(n_x = 8\), with which the result in Fig. [\[square-columns\]](#square-columns){reference-type="ref" reference="square-columns"} is obtained. In this case, the radius of the column \(r = 0.378 a\) is relatively large compared to the size of the unit cell. Therefore the structure is smooth and well restored with the integral mesh \(\alpha / 5\) for \(n_x = 8\). However, when the radius of the column is smaller, the basis set has to detect a finer structure in integration and we need a larger number of basis elements, which implies slower convergence. Another potential source of bad convergence is a sharp structure such as a corner of a rectangular rod, and we should be careful when such a structure is treated in the present method. Next, we choose a simple scaffold structure, namely a three-dimensional square rod structure studied by Sözüer and Haus, as an example system for the 3D Gaussian basis set. We assume the width of the rod as \(0.26 a\) and the relative permittivity \(\varepsilon = 13.0\), which is the same situation as in the study of Dobson *et al.*. The result shown in Fig. [\[cubic-scaffold\]](#cubic-scaffold){reference-type="ref" reference="cubic-scaffold"} is obtained with the same setting as in the empty lattice case \[the solid line in Fig. [\[empty-lattice\]](#empty-lattice){reference-type="ref" reference="empty-lattice"}(b)\]. As known, there exists full gap between the fifth and sixth bands. As in the case of the empty lattice, there is difficulty in the \(\omega\rightarrow 0\) and \(k\rightarrow 0\) limit, however, except that point, the global profile of the calculated band structure, such as the existence of the full gap, is consistent with that of the previous works. # Section Chern numbers and edge states in a 3D photonic crystal In this section, we demonstrate the topologically protected edge states of EM waves associated with the finite section Chern number in a 3D photonic crystal with broken SIS. In a 3D periodic system, the wave vector has three components, \(k_1\), \(k_2\), and \(k_3\), and the section Chern number is defined using two of them, regarding the remaining one as a free parameter. If we fix \(k_3\), the section Chern number \(C_{n} (k_3)\) is defined as \[\begin{aligned} C_{n} (k_3) &= \frac{1}{2\pi \mathrm{i}} \int \mathrm{d}k_1 \mathrm{d}k_2 \thickspace B_{n,12} (\bm{k}), \\ B_{n,12} (\bm{k}) &= \partial_{k_1} \mathcal{A}_{n,2} \left( \bm{k} \right)- \partial_{k_2} \mathcal{A}_{n,1} \left( \bm{k} \right), \label{secChern_def} \\ \mathcal{A}_{n,i} (\bm{k}) &= \langle \psi_n (\bm{k}) | \partial_{{k}_i} | \psi_n (\bm{k}) \rangle, \end{aligned}\] and those defined by fixing \(k_1\) or \(k_2\) are similarly given by permutations of the components \((k_1,k_2,k_3)\rightarrow (k_2,k_3,k_1) \rightarrow (k_3,k_1,k_2)\). In order to make \(C_{n} (k_3)\) well-defined, the gap should remain finite over the entire \(k_3\) constant plane. In general, the section Chern number takes any value, but the symmetry of a given system induces restriction. If a system has TRS, the eigenstate for \(\bm{k}\) is related to that for \(-\bm{k}\) by the bosonic time-reversal operation \(\mathcal{T}\), as \(\mathcal{T}|\psi_n(\bm{k})\rangle = |\psi_n(-\bm{k})\rangle^*\), which leads to \(\bm{B}_n (-\bm{k})=-\bm{B}_n (\bm{k})\). Then, we have \(C_{n} (k_3)=0\) for \(k_3=0\) and \(\pi\), because of \(B_n (-k_1,-k_2,0)=-B_n (k_1,k_2,0)\) and \(B_n (-k_1,-k_2,-\pi(=\negthickspace \pi))=-B_n (k_1,k_2,\pi)\), and the definition Eq. [\[secChern_def\]](#secChern_def){reference-type="eqref" reference="secChern_def"}. If there is SIS in addition to TRS, an additional restriction may arise, that is, SIS gives \(|\psi_n(-\bm{k})\rangle=|\psi_n(\bm{k})\rangle\), which indicates \(\bm{B}_n (-\bm{k})=\bm{B}_n (\bm{k})\), and if this relation is combined with \(\bm{B}_n (-\bm{k})=-\bm{B}_n (\bm{k})\) required by TRS, we finally obtain \(\bm{B}_n (\bm{k})=-\bm{B}_n (\bm{k})=0\). Therefore, SIS has to be broken for a nonzero section Chern number if the system has TRS. Owing to the topological nature of \(C_{n} (k_3)\), when \(C_{n} (k_3)\) changes as a function of \(k_3\), there should be a gap closing point somewhere on the corresponding \(k_3\) constant plane. On the other hand, as we have noted, the section Chern number is always zero at \(k_3=0\) and \(\pi\) with TRS. These facts indicate that the finite section Chern number requires existence of a gap closing point. Typically, gap closing occurs on isolated points in the Brillouin zone and the linear dispersion appears around those degeneracy points. A degeneracy point with linear dispersion is named as a Weyl point. Note that \(k_z\) appearing in the section Chern number \(C_{n} \left(k_z\right)\) is related to the fixed \(k_z\) of an incident wave on the photonic crystal. ## Honeycomb array of air hole columns Now, we consider to realize Weyl points in photonic crystals. In specific, we need a system in which the Weyl point related physics is readily accessible. More specifically, in order to observe chiral edge modes associated with the finite section Chern number \(C_{n}(k_3)\), it is required to have a clear gap on the entire two-dimensional plane with fixed \(k_3\). Since a clear gap is required only on a fixed \(k_3\) plane, a possible strategy to achieve our goal is to start with a two-dimensional system having "pseudo-gap," a frequency region filled with only a few bands, and then to apply appropriate three-dimensional modifications to the system. Because we are handling photonic crystals, the band structure can be modified by controlling the background medium. However, our experience tells us that careless modifications prone to fail, namely, even if Weyl points are successively generated, they are often masked by the other dispersive bands. Hence, the careful design is important and this is what we discuss in the following. In order to realize easily accessible Weyl points, we consider a photonic crystal of hollow columns aligned on the honeycomb lattice as depicted in Fig. [\[hollow-columns\]](#hollow-columns){reference-type="ref" reference="hollow-columns"}. The band structure of this system has some similarity to that of the triangular lattice consisting of dielectric columns. This is because the complementary region of the hollow honeycomb lattice composes a triangular lattice. Despite this similarity, we choose to use the hollow honeycomb lattice since it turns out that the existence of the sublattice structure of the honeycomb lattice is advantageous in the following discussion. Without three-dimensional modulation, the second and third TE modes of the hollow honeycomb lattice form 2D Dirac cones at K and K' points, and TM modes do not mask these TE Dirac cones. Owing to SIS of the hollow honeycomb lattice, these TE Dirac cones are actually line degeneracies extending in the \(k_z\) direction in the three-dimensional Brillouin zone. Then, SIS should be broken to transform the line degeneracy into Weyl points, and therefore, we reshape each hollow column into a hollow helix. Note that it is possible to modulate columns differently on sublattices A and B of the honeycomb lattice. ## Weyl points and change in section Chern number Now, we consider to apply the same twist on the both hollows at sublattices A and B (Fig. [\[bulk-equal\]](#bulk-equal){reference-type="ref" reference="bulk-equal"}). By transforming columns into helices, the SIS is broken and the line degeneracy is lifted except K and K' points, which implies emergence of the Weyl points. This emergence is explained in terms of the point group symmetry as follows. Before the twist is applied, the system has \(D_{6h}\) symmetry, and the group of \(k\) is \(C_{3v}\) on the line parallel to the \(k_z\) axis passing through K and K' points, whereas it is \(D_{3h}\) for K and K' points. Since both of \(C_{3v}\) and \(D_{3h}\) contain two-dimensional representations, twofold degeneracy is allowed on the whole line, which supports existence of the line degeneracy without twist. Then, after applying the twist, the reflection symmetry with respect to a plane including the \(z\) axis is broken and the group of \(k\) on the line parallel to the \(k_z\) axis passing through K and K' points turns to \(C_3\), except K and K' points. This implies the line degeneracy dissolves except K and K' points since \(C_3\) has no multidimensional representation. For K and K' points, the group of \(k\) changes from \(D_{3h}\) to \(D_3\). By the compatibility relation, 2D representations of \(D_{3h}\) are connected to 2D representations of \(D_3\), and therefore, the degeneracies at K and K' points survive even with the twist. The surviving degeneracy points are two Weyl points with the same chirality since these points are related by two-fold rotation around the \(z\) axis. Then, the section Chern number \(C_{n}(k_z)\) changes by \(\pm 2\) when \(k_z\) crosses \(k_z=0\), because of *two* Weyl points on the \(k_z=0\) plane \[Figs. [\[bulk-equal\]](#bulk-equal){reference-type="ref" reference="bulk-equal"}(b) and [\[bulk-equal\]](#bulk-equal){reference-type="ref" reference="bulk-equal"}(c)\]. The exactly same argument also applies for the \(k_z=\pi\) plane, but it is not possible to resolve Weyl points on the \(k_z=\pi\) plane, since the bands on which we focus merge into the other bands as we increase \(k_z\) from \(0\) to \(\pi\). Note that the observation of \(\pm 2\) jump in \(C_{n}(k_z)\) at \(k_z=0\) does not contradict with the former statement that \(C_{n}(0)\) should be \(0\), since at that time we had assumed that the gap is finite on entire \(k_z=0\) plane, and this assumption is invalid in this case. Next, we consider to change the twist of helix only for the hollow on the sublattice B \[Fig. [\[bulk-differentr\]](#bulk-differentr){reference-type="ref" reference="bulk-differentr"}(a)\]. With this modification, the twofold rotational symmetry around the \(k_z\) axis and the axis perpendicular to \(k_z\) axis are broken and then, there is no reason to have degeneracy on the \(k_z=0\) plane. However, the Weyl points found in Fig. [\[bulk-equal\]](#bulk-equal){reference-type="ref" reference="bulk-equal"}(c) do not simply disappear because Weyl points are topologically stable and only disappear by annihilating as a pair of Weyl points with opposite chiralities. Our numerical calculation shows that the Weyl points move to \(\pm k_z\) directions as we increase the difference of the radius of the helices. When one of the Weyl points moves in \(+k_z\) direction, the other one moves in \(-k_z\) direction due to TRS. In this case, the dispersion is fully gapped on the \(k_z=0\) plane, and we have \(C_{n}(0)=0\), which is consistent with the former statement. The section Chern number becomes finite for large enough \(k_z\) as expected from the existence of the Weyl points. ## Edge modes In order to investigate edge modes, we perform calculations using the developed Gaussian basis elements on the photonic crystal truncated in the \(\vec{a}_1\) direction \[see Fig. [\[edge-mode-same\]](#edge-mode-same){reference-type="ref" reference="edge-mode-same"}(b) for the definition of \(\vec{a}_1\)\] with finite width (eight unit cells in specific). At the interface, we place a material with a smaller dielectric constant so as to prevent the light from evading to outside. This is necessary because the momentum/frequency region we are focusing on lies above the light cone (for the same dielectric constant with the hollow region), which means that the light is able to escape from the system. In Figs. [\[edge-mode-same\]](#edge-mode-same){reference-type="ref" reference="edge-mode-same"} and [\[edge-mode-differentr\]](#edge-mode-differentr){reference-type="ref" reference="edge-mode-differentr"}, the dispersion relations of the finite width system are compared with the bulk dispersions projected on the surface for several values of \(k_z\). We find several bands apart from the bulk contribution, which signals the existence of edge modes. In fact, it is confirmed that the non-bulk bands are localized at the interfaces by examining the eigenvector for each mode. The eigenvectors also tell us on which side, left or right, each mode localizes. In Figs. [\[edge-mode-same\]](#edge-mode-same){reference-type="ref" reference="edge-mode-same"} and [\[edge-mode-differentr\]](#edge-mode-differentr){reference-type="ref" reference="edge-mode-differentr"}, the red color corresponds to the states on the left edge, while the green color to the states on the right edge. The color spectrum is determined by the moment of the eigenfield \(\langle r_1 \rangle_{\scriptscriptstyle \boldsymbol{H}_{n,k}} = \int \mathrm{d}^3 r \thickspace r_1 \vec{H}_{n,k}^2 \negthinspace \left(\boldsymbol{r}\right)\). As we change \(k_z\), the way that the edge modes connect bulk bands changes in accordance with the change of the sum \(\sum_{n=1}^9 C_{n}(k_z)\), which confirms the bulk-edge correspondence. The same argument also applies to gaps other than the 9th gap with nonzero total section Chern number. When the fixed momentum is conserved for \(k_z\) with \(\sum_{n=1}^{i} C_{n} \negthinspace (k_z)\) being nonzero, the wave propagates unidirectionally along the interface. This photonic crystal has the structure similar to the honeycomb array of twisted waveguides studied by Rechtsman *et al.*, where evanescently coupled modes are considered. In the system of evanescently coupled waveguides, the \(z\) axis, which is the propagation direction of a waveguide mode, is regarded as the *temporal* axis and the system is spatially two-dimensional. There, the twist of a waveguide becomes a temporal periodic modulation for waveguide modes and the appearance of the chiral edge modes is attributed to the Chern number for the *Floquet* band structure. On the other hand, we consider, in this paper, TE-like modes in the spatially three-dimensional photonic crystal and relate the unidirectional propagation of them to the finite value of section Chern numbers. Depending on the dielectric constant of the material, there might exist some wave guide modes near the frequency of the Weyl point and it leads to coupling between edge modes and wave guide modes that extend into bulk. In such a case, it is possible to suppress the mixing between the edge modes and the wave guide modes by shifting the relative positions of them in the frequency space. The relative shift can be achieved by changing the pitch of the helix while keeping its twist unchanged because the dispersion of wave guide modes is easily affected by the periodic length in the \(z\) direction, whereas that of TE-like modes is not affected so much. # Summary We proposed the Gaussian basis sets for the calculation of EM fields in 2D and 3D photonic crystals. In the formulation with spatially localized basis elements, the wave number appears as a boundary condition in the eigenequation and the eigenequation strictly becomes periodic in the wave number, which is advantageous for the Chern number calculation. In addition, the localized property of the Gaussian basis element becomes effective for the consideration of finite size or interface effects, which usually requires larger systems, by utilizing iterative algorithms. Besides, in three-dimensional cases, the Gaussian basis element can easily be accommodated to the divergence-free constraint of the Maxwell equations due to its simplicity for differentiation operation. For further improvement, it would be expected to optimize the mesh alignment or the localization factor of the Gaussian basis elements for each structure. We demonstrated the bulk-edge relation between the section Chern number and chiral edge modes in the SIS broken 3D photonic crystal with TRS. It was confirmed that chiral edge modes in the \(n\)th gap at each \(k_z\) reflects the total of the section Chern number below the \(n\)th gap \(C_{n} \left(k_z\right)\). A system with a finite section Chern number is expected to lead to some applications for wave-packet dynamics. For simplicity, we take a situation that the Weyl points exist on the \(k_z=0\) plane. When a wave packet is injected into the finite width system of the SIS broken 3D photonic crystal, a wave packet composed of the Bloch states with positive \(k_z\) does propagate on the one side, but does not on the other side. If we use the Bloch states with negative \(k_z\), the side that allows propagation changes. In this way, the system is useful to filter wave packets with fixed sign of \(k_z\).
{'timestamp': '2016-09-15T02:07:27', 'yymm': '1607', 'arxiv_id': '1607.04714', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04714'}
# Introduction The DC conductivity formula obtained by the Fermi-liquid theory is shown to be equivalent to the one obtained by the linearized Boltzmann equation in the textbook . This proof is done for the case of spherical Fermi surface. Its extension to the case of anisotropic Fermi surface is done by Maebashi and Fukuyama . The Maebashi-Fukuyama formula is formal but exact. On the other hand, Taylor  derived a formal but exact solution of the linearized Boltzmann equation. These two works employ opposite strategies. The Maebashi and Fukuyama formula is on the basis of the Fermi-surface harmonics which is the polynomial of the velocity. Then the interaction is expanded in terms of the velocity. The Taylor formula is on the basis of the scattering eigenfunction. In this formula the interaction is diagonalized but the velocity is expanded in terms of the scattering eigenfunction. By these two formulae we know the exact forms of the DC conductivity of anisotropic Fermi liquids. Consequently it becomes evident that there is no room for the current-vertex-correction discussed in the fluctuation-exchange (FLEX) approximation . # Maebashi-Fukuyama formula We start from the Fermi-liquid formula  for the DC conductivity \[\sigma_{xx} = 2 e^2 \sum_{\bf p} \sum_{\bf p'} v_x({\bf p}) \big(K''\big)^{-1}_{{\bf p}{\bf p'}} \ v_x({\bf p'}) \Big(-{ \partial f_{\bf p} \over \partial \epsilon_{\bf p} } \Big), \label{FL}\] which is (2.34) of and equivalent to the result of the linearized Boltzmann equation, (58) of. Using the Fermi-surface harmonics , \(\psi_L({\bf p})\), ([\[FL\]](#FL){reference-type="ref" reference="FL"}) is written as \[\sigma_{xx} = 2 e^2 \sum_{\bf p} \sum_{\bf p'} \psi_1({\bf p}) \sum_L \sum_{L'} \big(K''\big)^{-1}_{LL'} \psi_L({\bf p}) \psi_{L'}({\bf p'}) \psi_1({\bf p'}) \langle v_x^2 \rangle \Big(-{ \partial f_{\bf p} \over \partial \epsilon_{\bf p} } \Big), \label{anisoFL}\] where \(\psi_1({\bf p}) = v_x({\bf p}) / \sqrt{\langle v_x^2 \rangle}\) and \[\big(K''\big)^{-1}_{{\bf p}{\bf p'}} = \sum_L \sum_{L'} \big(K''\big)^{-1}_{LL'} \psi_L({\bf p}) \psi_{L'}({\bf p'}).\] The normalization factor is defined as \[\langle v_x^2 \rangle = {1 \over N(\epsilon)} \sum_{\bf p} v_x({\bf p})^2 \ \delta\big( \epsilon-\epsilon_{\bf p} \big),\] with \[N(\epsilon) = \sum_{\bf p} \delta\big( \epsilon-\epsilon_{\bf p} \big).\] By the orthonormal condition \[{1 \over N(\epsilon)} \sum_{\bf p} \psi_L({\bf p}) \psi_{L'}({\bf p}) \delta\big( \epsilon-\epsilon_{\bf p} \big) = \delta_{LL'}, \label{orthonormal}\] ([\[anisoFL\]](#anisoFL){reference-type="ref" reference="anisoFL"}) is written as \[\sigma_{xx} = 2 e^2 \int d\epsilon N(\epsilon) \int d\epsilon' N(\epsilon') \big(K''\big)^{-1}_{11} \langle v_x^2 \rangle \Big(-{ \partial f \over \partial \epsilon } \Big).\] Here we have used the relation \[\sum_{\bf p} = \int d\epsilon \sum_{\bf p} \delta\big( \epsilon-\epsilon_{\bf p} \big).\] Introducing the memory-function matrix \[\big(M''\big)^{-1}_{11} = \int d\epsilon' N(\epsilon') \big(K''\big)^{-1}_{11},\] we obtain \[\sigma_{xx} = 2 e^2 \int d\epsilon N(\epsilon) \big(M''\big)^{-1}_{11} \langle v_x^2 \rangle \Big(-{ \partial f \over \partial \epsilon } \Big). \label{M}\] Finally ([\[M\]](#M){reference-type="ref" reference="M"}) is written as \[\sigma_{xx} = 2 e^2 \sum_{\bf p} v_x({\bf p}) \big(M''\big)^{-1}_{11} v_x({\bf p}) \Big(-{ \partial f_{\bf p} \over \partial \epsilon_{\bf p} } \Big).\] In conclusion the DC conductivity is determined by the memory-function matrix . The microscopic calculation of the memory-function matrix based on the Fermi-liquid theory is carried out by Maebashi and Fukuyama . To obtain a finite DC conductivity we have to take the contribution of the Umklapp process into the calculation of \(M''\). Next we have to perform the matrix inversion from \(M''\) to \((M'')^{-1}\). In the previous subsection it is clarified that the DC conductivity is determined by the memory-function matrix. In the Boltzmann equation it appears in the collision term. Thus we investigate the collision term after Allen  to obtain the feeling of the matrix structure. The collision term \(I_{\bf p}\) for \(n_{\bf p}\) is given by \[I_{\bf p} = \sum_{\bf p'} K''_{{\bf p}{\bf p'}} \Big\{ n_{\bf p'}-n_{\bf p} \Big\}, \label{C_p}\] as (47) of. Using the Fermi-surface harmonics ([\[C_p\]](#C_p){reference-type="ref" reference="C_p"}) is written as \[I_{\bf p} = \sum_{\bf p'} \sum_L \sum_{L'} K''_{LL'} \psi_L({\bf p}) \psi_{L'}({\bf p'}) \sum_{L''} n_{L''} \Big\{ \psi_{L''}({\bf p'})-\psi_{L''}({\bf p}) \Big\}. \label{C_L}\] By the orthonormal condition ([\[orthonormal\]](#orthonormal){reference-type="ref" reference="orthonormal"}) and \(\psi_0({\bf p}) = 1\), ([\[C_L\]](#C_L){reference-type="ref" reference="C_L"}) is written as \[I_{\bf p} = \int d\epsilon' N(\epsilon') \sum_L \sum_{L'} \sum_{L''} K''_{LL'} \ n_{L''} \ \psi_L({\bf p}) \Big\{ \delta_{L'L''}-\delta_{L'0} \psi_{L''}({\bf p}) \Big\}. \label{C_L'}\] Introducing the memory-function matrix \[M''_{LL'} = \int d\epsilon' N(\epsilon') K''_{LL'},\] ([\[C_L\'\]](#C_L'){reference-type="ref" reference="C_L'"}) is written as \[I_{\bf p} = \sum_L \psi_L({\bf p}) \Big\{ \sum_{L'} M''_{LL'} \ n_{L'}- \sum_{L''}\sum_{L'} C_{L''L'L} \ M''_{L''0} \ n_{L'} \Big\}, \label{C_L''}\] where \(C_{LL'L''}\) is the Clebsch-Gordan coefficient defined by \[\psi_{L}({\bf p}) \psi_{L'}({\bf p}) = \sum_{L''} C_{LL'L''} \ \psi_{L''}({\bf p}). \label{CG}\] Thus we obtain the collision term \(I_L\) for \(n_L\) as \[I_L = \sum_{L'} M''_{LL'} \ n_{L'}- \sum_{L''}\sum_{L'} C_{L''L'L} \ M''_{L''0} \ n_{L'}.\] For anisotropic Fermi surfaces \(n_L\) couples to the other \(n_{L'}\)s via the memory-function matrix so that we have to perform the matrix inversion, \(M'' \rightarrow (M'')^{-1}\), to obtain the conductivity. The case of the spherical Fermi surface is special and we do not have to perform the matrix inversion. By the symmetry \(K''_{{\bf p}{\bf p'}}\) only depends on the angle between \({\bf p}\) and \({\bf p'}\) so that the memory-function matrix becomes a diagonal matrix. Since \(C_{0L'L''} = \delta_{L'L''}\) which is obtained from ([\[CG\]](#CG){reference-type="ref" reference="CG"}) by using \(\psi_0({\bf p}) = 1\), we obtain \[I_L = 4\pi \Big\{ w_L-w_0 \Big\} n_L,\] where we put \(M''_{LL'} = 4\pi w_L \delta_{LL'}\). This is the result shown in. The DC conductivity is determined by \(n_1\) and the transport life-time \(\tau\) is given by \[{1 \over \tau} = 4\pi \Big\{ w_0-w_1 \Big\}. \label{Peierls}\] The use of the vector mean free path \({\bf \Lambda}\) is convenient  for the discussion of the symmetry property of the conductivity tensor. In the following all the vectors are defined on the Fermi surface. The conductivity tensor \({\boldsymbol \sigma}\) is expressed in the dyadic form as  \[{\boldsymbol \sigma} = 2 e^2 \sum_{\bf p} {\bf v}_{\bf p} \ {\bf \Lambda}_{\bf p} \Big(-{ \partial f_{\bf p} \over \partial \epsilon_{\bf p} } \Big), \label{dyadic}\] where \[{\bf \Lambda}_{\bf p} = \sum_{\bf p'} \big(K''\big)^{-1}_{{\bf p}{\bf p'}} \ {\bf v}_{\bf p'}, \label{Lambda}\] with \(\displaystyle {\bf v}_{\bf p} = \big( v_x({\bf p}), v_y({\bf p}), v_z({\bf p}) \big)\). For systems with cubic symmetry the conductivity tensor reduces to a scalar \(\sigma\) . This means the integrand in ([\[dyadic\]](#dyadic){reference-type="ref" reference="dyadic"}) is proportional to the scalar product \(\displaystyle {\bf v}_{\bf p} \cdot {\bf v}_{\bf p}\). Thus introducing the anisotropic life time \(\tau_{\bf p}\), the vector mean free path is written as \[{\bf \Lambda}_{\bf p} = \tau_{\bf p} \ {\bf v}_{\bf p}. \label{cubic}\] Namely, the direction of \({\bf \Lambda}_{\bf p}\) is the same as that of \({\bf v}_{\bf p}\). Substituting ([\[cubic\]](#cubic){reference-type="ref" reference="cubic"}) into ([\[Lambda\]](#Lambda){reference-type="ref" reference="Lambda"}) we obtain \[\tau_{\bf p} = \sum_{\bf p'} \big(K''\big)^{-1}_{{\bf p}{\bf p'}} \ { {\bf v}_{\bf p} \cdot {\bf v}_{\bf p'} \over {\bf v}_{\bf p} \cdot {\bf v}_{\bf p} }. \label{aniso-tau}\] This relation ([\[aniso-tau\]](#aniso-tau){reference-type="ref" reference="aniso-tau"}) is a mere rewriting and has no practical use. Using this anisotropic life time the conductivity is expressed as \[\sigma_{xx} = 2 e^2 \sum_{\bf p} v_x({\bf p}) \ \tau_{\bf p} \ v_x({\bf p}) \Big(-{ \partial f_{\bf p} \over \partial \epsilon_{\bf p} } \Big). \label{simple}\] The Fermi-liquid formula by Yamada and Yosida corresponds to this case. The appearance of ([\[simple\]](#simple){reference-type="ref" reference="simple"}) is simple but it is none other than ([\[FL\]](#FL){reference-type="ref" reference="FL"}). Actually \[\sigma_{xx} = 2 e^2 \sum_{\bf p} { {\bf v}_{\bf p} \cdot {\bf v}_{\bf p} \over 3 } \ \tau_{\bf p} \Big(-{ \partial f_{\bf p} \over \partial \epsilon_{\bf p} } \Big) = 2 e^2 \sum_{\bf p} \sum_{\bf p'} \big(K''\big)^{-1}_{{\bf p}{\bf p'}} \ { {\bf v}_{\bf p} \cdot {\bf v}_{\bf p'} \over 3 } \Big(-{ \partial f_{\bf p} \over \partial \epsilon_{\bf p} } \Big),\] in the case of cubic symmetry. For systems without cubic symmetry the proportional relation ([\[cubic\]](#cubic){reference-type="ref" reference="cubic"}) does not hold  so that we have to use the Maebashi-Fukuyama formula discussed above. For systems with spherical symmetry \(\tau_{\bf p}\) reduces to a constant \(\tau\). In this case the inverse relation of ([\[Lambda\]](#Lambda){reference-type="ref" reference="Lambda"}) \[{\bf v}_{\bf p} = \sum_{\bf p'} K''_{{\bf p}{\bf p'}} \ {\bf \Lambda}_{\bf p'},\] reduces to \[{\bf v}_{\bf p} = \tau \sum_{\bf p'} K''_{{\bf p}{\bf p'}} \ {\bf v}_{\bf p'}.\] Thus we obtain \[{1 \over \tau} = \sum_{\bf p'} K''_{{\bf p}{\bf p'}} \ \big( {\bf {\hat p}} \cdot {\bf {\hat p'}} \big), \label{life-time}\] with \({\bf v}_{\bf p} = {\bf p} / m\) and \({\bf {\hat p}} = {\bf p}/|{\bf p}|\). Since \[K''_{{\bf p}{\bf p'}} = {1 \over \tau_0} \delta_{{\bf p}{\bf p'}}-W_{{\bf p}{\bf p'}},\] and \[{1 \over \tau_0} = \sum_{\bf p'} W_{{\bf p}{\bf p'}},\] ([\[life-time\]](#life-time){reference-type="ref" reference="life-time"}) is written as \[{1 \over \tau} = \sum_{\bf p'} W_{{\bf p}{\bf p'}} \big( 1-\cos\theta \big), \label{constant}\] with \(\cos\theta = {\bf {\hat p}} \cdot {\bf {\hat p'}}\). This expression ([\[constant\]](#constant){reference-type="ref" reference="constant"}) is none other than ([\[Peierls\]](#Peierls){reference-type="ref" reference="Peierls"}). # Taylor formula Taylor  introduced the vector mean free path \({\bf \Lambda}\) and wrote down the DC conductivity tensor \({\boldsymbol \sigma}\) in the dyadic form as \[{\boldsymbol \sigma} = 2 e^2 \sum_{\bf p} {\bf v}({\bf p}) \ {\bf \Lambda}({\bf p}) \Big(-{ \partial f_{\bf p} \over \partial \epsilon_{\bf p} } \Big).\] In the following all the vectors are defined on the Fermi surface. The vector mean free path is determined by \[{\bf v}({\bf p}) = \sum_{\bf p'} Q({\bf p},{\bf p'}) \big[ {\bf \Lambda}({\bf p})-{\bf \Lambda}({\bf p'}) \big]. \label{vector}\] The Fermi-liquid theory works  on the calculation of the collision operator \(Q\). Hereafter we only consider the \(x\)-component of ([\[vector\]](#vector){reference-type="ref" reference="vector"}) \[v_x({\bf p}) = \sum_{\bf p'} Q({\bf p},{\bf p'}) \big[ \Lambda_x({\bf p})-\Lambda_x({\bf p'}) \big]. \label{x-com}\] Introducing the inner product \[(a,b) \equiv \sum_{\bf p} \Big(-{ \partial f_{\bf p} \over \partial \epsilon_{\bf p} } \Big) a({\bf p})b({\bf p}),\] the DC conductivity is compactly written as[^1] \[\sigma_{xx} = 2 e^2 \big( v_x, \Lambda_x \big). \label{DC}\] Loosely ([\[x-com\]](#x-com){reference-type="ref" reference="x-com"}) is written as \[v_x = \Big( {1 \over \tau}-{\hat Q} \Big) \Lambda_x, \label{matrix-x}\] where \[{\hat Q} \Lambda_x \equiv \sum_{\bf p'} Q({\bf p},{\bf p'}) \Lambda_x({\bf p'}),\] and \[{1 \over \tau} \Lambda_x \equiv \sum_{\bf p'} {1 \over \tau({\bf p})} \delta_{{\bf p}{\bf p'}} \Lambda_x({\bf p'}),\] with \[{1 \over \tau({\bf p})} \equiv \sum_{\bf p'} Q({\bf p},{\bf p'}).\] The inverse of ([\[matrix-x\]](#matrix-x){reference-type="ref" reference="matrix-x"}) is \[\Lambda_x = \Big( {1 \over \tau}-{\hat Q} \Big)^{-1} v_x = \tau v_x + \tau {\hat Q} \tau v_x + \tau {\hat Q} \tau {\hat Q} \tau v_x + \cdot\cdot\cdot. \label{inversion}\] The perturbative expansion for the DC conductivity is shown in Fig. 1. If we can find the scattering eigenfunction \(a_s({\bf p})\) which satisfies \[\tau({\bf p}) \sum_{\bf p'} Q({\bf p},{\bf p'}) a_s({\bf p'}) = \alpha_s a_s({\bf p}), \label{Q-alpha}\] with the eigenvalue \(\alpha_s\), the inversion ([\[inversion\]](#inversion){reference-type="ref" reference="inversion"}) is readily accomplished. To proceed we expand the velocity in terms of the scattering eigenfunction as \[\tau({\bf p}) v_x({\bf p}) = \sum_s \beta_s a_s({\bf p}). \label{v-beta}\] Employing ([\[Q-alpha\]](#Q-alpha){reference-type="ref" reference="Q-alpha"}) and ([\[v-beta\]](#v-beta){reference-type="ref" reference="v-beta"}) in ([\[inversion\]](#inversion){reference-type="ref" reference="inversion"}) we obtain[^2] \[\Lambda_x({\bf p}) = \sum_s \big[ 1 + \alpha_s + \alpha_s^2 + \cdot\cdot\cdot \big] \beta_s a_s({\bf p}), \label{Lambda-ab}\] To proceed further we have to establish the orthonormality of the scattering eigenfunction. First we rewrite ([\[Q-alpha\]](#Q-alpha){reference-type="ref" reference="Q-alpha"}) as \[\sum_{\bf p'} \sqrt{\tau({\bf p})} Q({\bf p},{\bf p'})\sqrt{\tau({\bf p'})}\ { a_s({\bf p'}) \over \sqrt{\tau({\bf p'})} } = \alpha_s \ { a_s({\bf p}) \over \sqrt{\tau({\bf p})} },\] Since \(\sqrt{\tau({\bf p})}Q({\bf p},{\bf p'})\sqrt{\tau({\bf p'})}\) is a Hermitian matrix[^3], the function \(a_s({\bf p}) / \sqrt{\tau({\bf p})}\) constitutes an orthogonal basis. Thus the orthonormal condition is written as \[\Big( {a_s \over \sqrt{\tau}}, {a_{s'} \over \sqrt{\tau}} \Big) = \Big( a_s, {a_{s'} \over \tau} \Big) = \delta_{ss'}. \label{orthonormal-a}\] Employing ([\[v-beta\]](#v-beta){reference-type="ref" reference="v-beta"}) and ([\[Lambda-ab\]](#Lambda-ab){reference-type="ref" reference="Lambda-ab"}) in ([\[DC\]](#DC){reference-type="ref" reference="DC"}) we obtain \[\sigma_{xx} = 2 e^2 \sum_{s'} \sum_s \beta_{s'} {1 \over 1-\alpha_s } \beta_s \Big( {a_{s'} \over \tau}, a_s \Big).\] Finally making use of the orthonormal condition ([\[orthonormal-a\]](#orthonormal-a){reference-type="ref" reference="orthonormal-a"}) we obtain the Taylor formula for the DC conductivity \[\sigma_{xx} = 2 e^2 \sum_s \beta_{s} {1 \over 1-\alpha_s } \beta_s.\] The repeated scatterings shown in Fig. 1 result in the renormalization of the life time \(\tau\). # Current-vertex-correction in FLEX approximation The purpose of this note is to clarify the exact expression for the DC conductivity of anisotropic Fermi liquids. Consequently it becomes evident that there is no room for the current-vertex-correction discussed in the FLEX approximation . Other deficiencies of the FLEX approximation have been also discussed previously . [^1]: See, for example, Ashcroft and Mermin: *Solid State Physics* (1976) Chap. 16-PROBLEM 4. [^2]: \[\tau {\hat Q} \tau v_x = \tau({\bf p}) \sum_{\bf p'} Q({\bf p},{\bf p'}) \sum_s \beta_s a_s({\bf p'}) = \sum_s \beta_s \alpha_s a_s({\bf p}), \nonumber\] and so on. [^3]: \(Q({\bf p},{\bf p'})\) itself is a Hermitian matrix.
{'timestamp': '2016-08-24T02:02:23', 'yymm': '1607', 'arxiv_id': '1607.04705', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04705'}
null
null
null
null
# Appendix: \(\< TT \>\) at 2-loops The holographic formula for the power spectrum reads, \[\label{power_sp_h} \Delta^2_{{\cal R}}(q) = -\frac{q^3}{4 \pi^2} \frac{1}{{\rm Im} \<\!\<T(q)T(-q)\>\!\>} ,\] where \(T=T^i_i\) is the trace of the energy momentum tensor and the double bracket notation indicates that the momentum conserving delta function (times \((2 \pi)^3\)) has been removed. The imaginary part in ([\[power_sp_h\]](#power_sp_h){reference-type="ref" reference="power_sp_h"}) is taken after analytic continuation \[\label{analytic} q \to-i q, N \to-i N,\] where \(q\) is the magnitude of the momentum. This formula was derived in using the domain-wall/cosmology correspondence and also follows from the wave-function of the universe approach. There is a similar holographic formula for the tensor power spectrum involving the transverse traceless part of the 2-point function of \(T_{ij}\). This class of theories we consider has the important property that if one promotes \(g_{YM}^2\) to a new field that transforms appropriately under conformal transformation, the theory becomes conformally invariant. We say that the theory has a "generalized conformal structure". This is not a bona fide symmetry of the theory but, nevertheless, controls many of its properties. The generalised conformal structure implies that the 2-point function of the energy momentum tensor, to leading order in the large \(N\) limit (planar diagrams), is given by \[\begin{aligned} \<\!\< T(q) T(-q) \>\!\> = q^3 N^2 f(\geff^2), \end{aligned}\] where \(\geff^2=\gYM^2 N/q\) is the effective dimensionless 't Hooft coupling constant and \(f(\geff^2)\) is function of \(\geff^2\). The overall factor of \(q^3\) reflects the fact that the energy momentum has dimension 3 in three dimensions and the overall factor of \(N^2\) is because this is the leading order term in the large \(N\) limit. Under the analytic continuation ([\[analytic\]](#analytic){reference-type="ref" reference="analytic"}), \[\geff^2(q) \to \geff^2(q), \quad N^2 q^3 \to-i N^2 q^3\] and therefore for this class of theories, \[\label{power_sp3} \Delta^2_{{\cal R}}(q) = \frac{1}{4 \pi^2 N^2 f(g_{eff}^2)},\] which is the formula we used in the main text. Perturbation theory is valid when \(g_{\rm eff}^2 \ll 1\). In the perturbative regime, the function \(f(\geff^2)\) is given by \[\label{app:2loop} f(\geff^2) = f_0 \left( 1-f_1 \, \geff^2 \ln \geff^2 + f_2 \, \geff^2 + O(\geff^4) \right).\] The function \(f_0\) is determined by a 1-loop computation, while \(f_1\) and \(f_2\) come from 2-loops. The presence of the logarithm is due to UV and IR divergences in the computation of the 2-point function of the energy momentum tensor. The 1-loop computation was done in and we summarize the 2-loop computation of here. Since this a gauge theory we first need to gauge fix. The gauged fixed action is \[\begin{aligned} S &=& \frac{1}{\gYM^2} \int d^3 x \, \tr \left[ \frac{1}{2} F_{ij} F^{ij} + (\partial^i A_i)^2 + 2 \partial^i \bar{c} \mD_i c \right. \nn \\ && + (\mD \Phi^M)^2 + 2 \bar{\psi}^L \mD \psi^L + 2 \sqrt{2} \, \mu_{M L_1 L_2} \Phi^M \bar{\psi}^{L_1} \psi^{L_2} \nn \\ && \left. + \frac{1}{6} \lambda_{M_1 M_2 M_3 M_4} \Phi^{M_1} \Phi^{M_2} \Phi^{M_3} \Phi^{M_4} \right]. \end{aligned}\] where \(b, c\) is the ghost sector. The energy-momentum tensor is obtained by coupling the theory to a background metric \(g_{ij}\) and then using \(T_{ij} = (2/\sqrt{g}) (\delta S/\delta g^{ij}) |_{g_{ij}=\delta_{ij}}\). This procedure defines a unique energy momentum tensor, except that there is a choice of how to couple the scalars to gravity. One may include the non-minimal coupling \(1/(2 \gYM^2) \sum_M The variation of this term contributes an ``improvement term'' to the energy momentum tensor. If\) The energy momentum obtained in this fashion is given by \[T_{ij} = T^A_{ij} + T^{g.f.}_{ij} + T^{gh}_{ij} + T^{\psi}_{ij} + T^{\Phi}_{ij} + T^{Y}_{ij},\] where the different terms denote, respectively, the contribution of the gauge fields, the gauge-fixing term, the ghost sector, the fermions, the scalars and the Yukawa interactions. These are explicitly given by \[\begin{aligned} \label{T} T^{A}_{ij} &=& \frac{1}{\gYM^2} \tr \left[ F_{ik} F_{jk}-\delta_{ij} \frac{1}{4} F_{kl} F_{kl} \right], \\ T^{g.f.}_{ij} &=& \frac{1}{ \gYM^2} \tr \left[ A_i \partial_j \left( \partial_k A_k\right) + A_j \partial_i \left( \partial_k A_k\right) \right. \nonumber \\ &&\left.-\delta_{ij} \left( A_k\ \partial_k \partial_l A_l + \frac{1}{2} (\partial_k A_k) (\partial_l A_l) \right)\right], \nn \\ T^{gh}_{ij} &=& \frac{1}{\gYM^2} \tr \left[ \partial_i \bar{c} \mD_j c + \partial_j \bar{c} \mD_i c-\delta_{ij} \partial_k \bar{c} \mD_k c\right], \nn \\ T^{\psi}_{ij} &=& \frac{1}{\gYM^2} \tr \left[ \frac{1}{2} \bar{\psi}^L \gamma_{(i} \oa{\mD}_{j)} \psi^L-\delta_{ij} \frac{1}{2} \bar{\psi}^L \gamma_k \oa{\mD}_k \psi^L \right] ,\nn \\ T^{\Phi}_{ij} &=& \frac{1}{\gYM^2} \tr \left[ \mD_i \Phi^M \mD_j \Phi^M-\delta_{ij} \left( \frac{1}{2} (\mD \Phi^M)^2 \right. \right. \nonumber \\ && \left. \qquad \qquad + \frac{1}{4!} \lambda_{M_1 M_2 M_3 M_4} \Phi^{M_1} \Phi^{M_2} \Phi^{M_3} \Phi^{M_4} \right) \nonumber \\ && \left. \qquad \qquad + \, T^{Y}_{ij} &=& \frac{1}{\gYM^2} \tr \left[-\delta_{ij} \, \, \mu_{M L_1 L_2} \Phi^M \bar{\psi}^{L_1} \psi^{L_2} \right]. \nn \end{aligned}\] The topology of the 2-loop diagrams that need to be computed is given in Fig. [\[Fig.TT2L\]](#Fig.TT2L){reference-type="ref" reference="Fig.TT2L"}. The 2-point function has a UV divergence that can be cancelled by the counterterm \[a_{CT} \int d^3 x \sqrt{g} R,\] where \(R\) is the curvature scalar of the background metric \(g_{ij}\), with appropriately chosen \(a_{CT}\). As usual, this process introduces a renormalization scale \(\mu\), which leads to scheme dependence. The correlator also has an IR divergence (unless all scalars are minimal), which we regulated with an IR cut-off, \(\mu_*\). This leads to the following result for \(\ln \beta\), \[\ln \beta=\ln \frac{1}{|g|}-\frac{a_0}{f_1}-\left(\ln \frac{q_*}{\mu} + \frac{64/\pi^2}{f_1 \mN_{(B)} }\Sigma_\Phi \ln \frac{\mu}{\mu_*}\right),\] where we have made use of the definition of \(g\), \(g q_* = f_1 \gYM^2 N\). The renormalization scale \(\mu\) is arbitrary and so is the pivot scale \(q_*\). We fix this scheme dependence by setting, \(\mu=q_*\). An alternative scheme is to set \(\mu\) to the (inverse of the) smallest scale in the data (i.e. equal to 0.17 Mpc\({}^{-1}\)). We have checked that the results we present in the main text are not sensitive to the choice of scheme, except possibly at very low \(l\)'s. As argued in the main text, this is precisely the regime where one should not trust the perturbative computation. Regarding the IR divergence now. It was argued in that superrenormalizable theories with a dimensionful coupling constant (as in our case) are non-perturbative IR finite, with \(\gYM^2\) providing the IR cut-off. We therefore set \(\mu_*= c \gYM^2\), where \(c\) is a number that can only be determined non-perturbatively. This leads to our final formula for \(\ln \beta\), \[\label{app:final_log} \ln \beta=\ln \frac{1}{|g|}-\frac{a_0}{f_1}-\frac{64/\pi^2}{ f_1 \mN_{(B)} }\Sigma_\Phi \ln \frac{N f_1}{c g}\] In the main text we set \(c=1\) but we also checked that the results do not change qualitatively if we change \(c\). An alternative way to deal with the IR issues is to get \(\mu_*\) equal to the (inverse of the) largest scale in the data (i.e. \(1.4 \times 10^{-4}\) Mpc\({}^{-1}\)). As in the case of scheme dependence, we have checked that the results are not very sensitive to how we treat \(\mu_*\), except possibly at very low \(l\)s.
{'timestamp': '2017-01-04T02:03:13', 'yymm': '1607', 'arxiv_id': '1607.04878', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04878'}
# Introduction The study of mixtures of Bose-Einstein condensates (BECs) has opened rich opportunities for novel experimental and theoretical investigations. Mixtures of ultracold atoms offer great flexibility thanks to the variety of atomic species and the additional degree of freedom related to the hyperfine structure (for a recent overview see ). For a weakly interacting mixture of two BECs, the ground state of the system can either be a miscible mixture of the two components or a phase separated configuration. Nevertheless, the stability of mixtures very close to the critical region is sensitive to other effects, such as asymmetries in the trapping potential. Moreover, for systems in which the intracomponent coupling constants do not exactly coincide, one of the two components will experience a positive buoyancy and will "float" on the other. Previous experiments involving two internal states of rubidium were affected by both of these problems hence setting strong limits to explore the many-body properties of miscible binary BECs. In particular, such conditions prevent the study of the static and dynamic response of an unpolarized system close to the transition between the miscible and immiscible phases, where interaction effects are particularly important despite the weakly interacting nature of the gas. Here we report on the first measurement of the spin-dipole (SD) polarizability of a two-component BEC, as well as the frequency of the SD oscillation, by using an ultracold mixture of the \(|3^2S_{1/2}, F=1, m_F=+1\rangle \equiv \ket{\uparrow}\) and \(|3^2S_{1/2}, F=1, m_F=-1\rangle \equiv \ket{\downarrow}\) states of atomic sodium. The polarizability characterizes in a fundamental way the thermodynamic behavior of binary ultracold gases and exhibits a divergent behavior at the transition between the miscible and immiscible phases, with the occurrence of important spin fluctuations. On the other hand, the SD oscillation is the simplest collective excitation supported by the system in the presence of harmonic trapping and is characterized by the motion of the two components with opposite phase around equilibrium. The SD oscillation is the analog of the famous giant dipole resonance of nuclear physics, where neutrons and protons oscillate with opposite phase. Actually, collective modes are a popular subject of research in quantum many body systems (see, *e.g.*, ) where experiments are able to determine the corresponding frequencies with high precision, providing a good testbed for detailed comparison with theory and an accurate determination of the relevant interaction parameters. Collective dynamics has been already investigated in quantum binary mixtures of atomic gases like repulsive gases of Fermi atoms, Bose-Bose and Bose-Fermi mixtures as well as Bose-Fermi superfluid mixtures. In the case of Bose-Bose mixtures both the polarization and the SD oscillation frequency are predicted to be crucially sensitive to the difference between the value of the intra and intercomponent interactions which is particularly small in our case. The dramatic change of the density profile of the trapped gas, caused by a small displacement of the minima of the trapping potentials of the two species near the miscible-immiscible phase transition, was first investigated theoretically in. Our mixture is not subject to buoyancy as \(g_{\uparrow\uparrow}=g_{\downarrow\downarrow}\equiv g\) and is on the miscible side \(g_{\uparrow\downarrow} < g\) near the boundary of the phase transition (\(g\) and \(g_{\uparrow\downarrow}\) are respectively the intra and intercomponent coupling constants). The fact that \((g-g_{\uparrow\downarrow}) / g \simeq 7 \, \%\), as given by the scattering lengths \(a_{\uparrow\uparrow}=a_{\downarrow\downarrow}=54.54(20) a_0\) and \(a_{\uparrow\downarrow}=50.78(40)a_0\), where \(a_0\) is the Bohr radius, ensures the stability of the mixture and, together with the absence of buoyancy, allows us to overcome the ultimate limits to measure both the polarizability and SD oscillation frequency. # Mixture preparation Our experiment is based on the apparatus introduced in and starts with a nearly pure BEC of \(^{23}\)Na atoms in the \(\ket{\downarrow}\) state in a crossed optical dipole trap with frequencies \(\left[ \omega_x,\omega_y,\omega_z \right] / 2 \pi = \left[47.7(2),207.2(3),156.8(2) \right] \, \text{Hz}\). The magnetic fields along the three spatial directions are calibrated with a precision of \(1 \, \text{mG}\) using RF spectroscopy techniques. The first step towards the creation of the spin mixture is to perform a Landau--Zener transition to the \(|F=1, m_F=0\rangle \equiv \ket 0\) state with nearly \(100 \, \%\) transfer efficiency. This is realized at a magnetic field of \(100 \, \text{G}\) to isolate a two-level system exploiting the quadratic Zeeman shifts. The second step consists in inducing a Rabi oscillation among the three Zeeman sublevels to obtain a 50/50 spin mixture of \(\ket{\downarrow}\) and \(\ket{\uparrow}\). The bias field along \(\hat x\) is taken small enough to allow us to neglect the quadratic Zeeman shifts compared to the Rabi frequency and is kept on during the whole experimental sequence following the Rabi pulse. The number of atoms in each spin component is \(N_\uparrow = N_\downarrow \simeq 10^6\) and the total chemical potential of the cloud is \(\mu_{\text{tot}}/k_{\text{B}} \simeq 200 \, \text{nK}\). Fig. [\[Fig1\]](#Fig1){reference-type="ref" reference="Fig1"}(a) shows typical absorption images of the spinor BEC after a \(10 \, \text{ms}\) Stern--Gerlach (SG) expansion in a magnetic field gradient along \(\hat z\). In order to prevent the decay of the mixture to \(|0\rangle\) by spin changing collisions, we lift this level by \(\sim 10 \, \text{kHz}\) using blue detuned microwave dressing on the transition to \(|F=2, m_F=0\rangle\) (see Fig. [\[Fig1\]](#Fig1){reference-type="ref" reference="Fig1"}(b)). # Spin-dipole polarizability The SD polarizability of a spin mixture describes the ability of the system to adapt itself to a displacement in opposite direction of the trapping potentials of the two components. After realizing a fully overlapped configuration, we adiabatically apply a magnetic field gradient \(B'_x\) along \(\hat x\) using a pair of coils in anti-Helmholtz configuration. The gradient is controlled with a resolution at the level of \(4 \, \text{mG/cm}\). This displaces the minima of the trapping potentials such that \(V_{\uparrow,\downarrow}=m\omega_x^2(x\pm x_0)^2/2\) where \(x_0 = g_F \mu_B B'_x/( m \omega_x^2)\) (\(g_F\) is the Landé factor, \(\mu_B\) the Bohr magneton and \(m\) the atomic mass). The SD polarizability is defined as \[\mathcal{P}(x_0) \equiv \frac{d(x_0)}{2 x_0},\] where \(d=x_\downarrow-x_\uparrow\) is the *in-situ* relative displacement between the centers of mass of each component (see Fig. [\[Fig2\]](#Fig2){reference-type="ref" reference="Fig2"}(a)). After a \(2 \, \text{ms}\) SG expansion, we measure \(d\) by fitting each spin component density distribution to independent Thomas--Fermi (TF) profiles to extract their centers \(x_{\uparrow,\downarrow}\). The individual density profiles are not exactly TF-like, but we verified, using a Gross--Pitaevskii equation (GPE) simulation, that this approximate fitting procedure results in an overestimation of \(\mathcal{P}\) by at most 6%. Later in the text we discuss the additional correction to the measurement of \(\mathcal{P}\) related to interactions during the SG expansion. Fig. [\[Fig2\]](#Fig2){reference-type="ref" reference="Fig2"}(a) shows the experimental results where the value of \(x_0\) is estimated after calibrating \(B'_x\). We observe that all data points strongly deviate from the prediction \(d=2 x_0\) for a mixture without intercomponent interactions (green solid line), revealing the large SD polarizability of the system. We use a second experimental protocol to determine the polarizability which will later prove to be useful for measuring the SD oscillation frequency. It consists in realizing the Rabi pulse \(\ket 0 \rightarrow \ket{\uparrow,\downarrow}\) in the presence of a magnetic field gradient. As the minima of the trapping potentials for the \(\ket{\uparrow}\) and \(\ket{\downarrow}\) states are shifted by \(\mp x_0\) with respect to the initial state \(|0 \rangle\), this makes the two components oscillate out of phase after the Rabi pulse. The *in-situ* time evolution of the relative displacement \(D(t)=x_\downarrow(t)-x_\uparrow(t)\) is expected to be given by \(D(x_0,t) = d(x_0) \left[1-\cos \left[ \omega(x_0) t \right] \right]\). Measurements of such oscillations after a SG expansion of \(t_{\text{SG}} = 10 \, \text{ms}\) for different values of \(x_0\) varying the magnetic field gradient are reported in Fig. [\[Fig2\]](#Fig2){reference-type="ref" reference="Fig2"}(b). After the SG expansion, the displacement between the spin components is given by \(D_{\text{SG}}(x_0,t,t_{\text{SG}}) = D(x_0,t) + \partial_t D (x_0,t) \, t_{\text{SG}}\) such that we analyze the data by fitting it with the following function: \[\label{SGEvol} D_{\text{SG}} = A(x_0,t_{\text{SG}}) \cos \left[ \omega(x_0) t + \phi(x_0,t_{\text{SG}}) \right] + d(x_0),\] where \(A(x_0,t_{SG} ) =-d(x_0) \sqrt{1+\omega^2(x_0) t^2_{\text{SG}} }\) and \(\phi(x_0,t_{\text{SG}}) = \arctan \left[ \omega(x_0) t_{\text{SG}}\right]\). Eq. ([\[SGEvol\]](#SGEvol){reference-type="ref" reference="SGEvol"}) allows us to extract the value of \(d(x_0)\) neglecting here again intercomponent interactions during the expansion. Fig. [\[Fig3\]](#Fig3){reference-type="ref" reference="Fig3"} shows the SD polarizability as a function of \(x_0/R_x\) (\(R_x\) is the TF radius along \(\hat x\)) using the data of Fig. [\[Fig2\]](#Fig2){reference-type="ref" reference="Fig2"}. We notice a strong nonlinear dependence of the polarizability on the separation between the two trapping potential minima, which is maximal in the linear limit (\(x_0 \rightarrow 0\)) and tends to \(1\) for large separation (\(x_0\gg R_x\)). In the same figure, we also plot the theoretical predictions obtained within the local density approximation (LDA) and the numerical integration of the GPE performed with the experimental parameters. We identify three regions along \(\hat x\), with the outer two regions occupied by either the \(\ket{\uparrow}\) or \(\ket{\downarrow}\) component and the inner region occupied by both of them. In the linear limit (\(x_0 \to 0\)) the LDA gives the result \[\label{LDAP} \mathcal{P}(x_0 \rightarrow 0) = \frac{g+g_{\uparrow\downarrow}}{g-g_{\uparrow\downarrow}},\] for the polarizability [^1], pointing out its divergent behavior near the phase transition occurring at \(g_{\uparrow\downarrow}=g\). The agreement between the LDA and the GPE is excellent except in the region of small minima separation where the LDA becomes less and less adequate because of the large value of the spin healing length \(\hbar/\sqrt{2mn(g-g_{\uparrow\downarrow})}\) (\(n\) is the total density of the cloud). In general, we observe a good agreement between the theoretical predictions and the experimental data. In particular, the huge effect on the polarizability caused by the vicinity to the miscible-immiscible phase transition is clearly revealed and the scaling with \(x_0/R_x\) is well reproduced. The data analysis presented so far has however been performed neglecting interactions between the spin components during the SG expansion. Indeed, GPE simulations of the expansion in the presence of interactions show that the experimentally measured polarizability is overestimated by \(5 \%\) (\(30 \%\)) for the \(2 \, \text{ms}\) (\(10 \, \text{ms}\)) SG expansion. This explains the remaining difference between the experimental points of Fig. [\[Fig3\]](#Fig3){reference-type="ref" reference="Fig3"} and the theoretical predictions. # Spin-dipole oscillation A useful estimate of the SD frequency is obtained by employing a sum rule approach based on the ratio \(\hbar^2 \omega_{\text{SD}}^2 = M_1/M_{-1}\), where \(M_1=N \hbar^2/2m\) (\(N=N_\uparrow+N_\downarrow\)) is the model independent energy weighted sum rule relative to the SD operator \(\sum_{i} (x_{i\downarrow}-x_{i\uparrow})\), and \(M_{-1}=N\mathcal{P}(x_0 \rightarrow 0)/(2m\omega_x^2)\) is the inverse energy weighted sum rule fixed according to linear response theory by the linear SD polarizability . This leads to the following relation between the SD frequency and polarizability \[\label{relationPolSD} \omega_{\text{SD}}= \frac{\omega_x}{\sqrt{\mathcal{P}(x_0 \rightarrow 0)}} \,.\] Using the LDA expression ([\[LDAP\]](#LDAP){reference-type="ref" reference="LDAP"}) for the polarizability one derives the following prediction for the SD frequency \[\label{SD_ratio} \omega_{\text{SD}}= \sqrt{ \frac{g-g_{\uparrow\downarrow}}{g+g_{\uparrow\downarrow}}} \, \omega_x \,.\] The same result can be directly obtained by generalizing the hydrodynamic theory developed in for density oscillations to the case of SD oscillations. Eq. ([\[SD_ratio\]](#SD_ratio){reference-type="ref" reference="SD_ratio"}) explicitly points out the crucial role played by the intercomponent coupling constant \(g_{\uparrow\downarrow}\) in softening the frequency of the SD mode with respect to the value \(\omega_x\) characterizing the frequency of the in-phase center-of-mass oscillation. We check, using time-dependent GPE simulations of the SD oscillations for our experimental parameters, that the sum rule prediction (4) provides \(\omega_{SD}\) with an accuracy better than \(1\%\) when substituting the value of the static SD polarizability \(\mathcal P(x_0\rightarrow0)\) from the GPE. This demonstrates that an accurate SD frequency measurement can be used to determine the value of the SD polarizability. As shown on Fig. [\[Fig2\]](#Fig2){reference-type="ref" reference="Fig2"}(b), the Rabi pulse in the presence of a magnetic field gradient gives rise to the excitation of SD oscillations whose frequency can be extracted as a function of the induced displacement \(x_0\). A first estimate of the SD frequency is obtained considering that \(\omega_{\text{SD}} = \omega(x_0 \rightarrow 0)\). Indeed, for large values of \(x_0\), the oscillation frequency \(\omega(x_0)\) approaches \(\omega_x\) while it decreases to \(\omega_{\text{SD}}\) as \(x_0 \rightarrow 0\). Since in the small \(x_0\) limit the amplitude of the oscillation tends to zero, we perform a linear fit to the curve of \(\omega(x_0)/\omega_x\) as a function of the oscillation amplitude \(A(x_0)\) and extract \(\omega_{\text{SD}}/\omega_x=0.18(1)\) from the y-intercept of the linear fit (see Fig. [\[Fig4\]](#Fig4){reference-type="ref" reference="Fig4"}(a)). This method shows a good agreement with the LDA prediction Eq. ([\[SD_ratio\]](#SD_ratio){reference-type="ref" reference="SD_ratio"}) \(\omega_{\text{SD}}/\omega_x = 0.189(15)\) and with the GPE simulations yielding \(\omega_{\text{SD}}/\omega_x = 0.213(17)\) (uncertainties take into account error bars on the value of the coupling constants ). The different values of \(\omega_{\text{SD}}\) from the LDA and GPE calculations have the same origin as the one discussed in the case of the polarizability and are due to the large value of the spin healing length in the vicinity of the quantum phase transition. An alternative and more efficient way to excite the SD mode and to measure its frequency consists in first creating two perfectly overlapped spin states where \(B'_x = 0\) (\(x_0=0\)) and then applying a magnetic field gradient \(B'_x = 0.1 \, \text{G/cm}\) (\(x_0 = 1.3 \, \mu \text{m}\)) for \(3 \, \text{ms} \ll 2 \pi / \omega(x_0)\) before restoring \(B'_x = 0\). This leads to an *in-situ* dipole oscillation shown in Fig. [\[Fig4\]](#Fig4){reference-type="ref" reference="Fig4"}(b) after \(10 \, \text{ms}\) of SG expansion. We measure \(\omega_{\text{SD}}/\omega_x = 0.218(2)\) which is slightly larger than the previous estimate based on the data of Fig. [\[Fig2\]](#Fig2){reference-type="ref" reference="Fig2"}(b) and shows better agreement with the prediction from the GPE simulations. For a precise determination of \(\omega_{\text{SD}}\), it is important to ensure that the SD mode has a small *in-situ* amplitude: here we estimate \(D_{\text{SD}} = D_{\text{SG}} / \sqrt{1+\omega^2_{\text{SD}} t^2_{\text{SG}}} = 5.4 \, \mu\text{m}\) which is relatively small compared to the TF radius \(R_x=40 \, \mu\text{m}\). In all experiments, we observe oscillations without noticeable damping on very long timescales (they are ultimately limited by the cloud lifetime). Indeed, the maximal relative velocities of the two superfluid components \(v_{\text{max}} = 1.2 \, \text{mm/s}\) for the data of Fig. 2(b), and \(v_{\text{max}} = 0.4 \, \text{mm/s}\) for the data of Fig. 4(b) are smaller than the critical velocity for the dynamical counterflow instability \(v_{\text{cr}} = \sqrt{\mu_{\text{tot}}(1-g_{\uparrow\downarrow}/g)/2m} = 1.8 \, \text{mm/s}\). # Conclusion In conclusion, we reported on the experimental measurements of the polarizability and of the frequency of the SD oscillation in a two-component BEC of sodium. Because of the vicinity to the miscible-immiscible quantum phase transition both quantities are very sensitive to the value of the intercomponent interaction and their behavior deviates by large factors from the values predicted in the absence of intercomponent interaction. This represents a major difference with respect to other available superfluid quantum mixtures, like the Bose-Fermi mixtures of lithium gases, where the role played by the intercomponent interaction is much less crucial. Similarly to the case of our mixture is characterized by two interacting superfluids oscillating with opposite phase and the observed SD oscillation is undamped for small amplitude as a consequence of superfluidity. For large amplitude motion the Landau's critical velocity will, however, behave very differently, being very sensitive to the value of the intercomponent interaction. Another interesting feature concerns the behavior of the SD oscillation at finite temperature. While the damping of the SD oscillation was actually observed in the old experiments of carried out on a normal Fermi gas, understanding the behavior of the collective modes in the presence of both a condensed (superfluid) and thermal (non-superfluid) components remains extremely challenging. Other topics of interest concern the experimental realization of magnetic solitons and the inclusion of coherent coupling between the two spin components. The Bose mixtures realized and investigated here then represent an ideal platform to explore important equilibrium and dynamic properties of binary superfluids. [^1]: The SD polarizability Eq. ([\[LDAP\]](#LDAP){reference-type="ref" reference="LDAP"}) should not be confused with the magnetic polarizability \(\chi_M = 1/\left[n(g-g_{\uparrow\downarrow})\right]\) which is defined in uniform matter in terms of the energy cost \(\delta E = M^2/(2 \chi_M)\) associated with a small polarization \(M=(N_\uparrow-N_\downarrow)/(N_\uparrow+N_\downarrow)\) of the gas.
{'timestamp': '2016-12-14T02:04:40', 'yymm': '1607', 'arxiv_id': '1607.04574', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04574'}
null
null
# Introduction There are some techniques that have been used to recover the losses of entanglement and protect it from decoherence. Among of these techniques is the quantum purification. Moreno et. al., showed that, it is possible to improve the coherence by using nested environments. Xiao et. al displayed that the teleportation of Fisher information can be enhanced by using partial measurements. The possibility of purifying and concentrating entanglement by using local filtering is discussed by Yashodamma et. al,. The immunity and protecting the entanglement of accelerated qubit-qutrit system by using the local filtering technique is investigated by Metwally. Recently, it has been shown that the local weak-reverse measurements,(WRM) can be used as a technique to protect and improve the correlation in qubit-qutrit systems. Xiao et. al., showed that, the entanglement losses caused by amplitude damping coherence can be retrieved by using the WRM. It is well known that, the accelerated systems lose some of its correlations, and consequently, their efficiencies to perform quantum information tasks decrease. This decay depends on the value of the acceleration and the dimensional of the accelerated system. Therefore we are motivated to investigate the possibility of using the weak-reverse measurement's technique to improve some properties of accelerated system. The main task of this contribution is quantifying the amount of entanglement between the two partners in the presence of weak-reverse measurements, investigating the dynamics of the accelerated/non accelerated information ( which is encoded on the accelerated/non accelerated particle) and the coherent information between the partners. The outline of this manuscript is described as follows: The suggested protocol is described in Sec.2. The suggested system and its analytical solutions are given in Sec.\(3\). The effect of the weak-reverse'strengths on the degree of entanglement is discussed in Sec.\(4\)., where different initial states settings are considered. The dynamics of the accelerated, non-accelerated and coherent information is investigated in Sec.\(5\). Finally, a conclusion is given in Sec.\(6\). # Model Our idea is to consider two partners, Alice and Bob, who share a two-qubit or two-qutrit systems. These systems are initially prepared in maximum or partial entangled states. Here, we consider only Alice's particle is accelerated, which may be a qubit or a qutrit. The partners decided in advance to use the weak-reverse measurement to protect their quantum communication channel. In the following steps, we describe the proposed protocol: 1. Weak measurements:\ Here, both users Alice and Bob perform the weak measurements on their own particles, either qubit or qutrit. We assume that, the two partners share initially a state defined by \(\rho_{ab}^{ini}\). After performing the weak measurements, the partners share a new state defined by \(\rho_{ab}^{out}\), \[\rho_{ab}^{out}=\mathcal{W}_j\mathcal{W}_j\left(\rho_{ab}^{ini}\right)\mathcal{W}_j^{\dagger}\mathcal{W}_j^{\dagger}, ~i,j=q,t,\] where \(\mathcal{W}_i\) are the weak measurements which are defined by the following operators \[\begin{aligned} \label{Wm} \mathcal{W}_q&=&\ket{0}\bra{0}+\sqrt{1-\alpha_q^{(\ell)}}~\ket{1}\bra{1}~\quad\mbox{(for qubit)},\nonumber\\ \mathcal{W}_t&=&\ket{0}\bra{0}+\sqrt{1-\alpha_t^{(1)}}~\ket{1}\bra{1}+\sqrt{1-\alpha_t^{(2)}}~\ket{2}\bra{2}~\quad\mbox{(for-qutrit)}, \end{aligned}\] and \(\alpha_q^{(\ell)}\), \(\alpha_t^{(\ell)}, \ell=1,2\) are the strengths of the weak measurements of the qubit and the qutrit, respectively. 2. Acceleration step:\ It is assumed that, only Alice' particle is moving with a uniform acceleration meanwhile Bob's partial is assumed to be inertial. If the shared particles are of fermions types, then in the Minkowski frame the qubit system is transformed in the Rindler frame as, \[\begin{aligned} \ket{0_M}&=&\cos r\ket{0}_I\ket{0}_{II}+\sin r\ket{1}_I\ket{1}_{II},\quad \ket{1_M}=\ket{1}_{I}\ket{0}_{II} \end{aligned}\] while, for the qutrit system, the vacuum, the spin up and spin down states are transformed into Rindler space as, \[\begin{aligned} \ket{0_M}&=&\cos^2r\ket{0}_{I}\ket{0}_{II}+e^{i\phi}\sin r\cos r(\ket{\mathcal{U}}_{I}\ket{\mathcal{D}}_{II}+\ket{\mathcal{D}}_{I}\ket{\mathcal{U}}_{II}) +e^{2i\phi}sin^2r\ket{\mathcal{D}}_{I}\ket{\mathcal{P}}_{II}, \nonumber\\ \ket{\mathcal{U}_M}&=&cos r\ket{\mathcal{U}}_I\ket{0}_{II}+e^{i\phi}\sin r\ket{\mathcal{P}}_I\ket{\mathcal{U}}_{II}, \nonumber\\ \ket{\mathcal{D}_M}&=&cos r\ket{\mathcal{D}}_I\ket{0}_{II}-e^{i\phi}\sin r\ket{\mathcal{P}}_I\ket{\mathcal{D}}_{II}, \end{aligned}\] where \(\ket{\mathcal{U}}, \ket{\mathcal{D}}\) and \(\ket{\mathcal{P}}\) are the spin up, spin down and pair states, respectively. The acceleration \(r\) is defined such that \(\tan r=Exp[-\pi\omega\frac{c}{a}]\),  \(0\leq r\leq \pi/4\), \(-\infty\leq a\leq\infty\), \(\omega\) is the frequency, \(c\) is the speed of light and \(\phi\) is the phase space which can be absorbed in the definition of the operators. After tracing the particle in the second region (II), the final state represents the accelerated quantum channel between Alice and Bob, \(\rho_{ab}^{acc}\). 3. Reverse measurement step:\ In this step, the users apply the reverse measurement operations on the accelerated state \(\rho_{ab}^{acc}\) to obtain the final state as, \[\label{Final} \rho_{ab}^{Final}=\mathcal{R}_i\mathcal{R}_i\left(\rho_{ab}^{acc}\right)\mathcal{R}^{\dagger}_{i}\mathcal{R}^\dagger_{i},\quad i=q,t,\] where \[\begin{aligned} \label{Rm} \mathcal{R}_q&=&\sqrt{1-\beta_q^{(\ell)}}~\ket{0}\bra{0}+\ket{1}\bra{1},\quad \ell=1,2 \nonumber\\ \mathcal{R}_t&=&\sqrt{(1-\beta_t^{(1)})(1-\beta_t^{(2)})}~\ket{0}\bra{0}+\sqrt{1-\beta_t^{(1)}}~\ket{1}\bra{1} +\sqrt{1-\beta_t^{(2)}}~\ket{2}\bra{2}, \end{aligned}\] and \(\beta_q^{(\ell)}\), \(\beta_t^{(\ell)}, \ell=1,2\) are the strengths of the reverse measurement operations. Some properties of the final state Eq.([\[Final\]](#Final){reference-type="ref" reference="Final"}), such as the behavior of entanglement, accelerated local non-local information are examined. Particularly, the effect of the initial state settings and the local operations strengths on these properties. ## The suggested Systems 1. **Two-qubit system\ ** In this section, it is assumed that the partners, Alice and Bob, share a partially entangled state, which is called \(X-\) state. This state can be written, using the computational basis as, \[\begin{aligned} \label{x-state} \rho_x&=&\mathcal{B}_1\left(\ket{00}\bra{00}+\ket{11}\bra{11}\right)+\mathcal{B}_2\left(\ket{00}\bra{11}+\ket{11}\bra{00}\right) \nonumber\\ &+&\mathcal{B}_3\left(\ket{01}\bra{01}+\ket{10}\bra{10}\right)+\mathcal{B}_4\left(\ket{01}\bra{10}+\ket{10}\bra{01}\right) \end{aligned}\] where \(\mathcal{B}_1=\frac{1}{4}(1+c_{33}), \mathcal{B}_2=\frac{1}{4}(c_{11}+c_{22}), \mathcal{B}_3=\frac{1}{4}(1-c_{33}),\) and \(\mathcal{B}_4=\frac{1}{4}(c_{11}-c_{22})\), and \(c_i,i=1,2,3\) are the diagonal elements of \(3\times 3\) cross dyadic. The state ([\[x-state\]](#x-state){reference-type="ref" reference="x-state"}) can be described by its Bloch vectors, \(\row{s}_j=0,j=1,2\) and a dyadic \(\dyadic{C}\) with \(c_{ij}=0\) for \(i\neq j\). From this state, one can get maximum entangled classes, MES. For example, if we set \(c_{11}=c_{22}=c_{33}=-1\), one gets the singled state, \(\rho_{\psi^-}=\ket{\psi^-}\bra{\psi^-}\). The Werener state can be obtained if we set \(c_{11}=c_{22}=c_{33}=-x\), etc. As mentioned earlier, only Alice's particle will be accelerated. Due to the acceleration the degree of entanglement decreases. Therefore, the partners try to improve the degree of entanglement by applying the weak-reverse measurements. At the end of the protocol, the partners share the state. \[\begin{aligned} \label{Final-qubit} \rho_{ab}^{Final}&=&\frac{1}{\mathcal{N}_q}\Bigl\{\mathcal{B}_1\ket{00}\bra{00}+\tilde\mathcal{B}_2\ket{00}\bra{11} +\tilde\mathcal{B}_3\ket{01}\bra{01}+\tilde\mathcal{B}_4\ket{01}\bra{10} \nonumber\\ &+&\mathcal{B}_5\ket{10}\bra{10}+\tilde\mathcal{B}_6\ket{10}\bra{01} +\tilde\mathcal{B}_7\ket{11}\bra{11}+\tilde\mathcal{B}_8\ket{11}\bra{00}\Bigr\} \end{aligned}\] where, \[\begin{aligned} \tilde\mathcal{B}_1&=&c_1^2\mathcal{B}_1(1-\beta_q^{(1)})(1-\beta_q^{(2)}),\quad \nonumber\\ \tilde\mathcal{B}_2&=&c_1\mathcal{B}_2\sqrt{1-\beta_q^{(1)}}\sqrt{1-\beta_q^{(2)}}\sqrt{1-\alpha_q^{(1)}}\sqrt{1-\alpha_q^{(2)}}, \nonumber\\ \tilde\mathcal{B}_3&=&c_1^2\mathcal{B}_3(1-\beta_q^{(1)})(1-\alpha_q^{(2)}),\quad \nonumber\\ \tilde\mathcal{B}_4&=&c_1\mathcal{B}_4\sqrt{1-\beta_q^{(1)}}\sqrt{1-\beta_q^{(2)}}\sqrt{1-\alpha_q^{(1)}}\sqrt{1-\alpha_q^{(2)}}, \nonumber\\ \tilde\mathcal{B}_5&=&(1-\beta_q^{(2)})\left[s_1^2\mathcal{B}_1+(1-\alpha_q^{(1)})\mathcal{B}_3\right],\quad \tilde\mathcal{B}_6=\tilde\mathcal{B}_4 \nonumber\\ \tilde\mathcal{B}_7&=&(1-\alpha_q^{(1)})(1-\alpha_q^{(2)}) \mathcal{B}_1,~\quad \tilde\mathcal{B}_8=\tilde\mathcal{B}_2, \end{aligned}\] and \(\mathcal{N}_q=\tilde\mathcal{B}_1+\tilde\mathcal{B}_3+\tilde\mathcal{B}_5+\tilde\mathcal{B}_6\) is the normalization factor and \(c_1=\cos r, s_1=\sin r\). 2. Two-qutrit system:\ In this section, we investigate the effect of the local-weak-reverse operations on a system consists of two qutrits, where it is assumed that only Alice's qutrit will be accelerated. In the computational basis, the initial sate of this system can be written as , \[\ket{\psi_{t}}=\frac{1}{\sqrt{2+\gamma^2}}\left(\ket{00}+\ket{11}+\gamma\ket{22}\right),\] where it turns into a maximum entangled state for \(\gamma=1\). The users apply the protocol which is described in Sec.2. At the end the partners share the following state, \[\begin{aligned} \label{Final-qutrit} \rho_{ab}^{Final}&=&\frac{1}{\mathcal{N}_t}\Bigl\{\mathcal{D}_1\ket{00}\bra{00}+\mathcal{D}_2\ket{00}\bra{11}+\mathcal{D}_3\ket{10}\bra{10}+ \mathcal{D}_4\ket{11}\bra{00}+\mathcal{D}_5\ket{11}\bra{11} \nonumber\\ &&+\mathcal{D}_6\ket{20}\bra{20}+\mathcal{D}_7\ket{22}\bra{00}+\mathcal{D}_8\ket{22}\bra{11}+\mathcal{D}_9\ket{22}\bra{22}+\mathcal{D}_{10}\ket{00}\bra{22} \nonumber\\&& +\mathcal{D}_{11}\ket{11}\bra{22}\Bigr\}, \end{aligned}\] where \[\begin{aligned} \mathcal{D}_1&=&c_1^2\mathcal{R}_{00}^2\mathcal{A}_1, \quad \mathcal{D}_2=c_1^3\mathcal{R}_{00}\mathcal{R}_{11}\mathcal{A}_2, \quad \mathcal{D}_3=c_1^2s_1^2\mathcal{R}_{10}^2\mathcal{A}_1, \nonumber\\ \mathcal{D}_4&=&c_1^3\mathcal{R}_{00}\mathcal{R}_{11}\mathcal{A}_4,\quad \mathcal{D}_5=c_1^3\mathcal{R}_{11}^2\mathcal{A}_5,\quad \mathcal{D}_6=c_1^2s_1^2\mathcal{R}_{20}^2\mathcal{A}_1, \nonumber\\ \mathcal{D}_7&=&c_1^3\mathcal{R}_{22}\mathcal{R}_{00}\mathcal{A}_7,\quad \mathcal{D}_8=c_1^2\mathcal{R}_{22}\mathcal{R}_{11}\mathcal{A}_8,\quad \mathcal{D}_9=c_1^2\mathcal{R}_{22}^2\mathcal{A}_9, \nonumber\\ \mathcal{D}_{10}&=&\mathcal{R}_{00}\mathcal{R}_{22}\mathcal{A}_3c^3,\quad \mathcal{D}_{11}=\mathcal{R}_{11}\mathcal{R}_{22}\mathcal{A}_6c^2,\quad \end{aligned}\] The normalization factor is given by \(\mathcal{N}_t=\mathcal{D}_1+\mathcal{D}_3+\mathcal{D}_5+\mathcal{D}_6+\mathcal{D}_9\) and the coefficients \(\mathcal{A}_i, i=1..9\) are given by, \[\begin{aligned} \mathcal{A}_1&=&\frac{1}{2+\gamma^2},~\quad\quad \mathcal{A}_2=\frac{\sqrt{1-\alpha_t^{(1)}}\sqrt{1-\alpha_t^{(2)}}}{2+\gamma^2},\quad \mathcal{A}_3=\frac{\gamma\sqrt{1-\alpha_t^{(1)}}\sqrt{1-\alpha_t^{(2)}}}{2+\gamma^2},~\quad\quad \nonumber\\ \mathcal{A}_4&=&\mathcal{A}_2,~\quad\quad \mathcal{A}_5=(2+\gamma^2)\mathcal{A}_2^2,~\quad\quad \mathcal{A}_6=\gamma\mathcal{A}_2\sqrt{1-\alpha_t^{(1)}}\sqrt{1-\alpha_t^{(2)}},\quad \nonumber\\ \mathcal{A}_7&=&\frac{\gamma}{2+\gamma^2}\sqrt{1-\alpha_t^{(1)}}\sqrt{1-\alpha_t^{(2)}},~\quad \mathcal{A}_8=\frac{\gamma(1-\alpha_t^{(1)})}{2+\gamma^2}\sqrt{1-\alpha_t^{(2)}}\sqrt{1-\alpha_t^{(2)}},\quad \nonumber\\ \mathcal{A}_9&=&\frac{\gamma^2}{2+\gamma^2}(1-\alpha_t^{(1)})(1-\alpha_t^{(2)}), \end{aligned}\] and \(\mathcal{R}_{ij},i,j=0,1,2\) are given by \[\begin{aligned} \mathcal{R}_{00}&=&\sqrt{(1-\beta_1^{(1)})(1-\beta_2^{(1)})}\sqrt{(1-\beta_1^{(2)})(1-\beta_2^{(2)})},\quad \nonumber\\ \mathcal{R}_{01}&=&\sqrt{1-\beta_1^{(2)}}\sqrt{(1-\beta_1^{(1)})(1-\beta_2^{(1)})},\quad \nonumber\\ \mathcal{R}_{02}&=&\sqrt{1-\beta_2^{(2)}}\sqrt{(1-\beta_1^{(1)})(1-\beta_2^{(1)})},\quad \nonumber\\ \mathcal{R}_{10}&=&\sqrt{1-\beta_1^{(1)}}\sqrt{(1-\beta_1^{(2)})(1-\beta_2^{(2)})}, \nonumber\\ \mathcal{R}_{11}&=&\sqrt{1-\beta_1^{(1)}}\sqrt{1-\beta_1^{(2)}},~\quad\quad \mathcal{R}_{12}=\sqrt{1-\beta_1^{(1)}}\sqrt{1-\beta_2^{(2)}}, \nonumber\\ \mathcal{R}_{20}&=&\sqrt{1-\beta_2^{(1)}}\sqrt{(1-\beta_1^{(2)})(1-\beta_2^{(2)})},~\quad\quad \mathcal{R}_{22}=\sqrt{1-\beta_2^{(1)}}\sqrt{1-\beta_2^{(2)}}. \end{aligned}\] # Entanglement To quantify the survival degree of entanglement \(\mathcal{E}\) contained in the system, we use the negativity as a measure. The negativity is defined as, \[\mathcal{E}=max\left(0,\sum_i{\lambda_i}\right),\] where \(\lambda_i,\) are the eigenvalues of the partial transpose of \(\rho^{T_a}_{ab}\). Fig.(1a), displays the behavior of entanglement of the final state ([\[Final-qubit\]](#Final-qubit){reference-type="ref" reference="Final-qubit"}), where it is assumed that, the partners share initially a singlet state i.e., \(c_{11}=c_{22}=c_{33}=1\) and we have set \(\alpha_q^{(1)}=\alpha_q^{(2)}=\beta_q^{(1)}=\beta_q^{(2)}=\alpha_q\). It is clear that, for \(r=0\) the degree of entanglement is maximum i.e., (\(\mathcal{E}=1)\). The general behavior shows that, the entanglement decreases as the acceleration parameter, \(r\) increases. On the other hand, as one increases the strengths of the local operations, the entanglement increases. However, for further values of \(\alpha\in[0.75,1]\), the entanglement decreases. The behavior of entanglement of a system initially prepared in a partially entangled state of Werner type is displayed in Fig.(1b). The general behavior is similar to that for the MES (Fig.(1a)), but with smaller decay rate. On the other hand, for \(\alpha\in[0.75,1]\), the entanglement re-appears again while, it vanishes completely for the MES. The dynamics of entanglement under the effect of the local weak-reverse measurement for a system is initially prepared in a two-qutrit system is described in Fig.(2). The general behavior is similar to that predicted in Figs.\((1)\), where the entanglement decays as \(r\) increases. However, the upper and lower bounds of entanglement depend on initial state settings. It is clear that, for the two-qutrit system the entanglement increases slowly compared with that shown in Fig.(1). Also, as displayed in Fig.(2b), the increasing rate of entanglement for PES is larger than that shown for MES. From the previous figures, one concludes that the weak and reverse measurements can recover the loss of entanglement for small values of accelerations as the local operations' strengths increase for different intervals. The length of these intervals depends on the initial states settings: two-qubit, or two qutrits, maximum / partial entangled states. The increasing rate of entanglement is much larger for systems that are initially prepared in partial entangled states. The partially entangled 2- qubits state is more robust than the partially entangled two-qutrit state. # Local and Non-local Information ## Two-qubit system In this subsection, we quantify Alice's and Bob's information, which represents the accelerated \((\mathcal{I}_a)\) and non-accelerated,(\(\mathcal{I}_b\)) information, respectively. Moreover, the amount of the non-local information between the partners defined by the coherent information \((\mathcal{I}_{coh})\) will be quantified also. In an explicit form, theses three types of information can be written as, \[\begin{aligned} \mathcal{I}_{a}&=&-\frac{\tilde\mathcal{B}_1+\tilde\mathcal{B}_3}{\mathcal{N}_{q}}\log\left(\frac{\tilde\mathcal{B}_1+\tilde\mathcal{B}_3}{\mathcal{N}_{q}}\right)- \frac{\tilde\mathcal{B}_5+\tilde\mathcal{B}_7}{\mathcal{N}_{q}}\log\left(\frac{\tilde\mathcal{B}_5+\tilde\mathcal{B}_7}{\mathcal{N}_{q}}\right) \nonumber\\ \mathcal{I}_{b}&=&-\frac{\tilde\mathcal{B}_1+\tilde\mathcal{B}_5}{\mathcal{N}_{q}}\log\left(\frac{\tilde\mathcal{B}_1+\tilde\mathcal{B}_5}{\mathcal{N}_q}\right)- \frac{\tilde\mathcal{B}_3+\tilde\mathcal{B}_7}{\mathcal{N}_{q}}\log\left(\frac{\tilde\mathcal{B}_3+\tilde\mathcal{B}_7}{\mathcal{N}_{q}}\right) \nonumber\\ I_{coh}&=&\sum_{i=1}^{4}\mu_i\log(\mu_i), \end{aligned}\] where \[\begin{aligned} \mu_{1,2}&=&\frac{1}{2\mathcal{N}_{q}}\Bigl\{(\tilde\mathcal{B}_1+\tilde\mathcal{B}_7) \pm\sqrt{(\tilde\mathcal{B}_1-\tilde\mathcal{B}_7)^2+4\tilde\mathcal{B}_2\tilde\mathcal{B}_8}\Bigr\}, \nonumber\\ \mu_{3,4}&=&\frac{1}{2\mathcal{N}_{q}}\Bigl\{(\tilde\mathcal{B}_3+\tilde\mathcal{B}_5) \pm\sqrt{(\tilde\mathcal{B}_3-\tilde\mathcal{B}_5)^2+4\tilde\mathcal{B}_4\tilde\mathcal{B}_6}\Bigr\}, \end{aligned}\] Fig.(3) displays the behavior of the coherent information and the accelerated local information, where we set \(\alpha_q^{(1)}=\alpha_q^{(2)}= \beta_q^{(1)}=\beta_q^{(2)}=\alpha_q\). At zero acceleration, both types of information are maximum. As Alice's particle is accelerated, the coherent information decreases. However, for a fixed value of \(r\), the coherent information increases as the strengths of the local operations increase. Meanwhile, the accelerated local information increases at the expense of the coherent information. For larger values of \(\alpha_q\), the accelerated information vanishes while the coherent information increases. This coherent information, no longer represents quantum information but it describes a classical information, because the two particle are almost separable for larger values of \(\alpha_q\) as displayed in Fig.(1). In Fig.(4), we scrutinize the effect of the WRM on the local and non local information, as well as on the coherent information. The behavior of the accelerated information \(\mathcal{I}_a\) and the non-accelerated information \(\mathcal{I}_b\) are described in Fig.(4a), where it is assumed that, the system is either initially prepared in a maximum entangled state (MES) or in a partial entangled state (PES). At zero acceleration \((r=0)\) the information which is encoded in Alice's and Bob's qubit is maximum i.e. \(\mathcal{I}_a=\mathcal{I}_b=1~ bit\). For small values of \(r\in[0,2.5]\), \(I_a\) and \(\mathcal{I}_b\) are remaining maximum \((1~bit)\). For larger values of \(r\), the accelerated information \(\mathcal{I}_a\) decreases as the acceleration \(r\) increases, while Bob's information (non-accelerated information) is slightly affected due to the local WRM. Moreover, the decay rate of the accelerated information for a system prepared initially in a partial entangled state is smaller than that displayed for systems that are initially prepared in MES. This shows that, the possibility of protecting the accelerated local information encoded in PES by the WRM is much better than protecting that coded in MES. The effect of the local WRM on the accelerated local information, \(\mathcal{I}_a\) and the coherent information \(\mathcal{I}_{coh}\) is depicted in Fig.(4b), where it is assumed that the system is initially prepared in the MES. In these calculations, we consider that the strengths of the local weak measurement and the reverse measurements are equal \((\alpha_q=0.5)\). The behavior of information shows that, both types of the information are maximum at \(r=0\). As \(r\) increases, the accelerated local information \(\mathcal{I}_a\) has maximum values for \(r\in[0,2.5]\), while the coherent information decreases for any value of \(r>0\). It is clear that, as the strengths of the WRM increase, both types of information decrease. However the decay rate of \(\mathcal{I}_{coh}\) is much larger than that for accelerated local information, \(\mathcal{I}_a\). For larger values of \(r\) and \(\alpha_q\), the coherent information increases at the expense of the local accelerated information. From these figures, one concludes that the accelerated information for small values of acceleration can be protected by using the weak-reverse measurement. The possibility of protecting the accelerated information which is encoded in partial entangled states is larger than that encoded in a maximum entangled states. The local measurement has a very slight effect on the non-accelerated local information. For larger values of accelerations, the coherent information increases at the expense of the accelerated information ## Two-qutrit system Fig.(5) shows the behavior of the coherent information \(\mathcal{I}_{coh}\) and the accelerated information \(\mathcal{I}_a\) for a system initially prepared in an entangled maximum two-qutrit system. It is clear that, both types of information decrease as the acceleration \(r\) increases. The decay rate is larger for the coherent information compared with that depicted for the accelerated information. However, as the strengths of the local operations increase both types of information are improved. For \(\alpha_t\in[0.8,~1]\), the coherent information increases at the expense of the accelerated information. The dynamics of different types of information coded in a system initially prepared in a maximum two-qutrit entangled state is described in Fig.(6). The behavior of the accelerated local information \(\mathcal{I}_a\) and the non-accelerated local information, \(\mathcal{I}_b\) is displayed in Fig.(6a), where it is assumed that, the system initially prepared in MES or PES. It is clear that, the accelerated information for a system is initially prepared in a MES is almost stable and it is a maximum for any value of \(r\). On the other hand, starting from PES, the initial values of the accelerated information is smaller than that depicted for MES. However, the upper bounds of \(\mathcal{I}_a\) increase as the acceleration \(r\) increases. These upper bounds become larger than that depicted for MES for \(r\in[0,~0.8]\). For non-accelerated information the behavior is completely different from those shown for the two-qubit system, where it is always smaller than the accelerated information. Moreover, it decreases as \(r\) increases. Fig(6b) shows the behavior of the coherent information and the accelerated information for different values of the local operation's strengths, with initial system prepared in a MES of the two-qutrit system. It is clear that, \(\mathcal{I}_{coh}\) decreases as \(\alpha_t\) and \(r\) increase. For smaller values of \(\alpha_t\), the coherent information decreases faster than that displayed for larger values of \(\alpha_t\). On the other hand, for larger values of \(\alpha_t\) and \(r\), the coherent information increases to reach its upper bounds as \(r\to \infty\). Also, the general behavior of the accelerate information shows that, \(\mathcal{I}_a\) decreases as the strengths of the local operations increases. However, the decay rate is smaller than the coherent information. This explain the decay of the non-accelerated information displayed in Fig.(6a). From these figures one can extract some important facts. The local information which is encoded in the accelerated part of the composite system can be protect by using the WRM. The local accelerated information which is encoded in a partial entangled state can be improved even for larger accelerations. The non-accelerated local information is very sensitive to the local operations (WRM), compared with that shown for the two-qubit system. The decay rate of accelerated information is much smaller than that shown for the coherent information. For larger values of accelerations, the coherent information increases at the expense of the non-accelerated local information. # Conclusion In this contribution, we assume that two users, Alice and Bob, share 2-qubit or 2-qutrit states as a communication channels to perform some quantum information tasks. Alice is allowed to accelerate her qubit/qutrit. Subsequently, the coherence of the communication channel between the partners decreases. Thus, we try to investigate the possibility of utilizing the weak-reverse technique to improve the quantum correlations between the users. It is shown that, the initial state settings, the values of the acceleration and the strengths of the weak-reverse operations represent control keys to the behavior of entanglement, local and non-local information. It is demonstrated that, for small values of acceleration the improving rate of the quantum correlation of the communication channels increases as the strengths of the weak-reverse operation increase. However, as one increases the local operations' strengths further the correlation decreases. Our results show that, starting from less entangled state the improving rate of correlations is better than that depicted for maximum entangled states. On the other hand, the increasing rate of the correlation for the two-qutrit system is larger for large values of the acceleration comparing with that displayed for the two-qubit-system. Additionally, the behavior of the local information which includes the accelerated and non-accelerated information is assayed. It is shown that, for the two-qubit system, the non-accelerated information is almost stable and slightly decreases for systems initially prepared in maximum entangled state. But the accelerated information decreases as the acceleration increases. Moreover, the encoded information in a system initially prepared in partial entangled state is more robust than that encoded in a system initially prepared in maximum entangled states. Theses results are dramatically changed for the two-qutrit systems, where the non-accelerated information is more fragile than the accelerated information. Also, the encoded information in the partial entangled state of the two-qutrit system is improved as the acceleration increases under weak-reverse measurements. Finally, the general behavior of the coherent information displays that. it decreases as the acceleration increases. For a fixed value of acceleration, the coherent information increases as the strength of the weak-reverse measurement increases. However, for the two-qubit system, larger values of these strengths cause a sudden decay of information for small values of the accelerations and gradually decay for larger values of these accelerations. Meanwhile, for the two-qutrit system, the coherent information slightly increases at larger values of the weak-reverse measurements' strengths. The changes happened at the expense of the local information. *In conclusion*, it is possible to improve and protect the accelerated communication channel by using weak-reverse measurements. The possibility of protecting the local encoded information on an accelerated 2-qutrit system is larger than that displayed for 2-qubit system. The restrained decay rate of the coded information on accelerated partial entangled state is better than that displayed for maximum entangled state. The encoded information in the accelerated part of the two-qutrit system can be improved by using the local operations, while it decays for the two-qubit system. Therefore, to perform quantum key distribution or quantum coding protocols using accelerated system, it is better to use partial entangled 2-qutrit systems with larger accelerations.
{'timestamp': '2016-07-19T02:08:49', 'yymm': '1607', 'arxiv_id': '1607.04916', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04916'}
null
null
null
null
null
null
# Introduction The problem of sampling complex free energy landscapes that exhibit long lived metastable states separated by large barriers is of great current interest. The vast literature on the subject is a clear evidence of its pressing relevance. Roughly speaking, two classes of methods can be identified. In one, a set of collective variables (CVs) that depend on the microscopic coordinates of the system is chosen, and a bias that depends on these chosen collective variables is constructed so as to speed up sampling. Examples of this approach are Umbrella Sampling, Local Elevation, Metadynamics and more recently Variationally Enhanced Sampling. However, identifying the appropriate CVs can at times require a lengthy, if instructive, process. The other set of methods can be classified under the generic name of tempering. The precursor of this approach is called Parallel Tempering (PT). In PT \(M\) replicas of the same system at different temperatures are run in parallel. Periodically a Monte Carlo test is made and, if the test is successful, configurations are exchanged between replicas. The rationale for this approach is that at high temperature it is easier for the system to move from basin to basin and this information is carried down to the colder temperature by the Monte Carlo exchange process. This idea has been generalized and several tempering schemes have been proposed in which the Hamiltonians in each replicas are progressively modified in order to favor the process of barrier crossing. Statistics is then collected in the unmodified replica. Since it is relevant for what follows, we highlight the recent proposal made by Voth and collaborators and by our group to use artificially induced quantum effects to favor sampling. In this approach quantum effects are described by the isomorphism that is generated by the use of the Path Integral representation of Quantum Statistical Mechanics. In this popular isomorphism each particle is mapped onto a polymer of \(P\) beads interacting between themselves through a harmonic potential while the interaction between beads belonging to different polymers is appropriately reduced. The success of these attempts have stimulated us to reduce this approach to its bare essentials and somehow to generalize it. In order to reduce its computational cost the number of beads will be reduced to two, thus we shall abandon any pretense of describing each replicas as a representation, albeit approximate, of a quantum state. Having given ourselves such a freedom we shall also change the interaction between the beads. Thus our tempering scheme will be composed of \(M\) replicas; in each replica every particle becomes a dimer, and the interaction that holds together the dimers is now anharmonic. By tightening the potential that holds the dimers together we can progressively come close to the Boltzmann distribution of the system. As in the Feynman's isomorphism, this is reached when the intra-dimer interaction becomes infinitely strong. Like in Ref. a key to the success of this method is the use of Metadynamics to increase the dimer length fluctuations. Thus we shall refer to this method as Dimer Metadynamics (DM). # Method As discussed in the Introduction, our method is inspired by previous attempts to use quantum effects to overcome sampling bottlenecks. However here we prefer to derive our formulas without making explicit reference to the Path Integral representation of Quantum Statistical Mechanics. We want to sample a Boltzmann distribution whose partition function is written as: \[\begin{aligned} \label{z1} Z_0 = \int dR \: \mbox{e}^{-\beta V(R)} \end{aligned}\] where \(N\) particles of coordinates \(R=\vec{r}_i, \: i=1...N\) interact via the \(V(R)\) potential at the inverse temperature \(\beta\). Let us now introduce \(3N\) new coordinates \(X = \lbrace \vec{x}_i \rbrace_{i=1}^{N}\) and a function \(F_\sigma (X)\) that depends parametrically on \(\sigma\) and it is such that \(\int dX \mbox{e}^{-\beta F_\sigma (X)}=1\). We can now rewrite \(Z_0\) as \[\begin{aligned} \label{zrew} Z_0 = \int dRdX \: \mbox{e}^{-\beta F_\sigma (X)}\mbox{e}^{-\beta V(R)} \end{aligned}\] and making the coordinate transformation \(X = R_1-R_2\) and \(R = (R_1 + R_2)/2\) we obtain \[\begin{aligned} \label{zrew2} Z_0 = \int dR_1dR_2 \: \mbox{e}^{-\beta F_\sigma (R_1-R_2)}\mbox{e}^{-\beta V(\frac{R_1 + R_2}{2})} \:. \end{aligned}\] Of course the partition function in Eq. [\[zrew2\]](#zrew2){reference-type="eqref" reference="zrew2"} is fully equivalent to the Boltzmann distribution and the doubling of the coordinates is only apparent. We now choose a family of functions \(F_\sigma\) such that \[\begin{aligned} \label{dfam} \lim_{\sigma\rightarrow 0} \mbox{e}^{-\beta F_\sigma (R_1-R_2)} = \delta (R_1-R_2) \end{aligned}\] where \(\delta(R_1-R_2)\) is the Dirac's delta. We shall postpone the choice of \(F_\sigma\) to later, in the meantime we note that if Eq. [\[dfam\]](#dfam){reference-type="eqref" reference="dfam"} holds for \(\sigma \rightarrow 0\), we can make the approximation \(V(\frac{R_1+R_2}{2}) \simeq \frac{V(R_1)}{2}+\frac{V(R_2)}{2}\) and rewrite the partition function as: \[\begin{aligned} \label{zrew3} Z_\sigma = \int dR_1dR_2 \: \mbox{e}^{-\frac{1}{2}V(R_1)}\mbox{e}^{-\beta F_\sigma(R_1-R_2)}\mbox{e}^{-\frac{1}{2}V(R_2)} \end{aligned}\] that is fully equivalent to the Boltzmann distribution (Eq. [\[z1\]](#z1){reference-type="eqref" reference="z1"}) in the limit \(\sigma \rightarrow 0\), however, contrary to Eq. [\[zrew2\]](#zrew2){reference-type="eqref" reference="zrew2"} here the degrees of freedom are doubled in earnest. This partition function describes a system of dimers bound by \(F_\sigma (R_1-R_2)\) and interacting via the reduced potential \(\frac{V(R)}{2}\). At large values of \(\sigma\) the dimers are loosely coupled and the two images of the system \(R_1\) and \(R_2\) can explore rather different configurations. By reducing \(\sigma\) the coupling becomes stronger and stronger until for \(\sigma\rightarrow 0\) the Boltzmann limit is approached. We now devise a set of replicas of which the first one is \(Z_0\) whose \(\sigma\) we denote by \(\sigma_0\) and the others are of the type in Eq. [\[zrew3\]](#zrew3){reference-type="eqref" reference="zrew3"} with progressively large values of \(\sigma\), starting with \(\sigma=\sigma_1\). For reasons that will become apparent later we take \(\sigma_0 = \sigma_1\). The idea is then to set up a replica exchange scheme in which systems of different \(\sigma_i, \: i=1...,M\) are run in parallel and periodically Monte Carlo attempts at swapping the configurations are made. In the tempering scheme proposed here the Monte Carlo test between neighboring replicas reads: \[\begin{aligned} p_{i,i+1}= \min \left[ 1, \mbox{e}^{-\beta\Delta E } \right] \label{pswp} \end{aligned}\] with: where the subscript in the many-body coordinates is the bead index of the dimer and the superscript identifies the replica index. The Monte Carlo test between \(Z_0\) and \(Z_{\sigma_1}\) is given by: with: and we use the fact that we have chosen \(\sigma_0 = \sigma_1\). We now turn to the choice of \(F_\sigma (R_1-R_2)\). We use the form \[\begin{aligned} F_\sigma (R_1-R_2 ) = \sum_{i=1}^N f_\sigma (\vec{r}_i^{\,1}-\vec{r}_i^{\,2}) \end{aligned}\] where \(\vec{r}_i^{\,1}\) and \(\vec{r}_i^{\,2}\) are the coordinates of atom \(i\) that has been split into the two beads \(1\) and \(2\) (see Eq. [\[zrew2\]](#zrew2){reference-type="eqref" reference="zrew2"}). For our approach to work it is necessary that for \(\sigma \rightarrow 0\) the behavior of \(\mbox{e}^{-f_\sigma (r)}\) is \(\delta\)-function like. One such class of functions can be obtained by considering for \(0 < q \le 1\) the following representations of the three dimensional delta functions \[\begin{aligned} \delta (r) = \lim_{\sigma\rightarrow 0} \frac{\mbox{e}^{-\left[\left(1+\frac{r^2}{2q\sigma^2}\right)^q-1\right]}}{Z_{\sigma}^q} \end{aligned}\] where the normalization constant \(Z_\sigma^q\) is given by \[\begin{aligned} Z_\sigma^q = 4\pi \int dr\: r^2 \mbox{e}^{-\left[\left(1+\frac{r^2}{2q\sigma^2}\right)^q-1\right]} \end{aligned}\] Using Eq. [\[dfam\]](#dfam){reference-type="eqref" reference="dfam"} and neglecting the immaterial constant \(\log Z_\sigma^q\) we get for \(f_\sigma^q(r)\) \[\begin{aligned} f_\sigma^q (r) = \left(1+\frac{r^2}{2q\sigma^2}\right)^q-1 \end{aligned}\] The choice of \(f_\sigma^q (r)\) determines the potential with which two beads interact. For \(q=1\) one has \[\begin{aligned} \label{fsgauss} f_\sigma^{(1)} (\vec{r}\,) = \frac{r^2}{2\sigma^2} \end{aligned}\] as in the standard Path Integral isomorphism. Otherwise for \(0 < q < 1\) all \(f_\sigma^q (r)\) exhibit a quadratic behavior \(\frac{r^2}{2\sigma^2}\) at small \(r\) and a slower growth at larger distances. In particular, for \(q=\frac{1}{2}\) \(f_\sigma^q (r)\) grows linearly with \(r\). The transition between small-\(r\) and large-\(r\) regimes is controlled by the parameter \(\sigma\), that in the spirit of the present work has no physical meaning and is only a tempering parameter. As \(q \rightarrow 0\) the asymptotic behavior is even slower and in the limit becomes logarithmic. However in this case \(Z_\sigma^q \rightarrow \infty\) and the system becomes unstable. In this first application of the method we choose \(q=\frac{1}{2}\). This value is possibly not optimal but it is better than \(q=1\). In our experimentation we found that \(q=0.4\) and \(q=0.3\) are also viable options. However the advantages did not seem so great as to warrant abandoning the more aesthetically pleasing \(q=\frac{1}{2}\) choice. In this respect we find amusing to note that also the quark-quark interaction has a linear asymptotic behavior. The final and essential ingredient of our approach is the use of Metadynamics. Following Ref. we shall combine the replica exchange scheme described above with Well-Tempered Metadynamics. In particular we shall use as CV the elastic energy per particle stored in the dimer \[\begin{aligned} \label{scv} s=\frac{1}{N \beta}\sum_{i=1}^{N}\left[\left(1+\frac{\left(\vec{r}_i^{\,2}-\vec{r}_i^{\,1}\right)^2}{2q\sigma^2}\right)^q-1\right] \end{aligned}\] The role of Metadynamics is to enhance the fluctuations of \(s\) since in a Well-Tempered Metadynamics that uses \(\gamma\) as boosting parameter the probability distribution of the biased variable, \(p_b(s)\) is related to that in the unbiased ensemble by the relation \(p_b(s) \propto \left[p(s)\right]^{1/\gamma}\). In our case since \(s\) is related to the elastic energy, configurations in which the dimer is highly stretched are more likely to be observed. # Results Before discussing the two applications presented here we want briefly to substantiate the assertion made earlier that the choice \(q=\frac{1}{2}\) is more efficient than \(q=1\). For this reason we consider once more the case of Alanine Dipeptide, a classical simple example on which new sampling methods are often tested. In [\[ala2pic\]](#ala2pic){reference-type="ref" reference="ala2pic"} we show the results obtained with the DM method with \(q=\frac{1}{2}\) where only 6 replicas where required while in contrast, for \(q=1\) we had to use 10 replicas. We now tackle a simple, yet challenging two dimensional system composed of 7 atoms interacting via Lennard-Jones (LJ) potential. This cluster is known to have three metastable configurations that can be represented as local minima in the free energy expressed as a function of the second and third momentum of the coordination number as shown in [\[figljfes\]](#figljfes){reference-type="ref" reference="figljfes"}. Standard MD is not able to sample the metastable states of this simple cluster in practical times and enhanced sampling methods have been used to study it. In [\[figljfes\]](#figljfes){reference-type="ref" reference="figljfes"} the free energy surface of this LJ cluster at \(T = 0.1\) has been computed with DM using 5 replicas with \(\sigma_i =\)`<!-- -->`{=html}0.01, 0.01, 0.05, 0.2 and 0.6, where \(\sigma_0 = \sigma_1\) as discussed before and LJ units are used. The timestep was \(dt = 0.01\) and the simulation ran for \(10^8\) steps. Swaps between replicas were tried every 500 steps and the Metadynamics Gaussians were deposited every 200 steps with initial height of \(0.1\) and width depending on the replica index, \(\sigma_g =\) 5.7, 3.6, 2.1 and 1.4; the bias factor was \(\gamma = 2.71\). These results are compared to a \(10^8\) steps long Metadynamics simulation with timestep \(dt = 0.005\) and bias factor \(\gamma = 10\), where every \(500\) steps the second and third momentum of the coordination number are biased with Gaussians of initial height \(w_0= 0.01\), width \(\sigma = (0.02; 0.02)\). As shown in [\[geom\]](#geom){reference-type="ref" reference="geom"} DM dramatically enhances the sampling of the long distances tail of the probability distribution of the dimer length. Analogously to the Path Integral case, this effect increases the delocalization of the particle and indeed even without replica exchange DM can locate all of the four minima in the free energy surface ([\[figljfes\]](#figljfes){reference-type="ref" reference="figljfes"}(c)). We consider now the more complex case of Alanine Tripeptide in vacuum as described by the Charmm22\(^\star\) forcefield, a protein that can have different conformations separated by moderately high energy barriers. These results were obtained with DM using 4 replicas with interaction strengths \(\sigma =\) 0.01, 0.05, 0.08, and 0.15 Å  plus one additional Boltzmann replica (Eq. [\[zrew2\]](#zrew2){reference-type="eqref" reference="zrew2"}) with \(\sigma_0=\sigma_1\). The temperature was \(T = 300\) K and the simulation ran for 100 ns with a timestep of \(dt = 2\) fs. Each 2 ps a Gaussian was deposited with initial height \(w_0 = 300\) K and bias factor \(\gamma = 4.3\), the standard deviations of the Gaussians depended on the replica index and were \(\sigma_g =\) 33.5, 22.3, 13.4, 6.7 and 4.5 meV. Swaps between replicas were attempted every 5 ps. We calculate a reference free energy surface by using the Variationally Enhanced Sampling (VES) of Ref. in the Well-Tempered variant of Ref. . We employed as CVs the three dihedral angles \(\Phi_{1},\Phi_{2},\Phi_{3}\) and expanded the bias potential in a Fourier series of size 6 for each CV, resulting a total number of 2196 basis functions. To optimize the VES functional we used the method of Bach. The coefficients were updated every 1 ps and a fixed step size of 0.08 kJ/mol was used. In order to achieve a Well-Tempered distribution we use a bias factor of 10 and the target distribution was updated every 500 ps. The simulation cost was equivalent to that of a 50 ns Molecular Dynamics run. The results of these two calculations are compared in [\[ala1d\]](#ala1d){reference-type="ref" reference="ala1d"} in which the free energy as function of the dihedral angle \(\phi_1\) and of \(\phi_2\) are shown along with the definition of the angles. The agreement between the two calculations is excellent and we note that strictly speaking use of the rigorously Boltzmannian replica \(Z_0\) is not necessary. Also the correlations between \(\phi_1\) and \(\phi_2\) ([\[ala2d\]](#ala2d){reference-type="ref" reference="ala2d"}) are well represented as well as the location and population of the different conformers. We underline the fact that in DM as opposed to VES no CVs need to be introduced. # Conclusions Taking inspiration from de Broglie swapping Metadynamics and from Ref. , we have used artificial delocalization effects to enhance sampling of Boltzmann systems. The delocalization has been obtained by mapping each particle into a dimer in which atoms are bound by an anharmonic potential. The computational cost relative to previous simulation methods is reduced. In fact here we deal only with dimers and not with polymers as in Ref. and also the more gentle behavior of \(f_\sigma^{\frac{1}{2}}(r)\) at large distances favors conformational swaps reducing the number of replicas needed. Like previous Path-Integral-based methods and PT, DM does not require choosing a CVs. Furthermore it offers a natural way of enhancing sampling of only a part of the system. For instance the conformational landscape of a mobile loop in a large protein could be selectively targeted.
{'timestamp': '2016-07-19T02:06:58', 'yymm': '1607', 'arxiv_id': '1607.04846', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04846'}
{'timestamp': '2016-07-18T02:08:40', 'yymm': '1607', 'arxiv_id': '1607.04580', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04580'}
null
null
null
null
null
null
# Introduction The classical problem to determine the rational points of the modular curve \(X_0(p)\), when \(p\) is a prime number, has been solved by Mazur in his paper in 1975. In his work, Mazur proved, for \(p>37\), that the only rational points of the curve \(X_0(p)\) are cusps and CM points. In the same paper Mazur also refers to the interesting case of the modular curve \(X_0^+(p)\). This curve is the quotient of \(X_0(p)\) by the Atkin-Lehner involution \(w_p\). See Galbraith. For these curves a similar result is expected: for sufficiently large \(p\) the only rational points of the curve \(X_0^+(p)\) are cusps and CM points. The genus \(g\) of \(X_0^+(p)\) growth with \(p\), so for a fixed \(g\) we have only finitely many curves. The cases \(g\) equal to \(0\) or \(1\) are well known. If \(g\) is equal to \(0\), then \(X_0^+(p)\) is isomorphic to \(\PP^1\) and there are infinitely many rational points. If \(g\) is equal to \(1\), then \(X_0^+(p)\) is an elliptic curve whose Mordell-Weil group is torsion-free and has rank \(1\). Let us focus on the case \(g>1\). Galbraith in his papers and describes a method to find explicit equations for the curves \(X_0^+(p)\). If \(X_0^+(p)\) is hyperelliptic he uses an *ad hoc* technique. We recall that by that the curve \(X_0^+(p)\) is hyperelliptic if and only if \(g=2\). If \(X_0^+(p)\) is not hyperelliptic, Galbraith uses the canonical embedding \(\varphi\colon X_0^+(p) \hookrightarrow \PP^{g-1}\) to find equations for the curves. He finds explicit equations for all curves \(X_0^+(p)\) with \(g\le 5\). Here we present the following result. Equations for the thirteen curves of genus \(6\) or \(7\) are listed in Section [6](#sec:listeq){reference-type="ref" reference="sec:listeq"}. The equations that we find have very small coefficients and have good reduction for each prime number \(\ell\neq p\). The technique used here play an important role in in finding explicit equations defining the modular curves \(X_{ns}^+(p)\). These curves are the modular curves associated to the normalizer of a nonsplit Cartan subgroup of the general linear group over a finite field with \(p\) elements. Few is still known about them and they are harder to study with respect to modular curves \(X_0^+(p)\) considered here. Galbraith uses the equations of the curves \(X_0^+(p)\) to look for rational points \(P\) up to some bounds on the naive projective height \(H\). See Silverman for more details about the naive height on projective space. Using a projection onto a plane model, possibly singular, we check that if \(X_0^+(p)\) has genus \(g=6\) or \(7\), then it has no rational points \(P\), except the cusps and the CM points, for which \(H(P)\le 10^4\). When the jacobian variety \(J_0^+(p)\) of \(X_0^+(p)\) has an isogeny factor over \(\QQ\) of dimension 1, we improve this bound significantly. In particular we prove the following result. # The method to get explicit equations {#sec:algoeq} In this section we show how to get explicit equations of the canonical model for modular curves. The final output is a list of quadrics with very small integer coefficients. Let \(\Gammait\) be a congruence subgroup of \(\SL_2(\ZZ)\), let \(\cfs_2(\Gammait)\) be the \(\CC\)-vector space of the cusp forms of weight \(2\) with respect to \(\Gammait\), and let \(f_1,\ldots,f_g\) be a \(\CC\)-basis of \(\cfs_2(\Gammait)\) where \(g\) is the genus of the modular curve \(X(\Gammait)\). If \(g>2\), we use the canonical embedding to get the canonical model for \(X(\Gammait)\). We know that \(X(\Gammait)\) is isomorphic as compact Riemann surfaces to \(\Gammait \backslash \HH^*\), where \(\HH^*\) is the extended complex upper half plane. We also know that \(\cfs_2(\Gammait)\) is isomorphic to the \(\CC\)-vector space of holomorphic differentials \(\Omega^1(X(\Gammait))\) with respect to the map \(f(\tau)\mapsto f(\tau)d\tau\). See Diamond and Shurman for more details about these results. Using these isomorphisms, the canonical embedding can be expressed in the following way \[\begin{aligned} \varphi\colon X(\Gammait) & \to \PP^{g-1}(\CC) \\ \Gammait\tau & \mapsto (f_1(\tau):\ldots:f_g(\tau)), \end{aligned}\] where \(\tau\in\HH^*\). The Enriques-Petri Theorem (see Griffith, Harris or Saint-Donat ), states that the canonical model of a complete nonsingular non-hyperelliptic curve is entirely cut out by quadrics and cubics. Moreover, it is cut out by quadrics if it is neither trigonal, nor a quintic plane curve with genus exactly \(6\). When \(X(\Gammait)\) can be defined over \(\QQ\), we can look for equations defined over \(\QQ\). The Enriques-Petri Theorem is proved over algebraically closed fields, but its proof can be suitable modified to work over \(\QQ\). Alternatively, if we find equations defined over \(\QQ\), then we can check by MAGMA that their zero locus \(Z\) is an algebraic curve with the same genus as \(X(\Gammait)\). An application of the Hurwitz genus formula for genus \(g>1\), to the morphism \(\varphi\colon X(\Gammait)\to Z\) implies that \(\varphi\) is an isomorphism. Finding equations for \(Z\) is equivalent to finding generators of the homogeneous ideal \(\mathcal{I}\) defining it in \(\PP^{g-1}\). Let \(\mathcal{I}_d\) be the set of homogeneous elements of \(\mathcal{I}\) of degree exactly \(d\). We explain how to find generators of \(\mathcal{I}\) that belong to \(\mathcal{I}_d\) for a fixed \(d\). By the Enriques-Petri Theorem and by Hasegawa and Hashimoto, we know that if \(g>2\), the ideal \(\mathcal{I}\) is generated by elements in \(\mathcal{I}_d\) for \(d=2,3\). Now, we fix the degree \(d\) and we suppose that we know the first \(m\) Fourier coefficients of the \(q\)-expansions of a basis \(\mathcal{B}=\{f_1,\dots,f_g\}\) of \(\Omega^1(X(\Gammait))\), where \(m>d(2g-2)\). This condition on \(m\) guarantees that if we have a polynomial \(F\) with rational coefficients and \(g\) unknowns such that \(F(f_1,\ldots,f_g)\equiv 0 \pmod{q^{m+1}}\), then \(F(f_1,\ldots,f_g)=0\). See for more details. We also assume that the Fourier coefficients of the basis \(\mathcal{B}\) are algebraic integers. There is a number field \(K\), of degree \(D\) over \(\QQ\), which contains all the coefficients of all the elements of \(\mathcal{B}\). Moreover, if the Fourier coefficients are algebraic integers, they have integer coordinates with respect to a suitable chosen basis of \(K\) over \(\QQ\). If \(a_n(f_i)\) is the \(n\)-th Fourier coefficient of \(f_i\), we denote the \(k\)-th coordinate of \(a_n(f_i)\) in \(K\) by \(c(i,n,k)\in\ZZ\), where \(k=1,\ldots,D\). We can associate to \(\mathcal{B}\) a matrix \(M\) whose integer entries are the coordinates of the Fourier coefficients where each row \(i\) corresponds to an element \(f_i\) of \(\mathcal{B}\). An explicit way to set the entries of \(M=(u_{ij})\) is to choose \(u_{ij}=c(i,n,k)\), where \(j=(n-1)D+k,\) for \(i=1,\ldots,g\), for \(n=1,\dots,m\) and for \(k=1,\ldots,D\). The matrix \(M\) has \(g\) rows and \(mD\) columns, where \(m\) is the number of the first Fourier coefficients used and \(g\) is the cardinality of the basis. Since \(m>d(2g-2)\) and \(g>0\), we always have \(m\geq g\), hence the rank of \(M\) is \(g\). To find equations defining \(X(\Gammait)\) we compute all the monomials \(F_j\) of degree \(d\) where the indeterminates are the elements of \(\mathcal{B}\). The elements of a \(\ZZ\)-basis of the space \(S\) of the solutions of the homogeneous linear system in the unknowns \(F_j\) are the coefficients of the desired equations. We are interested in reducing the size of the coefficients of these equations and minimizing the number of primes \(\ell\) such that the model has bad reduction modulo \(\ell\). To reduce the size of the coefficients, we apply the LLL-algorithm first to \(M\) and then to the \(\ZZ\)-basis of \(S\). We know that if the rank of \(M\) modulo \(\ell\) is less than \(g\), then the canonical model of the curve that we find, is singular modulo \(\ell\). We say that \(M\) is *optimal*, if the rank of \(M\) modulo \(\ell\) is exactly \(g\) for each prime \(\ell\). In Algorithm [\[alg:badprimes\]](#alg:badprimes){reference-type="ref" reference="alg:badprimes"} below, within a more general setting, we describe how to modify \(\mathcal{B}\) to make \(M\) optimal. Now, we describe how to find the primes \(\ell\) such that the canonical model has bad reduction. Let \(\mathcal{I}_{\text{jac}}\) be the ideal generated by the polynomials defining the curve and by all the determinants of order \(g-2\) of the jacobian matrix of the curve. We compute the elimination ideals \(\mathcal{J}_i:=\mathcal{I}_{\text{jac}}\cap \QQ[x_i]\), for \(i=1,\ldots,g\). If \(\mathcal{J}_i\neq 0\), it turns out to be generated by \(\lambda x_i^n\) with \(\lambda,n\in\ZZ_{>0}\). The curve has bad reduction modulo any prime \(\ell\) such that \(\ell\mid \lambda\). ## Algorithm Let \(g\) and \(m\) be positive integers such that \(g\leq m\) and let \(v_1,\ldots,v_g\in\ZZ^m\) be linearly independent vectors over \(\QQ\). We describe a method to find a basis over \(\ZZ\) of the lattice \(L:=\Span_{\QQ}\{v_1,\ldots,v_g\} \cap \ZZ^m\). Let \(L':=\Span_{\ZZ}\{v_1,\ldots,v_g\}\), so \(L'\) is a subgroup of \(L\), and let \(J:=[L:L']\). We want to modify \(L'\) and its basis until we have \(J=1\) and so \(L'=L\). Let \(M\in\ZZ^{m\times g}\) be the matrix whose columns are the vectors \(v_1,\ldots,v_g\). It follows from Lemma [\[lem:pdep\]](#lem:pdep){reference-type="ref" reference="lem:pdep"} above that if a prime \(p\) divides the index \(J\), then \(p\) divides the determinant \(\Delta\) of any \(g\times g\) submatrix of \(M\). # Computing the expected rational points We have a simple moduli interpretation of certain rational points of the modular curves \(X_0^+(p)\). The *expected* rational points on \(X_0^+(p)\) are the points corresponding to the unique cusp or to elliptic curves with complex multiplication such that \(p\) is split or ramified inside the endomorphism ring of the elliptic curve itself. We call a rational point *exceptional* if it is not one of the expected ones. Here we describe a way to find numerically the expected rational points on a modular curve \(X_0^+(p)\), for \(p\) an odd prime. We assume to know the first \(m\) Fourier coefficients of a basis \(f_1,\ldots,f_g\) of \(\Omega^1(X_0^+(p))\), where \(g\) is the genus of the modular curve. It is well known, see Stark, that the orders in imaginary quadratic number fields with class number \(1\) are the ones with discriminant \[\Delta=-3,-4,-7,-8,-11,-12,-16,-19,-27,-28,-43,-67,-163.\] The corresponding orders can be written in the form \(\ZZ+\ZZ\tau\), where \[\tau=\begin{cases} \frac{1+i\sqrt{|\Delta|}}{2} & \text{if }\Delta \text{ is odd}, \\ \frac{i\sqrt{|\Delta|}}{2} & \text{if }\Delta \text{ is even}. \end{cases}\] Let \(E\) be an elliptic curve with complex multiplication such that the discriminant \(\Delta_E\) of its endomorphism ring \(\mathcal{O}_E\) is a class number 1 discriminant. This means that \(\Delta_E\) is in the list above. If \(\Delta_E\) is a square modulo \(p\) we can associate to \(E\) a point on \(X_0^+(p)\), see Galbraith. The unique cusp is always rational. Let \(\{1,\tau_E\}\) be the basis of the order \(\mathcal{O}_E\) such that \(\tau_E\) is defined as above. There is a suitable element \(\hat{\tau}\) in the \(\SL_2(\ZZ)\)-orbit of \(\tau_E\) in \(\HH\) such that \(P=(f_1(\hat{\tau}):\ldots:f_g(\hat{\tau}))\) is a rational point for the modular curve. If we know \(\hat{\tau}\), we can evaluate \(f_1(\hat{\tau}),\ldots,f_g(\hat{\tau})\) numerically using the \(q\)-expansions. If \(m\) is large enough, it is quite easy recognize the rational coordinates of \(P\). To compute the coordinates of the cusp it is enough take \(\hat{\tau}=i\infty\), this means to set \(q=0\) in the \(q\)-expansions. Now we explain how to compute \(\hat{\tau}\) for the CM points. We recall the moduli interpretation of the points of \(X_0^+(p)\). Let \(E'\) be an elliptic curve and let \(C\) be a cyclic subgroup of order \(p\) of the \(p\)-torsion subgroup \(E'[p]\). We know that a point on \(X_0^+(p)\) is an unordered pair \(\{(E',C),(E'/C,E'[p]/C)\}\), where \(E'/C\) is the unique elliptic curve, up to isomorphism, that is the image of the unique isogeny of \(E'\) with kernel \(C\). Over \(\CC\) this is \(\{(\CC/\Lambda,\frac{1}{p}\ZZ+\Lambda),(\CC/\Lambda',\frac{1}{p}\ZZ+\Lambda')\}\) for some \(\tau\in\HH\) such that \(\Lambda=\ZZ+\tau\ZZ\) and \(\Lambda'=\ZZ+\frac{-1}{p\tau}\ZZ\). If \(\Delta_E\) is a square modulo \(p\), we know there is a principal prime ideal \(\mathfrak{p}\) of \(\mathcal{O}_E\) such that \((p)=\mathfrak{p}\bar{\mathfrak{p}}\). Let \(\alpha\) and \(\bar{\alpha}\) be generators of \(\mathfrak{p}\) and \(\bar{\mathfrak{p}}\) respectively. The cyclic subgroups \(C\) and \(E'[p]/C\) are the kernels of the multiplication by \(\alpha\) and \(\bar{\alpha}\). Let \((E',C)=(\CC/\Lambda_E,\frac{1}{\bar{\alpha}}\ZZ+\Lambda_E)\), where \(\Lambda_E=\ZZ+\tau_E\ZZ\). The point \(\{(\CC/\Lambda_E,\frac{1}{\bar{\alpha}}\ZZ+\Lambda_E),(\CC/\Lambda_E,\frac{1}{\alpha}\ZZ+\Lambda_E)\}\) is rational on \(X_0^+(p)\). Now, we want to find \(\hat{\tau}\) such that \((\CC/\Lambda_E,\frac{1}{\bar{\alpha}}\ZZ+\Lambda_E)=(\CC/\hat{\Lambda},\frac{1}{p}\ZZ+\hat{\Lambda})\), where \(\hat{\Lambda}=\ZZ+\hat{\tau}\ZZ\). Let \(c,d\in\ZZ\) such that \(\alpha=c\tau_E+d\). Then we can find \(a,b\in\ZZ\) such that \(ad-bc=1\) and so we have a matrix \(\hat{\gamma}=\begin{pmatrix} a & b \\ c & d \end{pmatrix}\in\SL_2(\ZZ)\). The transformation \(\hat{\tau}=\hat{\gamma} \tau_E\) correspond to the isogeny \(z+\Lambda_E\mapsto \frac{z}{\alpha}+\hat{\Lambda}\) for every \(z+\Lambda_E\in\CC/\Lambda_E\). Hence \(\frac{1}{\bar{\alpha}}+\Lambda_E\mapsto \frac{1}{\alpha\bar{\alpha}}+\hat{\Lambda}=\frac{1}{p}+\hat{\Lambda}\) and the group \(\frac{1}{\bar{\alpha}}\ZZ+\Lambda_E\) maps to \(\frac{1}{p}\ZZ+\hat{\Lambda}\) and we are done. # Computing the modular parametrization {#sec:compEmap} In this section we assume to know the first \(m\) Fourier coefficients of the modular forms involved, where \(m\) is \"large enough\" for our purposes. Let \(p\) be a prime number and let \(J_0(p)\) be the jacobian variety of the modular curve \(X_0(p)\). If there is an isogeny factor \(E\) over \(\QQ\) of dimension \(1\) of \(J_0(p)\), we know there is a non-constant morphism \(\phi\colon X_0(p)\to E\) defined over \(\QQ\). Let \(f\) be the normalized eigenform in \(\cfs_2(\Gammait_0(p))\) associated to the isogeny class of \(E\). Let \(\mathds{T}\) be the Hecke algebra over \(\ZZ\) and let \(I_f=\{T\in\mathds{T}: Tf=0\}\). We know that if the Fourier coefficients of \(f\) belong to \(\QQ\), the associated abelian variety \(J_0(p)/I_f J_0(p)\) is an elliptic curve called *optimal curve* (or *strong Weil curve*). See Diamond and Shurman for more details about these topics. If \(E=J_0(p)/I_f J_0(p)\), then we call \(\phi\) the *modular parametrization with respect to \(X_0(p)\)*. The degree of the morphism \(\phi\) is called the *modular degree*. Using the identification of \(X_0(p)\) with \(\Gammait_0(p)\backslash\HH^*\), we can write \[\begin{aligned} \phi\colon \Gammait_0(p)\backslash\HH^* &\to E(\CC) \\ \Gammait_0(p)\tau &\mapsto \phi(\tau)=(\phi_x(\tau),\phi_y(\tau)), \end{aligned}\] where \(\tau\in\HH^*\) and \(\phi_x(\tau)\) and \(\phi_y(\tau)\) are Fourier series with rational coefficients in the indeterminate \(q=e^{2\pi i \tau}\). Let \(X_0^+(p)\) a modular curve of genus \(g\) with an isogeny factor \(E\) over \(\QQ\) of dimension \(1\) of the jacobian variety \(J_0^+(p)\). We denote by \(f_1,\ldots,f_g\) some \(g\) linearly independent newforms of \(\cfs_2(\Gammait_0(p))\) with eigenvalue \(+1\) with respect to the Atkin-Lehner operator \(w_p\). This is a basis of \(\cfs_2(\Gammait^+_0(p))\). The modular parametrization \(\phi_+\colon X_0^+(p)\to E\) is locally given by four homogeneous polynomials \(p_x,q_x,p_y,q_y\in\ZZ[x_1,\ldots,x_g]\), such that \(p_x\) and \(q_x\) have the same degree, \(p_y\) and \(q_y\) have the same degree and \[\begin{aligned} \label{eq:sistema} \begin{sistema} \phi_x(\tau)= \frac{p_x(f_1(\tau),\ldots,f_g(\tau))}{q_x(f_1(\tau),\ldots,f_g(\tau))} \\ \phi_y(\tau)= \frac{p_y(f_1(\tau),\ldots,f_g(\tau))}{q_y(f_1(\tau),\ldots,f_g(\tau))}, \end{sistema} \end{aligned}\] for every \(\tau\) representative of \(\Gammait_0(p)\tau\) chosen in a suitable open set of \(X_0(p)\). To find these polynomials explicitly we use some linear algebra. The main difficulty is that we don't know exactly the degrees of the polynomials \(p_x,q_x,p_y,q_y\), but we know the degree of the morphism \(\phi_+\). In Cremona or in the LMFDB online database we can find the value of \(\deg\phi\) and we have \(\deg\phi_+=\frac{1}{2}\deg\phi\). For increasing degrees from \(1\) to \(\deg{\phi_+}\), we apply the following procedure to look for nontrivial relations. We rewrite the first equation as \[\label{eq:ctrimg} p_x(f_1(\tau),\ldots,f_g(\tau))-q_x(f_1(\tau),\ldots,f_g(\tau))\phi_x(\tau)=0.\] Let \(r=\begin{pmatrix} g +d_x-1 \\ d_x \end{pmatrix}\) be the number of monomials of degree \(d_x\). Let \(\mathcal{I}=\{F_1,\ldots,F_r\}\) be the set of the monomials obtained as product of \(d_x\) elements of the basis \(f_1,\ldots,f_g\), where repetitions are allowed, and let \(\mathcal{J}=\{G_1,\ldots,G_r\}\) be the set with the same elements of \(\mathcal{I}\) multiplied by \(\phi_x\). Since the Fourier coefficients of \(\phi_x,f_1,\ldots,f_g\) are algebraic numbers, the coefficients of the elements of \(\mathcal{I}\) and \(\mathcal{J}\) are also algebraic numbers and there is a number field \(K_f\) of finite degree over \(\QQ\), which contains all these Fourier coefficients. Multiplying the elements of \(\mathcal{I}\) and \(\mathcal{J}\) by a suitable algebraic number, we can assume that the Fourier coefficients belong to the ring of integers of \(K_f\) and that they have integer coordinates with respect to a suitable basis of \(K_f\). Hence, there are \(\alpha_1,\ldots,\alpha_r,\beta_1,\ldots,\beta_r\in\ZZ\) such that \[\begin{sistema} \alpha_1 F_1+\ldots+\alpha_r F_r+\beta_1 G_1+\ldots+\beta_r G_r=0, \\ \alpha_1 F_1+\ldots+\alpha_r F_r\neq 0, \\ \beta_1 G_1+\ldots+\beta_r G_r\neq 0. \end{sistema}\] These \(\alpha_i\)'s and \(\beta_i\)'s are the coefficients of \(p_x\) and \(q_x\) respectively. Different choices of \(\alpha_1,\ldots,\alpha_r,\beta_1,\ldots,\beta_r\) correspond to the same map defined on different open sets. The same method is applied to the second equation of ([\[eq:sistema\]](#eq:sistema){reference-type="ref" reference="eq:sistema"}) to find the coefficients of \(p_y\) and \(q_y\). # An estimation on heights {#sec:upbnd} Let \(n\) be a positive integer and let \(X\) be a complete nonsingular curve defined over \(\QQ\) in \(\PP^{n-1}\) and with an elliptic curve \(E\) as isogeny factor over \(\QQ\) of its jacobian variety. In this section we suppose to know a Weierstrass equation of \(E\), an explicit formula for the morphism \(\phi\colon X \to E\) defined over \(\QQ\) and generators for the Mordell-Weil group of \(E\). Writing \(\phi\) in more explicit terms, we have \[\phi(x_1,\ldots,x_n)=\left(\frac{p_x(x_1,\ldots,x_n)}{q_x(x_1,\ldots,x_n)}, \frac{p_y(x_1,\ldots,x_n)}{q_y(x_1,\ldots,x_n)}\right),\] where \(p_x,q_x\) are homogeneous polynomials with integer coefficients of degree \(d_x\) and \(p_y,q_y\) are homogeneous polynomials with integer coefficients of degree \(d_y\). Let \(r\) be the number of monomials of degree \(d_x\) and let \(\alpha_i\) and \(\beta_i\), for \(i=1,\ldots, r\), be the coefficients of \(p_x\) and \(q_x\) respectively. Moreover, let \(\alpha:=\log\max\left\{\sum_i\left|\alpha_i\right|,\sum_i\left|\beta_i\right|\right\}\) a constant that we use in the Proposition [\[pro:upbnd\]](#pro:upbnd){reference-type="ref" reference="pro:upbnd"} below. In the paper, Silverman gives an estimate of the difference between the canonical height and the Weil height on an elliptic curve. We use the Theorem 1.1 of to prove the following proposition. # Tables {#sec:listeq} In [6.1](#subsec:eq){reference-type="ref" reference="subsec:eq"} we list the equations for the canonical model of the modular curves \(X_0^+(p)\) for primes \(p\) such that the genus \(g\) is \(6\) or \(7\). We also list the expected rational points. The models all have good reduction modulo every prime \(\ell\neq p\). In [6.2](#subsec:modpar){reference-type="ref" reference="subsec:modpar"} we give an explicit morphism \(\phi_+\colon X_0^+(p)\to E\), whenever the jacobian variety \(J_0^+(p)\) of \(X_0^+(p)\) admits a \(1\)-dimensional isogeny factor \(E\) over \(\QQ\). ## Equations {#subsec:eq} Here we list the genus, the equations of the canonical model and the expected rational points with the corresponding discriminants. - Curve \(X_0^+(163)\). Genus \(g=6\). Equations of the canonical model in \(\PP^{5}\) \[\begin{sistema} x_1x_5+x_2x_3+x_2x_4-x_2x_5+x_2x_6=0 \\ x_1^2-x_1x_2+x_1x_5-x_2x_5+x_3^2+x_3x_4=0 \\ x_2x_4-x_2x_5+x_3x_5=0 \\-x_1^2+x_1x_2-x_1x_4+x_2x_4 + x_3x_6=0 \\ x_1x_3+x_1x_5+x_1x_6+x_2x_3-x_2x_5+x_4x_5=0 \\-x_1^2+x_1x_2+x_1x_3+x_1x_4-x_1x_5+x_1x_6+x_2x_5-x_3^2+x_4^2+x_5x_6=0. \end{sistema}\] \[\begin{array}{|cc|cc|} \toprule \text{Rational point} & \text{Disc.} & \text{Rational point} & \text{Disc.} \\ \midrule (0:0:0:0:0:1) & \text{cusp} & (0:1:1:0:1:0) &-12 \\ (24:10:13:15:-50:42) &-163 & (0:1:0:0:0:0) &-11 \\ (1:0:1:-2:0:-1) &-67 & (0:0:1:-1:0:0) &-8 \\ (1:1:2:-2:2:0) &-28 & (1:1:0:0:0:0) &-7 \\ (1:1:0:1:1:-1) &-27 & (2:-1:3:-4:-1:-2) &-3 \\ (0:0:0:0:1:0) &-19 & & \\ \bottomrule \end{array}\] - Curve \(X_0^+(193)\). Genus \(g=7\). Equations of the canonical model in \(\PP^{6}\) \[\begin{sistema}-x_1^2-x_1x_4-x_1x_6-x_1x_7+x_2x_7+x_3^2+x_3x_4=0 \\ x_1x_2+x_1x_4-x_1x_5+x_1x_7-x_2x_3-x_2x_4-x_2x_7+x_3x_5=0 \\ x_1x_2-x_1x_5-x_2x_7+x_3x_6=0 \\-x_1x_2-x_1x_3+2x_1x_5+x_1x_6+2x_2x_3+2x_2x_4-x_3^2+x_4^2+x_4x_5=0 \\ x_1^2-x_1x_2+2x_1x_4+x_1x_5+2x_1x_6+x_1x_7-x_2x_7-x_3^2+x_4^2+x_4x_6=0 \\ x_1x_3-x_1x_4-x_1x_5-2x_1x_6-x_1x_7-2x_2x_3-2x_2x_4+2x_2x_7+x_3^2-x_4^2+ x_4x_7=0 \\ x_1^2+x_1x_3+x_1x_4-2x_1x_5+x_1x_7-x_2x_3-3x_2x_4-x_2x_5-x_2x_7-x_4^2+x_5^2=0 \\-x_1^2+2x_1x_2-x_1x_4-x_1x_5-2x_1x_6-x_1x_7-x_2x_3-x_2x_4+x_2x_7+x_3^2-x_4^2+ x_5x_6=0 \\-x_1x_3+x_1x_5+x_1x_6+2x_2x_3+2x_2x_4-x_2x_7-x_3^2+x_4^2+x_5x_7=0 \\-x_1x_4-x_1x_5-x_1x_7+x_2x_3-x_2x_4-x_2x_6+2x_2x_7-x_3x_7-x_4^2+x_6^2=0. \end{sistema}\] \[\begin{array}{|cc|cc|} \toprule \text{Rational point} & \text{Disc.} & \text{Rational point} & \text{Disc.} \\ \midrule (0:0:0:0:0:0:1) & \text{cusp} & (1:0:-1:0:0:0:0) &-12 \\ (1:1:1:-1:2:0:-1) &-67 & (0:1:0:0:0:1:0) &-8 \\ (0:0:0:1:-1:-1:1) &-43 & (0:1:0:0:0:0:0) &-7 \\ (0:1:2:-2:0:0:0) &-28 & (1:0:0:1:0:-1:-1) &-4 \\ (0:1:0:0:1:0:0) &-27 & (3:2:3:0:2:0:0) &-3 \\ (1:0:0:-1:0:-1:1) &-16 & & \\ \bottomrule \end{array}\] - Curve \(X_0^+(197)\). Genus \(g=6\). Equations of the canonical model in \(\PP^{5}\) \[\begin{sistema}-x_1x_2+x_1x_4-x_1x_5-x_1x_6+x_2x_3=0 \\ x_1^2-x_1x_3-x_1x_4+x_1x_5+x_2x_4=0 \\ 2x_1x_2-2x_1x_3-2x_1x_4+x_1x_5+x_1x_6-x_2^2-x_2x_6+x_3^2+x_3x_4+x_4^2=0 \\-x_1^2+x_1x_2+x_1x_3+x_1x_6-x_2^2-x_2x_6+x_4x_5=0 \\ x_1x_2-x_1x_3-x_1x_4+x_3x_5+x_4x_6=0 \\-x_1x_2+x_1x_3+x_1x_4-x_1x_6+x_2x_5-x_2x_6-x_3x_5+x_3x_6+x_5^2=0. \end{sistema}\] \[\begin{array}{|cc|cc|} \toprule \text{Rational point} & \text{Disc.} & \text{Rational point} & \text{Disc.} \\ \midrule (0:0:0:0:0:1) & \text{cusp} & (1:1:1:0:0:0) &-19 \\ (1:-1:3:2:6:-6) &-163 & (1:0:0:0:-1:1) &-16 \\ (1:1:1:1:0:1) &-43 & (1:0:0:1:0:1) &-7 \\ (1:0:2:-1:0:-1) &-28 & (1:0:2:0:1:-1) &-4 \\ \bottomrule \end{array}\] - Curve \(X_0^+(211)\). Genus \(g=6\). Equations of the canonical model in \(\PP^{5}\) \[\begin{sistema}-x_1^2-x_1x_4-x_1x_5+x_2x_3-x_2x_6+x_3x_4=0 \\-x_1x_3-x_1x_4-x_1x_5+x_2x_3-x_2x_6+x_3x_5=0 \\ x_1^2-x_1x_2+x_1x_3+x_1x_4-x_2x_3+x_3x_6=0 \\ 2x_1^2-2x_1x_2+x_1x_3+3x_1x_4+x_1x_5-2x_2x_3-x_2x_4+2x_2x_6+x_4^2=0 \\ x_1x_3+x_1x_4+2x_1x_5-x_1x_6-2x_2x_3-x_2x_5+2x_2x_6+x_4x_5=0 \\ x_1^2-2x_1x_2+x_1x_3+2x_1x_4+x_1x_5-x_1x_6-2x_2x_3-x_2x_4+x_2x_5+2x_2x_6-x_4x_6+ x_5^2=0. \end{sistema}\] \[\begin{array}{|cc|cc|} \toprule \text{Rational point} & \text{Disc.} & \text{Rational point} & \text{Disc.} \\ \midrule (0:0:0:0:0:1) & \text{cusp} & (0:1:1:0:0:1) &-12 \\ (1:0:-1:-1:1:-1) &-67 & (0:0:1:0:0:0) &-8 \\ (2:1:2:0:0:-2) &-28 & (0:1:0:0:0:0) &-7 \\ (1:0:1:-1:-1:-1) &-27 & (2:-3:-1:-2:4:1) &-3 \\ \bottomrule \end{array}\] - Curve \(X_0^+(223)\). Genus \(g=6\). Equations of the canonical model in \(\PP^{5}\) \[\begin{sistema}-x_1^2+x_1x_4-x_2x_3-x_2x_4+x_2x_5=0 \\ x_1^2+2x_1x_3-x_1x_4+x_1x_5+x_2^2+2x_2x_4+x_2x_6+x_3^2=0 \\-x_1^2-x_1x_2-x_1x_3+x_1x_4-x_2x_4+x_2x_6+x_3x_4=0 \\ x_1^2-x_1x_2+2x_1x_5-x_1x_6+x_2^2+x_2x_4+2x_2x_6+x_3x_5=0 \\-x_1^2-x_1x_2+x_1x_3+x_1x_4+x_2x_6-x_3x_6+x_4x_5=0 \\-2x_1^2-2x_1x_2+x_1x_3+3x_1x_4+2x_1x_5+x_1x_6+x_2^2+4x_2x_6-x_3x_6-x_4^2-x_4x_6+ x_5^2=0. \end{sistema}\] \[\begin{array}{|cc|cc|} \toprule \text{Rational point} & \text{Disc.} & \text{Rational point} & \text{Disc.} \\ \midrule (0:0:0:0:0:1) & \text{cusp} & (0:1:1:-1:0:0) &-12 \\ (2:-2:2:3:6:7) &-163 & (0:0:0:1:0:-1) &-11 \\ (1:0:-2:1:0:1) &-67 & (2:-3:-3:-1:-6:2) &-3 \\ (1:0:0:1:0:1) &-27 & & \\ \bottomrule \end{array}\] - Curve \(X_0^+(229)\). Genus \(g=7\). Equations of the canonical model in \(\PP^{6}\) \[\begin{sistema}-x_1x_5+x_2x_4-x_2x_6+x_2x_7=0 \\ x_1x_2-x_1x_3-x_1x_4-x_1x_5+x_2x_3+x_2x_4-x_2x_6+x_3^2+x_3x_4=0 \\-x_1x_5+x_2^2+x_2x_3+x_2x_4-x_2x_5-x_2x_6+x_3x_5=0 \\ x_1x_2-x_1x_3-x_1x_4-x_1x_6-x_2^2+x_2x_4+x_3^2+x_3x_6=0 \\ x_1x_2+x_1x_3-x_1x_6-x_1x_7-x_2^2-x_2x_3+x_2x_6+x_3x_7=0 \\-2x_1x_2+x_1x_4+x_1x_5+x_1x_6+x_1x_7-x_2x_3-x_2x_4+x_2x_6-x_3^2+x_4^2=0 \\-x_1x_3+x_1x_6-x_2^2-x_2x_3-x_2x_4+x_2x_5+x_4x_5=0 \\-x_1x_2+x_1x_5+x_1x_7+x_2x_4-x_2x_5+x_5x_6=0 \\ x_1x_2-2x_1x_5-x_1x_6-x_1x_7+x_2x_3+2x_2x_4-x_2x_5-x_2x_6+x_3^2-x_4x_6+x_5x_7=0 \\ x_1x_5-x_2x_3-x_2x_4-x_3^2-x_4x_7+x_6^2=0. \end{sistema}\] \[\begin{array}{|cc|cc|} \toprule \text{Rational point} & \text{Disc.} & \text{Rational point} & \text{Disc.} \\ \midrule (0:0:0:0:0:0:1) & \text{cusp} & (0:1:-1:0:0:0:0) &-12 \\ (1:1:0:1:1:0:0) &-43 & (1:0:0:0:0:0:0) &-11 \\ (1:0:1:0:0:1:0) &-27 & (2:0:1:-1:0:1:0) &-4 \\ (0:0:0:0:1:0:0) &-19 & (2:-3:-1:0:-6:-4:0) &-3 \\ (0:0:1:-1:0:-1:0) &-16 & & \\ \bottomrule \end{array}\] - Curve \(X_0^+(233)\). Genus \(g=7\). Equations of the canonical model in \(\PP^{6}\) \[\begin{sistema}-x_1x_2-x_1x_3-x_1x_4+x_2^2-x_2x_3-x_2x_5-x_2x_6+x_2x_7=0 \\ x_1x_2+x_1x_4-x_2^2+x_2x_6+x_3x_5=0 \\ x_1^2-2x_1x_2-x_1x_3+x_1x_5+2x_2^2-2x_2x_3-x_2x_4-2x_2x_5-x_2x_6+x_3x_6=0 \\-2x_1x_3-x_1x_6+x_2^2-x_2x_3-x_2x_4-x_2x_5+x_3x_7=0 \\ x_1^2-2x_1x_2+x_1x_4+x_1x_5+2x_2^2-2x_2x_3-2x_2x_5-x_2x_6+x_3x_4+x_4^2=0 \\ x_1x_3+x_1x_6-x_2^2+x_2x_3+x_2x_5+x_4x_5=0 \\ x_1x_2+x_1x_3-x_1x_7-x_2^2+x_2x_3-x_2x_4+x_2x_5+x_2x_6+x_4x_6=0 \\-x_1x_2-x_1x_3-x_1x_4+x_1x_5-x_1x_6-x_1x_7+2x_2^2-x_2x_3-x_2x_4-2x_2x_5-x_2x_6+ x_5^2=0 \\-x_1^2+x_1x_2+2x_1x_3-x_1x_5+x_1x_6+x_1x_7-2x_2^2+2x_2x_3+x_2x_4+2x_2x_5+x_4x_7+ x_5x_6=0 \\-x_1^2+x_1x_2-x_1x_5+x_2x_4-x_2x_6+x_5x_7+x_6^2=0. \end{sistema}\] \[\begin{array}{|cc|cc|} \toprule \text{Rational point} & \text{Disc.} & \text{Rational point} & \text{Disc.} \\ \midrule (0:0:0:0:0:0:1) & \text{cusp} & (0:0:1:0:0:0:0) &-8 \\ (1:0:2:-2:1:0:2) &-28 & (1:0:0:0:-1:0:0) &-7 \\ (0:0:1:-1:0:0:0) &-19 & (1:1:0:-2:1:2:1) &-4 \\ (1:1:0:0:1:0:1) &-16 & & \\ \bottomrule \end{array}\] - Curve \(X_0^+(241)\). Genus \(g=7\). Equations of the canonical model in \(\PP^{6}\) \[\begin{sistema} x_1^2+x_1x_4-x_2^2-x_2x_4+x_2x_6=0 \\ x_1x_5-x_2x_3-x_2x_5+x_2x_7=0 \\ x_1x_2+2x_1x_3+x_1x_6-x_1x_7-x_2^2+x_3^2=0 \\ x_1x_3-x_1x_7+x_2x_4+x_3x_4=0 \\-x_1x_4-x_1x_5-x_2x_4+x_2x_5-x_3x_5-x_4^2+x_4x_5=0 \\-x_1^2+x_1x_2-x_1x_3+x_1x_4+x_1x_5+x_1x_7+x_2x_3-2x_2x_4-x_2x_5+x_3x_5-x_3x_6+ x_4x_6=0 \\-x_1^2-x_1x_3+x_1x_4-x_1x_6+x_2^2+x_2x_3-x_2x_4+x_3x_5-x_3x_6-x_3x_7+x_4x_7=0 \\-x_1x_2+x_1x_3-3x_1x_4-x_1x_5+x_1x_6+x_2^2+x_2x_4-x_3x_5+x_3x_6-x_4^2+x_5^2=0 \\-x_1^2-x_1x_3+x_1x_4+x_1x_5-x_1x_6+x_2^2+x_2x_3-2x_2x_4-x_2x_5+x_3x_5-x_3x_6-x_3x_7+x_5x_6=0 \\-x_1^2+x_1x_2-x_1x_3+x_1x_4+2x_1x_5+x_1x_7+x_2x_3-2x_2x_4-2x_2x_5+x_5x_7+ x_6^2=0. \end{sistema}\] \[\begin{array}{|cc|cc|} \toprule \text{Rational point} & \text{Disc.} & \text{Rational point} & \text{Disc.} \\ \midrule (0:0:0:0:0:0:1) & \text{cusp} & (1:1:0:0:1:0:0) &-12 \\ (1:0:-2:-1:0:0:0) &-67 & (0:0:0:1:1:0:0) &-8 \\ (0:1:-1:0:0:1:-1) &-27 & (1:-1:-2:0:0:0:-2) &-4 \\ (1:1:0:0:0:0:0) &-16 & (3:1:-4:-6:-3:4:2) &-3 \\ \bottomrule \end{array}\] - Curve \(X_0^+(257)\). Genus \(g=7\). Equations of the canonical model in \(\PP^{6}\) \[\begin{sistema}-x_1^2-x_1x_2+x_1x_4-x_2^2-x_2x_4+x_2x_5=0 \\-x_1^2-2x_1x_2+x_1x_3+x_1x_4-x_1x_5-2x_2^2-x_2x_4+x_2x_6=0 \\-x_1x_2-x_1x_6-x_2x_3+x_2x_7=0 \\ x_1x_2-x_1x_3+x_1x_6+x_2^2+x_2x_4+x_3x_4=0 \\-x_1x_3-x_1x_4+x_1x_6-x_1x_7+x_2^2-x_2x_3+x_2x_4+x_3x_5=0 \\ x_1^2+x_1x_2-x_1x_3-x_1x_4+2x_2^2-x_2x_3+2x_2x_4-x_3x_6+x_4x_5=0 \\ x_1^2+x_1x_2-2x_1x_4+x_1x_5-x_1x_7+2x_2^2-x_2x_3+x_2x_4+x_3^2-x_3x_6-x_3x_7+ x_4x_6=0 \\ x_1^2-x_1x_3-x_1x_4-x_1x_5-x_1x_6+x_2^2-x_2x_3+x_2x_4-x_3^2-x_3x_6+x_3x_7+ x_5^2=0 \\ x_1^2+x_1x_3-2x_1x_4-2x_1x_7+x_2^2-x_2x_3+x_2x_4+x_3^2-x_3x_6-x_3x_7+x_4x_7+ x_5x_6=0 \\-2x_1x_2+2x_1x_3-2x_1x_6-x_1x_7-2x_2^2-x_2x_4-x_3x_6+x_4x_7-x_5x_7+x_6^2=0. \end{sistema}\] \[\begin{array}{|cc|cc|} \toprule \text{Rational point} & \text{Disc.} & \text{Rational point} & \text{Disc.} \\ \midrule (0:0:0:0:0:0:1) & \text{cusp} & (0:0:1:0:0:0:1) &-11 \\ (1:0:0:1:0:0:-1) &-67 & (0:0:0:1:0:0:0) &-8 \\ (0:1:0:-1:0:1:0) &-16 & (2:-1:0:1:0:1:0) &-4 \\ \bottomrule \end{array}\] - Curve \(X_0^+(269)\). Genus \(g=6\). Equations of the canonical model in \(\PP^{5}\) \[\begin{sistema} x_1x_4+x_2^2+x_2x_3+x_3^2+x_3x_5=0 \\-x_1x_2-x_1x_5+x_2^2+x_2x_3+x_2x_5+x_3x_6=0 \\ x_1x_2+2x_1x_4+x_1x_6+x_3^2+x_3x_4+x_4^2=0 \\ x_1x_2-x_1x_4+x_1x_5-x_2^2-x_2x_3+x_2x_4+x_3x_4+x_4x_5=0 \\ x_1x_2+x_1x_3+x_1x_4+x_1x_5+x_1x_6-x_2x_3-x_2x_5+x_4x_6=0 \\-x_1x_2-x_1x_3-2x_1x_4-x_1x_5-x_1x_6-x_2^2+2x_2x_5-x_2x_6-x_3^2+x_5^2=0. \end{sistema}\] \[\begin{array}{|cc|cc|} \toprule \text{Rational point} & \text{Disc.} & \text{Rational point} & \text{Disc.} \\ \midrule (0:0:0:0:0:1) & \text{cusp} & (1:1:0:-1:0:0) &-16 \\ (1:-1:0:-1:1:2) &-67 & (1:0:0:0:0:0) &-11 \\ (1:1:0:-1:-1:0) &-43 & (1:-1:2:-3:0:0) &-4 \\ \bottomrule \end{array}\] - Curve \(X_0^+(271)\). Genus \(g=6\). Equations of the canonical model in \(\PP^{5}\) \[\begin{sistema} x_1^2+2x_1x_2+x_1x_5-x_1x_6-x_2x_4-2x_2x_5-x_2x_6+x_3x_5=0 \\ x_1^2+x_1x_2-x_1x_4-x_1x_6-x_2x_3-x_2x_4-x_2x_5-x_2x_6+x_4^2=0 \\-x_1^2-2x_1x_2-x_1x_5+x_1x_6+x_2x_3+x_2x_4+x_2x_5+x_4x_5=0 \\ x_1x_2-x_1x_4-x_1x_6-x_2x_3+x_2x_6+x_3^2+x_4x_6=0 \\ x_1^2+3x_1x_2+2x_1x_3+x_1x_5-3x_1x_6-2x_2x_4-x_2x_5+x_5^2=0 \\-x_1x_2-x_1x_3-x_1x_4-x_1x_5+x_1x_6-x_2x_3+x_2x_4+x_2x_5+x_2x_6+x_3x_4+x_5x_6=0. \end{sistema}\] \[\begin{array}{|cc|cc|} \toprule \text{Rational point} & \text{Disc.} & \text{Rational point} & \text{Disc.} \\ \midrule (0:0:0:0:0:1) & \text{cusp} & (0:1:0:0:0:0) &-19 \\ (1:0:1:1:0:1) &-43 & (1:0:1:0:0:1) &-12 \\ (0:1:1:1:-1:0) &-27 & (3:-2:-5:4:-4:-3) &-4 \\ \bottomrule \end{array}\] - Curve \(X_0^+(281)\). Genus \(g=7\). Equations of the canonical model in \(\PP^{6}\) \[\begin{sistema}-x_1x_2+x_1x_3+x_2x_3+x_2x_4-x_3x_4+x_3x_6=0 \\-x_1x_6-x_2x_3-x_2x_4-x_2x_5+x_3^2+x_3x_5+x_3x_7=0 \\ x_1x_2+x_1x_7+2x_3x_4-x_3x_5+x_4^2=0 \\-x_1x_2-2x_1x_3-x_1x_4-x_1x_5-x_1x_7-x_2x_3+2x_3x_5+x_4x_5=0 \\ x_1x_2-x_1x_3+x_2x_5+2x_3x_4+x_4x_6=0 \\ 2x_1x_3+2x_1x_4+x_1x_5+x_1x_6+x_1x_7-x_2^2+x_2x_3+x_2x_4+x_2x_5+x_3x_4-2x_3x_5+ x_4x_7=0 \\ x_1x_2+2x_1x_3+x_1x_4+x_1x_5+x_1x_6+x_1x_7+x_2x_3-x_2x_4-x_3x_5+x_5^2=0 \\ x_1x_2-x_1x_3+x_2^2-x_2x_3-x_2x_4-x_2x_5+x_3x_5+x_5x_6=0 \\ x_1x_2-x_1x_3-x_1x_4-x_1x_7+x_2^2+x_2x_6+x_3x_5+x_5x_7=0 \\ x_1x_2-x_1x_3+x_1x_6-x_2x_3-x_2x_4-x_2x_7+2x_3x_4+x_6^2=0. \end{sistema}\] \[\begin{array}{|cc|cc|} \toprule \text{Rational point} & \text{Disc.} & \text{Rational point} & \text{Disc.} \\ \midrule (0:0:0:0:0:0:1) & \text{cusp} & (1:0:0:0:0:0:0) &-16 \\ (2:-5:-5:-1:1:5:-3) &-163 & (1:0:0:1:0:0:-1) &-8 \\ (0:1:-1:1:1:1:-1) &-43 & (0:0:1:0:0:0:-1) &-7 \\ (0:0:1:-2:0:-2:-1) &-28 & (1:2:-2:2:2:0:-2) &-4 \\ \bottomrule \end{array}\] - Curve \(X_0^+(359)\). Genus \(g=6\). Equations of the canonical model in \(\PP^{5}\) \[\begin{sistema} x_1x_2-x_1x_5 + x_1x_6 + x_3x_4=0 \\ x_1^2 + x_1x_2 + x_1x_3-2x_1x_4-x_1x_5 + x_2x_4 + x_3x_5=0 \\-3x_1x_2-x_1x_3 + 2x_1x_4 + 2x_1x_5-x_1x_6 + x_2^2-x_2x_3-x_2x_4-x_3^2 + x_3x_6=0 \\ x_1^2 + 3x_1x_2 + x_1x_3-2x_1x_4-2x_1x_5 + x_1x_6-x_2^2 +x_2x_4 + x_3^2 + x_4x_5=0 \\-x_1^2-2x_1x_2-2x_1x_3 + 3x_1x_4 + 3x_1x_5-x_6 + x_2^2-2x_2x_4-x_2x_5-x_3^2 + x_4x_6=0 \\ x_1^2 +2x_1x_3-2x_1x_4-2x_1x_5 + x_2^2 +x_2x_4-x_2x_6 + x_5^2=0. \end{sistema}\] \[\begin{array}{|cc|cc|} \toprule \text{Rational point} & \text{Disc.} & \text{Rational point} & \text{Disc.} \\ \midrule (0:0:0:0:0:0:1) & \text{cusp} & (1:0:0:0:0:0:0) &-28 \\ (2:-5:-5:-1:1:5:-3) &-163 & (1:0:0:1:0:0:-1) &-19 \\ (0:1:-1:1:1:1:-1) &-67 & (0:0:1:0:0:0:-1) &-7 \\ (0:0:1:-2:0:-2:-1) &-43 & & \\ \bottomrule \end{array}\] ## Modular parametrization {#subsec:modpar} Here we list the isogeny factor \(E\) of \(J_0^+(p)\), a generator \(P_0\) of the Mordell Weil group, an explicit formula for the modular parametrization \(\phi_+\colon X_0^+(p)\to E\) and the modular degree \(\deg(\phi_+)\) of the morphism \(\phi_+\). - Curve \(X_0^+(163)\). Elliptic curve \(E: y^2+y=x^3-2x+1\).\ Mordell-Weil generator \(P_0=(1,0)\). Modular degree \(\deg(\phi_+)=3\).\ Modular parametrization: \(\phi_+(x_1:x_2:x_3:x_4:x_5:x_6)=(P_x,P_y)\), \[P_x =\frac{-2x_3-x_4+x_5-x_6}{-x_3+x_5}; \, P_y =\frac{-x_1^2+x_1x_2+2x_1x_3+ 2x_1x_4+2x_1x_6-x_3^2+x_4^2}{x_1x_3-x_1x_5}.\] - Curve \(X_0^+(197)\). Elliptic curve \(E: y^2+y=x^3-5x+4\).\ Mordell-Weil generator \(P_0=(1,0)\). Modular degree \(\deg(\phi_+)=5\).\ Modular parametrization: \(\phi_+(x_1:x_2:x_3:x_4:x_5:x_6)=(P_x,P_y)\), \[\begin{aligned} \num(P_x)&=-38x_1^2-25x_1x_2-6x_1x_3+13x_1x_4-61x_1x_5+11x_1x_6+22x_2^2+ \\ &+10x_2x_5+20x_2x_6+17x_3x_5-11x_3x_6, \\ \den(P_x)&=-36x_1^2-34x_1x_2+11x_1x_3+30x_1x_4-27x_1x_5-8x_1x_6+12x_2^2+ \\ &+5x_2x_5+11x_2x_6, \\ \num(P_y) &=1715x_1^3-899x_1^2x_2-2691x_1^2x_3-1013x_1^2x_4-2889x_1^2x_5-81x_1^2x_6+ \\ &+1875x_1x_2^2+2476x_1x_2x_5+1176x_1x_2x_6-669x_1x_3x_5+1224x_1x_3x_6+ \\ &+2205x_1x_5x_6-1306x_1x_6^2-384x_2^2x_5-384x_2x_5x_6, \\ \den(P_y)&= 961x_1^3-49x_1^2x_2+3875x_1^2x3-632x_1^2x_4+1195x_1^2x_5+356x_1^2x_6+ \\ &+593x_1x_2^2+655x_1x_2x_5+432x_1x_2x_6-839x_1x_3^2-1680x_1x_3x_4+ \\ &+154x_1x_3x_5+384x_1x_5^2. \end{aligned}\] - Curve \(X_0^+(229)\). Elliptic curve \(E: y^2+xy=x^3-2x-1\).\ Mordell-Weil generator \(P_0=(-1,1)\). Modular degree \(\deg(\phi_+)=4\).\ Modular parametrization: \(\phi_+(x_1:x_2:x_3:x_4:x_5:x_6:x_7)=(P_x,P_y)\), \[\begin{aligned} P_x &=\frac{-x_2-x_5+x_6}{x_2+x_5}, \\ \num(P_y)&=x_1^2+3x_1x_2+7x_1x_3+6x_1x_4+x_1x_5-x_1x_6-4x_1x_7+ \\ &-2x_2x_3-2x_2x_4+3x_2x_6-2x_3^2+x_4x_6-2x_4x_7-x_6x_7, \\ \den(P_y)&=x_1^2+5x_1x_2+x_1x_4+2x_1x_5-x_1x_6-x_1x_7-x_3^2-x_4x_7+x_6^2. \end{aligned}\] - Curve \(X_0^+(269)\). Elliptic curve \(E: y^2+y=x^3-2x-1\).\ Mordell-Weil generator \(P_0=(-1,0)\). Modular degree \(\deg(\phi_+)=3\).\ Modular parametrization: \(\phi_+(x_1:x_2:x_3:x_4:x_5:x_6)=(P_x,P_y)\), \[P_x =\frac{x_4+x_6}{x_2}, \qquad P_y =\frac{x_1x_5-x_2x_3+x_3x_4}{x_1x_2}.\] - Curve \(X_0^+(359)\). Elliptic curve \(E: y^2+xy+y=x^3-x^2-7x+8\).\ Mordell-Weil generator \(P_0=(2,-1)\). Modular degree \(\deg(\phi_+)=4\).\ Modular parametrization: \(\phi_+(x_1:x_2:x_3:x_4:x_5:x_6)=(P_x,P_y)\), \[\begin{aligned} P_x &=\frac{-4x_1^2 + 3x_1x_2-4x_1x_3-x_1x_5 + 3x_1x_6 + x_2x_4 + x_4^2}{-2x_1^2+ x_1x_2-2x_1x_3 + x_1x_6}, \\ P_y &=\frac{-5x_1^2 + 4x_1x_2-6x_1x_3 + 5x_1x_4 + 4x_1x_6-x_2x_4-x_2x_5-x_2x_6 + x_3^2 + 2x_4^2}{3x_1^2 + 3x_1x_3-2x_1x_4-2x_1x_5 + x_2x_4}. \end{aligned}\]
{'timestamp': '2016-07-18T02:07:59', 'yymm': '1607', 'arxiv_id': '1607.04558', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04558'}
null
null
# Introduction {#sec:intro} Hyperspectral imaging systems have become increasingly popular for a variety of applications, including remote sensing and biomedical analytics. With their dense, contiguous and narrow-band spectral sampling in visible through short-wave infrared regions of the electromagnetic spectrum, they provide rich spectral characterization of the objects that are dominant in the pixels of the hyperspectral image. The capability of hyperspectral data to accurately capture material-specific properties makes them an attractive choice for characterizing vegetation mapping invasive and endangered vegetation, (e.g., for ecological and precision agriculture applications), detection and characterization of physiological conditions and other related biomedical applications. In addition to such applications where hyperspectral data has seen a widespread popularity and acceptance, we suggest that hyperspectral data is also particularly suited for enhanced computer vision applications as they relate to scene understanding, biometrics, and person re-dentification, by virtue of their robust characterization of material properties. Person re-identification--the task of recognizing a person separated in location and time, has emerged as an important application of multi-camera surveillance systems. Despite algorithmic advances built upon traditional computer vision imaging systems, person reidentification remains a difficult problem due to various challenges, including variations in illumination conditions, pose and viewpoints. Advances in this area include pose, viewpoint and illumination invariant feature extraction, as well as the design of appropriate similarity metrics. A majority of these methods take into account global appearance of a person, such as a weighted color histogram. Local spatial information is often extracted by analyzing cropped regions around the face. In this paper, we assert that hyperspectral imagery is potentially very beneficial for the task of person re-identification in a multi-camera surveillance scenario---specifically, spectral content can serve as a powerful descriptor that can be used by itself or in conjunction with classical spatial and statistical features derived for re-identification. The spectral reflectance *signature* will demonstrate variability across materials (e.g. clothing) and across skin between different people. While it is expected that traditional computer vision imaging systems can characterize differences in clothing and other traditional descriptors, which may be highly separable between people in the visible wavelength regime, subtle differences due to individual skin physiology can be better characterized by hyperspectral imagery. With this in mind, we developed a pilot study and acquired hyperspectral imagery from fifteen subjects in the visible and very near infrared region of the electromagnetic spectrum at two different locations and at different times of the day (morning versus afternoon). We utilized a simple distance metric appropriate for hyperspectral data and reported re-identification performance using spectral features derived from superpixel patches over the skin, to demonstrate the efficacy of such data for person reidentification. This was compared with reidentification undertaken with color imagery. The outline of this paper is as follows. In section 2, we provide details of the data acquisition, and describe the data that was acquired. In section 3, we describe the approach utilized to setup and quantify re-identification performance with this data. In section 4, we provide experimental results for the re-identification task, and in section 5, we provide concluding remarks. # Hyperspectral Data Acquisition for Person Reidenitifcation {#sec:data} With the goal of demonstrating the efficacy of hyperspectral data for person re-identification, we acquired hyperspectral data from \(15\) people under a traditional multi-camera surveillance scenario at two different locations and two different times of the day (morning and afternoon). The hyperspectral images were acquired using a Headwall Photonics\(^\text{TM}\) hyperspectral imager---the spatial size of each image was \(1004\times400\) and each pixel had \(325\) contiguous spectral bands spanning the visible and near-infrared spectrum from \(400nm-1000nm\) uniformly, with a full-width half-maximum (FWHM) bandwidth of \(1.8nm\). In order to evaluate the performance of a re-identification model, the dataset must represent commonly encountered confounding factors such as viewpoint and illumination variation. Hence, for each person, the images were acquired at two different times of the day (morning and afternoon), and for each of these two acquisitions, the viewpoints, and location of the camera were different. This data collection was carried out in two batches---in the series of 15 hyperspectral images acquired in the morning, the camera orientation was fixed relative to the background in the scene, and each person was asked to arbitrarily stand somewhere in the scene and assume a natural (and arbitrary pose). Again, in the afternoon, another series of 15 images were acquired with the sample 15 people, but at a different location, and with a different orientation of the camera. For the purpose of quantifying the potential of hyperspectral images for person re-identification, we treat the morning images from the 15 people as the gallery set, and the afternoon images from those people as the probe set. Also, in addition to the camera having two different orientations and locations in the morning and afternoon, each person was asked to stand arbitrarily in the scene in order to get arbitrary viewpoints. Due to the temporal difference of the captured data, variations in physiological conditions are also represented in this dataset. Fig. [\[fig:spectra\]](#fig:spectra){reference-type="ref" reference="fig:spectra"} depicts spectral signatures from the background objects and various parts of the head from one of the \(15\) people. It also depicts spectral signatures corresponding to skin pixels (from the face and hands) for two people (Person 12 and Person 13) in the gallery (AM) and probe (PM) sets. This suggests that hyperspectral data can effectively characterize variability between people when using pixels over skin objects. # Spectral Angle based Person Re-identification {#sec:method} In this paper, we focus on the utility of spectral information from skin pixels for person re-identification. Our approach consists of the following steps---we manually annotated small patches of "skin" pixels (on the face and hands) for each person in the gallery and probe sets. We carried out superpixel segmentation to match mean spectra from skin superpixels to avoid pixel mixing that would result from rectangular windows that are traditionally employed. By averaging spectral signatures over superpixels, we stabilize spectral response from local variability and noise, while at the same time avoiding inadvertent mixing that would have resulted had we chosen rectangular windows. Specifically, we utilized the entropy rate superpixel algorithm which ensures compact and homogenous clusters of similar sizes. We took superpixels that intersected with our "ground-truth" skin patches in each image and used those to match each probe image with the gallery. A spectral angle distance was used as the similarity metric for the re-identification task. If \(S_i\) and \(S_j\) represent superpixels in the gallery and probe set respectively, then a spectral angle distance \(d_{sa} (S_i,S_j)\) between the two superpixels can be defined as: \[\label{eq:SAD} {d_{sa} (S_i,S_j)} = {\frac{{{\hat{{\mathbf{x}}}}_{S_i}^T}{{\hat{{\mathbf{x}}}}_{S_j}}}{{\parallel {{\hat{{\mathbf{x}}}}_{S_i}} \parallel} {\parallel {{\hat{{\mathbf{x}}}}_{S_j}} \parallel}}}\] where, \({\hat{{\mathbf{x}}}}_{S_i}\) is the mean spectral signature of region \({S_i}\). In addition to facilitating a computationally simple comparison, using spectral angle as a similarity metric with spectral signature features provides us with an added advantage in that the metric exhibits illumination invariance. Average spectral angle distances between skin superpixels from the probe image and every image in the gallery set are used to re-identify the probe set. Note that the superpixels that result from the entropy rate superpixel algorithm modified to use spectral angle based similarity are highly effective at oversegmenting the image into uniform sized clusters while preserving borders, and are hence appropriate for spectral matching via the spectral angle distance. In a practical implementation, matching skin superpixels between probe and gallery sets can be easily accomplished by first utilizing a face detection algorithm to identify the skin superpixels in all the images. For this closed set re-identification problem, we report performance via the cumulative matching characteristic (CMC) curve. The CMC curve represents the expectation of finding the correct match in the top *r* matches. In other words, a rank-*r* recognition rate shows the percentage of the test images that are correctly recognized from the top *r* matches in the gallery set. The rank-1 value on this curve indicates the true identification performance, while the rank-\(N\) score (\(N\) being the number of images in the gallery) will be \(100\%\) for closed-set re-identification, with the curve monotonically increasing from 1 through \(N\). Comparison between hyperspectral and RGB (color) images based on this approach is reported in the CMC curve in Fig. [\[fig:CMC\]](#fig:CMC){reference-type="ref" reference="fig:CMC"}. We observe that hyperspectral images not only provides a superior rank-1 performance compared to when using (RGB) color images, the CMC curve for hyperspectral images increases much faster as a function of the rank, with hyperspectral data providing as much as \(20\%\) better identification performance. # Conclusions, Caveats and Future Work We conclude from the data and the results that hyperspectral images have potential to enhance re-identification performance in multi-camera surveillance systems---the preliminary dataset and resulting CMC curves suggest the "value" added by considering the narrow-band spectral information for re-identification as opposed to 3-channel color images. By being able to distinguish subtle spectral variations between people (e.g. via the spectral signatures of their skin), it may also enable long-duration re-identification wherein significant time may have elapsed between the person reappearing in the collective field of view of the system. We acknowledge an important caveat with results presented here---the sample size (30 images total in gallery and probe sets) is small---overall performance will drop when the number of people are added to the re-identification problem. However, we expect that spectral signature as a descriptor to characterize skin pixels effectively is likely to result in superior performance compared to traditional (RGB) color systems. Finally, we note that the approach to re-identification can be enhanced by complementing spectral signatures with other descriptors (e.g. spatial information). In ongoing work, we are expanding the gallery and probe sets, and are exploring strategies to fuse information provided from spectral signatures with currently established color image based feature descriptors for effective re-identification.
{'timestamp': '2016-07-18T02:09:19', 'yymm': '1607', 'arxiv_id': '1607.04609', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04609'}
null
null
# Introduction Measurements on separated subsystems in a joint entangled state may display correlations that cannot be mimicked by local hidden variable (LHV) models. These correlations are termed nonlocal and are detected by violating Bell inequalities. In recent years it has become clear that non-locality is interesting not only for fundamental reasons, but also as a resource for device-independent (DI) quantum information tasks such as quantum key distribution or random number generation [@colbeck; @pironionature]. Thus, violations of Bell inequalities are not only indicators of non-locality, but can also be used to make qualitative and quantitative statements about operationally relevant quantum properties. Traditionally, the problem of constructing Bell inequalities has been addressed from the point of view of deriving constraints satisfied by LHV models. Following this standard approach, the inequalities are derived using well-known techniques in convex geometry. Indeed, the set of correlations admitting LHV models defines a polytope, i.e., a bounded convex set with a finite number of vertices. These vertices correspond to local deterministic assignments, while the facets are the desired Bell inequalities. Facet (or tight) Bell inequalities provide necessary and sufficient criteria to detect the non-locality of given correlations. Clauser-Horne-Shimony-Holt (CHSH)  and Collins-Gisin-Linden-Massar-Popescu (CGLMP)  Bell inequalities are examples thereof. Although such facet Bell inequalities are optimal detectors of non-locality, they are not necessarily optimal for inferring specific quantum properties in the DI setting. For instance, in a scenario where two binary measurements are performed on two entangled subsystems, it is well known that the violation of the CHSH inequality is a necessary and sufficient condition for non-locality. But certain "non-facet\" Bell inequalities are better certificates of randomness than the CHSH one when the two quantum systems are partially entangled. The main aim of this work is to introduce Bell inequalities valid for an arbitrary number of measurements and outcomes whose maximal quantum violation, usually referred to as the *Tsirelson bound*, is attained by maximally entangled states \[|\psi^+_d \rangle=(1/\sqrt{d})\sum_{i=0}^{d-1}|ii\rangle.\] This is a desirable property since these states have particular features such as perfect correlations between outcomes of local measurements in the same bases, and therefore many quantum information protocols rely on them. In the particular case of two measurements CHSH is the simplest example of a Bell inequality with the above property, but others are known (see also results for many settings ). Our construction works, however, for arbitrary numbers of measurements and outcomes, and, crucially, the Tsirelson bound of the resulting Bell inequalities can be computed *analytically*. In the case where only two measurements are made on each subsystem, all facet Bell inequalities are known for a small number of outputs and they are of the CGLMP form. However, they are not maximally violated by the maximally entangled states of two qudits (except in the case \(d=2\) corresponding to the CHSH inequality). Thus, we should not expect our Bell inequalities to be tight, and indeed they are not. This implies that we cannot use standard tools from convex geometry to construct them. In fact, no quantum property is used for the construction of tight Bell inequalities like the CGLMP one and, in this sense, it is not surprising that their maximal violation does not require maximal entanglement. Our approach is completely different: it starts from quantum theory and exploits the symmetries and perfect correlations of maximally entangled states to derive a Bell inequality (cf. Ref. for a similar method). It exploits sum of squares decompositions of Bell operators, which is used to determine their Tsirelson bound. Thus, contrary to any previous derivation of Bell inequalities, quantum theory becomes a key ingredient of our method. Our results provide new insight into the structure of the boundary of the set of quantum correlations (see discussion in Appendix [\[appendixstructure\]](#appendixstructure){reference-type="ref" reference="appendixstructure"}). In addition, our Bell inequalities have the potential to be used in DI quantum information protocols such as random number generation, quantum key distribution or to self-testing (see more detailed discussion in Section [5](#secappli){reference-type="ref" reference="secappli"}). # Preliminaries We consider a Bell scenario with two distant parties \(A\) and \(B\) performing one of \(m\) measurements \(A_x\) and \(B_y\) with \(d\) outcomes on their share of some physical system. We label the measurements and outcomes as \(x,y \in \{1,\ldots, m\}\) and \(a,b \in \{0,\ldots, d-1\}\). The correlations obtained in this experiment are described by \((md)^2\) joint probabilities \(P(A_x=a,B_y=b)\) that \(A\) and \(B\) obtain \(a\) and \(b\) upon performing the \(x\)th and \(y\)th measurement, respectively. These probabilities are ordered into a vector \(\vec{p}:=\{P(A_x=a,B_y=b)\}_{a,b,x,y} \in \mathbbm{R}^{(md)^2}.\) Importantly, the set of allowed vectors \(\vec{p}\) varies depending on the physical principle they obey. If the measurements define spacelike separated events, the observed correlations should obey the *no-signalling principle*, which prevents any faster-than-light communication among the parties. These correlations form a convex polytope denoted \(\mathcal{N}\). Contained in this set is the set of quantum correlations \(\mathcal{Q}\) which is formed by those \(\vec{p}\) whose components can be written as \(P(A_x=a,B_y=b)=\langle \psi|P_a^{(x)}\otimes P_b^{(y)}|\psi\rangle\), where \(|\psi\rangle\) is some state in a product Hilbert space \(H_A\otimes H_B\) of unconstrained dimension, and \(\{P_a^{(x)}\}\) and \(\{P_b^{(y)}\}\) are projection operators defining, respectively, Alice's and Bob's measurements. Finally, the set of correlations admitting LHV models, denoted \(\mathcal{L}\), contains those \(\vec{p}\) that can be written as a convex sum of product deterministic correlations \(P(A_x=a,B_y=b)=P(A_x=a)P(B_y=b)\) with \(P(A_x=a),P(B_y=b)=0,1\) for all \(x,y\). Bell was the first to prove that not all quantum correlations admit an LHV model. To this end, he used the concept of a Bell inequality \(I\leq C_b\), where \(I\) is the so-called Bell expression that is a linear combination of the \((md)^2\) joint probabilities of the form \[\label{BE} I:=\sum_{abxy} I_{abxy} P(A_x=a,B_y=b),\] and \(C_b=\max_{\vec{p} \in \mathcal{L}}I\) is its classical bound. The quantum or Tsirelson bound of \(I\) is the maximum value \(Q_b=\max_{\vec{p} \in \mathcal{Q}}I\) that it can achieve for quantum correlations. A Bell expression \(I\) gives rise to a proper Bell inequality---one that is violated by quantum theory---if \(C_b < Q_b\). If \(\vec{p}\) violates a Bell inequality, the correlations described by \(\vec{p}\) are termed nonlocal. Finally, one defines \(NS_b=\max_{\vec{p} \in \mathcal{N}}I\) as the maximum value of \(I\) over no-signalling correlations. For most of the known Bell inequalities \(NS_b>Q_b>C_b\). Let us stress that although \(\mathcal{Q}\) is convex, it is not a polytope. More importantly, the boundary of \(\mathcal{Q}\) remains unknown despite several attempts to characterize it analytically (see, nevertheless, ). This clearly makes the derivation of Tsirelson bounds a hard task. Given a Bell inequality, there is no procedure that guarantees finding its quantum bound, and it was achieved analytically only in a handful of cases. There is, however, a practical approximation scheme based on semidefinite programming, which consists in a hierarchy of sets \(\mathcal{Q}_{1} \supseteq \mathcal{Q}_{2} \supseteq \dots \supseteq \mathcal{Q}_{k} \supseteq \dots\) converging to \(\mathcal{Q}\) as \(k \to \infty\), and allows one to bound \(Q_b\) from above (see also ). Although for small Bell scenarios this method yields good numerical bounds (often tight), it becomes computationally expensive for scenarios involving a large number of measurements or outcomes. # Class of Bell expressions {#secclass} Our aim now is to introduce a family of Bell expressions, whose maximal quantum value is attained by the *two-qudit* maximally entangled state \(|\psi^+_d \rangle\). To derive them, we start from the premise that their maximal quantum values are obtained when Alice and Bob perform the optimal CGLMP measurements introduced in (cf. Appendix [\[appendixmeasurements\]](#appendixmeasurements){reference-type="ref" reference="appendixmeasurements"}). This choice stems from the fact that these measurements generalize the CHSH measurements (\(d=2\)) to arbitrary dimensions and they lead to non-local correlations that are most robust to noise  or for \(m=2\) give a stronger statistical test . The probabilities \(P(A_x=a,B_y=b)\) obtained when using the optimal CGLMP measurements on \(|\psi^+_d \rangle\) have several symmetries. For instance, they only depend on the difference \(a-b=k\mod d\). If we impose that our Bell expressions respect this symmetry, the probabilities \(P(A_{x} = j + k \text{ mod }d,\text{ } B_{y} = j)\) should be treated equally for all \(j\), i.e., the Bell expressions should be linear combinations of \(P(A_{x} = B_{y} + k) := \sum_{j = 0}^{d-1} P(A_{x} = j + k \text{ mod }d,\text{ } B_{y} = j)\). Taking into account all symmetries, a generic form for our Bell expressions is \[\label{Bellproba} I_{d,m}:=\sum_{k=0}^{\left \lfloor d/2\right \rfloor-1}\left( \alpha_k \mathbbm{P}_k-\beta_k \mathbbm{Q}_k \right ),\] where \(\mathbbm{P}_k := \sum_{i=1}^{m}[P(A_i=B_i+k)+P(B_i=A_{i+1}+k)]\), \(\mathbbm{Q}_k := \sum_{i=1}^{m}[P(A_i=B_i-k-1)+P(B_i=A_{i+1}-k-1)]\) with \(A_{m+1}:=A_1+1\). The parameters \(\alpha_k\) and \(\beta_k\) are our degrees of freedom. Taking, e.g., \(\alpha_k = \beta_k = 1-2k/(d-1)\) for \(m=2\), one recovers the CGLMP Bell inequalities. To exploit the symmetries inherent in Bell inequalities, we often write them in terms of correlators instead of probabilities. As we consider an arbitrary number of outcomes, we appeal to the notion of generalized correlators (see, e.g., Ref. and Ref. for other options). These are complex numbers that are defined through the two-dimensional Fourier transform of the probabilities \(P(A_x=a,B_y=b)\): \[\langle A_x^kB_y^l\rangle=\sum_{a,b=0}^{d-1}\omega^{ak+bl}P(A_x=a,B_y=b), \label{correlators}\] where \(\omega = \text{exp}(2\pi i /d)\), \(k,l\in \{0,\ldots,d-1\}\), and \(\{A_x^k\}_k\) and \(\{B_y^l\}_l\) can be thought of as measurements with outcomes labelled by roots of unity \(\omega^i\) \((i=0,\ldots,d-1)\). For quantum correlations \(\vec{p}\), the correlators \(\langle A_x^kB_y^l\rangle\) are average values of the tensor product of the operators \(A_x^{k}=\sum_{a=0}^{d-1}\omega^{ak}P_a^{(x)}\quad \mathrm{and} \quad B_y^{l}=\sum_{b=0}^{d-1}\omega^{bl}P_b^{(y)}\) in the state \(\ket{\psi}\). Note that they are unitary, their eigenvalues are the roots of unity, and they satisfy \((A_x^{k})^{\dagger}=A_x^{d-k}\) and \((B_y^{l})^{\dagger} = B_{y}^{d-l}\) for any \(k,l\). Now, exploiting ([\[correlators\]](#correlators){reference-type="ref" reference="correlators"}), expression ([\[Bellproba\]](#Bellproba){reference-type="ref" reference="Bellproba"}) can be rewritten as \[\label{Bellcorr} \widetilde{I}_{d,m}= \sum_{i=1}^m\sum_{l=1}^{d-1}\langle A_i^l\bar{B}_i^l \rangle,\] where, for clarity, the change of variables \(\bar{B}_i^l = a_l B_i^{d-l}+a_l^{*}B_{i-1}^{d-l}\) with \(a_l=\sum_{k=0}^{\lfloor d/2\rfloor-1}(\alpha_k\omega^{-kl}-\beta_k\omega^{(k+1)l})\) was introduced on Bob's side. Due to the convention \(A_{m+1}=A_1+1\), the term \(\bar{B}_1^l\) is defined as \(\bar{B}_1^l=a_lB_{1}^{d-l}+a_l^* \omega^lB_{m}^{d-l}\). For simplicity, in ([\[Bellcorr\]](#Bellcorr){reference-type="ref" reference="Bellcorr"}) we ignored the irrelevant scalar term corresponding to \(l=0\) and rescaled the expression. Below we denote the classical, quantum and no-signaling bound of \(\widetilde{I}_{d,m}\) by \(\widetilde{C}_b\), \(\widetilde{Q}_b\) and \(\widetilde{NS}_b\), respectively. Our aim now is to fix the free parameters \(\alpha_k\) and \(\beta_k\) according to the quantum property we need: maximal violation by the maximally entangled state \(| \psi^+_d \rangle\). At this point, it is instructive to look at the specific example of the CHSH Bell expression (\(m=2\), \(d=2\)). In the notation ([\[Bellcorr\]](#Bellcorr){reference-type="ref" reference="Bellcorr"}) the CHSH Bell expression \(\langle A_1B_1\rangle + \langle A_1 B_2\rangle + \langle A_2 B_1\rangle-\langle A_2 B_2\rangle\) reads \(\widetilde{I}_{2,2}=\langle A_1 \bar{B}_1\rangle +\langle A_2 \bar{B}_2\rangle,\) where \(\bar{B}_1=(B_1 + B_2)/\sqrt{2}\), \(\bar{B}_2=(B_1-B_2)/\sqrt{2}\). Then, for the optimal measurements leading to the Tsirelson bound of \(\widetilde{I}_{2,2}\), we have \(\bar{B}_1=A_1^*\) and \(\bar{B}_2=A_2^*\). This reflects the property that for the maximally entangled state \[\label{Raimat} M\otimes N\ket{\psi_d^+}=\mathbbm{1}\otimes NM^{T}\ket{\psi_d^+},\quad \forall M,N.\] This condition implies that a measurement by Alice is perfectly correlated with its complex conjugate by Bob. Our intuition to derive Bell inequalities detecting maximal entanglement is to impose this property for any \(m\) and \(d\): we choose the parameters \(\alpha_k\) and \(\beta_k\) such that \[\label{conditions} \bar{B}_{i}^{l} = (A_{i}^{l})^{*}\] hold for \(l = 1,\ldots, d-1\) and \(i = 1, \ldots, m\) with the initial operators \(\{P_a^{(x)}\}\) and \(\{P_b^{(y)}\}\) being the optimal CGLMP operators. Conditions ([\[conditions\]](#conditions){reference-type="ref" reference="conditions"}) give rise to a set of linear equations for \(\alpha_k\) and \(\beta_k\) which yields (see Appendix [\[appendixcoefficients\]](#appendixcoefficients){reference-type="ref" reference="appendixcoefficients"} for details) \[\label{alpha} \alpha_k= \frac{1}{2d}\tan\left(\frac{\pi}{2m}\right) \left[g(k)-g\left(\left\lfloor \frac{d}{2} \right\rfloor\right)\right],\] \[\label{beta} \beta_k= \frac{1}{2d}\tan\left(\frac{\pi}{2m}\right) \left[g\left(k + 1-\frac{1}{m}\right) + g\left(\left\lfloor \frac{d}{2} \right\rfloor\right)\right]\] with \(g(x):=\cot(\pi(x+ 1/2m)/d)\). To sum up, our class of Bell expressions is given by \(I_{d,m}\) ([\[Bellproba\]](#Bellproba){reference-type="ref" reference="Bellproba"}) or equivalently by \(\widetilde{I}_{d,m}\) ([\[Bellcorr\]](#Bellcorr){reference-type="ref" reference="Bellcorr"}), with coefficients ([\[alpha\]](#alpha){reference-type="ref" reference="alpha"}) and ([\[beta\]](#beta){reference-type="ref" reference="beta"}). We arrived at it by writing the most general Bell expression satisfying the symmetry of CGLMP correlations, re-writing these Bell expressions in the simple form ([\[Bellcorr\]](#Bellcorr){reference-type="ref" reference="Bellcorr"}) through a change of variable on Bob's side, and then imposing the conditions ([\[conditions\]](#conditions){reference-type="ref" reference="conditions"}) that take into account the symmetries of the maximally entangled state, as CHSH does for two binary measurements. # Properties of the novel Bell expressions {#secproperties} We now analyze the main properties of our Bell expressions: we compute all the relevant bounds \(\widetilde{C}_b\), \(\widetilde{Q}_b\), \(\widetilde{NS}_b\), and show that \(\widetilde{C}_b<\widetilde{Q}_b<\widetilde{NS}_b\) for any \(d\) and \(m\). For clarity we only include sketches of proofs (see Appendices [\[appendixclassical\]](#appendixclassical){reference-type="ref" reference="appendixclassical"}, [\[appendixtsirelson\]](#appendixtsirelson){reference-type="ref" reference="appendixtsirelson"} and [\[appendixns\]](#appendixns){reference-type="ref" reference="appendixns"} for details). Let us begin with the classical bound. Importantly, the resulting Bell inequality \(\widetilde{I}_{d,m}\leq \widetilde{C}_b\) is violated by quantum theory; one can reach the value \(\widetilde{I}_{d,m}=m(d-1)\) by applying the CGLMP measurements on \(\ket{\psi^+_d}\). This is seen by using Eq. ([\[conditions\]](#conditions){reference-type="ref" reference="conditions"}), the unitarity of \(A_{i}^{k}\), and the symmetries of the maximally entangled states ([\[Raimat\]](#Raimat){reference-type="ref" reference="Raimat"}). Then, all the correlators in ([\[Bellcorr\]](#Bellcorr){reference-type="ref" reference="Bellcorr"}) equal one, yielding the quantum violation of \(m(d-1)\). This violation is optimal and defines the tight Tsirelson bound of \(\widetilde{I}_{d,m}\). A few remarks are in order. First, it is not difficult to see that \(\widetilde{Q}_b>\widetilde{C}_b\) for any \(m,d\geq 2\), meaning that all our Bell inequalities are nontrivial (cf. Appendix [\[appendixscalings\]](#appendixscalings){reference-type="ref" reference="appendixscalings"}). Second, let us elaborate on how the SOS works in the case of two measurements, \(m=2\), which justifies the choice of conditions ([\[conditions\]](#conditions){reference-type="ref" reference="conditions"}). For \(m=2\), the second part of the SOS decomposition [\[sos\]](#sos){reference-type="eqref" reference="sos"} vanishes. For the optimal CGLMP measurements both sides of ([\[sos\]](#sos){reference-type="ref" reference="sos"}) must yield zero when applied to \(\ket{\psi^+_d}\), which stems from conditions ([\[Raimat\]](#Raimat){reference-type="ref" reference="Raimat"}) and ([\[conditions\]](#conditions){reference-type="ref" reference="conditions"}). This allows one to grasp the intuition behind conditions ([\[conditions\]](#conditions){reference-type="ref" reference="conditions"}), i.e., they allow one to construct in a quite direct way an SOS decomposition ([\[sos\]](#sos){reference-type="ref" reference="sos"}), in which all operators \(P_{ik}\) are polynomials of the measurement operators \(A_{i}^{k}\) and \(B_{i}^{k}\) of order one, significantly facilitating the computation of the Tsirelson bound. For the CHSH Bell inequality, one observes the same effect, as these same properties of the optimal state and measurements allow the Bell operator \(\mathcal{B}_{\text{CHSH}} = A_1 \otimes B_1 + A_1 \otimes B_2 + A_2 \otimes B_1-A_2 \otimes B_2\) to have the decomposition: \(2\sqrt{2}\mathbbm{1}-\mathcal{B}_{\text{CHSH}} = ( P_{1}^{\dagger}P_1 + P_{2}^{\dagger}P_2)/\sqrt{2},\) with \(P_1 = (1/\sqrt{2})\mathbbm{1}\otimes (B_1 +B_2)-A_1 \otimes \mathbbm{1}\), and \(P_2 = (1/\sqrt{2})\mathbbm{1} \otimes(B_1-B_2)-A_2 \otimes \mathbbm{1}\). Thus, our construction generalizes this quantum aspect of the CHSH Bell operator. For larger number of measurements, \(m>2\), the first part of the SOS decomposition is not enough and one has to add "by hand" the extra term in which all \(T_{ik}\)'s are also of order one in \(B_i^k\). Note that for two measurements, our Bell expressions coincide with those introduced in and then rederived in using a different approach. Moreover, the Tsirelson bounds of these Bell inequalities was computed in Refs. exploiting other techniques, and it was proven in that they are not tight. On the other hand, for \(d=2\) and any \(m\), our class recovers the well-known chained Bell inequalities. We finally notice that the alternative generalization of the CHSH Bell inequality to three measurements and outcomes given in was also found to be maximally violated by \(\ket{\psi^+_3}\). Let us eventually compute the no-signalling bound of our Bell expressions. Again, it is not difficult to see that \(\widetilde{NS}_b>\widetilde{Q}_b\) for any \(m,d\geq 2\) (see Appendix [\[appendixscalings\]](#appendixscalings){reference-type="ref" reference="appendixscalings"} for the proof and scalings of \(\widetilde{C}_b\), \(\widetilde{Q}_b\) and \(\widetilde{NS}_b\) with \(m\) and \(d\)). # Applications to device-independent protocols {#secappli} A natural application for our Bell inequalities is self-testing---a DI protocol in which a state and measurements performed on it are certified up to local isometries, based on the nonlocal correlations they produce. To perform self-testing, the correlations \(\vec{p}\) maximally violating the given Bell inequality must be unique, i.e., attained, up to local isometries, by certain state and measurements. This is generally hard to prove. There exists, however, a numerical method for self-testing. We applied it to the simplest case \(m=2\) and \(d = 3\), and the results are plotted in Figure [\[swap3\]](#swap3){reference-type="ref" reference="swap3"}. It shows that one can self-test the maximally entangled state of two qutrits \(\ket{\psi^+_3} =(\ket{00} + \ket{11} + \ket{22})/\sqrt{3}\) with our inequalities. An open question is whether one can generalize this result to any dimension. Our inequalities could then be applied in DI random number generation protocols [@colbeck; @pironionature; @pironiomassar]. Indeed, if \(\vec{p}\) maximally violating \(\widetilde{I}_{d,m}\) is unique, one can apply the method of and use the symmetries of the Bell expressions to guarantee a dit of perfect randomness. This, by increasing the dimension \(d\), would result in unbounded randomness expansion. Our inequalities could also find applications in DI quantum key distribution. An advantage that our inequalities have over CGLMP in that scenario is that, as said before, the maximal violation is obtained for the maximally entangled state. This state can produce perfect correlations between the users, which reduces the error-correcting phase of the protocol and can lead to better key generation rates. We study this question in Appendix [\[diqkd\]](#diqkd){reference-type="ref" reference="diqkd"}. Numerically, we find that for \(m=2\) and \(d=3\), our inequalities lead to higher key rates than CGLMP for levels of white noise up to \(4.2\) percent. While this advantage is not very large, we believe it grows with the dimension of the systems, at least in the noiseless case. Moreover, it is known that maximally entangled states are much simpler to prepare experimentally than fine-tuned partially entangled states. It would be interesting to confirm these conjectures in a future work focused on DIQKD. # Conclusions In this work, we introduced a new technique allowing to construct Bell inequalities with arbitrary numbers of measurements and outcomes that are maximally violated by the maximally entangled states. It exploits the SOS decompositions of Bell operators and, crucially, allows one to compute analytically their Tsirelson bounds. Our results are general as, unlike previous works, we do not consider a particular Bell scenario, but allow for arbitrary number of measurements \(m\) and outcomes \(d\). Our inequalities can be seen as the "quantum" or the DI-oriented generalization of CHSH Bell inequality, in the same spirit as the CGLMP inequality generalizes the CHSH one classically.
{'timestamp': '2017-08-01T02:14:52', 'yymm': '1607', 'arxiv_id': '1607.04578', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04578'}
# Introduction [\[sec1\]]{#sec1 label="sec1"} The on going search for new physics (NP) is mostly inspired by the shortcomings of the SM in addressing some of the fundamental questions in modern particle physics, such as the hierarchy problem, the flavor patterns in the fermionic sector and dark matter. Some of these unresolved issues may be closely related and may have TeV-scale origins, thus inspiring the search for TeV-scale NP, both theoretically and experimentally. Indeed, two seemingly unrelated interesting measurements of both the ATLAS and the CMS collaborations at CERN, have been recently reported: 1. A possible \((2-4) \sigma\) (local) excess in the diphoton invariant mass distribution around 750 GeV, corresponding to a signal cross-section roughly in the range \(\sigma(pp \to \gamma \gamma) \sim 3-13\) fb \((1\sigma)\), see e.g.,. The interpretation of this excess signal has a slight preference to a spin 0 resonance, produced via gluon-fusion and having a total width ranging from sub-GeV to 45 GeV, with a more significant signal obtained in the ATLAS analysis for a scalar with a total width \(\Gamma \sim 45~{\rm GeV}\). 2. A possible \((1-2.5) \sigma\) excess in the measurement of the LFV decay \(h \to \tau \mu\) of the 125 GeV light Higgs. In particular, the CMS collaboration finds \(BR(h \to \tau \mu) = 0.84\%^{+0.39\%}_{-0.37\%}\), while the ATLAS collaboration finds \(BR(h \to \tau \mu) = (0.53 \pm 0.51)\%\). Whether or not these two measurements are confirmed, it emphasizes the importance of the current efforts in the search for NP, since it provides an interesting manifestation/example of the exciting possibility that the building blocks of new TeV-scale physics may have rather non-conventional properties, potentially with important repercussions for both flavor and the hierarchy problems. For example, the new heavy scalar particle, \(S\), responsible for the 750 GeV \(\gamma \gamma\) excess, should have a rather narrow width and suppressed decay rates into "conventional\" channels such as \(S \to WW,~ZZ,~t \bar t\), for which no excess signals has been observed within the currently available sensitivity of the corresponding LHC searches. In addition, such a heavy scalar \(S\) is most likely related to the light 125 GeV scalar state and, therefore, might also be involved in flavor changing (FC) transitions in the fermionic sector. Such properties of the would be new 750 GeV resonating particle are, therefore, very challenging to accommodate in models beyond the SM, in particular, in supersymmetric models or in models that involve extra space-time dimensions, which seem to have a more fundamental origin and, therefore, likely linked to physics at higher energy scales. Nonetheless, we will show in this paper that a certain class of low-energy effective 2HDM frameworks with a 4th generation of heavy chiral fermions may be interesting candidates for such "exotic\" TeV-scale NP. In particular, since no evidence for such fundamentally structured theories has yet been seen, a frequently adopted phenomenological approach in studies of NP, is to construct TeV-scale models which require a UV completion and may, thus, be viewed as low energy effective frameworks for the underlying dynamics. Such models are useful as a guide for the exploration and model building of more fundamental theories and they often include new heavy fermionic and scalar states with sub-TeV masses. One of the simplest variants of an effective low-energy NP candidate, which dates back to the 1980's, is the SM with an additional 4th generation of fermions; the so called SM4 (for useful reviews see e.g., ). Indeed, since three generations of chiral fermions have been observed in nature, it is natural to ask why not four generations of chiral fermions? It is quite interesting that this simple extension of the SM may address some of the theoretical challenges in particle physics, such as: electroweak symmetry breaking (EWSB) and the hierarchy problem, the CP-violation and the strength of the first-order phase transition needed to explain the origin of matter-anti matter asymmetry in the universe, and flavor physics. As is well known, the SM4 (i.e., with four generations of fermions and one Higgs doublet) is now excluded, since it cannot accommodate the measured SM-like properties of the 125 GeV scalar, see e.g.,, primarily due to an \({\cal O}(10)\) enhancement in the gluon-fusion light Higgs production mechanism from diagrams with \(t^\prime\) and \(b^\prime\) in the loops; see, however, W.-S. Hou in. This fact, along with the rather stringent direct limits on the masses of such heavy quarks (to be discussed later), has led to a common belief that generic extensions to the SM with heavy chiral 4th generation fermions \(t^\prime,b^\prime,\nu^\prime,\tau^\prime\) are excluded. However, as was suggested by us a few years ago and will be demonstrated again here, this is not the case when the heavy 4th generation chiral sector is embedded in frameworks with an extended Higgs sector (see also ). Indeed, an extended Higgs sector in the context of 4th generation heavy fermions may come in handy for further addressing flavor problems and the strength of the EW phase transition required for baryogenesis. In particular, we will consider in this paper a version of a 2HDM introduced by us in -the so called 4G2HDM of type I, where a chiral 4th generation doublet of heavy fermions (quark and lepton) is added and is coupled *only* to one of the scalar doublets (the "heavy\" doublet), while the SM 1st-3rd generations fermions are coupled only to the other doublet (the "light\" doublet). We will show in this paper that this 4G2HDM is a well motivated and valid low-energy model, which is compatible with the 125 GeV signals (see also ), with PEWD and with the existing direct bounds on the heavy fermions, and at the same time can also accommodate the recent indications for a new 750 GeV scalar resonance in the \(\gamma \gamma\) channel. As was shown in, the price to pay when adding another heavy SM-like chiral fermion doublet is that such constructions posses a nearby threshold/cutoff at the several TeV scale, which is manifest (as Landau poles) in the evolution of the Yukawa and Higgs potential couplings. Indeed, the large Yukawa couplings of the heavy chiral fermions can be thought of as a reflection of an underlying TeV-scale strong dynamics, so that the 4G2HDM framework should be viewed as a low energy (i.e., sub-TeV) effective model of an underlying strongly interacting sector. In particular, if the new heavy chiral fermions are viewed as the agents of EWSB (and are, therefore, linked to strong dynamics at the nearby TeV-scale, see e.g., ), then more Higgs particles, which may be composites of these 4th generation fermions, are expected at the sub-TeV regime.\(^{[1]}\) In such scenarios the resulting low-energy effective theory may contain more than a single composite Higgs field and may thus resemble a two (or more) Higgs doublet framework (for other related studies of the phenomenology of multi-Higgs 4th generation models see e.g., ). The purpose of this work is to revisit the 4G2HDM of, studying its compatibility with the updated measurements of the 125 GeV light Higgs signals and with PEWD. We will also confront our model with the 750 GeV \(\gamma \gamma\) excess and study its compatibility with a sub-percent branching ratio of the light Higgs in the FC decay channel \(h \to \tau \mu\). Indeed, many interesting and exotic constructions beyond the SM have been suggested as possible explanations of the 750 GeV \(\gamma \gamma\) excess (too many to be cited here); in most cases involving new degrees of freedom beyond just the 750 GeV resonating particle. In particular, the relevance of 2HDM frameworks to the 750 GeV \(\gamma \gamma\) excess has been intensively studied in the past several months, where it was shown that the simplest 2HDM extension to the SM, in which no additional heavy degrees of freedom are added (i.e., beyond the extended scalar sector), cannot accommodate the necessary enhancement in \(\sigma(pp \to H(750) \to \gamma \gamma)\), see e.g.,. Consequently, extended 2HDM models with TEV scale vector-like (VL) fermions have been suggested for addressing the 750 GeV resonance signal. The upshot of these studies is that, the needed enhancement in the 1-loop production and decay channels \(gg \to H(750)\) and \(H(750) \to \gamma \gamma\), requires several copies of VL fermions and/or VL fermions with charges appreciably larger than those of the SM fermions, unless their Yukawa couplings are much larger than one. The 4G2HDM considered in this work is, therefore, conceptually simpler, relying on new heavy fermionic degrees of freedom with properties similar to the SM fermions in a model that already exists in the literature. The paper is organized as follows: in section 2 we describe the type I 4G2HDM and we layout the physical parameters that are used in the numerical analysis. In section 3 we show our results and in section 4 we discuss their phenomenological consequences. In section 5 we discuss our results and summarize. # The 4G2HDM: a 2HDM with 4th generation fermions [\[sec2\]]{#sec2 label="sec2"} Motivated by the idea that TeV-scale scalar degrees of freedom may emerge as composites associated with heavy fermions, we assume that the low-energy (sub-TeV) effective framework is parameterized by a 2HDM with a chiral SM-like 4th generation of heavy fermions. Specifically, the model is constructed following, such that one of the Higgs fields (\(\phi_h\)-the "heavier\" field) couples only to the new heavy 4th generation fermionic fields, while the second Higgs field (\(\phi_\ell\)-the "lighter\" field) is responsible for the mass generation of all other (lighter) fermions (i.e., the 1st-3rd generation SM fermions). In this model, named in the 4G2HDM of type I (here we will refer to it simply as the 4G2HDM), the Yukawa interaction Lagrangian can be realized in terms of a \(Z_2\)-symmetry under which the fields transform as follows: \[\begin{aligned} \Phi_{\ell}\to-\Phi_{\ell},~ \Phi_{h}\to+\Phi_{h},~ F_{L}\to+F_{L}~, f_{R}\to-f_{R}\;(f={\rm SM~fermions}),~f^\prime_{R}\to +f^\prime_{R}\;(f^\prime={\rm 4th~gen.~fermions})~, \label{eq:z2} \end{aligned}\] where \(F_L\) and \(f_{R},f_R^\prime\) are the SU(2) fermion (quark or lepton) doublets and singlets, respectively, and \(\Phi_{\ell,h}\) are the two Higgs doublets \(\Phi_i =\left( \phi^{+}_i,\frac{v_i+\phi^{0}_i}{\sqrt{2}} \right)\), \(i=\ell,h\). The Yukawa potential that respects the above \(Z_2\)-symmetry is: \[\begin{aligned} \mathcal{L}_{Y}=-\bar{F}_{L} \left( \Phi_{\ell} Y_d^f \cdot \left( I-{\cal I} \right) + \Phi_{h}Y_d^f \cdot {\cal I} \right) f_{d,R}-\bar{F}_{L} \left( \tilde\Phi_{\ell} Y_u^f \cdot \left( I-{\cal I} \right) + \Phi_{h} Y_u^f \cdot {\cal I} \right) f_{u,R} + h.c.\mbox{ ,} \label{eq:LY4G} \end{aligned}\] where \(f_{u,R}\) and \(f_{d,R}\) are the up and down-type SU(2) fermion singlets (quark or lepton of all four generations), \(I\) is the identity matrix and \({\cal I}\) is the diagonal \(4\times4\) matrix \({\cal I} \equiv {\rm diag}\left(0,0,0,1\right)\). The scalar sector contains five massive states: a charged scalar \(H^+\), a CP-odd state \(A\) and two CP-even scalars \(h,H\), so that \(h\) is the lighter one, corresponding to the observed 125 GeV Higgs boson. These physical states are related to the components of the two SU(2) scalar doublets via: \[\begin{aligned} H= s_\alpha {\rm Re} \left( \phi_h^0 \right) + c_\alpha {\rm Re}\left( \phi_\ell^0 \right) ~&,&~ A= s_\beta {\rm Im} \left( \phi_\ell^0 \right)-c_\beta {\rm Im}\left( \phi_h^0 \right) ~, \nonumber \\ h= c_\alpha {\rm Re}\left(\phi_h^0\right)-s_\alpha {\rm Re}\left(\phi_\ell^0\right) ~&,&~ H^+= s_\beta \phi_\ell^+-c_\beta \phi_h^+ ~, \label{Higgsangles} \end{aligned}\] where \(s_\alpha(c_\alpha)=\sin\alpha(\cos\alpha)\), \(\alpha\) being the Higgs mixing angle in the CP-even sector and \(s_\beta(c_\beta)=\sin\beta(\cos\beta)\), where \(\tan\beta \equiv v_h/v_\ell\) is the ratio between the VEV's of the heavy and light Higgs fields. The Yukawa Higgs-quark-quark interactions in the 4G2HDM are (similar terms can be written for the leptons): \[\begin{aligned} {\cal L}(h q_i q_j) &=& \frac{g}{m_W \sin 2\beta} \bar q_i \left\{ m_{q_i} s_\alpha s_\beta \delta_{ij}-\cos(\beta-\alpha) \cdot \left[ m_{q_i} \Sigma_{ij}^q R + m_{q_j} \Sigma_{ji}^{q \star} L \right] \right\} q_j h \label{Sff1}~, \\ {\cal L}(H q_i q_j) &=& \frac{g}{m_W \sin 2\beta} \bar q_i \left\{-m_{q_i} c_\alpha s_\beta \delta_{ij} + \sin(\beta-\alpha)\cdot \left[ m_{q_i} \Sigma_{ij}^q R + m_{q_j} \Sigma_{ji}^{q \star} L \right] \right\} q_j H ~, \\ {\cal L}(A q_i q_j) &=&-i I_q \frac{g}{m_W \sin 2\beta} \bar q_i \left\{ m_{q_i} s_\beta^2 \gamma_5 \delta_{ij}-\left[ m_{q_i} \Sigma_{ij}^q R-m_{q_j} \Sigma_{ji}^{q \star} L \right] \right\} q_j A ~, \end{aligned}\] \[\begin{aligned} {\cal L}(H^+ u_i d_j) = \sqrt{2} \frac{g}{ m_W \sin 2\beta} \bar u_i \left\{ \left[ m_{d_j} s_\beta^2 \cdot V_{u_id_j}-m_{d_k} V_{ik} \Sigma^{d}_{kj} \right] R + \left[-m_{u_i} s_\beta^2 \cdot V_{u_id_j} + m_{u_k} \Sigma^{u \star}_{ki} V_{kj} \right] L \right\} d_j H^+ \label{Sff2}~, \end{aligned}\] where \(V\) is the \(4 \times 4\) CKM matrix, \(q=d\) or \(u\) for down or up-quarks with \(I_d=-1\) and \(I_u=+1\), respectively, and \(R(L)=\frac{1}{2}\left(1+(-)\gamma_5\right)\). Also, \(\Sigma^d\) and \(\Sigma^u\) are new mixing matrices where all FCNC effects of the 4G2HDM are encoded. They are obtained after diagonalizing the quark mass matrices and, therefore, depend on the rotation (unitary) matrices of the right-handed down and up-quarks \(D_R\) and \(U_R\), respectively. In particular, for \({\cal I} \equiv {\rm diag}\left(0,0,0,1\right)\) in Eq. [\[eq:LY4G\]](#eq:LY4G){reference-type="ref" reference="eq:LY4G"}, we have (see ):\(^{[2]}\) \[\begin{aligned} \Sigma_{ij}^d = D_{R,4i}^\star D_{R,4j} ~,~ \Sigma_{ij}^u =U_{R,4i}^\star U_{R,4j} ~. \label{sigma} \end{aligned}\] The Yukawa structure and couplings defined by Eqs. [\[eq:LY4G\]](#eq:LY4G){reference-type="ref" reference="eq:LY4G"}-[\[sigma\]](#sigma){reference-type="ref" reference="sigma"} is assumed to be copied to the leptonic sector, see. In the following sections [\[sec3\]](#sec3){reference-type="ref" reference="sec3"} and [\[sec4\]](#sec4){reference-type="ref" reference="sec4"}, for illustrative purposes (and without loss of generality), we will set \(\Sigma^{d,u} \to {\rm diag}\left(0,0,0,1\right)\) in both the quark and lepton sectors, so that FCNC effects (in particular, between the 4th generation fermions and the SM fermions) are "turned off\". In fact, from the phenomenological point of view, it is sufficient to assume that \(\Sigma^{u}_{34,43} \to 0\) (i.e., forbidding the decay \(t^\prime \to t h\)) and \(V_{i4,4i} \to 0\) (\(i=1,2,3\), thus forbidding the decays \(t^\prime \to d_i W\) and \(b^\prime \to u_i W\) with \(d_i=d,s,b\) and \(u_i=u,c,t\)) in order to accommodate relatively light \(t^\prime\) and \(b^\prime\) with masses as low as 350 GeV, since the existing stringent exclusion limits of \(m_{t^\prime},m_{b^\prime} \:\raisebox{-0.5ex}{\)\stackrel{\textstyle>}{\sim}\(}\: 700\) GeV, are based on searches that assume 100% branching ratios of the 4th generation quarks into one of the channels: \(t^\prime \to th,tZ,d_i W\) and \(b^\prime \to Zb,u_iW\). We will, therefore, assume that the dominant \(t^\prime\) and \(b^\prime\) decays are into one of the FC channels \(t^\prime \to u_ih\) and \(b^\prime \to d_i h\) (\(u_i=u,c\) and \(d_i=d,s,b\)), due to small FCNC entries in \(\Sigma^{u,d}\) (which have no effect on the results presented in sections [\[sec3\]](#sec3){reference-type="ref" reference="sec3"} and [\[sec4\]](#sec4){reference-type="ref" reference="sec4"}), in which case small off-diagonal CKM entries \(V_{14,41}\) and/or \(V_{24,42}\) are also allowed as long as \(BR(t^\prime \to d_i W),~BR(b^\prime \to u_iW) \:\raisebox{-0.5ex}{\)\stackrel{\textstyle<}{\sim}\(}\: 0.5\). Such flavor structures, may have interesting phenomenological implications, as will be discussed in section [\[sec5\]](#sec5){reference-type="ref" reference="sec5"}. The 2HDM scalar sector is parameterized by seven free parameters (after minimization of the potential), which, in the so called "physical basis\", can be chosen as the four physical Higgs masses (\(m_h,~m_H,~m_A,~m_{H^+}\)), the two angles \(\beta\) and \(\alpha\) and one parameter from the scalar potential, which is needed in order to specify the scalar couplings, in particular, \(hH^+H^-\) (which enters in the 1-loop \(h \to \gamma \gamma\) decay), \(HH^+H^-\) (which enters the 1-loop \(H \to \gamma \gamma\) decay) and \(Hhh\) (required for the decay \(H \to hh\)). In the physical basis, these scalar couplings can be written at tree-level as (see e.g., ): \[\begin{aligned} \lambda_{Hhh}=-\frac{\cos(\alpha-\beta)}{2 v \sin 2\beta} \left[ \sin 2 \alpha \left(m_h^2+2m_H^2 \right)-\left( 3 \sin 2 \alpha-\sin 2 \beta \right) \frac{m_{\ell h}^2}{s_\beta c_\beta} \right] \label{lam1} ~, \end{aligned}\] \[\begin{aligned} \lambda_{hH^+H^-}=-\frac{1}{2 v \sin 2\beta} \left[ \left( \cos(\alpha-3\beta)+3 \cos(\alpha+\beta) \right) m_h^2-4 \sin 2 \beta \sin(\alpha-\beta) m_{H^\pm}^2-4 \cos(\alpha+\beta) \frac{m_{\ell h}^2}{s_\beta c_\beta} \right] \label{lam2} ~, \end{aligned}\] \[\begin{aligned} \lambda_{HH^+H^-}=-\frac{1}{2 v \sin 2\beta} \left[ \left( \sin(\alpha-3\beta)+3 \sin(\alpha+\beta) \right) m_H^2 + 4 \sin 2 \beta \cos(\alpha-\beta) m_{H^\pm}^2-4 \sin(\alpha+\beta) \frac{m_{\ell h}^2}{s_\beta c_\beta} \right] \label{lam3} ~, \end{aligned}\] where \(m_{\ell h}^2\) is a mass-like term, \(m_{\ell h}^2 \Phi_\ell^\dagger \Phi_h + h.c.\), which softly breaks the above \(Z_2\)-symmetry (i.e., \(\Phi_{\ell}\to-\Phi_{\ell},~ \Phi_{h}\to+\Phi_{h}\)), and which can be used to specify the above tree-level scalar couplings. However, since the working assumption of the 4G2HDM is that the scalar sector may be strongly interacting at the near by few TeV scale, the scalar potential is expected to be subject to significant renormalization and threshold effects. Thus, the above scalar couplings are expected to deviate from their tree-level values, depending on the details of the UV completion and on the masses of the heavy degrees of freedom of this model, see e.g.,. As an example, consider the 1-loop corrections to the \(Hhh\) coupling \(\lambda_{Hhh}\), for \(|\alpha| \to \pi/2\), in which case there is no mixing between the light and heavy Higgs fields (see Eq. [\[Higgsangles\]](#Higgsangles){reference-type="ref" reference="Higgsangles"}), as required in order to accommodate the 750 GeV \(\gamma \gamma\) excess in the 4G2HDM (see section [\[sec4\]](#sec4){reference-type="ref" reference="sec4"}). In this limit, the Yukawa couplings of the 4th generation fermions to the light Higgs state \(h\) (i.e., \(t^\prime t^\prime h\)) vanish (see Eq. [\[Sff1\]](#Sff1){reference-type="ref" reference="Sff1"} and Table [4](#tab4){reference-type="ref" reference="tab4"}) and we find that the dominant effect arises from the 1-loop triangle diagram with the charged Higgs exchange in the loop, giving a "renormalized\" \(Hhh\) coupling \(\bar\lambda_{Hhh} \equiv a_{Hhh} \lambda_{Hhh}\), with: \[\begin{aligned} a_{Hhh} \approx 1+ \frac{m_{\ell h}^4}{m_H^2 v^2} \frac{\left( 1-2c_\beta^2 \frac{m_{H^+}^2}{m_{\ell h}^2} \right) \left(1+c_\beta^2 \frac{m_{H}^2}{m_{\ell h}^2}-2s_\beta^2 \frac{m_{H^+}^2}{m_{\ell h}^2} \right)}{2 \pi^2 (\sin2\beta)^2} I\left(m_h,m_H,m_{H^+} \right)~, \end{aligned}\] where \(I\left(m_h,m_H,m_{H^+} \right)\) is the charged Higgs triangle loop integral, given by: \[\begin{aligned} I\left(m_h,m_H,m_{H^+} \right) =-\int_0^1 dx \int_0^{1-x} dy ~ \frac{1}{ (x+y) (x+y-1) m_h^2-xy m_H^2 + m_{H^+}^2 } ~. \end{aligned}\] In particular, one roughly finds \(|a_{Hhh}| \in \left\{0, 2\right\}\) when \(m_{H^+} \in \left\{500~{\rm GeV}, 1~{\rm TeV} \right\}\) and with \(m_H =750\) GeV, \(m_h =125\) GeV and \(m_{\ell h} \sim {\cal O}(1~{\rm TeV})\). For example, \(a_{Hhh} \sim-0.15\) for \(m_{H^+} = m_H = 750\) GeV and \(m_{\ell h} =1.2\) TeV. In what follows we will, therefore, define the "renormalized\" scalar couplings as: \(\bar\lambda_i \equiv a_i \lambda_i\), where \(\lambda_i\) (\(i=Hhh,~ hH^+H^-,~HH^+H^-\)) are the corresponding tree-level couplings in Eqs. [\[lam1\]](#lam1){reference-type="ref" reference="lam1"}-[\[lam3\]](#lam3){reference-type="ref" reference="lam3"}, and \(a_i\) will be treated as free-parameters in the fit that will be varied in the range \(|a_i| \in \left\{0, 2\right\}\). # The 125 GeV Higgs signals and PEWD [\[sec3\]]{#sec3 label="sec3"} The measured signals of the 125 GeV Higgs particle, which in the 4G2HDM is the light Higgs \(h\), and PEWD impose stringent constraints on the free parameter space of the 4G2HDM. For the 125 GeV Higgs signals we use the measured values of the "signal strength\" parameters, which are defined as the ratio between the measured rates and their SM expectation. In particular, for a specific production and decay channel \(i \to h \to f\), the signal strength is defined as: \[\begin{aligned} \mu_{i}^f \equiv \mu_i \cdot \mu^f ~, \end{aligned}\] with \[\begin{aligned} \mu_i = \frac{\sigma(i \to h)}{\sigma(i \to h)_{SM}} = k_i^2~,~ \mu^f = \frac{BR(h \to f)}{BR(h \to f)_{SM}} = \frac{k_f^2}{R^T} ~, \end{aligned}\] where \(k_j\) is the 4G2HDM coupling involved in \(j \to h\) or \(h \to j\) production or decay processes, normalized by its SM value, and \(R^T\) is the ratio between the total width of \(h\) in the 4G2HDM and the total width of the SM 125 GeV Higgs. In particular, \[\begin{aligned} k_j \equiv \frac{k_j^{4G2HDM}}{k_j^{SM}} ~,~ R^T \equiv \frac{\Gamma_{h_{4G2HDM}}^{Total}}{\Gamma_{h_{SM}}^{Total}}~, \end{aligned}\] so that \(\mu_i^f = k_i^2 k_f^2 / R^T\). In Table [1](#tab1){reference-type="ref" reference="tab1"} we list the latest combined ATLAS and CMS six parameter fit from RUN1, of the measured values for \(\mu_{gg}^{\gamma \gamma},~\mu_{gg}^{WW^\star}, ~\mu_{gg}^{ZZ^\star},~\mu_{gg}^{bb},~\mu_{gg}^{\tau \tau}\) and \(\mu_V/\mu_{gg}\), where \(\mu_V\) stands for Higgs production via vector-boson fusion (VBF) or in association with a vector-boson (VH).\(^{[3]}\). We also write in Table [1](#tab1){reference-type="ref" reference="tab1"} the model predictions for the various signal strengths in terms of the normalized couplings defined above. where \(S_{exp} = 0.06\) and \(T_{exp} = 0.1\) are the best fitted (central) values, \(\sigma_S = 0.09, \sigma_T = 0.07\) are the corresponding standard deviations and \(\rho=0.91\) is the (strong) correlation factor between S and T. We thus perform a random ("blind\") scan of the relevant parameter space, imposing compatibility at 95% CL of the 4G2HDM with the measured 125 GeV Higgs signals listed above and with the best fitted values of \(S\) and \(T\) using Eqs. [\[SandT\]](#SandT){reference-type="ref" reference="SandT"} and [\[ST2\]](#ST2){reference-type="ref" reference="ST2"}. In particular, we fix \(m_H=750\) GeV (for compatibility with the recent 750 GeV \(\gamma \gamma\) signal, see next section) and scan the rest of the parameters over the following ranges: \[\begin{aligned} \alpha \in \left[-\frac{\pi}{2},\frac{\pi}{2} \right] ~,~ \tan\beta \in \left[ 0.4, 10 \right]~,~ a_i\in \left[-2,2 \right] ~ (i=hH^+H^-,HH^+H^-,Hhh) ~, \nonumber \end{aligned}\] \[\begin{aligned} m_{\ell h}^2 \in \left[-\left(2~ {\rm TeV} \right)^2,\left(2~ {\rm TeV} \right)^2 \right] ~,~ m_{A,H^+} \in \left[ 300~ {\rm GeV} ,1.5~ {\rm TeV} \right] ~, \nonumber \end{aligned}\] \[\begin{aligned} m_{t^\prime,b^\prime} \in \left[ 350~{\rm GeV}, 500~{\rm GeV} \right] ~,~ m_{\nu^\prime,\tau^\prime} \in \left[ 200~{\rm GeV}, 1200~{\rm GeV} \right] ~. \end{aligned}\] We find two types of possible 4G2HDM "solutions\": In both cases above, \(m_A,~m_{H^+}\) and the 4th generation fermion masses can have values spanning over the entire scan ranges. In Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"} we plot the resulting distributions of the relevant parameter space in the \(S-T\), \(\tan\beta-\sin\alpha\) and \(\Delta m_{\ell^\prime}-\Delta m_{q^\prime}\) planes, where \(\Delta m_{\ell^\prime} \equiv m_{\nu^\prime}-m_{\tau^\prime}\) and \(\Delta m_{q^\prime} \equiv m_{t^\prime}-m_{b^\prime}\). We also show in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"} the resulting predicted 125 GeV Higgs signal strengths for the two cases above, which, as can seen, have different characteristics. We next discuss the compatibility of the above two 4G2HDM solutions with the recently observed 750 GeV \(\gamma \gamma\) excess. # The 4G2HDM and the 750 GeV \(\gamma \gamma\) resonance [\[sec4\]]{#sec4 label="sec4"} We search here for the portion of parameter space of the two 4G2HDM cases found in the previous section, that survive once the 4G2HDM is also required to accommodate the 750 GeV \(\gamma \gamma\) excess, which is being interpreted here as the decay of one or both of the heavy neutral Higgs (i.e., assumed to have masses \(\sim 750\) GeV) \(H \to \gamma \gamma\) and/or \(A \to \gamma \gamma\). Given the exploratory nature of our study, we will simplify our analysis at this point, assuming that the scalar spectrum have the characteristics of the so-called decoupling limit (see e.g., ). In particular, we assume that it is split into 2 typical scales: \(m_{light} \sim 125\) GeV, corresponding to the observed light Higgs and \(m_{heavy} \sim 750\) GeV around which the three heavy Higgs masses lie, i.e., \(m_H,m_A,m_{H^+} \sim 750\) GeV. Even though we find a wider range of allowed masses for the non-resonant heavy scalar states (i.e., for \(m_A\) and \(m_{H^+}\), see below) that can accommodate the 750 GeV signal, the choice \(m_H,~m_A,m_{H^+} \sim 750\) GeV will suffice for conveying our point: that the 750 GeV resonance in the \(\gamma\gamma\) channel can be accommodated by one of the heavy scalars of the 4G2HDM without any conflict with other existing relevant data. Indeed, if this measurement will be eventually confirmed, then it will be instructive to study the 4G2HDM within a wider range of the relevant parameter space. We, thus, re-scan the 4G2HDM parameter space corresponding to two 4G2HDM cases found in the previous section, where now \(m_H\), \(m_A\) and \(m_{H^+}\) are varied within a 30 GeV mass range around 750 GeV, i.e., \(m_{H,A,H^+} \in 750 \pm 30\) GeV. The scan is performed with the following additional "filters\"/requirements (i.e., in addition to the requirement of compatibility with PEWD and with the measured 125 GeV Higgs signals, as outlined in the previous section): - Reproducing the 750 GeV \(\gamma \gamma\) excess within the range \(3 ~{\rm fb} < \sigma(pp \to H/A \to \gamma \gamma) < 13~{\rm fb}\). We find that the (by far) dominant \(H\) and/or \(A\) production mechanism is the gluon-fusion one \(gg \to H/A\), so that all the relevant cross-sections \(\sigma(pp \to H/A \to f)\) are calculated in the narrow width approximation via: \[\begin{aligned} \sigma(pp \to H/A \to f) = \frac{C_{gg}}{s m_{H/A}} \Gamma(H/A \to gg) BR(H/A\to f) ~, \end{aligned}\] where \(\sqrt{s} =8\) or \(13\) TeV and \(C_{gg}\) is the gluon luminosity: \[\begin{aligned} C_{gg} = \frac{\pi^2}{8} \int_{m_{H/A}^2/s}^1 \frac{dx}{x} g(x) g\left( \frac{m_{H/A}^2}{sx} \right) ~, \end{aligned}\] giving \(C_{gg} \sim 2140(175)\) at \(\sqrt{s} =13(8)\) TeV, see. - The resonating scalar which produces the 750 GeV \(\gamma \gamma\) excess is required to have a width smaller than 45 GeV, i.e., \(\Gamma_{H/A} < 45\) GeV. - We impose the existing experimental bounds on the production and decays of the heavy neutral scalars \(H\) and \(A\), as obtained at the 8 and 13 TeV LHC runs (in particular when applied to \(m_H,m_A \sim 750\) GeV) in all other channels which are relevant to our study: \(pp \to W^+W^-,~ZZ,~t \bar t,~\tau \tau,~b \bar b,~hh,~hZ\). In particular, we use the 95% CL bounds in Table [2](#tab2){reference-type="ref" reference="tab2"} quoted in. ::: In particular, if \(m_H,m_A > m_{q^\prime}/2\), then \(H/A \to \bar q^\prime q^\prime\) is open and typically dominates, having a branching ratio of \({\cal O}(1)\) (see Table [4](#tab4){reference-type="ref" reference="tab4"}). In that case, we find that within the 4G2HDM parameter space discussed here, the corresponding resonance cross-sections for \(\bar q^\prime q^\prime\) production at the 13 TeV LHC are typically \(\sigma(pp \to H \to q^\prime q^\prime) \sim {\cal O}(10)\) \[pb\] and \(\sigma(pp \to A \to q^\prime q^\prime) \sim {\cal O}(0.1)\) \[pb\], (both \(H\) and \(A\) produced through gluon-fusion \(gg \to H/A\)), so that in the case of \(H \to q^\prime q^\prime\) (see Table [4](#tab4){reference-type="ref" reference="tab4"}), this is about an order of magnitude larger than the QCD (continuum) \(\bar q^\prime q^\prime\) production rate. Therefore, if the 750 GeV \(\gamma \gamma\) resonance persists, one should also expect an observable resonance signal at least in the \(H \to \bar q^\prime q^\prime\) channel. Let us, therefore, briefly investigate the signal \(H \to \bar q^\prime q^\prime\) under more general grounds, i.e., when \(m_H > m_{q^\prime}/2\) but not necessarily \(m_H \sim 750\) GeV. For example, in the case of \(H \to \bar t^\prime t^\prime\), the \(t^\prime\) will further decay either via the FC channels \(t^\prime \to uh\) (\(u=u\) or \(c\)) or via the 3-body decay \(t^\prime \to b^\prime W \to dhW\) (\(d=d,s,b\)), where \(b^\prime W\) are either off-shell or on-shell (i.e., when \(m_{t^\prime} > m_{b^\prime}+m_W\), see Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}). If the former case (i.e., \(t^\prime \to uh\)) dominates, then the resulting resonance signal should be searched for in \(pp \to \bar t^\prime t^\prime \to (jh)_{t^\prime}(jh)_{t^\prime}\) (\(j\) is a light jet), while if the 3-body \(t^\prime\) decay dominates then \(pp \to \bar t^\prime t^\prime \to (jhW^+)_{t^\prime}(jhW^-)_{t^\prime}\). In either case, the SM-like light Higgs (\(h\)) further decays into \(b \bar b\) or \(WW\) with SM rates, giving rise to resonance signatures of the form \(pp \to (nj+mb+\ell W)_H\), with \((n,m,\ell)=(2,4,0),(2,0,4),(2,2,2),(2,4,2),(2,2,4),(2,0,6),(0,2,6),(0,4,4),(0,6,2)\) and with unique kinematic features that distinguishes them from more conventional signatures. Similar signals are also expected for \(H \to \bar b^\prime b^\prime\). We recognize that these type of signals are very challenging and may require new strategies, in particular, for reconstructing the parent \(q^\prime\)'s in such a high jet-multiplicity environment. The decay pattern of the charged Higgs may also change in the 4G2HDM, in particular for the case when \(m_{H^+} > m_{t^\prime} + m_{b^\prime}\), for which the decay of \(H^+\) into a pair of heavy 4th generation fermions can dominate (see BMP1 in Table [4](#tab4){reference-type="ref" reference="tab4"}). In particular, taking \(m_{t^\prime} \sim m_{b^\prime} \equiv m_{q^\prime}\) and assuming that \(H^+\) is sufficiently heavier than \(2m_{q^\prime}\), so that we can ignore corrections of \({\cal O}(4m_{q^\prime}^2/m_{H^+}^2)\) in the phase-space factors, we have in the 4G2HDM: \[\begin{aligned} R_{t^\prime b^\prime/tb} &\equiv& \frac{\Gamma(H^+ \to t^\prime b^\prime)}{\Gamma(H^+ \to t b)} \sim 2 \frac{m_{q^\prime}^2}{m_t^2} \cot^4\beta ~, \\ R_{t^\prime b^\prime/Wh} &\equiv& \frac{\Gamma(H^+ \to t^\prime b^\prime)}{\Gamma(H^+ \to Wh)} \sim 12 \frac{m_{q^\prime}^2}{m_{H^+}^2} \left(\frac{\cot\beta}{\cos(\beta-\alpha)}\right)^2 ~. \end{aligned}\] Thus, for \(\alpha \sim-\pi/2\), \(\tan\beta \sim 0.5\) (i.e., \(\cos(\beta-\alpha) \sim-0.45\)), \(m_{q^\prime} \sim 350\) GeV (i.e., values of the 4G2HDM parameter space that can accommodate the 750 GeV \(\gamma \gamma\) signal) and taking \(m_{H^+} \sim {\cal O}(1)\) TeV, we obtain: \(R_{t^\prime b^\prime/tb} \sim {\cal O}(100)\) and \(R_{t^\prime b^\prime/Wh} \sim {\cal O}(10)\), in which case \(BR(H^+ \to t^\prime b^\prime) \sim 1\) (e.g., as in the case of BMP2), leading to some interesting signatures of the heavy charged Higgs at the LHC. In particular, the dominant production channels of \(H^+\) at the LHC are \(gg/gb \to H^+ b \bar t, H^+ W^-/ H^+ \bar t\), with a typical cross-section of \(\sim 100\) fb when \(\tan\beta \sim 1\). The subsequent \(H^+\) decay to a pair of 4th generation heavy fermions with \(BR(H^+ \to t^\prime \bar b^\prime) \sim 1\) will, thus, lead to new \(H^+\) signals, e.g., \(pp \to t (t^\prime b^\prime)_{H^+} \to (b W)_t (jh)_{t^\prime} (jh)_{b^\prime}\), again with the typical 4G2HDM heavy fermion high jet-multiplicity signatures of the form \(pp \to nj+mb+\ell W\). This is in contrast to "standard\" 2HDM frameworks where the heavy charged Higgs will dominantly decay to \(Wh\) and/or \(tb\) (see BMP1 and BMP3), leading to a lower multiplicity of jets in the final state. As noted earlier, a wider range of solutions exist (which are not being discussed here) to all data and filters mentioned above (i.e., including the 750 GeV \(\gamma \gamma\) resonance), in which lighter pseudoscalar \(A\) and charged Higgs \(H^{+}\) are allowed, with masses as low as \(300\) GeV. In such 4G2HDM scenarios, the heavy 4th generation quarks (and leptons) can have substantial decay rates in channels involving also the heavy Higgs species, i.e., \(t^\prime \to H^+ d, Au\) (\(d=d,s,b\) and \(u=u,c)\) and \(b^\prime \to H^+ u, Ad\) (\(d=d,s,b\) and \(u=u,c)\), followed by \(H^+ \to W^+ h, t \bar b\) and \(A \to hZ, t \bar t\). Indeed, such decay patterns can also lead to some un-explored collider signatures of the 4G2HDM. We leave the discussion of the phenomenology of such wider range of 4G2HDM scenarios to a later work. Finally, we wish to comment on the flavor violating structure of the 4G2HDM and its compatibility with the recently reported indications of the LFV decay of the 125 GeV light Higgs \(h \to \tau \mu\). Writing the LFV couplings of \(h\) in a general form: \[\begin{aligned} {\cal L}(h f_i f_j) = {\cal S}_{ij} + {\cal P}_{ij} \gamma_5 ~, \end{aligned}\] one obtains: \[\begin{aligned} \Gamma(h \to \bar f_i f_j + \bar f_j f_i) = \frac{m_h}{4 \pi} \left( |{\cal S}_{ij}|^2 + |{\cal P}_{ij}|^2 \right) ~. \end{aligned}\] In our 4G2HDM we have for the case of the LFV decay \(h \to \tau \mu\) (neglecting terms of \({\cal O}(m_\mu/m_\tau)\), see Eq. [\[Sff1\]](#Sff1){reference-type="ref" reference="Sff1"}): \[\begin{aligned} |{\cal S}_{\tau \mu}| = |{\cal P}_{\tau \mu}| \sim \frac{g}{4} \frac{m_\tau}{m_W} f(\beta,\alpha) \end{aligned}\] where we have defined \(\Sigma^\ell_{32} = \Sigma^\ell_{23} \equiv \begin{eqnarray} f(\beta,\alpha) = \frac{\cos(\beta-\alpha)}{s_\beta c_\beta} ~. \end{eqnarray} Requiring now that\)BR(h ) %\(we find: \begin{eqnarray} | f(\beta,\alpha) \end{eqnarray} Thus, since for the values of\[and\]that were found to be compatible with all data considered in the previous sections, we find\)\|f(,)\| \~1-5\(, and specifically\)f(,) \~1\(for\)-/2\(and\)\~0.5\(, as required in order to accommodate the 750 GeV\)\(resonance (see previous section), the 4G2HDM with\)\| \(BR(h \to \tau \mu) \:\raisebox{-0.5ex}{\)\stackrel{\textstyle<}{\sim}\(}\: 1\%\) if it persists. # Summary [\[sec6\]]{#sec6 label="sec6"} We have revisited a class of models beyond the SM, suggested by us a few years ago in, which put together an additional Higgs doublet with a heavy chiral 4th generation quark and lepton doublet and which have several important and attractive theoretical features. In particular, we focused on the so-called 4G2HDM of type I (in ), in which a discrete \(Z_2\) symmetry couples the "heavy\" scalar doublet only to the heavy 4th generation fermions and the "light\" one to the lighter SM fermions. We have confronted this model with PEWD, with the measured 125 GeV light Higgs signals and also studied its compatibility with the recent indication of a 750 GeV \(\gamma \gamma\) resonance and with the current LHC bounds on heavy scalar resonances in other relevant channels. We found that the CP-even heavy Higgs state of the 4G2HDM with a mass \(\sim 750\) GeV can accommodate the measured \(750\) GeV excess for a rather unique choice of the parameter space: \(\tan\beta \sim 0.5\), \(\alpha \sim-\pi/2\) (the Higgs mixing angle) and with heavy chiral fermion masses \(m_{t^\prime,b^\prime} \:\raisebox{-0.5ex}{\)\stackrel{\textstyle<}{\sim}\(}\: 400\) GeV and \(m_{\nu^\prime,\tau^\prime} \:\raisebox{-0.5ex}{\)\stackrel{\textstyle>}{\sim}\(}\: 900\) GeV. We have shown that the heavy chiral quarks (and leptons) of the 4G2HDM may have FCNC decays into the light 125 GeV Higgs plus a light-quark jet, \(q^\prime \to j h\), with branching ratios of \({\cal O}(1)\), thus leading to some un-explored signatures of \(q^\prime \bar q^\prime\) production at the LHC and, therefore, being consistent with the current direct bounds on the masses of new heavy fermions. Indeed, new and rich phenomenology in \(q^\prime\)-heavy Higgs systems is expected, including possible resonance production of \(q^\prime q^\prime\) pairs via either the heavy neutral or heavy charged Higgs particles of the 4G2HDM, which leads to high jet-multiplicity signatures, with or without charged leptons, of the form \(\bar q^\prime q^\prime \to nj + mb + \ell W\), with \(n+m+\ell=6-8\) and unique kinematic features which are related to the resonating heavy scalar and the decay pattern of the heavy quarks. The reconstruction of the \(q^\prime q^\prime\) pairs in such high jet-multiplicity signals is very challenging and require more thought and possibly new search strategies. We also show that the recent indication of a percent-level branching ratio in the LFV decay of the 125 GeV Higgs \(h \to \tau \mu\), if it persists, can be readily addressed within the distinct flavor structure of the 4G2HDM. We thank Pier Paolo Giardino for useful conversations. The work of AS was supported in part by the US DOE contract \#DE-SC0012704.
{'timestamp': '2016-07-19T02:00:14', 'yymm': '1607', 'arxiv_id': '1607.04643', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04643'}
null
null
null
null
# Introduction {#sec:intro} Galaxy clusters form from large primordial density fluctuations that have collapsed and virialised by the present epoch, with more massive clusters forming from larger and rarer fluctuations. This makes them especially sensitive to fundamental cosmological parameters, such as the matter density, the amplitude of the matter power spectrum and the equation of state of dark energy. The observable properties of a galaxy cluster result from a non-trivial interplay between gravitational collapse and astrophysical processes. The diverse range of formation histories of the cluster population leads to scatter in the observable-mass scaling relations and, as surveys select clusters based on an observable, this can lead to a biased sample of clusters, resulting in systematics when using them as a cosmological probe. Many previous studies have shown that the relationship between a cluster observable, such as its temperature or X-ray luminosity, and a quantity of interest for cosmology, e.g. its mass, has a smaller scatter for more massive, dynamically relaxed objects. Therefore, the fundamental requirement when probing cosmological parameters with galaxy clusters is a sample of relaxed, massive clusters with well calibrated mass-observable scaling relations. However, galaxy clusters are rare objects, becoming increasingly rare with increasing mass, and to observe a sample large enough to be representative of the underlying population requires a survey with significant size and depth. Currently ongoing and impending observational campaigns, such as the Dark Energy Survey, *eRosita*, *Euclid*, [SPT-3G]{.smallcaps} and Advanced ACTpol, will be the first to have sufficient volume to yield significant samples of massive clusters. Due to their rarity, the majority of these massive clusters will be at high redshift and it is therefore critical to understand how the cluster observables and their associated scatter evolve. Additionally, the most massive clusters will be the brightest and easiest to detect objects at high redshift, making it vital to understand the selection function of the chosen cluster observable and whether the most massive clusters are representative of the underlying cluster population. Theoretical modelling of the formation of clusters and their observable properties is required to understand these issues and to further clusters as probes of cosmology. Due to the range of scales involved in cluster formation, the need to incorporate astrophysical processes and to self-consistently predict observable properties, cosmological hydrodynamical simulations are the only viable option. Recent progress in the modelling of large-scale structure formation has been driven mainly by the inclusion of supermassive black holes and their associated Active Galactic Nucleus (AGN) feedback, which has been shown to be critical for reproducing many cluster properties. A number of independent simulations are now able to produce realistic clusters that simultaneously reproduce many cluster properties in good agreement with the observations. Results from the recent BAryons and HAloes of MAssive Systems (BAHAMAS) simulations have shown that by calibrating the subgrid model for feedback to match a small number of key observables, in this case the global galaxy stellar mass function and the gas fraction of clusters, simulations of large-scale structure are now able to reproduce many observed scaling relations and their associated scatter over two decades in halo mass. However, full gas physics simulations of large-scale structure formation, with sufficient resolution, are still computationally expensive. This has limited previous studies to either small samples with \(<50\) objects or to volumes of \(596\,\mathrm{Mpc}\), all of which are too small to contain the representative sample of massive clusters that is required for cosmological studies above \(z=0\). This paper introduces the Virgo consortium's MACSIS project, a sample of \(390\) massive clusters selected from a large volume dark matter simulation and resimulated with full gas physics to enable self-consistent observable predictions. The simulations extend the BAHAMAS simulations to the most massive clusters expected to form in a \(\Lambda\rm{CDM}\) cosmology. In this paper we study the cluster scaling relations and their evolution. We combine the MACSIS and BAHAMAS simulations to produce a sample that spans the complete mass range and that can be studied to high redshift, using the progenitors of the MACSIS sample. We also select the hottest clusters from the combined sample and a relaxed subset of them to examine the impact of such selections on the scaling relations and their evolution. We then study the gas profiles to further understand the differences between the samples. This paper is organised as follows. In Section [2](#sec:MACsamp){reference-type="ref" reference="sec:MACsamp"} we introduce the MACSIS sample and discuss the parent dark matter simulation from which the sample was selected, the selection criteria used, the model used to resimulate the haloes, how we produced the observable quantities and the three samples we use in this work. In Section [3](#sec:screlations){reference-type="ref" reference="sec:screlations"} we investigate how the scaling relations evolve and how this evolution changes when a hot cluster sample or relaxed, hot cluster sample is selected. We then study the hot gas profiles to understand the differences in the evolution of the relations for the different samples in Section [4](#sec:gasprofs){reference-type="ref" reference="sec:gasprofs"}. Finally, in Section [5](#sec:sad){reference-type="ref" reference="sec:sad"} we discuss our results and summarise our main findings. # Parent simulation and sample selection {#sec:MACsamp} In this section we describe the parent simulation, the selection of the MACSIS sample, the baryonic physics used in the resimulation of the sample and the calculation of the observable properties of the resimulated clusters. Additionally, we describe how MACSIS and BAHAMAS clusters were selected to produce the combined sample and the cuts made to yield a hot sample and its relaxed subset. ## The parent simulation To obtain a population of massive clusters we require a simulation with a very large volume \((> 1\,\rm{Gpc}^{3})\). With current computational resources it is unfeasible to simulate such a volume with hydrodynamics and the required gas physics, such as radiative cooling, star formation and feedback, at a resolution high enough to accurately capture the cluster properties. An alternative option is to apply the zoomed simulation technique to a representative sample of objects from a larger volume. Therefore, we select a sample of massive haloes from a dark matter only simulation that has sufficient volume to yield a population of massive clusters and the resolution to ensure they are well characterized. We label this simulation the 'parent' simulation. The parent simulation is a periodic cube with a side length of \(3.2\,\rm{Gpc}\). Its cosmological parameters are taken from the Planck 2013 results combined with baryonic acoustic oscillations, WMAP polarization and high multipole moments experiments and are \(\Omega_{\rm{b}}=0.04825\), \(\Omega_{\rm{m}}=0.307\), \(\Omega_{\Lambda}=0.693\), \(h\equiv H_0/(100\,\rm{km}\,\rm{s}^{-1}\,\rm{Mpc}^{-1})=0.6777\), \(\sigma_{8}=0.8288\), \(n_{\rm{s}}=0.9611\) and \(Y=0.248\). We note that there are minor differences between these values and the Planck-only cosmology used for the BAHAMAS simulations, but this has negligible impact on the results presented here. The simulation contained \(N=2520^3\) dark matter particles that were arranged in an initial glass-like configuration and then displaced according to second-order Lagrangian perturbation theory \((\mathrm{2LPT})\) using the [ic_2lpt_gen]{.smallcaps} code and the public Gaussian white noise field *Panphasia*. [^1] The particle mass of this simulation is \(m_{\rm{DM}}=5.43\times10^{10}\,\mathrm{M}_{\rm{\odot}}/h\) and the comoving gravitational softening length was set to \(40\,\mathrm{kpc}/h\). The simulation was evolved from redshift \(z=127\) using a version of the Lagrangian TreePM-SPH code [p-gadget3]{.smallcaps}. Haloes were identified at \(z=0\) using a *Friends-of-Friends* (FoF) algorithm with a standard linking length of \(b=0.2\) in units of the mean interparticle separation. We plot the FoF mass function of the parent simulation at \(z=0\) in Fig. [\[fig:massfunc\]](#fig:massfunc){reference-type="ref" reference="fig:massfunc"}. We compare it to the published relations of, , and. We plot the scaled differential mass function \[f(\sigma) = \frac{M}{\bar{\rho}}\frac{dn}{d\ln\sigma^{-1}}(M,z)\:,\] where \(M\) is halo mass, \(\bar{\rho}\) is the mean density of the Universe at \(z=0\), \(n\) is the number of haloes per unit volume, and \(\sigma^{2}\) is the variance of the linear density field when smoothed with a top-hat filter. We plot the mass function as a function of the variable \(\ln(\sigma^{-1})\) as it is insensitive to cosmology. For \(\ln(\sigma^{-1}) < 0.3\) we find that all of the mass functions show reasonable agreement with differences of \(\sim5-10\%\) between them, with the small differences likely due to the mass function not being exactly universal. However, for larger values the mass functions begin to diverge, as the parent simulation has an excess of massive clusters compared to the other simulations. This is likely due to two effects. First, the MACSIS simulation is the only one to use \(\mathrm{2LPT}\) when generating the initial conditions. It has been shown that not using \(\mathrm{2LPT}\) results in a significant underestimation of the abundance of the rarest objects. The second effect is simply statistics: even in a very large volume there are still low numbers of the rarest and most massive clusters, where there is likely to be significant variance between the simulation volumes. ## The MACSIS sample To select the MACSIS sample, all haloes with \(M_{\rm{FoF}}>10^{15}\,\mathrm{M}_{\odot}\) were grouped in logarithmically spaced bins, with \(\Delta\log_{10}\,M_{\rm{FoF}}=0.2\). If a bin contained less than one hundred haloes then all of the objects in that bin were selected. For bins with more than one hundred objects the bin was then further subdivided into bins of 0.02 dex and ten objects from each sub-bin were then selected at random. The subdividing of the bins ensured that our random selection was not biased to low masses by the steep slope of the mass function. This selection procedure results in a sample of \(390\) haloes that is mass limited above \(10^{15.6}\,\rm{M}_{\odot}\) and randomly sampled below this limit. Table [1](#tab:smpcomp){reference-type="ref" reference="tab:smpcomp"} shows the fraction of haloes selected from the parent simulation in each mass bin. We have compared the properties of the selected haloes with those of the underlying population and found the MACSIS sample to be representative. Additionally, in Appendix [\[app:seleff\]](#app:seleff){reference-type="ref" reference="app:seleff"} we demonstrate that selecting by a halo's FoF mass does not bias our results when binning clusters by their \(M_{500}\). [\[tab:smpcomp\]]{#tab:smpcomp label="tab:smpcomp"} Due to current computational constraints the BAHAMAS simulations are limited to periodic cubes with a side length of \(596\,\mathrm{Mpc}\). There are very few clusters with a mass greater than \(10^{15}\,\mathrm{M}_{\odot}\) in a volume of this size, and those that are present may be affected by the loss of power from large-scale modes that are absent due to their wavelengths being greater than the box size. The zoom simulations of the MACSIS project provide an extension to the BAHAMAS periodic simulations. They provide the most massive clusters and allow the mass-observable scaling relations to be studied across the complete cluster mass range. We use the zoomed simulation technique to re-simulate the chosen sample at increased resolution. We perform both DM only and full gas physics re-simulations. The Lagrangian region for every cluster was selected so that its volume was devoid of lower resolution particles beyond a cluster centric radius of \(5r_{200}\).[^2] The resolution of the Lagrangian region was increased such that the particles in the DM only simulations had a mass of \(m_{\rm{DM}}=5.2\times10^9\,\mathrm{M}_{\odot}/h\) and in the hydrodynamic re-simulations the dark matter particles had a mass of \(m_{\rm{DM}}=4.4\times10^9\,\mathrm{M}_{\odot}/h\) and the gas particles had an initial mass of \(m_{\rm{gas}}=8.0\times10^8\,\mathrm{M}_{\odot}/h\). In all simulations the Plummer equivalent gravitational softening length for the high-resolution particles was fixed to \(4\mathrm{kpc}/h\) in comoving units for \(z>3\) and in physical coordinates thereafter. The smoothed particle hydrodynamics interpolation used \(48\) neighbours and the minimum smoothing length was set to one tenth of the gravitational softening. A schematic view of the zoom approach is shown in Fig. [\[fig:zoomed\]](#fig:zoomed){reference-type="ref" reference="fig:zoomed"}. The resolution and softening of the zoom re-simulations were deliberately chosen to match the values of the periodic box simulations of the BAHAMAS project, which is a calibrated version of the OWLS code, which was also used for cosmo-OWLS. The subgrid models for feedback from star formation and AGN used in the BAHAMAS simulations was calibrated to obtain a good fit to the observed galaxy stellar mass function and the amplitude of the gas fraction-total mass relation, respectively, at \(z=0\). Without any further tuning, the simulations then produce a population of groups and clusters that shows excellent agreement with the observations for a range of galaxy-halo, hot gas-halo and galaxy-hot gas relations. ## Baryonic physics The BAHAMAS simulations were run with a version of [p-gadget3]{.smallcaps} that has been heavily modified to include new subgrid physics as part of the OWLS project. We now briefly describe the subgrid physics, but refer the reader to, and for greater detail, including the impact of varying the free parameters in the model and the calibration strategy. Radiative cooling is calculated on an element-by-element basis following, interpolating the rates as a function of density, temperature and redshift from pre-computed tables generated with [cloudy]{.smallcaps}. It accounts for heating and cooling due to the primary cosmic microwave background and a ultra-violet/X-ray background. The background due to reionization is assumed to switch on at \(z=9\). Star formation is modelled stochastically in a way that by construction reproduces the observations, as discussed in. Lacking the resolution and physics to correctly model the cold interstellar medium, gas particles with a density \(n_{\rm{H}} > 0.1\,\rm{cm}^{-3}\) follow an imposed equation of state with \(P\propto\rho^{4/3}\). These gas particles then form stars at a pressure-dependent rate that reproduces the observed Kennicutt-Schmidt law. Stellar evolution and the resulting chemical enrichment are implemented using the model of, where \(11\) chemical elements (H, He, C, N, O, Ne, Mg, Si, S, Ca and Fe) are followed. The mass loss rates are calculated assuming Type Ia and Type II supernovae and winds from massive and asymptotic giant branch stars. Stellar feedback is implemented via the kinetic wind model of. The BAHAMAS simulations used the calibrated mass-loading factor of \(\eta_{\rm{w}}=2\) and wind velocity \(v_{\rm{w}}=300\,\rm{km/s}\). This corresponds to \(20\) percent of available energy from Type II supernovae, assuming a IMF, and yields an excellent fit to the observed galaxy mass function. The seeding, growth and feedback from supermassive black holes (BH) is implemented using the prescription of, a modified version of the method developed by. A FoF algorithm is run on-the-fly and BH seed particles, with \(m_{\mathrm{BH}}=10^{-3}m_{\mathrm{gas}}\), are placed in haloes that contain at least \(100\) DM particles, which corresponds to a halo mass of \(\sim5\times10^{11}\,\rm{M}_{\odot}\). BHs grow via Eddington-limited accretion of gas at the Bondi-Hoyle-Littleton rate, with a boost factor that is a power-law of the local density for gas above the star formation density threshold. They also grow by direct mergers with other BHs. A fraction, \(\epsilon\), of the rest mass energy of the accreted gas is then used to heat \(n_{\rm{heat}}\) neighbour particles by increasing their temperature by \(\Delta T_{\rm{heat}}\). Changes to these parameters have a significant impact on the hot gas properties of clusters. The calibrated values of these parameters in the BAHAMAS simulations are \(n_{\rm{heat}}=20\) and \(\Delta T_{\rm{heat}}=10^{7.8}\,\rm{K}\). The feedback efficiency \(\epsilon=\epsilon_{\rm{r}}\epsilon_{\rm{f}}\), where \(\epsilon_{\rm{r}}=0.1\) is the radiative efficiency and \(\epsilon_{\rm{f}}=0.15\) is the fraction of \(\epsilon_{\rm{r}}\) that couples to the surrounding gas. The choice of the efficiency, assuming it is non-zero, is generally of little consequence as the feedback establishes a self-regulating scenario, but determines the black hole masses. ## Calculating observable properties Previous studies have shown that there can be significant biases in the observable properties of clusters due to issues such as multi-temperature structures and gas inhomogeneities. Therefore, when investigating cluster properties it is critical that, as far as possible, we make a like-with-like comparison with the observations. Following, we do this by producing synthetic observational data for each cluster and analysing it in a manner similar to what is done for real data. Using the particle's temperature, density and metallicity, where the metallicity is smoothed over a particle's neighbours, we first compute a rest-frame X-ray spectrum in the \(0.05-100.0\,\rm{keV}\) band for all gas particles, using the Astrophysical Plasma Emission Code via the [pyatomdb]{.smallcaps} module with atomic data from [atomdb]{.smallcaps} v3.0.2. A particle's spectrum is a sum of the individual spectra for each chemical element tracked by the simulations, scaled by the particle's elemental abundance. We ignore particles with a temperature lower than \(10^5\,\rm{K}\) as they make a negligible contribution to the total X-ray emission. We then estimate the density, temperature and metallicity of the hot gas in \(25\) logarithmically spaced radial bins by fitting a single temperature [apec]{.smallcaps} model, with a fixed metallicity, to the summed spectra of all particles that fall within that radial bin. We then scale the spectra by the relative abundance of the heavy elements as the fiducial spectra assume solar abundance. The spectra have an energy resolution of \(150\,\rm{eV}\) in the range \(0.05-10.0\,\rm{keV}\) and are logarithmically spaced between \(10.0-100.0\,\rm{keV}\). To get a closer match to the observations, we multiply the spectra by the effective area of *Chandra*. To derive temperature and density profiles of a cluster, we fit the spectrum in the range \(0.5-10.0\,\rm{keV}\) for each radial bin with a single temperature model using a least-squares approach. The temperature and density profiles derived from the X-ray spectra are then used to perform a hydrostatic mass analysis of the cluster. The profiles are fit with the density and temperature models proposed by to produce a hydrostatic mass profile. We then derive various mass and radius estimates, such as \(M_{500}\) and \(r_{500}\), from the hydrostatic mass profiles. With these estimates we calculate quantities, such as \(M_{\rm{gas}}\) or \(Y_{\rm{SZ}}\), by summing the properties of the particles that fall within the set. Core-excised quantities are calculated in the radial range \(0.15-1.0\) of the aperture. Luminosities are calculated by integrating the spectra of all particles within the aperture in the requisite energy band, for example, bolometric luminosities are calculated in the range \(0.05-100.0\,\rm{keV}\). Averaged X-ray temperatures are calculated by fitting a single temperature model to the sum of the spectra of all particles within the aperture. We repeat this analysis for all clusters in the combined sample at all redshifts of interest. All quantities derived in this manner are labelled with the sub-script 'spec'. ## Cluster sample selection We select clusters from MACSIS and BAHAMAS to form a 'combined' sample with which we can investigate the cluster scaling relations. We perform our analysis at \(z=0.0,0.25,0.5,1.0\) and \(1.5\). We create this sample at each redshift by selecting all clusters with a mass \(M_{500,\rm{spec}}\geq10^{14}\,\rm{M}_{\odot}\). Additionally, we introduce a mass cut at every redshift below which we remove any MACSIS clusters. For example, at \(z=0\) (\(z=1\)) this cut is made at \(M_{500,\mathrm{spec}}=10^{14.78}\,\rm{M}_{\odot}\) (\(M_{500,\mathrm{spec}}=10^{14.3}\,\rm{M}_{\odot}\)). This removes a tail of clusters with low \(M_{500,\mathrm{spec}}\), but have high \(M_{\mathrm{FoF}}/M_{500,\mathrm{spec}}\) ratios (see Appendix [\[app:seleff\]](#app:seleff){reference-type="ref" reference="app:seleff"}). For the luminosity-temperature relation, we use the temperature-mass relation of the combined sample to convert the mass cut into a temperature cut. At \(z=0\) this results in a sample of \(1294\) clusters, containing \(1098\) clusters from BAHAMAS and \(196\) MACSIS clusters, and at \(z=1\) a sample of \(225\) clusters, \(99\) from BAHAMAS and \(126\) from MACSIS. The MACSIS clusters enable the investigation of the behaviour of the most massive clusters at low redshift. These clusters are commonly selected in cosmological analyses because their deep potentials are expected to reduce the impact of non-gravitational processes and as the brightest clusters they require shorter exposures. We select a hot, and therefore massive, cluster sample by selecting all clusters in the combined sample with a core-excised X-ray temperature greater than \(5\,\rm{keV}\). At \(z=0\) (\(z=0.5\)) this yields a sample of \(244\) (\(186\)) clusters, with \(190\) (\(173\)) coming from the MACSIS sample. Finally, we examine the impact of selecting a relaxed subset of the hot cluster sample. Theoretically, there are many ways to define a relaxed halo. For this study we use the following criteria \[X_{\rm{off}} < 0.07\,;~f_{\rm{sub}} < 0.1~\rm{and}~\lambda < 0.07\,, \nonumber\] where \(X_{\rm{off}}\) is distance between the cluster's minimum gravitational potential and centre of mass, divided by its virial radius; \(f_{\rm{sub}}\) is the mass fraction within the virial radius that is bound to substructures; and \(\lambda\) is the spin parameter for all particles inside \(r_{200}\). These criteria are not designed to select a small subset that comprises the most relaxed objects, but to simply remove those clusters that are significantly disturbed. This results in a subsample at \(z=0\) (\(z=0.5\)) that contains \(213\) (\(117\)) clusters, with \(177\) (\(111\)) coming from the MACSIS sample. # The scaling relations of massive clusters {#sec:screlations} In this section we present our main results, measuring the scaling relations of our cluster samples across a range of redshifts. ## Comparison to observational data Fig. [\[fig:observations\]](#fig:observations){reference-type="ref" reference="fig:observations"} shows the gas mass, \(M_{\rm{gas},500,\rm{spec}}\), the integrated Sunyaev-Zel'dovich (SZ) signal, \(Y_{\rm{SZ}}\), measured in a \(5r_{500,spec}\) aperture as a function of estimated total mass, \(M_{500,\rm{spec}}\), (at \(z=0\) and \(z=1\)) and the core-excised bolometric X-ray luminosity, \(L^{\rm{X,ce}}_{500,\rm{spec}}\), as a function of core-excised X-ray temperature, \(T^{\rm{X,ce}}_{500,\rm{spec}}\), for the combined sample. We compare the sample to the relevant observational data. At all redshifts the MACSIS sample provides a consistent extension to the BAHAMAS clusters with similar scatter. At low redshift, have shown that the BAHAMAS sample shows good agreement with the observed median relations and shows similar intrinsic scatter. The MACSIS sample continues this agreement to observed high-mass clusters, though there are significantly fewer clusters to compare against. In detail, it appears that the \(M_{500,\mathrm{gas,spec}}\)-\(M_{500,\mathrm{spec}}\) and \(L^{\mathrm{X,ce}}_{500,\mathrm{spec}}\)-\(T^{\mathrm{X,ce}}_{500,\mathrm{spec}}\) relations are slightly steeper than observed. However, we would exercise caution as we have not applied the same selection criteria as was used for the observational X-ray analyses. At high redshift observational data becomes sparse and currently only SZ surveys have detected a reasonable number of clusters. At \(z=1\) these clusters are all significantly more massive than any cluster in the BAHAMAS volume. However, the progenitors of the very massive MACSIS clusters provide a sample that can be compared with these observations. We find that the median relation shows good agreement with the observations and the intrinsic scatter of the clusters about the median relation is consistent with the scatter in the observations. Overall, we find that all quantities computed in a like-with-like manner show good agreement with the observations. [\[tab:z0bestfit\]]{#tab:z0bestfit label="tab:z0bestfit"} [\[tab:Ppro_fit\]]{#tab:Ppro_fit label="tab:Ppro_fit"} At \(z=1\) the median profiles of the three samples are in closer agreement with each other, because the minimum mass limit of \(M=10^{14}\,\rm{M}_{\odot}\) causes the samples to converge at high redshift. We compare our median pressure profiles with the observed profile of. They find a median pressure profile that is in good agreement with the median profiles, but it is most consistent with the relaxed hot sample of massive clusters. The pressure profile of the relaxed subset shows very little evolution between \(z=1\) and \(z=0\), except for the core where the increasing density leads to an increased pressure with decreasing redshift. The hot sample shows an increased pressure in the core with decreasing redshift, due to the increased density, but a negative change in pressure from \(z=1\) to \(z=0\) at larger radii. The combined sample shows a negative pressure change between \(z=1\) and \(z=0\) at all radii. The decreased pressure with decreasing redshift is caused by the decrease in density from \(z=1\) to \(z=0\). ## Entropy Profiles The median entropy profiles are shown in the bottom right panel of Fig. [\[fig:enty_prof\]](#fig:enty_prof){reference-type="ref" reference="fig:enty_prof"} and they have been normalized by the predicted self-similar entropy. We note that we define entropy as \[K_{\Delta}\equiv\frac{k_{\rm{B}}T_{\Delta}}{n^{2/3}_{e,\Delta}}\,\] where \(n_e\) is the electron number density and \(\Delta\) is the chosen overdensity relative to the critical density of the Universe. At \(z=0\) the the combined sample shows a higher normalization compared to the hot sample and its relaxed subset. This is due to it lower density profile and higher temperature profile. The gradients of the hot sample and the relaxed subset profiles steepen in the centre due to the accretion of low entropy gas. We compare with the observed median profiles of and, and the baseline profile of derived from non-radiative SPH simulations. The combined sample is in good agreement with the observations and tends to the non-radiative predictions at large radii. At \(z=1\) the three samples are in reasonable agreement with each other, all having a similar shape with the hot sample showing a marginally lower normalization. This change from \(z=0\) is in agreement with the evolution in their density and temperature profiles. We compare the profiles to the observations of. The combined and relaxed hot sample show good agreement with the observed profile for \(r<0.5r_{500,\rm{spec}}\), but over predict the entropy at larger radii. In contrast the median profile of the hot sample is consistent with the observations at large radii, but under predicts the entropy in the centre of the cluster. The departure from self-similarity for the three samples is due to a combination of the evolution in their temperature and density profiles. The relaxed hot sample shows a mild increase in entropy from \(z=1\) to \(z=0\) at large radii, due to change in its temperature profile, and a decrease in entropy in the core due to the increase in density at \(z=0\). The increased normalization of the hot sample's temperature profile at \(z=0\) compared to \(z=1\) leads to an increased entropy profile with decreasing redshift, except in the core. The combined sample shows an increase in entropy at all radii at \(z=0\) compared to \(z=1\) and is produced by the decreased density and increased temperature with decreasing redshift. # Summary & Discussion {#sec:sad} In this work we have presented the MACSIS clusters, a sample of 390 zoomed simulations of the most massive and rarest clusters run with the state-of-the-art, calibrated baryonic physics model from the BAHAMAS project that yields realistic clusters. Such massive clusters are absent from the BAHAMAS simulation volumes of \(596\,\mathrm{Mpc}\) as the simulated volume is too small. After introducing the selection of the sample from the parent \(3.2\,\mathrm{Gpc}\) volume simulated with the *Planck* 2013 cosmology, and demonstrating the agreement of the properties of our massive cluster sample with the properties of observed massive clusters, we examined the evolution of the cluster scaling relations and the evolution of the cluster gas profiles. By combining the MACSIS sample with the clusters in the BAHAMAS volume, we were able to examine the cluster scaling relations over the full observed mass range for the first time. Additionally, the MACSIS clusters enabled the study of the evolution of the cluster scaling relations to unprecedentedly high redshifts. Finally, the MACSIS sample enabled clusters to be selected in ways which mimic a cosmological study, such as selecting the hottest clusters, to examine if the scaling relations of such objects evolve differently from the underlying cluster population. Our main results are: - As shown in Fig. [\[fig:observations\]](#fig:observations){reference-type="ref" reference="fig:observations"}, the MACSIS simulations yield realistic massive clusters at low redshift and their progenitors are in good agreement with the limited observational data that is available at high redshift (i.e. \(z=1\)). - Scaling relations for the combined sample that spans the full observed cluster mass range show significant deviations from the simple self-similar theory (see Figs. [\[fig:MgMsr\]](#fig:MgMsr){reference-type="ref" reference="fig:MgMsr"}-[\[fig:LxTxsr\]](#fig:LxTxsr){reference-type="ref" reference="fig:LxTxsr"}). Both the slope of the relations and the redshift evolution of the normalization are significantly affected by non-gravitational physics. The low redshift relations are in good agreement with observations and with most previous simulation work. - The main drivers of non-self-similar evolution are AGN feedback, non-thermal pressure support and a mild mass dependence of the spectroscopic temperature bias. Shallower potentials of clusters that are less massive or form at lower redshifts allows feedback from AGN to eject more gas. Non-thermal pressure lowers a cluster's temperature for a given potential and is more important in more massive clusters that have had less time to thermalise. We found that the spectroscopic temperature bias increases for the most massive clusters. - With the exception of the luminosity-temperature relation, we found the scatter about the best-fit scaling relations is insensitive to mass and redshift for all of the cluster samples. - Selecting a hot cluster sample, i.e. core-excised spectroscopic temperatures \(k_{\rm{B}}T^{\mathrm{X,ce}}_{500\mathrm{spec}}\geq5\,\rm{keV}\), significantly alters the scaling relations and their evolution. Excluding the spectroscopic temperature-total mass relation, we find that the scaling relations of the hot cluster sample evolve in a much more self-similar manner. After accounting for the expected self-similar evolution with redshift, we find that the normalizations are consistent with no evolution. The slopes of the best-fit relations at each redshift are also broadly consistent with the slopes predicted by self-similar theory. However, the spectroscopic temperature-total mass relation of the hot sample deviates further from self-similarity than the combined sample. Selecting hot clusters removes the less massive clusters from the sample, so the hot sample is dynamically younger than the combined sample as more massive clusters form later in the hierarchical merger scenario. This increases the average level of non-thermal support in the hot sample, leading to a flatter spectroscopic temperature-total mass relation. Additionally, the spectroscopic temperature bias flattens the relation for the most massive clusters and this has a larger impact in a sample of only hot clusters. - Selecting a relaxed subset of hot clusters, where the most dynamically disturbed objects are removed, leads to a small reduction in the scatter for most scaling relations. Removing the most disturbed objects also leads to a reduction in the level of non-thermal support in the sample compared to the complete hot sample. This leads to steeper slope of the spectroscopic temperature-total mass relation compared to the hot sample and a value that is closer to the self-similar prediction. - The median hot gas profiles of the combined sample in general shows good agreement with observed radial profiles. The low redshift data is in very good agreement, while the data at \(z=1\) shows reasonable agreement with the relaxed hot sample. - Comparison of the hot gas profiles at \(z=0\) and \(z=1\) show evolution different from self-similar prediction (see Figs. [\[fig:gas_prof\]](#fig:gas_prof){reference-type="ref" reference="fig:gas_prof"}-[\[fig:enty_prof\]](#fig:enty_prof){reference-type="ref" reference="fig:enty_prof"}). The combined sample shows a decreasing density profile with decreasing redshift, suggesting the impact of AGN feedback. Selecting a sample of hot clusters produces a median density profile that evolves in much more self-similar manner. The combined and hot samples have a median temperature profile that increases with decreasing redshift. This is likely driven by decreasing importance of non-thermal pressure support with decreasing redshift. Selecting relaxed hot cluster sample produces a median profile that evolves in better agreement with the self-similar prediction. MACSIS enables the study of the observable properties of the most massive and rarest galaxy clusters. We have demonstrated that their progenitors provide a good match to the currently limited observational data at high redshift and that their observable properties evolve in a significantly more self-similar manner than for lower-mass and less-relaxed clusters. We have shown how the selection function can impact the derived scaling relations and radial profiles. The size of the parent simulation enables the creation of synthetic lightcones with an area comparable to currently ongoing surveys. This will allow the impact of selection biases to be fully examined and the covariance of observable properties to be studied. Another route for future work is to improve our understanding of structure in the ICM, as the limited resolution and traditional SPH scheme used in this work limits our ability resolve structures and understand their impact on observable properties.
{'timestamp': '2016-11-16T02:03:59', 'yymm': '1607', 'arxiv_id': '1607.04569', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04569'}
null
null
# Introduction Due to their ease of fabrication and integration as well as their large area and high quality factor of resonances, photonic crystal slabs with one or two dimensional periodicity have been widely used in many applications, such as filters, lasers, and sensors. For more efficient utilization of light, it is often desirable to achieve highly directional out-of-plane coupling of light from photonic crystal slabs, in which light predominantly radiates to only one side of the slab. This would eliminate the need of a back-reflection mirror in high-power photonic crystal surface emitting lasers (PCSELs), where fabrication uncertainties in the laser wavelength and mirror-cavity distance currently make reliably achieving high slope efficiency difficult. This could also lead to increased efficiency of grating couplers for silicon photonics and light detection and ranging (LIDAR) devices. Previous designs of grating couplers have achieved a top-down asymmetry ratio (defined as the ratio of power going to the top and to the bottom) of up to 50:1, but they typically make use of a substrate reflector or involve multiple layers and grooves, which complicate fabrication and could be difficult to scale to larger areas if desired. Asymmetric out-of-plane emission from photonic crystal defect cavities of 4:1 has also been demonstrated, but all these works were guided primarily by numerical optimization. It is thus important to gain an understanding of the fundamental bounds on asymmetric radiation, and use such bounds as a guide to design stronger asymmetries. Closely related to highly directional radiation is achieving perfect absorption of fields incident from a single side of a weakly-absorbing photonic structure. This can be viewed as the time-reversal partner of the single-sided radiation emission process. An increased absorption efficiency could be important for improving the performance of many devices, including modulators, photodetectors, solar cells. However, the single-pass absorption of a thin absorbing layer in air is at most 50\(\%\). By combining electric and magnetic responses [or utilizing material anisotropy]{style="color: black"}, it is possible to design metamaterial perfect absorbers with near unity absorptance, but such designs can be difficult to implement at optical frequencies. Recent work achieving perfect absorption in photonic crystal structures has either employed illumination from both sides and used the interference between the beams---analogous to a time-reversed laser---to achieve coherent perfect absorption, or employed a back-reflection mirror and critical coupling to resonances to approach perfect absorption. Alternatively, specific surface textures can be designed to enhance light trapping and subsequent absorption, but a back-reflection mirror is still required to keep the photons inside the absorbing layer. In general, however, two-sided illumination can often be challenging to implement in realistic systems, while backing mirrors are often either lossy (e.g. metallic mirrors) or require additional fabrication efforts (e.g. distributed Bragg reflectors). Therefore, the possibility of achieving perfect absorption of fields incident from a single side, without the aid of backing mirrors, is highly attractive and could open up many engineering possibilities. [One recent approach achieving this is to utilize accidental degeneracies of critically coupled modes with opposite symmetries, but such an approach requires aligning the frequencies and quality factors of multiple resonances.]{style="color: black"} Here, we first design structures with highly directional radiation, and then consider the time-reversal scenario at critical coupling to realize devices with high absorption efficiency. Moreover, away from the strongly-absorbing resonance frequency, light can be mostly transmitted through the designed devices, which could have important applications in multi-junction solar cells. Previous work has used a temporal coupled-mode theory (TCMT) formalism with a single resonance and two coupling ports (one on each side of the slab) to examine bounds of asymmetric radiation from photonic crystal slabs. There, they reached the conclusion that the asymmetry ratio is bounded by \((1+r)/(1-r)\), where \(r\) is the background amplitude reflection coefficient. For index contrasts found in realistic materials and at optical frequencies, this bound limits the proportion of radiation going to one side of the photonic crystal slab to around \(90\%\), even for the high index contrast between silicon and air. For a smaller index contrast, this will be even more significantly different from perfect directional radiation. We find, however, that in more general scenarios, the bounds in Ref. can be greatly surpassed. In a periodic photonic structure, the natural choice of mode basis is the momentum-conserving Bloch-wave basis. As shown in Fig. [\[fig:tcmt\]](#fig:tcmt){reference-type="ref" reference="fig:tcmt"}(a), for general incident directions (nonzero in-plane momentum \(\vec{k}_\parallel\)) in asymmetric structures, the time-reversal operation relates the resonance at \(\vec{k}_\parallel\) to the resonance at \(-\vec{k}_\parallel\), suggesting that a two-resonance, four-port model is required to impose time-reversal constraints in the more general case. [Moreover, reciprocity automatically ensures that the two resonances share identical frequencies, eliminating the need for exquisite degenerate-frequency alignment to achieve multi-resonant responses.]{style="color: black"} Only when the system under consideration possesses certain symmetries---either in the structure (\(C_2^z\), i.e. \(180^\circ\) rotation around the out-of-plane axis) or in the incident field (normal incidence)---can we use the simplified model with only a single independent port on each side of the slab. In this article we show that the general two-resonance, four-port TCMT formalism, widely applicable to periodic structures with arbitrary geometry or in-plane momentum, enables bounds with significantly higher (sometimes even infinitely high) radiation asymmetry for realistic materials when \(C_2^z\) symmetry of the structure is broken. As an example, we apply this formalism to inversion-symmetric (\(P\)-symmetric) structures without \(C_2^z\) symmetry. Through numerical examples, we show that a top-down asymmetry ratio exceeding \(10^4\) can be achieved by tuning the resonance frequency to coincide with the perfectly transmitting frequency on the Fabry-Perot background. The emission direction can also be rapidly switched from top to bottom by tuning the \(\vec{k}_\parallel\) vector or frequency. These results provide important design principles for PCSELs, grating couplers, LIDARs, and many other applications that could benefit from directional emission and rapid tuning. In addition, we derive analytical expressions for the transmission spectrum and discuss features such as full transmission or reflection. We then show that such highly asymmetric coupling to the two sides of the photonic crystal slab can also be employed to achieve perfect absorption of light incident from one side of the slab, without the need of back-reflection mirrors as in previous designs. # Temporal Coupled-Mode Theory Formalism We start by considering arbitrary photonic crystal slab structures embedded in a uniform medium (identical substrate and superstrate). We assume weak coupling, linearity, energy conservation and time-reversal symmetry in the system, and we consider frequencies below the diffraction limit so that higher-order diffractions are not present. A plane wave with in-plane momentum \(\vec{k}_\parallel=(k_x,k_y)\) incident from the top \[port 1, see Fig. [\[fig:tcmt\]](#fig:tcmt){reference-type="ref" reference="fig:tcmt"}(a)\] will only couple to resonances and outgoing waves with the same \(\vec{k}_\parallel\) (conservation of Bloch momentum). We shall consider the typical case where there is a single resonance at \(\vec{k}_\parallel\) near the frequencies of interest, with the transmission spectrum consisting of a Fabry-Perot background and sharp resonant features, as in Fig. [\[fig:tcmt\]](#fig:tcmt){reference-type="ref" reference="fig:tcmt"}(b). To describe time-reversal symmetry constraints for general geometries and incident angles, we need to include the resonance at \(-\vec{k}_\parallel\) in our description as well, resulting in a two-resonance, four-port model. Although conservation of Bloch momentum means that each input port only excites either the \(\vec{k}_\parallel\) or \(-\vec{k}_\parallel\) resonance, the two resonances still influence one another indirectly via the time-reversal symmetry constraints on the coupling matrices between resonances and ports (\(K\) and \(D\) in Eq. ([\[eq:TCMT1\]](#eq:TCMT1){reference-type="ref" reference="eq:TCMT1"})), as described below. Writing down expressions consistent with momentum conservation and time-reversal symmetry, we obtain the TCMT equations \[\begin{aligned} \frac{d\textbf{A}}{dt}&=\left(j\omega-\frac{1}{\tau}-\frac{1}{\tau_{nr}}\right)\textbf{A}+K^{\rm T}|s_+\rangle,\; |s_-\rangle=C|s_+\rangle+D\textbf{A},\label{eq:TCMT1}\\ C&=e^{j\phi}\begin{pmatrix} 0 & r & 0 & jt\\r & 0 & jt & 0\\0 & jt & 0 & r\\jt & 0 & r & 0\\ \end{pmatrix},\; D=K\sigma_x=\begin{pmatrix} 0 & d_1\\d_2 & 0\\0 & d_3\\d_4 & 0\\ \end{pmatrix},\label{eq:TCMTcoeff} \end{aligned}\] where \(\textbf{A}=(A_1,A_2)^{\rm T}\) are the amplitudes of the two resonances (at \(\vec{k}_\parallel\) and \(-\vec{k}_\parallel\) respectively), \(\omega\) is the resonance frequency shared by both resonances, \(\tau\) is the radiative \(e^{-1}\)-decay lifetime (\(\omega\), \(\tau\) are identical for the two resonances due to reciprocity), \(\tau_{nr}\) is the \(e^{-1}\)-decay lifetime for nonradiative processes such as absorption, and \(|s_{+}\rangle=(s_{1+},s_{2+},s_{3+},s_{4+})^{\rm T}\), \(|s_{-}\rangle=(s_{1-},s_{2-},s_{3-},s_{4-})^{\rm T}\) are the amplitudes of the incoming and outgoing waves. \(C\) is the scattering matrix for the direct (non-resonant) transmission and reflection through the slab (namely, the Fabry-Perot background). Energy conservation and reciprocity constrain \(C\) to be unitary and symmetric. For identical substrates and superstrates, \(C\) takes the form in Eq. ([\[eq:TCMTcoeff\]](#eq:TCMTcoeff){reference-type="ref" reference="eq:TCMTcoeff"}), where \(t\) and \(r\) are real numbers satisfying \(r^2+t^2=1\) that characterize the Fabry-Perot background, and the phase \(\phi\) depends on the choice of reference plane position. \(K\) and \(D\) are the coupling matrices in and out of the resonances. Time reversal flips the two resonances, so instead of the usual relation \(D=K\), here we have \(D=K\sigma_x\), where \(\sigma_x\) is the \(2\times 2\) \(X\)-Pauli matrix acting to flip the resonances. We note that an alternative (and equivalent) formalism is to adopt a basis in which the underlying modes are time-reversal invariant (for which the standard multimode treatment is adequate), by superimposing the resonance at \(\vec{k}_\parallel\) with its time-reversal partner at \(-\vec{k}_\parallel\). This is to be contrasted with a basis change for ports, as discussed in Ref.. Detailed derivations of this and the following expressions are given in [Supplement 1](link). Energy conservation and time-reversal symmetry impose constraints on the coefficients. Energy conservation requires \[\begin{aligned} \label{eq:econs} |d_2|^2+|d_4|^2=|d_1|^2+|d_3|^2=\frac{2}{\tau}, \end{aligned}\] while time-reversal symmetry gives the constraint \(D=K\sigma_x\) and the two independent equations \[\begin{aligned} \label{eq:fulld1} e^{j\phi}(rd_2^*+jtd_4^*)+d_1&=0,\\ e^{j\phi}(jtd_2^*+rd_4^*)+d_3&=0.\label{eq:fulld2} \end{aligned}\] In the following, we shall fix the phase \(\phi\) to be 0 by appropriately choosing the location of our reference plane. Eqs. ([\[eq:econs\]](#eq:econs){reference-type="ref" reference="eq:econs"}, [\[eq:fulld1\]](#eq:fulld1){reference-type="ref" reference="eq:fulld1"}, [\[eq:fulld2\]](#eq:fulld2){reference-type="ref" reference="eq:fulld2"}) impose constraints on the values and phases of the couplings, and hence constrain the transmission spectrum and set bounds on the asymmetric coupling ratios. From the preceding equations, we can derive an expression for the transmission spectrum that only depends on the frequencies and decay rates of the resonances and the transmission and reflection coefficients of the direct Fabry-Perot pathway. The full scattering matrix including the direct pathway and resonance pathway is given by Eq. ([\[eq:scatmat\]](#eq:scatmat){reference-type="ref" reference="eq:scatmat"}) in [Supplement 1](Link). The power reflection and transmission coefficient for a wave incident from port 1 correspond to the amplitude squared of the (1,2), (1,4) element of the scattering matrix, given by \[\begin{aligned} \label{eq:R} R&=|S_{12}|^2=\left|e^{j\phi}r+\frac{d_1d_2}{j(\omega-\omega_0)+\frac{1}{\tau}+\frac{1}{\tau_{nr}}}\right|^2,\\\label{eq:T} T&=|S_{14}|^2=\left|e^{j\phi}jt+\frac{d_1d_4}{j(\omega-\omega_0)+\frac{1}{\tau}+\frac{1}{\tau_{nr}}}\right|^2, \end{aligned}\] and the power reflection coefficient can be rewritten in the lossless limit \(\tau_{nr}\rightarrow\infty\) as \[\begin{aligned} \label{eq:nolossR} R&=\frac{\left[r(\omega-\omega_0)\pm\sqrt{\frac{4}{\tau_1\tau_2}-\frac{r^2}{\tau^2}-\frac{2}{\tau\sigma}-\frac{1}{\sigma^2 r^2}}\right]^2+\left(\frac{1}{\sigma r}\right)^2}{(\omega-\omega_0)^2+\frac{1}{\tau^2}}, \end{aligned}\] where we have written \(\tau_i=2/|d_i|^2\), \(1/\sigma=1/\tau_1-1/\tau_4\) to simplify the expression. This expression provides general conditions for reaching full transmission or reflection with the Fano resonance. As shown in [Supplement 1](link), full transmission \(R=0\) can only occur when the coupling rates satisfy \(\tau_1=\tau_4\), \(\tau_2=\tau_3\) (\(P\)-symmetric coupling), consistent with the transmission spectrum shown in Fig. [\[fig:tcmt\]](#fig:tcmt){reference-type="ref" reference="fig:tcmt"}(b). Full reflection \(R=1\) can only occur when the coupling rates satisfy \(\tau_1=\tau_2\), \(\tau_3=\tau_4\) (\(C_2^z\) symmetric coupling), consistent with the results in Ref. . Note that for structures that do not have \(P\) or \(C_2^z\) symmetry, it is still possible for the coupling rates for resonances to be \(P\) or \(C_2^z\) symmetric, leading to full transmission/reflection features in the frequency spectrum (see for example [Supplement 1](link), Fig. [\[fig:multispec\]](#fig:multispec){reference-type="ref" reference="fig:multispec"}). # General Bounds on Asymmetric Coupling Rates We now derive bounds on the achievable asymmetry of coupling to the top and bottom based on Eqs. ([\[eq:fulld1\]](#eq:fulld1){reference-type="ref" reference="eq:fulld1"}, [\[eq:fulld2\]](#eq:fulld2){reference-type="ref" reference="eq:fulld2"}) derived from time-reversal symmetry. Denote \(|d_4/d_2|=a_r\), \(|d_3/d_1|=a_\ell\), and define the asymmetric coupling ratios on the right (\(\vec{k}_\parallel\)) and left (\(-\vec{k}_\parallel\)) directions of the resonator as \(a_r^2\) and \(a_\ell^2\) (the ratio of the power going to bottom and top). By taking the ratio of Eqs. ([\[eq:fulld1\]](#eq:fulld1){reference-type="ref" reference="eq:fulld1"}, [\[eq:fulld2\]](#eq:fulld2){reference-type="ref" reference="eq:fulld2"}), we find \[\begin{aligned} a_\ell^2=\left|\frac{jt+ra_re^{j\theta}}{r+jta_re^{j\theta}}\right|^2=\frac{t^2+r^2a_r^2+2tra_r\sin\theta}{r^2+t^2a_r^2-2tra_r\sin\theta}, \end{aligned}\] where \(\theta=\arg(d_2)-\arg(d_4)\) characterizes the phase difference between \(d_2\) and \(d_4\). This gives the bound \[\begin{aligned} \label{eq:fullbound} \left|\frac{t-ra_r}{r+ta_r}\right|\leq a_\ell\leq \left|\frac{t+ra_r}{r-ta_r}\right|. \end{aligned}\] Therefore, the amount of achievable asymmetry to top and bottom on the left is bounded by that on the right, and vice versa. Note that in general the phase here can be tuned through a \(2\pi\) cycle, so the bounds---even up to infinitely high asymmetry ratio---should be saturable for appropriate parameter choices. [The coefficients that enter the bounds are the transmission/reflection coefficients (\(t,r\)) of the direct process (Fabry-Perot background), as opposed to the total transmission/reflection including the resonant pathway.]{style="color: black"} If the structure has \(C_2^z\) symmetry, the two channels on the top and bottom will be constrained to have the same coupling rates, so \(d_1=d_2\), \(d_3=d_4\), \(a_\ell=a_r\). Plugging this into Eq. ([\[eq:fullbound\]](#eq:fullbound){reference-type="ref" reference="eq:fullbound"}), we find the same bound as in Ref. : \(\frac{1-r}{1+r}\leq a_\ell^2=a_r^2\leq\frac{1+r}{1-r},\) which shows the consistency of our approach. In typical photonic crystal systems at optical frequencies, the index contrast between the slab and the background medium is limited to around 3, which constrains the interface reflection coefficient to be less than 0.5 for most incident angles. This results in the Fabry-Perot direct pathway reflection coefficient \(r=\sqrt{1-t^2}\) being considerably smaller than 1, so strong asymmetry in the decay rates is difficult to achieve for \(C_2^z\)-symmetric structures without the use of an additional back-reflecting mirror. The general bound Eq. ([\[eq:fullbound\]](#eq:fullbound){reference-type="ref" reference="eq:fullbound"}) suggests, however, that much stronger asymmetry can be achieved if we break the \(C_2^z\) symmetry of the system. A simple example is when the structure possesses inversion symmetry \(P\) but breaks \(C_2^z\) symmetry, as shown in Fig. [\[fig:bound\]](#fig:bound){reference-type="ref" reference="fig:bound"}(a). In this case, the decay rates must satisfy \(d_1=d_4\), \(d_2=d_3\), \(a_\ell=1/a_r\), and Eq. ([\[eq:fullbound\]](#eq:fullbound){reference-type="ref" reference="eq:fullbound"}) becomes \[\begin{aligned} \label{eq:tbound} \frac{1-t}{1+t}\leq a_\ell^2=\frac{1}{a_r^2}\leq \frac{1+t}{1-t}. \end{aligned}\] For any index contrast, due to the up-down symmetry of the background material, the Fabry-Perot background will always have frequencies with full transmission, as exemplified by the red circles in Fig. [\[fig:tcmt\]](#fig:tcmt){reference-type="ref" reference="fig:tcmt"}(b). Therefore, by tuning the resonance frequency to such points, the lower and upper bounds of Eq. ([\[eq:tbound\]](#eq:tbound){reference-type="ref" reference="eq:tbound"}) approach \(0\) and \(+\infty\). Moreover, the bound can be saturated for appropriate choices of structural parameters and wavevectors, yielding arbitrarily high asymmetric decay rates of the photonic structure to top and bottom directions. [We note that similar design principles of breaking \(C_2^z\) symmetry to achieve higher radiation asymmetry have also been realized in Ref. using the different design intuition of destructive interference.]{style="color: black"} To verify these analytical results, we perform numerical simulations using the finite difference time domain (FDTD) method with a freely available software package. We extract the coupling rate to top or bottom by monitoring the field amplitude at reference planes placed in the far-field, and determine the Fabry-Perot background transmissivity from plane wave excitation calculations. The results are given in Fig. [\[fig:bound\]](#fig:bound){reference-type="ref" reference="fig:bound"}(b), where each data point in the figure (blue crosses) represents the maximal asymmetric coupling ratio searching over all \(\vec{k}\) points in the Brillouin zone, for \(P\)-symmetric structural parameter choices (Fig. [\[fig:bound\]](#fig:bound){reference-type="ref" reference="fig:bound"}(a)) with \(h=1.5a\), \(w=0.45a\), \(n_0=1.45\), and varying \(n_d\) and \(d\) (see [Supplement 1](Link) for more details). We can see that all data points obey the bound Eq. ([\[eq:tbound\]](#eq:tbound){reference-type="ref" reference="eq:tbound"}) derived above (red solid lines). Moreover, this bound can be saturated for each value of the background transmission coefficient by appropriate optimization of the structural parameters and in-plane momentum. The blue crosses that do not saturate the bound are structures with very little perturbation from \(C_2^z\) symmetry due to the choice of structural parameters. # Examples of Highly Asymmetric Radiation In this section, we provide numerical examples of strong asymmetry that highlight two features of the extreme data points in Fig. [\[fig:bound\]](#fig:bound){reference-type="ref" reference="fig:bound"} that are not obvious from the preceding data: it is possible to achieve high asymmetry even at the point of highest quality factor, and it is possible to achieve rapid tuning of the direction of asymmetry by slightly changing the frequency. As the form \((1+t)/(1-t)\) of the bound (for \(P\)-symmetric structures) suggests, strong asymmetry can be achieved when the resonance frequency coincides with locations of large transmissivity on the Fabry-Perot background, for any refraction index contrast. We optimize over the structural parameters shown in Fig. [\[fig:bound\]](#fig:bound){reference-type="ref" reference="fig:bound"}(a) to find examples of high asymmetry in coupling to the top and bottom. This example consists of the second transverse electric (TE) polarization band (nonzero \(E_z, E_x, H_y\), classified by mirror-symmetry with respect to the \(x-z\) plane) of a 1D photonic crystal with structural parameters \(h=1.5a\), \(w=0.45a\), \(n_0=n_d=1.45\), \(d=0.3a\), as defined in Fig. [\[fig:bound\]](#fig:bound){reference-type="ref" reference="fig:bound"}(a). See [Supplement 1](link) for a plot of the band structure. The resulting asymmetry ratio and quality factor (\(Q\)) as a function of the in-plane \(k_x\), along with the radiation field distribution at maximal asymmetry, are shown in Fig. [\[fig:maxasym\]](#fig:maxasym){reference-type="ref" reference="fig:maxasym"}(a). The resonance frequency lies very close to a point of full transmission on the Fabry-Perot background, and exhibits an asymmetry exceeding \(10^4\) at the \(\vec{k}_\parallel\) point of largest asymmetry as well as an asymmetry over \(300\) at the point of highest quality factor. It may therefore be possible to produce a laser that preferentially emits to the top or bottom using the principles discussed above. Another application of our results is the rapid steering of the direction of light emission by slight tuning of the frequency, which could be useful for LIDARs or antennas. We design such a structure by perturbing a bound state in the continuum (BIC). BICs are localized solutions embedded in the radiation continuum, where due to destructive interference of the amplitude for decay between outgoing wave channels, the quality factor of a resonance above the light line approaches infinity. In previous work, the photonic structures were chosen to have both \(P\) and \(C_2^z\) symmetries. With a perturbation that breaks \(C_2^z\) but preserves \(P\), the peak quality factor will be finite but still very high. We expect that this symmetry breaking will also split the momenta where radiation towards the top and towards the bottom vanish, thereby creating strong asymmetry in the two directions, with the extrema separated only by a small \(\vec{k}_\parallel\). We choose \(h=1.5a\), \(w=0.45a\), \(n_0=1.45\), \(d=0.1a\), \(n_d=1.1\), again examining the second TE band. The resulting asymmetry ratio and frequency are shown in Fig. [\[fig:maxasym\]](#fig:maxasym){reference-type="ref" reference="fig:maxasym"}(b). The asymmetric coupling flips from mostly radiating to the top to mostly radiating to the bottom (by a factor of \(10^4\)) when \(k_x\) is changed by as small as \(0.05\times 2\pi c/a\) or equivalently, when the frequency is changed by \(3\times 10^{-4}\times 2\pi c/a\). The radiative quality factor of these resonances are on the order of \(10^6\), so these two bands will be well-separated in emission. One can thereby envision rapid tuning of the emission direction by changing the frequency of radiation slightly. Moreover, the high \(Q\) of these resonances will enable long propagation lengths for collimated emission from large areas that is important for LIDAR applications, complementing the low-\(Q\) designs of conventional grating couplers. # Perfect Absorption with Single-Sided Illumination and No Backing Mirrors We now discuss achieving perfect absorption in photonic crystal structures by combining the highly asymmetric coupling to different channels and matched radiative and non-radiative quality factors. Previous work on achieving perfect absorption has utilized metamaterial responses, interference between multiple incident directions, or a backing mirror to confine and trap light. On the other hand, our results on achieving highly directional coupling suggest that it may be possible to achieve near-perfect absorption with single-sided illumination from the direction with strong coupling, only using dielectric structures and without the need of any backing mirrors. Intuitively, since the radiation coupling to one of the emission channels is strongly suppressed, there is only one direction to which the excited resonance can radiatively decay into. Appropriate tuning of the quality factor can then result in destructive interference towards this direction, thus achieving perfect absorption of incident waves. To incorporate material loss in our description, we include a non-radiative decay channel, setting \(\tau_{nr}\) to be finite in Eq. ([\[eq:TCMT1\]](#eq:TCMT1){reference-type="ref" reference="eq:TCMT1"}). We assume that the loss rate is small, and that the direct transmission pathway is not affected by the loss. We start from a \(P\)-symmetric structure and incident direction where the coupling rates have a large asymmetry ratio. The parameters are chosen such that \(t\) is close to 1, and the asymmetry ratio saturates the bound \(|d_2|/|d_1|=\sqrt{(1+t)/(1-t)}\). As shown in [Supplement 1](Link), \(P\)-symmetry and the time-reversal constraints Eqs. ([\[eq:fulld1\]](#eq:fulld1){reference-type="ref" reference="eq:fulld1"}, [\[eq:fulld2\]](#eq:fulld2){reference-type="ref" reference="eq:fulld2"}) imply that \(d_1=d_4=\sqrt{(1-t)/\tau}\exp(-3\pi j/4)\), \(d_2=d_3=\sqrt{(1+t)/\tau}\exp(-\pi j/4)\). Plugging this into the expressions for the power transmission and reflection coefficients, we find that on resonance \(\omega=\omega_0\) and at critical coupling \(\tau=\tau_{nr}\), \[\begin{aligned} R_{14}&=R_{23}=\frac{r^2}{4},\quad T_{14}=\left(\frac{1+t}{2}\right)^2,\quad T_{23}=\left(\frac{1-t}{2}\right)^2,\\ A_{14}&=1-R_{14}-T_{14}=\frac{1-t}{2},\quad A_{23}=\frac{1+t}{2}. \end{aligned}\] In the limit where \(t\rightarrow 1\), \(r\rightarrow 0\), it follows that light incident from port 1 or 4 will be completely transmitted (\(T_{14}=1\)), while light incident from port 2 or 3 will be completely absorbed (\(A_{23}=1-R_{23}-T_{23}=1\)). A schematic of the resulting transmission and absorption characteristics is shown in Fig. [\[fig:absorb\]](#fig:absorb){reference-type="ref" reference="fig:absorb"}(a). To verify these analytical results, we performed numerical simulations of the transmission and reflection spectrum with the rigorous coupled-wave analysis (RCWA) method using a freely available software package. The structural parameters are identical to the simulation in Fig. [\[fig:maxasym\]](#fig:maxasym){reference-type="ref" reference="fig:maxasym"}(a), with the difference being the addition of loss in the system. There is a slight shift of the resonance location relative to the Fabry-Perot background due to the different discretization schemes used in the FDTD and RCWA methods. Fig. [\[fig:absorb\]](#fig:absorb){reference-type="ref" reference="fig:absorb"} shows the simulation results. As shown in Fig. [\[fig:absorb\]](#fig:absorb){reference-type="ref" reference="fig:absorb"}(b), when no loss is present (top panel), the transmissivity is close to 1 in the vicinity of the resonance, reaching full transmission \(T=1\) at a single point on the Fano resonance as required by \(P\)-symmetry of the structure (see end of Section 2 for a discussion), and the reflectance is close to 0. With the addition of loss and with waves incident from the port with stronger coupling (incident direction is \(\theta=3.5^\circ\) from normal, middle panel), the transmittance is reduced, and at critical coupling the transmittance drops to 0 for the resonance frequency, resulting in full-absorption of the incoming waves. On the other hand, for the same lossy structure and for the opposite incident direction (bottom panel), there is negligible absorption and most of the waves are transmitted. In Fig. [\[fig:absorb\]](#fig:absorb){reference-type="ref" reference="fig:absorb"}(c), we show the maximum absorption for a given incident angle as a function of loss, clearly showing a peak of near-perfect absorption at critical coupling. [For other incident angles near that of maximum absorption, the \(Q\)-matching condition and asymmetric radiation condition are still approximately satisfied, giving rise to high absorption, and the absorption peak will shift to different frequencies following the band dispersion, as shown in [Supplement 1](link), Fig. [\[fig:varyangle\]](#fig:varyangle){reference-type="ref" reference="fig:varyangle"}.]{style="color: black"} Numerically, we find that when light is incident from the port with stronger coupling, the absorption can be as high as \(A_{23}=99.8\%\); when light is incident from the port with weaker coupling, the transmission is \(T_{14}=99.4\%\) while absorption is only \(A_{14}=0.5\%\) (this is roughly equal to the single-pass absorption rate of \(0.4\%\) in our simulations). The numerical simulations show excellent agreement with our TCMT predictions for a background Fabry-Perot transmissivity of \(t=0.996\) (Fig. [\[fig:absorb\]](#fig:absorb){reference-type="ref" reference="fig:absorb"}(b)); the small difference comes from the contribution of absorption to the Fabry-Perot background. Further numerical optimization placing the resonance frequency closer to the frequency of full transmission on the Fabry-Perot background could further increase on-resonance absorption in the desired port. These results are widely applicable to many different absorbing materials. The wide range of achievable resonance quality factors, as discussed in the preceding section, implies that for both strong and weak absorbers, structures can be designed such that there is highly asymmetric coupling and critical coupling, yielding perfect absorption in the system. While our simulations were performed assuming a material with a spatially uniform absorption profile, the generality of the TCMT formalism ensures that it is also applicable to scenarios with only a thin active layer with absorption, such as with 2D materials. # Discussion and Conclusion In conclusion, we developed a temporal coupled-mode theory formalism for general dielectric photonic crystal slab structures with arbitrary in-plane wavevectors, adequately taking into account the time-reversal-symmetry related pair of resonances and coupling channels. Using this formalism, we derived general bounds on the asymmetric radiation rates to the top and bottom of a photonic crystal slab. We then used the intuitions developed from these bounds to show examples of highly asymmetric radiation from inversion-symmetric photonic crystal slabs, demonstrating strong asymmetry, rapid tuning, and a variety of quality factors for different applications. Moreover, we showed how the highly asymmetric coupling to the top and bottom of photonic crystal slabs can be used to achieve perfect absorption for light incident from a single side, for a single-pass absorption rate of less than \(0.5\%\), without the need of a back-reflection mirror as in conventional setups. The highly directional radiation could greatly benefit applications such as PCSELs, grating couplers, and LIDARs, while achieving perfect absorption without the need of back-reflection mirrors could increase the efficiency and simplify the design of photodetectors and solar cells. [While our numerical examples focused on a particular structural design, the general principle of breaking \(C_2^z\) symmetry is applicable to a wide range of structures. We now briefly discuss how to implement such structures using readily available fabrication techniques. For example, gratings with slanted walls share the same structural symmetries as those in Fig. [\[fig:bound\]](#fig:bound){reference-type="ref" reference="fig:bound"}(a), and thus can approach perfect single-sided radiation and absorption as well. Such gratings can be fabricated using focused ion beam milling, angled-etching with Faraday cages, or inclined lithography. More generally, almost any of the techniques for fabricating blazed gratings or constructing 3D photonic crystals (e.g. layer-by-layer lithography or holographic lithography) could also be employed in a simplified form to make an asymmetric coating. A wide range of structures breaking \(C_2^z\) symmetry and achieving high asymmetry can thus be easily realized with these different techniques.]{style="color: black"} Our work provides new design principles for achieving highly directional radiation and perfect absorption in photonics, and could be extended to systems where there are nonlinearities, gain and loss, different substrates and superstrates, and non-reciprocal structures with magneto-optical effects. Our work can also be generalized to other systems characterized by temporal coupled-mode theory, such as in-plane chiral meta-surfaces, asymmetric ring resonators, and scattering from nano-plasmonic structures.
{'timestamp': '2016-09-23T02:01:43', 'yymm': '1607', 'arxiv_id': '1607.04774', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04774'}
null
null
# Introduction {#sec:intro} The discovery of a light Higgs boson  has opened a major new avenue in experimental and theoretical particle physics: comprehensive tests of a possible non-minimal fundamental scalar sector, for which there exists a plethora of motivations. While there has been a lot of progress in developing combined Higgs and gauge analysis strategies for the LHC Run II , there exists no general and proven analysis framework even for a Higgs portal model . The announcement of an excess seen in the di-photon spectrum by both ATLAS and CMS , if confirmed by future data, suggests such an extended scalar sector. The anomaly has led to an excessive number of publications, so we feel that adding one more, and hopefully useful publication can be justified somehow[^1]. Early studies of the anomaly in an effective theory framework can be found in Ref. . Intriguingly, an additional scalar is not sufficient to explain the signal in complete models. The new scalar's sizable couplings to photons and gluons need to be induced by relatively light new particles . For example in supersymmetric models, vector-like matter added to the MSSM or non-trivial signatures in the NMSSM are necessary for a successful explanation of the excess . In models in which these new states are connected to the SUSY breaking sector the new scalar can be identified with the sgoldstino, implying a very low SUSY breaking scale . Other extended spacetime symmetries give rise to dilaton  and radion interpretations , which imply similarly unintended consequences, such as low ultraviolet (UV) scales or a very large curvature of the extra dimension. Extra dimensional scalars not directly related to the compactification circumvent this problem  and can explain the localization of extra dimensional fermions, which makes the new scalar a localizer field . Related models, which consider the electroweak scale (or the TeV scale) arising from composite dynamics are less constrained than the MSSM, due to the large number of potential scalar resonances and fermionic quark partners . The possibility of the new resonance to be a spin 2 particle, associated with a higher dimensional theory of gravity is strongly constrained by dilepton searches and just like the radion implies sizable curvature terms . The large width of \(\Gamma_S=45\) GeV, as reported by ATLAS, can be addressed in some models ; while such a large width only slightly increases the statistical significance, if it is true, background interference effects are important . In this case, it is well motivated to assume that the new scalar provides a portal to a dark sector, inducing a sizable width through invisible decays . Alternatively, it might be the sign of cascade decays or other explanations not based on a single scalar resonance, which lead to cusps and endpoint structures that can fake a large width . The very minimal, yet not perturbatively realizable assumption of photon fusion induced production can not explain a large width . The new resonance could also be related to the various, persistent flavor anomalies , to the mechanism behind the electroweak phase transition , the strong CP problem , or an underlying string theory . Finally, a variety of models, motivated by different extensions of the Standard Model (SM) not fitting in the above categories, and further measurements testing the properties of the new resonance have been proposed .In spite of all these considerations, the most obvious question is whether such an additional, likely scalar resonance can be part of an extended Higgs sector ; in other words, if the new scalar can form a Higgs portal, possibly to a new sector. To answer this question we will remain agnostic about the underlying physics, but assume that a resonantly produced narrow scalar singlet is responsible for the excess. We couple the new scalar to the SM through an effective Lagrangian. This assumption exactly corresponds to recent developments on how to describe deviations from the Standard Model Higgs sector at the LHC . The combined Higgs portal Lagrangian is organized by the field content, the symmetry structure, and the mass dimension. This way we can contrast the apparent absence of dimension--six effects in the range \(\Lambda \approx 300~...~500\) GeV for the SM-like Higgs and gauge sector  with the need for higher-dimensional operators coupling to the new scalar with \(\Lambda \lesssim 1\) TeV. ## Theoretical framework {#sec:lag} The most general linear effective Lagrangian up to dimension six and built from Standard Model particles and a new scalar singlet reads \[\begin{aligned} \ensuremath{\mathcal{L}} = \ensuremath{\mathcal{L}}_\text{SM} + \ensuremath{\mathcal{L}}_{\text{dim-6}}^H + \ensuremath{\mathcal{L}}_{\text{dim}\leq 5}^S+\ensuremath{\mathcal{L}}_{\text{dim-6}}^S \;. \label{eq:lag1} \end{aligned}\] Here, \(\ensuremath{\mathcal{L}}_\text{SM}\) stands for the renormalizable SM Lagrangian, while \(\ensuremath{\mathcal{L}}_{\text{dim-6}}^H\) contains the dimension--six operators made out of SM fields. Adopting the basis of our set of Higgs legacy papers  it reads \[\begin{aligned} \ensuremath{\mathcal{L}}_\text{dim-6}^H = & \frac{f_{BB}}{\Lambda^2} \phi^{\dagger} \hat{B}_{\mu \nu} \hat{B}^{\mu \nu} \phi + \frac{f_{WW}}{\Lambda^2} \phi^{\dagger} \hat{W}_{\mu \nu} \hat{W}^{\mu \nu} \phi-\frac{\alpha_s }{8 \pi} \frac{f_{GG}}{\Lambda^2} \phi^\dagger \phi G_{\mu\nu}^a G^{a\mu\nu} + \frac{f_{WWW}}{\Lambda^2} \text{tr} \left( \hat{W}_{\mu \nu} \hat{W}^{\nu \rho} \hat{W}_\rho{}^\mu \right) \notag \\ & + \frac{f_B}{\Lambda^2} (D_{\mu} \phi)^{\dagger} \hat{B}^{\mu \nu} (D_{\nu} \phi) + \frac{f_W}{\Lambda^2} (D_{\mu} \phi)^{\dagger} \hat{W}^{\mu \nu} (D_{\nu} \phi) + \frac{f_{\phi,2}}{\Lambda^2} \frac{1}{2} \partial^\mu ( \phi^\dagger \phi ) \partial_\mu ( \phi^\dagger \phi ) \notag \\ & + \left(\frac{f_\tau m_\tau}{v \Lambda^2} (\phi^\dagger\phi)(\bar L_3 \phi e_{R,3}) + \frac{f_b m_b}{v \Lambda^2} (\phi^\dagger\phi)(\bar Q_3 \phi d_{R,3}) + \frac{f_t m_t}{v \Lambda^2} (\phi^\dagger\phi)(\bar Q_3 \tilde \phi u_{R,3})+\text{h.c.}\right) \; . \label{eq:ourleff} \end{aligned}\] The Higgs covariant derivative is \(D_\mu\phi= \left(\partial_\mu+ i g' B_\mu/2 + i g \sigma_a W^a_\mu/2 \right)\phi\), and the field strengths are \(\hat{B}_{\mu \nu} = i g' B_{\mu \nu}/2\) and \(\hat{W}_{\mu\nu} = i g\sigma^a W^a_{\mu\nu}/2\) in terms of the Pauli matrices \(\sigma^a\). The \(SU(2)_L\) and \(U(1)_Y\) gauge couplings are \(g\) and \(g^\prime\), respectively. While the minimum independent set consists of 59 baryon number conserving operators, barring flavor structure and Hermitian conjugation , we follow the definition of the relevant operator basis describing Higgs coupling and triple gauge boson vertex (TGV) modifications at the LHC in Ref. . In our construction we assume a narrow, CP-even Higgs, focusing on the minimal, Yukawa-like, couplings to the heavy fermions. We use the equations of motion to rotate to a basis where there are no blind directions linked to electroweak precision data. That way, we neglect all operators contributing to electroweak precision observables at tree level in our LHC analysis. For the Standard Model fit  we omit the operator \((\phi^\dagger \phi )^3\), which only contributes to the rather poorly measured triple Higgs coupling. In the appendix we argue why even in the presence of an additional, mixing scalar, this operator will not add any extra relevant features to the fit.Moving to the new scalar Lagrangian terms, we assume in the following that the additional singlet does not develop a VEV, or that the Lagrangian can be re-defined such that the VEV vanishes . The effective Lagrangian of such an additional singlet scalar can be divided into two pieces. Following Refs.  we first write down a set of non-redundant, independent operators up to dimension five, \[\begin{aligned} \ensuremath{\mathcal{L}}_{\text{dim}\leq 5}^S =& \frac{1}{2}\partial_\mu S \, \partial^\mu S-a_1 S-\frac{M_S^2}{2}S^2-a_3 S^3-a_4 S^4-\frac{f_5^S}{\Lambda} S^5\notag\\ &-\mu_S S\phi^\dagger\phi -\frac{\lambda_{SH}}{2}S^2\phi^\dagger\phi -\frac{f_1^S}{\Lambda}S ( \phi^\dagger\phi )^2 -\frac{f_3^S}{\Lambda} S^3 \phi^\dagger\phi \notag\\ &+\frac{\alpha_s}{4\pi}\frac{f_{GG}^S}{\Lambda}SG_{\mu\nu}^aG^{a\;\mu\nu} +\frac{\alpha}{4\pi c_w^2}\frac{f_{BB}^S}{\Lambda}SB_{\mu\nu}B^{\mu\nu} +\frac{\alpha}{4\pi s_w^2}\frac{f_{WW}^S}{\Lambda}SW_{\mu\nu}^aW^{a\;\mu\nu} \notag \\ &\left(-\frac{f_d^S}{\Lambda}S\bar{Q}_L\phi d_R-\frac{f_u^S}{\Lambda}S\bar{Q}_L\tilde\phi u_R -\frac{f_\ell^S}{\Lambda}S\bar{L}_L\phi \ell_R+\text{h.c.}\right)\;\;\;. \label{eq:lagd5} \end{aligned}\] To be fully consistent with the Standard Model Lagrangian we could then add all dimension--six operators including at least one power of the new singlet scalar. The corresponding set of additional operators can be written as  \[\begin{aligned} \ensuremath{\mathcal{L}}_{\text{dim-6}}^S =& \frac{f^{SS}_\phi}{\Lambda^2} \phi^\dagger\phi \partial_\mu S \partial^\mu S-\frac{f_6^S}{\Lambda^2} S^6-\frac{f_4^S}{\Lambda^2} S^4 \phi^\dagger\phi-\frac{f_2^S}{\Lambda^2} S^2 (\phi^\dagger\phi )^2 \notag \\ &+\frac{f_{GG}^{SS}}{\Lambda^2}S^2 G_{\mu\nu}^aG^{a\;\mu\nu} +\frac{f_{BB}^{SS}}{\Lambda^2}S^2B_{\mu\nu}B^{\mu\nu} +\frac{f_{WW}^{SS}}{\Lambda^2}S^2W_{\mu\nu}^aW^{a\;\mu\nu} \notag \\ &\left(-\frac{f_d^{SS}}{\Lambda^2}S^2\bar{Q}_L\phi d_R-\frac{f_u^{SS}}{\Lambda^2}S^2\bar{Q}_L\tilde\phi u_R-\frac{f_\ell^{SS}}{\Lambda^2}S^2\bar{L}_L\phi \ell_R+\text{h.c.}\right)\;\;. \label{eq:lagd6} \end{aligned}\] Nevertheless, given the singlet nature of the new scalar and neglecting lepton number violation, all dimension--six operators including the singlet are quadratic in the field \(S\). Consequently, their phenomenological effects will be contributions to the mass terms (\(f_2^S/\Lambda^2\)), re-definitions of the \(S\) field to recover canonical kinetic terms (\(f^{SS}_\phi/\Lambda^2\)), and the contributions to several vertices including two or more heavy scalars. After scalar-Higgs mixing, the two operators \(f^{SS}_\phi/\Lambda^2\) and \(f^S_2/\Lambda^2\) will contribute to the \(SHH\) interaction as well. However, all these phenomenology features are already taken into account in our analysis by the free parameters in the dimension--five Lagrangian. Therefore, we neglect for the time being the explicit features induced by Eq.[\[eq:lagd6\]](#eq:lagd6){reference-type="eqref" reference="eq:lagd6"}. We give more details on the effective Lagrangian and the Higgs portal mixing in the Appendix. ## Analysis framework {#sec:sfitter} The set of analyses presented here are derived using the [SFitter]{.smallcaps} framework. [SFitter]{.smallcaps} allows us to study multi-dimensional parameter spaces in the Higgs sector , the gauge sector  and in new physics models like supersymmetry . The fit procedure uses Markov chains to create an exclusive, multidimensional log-likelihood map, based on the available measurements and including all the relevant uncertainties and correlations. The construction of a profile likelihood with flat theory uncertainties leads to the [RFit]{.smallcaps} scheme . The statistic uncertainties on the measurements, both for event rates and kinematic distributions, follow Poisson statistics, as do the background uncertainties. All systematic uncertainties are described by Gaussian distributions and can be correlated between the relevant channels. We show log-likelihood projections on two-dimensional planes after profiling over all other parameters. Here, red-yellow regions will illustrate points within \(\Delta (-2\log \mathcal{L})=2.3\) of the best fit point log-likelihood (\(1\sigma\) in the Gaussian approximation), green regions indicate \(\Delta (-2\log {\cal L})=6.18\) (\(2\sigma\) in the Gaussian limit), and black dots imply the \(\Delta (-2\log \mathcal{L})=5.99\) exclusion limits (95% CL in the Gaussian case). The implementation of experimental results in the [SFitter]{.smallcaps} framework is described in Ref.  for the Higgs measurements and in Ref.  for anomalous triple gauge boson coupling measurements. For the triple gauge boson vertex (TGV) analyses[^2] the correlation of the theory uncertainties between the different bins of a given kinematic distribution is taken into account by flat profiled nuisance parameters , while for the different Higgs channels the theory uncertainties are considered uncorrelated without a sizable impact on the shown results . For the Higgs portal analysis we take into account the constraints on a possible new resonance based on the data listed in Tab. [1](#tab:750_data){reference-type="ref" reference="tab:750_data"}. For the new resonance we only implement inclusive measurements assuming a narrow width. # Higgs portal analysis {#sec:ana} In the following we will use the [SFitter]{.smallcaps} effective Lagrangian framework to analyze a new gluon-fusion produced resonance in combination with the electroweak gauge and Higgs sectors at the weak scale. In other words, we ask the question whether such a new particle could be part of an extended Higgs sector and what the allowed parameter space is. In Sec. [2.1](#sec:ana_heavy){reference-type="ref" reference="sec:ana_heavy"} we only include the dimension--five operators given in Eq.[\[eq:lagd5\]](#eq:lagd5){reference-type="eqref" reference="eq:lagd5"}, restricting the analysis to the data in Tab. [1](#tab:750_data){reference-type="ref" reference="tab:750_data"}. In Sec. [2.2](#sec:ana_combined){reference-type="ref" reference="sec:ana_combined"} we combine this analysis with the Higgs-electroweak measurements and the [SFitter]{.smallcaps} results induced by the dimension--six Lagrangian in Eq.[\[eq:ourleff\]](#eq:ourleff){reference-type="eqref" reference="eq:ourleff"}. Finally we link the size of different operators to a common origin in Sec. [2.3](#sec:ana_magic){reference-type="ref" reference="sec:ana_magic"}. ## Heavy scalar fit {#sec:ana_heavy} As a first step we analyze only the measurements for the heavy scalar, as listed in Tab. [1](#tab:750_data){reference-type="ref" reference="tab:750_data"}. In Fig. [\[fig:5param_750\]](#fig:5param_750){reference-type="ref" reference="fig:5param_750"} we use this data to determine the five parameters \[\begin{aligned} \{ \, f^S_{WW}, f^S_{BB}, f^S_{GG}, \sin\alpha, c_{SHH} \, \} \;. \end{aligned}\] In our parametrization \(c_{SHH}\) accounts for the independent contributions to the \(SHH\) vertex from the dimension--five Lagrangian terms beyond the terms generating the mixing, as discussed in the Appendix. The best fit point for this analysis has \(-2\log L=8.9\), while the SM point leads to \(-2\log L=28.2\), within a \(3.1\sigma\) range for a 5-parameter study (in the Gaussian limit). In the upper left panel of Fig. [\[fig:5param_750\]](#fig:5param_750){reference-type="ref" reference="fig:5param_750"} we can see that within the displayed range of parameters both \(f^S_{WW}\) and \(f^S_{BB}\) are strongly correlated, and they present a flat direction. The correlation reflects the fact that they are the only Wilson coefficients contributing to the di-photon decay of the new scalar at tree level. Because \(f^S_{WW}\) is constrained through the decay \(S \rightarrow WW\), the di-photon excess cannot be accommodated through this coupling only. Due to that we find that \(|f^S_{BB}/\Lambda| > 2~{\ensuremath\rm TeV^{-1}}\) in the upper-center panel. This is caused by the fact that \(f^S_{BB}\) does not contribute to the \(SWW\) vertex, and in addition its contribution to the \(SZZ\) vertex is suppressed by the weak mixing angle. This allows us to explain the observed excess without getting into conflict with the exclusion bounds, what makes \(f^S_{WW}=0\) compatible with the best fit point, as shown in the upper-right panel. Moving on to the mixing angle, we find \(\sin \alpha < 0.15\) at 95% CL and for the displayed ranges of \(f^S_{BB}\) and \(f^S_{WW}\) in the lower-left panel. This bound comes from the absence of a heavy scalar signal in \(WW\) and \(ZZ\), but also in di-jet, \(t\bar{t}\), \(\tau\bar{\tau}\), and \(hh\) decay channels. It is linked to maximum assumed values for \(f^S_{BB}\) and \(f^S_{WW}\), because a larger mixing angle can be partially compensated by larger Wilson coefficients \(f^S_{BB}\)+\(f^S_{WW}\). For large values the di-photon branching ratio of the heavy scalar can exceed 50%, while the remaining decay channel modes are suppressed, allowing \(\sin \alpha\) to increase without conflicting with data. If we allow for extreme values of \(f^S_{BB}/\Lambda+f^S_{WW}/\Lambda \sim 250~{\ensuremath\rm TeV^{-1}}\), the upper bound on \(\sin \alpha\) goes up to \(0.3\). In the lower-center panel we again observe two distinct regions in \(f^S_{GG}\). The vertical region with \(f^S_{GG}/\Lambda < 1.5\) TeV\(^{-1}\) is characterized by a large branching ratio for \(S\rightarrow \gamma \gamma\), linked to large values of \(f^S_{BB}+f^S_{WW}\). The horizontal region with \(f^S_{GG}/\Lambda = 1.5~...~10\) TeV\(^{-1}\) is characterized by a large production rate for the new scalar and a total decay width driven by \(f^S_{GG}\). The upper limit on \(f^S_{GG}\) is set by di-jet searches, and the mixing in this regime has to be small to respect the limits from other decay channels. Finally, in the lower-right panel we show the correlation between the mixing angle and \(c_{SHH}\) from the limit on the decay \(S\rightarrow HH\). Fixing \(c_{SHH} = 0\) and generating the \(SHH\) interaction through the mixing angle alone has no effect on any of the other correlations.We proceed with an analysis allowing the new scalar to couple to the two fermions for which there are direct searches available. The analysis now includes \[\begin{aligned} \{ \, f^S_{WW}, f^S_{BB}, f^S_{GG}, \sin\alpha, c_{SHH}, f^S_{t}, f^S_{\tau} \, \} \;. \end{aligned}\] A selection of results is shown in Fig. [\[fig:7param_750\]](#fig:7param_750){reference-type="ref" reference="fig:7param_750"}. The fermionic Wilson coefficients \(f_t^S\) and \(f_\tau^S\) are constrained by \(t\bar{t}\) and \(\tau^+\tau^-\) resonance searches, as well as from an upper limit \(\Gamma_S < 25\) GeV which we assume throughout our analysis and which sets hard limits on \(f_t^S\) and \(f_\tau^S\). The best fit point of this run is only mildly better than before, \(-2\log L=8.3\). The limits on these two fermion couplings are stronger for smaller \(f_{BB}^S\), as illustrated for \(f_t^S\) in the upper-left panel, and \(f_{\tau}^S\) in the upper-center one. The reason is that in those regions the partial decay width to photons becomes small, and the required di-photon branching ratio translates into small fermionic couplings. Conversely, larger fermionic Wilson coefficients now allow for best fit regions with large \(f^S_{GG}\) and \(f^S_{BB}\) at the same time, as shown in the upper-right panel of Fig. [\[fig:7param_750\]](#fig:7param_750){reference-type="ref" reference="fig:7param_750"}. This is the main difference with respect to the reduced analysis shown in Fig. [\[fig:5param_750\]](#fig:5param_750){reference-type="ref" reference="fig:5param_750"}. The rest of correlations remain qualitatively unchanged, as can be seen in the lower panels of Fig. [\[fig:7param_750\]](#fig:7param_750){reference-type="ref" reference="fig:7param_750"}. In particular the upper 95% CL limit on the mixing angle is still \(\sin \alpha < 0.15\).In passing we note that all the results shown so far have been derived assuming a CP-even new scalar. Nevertheless, for the analysis up to this point the results remain unchanged when instead we assume a heavy CP-odd scalar. ## Combined Higgs portal fit {#sec:ana_combined} Next, we discuss the results for the general scenario, where we constrain the 17 parameters \[\begin{aligned} \{ \, f^S_{WW}, f^S_{BB}, f^S_{GG}, \sin\alpha, f^S_t, f^S_b, f^S_\tau, f_{WW}, f_{BB}, f_{GG}, f_{W}, f_B, f_{\phi,2}, f_{WWW}, f_t, f_b, f_\tau \, \} \end{aligned}\] from the combined measurements in the electroweak-Higgs, and the heavy scalar sector. We have fixed \(c_{SHH}=0\) given its minor impact on the fit results. In this case the best fit value has a likelihood of \(-2\log L=242.0\), for an analysis containing \(252\) measurements, while the Standard Model point leads to \(-2\log L=273.9\). In Fig. [\[fig:17param_750\]](#fig:17param_750){reference-type="ref" reference="fig:17param_750"} we show a reduced selection of correlations between Wilson coefficients. When adding the heavy scalar to the combined Higgs and gauge boson analysis, the potentially largest change in the results appears for \(f_W\) and \(f_B\). The twofold reason is illustrated in detail in the Appendix. First, focusing on the electroweak-Higgs phenomenology, while the contribution of \(f_W\) and \(f_B\) to the Higgs vertices is now weighted by the cosine of the mixing angle, their contribution to the triple gauge boson vertex is not. This generates a different pattern of Higgs-TGV correlations once we add the new scalar. Second, the mixing of the Higgs boson with the heavy scalar allows \(f_W\) and \(f_B\) to generate genuinely new Lorentz structure contributions to the \(SWW\), \(SZZ\) and \(SZ\gamma\) vertices, on top of the contributions from the rest of dimension--five and dimension--six operators. The first effect turns out to be negligible, and given the small allowed size for the mixing angle, the electroweak-Higgs measurements are not precise enough to be sensitive to the scalar mixing contributions. Conversely, the second effect is more important. The mild preference for non-zero \(f_W\) values from the electroweak-Higgs measurements  causes the best fit regions to generate the new contribution to the decays \(S \to WW, ZZ, Z\gamma\). These channels can be then better fit suppressing them further with a smaller mixing angle. The addition of the dimension--six operator causes then the upper bound on the scalar mixing to be mildly reduced with respect to the results in the previous section: now \(\sin \alpha < 0.10\) at 95% CL. This can be observed in the left panel of Fig. [\[fig:17param_750\]](#fig:17param_750){reference-type="ref" reference="fig:17param_750"}. Apart from this effect, the small mixing angle causes a lack of sizable correlations between both the new scalar sector and the electroweak-Higgs sector. Consequently, the results and two-dimensional planes involving dimension--five operators are very similar to the ones shown in Fig. [\[fig:7param_750\]](#fig:7param_750){reference-type="ref" reference="fig:7param_750"}. The planes involving dimension--six operators remain unchanged with respect to the results shown in Ref. , something that we illustrate in the center and right panels of Fig. [\[fig:17param_750\]](#fig:17param_750){reference-type="ref" reference="fig:17param_750"} for two of the dimension--six correlations. ## A common origin of operators {#sec:ana_magic} When we split a common scalar potential for two mixing states into a set of dimension--five and dimension--six operators, the question becomes how different the higher-dimensional effects in the light and heavy scalar couplings can really be. In this section we assume that the set of heavy scalar couplings are directly tied to their Higgs-like counter parts, \[\begin{aligned} {9} \frac{f_{GG}}{\Lambda^2} &=-2 \frac{f_{GG}^S}{\Lambda} \left|\frac{f_{GG}^S}{\Lambda}\right| \qquad \qquad \qquad & \frac{f_{f}}{\Lambda^2} &=-\frac{v}{m_f}\frac{f_{f}^S}{\Lambda}\left|\frac{f_{f}^S}{\Lambda}\right| \notag\\ \frac{f_{BB}}{\Lambda^2} & =-\frac{1}{4\pi^2}\frac{f_{BB}^S}{\Lambda}\left|\frac{f_{BB}^S}{\Lambda}\right| & \frac{f_{WW}}{\Lambda^2} & =-\frac{1}{4\pi^2}\frac{f_{WW}^S}{\Lambda}\left|\frac{f_{WW}^S}{\Lambda}\right| \;, \label{eq:prior1} \end{aligned}\] for \(f = b,t,\tau\). The relative signs and pre-factors ensure that the underlying new physics scales are consistent, as defined in Eq.[\[eq:lagd5\]](#eq:lagd5){reference-type="eqref" reference="eq:lagd5"}. For the fermion case, this is motivated by the need to have a minimal flavor violating structure in both dimension--five and dimension--six operators to avoid large flavor changing neutral currents . In a Bayesian language this approach would correspond to a Dirichlet prior, for example employed in the dark matter fit of Ref. , with an exponent parameter \(\alpha\gg1\).After imposing the relations in Eq.[\[eq:prior1\]](#eq:prior1){reference-type="eqref" reference="eq:prior1"}, we proceed to perform the combined Higgs, triple gauge boson vertex and heavy scalar analysis spanning the 11 free parameters \[\begin{aligned} \{ \, f^S_{WW}, f^S_{BB}, f^S_{GG}, f^S_t, f^S_b, f^S_\tau, \sin\alpha, f_{W}, f_B, f_{\phi,2}, f_{WWW} \, \} \;. \end{aligned}\] We have again fixed \(c_{SHH}=0\), while \(f_{WW}\), \(f_{BB}\), \(f_{GG}\), \(f_t\), \(f_b\) and \(f_\tau\) are set from Eq.[\[eq:prior1\]](#eq:prior1){reference-type="eqref" reference="eq:prior1"}. Interestingly, the best fit point is \(-2\log L=242.6\), i. e. within the analysis precision very close to the best fit point of the previous general scenario. This illustrates one of the most important conclusions: when dimension--five and dimension--six operators of a similar type are imposed to be related, there are still regions in the new physics parameter space which can accommodate the di-photon anomaly while respecting the constraints from the electroweak-Higgs measurements.In Fig. [\[fig:11param_750\]](#fig:11param_750){reference-type="ref" reference="fig:11param_750"} we again show a selection of two-dimensional correlations. In the upper-left panel we start with tight constraints on \(f_{BB}^S\) and also on \(f_{GG}^S\). Now \(f_{BB}^S/\Lambda\) no longer presents an unconstrained direction, as the reduced allowed region for values around \(-10~{\ensuremath\rm TeV^{-1}}\) is limited from the constraint that \(f_{BB}\) and hence \(f_{BB}^S\) is constrained by the Higgs measurements. Because of the minus signs in Eq.[\[eq:prior1\]](#eq:prior1){reference-type="eqref" reference="eq:prior1"} the region of allowed values for both \(f_{BB}^S\) and \(f_{WW}^S\) corresponds to the solution that flips the sign of the \(H\gamma\gamma\) vertex while respecting its measured size , as seen in the upper-center panel. In the case of \(f_{GG}^S\) and \(f_{GG}\) the several best fit regions are due to the measurement of a SM--like Higgs boson in gluon fusion production, the interference between \(f_{GG}\) and \(f_t\) , and the heavy scalar anomaly that excludes \(f_{GG}\) null values. As seen in the upper-right panel, the stronger constraints on \(f_{BB}^S\) are directly translated into a stringent 95\(\%\) CL bound on the mixing angle, \(\sin\alpha<0.02\). In the lower-left panel we show the impact of Eq.[\[eq:prior1\]](#eq:prior1){reference-type="eqref" reference="eq:prior1"} on \(f_t^S\). The fact that in this analysis \(f_{BB}^S\) is more constrained than in the general scenario implies that \(f_t^S\) is constrained to order one values, as expected from the \(f_{BB}^S\) vs \(f_t^S\) correlation in Fig. [\[fig:7param_750\]](#fig:7param_750){reference-type="ref" reference="fig:7param_750"}. The solution for \(f_t\) that flips the sign of the Higgs-Yukawa present in Ref.  is excluded through \(f_t^S\). This reduces the number of allowed regions for \(f_{GG}\), as compared to the electroweak-Higgs fit . In the case of \(f_\tau^S\) and \(f_b^S\), the allowed regions are limited by the \(H\tau^+\tau^-\) and \(Hb\bar{b}\) measurements. The \(v/m_f\) factors in Eq.[\[eq:prior1\]](#eq:prior1){reference-type="eqref" reference="eq:prior1"} lead to reduced allowed ranges for \(f_\tau^S\) in comparison to the previous general scenario. We illustrate in the lower-center panel the allowed region for two of the dimension--six operators not involved in the simplifications of Eq.[\[eq:prior1\]](#eq:prior1){reference-type="eqref" reference="eq:prior1"}, \(f_W\) and \(f_B\). They remain unaltered with respect to the general analysis or the electroweak-Higgs results . Conversely, in the lower-right panel of Fig. [\[fig:11param_750\]](#fig:11param_750){reference-type="ref" reference="fig:11param_750"} we illustrate the two parameter regions \(f_{BB}\) vs. \(f_{WW}\). There we see how the SM solution observed in the electroweak-Higgs analysis is now disfavored with respect to positive values for the Wilson coefficients.In this section we have illustrated the results of a constrained scenario imposing hard relations between the heavy scalar and Higgs operators in Eq.[\[eq:prior1\]](#eq:prior1){reference-type="eqref" reference="eq:prior1"}. Realistically, we would expect such relations to not be as strict. We therefore checked that relaxing Eq.[\[eq:prior1\]](#eq:prior1){reference-type="eqref" reference="eq:prior1"} and allowing for order-one variations does not qualitatively change our conclusions. Numerically, the bound on the mixing angle \(\sin \alpha\) becomes weaker once the relation \(f_{BB}\propto f_{BB}^S |f_{BB}^S|\) is relaxed. # Epitaph {#sec:summary} We have developed the framework to perform a combined analysis of the electroweak-Higgs sector extended with a new scalar to test Higgs portal scenarios. The theoretical description we have studied is that of a linear effective Lagrangian extended with the addition of a singlet scalar. The key question we face is the test of the portal structure hypothesis for an extended scalar sector. With that purpose we include a large set of Higgs event rates and kinematic distributions, combined with the recently implemented LHC triple gauge boson vertex distributions . As a test of a Higgs portal scenario we study the possibility that a di-photon anomaly recently observed at the LHC  could be part of an extended Higgs sector. For that we include a selection of relevant experimental searches for heavy resonances as listed in Tab. [1](#tab:750_data){reference-type="ref" reference="tab:750_data"}.Analyzing first the new scalar sector only, we recover the result that a non-zero value for a reduced set of singlet scalar effective operators (\(f_{GG}^S\), \(f_{BB}^S+f_{WW}^S\)) fits the observed anomaly in the di-photon channel, without conflicting with the lack of other positive observations, see Fig. [\[fig:5param_750\]](#fig:5param_750){reference-type="ref" reference="fig:5param_750"}. The mixing angle of the new singlet state with the Higgs boson can be sizable, the upper bound we find in the analysis is \(\sin(\alpha)<0.15\) at the 95% CL. The addition of fermionic dimension--five operators increases the allowed parameter space regions for the bosonic operators. However it has no impact on the maximum allowed mixing angle value, see Fig. [\[fig:7param_750\]](#fig:7param_750){reference-type="ref" reference="fig:7param_750"}. We then extend the analysis combining the new scalar sector with the electroweak-Higgs sector, using the Lagrangian description based on the dimension--six operators in Eq.[\[eq:ourleff\]](#eq:ourleff){reference-type="eqref" reference="eq:ourleff"}. In this extended scenario the upper bound on the mixing angle is further reduced in order to suppress the new dimension-six contributions to the heavy scalar non-observed decays. The upper bound is now \(\sin(\alpha)<0.1\) at 95% CL, with a size still compatible with Higgs portal hypothesis. Beyond this change, the maximum allowed mixing angle reduces the correlations between the Higgs-electroweak phenomenology and the hypothetical heavy scalar interactions. This leads to results that in the most general scenario are very similar to the ones of the individual Higgs-electroweak , and heavy scalar analysis, respectively. Motivated by a scalar portal scenario we define and test a hypothesis for a unique origin of the dimension--five and dimension--six operators studied in the analysis. Imposing Eq.[\[eq:prior1\]](#eq:prior1){reference-type="eqref" reference="eq:prior1"} we find new physics regions of parameters that fit the di-photon anomaly while being consistent with the lack of deviations measured on the electroweak-Higgs measurements. The upper bound on the mixing angle is reduced in this case to \(\sin(\alpha)<0.02\), due to the strong constraints on the operators modifying \(h\rightarrow \gamma\gamma\).
{'timestamp': '2016-07-18T02:08:07', 'yymm': '1607', 'arxiv_id': '1607.04562', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04562'}
# Introduction Determining the properties and phases of cold and dense nuclear matter, as found for example in the interior of neutron stars, is a difficult task, because neither lattice methods nor perturbation theory can be applied in this regime of QCD. Holographic methods allow the study of strongly-coupled QCD-like theories as a function of temperature and baryon chemical potential, to hopefully provide some useful qualitative information about the rich variety of phases and transitions that might be expected in QCD. Pioneering studies on cold and dense nuclear matter in holographic QCD have used the Sakai-Sugimoto model, as this is the preeminent example of a string theory realization of a holographic QCD-like theory. The two flavour version has a low energy description as a five-dimensional \(U(2)\) gauge theory with a Chern-Simons term, in which baryons correspond to topological solitons in the bulk. As baryon number is identified with the instanton number of the gauge field, the study of baryons requires the construction of Yang-Mills-Chern-Simons solitons in curved space with a prescribed instanton number. It is a significant challenge to calculate the required solitons, even numerically, and to date only the single baryon solution has been computed. The computation in the baryon number one sector is facilitated by the application of continuous symmetries that are not present for higher baryon numbers, or at finite baryon number density. As a consequence, there are currently no numerical solutions available at finite density, either in the full Sakai-Sugimoto model or its five-dimensional version. This absence of solutions has motivated the use of a number of approximate methods to study the Sakai-Sugimoto model at finite density. Often this involves the use of point-like instantons, but below we discuss some alternative approaches that share some common features with the effective kink model to be introduced in the present paper. An obvious way to simplify the problem at hand is to assume spatial homogeneity in the non-holographic spatial directions, so that the construction of the required finite density static solitons in the five-dimensional theory reduces to an effective one-dimensional problem. However, it is easy to show that there are no smooth spatially homogeneous gauge potentials that yield a non-zero baryon number density and have finite energy per baryon. To circumvent this problem, the approach in imposes homogeneity through the use of singular configurations. This reduces the problem to the determination of a profile function that specifies the singluar gauge potential as a function of the holographic coordinate. An alternative way to obtain spatial homogeneity is to replace the three-dimensional Euclidean spatial slice by a three-sphere. As there are smooth finite energy spatially homogeneous gauge potentials on the three-sphere, this yields a finite baryon number density that can be varied by changing the radius of the three-sphere. However, in the studies in the dependence on the holographic direction is then assumed to be a fixed self-dual form, so an effective one-dimensional kink model is not obtained and the possibility of different types of behaviour in the holographic direction is excluded. The approach taken in the present paper may be viewed as a hybrid of the above two methods. We apply a homogeneous approximation at the level of the field strength, rather than at the level of the gauge potential, in a way that produces an expression for the field strength that is very similar to that obtained on the three-sphere. However, we retain an arbitrary dependence on the holographic coordinate that leads to the derivation of an effective kink theory in the holographic direction. As the baryon chemical potential is increased this kink theory yields a first order phase transition, just below the baryon mass, to lightly bound nuclear matter, as found in QCD. We explain why our homogeneous approximation should be viewed as a smeared crystal that is expected to provide a lower bound on the energy, as it is an unattainable idealization that distributes the energy perfectly equally in space, which a true crystal is unable to match. We provide a justification for this expectation based on the simpler case of the Skyrme crystal. Homogeneous approximations are often employed to study field theory at very high density and usually overestimate the true energy because of the existence of symmetry breaking negative modes. It is important to note that our smearing approximation is in contrast to this standard application of homogeneity, as smearing lowers the energy to provide an underestimate that is expected to give valid predictions even for the small to moderate densities at which bound nuclear matter is formed. As baryon number density is increased it is natural to expect that the bulk soliton explores more of the holographic direction and this has been interpreted as a holographic realization of the quarkyonic phase, where there is a quark Fermi sea with a baryonic Fermi surface. It has also been proposed that this provides a mechanism for approximate chiral symmetry restoration. However, the details of the way in which the bulk soliton expands into the holographic direction are unknown and several different possibilities have been suggested. One proposal is a dyonic salt phase, in which the bulk soliton splits into constituents at high density. This suggestion is motivated by a point particle approximation and the analogy with calorons, which are flat space self-dual instantons that can split into monopole constituents if a periodic direction is smaller than the size of an instanton. We find that our effective kink model provides an explicit realization of the dyonic salt phase, as the kink splits into a pair of half-kink constituents as the density is increased. Furthermore, there is a simple explanation for this splitting into constituents in terms of the novel potential that appears in the effective kink model and its explicit dependence on the holographic coordinate. A second proposal, based on the use of approximations involving flat space calorons and dilute instantons, is a baryonic popcorn phase, in which a sequence of transitions takes place where the soliton crystal develops additional layers in the holographic direction with increasing density. The kink model is also able to capture the phenomenon of baryonic popcorn and displays a first order phase transition that generates an additional soliton layer in the holographic direction that is energetically preferred over the dyonic salt phase. The fact that the phase transition of the first pop into baryonic popcorn occurs before the dyonic salt phase agrees with the results of full numerical simulations of low-dimensional toy models, based on sigma model instantons rather than Yang-Mills intantons. However, we find that the kink model predicts only one pop, rather than the sequence of pops suggested by previous point particle and flat space approximations. Once again, the form of the potential that appears in the effective kink model provides a simple explanation of the absence of additional pops once the first pop has taken place. In the kink model, the two layers produced by the single pop form the surface of a soliton bag that increases in size as the baryon chemical potential is increased. The interior of the bag is filled with abelian electric potential and the instanton charge density is localized on the surface of the bag. Our soliton bag is therefore very similar to the magnetic bag that approximates a large number of coincident non-abelian BPS monopoles. A magnetic bag also has the topological charge density localized on its surface, with the interior and the exterior of the bag corresponding to different values of the modulus of the Higgs field. In our soliton bag the abelian electric potential plays the role of the modulus of the Higgs field, taking different values inside and outside the bag. Note that our soliton bag is qualitatively different from the previously proposed instanton bag, in which monopole walls are embedded into the Sakai-Sugimoto model to produce a bag where the interior is filled with instanton charge density, rather than this density being localized on the surface of the bag. # An effective holographic kink theory In our study of the Sakai-Sugimoto model we choose to work with the Yang-Mills theory that results from the non-abelian Dirac-Born-Infeld action at leading order in \(\alpha'\). This simplification has the advantage that we do not need to choose a particular proposal for the non-abelian form of the Dirac-Born-Infeld action. The five-dimensional \(U(2)\) Yang-Mills gauge theory version of the Sakai-Sugimoto model involves a spacetime with a warped metric of the form \[\label{metric} ds^2 = H(z) \,dx_{\mu} dx^{\mu} + \frac{1}{H(z)}dz^2,\] where \(x_\mu,\) with \(\mu=0,1,2,3,\) are the coordinates of four-dimensional Minkowski spacetime and \(z\) is the spatial coordinate in the additional holographic direction. The warp factor in this expression is \(H(z)=(1+z^2)^{2/3}\) and the signature of the metric is \((-,+,+,+,+)\). In this section we follow the notation of, in which the hermitian \(U(2)\) gauge potential is decomposed into a non-abelian \(SU(2)\) component \(A\) and an abelian \(U(1)\) component \(\hat A\), with lowercase latin indices (excluding \(z\)) running over the three non-holographic spatial directions \(i=1,2,3\), whilst uppercase latin indices run over all four spatial directions, \(I=1,2,3,z.\) We immediately restrict to the case of time independent fields, where the appropriate static ansatz is \[{A}_0=0, \qquad A_I={A}_I(x_{J}), \qquad \hat{A}_0=\hat{A}_0(x_{J}), \qquad \hat{A}_{I} = 0.\] Then, in units of \(N_c\lambda/(216\pi^3)\), where \(N_c\) is the number of colours and \(\lambda\) is the 't Hooft coupling, the static energy of the five-dimensional Sakai-Sugimoto model is given by \[E=\int\bigg\{-\frac{(\partial_i\hat A_0)^2}{2H^{1/2}}-\frac{H^{3/2}}{2}(\partial_z\hat A_0)^2 +\frac{\mbox{Tr}(F_{ij}^2)}{2H^{1/2}}+H^{3/2}\mbox{Tr}(F_{iz}^2) +\frac{4}{\Lambda}\hat A_0\epsilon_{ijk}\mbox{Tr}(F_{iz}F_{jk})\bigg\}\,d^3x\,dz, \label{energy}\] where \(\Lambda=8\lambda/(27\pi)\) is the rescaled 't Hooft coupling. The final term in the above expression arises from the Chern-Simons term. Baryon number is identified with the \(SU(2)\) instanton number of the soliton \[N=\frac{1}{8\pi^2}\int\epsilon_{ijk}\mbox{Tr}(F_{iz}F_{jk})\,d^3x\,dz, \label{charge}\] and this provides the energy bound \(E\ge 8\pi^2 |N|.\) From the final term in ([\[energy\]](#energy){reference-type="ref" reference="energy"}) we see that the instanton charge density sources the abelian electric field. In the gauge \(A_z=0\) the holonomy that carries the pion degrees of freedom \[U({\bf x})={\cal P}\mbox{exp}\bigg(i\int_{-\infty}^\infty A_z({\bf x},z)\,dz\bigg),\] appears in the boundary condition for \(A_i\), namely, \[A_i({\bf x},z)\to R_i({\bf x}) \quad \mbox{ as }\quad z\to\infty, \qquad \mbox{ where } \quad R_i({\bf x})=i(\partial_iU)U^{-1}. \label{bc}\] The holonomy \(U({\bf x}):{\mathbb R}^3\to SU(2)\) carries the topology of the field configuration due to the equality \(N=\mbox{deg}\,U\in\mathbb{Z}=\pi_3(SU(2))\), where the compactification of \(\mathbb{R}^3\) is a consequence of the fact that \(U({\bf x})\) tends to the identity matrix as \(|{\bf x}|\to \infty.\) Note that the pure gauge currents, \(R_i({\bf x})\), satisfy the zero curvature condition \[\partial_iR_j-\partial_jR_i+i[R_i,R_j]=0. \label{zerocurvature}\] The first step in deriving our effective kink model is to work in the gauge \(A_z=0\) and apply the separable approximation \[A_i({\bf x},z)=R_i({\bf x})\psi(z), \label{sep}\] where the real function \(\psi(z)\) satisfies the boundary conditions \(\psi(-\infty)=0\) and \(\psi(\infty)=1\), in order to reproduce the correct boundary condition ([\[bc\]](#bc){reference-type="ref" reference="bc"}). Applying the separable approximation ([\[sep\]](#sep){reference-type="ref" reference="sep"}), and making use of the zero curvature relation, yields the gauge field strength \[F_{ij}=-i[R_i,R_j]\psi(1-\psi), \qquad F_{iz}=-R_i\psi'. \label{homog}\] Our second step is to make a homogeneous approximation by assuming that the gauge field strength is independent of \({\bf x}\). Explicitly, we set \(R_i=-\beta\sigma_i/2\), where \(\sigma_i\) denote the Pauli matrices and \(\beta\) is a real constant that we take to be positive, as the sign of \(\beta\) will turn out to be equal to the sign of the baryon number density. It is crucial that we apply the homogeneous approximation at the level of the field strength and not at the level of the gauge potential, since there are no homogeneous gauge potentials that yield the formulae ([\[homog\]](#homog){reference-type="ref" reference="homog"}). This is because the zero curvature condition ([\[zerocurvature\]](#zerocurvature){reference-type="ref" reference="zerocurvature"}) is obviously violated by restricting to constant non-commuting currents. By imposing homogeneity of the field strength, rather than the gauge potential, we have been able to incorporate the zero curvature relation before restricting to homogeneous fields that violate it. If a homogeneous approximation is applied directly at the level of the gauge potential then the zero curvature relation ([\[zerocurvature\]](#zerocurvature){reference-type="ref" reference="zerocurvature"}) implies that the field is topologically trivial. The only way to reintroduce instanton charge is to incorporate it via a discontinuous gauge potential , with the discontinuity being the source of the instanton charge. Essentially, the winding of the field is then moved from infinity onto a discontinuity in the bulk. The advantage of our continuous approach is that we avoid the need to deal directly with the non-abelian gauge potential, by restricting attention to the continuous physical fields. In summary, our homogeneous approximation is \[F_{ij}=\frac{1}{2}\beta^2\psi(1-\psi)\epsilon_{ijk}\sigma_k, \qquad F_{iz}=\frac{1}{2}\beta \psi'\sigma_i, \qquad \hat A_0=\omega(z). \label{ansatz}\] These expressions are very similar to those that appear if the spatial slice \(\mathbb{R}^3\) is replaced by \(S^3\) with a finite radius, as studied in. In the case of the three-sphere, smooth spatially homogeneous gauge potentials exist and directly generate relations analogous to ([\[ansatz\]](#ansatz){reference-type="ref" reference="ansatz"}), with \(\beta\) related to the inverse of the radius of the three-sphere. However, in the work in the analogue of the kink function \(\psi(z)\) is taken to be a fixed self-dual form, so an effective kink model is not obtained. Taking a fixed form for the kink function prevents a study of the way in which the bulk soliton explores the holographic direction with increasing density and the associated different phases. This is the main purpose of the present paper. It might be interesting to repeat the analysis presented here for the case of the three-sphere with an unfrozen kink field, where genuine homogeneous gauge potentials underpin the approximation. To begin with, we work in the canonical ensemble with fixed baryon number density and no explicit baryon chemical potential. The boundary condition on the real function \(\omega(z)\) is therefore \(\omega(\pm\infty)=0,\) because in holographic QCD the boundary value of \(\hat A_0\) is proportional to the baryon chemical potential. At finite density, the true spatial distribution of the fields in the non-holographic directions is expected to form a soliton crystal. Our homogeneous approximation may be viewed as a smeared version of the crystal, and we expect that this approximation provides a lower bound on the true crystal energy, as homogeneity is an unattainable idealization. A justification for this expectation is provided in the appendix, where we consider the related, though simpler, case of the Skyrme crystal. Substituting the expressions ([\[ansatz\]](#ansatz){reference-type="ref" reference="ansatz"}) into ([\[energy\]](#energy){reference-type="ref" reference="energy"}) and ([\[charge\]](#charge){reference-type="ref" reference="charge"}) gives the energy per unit 3-volume \[{\cal E}=\frac{1}{2}\int_{-\infty}^\infty\bigg\{ 3\beta^2H^{3/2}\psi'^2+\frac{3}{H^{1/2}}\beta^4\psi^2(1-\psi)^2-H^{3/2}\omega'^2 +\frac{24}{\Lambda}\beta^3\omega\psi(1-\psi)\psi'\bigg\}\,dz \label{henergy}\] and the baryon number density (ie. instanton number per unit 3-volume) \[\rho=\int_{-\infty}^\infty\frac{3\beta^3}{8\pi^2}\psi(1-\psi)\psi'\,dz =\frac{\beta^3}{16\pi^2}. \label{density}\] The field equation for \(\omega\) that follows from the variation of ([\[henergy\]](#henergy){reference-type="ref" reference="henergy"}) is \[\big(H^{3/2}\omega'\big)'=\frac{12\beta^3}{\Lambda}\bigg(\frac{1}{3}\psi^3-\frac{1}{2}\psi^2\bigg)'. \label{omegaeqn}\] As energy is minimized by a bulk configuration that is localized around \(z=0\), we impose the symmetry that both the instanton charge density and \(\omega\) are even functions of \(z\), that is, \(\omega(-z)=\omega(z)\) and \(\psi(-z)=1-\psi(z)\). We can therefore restrict our discussion to the region \(z\ge 0\), together with the boundary conditions \(\omega'(0)=0\) and \(\psi(0)=\frac{1}{2}\). Integrating ([\[omegaeqn\]](#omegaeqn){reference-type="ref" reference="omegaeqn"}) once and applying these boundary conditions at \(z=0\) yields \[H^{3/2}\omega'=\frac{12\beta^3}{\Lambda}\bigg(\frac{1}{3}\psi^3-\frac{1}{2}\psi^2+\frac{1}{12}\bigg). \label{wpp}\] Taking the limit of this equation as \(z\to\infty\) gives \[\lim_{z\to\infty}z^2\omega'=-\frac{\beta^3}{\Lambda}=-\frac{16\pi^2\rho}{\Lambda}.\] We therefore obtain the usual holographic result, relating the coefficient of the asymptotic behaviour to the baryon number density, \[\omega=\frac{16\pi^2\rho}{\Lambda z}+o\bigg(\frac{1}{z}\bigg). \label{asymptotic}\] We can rewrite the final term in ([\[henergy\]](#henergy){reference-type="ref" reference="henergy"}) by applying an integration by parts to obtain the identity \[\int_{-\infty}^\infty\omega\psi(1-\psi)\psi'\,dz =\int_{-\infty}^\infty\omega'\bigg(\frac{1}{3}\psi^3-\frac{1}{2}\psi^2+\frac{1}{12}\bigg)\,dz.\] Using this result, together with ([\[wpp\]](#wpp){reference-type="ref" reference="wpp"}), we obtain \[{\cal E}=\int_0^\infty\bigg\{ 3\beta^2H^{3/2}\psi'^2+\frac{3}{H^{1/2}}\beta^4\psi^2(1-\psi)^2 +\frac{144\beta^6}{\Lambda^2H^{3/2}}\bigg(\frac{1}{3}\psi^3-\frac{1}{2}\psi^2+\frac{1}{12}\bigg)^2 \bigg\}\,dz, \label{psionly}\] where all reference to \(\omega\) has been eliminated. This is the energy of our effective kink model. Note that imposing the flat space self-duality equation \(F_{IJ}=\frac{1}{2}\epsilon_{IJKL}F_{KL}\) on the homogeneous fields ([\[ansatz\]](#ansatz){reference-type="ref" reference="ansatz"}) yields the first order equation \[\psi'=\beta\psi(1-\psi),\] with kink solution \[\psi=\frac{1}{1+e^{-\beta z}}. \label{selfdual}\] This is the self-dual form of the kink function that is assumed to hold for all densities in. In the flat space limit \(H=1\) with no Chern-Simons term (\(\Lambda=\infty\)), the energy ([\[psionly\]](#psionly){reference-type="ref" reference="psionly"}) of the self-dual solution ([\[selfdual\]](#selfdual){reference-type="ref" reference="selfdual"}) is \({\cal E}=\beta^3/2=8\pi^2\rho\), so the BPS behaviour of flat space instantons is recovered within our homogeneous approximation. The field equation that follows from the variation of ([\[psionly\]](#psionly){reference-type="ref" reference="psionly"}) is \[\psi''+\frac{3H'}{2H}\psi' +\frac{\beta^2}{H^2}\psi(1-\psi)(2\psi-1) +\frac{48\beta^4}{\Lambda^2H^{3}}\psi(1-\psi)\bigg(\frac{1}{3}\psi^3-\frac{1}{2}\psi^2+\frac{1}{12}\bigg)=0. \label{psieqn}\] We solve this equation numerically using a gradient flow algorithm and the change of variable \(z=\tan u,\) to map the infinite range of \(z\) onto a finite range of \(u.\) All the numerical results presented in this paper are computed with the fiducial value \(\Lambda=10.\) In Figure [\[fig-energy\]](#fig-energy){reference-type="ref" reference="fig-energy"} the blue curve shows a plot of the ratio of the energy to the lower bound, that is \({\cal E}/(8\pi^2\rho)\), as a function of the baryon number density \(\rho\). We see that there is a non-zero optimum density \(\rho_\star= 0.027\), corresponding to the analogue of the nuclear matter density in QCD. As described below, this optimal density is associated with a critical value of the baryon chemical potential at which there is a first order phase transition to an equilibrium density of baryons. In the above we have worked in the canonical ensemble with fixed baryon number density, but an alternative is to work in the grand canonical ensemble with fixed baryon chemical potential \(\mu\). Applying the standard holographic dictionary, in the grand canonical ensemble a baryon chemical potential corresponds to a non-zero boundary value for \(\hat A_0\). With our chosen normalizations, a vectorial chemical potential \(\mu\) corresponds to the boundary condition \[\omega(\pm\infty)=-\frac{\mu \Lambda}{32\pi^2}. \label{chempot}\] This is because a shift in \(\omega\) by this constant value returns us to the previous boundary condition \(\omega(\pm\infty)=0\), but from ([\[henergy\]](#henergy){reference-type="ref" reference="henergy"}) we see that the energy then transforms into the grand potential (also known as the Landau free energy) \(\Omega={\cal E}-\mu\rho.\) Minimizing \(\Omega\) at fixed \(\mu\) we obtain the function \(\Omega(\mu)\) displayed as the blue curve in Figure [\[fig-freeenergy\]](#fig-freeenergy){reference-type="ref" reference="fig-freeenergy"}. The associated relation between the chemical potential \(\mu\) and the baryon number density \(\rho\) is displayed as the blue curve in Figure [\[fig-chempot\]](#fig-chempot){reference-type="ref" reference="fig-chempot"}. These curves show that a non-zero value for \(\rho\) is obtained for \(\mu\ge\mu_\star,\) where \(\mu_\star=98\) (for \(\Lambda=10\)) is the critical value of the chemical potential at which the density is indeed equal to \(\rho_\star\). For attraction between baryons, the critical value of the chemical potential \(\mu_\star\) must be less than the baryon mass \(M_B.\) In the units we are using, the self-dual single instanton approximation to the baryon yields the formula \[M_B=2\pi^2\bigg(4+\frac{32}{\Lambda}\sqrt{\frac{2}{15}}\bigg), \label{baryonmass}\] where terms of order \(1/\Lambda^2\) have been neglected. Substituting our fiducial value \(\Lambda=10\) into ([\[baryonmass\]](#baryonmass){reference-type="ref" reference="baryonmass"}) gives \(M_B=102,\) so indeed our numerically computed value \(\mu_\star=98\) is less then \(M_B\) and baryons form bound states. The percentage binding energy per nucleon is given by \[\Delta = (M_B-\mu_\star)/M_B\times 100\%,\] yielding \(\Delta=4\%\) for the chosen value \(\Lambda=10.\) In the limit \(\Lambda\to\infty\) the BPS result is recovered, \(M_B\to 8\pi^2\) and \(\mu_\star\to M_B\), giving zero binding energy. Although we expect holographic QCD to provide only qualitative information about QCD with three colours, the above results imply that there must be a value of \(\Lambda\) greater than 10 at which a realistic nuclear binding energy \(\Delta \sim 0.9\%\) is obtained, and indeed we compute that the appropriate value is \(\Lambda\sim 18\). Returning to the representative value \(\Lambda=10\), in the top row in Figure [\[fig-psi\]](#fig-psi){reference-type="ref" reference="fig-psi"} we display the kink profile function \(\psi(z)\) for the densities \(\rho=0.1\) (left column) and \(\rho=2\) (right column). In the middle row we plot the corresponding instanton number densities per unit 3-volume, and in the final row we display the abelian electric potential \(\omega(z)\). We see that at high density the kink splits into two half-kink constituents. This splitting of the soliton layer into a pair of constituents is the predicted dyonic salt phenomenon, corresponding to the fact that a periodic instanton (a caloron) splits into monopole constituents at high density. Note that a half-kink cannot exist in isolation as a finite energy configuration because the profile function \(\psi(z)\) of a half-kink interpolates between values that differ by \(\frac{1}{2}\), but for finite energy this difference must be equal to \(\pm1\) or \(0\). This mirrors the caloron situation, where the caloron splits into monopole constituents that are not finite energy configurations in isolation but together combine to form a well-defined periodic instanton. The form of the energy ([\[psionly\]](#psionly){reference-type="ref" reference="psionly"}) of the effective kink model provides a simple explanation for the split into half-kink constituents, as follows. The kink model contains a derivative term plus two potential terms, which are novel in a kink theory due to the form of the explicit dependence on the holographic spatial coordinate. The first potential term has an explicit spatial dependence that grows like \(z^{-2/3}\) for large \(z\). Such a term does not decay fast enough to be integrable, so the associated multiplying factor must tend to zero as \(z\to\pm\infty.\) This enforces the boundary conditions \(\psi(\pm\infty)\in\{0,1\}.\) The second potential term has an explicit spatial dependence that grows like \(z^{-2}\), which decays fast enough that finite energy considerations impose no conditions on the boundary values of \(\psi(z)\) from this term. However, as this term has a coefficient of \(\beta^6\) then at high density it is, at least locally, of more relevance than the first potential term, which has a coefficient of only \(\beta^4\). The second potential term is minimized by the value \(\psi=\frac{1}{2}\), as it vanishes at this value. Putting all this toegther we see that as the density increases the second potential term induces the kink field \(\psi(z)\) to remain close to the value \(\frac{1}{2}\) over an increasingly large range of \(z\), although the finite energy requirement of the first potential term always forces the kink boundary conditions to ultimately be attained. This is the simple reason why the kink splits into a pair of half-kinks. Previous results on low-dimensional models suggest that a baryonic popcorn transition appears before the formation of dyonic salt. In other words, at densities high enough to split the kink into a pair of half-kink constituents, we expect that there is a competing solution that describes a double layer and has lower energy than the single layer considered in this section. In the following section we examine this possibility by constructing a double layer configuration and calculating its energy as a function of baryon number density. # Baryonic popcorn and soliton bags The effective kink model, with energy given by ([\[psionly\]](#psionly){reference-type="ref" reference="psionly"}), has an anti-kink solution with boundary conditions \(\psi(-\infty)=1\) and \(\psi(\infty)=0.\) It is obtained from the kink solution by the transformation \(\psi\mapsto 1-\psi\) and has the same energy per unit volume as the kink but has a negative baryon number density. However, if the ansatz ([\[ansatz\]](#ansatz){reference-type="ref" reference="ansatz"}) is modified by the replacement \(F_{ij}\mapsto-F_{ij}\) then the anti-kink now has a positive baryon number density, and this provides an equivalent anti-kink formulation of the single layer described in the previous section in terms of a kink. We can construct a double layer configuration by gluing together two single layers in a continuous manner. Explicitly, we apply the approximation ([\[ansatz\]](#ansatz){reference-type="ref" reference="ansatz"}) for \(z\ge 0\) with kink boundary conditions on the half-line, \(\psi(0)=0\) and \(\psi(\infty)=1.\) For \(z\le 0\) we take the same approximation ([\[ansatz\]](#ansatz){reference-type="ref" reference="ansatz"}) after the replacement \(F_{ij}\to-F_{ij}\) with anti-kink boundary conditions on this half-line, \(\psi(-\infty)=1\) and \(\psi(0)=0.\) This corresponds to taking both \(\psi(z)\) and \(\omega(z)\) to be even functions of \(z\) and gives a continuous instanton charge density. The instanton charge density vanishes at the join at \(z=0\), and although it is continuous at this point it is not smooth there. Note that the parity of \(\omega\) is as before, in agreement with the requirement that we consider a vectorial baryon chemical potential. The baryon number density for this two layer configuration is \[\rho=\int_{0}^\infty\frac{6\beta^3}{8\pi^2}\psi(1-\psi)\psi'\,dz =\frac{\beta^3}{8\pi^2},\] which is twice the value ([\[density\]](#density){reference-type="ref" reference="density"}) for the single layer. Integrating the field equation ([\[omegaeqn\]](#omegaeqn){reference-type="ref" reference="omegaeqn"}) for \(\omega\), and imposing the new boundary condition \(\psi(0)=0\) and \(\omega'(0)=0\), gives for \(z\ge 0\) \[H^{3/2}\omega'=\frac{12\beta^3}{\Lambda}\bigg(\frac{1}{3}\psi^3-\frac{1}{2}\psi^2\bigg). \label{wpp2}\] Taking the limit of this equation as \(z\to\infty\) provides the relation \[\lim_{z\to\infty}z^2\omega'=-\frac{2\beta^3}{\Lambda}=-\frac{16\pi^2\rho}{\Lambda},\] again reproducing the correct asymptotic behaviour ([\[asymptotic\]](#asymptotic){reference-type="ref" reference="asymptotic"}). As in the single layer case, an integration by parts yields the following expression for the double layer energy, \[{\cal E}=\int_0^\infty\bigg\{ 3\beta^2H^{3/2}\psi'^2+\frac{3}{H^{1/2}}\beta^4\psi^2(1-\psi)^2 +\frac{144\beta^6}{\Lambda^2H^{3/2}}\bigg(\frac{1}{3}\psi^3-\frac{1}{2}\psi^2\bigg)^2 \bigg\}\,dz, \label{psionly2}\] as a function of \(\psi\) only. We obtain the double layer solution by numerically solving the static field equation for \(\psi(z)\) that follows from the variation of this energy. The red curve in Figure [\[fig-energy\]](#fig-energy){reference-type="ref" reference="fig-energy"} is a plot of the ratio of the energy to the lower bound, as a function of the baryon number density, for the double layer solution. From this figure we see that the double layer has a lower energy than the single layer beyond the critical density \(\rho_2= 0.22.\) This is the critical density for a homogeneous baryonic popcorn transition to a double layer, and is well below the density at which dyonic salt appears. This result is therefore in agreement with the previous low-dimensional studies mentioned earlier, where a popcorn transition also appears before the dyonic salt phase. In the top row in Figure [\[fig-psi2\]](#fig-psi2){reference-type="ref" reference="fig-psi2"} we display the double layer profile function \(\psi(z)\) for the densities \(\rho=2\) (left column) and \(\rho=10\) (right column). In the middle row we plot the corresponding instanton number densities per unit 3-volume, and in the final row we display the abelian electric potential \(\omega(z)\). Although there are some qualitative similarities between a double layer and a single layer that has split into two half-layer constituents, the crucial distinction is that a half-layer cannot exist in isolation as a finite energy configuration. There are, of course, significant quantitative differences, including the fact that the double layer configuration has a much lower energy at high densities. The free energy of the double layer as a function of the chemical potential is shown as the red curve in Figure [\[fig-freeenergy\]](#fig-freeenergy){reference-type="ref" reference="fig-freeenergy"}. This reveals that within the grand canonical ensemble there is a critical value of the chemical potential, given by \(\mu\ge\mu_2=185\) for \(\Lambda=10\), above which the free energy of the double layer is less than that of a single layer. This results in the baryonic popcorn first order phase transition seen in Figure [\[fig-chempot\]](#fig-chempot){reference-type="ref" reference="fig-chempot"} at \(\mu=\mu_2\), where the configuration pops from the single layer (given by the blue curve) to the double layer (given by the red curve). Note that in Figure 3 we only plot the single layer and double layer portions of the curve for the range of chemical potentials at which each is the global minimum of the free energy. The curves extend beyond these segments to physically irrelevant regions in which they are no longer global minima. Given the results in the previous section, where a single layer splits into half-layer constituents at high density, one might naively expect a similar phenomenon to take place for a double layer, with each layer splitting into half-layer constituents at high density. However, the right column in Figure [\[fig-psi2\]](#fig-psi2){reference-type="ref" reference="fig-psi2"} demonstrates that the double layer does not split, even at very high densities. This has a simple explanation, again obtained by examining the form of the novel potential terms in the effective energy ([\[psionly2\]](#psionly2){reference-type="ref" reference="psionly2"}). This time, we see that the final potential term in this effective energy is minimal only at \(\psi=0\) (for \(\psi\in[0,1]\)). Therefore, unlike the case of a single kink, the final potential term does not induce the kink or anti-kink to stay close to any new value as the density is increased. Hence the double layer does not split, but rather the two layers simply increase their separation as the density increases. As the density increases and the two layers move further apart, it would be reasonable to surmise that further pops would occur, since there now appears to be available space around \(z=0\) to generate new layers. However, this simple view ignores the fact that the region between the two layers is not empty but instead is full of abelian electric potential, as demonstrated by the plots in the bottom row of Figure [\[fig-psi2\]](#fig-psi2){reference-type="ref" reference="fig-psi2"}. The two layers form the surface of a soliton bag, with the interior of the bag associated with an approximately constant non-zero value of the abelian electric potential. The abelian electric potential decays to zero outside the bag, with the surface of the bag being the transition region where both the instanton charge density and the electric field are localized. This is similar to the magnetic bag description of large charge non-abelian monopoles, where the surface of the bag separates regions of zero and non-zero values for the modulus of the Higgs field. As the system under consideration is periodic in three spatial directions (approximated by homogeneity) then the surface of the bag is not a single connected component, like the magnetic bag in three-dimensional Euclidean space, but instead consists of two disconnected components corresponding to the top and bottom of the bag. This is why the soliton bag requires two layers. The creation of more layers through further baryonic popcorn transitions would produce a kind of multi-layer bag. In the monopole context, the possibility of multi-layer magnetic bags has been investigated in with the conclusion that these typically consist of only a single bag surrounded by layers of isolated unit charge monopoles. Any attempt to create a multi-bag configuration automatically rules out an interior bag that carries any significant fraction of the total magnetic charge. The soliton bag description therefore suggests the absence of additional baryonic popcorn transitions. Further support for this view is obtained by extending our anti-kink plus kink double layer approximation, denoted \(\bar K K\), to additional layers by the inclusion of more anti-kinks/kinks. For example, a 4-layer \(\bar K K\bar K K\) approximation has been studied where all contiguous anti-kinks and kinks are joined in the same continuous manner as in the double layer approximation. The location of the joins, together with the fraction of instanton charge carried by each layer, are allowed to vary and the resulting energy minimizing configurations computed. As expected from the similar monopole magnetic bag story, these computations yield only signficiant instanton charge density in the outer layer and the inner layer is irrelevant. A recent study, applying a different approximation, also reached the same conclusion that a double layer is the preferred configuration at high density and additional layers are not generated. Finally, it is important to note that our soliton bag is qualitatively different from the instanton bag proposed in as a description of the high density phase in the Sakai-Sugimoto model. The instanton bag is obtained via an initial compactification of one of the spatial directions to allow the embedding of a monopole wall. A pair of monopole walls are then patched together to form the surface of the instanton bag, with the result that the interior is filled with instanton charge density. This contrasts with our soliton bag, where the instanton charge density is localized on the surface of the bag. Our kink approximation, with an appropriate shape for the kink profile function, could produce a bag filled with instanton charge density, but such a shape does not appear when the profile function is obtained by minimizing the energy of the effective kink model. # Conclusion It is an open problem to understand the phases of cold and dense holographic nuclear matter as a function of baryon number density. The distribution of baryonic matter in both the non-holographic and holographic directions is unknown and this has led to a number of different approximate descriptions and suggestions for phenomena that might occur. In this paper we have assumed homogeneity in the non-holographic spatial directions to investigate the distribution in the holographic direction. As the holographic coordinate corresponds to an energy scale in the boundary theory, then understanding this aspect is likely to be the key to a holographic description of the baryonic Fermi surface of a quarkyonic phase. Our homogeneous holographic nuclear matter is described by an effective holographic kink theory, which we have shown is capable of a simultaneous realization of a number of previously suggested phases, including dyonic salt, baryonic popcorn and soliton bags. An advantage of this unifying description is that we can compare the various phases and determine which phase is preferred as the baryon chemical potential varies. The effective kink model also successfully reproduces the QCD behaviour of a first order phase transition to lightly bound nuclear matter, at a value of the baryon chemical potential that is just below the baryon mass, in agreement with QCD. In holographic QCD the spectrum of fluctuations in the holographic direction determines the masses of the vector and axial vector mesons. As the background solution changes with increasing baryon number density then so does the spectrum, and this has been proposed as a mechanism for approximate chiral symmetry restoration, as the masses of vector and axial vector mesons could converge with increasing density. However, as the details of the background solution are unavailable, to date the investigations of this issue have been mostly qualitative. It would be interesting to study this aspect within our effective kink model and to compute some quantitative data. Even within our homogeneous approximation, this is still a non-trivial calculation, because of the mixing of various modes, but it should be a tractable problem that we hope to address in the near future. Other avenues for future research include extending our approach to the finite temperature regime and investigating how kinky holographic nuclear matter responds to external electric and magnetic fields by introducing new boundary conditions for \(\hat A_\mu\), as in. Given that our soliton bag has some features in common with the magnetic bag description of a large number of coincident \(SU(2)\) magnetic monopoles, it could be interesting to study both problems for higher rank gauge groups, as new features certainly emerge for non-abelian monopoles beyond \(SU(2)\). In this paper we have restricted our investigations to the effective five-dimensional Yang-Mills-Chern-Simons version of the Sakai-Sugimoto model. However, it would be a simple matter to employ our ansatz directly in the full string theory version of the Sakai-Sugimoto model, with the usual caveat that a prescription must be employed to deal with the non-abelian Dirac-Born-Infeld action. We expect the same kind of behaviour as in the five-dimensional Yang-Mills-Chern-Simons version of the theory. # Appendix {#appendix .unnumbered} In this appendix we consider the Skyrme model and explain how the homogeneous approximation may be thought of as a smeared version of the Skyrme crystal. In particular, we show that the homogeneous approximation is an unattainable idealization that provides a lower bound on the energy of the true Skyrme crystal. In terms of the hermitian currents \(R_i=i(\partial_i U)U^{-1}\) the static energy of the Skyrme model (in Skyrme units) is \[E=\int\bigg( \frac{1}{2}\mbox{Tr}(R_iR_i)-\frac{1}{16}\mbox{Tr}([R_i,R_j][R_i,R_j])\bigg)\,d^3x \label{skyrmeenergy}\] and the baryon number is \[B=\frac{i}{24\pi^2}\int \epsilon_{ijk} \mbox{Tr}(R_iR_jR_k) \,d^3x. \label{skyrmebaryon}\] The Faddeev-Bogomolny bound is \(E\ge 12\pi^2 B\), but for non-zero \(B\) this bound cannot be attained, as it requires that the Skyrme field \(U({\bf x})\) is an isometry from \(\mathbb{R}^3\) to \(SU(2)\), which are two spaces that are not isometric. The solution of the Skyrme model that is closest to the bound is the triply periodic Skyrme crystal with \(E/B=12\pi^2\times 1.04.\) This is a cubic lattice that contains four Skyrmions within a cube of side length \(L=4.7\), and hence has a baryon number density \({\cal B}=4/L^3=0.04.\) Substituting our idealized homogeneous approximation \(R_i=-\beta\sigma_i/2\) into ([\[skyrmeenergy\]](#skyrmeenergy){reference-type="ref" reference="skyrmeenergy"}) and ([\[skyrmebaryon\]](#skyrmebaryon){reference-type="ref" reference="skyrmebaryon"}) gives the energy per unit volume \({\cal E}\) and the baryon number density \({\cal B}\) to be \[{\cal E}=\frac{3}{4}\beta^2+\frac{3}{16}\beta^4, \qquad\qquad {\cal B}=\frac{\beta^3}{16\pi^2}.\] Using these expressions, we recover the energy bound by the simple manipulation \[{\cal E}=\frac{3}{4}\bigg(\beta-\frac{1}{2}\beta^2\bigg)^2+\frac{3}{4}\beta^3\ge \frac{3}{4}\beta^3=12\pi^2{\cal B}.\] Within the homogeneous approximation the bound is attained by \(\beta=2\) with a corresponding baryon number density \({\cal B}=1/(2\pi^2)=0.05.\) These values provide a good estimate of both the energy and the baryon number density of the Skyrme crystal. The homogeneous approximation generates a lower bound for the true energy of the Skyrme crystal because there are no Skyrme fields that generate the idealized homogeneous currents required to attain the bound.
{'timestamp': '2016-07-19T02:06:27', 'yymm': '1607', 'arxiv_id': '1607.04832', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04832'}
null
null
# Introduction {#int} The study of hypernuclei has been attracting great interest of nuclear physics community in providing the information from nucleon-nucleon (NN) interaction to hyperon-nucleon (YN) and hyperon-hyperon (YY) interactions. Due to the injection of hyperons a new dimension is added in normal nuclear system and hyperons serve as a potential probe for exploring many nuclear properties in domain of strangeness . However, hyperon-nucleon interaction is weaker than nucleon-nucleon but it is imperative as well as important to describe the nuclear many-body system with strangeness. Various theoretical approaches Skyrme Hartree Fock , relativistic mean field , cluster, variational, diffusion Monte Carlo , and G-matrix  have been employed by scientific community to estimate the strength of hyperon-nucleon as well hyperon-hyperon interactions. Further, these models have established themselves as very effective in testing the existence of bound hypernuclei and the stability of nucleonic core against hyperon(s) addition or the occurrence of exotic strange matter which facilitate the path toward multi-strange systems. Magic numbers in nuclear physics are certain neutron and proton numbers in atomic nuclei, in which higher stability in the ground state is observed than in the neighbouring nuclides and are most abundant in nature. The various experimental signatures that show discontinuity at magic numbers are: the energy required for the separation of one and two nucleons, the energies of alpha and beta transitions, pairing energy and the excitation of low-lying vibrational . The separation energy is sensitive to the collective or single particle inter play and provides a sufficient information about the nuclear structure effects. The discovery of magic numbers paved the way to great progress in understanding of nuclear structure and these numbers became the cornerstones for future theoretical developments in nuclear physics. It is worthy to mention that the several signatures are seen for the evolution of the magic gaps along the nuclear chart including superheavy region . The quest for proton or neutron magic numbers in the elusive mass region of superheavy nuclei is of utmost importance as the mere existence of superheavy nuclei is the result of the interplay between the attractive nuclear force (shell effects) and the large disruptive coulomb repulsion between the protons that favours the fission . It is well established that 2, 8, 20, 28, 50, 82 and 126 are the nucleonic magic number. In addition to this, Z = 120 and N = 172, 184 are predicted to be next magic number by various theoretical models in superheavy mass region . These predictions have been made on the basis of separation energy, shell gaps, pairing energy and shell correction energy etc. It may therefore relevant to extend the line of thought to the hypernuclear chart. It is well known that the spin-orbit interaction in \(\Lambda\) channel is weaker than nucleonic sector and thus the \(\Lambda\) magic numbers are expected to be close to the harmonic oscillator ones: 2, 8, 20, 40 and 70. In this paper, our main motive is to make an extensive investigation to search the \(\Lambda\) magic number within the RMF approach and to obtain the stability of triply magic system with doubly magic core. The magic numbers in nuclei are characterized by a large shell gap in single-particle energy levels. This means that the nucleon in the lower level has a comparatively large value of energy than that on higher level giving rise to more stability. The extra stability corresponding to certain number can be estimated from the sudden fall in the separation energy. The \(\Lambda\) separation energy is considered to be one of the key quantity to reveal the nuclear response to the addition of lambda hyperon. Therefore, in present work, we obtain the the binding energy per particle and one-lambda as well as two-lambda separation energies for considered multi-hypernuclei. Moreover, two-lambda shell gaps is also calculated to make the clear presentation of the magic number which also support to the two-lambda separation energy. To mark the \(\Lambda\) shell gaps, single-particle energy levels are analyzed that may correspond to \(\Lambda\) magic number. In addition, to analyze the structural distribution as well as impact of \(\Lambda\) hyperon on bubble structure for considered nuclei, total (nucleon plus \(\Lambda\)) density is reported. Nucleon and lambda mean field and spin-orbit interaction potentials are also observed. On the basis of binding energy per particle, the stability of triply magic hypernuclei is reported. RMF theory has been quite successful for studying the infinite nuclear systems and finite nuclei including the superheavy mass region . It is quite successful to study the equation of state for infinite nuclear matter as well as pure neutron matter, where the existence of strange baryons is expected . In this context, addition of strangeness degree of freedom to RMF formalism is obvious for the suitable expansion of the model and such type of attempts have already been made . RMF explains not only the structural properties of singly strange hypernuclei but also provides the details study of multi-strange systems containing several \(\Lambda\)'s, \(\Sigma\)'s or \(\Xi\)'s. In fact, RMF explains spin-orbit interaction very nicely in normal nuclei as well as hypernuclei. The contribution of spin-orbit interaction is very crucial in emerging the magic number in nucleonic sector and the same is expected in strange sector. The paper is organized as follows: A brief introduction on hypernuclei and magic number is given in Section [1](#int){reference-type="ref" reference="int"}. Section [2](#form){reference-type="ref" reference="form"} gives a brief description of relativistic mean field formalism for hypernuclei with inclusion of \(\Lambda N\) and \(\Lambda \Lambda\) interactions. The results are presented and discussed in Section [3](#resu){reference-type="ref" reference="resu"}. The paper is summarized in Section [4](#sum){reference-type="ref" reference="sum"}. # Formalism {#form} Relativistic mean field theory has been applied succesfully to study the structural properties of normal nuclei as well as hypernuclei . The suitable expansion to hypernuclei has been made by including the lambda-baryon interaction Lagrangian with effective \(\Lambda\)N potential. The total Lagrangian density for single-\(\Lambda\) hypernuclei has been given in many Refs. . For dealing the multi-lambda hypernuclei in quantitative way, the additional strange scalar (\(\sigma^*\)) and vector (\(\phi\)) mesons have been included which simulate the \(\Lambda \Lambda\) interaction  . Now, the total Lagrangian density can be written as \[\begin{aligned} \mathcal{L}&=&\mathcal{L}_N+\mathcal{L}_\Lambda+\mathcal{L}_{\Lambda\Lambda} \;, \end{aligned}\] \[\begin{aligned} {\cal L}_N&=&\bar{\psi_{i}}\{i\gamma^{\mu} \partial_{\mu}-M\}\psi_{i} +{\frac12}(\partial^{\mu}\sigma\partial_{\mu}\sigma-m_{\sigma}^{2}\sigma^{2}) -{\frac13}g_{2}\sigma^{3} \nonumber \\ &&-{\frac14}g_{3}\sigma^{4}-g_{s}\bar{\psi_{i}}\psi_{i}\sigma -{\frac14}\Omega^{\mu\nu}\Omega_{\mu\nu} +{\frac12}m_{\omega}^{2}\omega^{\mu}\omega_{\mu} \nonumber \\ &&-g_{\omega }\bar\psi_{i}\gamma^{\mu}\psi_{i}\omega_{\mu} -{\frac14}B^{\mu\nu}B_{\mu\nu} +{\frac12}m_{\rho}^{2}{\vec{\rho}^{\mu}}{\vec{\rho}_{\mu}} -{\frac14}F^{\mu\nu}F_{\mu\nu} \nonumber \\ &&-g_{\rho}\bar\psi_{i}\gamma^{\mu}\vec{\tau}\psi_{i}\vec{\rho^{\mu}} -e\bar\psi_{i}\gamma^{\mu}\frac{\left(1-\tau_{3i}\right)}{2}\psi_{i}A_{\mu}\;, \nonumber \\ \mathcal{L}_{\Lambda}&=&\bar\psi_\Lambda\{i\gamma^\mu\partial_\mu-m_\Lambda\}\psi_\Lambda-g_{\sigma\Lambda}\bar\psi_\Lambda\psi_\Lambda\sigma -g_{\omega\Lambda}\bar\psi_\Lambda\gamma^{\mu}\psi_\Lambda \omega_\mu \;, \nonumber \\ \mathcal{L}_{\Lambda\Lambda}&=&{\frac12}(\partial^{\mu}\sigma^*\partial_{\mu}\sigma^* -m_{\sigma^*}^{2}\sigma^{*{2}})-{\frac14}S^{\mu\nu}S_{\mu\nu} +{\frac12}m_{\phi}^{2}\phi^{\mu}\phi_{\mu} \nonumber \\ &&-g_{\sigma^*\Lambda}\bar\psi_\Lambda\psi_\Lambda\sigma^*-g_{\phi\Lambda}\bar\psi_\Lambda\gamma^\mu\psi_\Lambda\phi_\mu \;, \end{aligned}\] where \(\psi\) and \(\psi_\Lambda\) denote the Dirac spinors for nucleon and \(\Lambda\)-hyperon, whose masses are M and \(m_\Lambda\), respectively. Because of zero isospin, the \(\Lambda\)-hyperon does not couple to \({\rho}\)-mesons. The quantities \(m_{\sigma}\), \(m_{\omega}\), \(m_{\rho}\), \(m_{\sigma^*}\), \(m_{\phi}\) are the masses of \(\sigma\), \(\omega\), \(\rho\), \(\sigma^*\), \(\phi\) mesons and \(g_s\), \(g_{\omega}\), \(g_{\rho}\), \(g_{\sigma\Lambda}\), \(g_{\omega\Lambda}\), \(g_{\sigma^*\Lambda}\), \(g_{\phi\Lambda}\) are their coupling constants, respectively. The nonlinear self-interaction coupling of \({\sigma}\) mesons is denoted by \(g_2\) and \(g_3\). The total energy of the system is given by \(E_{total} = E_{part}(N,\Lambda)+E_{\sigma}+E_{\omega}+E_{\rho} +E_{\sigma^*}+E_{\phi}+E_{c}+E_{pair}+E_{c.m.},\) where \(E_{part}(N,\Lambda)\) is the sum of the single-particle energies of the nucleons (N) and hyperon (\(\Lambda\)). The energies parts \(E_{\sigma}\), \(E_{\omega}\), \(E_{\rho}\), \(E_{\sigma^*}\), \(E_{\phi}\), \(E_{c}\), \(E_{pair}\) and \(E_{cm}\) are the contributions of meson fields, Coulomb field, pairing energy and the center-of-mass energy, respectively. In present work, for meson-baryon coupling constant, NL3\* parameter set is used through out the calculations . To find the numerical values of used \(\Lambda-\)meson coupling constants, we adopt the nucleon coupling to hyperon couplings ratio defined as; \(R_\sigma=g_{\sigma\Lambda}/g_s\), \(R_\omega=g_{\omega\Lambda}/g_\omega\), \(R_{\sigma^*}=g_{\sigma^*\Lambda}/g_s\) and \(R_\phi=g_{\phi\Lambda}/g_\omega\). The relative coupling values are used as \(R_\omega=2/3\), \(R_\phi=-\sqrt{2}/3\), \(R_\sigma=0.621\) and \(R_{\sigma^*}=0.69\) . In present calculations, we use the constant gap BCS approximation to include the pairing interaction and the centre of mass correction is included by \(E_{cm}=-(3/4)41A^{-1/3}\). # RESULTS AND DISCUSSIONS {#resu} Before taking a detour on searching the \(\Lambda\) magic behaviour in multi-\(\Lambda\) hypernuclei, first we see the effects of introduced \(\Lambda\) hyperon on normal nuclear core; how the binding energy and radii of normal nuclear system is affected by addition of \(\Lambda\)'s ? To analyze this, we consider a list of normal nuclei covering a range from light to superheavy mass region i.e. \(^{16}O\) to \(^{378}120\). Total binding energy (BE), binding energy per particle (BE/A), lambda binding energy (\(B_\Lambda\)) for \(s-\) and \(p-\) state and radii for considered core nuclei and corresponding hypernuclei are tabulated in Table [\[tab1\]](#tab1){reference-type="ref" reference="tab1"}. The calculated \(B_\Lambda\) is compared with available experimental data and we found a close agreement between them. This means the used parameter set is quite efficient to reproduce the experimental binding energy and ofcourse we can use it to make more calculations related to magicity in hypernuclei. Since, we are dealing with closed shell hypernuclei and hence our RMF calculations is restricted to spherical symmetric. The addition of \(\Lambda\) hyperon to normal nuclei enhances the binding and shrinks the core of the system. This is because of glue like role of \(\Lambda\) hyperon that residing on the \(s-\)state for most of the time. These observations are shown in Table [\[tab1\]](#tab1){reference-type="ref" reference="tab1"}. Binding energy of hypernuclei are larger than their normal counter parts and a reduction in total radius (\(r_{total}\)) of hypernuclei is observed, that means the \(\Lambda\) particle makes the core compact with increasing binding. For example, the total radius of \(^{16}O\) and \(^{209}Pb\) is 2.541 and 5.624 fm, which reduce to 2.536 and 5.616 fm by addition of single \(\Lambda\) into the core of \(O\) and \(Pb\), respectively. Moreover, for the sake of comparison with experimental data, binding energy and radii of the hypernuclei produced by replacing the neutrons means having a constant baryon number are also framed in Table [\[tab1\]](#tab1){reference-type="ref" reference="tab1"} and the shrinkage effect is also noticed. This results show that an important impact of \(\Lambda\) hyperon on binding as well as size of the system. The increasing value of \(B_\Lambda\) for \(s-\)state from light to superheavy hypernuclei confirming the potential depth of \(\Lambda-\)particle in nuclear matter which would be-28 MeV  . [\[tab1\]]{#tab1 label="tab1"} A peak of two-lambda shell gaps indicates the drastic change of the two-lambda separation energies; which is used as one of the significant signature of magic number. The two-lambda shell gaps \(\delta_{2\Lambda}\), for all considered hypernuclei as a function of added \(\Lambda\) hyperons are shown in Fig [\[s-gaps\]](#s-gaps){reference-type="ref" reference="s-gaps"}. A peak at certain \(\Lambda\) number suggests the existence of lambda shell closure. However, the quality of magic number is represented by sharpness as well as magnitude of the peak. Figure. [\[s-gaps\]](#s-gaps){reference-type="ref" reference="s-gaps"} reveals that the magnitude of the peak is found to be largest at \(\Lambda=\) 2, 8, 20, 40 indicating the strong shell closures. Further, the peaks appeared at \(\Lambda=\) 14, 18, 28, 34, 50, 58, 70 and 82 indicate the respective lambda magic number. Moreover, a peak with a very small magnitude is also appeared at \(\Lambda=\) 68 due to closureness of subshell (2\(d_{3/2}\)) revealing \(\Lambda\) semi-magic number. A peak with small magnitude is seen at \(\Lambda\)= 28 representing a feeble lambda magic number, contrary to nucleonic magic number. Pronounced peak is appeared at \(\Lambda=\) 34 and 58 indicating a strong \(\Lambda\) closed shells. ## Density profile and bubble structure A hypernucleus is a composed system of nucleons and hyperons and hence the gross structure of hypernucleus can be described by density distribution of nucleons as well as hyperons. It is well known and has mentioned earlier that the addition of a \(\Lambda\) hyperon makes the nuclear core compact with increasing binding as well as density. Therefore, it is important to study the effects of large number of added \(\Lambda\) hyperons on the nuclear density. Due to addition of hyperons, the magnitude of total density increases with increasing number of \(\Lambda\)'s as shown in Fig. [\[density\]](#density){reference-type="ref" reference="density"}. On view the density profile, one can examine the most interesting feature of nuclei i.e. bubble structure, which measure the depression of central density and has already been observed in light to superheavy mass region . It is to be noticed that several factors, including pairing correlations  , tensor force  and dynamic shape fluctuations  turn out to have influence on depression of central density. The exotic structures like bubble and halo have been recently studied in \(\Lambda-\)hypernuclei . Owing to weaker \(\Lambda \Lambda\) attraction compared to the nucleon-nucleon one the lambda hyperons are more diffuse in a nucleus than nucleons and thus generating a hyperon density about \(1/3\) smaller than the nucleonic density. Thus, it becomes quite important to look for the effect of large number of hyperons on neutron and proton density distributions. Since, there is no change of nucleon number and hence no anomalous effect of introduced \(\Lambda\)'s on neutron, proton densities is observed, individually. But the total density of the system is largely affected due to increasing number of \(\Lambda\)'s into the core. In considered multi-hypernuclei, the nucleonic core of some of them shows the depression of central density for example \(^{16}O\), \(^{90}Zr\), \(^{292}120\), \(^{304}120\) and \(^{378}120\) as predicted earlier also . It is found that the injected \(\Lambda\)'s reduce the depression of central density. For example, the depression of central part in \(^{16}O\) is reduced by injection of 2 \(\Lambda\)'s and further more by 8 \(\Lambda\)'s. The \(\Lambda\) particle attracts the nucleons towards the centre enhancing the central density and as a result remove the bubble structure partially or fully as reflected in Fig. [\[density\]](#density){reference-type="ref" reference="density"}. Therefore, it is one of the important implication of \(\Lambda\) particle to the nuclear system. Beyond the bubble structure, no anomalous behaviour of total density (core + \(\Lambda\)) in triply magic system is reported. ## Spin-orbit interaction and mean field potentials The spin-orbit interaction plays a significant role in reproducing the results quantitatively. It is the beauty of RMF in which the spin-orbit splitting is built-in naturally with exchange of scalar and vector mesons and thus describe a nuclear fine structure. It is not limited only to nuclei or superheavy nuclei but appears in hypernuclei also, however the strength of interaction is weaker than normal nuclei . It is clearly seen from Figs. [\[potentials\]](#potentials){reference-type="ref" reference="potentials"} that the spin-orbit potential for lambda hyperon is weaker than their normal counter parts and our results are consistent with theoretical predictions and experimental measurements . Here, nucleon (\(V^N_{so}\)) and lambda (\(V^\Lambda_{so}\)) spin-orbit interaction potentials are calculated for considered triply magic multi-hypernuclei. It is also conclude that the addition of \(\Lambda\)'s affects the nucleon as well as \(\Lambda\) spin-orbit potential to a great extent. The nucleon (\(V_N\)) and lambda (\(V_\Lambda\)) mean field potentials are also investigated and plotted as a function of radial parameter shown in Fig. [\[meanpot\]](#meanpot){reference-type="ref" reference="meanpot"}. The total depth of \(\Lambda\) mean field potential is found to be around-30 MeV, which is in agreement with existing experimental data . It is to mention that the additions of \(\Lambda\)'s affects the depth of both nucleon as well as lambda mean potentials. The nucleonic potential depth in multi-lambda hypernuclei is approx to-80 MeV to-90 MeV. The shape of lambda potential looks like to be same as nucleonic potential and only the amount of depth is different. It is also to be noticed that the nucleonic potential looks like as V-shape type and shows the maximum depth around-90 MeV at r = 4 fm, while this amount of depth reaches to-70 MeV at r = 0 fm for \(^{292}\)`<!-- -->`{=html}120. It indicates a relatively low concentration of the particles at central region (r = 0) which is a direct consequence of depression of central density so-called bubble structure. ## Single-particle energies Any kinds of change in a system can be observed from their single-particle energy levels. To analyze the impact of \(\Lambda\) hyperon on nucleon single-particle energy levels, the filled neutron and proton levels for Ca hypernuclei are plotted as a function of added hyperons as shown in Fig. [\[levels-ca\]](#levels-ca){reference-type="ref" reference="levels-ca"}. Figure [\[levels-ca\]](#levels-ca){reference-type="ref" reference="levels-ca"} reveals that the neutron and proton energy levels goes dipper with addition of \(\Lambda\)'s as a result increase the stability of the system. The added hyperons increase the nucleon separation energy and as a result form a more bound system with increasing binding then their normal counter parts, which also leads to an extension of drip-line . For example, neutron \(s_{1/2}\)(n) level has an energy of about-54.037 MeV for the core of Ca hypernucleus, while this amount reaches to-62.499 MeV for \(^{48+18\Lambda}Ca\) system with 18 \(\Lambda\)'s. Also, a same trend is observed for proton levels where, \(s_{1/2}\)(p) has an energy 51.787 MeV for the core of Ca hypernuclei and this value reaches to-60.0477 MeV with addition of 18 \(\Lambda\)'s. This results show that the \(\Lambda\) hyperons draw the nuclear system towards more stability with increasing strangeness. Moreover, the same trend of neutron and proton energy levels is observed for other multi hypernuclei where both the levels would go dipper with increasing number of \(\Lambda\) hyperon to nucleonic core but we do not make a plot for the same. A inversion of proton levels is seen, where \(d_{3/2}\) fill faster than \(s_{1/2}\) and this type of filling is also observed in lambda levels. Further, we analyze the lambda single-particle energy levels for \(^{48+n\Lambda}Ca\), \(^{208+n\Lambda}Pb\) and \(^{304+n\Lambda}120\) hypernuclei to extract the lambda shell gaps for confirming the \(\Lambda\) magic number. The lambda energy levels as function of added \(\Lambda\)'s are given in Fig. [\[levels-ca\]](#levels-ca){reference-type="ref" reference="levels-ca"}, [\[levels-pb\]](#levels-pb){reference-type="ref" reference="levels-pb"}, [\[levels-120\]](#levels-120){reference-type="ref" reference="levels-120"}. The filling of \(\Lambda\)'s is same as the nucleons following the shell model scheme with lambda spin-orbit interaction potential. It is observed that the single-particle gap of spin-orbit splitting in lambda levels is smaller than the nucleons due to weakening strength of lambda spin-orbit interaction. By analyzing the lambda single-particle energy levels of Ca hypernuclei it is found that large energy gap exist in \(1d_{5/2}\) to \(1d_{3/2}\) or \(2s_{1/2}\) and that's why lambda magic number 14 is emerged. Further, \(2s_{1/2}\) and \(1d_{3/2}\) are very much close to each other due to weaker strength of \(\Lambda\) spin-orbit interaction. However, \(\Lambda=\) 20 is clearly seen due to large energy gap by \(f_{7/2}\) to lower orbital. In case of Pb, the large shell gaps for \(\Lambda =\) 2, 8, 18, 20, 28, 34, 40, 50, 58, 70 and 82 is appeared. However, the single-particle gap for lambda number 28 is not so strong as compared to others suggesting the feeble magic number. The inversion of normal level scheme is noticed and the higher levels fill faster than lower one and hence this types of filling is responsible to emerge the new more magic number. For example, the filling of \(1d_{3/2}\) before \(2s_{1/2}\) shows a shell gap at \(\Lambda=\) 18. Along the similar line, due to inversion between {\(1f_{5/2}\), \(2p_{3/2}\)} and {\(1g_{7/2}\), \(2d_{5/2}\)} the \(\Lambda\) closed shells 34 and 58 is observed, respectively. In case of superheavy multihypernuclei, large single-particle shell gaps are appeared for lambda number 2, 8, 18, 20, 34, 40, 50, 58, 70 and 82. It is quite worth to notice that pronounced energy gaps is noticed in \(^{208}Pb\) and \(^{304}120\) at \(\Lambda=\) 34, 58 are being suggested to be strong \(\Lambda\) shell closure. The sharp peaks observed in \(\delta_{2\Lambda}\) at \(\Lambda=\) 2, 8, 20, 34, 40 and 58 is clearly reflected from lambda single-particle energies, where a large energy gap is exist for the filling of these number of lambda hyperons. ::: [\[tab2\]]{#tab2 label="tab2"} ## Magicity Various signatures of the evolution of magic shell gaps have been discovered across the nuclear landscape during the past few decades  such as (i) A large binding energy then neighbouring nuclides, (ii) Sudden fall at separation energy, (iii) A large shell gap, etc. It becomes therefore quite relevant to extend the prediction of magic numbers to the hypernuclear chart. Looking for the magic behaviour firstly, we emphasize on binding of some selected nuclei whose nucleonic core is doubly magic such as \(O\), \(Ca\), \(Ni\), \(Zr\), \(Sn\), \(Pb\) and \(120\). Some certain number of \(\Lambda\) hyperon binds the nuclear core with maximum stability that may correspond to \(\Lambda\) magic number in hypernuclei and might form a triple magic system with doubly magic nuclear core as initially discussed in Ref.  and recently in Ref. . For example, \(\Lambda=\) 2, produce a maximum binding for \(^{16}O\) and \(^{48}Ca\) reveal the maximum binding with \(\Lambda=\) 8. Also, \(^{378}120\) shows a peak binding on addition of 90 \(\Lambda\)'s. In this way, we extract some lambda number which are 2, 8, 18, 20, 40, 70, 90 produce a maximum stability for their particular system and suppose to be \(\Lambda\) magic number as much as close to harmonic oscillator number. But several other strong signatures are exist to identify the magicity and we make the analysis in this direction to look out the correct \(\Lambda\) magic number. After analyzing the \(S_\Lambda\) and \(S_{2\Lambda}\), for considered light to superheavy mass multi-hypernuclei it is noticed that a sudden fall is observed at \(\Lambda=\) 2, 8, 14, 20, 28, 34, 40, 50, 58, 68, 70 and 82. And hence, the analysis suggest that these numbers are supposed to be \(\Lambda\) magic number in multi-lambda hypernuclei and form a triply magic system having doubly magic nucleonic core. In order to identify the \(\Lambda\) magic number strongly, two-lambda shell gaps is examined which provide more strong signature of magicity and favoured \(S_{2\Lambda}\). Pronounced peak in two-lambda shell gap is observed at \(\Lambda=\) 2, 8, 20 and 40 indicating a strong shell closure. The peaks observed with a significant magnitude at \(\Lambda=\) 14, 18, 34, 50, 58, 70 and 82 indicating a shell closure also. Further to testify, we look for the lambda shell gaps by examining the single-particle energy levels. We noticed that a large single-particle gap is appeared in \(^{208}Pb\) hypernucleus at \(\Lambda=\) 2, 8, 18, 20, 28, 34, 40, 50, 58 and 70 confirming the lambda magic number. The analysis of single-particle energy levels for \(^{378}120\) multi hypernuclei clear the results by showing a large energy gap in 2, 8, 18, 20, 34, 40, 50, 58, 70 and 82. It is to mention that a significant shell gap is observed for 34 and 58 suggesting a strong \(\Lambda\) shell closure. The inversion of normal level scheme is responsible to emerge the \(\Lambda\) magic number 34 and 58. The experimentally confirmation of nucleonic shell closure of 34 supports our predictions . The nucleonic number 14, 16, 18 and 32 have also been in discussion and expected to be shell closure . In addition, nucleon number 16 and 32 have also been experimentally confirmed in exotic nuclei as a new neutron magic number . Therefore, it is concluded that the \(\Lambda\) magicity quite resembles with the nuclear magicity and it is expected that our predictions might be used as significant input to make the things clear regarding new sub shell closure. The predicted \(\Lambda\) magic number in multi-hypernuclei are framed in Table [1](#tab2){reference-type="ref" reference="tab2"}. The present lambda magic numbers are quite agreeable with the prediction of Ref.  and Ref.  where Bruckner-Hartee Fock calculations using the lambda density functional have been made. It is clear from the plot of \(\delta_{2\Lambda}\), that 34, 58 and 70 have peaks of great magnitude, while 68 in a feeble magnitude and suppose to be subshell closure. Moreover, strong nucleonic magic number 28 is observed very feeble in lambda magicity. It is to mention that the lambda number 14 is appeared in medium mass hypernuclei, on contrary to this 18 is observed in superheavy mass multi hypernuclei. # Summary and conclusion {#sum} In summary, we have suggested the possible \(\Lambda\) magic number i.e. 2, 8, 14, 18, 20, 28, 34, 40, 50, 58, 68, 70, 82 in multi-\(\Lambda\) hypernuclei within the relativistic mean field theory with effective \(\Lambda\)N as well as \(\Lambda\)\(\Lambda\) interactions. The survey of \(\Lambda\) magic number is made on the basis of binding energy, one-and two-lambda separation energies \(S_\Lambda\), \(S_{2\Lambda}\), and two-lambda shell gaps \(\delta_{2\Lambda}\). It is noticed that pronounced single particle energy gap is observed for lambda number 34 and 58 in Pb and superheavy multi hypernuclei representing the strong \(\Lambda\) magic number. Our predictions are strongly supported by nuclear magicity, where nucleon number 34 is experimentally confirmed as a neutron shell closure . It is expected that the weakening strength of lambda spin-orbit interaction potential is responsible for emerging the new lambda magic number. The predicted \(\Lambda\) magic numbers are in remarkable agreement with earlier predictions  and hypernucler magicity quite resembles with nuclear magicity. In the analogy of nuclear stability, we noticed a similar pattern of binding energy per particle in hypernuclear regime and Ni hypernucleus with 8 \(\Lambda\)'s is found to be most tightly bound triply magic system in hypernuclear landscape. The addition of \(\Lambda\) hyperons have significant impact on nucleon distribution and remove the bubble structure partially or fully. The spin-orbit interaction potentials and mean field potentials is also studied for predicted triply magic hypernuclear systems and the added \(\Lambda\)'s affect both the potential to a large extent. The present results may be used as a significant input to produce the triply magic hypernuclei in laboratory in future. It is also concluded that the addition of \(\Lambda\) hyperons draw the nuclear system towards more stability with increasing strangeness. We noticed that the core of superheavy nuclei has more affinity to absorb large number of hyperons. This means such systems are able to simulate the strange hadronic matter containing large number of heavy hyperons such as \(\Sigma\)'s and \(\Xi\)'s including several \(\Lambda\)'s and the formation of such systems has large implication in nuclear-astrophysics.
{'timestamp': '2016-07-19T02:07:17', 'yymm': '1607', 'arxiv_id': '1607.04865', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04865'}
# Introduction In the last decade, many studies have permitted a deeper understanding of the relationship between gravity and gauge theories from the point of view of scattering amplitudes (see for a comprehensive review). One of the most interesting results is color-kinematics duality, which allows the construction of gravity amplitudes by replacing color factors by a second copy of the kinematic numerators. This double copy structure has a historic antecedent in the Kawai, Lewellen, and Tye (KLT) relations, showing how, at tree level, closed string amplitudes admit a decomposition in terms of products of open string amplitudes. Similar structures have been found in various other setups. It seems clear that, at the level of scattering amplitudes, there is a sense in which gravity can be considered the "square" of a gauge theory. Given this double copy structure, a natural question to ask is how certain properties of gauge theory amplitudes translate into the gravitational side. One of these is the existence of radiation zeros. This is a peculiar feature of certain scattering processes where one or more massless gauge bosons are radiated, consisting in the vanishing of the amplitude for certain phase space configurations. The phenomenon was first identified in processes involving gauge bosons trilinear couplings, in particular \(u\bar{d}\rightarrow W^{+}\gamma\). It has been experimentally observed both at the Tevatron and LHC. Their existence has been also studied in graviton photoproduction. These so-called Type-I zeros appear for momentum configurations satisfying the constraint \(Q_{i}=\kappa \, p_{i}\cdot k\), where \(k\) is the momentum of the gauge boson, \(Q_{i}\) and \(p_{i}\) are the charge and momenta of the other particles, and \(\kappa\) is a numerical constant. In Ref. it was realized that zeros in the amplitude may also occur when the spatial momenta of the particles involved in the process lie on the same plane. These Type-II or planar zeros have been identified in the processes \(e^{+}e^{-}\rightarrow W^{+}W^{-}\gamma\) and \(e^{+}e^{-}\rightarrow \tau\bar{\tau}\gamma\), in both cases in the soft photon limit. So far, the only study of planar zeros beyond the soft limit has been carried out in the interesting work , where the five parton amplitude in QCD was analyzed. Using the maximally helicity violating (MHV) formalism, planar zeros were found both for the \(gg\rightarrow ggg\) and \(q\bar{q}\rightarrow ggg\) processes. In the case of the five-gluon amplitude for general color factors the planar zero condition depends on the color quantum numbers of incoming and outgoing gluons. The present article has a double aim. One is to study the conditions for the emergence of planar zeros in the five-gluon amplitude. We show that planar zeros are a "projective\" property of the amplitude, in the sense that they are preserved by a simultaneous rescaling of the stereographic coordinates labeling the flight directions of the three outgoing gluons. In terms of stereographic coordinates, we find that the existence of planar zeros is determined by a cubic algebraic curve whose integer coefficients are given in terms of the color factors. In the case of SU(\(N\)) gauge groups, we find that the casuistic of curves obtained for different color configurations gets broader as the rank \(N\) increases, starting with the case of SU(2) where no physical zeros are found for external particles with well-defined color quantum numbers. Our second target consists in exploiting color-kinematics duality to study planar zeros in the gravitational case, where we find that the five-graviton amplitude vanishes whenever the process is planar. This can be understood applying the BCJ prescription to the equation determining the zeros in the gauge case. By replacing color factors with kinematic numerators satisfying color-kinematics duality, the condition for the planar zero is seen to be identically satisfied without further kinematic constraints. The plan of the paper is as follows. In Section [2](#sec:5-gluon){reference-type="ref" reference="sec:5-gluon"} we review the calculation of the five gluon amplitude using the MHV formalism. Section [3](#sec:gluon_planar){reference-type="ref" reference="sec:gluon_planar"} is devoted to the conditions for planar zeros in the gauge case, while in Section [4](#sec:permutations){reference-type="ref" reference="sec:permutations"} we study the transformation of the loci of planar zeros under permutations of the color labels of the external gluons. In Section [5](#sec:graviton){reference-type="ref" reference="sec:graviton"} the graviton amplitude is obtained using color-kinematics duality and the condition for the existence of amplitude zeros is obtained. Finally, in Section [6](#sec:conclusions){reference-type="ref" reference="sec:conclusions"} we summarize our conclusions. # The five-gluon amplitude {#sec:5-gluon} In this section we revisit the construction of the five-gluon amplitude \[\begin{aligned} g(p_{1},a_{1})+g(p_{2},a_{2})\longrightarrow g(p_{3},a_{3})+g(p_{4},a_{4})+g(p_{5},a_{5}), \end{aligned}\] where we take all momenta incoming. The tree level amplitude is computed in terms of 15 nonequivalent trivalent diagrams, leading to the expression \[\begin{aligned} \mathcal{A}_{5}&=g^{3}\left({c_{1}n_{1}\over s_{12}s_{45}}+{c_{2}n_{2}\over s_{23}s_{15}}+{c_{3}n_{3}\over s_{34}s_{12}} +{c_{4}n_{4}\over s_{45}s_{23}}+{c_{5}n_{5}\over s_{15}s_{34}}+{c_{6}n_{6}\over s_{14}s_{25}} +{c_{7}n_{7}\over s_{13}s_{25}}+{c_{8}n_{8}\over s_{24}s_{13}} \right. \nonumber \\[0.2cm] &+ \left.{c_{9}n_{9}\over s_{35}s_{24}}+{c_{10}n_{10}\over s_{14}s_{35}}+{c_{11}n_{11}\over s_{15}s_{24}} +{c_{12}n_{12}\over s_{12}s_{35}}+{c_{13}n_{13}\over s_{23}s_{14}}+{c_{14}n_{14}\over s_{25}s_{34}} +{c_{15}n_{15}\over s_{13}s_{45}}\right), \label{eq:amplitude_gluon_general} \end{aligned}\] where we have introduced the kinematic invariants \[\begin{aligned} s_{ij}=(p_{i}+p_{j})^{2}=2 \, p_{i}\cdot p_{j}, \hspace*{1cm} i<j. \end{aligned}\] The color factors in Eq. [\[eq:amplitude_gluon_general\]](#eq:amplitude_gluon_general){reference-type="eqref" reference="eq:amplitude_gluon_general"} are given by[^1] \[\begin{aligned} c_{1} &= f^{a_{1}a_{2}b} f^{ba_{3}c} f^{ca_{4}a_{5}}, \hspace*{1cm} c_{2}=f^{a_{1}a_{5}b} f^{ba_{4}c} f^{ca_{3}a_{2}}, \nonumber \\[0.2cm] c_{3} &= f^{a_{3}a_{4}b} f^{ba_{5}c} f^{ca_{1}a_{2}}, \hspace*{1cm} c_{4}=f^{a_{4}a_{5}b} f^{ba_{1}c} f^{ca_{2}a_{3}}, \nonumber \\[0.2cm] c_{5} &= f^{a_{5}a_{1}b} f^{ba_{2}c} f^{ca_{3}a_{4}}, \hspace*{1cm} c_{6}=f^{a_{1}a_{4}b} f^{ba_{3}c} f^{ca_{5}a_{2}}, \nonumber \\[0.2cm] c_{7} &= f^{a_{1}a_{3}b} f^{ba_{4}c} f^{ca_{5}a_{2}}, \hspace*{1cm} c_{8}=f^{a_{1}a_{3}b} f^{ba_{5}c} f^{ca_{4}a_{2}}, \label{eq:color_factors} \\[0.2cm] c_{9} &= f^{a_{3}a_{5}b} f^{ba_{1}c} f^{ca_{2}a_{4}}, \hspace*{1cm} c_{10}=f^{a_{4}a_{1}b} f^{ba_{2}c} f^{ca_{3}a_{5}}, \nonumber \\[0.2cm] c_{11} &= f^{a_{1}a_{5}b} f^{ba_{3}c} f^{ca_{4}a_{2}}, \hspace*{1cm} c_{12}=f^{a_3a_5b} f^{ba_4c} f^{ca_1a_2}, \nonumber \\[0.2cm] c_{13} &= f^{a_{1}a_{4}b} f^{ba_{5}c} f^{ca_{3}a_{2}}, \hspace*{1cm} c_{14}=f^{a_{5}a_{2}b} f^{ba_1c} f^{ca_{3}a_{4}}, \nonumber \\[0.2cm] c_{15} &= f^{a_{1}a_{3}b} f^{ba_{2}c} f^{ca_{4}a_{5}}, \nonumber \end{aligned}\] and satisfy nine independent Jacobi identities \[\begin{aligned} c_{3}-c_{5}+c_{14}&=0, \hspace*{1.15cm} c_{3}-c_{1}-c_{12}=0, \nonumber \\[0.2cm] c_{4}-c_{1}+c_{15}&=0, \hspace*{1.15cm} c_{4}+c_{2}-c_{13}=0, \nonumber \\[0.2cm] c_{5}+c_{2}-c_{11}&=0, \hspace*{1cm} c_{13}-c_{6}+c_{10}=0, \label{eq:jacobi_id}\\[0.2cm] c_{14}-c_{7}+c_{6}&=0, \hspace*{1.15cm} c_{7}-c_{8}+c_{15}=0, \nonumber \\[0.2cm] c_{8}-c_{9}-c_{11}&=0, \hspace*{0.85cm} (c_9-c_{10}+c_{12}=0). \nonumber \end{aligned}\] On general grounds, the amplitude can be written in terms of color-ordered amplitudes as \[\begin{aligned} \mathcal{A}_{5}=g^{3}\sum_{\sigma\in S_{4}}c[1,\sigma(2,3,4,5)]A_{5}[1,\sigma(2,3,4,5)], \end{aligned}\] where the sum is over noncyclic permutations of the external legs. However, the set of color-ordered amplitudes is over complete, a fact expressed by the Kleiss-Kuijf relations . In the case of the five-point amplitude, there are \(5\times 4\) different ways of choosing a basis in the space of independent color structures \(\mathsf{TCS}_{5}\). We select one of these basis by fixing the incoming gluons (see Fig. [\[fig:KK-rel\]](#fig:KK-rel){reference-type="ref" reference="fig:KK-rel"}), so the five-gluon amplitude in Eq. [\[eq:amplitude_gluon_general\]](#eq:amplitude_gluon_general){reference-type="eqref" reference="eq:amplitude_gluon_general"} can be re-expressed in terms of \(3!\) color ordered amplitudes according to \[\begin{aligned} \mathcal{A}_{5}=g^{3}\sum_{\sigma\in S_{3}}c[1,2,\sigma(3,4,5)]A_{5}[1,2,\sigma(3,4,5)], \label{eq:kleiss-kuijf_exp} \end{aligned}\] where the subamplitudes are explictly given in terms of the numerators \(n_{i}\) by \[\begin{aligned} A_{5}[1,2,3,4,5]&={n_{1}\over s_{12}s_{45}}-{n_{2}\over s_{23}s_{15}}+{n_{3}\over s_{34}s_{12}}+{n_{4}\over s_{45}s_{23}} +{n_{5}\over s_{15}s_{34}}, \nonumber \\[0.2cm] A_{5}[1,2,3,5,4]&=-{n_{1}\over s_{12}s_{45}}-{n_{13}\over s_{23}s_{14}}+{n_{12}\over s_{35}s_{12}}-{n_{4}\over s_{45}s_{23}} +{n_{10}\over s_{14}s_{35}}, \nonumber \\[0.2cm] A_{5}[1,2,4,3,5]&=-{n_{12}\over s_{12}s_{35}}-{n_{11}\over s_{24}s_{15}}-{n_{3}\over s_{34}s_{12}}+{n_{9}\over s_{35}s_{24}}-{n_{5}\over s_{15}s_{34}}, \\[0.2cm] A_{5}[1,2,4,5,3]&={n_{12}\over s_{12}s_{35}}-{n_{8}\over s_{24}s_{13}}-{n_{1}\over s_{45}s_{12}}-{n_{9}\over s_{35}s_{24}}-{n_{15}\over s_{13}s_{45}}, \nonumber \\[0.2cm] A_{5}[1,2,5,3,4]&=-{n_{3}\over s_{12}s_{34}}-{n_{6}\over s_{25}s_{14}}-{n_{12}\over s_{35}s_{12}}+{n_{14}\over s_{34}s_{25}}-{n_{10}\over s_{14}s_{35}}, \nonumber \\[0.2cm] A_{5}[1,2,5,4,3]&={n_{3}\over s_{12}s_{34}}-{n_{7}\over s_{25}s_{13}}+{n_{1}\over s_{12}s_{45}}-{n_{14}\over s_{34}s_{25}} +{n_{15}\over s_{13}s_{45}}. \nonumber \end{aligned}\] Going back to the expression for the color factors in Eq. [\[eq:color_factors\]](#eq:color_factors){reference-type="eqref" reference="eq:color_factors"}, these partial amplitudes are respectively associated with the six color factors \(c_{7}\), \(c_{8}\), \(c_{6}\), \(c_{13}\), \(c_{11}\), and \(c_{2}\). At this point we can exploit the generalized gauge freedom in the definition of the numerators to implement color-kinematics duality, so the numerators \(n_{i}\) mimic the Jacobi identities [\[eq:jacobi_id\]](#eq:jacobi_id){reference-type="eqref" reference="eq:jacobi_id"}. Solving the corresponding equations we can eliminate \(n_{7}\) to \(n_{15}\) finding the following solution for the numerators in terms of the basis of color-ordered amplitudes \[\begin{aligned} n_{1}&=-n_{12}=n_{15}=s_{12}s_{45}A_{5}[1,2,3,4,5], \nonumber \\[0.2cm] n_{2}&=n_{3}=n_{4}=n_{5}=n_{11}=n_{13}=n_{14}=0, \nonumber \\[0.2cm] n_{6}&=n_{7}=n_{10}=s_{14}s_{35}A_{5}[1,2,3,5,4]+s_{14}(s_{35}+s_{45})A_{5}[1,2,3,4,5], \\[0.2cm] n_{8}&=n_{9}=s_{14}s_{35}A_{5}[1,2,3,5,4]+(s_{14}s_{35}+s_{14}s_{45}+s_{12}s_{45})A_{5}[1,2,3,4,5]. \nonumber \end{aligned}\] Color-kinematics duality is independent of the polarization of the gluons. Here we are going to use the MHV formalism and assign negative helicity to the incoming gluons. Using the Parke-Taylor formula we have \[\begin{aligned} A_{5}[1^{-},2^{-},\sigma(3^{+},4^{+},5^{+})]=i{\langle 12\rangle^{4}\over \langle 12\rangle\langle 2\sigma(3)\rangle \langle \sigma(3)\sigma(4)\rangle \langle \sigma(4)\sigma(5)\rangle\langle \sigma(5)1\rangle}, \end{aligned}\] for any permutation \(\sigma\in S_{3}\) of the last three indices. Expressing in addition the kinematic invariants in terms of spinors, \(s_{ij}=\langle ij\rangle[ji]\), we arrive at the following expressions for the numerators \[\begin{aligned} n_{1}&=-n_{12}=n_{15}=i{\langle 12\rangle^{4}[21][54]\over \langle 23\rangle\langle 34\rangle \langle 51\rangle}, \nonumber \\[0.2cm] n_{6}&=n_{7}=n_{10}=i{\langle 12\rangle^{4}[14][52]\over \langle 23\rangle\langle 34\rangle \langle 51\rangle}, \nonumber \\[0.2cm] n_{8}&=n_{9}=i{\langle 12\rangle^{4}[24][51]\over \langle 23\rangle\langle 34\rangle \langle 51\rangle}, \hspace*{1cm} \label{eq:numerators} \\[0.2cm] n_{2}&=n_{3}=n_{4}=n_{5}=n_{11}=n_{13}=n_{14}=0. \nonumber \end{aligned}\] With this result, the five-gluon amplitude can be written as \[\begin{aligned} \mathcal{A}_{5}&=-ig^{3}\langle 12\rangle^{3}\left({c_{2}\over \langle 23\rangle\langle 34\rangle\langle 45\rangle\langle 51\rangle} +{c_{6}\over \langle 25\rangle\langle 53\rangle\langle 34\rangle\langle 41\rangle} +{c_{7}\over \langle 25\rangle\langle 54\rangle\langle 43\rangle\langle 31\rangle}\right. \nonumber \\[0.2cm] &\left.+{c_{8}\over \langle 24\rangle\langle 45\rangle\langle 53\rangle\langle 31\rangle} +{c_{11}\over \langle 24\rangle\langle 43\rangle\langle 35\rangle\langle 51\rangle} +{c_{13}\over \langle 23\rangle\langle 35\rangle\langle 54\rangle\langle 41\rangle}\right). \label{eq:5g_amplitude_spinors} \end{aligned}\] Alternatively, this expression can be obtained from Eq. [\[eq:kleiss-kuijf_exp\]](#eq:kleiss-kuijf_exp){reference-type="eqref" reference="eq:kleiss-kuijf_exp"} by a direct application of the Parke-Taylor formula. The spinor products appearing in the five-gluon amplitude can now be recast in terms of momenta. Working in the center-of-mass frame, the incoming momenta take the form \[\begin{aligned} p_{1}={\sqrt{s}\over 2}(1,0,0,1), \hspace*{1cm} p_{2}={\sqrt{s}\over 2}(1,0,0,-1). \end{aligned}\] On the other hand, for the three outgoing gluons their spatial momenta are parametrized using stereographic coordinates \(\zeta_{a}\in \mathbb{C}\) (with \(a=3,4,5\)) according to \[\begin{aligned} p_{a}=-\omega_{a}\left(1,{\zeta_{a}+\overline{\zeta}_{a}\over 1+\zeta_{a}\overline{\zeta}_{a}},i{\overline{\zeta}_{a}-\zeta_{a}\over 1+\zeta_{a}\overline{\zeta}_{a}},{\zeta_{a}\overline{\zeta}_{a}-1\over 1+\zeta_{a}\overline{\zeta}_{a}}\right), \end{aligned}\] where the global minus sign reflects that all momenta are taken entering the diagram. The stereographic coordinates \(\zeta_{a}\) are related to the rapidity \(y_{a}\) and the azimuthal angle \(\phi_{a}\) by \[\begin{aligned} \zeta_{a}=e^{y_{a}+i\phi_{a}}. \end{aligned}\] # Gauge planar zeros {#sec:gluon_planar} We focus now on planar five-gluon scattering with general color quantum numbers. Since the incoming particles travel along the \(z\) axis, without loss of generality we can take all momenta lying on the \(xz\)-plane. This means that \(p_{a}^{y}=0\) and therefore \(\zeta_{a}\) has to be real and the outgoing momenta read \[\begin{aligned} p_{a}=-{\omega_{a}\over 1+\zeta_{a}^{2}}(1+\zeta_{a}^{2},2\zeta_{a},0,\zeta_{a}^{2}-1). \end{aligned}\] Alternatively, the planarity condition implies that all emitted particles have azimuthal angles with either \(\phi_{a}=0\) or \(\phi_{a}=\pi\). Implementing momentum conservation \(p_{1}+\ldots+p_{5}=0\) gives three independent equations that determine the gluon energies \(\omega_{a}\) in terms of the center-of-mass energy \(\sqrt{s}\) and the flight directions of the outgoing gluons labelled by \(\zeta_{a}\), \[\begin{aligned} \omega_{3}&={\sqrt{s}\over 2}{(1+\zeta_{3}^{2})(1+\zeta_{4}\zeta_{5})\over (\zeta_{3}-\zeta_{4})(\zeta_{3}-\zeta_{5})}, \nonumber \\[0.2cm] \omega_{4}&={\sqrt{s}\over 2}{(1+\zeta_{4}^{2})(1+\zeta_{3}\zeta_{5})\over (\zeta_{4}-\zeta_{3})(\zeta_{4}-\zeta_{5})}, \label{eq:energies_dehomog}\\[0.2cm] \omega_{5}&={\sqrt{s}\over 2}{(1+\zeta_{5}^{2})(1+\zeta_{3}\zeta_{4})\over (\zeta_{5}-\zeta_{3})(\zeta_{5}-\zeta_{4})}. \nonumber \end{aligned}\] Furthermore, the finite positive energy condition \(0\leq \omega_{a}<\infty\) imposes constraints on the possible values of \(\zeta_{a}\). In particular, let us remark that finite energy implies that \(\zeta_{a}\neq \zeta_{b}\) for \(3\leq a<b\leq 5\). Using this parametrization, the amplitude [\[eq:5g_amplitude_spinors\]](#eq:5g_amplitude_spinors){reference-type="eqref" reference="eq:5g_amplitude_spinors"} takes the form \[\begin{aligned} \mathcal{A}_{5}&={2ig^{3}\over\sqrt{s}}{(\zeta_{3}-\zeta_{4})(\zeta_{3}-\zeta_{5})(\zeta_{4}-\zeta_{5})\over (1+\zeta_{3}\zeta_{4})(1+\zeta_{3}\zeta_{5})(1+\zeta_{4}\zeta_{5})}\left[-c_{2}{\zeta_{5}-\zeta_{3}\over \zeta_{3}}-c_{6}{\zeta_{4}-\zeta_{5}\over \zeta_{5}}\right. \nonumber \\[0.2cm] &\left.+c_{7}{\zeta_{3}-\zeta_{5}\over \zeta_{5}}-c_{8}{\zeta_{3}-\zeta_{4}\over \zeta_{4}} +c_{11}{\zeta_{5}-\zeta_{4}\over\zeta_{4}}+c_{13}{\zeta_{4}-\zeta_{3}\over \zeta_{3}} \right]. \label{eq:amplitude_zetas} \end{aligned}\] In order to find the zeros of the amplitude, we notice that the finite energy condition implies that the prefactor can never vanish. As a consequence, we find the following equation depending on the color factors \[\begin{aligned} c_{2}{\zeta_{5}-\zeta_{3}\over \zeta_{3}} +c_{6}{\zeta_{4}-\zeta_{5}\over \zeta_{5}}-c_{7}{\zeta_{3}-\zeta_{5}\over \zeta_{5}} \hspace*{3cm}\nonumber \\[0.2cm] +c_{8}{\zeta_{3}-\zeta_{4}\over \zeta_{4}}-c_{11}{\zeta_{5}-\zeta_{4}\over\zeta_{4}}-c_{13}{\zeta_{4}-\zeta_{3}\over \zeta_{3}}=0. \label{eq:planar_cond1} \end{aligned}\] The planar zero condition just derived is a homogeneous equation of vanishing degree. Since the amplitude [\[eq:amplitude_zetas\]](#eq:amplitude_zetas){reference-type="eqref" reference="eq:amplitude_zetas"} diverges whenever any of the \(\zeta_{a}\) vanishes, we can multiply the previous equation by \(\zeta_{3}\zeta_{4}\zeta_{5}\) without generating spurious solutions in the physical region. Taking projective coordinates \[\begin{aligned} (\zeta_{3},\zeta_{4},\zeta_{5})=\lambda(1,U,V), \hspace*{1cm} \lambda,U,V\neq 0 \label{eq:projective_coord} \end{aligned}\] the planar zeros of the five-gluon amplitude are determined by the loci defined by the following equation \[\begin{aligned} c_{7}U-c_{8}V-c_{6}U^{2}+c_{11}V^{2}+(c_{2}+c_{6}-c_{7}+c_{8}-c_{11}-c_{13})UV+c_{13}U^{2}V-c_{2}UV^{2}=0. \label{eq:cubic_eq} \end{aligned}\] Moreover, this equation is homogeneous in the color factors and therefore independent of the normalization of the gauge group generators. Since there exists a normalization of the generators that makes all structure constants integer numbers, the planar zeros are determined by a cubic curve with integer coefficients. In terms of the projective coordinates [\[eq:projective_coord\]](#eq:projective_coord){reference-type="eqref" reference="eq:projective_coord"}, the energies of the outgoing particles take the form \[\begin{aligned} \omega_{3}&={\sqrt{s}\over 2}{(1+\lambda^{2})(1+\lambda^{2}UV)\over \lambda^{2}(1-U)(1-V)}, \nonumber \\[0.2cm] \omega_{4}&={\sqrt{s}\over 2}{(1+\lambda^{2}U^{2})(1+\lambda^{2}V)\over \lambda^{2}(U-1)(U-V)}, \label{eq:omegas_UV}\\[0.2cm] \omega_{5}&={\sqrt{s}\over 2}{(1+\lambda^{2}V^{2})(1+\lambda^{2}U)\over \lambda^{2}(V-1)(V-U)}.\nonumber \end{aligned}\] We have seen already that in order to keep the amplitude finite we have to exclude \(U=0\) and \(V=0\) from the physical region. Now, energy finiteness further demands that \(U\neq 1\), \(V\neq 1\), and \(U\neq V\). By requiring \(\omega_{a}\geq 0\) we find that, for example, the region \(U>0\), \(V>0\) has to be considered unphysical as well. Indeed, if this is the case all three numerators in [\[eq:omegas_UV\]](#eq:omegas_UV){reference-type="eqref" reference="eq:omegas_UV"} are positive whereas the three denominators cannot have the same sign simultaneously. As a consequence, at least one of the energies has to be negative and the configuration is excluded. Studying the values of \(U\) and \(V\) in which the three energies are simultaneously positive for a given \(\lambda\), we arrive at the physical regions shown in Fig. [\[fig:physical_regions\]](#fig:physical_regions){reference-type="ref" reference="fig:physical_regions"}. Notice that the plot is symmetric under the exchange \(U\leftrightarrow V\). The conformal structure of the equation defining the planar zeros indicates that each solution of Eq. [\[eq:cubic_eq\]](#eq:cubic_eq){reference-type="eqref" reference="eq:cubic_eq"} can be realized in infinitely many physical setups, depending on the value of the parameter \(\lambda\). Notice that the boundaries of the allowed regions depend on \(\lambda\) as well, so they move as this parameter varies, while the position of the zeros, being a projective invariant, remain fixed. A particularly interesting regime is the soft limit, in which one or various of the emitted gluon energies tend to zero. From Eq. [\[eq:omegas_UV\]](#eq:omegas_UV){reference-type="eqref" reference="eq:omegas_UV"} we see that the points in the \((U,V)\) plane for which \(\omega_{a}\) vanish are given by \[\begin{aligned} UV&=-{1\over \lambda^{2}} \hspace*{1cm} (\omega_{3}=0), \nonumber \\[0.2cm] V&=-{1\over \lambda^{2}} \hspace*{1cm} (\omega_{4}=0), \\[0.2cm] U&=-{1\over \lambda^{2}} \hspace*{1cm} (\omega_{5}=0), \nonumber \end{aligned}\] which are indicated by the dashed lines in Fig. [\[fig:physical_regions\]](#fig:physical_regions){reference-type="ref" reference="fig:physical_regions"}. On general grounds, a planar zero corresponding to a point of the cubic [\[eq:cubic_eq\]](#eq:cubic_eq){reference-type="eqref" reference="eq:cubic_eq"} can be physically captured in the soft limit provided there is a value of \(\lambda\) for which this point collides against any of the "soft" lines defining the boundaries of the physical region. The first example to analyze is the case of incoming gluons in a singlet state for arbitrary gauge group, already studied in. Using the fact that \(f^{da_{3}b}f^{ba_{4}c}f^{ca_{5}d}\sim f^{a_{3}a_{4}a_{5}}\), we find \[\begin{aligned} c_{2}=c_{6}=-c_{7}=c_{8}=-c_{11}=-c_{13}=-f^{a_{3}a_{4}a_{5}}. \end{aligned}\] The cubic equation then reads \[\begin{aligned} U+V+U^{2}+V^{2}-6 UV +U^{2}V+UV^{2}=0. \label{eq:cubid_eq_singlet} \end{aligned}\] The associated algebraic curve is represented in Fig. [\[fig:singlet_case\]](#fig:singlet_case){reference-type="ref" reference="fig:singlet_case"}. Comparing with Fig. [\[fig:physical_regions\]](#fig:physical_regions){reference-type="ref" reference="fig:physical_regions"} we see that for small enough \(\lambda\) there is indeed a large part of the curve lying on physically allowed regions. In particular, for \(\lambda<1\) there are solutions with arbitrarily large \(|U|\) and \(|V|\). We study next the loci defined by Eq. [\[eq:cubic_eq\]](#eq:cubic_eq){reference-type="eqref" reference="eq:cubic_eq"} for SU(\(N\)) gauge groups with different ranks and various color configurations: #### SU(\(\mathbf{2}\)). In the case of SU(2) it is easy to write a closed expression for the color factors \[\begin{aligned} c_{2}&=\delta^{a_{3}a_{4}}\epsilon^{a_{2}a_{5}a_{1}}-\delta^{a_{2}a_{4}}\epsilon^{a_{3}a_{5}a_{1}}, \nonumber \\[0.2cm] c_{6}&=\delta^{a_{5}a_{3}}\epsilon^{a_{2}a_{4}a_{1}}-\delta^{a_{2}a_{3}}\epsilon^{a_{5}a_{4}a_{1}}, \nonumber \\[0.2cm] c_{7}&=\delta^{a_{1}a_{4}}\epsilon^{a_{3}a_{5}a_{2}}-\delta^{a_{3}a_{4}}\epsilon^{a_{1}a_{5}a_{2}}, \nonumber \\[0.2cm] c_{8}&=\delta^{a_{2}a_{5}}\epsilon^{a_{4}a_{1}a_{3}}-\delta^{a_{4}a_{5}}\epsilon^{a_{2}a_{1}a_{3}}, \\[0.2cm] c_{11}&=\delta^{a_{4}a_{3}}\epsilon^{a_{2}a_{5}a_{1}}-\delta^{a_{2}a_{3}}\epsilon^{a_{4}a_{5}a_{1}}, \nonumber \\[0.2cm] c_{13}&=\delta^{a_{4}a_{5}}\epsilon^{a_{1}a_{2}a_{3}}-\delta^{a_{1}a_{5}}\epsilon^{a_{4}a_{2}a_{3}}, \nonumber \end{aligned}\] where a convenient normalization of the gauge group generators has been chosen. In principle, the color factors can only take the values \(0,\pm 1\), and \(\pm 2\), since each term on the right-hand side of these equations is either \(0\) or \(\pm 1\). However, the case \(\pm 2\) is excluded. The reason is that due to the structure of indices of the Levi-Civita tensor, sharing the last two entries, they cannot have oposite signs. As a consequence, they cannot add up and we conclude that for SU(2) the color factors satisfy \(c_{i}=0,\pm 1\). An exploration of the possible external color numbers show that there are no solutions containing physical points. We illustrate this with a few examples. Our first case has color structure \((a_{1},a_{2},a_{3},a_{4},a_{5})=(2,3,1,1,1)\), giving the same value for all color factors \[\begin{aligned} c_{2}=c_{6}=c_{7}=c_{8}=c_{11}=c_{13}=1. \end{aligned}\] The resulting cubic equation completely factorizes as \[\begin{aligned} (U-1)(V-1)(U-V)=0. \label{eq:curve1_SU(2)} \end{aligned}\] We see that the three solutions lie outside the physical region and as a consequence there are no planar zeros for this gauge configuration. Next we try \((a_{1},a_{2},a_{3},a_{4},a_{5})=(2,2,2,1,3)\), which corresponds to color factors \[\begin{aligned} c_{2}=c_{7}=c_{8}=c_{13}=0, \hspace*{1cm} c_{6}=-c_{11}=1. \end{aligned}\] In this case the equation for the zeros becomes quadratic and factorizes as \[\begin{aligned} (U-V)^{2}=0. \end{aligned}\] The geometric loci of zeros coincide again with the unphysical region corresponding to two particles in the final state with infinite energy. As a last example, we take \((a_{1},a_{2},a_{3},a_{4},a_{5})=(1,2,2,2,3)\), which gives \[\begin{aligned} c_{2}=c_{8}=c_{11}=c_{13}=0, \hspace*{1cm} c_{6}=c_{7}=1. \end{aligned}\] In this case the cubic again degenerates into a quadratic equation \[\begin{aligned} U(U-1)=0, \end{aligned}\] which has no physical solutions. To summarize, a scan of possible values of the external color numbers show that the only curves obtained in this case coincide with unphysical regions in the plot of Fig. [\[fig:physical_regions\]](#fig:physical_regions){reference-type="ref" reference="fig:physical_regions"}, \(U=0,1\), \(V=0,1\) or \(U=V\). The only possibility for planar zeros in this case is to consider external states without well-defined color numbers, such as the singlet case studied above. #### SU(\(\mathbf{3}\)). We work out a first example where we take color quantum numbers \((a_{1},a_{2},a_{3},a_{4},a_{5})=(7,7,6,1,5)\) and color factors \[\begin{aligned} c_{2}=-c_{7}=c_{8}=-c_{13}=2, \hspace*{1cm} c_{6}=-c_{11}=-1. \end{aligned}\] The planar zeros are given by the factorized cubic \[\begin{aligned} (U+V-2)(U+V-2UV)=0. \label{eq:curve_SU(3)_ex1} \end{aligned}\] This is a hyperbola together with its tangent at \((U,V)=(1,1)\) (see the left panel of Fig. [\[fig:SU(3)\_case1\]](#fig:SU(3)_case1){reference-type="ref" reference="fig:SU(3)_case1"}). The loci has nonvanishing intersection with the physically allowed region in the \(UV\) plane for appropriate values of \(\lambda\). A different hyperbola is obtained for \((a_{1},a_{2},a_{3},a_{4},a_{5})=(1,4,1,2,6)\) with \[\begin{aligned} c_{2}=-c_{11}=-1,\hspace*{1cm} c_{6}=-4, \hspace*{1cm} c_{7}=c_{8}=0, \hspace*{1cm} c_{13}=-2. \end{aligned}\] The equation determining the zeros also factorizes in this case, giving \[\begin{aligned} (2U-V)(-2U+V+UV)=0. \label{eq:second_hyp_SU(2)} \end{aligned}\] Again, we have a hyperbola and one of its tangents, this time at the origin. The curves are shown in the RHS panel of Fig. [\[fig:SU(3)\_case1\]](#fig:SU(3)_case1){reference-type="ref" reference="fig:SU(3)_case1"}. As in the SU(2) cases all examples explored for the gauge group SU(3) show factorization of the cubic equation. In this latter case, however, not only the type of curves is enlarged to include hyperbolas which were absent for SU(2), but the curves contain physical points. In addition, considering quantum numbers in a SU(2) subgroup of SU(3) generates the curves obtained for the former group. #### SU(5). Enlarging the gauge group to SU(5) brings more general types of cubic algebraic curves. This is for example the case taking \((a_{1},a_{2},a_{3},a_{4},a_{5})=(17,19,19,18,23)\). The resulting color factors are \[\begin{aligned} c_{2}=c_{13}=0, \hspace*{1cm} c_{6}=c_{8}=2, \hspace*{1cm} c_{7}=c_{11}=1. \end{aligned}\] Since \(c_{2}\) and \(c_{13}\) vanish, it results in the following quadratic equation determining the planar zeros \[\begin{aligned} U-2 U^{2}-2 V + 2 U V + V^{2}=0. \label{eq:alg_curve1_SU(5)} \end{aligned}\] Unlike the examples encountered for SU(2) and SU(3), this curve does not factorize and corresponds to the hyperbola shown in the LHS panel of Fig. [\[fig:plotsSU(5)1\]](#fig:plotsSU(5)1){reference-type="ref" reference="fig:plotsSU(5)1"}. A second interesting example is provided by \((a_{1},a_{2},a_{3},a_{4},a_{5})=(19,18,23,17,19)\). The corresponding color factors are \[\begin{aligned} c_{2}=c_{11}=0, \hspace*{1cm} c_{6}=c_{8}=2, \hspace*{1cm} c_{7}=c_{13}=1. \end{aligned}\] The resulting equation for the zero \[\begin{aligned} U-2 U^{2}-2 V + 2 U V + U^{2} V=0 \label{eq:SU(5)curve_ex2} \end{aligned}\] is the cubic curve shown in the RHS panel of Fig. [\[fig:plotsSU(5)1\]](#fig:plotsSU(5)1){reference-type="ref" reference="fig:plotsSU(5)1"}. As a last example we take \((a_{1},a_{2},a_{3},a_{4},a_{5})=(19,19,18,23,17)\), with color factors \[\begin{aligned} c_{2}=-c_{7}=-2, \hspace*{1cm} c_{6}=c_{8}=-c_{11}=-c_{13}=1. \end{aligned}\] We get the cubic curve \[\begin{aligned} 2U-U^{2}-V-U^{2}V-V^{2}+2UV^{2}=0, \label{eq:curveSU(3)Ex3} \end{aligned}\] which, as shown in Fig. [\[fig:plotSU(5)Ex3\]](#fig:plotSU(5)Ex3){reference-type="ref" reference="fig:plotSU(5)Ex3"}, contains a singular point at \((U,V)=(1,1)\). We see how SU(5) provides more general types of curves than the ones found for unitary groups of lower rank. We also have to take into account that SU(5) contains SU(3) and SU(2) subgroups. Using the standard generators (see, for example, ) these subgroups are respectively generated by \(\{T^{1},\ldots,T^{8}\}\) and \(\{T^{21}, T^{22},T^{23}\}\). Thus, setting the external indices in the subsets \((1,\ldots,8)\) or \((21,22,23)\) we recover previous examples. For instance, \((a_{1},a_{2},a_{3},a_{4},a_{5})=(7,7,6,1,5)\) gives the curve shown in the LHS panel of Fig. [\[fig:SU(3)\_case1\]](#fig:SU(3)_case1){reference-type="ref" reference="fig:SU(3)_case1"}, whereas \((a_{1},a_{2},a_{3},a_{4},a_{5})=(22,23,21,21,21)\) reproduces Eq. [\[eq:curve1_SU(2)\]](#eq:curve1_SU(2)){reference-type="eqref" reference="eq:curve1_SU(2)"}. # Planar zeros and color permutations {#sec:permutations} It is interesting to see how the zeros here investigated transform under permutations of the color quantum numbers of the external particles. We begin considering those permutations preserving the choice of amplitudes basis implied by Eq. [\[eq:kleiss-kuijf_exp\]](#eq:kleiss-kuijf_exp){reference-type="eqref" reference="eq:kleiss-kuijf_exp"}. These are elements of \(S_{3}\) permuting the color indices of the three outgoing gluons (see Fig. [\[fig:KK-rel\]](#fig:KK-rel){reference-type="ref" reference="fig:KK-rel"}). In order to find the action of these permutations on the geometric loci of planar zeros, we can see how the color factors [\[eq:color_factors\]](#eq:color_factors){reference-type="eqref" reference="eq:color_factors"} transform under permutations of the \((a_{3},a_{4},a_{5})\) color indices. Here instead we use a more geometric approach and work with the homogenization of the cubic equation [\[eq:cubic_eq\]](#eq:cubic_eq){reference-type="eqref" reference="eq:cubic_eq"} \[\begin{aligned} c_{7}Z^{2}U-c_{8}Z^{2}V-c_{6}ZU^{2}+c_{11}ZV^{2}+(c_{2}+c_{6}-c_{7}+c_{8}-c_{11}-c_{13})ZUV \nonumber \\[0.2cm] +c_{13}U^{2}V-c_{2}UV^{2}=0. \label{eq:cubic_eq_hom} \end{aligned}\] The group \(S_{3}\) acts passively by permutation of the coordinates \((Z,U,V)\). Let us discuss the geometrical meaning of these transformations. Equation [\[eq:cubic_eq_hom\]](#eq:cubic_eq_hom){reference-type="eqref" reference="eq:cubic_eq_hom"} is defined in the whole projective plane, which is covered by the three affine patches centered at \((1,0,0)\), \((0,1,0)\), and \((0,0,1)\). The group \(S_{3}\) is generated by \(S_{2}\) transformations interchanging the two coordinates within each patch, together with cyclic permutations of the three patches. This defines a coset decomposition of \(S_{3}\) with respect to its cyclic subgroup generated by \((123)\). The physically allowed regions shown in Fig. [\[fig:physical_regions\]](#fig:physical_regions){reference-type="ref" reference="fig:physical_regions"} correspond to the patch centered at \((1,0,0)\). It has been already pointed out that it is invariant under the interchange of the two coordinates \(U\leftrightarrow V\). Moreover, the corresponding plots in the other two affine coordinate patches are identical to this one. This is easy to see from Eq. [\[eq:energies_dehomog\]](#eq:energies_dehomog){reference-type="eqref" reference="eq:energies_dehomog"}, where it is glaring that cyclic permutations of the three patches only interchange the energies of the three outgoing gluons. Thus, the positivity conditions remain algebraically the same in any of the three affine patches. The final conclusion is that \(S_{3}\) only acts on the axes labels of the plot in Fig. [\[fig:physical_regions\]](#fig:physical_regions){reference-type="ref" reference="fig:physical_regions"}. This is a passive version of the fact that the energies of the outgoing gluons are determined by momentum conservation alone and that the color structures play no role in it. Applying a permutations of \(S_{3}\) to Eq. [\[eq:cubic_eq_hom\]](#eq:cubic_eq_hom){reference-type="eqref" reference="eq:cubic_eq_hom"}, we find that the color factors transform under the (six-dimensional) regular representation of the group: in particular, if we write \(\boldsymbol{C}=(c_{2},c_{6},c_{7},c_{8},c_{11},c_{13})^{T}\) the group \(S_{3}\) acts through the matrices \[\begin{aligned} (1)(2)(3)&=\left( \begin{array}{cccccc} 1 & 0 & 0 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 & 0 & 0 \\ 0 & 0 & 1 & 0 & 0 & 0 \\ 0 & 0 & 0 & 1 & 0 & 0 \\ 0 & 0 & 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 0 & 0 & 1 \end{array} \right), \hspace*{1.3cm} (123)= \left( \begin{array}{cccccc} 0 & 0 & 0 & 1 & 0 & 0 \\ 1 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 1 \\ 0 & 1 & 0 & 0 & 0 & 0 \\ 0 & 0 & 1 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 1 & 0 \end{array} \right), \nonumber \\[0.2cm] (132)&=\left( \begin{array}{cccccc} 0 & 1 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 1 & 0 & 0 \\ 0 & 0 & 0 & 0 & 1 & 0 \\ 1 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 1 \\ 0 & 0 & 1 & 0 & 0 & 0 \end{array} \right),\hspace*{1cm} (12)(3)= \left( \begin{array}{cccccc} 0 & 0 & 0 & 0 & 1 & 0 \\ 0 & 0 & 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 1 \\ 1 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 1 & 0 & 0 \end{array} \right), \hspace*{1cm} \label{eq:regular_rep_S3}\\[0.2cm] (13)(2)&=\left( \begin{array}{cccccc} 0 & 0 & 1 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 1 \\ 1 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 & 0 & 0 \\ 0 & 1 & 0 & 0 & 0 & 0 \end{array} \right),\hspace*{1cm} (1)(23)= \left( \begin{array}{cccccc} 0 & 0 & 0 & 0 & 0 & 1 \\ 0 & 0 & 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 & 0 & 0 \\ 1 & 0 & 0 & 0 & 0 & 0 \end{array} \right). \nonumber \end{aligned}\] Notice that the combination of color factors in the coefficient of the \(ZUV\) term itself transforms with the one-dimensional parity representation, \[\begin{aligned} \sigma(c_{2}+c_{6}-c_{7}+c_{8}-c_{11}-c_{13})=(-1)^{\pi(\sigma)}(c_{2}+c_{6}-c_{7}+c_{8}-c_{11}-c_{13}), \end{aligned}\] where \(\pi(\sigma)=0,1\) for even and odd permutations respectively. This is a consequence of the fact that \(ZUV\) is an invariant under the permutation group. Applying the transformations given by the matrices [\[eq:regular_rep_S3\]](#eq:regular_rep_S3){reference-type="eqref" reference="eq:regular_rep_S3"} to the curve [\[eq:cubid_eq_singlet\]](#eq:cubid_eq_singlet){reference-type="eqref" reference="eq:cubid_eq_singlet"} obtained in the case of the scattering of two gluons in a singlet state, \(\boldsymbol{C}=(1,1,-1,1,-1,-1)^{T}\), we find that it is invariant, since \(\boldsymbol{C}\) transforms with the parity of the permutation, \(\sigma(\boldsymbol{C})=(-1)^{\pi(\sigma)}\boldsymbol{C}\). This means the curve shown in the plot in Fig. [\[fig:singlet_case\]](#fig:singlet_case){reference-type="ref" reference="fig:singlet_case"} describes the planar zeros in all three coordinate patches. Incidentally, notice that the curve is invariant as well under the interchange of the two coordinates in any of the three corresponding plots. The curves presented in Section [3](#sec:gluon_planar){reference-type="ref" reference="sec:gluon_planar"} are expressed in the patch \((1,0,0)\). Under permutation of the two coordinates, some solutions remain invariant, such as Eq. [\[eq:curve_SU(3)\_ex1\]](#eq:curve_SU(3)_ex1){reference-type="eqref" reference="eq:curve_SU(3)_ex1"}, or get mapped into a different solution. We can also consider the transformation of the curves with respect to general color permutations. They form the group \(\mathsf{TCS}_{5}\) which is identified with the cyclic Lie operad \(\mathsf{Lie}(\!(5)\!)\). Its structure has been studied in Ref., where it was found that its action on the six independent color structures is given by the following six-dimensional representation of \(S_{5}\) \[\begin{aligned} \yng(3,1,1)\,\,\. \end{aligned}\] Unlike the transformations of \(S_{3}\) studied above, those in \(\mathsf{TCS}_{5}\backslash S_{3}\) do not act on the geometric loci of planar zeros by permutation of the projective coodinates \((Z,U,V)\). The most obvious example is provided by the interchange of the color indices of the incoming gluons \(a_{1}\), \(a_{2}\). From Eq. [\[eq:color_factors\]](#eq:color_factors){reference-type="eqref" reference="eq:color_factors"} we find that this transformation acts linearly on \(\boldsymbol{C}\) through the matrix \[\begin{aligned} (12)(3)(4)(5)=\left( \begin{array}{rrrrrr} 0 & 0 &-1 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 &-1 & 0 \\-1 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 &-1 \\ 0 &-1 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 &-1 & 0 & 0 \end{array} \right), \end{aligned}\] where we have used the cycle notation for the elements of \(S_{5}\). It is interesting to point out that this transformation leaves invariant the coefficient of the \(ZUV\) coefficient in [\[eq:cubic_eq_hom\]](#eq:cubic_eq_hom){reference-type="eqref" reference="eq:cubic_eq_hom"}. This last property is not shared by other transformations in \(\mathsf{TCS}_{5}\backslash S_{3}\). For example, \[\begin{aligned} (134)(25)=\left( \begin{array}{rrrrrr}-1 & 0 & 0 & 0 & 0 & 1 \\ 0 & 0 & 1 & 0 & 0 & 0 \\ 0 &-1 & 1 & 0 & 0 & 0 \\-1 & 0 & 0 & 0 & 1 & 0 \\-1 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 1 &-1 & 0 & 0 \end{array} \right), \hspace*{0.5cm} (1245)(3)=\left( \begin{array}{rrrrrr}-1 & 1 &-1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 1 \\-1 & 0 & 0 & 0 & 0 & 1 \\-1 & 0 &-1 & 1 & 0 & 1 \\ 0 & 1 &-1 & 0 & 0 & 0 \end{array} \right), \end{aligned}\] act on \(c_{2}+c_{6}-c_{7}+c_{8}-c_{11}-c_{13}\) respectively as \[\begin{aligned} (134)(25)(c_{2}+c_{6}-c_{7}+c_{8}-c_{11}-c_{13})&=c_{2}-c_{6}+c_{7}-c_{8}-c_{11}-c_{13}, \nonumber \\[0.2cm] (1245)(3)(c_{2}+c_{6}-c_{7}+c_{8}-c_{11}-c_{13})&=c_{2}-c_{6}-c_{7}+c_{8}-c_{11}+c_{13}. \end{aligned}\] Through its transformations of the color factors, \(\mathsf{TCS}_{5}\) acts on the curves determining the planar zeros. In fact, this action can be used to generate the whole orbit of projective curves associated with the permutations of the color quantum numbers of the interacting gluons. # Graviton planar zeros from color-kinematics duality {#sec:graviton} We turn now to the problem of planar zeros in the five-graviton tree level amplitude. The gravitational amplitude can be constructed from its gluon counterpart [\[eq:amplitude_gluon_general\]](#eq:amplitude_gluon_general){reference-type="eqref" reference="eq:amplitude_gluon_general"} using the BCJ prescription, \[\begin{aligned} -i\mathcal{M}_{5}&=\left({\kappa\over 2}\right)^{3}\left({n_{1}^{2}\over s_{12}s_{45}}+{n_{2}^{2}\over s_{23}s_{15}}+{n_{3}^{2}\over s_{34}s_{12}} +{n^{2}_{4}\over s_{45}s_{23}}+{n_{5}^{2}\over s_{15}s_{34}}+{n_{6}^{2}\over s_{14}s_{25}} +{n_{7}^{2}\over s_{13}s_{25}}+{n_{8}^{2}\over s_{24}s_{13}} \right. \nonumber \\[0.2cm] &+ \left.{n_{9}^{2}\over s_{35}s_{24}}+{n_{10}^{2}\over s_{14}s_{35}}+{n_{11}^{2}\over s_{15}s_{24}} +{n_{12}^{2}\over s_{12}s_{35}}+{n_{13}^{2}\over s_{23}s_{14}}+{n_{14}^{2}\over s_{25}s_{34}} +{n_{15}^{2}\over s_{13}s_{45}}\right), \end{aligned}\] provided the numerators \(n_{i}\) satisfy color-kinematics duality, with \(\kappa\) the gravitational coupling. Taking the graviton polarizations \((1^{-},2^{-},3^{+},4^{+},5^{+})\), we use our expression for the MHV gauge amplitude given in Eq. [\[eq:5g_amplitude_spinors\]](#eq:5g_amplitude_spinors){reference-type="eqref" reference="eq:5g_amplitude_spinors"} to write \[\begin{aligned} -i\mathcal{M}_{5} &=-i\left({\kappa\over 2}\right)^{3}\langle 12\rangle^{3}\left({n_{2}\over \langle 23\rangle\langle 34\rangle\langle 45\rangle\langle 51\rangle} +{n_{6}\over \langle 25\rangle\langle 53\rangle\langle 34\rangle\langle 41\rangle} +{n_{7}\over \langle 25\rangle\langle 53\rangle\langle 43\rangle\langle 31\rangle}\right. \nonumber \\[0.2cm] &\left.+{n_{8}\over \langle 24\rangle\langle 45\rangle\langle 53\rangle\langle 31\rangle} +{n_{11}\over \langle 24\rangle\langle 43\rangle\langle 35\rangle\langle 51\rangle} +{n_{13}\over \langle 23\rangle\langle 35\rangle\langle 54\rangle\langle 41\rangle}\right). \end{aligned}\] Using the form of the gauge theory numerators in Eq. [\[eq:numerators\]](#eq:numerators){reference-type="eqref" reference="eq:numerators"}, after a bit of algebra we arrive at the simpler expression \[\begin{aligned} -i\mathcal{M}_{5}&=-\left({\kappa\over 2}\right)^{3}{\langle 12\rangle^{7}[41][52]\over \langle 12\rangle\langle 14\rangle \langle 23\rangle \langle 25\rangle \langle 34\rangle \langle 35\rangle \langle 45\rangle} \left(1-{\langle 14\rangle\langle 25\rangle [42][51]\over \langle 15\rangle \langle 24\rangle[41][52]}\right). \end{aligned}\] The term inside the parenthesis can be further simplified taking into account the relation \(s_{ij}=\langle ij\rangle[ji]\), \[\begin{aligned} 1-{\langle 14\rangle\langle 25\rangle [42][51]\over \langle 15\rangle \langle 24\rangle[41][52]} =1-\left({\langle 14\rangle\langle 25\rangle\over \langle 15\rangle\langle 24\rangle}\right)^{2} \left({s_{15}s_{24}\over s_{14}s_{25}}\right), \end{aligned}\] which in turn can be written as a function of \(\Delta \phi_{45}\), the difference of azimuthal angles of particles \(4\) and \(5\) (see the Appendix of Ref. ), \[\begin{aligned} \left({\langle 14\rangle\langle 25\rangle\over \langle 15\rangle\langle 24\rangle}\right)^{2} \left({s_{15}s_{24}\over s_{14}s_{25}}\right)=e^{2i\Delta\phi_{45}}. \end{aligned}\] With this, the five-graviton tree level amplitude reads \[\begin{aligned} -i\mathcal{M}_{5}&=-\left({\kappa\over 2}\right)^{3}{\langle 12\rangle^{7}[41][52]\over \langle 12\rangle\langle 14\rangle \langle 23\rangle \langle 25\rangle \langle 34\rangle \langle 35\rangle \langle 45\rangle} \left(1-e^{2i\Delta\phi_{45}}\right). \label{eq:grav_ampl} \end{aligned}\] Given our choice of reference frame, planarity implies that for any two outgoing momenta their azimuthal angles must satisfy \(\Delta\phi_{ij}=0,\pi\). In both cases we find from [\[eq:grav_ampl\]](#eq:grav_ampl){reference-type="eqref" reference="eq:grav_ampl"} that the gravitational amplitude vanishes \[\begin{aligned} -i\mathcal{M}_{5}\Big|_{\rm planar}=0. \end{aligned}\] Unlike the gauge case where the cancellation condition depends on the color factors of the incoming particles, the graviton amplitude automatically vanishes in the limit of planar scattering. We now show that this is a consequence of color kinematic duality. In gauge theories we have seen that the vanishing of the amplitude in the planar case leads to a nontrivial condition on the momenta of the outgoing particles given in Eq. [\[eq:planar_cond1\]](#eq:planar_cond1){reference-type="eqref" reference="eq:planar_cond1"}. In fact, the numerators in Eq. [\[eq:numerators\]](#eq:numerators){reference-type="eqref" reference="eq:numerators"} have been chosen to satisfy color-kinematics duality, so we can obtain the planar zero condition for gravity by replacing the color factors \(\{c_{2},c_{6},c_{7},c_{8},c_{11},c_{13}\}\) in [\[eq:planar_cond1\]](#eq:planar_cond1){reference-type="eqref" reference="eq:planar_cond1"} with the corresponding numerators \(\{n_{2},n_{6},n_{7},n_{8},n_{11},n_{13}\}\). In terms of the stereographic coordinates, the latter are given by \[\begin{aligned} n_{6}&=n_{7}=is^{3\over 2}{(\zeta_{3}-\zeta_{5})\zeta_{5}\over \zeta_{3}(1+\zeta_{4}\zeta_{5})}, \nonumber \\[0.2cm] n_{8}&=is^{3\over 2}{(\zeta_{3}-\zeta_{5})\zeta_{4}\over \zeta_{3}(1+\zeta_{4}\zeta_{5})}, \\[0.2cm] n_{2}&=n_{11}=n_{13}=0. \nonumber \end{aligned}\] After this substitution, the condition for the existence of a zero in the amplitude is identically satisfied \[\begin{aligned} -n_{2}{\zeta_{5}-\zeta_{3}\over \zeta_{3}}-n_{6}{\zeta_{4}-\zeta_{5}\over \zeta_{5}}+n_{7}{\zeta_{3}-\zeta_{5}\over \zeta_{5}}-n_{8}{\zeta_{3}-\zeta_{4}\over \zeta_{4}} +n_{11}{\zeta_{5}-\zeta_{4}\over\zeta_{4}}+n_{13}{\zeta_{4}-\zeta_{3}\over \zeta_{3}} \nonumber \\[0.2cm] =is^{3\over 2}{(\zeta_{3}-\zeta_{5})\over \zeta_{3}(1+\zeta_{4}\zeta_{5})}\Big(-\zeta_{4}+\zeta_{5}+\zeta_{3}-\zeta_{5}-\zeta_{3}+\zeta_{4}\Big)=0. \label{eq:planar_cond1_grav} \end{aligned}\] This implies that, in the gravitational case, planarity is enough to make the amplitude vanish, without additional kinematic conditions to be satisfied by the stereographic coordinates of the outgoing gravitons. # Closing remarks {#sec:conclusions} We have studied the presence of planar zeros in both Yang-Mills theories and gravity. For the case of gauge theories, we have represented in Fig. [\[fig:vector_plots\]](#fig:vector_plots){reference-type="ref" reference="fig:vector_plots"} the kinematics on the interaction plane for two typical planar zeros within the same color configuration. By varying the value of \(\lambda\equiv \zeta_{3}\) while keeping \(U\equiv\zeta_{4}/\zeta_{3}\) and \(V\equiv \zeta_{5}/\zeta_{3}\) constant, these processes can be deformed into a different one with the emission, for example, of one or more soft gluons while the total amplitude remains equal to zero. This happens because planar zeros live in the projective \(U\)-\(V\) plane and are therefore invariant under a simultaneous rescaling of the three outgoing stereographic coordinates. Without loss of generality we considered the situation in which the scattering takes place in the \(y=0\) plane. Planar zeros on a different interaction plane can be obtained by applying rotations to the solutions studied here. In particular, the Lorentz group acts on the stereographic coordinates parametrizing the direction of the momenta through \(\mbox{SL}(2,\mathbb{C})\) transformations \[\begin{aligned} \zeta_{k}'={a\zeta_{k}+b\over c\zeta_{k}+d}, \hspace*{2cm} ad-bc=1, \end{aligned}\] where for the incoming particles we have \(\zeta_{1}=\infty\) and \(\zeta_{2}=0\). Rotations can be spotted by looking for transformations leaving invariant the energies [\[eq:energies_dehomog\]](#eq:energies_dehomog){reference-type="eqref" reference="eq:energies_dehomog"}, together with those of the incoming particles. They are given by \[\begin{aligned} \left( \begin{array}{cc} a & b \\[0.2cm] c & d \end{array} \right)=\left( \begin{array}{cc} \sqrt{1-\end{array} \right). \end{aligned}\] For real \(| angle\)`<!-- -->`{=html}2\(with respect to the\)x\(-axis. Alternatively, for\)\| a rotation of angle \(\sin\phi'=\tanh{2\chi}\) around the \(y\)-axis. With the results here presented we have shed some light on the origin of the planar zeroes present in Yang-Mills scattering amplitudes. Our results can be generalized to an arbitrary number of external legs at Born level. It will be worth further investigating the effect of quantum corrections. We have also connected, via the BCJ duality, these zeroes to the corresponding ones in gravity. We are currently studying how this picture is modified when the scattering of open and closed strings is considered.
{'timestamp': '2016-09-02T02:04:08', 'yymm': '1607', 'arxiv_id': '1607.04605', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04605'}
# Introduction After the discovery of the extraordinary properties of graphene, the next challenge is to develop mechanisms that allow the enhancement and modulation of such properties. Along these lines, strain engineering is currently one of the trending topics in graphene science because of the possibility to induce new physical phenomena by means of mechanical strain. Examples are modifications on the Fermi velocity, the modulation of Landau levels spectra, the generation of pseudo-magnetic fields, the modulation of the electrical, and thermal conductivities, just to name a few. Furthermore, with the recent advances in experimental techniques to apply strain, there are different reports of uniaxial, biaxial, and shear strain in graphene. Interestingly, it has been shown that uniaxial strain can be applied in a controlled, reversible and non destructive way, making it of particular interest. Two of the most studied properties of uniaxially strained graphene are its electronic and vibrational structure. Currently, it is well know that in uniaxially strained graphene the crossing point of the valence and conduction bands at the Fermi level, the so called Dirac point, shifts away from the corner of the Brillouin zone (the \({\bf K}\) point), with no band gap opening. Among the vibrational structure of graphene, the \(E_{2g}\) phonon mode at the center of the Brillouin zone (the \(\Gamma\) point), is particularly interesting because it is responsible for the G-band in the Raman spectroscopy. Under uniaxial strain the \(E_{2g}\) phonon mode is split in two modes, one parallel and the other perpendicular to the axis of the applied strain. That effect is useful to characterize the direction and strength of the uniaxial strain via Raman spectroscopy. Even more, the full phonon dispersion, Grüneisen parameters, and the origin of the phonon instability at the ideal strength have been studied in uniaxially strained graphene using first principles calculations. Regardless the level of understanding of the electronic and vibrational structure of uniaxially strained graphene, some fundamental and important microscopic properties like the electron-phonon (e-ph) coupling needs to be studied in detail. In pristine graphene the e-ph coupling induces strong anomalies in the phonon dispersion, contributes to the intrinsic electronic resistivity, is responsible for most of the linewidth in the Raman G-band, and determines the scattering rules for the double resonance Raman 2D-band. Even more interesting could be the possibility to induce electron-phonon superconductivity by means of atomic-decorating, heavy doping, and a combination of doping and biaxial tensile strain. Therefore, in order to have a deep understanding of the effects of uniaxial strain on the vibrational, thermal and transport properties, a detailed study of the e-ph coupling in uniaxially strained graphene is mandatory. A key feature of the e-ph coupling are Kohn anomalies: anomalous behavior in the phonon dispersion due to an electronic screening of the ionic vibration, which are fully determined by the geometry of the Fermi surface. In graphene the Fermi surface is the Dirac point, thus the Kohn anomalies take place only at the \({\Gamma}\) and \({\bf K}\) points, which are shown as a discontinuity in the frequency derivative of the phonon dispersion of the highest optical (HO) branches. Therefore, the e-ph coupling is localized on the transversal (TO) and longitudinal (LO) optical branches at \({\Gamma}\) in the \(E_{2g}\) phonon mode. Meanwhile, at \({\bf K}\) the e-ph coupling is almost entirely localized on the TO \(A'_1\) phonon mode, with a very small contribution from the double degenerated \(E'\) phonon mode on the LO and the longitudinal acoustic (LA) phonon branches, which is usually neglected in the study of properties related to the e-ph coupling. In uniaxially strained graphene, the loss of hexagonal symmetry and the shift of the Dirac point is expected to have an effect on the e-ph coupling but, to the best of the authors knowledge, this is not yet reported. In this work we have employed first principles density functional theory (DFT) calculations to systematically study the effects of uniaxial strain along the armchair (AC) and zigzag (ZZ) directions on the Kohn anomalies and e-ph coupling in graphene. In particular, we determine the displacement of the Kohn anomaly from \({\bf K}\), its frequency softening, vibrational modes and e-ph coupling. [^1] We show that uniaxial strain induce a substantial enhancement of the e-ph coupling in the LA branch around \({\bf K}\) with respect to pristine graphene, generating a non-negligible Kohn anomaly. This paper is organized as follow: In Sec. [2](#methodology){reference-type="ref" reference="methodology"} we describe the computational details of our first principles calculations. In Sec. [3.1](#res-struct){reference-type="ref" reference="res-struct"} we present the changes in the bond length and average force constants that will be useful for forthcoming discussions. The Kohn anomalies in the phonon dispersion, its vibrational phonon modes and frequency shifts are shown in Sec. [3.2](#results-Kohn anomaly){reference-type="ref" reference="results-Kohn anomaly"}. An analysis of the uniaxial strain effects in the e-ph coupling is discussed in Sec. [3.3](#e-ph coupling-res){reference-type="ref" reference="e-ph coupling-res"}. In Sec. [4](#conclusions-sec){reference-type="ref" reference="conclusions-sec"} we summarize our main findings. Finally, we include four Appendix sections with several useful relations used along this article, concerning the structural properties of uniaxially strained graphene, the Kohn anomaly shift, the classical atomic displacement of the discussed phonon modes, and the calculated e-ph coupling quantities. # Computational details {#methodology} The present calculations were performed within DFT, in the framework of the Mixed Basis Pseudopotential approach (MBPP). Core electrons were replaced by norm conserving pseudopotentials with non-linear core-corrections included. Valence states were represented by a combination of \(s\) and \(p\) type localized functions at each atomic sites, complemented with plane waves up to a kinetic energy of 25 Ry. The exchange-correlation functional was treated with the PBE parameterization of the generalized gradient approximation. During the structural optimization, the carbons positions were relaxed until the interatomic forces were 0.0001 Ry/Bohr or less. For phonon and e-ph coupling calculations we employed the density functional perturbation theory as implemented in the MBPP code. Special attention was paid to the integration in the irreducible Brillouin zone with a 72\(\times\)`<!-- -->`{=html}72\(\times\)`<!-- -->`{=html}1 Monkhorst-Pack \(k\)-point mesh and a small Gaussian broadening of 0.10 eV. This was needed in order to avoid electronic smearing effects on the Kohn anomalies and at the same time obtain converged phonon frequencies. Dynamical matrices were calculated using 12\(\times\)`<!-- -->`{=html}12\(\times\)`<!-- -->`{=html}1 and 9\(\times\)`<!-- -->`{=html}9\(\times\)`<!-- -->`{=html}1 \(q\)-points grids for pristine and uniaxial strained graphene, respectively. Full phonon dispersion and force constants were obtained via standard Fourier interpolation. In order to resolve the Kohn anomalies on the phonon dispersion we also computed several low-symmetry \(q\)-points corresponding to the full \(q\)-grid of 72\(\times\)`<!-- -->`{=html}72\(\times\)`<!-- -->`{=html}1. For the evaluation of the e-ph coupling properties we used a denser \(k\)-grid of 144\(\times\)`<!-- -->`{=html}144\(\times\)`<!-- -->`{=html}1, within a Gaussian broadening varying from 0.05 to 0.30 eV which, however, does not affect our final results. To simulate a single atomic layer, we used the supercell approach and we left at least 12 \({\rm \AA}\) of vacuum space between successive layers to avoid spurious supercell effects on the electronic states and phonon frequencies. # Results and Discussion ## Structural properties {#res-struct} For pristine graphene we have obtained a lattice parameter of 2.465 Å, which corresponds to a bond length of 1.423 Å. Taking the derivative of the acoustic phonon branches in the limit of \({\bf q}\rightarrow 0\), we estimate a Young modulus of 369 N/m and a Poisson's ratio of 0.182. The calculated elastic constants values are in agreement with the previously experimental and computational reported values. For instance, Politano *et al.* perform phonon dispersion measurements from macroscopic graphene samples, and estimate a Young modulus of 342 N/m and a Poisson's ratio of 0.19 from the sound velocities of the TA and LA phonon branches. In the context of previous DFT-based reports, the Young Modulus value varies from 344 to 356 N/m, and the Poisson's ratio from 0.162 to 0.186, depending on the exchange-correlation functional and other numerical approximations. As we described in the Appendix [\[uniax-strain\]](#uniax-strain){reference-type="ref" reference="uniax-strain"}, the structural properties of graphene under ZZ and AC strain are defined by the relation between the parallel or applied strain \(\varepsilon_{\parallel}\), the perpendicular contraction \(\varepsilon_{\bot}\), and the C-C distances \(\alpha\) and \(\beta\) (see Fig. [\[esquema\]](#esquema){reference-type="ref" reference="esquema"} for the definition of strain directions, and the real and reciprocal lattice). In Fig. [\[struct-prop\]](#struct-prop){reference-type="ref" reference="struct-prop"}(a) we present the computed values for \(\varepsilon_{\bot}\) as a function of \(\varepsilon_{\parallel}\), and for reference we have included the linear dependence for a constant Poisson's ratio. From that, it clearly shows a non-linear behavior for \(\varepsilon_{\parallel} > 2 \%\), which indicates a non-constant Poisson's ratio, in agreement with previous works. Hereinafter, for simplicity \(\varepsilon_{\parallel}\) will be referred only as strain. The changes in the interatomic C-C distances \(\alpha\) and \(\beta\) are show in Fig. [\[struct-prop\]](#struct-prop){reference-type="ref" reference="struct-prop"}(b). Although we consider only tensile strain, the C-C distances do not increase in all cases. For ZZ strain there is a small contraction in \(\alpha\), corresponding to the bond perpendicular to the direction of the applied strain. In a classical picture, the contraction of \(\alpha\) should increase the force constant related to this bond, contrary of what is expected when a material is under tension. To corroborate this picture, we calculate the average force constant related to atom-atom bonds, defined by \[\label{forcecte} I(b) = \sqrt{\frac{1}{3}\sum_{i j}\Phi_{i j}^2 (b)},\] where \(\Phi_{i j}(b)\) represents the force constant matrix assigned to a bond \(b\). The respective \(I(\alpha)\) and \(I(\beta)\) are shown in Fig. [\[force-ctes\]](#force-ctes){reference-type="ref" reference="force-ctes"} for both ZZ and AC uniaxial strain. In all cases we found that the dominant change in \(I(b)\) comes from the longitudinal component of the force constant. Just as expected from the change in the length of the C-C bonds, all the average force constants decrease, except for a small hardening in \(I(\alpha)\) under ZZ strain. Such behavior is a key feature in the forthcoming discussion of the phonon frequency shift for the Kohn anomaly. ## Kohn anomalies {#results-Kohn anomaly} In order to determine the position of the Kohn anomalies under uniaxial strain, we need to determine the distance \(\Delta\) between the Dirac point and the \({\bf K}\) point in the electronic structure, as we described in Appendix [\[Dirac-cone-shift\]](#Dirac-cone-shift){reference-type="ref" reference="Dirac-cone-shift"}. We estimate the evolution of \(\Delta\) as a function of the applied strain by an interpolation of the electronic bands at the Fermi level (see Fig. [\[dp-shift\]](#dp-shift){reference-type="ref" reference="dp-shift"}(a)). We found that \(\Delta\) is bigger for strain in the AC direction than in ZZ, although for strains lower than 3 % it is almost independent of the strain direction. Then, the position of the Kohn anomaly should be at the phonon nesting vector \({\bf q}_{ZZ}\) or \({\bf q}_{AC}\), presented for the unit cell of the reciprocal space in Figs. [\[dp-shift\]](#dp-shift){reference-type="ref" reference="dp-shift"}(b) and [\[dp-shift\]](#dp-shift){reference-type="ref" reference="dp-shift"}(c) for the ZZ and AC strain, respectively. The phonon dispersion around the Kohn anomaly in uniaxially strained graphene for \(\varepsilon = 5 \%\), is show in Fig. [\[ph-disp\]](#ph-disp){reference-type="ref" reference="ph-disp"}. For an easy reference and comparison, each branch and its respective phonon mode will be identified by its polarization in pristine graphene: LO, TO, and LA. As general trends, at \({\Gamma}\) we can observe the splitting of the \(E_{2g}\) phonon mode (see Fig. [\[ph-disp\]](#ph-disp){reference-type="ref" reference="ph-disp"}(a)), and that the derivative discontinuity of the HO branches depends on the chosen direction along the Brillouin zone. Around \({\bf K}\), the Kohn anomaly in the HO branch shows the expected shift according to our estimation for \(\Delta\) (dotted line in Fig. [\[ph-disp\]](#ph-disp){reference-type="ref" reference="ph-disp"}(b)). More interesting is the new derivative discontinuity on the LA branch at approximately 125 and 129 meV for the AC and ZZ strain, respectively. The fact that such discontinuities occur at the nesting vector that connect two Dirac points (\({\bf q}_{ZZ}\) and \({\bf q}_{AC}\)) is a direct indication of a Kohn anomaly in the LA branch. This is confirmed in Sec. [3.3](#e-ph coupling-res){reference-type="ref" reference="e-ph coupling-res"} with the analysis of e-ph coupling in the LA branch. For a further discussion of the phonon modes at the Kohn anomaly, we first focus in the \({\Gamma}\) point. As has been reported previously, the splitting of the \(E_{2g}\) phonon mode results in two modes with eigenvectors which are perpendicular (with smaller softening) and parallel to the strain direction (see Fig. [\[gamma-modes\]](#gamma-modes){reference-type="ref" reference="gamma-modes"}(a)). This effect is measured in Raman spectroscopy via the G-band, and because its relevance in graphene characterization, we adopt the same nomenclature that identify as \({\rm G}^+\) (\({\rm G}^-\)) the band with smaller (higher) softening. The phonon frequency shift for the \({\rm G}^+\) and \({\rm G}^-\) bands are show in Fig. [\[gamma-modes\]](#gamma-modes){reference-type="ref" reference="gamma-modes"}(b). The present results for the shift and splitting of the G-band are in good agreement with previous theoretical and experimental reports. However, it is important to mention that there are a wide range of reported values because the different set ups and conditions to induce strain on graphene, as well as other effects like substrate interaction, temperature, and the number of graphene layers. Beside that, in the studied range of strain we obtain an almost linear softening in \({\rm G}^+\) and \({\rm G}^-\), which becomes independent of the strain direction for deformations lower than 2 %. For the anomalies at \({\bf q}_{ZZ}\) and \({\bf q}_{AC}\) in the TO and LA branches, we found a polarization of the phonon eigenvectors \(\eta_{\kappa s}^{{\bf q}\nu}\) as a function of the strain, such that the atoms moves on ellipses with the mayor axis parallel (LA) and perpendicular (TO) to the strain direction, and whose eccentricity approaches to one as the strain increases, until the ellipses become almost straight lines (see Appendix [\[app-atom-disp\]](#app-atom-disp){reference-type="ref" reference="app-atom-disp"} for a proper description of the classical atomic displacement in graphene). During this evolution the phase difference \(\varPhi\) between the atomic displacements along the \(x\) and \(y\) direction is \(\pm \pi/2\), the major and minor axes of the ellipses are defined by the magnitude of the phonon eigenvectors, and the relations \(\left|\eta_{x}^{TO}\right| = \left|\eta_{y}^{LA}\right|\) and \(\left|\eta_{y}^{TO}\right| = \left|\eta_{x}^{LA}\right|\) are always fulfilled. This means that the Kohn anomaly shift from the high symmetry point \({\bf K}\) induces a mixing of the phonon eigenvectors \(\eta_{\kappa s}^{TO}\) and \(\eta_{\kappa s}^{LA}\), which belong to the same irreducible representation of the point group of \({\bf q}_{ZZ}\) and \({\bf q}_{AC}\), as in pristine graphene for \({\bf q}\) points outside the high symmetry points \(\Gamma\), \({\bf K}\), and \({\bf M}\). Thus the classical atomic displacement on each anomaly are in mutually perpendicular ellipses, but with the same magnitudes for the major and minor axes. The magnitude of the phonon eigenvectors and the eccentricity of the resulting ellipses for the Kohn anomaly in the TO branch are show in Fig. [\[eigenvectK\]](#eigenvectK){reference-type="ref" reference="eigenvectK"} with the norm \(\sqrt{\left|\eta_{x}\right|^2+\left|\eta_{y}\right|^2}=1\) assumed for simplicity. Within the ZZ (AC) strain along the \(x(y)\) Cartesian axis (see Fig. [\[esquema\]](#esquema){reference-type="ref" reference="esquema"}), it is clear that the phonon eigenvectors tend to align in the strain direction, especially for ZZ strain where the eccentricity approaches one faster than for the AC strain, resulting in a straight line displacement. Using the same nomenclature as that for the splitting of the G-band (\({\rm G}^{-}\) and \({\rm G}^{+}\)) for the anomalies at the TO and LA branches, we will employ the \(+\)(\(-\)) superindex to indicate that the phonon mode has eigenvectors perpendicular (parallel) to the strain direction and the smaller (higher) softening. A schematic representation of the \({\rm TO}^{+}\) and \({\rm LA}^{-}\) modes and the behavior of the phonon frequency shift are show in Fig. [\[q-modes\]](#q-modes){reference-type="ref" reference="q-modes"}. Unlike the phonon frequency shift in \({\rm G}^{-}\), \({\rm G}^{+}\), and \({\rm LA}^{-}\), in the case of \({\rm TO}^{+}\) the phonon softening is non-linear and becomes nearly constant starting from 2 % of ZZ strain. In \({\rm TO}_{ZZ}^{+}\) the atoms move along the AC direction, inducing a large distortion of the \(\alpha\) bond. Thus, the constant frequency softening is a consequence of the very small increment of the force constant for the \(\alpha\) bond, whose length remains almost constant under ZZ strain (see Fig. [\[struct-prop\]](#struct-prop){reference-type="ref" reference="struct-prop"} and Fig. [\[force-ctes\]](#force-ctes){reference-type="ref" reference="force-ctes"}). In \({\rm TO}_{AC}^{+}\) the atoms move along the ZZ direction, the atomic distortion is not along the \(\alpha\) bond, and therefore the frequency softening is not yet constant as in \({\rm TO}_{ZZ}^{+}\). ## Electron-phonon coupling {#e-ph coupling-res} The computed values for the average e-ph coupling matrix-element square over the Fermi surface in pristine graphene are \(\langle g_{{\Gamma},G}^2\rangle = 0.0400 {\rm \ eV}^2\) and \(\langle g_{{\bf K},TO}^2\rangle = 0.0989 {\rm \ eV}^2\), which are in excellent agreement with previously reported values. We also obtain a value of \(0.0037 {\rm \ eV}^2\) for the double degenerate LO and LA branches at \({\bf K}\), which is very small in comparison with the TO branch. The effect of uniaxial strain on the e-ph coupling matrix element square over the Fermi surface is shown in Fig. [\[epc-fig\]](#epc-fig){reference-type="ref" reference="epc-fig"}. We report the evolution of \(\langle g^2\rangle\) for \({\rm G}^+\) and \({\rm G}^-\) at the \({\Gamma}\) point (Fig. [\[epc-fig\]](#epc-fig){reference-type="ref" reference="epc-fig"}a), meanwhile for the \({\bf q}_{ZZ}\) and \({\bf q}_{AC}\) points we analyze the TO, LO and LA branches (Fig. [\[epc-fig\]](#epc-fig){reference-type="ref" reference="epc-fig"}b). In the case of the LO branch at \({\bf q}_{ZZ}\) and \({\bf q}_{AC}\) we found that \(\langle g^2\rangle\) remains practically constant for both ZZ and AC strain, and for clarity it has not been included in Fig. [\[epc-fig\]](#epc-fig){reference-type="ref" reference="epc-fig"}. At the \({\Gamma}\) point, we found that after the splitting of the \(E_{2g}\) phonon mode under uniaxial strain, the e-ph coupling in \({\rm G}^+\) (\({\rm G}^-\)) slightly increases (decreases) with almost no dependence on the strain direction. The overall change at \({\Gamma}\) for 5% of uniaxial strain, considering the sum of both \({\rm G}^+\) and \({\rm G}^-\), shows a small reduction in \(\langle g_{\Gamma}^2\rangle\) of 0.0017 eV\(^2\), which corresponds to-2% with respect to the value for the pristine case. A more complex behavior takes place in \({\bf q}_{ZZ}\) and \({\bf q}_{AC}\) as a function of uniaxial strain: the \(\langle g^2\rangle\) in the TO branch starts to decrease with a non-linear dependence, while in the LA branch the e-ph coupling increases with almost the same rate as the corresponding one of the TO. In the same way as the phonon eigenvectors polarization behave under uniaxial strain, these changes occur faster in ZZ than AC, with the LA branch overcoming the TO after 3% of ZZ strain. However, if we take into account the sum of all the branches in \({\bf q}_{ZZ}\) and \({\bf q}_{AC}\), the total \(\langle g_{{\bf q}}^2\rangle\) for ZZ and AC strains are very similar, and increases only by 2% with respect to the pristine case. Therefore, considering both the \({\Gamma}\) and \({\bf q}_{ZZ}\) and \({\bf q}_{AC}\) contributions to the e-ph coupling, we have that the total \(\langle g^2\rangle\) in uniaxially strained graphene remains practically constant. To understand the trend shown in Fig. [\[epc-fig\]](#epc-fig){reference-type="ref" reference="epc-fig"}, it is important to note that \(\langle g^2\rangle \sim (\delta V)^2/\omega\) (see eqs. [\[gelements\]](#gelements){reference-type="ref" reference="gelements"} and [\[g2\]](#g2){reference-type="ref" reference="g2"}). Thus, considering only the contribution of \(\omega\), due to the phonon softening of the Kohn anomalies, one would expected an increment in \(\langle g^2\rangle\). However, from Fig. [\[epc-fig\]](#epc-fig){reference-type="ref" reference="epc-fig"} we can see that \(\langle g^2\rangle\) decreases for some phonon modes as a function of strain. On the other hand, it should be noted that for ZZ and AC strain the pattern of the atomic vibrations in the \(\Gamma\) point remains the same as in pristine graphene, even though the atomic distances \(\alpha\) and \(\beta\) are not equal. Meanwhile for \({\bf q}_{\rm ZZ}\) and \({\bf q}_{\rm AC}\), as a result of the Kohn anomaly shift from the high symmetry point \({\bf K}\), the mixing of the phonon eigenvectors for the TO and LA branches induces an important modification of the atomic vibration with respect to pristine graphene. Therefore, the behavior of \(\langle g^2\rangle\) as function of uniaxial strain is mainly due to the change in the polarization of the phonon eigenvectors, which contributes to the enhancement(reduction) of the e-ph coupling in the LA(TO) branch. Regarding the anomaly in the phonon dispersion for the LA branch at \({\bf q}_{ZZ}\) and \({\bf q}_{AC}\) discussed in Sec. [3.2](#results-Kohn anomaly){reference-type="ref" reference="results-Kohn anomaly"}, based on the substantial increment of \(\langle g^2\rangle\), it could be assigned to an emergent Kohn anomaly in uniaxially strained graphene. This feature is a major difference in the e-ph coupling between pristine and uniaxially strained graphene, due to the presence of a new intervalley phonon-scattering channel for electronic states close to the Dirac point, in addition to the TO branch. On the other hand, it was previously reported that in comparison to many-body theories which includes electronic correlation effects, standard DFT underestimates the e-ph coupling of the \(E_{2g}\) and \(A_1'\) phonon modes of graphene. Therefore, the inclusion of such many-body effects could give rise to a stronger Kohn anomaly and e-ph coupling than our results. However, the use of linear response theory to compute the phonon dispersion and e-ph coupling are, at present, not implemented in many-body methodologies such as GW. Here we have shown that uniaxial strain induces a non-negligible Kohn anomaly even for small strain rates, which opens the possibility to be experimentally observed. It would be even more interesting to evaluate the contribution of this anomaly to those graphene properties which depend on the e-ph coupling. For example, it could be important to determine if this new Kohn anomaly contributes to the splitting of the double resonance Raman scattering 2D-band, or to the intrinsic electronic resistivity, where until now only the optical \(A'_1\) intervalley phonon mode was considered. # Conclusions {#conclusions-sec} We have performed a first principles study of the structural properties, Kohn anomalies, and e-ph coupling for uniaxially strained graphene in the ZZ and AC directions. For ZZ strain we found a small contraction of the bond perpendicular to the strain direction, that increases the corresponding force constant. Evaluating the shift of the Dirac point from \({\bf K}\), the phonon nesting vectors \({\bf q}_{ZZ}\) and \({\bf q}_{AC}\) were calculated. Analyzing the phonon dispersion we found that a Kohn anomaly in \({\bf q}_{ZZ}\) and \({\bf q}_{AC}\) emerges as a function of the uniaxial strain, in the LA branch. For both, the original Kohn anomaly in the TO branch and the new anomaly in the LA branch, there is a polarization of the phonon eigenvectors as induced by the strain, in directions parallel and perpendicular to the applied strain, in the same way as is known to occur for \({\rm G}^+\) and \({\rm G}^-\) at \({\Gamma}\). The softening of frequency in the Kohn anomaly shows a linear behavior as a function of the strain, except for the TO branch which shows a non-linear softening, and becomes almost constant for ZZ strains higher than 3%. From the analysis of the average e-ph coupling matrix element square over the Fermi surface as a function of the uniaxial strain, for the \({\Gamma}\) point we found that the strain has a small effect on the \({\rm G}^+\) and \({\rm G}^-\) phonon modes. For the TO branch there is a reduction of \(\langle g^2\rangle\) at \({\bf q}_{ZZ}\) and \({\bf q}_{AC}\), while for the LA branch there is a large enhancement of the e-ph coupling as a function of strain. Such behavior is mainly a consequence of the change in the polarization of the phonon eigenvectors because of mixing of the LA and TO modes induced by the uniaxial strain. Finally, it is important to emphasize that uniaxial strain in graphene induces a Kohn anomaly and enhancement of the e-ph coupling in the LA phonon branch, in contrast with the view commonly assumed that the e-ph coupling around the \({\bf K}\) point is present only in the TO phonon branch. [^1]: We use the definition of e-ph coupling in graphene following the work of Piscanec et al. The details are explained in Appendix [\[epc-definition\]](#epc-definition){reference-type="ref" reference="epc-definition"}.
{'timestamp': '2016-07-19T02:01:11', 'yymm': '1607', 'arxiv_id': '1607.04672', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04672'}
null
null
# Introduction. {#intro} # Introduction A central modeling primitive in mathematical optimization is the disjunctive constraint: any feasible solution must satisfy at least one of some fixed, finite collection of alternatives. This type of constraint is general enough to capture structures as diverse as boolean satisfiability, complementarity constraints, special ordered sets, and (bounded) integrality. The special case of polyhedral disjunctive constraints corresponds to the form \[\label{disconst} x \in \bigcup\nolimits_{i=1}^d P^i,\] where we have that each \(P^i\subseteq \mathbb{R}^n\) is a polyhedron. In this work, we will focus on \(\mathscr{V}\)-polyhedra; that is, we have a description of the \(P^i\) in terms of their extreme points \(\operatorname{ext}(P^i)\). We are particularly interested in the case where constraint [\[disconst\]](#disconst){reference-type="eqref" reference="disconst"} is *primitive*, or a basic building block for a much more complex optimization problem. For this reason, we are interested in modeling it in a generic, composable way. In particular, if [\[disconst\]](#disconst){reference-type="eqref" reference="disconst"} is embedded in a larger, more complex optimization problem \[\label{eqn:larger-optimization-problem} \min_{(x,y) \in Q: \eqref{disconst}} f(x,y),\] we hope for a mathematical description sufficiently structured such that we may use more advanced algorithmic approaches, beyond naïve enumeration, to solve [\[eqn:larger-optimization-problem\]](#eqn:larger-optimization-problem){reference-type="eqref" reference="eqn:larger-optimization-problem"}. In particular, \(Q\) could be described by any number of different types of constraints: linear inequalities, conic constraints, integrality conditions, or additional disjunctive constraints. In this context, it is well known that merely constructing the convex hull \(\operatorname{Conv}(\bigcup_{i=1}^d P^i)\) is not sufficient for solving [\[eqn:larger-optimization-problem\]](#eqn:larger-optimization-problem){reference-type="eqref" reference="eqn:larger-optimization-problem"}; we will need a formulation for \(\bigcup_{i=1}^d P^i\) directly. Mixed-integer programming (MIP) has emerged as an incredibly expressive modeling methodology, with advanced computational methods capable of solving many problems of practical interest, often at very large scale. Constraint [\[disconst\]](#disconst){reference-type="eqref" reference="disconst"} with polyhedral sets \(P^i\) is particularly well-suited for a mixed-integer programming approach. Indeed, standard formulations for [\[disconst\]](#disconst){reference-type="eqref" reference="disconst"} were presented in, and are *ideal*, or as strong as possible with respect to their continuous linear programming relaxations (see Section [5.1](#ss:formulation-definitions){reference-type="ref" reference="ss:formulation-definitions"} for a formal definition). However, this formulation requires introducing \(d\) auxiliary binary variables, which may be impractically large, especially in the context of the larger problem [\[eqn:larger-optimization-problem\]](#eqn:larger-optimization-problem){reference-type="eqref" reference="eqn:larger-optimization-problem"}. However, it is sometimes possible to construct ideal formulations with considerably fewer auxiliary variables. In particular, a string of recent work has presented ideal formulations for certain highly structured constraints such as SOS2 with only \(\mathscr{O}(\log(d))\) auxiliary binary variables and additional constraints (excluding variable bounds). Moreover, these formulations have proven practically useful, and indeed are the most performant by a significant margin for a large swath of instances of the problem classes to which they have been applied. However, building these formulations is complex and ad-hoc, hindering the construction, analysis, and implementation of such formulations for new constraints. One relatively generic and versatile approach to construct small ideal formulations is the *independent branching* (IB) scheme framework introduced by Vielma and Nemhauser. The approach is to find some (particularly structured) polyhedra \(Q^{1,j}\) and \(Q^{2,j}\) such that [\[disconst\]](#disconst){reference-type="eqref" reference="disconst"} can be rewritten as \[\label{eqn:disjunction-via-IB} \bigcup\nolimits_{i=1}^d P^i = \bigcap\nolimits_{j=1}^t \left(Q^{1,j} \cup Q^{2,j}\right).\] This represents the disjunctive constraint in term of a series of simple choices between two alternatives. Given such a representation, it is often straightforward to construct a simple, small, and ideal formulation for [\[disconst\]](#disconst){reference-type="eqref" reference="disconst"} by formulating each of the \(t\) alternatives separately, and then combining them. Furthermore, when the polyhedra \(P^i\) are \(\mathscr{V}\)-polyhedra, the construction of the independent branching scheme-based formulation is purely combinatorial, based on the extreme points that are shared between the different polyhedra \(P^i\). As we will see, we can therefore approach formulating [\[disconst\]](#disconst){reference-type="eqref" reference="disconst"} combinatorially, by studying the shared structure amongst the extreme points. In this work we generalize and provide a systematic study of the applicability and limitations of the independent branching approach. The contributions of this work can be categorized in the following way. 1. We generalize the notion of independent branching schemes to allow for multiple alternatives, and provide an exact characterization of when there exists *any* independent branching representation for [\[disconst\]](#disconst){reference-type="eqref" reference="disconst"}, in terms of the graphical representation of the shared extreme points amongst the polyhedra \(P^i\). In particular: 1. We demonstrate that the widely-used cardinality constraints cannot be expressed by any independent branching scheme with few alternatives. We argue that this negative result provides theoretical justification for the practical observation that both MIP formulations and simple constraint branching schemes struggle with modeling cardinality constraints effectively. 2. We show that arbitrary piecewise linear functions in the plane can be modeled with at most three alternatives, and provide a polynomial-time verifiable condition for representability with two alternatives. 3. We argue that nonconvex polygonal set avoidance constraints are always representable with two alternatives. 2. We provide an exact characterization for when there exists a two-alternative independent branching representation for [\[disconst\]](#disconst){reference-type="eqref" reference="disconst"} of size \(t\), in terms of the classical biclique covering problem. This relation allows the *algorithmic* construction of small independent branching formulations for [\[disconst\]](#disconst){reference-type="eqref" reference="disconst"}. In particular, we study and apply simple properties of biclique covers and their composition to systematically construct explicit descriptions of small independent branching formulations for special structures. 3. We apply our framework to a variety of constraints of the form [\[disconst\]](#disconst){reference-type="eqref" reference="disconst"} to give an indication of the expressive power of the IB scheme approach and the advantage of using biclique cover techniques to construct formulations. In particular: 1. We review and develop several generic properties of biclique covers that lead to systematic construction techniques. Using these techniques we construct explicit, small (i.e. logarithmic in \(d\)), ideal formulations for generalizations of the special ordered sets of Beale and Tomlin, piecewise linear functions over arbitrary 2-dimensional grid triangulations, and outer-approximating discretizations of multilinear terms. This last formulation generalizes the popular logarithmically-sized formulation of Misener et al. for nonconvex quadratic optimization, which we show is not ideal in general. 2. We provide matching lower bounds for these constructions, showing that they are asymptotically optimal with respect to the size of *any* possible MIP formulation. # Preliminaries: Definitions, notation, and nomenclature {#prelimsec} A (bounded) \(\mathscr{V}\)-polyhedra (or polyhedra in \(\mathscr{V}\)-form) is a set \(P \subset \mathbb{R}^n\) that can be expressed as \[P = \operatorname{Conv}(V) \defeq \left\{\sum_{v \in V} \lambda_v v: \lambda \in \Delta^V \right\}\] for some finite set of vectors \(V \subset \mathbb{R}^n\), where \(\Delta^V \defeq \{\lambda \in \mathbb{R}_{+}^V: \sum_{v \in V} \lambda_v = 1\}\) is the standard simplex. According to the celebrated Minkowski-Weyl Theorem (e.g. ), any polyhedral disjunctive constraint [\[disconst\]](#disconst){reference-type="eqref" reference="disconst"} can be expressed as the union of \(\mathscr{V}\)-polyhedra in terms of their extreme points \(\operatorname{ext}(P^i)\)[^1]. When the disjunctive constraint [\[disconst\]](#disconst){reference-type="eqref" reference="disconst"} is a union of \(\mathscr{V}\)-polyhedra, it suffices to consider only the combinatorial structure of the extreme points of the polyhedra \(P^i\). To see why, consider \(J = \bigcup_{i=1}^d \operatorname{ext}(P^i)\) as the *ground set* and take \(\mathscr{S} = \{\operatorname{ext}(P^i)\}_{i=1}^d \subseteq 2^J\) as the collection of extreme points for each of the polyhedra. We can then define a corresponding disjunctive constraint that is purely combinatorial on the sets \(\mathscr{S}\). Combinatorial disjunctive constraints may also appear as natural primitive constraints that do not explicitly arise from unions of \(\mathscr{V}\)-polyhedra, as we will see in Sections [4.2](#soskintrosec){reference-type="ref" reference="soskintrosec"} and [4.3](#cardintrosec){reference-type="ref" reference="cardintrosec"}. However, when they do arise from unions of \(\mathscr{V}\)-polyhedra, it is straightforward to construct a corresponding formulation for [\[disconst\]](#disconst){reference-type="eqref" reference="disconst"} as \[\label{eqn:disaggregated-CDC-to-DC} \left\{\sum_{v \in J} \lambda_v v \:: \: \lambda \in \operatorname{CDC}(\mathscr{S}) \right\}.\] One advantage of this approach is that formulation [\[eqn:disaggregated-CDC-to-DC\]](#eqn:disaggregated-CDC-to-DC){reference-type="eqref" reference="eqn:disaggregated-CDC-to-DC"} allows us to divorce the problem-specific data (i.e. the values \(v \in J\)) from the underlying combinatorial structure encapsulated in \(\operatorname{CDC}(\mathscr{S})\). As such, we can construct a single, strong formulation for a given structure \(\operatorname{CDC}(\mathscr{S})\) and this formulation will remain valid for transformations of the data, so long as this transformation sufficiently preserves the combinatorial structure of \(\operatorname{CDC}(\mathscr{S})\). For instance, if \(\set{P^i}_{i=1}^d\) are the polyhedra for the original constraint represented by \(\operatorname{CDC}(\mathscr{S})\), and \(\{\hat{P}^i\}_{i=1}^d\) are those associated with the new data, then a sufficient condition for the formulation of \(\operatorname{CDC}(\mathscr{S})\) yielding a valid formulation for \(\bigcup_{i=1}^d \hat{P}^i\) is the existence of a bijection \(\pi:J \to \hat{J}\) (with \(J = \bigcup\nolimits_{i=1}^d \operatorname{ext}(P^i)\) and \(\hat{J} = \bigcup\nolimits_{i=1}^d \operatorname{ext}(\hat{P}^i)\)) such that \[\label{biyectioncond} v \in \operatorname{ext}(P^i) \Longleftrightarrow \pi(v) \in \operatorname{ext}(\hat{P}^i) \quad \forall i \in \llbracket d \rrbracket, \: v \in J,\] where \(\llbracket d \rrbracket\defeq \{1,\ldots,d\}\). In this way, we can construct a single small, strong formulation for \(\operatorname{CDC}(\mathscr{S})\), and use it repeatedly for many different "combinatorially equivalent" instances of the same constraint. We note that one subtle disadvantage of this data-agnostic approach is that, even if condition [\[biyectioncond\]](#biyectioncond){reference-type="eqref" reference="biyectioncond"} is satisfied, the resulting formulation for \(\bigcup_{i=1}^d \hat{P}^i\) may be larger than necessary. An extreme manifestation of this would be when the new polyhedra \(\{\hat{P}^i\}_{i=1}^d\) are such that \(\hat{P}_i\subseteq \hat{P}_1\) for all \(i\in \llbracket d \rrbracket\). In this case, \(\bigcup\nolimits_{i=1}^d \hat{P}^i = \hat{P}^1\), and so formulating this does not require a MIP formulation at all. Less pathological cases could occur where some subset of the disjunctive sets become redundant after changing the problem data. However, we note that in many of the applications considered in this work, the combinatorial representation leads to redundancy of this form only in rare pathological cases (e.g. Sections [4.1](#sos2introsec){reference-type="ref" reference="sos2introsec"} and [4.5](#gridintrosec){reference-type="ref" reference="gridintrosec"}). In the remaining cases we will take care to consider, for example, the geometric structure of the data before constructing the disjunctive constraint (e.g. Section [7.2](#ss:partitions-of-the-plane){reference-type="ref" reference="ss:partitions-of-the-plane"}). Finally, we also note that if we wish to model the case where the polyhedra \(P^i\) are unbounded, a result of Jeroslow and Lowe tells us that we may only construct a (binary) MIP formulation for [\[disconst\]](#disconst){reference-type="eqref" reference="disconst"} if the recession cones coincide for each \(P^i\). In the case this condition is met, we may formulate [\[disconst\]](#disconst){reference-type="eqref" reference="disconst"} with \[\label{eqn:disaggregated-CDC-to-DC-with-rays} \left\{\sum_{v \in J} \lambda_v v + \sum_{r \in R} \mu_r r \:: \: \lambda \in \operatorname{CDC}(\mathscr{S}), \: \mu \in \mathbb{R}^R_{+} \right\},\] where \(R\) is the shared set of extreme rays for each of the \(P^i\). Therefore, we will restrict our attention to the case where each of the \(P^i\) are bounded, as formulating the unbounded case is a straightforward extension. In the remainder of the paper, we will make the following assumptions on \(\mathscr{S}\) that are without loss of generality. We will say that a set \(S \subseteq J\) is a *feasible set* with respect to \(\operatorname{CDC}(\mathscr{S})\) if \(Q(S) \subseteq \operatorname{CDC}(\mathscr{S})\) (equivalently, if \(S \subseteq T\) for some \(T \in \mathscr{S}\)) and that it is an *infeasible set* otherwise. # Motivating examples {#sec:motivating-examples} As mentioned, we can represent any polyhedral disjunctive constraint [\[disconst\]](#disconst){reference-type="eqref" reference="disconst"} as the union of \(\mathscr{V}\)-polyhedra. However, there are many disjunctive constraints for which the \(\mathscr{V}\)-form of [\[disconst\]](#disconst){reference-type="eqref" reference="disconst"} is especially natural. We now present some as running examples that we will return to throughout. One motif to appear repeatedly will be the graph of a continuous piecewise linear function. That is, given some bounded domain \(\Omega \subset \mathbb{R}^n\) and some polyhedral partition \(\bigcup_{i=1}^d P^i = \Omega\) (i.e. the relative interiors do not overlap[^2]), we are interested in modeling a continuous function \(f: \Omega \rightarrow \mathbb{R}\) such that \(x \in P^i \Longrightarrow f(x) = a^i \cdot x + b_i\) for some appropriate \(a^i \in \mathbb{R}^n\) and \(b \in \mathbb{R}\). In order to model the graph \(\operatorname{gr}(f;\Omega) \defeq \{(x,f(x)): x \in \Omega\}\), we can construct a formulation for \(\operatorname{CDC}(\mathscr{S})\) (with \(\mathscr{S} = \{\operatorname{ext}(P^i)\}_{i=1}^d\)) and express \[\label{eqn:pwl-graph} \operatorname{gr}(f) = \left\{ \sum_{v \in J} \lambda_v (v,f(v)): \lambda \in \operatorname{CDC}(\mathscr{S}) \right\},\] where we will use the notation \(\operatorname{gr}(f) \equiv \operatorname{gr}(f;\Omega)\) when \(\Omega\) is clear from context.[^3] ## Univariate piecewise linear functions and the SOS2 constraint {#sos2introsec} Consider a univariate (nonconvex) piecewise linear function \(f\) characterized by \(N\) breakpoints \(x^1 < x^2 < \ldots < x^N\). We may model the graph of this function via [\[eqn:pwl-graph\]](#eqn:pwl-graph){reference-type="eqref" reference="eqn:pwl-graph"}, where \(\mathscr{S} = \{\{x^j,x^{j+1}\}: j \in \llbracket N-1 \rrbracket\}\). As long as the ordering of the breakpoints is preserved, it is easy to see that condition [\[biyectioncond\]](#biyectioncond){reference-type="eqref" reference="biyectioncond"} will be satisfied for any transformations of the problem data. Furthermore, the only case in which knowledge of the specific data \(\{(x^j,f(x^j))\}_{j=1}^n\) allows the simplification of the original disjunctive representation of \(\operatorname{gr}(f)\) is when \(f\) is affine in on two adjacent intervals, e.g. affine over \([x^j,x^{j+2}]\) for some \(j \in \llbracket N-2 \rrbracket\). Therefore, the potential disadvantage of disregarding the specific data when formulating the constraint occurs only in rare pathological cases which are easy to detect. For this reason, we will strip out the problem data and instead express the constraint with respect to the indices \(j\) of the vertices \(\{x^j\}_{j=1}^N\). That is, we take \(J = \llbracket N \rrbracket\) and write \(\mathscr{S} = \{\{\tau,\tau+1\}: \tau \in \llbracket N-1 \rrbracket\}\) to emphasize the data independence of the constraint and highlight the combinatorial structure. In this form, we can recognize the special ordered set of type 2 (SOS2\((N)\)) constraint of Beale and Tomlin, which requires that at most two components of \(\lambda\) may be nonzero, and that these nonzero components must be consecutive in the ordering on \(J\). ## SOS\(k\) {#soskintrosec} A generalization of the special ordered sets considers the case where at most \(k\) consecutive components of \(\lambda\) may be nonzero at once. In particular, if \(J = \llbracket N \rrbracket\), we have \(\mathscr{S} = \{\{\tau,\tau+1,\ldots,\tau+k-1\}: \tau \in \llbracket N-k + 1 \rrbracket\}\). This constraint may arise, for example, in chemical process scheduling problems, where an activated machine may only be on for \(k\) consecutive time units and must produce a fixed quantity during that period. ## Cardinality constraints {#cardintrosec} An extremely common constraint in optimization is the cardinality constraint of degree \(\ell\), where at most \(\ell\) components of \(\lambda\) may be nonzero. This corresponds to \(\mathscr{S} = \{I \subseteq J: |I| = \ell\}\). A particularly compelling application of the cardinality constraint is in portfolio optimization, where it is often advantageous to limit the number of investments to some fixed number \(\ell\) to minimize transaction costs, or to allow differentiation from the performance of the market as a whole. ## Discretizations of multilinear terms {#multilinearsec} Consider a multilinear function \(f(x_1,\ldots,x_\eta) = \prod_{i=1}^\eta x_i\) defined over some box domain \(\Omega \defeq [l,u] \subset \mathbb{R}^\eta\). This function appears often in optimization models, but is nonconvex, and often leads to problems which are difficult to solve to global optimality in practice. As a result, computational techniques will often "relax" the graph of the function \(\operatorname{gr}(f)\) with a convex outer approximation, which is easier to optimize over.[^4] For the bilinear case (\(\eta=2\)), the well-known McCormick envelope describes the convex hull of \(\operatorname{gr}(f)\). Although traditionally stated in an inequality description, we may equivalently describe the convex hull via its four extreme points, which are readily available in closed form. For higher-dimensional multilinear terms, the convex hull has \(2^\eta\) extreme points, and can be constructed in a similar manner (e.g. see equation (3) in  and the associated references). Misener et al. propose a computational technique for optimizing problems with bilinear terms where, instead of modeling the graph over a single region \(\Omega = [l,u] \subset \mathbb{R}^2\), they discretize the region in a regular fashion and apply the McCormick envelope to each subregion. They model this constraint as a union of polyhedra, where each subregion enjoys a tighter relaxation of the bilinear term. Additionally, they propose a logarithmically-sized formulation for the union. However, it is not ideal (see Appendix [\[app:misener\]](#app:misener){reference-type="ref" reference="app:misener"}), it only applies for bilinear terms (\(\eta=2\)), and it is specialized for a particular type of discretization (namely, only discretizing along one component \(x_1\), and with constant discretization widths \(\{h^1_{j+1}-h^1_{j}\}_{j=1}^{d_1-1}\)). For a more general setting, we have that the extreme points of the convex hull of the graph \(\operatorname{Conv}(\operatorname{gr}(f))\) are given by \(\{(x,f(x)): x \in \operatorname{ext}(\Omega)\}\), where it is easy to see that \(\operatorname{ext}(\Omega) = \prod_{i=1}^\eta \{l_i,u_i\}\). Consider a grid imposed on \([l,u] \subset \mathbb{R}^\eta\); that is, along each component \(i \in \llbracket \eta \rrbracket\), we partition \([l_i,u_i]\) along the points \(l_i \equiv h_1^i < h_2^i < \cdots < h_{d_i-1}^i < h_{d_i}^i \equiv u_i\). This yields \(\prod_{i=1}^\eta \bra{d_i-1}\) subregions; denote them by \(\mathscr{R} \defeq \left\{\prod_{i=1}^\eta [h^i_{k_i},h^i_{k_i+1}]: k \in \prod_{i=1}^\eta \llbracket d_i-1 \rrbracket\right\}\). We can then take the polyhedral partition of \(\Omega\) given by \(P^R \defeq \operatorname{Conv}(\operatorname{gr}(f;R))\) for each subregion \(R\), the sets as \(\mathscr{S} = \left\{\operatorname{ext}(P^R)\right\}_{R \in \mathscr{R}}\), and the ground set as \(J = \bigcup\{S \in \mathscr{S}\}\). In particular, we have that \(J = \prod_{i=1}^\eta \{h^i_1,\ldots,h^i_{d_i}\}\). Analogously to the notational simplification we took with the SOS2 constraint, for the remainder we will take \(J=\prod_{i=1}^\eta \llbracket d_i \rrbracket\) and \(\mathscr{S}=\set{\prod_{i=1}^\eta \set{k_i,k_i+1}: k \in \prod_{i=1}^\eta \llbracket d_i-1 \rrbracket}\). We also note that condition [\[biyectioncond\]](#biyectioncond){reference-type="eqref" reference="biyectioncond"} is satisfied as long as the ordering of the discretization is respected along each dimension. ## Piecewise linear functions in the plane and grid triangulations {#gridintrosec} Consider a (potentially nonconvex) region \(\Omega \subset \mathbb{R}^2\). We would like to model a (also potentially nonconvex) piecewise linear function \(f\) with domain over \(\Omega\). Take \(\{P^i\}_{i=1}^d\) as the set of pieces of the domain, and the corresponding ground set as \(J = \bigcup_{i=1}^d \operatorname{ext}(P^i)\) and sets as \(\mathscr{S} = \{\operatorname{ext}(P^i)\}_{i=1}^d\). We may then model the piecewise linear function via the graph representation [\[eqn:pwl-graph\]](#eqn:pwl-graph){reference-type="eqref" reference="eqn:pwl-graph"}. An important special case occurs when the function \(f\) is affine over a triangulation of a grid similar to the one used for multilinear terms. A description of the associated CDC that emphasizes its combinatorial structure can be obtained through the same simplification from Section [4.4](#multilinearsec){reference-type="ref" reference="multilinearsec"} as follows. Consider a rectangular region in the plane \(\Omega = [1,M] \times [1,N]\), and the regular grid points \(J = \{1,\ldots,M\} \times \{1,\ldots,N\}\). A *grid triangulation* \(\mathscr{S}\) of \(\Omega\) is then a set \(\mathscr{S}\) where: - Each \(S \in \mathscr{S}\) is a triangle: \(|S|=3\). - \(\mathscr{S}\) partitions \(\Omega\): \(\bigcup_{S \in \mathscr{S}} \operatorname{Conv}(S) = \Omega\) and \(\operatorname{relint}(\operatorname{Conv}(S)) \cap \operatorname{relint}(\operatorname{Conv}(T)) = \emptyset\) for each distinct \(S, T \in \mathscr{S}\). - \(\mathscr{S}\) is on a regular grid: \(S \subset J\) for each \(S \in \mathscr{S}\), and \(||v-w||_\infty \leq 1\) for each \(v,w \in S\). As concrete examples, consider Figure [\[fig:triangulations\]](#fig:triangulations){reference-type="ref" reference="fig:triangulations"}, where we depict three different triangulations with \(M=N=3\). Grid triangulations are often used to model bivariate, non-separable, piecewise linear functions using [\[eqn:pwl-graph\]](#eqn:pwl-graph){reference-type="eqref" reference="eqn:pwl-graph"} . Furthermore, any formulation constructed for a given grid triangulation can be readily applied to any other grid triangulation obtained by shifting the grid points in the plane, so long as the resulting triangulation is strongly isomorphic to, or compatible with, the original triangulation. ## Obstacle avoidance {#obstacleavoidancesec} Consider an unmanned aerial vehicle (UAV) which you would like to navigate through an area with fixed obstacles. At any given time, you wish to impose the constraint that the location of the vehicle \(x \in \mathbb{R}^2\) must lie in some (nonconvex) region \(\Omega \subset \mathbb{R}^2\), which is the plane, less any obstacles in the area. MIP formulations of this constraint has received interest as a useful primitive for path planning. We may model \(x \in \Omega\) by partitioning the region \(\Omega\) with polyhedra such that \(\Omega = \bigcup_{i=1}^d P^i\). Traditional approaches to modeling constraints [\[disconst\]](#disconst){reference-type="eqref" reference="disconst"} of this form use a linear inequality description for each of the polyhedra \(P^i\) and construct a corresponding big-\(M\) formulation, which will not be ideal in general. In the \(\mathscr{V}\)-polyhedra framework, we will instead take the \(P^i\) as \(\mathscr{V}\)-polyhedra that partition \(\Omega\), and be able to construct small, ideal formulations. # MIP formulations for combinatorial disjunctive constraints Using standard MIP formulation techniques, we now present formulations for \(\operatorname{CDC}(\mathscr{S})\) for comparison with our approach. In particular, we will argue that the framework we will present later can lead to formulations that are smaller (in terms of the number of auxiliary variables), and enjoy other favorable properties we enumerate in Section [6](#ss:IB-schemes){reference-type="ref" reference="ss:IB-schemes"}. ## MIP formulations: Definitions, size, and strength {#ss:formulation-definitions} Formally, we say that a (binary) MIP formulation \(F\) for a constraint \(x \in Q \subseteq \mathbb{R}^{n_1}\) is the composition of linear inequalities (the *linear programming (LP) relaxation*, or just *relaxation* in the context of this work) \[R \defeq \left\{ (x,y,z) \in [l^x,u^x] \times [l^y,u^y] \times [0,1]^{n_3}: Ax + By + Cz \leq d \right\}\] with (binary) integrality conditions \[\label{eqn:generic-MIP-formulation} F \defeq R \cap \left(\mathbb{R}^{n_1} \times \mathbb{R}^{n_2} \times \{0,1\}^{n_3}\right)\] such that \(\operatorname{Proj}_x(F) = Q\).[^5] Finally, we assume \(R\) is line-free (i.e. has at least one extreme point), which is satisfied by essentially all practical formulations. Throughout, we will be interested in ways of understanding both the strength of a given formulation, as well as in quantifying the size or complexity of a formulation. We say that a formulation is *ideal* if each extreme point of the relaxation naturally satisfies the integrality conditions, i.e. \(\operatorname{Proj}_z(\operatorname{ext}(R)) \subseteq \{0,1\}^{n_3}\). The choice of name is apt, as this is the strongest possible MIP formulation we can expect.[^6] As a measure of the complexity of the formulation, we count the number of auxiliary continuous variables \(y\) and continuous binary variables \(z\) used by the formulation, as well as the number of inequalities in our description of \(R\). We ignore the size of \(x\), since this is intrinsic to the constraint we wish to model. We will say that a formulation is *extended* if there are auxiliary continuous variables \(y\) in the representation \(F\) (that is, \(n_2 > 0\)) and *non-extended* otherwise (\(n_2 = 0\)). Furthermore, as suggested by the definition of \(R\), we distinguish between variable bounds (e.g. \(l^y \leq y \leq u^y\)) and general inequalities (\(Ax + By + Cz \leq d\)), as modern MIP solvers are able to incorporate variable bounds with minimal extra computational cost. Finally, we note that from now on (with a single exception), any statements regarding formulation properties or size will be with respect to the combinatorial disjunctive constraint \(\operatorname{CDC}(\mathscr{S})\), rather than the \(\mathscr{V}\)-polyhedral disjunctive constraints itself. In other words, all statements are with respect a fixed combinatorial structure, with no possible simplifications from the specific data as discussed in Section [3](#prelimsec){reference-type="ref" reference="prelimsec"}. The sole exception will be in Section [7.2](#ss:partitions-of-the-plane){reference-type="ref" reference="ss:partitions-of-the-plane"}, where we will make use of some geometric properties inherent in a given realization of the data. ## Existing formulations for combinatorial disjunctive constraints A standard formulation for \(\operatorname{CDC}(\mathscr{S})\) adapted from Jeroslow and Lowe is This formulation has \(d=|\mathscr{S}|\) auxiliary binary variables, \(\sum_{S \in \mathscr{S}} |S|\) auxiliary continuous variables, and no general inequalities. Additionally, it is ideal. Using Proposition 9.3 from, we can construct an ideal MIP formulation with fewer auxiliary binary variables: where \(\{h^S\}_{S \in \mathscr{S}} \subseteq \{0,1\}^r\) is some set of distinct binary vectors. This formulation is actually a generalization of [\[eqn:jeroslow-formulation\]](#eqn:jeroslow-formulation){reference-type="eqref" reference="eqn:jeroslow-formulation"}, which we recover if we take \(h^S = \textbf{e}^S \in \mathbb{R}^\mathscr{S}\) as the canonical unit vectors. If instead we take \(r\) to be as small as possible (while ensuring that the vectors \(\{h^S\}_{S \in \mathscr{S}}\) are distinct), we recover \(r = \lceil \log_2(d) \rceil\). Therefore, formulation [\[eqn:prop-9.3\]](#eqn:prop-9.3){reference-type="eqref" reference="eqn:prop-9.3"} yields an ideal extended formulation for [\[disconst\]](#disconst){reference-type="eqref" reference="disconst"} with \(\lceil \log_2(d) \rceil\) auxiliary binary variables, \(\sum_{S \in \mathscr{S}} |S|\) auxiliary continuous variables, and no general inequalities. The following corollary shows that this is the smallest number of auxiliary binary variables we may hope for. See Appendix [\[app:log-bound\]](#app:log-bound){reference-type="ref" reference="app:log-bound"}. The formulations thus far have been extended formulations, as they are constructed by formulating each polyhedra \(Q\bra{S}\) separately and then aggregating them, rather than working with the combinatorial structure underlying the shared extreme points. Therefore, each of these formulations requires a copy of the multiplier \(\gamma^S_v\) for each set \(S \in \mathscr{S}\) for which \(v \in S\), and so \(\sum_{i=1}^d |\mathscr{S}|\) auxiliary continuous variables total. In contrast, we can construct non-extended formulations for \(\operatorname{CDC}(\mathscr{S})\) that work directly on the \(\lambda\) variables and the underlying combinatorial structure of \(\mathscr{S}\). An example of a non-extended formulation for CDC is the widely used ad-hoc formulation (see and the references therein) given by This formulation is not necessarily ideal, and it requires no auxiliary continuous variables, \(d\) auxiliary binary variables, and \(|J|\) general inequalities. In summary, we have seen an ideal extended formulation [\[eqn:prop-9.3\]](#eqn:prop-9.3){reference-type="eqref" reference="eqn:prop-9.3"} for \(\operatorname{CDC}(\mathscr{S})\) with relatively few auxiliary binary variables, but relatively many auxiliary continuous variables. On the other end of the spectrum, we have a non-extended formulation [\[eqn:disaggregated\]](#eqn:disaggregated){reference-type="eqref" reference="eqn:disaggregated"} with no auxiliary continuous variables, but which requires relatively many auxiliary binary variables and which may fail to be ideal. However, we know that in special cases we can construct ideal, non-extended formulations with only \(\mathscr{O}(\log(d))\) auxiliary variables and constraints (e.g. SOS1, SOS2, and particular 2-dimensional grid triangulations ). This work provides a framework for constructing such *small, strong*, non-extended MIP formulations for \(\operatorname{CDC}(\mathscr{S})\), which are automatically ideal, and in the best case will have \(\mathscr{O}(\log(d))\) auxiliary binary variables and general inequality constraints. # Independent branching schemes {#ss:IB-schemes} Vielma and Nemhauser introduced the notion of an *independent branching scheme* as a natural framework for constructing formulations for combinatorial disjunctive constraints. The independent branching scheme is a logically equivalent way of expressing a CDC in terms of a conjunction of dichotomies: that is, as a series of choices between two (simple) options. This approach is parsimonious: if you are given an independent branching scheme for a particular CDC, it is straightforward to construct an ideal formulation whose size is on the order of the number of dichotomies. For our purposes, we present a generalized notion, where we allow potentially more than two alternatives. We say that such an IB scheme has *depth \(t\)*, and that each \(j \in \llbracket t \rrbracket\) yields a corresponding *level* of the IB scheme \(\bigcup_{i=1}^k Q(L^j_i)\), given by the \(k\) *alternatives* \(Q(L^j_i)\). An equivalent way of understanding these representations, which we will be using for the remainder of this work, is by eschewing the polyhedra \(Q(L^j_i)\) and working directly on the underlying set \(L^j_i\). That is, a valid \(k\)-way IB scheme satisfies the condition that \[T \subseteq J \text{ is a feasible set} \Longleftrightarrow \forall j \in \llbracket t \rrbracket, \: \exists i \in \llbracket k \rrbracket \text{ s.t. } T \subseteq L^j_i.\] First, we observe that, due to our assumption that \(\mathscr{S}\) covers the ground set, we have that for each element \(v \in J\) and level \(j \in \llbracket t \rrbracket\), there will be at least one alternative \(i \in \llbracket k \rrbracket\) such that \(v \in L^j_i\). We will use this extensively in the analysis to come, as it simplifies some otherwise tedious case analyses. Second, we see that this definition can capture potential schemes with a variable number of alternatives in each level by adding empty alternatives \(L^j_i=\emptyset\), provided we take \(k\) as the maximum number of alternatives for all levels. For notational simplicity, we say that a 2-way IB scheme is a *pairwise IB scheme*, and in this case we write the sets as \(\{(L^{j},R^{j})\}_{j=1}^t\) as in. In contrast, we will call the case with \(k > 2\) a *multi-way IB scheme*. In this form, we have replaced the monolithic constraint \(\operatorname{CDC}(\mathscr{S})\) by \(t\) constraints, each of which require the selection between \(k\) alternatives. We may then use standard techniques to construct a corresponding mixed-integer formulation. If \(S \in \mathscr{S}\), it is obviously a feasible set, and so we have immediately that \(S\) is an independent set in \(H^c_\mathscr{S}\). If it is not maximal, then we could add some \(v \in J \backslash S\) and maintain feasibility, which would violate our irredundancy assumption (i.e. \(S \cup \{v\} \subseteq S' \in \mathscr{S}\), \(S\in \mathscr{S}\), and \(S \subsetneq S \cup \{v\} \subseteq S'\)). If \(S\) is a maximal independent set in \(H^c_\mathscr{S}\), then it must be a feasible set with respect to \(\mathscr{S}\) as well. As it is maximal, there is no set \(\hat{S} \in \mathscr{S}\) with \(S \subsetneq \hat{S}\), and so we must have \(S \in \mathscr{S}\) as well. To show the "if" direction, it suffices to show the validity of [\[basicIB\]](#basicIB){reference-type="eqref" reference="basicIB"}. First note that every minimally infeasible set \(E^j \in \mathcal{E}_\mathscr{S}\) is rendered infeasible by level \(j\), which implies that every infeasible set is rendered infeasible as well. Then note that for any \(S \in \mathscr{S}\) and for any \(j\in \llbracket t \rrbracket\), we have \(E^j \not\subseteq S\), so there exists \(i \in \llbracket \abs{E^j} \rrbracket\) such that \(e^j_i\in E^j \setminus S\). Hence, \(S \in L^j_i\) and \(S\) is feasible for level \(j\). To show the "only if" direction, assume for a contradiction that there exists a \(k\)-way IB scheme with \(k\leq r\bra{H^c_\mathscr{S}}-1\). Take a minimal infeasible set \(E =\{e_1,\ldots,e_r\} \in \mathcal{E}_\mathscr{S}\), where \(r = r(H_\mathscr{S}^c)\). Then take \(j \in \llbracket t \rrbracket\) as a level of the IB scheme that renders \(E\) infeasible. By the minimality of \(E\), we have that, for all \(\ell \in \llbracket r \rrbracket\), there exists some \(i(\ell) \in \llbracket k \rrbracket\) such that \(E(\ell) \defeq E \setminus \set{e_\ell} \subseteq L^j_{i(\ell)}\). As \(k < r\), we may apply the pigeonhole principle to see that there must exist some distinct \(\ell^1,\ell^2 \in \llbracket r \rrbracket\) such that \(i(\ell^1)=i(\ell^2)\), and such that \(E(\ell^1) \subseteq L^j_{i(\ell^1)}\) and \(E(\ell^2) \subseteq L^j_{i(\ell^2)}\). As \(E = E(\ell^1) \cup E(\ell^2)\) and \(L^j_{i(\ell^1)} = L^j_{i(\ell^2)}\), this implies that \(E \subseteq L^j_{i(\ell^1)}\), which contradicts our supposition that level \(j\) rendering \(E\) infeasible. Throughout, we will say that \(\operatorname{CDC}(\mathscr{S})\) is \(k\)-way IB-representable (or pairwise IB-representable for \(k=2\)) if it admits a \(k\)-way IB scheme. ## Cardinality constraints {#ss:cardinality} Our first application of Theorem [\[thm:representability\]](#thm:representability){reference-type="ref" reference="thm:representability"} is to derive a strong restriction on the existence of multi-way IB schemes for the cardinality constraint. Direct from Theorem [\[thm:representability\]](#thm:representability){reference-type="ref" reference="thm:representability"} by observing that the conflict hypergraph has rank \(\ell+1\). We observe that the IB scheme [\[basicIB\]](#basicIB){reference-type="eqref" reference="basicIB"}, when applied to the cardinality constraint, is a natural MIP formulation for the "conjunctive normal form", and is unlikely to be practical for even moderately large \(\ell\). In addition, both specialized constraint branching schemes for cardinality constraints and the binary variable branching induced by standard formulations for cardinality constraints are quite imbalanced. The existence of a pairwise independent branching scheme for cardinality constraints would likely have finally produced the sought-after balanced constraint branching. However, Corollary [\[prop:cardinality\]](#prop:cardinality){reference-type="ref" reference="prop:cardinality"} implies that such a balanced constraint branching cannot be produced via IB schemes, or equivalently by constraint branchings that do not use general inequalities (i.e. are only multi-variable branchings). ## Polygonal partitions of the plane {#ss:partitions-of-the-plane} Consider a (nonconvex) bounded region in the plane \(\Omega \subset \mathbb{R}^2\) that describes all possible locations for a UAV, as described in Section [4.6](#obstacleavoidancesec){reference-type="ref" reference="obstacleavoidancesec"}. Assume that \(\Omega\) can be partitioned into polyhedra \(\{P^i\}_{i=1}^d\) such that \(\bigcup_{i=1}^d P^i = \Omega\) and \(\operatorname{relint}(P^i) \cap \operatorname{relint}(P^j) = \emptyset\) for each distinct \(i,j \in \llbracket d \rrbracket\). We note that this partition will not, in general, be unique, and its selection can have a significant effect on questions of representability or formulation size. Figure [\[fig:partitioning\]](#fig:partitioning){reference-type="ref" reference="fig:partitioning"} illustrates this for a convex region with a "hole." The figure shows three ways to partition the resulting nonconvex region into convex polyhedra. Once this partition is fixed, we describe the associated polyhedra in \(\mathscr{V}\)-form, and so the corresponding combinatorial disjunctive constraint is given by \(\mathscr{S} = \{\operatorname{ext}(P^i)\}_{i=1}^d\) and \(J = \bigcup\{S \in \mathscr{S}\}\). We additionally forbid polyhedra with "internal vertices" by requiring that \[\label{intvercon} v \in P^i \Longleftrightarrow v \in \operatorname{ext}(P^i) \quad \forall i \in \llbracket d \rrbracket, v \in J,\] so that \(\mathscr{S}\) corresponds to the maximal elements of a polyhedral complex. For example, the second and third partitions in Figure [\[fig:partitioning\]](#fig:partitioning){reference-type="ref" reference="fig:partitioning"} satisfy this condition, while the first does not. In this setting, minimal infeasible sets have a natural characterization. Take some minimal infeasible hyperedge \(E \in \mathcal{E}_\mathscr{S}\) of \(H^c_\mathscr{S}\), assuming for contradiction that \(r = |E| > 3\), and label the points \(E = \{v^i\}_{i=1}^r\). First, we show that the points may not be in general position, i.e. that w.l.o.g. \(v^r \in \operatorname{Conv}(\{v^i\}_{i=1}^{r-1})\). Then, we argue that the points not being in general position implies that \(\{v^i\}_{i=1}^{r-1}\) is also an infeasible set, violating the minimality condition. Assume for contradiction that the points are in general position; that is, that none can be written as a convex combination of the others. This implies that \(\operatorname{ext}(\operatorname{Conv}(E)) = E\). Assume that the ordering \(\{v^1,\ldots,v^r\}\) forms a path around the edges of \(\operatorname{Conv}(E)\); that is, \(v^i\) and \(v^j\) both lie on an edge of \(\operatorname{Conv}(E)\) if and only if \(|i-j|=1\) or \(\{i,j\}=\{1,r\}\). Choose some set \(S^1 \in \mathscr{S}\) and some \(2 < j < r\) such that \(v^1,v^j \in S^1\) and \(v^2 \not\in S^1\); the associated polyhedron is \(P^1\). Such a set exists, else \(E\) is not a minimal infeasible set (choose instead \(E \backslash \{v^2\}\)). Now choose \(S^2 \in \mathscr{S}\) such that \(v^2, v^r \in S^2\); the associated polyhedron is \(P^2\). Such as set exists, as \(\{v^2,v^r\} \subsetneq E\) and \(E\) is minimal. As the nodes \(v^1,v^2,v^j,v^r\) are interlaced along the boundary of \(\operatorname{Conv}(E)\), we have that \(\operatorname{Conv}(\{v^1,v^j\}) \cap \operatorname{Conv}(\{v^2,v^r\}) \subset \operatorname{Conv}(E)\) is nonempty. As each of the four points is on the boundary of \(\operatorname{Conv}(E)\), and the points are in general position, it follows that \(\operatorname{Conv}(\{v^1,v^j\}) \cap \operatorname{Conv}(\{v^2,v^r\}) = \operatorname{relint}(\operatorname{Conv}(\{v^1,v^j\})) \cap \operatorname{relint}(\operatorname{Conv}(\{v^2,v^r\}))\). Therefore, there must exist some point \(y\) with \(y \in \operatorname{relint}(\operatorname{Conv}(\{v^1,v^j\})) \subseteq \operatorname{relint}(P^1)\) and \(y \in \operatorname{relint}(\operatorname{Conv}(\{v^2,v^r\})) \subseteq \operatorname{relint}(P^2)\). However, this implies that \(\operatorname{relint}(P^1) \cap \operatorname{relint}(P^2) \neq \emptyset\), which contradicts the assumption that our sets partition the region \(\Omega\). Finally, it just remains to show that \(\{v^i\}_{i=1}^{r-1}\) is also an infeasible set, and therefore \(\{v^i\}_{i=1}^r\) cannot be a minimal infeasible set. Assume for contradiction that it is not: i.e. that there exists some \(j\) such that \(\{v^i\}_{i=1}^{r-1} \subseteq \operatorname{ext}(P^j)\). But this implies that \(v^r \in J\) and \(v^r \in \operatorname{Conv}(E) \subseteq P^j\), yet \(v^r \not\in \operatorname{ext}(P^j)\), a contradiction of the internal vertices assumption. In other words, every polyhedral partition of the plane is \(3\)-way independent branching-representable, and pairwise IB representability can be checked in polynomial time (for example, by enumerating the subsets of \(J\) of cardinality 3). To illustrate, in Figure [\[fig:partitioning\]](#fig:partitioning){reference-type="ref" reference="fig:partitioning"} we depict the three possible cases for a partition with respect to Theorem [\[thm:partitioning\]](#thm:partitioning){reference-type="ref" reference="thm:partitioning"}: 1) it does not satisfy the internal vertices condition, 2) it admits a pairwise IB scheme (\(r(H^c_\mathscr{S})=2\)), or 3) it does not admit a pairwise IB scheme, but does admit a 3-way IB scheme (\(r(H^c_\mathscr{S})=3\)). Furthermore, we can argue that we can always represent a obstacle avoidance constraint in such a way that it admits a pairwise IB scheme. Inspecting Figure [\[fig:partitioning\]](#fig:partitioning){reference-type="ref" reference="fig:partitioning"}, we see that the region \(\Omega\) is the same in each, and it is only the partition of \(\Omega\) that can potentially lead to constraints that are not pairwise IB-representable. Therefore, the obstacle avoidance constraint is invariant to the specification of the partition, and if any polyhedral partitioning exists, then it is always possible to construct one that satisfies the conditions of Theorem [\[thm:partitioning\]](#thm:partitioning){reference-type="ref" reference="thm:partitioning"}.[^7] ## SOS2 {#ss:SOS2} Our first example of a constraint that is always pairwise IB-representable is the SOS\(2(N)\) constraint. Recall that \(J = \llbracket N \rrbracket\) and \(\mathscr{S} = \{\{\tau,\tau+1\}: \tau \in \llbracket N-1 \rrbracket\}\) for SOS2. Then \(\mathcal{E}_\mathscr{S}=\set{\{\tau,\tau+t\}: \tau,\tau+t \in \llbracket N \rrbracket, \: t \geq 2}\), \(r\bra{H^c_\mathscr{S}}=2\), and formulation [\[basicIB\]](#basicIB){reference-type="eqref" reference="basicIB"} has depth \(t=\binom{N}{2}-N+1\). However, Vielma and Nemhauser construct a pairwise IB scheme for SOS\(2(N)\) constraints of depth logarithmic in \(N\). The construction is built around a Gray code, or sequence of distinct binary vectors \(\{h^i\}_{i=1}^{N-1} \subseteq \{0,1\}^{\lceil \log_2(N-1) \rceil}\) where each adjacent pair \((h^i,h^{i+1})\) differs in exactly one component. Notationally, here and throughout, take \(h^0 \defeq h^1\) and \(h^N \defeq h^{N-1}\). The pairwise IB scheme is then given by \[L^j = \left\{\tau \in \llbracket N \rrbracket: h^{\tau-1}_j = 1 \text{ or } h^{\tau}_j = 1 \right\}, \quad R^j = \left\{\tau \in \llbracket N \rrbracket: h^{\tau-1}_j = 0 \text{ or } h^{\tau}_j = 0 \right\} \quad\quad \forall j \in \llbracket \lceil \log_2(N-1) \rceil \rrbracket.\] We observe that the resulting formulation matches the lower bound from Proposition [\[prop:log-bound\]](#prop:log-bound){reference-type="ref" reference="prop:log-bound"} with respect to the number of auxiliary binary variables and is significantly smaller than formulation [\[basicIB\]](#basicIB){reference-type="eqref" reference="basicIB"}. Indeed, formulation [\[basicIB\]](#basicIB){reference-type="eqref" reference="basicIB"} is likely to be unnecessarily large for the other constraints we consider as well, so we turn our attention to finding smaller IB schemes in Section [8](#PISEC){reference-type="ref" reference="PISEC"}. We end this section by applying Theorem [\[thm:representability\]](#thm:representability){reference-type="ref" reference="thm:representability"} to succinctly prove the pairwise IB-representability of two other constraints. ## Other pairwise IB-representable constraints #### SOS\(k\) {#sosk} In this case \(\mathcal{E}_\mathscr{S}=\set{\{\tau,\tau+t\}: \tau,\tau+t \in \llbracket N \rrbracket, \: t\geq k+1}\) and \(r\bra{H^c_\mathscr{S}}=2\). #### Grid triangulations We show that \(r\bra{H^c_\mathscr{S}}=2\) by seeing that for any infeasible set \(T \subseteq J\) there exist some distinct \(v,w\in T\) such that \(\set{v,w}\) is infeasible. Indeed, if there are some \(v,w \in T\) such that \(||v-w||_\infty > 1\), then there does not exist any triangle on the grid that contains both, so \(\{v,w\}\) is also an infeasible set. Otherwise, we have that \(T \subset \{r,r+1\}\times\{s,s+1\}\) for some \(r,s\), and that \(T\) contains elements in both of the triangles in this square. For each of the two triangles, we can select an element of \(T\) that is not contained in the other triangle, which yields an infeasible pair contained in \(T\). Therefore, any grid triangulation is pairwise IB-representable. #### Discretization of multilinear terms Similarly as for the grid triangulation case, each infeasible set \(T \subseteq J\) must necessarily contain two elements \(v,w \in T\) with \(||v-w||_\infty > 1\), and so we have that \(T\) can be reduced to the infeasible pair \(\{v,w\}\). Therefore, \(r(H^c_\mathscr{S})=2\), and so any discretization of this form is pairwise IB-representable. # Pairwise independent branching schemes {#PISEC} The pairwise independent branching scheme framework was initially introduced by Vielma and Nemhauser, where it was used to model particularly structured piecewise linear functions. In the remainder of this work, we will focus on pairwise IB schemes and offer a complete picture of their expressive powers, along with an algorithmic framework for constructing them. ## Graphical representations of pairwise IB-representable CDCs {#ss:graph-representation} From our covering assumption \(J = \bigcup\set{S\in \mathscr{S}}\), we can see that \(\abs{E}\geq 2\) for each \(E\in\mathcal{E}_\mathscr{S}\). By applying Theorem [\[thm:representability\]](#thm:representability){reference-type="ref" reference="thm:representability"}, we then immediately have that \(\operatorname{CDC}(\mathscr{S})\) is pairwise IB-representable if and only if \(H^c_\mathscr{S}\) is (equivalent to) a graph. Along this line, for any constraint we may define a *conflict graph* for any \(\operatorname{CDC}(\mathscr{S})\) as \(G^c_\mathscr{S} \defeq (J,\bar{E})\), where \(\bar{E} = \bar{E}_\mathscr{S} \defeq \{\set{u,v} \in [J]^2: u \neq v, \: \{u,v\} \text{ is an infeasible set}\}\) is the set of all infeasible pairs of elements of \(J\). Checking for pairwise IB-representability then reduces to verifying if \(\mathcal{E}_\mathscr{S}=\bar{E}_\mathscr{S}\). The following corollary of Theorem [\[thm:representability\]](#thm:representability){reference-type="ref" reference="thm:representability"} shows that this can also be verified by only working with \(G^c_\mathscr{S}\). If \(\operatorname{CDC}(\mathscr{S})\) is pairwise IB-representable, then \(G_\mathscr{S}^c\) is equivalent to \(H^c_\mathscr{S}\). By applying Theorem [\[thm:representability\]](#thm:representability){reference-type="ref" reference="thm:representability"}, the maximal independent sets of \(G_\mathscr{S}^c\) are exactly the elements of \(\mathscr{S}\). For the converse, assume for a contradiction that \(\mathscr{S}\) is exactly the maximal independent sets of \(G_\mathscr{S}^c\), but that there exists some \(E\in\mathcal{E}_\mathscr{S}\) with \(\abs{E}\geq 3\). By the minimal infeasibility of \(E\), we have that \(\set{r,s}\notin \bar{E}_\mathscr{S}\) for any distinct \(r,s\in E\), and therefore \(E\) is an independent set in \(G_\mathscr{S}^c\). This implies that \(E\) is contained in a maximal independent \(S\). By assumption, \(S\in \mathscr{S}\), which contradicts the infeasibility of \(E\). Therefore, verifying general pairwise IB-representability reduces to enumerating the maximal independent sets of \(G_\mathscr{S}^c\) and identifying them to exactly the sets \(\mathscr{S}\). As an example, we can see that, for cardinality constraint of degree \(\ell\) with \(2 \leq \ell < |J|\), the only maximal independent set of \(G_\mathscr{S}^c\) is the entire ground set \(J\), which certainly cannot be identified with \(\mathscr{S} = \{S \subset J: |S| = \ell \}\). ## Representation at a given depth Once a CDC has been shown to be pairwise IB-representable, a natural next question is: what is the smallest possible depth at which we may construct an IB scheme? That is, we ask if there exists a pairwise IB scheme for \(\operatorname{CDC}(\mathscr{S})\) of some given depth \(t\). The answer to this question reduces to the existence of a graphical decomposition of the conflict graph \(G_\mathscr{S}^c\). The following theorem formalizes the equivalence between biclique covers and pairwise IB schemes. For the first part, take \(\bar{E}\) as the edge set of \(G_\mathscr{S}^c\). To see that any \(S \in \mathscr{S}\) is feasible for the IB scheme [\[eqn:biclique-to-IBnew\]](#eqn:biclique-to-IBnew){reference-type="eqref" reference="eqn:biclique-to-IBnew"}, note that if \(S\not\subseteq L^j\) and \(S\not\subseteq R^j\), then there exist some \(u\in A^j \cap S\) and \(v\in B^j \cap S\). However, this implies that \(\set{u,v}\in A^j * B^j \subseteq \bar{E}\), which is a contradiction of feasibility as \(\set{u,v}\subseteq S\) and \(S \in \mathscr{S}\). Furthermore, as \(\{(A^j,B^j)\}_{j=1}^t\) is a biclique cover of \(G^c_\mathscr{S}\), for every \(\set{u,v}\in \bar{E}\) we have that there exists some level \(j\in \sidx{t}\) such that w.l.o.g. \(u\in A^j\) and \(v\in B^j\). This implies that \(u\notin L^j\) and \(v\notin R^j\) by their construction, and as \(\operatorname{CDC}(\mathscr{S})\) is pairwise IB-representable, then any infeasible set for \(\operatorname{CDC}(\mathscr{S})\) is also infeasible for the proposed IB scheme. Therefore, [\[eqn:biclique-to-IBnew\]](#eqn:biclique-to-IBnew){reference-type="eqref" reference="eqn:biclique-to-IBnew"} is a valid pairwise IB scheme. For the second part, note that \(A^j \cap B^j = \emptyset\) for all \(j \in \llbracket t \rrbracket\), and that the covering portion of Assumption [\[basicassumption\]](#basicassumption){reference-type="ref" reference="basicassumption"} implies that \(L^j\cup R^j=J\). Therefore, it only remains to show that \(\bar{E} = \bigcup_{j=1}^t \bar{E}^j\), where \(\bar{E}^j = A^j * B^j\). For that, first note that as \(L^j\cup R^j=J\), we have that \(A^j=R^j\setminus L^j\) and \(B^j=L^j\setminus R^j\). The containment \(\bar{E} \subseteq \bigcup_{j=1}^t \bar{E}^j\) then follows by noting that, as \(\{(L^j,R^j)\}_{j=1}^t\) is a valid pairwise IB scheme, each minimal infeasible set \(\{u,v\}\in \bar{E}\) has some level \(j \in \llbracket t \rrbracket\) such that \(\{u,v\}\not\subseteq L^j\) and \(\{u,v\}\not\subseteq R^j\). Then, as \(L^j\cup R^j=J\), we have (w.l.o.g.) that \(u \in L^j \backslash R^j \equiv B^j\) and \(b \in R^j \backslash L^j \equiv A^j\), and so \(\{a,b\} \in \bar{E}^j\). For the reverse containment \(\bigcup_{j=1}^t \bar{E}^j \subseteq \bar{E}\), take some arbitrary \(j \in \llbracket t \rrbracket\) and some edge \(\{a,b\} \in \bar{E}^j\). From the definition of our biclique cover, we have that w.l.o.g. \(a \in A^j \equiv R^j \backslash L^j\) and \(b \in B^j \equiv L^j \backslash R^j\). Therefore, \(\{a,b\}\) is an infeasible set for the IB scheme, and thus for \(\operatorname{CDC}(\mathscr{S})\) as well, and so \(\{a,b\} \in \bar{E}\). We can now naturally frame the problem of finding a minimum depth pairwise IB scheme as the minimum biclique cover problem. Unfortunately, the decision version of this problem is known to be NP-complete and inapproximable within a factor of \(|J|^{1/3-\epsilon}\) if \(P \neq NP\) , even for bipartite graphs. However, we note that it is simple to construct a MIP feasibility problem for finding a pairwise IB scheme of a given depth \(t\), which gives us a way to algorithmically find the smallest pairwise IB scheme for a specific (fixed) CDC. We present such a formulation in Proposition [\[prop:feasibility-IP\]](#prop:feasibility-IP){reference-type="ref" reference="prop:feasibility-IP"} in Appendix [\[app:feasibility-IP\]](#app:feasibility-IP){reference-type="ref" reference="app:feasibility-IP"}. Additionally, Cornaz and Fonlupt present a MIP formulation (with an exponential number of constraints that can be efficiently separated) to find the minimum level biclique cover of a graph. Furthermore, we can restate the MIP formulation from (which is a special case of [\[eqn:multiway-formulation\]](#eqn:multiway-formulation){reference-type="eqref" reference="eqn:multiway-formulation"} with \(k=2\)) in terms of biclique covers of \(G_\mathscr{S}^c\). as seen on the right side of Figure [\[fig:sos3\]](#fig:sos3){reference-type="ref" reference="fig:sos3"}. The corresponding MIP formulation is \[\begin{aligned} {2} \lambda_1 + \lambda_8 + \lambda_9 + \lambda_{10} &\leq z_1 \quad\quad \lambda_4 + \lambda_5 &\leq 1-z_1 \\ \lambda_1 + \lambda_2 + \lambda_{10} &\leq z_2 \quad\quad \lambda_5 + \lambda_6 + \lambda_7 &\leq 1-z_2 \\ \lambda_1 + \lambda_2 + \lambda_3 + \lambda_9 + \lambda_{10} &\leq z_3 \quad\quad \lambda_6 &\leq 1-z_3 \\ \lambda_1 + \lambda_2 + \lambda_3 + \lambda_4 &\leq z_4 \quad\quad \lambda_7 + \lambda_8 + \lambda_9 + \lambda_{10} &\leq 1-z_4 \\ (\lambda,z) &\in \Delta^{10} \times \{0,1\}^4. \end{aligned}\] The ad-hoc construction for SOS\(3(6)\) suggests a more general construction for SOS\(k(N)\) when \(k \leq N/2\) (assume for convenience that \(N\) is even). Consider the sets given by \[A^j = \{1,\ldots,j\} \cup \{j+N/2+k,\ldots,N\}, \quad\quad\quad B^j = \{j+k,\ldots,j+N/2\}\] for each \(j \in \llbracket N/2 \rrbracket\). It is straightforward to see that this yields a biclique cover of the conflict graph for SOS\(k(N)\) of depth \(N/2\). Therefore, with this simple operation, we have constructed an ideal formulation for SOS\(k(N)\) with size strictly smaller than \(N\), the size of the naïve non-extended formulation [\[eqn:disaggregated\]](#eqn:disaggregated){reference-type="eqref" reference="eqn:disaggregated"}. Based on the second example [\[eqn:sos3-6-biclique\]](#eqn:sos3-6-biclique){reference-type="eqref" reference="eqn:sos3-6-biclique"}, we know that this construction is, in general, not the smallest possible. In Section [9.4.2](#ss:sosk){reference-type="ref" reference="ss:sosk"}, we will see how we can systematically construct small biclique covers (and MIP formulations) for SOS\(k\)(\(N\)) with arbitrary \(k\) and \(N\), using techniques we will now develop. ## Systematic construction of biclique covers As discussed in Section [7.3](#ss:SOS2){reference-type="ref" reference="ss:SOS2"}, there exists an IB scheme for the SOS2 constraint of optimal depth that can be constructed using a Gray code. The following proposition shows how the validity of this scheme can easily be proven by reinterpreting it via a biclique cover. For the SOS2(\(N\)) constraint we have that \(\bar{E}_\mathscr{S}=\set{\{r,s\} \in \llbracket N \rrbracket^2: r+2 \leq s}\). Take any infeasible pair \(\{r,s\}\in \bar{E}_\mathscr{S}\). As \(r + 2 \leq s\), we conclude that \(r-1<r<s-1<s\), and so it must be that \(h^{r-1}, h^r \neq h^{s-1},h^s\). The set of components which flip values between the two pairs of adjacent codes \((h^{r-1},h^r)\) and \((h^{s-1},h^s)\) is \(I = \{ j \in \llbracket \lceil \log_2(N-1) \rceil \rrbracket: h^{r-1}_j \neq h^r_j \text{ or } h^{s-1}_j \neq h^s_j\}\), and \(|I| \leq 2\) as we have selected a Gray code. Now it must be the case that there is some component \(j \in \llbracket \lceil \log_2(N-1) \rceil \rrbracket \backslash I\) wherein \(h^{r-1}_j=h^r_j \neq h^{s-1}_j=h^s_j\), else we conclude that two of the vectors \(h^i=h^{\ell}\) coincide for some \(i \in \{r-1,r\}\) and \(\ell \in \{s-1,s\}\), a contradiction of their uniqueness. Then \(\{r,s\} \in E^j\), i.e. it is covered by the \(j\)-th level of the biclique. Furthermore, we observe that no edges of the form \(\{r,r+1\}\) will be contained in the biclique cover, as it is not possible that \(h_j^{r-1} = h_j^r = 0\) (resp. \(=1\)) and \(h_j^r = h_j^{r+1} = 1\) (resp. \(=0\)) simultaneously. Interestingly, we can also view this construction recursively if we use a specific Gray code known as the binary reflected Gray code . For SOS2(\(2^k\)), we will take \(E^k\) as the edge set for the corresponding conflict graph. First, with \(k=1\), \(d=2^k=2\), \(N=2^{k}+1=3\), then \(E^1 = \{\{1,3\}\}\). A complete biclique cover is given by the single biclique \(A^{1,1}=\{1\}\) and \(B^{1,1} = \{3\}\). As we see in Figure [\[fig:sos2-recursive\]](#fig:sos2-recursive){reference-type="ref" reference="fig:sos2-recursive"}, we can construct a biclique cover for SOS2 with \(k=2\) by stitching together two copies of the biclique. We construct two copies of the node set for \(k=1\), invert the second, and identify the last node from the first set with the first node with the second set. Then we can readily construct a mapping of the biclique \((A^{1,1},B^{1,1})\) for \(k=1\) to a biclique for \(k=2\), using the node identification, as \(A^{2,1} = \{1,5\}\) and \(B^{2,1} = \{3\}\). This will cover all edges in \(E^2\) with both incident nodes in the first half of the nodes, or both in the second half of the nodes (along with some other edges in \(E^2\), as well). To cover all edges with one adjacent node in the first half, and the other in the second half, we construct a second biclique of the form \(A^{2,2} = \{1,2\}\) and \(B^{2,2} = \{4,5\}\). We can repeat this construction with \(k=3\) to get the three level biclique cover \[\begin{aligned} {3} A^{3,1} &= \{1,5,9\}, \quad\quad &B^{3,1} &= \{3,7\} \\ A^{3,2} &= \{1,2,8,9\}, \quad\quad &B^{3,2} &= \{4,5,6\} \\ A^{3,3} &= \{1,2,3,4\}, \quad\quad &B^{3,2} &= \{6,7,8,9\}. \end{aligned}\] Iterating this construction gives a biclique cover for \(E^{k+1}\) as \(\{(A^{k+1,i},B^{k+1,i})\}_{i=1}^{k+1}\), where \[\begin{aligned} {3} A^{k+1,i} &= \bigcup_{u \in A^{k,i}} \left\{u, 2^{k+1}+2-u\right\}, \quad\quad &B^{k+1,i} =& \bigcup_{v \in B^{k,i}} \left\{v,2^{k+1}+2-v\right\} \quad \forall i \in \llbracket k \rrbracket \\ A^{k+1,k+1} &= \left\{1,\ldots,2^k\right\}, \quad\quad &B^{k+1,k+1} =& \left\{ 2^k + 2, \ldots, 2^{k+1}+1 \right\}. \end{aligned}\] In fact, we can readily state this recursive construction in a more general form, where we adapt a biclique cover for one graph into a biclique cover for another graph that is created in some specific way. In the remainder of this work, we will see how we may apply similar graphical results to systematically construct small biclique covers for the conflict graphs of constraints by exploiting their specific structure. ## Biclique covers for graph products and discretizations of multilinear terms Consider the discretization of multilinear terms described in Section [4.4](#multilinearsec){reference-type="ref" reference="multilinearsec"}, given by \(J=\prod_{i=1}^\eta \llbracket d_i \rrbracket\) and \(\mathscr{S}=\set{\prod_{i=1}^\eta \set{k_i,k_i+1}: k \in \prod_{i=1}^\eta \llbracket d_i-1 \rrbracket}\). We can interpret this constraint as a \(\eta\)-dimensional version of the SOS2 constraint, or as the Cartesian product of \(\eta\) SOS2 constraints. This can be formalized through the following definition and straightforward lemma. Using this characterization, we can easily construct an IB scheme for discretizations of multilinear terms by taking the graph products of IB schemes for the SOS2 constraint. We note that, since \(|\mathscr{S}| = \prod_{i=1}^\eta (d_i-1)\), by Proposition [\[prop:log-bound\]](#prop:log-bound){reference-type="ref" reference="prop:log-bound"} this construction yields a formulation that is asymptotically optimal (with respect to number of auxiliary binary variables) for any possible MIP formulation, up to an additive factor of at most \(\eta\). Furthermore, we can specialize this to the bilinear case studied by Misener et al.. This result yields an ideal MIP formulation for the outer-approximation of bilinear terms with \(\lceil \log_2(m) \rceil\) auxiliary binary variables, \(2(m+1)\) auxiliary continuous variables (the \(\lambda\) variables, one for element in \(J\)), and \(2\lceil \log_2(m) \rceil\) general inequality constraints. In contrast, the logarithmic formulation from Misener et al. has \(\lceil \log_2(m) \rceil\) auxiliary binary variables, \(2\lceil \log_2(m) \rceil + 1\) auxiliary continuous variables, at least \(2\lceil \log_2(m) \rceil + 6\) general inequality constraints, and is not ideal in general (see Appendix [\[app:misener\]](#app:misener){reference-type="ref" reference="app:misener"}). Therefore, we gain an ideal formulation with a naturally induced constraint branching at the price of a modest number of additional auxiliary continuous variables. Furthermore, our formulation generalizes readily to a discretization along the second dimension (\(d_2 > 1\)), for non-uniform discretizations, and for higher dimensional multilinear functions (\(\eta > 2\)). ## Completing biclique covers via graph unions Another useful graphical technique for our heuristic constructions will be to combine together biclique covers, each of which is designed to cover a substructure of the constraint. For example, the conflict graph of a grid triangulation of the plane is equivalent to the conflict graph of a \(2\)-dimensional grid discretization of multilinear terms, with one extra edge added for each subrectangle in the grid. Therefore, a biclique cover of a grid triangulation can be obtained from a biclique cover of a \(2\)-dimensional discretization of multilinear terms (i.e. from Corollary [\[prop:multilinear-IB-scheme\]](#prop:multilinear-IB-scheme){reference-type="ref" reference="prop:multilinear-IB-scheme"}) by completing it with some number additional bicliques that cover those extra edges. This construction can be formalized in the following way. We can apply Lemma [\[lemma:graph-union\]](#lemma:graph-union){reference-type="ref" reference="lemma:graph-union"} to construct biclique covers for the grid triangulations depicted in Figure [\[fig:triangulations\]](#fig:triangulations){reference-type="ref" reference="fig:triangulations"}. First, we apply the biclique cover construction from Corollary [\[prop:multilinear-IB-scheme\]](#prop:multilinear-IB-scheme){reference-type="ref" reference="prop:multilinear-IB-scheme"} to cover all edges not sharing a subrectangle. This is depicted in the first two subfigures of each row in Figure [\[fig:triangulation-IBS\]](#fig:triangulation-IBS){reference-type="ref" reference="fig:triangulation-IBS"}. To cover the remaining 4 edges created by the triangulation, we see that the number of additional levels needed is dependent on the combinatorial structure. Additionally, in all three cases we can verify through Proposition [\[prop:feasibility-IP\]](#prop:feasibility-IP){reference-type="ref" reference="prop:feasibility-IP"} that the resulting biclique cover is of the smallest possible depth. The first example is the "Union Jack" triangulation for \(N=M=3\) and the results in show that for this triangulation the biclique cover from Corollary [\[prop:multilinear-IB-scheme\]](#prop:multilinear-IB-scheme){reference-type="ref" reference="prop:multilinear-IB-scheme"} can be completed with a single additional biclique cover for any \(N\) and \(M\). The second triangulation is a K1 triangulation for \(N=M=3\), and an early version of showed that for this triangulation the biclique can always be completed with two additional bicliques (See Section [9.5](#chromaticsec){reference-type="ref" reference="chromaticsec"} for a generalization of these results). In contrast, for generic triangulations such as the third one, it was not previously known if the biclique cover can always be completed with fewer than the trivial \((M-1)(N-1)\) levels needed to cover each "diagonal" edge with its own additional biclique. First, we can adapt Proposition [\[prop:simple-pIBS\]](#prop:simple-pIBS){reference-type="ref" reference="prop:simple-pIBS"} to cover the extra edges with stars, but in general this will result in \(\Theta(M\cdot N)\) stars, and hence the same number of additional levels. To reduce this, we need a way to *stick* the stars together into more complicated bicliques. In section [9.5](#chromaticsec){reference-type="ref" reference="chromaticsec"} we will see how we may use graph colorings for a broad class of triangulations (subsuming the Union Jack and K1 triangulations as special cases), to cover the extra edges with either one or two additional bicliques. In general, it turns out that we may cover the remaining edges for any grid triangulation with a constant number of additional levels by applying the simple following lemma. The strength of Lemma [\[stickinglemma\]](#stickinglemma){reference-type="ref" reference="stickinglemma"} comes from the fact that many CDCs of practical interest have a local structure (i.e. sets in \(\mathscr{S}\) have small cardinality, or, equivalently, the minimum degree of the conflict graph is close to the total number of nodes). In this case, the condition of Lemma [\[stickinglemma\]](#stickinglemma){reference-type="ref" reference="stickinglemma"} will hold for families of stars centered at nodes that are located "sufficiently far apart." ### Grid triangulations of the plane {#ss:grid-triangulations} We may now present a biclique cover construction for generic grid triangulations, with no further assumptions on the structure of the triangles such as in, whose depth scales like \(\log_2(M) + \log_2(N) + \mathscr{O}(1)\). In the same way as depicted in Figure [\[fig:triangulation-IBS\]](#fig:triangulation-IBS){reference-type="ref" reference="fig:triangulation-IBS"}, we construct the biclique cover by using Lemma [\[lemma:graph-union\]](#lemma:graph-union){reference-type="ref" reference="lemma:graph-union"} to complete the construction of Corollary [\[prop:multilinear-IB-scheme\]](#prop:multilinear-IB-scheme){reference-type="ref" reference="prop:multilinear-IB-scheme"}. For this, we will use the following corollary of Lemma [\[stickinglemma\]](#stickinglemma){reference-type="ref" reference="stickinglemma"} that shows how to combine certain stars centered at sufficiently separated nodes. Direct from Lemma [\[stickinglemma\]](#stickinglemma){reference-type="ref" reference="stickinglemma"} by taking \(u \in J\) and the family of bicliques \(\{(A(w),B(w)\}_{w \in J \cap (u + 3\mathbb{Z}^2)}\) and noting that, if \(u,v\in J \cap (u+3\mathbb{Z}^2)\), then \(||u-v||_\infty \geq 3\), and so \((A(u),B(v))\) is also a biclique for \(G^c_\mathscr{S}\). Figure [\[fig:pwl\]](#fig:pwl){reference-type="ref" reference="fig:pwl"} shows two possible bicliques that can be obtained from Corollary [\[stencil1coro\]](#stencil1coro){reference-type="ref" reference="stencil1coro"}. We can now use Lemma [\[lemma:graph-union\]](#lemma:graph-union){reference-type="ref" reference="lemma:graph-union"} and Corollary [\[stencil1coro\]](#stencil1coro){reference-type="ref" reference="stencil1coro"}, along with the biclique cover derived in Corollary [\[prop:multilinear-IB-scheme\]](#prop:multilinear-IB-scheme){reference-type="ref" reference="prop:multilinear-IB-scheme"}, to obtain a biclique cover for any triangulation with an asymptotically optimal number of levels. for all \(j \in \llbracket \lceil \log_2(M-1) \rceil \rrbracket\), \(j' \in \llbracket \lceil \log_2(N-1) \rceil \rrbracket\), and \(u \in \{0,1,2\}^2\). Let \(G^1 \times G^2 = (J',E')\). Consider the bijection \(f:J\to J'\) given by \(f(u)=(\div(u,k),\operatorname{mod}(u,k))\), where \(\div(u,k)=\lfloor u/k \rfloor\) and \(\operatorname{mod}(u,k)=u-k \div(u,k)\) are the quotient and remainder of the division of \(u\) by \(k\), so that \(f^{-1}\bra{m,r}=km+r\). We have that \(\set{(m,r),(m',r')} \in E'\) if and only if \(\set{m,m'} \in E^1\), which in turn is equivalent to \(|m-m'| \geq 2\). Therefore, for any \(\set{(m,r),(m',r')} \in E'\), we have \[\abs{f^{-1}(m,r)-f^{-1}(m',r')}= |(km+r)-(km'+r')| = |k(m-m')+(r-r')| \geq k|m-m'| + |r-r'| \geq 2k,\] and hence \(\set{f^{-1}(m,r),f^{-1}(m',r')} \in \bar{E}\), i.e. \(\hat{E} \subseteq \bar{E}\). For the second condition, see that if \(u,v\in J\) are such that \(\abs{u-v}\geq 2k\), then \(|\div(u)-\div(v)| \geq 2\), and therefore \(\set{f(u),f(v)} \in E'\). We can then cover the remaining edges with the following bicliques obtained by stitching together families of sufficiently separated stars. Direct from Lemma [\[stickinglemma\]](#stickinglemma){reference-type="ref" reference="stickinglemma"} by considering the family of bicliques \(\{(A(w),B(w)\}_{w \in J \cap (u+3k\mathbb{Z})}\) and noting that, for distinct \(u,v\in J \cap (u+3k\mathbb{Z})\), \(|u-v| \geq 3k\), and so \((A(u),B(v))\) is also a biclique for \(G^c_\mathscr{S}\). Finally, we can combine both classes of bicliques with Lemma [\[lemma:graph-union\]](#lemma:graph-union){reference-type="ref" reference="lemma:graph-union"} to construct a complete biclique cover for SOS\(k(N)\). See Figure [\[fig:sos3-26\]](#fig:sos3-26){reference-type="ref" reference="fig:sos3-26"} for an example of the resulting construction. Take \(G^1 \defeq (\llbracket N/k \rrbracket,E^1)\) as the conflict graph for SOS2(\(N/k\)), \(G^2 \defeq (\{0,\ldots,k-1\},\emptyset)\) as the empty graph on \(k\) nodes, and \(G^3 \defeq \bra{J, \bigcup_{j'=1}^{3k} A^{2,j'} * B^{2,j'}}\). Let \(\hat{G}\) be the subgraph of \(G_\mathscr{S}^c\) from Lemma [\[soskproduct\]](#soskproduct){reference-type="ref" reference="soskproduct"}, which is isomorphic to \(G^1 \times G^2\) through the bijection \(g:\llbracket N/k \rrbracket\times \{0,\ldots,k-1\}\to J\) with \(g\bra{m,r}=km+r\). Then we have that \(G^c_\mathscr{S}= \hat{G}\cup G^3\), after applying Lemma [\[soskproduct\]](#soskproduct){reference-type="ref" reference="soskproduct"} and using the fact that the edges of \(G^3\) contain the edges of \(G^c_\mathscr{S}\) not included in \(\hat{G}\). The result then follows from Lemma [\[lemma:graph-product\]](#lemma:graph-product){reference-type="ref" reference="lemma:graph-product"}, Lemma [\[lemma:graph-union\]](#lemma:graph-union){reference-type="ref" reference="lemma:graph-union"}, Lemma [\[soskproduct\]](#soskproduct){reference-type="ref" reference="soskproduct"}, and Corollary [\[stencil1coro\]](#stencil1coro){reference-type="ref" reference="stencil1coro"}. We note that, when \(k = \mathscr{O}(\log(N))\), this biclique cover yields a MIP formulation that is asymptotically tight (with respect to the number of auxiliary binary variables) with our lower bound of \(\lceil \log_2(N-k+1) \rceil\) from Proposition [\[prop:log-bound\]](#prop:log-bound){reference-type="ref" reference="prop:log-bound"}. We can also show an absolute lower bound of depth \(k\) for any biclique cover for SOS\(k\). This implies that when \(k = \omega(\log(N))\), although the formulation from Theorem [\[thm:sosk\]](#thm:sosk){reference-type="ref" reference="thm:sosk"} is not tight with respect to the lower bound from Proposition [\[prop:log-bound\]](#prop:log-bound){reference-type="ref" reference="prop:log-bound"}, it is asymptotically the smallest possible formulation in the pairwise IB framework. Define \(\gamma \defeq \min\{k, N-k\}\) and consider any possible biclique cover \(\{(A^j,B^j)\}_{j=1}^t\). The biclique cover must separate the edges \(\{(\tau,\tau+k)\}_{\tau=1}^{\gamma}\). Consider a level \(j\) of the biclique cover that contains edge \(\{\tau,\tau+k\}\) for some \(\tau \in \llbracket \gamma \rrbracket\); w.l.o.g., \(\tau \in A^j\) and \(\tau+k \in B^j\). Consider the possibility that the same level \(j\) separates another such edge in the set, e.g. \((\tau',\tau'+k)\) for \(\tau' \in \llbracket \gamma \rrbracket\), where w.l.o.g. \(\tau < \tau'\). That would imply that either \(\tau' \in A^j\) or \(\tau' \in B^j\). In the case that \(\tau' \in A^j\), we have that \(\bar{E}^j\) contains the edge \(\{\tau',\tau+k\}\). However, since \(|(\tau+k)-\tau'| = \tau+k-\tau' < \tau+k-\tau = k\), this implies that the biclique cover separates a feasible edge, a contradiction. In the case where \(\tau' \in B^j\), we have that \(\bar{E}^j\) contains the edge \(\{\tau,\tau'\}\), and as \(\tau'-\tau < k \leq \gamma\) from the definition of our set of edges, a similar argument holds. Therefore, each edge \(\{\{\tau,\tau+k\}\}_{\tau=1}^{\gamma}\) must be uniquely contained in some level of the biclique cover, giving the result. Furthermore, when \(k = \lfloor N/2 \rfloor\), this proposition gives a lower bound on the depth of a biclique cover that is asymptotically tight with the upper bound of \(N\) from Proposition [\[prop:simple-pIBS\]](#prop:simple-pIBS){reference-type="ref" reference="prop:simple-pIBS"}. In other words, in this particular regime, we have that the SOS\(k\) constraint admits a pairwise IB-based formulation, but only one that is relatively large (\(\Omega(|\mathscr{S}|) = \Omega(|J|)\) auxiliary binary variables and constraints). ## Covering edges with a chromatic characterization of bicliques {#chromaticsec} Our final contribution is to adapt a result of Cornaz and Fonlupt  that gives a chromatic characterization of the set of edges that can be covered by a biclique. Using this result, we can sometimes exploit the structure of a grid triangulations to obtain biclique covers that are smaller than those derived in Theorem [\[thm:pwl\]](#thm:pwl){reference-type="ref" reference="thm:pwl"}. Let \(G^x \defeq (\llbracket M \rrbracket,E^x)\) and \(G^y \defeq (\llbracket N \rrbracket,E^y)\) be the conflict graphs for the SOS2(\(M\)) and SOS2(\(N\)) constraints, respectively. Furthermore, let \(G^{3} \defeq (\llbracket M \rrbracket \times \llbracket N \rrbracket,(A^{3,even} * B^{3,even}) \cup (A^{3,odd} * B^{3,odd}))\). Then we may see that \(G^c_\mathscr{S} = (G^x \times G^y) \cup G^3\) by noting that all *diagonal* edges of \(\bar{E}\) (i.e. those of the form \(\set{w,w+v}\in \bar{E}\) for \(w\in J\) and \(v\in \set{-1,1}^2\)) are included in \(G^3\), and that \(G^3\) is a subgraph of \(G^c_\mathscr{S}\). The first part of the theorem then follows from Lemma [\[lemma:graph-product\]](#lemma:graph-product){reference-type="ref" reference="lemma:graph-product"}, Lemma [\[lemma:graph-union\]](#lemma:graph-union){reference-type="ref" reference="lemma:graph-union"}, and Theorem [\[cornaztheo\]](#cornaztheo){reference-type="ref" reference="cornaztheo"}. For the sufficient condition, w.l.o.g. consider the case where \(s=even\) and \(r=odd\). Define \(p_{even}:E_{even}\to \set{0,1}\) as \(p_{even}(e) = \mathds{1}[e \in \bar{E}]\). The result will follow from Theorem [\[cornaztheo\]](#cornaztheo){reference-type="ref" reference="cornaztheo"} by showing that [\[odddegreecond\]](#odddegreecond){reference-type="eqref" reference="odddegreecond"} satisfies condition 2 in the equivalence of the theorem. Let \(E'_{even}=F_{even}\cup\bar{F}_{even}\), and assume for contradiction that there exists \(C\in \mathcal{C}\bra{E'_{even}}\) such that \(\sum\nolimits_{u\in C} p_{even}(u)\) is odd. If \(\abs{C}=4\), we may assume without loss of generality that \(V(C)=\set{u,u+(1,1),u+(1,-1),u+(2,0)}\subset J_{even}\) for some \(u\in J_{even}\). Then \(v=u+(0,1)\in J_{odd}\) is such that \(d_{E_{odd}}(v)\) is odd, a contradiction of [\[odddegreecond\]](#odddegreecond){reference-type="eqref" reference="odddegreecond"}. If \(\abs{C}>4\), note that \(C\in \mathcal{C}\bra{E_{even}}\) and that there exists \(e\in C\) such that \(p_{even}(e)=1\). In addition, there exists \(C_0\in \mathcal{C}\bra{E_{even}}\) such that \(e\in C_0\), \(\abs{C_0}=4\), \(C_1=\bra{C_0\cup C}\setminus \bra{C_0\cap C}\in \mathcal{C}\bra{E_{even}}\) and \(\operatorname{Conv}(C_1) \subsetneq \operatorname{Conv}(C)\). If \(\sum\nolimits_{u\in C_0} p_{even}(u)\) is odd, we may make the same argument above as \(|C_0|=4\) to derive a contradiction of [\[odddegreecond\]](#odddegreecond){reference-type="eqref" reference="odddegreecond"}. If not, then \(C_1\) and \(C\) have the same parity, and therefore \(\sum\nolimits_{u\in C_1} p_{even}(u)\) is odd. We may then repeat this shrinking procedure recursively on \(C_1\) until either the \(4\)-cycle \(C_0\) has odd parity, or \(C_1\) is itself a \(4\)-cycle. In either case, we have a \(4\)-cycle with odd parity, which implies some \(u\in J_{odd}\) that violates [\[odddegreecond\]](#odddegreecond){reference-type="eqref" reference="odddegreecond"}, giving the result. We note that we may use this coloring characterization to recover the biclique covers for both the Union Jack and K1 triangulation example in Figure [\[fig:triangulation-IBS\]](#fig:triangulation-IBS){reference-type="ref" reference="fig:triangulation-IBS"}. For the Union Jack example, we have that \[\bar{F}_{even} = \emptyset, \quad\quad \bar{F}_{odd} = \left\{\{(1,2),(2,1)\},\:\{(2,1),(3,2)\},\:\{(1,2),(2,3)\},\:\{(2,3),(3,2)\}\right\},\] and so we may apply the simplification in Theorem [\[thm:pwl2\]](#thm:pwl2){reference-type="ref" reference="thm:pwl2"} to construct a biclique cover of depth \(\log_2(2)+\log_2(2)+1=3\). Indeed, the original formulation of Vielma and Nemhauser  for the Union Jack triangulation can be reinterpreted analogously through the chromatic characterization of Theorem [\[thm:pwl2\]](#thm:pwl2){reference-type="ref" reference="thm:pwl2"}. For the K1 triangulation example, we have \[\bar{F}_{even} = \left\{ \{(1,3),(2,2)\},\:\{(2,2),(3,1)\} \right\}, \quad\quad \bar{F}_{odd} = \left\{\{(1,2),(2,1)\},\:\{(2,3),(3,2)\} \right\},\] giving a biclique cover of depth \(\log_2(2)+\log_2(2)+2=4\). Furthermore, we close by noting that the sufficient condition in Theorem [\[thm:pwl2\]](#thm:pwl2){reference-type="ref" reference="thm:pwl2"} is, in general, not necessary. For example, in Figure [\[fig:coloring-example\]](#fig:coloring-example){reference-type="ref" reference="fig:coloring-example"} we see a grid triangulation that does not satisfy [\[odddegreecond\]](#odddegreecond){reference-type="eqref" reference="odddegreecond"}, but for which there exists a coloring given by \(\{f_{even},f_{odd}\}\) that satisfies [\[pwlcolorcond\]](#pwlcolorcond){reference-type="eqref" reference="pwlcolorcond"}. That is, \((2,3)\) has odd degree (i.e. \(d_{F_{odd}}((2,3)) = 1\)), and so the sufficient condition is not satisfied. However, a coloring satisfying the conditions of Theorem [\[thm:pwl2\]](#thm:pwl2){reference-type="ref" reference="thm:pwl2"} exists. This offers a generalization of the result that originally appeared in a preliminary version of, which only showed the sufficient condition [\[odddegreecond\]](#odddegreecond){reference-type="eqref" reference="odddegreecond"}.
{'timestamp': '2018-05-25T02:11:17', 'yymm': '1607', 'arxiv_id': '1607.04803', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04803'}
# Introduction Recently, we proposed Prototypes as a way to represent knowledge on the web [^1]. That paper has its focus on theoretical aspects and analysis. In this resource paper we describe the tools we developed to deploy prototypes. First, in [2](#sec:standalone){reference-type="ref" reference="sec:standalone"} we present the implementation of a knowledge base, based on the implementation of a Java interface, which can be used for storing prototypes. Then, we reuse this system to show how the knowledge base can be used in remote and distributed settings ([\[sec:distributed\]](#sec:distributed){reference-type="ref" reference="sec:distributed"}). Each of the sections includes some amount of benchmarking to give the reader an impression of the practical re-usability of the provided solutions. We do assume that the reader has some familiarity with the ideas behind prototypes. The implementation and the code used for the benchmarks is licensed under the LGPLv3 license and can be downloaded from <https://github.com/miselico/knowledgebase>. # A Standalone Knowledge Base {#sec:standalone} A prototype knowledge base (KB) consists of a collection of prototypes. To mirror this, the IKnowledge interface, which we define as the basis for a KB, has only one method which must[^2] be implemented. The signature of this method is `Optional<? extends Prototype> isDefined(ID id);`. The provided Java source code contains five implementations of this interface, namely EmptyKnowledgeBase : is a KB without any content. However, as per the definition, it still contains the empty prototype \(\emptyProtoID\). PredefinedKB : is a KB containing string and integer constants and is described in more details in this section. KnowledgeBase : stores prototypes. It can be constructed using another `IKnowledgeBase` as a basis. The underlying basis will be queried in case the requested prototype is not directly defined in the `KnowledgeBase`. RemoteKB : gets its prototypes from a remote KB. This implementation is described further in [\[sec:remote\]](#sec:remote){reference-type="ref" reference="sec:remote"}. ChainedKB : is an `IknowledgeBase` which connects multiple `IKnowledgeBase`s together. The KBs are checked in turn until one where the Prototype is defined is found. If none is found, an empty `Optional` is returned, indicating that no Prototype could be found. Each prototype consists of four components, namely This structure is closely mimicked in our implementation. The IDs are essentially represented using String types and the change sets using multimaps (i.e., maps which can associate multiple values for a given key). The formal definition allows the creation of a prototype which would remove all values for a given property. According to the theoretical definition, that computation would involve the set of all possible IDs by enumeration, which is unfeasible. Hence [^1]: Please use that paper as a reference for all definitions. [^2]: Other methods are Java 8 default methods.
{'timestamp': '2016-07-19T02:05:47', 'yymm': '1607', 'arxiv_id': '1607.04809', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04809'}
# Introduction #### In recent years, many authors have worked on fractional integral inequalities by using different fractional integral operator such as Riemann-Liouville, Hadamard, Saigo and Erdelyi-Kober, see. In S. Kilinc and H. Yildirim establish new generalized k-fractional integral inequalities involving Gauss hypergeometric function related to Chebyshev functional. In authors gave the following fractional integral inequalities, using the Hadamard and Riemann-Liouville fractional integral for extended Chebyshev functional. #### The main objective of this paper is to establish some Chebyshev type inequalities and some other inequalities using generalized k-fractional integral operator. The paper has been organized as follows. In Section 2, we define basic definitions related to generalized k-fractional integral operator. In section 3, we obtain Chebyshev type inequalities using generalized k-fractional. In Section 4, we prove some inequalities for positive continuous functions. # Preliminaries #### In this section, we present some definitions which will be used later discussion. where, the function \(_{2}F_{1}(-)\) in the right-hand side of (2.4) is the Gaussian hypergeometric function defined by \[_{2}F_{1} (a, b; c; t)=\sum_{n=0}^{\infty}\frac{(a)_{n}(b)_{n}}{(c)_{n}} \frac{t^{n}}{n!},\] and \((a)_{n}\) is the Pochhammer symbol\ \[(a)_{n}=a(a+1)...(a+n-1)=\frac{\Gamma(a+n)}{\Gamma(a)}, \,\,\,(a)_{0}=1.\] Consider the function \[\begin{split} F(t,\tau)&= \frac{(k+1)^{\mu+\beta+1}t^{(k+1)(-\alpha-\beta-2\mu)}}{\Gamma (\alpha)}\tau^{(k+1)\mu}\\ &(t^{k+1}-\tau^{k+1})^{\alpha-1} \times _{2}F_{1} (\alpha+ \beta+\mu,-\eta; \alpha; 1-(\frac{\tau}{t})^{k+1})\\ &=\sum_{n=0}^{\infty}\frac{(\alpha+\beta+\mu)_{n}(-n)_{n}}{\Gamma(\alpha+n)n!}t^{(k+1)(-\alpha-\beta-2\mu-\eta)}\tau^{(k+1)\mu}(t^{k+1}-\tau^{k+1})^{\alpha-1+n}(k+1)^{\mu+\beta+1}\\ &=\frac{\tau^{(k+1)\mu}(t^{k+1}-\tau^{k+1})^{\alpha-1}(k+1)^{\mu+\beta+1}}{t^{k+1}(\alpha+\beta+2\mu)\Gamma(\alpha)}+\\ &\frac{\tau^{(k+1)\mu}(t^{k+1}-\tau^{k+1})^{\alpha}(k+1)^{\mu+\beta+1}(\alpha+\beta+\mu)(-n)}{t^{k+1}(\alpha+\beta+2\mu+1)\Gamma(\alpha+1)}+\\ &\frac{\tau^{(k+1)\mu}(t^{k+1}-\tau^{k+1})^{\alpha+1}(k+1)^{\mu+\beta+1}(\alpha+\beta+\mu)(\alpha+\beta+\mu+1)(-n)(-n+1)}{t^{k+1}(\alpha+\beta+2\mu+1)\Gamma(\alpha+2)2!}+... \end{split}\] It is clear that \(F(t,\tau)\) is positive because for all \(\tau \in (0, t)\), \((t>0)\) since each term of the (2.6) is positive. # Fractional Integral Inequalities for Extended Chebyshev Functional In this section, we establish some Chebyshev type fractional integral inequalities by using the generalized k-fractional integral (in terms of the Gauss hypergeometric function) operator. The following lemma is used for the our main result. **Proof**: Since \(f\) and \(g\) are synchronous on \([0,\infty[\) for all \(\tau \geq 0\), \(\rho\geq 0\), we have \[(f(\tau)-f(\rho)) (g(\tau)-g(\rho))\geq 0.\] From (3.2), \[f(\tau)g(\tau)+f(\rho)g(\rho)\geq f(\tau)g(\rho)+f(\rho)g(\tau).\] Now, multiplying both side of (3.3) by \(\tau^{k}x(\tau)F(t,\tau)\), \(\tau \in (0,t)\), \(t>0\). Then the integrating resulting identity with respect to \(\tau\) from \(0\) to \(t\), we obtain by definition (2.4) \[\begin{split} &I^{\alpha,\beta,\eta,\mu}_{t,k}(xfg)(t)+f(\rho)g(\rho) I^{\alpha,\beta,\eta,\mu}_{t,k}(x)(t)\\ &I^{\alpha,\beta,\eta,\mu}_{t,k}(yg)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(xf)(t)+f(\rho)I^{\alpha,\beta,\eta,\mu}_{t,k}(xg)(t). \end{split}\] Now, multiplying both side of (3.4) by \(\rho^{k}y(\rho)F(t,\rho)\), \(\rho \in (0,t)\), \(t>0\), where \(F(t,\rho)\) defined in view of (2.6). Then the integrating resulting identity with respect to \(\rho\) from \(0\) to \(t\), we obtain by definition (2.4) \[\begin{split} &I^{\alpha,\beta,\eta,\mu}_{t,k}y(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(xfg)(t)+ I^{\alpha,\beta,\eta,\mu}_{t,k}(yfg)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(x)(t)\\ &\geq g(\rho)I^{\alpha,\beta,\eta,\mu}_{t,k}(xf)(t)+I^{\alpha,\beta,\eta,\mu}_{t,k}(yf)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(xg)(t). \end{split}\] This complete the proof of (3.1)\ Now, we gave our main result here. **Proof**: To prove above theorem, putting \(x=p, \ y=q\), and using lemma 3.1, we get \[\begin{split} &I^{\alpha,\beta,\eta,\mu}_{t,k}p(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(qfg)(t)+I^{\alpha,\beta,\eta,\mu}_{t,k}q(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(pfg)(t)\geq \\ &I^{\alpha,\beta,\eta,\mu}_{t,k}(pf)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(qg)(t)+I^{\alpha,\beta,\eta,\mu}_{t,k}(qf)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(pg)(t). \end{split}\] Now, multiplying both side by (3.7) \(I^{\alpha,\beta,\eta,\mu}_{t,k}r(t)\), we have \[\begin{split} &I^{\alpha,\beta,\eta,\mu}_{t,k}r(t)\left[I^{\alpha,\beta,\eta,\mu}_{t,k}p(t) I^{\alpha,\beta,\eta,\mu}_{t,k}(qfg)(t)+I^{\alpha,\beta,\eta,\mu}_{t,k}q(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(pfg)(t)\right]\geq \\ &I^{\alpha,\beta,\eta,\mu}_{t,k}r(t)\left[I^{\alpha,\beta,\eta,\mu}_{t,k}(pf)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(qg)(t)+I^{\alpha,\beta,\eta,\mu}_{t,k}(qf)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(pg)(t)\right], \end{split}\] putting \(x=r, y=q\), and using lemma 3.1, we get \[\begin{split} &I^{\alpha,\beta,\eta,\mu}_{t,k}r(t) I^{\alpha,\beta,\eta,\mu}_{t,k}(qfg)(t)+I^{\alpha,\beta,\eta,\mu}_{t,k}q(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(rfg)(t)\geq \\ &I^{\alpha,\beta,\eta,\mu}_{t,k}(rf)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(qg)(t)+I^{\alpha,\beta,\eta,\mu}_{t,k}(qf)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(rg)(t), \end{split}\] multiplying both side by (3.9) \(I^{\alpha,\beta,\eta,\mu}_{t,k}p(t)\), we have \[\begin{split} &I^{\alpha,\beta,\eta,\mu}_{t,k}p(t)\left[I^{\alpha,\beta,\eta,\mu}_{t,k}r(t) I^{\alpha,\beta,\eta,\mu}_{t,k}(qfg)(t)+I^{\alpha,\beta,\eta,\mu}_{t,k}q(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(rfg)(t) \right]\geq\\ &I^{\alpha,\beta,\eta,\mu}_{t,k}p(t)\left[I^{\alpha,\beta,\eta,\mu}_{t,k}(rf)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(qg)(t)+I^{\alpha,\beta,\eta,\mu}_{t,k}(qf)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(rg)(t)\right]. \end{split}\] With the same arguments as before, we can write \[\begin{split} &I^{\alpha,\beta,\eta,\mu}_{t,k}q(t)\left[I^{\alpha,\beta,\eta,\mu}_{t,k}r(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(pfg)(t)+I^{\alpha,\beta,\eta,\mu}_{t,k}p(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(rfg)(t)\right]\geq\\ &I^{\alpha,\beta,\eta,\mu}_{t,k}q(t)\left[I^{\alpha,\beta,\eta,\mu}_{t,k}(rf)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(pg)(t)+I^{\alpha,\beta,\eta,\mu}_{t,k}(pf)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(rg)(t)\right]. \end{split}\] Adding the inequalities (3.8), (3.10) and (3.11), we get required inequality (3.6).\ Here, we give the lemma which is useful to prove our second main result. **Proof**: Now multiplying both side of (3.4) by \[\begin{split} &\frac{(k+1)^{\upsilon+\delta+1}t^{(k+1)(-\delta-\gamma-2\upsilon)}}{\Gamma (\gamma)}\rho^{(k+1)\upsilon}y(\rho)\\ &(t^{k+1}-\rho^{k+1})^{\gamma-1} \times _{2}F_{1} (\gamma+ \delta+\upsilon,-\zeta; \gamma; 1-(\frac{\rho}{t})^{k+1})\rho^{k} \end{split}\] which remains positive in view of the condition stated in (3.12), \(\rho \in (0,t)\), \(t>0\), we obtain \[\begin{split} &\frac{(k+1)^{\upsilon+\delta+1}t^{(k+1)(-\delta-\gamma-2\upsilon)}}{\Gamma (\gamma)}\rho^{(k+1)\upsilon}y(\rho)\\ &(t^{k+1}-\rho^{k+1})^{\gamma-1} \times _{2}F_{1} (\gamma+ \delta+\upsilon,-\zeta; \gamma; 1-(\frac{\rho}{t})^{k+1})\rho^{k} I^{\alpha,\beta,\eta,\mu}_{t,k}(xfg)(t)\\ &+\frac{(k+1)^{\upsilon+\delta+1}t^{(k+1)(-\delta-\gamma-2\upsilon)}}{\Gamma (\gamma)}\rho^{(k+1)\upsilon}y(\rho)f(\rho)g(\rho)\\ &(t^{k+1}-\rho^{k+1})^{\gamma-1} \times _{2}F_{1} (\gamma+ \delta+\upsilon,-\zeta; \gamma; 1-(\frac{\rho}{t})^{k+1})\rho^{k} I^{\alpha,\beta,\eta,\mu}_{t,k}x(t)\geq \\ &\frac{(k+1)^{\upsilon+\delta+1}t^{(k+1)(-\delta-\gamma-2\upsilon)}}{\Gamma (\gamma)}\rho^{(k+1)\upsilon}y(\rho)g(\rho)\\ &(t^{k+1}-\rho^{k+1})^{\gamma-1} \times _{2}F_{1} (\gamma+ \delta+\upsilon,-\zeta; \gamma; 1-(\frac{\rho}{t})^{k+1})\rho^{k} I^{\alpha,\beta,\eta,\mu}_{t,k}(xf)(t)\\ &+\frac{(k+1)^{\upsilon+\delta+1}t^{(k+1)(-\delta-\gamma-2\upsilon)}}{\Gamma (\gamma)}\rho^{(k+1)\upsilon}y(\rho)f(\rho)\\ &(t^{k+1}-\rho^{k+1})^{\gamma-1} \times _{2}F_{1} (\gamma+ \delta+\upsilon,-\zeta; \gamma; 1-(\frac{\rho}{t})^{k+1})\rho^{k} I^{\alpha,\beta,\eta,\mu}_{t,k}(xg)(t), \end{split}\] then integrating (3.14) over (0,t), we obtain \[\begin{split} &I^{\alpha,\beta,\eta,\mu}_{t,k}(xfg)(t)I^{\gamma,\delta,\zeta,\upsilon}_{t,k}y(t)+ I^{\alpha,\beta,\eta,\mu}_{t,k}(x)(t)I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(yfg)(t)\\ &\geq I^{\alpha,\beta,\eta,\mu}_{t,k}(xf)(t)I^{\gamma,\delta,\zeta,\upsilon}_{t,k}yg(t) +I^{\alpha,\beta,\eta,\mu}_{t,k}(xg)(t)I^{\gamma,\delta,\zeta,\upsilon}_{t,k}yf(t), \end{split}\] this ends the proof of inequality (3.12). **Proof**: To prove above theorem, putting \(x=p, \ y=q\), and using lemma 3.3 we get \[\begin{split} &I^{\alpha,\beta,\eta,\mu}_{t,k}p(t) I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(qfg)(t)+I^{\gamma,\delta,\zeta,\upsilon}_{t,k}q(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(pfg)(t)\geq \\ &I^{\alpha,\beta,\eta,\mu}_{t,k}(pf)(t)I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(qg)(t)+I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(qf)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(pg)(t). \end{split}\] Now, multiplying both side by (3.17) \(I^{\alpha,\beta,\eta,\mu}_{t,k}r(t)\), we have \[\begin{split} &I^{\alpha,\beta,\eta,\mu}_{t,k}r(t)\left[I^{\alpha,\beta,\eta,\mu}_{t,k}p(t) I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(qfg)(t)+I^{\gamma,\delta,\zeta,\upsilon}_{t,k}q(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(pfg)(t)\right]\geq \\ &I^{\alpha,\beta,\eta,\mu}_{t,k}r(t)\left[I^{\alpha,\beta,\eta,\mu}_{t,k}(pf)(t)I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(qg)(t)+I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(qf)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(pg)(t)\right], \end{split}\] putting \(x=r, \ y=q\), and using lemma 3.3, we get \[\begin{split} &I^{\alpha,\beta,\eta,\mu}_{t,k}r(t) I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(qfg)(t)+I^{\gamma,\delta,\zeta,\upsilon}_{t,k}q(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(rfg)(t)\geq \\ &I^{\alpha,\beta,\eta,\mu}_{t,k}(rf)(t)I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(qg)(t)+I^{\gamma,\delta,\zeta,\upsilon}_{t,k}qf)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(rg)(t), \end{split}\] multiplying both side by (3.19) \(I^{\alpha,\beta,\eta,\mu}_{t,k}p(t)\), we have \[\begin{split} &I^{\alpha,\beta,\eta,\mu}_{t,k}p(t)\left[I^{\alpha,\beta,\eta,\mu}_{t,k}r(t) I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(qfg)(t)+I^{\gamma,\delta,\zeta,\upsilon}_{t,k}q(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(rfg)(t)\geq \right]\\ &I^{\alpha,\beta,\eta,\mu}_{t,k}p(t)\left[I^{\alpha,\beta,\eta,\mu}_{t,k}(rf)(t)I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(qg)(t)+I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(qf)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(rg)(t)\right]. \end{split}\] With the same argument as before, we obtain \[\begin{split} &I^{\alpha,\beta,\eta,\mu}_{t,k}q(t)\left[I^{\alpha,\beta,\eta,\mu}_{t,k}r(t)I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(pfg)(t)+I^{\gamma,\delta,\zeta,\upsilon}_{t,k}p(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(rfg)(t)\right]\geq \\ &I^{\alpha,\beta,\eta,\mu}_{t,k}q(t)\left[I^{\alpha,\beta,\eta,\mu}_{t,k}(rf)(t)I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(pg)(t)+(pf)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(rg)(t)\right]. \end{split}\] Adding the inequalities (3.18), (3.20) and (3.21), we follows the inequality (3.16). # Other fractional integral inequalities In this section, we proved some fractional integral inequalities for positive and continuous functions which as follows: **Proof**: Since \(f\), \(g\) and \(h\) be three positive and continuous functions on \([0,\infty[\) by (4.1), we can write \[\begin{split} &f(\tau)g(\tau)h(\tau)+f(\rho)g(\rho)h(\rho)+f(\tau)g(\tau)h(\rho)+f(\rho)g(\rho)h(\tau)\\ &\geq f(\tau)g(\rho)h(\tau)+f(\tau)g(\rho)h(\rho)+f(\rho)g(\tau)h(\tau)+f(\rho)g(\tau)h(\rho). \end{split}\] Now, multiplying both side of (4.3) by \(\tau^{k}x(\tau)F(t,\tau)\), \(\tau \in (0,t)\), \(t>0\). Then the integrating resulting identity with respect to \(\tau\) from \(0\) to \(t\), we obtain by definition (2.4) \[\begin{split} &I^{\alpha,\beta,\eta,\mu}_{t,k}(xfgh)(t)+f(\rho)g(\rho)h(\rho)I^{\alpha,\beta,\eta,\mu}_{t,k}x(t)+g(\tau)h(\rho)I^{\alpha,\beta,\eta,\mu}_{t,k}(xf)(t)\\ &+f(\rho)g(\rho)I^{\alpha,\beta,\eta,\mu}_{t,k}(xh)(t)\geq g(\rho)I^{\alpha,\beta,\eta,\mu}_{t,k}(xfh)(t)+g(\rho)h(\rho)I^{\alpha,\beta,\eta,\mu}_{t,k}(xf)(t)\\ &+f(\rho)I^{\alpha,\beta,\eta,\mu}_{t,k}(xgh)(t)+f(\rho)h(\rho)I^{\alpha,\beta,\eta,\mu}_{t,k}(xg)(t). \end{split}\] Now multiplying both side of (4.4) by \[\begin{split} &\frac{(k+1)^{\upsilon+\delta+1}t^{(k+1)(-\delta-\gamma-2\upsilon)}}{\Gamma (\gamma)}\rho^{(k+1)\upsilon}x(\rho)\\ &(t^{k+1}-\rho^{k+1})^{\gamma-1} \times _{2}F_{1} (\gamma+ \delta+\upsilon,-\zeta; \gamma; 1-(\frac{\rho}{t})^{k+1})\rho^{k} \end{split}\] which remains positive in view of the condition stated in (4.2), \(\rho \in (0,t)\), \(t>0\) and integrating resulting identity with respective \(\rho\) from \(0\) to \(t\), we obtain \[\begin{split} &I^{\alpha,\beta,\eta,\mu}_{t,k}(xfgh)(t)I^{\gamma,\delta,\zeta,\upsilon}_{t,k}x(t)+I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(xfgh)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}x(t)\\ &+I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(xh)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(xgf)(t)+I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(xfg)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(xh)(t)\\ &\geq I^{\gamma,\delta,\zeta,\upsilon}_{t,k}xg(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(xfh)(t)+I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(xgh)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(xf)(t)\\ &+I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(xf)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(xgh)(t)+I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(xfh)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(xg)(t).\end{split}\] which implies the proof inequality 4.2.\ Here, we give another inequality which is as follows. **Proof**: Multiplying both side of (4.3) by \(\tau^{k}x(\tau)F(t,\tau)\), \(\tau \in (0,t)\), \(t>0\), where \(F(t,\tau)\) defined by (2.6). Then the integrating resulting identity with respect to \(\tau\) from \(0\) to \(t\), we obtain by definition (2.4) \[\begin{split} &I^{\alpha,\beta,\eta,\mu}_{t,k}(xfgh)(t)+f(\rho)g(\rho)h(\rho)I^{\alpha,\beta,\eta,\mu}_{t,k}x(t)+g(\tau)h(\rho)I^{\alpha,\beta,\eta,\mu}_{t,k}(xf)(t)\\ &+f(\rho)g(\rho)I^{\alpha,\beta,\eta,\mu}_{t,k}(xh)(t)\geq g(\rho)I^{\alpha,\beta,\eta,\mu}_{t,k}(xfh)(t)+g(\rho)h(\rho)I^{\alpha,\beta,\eta,\mu}_{t,k}(xf)(t)\\ &+f(\rho)I^{\alpha,\beta,\eta,\mu}_{t,k}(xgh)(t)+f(\rho)h(\rho)I^{\alpha,\beta,\eta,\mu}_{t,k}(xg)(t). \end{split}\] Now multiplying both side of (4.8) by \[\begin{split} &\frac{(k+1)^{\upsilon+\delta+1}t^{(k+1)(-\delta-\gamma-2\upsilon)}}{\Gamma (\gamma)}\rho^{(k+1)\upsilon}y(\rho)\\ &(t^{k+1}-\rho^{k+1})^{\gamma-1} \times _{2}F_{1} (\gamma+ \delta+\upsilon,-\zeta; \gamma; 1-(\frac{\rho}{t})^{k+1})\rho^{k} \end{split}\] which remains positive in view of the condition stated in (4.7), \(\rho \in (0,t)\), \(t>0\) and integrating resulting identity with respective \(\rho\) from \(0\) to \(t\), we obtain \[\begin{split} &I^{\alpha,\beta,\eta,\mu}_{t,k}(xfgh)(t)I^{\gamma,\delta,\zeta,\upsilon}_{t,k}y(t)+I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(yfgh)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}x(t)\\ &+I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(yh)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(xgf)(t)+I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(yfg)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(xh)(t)\\ &\geq I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(yg)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(xfh)(t)+I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(ygh)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(xf)(t)\\ &+I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(yf)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(xgh)(t)+I^{\gamma,\delta,\zeta,\upsilon}_{t,k}(yfh)(t)I^{\alpha,\beta,\eta,\mu}_{t,k}(xg)(t). \end{split}\] which implies the proof inequality 4.7.
{'timestamp': '2016-07-19T02:03:40', 'yymm': '1607', 'arxiv_id': '1607.04727', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04727'}
# Introduction Advancement in the current technology of trapping and controlling single atoms cooled down to ultra low temperatures has opened new research directions in quantum optics and quantum communication. Spatial configurations of linear atomic chains or two-dimensional atomic lattices have been engineered and have been widely applied in various experimental setups . Recently, the subject of utilizing super-cold atoms as highly reflecting mirrors has gained much attention. In particular, it has been demonstrated that a collection of cold atoms trapped near the surface of a one-dimensional waveguide can form a nearly perfect mirror for the radiation incident on the atoms . The waveguide represents a photonic channel which enhances the electromagnetic field to which the atoms are coupled thereby leading to a strong collective behavior of the atoms. As a consequence, a large part of the incident light is directed, reflected back, to the medium from which it originated. This striking mirror property of atoms is in contrary to the usual observation where atoms absorb/scatter all or most of the incident energy. Another kind of systems that can exhibit mirror properties or equivalently a highly directive radiative properties are atoms chirally coupled to a waveguide . Chirality in atom-waveguide coupling is an effect associated with a broken symmetry of emission of photons from the atoms into the right and left propagating modes of the waveguide. As a result, the emitted photons are channeled into one of the two directions of the waveguide. It has been shown that the chiral property of the emission can enhance entanglement between two distant atoms . Directive radiative properties have also been demonstrated for a single atom trapped at front of a distant dielectric mirror . It has been demonstrated both theoretically and experimentally that the atom can behave as an optical mirror effectively forming, together with the dielectric mirror, a Fabry-Pérot cavity. Related studies have shown that An atom mirror cannot only serve as a single mirror for a one-dimensional cavity, but also could be arranged to behave like a high-finesse cavity . In the course of previous work on directional emission the underlying atoms independently couple to a one-dimensional field of a waveguide, or nano-cavity or a nano-fibre. Although systems involving independent atoms exhibit interesting directional properties, there can be similar features created by an open system of atoms in which the atoms are coupled to a common three-dimensional field. It was Dicke  who pointed out that a collection of a large number of atoms coupled to a common EM field can radiate collectively such that the spontaneous radiation can be enhanced in certain directions. Since then there have been many studies of the collective radiative properties of multi-atom system demonstrating the dependence of the emitted radiation on the number of atoms and the geometry of the emitting system . In this paper, we investigate radiative properties of an open system of a line of few atoms and demonstrate that the dipole-dipole interaction between atoms may lead to a collective ordering of the emission along some specific directions. To determine directions of the emission, we introduce the directivity function of the emitted radiation field and study the dependence of the function on the distance between the atoms. We analyze the directional properties of the radiation field for two different configurations of the atomic chains, one mimicking an atom in front of a mirror and the other an atom inside a cavity. In the first, we choose the left-hand-side atom of the chain as a probe atom and examine conditions under which the system may radiate only to those modes whose propagation vectors lie within a small solid angle about the inter atomic axis and oriented in one of the two directions of the inter atomic axis, either right (forward) or left (backward) direction. Such a system can be regarded as an atom in front of perfectly reflecting or perfectly transmitting atomic mirror. In the second arrangement, we choose the middle atom of the chain as a probe atom and examine conditions under which the system may radiate only to those modes whose propagation vectors lie within a small solid angle about the inter atomic axis. Such a scheme can be regarded as an atom inside a single-mode cavity. We find that there is a threshold for the inter atomic distances above which a highly directional emission could be achieved. Below the threshold the emission is spherically symmetric. However, we find a population trapping in one of the mirror atoms. For the cavity-like arrangement, the directivity function depends strongly on the number of atoms contained in the chain and the distance between them. For a \(3\)-atom chain and atomic distances above the threshold, two radiative modes of different spatial directions are available for the emission, one in the direction parallel and the other in the direction normal to the inter-atomic axis. Below the threshold, the system can radiate only to the mode normal to the inter atomic axis. For a \(5\)-atom chain and distances above the threshold, only one mode is available for emission, either normal or parallel to the atomic line. Thus, there exist ranges of the inter atomic distances under which the atomic chain exhibits features characteristic of a single-mode cavity. The paper is organized as follows. In Sec. [2](#sec2){reference-type="ref" reference="sec2"}, we describe the master equation of the density operator of the system and the mathematical approach used in the evaluation of the density matrix elements. We introduce the definitions of the directivity function, reflection and transmission coefficients of the radiation field emitted by a chain of atoms. In Sec. [3](#sec3){reference-type="ref" reference="sec3"}, we examine the conditions for the mirror-type behavior of short chains composed of \(3\) and \(5\) atoms. We observe the transient transfer of the population between the atoms and the transient directivity function for an initial condition in which the probe atom is prepared in its excited state. Then, we examine the directivity function of the stationary field when the probe atom is driven by a continuous wave (cw) laser field. Section [4](#sec4){reference-type="ref" reference="sec4"} is devoted to the problem of a cavity formation with atomic mirrors. We are particularly interested in the possibility of the system to concentrate the radiation along the inter atomic axis and thus to behave as a single-mode cavity. Polar diagrams are given to illustrate the mirror and cavity-like features of atomic chains and to show how the features are sensitive to distances between the atoms. The results are summarized in Sec. [5](#sec5){reference-type="ref" reference="sec5"}. The paper concludes with an Appendix6A in which we give details of the derivation of the atomic correlation functions in terms of the populations of the collective states of a three-atom system and the coherences between them and Appendix6B where the calculation of equations of motion has been presented. # Radiative properties of a chain of atoms {#sec2} We consider a system composed of \(N\) identical two-level atoms located at fixed positions \(\vec{r}_{i}\) and coupled to the three-dimensional electromagnetic field whose modes are initially in a vacuum state \(\ket{\{0\}}\). Each atom has an excited state \(\ket{e_{i}}\) and a ground state \(\ket{g_{i}}\) separated by energy \(\hbar\omega_{0}\) and connected by a transition dipole moment \(\vec{\mu}\). The atoms are arranged in a line, and we consider two cases shown in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}. In the first case, illustrated in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}(a), we assume that the left-side atom, chosen as a "probe\" atom is separated from its next-neighbor by a distance \(r_{0}\) which is larger than the separation \(r_{m}\) between the remaining \(N-1\) atoms, \(r_{0}> r_{m}\). If the probe atom is excited into its upper level, it will spontaneously decay into the ground state emitting the radiation field that can be absorbed by the chain of closely located atoms and then re-emitted by the atoms towards the probe atom. Thus, the chain of closely located atoms could act as a mirror, directing the emitted radiation into a cone about the interatomic axis and turned towards the probe atom. In the second case, illustrated in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}(b), we assume that the middle atom of the chain is separated from its adjacent neighbors by a distance \(r_{0}\), which is much larger that the separation \(r_{m}\) between the remaining atoms. This arrangement may model a situation of an atom located inside a cavity whose mirrors are formed by two chains of equally distant atoms. In practice this scheme could be realized by extending the recently demonstrated scheme involving two superconducting qubits coupled to a 1D field  to the case of three or five qubits coupled to a 2D field. In the experiment of Ref. , effective separations of \(\lambda\) and \(3\lambda/4\) were achieved between the fixed qubits by changing the qubit transition frequencies. ## Master equation When the system is coupled to a reservoir the state of the total system, the chain of atoms plus the reservoir field, is described by the density operator \(\rho_{T}\). The reduced density operator describing the properties of only the chain of atoms is obtained by tracing the total density operator \(\rho_{T}\) over the states of the reservoir, \(\rho ={\rm Tr}_{R}\rho_{T}\). The master equation describing the time evolution of the reduced density operator has the form  \[\begin{aligned} \frac{\partial \rho}{\partial t} &=-\frac{i}{\hbar}\left[H_{0}+H_{L}+H_{dd},\rho\right] \nonumber\\ &-\frac{1}{2}\sum_{i=1}^N\gamma \left( \left[S_{i}^{+},S_{i}^{-}\rho\right] + \textrm{H.c.}\right)\nonumber\\ &-\frac{1}{2}\sum_{i\neq j =1}^N\gamma_{ij} \left( \left[S_{i}^{+},S_{j}^{-}\rho\right] + \textrm{H.c.}\right) ,\label{q1} \end{aligned}\] where \(\gamma\) is the spontaneous emission damping rate of the individual atoms, equal to the Einstein \(A\) coefficient, \(S_{i}^{+}=\ket{e_{i}}\bra{g_{i}}\) and \(S_{i}^{-}=\ket{g_{i}}\bra{e_{i}}\) are the dipole raising and lowering operators of atom \(i\), and \(\gamma_{ij}\) is the collective damping rate \[\begin{aligned} \gamma_{ij} &= \frac{3}{2}\gamma\left\{\left[1-\left(\hat{\mu}\cdot\hat{r}_{ij}\right)^{2}\right]\frac{\sin\eta_{ij}}{\eta_{ij}}\right. \nonumber\\ &\left. + \left[1-3\left(\hat{\mu}\cdot\hat{r}_{ij}\right)^{2}\right]\left(\frac{\cos\eta_{ij}}{\eta_{ij}^2}-\frac{\sin\eta_{ij}}{\eta_{ij}^3}\right)\right\}, \label{q2} \end{aligned}\] with \[\begin{aligned} \eta_{ij} = k\,r_{ij}=2\pi r_{ij}/\lambda ,\quad \vec{r}_{ij} = r_{ij}\hat{r}_{ij}=\vec{r}_{j}-\vec{r}_{i}, \end{aligned}\] in which \(r_{ij}\) is the distance between atoms \(i\) and \(j\), \(\hat{r}_{ij}\) is the unit vector in the direction \(\vec{r}_{ij}\), and \(\lambda\) is the resonant wavelength. The master equation ([\[q1\]](#q1){reference-type="ref" reference="q1"}) describes the atomic dynamics under the Born-Markov and rotating-wave approximations , \(H_{0}\) is the Hamiltonian describing the free energy of the atoms \[\begin{aligned} H_{0} = \hbar\sum_{i=1}^{N}\omega_{0}S^{+}_{i}S_{i}^{-} ,\label{q3} \end{aligned}\] \(H_{L}\) is the Hamiltonian describing the interaction of the probe atom with an external driving field of frequency \(\omega_{L}\): \[\begin{aligned} H_{L} = \frac{1}{2}\hbar\Omega_{0}\left(S_{1}^{+}e^{-i\omega_{L}t} + S_{1}^{-}e^{i\omega_{L}t}\right) ,\label{q3a} \end{aligned}\] where \(\Omega_{0}\) is the Rabi frequency of the driving field, and \(H_{dd}\) is the Hamiltonian describing the dipole-dipole interaction between the atoms \[\begin{aligned} H_{dd} = \hbar\sum_{i\neq j=1}^{N}\Omega_{ij}\left(S_{i}^{+}S_{j}^{-} + S_{j}^{+}S_{i}^{-}\right) ,\label{q4} \end{aligned}\] where \(\Omega_{ij}\) is the dipole-dipole interaction strength between atoms \(i\) and \(j\), defined by \[\begin{aligned} \Omega_{ij} &= \frac{3}{4}\gamma\left\{\left[1-3\left(\hat{\mu}\cdot\hat{r}_{ij}\right)^{2}\right]\left(\frac{\sin\eta_{ij}}{\eta_{ij}^2}+\frac{\cos\eta_{ij}}{\eta_{ij}^3}\right)\right. \nonumber\\ &\left.-\left[1-\left(\hat{\mu}\cdot\hat{r}_{ij}\right)^{2}\right]\frac{\cos\eta_{ij}}{\eta_{ij}}\right\}.\label{q5} \end{aligned}\] The parameters \(\gamma_{ij}\) and \(\Omega_{ij}\) depend on the separation between the atoms. For large separations, \(\eta_{ij}\gg 1\), and then both coupling parameters approach zero. For \(\eta_{ij}\ll 1\) the parameter \(\gamma_{ij}\) reduces to \(\gamma\) while \(\Omega_{ij}\) becomes large and strongly dependent on \(r_{ij}\). It is well known that \(\Omega_{ij}\) plays the important role in the collective behavior of multi-atom systems and we shall see that it has important effect on the distribution of the radiation field emitted by a chain of atoms. The calculation of the equations of motion for atomic populations and coherences for a time-dependent state vector has been outlined briefly in Appendix6B. ## Directivity function, reflection and transmission coefficients {#sec2a} The intensity of the radiation field emitted at time \(t\) in the direction specified by the polar angle \(\theta\) between the direction of observation \(\vec{R}\) and the direction of the atomic axis \(\vec{r}_{ij}\) can be expressed in terms of the correlation functions of the atomic dipole operators as \[\begin{aligned} I(\theta,t) = u(\phi)\sum_{i,j=1}^{N}\gamma \langle S_{i}^{+}(t)S_{j}^{-}(t)\rangle e^{ikr_{ij}\cos\theta} ,\label{q6} \end{aligned}\] where \(u(\phi)=(3/8\pi)\sin^{2}\phi\) is the radiation pattern of a single atomic dipole, with \(\phi\) the angle between the observation direction \(\vec{R}\) and the direction of the atomic transition dipole moment \(\vec{\mu}_{i}\). Since the radiation intensity \(I(\theta,t)\) is symmetric about the interatomic axis, it defines a two-dimensional surface called the polar radiation pattern of the emitting system. We may introduce the *directivity* function determining of how effective the atoms are in converging the emitted radiation into a small solid angle centered about the direction \(\theta\). The directivity function \(D(\theta,t)\) in the direction \(\theta\) at time \(t\) is defined as the ratio of the radiation intensity \(I(\theta,t)\) emitted in the direction \(\theta\) divided by the total radiation intensity \(I(t)\): \[\begin{aligned} D(\theta, t) = \frac{u(\phi)}{I(t)}\sum_{i,j=1}^{N}\gamma \langle S_{i}^{+}(t)S_{j}^{-}(t)\rangle e^{ikr_{ij}\cos\theta} ,\label{q9} \end{aligned}\] where the total radiation intensity \(I(t)\) at time \(t\) is obtained by integrating \(I(\theta,t)\) over \(\theta\): \[\begin{aligned} I(t) = \int I(\theta,t)d\theta = \sum_{i,j=1}^{N}\gamma_{ij}\langle S_{i}^{+}(t)S_{j}^{-}(t)\rangle ,\label{q10} \end{aligned}\] in which \(\gamma_{ii}=\gamma_{jj}=\gamma\) and \(\gamma_{ij}\, (i\neq j)\) is given in Eq. ([\[q2\]](#q2){reference-type="ref" reference="q2"}). The directivity function is a measure of how effective the system is in concentrating the radiation in a given direction. It is equivalent to the probability density of detecting a fluorescence photon traveling in the direction \(\theta\). Since our interest is in situations where a chain of atoms works as an atomic mirror, an important factor is the ability of producing highly directional patterns of the radiation concentrated in one of the two directions along the interatomic axis, either \(\theta=0\) or \(\theta=\pi\). Following the arrangement illustrated in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}(a), \(D(\theta=\pi,t)\) describes the radiation field emitted along the atomic axis in the direction towards the probe atom, the \"backward\" direction. Thus, it would correspond to the reflection coefficient. On the other hand, the directivity \(D(\theta=0,t)\) describes the radiation field emitted along the atomic axis in the direction away from the probe atom, the "forward\" direction. Therefore, it would correspond to the transmission coefficient of the atomic mirror. Thus, we may define the *reflection* coefficient of the chain of atoms as the ratio of the radiation intensity emitted in the direction \(\theta= \pi\) to the total radiation intensity \[\begin{aligned} R(t) \equiv D(\theta\!=\!\pi, t) = \frac{u(\phi)}{I(t)}\!\sum_{i,j=1}^{N}\!\gamma \langle S_{i}^{+}(t)S_{j}^{-}(t)\rangle e^{ikr_{ij}}.\label{q12} \end{aligned}\] The reflection coefficient is a measure of how effective the atoms are in concentrating the radiation about one side of the interatomic axis, i.e. about the direction \(\theta =\pi\). The coefficient \(R(t)\) is equivalent to the probability density of detecting a fluorescence photon traveling in the direction \(\theta =\pi\). Similarly, we can define the *transmission* coefficient of the chain as the ratio of the radiation intensity emitted in the direction \(\theta=0\) to the total intensity of the field emitted \[\begin{aligned} T(t) \equiv D(\theta\!=\!0,t) = \frac{u(\phi)}{I(t)}\!\sum_{i,j=1}^{N}\!\gamma \langle S_{i}^{+}(t)S_{j}^{-}(t)\rangle e^{-ikr_{ij}}.\label{q11} \end{aligned}\] Obviously, \(T(t)=1\) would correspond to complete transmission whereas \(R(t)=1\) would correspond to complete reflection of the radiation field emitted along the atomic axis. ## Directional properties of the radiation field {#sec2b} The quantity of central interest is the directivity function which can be determined from analyzing the polar radiation pattern of a chain of atoms. We may determine general conditions under which the radiation pattern of \(N\) atoms would be highly non spherical and its maximum is concentrated along the inter atomic axis. The conclusions will serve as reference for choosing distances between the atoms and for calculations of the atomic populations and correlations. The expression ([\[q6\]](#q6){reference-type="ref" reference="q6"}) can be written as a sum of \(N\) terms \[\begin{aligned} I(\theta,t) = \sum_{i<j=1}^{N}I_{ij}(\theta,t) ,\label{q7} \end{aligned}\] where \[\begin{aligned} I_{ij}(\theta,t) &= u(\phi)\gamma \left\{\frac{1}{N\!-\!1}\!\left(\langle S_{i}^{+}(t)S_{i}^{-}(t)\rangle\!+\!\langle S_{j}^{+}(t)S_{j}^{-}(t)\rangle\right)\right. \nonumber\\ &\left. +\, 2{\rm Re}\{\langle S_{i}^{+}(t)S_{j}^{-}(t)\rangle\}\cos\left(kr_{ij}\cos\theta\right)\right. \nonumber\\ &\left.-\, 2{\rm Im}\{\langle S_{i}^{+}(t)S_{j}^{-}(t)\rangle\}\sin\left(kr_{ij}\cos\theta\right)\right\}.\label{q8} \end{aligned}\] We see that the contribution of the atoms to the intensity occurs in pairs of different combinations of the atoms. Therefore, the system of radiating atoms can be considered as made up of a number of short two-atom elements and the total intensity is obtained by summing up the intensities of the fields produced by all the elements. If we wish a short chain of atoms to work like a mirror with a large convergence and reflectivity of the radiation emitted by a probe atom located at either end of the chain, we should arrange the atoms such that the total field emitted (scattered) could be highly focused along the interatomic axis with a pronounced maximum in the backward direction \(\theta = \pi\) and a minimum, preferably zero emission in the forward direction \(\theta = 0\) with respect to the line center. In order to find the conditions for concentrating the radiation in the direction \(\theta= \pi\), let us examine the intensity ([\[q8\]](#q8){reference-type="ref" reference="q8"}) in more details. From Eq. ([\[q8\]](#q8){reference-type="ref" reference="q8"}) it is seen that there are three terms determining the radiation pattern. The first term in Eq. ([\[q8\]](#q8){reference-type="ref" reference="q8"}) is just the sum of the populations of the two atoms involved, the probabilities that the atoms are in their excited states. This term is independent of \(\theta\) and therefore contributes uniformly in all directions. The second term depends on \(\theta\) and varies as \(\cos\left(kr_{ij}\cos\theta\right)\) with an amplitude equal to the real part of atomic correlations. Consequently, this term could contribute to the radiation pattern only if the correlations between the atoms would have nonzero real part, \({\rm Re}\{\langle S_{i}^{+}S_{j}^{-}\rangle\}\neq 0\). However, the cosine term, although dependent on \(\theta\), would produce intensity maxima in the \(\theta =0, \pi/2, 3\pi/2\) and \(\pi\) directions. As such this term is not effective in concentrating the radiation along the interatomic axis. This simple argument suggests that the atoms should be arranged such that the cosine term vanishes. This could be achieved when the correlations between the atoms have zero real part. The third term contributing to the radiation intensity ([\[q8\]](#q8){reference-type="ref" reference="q8"}) varies as \(\sin\left(kr_{ij}\cos\theta\right)\) and hence can affect the radiation intensity in a decidedly different way than the cosine term. An important difference is that \(\sin\left(kr_{ij}\cos\theta\right)\) vanishes for \(\theta =\pi/2\) and \(3\pi/2\). This means that the sine term does not contribute to the radiation emitted in the direction perpendicular to the interatomic axis. Consequently, an appreciable concentration of the radiation could be achieved along the interatomic axis by choosing proper distances \(r_{ij}\) between the atoms at which \(\sin\left(kr_{ij}\right)=\pm 1\). Furthermore, since the sine is an odd function, it follows that \(\sin\left(kr_{ij}\cos 0^{\circ}\right)=-\sin\left(kr_{ij}\cos 180^{\circ}\right)\), which means that independent of the separation between the atoms a maximum in the backward direction is always accompanied by a minimum in the forward direction. Obviously, the sine term could influence the angular distribution of the radiation intensity only if the atomic correlations would have nonzero imaginary parts, \({\rm Im}\{\langle S_{i}^{+}S_{j}^{-}\rangle\}\neq 0\). In addition, the sign of the imaginary part of the atomic correlations dictates the choice of distances between the atoms, at which the emission would be enhanced in the backward direction (high reflection) and reduced in the forward direction (low transmission). Thus, if \({\rm Im}\{\langle S_{i}^{+}S_{j}^{-}\rangle\} > 0\), the sine term will display a maximum in the backward direction for atomic distances at which \(\sin(kr_{ij})=1\). This condition is satisfied when the atomic separations are \[\begin{aligned} r_{ij} = \frac{1}{4}(2n+1)\lambda ,\quad n\in\{0,2,4,\ldots \} \end{aligned}\] However, if the coefficient of the sine term is negative, \({\rm Im}\{\langle S_{i}^{+}S_{j}^{-}\rangle\} < 0\), a different choice of the distances is required at which \(\sin(kr_{ij})=-1\). This condition is satisfied when \[\begin{aligned} r_{ij} = \frac{3}{4}(2n+1)\lambda ,\quad n\in\{0,2,4,\ldots \} \end{aligned}\] We see that there is a lower bound imposed on the distances between the atoms, either \(r_{ij}=\lambda/4\) or \(r_{ij}=3\lambda/4\), above which a one-sided emission along the interatomic axis can be achieved, i.e., a maximum of radiation in the direction \(\theta = 0\). For distances smaller that the lower bound, the one-sided emission is expected to be significantly reduced. In the following, we limit ourselves to short chains containing only \(N=3\) and \(N=5\) atoms. Also in the absence of the driving field, \(\Omega_{0}=0\), the \(N=3\) case can be solved in closed form yielding simple expressions. In the Appendix6A, we outline the calculation of atomic correlation functions in the collective state basis. ## Mathematical approach To study the radiative behavior of a chain of interacting atoms, we require the time evolution of the populations of the atoms and the coherence between them. These are given by the diagonal and off-diagonal density matrix elements, respectively. If the space of the atomic system is spanned in the basis of the eigenstates of the free Hamiltonian \(H_{0}\), we readily find that the basis is composed of \(2^{N}\) state vectors, i.e. for a chain composed of \(N=3\) atoms, the basis is composed of \(8\) vectors \(\ket{i_{1}j_{2}k_{3}}\), whereas for \(N=5\) atoms it is composed of \(64\) vectors \(\ket{i_{1}j_{2}k_{3}l_{4}m_{5}},\, \{i,j,k,l,m\} \in\{ g,e\}\). Hence, in the simplest case of \(N=3\) atoms, the master equation ([\[q1\]](#q1){reference-type="ref" reference="q1"}) provides us with a system of \(64\) coupled linear equations to be solved, in principle, a \(63\times 63\) matrix to be diagonalized. For \(N=5\) we get a system of \(4096\) coupled linear equations whose solution requires a \(4095\times 4095\) matrix to be diagonalized. Needless to say, it is not easily accomplished. Therefore, we shall use numerical methods. To get solutions for the density matrix elements, the following approach will be taken. From the master equation ([\[q1\]](#q1){reference-type="ref" reference="q1"}), we find equations of motion for the density matrix elements, which can be written in a matrix form as \[\dot{\vec{X}}\left( t\right) = {\cal A} \vec{X}\left( t\right) +\vec{R}, \label{q17}\] where \(\vec{X}\left( t\right)\) is a column vector composed of the density matrix elements, \(\vec{R}\) is a column vector composed of the inhomogenous terms, and \({\cal A}\) is a matrix of the coefficients appearing in the equations of motion of the density matrix elements. A direct integration of Eq. ([\[q17\]](#q17){reference-type="ref" reference="q17"}) leads to the following formal solution for \(\vec{X}\left( t\right)\) \[\begin{aligned} \vec{X}\left( t\right) =\vec{X}\left( t_{0}\right) e^{{\cal A}t}-\left(1-e^{{\cal A}t}\right) {\cal A}^{-1}\vec{R}, \label{q18} \end{aligned}\] where \(t_{0}\) is the initial time. Because the determinant of the matrix \({\cal A}\) is different from zero, there exists a complex invertible matrix \({\cal U}\) which diagonalizes \({\cal A}\), and \(w = {\cal U}^{-1}{\cal A}{\cal U}\) is the diagonal matrix of complex eigenvalues. By introducing new vectors \(\vec{Y}={\cal U}^{-1}\vec{X}\) and \(\vec{T}={\cal U}^{-1}\vec{R}\), we can rewrite ([\[q18\]](#q18){reference-type="ref" reference="q18"}) as \[\vec{Y}\left( t\right) =\vec{Y}\left( t_{0}\right) e^{w t}-\left(1-e^{w t}\right) w^{-1}\vec{T}, \label{q19}\] or in component form \[Y_{n}\left( t\right) =Y_{n}\left( t_{0}\right) e^{w_{n}t}-\sum_{m=1}^{s}\!\left( w^{-1}\right)_{nm}\!\left( 1-e^{w_{m}t}\right) T_{m}, \label{q20}\] in which \(s=63\) for the case of \(N=3\) atoms and \(s=4095\) for \(N=5\) atoms. To obtain solutions for \(X_{n}\left( t\right)\) we must determine the eigenvalues \(w_{n}\) and eigenvectors \(Y_{n}\left( t\right) ,\) which are readily evaluated by a numerical diagonalization of the matrix \({\cal A}\). The steady-state values of the density matrix elements can be found from Eq. ([\[q20\]](#q20){reference-type="ref" reference="q20"}) by taking \(t\rightarrow \infty\), or more directly by setting the left-hand side of Eq. ([\[q17\]](#q17){reference-type="ref" reference="q17"}) equal to zero, and then \[\vec{X}\left( \infty \right) =-{\cal A}^{-1}\vec{R}, \label{q21}\] or in component form \[X_{n}\left( \infty \right) =-\sum_{m=1}^{s}\left( {\cal A}^{-1}\right)_{nm}R_{m}. \label{q22}\] In what follows, we shall use the solutions ([\[q20\]](#q20){reference-type="ref" reference="q20"}) and ([\[q22\]](#q22){reference-type="ref" reference="q22"}) to illustrate the radiative properties of the atomic chains. In particular, we calculate the transient populations of the atoms and the steady-state directivity function. # Radiating atom at front of an atomic mirror {#sec3} Consider first the configuration illustrated in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}(a), a single atom distance \(r_{12}\) from a finite-size chain of closely located atoms. This arrangement could constitute a radiating atom at front of an atomic mirror. In order to find the mirror-like characteristics of the chain that could be inferred from the radiative properties of the systems, we examine the directivity function of the radiation field emitted by the system. The directivity function, which is determined by the angular distribution of the radiation intensity, Eq. ([\[q7\]](#q7){reference-type="ref" reference="q7"}), depends on temporal and spatial factors, the time of the evolution of the system and the separation between the atoms. It is clear from Eq. ([\[q7\]](#q7){reference-type="ref" reference="q7"}) that in general the temporal and spatial factors cannot be separated. The angular distribution of the radiation field at a given time can be different for different separations between the atoms. Moreover, the transient radiation intensity is a sensitive function of the initial atomic conditions. Even at the initial time \(t=0\) the angular distribution of the radiation intensity can depend on the separation \(r_{ij}\) if the system was prepared in a state with nonzero interatomic correlations, either \({\rm Re}\{\langle S_{i}^{+}S_{j}^{-}\rangle\}\) or \({\rm Im}\{\langle S_{i}^{+}S_{j}^{-}\rangle\}\) different from zero. In order to study the angular distribution of the radiation field, it is important to understand the radiative behavior of individual atoms in the chain. Therefore, we first consider the evolution of the populations of the atoms. Following this discussion, we display the variation of the directivity function with time for different distances between the atoms. We assume that the probe atom is initially excited and separated from the front atom of the "mirror\", its nearest neighbor, at a distance much larger than the separation between the "mirror\" atoms. The method of preparing the probe atom in the excited state will not concern us but it might be done with a short laser pulse, for example. ## Transient transfer of the population {#sec3a} Our interest is in the time evolution of the directivity function of the emitted field for an initially excited probe atom, the case illustrated in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}(a). In order to study this evolution, we first look at the time evolution of the populations of the atoms for chains composed of \(N=3\) and \(N=5\) atoms. The aim is to first determine and then optimize the conditions which not only allow the radiation emitted by the probe atom to be emitted back (reflected) to it but also to achieve a strong convergence of the radiated field along the interatomic axis. We evaluate the transient populations of the atoms assuming there is no driving field \((\Omega_{0}=0)\) and that initially the probe atom was in its excited state. The corresponding results for the time evolution of the atomic populations for different distances between the atoms are presented in Figs. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"} and [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}. Figure [\[fig2\]](#fig2){reference-type="ref" reference="fig2"} shows the transient transfer of the population between \(N=3\) atoms; the initially excited probe atom and a group of two atoms forming a mirror. The populations are computed for several different sets of distances \((r_{12}, r_{23})\) between the atoms. For distances \(r_{ij}>\lambda/5\) the initial population of the probe atom decays exponentially in time whereas the populations of the mirror atoms \(2\) and \(3\) build up with small oscillations. However, there is no tendency of the mirror atoms to transfer their populations back to the probe atom. A small oscillatory behavior of the population \(\rho_{11}(t)\) can be seen for short times, \(t<5/\gamma\), but it is not periodic. A periodic oscillatory behavior is observed for the populations of atoms \(2\) and \(3\) that the mirror atoms periodically exchange their populations without transferring it to the probe atom. The reason is that the distance \(r_{23}\) is much smaller than \(r_{12}\), resulting in a stronger dipole-dipole interaction between the mirror atoms, \(\Omega_{23}\gg \Omega_{12}\). For smaller distances, the population of the probe atom begins to show periodic oscillations, with the population actually oscillating between the probe atom and only the rear atom of the mirror. Except for very short times, the population of the middle atom \(2\) remains almost constant during the evolution of the system. Note that the transfer of the population between the atoms is not complete since the populations remain nonzero, \(\rho_{ii}(t)\neq 0\) for all \(t > 0\). It is interesting that atom \(2\), the front atom of the mirror, appears as an mediator in the population exchange between the probe atom and the rear atom of the mirror. In other words, the front atom of the mirror dictates the direction of the emission. Despite of the possibility to emit in any spatial direction, the probe atom "prefers\" to radiate towards the atomic mirror and vice versa, the mirror atoms prefer to radiate towards the probe atom. We now consider a chain containing \(N=5\) atoms with the probe atom initially excited and located at a distance \(r_{12}\) from the front of a line of four equally separated atoms. The transient transfer of the population between the atoms for several different sets of distances, \((r_{12}, r_{i,i+1}), \, i \in \{2,3,4 \}\) is shown in Fig. [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}. For large distances the situation is similar to that of three atoms, the population rapidly escapes from the system leaving the atoms unpopulated over a short evolution time. For small distances, we observe periodic oscillations of the populations with the periodicity, as before for three atoms, determined by the dipole-dipole interaction strength. However, unlike the three atoms case, the population can be completely transferred from the mirror atoms back to the probe atom such that the mirror atoms are unpopulated at times when the population of the probe atom is maximal. This is a substantial difference compared to the case of \(N=3\) atoms where a part of the population was trapped by one of the mirror atoms. Clearly, longer chains are more effective in the complete transfer of an excitation from the probe atom to the mirror atoms and vice versa. ## Transient directivity function {#sec3b} The periodic exchange of the populations between the atoms, in particular, between the atoms forming an atomic mirror, can lead to the emission of the atoms into certain preferred directions. In order to demonstrate this behavior, we examine the time variation of the directivity function \(D(\theta,t)\) of the field radiated by the group of mirror atoms only. Of course it might be argued that it could be hard for experiments to detect the field radiated from a fraction of atoms, but we would like to see under which conditions the field radiated by the group of mirror atoms could be concentrated in the preferred direction along the interatomic axis, either \(\theta=0\) or \(\theta =\pi\). With a high concentration of the radiated field into a small solid angle centered around \(\theta=0\), the atoms could be regarded as a highly reflecting mirror scattering the field back towards the probe atom. The directionality function of the field emitted by two (mirror) atoms and detected on a sphere around the line of three atoms is shown in Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}. We keep the distance of the probe atom from the front of the mirror atoms at \(r_{12}=\lambda/2\) and consider the variation of \(D(\theta,t)\) with the distance between the mirror atoms at two different times. It is seen that the emission is concentrated along the interatomic axis. For a distance between the mirror atoms \(r_{23}=\lambda/4\), the directivity function is significantly enhanced in one direction, that the emission is almost one-sided either in \(\theta=0\) or \(\theta=\pi\). The directivity function oscillates in time in such a manner that it develops a maximum in the direction \(\theta =0\) at later times in which it was a minimum at earlier times and vice versa. The oscillation of the directivity results from the oscillation of the population between the mirror atoms, as we have see above in Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}. The beam width or equivalently the solid angle inside which the emitted field is concentrated remains almost constant. One can also notice from Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"} that the ability of the coupled atoms to concentrate the radiation in one direction decreases with a decreasing distance between the atoms. It is clearly seen that for distances \(r_{23}\leq \lambda/8\) the directivity function is almost symmetrical along the interatomic axis. This result is consistent with the general property of the radiation pattern discussed in Sec. [2.2](#sec2a){reference-type="ref" reference="sec2a"} that there is a minimal value of the distance between the atoms \((r_{ij}=\lambda/4)\) above which the sine term can be maximal in the \(\theta=0\) direction. For distances smaller than the minimal value the contribution of the sine term is necessarily smaller resulting in the reduction of the ability of the system to concentrate the emission in one direction. The one-sided emission and then the reflectivity coefficient can be enhanced by increasing the number of atoms forming the atomic mirror. This is demonstrated in Fig. [\[fig5\]](#fig5){reference-type="ref" reference="fig5"} which shows the polar diagram of the directivity function \(D(\theta,t)\) for the case of four atoms forming the atomic mirror. As above for three atoms, we keep the probe atom at a fixed distance from the front of the atomic mirror, \(r_{12}=\lambda/2\), and consider the variation of \(D(\theta,t)\) with the distance between the mirror atoms. It can be seen that similar to the case of two atoms the emission is concentrated mainly along the interatomic axis. The one-sided emission is significantly enhanced and persists even for small distances. Small peaks can be seen at about \(2\pi/3\) and \(4\pi/3\) degrees. It is easy to verify that the function \(\sin\left(kr_{ij}\cos\theta\right)\) when evaluated for \(r_{25}=3\lambda/8\), corresponding to the distance between the front and rear atoms of the mirror, attains its maximal value of \(\sin\left(kr_{ij}\cos\theta\right)= 1\) for \(\cos\theta= 2/3\), which corresponds to directions \(\theta \approx 130^{\circ}\). ## Directivity function of the stationary field {#sec3c} In practice a photo detector located at some position \(\vec{R}\) in the far field zone of the radiation field emitted by the atoms would detect the field emitted by the entire set of atoms rather than a fraction of selected atoms only. It is a consequence of the fact that the fields from the probe atom and the mirror atoms are unresolved at the detector. Therefore, we now consider the directivity function of the field radiated by the entire system of atoms including the field radiated from the probe atom. Moreover, we assume that initially all atoms were in their ground states and then the probe atom was exposed to the incident weak laser light. We shall look into the radiation pattern of the emitted field in the steady-state limit, \(t\rightarrow \infty\), and compute the directivity function \(D(\theta) \equiv \lim_{t\rightarrow\infty}D(\theta, t)\) for equally distant as well as for non-equally distant \(N=3\) and \(N=5\) atoms. Since the directivity function involves the contribution from the populations of the atoms, that a significant directionality can be obtained when the populations are small, we shall assume that the driving field is weak so that the Rabi frequency \(\Omega\) much smaller than \(\gamma\). The directivity function for the radiation field emitted by \(N=3\) atoms is shown in Fig. [\[fig6\]](#fig6){reference-type="ref" reference="fig6"}, where frame (a) is for equally distant atoms, while frame (b) is for unequally distant atoms. Unfortunately, as one can see from the figure, in both cases that the stationary field radiated by the system is concentrated in directions \(\theta=\pi/3\) and \(\theta = 5\pi/3\), which significantly departure from the direction of the interatomic axis. Therefore, the system of three equally or unequally separated atoms with the continuously driven probe atom cannot be regarded as suitable for the mirror-type behavior of the maximal directivity along the interatomic axis. Figure [\[fig7\]](#fig7){reference-type="ref" reference="fig7"} shows the corresponding situation for the case of \(N=5\) atoms. It is seen that in the case of four atoms forming the mirror the directivity function depends crucially on the separation between the atoms. In particular when the distance of the probe atom from the front of the mirror atoms is \(r_{12}=\lambda/4\), the radiated field modes available for the emission are only those contained inside a cone centered about the direction \(\theta =\pi\). Thus, at that particular distance the chain of four undriven atoms acts as a perfectly reflecting mirror by "pushing\" the radiation in the backward direction. It can be seen that the optimal conditions for the one-sided emission with a maximum in the direction \(\theta =\pi\) are achieved when the probe atom is separated from the front of the mirror atoms by \(r_{12}=\lambda/4\) and the mirror atoms are themselves separated by \(r_{ij}=\lambda/6\). In this case, the directivity function is nonzero only in directions contained inside a cone limited by \(\theta\leq \pm\pi/3\). Thus, the emission is entirely one-sided with no radiation in the \(\theta=0\) direction. This means that the transmission coefficient \(T=0\), therefore this type of behavior can be regarded as a mirror type with perfect reflectivity. Of course, the reflectivity is accompanied by losses in the sense that the radiation is emitted into a cone of a finite solid angle \(2\theta\) determining the beamwidth of the radiated field. The solid angle subtended by the cone seen in Fig. [\[fig7\]](#fig7){reference-type="ref" reference="fig7"}(c) is \(2\theta = 120^{\circ}\). It was pointed out in  that a strong directivity of the emitted radiation along the interatomic axis can be obtained in a chain of independent atoms, i.e. in the absence of the dipole-dipole interaction. However, the one sided emission seen in frames (c) and (d) of Fig. [\[fig7\]](#fig7){reference-type="ref" reference="fig7"} requires a non-zero dipole-dipole interaction. This is illustrated in Fig. [\[fig7a\]](#fig7a){reference-type="ref" reference="fig7a"} which shows the directivity function for the same parameters as in Fig. [\[fig7\]](#fig7){reference-type="ref" reference="fig7"}(d) but for independent atoms. Clearly, in the absence of the dipole-dipole interaction between the atoms, the emission is strongly directional, but is symmetrical in the \(\theta=0\) and \(\pi\) directions. In summary of this section, we have seen that the directivity of the emission and the mirror like behavior of a line of atoms depends on the distance between the atoms and the nature of the excitation. At particular distances between the atoms, the system can radiate along the interatomic axis in only one direction \(\theta=\pi\), which can be interpreted as a perfect reflection of the radiation towards the probe atom. # Cavity formation with atomic mirrors {#sec4} We now wish to create a cavity with atomic mirrors. For this purpose, we place a probe atom between a pair of finite-size chains of closely located atoms, as illustrated in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}(b). These would constitute a radiating atom located inside a one-dimensional cavity. We determine the parameter ranges in which the two chains of \((N-1)/2\) atoms could act as mirrors to the field emitted by the probe atom. We examine two systems containing different numbers of atoms. First, we consider the simplest system composed of \(N=3\) equidistant atoms in a line. Then, we extend our discussion to a larger system composed of \(N=5\) atoms. We illustrate our considerations by examining the time evolution of the atomic populations. For the initial conditions we choose the middle (probe) atom to be in its excited state and the other atoms to be in their ground states. We also analyze the directivity function for features indicative of cavity-type modifications of the radiation pattern of the radiation field. ## Transient regime {#sec4a} The simplest system which could exhibit features characterizing a cavity formed by atomic mirrors is a chain of \(N=3\) equidistant atoms. Let us first examine the process of population transfer between the middle (probe) atom of the chain and the side (mirror) atoms. The time evolution of the populations of the atoms is illustrated in Fig. [\[fig8\]](#fig8){reference-type="ref" reference="fig8"}, where we present the dependence of the population transfer on the distance between the atoms. It is seen that the population is periodically transferred between the probe atom and the mirror atoms. The transfer occurs with frequency determined by the dipole-dipole coupling \(\Omega_{12}\, (=\Omega_{23})\) between the probe and mirror atoms. The oscillations show an interesting behavior that the population transfer is not affected (modulated) by the presence of the dipole-dipole coupling \(\Omega_{13}\) between the mirror atoms. The oscillations are accompanied by a steady decay of the populations. However, depending on the distance between the atoms, the populations may not decay to zero but rather to long-lived non-zero values. It is clearly seen from the figure that for atomic distances \(r_{ij}<\lambda/4\), i.e. the spacing between the mirror atoms \(r_{13}<\lambda/2\), a significant part of the initial population remains trapped in the probe atom from which it decays very slowly. We have thus a situation similar to that one encounters in a Fabry-Pérot cavity composed of a pair of parallel plate mirrors . In the cavity, the number of modes for interaction with an atom decreases with an decreasing separation \(L\) between the mirrors. For \(L<\lambda/2\) the number of modes is suppressed resulting in the suppression of the radiation from the atom. Figure [\[fig9\]](#fig9){reference-type="ref" reference="fig9"} illustrates the cavity-like situation involving \(5\) atoms. Here, each of the cavity mirrors is formed with two atoms. The results are similar to those we encountered for the case of \(3\) atoms, Fig. [\[fig8\]](#fig8){reference-type="ref" reference="fig8"}, however, one can see some interesting differences. Again, trapping of the population for small distances between the atoms is evident. It is worth noting that the system has a tendency to trap the population in the mirror atoms rather than in the probe atom. It can also be noted that for large distances the probe atom exchanges the population with the next neighbors, the front atoms rather than with the rear atoms forming the mirrors. However, for small distances, the probe atom exchanges the population with the rear atoms leaving the populations of the front atoms almost constant in time. Similar effects appeared to the population transfer in the system composed of a probe atom located at front of an atomic mirror, see Sec. [3.1](#sec3a){reference-type="ref" reference="sec3a"}. ## Stationary regime {#sec4b} We now turn to the problem of determining the conditions under which the two chains of atoms could act like cavity mirrors to select modes centered about the atomic axis as the only modes available for the emission. We assume that the probe atom is continuously driven by a coherent laser field and consider the field in the steady state limit. Figure [\[fig10\]](#fig10){reference-type="ref" reference="fig10"} illustrates the effect of decreasing distance between atoms on the directivity function \(D(\theta)\) of the stationary field for the cavity-like situation involving \(3\) atoms. It is seen that the directivity function is very sensitive to the distances between the atoms. For large distances, the system radiates along the cavity axis as well as in the direction normal to the cavity axis. When the atomic separations are reduced below \(\lambda/4\), which corresponds to the mirror atoms spacing less than \(\lambda/2\), we see a cancellation of the radiation along the interatomic axis. Thus, the system turns off the radiation along the cavity axis. It radiates only in the directions normal to the cavity axis, the property which clearly is not characteristic of a cavity. Turning next to the case of \(5\) atoms, we plot in Fig. [\[fig11\]](#fig11){reference-type="ref" reference="fig11"} the directivity function for several different distances between the atoms. We see that the directivity function differs significantly from what we observed for the case of three atoms. Thus, the effect of additional atoms in forming the cavity mirrors is clearly more pronounced on the directivity function of the stationary field than on the transient population of the atoms. First we note from the figure that there are distances between the atoms at which the directivity function is nonzero only along the interatomic axis \[ Figs. [\[fig11\]](#fig11){reference-type="ref" reference="fig11"}(b) and [\[fig11\]](#fig11){reference-type="ref" reference="fig11"}(c) \]. It can also be noticed that there is an optimal distance between the mirror atoms at which the solid angle subtended by the cone is minimal, or equivalently the width of the cavity axial mode is very narrow. As seen from Fig. [\[fig11\]](#fig11){reference-type="ref" reference="fig11"}(c), a reduction of the separation may result in an increase of the beamwidth. We also see quite clearly that larger separations between the mirror atoms, additional directions of emission appear \[ Fig. [\[fig11\]](#fig11){reference-type="ref" reference="fig11"}(d) \]. # Summary {#sec5} We have studied directional properties of the radiation field emitted by a chain of closely located and dipole-dipole interacting two-level atoms. Two geometrical configurations of atoms in the chain have been studied: A probe atom in front of a finite-size chain of closely located atoms, and a probe atom between a pair of chains of closely located atoms. We call the earlier case as an atom in front of an atomic mirror, and the former as an atom inside a cavity whose the mirrors are formed by the two chains of atoms. We have found that it is possible to account for certain mirror and cavity-like features, such as We have examined the conditions for one-sided emission centered about the interatomic axis, and have found a lower bound for the distance between the atoms above which a one-sided emission along the interatomic axis can be achieved. The one-sided emission focused into a cone about the interatomic axis and oriented in the backward direction can be regarded as a mirror type with perfect reflectivity. For atomic distances smaller than the lower bound, a part of the population can be trapped in the probe atom indicating that at these distances there are no radiative modes available for the emission. This is a situation similar to that one encounters in a Fabry-Pérot cavity composed of a pair of parallel plate mirrors . Polar diagrams have been presented showing the variation of the directivity function of the time dependent as well as the stationary fields with the separation between the atoms. We have found that the directional properties of the radiation field are generally more manifested in the stationary field when the probe atom is continuously driven by a coherent laser field. The control of directionality of photon emission could have applications in quantum information processing in realizing a directional quantum network. For example, the mirror-like behavior of the chain could be used to realize a chiral quantum network, where atoms interact via one-way emission . It could also be used as an optical reflector or an optical mirror to create, together with a distant dielectric mirror, a Fabry-Pérot cavity . Q. G. acknowledges support from 6th research grant (proposal number 3775) from Jazan University, Saudi Arabia.
{'timestamp': '2016-07-18T02:07:26', 'yymm': '1607', 'arxiv_id': '1607.04519', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04519'}
null
null
null
null
null
null
null
null
# Results and Discussion Consider the optical response of a symmetric quadrumer when it is excited by a plane wave, whose propagation direction and electric field polarization lie in the plane of the quadrumer (\(s\)-polarization), as shown in Fig. [\[fig:schematic\]](#fig:schematic){reference-type="ref" reference="fig:schematic"}a. The conventional magnetic response to such an excitation is supported by a collective circulation of electric displacement current around all four nanoparticles. However, if the nanoparticles are made of a high-index dielectric material such as silicon, the individual nanoparticles are also able to couple directly with the applied magnetic field, sustaining internal circulating polarization currents. In this sense, a silicon quadrumer can produce magnetic responses to both electric *and* magnetic components of the applied plane wave, as is depicted in Fig. [\[fig:schematic\]](#fig:schematic){reference-type="ref" reference="fig:schematic"}b. In this figure, the red arrows denote an electric dipole, and the blue circular arrows denote the circulation of displacement current that supports an out-of-page magnetic dipole. Importantly, an interplay may be induced between the two magnetic responses through the locally-enhanced, \(z\)-polarized magnetic field sustained by the collective circulation of displacement current, which establishes a coupling channel to the magnetic responses of the individual nanoparticles. We begin by understanding this magnetic interplay, for which we will model the interactions between nanoparticles using the *Coupled Electric and Magnetic Dipole Approximation*  (CEMDA). The mathematical description of this dipole model in free space is based on the following equations: where \(\mathbf{p}_i\) (\(\mathbf{m}_i\)) is the electric (magnetic) dipole moment of the \(i^{\mathrm{th}}\) particle, \(\hat{G}_{0}(\mathbf{r}_{i},\mathbf{r}_{j})\) is the free space dyadic Greens function between the \(i^{\mathrm{th}}\) and \(j^{\mathrm{th}}\) dipole, \(\alpha_{E}\) (\(\alpha_{H}\)) is the electric (magnetic) polarizability of a particle, \(c_0\) is the speed of light and \(k\) is the free-space wavenumber. In Fig. [\[fig:CSTvsCEMDA\]](#fig:CSTvsCEMDA){reference-type="ref" reference="fig:CSTvsCEMDA"}a, we show that this dipole model is accurate in modeling full-wave simulations of the silicon nanosphere quadrumers performed using CST Microwave Studio, confirming that, in this spectral range, the optical response of such nanoparticle quadrumers is dominated by dipole interactions among the individual particles. This is the case because higher order coupling in such systems will occur only with smaller interparticle separations  (see also the Supporting Information). Fig. [\[fig:CSTvsCEMDA\]](#fig:CSTvsCEMDA){reference-type="ref" reference="fig:CSTvsCEMDA"}b shows the electric field distribution at the collective magnetic resonance of the cluster, confirming that the response of the silicon quadrumer includes a collective circulation of electric field around the particles. However, we can also see the additional dynamic produced by dielectric nanoparticles: the magnetic response induced by circulating transverse electric dipoles is accompanied by a magnetic response from each of the individual nanoparticles. This is the unique effect we are interested in. The magnetic response in an individual silicon nanosphere is an internal circulation of polarization current, which can be seen in Fig. [\[fig:CSTvsCEMDA\]](#fig:CSTvsCEMDA){reference-type="ref" reference="fig:CSTvsCEMDA"}b. In the CEMDA we choose to homogenize this circulating current distribution into a discrete source of magnetization current; the magnetic dipole. This implies that the key to strong interaction between individual and collective magnetic responses resides in the strong coupling between induced electric and magnetic dipoles in each inclusion. To describe this coupling, we can consider the eigenmodes of the quadrumer system. As we will show, the eigenmodes of the full, electromagnetic, system can be constructed from the eigenmodes of the decoupled electric and magnetic dipole systems. Therefore, we are going to break apart our dipole system of Eq. [\[eq:dipole equations\]](#eq:dipole equations){reference-type="ref" reference="eq:dipole equations"} into two decoupled equations: one equation for the electric dipoles and one equation for the magnetic dipoles We can then define two sets of eigenmodes from these equations: one for the electric dipoles and one for the magnetic dipoles. Using state notation, we can refer to the electric dipole eigenmodes as \(\left | \mathbf{e} \right \rangle\) and the magnetic dipole eigenmodes as \(\left | \mathbf{h} \right \rangle\). The eigenmode approach in the dipole approximation offers a huge simplification when combined with symmetry, because it allows us to determine certain eigenmodes without calculation. The key is that, for both electric and magnetic dipole systems, the dipole approximation restricts the number of eigenmodes for each equation to twelve (three dimensions by four dipoles). Each of these eigenmodes can only transform according to a single irreducible representation of the quadrumer's symmetry group, referred to as \(\mathrm{D_{4h}}\). Further details regarding the \(\mathrm{D_{4h}}\) symmetry group's irreducible representations and the implications for eigenmodes are provided in the Supporting Information. For the analysis here, it suffices to say that there are only eight irreducible representations in \(\mathrm{D_{4h}}\) for twelve eigenmodes, and therefore the eigenspace associated with certain irreducible representations must be one-dimensional. Moreover, given that any vector in a one-dimensional space is an eigenvector by default, we are able to derive a number of eigenmodes by simply finding dipole moment profiles that transform according to certain irreducible representations. Eight such dipole moment profiles are shown in Fig. [\[fig:eigenmodes\]](#fig:eigenmodes){reference-type="ref" reference="fig:eigenmodes"}. Each of these dipole moment profiles is the sole basis vector for a single irreducible representation and is therefore an eigenmode of the electric or magnetic dipole equations in Eq. [\[eq:decoupled equations\]](#eq:decoupled equations){reference-type="ref" reference="eq:decoupled equations"}, irrespective of wavelength or the choice of material, size, or any parameter which conserves the symmetry of the quadrumer. It is worth noting that this same procedure for finding eigenmodes is applicable to other symmetries when using the dipole approximation, particularly the \(\mathrm{D}_{n\mathrm{h}}\) symmetry groups for small values of \(n\). We need to now consider the interaction between electric and magnetic dipole systems. An eigenmode of either electric or magnetic dipoles can be substituted into Eq. [\[eq:dipole equations\]](#eq:dipole equations){reference-type="ref" reference="eq:dipole equations"} to determine the resulting state of magnetic or electric dipoles (\(| \mathbf{m_{(e)}} \rangle\) or \(| \mathbf{p_{(h)}} \rangle\), respectively) that it will induce due to magneto-electric coupling: However, it has been shown that the operation describing the coupling between states of electric and magnetic dipole moments, must commute with the geometry's symmetry operations.  Subsequently the \(\left | \mathbf{e}\right \rangle\) and \(| \mathbf{m_{(e)}} \rangle\) dipole moments must both transform according to the same irreducible representation, and similarly for the \(\left | \mathbf{h}\right \rangle\) and \(| \mathbf{p_{(h)}} \rangle\) dipole moments. Therefore, if we consider the eigenmodes of electric and magnetic dipoles in Fig. [\[fig:eigenmodes\]](#fig:eigenmodes){reference-type="ref" reference="fig:eigenmodes"}, the requirement for symmetry conservation specifies which electric dipole eigenmode can couple into which (if any) magnetic dipole eigenmode, and *vice versa*. As a result, the \(\mathrm{A_{2g}}\) or \(\mathrm{B_{1g}}\) eigenmodes (*cf.* Fig. [\[fig:eigenmodes\]](#fig:eigenmodes){reference-type="ref" reference="fig:eigenmodes"}) are only able to magneto-electrically couple into the other \(\mathrm{A_{2g}}\) or \(\mathrm{B_{1g}}\) eigenmode, whereas the \(\mathrm{A_{1g}}\) and \(\mathrm{B_{2g}}\) eigenmodes are not able to couple into any eigenmodes due to a symmetry mismatch. This conclusion allows us to take our analysis of the decoupled electric and magnetic dipole equations, and use it to determine the full, *electromagnetic*, eigenmodes of the real system. Indeed, we would not necessarily expect an eigenmode of this complex scattering system to be a current distribution described by purely electric or magnetic dipoles; we would expect it to be a combination of both. If we now restrict ourselves to considering only the \(\mathrm{A_{2g}}\) (magnetic-like) eigenmodes of an all-dielectric quadrumer, the two eigenmodes from the decoupled electric and magnetic dipole equations (*cf.* Fig. [\[fig:eigenmodes\]](#fig:eigenmodes){reference-type="ref" reference="fig:eigenmodes"}) form basis vectors for the \(\mathrm{A_{2g}}\) eigenmodes, \(| \mathbf{v}_x \rangle\), of the electromagnetic system. \[\begin{aligned} | \mathbf{v}_x \rangle = a_x \left ( \!\! \begin{array}{c} | \mathbf{e} \rangle \\ 0 \end{array} \!\!\right) + b_x \left (\!\! \begin{array}{c} 0 \\ | \mathbf{h} \rangle \end{array} \!\! \right)\,,\label{eq:EMeigenmode} \end{aligned}\] where \(a_x\) and \(b_x\) are complex scalars. Notably, we should expect to have two distinct electromagnetic eigenmodes here because there are two distinct basis vectors and, subsequently, a two-dimensional eigenspace for the quadrumer's response. To derive expressions for \(a_x\) and \(b_x\), we can write the electromagnetic eigenvalue equation for Eq. [\[eq:dipole equations\]](#eq:dipole equations){reference-type="ref" reference="eq:dipole equations"}: where \(\lambda_x\) is the eigenvalue of the electromagnetic eigenmode. However, Eqs. [\[eq:eigenmode equations\]](#eq:eigenmode equations){reference-type="ref" reference="eq:eigenmode equations"}a and [\[eq:eigenmode equations\]](#eq:eigenmode equations){reference-type="ref" reference="eq:eigenmode equations"}b are not independent[^1] because one can be obtained from the other using duality transformations.  In other words, a solution for \(a_x\) and \(b_x\), that satisfies *either* Eq. [\[eq:eigenmode equations\]](#eq:eigenmode equations){reference-type="ref" reference="eq:eigenmode equations"}a or Eq. [\[eq:eigenmode equations\]](#eq:eigenmode equations){reference-type="ref" reference="eq:eigenmode equations"}b, must also satisfy the complete Eq. [\[eq:eigenmode equations\]](#eq:eigenmode equations){reference-type="ref" reference="eq:eigenmode equations"} and describe an electromagnetic eigenmode of Eq. [\[eq:EMeigenmode\]](#eq:EMeigenmode){reference-type="ref" reference="eq:EMeigenmode"}. So, if the basis vectors \(\left | \mathbf{e}\right \rangle\) and \(\left | \mathbf{h} \right \rangle\) are normalized to a magnitude of one, we can project Eq. [\[eq:eigenmode equations\]](#eq:eigenmode equations){reference-type="ref" reference="eq:eigenmode equations"}a and Eq. [\[eq:eigenmode equations\]](#eq:eigenmode equations){reference-type="ref" reference="eq:eigenmode equations"}b onto \(\mathbf{e}_{i}\) and \(\mathbf{h}_{i}\) (respectively) and sum over all \(i\), to obtain the two solutions for Eq. [\[eq:eigenmode equations\]](#eq:eigenmode equations){reference-type="ref" reference="eq:eigenmode equations"}: \[\begin{aligned} \frac{a_1}{b_1} &= \frac{ \underset{i}{\sum} \, \mathbf{e}_i^{*}\cdot \mathbf{p_{(h)}}_{i} } {\alpha_{\scriptscriptstyle E} \epsilon_0 (\lambda_{ \mathbf{e} }-\lambda_1)} \;, \label{eq:result1} \\[1 ex] \frac{b_2}{a_2} &= \frac{\underset{i}{\sum} \mathbf{h}_i^{*}\cdot \mathbf{m_{(e)}}_{i} } {\alpha_{\scriptscriptstyle H}(\lambda_{ \mathbf{h}}-\lambda_2)} \;, \label{eq:result2} \end{aligned}\] where \(\lambda_{ \mathbf{e} }\) and \(\lambda_{ \mathbf{h}}\) are the eigenvalues of \(| \mathbf{e} \rangle\) and \(| \mathbf{h} \rangle\) in the decoupled electric and magnetic dipole equations in Eq. [\[eq:decoupled equations\]](#eq:decoupled equations){reference-type="ref" reference="eq:decoupled equations"}: The two ratios in Eq. [\[eq:result1\]](#eq:result1){reference-type="ref" reference="eq:result1"} and Eq. [\[eq:result2\]](#eq:result2){reference-type="ref" reference="eq:result2"} each describe a distinct eigenmode for the electromagnetic system according to Eq. [\[eq:EMeigenmode\]](#eq:EMeigenmode){reference-type="ref" reference="eq:EMeigenmode"}. It is worth noting that, if either numerator or denominator go to zero in Eq. [\[eq:result1\]](#eq:result1){reference-type="ref" reference="eq:result1"} or Eq. [\[eq:result2\]](#eq:result2){reference-type="ref" reference="eq:result2"}, the starting basis vectors of the decoupled electric or magnetic dipole equations were already eigenmodes of the full system. In regard to the \(\mathrm{A_{2g}}\) (magnetic-like) eigenmodes of an all-dielectric quadrumer, our derivations show that, when both eigenmodes are far from resonance and magneto-electric coupling can be neglected, one eigenmode will be purely composed of electric dipoles and the other will be purely composed of magnetic dipoles. However, as they approach resonance, magneto-electric coupling cannot be neglected and the two ratios of \(a_x\) and \(b_x\) will need to be calculated to determine their form as electromagnetic eigenmodes (see Fig. [\[fig:interference\]](#fig:interference){reference-type="ref" reference="fig:interference"}c). In such a situation, the two eigenmodes, \(| \mathbf{v}_1\rangle\) and \(| \mathbf{v}_2\rangle\), are explicitly nonorthogonal to each other with the overlap defined by: \[\begin{aligned} \langle \mathbf{v}_1|\mathbf{v}_2 \rangle = a_1 ^* a_2 + b_1 ^* b_2\. \label{eq:overlap} \end{aligned}\] We can therefore expect interference between these eigenmodes as they approach resonance.  Indeed, as shown in Fig. [\[fig:interference\]](#fig:interference){reference-type="ref" reference="fig:interference"}, the Fano resonance feature in the silicon quadrumer of Fig. [\[fig:CSTvsCEMDA\]](#fig:CSTvsCEMDA){reference-type="ref" reference="fig:CSTvsCEMDA"} is produced entirely by the interference between these two eigenmodes. To serve as a broader qualification of this magnetic Fano resonance feature, we also investigate its parameter dependence. In Fig. [\[fig:sweep\]](#fig:sweep){reference-type="ref" reference="fig:sweep"}a and [\[fig:sweep\]](#fig:sweep){reference-type="ref" reference="fig:sweep"}b, we vary the size of the gap between nanoparticles and the size of each individual nanoparticle, respectively. It can be seen that the Fano resonance is dependent on a number of these coupling parameters, but in a way that is different to electric Fano resonances in plasmonic oligomers. Perhaps most notably, the magnetic Fano resonance is significantly shifted with small changes in size of the constituent nanoparticles, and closely spaced nanoparticles are favorable for electric Fano resonances in plasmonic oligomers, but can destroy the magnetic Fano resonance here.  More specifically, the magnetic Fano resonance feature is heavily dependent on the spacing between neighboring nanoparticles; increasing or decreasing the spacing can quickly diminish the Fano feature. On the other hand, varying the size of the particles while holding the gap between particles constant is able to conserve the Fano resonance and shift it spectrally (shown in the shaded region of Fig. [\[fig:sweep\]](#fig:sweep){reference-type="ref" reference="fig:sweep"}b). This behavior can be expected because the action of increasing the size of the particles while holding the gap constant is very similar to uniformly scaling Maxwell's equations, and silicon permittivity has relatively minor dispersion in this spectral range.  Otherwise, the final parameter we consider in Fig. [\[fig:sweep\]](#fig:sweep){reference-type="ref" reference="fig:sweep"}c, is the angle of incidence to address the practical limitations for in-plane excitation of nanoscale optical structures. It can be seen that the Fano resonance will persist up to roughly 45 incidence, beyond which the magnetic response is dominated by the normal-incidence responses. To verify and validate the theoretical arguments presented above, we look for the existence of magnetic Fano resonances in a high index cluster experimentally. In this regard, one practical option is to mimic the scattering properties of silicon nanoparticles using MgO-\(\mathrm{TiO_2}\) ceramic spheres characterized by dielectric constant of 16 and dielectric loss factor of \((1.12 - 1.17)\times10^{-4}\), measured at \(9-12\) GHz. These ceramic spheres in the microwave range therefore have very similar properties to silicon nanospheres in the optical range and they are subsequently a useful macroscopic platform on which to prototype silicon nanostructures. Here, they allow us to perform a 'proof of concept' investigation into the properties of an isolated quadrumer with much more signal than would be expected from a single silicon nanosphere quadrumer. Indeed, such spheres have been used previously to predict the behavior of silicon nanoantennas.  The MgO-\(\mathrm{TiO_2}\) quadrumer consists of four dielectric spheres with diameter \(d=15\mathrm{mm}\), and the size of the gap between the particles is \(s=5\, \mathrm{ mm}\). The experimentally measured, and numerically calculated, total scattering of the quadrumer structure are shown in Fig. [\[fig:experiment\]](#fig:experiment){reference-type="ref" reference="fig:experiment"}. It can clearly be seen that a magnetic Fano resonance is produced at 5.4 GHz, in both simulation and experiment. This is the first example of a magnetic-magnetic Fano resonance in a single symmetric metamolecule. Notably, this Fano resonance occurs in a spectral range where the single particle is not at resonance, which demonstrates its collective nature. Indeed, it appears near the intersection of the single particle's electric and magnetic scattering contributions, reflecting that the overlap of eigenmodes (Eq. [\[eq:overlap\]](#eq:overlap){reference-type="ref" reference="eq:overlap"}) is dependent on both electric and magnetic dipole polarizabilities. # Conclusions We have presented a comprehensive study of the interplay between the collective optically-induced magnetic responses of all-dielectric quadrumers and the individual magnetic responses of their constituent dielectric nanoparticles. We have been able to establish the theoretical basis behind the interaction between collective and individual magnetic responses in all-dielectric structures, providing a quantitative prediction of the interference between the quadrumer's magnetic responses leading to magnetic-magnetic Fano resonance features. We have also been able to experimentally observe the existence of a sharp magnetic-magnetic Fano resonance in a dielectric quadrumer. Such Fano resonance features demonstrate the unique potential that suitably designed dielectric nanoclusters have for scattering engineering at the nanoscale, opening exciting and unexplored opportunities for dielectric nanophotonics. # Methods To fasten together the MgO-\(\mathrm{TiO_2}\) ceramic spheres for the experiment, we used a custom holder made of a styrofoam material with dielectric permittivity of 1 (in the microwave frequency range). To approximate plane wave excitation, we employed a rectangular horn antenna (TRIM \(0.75-18\) GHz; DR) connected to the transmitting port of a vector network analyzer (Agilent E8362C). The quadrumer was then located in the far-field of the antenna, at a distance of approximately 2.5 m, and a second horn antenna (TRIM \(0.75-18\) GHz; DR) was used as a receiver to observe the transmission through the quadrumer. The extinction measurement shown in Fig. [\[fig:experiment\]](#fig:experiment){reference-type="ref" reference="fig:experiment"}b was then obtained as the difference between the measured transmission and unity transmission (*i.e.*, with no quadrumer). For the theory, CST simulations were performed assuming plane wave excitation on a quadrumer located in free space. The CEMDA simulations, seen in Fig. [\[fig:CSTvsCEMDA\]](#fig:CSTvsCEMDA){reference-type="ref" reference="fig:CSTvsCEMDA"}a, used electric and magnetic dipole polarizabilities that were derived from the scattering coefficients of Mie theory.  In all simulations, silicon permittivity data was taken from Palik's Handbook  and the MgO-\(\mathrm{TiO_2}\) spheres were assumed be dispersionless with a dielectric constant of 16 and dielectric loss factor of \((1.12 - 1.17)\times10^{-4}\). # Results and Discussion Consider the optical response of a symmetric quadrumer when it is excited by a plane wave, whose propagation direction and electric field polarization lie in the plane of the quadrumer (\(s\)-polarization), as shown in Fig. [\[fig:schematic\]](#fig:schematic){reference-type="ref" reference="fig:schematic"}a. The conventional magnetic response to such an excitation is supported by a collective circulation of electric displacement current around all four nanoparticles. However, if the nanoparticles are made of a high-index dielectric material such as silicon, the individual nanoparticles are also able to couple directly with the applied magnetic field, sustaining internal circulating polarization currents. In this sense, a silicon quadrumer can produce magnetic responses to both electric *and* magnetic components of the applied plane wave, as is depicted in Fig. [\[fig:schematic\]](#fig:schematic){reference-type="ref" reference="fig:schematic"}b. In this figure, the red arrows denote an electric dipole, and the blue circular arrows denote the circulation of displacement current that supports an out-of-page magnetic dipole. Importantly, an interplay may be induced between the two magnetic responses through the locally-enhanced, \(z\)-polarized magnetic field sustained by the collective circulation of displacement current, which establishes a coupling channel to the magnetic responses of the individual nanoparticles. We begin by understanding this magnetic interplay, for which we will model the interactions between nanoparticles using the *Coupled Electric and Magnetic Dipole Approximation*  (CEMDA). The mathematical description of this dipole model in free space is based on the following equations: where \(\mathbf{p}_i\) (\(\mathbf{m}_i\)) is the electric (magnetic) dipole moment of the \(i^{\mathrm{th}}\) particle, \(\hat{G}_{0}(\mathbf{r}_{i},\mathbf{r}_{j})\) is the free space dyadic Greens function between the \(i^{\mathrm{th}}\) and \(j^{\mathrm{th}}\) dipole, \(\alpha_{E}\) (\(\alpha_{H}\)) is the electric (magnetic) polarizability of a particle, \(c_0\) is the speed of light and \(k\) is the free-space wavenumber. In Fig. [\[fig:CSTvsCEMDA\]](#fig:CSTvsCEMDA){reference-type="ref" reference="fig:CSTvsCEMDA"}a, we show that this dipole model is accurate in modeling full-wave simulations of the silicon nanosphere quadrumers performed using CST Microwave Studio, confirming that, in this spectral range, the optical response of such nanoparticle quadrumers is dominated by dipole interactions among the individual particles. This is the case because higher order coupling in such systems will occur only with smaller interparticle separations  (see also the Supporting Information). Fig. [\[fig:CSTvsCEMDA\]](#fig:CSTvsCEMDA){reference-type="ref" reference="fig:CSTvsCEMDA"}b shows the electric field distribution at the collective magnetic resonance of the cluster, confirming that the response of the silicon quadrumer includes a collective circulation of electric field around the particles. However, we can also see the additional dynamic produced by dielectric nanoparticles: the magnetic response induced by circulating transverse electric dipoles is accompanied by a magnetic response from each of the individual nanoparticles. This is the unique effect we are interested in. The magnetic response in an individual silicon nanosphere is an internal circulation of polarization current, which can be seen in Fig. [\[fig:CSTvsCEMDA\]](#fig:CSTvsCEMDA){reference-type="ref" reference="fig:CSTvsCEMDA"}b. In the CEMDA we choose to homogenize this circulating current distribution into a discrete source of magnetization current; the magnetic dipole. This implies that the key to strong interaction between individual and collective magnetic responses resides in the strong coupling between induced electric and magnetic dipoles in each inclusion. To describe this coupling, we can consider the eigenmodes of the quadrumer system. As we will show, the eigenmodes of the full, electromagnetic, system can be constructed from the eigenmodes of the decoupled electric and magnetic dipole systems. Therefore, we are going to break apart our dipole system of Eq. [\[eq:dipole equations\]](#eq:dipole equations){reference-type="ref" reference="eq:dipole equations"} into two decoupled equations: one equation for the electric dipoles and one equation for the magnetic dipoles We can then define two sets of eigenmodes from these equations: one for the electric dipoles and one for the magnetic dipoles. Using state notation, we can refer to the electric dipole eigenmodes as \(\left | \mathbf{e} \right \rangle\) and the magnetic dipole eigenmodes as \(\left | \mathbf{h} \right \rangle\). The eigenmode approach in the dipole approximation offers a huge simplification when combined with symmetry, because it allows us to determine certain eigenmodes without calculation. The key is that, for both electric and magnetic dipole systems, the dipole approximation restricts the number of eigenmodes for each equation to twelve (three dimensions by four dipoles). Each of these eigenmodes can only transform according to a single irreducible representation of the quadrumer's symmetry group, referred to as \(\mathrm{D_{4h}}\). Further details regarding the \(\mathrm{D_{4h}}\) symmetry group's irreducible representations and the implications for eigenmodes are provided in the Supporting Information. For the analysis here, it suffices to say that there are only eight irreducible representations in \(\mathrm{D_{4h}}\) for twelve eigenmodes, and therefore the eigenspace associated with certain irreducible representations must be one-dimensional. Moreover, given that any vector in a one-dimensional space is an eigenvector by default, we are able to derive a number of eigenmodes by simply finding dipole moment profiles that transform according to certain irreducible representations. Eight such dipole moment profiles are shown in Fig. [\[fig:eigenmodes\]](#fig:eigenmodes){reference-type="ref" reference="fig:eigenmodes"}. Each of these dipole moment profiles is the sole basis vector for a single irreducible representation and is therefore an eigenmode of the electric or magnetic dipole equations in Eq. [\[eq:decoupled equations\]](#eq:decoupled equations){reference-type="ref" reference="eq:decoupled equations"}, irrespective of wavelength or the choice of material, size, or any parameter which conserves the symmetry of the quadrumer. It is worth noting that this same procedure for finding eigenmodes is applicable to other symmetries when using the dipole approximation, particularly the \(\mathrm{D}_{n\mathrm{h}}\) symmetry groups for small values of \(n\). We need to now consider the interaction between electric and magnetic dipole systems. An eigenmode of either electric or magnetic dipoles can be substituted into Eq. [\[eq:dipole equations\]](#eq:dipole equations){reference-type="ref" reference="eq:dipole equations"} to determine the resulting state of magnetic or electric dipoles (\(| \mathbf{m_{(e)}} \rangle\) or \(| \mathbf{p_{(h)}} \rangle\), respectively) that it will induce due to magneto-electric coupling: However, it has been shown that the operation describing the coupling between states of electric and magnetic dipole moments, must commute with the geometry's symmetry operations.  Subsequently the \(\left | \mathbf{e}\right \rangle\) and \(| \mathbf{m_{(e)}} \rangle\) dipole moments must both transform according to the same irreducible representation, and similarly for the \(\left | \mathbf{h}\right \rangle\) and \(| \mathbf{p_{(h)}} \rangle\) dipole moments. Therefore, if we consider the eigenmodes of electric and magnetic dipoles in Fig. [\[fig:eigenmodes\]](#fig:eigenmodes){reference-type="ref" reference="fig:eigenmodes"}, the requirement for symmetry conservation specifies which electric dipole eigenmode can couple into which (if any) magnetic dipole eigenmode, and *vice versa*. As a result, the \(\mathrm{A_{2g}}\) or \(\mathrm{B_{1g}}\) eigenmodes (*cf.* Fig. [\[fig:eigenmodes\]](#fig:eigenmodes){reference-type="ref" reference="fig:eigenmodes"}) are only able to magneto-electrically couple into the other \(\mathrm{A_{2g}}\) or \(\mathrm{B_{1g}}\) eigenmode, whereas the \(\mathrm{A_{1g}}\) and \(\mathrm{B_{2g}}\) eigenmodes are not able to couple into any eigenmodes due to a symmetry mismatch. This conclusion allows us to take our analysis of the decoupled electric and magnetic dipole equations, and use it to determine the full, *electromagnetic*, eigenmodes of the real system. Indeed, we would not necessarily expect an eigenmode of this complex scattering system to be a current distribution described by purely electric or magnetic dipoles; we would expect it to be a combination of both. If we now restrict ourselves to considering only the \(\mathrm{A_{2g}}\) (magnetic-like) eigenmodes of an all-dielectric quadrumer, the two eigenmodes from the decoupled electric and magnetic dipole equations (*cf.* Fig. [\[fig:eigenmodes\]](#fig:eigenmodes){reference-type="ref" reference="fig:eigenmodes"}) form basis vectors for the \(\mathrm{A_{2g}}\) eigenmodes, \(| \mathbf{v}_x \rangle\), of the electromagnetic system. \[\begin{aligned} | \mathbf{v}_x \rangle = a_x \left ( \!\! \begin{array}{c} | \mathbf{e} \rangle \\ 0 \end{array} \!\!\right) + b_x \left (\!\! \begin{array}{c} 0 \\ | \mathbf{h} \rangle \end{array} \!\! \right)\,,\label{eq:EMeigenmode} \end{aligned}\] where \(a_x\) and \(b_x\) are complex scalars. Notably, we should expect to have two distinct electromagnetic eigenmodes here because there are two distinct basis vectors and, subsequently, a two-dimensional eigenspace for the quadrumer's response. To derive expressions for \(a_x\) and \(b_x\), we can write the electromagnetic eigenvalue equation for Eq. [\[eq:dipole equations\]](#eq:dipole equations){reference-type="ref" reference="eq:dipole equations"}: where \(\lambda_x\) is the eigenvalue of the electromagnetic eigenmode. However, Eqs. [\[eq:eigenmode equations\]](#eq:eigenmode equations){reference-type="ref" reference="eq:eigenmode equations"}a and [\[eq:eigenmode equations\]](#eq:eigenmode equations){reference-type="ref" reference="eq:eigenmode equations"}b are not independent[^2] because one can be obtained from the other using duality transformations.  In other words, a solution for \(a_x\) and \(b_x\), that satisfies *either* Eq. [\[eq:eigenmode equations\]](#eq:eigenmode equations){reference-type="ref" reference="eq:eigenmode equations"}a or Eq. [\[eq:eigenmode equations\]](#eq:eigenmode equations){reference-type="ref" reference="eq:eigenmode equations"}b, must also satisfy the complete Eq. [\[eq:eigenmode equations\]](#eq:eigenmode equations){reference-type="ref" reference="eq:eigenmode equations"} and describe an electromagnetic eigenmode of Eq. [\[eq:EMeigenmode\]](#eq:EMeigenmode){reference-type="ref" reference="eq:EMeigenmode"}. So, if the basis vectors \(\left | \mathbf{e}\right \rangle\) and \(\left | \mathbf{h} \right \rangle\) are normalized to a magnitude of one, we can project Eq. [\[eq:eigenmode equations\]](#eq:eigenmode equations){reference-type="ref" reference="eq:eigenmode equations"}a and Eq. [\[eq:eigenmode equations\]](#eq:eigenmode equations){reference-type="ref" reference="eq:eigenmode equations"}b onto \(\mathbf{e}_{i}\) and \(\mathbf{h}_{i}\) (respectively) and sum over all \(i\), to obtain the two solutions for Eq. [\[eq:eigenmode equations\]](#eq:eigenmode equations){reference-type="ref" reference="eq:eigenmode equations"}: \[\begin{aligned} \frac{a_1}{b_1} &= \frac{ \underset{i}{\sum} \, \mathbf{e}_i^{*}\cdot \mathbf{p_{(h)}}_{i} } {\alpha_{\scriptscriptstyle E} \epsilon_0 (\lambda_{ \mathbf{e} }-\lambda_1)} \;, \label{eq:result1} \\[1 ex] \frac{b_2}{a_2} &= \frac{\underset{i}{\sum} \mathbf{h}_i^{*}\cdot \mathbf{m_{(e)}}_{i} } {\alpha_{\scriptscriptstyle H}(\lambda_{ \mathbf{h}}-\lambda_2)} \;, \label{eq:result2} \end{aligned}\] where \(\lambda_{ \mathbf{e} }\) and \(\lambda_{ \mathbf{h}}\) are the eigenvalues of \(| \mathbf{e} \rangle\) and \(| \mathbf{h} \rangle\) in the decoupled electric and magnetic dipole equations in Eq. [\[eq:decoupled equations\]](#eq:decoupled equations){reference-type="ref" reference="eq:decoupled equations"}: The two ratios in Eq. [\[eq:result1\]](#eq:result1){reference-type="ref" reference="eq:result1"} and Eq. [\[eq:result2\]](#eq:result2){reference-type="ref" reference="eq:result2"} each describe a distinct eigenmode for the electromagnetic system according to Eq. [\[eq:EMeigenmode\]](#eq:EMeigenmode){reference-type="ref" reference="eq:EMeigenmode"}. It is worth noting that, if either numerator or denominator go to zero in Eq. [\[eq:result1\]](#eq:result1){reference-type="ref" reference="eq:result1"} or Eq. [\[eq:result2\]](#eq:result2){reference-type="ref" reference="eq:result2"}, the starting basis vectors of the decoupled electric or magnetic dipole equations were already eigenmodes of the full system. In regard to the \(\mathrm{A_{2g}}\) (magnetic-like) eigenmodes of an all-dielectric quadrumer, our derivations show that, when both eigenmodes are far from resonance and magneto-electric coupling can be neglected, one eigenmode will be purely composed of electric dipoles and the other will be purely composed of magnetic dipoles. However, as they approach resonance, magneto-electric coupling cannot be neglected and the two ratios of \(a_x\) and \(b_x\) will need to be calculated to determine their form as electromagnetic eigenmodes (see Fig. [\[fig:interference\]](#fig:interference){reference-type="ref" reference="fig:interference"}c). In such a situation, the two eigenmodes, \(| \mathbf{v}_1\rangle\) and \(| \mathbf{v}_2\rangle\), are explicitly nonorthogonal to each other with the overlap defined by: \[\begin{aligned} \langle \mathbf{v}_1|\mathbf{v}_2 \rangle = a_1 ^* a_2 + b_1 ^* b_2\. \label{eq:overlap} \end{aligned}\] We can therefore expect interference between these eigenmodes as they approach resonance.  Indeed, as shown in Fig. [\[fig:interference\]](#fig:interference){reference-type="ref" reference="fig:interference"}, the Fano resonance feature in the silicon quadrumer of Fig. [\[fig:CSTvsCEMDA\]](#fig:CSTvsCEMDA){reference-type="ref" reference="fig:CSTvsCEMDA"} is produced entirely by the interference between these two eigenmodes. To serve as a broader qualification of this magnetic Fano resonance feature, we also investigate its parameter dependence. In Fig. [\[fig:sweep\]](#fig:sweep){reference-type="ref" reference="fig:sweep"}a and [\[fig:sweep\]](#fig:sweep){reference-type="ref" reference="fig:sweep"}b, we vary the size of the gap between nanoparticles and the size of each individual nanoparticle, respectively. It can be seen that the Fano resonance is dependent on a number of these coupling parameters, but in a way that is different to electric Fano resonances in plasmonic oligomers. Perhaps most notably, the magnetic Fano resonance is significantly shifted with small changes in size of the constituent nanoparticles, and closely spaced nanoparticles are favorable for electric Fano resonances in plasmonic oligomers, but can destroy the magnetic Fano resonance here.  More specifically, the magnetic Fano resonance feature is heavily dependent on the spacing between neighboring nanoparticles; increasing or decreasing the spacing can quickly diminish the Fano feature. On the other hand, varying the size of the particles while holding the gap between particles constant is able to conserve the Fano resonance and shift it spectrally (shown in the shaded region of Fig. [\[fig:sweep\]](#fig:sweep){reference-type="ref" reference="fig:sweep"}b). This behavior can be expected because the action of increasing the size of the particles while holding the gap constant is very similar to uniformly scaling Maxwell's equations, and silicon permittivity has relatively minor dispersion in this spectral range.  Otherwise, the final parameter we consider in Fig. [\[fig:sweep\]](#fig:sweep){reference-type="ref" reference="fig:sweep"}c, is the angle of incidence to address the practical limitations for in-plane excitation of nanoscale optical structures. It can be seen that the Fano resonance will persist up to roughly 45 incidence, beyond which the magnetic response is dominated by the normal-incidence responses. To verify and validate the theoretical arguments presented above, we look for the existence of magnetic Fano resonances in a high index cluster experimentally. In this regard, one practical option is to mimic the scattering properties of silicon nanoparticles using MgO-\(\mathrm{TiO_2}\) ceramic spheres characterized by dielectric constant of 16 and dielectric loss factor of \((1.12 - 1.17)\times10^{-4}\), measured at \(9-12\) GHz. These ceramic spheres in the microwave range therefore have very similar properties to silicon nanospheres in the optical range and they are subsequently a useful macroscopic platform on which to prototype silicon nanostructures. Here, they allow us to perform a 'proof of concept' investigation into the properties of an isolated quadrumer with much more signal than would be expected from a single silicon nanosphere quadrumer. Indeed, such spheres have been used previously to predict the behavior of silicon nanoantennas.  The MgO-\(\mathrm{TiO_2}\) quadrumer consists of four dielectric spheres with diameter \(d=15\mathrm{mm}\), and the size of the gap between the particles is \(s=5\, \mathrm{ mm}\). The experimentally measured, and numerically calculated, total scattering of the quadrumer structure are shown in Fig. [\[fig:experiment\]](#fig:experiment){reference-type="ref" reference="fig:experiment"}. It can clearly be seen that a magnetic Fano resonance is produced at 5.4 GHz, in both simulation and experiment. This is the first example of a magnetic-magnetic Fano resonance in a single symmetric metamolecule. Notably, this Fano resonance occurs in a spectral range where the single particle is not at resonance, which demonstrates its collective nature. Indeed, it appears near the intersection of the single particle's electric and magnetic scattering contributions, reflecting that the overlap of eigenmodes (Eq. [\[eq:overlap\]](#eq:overlap){reference-type="ref" reference="eq:overlap"}) is dependent on both electric and magnetic dipole polarizabilities. # Conclusions We have presented a comprehensive study of the interplay between the collective optically-induced magnetic responses of all-dielectric quadrumers and the individual magnetic responses of their constituent dielectric nanoparticles. We have been able to establish the theoretical basis behind the interaction between collective and individual magnetic responses in all-dielectric structures, providing a quantitative prediction of the interference between the quadrumer's magnetic responses leading to magnetic-magnetic Fano resonance features. We have also been able to experimentally observe the existence of a sharp magnetic-magnetic Fano resonance in a dielectric quadrumer. Such Fano resonance features demonstrate the unique potential that suitably designed dielectric nanoclusters have for scattering engineering at the nanoscale, opening exciting and unexplored opportunities for dielectric nanophotonics. # Methods To fasten together the MgO-\(\mathrm{TiO_2}\) ceramic spheres for the experiment, we used a custom holder made of a styrofoam material with dielectric permittivity of 1 (in the microwave frequency range). To approximate plane wave excitation, we employed a rectangular horn antenna (TRIM \(0.75-18\) GHz; DR) connected to the transmitting port of a vector network analyzer (Agilent E8362C). The quadrumer was then located in the far-field of the antenna, at a distance of approximately 2.5 m, and a second horn antenna (TRIM \(0.75-18\) GHz; DR) was used as a receiver to observe the transmission through the quadrumer. The extinction measurement shown in Fig. [\[fig:experiment\]](#fig:experiment){reference-type="ref" reference="fig:experiment"}b was then obtained as the difference between the measured transmission and unity transmission (*i.e.*, with no quadrumer). For the theory, CST simulations were performed assuming plane wave excitation on a quadrumer located in free space. The CEMDA simulations, seen in Fig. [\[fig:CSTvsCEMDA\]](#fig:CSTvsCEMDA){reference-type="ref" reference="fig:CSTvsCEMDA"}a, used electric and magnetic dipole polarizabilities that were derived from the scattering coefficients of Mie theory.  In all simulations, silicon permittivity data was taken from Palik's Handbook  and the MgO-\(\mathrm{TiO_2}\) spheres were assumed be dispersionless with a dielectric constant of 16 and dielectric loss factor of \((1.12 - 1.17)\times10^{-4}\).
{'timestamp': '2016-07-18T02:08:54', 'yymm': '1607', 'arxiv_id': '1607.04592', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04592'}
# Introduction The original motivation for warped extra-dimensions was to address the hierarchy problem, so that the fundamental scale of gravity is exponentially reduced along the extra dimension, from the Planck mass scale to the TeV scale. Thus, the TeV scale becomes the natural scale of the Higgs sector if this one is localized near the TeV boundary of the extra dimension, as first introduced by Randall and Sundrum (RS). If SM fields are allowed to propagate in the extra dimension, the scenario can also address the flavor puzzle of the SM, explaining fermion masses and mixings from the geographical location of fields along the extra dimension. However, processes mediated by the heavy resonances of the 5D bulk fields, Kaluza-Klein (KK) modes, generate dangerous contributions to electroweak and flavor observables (including dangerous deviations to the \(Zb\bar{b}\) coupling), pushing the KK mass scale to \(5-10\) TeV. A popular mechanism to lower the KK scale involves using a custodial gauge \(SU(2)_R\) symmetry, which ensures a small contribution to electroweak precision parameters, lowering the KK scale bound to about 3 TeV. Alternatively, one can study scenarios in which the metric is slightly modified from the RS metric background (\(AdS_5\)). This can be achieved quite naturally from the backreaction on the metric caused by a 5D scalar field stabilizing the original \(AdS_5\) warped background. When the 5D Higgs is sufficiently leaking into the bulk and when the metric background is modified near the TeV boundary, the scenario allows for KK scales as low as 1-2 TeV, with precision electroweak and flavor constraints under control. An inconvenience is that these scenarios are typically hard to probe experimentally as the couplings of all particles are very suppressed. Still, it has been shown that it can still lead to interesting deviations in Higgs phenomenology, as the Higgs couplings can receive sufficient radiative corrections from the many KK fermions of the model. Run 2 LHC data at \(\sqrt{s}=13\) TeV shows signals of a new resonance in the diphoton distribution at an invariant mass of 750 GeV with a 3.9 significance at ATLAS, with 3.2 fb\(^{-1}\) and 3.4 combined significance at CMS, with 2.6 fb\(^{-1}\)(combining run 1 and run 2 results). ATLAS reports 14 events and CMS, 10. The experimental data is summarized in Table [2](#tab:siglims){reference-type="ref" reference="tab:siglims"}. ) and the covariant derivative is \(D_M= \partial_M + i g_5 A_M\) with [\[Aexpansion\]]{#Aexpansion label="Aexpansion"} A_M=( ) and CP-odd and charged Higgs part is [\[Piexpansion\]]{#Piexpansion label="Piexpansion"} =( ) with the weak angle defined like in the SM, i.e. \(s_W/c_W=g'_5/g_5\), where \(g_5\) and \(g_5'\) are the 5D coupling constants of \(SU(2)_L\) and \(U(1)_Y\). The extraction of degrees of freedom in this context has been performed in and we outline here the main results. The effect of brane kinetic terms in the Higgs sector is new and its derivation is outlined in the Appendix. The 5D equations of motion for all these fields are coupled (except for the case of the real Higgs excitation \(h(x,y)\)) and in order to decouple them, one can partially fix the gauge, or add a gauge fixing term to the previous 5D action. For example, in the CP-odd case, the fields \(Z_\mu(x,y)\), \(Z_5(x,y)\) and \(\Pi_z(x,y)\) must be unmixed. The partial gauge fixing constraint[^2] \^Z\_-M\^2_z(y) \_z +(e\^-2Z_5)'=0 manages to decouple the fields \(Z_\mu\) from \(Z_5\) and \(\Pi_z\) in the bulk. We defined here \(M_z(y)=\frac{g_5}{2c_W} v(y) e^{-\sigma(y)}\). However, the presence of the Higgs brane kinetic terms, proportional to \(d_i\) in the action, forces us to extend the gauge choice on the branes, producing a lifting of \(Z_5\) field so that the decoupling is maintained at the boundaries[^3]. The appropriate boundary condition at the IR brane is Z_5(x,y_1) =- M\^2_z(y_1) e\^2(y_1) \_z(x,y_1) where \(y_1\) denotes the position of the boundary (note that if the brane kinetic term parameter \(d_1\) tends to zero, the condition on \(Z_5\) becomes Dirichlet, as expected). With this type of gauge choice, the 5D fields \(Z_\mu\), \(W_\mu\) and \(A_\mu\) have independent 5D equations of motion. In order to extract the effective 4D degrees of freedom, we expand the gauge fields as \(Z_\mu(x,y)=Z^n_\mu(x) f^n_z(y)\), \(W_\mu(x,y)=W^n_\mu(x) f^n_w(y)\) and \(A_\mu(x,y)=A^n_\mu(x) f^n_\gamma(y)\) (summation over \(n\) is understood) and where \(Z^0_\mu(x)\), \(W^0_\mu(x)\) and \(A^0_\mu(x)\) are the \(Z\), \(W\) and \(\gamma\) gauge bosons of the SM. The extradimensional profiles \(f^n_z(y)\), \(f^n_w(y)\) and \(f^n_\gamma(y)\) are solutions of [\[gaugeeq\]]{#gaugeeq label="gaugeeq"} (e\^-2 f'\_a)'+(m\^2\_n-M_a\^2(y)) f_a =0 where \(a=z,w,\gamma\) and where \(M_z(y)=\frac{g_5}{2c_W} v(y) e^{-\sigma(y)}\), as defined before, \(M_w(y)=\frac{g_5}{2} v(y) e^{-\sigma(y)}\), and \(M_\gamma=0\). The boundary conditions for these profiles are[^4] [\[gaugebc\]]{#gaugebc label="gaugebc"} M_a\^2 f_a(y_i) =-e\^-2f'\_a(y_i). The CP-even Higgs field is expanded as \(h(x,y)=h^n(x) h^n_y(y)\) and the equations for the Higgs profiles are, with \(h_y \equiv h_y^n(y)\): (e\^-4 h'\_y)'+(m\^2\_h_n-\^2\_bulk) h_y =0, where \(\displaystyle \mu^2_{bulk} = \frac{\partial^2 V }{\partial H^2}\Big|_{H=v}\). The boundary conditions are (\^2\_brane_i- m\^2\_h_n e\^2) h_y =-e\^-4 h'\_y, with \(\displaystyle \mu^2_{brane_i} = \frac{\partial^2 \lambda_i }{\partial H^2}\Big|_{H=v}\). Note that the mode \(h^0(x)\) is interpreted as the SM Higgs boson. There are still some degrees of freedom left, and their 5D equations of motion still happen to be mixed. One of the coupled systems involves \(Z_5\) and \(\Pi_z\) and the other coupled system involves \(\Pi^{\pm}\) and \(W_5^\pm\). In order to disentangle these systems one must perform a mixed expansion, so that the decoupling of fields will happen KK level by KK level. The mixed expansions are, in the CP-odd sector, [\[Z5decomp\]]{#Z5decomp label="Z5decomp"} Z_5(x,y)&=& G\^n(x) + \_n(x)X\_(y)\ \_z(x,y)&=& G\^n(x) + \_n(x) X\_'(y), [\[Pidecomp\]]{#Pidecomp label="Pidecomp"} and in the charged scalar sector they are W_5\^(x,y) &=& G_n\^(x) + \^\_n(x)X\_(y)  \ \^(x,y)&=& G_n\^(x) +\^\_n(x) X\_'(y).  where \(M_z(y)\) and \(M_w(y)\) were defined below Eq ([\[gaugeeq\]](#gaugeeq){reference-type="ref" reference="gaugeeq"}). The effective 4D physical fields are the tower of CP-odd neutral scalars \(\Pi_n(x)\) and the tower of charged scalars \(\Pi^\pm_n(x)\). Their associated extra-dimensional profiles \(X_\pi(y)\) and \(X_\pm(y)\) obey the equations ( X_a' )'+ (-1 ) e\^2 X_a=0 [\[Xeq\]]{#Xeq label="Xeq"} where \(M_a(y)=(M_z(y),M_{w}(y))\) and \(X_a=(X_\pi,X_{\pm})\). The boundary conditions are [\[Xbc\]]{#Xbc label="Xbc"}  X'\_a=-X_a, and note that vanishing Higgs brane kinetic terms implies Dirichlet boundary conditions for \(X_a\). We checked that these bulk equations agree with , the only new addition being the boundary conditions imposed by the presence of Higgs brane kinetic terms. In order for these 4D scalars to be canonically normalized, we require dy e\^2 =1 and this condition includes the effect of Higgs brane kinetic terms. The remaining 4D fields are \(G_n(x)\) and \(G_n^\pm(x)\), which are Goldstone bosons at each KK level. The profile wavefunctions \(f_{G_a}(y)\) obey the same differential equations as the gauge profiles, Eq.([\[gaugeeq\]](#gaugeeq){reference-type="ref" reference="gaugeeq"}), as well as the same boundary conditions, Eq.([\[gaugebc\]](#gaugebc){reference-type="ref" reference="gaugebc"}). The spectrum is thus identical to the gauge bosons spectrum level by level. These fields appear in the effective 4D action coupled to (\(\partial^\mu Z^n_\mu\)) or (\(\partial^\mu W^n_\mu\)), and of course there is a leftover gauge freedom allowing us to gauge them away (i.e., they are pure gauge). We wish to identify the lightest CP-odd scalar \(\Pi_0(x)\) with the observed diphoton peak at the LHC, so that we need to fix its mass \(m_{\Pi_0}= 750\) GeV. In order to have an idea of the effects of the Higgs brane kinetic terms on the CP-odd scalar spectrum, we consider two different parameter points, one in which the RS background metric is recovered with \(\nu=10\) and \(y_s=4\times y_{1}\) (\(\nu\) is the exponent appearing in the modified metric and if relatively large, the location of spurious singularity is sent away from the boundary, recovering essentially the \(AdS_5\) metric). The other case is the situation where the metric modification allows for TeV size KK masses, which are safe from precision electroweak constrains. The parameters chosen there are \(\nu=0.5\) and \(y_s=1.04\times y_{1}\). In both parameter points, we fix the KK mass of the first gluon excitation to be \(1500\) GeV[^5]. In Fig. [\[fig:Xmassplot\]](#fig:Xmassplot){reference-type="ref" reference="fig:Xmassplot"} we show the spectrum of the first 4 KK levels of CP-odd Higgs bosons \(\Pi_n(x)\), for \(n=0,1,2,3\), as a function of the brane kinetic term \(d_{1}\) in units of the curvature \(k\sim M_{Pl}\). The effects of the UV localized brane kinetic term are warped suppressed and so we do not consider them here anymore. We can see that in the RS limit, it is not possible to reduce sufficiently the lightest CP-odd mass to 750 GeV, as its mass tends asymptotically from about \(2500\) GeV without brane kinetic term to about \(1750\) GeV for large brane kinetic terms. On the other hand, with the modified metric it becomes possible to reduce greatly the lightest CP-odd mass with relatively small brane kinetic term coefficients (which in this particular case tends asymthotically to about 400 GeV). Parameter points such that the metric modification lies between the two considered will actually have an intermediate behavior, with a lightest CP-odd mass having increasing asymptotic values as one recovers the RS background. Finally also note that the spectrum for the charged scalars is essentially the same since their differential equations and boundary conditions are identical except for the functions \(M_z(y)\) and \(M_w(y)\), which differ by about \(10\%\). \(M_w(y)\) produces a deviation from the CP-odd scalar spectrum of less than \(5\%\). This means that the scenario under consideration should also contain a lightest charged Higgs scalar with a mass of about \(750\) GeV also. The next question to ask is how big is the effect of the Higgs brane kinetic term on the gauge bosons, and in particular on the lowest ones, i.e. the SM \(W\) and \(Z\) bosons. These terms represent an additional (brane localized) contribution to the mass of the gauge bosons. In principle their mass is generated here from a bulk Higgs mechanism, unless the brane kinetic terms are overly important (not the limit we are working with here). We can quickly estimate its effect on the lowest lying gauge fields. These are essentially flat (like all gauge zero modes) and thus their wave function is \(f^0_z\sim 1/\sqrt{y_{1}}\). The contribution of a brane localized mass squared term is \(\delta m^2_z \mathrel{\raise.3ex\hbox{\)<\(\kern-.75em\lower1ex\hbox{\)\sim\(}}} d_{1}\ v^2/y_{1}\sim d_{1}\times 700\) GeV\(^2\). For IR brane kinetic term coefficient \(d_{1}\) of \({\cal O}(1)\), this represents naively at most some \(10\%\) contribution to the overall mass squared of either \(W\) or \(Z\). In the particular case of the modified metric with a brane coefficient \(d_{1}\simeq 0.51\), and metric parameters \(\nu =0.5\) and \(y_s=1.04 \times y_1\) (which produces a light CP-odd scalar of \(750\) GeV), the exact numerical effect on the zero mode gauge boson masses (\(W\) and \(Z\)) is a shift of \(3\) GeV with respect to the no-brane-kinetic-term limit. Of course, in the presence of brane kinetic terms, one redefines the VEV normalization constant, and the value of \(g_5\), in order to correctly account for the SM gauge boson masses and electroweak couplings. # CP-odd Higgs couplings {#sec:cpodd} As a \(750\) GeV CP-odd Higgs scalar \(\Pi_0(x)\) is allowed in the spectrum, thanks to the effect of small brane localized Higgs kinetic terms, we now study its couplings to SM particles in order to see if the observed excess at the LHC can be associated with this excitation. Of course being a CP-odd scalar its tree-level couplings to \(ZZ\) and \(WW\) are zero, making it an ideal candidate for the observed exotic events. We thus need to focus on its tree-level couplings to fermions (and top quark in particular), to \(Zh\) (where \(h\) is the \(125\) GeV Higgs) and to its radiative couplings to photons and gluons. We study these in the subsequent subsections. ## Fermion couplings {#subsec:fermion} The couplings of \(\Pi_0(x)\) to fermions arise from two sources in the action. First source comes from the 5D Higgs Yukawa couplings, and second, from the gauge fermion couplings. This is because the physical field \(\Pi_0(x)\) contains some of CP-odd Higgs scalar, and some of \(Z_5\) excitation, where \(Z_5\) is the fifth component of the 5D vector boson \(Z_M\). However the 5D Yukawa coupling allows for direct coupling of \(\Pi_0(x)\) to two zero-mode fermions, whereas the gauge-fermion coupling allows only couplings between fermion zero-modes and higher KK fermion levels. As we will see, it is important to keep both couplings, since after electroweak symmetry breaking the physical SM fermions (top quarks in particular) are mostly zero-modes but also contain a small amount of higher KK excitations, and could thus inherit some of the original gauge-fermion coupling, especially if the tree-level Yukawa coupling between \(\Pi_0(x)\) and zero-mode top quarks is suppressed (as it can be). The relevant terms in the action are the 5D Higgs Yukawa couplings and the fermion gauge interaction term, S\_ff_0 d\^4xdy where \(Q,U,D\) represent the 5D fermion \(SU(2)_L\) doublets, up-type and down-type singlets (with generation indices and isospin indices suppressed). The kinetic terms contain the 5D covariant derivative and from them we extract the terms containing the CP-odd component \(Z_5(x,y)\), and from the Higgs Yukawa couplings we extract the terms containing the CP-odd Higgs component \(\Pi_z(x,y)\). We follow the approach of and compute these couplings by considering only the effects of three full KK levels, i.e. computing \(21 \times 21\) fermion Yukawa coupling matrices (with 3 \(up\) and 3 \(down\) families, each containing zero modes and 3 KK levels with an \(SU(2)_L\) doublet and a singlet in each level, i.e., 3 zero modes plus \(3\times 3 \times 2\) KK modes). Note that we are interested in the couplings of the \(750\) GeV CP-odd scalar \(\Pi_0(x)\) to SM fermions (top quarks primarily), but we also need its couplings with the rest of KK fermions, since these interactions will be crucial to generate large enough radiative couplings to photons and gluons. We first write the effective 4D up-type quark mass matrix as ( ) M_u ( ) in a basis where \(q^0_L(x)\) and \(u_R^0(x)\) represent three zero-mode flavors each (doublets and singlets of \(SU(2)_L\)), and \(Q_L(x)\) and \(Q_R(x)\) represent three flavors and three KK levels of the vector-like KK up-type doublets, and \(U_L(x)\) and \(U_R(x)\) represent three flavors and three KK levels of vector-like KK up-type singlets. The mass matrix is thus M_u= ( ) [\[mumatrix\]]{#mumatrix label="mumatrix"} with the down sector mass matrix \({M}_d\) computed in the same way. The submatrices are obtained by evaluating the overlap integrals [\[YYY1KK\]]{#YYY1KK label="YYY1KK"} y\^0_u= \_0\^y_1 dy e\^-4(y) q\^0,i\_L(y)u\^0,j\_R(y)\ Y\^qU= \_0\^y_1 dy e\^-4(y) q\^0,i\_L(y)U\^n,j\_R(y)\ Y\^Qu = \_0\^y_1 dy e\^-4(y) Q\^m,i\_L(y)u\^0,j\_R(y)\ Y_1 = \_0\^y_1 dy e\^-4(y) Q\^m,i\_L(y)U\^n,j\_R(y)\ Y_2 = \_0\^y_1 dy e\^-4(y) Q\^m,i\_R(y)U\^n,j\_L(y) , where the indices \(m\) and \(n\) track the KK level and \(i,j=1,2,3\) are 5D flavor indices. The diagonal matrices \((M_Q)_{9\times 9}\) and \((M_U)_{9\times 9}\) are constructed with the masses of all the KK quarks involved. The masses and the profiles of the KK fermions appearing in these overlap integrals (\(Q_L(y)\), \(Q_R(y)\), \(U_L(y)\) and \(U_R(y)\)) are obtained by solving differential equations for the fermion profiles \_y(e\^(2c-1)(y)\_y( e\^-(c+2)(y)))f(y)+e\^(c-1)(y) m_n\^2 f(y) =0 [\[fermionprofile\]]{#fermionprofile label="fermionprofile"} where \(f(y)\) is the KK profile. The mass eigenvalues \(m_n\) are found by imposing Dirichlet boundary conditions on the wrong chirality modes. As mentioned before, we have included 3 full KK levels so that the mass matrices in the gauge basis are \(21 \times 21\) dimensional matrices, which are not diagonal. One needs to diagonalize them, and by doing so, move to the quark physical basis where all the fermions couplings can then be extracted. In the CP-odd scalar sector, we can write the effective 4D Yukawa-type couplings to fermions in the same gauge basis as before ( ) **Y**\_( )  \_0(x) where now the \(21 \times 21\) coupling matrix \({\bf Y}_{\pi}\) is given by \_= ( ) [\[Ypimatrix\]]{#Ypimatrix label="Ypimatrix"} The submatrices are obtained by the overlap integrals [\[YYPI\]]{#YYPI label="YYPI"} y\^0\_qu &=& i \_0\^y_1 dy e\^-3(y) q\^0,i\_L(y)u\^0,j\_R(y)\ Y\_qU&=& i \_0\^y_1 dy e\^-3(y) q\^0,i\_L(y)U\^n,j\_R(y)\ Y\_Qu& =& i \_0\^y_1 dy e\^-3(y) Q\^m,i\_L(y)u\^0,j\_R(y)\ Y\^\_1& =& i \_0\^y_1 dy e\^-3(y) Q\^m,i\_L(y)U\^n,j\_R(y)\ Y\^\_2& =& i \_0\^y_1 dy e\^-3(y) Q\^m,i\_R(y)U\^n,j\_L(y)  , and a\_qQ&=& \_0\^y_1 dy e\^-2(y) q\^0,i\_L(y)Q\^n,j\_R(y)\ a\_uU&=& \_0\^y_1 dy e\^-2(y) u\^0,i\_R(y)U\^n,j\_L(y)\ a\_QQ&=& \_0\^y_1 dy e\^-2(y) Q\^m,i\_L(y)Q\^n,j\_R(y)\ a\_UU&=& \_0\^y_1 dy e\^-2(y) U\^m,i\_R(y)U\^n,j\_L(y) where the \(g^{5D}_{L,R}\) coupling are given by g\^5D\_L = g\^5D\_W (T_3-Q_q \^2 \_W ) g\^5D\_R = g\^5D\_W Q_q \^2 \_W, with \(Q_q\) the charge of the quark, (here \({2 \over 3}\)), \(\theta_W\) the weak angle and \(T_3 = {1\over 2}\). Note that when the interaction originates in the 5D Yukawa couplings, the profile to use is the one coming from the CP-odd Higgs component, i.e proportional to \(X_\pi'(y)\). When the interaction originates in the gauge fermion coupling and thus comes from the \(Z_5\) component, the profile to use is proportional to \(X_\pi(y)\), with \(X_\pi\) being the solution of Eq. ([\[Xeq\]](#Xeq){reference-type="ref" reference="Xeq"}), using the decompositions of Eqs. ([\[Z5decomp\]](#Z5decomp){reference-type="ref" reference="Z5decomp"}) and ([\[Pidecomp\]](#Pidecomp){reference-type="ref" reference="Pidecomp"}). When the fermion matrix in ([\[mumatrix\]](#mumatrix){reference-type="ref" reference="mumatrix"}) is diagonalized, the coupling matrix of fermions with the CP-odd field \(\Pi_0(x)\) in ([\[Ypimatrix\]](#Ypimatrix){reference-type="ref" reference="Ypimatrix"}) is rotated, and we can then extract all the physical Yukawa couplings. All these couplings are needed later in order to compute the radiative couplings of \(\Pi_0(x)\) with gluons and photons. We first analyze the very important Yukawa coupling between \(\Pi_0(x)\) and top quarks, as this coupling might dominate the decays of the CP-odd scalar. This coupling comes essentially from the entry \((y^0_{\pi qu})_{ 33}\) (before rotation to the physical basis) although it receives small corrections after going to the physical basis. We focus on \((y^0_{\pi qu})_{ 33}\) which comes from the overlap integral [\[YPtt\]]{#YPtt label="YPtt"} (y\^0\_qu)\_33 &=& i \_0\^y_1 dy e\^-3(y) q\^0\_t(y)u\^0\_t(y) In the RS limit, the warp factor is \(\sigma(y)=ky\), and the top profiles are \(q^{0}_{t}(y)= f(c_q) e^{(2-c_q)ky}\) and \(u^{0}_{t}(y)=f(-c_u) e^{(2+c_u)ky}\), where \(f(x)\) is a normalization factor. We also have \(M_z(y)=\frac{g_5}{2c_W} v_0 e^{(a-1)ky}\), with \(v_0\) a constant factor, so that the previous overlap integral in this limit reads (y\^0\_qu)\_33 = i \_0\^y_1 dy e\^(2-a-c_q+c_u)k y X\_'(y) We integrate this by parts to find (y\^0\_qu)\_33 &=& -i \_0\^y_1 dy (2-a-c_q+c_u) e\^(2-a-c_q+c_u)k y X\_(y) + BT      where \(BT= i \frac{(Y^{5D}_u)_{33}}{\sqrt{2k}} \frac{2 c_W}{g_5}\frac{f(c_q) f(-c_u)}{v_0 m^2_{\pi_0}}\ e^{(2-a-c_q+c_u)k y} X_\pi(y)\Big|_0^{y_1}\) is a boundary term. Note that the profile \(X_\pi(y)\) has vanishing boundary conditions in the absence of Higgs localized brane kinetic terms. In that limit we can see that the coupling of the CP-odd scalar can actually vanish, when \((2-a-c_q+c_u)=0\) . Note also that the Higgs localizer parameter \(a\) is, in this RS limit, \(a \mathrel{\raise.3ex\hbox{\)>\(\kern-.75em\lower1ex\hbox{\)\sim\(}}} 2\) and the bulk parameters \(c_q\) and \(c_u\) are defined such that, for example, charm or bottom quarks are assigned values more or less \(c_q \in (0.45,0.55)\) and \(c_u\in (-0.5,-0.6)\), whereas for top we have \(c_{q_3}\sim 0.45\) and \(c_{u_3}>-0.45\). This means that we expect the term \((2-a-c_q+c_u)\) to vanish, in the limit of \(a\sim 2\), when \(c_q-c_u \sim 0\), so that the suppression in this case seems only possible for the top quark, where both \(c_q\) and \(c_u\) could be small. Of course when the metric background is modified, and the boundary conditions include the brane kinetic terms there will be deviations from the previous values. Nevertheless it is clear that the Yukawa coupling of the CP-odd scalar field to top quarks can have highly suppressed values. Another way to see this is to consider the overlap integral in Eq. ([\[YPtt\]](#YPtt){reference-type="ref" reference="YPtt"}). Because the profile \(X_\pi(y)\) vanishes at the boundaries (or \(almost\) vanishes, for small brane kinetic terms), then its derivative \(X'_\pi(y)\) will have a node in the bulk, and therefore will change sign. That means that there can be some parameter choice for which it is possible for the overlap integral to vanish, since the fermion zero mode profiles have no nodes in the bulk. This feature is clearly seen in Fig. [\[fig:yPiplot\]](#fig:yPiplot){reference-type="ref" reference="fig:yPiplot"}, where we plot the absolute value of the Yukawa couplings between zero-mode fermions and both the Higgs and the CP-odd scalar \(\Pi_0(x)\).[^6] The couplings shown are relative to the 5D bulk Higgs Yukawa coupling \(Y_5\) and are plotted as functions of the fermion bulk mass parameter \(c_q\) and \(c_u\) (for the case where we take \(c_q=-c_u\), for simplicity), for different overall KK scales. We observe that the CP-odd Yukawa couplings are fairly similar to the Higgs Yukawa couplings (i.e. exponentially sensitive to UV localization and then top-like when the zero mode is IR localized) except that there is a range of parameters where the coupling vanishes. Interestingly enough, this suppression happens for preferred values of the top quark bulk mass parameters. This means that the existence of suppressed couplings to top quarks of the CP-odd \(\Pi_0\) is a natural possibility in this scenario. ## Radiative couplings to photons and gluons {#subsec:gggamgam} Just like in the Higgs boson case, the radiative couplings of \(\Pi_0(x)\) to gluons and photons will depend on the physical Yukawa couplings \(y_{nn}\) between \(\Pi_0\) and the fermions (zero modes and KK modes) running in the loop, as well as on the fermion masses \(m_n\) (the eigenvalues of the mass matrix in Eq.([\[mumatrix\]](#mumatrix){reference-type="ref" reference="mumatrix"})). The real and imaginary parts of the couplings are associated with different loop functions, \(A^S_{1/2}\) and \(A^P_{1/2}\), as they generate the two operators \(\Pi_0G_{\mu\nu}G^{\mu\nu}\) and \(\Pi_0 G_{\mu\nu}\tilde{G}^{\mu\nu}\).[^7] The production cross section through gluon fusion is \_gg_0 = \_s\^2 m\_\_0\^2  and the decay widths to gluons and photons are \_\_0gg =  \_\_0 =  where \(\alpha_s\) and \(\alpha\) are the strong and weak coupling constants, \(N_c\) is the number of colors and \(Q_n\) is the charge of the fermion, and where c\^S_n =() A\^S\_1/2(\_n) and c\^P_n = Im() A\^P\_1/2(\_n) with \(\ \tau_n = m^2_{\Pi_0}/4m^2_n\ \) and with the loop functions defined as [\[eq:loopfunction\]]{#eq:loopfunction label="eq:loopfunction"} A\^S\_1/2() &=& 3\^-2,\ A\^P\_1/2() &=&-3 f()/, and with f() = { ------------------------------------------------------------------------------------------------------------------- --------------- \(\left[\arcsin\sqrt{\tau}\right]^2\) \((\tau\le 1)\) \(-{1\over4}\left[\ln\left({1+ \sqrt{1-\tau^{-1}}\over 1- \sqrt{1-\tau^{-1}}}\right)-i\pi\right]^2\) \((\tau> 1)\). ------------------------------------------------------------------------------------------------------------------- --------------- . For heavy KK quarks with masses \(m_n\) much greater than the CP-odd mass \(m_{\Pi_0}\) (i.e. when \(\tau\) is very small) the loop functions are essentially constant, as they behave asymptotically as \(\displaystyle \lim_{\tau \to 0} A^S_{1/2} =1 \hspace{.2cm} \mbox{and} \hspace{.2cm} \lim_{\tau \to 0} A^P_{1/2} =3/2.\) On the other hand, for light quarks (all the SM quarks except top and bottom), the loop functions essentially vanish asymptotically as \(\displaystyle \lim_{\tau \to \infty} A^S_{1/2} = \lim_{\tau \to \infty} A^P_{1/2} = 0.\) Moreover, we investigate a parameter region where the couplings of \(\Pi_0\) to top quarks are highly suppressed. This means that the production mechanism must rely exclusively on the heavy KK fermions running in the loop and as we have seen, this coupling depends on the ratio \(\frac{y_{nn}}{m_n}\) between the physical Yukawa coupling and the mass of the fermion running in the loop. To have an idea of the relative contribution of each of these KK fermions in the loop, in Fig. [\[fig:yPiKKplot\]](#fig:yPiKKplot){reference-type="ref" reference="fig:yPiKKplot"} we plot the mass normalized Yukawa couplings of Standard Model Higgs with top quarks, Higgs with first KK fermion and of \(\Pi_0\) to first KK fermion, for different values of the KK scale. As expected we see that the \(c_q\) dependence is mild (i.e. all KK fermions of any flavor will couple with similar strength) and also, as expected, we observe that the mass normalized couplings are quite suppressed with respect to the SM top quark case. Still the multiplicity of KK fermions is high, since there are 6 families of quarks and 3 families of charged leptons (the latter run in the diphoton loop), and for each family there are a few KK levels that give important contributions to the rate. A numerical scan of the couplings, including all families and 3 full KK levels is computationally too intensive, so in order to produce the couplings plotted in Fig. [\[fig:yPiKKplot\]](#fig:yPiKKplot){reference-type="ref" reference="fig:yPiKKplot"} we performed an approximation, sufficient for the purposes of the graph. The KK fermion Yukawa couplings plotted neglect mixings between different KK levels and different flavors, and with the zero mode fermions. They are obtained as follows. Consider the \(2\times2\) KK mass matrix ( ) M_u ( ), where \(Q_L(x)\) and \(Q_R(x)\) represent here a single flavor and a single KK level of the vector-like KK up-type doublets, and similarly for \(U_L(x)\) and \(U_R(x)\), vector-like KK up-type singlets. The mass matrix is thus M_u= ( ), [\[QUmatrix\]]{#QUmatrix label="QUmatrix"} where the diagonal entries are the KK masses (large) whereas the off-diagonal entries are coming from Yukawa couplings and are therefore smaller. In order to give a simple estimate, we take for simplicity the fermion bulk mass parameters as \(c_q=-c_u\), and the bulk Higgs Yukawa \(Y^{5D}\) to be real, which leads to \(Y_1=Y_2\) and \(m_Q=m_U=m_{KK}\), with the masses and profiles obtained by solving Eq. ([\[fermionprofile\]](#fermionprofile){reference-type="ref" reference="fermionprofile"}). With the KK fermion profiles one obtains the off-diagonal entries Y_1= \_0\^y_1 dy e\^-4(y) Q\_L(y)U\_R(y). The matrix that diagonalizes ([\[QUmatrix\]](#QUmatrix){reference-type="ref" reference="QUmatrix"}) in this simple limit (\(c_q=-c_u\)) is \(\left(\begin{matrix}-\frac{1}{\sqrt{2}} & \frac{1}{\sqrt{2}} \\ \frac{1}{\sqrt{2} }& \frac{1}{\sqrt{2}} \end{matrix}\right)\) and the eigenvalues are \(m_1=m_{KK}-Y_1\) and \(m_2=m_{KK}+Y_1\). Now we apply this rotation to the CP-odd Yukawa coupling matrix Y\_= ( ), where Y\_& =& i \_0\^y_1 dy e\^-3(y) Q\_L(y)U\_R(y) , and where for simplicity we have neglected gauge couplings terms compared to IR Yukawa terms (safe assumption when \(Y^{5D}\) is large). After diagonalization, we obtain the two physical couplings between \(\Pi_0\) and the KK fermions. When we normalize the couplings by the two eigenmasses and add the two contributions,[^8] we obtain \_i=1\^2 =+ =-2 . The last expression corresponds to the *mass normalized* Yukawa couplings of \(\Pi_0\) plotted in Fig. [\[fig:yPiKKplot\]](#fig:yPiKKplot){reference-type="ref" reference="fig:yPiKKplot"}, and this describes very closely the behavior of the couplings obtained in the full flavor calculation. The parametric dependence of these couplings is \(Y_{5D}^2 v/m_{KK}^2\), so that if \(Y^{5D}\sim 3\) we expect mass normalized couplings of order \((10^{-3}-10^{-4})\) GeV\(^{-1}\), if the overlap integral is of \({\cal O}(1)\). Since all the profiles of the integral are IR localized, one expects that integral to be \({\cal O}(1)\), although the precise numerical result varies between \(0.5\) and \(0.05\), depending on the values of the \(c_q\) parameter, as shown in the plots. All in all it seems likely that after taking into consideration all the fermion flavors, and for a KK scale of order \(1-2\) TeV, the overall KK fermion contribution to the radiative couplings of \(\Pi_0(x)\) to photons and gluons can be close to the top quark contribution to the gluon and photon couplings of the Higgs in the SM model. ## \(\Pi_0Zh\) coupling {#subsec:zh} The coupling between the CP-odd scalar, the \(Z\) boson and the Higgs will be extracted from the kinetic operator of the 5D Higgs, d\^4x dy e\^-2 D_H\^D\^H (1+ (y-y_i) ). Expanding the SM-like Higgs mode using Eq. ([\[Hexpansion\]](#Hexpansion){reference-type="ref" reference="Hexpansion"}) as well as the SM-like \(Z_\mu\) and the \(750\) GeV \(\Pi_0\) using Eqs. ([\[Aexpansion\]](#Aexpansion){reference-type="ref" reference="Aexpansion"}) and ([\[Piexpansion\]](#Piexpansion){reference-type="ref" reference="Piexpansion"}), we can obtain the coefficient \(g_{\Pi hZ}\) of the operator \(Z^\mu(x)\Big(h(x)\partial_\mu\Pi_0(x)+\Pi_0(x)\partial_\mu h(x)\Big)\) g\_hZ = dy e\^-2 v(y)h(y) f_z(y) (1+ (y-y_i) ) Now since \(M_z(y)=\frac{g_5}{2c_W}e^{-\sigma} v(y)\), \(h(y)\sim v(y)/v_4\), and \(f_z\simeq 1/\sqrt{y_1}\) we can write g\_hZ (X(y_1)+ X'(y_1))=0, where we have used the boundary conditions for the profile \(X(y)\) (see Eq.([\[Xbc\]](#Xbc){reference-type="ref" reference="Xbc"})) and assumed no UV brane kinetic term (\(d_0=0\)). The coupling should thus vanish in the limit of flat \(Z\) boson profile \(f_z(y)\), and when the nontrivial Higgs VEV \(v(y)\) is proportional to the Higgs scalar profile \(h(y)\). Corrections to these limits scale as \(v_4^2/m_{KK}^2\) and \(m_h^2/m_{KK}^2\) in the RS case, and so we expect the overall coupling to be highly suppressed. The partial width for the decay \(\Pi_0 \to h Z\) is (\_0 h Z)= (m\^2_h,m\^2\_;m\^2_Z) where \(m_Z\), \(m_\pi\) and \(m_h\) are the masses of the particles involved and where \(\lambda(x,y;z) = (1-x/z-y/z)^2-4 x y/z^2\). With the masses \(m_Z=91\) GeV, \(m_\pi=750\) GeV and \(m_h=125\) GeV, the width becomes \(\Gamma(\Pi_0 \to h Z) \sim (900\ g^2_{\Pi hZ})\) GeV. For example we compute numerically \(g_{\Pi hZ}\) for three different values of \(M_{KKglu_1}\), and with \(m_{\Pi_0}=750\) GeV and find Note that the couplings and widths are small but we observe that the partial width becomes larger as the KK mass scale is increased. Since we need the partial width \(\Gamma(\Pi_0 \to \gamma \gamma)\) to be similar to that of a \(750\) GeV Higgs, i.e. \(\Gamma(h_{750GeV} \to \gamma \gamma) \sim 10^{-5}\) GeV, we expect that at \(M_{KKglu_1} =1400\) GeV the \(Zh\) signal should start putting too much pressure on the allowed parameter space. This is confirmed in the full numerical analysis presented in the next section. ## Estimates and numerical results {#subsec:numerical} With all the previous ingredients one can estimate the viability of this scenario in terms of the possible diphoton excess. Let's choose the KK scale such that \(M_{KKglu_1}=1400\) GeV; when the bulk mass parameters of the top quark are around \(|c_{u_3}|\sim 0.35\) we know that the top quark will have highly suppressed couplings to \(\Pi_0\), as shown in the third panel of Fig. [\[fig:yPiplot\]](#fig:yPiplot){reference-type="ref" reference="fig:yPiplot"}. At the same time, the couplings of the KK tops (as well as all other KK quarks) will have relatively strong Yukawa couplings to \(\Pi_0\) (third panel of Fig. [\[fig:yPiKKplot\]](#fig:yPiKKplot){reference-type="ref" reference="fig:yPiKKplot"}), so that the contribution of each of them to the radiative coupling of \(\Pi_0\) to gluons is about an order of magnitude smaller than the top contribution to the \(h-glu-glu\) coupling of the SM. Thus we could estimate that the overall contribution of all flavors and KK excitations can make up for the suppressed coupling, so that the production cross section of \(\Pi_0\) is similar to that of a \(750\) GeV SM-like Higgs. The production cross section of a \(750\) GeV Higgs through gluon fusion, at the LHC running at 13 TeV is \(497\) fb , so, roughly, this could be assumed for the \(\Pi_0\) production cross section. Since the \(\Pi_0\) decays to top quarks and \(hZ\) are suppressed in this parameters space point, and its decays to \(WW\) and \(ZZ\) can only be radiative via the CP-odd gauge boson kinetic operator, the main decay channel is into gluons so that the branching of the diphoton channel should be very roughly Br(\_0 ) \~ where \(N_\gamma\) and \(N_{glu}\) are the multiplicities of states running in the \((\Pi_0 \gamma \gamma)\) loop and in the \((\Pi_0\ glu\ glu)\) loop respectively. In the diphoton loop there are \(3\) extra families of charged lepton KK excitations making the multiplicity of states greater. If their multiplicity and their Yukawa couplings can partially make up for the color factor of 8, then the diphoton cross section might become of \({\cal O}\) (1 fb), as hinted by the December 2015 LHC data. To complete the analysis, we perform a full numerical computation of production and branching ratios in a setup where we consider an effective 4D scenario including three full KK levels for all fields, i.e., we consider \(21 \times 21\) fermion mass matrices, which we diagonalize in order to obtain the physical Yukawa couplings. We choose a set of \(c\)-parameters and 5D Yukawa entries such that the SM masses and mixings are reproduced; the specific flavor choice for these parameters should not affect much the overall results since these depend on overlap integrals between IR localized fields, with very loose \(c\)-dependence. We choose the background metric parameters so that precision electroweak bounds are kept at bay, i.e. \(\nu=0.5\) and \(y_s=1.04 y_1\). Two average 5D Yukawa scales are considered, \(Y_{5D} \simeq 3\) and \(Y_{5D}\simeq 2\), to show the dependence on this parameter, and we also consider two different KK mass scales, \(M_{KKglu_1}=1000\) GeV and \(M_{KKglu_1}=1300\) GeV, which turn out to lead to successful signal generation. In order to see how tuned is the choice of top \(c\)-parameter, we plot the production cross section of the CP odd resonance, followed by decays into \(\gamma \gamma\), \(t\bar{t}\) and \(Zh\), as functions of \(c_{u_3}\) (the bulk mass parameter of the 5D singlet top quark), with the doublet bulk mass parameter fixed at \(c_{q_3}=0.4\). (This value ensures typically suppressed bounds from \(Zb_L{\bar b}_L\) bounds ). The results are shown in Figs. [\[fig:prodplot1\]](#fig:prodplot1){reference-type="ref" reference="fig:prodplot1"} and [\[fig:prodplot2\]](#fig:prodplot2){reference-type="ref" reference="fig:prodplot2"}, where we have chosen two KK scales, \(M_{KKglu_1}=1000\) GeV and \(M_{KKglu_1}=1300\) GeV. In both cases we show results for \(Y_{5D}\simeq 3\) and \(Y_{5D}\simeq 2\) to illustrate the sensitivity on this bulk parameter, crucial for enhancing the radiative couplings of \(\Pi_0\). When the KK scale is smaller and 5D Yukawa couplings are larger, the production of top pairs and gluon pairs can be too large. By reducing the 5D Yukawa couplings, enough signal can be generated with the dijets and top pairs under control, as well as the \(Zh\) decay. The values of \(|c_{u_3}|\) must be in the region \(0.2\mathrel{\raise.3ex\hbox{\)<\(\kern-.75em\lower1ex\hbox{\)\sim\(}}} |c_{u_3}| \mathrel{\raise.3ex\hbox{\)<\(\kern-.75em\lower1ex\hbox{\)\sim\(}}} 0.5\). For slightly smaller KK scales one expects a similar behavior, but such that 5D Yukawa couplings should be smaller in order to suppress overproduction of \(\Pi_0\) particles. This leads to the question of how large can the KK scale be and still manage to produce enough signal. We observe that at \(M_{KKglu_1}=1300\) GeV, larger 5D Yukawa couplings (\(Y_{5d} \geq 3\)) are required in order to have enough signal production. The branching fraction into \(hZ\) starts to be problematic, and actually becomes worse for heavier KK scales. The signal production requires that the value of \(|c_{u_3}|\) be located around the point where the top Yukawa couplings of \(\Pi_0\) are suppressed, in this case around \(|c_{u_3}|\sim 0.3\). We conclude therefore that in order to explain the \(750\) GeV diphoton, the KK scale must be \(M_{KKglu_1} \leq 1300\) GeV, with 5D Yukawa couplings \(Y_{5D}\mathrel{\raise.3ex\hbox{\)<\(\kern-.75em\lower1ex\hbox{\)\sim\(}}} 3\). For those values, the signal is quite generic (i.e. small, but typical, \(c_{u_3}\) values are required), as long as the \(\Pi_0\) mass is set to \(750\) GeV with appropriate boundary kinetic terms. # Discussion {#sec:discussion} We performed an analysis of the scalar sector of warped space models to investigate whether the *minimal* model can accommodate a resonance at 750 GeV with the properties observed in the diphoton resonance at the CMS and ATLAS. We show that in the simplest extra dimensional extension of the SM, that is with a 5D Higgs doublet living in the bulk, the lowest pseudoscalar KK excitation can be responsible for the signal observed at 750 GeV. We emphasize that, unlike other explanations relying on scalar fields in warped models, ours does not introduce any new fields or representations, but relies on Higgs brane kinetic terms to lower the KK mass of the lightest CP odd Higgs resonance. This makes the model extremely constrained, with the only new parameter being the IR brane kinetic coefficient \(d_{1}\). The lightest CP odd excitation, a mixture of the 5D Higgs field and \(Z_5\), does not decay at tree level into \(WW\) or \(ZZ\), and, over a range of the parameter space, can have suppressed couplings to the top quark, thus a small decay width into \(t {\bar t}\). The production through gluon fusion can be loop-enhanced through the effects of the usual KK fermion modes, and so can the diphoton decay. The coupling to \(Zh\) is also suppressed although starts increasing dangerously for KK masses above \(1500\) GeV. We also showed that in \(AdS_5\) spaces (RS-type models) with a KK scale as low as \(1500\) GeV, the lowest CP-odd scalar cannot have a mass at 750 GeV, but if the metric differs slightly from the \(AdS_5\) metric (generalized warped metric spaces), the Higgs brane kinetic terms can produce a viable CP-odd scalar of mass 750 GeV. Within these modified metric scenarios, and for KK mass scales at around 1 TeV (consistent with precision electroweak bounds) this CP odd resonance obeys the (current) experimental constraints. We analyzed its production and decay for several values of the lowest KK gluon mass (\(M_{KKglu_1}\)), and show that consistency with the data requires \(M_{KKglu_1} \mathrel{\raise.3ex\hbox{\)<\(\kern-.75em\lower1ex\hbox{\)\sim\(}}} 1300\) GeV. Our analysis is quite general, and should be valid even in the absence of a signal at 750 GeV. The general conclusion to be taken from our analysis here is that warped space models, without any new particles, can explain a (relatively) light diphoton resonance at the LHC. Should a diphoton excess be found at higher values, even the RS model might accommodate such a state without the need modify the metric. Among the general features of the light CP-odd scalar resonance, resulting from the 5D Higgs doublet are the fact that its coupling to top pairs can be suppressed for appropriate top bulk mass parameters. Also its coupling to \(Zh\) is generically suppressed due to the boundary conditions of the CP odd state. In addition, the model predicts that the spectrum for the CP odd and the charged scalars is essentially the same since their differential equations and boundary conditions are almost identical. This means that the lightest charged Higgs boson is expected to have a mass very close to the pseudoscalar mass, so about 750 GeV, in the scenario in which the latter is the diphoton resonance. If the diphoton resonance ends up being a real particle, and not a statistical fluctuation, the prediction from this scenario is that \(ZZ\) and \(WW\) should not be seen, whereas top pair production, dijet production and \(Zh\) signal should be around the corner. Moreover a search for the charged scalars at around \(750\) GeV might prove useful in order to disentangle the different models explaining the resonance. # Appendix In this section we consider the effect of brane localized kinetic terms associated with the 5D Higgs doublet and also with the gauge bosons. For simplicity, let's consider a 5D toy model with a Higgs scalar \(H(x,y)\) charged under a local \(U(1)\), defined by the following action: S&=&d\^4xdy  (- F\^2\_MN + \|D\^M H\|\^2-V(H) )\ &&   + \_i d\^4xdy  (y-y_i) ( r_i F\^2\_MN + d_i \|D\^M H\|\^2-\_i(H))) [\[5Daction\]]{#5Daction label="5Daction"} where \(F_{MN} = \partial _M A_N-\partial_N A_M\) and for simplicity we set gauge coupling constant to unity in the appropriate mass dimensions. The background spacetime metric is assumed to take the form ds\^2=e\^-2(y)\_ dx\^dx\^-dy\^2 where \(\sigma(y)\) is the warp factor. We are interested in studying the effective 4D perturbative spectrum of the 5D Higgs field and the 5D gauge boson, around a nontrivial Higgs vacuum profile solution \(<H>=v(y)\). H(x,y)= (v(y) + h(x,y)) e\^i (x,y) In particular we are interested in the CP-odd Higgs perturbations \(\pi(x,y)\), whose equations of motion are coupled with the gauge boson perturbations. The equations read && (1+r_i \_i) \_\^A\_-( (1+r_i \_i)e\^-2 A'\_)' + (1+d_i \_i) M_A\^2 A\_+\ &&\_((1+d_i \_i) M_A\^2 -(1+r_i \_i)\^A\_-((1+r_i \_i)e\^-2 A_5)') = 0\ &&(1+r_i \_i) \_\^A_5-(1+r_i \_i) \^A'\_+ (1+d_i \_i)M_A\^2 ('-A_5)=0\ &&(1+d_i \_i) \_\^-(1+d_i \_i) \^A\_+ M_A\^-2((1+d_i \_i) M_A\^2 e\^-2 ('-A_5))'= 0 where \(M_A= v(y) e^{-\sigma}\) and where \(\displaystyle d_i \delta_i \equiv \sum_i d_i \delta(y-y_i)\) and \(\displaystyle r_i \delta_i \equiv \sum_i r_i \delta(y-y_i)\). We fix partially the 5D gauge by imposing (1+d_i \_i) M_A\^2 -(1+r_i \_i)\^A\_-((1+r_i \_i)e\^-2 A_5)' = 0. [\[gaugefix\]]{#gaugefix label="gaugefix"} The previous gauge fixing equation reads in the bulk: M_A\^2 -\^A\_-(e\^-2 A_5)' = 0. [\[bulkgaugefix\]]{#bulkgaugefix label="bulkgaugefix"} Note that if we evaluate the bulk constrain Eq. ([\[bulkgaugefix\]](#bulkgaugefix){reference-type="ref" reference="bulkgaugefix"}) at \(y=y_1-\epsilon\) (i.e. right before the IR brane), we obtain: \^A\_\|\_y_1- = M_A\^2 -(e\^-2 A_5)'\|\_y_1-. [\[epsiloncondition\]]{#epsiloncondition label="epsiloncondition"} On the other hand, the effect of the delta functions in Eq.([\[gaugefix\]](#gaugefix){reference-type="ref" reference="gaugefix"}) is to produce a discontinuity in the 5D field \(A_5\) at the brane location as d_1 M_A\^2 -r_1 \^A\_\|\_y_1- = \_y_1-\^y_1, and similarly for the UV brane. We can thus multiply ([\[epsiloncondition\]](#epsiloncondition){reference-type="ref" reference="epsiloncondition"}) by \(r_1\) and use it in the previous equation, and find the necessary boundary condition between \(\pi\) and \(A_5\), which ensures that \(A_\mu\) is completely decoupled, even on the brane. We find (d_1-r_1) M\^2_A + r_1 (e\^2 A_5)' . \|\_y_1- =-.e\^-2 A_5\|\_y_1- [\[boundaryconditions\]]{#boundaryconditions label="boundaryconditions"} where we have taken \(A_5\) to vanish exactly on the brane, but it jumps right before the boundary. Inserting the gauge choice in the coupled equations of motion, one manages to decouple the gauge modes \(A_\mu\) (in both the bulk and the branes) with a bulk equation \_\^A\_-(e\^-2 A'\_)' + M\_A\^2 A\_ and jump condition on \(A'_\mu\) r_1 \_\^A\_+ d_1 M\_A\^2 A\_\|\_y_1-=-. e\^-2 A'\_\|\_y_1- where \(A'_\mu\) again vanishes exactly on the brane, but has a jump right before it. We separate variables A\_(x,y)= V\^4d\_(x) V_y(y) and find the separated equations for the gauge boson tower become && \_\^V\^4d\_(x)+m\^2_A V\^4d\_(x)=0\ && ( e\^-2 V'\_y)' + ( m\^2_A-M_A\^2 ) V_y =0 with jump conditions on \(V'_y\) (d_i M\_A\^2-r_i m\^2_A) V_y \|\_y_1-=-. e\^-2 V'\_y\|\_y_1- where the 4D effective mass \(m^2_A\) is the constant of separation of variables. The remaining equations are, in the bulk, && \_\^A_5+ M_A\^2 ('-A_5)-(M_A\^2 )' +( (e\^-2 A_5)' )' =0\ && \_\^+ M_A\^-2 ( M_A\^2 e\^-2 ('-A_5) )'-M_A\^2 + (e\^-2 A_5)'= 0[\[a5pimixedequation\]]{#a5pimixedequation label="a5pimixedequation"} and the fields must verify the boundary conditions of Eq.([\[boundaryconditions\]](#boundaryconditions){reference-type="ref" reference="boundaryconditions"}). We now perform a mixed separation of variables: A_5(x,y)&=& G(x) g(y) + \_x(x) (y) [\[expansion1\]]{#expansion1 label="expansion1"}\ (x,y)&=& G(x) h(y) + \_x(x) which is to say that both \(A_5(x,y)\) and \(\pi(x,y)\) contain each some Goldstone and CP-odd degree of freedom. The profiles \(g(y)\), \(\eta(y)\), \(h(y)\) and \(quantify how much of each they contain. Of course the functions\)g\(and\)h\(are related to each other, as well as\[and\) With the choice h(y)&=&\ g(y)&=&\ (y)&=&X(y)\ and using the mixed separation of variables in ([\[expansion1\]](#expansion1){reference-type="ref" reference="expansion1"}) and ([\[expansion2\]](#expansion2){reference-type="ref" reference="expansion2"}), the mixed equations of motion in ([\[a5pimixedequation\]](#a5pimixedequation){reference-type="ref" reference="a5pimixedequation"}) decouple and we obtain && X(y) \_\^\_x(x) +\_x(x) e\^2 X(y)-\_x(x) '= 0\ &&K(y) \_\^G(x)+ G(x) = 0. Once separated, we obtain, for the CP odd physical scalars &&\_\^\_x(x) + m\_\^2 \_x(x) =0\ &&(M_A\^-2 X')'+ e\^2 (-1) X=0, with boundary conditions d_i X' =-X and for the Goldstone modes, &&\_\^G(x)+ m_G\^2 G(x)=0\ &&(K'e\^-2)'+ (m_G\^2-M_A\^2) K=0, with boundary conditions (d_i M_A\^2-r_i m\^2_G ) K =-e\^-2 K'. Note that both the equations and boundary conditions for the Goldstone bosons are identical to the ones for the gauge boson tower, as should be, so that they can then be gauged away level by level with the remaining gauge fixing freedom. # Introduction The original motivation for warped extra-dimensions was to address the hierarchy problem, so that the fundamental scale of gravity is exponentially reduced along the extra dimension, from the Planck mass scale to the TeV scale. Thus, the TeV scale becomes the natural scale of the Higgs sector if this one is localized near the TeV boundary of the extra dimension, as first introduced by Randall and Sundrum (RS). If SM fields are allowed to propagate in the extra dimension, the scenario can also address the flavor puzzle of the SM, explaining fermion masses and mixings from the geographical location of fields along the extra dimension. However, processes mediated by the heavy resonances of the 5D bulk fields, Kaluza-Klein (KK) modes, generate dangerous contributions to electroweak and flavor observables (including dangerous deviations to the \(Zb\bar{b}\) coupling), pushing the KK mass scale to \(5-10\) TeV. A popular mechanism to lower the KK scale involves using a custodial gauge \(SU(2)_R\) symmetry, which ensures a small contribution to electroweak precision parameters, lowering the KK scale bound to about 3 TeV. Alternatively, one can study scenarios in which the metric is slightly modified from the RS metric background (\(AdS_5\)). This can be achieved quite naturally from the backreaction on the metric caused by a 5D scalar field stabilizing the original \(AdS_5\) warped background. When the 5D Higgs is sufficiently leaking into the bulk and when the metric background is modified near the TeV boundary, the scenario allows for KK scales as low as 1-2 TeV, with precision electroweak and flavor constraints under control. An inconvenience is that these scenarios are typically hard to probe experimentally as the couplings of all particles are very suppressed. Still, it has been shown that it can still lead to interesting deviations in Higgs phenomenology, as the Higgs couplings can receive sufficient radiative corrections from the many KK fermions of the model. Run 2 LHC data at \(\sqrt{s}=13\) TeV shows signals of a new resonance in the diphoton distribution at an invariant mass of 750 GeV with a 3.9 significance at ATLAS, with 3.2 fb\(^{-1}\) and 3.4 combined significance at CMS, with 2.6 fb\(^{-1}\)(combining run 1 and run 2 results). ATLAS reports 14 events and CMS, 10. The experimental data is summarized in Table [2](#tab:siglims){reference-type="ref" reference="tab:siglims"}. ) and the covariant derivative is \(D_M= \partial_M + i g_5 A_M\) with [\[Aexpansion\]]{#Aexpansion label="Aexpansion"} A_M=( ) and CP-odd and charged Higgs part is [\[Piexpansion\]]{#Piexpansion label="Piexpansion"} =( ) with the weak angle defined like in the SM, i.e. \(s_W/c_W=g'_5/g_5\), where \(g_5\) and \(g_5'\) are the 5D coupling constants of \(SU(2)_L\) and \(U(1)_Y\). The extraction of degrees of freedom in this context has been performed in and we outline here the main results. The effect of brane kinetic terms in the Higgs sector is new and its derivation is outlined in the Appendix. The 5D equations of motion for all these fields are coupled (except for the case of the real Higgs excitation \(h(x,y)\)) and in order to decouple them, one can partially fix the gauge, or add a gauge fixing term to the previous 5D action. For example, in the CP-odd case, the fields \(Z_\mu(x,y)\), \(Z_5(x,y)\) and \(\Pi_z(x,y)\) must be unmixed. The partial gauge fixing constraint[^10] \^Z\_-M\^2_z(y) \_z +(e\^-2Z_5)'=0 manages to decouple the fields \(Z_\mu\) from \(Z_5\) and \(\Pi_z\) in the bulk. We defined here \(M_z(y)=\frac{g_5}{2c_W} v(y) e^{-\sigma(y)}\). However, the presence of the Higgs brane kinetic terms, proportional to \(d_i\) in the action, forces us to extend the gauge choice on the branes, producing a lifting of \(Z_5\) field so that the decoupling is maintained at the boundaries[^11]. The appropriate boundary condition at the IR brane is Z_5(x,y_1) =- M\^2_z(y_1) e\^2(y_1) \_z(x,y_1) where \(y_1\) denotes the position of the boundary (note that if the brane kinetic term parameter \(d_1\) tends to zero, the condition on \(Z_5\) becomes Dirichlet, as expected). With this type of gauge choice, the 5D fields \(Z_\mu\), \(W_\mu\) and \(A_\mu\) have independent 5D equations of motion. In order to extract the effective 4D degrees of freedom, we expand the gauge fields as \(Z_\mu(x,y)=Z^n_\mu(x) f^n_z(y)\), \(W_\mu(x,y)=W^n_\mu(x) f^n_w(y)\) and \(A_\mu(x,y)=A^n_\mu(x) f^n_\gamma(y)\) (summation over \(n\) is understood) and where \(Z^0_\mu(x)\), \(W^0_\mu(x)\) and \(A^0_\mu(x)\) are the \(Z\), \(W\) and \(\gamma\) gauge bosons of the SM. The extradimensional profiles \(f^n_z(y)\), \(f^n_w(y)\) and \(f^n_\gamma(y)\) are solutions of [\[gaugeeq\]]{#gaugeeq label="gaugeeq"} (e\^-2 f'\_a)'+(m\^2\_n-M_a\^2(y)) f_a =0 where \(a=z,w,\gamma\) and where \(M_z(y)=\frac{g_5}{2c_W} v(y) e^{-\sigma(y)}\), as defined before, \(M_w(y)=\frac{g_5}{2} v(y) e^{-\sigma(y)}\), and \(M_\gamma=0\). The boundary conditions for these profiles are[^12] [\[gaugebc\]]{#gaugebc label="gaugebc"} M_a\^2 f_a(y_i) =-e\^-2f'\_a(y_i). The CP-even Higgs field is expanded as \(h(x,y)=h^n(x) h^n_y(y)\) and the equations for the Higgs profiles are, with \(h_y \equiv h_y^n(y)\): (e\^-4 h'\_y)'+(m\^2\_h_n-\^2\_bulk) h_y =0, where \(\displaystyle \mu^2_{bulk} = \frac{\partial^2 V }{\partial H^2}\Big|_{H=v}\). The boundary conditions are (\^2\_brane_i- m\^2\_h_n e\^2) h_y =-e\^-4 h'\_y, with \(\displaystyle \mu^2_{brane_i} = \frac{\partial^2 \lambda_i }{\partial H^2}\Big|_{H=v}\). Note that the mode \(h^0(x)\) is interpreted as the SM Higgs boson. There are still some degrees of freedom left, and their 5D equations of motion still happen to be mixed. One of the coupled systems involves \(Z_5\) and \(\Pi_z\) and the other coupled system involves \(\Pi^{\pm}\) and \(W_5^\pm\). In order to disentangle these systems one must perform a mixed expansion, so that the decoupling of fields will happen KK level by KK level. The mixed expansions are, in the CP-odd sector, [\[Z5decomp\]]{#Z5decomp label="Z5decomp"} Z_5(x,y)&=& G\^n(x) + \_n(x)X\_(y)\ \_z(x,y)&=& G\^n(x) + \_n(x) X\_'(y), [\[Pidecomp\]]{#Pidecomp label="Pidecomp"} and in the charged scalar sector they are W_5\^(x,y) &=& G_n\^(x) + \^\_n(x)X\_(y)  \ \^(x,y)&=& G_n\^(x) +\^\_n(x) X\_'(y).  where \(M_z(y)\) and \(M_w(y)\) were defined below Eq ([\[gaugeeq\]](#gaugeeq){reference-type="ref" reference="gaugeeq"}). The effective 4D physical fields are the tower of CP-odd neutral scalars \(\Pi_n(x)\) and the tower of charged scalars \(\Pi^\pm_n(x)\). Their associated extra-dimensional profiles \(X_\pi(y)\) and \(X_\pm(y)\) obey the equations ( X_a' )'+ (-1 ) e\^2 X_a=0 [\[Xeq\]]{#Xeq label="Xeq"} where \(M_a(y)=(M_z(y),M_{w}(y))\) and \(X_a=(X_\pi,X_{\pm})\). The boundary conditions are [\[Xbc\]]{#Xbc label="Xbc"}  X'\_a=-X_a, and note that vanishing Higgs brane kinetic terms implies Dirichlet boundary conditions for \(X_a\). We checked that these bulk equations agree with , the only new addition being the boundary conditions imposed by the presence of Higgs brane kinetic terms. In order for these 4D scalars to be canonically normalized, we require dy e\^2 =1 and this condition includes the effect of Higgs brane kinetic terms. The remaining 4D fields are \(G_n(x)\) and \(G_n^\pm(x)\), which are Goldstone bosons at each KK level. The profile wavefunctions \(f_{G_a}(y)\) obey the same differential equations as the gauge profiles, Eq.([\[gaugeeq\]](#gaugeeq){reference-type="ref" reference="gaugeeq"}), as well as the same boundary conditions, Eq.([\[gaugebc\]](#gaugebc){reference-type="ref" reference="gaugebc"}). The spectrum is thus identical to the gauge bosons spectrum level by level. These fields appear in the effective 4D action coupled to (\(\partial^\mu Z^n_\mu\)) or (\(\partial^\mu W^n_\mu\)), and of course there is a leftover gauge freedom allowing us to gauge them away (i.e., they are pure gauge). We wish to identify the lightest CP-odd scalar \(\Pi_0(x)\) with the observed diphoton peak at the LHC, so that we need to fix its mass \(m_{\Pi_0}= 750\) GeV. In order to have an idea of the effects of the Higgs brane kinetic terms on the CP-odd scalar spectrum, we consider two different parameter points, one in which the RS background metric is recovered with \(\nu=10\) and \(y_s=4\times y_{1}\) (\(\nu\) is the exponent appearing in the modified metric and if relatively large, the location of spurious singularity is sent away from the boundary, recovering essentially the \(AdS_5\) metric). The other case is the situation where the metric modification allows for TeV size KK masses, which are safe from precision electroweak constrains. The parameters chosen there are \(\nu=0.5\) and \(y_s=1.04\times y_{1}\). In both parameter points, we fix the KK mass of the first gluon excitation to be \(1500\) GeV[^13]. In Fig. [\[fig:Xmassplot\]](#fig:Xmassplot){reference-type="ref" reference="fig:Xmassplot"} we show the spectrum of the first 4 KK levels of CP-odd Higgs bosons \(\Pi_n(x)\), for \(n=0,1,2,3\), as a function of the brane kinetic term \(d_{1}\) in units of the curvature \(k\sim M_{Pl}\). The effects of the UV localized brane kinetic term are warped suppressed and so we do not consider them here anymore. We can see that in the RS limit, it is not possible to reduce sufficiently the lightest CP-odd mass to 750 GeV, as its mass tends asymptotically from about \(2500\) GeV without brane kinetic term to about \(1750\) GeV for large brane kinetic terms. On the other hand, with the modified metric it becomes possible to reduce greatly the lightest CP-odd mass with relatively small brane kinetic term coefficients (which in this particular case tends asymthotically to about 400 GeV). Parameter points such that the metric modification lies between the two considered will actually have an intermediate behavior, with a lightest CP-odd mass having increasing asymptotic values as one recovers the RS background. Finally also note that the spectrum for the charged scalars is essentially the same since their differential equations and boundary conditions are identical except for the functions \(M_z(y)\) and \(M_w(y)\), which differ by about \(10\%\). \(M_w(y)\) produces a deviation from the CP-odd scalar spectrum of less than \(5\%\). This means that the scenario under consideration should also contain a lightest charged Higgs scalar with a mass of about \(750\) GeV also. The next question to ask is how big is the effect of the Higgs brane kinetic term on the gauge bosons, and in particular on the lowest ones, i.e. the SM \(W\) and \(Z\) bosons. These terms represent an additional (brane localized) contribution to the mass of the gauge bosons. In principle their mass is generated here from a bulk Higgs mechanism, unless the brane kinetic terms are overly important (not the limit we are working with here). We can quickly estimate its effect on the lowest lying gauge fields. These are essentially flat (like all gauge zero modes) and thus their wave function is \(f^0_z\sim 1/\sqrt{y_{1}}\). The contribution of a brane localized mass squared term is \(\delta m^2_z \mathrel{\raise.3ex\hbox{\)<\(\kern-.75em\lower1ex\hbox{\)\sim\(}}} d_{1}\ v^2/y_{1}\sim d_{1}\times 700\) GeV\(^2\). For IR brane kinetic term coefficient \(d_{1}\) of \({\cal O}(1)\), this represents naively at most some \(10\%\) contribution to the overall mass squared of either \(W\) or \(Z\). In the particular case of the modified metric with a brane coefficient \(d_{1}\simeq 0.51\), and metric parameters \(\nu =0.5\) and \(y_s=1.04 \times y_1\) (which produces a light CP-odd scalar of \(750\) GeV), the exact numerical effect on the zero mode gauge boson masses (\(W\) and \(Z\)) is a shift of \(3\) GeV with respect to the no-brane-kinetic-term limit. Of course, in the presence of brane kinetic terms, one redefines the VEV normalization constant, and the value of \(g_5\), in order to correctly account for the SM gauge boson masses and electroweak couplings. # CP-odd Higgs couplings {#sec:cpodd} As a \(750\) GeV CP-odd Higgs scalar \(\Pi_0(x)\) is allowed in the spectrum, thanks to the effect of small brane localized Higgs kinetic terms, we now study its couplings to SM particles in order to see if the observed excess at the LHC can be associated with this excitation. Of course being a CP-odd scalar its tree-level couplings to \(ZZ\) and \(WW\) are zero, making it an ideal candidate for the observed exotic events. We thus need to focus on its tree-level couplings to fermions (and top quark in particular), to \(Zh\) (where \(h\) is the \(125\) GeV Higgs) and to its radiative couplings to photons and gluons. We study these in the subsequent subsections. ## Fermion couplings {#subsec:fermion} The couplings of \(\Pi_0(x)\) to fermions arise from two sources in the action. First source comes from the 5D Higgs Yukawa couplings, and second, from the gauge fermion couplings. This is because the physical field \(\Pi_0(x)\) contains some of CP-odd Higgs scalar, and some of \(Z_5\) excitation, where \(Z_5\) is the fifth component of the 5D vector boson \(Z_M\). However the 5D Yukawa coupling allows for direct coupling of \(\Pi_0(x)\) to two zero-mode fermions, whereas the gauge-fermion coupling allows only couplings between fermion zero-modes and higher KK fermion levels. As we will see, it is important to keep both couplings, since after electroweak symmetry breaking the physical SM fermions (top quarks in particular) are mostly zero-modes but also contain a small amount of higher KK excitations, and could thus inherit some of the original gauge-fermion coupling, especially if the tree-level Yukawa coupling between \(\Pi_0(x)\) and zero-mode top quarks is suppressed (as it can be). The relevant terms in the action are the 5D Higgs Yukawa couplings and the fermion gauge interaction term, S\_ff_0 d\^4xdy where \(Q,U,D\) represent the 5D fermion \(SU(2)_L\) doublets, up-type and down-type singlets (with generation indices and isospin indices suppressed). The kinetic terms contain the 5D covariant derivative and from them we extract the terms containing the CP-odd component \(Z_5(x,y)\), and from the Higgs Yukawa couplings we extract the terms containing the CP-odd Higgs component \(\Pi_z(x,y)\). We follow the approach of and compute these couplings by considering only the effects of three full KK levels, i.e. computing \(21 \times 21\) fermion Yukawa coupling matrices (with 3 \(up\) and 3 \(down\) families, each containing zero modes and 3 KK levels with an \(SU(2)_L\) doublet and a singlet in each level, i.e., 3 zero modes plus \(3\times 3 \times 2\) KK modes). Note that we are interested in the couplings of the \(750\) GeV CP-odd scalar \(\Pi_0(x)\) to SM fermions (top quarks primarily), but we also need its couplings with the rest of KK fermions, since these interactions will be crucial to generate large enough radiative couplings to photons and gluons. We first write the effective 4D up-type quark mass matrix as ( ) M_u ( ) in a basis where \(q^0_L(x)\) and \(u_R^0(x)\) represent three zero-mode flavors each (doublets and singlets of \(SU(2)_L\)), and \(Q_L(x)\) and \(Q_R(x)\) represent three flavors and three KK levels of the vector-like KK up-type doublets, and \(U_L(x)\) and \(U_R(x)\) represent three flavors and three KK levels of vector-like KK up-type singlets. The mass matrix is thus M_u= ( ) [\[mumatrix\]]{#mumatrix label="mumatrix"} with the down sector mass matrix \({M}_d\) computed in the same way. The submatrices are obtained by evaluating the overlap integrals [\[YYY1KK\]]{#YYY1KK label="YYY1KK"} y\^0_u= \_0\^y_1 dy e\^-4(y) q\^0,i\_L(y)u\^0,j\_R(y)\ Y\^qU= \_0\^y_1 dy e\^-4(y) q\^0,i\_L(y)U\^n,j\_R(y)\ Y\^Qu = \_0\^y_1 dy e\^-4(y) Q\^m,i\_L(y)u\^0,j\_R(y)\ Y_1 = \_0\^y_1 dy e\^-4(y) Q\^m,i\_L(y)U\^n,j\_R(y)\ Y_2 = \_0\^y_1 dy e\^-4(y) Q\^m,i\_R(y)U\^n,j\_L(y) , where the indices \(m\) and \(n\) track the KK level and \(i,j=1,2,3\) are 5D flavor indices. The diagonal matrices \((M_Q)_{9\times 9}\) and \((M_U)_{9\times 9}\) are constructed with the masses of all the KK quarks involved. The masses and the profiles of the KK fermions appearing in these overlap integrals (\(Q_L(y)\), \(Q_R(y)\), \(U_L(y)\) and \(U_R(y)\)) are obtained by solving differential equations for the fermion profiles \_y(e\^(2c-1)(y)\_y( e\^-(c+2)(y)))f(y)+e\^(c-1)(y) m_n\^2 f(y) =0 [\[fermionprofile\]]{#fermionprofile label="fermionprofile"} where \(f(y)\) is the KK profile. The mass eigenvalues \(m_n\) are found by imposing Dirichlet boundary conditions on the wrong chirality modes. As mentioned before, we have included 3 full KK levels so that the mass matrices in the gauge basis are \(21 \times 21\) dimensional matrices, which are not diagonal. One needs to diagonalize them, and by doing so, move to the quark physical basis where all the fermions couplings can then be extracted. In the CP-odd scalar sector, we can write the effective 4D Yukawa-type couplings to fermions in the same gauge basis as before ( ) **Y**\_( )  \_0(x) where now the \(21 \times 21\) coupling matrix \({\bf Y}_{\pi}\) is given by \_= ( ) [\[Ypimatrix\]]{#Ypimatrix label="Ypimatrix"} The submatrices are obtained by the overlap integrals [\[YYPI\]]{#YYPI label="YYPI"} y\^0\_qu &=& i \_0\^y_1 dy e\^-3(y) q\^0,i\_L(y)u\^0,j\_R(y)\ Y\_qU&=& i \_0\^y_1 dy e\^-3(y) q\^0,i\_L(y)U\^n,j\_R(y)\ Y\_Qu& =& i \_0\^y_1 dy e\^-3(y) Q\^m,i\_L(y)u\^0,j\_R(y)\ Y\^\_1& =& i \_0\^y_1 dy e\^-3(y) Q\^m,i\_L(y)U\^n,j\_R(y)\ Y\^\_2& =& i \_0\^y_1 dy e\^-3(y) Q\^m,i\_R(y)U\^n,j\_L(y)  , and a\_qQ&=& \_0\^y_1 dy e\^-2(y) q\^0,i\_L(y)Q\^n,j\_R(y)\ a\_uU&=& \_0\^y_1 dy e\^-2(y) u\^0,i\_R(y)U\^n,j\_L(y)\ a\_QQ&=& \_0\^y_1 dy e\^-2(y) Q\^m,i\_L(y)Q\^n,j\_R(y)\ a\_UU&=& \_0\^y_1 dy e\^-2(y) U\^m,i\_R(y)U\^n,j\_L(y) where the \(g^{5D}_{L,R}\) coupling are given by g\^5D\_L = g\^5D\_W (T_3-Q_q \^2 \_W ) g\^5D\_R = g\^5D\_W Q_q \^2 \_W, with \(Q_q\) the charge of the quark, (here \({2 \over 3}\)), \(\theta_W\) the weak angle and \(T_3 = {1\over 2}\). Note that when the interaction originates in the 5D Yukawa couplings, the profile to use is the one coming from the CP-odd Higgs component, i.e proportional to \(X_\pi'(y)\). When the interaction originates in the gauge fermion coupling and thus comes from the \(Z_5\) component, the profile to use is proportional to \(X_\pi(y)\), with \(X_\pi\) being the solution of Eq. ([\[Xeq\]](#Xeq){reference-type="ref" reference="Xeq"}), using the decompositions of Eqs. ([\[Z5decomp\]](#Z5decomp){reference-type="ref" reference="Z5decomp"}) and ([\[Pidecomp\]](#Pidecomp){reference-type="ref" reference="Pidecomp"}). When the fermion matrix in ([\[mumatrix\]](#mumatrix){reference-type="ref" reference="mumatrix"}) is diagonalized, the coupling matrix of fermions with the CP-odd field \(\Pi_0(x)\) in ([\[Ypimatrix\]](#Ypimatrix){reference-type="ref" reference="Ypimatrix"}) is rotated, and we can then extract all the physical Yukawa couplings. All these couplings are needed later in order to compute the radiative couplings of \(\Pi_0(x)\) with gluons and photons. We first analyze the very important Yukawa coupling between \(\Pi_0(x)\) and top quarks, as this coupling might dominate the decays of the CP-odd scalar. This coupling comes essentially from the entry \((y^0_{\pi qu})_{ 33}\) (before rotation to the physical basis) although it receives small corrections after going to the physical basis. We focus on \((y^0_{\pi qu})_{ 33}\) which comes from the overlap integral [\[YPtt\]]{#YPtt label="YPtt"} (y\^0\_qu)\_33 &=& i \_0\^y_1 dy e\^-3(y) q\^0\_t(y)u\^0\_t(y) In the RS limit, the warp factor is \(\sigma(y)=ky\), and the top profiles are \(q^{0}_{t}(y)= f(c_q) e^{(2-c_q)ky}\) and \(u^{0}_{t}(y)=f(-c_u) e^{(2+c_u)ky}\), where \(f(x)\) is a normalization factor. We also have \(M_z(y)=\frac{g_5}{2c_W} v_0 e^{(a-1)ky}\), with \(v_0\) a constant factor, so that the previous overlap integral in this limit reads (y\^0\_qu)\_33 = i \_0\^y_1 dy e\^(2-a-c_q+c_u)k y X\_'(y) We integrate this by parts to find (y\^0\_qu)\_33 &=& -i \_0\^y_1 dy (2-a-c_q+c_u) e\^(2-a-c_q+c_u)k y X\_(y) + BT      where \(BT= i \frac{(Y^{5D}_u)_{33}}{\sqrt{2k}} \frac{2 c_W}{g_5}\frac{f(c_q) f(-c_u)}{v_0 m^2_{\pi_0}}\ e^{(2-a-c_q+c_u)k y} X_\pi(y)\Big|_0^{y_1}\) is a boundary term. Note that the profile \(X_\pi(y)\) has vanishing boundary conditions in the absence of Higgs localized brane kinetic terms. In that limit we can see that the coupling of the CP-odd scalar can actually vanish, when \((2-a-c_q+c_u)=0\) . Note also that the Higgs localizer parameter \(a\) is, in this RS limit, \(a \mathrel{\raise.3ex\hbox{\)>\(\kern-.75em\lower1ex\hbox{\)\sim\(}}} 2\) and the bulk parameters \(c_q\) and \(c_u\) are defined such that, for example, charm or bottom quarks are assigned values more or less \(c_q \in (0.45,0.55)\) and \(c_u\in (-0.5,-0.6)\), whereas for top we have \(c_{q_3}\sim 0.45\) and \(c_{u_3}>-0.45\). This means that we expect the term \((2-a-c_q+c_u)\) to vanish, in the limit of \(a\sim 2\), when \(c_q-c_u \sim 0\), so that the suppression in this case seems only possible for the top quark, where both \(c_q\) and \(c_u\) could be small. Of course when the metric background is modified, and the boundary conditions include the brane kinetic terms there will be deviations from the previous values. Nevertheless it is clear that the Yukawa coupling of the CP-odd scalar field to top quarks can have highly suppressed values. Another way to see this is to consider the overlap integral in Eq. ([\[YPtt\]](#YPtt){reference-type="ref" reference="YPtt"}). Because the profile \(X_\pi(y)\) vanishes at the boundaries (or \(almost\) vanishes, for small brane kinetic terms), then its derivative \(X'_\pi(y)\) will have a node in the bulk, and therefore will change sign. That means that there can be some parameter choice for which it is possible for the overlap integral to vanish, since the fermion zero mode profiles have no nodes in the bulk. This feature is clearly seen in Fig. [\[fig:yPiplot\]](#fig:yPiplot){reference-type="ref" reference="fig:yPiplot"}, where we plot the absolute value of the Yukawa couplings between zero-mode fermions and both the Higgs and the CP-odd scalar \(\Pi_0(x)\).[^14] The couplings shown are relative to the 5D bulk Higgs Yukawa coupling \(Y_5\) and are plotted as functions of the fermion bulk mass parameter \(c_q\) and \(c_u\) (for the case where we take \(c_q=-c_u\), for simplicity), for different overall KK scales. We observe that the CP-odd Yukawa couplings are fairly similar to the Higgs Yukawa couplings (i.e. exponentially sensitive to UV localization and then top-like when the zero mode is IR localized) except that there is a range of parameters where the coupling vanishes. Interestingly enough, this suppression happens for preferred values of the top quark bulk mass parameters. This means that the existence of suppressed couplings to top quarks of the CP-odd \(\Pi_0\) is a natural possibility in this scenario. ## Radiative couplings to photons and gluons {#subsec:gggamgam} Just like in the Higgs boson case, the radiative couplings of \(\Pi_0(x)\) to gluons and photons will depend on the physical Yukawa couplings \(y_{nn}\) between \(\Pi_0\) and the fermions (zero modes and KK modes) running in the loop, as well as on the fermion masses \(m_n\) (the eigenvalues of the mass matrix in Eq.([\[mumatrix\]](#mumatrix){reference-type="ref" reference="mumatrix"})). The real and imaginary parts of the couplings are associated with different loop functions, \(A^S_{1/2}\) and \(A^P_{1/2}\), as they generate the two operators \(\Pi_0G_{\mu\nu}G^{\mu\nu}\) and \(\Pi_0 G_{\mu\nu}\tilde{G}^{\mu\nu}\).[^15] The production cross section through gluon fusion is \_gg_0 = \_s\^2 m\_\_0\^2  and the decay widths to gluons and photons are \_\_0gg =  \_\_0 =  where \(\alpha_s\) and \(\alpha\) are the strong and weak coupling constants, \(N_c\) is the number of colors and \(Q_n\) is the charge of the fermion, and where c\^S_n =() A\^S\_1/2(\_n) and c\^P_n = Im() A\^P\_1/2(\_n) with \(\ \tau_n = m^2_{\Pi_0}/4m^2_n\ \) and with the loop functions defined as [\[eq:loopfunction\]]{#eq:loopfunction label="eq:loopfunction"} A\^S\_1/2() &=& 3\^-2,\ A\^P\_1/2() &=&-3 f()/, and with f() = { ------------------------------------------------------------------------------------------------------------------- --------------- \(\left[\arcsin\sqrt{\tau}\right]^2\) \((\tau\le 1)\) \(-{1\over4}\left[\ln\left({1+ \sqrt{1-\tau^{-1}}\over 1- \sqrt{1-\tau^{-1}}}\right)-i\pi\right]^2\) \((\tau> 1)\). ------------------------------------------------------------------------------------------------------------------- --------------- . For heavy KK quarks with masses \(m_n\) much greater than the CP-odd mass \(m_{\Pi_0}\) (i.e. when \(\tau\) is very small) the loop functions are essentially constant, as they behave asymptotically as \(\displaystyle \lim_{\tau \to 0} A^S_{1/2} =1 \hspace{.2cm} \mbox{and} \hspace{.2cm} \lim_{\tau \to 0} A^P_{1/2} =3/2.\) On the other hand, for light quarks (all the SM quarks except top and bottom), the loop functions essentially vanish asymptotically as \(\displaystyle \lim_{\tau \to \infty} A^S_{1/2} = \lim_{\tau \to \infty} A^P_{1/2} = 0.\) Moreover, we investigate a parameter region where the couplings of \(\Pi_0\) to top quarks are highly suppressed. This means that the production mechanism must rely exclusively on the heavy KK fermions running in the loop and as we have seen, this coupling depends on the ratio \(\frac{y_{nn}}{m_n}\) between the physical Yukawa coupling and the mass of the fermion running in the loop. To have an idea of the relative contribution of each of these KK fermions in the loop, in Fig. [\[fig:yPiKKplot\]](#fig:yPiKKplot){reference-type="ref" reference="fig:yPiKKplot"} we plot the mass normalized Yukawa couplings of Standard Model Higgs with top quarks, Higgs with first KK fermion and of \(\Pi_0\) to first KK fermion, for different values of the KK scale. As expected we see that the \(c_q\) dependence is mild (i.e. all KK fermions of any flavor will couple with similar strength) and also, as expected, we observe that the mass normalized couplings are quite suppressed with respect to the SM top quark case. Still the multiplicity of KK fermions is high, since there are 6 families of quarks and 3 families of charged leptons (the latter run in the diphoton loop), and for each family there are a few KK levels that give important contributions to the rate. A numerical scan of the couplings, including all families and 3 full KK levels is computationally too intensive, so in order to produce the couplings plotted in Fig. [\[fig:yPiKKplot\]](#fig:yPiKKplot){reference-type="ref" reference="fig:yPiKKplot"} we performed an approximation, sufficient for the purposes of the graph. The KK fermion Yukawa couplings plotted neglect mixings between different KK levels and different flavors, and with the zero mode fermions. They are obtained as follows. Consider the \(2\times2\) KK mass matrix ( ) M_u ( ), where \(Q_L(x)\) and \(Q_R(x)\) represent here a single flavor and a single KK level of the vector-like KK up-type doublets, and similarly for \(U_L(x)\) and \(U_R(x)\), vector-like KK up-type singlets. The mass matrix is thus M_u= ( ), [\[QUmatrix\]]{#QUmatrix label="QUmatrix"} where the diagonal entries are the KK masses (large) whereas the off-diagonal entries are coming from Yukawa couplings and are therefore smaller. In order to give a simple estimate, we take for simplicity the fermion bulk mass parameters as \(c_q=-c_u\), and the bulk Higgs Yukawa \(Y^{5D}\) to be real, which leads to \(Y_1=Y_2\) and \(m_Q=m_U=m_{KK}\), with the masses and profiles obtained by solving Eq. ([\[fermionprofile\]](#fermionprofile){reference-type="ref" reference="fermionprofile"}). With the KK fermion profiles one obtains the off-diagonal entries Y_1= \_0\^y_1 dy e\^-4(y) Q\_L(y)U\_R(y). The matrix that diagonalizes ([\[QUmatrix\]](#QUmatrix){reference-type="ref" reference="QUmatrix"}) in this simple limit (\(c_q=-c_u\)) is \(\left(\begin{matrix}-\frac{1}{\sqrt{2}} & \frac{1}{\sqrt{2}} \\ \frac{1}{\sqrt{2} }& \frac{1}{\sqrt{2}} \end{matrix}\right)\) and the eigenvalues are \(m_1=m_{KK}-Y_1\) and \(m_2=m_{KK}+Y_1\). Now we apply this rotation to the CP-odd Yukawa coupling matrix Y\_= ( ), where Y\_& =& i \_0\^y_1 dy e\^-3(y) Q\_L(y)U\_R(y) , and where for simplicity we have neglected gauge couplings terms compared to IR Yukawa terms (safe assumption when \(Y^{5D}\) is large). After diagonalization, we obtain the two physical couplings between \(\Pi_0\) and the KK fermions. When we normalize the couplings by the two eigenmasses and add the two contributions,[^16] we obtain \_i=1\^2 =+ =-2 . The last expression corresponds to the *mass normalized* Yukawa couplings of \(\Pi_0\) plotted in Fig. [\[fig:yPiKKplot\]](#fig:yPiKKplot){reference-type="ref" reference="fig:yPiKKplot"}, and this describes very closely the behavior of the couplings obtained in the full flavor calculation. The parametric dependence of these couplings is \(Y_{5D}^2 v/m_{KK}^2\), so that if \(Y^{5D}\sim 3\) we expect mass normalized couplings of order \((10^{-3}-10^{-4})\) GeV\(^{-1}\), if the overlap integral is of \({\cal O}(1)\). Since all the profiles of the integral are IR localized, one expects that integral to be \({\cal O}(1)\), although the precise numerical result varies between \(0.5\) and \(0.05\), depending on the values of the \(c_q\) parameter, as shown in the plots. All in all it seems likely that after taking into consideration all the fermion flavors, and for a KK scale of order \(1-2\) TeV, the overall KK fermion contribution to the radiative couplings of \(\Pi_0(x)\) to photons and gluons can be close to the top quark contribution to the gluon and photon couplings of the Higgs in the SM model. ## \(\Pi_0Zh\) coupling {#subsec:zh} The coupling between the CP-odd scalar, the \(Z\) boson and the Higgs will be extracted from the kinetic operator of the 5D Higgs, d\^4x dy e\^-2 D_H\^D\^H (1+ (y-y_i) ). Expanding the SM-like Higgs mode using Eq. ([\[Hexpansion\]](#Hexpansion){reference-type="ref" reference="Hexpansion"}) as well as the SM-like \(Z_\mu\) and the \(750\) GeV \(\Pi_0\) using Eqs. ([\[Aexpansion\]](#Aexpansion){reference-type="ref" reference="Aexpansion"}) and ([\[Piexpansion\]](#Piexpansion){reference-type="ref" reference="Piexpansion"}), we can obtain the coefficient \(g_{\Pi hZ}\) of the operator \(Z^\mu(x)\Big(h(x)\partial_\mu\Pi_0(x)+\Pi_0(x)\partial_\mu h(x)\Big)\) g\_hZ = dy e\^-2 v(y)h(y) f_z(y) (1+ (y-y_i) ) Now since \(M_z(y)=\frac{g_5}{2c_W}e^{-\sigma} v(y)\), \(h(y)\sim v(y)/v_4\), and \(f_z\simeq 1/\sqrt{y_1}\) we can write g\_hZ (X(y_1)+ X'(y_1))=0, where we have used the boundary conditions for the profile \(X(y)\) (see Eq.([\[Xbc\]](#Xbc){reference-type="ref" reference="Xbc"})) and assumed no UV brane kinetic term (\(d_0=0\)). The coupling should thus vanish in the limit of flat \(Z\) boson profile \(f_z(y)\), and when the nontrivial Higgs VEV \(v(y)\) is proportional to the Higgs scalar profile \(h(y)\). Corrections to these limits scale as \(v_4^2/m_{KK}^2\) and \(m_h^2/m_{KK}^2\) in the RS case, and so we expect the overall coupling to be highly suppressed. The partial width for the decay \(\Pi_0 \to h Z\) is (\_0 h Z)= (m\^2_h,m\^2\_;m\^2_Z) where \(m_Z\), \(m_\pi\) and \(m_h\) are the masses of the particles involved and where \(\lambda(x,y;z) = (1-x/z-y/z)^2-4 x y/z^2\). With the masses \(m_Z=91\) GeV, \(m_\pi=750\) GeV and \(m_h=125\) GeV, the width becomes \(\Gamma(\Pi_0 \to h Z) \sim (900\ g^2_{\Pi hZ})\) GeV. For example we compute numerically \(g_{\Pi hZ}\) for three different values of \(M_{KKglu_1}\), and with \(m_{\Pi_0}=750\) GeV and find Note that the couplings and widths are small but we observe that the partial width becomes larger as the KK mass scale is increased. Since we need the partial width \(\Gamma(\Pi_0 \to \gamma \gamma)\) to be similar to that of a \(750\) GeV Higgs, i.e. \(\Gamma(h_{750GeV} \to \gamma \gamma) \sim 10^{-5}\) GeV, we expect that at \(M_{KKglu_1} =1400\) GeV the \(Zh\) signal should start putting too much pressure on the allowed parameter space. This is confirmed in the full numerical analysis presented in the next section. ## Estimates and numerical results {#subsec:numerical} With all the previous ingredients one can estimate the viability of this scenario in terms of the possible diphoton excess. Let's choose the KK scale such that \(M_{KKglu_1}=1400\) GeV; when the bulk mass parameters of the top quark are around \(|c_{u_3}|\sim 0.35\) we know that the top quark will have highly suppressed couplings to \(\Pi_0\), as shown in the third panel of Fig. [\[fig:yPiplot\]](#fig:yPiplot){reference-type="ref" reference="fig:yPiplot"}. At the same time, the couplings of the KK tops (as well as all other KK quarks) will have relatively strong Yukawa couplings to \(\Pi_0\) (third panel of Fig. [\[fig:yPiKKplot\]](#fig:yPiKKplot){reference-type="ref" reference="fig:yPiKKplot"}), so that the contribution of each of them to the radiative coupling of \(\Pi_0\) to gluons is about an order of magnitude smaller than the top contribution to the \(h-glu-glu\) coupling of the SM. Thus we could estimate that the overall contribution of all flavors and KK excitations can make up for the suppressed coupling, so that the production cross section of \(\Pi_0\) is similar to that of a \(750\) GeV SM-like Higgs. The production cross section of a \(750\) GeV Higgs through gluon fusion, at the LHC running at 13 TeV is \(497\) fb , so, roughly, this could be assumed for the \(\Pi_0\) production cross section. Since the \(\Pi_0\) decays to top quarks and \(hZ\) are suppressed in this parameters space point, and its decays to \(WW\) and \(ZZ\) can only be radiative via the CP-odd gauge boson kinetic operator, the main decay channel is into gluons so that the branching of the diphoton channel should be very roughly Br(\_0 ) \~ where \(N_\gamma\) and \(N_{glu}\) are the multiplicities of states running in the \((\Pi_0 \gamma \gamma)\) loop and in the \((\Pi_0\ glu\ glu)\) loop respectively. In the diphoton loop there are \(3\) extra families of charged lepton KK excitations making the multiplicity of states greater. If their multiplicity and their Yukawa couplings can partially make up for the color factor of 8, then the diphoton cross section might become of \({\cal O}\) (1 fb), as hinted by the December 2015 LHC data. To complete the analysis, we perform a full numerical computation of production and branching ratios in a setup where we consider an effective 4D scenario including three full KK levels for all fields, i.e., we consider \(21 \times 21\) fermion mass matrices, which we diagonalize in order to obtain the physical Yukawa couplings. We choose a set of \(c\)-parameters and 5D Yukawa entries such that the SM masses and mixings are reproduced; the specific flavor choice for these parameters should not affect much the overall results since these depend on overlap integrals between IR localized fields, with very loose \(c\)-dependence. We choose the background metric parameters so that precision electroweak bounds are kept at bay, i.e. \(\nu=0.5\) and \(y_s=1.04 y_1\). Two average 5D Yukawa scales are considered, \(Y_{5D} \simeq 3\) and \(Y_{5D}\simeq 2\), to show the dependence on this parameter, and we also consider two different KK mass scales, \(M_{KKglu_1}=1000\) GeV and \(M_{KKglu_1}=1300\) GeV, which turn out to lead to successful signal generation. In order to see how tuned is the choice of top \(c\)-parameter, we plot the production cross section of the CP odd resonance, followed by decays into \(\gamma \gamma\), \(t\bar{t}\) and \(Zh\), as functions of \(c_{u_3}\) (the bulk mass parameter of the 5D singlet top quark), with the doublet bulk mass parameter fixed at \(c_{q_3}=0.4\). (This value ensures typically suppressed bounds from \(Zb_L{\bar b}_L\) bounds ). The results are shown in Figs. [\[fig:prodplot1\]](#fig:prodplot1){reference-type="ref" reference="fig:prodplot1"} and [\[fig:prodplot2\]](#fig:prodplot2){reference-type="ref" reference="fig:prodplot2"}, where we have chosen two KK scales, \(M_{KKglu_1}=1000\) GeV and \(M_{KKglu_1}=1300\) GeV. In both cases we show results for \(Y_{5D}\simeq 3\) and \(Y_{5D}\simeq 2\) to illustrate the sensitivity on this bulk parameter, crucial for enhancing the radiative couplings of \(\Pi_0\). When the KK scale is smaller and 5D Yukawa couplings are larger, the production of top pairs and gluon pairs can be too large. By reducing the 5D Yukawa couplings, enough signal can be generated with the dijets and top pairs under control, as well as the \(Zh\) decay. The values of \(|c_{u_3}|\) must be in the region \(0.2\mathrel{\raise.3ex\hbox{\)<\(\kern-.75em\lower1ex\hbox{\)\sim\(}}} |c_{u_3}| \mathrel{\raise.3ex\hbox{\)<\(\kern-.75em\lower1ex\hbox{\)\sim\(}}} 0.5\). For slightly smaller KK scales one expects a similar behavior, but such that 5D Yukawa couplings should be smaller in order to suppress overproduction of \(\Pi_0\) particles. This leads to the question of how large can the KK scale be and still manage to produce enough signal. We observe that at \(M_{KKglu_1}=1300\) GeV, larger 5D Yukawa couplings (\(Y_{5d} \geq 3\)) are required in order to have enough signal production. The branching fraction into \(hZ\) starts to be problematic, and actually becomes worse for heavier KK scales. The signal production requires that the value of \(|c_{u_3}|\) be located around the point where the top Yukawa couplings of \(\Pi_0\) are suppressed, in this case around \(|c_{u_3}|\sim 0.3\). We conclude therefore that in order to explain the \(750\) GeV diphoton, the KK scale must be \(M_{KKglu_1} \leq 1300\) GeV, with 5D Yukawa couplings \(Y_{5D}\mathrel{\raise.3ex\hbox{\)<\(\kern-.75em\lower1ex\hbox{\)\sim\(}}} 3\). For those values, the signal is quite generic (i.e. small, but typical, \(c_{u_3}\) values are required), as long as the \(\Pi_0\) mass is set to \(750\) GeV with appropriate boundary kinetic terms. # Discussion {#sec:discussion} We performed an analysis of the scalar sector of warped space models to investigate whether the *minimal* model can accommodate a resonance at 750 GeV with the properties observed in the diphoton resonance at the CMS and ATLAS. We show that in the simplest extra dimensional extension of the SM, that is with a 5D Higgs doublet living in the bulk, the lowest pseudoscalar KK excitation can be responsible for the signal observed at 750 GeV. We emphasize that, unlike other explanations relying on scalar fields in warped models, ours does not introduce any new fields or representations, but relies on Higgs brane kinetic terms to lower the KK mass of the lightest CP odd Higgs resonance. This makes the model extremely constrained, with the only new parameter being the IR brane kinetic coefficient \(d_{1}\). The lightest CP odd excitation, a mixture of the 5D Higgs field and \(Z_5\), does not decay at tree level into \(WW\) or \(ZZ\), and, over a range of the parameter space, can have suppressed couplings to the top quark, thus a small decay width into \(t {\bar t}\). The production through gluon fusion can be loop-enhanced through the effects of the usual KK fermion modes, and so can the diphoton decay. The coupling to \(Zh\) is also suppressed although starts increasing dangerously for KK masses above \(1500\) GeV. We also showed that in \(AdS_5\) spaces (RS-type models) with a KK scale as low as \(1500\) GeV, the lowest CP-odd scalar cannot have a mass at 750 GeV, but if the metric differs slightly from the \(AdS_5\) metric (generalized warped metric spaces), the Higgs brane kinetic terms can produce a viable CP-odd scalar of mass 750 GeV. Within these modified metric scenarios, and for KK mass scales at around 1 TeV (consistent with precision electroweak bounds) this CP odd resonance obeys the (current) experimental constraints. We analyzed its production and decay for several values of the lowest KK gluon mass (\(M_{KKglu_1}\)), and show that consistency with the data requires \(M_{KKglu_1} \mathrel{\raise.3ex\hbox{\)<\(\kern-.75em\lower1ex\hbox{\)\sim\(}}} 1300\) GeV. Our analysis is quite general, and should be valid even in the absence of a signal at 750 GeV. The general conclusion to be taken from our analysis here is that warped space models, without any new particles, can explain a (relatively) light diphoton resonance at the LHC. Should a diphoton excess be found at higher values, even the RS model might accommodate such a state without the need modify the metric. Among the general features of the light CP-odd scalar resonance, resulting from the 5D Higgs doublet are the fact that its coupling to top pairs can be suppressed for appropriate top bulk mass parameters. Also its coupling to \(Zh\) is generically suppressed due to the boundary conditions of the CP odd state. In addition, the model predicts that the spectrum for the CP odd and the charged scalars is essentially the same since their differential equations and boundary conditions are almost identical. This means that the lightest charged Higgs boson is expected to have a mass very close to the pseudoscalar mass, so about 750 GeV, in the scenario in which the latter is the diphoton resonance. If the diphoton resonance ends up being a real particle, and not a statistical fluctuation, the prediction from this scenario is that \(ZZ\) and \(WW\) should not be seen, whereas top pair production, dijet production and \(Zh\) signal should be around the corner. Moreover a search for the charged scalars at around \(750\) GeV might prove useful in order to disentangle the different models explaining the resonance. # Appendix In this section we consider the effect of brane localized kinetic terms associated with the 5D Higgs doublet and also with the gauge bosons. For simplicity, let's consider a 5D toy model with a Higgs scalar \(H(x,y)\) charged under a local \(U(1)\), defined by the following action: S&=&d\^4xdy  (- F\^2\_MN + \|D\^M H\|\^2-V(H) )\ &&   + \_i d\^4xdy  (y-y_i) ( r_i F\^2\_MN + d_i \|D\^M H\|\^2-\_i(H))) [\[5Daction\]]{#5Daction label="5Daction"} where \(F_{MN} = \partial _M A_N-\partial_N A_M\) and for simplicity we set gauge coupling constant to unity in the appropriate mass dimensions. The background spacetime metric is assumed to take the form ds\^2=e\^-2(y)\_ dx\^dx\^-dy\^2 where \(\sigma(y)\) is the warp factor. We are interested in studying the effective 4D perturbative spectrum of the 5D Higgs field and the 5D gauge boson, around a nontrivial Higgs vacuum profile solution \(<H>=v(y)\). H(x,y)= (v(y) + h(x,y)) e\^i (x,y) In particular we are interested in the CP-odd Higgs perturbations \(\pi(x,y)\), whose equations of motion are coupled with the gauge boson perturbations. The equations read && (1+r_i \_i) \_\^A\_-( (1+r_i \_i)e\^-2 A'\_)' + (1+d_i \_i) M_A\^2 A\_+\ &&\_((1+d_i \_i) M_A\^2 -(1+r_i \_i)\^A\_-((1+r_i \_i)e\^-2 A_5)') = 0\ &&(1+r_i \_i) \_\^A_5-(1+r_i \_i) \^A'\_+ (1+d_i \_i)M_A\^2 ('-A_5)=0\ &&(1+d_i \_i) \_\^-(1+d_i \_i) \^A\_+ M_A\^-2((1+d_i \_i) M_A\^2 e\^-2 ('-A_5))'= 0 where \(M_A= v(y) e^{-\sigma}\) and where \(\displaystyle d_i \delta_i \equiv \sum_i d_i \delta(y-y_i)\) and \(\displaystyle r_i \delta_i \equiv \sum_i r_i \delta(y-y_i)\). We fix partially the 5D gauge by imposing (1+d_i \_i) M_A\^2 -(1+r_i \_i)\^A\_-((1+r_i \_i)e\^-2 A_5)' = 0. [\[gaugefix\]]{#gaugefix label="gaugefix"} The previous gauge fixing equation reads in the bulk: M_A\^2 -\^A\_-(e\^-2 A_5)' = 0. [\[bulkgaugefix\]]{#bulkgaugefix label="bulkgaugefix"} Note that if we evaluate the bulk constrain Eq. ([\[bulkgaugefix\]](#bulkgaugefix){reference-type="ref" reference="bulkgaugefix"}) at \(y=y_1-\epsilon\) (i.e. right before the IR brane), we obtain: \^A\_\|\_y_1- = M_A\^2 -(e\^-2 A_5)'\|\_y_1-. [\[epsiloncondition\]]{#epsiloncondition label="epsiloncondition"} On the other hand, the effect of the delta functions in Eq.([\[gaugefix\]](#gaugefix){reference-type="ref" reference="gaugefix"}) is to produce a discontinuity in the 5D field \(A_5\) at the brane location as d_1 M_A\^2 -r_1 \^A\_\|\_y_1- = \_y_1-\^y_1, and similarly for the UV brane. We can thus multiply ([\[epsiloncondition\]](#epsiloncondition){reference-type="ref" reference="epsiloncondition"}) by \(r_1\) and use it in the previous equation, and find the necessary boundary condition between \(\pi\) and \(A_5\), which ensures that \(A_\mu\) is completely decoupled, even on the brane. We find (d_1-r_1) M\^2_A + r_1 (e\^2 A_5)' . \|\_y_1- =-.e\^-2 A_5\|\_y_1- [\[boundaryconditions\]]{#boundaryconditions label="boundaryconditions"} where we have taken \(A_5\) to vanish exactly on the brane, but it jumps right before the boundary. Inserting the gauge choice in the coupled equations of motion, one manages to decouple the gauge modes \(A_\mu\) (in both the bulk and the branes) with a bulk equation \_\^A\_-(e\^-2 A'\_)' + M\_A\^2 A\_ and jump condition on \(A'_\mu\) r_1 \_\^A\_+ d_1 M\_A\^2 A\_\|\_y_1-=-. e\^-2 A'\_\|\_y_1- where \(A'_\mu\) again vanishes exactly on the brane, but has a jump right before it. We separate variables A\_(x,y)= V\^4d\_(x) V_y(y) and find the separated equations for the gauge boson tower become && \_\^V\^4d\_(x)+m\^2_A V\^4d\_(x)=0\ && ( e\^-2 V'\_y)' + ( m\^2_A-M_A\^2 ) V_y =0 with jump conditions on \(V'_y\) (d_i M\_A\^2-r_i m\^2_A) V_y \|\_y_1-=-. e\^-2 V'\_y\|\_y_1- where the 4D effective mass \(m^2_A\) is the constant of separation of variables. The remaining equations are, in the bulk, && \_\^A_5+ M_A\^2 ('-A_5)-(M_A\^2 )' +( (e\^-2 A_5)' )' =0\ && \_\^+ M_A\^-2 ( M_A\^2 e\^-2 ('-A_5) )'-M_A\^2 + (e\^-2 A_5)'= 0[\[a5pimixedequation\]]{#a5pimixedequation label="a5pimixedequation"} and the fields must verify the boundary conditions of Eq.([\[boundaryconditions\]](#boundaryconditions){reference-type="ref" reference="boundaryconditions"}). We now perform a mixed separation of variables: A_5(x,y)&=& G(x) g(y) + \_x(x) (y) [\[expansion1\]]{#expansion1 label="expansion1"}\ (x,y)&=& G(x) h(y) + \_x(x) which is to say that both \(A_5(x,y)\) and \(\pi(x,y)\) contain each some Goldstone and CP-odd degree of freedom. The profiles \(g(y)\), \(\eta(y)\), \(h(y)\) and \(quantify how much of each they contain. Of course the functions\)g\(and\)h\(are related to each other, as well as\]and\) With the choice h(y)&=&\ g(y)&=&\ (y)&=&X(y)\ and using the mixed separation of variables in ([\[expansion1\]](#expansion1){reference-type="ref" reference="expansion1"}) and ([\[expansion2\]](#expansion2){reference-type="ref" reference="expansion2"}), the mixed equations of motion in ([\[a5pimixedequation\]](#a5pimixedequation){reference-type="ref" reference="a5pimixedequation"}) decouple and we obtain && X(y) \_\^\_x(x) +\_x(x) e\^2 X(y)-\_x(x) '= 0\ &&K(y) \_\^G(x)+ G(x) = 0. Once separated, we obtain, for the CP odd physical scalars &&\_\^\_x(x) + m\_\^2 \_x(x) =0\ &&(M_A\^-2 X')'+ e\^2 (-1) X=0, with boundary conditions d_i X' =-X and for the Goldstone modes, &&\_\^G(x)+ m_G\^2 G(x)=0\ &&(K'e\^-2)'+ (m_G\^2-M_A\^2) K=0, with boundary conditions (d_i M_A\^2-r_i m\^2_G ) K =-e\^-2 K'. Note that both the equations and boundary conditions for the Goldstone bosons are identical to the ones for the gauge boson tower, as should be, so that they can then be gauged away level by level with the remaining gauge fixing freedom. [^1]: In order to avoid tachyons and/or ghosts, the sign of the purely brane localized brane kinetic terms will be kept positive, i.e. \(d_i>0\). [^2]: There is still be some gauge freedom left, so that the towers of 4D Goldstone bosons that appear can be gauged away. [^3]: In the absence of brane kinetic terms, \(Z_5\) must have vanishing boundary conditions (Dirichlet) if \(Z_\mu\) is to have Neumann conditions and thus develop a zero mode KK excitation in the effective 4D theory. [^4]: We ignore here possible brane localized gauge kinetic terms and keep only the effects from Higgs brane kinetic terms. We include everything in the derivation outlined in the Appendix. [^5]: Of course, this RS point is presented for comparison only, since such light KK masses will produce too large deviations in the precision electroweak observables. [^6]: We are actually plotting the values defined in Eqs. ([\[YYPI\]](#YYPI){reference-type="ref" reference="YYPI"}) and ([\[YYY1KK\]](#YYY1KK){reference-type="ref" reference="YYY1KK"}), i.e. the zero mode Yukawa couplings before going to the fermion mass basis. In that basis, the couplings will inherit a small correction due to mixing with heavy KK fermions, so that the exact cancellation of the coupling will be replaced by a strong suppression. [^7]: The Yukawa couplings of \(\Pi_0\) are mostly imaginary and thus the dominant contribution will come, as expected, from the operator \(\Pi_0 G_{\mu\nu}\tilde{G}^{\mu\nu}\). Still, small real Yukawa coupling components are generated when going to the fermion mass basis, and so we keep the general formalism in our formulas. [^8]: One needs to add the two contributions since there is a cancellation happening level by level. [^9]: In order to avoid tachyons and/or ghosts, the sign of the purely brane localized brane kinetic terms will be kept positive, i.e. \(d_i>0\). [^10]: There is still be some gauge freedom left, so that the towers of 4D Goldstone bosons that appear can be gauged away. [^11]: In the absence of brane kinetic terms, \(Z_5\) must have vanishing boundary conditions (Dirichlet) if \(Z_\mu\) is to have Neumann conditions and thus develop a zero mode KK excitation in the effective 4D theory. [^12]: We ignore here possible brane localized gauge kinetic terms and keep only the effects from Higgs brane kinetic terms. We include everything in the derivation outlined in the Appendix. [^13]: Of course, this RS point is presented for comparison only, since such light KK masses will produce too large deviations in the precision electroweak observables. [^14]: We are actually plotting the values defined in Eqs. ([\[YYPI\]](#YYPI){reference-type="ref" reference="YYPI"}) and ([\[YYY1KK\]](#YYY1KK){reference-type="ref" reference="YYY1KK"}), i.e. the zero mode Yukawa couplings before going to the fermion mass basis. In that basis, the couplings will inherit a small correction due to mixing with heavy KK fermions, so that the exact cancellation of the coupling will be replaced by a strong suppression. [^15]: The Yukawa couplings of \(\Pi_0\) are mostly imaginary and thus the dominant contribution will come, as expected, from the operator \(\Pi_0 G_{\mu\nu}\tilde{G}^{\mu\nu}\). Still, small real Yukawa coupling components are generated when going to the fermion mass basis, and so we keep the general formalism in our formulas. [^16]: One needs to add the two contributions since there is a cancellation happening level by level.
{'timestamp': '2016-07-18T02:07:40', 'yymm': '1607', 'arxiv_id': '1607.04534', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04534'}
# Proof of Theorem [\[jamesdreamcometrue\]](#jamesdreamcometrue){reference-type="ref" reference="jamesdreamcometrue"} {#doyoulift} In this section, we provide a complete proof of the main theorem of this paper, namely Theorem [\[jamesdreamcometrue\]](#jamesdreamcometrue){reference-type="ref" reference="jamesdreamcometrue"}. The basic strategy has two steps: lift from \(\ell^2\) to \(\ell^\infty\) (see Section [1.1](#allthewaytonight){reference-type="ref" reference="allthewaytonight"}), and lift from \(\ell\) to \(\ell^2\) (see Section [1.2](#bataneye){reference-type="ref" reference="bataneye"}). Considerable care must be taken in dealing with the cases where \(\ell = 2, 3\), so we treat these situations separately (see Sections [1.3](#booyah){reference-type="ref" reference="booyah"} and [1.4](#thisishisbananas){reference-type="ref" reference="thisishisbananas"}). We execute this strategy as follows: ## Lifting from \(\ell^2\) to \(\ell^\infty\) for \(\ell \ge 3\), and from \(8\) to \(2^\infty\) {#allthewaytonight} ## Lifting from \(\ell\) to \(\ell^2\) for \(\ell \ge 5\) {#bataneye} ## Lifting from \(4\) to \(8\) {#booyah} ## Lifting from 2 to 4 and from 3 to 9 {#thisishisbananas} ### Idea of Proof {#idea-of-proof .unnumbered} The argument will proceed by induction on \(g\). We start by verifying the base case \(g = 2\) in Lemma [\[lemma:base-case-low-lifting\]](#lemma:base-case-low-lifting){reference-type="ref" reference="lemma:base-case-low-lifting"}. We then inductively assume this holds for \(g-1\) and prove it for \(g\). We use the inductive hypothesis to construct a particular element lying in \(H(\ell^2)\) in Lemma [\[lemma:particular-matrices-in-preimage\]](#lemma:particular-matrices-in-preimage){reference-type="ref" reference="lemma:particular-matrices-in-preimage"}. Then, we use Lemma [\[lemma:code\]](#lemma:code){reference-type="ref" reference="lemma:code"} to show that conjugates of this particular element generate \(\mathfrak{sp}_{4}(\mathbb Z/ \ell^2 \mathbb Z)\), embedded as in  [\[abovesubspace\]](#abovesubspace){reference-type="eqref" reference="abovesubspace"}. We finally translate around this copy of \(\mathfrak{sp}_{4}(\mathbb Z/ \ell^2 \mathbb Z)\) to obtain that \(H(\ell^2) = \on{Sp}_{2g}(\mathbb Z/\ell^2 \mathbb Z)\). The following lemma deals with the base case: This concludes the proof of Lemma [\[lemma:base-case-low-lifting\]](#lemma:base-case-low-lifting){reference-type="ref" reference="lemma:base-case-low-lifting"}. ◻ Next, in Lemma [\[lemma:code\]](#lemma:code){reference-type="ref" reference="lemma:code"}, we show that we can conjugate the matrices \(\matone\) and \(\mattwo\) from Lemma [\[lemma:particular-matrices-in-preimage\]](#lemma:particular-matrices-in-preimage){reference-type="ref" reference="lemma:particular-matrices-in-preimage"} to obtain all of \(\on{Sp}_4(\mathbb Z/\ell \mathbb Z)\). This concludes the proof of Lemma [\[lemma:code\]](#lemma:code){reference-type="ref" reference="lemma:code"} ◻ ## Finishing the Proof We are now in position to complete the proof of Theorem [\[jamesdreamcometrue\]](#jamesdreamcometrue){reference-type="ref" reference="jamesdreamcometrue"} # Introduction Let \(g\) be a positive integer, and for a ring \(R\), denote by \(\mathrm{Sp}_{2g}(R)\) the group of \(2g \times 2g\) symplectic matrices over \(R\). Let \(\mathbb Z_\ell\) denote the ring of \(\ell\)-adic integers, and consider the natural projection map \(\on{Sp}_{2g}(\mathbb Z_\ell) \to \on{Sp}_{2g}(\mathbb Z/\ell \mathbb Z)\). In this paper, we show that when \(g > 1\), there are no proper closed subgroups of \(\on{Sp}_{2g}(\mathbb Z_\ell)\) that surject via this projection map onto all of \(\on{Sp}_{2g}(\mathbb Z/\ell \mathbb Z)\). The case \(g = 1\), in which \(\on{Sp}_2 = \SL_2\), is well-understood. Indeed, as proven in , if \(\ell \geq 5\) and \(H_\ell \subset \SL_2(\mathbb Z_\ell)\) is a closed subgroup that surjects onto \(\SL_2(\mathbb Z/\ell \mathbb Z)\), then \(H_\ell = \SL_2(\mathbb Z_\ell)\). The corresponding result for \(\ell \in \{2,3\}\) simply does not hold: in each of these cases, there are nontrivial subgroups of \(\on{Sp}_{2g}(\mathbb Z / \ell^2 \mathbb Z)\) that surject onto \(\SL_2(\mathbb Z/\ell \mathbb Z)\). See  for exercises outlining a proof, and also see for more comprehensive descriptions  for the case \(\ell = 2\) and  for the case \(\ell = 3\). The objective of the present article is to generalize , to hold for all \(g \geq 2\). Our main theorem is stated as follows: The rest of this paper is organized as follows. In Section [3.1](#subsection:stimpy){reference-type="ref" reference="subsection:stimpy"}, we introduce the basic definitions and properties of the symplectic group. Next, in Section [3.2](#subsection:abelianizations){reference-type="ref" reference="subsection:abelianizations"}, we compute the commutator subgroups of \(\on{Sp}_{2g}(\mathbb Z_\ell)\) and \(\on{Sp}_{2g}(\mathbb Z/\ell^k \mathbb Z)\) for every prime number \(\ell\) and positive integer \(k\). Finally, in Section [1](#doyoulift){reference-type="ref" reference="doyoulift"}, we prove Theorem [\[jamesdreamcometrue\]](#jamesdreamcometrue){reference-type="ref" reference="jamesdreamcometrue"}. # Background on Symplectic Groups In this section, we first detail the basic definitions and properties of symplectic groups, and then proceed to prove a number of lemmas that are used in our proof of Theorem [\[jamesdreamcometrue\]](#jamesdreamcometrue){reference-type="ref" reference="jamesdreamcometrue"}. ## Symplectic Groups {#subsection:stimpy} Fix a commutative ring \(R\), let \(\operatorname{Mat}_{2g \times 2g}(R)\) denote the space of \(2g \times 2g\) matrices with entries in \(R\), and let \(\Omega_{2g} \in \operatorname{Mat}_{2g \times 2g}(R)\) be defined by \[\Omega_{2g} \defeq \left[\begin{array}{c|c} 0 & \mathrm{id}_g \\ \hline-\mathrm{id}_g & 0\end{array}\right],\] where \(\mathrm{id}_g\) denotes the \(g \times g\) identity matrix. We define the symplectic group \(\on{Sp}_{2g}(R)\) as the set of \(M \in \SL_{2g}(R)\) so that \(M^T \Omega_{2g} M = \Omega_{2g}\). In the proof of Theorem [\[jamesdreamcometrue\]](#jamesdreamcometrue){reference-type="ref" reference="jamesdreamcometrue"}, we will make heavy use of the "Lie algebra" \(\mathfrak{sp}_{2g}(R)\), which is defined by \[\begin{aligned} \mathfrak{sp}_{2g}(R) &\defeq \{M \in \operatorname{Mat}_{2g \times 2g}(R): M^T \Omega_{2g} + \Omega_{2g} M = 0 \}. \end{aligned}\] It is easy to see that \(M^T \Omega_{2g} + \Omega_{2g}M = 0\) is equivalent to \(M\) being a block matrix with \(g \times g\) blocks of the form \[M = \left[\begin{array}{c|c} A & B \\ \hline C &-A^T \end{array}\right],\] where \(B\) and \(C\) are symmetric. In what follows, we specialize to studying symplectic groups over \(R = \mathbb Z\), \(R = \mathbb Z_{\ell}\), or \(R = \mathbb Z / \ell^k \mathbb Z\) for \(\ell\) a prime number and \(k\) a positive integer. We will adhere to the following notational conventions: - Let \(H_{\ell} \subset \on{Sp}_{2g}(\mathbb Z_\ell)\) be a closed subgroup. - Let \(H(\ell^k) \subset \on{Sp}_{2g}(\mathbb Z/\ell^k \mathbb Z)\) be the mod-\(\ell^k\) reduction of \(H_\ell\). - Notice that the map \(S \mapsto \mathrm{id}_{2g} + \ell^k S\) gives an isomorphism of groups \[\mathfrak{sp}_{2g}(\mathbb Z/ \ell \mathbb Z) \simeq \ker(\on{Sp}_{2g}(\mathbb Z/\ell^{k+1} \mathbb Z) \to \on{Sp}_{2g}(\mathbb Z/ \ell^k \mathbb Z))\] for every \(k \geq 1\). We will use the Lie algebra notation \(\mathfrak{sp}_{2g}(\mathbb Z/\ell\mathbb Z)\) to denote the above kernel when we want to think of its elements additively, and we will use the kernel notation when we want to view its elements multiplicatively. - For any group \(G\), let \([G,G]\) be its commutator subgroup, and let \(G^{\ab} = G/[G,G]\) be its abelianization. ## Commutators {#subsection:abelianizations} We shall now compute the abelianizations of \(\on{Sp}_{2g}(\bz_\ell)\) and \(\on{Sp}_{2g}(\bz / \ell^k \mathbb Z)\) for every integer \(g \geq 2\), prime number \(\ell\), and positive integer \(k\). It will first be convenient for us to compute the abelianization \(\on{Sp}_{2g}(\mathbb Z)^{\ab}\). Using Lemma [\[theorem:commutator-sp-z\]](#theorem:commutator-sp-z){reference-type="ref" reference="theorem:commutator-sp-z"}, we can now compute the abelianizations of all aforementioned groups.
{'timestamp': '2017-03-28T02:05:31', 'yymm': '1607', 'arxiv_id': '1607.04698', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04698'}
null
null
# Introduction Pre-class preparation positions students to get the most out of face-to-face class time. This is particularly important for active learning classrooms, where students rely on their prior learning to participate in peer discussion and construct their own further understanding of key concepts. The standard approach to foster student preparedness is the assignment of pre-class reading. This is an evidence-based practice: in courses using pre-assignments coupled with quizzes, students complete the reading more consistently, ask more---and more challenging---questions during class, and perform better on course assessments and exams. Notably, a common thread among most studies is in the use of a relatively passive modality in pre-class assignments---textbook reading or watching videos. While it is possible that students may actively engage---by taking notes, developing questions, and making connections with the material---there is no guarantee that they are using these types of strategies. Additionally, they may have low intrinsic motivation to deeply engage, given the minimal opportunity for inquiry and exploration. As a possible solution, multimedia learning modules used for pre-lecture assignments were shown to promote learning and to improve student attitudes. Thus, there is a clear opportunity to deliberately use non-passive tools in the context of pre-class preparation assignments to support active engagement, inquiry, and other positive student outcomes. Simulations offer one such tool. The PhET Interactive Simulations have been carefully designed with dual goals of student engagement and learning. The results have been positive: PhETs promote learning, especially when students are exploring in a manner driven by their own questioning. Additionally, the nature of PhET-based assignments that promote productive engagement has been investigated. In interviews, maximal engagement was seen in the presence of minimal---but non-zero---guidance. A separate in-class study provided students with either light-, moderate-, or heavy-guidance PhET-based assignments; students in the light guidance condition explored the PhET more and paid more attention to their interactions with the simulation. Adams *et al.* found that students can learn from PhETs at home, in an unstructured environment, and that moderate, question-driven guidance may be best in this context. ::: # Discussion From pre-to post-test, students learned during the pre-class assignments, with a pedagogically significant effect size. This indicates that the assignments were indeed successful in helping prepare students for class. Although we hypothesized that PhETs could contribute to learning during the pre-class assignments, our analysis revealed no effect on our measures of learning relative to the textbook-only assignments. It could be that, for the at-home pre-class assignment, our light guidance did not provide the scaffolding students would need to productively explore within the PhET activity. If this were the case, the Text then PhET treatment, for which reading the textbook before the PhET activity may provide extra scaffolding, may have had a better chance of providing the structure students would need for productive engagement with the PhET. However, no learning effect was seen whether students were constrained to do the PhET activity before or after the textbook reading, so it may be that, no matter the order, the level of guidance was too light for this assignment. In their study of PhET activities in unstructured environments, Adams *et al.* were unable to create effective activities with light guidance; it may be that more direct instruction is necessary for pre-class activities as well. It is possible that, although the assignment was constrained, students referred to the textbook through all parts of the assignment. Then, by also having a textbook assignment, productive exploration may have been undermined. Because the textbook part of the assignment is important to students' preparation for the class---so that they have encountered basic definitions and concepts---it may be difficult to disentangle the learning specific to the PhET activity. It may also be that our assessment questions, while appropriate for the topic in general, did not target the specific concepts addressed in the PhETs, or that these particular PhETs did not provide exploration opportunities at a level appropriate for this audience. Including a PhET activity in the pre-class assignment influenced both student engagement with and perceptions of the learning benefits of the assignments. When PhET was included, more students reported the pre-class assignments as enjoyable and as many students as the text-only condition rated them as beneficial to their learning. Additionally, when PhET was included in the Resonance topic, students reported spending 5--10 minutes longer with the assignments. This is notable, as Resonance was the most difficult of the topics and was encountered late in the semester, when competing priorities and general fatigue may have prevented students from engaging. The affective benefits of including PhET may have helped to sustain the motivation of these students. Interestingly, the reported time spent on the PhET activity did not seem to depend on the ordering of activities. If completing the textbook reading before the PhET activity affected how students interacted with the PhET, it did not show up here. Overall, students in both PhET conditions reported spending marginally less time on the textbook part of the assignment. This is odd, because students in both the Text or Text then PhET treatment did not know if they were going to have a PhET activity, so we would expect the textbook engagement for these treatments to match. Since their perceived time spent on the assignments was recorded after finishing all activities, it appears that having PhET as part of the assignment altered how students thought about the assignment as a whole. There was an interesting correlation between time spent on the PhET and perceptions of the pre-class assignments. It could be that students who enjoyed the assignment, topic, or PhET more were pre-disposed to spend more time with the simulation, or it could be that students who spent more time with the PhET ended up enjoying the assignment more. More study is needed to understand how student engagement and attitudes interact. No effect of PhET engagement on learning was observed, indicating that the extra time spent did not translate into learning gains on our measures. # Conclusion PhETs are not a magic bullet: They have been shown to promote learning with carefully designed activities in particular contexts. This report shows that even purposeful inclusion of PhETs in pre-class assignments does not automatically translate into learning. In the case of pre-class assignments, more study is required to see if it is possible to design supplemental PhET activities that improve student preparation beyond that from a text-only assignment. The inclusion of PhET did have an effect on both student engagement with the pre-class assignment and student perceptions of the assignments. These affective results suggest that PhETs did contribute meaningfully to these assignments. Sustaining student motivation can be a challege, so the potential for PhETs to contribute in this domain is an important consideration.
{'timestamp': '2016-07-18T02:08:46', 'yymm': '1607', 'arxiv_id': '1607.04588', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04588'}
null
null
# Introduction {#sec:intro} As we enter the era of precision LHC phenomenology, where NNLO QCD calculations are becoming the standard for many processes, the influence of electroweak corrections is increasingly relevant. A complete treatment of these inevitably requires the inclusion of diagrams with initial--state photons, with corresponding photon parton distribution function (PDF) introduced in analogy to the more commonly considered PDFs of the quarks and gluons . As discussed recently in  the photon--initiated contribution may be significant for the production of lepton, \(W\) boson and top quark pairs at higher invariant masses, and hence higher parton \(x\). Such processes are of much phenomenological interest, being particularly sensitive to electroweak corrections and the PDFs, as well BSM physics; high mass lepton pair production, for example, is an irreducible background to the Drell--Yan production of a new \(Z'\) boson. The issue of how to extract the photon PDF, and what the uncertainties associated with this are, is therefore crucial to any complete discussion of these processes. A range of approaches to this problem have been taken by the global PDF fitting groups: the first attempt in the MRST2004QED set  fixed the functional form of the photon PDF by taking a simple model for photon emission from the valence quarks, while CT14QED  generalized this to allow freedom in the overall normalization, which can then be extracted from data. An alternative approach is taken in NNPDF2.3QED  (and more recently NNPDF3.0QED ), where instead the photon is parameterised freely, as in the case of the quarks and gluons, and fitted to a selection of inclusive data. In this case, the corresponding PDF uncertainties due to the quite unconstraining data considered in these fits are very large. However, these previous approaches in fact omit an important physical distinction between the photon and the quarks and gluons. The crucial difference is that QED, in contrast to QCD, corresponds to a long range force that does not suffer from the issue of non--perturbativity at low scales. Thus, a proton may coherently emit a photon (\(p\to p\gamma\)): such a process is experimentally extremely well measured, being governed by the well known electric and magnetic proton form factors for coherent photon emission, and is expected to constitute the dominant component of the input photon PDF. This is accounted for in the approach of  (see also ), where it is shown that the photon PDF is then determined to a relatively high degree of accuracy. It is important to emphasise that the inclusion of this effect is not a theoretical assumption: indeed, exclusive lepton and \(W\) boson pair production, due to precisely this initial--state coherent photon emission have been observed by both ATLAS  and CMS  at the LHC. Such processes contribute by definition to the corresponding inclusive observables. It is therefore important to consider the consequences of these physical considerations, and the approach which derives from it for describing the photon PDF; we will consider for concreteness in this paper the cases of high mass lepton and \(W\) boson pair production. Here, the \(\sim \alpha^2\) suppression in the initial--state \(\gamma\gamma\) luminosity may be overcome by the enhancement of the \(t\)--channel photon--induced process at higher masses. However, for certain sets the most significant effect is due to the PDFs themselves: the NNPDF2.3QED  set in particular predicts a sharper decrease in the quark (and gluon) densities compared to the central photon value, albeit within sizeable PDF uncertainties in the latter case. In such a situation, for both processes the photon--initiated contribution is found to be potentially sizeable and even dominant at high invariant mass . The case of \(t\overline{t}\) production has also recently been discussed in , where it is again shown that the NNPDF2.3QED set is consistent with a sizeable photon--initiated contribution at larger invariance masses, as well as forward rapidities. Although for the sake of brevity we will not deal with this explicitly here, our results can be readily extended to such a process. Given these findings, it is natural to consider what the prediction is for these processes within the approach of  (see also ). In other words, what are the consequences of this dominantly coherent photon input PDF for the size and uncertainties of the photon PDF at higher \(x\), and what are the implications for these high mass production processes? This is the question we consider in this paper: we will show that this approach in fact predicts that the photon PDF is well constrained by the dominantly coherent input requirement, with at higher \(x\) any unknown incoherent contribution, already expected to be small, being further kinematically suppressed. We will demonstrate that this follows from the relatively simple nature of the photon DGLAP evolution, which due to the small size of the coupling \(\alpha\) may to very good approximation be solved exactly . Within this approach, we will find that the decrease in the \(\gamma\gamma\) luminosity is qualitatively similar to the quark and gluon cases. We will also discuss how the recent ATLAS measurement of high mass lepton pair production  and the corresponding extraction of the photon PDF presented in this analysis lends qualitative support to our results. The production of lepton and \(W\) boson pairs are processes of much phenomenological interest at both the LHC and at a Future Circular Collider (FCC). With this in mind, we will consider cross section predictions for both of these processes at 13 and 100 TeV. In the case of lepton pair production we will show that the DY contribution is expected to be dominant out to very high masses, with a relatively small contribution from the photon---initiated process (\(\sim 10\)% for our choice of cuts). For \(W\) pair production, on the other hand, the relative contribution from the photon--initiated process is more significant, and at higher masses it is comparable in size to the standard QCD--initiated process. In this case a very careful consideration of the uncertainties associated with the photon PDF is therefore essential. Finally, during the last stages of preparing this manuscript the paper  appeared. While the overall approach to treating the photon PDF and the details of the analysis are quite distinct, this work also includes the dominantly coherent input component. We may therefore expect the general conclusions to be consistent with our findings. Although we will not provide a detailed investigation of this question here, we present a brief comparison to the results of this approach. As we shall see, the predicted photon PDF does indeed quite closely coincide with our results. Thus, we will expect comparable results to hold for the cases of high mass lepton and \(W\) boson pair production when using the `LUXqed` set. The outline of this paper is as follows. In Section [2](#sec:photongen){reference-type="ref" reference="sec:photongen"} we describe our approach to modelling the photon PDF, demonstrating in Section [2.1](#sec:photondglap){reference-type="ref" reference="sec:photondglap"} how the DGLAP equation for the photon PDF may be solved, and then discussing in Section [2.2](#sec:input){reference-type="ref" reference="sec:input"} how the input photon PDF may be described. In Section [3](#sec:photonhighx){reference-type="ref" reference="sec:photonhighx"} we discuss the implications for the photon, in particular at higher \(x\), and compare our results with the other available PDFs, concentrating on the NNPDFQED3.0 set. In Section [4](#sec:dy){reference-type="ref" reference="sec:dy"} we present predictions for high mass lepton and \(W\) boson pair production at the LHC and \(\sqrt{s}=100\) TeV FCC. Finally, in Section [5](#sec:conc){reference-type="ref" reference="sec:conc"} we conclude. # The photon PDF: general considerations {#sec:photongen} ## Solving the DGLAP equation {#sec:photondglap} The starting point for any discussion of the photon PDF is the corresponding DGLAP evolution equation for the distribution \(\gamma(x,Q^2)\). At LO in \(\alpha\) and \(\alpha_S\) this is given by[^1] \[\label{dglap} \frac{\partial \gamma(x,Q^2)}{\partial \ln Q^2}=\frac{\alpha(Q^2)}{2\pi}\!\int_x^1\!\frac{dz}z\! \left(P_{\gamma\gamma}(z)\gamma(\frac xz,Q^2) +\sum_q e^2_qP_{\gamma q}(z)q(\frac xz,Q^2)\right)\;.\] Here \(P_{\gamma q}(z)\) is the \(q\to \gamma\) splitting function, and \(P_{\gamma\gamma}\) corresponds to the virtual self--energy correction to the photon propagator, given by \[\label{pgg} P_{\gamma\gamma}(z)=-\frac{2}{3}\left[N_c\sum_q e^2_q +\sum_l e^2_l\right]\delta(1-z)\;,\] where \(q\) and \(l\) denote the active quark and lepton flavours in the fermion loop. As the virtual correction ([\[pgg\]](#pgg){reference-type="ref" reference="pgg"}) is proportional to an overall delta function the corresponding contribution to ([\[dglap\]](#dglap){reference-type="ref" reference="dglap"}) is proportional to the photon PDF evaluated at \(x\). Therefore, if we ignore the small effect that the photon PDF has on the evolution of the quark and gluons (as discussed in , these generally give less than a 0.1% correction to the photon), which enter at higher orders in \(\alpha\), then ([\[dglap\]](#dglap){reference-type="ref" reference="dglap"}) can be solved exactly, giving  \[\begin{aligned} \nonumber \gamma(x,\mu_F^2)&=\gamma(x,Q_0^2)\,S_{\gamma}(Q_0^2,\mu_F^2)+\int_{Q_0^2}^{\mu_F^2}\frac{\alpha(Q^2)}{2\pi}\frac{dQ^2} {Q^2}\int_x^1\frac{dz}z \;\sum_q e^2_qP_{\gamma q}(z)q(\frac xz,Q^2)\,S_{\gamma}(Q^2,\mu_F^2)\;,\\ \label{gampdf} &\equiv \gamma^{{\rm in}}(x,\mu^2)+\gamma^{\rm evol}(x,\mu^2)\;, \end{aligned}\] where \(\gamma(x,Q_0^2)\) is the input PDF at the scale \(Q_0\), and we have introduced the photon Sudakov factor \[\label{sudgam} S_{\gamma}(Q_0^2,\mu_F^2)=\exp\left(-\frac{1}{2}\int_{Q_0^2}^{\mu_F^2}\frac{{\rm d}Q^2}{Q^2}\frac{\alpha(Q^2)}{2\pi}\int_0^1 {\rm d} z\sum_{a}\,P_{a\gamma}(z)\right)\;.\] Here \(P_{q(l)\gamma}(z)\) is the \(\gamma\) to quark (lepton) splitting function, given by \[P_{a\gamma}(z)=N_a\left[z^2+(1-z)^2\right]\;,\] where \(N_a=N_c e_q^2\) for quarks and \(N_a=e_l^2\) for leptons, while the factor of \(1/2\) in ([\[sudgam\]](#sudgam){reference-type="ref" reference="sudgam"}) is present to avoid double counting over the quark/anti--quarks (lepton/anti--leptons). Written in this form, the physical interpretation of the Sudakov factor is clear: it represents the Poissonian probability for no parton emission from the photon during its evolution from the low scale \(Q_0\) up to the hard scale \(\mu_F\). Thus, the photon PDF ([\[gampdf\]](#gampdf){reference-type="ref" reference="gampdf"}) at \(\mu_F\) can be written as the sum of a contribution from low--scale emission of a photon, with no further branching, and a term due to higher scale DGLAP emission from quarks[^2]. For the purposes of the discussion in this paper, the crucial point is that when considering the photon PDF and its corresponding uncertainty at some given \(x\) and \(\mu_F^2\) value, the contributions to this from the input photon distribution, at the starting scale \(Q_0\), and from the DGLAP evolution term due to high scale emission from the quarks, are completely separated; this will greatly simplify the discussion which follows, and allow some fairly simple and robust conclusions to be drawn. ## The input distribution {#sec:input} The photon PDF has been separated in ([\[gampdf\]](#gampdf){reference-type="ref" reference="gampdf"}) into an input component at \(Q_0\) and an evolution component, due to high scale \(q\to q\gamma\) emission. While the latter quantity is given in terms of the generally well determined quark PDFs, the former quantity is on the face of it completely unknown. Thus the uncertainty on the photon distribution at some scale \(\mu_F\) is given quite directly in terms of the uncertainty on the starting distribution \(\gamma^{\rm in}\), and it is this object which we are principally interested in. It is perfectly possible to simply treat this as an unknown quantity in a global fit, i.e. in the same way as the quarks and gluons. This is the approach taken in the latest NNPDFQED fit , where the freely parameterised photon is fitted to DIS and a small set of LHC data, namely \(W,Z\) and high/low--mass Drell--Yan production (more precisely this is achieved by Bayesian reweighting, see  for full details). Due to the generally small contribution from photon--initiated processes, the constraining power of this data is quite limited, and the corresponding PDF uncertainties are large. However, by treating the photon PDF identically to the quark and gluons, a significant part of the available experimental information is in fact being thrown away . The crucial difference is that QED, in contrast to QCD, corresponds to a long range force that does not suffer from the issue of non--perturbativity at low scales. Thus a proton may coherently emit a photon (\(p\to p\gamma\)) at low scale \(Q<Q_0\sim\) 1 GeV, and this will contribute directly to the input component in ([\[gampdf\]](#gampdf){reference-type="ref" reference="gampdf"}). Such a process is experimentally extremely well measured, being governed by the well known electric and magnetic proton form factors for coherent photon emission. In particular we have \[\label{gamcoh} \gamma_{\rm coh}(x,Q_0^2)=\frac{1}{x}\frac{\alpha}{\pi}\int_0^{Q^2<Q_0^2}\!\!\frac{{\rm d}q_t^2 }{q_t^2+x^2 m_p^2}\left(\frac{q_t^2}{q_t^2+x^2 m_p^2}(1-x)F_E(Q^2)+\frac{x^2}{2}F_M(Q^2)\right)\;,\] where \(q_t\) is the transverse momentum of the emitted photon, and \(Q^2\) is the modulus of the photon virtuality, given by \[\label{qmincoh} Q^2=\frac{q_t^2+x^2m_p^2}{1-x}\;,\] The functions \(F_E\) and \(F_M\) are \[\label{form1} F_M(Q^2)=G_M^2(Q^2)\;,\qquad F_E(Q^2)=\frac{4m_p^2 G_E^2(Q^2)+Q^2 G_M^2(Q^2)}{4m_p^2+Q^2}\;,\] with \[\label{form2} G_E^2(Q^2)=\frac{G_M^2(Q^2)}{7.78}=\frac{1}{\left(1+Q^2/0.71 {\rm GeV}^2\right)^4}\;,\] in the dipole approximation, where \(G_E\) and \(G_M\) are the 'Sachs' form factors, which have been very precisely measured in a range of elastic \(ep\) scattering experiments[^3]. As the contribution to \(ep\) scattering for low photon \(Q^2\) is dominantly coherent, we expect ([\[gamcoh\]](#gamcoh){reference-type="ref" reference="gamcoh"}) to give the dominant contribution to the input photon PDF. Thus the input photon distribution is in fact already well determined. More precisely, in general there will also be some contribution from incoherent emission (\(\gamma p \to \gamma X\)), where the proton dissociates after the scattering process. That is, we have \[\label{inputdef} \gamma(x,Q_0^2)=\gamma_{\rm coh}(x,Q_0^2)+\gamma_{\rm incoh}(x,Q_0^2)\;,\] where the second term corresponds to this incoherent input; it is this combined input PDF, including both coherent and incoherent components, which corresponds to the freely parameterised NNPDF distribution described above. In general, as recently discussed in  this incoherent contribution may be constrained from experimental data on \(F_2\) and \(F_L\), however for our considerations it is sufficient to use a simplified model which gives an upper bound on such a contribution. Thus, following  we model this emission process as being due to one photon emission from the valence quarks in the leading--logarithmic approximation; such an approach is also taken in  to model the photon PDFs, although in these cases no coherent component is included. We take[^4] \[\label{gamincoh} \gamma_{\rm incoh}(x,Q_0^2)=\frac{\alpha}{2\pi}\int_x^1\frac{{\rm d}z}{z}\left[\frac{4}{9}u_0\left(\frac{x}{z}\right)+\frac{1}{9}d_0\left(\frac{x}{z}\right)\right]\frac{1+(1-z)^2}{z}\int^{Q_0^2}_{Q^2_{\rm min}}\frac{{\rm d}Q^2}{Q^2+m_q^2}\left(1-G_E^2(Q^2)\right)\;,\] where \[\label{qminincoh} Q^2_{\rm min}=\frac{x}{1-x}\left(m_\Delta^2-(1-x)m_p^2\right)\;,\] accounts for the fact that the lowest proton excitation is the \(\Delta\)--isobar, and the final factor \((1-G_E^2(Q^2))\) corresponds to the probability to have no intact proton in the final state (which is already included in the coherent component). Here \(m_q=m_d\)(\(m_u\)) when convoluted with \(d_0\)(\(u_0\)), and the current quark masses are taken. Crucially, as the quark distributions are frozen for \(Q<Q_0\), this represents an upper bound on the incoherent contribution. If we consider the momentum fraction \[p_\gamma = \int {\rm d}x\, x\gamma(x,Q_0^2)\;,\] carried by the photon at the starting scale \(Q_0^2=2\,{\rm GeV}^2\), then even for this upper bound we find \[\label{pgam} p_\gamma^{\rm coh} =0.15\% \quad\quad p_\gamma^{\rm incoh.}=0.05\%\;,\] that is we expect \(p_\gamma^{\rm incoh.} \ll p_\gamma^{\rm coh.}\), consistent with the general expectation that the emission process for low \(Q^2\) photons should be dominantly coherent. As the coherent input is quite precisely determined, the uncertainty on the input photon PDF is essentially purely due to the incoherent term. Being maximally conservative, we can consider a range of incoherent inputs, with the lower bound being simply setting \(\gamma^{\rm incoh.}(x,Q_0^2)=0\), and the upper bound calculated as described above. Taking this as our uncertainty band, we then expect from ([\[pgam\]](#pgam){reference-type="ref" reference="pgam"}) a \(\sim \pm 10-15\%\) uncertainty on the photon PDF at the starting scale \(Q_0\). On the other hand, at higher scales as the contribution from the evolution term in ([\[gampdf\]](#gampdf){reference-type="ref" reference="gampdf"}) becomes more significant, this uncertainty will be smaller; we will show this explicitly in the following section, where we will see that for phenomenologically relevant scales the uncertainty due to the incoherent input shrinks to \(\sim \pm 5\%\). It should be emphasised that this is a relatively conservative estimate of the uncertainty on the photon PDF due to the incoherent input component. In particular, it is possible and desirable to further constrain this incoherent input in a global fit, while a more complete treatment accounting for example for the (\(\Delta\)\...) resonant contribution to the incoherent input, and more generally constraining this with the existing \(ep\) scattering data will also further reduce this uncertainty. Nonetheless, even within this simplified and conservative approach, we can see that the corresponding uncertainty on the photon PDF is already under relatively good control. In contrast, the NNPDF3.0QED set  gives \[p_{\gamma}=(1.26\pm 1.26)\%\;,\] that is, a \(\sim 100\%\) uncertainty, with a central value which lies much higher than that expected from the simple physical considerations above; we will see the impact of this is in the following sections. # The photon PDF: results {#sec:photonhighx} In the previous section we demonstrated how the photon PDF at the starting scale is already quite precisely determined in terms of the form factors for coherent \(p\to p \gamma\) emission. We will now demonstrate the impact this result has on the photon PDF at higher \(x\) values. We show in Fig. [\[fig:gam2q\]](#fig:gam2q){reference-type="ref" reference="fig:gam2q"} (left) the contributions from the coherent ([\[gamcoh\]](#gamcoh){reference-type="ref" reference="gamcoh"}) and incoherent ([\[gamincoh\]](#gamincoh){reference-type="ref" reference="gamincoh"}) components of the photon PDF at the starting scale \(Q_0^2=2\, {\rm GeV}^2\), where as described above the incoherent term corresponds to an upper bound on this contribution. Here, and in all results which follow, we take treat the evolution ([\[dglap\]](#dglap){reference-type="ref" reference="dglap"}) at NLO in \(\alpha_s\). We make use of the MMHT2014NLO  set for the quark PDFs in the incoherent component, as well as for the PDFs in the evolution of the photon. As expected from the previous discussion, the incoherent component is smaller than the well determined coherent component, and constitutes \(\sim 25\%\) of the total photon PDF at intermediate values of \(x\), consistent with ([\[pgam\]](#pgam){reference-type="ref" reference="pgam"}). However, interestingly the ratio of incoherent to coherent is found to decrease with increasing \(x\), such that in the higher \(x\gtrsim 0.1\) region, the coherent component is particularly dominant. This is due in part to the decreasing phase space for photon emission from the individual quarks, with the range of the \(z\) integral in ([\[gamincoh\]](#gamincoh){reference-type="ref" reference="gamincoh"}) decreasing with increasing \(x\). In addition to this, another physical effect is playing a role, due to the minimum photon virtuality \(Q^2_{\rm min}\), given by ([\[qmincoh\]](#qmincoh){reference-type="ref" reference="qmincoh"}) and ([\[qminincoh\]](#qminincoh){reference-type="ref" reference="qminincoh"}). The kinematic minimum ([\[qmincoh\]](#qmincoh){reference-type="ref" reference="qmincoh"}) follows simply from the on--shellness requirement for the outgoing proton, and similarly in ([\[qminincoh\]](#qminincoh){reference-type="ref" reference="qminincoh"}) for the outgoing \(\Delta\) resonance. In the latter case this corresponds to the mass of the lowest lying resonance above the proton: for higher mass resonance production, the kinematic minimum will be larger still. The effect of this is shown in Fig. [\[fig:gam2q\]](#fig:gam2q){reference-type="ref" reference="fig:gam2q"} (right). Due to the higher mass of the dissociating state in the latter incoherent case, the minimum photon virtuality can be quite large for higher \(x\), with the effect that contribution from low--scale incoherent photon emission becomes kinematically limited; by construction we must have \(Q^2<Q_0^2\sim 1\,{\rm GeV^2}\), while the contribution for photon \(Q^2>Q_0^2\) is given by the evolution component in ([\[gampdf\]](#gampdf){reference-type="ref" reference="gampdf"}), in terms of the relatively well constrained quark PDFs. This effect can be seen in Fig. [\[fig:gam2q\]](#fig:gam2q){reference-type="ref" reference="fig:gam2q"} (left) in the high \(x\) region, where the coherent component becomes completely dominant. In Fig. [\[fig:pdfcomp\]](#fig:pdfcomp){reference-type="ref" reference="fig:pdfcomp"} (right) we show the photon PDF at \(\mu_F=2\) TeV, corresponding to \(x\sim 0.2\) at the LHC, with the contributions from the input coherent and incoherent, and evolution components shown explicitly. We also show in Fig. [\[fig:pdfcomp\]](#fig:pdfcomp){reference-type="ref" reference="fig:pdfcomp"} (left) the corresponding PDF for the lower scale \(\mu_F=10\) GeV, to give an indication of the influence of the photon evolution on these different components. We can see that the effect of evolution is as expected to further decrease the contribution from the incoherent input, which is already \(\lesssim 10\%\) of the total photon at \(\mu_F=100\) GeV, and even lower for \(\mu_F=2\) TeV. Thus in the cross section results which follow we expect a conservative \(\sim \pm 5\%\) uncertainty due to this effect. The coherent contribution, even at the higher scale \(\mu_F=2\) TeV is \(\sim 20\%\) at \(x\sim 0.01\), and increases to \(\sim 50\%\) at higher \(x\); for \(\mu_F=100\) GeV it is larger still. The `LUXqed` prediction is shown and is found to coincide quite closely with our result, although some deviation is visible, in particular at higher \(x\). The NNPDF3.0 distribution, with the corresponding \(68\%\) C.L. uncertainty bands, is also shown in Fig. [\[fig:pdfcomp\]](#fig:pdfcomp){reference-type="ref" reference="fig:pdfcomp"}: here, and in all results which follow, we take the NLO set with \(\alpha_s(M_Z)=0.118\). For the lower \(x\) region the contribution from the high scale \(q\to q\gamma\) evolution component in ([\[gampdf\]](#gampdf){reference-type="ref" reference="gampdf"}) is dominant, and as a result the corresponding uncertainties are under reasonable control[^5]. As \(x\) increases, however, the phase space for the DGLAP \(q\to q\gamma\) emission process decreases, and the contribution from the coherent photon input becomes more important. This effect is evident in the NNPDF set, where the increasing contribution from the poorly determined input photon leads to a rapidly increasing uncertainty as \(x\) increases. In Fig. [\[fig:pdflumi\]](#fig:pdflumi){reference-type="ref" reference="fig:pdflumi"} we show the corresponding PDF luminosities, defined as \[\label{lumi} \frac{{\rm d}\mathcal{L}_{ij}}{{\rm d} \ln M_X^2}= \frac{M_X^2}{s}\int^1_{\tau} \frac{{\rm d}x_1}{x_1} \,f_i(x_1,M_X^2)f_j(\tau/x_1,M_X^2)\;,\] where \(\tau=M_X^2/s\) and \(f_i\) is the corresponding PDF for parton \(i\). As well as the \(\gamma\gamma\) case discussed above, we also show for comparison the \(qq\), \(q\overline{q}\) (defined in both cases as a uniform sum over the 5 corresponding quark flavours) and \(gg\) cases, using the same NNPDF set. For our prediction, we now for illustration include an uncertainty band due to varying the incoherent component between \(x\gamma(x,Q_0)=0\) and the upper bound of ([\[gamincoh\]](#gamincoh){reference-type="ref" reference="gamincoh"}), although in the plots this is essentially invisible within the width of the central curves. Other uncertainties, due for example to the quark (and at higher orders, gluon) PDFs entering the photon evolution in ([\[dglap\]](#dglap){reference-type="ref" reference="dglap"}), the use of the dipole approximation ([\[form2\]](#form2){reference-type="ref" reference="form2"}) for the elastic form factor and the choice of \(Q_0\) in ([\[gampdf\]](#gampdf){reference-type="ref" reference="gampdf"}) are not included here. These effects are expected to be generally subleading in comparison to that due to the incoherent input, and will be omitted in the results which follow. Nonetheless, it is worth bearing in mind that the effect of these will be to increase the total uncertainty on the photon PDF somewhat, which should be accounted for in a complete analysis; for the current purposes, however, this is not necessary. The `LUXqed` prediction is shown and is again found to coincide quite closely with our result, with some deviation at higher \(M_X\). The same increase in Fig. [\[fig:pdflumi\]](#fig:pdflumi){reference-type="ref" reference="fig:pdflumi"} in the NNPDF uncertainty band at high \(M_X\) for the \(\gamma\gamma\) case is clear. However, interestingly we can see that the trend in the central value of the NNPDF \(\gamma\gamma\) luminosity is remarkably different compared to the other partons, with the former decreasing much less rapidly at high \(M_X\), i.e. high \(x\). On the other hand, our prediction shows no such significant difference, and roughy follows the same trend as in the quarks. As discussed in  some steepening of the PDF luminosities for the QCD partons may be expected due to the differing behaviours of \(\alpha_{\rm QED}\) and \(\alpha_s\) at higher scales. However this effect, which is indeed observable in particular upon comparison of our result for the \(\gamma\gamma\) and the \(gg\) luminosity, is relatively small and cannot explain the difference seen in the NNPDF case. We are therefore led to conclude that this potentially significant difference is an artefact of the large uncertainties in the NNPDF photon PDF; the physically motivated photon PDF of our approach, which lies towards the lower end of the NNPDF uncertainty band, displays no significant difference in behaviour at higher \(x\) compared to the quarks and gluons. It is therefore in this higher \(x\) region that the importance of including all available information about the photon PDF is clearest; by excluding the additional input which comes from considering the physics of the dominantly coherent photon emission process at the starting scale \(Q_0\), the corresponding PDF uncertainties are dramatically over--inflated. By including this information, as in Section [2.2](#sec:input){reference-type="ref" reference="sec:input"}, the predicted photon PDF at higher \(x\) is determined quite precisely to lie close to the lower edge of the NNPDF uncertainty band. It has for example been pointed out in  that the upper limits on the NNPDF photon PDF predict potentially sizeable photon--initiated contributions to the Drell--Yan cross section at high mass, with large corresponding PDF uncertainties. From the above considerations, however, we expect this not to be the case: we will consider this in more detail in the following section. # Cross section predictions {#sec:dy} As discussed in the introduction, the photon--initiated contribution to lepton and \(W\) boson pair production may be particularly significant at higher mass, where the production cross sections are relatively enhanced due to the \(t\)--channel nature of the corresponding Feynman diagrams. We therefore consider predictions for both these processes at the LHC and FCC in this section. We use our own implementation of these processes, with the corresponding LO cross sections given by (see e.g. ) \[\begin{aligned} \nonumber \frac{{\rm d}\sigma}{{\rm d}\cos\theta^*}(\gamma\gamma\to W^+W^-) &= \frac{\pi\alpha^2\beta}{\hat{s}}\frac{19-6\beta^2(1-\beta^2)+2(8-3\beta^2)\beta^2\cos^2\theta^*+3\beta^4\cos^4\theta^*}{(1-\beta^2\cos^2\theta^*)^2} \,, \\ \frac{{\rm d}\sigma}{{\rm d}\cos\theta^*}(\gamma\gamma\to l^+l^-) &= \frac{2\pi\alpha^2\beta}{\hat{s}}\frac{1+2\beta^2(1-\beta^2)(1-\cos^2\theta^*)-\beta^4\cos^4\theta^*}{(1-\beta^2\cos^2\theta^*)^2}\,, \end{aligned}\] where \(\beta=(1-4m^2/\hat{s})^{1/2}\), with \(m=m_{W},m_{l}\), and \(\theta^*\) is the angle of the outgoing particles with respect to the photons in the \(\gamma\gamma\) C.M. frame. As we are interested in the high mass regime, in the lepton case we will for concreteness take the massless limit, \(m_l=0\), in what follows. In Fig. [\[fig:cross13\]](#fig:cross13){reference-type="ref" reference="fig:cross13"} we show the lepton pair production cross section via the photon--initiated production and Drell--Yan production mechanisms, for lepton \(|\eta|<2.5\) and \(p_\perp> 20\) GeV. The former is shown using both the approach of Section [2.2](#sec:input){reference-type="ref" reference="sec:input"} and with the NNPDF3.0QED set, while the latter is calculated at NLO in \(\alpha_s\) with `MCFM`  using MMHT2014NLO  PDFs; the results which follow are not affected significantly by NNLO corrections. For the curve corresponding to our approach, we show an uncertainty band due to varying the incoherent component between \(x\gamma(x,Q_0)=0\) and the upper bound of ([\[gamincoh\]](#gamincoh){reference-type="ref" reference="gamincoh"}). The 68% PDF uncertainty bands are shown for the `MCFM` predictions in all cases which follow, although for certain distributions this is sufficiently small that it is not visible on the plots. For the LHC predictions in Fig. [\[fig:cross13\]](#fig:cross13){reference-type="ref" reference="fig:cross13"} (left), we can see that for higher \(M_{ll}\) the photon--initiated cross section predicted by the NNPDF set may be comparable in size and even larger than the Drell--Yan cross section, within the increasingly large PDF uncertainty bands. This was recently discussed in , where it was pointed out that the potential dominance of the photon--initiated NNPDF prediction induced a large uncertainty in the predicted cross section for high mass lepton pair production; this could, for example, have an impact on searches for new heavy particles decaying to lepton pairs. However, it is our finding that this is not the case. In particular, we can see from Fig. [\[fig:cross13\]](#fig:cross13){reference-type="ref" reference="fig:cross13"} that even up to the highest \(M_{ll}\) values the predicted contribution from the photon--initiated process is fairly small, \(\sim 10\%\) of the Drell--Yan. This result is entirely consistent with the expectations from Fig. [\[fig:pdfcomp\]](#fig:pdfcomp){reference-type="ref" reference="fig:pdfcomp"}. Thus we expect no significant contamination from the photon--initiated process. For the FCC case shown in Fig. [\[fig:cross13\]](#fig:cross13){reference-type="ref" reference="fig:cross13"} (right), which was recently discussed in , a similar trend is seen. Moreover, it is worth emphasising that for both the LHC and FCC cases, tighter cuts on the lepton transverse momentum \(p_\perp\) and pseudorapidity \(\eta\) will further decrease the relative contribution from the photon--initiated process, which being due to the \(t\) and \(u\) channel diagrams is more strongly peaked towards low \(p_\perp\) and high \(\eta\). In Fig. [\[fig:crossw\]](#fig:crossw){reference-type="ref" reference="fig:crossw"} we show predictions for the \(W\) boson pair production cross sections, again at the LHC and FCC. We impose the same cuts on the \(W\) boson pseudorapidities, and include no further decays, as in , for the sake of comparison. `MCFM`  with MMHT2014NLO  PDFs is used to generate the the QCD \(WW\) production process at NLO in \(\alpha_s\), with the \(gg\)--initiated box contribution also included. Again a similar trend is clear, with the NNPDF set predicting potentially a completely dominant photon--initiated contribution at higher masses, within very large uncertainties. However, for the LHC our approach predicts instead that the standard QCD--initiated is dominant, apart from at the very highest masses. On the other hand for the FCC this is no longer the case: over the mass range considered the \(\gamma\gamma\) and QCD--initiated contributions are generally expected to be comparable in size. In this case a very careful treatment of the photon PDF uncertainties will be essential. It should be emphasised that the predicted cross sections within our approach lie entirely within the NNPDF uncertainty bands, and are therefore completely consistent with these. The issue is simply that the NNPDF approach, by omitting the physical constraints on the photon input described in Section [2.2](#sec:input){reference-type="ref" reference="sec:input"}, allows in principle unphysically large input photon distributions, the effect of which becomes increasingly dominant at higher \(x\), where the contribution from the DGLAP \(q\to q\gamma\) emission from the quarks becomes smaller. On the other hand, the NNPDF starting distribution effectively parameterises the contribution from both the coherent and incoherent input components as in ([\[inputdef\]](#inputdef){reference-type="ref" reference="inputdef"}), but without any further constraints, and therefore we fully expect consistency within both approaches, once the PDF uncertainties have been properly included; it is encouraging to find that this is indeed the case. Moreover, as further data from the LHC is included in the NNPDF fit, we fully expect this consistency to continue as the PDF uncertainties decrease. Indeed, this expectation is supported by the recent ATLAS measurement  of high mass lepton pair production in the \(116<M_{ll}<1500\) GeV region at \(\sqrt{s}=8\) TeV, corresponding to \(0.015\lesssim x \lesssim 0.2\). We show the predicted cross section for this mass region in Fig. [\[fig:cross8\]](#fig:cross8){reference-type="ref" reference="fig:cross8"}, and we can see that for larger \(M_{ll}\) quite significant photon--initiated contributions are allowed within the NNPDF uncertainty bands. However, in  a Bayesian reweighting exercise is performed, and it is found that a photon PDF which lies systematically on the lower end of the NNPDF2.3 uncertainty band is preferred, with greatly reduced uncertainties, see in particular Fig. 13 of [^6]. While a full refit would be required to make a completely firm conclusion, this nonetheless provides a strong indication that the larger photon--initiated cross section predictions allowed by the higher \(x\) NNPDF photon uncertainty band are already disfavoured by this ATLAS data, while our result, which predict lower photon--initiated cross sections, are qualitatively consistent. Further experimental support for this result at the LHC can be found in the ATLAS  and CMS  measurements of exclusive two--photon induced \(W^+W^-\) production. Here, events are selected by demanding a dilepton vertex with no additional associated charged tracks within the tracker acceptance (\(\eta<\) 2.4). As part of these measurements, a larger data sample of \(\mu^+\mu^-\) events is also selected with the same track veto applied. Crucially, such a veto effectively isolates the photon--initated contribution; the standard Drell--Yan production process will dominantly produce additional tracks centrally, and the remaining contribution can be suppressed with further cuts and then subtracted using MC simulation. The photon--initiated cross section for this semi--exclusive case was considered in , where it was demonstrated how ([\[gampdf\]](#gampdf){reference-type="ref" reference="gampdf"}) could be relatively simply modified to account for such a rapidity gap veto. Predictions for semi--exclusive \(\mu^+\mu^-\) production within the ATLAS/CMS event selection were presented, and it was shown that the data are completely consistent with the approach described in this paper, but are in strong tension with the higher cross sections allowed by the NNPDF2.3 set. # Conclusion {#sec:conc} Processes with initial--state photons are becoming increasingly relevant for phenomenology at the LHC, in particular as the requirement for high precision becomes standard. Given this, it is becoming increasingly important to constrain as precisely as possible the corresponding photon PDF. This is particularly so for processes such as high mass lepton and \(W\) boson production, where the contribution from the photon--initiated process may be significant. In this paper we have described how the photon PDF may we be quite precisely constrained by relatively simple considerations about the nature of the initial--state photon, see . In particular, QED, in contrast to QCD, corresponds to a long range force that does not suffer from the issue of non--perturbativity at low scales, and thus a proton may coherently emit a photon (\(p\to p\gamma\)) at low scales. This process is already experimentally very well measured in \(ep\) collisions, and is governed by the well known electric and magnetic proton form factors for coherent photon emission. Taking these as input, we have shown how the photon PDF can already be well constrained with even a quite conservative model estimation for an upper limit on the remaining incoherent contribution to the photon at the starting scale: the uncertainty due to this is found to be generally \(\sim 10\%\) or less, with the precise value depending on the scale and parton \(x\). A more detailed treatment of the (resonant and non--resonant) incoherent contribution, and the inclusion of further constraints within the context of a global fit will reduce this even further. Indeed, as discussed in the recent analysis of , it is possible to place quite precise constraints by using low \(Q^2\) inelastic structure function data. In this work, a dominantly coherent component is also included, and we have seen that the resultant photon PDF lies quite close to our prediction. These constraints are not included in the currently available global fits  and may lead to a signifiant overestimation in the uncertainty in the photon PDF. This can have important phenomenological implications, and indeed this is the case for high mass lepton and \(W\) boson pair production, which have recently been discussed in . Here, it was found that the NNPDF2.3QED  set, which freely parameterises the input photon and fits to a selection of DIS and LHC data, is consistent with strongly dominant photon--initiated contributions to both of these processes at higher masses, at the LHC and FCC. In this paper we have shown that this is not expected to be the case. We have found that the DY contribution to lepton pair production is dominant out to very high masses, while for \(W\) boson pair production the QCD process is generally larger, although for higher masses the \(\gamma\gamma\)--initiated contribution is comparable in size. In this latter case, therefore, a precise treatment of the photon PDF will be essential.
{'timestamp': '2016-10-11T02:10:01', 'yymm': '1607', 'arxiv_id': '1607.04635', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04635'}
# Introduction Jovian Trojans are actually large populations of minor bodies in the 1:1 mean motion resonance (MMR) with Jupiter, librating around \(L_4\) and \(L_5\) Lagrangian points. In general, there are two classes of theories explaining their origin: i) a theory in the framework of accretion model (e.g. Goldreich 2004, Lyra et al. 2009) and ii) a capture of bodies located in libration zones during a migration of giant planets (Morbidelli et al. 2005, Morbidelli et al. 2010, Nesvorný et al. 2013), which is preferred in our solar system. Since the librating regions are very stable in the current configuration of planets and they are surrounded by strongly chaotic separatrices, bodies from other source regions (e.g. Main belt, Centaurs, Jupiter family comets) cannot otherwise enter the libration zones and Jupiter Trojans thus represent a rather primitive and isolated population. Several recent analyses confirmed the presence of several families among Trojans (e.g. Nesvorný et al. 2015, Vinogradova, 2015). The Trojan region as such is very favourable for dynamical studies of asteroid families, because there is no significant systematic Yarkovsky drift in semimajor axis due to the resonant dynamics. On the other hand, we have to be aware of boundaries of the libration zone, because ballistic transport can cause a partial depletion of family members. At the same time, as we have already shown in Brož & Rozehnal (2011), no family can survive either late phases of a slow migration of Jupiter, or Jupiter "jump", that results from relevant scenarios of the Nice model (Morbidelli et al. 2010). We thus focus on post-migration phase in this paper. We feel the need to evaluate again our previous conclusions on even larger datasets, that should also allow us to reveal as-of-yet unknown structures in the space of proper elements or unveil possible relations between orbital and physical properties (e.g. albedos, colours, diameters) of Jovian Trojans. In Section [2](#sec:obsdata){reference-type="ref" reference="sec:obsdata"} we use new observational data to compute appropriate resonant elements. In Section [3](#sec:physchar){reference-type="ref" reference="sec:physchar"} we use albedos obtained by Grav et al. (2012) to derive size-frequency distributions and distribution of albedos, which seem to be slightly dependent on the proper inclination \(I_{\rm p}\). In Section [4](#sec:groups){reference-type="ref" reference="sec:groups"} we identify families among Trojans with our new "randombox" method. We discuss properties of statistically significant families in Section [5](#sec:props){reference-type="ref" reference="sec:props"}. Then we focus mainly on the Hektor family because of its unique D--type taxonomical classification, which is the first of its kind. We also discuss its long-term dynamical evolution. In Section [6](#sec:coll_model){reference-type="ref" reference="sec:coll_model"} we simulate collisional evolution of Trojans and we estimate the number of observable families among Trojans. Finally, in Section [7](#sec:SPH){reference-type="ref" reference="sec:SPH"} we simulate an origin of the Hektor family using smoothed-particle hydrodynamics and we compare results for single and bilobed targets. Section [8](#sec:conclusions){reference-type="ref" reference="sec:conclusions"} is devoted to Conclusions. # New observational data {#sec:obsdata} ## Resonant elements We computed resonant elements, i.e. the averaged semimajor axis \(\bar{a}\), libration amplitude \(\Delta a_{\rm p}\), eccentricity \(e_{\rm p}\) and inclination \(I_{\rm p}\) of 3907 Trojans in \(L_4\) cloud and 1945 Trojans in \(L_5\) cloud. As an input, we used osculating elements listed in AstOrb catalogue (Bowell et al. 2002), released in July 2014. A detailed description of the resonant elements computation can be found in Brož & Rozehnal (2011). Positions of Trojans in the space of proper elements (\(a_{\rm p}\), \(I_{\rm p}\)), where \(a_{\rm p} = \bar{a} + \Delta a_{\rm p}\), and (\(e_{\rm p}\), \(I_{\rm p}\)), calculated with a suitably modified version of the SWIFT integrator (Levison & Duncan, 1994), are presented graphically in Figure [\[trojans_L4_arer_sizes\]](#trojans_L4_arer_sizes){reference-type="ref" reference="trojans_L4_arer_sizes"}, together with their sizes and albedos. [^1] ## WISE and AKARI albedos and diameters To construct size-frequency distributions of the whole \(L_4\) and \(L_5\) Trojan populations and later of individual families, we mostly used WISE albedos and diameters derived by Grav et al. (2012). We also compared the respective values to AKARI, as reported by Usui et al. (2011).[^2] We used albedo values of 1609 Trojans in both \(L_4\) and \(L_5\) clouds obtained by WISE; about one third of these albedos were obtained during cryo-phase, the rest were measured in post-cryo-phase (see Grav et al., 2011). # Physical characterisation of Trojan populations {#sec:physchar} ## Albedo distribution and taxonomy The values of visible albedos \(p_{\rm V}\) of Trojans derived by Grav et al. (2012) vary in the range from \(p_{\rm V} = 0.025\) to \(p_{\rm V} \simeq 0.2\). Distributions of albedos are qualitatively the same for both \(L_4\) and \(L_5\) populations. The median albedo of WISE sample is \(\widetilde{p}_{\rm v}=0.072 \pm 0.017\) for \(L_4\) and \(\widetilde{p}_{\rm v}=0.069 \pm 0.015\) for \(L_5\). These values of visible albedos mostly correspond to C or D taxonomical classes in Tholen taxonomic classification scheme (Mainzer et al., 2011). However, there is a significant presence of small asteroids \((D<15\,\rm km)\) with apparently high albedo---almost 20 % of asteroids in \(L_4\) and 13 % of asteroids in \(L_5\) have albedo \(p_{\rm V} > 0.10\). As stated in Grav et al. (2012), this is probably not a physical phenomenon, it is rather due to the fact that for small diameters the photon noise contribution becomes too significant. When we compute the median albedo from AKARI data, we realize that its value is slightly lower (\(\widetilde{p}_{\rm v}=0.054 \pm 0.005\)) than that from WISE, but when we compute the median from WISE values for the same asteroids which are listed in AKARI catalogue, we obtain a similar value (\(\widetilde{p}_{\rm v}=0.061 \pm 0.012\)). What is more serious, AKARI and WISE data differ considerably for large asteroids with \(D>100\,\rm km\)---the average difference between albedos is \(|p_{\rm V_{AKARI}}-p_{\rm V_{WISE}}| = 0.02\). The same difference we see in derived diameters. These discrepancies may be caused for example by limitations of the thermal model (cf. spheres in NEATM models). Hereinafter, we prefer to use the WISE data when available, because they represent orders of magnitude larger sample than AKARI. When we split Trojan asteroids according to their albedo into two rather artificial subpopulations with \(p_{\rm V} < 0.08\) and \(p_{\rm V} > 0.08\) respectively, and then we compute distributions of these subpopulations with respect to the resonant inclination \(I_{\rm p}\), we get two different pictures. As can be seen in Figure [\[albedo_separation_L4_hist\]](#albedo_separation_L4_hist){reference-type="ref" reference="albedo_separation_L4_hist"}, most bodies have resonant inclinations \(I_{\rm p} < 15^{\circ}\), but there are 77 % of bodies with higher albedo with \(I_{\rm p} < 15^{\circ}\), while only 55 % of the population with lower albedo is located in the same range of inclinations. This is a similar phenomenon as described by Vinogradova (2015), who reported different upper limits in inclinations for different taxonomical types obtained mostly from SDSS colour data. ## Size-frequency distributions {#sec:SFDs} The WISE data (Grav et al. 2011, 2012) provide very useful source of information on diameters we need to construct size-frequency distributions (SFDs) of Trojan populations in \(L_4\) and \(L_5\). However, the sample measured by WISE is not complete. In our previous work (Brož and Rozehnal, 2011), we constructed the SFDs assuming a constant albedo which we set to be equal to the median albedo of Trojans that was measured back then. Since the number of measurements was very low (several tens), this was the only reasonable way. Now we choose another method to construct more reliable SFDs. As we calculated resonant elements for more than 5800 Trojans and we have more than one quarter of appropriate albedos, we constructed the SFDs by assigning albedos randomly from the observed WISE distribution to the remaining Trojans, whose albedo was not measured. To avoid a bias, we compared different SFDs constructed with different random generator seeds and we realized that the overall shape of SFDs does not change noticeably, the slope \(\gamma\) varies in the range of \(\pm 0.1\) at most. The SFDs we constructed this way are shown in Figure [\[SFD_WISE\]](#SFD_WISE){reference-type="ref" reference="SFD_WISE"}. The SFDs for the \(L_4\) and \(L_5\) clouds look slightly different, especially in the size range from 60 km to 100 km. This part of the SFD is not influenced by the Eurybates family, the largest family among Trojans, because all its members have diameters \(D<50\,\rm{km}\). We used these SFDs to determine the ratio of the number of asteroids in \(L_4\) and \(L_5\) clouds. There are 2746 asteroids with diameter \(D>8\,\rm km\) in \(L_4\) and 1518 asteroids in \(L_5\). When we remove all family members with diameters \(D>8\,\rm km\), we have 2436 asteroids in \(L_4\) and 1399 in \(L_5\). However, this sample may be still influenced by debris produced by catastrophic disruptions of small bodies (\(D\geq 50\,\rm km\)), which need not to be seen as families. Counting only asteroids with diameter \(D>20\,\rm km\), which corresponds to the absolute magnitude \(H\simeq12\), and removing family members, we get the ratio \(N_{L_4}/N_{L_5}=1.3 \pm 0.1\). As this is entirely consistent with value of Nesvorný et al. (2013), which was derived for Trojans with \(H>12\), and with Grav et al. (2012), whose estimate is \(N_{L_4}/N_{L_5} = 1.4 \pm 0.2\), we can confirm a persisting asymmetry between the number of \(L_4\) and \(L_5\) Trojans in new data. Although for bodies with diameter \(D>100\,\rm km\), the \(L_5\) cloud has more asteroids than \(L_4\), the total number of these bodies is of the order of 10, so this is just an effect of small-number statistics and does not affect the \(N_{L_4}/N_{L_5}\) ratio much. # Families detection methods {#sec:groups} A brief inspection of the resonant-element space (\(a_{\rm p}\), \(e_{\rm p}\), \(I_{\rm p}\)) (see Figure [\[trojans_L4_arer_sizes\]](#trojans_L4_arer_sizes){reference-type="ref" reference="trojans_L4_arer_sizes"}), reveals several locations with higher concentrations of bodies. These could be collisional families, created by a disruption of a parent body during a random collision, but they could also originate randomly by chaotic diffusion and due to effects of secular and high-order resonances. To be regarded as a family, the cluster must comply with, inter alia, the following criteria: i) it must be concentrated in the space of proper elements; ii) the cluster must have the SFD different from that of the whole \(L_4\) and \(L_5\) population; iii) the last criterion is usually spectral, or at least, albedo homogeneity of family members, but so far, there are not enough sufficiently accurate data for Trojans, especially for bodies with diameters \(D< 50\,\rm km\), which usually form a substantial part of Trojan families. Therefore we cannot perform any detailed spectral analysis in this work. We analyzed the space of resonant elements both in terms of mutual distances among bodies and in terms of statistical probability that clusters are *not* random. ## Randombox method Besides the commonly used hierarchical clustering method (HCM, Zappalà et al., 1990), we applied a "randombox" method, based on numerical Monte-Carlo simulations. This method allows us to compute the statistical significance of the clusters, i.e. the probability that the cluster is a random concentration of bodies in the space of proper elements (\(a_{\rm p}, e_{\rm p}, \sin{I_{\rm p}}\)). We divided the space of proper elements into equally sized "boxes" with dimensions \(\Delta a_{\rm p} = 0.025\, \rm au\), \(\Delta e_{\rm p} = 0.2\) and \(\Delta \sin{I_{\rm p}} = 0.025\). Then we created \(N = 100,000\) random distributions of the same number of bodies which are observed together in the given box and two adjacent boxes (in the direction of the \(y\)-axis, cf. Figure [\[Randombox\]](#Randombox){reference-type="ref" reference="Randombox"}), and we counted number of positive trials \(N^+\), for which the randomly generated number of bodies in the central box was larger than the observed one. From here we can calculate the probability \(P_{\rm rnd}\), that the observed number of bodies in the box is random: \(P_{\rm rnd} = N^+ / N\). Alternatively, one can also use our analytical formula: \[\label{randombox_p} p_{\rm rnd} = \frac{\sum _{k=n_2} ^n C(n,k)V^\prime (n_{\rm box}-1,n-k)}{V^\prime(n_{\rm box},n)}\,,\] where \(n\) denotes the total number of bodies, \(n_{\rm box}\) is the total number of boxes (3 in our case), \(n_2\) is the observed number of bodies in the middle box, \(k\) is the number of observed bodies in the current box, \(C(n,k)\) are combinations without repetitions, i.e. the total number of trials to select \(k\) bodies observed in the current box from the total number of \(n\) bodies; \(V^\prime (n_{\rm box}-1,n-k)\) are variations with repetitions, i.e. the total number of trials to distribute the remaining bodies into the remaining boxes; and \(V^\prime(n_{\rm box},n)\) are also variations with repetitions, i.e. the total number of trials to distribute all \(n\) bodies into all \(n_{\rm box}\) boxes. We verified the results of the analytical formula ([\[randombox_p\]](#randombox_p){reference-type="ref" reference="randombox_p"}) by the MC method. We plot the results in Figure [\[Randombox\]](#Randombox){reference-type="ref" reference="Randombox"} for both the \(L_4\) and \(L_5\) clouds. In comparison with Figure [\[trojans_L4_arer_sizes\]](#trojans_L4_arer_sizes){reference-type="ref" reference="trojans_L4_arer_sizes"}, one can see that for all clusters we identified as families the probability \(P_{\rm rnd}\) varies between \(2\cdot 10^{-3}\) and \(5\cdot 10^{-5}\), i.e. the probability that clusters are random fluctuations is indeed very low. We also re-evaluated all families identified by the hierarchical clustering method using the "randombox" method, which makes our decision whether the cluster is a real family much more quantitative. ## Hierarchical clustering method {#sec:HCM} We also used the HCM independently to extract significant clusters. Families identified by both the "randombox" and HCM methods are listed in Table [\[fams_list\]](#fams_list){reference-type="ref" reference="fams_list"}. For each family, we constructed a dependence of the number of members of the cluster \(N_{\rm{memb}}\) on the cutoff velocity \(v_{\rm{cutoff}}\). Because the number of members of a real collisional family rises first slowly with rising \(v_{\rm{cutoff}}\) (Brož and Rozehnal, 2011) ---in contrast with random clusters which are merging very quickly with the background---the constructed dependence allows us to guess a realistic number of family members \(N_{\rm{memb}}\). For all families listed in Table [\[fams_list\]](#fams_list){reference-type="ref" reference="fams_list"} we were convinced that they fulfill this criterion. However, we cannot distinguish possible interlopers this way, and it is also possible that some fraction of family members with high \(v_{\rm{cutoff}}\) (so called halo, as in Brož and Morbidelli, 2013) remains unidentified in the surrounding background. # Properties of statistically significant families {#sec:props} ## Eurybates As we have already demonstrated in Brož and Rozehnal (2011), the family associated with asteroid (3548) Eurybates is the largest collisional family, and it is the only family among Trojans with the parent body size \(D_{\rm{PB}} > 100\,\rm{km}\), which originated by a catastrophic disruption (this means that the mass ratio of the largest remnant to the parent body \(M_{\rm{LR}}/M_{\rm{PB}}< 0.5\)). Using new albedos derived by Grav et al. (2012), we recalculated the overall SFD slope of the family to be \(\gamma =-3.4 \pm 0.1\). As the WISE sample provides albedos for only about 1/5 of the family members, we calculated two values of \(\gamma\): the first one assuming that remaining asteroids have a constant albedo \(p_{\rm V} = 0.06\), the second one by assigning albedos randomly from the WISE distribution, as described in Section [3.2](#sec:SFDs){reference-type="ref" reference="sec:SFDs"}. Both values are equal within their errorbars. The new slope \(\gamma\) is significantly steeper than our previous calculation (\(\gamma =-2.5 \pm 0.1\)), derived with the assumption of a constant albedo of all members of the family. The lower value was most likely caused by a significant observational incompleteness in the size range from \(D=12\,\rm km\) to \(D=30\,\rm km\). We also derived the new value of the parent body diameter, which is still above the limit of 100 km. An extrapolation of the SFD by a power law gives the value \(D_{\rm{PB}}\simeq 140\,\rm{km}\). By fitting the synthetic SFDs from SPH simulations (Durda et al., 2007), we obtained the value \(D_{\rm{PB(SPH)}}\simeq155\,\rm{km}\). ## Hektor---the first D--type family {#subsec:Hektor} Since asteroid (624) Hektor is a close binary with a satellite (Marchis et al. 2014), i.e. an exceptional object, we want to address its association with the family. The cluster around the largest Trojan asteroid appears in the space of proper elements as a relatively compact group, which is limited particularly in proper inclinations, \(I_{\rm{p}}\in\langle 18.13^\circ;19.77^\circ\rangle\), and with resonant semimajor axes located in the interval \(a_{\rm{p}}\in\langle 5.234;5.336\rangle\) au. The number of members of this group slowly increases with increasing cutoff velocity up to \(v_{\rm{cutoff}} \simeq 110\,\rm{m}\,\rm{s^{-1}}\), above which it quickly joins the background. With our randombox method, we estimated the probability that the family is just a random fluke to be as low as \(P_{\rm{rnd}}\simeq2\cdot10^{-3}\). The nominal diameter of asteroid (624) Hektor derived from its albedo is 164 km (Grav et al., 2012), but the albedo measured by AKARI \(p_V = 0.034 \pm 0.001\) (Usui et al., 2011) totally differs from that measured by WISE, \(p_V = 0.087 \pm 0.016\). and these values do not match even within the error limits. This may be caused by applying a thermal model assuming spheres to the bilobed shape of the asteroid (Marchis et al., 2014). We hence do not determine Hektor's diameter from its albedo, but from fits of Marchis et al. (2014), which effective value \(D = (250 \pm 26)\,\rm km\) is suitable within its uncertainty for all possible geometries (convex, bilobe and binary). For other bodies in family we use a nominal value \(p_{\rm V}=0.072\), which is the median of WISE measurements. Asteroid (624) Hektor is often classified as D-type (e.g. Cruikshank et al., 2001, Emery et al., 2006, Emery et al., 2011). We tried to evaluate taxonomical classification of other family members and we have found colours for two more expected family members in SDSS-MOC vers. 4 (Ivezic et al., 2002): asteroids (65000) 2002 AV63 and (163702) 2003 FR72. Even though the photometric noise in individual bands is not negligible (\(\sigma_i = 0.02\,{\rm mag}\) up to \(\sigma_u = 0.12\,{\rm mag}\)) both of them are D-types, with principal components (aka slopes) \({\rm PC}_1 > 0.3\). This seems to support the D-type classification of the whole family. We also tried to constrain the taxonomic classification of the family members by comparing their infrared albedos \(p_{\rm IR}\) and visual albedos \(p_{\rm V}\) as described in Mainzer et al. (2011), but there are no data for family members in the W1 or W2 band of the WISE sample, which are dominated by reflected radiation. The fact that we observe a collisional family associated with a D--type asteroid is the main reason we use word "exceptional" in connection with the Hektor family. As we claimed in Brož et al. (2013), in all regions containing a mixture of C--type and D--type asteroids (e.g. Trojans, Hildas, Cybeles), there have been only C--type families observed so far, which could indicate that disruptions of D--type asteroids leave no family behind, as suggested by Levison et al. (2009). Nevertheless, our classification of the Hektor family as D--type is not in direct contradiction with this conclusion, because Levison et al. (2009) were concerned with catastrophic disruptions, while we conclude below that the Hektor family originated from a cratering event, i.e. by an impactor with kinetic energy too small to disrupt the parent body. ### Simulations of long-term dynamical evolution {#sub:dyn_evol} To get an upper limit of the age of the Hektor family, we simulated a long-term evolution of *seven synthetic families* created for different breakup geometries. Our model included four giant planets on current orbits, integrated by the symplectic integrator SWIFT (Levison and Duncan, 1994), modified according to Laskar and Robutel (2001), with the time step of \(\Delta t = 91\,\rm days\) and time span 4 Gyr. We also accounted for the Yarkovsky effect in our simulations. Although in a first-order theory, it is not effective in zero-order resonances (it could just shift libration centre, but there is no systematic drift in semimajor axis) and the observed evolution of proper elements is mainly due to chaotic diffusion, in higher-order theories the Yarkovsky effect can play some role. In our model, we assumed a random distribution of spins and rotation periods (typically several hours), the bulk and surface density \(\rho_{\rm{bulk}} = \rho_{\rm{surf}} = 1.3\,\,\rm{g}\,\rm{cm^{-3}}\), the thermal conductivity \(K = 0.01\,\,\rm{W}\,\rm{m^{-1}}\,\rm{K^{-1}}\), the specific heat capacity \(C = 680\,\,\rm{J}\,\rm{kg^{-1}}\,\rm{K^{-1}}\), the Bond albedo \(A_{\rm B} = 0.02\) and the IR emissivity \(\epsilon = 0.95\). We created each synthetic family by assigning random velocities to 234 bodies (i.e. 3 times more than the number of the observed family members), assuming an isotropic velocity field with a typical velocity of \(70\,\,\rm{m}\,\rm{s^{-1}}\), corresponding to the escape velocity from parent body (Farinella et al., 1993). Here we assumed the velocity of fragments to be size independent. Possible trends in the ejection velocity field cannot be easily revealed in the (\(a, H\)) space in the case of the Hektor family, because of its origin by a cratering event--there is a large gap in the range between absolute magnitude of (624) Hektor (\(H = 7.20\)) and other bodies (\(H > 11.9\)), so we are not able to distinguish a simple Gaussian dispersion from the physical dependence (cf. Carruba et al. 2016). Either way, we are interested in the orbital distribution of *mostly* small bodies. Our assumption of size-independent ejection velocity is also in good agreement with results of SPH models (see Subsection [7.3](#subsec:velocity_fields){reference-type="ref" reference="subsec:velocity_fields"} and Figure [\[fig:SPH_v\_evol\]](#fig:SPH_v_evol){reference-type="ref" reference="fig:SPH_v_evol"}). To create a synthetic family in the same position as occupied by the observed Hektor family, we integrated the orbit of asteroid (624) Hektor with osculating elements taken from AstOrb catalogue (Bowell et al., 2002), until we got appropriate values of the true anomaly \(f\) and the argument of pericentre \(\omega\). We tried values of \(f\) ranging from \(0^\circ\) to \(180^\circ\) with the step of \(30^\circ\) and \(\omega\) always satisfying the condition \(f + \omega = 60^\circ\), i.e. we fixed the angular distance from the node to ensure a comparably large perturbations in inclinations. Initial positions of synthetic families members just after the disruption, compared to the observed Hektor family, are shown in Figure [\[Hektor_sim_init\]](#Hektor_sim_init){reference-type="ref" reference="Hektor_sim_init"}. To make a quantitative comparison of the distribution in the space of proper elements, we used a two-dimensional Kolmogorov--Smirnov test to compute KS distance of the synthetic family to the observed one with the output timestep of 1 Myr. The results for different initial geometries are shown in Figure [\[KS\]](#KS){reference-type="ref" reference="KS"}. Our two best fits corresponding to the lowest KS distance are displayed in Figure [\[Hektor_best_fit\]](#Hektor_best_fit){reference-type="ref" reference="Hektor_best_fit"}. As we can see from the image of the whole Trojan L4 population, Hektor seems to be near the outskirts of the librating region (cf. Figure [\[trojans_L4_arer_sizes\]](#trojans_L4_arer_sizes){reference-type="ref" reference="trojans_L4_arer_sizes"}). In Figure [\[Hektor_sim_init\]](#Hektor_sim_init){reference-type="ref" reference="Hektor_sim_init"}, we can note, that there are almost no observed asteroids in the shaded area with \(a_{\rm p}> 5.32\,\rm{au}\), but we can see some synthetic family members in the left panel of Figure [\[Hektor_best_fit\]](#Hektor_best_fit){reference-type="ref" reference="Hektor_best_fit"} (initial geometry \(f=0^\circ\), \(\omega = 60^\circ\)). On the other hand, when we look at right panel of Figure [\[Hektor_best_fit\]](#Hektor_best_fit){reference-type="ref" reference="Hektor_best_fit"} (initial geometry \(f=150^\circ\), \(\omega = 270^\circ\)), we can see that there are many fewer bodies in the proximity of the border of the stable librating region. One can also see the initial "fibre-like" structure is still visible on the left, but is almost dispersed on the right. Hence, we conclude that the geometry at which the disruption occurred is rather \(f=150^\circ\), \(\omega = 270^\circ\) and the corresponding age is between 1 and 4 Gyr. The second but less likely possibility is that the disruption could have occured more recently (0.1 to 2.5 Gyr) at \(f=0^\circ\), \(\omega = 60^\circ\). ### Parent body size from SPH simulations {#subsec:Durda_fit} We tried to estimate the parent body size of Hektor family and other families by the method described in Durda et al. (2007). To this point, we calculated a pseudo\(-\chi^2\) for the whole set of synthetic size-frequency distributions as given by the SPH simulations results (see Figure [\[Hektor_SPH_fit\]](#Hektor_SPH_fit){reference-type="ref" reference="Hektor_SPH_fit"}). Parent body sizes \(D_{\rm PB(SPH)}\) and mass ratios of the largest fragment and parent body \(M_{\rm{LF}}/M_{\rm{PB}}\) estimated by this method are listed in Table [\[fams_prop\]](#fams_prop){reference-type="ref" reference="fams_prop"}. The parent body size for Hektor family we derived from SPH simulations is \(D_{\rm PB(SPH)} = (260 \pm 10)\,\rm km\), the impactor diameter \(D_{\rm imp} = (24 \pm 2)\,\rm km\), the impactor velocity \(v_{\rm imp} = (4 \pm 1)\,\rm{km}\,\rm{s^{-1}}\) and the impact angle \(\varphi_{\rm imp} = (60^\circ \pm 15^\circ)\). We will use these values as initial conditions for simulations of collisional evolution below. ## 1996 RJ---extremely compact family In our previous work, we mentioned a small cluster associated with asteroid (9799) 1996 RJ, which consisted of just 9 bodies. With the contemporary sample of resonant elements we can confirm that this cluster is indeed visible. It is composed of 18 bodies situated near the edge of the librating zone on high inclinations, within the ranges \(I_{\rm p} \in \langle 31.38^\circ ; 32.27^\circ \rangle\) and \(a_{\rm{p}}\in\langle 5.225\,;\,5.238\rangle\) au. As it is detached from the background in the space of proper elements, it remains isolated even at high cutoff velocity \(v_{\rm{cutoff}}=160\, \rm{m\,s^{-1}}\). Unfortunately, we have albedos measured by WISE for just 4 members of this family. These albedos are not much dispersed. They range from \(p_{\rm{V}} = 0.079 \pm 0.019\) to \(p_{\rm{V}} = 0.109 \pm 0.029\) and, compared to the median albedo of the whole \(L_4\) population \(\widetilde {p_{\rm{V}}} = 0.072 \pm 0.017\), they seem to be a bit brighter, but this statement is a bit inconclusive. ## Arkesilaos This family is located on low inclinations \(I_{\rm p} \in \langle 8.52^\circ; 9.20^\circ \rangle\), in the range of \(a_{\rm{p}}\in\langle 5.230\,;\,5.304\rangle\) au. It is clearly visible in the space of proper elements, although this area of \(L_4\) cloud is very dense. Still, it is difficult to find the largest remnant of the parent body, because this region is populated mainly by small asteroids with absolute magnitudes \(H>12\). The only four asteroids with \(H<12\) are (2148) Epeios with \(H=10.7\), (19725) \(1999\,\rm WT_4\) with \(H=10.7\), (38600) \(1999\,\rm XR_{213}\) with \(H=11.7\) and (20961) Arkesilaos with \(H=11.8\). The only diameter derived from measured albedo is that of (2148) Epeios, which is \(D = (39.02 \pm 0.65)\,\rm km\). Diameters of remaining bodies were calculated from their absolute magnitude assuming albedo \(p_{\rm V} = 0.072\), which is the median of \(L_4\) Trojans. Although (20961) Arkesilaos has the diameter only \(D = (24 \pm 5)\,\rm km\), it is the only asteroid with \(H<12\), for which the associated family has a reasonable number of members \(N_{\rm memb}\) even for small values of the cutoff velocity \(v_{\rm cutoff}\) (see Section [4.2](#sec:HCM){reference-type="ref" reference="sec:HCM"}). As this is also the only larger body located near the center of the family in the space of proper elements, we treat (20961) Arkesilaos as the largest remnant of the parent body, whose diameter we estimate to be \(D_{\rm PB(SPH)} \simeq 87\,\rm km\). Given that the mass ratio of the largest remnant and the parent body, as derived from SPH simulations of Durda et al. (2007), is \(M_{\rm LR}/M_{\rm PB} \simeq 0.02\) only, it seems this family inevitably originated from a catastrophic disruption. ## Ennomos In our previous work, we reported a discovery of a possible family associated with asteroid (4709) Ennomos. With new data, we can still confirm that there is a significant cluster near this body, but when we take into account our "\(N_{\rm{memb}}(v_{\rm{cutoff}})\)" criterion described above, it turns out that the family is rather associated with asteroid (17492) Hippasos. It is a relatively numerous group composed of almost 100 bodies, situated near the border of the stable librating zone \(L_5\) at high inclinations, ranging from \(I_{\rm p} \in\langle 26.86^\circ;30.97^\circ \rangle\), and \(a_{\rm{p}}\in\langle 5.225; 5.338\rangle\) au. ## 2001 UV\(_{209}\) Using new data, we discovered a "new" family around asteroid (247341) \(2001\,\rm UV_{209}\), which is the second and apparently the last observable family in our sample. Similar to the Ennomos family, it is located near the border of the \(L_5\) zone on high inclinations \(I_{\rm p} \in \langle 24.02^\circ;26.56^\circ \rangle\) and \(a_{\rm{p}}\in\langle 5.218;5.320\rangle\) au. This family has an exceptionally steep slope of the SFD, with \(\gamma=-8.6 \pm 0.9\), which may indicate a recent collisional origin or a disruption at the boundary of the libration zone, which may be indeed size-selective as explained in Chrenko et al. (2015). # Collisional models of the Trojan population {#sec:coll_model} In order to estimate the number of collisional families among \(L_4\) Trojans, we performed a set of 100 simulations of the collisional evolution of Trojans with the Boulder code (Morbidelli et al., 2009) with the same initial conditions, but with different values of the random seed. ## Initial conditions We set our initial conditions of the simulations such that 4 Gyr of collisional evolution leads to the observed cumulative SFD of \(L_4\) Trojans (red curve in Figure [\[fig:Boulder_SFD\]](#fig:Boulder_SFD){reference-type="ref" reference="fig:Boulder_SFD"}). We constructed the initial synthetic SFD as three power laws with the slopes \(\gamma_{\rm a} = -6.60\) in the size range from \(D_1=117\,\rm{km}\) to \(D_{\rm max}=250\,\rm{km}\), \(\gamma_{\rm b} =-3.05\) from \(D_2=25\,\rm{km}\) to \(D_1\) and \(\gamma_{\rm c} =-3.70\) from \(D_{\rm min}=0.05\,\rm{km}\) to \(D_2\). The synthetic initial population was normalized to contain \(N_{\rm norm}=11\) asteroids with diameters \(D\geq D_1\). To calculate the target strength \(Q_D^*\), we used a parametric formula of Benz and Asphaug (1999): \[\label{Q*} Q_D^*=Q_0 R_{\rm PB}^a + B\rho_{\rm bulk}R_{\rm PB}^b,\] where \(R_{\rm PB}\) is the parent body radius in centimetres, \(\rho_{\rm bulk}\) its bulk density, which we set to be \(\rho_{\rm bulk} = 1.3\,\,\rm{g}\,\rm{cm^{-3}}\) for synthetic Trojans (cf. Marchis et al., 2014). As of constants \(a,\,b,\,B\,\rm{and}\,Q_0\) we used the values determined by Benz and Asphaug (1999) for ice at the impact velocity \(v_{\rm imp}=3\,\rm km\,s^{-1}\), which are: \(a=-0.39\), \(b=1.26\), \(B=1.2\,\,\rm erg\,cm^3\,g^{-2}\) and \(Q_0=1.6\cdot10^7\,\rm erg\,g^{-1}\). In our model, we take into account only Trojan vs Trojan collisions, as the Trojan region is practically detached from the main belt. Anyway, main-belt asteroids with eccentricities large enough to reach the Trojan region are usually scattered by Jupiter on a time scale significantly shorter than the average time needed to collide with a relatively large Trojan asteroid. We thus assumed the values of collisional probability \(P_{\rm i}=7.80\cdot 10^{-18}\,\,\rm km^{-2}\,yr^{-1}\) and the impact velocity \(v_{\rm imp}=4.66\,\,\rm km\,s^{-1}\) (Dell'Oro et al., 1998). Unfortunately, Benz and Asphaug (1999) do not provide parameters for ice at the impact velocities \(v_{\rm imp} > 3\,\rm km\,s^{-1}\). We also ran several simulations with appropriate values for basalt at impact velocity \(v_{\rm imp}=5\,\rm km\,s^{-1}\) (\(a=-0.36\), \(b=1.36\), \(B=0.5\,\,\rm erg\,cm^3\,g^{-2}\) and \(Q_0=9\cdot10^7\,\rm erg\,g^{-1}\)). Both models qualitatively exhibit the same evolution of SFD and they give approximately the same total numbers of disruptions and craterings occured, but for basalt the model gives three times fewer *observable* families originated by cratering than for ice. The results for the ice match the observation better, so we will further discuss the results for ice only. ## Long-term collisional evolution The results of our simulations of the collisional evolution are shown in Figure [\[fig:Boulder_SFD\]](#fig:Boulder_SFD){reference-type="ref" reference="fig:Boulder_SFD"}. Our collisional model shows only little changes above \(D>50\,\rm{km}\) over the last 3.85 Gyr (i.e. post-LHB phase only). Slopes of the initial synthetic population and the observed \(L_4\) population differ by \(\Delta \gamma < 0.1\) in the size range from 50 km to 100 km, while a relative decrease of the number of asteroids after 3.85 Gyr of collisional evolution is only about \(12\,\%\) in the same size range. Hence, we can consider this part of the Trojan population as a representative sample of the source population, which is not much affected by collisional evolution. Therefore, these Trojans provide very useful information about the source population, from which they were captured (as modeled in Nesvorný et al., 2013). ## An estimate of the number of observable families From our set of simulations, we also obtained the number of collisions leading to collisional families among \(L_4\) Trojans, namely catastrophic disruptions, where the mass ratio of the largest remnant and the parent body \(M_{\rm{LR}}/M_{\rm{PB}} < 0.5\), and cratering events, where \(M_{\rm{LR}}/M_{\rm{PB}} > 0.5\). As one can verify in Figure [\[fig:Boulder_families\]](#fig:Boulder_families){reference-type="ref" reference="fig:Boulder_families"}, these numbers are dependent on the diameter of the parent body \(D_{\rm PB}\). However, not all of these collisions produce families which are in fact observable (detectable). There are generally two possible obstacles in the detection of a family in the space of proper elements: i) somewhat more concentrated background population, due to which our detection methods (both "randombox" and HCM, see Chapter [4](#sec:groups){reference-type="ref" reference="sec:groups"}) may fail, if the number of observed fragments is too low in comparison with the background, and ii) an observational incompleteness, which means that in the case of Trojans, a substantial part of fragments with sizes \(D<10\,\rm km\) is still unknown, what again reduces a chance of a family detection. For these reasons, we constructed a criterion of observability that a synthetic family must fulfill in order to be detectable in the current conditions (i.e. we simulated a detection of synthetic families by the same methods we used to detect the real ones). The simplest criterion could be that a family must contain at least \(N_{\rm min} = 10\) fragments with diameter \(D\geq 10\,\rm{km}\). Within 100 simulations, there were 93 catastrophic disruptions of bodies with diameters \(D_{\rm{PB}}>100\,\rm{km}\), but only 50 of them produced more than 10 fragments with \(D\geq10\,\rm{km}\), see Figure [\[fig:Boulder_families\]](#fig:Boulder_families){reference-type="ref" reference="fig:Boulder_families"}. Hence, the probability that we would observe a collisional family originated by a catastrophic disruption of a parent body with \(D_{\rm{PB}}>100\,\rm{km}\) is *only* 0.50, which matches the observations (namely Eurybates family with \(D_{\rm PB(SPH)}\simeq155\,\,\rm km\), see Table [\[fams_prop\]](#fams_prop){reference-type="ref" reference="fams_prop"}). This value is also roughly consistent with our previous estimate based on the stationary model (Brož and Rozehnal, 2011), which gives the value 0.32 with new observational data. As one can also see in Figure [\[fig:Boulder_families\]](#fig:Boulder_families){reference-type="ref" reference="fig:Boulder_families"}, the number of cratering events is about one to two orders higher than the number of catastrophic disruptions, however, they do not produce enough fragments larger than \(D\geq10\,\rm{km}\). For the parent body size \(D_{\rm{PB}}>100\,\rm{km}\) there occurred almost 45,000 cratering events within 100 simulations which produced the largest fragment with \(D_{\rm LF}\geq1\,\rm km\), but only 10 of them fulfill our criterion of observability. Hence, the probability that we can observe a family originated by a cratering of a parent body with \(D_{\rm{PB}}>100\,\rm{km}\) is only 0.10, at least with contemporary data. From a statistical point of view, this can actually correspond to the Hektor family. As we have already demonstrated in Brož and Rozehnal (2011), the number of families is not significantly affected by chaotic diffusion or by a ballistic transport outside the libration zone. # SPH simulations of Hektor family {#sec:SPH} As we have already mentioned in Section [5.2](#subsec:Hektor){reference-type="ref" reference="subsec:Hektor"}, (624) Hektor is very interesting Trojan asteroid with possibly bilobed shape and a small moon. Diameters of (624) Hektor stated in Marchis et al. (2014) are as follows: equivalent diameter \(D_{\rm eq}=(250\pm26)\,\,\rm km\) for a convex model, the individual diameters of the lobes \(D_{\rm A}=(220\pm22)\,\,\rm km\), \(D_{\rm B}=(183\pm18)\,\,\rm km\) for a bilobed version. Estimated parameters of the moon are: the diameter \(D_{\rm m}=(12\pm3)\,\,\rm km\), the semimajor axis \(a_{\rm m}=(623\pm10)\,\,\rm km\), the eccentricity \(e_{\rm m}=(0.31\pm0.03)\) and the inclination (with respect to the primary equator) \(I_{\rm m}=(50\pm1)^\circ\). As we associate (624) Hektor with the collisional family, we would like to know, how the properties of the family are influenced by the shape of target body. We therefore performed a series of SPH simulations aiming to explain the origin of the Hektor family, for both cases of convex and bilobed shape of its parent body. ## Methods and initial conditions We simulated a collisional disruption using the smoothed-particle hydrodynamic code SPH5 (Benz and Asphaug, 1994). We performed two sets of simulations. In the first one, we simulated an impact on a single spherical asteroid. In the second, on a bilobed asteroid represented by two spheres positioned next to each other. The two touching spheres have a narrow interface, so that the SPH quantities do not easily propagate between them. In this setup, we are likely to see differences between sinlge/bilobed cases as clearly as possible. As for the main input parameters (target/impactor sizes, the impact velocity and the impact angle) we took the parameters of our best-fit SFDs, obtained by Durda et al. (2007) scaling method, see Section [5.2.2](#subsec:Durda_fit){reference-type="ref" reference="subsec:Durda_fit"} and Figure [\[Hektor_SPH_fit\]](#Hektor_SPH_fit){reference-type="ref" reference="Hektor_SPH_fit"}. To simulate a collision between the parent body and the impactor we performed a limited set of simulations: i) a single spherical basalt target with diameter \(D_{\rm PB} = 260\,\,\rm km\) vs a basalt impactor with diameter \(D_{\rm imp} = 48\,\,\rm km\); ii) the single basalt target \(D_{\rm PB} = 260\,\,\rm km\) vs an ice impactor (a mixture of ice and 30 % of silicates) with \(D_{\rm imp} = 64\,\,\rm km\) (impactor diameter was scaled to get the same kinetic energy); iii) a bilobed basalt target approximated by two spheres with diameters \(D_{\rm PB} = 200\,\,\rm km\) each (the total mass is approximately the same) vs a basalt impactor with \(D_{\rm imp} = 48\,\,\rm km\); iv) a single spherical ice target \(D_{\rm PB} = 260\,\,\rm km\) vs an ice impactor \(D_{\rm imp} = 38\,\,\rm km\) (impactor diameter was scaled to get the same ratio of the specific kinetic energy \(Q\) to the target strength \(Q^*_{\rm D}\)). The integration was controlled by the Courant number \(C = 1.0\), a typical time step thus was \(\Delta t \simeq 10^{-5}\,{\rm s}\), and the time span was \(t_{\rm stop} = 100\,{\rm s}\). The Courant condition was the same in different materials, using always the maximum sound speed \(c_{\rm s}\) among all SPH particles, as usually. We used \(N_{\rm SPH, st}=10^5\) SPH particles for the single spherical target and \(N_{\rm SPH, bt}=2\cdot 10^5\) for the bilobed one. For impactor \(N_{\rm SPH, i}=10^3\) SPH particles. We assumed the Tillotson equation of state (Tillotson, 1962) and material properties, which are listed in Table [1](#tab:materials){reference-type="ref" reference="tab:materials"}. [\[tab:materials\]]{#tab:materials label="tab:materials"} We terminated SPH simulations after 100 s from the impact. This time interval is needed to establish a velocity field of fragments and to complete the fragmentation. Then we handed the output of the SPH simulation as initial conditions to the N--body gravitational code Pkdgrav (Richardson et al., 2000), a parallel tree code used to simulate a gravitational reaccumulation of fragments. Unlike Durda et al. (2007), who calculated radii of fragments \(R\) from the smoothing length \(h\) as \(R=h/3\), we calculated fragments radii from their masses \(m\) and densities \(\rho\) as \(R = (m/(4\pi\rho))^{1/3}\). We ran Pkdgrav with the time step \(\Delta t = 5.0\,\,\rm s\) and we terminated this simulation after \(t_{\rm evol} = 3\) days of evolution. To ensure this is sufficiently long, we also ran several simulations with \(t_{\rm evol} = 5\) days, but we had seen no significant differences between final results. We used the nominal value for the tree opening angle, \({\rm d}\theta = 0.5\,{\rm rad}\), even though for the evolution of eventual moons it would be worth to use even smaller value, e.g. \({\rm d}\theta = 0.2\,{\rm rad}\). ## Resulting size-frequency distributions From the output of our simulations we constructed size-frequency distributions of synthetic families, which we compare to the observed one, as demonstrated in Figure [\[fig:SPH_SFD\]](#fig:SPH_SFD){reference-type="ref" reference="fig:SPH_SFD"}. As one can see, there are only minor differences between SFDs of families created by the impacts on the single and bilobed target, except the number of fragments with diameter \(D<5\,\rm km\), but this is mostly due to different numbers of SPH particles. However, there are differences between ice and basalt targets. Basalt targets provide generally steeper SFDs with smaller largest remnants than the ice target. To make the comparison of these synthetic initial SFDs to each other more realistic, we removed the fragments of the impactor from our synthetic families. This is because fragments of the impactor often do not remain in the libration zone. Note that this procedure does not subtitute for a full simulation of further evolution; it serves just for a quick comparison of the SFDs. To match the observed SFD of the Hektor family more accurately, we should perform a much larger set of simulations with different sizes of projectiles and also different compositions (mixtures of ice and basalt). However, material parameters of these mixtures are generally not known. Regarding the material constants of pure ice, we have them for the impact velocity \(v_{\rm imp} = 3\,\rm km\,s^{-1}\) only (Benz and Aspaugh, 1999). There are also some differences between SFDs of single and bilobe targets, so we should perform these simulations for each target geometry. However, we postpone these detailed simulations for future work; in this work we further analyse results of simulations with basalt targets and we focus on the evolution of the SFDs. It should be emphasized that the SFDs presented here correspond to very young synthetic families, hence they are not affected by any dynamical and collisional evolution yet. To reveal possible trends of the evolution by a ballistic transport and chaotic diffusion, we prepared initial conditions for the SWIFT integrator, similarly as described in Section [5.2.1](#sub:dyn_evol){reference-type="ref" reference="sub:dyn_evol"}, let the simulation run and monitored the corresponding evolution of the SFD. The results can be seen in Figure [\[fig:SPH_SFD_evol\]](#fig:SPH_SFD_evol){reference-type="ref" reference="fig:SPH_SFD_evol"}. The biggest difference between \(t=0\) and \(t=1\,\,\rm Myr\) is caused by a ballistic transport outside the libration zone---fragments (especially of the impactor) missing from the SFD at \(t=1\,\,\rm Myr\) were perturbed too much to remain in the libration zone, at least for a given impact geometry. We actually tested two impact geometries: in the direction tangential and perpendicular to the orbit. This may be important for the method we used in Section [5.2.2](#subsec:Durda_fit){reference-type="ref" reference="subsec:Durda_fit"} to derive a preliminary parent body size and other properties of the family. The SFDs obtained by Durda et al. (2007) were directly compared in their work to the main-belt families, however, there is a part of fragments among Trojans (in our case even the largest ones, see Figure [\[fig:SPH_SFD_evol\]](#fig:SPH_SFD_evol){reference-type="ref" reference="fig:SPH_SFD_evol"}), which cannot be seen in the space of resonant elements, because they do not belong to Trojans any more. Fortunately, values of pseudo-\(\chi^2\) we computed in Section [5.2.2](#subsec:Durda_fit){reference-type="ref" reference="subsec:Durda_fit"} depend rather weakly on the distribution of a few largest bodies. Even so, we plan to analyze SFDs of synthetic families more carefully in future works. ## Resulting velocity fields {#subsec:velocity_fields} In our \(N\)-body simulations, we used the model of isotropic disruption (Farinella et al., 1994). As we compared the synthetic family with the observed one (see Section [5.2.1](#sub:dyn_evol){reference-type="ref" reference="sub:dyn_evol"}), we simulated only the evolution of bodies with relatively low ejection velocities (\(v<200\,\rm m\,s^{-1}\)), because the observed family is confined by the cutoff velocity \(v_{\rm cutoff} = 110\,\rm m\,s^{-1}\). Very small fragments with higher velocities may be still hidden in the background. Here, we compare Farinella's model to the velocity fields of fragments from SPH simulations, see Figure [\[fig:SPH_v\_evol\]](#fig:SPH_v_evol){reference-type="ref" reference="fig:SPH_v_evol"}. We realized that Farinella's model is not offset substantially with respect to other velocity histograms, especially at lower velocities, \(v < 200\,{\rm m}\,{\rm s}^{-1}\). On the other side, there remained some fragments of the impactor with velocities \(v>2\,\rm km\,s^{-1}\) in our SPH simulations, which are not produced in the isotropic model. It does not affect a comparison of the synthetic and observed families in the space of proper elements, as these high-velocity fragments leaved the Trojan region in our case, but it does affect the SFD of the synthetic family. As a consequence, one should always analyse SFDs and velocity fields together. We also simulated a further evolution of the velocity field. After just 1 Myr of evolution, there remained no bodies with \(v>1.5\,\rm km\,s^{-1}\) in our impact geometries, and as one can see in Figure [\[fig:SPH_v\_evol\]](#fig:SPH_v_evol){reference-type="ref" reference="fig:SPH_v_evol"}, there was a rapid decrease in the number of fragments with initial \(v>300\,\rm m\,s^{-1}\). The resulting histogram is again similar to that of the simple isotropic model. ## Synthetic moons [\[tab:moons\]]{#tab:moons label="tab:moons"} In our simulation of the impact of basalt projectile on the bilobe-shape basalt target, we spotted two low-velocity fragments with original velocities \(130\,\rm m\,s^{-1}\) and \(125\,\rm m\,s^{-1}\), which were consequently captured as moons of the largest remnant. Their sizes and orbital parameters are listed in Table [2](#tab:moons){reference-type="ref" reference="tab:moons"}. These satellites were captured on orbits with high eccentricities (\(e=0.82\) and 0.64 respectively), which are much higher than the eccentricity of the observed moon determined by Marchis et al. (2014) (\(e=0.31\pm0.03\)). However, this could be partly caused by the fact, that we handed the output of (gravity free) SPH simulations to the gravitational N-body code after first 100 s. Hence, fragments leaving the parent body could move freely without slowing down by gravity. More importantly, we do not account for any long-term dynamical evolution of the moons (e.g. by tides or binary YORP). When compared to the observed satellite, the diameters of the synthetic moons are several times smaller. This is not too surprising, given that the results for satellite formation are at the small end of what can be estimated with our techniques (median smoothing length \(h = 2.3\,{\rm km}\); satellite radius \(r \simeq 1.2\,{\rm km}\)). The size of captured fragments could also be dependent on impact conditions as different impact angles, impactor velocities and sizes (as is the case for scenarios of Moon formation) which we will analyze in detail in the future and study with more focused simulations. # Conclusions {#sec:conclusions} In this paper, we updated the list of Trojans and their proper elements, what allowed us to update parameters of Trojan families and to discover a new one (namely \(2001\,\rm UV_{209}\) in \(L_5\) population). We focused on the Hektor family, which seems the most interesting due to the bilobed shape of the largest remnant with a small moon and also its D-type taxonomical classification, which is unique among the collisional families observed so far. At the current stage of knowledge, it seems to us there are no major inconsistencies among the observed number of Trojan families and their dynamical and collisional evolution, at least in the current environment. As usual, we "desperately" need new observational data, namely in the size range from 5 to 10 km, which would enable us to constrain the ages of asteroid families on the basis of collisional modeling and to decide between two proposed ages of Hektor family, 1 to 4 Gyr or 0.1 to 2.5 Gyr. As expected, there are qualitative differences in impacts on single and bilobed targets. In our setup, the shockwave does not propagate easily into the secondary, so that only one half the mass is totally damaged as one can see in Figure [\[fig:SPH\]](#fig:SPH){reference-type="ref" reference="fig:SPH"}. On the other hand, the resulting SFDs are not that different, as we would expect. Even so, there is a large parameter space, which is still not investigated (i.e. the impact geometry with respect to the secondary, secondary impacts, the position in the orbit). SPH simulations of impacts on bilobed or binary targets thus seem very worthy for future research. Our work is also a strong motivation for research of disruptions of weak bodies (e.g. comets), better understanding the cometary disruption scaling law and also for experimental determination of material constants, which appear in the respective equation of state. As a curiosity, we can also think of searching for the remaining projectile, which could be still present among Trojans on a trajectory substantially different from that of family. A substantial part of projectile momentum is preserved in our simulations, so we may turn the logic and we may assume the projectile most likely came from the Trojan region and then it should remain in this region too.
{'timestamp': '2016-07-19T02:01:54', 'yymm': '1607', 'arxiv_id': '1607.04677', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04677'}
# Introduction {#sec:intro} Discrete breathers are intrinsic localized modes that can emerge in networks of coupled nonlinear oscillators. They have been observed not only in Hamiltonian lattices but also in driven dissipative systems under certain conditions. Discrete breathers have been theoretically predicted or experimentally generated in a wide variety of physical systems such as Josephson junction arrays, coupled pendula chains, micro-and macro-mechanical cantilever arrays, granular crystals, nonlinear electrical lattices, and double-strand DNA models, just to cite a few instances. Up to now, breathers have been mainly studied for the case of a harmonic external excitation, while various types of periodic excitations are in principle possible, depending upon the physical context under consideration. Since there are infinitely many different wave forms, a quite natural question is to ask how the generation and dynamics of breathers are affected by the presence of a generic periodic excitation. In this present work, we show that a relevant quantity properly characterizing the effectiveness of zero-mean periodic excitations \(F(t)\) having equidistant zeros at controlling the generation and dynamics of discrete breathers is the *impulse* transmitted by the external excitation over a half-period (hereafter referred to simply as the excitation's impulse, \(I\equiv\int_{0}^{T/2}F(t)dt\), \(T\) being the period)\(-\) a quantity integrating the conjoint effects of the excitation's amplitude, period, and waveform. It is worth mentioning that the relevance of the excitation's impulse has been observed previously in quite different contexts, such as ratchet transport, adiabatically ac driven periodic (Hamiltonian) systems, driven two-level systems and periodically curved waveguide arrays, chaotic dynamics of a pump-modulation Nd:YVO\(_{4}\) laser, topological amplification effects in scale-free networks of signaling devices, and controlling chaos in starlike networks of dissipative nonlinear oscillators. The rest of this paper is organized as follows. In Sec. II we introduce the model and further comment on some of its main features. Section III provides numerical evidence for the essential role of the excitation's impulse at generating breathers and controlling their stability for the prototypical cases of a hard \(\phi^{4}\) potential and a sine-Gordon potential. A theoretical explanation of the effectiveness of the excitation's impulse in terms of energy and action is provided in Sec. IV. Finally, Sec. V is devoted to a discussion of the major findings and of some open problems. # Model system {#sec:model} The discrete nonlinear Klein--Gordon equation with linear coupling, linear damping, and external periodic excitation is one of the simplest equations where dissipative discrete breathers may arise:
{'timestamp': '2016-11-22T02:04:26', 'yymm': '1607', 'arxiv_id': '1607.04630', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04630'}
# Introduction {#sec:intro} The CDM paradigm, within which the majority of matter in the Universe is weakly interacting, has enjoyed great success in explaining astrophysical and cosmological data. The inclusion of CDM in the now standard \(\Lambda\)CDM cosmological model makes clear predictions for, among other things, the mass function of gravitationally collapsed structures (e.g. @Jenkins0005260 [@Evrard0110246; @Tinker0803.2706]), the ratio of baryonic mass to CDM in galaxy clusters (e.g. @Eke9708070 [@Kay0407058; @Borgani0906.4370]), and the distribution of mass within the structures that form hierarchically in the Universe (e.g. @Bullock9908159 [@Navarro0311231; @Navarro9611107; @Gao0711.0746]). Observations testing the predicted mass function and cluster baryon fraction have largely validated CDM (e.g. @White1993Natur.366..429W [@Bahcall9803277; @Reiprich0111285; @Ettori0211335; @Ettori0904.2740; @Allen0405340; @Allen0706.0033; @Allen1103.4829; @Vikhlinin0812.2720; @Mantz0909.3098; @Mantz1407.4516; @Rozo0902.3702; @Sehgal1010.1025; @Benson1112.5435; @Hasselfield1301.0816; @Planck1303.5080; @Planck1502.01597]). Observational tests of the distribution of mass within bound structures is in some ways more challenging, requiring spatially resolved measurements of the gravitational potential of objects whose mass is predominately dark. Nevertheless, an extensive body of work now exists, using primarily X-ray and gravitational lensing observations of galaxy clusters. In this work, we revisit the subject using *Chandra* X-ray observations of a sample of massive, highly dynamically relaxed clusters. Although not representative of the cluster population at large, these systems are an ideal laboratory for deriving three-dimensional mass profiles based on observations of the intracluster medium (ICM) because departures from hydrostatic equilibrium and systematics due to projection are minimized. Both simulations and direct calibration using weak gravitational lensing indicate that the overall bias in *Chandra* X-ray mass estimates for this sample is small (\(\ {\raise-.75ex\hbox{\)\buildrel<\over\sim\(}}\ 10\) per cent; @Nagai0609247 [@Applegate1509.02162]). The specific features of the cluster mass distribution that we consider here are (1) the concentration parameter of cluster mass profiles, and its dependence on mass and redshift; and (2) the shape of mass profiles, in particular departures from the baseline model defined by @Navarro9611107 (hereafter NFW). The selection of the sample of massive, relaxed clusters employed here is detailed in the first paper in this series (@Mantz1502.06020; hereafter Paper ). Papers , and employ *Chandra* data for this sample to respectively constrain cosmological parameters, through gas-mass fraction measurements; scaling relations and thermodynamics of the ICM; and the average bias of the X-ray mass determinations. Section [2](#sec:data){reference-type="ref" reference="sec:data"} reviews our procedure for constraining cluster mass profiles from X-ray data, for which complete details are available in Papers  and , and introduces the specific mass models employed in this work. In Section [3](#sec:results){reference-type="ref" reference="sec:results"}, we review the specific predictions of CDM for cluster mass profile concentrations and shapes, present our results, and compare them to others in the literature. Section [4](#sec:conclusions){reference-type="ref" reference="sec:conclusions"} summarizes our conclusions. We assume a concordance flat \(\Lambda\)CDM cosmology with \(h=0.7\) and \(\mysub{\Omega}{m}=0.3\) throughout. Unless otherwise noted, quoted parameter uncertainties correspond to the maximum-likelihood (i.e. shortest) interval enclosing 68.3 per cent posterior probability, and best-fitting values are the posterior modes. # Data and Analysis {#sec:data} This work employs the sample of 40 massive, dynamically relaxed galaxy clusters identified in Papers  and (Table [\[tab:thetable\]](#tab:thetable){reference-type="ref" reference="tab:thetable"}). For our purposes, dynamical relaxation was defined quantitatively in terms of X-ray image features, specifically the sharpness of the surface brightness peak and the alignment and symmetry of a series of standardized cluster isophotes; in addition all clusters in the sample must have an emission-weighted temperature of \(>5\) keV outside of their cores. The procedure for cleaning and fitting the *Chandra* data for these clusters is described in detail in Paper ; in Paper  the data reduction was updated to use a newer version of the *Chandra* calibration files (specifically caldb[^1] 4.6.2), and we use that version of the data reduction here. A complete list of the specific OBSIDs included in the analysis can also be found in that work. Note that, while the data for each of these 40 clusters can constrain the 2-parameter NFW mass model, the 3-parameter models discussed below can be constrained for only a subset of entire sample (see Sections [3.2](#sec:gnfw){reference-type="ref" reference="sec:gnfw"} and [3.3](#sec:einasto){reference-type="ref" reference="sec:einasto"}). As in our previous papers, the spectral analysis of the X-ray data here involves a deprojection, assuming spherical symmetry and hydrostatic equilibrium, in order to infer the 3-dimensional properties of the hot gas and gravitating matter in the clusters. We consider three parametric models for the total mass profiles of clusters: the standard NFW profile, whose results were previously presented in Papers  and ; a generalized NFW (GNFW) profile, in which the power-law slope of the density profile at small radii is a free parameter; and the Einasto profile, which provides a more accurate description of simulated halos than the other two options, particularly at small radii. Up to normalizing factors, the density profiles associated with these models are \[\begin{aligned} \label{eq:profmodels} \mysub{\rho}{NFW}(r) &\propto& \left(\frac{r}{\mysub{r}{s}}\right)^{-1} \left(1 + \frac{r}{\mysub{r}{s}}\right)^{-2}, \\ \mysub{\rho}{GNFW}(r) &\propto& \left(\frac{r}{\mysub{r}{s}}\right)^{-\beta} \left(1 + \frac{r}{\mysub{r}{s}}\right)^{-(3-\beta)}, \nonumber\\ \mysub{\rho}{Einasto}(r) &\propto& \exp \left\{-\frac{2}{\alpha}\left[\left(\frac{r}{r_{-2}}\right)^\alpha-1\right] \right\}, \nonumber \end{aligned}\] where and \(r_{-2}\) are scale radii. Figure [\[fig:demo\]](#fig:demo){reference-type="ref" reference="fig:demo"} shows how the shapes of the GNFW and Einasto models differ from the NFW profile. Note that, while the NFW model is a special case of the GNFW profile with \(\beta=1\), no value of \(\alpha\) makes the Einasto profile completely equivalent to NFW, although values of \(\alpha \approx 0.2\)--\(0.3\) can closely approximate it over the radial range probed by our data (roughly \(0.1 \ {\raise-.75ex\hbox{\)\buildrel<\over\sim\(}}\ r/\mysub{r}{s} \ {\raise-.75ex\hbox{\)\buildrel<\over\sim\(}}\ 1.5\), in terms of the NFW scale radius). The spectral analysis itself uses xspec, within which we have modified the clmass model to include the GNFW and Einasto mass profiles in addition to NFW. This model treats the cluster ICM as a series of concentric, isothermal spherical shells of gas, in hydrostatic equilibrium with the gravitational potential given by one of the above parametric mass profiles. We simultaneously model absorption from Galactic hydrogen and, where necessary, the presence of foreground emission from the Galactic halo and/or the local bubble. Complete details of these aspects of the analysis can be found in Paper . For every cluster, we identify and exclude from the analysis a central region, typically 50 kpc in radius, where visible features in the ICM indicate that the assumptions of spherical symmetry and hydrostatic equilibrium are not met. This limits the radial range over which we can draw conclusions about the shape of the mass profiles, in particular excluding the central galaxies of each cluster, where the contribution of cold gas and stars to the total mass may be significant (e.g. @Newman1209.1391). The outer radius used in our analysis is set by the need to measure temperature profiles without background modelling dominating the uncertainties (see Paper ). The specific radial range for each cluster over which the mass profiles are constrained is listed in Table [\[tab:thetable\]](#tab:thetable){reference-type="ref" reference="tab:thetable"}. For each cluster and each mass model considered, our spectral analysis produces a Markov chain of the parameters of interest, for which the density of samples is proportional to the joint posterior density of the parameters. Rather than summarize these results by, e.g., a mean and standard deviation, we continue to use the full set of samples when constraining properties of the ensemble of clusters (e.g. the concentration-mass relation) in Section [3](#sec:results){reference-type="ref" reference="sec:results"}. In detail, the samples for each cluster are importance weighted and then directly numerically integrated over to provide a likelihood for the hierarchical model being fitted. This is especially important for constraining the distributions of the GNFW and Einasto shape parameters, as the posterior distributions of these parameters for individual clusters are frequently poorly constrained and highly asymmetric (Table [\[tab:thetable\]](#tab:thetable){reference-type="ref" reference="tab:thetable"}). We additionally calculate central cooling luminosities (\(\mysub{L}{cool}\)) for each cluster in the sample. These we define as the bolometric luminosity emitted within the radius where the cooling time equals 5 Gyr. The cooling time is calculated as \[t_\mathrm{c} = \frac{3\, n_\mathrm{tot} \,kT}{2\, n_\mathrm{e} n_\mathrm{H} \,\Lambda(T,Z)},\] where \(T\) is the ICM temperature; \(n_\mathrm{e}\), \(n_\mathrm{H}\) and \(n_\mathrm{tot}\) are the number densities of electrons, protons and all particles; \(Z\) is the metallicity; and \(\Lambda(T,Z)\) is the cooling function. These calculations are based on spectral analyses of each cluster that do not assume equilibrium or a particular underlying mass profile, using instead the projct model in xspec, as presented in Paper . Like total luminosity, \(\mysub{L}{cool}\) displays an overall trend with mass; fitting a power-law plus log-normal intrinsic scatter model, we find \[\label{eq:Lcool-scaling} \frac{\mysub{L}{cool}}{10^{45}\unit{erg}\unit{s}^{-1}} = \left(2.9\pm0.3\right) \left(\frac{M_{200}}{10^{15}\ensuremath{\, M_{\odot}}}\right)^{1.3\pm0.3},\] with a log-normal scatter of \(0.70\pm0.09\). # Results and Discussion {#sec:results} This section presents our results and discusses them in the context of CDM predictions and other measurements in the literature. In Section [3.1](#sec:cm){reference-type="ref" reference="sec:cm"}, we assume the NFW mass model and investigate the mass and redshift dependence of cluster concentrations. Sections [3.2](#sec:gnfw){reference-type="ref" reference="sec:gnfw"} and [3.3](#sec:einasto){reference-type="ref" reference="sec:einasto"} respectively consider the GNFW and Einasto generalizations to the NFW profile. ## NFW Concentration--Mass Relation {#sec:cm} The concentration of mass within an NFW halo can be described by the parameter \[c = \frac{r_{200}}{\mysub{r}{s}},\] where \(r_{200}\) is the radius within which the mean enclosed density is 200 times the critical density of the Universe at a cluster's redshift. Numerous dark-matter-only simulations have shown that the concentration of halos is expected to vary with both redshift and mass on average, and to have an intrinsic scatter of \(\sim20\) per cent driven primarily by differences in accretion history. Early work suggested a scaling relation similar to \(c \propto (1+z)^{-1} \, M^{-0.13}\). More recent work revealed that the dependence on both mass and redshift is reduced for the most massive (i.e., cluster-scale) halos (e.g. @Gao0711.0746 [@Klypin1002.3660; @Prada1104.5130]), largely because these halos are preferentially recently formed (e.g. @Ludlow1312.0945). have shown that selecting the most dynamically relaxed halos from simulations recovers a power-law with mass out to the largest masses. Our current understanding, at least on the basis of N-body simulations, is that concentrations for individual clusters grow steadily as they accrete through minor mergers, but that sufficiently disruptive mergers can "reset" the concentration parameter to lower values (\(c\sim3\)), with the overall trends largely reflecting differences in the typical accretion histories of halos as a function of mass and redshift. By construction, our sample contains only the most dynamically relaxed, massive clusters at any redshift. We therefore expect their accretion histories at the time of observation to be similar, and in particular that the time since the last major merger is significant in dynamical terms. As a consequence, we do not necessarily expect the redshift dependence found in simulations for all halos to apply to our sample, since it must partially reflect evolution in the rate of major mergers that will not be present in our sample. It therefore makes sense to consider models where both the dependence on mass and redshift are free parameters, \[\label{eq:cm} c = c_0 \, \left(\frac{1+z}{1.35}\right)^{\kappa_\zeta} \left( \frac{M_{200}}{10^{15}\ensuremath{\, M_{\odot}}} \right)^{\kappa_m}.\] Our model also includes as a free parameter a log-normal intrinsic scatter in concentration, \(\sigma_{\ln c}\), which is assumed to have the same value at all masses and redshifts. The concentrations and masses of our clusters, derived from fitting the NFW profile, are shown in the left panel of Figure [\[fig:cM\]](#fig:cM){reference-type="ref" reference="fig:cM"}. Note that the measurement correlation between \(M_{200}\) and \(c\) is typically strong, as illustrated by a joint constraint ellipse for the case of RX J1347.5\(-\)`<!-- -->`{=html}1145, and that this correlation is accounted for by our fitting method (Section [2](#sec:data){reference-type="ref" reference="sec:data"}). Fitting the model of Equation [\[eq:cm\]](#eq:cm){reference-type="ref" reference="eq:cm"}, we find \(\kappa_m =-0.16 \pm 0.07\) and \(\kappa_\zeta =-0.17 \pm 0.26\). A dependence on mass, consistent with the expectation from simulations, is detected at \(\sim2\sigma\) significance, while the redshift dependence is consistent with zero. We find an intrinsic scatter of \(\sigma_{\ln c}=0.16\pm 0.03\), somewhat smaller that the \(\sim0.2\) scatter found for all halos in simulations. The results are consistent with the notion that our sample consists of highly relaxed clusters, for which we expect a power-law with mass to hold even at the highest masses, with a smaller intrinsic scatter than the full population. This is reinforced by the fact that our concentrations are on average greater than those in simulations at the same masses. Selecting relatively relaxed halos from simulations based on the absence of substructure and the ratio of kinetic to potential energy, find median concentrations of 4--5 in the mass range considered here, consistent with our observed clusters being typically more relaxed than the halos they selected.[^2] These authors also find a clear correlation between halo concentration and the time since "formation", operationally defined as the time when the halo reaches half of its final mass (see also @Wechsler0108151). Taking their \(z=0\) and \(z=1\) results as a guide, our measured concentrations in the range 5--8 place the clusters' formation times approximately between redshifts 1--2. Our selection of the most relaxed systems is also plausibly responsible for a reduced redshift dependence (\(\kappa_\zeta\) consistent with zero) compared with all clusters in simulations. The right panel of Figure [\[fig:cM\]](#fig:cM){reference-type="ref" reference="fig:cM"} shows the residuals from the best-fitting concentration--mass relation as a function of central cooling luminosity (as a proxy for the importance of non-gravitational physics), with the latter also expressed as a residual from the best-fitting \(\mysub{L}{cool}\)--mass relation (Equation [\[eq:Lcool-scaling\]](#eq:Lcool-scaling){reference-type="ref" reference="eq:Lcool-scaling"}). While the two sets of residuals are not strongly correlated, it may be worth noting the two clusters with large positive residuals in both concentration and cooling luminosity, RX J1347.5\(-\)`<!-- -->`{=html}1145 and SPT-CL J2344\(-\)`<!-- -->`{=html}4242. If the relatively efficient cooling at work in the centers of these clusters has an effect on the concentration of the dark matter as well as the gas, boosting the value of \(c\), this would not be represented in the dark-matter-only simulations used to predict the c--M relation. (The closest analogs of these systems in simulations may, of course, still have greater than average concentration.) Concentration--mass estimates based on X-ray data, mostly for low-redshift clusters, show broad agreement with the mass dependence predicted from N-body simulations. More recently, measurements based on weak gravitational lensing have become possible. Weak lensing shear alone can constrain concentrations only for stacks or ensembles of clusters, but individual cluster constraints are possible in combination with strong lensing or magnification. Our own analysis of weak lensing data for a subset of the clusters used in this work finds an ensemble average concentration of \(3.0^{+4.4}_{-1.8}\) (Paper ), consistent with our X-ray results here within the statistical uncertainties. Results using strong and weak lensing by X-ray selected clusters and from stacked analyses of optically selected clusters are in broadly good agreement with our results. However, an arguably better comparison is with the lensing analyses of the CLASH sample of clusters (\(0.19<z<0.89\)), which mostly consists of relaxed or nearly relaxed clusters (by our definition). Using CLASH strong+weak lensing, found \(\kappa_m=-0.32\pm0.18\) and \(\kappa_\zeta=-0.14\pm0.52\), consistent with our results; incorporating magnification, found \(\kappa_m=-0.44\pm0.19\) while fixing \(\kappa_\zeta=-0.668\). We conclude that our concentration--mass estimates for relaxed clusters are in good agreement with simulations and previous measurements, given the expectation that we have selected dynamically similar, and, in particular, the most dynamically relaxed, clusters at all redshifts. Direct tests of that selection on simulated cluster images would not be straightforward, given the inability of current hydrodynamic simulations to reproduce the cores of relaxed clusters. However, the sensitivity of our results to a few strongly cooling clusters motivates doing so in future work, as well as expanding the relaxed cluster sample if possible. ## Generalized NFW Mass Profiles {#sec:gnfw} We next turn to constraints on the inner slope of the cluster mass profiles, as parametrized by the GNFW model. While a profile of the Einasto form is now known to provide a better description of simulated clusters over a large range in radii, the GNFW model still represents a simple modification of the baseline NFW model, and one that is arguably easier to interpret in light of the limited radial range of our data. N-body simulations generally place values of the inner slope of the GNFW profile, \(\beta\) (Equation [\[eq:profmodels\]](#eq:profmodels){reference-type="ref" reference="eq:profmodels"}), in the range 1.0--1.5, with \(\beta=1\) corresponding to the NFW model. Unfortunately, due to its greater complexity compared to the NFW profile, the GNFW model cannot be fully constrained for all of the clusters in our sample using the current data. In practice, the scale radius parameter becomes unconstrained, taking on very large values, while the inner slope becomes relatively steep, taking values roughly appropriate for the average of the profile in the observed region (i.e., \(\sim-2\)). A straightforward way to eliminate these cases is to adopt the outer radius where the data can constrain the mass profile, \(\mysub{r}{out}\), as an upper limit for the scale radius (when fitting the NFW profile, all scale radii are comfortably within the measured range). Those clusters for which the 95.4 per cent confidence interval for \(\mysub{r}{s}\) includes values \(>\mysub{r}{out}\) are removed from this portion of the analysis. We have checked that making this cut based on the 68.3 per cent confidence interval for \(\mysub{r}{s}\) does not change our results apart from tightening the constraints slightly (because fewer clusters are excluded). Thus, while this procedure at some level introduces the prior that the mass profiles not be too far from NFW, our results are not very sensitive to it. The left panel of Figure [\[fig:gnfw\]](#fig:gnfw){reference-type="ref" reference="fig:gnfw"} shows these individual constraints (see also Table [\[tab:thetable\]](#tab:thetable){reference-type="ref" reference="tab:thetable"}). Using the 20 clusters that pass this check, we fit a model for the average value of \(\beta\) and its cluster-to-cluster scatter, which we assume to be described by a Gaussian distribution with mean \(\bar{\beta}\) and width \(\sigma_\beta\). The joint constraints on these parameters are shown in the right panel of Figure [\[fig:gnfw\]](#fig:gnfw){reference-type="ref" reference="fig:gnfw"}, and are consistent with the NFW value of \(\beta=1\). The marginalized constraints are \(\bar{\beta} = 1.02 \pm 0.08\) and \(\sigma_\beta = 0.22 \pm 0.07\). We emphasize that the data used to fit this model do not extend all the way into the cluster centers, although they do include radii down to \(\sim 0.1\,\mysub{r}{s}\) (see Section [2](#sec:data){reference-type="ref" reference="sec:data"}). Our analysis indicates that, on average, the NFW model provides a good description of the most relaxed clusters down to small radii. However, it also shows significant intrinsic scatter in \(\beta\), at the \(\sim20\) per cent level, even among the most relaxed clusters. Earlier work by and using X-ray observations reached similar conclusions on the mean value of \(\beta\), respectively finding \(\bar{\beta}=0.88^{+0.26}_{-0.31}\) and \(0.98<\bar{\beta}<1.19\) (both 95 per cent confidence intervals). Neither of these authors fit for intrinsic scatter explicitly, but their results also pointed to its presence, to the extent that the scatter in individual-cluster constraints on \(\beta\) was not consistent with measurement uncertainties alone. More recently, also found inner mass profile slopes of massive clusters to be consistent with CDM-only predictions, using a combination of strong and weak gravitational lensing constraints, at radii where stars do not dominate the mass budget. ## Einasto Mass Profiles {#sec:einasto} In detail, the cluster mass profile shapes in simulations have been found to be better described by a continuously changing power-law, such as the Einasto form (Equation [\[eq:profmodels\]](#eq:profmodels){reference-type="ref" reference="eq:profmodels"}), than by the NFW or GNFW models (e.g. @Navarro0810.1522). Values of the shape parameter, \(\alpha\), seen in simulations are mass and redshift dependent; for massive clusters at \(z\ltsim1\), average values of \(\sim0.2\)--0.3 are typical. While we do not expect our data to be capable of distinguishing the GNFW and Einasto models, due to the limited radial range covered, it is nonetheless interesting to see how values of \(\alpha\) measured from the X-ray data compare to these expectations. Similarly to the GNFW case, degeneracies between model parameters prevent us from constraining the Einasto model for every cluster in our sample. We follow the same procedure as in Section [3.2](#sec:gnfw){reference-type="ref" reference="sec:gnfw"} to remove from the analysis clusters where the scale radius of the Einasto model is not constrained. Unlike the GNFW case, the constraints on \(\alpha\) for several of the remaining clusters are consistent with either the lower or upper bound of the uniform prior we adopt in the analysis, \(0.01\leq\alpha\leq0.75\) (left panel of Figure [\[fig:ein\]](#fig:ein){reference-type="ref" reference="fig:ein"}). While this makes our results dependent on the width of this prior, in practice the effect is small; for example, our results below are nearly identical to those we would obtain with a tighter prior of \(0.01\leq\alpha\leq0.5\). As in the previous section, we fit a Gaussian model for the intrinsic distribution of Einasto shape parameters, parametrized by \(\bar{\alpha}\) and \(\sigma_\alpha\). The results are shown in the right panel of Figure [\[fig:ein\]](#fig:ein){reference-type="ref" reference="fig:ein"}; the marginalized constraints are \(\bar{\alpha} = 0.29 \pm 0.04\) and \(\sigma_\alpha = 0.12 \pm 0.04\). Although the intrinsic scatter is relatively large, \(\sim 40\) per cent, the mean value of the shape parameter is within the range expected for the most massive galaxy clusters at low redshifts. Note that, for this value of \(\alpha\), the density profile is relatively similar to NFW over the range probed (Figure [\[fig:demo\]](#fig:demo){reference-type="ref" reference="fig:demo"}); hence, this result is consistent with the value \(\beta \approx 1\) arrived at when fitting the GNFW model. From their analysis of X-ray data for 11 clusters, found \(0.14<\bar{\alpha}<0.26\) at 95 per cent confidence, with a best-fitting value of \(\sim0.2\). This is somewhat smaller than our result, although we note that their cluster sample is also lower in mass on average (\(M_{200}\sim2\E{14}\ensuremath{\, M_{\odot}}\)), such that these results are not necessarily in conflict. # Conclusion {#sec:conclusions} We present constraints on mass profile models for massive, relaxed galaxy clusters based on an analysis of X-ray data from *Chandra*. This analysis assumes hydrostatic equilibrium between the ICM and the gravitational potential, and includes data only at cluster radii where we are confident that both departures from equilibrium and systematics due to background modeling are minimal. Assuming the NFW mass profile model, the measured concentration--mass relation has a power-law slope with mass of \(\kappa_m=-0.16\pm0.07\), consistent with CDM simulations of cosmological structure formation. The measured relation is consistent with being constant as a function of redshift, a feature that is not seen in simulations; however, this is plausibly the result of our selection of the most dynamically relaxed clusters at a given redshift. Simulations including gas physics, on which the same selection procedure is replicated, would be required to perform a completely fair comparison. We detect an intrinsic scatter of \(\sigma_{\ln c}=0.16\pm0.03\) in the concentration--mass relation. Two clear high-concentration outliers from the mean relation also have the largest central cooling luminosities in the sample, suggesting a role for baryonic physics in the scatter. However, the remaining clusters do not support a simple trend between concentration and cooling luminosity at fixed mass. When fitting a GNFW mass profile, where the slope at small radii is a free parameter, we find an average value that is consistent with the NFW model, \(\bar{\beta}=1.02\pm0.08\). However, there is significant cluster-to-cluster scatter (\(\sigma_\beta=0.22\pm0.07\)), even within the relaxed sample of clusters analyzed here. For the Einasto profile, we similarly measure a mean value, \(\bar{\alpha}=0.29\pm0.04\), which is consistent with CDM predictions for clusters of the mass studied here. In this case, we measure a larger fractional intrinsic scatter, \(\sigma_\alpha=0.12\pm0.04\). Overall, our results confirm that the mass distribution within the most massive halos is in agreement with CDM predictions. There are clear opportunities to improve this analysis in the future, both by finding additional highly relaxed clusters and obtaining deeper X-ray data (especially to better constrain 3-parameter models like the GNFW and Einasto models). Although it is beyond the scope of this work, X-ray data for dynamically relaxed clusters at intermediate radii can potentially be combined with weak lensing (at larger radii, \(\ {\raise-.75ex\hbox{\)\buildrel>\over\sim\(}}\ r_{500}\)), and strong lensing or stellar velocity dispersions (at small radii, \(\ {\raise-.75ex\hbox{\)\buildrel<\over\sim\(}}\ 0.5r_{2500}\)), in the vein of e.g.Ẇhile these additional data sets have their own challenges in the form of projection effects and velocity anisotropy, a careful combination for an appropriately chosen cluster sample potentially provides the most complete possible view of the mass distribution within clusters.
{'timestamp': '2016-07-26T02:17:53', 'yymm': '1607', 'arxiv_id': '1607.04686', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04686'}
null
null
# Introduction ## The notion of projective and affine geometry plays a crucial role in the emerging theory of \(\mathbb{F}_1\), the field with one element. Combinatorially, following Tits, such a projective space \(\mathbb{P}\) of dimension \(m\) is a complete graph on \(m + 1\) vertices endowed with the full subgraph structure. Its automorphism group is isomorphic to the symmetric group \(\texttt{S}_{m + 1}\), and the latter is the Weyl group of the projective general linear group of any projective space of dimension \(m\) over any field. In, Kapranov and Smirnov describe the same theory from the viewpoint of \(\mathbb{F}_1\)-vector spaces, and they end up with \(\texttt{S}_{m + 1}\) as the group of \((m + 1)\times(m + 1)\)-permutation matrices acting as linear automorphisms. In Deitmar's scheme theory, an affine space (of dimension \(m\)) over \(\mathbb{F}_1\) is defined as \(\texttt{Spec}(\mathbb{F}_1[Y_1,\ldots,Y_m])\), where \(\mathbb{F}_1[Y_1,\ldots,Y_m]\) is the free abelian monoid generated by \(Y_1,\ldots,Y_m\) (usually also containing an extra element \(0\)), and the \(\texttt{Spec}\)-construction generalizes naturally from commutative unital rings to commutative multiplicative monoids with an extra absorbing element \(0\). As noticed by the author in (e.g.) \| see also \| there is also a natural \(\texttt{Proj}\)-construction in this theory, and hence \(\texttt{Proj}(\mathbb{F}_1[Y_1,\ldots,Y_{m + 1}])\) is the "Algebro-Geometric version" of \(\mathbb{P}\). Both notions fit perfectly together: the \(m + 1\) closed points of \(\texttt{Proj}(\mathbb{F}_1[Y_1,\ldots,Y_{m + 1}])\) correspond to the vertices of the graph, and linear subspaces of dimension \(1\) indeed contain precisely two closed points. ## In most of the known scheme theories in characteristic \(1\), a scheme essentially consists of a triple \((S,X,\gamma)\), where \(S\) is a Deitmar scheme, \(X\) a Grothendieck scheme, and \(\gamma\) a map \(S \mapsto X\) which comes from a form of base extension from \(\mathbb{F}_1\) to \(\mathbb{Z}\). (See for instance Deitmar, Connes and Consani, and Thas.) As such, Deitmar schemes play a very important role in Absolute Arithmetic. In general, many phenomena which are invisible at the \(\mathbb{F}_1\)-level, only become visible after having applied a base extension functor (say, to a commutative field) \| Manin coined this principle with the term "acquiring flesh." This principle is essential for the present note and we study it in the guise of the "Hidden Points Principle." ## {#issue} In, the authors define the "field extension" \(\mathbb{F}_{1^{\ell}} \Big/ \mathbb{F}_1\) (for any integer \(\ell \geq 1\)) as the cyclic group \(\mu_{\ell}\) with an extra element \(0\), no addition and usual multiplication. In coordinates, elements of the vector space \(V(m + 1,\mathbb{F}_{1^{\ell}})\) are \((m + 1)\)-tuples with at most one nonzero entry \(a \in \mathbb{F}_{1^{\ell}}\); there is one vector (the zero vector), and there are \((m + 1)\ell\) vector lines (which are not necessarily different \| see §§[3.1](#ghost){reference-type="ref" reference="ghost"}), and each contains only the zero vector. Everything passes in a straightforward way to affine spaces, and there arises a multiple occurrence of the aforementioned principle: - according to Deitmar's theory, affine \(\mathbb{F}_1\)-schemes only have one closed point, and the number of lines equals the dimension; - in the \(\texttt{Proj}\)-setting, the topology of \(\texttt{Proj}(\mathbb{F}_1[Y_1,\ldots,Y_{m + 1}])\) and\ \(\texttt{Proj}(\mathbb{F}_{1^{\ell}}[Y_1,\ldots,Y_{m + 1}])\) is the same, so the latter also has \(m + 1\) closed points. On the other hand, according to Kurokawa, a \(\mathbb{Z}\)-scheme \(\mathcal{X}\) is *of \(\mathbb{F}_1\)-type* if it comes with a counting polynomial \(N_{\mathcal{X}}(T) \in \mathbb{Z}[T]\) (such that for any finite field \(\mathbb{F}_q\), we have that \(\vert \mathcal{X}_q \vert = N_{\mathcal{X}}(q)\)). According to Connes and Consani (see also Deitmar and Manin and Marcolli ), we should have that - The equality \(\vert \mathcal{X}_{\mathbb{F}_{1^{\ell}}} \vert = N(\ell + 1)\) holds. So affine and projective spaces over \(\mathbb{F}_{1^{\ell}}\) in Deitmar's setting have a number of hidden points which should be manifested through the counting polynomial, and which are invisible for the Zariski topology at the \(\mathbb{F}_1\)-level. (Put more bluntly: *the combinatorial side and the algebro-geometric side do not agree on field extensions*.) In this note, we present a way to handle this problem, and show that the algebro-geometric side and the combinatorial one actually *do* agree. # The functor of points Let \(k\) be a field, and consider an affine algebraic variety \(V(\mathcal{A})\) with coordinate ring \(A = k[t_1,\ldots,t_m]/\mathcal{A}\). Let \(\gamma: k[t_1,\ldots,t_m] \mapsto A\) be the natural projection with kernel \(\mathcal{A}\). Define \(u_i := \gamma(t_i)\) for all \(i\). If \[\varphi:\ A\ \mapsto\ k\] is any morphism, then \((\varphi(u_1),\ldots,\varphi(u_m))\) is a \(k\)-point of \(V(\mathcal{A})\). Vice versa, let \(\overline{x} = (x_1,\ldots,x_m)\) be any \(k\)-point of \(V(\mathcal{A})\). Then the following map defines a morphism \(A \mapsto k\): \[\varphi_{\overline{x}}:\ f\ \mapsto\ f(\overline{x}).\] ## Correspondence {#correspondence .unnumbered} For any morphism \(\varphi: A \mapsto k\), we have that \[\varphi \ \equiv \ \varphi_{\overline{x}},\] where \(\overline{x} = (\varphi(u_1),\ldots,\varphi(u_m))\).\ This classical insight, which sees \(k\)-points as \(k\)-morphisms, also works for schemes in general. This is the motivation for the definition of *\(\mathbb{F}_{1^\ell}\)-rational point* of an \(\mathbb{F}_1\)-scheme \(\mathcal{X}\), which is used in, e.g.,: \[\mathcal{X}(\mathbb{F}_{1^\ell}) = \texttt{Hom}(\texttt{Spec}(\mathbb{F}_{1^\ell}),\mathcal{X}).\] When \(\mathcal{X} = \texttt{Spec}(A)\) is an affine \(\mathbb{F}_1\)-scheme, we also have \[\mathcal{X}(\mathbb{F}_{1^\ell}) \cong \texttt{Hom}(A,\mathbb{F}_{1^\ell}).\] # Spaces over \(\mathbb{F}_{1^{\ell}}\) {#spa-ext} It is our goal to find a common picture to the different models of affine and projective \(\mathbb{F}_{1^{\ell}}\)-spaces. Our starting point is the projection map \[V(m + 1,\mathbb{F}_{1^{\ell}})\ \ \mapsto\ \ \texttt{Proj}(\mathbb{F}_{1^\ell}[Y_1,\ldots,Y_{m + 1}]).\] Following Kapranov and Smirnov and, \(V = V(m + 1,\mathbb{F}_{1})\) has only one vector (corresponding to the zero vector), and \(m + 1\) directions, which after base extension to a field, become vector lines. Each is given a coordinate of type \((-,1,-)\), and the linear maps act on these directions. Defining \(\mathbf{PG}(m,\mathbb{F}_1)\) from \(V\) in the usual way, each direction defines a projective point and the zero vector gets lost, so that automorphisms of \(V\) carry over in a faithful manner to automorphisms of \(\mathbf{PG}(m,\mathbb{F}_1)\). ## Over \(\mathbb{F}_{1^{\ell}}\) (ghost directions) {#ghost} Kapranov and Smirnov consider all directions of the form \((-,\alpha,-)\) with \(\alpha \in \mu_{\ell}\) to define \(V(m + 1,\mathbb{F}_{1^\ell})\). In fact, as any direction corresponds to a vector line, no new directions are introduced, but rather cyclotomic coordinates get into the picture. Making the transition to projective space as above, we get the same set of points (see also Proposition [\[topell1\]](#topell1){reference-type="ref" reference="topell1"}), as proportionality kills the cyclotomy (on the geometric level). If \(X\) is a \(k\)-scheme, where \(k\) is a field, then the closed points of \(X\) represent orbits of the Galois group \(\texttt{Gal}(\overline{k}/k)\), and all \(\overline{k}\)-rational points of \(X \times_{k}\overline{k}\) are contained in the union of these orbits. So there is a natural map \[\alpha: X(\overline{k}) \longrightarrow X,\] sending closed points of \(X(\overline{k})\) to closed points of \(X\), which is neither injective nor surjective in general. Going back to the spaces \(\mathbf{PG}(n,\mathbb{F}_{1^\ell})\) of above, we see the closed points of \(\mathbf{PG}(n,\mathbb{F}_{1^\ell})\) as orbits of \(\texttt{Gal}(\overline{\mathbb{F}_1}/\mathbb{F}_{1^\ell})\). A stalk at an arbitrary closed point of \(\mathbf{PG}(n,\mathbb{F}_{1^\ell})\) is isomorphic to \[\mu_\ell \times \mathbb{F}_1[X_1,\ldots,X_n].\] So on the algebraic level we can see the extension of the ground field \| we consider the stalk as consisting, besides \(0\), of \(\ell\) distinct copies of \(\mathbb{F}_1[X_1,\ldots,X_n] \setminus \{0\}\) equipped with a sharply transitive \(\mu_\ell\)-action, which is in accordance with the classical picture. Similar conclusions can be made for the affine spaces \(\texttt{Spec}(\mathbb{F}_{1^\ell}[X_1,\ldots,\\ X_n])\) (where the cyclotomic directions are ghost directions defined on the associated vector space, but are not defined in the Deitmar scheme). For any finite field \(\mathbb{F}_p\), where \(p = 1\) is allowed, we have a natural identification \[V(mn,\mathbb{F}_p)\ \overset{\sim}{\mapsto}\ V(m,\mathbb{F}_{p^n}),\] and so we have \[V(m,\mathbb{F}_{1^\ell})\ \overset{\sim}{\mapsto}\ V(m\ell,\mathbb{F}_1)\ \overset{\otimes \mathbb{F}_p}{\mapsto}\ V(m\ell,\mathbb{F}_p)\ \overset{\sim}{\mapsto}\ V(m,\mathbb{F}_{p^\ell}),\] which justifies Kapranov and Smirnov's approach. Here, the base extension "\(\otimes_{\mathbb{F}_1}\mathbb{F}_p\)" can be defined using, e.g., the affine space: with \(A = \mathbb{F}_1[X_1,\ldots,X_{m\ell}]\) we have \[A \otimes_{\mathbb{F}_1}\mathbb{F}_p \ =\ (A \otimes_{\mathbb{F}_1} \mathbb{Z}) \otimes_{\mathbb{Z}} \mathbb{F}_p\ =\ \mathbb{Z}[A] \otimes_{\mathbb{Z}} \mathbb{F}_p,\] where \(\mathbb{Z}[A]\) is the monoidal ring defined by \(A\). ## Hidden Points Principle The following trivial observation reveals the Hidden Points Principle at the topological level. *Proof*.It suffices to observe that for any \(\gamma \in \mu_\ell\) and any \(i \in \{0,1,\ldots,n\}\), we have \[(X_i) = (\gamma X_i),\] whence the topology is the same as (= homeomorphic to) that of \(\mathbf{PG}(n,\mathbb{F}_1)\). \(\blacksquare\)\ Due to this fact, we need to introduce "hidden points," which cannot be detected by the classical way in the Zariski topology. We can formulate the Hidden Points Principle as follows (for a Deitmar scheme \(\mathcal{X}\) defined over \(\mathbb{F}_{1^\ell}\)): \[\vert \texttt{Hom}(\texttt{Spec}(\mathbb{F}_{1^\ell}),\mathcal{X}) \vert\ \gg \ \vert \mathcal{X}_{\mathbb{F}_{1^\ell}} \vert,\] where on the right hand side, \(\mathcal{X}_{\mathbb{F}_{1^\ell}}\) is the set of \(\mathbb{F}_{1^\ell}\)-rational points defined as closed points (i.e., coming from (next-to-) maximal ideals). In fact, the ratio \[f(\mathcal{X},\ell) := \frac{\vert \texttt{Hom}(\texttt{Spec}(\mathbb{F}_{1^\ell}),\mathcal{X})\vert}{\vert \mathcal{X}_{\mathbb{F}_{1^\ell}}\vert}\] is strictly increasing for a fixed \(\mathcal{X}\), as a function of \(\ell \in \mathbb{N} \setminus \{0\}\).\ In this note, we will look for a minimal enlargement of Deitmar schemes which reveals the hidden points in Deitmar schemes such as \(\texttt{Spec}(\mathbb{F}_{1^\ell}[X_1,\ldots,X_m])\) and \(\texttt{Proj}(\mathbb{F}_{1^\ell}[X_1,\ldots,X_m])\), and which resolves the issues of §§[1.3](#issue){reference-type="ref" reference="issue"}. ## The combinatorial picture \| I ### {#section-2} Let \(\Gamma = (V,E)\) be a graph (undirected), with \(V\) the vertex set and \(E\) the edge set. A *cover* of \(\Gamma\) is a graph \(\widetilde{\Gamma} = (\widetilde{V},\widetilde{E})\) together with a surjective graph morphism \[\gamma: \widetilde{\Gamma} \ \mapsto\ {\Gamma}\] such that locally \(\gamma\) is a bijection (that is, for any vertex \(v \in \widetilde{V}\), \(\gamma\) induces a bijection between the edges incident with \(v\) and the edges incident with \(\gamma(v)\)). If there is a positive integer \(m\) such that any vertex of \(\Gamma\) has a fiber of size \(m\), then \((\widetilde{\Gamma},\gamma)\) is an *\(m\)-fold cover*. Now let \(A = (V,E)\) and \(B = (V',E')\) be two graphs. The *lexicographic product* \(A[B]\) of \(A\) and \(B\), also denoted by \(A \cdot B\), is the graph defined on the vertex set \(V \times V'\), such that \((v,v') \sim (w,w')\) if either \(v \sim w\), or (\(v = w\) and \(v' \sim w'\)). We allow \(B\) to be a directed. The lexicographic product \(A \cdot B\) is in general not commutative, and, considered as the morphism defined by \[\pi: A\ \mapsto\ B:\ (v,w) \ \mapsto\ v,\] not a cover of \(A\). On the other hand, each vertex fiber has a constant size \(\vert V' \vert\), and the morphism *is* surjective. Let \(G\) be any group, and \(S\) a (minimal) generating set of \(G\). The *Cayley graph* \(\Gamma(G,S)\) of \(G\) with respect to \(S\) is the directed graph with vertex set \(G\), and directed edges of type \((v,sv)\), with \(s \in S\). (With any fixed \(s\) corresponds a color, but that is not important here.) ## Example {#example .unnumbered} Let \(G = \mu_n\) be any cyclic group, with \(n \in \mathbb{N}^{\times} \cup \{ \vert \mathbb{N} \vert \}\), and let \(S = \{ s\}\), with \(s\) a generator. Then \(\Gamma(\mu_n,S)\) is a directed cycle on \(n\) vertices. ### Combinatorial picture One way of introducing a combinatorial model for \(\mathbb{P}^m(\mathbb{F}_{1^\ell})\) could be to define it as \(K_{m + 1}[\Gamma(\mu_{\ell}),\{s\}]\). This model essentially corresponds to the Kapranov-Smirnov model of vector spaces over extensions of \(\mathbb{F}_1\). ![image](Pic_arxiv) ### Automorphism group The automorphism group of \(K_{m + 1}[\Gamma(\mu_{\ell}),\{s\}]\) is easily seen to be isomorphic to \(\mu_{\ell} \wr \texttt{S}_{m + 1}\). The picture presented in this section is certainly richer than the topological model: it now possesses the right automorphism group, but the counting polynomial is not the predicted one. # Frames {#frame} Let \(S\) be any monoid, written multiplicatively and with \(0\), and \(m \in \mathbb{N}^{\times}\). The *affine frame* \(\mathcal{A}(m,S)\) is defined as the set \(\{ (s_1,\ldots,s_m) \vert s_i \in S \}\). Let \((0,\ldots,0) =: \omega\). If \(S \setminus \{0\}\) is also a group, the *projective frame* \(\mathcal{P}(m-1,S)\) is defined as \(\mathcal{A}(m,S) \setminus \omega\) modulo proportionality (so an element of \(\mathcal{P}(m-1,S)\) is a class \((s_1:\cdots:s_m) := \{ \sigma(s_1,\ldots,s_m) \vert \sigma \in S, \sigma \ne 0 \}\), with \((s_1,\ldots,s_m) \ne \omega\)). If \(m = 0\), \(\mathcal{P}(m-1,S) = \emptyset\); if \(m = 1\), \(\mathcal{P}(m-1,S)\) is a point, and \[\mathcal{P}(1,S) = \{(1: 0),(0: 1)\} \cup \{(1: s) \vert s \in S^{\times}\}.\] ## Frames over \(\mathbb{F}_{1^\ell}\) Let \(\ell \in \mathbb{N}^{\times}\). The affine, respectively projective, frame *over* \(\mathbb{F}_{1^\ell}\) of *dimension* \(d\) is \(\mathcal{A}(d,\mathbb{F}_{1^\ell})\), respectively \(\mathcal{P}(d,\mathbb{F}_{1^\ell})\). (Here, \(d \in \mathbb{N}\), respectively \(d \in \mathbb{N} \cup \{-1\}\).) We will see \(\mathcal{A}(d,\mathbb{F}_{1^\ell})\) and \(\mathcal{P}(d,\mathbb{F}_{1^{\ell}})\) as models of respectively affine and projective spaces over \(\mathbb{F}_{1^\ell}\). ## Subframes An (affine) *subframe* of \(\mathcal{A}(d,\mathbb{F}_{1^{\ell}})\) is any subset of tuples \((x_1,\ldots,x_d)\) in which a fixed number \(e\) of coordinates (\(0 \leq e \leq d\)), are fixed elements of \(\mathbb{F}_{1^\ell}\), and the others take all possible values. Its *dimension* is \(d-e\). Similarly as in the previous subsection, one defines projective subframes of dimension \(d-e-1\) of \(\mathcal{P}(d-1,\mathbb{F}_{1^\ell})\) from affine subframes of dimension \(d-e\) of \(\mathcal{A}(d,\mathbb{F}_{1^\ell})\) endowed with the proportionality relation. ## The combinatorial picture \| II. From spaces to frames Let \(\mathbb{A}^m_{1^\ell} =\\ \texttt{Spec}(\mathbb{F}_{1^\ell}[X_1,\ldots,X_m])\) be the Deitmar scheme of \(m\)-dimensional affine space over \(\mathbb{F}_{1^\ell}\). According to the point of view of morphisms, a (rational) point should correspond to a morphism \[\gamma: \mathbb{F}_{1^\ell}[X_1,\ldots,X_m]\ \mapsto\ \mathbb{F}_{1^\ell}.\] Obviously, any such \(\gamma\) is completely determined by \((\gamma(X_1),\ldots,\gamma(X_m))\), and conversely, any such choice gives us a morphism. So \(\texttt{Hom}(\texttt{Spec}(\mathbb{F}_{1^\ell}),\mathbb{A}_1^m)\) is in a natural bijective correspondence with the affine frame \(\mathcal{A}(m,\mathbb{F}_{1^\ell})\), and \[\vert \texttt{Hom}(\texttt{Spec}(\mathbb{F}_{1^\ell}),\mathbb{A}_1^m)\vert = (\ell + 1)^m.\] Similar remarks hold for projective space Deitmar schemes and projective frames. In the latter case, we have \[\vert \texttt{Hom}(\texttt{Spec}(\mathbb{F}_{1^\ell}),\mathbb{P}_1^m)\vert = \sum_{i = 0}^m(\ell + 1)^i.\] ### Example (case \(\ell = 1\)) The number of rational points of \(\mathcal{P}(d,\mathbb{F}_1)\) is \(2^{d} + 2^{d-1} + \cdots + 1\). Amongst these are the \(d + 1\) closed points of the underlying Deitmar scheme. ## Automorphism group Obviously, since we do not have any addition in \(\mathbb{F}_{1^\ell}\), we have that the automorphism group (defined in a natural way) of \(\mathcal{A}(d,\mathbb{F}_{1^\ell})\), respectively \(\mathcal{P}(d,\mathbb{F}_{1^\ell})\), is \[\mu_\ell \wr \texttt{S}_d, \ \mathrm{respectively}\ \Big(\mu_\ell \wr \texttt{S}_{d + 1}\Big) \Big/ \mathbb{F}_{1^\ell}^\times.\] This picture is very interesting: in case of affine and projective spaces over real fields \(k\), the respective automorphism groups act transitively on the \(k\)-linear subspaces of fixed dimension, and in particular on the \(k\)-points. This is not at all the case here; observe the following. (Denote the affine group by \(\mathbf{AGL}_d(1^\ell)\) and the projective group by \(\mathbf{PGL}_{d + 1}(1^\ell)\).) So in this setting, the original closed Deitmar points form one orbit under the full automorphism group. # Frames and enlarged Deitmar schemes As we have seen, the data \((K_{m}[\Gamma(\mu_{\ell},S)],\mathcal{P}(m-1,\mathbb{F}_{1^\ell}))\) reveals a number of hidden properties of (say) projective spaces \(\texttt{Proj}(\mathbb{F}_{1^\ell}[X_1,\ldots,X_m])\). As the points of an \(\mathbb{F}_{1^\ell}\)-frame naturally correspond to \(\mathbb{F}_{1^\ell}\)-rational points of the corresponding space, we want to enlarge Deitmar schemes such that points of the frames associated to affine/projective/etc. spaces appear as closed points. Those can be seen if we (minimally) enlarge the category of Deitmar schemes in the following way, by using congruences. ## Congruences and congruence ideals Let \(M\) be a multiplicative commutative monoid. A *congruence* \(\mathcal{C} \subseteq M \times M\) on \(M\) is an equivalence relation with the additional property that if \((u,v) \in \mathcal{C}\), then \((mu,mv) \in \mathcal{C}\) for each \(m \in M\). If \((u,v) \in \mathcal{C}\), we also write \(u \sim v\). If \(\mathcal{C}\) is a congruence on \(M\), one shows that \(M/\mathcal{C}\) (or \(M/\sim\)) naturally inherits the structure of a monoid. Note that in general the union of two congruences \(\mathcal{C}, \mathcal{C}'\) is not a congruence (as transitivity does not necessary hold anymore). (The congruence generated by \(\mathcal{C}, \mathcal{C}'\) is essentially the transitive closure.) The intersection *is* a congruence, though. Any ideal \(I\) in the monoid \(M\) gives rise to a congruence, called "Reese congruence\" and denoted by \(\mathcal{C}_I\); if \(m \in M\setminus I\), the class \([m]\) is just \(\{m\}\); if \(a, b \in I\), then \([a] = [b]\). So \(I\) defines one class, which we will also denote by \([I]\). If \(M\) has an absorbing element \(0\), then \(0 \in I\), so in that case \(a \sim 0 \sim b\). From now on we keep assuming that \(M\) has a \(0\). A congruence \(\mathcal{C}\) is *prime* if from \((ab,ac) \in \mathcal{C}\) follows that either \((b,c) \in \mathcal{C}\) or \((a,0) \in \mathcal{C}\). (Alternatively, a congruence \(\mathcal{C}\) on \(M\) is *prime* if \(M/\mathcal{C}\) is integral, i.e., if it satisfies the cancellation property.) Now let \(\frak{p}\) be a prime ideal in the monoid \(M\), so that \(0 \in \frak{p}\). Let \(a, b, c\) be elements in \(M\), and suppose \(ab = ac\) in \(M/\mathcal{C}_{\frak{p}}\). If \(ab \sim ac \in \frak{p}\), then either \(a \in \frak{p}\), or \(a \not\in \frak{p}\) and \(b, c \in \frak{p}\). (If \(ab = ac \not\in \frak{p}\), then \(a, b, c \ne 0\).) If \(M\) is supposed to be integral, then \(\mathcal{C}_\frak{p}\) is a prime ideal of \(M\). ## Example: affine space Consider \(\texttt{Spec}(\mathbb{F}_{1^\ell}[X_1,\ldots,X_m])\), Deitmar version. Put \(A := \mathbb{F}_{1^\ell}[X_1,\ldots,X_m]\). If \(\frak{p}\) is a prime ideal, then \(\mathcal{C}_{\frak{p}}\) is a prime congruence since \(A\) is integral. ## New Zariski topology \(\texttt{Spec}^c(M)\) In Deitmar scheme theory, any set of elements \(S\) in \(\mathbb{F}_1[X_1,\ldots,X_m]\) defines an ideal \(I(S)\), and the closed set corresponding to \(I(S)\) classically contains the solutions of all the equations defined by its elements. Only equations of the form \(P = 0\) arise, with \(P\) an element in \(\mathbb{F}_1[X_1,\ldots,X_m]\). We need to allow equations of the form \(P = 1\), and more generally, over \(\mathbb{F}_{1^\ell}\), of the form \(P = \mu\). Let \(\texttt{Spec}^c(M)\) be the set of prime congruences on \(M\). Then the Zariski topology it comes with (and denoted in the same way) is defined by the closed sets \[C(I) := \{ \frak{p}\ \vert\ I \subseteq \frak{p}, \frak{p}\ \text{prime congruence} \},\] with \(I\) any congruence on \(M\). The key is that the closed set topology is generated by the sets \[\mathcal{D}(a,b) := \{ \frak{p}\ \vert\ (a,b) \in \frak{p}, \frak{p}\ \text{prime congruence} \}.\] ## Localization Let \(\frak{p}\) be any prime congruence. Define \(S_{\frak{p}}\) as the monoid multiplicatively generated by the set \[\{ b-a\ \vert\ (a,b) \in A \times A,\ a \not\sim_{\frak{p}} b \}.\] Then \(A_{\frak{p}} := S_{\frak{p}}^{-1}A\) is defined as the monoid naturally generated by the fractions \(\{ \frac{u}{v}\ \vert\ u \in A, v \in S_{\frak{p}} \}\). Here, we identify \(u/v\) with \(u'/v'\) if there is a \(w \in S_{\frak{p}}\) such that \(wuv' = wu'v\). ## Sections and structure sheaf Let \(U\) be an open set. Then \(\mathcal{O}_X(U)\) is defined as the set of maps \[s: U \ \mapsto\ \coprod_{\frak{p} \in U}A_{\frak{p}}\] such that for each prime \(\frak{p} \in U\), \(s(\frak{p}) \in A_{\frak{p}}\), and for which there is a neighborhood \(V\) of \(\frak{p}\) in \(U\) such that for all \(\frak{q} \in V\), \(s(\frak{q})\) is locally a fraction. This means that there are \(a \in A\) and \(f \in \bigcap_{\frak{q} \in V} S_{\frak{q}}\) so that \(s = a/f\) over \(V\). Note that \(\mathcal{O}_X(U)\) carries the structure of a monoid. ## Example: maximal ideals (closed points) for affine spaces Consider again \(A := \mathbb{F}_{1^\ell}[X_1,\ldots,X_m]\). For each coordinate monomial \(P\) and each \(\mu \in \mathbb{F}_{1^\ell}\), we consider the *base congrence* on \(A\) generated by \(P \sim \mu\). If, for example, \(P = X_j\), the congruence is \[\{ (QX_j^n,Q\mu^n) \vert Q \in A, n \in \mathbb{N}\} \cup \{ (Q\mu^n,QX_j^n) \vert Q \in A, n \in \mathbb{N} \}.\] It is easy to see that the maximal congruences are precisely those generated by base congruences of the following form \[\{ (X_i,\mu_i)\ \vert\ i \in \{1,\ldots,n \}, \mu_i \in \mathbb{F}_{1^\ell} \}.\] In other words, the maximal congruences correspond to the rational points/frame points of \(\mathcal{A}(m,\mathbb{F}_{1^\ell})\) as described in §[4](#frame){reference-type="ref" reference="frame"}. ## General congruence schemes and \(\texttt{Proj}^c\)-schemes For the theory of general congruence schemes and more details, we refer to Deitmar's paper. As for the \(\texttt{Proj}^c\)-construction, this is tersely described in in the case of projective spaces. One defines the *irreducible congruence* \(\texttt{Irr}^c\) on \(A = \mathbb{F}_{1^\ell}[X_1,\ldots,X_m]\) as \[\texttt{Irr}^c := \Big\langle X_1 \sim 0, \ldots, X_m \sim 0 \Big\rangle,\] leaves it out, and applies a construction which is similar to that in Deitmar scheme theory. The next-to-maximal congruences in \(A\) correspond to the closed points, and they correspond to the elements of the projective frame \(\mathcal{P}(m-1,\mathbb{F}_{1^\ell})\). ## Maximal points vs. closed points In the congruence setting of this section, one observes that for affine and projective spaces the set of rational points *coincides* with the set of closed points, a phenomenon which for proper fields only occurs for algebraically closed fields. ## Absolute and geometric \(\mathbb{F}_1\)-Frobenius endomorphisms Consider the algebraic closure \(\overline{\mathbb{F}_1}\) of \(\mathbb{F}_1\); by definition, it consists of all complex roots of unity, together with an absorbing element \(0\), foreseen with multiplication. We define the *\(\mathbb{F}_1\)-Frobenius endomorphism* of degree \(n \in \mathbb{N}\), denoted \(\mathrm{Fr}^n_1\), to be the map \[\mathrm{Fr}_1^n: \overline{\mathbb{F}_1} \longrightarrow \overline{\mathbb{F}_1}: x \longrightarrow x^n.\] Elements of \(\mathbb{F}_{1^{d}} \cong \mu_d \cup \{0\} \leq \overline{\mathbb{F}_1}\) are characterized by the fact that they are the solutions of \[{\mathrm{Fr}_1^{d + 1}}(x) = x,\] which is analogous to the fact that elements of finite fields \(\mathbb{F}_{q^d} \leq \overline{\mathbb{F}_q}\) are singled out as fixed points of \(\mathrm{Fr}^d\). The absolute \(\mathbb{F}_1\)-Frobenius map \(\mathrm{Fr}_1^{d + 1}\) with \(d \in \mathbb{N}^\times\) acts on an extended Deitmar scheme \(\mathcal{Y}\) of finite type over \(\mathbb{F}_{1^d}\) by acting trivially on the topology, and as \(\mathrm{Fr}_1^{d + 1}\) on the structure scheaf. Passing to the scheme \(\overline{\mathcal{Y}} := \mathcal{Y} \otimes_{\texttt{Spec}(\mathbb{F}_{1^d})}\texttt{Spec}(\overline{\mathbb{F}_1})\), we define the *geometric \(\mathbb{F}_1\)-Frobenius map* (still in the same degree) as the map \(\mathrm{Fr}_{\mathrm{geom}}^{d + 1}: \overline{\mathcal{Y}} \mapsto \overline{\mathcal{Y}}\) with local comorphisms \(y \otimes f \mapsto y^{d + 1} \otimes f\). It acts as \[\mathbf{x} \mapsto \mathbf{x}^{d + 1}\] on frame points. # Zeta functions In the inspiring note, Kurokawa says that a \(\mathbb{Z}\)-scheme \(\mathcal{X}\) is of *\(\mathbb{F}_1\)-type* if there exists a polynomial \(N_\mathcal{X}\) in \(\mathbb{Z}[X]\) such that, for each prime power \(q\), we have \[\vert \mathcal{X}_q \vert = N_{\mathcal{X}}(q).\] (Through one of Tate's conjectures, such schemes would come with a mixed Tate motive, a feature which one sees more clearly in the appropriate Grothendieck ring \| see for instance for a detailed discussion.) In fact, originally, in, a \(\mathbb{Z}\)-scheme (of finite type) is of \(\mathbb{F}_1\)-type if its arithmetic zeta function \(\zeta_{\mathcal{X}}(s)\) can be expressed in the form \[\zeta_{\mathcal{X}}(s) = \prod_{k = 0}^n\zeta(s-k)^{a_k}\] with the \(a_k\)s in \(\mathbb{Z}\). (Here, \(s\) is a complex variable \| cf. the remark right after the statement of Theorem [\[Kuro\]](#Kuro){reference-type="ref" reference="Kuro"}.) The definition through the counting polynomial then is derived from the following result. (In the next theorem, \(\zeta(\cdot)\) is the classical Riemann zeta.) For details about matters of convergence, we refer to Theorem [\[eqext\]](#eqext){reference-type="ref" reference="eqext"} below. Kurokawa defines the *\(\mathbb{F}_1\)-zeta function* of a \(\mathbb{Z}\)-scheme \(\mathcal{X}\) of \(\mathbb{F}_1\)-type as \[\zeta_{\mathcal{X}\vert \mathbb{F}_1}(s) := \prod_{k = 0}^n(s-k)^{-a_k}\] with the \(a_k\)s as above. The *Euler characteristic* is \[\vert X_{\mathbb{F}_1} \vert := \sum_{k = 0}^na_k = N_{\mathcal{X}}(1).\] To fit this formula into **Count**, put \[\mathbb{F}_{1^0} = \mathbb{F}_{1^1}.\] (Although this seems rather artificial, it is also rather interesting that one makes **Count** work by putting \(0 = 1\)!) ## Schemes of \(\mathbb{F}_1\)-type vs. schemes defined over \(\mathbb{F}_1\) {#VS} As Deitmar envisioned in, commutative monoids with a \(1\) and \(0 \ne 1\) should/could be seen as "\(\mathbb{F}_1\)-algebras." For each such monoid \(A\), we have an embedding \(\mathbb{F}_1 \hookrightarrow A\), so a projection \(\texttt{Spec}(A) \twoheadrightarrow \texttt{Spec}(\mathbb{F}_1)\). So \(\texttt{Spec}(A)\) is "defined over \(\mathbb{F}_1\)." (And one can make similar observations for general monoidal schemes.) Considering a \(k\)-algebra \(\mathcal{A}\) for a field \(k\), one can realize \(\mathcal{A}\) as a quotient of a polynomial ring over \(k\) with the kernel of an appropriate surjective morphism to \(\mathcal{A}\). In particular, if \[\iota: k\ \hookrightarrow\ R\] is an embedding of a field in a unital commutative ring, then \(R\) comes with a \(k\)-algebra structure, so \(R\) is a quotient of a polynomial ring over \(k\). And any unital commutative ring is a \(\mathbb{Z}\)-algebra, of course. Over \(\mathbb{F}_1\), that is to say, in the setting of monoidal schemes, this situation is less natural, as some unital commutative monoids with \(0\) can only be realized as a quotient of a polynomial ring over \(\mathbb{F}_1\) and a *congruence* (instead of a monoidal ideal). For instance, although we have an embedding \(\mathbb{F}_1 \hookrightarrow \mathbb{Z}\), the latter, as a monoid, *needs* a congruence which is *not* associated to a monoidal ideal if one wants to see it as such a quotient. Whereas over \(\mathbb{F}_{1^2}\), \(\mathbb{Z}\) is indeed even *isomorphic* to a polynomial ring. So it seems more natural to say that \(\texttt{Spec}(\mathbb{Z})\) is "defined over \(\mathbb{F}_{1^2}\)," instead than over \(\mathbb{F}_1\). But still, strictly speaking, if \(\texttt{Spec}(A)\) is defined over \(\mathbb{F}_{1^\ell}\), it should also be defined over \(\mathbb{F}_1\) as well. With this discussion in mind, one can then decide when a \(\mathbb{Z}\)-scheme is defined over \(\mathbb{F}_1\), or an extension. One should make a distinction in any case with Kurokawa's \(\mathbb{Z}\)-schemes "of \(\mathbb{F}_1\)-type." If \(\mathcal{X}\) is a \(\mathbb{Z}\)-scheme of \(\mathbb{F}_1\)-type à la Kurokawa, then any reasonable definition in the same vein of \(\mathbb{Z}\)-schemes over extensions of \(\mathbb{F}_1\) will lead to the property that "of \(\mathbb{F}_{1^m}\)-type" implies "of \(\mathbb{F}_{1^n}\)-type" if \(m\) divides \(n\), but not necessarily the other way around. (This phenomenon will manifest itself in the next subsection.) So the situation is quite different (and more general) than the \(\mathbb{F}_1\)-algebra viewpoint. ## Dirichlet series A *Dirichlet series* is a series of the form \[D(s) = \sum_{m = 1}^\infty\gamma(m)m^{-s}\] with \(s\) a complex variable and \(\gamma: \mathbb{N} \to \mathbb{C}\) a function. It can be shown that if \(\sum_{m = 1}^\infty\vert \gamma(s)m^{-s}\vert\) does not diverge for all \(s\), there exists a real number \(\kappa\) such that \(\sum_{m = 1}^\infty\vert \gamma(s)m^{-s}\vert\) converges for all \(s\) with \(\Re(s) > \kappa\) (that is, the Dirichlet series is absolutely convergent for these values). (The real part of a complex number \(c\) is denoted by \(\Re(c)\).) We will use the next well known result later on. ## Schemes of \(\mathbb{F}_{1^\ell}\)-type Let \(V\) be nonsingular projective variety over \(\mathbb{F}_q\), the finite field with \(q\) elements. The *local zeta function* (or *congruence zeta function*) of \(V\) is given by \[\zeta_V(s) = Z(V,s)\ := \ \exp\Big(\sum_{m = 1}^\infty \frac{\vert V_{q^m} \vert}{m}q^{-sm}\Big).\] Sometimes \(Z(V,s)\) is also denoted by \(Z(V,q^{-s})\). The variable transformation \(u = q^{-s}\) gives \(Z(V,u)\) as a formal power series in \(u\). To stress that the local zeta function is expressed "with respect to \(\mathbb{F}_q\)," we also write \(\zeta_{V \vert \mathbb{F}_q}(s)\) instead of \(\zeta_{V}(s)\). Let \(\ell\) be a positive nonzero integer. We will say that a \(\mathbb{Z}\)-scheme \(\mathcal{X}\) is *of \(\mathbb{F}_{1^\ell}\)-type* if there exists a polynomial \(N_{\mathcal{X}}(Y) = \sum_{k = 0}^na_kY^k\) such that for all primes \(p\) and all extensions \(\mathbb{F}_{p^{\ell m}}\) of \(\mathbb{F}_{p^\ell}\), we have \[\vert \mathcal{X}_{\mathbb{F}_{p^{\ell m}}} \vert = N_{\mathcal{X}}(p^{\ell m}).\] In the next theorem, for a complex number \(c\), the modulus will be denoted by \(\vert c\vert\). Also, \(\log(\cdot)\) will denote the principal value complex logarithm. *Proof*.We first show that (ii) and (iii) are equivalent. Observe that, with \(q = p^\ell\), \[\label{first} \log(\zeta_{\mathcal{X}\vert \mathbb{F}_{q}}(s)) = \ \sum_{m = 1}^\infty \frac{\vert \mathcal{X}_{q^m} \vert}{m}q^{-sm}\] while \[\begin{aligned} \label{second} \log\Big(\prod_{k = 0}^n(1-p^{\ell(k-s)})^{-a_k}\Big) &= \sum_{k = 0}^n-a_k\log(1-p^{\ell(k-s)}) \nonumber \\ &= \sum_{k = 0}^na_k\Big(\sum_{m = 1}^{\infty}\frac{p^{\ell(k-s)m}}{m} \Big) \nonumber \\ &= \sum_{m = 1}^\infty \frac{\Big(\sum_{k = 0}^na_k{p^{\ell mk}}\Big)}{m}q^{-sm}. \end{aligned}\] (Note that we can indeed expand \(\log(1-p^{\ell(k-s)})\) in its Mercator series since \(\vert p^{\ell(k-s)} \vert < 1\).) Then assuming (ii) (equality of the expressions in ([\[first\]](#first){reference-type="ref" reference="first"}) and ([\[second\]](#second){reference-type="ref" reference="second"})), equality of the coefficients in both sides leads to (iii). Also, after assuming (iii), substituting \(\vert \mathcal{X}_{\mathbb{F}_{p^{\ell m}}} \vert = N_{\mathcal{X}}(p^{\ell m})\) for all \(p^{\ell m}\) in ([\[first\]](#first){reference-type="ref" reference="first"}) leads to (ii) (using ([\[second\]](#second){reference-type="ref" reference="second"})). Now assume (ii). Then \[\begin{aligned} \prod_{\text{\(p\) prime}}\zeta_{\mathcal{X}\vert\mathbb{F}_{p^\ell}} &= \prod_{\text{\(p\) prime}}\Big(\prod_{k = 0}^n(1-p^{\ell(k-s)})^{-a_k}\Big) \nonumber \\ &= \prod_{k = 0}^n\Big(\prod_{\text{\(p\) prime}}(1-p^{\ell(k-s)})^{-a_k}\Big) \nonumber \\ &= \prod_{k = 0}^n\zeta(\ell(s-k))^{a_k}, \end{aligned}\] which yields (i).\ Then assume (i). We obtain \[\begin{aligned} \prod_{\text{\(p\) prime}}\zeta_{\mathcal{X}\vert \mathbb{F}_p^\ell} &= \prod_{k = 0}^n\zeta(\ell(s-k))^{a_k} \nonumber \\ &= \prod_{k = 0}^n\Big(\prod_{\text{\(p\) prime}}(1-p^{\ell(k-s)})^{-a_k}\Big) \nonumber \\ &= \prod_{\text{\(p\) prime}}\Big(\prod_{k = 0}^n(1-p^{\ell(k-s)})^{-a_k}\Big). \end{aligned}\] Taking complex logarithms of both sides yields: \[\begin{aligned} \sum_{\text{\(p\) prime}}\Big(\sum_{m = 1}^\infty \frac{\vert \mathcal{X}_{q^m} \vert}{m}q^{-sm}\Big) &= \sum_{\text{\(p\) prime}}\Big(\sum_{k = 0}^n-a_k\log(1-p^{\ell(k-s)})\Big) \nonumber \\ &= \sum_{\text{\(p\) prime}}\bigg(\sum_{k = 0}^na_k\Big(\sum_{m = 1}^{\infty}\frac{p^{\ell(k-s)m}}{m} \Big)\bigg). \end{aligned}\] It follows that \[\begin{aligned} \sum_{p^m\ \text{prime power}}\frac{\vert \mathcal{X}_{p^{\ell m}} \vert}{m}p^{-\ell ms} &= \sum_{p^m\ \text{prime power}}\Big(\sum_{k = 1}^n a_kp^{\ell mk} \Big)p^{-\ell ms}. \end{aligned}\] The uniqueness of the coefficients in the Dirichlet series lead to (iii). \(\blacksquare\)\ Define the *\(\mathbb{F}_{1^\ell}\)-zeta function* of a \(\mathbb{Z}\)-scheme \(\mathcal{X}\) of \(\mathbb{F}_{1^\ell}\)-type as \[\zeta_{\mathcal{X}\vert \mathbb{F}_{1^\ell}}(s) := \prod_{k = 0}^n(\ell(s-k))^{-a_k}\] with the \(a_k\)s as in the previous theorem. Similarly as Theorem [\[Kurolimit\]](#Kurolimit){reference-type="ref" reference="Kurolimit"}, we have: *Proof*.We have \[\zeta_{X\vert \mathbb{F}_{1^\ell}}(s)(p-1)^{\# X(\mathbb{F}_1)} = \prod_{k = 0}^n\big(\frac{1-p^{\ell(k-s)}}{p-1}\big)^{-a_k}\] so that for the limit we get \[\begin{aligned} \lim_{p \longrightarrow 1}\zeta_{X\vert \mathbb{F}_{1^\ell}}(s)(p-1)^{\# X(\mathbb{F}_1)} &= &\prod_{k = 0}^n\big(\lim_{p \longrightarrow 1}\frac{1-p^{\ell(k-s)}}{p-1}\big)^{-a_k}\nonumber \\ &= &\prod_{k = 0}^n(\ell(s-k))^{-a_k}\nonumber \\ &= &\zeta_{X\vert \mathbb{F}_{1^\ell}}(s). \end{aligned}\] \(\blacksquare\) # Schemes of \(\mathbb{F}_1\)-type versus schemes of \(\mathbb{F}_{1^\ell}\)-type In general, \(\mathbb{Z}\)-schemes of \(\mathbb{F}_{1^\ell}\)-type are not necessarily of \(\mathbb{F}_{1^m}\)-type, with \(m\) a natural divisor of \(\ell\) (cf. Remark [6.1](#VS){reference-type="ref" reference="VS"}). In particular, they need not be of \(\mathbb{F}_1\)-type, as the following example shows. ## {#section-3} Let \(X^2 + uX + v \in \mathbb{Z}[X,X_2,\ldots,X_m]\) and \(u\) odd, \(\mathrm{gcd}(u,v) = 1\). It is easy to see that over \(\mathbb{F}_{p^2}\), \(p\) any prime, the number of points is \(2(p^2 + 1)^m\). Taking \((u,v)\) such that \(X^2 + uX + v\) is not reducible over every prime field \(\mathbb{F}_p\), we obtain a scheme of \(\mathbb{F}_{1^2}\)-type which is not of \(\mathbb{F}_1\)-type. Its \(\mathbb{F}_{1^2}\)-counting polynomial is \[2(Y + 1)^m = 2\sum_{k = 0}^m\begin{pmatrix}m \\ k \end{pmatrix}Y^{m-k},\] so its \(\mathbb{F}_{1^2}\)-zeta function is \[\prod_{k = 0}^m(2(s-k))^{-2\begin{pmatrix}m \\ k \end{pmatrix}}.\] ## {#section-4} Let \(\mathcal{X}\) be a \(\mathbb{Z}\)-scheme of finite type defined over \(\mathbb{F}_1\). By definition, it means that \(\mathcal{X}\) is a \(\mathbb{Z}\)-scheme of finite type which arises from Deitmar base extension to \(\mathbb{Z}\) from a Deitmar scheme. Then Deitmar showed in that there exists a natural number \(e\) and a polynomial \(N(Y) \in \mathbb{Z}[Y]\) such that for every prime power \(q\) for which \(\mathrm{gcd}(q-1,e) = 1\), one has \(\#\mathcal{X}(\mathbb{F}_q) = N(q)\). It might be interesting to consider \(\mathbb{Z}\)-schemes \(\mathcal{X}\) which come with a counting polynomial \(N_{\mathcal{X}}(Y)\) such that for some positive integer \(\ell\), we have that \[\#\mathcal{X}(\mathbb{F}_{p^\ell}) = N_{\mathcal{X}}(p^\ell),\] where \(p\) varies over the primes, except possibly a *finite number of exceptions*. (When \(\ell = 1\), such schemes are related to mixed Tate motives.)\ Constructing examples is easy. Consider, for example, the affine conics with equation \[\mathcal{C}(u,v): X^2-uY^2 = v,\] with \(u, v \in \mathbb{Z}\) (and \(u, v\) fixed). Over a finite field \(\mathbb{F}_{2^n}\) we have that \(\mathcal{C}(u,v)\) has \(2^n + 1\) points. If \(p\) is an odd prime, we have that \(\vert \mathcal{C}(u,v)\vert_{\mathbb{F}_{p^n}}\) is \(p^n-1\) if \(u\) is a square, and \(p^n + 1\) if \(u\) is not. If \(u\) would be a square in \(\mathbb{F}_p\), it is also a square in \(\mathbb{F}_{p^2}\), so \(\vert \mathcal{C}(u,v) \vert_{\mathbb{F}_p} = p-1\) and \(\vert \mathcal{C}(u,v) \vert_{\mathbb{F}_{p^2}} = p^2-1\). If \(u\) is not a square in \(\mathbb{F}_p\), we have \(\vert \mathcal{C}(u,v)\vert_{\mathbb{F}_p} = p + 1\), but in the quadratic extension \(\mathbb{F}_{p^2}\) we know that \(u\) *is* a square, so that \(\vert \mathcal{C}(u,v) \vert_{\mathbb{F}_{p^2}} = p^2-1\). In other words, we have a fixed counting polynomial for all fields \(\mathbb{F}_{p^2}\) (\(p\) any prime) in odd characteristic, but not necessarily over \(\mathbb{F}_p\). ## {#section-5} The *Grothendieck ring of schemes* of finite type over a field \(k\), denoted as \(K_0(\texttt{Sch}_{k})\), is generated by the isomorphism classes of schemes \(\mathcal{X}\) of finite type over \(\mathbb{F}_1\), \([\mathcal{X}]_{k}\), with the relation \[_{k}= [\mathcal{X}\setminus \mathcal{Y}]_{k} + [\mathcal{Y}]_{k}\] for any closed subscheme \(\mathcal{Y}\) of \(\mathcal{X}\) and with the product structure given by \[_{k}\cdot[\mathcal{Y}]_{k}= [\mathcal{X}\times_{k}\mathcal{Y}]_{k}.\] If it is clear what \(k\) is, we will not write the index. Denote by \({\mathbb{L}} = [\mathbb{A}^1_{k}]\) the class of the affine line over \(k\). By the first defining property, one obtains typical decompositions such as \[_k = \mathbb{L}^n + \cdots + \mathbb{L} + 1.\] Let \(\texttt{Sch}_{\mathbb{F}_{q}}\) be the category of schemes of finite type over the finite field \(\mathbb{F}_{q}\). An *additive invariant* (or "multiplicative Euler-Poincaré characteristic") from \(\texttt{Sch}_{\mathbb{F}_{q}}\) to \(\mathbb{Z}\) is a map \(\alpha\) which satisfies the following properties: - if \(X, Y\) are in \(\texttt{Sch}_{\mathbb{F}_{q}}\) and \(X \cong Y\), then \(\alpha(X) = \alpha(Y)\); - if \(X\) is in \(\texttt{Sch}_{\mathbb{F}_{q}}\) and \(C\) is a closed subscheme, then \(\alpha(X) = \alpha(C) + \alpha(X \setminus C)\); - if \(X, Y\) are in \(\texttt{Sch}_{\mathbb{F}_{q}}\), then \(\alpha(X \times Y) = \alpha(X)\cdot\alpha(Y)\). An additive invariant \(\alpha\) factors uniquely through a ring morphism from \(K_0(\texttt{Sch}_{\mathbb{F}_{q}})\) to \(\mathbb{Z}\) by letting \[\alpha(\sum_i[X_i]) = \sum_i\alpha[X_i].\] An element \(\gamma\) in \(K_0(\texttt{Sch}_{\mathbb{F}_q})\), \(\mathbb{F}_q\) a finite field, is said to be *polynomial-count* (according to ) if there exists a polynomial \(P(Y) = \sum_ia_iY^i \in \mathbb{C}[Y]\) such that for every finite extension \(\mathbb{F}_{q^n}\Big/\mathbb{F}_q\), we have that \(\vert \gamma(\mathbb{F}_q) \vert_{\mathbb{F}_{q^n}} = P(q^n)\). By, it follows that then \(P(Y) \in \mathbb{Z}[Y]\). Note that \(\vert \cdot \vert_{\mathbb{F}_{q^n}}\) is well defined on \(K_0(\texttt{Sch}_{\mathbb{F}_{q^n}})\) by extension from the function \(\vert \cdot \vert_{\mathbb{F}_{q^n}}\) on \(\texttt{Sch}_{\mathbb{F}_{q^n}}\), using the fact that the latter function is an additive invariant from \(\texttt{Sch}_{\mathbb{F}_{q^n}}\) to \(\mathbb{Z}\). So in terms of, by Theorem [\[eqext\]](#eqext){reference-type="ref" reference="eqext"} a \(\mathbb{Z}\)-scheme \(\mathcal{X}\) is of \(\mathbb{F}_{1^\ell}\)-type precisely if \([\mathcal{X}_{\mathbb{F}_{p^\ell}}] \in K_0(\texttt{Sch}_{\mathbb{F}_{p^\ell}})\) is polynomial-count for all primes \(p\), where the corresponding counting polynomials *are the same for all primes \(p\)*. Also in terms of, this means that \([\mathcal{X}_{p^\ell}]\) is *zeta equivalent* to an element \(P_{\mathcal{X}}\) in \(\mathbb{Z}[\mathbb{L}]\) for all primes \(p\), where \(P_{\mathcal{X}}\) is independent of the prime \(p\). # Examples For affine and projective spaces, we obtain the following zeta functions (over \(\mathbb{F}_p^\ell\) and \(\mathbb{F}_{1^\ell}\), with \(\ell \in \mathbb{N}^\times\)): \[\left\{\begin{array}{ccc} \zeta_{\mathbb{A}^n\vert \mathbb{F}_{p^\ell}}(s) &= &\Big({1-p^{\ell(n-s)}}\Big)^{-1};\\ & & \\ \zeta_{\mathbb{A}^n\vert \mathbb{F}_{1^\ell}}(s) &= &\Big({\ell(s-n)}\Big)^{-1}, \end{array}\right.\] and \[\left\{ \begin{array}{ccc} \zeta_{\mathbb{P}^n\vert \mathbb{F}_{p^\ell}}(s) &= &\Big({(1-p^{-\ell s})(1-p^{\ell(1-s)})\cdots(1-p^{\ell(n-s)})}\Big)^{-1}; \\ & & \\ \zeta_{\mathbb{P}^n\vert \mathbb{F}_{1^\ell}}(s) &= &\Big({\ell^{n + 1}s(s-1)\cdots(s-n)}\Big)^{-1}. \end{array}\right.\] We provide some more examples. In, the authors have defined and studied a functor \(\mathcal{F}_k\) for each field finite \(k\), including \(\mathbb{F}_1\), which maps any "loose graph" \(\Gamma\) (which is a more general version of a graph, where edges with \(0\) or \(1\) vertices are allowed) to a \(k\)-scheme \(\mathcal{F}(\Gamma) \otimes_{\mathbb{F}_1}k\). Here, \(\mathbb{F}_1\)-schemes are Deitmar schemes with the additional minimal congruences considered in this paper. It is shown in that any such scheme comes with a counting polynomial, which is in fact independent of the field. So these schemes are of \(\mathbb{F}_1\)-type, so also of \(\mathbb{F}_{1^\ell}\)-type for any positive integer \(\ell\). For loose trees, precise calculations are made in to obtain the counting polynomials. The outcome is as follows. Let \(\Gamma\) be a loose tree. We use the following notation: - \(D\) is the set of degrees \(\{d_1, \ldots, d_m \}\) of \(V(\Gamma)\) such that \(1 < d_1 < d_2 < \ldots < d_m\); - \(n_i\) is the number of vertices of \(\Gamma\) with degree \(d_i\), \(1\leq i \leq m\); - \(\displaystyle I= \sum_{i=1}^m n_i-1\); - \(E\) is the number of vertices of \(\Gamma\) with degree \(1\). Then, \[\big[\Gamma\big]_{_{\mathbb{F}_1}} = \displaystyle\sum_{i = 1}^m n_i\underline{\mathbb{L}}^{d_i}-I\cdot\underline{\mathbb{L}} + I + E.\] This formula is expressed in terms of the Grothendieck ring of \(\mathbb{F}_1\)-schemes of finite type (see ). The symbol \(\underline{\mathbb{L}}\) denotes the class of the affine line over \(\mathbb{F}_1\). (The same polynomial arises after base extension to fields.) For *each* positive integer \(\ell\), the \(\mathbb{F}_{1^\ell}\)-zeta function is thus given by \[\zeta^{\mathbb{F}_{1^\ell}}_{\Gamma}(s)\ \ =\ \ \frac{(\ell(s-1))^I}{(\ell s)^{E + I}}\cdot\displaystyle \prod_{k = 1}^m(\ell(s-k))^{-n_k}.\] If \(D = \{d_1 = d \}\), \(d > 1\) and \(n_1 = 1\), then \(I = 0\) and \(E = 0\). This is the loose graph of affine \(d\)-space, and we obtain the \(\mathbb{F}_{1^\ell}\)-zeta function for \(\mathbb{A}^d\) as above. (For the affine line, put \(D = \emptyset\), \(I =-1\) and \(E = 1\).)
{'timestamp': '2016-07-18T02:07:18', 'yymm': '1607', 'arxiv_id': '1607.04513', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04513'}
null
null
# Introduction Suppose that \(\ensuremath{\textup{\textsf{F}}} = \ensuremath{\textup{\textsf{F}}}_q\) is a finite field with size \(q\). Let \(m\) and \(n\) be integers such that \(0 \le m \le n\). The paper that introduced Wiedemann's algorithm  also includes a proof of the following claim---which concerns an \(n \times n\) matrix obtained by appending an additional set of row vectors to a matrix \(A \in \matgrp{\ensuremath{\textup{\textsf{F}}}}{m}{n}\) with maximal rank:[^1] > **Theorem \(\text{\textbf{1}}'\) \[Wiedemann\]**: *Numbers \(\epsilon > 0\) and \(c_1\) exist, both independent of \(q\), with the following property: For any integers \(n > m \ge 0\) a random procedure exists for generating \(n-m\) row vectors with length \(n\) such that if \(A\) is an \(m \times n\) matrix of rank \(m\), then with probability at least \(\epsilon\), the resulting \(n \times n\) matrix is nonsingular and the total Hamming weight of the generated rows is at most \(1 + c_1 n \log n\).* Unfortunately the unknown constants \(\epsilon\) and \(c_1\) are neither supplied nor estimated. Furthermore, it seems that if the proof in  is applied without change in order to determine these values then either \(c_1\) must be so large or \(\epsilon\) so tiny that the result is of limited practical interest. This is, somewhat, rectified in Section [2](#sec:modified_proof){reference-type="ref" reference="sec:modified_proof"}: While the outline of Wiedemann's argument is maintained, along with the details of several steps, several other components are revised or replaced entirely in order to remove unnecessary bounds on various parameters and to simplify the estimation of the unknown parameters \(\epsilon\) and \(c_1\). The bound \(\epsilon\) is also increased by adding another \(\ell\) rows to the resulting matrix \(B \in \matgrp{\ensuremath{\textup{\textsf{F}}}}{(n+\ell)}{n}\) rows; here, \(\ell\) depends on the size of the field \(\ensuremath{\textup{\textsf{F}}}\). In particular, the following result is obtained. The probability bounds listed above are quite arbitrary. The parameter \(\sigma\) does not depend on this probability. Formulas for \(\ell\), \(\tau\) and \(\upsilon\), depending on the field size \(q\) and an arbitrarily small failure probability \(\epsilon\), are given in Section [2](#sec:modified_proof){reference-type="ref" reference="sec:modified_proof"}. A second result, which is also proved in Section [2](#sec:modified_proof){reference-type="ref" reference="sec:modified_proof"}, establishes that the constant \(\sigma\), mentioned above, can be made arbitrarily close to \(6 \left(1-\frac{1}{q}\right)\), provided that the minimum field size \(q\) and the constant \(\upsilon\) are both increased---at the cost of increasing the constants \(\ell\), \(\tau\) and \(\upsilon\) (but not \(\sigma\)) that are listed. Once again, the probability bounds here are quite arbitrary, and probability bounds that are closer to one can be obtained by applications of the same techniques, at the cost of increasing the values of the constants \(\tau\) and \(\upsilon\). This is work in progress. Future versions of this report will document progress in establishing that this yields an efficient matrix preconditioner, to bound the number of nontrivial nilpotent blocks of a conditioned matrix without lowering matrix rank, for matrices over small finite fields. # A Modified Proof of Wiedemann's Result {#sec:modified_proof} This section describes modifications to Wiedemann's argument needed to establish Theorems [\[thm:constants_for_conditioner\]](#thm:constants_for_conditioner){reference-type="ref" reference="thm:constants_for_conditioner"} and [\[thm:second_constants_for_conditioner\]](#thm:second_constants_for_conditioner){reference-type="ref" reference="thm:second_constants_for_conditioner"}. ## Getting Started---and Improving Reliability by Adding Rows {#ssec:getting_started} Suppose that \(m\) and \(n\) are positive integers such that \(0 \le m \le n\) and \(A \in \matgrp{\ensuremath{\textup{\textsf{F}}}}{m}{n}\) is a matrix with maximal rank \(m\), where \(\ensuremath{\textup{\textsf{F}}} = \ensuremath{\textup{\textsf{F}}}_q\) is a finite field with size \(q\). Following Wiedemann's argument, let us begin by assuming that \(q \le n^2\) and suppose, as well, that \[\label{eq:definition_of_k_and_c2} n = m + k + c_2\] where \(c_2\) will be defined later. Suppose first that \(k \le c_3 \ln n\), where \(c_3\) is another constant to be chosen later, and that the remaining \(n-m+\ell\) rows of an \((n + \ell) \times n\) matrix \(B\) are chosen uniformly and independently from \(\matgrp{\ensuremath{\textup{\textsf{F}}}}{1}{n}\). The following lemma, which is easily proved, bounds the probability that the rank of \(B\) is less than \(n\) in this case. Since \(n-m = k + c_2\), the expected number of nonzero entries in these rows is \[\left( \frac{q-1}{q} \right) (k + c_2 + \ell)n < \left( \frac{q-1}{q} \right) \cdot \left( c_3 n \ln n + (c_2 + \ell) n \right)\] when \(k \le c_3 \ln n\). With that noted let us suppose, instead, that \(k > c_3 \ln n\). Following Wiedemann once again, let \[\label{eq:definition_of_z} z = 1-\frac{c_3 \ln n}{k} > 0.\] Suppose that, for the initial \(k\) rows, each entry is set to zero with probability \(z\). The remaining unset entries are then chosen uniformly and independently from \(\ensuremath{\textup{\textsf{F}}}\). The expected number of nonzero entries in these rows is then \[\left( \frac{q-1}{q} \right) c_3 n \ln n.\] If the entries of another \(c_2 + \ell\) rows are chosen uniformly and independently from \(\ensuremath{\textup{\textsf{F}}}\) then the expected number of nonzero entries in these rows is \[\left( \frac{q-1}{q} \right) (c_2 + \ell) n.\] Consequently the expected number of nonzero entries in all these rows is less than \[\left( \frac{q-1}{q} \right) \cdot \left( c_3 n \ln n + (c_2 + \ell)n \right)\] in this case as well. The bulk of the rest of Wiedemann's argument concerns the derivation of an upper bound for the probability that \(B\) has rank less than \(n\) if \(k > c_3 \ln n\) and the rows of \(B\) are chosen in this way. Following Wiedemann, let \(\rho\) be the probability that the rows of \(A\), together with the first \(k\) (sparse) rows generated using the above process, are linearly dependent---that is, the probability that the space spanned by these vectors has dimension less than \(m+k\). Wiedemann shows that \[\label{eq:bound_for_rho} \rho \le \rho_0 + \rho_1\] where \[\label{eq:definition_of_rho0} \rho_0= \sum_{1 \le j \le k\beta_0} \binom{k}{j} (q-1)^j \left( q^{-1} + \frac{(q-1)}{q} (nq)^{-c_4 \beta} \right)^{n-m},\] and \[\label{eq:definition_of_rho_1} \rho_1 = \sum_{k \beta_0 < j \le k} \binom{k}{j} (q-1)^j \left( q^{-1} + \frac{(q-1)}{q} (nq)^{-c_4 \beta} \right)^{n-m},\] when \[\label{eq:definition_of_c4} c_4 = \frac{c_3}{3},\] and when \[\label{eq:definition_of_beta} \beta = \frac{j}{k}\] in the above expressions, and where \(\beta_0\) is yet another constant to be defined later. A useful bound for \(\rho_1\) is next obtained: Assuming that \[\label{eq:constraints_on_c2_and_c4} c_2 \ge \frac{3}{2} \ge \log_2 e \qquad \text{and} \qquad c_4 \ge \frac{2}{\beta_0},\] Wiedemann establishes that \[\label{eq:bound_for_rho1} \rho_1 \le 2 q^{-c_2}.\] Wiedemann continues by observing that \[\label{eq:first_bound_for_rho0} \rho_0 \le \sum_{1 \le j < k \beta_0} \left( \frac{1}{\beta^{\beta} (1-\beta)^{1-\beta}} f(q) \right)^k,\] where \[\label{eq:definition_of_f} f(x) = (x-1)^{\beta} \left( x^{-1} + (1-x^{-1}) (nx)^{-c_4 \beta} \right).\] ## Getting to the Next Step by a Different Route {#ssec:next_step} Wiedemann continues by using the above to establish that \[\label{eq:big_goal} \rho_0 \le \sum_{1 \le j < k \beta_0} \beta^{k \beta}.\] Unfortunately, Wiedemann's involves a Taylor series approximation that seems only to be accurate for a limited range of values, and might suggest that either \(\beta\) must be tiny or \(c_4\) must be huge in order for it to be applicable. The argument that follows is, therefore, quite different from given by Wiedemann . With that noted, consider the equation at lines [\[eq:first_bound_for_rho0\]](#eq:first_bound_for_rho0){reference-type="eqref" reference="eq:first_bound_for_rho0"} and [\[eq:definition_of_f\]](#eq:definition_of_f){reference-type="eqref" reference="eq:definition_of_f"} once again. These imply that \[\label{eq:what_is_known_1} \begin{split} \rho_0 &\le \sum_{1 \le j < k \beta_0} \left(\frac{(q-1)^{\beta}}{\beta^{\beta} (1-\beta)^{1-\beta}} \left( \frac{1}{q} + \frac{q-1}{q} (nq)^{-c_4 \beta} \right) \right)^k \\ &\le \sum_{1 \le j < k \beta_0} \left(\frac{(q-1)^{\beta}}{\beta^{\beta} (1-\beta)^{1-\beta}} \left( \frac{1}{q} + \frac{q-1}{q} \left( \frac{\beta}{q} \right)^{c_4 \beta} \right)\right)^k \end{split}\] since \(n^{-1} \le k^{-1} \le \frac{j}{k} = \beta\). It follows from this that \[\label{eq:what_is_known_2} \rho_0 < \sum_{1 \le j < k \beta_0} \left(\frac{1}{\beta^{\beta} (1-\beta)^{1-\beta}} \left( \frac{(q-1)^{\beta}}{q} + \frac{(q-1) q^{(1-c_4)\beta}}{q} \beta^{c_4 \beta} \right) \right)^k.\] ### Bounding Terms in \(\boldsymbol{\rho_0}\) When \(\boldsymbol{\beta}\) is Extremely Small {#ssec:beta_is_tiny} Note that \(\displaystyle{\lim_{\beta \rightarrow 0^{+}}} \beta^{\beta} = \displaystyle{\lim_{\beta \rightarrow 0^{+}}} (1-\beta)^{1-\beta} = 1\), and \(\beta^{\beta} < 1\) when \(0 < \beta < 1\). The bounds given at lines [\[eq:what_is_known_1\]](#eq:what_is_known_1){reference-type="eqref" reference="eq:what_is_known_1"} and [\[eq:what_is_known_2\]](#eq:what_is_known_2){reference-type="eqref" reference="eq:what_is_known_2"} can be simplified by establishing a lemma like the following, allowing the factor \((1-\beta)^{-(1-\beta)}\) to be replaced by a factor \(\beta^{\gamma \beta}\), for a negative constant \(\gamma\), when \(\beta\) is small. Suppose, now, that \(\gamma \in \ensuremath{\mathbb{Q}}\) is a negative constant such that \((1-\beta)^{-(1-\beta)} \le \beta^{\gamma \beta}\) when \(0 < \beta \le \Delta\) for a positive constant \(\Delta\); it follows by the above lemma that one might choose \(\gamma =-\frac{11}{25}\) if \(\Delta = \frac{5}{43}\), and that one might choose \(\gamma =-\frac{23}{40}\) if \(\Delta = \frac{1}{5}\). It would follow from the inequality at line [\[eq:what_is_known_2\]](#eq:what_is_known_2){reference-type="eqref" reference="eq:what_is_known_2"} that \[\label{eq:what_is_known_3} \rho_0 \le \sum_{1 \le j < k \beta_0} \left(\beta^{(\gamma-1)\beta} \left( \frac{(q-1)^{\beta}}{q} + \frac{(q-1)q^{(1-c_4)\beta}}{q} \beta^{c_4 \beta} \right)\right)^k\] This can be further simplified by bounding \((q-1)^{\beta}\) by \(\beta^{\delta \beta}\) for a small positive constant \(\delta\): Suppose next that \(\delta \in \ensuremath{\mathbb{Q}}\) is constant that is less than or equal to \(0\) such that \((q-1)^{\beta} \le \beta^{\delta \beta}\) when \(0 < x \le \Delta\); one can certainly choose \(\delta = 0\) if \(q = 2\), and it follows by the above lemma that if \(q=3\) then one can choose \(\delta =-\frac{9}{20}\) when \(\Delta = \frac{1}{5}\). It is already necessary for the argument being developed that \(c_4 \ge 1\), so that (when \(q \ge 2\)) \(q^{(1-c_4)\beta} \le 1\). It would therefore follow from the bound at line [\[eq:what_is_known_3\]](#eq:what_is_known_3){reference-type="eqref" reference="eq:what_is_known_3"} that \[\label{eq:what_is_known_4} \rho_0 \le \sum_{1 \le j < k \beta_0} \left(\beta^{(\gamma-1)\beta} \left( \frac{\beta^{\delta \beta}}{q} + \frac{(q-1)}{q} \beta^{c_4 \beta}\right)\right)^k.\] Once again suppose that---as in Wiedemann's original argument---we wish to show that \[\rho_0 \le \sum_{1 \le j < k \beta_0} \beta^{k\beta}.\] Then it follows from the above that it is sufficient to show that \[\label{eq:what_is_needed_1} \beta^{(\gamma-1) \beta} \left( \frac{\beta^{\delta \beta}}{q} + \frac{(q-1)}{q} \beta^{c_4 \beta} \right) \le \beta^{\beta} \qquad \text{when \(0 < \beta = \frac{j}{k} \le \beta_0\),}\] that is, that \(f_1(\beta, q, c, \gamma, \delta) \ge 0\), when \(0 < \beta \le \beta_0\) and \(c = c_4\), for \[\label{eq:first_general_approximation} f_1(\beta, q, c, \gamma, \delta) = \beta^{\beta}-\beta^{\beta (\gamma-1)} \left( \frac{\beta^{\beta \delta}}{q} + \frac{(q-1)}{q} \beta^{\beta c} \right).\] Since \({\displaystyle{\lim_{\beta \rightarrow 0^{+}}}} \beta^{\beta} = 1\), \[\lim_{\beta \rightarrow 0^{+}} f_1(\beta, q, c, \gamma, \delta) = 1-\left( {\textstyle{\frac{1}{q}}} + {\textstyle{\frac{q-1}{q}}} \right) = 0.\] It therefore suffices to check the condition at line [\[eq:new_requirement_for_c\]](#eq:new_requirement_for_c){reference-type="eqref" reference="eq:new_requirement_for_c"} and to confirm that \(f_1(\Delta) \ge 0\) and \(f_1(\widehat{\Delta}) < 0\), for \(0 < \Delta < \widehat{\Delta} \le \frac{1}{e}\), in order to establish that \(f_1(\beta) \ge 0\) when \(0 < \beta \le \Delta\)---so that \[\label{eq:first_target} \sum_{1 \le j \le k \Delta} \left( \frac{(q-1)^{\beta}}{\beta^{\beta} (1-\beta)^{1-\beta}} \left( \frac{1}{q} + \frac{q-1}{q} (nq)^{-c_4 \beta} \right) \right)^k \le \sum_{1 \le j \le k \Delta} \beta^{k\beta}.\] ### Bounding Terms in \(\boldsymbol{\rho_0}\) When \(\boldsymbol{\beta}\) is Larger {#ssec:beta_is_moderate} The process described in Subsection [2.2.1](#ssec:beta_is_tiny){reference-type="ref" reference="ssec:beta_is_tiny"} can only be used to establish the inequality at line [\[eq:first_target\]](#eq:first_target){reference-type="eqref" reference="eq:first_target"}, above, for small positive constants \(\Delta\) that are generally much smaller than the desired bound \(\beta_0 = \frac{2}{c_4}\). However, a complementary process---which in turn, does not seem to be useful to establish the above inequalities when \(\beta\) is extremely close to zero---can (at least, sometimes) be used to establish that these inequalities hold for larger \(\Delta\) as well. Once again, recall that \[\frac{(q-1)^{\beta}}{\beta^{\beta} (1-\beta)^{1-\beta}} \left( \frac{1}{q} + \frac{q-1}{q} (nq)^{-c_4 \beta} \right) \le \frac{(q-1)^{\beta}}{\beta^{\beta} (1-\beta)^{1-\beta}} \left( \frac{1}{q} + \frac{q-1}{q} \left( \frac{\beta}{q} \right)^{c_4 \beta} \right).\] Suppose, now, that \(\eta\) is a constant such that \(0 < \eta < 1\). Then it certainly follows that \[\frac{(q-1)^{\beta}}{\beta^{\beta} (1-\beta)^{1-\beta}} \left( \frac{1}{q} + \frac{q-1}{q} (nq)^{-c_r \beta} \right) \le \beta^{\beta}\] when \(\Gamma_L \le \beta \le \Gamma_H\), for positive constants \(\Gamma_L\) and \(\Gamma_H\), if \[\label{eq:first_new_relationship} \frac{(q-1)^{\beta}}{\beta^{\beta} (1-\beta)^{1-\beta}} \cdot \frac{1}{q} \le \eta \beta^{\beta}\] and \[\label{eq:second_new_relationship} \frac{(q-1)^{\beta}}{\beta^{\beta} (1-\beta)^{1-\beta}} \cdot \frac{q-1}{q} \left(\frac{\beta}{q}\right)^{c_4 \beta} \le (1-\eta) \beta^{\beta}\] when \(\Gamma_L \le \beta \le \Gamma_H\) as well. Since \(q > 1\) and \(0 < \beta < 1\), the inequality at line [\[eq:first_new_relationship\]](#eq:first_new_relationship){reference-type="eqref" reference="eq:first_new_relationship"} holds if and only if \[\frac{\eta \beta^{2\beta} (1-\beta)^{1-\beta} q}{(q-1)^\beta} \ge 1.\] Considering logarithms, one can see that this is the case for \(\Gamma_L \le \beta \le \Gamma_H\) if and only if \(F_1(\beta) \ge 0\) for \(\Gamma_L \le \beta \le \Gamma_H\), where \[\label{eq:defiinition_of_F1} F_1(x) = 2x\ln x + (1-x) \ln (1-x)-x \ln (q-1) + \ln q + \ln \eta.\] Similarly, the inequality at line [\[eq:second_new_relationship\]](#eq:second_new_relationship){reference-type="eqref" reference="eq:second_new_relationship"} is satisfied if and only if \[\frac{(1-\eta) \beta^{(2-c_4)\beta} (1-\beta)^{1-\beta} q^{c_4\beta + 1}}{(q-1)^{\beta+1}} \ge 1.\] Considering logarithms, one can see that this is the case for \(\Gamma_L \le \beta \le \Gamma_H\) if and only if \(F_2(\beta) \ge 0\) for \(\Gamma_L \le \beta \le \Gamma_2\), where \[\label{eq:definition_of_F2} F_2(x) = (2-c_4)x\ln x + (1-x)\ln(1-x) + (c_4x+1) \ln q-(x+1)\ln(q-1) + \ln (1-\eta).\] Note next that \[F_1'(x) = 1 + 2 \ln x-\ln (1-x)-\ln (q-1)\]---which is independent of \(\eta\)---and \[F_1''(x) = \frac{2}{x} + \frac{1}{1-x}\]---which is positive if \(0 < x < 1\). Consequently it it suffices to check that \(F_1'(\beta_0) < 0\) in order to establish that \(F_1'(x) < 0\) for \(0 < x \le \beta_0\), so that \(F_1(x)\) is decreasing over the interval \(\Delta < x \le \beta_0\). If this is the case, and \(\Delta \le \Gamma_L < \Gamma_H \le \beta_0\), then it suffices to check that \(F_1(\Gamma_H) \ge 0\) in order to confirm that \(F_1(\beta) \ge 0\) for \(\Gamma_L \le \beta \le \Gamma_H\) as well. Note as well that \[F_2'(x) = (1-c_4) + (2-c_4)\ln x-\ln (1-x) + c_4 \ln q-\ln (q-1)\]---which is also independent of \(\eta\)---and that \[F_2''(x) = \frac{2-c_4}{x} + \frac{1}{1-x}\]---which is negative (for \(c_4 > 2\)) if \(0 < x < \frac{2-c_4}{1-c_4} = 1 + \frac{1}{1-c_4}\), zero if \(x = 1 + \frac{1}{1-c_4}\), and positive if \(x > 1 + \frac{1}{1-c_4}\). Consequently if \(\beta_0 \le 1 + \frac{1}{1-c_4}\) then it suffices to check that \(F2'(\beta_0) \ge 0\) in order to establish that \(F_2\) is increasing over the interval \(\Delta \le x \le \beta_0\). If \(\Delta \le 1 + \frac{1}{1-c_4} \le \beta_0\) then it suffices to check that \(F_2'\left(1 + \frac{1}{1-c_4}\right) \ge 0\) in order to confirm that \(F_2\) is increasing over this interval. If \(1 + \frac{1}{1-c_4} \le \Delta\) then it suffices to check that \(F_2'(\Delta) \ge 0\) in order to confirm this. In any case, if this has been confirmed and \(\Delta \le \Gamma_L < \Gamma_H \le \beta_0\), then it suffices to check that \(F_2(\Gamma_L) \ge 0\) in order to confirm that \(F_2(\beta) \ge 0\) for \(\Gamma_L \le \beta \le \Gamma_H\). The desired inequality can now be established, for \(\Delta \le \beta \le \beta_0\), by breaking this interval into one or more subintervals, and using the above process with various choices of \(\eta\) to confirm that \(F_1\) and \(F_2\) are both non-negative over each subinterval. ### Establishing That \(f\) is a Decreasing Function {#ssec:f_is_decreasing} Once again, consider the function \[f(x) = (x-1)^{\beta} (x^{-1} + (1-x^{-1}) (nx)^{-c_4 \beta})\] as defined at line [\[eq:definition_of_f\]](#eq:definition_of_f){reference-type="eqref" reference="eq:definition_of_f"}. Wiedemann establishes that if \(0 < \beta \le \beta_0 < {\textstyle{\frac{1}{4}}}\), \(n \ge 1\), \(q \ge 2\) and \(c_4 > {\textstyle{\frac{4}{\ln 2}}}\), then \(f\) is a non-increasing function, so that \(f(q) \le f(2)\) for \(q \ge 2\)---as needed to establish that results like the above hold for larger finite fields as well. Unfortunately, this argument requires both \(c_4\) and \(c_3\) to assume larger values than are either necessary or desirable. However, Wiedemann's argument can be modified in a straightforward way to establish the following. ### Application of These Processes {#ssec:bounding_rho0} The processes described in Subsections [2.2.1](#ssec:beta_is_tiny){reference-type="ref" reference="ssec:beta_is_tiny"} and [2.2.2](#ssec:beta_is_moderate){reference-type="ref" reference="ssec:beta_is_moderate"} and the result established in Section [2.2.3](#ssec:f_is_decreasing){reference-type="ref" reference="ssec:f_is_decreasing"} can now be applied to establish the following. It now follows that \[\label{eq:fourth_bound_for_rho0_when_q=2} \rho_0 \le \sum_{1 \le j < k \beta_0} \beta^{k\beta} \quad \text{if \(q = 2\), \(\beta_0 = \frac{6}{43}\), and \(c_4 \ge \frac{43}{3}\).}\] In the above lemma the upper limit, \(\beta_0 = \frac{6}{43}\) for \(\beta\), has been chosen so that \(\beta_0 = \frac{2}{c_4}\)---in order to match the constraint between \(\beta_0\) and \(c_4\) as shown at line [\[eq:constraints_on_c2_and_c4\]](#eq:constraints_on_c2_and_c4){reference-type="eqref" reference="eq:constraints_on_c2_and_c4"}. A plot of the function \(\beta^{\beta}-\frac{1}{\beta^{\beta} (1-\beta)^{1-\beta}} \left(\frac{1}{2} + \frac{1}{2} \left(\frac{\beta}{2}\right)^{\frac{43}{3}\beta}\right)\), for \(0 < \beta \le \frac{6}{43}\), is shown in Figure [\[fig:plot1\]](#fig:plot1){reference-type="ref" reference="fig:plot1"}. As this may suggest, the above result result can be improved slightly---but not by very much: The inequality at line [\[eq:fourth_bound_for_rho0_when_q=2\]](#eq:fourth_bound_for_rho0_when_q=2){reference-type="eqref" reference="eq:fourth_bound_for_rho0_when_q=2"}, above, is not satisfied if \(c_4\) is decreased to \(14\) and \(\beta_0\) increased to \(\frac{1}{7}\). In order to see one more example of this process let us consider the case that \(q \ge 3\). An application of the technique described above establishes the following. It now follows that \[\label{final_bound_for_rho0_when_q=3} \rho_0 \le \sum_{1 \le j < k \beta_0} \beta^{k\beta} \quad \text{if \(q = 3\), \(\beta_0 = \frac{1}{4}\), and \(c_4 \ge 4\).}\] A plot of the function \(\beta^{\beta}-\frac{2^{\beta}}{\beta^{\beta} (1-\beta)^{1-\beta}} \left( \frac{1}{3} + \frac{2}{3} \left(\frac{\beta}{3}\right)^{8 \beta}\right)\), when \(0 < \beta \le \frac{1}{4}\), is shown in Figure [\[fig:plot2\]](#fig:plot2){reference-type="ref" reference="fig:plot2"}. Once again, this suggests that the above result cannot be improved by very much. Appendix [\[sec:continuation_of_argument\]](#sec:continuation_of_argument){reference-type="ref" reference="sec:continuation_of_argument"} include details of analyses for additional field sizes as well---as summarized in Figure [\[fig:choices_of_c4\]](#fig:choices_of_c4){reference-type="ref" reference="fig:choices_of_c4"} on page . A Maple worksheet, that can be used to check these details, is available online at It follows from this that \[\rho_0 \le \sum_{1 \le j < k \beta_0} \beta^{k\beta}\] for each of the choices of \(q\), \(\beta_0\) and \(c_4\) given in Figure [\[fig:choices_of_c4\]](#fig:choices_of_c4){reference-type="ref" reference="fig:choices_of_c4"}. ## Asymptotic Results: Choice of Field Size {#ssec:asymptotics_1} The objective of this next subsection is to identify bounds on the sizes of primes allowing the inequality at line [\[eq:big_goal\]](#eq:big_goal){reference-type="eqref" reference="eq:big_goal"} to be established when \(c_4\) is closer to \(2\). Suppose, in particular, that \(N\) is an integer such that \(N \ge 18\) and that \[\label{eq:asymptotic_constraints_on_constants} c_4 = 2 + {\textstyle{\frac{1}{N}}} \qquad \text{and} \qquad q \ge 16N + 9.\] Consider now the function \[\label{eq:definition_of_g} \begin{split} g(\beta) &= \frac{\left( \frac{(q-1)^{\beta}}{\beta^{\beta} (1-\beta)^{1-\beta}} \right) \left( \frac{1}{q} + \left(\frac{q-1}{q}\right) \left( \frac{\beta}{q} \right)^{c\beta} \right)}{\beta^{\beta}} \\ &= \frac{(q-1)^{\beta}}{\beta^{2\beta} (1-\beta)^{1-\beta}} \left( \frac{1}{q} + \left(\frac{q-1}{q}\right) \left( \frac{\beta}{q} \right)^{c_4 \beta} \right) \end{split}\] noting---by the inequality at line [\[eq:what_is_known_1\]](#eq:what_is_known_1){reference-type="eqref" reference="eq:what_is_known_1"}---that the inequality at line [\[eq:big_goal\]](#eq:big_goal){reference-type="eqref" reference="eq:big_goal"} is satisfied if \(g(\beta) \le 1\) when \(0 < \beta \le \beta_0 = \frac{2}{c_4}\). Consider, as well the function \[\label{eq:definition_of_h} \begin{split} h(\beta) &= \ln g(\beta) \\ &= \beta \ln (q-1)-2 \beta \ln \beta-(1-\beta) \ln (1-\beta) \\ &\hspace*{1 true in} + \ln \left( \frac{1}{q} + \left( \frac{q-1}{q} \right) \left(\frac{\beta}{q} \right)^{c_4 \beta} \right) \end{split}\] observing that \[\label{eq:definition_of_hprime} h'(\beta) = \ln (q-1)-2 \ln \beta + \ln (1-\beta)-1 + \frac{H(\beta)}{H(\beta)+1} (c_4 + c_4 \ln \beta-c_4 \ln q)\] where \[\label{eq:definition_of_H} H(\beta) = (q-1) \left( \frac{\beta}{q}\right)^{c_4 \beta}.\] A different approach is required for larger values of \(\beta\) because \(h'(\beta)\) is eventually positive. Note that \(g(\beta) = g_1(\beta) + g_2(\beta)\), where \[\label{eq:definition_of_g1} g_1(\beta) = \frac{(q-1)^{\beta}}{\beta^{2\beta} (1-\beta)^{1-\beta}} \cdot \frac{1}{q} = q^{-1} (q-1)^{\beta} \cdot \frac{(1-\beta)^{\beta-1}}{\beta^{2\beta}}\] and \[\label{eq:definition_of_g2} g_2(\beta) = \frac{(q-1)^{\beta}}{\beta^{2\beta} (1-\beta)^{1-\beta}} \cdot \left( \frac{q-1}{q} \right) \left( \frac{\beta}{q} \right)^{c_4 \beta} =\left( 1-\frac{1}{q}\right) \frac{\beta^{\beta/N}}{ \left( \frac{q^{c_4-1}}{1-\beta} \right)^{\beta} (1-\beta)}.\] Once again the functions \(h(\beta)\), \(h'(\beta)\) and \(H(\beta)\), shown at lines [\[eq:definition_of_h\]](#eq:definition_of_h){reference-type="eqref" reference="eq:definition_of_h"}--[\[eq:definition_of_H\]](#eq:definition_of_H){reference-type="eqref" reference="eq:definition_of_H"}, are of use to prove the desired result for larger values of \(\beta\). The following is now a straightforward consequence of Lemmas [\[lem:first_asymptotic_interval\]](#lem:first_asymptotic_interval){reference-type="ref" reference="lem:first_asymptotic_interval"}--[\[lem:third_asymptotic_interval\]](#lem:third_asymptotic_interval){reference-type="ref" reference="lem:third_asymptotic_interval"}. ## Splitting the Sum to Get a Better Bound for \(\boldsymbol{\rho_0}\) {#ssec:splitting_sum} Let \(\epsilon\) be a positive constant. Suppose now that \(\Delta\) is a positive integer whose depends on \(\epsilon\) and the field size \(q\). It follows from the above (for appropriate choices of \(q\), \(\beta_0\) and \(c_4\)) that \[\begin{aligned} \rho_0 &\le \sum_{1 \le j < k \beta_0} \beta^{k \beta} \\ &= \sum_{1 \le j < k \beta_0} \left(\frac{j}{k}\right)^j & \tag{since \(\beta = \frac{j}{k}\)} \\ &= \zeta+ \theta, \end{aligned}\] where \[\label{eq:defn_of_zeta} \zeta = \sum_{1 \le j \le \Delta} \left( \frac{j}{k} \right)^j\] and \[\label{eq:defn_of_theta} \theta = \sum_{\Delta < j < k \beta_0} \left( \frac{j}{k} \right)^j.\] It follows from the above that \[\begin{aligned} \theta &\le \sum_{\Delta < j \le k \beta_0} \left( \frac{j}{k} \right) ^j\\ &\le \sum_{\Delta < j \le k \beta_0} \beta_0^j \\ &\le \beta_0^{\Delta + 1} \sum_{j \ge 0} \beta_0^j \\ &= \frac{\beta_0^{\Delta+1}}{1-\beta_0} \\ &\le {\textstyle{\frac{1}{10}}} \epsilon \end{aligned}\] provided that \[\Delta \ge \frac{\ln (\epsilon^{-1}) + \ln 10-\ln (1-\beta_0)}{\ln (1/\beta_0)}-1.\] Choices of \(\Delta\) that satisfy this inequality for various field sizes (and corresponding choices of \(\beta_0\)) are shown in Figure [\[fig:choices_of_Delta\]](#fig:choices_of_Delta){reference-type="ref" reference="fig:choices_of_Delta"} on page . When \(N \ge 18\) and \(\beta_0 = \frac{2N}{2N+1}\) (as in Subsection [2.3](#ssec:asymptotics_1){reference-type="ref" reference="ssec:asymptotics_1"}) \[\begin{aligned} \frac{\ln (\epsilon^{-1}) + \ln 10-\ln(1-\beta_0)}{\ln(1/\beta_0}-1 &= \frac{\ln (\epsilon^{-1} + \ln 10-\ln \left( {\textstyle{\frac{1}{2N+1}}}\right)}{\ln \left(1 + {\textstyle{\frac{1}{2N}}}\right)}-1 \\ &\le \frac{\ln (\epsilon^{-1}) + \ln 10 + \ln (2N+1)}{{\textstyle{\frac{1}{2N+1}}}}-1 \tag{since \(\ln \left(1 + \frac{1}{2N}\right) \ge \frac{1}{2N+1}\)} \\ &= (2N+1) \ln (\epsilon^{-1}) + (2N+1) \ln (2N+1) + (2N+1) \ln 10-1. \end{aligned}\] It therefore suffices to ensure that \[\Delta \ge \lceil (2N+1) \ln (\epsilon^{-1}) + (2N+1) \ln (2N+1) + (2N+1) \ln 10-1 \rceil\] in this case. It also follows from the above that \[\begin{aligned} \zeta &\le \sum_{1 \le j \le \Delta} \left( \frac{j}{k} \right)^j \\ &\le \sum_{1 \le j \le \Delta} \left( \frac{\Delta}{k} \right)^j \\ &< \sum_{j \ge 1} \left( \frac{\Delta}{k} \right)^j \\ &= \frac{\Delta/k}{1-\Delta/k} \\ &\le \textstyle{\frac{4}{5}} \epsilon \end{aligned}\] provided that \(k \ge \lceil (\frac{5}{4} \epsilon^{-1} + 1) \Delta \rceil\), and this is the case if \(k \ge \lceil (\frac{5}{4} \epsilon^{-1} + 1) (\widehat{\Delta} + 1) \rceil\), where \(\Delta = \lceil \widehat{\Delta} \rceil\). Suitable choices of \(k\) for small field sizes are as shown in Figure [\[fig:choices_of_k\]](#fig:choices_of_k){reference-type="ref" reference="fig:choices_of_k"} on page . When \(N \ge18\), \(c_4 = 2 + \frac{1}{N}\), and \(\beta_0 =\frac{2}{c_4} = \frac{2N}{2N+1}\), it suffices that \[\begin{gathered} \label{eq:asymptotic_bound_for_k} k \ge \left\lceil \left({\textstyle{\frac{5}{4}}} \epsilon^{-1} + 1\right) ((2N+1) \ln (\epsilon^{-1}) + (2N+1) \ln (2N+1) + (2N+1) \ln 10 \right\rceil \\ \in \Theta(\epsilon^{-1} N(\ln (\epsilon^{-1}) + \ln N)). \end{gathered}\] In particular, when \(\epsilon = \frac{1}{10}\), it suffices to ensure that \[\label{eq:asymptotic_bound_for_k_with_specific_failure_probability} k \ge \left\lceil (2N+1) \ln (2N+1) = {\textstyle{\frac{167}{5}}} \ln (2N+1) \right\rceil.\] It now follows that \(\rho_0 \le \frac{1}{10} \epsilon + \frac{4}{5} \epsilon = \frac{9}{10} \epsilon\) provided that this constraint on \(k\) can be satisfied. ## Completion of the Analysis for the Case \(\boldsymbol{q \le n^2}\) Suppose next that one wishes to ensure that \(\rho_1 \le \frac{1}{20}\epsilon\), so that \(\rho \le \frac{19}{20} \epsilon\). If the constraint on \(k\), described above, can be satisfied, then it follows by the inequality at line [\[eq:bound_for_rho1\]](#eq:bound_for_rho1){reference-type="eqref" reference="eq:bound_for_rho1"} that it suffices to choose \(c_2\) such that \(2q^{-c_2} \le \frac{\epsilon}{20}\), that is, \[c_2 \ge \log_q (40 \epsilon^{-1}) = \frac{\ln (\epsilon^{-1}) + \ln 40}{\ln q}.\] It remains to choose \(c_2 + \ell\) rows of the \((n + \ell) \times n\) matrix \(B\) uniformly and independently from \(\matgrp{\ensuremath{\textup{\textsf{F}}}}{1}{n}\). Suppose that that the set of rows of the original matrix \(A\), and the \(k\) "sparse" rows selected as described above, are linearly independent. Then it follows by Lemma [\[lem:dense_selection_of_vectors\]](#lem:dense_selection_of_vectors){reference-type="ref" reference="lem:dense_selection_of_vectors"} that (once the remaining rows of \(B\) are chosen uniformly and independently) the rank of \(B\) is less than \(n\) with probability at most \(q^{-\ell}\). Furthermore, if \(q \ge 3\) then the top \(n\) rows of \(B\) are linearly independent---so that we can set \(\ell = 0\)---with probability at least \(1-\frac{1}{q-1}\). Thus, if we wish to ensure that this probability is at most \(\frac{1}{20}\epsilon\) then it suffices to ensure that either \[\ell \ge \log_q (20 \epsilon^{-1}) = \textstyle{\frac{\ln (\epsilon^{-1}) + \ln 20}{\ln q}} \quad \text{or} \quad q \ge 20 \epsilon^{-1} + 1 \text{ (and \(\ell = 0\)).}\] It now follows that if \(c_4\), \(c_2\) and \(\ell\) have all been chosen as described above then the probability that \(B\) has rank less than \(n\) is at most \[\rho + \frac{\epsilon}{20} (1-\rho) \le \rho + \frac{\epsilon}{20} \le \frac{19 \epsilon}{20} + \frac{\epsilon}{20} = \epsilon.\] Choices of \(c_2\) and \(\ell\) satisfying the above constraints, along with the constraints at line [\[eq:constraints_on_c2_and_c4\]](#eq:constraints_on_c2_and_c4){reference-type="eqref" reference="eq:constraints_on_c2_and_c4"}, are shown in Figure [\[fig:choices_of_c2_and_ell\]](#fig:choices_of_c2_and_ell){reference-type="ref" reference="fig:choices_of_c2_and_ell"}, for the case that \(\epsilon = \frac{1}{10}\). It now suffices to set \(\sigma = c_3 \left( \frac{q-1}{q} \right)= 3 c_4 \left( \frac{q-1}{q} \right)\), for \(c_4\) as given in Figure [\[fig:choices_of_c4\]](#fig:choices_of_c4){reference-type="ref" reference="fig:choices_of_c4"} and \(\tau = c_2 + \ell\), for \(c_2\) and \(\ell\) as given in Figure [\[fig:choices_of_c2_and_ell\]](#fig:choices_of_c2_and_ell){reference-type="ref" reference="fig:choices_of_c2_and_ell"}, in order to establish the claim in Theorem [\[thm:constants_for_conditioner\]](#thm:constants_for_conditioner){reference-type="ref" reference="thm:constants_for_conditioner"} when \(k = n-m-c_2\) is greater than or equal to the lower bound given in Subsection [2.4](#ssec:splitting_sum){reference-type="ref" reference="ssec:splitting_sum"} and \(q \le n^2\), for \(c_2\) as given above. Setting \(\upsilon\) to be the sum of \(\ell\) and the lower bound for \(k\), described in Subsection [2.4](#ssec:splitting_sum){reference-type="ref" reference="ssec:splitting_sum"}, suffices to establish these claims when \(n-m-c_2\) is less than the lower bound for \(k\) and \(q \le n^2\) as well, for \(c_2\) as above. The values shown in Figure [\[fig:summary_table\]](#fig:summary_table){reference-type="ref" reference="fig:summary_table"} have been obtained using these equations. Similarly, Theorem [\[thm:second_constants_for_conditioner\]](#thm:second_constants_for_conditioner){reference-type="ref" reference="thm:second_constants_for_conditioner"} can be obtained by setting \(c_4 = 2 + \frac{1}{N}\) for \(N \ge 18\), \(c_3 = 3 c_4 = 6 + \frac{3}{N}\), \(\sigma = c_3 (1-\frac{1}{q})\), \(c_2 = \left\lceil \frac{\ln (40 \epsilon^{-1})}{\ln q} \right\rceil\), \(\ell = \left\lceil \frac{\ln 20 \epsilon^{-1}}{ln q} \right\rceil\) if \(\ell \le 20 \epsilon^{-1} + 1\), setting \(\ell = 0\) otherwise, setting \(\tau = c_2 + \ell\), and setting \(\upsilon\) to the sum of \(\ell\) and the lower bound for \(k\) described in Subsection [2.4](#ssec:splitting_sum){reference-type="ref" reference="ssec:splitting_sum"}, above. ## Analysis for the Case \(\boldsymbol{q > n^2}\) A slight variant of the argument from Wiedemann  can be applied when \(q > n^2\). Since dense linear algebra is certainly adequate for computations on small matrices it will be assumed that \(n \ge 7\), so that \(q \ge 49\). Suppose, once again, that \(A\) is an \(m \times n\) matrix over a field \(\ensuremath{\textup{\textsf{F}}}\) that is either infinite or has size \(q > n^2\). Let \(\widehat{q}\) be the largest power of a prime that is less than or equal to \(n^2\). It suffices to apply the above construction, using the choices of \(c_4\), \(c_3\), \(\beta_0\), \(c_2\) and \(\ell\) appropriate for a field with size \(\widehat{q}\) (so that \(\ell = 0\))---except that, after choosing entries of rows that might be nonzero, the remaining entries of the matrix \(B\), to be filled in, should be chosen uniformly and independently from a finite subset \(S\) of \(\ensuremath{\textup{\textsf{F}}}\) with size at least \(n^2\), rather than from the finite field with size \(\widehat{q}\). In order to see that this process is reliable, consider yet another matrix---namely, a matrix \(\widehat{B}\) obtained by placing a distinct indeterminate into each row entry that is assigned a value from \(S\), above, instead of \(0\). Let us denote the indeterminate placed into the \(i^{\text{th}}\) new row, in column \(j\), by \(z_{i, j}\). Since \(\ell = 0\) this results in an \(n \times n\) matrix whose entries are elements of \(\ensuremath{\textup{\textsf{F}}}\) and indeterminates. Let \(\widehat{f}\) be the determinant of this matrix---a multivariate polynomial with total degree at most \(n-m\), since only \(n-m\) rows include indeterminates, and each entry of such a row has total degree at most one. Since the rows of the \(m \times n\) matrix \(A\) are linearly independent, there exists a sequence of integers \(i_1, i_2, \dots, i_m\) such that \[1 \le i_1 < i_2 < \dots < i_m \le n\] and such that the \(m \times m\) submatrix, including columns \(i_1, i_2, \dots, i_m\), is nonsingular. Permuting rows of \(A\) as needed, we may assume without loss of generality that the entry of the \(j^{\text{th}}\) row of \(A\) in column \(i_j\) is nonzero, for \(1 \le j \le m\). Consequently, if the \(j^{\text{th}}\) row of \(A\) was replaced by a row whose \(i_j^{\text{th}}\) entry is \(1\) and whose other entries are \(0\), this would result in an \(n \times m\) matrix \(\widehat{A}\) whose rows are linearly independent as well. Indeed, the \(m \times m\) submatrix including the entries in columns \(i_1, i_2,\dots, i_m\) would have determinant \(1\). Similarly, the rows of this matrix are linearly independent when the entries are viewed as elements of the finite field \(\ensuremath{\textup{\textsf{F}}}_{\widehat{q}}\) with size \(\widehat{q}\), instead of as elements of \(\ensuremath{\textup{\textsf{F}}}\). Let us call this matrix (an \(m \times n\) matrix with entries in \(\ensuremath{\textup{\textsf{F}}}_{\widehat{q}}\)) \(\widetilde{A}\). Note that the process, described above, to produce new rows to obtain \(B\) from \(A\), is independent of the entries in the rows of \(A\)---it only depends on the number \(m\) of rows and \(n\) of columns of \(A\). With that noted, let us consider yet another \(n \times n\) matrix, namely a matrix with entries in \(F_{\widehat{q}}\) whose first \(m\) rows are the rows of \(\widetilde{A}\) and whose remaining rows are produced by initially deciding to set the same entries of rows to \(0\) as for the new rows of \(B\), and whose remaining entries are chosen uniformly and independently from \(\ensuremath{\textup{\textsf{F}}}_{\widehat{q}}\). It follows by the analysis for the case \(q \le n^2\) that this matrix is nonsingular with some probability \(\sigma \ge \frac{8}{9}\). Let us suppose that this is the case. Then there must exist a set of column indices \(i_{m+1}, i_{m+2}, \dots, i_n\) such that \[1 \le i_{m_1} < i_{m+2}, \dots, i_n \le n, \quad \{ i_1, i_2, \dots, i_m \} \cup \{i_{m+1}, i_{m+2}, \dots, i_n \} = \{1, 2, \dots, n \},\] and the entries of \(\widetilde{A}\) in row \(j\) and column \(i_j\) are all nonzero, for \(1 \le j \le n\). Consequently if \(m+1 \le j \le n\) then the entry of \(\widehat{B}\) in row \(j\) and column \(i_j\) is an indeterminate, \(z_{j-m, i_j}\), rather than zero. It now follows that the above polynomial \(\widehat{f}\) is not identically zero (in this case): For if one sets the value of each indeterminate \(z_{j-m, i_j}\) to be \(1\) and one sets the value of all other indeterminates to be \(0\), then the value of this polynomial is the product of \(\pm 1\) and the determinant of the submatrix of \(A\) including the entries in columns \(i_1, i_2, \dots, i_m\)---which is nonsingular, as noted above. It now follows by an application of the Schwartz-Zippel lemma  that the above matrix \(B\) is singular, in this particular case, with probability at most \(\frac{n-m}{|S|} \le \frac{n}{|S|} \le \frac{1}{n}\). It follows that the overall probability that \(B\) is singular is at most \[(1-\sigma) + \textstyle{\frac{\sigma}{n}} \le \textstyle{\frac{1}{9} + \frac{8}{9n}},\] as needed to complete the proofs of Theorems [\[thm:constants_for_conditioner\]](#thm:constants_for_conditioner){reference-type="ref" reference="thm:constants_for_conditioner"} and [\[thm:second_constants_for_conditioner\]](#thm:second_constants_for_conditioner){reference-type="ref" reference="thm:second_constants_for_conditioner"}. [^1]: Wiedemann attributes much of the proof of this claim to an anonymous referee who is thanked for allowing this work to be included.
{'timestamp': '2016-07-18T02:07:18', 'yymm': '1607', 'arxiv_id': '1607.04514', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04514'}
# A numerical example {#Sec-6} For the sake of the discussion in this section, we adopt the view of the Financial Stability Board and the International Monetary Fund that the systemic risk can be caused by impairment of all or parts of the financial system, and more formally, we call the risk factor \(j\in\{1,\ldots,l+m\}\) 'systemic', if \(c_{i,j}=1\) for at least two distinct r.c.'s \(i\in\{1,\ldots,n\}\). Also, we call the risk factor \(j\in\{1,\ldots,l+m\}\) 'idiosyncratic', if \(c_{i,j}=1\) for only one risk component \(i\in\{1,\ldots,n\}\). Consider obligors in a default risk portfolio, each of which is exposed to exactly three distinct categories of fatal risk factors, e.g., systemic with fully-commonotonic occurrence times of the r.f.'s (category A), systemic with conditionally independent occurrence times of the r.f.'s (category B) and idiosyncratic with independent occurrence times of the r.f.'s (category C). We assume that the risk factors from distinct risk categories are independent and that the hitting times (or occurrences) of defaults of the r.c.'s are exponentially-distributed with random parameters distributed gamma. In fact, the future lifetime r.v. of the \(i\)-th r.c., \(i=1,\ldots,n\) has exponential distribution with the random parameter \(\sigma_i^{-1}\sum_{j=1}^{l+m}c_{i,j}\Lambda_j\), where \(\Lambda_j\) are distributed gamma with unit rate parameters. Then Definition [\[MRF-stoch-Def\]](#MRF-stoch-Def){reference-type="ref" reference="MRF-stoch-Def"} readily implies that the joint default times of the aforementioned r.c.'s has d.d.f. ([\[ddf-gp\]](#ddf-gp){reference-type="ref" reference="ddf-gp"}). To illustrate the effect of the dependence structure on the joint default probability we further set the dimension to \(n=2\) and specialize the set-up above along the lines in Section 16.8 of Engelmann and Rauhmeier (2011) as well as employing the 2014's Annual Global Corporate Default Study and Rating Transitions of Standard & Poor's (Standard & Poor's (2015)). More specifically, we assume the existence of six r.f.'s and set each \(\mu:=\mathbf{E}[\Lambda_j]\equiv 1/1.8\), then fix the time horizon to \(15\) years and choose the corresponding default probability, \(p\) say, to be equal to \(0.3198\) (on par with the 'B' credit rating of highly speculative entities). This yields the multivariate probability structure of Definition [\[MRF-stoch-Def\]](#MRF-stoch-Def){reference-type="ref" reference="MRF-stoch-Def"} with identically distributed margins having the parameters \(\sigma_i\equiv \sigma=122.39\) and \(Then we explore three different exposures of the obligors to the systemic and idiosyncratic r.f.'s. The distinct exposures are stipulated by appropriate choices of the\)c\(parameters gathered by matrices\)A_c\^(k), k=1, 2, 3\(. We compare the aforementioned three exposures with the reference case in which no systemic risk presents, that is the joint d.d.f. of default times is a bivariate Pareto with independent margins. We note in passing that the expressions for the d.d.f.'s below readily follow from Theorem \ref{def-gp}, whereas the values of the Pearson correlation coefficient are in non-trivial cases obtained with the help of Theorem \ref{joint_m}. \begin{itemize} \item [Case (1).] Only the systemic (category A) and idiosyncratic (category C) risks present. The exposure is represented schematically with the use of the following matrix, in which the rows and the columns represent r.c.'s and r.f.'s, respectively \[ A_c^{(1)}=\left( \begin{array}{cccc||cccc} 1 & 1 & 1 & 1 & 1 & 1 &0 & 0\\ 1 & 1 & 1 & 1 & 0 & 0& 1 & 1\\ \end{array} \right). \] The joint d.d.f. of the risk components is given by \[ \overline{F}^{(1)}(x_1,x_2)=\left( 1+\frac{\max(x_1,x_2)}{\sigma} \right)^{-4\mu} \left( 1+\frac{x_1}{\sigma} \right)^{-2\mu}\left( 1+\frac{x_2}{\sigma} \right)^{-2\mu} , \] where\)x_1,x_2\(are all in\)\_+\(. This is obviously the d.d.f. of the bivariate Pareto distribution of Asimit et al. (2010). In this r.p., the Pearson correlation coefficient between the r.c.'s is\)`<!-- -->`{=html}0.36\(. \end{itemize} \noindent \begin{itemize} \item [Case (2).] There are four conditionally independent r.f.'s (category B) and two uncorrelated idiosyncratic r.f.'s (category C). The exposure is gathered by the following block matrix \[ A_c^{(2)}=\left( \begin{array}{c|cccc|cccc} ~ & 1 & 1 & 1 & 1 & 1 & 1 &0 & 0\\ ~ & 1 & 1 & 1 & 1 & 0 & 0& 1 & 1\\ \end{array} \right). \] The joint d.d.f. of the risk components is given by \begin{eqnarray*} &&\overline{F}^{(2)}(x_1,x_2)\\ &&=\left( 1+\frac{x_1+x_2}{\sigma} \right)^{-4\mu} \left( 1+\frac{x_1}{\sigma} \right)^{-2\mu} \left( 1+\frac{x_2}{\sigma} \right)^{-2\mu}, \end{eqnarray*} where\)x_1,x_2\(are all in\)\_+\(. This case corresponds to the bivariate Pareto model of Su \& Furman (2016). In this r.p., the Pearson correlation coefficient between the r.c.'s is\)`<!-- -->`{=html}0.14\(. \item [Case (3).] The r.p. admits the most general form that is proposed in the current paper. Namely r.f.'s from all three categories (A, B and C) present. The exposure block matrix is given by \[ A_c^{(3)}=\left( \begin{array}{cc|cc|ccccc} 1 & 1 & 1 & 1 & 1 & 1 &0 & 0\\ 1 & 1 & 1 & 1 & 0 & 0& 1 & 1\\ \end{array} \right). \] The joint d.d.f. of the risk components is \[ \overline{F}^{(3)}(x_1,x_2)=\left( 1+\frac{\max(x_1,x_2)}{\sigma}\right)^{-2\mu} \left( 1+\frac{x_1+x_2}{\sigma}\right)^{-2\mu} \left( 1+\frac{x_1}{\sigma} \right)^{-2\mu}, \left( 1+\frac{x_2}{\sigma} \right)^{-2\mu}, \] where\)x_1,x_2\(are all in\)\_+\(. In this r.p., the Pearson correlation coefficient between the r.c.'s is equal to\)`<!-- -->`{=html}0.23\(. \end{itemize} \subsection{Expected times of the first default} The left panel of Figure \ref{fig:CTEmin} depicts the values of\)CTE_q\[X\_-\]\(for\)qCTE\^(3)\_q\[X\_-\]CTE\^(2)\_q\[X\_-\]CTE\^()\_q\[X\_-\] for all \(q\in[0,\ 1)\) and \(X_-\in\mathcal{X}\). This conforms to Figure [\[fig:CTEmin\]](#fig:CTEmin){reference-type="ref" reference="fig:CTEmin"} (left panel), which hints that the r.p.'s with more significantly correlated r.c.'s enjoy higher, and thus more favourable, occurrence times of the first default. The downside of high correlations is elucidated in Figure [\[fig:minCase\]](#fig:minCase){reference-type="ref" reference="fig:minCase"}, in which we leave the probability of default \(p\) to be equal to \(0.3198\) ('B' rating), but vary the \(\mu\) parameter that stipulates the effect of the risk factors. In this respect, we observe that the r.p.'s with stronger correlations between r.c.'s are more sensitive to the changes in the \(\mu\) parameter, and therefore such r.p.'s must be monitored and stress-tested more frequently. ## Expected times of the last default Figure [\[fig:CTEmin\]](#fig:CTEmin){reference-type="ref" reference="fig:CTEmin"} (right panel) depicts the values of \(CTE_q[X_+]\) for \(q\in[0,\ 1),\) \(X_+\in\mathcal{X}\) and portfolios (1) to (3) as well as the reference portfolio \((\perp)\). We have that \[\overline{F}^{(\perp)}_+\geq_{st}\overline{F}^{(2)}_+\geq_{st} \overline{F}^{(3)}_+ \geq_{st} \overline{F}^{(1)}_+,\] and hence \[CTE^{(\perp)}_q[X_+]\geq CTE^{(2)}_q[X_+]\geq CTE^{(3)}_q[X_+] \geq CTE^{(1)}_q[X_+]\] for all \(q\in[0,\ 1)\) and \(X_+\in\mathcal{X}\). This conforms with the right panel of Figure [\[fig:CTEmin\]](#fig:CTEmin){reference-type="ref" reference="fig:CTEmin"}. Unlike in the case of the first default, we observe that if the time of the last default is of interest and the distributions of the r.c.'s are fixed, then assuming stronger correlations between r.c.'s yields a more conservative assessment of the expected time of the last default. ## Solvency bonus indices Figure [\[fig:econCTE\]](#fig:econCTE){reference-type="ref" reference="fig:econCTE"} depicts the solvency bonus indices for portfolios (1) to (3). As expected, stronger dependencies between default times are associated with higher values of \(\beta\). # Conclusions {#Sec-7} The latest Solvency II directives require the insurers to recognize the interdependencies within and among different liability classes. As such, the new paradigm arguably brings to an end the indisputable role of the assumption of independence that has shaped both research and applications in actuarial science and general quantitative risk management in the 20th century. The choice of an appropriate probability dependence model is however not an easy call. Indeed, while there exists only one way to describe stochastic independence, the forms of stochastic dependence are infinite. In this paper we have introduced a new class of Multiple Risk Factor dependence structures. On the one hand, these structures emerge as an extension of the popular CreditRisk\(^+\) approach, and as such they formally describe default risk portfolios exposed to an arbitrary number of fatal risk factors with conditionally exponential hitting times that can be independent, positively dependent and even fully comonotonic. On the other hand, the MRF structures can be viewed as a quite general family of multivariate probability distributions with Pareto of the 2nd kind univariate margins, and in this role, they can model risk portfolios of (insurance) losses with heavy tailed and positively dependent risk components. It often happens in mathematical sciences that generalizing an object highlights the essence of the matter and helps to understand it better. By generalizing the classical multivariate Pareto distribution of Arnold (1983, 2015), among others, in this paper we have extended the range of the attainable Pearson correlations to the entire \([0,\ 1]\) interval and complemented some formal dependence analysis. As factor models have been known to produce under-correlated default times, the MRF structures introduced and studied herein may provide a possible route to 'inject the required amounts of extra correlation', and they may thereby become of interest to banks, credit unions and insurance companies. In addition, as the new realms of excessively prudent risk management make particular effort to model non-hedgeable and heavy tailed risks, the family of MRF multivariate distributions may be of interest well beyond the context of credit risk.
{'timestamp': '2016-07-19T02:04:00', 'yymm': '1607', 'arxiv_id': '1607.04739', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04739'}
# Introduction {#introduction .unnumbered} Let \(\mathcal{P}\subset\mathbb{R}^d\) be a \(d\)-dimensional lattice polytope, that is, a convex polytope all of whose vertices belong to \(\mathbb{Z}^d\). Let \(\operatorname{Vol}(\mathcal{P})\) denote the *normalized volume* of \(\mathcal{P}\), which is \(d!\) times the Euclidean volume (Lebesgue measure) of \(\mathcal{P}\). For \(k\in \mathbb{Z}_{>0}\), the *lattice point enumerator* \(i(\mathcal{P},k)\) counts the number of lattice points in \(k\mathcal{P}=\{k\alpha:\alpha\in \mathcal{P}\}\), the \(k\)th dilation of \(\mathcal{P}\). That is, \[i(\mathcal{P},k)=\#(k\mathcal{P}\cap \mathbb{Z}^d), \ \ k\in\mathbb{Z}_{>0}.\] Provided that \(\mathcal{P}\) is a lattice polytope, it is known that \(i(\mathcal{P},k)\) is a polynomial in the variable \(k\) of degree \(d\) (). The *Ehrhart Series* for \(\mathcal{P}\), \(\Ehr{\mathcal{P}}{z}\), is the rational generating function \[\Ehr{\mathcal{P}}z= 1+\sum_{k\geq 1}i(\mathcal{P},k)z^k=\frac{\delta(\mathcal{P},z)}{(1-z)^{d+1}}\] where \(\delta(\mathcal{P},z)=1+\delta_1z+\delta_2z^2+\cdots+\delta_dz^d\) is the *\(\delta\)-polynomial* of \(\mathcal{P}\) (cf. ). The \(\delta\)-polynomial is endowed with the following properties: - \(\delta_0=1\), \(\delta_1=i(\mathcal{P},1)-(d+1)\), and \(\delta_d=\#(\mathcal{P}\setminus \partial\mathcal{P}\cap \mathbb{Z}^d)\); - \(\delta_i\geq 0\) for all \(0\leq i\leq d\) (); - If \(\delta_d\neq 0\), then \(\delta_1\leq \delta_i\) for each \(0\leq i\leq d-1\) (). For proofs of the first three properties of the coefficients, the reader should consult or. The Ehrhart series and \(\delta\)-polynomials for polytopes have been studied extensively. For a detailed background on these topics, please refer to. Given two polytopes \(\mathcal{P}_1\) and \(\mathcal{P}_2\) in \(\mathbb{R}^d\), we say that \(\mathcal{P}_1\) and \(\mathcal{P}_2\) are *unimodularly equivalent* if there exists a unimodular matrix \(U\in\mathbb{Z}^{d\times d}\) (i.e. \(\det(U)=\pm 1\)) and an integral vector \({\boldsymbol v} \in \mathbb{Z}^d\), such that \(\mathcal{P}_2=f_U(\mathcal{P}_1)+{\boldsymbol v}\), where \(f_U\) is the linear transformation defined by \(U\), i.e., \(f_U({\bf v}) = {\bf v} U\) for all \({\bf v} \in \mathbb{R}^d\).. We write \(\mathcal{P}_1\cong \mathcal{P}_2\) in the case of unimodular equivalence. It is clear that if \(\mathcal{P}_1\cong\mathcal{P}_2\), then \(\delta(\mathcal{P}_1,z)=\delta(\mathcal{P}_2,z)\). We say that a lattice polytope \(\mathcal{P}\) is *reflexive* if the origin is the unique interior lattice point of \(\mathcal{P}\) and its dual polytope \[\mathcal{P}^\vee=\left\{y\in\mathbb{R}^d \: \langle x,y \rangle\leq 1 \mbox{ for all } x\in\mathcal{P} \right\}\] is a lattice polytope. Moreover, it follows from that the following statements are equivalent: - \(\mathcal{P}\) is unimodularly equivalent to some reflexive polytope; - \(\delta(\mathcal{P},z)\) is of degree \(d\) and is symmetric, that is \(\delta_i=\delta_{d-i}\) for \(0\leq i \leq \lfloor \frac{d}{2} \rfloor\). A polytope \(\mathcal{P}\) is called *self dual* if \(\mathcal{P}\) is unimodularly equivalent to its dual polytope \(\mathcal{P}^\vee\). This is an extremely rare property in reflexive polytopes, especially for reflexive simplices. There are two families known self dual reflexive simplices. The first such family is given in and the second family is given in. A construction for self dual reflexive polytopes is given in, though these polytopes are not simplicial and hence not simplices. In this paper, we provide a new family of self dual reflexive simplices \(Q_n\) with small volume. We now define a family of reflexive simplices which are self dual. For \(n \geq 2\), let \(Q_n\) denote the \(n-1\) dimensional simplex with \(\mathcal{V}\)-representation \[Q_n:=\mathrm{conv} \begin{bmatrix} 1 & 1-n & 0 & 0 & \cdots & 0\\ 1 & 1 & 2-n & 0 & \cdots & 0\\ 1 & 1 & 1 & 3-n & \cdots & 0\\ \vdots & \vdots & \vdots & & \ddots & \vdots\\ 1 & 1 & 1 & \cdots & 1 &-1\\ \end{bmatrix}\] where we use the convention that the \(x_{n-1}\) coordinate is given by the first row and the \(x_1\) is given by the last row. We will adopt for simplicity the notation \(C_{i}\) where \(i\in\{0,\cdots,n-1\}\) for each column (vertex) such that \(Q_n=\mathrm{conv}[C_{0} \ C_{1} \ \cdots \ C_{n-1}]\). We have the following theorem. It behooves us to introduce an \(\mathcal{H}\)-representation for the simplex to compute its dual polytope. We now give such a representation. By, and Proposition [\[hrepQn\]](#hrepQn){reference-type="ref" reference="hrepQn"}, it is clear that \(Q_n^\vee =-Q_n\). Therefore, we have shown Theorem [\[selfdual\]](#selfdual){reference-type="ref" reference="selfdual"}. Moreover, the self dual reflexive simplex of \(Q_n\) has an interesting \(\delta\)-polynomial and a special triangulation
{'timestamp': '2017-03-06T02:07:47', 'yymm': '1607', 'arxiv_id': '1607.04871', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04871'}
null
null
# Introduction We study \(r\)-nets, a powerful tool in computational and metric geometry, with several applications in approximation algorithms. An \(r\)-net for a metric space \((X,\norm{\cdot}), \, |X|=n\) and for numerical parameter \(r\) is a subset \(R\subseteq X\) such that the closed \(r/2\)-balls centered at the points of \(R\) are disjoint, and the closed \(r\)-balls around the same points cover all of \(X\). We define approximate \(r\)-nets analogously. Formally, #### Previous Work. Finding \(r\)-nets can be addressed naively by considering all points of \(X\) unmarked and, while there remains an unmarked point \(p\), the algorithm adds it to \(R\) and marks all other points within distance \(r\) from \(p\). The performance of this algorithm can be improved by using grids and hashing. However, their complexity remains too large when dealing with big data in high dimension. The naive algorithm is quadratic in \(n\) and the grid approach is in \(O(d^{d/2} n)\), hence it is relevant only for constant dimension \(d\) . In, they show that an approximate net hierarchy for an arbitrary finite metric can be computed in \(O(2^{ddim}n \log n)\), where \(ddim\) is the doubling dimension. This is satisfactory when doubling dimension is constant, but requires a vast amount of resources when it is high. When the dimension is high, there is need for algorithms with time complexity polynomial in \(d\) and subquadratic in \(n\). One approach, which computes \((1+\epsilon)r\)-nets in high dimension is that of, which uses the Locality Sensitive Hashing (LSH) method of. The resulting time complexity is \(\tilde{O}(dn^{2-\Theta(\epsilon)})\), where \(\epsilon>0\) is quite small and \(\tilde{O}\) hides polylogarithmic factors. In general, high dimensional analogues of classical geometric problems have been mainly addressed by LSH. For instance, the approximate closest pair problem can be trivially solved by performing \(n\) approximate nearest neighbor (ANN) queries. For sufficiently small \(\epsilon\), this costs \(\tilde{O}(dn^{2-\Theta(\epsilon)})\) time, due to the complexity factor of an LSH query. Several other problems have been reduced to ANN queries. Recently, Valiant, presented an algorithm for the approximate closest pair problem in time \(\tilde{O}(dn^{2-\Theta(\sqrt{\epsilon})})\). This is a different approach in the sense that while LSH exploits dimension reduction through random projections, the algorithm of is inspired by high dimensional phenomena. One main step of the algorithm is that of projecting the pointset up to a higher dimension. #### Our Contribution. We present a new randomized algorithm that computes approximate \(r\)-nets in time subquadratic in \(n\) and polynomial in the dimension, and improves upon the complexity of the best known algorithm. Our method does not employ LSH and, with probability \(1-o(1)\), it returns \(R\subset X\), which is a \((1+\epsilon)r\)-net of \(X\). We reduce the problem of an approximate \(r\)-net for arbitrary vectors (points) under Euclidean distance to the same problem for vectors on the unit sphere. Then, depending on the magnitude of distance \(r\), an algorithm handling "small\" distances or an algorithm handling "large\" distances is called. These algorithms reduce the Euclidean problem of \(r\)-nets on unit vectors to finding an \(r\)-net for unit vectors under inner product (Section [3](#SGeneral){reference-type="ref" reference="SGeneral"}). This step requires that the multiplicative \(1+\epsilon\) approximation of the distance corresponds to an additive \(c\epsilon\) approximation of the inner product, for suitable constant \(c>0\). Next, we convert the vectors having unit norm into vectors with entries \(\{-1, +1\}\) (Section [2](#SInner){reference-type="ref" reference="SInner"}). This transformation is necessary in order to apply the Chebyshev embedding of, an embedding that damps the magnitude of the inner product of "far\" vectors, while preserving the magnitude of the inner product of "close\" vectors. For the final step of the algorithm, we first apply a procedure that allows us to efficiently compute \((1+\epsilon)\)-nets in the case where the number of "small\" distances is large. Then, we apply a modified version of the `Vector Aggregation` algorithm of, that exploits fast matrix multiplication, so as to achieve the desired running time. In short, we extend Valiant's framework and we compute \(r\)-nets in time \(\tilde{O}(dn^{2-\Theta(\sqrt{\epsilon})})\), thus improving on the exponent of the LSH-based construction, when \(\epsilon\) is small enough. This improvement by \(\sqrt{\epsilon}\) in the exponent is the same as the complexity improvement obtained in over the LSH-based algorithm for the approximate closest pair problem. Our study is motivated by the observation that computing efficiently an \(r\)-net leads to efficient solutions for several geometric problems, specifically in approximation algorithms. In particular, our extension of \(r\)-nets in high dimensional Euclidean space can be plugged in the framework of . The new framework has many applications, notably the \(k\)th nearest neighbor distance problem, which we solve in \(\tilde{O}(dn^{2-\Theta(\sqrt{\epsilon})})\). #### Paper Organization. Section [2](#SInner){reference-type="ref" reference="SInner"} presents an algorithm for computing an approximate net with respect to the inner product for a set of unit vectors. Section [3](#SGeneral){reference-type="ref" reference="SGeneral"} translates the problem of finding \(r\)-nets under Euclidean distance to the same problem under inner product. In Section [4](#Sapps){reference-type="ref" reference="Sapps"}, we discuss applications of our construction and possible future work. Omitted proofs are included in the Appendices. We use \(\norm{\cdot}\) to denote the Euclidean norm \(\norm{\cdot}_2\) throughout the paper. # Points on a sphere under inner product {#SInner} In this section, we design an algorithm for constructing an approximate \(\rho\)-net of vectors on the sphere under inner product. To that end, we reduce the problem to constructing an approximate net under absolute inner product for vectors that lie on the vertices of a unit hypercube. Since our ultimate goal is a solution to computing \(r\)-nets with respect to Euclidean distance, we allow additive error in the approximation, which under certain assumptions, translates to multiplicative error in Euclidean distance. In the following, we define rigorously the notion of approximate \(\rho\)-nets under inner product. One relevant notion is that of \(\epsilon\)-kernels. In \(\epsilon\)-kernels, one is interested in finding a subset of the input pointset, which approximates its directional width. Such constructions have been extensively studied when the dimension is low, due to their relatively small size. ## Crude approximate nets In this subsection we develop our basic tool, which is based on the Vector Aggregation Algorithm by. This tool aims to compute approximate \(\rho\)-nets with multiplicative error, as opposed to what we have set as our final goal for this section, namely to bound additive error. Moreover, in the context of this subsection, two vectors are close to each other when the magnitude of their inner product is large, and two vectors are far from each other when the magnitude of their inner product is small. Let \(|\langle \cdot,\cdot \rangle|\) denote the magnitude of the inner product of two vectors. `Vector Aggregation` follows the exposition of. The main difference is that, instead of the "compressed" matrix \(Z^T Z\), we use the form \(X^T Z\), where \(Z\) derives from vector aggregation. Both forms encode the information in the Gram matrix \(X^T X\). The matrix \(X^TZ\) is better suited for our purposes, since each row corresponds to an input vector instead of an aggregated subset; this extra information may be useful in further problems. For the case of pointsets with many "small\" distances, we rely crucially on the fact that the expected number of near neighbors for a randomly chosen point is large. So, if we iteratively choose random points and delete these and their neighbors, we will end up with a pointset which satisfies the property of having sufficiently few "small\" distances. Then, we apply `Vector Aggregation`. ## Approximate inner product nets In this subsection, we show that the problem of computing \(\rho\)-nets for the inner product of unit vectors reduces to the less natural problem of Definition [\[DfnMagnInnNet\]](#DfnMagnInnNet){reference-type="ref" reference="DfnMagnInnNet"}, which refers to the magnitude of the inner product. The first step consists of mapping the unit vectors to vectors in \(\{-1,1\}^{d'}\). The mapping is essentially Charikar's LSH scheme. Then, we apply the Chebyshev embedding of  in order to achieve gap amplification, and finally we call algorithm `Crude ApprxNet`, which will now return a proper \(\rho\)-net with additive error. The following theorem provides a randomized embedding that damps the magnitude of the inner product of "far\" vectors, while preserving the magnitude of the inner product of "close\" vectors. The statement is almost verbatim that of  except that we additionally establish an asymptotically better probability of success. The proof is the same, but since we claim stronger guarantees on success probability, we include the complete proof in Appendix [\[AppThmCheb\]](#AppThmCheb){reference-type="ref" reference="AppThmCheb"}. # Approximate nets in high dimensions {#SGeneral} In this section, we translate the problem of computing \(r\)-nets in \(({\mathbb R}^d,\|\cdot \|)\) to the problem of computing \(\rho\)-nets for unit vectors under inner product. One intermediate step is that of computing \(r\)-nets for unit vectors under Euclidean distance. ## From arbitrary to unit vectors In this subsection, we show that if one is interested in finding an \(r\)-net for \(({\mathbb R}^d,\|\cdot \|)\), it is sufficient to solve the problem for points on the unit sphere. One analogous statement is used in, where they prove that one can apply a randomized mapping from the general Euclidean space to points on a unit sphere, while preserving the ratio of distances for any two pairs of points. The claim derives by the simple observation that an \(r\)-net in the initial space can be approximated by computing an \(\epsilon r/c\)-net on the sphere, where \(c\) is the maximum norm of any given point envisaged as vector. Our exposition is even simpler since we can directly employ the analogous theorem from. ## Approximate nets under Euclidean distance In this subsection, we show that one can translate the problem of computing an \(r\)-net for points on the unit sphere under Euclidean distance, to finding an \(r\)-net for unit vectors under inner product as defined in Section [2](#SInner){reference-type="ref" reference="SInner"}. Moreover, we identify the subset of the \(r\)-net which contains the centers that are approximately far from any other point. Formally, If \(r\) is greater than some constant, the problem can be immediately solved by the law of cosines. If \(r\) cannot be considered as constant, we distinguish cases \(r\geq 1/n^{0.9}\) and \(r <1/n^{0.9}\). The first case is solved by a simple modification of an analogous algorithm in. The second case is not straightforward and requires partitioning the pointset in a manner which allows computing \(r\)-nets for each part separately. Each part has bounded diameter which implies that we need to solve a "large \(r\)\" subproblem. Let us now present an algorithm which translates the problem of finding an \(r\)-net for \(r<1/n^{0.9}\) to the problem of computing an \(r\)-net for \(r\geq 1/n^{0.9}\). The main idea is that we compute disjoint subsets \(S_i\), which are far enough from each other, so that we can compute \(r\)-nets for each \(S_i\) independently. We show that for each \(S_i\) we can compute \(T_i \subseteq S_i\) which has bounded diameter and \(T_i'\subseteq S_i\) such that \(T_i\), \(T_i'\) are disjoint, each point in \(T_i\) is far from each point in \(T_i'\), and \(|T_i'|\leq 3|S_i|/4\). It is then easy to find \(r\)-nets for \(T_i\) by employing the ApprxNet(Large radius) algorithm. Then, we recurse on \(T_i'\) which contains a constant fraction of points from \(|S_i|\). Then, we cover points in \(S_i \setminus(T_i \cup T_i')\) and points which do not belong to any \(S_i\). We now present an algorithm for an \((1+\epsilon)r\)-net for points in \({\mathbb R}^d\) under Euclidean distance. # Applications and Future work {#Sapps} Concerning applications, in, they design an approximation scheme, which solves various distance optimization problems. The technique employs a grid-based construction of \(r\)-nets which is linear in \(n\), but exponential in \(d\). The main prerequisite of the method is the existence of a linear-time decider (formally defined in Appendix [\[Aframework\]](#Aframework){reference-type="ref" reference="Aframework"}). The framework is especially interesting when the dimension is constant, since the whole algorithm costs time linear in \(n\) which, for some problems, improves upon previously known near-linear algorithms. When the dimension is high, we aim for polynomial dependency on \(d\), and subquadratic dependency on \(n\). Let us focus on the problem of approximating the *\(k\)th nearest neighbor distance*. Now we present an approximate decider for the problem above. This procedure combined with the framework we mentioned earlier, which employs our net construction, results in an efficient solution for this problem in high dimension. To the best of our knowledge, this is the first high dimensional solution for this problem. Setting \(k=n\) and applying Theorem [\[KND\]](#KND){reference-type="ref" reference="KND"} one can compute the *farthest nearest neighbor* in \(\tilde{O}(dn^{2-\Theta(\sqrt{\epsilon})})\) with high probability. Concerning future work, let us start with the problem of finding a greedy permutation. A permutation \(\Pi = <\pi_1, \pi_2,\dots >\) of the vertices of a metric space \((X, \norm{\cdot})\) is a *greedy permutation* if each vertex \(\pi_i\) is the farthest in \(X\) from the set \(\Pi_{i-1} = <{\pi_1,\dots, \pi_{i-1}}>\) of preceding vertices. The computation of \(r\)-nets is closely related to that of the greedy permutation. The \(k\)-center clustering problem asks the following: given a set \(X \subseteq {\mathbb R}^d\) and an integer \(k\), find the smallest radius \(r\) such that \(X\) is contained within \(k\) balls of radius \(r\). By, a simple modification of our net construction implies an algorithm for the \((1+\epsilon)\) approximate greedy permutation in time \(\tilde{O}(d n^{2-\Theta(\sqrt{\epsilon})} \log \Phi)\) where \(\Phi\) denotes the spread of the pointset. Then, approximating the greedy permutation implies a \((2+\epsilon)\) approximation algorithm for \(k\)-center clustering problem. We expect that one can avoid any dependencies on \(\Phi\). The Corollaries below follow from Theorem [\[ApprxNet\]](#ApprxNet){reference-type="ref" reference="ApprxNet"}, Lemma 3.5 and Lemma 2.1.
{'timestamp': '2017-05-09T02:03:40', 'yymm': '1607', 'arxiv_id': '1607.04755', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04755'}
null
null
# Introduction Lately there was a growing interest in studying self-similarity and fractal properties of graphs, which is largely inspired by applications in biology, sociology and chemistry. Such studies often employ statistical physics methods that use ideas from graph theory and general topology, but are not intended to approach the problems under consideration in a rigorous mathematical way. Several studies that translate certain notions of topological dimension theory to graphs using combinatorial methods are also known. However, to the best of our knowledge a rigorous combinatorial theory that defines and studies graph-theoretical analogues of topological fractals still has not been developed. In this paper we introduce and study graph analogues of Lebesgue and Hausdorff dimensions of topological spaces from the graph-theoretical point of view. We show that they are closely related to well-known graph characteristics such as rank dimension and Prague (or Nešetřil-Rödl) dimension . It allowed us to define fractal graphs and determine fractality of some graph classes. In particular, it occurred that fractal graphs in some sense could be considered as generalizations of class 2 graphs. We demonstrate that various properties of dimensions of compact topological spaces have graph-theoretical analogues. Moreover, we also show how these relations allow for reverse transfer of combinatorial results to the general topology by proving a new property of general compact metric spaces using machinery from theory of hypergraphs. Finally, we show how Hausdorff (and Prague) dimension of graphs is related to Kolmogorov complexity. This relation allowed us to find a lower bound for Prague dimension for almost all graphs using incompressibility method from the theory of Kolmogorov complexity. # Basic definitions and facts from measure theory, dimension theory and graph theory {#sec:examples} Let \(X\) be a compact metric space. A family \(\mathcal{C} = \{C_{\alpha}: \alpha \in A\}\) of open subsets of \(X\) is a *cover*, if \(X = \bigcup_{\alpha \in A} C_{\alpha}\). A cover \(\mathcal{C}\) is *\(k\)-cover*, if every \(x\in X\) belongs to at most \(k\) sets from \(\mathcal{C}\); *\(\epsilon\)-cover*, if for every set \(C_i\in \mathcal{C}\) its diameter \(diam(C_i)\) does not exceed \(\epsilon\); \((\epsilon, k)\)-cover, if it is both \(\epsilon\)-cover and \(k\)-cover. *Lebesgue dimension* (*cover dimension*) \(dim_L(X)\) of the space \(X\) is the minimal integer \(k\) such that for every \(\epsilon > 0\) there exists \((\epsilon, k+1)\)-cover of \(X\). Let \(\mathcal{F}\) be a semiring of subsets of a set X. A function \(m:\mathcal{F}\rightarrow \mathbb{R}^+_0\) is *a measure*, if \(m(\emptyset) = 0\) and for any disjoint sets \(A,B\in \mathcal{F}\) \(m(A\cup B) = m(A) + m(B)\). Let now \(X\) be a subspace of an Euclidean space \(\mathbb{R}^d\). *Hyper-rectangle* \(R\) is a Cartesian product of semi-open intervals: \(R = [a_1,b_1)\times\dots\times [a_d,b_d)\), where \(a_i < b_i\), \(a_i,b_i \in \mathbb{R}\); the *volume* of a the hyper-rectangle \(R\) is defined as \(vol(R) = \prod_{i=1}^d (b_i-a_i)\). The *\(d\)-dimensional Jordan measure* of the set \(X\) is the value \(\mathcal{J}^d(X) = \inf\{\sum_{R\in \mathcal{C}} vol(R)\},\) where infimum is taken over all finite covers \(\mathcal{C}\) of \(X\) by disjoint hyper-rectangles. The *\(d\)-dimensional Lebesgue measure* of a measurable set \(\mathcal{L}^d(X)\) is defined analogously, with the additional condition that infimum is taken over all countable covers \(\mathcal{C}\) of \(X\) by (not necessarily disjoint) hyper-rectangles. Let \(s >0\) and \(\epsilon > 0\). Consider the parameter \(\mathcal{H}^s_{\epsilon}(X) = \inf\{\sum_{C\in \mathcal{C}}diam(C)^s\},\) where infimum is taken over all \(\epsilon\)-covers of \(X\). The *\(s\)-dimensional Hausdorff measure* of the set \(X\) is defined as \(\mathcal{H}^s(X) = \lim_{\epsilon \rightarrow 0} \mathcal{H}^s_{\epsilon}(X)\). The aforementioned measures are related as follows. If Jordan measure of the set \(X\) exists, then it is equal to its Lebesgue measure. For Borel sets Lebesgue measure and Hausdorff measure are equivalent in the sense, that for any Borel set \(Y\) we have \(\mathcal{L}^d(Y) = C_d\mathcal{H}^d(Y)\), where \(C_d\) is a constant depending only on \(d\). *Hausdorff dimension* \(dim_H(X)\) of the set \(X\) is the value \[\label{hdimtop} dim_H(X) = \inf\{s \geq 0 :\mathcal{H}^s(X) < \infty\}.\] Lebesgue and Hausdorff dimension of \(X\) are related as follows: \[\label{lebvshaus} dim_L(X) \leq dim_H(X).\] The set \(X\) is *a fractal*, if the inequality ([\[lebvshaus\]](#lebvshaus){reference-type="ref" reference="lebvshaus"}) is strict. Let \(G=(V(G),E(G))\) be a simple graph. The family of subgraphs \(\mathcal{C} = \{C_1...,C_m\}\) of \(G\) is a *cover*, if every edge \(uv\in E(G)\) belongs to at least one subgraph from \(\mathcal{C}\). A cover \(\mathcal{C}\) is *\(k\)-cover*, if every vertex \(v\in V(G)\) belongs to at most \(k\) subgraphs of \(\mathcal{C}\); *clique cover*, if all subgraphs \(C_i\) are cliques. A set \(W\subseteq V(G)\) *separates* vertices \(u,v\in V(G)\), if \(|W\cap \{u,v\}| = 1\). For a hypergraph \(\mathcal{H} = (\mathcal{V}(\mathcal{H}),\mathcal{E}(\mathcal{H}))\), its *rank* \(r(\mathcal{H})\) is the maximal size of its edges. A hypergraph \(\mathcal{H}\) is *strongly \(k\)-colorable*, if for every vertex a color from the set \(\{1...,k\}\) can be assigned in such a way that vertices of every edge receive different colors. Intersection graph \(L = L(\mathcal{H})\) of a hypergraph \(\mathcal{H}\) is a simple graph with a vertex set \(V(L) = \{v_E: E\in \mathcal{E}(\mathcal{H})\}\) in a bijective correspondence with the edge set of \(\mathcal{H}\) and two distinct vertices \(v_E,v_F\in V(L)\) being adjacent, if and only if \(E\cap F \ne \emptyset\). The following theorem establishes a connection between intersection graphs and clique \(k\)-covers: *Rank dimension* \(dim_R(G)\) of a graph \(G\) is the minimal \(k\) such that \(G\) satisfies conditions of Theorem [\[thm:covervsinter\]](#thm:covervsinter){reference-type="ref" reference="thm:covervsinter"}. In particular, graphs with \(dim_R(G)=1\) are disjoint unions of cliques (such graphs are called *equivalence graphs* or *\(M\)-graphs* ), and graphs with \(dim_R(G)=2\) are line graphs of multigraphs. *Categorical product* of graphs \(G_1\) and \(G_2\) is the graph \(G_1 \times G_2\) with the vertex set \(V(G_1\times G_2) = V(G_1)\times V(G_2)\) with two vertices \((u_1,u_2)\) and \((v_1,v_2)\) being adjacent whenever \(u_1v_1\in E(G_1)\) and \(u_2v_2\in E(G_2)\). *Prague dimension* \(dim_P(G)\) is the minimal integer \(d\) such that \(G\) is an induced subgraph of a categorical product of \(d\) complete graphs. *Equivalent cover* \(\mathcal{M} = \{M_1...,M_k\}\) of the graph \(G\) consists of spanning subgraphs such that each subgraph \(M_i\) is an equivalence graph. Equivalent cover is *separating*, if every two distinct vertices of \(G\) are separated by one of connected components in some subgraph from \(\mathcal{M}\). Relations between Prague dimension, clique covers, vertex labeling and intersection graphs are described by the following theorem: Numbers of vertices and edges of a graph \(G\) are denoted by \(n\) and \(m\), respectively. A subgraph of \(G\) induced by vertex subset \(U\subseteq V(G)\) is denoted as \(G[U]\). For two graphs \(G_1\) and \(G_2\) the notation \(G_1 \leq G_2\) indicates, that \(G_1\) is an induced subgraph of \(G_2\). # Lebesgue dimension of graphs {#lebgraph} Lebesgue dimension of a metric space is defined through \(k\)-covers by sets of arbitrary small diameter. It is natural to transfer this definition to graphs using graph \(k\)-covers by subgraphs of smallest possible diameter, i.e. by cliques. Thus by Theorem [\[thm:covervsinter\]](#thm:covervsinter){reference-type="ref" reference="thm:covervsinter"} we define Lebesgue dimension of a graph through its rank dimension: \[\label{kdimeqleb} dim_L(G) = dim_R(G)-1.\] An analogy between Lebesgue and rank dimensions is futher justified by the following Proposition [\[topspacehyper\]](#topspacehyper){reference-type="ref" reference="topspacehyper"}, that states that any compact metric spaces of bounded Lebesgue measure could be approximated by intersection graphs of (infinite) hypergraphs of bounded rank. To prove it, we will use the following fact: So, \(dim_L(X)\leq k\) whenever for any \(\epsilon > 0\) there is a well-defined hypergraph \(\mathcal{H}(\epsilon)\) of \(rank(\mathcal{H}(\epsilon)) \leq k\) with edges in bijective correspondence with points of \(X\) such that two points are close if and only if corresponding edges intersect. # Graph measure and Hausdorff dimension of graphs {#measuregraph} In order to rigorously define a graph analogue of Hausdorff dimension, we need to define first a corresponding measure. Note that in any meaningful finite graph topology every set is a Borel set. As mentioned above, for measurable Borel sets in \(\mathbb{R}^n\) Jordan, Lebesgue and Hausdorff measures are equivalent. Thus further we will work with a graph analogue of Jordan measure. It is known, that every graph is isomorphic to an induced subgraph of a categorical product of complete graphs. Consider a graph \(G\) embedded into a categorical product of \(d\) complete graphs \(K^1_{n_1}\times\dots \times K^d_{n_d}\). Without loss of generality we may assume that \(n_1 =... = n_d = n\), i.e. \[\label{graphembed} G\cong G' \leq S = (K_n)^d,\] Suppose that \(V(K^n) = \{1,\dots,n\}\). The graph \(S\) will be referred to as *a space* of dimension \(d\) and \(G'\) as an embedding of \(G\) into \(S\). It is easy to see, that by definition every vertex \(v\in S\) is a vector \(v = (v_1...,v_d)\) with \(v_i\in [n]\), and two vertices \(u\) and \(v\) are adjacent in \(S\) if and only if \(v_r\ne u_r\) for every \(r\in [d]\). *Hyper-rectangle* \(R = R(J_1...,J_d)\) is a subgraph of \(S\), that is defined as follows: for every \(i=1...,d\) choose a non-empty subset \(J_i \subseteq [n]\), then \(R = K_n[J_1]\times\dots\times K_n[J_d]\). The *volume* of a hyper-rectangle \(R\) is the value \(vol(R) = |V(R)| = \prod_{i=1}^d |J_i|\). The family \(\mathcal{R} = \{R^1...,R^m\}\) of hyper-rectangles is a *rectangle co-cover* of \(G'\), if the subgraphs \(R^i\) are pairwise vertex-disjoint, \(V(G')\subseteq \bigcup_{i=1}^m V(R^i)\) and \(\mathcal{R}\) covers all non-edges of \(G'\), i.e. for every \(x,y\in V(G')\), \(xy\not\in E(G')\) there exists \(j\in[m]\) such that \(x,y\in V(R^j)\). We define *\(d\)-volume* of a graph \(G\) as \[\label{grvol} vol^d(G) = \min_{G'}\min_{\mathcal{R}} \sum_{R\in \mathcal{R}} vol(R),\] where the first minimum is taken over all embeddings \(G'\) of \(G\) into \(d\)-dimensional spaces \(S\) and the second minimum-over all rectangle co-covers of \(G'\) (see Fig. [\[fig:embedP4\]](#fig:embedP4){reference-type="ref" reference="fig:embedP4"}). We define a *\(d\)-measure* of a graph \(F\) as follows: \[\label{neasuregraph} \mathcal{H}^d(F) = \left \{ \begin{array}{lll} vol^d(\overline{F}), \text{ if } \overline{F} \text{ can be represented as (\ref{graphembed})}; \\ +\infty, \text{ otherwise } \end{array} \right.\] In the remaining part of this section we will prove, that \(\mathcal{H}^d\) indeed satisfy the property of a measure, i.e. \(\mathcal{H}^d(F^1\cup F^2) = \mathcal{H}^d(F^1) + \mathcal{H}^d(F^2)\), where \(F^1\cup F^2\) is a disjoint union of graphs \(F^1\) and \(F^2\). Let \(W^1,W^2\subseteq V(S)\). We write \(W_1\sim W_2\), if every vertex from \(W_1\) is adjacent to every vertex from \(W_2\). Denote by \(P_k(W_1)\) *the \(k\)-th projection* of \(W_1\), i.e. the set of all \(k\)-coordinates of vertices of \(W_1\): \[J_k(W_1) = \{v_k: v\in W_1\}.\] In particular, \(P_k(R(J_1...,J_d)) = J_k\). The following proposition follows directly from the definition of \(S\) Assume that \(\mathcal{R} = \{R^1...,R^m\}\) is a minimal rectangle co-cover of a minimal embedding \(G'\), i.e. \(vol^d(G) = \sum_{R\in \mathcal{R}} vol(R)\). Further we will demonstrate, that \(\mathcal{R}\) has a rather simple structure. Let \(J_k^i = P_k(R^i)\). Following the analogy with Hausdorff dimension of topological spaces ([\[hdimtop\]](#hdimtop){reference-type="ref" reference="hdimtop"}), we define a *Hausdorff dimension* of a graph \(G\) as \[\label{hdimgraph} dim_H(G) = \min\{s \geq 0 :\mathcal{H}^s(G) < \infty\}-1.\] Thus, Hausdorff dimension of a graph can be identified with a Prague dimension of its complement minus 1. # Relations with Kolmogorov complexity Let \(\mathbb{B}^*\) be the set of all finite binary strings and \(\Phi:\mathbb{B}^* \rightarrow \mathbb{B}^*\) be a computable function. A *Kolmogorov complexity* \(K_{\Phi}(s)\) of a binary string \(s\) with respect to \(\Phi\) is defined as a minimal length of a string \(s'\) such as \(\Phi(s') = s\). Since Kolmogorov complexities with respect to any two functions differ only by an additive constant, it is usually assumed that some canonical function \(\Phi\) is fixed, and Kolmogorov complexity is denoted simply by \(K(s)\). Thus, informally \(K(s)\) could be described as a length of a shortest encoding of the string \(s\), that allows to completely reconstruct it. Analogously, for two strings \(s,t\in \mathbb{B}^*\), a *conditional Kolmogorov complexity* \(K(s|t)\) is a a length of a shortest encoding of \(s\), if \(t\) is known in advance. More information on properties of Kolmogorov complexity can be found in. Every graph \(G\) can be naturally encoded using the string representation of an upper triangle of its adjacency matrix. Kolmogorov complexity of a graph could be defined as a Kolmogorov complexity of that string. It gives estimations \(K(G) = O(n^2)\), \(K(G|n) = O(n^2)\). Alternatively, \(n\)-vertex connected labeled graph can be represented as a list of edges with ends of each edge encoded using their binary representations concatenated with a binary representation of \(n\). It gives estimations \(K(G) \leq 2m\log(n) + \log(n) = O(m\log(n))\), \(K(G|n) \leq 2m\log(n) = O(m\log(n))\). Further in this section for simplicity we will consider connected graphs (for disconnected graphs all considerations below could be applied to every connected component). Let \(dim_H(G) = dim_P(\overline{G})-1 = d-1\) and \(\mathcal{H}^d(G) = h\). Then by Corollary [\[cor:minvolspace\]](#cor:minvolspace){reference-type="ref" reference="cor:minvolspace"} \(\overline{G}\) is an induced subgraph of a product \[\label{graphembedmin} K_{p_1}\times\dots \times K_{p_d},\] where \(h = p_1\cdot\dots \cdot p_d\). So, by Theorem [\[thm:pdimcharact\]](#thm:pdimcharact){reference-type="ref" reference="thm:pdimcharact"}, \(G\) and \(\overline{G}\) could be encoded using a collection of vectors \(\phi(v) = (\phi_1(v),\dots,\phi_d(v))\), \(v\in V(G)\), \(\phi_j(v)\in [p_j]\). Such encoding could be stored as a string containing binary representations of coordinates \(\phi_j(v)\) using \(\log(p_j)\) bits concatenated with a binary representations of \(n\) and \(p_j\), \(j=1...,n\). The length of this string is \((n+1)\sum_{j=1}^d \log(p_j) + \log(n)\). Analogously, if \(n\) and \(p_j\) are given, then the length of encoding is \(n\sum_{j=1}^d \log(p_j)\). Thus, the following estimation is true: Let \(p^* = \max_j p_j\). Then we have \(K(G)\leq (n+1)d\log(p^*) + \log(n),\[K(G|n,p_1...,p_d)\leq nd\log(p^*).\) By minimality of the representation ([\[graphembedmin\]](#graphembedmin){reference-type="ref" reference="graphembedmin"}), we have \(p^* \leq n\). Thus \(K(G) = O(dn\log(n))\), \(K(G|n,p_1...,p_d) = O(dn\log(n))\). So, Hausdorff (and Prague) dimension could be considered as a measure of descriptive complexity of a graph. In particular, for graphs with a small Hausdorff dimension ([\[eq:KolG\]](#eq:KolG){reference-type="ref" reference="eq:KolG"})-([\[eq:KolcondG\]](#eq:KolcondG){reference-type="ref" reference="eq:KolcondG"}) give better estimation of their Kolmogorov complexity than the standard estimations mentioned above. Relations between Hausdorff (Prague) dimension and Kolmogorov complexity could be used to derive lower bound for Hausdorff (and Prague) dimension in a typical case. More rigorously, let \(X\) be a graph property and \(\mathcal{P}_n(X)\) be the set of labeled n-vertex graphs having \(X\). Property \(X\) holds for *almost all graphs*, if \(|\mathcal{P}_n(X)|/2^{\binom{n}{2}} \rightarrow 1\) as \(n \rightarrow \infty\). We will use the following lemma: Then the following theorem is true: Considerations above and proof of Theorem [\[thm:almostalldim\]](#thm:almostalldim){reference-type="ref" reference="thm:almostalldim"} imply that for every \(\epsilon > 0\) and \(n\)-vertex graph \(G\) with sufficiently large \(n\), \(dim_H(G) \geq \frac{1}{(1+\epsilon)\log(p^*)}\frac{K(G)}{n}\). Interestingly, similar relations between Kolmogorov complexity and analogues of Hausdorff dimension hold for other objects. In particular, for Cantor space \(\mathcal{C}\) (the space of all infinite 0-1 sequences) it is proved in that Kolmogorov complexity and effective (or constructive) Hausdorff dimension \(dim_H(s)\) of each sequence \(s\) are related as follows: \(dim_H(s)= \liminf\limits_{n \rightarrow \infty} \frac{K(s_n)}{n}\) (here \(s_n\) is the prefix of \(s\) of length \(n\)). Similar estimations are known for other variants of Hausdorff dimension. # Fractal graphs {#fracgraph} Importantly, the relation ([\[lebvshaus\]](#lebvshaus){reference-type="ref" reference="lebvshaus"}) between Lebesgue and Hausdorff dimensions of topological spaces remains true for graphs. Definitions of dimensions immediately imply, that both Lebesgue and Hausdorff dimensions are monotone with respect to induced subgraphs. Analogously to the definition of fractals for topological spaces, we say, that a graph \(G\) is a *fractal*, if \(dim_L(G) < dim_H(G)\), i.e. \(dim_R(G) < dim_P(\overline{G})\). The following proposition provides a first non-trivial example of fractal graphs As another example, consider Sierpinski gasket graphs \(S_n\). This class of graphs is associated with the Sierpinski gasket-well-known topological fractal with a Hausdorff dimension \(\log(3)/\log(2)\approx 1.585\). Edges of \(S_n\) are line segments of the \(n\)-th approximation of the Sierpinski gasket, and vertices are intersection points of these segments (Fig. [\[fig:serpGasket\]](#fig:serpGasket){reference-type="ref" reference="fig:serpGasket"}). Sierpinski gasket graphs can be defined recursively as follows. Consider tetrads \(T_n = (S_n,x_1,x_2,x_3)\), where \(x_1,x_2,x_3\) are distinct vertices of \(S_n\) called *contact vertices*. The first Sierpinski gasket graph \(S_1\) is a triangle \(K_3\) with vertices \(x_1,x_2,x_3\), the first tetrad is defined as \(T_1 = (S_1,x_1,x_2,x_3)\). The \((n+1)\)-th Sierpinski gasket graph \(S_{n+1}\) is constructed from 3 disjoint copies \((S_n,x_1,x_2,x_3)\), \((S'_n,x'_1,x'_2,x'_3)\), \((S''_n,x''_1,x''_2,x''_3)\) of \(n\)-th tetrad \(T_n\) by gluing together \(x_2\) with \(x'_1\), \(x'_3\) with \(x''_2\) and \(x_3\) with \(x''_1\); the corresponding \((n+1)\)-th tetrad is \(T_{n+1} = (S_{n+1},x_1,x'_2,x''_3)\). It is important to emphasize that Lebesgue and Hausdorff dimensions of Sierpinski gasket graphs agree with the corresponding dimensions of Sierpinski gasket fractal. As a contrast, note that rectangular grid graphs (cartesian products of 2 paths) are not fractals (by König theorem since they are bipartite), just like rectangles are not fractals in \(\mathbb{R}^2\). # Acknowlegements The work of LB was partially supported by the NSF grant DMS-1600568
{'timestamp': '2016-07-26T02:06:12', 'yymm': '1607', 'arxiv_id': '1607.04703', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04703'}
null
null
# Introduction In information theory, the *index coding* problem  is the following: A sender wishes to *broadcast* over a noiseless channel an \(n\)-symbol string \(x\in \mathbb{F}^n\) to a group of \(n\) receivers \(R_1,\ldots,R_n\), each equipped with some *side information*, namely, a subvector \(x_{K_i}\) of \(x\) indexed by a subset \({K_i} \subseteq \{x_1,\ldots ,x_n\}\). The index coding problem asks what is the minimum length \(m\) of a broadcast message that allows each receiver \(R_i\) to retrieve the \(i\)th symbol \(x_i\), given his side-information \(x_{K_i}\) and the broadcasted message. The side information of the receivers can be modeled by a directed graph \(\mathcal{K}_n\), in which \(R_i\) observes
{'timestamp': '2017-02-17T02:04:40', 'yymm': '1607', 'arxiv_id': '1607.04842', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04842'}