Datasets:

ccdv

/

arxiv-classification

like 19

Stochastic Input Models in Online Computing arXiv:1709.07601v1 [cs.DS] 22 Sep 2017 Yasushi Kawase∗ Tokyo Institute of Technology, Tokyo, Japan kawase.y.ab@m.titech.ac.jp Abstract In this paper, we study twelve stochastic input models for online problems and reveal the relationships among the competitive ratios for the models. The competitive ratio is defined as the worst ratio between the expected optimal value and the expected profit of the solution obtained by the online algorithm where the input distribution is restricted according to the model. To handle a broad class of online problems, we use a framework called request-answer games that is introduced by Ben-David et al. The stochastic input models consist of two types: known distribution and unknown distribution. For each type, we consider six classes of distributions: dependent distributions, deterministic input, independent distributions, identical independent distribution, random order of a deterministic input, and random order of independent distributions. As an application of the models, we consider two basic online problems, which are variants of the secretary problem and the prophet inequality problem, under the twelve stochastic input models. We see the difference of the competitive ratios through these problems. 1 Introduction In online computing, we are given a sequence of requests, and we need to choose an action in each step based on the current information without knowing the full information which will be completely obtained in the future [6, 25]. Since an online algorithm is forced to make decisions without knowing the entire inputs, they may later turn out not to be optimal. The quality of an online algorithm is usually measured by the competitive ratio [35], which is the worst ratio between the optimal value and the profit of the solution obtained by the online algorithm. However, for practical application, it is too pessimistic because the worstcases rarely occur in real-world. Thus, an average-case analysis is more suitable than the worst one in such a situation. In order to analyze average performance, we need to assume some distribution of inputs. Hence, in this paper, we consider stochastic input models. In 2004, Hajiaghayi, Kleinberg, and Parkes [16] studied the case where requests are drawn independently from some unknown distribution and the case where the requests sequence is determined by picking a multi-set of requests and then permuting them randomly. However, we can define various stochastic input models, and we need to consider which should we use depending on analysis objective. Various studies related to online problems with a stochastic input have been extensively conducted on statistics literature. A most famous example of the problems is secretary problem [14, 15, 34]. In the (classical) secretary problem, a decision-maker is willing to hire the best secretary out of n applicants that arrive in a random order, and the goal is to maximize the probability of choosing the best applicant. As each applicant appears, it must be either selected or rejected, and the decision is irrevocable. It is assumed that the decision must be based only on the relative ranks (without ties) of the applicants seen so far and the number of applicants n. It is well known that one can succeed with the optimal probability 1/e by the following algorithm: observe the first n/e applicants without selecting, and then select the next applicant who is the best among the observed applicants [12]. ∗ Supported by a Grant-in-Aid for Young Scientists (B) (No. 16K16005). 1 The above setting is also referred to the no-information case because there is no information about the value of applicants. When the decision-maker is allowed to observe the actual values of the applicants, the problem is called the full-information secretary problem if the values are chosen independently and identically from a known distribution. Also, if the values are chosen independently and identically from a distribution with an unknown parameter, it is called the partial-information secretary problem. Moreover, when the values are chosen independently from known (not necessarily identical) distributions, the problem is studied as prophet inequality. Krengel, Sucheston, and Garling [27] provided a tight 2-competitive algorithm: select the first value that is at least half of the expected maximum. There are also a huge number of researches on the stochastic input models in online learning literature, such as online (convex) optimization, online prediction, online classification, and multi-armed bandit [10, 9, 18]. In many studies in the literature, it is assumed that input is drawn from an unknown distribution and one repeatedly performs the same task. The main purpose of this paper is to compare stochastic input models in the sense of competitive ratios. As a previous result, Mehta [32] provided a relation between four models: unknown deterministic input, random order of an unknown deterministic input, unknown i.i.d., and known i.i.d. However, he considered only the online allocation problems. On the other hand, this paper studies a more general form of online problems through request-answer games. Our results In this paper, we provide twelve stochastic input models for online problems and reveal the relationships among the competitive ratios for the models. The models consist of two types: known distribution and unknown distribution. For each type, we consider six classes of distributions: dependent distributions (dep), deterministic input (det), independent distributions (ind), identical independent distribution (i.i.d.), random order of a deterministic input (r.d.), and random order of independent distributions (r.i.). The competitive ratio is defined as the worst ratio between the expected optimal value and the expected profit of the solution obtained by the online algorithm where the input (request) distribution is restricted according to the model. We will denote by CRmodel (resp. CRmodel ) the competitive ratio with unknown (resp. known) distribution model. For example, CRdet , CRr.d. , CRi.i.d. , CRi.i.d. represent competitive ratios for unknown deterministic input, random order of an unknown deterministic input, unknown i.i.d., and known i.i.d., respectively. Then, our results can be summarized as Figure 1. CRind CRdep CRdet CRdet CRi.i.d. CRi.i.d. CRr.d. CRr.d. CRr.i. CRr.i. CRind CRdep Figure 1: The relationships of the competitive ratios. Each arrow represents an inequality (the value at the head of each arrow is at most the one at its tail) and ratios in the same region have the same value. Related work Ben-David et al. [5] introduced a general framework of online problems, which is called request-answer games. They compared oblivious, adaptive-online, and adaptive-offline adversaries. The request-answer games can be seen as the following two-person zero-sum games. The first player seeks an online algorithm that minimizes 2 the competitive ratio; the second player seeks a request sequence that maximizes the competitive ratio. We can view a randomized online algorithm as a mixed strategy of the first player. By using von Neumann’s minimax theorem [38] and Loomis’s lemma [29], we can get a technique to obtain an upper bound of the competitive ratio. This technique is called Yao’s principle [39] and used for various online problems. Many generalizations of the secretary problem, i.e., problems in the random order model, have been studied in the competitive analysis literature. As a natural generalization, Hajiaghayi, Kleinberg, and Parkes [16] introduced the multiple-choice secretary problem and Kleinberg [23] provided an asymptotically optimal competitive algorithm when the objective is to maximize the sum of the k choices. Babaioff, Immorlica, and Kleinberg [4] introduced the matroid secretary problem. In this model, the set of selected applicants must be an independent set in an underlying matroid. Lachish [28] provided O(log log r)-competitive algorithm where r is the rank of given matroid. Korula and Pál [26] presented 8-competitive algorithm for bipartite matching and then Kesselheim et al. [22] provide a tight e-competitive algorithm. Babaioff et al. [2] introduced knapsack secretary problem and proposed 10e-competitive algorithm. For more details, see Dinitz [11] and Babaioff et al. [3]. There are also a number of researches for generalizations of prophet inequalities. Hajiaghayi, Kleinberg, and Sandholm [17] considered multiple-choice variant and they provided a prophet inequality when the objective is to maximize the sum of the k choices. Kleinberg and Weinberg [24] studied the matroid constrained version and gave a tight 2-competitive algorithm. Rubinstein and Singla [33] introduced framework for combinatorial valuation functions. Moreover, Esfandiari et al. [13] introduced a prophet secretary problem, which is a natural combination of the prophet inequality problem and the secretary problem. In their setting, the values are chosen independently from known (not necessarily identical) distributions but arrive in a random order. Paper Organization Section 2 describes formal definitions of our models via request-answer games. Section 3 discusses some variants of the secretary problem and the prophet inequality problem for our models. In Section 4, we give the proof of our results shown in Figure 1 and show that we cannot simplify the relations. 2 Model In this section, we introduce request-answer games with stochastic input models. A request-answer system is a tuple (R, A, f, n) where each component is defined as follows. There is a set of requests R and a set of answers A. Throughout the paper, we assume that the sets R and A are finite. A positive integer n represents the number of rounds. Let f : Rn × An → R+ denote a utility function, where R+ is the set of nonnegative real numbers. A deterministic online algorithm ALG is a sequence of functions gi : Ri → A (i = 1, 2, . . . , n). Given a deterministic online algorithm ALG = (g1 , . . . , gn ) and a request sequence r = (r1 , . . . , rn ) ∈ Rn , the output is an answer sequence ALG[r] = (a1 , . . . , an ) ∈ An where ai = gi (r1 , . . . , ri ) for i = 1, . . . , n. The utility incurred by ALG on r, denoted by ALG(r) is defined as ALG(r) = f (r, ALG[r]). A randomized online algorithm ALG is defined as a probability distribution over the set of all deterministic online algorithms A = {(g1 , . . . , gn ) | gi : Ri → A (i = 1, . . . , n)}. The utility incurred by ALG on r, denoted by ALG(r) is defined as ALG(r) = Ex [f (r, ALGx [r])] where we use Ex to denote the expectation with respect to the distribution over the set A = {ALGx }, which defines ALG. We will denote by A the set of randomized algorithms. We remark that A is a convex compact set since A is a finite set by the assumption that R and A are finite. 3 The performance of an online algorithm is measured by the competitive ratio—the ratio between its value and the optimal value for the worst request sequence. Note that we only discuss in this paper the oblivious adversary, i.e., the request sequence does not depend on the randomized results of the algorithm. This is because we want to consider weaker adversaries than the standard model and the adaptive adversary and stochastic input models are incompatible in most cases. Given a request sequence r ∈ Rn , the optimal offline utility on r is defined as OPT(r) = max{f (r, a) | a ∈ An }. Let Sn be the set of permutations of [n], where [n] = {1, . . . , n}. For a permutation σ ∈ Sn and a sequence x = (x1 , . . . , xn ), let us denote xσ = (xσ(1) , . . . , xσ(n) ). Definition 2.1 (Competitive Ratios). For a request-answer game (R, A, f, n), a randomized online algorithm ALG ∈ A is ρ-competitive for a distribution D over the request sequences Rn if Er∼D [OPT(r)] ≤ ρ · Er∼D [ALG(r)]. Let D be a compact subset1 of the distributions over the request sequences Rn . We define the competitive ratio of (R, A, f, n) for unknown distribution in D as CRD (R, A, f, n) = inf sup ALG∈A D∈D Er∼D [OPT(r)] Er∼D [ALG(r)] (1) where we define 0/0 = 1. In other words, there exists a randomized online algorithm that is CRD (R, A, f, n)competitive for any D ∈ D, and this is best possible. Similarly, we define the competitive ratio of (R, A, f, n) for known distribution in D as Er∼D [OPT(r)] . D∈D ALG∈A Er∼D [ALG(r)] CRD (R, A, f, n) = sup inf (2) Namely, for any D ∈ D, there exists a CRD (R, A, f, n)-competitive algorithm, and this is best possible. We often omit the argument (R, A, f, n) if there is no confusion. The value of the competitive ratio is at least 1 and smaller is better. It should be noted that, for minimization problems, a randomized online algorithm ALG is ρ-competitive for a distribution D over the request sequences Rn if Er∼D [ALG(r)] ≤ ρ · Er∼D [OPT(r)]. We can also define the competitive ratio of a minimization version of request-answer games in the same manner. By the max-min inequality, the following inequality holds. Lemma 2.2. For any request-answer game (R, A, f, n) and a set of distributions D over the request sequences Rn , we have CRD (R, A, f, n) ≥ CRD (R, A, f, n). Moreover, if D is a compact convex set, the equality in the above inequality holds. Lemma 2.3. For any request-answer game (R, A, f, n) and a compact convex set of distributions D over the request sequences Rn , we have CRD (R, A, f, n) = CRD (R, A, f, n). 1 We assume compactness to simplify the exposition. 4 Proof. Since CRD (R, A, f, n) ≥ CRD (R, A, f, n) holds by Lemma 2.2, it is sufficient to prove CRD (R, A, f, n) ≤ CRD (R, A, f, n). Let c = CRD (R, A, f, n) and let h(D, ALG) = Er∼D [OPT(r) − c · ALG(r)]. Since [OPT(r)] , it holds that for any ALG ∈ A there exists D ∈ D c = CRD (R, A, f, n) = inf ALG∈A supD∈D EEr∼D r∼D [ALG(r)] such that h(D, ALG) ≥ 0. Thus we have inf sup h(D, ALG) ≥ 0. ALG∈A D∈D Moreover, by von Neumann’s minimax theorem [38], we have sup h(D, ALG) = sup inf ALG∈A D∈D inf h(D, ALG) D∈D ALG∈A since A and D are compact convex sets, and h(D, ALG) is linear by the linearity of expectations. Thus, supD∈D inf ALG∈A h(D, ALG) ≥ 0 and hence there exists D∗ ∈ D such that h(D∗ , ALG) ≥ 0 for any ALG ∈ A. Therefore, we obtain Er∼D [OPT(r)] D∈D ALG∈A Er∼D [ALG(r)] Er∼D∗ [OPT(r)] ≥ inf ALG∈A Er∼D∗ [ALG(r)] CRD (R, A, f, n) = sup inf ≥ c = CRD (R, A, f, n), which proves the lemma. P A distribution over a (finite) domain Ω is a function D : Ω → [0, 1] such that x∈Ω D(x) = 1. We denote by ∆(Ω) the set of all such distributions. For x ∈ Ω, let δx be a distribution such that δx (x) = 1 and δx (y) = 0 for any x, y ∈ Ω such that y 6= x. In this paper, we consider the following six classes of distributions. 1. Dependent distribution: ∆dep (Rn ) = ∆(Rn ). 2. Deterministic input : ∆det (Rn ) = {δr | r ∈ Rn }. 3. Independent distribution: ∆ind (Rn ) = ( D ∈ ∆(Rn ) D(r) = 4. Identical independent distribution: ( n ∆i.i.d.(R ) = n D ∈ ∆(R ) D(r) = n Y i=1 ) Di (ri ), Di ∈ ∆(R) . n Y i=1 ′ ′ ) D (ri ), D ∈ ∆(R) . 5. Random order of deterministic input : ∆r.d. (Rn ) = {ψr | r ∈ Rn } where ψr ∈ ∆(Rn ) (r ∈ Rn ) is a distribution such that ψr (r ′ ) = |{σ ∈ Sn | r ′ = r σ }|/n! for r ′ ∈ Rn . Here, recall that r σ = (rσ(1) , . . . , rσ(n) ). 6. Random order of independent distribution: ( n ∆r.i. (R ) = n D ∈ ∆(R ) D(r) = 5 n X Y σ∈Sn i=1 ) Dσ(i) (ri )/n!, Di ∈ ∆(R) . We will omit the argument Rn when no confusion can arise. In addition, we abbreviate CR∆dep (Rn ) (R, A, f, n) to CRdep and also we abbreviate the other competitive ratios in the same way. Now, we can formally state our main results, which was depicted in Figure 1. Theorem 2.4. For any request-answer game (R, A, f, n), we have CRdet ≤ CRi.i.d. ≤ CRi.i.d. ≤ CRr.d. = CRr.i. ≤ CRdep = CRdet = CRind = CRdep , CRdet ≤ CRr.d. ≤ CRr.i. ≤ CRr.i. , (4) CRi.i.d. ≤ CRind ≤ CRdep . (6) CRi.i.d. ≤ CRr.i. , 3 (3) (5) Examples of the Stochastic Input Models Before proving our main results, we see some examples of the stochastic input models. 3.1 Online selection problem Suppose that a decision-maker sequentially observes a sequence of random variables X1 , X2 , . . . , Xn and is allowed to choose only one number, which can be done only upon receiving that number. The goal is to maximize the expectation of the chosen value. We formalize the problem as a request-answer game. Let Rm = {0, 1, . . . , m} (where m is a sufficiently large integer), A = {0, 1}, and ( ri∗ (if {i∗ } = {i | ai = 1}), f (r, a) = 0 (otherwise). ai = 1 (resp. ai = 0) represents that the decision-maker selects (resp. rejects) ith value. The problem in the known independent distributions model is introduced by Krengel and Sucheston, and it is eagerly studied under the name of the prophet inequalities. It is well known that this problem is 2-competitive when n ≥ 2, i.e., lim supm→∞ CR∆ind (Rnm ) = 2 for any n ≥ 2 [21]. Hill and Kertz [19] gave a better upper bound of the competitive ratio in the known i.i.d. model and the current best upper bound is 1.354 [1]. For the known dependent distributions model, Hill and Kertz [20] proved that there exists no constant competitive algorithm, i.e., lim supn,m→∞ CR∆dep (Rnm ) = ∞. By Theorem 2.4, this implies that lim sup CR∆dep (Rnm ) = lim sup CR∆ind (Rnm ) = lim sup CR∆det (Rnm ) = ∞. n,m→∞ n,m→∞ n,m→∞ For random order of known independent distributions model (prophet secretary problem), Esfandiari et al. [13] provided e/(e − 1)-competitive algorithm and there is no online algorithm that can achieve a competitive ratio better than 15/11. Hence, it holds that 15/11 ≤ lim supn,m→∞ CR∆r.i. (Rnm ) ≤ e/(e − 1). When the given input is random order or i.i.d. distribution, it is easy to see that the secretary algorithm is e-competitive. On the other hand, Stewart [36] proves that no algorithm can select the best number with probability 1/e − o(1) when X1 , X2 , . . . , Xn are chosen i.i.d. from a uniform distribution on the interval (α, β) and (α, β) is chosen from a certain Pareto distribution. Transforming the variables Xi to M Xi , where M is a sufficient large number, implies that there is no online algorithm that is better than e-competitive. Thus we have lim supn,m→∞ CR∆r.d. (Rnm ) = lim supn,m→∞ CR∆r.i. (Rnm ) = lim supn,m→∞ CR∆i.i.d. (Rnm ) = e. For the other models, we can easily check that CR∆r.d. (Rnm ) = CR∆det (Rnm ) = 1 for any n, m because we can deterministically pick the maximum number. 6 3.2 Odds problem We show that the odds problem—a generalization of the classical secretary problem—can be formulated as the known identical distribution model. In the odds problem, a decision-maker sequentially observes a sequence of independent 0/1 random variables X1 , X2 , . . . , Xn , where Pr[Xi = 1] = pi (1) and Pr[Xi = 0] = pi (0) (= 1 − pi (1)). We say success if Xi = 1 and failure if Xi = 0. The goal is to find an optimal stopping rule to maximize the probability of win—the probability of obtaining the last success. Note that the special case pi = 1/i corresponds to the classical secretary problem and several generalizations of the odds problem have been studied [8, 37, 30, 31]. Let us consider a request-answer game (R, A, f, n) where R = {0, 1}, A = {0, 1}, and ( 1 (if r = (0, . . . , 0) or {i | ai = 1} = arg max{i | ri = 1}), f (r, a) = 0 (otherwise). Qn Then, for a distribution D∗ such that D∗ (r) = i=1 pi (ri ), the expected value Er∼D∗ [ALG(r)] is the probability of win2 . Note that OPT(r) = 1 for any r ∈ Rn and hence the competitive ratio is the inverse value of the probability of win. Bruss [7] proved that ALG∗ = (g1∗ , . . . , gn∗ ) is e-competitive when ( Pn pj (1) 1 (ri = 1 and j=i+1 pj (0) < 1), ∗ gi (r1 , . . . , ri ) = 0 (otherwise) and this is asymptotically best possible. Moreover, no online algorithm can achieve a competitive ratio better than e even if the input is a known i.i.d. distribution when n goes to infinity (see Lemma A.1 in Appendix). Namely, we can conclude that lim sup CRind = lim sup CRi.i.d. = e. n→∞ n→∞ Next, let us consider the following distribution D∗ :   . . . , 1, 0, . . . , 0), i = 1, 2, . . . , n), 1/n (r = (1, | {z } | {z } ∗ D (r) = i n−i  0 (otherwise). Then, it is easy to see that any algorithm wins with probability at most 1/n. On the other hand, selecting a variable uniformly at random is an n-competitive algorithm. Thus, we obtain CRdep = CRdet = CRind = CRdep = n. For known deterministic inputs or known random order distribution, it holds that CRdet = CRr.d. = 1 because the algorithm can distinguish whether a current success is the last one or not. For the other models CRi.i.d. , CRr.d. , CRr.i. , and CRr.i. , we can observe, by Theorem 2.4, that the values are at least e when n goes to infinity and they are at most n. However, any good bounds are not known, to the best of the author’s knowledge. 4 Relating the Competitive Ratios In this section, we give the proof of Theorem 2.4 and we show that the relations (3)–(6) cannot be merged into one sequence. 2 More precisely, we add extra probability Pr[r = (0, . . . , 0)] to simplify later discussion. 7 4.1 Proof of Theorem 2.4 Let us start with an easy part. Lemma 4.1. For any request-answer game (R, A, f, n), we have CR∆det (Rn ) = 1. Proof. As the online algorithm knows the whole request sequence in advance, it can output the optimal answer sequence. More precisely, we get CRdet = Er′ ∼δr [OPT(r ′ )] ′ δr ∈∆det (Rn ) ALG∈A Er ′ ∼δr [ALG(r )] sup inf OPT(r) ALG(r) maxa∈An f (r, a) = sup r∈Rn maxALG∈A ALG(r) = sup inf r∈Rn ALG∈A = sup r∈Rn maxa∈An f (r, a) = 1. maxa∈An f (r, a) As the competitive ratios are at least 1, the lemma implies that CRdet is the smallest one. We use the convex hull of a set of distributions to prove the rest part of the theorem. For a set of distributions D, the convex hull is the set ( ) X X conv(D) = λD · D λD = 1 and λD ≥ 0 (∀D ∈ D) . D∈D D∈D The following lemma connects the unknown distribution model and the known distribution model. Lemma 4.2. For any compact set of distributions D, we have CRD = CRconv(D) = CRconv(D) . Proof. CRconv(D) = CRconv(D) holds by Lemma 2.3 since conv(D) is a compact convex set. Since D ⊆ conv(D), we have inf sup ALG∈A D∈D Er∼D [OPT(r)] Er∼D [OPT(r)] ≤ inf . sup Er∼D [ALG(r)] ALG∈A D∈conv(D) Er∼D [ALG(r)] (7) Thus, it is sufficient to prove the reverse inequality of (7). Let us fix ALG ∈ A. Let D1 , D2 , · · · ∈ conv(D) be a sequence of distributions satisfying lim i→∞ Er∼Di [OPT(r)] Er∼D [OPT(r)] . = sup Er∼Di [ALG(r)] D∈conv(D) Er∼D [ALG(r)] By the definition for each i ∈ {1, 2, . . . },there exist coefficients λiD such that Di = P of thei convex closure, P i i D∈D λD = 1, and λD ≥ 0 (∀D ∈ D). Thus, we have D∈D λD · D, P i Er∼Di [OPT(r)] n D (r) · OPT(r) = Pr∈R i Er∼Di [ALG(r)] r∈Rn D (r) · ALG(r) P P λi · D(r) · OPT(r) n = Pr∈R PD∈D D i D∈D λD · D(r) · ALG(r) r∈Rn P λi · Er∼D [OPT(r)] = PD∈D D i D∈D λD · Er∼D [ALG(r)] ≤ sup D∈D Er∼D [OPT(r)] . Er∼D [ALG(r)] 8 Hence we obtain i [OPT(r)] E Er∼D [OPT(r)] Er∼D [OPT(r)] = lim r∼D . ≤ sup i→∞ i E [ALG(r)] E [ALG(r)] D∈D Er∼D [ALG(r)] r∼D D∈conv(D) r∼D sup By taking infimum over ALG ∈ A, we have Er∼D [OPT(r)] Er∼D [OPT(r)] ≤ inf sup , ALG∈A D∈D Er∼D [ALG(r)] ALG∈A D∈conv(D) Er∼D [ALG(r)] inf sup which proves the lemma. Next, we examine relationships between our intended sets of distributions. Lemma 4.3. For any request-answer game (R, A, f, n), we have conv(∆i.i.d. ) ⊆ conv(∆r.d. ) = conv(∆r.i. ) ⊆ conv(∆det ) = conv(∆ind ) = ∆dep = conv(∆dep ). Proof. We first prove that conv(∆det ) = conv(∆ind ) = conv(∆dep ) = ∆dep . By the definition, we have ∆det ⊆ ∆ind ⊆ ∆dep . P Also, for D ∈ ∆dep , it holds that D = r∈Rn D(r)δr and hence ∆dep ⊆ conv(∆det ). Thus, ∆dep = conv(∆dep ) = conv(∆ind ) = conv(∆det ) holds. Next, we observe that conv(∆r.d. (Rn )) ⊆ ∆dep (Rn ). As ∆dep (Rn ) = conv(∆dep (Rn )), it is sufficient to prove that ∆r.d. (Rn ) ⊆ ∆dep (Rn ). Let ψr ∈ ∆r.d. (Rn ). Recall that ψr (r ′ ) = |{σ ∈ Sn | r ′ = r σ }|/n! for r ′ ∈ Rn . Thus we obtain ψr = X |{σ ∈ Sn | r ′ = r σ }| · δr′ ∈ ∆dep (Rn ). n! ′ n r ∈R Furthermore, we show conv(∆r.d. ) = conv(∆ r.i. Q).n Since ∆r.d. ⊆ ∆r.i. , it is sufficient to prove that ∆r.i. ⊆ P conv(∆r.d. ). Let D ∈ ∆r.i. where D(r) = σ∈Sn i=1 Dσ(i) (ri )/n! and Di ∈ ∆(R). Also, let D′ ∈ ∆ind be a Qn distribution such that D′ (r) = i=1 Di (ri ). Then we have X D= D′ (r ′ )ψr′ ∈ conv(∆r.d. ). r ′ ∈Rn n n n Finally, we show r.d. (R )). Let D ∈ ∆i.i.d. (R ) and D(r) = Qn ) ⊆ conv(∆ Qn that′ ∆i.i.d.(R ′ σ we have D(r) = i=1 D (ri ) = i=1 D (rσ(i) ) = D(r ) for any σ ∈ Sn and hence D= X D(r)δr = r∈Rn X 1 X D(r σ )δrσ n! n σ∈Sn r∈R X X δr σ X X D(r)δrσ = D(r) = n! n! n n r∈R σ∈Sn = X r∈Rn = X r∈Rn D(r) σ∈Sn r∈R ′ X |{σ ∈ Sn | r = r σ }| · δr ′ n! ′ n r ∈R D(r) · ψr ∈ conv(∆r.d. (Rn )), which proves the claim. 9 ! Qn i=1 D′ (ri ). Then We remark that ∆i.i.d. (Rn ) ( conv(∆i.i.d. (Rn )) in general. For example, if R = {0, 1} and n = 2, then + 21 δ(1,1) ∈ conv(∆i.i.d. (Rn )) while 21 δ(0,0) + 12 δ(1,1) 6∈ ∆i.i.d. (Rn ). Now we are ready to prove Theorem 2.4. By Lemmas 4.2 and 4.3, we obtain 1 2 δ(0,0) CRi.i.d. ≤ CRr.d. = CRr.i. ≤ CRdep = CRdet = CRind = CRdep since D ⊆ D′ implies CRD ≤ CRD′ . Moreover, CRi.i.d. ≤ CRi.i.d. holds by Lemma 2.2 and CRi.i.d. ≥ 1 (= CRdet ) by the definition of competitive ratio. Thus, (3) holds. Also, (4) holds by ∆r.d. ⊆ ∆r.i. , (5) holds by ∆i.i.d. ⊆ ∆r.i. , and (6) holds by ∆i.i.d. ⊆ ∆ind ⊆ ∆dep . 4.2 Inequalities for incomparable ratios We show that the relations (3)–(6) cannot be merged into one sequence. To prove this, we observe some examples of request-answer games. Lemma 4.4. For each following statement, there exists a request-answer game (R, A, f, n) that satisfies it: (a) CRr.i. < CRi.i.d. , (b) CRind < CRi.i.d. , (c) CRr.d. < CRi.i.d. , (d) CRi.i.d. < CRr.d. , and (e) CRr.d. < CRind . Proof. (a) CRr.i. < CRi.i.d.. This inequality holds for an online selection problem instance since CRr.i. ≤ e/(e − 1) (≈ 1.582) and CRi.i.d. ≥ e (≈ 2.718) when n and m go to infinity (see Section 3.1). (b) CRind < CRi.i.d.. This inequality holds for an online selection problem instance since CRind ≤ 2 and CRi.i.d. ≥ e when n and m go to infinity (see Section 3.1). (c) CRr.d. < CRi.i.d. . This inequality holds for an odds problem instance since CRr.d. → 1 and CRi.i.d. → e as n goes to infinity (see Section 3.2). To prove the other cases, we consider request-answer games of the form G = (R, A, fS , n) where S ⊆ Rn , A = {0, 1}, and ( 1 ((r ∈ S and a1 = 1) or (r 6∈ S and a1 = 0)), fS (r, a) = 0 (otherwise). By the definition of the game, the task of the online algorithm is to predict r ∈ S or not. The prediction is answered as a1 . The profit is one if the prediction is correct and is zero otherwise. Here, we remark that OPT(r) = 1 for any r ∈ Rn . (d) CRi.i.d. < CRr.d. . We consider the game G with R = {1, 2, . . . , n} and S = {(σ(1), . . . , σ(n)) ∈ Rn | σ ∈ Sn , σ(n) > σ(n − 1)}. Then, for any i.i.d. distribution, the request sequence does not belong to S with high probability. Thus, by predicting that the request sequence does not belong to S, one can obtain an expected profit 1 − o(1). On the other hand, when the input distribution is ψ(1,2,...,n) , we cannot predict that the request sequence belongs to S or not with probability better than 1/2. Thus, CRi.i.d. = 1 + o(1) and CRr.d. = 2 and hence the inequality holds. (e) CRr.d. < CRind . We consider the game G with R = {0, 1} and S = {(0, 0, . . . , 0, 1)}. Then, for any random order distribution, the request sequence belong to S with probability at most 1/n. Thus, by predicting that the request sequence does not belong to S, one canQ get an expected profit 1 − 1/n. On the other hand, when the input distribution D ∈ ∆ind satisfies D(r) = ni=1 Di (ri ) where D1 (0) = · · · = Dn−1 (0) = 1 and Dn (0) = Dn (1) = 1/2, we cannot predict that the request sequence belongs to S or not with probability better than 1/2. Thus, CRr.d. = n/(n − 1) and CRind = 2 and hence the inequality holds. 10 References [1] M. Abolhassani, S. Ehsani, H. Esfandiari, M. HajiAghayi, R. Kleinberg, and B. Lucier. Beating 1 − 1/e ordered prophets. In Proceedings of STOC, pages 16–71, 2017. [2] M. Babaioff, N. Immorlica, D. Kempe, and R. Kleinberg. A knapsack secretary problem with applications. In Proceedings of APPROX and RANDOM, pages 16–28, 2007. [3] M. Babaioff, N. Immorlica, D. Kempe, and R. Kleinberg. Online auctions and generalized secretary problems. ACM SIGecom Exchange, 7(2):7:1–7:11, 2008. [4] M. Babaioff, N. Immorlica, and R. Kleinberg. Matroids, secretary problems, and online mechanisms. In Proceedings of SODA, pages 434–443, 2007. [5] S. Ben-David, A. Borodin, R. M. Karp, G. Tardos, and A. Wigderson. On the power of randomization in on-line algorithms. Algorithmica, 11(1):2–14, 1994. [6] A. Borodin and R. El-Yaniv. Online computation and competitive analysis. Cambridge university press, 2005. [7] F. T. Bruss. Sum the odds to one and stop. The Annals of Probability, 28(3):1384–1391, 2000. [8] F. T. Bruss and D. Paindaveine. Selecting a sequence of last successes in independent trials. Journal of Applied Probability, 37(02):389–399, 2000. [9] S. Bubeck and N. Cesa-Bianchi. Regret analysis of stochastic and nonstochastic multi-armed bandit problems. Foundations and Trends in Machine Learning, 5(1):1–122, 2012. [10] N. Cesa-Bianchi and G. Lugosi. Prediction, learning, and games. Cambridge university press, 2006. [11] M. Dinitz. Recent advances on the matroid secretary problem. SIGACT News, 44(2), 2013. [12] E. B. Dynkin. The optimum choice of the instant for stopping a Markov process. Soviet Mathematics Doklady, 4:627–629, 1963. [13] H. Esfandiari, M. Hajiaghayi, V. Liaghat, and M. Monemizadeh. Prophet secretary. In Proceedings of ESA, pages 496–508, 2015. [14] T. S. Ferguson. Who solved the secretary problem? Statical Science, 4(3):282–289, 1989. [15] P. R. Freeman. The secretary problem and its extensions: A review. International Statistical Review, 51(2):189–206, 1983. [16] M. Hajiaghayi, R. Kleinberg, and D. C. Parkes. Adaptive limited-supply online auctions. In Proceedings of EC, pages 71–80. ACM, 2004. [17] M. Hajiaghayi, R. Kleinberg, and T. Sandholm. Automated online mechanism design and prophet inequalities. In Proceedings of AAAI, volume 7, pages 58–65, 2007. [18] E. Hazan. Introduction to Online Convex Optimization. Foundations and Trends in Optimization Series. 2016. [19] T. P. Hill and R. P. Kertz. Comparisons of stop rule and supremum expectations of i.i.d. random variables. The Annals of Probability, 10(2):336–345, 1982. [20] T. P. Hill and R. P. Kertz. Stop rule inequalities for uniformly bounded sequences of random variables. Transactions of the American Mathematical Society, 278(1):197–207, 1983. [21] T. P. Hill and U. Krengel. A prophet inequality related to the secretary problem. Strategies for Sequential Search and Selection in Real Time, pages 209–215, 1992. 11 [22] T. Kesselheim, K. Radke, A. Tönnis, and B. Vöcking. An optimal online algorithm for weighted bipartite matching and extensions to combinatorial auctions. In Proceedings of ESA, pages 589–600, 2013. [23] R. Kleinberg. A multiple-choice secretary algorithm with applications to online auctions. In Proceedings of SODA, pages 630–631, 2005. [24] R. Kleinberg and S. M. Weinberg. Matroid prophet inequalities. In Proceedings of STOC, pages 123–136, 2012. [25] D. Komm. An Introduction to Online Computation: Determinism, Randomization, Advice. Springer International Publishing, 2016. [26] N. Korula and M. Pál. Algorithms for secretary problems on graphs and hypergraphs. In Proceedings of ICALP, pages 508–520, 2009. [27] U. Krengel and L. Sucheston. Semiamarts and finite values. Bulletin of the American Mathematical Society, 83(4):745–747, 1977. [28] O. Lachish. O(log log rank) completitive-ratio for the matroid secretary problem. In Proceedings of FOCS, pages 326–335, 2014. [29] L. H. Loomis. On a theorem of von neumann. In Proceedings of the National Academy of Sciences of the U.S.A., volume 32, pages 213–215, 1946. [30] T. Matsui and K. Ano. A note on a lower bound for the multiplicative odds theorem of optimal stopping. Journal of Applied Probability, 51(03):885–889, 2014. [31] T. Matsui and A. Katsunori. Lower bounds for bruss’ odds problem with multiple stoppings. Mathematics of Operations Research, 41(2):700–714, 2016. [32] A. Mehta. Online matching and ad allocation. Foundations and Trends in Theoretical Computer Science, 8(4):265–368, 2013. [33] A. Rubinstein and S. Singla. Combinatorial prophet inequalities. In Proceedings of SODA, pages 1671– 1687, 2017. [34] S. M. Samuels. Secretary problems. In Handbook of Sequential Analysis. Marcel Dekker, 1991. [35] D. D. Sleator and R. E. Tarjan. Amortized efficiency of list update and paging rules. Communications of the Association for Computing Machinery, 28(2):202–208, 1985. [36] T. J. Stewart. Optimal selection from a random sequence with learning of the underlying distribution. Journal of the American Statistical Association, 73(364):775–780, 1978. [37] M. Tamaki. Sum the multiplicative odds to one and stop. Journal of Applied Probability, 47(03):761–777, 2010. [38] J. von Neumann. Zur Theorie der Gesellschaftsspiele. Mathematische Annalen, 100:295–320, 1928. [39] A. Yao. Probabilistic computations: Toward a unified measure of complexity. In Proceedings of FOCS, pages 222–227. IEEE, 1977. 12 A Omitted Proof Lemma A.1. No online algorithm can achieve a competitive ratio better than e even if the input is a known i.i.d. distribution. √ √ Proof. Let us assume that Pr[Xi = 1] = 1/ n and Pr[Xi = 0] = 1 − 1/ n for each i = 1, . . . , n. Consider the following linear programming (LP): max s.t. Pn  1− √1 n n−i qi Pi−1 √ n · qi ≤ 1 − j=1 qj qi ≥ 0 i=1 (i ∈ [n]), (i ∈ [n]) where [n] = {1, 2, . . . , n}. We claim that the optimal value for the LP gives an upper bound of the probability of win for the problem. Let qi be the probability of selecting the ith random variable as success. Then the probability of win is n−i n  X 1 1− √ qi n i=1  because if Xi+1 = · · · = Xn = 0 with probability 1 − √1 n n−i . Also, pij must satisfy the following relation:  pi ≤ Pr[1, . . . , i − 1 are not selected] · Pr[Xi = 1] = 1 − Thus, the LP present an upper bound of the success probability. To evaluate the optimal value, now we consider the dual problem: Pn min i=1 ri  n−i Pn √ s.t. n · ri + j=i+1 rj ≥ 1 − √1n ri ≥ 0 i−1 X  1 qj  · √ . n j=1 (i ∈ [n]), (i ∈ [n]). √ n−i−1 √ n−i n−i − n · (1 − 1/ n) , 0}. Then it is easy to check that ri∗ is a feasible Let ri∗ = max{ √1n · (1 − 1/ n) solution for the dual LP. Thus, the optimal value of the LP (and the dual LP) is upper bounded by n X ri∗ = n X max i=1 i=1 = n X   √ n−i n − i √ n−i−1 1 √ · (1 − 1/ n) − ,0 · (1 − 1/ n) n n   √ n−i n − i √ n−i−1 1 √ · (1 − 1/ n) · (1 − 1/ n) − n n √ i=n−⌊ n⌋+1 √ ⌊√n⌋ √ ⌊√n⌋ √ ⌊√n⌋ 1 1 − (1 − 1/ n) 1 1 − (1 − 1/ n) − (1 − 1/ n) √ =√ · − · n 1 − (1 − 1/ n) n 1/n √ √ ⌊ n⌋ 1 (n → ∞). → = (1 − 1/ n) e Hence, 1/e is an upper bound of the probability of win for the problem. Moreover, since Pr[X1 = · · · = Xn = 0] → 0 as n → ∞, there exists no online algorithm that can achieve a competitive ratio better than e. 13 

8cs.DS

arXiv:1704.00990v2 [math.CO] 7 Apr 2017 TESTING ISOMORPHISM OF CENTRAL CAYLEY GRAPHS OVER ALMOST SIMPLE GROUPS IN POLYNOMIAL TIME ILIA PONOMARENKO AND ANDREY VASIL’EV In memory of Sergei Evdokimov Abstract. A Cayley graph over a group G is said to be central if its connection set is a normal subset of G. It is proved that for any two central Cayley graphs over explicitly given almost simple groups of order n, the set of all isomorphisms from the first graph onto the second can be found in time poly(n). 1. Introduction In the present paper, we are interested in a special case of the following restriction of the Graph Isomorphism Problem to the class of Cayley graphs. Cayley Graph Isomorphism Problem. For two explicitly given finite groups G and G′ and two sets X ⊂ G and X ′ ⊂ G′ , construct the set Iso(Γ, Γ′ ), where Γ = Cay(G, X) and Γ′ = Cay(G, X ′ ). Here the input consists of the multiplication tables of G and G′ and the sets X and X ′ , whereas the output is either empty or given by an element of the set Iso(Γ, Γ′ ) and a generating set of the group Aut(Γ) (of size polynomial in the order n of the group G). Obviously, the Luks algorithm [12] solves the Cayley Graph Isomorphism Problem in polynomial time for every group G, whenever the set X is of constant sizes. If G is cyclic, then the problem with no restriction for X is also solvable in polynomial time (see [6] and [13]). It should be noted that if G = G′ and G is a so-called CI-group, then an obvious algorithm solves the Cayley Graph Isomorphism Problem in time polynomial in | Aut(G)| (for details, see [10]). The aforementioned special case is formed by the two following conditions imposed on the input graphs and groups. First, we assume that Γ is a central Cayley graph over G, which means that X is a normal subset of G, i.e., X g = X for every g ∈ G. Second, the group G is assumed to be almost simple, i.e., the socle of G is a nonabelian simple group. The same conditions are imposed on the graph Γ′ and group G′ . Even in this rather restrictive case, the problem is still nontrivial; at least the number of possible input graphs is exponential in n. Example. Let G = Sym(m) be a symmetric group of degree m. Then the number N (m) of central colored Cayley graphs over G is equal to 2p(m) , where p(m) is the number of all partitions of the set {1, . . . m}. Since p(m) is approximately √ equal to 2 m , the number N (m) is exponential in n = m!. By technical reasons, it is more convenient to deal with colored Cayley graphs. Such a graph is given by a partition P of the group G into k ≥ 2 classes X0 , . . . , Xk−1 with X0 = {1}, and can be thought as arc-colored complete graph with vertex set 1 2 ILIA PONOMARENKO AND ANDREY VASIL’EV G and the ith color class of arcs coinciding with the arc set of the Cayley graph Cay(G, Xi ), i = 0, . . . , k − 1. We say that P is the Cayley partition of this graph and denote the latter by Cay(G, P). In what follows, all Cayley graphs are assumed to be colored: the graph Cay(G, X) is treated as Cay(G, P) for k = 3 and X1 = X. Theorem 1.1. For any two central Cayley graphs Γ and Γ′ over explicitly given almost simple groups G and G′ of order n, the set Iso(Γ, Γ′ ) can be found in time poly(n). Corollary 1.2. The automorphism group of a central Cayley graph over an explicitly given almost simple group of order n can be found in time poly(n). The proof of Theorem 1.1 is a mix of combinatorial and permutation group techniques. Section 2 provides a relevant background for the combinatorial part including coherent configurations and Cayley schemes. In Section 3, we use a classification of regular almost simple subgroups of primitive groups [11] to prove (Lemma 3.2) that except for one special case, if K ≤ Sym(G) is a 2-closed primitive group containing regular almost simple subgroup, then (1) K = Sym(G) or K ≤ D(2, G), where D(2, G) is the subgroup of Sym(G) generated by the holomorph of G and the permutation σ : g 7→ g −1 , g ∈ G. We extend this result to non-primitive groups in Sections 4 and 5 by showing that in this case, either formula (1) holds or K is a nontrivial generalized wreath product (Theorems 4.1 and 5.1). We apply this fact in Section 6 to the automorphism group K of a central Cayley graph Γ over the group G to establish that (2) K S = Sym(S) or K S ≤ D(2, S), where S = Soc(G) is the socle of G and K S is the restriction to S of the setwise stabilizer of S in K. Note that if G is a symmetric group of degree at least 5, then the group D(2, G) is isomorphic to the group G ≀ Sym(2). Thus as a byproduct of (2), we obtain the following generalization of the result [9, Theorem 1.1] on the automorphism group of the Cayley graph Cay(G, X), where G is a symmetric group and X is the set of its transpositions. Theorem 1.3. Let G be a symmetric group of degree at least 5, X a proper normal subset of G \ Soc(G), and Γ = Cay(G, X). Then Aut(Γ) = D(2, G). In Sections 7 and 8, we develop algorithmic tools to find the above structure of the group K with the help of the related Cayley scheme and the group G. The main algorithm providing the proof of Theorem 1.1 is given in Section 9. Notation. The diagonal of the Cartesian product Ω × Ω is denoted by 1Ω . For s ⊆ Ω × Ω, set s∗ = {(β, α) : (α, β) ∈ s}. For α ∈ Ω and s ⊆ Ω × Ω, set αs = {β ∈ Ω : (α, β) ∈ s}. For ∆ ⊆ Ω and s ⊆ Ω × Ω, set s∆ = s ∩ (∆ × ∆). For a partition E of a set Ω and s ⊆ Ω × Ω, set sE to be the relation consisting of the pairs (∆, ∆′ ) ∈ E × E such that s meets ∆ × ∆′ . For a set S of binary relations, put S ∪ to be the set of all unions of relations from S. The symmetric and alternating groups on Ω are denoted by Sym(Ω) and Alt(Ω), respectively. TESTING ISOMORPHISM OF CENTRAL CAYLEY GRAPHS 3 For f ∈ Sym(Ω) and s ⊆ Ω × Ω, set sf = {(αf , β f ) : (α, β) ∈ s}. For a group G and its subgroup L, set Llef t and Lright to be the subgroups of Sym(G) induced by left and right multiplications of L, respectively, and L∗ = Llef t Lright . For a group G, set D(2, G) to be the subgroup of Sym(G) generated by the group Hol(G) = NSym(G) (Gright ) and the permutation σ : g 7→ g −1 . For a group K ≤ Sym(G) and a set H ⊆ G, the restriction to H of the setwise stabilizer of H in K is denoted by K H . For an imprimitivity system L of a transitive group K, set KL and K L to be, respectively, the intersection of all K H with H ∈ L and the permutation group induced by the action of K on L. For a group G and a permutation group K, set Reg(K, G) to be the set of all regular subgroups of K that are isomorphic to G. 2. Coherent configurations and Cayley schemes This section contains well-known basic facts on coherent configurations. All of them can be found in [7] and papers cited there. 2.1. Main definitions. Let Ω be a finite set and S a partition of Ω × Ω. The pair X = (Ω, S) is called a coherent configuration on Ω if the following conditions hold: (C1) 1Ω ∈ S ∪ , (C2) S ∗ = S, (C3) given r, s, t ∈ S, the number ctrs = |αr ∩ βs∗ | does not depend on the choice of the pair (α, β) ∈ t. The elements of Ω and S, and the numbers ctrs are called the points and basis relations, and the intersection numbers of X , respectively. The numbers |Ω| and |S| are called the degree and the rank of X . The coherent configuration X is said to be homogeneous if 1Ω ∈ S. Denote by Φ = Φ(X ) the set of Λ ⊆ Ω such that 1Λ ∈ S. The elements of Φ are called the fibers of X . In view of condition (C1), the set Ω is the disjoint union of all of them. Moreover, for each s ∈ S, there exist uniquely determined fibers Λ and ∆ such that s ⊆ Λ × ∆. Note that the coherent configuration X is homogeneous if and only if |Φ| = 1. Let e ∈ S ∪ be an equivalence relation and E the set of its classes. Given ∆ ∈ E denote by S∆ the set of all nonempty relations s∆ with s ∈ S. Then the pair X∆ = (∆, S∆ ) is a coherent configuration called the restriction of X to ∆. This enables to define the restriction of X to a set Λ ∈ Φ∪ : the corresponding equivalence relation is equal to the union of ∆ × ∆, where ∆ runs over the fibers contained in Λ. Another coherent configuration associated with e is obtained as follows. Denote by SE the set of all nonempty relations sE , s ∈ S. Then XE = (E, SE ) is a coherent configuration called the quotient of X modulo e. 4 ILIA PONOMARENKO AND ANDREY VASIL’EV 2.2. Combinatorial and algebraic isomorphisms. A bijection f : Ω → Ω′ is called the (combinatorial) isomorphism from X onto a coherent configuration X ′ = (Ω′ , S ′ ) if the set S ′ contains the relation sf for each s ∈ S. The set of all isomorphisms f is denoted by Iso(X , X ′ ). The group of all isomorphisms of X to itself contains a normal subgroup Aut(X ) = {f ∈ Sym(Ω) : sf = s, s ∈ S} called the automorphism group of X . Conversely, let G ≤ Sym(Ω) be a permutation group, and let S be the set of orbits of the component-wise action of G on Ω × Ω. Then the pair X = (Ω, S) is a coherent configuration; we say that X is associated with G and denote it by Inv(G). According to Wielandt [17], a permutation group G on Ω is said to be 2-closed if it is equal to its 2-closure G(2) = Aut(Inv(G)), or, equivalently, if G is an automorphism group of a family of binary relations on Ω (such a family can always be chosen as the set of basis relations of a coherent configuration on Ω). If G is 2-closed and L is an imprimitivity system of G, then the group GL is 2-closed. However, the group GL is not always 2-closed. A bijection ϕ : S → S ′ , r 7→ r′ is called an algebraic isomorphism from X onto X ′ if ′ ctrs = ctr′ s′ , (3) ′ r, s, t ∈ S. In this case, X and X are said to be algebraically isomorphic. Each isomorphism f from X onto X ′ induces an algebraic isomorphism ϕf : r 7→ rf between these configurations. The set of all isomorphisms inducing the algebraic isomorphism ϕ is denoted by Iso(X , X ′ , ϕ). In particular, (4) Iso(X , X , idS ) = Aut(X ), where idS is the identity mapping on S. An algebraic isomorphism ϕ induces a bijection from S ∪ onto (S ′ )∪ : the union r∪s∪· · · of basis relations of X is taken to r′ ∪s′ ∪· · · . This bijection is also denoted by ϕ. It preserves the equivalence relations e ∈ S ∪ ; moreover, the equivalence relations e and ϕ(e) have the same number of classes as well as the same multiset of their sizes. In this case, if E and E′ are the sets of classes of e and ϕ(e), respectively, and ∆ ∈ E, then ϕ induces the algebraic isomorphisms ′ ϕ∆ : X∆ → X∆ ′ , s 7→ ϕ(s)∆′ for a suitable ∆′ ∈ E′ . and ϕE : XE → XE′ ′ , sE 7→ ϕ(s)E′ 2.3. Direct sum and wreath product. Let X = (Ω, S) and X ′ = (Ω′ , S ′ ) be coherent configurations. Denote by Ω ⊔ Ω′ the disjoint union of Ω and Ω′ , and by S ⊞ S ′ the union of the set S ⊔ S ′ and the set of all relations ∆ × ∆′ and ∆′ × ∆ with ∆ ∈ Φ(X ) and ∆′ ∈ Φ(X ′ ). Then the pair X ⊞ X ′ = (Ω ⊔ Ω′ , S ⊞ S ′ ) is a coherent configuration called the direct sum of X and X ′ . The automorphism group of this configuration equals the direct product Aut(X ) × Aut(X ′ ) acting on the set Ω⊔Ω′ . Furthermore, if ϕ is an algebraic isomorphism from X ⊞X ′ to another coherent configuration, then the latter is also the direct sum Y ⊞ Y ′ and ϕ induces TESTING ISOMORPHISM OF CENTRAL CAYLEY GRAPHS 5 algebraic isomorphisms X → Y and X ′ → Y ′ coinciding with the restrictions of ϕ on Ω and Ω′ , respectively. Let X be a homogeneous coherent configuration, e ∈ S ∪ an equivalence relation, and E the set of classes of e. We say that X is the wreath product with respect to e if for each s ∈ S such that s 6⊆ e, [ Λ × ∆. s= (Λ,∆)∈sE In what follows, we always assume that the classes of e can be identified with the help of a family of the isomorphisms fΛ,∆ : XΛ → X∆ , Λ, ∆ ∈ E, such that (5) (sΛ )fΛ,∆ = s∆ , s ∈ S. In this case, X is isomorphic to the usual wreath product X∆ ≀ XE for all ∆ ∈ E (see [15, p.45]). The automorphism group of this coherent configuration is permutation isomorphic to the wreath product Aut(X∆ ) ≀ Aut(XE ) in imprimitive action. Furthermore, if ϕ is an algebraic isomorphism from the wreath product X with respect to e to another coherent configuration, then the latter is also the wreath product X ′ with respect to e′ = ϕ(e) and ϕ induces algebraic isomorphisms X∆ → ′ ′ X∆ ′ and XE → XE′ coinciding with the restrictions of ϕ on ∆ and E, respectively, ′ where E is the set of classes of e′ and ∆′ ∈ E′ . 2.4. Cayley schemes. A coherent configuration X = (Ω, S) is called the Cayley scheme over a group G if Ω = G and Gright ≤ Aut(X ). In this case, X is homogeneous and each basis relation s is the set of arcs of Cayley graph Cay(G, X), where X is the neighborhood of the identity of G in relation s. In particular, X can be treared as a color graph Cay(G, P), where classes of the Cayley partition P are the neighborhoods of the identity of G in basis relations of X . the the the the The class of Cayley schemes is closed with respect to taking restrictions and quotients. Namely, if X is a Cayley scheme over a group G and e ∈ S ∪ is an equivalence relation, then the class H of e containing the identity of G is a subgroup of G. Moreover, the set E of classes of e coincides with the right H-cosets of G. It follows that with XH = Cay(H, EH ) and XE = Cay(G/H, EG/H ), EH = {X ∈ E : X ⊂ H} and EG/H = {π(X) : X ∈ E}, where in the latter case, H is a normal subgroup of G and π : G → G/H is the canonical epimorphism. Assume that the Cayley scheme X is the wreath product with respect to an equivalence relation e ∈ S ∪ . Then for any two classes Λ, ∆ ∈ E, there exists a permutation f ∈ Gright taking Λ to ∆; set fΛ,∆ to be the restriction of f to Λ. Since all these f are automorphisms of X , the family {fΛ,∆ }Λ,∆∈E satisfies conditions (5). The Cayley scheme X is said to be central if Glef t ≤ Aut(X ) which is by definition of a Cayley scheme is equivalent to G∗ ≤ Aut(X ). One can see that X is central if and only if the colored Cayley graph associated with X is central. 6 ILIA PONOMARENKO AND ANDREY VASIL’EV 2.5. Partial order and the WL-algorithm. There is a natural partial order ≤ on the set of all coherent configurations on the same set Ω. Namely, given two coherent configurations X = (Ω, S) and X ′ = (Ω, S ′ ), we set X ≤ X ′ ⇔ S ∪ ⊆ (S ′ )∪ . The minimal and maximal elements with respect to this order are, respectively, the trivial and discrete coherent configurations. The first one is a unique coherent configuration TΩ with at most two basis relations: 1Ω and its complement to Ω×Ω (if Ω consists of at least two points). Every basis relation of the discrete configuration is a singleton. With respect to this order, the direct sum X ⊞ X ′ is the minimal coherent configuration on Ω ⊔ Ω′ , the restrictions of which to Ω and Ω′ are equal to X and X ′ , respectively. One can prove that given a set T ⊆ 2Ω×Ω , there exists a unique minimal coherent configuration X ′ such that every relation of T is the union of some basis relations of X ′ . This coherent configuration is called the coherent closure of T and can be constructed by the well-known Weisfeiler-Leman algorithm (WL-algorithm) [15, Section B] in time polynomial in sizes of T and Ω. To stress this fact, the coherent closure of T is denoted by WL(T ). For a color graph Γ with the set S of color classes, we set WL(Γ, T ) = WL(S ∪ T ) and write WL(Γ) instead of WL(Γ, ∅). It is important to note that the automorphism group of the coherent configuration WL(Γ, T ) is equal to the subgroup of Aut(Γ) leaving each relation of T fixed (as a set). This implies that if Γ is a Cayley graph over G, then the coherent configuration WL(Γ) is a Cayley scheme over G. Since any coherent configuration X can be considered as a color graph, we extend our notation to write WL(X , T ). Concerning the following statement, we refer to [14, Theorem 2.4]. Theorem 2.1. Let S and S ′ be m-sets of binary relations on an n-element set. Then given a bijection ψ : S → S ′ , one can check in time mnO(1) whether or not there exists an algebraic isomorphism ϕ : WL(S) → WL(S ′ ) such that ϕ|S = ψ. Moreover, if ϕ does exist, then it can be found within the same time. 3. Almost simple groups In this section, we collect several known facts on finite almost simple groups and deduce some auxiliary results to be used throughout the paper. Lemma 3.1. Let G be an almost simple group of order n. Then (i) |G/L| ≤ log n, where L = Soc(G), (ii) | Reg(K, G)| = nO(1) for every K ≤ Sym(G) containing G as a normal regular subgroup. Proof. From the description of the automorphism groups of simple groups (see, e.g., [4, Introduction]), it follows that | Aut(L)/L| ≤ log |L|. Therefore, statement (i) is a consequence of the inclusions L ≤ G ≤ Aut(G) ≤ Aut(L). The inclusions also imply that | Aut(G)| ≤ n log n. Since the centralizer of G in Sym(G) is of order n [16, Exercise 4.5’], we have |K| ≤ |NSym(G) (G)| ≤ |CSym(G) (G)| · | Aut(G)| ≤ n2 log n. TESTING ISOMORPHISM OF CENTRAL CAYLEY GRAPHS 7 On the other hand, the group G is 3-generated
[5]. Thus, the number of regular subgroups of K isomorphic to G is at most K  3 and statement (ii) follows. In the following statement, we use the classification of regular almost simple subgroups of a primitive group [11, Theorem 1.4]. Lemma 3.2. Let G be an almost simple group and G∗ ≤ K ≤ Sym(G). Suppose that K is primitive. Then one of the following holds: (i) K ≥ Alt(G), (ii) G = Soc(G) and K ≤ D(2, G), (iii) G = Sym(5) and K = Sp8 (2) is 2-transitive. Proof. Without loss of generality, we assume that neither K ≥ Alt(G), nor G = Soc(G) and K ≤ D(2, G). Then by aforementioned classification, exactly one of the following pairs (G, Soc(K)) occurs: (a) (Alt(p2 − 2), Alt(p2 + 1)), where p = 3 (mod 4) is prime, (b) (Sym(p − 2), Alt(p)) or (Sym(p − 2), Alt(p + 1)), where p ≥ 7 is prime, (c) the twelve pairs in the table below. N o. G Soc(K) 1 Alt(5) L2 (59) 2 Alt(7) Alt(11) 3 Alt(7) Alt(12) 4 Sym(5) Alt(9) 5 Sym(5) Sp4 (4) 6 Sym(5) Sp6 (2) N o. G Soc(K) 7 Sym(5) Ω+ 8 (2) 8 Sym(5) Sp8 (2) 9 L2 (16).4 Sp6 (4) 10 L2 (16).4 Ω+ 8 (4) 11 L3 (4).2 M23 12 L3 (4).2 M24 The assumption G∗ ≤ K, in particular, implies that |G|2 ≤ |K|. By straightforward check this excludes cases (a) with p > 3, (b) with p > 7, and cases 11 and 12 from the table. Similarly, the remaining cases in (a) and (b) as well as cases 1–4, and 9 are impossible because |G|2 must divide |K|. In cases 5, 6, 7, and 10 from the table, we check the maximal subgroups of K and show that none of them includes the subgroup isomorphic to G∗ . Indeed, in cases 5 and 6 none of the maximal subgroups contains Alt(5) × Alt(5) [2, Tables 8.14, 8.28, 8.29]. In case 10, information from [11, Table 2] shows that K is an extension of Ω+ 8 (4) by a field automorphism. This group contains the only (up-to conjugation) maximal subgroup with section isomorphic to L2 (16) × L2 (16) [2, Table 8.50], but the order of this subgroup is less than |G∗ |. In case 7, we make use of [11, Table 2] to see that K = Ω+ 8 (2). Again this group includes up to conjugation the only maximal subgroup M with section isomorphic to Alt(5)×Alt(5) (see [2, Table 8.50]). However, |M | = |G∗ | but M ≃ D(2, Alt(5)) is not isomorphic to G∗ ≃ Sym(5) × Sym(5). This leaves us with case 8 of the table where Soc(K) = Sp8 (2) = K and we arrive at case (iii) of the conclusion of the lemma.  8 ILIA PONOMARENKO AND ANDREY VASIL’EV 4. The structure of automorphism groups: the principal section 4.1. Preliminaries. Let G be a finite group. The automorphism group of every central Cayley graph over G contains a subgroup G∗ (see Notation). In this section, we establish some basic facts on the permutation groups K satisfying the following condition: (6) G∗ ≤ K ≤ Sym(G), where G is an almost simple group. We use a concept of the generalized wreath product of permutation groups introduced and studied in [8]. Namely, a transitive group K is the generalized wreath product if it has two imprimitivity systems L and U such that every block of L is contained in a block of U and Y KL = (KL )X . X∈U The generalized wreath product is said to be trivial if either L consists of singletons or U = {Ω}. When L = U, the group K is permutation isomorphic to the wreath product K X ≀ K L in imprimitive action, where X ∈ U. Theorem 4.1. Let G be an almost simple group, and let K be a 2-closed group satisfying condition (6). Then one of the following statements holds: (i) K = Sym(G) or G∗ E K, (ii) K is a nontrivial generalized wreath product. The proof of Theorem 4.1 is given in the end of Section 6. 4.2. The minimal block. Let K satisfy condition (6) and X a K-block containing the identity of G. Since K is a permutation group on G that includes Gright , the block X is a subgroup of G [16, Theorem 24.12]. Taking into account that Glef t also lies in K, we conclude that X is normal. Denote by L the intersection of all non-singleton K-blocks containing the identity of the group G. Then L is a K-block and we call it the minimal block of K. Lemma 4.2. Let K satisfy condition (6) and L the minimal block of K. Then L is normal subgroup of G including Soc(G). In particular, L is an almost simple group such that Soc(L) = Soc(G). Proof. According to the above remark, every K-block containing the identity of G is a normal subgroup of G. If the block is not a singleton, then this normal subgroup is nontrivial and hence contains Soc(G), because the group G is almost simple. Thus, the minimal block L being the intersection of nontrivial normal subgroups of G is a normal subgroup and contains Soc(G).  Denote by L the imprimitivity system containing L. Obviously, Orb(L∗ , G) = L. Recall that according to the definition, KL is a normal subgroup of K leaving each block of L fixed (as a set), and given X ∈ L, the group (L∗ )X ≤ Sym(X) is induced by right multiplications of L. Lemma 4.3. For any X ∈ L, the group (KL )X is primitive and contains (L∗ )X . TESTING ISOMORPHISM OF CENTRAL CAYLEY GRAPHS 9 Proof. The normality of L in G implies that the orbits of the action of L∗ on G coincide with the L-cosets. It follows that the permutation group induced by this action is contained in KL . This proves the second statement. To prove the first statement, in view of the transitivity of K, we may assume that X = L. Assume on the contrary that the group K ′ = (KL )L is not primitive. Then there exists a minimal non-singleton K ′ -block L′ < L. Taking into account that K ′ E K L , we conclude that L′′ = (L′ )k is a also K ′ -block for every k ∈ K [16, Proposition 6.2]. The imprimitivity system L′ of the group K ′ that contains L′ coincides with the imprimitivity system containing L′′ , for otherwise by the minimality of L′ and Lemma 4.2 applied for the group K ′ , one can choose the block L′′ so that 1 = L′ ∩ L′′ ⊇ Soc(L) = Soc(G) ) 1, a contradiction. Thus, L′ is an imprimitivity system of the group K L . Consequently, L′ is a non-singleton K-block strictly contained in L, which is impossible by the definition of L.  4.3. The wreath decomposition of KL . For every two sets X, X ′ ∈ L, we write X ∼ X ′ if the restriction epimorphisms (7) ′ (KL )X∪X → (KL )X ′ and (KL )X∪X → (KL )X ′ ′ are isomorphisms. In particular, the groups (KL )X and (KL )X are isomorphic. It is easily seen that ∼ is an equivalence relation on L. This relation is K-invariant, because KL is a normal subgroup of K. Denote by U the union of L-cosets belonging to the class of ∼ that contains L. Then U is obviously a K-block and hence is a normal subgroup of G. Thus, L E U E G. The imprimitivity system of the group K that contains the block U is denoted by U. We say that U/L is the principal K-section of G, and U and L are the associated partitions. Theorem 4.4. Let G be an almost simple group, K a 2-closed group satisfying condition (6), and U, L are the partitions associated with the principal K-section. Then the coherent configuration Inv(KL ) is the direct sum of the coherent configurations Inv(KL )Y , where Y ∈ U. In particular, Y (8) KL = (KL )Y , Y ∈U i.e., K is generalized wreath product. Proof. The subgroup KL of 2-closed group K is 2-closed too (see Subsection 2.2). Therefore, equality (8) follows from the first statement of the theorem, because the automorphism group of the direct sum equals the direct product of the summands (see Subsection 2.3). To prove the first statement, it suffices to verify that given X, X ′ ∈ Orb(KL , G), (9) X 6∼ X ′ ⇒ X × X ′ ∈ Orb(KL , X × X ′ ). ′ To this end, we note that the group (KL )X∪X is the subdirect product of the ′ transitive constituents M = (KL )X and M ′ = (KL )X . Therefore, there exist 10 ILIA PONOMARENKO AND ANDREY VASIL’EV uniquely determined normal subgroups H and H ′ of M and M ′ , respectively, and a group isomorphism ϕ : M/H → M ′ /H ′ such that (10) ′ (KL )X∪X = {(k, k ′ ) ∈ M × M ′ : ϕ(k) = k ′ }. Now if X 6∼ X ′ , then at least one of the epimorphisms (7) is not an isomorphism. Therefore, one of the groups H and H ′ , say H, is nontrivial. It follows that H being a normal subgroup of the primitive group M (Lemma 4.3) acts transitively ′ on X. By (10), this implies that (KL )X∪X contains the subgroup H × 1. Thus, (x, x′ )H×1 = X × {x′ } ′ for all x′ ∈ X ′ and hence the group (KL )X∪X is transitive on the set X × X ′ .  5. The normalizer of Soc(G)∗ in Sym(G) The goal of this section is to prove the following theorem that shows (as we will see) that the case U = G is very similar to the case where the group K is primitive. Theorem 5.1. Let G be an almost simple group, S = Soc(G), and N = NSym(G) (S ∗ ). Then (11) N = D(2, G). Clearly, N is a proper subgroup of Sym(G) and condition (6) is satisfied for K = N . In particular, the minimal block of N coincides with S. Lemma 5.2. In the above notation, CSym(G) (S ∗ ) = 1. Proof. Set C = CSym(G) (S ∗ ). Then obviously, C = CN (S ∗ ). Since N is transitive and C is normal in N , the orbits of C form an imprimitivity system of N . Denote by X the block of this system that contains the identity of G. We may assume that the block X is not a singleton, for otherwise C = 1 and we are done. Then S ⊆ X, because S is the minimal N -block. Since C is transitive on X and S is a block of C X , there exists c ∈ C such that S c = S and cS 6= 1. However, the latter is impossible, because cS centralizes the subgroups Slef t and Sright .  Proof of Theorem 5.1. By Lemma 5.2, there exists a monomorphism from N = NSym(G) (S ∗ ) to Aut(S ∗ ). Since the latter is isomorphic to the wreath product W = Aut(S) ≀ Sym(2), this monomorphism induces a monomorphism ϕ : N → W. Clearly, ϕ can be chosen so that the subgroups Slef t and Sright of the group N go to the subgroup Inn(S) × 1 of the group A = Aut(S) × 1 and to the subgroup 1 × Inn(S) of the group B = 1 × Aut(S), respectively. Set W1 = im(ϕ) and W0 = im(ϕ) ∩ (A × B). Then since the index of A × B in W equals 2 and W1 is not contained in A × B, we conclude that (12) |W1 : W0 | = 2. Note that the centralizer of the group ϕ(Sright ) ≤ B in the group W0 is contained in A, because the group B is almost simple with the socle ϕ(Sright ). Since the group Glef t ≤ N centralizes Sright , this proves the first of the two following inclusions (the second one can be proved in a similar way): ϕ(Glef t ) ≤ A and ϕ(Gright ) ≤ B. TESTING ISOMORPHISM OF CENTRAL CAYLEY GRAPHS 11 The first inclusion implies that ϕ(Glef t ) ≤ W0 ∩ A. The reverse inclusion follows from the fact that the centralizer of Glef t in Sym(G) is equal to Gright [16, Proposition 4.3]. Thus, we obtain the equalities: (13) ϕ(Glef t ) = W0 ∩ A and ϕ(Gright ) = W0 ∩ B. This immediately implies that ϕ(Glef t ) and ϕ(Gright ) are normal in W0 . This group has trivial center and hence can be identified with a subgroup of the direct product of the groups A′ = Aut(ϕ(Glef t )) and B ′ = Aut(ϕ(Gright )) (isomorphic to Aut(G)). It follows that ′ ′ ′ W 0 = W0 /ϕ(G∗ ) ≤ A × B , ′ where A = A′ /ϕ(Glef t ) and B = B ′ /ϕ(Gright ). Moreover in view of formulas (13), ′ ′ the group W 0 intersects each of the groups A and B trivially. Therefore, (14) |W 0 | = | Aut(G)/G|. Now, the inclusion ϕ(Aut(G)) ≤ W0 and formula (14) show that W0 = ϕ(Hol(G)). Formula (12) yields that |N : Hol(G)| = 2. Since the permutation σ : g 7→ g −1 lies in N , formula (11) holds.  6. Symmetric and normal types of the automorphism group Let G be an almost simple group, K be a 2-closed group satisfying condition (6), and U/L the principal K-section of G. Then the group L is almost simple (Lemma 4.2) and K L is primitive (Lemma 4.3). We say that K is of symmetric type if either K L ≥ Alt(L), or G = Sym(5) and K = Sp8 (2) is 2-transitive; if L = Soc(L) and K L ≤ D(2, L), the group K is said to be of normal type. The following statement is a straightforward consequence of Lemma 3.2 and the above definitions. Proposition 6.1. Let G be an almost simple group, and let K ≤ Sym(G) be a 2-closed group containing G∗ . Then K is of symmetric or normal type. Let us study a group of symmetric and normal types in detail. As the following statement shows, any group of symmetric type is, in fact, the wreath product in imprimitive action. In what follows, U and L are the partitions associated with the principal K-section of G. Theorem 6.2. Let K be a group of symmetric type. Then L=U and (KL )L = Sym(L). In particular, K is permutation isomorphic to the wreath product Sym(L) ≀ K L in imprimitive action. Proof. Let X, X ′ ∈ L and X ∼ X ′ . We claim that there exists a bijection f : X → X ′ , for which (15) sf ∈ Orb(KL , X × X ′ ), where sf = {(α, αf ) : α ∈ X} is the graph of f . Indeed, consider the group ′ M = (KL )X∪X . Since X ∼ X ′ , the group M acts faithfully on X and X ′ . As the group K is of symmetric type, each of these actions is 2-transitive. However, M has a unique faithful 2-transitive representation of degree d = |X|: this is obvious if M ≥ Alt(d) and follows from the classification of 2-transitive groups if M = Sp8 (2) 12 ILIA PONOMARENKO AND ANDREY VASIL’EV (see, e.g., [3, Table 7.4]). Consequently, among 2d point stabilizers Mβ , β ∈ X ∪X ′ , there are exactly d distinct, and also Mβ 6= Mγ whenever β and γ are distinct points in X. Therefore, for every α ∈ X there is the only α′ ∈ X ′ such that Mα = Mα′ = Mα,α′ . Thus, the required bijection f takes α to α′ . To prove that L = U, assume on the contrary that U contains a block X ∈ L other than L. Denote by f : L → X the bijection defined in the above claim for X = L and X ′ = X. Then by the assumption, the element f (1), where 1 is the identity of the group G, does not belong to L. The binary relation sf is invariant with respect to the group KL ≥ L∗ and hence for all x ∈ L, (x, f (1)x) = (1, f (1))xr = (x, f (x)) = (1, f (1))xl = (x, xf (1)), where xr and xl are the permutations of L∗ induced by the right and left multiplication by x. This implies that the element f (1) ∈ X ⊂ U centralizes L. However, this is impossible, because f (1) 6= 1 and the group U is almost simple. Let us prove the second equality. The 2-closedness of K implies that KL is 2-closed. Furthermore, in view of the 2-transitivity of the group (KL )X its the 2-closure equals Sym(X). By Theorem 4.4 and equality L = U, this implies that Y Y Y KL = (KL )(2) = ( (KL )X )(2) = ((KL )X )(2) = Sym(X). X∈L X Thus, (KL ) X∈L = Sym(X) for all X ∈ L and we are done. X∈L  U Theorem 6.2 shows that in the case of symmetric type, the group K = Sym(U ) is the largest possible. In the normal type case, the group K U is quite small. More exactly, the following statement holds. Theorem 6.3. Let K be a group of normal type. Then K U ≤ D(2, U ). Proof. By the hypothesis of the theorem, L = Soc(L) and K L ≤ D(2, L). By Lemma 4.2, the first equality implies that L = Soc(G) and hence (16) L = Soc(U ). The second inclusion implies that (L∗ )X is a characteristic subgroup of the group (KL )X for all X ∈ L. By the definition of U , this implies that (L∗ )U is a characteristic subgroup of the group (KL )U . However, the latter group is normal in K U . Thus, (L∗ )U E K U . It follows that K U is contained in the normalizer of (L∗ )U in Sym(U ). However, this normalizer is contained in D(2, U ) by Theorem 5.1 applied for G = U with taking into account equality (16).  Proof of Theorem 4.1. Let U/L be the principal K-section of the group G. By Proposition 6.1, the group K is of symmetric or normal type. Suppose first that U = G. Then statement (i) of Theorem 4.1 holds. Indeed, if K is of symmetric type, then K = Sym(G) (Theorem 6.2), whereas if K is of normal type, then K ≤ D(2, G) (Theorem 6.3) and the required statement follows from the fact that G∗ is normal in D(2, G). Finally, if U < G, then statement (ii) of Theorem 4.1 holds by Theorem 4.4.  TESTING ISOMORPHISM OF CENTRAL CAYLEY GRAPHS 13 Proof of Theorem 1.3. First, assume that the group K = Aut(Γ) is a nontrivial generalized wreath product. Note that S = Soc(G) is a unique proper normal subgroup in G and |G/S| = 2. Therefore the generalized wreath product must be a usual one and the group K is permutation isomorphic to the wreath product M ≀ C2 in imprimitive action for some group M ≤ Sym(S). It follows that G \ S is an orbit of the point stabilizer Kα , where α is the identity of G. Since X ⊂ G \ S is a union of some orbits of Kα , we conclude that X = G \ S, a contradiction. The group K is 2-closed as the automorphism group of a graph. The normality of X implies that K satisfies condition (6) with G = Sym(m) for m ≥ 5. Finally, K is not a nontrivial generalized wreath product by above, and K 6= Sym(G), because the graph Γ is neither complete nor empty. Thus, by Theorem 4.1, the group G∗ is normal in K and G∗ ≤ K ≤ D(2, G), hence K = G∗ or K = D(2, G), because |D(2, G) : G∗ | = 2. However, since X is a normal subset of a symmetric group, we have X = X −1 , so the graph Γ has the automorphism σ : g 7→ g −1 , g ∈ G. Thus, K = hG∗ , σi = D(2, G), as required.  7. Finding the principal section in a Cayley scheme 7.1. The main resut. Let X be a central Cayley scheme over an almost simple group G. Then the group K = Aut(X ) is 2-closed and satisfies condition (6). Therefore, K is of symmetric or normal type by Proposition 6.1. In this section, we develop an algorithmic technique to determine (with no K in hand) which of these cases occurs for the scheme X . The main result here is Theorem 7.1 below which immediately follows from Corollaries 7.3 and 7.5 proved in Subsections 7.2 and 7.3, respectively. Theorem 7.1. Given a central Cayley scheme X over an almost simple group G of order n, one can determine the type of K = Aut(X ) and find the principal K-section of G in time poly(n). 7.2. The case of symmetric type. For a group H ≤ G, denote by H and WL(X , H) the partition of G into the right H-cosets and the Cayley scheme WL(X , T ) with T = {1X : X ∈ H}, respectively. Recall that TX is the trivial coherent configuration on X. Denote by H0 the set of groups H such that Soc(G) ≤ H ≤ G and (17) WL(X , H) = ⊞ TX . X∈H Lemma 7.2. In the above notation, the following statements hold: (i) if the group K is of symmetric type, then the set H0 is nonempty and the minimal block L of K is the largest (by inclusion) element of H0 , (ii) K is of symmetric type if and only if H0 contains S = Soc(G), Proof. To prove statement (i), assume that the group K is of symmetric type. Set U and L to be the partitions associated with the principal K-section of G. Then by Theorem 6.2, we have L = U and Inv(K X ) = TX for all X ∈ L. By Theorem 4.4, this implies that (18) Inv(KL ) = ⊞ TX . X∈L 14 ILIA PONOMARENKO AND ANDREY VASIL’EV The minimality of the direct sum implies that WL(X , L) = Inv(KL ), which proves formula (17) for H = L and H = L, in particular, H0 is nonempty. If L is not the largest element of H0 , then there exists H ∈ H0 such that H \ L 6= ∅. It follows that K ≥ Sym(H) × idG\H , where idG\H is the identity subgroup of Sym(G \ H). Hence there is a permutation k ∈ K that moves the identity of G to H \ L and leaves all non-identity elements of L fixed. But this is impossible, because L is a K-block. To prove the necessity for statement (ii), let K be of symmetric type. Then formula (18) holds. Therefore, if S is the partition of G into the cosets of S, then S refines L and hence WL(X , S) = WL(WL(X , L), S) = ⊞ WL(TX , SX ) = ⊞ TX , X∈L X∈S where SX is the partition of X induced by S. Consequently, S ∈ H0 . Conversely, assume on the contrary that K is of normal type. Then K U ≤ D(2, U ) by Theorem 6.3. Therefore K S ≤ Hol(S). On the other hand, since S ∈ H0 , we have Sym(S) = K S ≤ Hol(S), a contradiction.  From statement (i) of Lemma 3.1, it follows that the number of groups H ≤ G containing Soc(G) is at most log n, where n is the order of G; in particular, |H0 | ≤ log n. Moreover, for each H, the coherent configuration WL(X , H) can be efficiently found by the WL-algorithm and condition (17) can be verified by checking at most |H|2 ≤ n2 basis relations. Therefore, the set H0 can be found in time poly(n). By statement (ii) of Lemma 7.2, this is enough to test efficiently whether or not K is of symmetric type, and if it so, then to find the minimal block L = U of the group K (statement (i) of the same lemma). Corollary 7.3. Given a central Cayley scheme X over an almost simple group G of order n, one can test in time poly(n) whether the group Aut(X ) is of symmetric type, and (if so) find the principal section of Aut(X ) within the same time. 7.3. The case of normal type. In view of Corollary 7.3 and Proposition 6.1, one can efficiently test whether the automorphism group K of a central Cayley scheme X is of normal type. Denote by H1 the set of all groups H such that Soc(G) ≤ H E G and (19) H ∗ × idG\H ≤ K, where the left-hand side denotes the subgroup of Sym(G) that leaves each point of G \ H fixed and coincides with H ∗ on H. Lemma 7.4. Suppose that the group K is of normal type and U/L is the principal K-section of G. Then L = Soc(G), the set H1 is nonempty, and U is the smallest element of H1 . Proof. Lemma 3.2 yields that L = Soc(G). From Theorem 4.4, it is easily follows that U ∈ H1 . Assume on the contrary that the group U is not the smallest in H1 . Then there is a group V ∈ H1 such that L ≤ U ∩ V < U . Take a non-identity element w ∈ U ∩V and denote by kU (respectively, kV ) the permutation on G acting on U (respectively, V ) by right multiplication by w and acting trivially outside U (respectively, V ). Then kU , kV ∈ K, the permutation k = kU kV−1 TESTING ISOMORPHISM OF CENTRAL CAYLEY GRAPHS 15 is not identity on U , and U k = U and k U∩V = idU∩V . Clearly, k U belongs to K U , and hence to D(2, U ) because K is of normal type. However, as is easily seen, the identity element is the only element of D(2, U ) that leaves all points of L ≤ U ∩ V fixed, a contradiction.  Again from statement (i) of Lemma 3.1 it follows that the number of groups H such that Soc(G) ≤ H E G is at most log n, and so is |H1 |. For every H and each k ∈ H ∗ × idG\H one can efficiently test whether k is an automorphism of X . Thus, Lemma 7.4 immediately implies the following statement. Corollary 7.5. Given a central Cayley scheme X over an almost simple group G of order n, one can test in time poly(n) whether the group Aut(X ) is of normal type, and (if so) find the principal section of Aut(X ) within the same time. 8. A majorant for the coset of isomorphisms Throughout this section, we assume that X is a central Cayley scheme over an almost simple group G and K = Aut(X ). The principal K-section of G and the associated partitions are denoted by U/L and U and L, respectively. The equivalence relations corresponding to the partitions U and L, are denoted by eU and eL . Let ϕ be an algebraic isomorphism from X onto a Cayley scheme X ′ over an almost simple group G′ . Assume that (20) ϕ(eU ) = eU′ and ϕ(eL ) = eL′ , ′ where in what follows, the group K , the principal section U ′ /L′ , the partitions U′ and L′ , and the equivalence relations eU′ and eL′ are defined for the scheme X ′ in a similar way. Lemma 8.1. In the above notation, |U| = |U′ | and |L| = |L′ |. Moreover, the groups K and K ′ either both of symmetric type, or both of normal type. Proof. The first statement follows from assumption (20). To prove the second one, we note that by Lemma 7.2 the group K is of symmetric type if and only if the scheme X is isomorphic to the wreath product TU ≀ Y, where Y is the quotient of X modulo the equivalence relation eL . Since algebraic isomorphisms respect wreath products, we are done.  For all Y ∈ U and Y ′ ∈ U′ , the algebraic isomorphism ϕ induces an algebraic isomorphism ϕY ,Y ′ : XY → XY′ ′ , that takes a relation sY to the relation s′Y ′ for all basis relations s ⊆ eU of the scheme X , where s′ = ϕ(s). It follows that if X ′ = X and ϕ is trivial, then ϕY,Y is trivial for all Y ∈ U. For each Y ∈ U, set ( Sym(Y ) DY = D(2, U Y ) ∩ Aut(XY ) if K is of symmetric type, otherwise, where, for brevity, U Y denotes the restriction of the permutation group Uright to the set Y . Note that the form of the group DY does not depend on Y ∈ U, and DY contains K Y for all Y (Theorems 6.2 and 6.3). Furthermore, if Z ∈ U, then any permutation of G∗ taking Y to Z induces a permutation isomorphism from the group DY onto the group DZ . 16 ILIA PONOMARENKO AND ANDREY VASIL’EV Lemma 8.2. (DY )f = DY f for all Y ∈ U and all f ∈ Iso(X , X ′ , ϕ). Proof. Without loss of generality, we may assume that both K and K ′ are of normal types (Lemma 8.1). Let Y ∈ U and f ∈ Iso(X , X ′ , ϕ). Then in view of (20), Aut(XY )f = Aut(XY′ ′ ), (21) where Y ′ = Y f . Thus, it suffices to verify that D(2, V )f = D(2, V ′ ), where V = U Y ′ and V ′ = (U ′ )Y . However, by the definition of U, we have V = U g for suitable g ∈ G and D(2, V ) = D(2, U gr ), where gr : U → V is the bijection induced by right multiplication by g. A similar statement holds for V ′ , U ′ , and a suitable bijection (g ′ )r . Therefore, without loss of generality, we may assume that Y = U and Y ′ = U ′ . Recall that U is an almost simple group with Soc(U ) = L, and also U ∗ ≤ K U . Therefore, Soc(K U ) = L∗ . The same is true with K, U , and L, replaced by K ′ , U ′ , ′ and L′ , respectively. Taking into account that f takes K U to (K ′ )U , we conclude that (L∗ )f = (L′ )∗ . This implies that f takes the normalizer of the group L∗ ≤ Sym(U ) to the normalizer of the group (L′ )∗ ≤ Sym(U ′ ). However by Theorem 5.1, these normalizers are equal to D(2, U ) and D(2, U ′ ). Thus, f takes the first of these groups to the second, and we are done.  Definition 8.3. Denote by Cϕ (X , X ′ ) the set of all bijections f : G → G′ taking U to U′ and satisfying the following conditions for every Y ∈ U: (22) f Y ∈ Iso(XY , XY′ f , ϕY ,Y f ) and (DY )f = DY f . Let us find the explicit form of the set C = Cϕ (X , X ′ ) when X = X ′ and ϕ = id. In this case, C is obviously a subgroup of Sym(G) preserving the partition U. Condition (22) means that f Y belongs to the intersection of Aut(XY ) and the normalizer of DY in Sym(Y ) for all f ∈ CU . This proves the first of the equalities Y DY = CU and C U = Sym(U); Y ∈U the second equality follows from the fact that any g ∈ G induces the permutation isomorphism from DY onto DY g that induces ϕY ,Y g . Thus, the group C is permutation isomorphic to the wreath product DU ≀ Sym(U) (in imprimitive action). For arbitrary X ′ and ϕ, an for each f ∈ Cϕ (X , X ′ ), we obviously have (23) Cϕ (X , X ′ ) = Cid (X , X )f. Thus if the set Cϕ (X , X ′ ) is not empty, then it can be given by a generator set of the group C = Cid (X , X ) and the bijection f . In the sense of the following statement, the set Cϕ (X , X ′ ) can be called a majorant of Iso(X , X ′ , ϕ). Theorem 8.4. Iso(X , X ′ , ϕ) ⊆ Cϕ (X , X ′ ). Moreover, the set Cϕ (X , X ′ ) can be found in time poly(n). Proof. The first statement immediately follows from Lemma 8.2. To prove the second one, it suffices to find the set C0 = {f0 ∈ Isoψ (XU , XU′ ′ ) : (DU )f0 = DU ′ }, TESTING ISOMORPHISM OF CENTRAL CAYLEY GRAPHS 17 where ψ = ϕU ,U ′ . Indeed, if this set is empty, then obviously so is the majorant Cϕ (X , X ′ ). On the other hand, if f0 ∈ C0 , then to construct the majorant given by formula (23) it suffices to find DU and the bijection f defined as follows: f Y = (gY )−1 f0 gY′ ′ , Y ∈ U, ′ where Y 7→ Y is an arbitrary bijection from U onto U′ taking U to U ′ , and the bijections gY : U → Y and gY′ ′ : U ′ → Y ′ are induced by the right multiplications by the elements g ∈ G and g ′ ∈ G′ such that Y = U g and Y ′ = U ′ g ′ . To find the sets DU and C0 , assume first that K is of symmetric type (recall that this can efficiently checked by Theorem 7.1). Then the coherent configurations XU and XU′ ′ are trivial. Thus, DU = Sym(U ) and for any bijection f0 : U → U ′ , C0 = Sym(U )f0 . Let now K be of normal type. Then DU ≤ D(2, U ) and DU ′ ≤ D(2, U ′ ). In particular, DU = D(2, U ) ∩ Aut(XU ) can be found in time poly(n). Furthermore, every element f0 ∈ C0 takes DU to DU ′ (Lemma 8.2), and induces a permutation group isomorphism from Uright onto a group V ′ ∈ Reg(DU ′ , U ′ ). By statement (ii) of Lemma 3.1, the set Reg(DU ′ , U ′ ) is of cardinality at most |U ′ |c ≤ nc for some constant c > 0, and all its elements can be found by exhaustive search of all 3-generated subgroups of the group DU ′ . Since for a fixed V ′ , there are at most | Aut(V ′ )| ≤ nc distinct elements f0 ∈ C0 taking Uright to V ′ , one can test in time poly(n), whether the set C0 is not empty and (if so) find it in the form C0 = DU f0 with arbitrary f0 ∈ C0 .  9. Proof of Theorem 1.1 9.1. Reduction to Cayley schemes. Let Γ be a central Cayley graph over an almost simple group G, {ei : i ∈ I} the set of color classes of Γ, and K = Aut(Γ). The principal K-section of G and the associated partitions are denoted by U/L and U and L, respectively. Set X = WL(Γ, {eU , eL }), where eU and eL are the equivalence relations corresponding to the partitions U and L. For other central Cayley graphs Γ′ , we use similar notation, e.g., G′ and K ′ denote the underlying group and the automorphism group of Γ′ , respectively. Lemma 9.1. Given central Cayley graphs Γ and Γ′ over almost simple groups G and G′ , respectively, one can construct in time poly(n) the Cayley schemes X and X ′ over the same underlying groups and check whether there exists a (unique) algebraic isomorphism ϕ : X → X ′ such that (24) ϕ(ei ) = e′i for all i ∈ I and ϕ(eU ) = eU′ , ϕ(eL ) = eL′ , and (if so) find ϕ within the same time. Moreover, K = Aut(X ), K ′ = Aut(X ′ ), and also (25) Iso(Γ, Γ′ ) = Iso(X , X ′ , ϕ). 18 ILIA PONOMARENKO AND ANDREY VASIL’EV Proof. By Theorem 7.1, the principal sections and hence the equivalence relations eU , eL and eU′ , eL′ can be found in time poly(n). Therefore the first part of the statement immediately follows from Theorem 2.1. To prove the second one, we observe that every f ∈ Iso(Γ, Γ′ ) takes the group K to the group K ′ . By the definition of the minimal block (Subsection 4.2) this implies that f takes eL to eL′ and hence takes eU to eU′ . Thus, the isomorphism f induces ϕ. This means that f ∈ Iso(X , X ′ , ϕ) and hence the left-hand side of (25) is contained in the right-hand side. Since the reverse inclusion is obvious, equality (25) is completely proved. Next, if Γ′ = Γ, then ϕ = id and equality (25) shows that K = Aut(X ). Similarly, K ′ = Aut(X ′ ).  9.2. Determining the coset of isomorphisms. Denote by π the canonical epimorphism from G onto G = G/L = L. Then π induces a map taking the set S of basis relations of the Cayley scheme X over G to the set S of basis relations of the ∪ quotient Cayley scheme X over G. In particular, π takes S ∪ to S . Set Γ to be the Cayley graph over G with color classes π(ei ), i ∈ I. For any set C of the bijections f : G → G′ taking eU to eU′ and eL to eL′ , we denote by C the set of bijections ′ f : G → G induced by f ∈ C. Theorem 9.2. Let Γ and Γ′ be central Cayley graphs over almost simple groups G and G′ , respectively. Assume that the algebraic isomorphism ϕ from Theorem 9.1 does exist. Then Iso(Γ, Γ′ ) = π −1 (C ∩ B), (26) ′ where B = Iso(Γ, Γ ), C = Cϕ (X , X ′ ), and the right-hand side consists of all f ∈ C for which f ∈ B. Proof. To prove that that the left-hand side of (26) is contained in the right-hand side, let f ∈ Iso(Γ, Γ′ ). Then the uniqueness of the principal sections implies that (eU )f = eU′ and (eL )f = eL′ . Therefore, the isomorphism f induces the algebraic isomorphism ϕ. By Theo′ rem 8.4, this implies that f ∈ C. Consequently, the induced bijection f : G → G belongs to the set C. Since obviously f ∈ B, we conclude that f belongs to the right-hand side of (26), as required. Conversely, let f belong to the right-hand side of (26). By formula (25) in Lemma 9.1, it suffices to verify that f induces ϕ. To this end, let s ∈ S. Assume first that s ⊆ eU . Then s equals the union of sY , Y ∈ U. Therefore since f ∈ C, conditions (22) are satisfied for all Y and hence [ [ [ sf = ( sY )f = (sY )f = ϕY,Y f (sY ) = ϕ(s). Y ∈U Y ∈U Y ∈U Now assume that s is outside the equivalence relation eU . Let us prove that [ [ X ×Y and sf = (27) s= X ′ × Y ′, X,Y ∈L, (X,Y )∈s X ′ ,Y ′ ∈L′ , (X ′ ,Y ′ )∈s′ where s = π(s) and s′ = π ′ (sf ). Since (eU )f = eU′ , the relation sf is outside the equivalence relation eU′ . Therefore, it suffices to verify the first equality of (27); denote the right-hand side of this equality by t. Clearly, s ⊆ t. Conversely, let TESTING ISOMORPHISM OF CENTRAL CAYLEY GRAPHS 19 (X, Y ) ∈ t for some (X, Y ) ∈ s. Since s is outside eU , we conclude that X × Y is a basis relation of the coherent configuration Inv(KL ) (Theorem 4.4). Since Inv(KL ) ≥ X , it follows that this basis relation is contained in a basis relation of X which equals s, because (X, Y ) ∈ s. Therefore, t ⊆ s. This completes the proof of (27) implying s = π −1 (s) and sf = (π ′ )−1 (s′ ). On the other hand, the graph isomorphism f ∈ B ∩ C induces an algebraic ′ isomorphism ϕ : X → X that coincides with the restriction of the algebraic isomorphism ϕ modulo eL . Thus, ϕ(s) = ϕ(π −1 (s)) = (π ′ )−1 (ϕ(s)) = (π ′ )−1 ((s)f ) = (π ′ )−1 (s′ ) = sf , as required.  9.3. The algorithm. In the algorithm below, the input is given by two central Cayley graphs Γ and Γ′ over almost simple groups G and G′ , respectively. It is assumed that these groups are presented by the multiplication tables. The output consists of the set Iso(Γ, Γ′ ), which is either empty or equals the set Aut(Γ)f for some f ∈ Iso(Γ, Γ′ ). Here, the group Aut(Γ) is presented by a generating set. Central Cayley graph isomorphism test Step 1. Find the principal sections of the automorphism groups of the (central Cayley) schemes WL(Γ) and WL(Γ′ ) (Theorem 7.1); denote by U, L and U′ , L′ the associated partitions of G and G′ , respectively. Step 2. Find the schemes X = WL(Γ, {eU , eL }) and X ′ = WL(Γ′ , {eU′ , eL′ }) and the algebraic isomorphism ϕ satisfying condition (24); if ϕ does not exist, output Iso(Γ, Γ′ ) = ∅. Step 3. Find the set C = Cϕ (X , X ′ ) (Theorem 8.4). Step 4. Using the graph isomorphism and coset intersection algorithms from [1], ′ find the set B = Iso(Γ, Γ ) and then the set B ′ = B ∩ C. Step 5. Output Iso(Γ, Γ′ ) = π −1 (B ′ ).  To complete the proof of Theorem 1.1, we show that the above algorithm correctly finds the set Iso(Γ, Γ′ ) in time poly(n). Note that every graph isomorphism f ∈ Iso(Γ, Γ′ ) induces an algebraic isomorphism ϕ satisfying condition (24). Therefore, the output at Step 2 is correct. Thus, the correctness of the output at Step 5 and hence of the algorithm immediately follows from Theorem 9.2. To estimate the running time, we note that all the steps except for Step 4 run in polynomial time (Theorem 7.1, Lemma 9.1, and Theorem 8.4). Furthermore, the graph isomorphism and coset intersection algorithms from [1] are applied at Step 4 to graphs with m = |L| vertices and to the cosets contained in√Sym(m), respectively. Each of these algorithms runs in time at most exp((log m)c m). Since m ≤ log n, the complexity of this step does not exceed √ exp((log m)c m) ≤ exp((log log n)c (log n)1/2 ) ≤ exp(log n) ≤ n for sufficiently large n and a suitable constant c > 0. Thus, the running time of the algorithm is polynomial in n, as required.  20 ILIA PONOMARENKO AND ANDREY VASIL’EV References [1] L. Babai, W. Kantor, and E. M. Luks, “Computational complexity and the classification of finite simple groups,” in: Proceedings of the 24th Ann. Symp. Found. Comput. Sci (1983), pp. 162–171. [2] J. Bray, J. Holt, and D. Roney-Dougal, The Maximal Subgroups of the Low-Dimensional Finite Classical Groups, Cambridge University Press, Cambridge (2013). [3] P. J. Cameron, Permutation Groups, Cambridge University Press (1999). [4] J. H. Conway, R. T. Curtis, S. P. Norton, R. A. Parker, and R. A. Wilson, An ATLAS of Finite Groups, Oxford University Press, Oxford (1985). [5] F. Dalla Volta and A. Lucchini, “Generation of almost simple groups,” J. Algebra, 178, No. 1, 194–223 (1995). [6] S. Evdokimov and I. Ponomarenko, “Recognizing and isomorphism testing circulant graphs in polynomial time,” Algebra Analiz, 15, No. 6, 1–34 (2003). [7] S. Evdokimov and I. Ponomarenko, “Permutation group approach to association schemes,” Eur. J. Combin., 30,1456–1476 (2009). [8] S. Evdokimov and I. Ponomarenko, “Schurity of S-rings over a cyclic group and generalized wreath product of permutation groups,” Algebra Analiz, 24, No. 3, 84–127 (2012). [9] A. Ganesan, “Automorphism group of the complete transposition graph,” J. Algebr. Combin., 42, No. 3, 793–801 (2015). [10] C. H. Li, “On isomorphisms of finite Cayley graphs – a survey,” Discrete Math., 256, 301-334 (2002). [11] M. W. Liebeck, Ch. E. Praeger, and J. Saxl, “Regular subgroups of primitive permutation groups,” Memoirs Amer. Math. Soc., 203, No. 952, 1–88 (2010). [12] E. M. Luks, “Isomorphism of graphs of bounded valence can be tested in polynomial time,” J. Comp. Sys. Sci., 25, 42–65 (1982). [13] M. Muzychuk, “A solution of the isomorphism problem for circulant graphs,” Proc. London Math. Soc., 88, 1–41 (2004). [14] I. Ponomarenko, “Bases of schurian antisymmetric coherent configurations and isomorphism test for schurian tournaments,” J. Math. Sci., 192, No. 3, 316–338 (2013). [15] B. Weisfeiler (editor), On Construction and Identification of Graphs, Lecture Notes Math., 558 (1976). [16] H. Wielandt, Finite Permutation Groups, Academic press, New York - London (1964). [17] H. Wielandt Permutation Groups Through Invariant Relations and Invariant Functions, Lect. Notes Dept. Math. Ohio St. Univ., Columbus (1969). St.Petersburg Department of the Steklov Mathematical Institute, St.Petersburg, Russia E-mail address: inp@pdmi.ras.ru Sobolev Institute of Mathematics, Novosibirsk, Russia Novosibirsk State University, Novosibirsk, Russia E-mail address: vasand@math.nsc.ru 

4math.GR

Extending LTE into the Unlicensed Spectrum: Technical Analysis of the Proposed Variants Mina Labib*, Vuk Marojevic*, Jeffrey H. Reed*, Amir I. Zaghloul*† *Virginia Tech, Blacksburg, VA, USA arXiv:1709.04458v1 [cs.NI] 13 Sep 2017 † US Army Research Laboratory, Adelphi, MD, USA Abstract The commercial success of the Long Term Evolution (LTE) and the resulting growth in mobile data demand have urged cellular network operators to strive for new innovations. LTE in unlicensed spectrum has been proposed to allow cellular network operators to offload some of their data traffic by accessing the unlicensed 5 GHz frequency band. Currently, there are three proposed variants for LTE operation in the unlicensed band, namely LTE-U, Licensed Spectrum Access (LAA), and MulteFire. This paper provides a comparative analysis of these variants and explains the current regulations of the 5 GHz band in different parts of the world. We present the technical details of the three proposed versions and analyze them in terms of their operational features and coexistence capabilities to provide an R&D perspective for their deployment and coexistence with legacy systems. Index Terms Long Term Evolution (LTE), LTE-Advanced, LTE-U, LAA, MulteFire, Spectrum sharing, Wi-Fi. I. I NTRODUCTION The demand in mobile traffic has been growing tremendously since the introduction of smartphones in 2007. Since then, cellular network operators have been looking for new technologies to meet the demand. At that time, the 4G Long Term Evolution (LTE) standard had almost been completed by the 3rd Generation Partnership Project (3GPP). LTE specifications were finalized by the 3GPP in March 2009 (3GPP Rel-8). Cellular network operators quickly deployed LTE, starting in December 2009 by the Swedish-Finnish operator (TeliaSonera), as the enabling technology to meet the demand for more wireless data. True to its namesake, LTE has been able to keep pace with the growing demand for capacity through several added features and modifications. 3GPP Rel-10 for LTE-Advanced (LTE-A) was finalized in June 2011 and fully meets the 4G system requirements. LTE-A includes features such 1 as carrier aggregation (CA), which allows mobile operators to aggregate several frequency chunks into a larger bandwidth. In October 2015, the 3GPP announced the plan to further evolve LTE for paving the path towards 5G through LTE-Advanced Pro (LTE-A Pro). LTE-A Pro refers to LTE enhanced with the new features included in the 3GPP specifications starting from Rel-13 (which was finalized in March 2016), Rel-14 (which was released in January 2017 and expected to be finalized before September 2017), and onwards. One of the salient features of LTE-A Pro is extending LTE into the 5 GHz unlicensed spectrum, comprising the frequency range between 5150 MHz and 5925 MHz. The 5 GHz band, which is also known as the U-NII (Unlicensed National Information Infrastructure) band, is currently utilized by various radar systems, in addition to Wireless Local Area Networks (WLAN), specifically the ones that are based on IEEE 801.11a/g/n/ac technologies, which are also referred to as Wi-Fi systems. Currently, there are three proposed variants of LTE in unlicensed bands [1]. The first is called LTEU and is developed by the LTE-U Forum to work with the existing 3GPP Releases 10/11/12. LTE-U was designed for quick launch in countries, such as the United States and China, that do not mandate implementing the listen-before-talk (LBT) technique. The second variant is Licensed Assisted Access (LAA) and has been standardized by the 3GPP in Rel-13. LAA adheres to the requirements of the LBT protocol, which is mandated in Europe and Japan. It promises to provide a unified global framework that complies with the regulatory requirements in the different regions of the world. Both variants, LTE-U and LAA, use licensed spectrum as the primary carrier for signaling (control channels) and to deliver data for users with high Quality-of-Service (QoS) requirements. Carrier aggregation is used to add secondary component carriers in the unlicensed spectrum to deliver data to users with best-effort QoS requirements. MulteFire is the third variant of LTE in unlicensed bands and has been proposed as a standalone version of LTE for small cells. This variant will use only the unlicensed spectrum as the primary and only carrier, and it will provide an opportunity for neutral hosts to deploy LTE in the future. As opposed to other survey papers on this topic, such as [2] and [3], this paper identifies the motivation for introducing different modes of LTE operation in the unlicensed spectrum, analyzes them with respect to regulations and coexistence capabilities, and identifies research issues and the way forward. The rest of the paper is organized as follows: Section II summarizes the current regulations for the unlicensed 5GHz band in the different parts of the world as the basis for our analysis. Section III provides a general overview of extending LTE into the unlicensed spectrum in terms of potentials and challenges. Sections IV-VI discuss the technical details of each of the proposed variants of LTE for unlicensed band operation. We present numerical results in Section VII, comparing the performance of the three variants in terms 2 of coexistence with Wi-Fi. Section VIII provides the conclusions and an R&D perspective on future deployment and coexistence of radios in unlicensed spectrum. II. R EGULATIONS IN THE 5 GH Z BAND The regulatory requirements to access the spectrum are different in every region in the world, and the same applies for the 5 GHz band. In general, there are several compliance rules that have been defined around the world for regulating the use of the unlicensed spectrum. These rules can be summarized as follows: • Limitations of the maximum transmission power and the maximum power spectrum density (PSD). • Use indoor only or use both indoor and outdoor. • Dynamic Frequency Selection (DFS). • Listen-Before-Talk (LBT). • Transmission Power Control (TPC). DFS is a mechanism that is specifically designed to avoid causing interference to non-IMT (International Mobile Telecommunications) systems, such as radars. According to the Federal Communications Commission (FCC) regulations in the United States, any device (working in certain sub-bands of the 5 GHz band) must sense the channel before using it, and sense it periodically to ensure there is no radar system using this channel. If a radar signal is detected, i.e. the received power levels is above a certain threshold, the operating channel must be vacated. The device must not utilize that channel for the non-occupancy period of thirty minutes [4]. LBT is a mechanism introduced for fair co-existence with other wireless communication systems (such as Wi-Fi). In Europe and Japan, there is a mandatory requirement to implement LBT when accessing the unlicensed spectrum. According to the European Telecommunications Standards Institute (ETSI), and based on the Load Based Equipment rules, any device that wants to access the unlicensed spectrum, needs to perform Clear Channel Assessment (CCA) beforehand, which translates to spectrum sensing for certain period (called CCA period and it is greater than 20 µs). If the detected energy is lower than a certain threshold (which equals -73 dBm/MHz for the case of a transmitter with an EIRP of 23 dBm and assuming receiver antenna gain of 0), the device can access the channel for a period called channel occupancy time (which should be less than 13 32 q, where q is selected by the manufacturer and it is in the range 4-32). Then the device has to stay idle for a minimum period of CCA multiplied by a number that is randomly selected between 1 and q [5]. 3 Table I captures the regulatory requirements in the major regions of the world [5]. The maximum transmission power in every sub-band is limited by regulatory requirements, which motivates designing the LTE in unlicensed bands for the small cell network deployment. III. B ENEFITS AND C HALLENGES OF LTE IN U NLICENSED BANDS Extending the use of LTE in the 5-GHz unlicensed band can achieve several benefits compared to Wi-Fi. The benefits can be summarized as [2]: • Better spectrum efficiency: LTE is using scheduled-based channel access compared to the contentionbased scheme used by Wi-Fi. That leads to offering more efficient multiuser channel access and improves improve system capacity. Recent simulation results have shown that both LTE-U and LAA can achieve twice the capacity offered by Wi-Fi [1]. • Larger coverage area: LTE uses the more effective 1/3 turbo coding to overcome low SINR, has a more robust control channel design, and implements the HARQ (Hybrid Automatic Repeat Request) protocol, which makes it more robust to interference. • Unified LTE network: Operators will be able to use a single platform for authentication, registration and management. • Security: LTE offers more security than Wi-Fi due to the enhanced authentication procedures. • Better user experience: LTE offers good mobility management (which is not well supported in WiFi), so user will experience less service interruptions during mobility. Furthermore, the switch from using LTE in the licensed bands to the unlicensed one will be transparent to the users. However, LTE faces several challenges if it is extended into the 5 GHz unlicensed band in terms of coexistence with both radar and Wi-Fi systems. Wi-Fi systems are widely deployed in the 5 GHz band, and it is crucial that LTE in unlicensed bands does not cause degradation of Wi-Fi performance. Wi-Fi was designed to operate in the unlicensed spectrum and employs Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA) to access the channel and to ensure fair coexistence with other technologies. Recently, the coexistence between LTE and Wi-Fi in unlicensed bands has become an important area of research, from both industry and academia. The authors in [6] have conducted system-level analysis to evaluate the performance of both LTE and Wi-Fi when working in the same band. They concluded that, if no modifications were done to LTE, the performance of Wi-Fi may be severely degraded, while the performance of LTE would remain almost unchanged. The authors in [7] suggest using the almost-blank subframe (ABS) feature, which was introduced in LTE Rel-10, to improve the coexistence between LTE 4 TABLE I R EGULATORY REQUIREMENTS IN D IFFERENT W ORLD R EGIONS . 5150-5250 MHz 5250-5350 MHz 53505470 5470-5725 MHz 5725-5850 MHz 5850-5925 MHz WAS/RLAN FWA (Allowing WAS is under ITS (Allowing WAS is under Indoor/Outdoor, TPC (max EIRP is 30 dBm, max PSD is consideration) Indoor/Outdoor, TPC (max EIRP is 33 dBm, max PSD is consideration) (1) Indoor/Outdoor, TPC (max EIRP is 36 dBm), DFS (none for 5850- 17 dBm/MHz), DFS, LBT 23 dBm/MHz), DFS, No LBT 5875 LBT U-NII-2C U-NII-3 Indoor/Outdoor, max Tx Power is 30 dBm, max PSD is 30 dBm U-NII-4 MHz Usage WAS/RLAN WAS/RLAN Under consideration Europe Rules Indoor only, max Tx Power is 23 dBm, max PSD is 10 dBm/MHz, No TPC, DFS, LBT Usage USA Rules No U-NII-1 Indoor/Outdoor, max Tx Power is 30 dBm, max PSD is 17 dBm/MHz, No TPC, No DFS, No LBT Usage Canada (mx EIRP is 23 dBm, max PSD is 10 dBm/MHz), DFS, LBT U-NII-2A Indoor/Outdoor, max Tx Power is 24 dBm, max PSD is 11 dBm/MHz, TPC, DFS, No LBT U-NII-2B (Under consideration) Indoor/Outdoor, max Tx Power is 24 dBm, max PSD is 11 dBm/Mhz, TPC, DFS, No LBT (2) in 500 KHz, No TPC, No DFS, No LBT RLAN RLAN RLAN Indoor only, max EIRP 200 mW, max PSD is 4 dBm/MHz, No Indoor/Outdoor, max Tx Power is 24 dBm, max PSD is 11 Indoor/Outdoor, max Tx Power is 24 dBm, max PSD is 11 TPC, No DFS, No LBT dBm/MHz, TPC, DFS, No LBT dBm/MHz, TPC, DFS, No LBT Usage RLAN RLAN RLAN RLAN Rules Indoor only, max EIRP 200 mW, max EIRP PSD is 10 mW/MHz, no DFS Indoor only, max EIRP 200 mW, max EIRP PSD is 10 mW/MHz, DFS Indoor/Outdoor, max Tx power is 250 mW, max EIRP PSD is 50 mW/MHz, DFS Indoor/Outdoor, max Tx power is 1W Usage WAS/RLANs WAS/RLANs Rules Indoor only, max EIRP 200 mW, max EIRP PSD is 10 dBm/MHz, TPC, DFS, no LBT Indoor only, max EIRP 200 mW, max EIRP PSD is 10 dBm/MHz, TPC, DFS, no LBT RLAN Indoor only, max Tx power depends on BW, no TPC, no DFS, LBT RLAN RLAN Indoor only, max Tx power Indoo/Outdoor, max Tx power Rules Brazil China Usage Japan Indoor only, TPC Rules depends on BW, TPC, DFS, LBT N/A N/A (Under consideration) RLAN Indoor/Outdoor, max Tx Power is 30 dBm, max PSD is 17 dBm/MHz, No TPC, No DFS, No LBT N/A N/A RLAN N/A N/A Under Consideration depends on BW, TPC, DFS, LBT Rules not formally issued N/A N/A N/A (1) In Europe, FWA is utilizing the 5725-5875 MHz range, and ITS is utilizing the 5855-5925 MHz range. (2) In Canada, RLAN is forbidden in the frequency range 5600-5650 MHz. Abbreviations: WAS: Wireless Access Systems, RLAN: Radio Local Area Networks, FWA: Fixed Wireless Access, U-NII: Unlicensed National Information Infrastructure. MHz), 5 and Wi-Fi technologies. ABS is a feature that allows the LTE base station (eNodeB) to transmit subframes that contain only the basic system information messages, and it is used for better coordination between macro-cells and small-cells. It can be considered as a sort of static muting mechanism. The authors in [8] show that a simple LBT algorithm will provide better coexistence performance than the static muting algorithm. Reference [3] surveys the coexistence mechanisms proposed for LTE and Wi-Fi systems in unlicensed bands. IV. LTE-U LTE-U is developed by the LTE-U Forum, which was formed in 2014 by Verizon, Alcatel-Lucent, Ericsson, Qualcomm Technologies Inc., and Samsung Electronics. The goal of LTE-U is to use the existing features in the latest LTE 3GPP specifications and adapt them to unlicensed operation in countries, such as the US and China, that do not mandate LBT. LTE-U supports supplemental downlink (SDL) only within the frequency bands 5150-5250 MHz and 5725-5850 MHZ, whereas the frequency bands 52505725 MHz has been reserved for future use. The last set of specifications for LTE-U (issued in October 2015) provides general technical guidelines and benchmarks for testing scenarios, but does not specify certain implementation mechanisms [9]. LTE-U specifications were designed for the case when a single eNodeB has access to licensed spectrum (called Primary Cell or PC) and unlicensed spectrum (called Secondary Cell or SC). Without modifying the 3GPP LTE specifications, there are several mechanisms that are used for LTE-U to better coexist with Wi-Fi [1], such as carrier selection, on-off switching, and Carrier-Sensing Adaptive Transmission (CSAT). A. Carrier Selection The eNodeB performs carrier selection at startup, periodically and based on performance triggers. Carrier selection implies scanning the spectrum and measuring the power level in each channel to find the channel that is free of interference. If all channels are occupied by other systems, the eNodeB choses the channel with the lowest detected signal power level. The eNodeB will continue monitoring channel activities and select a more suitable channel when available. Carrier selection algorithm was left to be implementation specific. B. On-Off Switching When the traffic demand is low, the small cell eNodeB can stop transmitting in the unlicensed spectrum and relies only on the licensed spectrum. Doing this will reduce the amount of interference to Wi-Fi 6 users. LTE-U specifications define two states for the LTE-U SC [10]: • Off-State: The SC stops any type of transmission. • On-States: The SC is either transmitting full LTE frames according to the 3GPP specifications or transmitting the LTE-U Discovery Signal (LDS). The LDS is transmitted by the SC at a certain subframe (subframe number 5) and with fixed time intervals defined by the LDS periodicity parameter (which can be either 40, 80, or 160 ms). LDS contains the physical signals and channels required for the LTE User Equipment (UE) to obtain time and frequency synchronization and to perform SC measurements. C. Carrier-Sensing Adaptive Transmission CSAT is a mechanism that allows the eNodeB to share the spectrum with other systems using the same channel in a Time Division Multiplex (TDM) manner. Qualcomm proposed CSAT as a spectrum sharing technique to be used with LTE-U [1]. CSAT, in concept, is a sort of adaptive muting algorithm, where the eNodeB initially and periodically senses the channel for relatively long time periods (anywhere in the rage between 0.5 and 200 ms). As a function of the channel activity and detection of Wi-Fi signals above the energy threshold level (which is -62 dBm), SC will adjust its duty cycle and define a time cycle for transmission. Since Wi-Fi stations use carrier sensing, they will be able to adjust their own transmissions in the periods when the duty cycle of LTE-U SC is off. The CSAT duty cycle can change over time based on channel usage, but the constraint values are [10]: • Minimum Off-State Period: 1 ms. • Maximum Off-State Period: Determined by the LDCS periodicity, which can be either 40, 80, or 160 ms. • Minimum On-State Period: 4 ms in case of available user data and 1 ms (LDS period), otherwise.. • Maximum On-State Period: 20 ms. • Energy detection threshold: -62 dBm. An obvious drawback of CSAT is the long latency that may not be suitable for real-time applications over Wi-Fi. Another drawback is that the eNodeB can extend its transmission until the duty cycle reaches 90% when it cannot detect the Wi-Fi signal (hidden node problem), which will lead to diminishing the Wi-Fi signal. Recently, the FCC approved that Qualcomm and Verizon perform small-scale testing for LTE-U in real-world scenarios at two different locations. If test results show that LTE-U can coexist fairly with Wi-Fi systems, LTE-U might be commercially deployed in 2017. 7 V. L ICENSED A SSISTED ACCESS (LAA) Licensed Assisted Access (LAA) was standarized in 3GPP Rel-13. The operating frequency band for LAA spans the frequency range 5150 MHz - 5925 MHz (channel numbers 46 and 47 in the 3GPP specifications). The current allowable bandwidths for LAA operation in unlicensed spectrum are 10 and 20 MHz [11]. Rel-13 defines LAA only for the downlink (DL). One of the main features of 3GPP Rel-14 is the introduction of enhanced-Licensed Assisted Access (eLAA), which includes uplink (UL) operation for LAA. LAA has a different frame structure type (Frame Sturcture Type 3) for operation in the unlicensed band. Frame Structure Type 3, similar to the one defined for FDD, has a duration of 10 ms and consists of 20 slots, each slot duration is 0.5 ms and each two adjacent slots form one subframe. Any subframe may be available for DL or UL transmission, however the transmission may or may not start at the boundary of the subframe, and it may or may not end at the boundary of the subframe [12]. This is due to the fact that the eNodeB has to sense the spectrum before transmitting and transmits only if the channel is free. Since one of the goals of LAA is ensuring fair coexistence with Wi-Fi, we will further examine the channel access procedures for eNodeB in the unlicensed spectrum as detailed in [13]. 3GPP Rel-13 has identified two different modes for the eNodeB to transmit in the unlicensed spectrum. The first mode is for transmitting the Physical Downlink Shared Channel (PDSCH), which is the channel that carries user data on the DL. The second mode is for transmitting a Discovery Reference Signal (DRS) without the PDSCH. For the case of transmitting the PDSCH, Figure 1 illustrates the main procedures that need to be followed by the eNodeB. These procedures are based on the LBT algorithm with random back-off and a variable contention window size (LBT Category 4), where the eNodeB generates a random number within the contention window size to identify the period it needs to sense the spectrum before transmission. The size of the contention window is variable as the device will increase the window size when it finds the spectrum occupied. The 3GPP specifications have identified four different priority access classes, which define the channel sensing parameters. It is worth mentioning that these parameters are very similar to those used for the different access categories in IEEE 802.11, with the aim of achieving fair coexistence among the two technologies. To elaborate more, the eNodeB starts by sensing the channel for a period called defer duration time Td , then selects a random number N that is uniformly distributed between 0 and CWp , where CWp is the contention window size that has a minimum and maximum values, which are determined by the channel access priority class. The eNodeB senses the channel for an additional period of N + 1 times 8 Fig. 1. eNodeB procedures to access the unlicensed spectrum before transmitting PDSCH. the slot duration, which equals 9 µs. If the channel is found to be idle during all the slot durations of Td and during the N + 1 slot durations, the eNodeB can start transmitting with a maximum duration of Tmcot,p (which ranges between 2 and 10 ms depending on the priority channel access class). The channel is determined to be idle during a certain time slot if the detected power is less than a certain energy detection threshold XT hresh for at least 4 µs of the slot duration time. If the channel is found to be busy during any time slot, the eNodeB continues sensing the channel for an additional defer duration time Td . The defer duration time Td consists of a duration Tf , which equals 16 µs, followed immediately by mp additional slot durations Tsl . Each slot duration Tsl is 9 µs and Tf includes an idle slot duration Tsl at its start, wile mp ranges between 1 and 7, where its value depends on the priority channel access class. For the choice of the contention window size CWp , the eNodeB starts with the minimum value CWmin and if the data was not received correctly by the users, the eNodeB chooses the next higher CWp , and so 9 on until reaching the maximum allowable value CWmax and continues using it until the data is correctly received. For the case of transmitting a DRS, the eNodeB needs to sense the channel for a period of 25 µs and if the channel is found to be idle for the entire period, the eNodeB can transmit a discovery signal for a maximum period of 1 ms. The DRS is identified for Frame Structure Type 3 and contains 12 OFDM symbols within one non-empty subframe. It carries the primary synchronization signal (PSS), secondary synchronization signal (SSS) and cell specific reference signal (CRS) [12]. The value of the energy detection threshold XT hresh is determined differently in two different cases. For the case where there is no other technology occupying the channel on a long term basis (as in the case of regulations in certain regions) and for 20 MHz bandwidth, XT hresh = min    −52dBm, (1)   Xr where Xr is the maximum energy detection threshold defined by regulation requirements in dBm (if defined). For the case where multiple technologies are allowed to share the channel and for the 20 MHz bandwidth,     −72dBm,             −62dBm,    XT hresh = max    min −62 − TA + (23 − PT x )dBm,                 (2) where TA is either 10 dBm in case of transmitting PDSCH or 5 dBm in case of transmitting DRS, and PTx is the maximum eNodeB output power in dBm. For UL transmission, which was defined in Rel-14, the UE also is allowed to use one of two modes to transmit in the unlicensed spectrum. These two modes of sensing the spectrum are similar to the ones used by the eNodeB. In most cases, the eNodeB informs the UE of the mode that it needs to use before accessing the spectrum along with the UL grant. It is worth mentioning that the 3GPP specifications have also identified the procedures for the eNodeB when accessing multiple channels in the unlicensed spectrum, allowing the eNodeB to choose the channel that has that lowest detected signal power level to 10 reduce interfering with other existing systems within the unlicensed spectrum. Qualcomm has performed a laboratory tests to evaluate the performance of the coexistence algorithm implemented in LAA. The results show that when an operator switches from Wi-Fi to LAA, the throughput for this operator increases by 100%, and even the throughput for the Wi-Fi operators increases by approximately 10% [1]. VI. M ULTE F IRE MulteFire was proposed as a standalone version of LTE for small cells. In December 2015, MulteFire alliance was formed by Qualcomm, Nokia, Ericsson and Intel to promote the MulteFire technology and several companies have joined the alliance since then. The first release of the technical specifications for MulteFire was issued in Jan. 2017 [14]. MulteFire, similar to Wi-Fi, relies only the unlicensed spectrum and can provide service to users with or without USIM (Universal Subscriber Identity Module) card. Hence, MulteFire will combine the benefits of the advanced LTE technology and the simplicity of Wi-Fi deployment [1]. MulteFire can be deployed either by traditional mobile operators or by neutral hosts. Accordingly, MulteFire specifies two different architectures: • Public Land Mobile Network (PLMN) access mode, which allows mobile network operators to extend their coverage into the unlicensed band, specially in case where licensed spectrum is not available at certain locations. • Neutral Host Network (NHN) access mode, which is similar to Wi-Fi, a self-contained network deployment that provides access to the Internet. Because of the nature of transmission in the unlicensed band and the need to adhere to the LBT requirements, MulteFire has introduced several modifications in the radio air interface compared to LTE. A. Downlink Operation As for LAA, a MulteFire eNodeB will need to perform LBT before transmitting any signal. The LBT procedure is similar to the one of LAA and has the same four channel access priority classes. Also similar to LAA, the eNodeB can transmit a DRS that contains critical data for synchronization and acquiring the system information. DRS for MulteFire is also 12 OFDM symbols long, but its structure is different when compared to LAA. The components of DRS for MulteFire are: • Primary Synchronization Signal (PSS): transmitted on the seventh OFDM symbol of the DRS subframe. 11 • Secondary Synchronization Signal (SSS): transmitted on the sixth OFDM symbol of the DRS subframe. • MulteFire Primary Synchronization Signal (MF-PSS): transmitted on the fourth OFDM symbol of the DRS subframe. • MulteFire Secondary Synchronization Signal (MF-SSS): transmitted on the third OFDM symbol of the DRS subframe. MF-PSS and MF-SSS support the UE in performing frequency/time synchronization and also allow differentiating between an LAA eNodeB and a MulteFire eNodeB. • Cell-specific reference signals (CRS). • Configurable channel state information reference signals (CSI-RS). • Master information broadcast (MIB-MulteFire) through the MulteFire Physical Broadcast Channel (MulteFire-PBCH), which is transmitted over six OFDM symbols (the fifth, eighth, ninth, tenth, eleventh and last OFDM symbol). • MulteFire system information broadcast (SIB-MulteFire) which is transmitted through PDSCH and carries information similar to the SIB1 and SIB2 messages of LTE Rel-13. MulteFire allows sending the DRS in two modes: • During the serving cell DRS measurement and timing configuration (DMTC) window, which can be up to 10 ms long and during which the UE expects to receive the DRS. The DMTC periodicity is 40, 80 or 160 ms. MulteFire will transmit the DRS during the DMTC window after sensing the channel for a period of 25 µs. • Just like for the downlink PDSCH, opportunistic transmission of DRS is allowed only on subframe number 0, after performing the LBT mechanism. B. Dynamic DL/UL Configuration MulteFire adopts a very flexible frame structure to dynamically adapt to the DL and UL traffic loads. Accordingly, the ratio between DL and UL transmission can vary from one frame to the next. The eNodeB will broadcast whether a subframe is DL or UL through the Common Physical Downlink Control Channel (C-PDCCH). C. Uplink Operation MulteFire uses Block Interleaved FDMA (B-IFDMA) as the UL transmission scheme, where the bandwidth is divided into N interlaces (N = 10 for 20 MHz, and N = 5 for 10 MHz), each interlace consists of 10 equally spaced physical resource blocks. MulteFire introduces two different formats for the 12 Physical Uplink Control Channel (PUCCH): Short-PUCCH (MF-sPUCCH) and Extended PUCCH (MFePUCCH). MF-sPUCCH is transmitted by the UE during the last four OFDM symbols of a DL subframe. In other words, the UE is allowed to transmit the MF-sPUCCH immediately on the gap between DL and UL transmission, and, therefore, does not need to perform LBT. This is allowed according to the ETSI regulations in the unlicensed band, as long as the UE is transmitting within 16 µs after the DL transmission. Due to its compact design, MF-sPUCCH will be used to carry small control information such as reception acknowledgments. On the other hand, MF-ePUCCH will be used by the UE to transmit large control information such as the channel state information (CSI). The UE will transmit MF-ePUCCH based on the UL resource assignment given by the eNodeB, the same way the UE transmits the Physical Uplink Shared Channel (PUSCH). Because of the huge demand for high data rates and for higher capacity, MulteFire has the potential to use spectrum opportunities and leverage LTE-A techniques to access the unlicensed spectrum. Furthermore, MulteFire supports mobility management and can provide a better user experience than Wi-Fi. However, MulteFire faces several challenges. Wi-Fi is already very widely used nowadays, not only in cellphones, but also in portable computing devices (laptops), so Wi-Fi chips are manufactured at low cost because of it high penetration. Furthermore, MulteFire will not be backward compatible with legacy LTE devices. Moreover, there are still lots of design issues need to be considered, such as the control channel design and their reliability in the unlicensed band. One of these issues is handling the unintentional interference generated among several eNodeBs in dense deployment as illustrated in [15]. VII. C OMPARATIVE A NALYSIS The objective of this analysis is to compare the coexistence performance of CSMA/CA (used in Wi-Fi), LBT (used in LAA and MulteFire), and CSAT (used in LTE-U). The performance is compared based on the normalized total time of transmission opportunities (TTTO) that is granted (free of collision) to an operator, who is either deploying Wi-Fi, LAA/MulteFire, or LTE-U. Table II shows the simulation parameters. TTTO is calculated based on the summation of transmission opportunity durations that an operator gets free of collision, normalized by the total analysis time. The coexistence performance of LAA/MulteFire or LTE-U is evaluated based on the average TTTO that a Wi-Fi operator will get (when adding an LAA/MulteFire or LTE-U eNodeB) compared to the case when adding another Wi-Fi operator. The simulation results are presented in Figure 2 and illustrate the following: When deploying three Wi-Fi operators, the average TTTO per operator is 33.01%. When one operator switches to LAA, the TTTO for each of the remaining operators increases to 36.728% on average. When one of three Wi-Fi operators 13 TABLE II S IMULATION PARAMETERS General Simulation Parameters Wi-Fi Specific Parameters LAA Specific Parameters LTE-U Specific Parameters Sampling Time 1 µs Simulation Time 100 s Traffic Model Full Buffer Access Category Video Minimum Contention Window 7 Maximum Contention Window 15 Arbitration Inter-Frame Spacing Number 2 Short Inter-Frame Spacing (SIFS) 16 µs DCF Inter-Frame Spacing (DIFS) 34 µs Data transmission time (based on maximum PPDU frame) 5.484 ms Acknowledgment Transmission Time 34 µs Priority Class 2 Minimum Contention Window 7 Maximum Contention Window 15 mp 1 Tmcot,p 3 ms On-State Period 12 ms Off-State Period 24 ms switches to LTE-U, this TTTO decreases to 30.95%. This means that LAA and MulteFire are better neighbors to Wi-Fi than Wi-Fi itself and than LTE-U. It is worth mentioning that the TTTO metric can be a good indicator of DL throughput for operators with the same technology, but this is not true for operators using different technologies. The throughput in this case will depend on the spectral efficiency of each technology and how efficient each technology can utilize the time when granted access to the channel. Although LAA and MulteFire perform the same in terms of TTTO, the throughput will be different for the two variants, since LAA uses the licensed spectrum to transmit the control messages. In our simulations, it was assumed that the received power of the other operators is always above the detection threshold level. In reality that may not always be true. Actually, since the detection threshold level is -62 dBm for LTE-U and -72 dBm for LAA/MulteFire, LAA/MulteFire operators will have larger detection range for Wi-Fi signal and outperform LTE-U even more in terms of fair coexistence. Table III and Figure 3 provide a comparison of the three proposed variants of LTE in unlicensed bands. Note that the radar chart of Figure 3 reflects the current state of the art. The shapes may change 14 Operator A Opeator B Operator C Average For WiFi Operators 40.00% 36.728% 35.00% 33.01% 30.953% 30.00% 25.00% 20.00% 15.00% 10.00% 5.00% 0.00% WiFi Fig. 2. WiFi WiFi WiFi Avg WiFi WiFi LTE-U WiFi Avg WiFi WiFi LAA WiFi Avg Performance comparison in terms of normalized total time of transmission opportunities (free of collision). as research, regulation and standardization evolves. VIII. C ONCLUSION This paper has discussed LTE in unlicensed spectrum as the next big milestone in the evolution of LTE. We have presented the different regulatory requirements for the 5 GHz unlicensed bands in different world regions and analyzed the benefits and the challenges for operating LTE in unlicensed spectrum. We have explained the technical details of the three proposed variants of LTE for unlicensed spectrum and have compared the coexistence mechanism used in these variants. Our numerical analysis has shown that LAA and MulteFire offer better coexistence to Wi-Fi than LTE-U. This paper has highlighted the differences and uniquenesses of each variant of LTE operating in unlicensed spectrum. LAA is the most unified solution and will be operable worldwide. LTE-U is less regulated and thus expected to be introduced first. MulteFire is most flexible and will be as simple to deploy as Wi-Fi. Moreover, MulteFire has the potential to play a big role in the future of wireless communications, especially for enterprises and industrial applications, as several major high-tech companies 15 LTE-U LAA MuLTEfire Faster Time to Market Simplicity of Deployment Backward Compatibility Fig. 3. Coexistence with Wi-Fi Unified Solution Comparison between the three variants of LTE in unlicensed bands. are currently collaborating to improve performance and simplify deployment and configuration. MulteFire also has the potential to enhance the future of the Internet of Things (IoT), as it can enable new IoT deployment scenarios, hence creating new business opportunities. It is expected that LTE in unlicensed bands will become more important and the technology be leveraged for LTE moving into additional unlicensed bands. This can be expected with the ongoing spectrum relocation in the US and worldwide. Moreover, history has shown that popular technologies (such as LTE or Wi-Fi) expand. For example, Wi-Fi has gone through several generations and expanded from the 2.4 GHz band to the 5 GHz, 60 GHz (WiGig) and former TV bands (Super-Wifi). We will likely see different variants of LTE in unlicensed bands in the market because of different regulations and incentives. This diversity will help gaining experience for next generation wireless networks. Furthermore, LTE and Wi-Fi will both evolve and once radars are relocated, there will be more room for improvement and efficiency to make better use of spectrum opportunities. LTE in unlicensed spectrum is not competing with Wi-Fi; they complement each other, have their pros and cons and may cooperate in the future. 16 TABLE III C OMPARISON BETWEEN THE THREE DIFFERENT VERSIONS OF LTE IN UNLICENSED BANDS LTE-U LAA MulteFire Devloped By LTE-U Forum 3GPP MulteFire Alliance Mode of Opera- Supplemental downlink (SDL) DL is supported in Rel-13, UL DL and UL (using Time Divi- tion only is standardized in Rel-14 sion Multiplexing (TDD)) 5150-5250 MHz and 5725-5850 Frequency Bands The 5 GHz, and the 3.5 GHz MHZ, whereas the frequency 5150 MHz- 5925 MHz (Chan- in the US. The 1.9 & 2.4 GHz bands 5250- 5725 MHz has nels 46 and 47) bands are expected to be sup- been reserved for future use. Bandwidths Allowed Purpose Coexistence Mechanisms 20 MHz. 10 MHz and 20 MHz. 10 MHz and 20 MHz. Use of the existing LTE specifi- Single unified global framework Standalone version that operates cations in countries that do not that complies with regulations without a primary carrier in li- impose LBT. in the different world regions. censed band. Carrier selection and LBT. LBT (similar to LAA). Carrier selection, on-off switching, Carrier-Sensing Adaptive Transmission (CSAT). Will likely be the first version Main Advantage ported in subsequent releases in the market; Samsung and Verzion announced their plans to implement it in 2016. Will be a single unified global framework for LTE in unlicensed bands worldwide. Combines benefits of LTE with simplicity of Wi-Fi. Backward compatibility Yes, since using the licensed spectrum as the primary carrier. No. with 3GPP Rel-9 R EFERENCES [1] Qualcomm Technologies, Inc., “LTE-U/LAA, MuLTEfire and Wi-Fi; making best use of unlicensed spectrum,” Sept. 2015. [2] R. Zhang, M. Wang, L. Cai, Z. Zheng, and X. Shen, “LTE-unlicensed: the future of spectrum aggregation for cellular networks,” IEEE Wireless Commun., vol. 22, no. 3, pp. 150–159, June 2015. [3] B. Chen, J. Chen, Y. Gao, and J. Zhang, “Coexistence of LTE-LAA and Wi-Fi on 5 GHz with Corresponding Deployment Scenarios: A Survey,” IEEE Communications Surveys Tutorials, vol. PP, no. 99, pp. 1–27, 2016. [4] FCC, “Revision of Part 15 of the Commissions rules to permit Unlicensed National Information Infrastructure (U-NII) devices in the 5 GHz band,” Federal Communications Commission, First Report and Order, April 2014. [5] 3GPP, “Study on Licensed-Assisted Access to Unlicensed Spectrum; (Release 13),” 3rd Generation Partnership Project (3GPP), TS 36.889, Jun. 2015. [Online]. Available: http://www.3gpp.org/dynareport/36889.htm [6] A. Cavalcante, E. Almeida, R. Vieira, F. Chaves, R. Paiva, F. Abinader, S. Choudhury, E. Tuomaala, and K. Doppler, 17 “Performance evaluation of LTE and Wi-Fi coexistence in unlicensed bands,” in Proc. IEEE VTC ’13-Spring, June 2013, pp. 1–6. [7] E. Almeida, A. Cavalcante, R. Paiva, F. Chaves, F. Abinader, R. Vieira, S. Choudhury, E. Tuomaala, and K. Doppler, “Enabling LTE/WiFi coexistence by LTE blank subframe allocation,” in Proc. IEEE ICC ’13, June 2013, pp. 5083–5088. [8] J. Jeon, H. Niu, Q. Li, A. Papathanassiou, and G. Wu, “LTE in the unlicensed spectrum: evaluating coexistence mechanisms,” in Proc. IEEE Globecom Workshops ’14, Dec 2014, pp. 740–745. [9] LTE-U Forum, “LTE-U SDL Coexistence Specifications V1.3,” Oct. 2015. [10] ——, “LTE-U CSAT Procedure TS V1.0,” Oct. 2015. [11] 3GPP, “Evolved Universal Terrestrial Radio Access (E-UTRA); User Equipment (UE) radio transmission and reception (Release 14),” 3rd Generation Partnership Project (3GPP), TS 36.101, 2017. [Online]. Available: http://www.3gpp.org/dynareport/36101.htm [12] ——, “Evolved Universal Terrestrial Radio Access (E-UTRA); Physical channels and modulation (Release 14),” 3rd Generation Partnership Project (3GPP), TS 36.211, 2017. [Online]. Available: http://www.3gpp.org/dynareport/36211.htm [13] ——, “Evolved Universal Terrestrial Radio Access (E-UTRA); Physical layer procedures (Release 14),” 3rd Generation Partnership Project (3GPP), TS 36.213, 2017. [Online]. Available: http://www.3gpp.org/dynareport/36213.htm [14] MulteFire Alliance, “MulteFire Release 1.0 Technical Paper: A new way to wireless,” Jan. 2017. [15] M. Labib, V. Marojevic, J. H. Reed, and A. I. Zaghloul, “Enhancing the Robustness of LTE Systems: Analysis and Evolution of the Cell Selection Process,” IEEE Communications Magazine, vol. 55, no. 2, pp. 208–215, February 2017. 

7cs.IT

Comparison of Volumes of Siegel Sets and Fundamental Domains for SLn(Z) arXiv:1604.03613v4 [math.GR] 23 Mar 2018 Gisele Teixeira Paula Instituto Nacional de Matemática Pura e Aplicada. Est. Dona Castorina, 110. CEP 22460-320. Rio de Janeiro-RJ. Brazil. Correspondence to be sent to: e-mail: giseletp@impa.br Abstract The purpose of this paper is to calculate explicitly the volumes of Siegel sets which are coarse fundamental domains for the action of SLn (Z) in SLn (R), so that we can compare these volumes with those of the fundamental domains of SLn (Z) in SLn (R), which are also computed here, for any n ≥ 2. An important feature of this computation is that it requires keeping track of normalization constants of the Haar measures. We conclude that the ratio between volumes of fundamental domains and volumes of Siegel sets grows super-exponentially fast as n goes to infinity. As a corollary, we obtained that this ratio gives a super-exponencial lower bound, depending only on n, for the number of intersecting Siegel sets. We were also able to give an upper bound for this number, by applying some results on the heights of intersecting elements in SLn (Z). Keywords: Arithmetic Groups, Siegel Sets, Coarse Fundamental Domains, Volumes. 1 Introduction Siegel sets were first introduced in the study of quadratic forms by Siegel [9] in 1939, with some results following from previous works of Hermite and KorkineZolotarreff. In a fundamental paper [2], Borel and Harish-Chandra have generalised this notion and used Siegel domains to prove finiteness of covolumes of non-cocompact arithmetic subgroups. The simple structure of Siegel sets, compared to those of the actual fundamental domains makes them appealing for applications. For example, in his recent paper [14], R. Young exploited their properties to obtain new results in geometric group theory. Still very little is known about the geometry of Siegel sets in general. In his book [7], Morris describes algebraically examples of Siegel sets not only for SLn (R), with n ≥ 2 , but also in the case of any semisimple Lie group G with a given Iwasawa decomposition. In this paper we recall one of the main properties of Siegel sets – the finiteness of their volumes. We evaluate these volumes explicitely in the basic case of Siegel sets for SLn (Z) in SLn (R) for any n ≥ 2. We then compare these volumes with the actual covolumes of SLn (Z). To this end, we have to deal with an essential difficulty related to the normalization of the Haar measure. 1 For calculating the volumes of Siegel sets, the main difficulty is to find a nice way to describe the region of integration, which we solve with an appropriate change of coordinates. Most of the volume computations that followed Siegel’s original approach were not careful about the normalization constants, just noting that they are computable and could be calculated from the proof. In Section 5, we follow Garret’s notes on Siegel’s method [4] to compute the volumes of the quotients SLn (Z)\SLn (R) for n ≥ 3 using induction and the volume of SL2 (Z)\SL2 (R), that is computed in [4]. Our main goal here is to keep a careful track of the normalization constants. The main tools we use are the Poisson Summation formula, the Iwasawa decomposition of G and the choice of a good Haar measure normalization on each group. At the end of the section we discuss the relation between the normalization of the measure we used and the canonical normalization that comes from the metric associated to the Killing form on sln (R). By comparing the volumes of Siegel sets and the volumes of fundamental domains of SLn (Z), we conclude that somewhat surprinsingly the ratio between them grows super-exponentially fast with n. As an application of the computations presented here, in Section 6 we show that given a Siegel set Σ of SLn (Z), we have an explicit lower bound for the number of elements γ ∈ SLn (Z) such that γΣ intersects Σ. This bound is given by the ratio between vol(Σ) and vol(SLn (Z)\SLn (R)) – see Corollary 6.1. We also give a proof that this result is consistent with a recent work of M. Orr [8], which generalizes a previous result of P. Habegger and J. Pila [5] on the height of such elements γ, motivated by the study of Shimura varieties and their unlikely intersections. More precisely, Orr’s result gives, as a corollary, an upper bound for the number of intersecting Siegel sets while our work provides a lower bound for this number (see Corollary 6.2). It would be interesting to compute the volumes of Siegel sets in other cases, for example for the action of well known Bianchi groups Γd = SL2 (Od ) on the hyperbolic three-dimensional space H3 . In this case we should have to deal with another difficulty when describing Siegel sets, because of the fact that as d grows the quotients Γd \H3 have a growing number of cusps. It would be worth doing these computations in the future, and then comparing them to the results obtained in this paper. 2 The Iwasawa decomposition of SLn (R) Let n ≥ 2, G = SLn (R) and Γ = SLn (Z). Consider the action of Γ by left translations on G and let K = SOn ; ( A= diag(a1 , . . . , an ); n Y ) ai = 1; ai > 0, for any i = 1, . . . , n ; i=1 N = {(nij )i,j ∈ G; nii = 1 and nij = 0 for i > j} . Lemma 2.1. The product map Φ : K × A × N −→ G 2 (k, a, n) 7→ kan is a homeomorphism. Proof. We can construct an inverse map for Φ by using the Gram-Schmidt orthonormalization process. Take g ∈ G and let x1 , . . . , xn be its columns. Then define inductively y1 , . . . , yn by x1 y1 = ; kx1 k i−1 yi = X yei , where yei = xi − hxi , yl iyl ; for i = 2, . . . , n. ke yi k l=1 Let e1 , . . . , en be the standard orthonormal basis of Rn . Then there exists an unique k ∈ SOn such that k(yi ) = ei , for any i = 1, . . . n. Therefore k(e yi ) = k(ke yi k yi ) = ke yi k k(yi ) = ke yi k ei , for any i = 1, . . . , n. So there is a diagonal matrix a = diag(ke y1 k , . . . , ke yn k), such that k(e yi ) = a(ei ), for any i = 1, . . . , n. Also, it is easy to see that yi ∈ hx1 , . . . , xi i , for any i = 1, . . . , n. Thus we have: g −1 yei = g −1 (xi − i−1 X hxi , yl iyl ) ∈ g −1 xi + g −1 hx1 , . . . , xi−1 i l=1 ⇒ g −1 yei ∈ ei + he1 , . . . , ei−1 i . From this, we conclude that there exists u ∈ N such that g −1 yei = uei , for every i = 1, . . . , n. Therefore, u−1 g −1 yei = ei = a−1 k(e yi ), for any i ⇒ u−1 g −1 = a−1 k ⇒ g = k −1 au−1 . It is easy to see now that det(a) = 1, so a ∈ A and thus we can define a continuous inverse map g ∈ G 7→ (k −1 , a, u−1 ) ∈ K × A × N . The previous lemma gives us the Iwasawa decomposition G = KAN of SLn (R). Note that K ∩ A = K ∩ N = A ∩ N = {I} and that for this Iwasawa decomposition, AN = N A and K(AN ) = (AN )K (see [7], page 148). 3 Haar measure on SLn (R) Given a locally compact Hausdorff topological group G, a left invariant Haar measure on G is, by definition, a regular Borel measure µ on G such that for all g ∈ G and all Borel sets E ⊂ G we have µ(gE) = µ(E). It is well known that every connected Lie group admits such a Haar measure. Moreover, it is unique up to scalar multiples. We can define analogously right-invariant Haar measures. See [13] for more results about Haar measures on Lie groups. Since G = SLn (R) is unimodular, i.e. the left and right invariant Haar measures coincide, and dg is invariant under left translation by elements of K and under right translation by elements of AN , we get that the Haar measure 3 of G in k, v, a coordinates is given by the product measure dg = dk du da, where da, du and dk are the Haar measures on A, N and K, repectively. This means that for every compactly supported and continuous function f on G, we have Z Z Z Z f (g)dg = f (kau)du da dk. G K A N It canY be proved by induction on n that the Haar measure on N is given by du = duij . It is usually convenient to change the order of integration on i<j the variables a and u, and to this end we can change the coordinates from u to v = aua−1 . Then v is also an upper triangular unipotent matrix of the form X ai uij Eij . v = Id + aj i<j≤n It is easily seen that dv = Y dvij = i<j Z Y ai duij . This gives us a i<j j Z Z Z f (g)dg = G Z Z Z f (kva)dv da dk = K A N f (kau) K N A Y ai da du dk. a i<j j Also for convenience, we change coordinates from au to k −1 auk in the last integral. This has Jacobian equal to 1 (for each k ∈ K), so we get: Z Z Z Z Y ai f (g)dg = f (auk) dk da du. a G N A K i<j j In this work, we will consider the Haar measure in K to be the following: it is easy to see that the isotropy group of en = (0, . . . , 0, 1) by the action of SOn in Sn−1 is isomorphic to SOn−1 . Then Sn−1 ∼ = SOn−1 \SOn . We have that the natural map π : SOn → Sn−1 is a Riemannian submersion if we rescale it by a factor of √12 . Thus 1 vol(SOn ) = 2 2 (n−1) vol(Sn−1 ) · vol(SOn−1 ). By using induction and the fact that vol(Sn−1 ) = vol(SOn ) = 2 1 4 n(n−1) n−1 vol(S n−2 ) · vol(S 1 n 2π 2 , Γ( n 2) ) . . . vol(S ) = 2 we obtain (n−1)( n 4 +1) It remains to define a Haar measure on A. We claim that da = i n Y π2 . Γ( 2i ) i=2 n−1 Y i=1 dai is ai such a measure. Indeed, let φ : A → Rn−1 be the map   a1 0 . . . 0 0  0 a2 · · ·  0 0     .. .. . . .. ..  .  . . . .  7→ (t1 , . . . , tn−1 ) = (log a1 , . . . , log an−1 ). a=  0 0 · · · an−1  0   n−1   Y   0 0 ··· 0 a−1 i i=1 4 As φ is a group isomorphism and Haar measure is preserved by isomorphisms n−1 n−1 Y Y dai we get that da = dti = is a Haar measure on A. ai i=1 i=1 Siegel Sets for SLn (R) 4 Definition 4.1. Let Γ be some group acting properly discontinuously on a topological space X. We call F ⊂ X a coarse fundamental domain for Γ if: • ΓF = X; • {γ ∈ Γ; γF ∩ F = 6 ∅} is finite. Definition 4.2. A Siegel set in SLn (R) is a set Σt,λ of the form Σt,λ = At Nλ K, where t, λ are positive real numbers,   ai ≤ t, for any i = 1, . . . , n At = a ∈ A; ai+1 and Nλ = {u ∈ N | |uij | ≤ λ, for any i, j = 1, . . . , n} . For certain parameters t, λ the Siegel sets Σt,λ are coarse fundamental domains for SLn (Z). Another important property is that they have finite volume. Siegel sets can be also defined in a more general way for lattices in other semisimple Lie groups, as it can be seen in Chapter 19 of Morris [7]. In many cases, a finite union of copies of Siegel sets glue together to form coarse fundamental domains for general lattices. In this section we compute the volumes of the Siegel sets in SLn (R). We will use the Haar measure on G given in Section 3 in v, a, k coordinates. Theorem 4.1. n(n2 −1) n(n−1) 1 t 6 vol(Σt,λ ) = vol(SOn )(2λ) 2 . 2 ((n − 1)!)2 Proof. Z Z vol(Σt,λ ) = |uij |≤λ ai ai+1 = vol(K)(2λ) Z Y n−1 Y dai ai dk ai ≤t K i<j aj i=1 n(n−1) 2 Z ai ai+1 Y Y ai n−1 Y dai . ≤t i<j aj i=1 ai To compute the integral over a1 , . . . , an (with the condition change variables from a1 , . . . , an to the variables bi = ai , for any i = 1, . . . , n − 1. ai+1 5 duij . 1≤i<j≤n n Y i=1 ai = 1), we By elementary computation, we get n−1 Y ai Y i(n−i) = bi . a i<j j i=1 Moreover, as ai = bi ai+1 , the Jacobian of the change of coordinates from ai to bi is 2a11 . The integral then becomes n−1 Y Z bi ≤t i=1 1 = 2 Z n−1 Y bi ≤t i=1 i(n−i) bi 1 1 Y dbi b1 a2 b2 a3 . . . bn−1 an 2a1 i≤n−1 2 [i(n−i)−1] bi Y i≤n−1 n(n −1) 2 n−1 1 t 6 1 Y tni−i . = dbi = 2 i=1 (ni − i2 ) 2 ((n − 1)!)2 Thus we get to n(n2 −1) n(n−1) 1 t 6 vol(Σt,λ ) = vol(SOn )(2λ) 2 . 2 ((n − 1)!)2 (1) Borel proves in [1] the following theorem: Theorem 4.2. For t ≥ √23 and λ ≥ 12 , one has Σt,λ Γ = G. Moreover, Σt,λ is a coarse fundamental domain for Γ in G. Corollary 4.1. The quotient Γ\G has finite volume, which satisfies 3 vol(Γ\G) ≺ ecn , as n → ∞, (2) for some positive constant c. Proof. It is clear that vol(Γ\G) < ∞, since Σt,λ has finite volume and it contains a fundamental domain for SLn (Z) if t ≥ √23 and λ ≥ 12 . Thus vol(Γ\G) ≤ vol(Σt,λ ) for these values of t and λ. By taking λ = 21 and t = √23 in formula (1), we get that n(n2 −1) vol(Σ √2 ,1 3 2 )=2 (n−1)( n 4 +1)−1 2 = 3 2 i n Y π 2  ( √3 ) 6 Γ( 2i ) ((n − 1)!)2 i=2 2n3 +3n2 +7n−24 12 n(n2 −1) 12 ((n − π n2 +n−2 2 1)!)2 n Y i Γ( ) 2 i=2 Using Stirling’s formula, this volume is easily seen to grow assymptotically like 3 ecn , for some positive constant c and this finishes the proof. On the other hand, as we will see in the next section, vol(SLn (Z)\SLn (R)) computed with respect to the same normalization of the Haar measure goes to zero as n grows. 6 5 Volume of SLn (Z)\SLn (R) It is a well-known fact that vol(SLn (Z)\SLn (R)) is finite. Our goal is to calculate it, with respect to the same normalization of the Haar measure used in the previous section. The whole computation follows the original approach of Siegel [11], but we have to be careful with the normalization constants. We use Poisson summation, induction and the previously known fact that √ vol(SL2 (Z)\SL2 (R)) = 2ζ(2), which can be proved in a similar way (see [4], being careful with respect to the different normalization of vol(SO2 ) we are considering). We will state first the Poisson Summation Formula, which will play a fundamental role in the computations, and for which the reader can refer to [12]. Given a lattice Λ in Rn , we define |Λ| to be the covolume of Λ, i.e. the volume of Rn /Λ and the dual lattice of Λ by Λ∗ = {y ∈ Rn ; hx, yi ∈ Z for any x ∈ Λ} . Theorem 5.1 (Poisson Summation Formula). Given any lattice Λ in Rn , a vector w ∈ Rn and an adimissible function f : Rn → R in L1 , we have X f (x + w) = t∈Λ x∈Λ Here, fˆ(t) = 1 X −2πihw,ti ˆ e f (t), |Λ| ∗ f (x)e2πihx,ti dx is the Fourrier transform of f and admissibilty of f means that there exist constants , δ > 0 such that |f (x)| and fˆ(x) R Rn are bounded above by (1 + |x|)−n−δ . Let then f ∈ L1 be an admissible function on Rn . We can ask f to be a C ∞ function with compact support. We then define F : G 7→ R by X F (g) = f (vg). v∈Zn Here we are considering the multiplication of line-vectors v ∈ Rn by elements of G by the right. Clearly, F is left Γ-invariant, as Zn Γ = Zn under the action of SLn (Z) on Rn by right multiplication of line vectors by the inverse elements of SLn (Z). R Consider Γ\G F (g)dg. We will use this integral to calculate vol(Γ\G). Let    h v Q = stabG (e) = ; h ∈ SLn−1 (R), v ∈ Rn−1 , 0 1 where e = (0, . . . , 0, 1) ∈ Rn , and write QZ = Q ∩ Γ. Using linear algebra over Z, note that [ [ Zn − {0} = `eγ, `>0 γ∈QZ \Γ where ` runs over positive integers. Then we can write Z Z Z F (g)dg = f (0)dg + Γ\G Γ\G X X Γ\G `>0 γ∈Q \Γ Z 7 f (`eγg)dg = vol(Γ\G)f (0) + XZ `>0 f (`eg)dg. QZ \G For the second equality note that a fundamental domain for QZ in G is the union of images of a fundamental domain for Γ in G by representatives of classes in QZ \Γ. In addition, the Schwartz condition on f ensures that the integral over QZ \G is finite. Indeed, in his article [10], Siegel proves that the function F (g) is integrable over a fundamental domain for Γ\G. On pages 344345 of [loc. cit.] we can see that the integrals over QZ \G are also convergent (for any fixed l ∈ N). We observe that although he uses different decomposition of G and normalization of Haar measures, this does not change the finiteness of the integrals. Write ) ! ( h ∗ n−1 ; h ∈ GLn−1 (R), det(h) > 0 and ∗ ∈ R ; P = 0 det1(h) 0  N = In−1 0 v 1  ;v ∈ R n−1  , NZ0 = N 0 ∩ Γ.  M=   h 0 ; h ∈ SLn−1 (R) , MZ = M ∩ Γ; 0 1  1   t n−1 In−1 0 A0 = ; t > 0 ; 0 t−1 Note that P = N 0 M A0 ⊃ N A, Q = N 0 M and G = N 0 M A0 K. However this time we have that N 0 M A0 intersects K non-trivially, i.e. this is not an Iwasawa decomposition. The product N 0 M A0 K projects on G with fiber SO(n − 1). Therefore we get the following Lemma 5.1. For every left G-invariant function Φ, we have Z Z 1 Φ(n0 ma0 k)dn0 dm da0 dk Φ(g)dg = vol(SOn−1 ) QZ \(N 0 M A0 K) QZ \G where dg is the Haar measure in G coming from its Iwasawa decomposition (as in Section 4). Here dn0 , dm, da0 and dk are the left Haar measures on N 0 , M , A0 and K, n−1 Y respectively. We see that dn0 = vi and that M is isomorphic to SLn−1 (R) i=1 and thus dm will appear as the measure of this group. This allows us to use induction in the calculations. On the other hand, A0 is isomorphic to R>0 via the isomorphism  1  0 t n−1 In−1 ∈ A0 7→ t ∈ R>0 . 0 t−1 Thus we have da0 = dt t where dt is the usual measure in R. Again it will be convenient to change the order of integration, by letting the variable a0 ∈ A0 to be the last one. This will give us d(a0 qa0−1 ) = tn dq, for q = 8    1 h v t n−1 In−1 0 0 n m ∈ N M . Indeed, for a = ∈ A and q = ∈ 0 1 0 t−1   n h t n−1 v , and thus the M -contribuction to the Q, we have a0 q(a0 )−1 = 0 1 n measure doesn’t change, but the N 0 -contribution is multiplied by (t n−1 )n−1 = tn and we get to d(a0 qa0−1 ) = tn dq as stated. R Then, if we require f to be K-invariant, the integral Γ\G F (g)dg becomes equal to 0 0 0 vol(Γ\G)f (0) +  X 1 vol(SOn−1 ) Z `>0 = vol(Γ\G)f (0) + f (`en0 ma0 k)dn0 dm da0 dk QZ \(N 0 M A0 K) vol(QZ \K) X vol(SOn−1 ) Z Z QZ `>0 \(N 0 M ) f (`en0 ma0 )tn da0 dn0 dm. A0 We have K ∩ QZ = SOn−1 (Z). Noting that SOn−1 (Z)\SOn , Sn−1 ∼ = SOn−1 \SOn ∼ = SOn−1 (Z)\SOn−1 we get vol(Sn−1 ) = vol(SOn−1 \SOn ) = vol(SOn−1 (Z)\SOn ) . vol(SOn−1 (Z)\SOn−1 ) As SOn (Z) acts properly and freely in SOn , for any n ∈ N we have that SOn −→ SOn (Z)\SOn is a finite covering with #SOn (Z) sheets, which gives us vol(SOn ) = #(SOn (Z))vol(SOn (Z)\SOn ). Altogether, we obtain: vol(QZ \K) = vol(SOn−1 (Z)\SOn ) = vol(Sn−1 )vol(SOn−1 ) . #(SOn−1 (Z)) As the integrand is invariant under N 0 M (en0 m = e, for any n0 ∈ N 0 and m ∈ M ) and the volume of NZ0 \N 0 is 1, this implies vol(Γ\G)f (0) + vol(QZ \K) X vol(SOn−1 ) `>0 Z Z QZ \(N 0 M ) f (`en0 ma0 )tn da0 dn0 dm A0 XZ vol(QZ \K) vol(SLn−1 (Z)\SLn−1 (R)) f (`ea0 )tn da0 vol(SOn−1 ) 0 A `>0 Z X vol(Sn−1 ) vol(SLn−1 (Z)\SLn−1 (R)) f (`ea0 )tn da0 = vol(Γ\G)f (0) + #(SOn−1 (Z)) A0 = vol(Γ\G)f (0) + `>0 By replacing a0 ∈ A0 by t ∈ R>0 and using the description of da0 , we get to XZ ∞ vol(Sn−1 ) dt vol(Γ\G)f (0) + vol(SLn−1 (Z)\SLn−1 (R)) f (`et)tn . #(SOn−1 (Z)) t 0 `>0 9 By replacing t by `t , we obtain X 1 vol(Sn−1 ) vol(Γ\G)f (0) + vol(SLn−1 (Z)\SLn−1 (R)) #(SOn−1 (Z)) `n `>0 ∞ Z f (et)tn 0 dt . t By using polar coordinates in Rn = {(v, t), v ∈ Sn−1 , t ∈ R>0 }, we get Z Z ∞ Z ∞ Z n−1 n dt n−1 vol(S = ) f (et)t f (v, t)t dtdv = f (x)dx = fˆ(0). t Sn−1 0 Rn 0 Thus what we get until now is the following Proposition 5.1. The initial integral becomes Z vol(SLn−1 (Z)\SLn−1 (R)) ζ(n)fˆ(0), F (g)dg = vol(Γ\G)f (0) + #(SOn−1 (Z)) Γ\G where ζ(n) = X 1 is the Riemman zeta function. ln l∈Z Corollary 5.1. The previous result allows us to compute explicitely the value of vol(Γ\G): vol(Γ\G) = n n−1 Y √ Y 1 vol(SLn−1 (Z)\SLn−1 (R)) ζ(n) = 2 ζ(i) . #(SOn−1 (Z)) #(SO i (Z)) i=2 i=1 Proof. For every g ∈ G, we are going to apply the Poisson summation formula to the lattice Λ = {vg; v ∈ Zn } in Rn , the vector w = 0 and the initial function f . Note that Λ∗ = {vg ∗ ; v ∈ Zn }, where g ∗ =>g −1 . Then we get X X F (g) = f (vg) = fˆ(vg ∗ ) = F̂ (g ∗ ), for any g ∈ G. v∈Zn v∈Zn The automorphism g 7→ g ∗ preserves the measure on G and stabilizes Γ, so we can do an analogous computation with the roles of f and fˆ reversed. Since R R ˆ fˆ(0) = f (0) and Γ\G F (g)dg = Γ\G F̂ (g)dg, we obtain vol(SLn−1 (Z)\SLn−1 (R)) vol(Γ\G)f (0) + ζ(n)fˆ(0) = #(SOn−1 (Z)) Z F (g)dg Γ\G Z = vol(SLn−1 (Z)\SLn−1 (R)) F̂ (g)dg = vol(Γ\G)fˆ(0) + ζ(n)f (0). #(SOn−1 (Z)) Γ\G By asking additionally that f is such that f (0) 6= fˆ(0) and using indution on n, we get to the desired result. We observe that for every i ∈ N, #(SOi (Z)) = 2i−1 i!. Indeed, the group SOi (Z) consists of monomial matrices whose nonzero entries are equal to ±1 and which have determinant equal to 1. The first condition gives us 2i i! matrices. Now if we look at the surjective group homomorphism det : B = {monomial matrices with nonzero entries ∈ {±1}} → {±1}, 10 we get B/Ker(det) ∼ = {±1}, which implies #(SOi (Z)) = #(Ker(det)) = #(B) 2i i! = = 2i−1 i!. 2 2 Thus we have proved the following Theorem 5.2. The explicit volume of Γ\G, by considering the Haar measures described in Section 3 is given by n Y n n−1 Y √ Y ζ(i) vol(SLn (Z)\SLn (R)) = 2 i=2 i=1 1 2i−1 i! ζ(i) i=2 = 2 n Y n2 −3n+1 2 . i! i=2 0 2 It is not difficult to see that this function goes to zero like e−c n as n grows, where c0 is a positive constant. It has a completely different behaviour from the volume growth of Siegel Sets described by formula (1). What we can conclude directly from all this is that although the geometry of a Siegel set is simpler than that of the actual fundamental domain for a lattice, their volumes can differ dramatically as n grows. Thus we should be careful if we want to replace fundamental domains of any lattice by simpler structures such as Siegel sets, due to the possibility that some of their relevant geometric features, e.g. volume, may have different behavior to that of fundamental domains. As a consequence of Sections 4 and 5, we obtain: Corollary 5.2. The ratio between volumes of the minimal Siegel sets Σ = Σ 21 , √2 for SLn (Z) and the actual fundamental domains for these groups in 3 SLn (R) is given by 2 vol(Σ) C(n) = = vol(Γ\G) 3 2n3 +9n2 +25n−30 12 n3 −n 12 ((n − 1)!)2 π n2 +n−2 4 n Y n−1 Y i=1 n Y i! i Γ( ) ζ(i) 2 i=2 i=2 . 3 Moreover, C(n) ∼ ec̃n for some constant c̃ that does not depend on n. A natural question arising here is the following: “How is our normalization of the Haar measure related to the canonical normalization defined by using the Killing form on sln (R)?” To answer to this question we can compare our formula with a result of Harder [6], who computed the volume of SLn (Z)\X, where X is the symmetric space SLn (R)/SOn . In order to do this comparison, note that by equation (5.2) we have n Y 1 Y √ n−1 2 ζ(i) 2i−1 i! i=2 vol(Γ\G) i=1 = . (3) vol(SLn (Z)\X) = i n Y vol(SOn ) π2 n (n−1)( +1) 4 2 Γ( 2i ) i=2 11 By Harder’s formula, we obtain that this volume in the canonical normalization is given by n−1 Y vol1 (SLn (Z)\X) = i! n Y i=1 i=2 (2π) n(n+3) 2 ζ(i) 2τ n! , (4) where τ = n if n is odd and τ = n − 1 if n is even. We see that these volumes differ by a factor given by 2 vol1 (SLn (Z)\X) = C1 (n) = vol(SLn (Z)\X) n2 −5n−2 −τ 4 n−1 Y 2 i! i=1 n!π n2 +5n+2 4 n  i , Y Γ 2 i=2 (5) where τ = n if n is odd and τ = n − 1 if n is even. We note that again by using Stirling’s formulas, we obtain that C1 (n) grows assymptotically with 2 n like eκn , for some positive constant κ. The same renormalization can be applied to (1) in order to obtain the volumes of Siegel sets in the symmetric spaces with respect to the standard normalization of the measure. 6 Bounding the number of intersecting domains Another relevant consequence of this work is the following corollary: Corollary 6.1. Let N be the cardinality of the set I := {γ ∈ Γ; γΣ ∩ Σ 6= ∅}, vol(Σ) where Σ = Σ 21 , √2 . Then N ≥ C(n) = vol(Γ\G) . 3 Proof. As Σ is a S Siegel set, it must contain a fundamental domain F for Γ. We affirm that Σ ⊂ γF. γ∈I Indeed, given x ∈ Σ, if x ∈ F, there is nothing to prove. If x ∈ / F, as the images of F tesselate SLn (R) we must have x ∈ γF, for some Id 6= γ ∈ Γ. As γF ⊂ γΣ, we obtain x ∈ γΣ ∩ Σ, and thus γ ∈ I. Therefore the inclusion above is true. From this we obtain N vol(Γ\G) = N vol(F) ≥ vol(Σ) and thus N ≥ C(n), as stated. In his recent work [8], Martin Orr shows in a more general setting that given a reductive algebraic group G defined over Q, a general Siegel set Σ ⊂ G(R) for some arithmetic subgroup Γ ⊂ G(Q), and θ ∈ G(Q), there exists an upper bound for the height of elements γ ∈ Γ such that θΣ ∩ γΣ 6= ∅. The height of an element is defined by: H(γ) = max H(γij ), 1≤i,j≤n where given a rational number a/b, H(a/b) is defined as the maximum of the absolute values of a and b. Orr shows that, given any element γ of the set ΣN,D := ΣΣ−1 ∩ {γ ∈ G(Q), detγ ≤ N and the denominators of γ are ≤ D} , 12 there exists some constant C1 , depending on the group G, on the Siegel set Σ and on the way the group G is embedded in some GLn (R), such that 2 H(γ) ≤ C1 N n Dn , where N = |detγ| and D is the maximum of the denominators of entries of γ. Note that for Γ = SLn (Z), the set I defined above is contained in ΣN,D . In this section we are going to compare this result with ours, i.e., to see what happens in the case when G = SLn (R) and Γ = SLn (Z). Note that in this case, for any γ ∈ Γ, we have N = |detγ| = 1 and also D = 1 because the entries of γ are all integers. Thus Orr’s result gives us, for this case, H(γ) ≤ C1 (n). By the definition, the height of an element γ ∈ SLn (Z) is equal to |γ|max . Therefore, his result turns to |γ|max ≤ C1 (n), for any γ ∈ ΣΣ−1 . By Example 1.6 on page 5 of [3], the set {γ ∈ SLn (Z); kγk ≤ C1 (n)} has 2 cardinality of assymptotic order cn C1 (n)(n −n) , with cn → 0 as n → ∞. Thus if we assume that n is sufficiently large, we can suppose that cn <  for some  > 0 fixed. Therefore, we have 2 |{γ ∈ SLn (Z); kγk ≤ C1 (n)}| ≺ C1 (n)(n −n) , where the notation f (n) ≺ g(n) used above means that there exists a positive constant C such that for sufficiently big n, we have f (n) ≤ Cg(n). Note that the result in [3] is proved for the Euclidean norm k.k in Mn×n and we know that kγk ≤ n |γ|max . Thus 2 |{γ ∈ SLn (Z); |γ|max ≤ C1 (n)}| ≺ (nC1 (n))(n −n) . We are going to show that C1 (n) ≤ e From this we obtain that |I| ≺ e n4 2 ln(n) n2 −n ln(n) 2 . . Hence we have: Corollary 6.2. For Γ = SLn (Z) in SLn (R) and I defined above, there exist constants c1 , c2 > 0 such that 3 ec1 n ≤ |I| ≤ ec2 n 4 ln(n) . In order to obtain the second inequality we adapt the proofs in [8] for the SLn (R) case, with the difference that we give explicit values for the constants. Definition 6.1. Let γ ∈ I. From this element, we can define: • A partition of {1, . . . , n} (with respect to γ) is a list of disjoint subintervals of {1, . . . , n}, which we call components, whose union is all of {1, . . . , n} and such that: 13 – γ is block upper triangular with respect to the chosen partition; – γ is not block upper triangular with respect to any other finer partition of {1, . . . , n}; 2 • A leading entry of γ is a pair (i, j) ∈ {1, . . . , n} such that γij is the leftmost non-zero entry of the i-th row of γ. For a concrete description of what are the possible partitions in the GL3 case see Section 3.2 of [8]. We will make use of the following lemma whose proof can be found in [8]: Lemma 6.1. If i, j are in the same component, then there exists a sequence of indices i1 , . . . , is such that i1 = i, is = j and (∗) For every p ≤ s − 1, either ip ≤ ip+1 or (ip , ip + 1) is a leading entry. In the proof of the following lemmas for the GLn case, Martin Orr uses the notation A  B meaning that there exists a constant C, depending on n, such that |A| ≤ C |B|. Our point here is to compute such constants so that we can make explicit the value of C1 (n). √ Lemma 6.2. If (i, j) is a leading entry of γ, then αj ≤ nβi . Proof. For any γ ∈ ΣΣ−1 , we can write γ = νβκα−1 µ−1 , with κ ∈ SOn , ν, µ ∈ N 12 and α, β ∈ A √2 . This gives us the equation γµα = νβκ. We will 3 compare the lengths of the i-th rows on each side of this equation. As κ ∈ SOn , multiplying by κ on the right does not change the length of each row. If we expand out lengths we obtain n X n n  X X 2 2 γiq µqp αp2 = νip βp . p=1 q=1 p=1 As ν is upper triangular, the non-zero terms on the right hand side of the last equation must have p ≥ i. By the definition of At , for all p ≥ i we have βp ≤ 1 βi ≤ βi , t(p−i) where in the second inequality we used that t = Since ν ∈ N 21 , |νip | ≤ 1 for any i, p. Alltogether, n X 2 2 νip βp ≤ p=1 n X p=1 2 2 νip βp ≤ n X √2 3 and p ≥ i imply 1 t(p−i) ≤ 1. βp2 ≤ (n − i)βi2 ≤ nβi2 . p≥i On the other hand, by looking at the left hand side of the equation, we obtain: n n X n X   X 2 γiq µqj αj ≤ γiq µqp αp2 . q=1 p=1 q=1 As (i, j) is a leading entry, we can only have γiq 6= 0 if q ≥ j. But as µ is upper triangular, µqj 6= 0 implies q ≤ j. Thus the only non-zero term in the first sum is the one for q = j and then we get n X  2 2 2 2 2 γiq µqj αj2 = γij µjj αj = γij αj . q=1 14 Note that γij 6= 0 and that as γ has integer entries, we must have |γij | ≥ 1, 2 which implies γij ≥ 1. Altogether, we obtain √ 2 αj2 ≤ αj2 γij ≤ nβi2 ⇒ αj ≤ nβi , from what we conclude the proof. Lemma 6.3. For all k ∈ {1, . . . , n}, αk ≤ √ nβk . Proof. We affirm that there must exist a leading entry (i, j) such that j ≤ k ≤ i. To prove this notice that as γ is invertible, there must exist i ≥ k such that the i-th row of γ contains a non-zero entry in the k-th column or to its left (otherwise the leftmost k columns of γ would have rank less than k). Choose j so that γij is the leading entry of γ in the i-th line and it will satisfy j ≤ k as claimed. By Lemma 6.2 and by the definition of At we obtain √ √ √ 1 1 αk ≤ (k−j) αj ≤ nβi ≤ n (i−k) βk ≤ nβk . t t √ Lemma 6.4. For all j ∈ {1, . . . , n}, βj ≤ ( n)n−1 αj . Proof. As α and β are diagonal with positive real entries, we have (by using Lemma 6.3 in the inequality) βj det(α) = βj n Y Y √ √ αk ≤ βj αj ( n)n−1 βk = ( n)n−1 αj det(β). k=1 But as det(β) = det(α) = 1, k6=j √ βj ≤ ( n)n−1 αj and the lemma is proved. √ 2 Lemma 6.5. If i and j are in the same component, βj ≤ ( n)n −n αi . Proof. We can apply Lemma 6.1 to obtain a sequence i1 = i, . . . , is = j such that for any p ∈ {1, . . . , s} , we have either ip ≤ ip+1 or (ip , ip+1 ) is a leading entry. We take this subsequence as the √ smallest possible. If ip ≤ ip+1 then as α ∈ At and ( n)n ≥ 1, we get √ αip ≥ tip+1 −ip ≥ 1 ⇒ αip+1 ≤ αip ≤ ( n)n αip . αip+1 On the other hand if (ip , ip+1 ) is a leading entry then by Lemmas 6.2 and 6.4 we have √ √ αip+1 ≤ nβip ≤ ( n)n αip . If we apply the last inequality successively we get to √ αj = αis ≤ ( n)n(s−1) αi . Now we just apply Lemma 6.4 and notice that s ≤ n to obtain √ √ 2 βj ≤ ( n))n−1 αj ≤ ( n)n −1 αi . 15 We write Q = {g ∈ G; g is block upper triangular according to the components of γ} ; L = {g ∈ G; g is block diagonal according to the components of γ} . We affirm that κ ∈ L. Indeed, as the matrices γ, µ, α, β and ν are in Q by the construction, we also have κ ∈ Q. On the other hand, if a matrix is block upper triangular and is also orthogonal, then it is block diagonal. Thus κ ∈ L. Lemma 6.6. If i, j ∈ {1, . . . , n}, then |γij | ≤ C1 (n) = n n2 −n 2 . Proof. Write γ = νβκα−1 µ−1 . Because α, β are diagonal, the pq-th entry of βκα−1 is βp κpq αq−1 . If p and q are not in the same component, as κ ∈ L, we get that κpq = 0. On the other hand, if they are in the same component, then by Lemma 6.5 √ 2 βp κpq αq−1 ≤ κpq ( n)n −1 . By the definition of SOn , |κ|max ≤ 1 for every κ ∈ SOn . Therefore √ 2 βp κpq αq−1 ≤ ( n)n −1 . As we have µ, ν ∈ N 12 , we have |µ|∞ , |ν|∞ ≤ 1. Altogether, we obtain √ 2 |γij | ≤ ( n)n −1 . Therefore we conclude the proof that H(γ) ≤ C1 (n), where √ 2 n2 −1 C1 (n) = ( n)n −1 = e 2 ln(n) and this finishes the proof of Corollary 6.2. Acknowledgements I would like to thank Professor Mikhail Belolipetsky for several suggestions on the development of this paper and also on the text. I also thank Paul Garret and Martin Orr for their very helpful works and for always answering my emails with good suggestions, and Cayo Dória for helping me to understand better some topics. Finally, I also thank the refferee for carefully reading the paper and for giving suggestions that improved the presentation of the results. References [1] Borel, A.: Introduction aux Groupes Arithmètiques. Hermann, Paris (1969). [2] Borel, A. and Harish-Chandra.: Arithmetic subgroups of algebraic groups. Ann. of Math. 75, 485–535 (1962). 16 [3] Duke, W.; Rudinick, Z. and Sarnak, P.: Density of Integer Points on Affine Homogeneous Varieties. Duke Math. J., Vol.71, No.1, 143-179 (1993). [4] Garret, P.: Volume of SLn (Z)\SLn (R) and Spn (Z)\Spn (R). Paul Garrett’s homepage http://www-users.math.umn.edu/∼garrett/m/v/volumes.pdf (2014). Accessed 07 November 2017. [5] Habegger, P. and Pila, J.: Some unlikely intersections beyond André–Oort. Compos. Math. 148, 1-27 (2012). [6] Harder, G.: A Gauss-Bonnet formula for discrete arithmetically defined groups. Ann. Sci. Ec. Norm. Supér. 4e série, tome 4, n° 3, 409-455 (1971). [7] Morris, D. W.: Introduction to Arithmetic Groups. Deductive Press, United States (2015). [8] Orr, M.: Height bounds arXiv:1609.01315v3 (2016). and the Siegel property. Preprint, [9] Siegel, C.L.: Einführung in die Theorie der Modulfunktionen n-ten Grades. Math. Ann 116, 617–657 (1939). [10] Siegel, C.L.: A Mean Value Theorem in Geometry of Numbers. Annals of Mathematics Vol 45, No 2 (1945). [11] Siegel, C.L.: Lectures on the Geometry of Numbers. Springer-Verlag, Berlin (1989). [12] Stein, E. M. and Weiss, G. L.: Introduction to Fourier Analysis on Euclidean Spaces. Princeton Math. Ser. 32, Princeton Univ. Press, Princeton, NJ (1971). [13] Venkataramana, T.N.: Lattices in Lie Groups. In Workshop on Geometric Group Theory, India (2010). [14] Young, R.: The Dehn function of SLn (Z). Ann. of Math. 177, 969–1027 (2013). 17 

4math.GR

1 Degrees-of-Freedom of the MIMO Three-Way Channel with Node-Intermittency Joachim Neu, Anas Chaaban, Aydin Sezgin, Mohamed-Slim Alouini Abstract arXiv:1708.08161v1 [cs.IT] 28 Aug 2017 The characterization of fundamental performance bounds of many-to-many communication systems in which participating nodes are active in an intermittent way is one of the major challenges in communication theory. In order to address this issue, we introduce the multiple-input multiple-output (MIMO) three-way channel (3WC) with an intermittent node and study its degrees-of-freedom (DoF) region and sum-DoF. We devise a non-adaptive encoding scheme based on zero-forcing, interference alignment and erasure coding, and show its DoF region (and thus sum-DoF) optimality for non-intermittent 3WCs and its sum-DoF optimality for (node-)intermittent 3WCs. However, we show by example that in general some DoF tuples in the intermittent 3WC can only be achieved by adaptive schemes, such as multi-hop or decode-forward relaying. This shows that non-adaptive encoding is sufficient for the non-intermittent 3WC and for the sum-DoF of intermittent 3WCs, but adaptive encoding is necessary for the DoF region of intermittent 3WCs. Our work contributes to a better understanding of the fundamental limits of multi-way communication systems with intermittency and the impact of adaptation therein. Index Terms Degrees-of-freedom, three-way channel, intermittency, multiple-input multiple-output (MIMO), interference alignment, zero forcing, relaying. I. I NTRODUCTION In multi-way communication scenarios multiple nodes communicate with each other, each acting as a source, a destination, and possibly a relay at the same time. This mode of communication is especially important for future systems employing full-duplex and device-to-device (D2D) communication, [2]–[4]. It is an important technique for efficient resource utilization that is expected to gain more prominence in future communication systems, especially with the rise of mesh networks, e.g., in industrial and vehicular networks. Parts of the results of this paper were presented at ISIT 2017 [1]. The work of J. Neu, a student at Technische Universität München (TUM), 80333 München, Germany, was conducted during a research internship at King Abdullah University of Science and Technology (KAUST). Email: joachim.neu@tum.de. A. Chaaban and M.-S. Alouini are with the Division of Computer, Electrical, and Mathematical Sciences and Engineering (CEMSE) at King Abdullah University of Science and Technology (KAUST), Thuwal, Saudi Arabia. Email: {anas.chaaban,slim.alouini}@kaust.edu.sa. A. Sezgin is with the Institute of Digital Communication Systems at the Ruhr-Universität Bochum (RUB), 44780 Bochum, Germany. Email: aydin.sezgin@rub.de. This work was supported in part by the German Research Foundation (DFG), under grant SE 1697/16. 2 1 1 2 1 1 1 τ 3 2 (a) no intermittency τ 1 3 (b) node-intermittency Fig. 1: MIMO 3WC with no intermittency such that all nodes are always connected (a) and node-intermittency where node 1 is available only τ fraction of the time (b) Several multi-way communication scenarios have been studied in the literature, [5]. It started with the twoway channel which was first studied by Shannon in [6] and subsequently in [7]–[10] for instance. Then, the scope was extended to two-way networks where two groups of nodes communicate with each other in a two-way fashion [11]–[13], and two-way relay networks where two nodes communicate with each other in a two-way fashion via a relay node [14]–[17]. This line of research has been further extended to multi-way networks where multiple nodes communicate with each other in a multi-way fashion, each node being a source and a destination at the same time [18]–[22], and to multi-way relay networks where the same communication as in multi-way networks takes place via a relay node [23]–[34]. A. The Three-Way Channel with Intermittency In this work, we focus on the multiple-input multiple-output (MIMO) three-way channel (3WC) which can be described as follows: Consider a system consisting of three terminals communicating with each other in a multi-way fashion, e.g., two D2D user terminals and a base station (BS) where the D2D users communicate with each other while exchanging signals with the BS (control signals or data). This 3WC is an extension of Shannon’s two-way channel [6] and has been studied in [20]–[22]. Therein, it is assumed that the three nodes are connected all the time (Fig. 1a). In practice however, connectivity might be intermittent. For instance, a pair of D2D users, commonly chosen to be nearby users [3], [4], might be both disconnected from the BS due to shadowing (Fig. 1b). Future generations of mobile communication systems (e.g., using mmWaves, where line-of-sight (LOS) propagation will become dominant) are expected to be heavily susceptible to this type of shadowing. In another scenario, the D2D users might operate in an underlay mode over a resource block used by the BS to communicate to cellular users (CUs). The BS connects to the D2D users whenever it does not communicate with a CU, and disconnects from the D2D users otherwise (see Fig. 2). Note that D2D users are chosen so that they cause/receive negligible interference (relative to the desired signals) to/from other nodes using the same resource block [35], [36]. In both cases the two links from the D2D users to the BS are jointly intermittent, i.e., both are available or blocked at the same time. We call this node intermittency and call the BS an intermittent node (from the D2D users’ perspective). Motivated by these scenarios, we study the 3WC with this type of intermittency. The impact of intermittency in various networks was studied in [37]–[40] for instance. 3 D2D ere interf gible negli BS D2D BS nce D2D (a) BS serving D2D users CU D2D (b) BS serving CU Fig. 2: A D2D pair sharing the same resources with a cellular user (CU), where the BS communicates with the D2D pair part of the time (a) and with the CU the rest of the time (b). The D2D pair is far enough from both the BS and the CU and thus causes/receives negligible interference (dotted). B. Degrees-of-Freedom and Intermittency In many previously mentioned works, the focus is generally on the capacity of the studied network. Since finding the capacity is elusive in most cases, some works (and also this work) focus on the degrees-offreedom (DoF) which provide a good capacity approximation at high signal-to-noise ratio (SNR) [41], thereby highlighting the interaction between the signals of the different nodes while diminishing the impact of noise. This is of interest since state-of-the-art wireless communication systems operate in a regime where they are essentially interference-limited rather than noise-limited. The concept of intermittency as a form of channel impairment also fits well with the philosophy of DoFs: With increasing SNR, signal components ‘harden’ in the sense that some allow (almost) noise- and interferencefree communication at a rate that scales with SNR on a logarithmic scale, while others are hopelessly burried in uncancelable interference and are therefore useless. This effect is exactly what the DoF perspective captures as it essentially counts available Shannon-Hartley communication ‘units’. Intermittency is the channel impairment that goes well together with this ‘all or nothing’ perspective, capturing the notion that some signal dimensions might become useless due to stochastic processes in the channel, such as shadowing of dominant LOS components. C. Scope of this Work Here, we study the impact of node-intermittency on the DoF region and sum-DoF, i.e., the capacity scaling versus SNR in a dB scale, of a full-duplex MIMO 3WC. In this MIMO network, each node generally has two independent messages, each intended for one of the remaining two nodes. We pay particular attention to the necessity (or the lack thereof) of adaptive encoding for reaching DoF region/sum-DoF. Adaptive encoding enables cooperative communication schemes by allowing the transmit signal of a node to depend on its previously received signals, which can be interpreted as a form of feedback. In contrast, with non-adaptive encoding transmit signals depend only on the messages to be sent, and cooperation (e.g., in the form of 4 relaying) is excluded. The issue of (non-)adaptive encoding has been studied for other networks earlier in [9]–[11] for instance. We assume that nodes have causal knowledge of the intermittency state of adjacent links. This can be obtained by estimating the connectivity from the receive signals (e.g., signal strength). Note that in some cases the intermittency state can be known without the need for estimation, occasionally even ahead of time, e.g., in the scenario of D2D/CU/BS users, where the BS knows its scheduling of users in advance and can anticipate when it will not be able to receive the D2D user signals. D. Outline and Overview of Results After introducing the details of the system model in Section II and highlighting the main results in Section III, we examine the (node-)intermittent 3WC in Section IV. We devise a non-adaptive encoding scheme based on zero-forcing (ZF), interference alignment (IA) and erasure coding (EC) and derive its achievable DoF region and sum-DoF. We present the genie-aided converse techniques used to derive DoF region and sum-DoF upper bounds, both under non-adaptive and adaptive encoding. We conclude that for the intermittent 3WC the presented non-adaptive scheme is sum-DoF optimal, so adaptation is not necessary to achieve sum-DoF. Then, we provide examples of adaptive relaying schemes that can achieve a DoF region point that no non-adaptive scheme can achieve. This shows that adaptive schemes can achieve strictly larger DoF regions, and therefore adaptation is required to achieve the DoF region of intermittent 3WCs. To complete the picture, we examine the nonintermittent 3WC as special case of intermittent 3WCs in Section V. We show that for the non-intermittent 3WC the presented scheme is DoF region optimal (and thus also sum-DoF optimal) and therefore adaptive encoding is not required. This reveals an interesting interplay between intermittency and adaptation. In Section VI we provide conclusive remarks and directions for future research. E. Notation Throughout the paper, we use xni := (xi,1 , . . . , xi,n ) for some index i, and xni,` := (xi,` , . . . , xi,n ). We use regular letters to denote scalar-valued quantities, boldface letters to denote vector- und matrix-valued quantities; lowercase letters for scalar and vector values (e.g., realizations of random variables), uppercase letters for matrix values and for random variables. The N × N identity matrix is denoted IN , the N × N all-zero matrix is 0N . We write X ∼ CN (0, Q) to indicate that X is a multivariate complex Gaussian random variable with zero mean and covariance matrix Q, and S ∼ Bern(τ ) to indicate that S is Bernoulli distributed with Pr[S = 1] = τ and Pr[S = 0] = 1 − τ =: τ . We write x+ to denote max{0, x} for some x ∈ R, and H † , H T , H H , and span(H) to denote the Moore-Penrose pseudo-inverse, the transpose, the Hermitian transpose, and the subspace spanned by the columns of the matrix H. By log(x) we denote the logarithm of x to base 2, by i → j the communication from node i to node j one-way, and by i ↔ j both i → j and j → i. By pX (x) we denote the probability density function (PDF) of random variable X. II. T HE MIMO T HREE -WAY C HANNEL WITH N ODE -I NTERMITTENCY Throughout this section we assume i, j, k ∈ {1, 2, 3} and mutually distinct, and n is the number of channel accesses. The MIMO 3WC with node-intermittency is comprised of three terminals 1, 2 and 3 communicating 5 with each other in full-duplex mode over a shared medium. Each node i has two messages wij and wik (wi := (wij , wik )) to be delivered to the remaining nodes j and k, and two messages ŵji and ŵki to be decoded from the receive signals (Fig. 3). Each message wij is a realization of the random variable Wij . All random variables Wij are independent. In the (node-)intermittent 3WC one link is always available and two links are jointly intermittent with probability of being available τ (Fig. 1b and 3). Our objects under investigation (sum-DoF and DoF region of the intermittent 3WC – both are introduced in the sequel) depend on how the numbers of antennas at each node relate to each other, i.e., whether the intermittent node has most, second most, or least antennas. To reduce the number of cases one has to analyze, yet study the system without loss of generality (w.l.o.g.), two symmetries come to mind that can be exploited: Either a) fix a certain node to be intermittent (e.g., node 1 is intermittent), and investigate all possible relations among the numbers of antennas, or b) fix a relation between the numbers of antennas, and allow any one of the three nodes to be intermittent. The respective remaining cases follow by renaming. Preliminary work [1] fixed node 1 to be intermittent and further assumed a relation on the numbers of antennas, which is not w.l.o.g. We fix node 1 to be intermittent, but allow for any combination of numbers of antennas, which is approach a) and w.l.o.g. Node i is equipped with Mi antennas that are used for reception and transmission simultaneously. We assume channel accesses of the nodes are synchronized and time-discretized: At time instance ` the transmit signal xi,` ∈ CMi is a realization of a random vector Xi,` satisfying the power constraint n X E[kXi,` k22 ] ≤ nP. (1) `=1 The receive signals yi,` ∈ CMi are (Fig. 4)    y1,` 0M1       y  = s H  2,`   ` 12    y3,` s` H13 s` H21 s` H31 0M2 H32 H23 0M3     x1,` z1,`          x  + z  .   2,`   2,`      x3,` z3,` (2) Hij ∈ CMj ×Mi represents the channel matrix from node i to node j and is constant over time and known to all nodes in advance. The elements of these matrices are drawn independently from the same continuous distribution, such that rank(Hij ) = min(Mj , Mi ) almost surely. zi,` ∈ CMi is a realization of the noise process Zi,` ∼ CN (0, σ 2 IMi ) independent and identically distributed (i.i.d.) with respect to (w.r.t.) `, and s` ∈ {0, 1} is a realization of the intermittency state process S` . S` ∼ Bern(τ ) is assumed to be i.i.d. w.r.t. `. The state sequence sn is causally known at all nodes (i.e., in time instance ` all nodes know s`−1 ). All random variables Hij , Zin and S n are mutually independent. We denote ρ := P σ2 and call it SNR (signal-to-noise-power-ratio) throughout the paper. The messages Wij are each uniformly distributed over Wij = {1, . . . , |Wij (ρ)|}. Using an encoding function Ei,` node i constructs xi,` either from (wij , wik ) (non-adaptive encoding) or from (wij , wik , yi`−1 ) (adaptive encoding). After n transmissions (where n is the code length), node i decodes its desired messages using a decoding function Fi to obtain (ŵji , ŵki ) = Fi (sn , yin , wij , wik ). Transmission is considered successful if all messages are recovered successfully (wij = ŵij ), otherwise an error is reported. The average over all messages of the error probability is denoted by Pe,n . 6 W12 , W13 Ŵ21 , Ŵ31 Node 1 S Node 2 W21 , W23 Node 3 Ŵ12 , Ŵ32 W31 , W32 Ŵ13 , Ŵ23 Fig. 3: Every node i in the MIMO 3WC with node-intermittency sends messages wij and wik to nodes j and k, respectively (with i, j, k ∈ {1, 2, 3} mutually distinct). W.l.o.g. node 1 is assumed to be intermittent. M1 Y1 X1 H12 Z1 H13 S S Z3 H21 X2 Y3 H23 H31 M2 M3 H32 Y2 Z2 X3 Fig. 4: AWGN channel model of the shared medium in the MIMO 3WC with node-intermittency, with channel inputs Xi and outputs Yi , channel gains Hij , noise Zi and intermittency state S (time index ` omitted) A rate tuple R(ρ) := (R12 (ρ), R13 (ρ), R21 (ρ), R23 (ρ), R31 (ρ), R32 (ρ)) ∈ R6+ with Rij (ρ) = log(|Wij (ρ)|) n (3) is said to be achievable if there exists a sequence of encoder-decoder pairs for increasing code length n, where Pe,n → 0 as n → ∞. The capacity region C(ρ) is the set of all achievable rate tuples. The DoF region D is the set of achievable DoF tuples d := (d12 , d13 , d21 , d23 , d31 , d32 ) ∈ R6+ (4) 7 defined as in [42]. Roughly speaking, if a rate tuple R(ρ) as a function of ρ is achievable, i.e., R(ρ) ∈ C(ρ) for all ρ > 0, then the DoF tuple d with dij = lim supρ→∞ Rij (ρ) log(ρ) is achievable. We denote DoF regions with D and use suitable subscripts when further restricting assumptions apply (e.g., under non-adaptive encoding). We P define the corresponding sum-DoF as dsum = maxd∈D i,j∈{1,2,3},i6=j dij . The DoF perspective only captures rate contributions that are non-vanishing relative to log(ρ) as we let ρ → ∞. It neglects vanishing rate portions f (ρ) that grow sublinear in log(ρ) and are therefore o [log(ρ)], i.e., where limρ→∞ f (ρ) log(ρ) = 0. III. M AIN R ESULTS In this section, we summarize the main results of this paper. We denote the DoF region of the intermittent 3WC under adaptive encoding by DI , the DoF region of the intermittent 3WC under non-adaptive encoding I I , and the sum-DoF by dIsum . Obviously, DA ⊆ DI . We denote the DoF region of the non-intermittent by DA 3WC by DN and the sum-DoF by dN sum . In Section IV-A we devise a non-adaptive encoding scheme whose I achievable DoF region we denote by DIB,A . We start with this achievable DoF region given in the following theorem. I satisfying the Theorem 1 (DoF Region Inner Bound for Node-Intermittent 3WC). All DoF tuples d ∈ DIB,A following set of inequalities are achievable in the node-intermittent 3WC using non-adaptive encoding: max{d12 + d13 , d21 + d31 } ≤ τ M1 max{d21 + τ d23 , d12 + τ d32 } ≤ τ M2 max{d31 + τ d32 , d13 + τ d23 } ≤ τ M3 max{d12 + d13 + τ d23 , d21 + d31 + τ d32 } ≤ τ max {M1 , M3 } max{d12 + d13 + τ d32 , d21 + d31 + τ d23 } ≤ τ max {M1 , M2 } max{d12 + d31 + τ d32 , d21 + d13 + τ d23 } ≤ τ max {M2 , M3 } min{d12 , d13 , d21 , d23 , d31 , d32 } ≥ 0 I Therefore, DIB,A constitutes an inner bound on the DoF region of the node-intermittent 3WC, such that I I DIB,A ⊆ DA ⊆ DI . Proof. This theorem follows by the construction in Section IV-A. Subsequently we investigate whether this non-adaptive scheme is optimal. For the intermittent 3WC it turns out to be sum-DoF optimal, establishing the sum-DoF of the intermittent 3WC and the fact that adaptive encoding is not necessary to achieve it. Theorem 2 (Sum-DoF of Node-Intermittent 3WC). A non-adaptive encoding scheme achieves the sum-DoF of the node-intermittent 3WC given by dIsum = 2τ min{M2 , M3 } + 2τ (M1 + M2 + M3 − min{M1 , M2 , M3 } − max{M1 , M2 , M3 }). Proof. The theorem follows from achievability and converse results developed in Sections IV-B1 and IV-B2 (Lemmas 1 and 2). 8 From a DoF region perspective however it turns out that non-adaptive schemes cannot be optimal, and therefore the presented scheme is not DoF region optimal in the intermittent 3WC. Theorem 3 (Necessity of Adaptive Encoding for DoF Region of Node-Intermittent 3WC). Adaptive encoding is required to achieve the DoF region of the node-intermittent 3WC, i.e. I 6= ∅ D I \ DA Proof. The theorem follows from an upper bound on d31 under non-adaptive encoding, presented in Section IV-C1, and counterexamples of adaptive schemes exceeding this bound, devised in Section IV-C2. Theorems 2 and 3 show that non-adaptive encoding is sufficient to achieve sum-DoF, but not sufficient to achieve the DoF region of the intermittent 3WC. This is particularly interesting in light of the fact, that nonadaptive encoding is sufficient to achieve the DoF region of the non-intermittent 3WC, which the presented non-adaptive scheme does. Theorem 4 (DoF Region of Non-Intermittent 3WC). The DoF region of the non-intermittent 3WC DN (with M1 ≥ M2 ≥ M3 w.l.o.g.) is given by max{d12 + d13 + d23 , d12 + d13 + d32 } ≤ M1 max{d21 + d31 + d32 , d21 + d31 + d23 } ≤ M1 max{d21 + d13 + d23 , d12 + d31 + d32 } ≤ M2 max{d31 + d32 , d13 + d23 } ≤ M3 min{d12 , d13 , d21 , d23 , d31 , d32 } ≥ 0 and achievable using a non-adaptive encoding scheme. Proof. The theorem follows from Lemmas 5 and 6 in Sections V-A and V-B. This complements earlier work [18] that characterized the sum-DoF of the non-intermittent 3WC and showed its achievability using non-adaptive encoding. Our results show that intermittency does not decrease sum-DoF as long as node 1 has the smallest number of antennas. Otherwise, intermittency does affect sum-DoF, which is affine-linearly increasing in τ . Fig. 5 and 6 plot sum-DoF of non-intermittent and node-intermittent 3WC against M3 , for different values of τ and M2 , respectively, in a scenario where node 1 has the largest number of antennas. The resulting graphs are piecewise linear with a change in slope at M3 = M2 . For M3 < M2 the sum-DoF of the 3WC is a constant depending only on M2 , for M3 > M2 it is linear in M3 . The slope of the sum-DoF of the intermittent 3WC is proportional to τ for M3 < M2 and proportional to τ for M3 > M2 . The sum-DoF of the intermittent 3WC with 0 < τ < 1 is a convex time-sharing combination of the two extreme cases τ = 0 and τ = 1 (Fig. 5). Note, that such a time-sharing combination is the best any non-adaptive coding scheme can achieve, when each node knows the intermittency state ahead of time and codes optimally for the respective state in each time instance `. The achievability scheme presented in this work does not use intermittency state information at all, yet achieves the same sum-DoF, averaging out intermittency state through 9 M1 = 10, M2 = 7 20 18 14τ + 18τ = 15 Sum-DoF 15 14 14 10 6τ + 14τ = 8 6 5 0 0 2 3WC τ = 0.00 τ = 0.25 τ = 1.00 4 6 8 10 M3 Fig. 5: Sum-DoF of 3WC (red) and intermittent 3WC (blue) with varying τ (marks show case τ = 0.25 as convex combination of extreme cases τ = 0 and τ = 1) M1 = 10, τ = 0.7 20 Sum-DoF 15 10 M2 M2 M2 M2 M2 M2 5 0 0 2 4 6 8 =2 =5 =9 =2 =5 =9 10 M3 Fig. 6: Sum-DoF of 3WC (red) and intermittent 3WC (blue) with varying M2 erasure coding. This shows that intermittency state information is dispensable for the non-adaptive scheme in this case. Furthermore, note that the sum-DoF of the 3WC is larger than the sum-DoF of the intermittent 3WC (Fig. 6). In the sequel we provide detailed derivations and proofs of our main results highlighted in this section. 10 Node 1 S3 ∼ Bern(τ3 ) Node 2 S2 ∼ Bern(τ2 ) Node 3 S1 ∼ Bern(τ1 ) Fig. 7: Rationale behind the aliases s1,` := 1, s2,` := s3,` := s` , τ1 := 1, τ2 := τ3 := τ introduced in Section IV-A: treating every link j ↔ k as potentially intermittent with intermittency state variable si,` (marginally distributed as Bern(τi )) allows to derive general expressions for, e.g., yi,` based on xj,` , xk,` , sj,` and sk,` , independent of whether i = 1, i = 2 or i = 3 IV. N ODE I NTERMITTENCY In this section we first introduce a non-adaptive ZF/IA/EC-based scheme and derive its achievable sumDoF and DoF region. We then show, using enhanced genie-aided bounds for both adaptive and non-adaptive encoding, that this scheme is sum-DoF optimal, but not DoF region optimal. Furthermore, no non-adaptive scheme can be DoF region optimal, since tighter outer bounds hold for non-adaptive schemes, that however can be exceeded by adaptive schemes, as we show by examples of multi-hop and decode-forward relaying. This establishes that adaptive encoding is necessary from a DoF region perspective, but non-adaptive encoding is sufficient to achieve sum-DoF. A. A Non-Adaptive Scheme Based on ZF, IA and EC In this subsection we present a non-adaptive transmission scheme based on ZF, IA and EC that provides an inner bound on sum-DoF and DoF region of 3WCs. Throughout this section we continue to assume i, j, k ∈ {1, 2, 3} and mutually distinct. For notational simplicity (capturing the symmetries inherent in the model thereby avoiding case distinctions), we introduce the following aliases that will be resubstituted towards the end of the section: s1,` := 1 s2,` := s3,` := s` τ1 := 1 τ2 := τ3 := τ (5) The idea behind these aliases is visualized in Fig. 7: Using these aliases we can, for instance, find a general expression for the receive signal yi,` of node i in terms of the transmit signals xj,` and xk,` of nodes j and k, and the intermittency states sk,` and sj,` (marginally distributed as Bern(τk ) and Bern(τj ), respectively) of the links j ↔ i and k ↔ i, respectively. The fact that 2 ↔ 3 is not intermittent, and 1 ↔ 2 and 1 ↔ 3 are jointly intermittent, is accounted for by the appropriate resubstitution at due time. [ZF] 1) Encoding: Each node splits each message wij into wij interference alignment, respectively. At node i, the four messages [q]n codewords xij [q] with symbols xij,` ∈ C [q] aij [IA] and wij [q] wij to be sent via zero-forcing and (q ∈ {ZF, IA}) are encoded into [q] each, for some vector lengths aij ∈ N0 . The symbols of these 11 TX space Z12 Z13 A21 TX space A31 Z21 Z23 A12 A32 Node 1 Node 2 Node 3 Z13 Z23 A31 A32 Z23 A21 A32 A12 τ A12 τ RX space Fig. 8: Visualization of transmit (TX) signal spaces at nodes 1 and 2 and receive (RX) signal space at node 3 depending on intermittency state of 1 ↔ 3 where the first and second row are received with probability τ and [IA] [ZF] τ , respectively (Zij denotes wij , Aji denotes wij , desired signals in red, interfering signals in blue) codewords are chosen i.i.d. CN (0, pi Ia[q] ) respectively, where pi is the power. The power constraints on xni ij stated in Section II are satisfied by choosing pi = P [ZF] aij + [IA] aij [ZF] + aik [IA] . (6) + aik For encoding, the codes are designed to employ EC to be able to tolerate a certain number of symbol erasures (e.g., caused by intermittency), by not using all codeword symbols for net user data, but deliberately adding some redundancy. The rates and DoFs are thereby reduced accordingly. While for the non-intermittent 3WC this additional layer of EC is not required, it is made use of for intermittent 3WCs to cope with intermittency. 2) Transmission: At time `, node i sends h X xi,` = [q] [q] [q] [q] Vij xij,` + Vik xik,` i (7) q∈{ZF,IA} [q] [q] where Vij ∈ CMi ×aij are pre-coding matrices with unit-norm column vectors. Zero-forcing is achieved by [ZF] choosing the Vij such that [ZF] Hik Vij [ZF] Such matrices Vij = 0. (8) [ZF] exist if node i has enough antennas to send aij streams to node j without interfering with node k, i.e. [ZF] aij ≤ (Mi − Mk )+ . (9) To avoid any overlap of the different transmit signal subspaces, we require furthermore that [ZF] aij [IA] [ZF] + aij + aik [IA] + aik ≤ Mi . (10) 12 3) Decoding: Node i receives (Fig. 8) yi,` = sk,` Hji xj,` + sj,` Hki xk,` + zi,` (11)   [IA] h i x X  jk,`  [q] [q] [q] [q] = sk,` Hji Vji xji,` + sj,` Hki Vki xki,` + Gi   + zi,` [IA] xkj,` q∈{ZF,IA} [IA] with Gi = [ sk,` Hji Vjk [IA] sj,` Hki Vkj [ZF] [ZF] [ZF] (12) [ZF] ]. Note that the terms Hji Vjk xjk,` and Hki Vkj xkj,` vanished due to the zero-forcing condition (8). The first summand represents the four desired signals from j → i and k → i, the second summand represents occasional interference from j ↔ k, and the third summand is noise. [q]n To decode a desired signal xji , node i zero-forces the remaining signals by multiplying yin with a suitable [q] [q] post-coder Tji ∈ Caji ×Mi with unit-norm row vectors satisfying: [q] [q]H Tji Tji [q] Tji h [q] Hji Vji [ZF] Hki Vki [IA] Hki Vki [q] = Ia[q] (13) ji i with q ∈ {ZF, IA} \ {q} Gi = 0 [q] [q] rank(Tji Hji Vji ) = aji (14) (15) Here, (14) ensures zero-forcing of the remaining three messages and the interference and (15) ensures post[q] coding without loss of meaningful signal dimensions. The existence of such post-coders Tji is guaranteed as long as the columns of h [ZF] Hji Vji [IA] [ZF] Hji Vji Hki Vki [IA] Hki Vki Gi i [IA] (16) [IA] are linearly independent. Let γi be the dimension of span(Hji Vjk ) ∩ span(Hki Vkj ). Then, the dimension [IA] [IA] of span(Gi ) is ajk + akj − γi , and the above linear independence is possible almost surely if we choose [ZF] aji [IA] [ZF] + aji + aki [IA] [IA] [IA] + aki + (ajk + akj − γi ) ≤ Mi . (17) [q] To minimize the impact of interference, we choose the pre-coders Vij such that all γi are maximized, i.e., we ‘maximally’ align the interference subspaces at the receivers. The γi cannot be chosen arbitrarily large, the dimension of the intersection of the interference subspaces (γi ) is upper bounded by the dimensions of the [IA] [IA] interference subspaces (ajk and akj ). Furthermore, γi needs to be smaller than the dimension of span(Hji )∩ span(Hki ), which is (Mj + Mk − Mi )+ almost surely. Therefore, we require that [IA] [IA] γi ≤ min{ajk , akj , (Mj + Mk − Mi )+ }. (18) After post-coding, node i is left with the signals [q] [q] [q] [q] [q] yji,` = sk,` Tji Hji Vji xji,` + Tji zi,` . (19) The resulting channel is an erasure-Gaussian MIMO channel with erasure probability τ k , i.e., a channel whose output is [q] [q] [q] [q] [q] Yji = Sk Tji Hji Vji Xji + Tji Zi (20) 13 [q] with random variables Sk ∼ Bern(τk ), Xji ∼ CN (0, pj Ia[q] ) and Zi ∼ CN (0, σ 2 IMi ). We treat this as a ji channel with state known causally to the receiver [43, ch. 7.4], such that for large n, the achievable rate over [q] [q] this channel is the mutual information between Xji and (Yji , Sk ): [q] [q] [q] [q] I(Xji ; Yji Sk ) = I(Xji ; Yji | Sk ) [q] (21) [q] [q] [q] = τk I(Xji ; Yji | Sk = 1) + τ k I(Xji ; Yji | Sk = 0)   pj [q] [q] [q]H H [q]H Tji = τk log det Ia[q] + 2 Tji Hji Vji Vji Hji ji σ (22) (23) [q] Due to (6) and (15), this leads to a DoF of τk aji , and the code that achieves it is an EC. [ZF] 4) Achievable DoF Region: For n large, i → j has a total of τk aij [ZF] dij = τk aij [IA] + τk aij DoFs per channel use, [IA] + τk aij . (24) [q] Collecting (9), (10), (17) and (18) as well as non-negativity of every aij , we obtain: [ZF] aij [ZF] aji [IA] [ZF] + aji + aki [IA] [ZF] + aij + aik [IA] [IA] [IA] + aik ≤ Mi (25) [IA] + aki + (ajk + akj − γi ) ≤ Mi [ZF] aij (26) ≤ (Mi − Mk )+ [IA] (27) [IA] γi ≤ min{ajk , akj , (Mj + Mk − Mi )+ } [q] 0 ≤ aij (28) (29) Using (24), we obtain: τj dij + τk dik ≤ τj τk Mi [ZF] [ZF] τj τi dji + τk τi dki + τj τk djk + τk τj dkj − τi τj τk ajk − τi τk τj akj − τi τj τk γi ≤ τi τj τk Mi [ZF] τi (γi + ajk ) ≤ djk (30) (31) (32) γi ≤ (Mj + Mk − Mi )+ (33) [ZF] aij [ZF] ≤ (Mi − Mk )+ min{aij , dij , γi } ≥ 0 (34) (35) Instantiating these constraints for every possible combination of i, j, k ∈ {1, 2, 3} mutually distinct, resubstituting all τi from (5), collecting the resulting bounds and eliminating redundant bounds, finally yields (for both (i, i) ∈ {(2, 3), (3, 2)}): [ZF] d12 + d13 ≤ τ M1 (36) di1 + τ dii ≤ τ Mi (37) [ZF] (38) d21 + d31 + τ d23 + τ d32 − τ a23 − τ a32 − τ γ1 ≤ τ M1 [ZF] d1i + τ dii + d1i + di1 − τ a1i [ZF] − τ ai1 − τ γi ≤ τ Mi (39) [ZF] (40) [ZF] (41) 0 ≤ (γ1 + aii ) ≤ dii 0 ≤ τ (γi + a1i ) ≤ d1i 14 [ZF] 0 ≤ τ (γi + ai1 ) ≤ di1 (42) 0 ≤ γ1 ≤ (M2 + M3 − M1 )+ (43) 0 ≤ γi ≤ (M1 + Mi − Mi )+ (44) [ZF] ≤ (M1 − Mi )+ (45) [ZF] ≤ (Mi − Mi )+ (46) [ZF] ≤ (Mi − M1 )+ (47) 0 ≤ a1i 0 ≤ ai1 0 ≤ aii In order to resolve the (.)+ expressions, we do the following for each of the twelve cases of Mi ≥ Mj +Mk ≥ Mj ≥ Mk and Mj + Mk ≥ Mi ≥ Mj ≥ Mk (for all possible combinations i, j, k ∈ {1, 2, 3} mutually distinct): 1) Instantiate the (.)+ expressions under the respective assumption on the numbers of antennas, therefore [ZF] some of the aij and γi will be forced to zero. [ZF] 2) Perform Fourier-Motzkin’s elimination to remove all remaining aij and γi and obtain the achievable DoF region. We then combine the resulting achievable DoF regions into the following compact formulation: max{d12 + d13 , d21 + d31 } ≤ τ M1 (48) max{d21 + τ d23 , d12 + τ d32 } ≤ τ M2 (49) max{d31 + τ d32 , d13 + τ d23 } ≤ τ M3 (50) max{d12 + d13 + τ d23 , d21 + d31 + τ d32 } ≤ τ max {M1 , M3 } (51) max{d12 + d13 + τ d32 , d21 + d31 + τ d23 } ≤ τ max {M1 , M2 } (52) max{d12 + d31 + τ d32 , d21 + d13 + τ d23 } ≤ τ max {M2 , M3 } (53) min{d12 , d13 , d21 , d23 , d31 , d32 } ≥ 0 (54) This region is achievable for tuples d with non-negative integer entries. Tuples with non-negative real entries (such as the corner points of the region) are first approximated by non-negative rationals which then can be achieved using symbol extension, as in [44]. I . This proves Theorem 1. The set of all DoF tuples d satisfying constraints (48) to (54) is denoted by DIB,A What is the rationale behind the DoF region inner bound (48) to (54)? The first three inequalities constrain the sum-DoF of outbound and inbound streams at each node, similar to cut-set bounds. The next three inequalities follow this rule: For each of the three links 1 ↔ 2, 1 ↔ 3 and 2 ↔ 3 there are two DoF variables, one for each direction (i.e., d12 and d21 , etc.). For each link pick one direction. There are eight such combinations. Whenever a combination contains both DoF variables that occur in a node’s outbound or inbound sum-DoF constraint, the number of antennas at this node appears in the max{.} operator at the right side of the inequality. This means that whenever a node’s index appears two times as left or two times as right index, this node’s index appears also in the max{.} on the right side. As can be seen above, there are six cases where each case applies to two nodes each, while for the third node it does not, because the third node’s index appears once as left and once as right index. There are two cases missing altogether, d12 + d31 + τ d23 and d21 + d13 + τ d32 , where the indices of all three nodes appear 15 once as left and once as right index. Depending on numbers of antennas, this achievable DoF region yields four bounds for the largest and two bounds for the second largest node from the inequalities (51) to (53), and two bounds for the third largest node from (48) to (50). The remaining bounds from (48) to (50) are inactive due to the tighter bounds from (51) to (53). Although we proved that the above region is achievable, it is still useful to provide a ‘recipe’ which describes [q] how a specific DoF tuple can be achieved. To obtain the actual allocation of signal dimensions aij for a DoF tuple d satisfying (48) to (54), proceed as follows: First, use as many ZF resources as possible. Only once the ZF dimensions are exhausted, assign IA resources and align as much of the resulting interference as possible. Throughout the process, account for redundancy required by EC to be able to tolerate intermittency. As an example (Fig. 9), assume (M1 , M2 , M3 , τ ) = (5, 7, 4, 0.5). There is one ZF dimension 1 → 2, three ZF dimensions 2 → 1 and two ZF dimensions 2 → 3, all other communication cannot be zero-forced. We try to achieve d = (0.5, 0, 0.5, 4, 0, 4). To transmit on average half a symbol per channel access full-duplex over the intermittent 1 ↔ 2, we use a rate 1 2 [ZF] [ZF] EC over one available ZF dimension in each direction (a12 = a21 = 1), [IA] [IA] the remaining two ZF dimensions 2 → 1 remain unused. IA is not required for 1 ↔ 2 (a12 = a21 = 0). No [ZF] communication 1 ↔ 3 takes place (a13 [IA] [ZF] = a13 = a31 [IA] = a31 = 0). To transmit four symbols per channel [ZF] access 2 → 3, we use the two available ZF dimensions for two of them (a23 = 2), and two IA dimensions [IA] (a23 = 2) that occupy a two-dimensional interference subspace at node 1. To transmit four symbols per channel [IA] [ZF] access 3 → 2, we use four IA dimensions (a32 = 4), since ZF is not possible (a32 = 0). All of them create interference at node 1, but this four-dimensional interference subspace can be aligned with the two-dimensional interference subspace caused by 2 → 3 (γ1 = 2). As a result, four of the five receive dimensions at node 1 are interference of 2 ↔ 3 communication, while the remaining was used for zero-forced and erasure-coded communication 2 → 1. At nodes 2 and 3 no interference is caused, such that no interference alignment takes place (γ2 = γ3 = 0). B. Sum-DoF The devised non-adaptive encoding scheme based on ZF, IA and EC is sum-DoF optimal in the intermittent 3WC, as we will prove in this section. 1) Lower Bounds: Using (36) to (47), we derive a lower bound on the sum-DoF of the intermittent 3WC with intermittent node 1. For each of the six cases of Mi ≥ Mj ≥ Mk (for all possible combinations i, j, k ∈ {1, 2, 3} mutually distinct), we solve the linear program maximizing sum-DoF using, e.g., the simplex algorithm. Definition 1. We denote dIsum,LB,A := max [d12 + d13 + d21 + d23 + d31 + d32 ] . d∈D I IB,A The resulting lower bounds for each of the different cases are listed in Table I, and are condensed into a single expression in the following lemma: Lemma 1 (Sum-DoF Lower Bound for Node-Intermittent 3WC). The sum-DoF dIsum,LB,A = 2τ min{M2 , M3 } + 2τ (M1 + M2 + M3 − min{M1 , M2 , M3 } − max{M1 , M2 , M3 }) ≤ dIsum 16 TX Z12 RX Z21 A32 A23 τ τ Node 1 Node 3 Node 2 TX Z21 RX Z12 Z23 A32 A23 TX A23 τ RX Z23 A32 τ A23 τ Z23 A32 τ Fig. 9: Visualization of transmit (TX) and receive (RX) signal spaces achieving d = (0.5, 0, 0.5, 4, 0, 4) under the assumption of (M1 , M2 , M3 , τ ) = (5, 7, 4, 0.5) (depending on intermittency state, where the first and second [ZF] [IA] row are received with probability τ and τ , respectively; Zij denotes wij , Aji denotes wij ) TABLE I: Achievable Sum-DoFs of Intermittent 3WC Case Sum-DoF Lower Bound M1 ≥ M2 ≥ M3 dI = 2M3 + 2τ M2 − 2τ M3 M2 ≥ M1 ≥ M3 dI = 2M3 + 2τ M1 − 2τ M3 M2 ≥ M3 ≥ M1 M1 ≥ M3 ≥ M2 M3 ≥ M1 ≥ M2 M3 ≥ M2 ≥ M1 sum,LB,A sum,LB,A dI sum,LB,A dI sum,LB,A dI sum,LB,A dI sum,LB,A = 2M3 = 2M2 + 2τ M3 − 2τ M2 = 2M2 + 2τ M1 − 2τ M2 = 2M2 is achievable in the node-intermittent 3WC and therefore constitutes a lower bound on the sum-DoF of the node-intermittent 3WC. 2) Upper Bounds: We first motivate the converse techniques used throughout this section and in the sequel provide rigorous proofs. The general approach for upper bounding the sum-DoF of the intermittent 3WC is as follows: Partition the DoF sum dij + dik + dji + djk + dki + dkj into two partial sums dij + dkj + dki and dik + djk + dji (Fig. 10), where wij , wkj , wki are to be decoded by node j and wik , wjk , wji are to be decoded by node k. There are three such partitions, and the partition is fully determined by choosing which node takes the role of node i. While nodes j and k function exclusively as source and sink in any one of the partial sums, node i is an intermediary node in both. Have an imaginary genie provide just enough side information to node j and k (hence the name ‘genie-aided’ bound), respectively, such that they can recover the receive signal yin of node i, then (assuming existence of a suitable coding scheme) nodes j and k can decode the additional messages wki and wji , respectively. The details of this decoding process and the required side information will be presented in due course. Here we 17 i j i k (a) j k (b) Fig. 10: Partition the DoF sum dij + dik + dji + djk + dki + dkj into two partial sums dij + dkj + dki to be decoded at node j (a) and dik + djk + dji to be decoded at node k (b), where nodes j and k are provided enough side information such that they can recover the receive signal yin of node i and from it decode the one message not originally intended for them (additional dki and dji DoFs, respectively). only remark, that the side information can serve for the following four purposes (some cases might not require some of the types of side information): 1) An additional message is required such that node j (or k) can decode the additional message wki (or wji ) from the recovered yin , because decoding requires knowledge of wi , and node j (or k) only knows wij (or wik ) from decoding its own receive signal yjn (or ykn ). 2) Node j (or k) might be incapable of capturing enough information about yin because Mj (or Mk ) is small. In this case, additional measurements about yin (or alternatively about the unknown ‘ingredient’ of interest, xnk or xnj , respectively) need to be provided by side information. 3) Node j (or k) might be incapable of capturing enough information about yin because of intermittency. In this case, additional measurements about yin for those time instances ` where s` = 0 need to be provided by side information. 4) For rather technical reasons a noise correction signal is required to accurately recover yin . However, the information contained in this signal about the three desired messages in question scales only o [log(ρ)]. Assuming reliable communication, the partial DoF sum of each three messages is necessarily upper bounded by a mutual information expression, using Fano’s inequality. Adding the two resulting bounds yields an upper bound on the sum-DoF. The challenge with this approach is two-fold: a) provide as little side information as possible to the respective nodes, b) use tight bounding when expanding the mutual information expression from Fano’s inequality. For the intermittent 3WC the partition that yields the tightest upper bound on the sum-DoF depends on the numbers of antennas. The node with largest number of antennas should take the role of the intermediary node i (Fig. 11). The order among the remaining two nodes decides about which side information to give to which node, to compensate for insufficient number of antennas or intermittency. We prove upper bounds for the three cases M1 ≥ M2 ≥ M3 , M2 ≥ M1 ≥ M3 and M2 ≥ M3 ≥ M1 , the remaining three cases go by renaming node 2 and 3. a) Case 1: M1 ≥ M2 ≥ M3 : For the case where M1 is the largest number of antennas, we develop two partial sums around nodes 3 and 2 (Fig. 11a and 11b), respectively. We start with the bound around node 2 (Fig. 11b), as it requires less side information and is therefore simpler to argue, and in the sequel extend the 18 1 2 1 3 2 3 (a) (b) 1 1 2 3 2 (c) 3 (d) Fig. 11: Partition of sum-DoF into partial sums for case M1 ≥ M2 ≥ M3 (a)/(b) and cases M2 ≥ M1 ≥ M3 and M2 ≥ M3 ≥ M1 (c)/(d). sn , w2 , y2n Enc. E1,` Dec. F2 x1,` Enc. E2,` Cancel x1,` H32 x3,` + z2,` w13 ŵ32 , ŵ12 `++ Invert x3,` + ŵ31 Dec. F1 y1,` x2,` † H32 z2,` , if s` = 1 Combine zcorr,` Fig. 12: Decoding of w31 at node 2 by iterative reconstruction of y1n from a-priori knowledge w2 and n observations (y2n , sn ) using side information (w13 , zcorr ) (grey box: a-priori knowledge and channel output, dashed boxes: side information): successively obtain x1,` , cancel its effect from y2,` to obtain a noisy version of x3,` , and combine this with x2,` and zcorr,` to finally obtain y1,` ; repeat for next `. basic technique to develop the bound around node 3 (Fig. 11a), which requires more side information and is therefore slightly more involved. Which side information does node 2 need to be able to recover y1n and decode w31 from it, assuming a scheme allowing every node to decode its desired messages with high probability? We present a suitable process depicted in Fig. 12. At the end of the transmission, node 2 has w2 , y2n , sn and xn2 , as shown on the top left of the figure. Using the decoder F2 it can decode (ŵ12 , ŵ32 ) = (w12 , w32 ) with high probability. Assume we provide w13 as side information, so that node 2 can decode messages intended for node 1 using F1 as soon as it obtains y1n , as shown on the bottom left of the figure. Node 2 can now obtain x1,1 from w1 using E1,1 . 19 From (s1 , y2,1 ) it can obtain H32 x3,1 + z2,1 using x1,1 . Since H32 is a tall matrix, a noisy version of x3,1 † can be obtained from H32 x3,1 + z2,1 using the pseudo-inverse H32 , i.e., † † z2,1 . (H32 x3,1 + z2,1 ) = x3,1 + H32 H32 Using the noisy version of x3,1 , node 2 can obtain a noisy version of y1,1 , i.e.,    † † z2,1 . z2,1 = s1 H21 x2,1 + s1 H31 x3,1 + s1 H31 H32 s1 H21 x2,1 + H31 x3,1 + H32 † Given a suitably formed noise correction term zcorr,1 := z1,1 − s1 H31 H32 z2,1 as side information, node 2 can finally obtain y1,1 , i.e. † z2,1 + zcorr,1 . y1,1 = s1 H21 x2,1 + s1 H31 x3,1 + s1 H31 H32 Then, this reconstruction cycle repeats for the next ` = 2, ..., n, where all previous y1`−1 are used to obtain x1,` using E1,` . From the four abstract types of side information introduced before, only two are required for the reconstruction and subsequently for this bound: a message and a noise correction term. No side information to compensate for insufficient number of antennas or intermittency is required. n In a nutshell, side information w13 and zcorr is provided to node 2 by a genie, defined as † Zcorr,` := Z1,` − S` H31 H32 Z2,` . Since the scheme ought to be reliable, we bound the sum rate of w12 , w32 and w31 using Fano’s inequality: side information n(R12 + R32 + R31 − ε(1) n ) ≤ I(W12 W32 W31 ; W2 Y2n S n z }| { n W13 Zcorr ) (a) n = I(W12 W32 W31 ; Y2n | W13 W2 S n Zcorr ) n X (b) n = I(W12 W3 ; Y2,` | Y2`−1 S n W13 W2 Zcorr ) `=1 n X   n n h(Y2,` | Y2`−1 S n W13 W2 Zcorr ) − h(Y2,` | Y2`−1 S n W1 W2 W3 Zcorr ) = `=1 n X   n ≤ h(Y2,` | S` ) − h(Y2,` | Y2`−1 S n W1 W2 W3 Zcorr X1,` X3,` ) (c) (d) = = (b) = `=1 n X `=1 n X `=1 n X [h(Y2,` | S` ) − h(Y2,` | S` Zcorr,` X1,` X3,` )] I(Zcorr,` X1,` X3,` ; Y2,` | S` ) [I(X1,` X3,` ; Y2,` | S` ) + I(Zcorr,` ; Y2,` | S` X1,` X3,` )] `=1 (e) = n X  =o[log(ρ)] z }|  I(X1,` X3,` ; Y2,` | S` ) + I(Zcorr,` ; Z2,` `=1 (f) ≤ n [τ M2 + τ M3 ] log(ρ) + no [log(ρ)] These steps are justified as follows: n (a) (W12 , W3 ) is independent of (W13 , W2 , S n , Zcorr )  { | S` ) 20 (b) Chain rule for mutual information (c) Conditioning reduces entropy n `−1 n (d) Y2,` is independent of (Y2`−1 , S `−1 , S`+1 , Zcorr , Zcorr,`+1 , W1 , W2 , W3 ) given (S` , X1,` , X3,` ) (e) Zcorr,` is independent of (X1,` , X3,` ) and Y2,` given (X1,` , X3,` ) is only noise with a mean Z2,` + const (f) (X1,` , X3,` ) Y2,` is a MIMO channel with min{M1 + M3 , M2 } = M2 DoFs if s` = 1, and min{0 + M3 , M2 } = M3 DoFs if s` = 0 Dividing both sides by n log(ρ) and letting ρ, n → ∞ we obtain d12 + d32 + d31 ≤ τ M2 + τ M3 . (55) We turn to the second partial sum, developed around node 3 (Fig. 11a). This bound is slightly more involved, as an additional type of side information is required which compensates for the small number of antennas M3 . Which side information does node 3 need to be able to recover y1n and decode w21 from it, assuming a scheme allowing every node to decode its desired messages with high probability? A suitable process is depicted in Fig. 13. At the end of the transmission, node 3 has w3 , y3n , sn and xn3 , as shown on the top left of the figure. Using the decoder F3 it can decode (ŵ13 , ŵ23 ) = (w13 , w23 ) with high probability. Assume we provide w12 as side information, so that node 3 can decode messages intended for node 1 using F1 as soon as it obtains y1n , as shown on the bottom left of the figure. Node 3 can now obtain x1,1 from w1 using E1,1 . From (s1 , y3,1 ) it can obtain H23 x2,1 + z3,1 using x1,1 . Assume we ‘virtually’ increase the number of antennas at node 3 so that it can fully observe x2,1 whenever node 1 can, by providing ỹ3,1 = H̃23 x2,1 + z̃3,1 as side information h i H if s1 = 1, with H̃23 ∈ C(M2 −M3 )×M2 such that rank( H̃23 ) = M2 . Note that if s1 = 0 then x2,1 does not 23 contribute to y1,1 . In this case, x2,1 does not need to be reconstructed, and therefore no side information to compensate for insufficient number of antennas is required. We define shortcuts to group receive signal y3n and side information ỹ3n into a joint signal ŷ3n , i.e., i h i h y3,1 H , ŷ3,1 := ỹ3,1 , Ĥ23 := H̃23 23 ẑ3,1 := h z3,1 z̃3,1 i . −1 ẑ3,1 if A matrix H̃23 satisfying rank(Ĥ23 ) = M2 exists almost surely and it allows to obtain x2,1 + Ĥ23 −1 s1 = 1. Assume we provide zcorr,1 = z1,1 − s1 H21 Ĥ23 ẑ3,1 as side information. Then node 3 can obtain y1,1 −1 from zcorr,1 if s1 = 0, and from x3,1 , x2,1 + Ĥ23 ẑ3,1 and zcorr,1 if s1 = 1, i.e., −1 y1,1 = s1 H21 (x2,1 + Ĥ23 ẑ3,1 ) + s1 H31 x3,1 + zcorr,1 = s1 H21 x2,1 + s1 H31 x3,1 + z1,1 . Using (y1,1 , w1 ) and the encoder E1,2 node 3 can obtain x1,2 and the cycle repeats, for ` = 2, ..., n. Finally, node 3 obtains y1n , and decodes w21 from (y1n , w1 , sn ). Side information to compensate for intermittency is not required. n In a nutshell, side information w12 , ỹ3n and zcorr is provided to node 3 by a genie, following the definitions Ỹ3,` := S` (H̃23 X2,` + Z̃3,` ), −1 Zcorr,` := Z1,` − S` (H21 Ĥ23 Ẑ3,` ), where Z̃3,` h H23 H̃23 i H̃23 ∈ C(M2 −M3 )×M2 s.t. rank(Ĥ23 ) = M2 , hZ i hY i 3,` 3,` ∼ CN (0, σ 2 IM2 −M3 ), Ẑ3,` := Z̃3,` , Ŷ3,` := Ỹ3,` . Ĥ23 := , 21 sn , w3 , y3n x1,` Enc. E1,` Dec. F3 Enc. E3,` Cancel x1,` H23 x2,` + z3,` w12 ŵ23 , ŵ13 `++ ỹ3,` Invert x3,` −1 ẑ3,` , if s` = 1 x2,` + Ĥ23 Dec. F1 ŵ21 Combine y1,` zcorr,` Fig. 13: Decoding of w21 at node 3 by iterative reconstruction of y1n from a-priori knowledge w3 and n observations (y3n , sn ) using side information (w12 , ỹ3n , zcorr ) (grey box: a-priori knowledge and channel output, dashed boxes: side information): successively obtain x1,` , cancel its effect from y3,` (using side information ỹ3,` ) to obtain a noisy version of x2,` , and combine this with x3,` and zcorr,` to finally obtain y1,` ; repeat for next `. Since the scheme ought to be reliable, we again bound the sum rate of w13 , w23 and w21 using Fano’s inequality, and following similar steps as before (for details see appendix A) we obtain d13 + d23 + d21 ≤ τ M2 + τ M3 . (56) The two bounds on partial DoF sums are combined into a sum-DoF upper bound for the case M1 ≥ M2 ≥ M3 : Proposition 1 (Sum-DoF of Node-Intermittent 3WC for M1 ≥ M2 ≥ M3 ). dIsum ≤ 2τ M2 + 2τ M3 Proof. The proposition follows by adding (55) and (56). b) Case 2: M2 ≥ M1 ≥ M3 : Since M2 is the largest number of antennas, we develop two partial sums around nodes 3 and 1 (Fig. 11c and 11d), respectively. The reasoning around node 3 in this case proceeds in close analogy to the bound around node 3 in the previous case, just with 1 and 2 interchanged. We provide as side information w21 (to allow for decoding using n F2 ), ỹ3n (to compensate for small number of antennas M3 ) and zcorr (a noise correction), defined as Ỹ3,` := S` (H̃13 X1,` + Z̃3,` ), −1 Zcorr,` := Z2,` − S` (H12 Ĥ13 Ẑ3,` ), where Z̃3,` h H13 H̃13 i H̃13 ∈ C(M1 −M3 )×M1 s.t. rank(Ĥ13 ) = M1 , hZ i hY i 3,` 3,` ∼ CN (0, σ 2 IM1 −M3 ), Ẑ3,` := Z̃3,` , Ŷ3,` := Ỹ3,` . Ĥ13 := , With this information, node 3 can construct x2,1 from w2 using E2,1 (since w23 is a message intended for node 3 and assumed to have been decoded from y3n using F3 , and w21 is side information), obtain a noisy version 22 of x1,1 from channel output y3,1 and side information ỹ3,1 as necessary for y2,1 (i.e., if s1 = 1), combine all relevant signals into y2,1 , encode w2 using E2,2 and y2,1 to obtain x2,2 , and continue this cycle for the next ` = 2, ..., n until y2n is complete, from which w12 can be decoded with the help of w2 using F2 . After similar steps as before (see appendix B for details) we obtain d13 + d23 + d12 ≤ τ M1 + τ M3 . (57) We turn to the second partial sum, developed around node 1 (Fig. 11d), where node 1 should be enabled to decode w32 . The main difference to the previous cases is that the link 3 ↔ 2 is always available, while the link 3 ↔ 1 is intermittent. Therefore, y1,` does not contain information about x3,` if s` = 0; this needs to be compensated for by side information, here ỹ1n defined as ỹ1,` := s` (H31 x3,` + z̃1,` ), which provides measurements of xn3 for those time instances where node 1 is intermittent, i.e., s` = 0. This is an instance of the fourth type of side information, that for previous bounds was not necessary, namely side information that compensates for intermittency. Given this side information and the customary side information (a message, to allow for decoding, and a noise correction signal), the reconstruction proceeds in analogy to the n to node 1, defined as previous cases. In a nutshell, the genie provides w23 , ỹ1n and zcorr † Zcorr,` := Z2,` − H32 H31 (S` Z1,` + S ` Z̃1,` ), Ỹ1,` := S ` (H31 X3,` + Z̃1,` ), with Z̃1,` ∼ CN (0, σ 2 IM1 ), Ẑ1,` := (Z1,` , S ` Z̃1,` ), Ŷ1,` := (Y1,` , Ỹ1,` ). At the end of the transmission, node 1 has w1 , y1n , sn and xn1 . It decodes (w21 , w31 ) from y1n using its decoder F1 , and gets w23 from side information. It generates x2,1 , then uses its channel output y1,1 (if s1 = 1) or side information ỹ1,1 (if s1 = 0) to obtain a noisy version of x3,1 , and with it y2,1 . From there the cycle repeats, until y2n is obtained and w32 can be decoded. After similar steps as before (see appendix C for details) we obtain d21 + d31 + d32 ≤ τ M1 + τ M3 . (58) Proposition 2 (Sum-DoF of Node-Intermittent 3WC for M2 ≥ M1 ≥ M3 ). dIsum ≤ 2τ M1 + 2τ M3 Proof. The proposition follows by adding (57) and (58). c) Case 3: M2 ≥ M3 ≥ M1 : Since M2 is still the largest number of antennas, we again develop two partial sums around nodes 3 and 1 (Fig. 11c and 11d), respectively. The only difference to the previous case is that this time M3 ≥ M1 , therefore the number of antennas at node 1 needs to be augmented ‘virtually’ to obtain sufficient measurements of xn3 , while node 3 remains unchanged. n We turn to the partial sum around node 3 and provide as side information w21 and zcorr with † Zcorr,` := Z2,` − S` (H12 H13 Z3,` ). 23 Note that x1,` contributes to y2,` only if s` = 1. In these instances, node 3 has sufficient information about x1,` from y3,` . If s` = 0, node 3 does not have information about x1,` , but x1,` does not contribute to y2,` anyhow, so node 3 does not need additional side information in these cases. Therefore, with the given side information, node 3 can construct x2,1 from w2 , obtain a noisy version of x1,1 from channel output y3,1 as necessary for y2,1 (i.e., if s1 = 1), generate y2,1 , and continue this cycle for the next ` = 2, ..., n until y2n is complete, from which w12 can be decoded with the help of w2 . After similar steps as before (see appendix D for details) we obtain d13 + d23 + d12 ≤ M3 . (59) We turn to the second partial sum, developed around node 1 (Fig. 11d), where node 1 should be enabled to decode w32 . Again the main difference to the previous cases is that the link 3 ↔ 2 is always available, while the link 3 ↔ 1 is intermittent. Therefore, y1,` does not contain information about x3,` if s` = 0; this effect of intermittency needs to be compensated for by side information, here ỹ1n (a formal definition follows). Furthermore, the number of antennas at node 1 needs to be increased to fully capture x3,` , here accomplished by side information y̆1n (a formal definition follows). In addition, the genie provides the message w23 and noise n to node 1. In a nutshell, side information variables are defined as correction zcorr Ỹ1,` := S ` (H31 X3,` + Z̃1,` ), Y̆1,` := H̆31 X3,` + Z̆1,` , −1 Zcorr,` := Z2,` − H32 Ĥ31 h S` Z1,` +S ` Z̃1,` Z̆1,` i , with auxiliary variables Z̃1,` ∼ CN (0, σ 2 IM1 ), Z̆1,` ∼ CN (0, σ 2 IM3 −M1 ), i h H , H̆31 ∈ C(M3 −M1 )×M3 s.t. rank(Ĥ31 ) = M3 , Ĥ31 := H̆31 31 Ŷ1,` := (Y1,` , Ỹ1,` , Y̆1,` ), Ẑ1,` := (Z1,` , S ` Z̃1,` , Z̆1,` ). At the end of the transmission, node 1 has w1 , y1n , sn and xn1 . It decodes (w21 , w31 ) using F1 , and gets w23 from side information. It generates x2,1 using E2,1 , then uses side information y̆1,1 and channel output y1,1 (if s1 = 1) or side information y̆1,1 and ỹ1,1 (if s1 = 0) to obtain a noisy version of x3,1 , and with it y2,1 using zcorr,1 . From there the cycle repeats, until y2n is obtained and w32 can be decoded using F2 . After similar steps as before (see appendix E for details) we obtain d21 + d31 + d32 ≤ M3 . Proposition 3 (Sum-DoF of Node-Intermittent 3WC for M2 ≥ M3 ≥ M1 ). dIsum ≤ 2M3 Proof. The proposition follows by adding (59) and (60). Lemma 2 (Sum-DoF Upper Bound for Node-Intermittent 3WC). dIsum ≤ 2τ min{M2 , M3 } + 2τ (M1 + M2 + M3 − min{M1 , M2 , M3 } − max{M1 , M2 , M3 }) Proof. The lemma follows by Propositions 1, 2 and 3, and symmetry of node 2 and 3. (60) 24 Using the achievability and converse results developed in Sections IV-B1 and IV-B2, we establish the sumDoF of the intermittent 3WC for which non-adaptive encoding is sufficient, i.e. dIsum,LB,A = 2τ min{M2 , M3 } + 2τ (M1 + M2 + M3 − min{M1 , M2 , M3 } − max{M1 , M2 , M3 }) = dIsum . This proves Theorem 2. C. DoF Region In this section we first derive an upper bound on d31 under the assumption of non-adaptive encoding. We then show adaptive schemes that can exceed this bound, namely multi-hop and decode-forward relaying. This proves that some DoF region points are only achievable by adaptive encoding schemes, hence adaptive encoding is in general required to achieve the DoF region of the intermittent 3WC. 1) Upper Bound on d31 under Non-Adaptive Encoding: Node 1 is able to decode w31 from its channel output (y1n , sn ) and a-priori knowledge w1 with high probability. We further provide w2 as side information and bound the rate of w31 using Fano’s inequality: side information n(R31 − εn ) ≤ I(W31 ; W1 Y1n S n z}|{ W2 ) (a) = I(W31 ; Y1n | W1 W2 S n ) n (b) X = I(W31 ; Y1,` | Y1`−1 W1 W2 S n ) `=1 n X   h(Y1,` | Y1`−1 W1 W2 S n ) − h(Y1,` | Y1`−1 W1 W2 W31 S n ) = `=1 n X   ≤ h(Y1,` | W2 S` ) − h(Y1,` | Y1`−1 W1 W2 W31 S n X2,` X3,` ) (c) `=1 (d) ≤ = n X `=1 n X (b) `=1 n X (e) `=1 n X = = [h(Y1,` | W2 S` ) − h(Y1,` | W2 S` X2,` X3,` )] I(X2,` X3,` ; Y1,` | W2 S` ) [I(X2,` ; Y1,` | W2 S` ) + I(X3,` ; Y1,` | X2,` W2 S` )] I(X3,` ; Y1,` | X2,` W2 S` ) `=1 (f) ≤ nτ M3 log(ρ) + nτ o [log(ρ)] (61) These steps are justified as follows: (a) W31 is independent of (W1 , W2 , S n ) (b) Chain rule for mutual information (c) Conditioning reduces entropy n (d) Y1,` is independent of (Y1`−1 , S `−1 , S`+1 , W1 , W31 ) given (S` , X2,` , X3,` ) (e) X2,` is independent of Y1,` given (W2 , S` ) because x2,` = E2,` (w2 ) under the assumption of non-adaptive encoding 25 2 1 w31 3 τ τ ŵ31 1 Fig. 14: All messages except w31 are fixed to 0, node 2 supports node 3 in communicating w31 to node 1 (f) X3,` Y1,` given X2,` is a MIMO channel with maximum DoF M3 and 0 for s` = 1 and s` = 0, respectively The following lemma provides the desired upper bound. Lemma 3 (Upper Bound on d31 under Non-Adaptive Encoding). The achievable DoFs of non-adaptive encoding schemes are restricted by d31 ≤ τ M3 . (62) Proof. We divide both sides of (61) by n log(ρ) and take ρ, n → ∞. 2) Adaptive Schemes Achieving d31 > τ M3 : We assume all messages are fixed to 0, except for w31 , which node 3 wants to convey to node 1, potentially with the help of node 2 (Fig. 14). If τ M2 ≥ M3 , multi-hop relaying can be used to achieve M3 DoFs: Assume a block-Markov transmission scheme over B blocks consisting of n channel accesses each. In each of the first B − 1 blocks, nM3 symbols are transmitted 3 → 2. In the final block B node 3 remains silent, as does node 2 in block b = 1. In each block b ≥ 2, node 2 retransmits the message it received in the previous block b − 1, using EC to protect against intermittency. If τ M2 ≥ M3 , 2 → 1 can carry all nM3 symbols it received in the previous block. Node 1 uses backwards-decoding to successively cancel the interference from 3 → 1 and decode the messages. While multi-hop relaying achieves d31 larger than τ M3 if τ M2 ≥ M3 , this condition can be relaxed: Consider using decode-forward relaying at node 2. This can be used to achieve min{τ M1 , τ (M2 + M3 ), M3 } DoFs and outperforms any non-adaptive scheme as soon as M1 > M3 . We derive the achievable DoF for d31 based on the well-known lower bound for decode-forward relaying [45]: C ≥ max min{I(X3 X2 ; Y1 S), I(X3 ; Y2 S | X2 )} pX 3 X 2 ≥ min{I(X3 X2 ; Y1 | S), I(X3 ; Y2 | X2 )} with X2 , X3 Gaussian = min{τ min{M1 , (M2 + M3 )}, M3 } log(ρ) + o [log(ρ)] = min{τ M1 , τ (M2 + M3 ), M3 } log(ρ) + o [log(ρ)] (63) The following lemma provides the desired lower bound. Lemma 4 (Lower Bound on d31 under Adaptive Encoding). Decode-forward relaying achieves d31 ≥ min{τ M1 , τ (M2 + M3 ), M3 }. Proof. We divide both sides of (63) by log(ρ) and take ρ → ∞. (64) 26 Note that if τ M2 > τ M3 and τ M1 > M3 , then we can transmit at M3 DoF from node 3 to node 1 using this adaptive scheme, compensating all the the loss due to intermittency. We proved in Lemma 3 that the DoF region point d = (0, 0, 0, 0, d31,A , 0) d31,A := min{τ M1 , τ (M2 + M3 ), M3 } is not achievable for any non-adaptive encoding scheme if M1 > M3 , while we showed in Lemma 4 that there exist adaptive schemes that achieve it. This proves Theorem 3, which states that adaptive encoding is in general required to achieve the DoF region of the intermittent 3WC. Theorems 2 and 3 show that non-adaptive encoding is sufficient to achieve sum-DoF, but not sufficient to achieve the DoF region of the intermittent 3WC. This is particularly interesting in light of the next section, where we show that adaptive encoding is not beneficial in the non-intermittent 3WC even from a DoF region perspective. V. N O I NTERMITTENCY The sum-DoF of the non-intermittent 3WC was investigated in [18]. We present the DoF region of the nonintermittent 3WC and show that the non-adaptive encoding scheme presented in Section IV-A is sufficient to achieve it; therefore, adaptive encoding is neither required from a sum-DoF nor from a DoF region perspective in the non-intermittent 3WC. The non-intermittent 3WC is a special case of the intermittent 3WC with τ = 1. We may assume without loss of generality M1 ≥ M2 ≥ M3 . A. Achievability From (48) to (54) we obtain with τ = 1: d12 + d13 + d23 ≤ M1 (65) d12 + d13 + d32 ≤ M1 (66) d21 + d31 + d32 ≤ M1 (67) d21 + d31 + d23 ≤ M1 (68) d21 + d13 + d23 ≤ M2 (69) d12 + d31 + d32 ≤ M2 (70) d31 + d32 ≤ M3 (71) d13 + d23 ≤ M3 (72) min{d12 , d13 , d21 , d23 , d31 , d32 } ≥ 0 (73) N Definition 2. We denote the set of all DoF tuples d satisfying constraints (65) to (73) by DIB,A . N Lemma 5 (DoF Region Inner Bound for Non-Intermittent 3WC). All d ∈ DIB,A are achievable in the nonN intermittent 3WC. Therefore, DIB,A constitutes an inner bound on the DoF region of the non-intermittent 3WC, i.e. N DIB,A ⊆ DN . 27 Proof. The lemma follows by construction in Section IV-A. B. Converses Besides cut-set outer bounds [46] on pairs of DoFs (involving two messages either intended for or originating at a certain node), we use tighter genie-aided outer bounds [18], [19], [21] on triplets of DoFs (involving two messages either intended for or originating at a certain node, and one message exchanged between the remaining two nodes): R13 + R23 ≤ M3 log(ρ) + o [log(ρ)] (74) R31 + R32 ≤ M3 log(ρ) + o [log(ρ)] (75) R21 + R31 + R32 ≤ min{M1 , M2 + M3 } log(ρ) + o [log(ρ)] (76) R21 + R31 + R23 ≤ min{M1 , M2 + M3 } log(ρ) + o [log(ρ)] (77) R12 + R32 + R13 ≤ M1 log(ρ) + o [log(ρ)] (78) R13 + R23 + R12 ≤ M1 log(ρ) + o [log(ρ)] (79) R12 + R32 + R31 ≤ M2 log(ρ) + o [log(ρ)] (80) R13 + R23 + R21 ≤ M2 log(ρ) + o [log(ρ)] (81) Note that if M2 + M3 ≤ M1 , then (76) and (77) are redundant given (75) and (81), therefore min{M1 , M2 + M3 } can be replaced with M1 in (76) and (77). Dividing these bounds by log(ρ) and taking ρ → ∞ yields to the following DoF region outer bounds: d21 + d31 + d32 ≤ M1 (82) d21 + d31 + d23 ≤ M1 (83) d12 + d32 + d13 ≤ M1 (84) d13 + d23 + d12 ≤ M1 (85) d12 + d32 + d31 ≤ M2 (86) d13 + d23 + d21 ≤ M2 (87) d13 + d23 ≤ M3 (88) d31 + d32 ≤ M3 (89) Note that (86) and (87) can also be obtained from (55) and (56), respectively, with τ = 1. N Definition 3. We denote the set of all DoF tuples d satisfying constraints (82) to (89) by DOB . N Lemma 6 (DoF Region Outer Bound for Non-Intermittent 3WC). None of the d 6∈ DOB are achievable in the N non-intermittent 3WC. Therefore, DOB constitutes an outer bound on the DoF region of the non-intermittent 3WC, i.e. N DN ⊆ DOB . Proof. The lemma follows by the DoF region outer bounds (82) to (89). 28 1 1 2 τ1 1 1 τ3 3 2 (a) 1-link interm. τ2 τ1 3 (b) 3-link interm. Fig. 15: Examples of different link-intermittent MIMO 3WCs C. DoF Region and Sum-DoF of Non-Intermittent 3WC The previous achievability and converse results establish the DoF region (and thus sum-DoF) optimality of non-adaptive schemes, rendering adaptive encoding dispensable, i.e. N N = DN = DOB DIB,A . This proves Theorem 4. From the DoF region of the 3WC and using the sum-DoF of the intermittent 3WC, we reproduce the sum-DoF of the 3WC given in [18]: Corollary 1 (Sum-DoF of Non-Intermittent 3WC). = 2M2 = dN dN sum sum,LB,A Proof. The statement follows as corollary of Theorems 2 and 4. VI. C ONCLUSION We introduced the MIMO 3WC with node-intermittency and studied its DoF region and sum-DoF. In particular, we devised a non-adaptive encoding scheme based on zero-forcing, interference alignment and erasure coding, and showed its DoF region (and thus sum-DoF) optimality for non-intermittent 3WCs and its sumDoF optimality for node-intermittent 3WCs. This shows that adaptive encoding is not required in those cases. However, we showed by example that in general there are DoF region points in the node-intermittent 3WC that can only be achieved by adaptive schemes, such as multi-hop or decode-forward relaying, making adaptive encoding a necessity. Our work contributes to a better understanding of the necessity of adaptive schemes such as relaying in multi-way communications with intermittency. As remarked in the introduction, node-intermittency is only one of a multitude of practically relevant intermittency scenarios. Links might be intermittent independently of each other, e.g., moving objects passing by only interrupt the link between the two D2D users from time to time, while the other links remain intact (Fig. 15a). Or all links being intermittent, but independently of each other, and with different probabilities (Fig. 15b). Here, we speak of link intermittency and intermittent links. Intermittent 3WCs with other intermittency models are interesting directions for future research. 29 A PPENDIX A. Sum-DoF Upper Bound for Intermittent 3WC with M1 ≥ M2 ≥ M3 (Part II) Since the scheme ought to be reliable, we bound the sum rate of w13 , w23 and w21 using Fano’s inequality: side information }| { z n n n n n(R13 + R23 + R21 − ε(2) n ) ≤ I(W13 W23 W21 ; W3 Y3 S W12 Ỹ3 Zcorr ) (a) n = I(W13 W23 W21 ; Y3n Ỹ3n | W12 W3 S n Zcorr ) n (b) X n = I(W13 W2 ; Ŷ3,` | Ŷ3`−1 S n W12 W3 Zcorr ) `=1 n h i X n n = h(Ŷ3,` | Ŷ3`−1 S n W12 W3 Zcorr ) − h(Ŷ3,` | Ŷ3`−1 S n W1 W2 W3 Zcorr ) `=1 n h i X n ≤ h(Ŷ3,` | S` ) − h(Ŷ3,` | Ŷ3`−1 S n W1 W2 W3 Zcorr X1,` X2,` ) (c) `=1 n h i (d) X = h(Ŷ3,` | S` ) − h(Ŷ3,` | S` Zcorr,` X1,` X2,` ) `=1 = (b) = n X I(Zcorr,` X1,` X2,` ; Ŷ3,` | S` ) `=1 n h X I(X1,` X2,` ; Ŷ3,` | S` ) + I(Zcorr,` ; Ŷ3,` | S` X1,` X2,` ) i `=1  (e) = =o[log(ρ)] }| z n X  I(X1,` X2,` ; Ŷ3,` | S` ) + I(Zcorr,` ; Ẑ3,`  `=1  { | S` )  (f) ≤ n [τ M2 + τ M3 ] log(ρ) + no [log(ρ)] These steps are justified as follows: n (a) (W13 , W2 ) is independent of (W12 , W3 , S n , Zcorr ) (b) Chain rule for mutual information (c) Conditioning reduces entropy n `−1 n , Zcorr,`+1 , W1 , W2 , W3 ) given (S` , X1,` , X2,` ) , Zcorr (d) Ŷ3,` is independent of (Ŷ3`−1 , S `−1 , S`+1 (e) Zcorr,` is independent of (X1,` , X2,` ) and Ŷ3,` given (X1,` , X2,` ) is only noise with a mean Ẑ3,` + const (f) (X1,` , X2,` ) (Y3,` , Ỹ3,` ) is a MIMO channel with min{M1 + M2 , M3 + (M2 − M3 )} = M2 DoFs if s` = 1, and min{M1 + M2 , M3 + 0} = M3 DoFs if s` = 0 Dividing both sides by n log(ρ) and letting ρ, n → ∞ we obtain d13 + d23 + d21 ≤ τ M2 + τ M3 . B. Sum-DoF Upper Bound for Intermittent 3WC with M2 ≥ M1 ≥ M3 (Part I) Since this scheme ought to be reliable, we bound the sum rate of w13 , w23 and w12 using Fano’s inequality: side information z }| { n n n n n(R13 + R23 + R12 − ε(1) n ) ≤ I(W13 W23 W12 ; W3 Y3 S W21 Ỹ3 Zcorr ) (a) n = I(W13 W23 W12 ; Y3n Ỹ3n | W21 W3 S n Zcorr ) 30 (b) = n X n I(W1 W23 ; Ŷ3,` | Ŷ3`−1 S n W21 W3 Zcorr ) `=1 = n h X i n n h(Ŷ3,` | Ŷ3`−1 S n W21 W3 Zcorr ) − h(Ŷ3,` | Ŷ3`−1 S n W1 W2 W3 Zcorr ) `=1 (c) ≤ n h X n h(Ŷ3,` | S` ) − h(Ŷ3,` | Ŷ3`−1 S n W1 W2 W3 Zcorr X1,` X2,` ) i `=1 n h i (d) X = h(Ŷ3,` | S` ) − h(Ŷ3,` | S` Zcorr,` X1,` X2,` ) `=1 = (b) = n X I(Zcorr,` X1,` X2,` ; Ŷ3,` | S` ) `=1 n h X i I(X1,` X2,` ; Ŷ3,` | S` ) + I(Zcorr,` ; Ŷ3,` | S` X1,` X2,` ) `=1  (e) = n X `=1  I(X1,` X2,` ; Ŷ3,`   =o[log(ρ)] }| { z | S` ) + I(Zcorr,` ; Ẑ3,` | S` )  (f) ≤ n [τ M1 + τ M3 ] log(ρ) + no [log(ρ)] These steps are justified as follows: n ) (a) (W1 , W23 ) is independent of (W21 , W3 , S n , Zcorr (b) Chain rule for mutual information (c) Conditioning reduces entropy n `−1 n (d) Ŷ3,` is independent of (Ŷ3`−1 , S `−1 , S`+1 , Zcorr , Zcorr,`+1 , W1 , W2 , W3 ) given (S` , X1,` , X2,` ) (e) Zcorr,` is independent of (X1,` , X2,` ) and Ŷ3,` given (X1,` , X2,` ) is only noise with a mean Ẑ3,` + const (f) (X1,` , X2,` ) (Y3,` , Ỹ3,` ) is a MIMO channel with min{M1 + M2 , M3 + (M1 − M3 )} = M1 DoFs if s` = 1, and min{M1 + M2 , M3 + 0} = M3 DoFs if s` = 0 Dividing both sides by n log(ρ) and letting ρ, n → ∞ we obtain d13 + d23 + d12 ≤ τ M1 + τ M3 . C. Sum-DoF Upper Bound for Intermittent 3WC with M2 ≥ M1 ≥ M3 (Part II) We bound the sum rate of w21 , w31 and w32 using Fano’s inequality: side information z }| { n n n n n(R21 + R31 + R32 − ε(2) ) ≤ I(W W W ; W Y S W Ỹ Z 21 31 32 1 23 1 corr ) n 1 (a) n = I(W21 W31 W32 ; Ŷ1n | W1 W23 S n Zcorr ) n X (b) n = I(W21 W3 ; Ŷ1,` | Ŷ1`−1 S n W1 W23 Zcorr ) `=1 n h i X n n = h(Ŷ1,` | Ŷ1`−1 S n W1 W23 Zcorr ) − h(Ŷ1,` | Ŷ1`−1 S n W1 W2 W3 Zcorr ) `=1 n h i X n ≤ h(Ŷ1,` | S` ) − h(Ŷ1,` | Ŷ1`−1 S n W1 W2 W3 Zcorr X2,` X3,` ) (c) `=1 31 (d) = n h X h(Ŷ1,` | S` ) − h(Ŷ1,` | S` Zcorr,` X2,` X3,` ) i `=1 = (b) = n X I(Zcorr,` X2,` X3,` ; Ŷ1,` | S` ) `=1 n h X i I(X2,` X3,` ; Ŷ1,` | S` ) + I(Zcorr,` ; Ŷ1,` | S` X2,` X3,` ) `=1  (e) = n X `=1 (b) =  I(X2,` X3,` ; Ŷ1,`   =o[log(ρ)] }| { z | S` ) + I(Zcorr,` ; Ẑ1,` | S` )  n h i X I(X2,` X3,` ; Y1,` | S` ) + I(X2,` X3,` ; Ỹ1,` | S` Y1,` ) `=1 (f) ≤ n [τ M1 + τ M3 ] log(ρ) + no [log(ρ)] These steps are justified as follows: n (a) (W21 , W3 ) is independent of (W1 , W23 , S n , Zcorr ) (b) Chain rule for mutual information (c) Conditioning reduces entropy n n `−1 , Zcorr,`+1 , W1 , W2 , W3 ) given (S` , X2,` , X3,` ) (d) Ŷ1,` is independent of (Ŷ1`−1 , S `−1 , S`+1 , Zcorr (e) Zcorr,` is independent of (X2,` , X3,` ) and Ŷ1,` given (X2,` , X3,` ) is only noise with a mean Ẑ1,` + const (f) (X2,` , X3,` ) Y1,` is a MIMO channel with min{M2 + M3 , M1 } = M1 DoFs if s` = 1, and 0 DoFs if s` = 0; (X2,` , X3,` ) Ỹ1,` is a MIMO channel with 0 DoFs if s` = 1 (because then Ỹ1,` = 0), and min{0 + M3 , M1 } = M3 DoFs if s` = 0 (because then Y1,` is noise, and Ỹ1,` is independent of X2,` ) Dividing both sides by n log(ρ) and letting ρ, n → ∞ we obtain d21 + d31 + d32 ≤ τ M1 + τ M3 . D. Sum-DoF Upper Bound for Intermittent 3WC with M2 ≥ M3 ≥ M1 (Part I) We bound the sum rate of w13 , w23 and w12 using Fano’s inequality: side information n(R13 + R23 + R12 − ε(1) n ) ≤ I(W13 W23 W12 ; W3 Y3n S n z }| { n W21 Zcorr ) (a) n = I(W13 W23 W12 ; Y3n | W21 W3 S n Zcorr ) n X (b) n = I(W1 W23 ; Y3,` | Y3`−1 S n W21 W3 Zcorr ) `=1 n X   n n = h(Y3,` | Y3`−1 S n W21 W3 Zcorr ) − h(Y3,` | Y3`−1 S n W1 W2 W3 Zcorr ) `=1 n X   n ≤ h(Y3,` | S` ) − h(Y3,` | Y3`−1 S n W1 W2 W3 Zcorr X1,` X2,` ) (c) (d) = = `=1 n X `=1 n X `=1 [h(Y3,` | S` ) − h(Y3,` | S` Zcorr,` X1,` X2,` )] I(Zcorr,` X1,` X2,` ; Y3,` | S` ) 32 (b) = n X [I(X1,` X2,` ; Y3,` | S` ) + I(Zcorr,` ; Y3,` | S` X1,` X2,` )] `=1   =o[log(ρ)] n }| { z X (e)   = I(X1,` X2,` ; Y3,` | S` ) + I(Zcorr,` ; Z3,` | S` ) `=1 (f) ≤ n [M3 ] log(ρ) + no [log(ρ)] These steps are justified as follows: n (a) (W1 , W23 ) is independent of (W21 , W3 , S n , Zcorr ) (b) Chain rule for mutual information (c) Conditioning reduces entropy n `−1 n (d) Y3,` is independent of (Y3`−1 , S `−1 , S`+1 , Zcorr , Zcorr,`+1 , W1 , W2 , W3 ) given (S` , X1,` , X2,` ) (e) Zcorr,` is independent of (X1,` , X2,` ) and Y3,` given (X1,` , X2,` ) is only noise with a mean Z3,` + const (f) (X1,` , X2,` ) Y3,` is a MIMO channel with min{M1 + M2 , M3 } = M3 DoFs if s` = 1, and min{M2 , M3 } = M3 DoFs if s` = 0 (because X1,` is independent of Y3,` ) Dividing both sides by n log(ρ) and letting ρ, n → ∞ we obtain d13 + d23 + d12 ≤ M3 . E. Sum-DoF Upper Bound for Intermittent 3WC with M2 ≥ M3 ≥ M1 (Part II) We bound the sum rate of w21 , w31 and w32 using Fano’s inequality: side information n(R21 + R31 + R32 − ε(2) n ) ≤ I(W21 W31 W32 ; W1 Y1n S n }| { z n W23 Ỹ1n Y̆1n Zcorr ) (a) n = I(W21 W31 W32 ; Ŷ1n | W1 W23 S n Zcorr ) n X (b) n I(W21 W3 ; Ŷ1,` | Ŷ1`−1 S n W1 W23 Zcorr ) = `=1 n h i X n n = h(Ŷ1,` | Ŷ1`−1 S n W1 W23 Zcorr ) − h(Ŷ1,` | Ŷ1`−1 S n W1 W2 W3 Zcorr ) `=1 n h i X n ≤ h(Ŷ1,` | S` ) − h(Ŷ1,` | Ŷ1`−1 S n W1 W2 W3 Zcorr X2,` X3,` ) (c) `=1 n h i (d) X = h(Ŷ1,` | S` ) − h(Ŷ1,` | S` Zcorr,` X2,` X3,` ) `=1 = (b) = n X I(Zcorr,` X2,` X3,` ; Ŷ1,` | S` ) `=1 n h X i I(X2,` X3,` ; Ŷ1,` | S` ) + I(Zcorr,` ; Ŷ1,` | S` X2,` X3,` ) `=1  (e) = n X `=1  I(X2,` X3,` ; Ŷ1,`   =o[log(ρ)] z }| { | S` ) + I(Zcorr,` ; Ẑ1,` | S` )  (f) ≤ n [M3 ] log(ρ) + no [log(ρ)] 33 These steps are justified as follows: n (a) (W21 , W3 ) is independent of (W1 , W23 , S n , Zcorr ) (b) Chain rule for mutual information (c) Conditioning reduces entropy n `−1 n (d) Ŷ1,` is independent of (Ŷ1`−1 , S `−1 , S`+1 , Zcorr , Zcorr,`+1 , W1 , W2 , W3 ) given (S` , X2,` , X3,` ) (e) Zcorr,` is independent of (X2,` , X3,` ) and Ŷ1,` given (X2,` , X3,` ) is only noise with a mean Ẑ1,` + const (f) (X2,` , X3,` ) (Y1,` , Ỹ1,` , Y̆1,` ) is a MIMO channel with min{M2 + M3 , M1 + 0 + (M3 − M1 )} = M3 DoFs if s` = 1, and min{M2 + M3 , 0 + M1 + (M3 − M1 )} = M3 DoFs if s` = 0 Dividing both sides by n log(ρ) and letting ρ, n → ∞ we obtain d21 + d31 + d32 ≤ M3 . R EFERENCES [1] A. Chaaban, A. Sezgin, and M.-S. Alouini, “On the Degrees-of-Freedom of the MIMO Three-Way Channel with Intermittent Connectivity,” Proc. IEEE Int. Symp. Inf. Theory (ISIT), 2017. [2] A. Sabharwal, P. Schniter, D. Guo, D. W. Bliss, S. Rangarajan, and R. Wichman, “In-band full-duplex wireless: Challenges and opportunities,” IEEE J. Sel. Areas Commun., vol. 32, no. 9, pp. 1637–1652, Sep. 2014. [3] M. N. Tehrani, M. Uysal, and H. Yanikomeroglu, “Device-to-device communications in 5G cellular networks: Challenges, solutions, and future directions,” IEEE Commun. Mag., pp. 86–92, May 2014. [4] A. Asadi, Q. Wang, and V. Mancuso, “A Survey on Device-to-Device Communication in Cellular Networks,” IEEE Commun. Surveys Tuts., vol. 16, no. 4, pp. 1801–1819, 4th quarter 2014. [5] A. Chaaban and A. Sezgin, “Multi-way communications: An information theoretic perspective,” Found. Trends Commun. Inf. Theory, vol. 12, no. 3-4, pp. 185–371, 2015. [6] C. E. Shannon, “Two-way communication channels,” in Proc. of 4th Berkeley Symp. Math. Stat. Probab., vol. 1, 1961, pp. 611–644. [7] A. P. Hekstra and F. M. J. Willems, “Dependence balance bounds for single output two-way channels,” IEEE Trans. Inf. Theory, vol. 35, no. 1, pp. 44–53, Jan. 1989. [8] T. S. Han, “A general coding scheme for the two-way channel,” IEEE Trans. Inf. Theory, vol. 30, no. 1, pp. 35–44, 1984. [9] L. R. Varshney, “Two way communication over exponential family type channels,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), July 2013, pp. 2795–2799. [10] L. Song, F. Alajaji, and T. Linder, “Adaptation is useless for two discrete additive-noise two-way channels,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), July 2016, pp. 1854–1858. [11] Z. Cheng and N. Devroye, “Two-way networks: When adaptation is useless,” IEEE Trans. Inf. Theory, vol. 60, no. 3, pp. 1793–1813, March 2014. [12] C. Suh, I.-H. Wang, and D. Tse, “Two-way Interference Channels,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), Cambridge, MA, USA, July 2012, pp. 2801–2805. [13] Z. Cheng and N. Devroye, “On constant gaps for the two-way Gaussian interference channel,” in Ann. Allerton Conf. Commun., Control, and Comput., Oct. 2012, pp. 1783–1789. [14] W. Nam, S.-Y. Chung, and Y. H. Lee, “Capacity of the Gaussian two-way relay channel to within 1/2 bit,” IEEE Trans. Inf. Theory, vol. 56, no. 11, pp. 5488–5494, Nov. 2010. [15] Y. Song, N. Devroye, H. R. Shao, and C. Y. Ngo, “Lattice coding for the two-way line network,” IEEE J. Sel. Areas Commun., vol. 32, no. 9, pp. 1707–1720, Sep. 2014. [16] R. Vaze and J. Robert W. Heath, “On the capacity and diversity-multiplexing tradeoff of the two-way relay channel,” IEEE Trans. Inf. Theory, vol. 57, no. 7, pp. 4219–4234, July 2011. [17] M. P. Wilson, K. Narayanan, H. D. Pfister, and A. Sprintson, “Joint physical layer coding and network coding for bidirectional relaying,” IEEE Trans. Inf. Theory, vol. 56, no. 11, pp. 5641–5654, Nov. 2010. [18] H. Maier, A. Chaaban, and R. Mathar, “Degrees of Freedom of the MIMO 3-way channel,” in Proc. Int. Symp. Inf. Theory Appl. (ISITA), Melbourne, Australia, Oct. 2014, pp. 625–629. 34 [19] ——, “Symmetric degrees of freedom of the MIMO 3-Way channel with MT × MR antennas,” in Proc. IEEE Inf. Theory Workshop (ITW), Hobart, Tasamania, Australia, Nov. 2014. [20] A. Chaaban, H. Maier, A. Sezgin, and R. Mathar, “Three-Way Channels With Multiple Unicast Sessions: Capacity Approximation via Network Transformation,” IEEE Trans. Inf. Theory, vol. 62, no. 12, pp. 7086–7102, Dec. 2016. [21] A. M. Elmahdy, A. El-Keyi, Y. Mohasseb, T. ElBatt, M. Nafie, and K. G. Seddik, “Asymmetric degrees of freedom of the full-duplex MIMO 3-way channel,” in Proc. IEEE Inf. Theory Workshop (ITW), Sep. 2016, pp. 469–473. [22] L. Ong, “Capacity results for two classes of three-way channels,” in Proc. IEEE Int. Symp. Commun. Inf. Tech. (ISCIT), Gold Coast, QLD, Oct. 2012. [23] D. Gündüz, A. Yener, A. Goldsmith, and H. V. Poor, “The multi-way relay channel,” IEEE Trans. Inf. Theory, vol. 59, no. 1, pp. 51–63, Jan. 2013. [24] L. Ong, C. M. Kellett, and S. J. Johnson, “On the equal-rate capacity of the AWGN multiway relay channel,” IEEE Trans. Inf. Theory, vol. 58, no. 9, pp. 5761–5769, Sep. 2012. [25] A. Sezgin, A. S. Avestimehr, M. A. Khajehnejad, and B. Hassibi, “Divide-and-conquer: Approaching the capacity of the two-pair bidirectional Gaussian relay network,” IEEE Trans. Inf. Theory, vol. 58, no. 4, pp. 2434–2454, Apr. 2012. [26] A. Zewail, Y. Mohasseb, M. Nafie, and H. El-Gamal, “The deterministic multicast capacity of 4-node relay networks,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), Istanbul, Turkey, July 2013. [27] B. Matthiesen, A. Zappone, and E. A. Jorswieck, “Resource allocation for energy-efficient 3-way relay channels,” IEEE Trans. Wireless Commun., vol. 14, no. 8, pp. 4454–4468, Aug. 2015. [28] H. Q. Ngo and E. G. Larsson, “Large-scale multipair two-way relay networks with distributed AF beamforming,” IEEE Commun. Lett., vol. 17, no. 12, pp. 1–4, 2013. [29] L. Ong, G. Lechner, S. J. Johnson, and C. M. Kellett, “The three-user Finite-field multi-way relay channel with correlated source,” IEEE Trans. Commun., vol. 61, no. 8, pp. 3125–3135, Aug. 2013. [30] R. Wang and X. Yuan, “MIMO multiway relaying with pairwise data exchange: A degrees of freedom perspective,” IEEE Trans. Signal Process., vol. 62, no. 20, pp. 5294–5307, Oct. 2014. [31] Y. Tian and A. Yener, “Degrees of freedom for the mimo multi-way relay channel,” IEEE Trans. Inf. Theory, vol. 60, no. 5, pp. 2495–2511, May 2014. [32] N. Lee and J. Chun, “Degrees of freedom for the MIMO Gaussian K-way relay channel: Successive network code encoding and decoding,” IEEE Trans. Inf. Theory, vol. 60, no. 3, pp. 1814–1821, Mar. 2014. [33] C. Wang, “Beyond one-way communication: Degrees of freedom of multi-way relay MIMO interference networks,” IEEE Trans. Wireless Commun., vol. 15, no. 10, pp. 7174–7186, Oct. 2016. [34] Z. Cheng, N. Devroye, and T. Liu, “The Degrees of Freedom of full-duplex bidirectional interference networks with and without a MIMO relay,” IEEE Trans. Wireless Commun., vol. 15, no. 4, pp. 2912–2924, April 2016. [35] P. Mach, Z. Becvar, and T. Vanek, “In-Band Device-to-Device Communication in OFDMA Cellular Networks: A Survey and Challenges,” IEEE Commun. Surveys Tuts., vol. 17, no. 4, pp. 1885–1922, Nov. 2015. [36] A. Celik, R. M. Radaydeh, F. S. Al-Qahtani, and M.-S. Alouini, “Joint interference management and resource allocation for Deviceto-Device (D2D) communications underlying Downlink/Uplink Decoupled (DUDe) heterogeneous networks,” in Proc. IEEE Int. Conf. Commun. (ICC), Paris, France, May 2017, to appear. [37] C. Karakus, I. H. Wang, and S. Diggavi, “Gaussian interference channel with intermittent feedback,” IEEE Trans. Inf. Theory, vol. 61, no. 9, pp. 4663–4699, Sep. 2015. [38] I. H. Wang, C. Suh, S. Diggavi, and P. Viswanath, “Bursty interference channel with feedback,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), July 2013, pp. 21–25. [39] A. Vahid, M. A. Maddah-Ali, and A. S. Avestimehr, “Binary fading interference channel with no CSIT,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), June 2014, pp. 666–670. [40] S. Y. Yeh and I. H. Wang, “Degrees of freedom of the bursty MIMO X channel without feedback,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), July 2016, pp. 1312–1316. [41] V. R. Cadambe and S. A. Jafar, “Interference alignment and the degrees of freedom of wireless X networks,” IEEE Trans. Inf. Theory, vol. 55, no. 9, pp. 3893–3908, Sep. 2009. [42] S. A. Jafar and S. Shamai, “Degrees of freedom region for the MIMO X channel,” IEEE Trans. Inf. Theory, vol. 54, no. 1, pp. 151–170, Jan. 2008. [43] A. E. Gamal and Y.-H. Kim, Network Information Theory. Cambridge University Press, 2011. 35 [44] A. Chaaban and A. Sezgin, “The approximate capacity region of the Gaussian Y-Channel via the deterministic approach,” IEEE Trans. Inf. Theory, vol. 61, no. 2, pp. 939–962, Feb. 2015. [45] T. M. Cover and A. El-Gamal, “Capacity theorems for the relay channel,” IEEE Trans. Inf. Theory, vol. IT-25, no. 5, pp. 572–584, Sep. 1979. [46] T. Cover and J. Thomas, Elements of Information Theory (Second Edition). John Wiley and Sons, Inc., 2006. 

7cs.IT

STDP Based Pruning of Connections and Weight Quantization in Spiking Neural Networks for EnergyEfficient Recognition Nitin Rathi, Priyadarshini Panda, Student Member, IEEE, and Kaushik Roy, Fellow, IEEE Abstract— Spiking Neural Networks (SNNs) with a large number of weights and varied weight distribution can be difficult to implement in emerging in-memory computing hardware due to the limitations on crossbar size (implementing dot product), the constrained number of conductance levels in non-CMOS devices and the power budget. We present a sparse SNN topology where non-critical connections are pruned to reduce the network size and the remaining critical synapses are weight quantized to accommodate for limited conductance levels. Pruning is based on the power law weight-dependent Spike Timing Dependent Plasticity (STDP) model; synapses between pre- and post-neuron with high spike correlation are retained, whereas synapses with low correlation or uncorrelated spiking activity are pruned. The weights of the retained connections are quantized to the available number of conductance levels. The process of pruning non-critical connections and quantizing the weights of critical synapses is performed at regular intervals during training. We evaluated our sparse and quantized network on MNIST dataset and on a subset of images from Caltech-101 dataset. The compressed topology achieved a classification accuracy of 90.1% (91.6%) on the MNIST (Caltech-101) dataset with 3.1x (2.2x) and 4x (2.6x) improvement in energy and area, respectively. The compressed topology is energy and area efficient while maintaining the same classification accuracy of a 2-layer fully connected SNN topology. Index Terms— Pruning, Spiking Neural Network, Spike Timing Dependent Plasticity, Unsupervised Learning, Weight Quantization. 1. INTRODUCTION Human brain consisting of 20 billion neurons and 200 trillion synapses is by far the most energy-efficient neuromorphic system with cognitive intelligence. The human brain consumes only ~20W of power which is nine orders of magnitude lower compared to a computer simulating human brain activity in real time [1]. This had led to the inspiration and development of Spiking Neural Networks (SNNs) which tries to mimic the behavior of human brain and process inputs in real time [2]. SNNs may provide an energy-efficient solution to perform neural computing. However, recent works have shown that to get reasonable accuracy compared to nonspiking Artificial Neural Networks (nANNs), the complexity and size of SNNs is enormous. In [3] to improve the classification accuracy for MNIST dataset by 12%, the number of neurons in 2-layer SNN had to be increased by 64x. The authors in [4] achieved an average accuracy of 98.6% for MNIST dataset with two hidden layers consisting of 800 neurons in each layer. The quest of making SNNs larger and deeper for higher accuracy have compromised their energy efficiency and introduced challenges as mentioned below: 1. Large SNNs implemented on emerging memristive crossbar structures [5] are limited by the crossbar size. Large crossbars suffer from supply voltage degradation, noise generated from process variations, and sneak paths [6, 7]. 2. SNNs with numerous synapses involve higher number of computations making them slower and energy inefficient. SNNs are driven by the synaptic events and the total computation, memory, communication, power, area, and speed scale with the number of synapses [2]. We propose a pruned and weight quantized SNN topology with self-taught Spike Timing Dependent Plasticity (STDP) based learning. STDP, in turn, is also used to classify synapses as critical and non-critical. The non-critical synapses are pruned from the network, whereas the critical synapses are retained and weight quantized. Such pruning of connections and weight quantization can lead to their efficient implementations in emerging cross-bar arrays such as resistive random access memories (R-RAMs) [8, 9], magnetic tunnel junctions [10], or domain-wall motion based magnetic devices [11]. Such cross-bars, even though suitable for implementing efficient dot-products required for neural computing, are constrained in size, because of non-idealities such as sneak paths, weight quantization, and parameter variations [6, 7]. The resulting sparse SNN can achieve 2-3x improvement in energy, 2-4x in area and 2-3x in testing speed. Synaptic pruning is commonly observed during the development of human brain. The elimination of synapses begins at the age of two and continues till adulthood, when synaptic density stabilizes and is maintained until old age [12]. From hardware implementation of neural networks, synapses are a costly resource and needs to be efficiently utilized for energy efficient learning. If synapses or connections are properly pruned, the performance decrease due to synaptic deletion is small compared to the energy savings [13]. This has motivated researchers to apply the technique of pruning [14] and weight quantization [15] to compress nANNs. Pruning and quantization performed on state-of-the-art network AlexNet trained for ImageNet dataset provided 7x benefit in energy efficiency along with 35x reduction in synaptic weight storage without any loss of accuracy [15]. The authors in [14] prune the connections of an nANN trained using backpropagation based on the Hessian of the loss function. The number of parameters were reduced Fully Connected membrane potential output spike threshold potential Excitatory Layer Input Layer Compressed Inhibitory Layer w1 refractory period rest potential input spikes reset potential w2 . . . w3 . time Fig. 1. Leaky-Integrate-and-Fire (LIF) model of a single neuron’s membrane potential dynamics in response to input spikes in SNN. by a factor of two while maintaining the same test accuracy. The supervised learning algorithm in [16] pruned the hidden layer neurons with low dominance to reduce network size. The network achieved similar performance with 4x less parameters for Fisher Iris problem compared to other spiking networks. The idea of pruning is based on identifying parameters with small saliency, whose deletion will have minimal effect on the error. These networks were trained with supervised learning algorithms, but in SNNs with unsupervised training it is difficult to calculate such parameters since there is no such defined error function. In real world, obtaining unlabeled images for unsupervised learning is much easier than gathering labelled images for supervised learning. The novelty of our approach lies in selftaught STDP based weight pruning where the connections to be pruned are decided based on their weights learnt by the unsupervised STDP algorithm. Connections having STDP weights above a threshold are considered critical while others are temporarily pruned. The threshold is fixed before training and it referred as pruning threshold. The critical connections are weight quantized to further reduce network complexity. The resulting compressed topology is energy-efficient while maintaining accuracy and alleviates the issues that constrain the scalability of crossbar structure, leading to robust design of neuromorphic systems. The rest of the paper is organized as follows. Section 2 provides background information on the neuron and the synapse models and the STDP learning algorithm employed in this work. The network topology and the training and testing schemes are also briefly discussed. Section 3 presents the proposed compression techniques; STDP based pruning and weight quantization and sharing. The experiments on the proposed topology are presented in Section 4. The results of the experiments are analyzed in Section 5. Conclusions are drawn in Section 6. 2. BACKGROUND 2.1. Neuron & Synapse Model and STDP Learning We employ the Leaky-Integrate-and-Fire (LIF) model [3] to simulate the membrane potential dynamics of a neuron in our spiking network model. Fig. 1 shows the change in membrane Excitatory Layer Input Layer Inhibitory Layer w . . . . . . wn . . . Pruning & . Weight Quantization . . . . . . w w w w . . . . . . . . . . . . w Fig. 2 SNN topology with lateral inhibition. Input to excitatory is fully connected which is later pruned. Excitatory to inhibitory is one-to-one connected, whereas inhibitory is backward connected to all the excitatory except the one it receives the connection from. Pruning is performed only on the input to excitatory connections. potential of a single post-neuron in response to input spikes (blue arrows) from pre-neurons. The membrane potential increases at the onset of a spike and exponentially decays towards rest potential in the absence of spiking activity. The post-neuron fires or emits a spike when its potential crosses the threshold and immediately its potential is set to a reset value. After firing the post-neuron goes into a period of inactivity known as refractory period during which it is abstained from spiking, irrespective of input activity as shown in Fig. 1. The connection between two neurons is termed a synapse and is modelled by the conductance change which is modulated by the synaptic weight (w). The synaptic weight between a pair of neurons increases (decreases) if the post-neuron fires after (before) the pre-neuron has fired. This phenomenon of synaptic plasticity where the weight change is dependent on the inter spike timing of pre- and post-neuron is termed STDP. We adopt the power law weight-dependent STDP model, where the weight change is exponentially dependent on the spike timing difference of the pre- and post-neuron (𝑡𝑝𝑟𝑒 − 𝑡𝑝𝑜𝑠𝑡 ) as well as the previous weight value [3]. 2.2. Network Topology The SNN topology for this work is shown in Fig. 2. It consists of input layer followed by excitatory and inhibitory layer. The input layer is fully connected to the excitatory layer, which in turn is one-to-one connected to the inhibitory layer. The number of neurons in the excitatory layer are varied to achieve better accuracy, whereas the number of neurons in the inhibitory layer is the same as the number in the excitatory layer. Each inhibitory neuron is backward connected to all the excitatory neurons except for the one from which it receives a connection from. Thus, the inhibitory layer provides lateral inhibition which discourages simultaneous firing of multiple excitatory neurons and promotes competition among them to learn different input features. The process of pruning and weight quantization is applied to the excitatory synapses to obtain the compressed topology as shown in Fig. 2. The fully connected topology serves as the baseline design and we compare the results for the compressed design with baseline. To ensure similar firing rates for all neurons in the excitatory layer we employ an adaptive membrane threshold mechanism called homoeostasis [3]. The threshold potential is expressed as 𝑉𝑡ℎ𝑟𝑒𝑠ℎ = 𝑉𝑡 + θ, where 𝑉𝑡 is a constant and θ is changed dynamically. θ increases every time a neuron fires and decays exponentially. If a neuron fires more often, then its threshold potential increases and it requires more inputs to fire again. This ensures that all neurons in the excitatory layer learn unique features and avoids few neurons from dominating the response pattern. 2.3. Training & Testing The connections from input to excitatory layer are trained using the STDP weight update rule to classify an input pattern. The training is unsupervised as we do not use any labels to update the weights. The weight update is given by the formula described in section 3.1. The input image is converted into a Poisson spike train based on individual pixel intensities. The excitatory neurons are assigned a class/label based on their average spiking activity over all the images. During testing, the class prediction is inferred by averaging the response of all excitatory neurons per input. The class represented by the neurons with the highest spiking rate is predicted as the image label. The prediction is correct if the actual label matches the one predicted by the SNN. This is similar to the approach followed in [3]. 3. COMPRESSION TECHNIQUES In this section, we describe the two compression techniques (pruning and weight quantization) employed in this work to convert the 2-layer fully connected SNN into a compact and sparse topology for digit and image recognition. 3.1. STDP Based Pruning Spike Timing Dependent Plasticity (STDP) is widely used as an unsupervised Hebbian training algorithm for SNNs. STDP postulates that the strength of the synapse is dependent on the spike timing difference of the pre- and post-neuron. The power law weight update for an individual synapse is calculated as 𝑤 =  × [𝑒 ( 𝑡𝑝𝑟𝑒 −𝑡𝑝𝑜𝑠𝑡  ) − 𝑜𝑓𝑓𝑠𝑒𝑡] × [𝑤𝑚𝑎𝑥 − 𝑤]𝜇 where 𝑤 is the change in weight,  is the learning rate, 𝑡𝑝𝑟𝑒 and 𝑡𝑝𝑜𝑠𝑡 are the time instant of pre- and post-synaptic spikes,  is the time constant, 𝑜𝑓𝑓𝑠𝑒𝑡 is a constant used for depression, 𝑤𝑚𝑎𝑥 is the maximum constrained imposed on the synaptic weight, 𝑤 is the previous weight value, 𝜇 is a constant which governs the exponential dependence on Decrease in weight (Depression) Increase in weight (Potentiation) - (ms) Fig. 3 Change in synaptic weight based on temporal correlation in pre- and post-synaptic spikes. ( = 0.002,  = 20 ms, 𝑜𝑓𝑓𝑠𝑒𝑡 = 0.4, 𝑤𝑚𝑎𝑥 = 1, 𝑤 = 0.5, 𝜇 = 0.9) previous weight value. The weight update is positive (potentiation) if the post-neuron spikes immediately after the pre-neuron and negative (depression) if the spikes are far apart (Fig. 3). We employ STDP to train the excitatory synapses as well as to classify them as critical or non-critical. The synapses whose weights do not increase for a set of inputs are likely to have not contributed towards learning and thus can be potential candidates for deletion. On the other hand, synapses with higher weights have most likely learned the input pattern and can be classified as critical (provided they were initialized with small weights). The characteristic features of the input is captured in connections with higher weights and are critical for correct classification. Thus, synapses with STDP trained weights ( 𝑤 + 𝑤) above pruning threshold are considered critical and all other synapses are marked as non-critical. The process of pruning and training is performed repeatedly by dividing the entire training set into multiple batches. After each batch the weights of all non-critical synapses are reduced to zero (they still remain in the network) and the network is trained with the next batch. The synapses with zero weight continue to participate in training, thus a non-critical synapse may become critical for different inputs. The process of reducing the weight to zero instead of eliminating the connection is essential for the network to learn the representation of inputs which appear in latter batches. The elimination of synapses will either make the network not learn the new representations or force the network to forget previous representations in order to learn new inputs. The process of retaining non-critical synapses with zero weight makes the network scalable. In the final training step when all the training images have been presented to the network any remaining non-critical connections are permanently removed from the network. The training starts with a fully connected network and the number of critical connections gradually decrease over time. At the end of training only the critical connections capturing the characteristic features of the inputs remain. Fully connected input to excitatory layer Divide training images into N batches. Each batch containing equal number of images Synaptic connections from input to excitatory layer trained using STDP for M batches. (M<N) Check learned synaptic weights. Is weight greater than pruning threshold? Yes No Mark the synapses as noncritical Mark the synapses as critical Quantize the weights of the critical synapses to k-levels (0, w1, w2 ,….., wk-1) w1 = Average of critical weights from 0 to (100/(k-1))th percentile weight. All synapses with critical weights between 0 to (100/(k-1))th percentile weight are assigned with w1 Reduce the weights of noncritical synapses to zero. Thus, temporarily removing them from network. w2 = Average of critical weights from (100/(k-1))th percentile to 2×(100/(k-1))th percentile weight. Similarly, all these synapses share the same weight w2 . . wk-1 = Average of critical weights from (k-1) ×(100/(k-1))th percentile to k×(100/(k-1))th percentile weight. Synapses with weights in this range are assigned w k-1 Yes Anymore training batches remaining? Train the critical and noncritical synapses with STDP for next batch of images. No Training ends. The noncritical synapses are pruned permanently. Sparse topology consists of only the quantized critical synapses Fig. 4 Flowchart of the proposed algorithm for compressing SNN using pruning and weight quantization. 3.2. Weight Sharing and Quantization The process of pruning reduces the overall connectivity, but as mentioned in section 1, SNNs with continuous weight values are difficult to implement in crossbar structures due to limitations on the number of available conductance levels in devices implementing the synapse. Weight sharing and quantization discretizes the weights to the available number of conductance levels. For example, network with 2-level weight quantization has only two values of weights: 0 (no connection) and w. All the synapses share the same weight (w) and the entire network can be represented as a sparse binary matrix. A 2-level weight quantized SNN can be implemented in crossbar architecture with a single fixed resistor [17], where w is the conductance of the resistor. The value of w is the average weight of all the critical connections trained using STDP. For example, we start with a network with n number of synapses and after training (with pruning) m critical synapses remain. The weights of the m critical synapses (w1, w2, ….., wm) are continuous and computed based on the STDP formula. The common weight value w is the average of w1 to wm. The average is calculated after each pruning step and all 1.0 0.8 0.6 (a) (b) (c) 0.4 0.2 (d) (e) (f) 0.0 Fig. 5 Rearranged weights of the connections from input to excitatory for (a) MNIST baseline; (b) MNIST pruning; (c) MNIST pruning and quantization; (d) Caltech 101 baseline; (e) Caltech 101 pruning; (f) Caltech 101 pruning and quantization the critical connections share the same average weight (w1 to wm is replaced with w). Like pruning, the process of weight quantization and sharing is performed repeatedly after each training batch. The value of w changes at every quantization step and the final value is obtained after training the network 5 for all the input batches. Similarly, the weights can be quantized to 3-levels: 0, w1, w2, where w1 (w2) is the low (high) conductance value. The conductance values are computed by calculating the 50th percentile or the median weight of all the critical connections. The lower conductance value w1 is the average of all weights between 0 and the median weight, w2 is the average of rest of the weights. The critical synapses with weights between 0 and the median weight are assigned w1. The critical synapses with weights between median weight and the maximum weight share the quantized vale of w2. The accuracy of the network is directly proportional to the number of quantization levels. The performance of the system improves with more number of conductance levels. In a quantized SNN most of the connections share the same weight which reduces the implementation complexity. Fig. 4 summarizes the proposed algorithm for achieving a pruned and weight quantized SNN. The 2-layer untrained network is initialized with full connectivity from input to excitatory layer. The weights are randomly assigned from a uniform distribution. The training images are divided into N batches of equal number of images. The excitatory synapses are trained with STDP weight update rule for M (M<N) training batches. The connections with current weights above the pruning threshold are classified as critical, rest of the connections are marked as non-critical. The non-critical connections are pruned by reducing their weights to zero. The weights of the critical synapses are quantized to the required number of conductance levels. The pruned and quantized network is trained with STDP weight update rule for the next training batch. The process of pruning and quantization is performed at regular intervals for all the remaining training batches. The training ends when all the batches have been presented to the network. The first pruning and quantization step is delayed for M batches to ensure proper detection of critical connections and to mitigate the bias due to random initialization of weights. The randomly initialized synapses require more training images to capture the input characteristic features. Once the input features have been captured the pruning can be performed more often (after every batch). Once the critical connections are identified, they more or less remain the same during training. So, the first pruning step is very crucial and more than one training batch is needed to identify the critical synapses. The baseline design is trained in a similar fashion with no pruning and quantization. All the training images are presented in one batch and the weights are trained using STDP. baseline baseline (a) (b) baseline baseline (c) (d) (e) (f) Fig. 6 Variation in network connectivity with pruning threshold for (a) MNIST; (b) Caltech 101. Classification accuracy for different network connectivity for (c) MNIST; (d)Caltech 101. Classification accuracy for different number of excitatory neurons for (e) MNIST; and (f) Caltech 101. 4. EXPERIMENTAL METHODOLOGY The proposed SNN topology is simulated in the open source spiking neuron simulator BRIAN implemented in Python [18]. BRIAN allows the modelling of biologically plausible neurons and synapses defined by differential equations. The parameters for the models are same as [3]. We tested our network for digit recognition on the MNIST dataset [19] and image recognition on a subset of images from the Caltech 101 dataset [20]. A. MNIST Dataset MNIST dataset contains 28×28-pixel sized grayscale images of digits 0-9. Thus, the input layer has 784 (28×28) neurons fully connected with 100 excitatory neurons. The dataset is divided into 60,000 training and 10,000 testing images. We further divide the 60,000 training images into batches of 5,000 images (N=12). The baseline design is trained with entire 60,000 images presented one after another. The compressed topology is initially trained for three training batches totaling 15,000 images (M=3). STDP based critical connections are weight quantized and the weights of the non-critical connections is reduced to zero. The pruned and quantized network is trained with the next training batch. The process of pruning and quantization is performed after every batch henceforth. The rearranged input to excitatory synaptic weights of the trained baseline topology with 100 excitatory neurons is shown in Fig. 5(a). Fig. 5(b) shows the rearranged (a) (b) Fig. 7 Classification accuracy for different network sparsity achieved by pruning the connections during training and before training for (a) MNIST; and (b) Caltech 101. synaptic weights of the same network compressed with pruning threshold=0.3 and having continuous weight values. The rearranged synaptic weights of the pruned and 2-level weight quantized network is shown in Fig. 5(c). B. Caltech 101 Dataset Caltech 101 dataset is a collection of images of objects belonging to 101 different categories. Each category consists of 40 to 800 of around 300×200-pixel sized RGB images. The dataset also provides annotations for the object in the image which we use to separate the object from the background. Unlike MNIST images, we preprocess the Caltech 101 images to obtain 28×28-pixel sized grayscale images. Maintaining the same image size across datasets ensures that we do not need to change the network parameters. Out of 101, we selected 10 categories (yin yang, saxophone, stop sign, wrench, revolver, Buddha, airplanes, pigeon, motorbikes, umbrella) and randomly divided the total images in each category with 80% training and 20% testing images. Since each category has different number of images we create copies of images so that each category has similar number of training and testing images. This is necessary to avoid categories with more images to dominate the learning in the network. The preprocessing steps involved converting the images to grayscale, averaging the pixels with Gaussian kernel of size 3×3 to suppress the noise and resizing the image to 28×28 pixels. All the preprocessing steps are performed using the OpenCV library [21] in python. The training set consists of 10,000 images with 1000 images per category. The 10,000 images are further divided into batches of 500 images (N=20). The baseline fully connected design is trained with entire 10,000 images. The compressed topology is initially trained with ten training batches totaling 5000 images (M=10). The critical connections are identified using STDP and weight quantized. The non-critical synapses are pruned. The pruned and quantized network is trained with all the remaining batches with pruning and quantization performed after every training batch. Fig. 5(d), (e) and (f) show the rearranged synaptic weights for the baseline, pruned and weight quantized topologies, respectively. Compression is performed with pruning threshold of 0.2 and 2-level weight quantization. 5. RESULTS & ANALYSES In this section, we analyze the results and compare the performance of compressed topology with the baseline design. The results are evaluated based on different parameters like pruning threshold and number of excitatory neurons. The removal of connections during training is compared with training a sparse network, both having similar connectivity. 5.1. Comparison with varying pruning threshold The network connectivity is a strong function of the pruning threshold; higher the threshold, sparser is the network. The network connectivity is defined as the ratio of the actual number of connections to the total number of possible connections. The total number of possible connections with 100 excitatory neurons is 78400 (784×100). The number of actual connections depend on the pruning steps. Fig. 6(a) and (b) shows the variation in final network connectivity with pruning threshold for MNIST and Caltech 101 datasets, respectively. The red dot in Fig. 6(a) and (b) with zero pruning threshold denotes the baseline design with no compression techniques applied. Ideally, the connectivity should be 1 since the connections are not pruned during training. The reduction in connectivity results from the inherent depression in the STDP learning rule. The further reduction in connectivity is achieved by increasing the pruning threshold. The compressed topologies are less sparse for low pruning threshold compared to baseline. This is due to weight quantization and sharing in early training stages. The shared weight is the average of all critical weights which is higher than almost half the critical weights. Thus, the average weight replaces half the STDP learnt weights which were supposed to be much lower. This reduces the effect of inherent STDP depression on these synapses and reduces the probability of their removal. Fig. 6(c) and (d) show the test accuracy for different network connectivity for MNIST and Caltech 101 datasets, respectively. The baseline topology has an accuracy of 81.6% (MNIST) and 84.2% (Caltech 101) which is consistent with the results shown in [3]. The highest 5.2. Comparison with varying number of neurons The change in classification accuracy with the number of excitatory neurons for MNIST and Caltech 101 datasets is shown in Figs. 6(e) and 6(f), respectively. The compressed topology has 27% connectivity for MNIST and 42% for Caltech 101 datasets, also the weights are quantized to 3levels. These parameters correspond to the best performance of compressed topology (Fig. 6(c-d)). The baseline design with 6400 neurons achieved an accuracy of 93.2% for MNIST and 94.2% for Caltech 101 datasets. The pruned topology achieved an accuracy of 91.5% with 8% connectivity and 92.8% with 12% connectivity for MNIST and Caltech 101 datasets, respectively. 5.3. Pruning while training The objective of pruning is to increase the sparsity in the network. This can be achieved in two ways: removing connections during training of a fully connected network or training a sparse network. The second approach is performed by randomly removing connections from a fully connected topology to produce a sparse network. In first case the connections are removed systematically based on some parameters whereas in the second approach the removal is completely random. Nevertheless, in both the cases the final network connectivity is same. Fig. 7(a) and (b) shows the classification accuracy with varying network connectivity for both the approaches for MNIST and Caltech 101 datasets, respectively. The network with initial pruning is trained similar to baseline with no compression techniques. The pruning while training is the approach followed in rest of the paper, where pruning is performed at regular intervals during training. The results for both the networks are shown for continuous weight distribution. Pruning the connections during training performs better since only the non-critical connections are removed. The network with initial sparsity is constructed by randomly removing connections. This shows that STDP successfully identifies the non-critical connections. 5.4. Reduction in spike count or energy The decrease in connectivity due to pruning leads to reduced spiking activity in the excitatory layer. The active power of a SNN is proportional to the firing activity in the network [2]. Thus, the energy can be quantified as the reduction in spike count of excitatory neurons during testing. Fig. 8 shows the normalized reduction in spiking activity or energy for compressed topology with respect to baseline. The pruned topology shows 3.1x and 2.2x improvement in energy Normalized Energy classification accuracy achieved for the compressed topology is 79.5% (MNIST) and 82.8% (Caltech 101). The accuracy degrades slightly compared to baseline but at the same time there is immense drop in network connectivity. The compressed topology is 75% (36%) sparser than the baseline topology for MNIST (Caltech 101) dataset. MNIST Caltech 101 Fig. 8 Normalized improvement in energy with pruning and weight quantization compared to baseline topology. whereas the 2-level weight quantized network achieves 2.4x and 1.92x improvement for MNIST and Caltech101 datasets, respectively. The compressed topology may achieve additional energy benefits from implementation in crossbar structures with low power devices. The emerging post-CMOS devices like MTJ, R-RAM and domain wall motion based devices consume very low power in idle state due to elimination of leakage. But these devices have limited number of programmable conductance levels. The compressed topology quantized to the available number of conductance levels can reap the energy benefits provided by these devices. The baseline design with continuous weight distribution is difficult to implement with these devices. 6. CONCLUSIONS In this paper, we propose two compression techniques, pruning and weight quantization to compress SNNs. Compressed SNNs not only provide energy benefits but also mitigate the issue of limited programmable conductance levels of post-CMOS devices for neuromorphic implementation. The novelty of our approach lies in fact that STDP learning rule is used to decide the network pruning and the weights of the critical connections are quantized to specific levels depending on device and technology requirements. The compressed topology is compared with the 2-layer fully connected topology for digit recognition with MNIST dataset and image recognition with Caltech 101 dataset. The proposed topology achieves 3.1x and 2.2x improvement in energy for MNIST and Caltech 101 datasets, respectively, compared to baseline fully connected SNN. The optimal compression parameters like pruning threshold and weight quantization levels are decided by performing multiple experiments with different images. Additionally, it is worth mentioning that the proposed topology reduced the training time by 3x and 2x for MNIST and Caltech 101 datasets, respectively, by achieving faster training convergence. ACKNOWLEDGEMENT The research was funded in part by National Science Foundation, Intel Corporation, Vannevar Bush Faculty Fellowship, and by the Center for Spintronics (C-SPIN) funded by DARPA/MARCO. References [1] D. S. Modha, “Introducing a brain-inspired computer,” Published online at http://www.research.ibm.com/articles/brain-chip.shtml, 2017. [2] P. A. Merolla, J. V. Arthur, R. Alvarez-Icaza, A. S. Cassidy, J. Sawada, F. Akopyan, B. L. Jackson, N. Imam, C. Guo, Y. Nakamura et al., “A million spiking-neuron integrated circuit with a scalable communication network and interface,” Science, vol. 345, no. 6197, pp. 668–673, 2014. [3] P. U. Diehl and M. Cook, “Unsupervised learning of digit recognition using spike-timing-dependent plasticity,” Frontiers in computational neuroscience, vol. 9, 2015. [4] J. H. Lee, T. Delbruck, and M. Pfeiffer, “Training deep spiking neural networks using backpropagation,” Frontiers in neuroscience, vol. 10, 2016. [5] P. Merolla, J. Arthur, F. Akopyan, N. Imam, R. Manohar, and D. S. Modha, “A digital neurosynaptic core using embedded crossbar memory with 45pj per spike in 45nm,” in Custom Integrated Circuits Conference (CICC), 2011 IEEE. IEEE, 2011, pp. 1–4. [6] B. Liu, W. Wen, Y. Chen, X. Li, C.R. Wu, and T.Y. Ho, “Eda challenges for memristor-crossbar based neuromorphic computing,” in Proceedings of the 25th edition on Great Lakes Symposium on VLSI. ACM, 2015, pp. 185–188. [7] E. Linn, R. Rosezin, C. Kugeler, and R. Waser, “Complementary resistive switches for passive nanocrossbar memories,” Nature materials, vol. 9, no. 5, p. 403, 2010. [8] M. Hu, H. Li, Y. Chen, Q. Wu, G. S. Rose, and R. W. Linderman, “Memristor crossbar-based neuromorphic computing system: A case study,” IEEE transactions on neural networks and learning systems, vol. 25, no. 10, pp. 1864–1878, 2014. [9] S. Park, H. Kim, M. Choo, J. Noh, A. Sheri, S. Jung, K. Seo, J. Park, S. Kim, W. Lee et al., “Rram-based synapse for neuromorphic system with pattern recognition function,” in Electron Devices Meeting (IEDM), 2012 IEEE International. IEEE, 2012, pp. 10–2. [10] P. Krzysteczko, J. Munchenberger, M. Schafers, G. Reiss, and A. Thomas, “The memristive magnetic tunnel junction as a nanoscopic synapse-neuron system,” Advanced Materials, vol. 24, no. 6, pp. 762– 766, 2012. [11] M. Sharad, C. Augustine, G. Panagopoulos, and K. Roy, “Spinbased neuron model with domain-wall magnets as synapse,” IEEE Transactions on Nanotechnology, vol. 11, no. 4, pp. 843–853, 2012. [12] P. R. Huttenlocher et al., “Synaptic density in human frontal cortex- developmental changes and effects of aging,” Brain Res, vol. 163, no. 2, pp. 195–205, 1979. [13] G. Chechik, I. Meilijson, and E. Ruppin, “Synaptic pruning in development: a computational account,” Neural computation, vol. 10, no. 7, pp. 1759–1777, 1998. [14] Y. LeCun, J. S. Denker, and S. A. Solla, “Optimal brain damage,” in Advances in Neural Information Processing Systems 2, 1990, pp. 598–605. [Online]. Available: http://papers.nips.cc/paper/250-optimalbrain-damage.pdf [15] S. Han, H. Mao, and W. J. Dally, “Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding,” arXiv preprint arXiv:1510.00149, 2015. [16] S. Dora, S. Sundaram, and N. Sundararajan, “A two stage learning algorithm for a growing-pruning spiking neural network for pattern classification problems,” in International Joint Conference on Neural Networks (IJCNN), 2015. IEEE, 2015, pp. 1–7. [17] H. Graf, L. Jackel, R. Howard, B. Straughn, J. Denker, W. Hubbard, D. Tennant, and D. Schwartz, “Vlsi implementation of a neural network memory with several hundreds of neurons,” in AIP conference proceedings, vol. 151, no. 1. AIP, 1986, pp. 182–187. [18] D. Goodman and R. Brette, “Brian: a simulator for spiking neural networks in python,” Frontiers in neuroinformatics, vol. 2, 2008. [19] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradientbased learning applied to document recognition,” Proceedings of the IEEE, vol. 86, no. 11, pp. 2278–2324, 1998. [20] L. Fei-Fei, R. Fergus, and P. Perona, “One-shot learning of object categories,” IEEE transactions on pattern analysis and machine intelligence, vol. 28, no. 4, pp. 594–611, 2006. [21] G. Bradski, “The opencv library.” Dr. Dobb’s Journal: Software Tools for the Professional Programmer, vol. 25, no. 11, pp. 120–123, 2000. 

9cs.NE

arXiv:0807.1835v4 [math.AC] 2 Sep 2015 CONTENT ALGEBRAS OVER COMMUTATIVE RINGS WITH ZERO-DIVISORS PEYMAN NASEHPOUR A BSTRACT. Let M be an R-module and c the function from M to the ideals of R defined by c(x) = ∩{I : I is an ideal of R and x ∈ IM}. M is said to be a content R-module if x ∈ c(x)M, for all x ∈ M. B is called a content R-algebra, if it is a faithfully flat and content R-module and it satisfies the Dedekind-Mertens content formula. In this article, we prove some new results for content modules and algebras by using ideal theoretic methods. C ONTENTS 1. Introduction 2. Content Modules and Algebras 3. Prime Ideals in Content Algebras 4. Rings and modules having few zero-divisors 5. Gaussian and Armendariz Algebras 6. Nilradical and Jacobson radical of content algebras 7. Acknowledgment References 1 3 7 8 10 12 14 14 1. I NTRODUCTION This preprint is a collection of some of the definitions and results that the author was working on them during his studentship at the Department of Mathematics and Computer Science at the University of Osnabrück. Though this preprint was never submitted for publication, but its definitions and results were reproduced and used in a couple of different papers composed by the author and were published in different journals. Since the author has cited to this preprint in some of his papers, it was necessary to update and edit it for the convenience of the readers of his papers. Throughout this paper, all rings are commutative with unit and all modules are assumed to be unitary. In this paper, we discuss the ideal theoretic properties of some special algebras called content algebras. This concept stems from Dedekind-Mertens content 1 2 PEYMAN NASEHPOUR formula for polynomial rings. For doing this, we need to know about content modules introduced in [22], which in the next section, we introduce them and then we start the main theme of the paper that is on content algebras over rings with zero-divisors. Our main goal is to show that a couple of ideal theoretic results of polynomial rings, also hold for content algebras. The class of content modules are themselves considerable and interesting in the field of module theory. For example, all projective modules are content and a kind of Nakayama lemma holds for content modules. Let R be a commutative ring with identity, and M a unitary R-module and the content function, c from M to the ideals of R defined by c(x) = \ {I : I is an ideal of R and x ∈ IM}. M is called a content R-module if x ∈ c(x)M, for all x ∈ M. In Section 2, we prove that if M is a content R-module and Jac(R) is the Jacobson radical of R and I, an ideal of R such that I ⊆ Jac(R), then IM = M implies M = (0). Also we introduce content and weak content algebras and mention some of their basic properties that we need them in the rest of the paper for the convinience of the reader. Let R be a commutative ring with identity and R′ an R-algebra. R′ is defined to be a content R-algebra, if the following conditions hold: (1) R′ is a content R-module. (2) (Faithfully flatness) For any r ∈ R and f ∈ R′ , the equation c(r f ) = rc( f ) holds and c(R′ ) = R. (3) (Dedekind-Mertens content formula) For each f and g in R′ , there exists a natural number n such that c( f )n c(g) = c( f )n−1 c( f g). In this section, also we prove that if R is a ring and S, a commutative monoid, then the monoid ring B = R[S] is a content R-algebra if and only if one of the following conditions satisfies: (1) For f , g ∈ B, if c( f ) = c(g) = R, then c( f g) = R. (2) (McCoy’s Property) For g ∈ B, g is a zero-divisor of B iff there exists r ∈ R − {0} such that rg = 0. (3) S is a cancellative and torsion-free monoid. In Section 3, we discuss about prime ideals of content and weak content algebras (Cf. [23]) and we show that in content extensions, minimal primes extend to minimal primes. More precisely, if B is a content R-algebra, then there is a correspondence between Min(R) and Min(B), with the function ϕ : Min(R) −→ Min(B) defined by p −→ pB. CONTENT ALGEBRAS OVER COMMUTATIVE RINGS WITH ZERO-DIVISORS 3 In Section 4, we introduce a family of rings and modules who have very few zerodivisors. It is a well-known result that the set of zero-divisors of a finitely generated module over a Noetherian ring is a finite union of the associated primes of the module [16, p. 55]. Rings having few zero-divisors have been introduced in [8]. We define that a ring R has very few zero-divisors, if Z(R) is a finite union of prime ideals in Ass(R). In this section, we prove that if R is a ring that has very few zero-divisors and B is a content R-algebra, then B has very few zero-divisors also. Another celebrated property of Noethering rings is that every ideal entirely contained in the set of its zero-divisors has a nonzero annihilator. A ring R has Property (A), if each finitely generated ideal I ⊆ Z(R) has a nonzero annihilator [15]. In Section 5, also we prove some results for content algebras over rings having Property (A) and then we discuss on rings and modules having few zero-divisors in more details. In Section 5, we discuss Gaussian and Armendariz content algebras that are natural generalization of the same concepts in polynomials rings. In this section we show that if B is a content R-algebra, then B is a Gaussian R-algebra iff for any ideal I of R, B/IB is an Armendariz (R/I)-algebra. This is a generalization of a result in [3]. In Section 6, we prove some results about Nilradical and Jacobson radical of content algebras, i.e., statements about content algebras over domainlike and presimplifiable rings. Also we show some results similar to what we have for the ring R(X ) = R[X ]S, where S = { f ∈ R[X ] : c( f ) = R}. Some of the results of the present paper can be generalized to monoid modules. Whenever it is possible, we bring those results, though their similar proofs are omitted. Unless otherwise stated, our notation and terminology will follow as closely as possible that of Gilmer [11]. 2. C ONTENT M ODULES AND A LGEBRAS In this section, first we give the definition of content modules and prove Nakayama lemma for them and then we introduce content and weak content algebras with some results that we need in the rest of the paper for the convinience of the reader. More on content modules and algebras can be found in [22] and [23]. Definition 1. Let R be a commutative ring with identity, and M a unitary R-module and the content function, c from M to the ideals of R defined by c(x) = \ {I : I is an ideal of R and x ∈ IM}. 4 PEYMAN NASEHPOUR M is called a content R-module if x ∈ c(x)M, for all x ∈ M, also when N is a non-empty subset of M, then by c(N) we mean the ideal generated by all c(x) that x ∈ N. Theorem 2. Nakayama Lemma for Content Modules: Let M be a content R-module and Jac(R) be the Jacobson radical of R and I be an ideal of R such that I ⊆ Jac(R). If IM = M, then M = (0). Proof. Let x ∈ M. Since M is a content R-module, x ∈ c(x)M, but IM = M, so x ∈ c(x)IM and therefore c(x) ⊆ c(x)I, but c(x) is a finitely generated ideal of R [22, 1.2, p. 51], so by Nakayama lemma for finitely generated modules, c(x) = (0) and consequently x = 0.  Corollary 3. Let P be a projective R-module and Jac(R) be the Jacobson radical of R and I be an ideal of R such that I ⊆ Jac(R). If IP = P, then P = (0). Proof. Any projective module is a content module [22, Corollary 1.4].  Lemma 4. Let M be an R-module. The following statements are equivalent: (1) M is a content R-module, i.e. x ∈ c(x)M, for all x ∈ M. T T (2) For any non-empty family of ideals {Ii } of R, (Ii )M = (Ii M). Moreover when M is a content R-module, c(x) is a finitely generated ideal of R, for all x ∈ M. Theorem 5. Let M be a content R-module. Then the following are equivalent: (1) M is flat. (2) For every r ∈ R and x ∈ M, rc(x) = c(rx). Moreover, M is faithfully flat iff M is flat and c(M) = R. Proof. Proof at [22, Corollary 1.6, p. 53] and [22, Remark 2.3(d), p. 56].  The application of the above theorem will appear in the next section on content algebras. Also with the help of the above theorem, we will describe some of the prime and primary submodules of faithfully flat and content modules. Definition 6. Let M be an R-module and P be a proper R-submodule of M. P is said to be a prime submodule of M, if rx ∈ P implies x ∈ P or rM ⊆ P, for each r ∈ R and x ∈ M. Definition 7. Let M be an R-module and P be a proper R-submodule of M. P is said to be a primary submodule of M, if rx ∈ P then x ∈ P or there exists a natural number n such that rn M ⊆ P, for each r ∈ R and x ∈ M. Theorem 8. Let M be a content and faithfully flat R-module and p be an ideal of R. Then pM is a primary (prime) R-submodule of M iff p is a primary (prime) ideal of R. CONTENT ALGEBRAS OVER COMMUTATIVE RINGS WITH ZERO-DIVISORS 5 Proof. Let p be a prime ideal of R and r ∈ R and x ∈ M such that rx ∈ pM. Therefore c(rx) ⊆ p and since c(rx) = rc(x) we have rc(x) ⊆ p and this means that c(x) ⊆ p or (r) ⊆ p and at last x ∈ pM or rM ⊆ pM. Notice that since M is a faithfully flat R-module, pM 6= M. The other assertions can be proved in a similar way.  Content algebras and later weak content algebras were introduced and discussed in [22] and [23] respectively. Content algebras are actually a natural generalization of (almost) polynomial rings [9]. Let R be a commutative ring with identity. For f ∈ R[X ], the content of f , denoted by c( f ), is defined as the R-ideal generated by the coefficients of f . One can easily check that c( f g) ⊆ c( f )c(g) for the two polynomials f , g ∈ R[X ] and may ask when the equation c( f g) = c( f )c(g) holds. Tsang, a student of Kaplansky, proved that if D is an integral domain and c( f ), for f ∈ D[X ], is an invertible ideal of D, then c( f g) = c( f )c(g), for all g ∈ D[X ]. Tsang’s guess was that the converse was true and the correctness of her guess was completely proved some decades later [17]. Though the equation c( f g) = c( f )c(g) does not hold always, a weaker formula always holds that is called the Dedekind-Mertens content formula [4]. Theorem 9. Dedekind-Mertens Lemma. Let R be a ring. For each f and g in R[X ], there exists a natural number n such that c( f )n c(g) = c( f )n−1 c( f g). For a history of Dedekind-Mertens lemma, refer to [13] and its combinatorial proof, refer to [6, Corollary 2]. Now we bring the definition of content algebras from [22, 6, p. 63]: Definition 10. Let R be a commutative ring with identity and R′ an R-algebra. R′ is defined to be a content R-algebra, if the following conditions hold: (1) R′ is a content R-module. (2) (Faithfully flatness) For any r ∈ R and f ∈ R′ , the equation c(r f ) = rc( f ) holds and c(R′ ) = R. (3) (Dedekind-Mertens content formula) For each f and g in R′ , there exists a natural number n such that c( f )n c(g) = c( f )n−1 c( f g). A good example of a content R-algebra is the group ring R[G] where G is a torsion-free abelian group [19]. This is actually a free R-module. For some examples of content Ralgebras that as R-modules are not free, one can refer to [22, Examples 6.3, p. 64]. Rush defined weak content algebras as follows [23, p. 330]: Definition 11. Let R be a commutative ring with identity and R′ an R-algebra. R′ is defined to be a weak content R-algebra, if the following conditions hold: (1) R′ is a content R-module. 6 PEYMAN NASEHPOUR (2) (Weak content formula) For all f and g in R′ , c( f )c(g) ⊆ rad(c( f g)) (Here rad(A) denotes the radical of the ideal A). Also he gave an equivalent condition for when an algebra that is a content module is a weak content algebra [23, Theorem 1.2, p. 330]: Theorem 12. Let R′ be an R-algebra such that R′ is a content R-module. The following are equivalent: (1) R′ is a weak content R-algebra. (2) For each prime ideal p of R, either pR′ is a prime ideal of R′ , or pR′ = R′ . It is obvious that content algebras are weak content algebras, but the converse is not true. For more on this interesting topic, refer to [10, Example 4.1]. We end our introductory section with the following results: Theorem 13. Let R be a ring and S be a commutative monoid. Then the following statements about the monoid algebra B = R[S] are equivalent: (1) (2) (3) (4) B is a content R-algebra. B is a weak content R-algebra. For f , g ∈ B, if c( f ) = c(g) = R, then c( f g) = R. (McCoy’s Property) For g ∈ B, g is a zero-divisor of B iff there exists r ∈ R − {0} such that rg = 0. (5) S is a cancellative and torsion-free monoid. Proof. (1) → (2) → (3) and (1) → (4) are obvious ([22] and [23]). Also according to [19] (5) implies (1). Therefore the proof will be complete if we prove that (3) and also (4) implies (5). (3) → (5): We prove that if S is not cancellative nor torsion-free then (3) cannot hold. For the moment, suppose that S is not cancellative, so there exist s,t, u ∈ S such that s +t = s + u while t 6= u. Put f = X s and g = (X t − X u ). Then obviously c( f ) = c(g) = R, while c( f g) = (0). Finally suppose that S is cancellative but not torsion-free. Let s,t ∈ S be such that s 6= t, while ns = nt for some natural n. Choose the natural number k minimal (k−i−1)s+it ). so that ns = nt. Then we have 0 = X ks − X kt = (X s − X t )(∑k−1 i=0 X Since S is cancellative, the choice of k implies that (k −i1 −1)s+i1t 6= (k −i2 −1)s+i2t (k−i−1)s+it 6= 0, and this completes the proof. for 0 ≤ i1 < i2 ≤ k − 1. Therefore ∑k−1 i=0 X In a similar way one can prove (4) → (5) [12, p. 82].  Remark 14. Let S be a commutative monoid and M be a nonzero R-module. It is trivial that M[S] is an R[S]-module. Let g ∈ M[S] and put g = m1 s1 + m2 s2 + · · · + mn sn , where m1 , . . ., mn ∈ M and s1 , . . . , sn ∈ S. We define the content of g to be the R-submodule of M CONTENT ALGEBRAS OVER COMMUTATIVE RINGS WITH ZERO-DIVISORS 7 generated by the coefficients of g, i.e. c(g) = (m1 , . . . , mn ). The following statements are equivalent: (1) S is a cancellative and torsion-free monoid. (2) For all f ∈ R[S] and g ∈ M[S], there exists a natural number k such that c( f )k c(g) = c( f )k−1 c( f g). (3) (McCoy’s Property) For all f ∈ R[S] and g ∈ M[S] − {0}, if f g = 0, then there exists an m ∈ M − {0} such that f · m = 0. Proof. (1) → (2) and (2) → (3) have been proved in [19] and [18] respectively. For (3) → (1) use the technique in the previous theorem.  3. P RIME I DEALS IN C ONTENT A LGEBRAS Let B be a weak content R-algebra such that for all m ∈ Max(R) (by Max(R), we mean the maximal ideals of R), we have mB 6= B, then by Theorem 12, prime ideals extend to prime ideals. Particularly in R-content algebras - that are faithfully flat R-modules by definition - primes extend to primes. We recall that when B is a content R-algebra, then g is a zero-divisor of B, iff there exists an r ∈ R − {0} such that rg = 0 [22, 6.1, p. 63]. Now we give the following theorem about associated prime ideals. We assert that by AssR (M), we mean the associated prime ideals of R-module M. Theorem 15. Let B be a content R-algebra and M a nonzero R-module. If p ∈ AssR (M) then pB ∈ AssB (M ⊗R B). Proof. Let p ∈ AssR (M), therefore 0 −→ R/p −→ M is an R-exact sequence. Since B is a faithfully flat R-module, we have the following B-exact sequence: 0 −→ B/pB −→ M ⊗R B with pB = Ann(x ⊗R 1B ). Since B is a content R-algebra, pB is a prime ideal of B.  We give a general theorem on minimal prime ideals in algebras. One of the results of this theorem is that in faithfully flat weak content algebras (including content algebras), minimal prime ideals extend to minimal prime ideals and more precisely, there is actually a correspondence between the minimal prime ideals of the ring and their extensions in the algebra. Theorem 16. Let B be an R-algebra with the following properties: (1) For each prime ideal p of R, the extended ideal pB of B is prime. (2) For each prime ideal p of R, pB ∩ R = p. Then the function ϕ : Min(R) −→ Min(B) given by p −→ pB is a bijection. 8 PEYMAN NASEHPOUR Proof. First we prove that if p is a minimal prime ideal of R, then pB is also a minimal prime ideal of B. Let Q be a prime ideal of B such that Q ⊆ pB. So Q ∩ R ⊆ pB ∩ R = p. Since p is a minimal prime ideal of R, we have Q ∩ R = p and therefore Q = pB. This means that ϕ is a well-defined function. Obviously the second condition causes ϕ to be one-to-one. The next step is to prove that ϕ is onto. For showing this, consider Q ∈ Min(B), so Q∩R is a prime ideal of R such that (Q∩R)B ⊆ Q and therefore (Q∩R)B = Q. Our claim is that (Q ∩ R) is a minimal prime ideal of R. Suppose p is a prime ideal of R such that p ⊆ Q ∩ R, then pB ⊆ Q and since Q is a minimal prime ideal of B, pB = Q = (Q ∩ R)B and therefore p = Q ∩ R.  Corollary 17. Let B be a weak content and faithfully flat R-algebra, then the function ϕ : Min(R) −→ Min(B) given by p −→ pB is a bijection. Proof. Since B is a weak content and faithfully flat R-algebra, then for each prime ideal p of R, the extended ideal pB of B is prime and also c(1B ) = R by [22, Corollary 1.6] and [23, Theorem 1.2]. Now consider r ∈ R, then c(r) = c(r · 1B) = r · c(1B ) = (r). Therefore  if r ∈ IB ∩ R, then (r) = c(r) ⊆ I. Thus for each prime ideal p of R, pB ∩ R = p. Corollary 18. Let R be a Noetherian ring. Then ϕ : Min(R) −→ Min(R[[X1, . . . , Xn ]]) given by p −→ p.(R[[X1, . . . , Xn]]) is a bijection. 4. R INGS AND MODULES HAVING FEW ZERO - DIVISORS For a ring R, by Z(R), we mean the set of zero-divisors of R. In [8], it has been defined that a ring R has few zero-divisors, if Z(R) is a finite union of prime ideals. We present the following definition to prove some other theorems related to content algebras. Definition 19. A ring R has very few zero-divisors, if Z(R) is a finite union of prime ideals in Ass(R). Theorem 20. Let R be a ring that has very few zero-divisors. If B is a content R-algebra, then B has very few zero-divisors also. Proof. Let Z(R) = p1 ∪ p2 ∪ · · · ∪ pn , where pi ∈ AssR (R) for all 1 ≤ i ≤ n. We will show that Z(B) = p1 B ∪ p2 B ∪ · · · ∪ pn B. Let g ∈ Z(B), so there exists an r ∈ R − {0} such that rg = 0 and so rc(g) = (0). Therefore c(g) ⊆ Z(R) and this means that c(g) ⊆ p1 ∪ p2 ∪ · · · ∪ pn and according to Prime Avoidance Theorem, we have c(g) ⊆ pi , for some 1 ≤ i ≤ n and therefore g ∈ pi B. Now let g ∈ p1 B ∪ p2 B ∪ · · · ∪ pn B so there exists an i such that g ∈ pi B, so c(g) ⊆ pi and c(g) has a nonzero annihilator and this means that  g is a zero-divisor of B. Note that pi B ∈ AssB (B), for all 1 ≤ i ≤ n. Remark 21. Let R be a ring and consider the following three conditions on R: CONTENT ALGEBRAS OVER COMMUTATIVE RINGS WITH ZERO-DIVISORS 9 (1) R is a Noetherian ring. (2) R has very few zero-divisors. (3) R has few zero-divisors. Then, (1) → (2) → (3) and none of the implications are reversible. Proof. For (1) → (2) use [16, p. 55]. It is obvious that (2) → (3). Suppose k is a field, A = k[X1 , X2, X3 , . . ., Xn , . . .] and m = (X1, X2 , X3, . . . , Xn , . . .) and at last I = (X12, X22 , X32, . . . , Xn2, . . .). Since A is a content k-algebra and k has very few zerodivisors, A has very few zero-divisors while it is not a Noetherian ring. Also consider the ring R = A/I. It is easy to check that R is a quasi-local ring with the only prime ideal m/I and Z(R) = m/I and finally m/I ∈ / AssR (R). Note that AssR (R) = 0. /  Now we bring the following definition from [15] and prove some other results for content algebras. Definition 22. A ring R has Property (A), if each finitely generated ideal I ⊆ Z(R) has a nonzero annihilator. Let R be a ring. If R has very few zero-divisors (for example if R is Noetherian), then R has Property (A) [16, Theorem 82, p. 56], but there are some non-Noetherian rings which have not Property (A) [16, Exercise 7, p. 63]. The class of non-Noetherian rings having Property (A) is quite large [14, p. 2]. Theorem 23. Let B be a content R-algebra such that R has Property (A). Then T (B) is a content T (R)-algebra, where by T (R), we mean total quotient ring of R. Proof. Let S′ = B − Z(B). If S = S′ ∩ R, then S = R − Z(R). We prove that if c( f ) ∩ S = 0, / ′ then f 6∈ S . In fact when c( f ) ∩ S = 0, / then c( f ) ⊆ Z(R) and since R has Property (A), c( f ) has a nonzero annihilator. This means that f is a zero-divisor of B and according to [22, Theorem 6.2, p. 64], the proof is complete.  Theorem 24. Let B be a content R-algebra such that the content function c : B −→ FId(R) is onto, where by FId(R), we mean the set of finitely generated ideals of R. The following statements are equivalent: (1) R has Property (A). (2) For all f ∈ B, f is a regular member of B iff c( f ) is a regular ideal of R. Proof. (1) → (2): Let R has Property (A). If f ∈ B is regular, then for all nonzero r ∈ R, r f 6= 0 and so for all nonzero r ∈ R, rc( f ) 6= (0), i.e. Ann(c( f )) = (0) and according to the definition of Property (A), c( f ) 6⊆ Z(R). This means that c( f ) is a regular ideal of R. Now let c( f ) be a regular ideal of R, so c( f ) 6⊆ Z(R) and therefore Ann(c( f )) = (0). This 10 PEYMAN NASEHPOUR means that for all nonzero r ∈ R, rc( f ) 6= (0), hence for all nonzero r ∈ R, r f 6= 0. Since B is a content R-algebra, f is not a zero-divisor of B. (2) → (1): Let I be a finitely generated ideal of R such that I ⊆ Z(R). Since the content function c : B −→ FId(R) is onto, there exists an f ∈ B such that c( f ) = I. But c( f ) is not a regular ideal of R, therefore according to our assumption, f is not a regular member of B. Since B is a content R-algebra, there exists a nonzero r ∈ R such that r f = 0 and this means that rI = (0), i.e. I has a nonzero annihilator.  Remark 25. In the above theorem the surjectivity condition for the content function c is necessary, because obviously R is a content R-algebra and the condition (2) is satisfied, while one can choose the ring R such that it does not have Property (A) [16, Exercise 7, p. 63]. Theorem 26. Let R have property (A) and B be a content R-algebra. Then Z(B) is a finite union of prime ideals in Min(B) iff Z(R) is a finite union of prime ideals in Min(R). Proof. The proof is similar to the proof of Theorem 20 by considering Theorem 16.  Please note that if R is a Noetherian reduced ring, then Z(R) is a finite union of prime ideals in Min(R) (Refer to [16, Theorem 88, p. 59] and [14, Corollary 2.4]). Now we generalize the definition of rings having very few zero-divisors in the following way and prove the monoid module version of the above theorem. Definition 27. An R-module M has very few zero-divisors, if ZR (M) is a finite union of prime ideals in AssR (M). Remark 28. Examples of modules having very few zero-divisors. If R is a Noetherian ring and M is an R-module such that AssR (M) is finite, then obviously M has very few zero-divisors. For example AssR (M) is finite if M is a finitely generated R-module [16, p. 55]. Also if R is a Noetherian quasi-local ring and M is a balanced big Cohen-Macaulay R-module, then AssR (M) is finite [7, Proposition 8.5.5, p. 344]. Theorem 29. Let R-module M have very few zero-divisors. If S is a commutative, cancellative, torsion-free monoid then the R[S]-module M[S] has very few zero-divisors also. Proof. The proof is similar to the proof of Theorem 20. 5. G AUSSIAN AND  A RMENDARIZ A LGEBRAS Definition 30. Let B be an R-algebra that is a content R-module. B is said to be a Gaussian R-algebra if c( f g) = c( f )c(g), for all f , g ∈ B. CONTENT ALGEBRAS OVER COMMUTATIVE RINGS WITH ZERO-DIVISORS 11 For example if B is a content R-algebra such that every nonzero finitely generated ideal of R is cancellation ideal of R, then B is a Gaussian R-algebra. Another example is given in the following remark: Remark 31. Let (R, m) be a quasi-local ring with m2 = (0). If B is a content R-algebra, then B is a Gaussian R-algebra. Proof. Let f , g ∈ B such that c( f ) ⊆ m and c(g) ⊆ m, then c( f g) = c( f )c(g) = (0), otherwise one of them, say c( f ), is R and according to Dedekind-Mertens content formula, we have c( f g) = c(g) = c( f )c(g).  Theorem 32. Let M be an R-module such that every finitely generated R-submodule of M is cyclic and S be a commutative, cancellative, torsion-free monoid. Then for all f ∈ R[S] and g ∈ M[S], c( f g) = c( f )c(g). Proof. Let g ∈ M[S] such that g = m1 g1 + m2 g2 + · · · + mn gn , where m1 , m2 , . . . , mn ∈ M and g1 , g2 , . . ., gn ∈ S. Then there exists an m ∈ M, such that c(g) = (m1 , m2 , . . . , mn ) = (m). From this, we can get mi = ri m and m = ∑ si mi , where ri , si ∈ R. Put d = ∑ si ri , then m = dm. Since S is an infinite set, it is possible to choose gn+1 ∈ S − {g1 , g2, . . . , gn } and put g′ = r1 g1 + r2 g2 + · · · + rn gn + (1 − d)gn+1 . One can easily check that g = g′ m and c(g′ ) = R and c( f g) = c( f g′ m) = c( f g′ )m = c( f )m = c( f )c(g).  Corollary 33. Let R be a ring such that every finitely generated ideal of R is principal and S be a commutative, cancellative, torsion-free monoid. Then R[S] is a Gaussian Ralgebra. For more about content formulas for polynomial modules, refer to [21] and [5]. In the next step, we define Armendariz algebras and show their relationships with Gaussian algebras. Armendariz rings were introduced in [24]. A ring R is said to be an Armendariz ring if for all f , g ∈ R[X ] with f = a0 + a1 X + · · · + an X n and g = b0 + b1 X + · · · + bm X m , f g = 0 implies ai b j = 0, for all 0 ≤ i ≤ n and 0 ≤ j ≤ m. This is equivalent to say that if f g = 0, then c( f )c(g) = 0 and our inspiration to define Armendariz algebras. Definition 34. Let B be an R-algebra such that it is a content R-module. We say B is an Armendariz R-algebra if for all f , g ∈ B, if f g = 0, then c( f )c(g) = (0). For example if B is a weak content R-algebra and R is a reduced ring, then B is an Armendariz R-algebra. Theorem 35. Let R be a ring and (0) a p-primary ideal of R such that p2 = (0) and B a content R-algebra. Then B is an Armendariz R-algebra. 12 PEYMAN NASEHPOUR Proof. Let f , g ∈ B, where f g = 0. If f = 0 or g = 0, then definitely c( f )c(g) = 0, otherwise suppose that f 6= 0 and g 6= 0, therefore f and g are both zero-divisors of B. Since (0) is a p-primary ideal of R, so (0) is a pB-primary ideal of B [23, R, p. 331] and therefore pB is the set of zero-divisors of B. So f , g ∈ pB and this means that c( f ) ⊆ p  and c(g) ⊆ p. Finally c( f )c(g) ⊆ p2 = (0). In order to characterize Gaussian algebras in terms of Armendariz algebras, we should mention the following useful remark. Remark 36. Let R be a ring and I an ideal of R. If B is a Gaussian R-algebra then B/IB is a Gaussian (R/I)-algebra also. Theorem 37. Let B be a content R-algebra. Then B is a Gaussian R-algebra iff for any ideal I of R, B/IB is an Armendariz (R/I)-algebra. Proof. (→) : According to the above remark, since B is a Gaussian R-algebra, B/IB is a Gaussian (R/I)-algebra and obviously any Gaussian algebra is an Armendariz algebra and this completes the proof. (←) : One can easily check that if B is an algebra such that it is a content R-module, then for all f , g ∈ B, c( f g) ⊆ c( f )c(g) [23, Proposition 1.1, p. 330]. Therefore we need to prove that c( f )c(g) ⊆ c( f g). Put I = c( f g), since B/IB is an Armendariz (R/I)-algebra and c( f g + IB) = I so c( f + IB)c(g + IB) = I and this means that c( f )c(g) ⊆ c( f g).  For more about Armendariz and Gaussian rings, one can refer to [3]. 6. N ILRADICAL AND JACOBSON RADICAL OF CONTENT ALGEBRAS Definition 38. A ring R is said to be domainlike if any zero-divisor of R is nilpotent, i.e. Z(R) ⊆ Nil(R) [2, Definition 9]. Theorem 39. If B is a content R-algebra, then R is domainlike iff B is domainlike. Proof. The ring R is domainlike iff (0) is a primary ideal of R [2, Lemma 10]. Also it is easy to prove that if B is a content R-algebra, then q is a p-primary ideal of R iff qB is a pB-primary ideal of B [R, p. 331].  In a similar way one can see: Remark 40. Let S be a commutative, cancellative and torsion-free monoid and M be an R-module. Then ZR (M) ⊆ Nil(R) iff ZR[S] (M[S]) ⊆ Nil(R[S]). Remark 41. If B is a weak content R-algebra, then Nil(B) = Nil(R)B, particularly R is a reduced ring iff B is a reduced ring. CONTENT ALGEBRAS OVER COMMUTATIVE RINGS WITH ZERO-DIVISORS 13 Proof. It is obvious that Nil(R)B ⊆ Nil(B). Also it is easy to prove that for all f ∈ B and natural number n, we have c( f )n ⊆ rad(c( f n )) and therefore if f ∈ B is nilpotent, then c( f ) ⊆ Nil(R) and consequently f ∈ Nil(R)B.  Definition 42. A ring R is called presimplifiable if any zero-divisor of R is a member of the Jacobson radical of R, i.e. Z(R) ⊆ Jac(R). For more about presimplifiable rings, one can refer to [2]. In the following, our aim is show when some of the content algebras are presimplifiable. For doing that, we need to know about localization of content algebras that have been discussed in [22, Section 3, p. 56-58]. Actually we are interested in the following special case of localization: Let B be a content R-algebra and S′ = { f ∈ B : c( f ) = R}. It is easy to check that S S′ = B − m∈Max(R) mB and S = S′ ∩ R = U (R), where by U (R), we mean the units of R. According to [22, Theorem 6.2, p. 64], it is clear that BS′ is also a content R-algebra and B is a subring of BS′ . This special content R-algebra has some interesting properties: Theorem 43. Let B be a content R-algebra such that S′ = { f ∈ B : c( f ) = R} and put R′ = BS′ , then the following statements hold: (1) (2) (3) (4) (5) The map ϕ : Max(R) −→ Max(R′ ), defined by I −→ IR′ is a bijection. Jac(R′ ) = Jac(R)R′ . U (R′ ) = { f /g : c( f ) = c(g) = R} The ring R′ is presimplifiable iff R is presimplifiable. The ring R′ is 0-dimensional iff R is 0-dimensional. Proof. The first proposition is actually a special case of [11, 4.8]. For the proof of the second proposition, notice that the Jacobson radical of a ring is the intersection of all maximal ideals. Now use [22, 1.2, p. 51]. It is obvious that if c( f ) = c(g) = R, then f /g is a unit of R′ . Now let f /g be a unit of R′ , where c(g) = R and assume that there exists a member of R′ , say f ′ /g′ with c(g′ ) = R, such that ( f /g) · ( f ′ /g′ ) = 1/1. According to McCoy’s property for content algebras, S′ ⊆ B − ZB (B). So f f ′ = gg′ and f f ′ ∈ S′ . This means that f ∈ S′ and the proof of the third proposition is complete. For the proof of the forth proposition, suppose R is presimplifiable and let f ∈ Z(R′ ). Therefore there exists a nonzero r ∈ R such that r f = 0 and so rc( f ) = (0). This means that c( f ) ⊆ Z(R). Since R is presimplifiable, c( f ) ⊆ Jac(R) and at last f ∈ Jac(R)R′ and according to (2), f ∈ Jac(R′ ). It is easy to check that if R′ is presimplifiable then R is presimplifiable also. For the proof of the fifth proposition note that a ring, say T , is 0-dimensional iff Min(T ) = Max(T ).  14 PEYMAN NASEHPOUR Theorem 44. Let B be a content R-algebra with the property that if f ∈ B with c( f ) = (a) where a ∈ R, then there exists an f1 ∈ B such that f = a f1 and c( f1 ) = R and put S′ = { f ∈ B : c( f ) = R} and R′ = BS′ . Then the idempotent members of R and R′ coincide. Proof. Let f /g be an idempotent member of R′ , where f , g ∈ B and c(g) = R. Therefore f g2 = g f 2 and since g is a regular member of B, we have f g = f 2 . So c( f 2 ) = c( f g) = c( f ), but c( f 2 ) ⊆ c( f )2 , therefore c( f )2 = c( f ). We know that every finitely generated idempotent ideal of a ring is generated by an idempotent member of the ring [G1, p. 63]. Therefore we can suppose that c( f ) = (e) such that e2 = e. On the other side we can find an f1 ∈ B such that f = e f1 and c( f1 ) = R. Consider e f1 /g = f /g = f 2 /g2 = e2 f12 /g2 . Since f1 and g are both regular, and e is idempotent, we have e = e f1 /g = f /g ∈ R.  Corollary 45. Let R be a ring and M a commutative, cancellative and torsion-free monoid and put S′ = { f ∈ R[M] : c( f ) = R} and R′ = BS′ . Then the idempotent members of R and R′ coincide. Definition 46. A commutative ring R is said to be a valuation ring if for any a and b in R either a divides b or b divides a ([16, p. 35]). Theorem 47. Let B be a content R-algebra with the property that if f ∈ B with c( f ) = (a) where a ∈ R, then there exists an f1 ∈ B such that f = a f1 and c( f1 ) = R and put S′ = { f ∈ B : c( f ) = R} and R′ = BS′ . If R is a valuation ring, then so is R′ = BS′ . Proof. Let f /g be a member of R′ , where f , g ∈ B and c(g) = R. Since c( f ) is a finitely generated ideal of R and R is a valuation ring, there exists an r ∈ R such that c( f ) = (r) and therefore there exists an f1 ∈ B such that f = r f1 and c( f1 ) = R. By considering this fact that f1 /g is a unit in R′ , it is obvious that R′ is also a valuation ring and the proof is complete.  7. ACKNOWLEDGMENT The author wishes to thank Prof. Winfried Bruns for his useful advice. R EFERENCES [1] E. P. Armendariz, A note on extensions of Baer and p.p.-rings, J. Austral. Math. Soc. 18 (1974), 470–473. [2] D. D. Anderson, M. Axtell, S. F. Forman and Joe Stickles, Commutative rings with domain-type properties, preprint. [3] D. D. Anderson and V. Camillo, Armendariz rings and Gaussian rings, Comm. Algebra. 26 (1998), 2265–2272. [4] J. T. Arnold and R. Gilmer, On the content of polynomials, Proc. Amer. Math. Soc. 40 (1970), 556– 562. CONTENT ALGEBRAS OVER COMMUTATIVE RINGS WITH ZERO-DIVISORS 15 [5] D. D. Anderson and B. G. Kang, Content formulas for polynomials and power series and complete integral closure, J. Algebra, 181 (1996), 82–94. [6] W. Bruns and A. Guerrieri, The Dedekind-Mertens formula and determinantal rings, Proc. Amer. Math. Soc. 127 (1999), no. 3, 657–663. [7] W. Bruns and J. Herzog, Cohen-Macaulay Rings, revised edn., Cambridge, 1998. [8] E. Davis, Overrings of commutative rings II. Integrally closed overrings, Trans. Amer. Math. Soc, 110 (1964), 196–212. [9] P. Eakin and J. Silver, Rings which are almost polynomial rings, Trans. Amer. Math. Soc. 174 (1974), 425–449. [10] N. Epstein and J. Shapiro, A Dedekind-Mertens theorem for power series rings, to appear in Proc. Amer. Math. Soc. [11] R. Gilmer, Multiplicative Ideal Theory, Marcel Dekker, New York, 1972. [12] R. Gilmer, Commutative Semigroup Rings, The University of Chicago Press, 1984. [13] W. Heinzer and C. Huneke, The Dedekind-Mertens Lemma and the content of polynomials, Proc. Amer. Math. Soc. 126 (1998), 1305–1309. [14] J. A. Huckaba, Commutative Rings with Zero Divisors, Marcel Dekker, 1988. [15] J. A. Huckaba and J. M. Keller, Annihilation of ideals in commutative rings, Pac. J. Math. 83 (1979), 375–379. [16] I. Kaplansky, Commutative Rings, Allyn and Bacon, Boston, 1970. [17] K. A. Loper and M. Roitman, The content of a Gaussian polynomial is invertible, Proc. Amer. Math. Soc. 133 (2005), 1267–1271. [18] N. H. McCoy, Remarks on divisors of zero, Amer. Math. Monthly, 49 (1942), 286–295. [19] D. G. Northcott, A generalization of a theorem on the content of polynomials, Proc. Cambridge Phil. Soc. 55 (1959), 282–288. [20] A. G. Naoum and A. S. Mijbass, Weak cancellation modules, Kyungpook Math. J., 37 (1997), 73–82. [21] P. Nasehpour and S. Yassemi, M-cancellation Ideals, Kyungpook Math. J., 40 (2000), 259–263. [22] J. Ohm and D. E. Rush, Content modules and algebras, Math. Scand. 31 (1972), 49–68. [23] D. E. Rush, Content algebras, Canad. Math. Bull. Vol. 21 (3) (1978), 329–334. [24] M. B. Rege and S. Chhawchharia, Armendariz rings, Proc. Japan Acad. Ser. A Math. Sci. 73, Number 1 (1997), 14–17. [25] H. Tsang, Gauss’ lemma, dissertation, University of Chicago, Chicago, 1965. U NIVERSIT ÄT O SNABR ÜCK , FB M ATHEMATIK /I NFORMATIK , 49069 O SNABR ÜCK , G ERMANY E-mail address: pnasehpo@uni-osnabrueck.de E-mail address: nasehpour@gmail.com 

0math.AC

Optimal experimental design that minimizes the width of simultaneous confidence bands arXiv:1704.03995v2 [math.ST] 26 Mar 2018 Satoshi Kuriki∗ and Henry P. Wynn† Abstract We propose an optimal experimental design for a curvilinear regression model that minimizes the band-width of simultaneous confidence bands. Simultaneous confidence bands for curvilinear regression are constructed by evaluating the volume of a tube about a curve that is defined as a trajectory of a regression basis vector (Naiman, 1986). The proposed criterion is constructed based on the volume of a tube, and the corresponding optimal design that minimizes the volume of tube is referred to as the tube-volume optimal (TV-optimal) design. For Fourier and weighted polynomial regressions, the problem is formalized as one of minimization over the cone of Hankel positive definite matrices, and the criterion to minimize is expressed as an elliptic integral. We show that the Möbius group keeps our problem invariant, and hence, minimization can be conducted over cross-sections of orbits. We demonstrate that for the weighted polynomial regression and the Fourier regression with three bases, the tube-volume optimal design forms an orbit of the Möbius group containing D-optimal designs as representative elements. Key words: D-optimality, Fourier regression, Hankel matrix, Möbius group, volumeof-tube method, weighted polynomial regression. 1 Introduction Suppose that we observe pairs of explanatory variables xi ∈ X and response variables yi ∈ R, i = 1, . . . , N. Here, X ⊂ R is a domain of explanatory variables, and typically, a segment of R. For such data, we consider the regression model yi = b⊤ f (xi ) + εi , εi ∼ N(0, σ 2 (x)) i.i.d., where b = (b1 , . . . , bn )⊤ is an unknown coefficient vector, and f (x) = (f1 (x), . . . , fn (x))⊤ , x ∈ X , is a piecewise smooth regression basis vector. For the problem of this paper, we ∗ The Institute of Statistical Mathematics, 10-3 Midoricho, Tachikawa, Tokyo 190-8562, Japan, Email: kuriki@ism.ac.jp † The London School of Economics and Political Science, Houghton Street, London WC2A 2AE, UK, Email: h.wynn@lse.ac.uk 1 assume that the variance function σ 2 (x) > 0 is known. When σ 2 (x) is not a constant, the regression model is called weighted . In the case of experimental data, the explanatory variables xi can be chosen arbitrarily within its domain X ⊂ R. The allocation of {x1 , . . . , xN } ⊂ X to optimize some target function is called optimal experimental design. For example, for D-optimality, we take a function det(Σ) with Σ = Var(bb), where bb is the ordinary least square (OLS) estimator of b. Here, N X 1 −1 Σ=M with M = f (xi )f (xi )⊤ 2 , σ (x ) i i=1 the information matrix. Following Kiefer and Wolfowitz (1959), in an optimal design, allocation {x1 , . . . , xN } is regarded as the probability measure over X with mass pi = 1/N at each point xi . We write this discrete probability measure as     x1 · · · xN xi . = 1 · · · N1 pi 1≤i≤N N Viewed from this point, the problem is formalized as that of optimization with respect to the probability measure over X . Most criteria in the literature including the criterion mentioned above are convex or concave functionals of the probability measure, and can be considered in the framework of convex analysis (Wynn, 1985; Pukelsheim, 2006). In this paper, we propose a new non-convex criterion based on simultaneous confidence bands. The pointwise confidence band is based on the confidence region for regressor b⊤ f (x) at a fixed point x. On the other hand, the simultaneous confidence band is the confidence region for the full regression curve {(x, b⊤ f (x)) | x ∈ X } ⊂ R2 . The standard form of the simultaneous confidence band of hyperbolic-type is of the form p b⊤ f (x) ∈ bb⊤ f (x) ± cα f (x)⊤ Σf (x), (1.1) where u ± v stands for the region (u − v, u + v). The threshold cα is determined so that the event (1.1) holds for all x ∈ X with given probability 1 − α (Working and Hotelling, 1929; Scheffé, 1959; Wynn and Bloomfield, 1971; Liu, 2010). The simultaneous confidence bands are useful when x cannot be determined in advance. (See van Dyk (2014) for an application in experimental particle physics.) As shown in the next section, cα is also a functional of the allocation {x1 , . . . , xN }, and hence, we can consider an optimal design that in some way minimizes both the threshold cα and f (x)⊤ Σf (x). In fact, from the general equivalence theorem of Kiefer and Wolfowitz (1959), the design measure that minimizes maxx∈X f (x)⊤ Σf (x) coincides with the D-optimal design. Therefore, we propose the use of cα as a criterion of optimal design, and consider the corresponding optimal design as the tube-volume optimal (TV-optimal) design. If a design is optimal under both the tube-volume criterion and the D-criterion, it becomes the universal optimal design to minimize the width of confidence bands (1.1). Indeed cα substantially depends on the design. In Table 4.3 of Section 4.3, we will show how the difference in design affects the width of the simultaneous confidence band. 2 From its definition, cα is a complicated function of Σ. However, when α is small, cα tends to a simpler function. This approximation is due to the volume-of-tube method used to construct simultaneous confidence bands in curvilinear regression curves (Naiman, 1986; Johansen and Johnstone, 1990; Sun and Loader, 1994; Lu and Kuriki, 2017). The volume-of-tube method is a methodology to approximate the probability of the maximum of a Gaussian random field (Sun, 1993; Kuriki and Takemura, 2001, 2009; Takemura and Kuriki, 2002; Adler and Taylor, 2007). As shown later, cα corresponds to the upper tail probability of the maximum of a Gaussian field, and hence, the volume-of-tube method works well. As concrete regression models, weighted polynomial and Fourier regressions are mainly covered here. In these models, we will see that there is a group referred to as the Möbius transform that keeps the tube-volume optimal design problem invariant. In general, a group action simplifies problems. (See Section 13 of Pukelsheim (2006) for invariant optimal experimental design.) The use of such group invariance is another subject of this paper. The outline of this paper is as follows. Section 2 summarizes the volume-of-tube formula to construct approximate simultaneous confidence bands, and formalizes the tubevolume criterion and the corresponding optimal design. Section 3 analyzes the tubevolume optimal designs for Fourier and weighted polynomial regressions. The Möbius group is proved to keep the optimization problem invariant, and hence can be used to reduce the dimension of the problem. Using this consideration, Section 4 identifies the tube-volume optimal design in the weighted polynomial regression and the Fourier regression when n = 3. Some proofs are given in Appendix. 2 2.1 Tube-volume optimal design Volume-of-tube formula for simultaneous confidence bands In this subsection, we briefly summarize the volume-of-tube method. This is a general methodology used to approximate the probability of the maximum of a smooth Gaussian random process or random field. Here, we describe how this method is used to determine threshold cα . As mentioned in Section 1, threshold cα should be determined as a solution c = cα of  b⊤  | b f (x) − b⊤ f (x)| p Pr < c, ∀x ∈ X = 1 − α, f (x)⊤ Σf (x) which is equivalent to 1   | (bb − b)⊤ f (x)| Pr max > c = α, 1 x∈X kΣ 2 f (x)k 1 1 where Σ 2 is a square-root matrix such that (Σ 2 )⊤ Σ 2 = Σ. 3 (2.1) (2.2) We define the normalized basis vector and its trajectory as 1 ψΣ (x) = Σ 2 f (x) 1 kΣ 2 f (x)k ,  γΣ = ±ψΣ (x) | x ∈ X , (2.3) respectively. From this definition, the trajectory is a subset of the (n − 1)-dimensional unit sphere: γΣ ⊂ Sn−1 = {u ∈ Rn | kuk = 1}. In particular, when X is a segment, γΣ is a curve on the unit sphere. Let Vol1 (·) denote the one-dimensional volume, that is, the length. Then, when c is large, the volume-of-tube method provides an approximation to the upper tail probability of the maximum in (2.2). Further, let χ2ν denote the chi-square random variable with ν degrees of freedom. Proposition 2.1. (i) As c → ∞,  
| (bb − b)⊤ f (x)| Vol1 (γΣ ) Pr χ22 > c2 . Pr max >c ∼ 1 x∈X 2π kΣ 2 f (x)k (2.4) (ii) For all c > 0, the left-hand side of (2.4) is bounded above by

Vol1 (γΣ ) Pr χ22 > c2 + χ(γΣ ) Pr χ21 > c2 , 2π (2.5) where χ(γΣ ) is the number of the connected components of the set γΣ (⊂ Sn−1 ). If we admit approximation (2.4), an approximate threshold cα can be determined from the equation
Vol1 (γΣ ) Pr χ22 > c2α = α. 2π This means that the smaller the value of Vol1 (γΣ ), the smaller is cα . The statement (ii) above is due to Naiman (1986). Alternative proofs of the inequality can be found in Johnstone and Siegmund (1989) and Takemura and Kuriki (2002). See Lu and Kuriki (2017) for a generalization of Naiman’s inequality. By equating (2.5) to be α, we have a conservative threshold for the simultaneous confidence band. 2.2 Tube-volume criterion From (2.5), we find that the smaller the value of Vol1 (γΣ ), the narrower is the width of the confidence band. In this subsection, we formalize the experimental design optimization problem of the allocation of explanatory variables to minimize Vol1 (γΣ ). Here, we give our assumptions on f (x). Assumption 2.2. f : X → Rn is a continuous and piecewise C 1 -function. Image f (X ) spans Rn . 4 From elementary geometry, the volume of γΣ in (2.3) is given by Z dψΣ (x) Vol1 (γΣ ) = 2 dx dx X  1  Z d Σ 2 f (x) =2 dx 1 X dx kΣ 2 f (x)k Z p (f (x)⊤ Σf (x))(g(x)⊤ Σg(x)) − (f (x)⊤ Σg(x))2 =2 dx f (x)⊤ Σf (x) X  1 
2 f (x)⊤ Σ f (x), g(x) Z det g(x)⊤ 2×2 dx, =2 ⊤ f (x) Σf (x) X (2.6) where g(x) = df (x)/dx. We call (2.6) the tube-volume (TV) criterion. Note that (2.6) is invariant with respect to scale Σ 7→ kΣ (k > 0). Because of this, we introduce the set of all nonnegative measures (not necessarily probability measures) on X denoted by P. Each element of ρ ∈ P corresponds to an experimental design. This is an extension of the design measure of Kiefer and Wolfowitz (1959). We also extend the set of moment matrices. Thus, the set of all non-singular information matrices is denoted by Z  1 ⊤ f (x)f (x) 2 dρ(x) ≻ 0 | ρ ∈ P M= σ (x) X Z  ⊤ = f (x)f (x) dρ(x) ≻ 0 | ρ ∈ P , X where “≻ 0” denotes positive definiteness. From its definition, M forms a convex cone. Our optimal design problem is formulated as minimizing Vol1 (γΣ ) in (2.6) subject to Σ−1 ∈ M. The lemma below is a direct consequence of this invariance.     xi xi 2 with variance function σ1 (x), and the design Lemma 2.3. The design qi 1≤i≤N pi 1≤i≤N 2 2 2 with variance function Pσ2 (x), qi = kpi σ2 (xi )/σ1 (xi ), give the same volume, where k > 0 is a constant so that i qi = 1. Proof. The information matrices M1 and M2 of the two designs satisfy M1 = kM2 . A difficulty with this optimization problem is that this is not a convex problem. Figure 2.1 depicts the volume Vol1 (γM (c)−1 ) for a mixing design connecting two Fourier designs with three bases ((3.1) with n = 3) with weights 1 − c and c :  1     1 1  − 3 0 13 − 12 0 12 ti = and . 1 1 1 1 1 1 pi i=1,2,3 3 3 3 3 3 3 The information matrix is    1 0 0 1 0 M(c) = (1 − c) 0 1 0 + c  √ 0√ 13 2+ 6 0 0 1 0 3 5 √ √  2+ 6 3 0 5 3 . We see that Vol1 (γM −1 (c) ) is not convex in c. c Figure 2.1: Vol1 (γM (c)−1 ) for a mixing design. 3 3.1 Tube-volume optimal design for polynomial and Fourier regressions Equivalence between weighted polynomial regression and Fourier regression From this section, we focus on Fourier regression and weighted polynomial regression. In linear optimal design theory, the Fourier regression model has been used as one of the standard models to see the performance of the proposed criteria. The weighted polynomial used here is closely connected to the Fourier regression model, as is shown later. The Fourier (trigonometric) regression has basis vector f (x) =fF (x) ( √ √ √ √ (1, 2 sin(2πx), 2 cos(2πx), 2 sin(4πx), . . . , 2 cos(2πmx))⊤ √ √ √ = √ ( 2 cos(πx), 2 sin(πx), 2 cos(3πx), . . . , 2 cos(π(n − 1)x))⊤ (n = 2m + 1), (n = 2m) (3.1) defined on the domain X = (−1/2, 1/2]. For the Fourier regression, we only deal with the constant variance σ 2 (x) = σF2 (x) ≡ 1. Although the Fourier regression is not used for even values of n in practice, we define it for the sake of consistency. The polynomial regression is a regression model with basis vector f (x) = fP (x) = (1, x, x2 , . . . , xn−1 )⊤ ∈ R[x]n . (3.2) Here, we set the domain X to be the whole real line R. For the polynomial regression, we assume the variance function of form σ 2 (x) = Q(x)n−1 , where Q(x) is an arbitrary 6 positive quadratic function. As a canonical form of this class of variance functions, we use σP2 (x) = (1 + x2 )n−1 . (3.3) Later, we introduce a parameterization for Q(x) (see (3.20)). In this subsection, we see that under the tube-volume criterion, the optimization problem for the Fourier regression is equivalent to that for the weighted polynomial regression. That is, the optimization problem in the Fourier regression can be translated to one in the weighted polynomial regression, and vice versa. The model we discuss is a special case of the model proposed by Dette, et al. (1999), Section 2.2, who study the D-optimality. Further, Dette and Melas (2003) make use of the connection between the weighted polynomial and Fourier regressions. We will return briefly to question on the D-optimality in Section 3.5 below. From the lemma below, the transformation x = tan(πt) connects the two regression models. Lemma 3.1. There exists an n × n non-singular matrix B such that for all x ∈ R and t ∈ (−1/2, 1/2] satisfying x = tan(πt), we have fF (t) = BfP (x)λ0 (x), λ0 (x) = 1/(1 + x2 )(n−1)/2 . (3.4) Proof. Note that sin(2πt) = 2x , 1 + x2 cos(2πt) = 1 − x2 , 1 + x2 dt = dx . π(1 + x2 ) (3.5) It is known that  k sin(2πkt) = sin2r+1 (2πt) cosk−2r−1 (2πt), (−1) 2r + 1 r=0   [k/2] X k r sin2r (2πt) cosk−2r (2πt) cos(2πkt) = (−1) 2r r=0 [(k−1)/2] X r  (Moriguti, et al., 1957, pages 186–187). Substituting the formulas for sin(2πt) and cos(2πt) in (3.5) and expressing fF (t) as a rational function in x, we have formula (3.4). To prove that B is non-singular, consider the integral Z ∞  Z 1/2 dx ⊤ 2 ⊤ fF (t)fF (t) dt = B fP (x)λ0 (x) fP (x) B⊤. (3.6) 2) π(1 + x −1/2 −∞ Here, we used (3.5). The left-hand side is the identity matrix In by standard orthogonality. Hence, it is enough to check that the integrals in the parentheses of the right-hand side exist. The matrix in the parentheses of the right-hand side of (3.6) is (B ⊤ B)−1 with (i, j) element 

 Γ n − i+j−1 Γ i+j−1 Z ∞  i+j−2 2 2 x (i + j is even), dx = (B ⊤ B)−1 (3.7) πΓ(n) i,j = 2 n  −∞ π(1 + x ) 0 (i + j is odd), 7 which exists for i, j ≤ n. Hence, B is non-singular. When n = 3, 4,  1 1 0 1 0 √ 0 ,  B = √0 2 2 1 √ 2 0 − 2 0    0 1 0 1 0 1 , 0 −3 0  3 0 −1 respectively. The set of information matrices for the polynomial regression Z ∞  ⊤ MP = fP (x)fP (x) dρ(x) ≻ 0 | ρ ∈ P (3.8) −∞ is referred to as the moment cone (Karlin and Studden, 1966). The set of information matrices for the Fourier regression is given by Z 1/2  ⊤ MF = fF (t)fF (t) dρ(t) ≻ 0 | ρ ∈ P −1/2 = {BMB ⊤ | M ∈ MP } = BMP B ⊤ . (3.9) The following lemma gives the equivalence of the Fourier regression and the polynomial regression as the optimization problem for the tube-volume criterion. Theorem 3.2. Let VolF (γM −1 ) and VolP (γM −1 ) be the length of γM −1 given in (2.6) with f (x) being fF (x) in (3.1), and fP (x) in (3.2), respectively. Then, it holds that VolP (γM −1 ) = VolF (γ(BM B⊤ )−1 ). Proof. The derivatives of fF (t) and fP (x) are denoted by gF (t) = dfF (t)/dt and gP (x) = dfP (x)/dx, respectively. Then,   

λ0 (x) λ̇0 (x) 1 0 . fF (t), gF (t) = B fP (x), gP (x) 0 dx 0 λ0 (x) dt Therefore, fF (t)⊤ (BMB ⊤ )−1 fF (t) = fP (x)⊤ M −1 fP (x) × λ0 (x)2 ,      

fP (x)⊤ fF (t)⊤ −1 ⊤ −1 M fP (x), gP (x) (BMB ) fF (t), gF (t) = det det gP (x)⊤ gF (t)⊤  dx 2 × λ0 (x)4 , dt 8 and  21  
fF (t)⊤ Z 1/2 det (BMB ⊤ )−1 fF (t), gF (t) gF (t)⊤ dt fF (t)⊤ (BMB ⊤ )−1 fF (t) −1/2  1 
2 fP (x)⊤ −1 Z ∞ det M fP (x), gP (x) gP (x)⊤ = dx. fP (x)⊤ M −1 fP (x) −∞ Theorem 3.2 and (3.9) imply that M ∈ MP is the minimizer of VolP (M) ⇔ M ′ = BMB ⊤ ∈ MF is the minimizer of VolP (M ′ ). That is, the optimization problems for the polynomial regression and the Fourier regression are mathematically equivalent. For example, the information matrix M = P ⊤ f (t )f i F i F (ti ) pi ∈ MF for the Fourier regression, and the information matrix for the polynomial regression X B −1 fF (ti )(B −1 fF (ti ))⊤ pi MP ∋ B −1 M(B ⊤ )−1 = i = X fP (xi )fP (xi )⊤ λ0 (xi )2 pi , xi = tan(πti ), i = X fP (xi )fP (xi )⊤ i 1 pi σP2 (xi ) give the same volume. This equivalence is stated in terms of design measure as follows.     ti xi Theorem 3.3. The design for the Fourier regression, and the design , pi 1≤i≤N pi 1≤i≤N xi = tan(πti ), for the weighted polynomial regression with variance function σP2 (x) give the same volume. If the former is tube-volume optimal in the Fourier regression, then so is the latter in the polynomial regression with variance σP2 (x), and vice versa. In this paper, the (discrete) uniform designs in the Fourier regression and their counterparts in the polynomial regression play important roles. It is known that, in the Fourier regression, the uniform design in which xi are allocated as equally spaced with equal weights is D-optimal (Guest, 1958). Because of the symmetry, it is conjectured that the uniform design is the tube-volume optimal design as well. In Section 4, we prove that this is true for n = 3, and conjecture that it is true for all n. 9 The n-point discrete uniform design for the Fourier regression symmetric about the origin is  0 n+1 i ti . (3.10) , t0i = − 1 n 2n n 1≤i≤n For later use, we provide the concrete forms of the information matrix M for the weighted polynomial designs with σ 2 (x) = σP2 (x) in (3.3),  0 xi , x0i = tan(πt0i ) with t0i given in (3.10). (3.11) 1 n 1≤i≤n Lemma 3.4. The information matrix M = (Mi,j ) of the weighted polynomial design (3.11) scaled such that M1,1 = 1 is given by  k+1

k+1   Γ √2 Γ n −
2 (k is even), M = (Mi,j ) = (mi+j−2 )1≤i,j≤n , mk = πΓ n − 12  0 (k is odd). For the proof, see Appendix A.1. When n = 3 and 4,     1 0 51 0 1 1 0 3 0 1 0 1  5 5 , M = 0 31 0  ,  1 1   0 0 1 5 5 0 1 1 3 0 5 0 1 (3.12) respectively. The key transform connecting Fourier and polynomial regressions was the tangent transform x = tan(πt). For the same purpose, generalized transforms x = q tan(π(t − θ)) + r, q 6= 0, can be used. This is a composite map of the tangent transform and the Möbius transform to be discussed below. 3.2 The Möbius group action on the moment cone In this subsection, we introduce the Möbius group (transformation) acting on the set of design measures and the set of information matrices in polynomial regression. We will show that the Möbius group action reduces the dimension of the minimization problem for the tube-volume criterion. For a recent paper in which the Möbius transformation acts on polynomials, see Mackey, et al (2015). The real Möbius transformation is defined on the extended real numbers R = R ∪ {±∞} as follows: ax + b (ad − bc 6= 0). x 7→ ϕ(x; a, b, c, d) = cx + d Here, we assume that a d ±∞ 7→ , − 7→ ±∞. c c 10 This forms a group with product ϕ(·; a′ a + b′ c, a′ b + b′ d, c′a + d′ c, c′ b + d′ d) = ϕ(·; a′ , b′ , c′ , d′ ) ◦ ϕ(·; a, b, c, d). (3.13) The inverse is ϕ−1 (·; a, b, c, d) = ϕ(·; d, −b, −c, a). The identity element is e = ϕ(·; a, 0, 0, a), a 6= 0. This is a subgroup of the complex Möbius group referred to as projective general linear group PGL(2, C). Now, let fP (x) = (1, x, . . . , xn−1 )⊤ be the polynomial basis. We define an n × n matrix A = A(a, b, c, d) as
fP ϕ(x; a, b, c, d) = λ(x; a, b, c, d)AfP (x), λ(x; a, b, c, d) = 1 . (cx + d)n−1 (3.14) We write the factor λ as λ(x; a, b, c, d) instead of λ(x; c, d) to clarify that this is an invariant function under the group action (3.13) in the sense that
λ ϕ(x; a, b, c, d); a′, b′ , c′ , d′ λ(x; a, b, c, d) = λ(x; a′ a + b′ c, a′ b + b′ d, c′ a + d′ c, c′ b + d′ d). (3.15) Proposition 3.5. The (i, j) element of A = A(a, b, c, d) is max(i−1,j−1) (A)i,j = X l=max(0,i+j−n−1)    n−i i−1 al bi−1−l c j−1−l d n+1−i−j+l. j−1−l l The proof is straightforward and omitted. When n = 3 and n = 4,   3   2 d 3cd2 3c2 d c3 2 d 2cd c bd2 2bcd + ad2 bc2 + 2acd ac2   A = A(a, b, c, d) = bd bc + ad ac ,  b2 d b2 c + 2abd 2abc + a2 d a2 c , (3.16) 2 2 b 2ab a b3 3ab2 3a2 b a3 respectively. The set A = {A(a, b, c, d) | ad − bc 6= 0} is a representation of general linear group GL(2, R) and hence forms a group (Gross and Holman, 1980). The proof of the proposition below is straightforward and omitted. Proposition 3.6. Set A forms a matrix algebraic group. The identity matrix is A(1, 0, 0, 1) = (−1)n−1 A(−1, 0, 0, −1) = In , and the inverse of A = A(a, b, c, d) is given by (ad − bc)−(n−1) A(d, −b, −c, a). Proposition 3.7. For A = A(a, b, c, d) ∈ A, det(A) = (ad − bc)n(n−1)/2 . 11 Proof. For different x1 , . . . , xn , we have A(fP (x1 ), . . . , fP (xn )) = (fP (y1 ), . . . , fP (yn ))diag((cxi + d)n−1 ), By taking determinants, Y det(A) × (−1)n (xi − xj ) = (−1)n 1≤i<j≤n Y (yi − yj ) 1≤i<j≤n Y yi = axi + b . cxi + d (cxi + d)n−1 . (3.17) 1≤i≤n Substituting Y Y  axi + b axj + b  Y (ad − bc)(xi − xj ) = − cxi + d cxj + d (cxi + d)(cxj + d) 1≤i<j≤n 1≤i<j≤n Q (xi − xj ) n(n−1)/2 Q 1≤i<j≤n = (ad − bc) n−1 1≤i≤n (cxi + d) (yi − yj ) = 1≤i<j≤n into (3.17) yields det(A) = (ad − bc)n(n−1)/2 . Proposition 3.8. The Möbius group is reparameterized as {ϕ(·; a, b, c, d) | ad − bc 6= 0} ={ϕ(·; q, r, 0, 1) ◦ ϕ(·; s, −t, t, s) | q 6= 0, s2 + t2 = 1} ⊔ {ϕ(·; q, r, 0, 1) ◦ ϕ(·; −s, t, t, s) | q 6= 0, s2 + t2 = 1}, where ⊔ is the disjoint union. Group A is reparameterized as A ={kA(q, r, 0, 1)A(s, −t, t, s) | k > 0, q 6= 0, s2 + t2 = 1} ⊔ {kA(q, r, 0, 1)A(−s, t, t, s) | k > 0, q 6= 0, s2 + t2 = 1}. det A(q, r, 0, 1) = q n(n−1)/2 and det A(±s, ∓t, t, s) = ±1 for s2 + t2 = 1. Proof. We have the following relations: and   ad − bc ac + bd , , 0, 1 ϕ(x; a, b, c, d) = ϕ ϕ(x; ±d, ∓c, c, d); ± 2 c + d 2 c2 + d 2  ad − bc ac + bd  A(±d, ∓c, c, d) A(a, b, c, d) = (c2 + d2 )(n−1)/2 A ± 2 , , 0, 1 . c + d 2 c2 + d 2 (c2 + d2 )(n−1)/2 The results in the proposition follow by letting k = (c2 + d2 )(n−1)/2 , q= ad − bc , c2 + d 2 r= ac + bd , c2 + d 2 12 s= √ c2 d , + d2 t= √ c2 c . + d2 The sets of transformations {ϕ(·; ±s, ∓t, t, s) | s2 + t2 = 1} and {ϕ(·; q, r, 0, 1) | q 6= 0} form subgroups of the Möbius group, which are isomorphic to the orthogonal group O(2, R) and the affine group acting on R, respectively. Theorem 3.9. Let A ∈ A and M ∈ MP be n × n matrices. Then, AMA⊤ ∈ MP . Moreover, AMP A⊤ = {AMA⊤ | M ∈ MP } = MP . That is, group A acts on the moment cone MP . P ⊤ Proof. Suppose that M = i fP (xi )fP (xi ) wi . Let yi = ϕ(xi ; a, b, c, d). Note that λ(y; d, −b, −c, a)λ(x; a, b, c, d) = (ad − bc)−(n−1) . Then, X AMA⊤ = AfP (xi )(AfP (xi ))⊤ wi i = X fP (yi )fP (yi )⊤ λ(xi ; a, b, c, d)−2 wi i = X i = X i fP (yi )fP (yi )⊤ (ad − bc)2(n−1) λ(yi ; d, −b, −c, a)2 wi fP (yi )fP (yi )⊤ vi ∈ MP , where vi = λ(xi ; a, b, c, d)−2wi = (ad − bc)2(n−1) λ(yi ; d − b, −c, a)2 wi . (3.18) Therefore, AMP A⊤ ⊂ MP . Because A is a group, AMP A⊤ = MP . The Möbius group action on the polynomial basis f (x) has been introduced by (3.14). Similarly, we define the Möbius group action on the variance function σ 2 (x) = Q(x)n−1 . This provides a parameterization for the variance function. Using σP2 (x) = (1 + x2 )n−1 in (3.3), for ad − bc 6= 0, we define σP2 (ϕ(x; a, b, c, d)) = σP2 (x; a, b, c, d)λ(x; a, b, c, d)2, or  σP2 (x; a, b, c, d) = (b2 + d2 ) + 2(ab + cd)x + (a2 + c2 )x2 (3.19) n−1 . (3.20) Note that σP2 (x) = σP2 (x; 1, 0, 0, 1). This is always positive because of ad − bc 6= 0. For ϕ(·; a′, b′ , c′ , d′ ) = ϕ(·; a0 , b0 , c0 , d0 ) ◦ ϕ(·; a, b, c, d), as well as (3.15), we have σP2 (ϕ(x; a, b, c, d); a0, b0 , c0 , d0 ) = σP2 (x; a′ , b′ , c′ , d′ )λ(x; a, b, c, d)2 . (3.21) The parameterization (3.20) with (a, b, c, d) is redundant, since Q(x) has only three parameters. The lemma below shows that the stabilizer keeping the variance σP2 (·; a, b, c, d) invariant is the orthogonal subgroup with dimension one. 13 Lemma 3.10. σP2 (·; a, b, c, d) = σP2 (·; a′ , b′ , c′ , d′ ) if and only if there exist s, t, s2 + t2 = 1 such that ϕ(·; a′, b′ , c′ , d′) = ϕ(·; ±s, ∓t, t, s) ◦ ϕ(·; a, b, c, d). Proof. By direct calculations, σP2 (·; a0 , b0 , c0 , d0) = σP2 (·; 1, 0, 0, 1) ⇔ (a0 , b0 , c0 , d0 ) = (±s, ∓t, t, s). On the other hand, “σP2 (·; a0 , b0 , c0 , d0 ) = σP2 (·; 1, 0, 0, 1)” ⇔ “σP2 (ϕ(·; a, b, c, d); a0, b0 , c0 , d0) = σP2 (ϕ(·; a, b, c, d); 1, 0, 0, 1)” ⇔ “σP2 (x; a′ , b′ , c′ , d′) = σP2 (x; a, b, c, d) with ϕ(·; a′ , b′ , c′ , d′ ) = ϕ(·; a0 , b0 , c0 , d0) ◦ ϕ(·; a, b, c, d)” by (3.21). Remark 3.11. When n = 2, |ad−bc|/{πσP2 (x; a, b, c, d)} is a probability density of Cauchy distribution family on x ∈ R. Noting that ϕ̇(x; a, b, c, d) = (ad − bc)λ(x; a, b, c, d)2 , and that a′ d′ − b′ c′ = (ad − bc)(a0 d0 − b0 c0 ), (3.21) reads |a′ d′ − b′ c′ | |a0 d0 − b0 c0 | |ϕ̇(x; a, b, c, d)| dx = dx. πσP2 (ϕ(x; a, b, c, d); a0 , b0 , c0 , d0 ) πσP2 (x; a′ , b′ , c′ , d′) This means that the Cauchy distribution family is closed under the Möbius transform (McCullagh, 1996). See also Kato and McCullagh (2014) for Cauchy families in directional statistics. 3.3 Canonical parameterizations for information matrices As we have shown in Section 2.1, the optimal design problem is optimization with respect R ⊤ 1 to matrix M over the set of information matrices M. Here, M = X f (x)f (x) σ2 (x) dρ(x) ∈ M and the design measure ρ ∈ P is one-to-many. For the sake of optimization, we need to parameterize the set M. We first consider MP in (3.8) for the polynomial regression, and then interpret the results in terms of MF in (3.9) for the Fourier regression. The structure of the moment cone MP is well-studied in the context of the classical moment problem. One canonical parameterization for MP is given in Chapter II, Section 3 of Karlin and Studden (1966). The statement is summarized in Proposition 3.1 of Kato and Kuriki (2013). Proposition 3.12. M ∈ MP is uniquely represented with 2n−1 parameters (w0 , . . . , wn−1 , x1 , . . . , xn−1 ) as M= n−1 X wi fP (xi )fP (xi )⊤ + w0 fP (±∞)fP (±∞)⊤ , wi > 0, i=1 −∞ < x1 < · · · < xn−1 < ∞, where we let fP (±∞) = (0, . . . , 0, 1)⊤ . 14 (3.22) Note that fP (±∞) = limx→±∞ fP (x)/xn−1 = limx→±∞ fP (x)/(1 + x2 )(n−1)/2 . Let x0 = ±∞. From the same argument of the proof of Theorem 3.9, by considering the Möbius transform xi 7→ (axi + b)/(cxi + d), i = 0, 1, . . . , n − 1, we find that the fixed point x0 = ±∞ in (3.22) can be moved to an arbitrary point in R. Theorem 3.13. Let x0 ∈ R∪{±∞} be fixed arbitrarily. M ∈ MP is uniquely represented with 2n − 1 parameters (w0 , . . . , wn−1 , x1 , . . . , xn−1 ) as M= n−1 X wi fP (xi )fP (xi )⊤ + w0 fP (x0 )fP (x0 )⊤ , wi > 0, i=1 xi 6= x0 , −∞ < x1 < · · · < xn−1 ≤ ∞, where we assume that fP (±∞) = (0, . . . , 0, 1)⊤ . The counterpart for the moment cone (3.9) for trigonometric functions is obtained using Lemma 3.1. Theorem 3.14. Let t0 ∈ (− 21 , 21 ] be fixed arbitrarily. M ∈ MF is uniquely represented with 2n − 1 parameters (w0 , . . . , wn−1 , t1 , . . . , tn−1 ) as M= n−1 X wi fF (ti )fF (ti )⊤ + w0 fF (t0 )fF (t0 ), wi > 0, i=1 1 1 ti 6= t0 , − < t1 < · · · < tn−1 ≤ . 2 2 A square matrix M = (mi,j ) is said to be Hankel if mi,j = mk,l when i + j = k + l. For example, matrices M in (3.12) and (4.1) are Hankel. Obviously, each M ∈ MP should be an n × n positive definite Hankel matrix. It is known that the converse is also true. Proposition 3.15. The moment cone MP in (3.8) is characterized as  MP = M ≻ 0 | M is Hankel . For the proof, see (9.1) of Karlin and Studden (1966), p. 199. This also gives a unique representation of MP with 2n − 1 parameters (m0 , m1 , . . . , m2n−2 ). Theorem 3.9 combined with Proposition 3.15 implies that group A acts on the cone of (positive definite) Hankel matrices. For the Möbius group action on Hankel matrices, see also Heinig and Rost (1989, 2010). 3.4 Invariance under the Möbius group In this subsection, we consider the polynomial regression. We formalized our optimal experimental design problem to find the minimizer M ∈ MP of Vol1 (γM −1 ) in (2.6). Theorem 3.16. For M ∈ MP and A ∈ A, Vol1 (γM −1 ) = Vol1 (γ(AM A⊤ )−1 ). 15 Theorem 3.16 and Theorem 3.9 imply that the minimizer of Vol1 (γM −1 ) with respect to M ∈ MP forms an orbit (or a union of orbits) on MP . Proof. Let y = ϕ(x) = ϕ(x; a, b, c, d) = (ax + b)/(cx + d). Then, fP (y) = fP (ϕ(x)) = AfP (x)λ(x), where A = A(a, b, c, d), λ(x) = λ(x; a, b, c, d). Taking derivatives with respect to x, ad − bc gP (y)ϕ̇(x) = AfP (x)λ̇(x) + AgP (x)λ(x), ϕ̇(x) = . (cx + d)2 Therefore, and   λ̇(x)

λ(x) ϕ̇(x)   fP (y), gP (y) = A fP (x), gP (x)   λ(x)  0 ϕ̇(x)  1 
2 fP (y)⊤ −1 M fP (y), gP (y) det gP (y)⊤  1 
2 λ(x)2 fP (x)⊤ −1 −1 ⊤ −1 (A M(A ) ) fP (x), gP (x) = det . gP (x)⊤ ϕ̇(x) By combining this with fP (y)⊤ M −1 fP (y) = fP (x)⊤ (A−1 M(A−1 )⊤ )−1 fP (x)λ(x)2 and dy = ϕ̇(x)dx, we have   12
fP (y)⊤ Z ∞ det M −1 fP (y), gP (y) gP (y)⊤ dy Vol1 (γM −1 ) = 2 fP (y)⊤ M −1 fP (y) −∞ 1  
2 fP (x)⊤ −1 −1 ⊤ −1 Z ∞ det (A M(A ) ) fP (x), gP (x) gP (x)⊤ dx =2 fP (x)⊤ (A−1 M(A−1 )⊤ )−1 fP (x) −∞ = Vol1 (γA−1 M (A−1 )⊤ ).    xi with variTheorem 3.17. The volumes of the weighted polynomial design pi 1≤i≤N   yi ance σP2 (x; a0 , b0 , c0 , d0), and the design , y = ϕ(xi ; a, b, c, d) with variance pi 1≤i≤N i σP2 (y; a′, b′ , c′ , d′ ) are the same, where (a′ , b′ , c′ , d′ ) is determined by ϕ(·; a′, b′ , c′ , d′) = ϕ(·; a0, b0 , c0 , d0 ) ◦ ϕ−1 (·; a, b, c, d). 16 Proof. In (3.18) of the proof of Theorem 3.9, let wi = pi /σP2 (xi ; a0 , b0 , c0 , d0 ). Then, by (3.21), 1 pi pi vi = = 2 . 2 2 λ(xi ; a, b, c, d) σP (xi ; a0 , b0 , c0 , d0 ) σP (yi ; a′ , b′ , c′ , d′) This means that information matrices M1 and M2 of the two designs satisfy M1 = AM2 A⊤ and hence have the same volume by Theorem 3.16. 3.5 D-optimal design for weighted polynomial regression We characterize the D-optimal design for the weighted polynomial regression as an orbit of the Möbius group action. We start from the fact that in the Fourier regression, the uniform design is D-optimal. Proposition 3.18 (Guest (1958)). In the Fourier regression with the basis (3.1), among the n-point discrete design, only the uniform design 0  i n+1 ti − θ , t0i = − , (3.23) 1 n 2n n 1≤i≤n
1 1 is D-optimal, where θ ∈ − 2n , 2n is an arbitrary constant. The information matrix at the optimal point is the identity In .   t Let MF be an information matrix of a Fourier design i . By making a change pi 1≤i≤n of variables yi = tan(πti ) and yi = ϕ(xi ; a, b, c, d), we have from (3.4), (3.14), and (3.19) that X MF = fF (ti )fF (ti )⊤ pi i  1 pi B ⊤ =B fP (yi )fP (yi ) 2 σ (y ) i P iX  λ(xi ; a, b, c, d)2 ⊤ =BA fP (xi )fP (xi ) 2 p A⊤ B ⊤ 2 i σ (x ; a, b, c, d)λ(x ; a, b, c, d) i i P i X ⊤ =BAMP A⊤ B ⊤ , where MP = X fP (xi )fP (xi )⊤ i 1 pi σP2 (xi ; a, b, c, d)   xi is the information matrix of the design for the weighted polynomial regression pi 1≤i≤n with variance function σP2 (x; a, b, c, d). Because det(MF ) = det(AB)2 det(MP ), the Doptimal problem for searching optimal ti and pi in the Fourier regression are equivalent to searching for optimal xi and pi in the weighted polynomial regression. Hence, Proposition 3.18 is translated into the weighted polynomial regression as follows. 17 Theorem 3.19. In the weighted polynomial regression of degree n − 1 with  variance x σP2 (x; a0 , b0 , c0 , d0), among the n-point discrete design, only the design 1i with n 1≤i≤n xi = ϕ−1 (ϕ(tan(πt0i ); s, −t, t, s); a0 , b0 , c0 , d0 )) is D-optimal, where t0i is given in (3.23), and s, t are arbitrary numbers such that s2 + t2 = 1. The information matrix at the D⊤ −1 ⊤ −1 ⊤ −1 optimal point is A−1 is given 0 (B B) (A0 ) , where A0 = A(a0 , b0 , c0 , d0 ), and (B B) in (3.7). Proof. Note that yi = tan(π(t0i − θ)) = cos(πθ) tan(πt0i ) − sin(πθ) = ϕ(tan(πt0i ); s, −t, t, s), sin(πθ) tan(πt0i ) + cos(πθ) where s = cos(πθ), t = sin(πθ). xi = ϕ−1 (yi ; a0 , b0 , c0 , d0). When (a0 , b0 , c0 , d0 ) = (1, 0, 0, 1), Theorem 3.19 reduces to Theorem 3.3 of Dette, et al. (1999). 4 Tube-volume optimal design for n = 3 In the previous section, we discussed the Fourier regression and the polynomial regression having the basis of (3.1) and (3.2), respectively, of a general dimension n. In this section, we treat the case n = 3. This is the simplest non-trivial case, because when n = 2, Vol1 (γM −1 ) = 2π irrespective of M. When n = 3, the problem is reduced to the minimization of   12    1 0 2 1 x x M −1  x 1  det  Z ∞ 0 1 2x x2 2x   Vol1 (γM −1 ) = 2 dx 1 −∞
1 x x2 M −1  x  x2 Z ∞ p h1 (x) =2 dx, −∞ h0 (x) where h1 (x) =m4 (−m0 m23 + m0 m2 m4 + 2m1 m2 m3 − m21 m4 − m32 ) + 4m3 (−m0 m2 m4 + m0 m23 − 2m1 m2 m3 + m21 m4 + m32 )x + 6m2 (m0 m2 m4 − m0 m23 + 2m1 m2 m3 − m21 m4 − m32 )x2 + 4m1 (m0 m23 − m0 m4 m2 − 2m1 m2 m3 + m21 m4 + m32 )x3 + m0 (m0 m2 m4 − m0 m23 + 2m1 m2 m3 − m21 m4 − m32 )x4 , h0 (x) =m2 m4 − m23 + 2(−m1 m4 + m2 m3 )x + (m0 m4 + 2m1 m3 − 3m22 )x2 + 2(−m0 m3 + m1 m2 )x3 + (m0 m2 − m21 )x4 , 18 with respect to   m0 m1 m2 M = m1 m2 m3  ≻ 0. m2 m3 m4 (4.1) The volume becomes an elliptic integral, which does not have an explicit expression in general. Moreover, the number of parameters to be optimized is four. (Note that the p integrand h1 (x)/h0 (x) is a homogeneous function in m0 , . . . , m4 ). We will solve this minimization problem using the Möbius invariance. 4.1 Orbital decomposition The Möbius group action defines an equivalent class on the moment cone MP . We define M0 ∼ M1 for M0 , M1 ∈ MP if M1 = AM0 A⊤ for some A ∈ A. The orbit passing through M is denoted by A(M) = {AMA⊤ | A ∈ A} = {M1 | M1 ∼ M}. The goal of this subsection is to prove the orbital decomposition of the polynomial moment cone MP . Let   1 0 v Mv = 0 v 0 . (4.2) v 0 1 Theorem 4.1. MP = G v∈(0, 13 ] A(Mv ) = G v∈[ 13 ,1) A(Mv ), where ⊔ is the disjoint union. Proof. This is a consequence of Lemmas 4.2 and 4.3 below. Lemma 4.2. For any M ∈ MP and 0 < v < 1, there exists A ∈ A such that AMA⊤ = Mv . That is, M ∼ Mv . Lemma 4.3. Mv′ ∼ Mv if and only if v ′ = v or v ′ = (1 − v)/(1 + 3v). The proofs of Lemmas 4.2 and 4.3 are given in Appendix A.2 and A.3, respectively. Note that the map 1−v v 7→ 1 + 3v defines a one-to-one correspondence between (1, 1/3) and (1/3, 1), and v = 1/3 is the fixed point of this map. The stabilizer of A at M ∈ MP is defined as AM = {A ∈ A | AMA⊤ = M}. This is a subgroup of A. 19 Theorem 4.4. When v 6= 1/3, A Mv     0 0 1   1 0 0 = {A(±1, 0, 0, ±1), A(0, ±1, ±1, 0)} = 0 ±1 0 , 0 ±1 0 ,   0 0 1 1 0 0 and when v = 1/3, AMv = {A(s, −t, t, s) | s2 + t2 = 1} ⊔ {A(−s, t, t, s) | s2 + t2 = 1}. In particular, dim AMv = ( 0 (v ∈ (0, 31 ) ∪ ( 13 , 1)), 1 (v = 13 ). Proof. The proof follows from the proof of Lemma 4.3. The details are omitted. Theorem 4.5. The dimension of orbit A(Mv ) passing through Mv is ( 4 (v ∈ (0, 13 ) ∪ ( 13 , 1)), dim A(Mv ) = 3 (v = 31 ). f = (m Proof. Let M = (mi+j−2 )i,j=1,2,3 and M e i+j−2 )i,j=1,2,3 be 3 × 3 Hankel matrices, and f = AMA⊤ . Picking up the let A = A(a, b, c, d) (3 × 3 matrix in (3.16)). Assume that M (1, 1), (1, 2), (1, 3), (2, 3), and (3, 3) elements and rearranging them, we have m e = F m, ⊤ ⊤ where m = (m0 , m1 , m2 , m3 , m4 ) , m e = (m e 0, m e 1, m e 2, m e 3, m e 4 ) , and   4 d 4cd3 6c2 d2 4c3 d c4 3 2  bd3 3bdc2 + 3ad2 c bc3 + 3adc2 ac3    2 2 ad 2+ 3bcd 2 2 2 2 2 2 2 2 2 b d 2cdb + 2ad b b c + 4abdc + a d 2cda + 2bc a a c F = F (a, b, c, d) =  .  3 b d cb3 + 3adb2 3bda2 + 3b2 ca da3 + 3bca2 a3 c  b4 4ab3 6a2 b2 4a3 b a4 Let mv = (1, 0, v, 0, 1)⊤. The tangent space of the orbit at column vectors of  0 0  ∂ ∂ ∂ ∂  0 1 F (1, 0, 0, 1)mv =  , , , 2v 0 ∂a ∂b ∂c ∂d  0 3v 4 0 Mv is spanned by the four  0 4 3v 0   0 2v  . 1 0 0 0 The rank of this matrix is 4 when v 6= 1/3 and 3 when v = 1/3. From Theorems 4.4 and 4.5, we see that dim AMv + dim A(Mv ) = 4 = dim A as expected (e.g., Kawakubo (1992)). 20 4.2 Minimization over cross-section From the orbital decomposition (Theorem 4.1) and the invariance of the volume on an orbit (Theorem 3.16), the optimization problem is reduced to the minimization of Vol1 (γMv−1 ) with respect to v, where Mv is defined in (4.2). The range of v is taken to be (0, 13 ] or [ 13 , 1). We write Vol1 (γMv−1 ) = len(v) shortly. From the definition (2.6), Z ∞ len(v) = Vol1 (γMv−1 ) = 2 s(x; v)dx, −∞ where s(x; v) = q 1−v2 v √ 1 + 6vx2 + x4 1 + ( v1 − 3v)x2 + x4 . Note that len(v) is an elliptic integral. The following is the main theorem of this section. Theorem 4.6. The minimizer of Vol1 (γM −1 ) in (2.6) over M ∈ MP is given if and only if M is in the orbit   1 0 13 M ∼ M1/3 = 0 13 0  . 1 0 1 3 p The minimum volume is 4π 2/3. Proof. Because of Theorem 4.1, it is enough to take the range v ∈ (0, 31 ]. We use the inequality z 1 √ ≥ 1 − , |z| < 1. (4.3) 2 1+z The equality holds iff z = 0. Noting that 1 1 (1 − 3v)x2  1  √ q = 1+ , ≥ 2 1 + x2 (1 + x2 )2 1 + 6vx2 + x4 (1 + x2 ) 1 − 2(1−3v)x 2 2 (1+x ) s(x; v) is bounded below by s(x; v) = q (1 + 1−v2 (1 + 6vx2 + x4 ) v ( v1 − 3v)x2 + x4 )(1 +  (1 − 3v)x2  1+ . (1 + x2 )2 x2 ) Therefore, len(v) is bounded below by len(v) = 2 Z ∞ s(x; v)dx. −∞ This integral can be evaluated by counting the residues. When v < 1/3, the poles are r r  (1 − v)(1 + 3v) (1 + v)(1 − 3v) i , − ±ix1 = ± 2 v v 21 i ±ix2 = ± 2 r (1 − v)(1 + 3v) + v r  (1 + v)(1 − 3v) , v and ±ix0 = ±i. Denote the residues for +ix1 , +ix2 , and +ix0 by Res(+ix1 ), Res(+ix2 ), and Res(+ix0 ), respectively. Then, the integral is evaluated as len(v) = 2 × 2πi(Res(+ix1 ) + Res(+ix2 ) + Res(+ix0 )) q q   v(1+3v) 1−v2 3 2 2π 3v + 6v − 5v + 8 v 1−v = . 2 4(1 + v) The derivative is 2 3 4 5 − 38v + 14v + 18v + 9v + 48v d p len(v) = −2π dv 8(1 + v)2 v(1 − v 2 ) q v(1−v) 1+3v . (4.4) By applying inequality (4.3), r 3 n o v(1 − v) v(1 − v) 19 3 2 =q v(1 − v)(1 + 3v) − 1 ≥ v(1 − v) 1 − 1 + 3v 2 2 4 9 v(1 − v)(1 + 3v) 4 (the equality holds iff v = 1/3), the numerator of (4.4) is bounded below by n o 19 3 v(1 − v)(1 + 3v) − 1 5 − 38v + 14v 2 + 18v 3 + 9v 4 + 48v v(1 − v) 1 − 2 2 4 2 3 4 = (1 − 3v)(5 − 23v + 53v − 12v − 108v + 81v 5 ), which is positive for 0 < v < 1/3. Therefore, has the unique minimum at v = 1/3. Since s(x; v) ≥ s(x; v), d len(v) dv < 0 for 0 < v < 1/3, and len(v) min len(v) ≥ min len(v) = len(1/3). v∈(0,1/3] v∈(0,1/3] Moreover, since s(x; v) = s(x; v) at v = 1/3, min len(v) ≤ len(1/3) = len(1/3). v∈(0,1/3] Therefore, min len(v) = len(1/3) = 4π v∈(0,1/3] p 2/3. Point v = 1/3 is the unique minimizer, because this is the unique minimizer of len(v). Figure 4.1 depicts the objective function len(v) and its lower bound len(v) for v ≤ 1/3. 22 Figure 4.1: len(v) (solid line) and its lower bound len(v) for v ≤ 1/3 (dashed line). As shown in (3.12), the information matrix M1/3 is the counterpart of the information matrix for the uniform design in the Fourier regression. Recall the decomposition of A(a, b, c, d) in Proposition 3.8. We already know from Theorem 4.4 that, for s2 + t2 = 1, A(s, −t, t, s)M1/3 A(s, −t, t, s)⊤ = A(−s, t, t, s)M1/3 A(−s, t, t, s)⊤ = M1/3 . Moreover,   A(q, r, 0, 1)M1/3 A(q, r, 0, 1)⊤ =  q2 3  q2 2 1 r + r 3  q2 2 2 r + r r(q + r 2 ) . 3 + r 2 r(q 2 + r 2 ) (q 2 + r 2 )2 Theorem 4.6 can be written in the following form. Theorem 4.7. The minimizer of Vol1 (γM −1 ) in (2.6) over M ∈ MP is given when and only when M is of the form:   q2 2 1 r + r 3   q2 2 2 M = k r + r r(q + r 2 ) , q 6= 0, k > 0. 3 q2 + r 2 r(q 2 + r 2 ) (q 2 + r 2 )2 3 Remark 4.8. The minimum tube-volume M ∈ MP is attained when and only when the curve  1 1 (γM −1 )+ = ψM −1 (x) = M − 2 f (x)/kM − 2 f (x)k | x ∈ X = (−∞, ∞) p forms a circle. Moreover, in that case, the circle length is 2π 2/3. Finally, we characterize the tube-volume optimal design as a three-point design. The polynomial design corresponding to the Fourier uniform design is given in (3.12). The 23 tube-volume optimal design is obtained as an orbit of the transformation passing through the design in (3.12). In the following, let  0  1  ti − 3 0 13 = , (4.5) 1 1 1 1 3 3 i=1,2,3 3 3 a three-point uniform design in the Fourier regression. 2 Theorem 4.9. In the weighted polynomial regression  with  variance function σP (x; a0 , b0 , c0 , xi d0 ), the three-point tube-volume optimal design is , where pi i=1,2,3 xi = ϕ−1 (tan(πt0i ); a, b, c, d), pi = k σP2 (xi ; a0 , b0 , c0 , d0 ) . σP2 (xi ; a, b, c, d) (4.6) Here, t0i is defined in P (4.5), a, b, c, d are arbitrarily given so that ad − bc 6= 0, and k > 0 is a constant so that pi = 1. The tube-volume optimal design includes the D-optimal designs as special cases where ϕ(·; a, b, c, d) = ϕ(·; s, −t, t, s) ◦ ϕ(·; a0, b0 , c0 , d0 ) holds for some s2 + t2 = 1. Proof. The  optimal design is the orbit of the Möbius group passing through the deyi , yi = tan(πt0i ), with variance function σP2 (y; 1, 0, 0, 1). By Theorem sign 1 3 i=1,2,3   x ′ ′ ′ ′ is an optimal design under 3.17 with (a , b , c , d ) = (1, 0, 0, 1), we see that 1i 3 i=1,2,3 variance function σP2 (x; a, b, c, d). By Lemma 2.3 with σ12 (x) = σP (x; a, b, c, d), σ22 (x) = σP (x; a0 , b0 , c0 , d0), we have pi in (4.6). The results for D-optimality is proved in Theorem 3.19. Theorem 4.10. In the Fourier regression, the three-point tube-volume optimal design is given as ) (1   −1 0 tan (q tan(π(t − θ)) + r) i π ti 2  , (4.7) = 1+(q tan(π(t0i −θ))+r)2 pi i=1,2,3 k 0 2 1+tan (π(t −θ)) i where k is a normalizing constant so that P i=1,2,3 pi = 1, q 6= 0,and r,θ are arbitrarily t0i − θ given. In particular, the uniform design (D-optimal design) is a tube1 i 3 volume optimal design. i=1,2,3 Proof. Let xi and pi be given in (4.6) when (a0 , b0 , c0 , d0 ) = (1, 0, 0, 1). σP2 (x; a0 , b0 , c0 ,   Then, t d0 ) = σP2 (x), and from Theorem 3.3, the optimal design is given by i , ti = π1 tan−1 (xi ). pi 24 Let (a, b, c, d) be chosen such that ϕ−1 (·; a, b, c, d) = ϕ(·; q, r, 0, 1) ◦ ϕ(·; s, −t, t, s) with s = cos(πθ), t = sin(πθ). Then, ϕ(tan(πt0i ); s, −t, t, s) = tan(π(t0i − θ)) and xi = q tan(π(t0i − θ)) + r, and we have ti = π1 tan−1 (xi ) in (4.7).
On the other hand, since ϕ(·; a, b, c, d) = ϕ(·; s, −t, t, s)◦ϕ ·; 1q , − rq , 0, 1 , σP2 (xi ; a, b, c, d)

2 = σP2 xi ; 1q , − qr , 0, 1 = 1 + q12 (xi − r)2 . Therefore, pi = k (1 + x2i )2 σP2 (xi ; 1, 0, 0, 1) = k
2 , σP2 (xi ; a, b, c, d) 1 + q12 (xi − r)2 which is equivalent to pi in (4.7). The uniform design corresponds to the case (q, r) = (1, 0). 4.3 Numerical comparisons Here we conduct a small numerical experiment to see the difference of the width of the simultaneous confidence band under optimal and non-optimal designs. The model we √ √ use is the Fourier regression f (x) = (1, 2 sin(2πx), 2 cos(2πx))⊤ , x ∈ X = (−1/2, 1/2] with the variance function σ 2 (x) = 1. Three designs   1   1  1 1  −√4 0√ − 4 0 41 − 3 0 31 , D2 = , D3 = 1+ 3 2− 3 1+4√3 D1 = 1 1 2 1 1 1 3 3 3 6 3 6 6 3 6   xi , the information matrix is given by Σ−1 = are compared. For each design pi i=1,2,3 P3 ⊤ 2 i=1 pi f (xi )f (xi ) /σ (xi ). Design D1 is the pD-optimal and tube-volume (TV-) optimal design with det(Σ) = 1 and Vol1 (γΣ ) = 4π 2/3 = 10.260. Design D2 is TV-optimal, but is not D-optimal with det(Σ) = 27/8 = 3.375. Design D3 is considered so that the value of det(Σ) is equal to that of D2 , and is neither D-optimal nor TV-optimal. The value of D-criterion det(Σ), the tube-volume criterion Vol1 (γΣ ), and the width wα (α = 0.1, 0.05, 0.01) of 100(1 − α)% simultaneous confidence bands are listed in Table 4.3. wα is estimated by the upper α quantile of the random variable maxx∈X |bb⊤ f (x) − b⊤ f (x)| by simulation with 10,000 replications. Table 4.1: Comparisons of optimal/non-optimal designs. design D1 D2 D3 D-opt TV-opt det(Σ) ✓ ✓ 1.000 ✗ ✓ 3.375 ✗ ✗ 3.375 Vol1 (γΣ ) w0.1 w0.05 w0.01 10.260 6.99 7.94 9.55 10.260 19.86 23.15 30.59 11.262 26.06 30.58 40.35 As the theorems state, design D1 has the shortest simultaneous confidence band. By comparing designs D2 and D3 having the same value of det(Σ), we see how the design affects the width of the simulations confidence bands. 25 4.4 Summary and remaining problems In this paper, we have proposed the tube-volume (TV) criterion Vol1 (γΣ ) in (2.6) in experimental design. If a design is tube-volume optimal and simultaneously, D-optimal minimizing maxx∈X f (x)⊤ Σ−1 f (x), the design is optimal that attains the minimum bandwidth of simultaneous confidence bands. Then, the proposed criterion was applied to Fourier regression model that is a standard model in linear optimal design theory, and weighted polynomial regression model that is mathematically equivalent to the Fourier regression model. The Möbius group keeps the tube-volume criterion invariant, whereas the subgroup O(2, R) of the Möbius group keeps the D-criterion invariant. Using the Möbius invariance, when n = 3, we found that the tube-volume optimal designs in the Fourier regression and the weighted polynomial regression form an orbit of the Möbius group. The tube-volume optimal designs contain D-optimal designs as special cases. This means that in the Fourier regression, the uniform design is a universal optimal design minimizing both tube-volume criterion and D-criterion. We conjecture that for all n, the tube-volume optimal design is characterized as an orbit of the Möbius group containing D-optimal designs. One supporting observation is that for small n (n ≤ 6), tube-volume local optimality at the D-optimal designs can be proved by direct calculations. That is, the Hessian matrix of the tube-volume criterion evaluated at the D-optimal design is positive semi-definite, and the null space of the Hessian matrix corresponds to the tangent space of the orbit of Möbius group action. However, the proof for general n remains outstanding. Throughout the paper, we just dealt with the case where the explanatory variable is one-dimensional. However, the volume-of-tube method works for the construction of the simultaneous confidence bands for regression with multidimensional explanatory variables except for the conservativeness (ii) of Proposition 2.1, and the volume-optimality is welldefined. For example, we can discuss the volume-optimality of the p-variate polynomial regression model with the basis vector f (x) = 1, (xi )1≤i≤p , (xi xj )1≤i≤j≤p , . . . , (xi1 · · · xid )1≤i1 ≤···≤id ≤p
⊤ ∈ R[x1 , . . . , xp ]( ). p+d d By the same argument as the univariate case, we can prove that the multivariate Möbius p p transform ϕ : R → R defined by   Ax + b A b ⊤ x = (x1 , . . . , xp ) 7→ ϕ(x; A, b, c, d) = ⊤ 6= 0 , det ⊤ c d c x+d where A ∈ Rp×p , b, c ∈ Rp×1, d ∈ R (e.g., Kato and McCullagh (2014)) remains the invariance (volume preserving property) Volp (γM −1 ) = Volp (γ(AM A⊤ )−1 ) of Theorem 3.16. However, the treatment of the multidimensional case (see, e.g., Lasserre (2009) for the moment cone) remains a future topic of research. 26 A Appendix: Proofs A.1 Proof of Lemma 3.4 Proof. Let tk = k/n − (n + 1)/(2n) and xk = tan(πtk ). The (i, j) element of M is n mi,j n 1 X (i−1)+(j−1) 1 1 X i+j−2 = xk sin (πtk ) cosn−i−j (πtk ). = n k=1 (1 + x2k )2(n−1) n k=1 Then, apply Lemma A.1 below. Lemma A.1. Let ti = i/n − d, i = 1, . . . , n, where d is a constant. Then, n Γ(k + 12 )Γ(n − k − 21 ) 1 X 2k sin (πti ) cos2n−2k−2 (πti ) = . n i=1 πΓ(n) Proof. From the duplication formula for the gamma function, it suffices to show that n 1 X 2k (2k)!(2n − 2k − 2)! −2n+2 2 . sin (πtj ) cos2n−2k−2 (πtj ) = n j=1 k!(n − k − 1)!(n − 1)! Let ω = eiπ/n , c = e−iπd . The left-hand side times (2i)2k 22n−2k−2 = 22n−2 (−1)k is n 1X j (cω − c−1 w −j )2k (cω j + c−1 w −j )2n−2k−2 n j=1   n 1 X X 2k (−1)s cs ω sj c−(2k−s) ω −(2k−s)j = n j=1 0≤s≤2k s   X 2n − 2k − 2 t tj −(2n−2k−2−t) −(2n−2k−2−t)j cω c ω . × t (A.1) 0≤t≤2n−2k−2 Here, the contribution of the summation with respect to j is ( n 1 (s + t = n − 1), 1 X (−2n+2+2s+2t)j ω = n j=1 0 (otherwise). Hence, (A.1) is equal to X (−1) max(0,2k−n+1)≤s≤min(2k,n−1) s  2k s   (2k)!(2n − 2k − 2)! 2n − 2k − 2 = (−1)k . n−1−s k!(n − k − 1)!(n − 1)! The last equality follows from (1.41) of Gould (2010). 27 A.2 Proof of Lemma 4.2 The proof of Lemma 4.2 is divided into three parts (lemmas). Lemma A.2. For M ∈ MP , there exist u, w > 1 such that   u 1 1 M ∼ 1 1 1  . 1 1 w Lemma A.3. For u, w > 1, there  u 1 1 √ exist ũ, ṽ, w̃ > 0, ṽ < ũw̃ such that    1 1 ũ 0 ṽ 1 1  ∼  0 ṽ 0  . 1 w ṽ 0 w̃ √ Lemma A.4. For ũ, ṽ, w̃ > 0, ṽ < ũw̃,  ũ 0 0 ṽ ṽ 0 there exist v ∈ (0, 1) such that  ṽ 0  ∼ Mv . w̃ Proof of Lemma A.2. We start from a canonical form in (3.22): M = w1 fP (x1 )fP (x1 )⊤ + w2 fP (x2 )fP (x2 )⊤ + w0 fP (±∞)fP (±∞)⊤ , 1/4 where fP (x) = (1, x, x2 )⊤ , fP (±∞) = (0, 0, 1)⊤ . For A = A(a, b, c, d) with a = w2 (x1 − 1/4 1/4 x2 ), b = −w2 x1 , c = 0, d = −w2 , we have   u 1 1 A 1 1 1  A⊤ = M, u = w1 d−4 + 1, w = w0 a−4 + 1. 1 1 w Proof of Lemma A.3. Let   u 1 1 M =  1 1 1 , 1 1 v   ũ 0 ṽ f =  0 ṽ 0  , M ṽ 0 w̃  d2 2cd c2 A = bd bc + ad ac . b2 2ab a2  f has a solution (a, b, c, d) such that ad − bc 6= 0. It We confirm that equation AMA⊤ = M is enough to show that under the assumption a, d 6= 0, a solution (a, b, c, d) satisfies (AMA⊤ )1,2 = 0, (AMA⊤ )2,3 = 0, ad − bc 6= 0. 28 (A.2) Solving (AMA⊤ )1,2 = 0 with respect to b yields b = −a c3 v + 3c2 d + 3cd2 + d3 , d3 f1 (c/d; u) (A.3) where f1 (c; u) = (c + 1)3 + (u − 1) if f1 (c/d; u) 6= 0. Substituting (A.3) into (AMA⊤ )2,3 , we have 0 = (AMA⊤ )2,3 = a{(v − 1)c4 + (u − 1)d4 + (c + d)4 }d6 f (c/d; u, v) , {d3 f1 (c/d; u)}3 (A.4) where f (c; u, v) =c6 (−2 + 3v − v 2 ) + c5 (−6 + 7v − uv 2 ) + c4 (−10 + 15v − 5uv) + c3 (−10u + 10v) + c2 (10 − 15u + 5uv) + c(6 − 7u + u2 v) + (2 − 3u + u2 ). Substituting (A.3) into ad − bc, we have d4 f2 (c/d; u, v) , 0 6= ad − bc = a 3 d f1 (c/d; u) f2 (c; u, v) = 4c + 6c2 + 4c3 + u + c4 v. In the numerator of (A.4), if (v − 1)c4 + (u − 1)d4 + (c + d)4 = 0, then c = d = 0, and hence, ad − bc = 0. Now, we examine whether for all u, v > 1, there is a real x such that f (x; u, v) = 0, f1 (x; u) 6= 0, f2 (x; u) 6= 0. (A.5) Once a solution x = x∗ is obtained, we have a solution that a, d 6= 0 are given arbitrarily, c = dx∗ , and b is determined from (a, c, d) in (A.3). Write f (·) = f (·; u, v) shortly. It is easily shown that when u, v > 1, neither f (0) =(u − 2)(u − 1) > 0, f (−1) = − (u − v)(u − 1)(v − 1) > 0, f (±∞) = − (v − 2)(v − 1) > 0 nor f (0) < 0, f (−1) < 0, f (±∞) < 0 is true, where f (±∞) = limc→±∞ c−6 f (c). This means that f (x; u, v) = 0 has a real solution x = x∗ . In order to check f1 (x∗ ; u) 6= 0 and f2 (x∗ ; u, v) 6= 0, we need to check whether f and f1 (or f2 ) have a common factor. For this purpose, we calculate the resultants R(f, f1 ) and R(f, f2 ): R(f, f1 ) = (u − 2)(u − 1)h(u, v)2, 29 R(f, f2 ) = h(u, v), where h(u, v) = −28 + 54u − 27u2 + 54v − 105uv + 54u2v − 27v 2 + 54uv 2 − 30u2v 2 + u3 v 3 . (A.6) (For resultant, see, e.g., Prasolov (2004). Applications in statistics can be found in Drton, et al. (2009).) As shown in Lemma A.5 later, in the region u, v > 1, h(u, n) = 0 iff u = 2. Therefore, when u 6= 2, we have established that x = x∗ satisfying (A.5) exists, and hence, a solution (a, b, c, d) satisfying (A.2) exists. When u = 2, (A.4) is reduced to 0= f (c; 2, v) (2 − v)c(−2 − 4c − 3c2 − c3 − c4 + c4 v) = f1 (c; 2)3 (2 + c)2 (1 + c + c2 )3 and c = 0 is a solution. From (A.3), b = −a/2, and a 6= 0, d 6= 0 are arbitrarily given. In fact, for A = A(a, −a/2, 0, d),     2 1 1 2d4 0 a2 d2 /2 . a2 d2 /2 0 A 1 1 1 A⊤ =  0 2 2 4 1 1 v a d /2 0 a (v − 7/8) Lemma A.5. Let h(u, v) be defined in (A.6). When u, v > 1, h(u, v) ≥ 0, and the equality holds iff u = v = 2. Proof. Fix u and consider h(u, v) as a function of v. Note that h(u, 1) = (u − 1)3 > 0. ∂ h(u, v) ∂v = 54 − 105u + 54u2 + (−54 + 108u − 60u2 )v + 3u3 v 2 . hv (u, v) = hv (u, 1) = 3u(u − 1)2 > 0. It is easy to see that hv (u, v) = 0 has a real solution iff 3/2 ≤ u ≤ 3. Therefore, when u ≤ 3/2 or u ≥ 3, hv (u, v) is always positive. Combined with h(u, 1) > 0, this means that h(u, v) > 0. Consider the case 3/2 < u < 3. The largest zero of hv (u, v) is p 9 − 18u + 10u2 + 3 (u − 1)3 (3 − u)(2u − 3) v∗ (u) = > 1. u3 At this point, the function h(u, v) takes a local minimum  27 h(u, v∗ (u)) = 6 (u − 1)4 (−54 + 108u − 72u2 + 18u3 − u4 ) u p − 2(3 − u)(2u − 3) (u − 1)(3 − u)(2u − 3)  =27(u − 1)4 (u − 2)2 (−54 + 108u − 72u2 + 18u3 − u4 ) p −1 + 2(3 − u)(2u − 3) (u − 1)(3 − u)(2u − 3) . It is easy to check that −54 + 108u − 72u2 + 18u3 − u4 > 0 for 3/2 < u < 3. Hence, h(u, v∗ (u)) ≥ 0 for 3/2 < u < 3 and the equality holds when u = 2. When u = 2, v∗ (2) = 2. 30 Proof of Lemma A.4. For A = A(a, b, c, d) with b = c = 0,    4  ũ 0 ṽ d ũ 0 a2 d2 ṽ a2 d2 ṽ 0 . A 0 ṽ 0  A⊤ =  0 ṽ 0 w̃ a2 d2 ṽ 0 a4 w̃ √ By letting a = w̃ −1/4 , d = ũ−1/4 , v = a2 d2 ṽ = ṽ/ ũw̃, we obtain the result. √ Remark A.6. In the proof of Lemma A.4, by letting d = ũ−1/4 , a = ṽ −1/2 / 3d = √ ũ1/4 / 3ṽ 1/2 , w = a4 w̃ = ũw̃/9ṽ 2 , we have another representative group element   1 0 31 0 1 0  , w > 1 . 3 9 1 0 w 3 A.3 Proof of Lemma 4.3 Proof. The (1, 1), (2, 1), (3, 2), (3, 3) components of the equation Mv′ = AMv A⊤ are (AMv A⊤ )1,1 = c4 + d4 + 6vc2 d2 , (AMv A⊤ )2,1 = ac3 + bd3 + 3vcd(bc + ad), (AMv A⊤ )3,2 = a3 c + b3 d + 3vab(bc + ad), (AMv A⊤ )3,3 = a4 + b4 + 6va2 b2 , respectively. By solving (AMv A⊤ )2,1 = 0, we have b=− ac(c2 + 3d2 v) d(d2 + 3c2 v) when d 6= 0. (c and d cannot be 0 simultaneously because ad − bc 6= 0.) Suppose first that d 6= 0. Substituting this into (AMv A⊤ )3,2 , we have (AMv A⊤ )3,2 = a3 c(c4 − d4 )(c4 + d4 + 6c2 d2 v)(−1 + 9v 2 ) . d2 (d2 + 3c2 v)3 This becomes 0 when (i) c = 0, (ii) c = ±d, or (iii) v = 1/3. We try to solve the equation (AMv A⊤ )1,1 = (AMv A⊤ )3,3 = 1 in each case. Note that a = 0 cannot be a solution, because b becomes 0 and ad − bc = 0. (i) If c = 0, then b becomes 0 and (AMv A⊤ )1,1 = d4 , (AMv A⊤ )3,3 = a4 . Hence, (a, b, c, d) = (±1, 0, 0, ±1) (four ways) are the solutions. In each case, v ′ = (AMv A⊤ )3,1 = v. (ii) If c = ±d, then b = ∓a, (AMv A⊤ )1,1 = 2c4 (1 + 3v) = 1 = (AMv A⊤ )3,3 = 2a (1 + 3v), and hence, c = ±a. Therefore,  (1, −1, 1, 1),    (1, −1, −1, −1), (a, b, c, d) = 1/(2 + 6v)1/4 × (−1, 1, 1, 1),    (−1, 1, −1, −1) 4 31 are solutions. In each case, v ′ = (AMv A⊤ )3,1 = 1−v . 1 + 3v (iii) When v = 1/3 and d 6= 0, we have b = −ac/d and (AMv A⊤ )1,1 = (c2 + d2 )2 = 1, (AMv A⊤ )3,3 = (a2 + b2 )2 = a4 (c2 + d2 ) a4 = = 1. d4 d4 √ √ 2 . When d = a, c = −b = ∓ 1 − a2 . When d = −a, c = b = 1 − a Hence, b = ± √ ± 1 − a2 . In summary, √ √ √ √

(a, b, c, d) = a, 1 − a2 , − 1 − a2 , a , a, − 1 − a2 , 1 − a2 , a , √ √ √ √

a, 1 − a2 , 1 − a2 , −a , a, − 1 − a2 , − 1 − a2 , −a . In each case, 1 = v. 3 (iii’) When v = 1/3 and d = 0, (AMv A⊤ )2,1 = ac3 = 0. Because ad − bc 6= 0, c = 6 0, ⊤ ⊤ 4 ⊤ 4 and a = 0. In this case, (AMv A )3,2 = 0, (AMv A )1,1 = c = 1, (AMv A )3,3 = b = 1. Hence, (a, b, c, d) = (0, ±1, ±1, 0) (four ways) are the solutions. In each case, v ′ = (AMv A⊤ )3,1 = v ′ = (AMv A⊤ )3,1 = 1 = v. 3 Acknowledgment The authors are grateful to Yasuhiro Omori for his helpful comments. References Adler, R. J. and Taylor, J. E. (2007). Random Fields and Geometry, Springer. Dette, H., Haines, L. M. and Imhof, L. (1999). Optimal designs for rational models and weighted polynomial regression, The Annals of Statistics, 27 (4), 1272–1293. Dette, H. and Melas, V. B. (2003). Optimal designs for estimating individual coefficients in Fourier regression models, The Annals of Statistics, 31 (5), 1669–1692. Drton, M., Sturmfels, B. and Sullivant, S. (2009). Lectures on Algebraic Statistics, Oberwolfach Seminars 39, Springer. van Dyk, D. A. (2014). The role of statistics in the discovery of a Higgs boson, Annual Review of Statistics and Its Application, 1, 41–59. 32 Gould, H. W. (2010). Combinatorial Identities: Table I: Intermediate Techniques for Summing Finite Series: From the seven unpublished manuscripts of H. W. Gould Edited and Compiled by Jocelyn Quaintance. http://www.math.wvu.edu/~ gould/Vol.4.PDF Gross, K. I. and Holman III, W. J. (1980). Matrix-valued special functions and representation theory of the conformal group, I: The generalized gamma function, Transactions of the American Mathematical Society, 258 (2), 319–350. Guest, P. G. (1958). The spacing of observations in polynomial regression, The Annals of Mathematical Statistics, 29 (1), 294–299. Heinig, G. and Rost, K. (1989). Matrices with displacement structure, generalized Bezoutians, and Möbius transformations, Operator Theory: Advances and Applications, 40, 203–230. Heinig, G. and Rost, K. (2010). Introduction to Bezoutians, Operator Theory: Advances and Applications, 199, 25–118. Johansen, S. and Johnstone, I. M. (1990). Hotelling’s theorem on the volume of tubes: Some illustrations in simultaneous inference and data analysis, The Annals of Statistics, 18 (2), 652–684. Johnstone, I. and Siegmund, D. (1989). On Hotelling’s formula for the volume of tubes and Naiman’s inequality, The Annals of Statistics, 17 (1), 184–194. Karlin, S. and Studden, W. (1966). Tchebycheff Systems: With Applications in Analysis and Statistics, Interscience Publishers, Wiley. Kato, N. and Kuriki, S. (2013). Likelihood ratio tests for positivity in polynomial regressions, Journal of Multivariate Analysis, 115, 334–346. Kato, S. and McCullagh, P. (2014). A characterization of a Cauchy family on the complex space, arXiv:1402.1905 [math.ST]. Kawakubo, K. (1992). The Theory of Transformation Groups, Oxford University Press. Kiefer, J. and Wolfowitz, J. (1959). Optimum designs in regression problems, The Annals of Mathematical Statistics, 30 (2), 271–294. Kuriki, S. and Takemura, A. (2001). Tail probabilities of the maxima of multilinear forms and their applications, The Annals of Statistics, 29 (2), 328–371. Kuriki, S. and Takemura, A. (2009). Volume of tubes and the distribution of the maximum of a Gaussian random field, Selected Papers on Probability and Statistics, American Mathematical Society Translations Series 2, Vol. 227, No. 2, 25–48. 33 Lasserre, J. B. (2009). Moments, Positive Polynomials and Their Applications, Imperial College Press. Liu, W. (2010). Simultaneous Inference in Regression, Chapman & Hall/CRC. Lu, X. and Kuriki, S. (2017). Simultaneous confidence bands for contrasts between several nonlinear regression curves, Journal of Multivariate Analysis, 155, 83–104. Mackey, D. S., Mackey, N., Mehl, C. and Mehrmann, V. (2015). Möbius transformations of matrix polynomials, Linear Algebra and Its Applications, 470, 120–184. McCullagh, P. (1996). Möbius transformation and Cauchy parameter estimation, The Annals of Statististics, 24 (2), 787–808. Moriguti, S., Udagawa, K. and Hitotumatu, S. (1957). Iwanami Sugaku Koshiki , II (in Japanese), Iwanami. Naiman, D. Q. (1986). Conservative confidence bands in curvilinear regression, The Annals of Statistics, 14 (3), 896–906. Prasolov, V. V. (2004). Polynomials, Algorithms and Computation in Mathematics 11, Springer. Pukelsheim, F. (2006). Optimal Design of Experiments, Classics in Applied Mathematics 50, SIAM. Scheffé, H. (1959). The Analysis of Variance, Wiley. Sun, J. (1993). Tail probabilities of the maxima of Gaussian random fields, The Annals of Probability, 21 (1), 34–71. Sun, J. and Loader, C. R. (1994). Simultaneous confidence bands for linear regression and smoothing, The Annals of Probability, 22 (3), 1328–1345. Takemura, A. and Kuriki, S. (2002). On the equivalence of the tube and Euler characteristic methods for the distribution of the maximum of Gaussian fields over piecewise smooth domains, The Annals of Applied Probability, 12 (2), 768–796. Working, H. and Hotelling, H. (1929). Applications of the theory of error to the interpretation of trends, Journal of the American Statistical Association, 24 (165), 73–85. Wynn, H. P. (1985). Jack Kiefer’s contributions to experimental design, in Jack Carl Kiefer Collected Papers III: Design of Experiments, L. D. Brown, O. Olkin, J. Sacks, H, P. Wynn eds, Springer and the Institute of Mathematical Statistics, 1985, pages xvii–xxiii. Wynn, H. P. and Bloomfield, P. (1971). Simultaneous confidence bands in regression analysis (with discussions), Journal of the Royal Statistical Society, Ser. B, 33 (2), 202–221. 34 

10math.ST

RELATIONAL EXCHANGEABILITY arXiv:1607.06762v2 [math.ST] 28 Apr 2017 HARRY CRANE AND WALTER DEMPSEY Abstract. A relationally exchangeable structure is a random combinatorial structure whose law is invariant with respect to relabeling its relations, as opposed to its elements. Aside from exchangeable random partitions, examples include edge exchangeable random graphs and hypergraphs, path exchangeable processes, and a range of other network-like structures that arise in statistical applications. We prove a de Finetti-type structure theorem for the general class of relationally exchangeable structures. 1. Introduction Consider a random sequence X = (X1 , . . . , Xn ) taking values in an at most countable set S. For concreteness, we may regard S as a set of species, so that X records the species for a random sample of animals from a certain population. From X, we define an equivalence relation ∼X on [n] := {1, . . . , n} by (1) i ∼X j if and only if Xi = X j . We write Π(X) = {B1 , B2 , . . .} to denote the set partition whose blocks B1 , B2 , . . . are the equivalence classes induced by ∼X . We sometimes write ∼Π in place of ∼X when convenient. Suppose now that X = (X1 , . . . , Xn ) is exchangeable, meaning that (Xσ(1) , . . . , Xσ(n) ) =D (X1 , . . . , Xn ) for all permutations σ : [n] → [n], where =D denotes equality in law. Exchangeability of X induces exchangeability on the equivalence relation Π := Π(X) as in (1) in the sense that Πσ =D Π for all permutations σ : [n] → [n], with Πσ defined by i ∼Πσ j if and only if σ−1 (i) ∼Π σ−1 ( j). Below we present a theory of general relationally exchangeable structures which contains the above example as a special case and generalizes Kingman’s correspondence for exchangeable random partitions [4]. To appreciate the generality in which we work, let S again be a countable set but now consider a sequence X = (X1 , . . . , Xn ) taking values in S × S, so Date: May 1, 2017. Key words and phrases. Edge exchangeable graph; Kingman’s correspondence; paintbox process; exchangeable random partition. H. Crane is partially supported by NSF grants CNS-1523785 and CAREER DMS-1554092. 1 2 HARRY CRANE AND WALTER DEMPSEY Figure 1. (a) Representation of sequence of ordered pairs in S = {a, b, c, d, e} with X1 = (a, b), X2 = (c, a), X3 = (d, e), X4 = (a, c). (b) Representation of sequence of ordered pairs in S = {a, b, c, d, e} with X10 = (c, a), X20 = (d, e), X30 = (a, c), X40 = (a, b). The object in (b) can be obtained from that in (a) by reordering (X1 , X2 , X3 , X4 ) as (X2 , X3 , X4 , X1 ). In both (a) and (b), the edges are labeled according to where they appear in sequence. that each Xi is an ordered pair (ci , ri ). For a concrete application, we can think of each Xi = (ci , ri ) as identifying the caller and receiver of a phone call sampled uniformly at random from a telephone call log. While the data comes in the form of a sequence X, the structure relating the calls can be represented by a network-like object as in Figure 1. Since the sequence X is exchangeable, the model assigns equal probability to structures that are isomorphic up to relabeling edges as in Figures 1(a) and 1(b). Just as we disregarded the species names in passing from X to its induced equivalence relation ∼X in (1), we may also disregard the vertex labels in Figure 1 to obtain an edge-labeled graph, as shown in Figure 2. The edgelabeled graph is defined formally as the equivalence class X  of all sequences X 0 that yield the same structure as X after removing vertex labels: (2) X  := {X 0 : [n] → S × S : ρ X 0 = X for some bijection ρ : S → S}, where here we overload notation by allowing the bijection ρ : S → S to act on S × S by (c, r) 7→ (ρ(c), ρ(r)), so that ρ X 0 := ρ ◦ X 0 : [n] → S × S is well-defined by composition of functions. Exchangeability of X immediately implies edge exchangeability of the edge-labeled graph associated to X  [1]. For another example, let S represent a set of scientists and let X = (X1 , . . . , Xn ) record names of coauthors on a collection of n scientific journal articles sampled uniformly from a database (e.g., arXiv). Each Xi is a finite subset {si,1 , . . . , si,ki } ⊂ S. The assumption of uniform sampling again makes X an exchangeable sequence, which in turn induces exchangeability on the associated edge-labeled hypergraph constructed by disregarding vertex labels in the induced hypergraph structure in a manner analogous to the construction of edge-labeled graphs from the equivalence class (2). For another example, let S be a set of Internet Protocol (IP) addresses and let each entry of X = (X1 , . . . , Xn ) correspond to the path taken by a message sent between two IP addresses over the Internet. Each Xi corresponds to RELATIONAL EXCHANGEABILITY 3 Figure 2. (a) Edge-labeled graph induced by disregarding vertex labels in Figure 1(a). (b) Edge-labeled graph induced by disregarding vertex labels in Figure 1(b). The graphs in (a) and (b) have equal probability under an edge exchangeable model. a path (si , ai,1 , . . . , ai,ki , ti ) from source si to target ti by passing through the intermediate nodes ai,1 , . . . , ai,ki , altogether indicating that the path traversed s to ai,1 to ai,2 and so on until passing from si,k1 to ti . If X was obtained, for example, by uniform random sampling of source-target pairs (si , ti ) and then by applying an algorithm, such as traceroute, to obtain a path between si and ti , then the sequence of paths X is exchangeable and, therefore, so is the induced path-labeled structure shown in Figure 3. Another example in the realm of networks is to take each Xi to be an r-step ego network, obtained, e.g., by snowball sampling a neighborhood of size r from a randomly chosen vertex. The relationally labeled structure is constructed by piecing together the neighborhoods obtained from the r-neighborhoods of n randomly chosen “egos” in this population. We omit the details of this example and instead move on to our general treatment. Myriad other data structures arise according to a similar recipe: let S be a set of elements and R be a set of relations on the elements in S; sample an exchangeable sequence X taking values in R; construct a structure from X (as in Figure 1) and obtain the corresponding relationally labeled structure by removing the vertex labels (as in Figure 2). The resulting structure is exchangeable with respect to relabeling of its relations, a property which we call relational exchangeability. We now define this notion formally and prove a generic structure theorem for infinite relationally exchangeable structures. The formal statement is given in Theorem 2.6. 2. Relational Exchangeability Let (S, S) be any Borel space. A countably infinite sequence of S-valued random variables X = (X1 , X2 , . . .) is exchangeable if X σ =D X for every permutation σ : N → N, where =D denotes equality in distribution and X σ := (Xσ(1) , Xσ(2) , . . .) is the reordering of X according to σ. By de Finetti’s theorem [3], the distribution of every such countably infinite sequence can be expressed as a mixture of independent, identically distributed (i.i.d.) sequences. In particular, with P(S) denoting the space of probability measures 4 HARRY CRANE AND WALTER DEMPSEY Figure 3. (a) and (b) both correspond to the structures induced by taking the equivalence class over paths between vertices. In this illustration, all paths have the same source vertex and each path is labeled by a different color, so that the structure in (b) can be obtained by recoloring the paths from (a) (i.e., green to orange, red to green, and orange to red). on (S, S), there is a unique probability measure φ on P(S) such that the distribution of X can be expressed as Z (3) P(X ∈ ·) = ν∞ (·)φ(dν), P(S) where ν∞ denotes the infinite product measure induced by ν on S∞ . We call the measure φ in (3) the de Finetti measure of X. de Finetti’s theorem figures into our treatment of relationally exchangeable structures, which we now define. Note that all of the examples in Section 1, and many more that could arise in practice, involve relational structures as they are defined in the coming section. Example 2.2 provides a concrete illustration. 2.1. Relational structures. Let α : N → N ∪{0} be a signature such that α( j) = 0 implies α(k) = 0 for all k ≥ j. An α-structure with domain A is a collection A = (RA , RA , . . .), where A is a set and each RA ⊆ Aα(j) is a relation 2 1 j of arity α(j) on A. We adopt the convention A0 := ∅ so that RA = ∅ whenever j α( j) = 0. We write dom A := A for the domain of A and Finα for the P set of all αA structures A with finite dom A ⊂ Z = {. . . , −1, 0, 1, . . .} such that ∞ j=1 |R j | < ∞, where |S| denotes the cardinality of a set S. Writing dom A ⊂ f Z to denote that dom A is a finite subset of Z, we have Finα := {A = (RA , RA 2 , . . .) 1 : dom A ⊂ f Z and ∞ X j=1 |RA | < ∞}. j RELATIONAL EXCHANGEABILITY 5 A j≥1 |R j | < ∞ implies that to every A ∈ Finα there is an rA := max{k ≥ 1 : RA , ∅} < ∞, making it convenient to sometimes express A ∈ k A Finα as (R1 , . . . , RA rA ) by omitting the infinite sequence of empty relations Condition P after rA . For any A∗ ⊇ A and any injection ρ : A∗ → A0 there is an induced action ρ(A) ρ(A) on A = (RA , RA , . . .) by A 7→ ρ(A) = (R1 , R2 , . . .) with dom ρ(A) = 2 1 ρ(A) := {ρ(a) : a ∈ A} and (4) ρ(A) (ρ(a1 ), . . . , ρ(aα(j) )) ∈ R j if and only if (a1 , . . . , aα(j) ) ∈ RA j for each j = 1, 2, . . .. For n ≥ 1 and any at most countable subset of finite α-structures R ⊆ Finα for which each A ∈ R has dom A ⊆ Z + := {1, 2, . . .}, we consider the set  of equivalence classes of R-valued sequences x : [n] → R obtained R[n] as follows. Any x : [n] → R determines a relationally labeled structure by removing the labels of the elements contained in each of the x(i) while maintaining the integrity of the overall structure induced by x. Formally, we define the relationally labeled structure induced by x as the equivalence class (5) x  := {x 0 : [n] → R : ρ x 0 = x for some bijection ρ : Z → Z}, where ρ x : [n] → R is defined by (ρ x)(i) = ρ(x(i)) as in (4) above. We write  as the set of all x constructed from some x : [n] → R in this way. R[n]  Remark 2.1 (Notation). Below we reserve lowercase letters x, y, . . . for generic (non-random) objects and uppercase letters X, Y, . . . for random objects. We use bold letters x, y, . . . for generic (non-random) functions x : [n] → R and X, Y, . . . for random functions X : [n] → R. Example 2.2. In this example, we demonstrate how the structures in Section 1 fit into the framework of α-structures. For this we write A = (RA , RA , . . .) as an 2 1 α-structure with dom A = A and α varying according to the context. For α(1) = 1 and α(k) = 0 for k ≥ 2, each α-structure A is determined by a subset RA ⊆ A. If we then choose R ⊆ Finα to consist of all singleton subsets of N, 1 so that each RA has the form {i}, i ≥ 1, then the equivalence class of x : [n] → R in 1 (5) corresponds to the equivalence relation induced by x as in (1). For α(1) = 2, α(k) = 0 for k ≥ 2, each α-structure A with dom A = A corresponds to a binary relation RA ⊆ A2 . Taking R ⊆ Finα to consist of all 1 A structures with R1 = {(i, j)} for i , j corresponds to sampling from the phone call database in Section 1: each sampled Xk = {(i, j)} corresponds to a caller-receiver pair (i, j). The equivalence class x  gives an edge-labeled (directed) graph as in Figure 2. (To get an undirected graph we would take each RA to consist of symmetric pairs 1 {(i, j), ( j, i)}.) Taking α( j) = j for all j ≥ 1 means that any α-structure is determined by a collection RA ⊆ A j of subsets of j-tuples for every j ≥ 1. If we take R ⊂ Finα j 6 HARRY CRANE AND WALTER DEMPSEY to consist only of those α-structures of the form RA = {(aσ(1) , aσ(2) , . . . , aσ(k) ) : k A permutations σ : [k] → [k]} for some k ≥ 1 and R j = ∅ for all j , k, then the elements A ∈ R can be used to represent the set of coauthors in a sample of articles. Also for α( j) = j for all j ≥ 1, if we take R ⊂ Finα to consist only of those α-structures of the form RA = {(a1 , a2 , . . . , ak )} for some k ≥ 1 and RA = ∅ for all j k j , k, then each A ∈ R corresponds to a path from a1 to ak , as discussed at the end of Section 1. Our discussion below specializes to the case of random structures labeled by the countable set N, which are those structures constructed just as in (5) but for a countable sequence x : N → R. (Note well the difference between the index set N and the domain of the structures in R, which we take to be Z + . The domain labeling is “quotiented out” in (5), while the indexing N remains, serving as the “edge labels” in the associated interpretation as an edge-labeled graph in Figure 2.)  to denote the set of such structures and we equip R with We write RN N the Borel σ-field associated to its product-discrete topology induced by the  and n ≥ 1, we define the restriction R x following metric. For any x  ∈ RN n   as the structure obtained by taking any x 0 ∈ x (i.e., x 0 : N → R of x  to R[n]  for which there exists ρ such that ρ x 0 = x), restricting its domain to [n] by x 0|[n] : [n] → R, i 7→ x 0 (i), and putting Rn x  := (x 0|[n] ) as in (5). By the  , it is clear that this is well-defined and does not depend on definition of RN  by the specific choice of x 0 ∈ x  . We then define the metric on RN d(x  , x 0 ) := 1/(1 + sup{n ≥ 1 : Rn x  = Rn x 0 }),  x  , x 0 ∈ RN ,  is compact. with the convention that 1/∞ = 0, under which RN  by For any permutation σ : N → N, we define the relabeling of x  ∈ RN σ 0 σ as the structure x  obtained by first choosing any x ∈ x  and putting x σ := (x 0 ◦σ) , where x 0 ◦σ : N → R is defined by usual composition of functions, (x 0 ◦σ)(i) := x 0 (σ(i)). It is once again clear that this does not depend on the specific choice of representative x 0 ∈ x  since the actions of ρ : N → N and σ : N → N commute for all x 0 , i.e., ρ(x 0 ◦σ) = (ρ x 0 ) ◦ σ.  is Definition 2.3 (Relational exchangeability). A random structure X  ∈ RN relationally exchangeable if X σ =D X  for all permutations σ : N → N. 2.2. Representation theorem. Our main theorem establishes a generic con . The struction for infinite relationally exchangeable structures X  in RN construction proceeds by sampling a sequence X1 , X2 , . . . conditionally i.i.d. in a related space R? and then modifying these observations to obtain a new sequence X † := (X1† , X2† , . . .). We then construct the equivalence class X † based on the sequence X1† , X2† , . . . as in (5). Since stating the theorem formally requires some new ideas and notation, we give the construction and key ideas of the proof prior to stating the result in Theorem 2.6. RELATIONAL EXCHANGEABILITY 7 As above, let α : N → N ∪{0} and R ⊆ Finα be an at most countable set of α-structures such that α(j) = 0 implies α(k) = 0 for all k ≥ j and each A ∈ R has dom A ⊂ f Z + . Since R is at most countable, we may fix an ordering R = {Rn }n≥1 of its elements. Given any A = (RA , RA , . . .) ∈ Finα and any 2 1 [0, 1]-valued sequence (Ti )i∈Z , we define TA :≡ T ◦ A as the α-structure TA with dom(TA) := {Ti : i ∈ dom A} and relations RTA given by j (6) (Ta1 , . . . , Taα( j) ) ∈ RTA j if and only if (a1 , . . . , aα(j) ) ∈ RA j for each j = 1, 2, . . .. We then define [ R[0,1] := {TRn : T = (Ti )i∈Z ∈ [0, 1]Z } n≥1 as the set of α-structures obtained by associating [0, 1]-valued labels to the elements of R. More generally, if (Ξi )i∈dom A is a collection of subsets Ξi ⊆ [0, 1], then we define ΞA by ΞA := {TA : Ti ∈ Ξi for each i ∈ dom A}. We equip R[0,1] with the σ-field on R[0,1] generated by all sets of the form ΞA with A ∈ R and Ξ = (Ξi )i∈dom A a collection of Borel subsets of [0, 1].  and any sequence ξ = (ξ ) From any x  ∈ RN i i∈Z of i.i.d. Uniform[0, 1] random variables, we write ξ x  to denote a random R[0,1] -valued sequence (Zi )i≥1 obtained by first taking any representative x 0 ∈ x  and then putting Zi = ξ x 0 (i) for each i ≥ 1, with ξ x 0 (i) defined as in (6). Since x  is fixed in this example, each ξ x 0 (i) corresponds to an assignment of Uniform[0, 1] random labels to the elements in the domain of x 0 (i). Since ξi , ξ j with probability 1 for all i , j, it is immediate that the distribution of ξ x  does not depend on the manner in which the representative x 0 is chosen. Now for  and a sequence ξ of any infinite relationally exchangeable structure X  ∈ RN Uniform[0, 1] random variables independent of X  , ξ X  = (Zi )i≥1 defines a random sequence obtained by putting Zi = ξ x 0 (i) for a representative x 0 ∈ x  on the event X  = x  . By exchangeability of X  , the sequence Z = (Zi )i≥1 obtained in this way is also exchangeable and de Finetti’s theorem implies that Z is distributed as an i.i.d. sequence from a random measure ν on R[0,1] , as described in (3). Given ν, we define the propensity of u ∈ [0, 1] in Z by (7) ν? (u) := ν({A ∈ R[0,1] : u ∈ dom A}), which equals the conditional probability of the event {u ∈ dom Z j } given ν, for each j ≥ 1. It is clear that each u ∈ [0, 1] appears in either 0, 1, or infinitely many of the relations (Zi )i≥1 with probability 1. First, if ν? (u) > 0, then the strong law of large numbers implies that u occurs in infinitely many of the relations Zi with proportion ν? (u). If ν? (u) = 0 and u ∈ Zi for some i, then the probability that u occurs in Z j is ν? (u) = 0 independently for each j , i and therefore u appears only once in Z with probability 1. 8 HARRY CRANE AND WALTER DEMPSEY Clearly the set U = {u : ν? (u) > 0} is at most countable since | dom(Zi )| < ∞ for each i = 1, 2, . . .. We can, therefore, order the elements of U as u1 , u2 , . . . such that ν? (u j ) ≥ ν? (u j+1 ) for j ≥ 1, breaking ties ν? (v) = ν? (w) as follows. For each A ∈ R, we define ν? (u; A) := ν({TA : T = (Ti )i∈Z ∈ [0, 1]Z such that u ∈ dom(TA)}), u ∈ [0, 1], to be the measure assigned to the subset of R[0,1] whose structure is consistent with A and which contains u in its domain. Assuming ν? (v) = ν? (w), we assign the smaller label to v if there is some k ≥ 1 such that ν? (v; R j ) = ν? (w; R j ) for all j < k and ν? (v; Rk ) > ν? (w; Rk ). If ν? (v; R j ) = ν? (w; R j ) for all j ≥ 1, then we label v and w in increasing order, i.e., if we are to assign labels j and j + 1 to v and w and if v < w, then we shall put u j = v and u j+1 = w; otherwise, we put u j = w and u j+1 = v. The ordering U = (u j ) j≥1 is thus uniquely determined by ν and the fixed ordering of R chosen at the outset. Now given U, for any A ∈ R[0,1] we define A? (suppressing the dependence on ν) by replacing each occurrence of u j ∈ U by j and replacing each occurrence of v0 < U in A by a unique non-positive integer z(v0 ) = 0, −1, −2, . . . so that for v0 , v00 < U and both in dom A with v0 ≤ v00 we have z(v0 ) ≤ z(v00 ) and the z(v0 ) are chosen to be the largest possible non-positive integers that satisfy this condition. (For example, if v1 < v2 < v3 are the only elements in dom A that are not in U, we assign z(v3 ) = 0, z(v2 ) = −1, and z(v1 ) = −2.) Every A ∈ R[0,1] thus corresponds to a unique such A? and we define R? as the set of all structures A? obtained in this way. Note that since R is at most countable, so is R? . Also, although we have constructed R? using U (and therefore ν), the set R? depends only on R, justifying the term R-simplex in our definition of FR in (8) below. The elements of R? are thus α-structures A? with dom A? ⊆ Z. We define the R-simplex FR by     X     ? : fB ≥ 0 and (8) FR :=  ( f ) f = 1 ,  B B B∈R     ? B∈R on which we equip the metric dFR ( f, f 0 ) := X | fB − fB0 |, f, f 0 ∈ FR , B∈R? and the associated Borel σ-field.  by first Any f ∈ FR determines a unique probability measure ε f on RN drawing X = (X1 , X2 , . . .) i.i.d. from (9) P(Xi = B | f ) = fB , B ∈ R? , and then constructing X † = (X1† , X2† , . . .) from X as follows. We initialize by putting m0 = 0. For each n ≥ 1, given mn−1 , we replace the non-positive elements in dom Xn according to the rule: (i) If dom Xn has no non-positive elements, then put mn = mn−1 and Xn† := Xn . RELATIONAL EXCHANGEABILITY 9 (ii) If dom Xn has non-positive elements 0, . . . , −k for k ≥ 0, then define Xn† by replacing each occurrence of −i, for 0 ≤ i ≤ k, in Xn by mn−1 −i, then putting mn = mn−1 −k−1 and keeping positive elements unchanged. See Example 2.5 below for an illustration. Note the distinction between the use of non-positive elements in construction of Xi and Xi† respectively. The non-positive elements of each Xi ∈ R? serve to denote the non-recurring particles that appear only within this particular relation. The non-positive elements of Xi† , on the other hand, serve to denote the non-recurring particles across all relations in the sequence X † . We define ε f to be the distribution of X † constructed by applying (5) to the sequence X†. From ν, we define f ν = ( fBν )B∈R? ∈ FR by (10) fBν := ν({A ∈ R[0,1] : A? = B}), B ∈ R? . This choice of f ν = ( fBν )B∈R? is uniquely determined by ν and the fixed ordering of R. Conversely, given f = ( fB )B∈R? ∈ FR we construct a measure ν f on R[0,1] to be the distribution of ξY from (6) for Y drawn from distribution (9) and ξ = (ξi )i∈Z i.i.d. Uniform[0, 1] independent of Y. Proposition 2.4 proves that the above procedure does not alter the random, relationally labeled structure. Proposition 2.4. Let X  be relationally exchangeable and let θ = (θi )i∈Z be i.i.d. Uniform[0, 1] independent of X  . Then (((θ X  )? )† ) = X  a.s., where (θ X  )? denotes the application of ? : R[0,1] → R? to each component of the sequence θ X  . Proof. Let θ = (θi )i∈Z be i.i.d. Uniform[0, 1] independently of X  . Each event X  = x  gives rise to a probability measure ν on R[0,1] through the de Finetti measure of θ x  ; see (3). Let U = (ui )i≥1 be the ordered subset of [0, 1] corresponding to the atoms of ν? as in (7). Since θ are i.i.d. Uniform[0, 1], we have P{θi , θ j for all i , j} = 1, which implies that distinct i, j ∈ Z are labeled distinctly in θ x  with probability 1. In particular, the non-positive labels in (θ x  )? account for all those elements that appear in only one entry of the sequence θ x  . It follows that ((θ x  )? )† ∈ x  with probability 1 and, thus, (((θ x  )? )† ) = x  with probability 1 for all possible outcomes X  = x  .  Example 2.5. To illustrate the above procedure, let R consist of the singleton sets ({i}), i ≥ 1, just as in the first part of Example 2.2, written as {1}, {2}, . . . for simplicity. Let U = (ui )i≥1 be a countable subset of [0, 1]. From R = {{1}, {2}, . . .}, we obtain R? = {{0}, {1}, . . .} since for any sequence T = (Ti )i∈Z and any {i}, i ≥ 1, the transformed relation T{i} ≡ {Ti } has either Ti ∈ U or Ti < U. If Ti = uk ∈ U, then {Ti }? = {k} for k ≥ 1; and if Ti = v < U, then {Ti }? = ({0}). Given ( f{0} , f{1} , . . .) ∈ FR , we generate a sequence X1 , X2 , . . . of singleton sets i.i.d. as in (9), from which we obtain X1† , X2† , . . . by reassigning occurrences of 0 10 HARRY CRANE AND WALTER DEMPSEY to the greatest negative integer that has not yet appeared in the sequence. For example, if the sequence begins {4}, {0}, {2}, {0}, {2}, {2}, {0}, . . ., then we reassign labels to obtain {4}, {0}, {2}, {−1}, {2}, {2}, {−2}, . . .. Delabeling according to (5) (or equivalently (1)) yields the equivalence classes {1}, {2}, {3, 5, 6}, {4}, {7}. For another example, let R consist of ordered pairs {(i, j)}, 1 ≤ i , j < ∞, so that we are in the phone call example of Section 1. Let U = (ui )i≥1 be a countable subset of [0, 1]. Then R = {{(i, j)} : i , j ≥ 1} and R? = {{(i, 0)} : i ≥ 1} ∪ {{(0, i)} : i ≥ 1} ∪ {{(0, −1)}, {(−1, 0)}} ∪ R since any T = (Ti )i∈Z and any {(i, j)} ∈ R is transformed by T{(i, j)} = {(Ti , T j )}, which may have Ti = ui0 ∈ U and T j = u j0 ∈ U in which case {(Ti , T j )}? = {(i0 , j0 )}. If Ti = ui0 ∈ U and T j < U, then {(Ti , T j )}? = {(i0 , 0)}. If Ti < U and T j = u j0 then {(Ti , T j )}? = {(0, j0 )}. If Ti , T j < U then {(Ti , T j )}? will equal {(0, −1)} or {(−1, 0)} depending on whether T j < Ti or Ti < T j , respectively. The above construction gives the following representation theorem. Theorem 2.6. Let α be a signature, R ⊆ Finα , and fix an ordering R = (Rn )n≥1 .  . Then there exists a Let X  be a relationally exchangeable random structure in RN probability measure φ on FR such that X  ∼ εφ , where Z (11) εφ (·) = ε f (·)φ(d f ). FR A canonical version of the measure φ in Theorem 2.6 can be constructed as in (10) above, and this is the measure we construct in the following proof. Proof. We proceed by constructing φ as in (10) and showing that Y † ∼ εφ satisfies Y † =D X  . To see this, we first let ψ be the de Finetti measure on P(R[0,1] ) (i.e., the space of probability measures on R[0,1] ) associated to ξ X  for ξ = (ξi )i∈Z i.i.d. Uniform[0, 1] independently of X  . In particular, the distribution of ξ X  is conditionally i.i.d. from ν ∼ ψ. The measure ψ determines a measure φ on FR through (10). Given f ∼ φ, we construct Y = (Y1 , Y2 , . . .) as conditionally i.i.d. from the distribution in (9) and Y † as in (i) and (ii) above. Let U = (ui )i≥1 be the atoms of θ Y † for θ = (θi )i∈Z i.i.d. Uniform[0, 1] independent of Y † , let ζ = (ζk )k≤0 be i.i.d. Uniform[0, 1] independently of everything else, and define ξ f = (ξi )i∈Z by ( ui , i ≥ 1, ξi = ζi , otherwise. First note that the conditional distribution of ξ f Y † given f is the same as the conditional law of θ Y † given ν f , since the ξi for i ≥ 1 were constructed from the atoms of θ Y † . Writing D(X | Y) to denote the conditional distribution RELATIONAL EXCHANGEABILITY 11 of X given Y, we thus have (12) D(ξ f Y † | f ) = D(θ Y † | ν f ) (13) D(ξ f Y † | f ) = D(θ X  | ν f ), and where (13) follows from the construction of φ from the de Finetti measure ψ of ξ X  . From (12) and (13) it follows that D((((ξ f Y † )? )† ) | f ) = D((((θ Y † )? )† ) | ν f ) D((((ξ f Y † )? )† ) and ? † | f ) = D((((θ X  ) ) ) | ν f ), and thus D((((θ Y † )? )† ) | ν f ) = D((((θ X  )? )† ) | ν f ). By Proposition 2.4, we have (((θ Y † )? )† ) = Y † a.s. and (((θ X  )? )† ) = X  , implying Y † =D X  as desired.  2.3. Special cases. As discussed in Example 2.5, the case in which R corresponds to the singleton sets {i} for i ≥ 1 gives FR equal to the ranked simplex     X     ↓ ∆ :=  ( f , f , . . .) : f ≥ f ≥ · · · ≥ 0, f ≥ 0, f = 1 .  0 1 2 0 1 i     i≥0 To see this note that R[0,1] is the set of singleton elements of [0, 1], i.e., R[0,1] := {{u} : u ∈ [0, 1]}. An exchangeable sequence X = (X1 , X2 , . . .) in R[0,1] gives rise to a random countable subset U ⊂ [0, 1] of elements that appear infinitely often among the Xi and its complement [0, 1] \ U consisting of all elements appearing at most once among the Xi . In the construction of X † from X described above, any occurrence Xi = {u} for u < U gives rise to Xi? = {0}, explaining why the definition of FR in (8) corresponds to the simplex of elements ( fi )i≥0 . Following Example 2.2, the case of edge exchangeable random (directed) graphs corresponds to R ⊂ Finα with α(1) = 2 and α(k) = 0 for k ≥ 2, with each A ∈ R having RA = {(i, j)}, i , j. In this case, R[0,1] corresponds to all 1 pairs (u, v) for u, v ∈ [0, 1]. An exchangeable sequence X in R[0,1] determines a random subset U ⊂ [0, 1]. Now, in each Xi there can be 0, 1, or 2 elements v < U. Occurrences of such elements are replaced by 0 and −1, as explained in Example 2.5. In general, we call the occurrences of any u < U in the sequence X blips. These elements are merely a ‘blip’ in the overall sequence X—they occur only for an instant and then never again. Our labeling convention in defining R? is that the non-positive labels correspond to the blips. Theorem 2.6 describes how blips arise in general relationally exchangeable structures. 12 HARRY CRANE AND WALTER DEMPSEY 3. Concluding Recent work in the statistical analysis of network data underscores the significance of relationally labeled structures in applications, as many data structures which are typically represented graphically, such as social networks and networks detailing email correspondence and professional collaborations, arise from a process by which interactions or relations accumulate within a population of otherwise indistinguishable individuals. The work in [1] focused on the case of edge and hyperedge exchangeable random graphs, but the additional examples in Section 1 involving repeated path sampling and snowball sampling are also highly relevant in networks applications. The representation theorem serves two immediate statistical purposes. First, the representation characterizes a general class of nonparametric statistical models of potential interest in the aforementioned applications. Second, it establishes that vertices arrive in size-biased random order in relationally exchangeable structures, explaining why the common assumption of exchangeable vertex labeling, as presented in graphon models [5], is not tenable in many applications. We reserve discussion of these practical implications for other work; see [1, 2]. References [1] H. Crane and W. Dempsey. Edge exchangeable models for network data. arXiv:1603.04571, 2015. [2] H. Crane and W. Dempsey. A framework for statistical network modeling. arXiv:1509.08185, 2015. [3] B. de Finetti. La prévision: ses lois logiques, ses sources subjectives. Annales de l’Institut Henri Poincaré, 7:1–68. [4] J. F. C. Kingman. Random partitions in population genetics. Proc. Roy. Soc. London Ser. A, 361(1704):1–20, 1978. [5] L. Lovász and B. Szegedy. Limits of dense graph sequences. J. Comb. Th. B, 96:933–957, 2006. Department of Statistics & Biostatistics, Rutgers University, 110 Frelinghuysen Avenue, Piscataway, NJ 08854, USA E-mail address: hcrane@stat.rutgers.edu URL: http://stat.rutgers.edu/home/hcrane Department of Statistics, University of Michigan, 1085 S. University Ave, Ann Arbor, MI 48109, USA E-mail address: wdem@umich.edu 

10math.ST

1 Joint Investment and Operation of Microgrid arXiv:1511.01984v1 [cs.SY] 6 Nov 2015 Hao Wang, Student Member, IEEE, and Jianwei Huang, Senior Member, IEEE Abstract—In this paper, we propose a theoretical framework for the joint optimization of investment and operation of a microgrid, taking the impact of energy storage, renewable energy integration, and demand response into consideration. We first study the renewable energy generations in Hong kong, and identify the potential benefit of mixed deployment of solar and wind energy generations. Then we model the joint investment and operation as a two-period stochastic programming program. In period-1, the microgrid operator makes the optimal investment decisions on the capacities of solar power generation, wind power generation, and energy storage. In period-2, the operator coordinates the power supply and demand in the microgrid to minimize the operating cost. We design a decentralized algorithm for computing the optimal pricing and power consumption in period-2, based on which we solve the optimal investment problem in period-1. We also study the impact of prediction error of renewable energy generation on the portfolio investment using robust optimization framework. Using realistic meteorological data obtained from the Hong Kong observatory, we numerically characterize the optimal portfolio investment decisions, optimal day-ahead pricing and power scheduling, and demonstrate the advantage of using mixed renewable energy and demand response in terms of reducing investment cost. Index Terms—Smart grid, microgrid, renewable energy, solar, wind, energy storage, demand response, investment, pricing. N OMENCLATURE Acronyms P1 P2 Pui EP1 RP1 RP2 Sets N H T Ω Period-1 problem for investment Period-2 problem for operation User i’s cost minimizing problem Equivalent problem of P1 Period-1 problem for robust optimization Period-2 problem for robust optimization Set of users Investment horizon Operational horizon Set of scenarios Indices i Index of users t Index of hours ω Index of scenarios This work is supported by a grant from the Research Grants Council of the Hong Kong Special Administrative Region, China, under Theme-based Research Scheme through Project No. T23-407/13-N. Part of the results have appeared in IEEE SmartGridComm 2014 [1]. H. Wang and J. Huang (corresponding author) are with the Network Communications and Economics Lab (NCEL), Department of Information Engineering, The Chinese University of Hong Kong, Shatin, Hong Kong SAR, China, e-mails: {haowang, jwhuang}@ie.cuhk.edu.hk. Parameters N T D cs cw ce B βo βi bt Di dti t di yit ω,t rmax ηsω,t ω,t ηw max rc rdmax ηc ηd SOCmin SOCmax DODmax eω,t s,min eω,t s,max eω,t w,min eω,t w,max Variables αs αw αe xω,t i rω,t q ω,t SOC ω,t rcω,t rdω,t Qω,t pω,t eω,t s eω,t w η̂sω,t ω,t η̂w Number of electricity users Number of hours in the operational horizon Number of days in the investment horizon Investment cost of solar power Investment cost of wind power Investment cost of energy storage Investment budget Coefficient of the operator’s cost Coefficient of user i’s discomfort cost Aggregate inelastic load in time slot t Total elastic load of user i Minimum power load of user i in time slot t Maximum power load of user i in time slot t Preferred power load of user i in time slot t Maximum renewable power in t and ω Solar power supply per unit capacity in t and ω Wind power supply per unit capacity in t and ω Maximum charging amount per unit capacity Maximum discharging amount per unit capacity Conversion efficiency of charging Conversion efficiency of discharging Minimum state-of-charge Maximum state-of-charge Maximum depth-of-discharge Minimum solar power prediction error in t and ω Maximum solar power prediction error in t and ω Minimum wind power prediction error in t and ω Maximum wind power prediction error in t and ω Generation capacity of solar power Generation capacity of wind power Capacity of energy storage User i’s elastic load in t and ω Renewable power supply in t and ω Grid power procurement in t and ω State-of-charge of battery in t and ω Charging amount in t and ω Discharging amount in t and ω Aggregate power supply in t and ω Day-ahead price in t and ω Solar power prediction error in t and ω Wind power prediction error in t and ω Actual solar power generation in t and ω Actual wind power generation in t and ω I. I NTRODUCTION Aiming at reducing greenhouse gas emissions and enhancing the power grid reliability, many countries are building 2 new power infrastructures known as the smart grid [2]. The major features of the smart grid include more distributed power generations (especially from renewable energy sources), smart charging/discharging of energy storage, two-way communications between the utility company and consumers for a better demand side management, and decentralized operations of power grid in the form of microgrids [2]. It’s essential to understand the impact of these new features, and how to make optimal economic and technology decisions on the planning and operation of the smart grid. Recently, there are many studies on power grid planning (e.g. [3], [4]), integration of renewable energy and energy storage (e.g. [5]–[8]), and demand response (e.g. [10]–[12]). However, the existing literature did not consider these important issues in a holistic fashion. For example, in [3]–[8], microgrid planning is studied without considering flexible load and microgrid operation. While in [9]–[12], only microgrid operation is studied under given microgrid facilities. However, all those new features including renewable energy, storage, and demand response affect the optimal planning and operation of the microgrid, and have to be taken into account at various different time scales. In this paper, we will jointly consider the optimal investment and operation of renewable generation, energy storage, and demand response optimization in the smart grid. In particular, this paper will focus on the mixed investment in renewable energy (both solar energy and wind energy) and energy storage. The optimal mix of solar and wind energy investment will depend on the stochastic nature of these two sources, which is highly location dependent. Hence we will rely on the meteorological data in Hong Kong to validate the practical relevance of our study. Energy storage provides flexibility in terms of coordinating supply and demand in the microgrid. Through smart charging and discharging of the energy storage, the microgrid operator is able to better utilize the renewable energy generation and reduce dependency on the main grid. The key question we want to answer is the following: What is the optimal investment portfolio? In this paper, we develop a theoretical framework that captures the economic impact of renewable energy, storage, and demand response in the smart grid, and derive the optimal investment strategy and optimal demand response scheme based on realistic data. The main contributions of this paper are as follows. • Correlation and scenarios of renewable energy: Based on the meteorological data acquired from the Hong Kong Observatory, we study the correlation between solar power and wind power at certain locations of Hong Kong, and suggest mixed renewable energy investment. • Framework development: We develop a theoretical framework that enables us to derive the optimal investment of mixed enewable generation and energy storage, and the optimal operation in a demand-responsive microgrid. The problem is challenging due to the coupling of decisions of investment and operation at different time scales. • Modeling and solution methods: We formulate the joint investment and operation problem as a two-period stochastic program. We design a distributed algorithm to attain the optimal power scheduling in period-2, and derive a single-level optimization formulation to solve the optimal investment portfolio in period-1. • Impact of uncertainty in renewable energy: We analyze the impact of the prediction error of renewable energy generation by the worst-case scenario analysis. • Case studies in Hong Kong: Numerical studies based on realistic meteorological data illustrate the optimal portfolio investment decisions, and demonstrate the benefits of mixed renewable investment and demand response in terms of saving investment. The remainder of this paper is organized as follows. We review the related work in Section II. Then we analyze the renewable power generation of Hong Kong in Section III, and formulate the system model as a two-period stochastic optimization problem in Section IV. We present the detailed models for period-1 and period-2 in Section V and VI, respectively. We propose the solution method in Section VII. In Section VIII, we analyze the impact of the prediction error of renewable energy generations on the energy portfolio investment. Numerical results are presented in Section IX. This paper is concluded in Section X. II. R ELATED WORK There are several related recent studies on power grid planning, integration of renewable energy, and demand response. Specifically, studies in [3] and [4] examined investment strategies on renewable energy generation through empirical (or numerical) approaches, without considering the power scheduling operation. Studies in [5] and [6] formulated cost minimization problems to determine the optimal investment of solar-storage system and wind-storage system, respectively. Wang et al. in [7] considered the optimal planning problem for mixed solar-wind energy in microgrids using robust optimization. Yang and Nehorai in [8] formulated a cost optimization problem to decide the optimal capacities for renewable energy generation and energy storage, and solved the problem in a distributed fashion. However, none of these studies took the proactive operation (such as demand response) into consideration. Optimal demand response for residential consumers and energy storage have been studied to derive the proper incentive schemes through either game theoretic models [9], [10] or optimization models [11], [12]. The key idea of these studies is to design incentive mechanisms such that costaware users schedule their elastic demands as responses to price changes. The existing literature focused on either renewable energy and energy storage investment at a large time scale (years), or power scheduling and demand response optimization under given energy capacity at a small time scale (hours and days). However, the decisions at these two different time scales are actually tightly coupled. The portfolio investment determines the time-varying power supply availability and power scheduling flexibility, and thus affects operator’s power dispatch and users’ demand response at a smaller time scale. Meanwhile, demand response can try to match the demand with the time varying renewable energy supply, and hence can maximize 3 TMT New Territories TPK SKG SHA SKG Kowloon TC KP Lantau Island Hong Kong Island WGL Meteorological station measuring solar radiation and wind speed Meteorological station measuring wind speed Fig. 1: Locations of meteorological stations the benefit of renewable energy and even reduce unnecessary investment expenditure at a large time scale. Therefore, the investment and operation at various time scales should be jointly optimized. Recently, joint optimization of investment and operation has been considered in [13]–[15] to study the wind power investment and network expansion at a transmission-level. Different from [13]–[15], our work aims to study the energy portfolio investment in microgrids at a distribution-level. Specifically, we consider a holistic configuration of the microgrid to incorporate the latest technologies of smart grid, which includes not only renewable energy but also energy storage and demand response. Moreover, we construct an investment portfolio that consists of different technologies (e.g., solar power, wind power, and storage), and provide a systematic framework to jointly determine the optimal portfolio investment strategy and optimal pricing scheme for demand response. III. S OLAR POWER AND WIND POWER IN H ONG KONG Aiming at studying the renewable power patterns in Hong Kong, we acquire meteorological data from the Hong Kong Observatory. The data include the hourly solar radiation in King’s Park (KP) of Hong Kong, and hourly wind speed at seven different locations (KP, TMT, TPK, SHA, SKG, TC, WGL) of Hong Kong, as shown in Fig. 1. Based on the data, we analyze the correlations of solar power and wind power generations across different locations of Hong Kong, which motivate us to study the mixed renewable energy investment in microgrids. A. Correlation between solar power and wind power Both solar power and wind power are intermittent power sources, and their stochastic features could be highly locational-dependent. Since Hong Kong is relatively small, we assume that the solar radiation is the same across Hong Kong and is represented by the data measured in KP. The wind power, however, has clearly different patterns at different locations. In this paper, we will focus on one year of meteorological data (from Sep. 1 2012 to Aug. 31 2013) to study the solar power and wind power productions. Based on the solar power model [16] and wind power model [17], we calculate the hourly solar and wind power productions in 365 days based on the measurement of data of solar radiation and wind speed.1 To study the statistical correlation between the hourly solar and wind power productions over one year, we calculate the sample correlation coefficient [18] as

P Y (k) − Ȳ k X(k) − X̄ ρX,Y = qP
2 qP
2 , X(k) − X̄ Y (k) − Ȳ k k where X and Y are data series with k = 1, ..., K terms, X̄ and Ȳ are the mean values of X and Y , respectively, and ρX,Y measures the correlation coefficient between X and Y . We substitute the one-year hourly solar power production into X, and the one-year hourly wind power production of each location into Y , and calculate their correlations. We find that the wind powers in four locations (KP, TPK, SHA, SKG) of Hong Kong have positive correlations with solar power, while the correlation is negative in two locations (TC, WGL), and the correlation is close to zero in location TMT. Motivated by the Markowitz portfolio selection theory in Finance [19], we will study the mixed investment strategy of solar power and wind power in the following two locations: TC and SKG. Specifically, TC and SKG are two representative examples for negative and positive correlations (with correlation coefficient −0.22 and 0.15) between solar and wind power generations, respectively. B. Scenario generation of solar power and wind power To study the mixed investment of solar power and wind power, we need to model the solar power and wind power generations. Usually, the renewable energy investment is made for years of operation. Therefore, we use one-year historical data to empirically model the distributions of solar and wind power generations, similar as [13], and assume that the future renewable generations in each year follow the same distribution of the one-year historical data. Each daily power production realization (solar power production in KP, wind power productions in TC and SKG) is called a scenario, and thus we obtain 365 scenarios for solar power and wind power respectively. As a large number of scenarios will reduce the computational tractability of the investment optimization problem, it is useful to choose a smaller subset of scenarios that can well approximate the original entire scenario set. Such technique has been widely used in economics and engineering research [20], [21] for the purpose of modeling stochastic processes. We applied the forward scenario reduction algorithm [21] to find a best scenario subset, and assign new probabilities to the smaller number scenarios. The key idea is to select a subset of scenarios to preserve, such that the corresponding reduced probability measure is the closest to the original measure. We set the number of preserved scenarios as 10,2 and generate 1 The technical parameters of the solar power model and wind power model are shown in the online technical report [31]. 2 The persevered number of scenarios depends on the tradeoff between performance and computational complexity in practice. 1 0.5 0.8 0.8 0.4 0.6 0.4 0.2 0 1 5 10 15 20 24 Wind Power (kW) 1 Wind Power (kW) Solar Power (kW) 4 0.6 0.4 0.2 0 1 5 Time (Hour) 10 15 20 0.3 0.2 0.1 0 1 24 5 Time (Hour) Fig. 2: Solar power scenarios in KP (per 1kW capacity) Fig. 3: Wind power scenarios in TC (per 1kW capacity) selected scenarios for the solar power production in KP (which we assume is the same as in TC and SKG since Hong Kong is relatively small geographically) and wind power productions in TC and SKG, shown in Fig. 2, 3, and 4. Therefore, we have a set of 10 scenarios denoted as Ω for the renewable energy generation, and each renewable generation scenario ω ∈ Ω consists of solar power and wind power productions in TC and SKG.3 Comparing Fig. 3 and 4 with Fig. 2, we can see that the solar power has a peak at noontime, while wind power productions show dramatic locational differences. Wind power in TC is often adequate during night time, while wind power in SKG reaches a higher output level during day time. Therefore, solar power and wind power have high locational dependence, which motivates us to study the optimal mixed investment of both cases. The data we use can be found at [32]. IV. S YSTEM OVERVIEW In this section, we present the system model for the joint investment and operation problem. Fig. 5 illustrates a typical microgrid, which connects to the main power grid, and consists of different local power supplies and responsive demand. Solar Power Wind Power Microgrid Energy Storage 10 15 20 24 Time (Hour) Fig. 4: Wind power scenarios in SKG (per 1kW capacity) energy storage device is also implemented in the microgrid to charge and discharge power properly to reduce the system operating cost. The demand side consists of a set of electricity users N = {1, ..., N }, and each user i ∈ N is equipped with a smart meter and energy scheduling module. The operator runs the microgrid, determines the investment in renewable energy and energy storage capacities at a large time scale (years), as well as energy prices at a small time scale (hours in one day). The users determine their energy consumptions based on the prices set by the operator (as the demand response). From the operator’s perspective, it needs to decide the optimal capacity investment and power scheduling. Fig. 6 depicts the investment and operation horizons. An investment horizon usually corresponds to several years. The operation horizon is one day, which includes T = {1, ..., T } of T time slots (say 24 hours). To model both investment and operation, we propose a two-period stochastic program that jointly optimizes capacity investment and power scheduling in the microgrid. Specifically, the period-1 problem is a longterm capacity investment problem, with the objective of minimizing the expected overall cost over an investment horizon H = {1, ..., D} of D days, subject to a budget constraint. The period-2 problem is a power scheduling problem, with the objective of minimizing the operating costs of both operator and users under a specific realization of renewable power generation within a smaller time window. Investment horizon Main Power Grid Operator 1 2 1 v s Smart Meters Demand Response Power Communications Fig. 5: System model Within the microgrid, there is a local generation system, consisting of solar and wind renewable power generation. An 3 For the detailed scenario generation and reduction, please refer to the technical report [31]. D 2 Operation horizon T Days Hours Fig. 6: Investment and operation horizons Note that the investment and operation problems are closely coupled, as the investment decision affects the operational cost, and the expected recurring operational cost affects the investment decision. Fig. 7 shows the connection between period1 investment problem and period-2 operational problem. In period-2, the optimal power scheduling is affected by both power supply and demand, and in particular the renewable power supply depends on the capacity investment decision made in period-1. On the other hand, the overall cost in period-1 includes both the one-time investment cost and the 5 Renewable generation capacity Period 1 (years): Investment price time slots to low price time slots. The corresponding discomfort (or inconvenience) [22] is X ω,t
2 Ci (xω xi − yit , (3) i ) = βi Period 2 (daily): Operation t∈T Capital Investment + Operating Cost where βi is the coefficient of discomfort cost, which indicates the sensitivity of user i towards its deviation from the preferred power consumption. Total cost minimization over investment horizon (incorporating operational decisions) Fig. 7: Two-periods optimization problem B. Operator’s model recurring operating cost of all D days in period-2. Hence the decisions in two periods are coupled. In the following, we will first formulate the operation problem in period-2, and then formulate the investment problem in period-1. V. P ERIOD -2 PROBLEM FOR MICROGRID OPERATIONS In this section, we first present the models of users and the operator, and then formulate the operator’s operating cost minimization problem. A. User’s model We classify each user’s load into two types: the elastic load and the inelastic load. The elastic load corresponds to the energy usage of those appliances such as electric vehicles, washing machines, and HVAC (heating, ventilation and air Conditioning) systems, as a user may shift the flexible load over time. The inelastic load corresponds to the energy usage of appliances such as lighting, refrigerators, and such load cannot be easily shifted over time. The demand response can only control the elastic load. We denote the corresponding ω,t ω,t decisions as xω is user i’s i = {xi , ∀t ∈ T }, where xi elastic energy consumption in time slot t ∈ T under renewable energy generation scenario ω. The elastic load scheduling for all users needs to satisfy the following two constraints: t dti ≤ xω,t ≤ di , ∀t ∈ T , i ∈ N , i X ω,t xi = Di , i ∈ N . (1) (2) t∈T Constraint (1) provides a minimum power consumption dti and t a maximum power consumption di for the user i in each time slot t.4 Constraint (2) corresponds to the total elastic energy demand Di for user i in the entire operation horizon. We further introduce a discomfort cost Ci (·), which meaω,t sures user i’s experience under xω i = {xi , ∀t ∈ T } which deviates from his preferred power consumption y i = {yit , ∀t ∈ T } under a time-independent flat price environment. We assume that the operator can control the demand to minimize the operating cost. Later in Section VII, we will discuss how such control can be realized through a properly designed pricing mechanism. If the operator sets time-varying prices, energy users will schedule their elastic loads to minimize users’ costs, i.e. shifting power consumption from high 4 Constraint (1) is general and can model the staring time and ending time of an elastic load by setting proper parameters. We assume that the operator can predict the renewable energy production scenario ω accurately at the beginning of an operation horizon (a day).5 In scenario ω and each time slot t, the operator determines the renewable power supply, conventional power procurement, and energy storage charging and discharging to meet the total users’ demand, which consists of the elastic power consumption xω i from each user i ∈ N and the aggregate inelastic load {bt , t ∈ T } of all the users. 1) Power supply: The renewable power supply r ω = ω,t {r , ∀t ∈ T } and conventional power procurement q ω = {q ω,t , ∀t ∈ T } should satisfy the following constraints: ω,t 0 ≤ rω,t ≤ rmax , ∀t ∈ T , (4) q ω,t ≥ 0, ∀t ∈ T , (5) ω,t in constraint (4) depends on the invested capacities where rmax of solar power αs and wind power αw , which are the operator’s decision variables in period-1. For each unit of invested solar capacity and wind capacity, the corresponding solar power and wind power in scenario ω and time t will be ω,t ω ω,t ηω s = {ηs , ∀t ∈ T } and η w = {ηw , ∀t ∈ T }, respectively. ω,t ω,t ω,t Hence we have rmax = ηs αs +ηw αw . Constraint (5) means that the operator can only purchase power from the main grid, but cannot sell power to the main grid, assuming that the main grid does not accommodate distributed generations in the microgrid. We assume that the main grid has adequate power to meet the demand of the microgrid, hence there is no upper-bound of q ω,t in (5). Different from the conventional power generation, the renewable power generation does not consume fuel sources, so we assume zero cost of generating renewable power [5]. Therefore, the operator will try to use as much renewable power as possible to meet the demand. 2) Energy storage: It’s well-known that energy storage (such as batteries) can smooth out the intermittent renewable power generation, match demand and supply by smart charge/discharge, and exploit time-varying energy generation costs for arbitrage. In our paper, energy storage is regarded as an energy asset in the investment portfolio, as it can complement time-varying renewable energy generation to provide relatively stable power supply, and can also work in parallel with demand response programs to help balance demand and 5 The short-run (day-ahead) renewable energy forecast can be quite accurate in practice [23], [24]. We will also consider the impact of prediction error in Section VIII. In addition, this paper focuses on the microgrid energy management instead of the transient dynamics and stability of the sysetem. Therefore, we assume that the microgrid operator is able to manage the intermittency of renewable generations, such that the stability of the microgrid system can always be guaranteed. 6 supply. We assume that the microgrid operator has decided in period-1 to install energy storage devices with a total capacity αe , and let SOC ω,t , rcω,t , and rdω,t denote the state-of-charge of the storage, charging amount, discharging amount in time slot t and scenario ω, respectively. First, the energy charging and discharging amounts are bounded, and satisfy the following constraints: 0 ≤ rcω,t ≤ αe rcmax , ∀t ∈ T , 0≤ rdω,t ≤ αe rdmax , ∀t ∈ T , (6) (7) where rcmax > 0 and rdmax > 0 denote the maximum charging and discharging amount per unit capacity of the energy storage, respectively. Hence, αe rcmax and αe rdmax indicate the maximum charging and discharging amount after the operator decides to deploy energy storage facilities with the capacity of αe . Second, there are power losses when electricity is charged into and discharged from the battery. We denote ηc ∈ [0, 1] and ηd ∈ [0, 1] as the conversion efficiencies of charging and discharging. Therefore, we obtain the energy storage dynamics of the microgrid i in time slot t as rω,t ηc rcω,t − d ∀t ∈ T , αe ηd αe ≤ SOCmax , ∀t ∈ T , SOC ω,t = SOC ω,t−1 + (8) SOCmin ≤ SOC ω,t (9) SOC ω,0 = SOC ω,T , (10) where SOC ω,t evolves with charging and discharging of the battery according to (8). It is shown in (9) that SOC ω,t is bounded between SOCmin and SOCmax , which are lower and upper bounds [25] for the level of energy storage in percentage, respectively. For example, we can set SOCmax as 100%, which means the battery can be charged to reach its full capacity. We can set SOCmin = 1 − DODmax , where DODmax is the maximum depth-of-discharge (DOD) allowed.6 Moreover, we set the terminal state-of-charge SOC ω,T at the end of each day to be equal to its initial value SOC ω,0 at the beginning of each day, such that the battery can be operated independently across days. 3) Operator’s cost: The power supply and demand should satisfy the following power balance constraint in time slot t. X ω,t rω,t + q ω,t + rdω,t = rcω,t + bt + xi , ∀t ∈ T , (11) i∈N ω ω,t We let Q = {Q i.e., , t ∈ T } denote the aggregate supply, Qω,t = rω,t + q ω,t ≥ 0, ∀t ∈ T , (12) and we can rewrite the power balance constraint (11) as follows: X ω,t Qω,t = bt + xi + rcω,t − rdω,t , ∀t ∈ T . (13) i∈N 6 DOD is defined as the ratio of maximum discharge to the battery capacity [26]. Usually, the lifetime of a battery can be measured by the number of charge-discharge cycles it can sustain at a given DoDmax . High DODmax causes fast depreciation of battery storage. Therefore, we set a low DODmax (thus a high SOCmin ) for the battery operation to reduce the impact of battery degradation, so as to make sure that the lifetime of battery is no shorter than the investment horizon. If the operator has enough renewable generation to meet ω,t the aggregate demand, i.e. rmax ≥ Qω,t , then there is no need to purchase any conventional power, i.e. q ω,t = 0. On the ω,t other hand, if rmax < Qω,t , the operator will first use all the ω,t ω,t renewable power r = rmax , and then purchase conventional ω,t ω,t ω,t power q = Q − rmax to meet the power deficit. The production cost of conventional power has a quadratic form [27], and thus we define the operator’s cost as Xh
+ i 2 ω,t Co (Qω ) = βo Qω,t − ηsω,t αs − ηw αw , (14) t∈T where (z)+ = max{z, 0} for any value z, and βo is the coefficient of the operator’s cost. C. The period-2 problem Next we state the period-2 problem, where the operator coordinates aggregate power supply Qω and schedules users’ power consumptions xω i to minimize the operating cost, which consists of the operator’s cost Co (Qω ) and all the users’ costs Ci (xω i ) as follows. P2: Operating cost minimization in period-2 X ω min C (Q ) + Ci (xω o i ) ω ω Q ,xi subject to i∈N Constraints (1), (2), (6)-(10), (13). Problem P2 is convex, and can be solved efficiently if a centralized optimization is possible. However, this may not be feasible in practice, as the operator cannot directly control users’ power consumptions {xω i , ∀i ∈ N }. We will discuss the design of a pricing scheme to derive the optimal power consumptions of users and the implementation of a decentralized algorithm in Section VII. VI. P ERIOD -1 PROBLEM FOR PORTFOLIO INVESTMENT In the period-1 investment problem, the operator needs to determine the capacities of the solar power, wind power, and energy storage facilities (αs , αw , and αe ) for the entire investment horizon, subject to a budget constraint B.7 These capacity decisions will determine the renewable power production and energy storage flexibility in each day of period-2, and consequently affect the power scheduling and operating cost. The operator wants to make optimal investment decisions to minimize the overall cost, including both the capital investment and the expected operating cost in period-2. The capital investment cost can be represented as CI (αs , αw , αe ) = cs αs + cw αw + ce αe , (15) where cs , cw and ce denote the investment costs of solar power, wind power and energy storage per kW , respectively. The investment cost covers all expenditures, e.g., deployment, installation and maintenance of photovoltaic panel for solar energy, turbine for wind energy, inverters, controllers, and cables. 7 we assume that the microgrid operator makes direct investment for the next long period of operation. 7 The daily expected operating cost Eω [f (·)] is a function of the invested capacities αs , αw and αe , X Eω∈Ω [f (αs , αw , αe , ω)] = πω f (αs , αw , αe , ω), (16) ω∈Ω where ω ∈ Ω denotes the renewable power scenario with a realization probability πω , which is obtained by the scenario reduction algorithm in Section II. Specifically, the operating cost function in scenario ω is the optimized objective value (i.e. minimized operating cost) of the period-2 problem in scenario ω: X ω f (αs , αw , αe , ω) = min [C (Q ) + Ci (xω (17) o i )], ω ω Q ,xi i∈N where the cost depends on the renewable power supply in scenario ω and users’ demand responses in period-2. The period-1 optimization problem is subject to a budget constraint and capacity constraints as follows: (18) min ω ω Ce (xω i ) + Ci (xi ) αs ≥ 0, αw ≥ 0, αe ≥ 0, (19) subject to Constraints (1) and (2), P1: Joint investment and operation in period-1 CI (αs , αw , αe ) + D · Eω∈Ω [f (αs , αw , αe , ω)] min subject to Pui : User i’s cost minimization problem cs αs + cw αw + ce αe ≤ B, where (18) indicates that the total investment expense cannot be greater than the budget B, and the capacity investment must be non-negative. To summarize, the period-1 problem is as follows. αs ,αw ,αe In particular, the operator can set day-ahead prices pω = {pω,t , t ∈ T } for the users in scenario ω, and let users choose the proper power scheduling accordingly to minimize their own costs. In the following, we first present a user’s total cost minimization problem given the prices. Then we discuss the operator’s optimal choices of prices so that the users’ power scheduling decisions coincide with the optimal solution of Problem P2. 1) User’s problem: In scenario ω, user i receives the price signals pω and schedules the power consumption xω i to minimize the total cost. Specifically, user i’s total cost consists of two parts: energy cost Ce and discomfort cost Ci . The energy cost8 of user i depends on the price and user i’s power which can be represented as P consumption, ω,t ω,t xi . Therefore, we have the following Ce (xω i ) = t∈T p total cost minimization problem for user i in scenario ω. Constraints (18) and (19), where the objective function consists of capital investment cost CI and expected operating cost D·Eω [f (·)] under all scenarios ω ∈ Ω over a total D days of operation in the entire investment horizon. xi where each user solves its optimal power consumption as a response to the price signal set by the microgrid operator. 2) Optimal pricing and decentralized algorithm: We denote pω∗ = {pω,t∗ , ∀t ∈ T } as the optimal pricing that induces the socially optimal power consumption {xω∗ i , ∀i ∈ N } in scenario ω (i.e. the optimal solution of Problem P2). We have the following theorem. Theorem 1. In each scenario ω, the optimal pricing scheme pω∗ that induces the socially optimal power consumptions xω∗ i for each user i satisfies the following relationship,  ω ) ω,t  ∂Co (Q , , when Qω,t∗ > rmax ∂Qω,t ω,t∗ Qω,t =Qω,t∗ (20) p = 0, when Qω,t∗ ≤ rω,t . max VII. S OLUTION M ETHOD To solve the above two-period stochastic programming problem, we start with solving the period-2 problem P2 using a distributed algorithm. Then we solve the period-1 problem P1 to obtain the optimal portfolio investment. A. Period-2: Optimal power scheduling As mentioned in Section V, it is not practical for the operator to solve Problem P2 centrally and control users’ power consumptions directly. Instead, the operator and users compute the price and power consumption in an iterative fashion, as shown in Fig. 8. Operator determines the optimal day-ahead prices Pricing Power consumption Consumers decide the optimal power scheduling Fig. 8: Iterative pricing and scheduling Theorem 1 motivates us to design a decentralized algorithm in Algorithm 1, where the operator sets the day-ahead prices and users respond to the prices by determining their power consumptions. At the beginning of each day, the operator and each user’s smart meter compute the hourly electricity prices and the corresponding hourly power consumptions iteratively for the whole-day operation. In Algorithm 1, we consider a sequence of diminishing stepsizes, γk ’s, which satisfy the P following conditions: limk→∞ γ(k) = 0 and limk→∞ k γ(k) = ∞.9 We can see that the decentralized Algorithm 1 requires minimum information exchange between the operator and users. The users only report their power consumptions to the operator, and the operator broadcasts the prices to all users based on the aggregate power load. There is no need for the users to directly coordinate with each other or to reveal their private information (such as cost function and consumption 8 The user cannot change the energy cost related to inelastic load, hence we do not consider that in user’s optimization problem. 9 Problem P2 is convex, and the corresponding decentralized algorithm that is executed once a day can coverage fast in the microgrid context with thousands of electricity users. 8 Algorithm 1 Decentralized algorithm in the microgrid Initialization: iteration index k = 0, error tolerance  > 0, t stepsize γ(k) > 0, predicted scenario ω, and xω,t i (0) = yi . 2: repeat 3: Operator: At the k-th iteration, the operator collects each user’s consumption and computes the aggregate supply Qω , and sets the prices pω,t (k) according to (20). 4: Users: Each user i solves Problem Pui by updating the power consumption xω,t i (k + 1) based on the price pω,t (k):   ∂Ci (xi ω (k)) ω,t ω,t ω,t + p (k) . x̂i (k + 1) = xi (k) − γ(k) ∂xω,t i (k) 1: 5: Project the power consumption on the feasible set by solving the following problem: min xω i (k+1) subject to 6: 7: 8: ω k xω i (k + 1) − x̂i (k + 1) k Constraints (1) and (2). k = k + 1; until k pω (k) − pω (k − 1) k≤ . end constraints). The prices set by the operator are the same for all users, and reflect the total power load without disclosing individual user’s power consumption. Induced by the optimal pricing scheme set by the microgrid operator, the optimal power consumption of each individual user is the socially optimal power consumption, which minimizes the social cost. Theorem 2. Algorithm 1 is a sub-gradient projection algorithm for solving Problem P2, and (with a diminishing stepsize) it converges to the socially optimal price and power consumption {pω∗ , xω∗ } for each ω. B. Period-1: Optimal energy portfolio investment After solving the period-2 problem P2, we solve the period1 problem P1 as follows. Theorem 3. The operator’s investment problem P1 is equivalent to the following optimization problem EP1: min (cs αs + cw αw + ce αe ) h i X + D · Eω Co (Qω ) + Ci (xω ) i subject to Constraints (1), (2), (6)-(10), (13), (18), (19), i∈N Varaibles: αs , αw , αe , Qω , xω , in which we assume that the operator can estimate the information about user’s power consumption behaviors through a survey or daily operations, and the parameters in (1) and (2) are known to the operator. Then the operator solves the equivalent problem EP1 for the optimal investment in a centralized manner. Note that EP1 is a convex quadratic program and can be solved efficiently using a standard interiorpoint method [28]. For the proofs of Theorems, please refer to the online technical report [31]. VIII. U NCERTAINTY IN R ENEWABLE G ENERATION In the previous sections, we have assumed that the operator can predict the solar power generation and wind power generation accurately. Nevertheless, the actual renewable energy generation may deviate from the predicted values due to uncertain weather conditions and different prediction methodologies. Therefore, it is necessary to study the impact of such prediction errors on the optimal operation and investment decisions. In this section, we use the worst-case uncertainty approach [29] to analyze such an issue, by assuming that the uncertain variables are assumed to be bounded in a given uncertainty set. Our aim is to maximize the performance by considering the worst case in the uncertainty set. Note that the robust optimization problem is formulated in period-2 to incorporate the impact of prediction errors of renewable power generations. The period-1 problem, on the other hand, minimizes the overall cost of investment and operation. The detailed formulation is presented in the following. We define the prediction errors for solar power generation and eω,t and wind power generation as eω,t w in time slot t s and each scenario ω. Then we can represent the actual solar ω,t as the power generation η̂sω,t and wind power generation η̂w summation of predicted generations and corresponding errors: η̂sω,t = ηsω,t + eω,t s , ω,t ω,t η̂w = ηw + eω,t w . We assume that the prediction errors are bounded in the following uncertainty sets in scenario ω: n o ω,t ω,t Esω = eω,t , (21) s,min ≤ es ≤ es,max , t ∈ T n o ω,t ω,t Ewω = eω,t , (22) w,min ≤ ew ≤ ew,max , t ∈ T ω,t where eω,t s,min and es,max denote the lower bound and upper bound of the solar power prediction error in time slot t and ω,t scenario ω, respectively; eω,t w,min and ew,max denote the lower bound and upper bound of the wind power prediction error in time slot t and scenario ω, respectively. ω,t ω ω,t Firstly, we let eω s = {es , t ∈ T } and ew = {ew , t ∈ T } denote the prediction error vectors in scenario ω for the solar power generation and wind power generation, respectively. We define the actual operating cost of the microgrid operator in period-2 as Xh
+ i 2 ω ω,t Ĉo (Qω , eω Qω,t − η̂sω,t αs − η̂w αw . s , e w ) = βo t∈T (23) Then, we aim to maximize the worst-case performance of the microgird operation, under all possible prediction errors Esω and Ewω , using robust optimization techniques [29]. Since the investment cost does not explicitly contain prediction errors, we will focus on the worst operating cost minimizing problem. Based on (23), we formulate the operating cost minimization problem in period-2 as a robust optimization problem denoted as RP2: 9 A. Optimal investment ω,∗ We substitute the worst-case prediction errors {eω,∗ s , ew } into RP2, and denote the minimized actual operating cost as X ω,∗ fˆ(αs , αw , αe , ω) = min [Ĉo (Qω , eω,∗ Ci (xω s , ew ) + i )]. ω ω Q ,xi i∈N (25) Based on the minimized actual operating cost (25) in period2, we write the worst-case overall cost minimizing problem in period-1 as Invested Capacity First, we study the optimal investment strategies of solar RP2: Robust optimization for operating cost minimization power, wind power, and energy storage in TC and SKG under " # X different budget constraints, as shown in Fig. 9 and Fig. 10, ω min max Ĉo (Qω , eω Ci (xω s , ew ) + i ) respectively. When the budget is tight (e.g., below 3 million ω ω ω ω Q ,xi es ,ew i∈N HKD), the investment priority at location TC is wind power as subject to Constraints (1), (2), (6)-(10), (13), (21), (22). it’s more economical. While the budget increases, investments To solve RP2, we first solve the inner maximization prob- in solar power and energy storage capacities also increase, lem of RP2, which aims to maximize the actual operating but wind power still dominates the energy portfolio. This is because the wind power output is higher than solar power ω cost with respect to prediction errors eω s and ew . Observing the objective function of RP2, we see that users’ total cost at location TC, and the demand response scheme as well as P energy storage can help better utilize the wind power. The ω i∈N Ci (xi ) is independent of prediction errors. Therefore, we focus on the first term which is the operator’s cost. Given optimal investment expense is 4.6 million HKD at TC, and the operator’s actual power scheduling Qω , we denote the even given more budget (e.g., greater than 5 million HKD), the optimal capacity investment remains the same. From Fig. 10, worst-case prediction errors as we see that the optimal investment strategy at location SKG is ω ω,∗ ω ω different, with the main difference that more investment is put {eω,∗ , e } = arg max Ĉ (Q , e , e ). (24) o s w s w ω ω ω eω s ∈Es , ew ∈Ew into solar power generation. This is because the wind power ω ω ω Since the operator’s actual cost Ĉo (Q , es , ew ) in (23) is a output at SKG is not as adequate as that at TC and solar ω ω power fits the daily demand better at SKG. As a result, wind convex function of eω s and ew (under a given Q ), the optimal ω solution must hit the boundary of the uncertainty sets Es and power generation becomes less competitive at SKG. When Ewω . Moreover, we observe that the actual cost of the operator the budget is less than 4 million HKD, all the investment goes ω to solar power, and when the budget further increases, the Ĉo (Qω , eω s , ew ) is a non-increasing function with respect to ω ω investments in wind power and energy storage increase. The es and ew . Thus we have the following theorem: optimal investment expense is 5.1 million HKD at SKG. Theorem 4. The optimal solutions of (24) hit the lower ω,∗ = {eω,t = bounds, i.e. eω,∗ s s,min , ∀t ∈ T } and ew 800 ω,t {ew,min , ∀t ∈ T }. Solar power (kW) 400 200 0 1 RP1: Robust optimization for overall cost minimization h i min CI (αs , αw , αe ) + D · Eω∈Ω fˆ(αs , αw , αe , ω) Wind power (kW) Energy storage (kWh) 600 2 3 4 5 6 7 8 Investment Budget (Million HKD) Fig. 9: Capacity investment at TC αs ,αw ,αe Constraints (18) and (19), which solves the optimal capacity investment under the worstcase prediction of renewable energy scenarios. Note that RP1 shares the same structure as Problem P1, and thus can be solved by the same methodology presented in Section VII. IX. S IMULATION R ESULTS We assume that the investment horizon includes D = 3650 days (10 years). We use the load curve in [30] as the preferred power consumption, and set βo = 0.005 and βi = 0.5 for i ∈ N . The investment costs of solar energy, wind energy and energy storage are set as cs = 12, 480 HKD per kW, cw = 7, 800 HKD per kW and ce = 1, 950 HKD per kWh. We obtain renewable power scenarios as discussed in Section II, with details described in the online technical report [31]. 500 Invested Capacity subject to Solar power (kW) Wind power (kW) Energy storage (kWh) 400 300 200 100 0 1 2 3 4 5 6 7 8 Investment Budget (Million HKD) Fig. 10: Capacity investment at SKG 10 B. Benefit of demand response 8 Solar power Wind power 6 C. Optimal power scheduling and pricing Next we focus on the numerical studies of the power scheduling in period-2. First, we plot the energy supply and power load at location TC of a typical day in Fig. 13 and Fig. 14, respectively. Note that we treat the aggregate supply from both solar power and wind power as renewable power in period-2. Fig. 13 shows that the renewable energy generation provides high energy supply, especially at night time during hour 7PM-5AM. Renewable energy generation drops drastically during hour 9AM-5PM, because location TC is mainly supplied by wind power and wind power produces less at day time. During this time period, the microgrid operator needs to compensate the loss of renewable energy generation through discharging the energy storage and purchasing energy from the main grid, as we can see that the energy storage level decreases. From Fig. 14, we see that the users respond to the day-ahead prices by optimizing the power load (demand), so that to shift the original peak power load during hour 9AM5PM to other time slots when there is more renewable energy available. Energy storage 300 Total investment 4 2 0 Without incentive With incentive Fig. 11: Energy portfolio investment at TC Energy scheduling (kWh) Optimal Investment (Million HKD) We study the benefit of adopting demand response schemes, and depict the results in Fig. 11 and Fig. 12 for locations TC and SKG, respectively. The investment budget in period-1 is set as 6 million HKD. At location TC, without incentives (in which case we keep the prices low and the same for 24 hours), a user i will choose the power consumption according to the preferred power consumption y i . In that case, it is optimal for the operator to use 5.1 million HKD budget for investment in period-1. With price incentives, the operator sets time dependent day-ahead prices so that to steer the users’ power consumptions to the socially optimal values. Comparing with the case without incentives, optimal demand response under incentives enables the operator to invest more wind power, which has a lower investment cost than solar power. Energy storage investment also decreases with incentives, because the elastic demand can be shifted proactively, which reduces the dependence on the energy storage. The total optimal investment expenditure reduces by 9.4% to 4.6 million HKD, which implies that the demand response may significantly reduce the system cost and avoid over-investment. The total optimal investment expenditure reduces by 6.1% to 5.1 million HKD, demonstrating the benefit of demand response. Renewable energy supply Main grid energy supply Battery energy level 250 200 150 100 50 0 1 5 10 15 20 24 Fig. 13: Energy supply and storage at TC 8 Solar power Wind power Energy storage Total investment 6 4 2 0 300 Power load (kW) Optimal Investment (Million HKD) Time (Hour) 250 200 150 100 50 Without incentive With incentive Fig. 12: Energy portfolio investment at SKG At location SKG, Fig. 12 shows that it is optimal to use 5.4 million HKD budget for investment in period-1 without incentives. However, with price incentives, optimal demand response under incentives enables the operator to invest more solar power but less energy storage, because the solar power fits the demand better compared with the wind power at SKG. 0 1 Original power load Optimized power load 5 10 15 20 24 Time (Hour) Fig. 14: Power load at TC We also study the energy supply and power load of a typical day at location SKG, shown in Fig. 15 and Fig. 16, respectively. Fig. 15 shows that the renewable energy generation provides high supply during day-time, especially 11 Pricing in SKG 1.5 1 0.5 5 10 15 Time (Hour) 20 24 Fig. 17: Day-ahead prices at TC and SKG 200 150 D. Impact of prediction errors 100 We take location TC as an example to study the impact of prediction errors on the optimal investment, as shown in Fig. 18. When there is zero prediction error of renewable generations, the optimal investment is 4.6 million HKD. With the increase of prediction errors, the optimal investment expense increases, as the operator over-invests to hedge the risk of renewable generation shortage caused by prediction errors. For example, when the error is 10%, the operator invests 4.1% over the 4.6 million HKD benchmark. This shows the importance of accurate predication for achieving an optimal investment decisions. 50 0 1 5 10 15 20 24 Time (Hour) Fig. 15: Energy supply and storage at SKG 300 Power load (kW) Pricing in TC 0 1 Renewable energy supply Main grid energy supply Battery energy level 250 2 250 200 150 100 Original power load Optimized power load 50 0 1 5 10 15 20 24 Time (Hour) Fig. 16: Power load at SKG Fig. 17 shows the internal day-ahead prices at TC and SKG. The prices indicate the marginal operating cost of the microgrid. We see that the microgrid operator does not charge the users when the demand can be satisfied by local renewable energy generation (e.g., hour 1AM-3AM at TC). At location TC, the operator charges high prices during day time, but low prices at night time. The reason is that the renewable portfolio consists of more wind power at TC, which has higher power output at night than that during day time. The operator needs to import more main grid power during hour 9AM6PM, which incurs operating cost. At location SKG, the price trend is opposite to the price at TC. Because more solar power is invested at SKG, which produces more renewable power during day time. However, during night time, when there is no solar power supply, the operator relies on main grid power, and thus charges relatively high prices at night. The day-ahead prices set by the operator provide efficiently incentivize users to shift their flexible loads. In Fig. 14 and Fig. 16, we see that users’ flexible loads deviate from the original patterns and are Optimal Investment (Million HKD) Energy scheduling (kWh) 300 shifted from high-price time slots to other low-price time slots at both TC and SKG. Day−ahead price (HKD/kWh) during hour 10AM-4PM. During other time slots, the microgrid operator needs to make sure all the demand can be satisfied, and thus purchases energy from the main grid. The microgrid operator also charges energy storage during high renewable supply periods (11AM-1PM), and discharges the energy storage afterwards as we can see that the energy storage level decreases after hour 7PM. Since location SKG has a large portion of solar power, and there is a power supply peak during day-time, we see from Fig. 16 that the original load during evening-time after hour 6PM is shifted to day-time when there is more renewable energy available. 6 11.1% 5.9% 7.5% 2.1% 4.1% 5 4 3 2 1 0 0 5 10 15 20 Prediction error (%) 25 Fig. 18: Impact of prediction errors E. Comparison with existing studies We add new simulations to compare our proposed method with existing studies in the simulation. Specifically, we consider four benchmark methods, which are motivated by [5], [6], [7] and [8], respectively. We first compute the optimal capacity investment for all the benchmark methods, and then calculate the corresponding investment cost and the actual operational cost using one-year renewable energy data. The simulation results (including the optimal capacity investment, investment cost, operational cost, and overall cost) are summarized in Table I. Benchmark 1 method focuses on the solar-storage investment, and benchmark 2 method focuses on the wind-storage 12 TABLE I: Performance comparison Methods Benchmark 1 Benchmark 2 Benchmark 3 Benchmark 4 Our method [5] [6] [7] [8] Demand response N N N N Y Solar power (kW) 304.4 N 77.8 71.8 42.4 Wind power (kW) N 536.9 494.6 481.4 503.5 Energy storage (kWh) 473.1 298.4 N 233.1 86.5 investment. We see that benchmark 2 achieves a much lower overall cost than benchmark 1 (7.6 vs. 18.9 Million HKD), because wind power is more abundant than solar power, and the investment cost of wind power per unit capacity is also lower than that of the solar power. Benchmark 3 and benchmark 4 show the mixed renewable energy investment with and without energy storage, respectively. We see that benchmark 4 achieves a lower overall cost than benchmark 3 (7.6 vs. 8.4 Million HKD), because the energy storage can help deal with the stochastic nature of renewable energy by proper charging and discharging. Our proposed method considers a comprehensive configuration of microgrids (solar power, wind power, and energy storage), and also optimizes demand response. Hence our proposed method has the minimum investment cost and operational cost, and achieves the minimum overall cost compared with all other benchmark methods. The simulation results in Table I demonstrate the advantage of our proposed method over the existing literature. X. C ONCLUSION We proposed a theoretical framework to study the joint investment and operation problem in the microgrid. The two-period stochastic program models the renewable energy uncertainty, and captures the coupled nature of investment and operation. With realistic meteorological data, our model provides the optimal investment decisions on renewable energy and energy storage capacities, and the optimal demand response scheme (pricing and power scheduling) in a microgrid. The simulation studies demonstrate the economic benefit of demand response and the impact of prediction accuracy. R EFERENCES [1] H. Wang and J. Huang, “Hybrid renewable energy investment in microgrid,” in Proc. of IEEE SmartGridComm, Venice, Italy, 2014. [2] X. Fang, S. Misra, G. Xue, and D. Yang, “Smart grid-the new and improved power grid: a survey,” IEEE Communications Surveys and Tutorials, 14: 944-980, 2012. [3] H. Xu, U. Topcu, S.H. Low, C.R. Clarke, and K.M. Chandy, “Loadshedding probabilities with hybrid renewable power generation and energy storage,” in Proc. of Allerton Conference, 2010. [4] Q. Fu et al., “Microgrid generation capacity design with renewables and energy storage addressing power quality and surety,” IEEE Trans. on Smart Grid, 3(4): 2019-2027, 2012. [5] Y. Ru, J. Kleissl, and S. Martinez, “Storage size determination for gridconnected photovoltaic systems,” IEEE Trans. on Sustainable Energy, 4(1): 68-81, 2013. [6] M. Dicorato, G. Forte, M. Pisani, and M. Trovato, “Planning and operating combined wind-storage system in electricity market,” IEEE Trans. on Sustainable Energy, 3(2): 209-217, 2012. [7] Z. Wang, B. Chen, J. Wang, J. Kim, and M.M. Begovic, “Robust optimization based optimal DG placement in microgrids,” IEEE Trans. on Smart Grid, 5(5): 2173-2182, 2014. [8] P. Yang and A. Nehorai, “Joint optimization of hybrid energy storage and generation capacity with renewable energy,” IEEE Trans. on Smart Grid, 5(4): 1566-1574, 2014. Investment cost (Million HKD) 4.7 4.8 4.8 5.1 4.6 Operational cost (Million HKD) 14.2 2.8 3.6 2.5 2.0 Overall cost (Million HKD) 18.9 7.6 8.4 7.6 6.6 [9] Y. Wang, W. Saad, Z. Han, H.V. Poor and T. Başar, “A game-theoretic approach to energy trading in the smart grid,” IEEE Trans. on Smart Grid, 5(3): 1439-1450, 2014. [10] A.H. Mohsenian-Rad, V.W.S. Wong, J. Jatskevich, R. Schober, and A. Leon-Garcia, “Autonomous demand-side management based on gametheoretic energy consumption scheduling for the future smart grid,” IEEE Trans. on Smart Grid, 1(3): 320-331, 2010. [11] N. Li, L. Chen, and S.H. Low, “Optimal demand response based on utility maximization in power networks,” in Proc. of IEEE PES General Meeting, 2011. [12] L. Qian, Y. Zhang, J. Huang, and Y. Wu, “Demand response management via real-time electricity price control in smart grids,” IEEE Journal on Selected Areas in Communications, 31(7): 1268-1280, 2013. [13] L. Baringo and A.J. Conejo, “Risk-constrained multi-stage wind power investment,” IEEE Trans. on Power Systems, 28(1): 401-411, 2013. [14] L. Baringo and A.J. Conejo, “Strategic wind power investment,” IEEE Trans. on Power Systems, 29(3): 1250-1260, 2014. [15] Y. Cao, M. He, Z. Wang, T. Jiang, and J. Zhang, “Multiple resource expansion planning in smart grids with high penetration of renewable generation,” in Proc. of IEEE SmartGridComm, 2012. [16] F.O. Hocaoglu, O.N. Gerek, and M. Kurban, “A novel hybrid (windphotovoltaic) system sizing procedure,” Solar Energy, 83: 2019-2028, 2009. [17] Y. Li, V.G. Agelidis, and Y. Shrivastava, “Wind-solar resource complementarity and its combined correlation with electricity load demand,” in Proc. of IEEE ICIEA, 2009. [18] J. Widén, “Correlations between large-scale solar and wind power in a future scenario for Sweden,” IEEE Trans. on Sustainable Energy, 2(2): 177-184, 2011. [19] D.G. Luenberger, “Investment science (2nd edition),” Oxford University Press, 2013. [20] A. Möller, W. Römisch, and K. Weber, “A new approach to O & D revenue management based on scenario trees,” Journal of Revenue and Pricing Management, 3(3): 265-276, 2004. [21] N. Gröwe-Kuska, H. Heitsch, and W. Römisch, “Scenario reduction and scenario tree construction for power management problems,” in Proc. IEEE Bologna Power Tech Conference, 2003. [22] L. Jiang and S.H. Low, “Multi-period Optimal Energy Procurement and Demand Response in Smart Grid with Uncertain Supply,” in Proc. of IEEE CDC, 2011. [23] R.H. Inman, H.T.C. Pedro, and C.F.M. Coimbra, “Solar forecasting methods for renewable energy integration,” Progress in Energy and Combustion Science, 39(6): 535-576, 2013. [24] A.M. Foley, P.G. Leahy, A. Marvuglia, and E.J. McKeogh, “Current methods and advances in forecasting of wind power generation,” Renewable Energy, 37(1): 1-8, 2012. [25] A. Mishra et al., “Integrating energy storage in electricity distribution networks,” in Proc. of ACM e-Energy, 2015. [26] X. Ke, N. Lu, and C. Jin, “Control and size energy storage systems for managing energy imbalance of variable generation resources,” IEEE Trans. on Sustainable Energy, 6(1): 70-78, 2015. [27] D. Kirschen and G. Strbac, “Fundamentals of power system economics,” Wiley, 2004. [28] S. Boyd and L. Vandenberghe, “Convex Optimization,” Cambridge Univ. Press, 2004. [29] A. Ben-tal and A. Nemirovski, “Selected topics in robust convex optimization,” Mathematical Programming, 112(1): 125-158, 2008. [30] G. Carpinelli et al., “Optimal integration of distributed energy storage devices in smart grids,” IEEE Trans. on Smart Grid, 4(2): 985-995, 2013. [31] H. Wang and J. Huang, Technical Report.[Online] http://jianwei.ie.cuhk.edu.hk/publication/MicrogridInvestment.pdf [32] Summary of Wind and Solar Meteorological Data in Hong Kong, http: //jianwei.ie.cuhk.edu.hk/energy.html 

3cs.SY

Are You Talking to Me? Reasoned Visual Dialog Generation through Adversarial Learning Qi Wu1 , Peng Wang2 , Chunhua Shen1 , Ian Reid1 , and Anton van den Hengel1 arXiv:1711.07613v1 [cs.CV] 21 Nov 2017 1 Australian Centre for Robotic Vision, The University of Adelaide, Australia 2 Northwestern Polytechnical University, China Abstract The Visual Dialogue task requires an agent to engage in a conversation about an image with a human. It represents an extension of the Visual Question Answering task in that the agent needs to answer a question about an image, but it needs to do so in light of the previous dialogue that has taken place. The key challenge in Visual Dialogue is thus maintaining a consistent, and natural dialogue while continuing to answer questions correctly. We present a novel approach that combines Reinforcement Learning and Generative Adversarial Networks (GANs) to generate more human-like responses to questions. The GAN helps overcome the relative paucity of training data, and the tendency of the typical MLE-based approach to generate overly terse answers. Critically, the GAN is tightly integrated into the attention mechanism that generates humaninterpretable reasons for each answer. This means that the discriminative model of the GAN has the task of assessing whether a candidate answer is generated by a human or not, given the provided reason. This is significant because it drives the generative model to produce high quality answers that are well supported by the associated reasoning. The method also generates the state-of-the-art results on the primary benchmark. Question Human-like Responses Machine-like Are there any large No tall buildings but large one or Yes there are. building nearby? two story buildings, and one clock is in front of looks like church of. With the clock does Yes, I think so because it’s made by I don’t know. it look expensive? stained glass. Do you see any signs for church? Yes, there is a sign with light on, Yes there are. but not clear enough. Figure 1: Human-like vs. Machine-like responses in a visual dialog. The human-like responses clearly answer the questions more comprehensively, and help to maintain a meaningful dialogue. a dialogue about an image. This is significant because it demands that the agent is able to answer a series of questions, each of which may be predicated on the previous questions and answers in the dialogue. Visual Dialogue thus reflects one of the key challenges in AI and Robotics, which is to enable an agent capable of acting upon the world, that we might collaborate with through dialogue. Due to the similarity between the VQA and Visual Dialog tasks, VQA methods [19, 40] have been directly applied to solve the Visual Dialog problem. The fact that the Visual Dialog challenge requires an ongoing conversation, however, demands more than just taking into consideration the state of the conversation thus far. Ideally, the agent should be an engaged participant in the conversation, cooperating towards a larger goal, rather than generating single 1. Introduction The combined interpretation of vision and language has enabled the development of a range of applications that have made interesting steps towards Artificial Intelligence, including Image Captioning [11, 34, 37], Visual Question Answering (VQA) [1, 22, 38], and Referring Expressions [10, 12, 41]. VQA, for example, requires an agent to answer a previously unseen question about a previously unseen image, and is recognised as being an AI-Complete problem [1]. Visual Dialogue [5] represents an extension to the VQA problem whereby an agent is required to engage in 1 word answers, even if they are easier to optimise. Figure 1 provides an example of the distinction between the type of responses a VQA agent might generate and the more involved responses that a human is likely to generate if they are engaged in the conversation. These more human-like responses are not only longer, they provide reasoning information that might be of use even though it is not specifically asked for. Previous Visual Dialog systems [5] follow a neural translation mechanism that is often used in VQA, by predicting the response given the image and the dialog history using the maximum likelihood estimation (MLE) objective function. However, because this over-simplified training objective only focus on measuring the word-level correctness, the produced responses tend to be generic and repetitive. For example, a simple response of ‘yes’,‘no’, or ‘I don’t know’ can safely answer a large number of questions and lead to a high MLE objective value. Generating more comprehensive answers, and a deeper engagement of the agent in the dialogue, requires a more engaged training process. A good dialogue generation model should generate responses indistinguishable from those a human might produce. In this paper, we introduce an adversarial learning strategy, motivated by the previous success of adversarial learning in many computer vision [3, 21] and sequence generation [4, 42] problems. We particularly frame the task as a reinforcement learning problem that we jointly train two sub-modules: a sequence generative model to produce response sentences on the basis of the image content and the dialog history, and a discriminator that leverages previous generator’s memories to distinguish between the humangenerated dialogues and the machine-generated ones. The generator tends to generate responses that can fool the discriminator into believing that they are human generated, while the output of the discriminative model is used as a reward to the generative model, encouraging it to generate more human-like dialogue. Although our proposed framework is inspired by generative adversarial networks (GANs) [9], there are several technical contributions that lead to the final success on the visual dialog generation task. First, we propose a sequential co-attention generative model that aims to ensure that attention can be passed effectively across the image, question and dialog history. The co-attended multi-modal features are combined together to generate a response. Secondly, and significantly, within the structure we propose the discriminator has access to the attention weights the generator used in generating its response. Note that the attention weights can be seen as a form of ‘reason’ for the generated response. For example, it indicates which region should be focused on and what dialog pairs are informative when generating the response. This structure is important as it allows the discriminator to assess the quality of the response, given the reason. It also allows the discriminator to assess the response in the context of the dialogue thus far. Finally, as with most sequence generation problems, the quality of the response can only be assessed over the whole sequence. We follow [42] to apply Monte Carlo (MC) search to calculate the intermediate rewards. We evaluate our method on the VisDial dataset [5] and show that it outperforms the baseline methods by a large margin. We also outperform several state-of-the-art methods. Specifically, our adversarial learned generative model outperforms our strong baseline MLE model by 1.87% on recall@5, improving over previous best reported results by 2.14% on recall@5, and 2.50% recall@10. Qualitative evaluation shows that our generative model generates more informative responses and a human study shows that 49% of our responses pass the Turing Test. We additionally implement a model under the discriminative setting (a candidate response list is given) and achieve the state-of-the-art performance. 2. Related work Visual dialog is the latest in a succession of vision-andlanguage problems that began with image captioning [11, 34, 37], and includes visual question answering [1, 22, 38]. However, in contrast to these classical vision-and-language tasks that only involve at most a single natural language interaction, visual dialog requires the machine to hold a meaningful dialogue in natural language about visual content. Mostafazadeh et al. [20] propose an Image Grounded Conversation (IGC) dataset and task that requires a model to generate natural-sounding conversations (both questions and responses) about a shared image. De Vries et al. [7] propose a GuessWhat game style dataset, where one person asks questions about an image to guess which object has been selected, and the second person answers questions in yes/no/NA. Das et al. [5] propose the largest visual dialog dataset, VisDial, by pairing two subjects on Amazon Mechanical Turk to chat about an image. They further formulate the task as a ‘multi-round’ VQA task and evaluate individual responses at each round in a retrieval or multiplechoice setup. Recently, Das et al. [6] propose to use RL to learn the policies of a ‘Questioner-Bot’ and an ‘AnswererBot’, based on the goal of selecting the right images that the two agents are talking, from the VisDial dataset. Concurrent with our work, Lu et al. [18] propose a similar generative-discriminative model for Visual Dialog. However, there are two differences. First, their discriminative model requires to receive a list of candidate responses and learns to sort this list from the training dataset, which means the model only can be trained when such information is available. Second, their discriminator only considers the generated response and the provided list of candidate responses. Instead, we measure whether the generated Weighted Sum CNN Discriminator + 𝑣෤ What color are the jeans? V: 7 × 7 × 512 Image I Question Q What color are the jeans? Question Q Weighted Sum LSTM + 𝑞෤ 𝑄: 𝐿 × 512 A woman riding on the back of a white horse. Does horse have saddle? Yes it does. How old it woman? I would say maybe 30s or 40s. MC search Is she wearing jeans? Yes she’s wearing jeans. LSTM T rounds of History H ((Caption),(𝑸𝟏 , 𝑨𝟏 ),…, (𝑸𝒕−𝟏 , 𝑨𝒕−𝟏 )) LSTM Human or Machine Reward + 𝑢෤ LSTM LSTM F C Weighted Sum LSTM Is she wearing boots? Yes she is. 𝑢𝑄𝐴 Answer A LSTM What color is saddle? Black and brown. What color is horses mane and tail? Both are bright white color. Decoder They are black. Co-attention Encoder LSTM LSTM 𝐹 softmax U: 𝑇 × 512 Sequential Co-attention Generator Figure 2: The adversarial learning framework of our proposed model. Our model is composed of two components, the first being a sequential co-attention generator that accepts as input image, question and dialog history tuples, and uses the co-attention encoder to jointly reason over them. The second component is a discriminator tasked with labelling whether each answer has been generated by a human or the generative model by considering the attention weights. The output from the discriminator is used as a reward to push the generator to generate responses that are indistinguishable from those a human might generate. response is valid given the attention weights which reflect both the reasoning of the model, and the history of the dialogue thus far. As we show in our experiments in Sec. 4, this procedure results in our generator producing more suitable responses. Dialog generation in NLP Text-only dialog generation [15, 16, 23, 30, 39] has been studied for many years in the Natural Language Processing (NLP) literature, and has leaded to many applications. Recently, the popular ‘Xiaoice’ produced by Microsoft and the ‘Its Alive’ chatbot created by Facebook have attracted significant public attention. In NLP, dialog generation is typically viewed as a sequence-to-sequence (Seq2Seq) problem, or formulated as a statistical machine translation problem [23, 30]. Inspired by the success of the Seq2Seq model [32] in the machine translation, [26, 33] build end-to-end dialog generation models using an encoder-decoder model. Reinforcement learning (RL) has also been applied to train a dialog system. Li et al. [15] simulate two virtual agents and handcraft three rewards (informativity, coherence and ease of answering) to train the response generation model. Recently, some works make an effort to integrate the Seq2Seq model and RL. For example, [2, 31] introduce real users by combining RL with neural generation. Li et al. in [16] were the first to introduce GANs for dialogue generation as an alternative to human evaluation. They jointly train a generative (Seq2Seq) model to produce response sequences and a discriminator to distinguish between human, and machine-generated responses. Although we also introduce an adversarial learning framework to the visual dialog generation in this work, one of the significant differences is that we need to consider the visual content in both generative and discriminative components of the system, where the previous work [16] only requires textual information. We thus designed a sequential co-attention mechanism for the generator and an attention memory access mechanism for the discriminator so that we can jointly reason over the visual and textual information. Critically, the GAN we proposed here is tightly integrated into the attention mechanism that generates human-interpretable reasons for each answer. It means that the discriminative model of the GAN has the task of assessing whether a candidate answer is generated by a human or not, given the provided reason. This is significant because it drives the generative model to produce high quality answers that are well supported by the associated reasoning. More details about our generator and discriminator can be found in Sections 3.1 and 3.2 respectively. Adversarial learning Generative adversarial networks [9] have enjoyed great successes in a wide range of applications in Computer Vision, [3, 21, 24], especially in image generation tasks [8, 43]. The learning process is formulated as an adversarial game in which the generative model is trained to generate outputs to fool the discriminator, while the discriminator is trained not to be fooled. These two models can be jointly trained end-to-end. Some recent works have applied the adversarial learning to sequence generation, for example, Yu et al. [42] backpropagate the error from the discriminator to the sequence generator by using policy gradient reinforcement learning. This 3. Adversarial Learning for Visual Dialog Generation In this section, we describe our adversarial learning approach to generating natural dialog responses based on an image. There are several ways of defining the visual based dialog generation task [7, 20]. We follow the one in [5], in which an image I, a ‘ground truth’ dialog history (including an image description C) H = (C, (Q1 , A1 ),...,(Qt−1 , At−1 )) (we define each QuestionAnswer (QA) pair as an utterance Ut , and U0 = C), and the question Q are given. The visual dialog generation model is required to return a response sentence  = [a1 ,a2 ,...,aK ] to the question, where K is the length (number of words) of the response answer. As in VQA, two types of models may be used to produce the response — generative and discriminative. In a generative decoder, a word sequence generator (for example, an RNN) is trained to fit the ground truth answer word sequences. For a discriminative decoder, an additional candidate response vocabulary is provided and the problem is re-formulated as a multi-class classification problem. The biggest limitation of the discriminative style decoder is that it only can produce a response if and only if it exists in the fixed vocabulary. Our approach is based on a generative model because a fixed vocabulary undermines the general applicability of the model, but also because it offers a better prospect of being extensible to the problem of generating more meaningful dialogue in future. In terms of reinforcement learning, our response sentence generation process can be viewed as a sequence of prediction actions that are taken according to a policy defined by a sequential co-attention generative model. This model is critical as it allows attention (and thus reasoning) to pass across image, question, and dialogue history equally. A discriminator is trained to label whether a response is human generated or machine generated, conditioned on the image, question and dialog attention memories. Considering here that as we take the dialog and the image as a whole into account, we are actually measuring whether the generated response can be fitted into the visual dialog. The output from this discriminative model is used as a reward to the previous generator, pushing it to generate responses that 𝑞1 𝑞2 𝑄 LSTM 𝑄 𝑞𝐿 ෥ 𝒒 … ෩ 𝒖 ෥) 𝑪𝒐𝑨𝒕𝒕𝒆𝒏𝟑 (𝑸, 𝒗′, 𝑣1 𝑣2 𝐻 𝐼 CNN 𝑈0 LSTM … model shows outstanding performance on several sequence generation problems, such as speech generation and poem generation. The work is further extended to more tasks such as image captioning [4, 28] and dialog generation [16]. Our work is also inspired by the success of adversarial learning, but we carefully extend it according to our application, i.e. the Visual Dialog. Specifically, we redesign the generator and discriminator in order to accept multi-modal information (visual content and dialog history). We also apply an intermediate reward for each generation step in the generator, more details can be found in Sec. 3.3. 𝑈𝑡−1 LSTM 𝑉 𝑣𝑁 … 𝑢1 𝑢2 𝑈 ෩ 𝑣′ 𝑢𝑇 … 𝑪𝒐𝑨𝒕𝒕𝒆𝒏𝟏 (𝑽, 𝑸, 𝟎) ෩ 𝒖 ෥, 𝒒 ෥) 𝑪𝒐𝑨𝒕𝒕𝒆𝒏𝟒 (𝒗′, ෥ 𝒗 ෩ 𝑸) 𝑪𝒐𝑨𝒕𝒕𝒆𝒏𝟐 (𝑼, 𝒗′, ෥ 𝒖 Figure 3: The sequential co-attention encoder. Each input feature is coattend by the other two features in a sequential fashion, using the Eq.1-3. The number on each function indicates the sequential order, and the final attended features ũ,ṽ and q̃ form the output of the encoder. are more fitting with the dialog history. In order to consider the reward at the local (i.e. word and phase) level, we use a Monte Carlo (MC) search strategy and the REINFORCE algorithm [36] is used to update the policy gradient. An overview of our model can be found in the Fig. 2. In the following sections, we will introduce each component of our model separately. 3.1. A sequential co-attention generative model We employ the encoder-decoder style generative model which has been widely used in the sequence generation problems. In contrast to text-only dialog generation problem that only needs to consider the dialog history, however, visual dialog generation additionally requires the model to understand visual information. And distinct from VQA that only has one round of questioning, visual dialog has multiple rounds of dialog history that need to be accessed and understood. It suggests that an encoder that can combine multiple information sources is required. A naive way of doing this is to represent the inputs - image, history and question separately and then concatenate them to learn a joint representation. We contend, however, that it is more powerful to let the model selectively focus on regions of the image and segments of the dialog history according to the question. Based on this, we propose a sequential co-attention mechanism [35]. Specifically, we first use a pre-trained CNN [29] to extract the spatial image features V = [v1 , . . . , vN ] from the convolutional layer, where N is the number of image regions. The question features is Q = [q1 , . . . , qL ], where ql = LST M (wl , ql−1 ), which is the hidden state of an LSTM at step l given the input word wl of the question. L is the length of the question. Because the history H is composed by a sequence of utterance, we extract each utterance feature separately to make up the dialog history features, i.e., U = [u0 , . . . , uT ], where T is the number of rounds of the utterance (QA-pairs). And each u is the last hidden state of an LSTM, which accepts the utterance words sequences as the input. Given the encoded image, dialog history and question feature V, U and Q, we use a co-attention mechanism to generate attention weights for each feature type using the other two as the guidance in a sequential style. Each coattention operation is denoted as x̃ = CoAtten(X, g1 , g2 ), which can be expressed as follows: Hi = αi softmax(WT Hi ), PM = i=1 αi xi , x̃ = tanh(Wx xi +Wg1 g1 +Wg2 g2 ), i = 1, . . . ,M, (1) (2) (3) where X is the input feature sequence (i.e., V , U or Q), and g1 , g2 ∈ Rd represent guidances that are outputs of previous attention modules. Here d is the feature dimension. Wx , Wg1 , Wg2 ∈ Rh×d and W ∈ Rh are learnable parameters. Here h denotes the size of hidden layers of the attention module. M is the input sequence length that corresponding to the N, L and T for different feature inputs. As shown in Fig. 3, in our proposed process, the initial question feature is first used to attend to the image. The weighted image features and the initial question representation are then combined to attend to utterances in the dialog history, to produce the attended dialog history (ũ). The attended dialog history and weighted image region features are then jointly used to guide the question attention (q̃). Finally, we run the image attention (ṽ) again, guided by the attended question and dialog history, to complete the circle. All three co-attended features are concatenated together and embedded to the final feature F : F = tanh(Weg [ṽ; ũ; q̃]) (4) where [; ] is a concatenation operator. Finally, this vector representation is fed to an LSTM to compute the probability of generating each token in the target using a softmax function, which forms the response Â. The whole generation process is denoted as π(Â|V,U,Q). 3.2. A discriminative model with attention memories Our discriminative model is a binary classifier that is trained to distinguish whether the input dialog is generated by humans or machines. In order to consider the visual information and the dialog history, we allow the discriminator to access to the attention memories in the generator. Specifically, our discriminator takes {ṽ, ũ, Q, Â} as the input, where ṽ, ũ are the attended image and dialog history features produced in the generative model1 , given the question Q. And  is the generated response in the generator. The Q- pair is further sent to an LSTM to obtain a vector 1 we also tested to use the question memory q̃, but we find the discriminator result is not as good as when using the original question input Q. representation uQ . All three features are embedded together and sent to a 2-way softmax function, which returns the probability distribution of whether the whole visual dialog is human-natural or not: O = tanh(Wed [ṽ; ũ; uQ ]) (5) P = softmax(O) (6) The probability of the visual dialog being recognised as a human-generated dialog is denoted as r({ṽ, ũ, Q, Â}). 3.3. Adversarial REINFORCE with an intermediate reward In adversarial learning, we encourage the generator to generate responses that are close to human generated dialogs, or, in our case, we want the generated response can fit into the visual dialog as good as possible. The policy gradient methods are used here to achieve the goal. The probability of the visual dialog being recognised as a humangenerated dialog by the discriminator (i.e., r({ṽ, ũ, Q, Â})) is used as a reward for the generator, which is trained to maximize the expected reward of generated response using the REINFORCE algorithm [36]: J(θ) = EÂ∼π(Â|V,U,Q) (r({ṽ, ũ, Q, Â})|θ) (7) Given the input visual information (V ), question (Q) and dialog history utterances (U ), the generator generates an response answer  by sampling from the policy. The attended visual (ṽ) and dialog (ũ) memories with the Q and generated answer  are concatenated together and fed to the discriminator. We further use the likelihood ratio trick [36] to approximate the gradient of Eq. 7: ∇J(θ) ≈ ∇ log π(Â|V,U,Q) · [r({ṽ, ũ, Q, Â}) − b] X =∇ log p(ak |V,U,Q,a1:k−1 ) · [r({ṽ, ũ, Q, Â}) − b] k (8) where p is the probability of the generated responses words, ak is the k-th word in the response. b denotes the baseline value. Following [16], we train a critic neural network to estimate the baseline value b by given the current state under the current generation policy π. The critic network takes the visual content, dialog history and question as input, encodes them to a vector representation with our co-attention model and maps the representation to a scalar. The critic neural network is optimised based on the mean squared loss between the estimated reward and the real reward obtained from the discriminator. The entire model can be trained end-to-end, with the discriminator updating synchronously. We use the human generated dialog history and answers as the positive examples and the machine generated responses as negative examples. Intermediate reward An issue in the above vanilla REINFORCE is it only considers a reward value for a finished sequence, and the reward associated with this sequence is used for all actions, i.e., the generation of each token. However, as a sequence generation problem, rewards for intermediate steps are necessary. For example, given a question ‘Are they adults or babies?’, the human-generated answer is ‘I would say they are adults’, while the machine-generated answer is ‘I can’t tell’. The above REINFORCE model will give the same low reward to all the tokens for the machinegenerated answer, but a proper reward assignment way is to give the reward separately, i.e., a high reward to the token ‘I’ and low rewards for the token ‘can’t’ and ‘tell’. Considering that the discriminator is only trained to assign rewards to fully generated sentences, but not intermediate ones, we propose to use the Monte Carlo (MC) search with a roll-out (generator) policy π to sample tokens. An N-time MC search can be represented as: {Â11:K , . . . ,ÂN 1:K } Ân1:k π = MC (Â1:k ; N ) (9) Ânk+1:K where = (a1 , . . . ,ak ) and are sampled based on the roll-out policy π and the current state. We run the roll-out policy starting from the current state till the end of the sequence for N times and the N generated answers share a common prefix Â1:k . These N sequences are fed to the discriminator, the average score rak = N 1 X r({ṽ, ũ, Q, Ân1:K }) N n=1 (10) of which is used as a reward for the action of generating the token ak . With this intermediate reward, our gradient is computed as: ∇J(θ) = ∇ X log p(ak |V,U,Q,a1:k−1 ) · [rak − b] (11) Algorithm 1 Training Visual Dialog Generator with REINFORCE Require: Pretrained generator Gen and discriminator Dis 1: for Each iteration do 2: # Train the generator Gen 3: for i=1, steps do 4: Sample (I,H,Q,A) from the real data 5: Sample (ṽ,ũ,Â) ∼ Genπ (·|I,H,Q) 6: Compute Reward r for (ṽ,ũ,Q,Â) using Dis 7: Evaluate ∇J(θ) with Eq. 8 or 11 depends on whether the intermediate reward (Eq. 10) is used 8: Update Gen parameter θ using ∇J(θ) 9: Update baseline parameters for b 10: Teacher-Forcing: Update Gen on (I,H,Q,A) using MLE 11: # Train the discriminator Dis 12: Sample (I,H,Q,A) from the real data 13: Sample (ṽ,ũ,Â) ∼ Genπ (·|I,H,Q) 14: Update Dis using (ṽ,ũ,Q,A) as positive examples and (ṽ,ũ,Q,Â) as negative examples 4. Experiments We evaluate our model on a recently published visual dialog generation dataset, VisDial [5]. Images in Visdial are all from the MS COCO [17], which contain multiple objects in everyday scenes. The dialogs in Visdial are collected by pairing 2 AMT works (a ‘questioner’ and an ‘answerer’) to chat with each other about an image. To make the dialog measurable, the image remains hidden to the questioner and the task of the questioner is to ask questions about this hidden image to imagine the scene better. The answerer sees the image and his task is to answer questions asked by the questioner. Hence, the conversation is more like multi-rounds of visual based question answering and it only can be ended after 10 rounds. There are 83k dialogs in the COCO training split and 40k in the validation split, for totally 1,232,870 QA pairs, in the Visdial v0.9, which is the latest available version thus far. Following [17], we use 80k dialogs for train, 3k for val and 40k as the test. k where we can see the intermediate rewards for each generation action are considered. Teacher forcing Although the reward returned from the discriminator has been used to adjust the generation process, we find it is still important to feed human generated responses to the generator for the model updating. Hence, we apply a teacher forcing [14, 16] strategy to update the parameters in the generator. Specifically, at each training iteration, we first update the generator using the reward obtained from the sampled data with the generator policy. Then we sample some data from the real dialog history and use them to update the generator, with a standard maximum likelihood estimation (MLE) objective. The whole training process is reviewed in the Alg. 1. 4.1. Evaluation Metrics Different from the previous language generation tasks that normally use BLEU, MENTOR or ROUGE score for evaluation, we follow [17] to use a retrieval setting to evaluate the individual responses at each round of a dialog. Specifically, at test time, besides the image, ground truth dialog history and the question, a list of 100 candidates answers are also given. The model is evaluated on retrieval metrics: (1) rank of human response, (2) existence of the human response in top-k ranked responses, i.e., recall@k and (3) mean reciprocal rank (MRR) of the human response. Since we focus on evaluating the generalization ability of our generator, we simply rank the candidates by the generative model’s log-likelihood scores. Image+Caption Question Human Answer CoAtt-G-MLE Ours A bathroom with a white bath tub, sink and large window. What color is the bathroom? Are there any people in there? Are there towels hanging? Is there any soap on the sink? What color are the towels? What kind of bathtub is it? Can you see anything out the bathroom window? Are there curtains on the window? Is the bathroom light on? Is there anything else on the sink? The walls are gray No No folded up I do n’t think so White A fancy rectangular No No Yes No White No No No soap White It ’s a tub No No Yes No Most white No No, on the floor I do n’t think so White It ’s a shower tub with a shower No, just the wall No curtains Yes No What color is the motorcycle? Is this on a busy street with shops and people? Is it daylight or night time? Is the photo in color? What color are the other cars? Are there any people walking? Can you tell what shops businesses they are? Do you see any traffic lights? Do you think the motorcycle should be parked on the sidewalk? Do you see any signs? It is black and white It looks like it is not Daytime Yes it is I see a white van and a blue Not that i can see Not really No, i do not Yes One, but only a picture White and blue No It ’s daytime Yes white and black no i ’m not sure No i do n’t No Yes It’s black and white No it is not It is daytime Yes One is blue and the other is white no, there are no people I ’m not sure , they are in the background No i do n’t No, it looks like it ’s parked I see a sign on the side of road Is the photo in color? How old does the man appear to be? What color wetsuit? What color surfboard? Do the rocks appear to be smooth or sharp? Is he close to the water? Does it appear to be a beach or private section? What color is the water dark or light blue? Does he have any shoes on? Does he appear to be wet or dry? Yes I would estimate late 30s Dark blue White and red I would guess they are smooth Moderately close Private area It is blurry so it appears black I ca n’t see his feet Dry Yes 20 ’s Black White with red Smooth No I ca n’t tell light blue I ca n’t see his feet Dry Yes I would say 20 ’s Black It ’s white with red They look smooth Yes I ca n’t tell It ’s light blue I ca n’t see his feet He looks dry A motorcycle, moped and a bus parked by the street. A man in a wet suit carrying a surfboard by some rocks. Figure 4: Qualitative results of our model (CoAtt-GAN-w/ Rinte -TF) comparing to human ground-truth answer and our baseline model. 4.2. Implementation Details To pre-process the data, we first lowercase all the texts, convert digits to words, and remove contractions, before tokenizing. The captions, questions and answers are further truncated to ensure that they are no longer than 40, 20 and 20, respectively. We then construct the vocabulary of words that appear at least 5 times in the training split, giving us a vocabulary of 8845 words. The words are represented as one-hot vector and 512-d embeddings for the words are learned. These word embeddings are shared across question, history, decoder LSTMs. All the LSTMs in our model are 1-layered with 512 hidden states. The Adam [13] optimizer is used with the base learning rate of 10−3 , further decreasing to 10−5 . We use 5-time Monte Carlo (MC) search for each token. The co-attention generative model is pretrained using the ground-truth dialog history for 30 epochs. We also pre-train our discriminator (for 30 epochs), where the positive examples are sampled from the ground-truth dialog, the negative examples are sampled from the dialog generated by our generator. The discriminator is updated after every 20 generator-updating steps. 4.3. Experiment results Baselines and comparative models We compare our model with a number of baselines and state-of-the-art models. Answer Prior [5] is a naive baseline that encodes answer options with an LSTM and scored by a linear classifier, which captures ranking by frequency of answers in the training set. NN [5] finds the nearest neighbor images and questions for a test question and its related image. The op- Model Answer Prior [5] NN [5] LF [5] HRE [5] HREA [5] MN [5] HCIAE [18] CoAtt-G-MLE CoAtt-GAN-w/o Rinte CoAtt-GAN-w/ Rinte CoAtt-GAN-w/ Rinte -TF MRR 0.3735 0.4274 0.5199 0.5237 0.5242 0.5259 0.5386 0.5411 0.5415 0.5506 0.5578 R@1 23.55 33.13 41.83 42.29 42.28 42.29 44.06 44.32 44.52 45.56 46.10 R@5 48.52 50.83 61.78 62.18 62.33 62.85 63.55 63.82 64.17 65.16 65.69 R@10 53.23 58.69 67.59 67.92 68.17 68.88 69.24 69.75 70.31 71.07 71.74 Mean 26.50 19.62 17.07 17.07 16.79 17.06 16.01 16.47 16.28 15.30 14.43 Table 1: Performance of generative methods on VisDial v0.9. Higher is better for MRR and recall@k, while lower is better for mean rank. tions are then ranked by their mean-similarity to answers to these questions. Late Fusion (LF) [5] encodes the image, dialog history and question separately and later concatenated together and linearly transformed to a joint representation. HRE [5] applies a hierarchical recurrent encoder [27] to encode the dialog history and the HREA [5] additionally adds an attention mechanism on the dialogs. Memory Network (MN) [5] maintains each previous question and answer as a ‘fact’ in its memory bank and learns to refer to the stored facts and image to answer the question. A concurrent work [18] proposes a HCIAE (HistoryConditioned Image Attentive Encoder) to attend on image and dialog features. From Table 1, we can see our final generative model CoAtt-GAN-w/ Rinte -TF performs the best on all the evaluation metrics. Comparing to the previous state-of-the-art model MN [5], our model outperforms it by 3.81% on R@1. We also produce better results than the HCIAE [18] model, Model LF [5] HRE [5] HREA [5] MN [5] SAN-QI [40] HieCoAtt-QI [19] AMEM [25] HCIAE-NP-ATT [18] Ours MRR 0.5807 0.5846 0.5868 0.5965 0.5764 0.5788 0.6160 0.6222 0.6398 R@1 43.82 44.67 44.82 45.55 43.44 43.51 47.74 48.48 50.29 R@5 74.68 74.50 74.81 76.22 74.26 74.49 78.04 78.75 80.71 R@10 84.07 84.22 84.36 85.37 83.72 83.96 86.84 87.59 88.81 Mean 5.78 5.72 5.66 5.46 5.88 5.84 4.99 4.81 4.47 Table 2: Performance of discriminative methods on VisDial v0.9. Higher is better for MRR and recall@k, while lower is better for mean rank. which is the previous best results that without using any discriminative knowledges. Figure 4 shows some qualitative results of our model. More results can be found in the supplementary material. Ablation study Our model contains several components. In order to verify the contribution of each component, we evaluate several variants of our model. • CoAtt-G-MLE is the generative model that uses our co-attention mechanism shown in Sec. 3.1. This model is trained only with the MLE objective, without any adversarial learning strategies. Hence, it can be used as a baseline model for other variants. • CoAtt-GAN-w/o Rinte is the extension of above CoAtt-G model, with an adversarial learning strategy. The reward from the discriminator is used to guide the generator training, but we only use the global reward to calculate the gradient, as shown in Equ. 8. • CoAtt-GAN-w/ Rinte uses the intermediate reward as shown in the Equ. 10 and 11. • CoAtt-GAN-w/ Rinte -TF is our final model which adds a ‘teacher forcing’ after the adversarial learning. Our baseline CoAtt-G-MLE model outperforms the previous attention based models (HREA, MN, HCIAE) shows that our co-attention mechanism can effectively encode the complex multi-source information. CoAtt-GAN-w/o Rinte produces slightly better results than our baseline model by using the adversarial learning network, but the improvement is limited. The intermediate reward mechanism contributes the most to the improvement, i.e., our proposed CoAttGAN-w/ Rinte model improves over our baseline by average 1%. The additional Teacher-Forcing model (our final model) brings the further improvement, by average 0.5%, achieving the best results. Discriminative setting We additionally implement a model for the discriminative task on the Visdial dataset [5]. In this discriminative setting, there is no need to generate a string, instead, a pre-defined answer set is given and the problem is formulated as a classification problem. We modify our model by replacing the response generation LSTM (can be treated as a multi-step classification process) as a single-step classifier. HCIAE-NP-ATT [18] is the origi- M1: Percentage of responses that pass the Turing Test M2: Percentage of responses that are evaluated as better or equal to human responses. MN [5] CoAtt-G-MLE Ours 0.39 0.46 0.49 0.36 0.42 0.45 Table 3: Human evaluation on 1000 sampled responses on VisDial v0.9 nal HCIAE model with a n-pair discriminative loss and a self-attention mechanism. AMEM [25] applies a more advanced memory network to model the dependency of current question on previous attention. Additional two VQA models [19, 40] are used for comparison. Table 2 shows that our model outperforms the previous baseline and stateof-the-art models on all the evaluation metrics. 4.4. Human study Above experiments verify the effectiveness of our proposed model on the Visdial [5] task. In this section, to check whether our model can generate more human-like dialogs, we conduct a human study. We randomly sample 1000 results from the test dataset in different length, generated by our final model, our baseline model CoAtt-G-MLE, and the Memory Network (MN)2 [5] model. We then ask 3 human subjects to guess whether the last response in the dialog is human-generated or machine-generated and if at least 2 of them agree it is generated by a human, we say it passed the Truing Test. Table 3 summarizes the percentage of responses in the dialog that passes the Turing Test (M1), we can see our model outperforms both the baseline model and the MN model. We also apply our discriminator model in Sec. 3.2 on these 1000 samples and it recognizes that nearly 70% percent of them as human-generated responses (random guess is 50%), which suggests that our final generator successfully fool the discriminator in this adversarial learning. We additionally record the percentage of responses that are evaluated as better than or equal to human responses (M2), according to the human subjects’ manual evaluation. As shown in Table 3, 45% of the responses fall into this case. 5. Conclusion Visual Dialog generation is an interesting topic that requires machine to understand visual content, natural language dialog and have the ability of multi-modal reasoning. More importantly, as a human-computer interaction interface for the further robotics and AI, apart from the correctness, the human-like level of the generated response is a significant index. In this paper, we have proposed an adversarial learning based approach to encourage the generator to generate more human-like dialogs. Technically, by combining a sequential co-attention generative model that can 2 we use the author provided code and pre-trained model provided on https://github.com/batra-mlp-lab/visdial jointly reason the image, dialog history and question, and a discriminator that can dynamically access to the attention memories, with an intermediate reward, our final proposed model achieves the state-of-art on VisDial dataset. A Turing Test fashion study also shows that our model can produce more human-like visual dialog responses. References [1] S. Antol, A. Agrawal, J. Lu, M. Mitchell, D. Batra, C. L. Zitnick, and D. Parikh. VQA: Visual Question Answering. In Proc. IEEE Int. Conf. Comp. Vis., pages 2425–2433, 2015. 1, 2 [2] N. Asghar, P. Poupart, J. Xin, and H. Li. Online sequence-tosequence reinforcement learning for open-domain conversational agents. arXiv preprint arXiv:1612.03929, 2016. 3 [3] X. Chen, Y. Duan, R. Houthooft, J. Schulman, I. Sutskever, and P. Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. In Advances in Neural Information Processing Systems, pages 2172–2180, 2016. 2, 3 [4] B. Dai, D. Lin, R. Urtasun, and S. Fidler. Towards diverse and natural image descriptions via a conditional gan. arXiv preprint arXiv:1703.06029, 2017. 2, 4 [5] A. Das, S. Kottur, K. Gupta, A. Singh, D. Yadav, J. M. Moura, D. Parikh, and D. Batra. Visual dialog. In Proc. IEEE Int. Conf. Comp. Vis., 2017. 1, 2, 4, 6, 7, 8 [6] A. Das, S. Kottur, J. M. Moura, S. Lee, and D. Batra. Learning cooperative visual dialog agents with deep reinforcement learning. Proc. IEEE Int. Conf. Comp. Vis., 2017. 2 [7] H. de Vries, F. Strub, S. Chandar, O. Pietquin, H. Larochelle, and A. Courville. Guesswhat?! visual object discovery through multi-modal dialogue. Proc. IEEE Conf. Comp. Vis. Patt. Recogn., 2017. 2, 4 [8] E. L. Denton, S. Chintala, R. Fergus, et al. Deep generative image models using a laplacian pyramid of adversarial networks. In Advances in neural information processing systems, pages 1486–1494, 2015. 3 [9] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial nets. In Proc. Advances in Neural Inf. Process. Syst., pages 2672–2680, 2014. 2, 3 [10] R. Hu, H. Xu, M. Rohrbach, J. Feng, K. Saenko, and T. Darrell. Natural language object retrieval. In Proc. IEEE Conf. Comp. Vis. Patt. Recogn., June 2016. 1 [11] A. Karpathy and L. Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In Proc. IEEE Conf. Comp. Vis. Patt. Recogn., pages 3128–3137, 2015. 1, 2 [12] S. Kazemzadeh, V. Ordonez, M. Matten, and T. L. Berg. Referit game: Referring to objects in photographs of natural scenes. In EMNLP, 2014. 1 [13] D. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7 [14] A. M. Lamb, A. G. A. P. GOYAL, Y. Zhang, S. Zhang, A. C. Courville, and Y. Bengio. Professor forcing: A new algorithm for training recurrent networks. In Advances In Neural Information Processing Systems, pages 4601–4609, 2016. 6 [15] J. Li, W. Monroe, A. Ritter, M. Galley, J. Gao, and D. Jurafsky. Deep reinforcement learning for dialogue generation. arXiv preprint arXiv:1606.01541, 2016. 3 [16] J. Li, W. Monroe, T. Shi, A. Ritter, and D. Jurafsky. Adversarial learning for neural dialogue generation. arXiv preprint arXiv:1701.06547, 2017. 3, 4, 5, 6 [17] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer, 2014. 6 [18] J. Lu, A. Kannan, J. Yang, D. Parikh, and D. Batra. Best of both worlds: Transferring knowledge from discriminative learning to a generative visual dialog model. arXiv preprint arXiv:1706.01554, 2017. 2, 7, 8 [19] J. Lu, J. Yang, D. Batra, and D. Parikh. Hierarchical question-image co-attention for visual question answering. In Proc. Advances in Neural Inf. Process. Syst., pages 289– 297, 2016. 1, 8 [20] N. Mostafazadeh, C. Brockett, B. Dolan, M. Galley, J. Gao, G. P. Spithourakis, and L. Vanderwende. Image-grounded conversations: Multimodal context for natural question and response generation. arXiv preprint arXiv:1701.08251, 2017. 2, 4 [21] A. Radford, L. Metz, and S. Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 2, 3 [22] M. Ren, R. Kiros, and R. Zemel. Image Question Answering: A Visual Semantic Embedding Model and a New Dataset. In Proc. Advances in Neural Inf. Process. Syst., volume 1, page 5, 2015. 1, 2 [23] A. Ritter, C. Cherry, and W. B. Dolan. Data-driven response generation in social media. In Proc. Conf. Empirical Methods in Natural Language Processing, pages 583–593. Association for Computational Linguistics, 2011. 3 [24] T. Salimans, I. Goodfellow, W. Zaremba, V. Cheung, A. Radford, and X. Chen. Improved techniques for training gans. In Advances in Neural Information Processing Systems, pages 2234–2242, 2016. 3 [25] P. H. Seo, A. Lehrmann, B. Han, and L. Sigal. Visual reference resolution using attention memory for visual dialog. arXiv preprint arXiv:1709.07992, 2017. 8 [26] I. V. Serban, A. Sordoni, Y. Bengio, A. C. Courville, and J. Pineau. Building end-to-end dialogue systems using generative hierarchical neural network models. In AAAI, pages 3776–3784, 2016. 3 [27] I. V. Serban, A. Sordoni, R. Lowe, L. Charlin, J. Pineau, A. C. Courville, and Y. Bengio. A hierarchical latent variable encoder-decoder model for generating dialogues. In AAAI, pages 3295–3301, 2017. 7 [28] R. Shetty, M. Rohrbach, L. A. Hendricks, M. Fritz, and B. Schiele. Speaking the same language: Matching machine to human captions by adversarial training. arXiv preprint arXiv:1703.10476, 2017. 4 [29] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 4 [30] A. Sordoni, M. Galley, M. Auli, C. Brockett, Y. Ji, M. Mitchell, J.-Y. Nie, J. Gao, and B. Dolan. A neural network approach to context-sensitive generation of conversational responses. arXiv preprint arXiv:1506.06714, 2015. 3 [31] P.-H. Su, M. Gasic, N. Mrksic, L. Rojas-Barahona, S. Ultes, D. Vandyke, T.-H. Wen, and S. Young. Continuously learning neural dialogue management. arXiv preprint arXiv:1606.02689, 2016. 3 [32] I. Sutskever, O. Vinyals, and Q. V. Le. Sequence to sequence learning with neural networks. In Proc. Advances in Neural Inf. Process. Syst., pages 3104–3112, 2014. 3 [33] O. Vinyals and Q. Le. A neural conversational model. arXiv preprint arXiv:1506.05869, 2015. 3 [34] O. Vinyals, A. Toshev, S. Bengio, and D. Erhan. Show and tell: A neural image caption generator. In Proc. IEEE Conf. Comp. Vis. Patt. Recogn., pages 3156–3164, 2014. 1, 2 [35] P. Wang, Q. Wu, C. Shen, and A. v. d. Hengel. The vqamachine: Learning how to use existing vision algorithms to answer new questions. Proc. IEEE Conf. Comp. Vis. Patt. Recogn., 2017. 4 [36] R. J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8(3-4):229–256, 1992. 4, 5 [37] Q. Wu, C. Shen, A. v. d. Hengel, L. Liu, and A. Dick. What Value Do Explicit High Level Concepts Have in Vision to Language Problems? In Proc. IEEE Conf. Comp. Vis. Patt. Recogn., pages 203–212, 2016. 1, 2 [38] Q. Wu, P. Wang, C. Shen, A. Dick, and A. v. d. Hengel. Ask Me Anything: Free-form Visual Question Answering Based on Knowledge from External Sources. In Proc. IEEE Conf. Comp. Vis. Patt. Recogn., pages 4622–4630, 2016. 1, 2 [39] Z. Xu, B. Liu, B. Wang, C. Sun, and X. Wang. Incorporating loose-structured knowledge into lstm with recall gate for conversation modeling. arXiv preprint arXiv:1605.05110, 2016. 3 [40] Z. Yang, X. He, J. Gao, L. Deng, and A. Smola. Stacked Attention Networks for Image Question Answering. In Proc. IEEE Conf. Comp. Vis. Patt. Recogn., pages 21–29, 2016. 1, 8 [41] L. Yu, P. Poirson, S. Yang, A. C. Berg, and T. L. Berg. Modeling context in referring expressions. In Proc. Eur. Conf. Comp. Vis., pages 69–85. Springer, 2016. 1 [42] L. Yu, W. Zhang, J. Wang, and Y. Yu. Seqgan: Sequence generative adversarial nets with policy gradient. In Proc. Conf. AAAI, pages 2852–2858, 2017. 2, 3 [43] H. Zhang, T. Xu, H. Li, S. Zhang, X. Huang, X. Wang, and D. Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. arXiv preprint arXiv:1612.03242, 2016. 3 

2cs.AI

Age-Minimal Online Policies for Energy Harvesting Sensors with Incremental Battery Recharges Ahmed Arafa1 , Jing Yang2 , Sennur Ulukus3 , and H. Vincent Poor1 1 arXiv:1802.02129v2 [cs.IT] 22 Feb 2018 2 Electrical Engineering Department, Princeton University School of Electrical Engineering and Computer Science, Pennsylvania State University 3 Department of Electrical and Computer Engineering, University of Maryland Abstract— A sensor node that is sending measurement updates regarding some physical phenomenon to a destination is considered. The sensor relies on energy harvested from nature to transmit its updates, and is equipped with a finite B-sized battery to save its harvested energy. Energy recharges the battery incrementally in units, according to a Poisson process, and one update consumes one energy unit to reach the destination. The setting is online, where the energy arrival times are revealed causally after the energy is harvested. The goal is to update the destination in a timely manner, namely, such that the long term average age of information is minimized, subject to energy causality constraints. The age of information at a given time is defined as the time spent since the latest update has reached the destination. It is shown that the optimal update policy follows a renewal structure, where the inter-update times are independent, and the time durations between any two consecutive events of submitting an update and having k units of energy remaining in the battery are independent and identically distributed for a given k ≤ B − 1. The optimal renewal policy for the case of B = 2 energy units is explicitly characterized, and it is shown that it has an energy-dependent threshold structure, where the sensor updates only if the age grows above a certain threshold that is a function of the amount of energy in its battery. I. I NTRODUCTION An energy harvesting sensor monitors some physical phenomenon and sends measurement updates about it to a destination. Updates are to be sent such that the long term average age of information is minimized. The age of information is the time spent since the freshest update has reached the destination. The sensor relies on energy harvested from nature to measure and send its updates, and is equipped with a finite B-sized battery to save its incoming energy. We characterize optimal online policies for this problem, where the sensor has only causal knowledge of the energy harvesting process. In this work, we connect results from the energy harvesting communication literature and the age of information minimization literature by using the age of information metric as a means to assess the performance of a single-user energy harvesting communication channel. The energy harvesting communication literature is broadly categorized into offline and online settings, depending on whether the energy arrival times/amounts are known prior to the start of communication. This research was supported in part by the National Science Foundation under Grants ECCS-1549881, ECCS-1647198, ECCS-1650299, CCF 1422111, and CNS 15-26608. Offline energy management works consider, e.g., single-user channels [1]–[4]; multiuser channels [5]–[9]; and multi hop and relay channels [10]–[14]. Recent online works include the near-optimal results for single-user and multiuser channels [15]–[18], systems with processing costs [19], and systems with general utilities [20]. Age of information minimization is generally studied in a queuing-theoretic framework, including a single source setting [21]; multiple sources [22]; variations of the single source setting such as randomly (out of order) arriving updates [23], update management and control [24], and nonlinear age metrics [25], [26]; multi hop networks [27]; broadcasting, multicasting, and multi streaming [28]–[30]; coding over erasures [31]; and caching systems [32]. Assessing the performance of energy harvesting communication systems by the age of information metric has recently gained some attention [33]–[40]. Except for [36], an underlying assumption in these works is that energy expenditure is normalized, i.e., it takes one energy unit to send an update to the destination. References [33], [34] study a system with an infinite-sized battery, with [33] considering online scheduling with random service times (time for the update to take effect), and [34] considering offline and online scheduling with zero service times. The offline policy in [34] is extended to fixed non-zero service times in [35] for single and multi hop settings, and to energy-controlled service times in [36]. The online policy in [34] is found by dynamic programming in a discrete-time setting, and was shown to be of a threshold structure, where an update is sent only if the age of information is higher than a certain threshold. Motivated by the results in the infinite battery case, [37] then analyzes the performance of threshold policies under a finite-sized battery and varying channel assumptions, yet with no claim of optimality. Reference [38] proves the optimality of threshold policies when the battery size is equal to one unit using tools from renewal theory; it also provides an asymptotically optimal update policy when the battery size grows infinitely large. In our recent work [39], we extend the results of [38] and formally prove the optimality of threshold policies for any finite-sized battery in an online setting where the battery is randomly fully recharged over time, i.e., whenever energy is harvested, it completely fills up the battery. An interesting result is recently reported in [40], where status updates send information, other than that related to measurements, in an energy harvesting single-user channel. In this work, we complement our results in [39] and study age-optimal online policies for an energy harvesting sensor with a finite battery with random incremental battery recharges; that is, energy is harvested in units as in [37], [38], as opposed to full chunks as in [39]. We extend the unit battery results of [38] and show that for a finite battery of size B, the optimal status update policy that minimizes the long term average age of information is a renewal policy: the times in between the two consecutive events where the sensor sends an update and has k energy units remaining in its battery, for some 0 ≤ k ≤ B − 1, are independent and identically distributed (i.i.d.). Further, we show that interupdate times are independent. Based on these results, we explicitly solve the case of B = 2 energy units, and formally prove, using optimization tools, that the optimal policy is an energy-dependent threshold policy: the sensor submits an update only if the instantaneous age of information is above a certain threshold that depends on the energy in its battery. age 0 s1 x1 Fig. 1. s2 x2 s3 t time x3 Example of the age evolution versus time with n(t) = 3. The age at time t is formally defined as a(t) , t − u(t) (3) where u(t) is the time stamp of the latest update received before time t. Let n(t) denote the total number of updates sent by time t. We are interested in minimizing the area under the age curve, see Fig. 1 for a possible sample path with n(t) = 3. At time t, this area is given by n(t) II. S YSTEM M ODEL AND P ROBLEM F ORMULATION We consider a sensor node that collects measurements from a physical phenomenon and sends updates to a destination over time. The sensor relies on energy harvested from nature to acquire and send its updates, and is equipped with a battery of finite size B to save its incoming energy. The sensor consumes one unit of energy to measure and send out an update to the destination. We assume that updates are sent over an errorfree link with negligible transmission times as in [34], [37]– [39]. Energy arrives (is harvested) one unit at a time, at times {t1 , t2 , . . . } according to a Poisson process of rate 1. Our setting is online in which energy arrival times are revealed causally over time; only the arrival rate is known a priori. Let si denote the time at which the sensor acquires (and transmits) the ith measurement update, and let E(t) denote the amount of energy remaining in the battery at time t. We then have the following energy causality constraint [1]
E s− ≥ 1, ∀i (1) i We assume that we begin with an empty battery at time 0, and that the battery evolves as follows over time

 (2) = min E s− E s− i−1 − 1 + A (xi ) , B i where xi , si −si−1 , and A(xi ) denotes the amount of energy harvested in [si−1 , si ). Note that A(xi ) is a Poisson random variable with parameter xi . We denote by F , the set of feasible transmission times {si } described by (1) and (2) in addition to an empty battery at time 0, i.e., E(0) = 0. The goal is to choose an online feasible transmission policy {si } (or equivalently {xi }) such that the long term average of the age of information experienced at the destination is minimized. The age of information is defined as the time elapsed since the latest update has reached the destination.
2 1X 2 1 r(t) , x + t − sn(t) 2 i=1 i 2 (4) and therefore the goal is to characterize the following quantity r̄ , min lim sup x∈F T →∞ 1 E [r(T )] T (5) where E(·) is the expectation operator. In the next section, we characterize the structure of the optimal policy. III. O PTIMAL S OLUTION S TRUCTURE : R ENEWAL T YPE P OLICIES In this section, we show that the optimal update policy that solves problem (5) has a renewal structure. Namely, we show that it is optimal to transmit updates in such a way that the inter-update delays are independent over time; and that the time durations in between the two consecutive events of transmitting an update and having k ≤ B − 1 units of energy left in the battery are i.i.d., i.e. these events occur at times that constitute a renewal process. We first introduce some notation. Let the pair (E(t), a(t)) represent the state of the system at time t. Fix k ∈ {0, 1, . . . , B −1}, and consider the state (k, 0), which means that the sensor has just submitted an update and has k units of energy remaining in its battery. Let li denote the time at which the system visits (k, 0) for the ith time. We use the term epoch to denote the time in between two consecutive visits to (k, 0). Observe that there can possibly be an infinite number of updates occurring in an epoch, depending on the energy arrival pattern and the update time decisions. For instance, in the ith epoch, which starts at li−1 , one energy unit may arrive at some time li−1 + τ1,i , at which the system goes to state (k + 1, τ1,i ), and then the sensor updates afterwards to get the system state back to (k, 0) again. Another possibility (if k ≥ 1) is that the sensor first updates at some time li−1 + xk,i , at which the system goes to state (k − 1, 0), and then two consecutive energy units arrive at times li−1 + τ1,i and li−1 +τ1,i +τ2,i , respectively, at which the system goes to state (k+1, τ1,i +τ2,i ), and then the sensor updates afterwards to get the system state back to (k, 0) again. Depending on how many energy arrivals occur in the ith epoch, how far apart from each other they are, and the status update times, one can determine the length of the ith epoch and how many updates it has. Observe that the update policy in the ith epoch may depend on the history of events (energy arrivals and transmission updates) that occurred in previous epochs, which we denote by Hi−1 . Our main result in this section shows that this is not the case, under some mild technical conditions, and that epoch lengths should be i.i.d. We first have the following definition. Definition 1 (Uniformly Bounded Policy) An online policy whose inter-update times, as a function of the energy arrival times, have a bounded second moment. We focus on uniformly bounded policies as per Definition 1. Such policies were also considered in [38] in the analysis of the B = 1 case. We now have the following theorem; the proof is in Appendix A. age age l0 τ2 τ1 Based on Theorem 1, the following corollary now follows. Corollary 1 In the optimal solution of problem (5), the interupdate times are independent. Proof: Observe that whenever an update occurs the system enters state (j, 0) for some j ≤ B − 1. The system then starts a new epoch with respect to state (j, 0). Since the choice of k energy units in Theorem 1 is arbitrary, the results of the theorem now tell us that the update policy in that epoch, and therefore its length, is independent of the past history, in particular the past inter-update lengths.  In the next section, we show how to use the results of Theorem 1 and Corollary 1 to provide an explicit solution for the case of B = 2 energy units. IV. T HE C ASE B = 2 Based on Corollary 1, we now introduce the following notation regarding the update policy in a given epoch. Starting from state (0, 0) at time l0 , the sensor has to wait for the first energy arrival in the epoch, which occurs after some time τ1 , and at which the system state becomes (1, τ1 ). Since the sensor now has energy, it schedules its next update at l0 + y1 (τ1 ), for some function y1 (·) to be optimally characterized. Now if another energy arrival occurs at time l0 + τ1 + τ2 , with τ2 > y1 (τ1 ) − τ1 , the sensor transmits the update as scheduled at l0 + y1 (τ1 ) and the system state returns to (0, 0) again. On the other hand, if this second energy arrival occurs relatively early, i.e., τ2 ≤ y1 (τ1 ) − τ1 , the system state becomes (2, τ1 + τ2 ) at l0 + τ1 + τ2 , and the sensor reschedules its update to be at l0 τ1 τ2 time ȳ2 (τ1 , τ2 ) y1 (τ1 ) Fig. 2. Age of information versus time under the two possible ways of updating starting from state (0, 0) at time l0 . On the left, the second energy arrival occurs late, and hence we have one energy arrival followed by one update, returning to state (0, 0) again at l0 + y1 (τ1 ). On the right, the second energy arrival occurs early, and hence we have two energy arrivals followed by one update, entering state (1, 0) at l0 + ȳ2 (τ1 , τ2 ). The yellow boxes represent energy units in the battery. age age l1 τ1 x1 Theorem 1 In the optimal solution of problem (5), any uniformly bounded policy is a renewal policy. That is, the sequence {li } denoting the times at which the system visits state (k, 0) forms a renewal process. time time l1 τ1 time y2 (τ1 ) Fig. 3. Age of information versus time under the two possible ways of updating starting from state (1, 0) at time l1 . On the left, the first energy arrival occurs late, and hence the sensor updates and enters state (0, 0) at l1 + x1 . On the right, the first energy arrival occurs early, and hence we have one energy arrival followed by an update, returning to state (1, 0) at l1 + y2 (τ1 ). The yellow boxes represent energy units in the battery. l0 + ȳ2 (τ1 , τ2 ) instead of l0 + y1 (τ1 ). Note that it is not clear so far whether ȳ2 (τ1 , τ2 ) depends only on the age τ1 + τ2 ; we leave it as a general function of the pair (τ1 , τ2 ) for now. The above two cases are illustrated in Fig. 2. Once the sensor has two energy units in its battery, it will eventually send an update making the system state become (1, 0) at some time l1 . The sensor then schedules its next update at l1 + x1 , for some x1 to be optimally characterized. If the first energy arrival after l1 occurs at time l1 + τ1 with τ1 > x1 , the sensor transmits the update at l+x1 as scheduled, whence the state becomes (0, 0). Note that by energy causality, x1 cannot depend on τ1 , and since it also does not depend on the past history before l1 (by Corollary 1), it is therefore a constant. On the other hand, if the first energy arrival occurs relatively early, i.e., τ1 ≤ x1 , the state becomes (2, τ1 ) at l1 + τ1 , and the sensor reschedules the update to be at l1 + y2 (τ1 ) instead of l1 + x1 . Note that it is not clear so far whether y2 (·) and ȳ2 (·, ·) are identical, since the former depends on only one random variable, as opposed to depending on two random variables in the latter; we optimally characterize both functions later on in the analysis. The above two cases are illustrated in Fig. 3. In summary, the optimal update policy for the case B = 2 in a given epoch is completely characterized by the constant x1 , and the functions y1 (·), y2 (·), and ȳ2 (·, ·). Since these represent the possible inter-update delays, we conclude by Corollary 1 that they do not depend on each other. We denote age age τ1 time τ2 y1 (τ1 ) τ1 Fig. 4. First possible way to return to state (0, 0). ȳ2 (τ1 , τ2 ) by R (x1 , y1 , y2 , ȳ2 ) and L (x1 , y1 , y2 , ȳ2 ) the area under the age curve in the epoch and its length, respectively, as a function of the policy (x1 , y1 , y2 , ȳ2 ). By Theorem 1 (and Corollary 1), one can use the strong law of large numbers of renewal processes [41] to reduce problem (5) to be an optimization over a single epoch as follows min x1 ,y1 ,y2 ,ȳ2 s.t. Fig. 5. τ1 Second possible way to return to state (0, 0). τ2 ȳ2 (τ1 , τ2 ) ∀τ ∀τ ȳ2 (τ1 , τ2 ) ≥ τ1 + τ2 , Fig. 6. ∀τ1 , τ2 x1 age E [R (x1 , y1 , y2 , ȳ2 )] E [L (x1 , y1 , y2 , ȳ2 )] x1 ≥ 0 y1 (τ ) ≥ τ, y2 (τ ) ≥ τ, time τ3 τ2 τ3 τ4 y2 (τ3 ) time x1 Third possible way to return to state (0, 0). (6) where the expectation is on the energy arrival patterns in the epoch. Note that the constraints on the functions y1 , y2 , and ȳ2 , represent the energy causality constraints. Next, in order to evaluate the expectations in the objective function, one needs to study the different patterns that can occur in a single epoch. We do so in the following subsection. A. Renewal State Analysis Consider the state (0, 0) as the renewal state1 , and without loss of generality assume that we start at time 0. Let us now state the possible ways of returning to that state. Note that the sensor has to wait for at least one energy arrival to update since it starts with no energy at state (0, 0). • The first way to return to (0, 0) is to receive an energy arrival after τ1 time units, and then update at y1 (τ1 ). This could happen if and only if the following energy arrival, occurring at τ2 time units after the first arrival, arrives after y1 (τ1 ) − τ1 . See Fig. 4. • The second way is to receive another energy arrival after the first one, before using the first energy unit to update. Then, submit the first update at ȳ2 (τ1 , τ2 ), which makes the state become (1, 0), and then submit another update after x1 time units. This could happen if and only if the following energy arrival, occurring at τ3 time units after the first update, is such that τ3 > x1 . See Fig. 5. • The third way is exactly as the second way, but with τ3 ≤ x1 , and hence the system goes to state (2, τ3 ) with 1 From Theorem 1, we know that both states (0, 0) and (1, 0) are renewal states. While we choose to perform our analysis using state (0, 0), we note that one can reach the same results if state (1, 0) is chosen instead. • the third energy arrival. Then, the sensor updates after y2 (τ3 ) time units from the first update (as opposed to x1 in the second way), which makes the state become (1, 0), and then finally submit a third update after x1 time units. As before, this could happen if and only if the following energy arrival, occurring at τ4 time units after the second update, is such that τ4 > x1 . See Fig. 6. In general, the mth way, m ≥ 3, begins exactly as in the second way by submitting the first update at ȳ2 (τ1 , τ2 ). Then, the second phase of the third way, namely, going from state (1, 0) to (2, τ3 ) to (1, 0) again, keeps repeating for m−2 times. By the end of these repetitions the system will be in state (1, 0). This is finally followed by the mth (and last) update after x1 time units. See Fig. 7. Based on the above, one can write the area under the age curve, R, in a single epoch as in equation (7) at the top of the next page2 . There, 1A = 1 if event A is true, and is 0 otherwise. Taking expectations and simplifying (mainly through using the fact that τi ’s are i.i.d.), we get   Z x1 1 1 2 x1 2 −τ x +e y2 (τ ) e dτ E [R] = 2 1 2   Z ∞ Z ∞ 0 1 y1 (τ )2 e−y1 (τ ) dτ × 1− e−y1 (τ ) dτ + 2 0 0 Z ∞ Z y1 (τ1 )−τ1 1 + ȳ2 (τ1 , τ2 )2 e−τ1 e−τ2 dτ1 dτ2 2 τ1 =0 τ2 =0 (9) 2 From now onwards, we drop the dependency on the tuple (x , y , y , ȳ ) 1 1 2 2 from R and L for convenience. mth triangle age ... τ1 y2 (τ4 ) y2 (τ3 ) ȳ2 (τ1 , τ2 ) Fig. 7. time τm+1 τ4 τ3 τ2 x1 General mth possible way to return to state (0, 0), m ≥ 3.   1 2 1 1 2 2 ȳ2 (τ1 , τ2 ) + x1 1τ2 ≤y1 (τ1 )−τ1 1τ3 >x1 R = y1 (τ1 ) 1τ2 >y1 (τ1 )−τ1 + 2 2 2   1 1 1 + ȳ2 (τ1 , τ2 )2 + y2 (τ3 )2 + x21 1τ2 ≤y1 (τ1 )−τ1 1τ3 ≤x1 1τ4 >x1 2 2 2   1 1 1 1 ȳ2 (τ1 , τ2 )2 + y2 (τ3 )2 + y2 (τ4 )2 + x21 1τ2 ≤y1 (τ1 )−τ1 1τ3 ≤x1 1τ4 ≤x1 1τ5 >x1 + 2 2 2 2 + ... L =y1 (τ1 )1τ2 >y1 (τ1 )−τ1 + (ȳ2 (τ1 , τ2 ) + x1 ) 1τ2 ≤y1 (τ1 )−τ1 1τ3 >x1 (7) + (ȳ2 (τ1 , τ2 ) + y2 (τ3 ) + x1 ) 1τ2 ≤y1 (τ1 )−τ1 1τ3 ≤x1 1τ4 >x1 + (ȳ2 (τ1 , τ2 ) + y2 (τ3 ) + y2 (τ4 ) + x1 ) 1τ2 ≤y1 (τ1 )−τ1 1τ3 ≤x1 1τ4 ≤x1 1τ5 >x1 + ... Equation (9) is justified in Appendix B. Similarly, the epoch length L is given by (8) at the top of this page, and its expectation is given by   Z x1 E [L] = x1 + ex1 y2 (τ )e−τ dτ   Z ∞ Z ∞0 −y1 (τ ) × 1− e dτ + y1 (τ )e−y1 (τ ) dτ 0 + Z ∞ τ1 =0 Z 0 y1 (τ1 )−τ1 ȳ2 (τ1 , τ2 )e−τ1 e−τ2 dτ1 dτ2 (10) τ2 =0 We use the above results to characterize the structure of the optimal policy for problem (6) in the next subsection. B. Optimal Solution for Problem (6): Threshold Policies We define the following parameterized problem to characterize the optimal solution of problem (6) p2 (λ) , min x1 ,y1 ,y2 ,ȳ2 s.t. E [R] − λE [L] x1 ≥ 0 y1 (τ ) ≥ τ, y2 (τ ) ≥ τ, ∀τ ∀τ ȳ2 (τ1 , τ2 ) ≥ τ1 + τ2 , ∀τ1 , τ2 (11) where the subscript 2 in p2 (λ) denotes the B = 2 case that we consider here. This approach has also been used in [42]. We now have the following lemma. (8) Lemma 1 p2 (λ) is decreasing in λ, and the optimal solution of problem (6) is given by λ∗ that solves p2 (λ∗ ) = 0. Proof: Let λ1 > 0, and let the solution of problem (11) (1) (1) (1) (1) for λ = λ1 , with the be given by x1 , y1 , y2 , ȳ2 corresponding average area under the age curve  in the epoch  and the average epoch length given by E R(1) and E L(1) , respectively. Now for some λ2 > λ1 , one can write h i h i p2 (λ1 ) = E R(1) − λ1 E L(1) i i h h > E R(1) − λ2 E L(1) ≥ p2 (λ2 ).  (1) (1) (1) (12)  (1) where the last inequality follows since x1 , y1 , y2 , ȳ2 is also feasible in problem (11) for λ = λ2 . Next, note that both problems (11) and (6) have the same feasible set. In addition, if p2 (λ) = 0, then the objective function of (6) satisfies E [R] /E [L] = λ. Hence, the objective function of (6) is minimized by minimizing λ ≥ 0 such that p2 (λ) = 0. Finally, by the first part of lemma, there can only be one such λ, which we denote λ∗ .  By Lemma 1, one can simply use a bisection method to find λ∗ that solves p2 (λ∗ ) = 0. This λ∗ certainly exists since p2 (0) > 0 and limλ→∞ p2 (λ) = −∞. We focus on problem (11) in the rest of this subsection, for which we introduce the following Lagrangian [43] Z ∞ L =E [R] − λE [L] − η1 x1 − γ1 (τ ) (y1 (τ ) − τ ) dτ 0 Z ∞ − γ2 (τ ) (y2 (τ ) − τ ) dτ Z0 ∞ Z ∞ − γ̄2 (τ1 , τ2 ) (ȳ2 (τ1 , τ2 ) − τ1 − τ2 ) dτ1 dτ2 0 0 (13) where η1 , γ1 (·), γ2 (·), γ̄2 (·, ·) are Lagrange multipliers. Using (9) and (10), we take the (functional) derivative of the Lagrangian with respect to ȳ2 (t1 , t2 ) and equate it to 0 to get γ̄2 (t1 , t2 ) (14) e−(t1 +t2 ) Now if t1 +t2 < λ, then ȳ2 (t1 , t2 ) has to be larger than t1 +t2 , for if it were equal, the right hand side of the above equation would be larger than the left hand side. By complementary slackness [43], we conclude that in this case γ̄2 (t1 , t2 ) = 0, and hence ȳ2 (t1 , t2 ) = λ. On the other hand, if t1 + t2 ≥ λ, then ȳ2 (t1 , t2 ) has to be equal to t1 + t2 , for if it were larger, then by complementary slackness γ̄2 (t1 , t2 ) = 0 and the right hand side of the above equation would be smaller than the left hand side. In conclusion, we have ( λ, t1 + t2 < λ ȳ2 (t1 , t2 ) = (15) t1 + t2 , t1 + t2 ≥ λ ȳ2 (t1 , t2 ) = λ + The above result says that starting from state (0, 0) the sensor has to wait at least for λ time units before submitting an update, provided that it received two consecutive energy units (without using the first one to send an update) in that epoch. If these two energy arrivals occur relatively early, i.e., t1 + t2 < λ, then the sensor updates exactly after λ time units from the beginning of the epoch. Otherwise, if t1 + t2 ≥ λ, then the sensor updates instantly after receiving the second energy unit. We coin this type of policies λ-threshold policy, where the sensor can only update if the age grows above a certain threshold λ. Such policies were first introduced in the solution of the case of B = 1 energy unit in [38], and have also appeared in the random full battery recharges analysis in [39]. We also note from the result in (15) that ȳ2 (t1 , t2 ) only depends on the age at the second energy arrival, t1 + t2 . Next, we take the derivative of the Lagrangian with respect to y2 (t) and equate to 0 to get y2 (t) = λ + γ2 (t) qex1 e−t (16) R∞ where q , 1 − 0 e−y1 (τ ) dτ . Note that q ∈ [0, 1] since y1 (τ ) ≥ τ . Following the same arguments as in the ȳ2 case, we get that ( λ, t<λ y2 (t) = (17) t, t≥λ That is, y2 is also a λ-threshold policy, and ȳ2 (t1 , t2 ) = y2 (t1 + t2 ). This settles the earlier question we posed at the beginning of this section of whether receiving two energy arrivals starting from state (0, 0) would lead to a different policy than receiving one energy arrival starting from state (1, 0); the optimal policy when the sensor has a full battery is only a function of the age at the time of receiving the second energy unit in the battery. Next, we take the derivative of the Lagrangian with respect to x1 and equate to 0 to get  Z x1  1 2 x1 x1 =λ + e λy2 (τ ) − y2 (τ ) e−τ dτ 2 0

1 η1 − 2 + λy2 x− + (18) 1 − y2 x1 2 q We now make an assumption that x1 >
λ, and verify that assumption below. Based on that, y2 x− 1 = x1 from (17). One can also use (17) to evaluate the integral in the above equation in terms of λ and x1 . After some algebraic manipulations, we get that for x1 > 0, η1 = 0 by complementary slackness, and the following holds   1 (19) x1 = log −λ 1 2 e − 2λ where log is the natural logarithm. It is direct to see from (19) that x1 > λ as assumed above. Finally, we take the derivative of the Lagrangian with respect to y1 (t) and equate to 0 to get  Z x1 1 1 x1 2 λy2 (τ ) − y2 (τ ) e−τ dτ + λx1 − x21 y1 (t) =λ + e 2 2 0 2 1  1 + y1 (t)2 − ȳ2 t, (y1 (t) − t)− 2 2   γ1 (t) − − λy1 (t) + λȳ2 t, (y1 (t) − t) + −y (t) (20) e 1 We now make another assumption that y1 (t) > λ, ∀t, and verify it below.  assumption, we conclude by  Based on this − = y1 (t). We substitute this in (15) that ȳ2 t, (y1 (t) − t) (20), and use (18) to get γ1 (t) (21) e−y1 (t) which verifies that y1 (t) > λ, ∀t, since x1 > λ. Similar to the arguments used in deriving (15) and (17), we conclude from (21) that y1 is an x1 -threshold policy given by ( x1 , t < x1 y1 (t) = (22) t, t ≥ x1 y1 (t) = x1 + Similar to the discussion regarding the equivalence of ȳ2 and y2 , we conclude from (22) that starting from state (0, 0) and receiving one energy unit is equivalent to starting from state (1, 0) and receiving no energy units; in both cases, the sensor has the same threshold x1 after which it can update. Using (15), (17), (19), and (22) we get that 1 p2 (λ) = λ2 + (λ + 1)e−λ + λ 2     1 2 1 −λ − e − λ + 1 log 2 e−λ − 21 λ2 (23) Long term average age 1.5 1 0.5 0 Optimal z=0 (uniform) z=1 z=2 Update policy Fig. 8. Comparison of the optimal policy for B = 2 to other policies: uniform updating, and energy-aware adaptive updating of [38]. It now remains to find λ∗ . Towards that, we first note that we have an upper bound on λ∗ given by 0.9012, the solution of the B = 1 case derived in [38]. We also have a lower bound of 0.5, which is the optimal solution in the case of having an infinite battery, also derived in [38]. Using bisection, we find that the optimal solution at which p2 (λ∗ ) = 0 is given by λ∗ ≈ 0.72, with the corresponding x∗1 ≈ 1.48. Observe that the fact that x∗1 is larger than λ∗ implies the intuitive behavior that the sensor is less eager to send an update if it has only one energy unit, compared to when it has a full battery of two energy units. C. Comparison to Other Policies We now compare the optimal result derived above with other schemes and system models in the literature. We first compare it to the energy-aware adaptive status update policy introduced and analyzed in [38]. In there, the sensor schedules its next update based on the amount of energy in its battery; if the energy is less than B/2, it schedules the next update after 1/(1 − β) time units, for some constant β < 1; if the energy is larger than B/2, it schedules the next update after 1/(1 + β) time units; and if the energy is exactly equal to B/2, it schedules the next update after 1 time unit. Then, if the sensor has no energy at its scheduled update time, it stays silent, and reschedules its following update accordingly after 1/(1−β) time units. We note that for β = 0, this energyaware status update policy transforms into a best effort uniform update policy, which is the optimal solution for the infinite battery case [38]. We also note that the constant β is chosen in [38] such that the policy is asymptotically optimal in the battery size. Specifically, it is chosen equal to z log B/B for some positive integer z that controls the policy’s asymptotic behavior. We compare our optimal policy to the energy-aware policy above for z = 2, z = 1, and z = 0 (uniform update policy) in Fig. 8. We see that it outperforms all of them. Finally, we compare the optimal policy to our recent results on an altered system model of the same problem [39]. There, the battery is fully recharged randomly over time, i.e., energy arrives in chunks of B energy units, as opposed to the incremental unit recharges considered in this work. We consider two situations of this random battery recharges to compare with. The first is when the Poisson arrival process is of unit rate, and the second is when it is of rate 1/2. The second case corresponds to an average recharge rate of B/2 = 1 energy unit per unit time, as considered in this work. From [39], the optimal long term average age for the first situation is given by r1∗ = 0.59. While the analysis in [39] is done for a Poisson arrival process of unit rate, it can be directly extended to account for that of rate 1/2; this gives the optimal long term average age for the second situation by r2∗ = 1.18, which is double r1∗ , since the average recharge rate is reduced to half. We conclude from this that while it is clearly better to have the battery recharged by 2 energy units, as opposed to only 1, every one time unit on average (r1∗ < λ∗ ), it is worse to be recharged by 2 energy units every 2 time units on average, as opposed to 1 energy unit per unit time (r2∗ > λ∗ ), although the recharge rate is the same. The latter conclusion for the second situation is due to the fact that the system with 1 energy unit recharge per unit time considered in this work gives more flexibility to the sensor on when to update compared to the system with 2 energy units recharge every 2 time units. This flexibility allows the sensor to submit updates more uniformly over time, which achieves better age by convexity of the square function that governs the areas of the triangles constituting the total area under the age curve to be minimized. V. C ONCLUSION AND F UTURE D IRECTIONS We have characterized optimal online policies for energy harvesting sensors with B-sized batteries that minimize the long term average age of information, subject to energy causality constraints. We have considered a noiseless channel where a transmission update consumes one energy unit and arrives instantaneously at the receiver. Under a Poisson energy arrival process with unit rate, energy units arrive at the sensor’s battery in an incremental fashion, i.e., one energy unit per arrival. We first have shown that the optimal status update policy has a renewal structure. Specifically, the times between the two consecutive events of submitting an update and having k energy units remaining in the battery afterwards, 0 ≤ k ≤ B − 1, are i.i.d. Then, we have thoroughly studied the specific scenario of B = 2 energy units and further shown that the optimal renewal policy has an energy-dependent threshold structure: the sensor submits an update only if the age of information surpasses a certain threshold which is a function of the energy available in its battery. From the analysis of the B = 2 case, it is amenable to show that threshold policies are also optimal for any B ≥ 3. One main difficulty in showing that is the combinatorial nature of how the different B random variables that govern the energy arrivals in between inter-updates are related. Similar to the approaches in [15]–[20], it is therefore of interest to study near-optimal renewal-type policies that provably perform within a constant gap from the optimal solution of problem (5) in future works. A PPENDIX A. Proof of Theorem 1 We prove this by showing that any given status update policy that is uniformly bounded according to Definition 1 is outperformed by a renewal policy as defined in the theorem. Let us consider the ith epoch (time between two consecutive visits to state (k, 0)); we introduce the following notation regarding the energy arrivals occurring in it. Let τ1,i denote the time until the first energy arrival after the epoch starts, and let there be j1 status updates after that energy arrival before a second energy arrival occurs. If j1 ≥ 1, then let τ2,i denote the time until the first energy arrival after the j1 th update. Otherwise, if j1 = 0, then let τ2,i denote the inter-arrival time between the first and the second energy arrivals in the epoch. Similarly, let there be j2 status updates after the second energy arrival before a third energy arrival occurs. If j2 ≥ 1, then let τ3,i denote the time until the first energy arrival after the j2 th update. Otherwise, if j2 = 0, then let τ3,i denote the interarrival time between the second and the third energy arrivals in the epoch. We continue defining τj,i ’s, j = 1, 2, . . . , until the epoch ends by retuning back to state (k, 0) again. Finally, in the event that the jth energy arrival in the epoch makes the battery full, then we wait until the first status update occurs after that event and denote by τj+1,i the time until the first energy arrival after that update, i.e., we do not account for energy arrivals that cause battery overflows. As noted before Theorem 1, there can possibly be an infinite number of updates before the system returns back to state (k, 0), depending on the energy arrival pattern and the update time decisions. For a given status update policy, one can enumerate all such patterns. For instance, following the above notation, the first pattern could be when the system goes from state (k, 0) to state (k + 1, τ1,i ) and then to state (k, 0) again; the second pattern could be when the system goes through the following sequence of states: (k, 0) − (k + 1, τ1,i ) − (k + 2, τ1,i +τ2,i )−(k+1, 0)−(k, 0); and so on. Let the vector τm,i contain all the τj,i ’s in the mth pattern. Note that this vector’s length varies with the pattern. For instance, we have τ1,i = τ1,i and τ2,i = [τ1,i , τ2,i ] for the above two pattern examples, respectively. For a given status update policy, one can also compute the probability of occurrence P of the mth pattern in ∞ the ith epoch, denoted by pm,i , with m=1 pm,i = 1. Let us also denote by Rm,i the area under the age curve in that epoch, given that it went through the mth pattern. Next, for a fixed history Hi−1 and a pattern m, let us group all the status updating sample paths that have the same τm,i and perform a statistical averaging over all of them to get the following average age in the ith epoch given that it went through the mth pattern R̂m,i (γm , Hi−1 ) , E [Rm,i |τm,i = γm , Hi−1 ] (24) Now for a given time T , let NT denote the number of epochs that have already started by time T . Then, we have E [Rm,i · 1i≤NT ] i h h i = EHi−1 Eτm,i R̂m,i (γm , Hi−1 ) · 1i≤NT Hi−1 (25) where equality follows since 1i≤NT is independent of τm,i given Hi−1 . Similarly, let xk,m,i denote the length of the ith epoch under the mth pattern, and define its (conditional) average as x̂k,m,i (γm , Hi−1 ) , E [xk,m,i |τm,i = γm , Hi−1 ] (26) Finally, we denote by Ri and xk,i the area under the age curve in the ith epoch and its length, respectively, irrespective of which pattern it went through. Next, note that by (4), the following holds ∞ ∞ 1 X 1 X r(T ) Ri 1i≤NT −1 ≤ Ri 1i≤NT ≤ T i=1 T T i=1 (31) Following similar analysis as in [38, Appendix C-1], one can show that E [RNT ] =0 (32) lim T →∞ T for any uniformly bounded policy as in Definition 1. Hence, the expected values of the upper and lower bounds in (31) are equal as T → ∞. P∞Hence, in the sequel, we derive a lower bound on T1 E [ i=1 Ri 1i≤NT ] and use the above note )] as T → ∞. to conclude that it is also a lowerPbound on E[r(T T ∞ Towards that end, note that E [ i=1 xk,i 1i≤NT ] ≥ T . Then, we have # "∞ P∞ X E [ i=1 Ri 1i≤NT ] 1 P∞ Ri 1i≤NT ≥ E (33) T E [ i=1 xk,i 1i≤NT ] i=1 We now proceed by lower bounding the right hand side of the above equation through a series of equations at the top of the next page. In there, (27) follows from (25) and the monotone convergence theorem, together with the fact P that E [Ri ] = P∞ p E [Rm,i ]; R∗ (Hi−1 ) is the minim,i m=1 ∞ p E [R̂m,i (γm ,Hi−1 )] m,i τ mum value of P∞m=1pm,i Eτ m,i[x̂k,m,i (γm ,Hi−1 )] ; and Rmin is m=1 m,i the minimum value of R∗ (Hi−1 ) over all possible epochs and their corresponding histories, i.e., the minimum over all iPand Hi−1 . This, together with the fact that E [xk,i ] = ∞ m=1 pm,i E [xk,m,i ], gives the last inequality. Observe that a policy achieving R∗ (Hi−1 ) is a policy which is a function of the possible energy arrival patterns in the ith epoch τm,i ’s only, since the history Hi−1 is fixed. Since the energy arrival process is Poisson with rate 1, it follows that the random vector τm,i consists of i.i.d. exponential random variables with parameter 1, and that {τm,i } are also independent across epochs. Therefore, if we repeat the policy that achieves Rmin over all epochs, we get a renewal policy where the epoch lengths are also i.i.d., and {li } forms a renewal process. This completes the proof. i i h hP P∞ ∞ P∞ m=1 pm,i Eτm,i R̂m,i (γm , Hi−1 ) · 1i≤NT Hi−1 i=1 EHi−1 E [ i=1 Ri 1i≤NT ] P P = (27) E[ ∞ E[ ∞ i=1 xk,i 1i≤NT ] i=1 xk,i 1i≤NT ]   P∞ P∞ P∞ m=1 pm,i Eτm,i [R̂m,i (γm ,Hi−1 )] P∞ · 1i≤NT Hi−1 i=1 EHi−1 m=1 pm,i Eτm,i [x̂k,m,i (γm , Hi−1 )] · m=1 pm,i Eτm,i [x̂k,m,i (γm ,Hi−1 )] P∞ = E [ i=1 xk,i 1i≤NT ] (28) i hP P∞ ∞ ∗ i=1 EHi−1 m=1 pm,i Eτi [x̂k,m,i (γ, Hi−1 )] · R (Hi−1 ) · 1i≤NT Hi−1 P (29) ≥ E[ ∞ i=1 xk,i 1i≤NT ] ≥ Rmin (30) B. Justification of (9) First, we have   1 2 E y1 (τ1 ) 1τ2 >y1 (τ1 )−τ1 2 Z ∞ Z ∞ 1 y1 (τ1 )2 e−τ1 e−τ2 dτ2 dτ1 = 2 τ1 =0 τ2 =y1 (τ1 )−τ1 Z ∞ 1 = y1 (τ1 )2 e−τ1 e−(y1 (τ1 )−τ1 ) dτ1 τ1 =0 2 Z ∞ 1 y1 (τ1 )2 e−y1 (τ1 ) dτ1 (34) = 2 τ1 =0 Next, let us define    1 2 1 2 ȳ2 (τ1 , τ2 ) + x1 1τ2 ≤y1 (τ1 )−τ1 α,E 2 2 (35) Since τm is independent of τ1 and τ2 for m ≥ 3, and are Pthey ∞ all i.i.d., we have that the term α gets multiplied by i=1 (1− e−x1 )i−1 e−x1 = 1 when we compute E[R]. Note that   E 1τ2 ≤y1 (τ1 )−τ1 = 1 − P [τ2 > y1 (τ1 ) − τ1 ] Z ∞ Z ∞ e−τ1 e−τ2 dτ2 dτ1 =1− τ =0 τ2 =y1 (τ1 )−τ1 Z 1∞ e−τ1 e−(y1 (τ1 )−τ1 ) dτ1 =1− Zτ1∞=0 =1− e−y1 (τ1 ) dτ1 (36) τ1 =0 and hence the term α can be expanded to   Z ∞ 1 2 −y1 (τ1 ) e dτ1 α = x1 1 − 2 τ1 =0 Z ∞ Z y1 (τ1 )−τ1 1 + ȳ2 (τ1 , τ2 )2 e−τ1 e−τ2 dτ1 dτ2 (37) 2 τ1 =0 τ2 =0 Next, let us define Z x1 1 y2 (τm )2 e−τm dτm , βm , 2 0 m≥3 (38) Now observe that, again by the fact that τi ’s are i.i.d., the terms βm ’s appear as follows when we take E[R] β3 E  1τ2 ≤y1 (τ1 )−τ1 +  = β3 E x1 ∞ X m=4 ∞ X 1τ2 ≤y1(τ1 )−τ1   = e β3 1 − Z ∞ βm E  e τ1 =0  1τ2 ≤y1 (τ1 )−τ1 i=0 dτ1 E [1τi ≤x1 ] i=3 1 − e−x1 −y1 (τ1 ) m Y 
i (39) where the second equality follows since βm is the same for all m. Equations (34), (37), and (39) yield E[R] in (9). R EFERENCES [1] J. Yang and S. Ulukus. Optimal packet scheduling in an energy harvesting communication system. IEEE Trans. Commun., 60(1):220– 230, January 2012. [2] K. Tutuncuoglu and A. Yener. Optimum transmission policies for battery limited energy harvesting nodes. IEEE Trans. Wireless Commun., 11(3):1180–1189, March 2012. [3] O. Ozel, K. Tutuncuoglu, J. Yang, S. Ulukus, and A. Yener. Transmission with energy harvesting nodes in fading wireless channels: Optimal policies. IEEE JSAC, 29(8):1732–1743, September 2011. [4] C. K. Ho and R. Zhang. Optimal energy allocation for wireless communications with energy harvesting constraints. IEEE Trans. Signal Process., 60(9):4808–4818, September 2012. [5] J. Yang, O. Ozel, and S. Ulukus. Broadcasting with an energy harvesting rechargeable transmitter. IEEE Trans. Wireless Commun., 11(2):571– 583, February 2012. [6] O. Ozel, J. Yang, and S. Ulukus. Optimal broadcast scheduling for an energy harvesting rechargebale transmitter with a finite capacity battery. IEEE Trans. Wireless Commun., 11(6):2193–2203, June 2012. [7] M. A. Antepli, E. Uysal-Biyikoglu, and H. Erkal. Optimal packet scheduling on an energy harvesting broadcast link. IEEE JSAC, 29(8):1721–1731, September 2011. [8] J. Yang and S. Ulukus. Optimal packet scheduling in a multiple access channel with energy harvesting transmitters. Journal of Commun. Networks, 14(2):140–150, April 2012. [9] K. Tutuncuoglu and A. Yener. Sum-rate optimal power policies for energy harvesting transmitters in an interference channel. Journal Commun. Networks, 14(2):151–161, April 2012. [10] C. Huang, R. Zhang, and S. Cui. Throughput maximization for the Gaussian relay channel with energy harvesting constraints. IEEE JSAC, 31(8):1469–1479, August 2013. [11] D. Gunduz and B. Devillers. Two-hop communication with energy harvesting. In Proc. IEEE CAMSAP, December 2011. [12] B. Gurakan and S. Ulukus. Cooperative diamond channel with energy harvesting nodes. IEEE JSAC, 34(5):1604–1617, May 2016. [13] B. Varan and A. Yener. Delay constrained energy harvesting networks with limited energy and data storage. IEEE JSAC, 34(5):1550–1564, May 2016. [14] A. Arafa, A. Baknina, and S. Ulukus. Energy harvesting two-way channels with decoding and processing costs. IEEE Trans. Green Commun. and Networking, 1(1):3–16, March 2017. [15] D. Shaviv and A. Ozgur. Universally near optimal online power control for energy harvesting nodes. IEEE JSAC, 34(12):3620–3631, December 2016. [16] H. A. Inan and A. Ozgur. Online power control for the energy harvesting multiple access channel. In Proc. WiOpt, May 2016. [17] A. Baknina and S. Ulukus. Energy harvesting multiple access channels: Optimal and near-optimal online policies. IEEE Trans. Commun. To appear. [18] A. Baknina and S. Ulukus. Optimal and near-optimal online strategies for energy harvesting broadcast channels. IEEE JSAC, 34(12):3696– 3708, December 2016. [19] A. Baknina and S. Ulukus. Online scheduling for energy harvesting channels with processing costs. IEEE Trans. Green Commun. and Networking, 1(3):281–293, September 2017. [20] A. Arafa, A. Baknina, and S. Ulukus. Energy harvesting networks with general utility functions: Near optimal online policies. In Proc. IEEE ISIT, June 2017. [21] S. Kaul, R. Yates, and M. Gruteser. Real-time status: How often should one update? In Proc. IEEE Infocom, March 2012. [22] R. Yates and S. Kaul. Real-time status updating: Multiple sources. In Proc. IEEE ISIT, July 2012. [23] C. Kam, S. Kompella, and A. Ephremides. Age of information under random updates. In Proc. IEEE ISIT, July 2013. [24] M. Costa, M. Codreanu, and A. Ephremides. On the age of information in status update systems with packet management. IEEE Trans. Inf. Theory, 62(4):1897–1910, April 2016. [25] A. Kosta, N. Pappas, A. Ephremides, and V. Angelakis. Age and value of information: Non-linear age case. In Proc. IEEE ISIT, June 2017. [26] Y. Sun, E. Uysal-Biyikoglu, R. Yates, C. E. Koksal, and N. B. Shroff. Update or wait: How to keep your data fresh. IEEE Trans. Inf. Theory, 63(11):7492–7508, November 2017. [27] A. M. Bedewy, Y. Sun, and N. B. Shroff. Age-optimal information updates in multihop networks. In Proc. IEEE ISIT, June, 2017. [28] Y. Hsu, E. Modiano, and L. Duan. Age of information: Design and analysis of optimal scheduling algorithms. In Proc. IEEE ISIT, June 2017. [29] J. Zhong, E. Soljanin, and R. D. Yates. Status updates through multicast networks. In Proc. Allerton, October 2017. [30] E. Najm and E. Telatar. Status updates in a multi-stream M/G/1/1 preemptive queue. Available Online: arXiv1801.04068. [31] R. Yates, E. Najm, E. Soljanin, and J. Zhong. Timely updates over an erasure channel. In Proc. IEEE ISIT, June 2017. [32] R. D. Yates, P. Ciblat, A. Yener, and M. A. Wigger. Age-optimal constrained cache updating. In Proc. IEEE ISIT, June 2017. [33] R. D. Yates. Lazy is timely: Status updates by an energy harvesting source. In Proc. IEEE ISIT, June 2015. [34] B. T. Bacinoglu, E. T. Ceran, and E. Uysal-Biyikoglu. Age of information under energy replenishment constraints. In Proc. ITA, February 2015. [35] A. Arafa and S. Ulukus. Age-minimal transmission in energy harvesting two-hop networks. In Proc. IEEE Globecom, December 2017. [36] A. Arafa and S. Ulukus. Age minimization in energy harvesting communications: Energy-controlled delays. In Proc. Asilomar, October 2017. [37] B. T. Bacinoglu and E. Uysal-Biyikoglu. Scheduling status updates to minimize age of information with an energy harvesting sensor. In Proc. IEEE ISIT, June 2017. [38] X. Wu, J. Yang, and J. Wu. Optimal status update for age of information minimization with an energy harvesting source. IEEE Trans. Green Commun. and Networking. To appear. [39] A. Arafa, J. Yang, and S. Ulukus. Age-minimal online policies for energy harvesting sensors with random battery recharges. In Proc. IEEE ICC, May 2018. [40] A. Baknina, O. Ozel, J. Yang, S. Ulukus, and A. Yener. Sending information through status updates. Available Online: arXiv:1801.04907. [41] S. M. Ross. Stochastic Processes. Wiley, 1996. [42] Y. Sun, Y. Polyanskiy, and E. Uysal-Biyikoglu. Remote estimation of the wiener process over a channel with random delay. In Proc. IEEE ISIT, June 2017. Longer version available: arXiv:1701.06734. [43] S. P. Boyd and L. Vandenberghe. Convex Optimization. Cambridge University Press, 2004. 

7cs.IT

COMPUTATIONAL POWER AND THE SOCIAL IMPACT OF ARTIFICIAL INTELLIGENCE Tim Hwang1 Machine learning is a computational process. To that end, it is inextricably tied to computational power - the tangible material of chips and semiconductors that the algorithms of machine intelligence operate on. Most obviously, computational power and computing architectures shape the speed of training and inference in machine learning, and therefore influence the rate of progress in the technology. But, these relationships are more nuanced than that: hardware shapes the methods used by researchers and engineers in the design and development of machine learning models. Characteristics such as the power consumption of chips also define where and how machine learning can be used in the real world. In a broader perspective, computational power is also important because of its specific geographies. Semiconductors are designed, fabricated, and deployed through a complex international supply chain. Market structure and competition among companies in this space influence the progress of machine learning. Moreover, since these supply chains are also considered significant from a national security perspective, hardware becomes an arena in which government industrial and trade policy has a direct impact on the fundamental machinery necessary for artificial intelligence (AI). This paper aims to dig more deeply into the relationship between computational power and the development of machine learning. Specifically, it examines how changes in computing architectures, machine learning methodologies, and supply chains might influence the future of AI. In doing so, it seeks to trace a set of specific relationships between this underlying hardware layer and the broader social impacts and risks around AI. On one hand, this examination shines a spotlight on how hardware works to exacerbate a range of concerns around ubiquitous surveillance, technological unemployment, and geopolitical conflict. On the other, it also highlights the potentially significant role that shaping the development of computing power might play in addressing these concerns. *** 1 Research Affiliate, MIT Media Lab. 1 Introduction 1 Part I Machine Learning and Computational Power The Tropes of Computational Power 6 7 Part II Specialization in Computational Power Impact: The Geography of Training and Inference 14 14 Part III Supply Chain and Computational Power Impact: The Geopolitics of Machine Learning 18 32 Part IV Computational Power as Data Substitute Impact: Computational Power and Economic Impact 36 36 Conclusion: Prospects for Governance 43 *** INTRODUCTION Machine learning is a computational process. To that end, it is inextricably tied to computational power - the tangible material of chips and semiconductors that the algorithms of machine intelligence operate on. Most obviously, computational power and computing architectures shape the speed of training and inference in machine learning, and therefore influence the rate of progress in the technology. But, these relationships are more nuanced than that: hardware shapes the methods used by researchers and engineers in the design and development of machine learning models. Characteristics such as the power consumption of chips also define where and how machine learning can be used in the real world. In a broader perspective, computational power is also important because of its specific geographies. Semiconductors are designed, fabricated, and deployed through a complex international supply chain. Market structure and competition among companies in this space influence the progress of machine learning. Moreover, since these supply chains are also considered significant from a national security perspective, hardware 2 becomes an arena in which government industrial and trade policy has a direct impact on the fundamental machinery necessary for artificial intelligence (AI). Despite this, many analyses of the social impact of the current wave of progress in AI have not substantively brought the dimension of hardware into their accounts. While a common trope in both the popular press and scholarly literature is to highlight the massive increase in computational power that has enabled the recent breakthroughs in machine learning, the analysis frequently goes no further than this observation around magnitude. This paper aims to dig more deeply into the relationship between computational power and the development of machine learning. Specifically, it examines how changes in computing architectures, machine learning methodologies, and supply chains might influence the future of AI. In doing so, it seeks to trace a set of specific relationships between this underlying hardware layer and the broader social impacts and risks around AI. On one hand, this examination shines a spotlight on how hardware works to exacerbate a range of concerns around ubiquitous surveillance, technological unemployment, and geopolitical conflict. On the other, it also highlights the potentially significant role that shaping the development of computing power might play in addressing these concerns. Part I will examine the role that computational power has played in the progress of machine learning, arguing that its impact has been somewhat flattened in recent accounts looking at the social impact of the technology. Part II will look at trends towards increasing specialization in the hardware used for machine learning, and its implications for control and privacy in the space. Part III will look at the semiconductor supply chain, and its implications for the geopolitics of machine learning. Part IV will examine research developments changing the balance between data and computational power in the workflow of machine learning, and its influence on the economic impact of the technology. It will then conclude with some remarks on the potential role of hardware as a lever for policy action in the space. PART I: MACHINE LEARNING AND COMPUTATIONAL POWER 3 AI has historically moved through multiple cycles of progress and optimism followed by setbacks and pessimism, so called “AI winters”.2 Present-day excitement around AI, and more specifically the recent breakthroughs in the subfield of machine learning, represent only the latest upswing in this historical pattern. Machine learning itself, the study of algorithms which improve themselves through data, is not a new domain of research. The fundamentals underlying the modern advances in the field were established by researchers in the 1950s and developed throughout the subsequent decades.3 However, neural networks - the specific technique of machine learning driving much of the commercial interest in AI today - were still considered a niche area of research only until relatively recently. As one popular account has put it, “for much of its history most computer scientists saw it [neural networks] as vaguely disreputable, even mystical.”4 It was recognized early in this history that the neural networks proposed during the 1950s and 1960s were limited by the comparatively minimal processing power available at the time.5 The continued growth of computational power, along with the accumulation of large datasets during the 1990s and 2000s, played a major role in revitalizing progress in neural networks and motivating significant investment within the field of AI more broadly. The field of computer vision, which focuses on advancing the ability for machines to extract understanding from images and video, offers one representative example on this point. “Traditional” approaches to these tasks in the 1990s and early 2000s focused on algorithms which specified a set of defined features that would be used to process and classify an image.6 For an account of this history, see generally John Markoff, Machines of Loving Grace: The Quest for Common Ground Between Humans and Robots (2015). 3 See, e.g., B. Farley & W. Clark, Simulation of self-organizing systems by digital computer, 4 Transactions of the IRE Professional Group on Information Theory 76–84 (1954). 4 Gideon Lewis-Kraus, The Great A.I. Awakening, The New York Times, December 14, 2016, https://www.nytimes.com/2016/12/14/magazine/the-great-ai-awakening.html (last visited Mar 20, 2018). 5 See Marvin Minsky & Seymour A. Papert, Perceptrons: An Introduction to Computational Geometry (1969) (noting these limitations). 6 See, e.g., D.G. Lowe, Object recognition from local scale-invariant features, 1150–1157 vol.2 (1999), http://ieeexplore.ieee.org/document/790410/ (last visited Mar 20, 2018) (describing the SIFT algorithm, one representative approach). 2 4 Neural networks, in contrast, learn the relevant features for classification rather than having them pre-designed into the algorithm.7 Expanding computational power and the availability of data changed the practice of computer vision. On the data front, the growth of the consumer web produced a vast library of images for machine learning systems to train on. ImageNet, an annotated dataset of 14 million images in 20 thousand categories assembled by workers on the Amazon Mechanical Turk platform, provided a common dataset for researchers to work with.8 Computational power also continued to increase in the 2000s, rising from 37 million transistors per chip in 2000 to 2.3 billion transistors per chip by 2009.9 This was augmented by the finding that a particular kind of computational architecture - the GPU - was particularly well-suited as a platform for neural networks, a development discussed in more detail in Part II.10 Leveraging both of these assets, neural networks were able to significantly surpass the performance of earlier techniques in the space.11 The ImageNet Large Scale Visual Recognition Challenge is a commonly cited marker of this transition. Hosted since 2010, the Challenge brings together researchers to compete in designing systems to solve a set of visual recognition tasks.12 Performance in 2010 and 2011, which featured teams using traditional techniques in the space, were never able to reduce error rates below 25%, with most teams showing much higher rates of error.13 See, e.g., Alex Krizhevsky, Ilya Sutskever & Geoffrey E. Hinton, ImageNet classification with deep convolutional neural networks, 60 Communications of the ACM 84–90 (2017) (demonstrating this learned feature approach). 8 Dave Gershgorn, The data that transformed AI research—and possibly the world Quartz, https://qz.com/1034972/the-data-that-changed-the-direction-of-ai-research-and-possiblythe-world/ (last visited Mar 20, 2018). 9 Moore’s Law: Transistors per microprocessor, Our World in Data, https://ourworldindata.org/grapher/transistors-per-microprocessor (last visited Mar 20, 2018). 10 See Rajat Raina, Anand Madhavan & Andrew Y. Ng, Large-scale deep unsupervised learning using graphics processors, 1–8 (2009), http://portal.acm.org/citation.cfm?doid=1553374.1553486 (last visited Mar 20, 2018). 11 Cf. From not working to neural networking, The Economist, 2016, https://www.economist.com/news/special-report/21700756-artificial-intelligence-boombased-old-idea-modern-twist-not (last visited Mar 20, 2018). 12 ImageNet Large Scale Visual Recognition Challenge (ILSVRC), http://www.imagenet.org/challenges/LSVRC/ (last visited Mar 20, 2018). 13 Id. 7 5 AlexNet, a system submitted to the competition by researchers Geoffrey Hinton, Ilya Sutskever, and Alex Krizhevsky in 2012, was both the first entry to apply neural networks in the Challenge and the first to achieve a below 25% error rate.14 The excitement around these results, and the margin of improvement over established techniques, led one researcher to state that the “Imagenet 2012 event was definitely what triggered the big explosion of AI today”.15 Computational power has for this reason been fundamental to the present-day breakthroughs in machine learning. Even if the necessary data been widely available at an earlier point historically, a lack of computational power would have effectively prevented neural networks from achieving their current level of performance. The Tropes of Computational Power The narrative of computational power and machine learning typically ends here. Recent pieces examining these technological breakthroughs have often focused more on the implications of what the technology can do, rather than the implications of how it is being done.16 To the extent that computational power is mentioned, it is typically addressed simply as an enabling factor in the emergence of machine learning. The prevailing attribute highlighted in these accounts has tended to be one of magnitude: the processing power of the chips running machine learning have been seen as their primary contribution. The 2016 White House report Preparing for the Future of Artificial Intelligence is illustrative. The paper focuses on computational power only in passing as one of the three factors enabling the present-day breakthroughs in machine learning. “[T]he availability of big data…dramatically improved machine learning approaches and algorithms…the capabilities of more powerful computers.”17 An Id. See supra note 8 (for a visualization of these results). 16 See, e.g., JURI Committee, European Civil Law Rules in Robotics (2016), available at http://www.europarl.europa.eu/RegData/etudes/STUD/2016/571379/IPOL_STU(2016)571 379_EN.pdf; Urs Gasser, AI and the Law: Setting the Stage, Medium (2017), https://medium.com/berkman-klein-center/ai-and-the-law-setting-the-stage-48516fda1b11 (last visited Mar 20, 2018). 17 White House National Science and Technology Council, Preparing for the Future of Artificial Intelligence 6 (2016), available at 14 15 6 accompanying paper released at the same time, The National Artificial Intelligence Research and Development Plan, highlights improved hardware for machine learning as a priority, but only to the extent that chips with higher levels of performance are needed to drive the technology forwards.18 Other reports from the European Union, civil society groups, and researchers on the topic of AI have followed a similar set of themes when considering the role of computational power.19 This narrative perhaps leaves out an important part of the story. Such a shorthand enables a focus on the numerous problematic ways that machine learning might be applied and the implications of those applications for justice, equity, and a host of other values. However, it also flattens out the role that computational power plays in these issues to simply that of a trigger for technological progress. This may miss the significant and nuanced ways that hardware influences the impact of AI on these broader values and social concerns. Computational power does more than simply make the present-day breakthroughs in machine learning possible. The medium is a significant message here: hardware actively shapes the landscape of what can be done with the technology of machine learning, and plays a significant role in influencing how it will evolve going forwards. The contours of computational power play a role in defining who has control over and access to the benefits of machine learning, and the actors that will play a role in its governance. It plays a role in the politics of the technology, both at the level of an individual citizen and in the broader competition between states. Computational power, in defining the speed at which machine learning models may trained and experimented on, shapes the speed at which the technology advances and therefore serves to define https://obamawhitehouse.archives.gov/sites/default/files/whitehouse_files/microsites/ostp/ NSTC/preparing_for_the_future_of_ai.pdf. 18 White House National Science and Technology Council, The National Artificial Intelligence Research and Development Strategic Plan 21 (2016). 19 See e.g., Royal Society (Great Britain), Machine learning: the power and promise of computers that learn by example (2017) available at https://royalsociety.org/~/media/policy/projects/machine-learning/publications/machinelearning-report.pdf; House of Commons Science and Technology Committee, Robotics and Artificial Intelligence (2016), https://publications.parliament.uk/pa/cm201617/cmselect/cmsctech/145/145.pdf; World Economic Forum, Assessing the Risk of Artificial Intelligence, Global Risks Report 2017, http://wef.ch/2izSQRP (last visited Mar 22, 2018); David Bollier, Artificial Intelligence: The Great Disruptor, Aspen Institute (2018) available at http://csreports.aspeninstitute.org/documents/AI2017.pdf. 7 its broader economic impact. These impacts turn on more than simply the amount of processing power available, but on the details of computational architecture, supply chains, and the co-evolution of the machine learning field itself. To that end, the evolving research and commercial ecosystem around hardware is more than just a sideshow: shifts in these underlying technologies have a significant place in understanding the impact of AI on society as a whole. Parts II, III, and IV work to map these many connections by drawing the lines from the changing landscape of computational power to the bigger social challenges surrounding AI. PART II: SPECIALIZATION IN COMPUTATIONAL POWER Computational power is not a simple matter of magnitude. The specific architecture of a chip plays a major role in determining whether or not it is effective in dealing with a given computational task. In general, the industry has tended towards increasingly specialized platforms for machine learning as the field continues to grow and attract commercial interest. In this sense, hardware has moved in a direction opposite to software: a shift towards narrower specialization in chips has proceeded even as the research field has been focused on building ever more general learning systems. Two dynamics shape this marketplace for machine learning hardware. One is an inverse relationship between performance and flexibility.20 While general purpose computing power can take on a wide range of tasks and can be easily configured to take on new tasks, it tends to be outpaced by hardware which is built for a specific purpose.21 However, this increased performance comes at a cost: specialized hardware accommodates a relatively smaller set of use cases and has an architecture which is less easy to change after it is deployed.22 These specialized platforms are also frequently more expensive than commodified general platforms.23 One overarching question is whether the demand for machine learning driven See generally Inside the Microsoft FPGA-based configurable cloud, Channel 9, https://channel9.msdn.com/Events/Build/2017/B8063 (last visited Mar 20, 2018) (discussing these trade-offs). 21 See Griffin Lacey, Graham W. Taylor & Shawki Areibi, Deep Learning on FPGAs: Past, Present, and Future, arXiv:1602.04283, 6 (2016), http://arxiv.org/abs/1602.04283 (last visited Feb 13, 2018) (discussing cost differentials). 22 Id. 23 Id. 20 8 products and the research community will tend over time to favor architectures that are more or less flexible given this tradeoff. A second important dynamic is that the hardware for training a machine learning model to accomplish a task can differ significantly from the hardware used to conduct inference with an already trained model. This is due to the different demands at each step of the machine learning workflow. For instance, energy consumption may matter for a computer vision system operating on a mobile device, though it may not matter when that computer vision system is being trained initially in a data center.24 Latency - the time delay between input and output of a system - might be a significant factor in a high-speed navigation context, where speed of inference would reduce the time needed for a course correction.25 However, as with energy consumption, latency may not prove to be as significant when the navigation system is being trained. These considerations influence what kinds of hardware are used at which points in the lifecycle of a machine learning system. They can be viewed as separate though overlapping markets, with hardware platforms being offered either for training or inference, and some offering support for both.26 Background: CPUs to GPUs Graphics processing units (GPUs) form the present-day backbone of the machine learning workflow.27 GPUs are the primary platform for both training and inference, and are widely used both for basic research and in the practical development and deployment of machine learning driven products in the marketplace.28 The outsize role that GPUs play in machine learning is the result of an unexpected historical convergence. As their name suggests, GPUs were 24 See Vivienne Sze et al., Efficient Processing of Deep Neural Networks: A Tutorial and Survey, arXiv:1703.09039, 5-6 (2017), http://arxiv.org/abs/1703.09039 (last visited Feb 13, 2018). 25 Id at 26. 26 See, e.g., Jeff Dean and Urs Hölzle, Build and train machine learning models on our new Google Cloud TPUs, Google (2017), https://www.blog.google/topics/google-cloud/googlecloud-offer-tpus-machine-learning/ (last visited Mar 20, 2018) (earlier generations of Google’s specialized machine learning chips were focused on inference, with the latest version supporting both inference and training). 27 See Deloitte, Hitting the Accelerator: The Next Generation of Machine-Leaning Chips (2017), available at https://www2.deloitte.com/content/dam/Deloitte/global/Images/infographics/technologyme diatelecommunications/gx-deloitte-tmt-2018-nextgen-machine-learning-report.pdf. 28 Id. at 1. 9 originally designed to support computer graphics and image processing applications.29 To accomplish this, GPUs feature an architecture which distributes computational tasks across a large number of cores to be processed in parallel.30 This is in contrast to central processing units (CPUs), which feature a smaller number of more powerful cores that are optimized for handling just a few tasks simultaneously.31 This parallel architecture allows the GPU to be uniquely well suited for machine learning applications. At its root, neural network training and inference relies on the execution of a large number of identical matrix multiplication calculations.32 This uniformity enables these operations to be parallelized and distributed across the many cores offered by the GPU.33 This enables the GPU to outperform CPU architectures which have comparatively more powerful processors but manage tasks in a more serial format.34 Increasing Specialization: FPGAs and ASICs The repurposing of GPUs as the primary hardware platform for machine learning reflects a selection among available technologies. As interest in machine learning has continued to grow, so has the notion of developing hardware entirely purpose-built for these applications become more attractive. Discussion within the industry has focused on the possibility of using field-programmable gate arrays (FPGAs) and application-specific integrated circuits (ASICs) as the next primary platforms for machine learning.35 FPGAs are distinct from CPUs and GPUs in that they do not run programs in stored memory. Instead, they are collection of standardized “logic blocks” whose relationships can be configured by a programmer Id. at 2. See supra note 10. 31 Id. 32 See supra note 22 at 12. 33 See supra note 10. 34 See Cade Metz, The Race To Build An AI Chip For Everything Just Got Real, WIRED, https://www.wired.com/2017/04/race-make-ai-chips-everything-heating-fast/ (last visited Feb 13, 2018) (describing some of the limitations of the CPU). 35 See, e.g., Karl Freund, Will ASIC Chips Become The Next Big Thing In AI?, Forbes, https://www.forbes.com/sites/moorinsights/2017/08/04/will-asic-chips-become-the-nextbig-thing-in-ai/ (last visited Mar 20, 2018). 29 30 10 once the chip is received from a manufacturer.36 ASICs are purpose-built chip boards which are specific to a purpose and cannot be easily reconfigured after they are manufactured.37 FPGAs and ASICs are particularly attractive in the context of machine learning inference. Both devices consume less energy than CPUs and GPUs, and their specialization allows for greater speed.38 These gains come at the loss of flexibility and an increased cost. FPGAs and ASICs cannot be as easily and quickly configured to run a wide range of tasks. Both are relatively more expensive when compared with CPUs and GPUs. This is particularly the case with ASICs, which are “bespoke” projects that are expensive and time-consuming to produce. This makes them costeffective as a platform only in significant quantities.39 However, these benefits may outweigh the costs, particularly in circumstances where a specific kind of machine learning inference is reliably needed at a mass scale. Project Catapult, an initiative launched by Microsoft, has shown high performance for FPGAs as the core computing unit in their data centers.40 FPGAs and ASICs have also been considered a promising approach in the autonomous vehicles context, where the tasks a machine learning system will need to take on will be relatively stable and where chips will be needed for a large number of vehicles.41 36 See What is an FPGA? Field Programmable Gate Array, https://www.xilinx.com/products/silicon-devices/fpga/what-is-an-fpga.html (last visited Mar 20, 2018). 37 Jeff Dean, Machine Learning for Systems and Systems for Machine Learning, NIPS 2017, available at http://learningsys.org/nips17/assets/slides/dean-nips17.pdf (noting the design challenges with ASICs). 38 See supra notes 19-20. 39 Id. 40 See Kalin Ovtcharov et al., Accelerating deep convolutional neural networks using specialized hardware, Microsoft Research Whitepaper, 2 (2015), available at https://www.microsoft.com/en-us/research/wpcontent/uploads/2016/02/CNN20Whitepaper.pdf. 41 See, e.g., ASIC might be the mainstream chip for autonomous driving, a chance for Chinese start-ups, VehicleTrend, https://www.vehicle-trend.com/Knowledge/201801081131.html (last visited Mar 21, 2018); Phil Kalaf, Self-Driving Cars, Wireless Data? It’s Time to Thank the Humble FPGA. IDS (2017), http://www.idsforward.com/wireless-datathank-humble-fpga/ (last visited Mar 21, 2018); Yu Wang et al, Reconfigurable Processor for Deep Learning in Autonomous Vehicles, https://www.itu.int/en/journal/001/Documents/itu2017-2.pdf (2017); Harsh Chauhan, Can Intel Dominate This Market by Overcoming This Smaller Rival? The Motley Fool (2017), https://www.fool.com/investing/2017/11/24/can-intel-dominate-this-market-by- 11 While FPGAs and ASICs seem to show promise in the inference context, they have traditionally had some limitations that have made them less attractive as platforms for training. For one, FPGAs and ASICs have tended to be less accurate, relying on “fixed point” computation or featuring comparatively limited floating-point performance.42 This has made them comparatively limited in performing the accurate level of calculation needed in the training process. These devices have also had limited external memory bandwidth, preventing them from efficiently conducting the matrix multiplication needed for training.43 However, the potential speed and energy consumption gains presented by FPGA and ASIC architectures have encouraged research which appears to be eliminating some of these limitations over time. In 2017, Intel researchers released software which maximizes data reuse and minimizes external memory bandwidth to boost training performance on FPGAs.44 The latest generation of Google’s “Tensor Processing Unit” (TPU), a specialized ASIC, supports both training and inference.45 The claimed improvements are quite significant. One recent talk from Google in 2017 noted that the TPU ASIC was able to execute training tasks at ten to fourteen times the speed of their previous production setups with a relatively smaller number of machines.46 Moving Forwards It remains unclear whether or not more specialized, less flexible hardware will unseat the preeminent place of the GPU in machine learning training and inference. Producers of FPGAs and ASICs are releasing performance benchmarks showing significant improvements over the GPU for both kinds of tasks. Google claims that its TPU ASIC is able to conduct overcoming-this.aspx (last visited Feb 20, 2018) (noting the application of FPGAs in the autonomous vehicles context). 42 See Brian Bailey, Machine Learning’s Growing Divide, Semiconductor Engineering, https://semiengineering.com/machine-learnings-growing-divide/ (last visited Mar 12, 2018) (reviewing these issues); Understanding Peak Floating-Point Performance Claims, https://www.altera.com/en_US/pdfs/literature/wp/wp-01222-understanding-peak-floatingpoint-performance-claims.pdf (last visited Feb 15, 2018) (reviewing the floating point issues in more depth). 43 Id. 44 Utku Aydonat et al., An OpenCL(TM) Deep Learning Accelerator on Arria 10, arXiv:1701.03534 [cs] (2017), http://arxiv.org/abs/1701.03534 (last visited Mar 12, 2018). 45 See supra note 34. 46 See supra note 34 (describing these improvements). 12 inference fifteen to thirty times faster than contemporary GPUs and CPUs.47 In a similar vein, Graphcore - one prominent startup focusing on specialized machine learning hardware - claims that eight of its proprietary “IPU” cards are equivalent to 128 contemporary GPUs.48 Despite this, benchmarking issues persist and it is challenging to evaluate these claims in a systematic way.49 For its part, GPU leader Nvidia has challenged the performance claimed by Google of its TPUs, noting that it failed to compare its chips against its latest generation of hardware.50 At the moment, the semiconductor industry does not yet have a common scheme for evaluating the performance of machine learning specialized hardware as it does in the CPU space.51 This is significant because the specific architecture of the neural network and how it is implemented can have a significant impact on reported performance.52 Though some commentators have framed the industry choices between GPUs, FPGAs, and ASICs as a mutually exclusive ones, it is not clear that this will be the case in practice.53 Even if FPGA and ASIC designs do not ultimately become a new standard for training and inference in machine learning writ large, it seems likely that they will become a natural option for certain applications of machine learning systems, particularly in the consumer products context. Recent moves by industry leaders seems to recognize this reality. Despite its leadership and championing of a GPU-focused model, Nvidia’s latest Drive PX product features a specialized “Deep Learning Accelerator” (DLA) module as it Norman P. Jouppi et al, In-Datacenter Performance Analysis of a Tensor Processing Unit, https://arxiv.org/ftp/arxiv/papers/1704/1704.04760.pdf (last visited Feb 15, 2018). 48 Graphcore Benchmarks, Presentation at NIPS 2017, available at https://cdn2.hubspot.net/hubfs/729091/NIPS2017/NIPS%2017%20%20benchmarks%20final.pdf?t=1521107772551. 49 See supra note 22, at 26-27 (describing the many influences on chip performance). 50 Jensen Huang, AI Drives the Rise of Accelerated Computing in Data Centers, Nvidia Blog, https://blogs.nvidia.com/blog/2017/04/10/ai-drives-rise-accelerated-computingdatacenter/ (Apr 10, 2017). 51 See, e.g., SPEC CPU 2017, https://www.spec.org/cpu2017/. 52 See supra note 22, at 26-27. 53 See, e.g., FPGA Based Deep Learning Accelerators Take on ASICs, The Next Platform (2016), https://www.nextplatform.com/2016/08/23/fpga-based-deep-learning-acceleratorstake-asics/ (last visited Feb 13, 2018); Does the future lie with CPU+GPU or CPU+FPGA?, Scientific Computing World, https://www.scientificcomputing.com/news/analysis-opinion/does-future-lie-cpugpu-or-cpufpga (last visited Feb 20, 2018). 47 13 attempts to cater to applications in autonomous vehicles.54 Nvidia has also open-sourced its designs for the DLA, a move likely to drive down the cost of this specialized hardware going forwards by enabling others to manufacture the same designs.55 These trends provide a framework for thinking about the economics of various machine learning applications, and how and where the technology might be used in practice. To that end, it begs the question of how these shifts in hardware specialization might influence the overall impact of machine learning on society and the governance of the technology. Impact: The Geography of Training and Inference Machine learning is not an abstract force, but a computational task that takes place somewhere. Hardware capabilities and the particular economics of processors are critical since they define the spatial dimensions of machine learning and what it likely to be applied towards. Power consumption defines whether or not machine learning computation can be done on a small, mobile device, or must have access to a reliable and continuous power source. High energy costs limit the ability to embed machine learning systems directly on a device. In the very least, it limits the application of machine learning to situations with sufficient connectivity for a device to communicate with a larger pool of computational power hosted in the “cloud.”56 Latency is also crucial in this respect. Even if a chip is able to operate at low power on a mobile device, it may be ineffectual for a particular use in the field if it is insufficiently responsive for the intended purpose. This acts as a bar to certain real-time or mission-critical uses of machine learning where an alternative cloud architecture would also produce similarly unacceptable levels of latency. Karl Freund, Why Nvidia is Building Its Own TPU, Forbes, https://www.forbes.com/sites/moorinsights/2017/05/15/why-nvidia-is-building-its-own-tpu (last visited Mar 20, 2017). 55 See Tom Simonite, To Compete With New Rivals, Chipmaker Nvidia Shares Its Secrets, WIRED, Sept 29, 2017, https://www.wired.com/story/to-compete-with-new-rivalschipmaker-nvidia-shares-its-secrets/ (last visited Feb 20, 2018). 56 See supra note 22, at 5-6 (discussing these different configurations). 54 14 Power consumption and latency are barriers to the application of machine learning within certain domains. This includes the placement of these systems on small devices with an untethered power source, and in low bandwidth situations with poor connectivity. The rise of ASICs and FPGAs - as well as ongoing improvements to GPUs - suggest that machine learning hardware will erode these limitations on the placement of machine learning systems over time, particularly for mass produced consumer products that have economies of scale. The Geography of Inference These developments offer a mixed blessing to those concerned about the harmful possibilities of machine learning inference. Machine learning can increasingly be integrated into a range of different products and services and used in situations where it was previously considered impractical to do so. For civil libertarians, FPGAs and ASICs enable the expansion of machine learning as a means of conducting surveillance: small, low power devices can now incorporate the advances of computer vision to recognize people and objects even in areas with low bandwidth. For those worried about the misuse of machine learning by “bad actors”, specialized hardware makes it more possible to benefit from the technology without reliance on cloud services where harmful activity might be more easily tracked and halted.57 Also concerning is the fact that the more inflexible architectures of FPGAs and ASICs might potentially make it more challenging to repair machine learning systems when flaws are discovered. A growing body of research continues to highlight the point both that machine learning systems frequently can render biased, discriminatory results, and are potentially vulnerable to malicious manipulation.58 Where a trained machine learning model is “hard wired” into a chip, the discovery that it has these flaws may make repair a more expensive and protracted process as it requires a For a review of these concerns, see, e.g., Miles Brundage et al, The Malicious Use of Artificial Intelligence (2018), available at https://maliciousaireport.com/. 58 See, e.g., Joy Buolamwini & Timnit Gebru, Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification, PMLR 81:1–15, 2018, available at http://proceedings.mlr.press/v81/buolamwini18a/buolamwini18a.pdf; Solon Barocas & Andrew D. Selbst, Big Data’s Disparate Impact (2016), https://papers.ssrn.com/abstract=2477899 (last visited Mar 21, 2018); Ian Goodfellow et al, Attacking Machine Learning with Adversarial Examples, OpenAI Blog (2017), https://blog.openai.com/adversarial-example-research/ (last visited Mar 21, 2018); Tom B. Brown et al., Adversarial Patch, arXiv:1712.09665 [cs] (2017), http://arxiv.org/abs/1712.09665 (last visited Mar 21, 2018). 57 15 replacement of the processor itself, rather than modification of software. This problem applies with particular force in an “embedded” setting where chips are sold and distributed with a product and there is no centralized means of changing their behavior once they have left the factory. At the same time, FPGAs and ASICs also raise the possibility that machine learning may be architected in a more robustly privacy-protecting ways going forwards. Since specialized computing power enables machine learning inference to be done on the device itself, it also opens the possibility that machine learning capabilities might be provided without ever having personal data leave a device. Consider a computer vision system which helps users quickly sort through their photos to find friends and family members. At present, the energy costs of inference might require that these photos be uploaded from a smartphone to a central server to be processed and tagged.59 Specialized hardware might enable an alternative architecture in which the machine learning model is embedded on the smartphone itself, such that the photos themselves do not need to be shared with a third-party to be analyzed by the system.60 This remains up the air as FPGAs and ASICs for machine learning enter the scene and attempt to find viable niches in the application of the technology. The economics of these hardware platforms will influence the viability of alternative architectures in various markets, and in doing so will inform whether advocates and policymakers are able to argue for the feasibility of more privacy sensitive approaches to machine learning going forwards. The Geography of Training It is important to recognize that the geography of training may look quite different from the geography of inference.61 As discussed above, FPGAs and ASICs have been traditionally somewhat limited as platforms for the training of machine learning systems. While the possibilities of using specialized hardware for training continue to be developed by Google See supra note 22 at 5-6. See, e.g., Ben Popper, Google’s new Clips camera is invasive, creepy, and perfect for a parent like me The Verge (2017), https://www.theverge.com/2017/10/5/16428708/googleclips-camera-privacy-parents-children (last visited Mar 21, 2018) (describing an architecture along these lines). 61 Nvidia, GPU-Based Deep Learning Inference: A Performance and Power Analysis, 4-5 (2015), available at https://www.nvidia.com/content/tegra/embeddedsystems/pdf/jetson_tx1_whitepaper.pdf (noting the differing demands of training and inference). 59 60 16 and other companies, the reality in the near-term seems to be that training will remain the province of the GPU for many researchers and practitioners. Combined with the fact that training is likely to continue being computationally intensive for the foreseeable future, it is likely that the creation of machine learning models will continue to happen within centralized data centers. FPGAs and ASICs therefore seem likely to have a differential impact on the geography of machine learning. In the past, pre-existing data center infrastructure, business models, and the energy consumption of existing processors tended to encourage an architecture where training and inference were situated in the same, centralized locations. As this specialized hardware matures, it seems likely to encourage a more distributed pattern in inference, permitting the application of machine learning “on device” in a broader set of contexts. At the same time, the continuing computational costs and energy requirements of training mean that the creation of machine learning systems will continue to largely take place in a relatively smaller number of central facilities. This geographic pattern has implications for the governance of machine learning. Training of the most complex, sophisticated models will continue to take place in a small number of locations among the set of actors who have the financial resources to maintain or rent the computational power necessary. However, once trained, machine learning models can be increasingly diffused and distributed. No doubt some types of machine learning models will continue to be offered “as a service,” with inference taking place in the cloud. However, FPGAs and ASICs open the door to inference no longer being tethered in this particular way. Simultaneously, these platforms - ASICs in particular - are more inflexible, making modifications after distribution more difficult. Whereas in the past it was more possible to repair flaws in machine learning models after training and deployment by directly modifying the model provided to many endpoints through the cloud, the specialization of hardware suggests an environment where these harms may be more challenging to rectify in the post-training phase. “Hard-wired” chips may be difficult to recall quickly, or otherwise difficult to modify when in the field. This may put increased pressure on companies to engage in more thorough pre-deployment checks and verification on these systems, rather than adopting a development approach that takes a more “launch-and-iterate” stance. From a governance perspective, these less easily rectifiable downstream harms may push regulators towards an approach that puts a 17 growing set of responsibilities on the entities creating and providing platforms for creating machine learning models to take precautions prior to wider distribution. It is worth recognizing that the continued development of the machine learning field may alter this balance over time. Progress continues to be made in the subfield of federated learning - which envisions an architecture in which many independent, distributed processors train locally and share updates to a model with one another.62 This work may become more practically feasible to implement as ASICs for machine learning mature and increasingly allow embedded training to happen on a device. Breakthroughs in one-shot learning, which would enable the effective training of models with a relatively smaller number of examples, might also lower the computational bar to executing training tasks in a more distributed way.63 PART III: SUPPLY CHAIN AND COMPUTATIONAL POWER CPUs, GPUs, FPGAs, and ASICs are all ultimately products in a complex global supply chain for semiconductors. Beyond simply connecting changing computational architectures to the social impact of machine learning as we did in the previous section, we can dig deeper to examine how the commercial specifics of its manufacture also has broader implications. Both the geography of semiconductor manufacturing and its place as a strategic asset in the context of national security make it likely that computing power will become an important arena in the geopolitics of AI. This seems to be particularly the case as China increasingly invests in becoming a leader in machine learning while continuing an ongoing effort to reshape the global semiconductor industry. The Semiconductor Supply Chain The hardware platforms discussed in Part II are just one facet of the much broader industry for semiconductors. Semiconductor chips, “tiny 62 See, e.g., Jakub Konečný et al., Federated Optimization: Distributed Machine Learning for On-Device Intelligence, arXiv:1610.02527 [cs] (2016), http://arxiv.org/abs/1610.02527 (last visited Mar 21, 2018). 63 See, e.g., Adam Santoro et al., One-shot Learning with Memory-Augmented Neural Networks, arXiv:1605.06065 [cs] (2016), http://arxiv.org/abs/1605.06065 (last visited Mar 21, 2018). 18 electronic device[s] comprised of billions of components that store, move, and process data” are the “enabling technology of the information age.”64 These chips give computers the power to run software applications, and are the key building block for a range of other devices “from cell phones and gaming systems to aircraft and industrial machinery to military equipment and weapons.”65 Not surprisingly given their broad scope of application, semiconductors are a massive global industry. In 2015, worldwide semiconductor sales were $335 billion, growing 15% since 2012.66 Many production steps are required to deliver a finished semiconductor chip. Some companies are “integrated device manufacturers”, or IDMs, which manage all aspects of semiconductor production from start to finish. This includes design, manufacturing, assembly, testing, and packaging.67 Companies adopting this model include Intel, Samsung, and Texas Instruments.68 However, many businesses specialize only in a particular part of this supply chain, contracting out tasks to other companies in the ecosystem as needed. Of particular importance in the discussion of machine learning hardware is the role of so-called “fabless foundries.” These businesses focus on the design of semiconductor chips, and contract out the manufacturing, often called “fabrication,” of the final product.69 Companies adopting this model include AMD, Broadcom, and Qualcomm.70 Many of the companies leading the development of machine learning specific hardware are “fabless”. This allows these businesses to avoid the massive capital outlay and expense of building and maintaining a chip “fab”. Building a single advanced plant for fabricating semiconductors can cost up to $20 billion.71 Michaela D. Platzer and John F. Sargent Jr., U.S. Semiconductor Manufacturing: Industry Trends, Global Competition, Federal Policy, Congressional Research Service 1 (2016) https://fas.org/sgp/crs/misc/R44544.pdf (last visited Feb 13, 2018). 65 Id. 66 Id. at 3. 67 Cf. Semiconductor Industry Association, Beyond Borders: The Global Semiconductor Value Chain 7 (2016), available at https://www.semiconductors.org/document_library_and_resources/trade/beyond_borders_t he_global_semiconductor_value_chain/. 68 Id. 69 Id. 70 Id. 71 TSMC Ready to Spend $20 Billion on its Most Advanced Chip Plant, Bloomberg.com, October 6, 2017, https://www.bloomberg.com/news/articles/2017-10-06/tsmc-ready-tospend-20-billion-on-its-most-advanced-chip-plant (last visited Mar 20, 2018). 64 19 The market for GPUs has been dominated by Nvidia, a fabless foundry. One industry analysis concluded that, as of the third quarter of 2017, Nvidia represented 72.8% of the market share for GPUs, with the rest being controlled by AMD, another fabless foundry.72 Both are headquartered in Santa Clara, California. The market for FPGAs has also been dominated by a small set of fabless foundries. In 2016, Xilinx led this segment with a market share of 53%.73 Altera, another FPGA specialist, was purchased by Intel in 2015 and accounted for 36% of the market.74 These were distantly followed by competitors Microsemi (7%) and Lattice Semiconductor (3%).75 This roughly holds stable from the market share in 2015.76 All of these companies are based in the United States. All but one, Lattice Semiconductor, are headquartered in California.77 Estimating market share in the context of specialized machine learning ASICs is more challenging. For one, the market is still emerging: no major player in the space is currently engaging in mass production and public sale of ASICs as a platform for machine learning. At the time of writing, Google is only distributing their TPU ASIC to a relatively small circle of researchers, and offers limited access to TPU computing cycles via its cloud services.78 In any case, because these chips are highly customized for particular purposes, it may be challenging to eventually define a single “market” which usefully groups together the different types of devices that might be fabricated as an ASIC for machine learning. Harsh Chauhan, Nvidia Is Running Away With the GPU Market The Motley Fool (2017), https://www.fool.com/investing/2017/12/06/nvidia-is-running-away-with-the-gpumarket.aspx (last visited Feb 20, 2018). 73 And the Winner of Best FPGA of 2016 is... | EE Times, EETimes, https://www.eetimes.com/author.asp?section_id=36&doc_id=1331443 (last visited Feb 20, 2018). 74 Id. 75 Id. 76 Id. 77 Xilinx, Corporate Locations, https://www.xilinx.com/about/contact/corporatelocations.html; Intel Programmable Solutions Group (PSG) Locations, https://www.altera.com/about/contact/contact/altera-hq.html; Microsemi, Locations, https://www.microsemi.com/locations; Lattice Semiconductor, Locations, http://www.latticesemi.com/About. 78 Google is giving a cluster of 1,000 Cloud TPUs to researchers for free, TechCrunch (2017), http://social.techcrunch.com/2017/05/17/the-tensorflow-research-cloud-programgives-the-latest-cloud-tpus-to-scientists/ (last visited Mar 21, 2018); Cloud TPUs - ML accelerators for TensorFlow, Google Cloud, https://cloud.google.com/tpu/ (last visited Mar 21, 2018). 72 20 Since the major leaders in machine learning hardware are “fabless”, they depend on a separate ecosystem of companies to provide the actual fabrication of the chips they design. These companies, called “pure play foundries” or simply “foundries”, are a highly consolidated marketplace. In 2016, Taiwan Semiconductor (TSMC) accounted for 59% of the global market for fabrication.79 Running significantly behind were GlobalFoundries (11%), United Microelectronics Corporation (UMC) (9%), and Semiconductor Manufacturing International Corporation (SMIC) (6%). 80 TSMC and UMC are based in Taiwan, with SMIC in China and GlobalFoundries in the US.81 While based in the US, GlobalFoundries is owned by the Emirate of Abu Dhabi through its state-owned investment arm Advanced Technology Investment Company (ATIC).82 There exists a network of somewhat stable relationships between this handful of leading “fabless” foundries that are designing much of the hardware that machine learning takes place on, and the small number of companies that do their manufacturing. In the GPU space, Nvidia contracts much of its high-performance GPU production to TSMC.83 In 2009, GlobalFoundries was spun-off from AMD as part of a transition of the latter towards a “fabless” model.84 As a result of this historical relationship, AMD has traditionally worked closely with GlobalFoundries, though recently Pure-Play Foundry Market Surges 11% in 2016 to Reach $50 Billion!, , http://www.icinsights.com/news/bulletins/PurePlay-Foundry-Market-Surges-11-In-2016To-Reach-50-Billion/ (last visited Feb 20, 2018). 80 Id. 81 TSMC, Business Contacts, http://www.tsmc.com/english/aboutTSMC/business_contacts.htm; UMC, Locations, http://www.umc.com/english/contact/index.asp; SMIC, About Us, http://www.smics.com/eng/about/about.php; GlobalFoundries, About Us, https://www.globalfoundries.com/about-us. 82 Mark LaPedus, ATIC takes control of GlobalFoundries | EE Times, EETimes, https://www.eetimes.com/document.asp?doc_id=1258215 (last visited Mar 21, 2018). 83 See Nvidia: TSMC will remain a ‘very important’ foundry partner, KitGuru (2015), https://www.kitguru.net/components/graphic-cards/anton-shilov/nvidia-tsmc-will-remaina-very-important-foundry-partner/ (last visited Feb 20, 2018); Ashraf Eassa, NVIDIA Corp.’s Relationship With Taiwan Semiconductor Manufacturing Is Deepening The Motley Fool (2017), https://www.fool.com/investing/2017/05/17/nvidia-corp-relationshiptaiwan-semiconductor.aspx (last visited Mar 21, 2018). 84 Benjamin Pimentel, GlobalFoundries created from AMD spin-off MarketWatch, https://www.marketwatch.com/story/globalfoundries-created-amd-spin-off-the (last visited Mar 21, 2018). 79 21 announced that they will be splitting their new GPU production between them and TSMC.85 FPGA production operates with a slightly different set of connections between chip designers and associated foundries. Xilinx has worked in the past with UMC though increasingly partners with TSMC on its more recent hardware.86 Altera was acquired by Intel in 2015 and relies in part on its corporate parent for fabrication services.87 Microsemi announced in 2013 that it too would work with Intel for its fabrication needs.88 Lattice Semiconductor works with both UMC and TSMC, along with some smaller foundries.89 The geographic distribution of these players is mirrored in the overall structure of the industry. US firms account for the largest share of the global market, accounting for 50% of semiconductor sales in 2016.90 However, the actual fabrication of semiconductor devices largely takes place outside of the United States. In 2015, about three-quarters of the world’s advanced semiconductor fabrication capacity was located in South Korea, Taiwan, Japan, and China.91 This continues a historical trend of production capacity moving from the US to the Asia-Pacific region. In 1980 the US accounted for 42% of global manufacturing capacity, a number which dropped consistently over subsequent years to 16% by 2007.92 US National Security and the Semiconductor Industry See AMD are splitting 7nm Zen 2 CPU and Vega GPU manufacturing between TSMC and GloFo | PCGamesN, , https://www.pcgamesn.com/amd-7nm-tsmc-globalfoundries (last visited Feb 20, 2018). 86 Dylan McGrath, Xilinx confirms: Samsung, TSMC in, UMC out at 28-nm | EE Times, EETimes, https://www.eetimes.com/document.asp?doc_id=1173112 (last visited Mar 21, 2018). 87 Can Intel Dethrone The Foundry Giants?, Semiconductor Engineering, https://semiengineering.com/intel-dethrone-foundry-giants/ (last visited Mar 21, 2018). 88 Microsemi, Microsemi Selects Intel Corporation Foundry Services for the Development of Digital Integrated Circuits, https://investor.microsemi.com/2013-05-01-MicrosemiSelects-Intel-Corporation-Foundry-Services-for-the-Development-of-Digital-IntegratedCircuits (last visited Mar 20, 2018). 89 Lattice Semiconductor, ISO 9001/ TS 16949/ ISO 14001 Certificates, http://www.latticesemi.com/en/Support/QualityAndReliability/ManufacturingPartners/ISO Certificates (last visited Mar 20, 2018). 90 See supra note 61 at 2. 91 Id. at 9. 92 Id. at 9-10. 85 22 Given the importance of semiconductors to the supply chain of consumer and military electronics, the industry has “[f]or decades” been considered relevant to national security within the US.93 Concerns about the shift of semiconductor manufacturing capacity to the Asia-Pacific region have also dominated this discussion for decades, beginning with the rise of the Japanese semiconductor industry in the 1970s.94 Worries about this transition has resulted in a range of different regulatory interventions in past decades to bolster and secure the US semiconductor industry, and block foreign access to the most cutting edge computational power. Export controls in high-performance computing have been one important way these concerns have manifested as policy. Under Executive Order 13222, the US Commerce Department regulates exports of high performance computers (HPC) to certain countries, which are grouped into tiers.95 Computers exceeding certain thresholds of processor performance to particular countries and end-users require prior government review for export.96 These exports may be blocked on national security and antiterrorism grounds.97 Fears around reliance on foreign manufacturing have also resulted in more extensive coordination between government and corporate actors. In 2004, the US Department of Defense (DOD) and the National Security Agency (NSA) initiated a “Trusted Foundry Program”, which arranged for long-term contracts with accredited US companies to ensure that the government would have guaranteed access to trusted chips for its needs.98 In practice, this program would center on a sole-source contract with IBM, which was deemed “the only U.S.-based company able to meet DOD and intelligence community needs for trusted leading-edge microelectronics.”99 While the program eventually expanded to include other trusted suppliers, one 2015 analysis by the Government Accountability Office (GAO) observed that IBM remained the only supplier with cutting-edge fabrication Supra note 61 at 21. Id. at 18-19. For a contemporaneous review, see Warren E. Davis & Daryl G. Hatano, The American Semiconductor Industry and the Ascendancy of East Asia, 27 California Management Review 128 (1985). 95 Exec. Order No. 13,222, 66 Fed. Reg. 44025 (Aug 22, 2001); Bureau of Industry and Security, US Department of Commerce, Legal Authority for the Export Administration Regulations, https://www.bis.doc.gov/index.php/documents/Export%20Administration%20Regulations %20Training/876-legal-authority-for-the-export-administration-regulations/file. 96 15 CFR § 774, Supplement No. 1. 97 Id. 98 Supra note 61 at 22. 99 Id. 93 94 23 facilities and that “use of accredited suppliers other than IBM has been minimal” as a result.100 This program is at present in transition. In 2014, IBM announced that it would transfer ownership of its foundries to the foreign owned GlobalFoundries.101 This sale followed significant losses for IBM on its foundry business. During 2013 and 2014 the company lost $700 million on its two primary chip fabrication facilities, and in the end IBM paid $1.5 billion to GlobalFoundries for the acquisition.102 This transfer was approved and included a multi-year contract with the DOD to provide semiconductors to the US government until 2023.103 However, the 2015 GAO analysis concluded simply that “there are no near-term alternatives to the foundry services formerly provided by IBM.”104 As of late 2018, efforts are ongoing to identify “new approaches to retain trustable, leading-edge capabilities”. 105 Broader coordination has also happened in the past. During the 1980s and 1990s, concern around rising Japanese dominance in the space motivated the launch of SEMATECH, a public-private research consortium of US semiconductor firms.106 Over subsequent years, $870 million in public subsidies from the Defense Advanced Research Projects Agency (DARPA) and augmented by matching funds from the participating companies focused on accelerating research into semiconductor manufacturing.107 In 1996, after a period of growing US market share in semiconductors, the directors of SEMATECH voted to stop receiving federal funding.108 Many factors enabled the resurgence of the US semiconductor industry in the late 1980s and early 1990s, and the ultimate impact of Marie A. Mak, Trusted Defense Microelectronics: Future Access and Capabilities Are Uncertain, Congressional Research Service 4 (2015), available at https://www.gao.gov/products/GAO-16-185T. 101 Supra note 61 at 22. 102 IBM-GlobalFoundries Deal Finalized | EE Times, EETimes, https://www.eetimes.com/document.asp?doc_id=1327029 (last visited Mar 18, 2018). 103 Doug Cameron, Pentagon Hires Foreign Chips Supplier, Wall Street Journal, June 5, 2016, https://www.wsj.com/articles/pentagon-takes-foreign-chips-partner-1465159332. 104 See supra note 95 at 4. 105 Id. 106 See Larry D. Browning & Judy C. Shetler, Sematech: Saving the U.S. Semiconductor Industry (2000). 107 See supra note 22 at 19. 108 Id. at 20. 100 24 SEMATECH has been disputed.109 However, consensus favors the positive role the consortium played. One 2003 review by the National Academies concluded that the public-private effort was “key among elements in the industry’s revival—contributing respectively to the restoration of financial health and product quality.”110 Recent US Focus in the Space: China Concerns around the robustness of the US semiconductor industry have persisted into the present as China has made a concerted push to advance its own semiconductor industry in recent years. In 2014, the State Council of China published National Guidelines for Development and Promotion of the Integrated Circuit (IC) Industry, a strategic plan which aims to establish national leadership across the semiconductor supply chain by 2030.111 The plan allocates $100 to $150 billion from public and private investment to support this effort.112 Goals declared later in 2015 would imply ambitions that “roughly all incremental foundry capacity installed globally over the next ten years would have to be in China.113 One result of this strategy has been a rapid expansion in acquisition activity by Chinese investment and technology companies. Some larger transactions in the space include a $2.3 billion acquisition of H3C, a Hong Kong subsidiary of Hewlett-Packard114, a $2.75 billion acquisition of NXP Semiconductor’s “Standard Products” division, and a $1.8 billion acquisition of Omnvision, which specializes in semiconductors for imaging applications.115 One analysis in late 2017 estimated that “the total volume of transactions of China’s semiconductor overseas M&As (completed) so far has exceeded US$11 billion.”116 It has also been reflected in a significant increase in the construction of new chip fabrication plants in the country. Id. National Academies, Government-Industry Partnerships for the Development of New Technologies 90 (2002), available at http://www.nap.edu/catalog/10584 (last visited Feb 15, 2018). 111 Christopher Thomas, A new world under construction: China and semiconductors, McKinsey & Company, https://www.mckinsey.com/global-themes/asia-pacific/a-newworld-under-construction-china-and-semiconductors (last visited Mar 18, 2018). 112 Id. 113 Id. 114 C. P. Yue & T. Lu, China’s Latest Overseas M&A in the Semiconductor Industry, 9 IEEE Solid-State Circuits Magazine 8–12 (2017). 115 Id. 116 Id. 109 110 25 Roughly 40% of front-end semiconductor fabs slated to begin operation in 2017 to 2020 worldwide are located in China.117 This effort has also included the recruitment of several high-profile leaders within the Taiwanese semiconductor industry. This includes the recent move in 2017 by Shih-wei Sun, former CEO of Taiwan’s UMC foundry, to China’s Tsinghua Unigroup.118 Similarly, Shang-Yi Chiang and Liang Mong-song, both research leaders at Taiwan’s TSMC, were recruited to China’s SMIC that same year.119 These moves have produced some alarm within the Taiwanese semiconductor industry. TSMC chairperson Morris Chang has voiced concerns that the industry may become a “one way road” where “talent only departs and never arrives.”120 These efforts are directed towards catching up: China has not traditionally been a leader in the semiconductor sector. While it accounts for a large percentage of worldwide consumption (57% in 2014), it largely relies on imports to meet this demand.121 As of 2015, China possessed only around 6% of the most advanced semiconductor fabs globally.122 One report noted that many fabrication plans continued to use “older technology and used equipment”, which reflected China’s focus on products that “do not require leading-edge semiconductors.”123 These recent efforts also follow on earlier, less successful initiatives by the Chinese government to accelerate the development of their national semiconductor base.124 However, the present effort differs in its focus on fostering a defined set of Dylan McGrath, China to house over 40% of semi fabs by 2020, EE Times Asia, https://www.eetasia.com/news/article/china-to-house-over-40-of-semi-fabs-by-2020 (last visited Mar 19, 2018). 118 Cheng Ting-Fang, China poaches more Taiwanese chip talent, Nikkei Asian Review, https://asia.nikkei.com/magazine/20171109/Business/China-poaches-more-Taiwanesechip-talent (last visited Feb 20, 2018); Alan Patterson, China Expected to Poach More Taiwan Chip Execs, EETimes, https://www.eetimes.com/document.asp?doc_id=1331144 (last visited Feb 20, 2018). 119 Id. 120 The China Post, TSMC boss slams government over China brain drain The China Post (2017), https://chinapost.nownews.com/20170125-11485 (last visited Mar 21, 2018). 121 See supra note 61 at 14. 122 Id. at 15. 123 Dieter Ernst, From Catching Up to Forging Ahead: China’s Policies for Semiconductors, East-West Institute (2015), https://www.ssrn.com/abstract=2744974 (last visited Mar 19, 2018). 124 Gordon Orr & Christopher Thomas, Semiconductors in China: Brave new world or same old story?, McKinsey & Company, https://www.mckinsey.com/industries/semiconductors/our-insights/semiconductors-inchina-brave-new-world-or-same-old-story (last visited Mar 21, 2018). 117 26 national “champions” to compete internationally, rather than spreading financial support more thinly throughout the economy.125 The US has been active in attempting to counter these efforts. In 2016, then Secretary of Commerce Penny Pritzker declared that the US would “not allow any nation to dominate this [the semiconductor] industry and impede innovation through unfair trade practices and massive, nonmarket-based state intervention.”126 An expert committee convened by the NSA and the Department of Energy (DOE) that same year agreed, “expressing significant concern that – absent aggressive action by the U.S. – the U.S. will lose leadership and not control its own future in HPC [highperformance computing].”127 Several steps were taken towards this end. President Obama authorized the National Strategic Computing Initiative in 2015, which directed the DOD, DOE, and the National Science Foundation to work to “preserve [the US] leadership role in creating HPC technology”.128 In 2015, the administration restricted the sales of advanced microprocessors to several research sites associated with the Chinese supercomputer Tianhe-2. 129 The administration also blocked the acquisition of the US based assets of Aixtron, a German semiconductor manufacturer, by China’s Fujian Grand Chip Investment Fund.130 Similar deals were canceled on the threat of Chips on their shoulders, The Economist, Jan 23, 2016, https://www.economist.com/news/business/21688871-china-wants-become-superpowersemiconductors-and-plans-spend-colossal-sums (last visited Mar 21, 2018). 126 U.S. Secretary of Commerce Penny Pritzker Delivers Major Policy Address on Semiconductors at Center for Strategic and International Studies, Department of Commerce (2016), https://www.commerce.gov/news/secretary-speeches/2016/11/ussecretary-commerce-penny-pritzker-delivers-major-policy-address (last visited Feb 20, 2018). 127 U.S. Leadership in High Performance Computing (HPC), NSA-DOE Technical Meeting on High Performance Computing, https://www.nitrd.gov/nitrdgroups/images/b/b4/NSA_DOE_HPC_TechMeetingReport.pdf (last visited Feb 13, 2018). 128 Office of Science and Technology Policy, Fact Sheet: National Strategic Computing Initiative (2015), https://obamawhitehouse.archives.gov/sites/default/files/microsites/ostp/nsci_fact_sheet.pd f. 129 Don Clark, U.S. Agencies Block Technology Exports for Supercomputer in China, Wall Street Journal, April 9, 2015, http://www.wsj.com/articles/u-s-agencies-block-technologyexports-for-supercomputer-in-china-1428561987 (last visited Mar 21, 2018). 130 Paul Mozur, Obama Moves to Block Chinese Acquisition of a German Chip Maker, The New York Times, December 2, 2016, 125 27 rejection by the Obama administration through the Committee on Foreign Investment in the United States (CFIUS), an inter-agency group “authorized to review transactions that could result in control of a U.S. business by a foreign person…in order to determine the effect of such transactions on the national security of the United States.”131 This included the sinking of a 2015 offer by China Resources Microelectronics and Hua Capital Management to acquire US-based Fairchild Semiconductor.132 The Trump administration has maintained the active stance of the Obama administration in limiting foreign acquisition of US semiconductor companies. In 2017, CFIUS blocked a proposed acquisition of Lattice Semiconductor, a leading FPGA producer, by China Venture Capital Fund Corporation (CVCF), a state-run investment arm.133 The decision cited “credible evidence that the foreign interest exercising control might take action that threatens to impair national security” and the “importance of semiconductor supply chain integrity to the U.S. government, and the use of Lattice products by the U.S. government.”134 In 2018, using the same authority and citing similar concerns, the Trump administration blocked a proposed $117 billion acquisition of Qualcomm by Singapore-based competitor Broadcom.135 The administration similarly blocked a $580 million offer to acquire Xcerra, a Massachusetts-based semiconductor company, by state-backed investment funds Sino IC Capital and Hubei Xinyan that same year.136 In late 2017, a bipartisan group proposed the https://www.nytimes.com/2016/12/02/business/dealbook/china-aixtron-obama-cfius.html (last visited Mar 18, 2018). 131 US Department of the Treasury, The Committee on Foreign Investment in the United States, https://www.treasury.gov/resource-center/international/Pages/Committee-onForeign-Investment-in-US.aspx (last accessed Mar 22, 2018). 132 Diane Bartz and Liana B. Baker, Fairchild rejects Chinese offer on U.S. regulatory fears, Reuters, February 16, 2016, https://www.reuters.com/article/us-fairchild-semico-ma/fairchild-says-china-resources-offer-not-superior-to-on-semis-idUSKCN0VP1O8 (last visited Mar 21, 2018).3 133 Timothy B. Lee, Trump blocks Chinese purchase of US chipmaker over national security Ars Technica (2017), https://arstechnica.com/tech-policy/2017/09/trump-blockschinese-purchase-of-us-chipmaker-over-national-security/ (last visited Feb 13, 2018). 134 Statement On The President’s Decision Regarding Lattice Semiconductor Corporation, Department of the Treasury, https://www.treasury.gov/press-center/pressreleases/Pages/sm0157.aspx (last visited Feb 13, 2018). 135 Kate O’Keeffe, Trump Orders Broadcom to Cease Attempt to Buy Qualcomm, Wall Street Journal, March 13, 2018, https://www.wsj.com/articles/in-letter-cfius-suggests-itmay-soon-recommend-against-broadcom-bid-for-qualcomm-1520869867 (last visited Mar 21, 2018). 136 Raymond Zhong, U.S. Blocks a Chinese Deal Amid Rising Tensions Over Technology, The New York Times, February 23, 2018, 28 Foreign Investment Risk Review Modernization Act (FIRRMA).137 The bill would expand CFIUS review to include a range of transactions beyond outright acquisition, and would raise review in situations where an emerging technology prospectively “could be essential to national security.” 138 Impact: Computational Power and the Geopolitics of Machine Learning From a geopolitical standpoint, the semiconductor industry finds itself in the middle of a perfect storm. First, semiconductor manufacturing capacity is considered a vital strategic asset, and the issue of who owns these businesses and their intellectual property is considered a national security matter. Second, the evolution of the industry has produced a bifurcated geography, with design-focused “fabless” foundries largely based in the US but with actual production taking place in the Asia-Pacific region. The result is that the semiconductor industry, like others, has been increasingly drawn into international politics as tensions escalate between the US and China. This broader backdrop suggests that the more specific technology of machine learning is likely to be a flashpoint in the larger context of semiconductor competition between the US and China. In July 2017, the Chinese State Council released its “New Generation AI Development Plan”, which sets out targets for development in AI and related industries. The plan declares AI to be “a new focus of international competition” and “a strategic technology that will lead in the future”.139 It plans for China to “firmly seize the strategic initiative in the new stage of international competition in AI development” and “effectively [protect] national security.”140 Specifically, the plan aims for China to be the world’s “primary” leader in the technology and for the technology to be a $150 billion industry in the country by 2030.141 As in the semiconductor context, https://www.nytimes.com/2018/02/23/technology/china-microchips-cfius-xcerra.html (last visited Mar 21, 2018). 137 Foreign Investment Risk Review Modernization Act of 2017, S. 2098, 115th Cong. (2017), https://www.congress.gov/bill/115th-congress/senate-bill/2098. 138 Id. 139 State Council of China, A Next Generation Artificial Intelligence Development Plan 2 (2017), available at https://www.newamerica.org/cybersecurity-initiative/blog/chinas-planlead-ai-purpose-prospects-and-problems/. 140 Id. 141 Id. at 6-7. 29 this plan echoes and builds on a series of earlier published policies focusing on robotics and other emerging technologies.142 There are a number of components to the State Council strategy. For one, the Chinese government intends to ramp up spending to support the development of AI in the country. Its Artificial Intelligence 2.0 program is poised to deploy “billions to develop AI for commercial and military use.” 143 This will also be supported by a system of “government guidance funds” that provide financial support to companies and new ventures working on AI and related technologies.144 The plan also contemplates recruiting and accelerating the training of leading AI researchers in the fields of “neural awareness, machine learning, automatic driving, intelligent robots, and other areas.”145 Advancing computational power appears as a consistent theme in these plans. The State Council plan highlights the lack of “high-end chips” as a factor in China lagging in the field of AI.146 The Ministry of Industry and Information Technology (MIIT), in elaborating on the State Council Plan, has explicitly laid out the “[development of] high-performance, scalable, and low-power cloud neural network chips for machine learning training applications” as a key priority.147 It also distinguishes between the cloud and the separate need to advance chips to support “terminal” embedded applications, where “low-power, high-performance neural network chips suitable for machine learning algorithms” are needed.148 By 2020, MIIT targets chips which will have “performance levels of 128 TFLOPS (16-bit floating point), and the energy efficiency ratio of more See, e.g., Central Committee of the Communist Party of China, The 13th Five-Year Plan for the Economic and Social Development of the People’s Republic of China 67 (2016), available at http://en.ndrc.gov.cn/newsrelease/201612/P020161207645765233498.pdf (highlighting AI as a target “strategic emerging industry” for development); State Council of China, Guiding Opinions on Actively Promoting the "Internet Plus" Initiative (2015), available at http://www.gov.cn/zhengce/content/2015-07/04/content_10002.htm (highlighting the development of AI as a priority). 143 Tom Simonite, The Trump Administration Can’t Stop China From Becoming an AI Superpower, WIRED, Jun 29, 2018, https://www.wired.com/story/america-china-aiascension/ (last visited Mar 21, 2018). 144 Yuan Yang, China fuels boom in domestic tech start-ups Financial Times (2017), https://www.ft.com/content/b63ee746-afc6-11e7-aab9-abaa44b1e130 (last visited Mar 21, 2018) (describing the government guidance funds ecosystem). 145 See supra note 133 at 14. 146 Id. at 4. 147 Translation: Chinese government outlines AI ambitions through 2020, New America, https://www.newamerica.org/cybersecurity-initiative/blog/chinas-plan-lead-ai-purposeprospects-and-problems/ (last visited Mar 21, 2018). 148 Id. 142 30 than 1 TFLOPS/Watt” with “terminal” chips having a similar efficiency.149 Consistent with these policies, in August 2017 China’s State Development & Investment Corporation led a $100 million round of funding in Cambricon, a startup focusing on the development of specialized machine learning hardware150. National policies depend critically on what can be practically controlled by a state.151 In a competitive environment, nations seek to control levers to shape the prospects of their adversaries, or at least the means of denying them access to key opportunities. Norms of open publication in research, the existence of widely distributed open-source toolkits, and the ever widening circle of those familiar with the practice of machine learning suggest that efforts to control research, software, or specialists in general will be challenging over time.152 Computational power, with its small number of actors and large, fixed production facilities, is likely to be a focal point for national competition in the space. International rivalry around the technology of machine learning will therefore in large part manifest itself in practice as international rivalry over the supply chain of computational power. This is particularly true in the US-China context, where US companies depend on the manufacturing capability of a Chinese neighbor Taiwan - and China desires access to more advanced computing power to support machine learning applications. Through market leaders Nvidia and Xilinx, Taiwanese foundries TSMC and UMC play a critical role in the manufacture of high-end GPUs and the FPGAs likely to support machine learning applications going forwards. As CFIUS seems poised to make it increasingly challenging for China to pursue acquisitions of US semiconductor companies, shaping Taiwanese trade and foreign investment policies in semiconductors will assume greater Id. Tom Simonite, China Challenges Nvidia’s Hold on Artificial Intelligence Chips | WIRED, https://www.wired.com/story/china-challenges-nvidias-hold-on-artificialintelligence-chips/ (last visited Mar 19, 2018). 151 Cf. James Scott, Seeing Like a State: How Certain Schemes to Improve the Human Condition Have Failed (New Ed edition ed. 1999). 152 See, e.g., arXiv, https://arxiv.org/ (open publication); Tensorflow, https://www.tensorflow.org/ (toolkit); Pytorch, http://pytorch.org/ (toolkit); Fast.ai, http://www.fast.ai/ (training); Deeplearning.ai, https://www.deeplearning.ai/ (training). 149 150 31 importance.153 The extent to which the US is able to successfully deny China access to advanced computing power, and the extent to which China is able to develop it domestically or acquire it otherwise, remains to be seen. The outcome will play a major role in determining the availability and sophistication of machine learning systems in different markets and among companies globally. These trends may be further exacerbated as the machine learning field itself continues to evolve. In particular, there exist a rapidly developing set of research areas within machine learning - simulation learning, self-play, and meta-learning - which seem poised to heighten the importance of computational power as the key element in advancing the effectiveness of machine learning models. Moreover, meta-learning, which seeks to automate the architecture of machine learning algorithms, may also enable computational power to substitute somewhat for a lack of ready research and engineering specialists in the space. PART IV: COMPUTATIONAL POWER AS DATA SUBSTITUTE Computational power is just one ingredient in the process of training effective machine learning systems. Machine learning also requires sufficient and appropriate data, the examples on which the system will learn a good representation of the task it is attempting to accomplish. The cost and availability of these inputs is crucial in thinking about the distributional impact of machine learning. Depending on the resources available to the practitioner and the presence of pre-existing data, acquiring both of these critical inputs may be more or less feasible for solving a particular problem. This influences both the set of actors that can generate novel machine learning systems, as well as the set of domains that machine learning can be effectively used in. While computational power is increasingly commoditized and available as a service, acquiring and cleaning data appropriate to training a system to solve a given task remains stubbornly costly and time-consuming in a number of domains. Insofar as commercial actors continue to be the 153 See supra note 113, Patterson (Quoting one industry analyst that “China will find it very tough to buy U.S. high-tech companies and difficult to leverage Chinese joint ventures or wholly-owned enterprises to access key U.S. intellectual property…We thus expect more senior Taiwan veterans to join China’s semiconductor industry as a second wave of talent moves to China.”) 32 primary investors in the development of machine learning, this may render a set of potential machine learning applications infeasible or unprofitable to invest in developing. However, one significant recent trend on this front has been a set of research successes which - in certain domains - are reducing a dependence on acquiring real-world data and increasing the role of computational power in the training of machine learning systems. This is significant since in these situations leveraging machine learning will be limited by the cost and availability of computing power, rather than the more fixed costs of gathering data in a specific domain. There are three arenas of research which seem particularly promising in this respect: simulation learning, selfplay, and meta-learning. Simulation Learning Simulation learning seeks to replace the need for collecting real-world data by training an agent in a virtual space. For example, rather than training a robot arm to grasp objects by having a real robot arm attempting to pick up many different types of objects in the physical world, this approach trains the machine by having it “experience” picking up virtual objects in a computer simulation.154 Once the system learns a good representation of the task, it can then be implemented to control a robot arm in the real world.155 Successes in using this approach have prompted a number of different simulation environments for training agents to be released in recent years.156 The continuing advancement of simulation learning changes the balance of data and computational power. From a practical perspective, Compare Xue Bin Peng et al, Generalizing from Simulation, OpenAI Blog (2017), https://blog.openai.com/generalizing-from-simulation/ (last visited Mar 21, 2018) with Sergey Levine, Deep Learning for Robots: Learning from Large-Scale Interaction, Google Research Blog, https://research.googleblog.com/2016/03/deep-learning-for-robotslearning-from.html (last visited Mar 21, 2018). 155 See, e.g., Xue Bin Peng et al., Sim-to-Real Transfer of Robotic Control with Dynamics Randomization, (2017), https://arxiv.org/abs/1710.06537 (last visited Mar 21, 2018); Josh Tobin et al., Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World, arXiv:1703.06907 [cs] (2017), http://arxiv.org/abs/1703.06907 (last visited Mar 21, 2018). 156 See Microsoft shares open source system for training drones, other gadgets to move safely on their own, Microsoft AI Blog (2017), https://blogs.microsoft.com/ai/microsoftshares-open-source-system-training-drones-gadgets-move-safely/ (last visited Mar 21, 2018); DeepMind, DeepMind Lab, https://deepmind.com/research/publications/deepmindlab/ (last visited Mar 21, 2018). 154 33 effective simulation learning removes the costs and logistical challenges of collecting data through sensors in the real world. In effect, it substitutes computational power for data as an input in machine learning. This is particularly promising since simulation learning might enable the collection of data far beyond what would be practical in the real world. Taking the example discussed above, a simulated approach may be able to run many more trials with virtual arms and virtual objects than the constraints of warehousing, maintenance, and energy in a real-world setting would allow. This at least theoretically holds out the promise of creating more effective machine learning systems which are trained on a more substantial simulated dataset at a faster rate. These gains may be limited to particular types of domains. One obvious limitation of this approach is that it will only be effective in situations in which the simulation acts as a sufficiently realistic proxy for the actual challenge. A simulation which has unrealistic assumptions will cause the agent to learn a poor representation that fails when it is finally deployed “in the field.”157 Simulating a robot operating in the physical world, for instance, may be relatively more straightforward than simulating group social behavior or ecological systems where underlying forces are not well understood. Self-Play Simulation learning avoids the need for collecting real-world data by substituting a virtual environment for an agent to interact with. Self-play avoids the need for real-world data by having an agent play itself in order to improve and acquire “experience” in a given task. It therefore represents another method which effectively substitutes computational power for data as an input in the machine learning process. At the time of writing, one notable demonstration of the use of selfplay as a technique for reducing a dependence on data has been AlphaZero, a system designed by Google’s DeepMind research lab which was able to achieve champion level skill at the board games of Go, chess, and shogi without the use of any pre-existing data of human games.158 This marks a departure from earlier versions of these systems, which required training on data around historical examples of Go games to achieve a certain level of See supra note 148 at 1. David Silver et al., Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm, arXiv:1712.01815 [cs] (2017), http://arxiv.org/abs/1712.01815 (last visited Mar 21, 2018). 157 158 34 proficiency.159 These self-play methods have also enabled advancements in transfer learning, allowing agents which learn skills in one environment to apply it in a new, novel one.160 Like simulation learning, self-play is also somewhat domain limited to situations in which the competitive action of two agents will generate the data necessary to train the system to solve the desired problem. But, in these contexts, computational power becomes a primary rate limiter of progress, rather than the availability of data. As in the simulation learning case, sufficient computational power enables a multitude of agents to rapidly play a number of games which potentially outstrips the data that is available from games played by humans. Meta-Learning Learning architectures - the system by which a machine learning system extracts features and acquires a good representation of the task it is intended to solve - are another key ingredient in the machine learning workflow. For the purposes of this discussion, these architectures are notable since they influence the efficiency of the training process, defining how successful a trained machine learning system will be at a task for a given amount of training examples given to it. These learning architectures have traditionally been hand-crafted by researchers and engineers developing these systems. Meta-learning seeks to automate this work by treating the task of designing these architectures as itself a machine learning task.161 This approach has seen a number of promising successes, with machine generated learning architectures outpacing the benchmarked performance of their human crafted counterparts at the same task.162 Meta-learning has also been used to Id. See, e.g., Trapit Bansal et al., Emergent Complexity via Multi-Agent Competition, arXiv:1710.03748 [cs] (2017), http://arxiv.org/abs/1710.03748 (last visited Mar 21, 2018). 161 See, e.g.,Marcin Andrychowicz et al., Learning to learn by gradient descent by gradient descent, arXiv:1606.04474 [cs] (2016), http://arxiv.org/abs/1606.04474 (last visited Mar 21, 2018). 162 See, e.g., Esteban Real et al., Regularized Evolution for Image Classifier Architecture Search, arXiv:1802.01548 [cs] (2018), http://arxiv.org/abs/1802.01548 (last visited Mar 21, 2018). 159 160 35 optimize the distributions of neural networks across CPUs and GPUs and improve training and inference performance.163 Advancements in these techniques again have implications for the role of computational power in the workflow of machine learning. For one, meta-learning might improve the performance of machine learning systems with a given dataset, or enable equivalent performance with less data. Secondly, the automation of learning architecture design may also increase the speed at which new machine learning systems can be designed by aiding human technicians in creating the appropriate architecture for a given task. Meta-learning does not offer a direct substitute to real-world data in the same way that simulation learning and self-play seem to. But, metalearning is parallel to these methods insofar as it potentially reduces the amount of data needed to train a system to successfully accomplish a task. Meta-learning also accentuates the need for computational power, as a training process is required for generating an effective learning architecture as well as in learning how to accomplish the task itself. Impact: Computational Power and Economic Impact These shifts in the research field have broader significance because of the role that the inputs of machine learning play in defining its economic impact. Machine learning must be cost-effective and equivalently proficient at accomplishing a given task before it can serve as a competitive substitute for a human executing the same task. However, achieving this level of proficiency requires a machine learning model to be trained - a process that typically requires both sufficient computational power and sufficient data. The availability of data and the cost of acquiring will vary depending on the domain. Existing market forces may already make this data freely available and structured for use. However, there are numerous reasons why this may not be the case. Legal restrictions rooted in copyright or privacy may make it expensive or challenging to acquire data.164 The specific domain may itself make the practical tasks around acquiring the right data difficult. This is important because the poor availability of data can make it Azalia Mirhoseini et al., Device Placement Optimization with Reinforcement Learning, arXiv:1706.04972 [cs] (2017), http://arxiv.org/abs/1706.04972 (last visited Feb 13, 2018). 164 See, e.g., Amanda Levendowski, How Copyright Law Can Fix Artificial Intelligence’s Implicit Bias Problem (2017), https://papers.ssrn.com/abstract=3024938 (last visited Mar 21, 2018). 163 36 challenging for machine learning to produce a system which is an effective automated substitute for a given task. In the very least, these barriers may make the application of machine learning economically unviable for certain tasks and domains. The advancements in machine learning research discussed above are significant because they reduce the dependence on data collection as a necessary component for training a machine learning system effectively. Simulation learning and self-play substitute computational power for data, enabling a system to learn a skill effectively without the physical costs and logistical requirements of real world collection. Meta-learning potentially boosts the efficiency of models by adding more computational power. These shifts have two economic impacts, one on the issue of technological unemployment, and the other on the issue of the industrial organization of machine learning. Technological Unemployment The economic literature examining the issue of technological unemployment in the context of machine learning has often abstracted away from the influence that inputs of data and compute have on the success of machine learning systems.165 One method is to rely on surveys of expert opinion about the likelihood that a given task is broadly subject to automation by machines.166 Common sources for these skills include the Dictionary of Occupational Titles (DOT) and its successor O*NET databases compiled by the US Department of Labor.167 These skills are usually articulated in a broad way, such as “[t]he ability to see objects in the presence of glare or bright lighting” or “[t]he ability to communicate information and ideas in speaking so others will understand.”168 See, e.g., Anton Korinek and Joseph E Stiglitz, Artificial Intelligence, WorkerReplacing Technological Change and Income Distribution, NBER Economics of Artificial Intelligence 2017, http://www.nber.org/conf_papers/f100963/f100963.pdf; David Autor & Anna Salomons, Robocalypse Now–Does Productivity Growth Threaten Employment?, NBER Economics of Artificial Intelligence 2017, http://papers.nber.org/conf_papers/f100969/f100969.pdf. 166 See, e.g., Carl Benedikt Frey & Michael A. Osborne, The future of employment: How susceptible are jobs to computerisation?, 114 Technological Forecasting and Social Change 254–280 (2017). 167 Dictionary of Occupational Titles, https://occupationalinfo.org/; O*Net Online, https://www.onetonline.org/. 168 O*Net, Sensory Abilities, https://www.onetonline.org/find/descriptor/browse/Abilities/1.A.4/; O*Net, Cognitive Abilities, https://www.onetonline.org/find/descriptor/browse/Abilities/1.A.1/. 165 37 However, examining whether or not machine learning would enable the automation of “the handling of irregular objects” in such an abstract way may fail to take into account the costs and practicality of assembling the data necessary to train a system to an adequate level of proficiency in a domain. It may also mask challenges in training a system to be sufficiently robust to accomplish this task across a broad set of sectors and contexts. The data needed to train a machine learning model to grasp an irregular object in a well-lit, controlled environment may differ considerably from the training data needed to teach it to do so underwater, or in the dark. Examining the availability and cost of these data inputs therefore becomes critical since it can determine the practical terms under which a human skill is subject to automation. The influence of data availability on machine learning performance suggests that the impact of machine learning on labor automation will proceed in a fragmentary, domain specific way. The data-replacing (or dataminimizing) techniques discussed above seem poised to accentuate this differential development. Though impressive, simulation learning and selfplay are not all-purpose; they are most appropriate as learning schemes in specific circumstances. Self-play is specific to a reinforcement learning context with competing agents, and simulation learning is most effective in situations where it is feasible to construct a virtual space analogous to the real-world problem being solved. However, within these applicable domains, the barrier to machine learning being applied to a situation may be fixed by the cost of computing power, rather than the unpredictable costs and occasionally absolute barriers to collecting data on a particular task. That may accelerate displacement in those domains. There are a scattered constellation of applicable domains suggested by these two techniques. Applications in robotics - in which an agent needs to accomplish a physical task in the real world, may be particularly benefitted by the acceleration made possible by simulation learning.169 Competitive marketplaces, from capital markets to programmatic advertising, may be places where techniques such as self-play work to significantly improve machine learning system performance. Military autonomy may also be similarly impacted as self-play and simulation learning together allow the training of drones and other robotics in a manner it would otherwise be See, e.g., Cade Metz, A.I. Researchers Leave Elon Musk Lab to Begin Robotics StartUp, The New York Times, November 6, 2017, https://www.nytimes.com/2017/11/06/technology/artificial-intelligence-start-up.html (last visited Mar 22, 2018). 169 38 prohibitively expensive to do in the real-world. Within these types of contexts, machine learning may begin to have a disproportionate impact. Meta-learning impacts the relationship between machine learning, automation, and labor in a related, but more general way. Meta-learning may produce an overall improvement in the ability for systems to be trained to an adequate level of proficiency on a task with less data. This may make certain applications at the margin more viable than they would otherwise be, particularly in situations where data is scarce. These techniques may also lower the costs to creating new machine learning systems where engineering and development costs would have previously made investment in developing a capability unprofitable. Certain domains may have plentiful data, but the costs of developing and tuning an appropriate learning architecture and the potential profit of offering a product in that domain may render the choice to actually invest in developing the product unattractive. By aiding human domain experts in architecting these systems, meta-learning impacts the engineering time and cost of developing new systems and in doing so broadens the set of arenas that machine learning will be deployed in. It is important to recognize that the processes by which labor is automated are complex, and not purely economic and technical. The ability for machine learning to proficiently accomplish a task is a necessary but not sufficient condition for that task to be actually automated in practice. Business practices can take a long time to respond to technological change.170 Work is a sociotechnical system - and many forces shape who is employed and how work is accomplished. Automation may depend less on the economic opportunities presented by the technology, and more on a complex fabric of local historical and cultural factors171. Nevertheless, these techniques may open opportunities to automate where it was previously technically infeasible or cost-prohibitive to do so. James Manyika et al, A Future That Works: Automation, Employment, and Productivity, McKinsey Global Institute 70 (2017), available at https://www.mckinsey.com/mgi/overview/2017-in-review/automation-and-the-future-ofwork/a-future-that-works-automation-employment-and-productivity (concluding that “automation will be a global force, but adoption will take decades and there is significant uncertainty on timing”). 171 See, e.g., Alex Rosenblat and Tim Hwang, Regional Diversity in Autonomy and Work: A Case Study from Uber and Lyft Drivers (2016), available at https://datasociety.net/pubs/ia/Rosenblat-Hwang_Regional_Diversity-10-13.pdf (highlighting regional variation in machine learning driven workplaces); Marc Levinson, The Box: How the Shipping Container Made the World Smaller and the World Economy Bigger, Chapters 5-6 (2008) (discussing local political circumstances shaping industrial outcomes in shipping automation). 170 39 Industrial Organization We can look at the economic impact of these research developments in two ways. Simulation learning, self-play, and meta-learning enable computational power to play a bigger role not just in defining the impact of machine learning on labor generally, but also in the competitive landscape between various firms seeking to offer products and services driven by AI, as well. To this end, computational power will play a role in shaping the industrial organization of machine learning: the types of competition, its level of consolidation, and the opportunities for new entrants. Two “ingredients” serve as barriers to entry in machine learning: the data necessary to train the systems, and the relatively small pool of qualified talent necessary to build effective learning architectures.172 Possessors of these assets have been able to generate some of the most impressive and groundbreaking machine learning systems, while competitors lacking these assets have been unable to compete on similar footing.173 These data-substituting and data-minimizing techniques are important because they shift the competitive balances between companies looking to lead in the latest generation of AI. For one, the advantages of data incumbency may be significantly eroded in certain markets. Expansive historical datasets on farm machinery to train agricultural robots, for instance, may be offset by sufficient computational power to simply simulate this farm machinery in a virtual space. As the research continues to develop, these may represent two viable paths towards developing and offering effective machine learning products in the space. The benefits of these research developments might then redound to the handful of large technology companies that have been already investing most aggressively in the latest generation of machine learning such as Google and Facebook. These are companies which already possess significant endowments of computational power. On that count, the improvement of simulation learning or self-play may enable these existing technology leaders to enter new industries and provide products to meet use 172 See, e.g., Cade Metz, Tech Giants Are Paying Huge Salaries for Scarce A.I. Talent, The New York Times, October 22, 2017, https://www.nytimes.com/2017/10/22/technology/artificial-intelligence-expertssalaries.html (last visited Mar 21, 2018). 173 See, e.g., Tom Simonite, Apple’s Privacy Pledge Complicates Its Push Into Artificial Intelligence, WIRED, Jul 14, 2017, https://www.wired.com/story/apple-ai-privacy/ (last visited Mar 21, 2018). 40 cases even where they lack training data. In the very least, it provides greater leverage to these technology companies in partnership negotiations with existing companies within a sector that do possess real world data. Meta-learning is likely to have a similar impact in practice, though in theory it should act as a leveler in the space. Meta-learning enables computational power to offset the advantages that a competitor may have in human machine learning expertise and talent. As the research matures and the techniques become increasingly commoditized, the autonomous construction of learning architectures may allow industries with extensive data and access to computing power to develop their own machine learning driven products effectively. But, it is unclear at what rate meta-learning will become more commodified and usable by non-specialists. Though it seems poised to accelerate development times and improve the performance of learning architectures, meta-learning itself at the moment continues to require technical expertise and computational power to leverage effectively. Established technology companies may very well stand to gain again from these research developments, given their existing endowments of compute and expertise. In this view, meta-learning will tend to act as a force multiplier for entities which already have machine learning talent on hand, rather than democratizing the technology to entities with no experts at all. These trends tend to suggest greater consolidation in the marketplace of companies offering products and services driven by machine learning. Simulation learning, self-play, and meta-learning all chip away at the potential advantage that non-AI focused industries might have in competing to offer their own machine learning systems: their data. In the very least, these trends suggest that one likely outcome is an ecosystem in which other industries rely on a handful of leading technology companies who provide these technologies as an “as a service” platform. This already seems to be in progress, as Google, Amazon, and a number of other companies compete to offer machine learning capabilities as a series of commodified APIs.174 CONCLUSION: PROSPECTS FOR GOVERNANCE See, e.g., Google, Google Cloud: Machine Learning at Scale, https://cloud.google.com/products/machine-learning/; Amazon, Machine Learning on AWS, https://aws.amazon.com/machine-learning/. 174 41 Computational power does more than simply enable the current generation of breakthroughs in machine learning. Trends in computational architecture, semiconductor supply chains, and the field of machine learning itself suggest the multifaceted ways in which hardware plays a role in influencing the immediate and long-term impact of the technology on society. These changes in the underlying hardware layer would appear to accentuate many existing concerns that have been raised around AI. Progression towards more specialized hardware seems to expand the privacy concerns presented by the technology, and present challenges to the nimble repair of these systems in the field. Evolving competition in the semiconductor industry seems likely to make machine learning a flashpoint in the broader economic conflicts between the US and China. Breakthroughs in simulation learning, self-play, and meta-learning seem likely to expand incumbent market power and potentially accelerate labor displacement. Simultaneously, seeing these issues through the lens of hardware also raises opportunities for governance. Computational power is intimately linked to a nexus of issues in civil rights, consumer safety, geopolitical competition, labor policy, and corporate power. In contrast to the software and data that the practice of machine learning relies on, hardware and its production is more centralized, more able to be monitored, and less able to evade efforts to govern it. Hardware therefore offers one tangible point of control for those seeking to find levers on which to base policy and effectively shape the social impact of the technology. Public policy to shape the markets and design of machine learning hardware might manifest in a range of different ways. Existing tools in the form of export controls and trade policy will play a significant role, especially given the specific political geography of the actors leading the design and actual fabrication of these chips. Industrial policy which aims to accelerate the development of a domestic semiconductor sector like SEMATECH, or otherwise ensure reliable access to computational power by companies and government actors working at the software layer of machine learning, will also play an important role. As FPGAs and ASICs broaden the potential range of contexts in which machine learning inference might be deployed, new regulations and industry standards might also serve to enshrine particular architectures for delivering machine learning products and services. 42 It is important to recognize that the current affordance that computational power provides to those thinking about the governance of AI may not be permanent. The machine learning field continues to evolve at a rapid rate, and new discoveries may change the relative importance and economics of computational power going forwards. Rapid advances in “one-shot learning”, which aim to enable machines to train effectively on smaller numbers of examples; or “transfer learning”, which works to allow machines to effectively use knowledge from one domain in another, may reduce the need for computational power. Indeed, a dependence on machine learning hardware and concerns about reliable access to it may itself motivate increased investment and more rapid progress on solving these types of technical problems. As with simulation learning and self-play, it is possible that these research areas may have a narrower influence, influencing only specific domains and settings in which machine learning is deployed. Much of what we currently understand about machine learning suggests that significant computational power will continue to be a key component in building effective systems.175 Regardless, any policy developed around such a fastmoving technology must take a potential future rewrite of our existing assumptions into account. Multiple efforts have been launched to construct a global structure of governance around the research, development, and application of AI.176 Many of these efforts are developing standards or sets of principles to guide the actions of the technology industry and the broader community of researchers and engineers working on machine learning.177 The aim has This is particularly the case given the increased performance that more data seems to promise. See Chen Sun et al., Revisiting Unreasonable Effectiveness of Data in Deep Learning Era, arXiv:1707.02968 [cs] (2017), http://arxiv.org/abs/1707.02968 (last visited Mar 21, 2018). 176 See, e.g., The Partnership on AI, http://partnershiponai.org/; The IEEE Global Initiative on Ethics of Autonomous and Intelligence Systems, https://ethicsinaction.ieee.org/; Building Agile Governance for AI & Robotics (BGI4AI), https://bgi4ai.org/ 177 See, e.g., Partnership on AI, Tenets, https://www.partnershiponai.org/tenets/; IEEE, 7000 - Model Process for Addressing Ethical Concerns During System Design, http://standards.ieee.org/develop/project/7000.html; Future of Life, Asilomar AI Principles, https://futureoflife.org/ai-principles/; The Conference Towards AI Network Society, http://www.soumu.go.jp/main_content/000507517.pdf; DJ Patil, A Code of Ethics for Data Science (2018), https://medium.com/@dpatil/a-code-of-ethics-for-data-sciencecda27d1fac1 (last visited Mar 22, 2018). For further sets of principles, see, e.g., Accenture, An Ethical Framework for Responsible AI and Robots, https://www.accenture.com/gben/company-responsible-ai-robotics; The Future Computed: Artificial Intelligence and its role in society, The Official Microsoft Blog (2018), 175 43 been to attempt to shape norms around how the technology is applied, and the ethical responsibilities around the use of data as an input in the training process. This is one approach, one which is likely to become increasingly difficult to enforce effectively as the set of actors that might use the technology for good or for ill, responsibly or irresponsibly, continues to expand. As this transition occurs, governance efforts focused on shaping the direction and availability of computational power may offer a useful addition to the toolkit of policy options, and an increasingly important one at that. https://blogs.microsoft.com/blog/2018/01/17/future-computed-artificial-intelligence-rolesociety/ (last visited Mar 21, 2018); Larry Dignan, IBM’s Rometty lays out AI considerations, ethical principles ZDNet, http://www.zdnet.com/article/ibms-rometty-laysout-ai-considerations-ethical-principles/ (last visited Mar 21, 2018). 44 

2cs.AI

1 Efficient Enumeration of Unidirectional Cuts for Technology Mapping of Boolean Networks Abstract—In technology mapping, enumeration of subcircuits or cuts to be replaced by a standard cell is an important step that decides both the quality of the solution and execution speed. In this work, we view cuts as set of edges instead of as set of nodes and based on it, provide a classification of cuts. It is shown that if enumeration is restricted to a subclass of cuts called unidirectional cuts, the quality of solution does not degrade. We also show that such cuts are equivalent to a known class of cuts called strong line cuts first proposed in [14]. We propose an efficient enumeration method based on a novel graph pruning algorithm that utilizes network flow to approximate minimum strong line cut. The runtimes for the proposed enumeration method are shown to be quite practical for enumeration of a large number of cuts. I. I NTRODUCTION T c d e f a b c d e f a b c d e f G3 G2 G1 G4 G2 G3 G5 G4 G7 G7 G6 G6 G8 G8 G0 G0 This research is supported by NSF PFI award no. 1237856 and NSF FRP award no. 1230401. a b G5 ECHNOLOGY mapping (TM) is a process of transforming a generic Boolean network, which is a network consisting of primitive gates (e.g. AND/OR), into an equivalent mapped network, that consists of a network of cells from a given technology library. Depending on the target implementation, the library cells can correspond to either lookup tables (LUT) in the case of an FPGA, or to a pre-designed set of standard cells in the case of an ASIC. The measures of delay, power, area, or some combination of them serve as an objective function to optimize during the transformation process. TM is formalized as a graph covering problem. A given Boolean network is represented as a directed acyclic graph (DAG) G = (V, E), which is referred to as the subject graph. The cells in the library, being single output Boolean functions, are also represented by DAGs, and each is referred to as a pattern graph. A feasible solution of TM is a complete covering of the subject graph with one or more of the pattern graphs (see Figure 1). A core step in this process is to identify a subgraph in the subject graph to match with one or more pattern graphs. In the structural approach to TM [24], the selection of subgraphs is achieved by computing a cut. The recent works on structural approaches to TM [4], [7], [16], [19] are based on a particular type of a cut, called minimal node-cut. A node-cut Cv associated with a node v ∈ V is a subset of nodes in the transitive fanin cone of v such that every path from the primary inputs to v includes a node in Cv . Note that in the existing literature, the definition of a node-cut restricts it to be minimal, i.e., no proper superset of a node cut is a legal node cut. In this article we will explicitly refer to a minimal node cut, if that is the case. By definition, A k-feasible minimal node cut of a node v is a minimal node cut of cardinality k [4]. A cut together with a node v defines a single-sink subgraph that is either directly mappable onto a k-input LUT in the case of an FPGA, or constitutes a match candidate for a NPNequivalent standard cell, for an ASIC design, provided such an entity exists in the given library. In the literature k-feasible node cuts are usually simply referred as k-feasible cuts as no other type of cut is usually considered for technology mapping [4], [7], [16], [19]. In Figure 1, associated with node G0, the sets {G6, G7}, and {G6, G8} are 2-feasible minimal node cuts, whereas {G2, G5, G7} is a 3-feasible minimal node cut. When the cut {G6, G7} is selected, the actual subgraph replaced by a cell in TM consists of gates {G8, G0}. Similarly, when the cut {G2, G5, G7} is chosen the subcircuit consisting of gates {G8, G6, G0} is replaced. G1 arXiv:1603.07371v1 [cs.DS] 23 Mar 2016 Niranjan Kulkarni, Sarma Vrudhula School of Computing, Informatics and Decisions Systems Engineering Arizona State University Email: vrudhula@asu.edu (a) h (b) h (c) h Fig. 1: a) A covered Boolean network. b) Its graph representation. c) The network mapped on the library gates. A node may have many associated k-feasible node cuts, which result in different coverings. In TM, the quality of a covering is usually taken to be the delay of the critical path, and/or the total area of the mapped circuit. Hence, to evaluate the quality of the complete covering, cuts also need to be evaluated. The quality of a cut is typically a combined measure of the subcircuit that will be replaced by a library cell and the subcircuit that feeds that cell. Since neither of 2 these can be evaluated without knowing the cut, enumeration of cuts is a core task in TM. Consequently, there has been a significant amount of effort devoted to the development of efficient algorithms for enumerating k-feasible node cuts [4], [5], [16], [22]. While structural technology mapping turned out to be very successful, minimal node cuts bear some bias that eventually limits the number of possible matches. As demonstrated in [18], this may result in excluding many feasible, high quality matches from the evaluation. The authors of [18] address this problem by constructing so-called supergates i.e. single output networks of library gates that are added to the library in order to increase the number of matches. This demonstrates that enhancing a match space could yield significant benefits for structural technology mapping. Existing work on TM focuses on minimal node cuts. However, including non-minimal node cuts can increase substantially increase the possible matches. Consider Figure 2. The node cut C = {a, b, x, d} is not minimal since {a, b, d} is also a (minimal) node cut. However C corresponds to the function ab + x + d. This happens to be a linearly separable (threshold) function, which would never be found by any cut enumerator that enumerates only minimal node cuts. Another representation of a cut, based on edges, is called a line cut, which includes both minimal and non-minimal node cuts. Definition 1: (a) A line cut is a minimal set of directed edges in a single sink DAG which when removed, eliminates all paths from any of source (primary input) to the sink i.e. produces an “S-T bipartition”. (b) A line cut is k-feasible if its cardinality (number of edges) is k or smaller. (c) A line cut is called a strong line cut [14] if no two edges in the line cut are on the same directed path. Note that line cuts and node cuts are simply two representations of same entity, and either form can be converted into other. They both partition the nodes of the DAG into two mutually exclusive sets S and T . In such an S-T partition, the set S of nodes must contain primary inputs and T must contain the sink node v. In Figure 2(a) the line cut {(x, z), (a, y), (b, y), (d, w)} (corresponding node cut being C = {a, b, x, d}) generates S = {a, b, c, d, e, x}, and T = {y, z, w}, and is unidirectional. In Figure 2(b), the line cut {(a, c), (b, g)} (corresponding node cut is {a, b}) has S = {a, b} and T = {c, g}, and is bidirectional since (b, g) is a “forward edge” and (c, b) is a “backward edge”. Note also that the minimal node cut {a, b, d} would identify ab + a0 b0 + d as a function to be replaced by some cell from the library. However this is neither a member of any well defined class of functions, e.g. threshold functions, nor it is a function that would be in a typical cell library. Thus, minimal cuts are not always the most useful or desirable. In addition, we show that bidirectional cuts are not necessary and discarding them does not degrade quality of TM with regard to critical path delay. Thus we need only enumerate unidirectional node cuts (minimal or non-minimal). We establish a one-to-one correspondence between unidirectional nod cuts and strong line cuts [14]. This correspondence is important because there exists a well established relation between strong line cuts in a DAG and independent sets in its corresponding line dependency graph (LDG) [14]. This allows for construction of efficient strong line cut enumeration techniques based on enumeration of independent sets. Since the latter is a problem well researched in computational graph theory, techniques exist that can be directly applied to enumerating line cuts by enumerating independent sets. Figure 3 shows classification of cuts and their relationships. The proposed technique for enumerating strong line cuts was used in [15] to find threshold functions in a logic circuit. Conventionally enumerated Line cuts Minimal node cuts Bidirectional Non-Minimal node cuts Unidirectional Unidirectional Bidirectional Discardable set Strong line cuts Fig. 3: Classification of cuts and their relationships Fig. 2: (a) Unidirectional node cut denoted as {a, b, x, d} (b) Bidirectional node cut denoted by {a, b} Figure 2 a) shows a line cut consisting of edges {(x, z), (a, y), (b, y), (d, w)}. The node cut corresponding to this line cut is C = {a, b, x, d}, which is not minimal. Since minimal and non-minimal node cuts are useful, it appears that line cuts are the cuts that should be enumerated. A cut can also be classified as being unidirectional or bidirectional. It is unidirectional if all the edges between S and T originate in S and terminate in T . Otherwise it is bidirectional. The main contributions of this article include: • Introduction of unidirectional node cuts and a proof showing that restricting the cut space to only unidirectional node cuts does not degrade the quality of mapping for delay. • Establishing the equivalence unidirectional node cuts and strong line cuts. • An efficient k-feasible strong line cut enumeration algorithm based on the relationship between a DAG and its LDG. • A general framework and the specific implementation of a pruning technique for k-feasible strong line cut enumeration. • An efficient implicit representation of bounded size MISs of a graph that allows for both unconstrained and con- 3 strained enumeration of such MISs. II. R ELATED W ORK The importance of cut computation in TM for FPGAs was first identified by the Cong et al. [5]. They developed a novel and elegant network flow based algorithm that directly identified a single, depth-optimal, k-feasible node cut, without enumerating cuts. Later, Pan et al. [21], [22] developed an efficient algorithm for enumerating cuts that avoided the large computational requirements of network flow. More recently, Ling et. al [16] developed a novel scheme for implicitly encoding cuts using Binary Decision Diagrams (BDD). This representation allowed for extraction of cuts when the value of a cut could be computed recursively. However, the authors admit that BDD approach is not very well suited for cut enumeration since non-cuts, which dominate cuts, are also implicitly included and need to be pruned during enumeration. Finding a computationally efficient way of encoding and enumerating cuts is of fundamental importance to technology mapping. Recently Takata et al. [25] proposed a top-down scheme for the procedure of [22] and demonstrated speedups of 3x-8x for larger k = 8, 9. Unfortunately, since the number of cuts of size at most k is of O(nk ), cut enumeration algorithms inherently suffer from poor scalability. To alleviate this problem, techniques for ranking and pruning of cuts were first proposed by Cong et al. in [7]. The basic observation of this work is that for certain optimization objectives it is possible to narrow the search down efficiently and extract depth-maximal or area-minimal cuts directly. Similar ideas, referred to as priority cuts, were proposed by Mischenko et al. in [19], where appropriate seeding of the procedure from [22] assured enumeration of only O(n2 ) priority cuts instead of O(nk ) cuts. These can be further sorted by quality, and pruned. An alternative approach to pruning was proposed by Chatterjee et al. in [4] where they introduced hierarchical partitioning of the cut space based on a novel concept that is similar to algebraic factorization. The authors showed that while complete factorization may still suffer from poor scalability, partial factorization of the cut space could yield good, practical solutions with very short runtimes. Takata [25] proposed a partial enumeration scheme that enumerates only a subset called label cuts. The scheme improves scalability of cut enumeration and guarantees to maintain the circuit’s depth, at the expense of small increase in the estimated network area. All the works identified above, and many others have demonstrated that structural technology mapping, the core of which involves cut enumeration, leads to far superior solutions than the traditional graph/tree matching based algorithms. Cut enumeration has also found uses in related applications such as re-synthesis through rewriting [17], application specific instruction set extension generation/optimization [6], hardware/software co-design [23], model checking in verification [3], and SAT problem preprocessing for simplification [10]. The use of a line dependency graph (LDG) derived from a DAG was proposed by Kagaris et al. [14] to compute the maximum strong cut in a circuit for the purpose of delay testing. Based on the observation that an LDG is a transitively-oriented graph, hence a comparability graph [13], they provide an efficient and elegant algorithm that computes a maximum independent set of the LDG using network flow. This set represents a maximum strong cut in the corresponding DAG. While their approach generated interest in the area of delay-testing, we will demonstrate that there is still greater opportunity for further exploration and exploitation of the DAG/LDG duality for strong cut enumeration. III. S TRONG LINE CUTS We now describe the relation between DAGs and their corresponding line dependency graphs (LDG). An LDG is an undirected graph derived from a DAG that encodes the transitive dependencies between DAG edges [14]. Each edge e of the DAG has a corresponding node v in the LDG. Two nodes of an LDG are connected if and only if their corresponding lines in the DAG are transitively dependent, i.e., there exists a path in the DAG from any source to any sink that contains both edges. Consequently, if an edge in DAG is transitively dependent on another edge, by definition the corresponding nodes in LDG will be neighbors. Since LDGs are by definition transitively oriented, they are also comparability graphs [13]. An independent set (IS) in a graph G(V, E) is a set S of vertices no two of which share an edge. A maximal independent set (MIS) is an independent set which is not a proper subset of any other independent set in the graph. Lemma 1: (From [14]) A strong line cut of a DAG forms a maximal independent set (MIS) in its corresponding LDG. Fig. 4 illustrates the relation between DAGs and LDGs established by Lemma 1, on an example that we will use throughout this article for illustration. The direct consequence of this lemma is that enumerating all k-feasible strong line cuts in a DAG is equivalent to enumerating all maximal independent sets of size ≤ k in the LDG. A B q p C r q r s E D t u t u p s F v (a) v (b) M Fig. 4: a) a DAG with strong cuts annotated. b) The corresponding maximal independent sets in LDG. A. Relationship between unidirectional node cuts and strong line cuts In this section, we show the equivalence between unidirectional node cuts and strong line cuts. We also establish the fact that if the cut space is restricted to unidirectional node cuts then the quality of technology mapping for minimizing delay remains same. In the following we restrict the DAG to be a transitive fanin cone of some gate in a circuit, since in TM only transitive 4 Fig. 5: a) Classification of edges in a bidirectional cut b) Replication after TM, (c) Classification of edges in corresponding unidirectional cut. fan-in cones of gates/nodes are considered for enumerating cuts. v refers to the root node whose transitive fan-in cone is being considered. Lemma 2: A strong line cut corresponds to a unidirectional set of edges crossing an S − T partition. Proof: For an arbitrary node u ∈ S, there exists a path u v. This is straightforward from the definition of DAG considered here which is the transitive fan-in cone of the node v. Assume that the S − T partition corresponding to a strong line cut Cv is bidirectional, i.e., there exists a directed edge (p, q) such that p ∈ T and q ∈ S. Then for some x ∈ S and y ∈ T , (x, y) ∈ Cv , there must exist a path x → y p. Since edge (p, q) exists, there must exist a path x → y p → q. Since q ∈ S, the root node v must be reachable from q through another edge in the cut Cv , say (r, s). Therefore we have a complete path that looks like x → y p→q r→s v. Note that edges (x, y) and (r, s) both belong to the cut. This is clearly a contradiction, since no two lines in the cut Cv should lie on same path. Also (x, y) 6= (r, s) because that would lead to a directed cycle x → y p → q x in the directed acyclic graph under consideration. Conversely, assume a unidirectional node cut in which all the edges are from S to T , and suppose the corresponding line cut is not a strong line cut. Then there must exist at least edges e1 = (x, y) and e2 = (u, w) in the line cut that are on the same path to node v (the output). Assume e1 precedes e2 in the path. By definition of a S − T partition, x ∈ S, y ∈ T , u ∈ S and w ∈ T . However, since y u, we have an edge starting from T and ending in S, which contradicts the assumption that it is a unidirectional node cut. Lemma 2 confirms that a strong line cut must be unidirectional and a unidirectional cut must be a strong line cut. Note that the cardinality of a strong line cut and the unidirectional node cut can be different. The reason we convert back from a strong line cut to a node cut (which is unidirectional) is that eventually a node cut is what is mapped onto a cell. A node cut form of a line cut would always require smaller library cell whether mapping is done using standard Boolean functions or threshold functions. Next we show that restricting node cuts to unidirectional node cuts will not increase the critical path delay when that is the objective function being minimized by TM. Note that in TM, the delay of path is the sum of the delays of gates in the path. We show that the set of paths to the output node v in a bidirectional cut is the same as those in the corresponding unidirectional cut. Figure 5(a) shows a classification of the edges in a bidirectional cut. T = X ∪Y , where X is the set of logic cones (node and all nodes in its fanin cone) whose output has a directed edge to some node in S, and Y is the set of nodes with no paths to S. Note v ∈ Y . TM would replicate X in S and then replace T with some appropriate cell in the library. This is depicted in Figure 5(b). Four types of edges can be identified in the S − T partition: (1) E1 are edges from S to X, (2) E2 are edges from X to S, (3) E3 are edges from S to Y , and (4) E4 are edges from X to Y . Now a path from input node in S to the output node v ∈ T can be one three types: E E 1 4 1) S =⇒ X =⇒ Y =⇒ v. E2 E3 2) S =⇒ S =⇒ Y =⇒ v. E3 3) S =⇒ Y =⇒ v. Note that every one of the above paths (sequence of nodes) in the graph of Figure 5(a) also exists in the graph shown in Figure 5(b). Now consider the corresponding unidirectional cut shown in Figure 5(c). Every path that exists in Figure 5(a) also exists in Figure 5(c), and visa versa. This shows that there is no disadvantage to retaining only unidirectional cuts. IV. C UT ENUMERATION Enumerating MISs is a key step in many computational graph theory problems [1], [2], [11]. In TM, because there is a fixed library of functions, MIS enumeration needs to be restricted to sets of size ≤ k. Without any restrictions, the number of MISs in arbitrary graphs grows exponentially in the size of the graph [2]. However, in TM, the size k of the MIS is bounded above by some constant, and independent of n, which is the size of the graph. Therefore the number of MISs of size ≤ k is ≤ nk . A brute force approach in which all subsets of size ≤ k are examined and those that are not an MIS are discarded, is not practical for even realistic values of n and k. Existing algorithms exploit specific properties of small MISs to facilitate enumeration [11]. We now describe a method that can significantly speedup existing MIS enumeration algorithms by pruning away many MISs that will not be part of the final solution. A. MIS pruning The LDG of a DAG encodes MISs, many of which have sizes > k. The basic idea in the pruning algorithm is to 5 (efficiently) transform an LDG into a new, smaller, and denser graph G0 which contains all the MISs of size ≤ k of the original LDG, and as few other (parasitic) MISs as possible. The objective is to construct a transformation which is computationally efficient and would significantly reduce the runtime of enumeration. The graph G0 to be constructed must satisfy the following conditions: every vertex v of G0 as well as every disconnected pair of vertices in G0 must independently be a part of some MIS of size ≤ k of the original graph G. This condition translates into two steps of the pruning algorithm. In the first step, for each vertex v we attempt to determine the size of the smallest MIS to which v belongs. If this MIS is of size ≤ k then v is included in G0 . The second step decides if any two disconnected vertices in G may safely share an edge in G0 , implying that they will not be part of any MIS. Again for each pair of disconnected vertices (u, v) we attempt to determine the size of the smallest MIS containing both of the vertices. If such MIS is of size > k then an edge (u, v) is added to G0 . This is the key step in the of following pruning algorithm. Input: A DAG D(VN , EN ), An LDG G = (V, E) and an integer k Output: Graph G0 = (V 0 , E 0 ) characterized above dl = Φ, el = Φ; for vertex v in G do λ = Min-MIS(D, G, v); if λ > k then S dl = dl v; end end for disconnected pair (u, v) in G such that u ∈ / dl and v∈ / dl do λ = Min-MIS(D, G, u, v); if λ > k then S el = el (u, v); end end S E0 = E el; V 0 = V − dl; return G0 (V 0 , E 0 ); Algorithm 1: Algorithm to prune MISs of an LDG There is no known polynomial procedure to compute the size of the smallest MIS that contains a given vertex or a pair of vertices for comparability graphs [9]. Hence we approximate the size of the minimum MIS containing a vertex v or a pair of vertices (u, v) of a given LDG by exploiting the duality between MISs in LDGs and strong line cuts in DAGs. The minimum MIS in an LDG is the minimum strong cut in the DAG. We use this fact to approximate the minimum MIS size in an LDG by means of a flow computation in its corresponding DAG. It is well known that a minimum s-t cut (min-cut) is equivalent to the maximum flow the sink t receives from the source s [8], [12]. The size of a cut is a sum of capacities of edges involved in it. If unit edge capacities are assigned, then the size of a cut is equivalent to number of edges. We note that an edge with a capacity of ∞ can never be part of a finite size cut. The size of the minimum MIS in an LDG containing a vertex v or a pair of vertices (u, v) is approximated by computing the min-cut in the DAG with unit edge capacities. The procedure (Alg. 2) assigns a capacity of ∞ to the dependent lines of the given line (e.g. corresponding to node v) and a capacity of 1 to all other lines. The capacities of edges attached to s and t are always ∞. This is because a min-cut must consist of circuit lines only. Finally it returns the size of the minimum s-t cut of the network (λ). Input: A DAG D(VN , EN ), LDG G and a line v Output: Approximate size of minimum MIS (strong cut) containing v for u in G do capacity[u] = 1; end for neighbor w of v in G do capacity[w] = ∞; end return Min-cut(D,capacity); Algorithm 2: Min-MIS procedure for single edge in DAG Lemma 3: Let Smin be the minimum strong cut containing line v. Let λ be the size of a min-cut of the network with capacities modified based on line v. Then λ ≤ |Smin |. Proof: The Min-MIS procedure never assigns a capacity of ∞ to any line l ∈ Smin . Thus |Smin | is finite. It immediately follows that Smin is an arbitrary s-t cut, and cannot be smaller than the min-cut of the network. Lemma 3 states that the size of the minimum strong cut is guaranteed to be greater than or equal to the size of a mincut. Hence if the result of Min-MIS is > k then the size of a minimum strong cut is also > k. Consequently, the vertex in G for which Min-MIS was computed to be > k can be safely discarded from G0 , or an unconnected pair of vertices for which Min-MIS was computed can be connected in G0 . As an example, consider the LDG in Fig. 4(b), and suppose we wish to check whether vertex p belongs to an MIS of size ≤ 2. This is equivalent to determining if line p belongs to minimum strong cut of size at most 2 in the DAG. We assign capacities to the edges, as shown in Fig. 6(a). Line p is assigned a capacity of 1 and its dependent lines, t and v, are assigned a capacity of ∞. After this, the s-t minimum cut size (λ) is determined. In this example, it is lines p, q and u, and its size is 3 — i.e. λ = 3. Theorem 1: Let S be an MIS in G such that |S| ≤ k. Then S is also an MIS in G0 . Proof: From Lemma 3 every vertex of S must be present in G0 and that no edge was added between any two of its vertices. Thus S is an independent set in G0 . Assuming S is not a MIS in G0 , there exists an independent set S 0 in G0 such that S ⊂ S 0 . It follows that S 0 is also an independent set in G contradicting the fact that S is an MIS in G. Lemma 4: Pruning algorithm runs in O(n3 ). 6 small k. Hence, even the simple recursive algorithm that we used is still efficient. More sophisticated approaches to enumerate MISs of size ≤ k [1], [11] could be used to improve runtime for with values of k. Their application on the transformed graph may further improve total runtimes as well as scalability of the solution. S ∞ A ∞ ∞ B q 1 r 1 p D t 1 1 s E u ∞ C 1 t F v u C. Results ∞ G ∞ v (b) (a) T Fig. 6: a) Formation of s−t Boolean network for determination of a cut containing line p. b) LDG from Fig. 4 pruned for k ≤ 2. Proof: Let n be the number of nodes, m be the number of edges in DAG. The second for loop in pruning algorithm runs in O(m2 ) and dominates the overall complexity. Determining the min-cut takes O(km) time [8]. Since k ≤ n and is independent of n, the pruning has fixed complexity of O(m3 ). We know that m ≤ ∆n where ∆ is maximum degree found in the DAG. For most of the circuits with limited fanin (and fanout) capacities ∆ can be regarded as a small constant independent of n. Hence the time complexity of the pruning procedure is O(n3 ). As a result of the pruning transformation we perform MIS enumeration on G0 instead of G. Note that not all MISs in G0 are of size k in G. However, our experiments demonstrate that using G0 instead of G to enumerate MISs significantly reduces the enumeration time. In fact, the enumeration runtimes we observe in our experiments are practical for all of the evaluated benchmark circuits, suggesting that the approximation of a Min-MIS used in pruning must be quite accurate. B. Enumerating MISs As stated earlier, there are many known MIS enumeration techniques in computational graph theory. In fact, the choice of enumeration algorithm is independent of MISs pruning. In our experiments, we used a recursive procedure that is basically a simplified form of the algorithm presented in [1]. The idea is to recursively enumerate MISs that contain a specific node and then those that do not contain that same node. Once we designate a node to be part of the MIS, none of its neighbors can be part of that MIS. Similarly, if a node is not a part of the MIS, any of its neighbors can be part of the MIS. Note, that due to the pruning transformation being approximate, G0 may still contain some MISs of size > k. Since we are interested only in maximal independent sets of size ≤ k, we simply discard, on the fly, the MISs whose size is greater than k. Unfortunately, the number of MISs of a graph increases exponentially as a function of its size [20]. However in practice we found that while running it on a pruned graph G0 , enumeration time is dominated by pruning time for sufficiently The procedures described in this article were implemented in C++ and run on a 2GHz PC with 2 GB of RAM. The results of these runs are summarized in Table I. We ran the simple cut enumeration algorithm after pruning the MISs, and enumerated all cuts in the ISCAS benchmark circuits. The starting point was an AND-INVERTER Graph (AIG) obtained from [19]. Note that MIS enumeration is not possible with any of the existing MIS enumeration schemes. In fact, even a graph with as few as a hundred nodes would not be practical. The proposed pruning procedure makes it possible to exhaustively enumerate large numbers of strong line cuts in reasonable time. We also evaluated the combined effect of constraining the cone size and increasing the value of k. The results demonstrate that for sufficiently small k, line cut enumeration is dominated by our pruning transformation and as such is practically polynomial. For larger values of k however the procedure could however benefit from a more efficient MIS enumeration procedures. V. C ONCLUSION In this work, we presented a novel cut enumeration framework that exploits duality between enumerable entities in DAGs and LDGs. Apart from resource efficient computational procedure, it also introduces into the area of technology mapping, the concept of k-feasible strong line cuts (or unidirectional cuts) that are distinct from conventional node cuts. The advantages are two-fold. On one hand they are enumerable with low per-unit computational effort. On the other hand they potentially open up a new space available to be explored by the technology mapper without degrading the quality of the mapping. Line cuts provide choices not available to node cut based technology mappers. More importantly line cuts inherently mitigate some of the structural bias of node cuts and unlike node cuts they guarantee a mappable cut of a circuit. R EFERENCES [1] J. Byskov. Algorithms for k-colouring and finding maximal independent sets. In Proc. Symp. on Discrete Algorithms, pages 456–457. PA, USA, 2003. [2] J. M. Byskov. Enumerating maximal independent sets with applications to graph colouring. Oper. Res. Lett., 32(6):547–556, Nov. 2004. [3] M. L. Case, A. Mishchenko, and R. K. Brayton. Cut-based inductive invariant computation. In Proc. IWLS’08, pages 172–179, 4–6 Jun. 2008. [4] S. Chatterjee, A. Mishchenko, and R. K. Brayton. Factor cuts. In Proc. ICCAD ’06, pages 143–150, 5–9 Nov. 2006. [5] J. Cong and Y. Ding. FlowMap: an optimal technology mapping algorithm for delay optimization in lookup-table based FPGA designs. IEEE Trans. CAD, 13(1):1–12, Jan. 1994. [6] J. Cong, G. Han, and Z. Zhang. Architecture and compiler optimizations for data bandwidth improvement in configurable processors. IEEE Trans. on VLSI Syst., 14(9):986–997, 2006. 7 TABLE I: Running times for enumeration circuit # inputs # nodes c432 c1355 c1908 c6288 c7552 36 41 33 32 207 356 1133 1793 4864 7233 K = 6, cone size = no #cuts pruning after pruning time (s) 4317 0.93 228,950 49.04 9,519,182 48.12 1,494,636 6085.06 5,949,016 182.96 limit enum. time (s) 0.02 3.22 9.43 190.86 17.43 K = 6, cone size = 300 #cuts pruning enum. after pruning time (s) time (s) 4253 0.98 0.02 129,994 15 2.5 2,099,375 18.49 5.02 737,587 480.83 63.96 2,521,039 55.48 10.75 [7] J. Cong, C. Wu, and Y. Ding. Cut ranking and pruning: enabling a general and efficient FPGA mapping solution. In Proc. FPGA’99, pages 29–35, New York, 21–23, Feb. 1999. ACM. [8] T. Cormen, C. Leiserson, R. Rivest, and C. Stein. Introduction to algorithms. MIT Press, Cambridge, MA, 2001. [9] D. Corneil and Y. Perl. Clustering and domination in perfect graphs. Discrete Applied Mathematics, 9(1):27–39, 1984. [10] N. Een, A. Mishchenko, and N. Sorensson. Applying logic synthesis for speeding up SAT. Lect. N. Comp. Sci., 4501:272, 2007. [11] D. Eppstein. Small maximal independent sets and faster exact graph coloring. Lect. N. Comp. Sci., pages 462–470, 2001. [12] L. Ford and D. Fulkerson. Flow in networks. Princeton University Press, Princeton, NJ, 1962. [13] M. Golumbic. Algorithmic graph theory and perfect graphs. NorthHolland, 2004. [14] D. Kagaris and S. Tragoudas. Maximum weighted independent sets on transitive graphs and applications. Integration, the VLSI Journal, 27(1):77–86, 1999. [15] N. Kulkarni, N. Nukala, and S. Vrudhula. Minimizing area and power of sequential cmos circuits using threshold decomposition. In Proceedings of the International Conference on Computer-Aided Design, ICCAD ’12, pages 605–612, New York, NY, USA, 2012. ACM. [16] A. C. Ling, J. Zhu, and S. D. Brown. BddCut: Towards scalable symbolic cut enumeration. In Proc. ASP-DAC’07, pages 408–413, 23–26 Jan. 2007. [17] A. Mishchenko, R. Brayton, J.-H. R. Jiang, and S. Jang. Scalable don’tcare-based logic optimization and resynthesis. In Proc. FPGA ’09, pages 151–160, NY, 21–23 Feb. 2009. ACM. [18] A. Mishchenko, S. Chatterjee, R. Brayton, X. Wang, and T. Kam. Technology mapping with boolean matching, supergates and choices. [19] A. Mishchenko, S. Cho, S. Chatterjee, and R. Brayton. Combinational and sequential mapping with priority cuts. In Proc. ICCAD’07, pages 354–361, 4–8 Nov. 2007. [20] J. Moon and L. Moser. On cliques in graphs. Israel Journal of Mathematics, 3(1):23–28, 1965. [21] P. Pan and C.-C. Lin. A new retiming-based technology mapping algorithm for LUT-based FPGAs. In Proc. FPGA ’98, pages 35–42, New York, 22–24 Feb. 1998. ACM. [22] P. Pan and C. L. Liu. Optimal clock period FPGA technology mapping for sequential circuits. ACM TODAES, 3(3):437–462, 1998. [23] J. Peddersen, S. Shee, A. Janapsatya, and S. Parameswaran. Rapid embedded hardware/software system generation. In Int. Conf. on VLSI Design, pages 111–116, 2005. [24] S. Chatterjee. On Algorithms for Technology Mapping, PhD Thesis. 2007. [25] T. Takata and Y. Matsunaga. An efficient cut enumeration for depthoptimum technology mapping for lut-based fpgas. In Proceedings of the 19th ACM Great Lakes Symposium on VLSI, GLSVLSI ’09, pages 351–356, New York, NY, USA, 2009. ACM. K = 10, cone size = #cuts pruning after pruning time (s) 382,697 2.22 26,194,458 29.25 1,311,442,759 36.46 150,921,850 379.75 3,113,268,991 193.59 100 enum. time (s) 1.60 286.06 6510.5 4570.26 7046.68 

8cs.DS

Evolution of Covariance Functions for Gaussian Process Regression using Genetic Programming Gabriel Kronberger and Michael Kommenda arXiv:1305.3794v2 [cs.NE] 22 May 2013 School of Informatics, Communications and Media, University of Applied Sciences Upper Austria, Softwarepark 11, 4232, Hagenberg {gabriel.kronberger,michael.kommenda}@fh-hagenberg.at Abstract. In this contribution we describe an approach to evolve composite covariance functions for Gaussian processes using genetic programming. A critical aspect of Gaussian processes and similar kernel-based models such as SVM is, that the covariance function should be adapted to the modeled data. Frequently, the squared exponential covariance function is used as a default. However, this can lead to a misspecified model, which does not fit the data well. In the proposed approach we use a grammar for the composition of covariance functions and genetic programming to search over the space of sentences that can be derived from the grammar. We tested the proposed approach on synthetic data from two-dimensional test functions, and on the Mauna Loa CO2 time series. The results show, that our approach is feasible, finding covariance functions that perform much better than a default covariance function. For the CO2 data set a composite covariance function is found, that matches the performance of a hand-tuned covariance function. Keywords: Gaussian Process, Genetic Programming, Structure Identification 1 Introduction The composition of covariance functions is a non-trivial task and has been described as a black art [2]. On the one hand, it is critical to tune the covariance function to the data set, that should be modeled, because this is the primary option to integrate prior knowledge into the learning process [10]; on the other hand a lot of experience and knowledge about the modeled system is required to do this correctly. Frequently, and especially for multi-dimensional data sets it is far from obvious how the covariance function should be structured. In this work we discuss the composition of covariance functions for Gaussian processes, that can be used for nonparametric machine learning tasks e.g., for regression or classification [10]. In this context a Gaussian process is used as a Bayesian prior over functions, relating the input variables to the target variable. Gaussian process regression allows modeling of non-linear functional dependencies through different covariance functions, and produces posterior probability distribution estimates for the target values instead of point estimates only. 2 1.1 Evolution of Covariance Functions for Gaussian Process Regression Our Contribution The aim of this paper is to describe the idea of using a grammar for covariance functions and genetic programming to search for a good covariance function for a given data set. We also describe our prototype implementation using grammarguided tree-based GP, and finally, present results as a proof-of concept. We have not yet evaluated the difficulty of this problem for genetic programming, and in particular, if GP suited well for this kind of problem. The results of our experiments indicate that the idea is feasible, producing good covariance functions for low-dimensional data sets. 1.2 Previous Work In a very recent contribution, the problem of structure identification for covariance functions has been approached using a grammar, or rather a set of rewriting rules, as a basis for searching over composite covariance functions for Gaussian processes [2]. This approach is actually very similar to our work; the main difference is, that in our work we use genetic programming to search over the set of possible structures, while Duvenaud et al. enumerate over composite functions, starting with standard functions. Another recent contribution discusses more flexible families of covariance functions, instead of composing covariance functions from simple terms [12]. Also related is earlier work that describes additive Gaussian processes [3], which are equivalent to a weighted additive composition of base kernels, but can be calculated efficiently. Genetic programming has been used previously to evolve kernel functions for SVMs with mixed results [4], [6]. The latest contribution found that genetic programming was able to “rediscover multiple standard kernels, but no significant improvements over standard kernels were obtained” [7]. These results can, however, not be transfered directly to Gaussian processes because of several major differences between Gaussian processes and SVMs. In particular, in the case of Gaussian processes hyper-parameters are optimized using a ML-II approach, in contrast to SVMs, where hyper-parameter values are usually tuned using crossvalidation and grid-search. Additionally, in contrast to all other previous work, simple embeddings of covariance functions by masking dimensions are supported. 2 Gaussian Processes A Gaussian process is a non-parametric model that produces predictions solely from the specified mean and covariance functions and the available training data [10]. The inference of function values f ∗ for observed input values X ∗ based on observations of y and X involves the calculation of the covariance matrices K(X, X) and K(X, X ∗ ) and inference from the multi-dimensional Gaussian shown in Equation 1.       y m(X) K(X, X) + σ 2 I K(X, X ∗) ∼ N , (1) f∗ m(X ∗ ) K(X ∗ , X) K(X ∗ , X ∗ ) Evolution of Covariance Functions for Gaussian Process Regression 3 The term σ 2 I is necessary to account for Gaussian distributed noise with variance σ 2 . From this definition it follows that the posterior for f ∗ is again a multi-dimensional Gaussian. For model selection and hyper-parameter learning the marginal likelihood p(y|X) must be calculated. The model is a multidimensional Gaussian so an analytical form of the likelihood can be derived. Calculation of the marginal likelihood requires a matrix inversion and, thus, has asymptotic complexity O(n3 ). Usually, the covariance function K(x, x′ ) has hyper-parameters that must optimized. This is often accomplished in a simple ML-II fashion, optimizing the hyper-parameters w.r.t. the likelihood using a quasi-Newton method (e.g., BFGS). Since the gradients of the marginal likelihood for the hyper-parameters can be determined with an additional computational complexity of O(n2 ) for each hyper-parameter, it is feasible to use gradient-based methods. The drawback is that the likelihood is typically multimodal, and especially for covariances with many hyper-parameters (e.g., ARD) the optimizer can converge to a local optimum. Thus, it is typically suggested to execute several random restarts. A better solution would be to include priors on the hyper-parameters and optimizing w.r.t. posterior distribution (MAP). However, this can only be accomplished using a MCMC approach which is computationally expensive. Frequently used covariance functions for Gaussian processes include the linear, polynomial, squared exponential (SE), rational quadratic (RQ) and the Matérn function. Covariance functions can be combined to more complex covariance functions, for instance as products or sums of different covariance functions [10]. 3 Genetic Programming Genetic programming generally refers to the automatic creation of computer programs using genetic algorithms [8]. The basic principle is to evolve variablelength structures, frequently symbolic expression trees, which represent potential solutions to the problem. One of the most prominent applications of genetic programming is symbolic regression, the synthesis of regression models without a predetermined structure. Genetic programming makes it possible to optimize the structure of solutions in combination with their parameters. Thus, it should also be possible to synthesize composite covariance functions with genetic programming. In the following, we use a grammar-guided genetic programming system to make sure that only valid covariance functions are produced. A good survey of grammar-guided genetic programming is given in [9]. 4 Grammar for Covariance Functions The grammar for covariance functions has been derived from the rules for the composition of kernels as e.g., discussed in [10]. It should be noted that the grammar shown below is not complete, meaning that several constructions that would 4 Evolution of Covariance Functions for Gaussian Process Regression lead to a valid covariance function are not possible1 . The following represents the grammar G(Cov) for covariance functions in EBNF notation2 : Cov -> "Prod" "(" Cov { Cov } ")" | "Sum" "(" Cov { Cov } ")" | "Scale" Cov | "Mask" BitVector Cov | TerminalCov . TerminalCov -> "SE" | "RQ" | "Matern1" | "Matern3" | "Matern5" | "Periodic" | "Linear" | "Constant" | "Noise" . BitVector -> "[" {"0" | "1" } "]" . The functions Prod and Sum produce the product and sum of multiple covariance functions, which can again be composite covariance functions. The scale operator can be used to add a scaling factor to any covariance function. The Mask operator selects a potentially empty subset of input variables from all possible input variables. The non-terminal symbol BitVector can be derived to a list of zeros and ones. The bit vector is used to mask selected dimensions in the data set, effectively reducing the dimensionality. The length of the bit mask has to match the total number of dimensions; this is checked when the resulting covariance function is evaluated. Finally, the non-terminal symbol TerminalCov can be derived to a range of default covariance functions. Currently, we only included isometric covariance functions, but other covariance functions can be added to the grammar easily. The grammar does not include the hyper-parameters, because they are not optimized by genetic programming. Instead, hyper-parameters are optimized for each potential solution, using a gradient-descent technique. 5 Experiments For the experiments we implemented Gaussian processes, a set of commonly used covariance functions, and the grammar for covariance functions in HeuristicLab3 [11] which already provides an implementation of grammar-guided treebased genetic programming. The aim of the experiments presented in this contribution is mainly to test the feasibility of the idea. Two different types of data sets are used for the experiments, and the forecasts of the synthesized covariance functions are compared to a set of default covariance functions and also to hand-tuned covariance functions. The first data set is the univariate Mauna Loa atmospheric CO2 time series. This data set has been chosen, because a hand-tuned covariance function for this data set is presented in [10]. For the second experiment we created several synthetic data sets sampled randomly from two-dimensional Gaussian process priors shown in Equation 2. The data generated from these functions are difficult to model with a single isometric covariance function. Multiple covariance functions have to be combined and the correct dimension masking vectors have 1 2 3 One example is vertical scaling of covariance functions: K ′ (x, x′ ) = a(x)K(x, x′ )a(x′ ) The grammar is largely based on the capabilities of the GPML package by Rasmussen and Nickisch, http://gaussianprocess.org/gpml/code. HeuristicLab version 3.3.8 is available from http://dev.heuristiclab.com/ Evolution of Covariance Functions for Gaussian Process Regression 5 to be identified. Each data set contains 882 samples of the function on a regular two-dimensional grid. SE+RQ(x, x′ ) = SE(x0 , x′0 ) + RQ(x1 , x′1 ) SE+Matérn(x, x′ ) = SE(x0 , x′0 ) + Matérn1(x1 , x′1 ) SE+Periodic(x, x′ ) = SE(x0 , x′0 ) + RQ(x1 , x′1 ) 5.1 (2) Genetic Programming Parameter Settings Training of Gaussian processes is computationally expensive, and because it is necessary to optimize the hyper-parameters for each evaluated covariance function the run time of the genetic programming algorithm grows quickly. Therefore, we used very restrictive parameter settings, in particular a small population size of only 50 individuals. All other parameter settings are shown in Table 1. Table 1. Genetic programming parameter settings for all experiments. Parameter Population size Max. length / height Initialization Parent selection Mutation rate ML-II iterations Offspring selection [1] Max. selection pressure Max. generations 5.2 Value 50 25 / 7 PTC2 gender-specific (proportional + random ) 15% 50 strict (success ratio = 1, comparison factor = 1) 100 20 Results on Mauna Loa CO2 Data Set The results for the CO2 time series are positive. The algorithm was able to consistently find covariance functions that fit well in the training period (1958 – 2004), accurate forecasts over the testing period (2004 – 2012). The structures of two exemplary solutions are shown in Equation 3. The first solution (K1) is actually very similar to the hand-tuned covariance solution proposed in [10]. The second covariance function is more complex and has only a slightly better likelihood. Unfortunately, genetic programming often leads to overly complex solutions which is a critical drawback of our approach. Both solutions have been found after only 800 evaluated solution candidates and achieve a negative loglikelihood of 129.8 and 116, respectively. The correlation coefficients for the forecasts in the test partition are above 0.99. Figure 1 shows the output of the first model. 6 Evolution of Covariance Functions for Gaussian Process Regression K1(x, x′ ) = SE(x, x′ ) + Periodic(x, x′ ) + Matérn1(x, x′ )+ SE(x, x′ ) + Matérn5(x, x′ ) + Const K2(x, x′ ) = Matérn3(x, x′ ) ∗ Perioric(x, x′ ) ∗ RQ(x, x′ ) ∗ (Matérn1(x, x′ ) + Matérn3(x, x′ ) + Matérn5(x, x′ ) + (3) Perioric(x, x′ ) + Linear(x, x′ )) ∗ (Matérn1(x, x′ ) + Matérn3(x, x′ ) + RQ(x, x′ )) Estimated Values (training) Estimated Values (test) 80 60 40 20 0 -20 -40 0 149 298 447 596 745 Fig. 1. The output and forecast for the Mauna Loa CO2 time series of a Gaussian process using the first evolved covariance function (K1) shown in Equation 3. 5.3 Results on Synthetic Data Sets The results for the synthetic two-dimensional data set are shown in Table 2. In this experiment we trained multiple Gaussian process models using several frequently used covariance functions. We trained many models using random restarts for each data set and covariance function, and report the best negative log-likelihood for each pair. As expected, the models with the isometric covariance functions do not fit well. In contrast, the composite covariance functions produced by genetic programming fit much better. For comparison, we also report the negative log-likelihood, that can be achieved with the optimal covariance function for each data set. In these experiments, the exact structure of the covariance could not be rediscovered, thus, the evolved functions are worse than the optimal solution. 6 Summary and Discussion In this contribution we described an approach for the synthesis of composite covariance function for Gaussian processes using grammar-guided genetic programming. In the proposed approach a set of commonly used covariance functions is Evolution of Covariance Functions for Gaussian Process Regression 7 Covariance Table 2. Best negative log-likelihood achieved for the three synthetic two-dimensional test functions, with default covariance functions and with evolved composite covariance functions. SE RQ Periodic Matérn Evolved Optimal Problem instance SE+RQ SE+Matérn SE+Periodic -204 -492 440 -272 -492 103 -221 -492 479 -27 31 304 -803 -760 -640 -2180 -2187 -2131 used to compose more complex covariance functions, using sums or products of several covariance functions. The set of valid covariance functions is defined via a grammar and genetic programming is used to search the space of possible derivations from this grammar. The hyper-parameters of covariance functions are not subject to the evolutionary search, but are optimized w.r.t. the likelihood using a standard gradient-descent optimizer (i.e., LBFGS). The proposed approach was tested on two types of low-dimensional problems as a proof of concept. We found, that for the univariate Mauna Loa CO2 time series it is possible to consistently find good covariance functions with genetic programming. The identified solutions perform as well as a hand-tuned covariance function for this problem. The results for our two-dimensional synthetic functions show that it is possible to find composite covariance functions, which perform much better than default covariance functions on these data sets. In contrast to previous work by the genetic programming community [7], which focused mainly on kernel synthesis for SVMs, this contribution discusses kernel synthesis for Gaussian processes, which are non-parametric fully Bayesian models. For Gaussian process models the hyper-parameters can be optimized with a standard gradient-descent approach, and it is not strictly necessary to execute cross-validation [10]. Previous work either used grid-search and crossvalidation to tune hyper-parameters, which is very computationally expensive, or did not consider hyper-parameter optimization at all. Additionally, we are using a grammar to compose covariance functions from simple covariance functions instead of evolving the full function. In the statistics community, a very recent contribution has also discussed the usage of grammars for the composition of covariance functions [2]. The main difference to this work is that here genetic programming is used to search over the derivations of the grammar. Another relevant difference is that the grammar in this contribution also supports simple embeddings through the masking function. It should be noted, that we have not yet analyzed if genetic programming is well suited for this task, and in particular we did not compare the approach to simple enumeration or random search. One question that remains for future work is whether composed covariance functions also work well for data sets with more variables. We have observed that 8 Evolution of Covariance Functions for Gaussian Process Regression simple covariance functions often work very well, and tuned covariance functions do not have a strong beneficial effect for these data sets. Another interesting topic for future research is to look at alternative ways for searching over the space of covariance functions defined by a grammar. Recently, an interesting approach has been described that uses variational methods for Bayesian learning of probabilistic context free grammars for this task [5]. This idea could be especially useful for Bayesian models such as Gaussian processes. Acknowledgments The authors would like to thank Jeffrey Emanuel for the initial idea leading to this contribution. This work has been supported by the Austrian Research Promotion Agency (FFG), on behalf of the Austrian Federal Ministry of Economy, Family and Youth (BMWFJ), within the program “Josef Ressel-Centers”. References 1. Affenzeller, M., Winkler, S., Wagner, S., Beham, A.: Genetic Algorithms and Genetic Programming: Modern Concepts and Practical Applications. Numerical Insights, CRC Press, Singapore (2009) 2. Duvenaud, D., Lloyd, J.R., Grosse, R., Tenenbaum, J.B., Ghahramani, Z.: Structure Discovery in Nonparametric Regression through Compositional Kernel Search. ArXiv e-prints (Feb 2013) 3. Duvenaud, D., Nickisch, H., Rasmussen, C.E.: Additive Gaussian processes. arXiv preprint arXiv:1112.4394 (2011) 4. Gagne, C., Schoenauer, M., Sebag, M., Tomassini, M.: Genetic programming for kernel-based learning with co-evolving subsets selection. In: Proc. of PPSN IX. LNCS, vol. 4193, pp. 1008–1017. Springer-Verlag (2006) 5. Hasegawa, Y., Iba, H.: Latent variable model for estimation of distribution algorithm based on a probabilistic context-free grammar. IEEE Transactions on Evolutionary Computation 13(4), 858–878 (Aug 2009) 6. Howley, T., Madden, M.G.: An evolutionary approach to automatic kernel construction. In: Proc. of ICANN 2006. LNCS, vol. 4132, pp. 417–426. Springer-Verlag (2006) 7. Koch, P., Bischl, B., Flasch, O., Beielstein, T., Weihs, C., Konen, W.: Tuning and evolution of support vector kernels. Evolutionary Intelligence 5, 153–170 (2012) 8. Koza, J.R.: Genetic programming: on the programming of computers by means of natural selection. MIT Press, Cambridge, MA, USA (1992) 9. McKay, R.I., Hoai, N.X., Whigham, P.A., Shan, Y., O’Neill, M.: Grammar-based genetic programming: a survey. Genetic Programming and Evolvable Machines 11(3/4), 365–396 (Sep 2010) 10. Rasmussen, C.E., Williams, C.K.: Gaussian Processes for Machine Learning. MIT Press (2006) 11. Wagner, S.: Heuristic optimization software systems – Modeling of heuristic optimization algorithms in the HeuristicLab software environment. Ph.D. thesis, Institute for Formal Models and Verification, Johannes Kepler University, Linz (2009) 12. Wilson, A.G., Prescott Adams, R.: Gaussian Process Covariance Kernels for Pattern Discovery and Extrapolation. ArXiv e-prints (Feb 2013) 

9cs.NE

DETECTING THE INTEGER DECOMPOSITION PROPERTY AND EHRHART UNIMODALITY IN REFLEXIVE SIMPLICES arXiv:1608.01614v2 [math.CO] 8 Dec 2017 BENJAMIN BRAUN, ROBERT DAVIS, AND LIAM SOLUS Abstract. A long-standing open conjecture in combinatorics asserts that a Gorenstein lattice polytope with the integer decomposition property (IDP) has a unimodal (Ehrhart) h∗ -polynomial. This conjecture can be viewed as a strengthening of a previously disproved conjecture which stated that any Gorenstein lattice polytope has a unimodal h∗ -polynomial. The first counterexamples to unimodality for Gorenstein lattice polytopes were given in even dimensions greater than five by Mustaţǎ and Payne, and this was extended to all dimensions greater than five by Payne. While there exist numerous examples in support of the conjecture that IDP reflexives are h∗ -unimodal, its validity has not yet been considered for families of reflexive lattice simplices that closely generalize Payne’s counterexamples. The main purpose of this work is to prove that the former conjecture does indeed hold for a natural generalization of Payne’s examples. The second purpose of this work is to extend this investigation to a broader class of lattice simplices, for which we present new results, data from computational experiments, and open problems. 1. Introduction Two well-studied properties of the coefficients of polynomials with non-negative integer coefficients are symmetry and unimodality. A polynomial a0 + a1 z + · · · + ad z d with ad 6= 0 is called symmetric if ai = ad−i for all i ∈ [d]. A polynomial a0 + a1 z + · · · + ad z d with ad 6= 0 is called unimodal if there exists an index j such that ai ≤ ai+1 for all i < j and ai ≥ ai+1 for all i ≥ j. Given a finitely generated C-algebra A = ⊕∞ i=0 Ai , we say A is standard if A is generated by A1 and semistandard if A is integral over the subalgebra generated by A1 . The Hilbert series of a graded algebra A is the generating function P ∞ j X j hj z i . dimC (Ai )z = H(A; z) := (1 − z)dim(A) i=0 P j The polynomial h(A; z) := j hj z is called the h-polynomial of A. When A is Cohen-Macaulay, each hj ∈ Z≥0 . An active topic of research is developing combinatorial interpretations of the h-polynomial coefficients and their distribution. It is known that the h-polynomial of A is symmetric when A is Gorenstein, but the unimodality property is more subtle. A long-standing open problem is the following conjecture posed by Brenti [5, Conjecture 5.1], inspired by a conjecture of Stanley [18]. It also appears in [10, Conjecture 1.5]. Conjecture 1.1 (Brenti, [5]). For a standard graded Gorenstein integral domain A, h(A; z) is unimodal. A subset P ⊂ Rn is a d-dimensional convex lattice polytope if P arises as the convex hull of finitely many points in Zn that together span an affine d-flat in Rn . The Ehrhart function of P is the lattice point enumerator i(P ; t) := |tP ∩ Zn |, where tP := {tp : p ∈ P } denotes the tth dilate of the polytope P . It is well-known [8] that i(P ; t) is a polynomial in t of degree d, and the corresponding Ehrhart series of P is the rational function X h∗ + h∗1 z + · · · + h∗d z d EhrP (z) := i(P ; t)z t = 0 , (1 − z)d+1 t≥0 where the coefficients h∗0 , h∗1 , . . . , h∗d are all nonnegative integers [17]. The polynomial h∗ (P ; z) := h∗0 + h∗1 z + · · · + h∗d z d is called the (Ehrhart) h∗ -polynomial of P . Analogously, the coefficient vector h∗ (P ) := (h∗0 , h∗1 , . . . , h∗d ) is called the (Ehrhart) h∗ -vector of P . We will often say P is h∗ -unimodal whenever h∗ (P ; z) is unimodal. Since h∗ (P ; z) arises via the enumeration of combinatorial data and has only nonnegative Date: 8 December 2017. 2010 Mathematics Subject Classification. Primary: 52B20, 05E40, 05A20, 05A15. Benjamin Braun was partially supported by grant H98230-16-1-0045 from the U.S. National Security Agency. Liam Solus was partially supported by an NSF Division of Mathematical Sciences Postdoctoral Fellowship (DMS - 1606407). 1 2 BENJAMIN BRAUN, ROBERT DAVIS, AND LIAM SOLUS integer coefficients, researchers have studied combinatorial interpretations of the h∗ -coefficients and their distribution. A current challenge in Ehrhart theory is to understand the geometric properties of a polytope that are necessary and/or sufficient for h∗ (P ; z) to be simultaneously symmetric and unimodal [3]. While Ehrhart series arise from a combinatorial context in polyhedral geometry, there is a well-known connection to commutative algebra. The cone over P is cone (P ) := spanR≥0 {(1, p) : p ∈ P } ⊂ R × Rn where we consider the new variable to be indexed at 0, i.e. x ∈ R×Rn = R1+n is written x = (x0 , x1 , . . . , xn ). Given a lattice polytope P in Rn , the semigroup algebra associated to P is C[P ] := C[xv : v ∈ cone (P )∩Z1+n ]. mn 0 We grade C[P ] by deg(xm 0 · · · xn ) = m0 . With this grading, C[P ] is a semistandard semigroup algebra, where the algebra generated by C[P ]1 is C[xv : v ∈ (1, P ) ∩ Z1+n ], i.e. the semigroup algebra generated by integer points in (1, P ). Given this, Conjecture 1.1 manifests itself in the setting of polytopes that have symmetric h∗ -polynomials and the integer decomposition property, which is defined as follows. Definition 1.2. A polytope P has the integer decomposition property, or is IDP, if for P every positive integer m and each x ∈ mP ∩ Zn , there exist m points x1 , . . . , xm ∈ P ∩ Zn for which x = xi . It is straightforward to verify that a lattice polytope P is IDP if and only if C[P ] is standard. Regarding symmetry, a lattice polytope P ⊆ Rn containing 0 in its interior is called reflexive if its polar P ∗ := {y ∈ Rn : xT y ≤ 1 for all x ∈ P } is also a lattice polytope. In [12], it is shown that a lattice polytope P ⊂ Rn is reflexive if and only if P is full-dimensional (i.e. d = n), the origin is in its interior, and h∗ (P ; z) is symmetric. Reflexive polytopes are a special case of the full-dimensional polytopes P ⊂ Rn for which h∗ (P ; z) is symmetric, known as Gorenstein polytopes. In this setting, Conjecture 1.1 translates to the following statement, which is often attributed to Hibi and Ohsugi [15]. Conjecture 1.3. If P is Gorenstein and IDP, then h∗ (P ; z) is unimodal. As in the case of graded Cohen-Macaulay algebras, unimodality of h∗ (P ; z) has proven to be a more elusive property to characterize than symmetry when using only the geometry of P . Since the symmetry property of h∗ (P ; z) reduces the number of inequalities to be verified by one-half, there has been much research into when h∗ (P ; z) is both symmetric and unimodal [3]. It was shown by Mustaţǎ and Payne [13] that there exist reflexive polytopes whose h∗ -polynomials are not unimodal. These results were then extended to reflexive simplices in every dimension greater than five by Payne [16]. This disproved the following (earlier) conjecture of Hibi [11]. (Disproved) Conjecture 1.4 (Hibi, [11]). If P is Gorenstein then h∗ (P ; z) is unimodal. Thus, Conjecture 1.3 is a reasonable strengthening of Conjecture 1.4 to consider in light of the counterexamples presented by Payne [16] and its previous proposal in the algebraic context by Brenti (Conjecture 1.1). Indeed, one may show that none of the counterexamples provided by Mustaţǎ and Payne are IDP. While the literature is ripe with examples of lattice polytopes supporting Conjecture 1.3, surprisingly, there has been little investigation as to whether or not the conjecture holds for reflexive lattice simplices that closely generalize the counterexamples constructed by Payne. The purpose of the present paper is to carefully identify a natural generalization of Payne’s examples, and prove that Conjecture 1.3 does indeed hold for this family of reflexive lattice simplices. The generalization of Payne’s examples that we will consider here is a subcollection Q of the reflexive lattice simplices whose associated toric varieties are weighted projective spaces. Consequently, each simplex ∆ ∈ Q can be associated to an integer partition q(∆) by using a classification system for reflexive simplices developed by Conrads [7]. The collection Q can be stratified by the number of parts used in the integer partitions q(∆). We will see that Payne’s examples live within the subcollection of Q for which q(∆) is an integer partition with two parts, making this subcollection a close generalization of Payne’s counterexamples. Our first main result is to show that Conjecture 1.3 holds for this subcollection. Our second contribution is to show that Conjecture 1.3 holds for many simplices in Q for which q(∆) is an integer partition with three parts, and to provide both computational and theoretical evidence of the challenges that exist for proving Conjecture 1.3 completely in this setting. The remainder of this paper is outlined as follows: In Subsection 2.1, we recall the counterexamples to Conjecture 1.4 presented by Payne in [16], and we present the generalizing family Q. In Subsections 2.2 DETECTING THE INTEGER DECOMPOSITION PROPERTY 3 and 2.3, we present a formula for the h∗ -polynomial and a characterization of IDP for all simplices in Q, respectively. In Section 3, we will prove some first results on IDP and h∗ -unimodality for simplices in Q whose associated integer partitions q(∆) have a fixed set of parts. In Section 4, we prove Conjecture 1.3 for our closest generalization of Payne’s examples; i.e., for all simplices ∆ ∈ Q whose associated partitions q(∆) have two parts. In Section 5, we extend the results of Section 4 by proving Conjecture 1.3 holds for simplices in Q whose integer partitions q(∆) contain three parts in a number of fundamental cases. We then end with a brief discussion of future directions in Section 6. 2. IDP and Ehrhart Theory for the Reflexive Simplices ∆(1,q) 2.1. The reflexive simplices ∆(1,q) . Given a polytope P ⊆ Rn with vertices v0 , . . . , vk , the height of a point x ∈ cone (P ) is its first coordinate x0 . If P is a lattice simplex with vertices v0 , v1 , . . . , vn , the fundamental parallelepiped of cone (P ) is ) ( n X λi (1, vi ) : 0 ≤ λi < 1 . ΠP := i=0 The fundamental parallelepiped tiles cone (P ) through nonnegative integer combinations of {(1, vi )}ni=0 . Consequently, if P is a lattice simplex then X (1) h∗ (P ; z) = z m0 . m∈ΠP ∩Zn In this paper, we will focus on a family of reflexive simplices Q that can be defined as follows. Let q = (q1 , q2 , . . . , qn ) be a weakly increasing sequence of positive integers satisfying the condition X qj | (1 + qi ) i6=j for all j = 1, . . . , n. For such a vector, the simplex ( ∆(1,q) := conv e1 , e2 , . . . , en , − n X i=1 qi ei ) , where ei ∈ Rn is the i-th standard basis vector, is reflexive. When the q-vector is understood, we will often P label the vertices of ∆(1,q) as vi := ei and v0 := − ni=1 qi ei . The collection of simplices Q is important in algebra and geometry since it is contained in the family of simplices whose associated toric varieties are weighted projective spaces [7]. Combinatorially, they are significant since they contain the counterexamples to Conjecture 1.4 developed by Payne [16]. In particular, Payne showed that for integers r ≥ 0, s ≥ 3, and k ≥ r + 2 the reflexive simplex ∆(1,q) for (2) q = (1, 1, . . . , 1, s, s, . . . , s) . | {z } | {z } sk−1 times r+1 times is not h∗ -unimodal. It follows from (1) that the normalized volume of ∆(1,q) (i.e. the value h∗ (∆(1,q) ; 1) = n! Vol(∆(1,q) )) is equal to 1 + q1 + · · · + qn [14, Proposition 4.4]. Consequently, the simplices ∆(1,q) are naturally stratified via their normalized volume by way of integer partitions; i.e., for every ∆(1,q) there exists a partition of the integer n! Vol(∆(1,q) ) − 1 with parts r1 < r2 < · · · < rk such that q = (r1x1 , r2x2 , . . . , rkxk ) := (r1 , r1 , . . . , r1 , r2 , r2 , . . . , r2 , . . . , rk , rk , . . . , rk ). | | {z } | {z } {z } x1 times x2 times xk times So as to speak formally from this perspective, we make the following definition. Definition 2.1. We say that both q and ∆(1,q) are supported by the vector r = (r1 , . . . , rk ) (or the integers therein) if there exist positive integers r1 < r2 < · · · < rk and x1 , . . . , xk such that q = (q1 , . . . , qn ) = (r1x1 , r2x2 , . . . , rkxk ). An r-vector is any vector of positive integers r = (r1 , . . . , rk ) in which r1 < r2 < · · · < rk . 4 BENJAMIN BRAUN, ROBERT DAVIS, AND LIAM SOLUS It follows that the natural generalization of Payne’s examples are those ∆(1,q) supported by two integers. In Section 4, we will prove Conjecture 1.3 for this generalization of Payne’s examples. In Section 3, we will prove some results about q-vectors with a fixed support r, and in Section 5, we will extend many of the results in Section 4 to those simplices supported on three integers. 2.2. The h∗ -polynomials for ∆(1,q) . In this subsection, we establish basic properties of ∆(1,q) with respect to h∗ -polynomials. We begin with a formula for h∗ (∆(1,q) ; z) in terms of the entries of the vector q. Theorem 2.2. The h∗ -polynomial of ∆(1,q) is h∗ (∆(1,q) ; z) = q1 +···+q X n z w(b) b=0 where w(b) := b − n  X i=1 Proof. Let v1 , v2 , . . . , vn , v0 denote the columns  1 1  1 0   0 1   0 0   .. ..  . . 0 qi b 1 + q1 + · · · + qn  . of 1 0 0 1 .. . 0 0 ··· ··· ··· ··· .. . 1 0 0 0 .. . 1 −q1 −q2 −q2 .. . · · · 1 −qn      .    By [14, Proposition 4.4] and Cramer’s rule, every point in Π∆(1,q) is of the form p := n X λi (1, vi ) i=0 where 0 ≤ λi < 1 for all i and λi = follow that for i ≥ 1 bi 1+q1 +···+qn for some bi = 0, 1, . . . , q1 + · · · + qn . If p ∈ Z1+n , then it must λi = qi λ0 − ⌊qi λ0 ⌋ . Hence, the choice of λ0 determines the point p, and we write b λ0 := . 1 + q1 + · · · + qn The height of the point corresponding to a given value of b is the first coordinate, which is easily computed to be  n  X qi b . b− 1 + q1 + · · · + qn i=1 Recalling equation (1) completes the proof.  2.3. The Integer Decomposition Property and ∆(1,q) . The following theorem (Theorem 2.3) provides a characterization of IDP reflexive ∆(1,q) in terms of the vector q. Our main tool for investigating ∆(1,q) is given by Corollary 2.4 below, a necessary but not sufficient relaxation of Theorem 2.3. Theorem 2.3. The reflexive simplex ∆(1,q) is IDP if and only if for every j = 1, . . . , n, for all b = 0, 1, . . . , qj − 1 satisfying P  X   1 + i6=j qi bqi − ≥2 (3) b qj qj i6=j there exists a positive integer c < b satisfying the following equations, where the first is considered for all 1 ≤ i ≤ n with i 6= j:       bqi cqi (b − c)qi (4) − = , and qj qj qj DETECTING THE INTEGER DECOMPOSITION PROPERTY (5) c  1+ P i6=j qi qj  − X  cqi  i6=j qj 5 = 1. Proof. Recall that IDP for a lattice simplex P is equivalent to the property that every lattice point in ΠP arises as a sum of lattice points in (1, P ). Let g ∈ Π∆(1,q) . Since ∆(1,q) is reflexive, we may subtract
(1, 0, 0, . . . , 0) from g until we reach a lattice point p on the boundary of cone ∆(1,q) . This point p must lie in the fundamental parallelepiped for a face of ∆(1,q) , hence ∆(1,q) is IDP if and only if every facet of ∆(1,q) is IDP. Observe that the facet with vertices {e1 , . . . , en } is unimodular, hence it is IDP. The remaining facets are of the form ) ( n X qi ei Fj := conv e1 , . . . , ej−1 , ej+1 , . . . , en , − i=1 for j = 1, . . . , n. Thus we areP looking for necessary and sufficient conditions for Fj to be IDP. For i ∈ [n]\{j}, n set vi := ei , and set v0 := − i=1 qi ei . Every lattice point in ΠFj is of the form X r= λi vi . 0≤i≤n i6=j Since the vector (1, ej ) is not a summand for r, and since r is an integer point, λ0 · (−qj ) must be an integer. Hence, λ0 = b/qj for some b in {0, 1, 2, . . . , qj − 1}. Note that if b = 0, r is in the fundamental parallelepiped for the facet with vertices {e1 , . . . , en }, and hence equal to the zero vector. We next show that for every choice of b between 1 and qj − 1, setting λ0 = b/qj yields a unique lattice point of ΠFj . Since every entry of r is integral and 0 ≤ λi < 1, it follows that   bqi bqi λi = . − qj qj Thus, as in the proof of Theorem 2.2, the value of λ0 determines the remaining λi ’s. When λ0 = b/qj , the height of the resulting point is X  bqi  bqi  b , + − qj qj qj 1≤i≤n i6=j which is an integer if and only if P   1 + 1≤i≤n qi X b bqi i6=j  + = b qj qj qj 1≤i≤n i6=j is an integer. Since ∆(1,q) is reflexive, we know that qj divides 1 + X qi , and thus this height is an integer. 0≤i≤n i6=j Hence, for every choice of b above, r is a lattice point given by   1+P q  P j k bqi i6=j i − i6=j b qj qj  k j  bq 1  − qj   ..  . k  j  bqj−1  − qj  r = rb :=   −b   k j  bq − qj+1  j   ..   j. k − bqqjn            .          6 BENJAMIN BRAUN, ROBERT DAVIS, AND LIAM SOLUS The simplex Fj is IDP if and only if for every rb at height greater than or equal to two, there exists an rc ∈ ΠFj at height one and an rd ∈ ΠFj such that rb − rc = rd . Given b and considering this vector equation entry-by-entry, this is equivalent to solving the following system with integers c and d between 1 and qj − 1:   P P  X   X    1 + i6=j qi 1 + i6=j qi bqi  dqi  (6) − − =d −1 + b qj qj qj qj i6=j  (7) (8) c  1+ P i6=j qi qj i6=j      bqi cqi dqi − = , i 6= j qj qj qj  − X  cqi  qj i6=j =1 Substituting (8) for the “1” on the left-hand side of (6), and substituting (7) for each of the “ j dqi qj k ” on the right-hand side of (6), yields b − c = d. Thus, we can reduce this system of equations to an equivalent system of equations:       bqi cqi (b − c)qi (9) − = , i 6= j qj qj qj (10) c  1+ P i6=j qj qi  − X  cqi  qj i6=j =1 Hence, ∆(1,q) is IDP if and only if each of its facets is IDP, which occurs if and only if for every j = 1, . . . , n, for all b = 1, . . . , qj − 1 corresponding to a lattice point at height at least two there exists a positive integer c solving (9) and (10).  At this point, we will frequently write a b to denote the fractional part of a/b. Corollary 2.4. Let ∆(1,q) be reflexive. If ∆(1,q) is IDP then for all j = 1, 2, . . . , n X  qi  1 = 1. + qj qj i6=j For any vector q of increasing positive integers that satisfies these equations for all j = 1, 2, . . . , n, the simplex ∆(1,q) is reflexive. However, this condition is necessary but not sufficient for ∆(1,q) to be IDP. Proof. By Theorem 2.3, ∆(1,q) is IDP if for all j = 1, . . . , n we have that for all b = 0, 1, . . . , qj − 1 satisfying equation (3) there exists 0 < c < b satisfying equations (4) and (5). Therefore, if b = 1 satisfies (3) then trivially no such c exists, meaning ∆(1,q) cannot be IDP. Thus, whenever ∆(1,q) is IDP, b = 1 must not satisfy (3). Moreover, since the left-hand side of (3) is equivalent to X  qi  1 + qj qj i6=j and q1 , q2 , . . . , qn > 0 then if (3) is not satisfied for b = 1 it must be that X  qi  1 = 1. + qj qj i6=j To see that these equations imply reflexivity of ∆(1,q) , note that the j-th equation verifies that the divisibility P condition qj | (1 + i6=j qi ) holds. As this holds for all j, we conclude that ∆(1,q) is reflexive. DETECTING THE INTEGER DECOMPOSITION PROPERTY 7 Finally, to show that this condition is not sufficient to establish IDP, consider the vector q = (2, 2, 15, 20, 20). It is straightforward to compute that the lattice points in ∆(1,q) are the columns of the following matrix.   1 0 0 0 0 −2 0 0 0 0 0 0 −1 −1 −1  0 1 0 0 0 −2 0 0 0 0 0 0 −1 −1 −1     0 0 1 0 0 −15 0 0 −1 −1 −2 −3 −7 −8 −9  .    0 0 0 1 0 −20 0 −1 −1 −2 −3 −4 −10 −11 −12  0 0 0 0 1 −20 0 −1 −1 −2 −3 −4 −10 −11 −12 Further, the point       1 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 0 0 1 −2 −2 −15 −20 −20   1/15 1/15 0 2/3 2/3 8/15               =      −1 −1 −8 −10 −10     ∈ 2∆(1,2,2,15,20,20) .   It is straightforward to verify that this point is not the sum of exactly two lattice points in ∆(1,2,2,15,20,20) , and thus this simplex is not IDP. However, (2, 2, 15, 20, 20) satisfies our linear system.  Remark 2.5. An equivalent formulation of Corollary 2.4 is that if q corresponds to a reflexive IDP simplex, then for each j = 1, . . . , n the sum X 1+ (qi mod qj ) i6=j must be equal to qj . In order for the reflexive condition to be satisfied, this sum must be equal to a multiple of qj . Thus, this necessary condition for IDP reflexive is a strengthening of the divisibility condition characterizing reflexivity. Remark 2.6. Corollary 2.4 implies that for IDP ∆(1,q) and b = 2, 3, . . . , qj − 1 the choice of c = 1 is always a potential solution to (4), as it satisfies equation (5). Example 2.7 (Revisiting Payne’s simplices). Recall that Payne presented the simplices ∆(1,q) with q as defined in equation (2) as counterexamples to Conjecture 1.4. Using Proposition 2.2, we see that s(k+r+1)−1 ∗ h (∆(1,q) ; z) = X z b−(r+1)⌊sb/s(k+r+1)⌋ b=0 from which it is straightforward to compute that h∗ (∆(1,q) ; z) = (1 + z k + z 2k + · · · + z (s−1)k )(1 + z + z 2 + · · · + z k+r ) . Non-unimodality for most of these simplices follows immediately from the conditions on r, b, and s. One can verify that ∆(1,q) is not IDP using the property that h∗1 = 1. Since this implies that the only lattice points in ∆(1,q) are the vertices and the unique interior point of ∆(1,q) . Hence the vector (1, 0, 0, . . . , 0) is the only vector in Π∆(1,q) at height 1. This prevents the existence of the “c” value required in Theorem 2.3. 3. q-Vectors With Fixed Support In Subsection 2.1, we established a stratification of the simplices ∆(1,q) in terms of their support vectors r. In this section, we present some first results on the IDP and h∗ -unimodality conditions for ∆(1,q) from this perspective. Our observations in this section center around the interplay between ∆(1,q) admitting an affine free sum decomposition into lower-dimensional ∆(1,q) and exhibiting IDP and/or h∗ -unimodality. As described below, admitting the former property can allow one to recursively detect the latter properties. In this section, we will show that only finitely many reflexive IDP ∆(1,q) with support vector r fail to admit such a free sum decomposition. This suggests that the stratification of ∆(1,q) by their support vectors is a desirable perspective from which to analyze Conjecture 1.3 for reflexive ∆(1,q) . In particular, we will use these results to prove Conjecture 1.3 for ∆(1,q) supported on two integers in Section 4. Our first two results in this section establish that there are infinitely many reflexive simplices supported by each r-vector, and that the r-vector and multiplicity vector (x1 , . . . , xk ) can be used to test our necessity condition for IDP. 8 BENJAMIN BRAUN, ROBERT DAVIS, AND LIAM SOLUS Proposition 3.1. Fix an r-vector r = (r1 , . . . , rk ) with gcd(r1 , . . . , rk ) = 1. There are infinitely many reflexive ∆(1,q) supported by r. Proof. By the positive solution to the Frobenius coin exchange problem [2, Chapter 1], there exists a positive Pk integer M such that for all m ≥ M , there exist positive integers x1 , . . . , xk such that m = 1 + i=1 xi ri . Thus, there exists a positive integer L such that for all ℓ ≥ L, there exist positive integers x1 , . . . , xk such Pk that ℓ lcm(r1 , . . . , rk ) = 1 + i=1 xi ri . For any such ℓ, for all i = 1, . . . , k we have that ri | ℓ lcm(r1 , . . . , rk ) = 1 + k X xj rj . j=1 Hence, there are infinitely many q = (r1x1 , r2x2 , . . . , rkxk ) such that the divisibility condition for reflexivity of ∆(1,q) is satisfied.  The following lemma is equivalent to Corollary 2.4. Lemma 3.2. If the r-vector r = (r1 , . . . , rk ) supports a reflexive IDP simplex ∆(1,q) then there exists a vector of positive integers x = (x1 , . . . , xk ) satisfying the k × k system of linear equations Rx = b where bi := ri − 1 and Rj,i   0 ri :=  ri mod rj if i = j if i < j if i > j . If a solution to this system of equations consisting of positive integers exists, then it corresponds to a reflexive ∆(1,q) . However, this condition is necessary but not sufficient to establish IDP. Proof. This lemma is a restatement of Corollary 2.4 using the notation (r1x1 , r2x2 , . . . , rkxk ).  It is important to observe that some reflexive IDP ∆(1,q) ’s admit decompositions into reflexive IDP ∆(1,q) ’s of smaller dimension, in the following sense. Let P, Q ⊂ Rn be two lattice polytopes. We say that P ⊕ Q := conv {P ∪ Q} is an affine free sum if, up to unimodular equivalence, P ∩ Q = {0} and the affine span of P and Q are orthogonal coordinate subspaces of Rn . Suppose further that P ⊂ Rn and Q ⊂ Rm are reflexive polytopes with 0 ∈ P and the vertices of Q labeled as v0 , v1 , . . . , vm . For every i = 0, 1, . . . , m, we define the polytope P ∗i Q := conv {(P × 0m ) ∪ (0n × Q − vi )} ⊂ Rn+m . The following theorem indicates that affine free sum decompositions can be detected from the q-vector defining ∆(1,q) . Theorem 3.3 (Braun, Davis [4]). If ∆(1,p) and ∆(1,q) are full-dimensional reflexive simplices with p = (p1 , . . . , pn ) and q = (q1 , . . . , qm ), respectively, then ∆(1,p) ∗0 ∆(1,q) is a reflexive simplex ∆(1,y) with y = Pn (p1 , . . . , pn , sq1 , . . . , sqm ) where s = 1 + j=1 pj . Moreover, if ∆(1,y) arises in this form, then it decomposes as a free sum. Further, if ∆(1,p) and ∆(1,q) are reflexive, IDP and h∗ -unimodal, then so is ∆(1,p) ∗0 ∆(1,q) . Thus, many reflexive, IDP, h∗ -unimodal ∆(1,q) ’s have these properties because they arise as affine free sums. If a reflexive IDP ∆(1,q) is an affine free sum of two reflexive IDP ∆(1,q) ’s of smaller dimension, then in order to infer h∗ -unimodality of the sum, this property of the summands must be separately verified. Nonetheless, it is reasonable to focus attention on those reflexive IDP ∆(1,q) ’s that do not arise as affine free sums, as these are fundamental examples that must be dealt with in any proof of Conjecture 1.3. Fortunately, as the following two results indicate, for each r-vector there are at most finitely many such reflexive IDP simplices to consider. Proposition 3.4. Fix an r-vector r = (r1 , . . . , rk ) for which ri | rk for all i ∈ {1, . . . , k − 1}. If q = (r1x1 , . . . , rkxk ) corresponds to a reflexive IDP simplex, then ∆(1,q) is an affine free sum of the simplices defined xk−1 xk−1 ) ) and (1xk ). Further, the simplices defined by the q-vectors (r1x1 , . . . , rk−1 by the q-vectors (r1x1 , . . . , rk−1 and (1xk ) are both IDP reflexive simplices. DETECTING THE INTEGER DECOMPOSITION PROPERTY 9 Proof. Using the notation of Lemma 3.2, observe that after canceling denominators in Rx = b our matrix equation has the form   0 r2 mod r1 r3 mod r1 · · · rk−1 mod r1 0 ..        r1 .  0 r3 mod r2 · · · rk−1 mod r2 r1 − 1   x1  .. ..   x   r − 1  ..  r1 2  . r2 0 . .    2     r3 − 1   x   3 . . . . = .    . . . .  r1 . . r2 . .   .. ..         . .  .  . . . . .. .. r .. ..   .. mod r k−1 k−2 rk − 1   xk  r1 r2 r3 ··· 0 0  r1 r2 r3 ··· rk−1 0 Notice that the upper-left block of this matrix and the first k − 1 entries of the right-hand vector form the system from Lemma 3.2 for the values r1 < r2 < · · · < rk−1 . Thus, the existence of a positive integer solution (x1 , . . . , xk ) to our matrix equation above implies that (x1 , . . . , xk−1 ) are the multiplicities of a qvector supported by (r1 , r2 , . . . , rk−1 ). Further, the final row of the matrix equation above implies that rk = Pk−1 xk−1 1 + i=1 xi ri , from which it follows that q = (r1x1 , . . . , rkxk ) arises from the free sum of p = (r1x1 , . . . , rk−1 ) xk and q = (1 ) by Theorem 3.3. It is straightfoward to see that the simplex defined by (1xk ) is always reflexive and IDP, so it only remains xk−1 to check that the simplex given by (r1x1 , . . . , rk−1 ) has the same properties. We know that ∆(1,q) = ∆(1,p) ∗0 xk−1 x1 ∆(1,h) where we set p := (r1 , . . . , rk−1 ) and h := (1xk ). By the definition of affine free sums, we know the free sum ∆(1,p) ∗0 ∆(1,h) is structured so that ∆(1,p) is a face of ∆(1,q) . We know that ∆(1,q) = ∆(1,p) ∗0 ∆(1,h) is IDP, and so each of its faces must also be IDP. The reflexive condition follows immediately from the analysis in the previous paragraph.  The final result in this section shows that for an r-vector such that ri ∤ rk for some i, there are at most finitely many such IDP reflexive simplices supported on that vector. Each of these simplices may or may not be decomposable as an affine free sum. Theorem 3.5. For a fixed support vector r, there exist only finitely many reflexive and IDP ∆(1,q) supported by r that do not admit a free sum decomposition into lower dimensional ∆(1,q) . Proof. Fix a support vector r = (r1 , . . . , rk ). By Proposition 3.4 we know that if ri | rk for all i ∈ {1, . . . , k − 1}, then any ∆(1,q) supported by r is an affine free sum of lower-dimensional ∆(1,q) . Thus, it remains to show that if ri ∤ rk for at least one i in {1, . . . , k − 1} then there are at most finitely many reflexive and IDP ∆(1,q) such that q is supported by r = (r1 , . . . , rk ). To see this, notice that if ∆(1,q) is reflexive and IDP and supported by r, then using the notation from Lemma 3.2 there is a positive integer vector x that satisfies Rx = b. The set of ∆(1,q) that are reflexive and IDP and supported by r corresponds to the integer points in Rk>0 ∩ (x + ker(R)). Since ker(R) is orthogonal to the image of RT , it follows from ri ∤ rk for some i that there exists a strictly positive vector a ∈ im(RT ) orthogonal to ker(R), namely the vector a obtained as the sum of the transpose of the k-th row of R and the transpose of the i-th row of R. Hence, Rk>0 ∩ (x + ker(R)) is bounded, and therefore contains at most a finite number of integer points.  4. q-Vectors With Support on Two Integers Using the results established in Sections 2 and 3, we are now able to prove our main theorem; i.e., we will prove that all IDP and reflexive ∆(1,q) supported on two integers are h∗ -unimodal. Let q = (r1x1 , r2x2 ) denote a q-vector supported on two vectors. By Lemma 3.2, in order to capture all IDP and reflexive simplices supported on two integers we need to consider two cases given by the possible divisibility relations between r1 and r2 . Table 1 summarizes the results of this section. For the two possible cases of q-vectors satisfying the conditions of Lemma 3.2, the table indicates if these simplices exhibit h∗ -unimodality, the IDP, and/or a known free sum decomposition. We now classify all IDP and reflexive simplices supported on two integers in Theorem 4.1. We see there are exactly two types of IDP and reflexive q-vectors, the former corresponding to the divisibility pattern r1 | r2 and the latter corresponding to r1 ∤ r2 . 10 BENJAMIN BRAUN, ROBERT DAVIS, AND LIAM SOLUS Case (1) r1 | r2 (2) r1 ∤ r2 h∗ -unimodal all vectors all vectors IDP all vectors all vectors known free sum decomposition all vectors none Table 1. IDP and h∗ -unimodality for q-vectors supported on two integers. The results are classified by the two divisibility cases of Lemma 3.2. Theorem 4.1. Let r < s be positive integers and let q = (r, r, . . . , r, s, s, . . . , s) = (rm , sx ) . | {z } | {z } m times x times Then ∆(1,q) is reflexive if and only if r | (1 + sx) and s | (1 + rm). Further, ∆(1,q) is IDP and reflexive if and only if either r 6= 1 with s = 1 + rm and x = r − 1 or r = 1 with s = 1 + m and x arbitrary. Proof. The claim regarding reflexivity follows immediately from the condition that ∆(1,q) is reflexive if and only if qj | (1 + q1 + · · · + qn ) for all j. To show that ∆(1,q) being reflexive IDP implies s = 1 + rm and x = r − 1 when r 6= 1, we apply Corollary 2.4. First, applying the corollary with qj = s we see that nro n s o 1 + mr 1 +m = 1, +x = s s s s and thus we have 1 + mr = s. Second, setting qj = r and applying the corollary yields   nso 1 nro 1+x 1 1 + mr +x = +x = +m = 1, r r r r r r from which it follows that x = r − 1. A similar argument shows that if r = 1, then 1 + m = s and x can be an arbitrary positive integer. Next, assume that s = 1 + mr and q is formed with x = r − 1, which Corollary 2.4 guarantees to be reflexive. To verify that ∆(1,q) is IDP in this case, we will directly apply Theorem 2.3. Consider the case where qj = r, hence b = 0, 1, . . . , r − 1. In this case, since 0 ≤ b < r, it follows that (3) reduces to   brm + b = b(rm − m + 1) − (r − 1)bm = b . b(rm − m + 1) − (r − 1) r Thus, in this case, our only option for the c value in equations (4) and (5) is c = 1. When qi = r it is immediate that cancellation implies (4). We therefore need to check that (4) holds when qi = rm + 1, c = 1, and b = 2, 3, . . . , r − 1, which follows since m + (b − 1)m = bm if and only if if and only if     (b − 1)mr + (b − 1) brm + b m+ = r r       rm + 1 (b − 1)(mr + 1) b(rm + 1) + = . r r r Next consider the case where qj = rm + 1, hence b = 0, 1, 2, . . . , rm. For this value of qj , (3) reduces to the following expression, which we denote by h(b):   br (11) h(b) := b − m mr + 1 It is straightforward to verify that when b = km + l with 0 ≤ l < m, we have h(0) = 0, we have h(km) = m when l = 0, and we have h(km + l) = l for l 6= 0, which follows directly from the following claim. DETECTING THE INTEGER DECOMPOSITION PROPERTY Claim: If k < r, then  kmr + r mr + 1  11 = k. This follows from the inequality k(mr + 1) < kmr + r < (k + 1)(mr + 1) . Thus, the viable candidates for c-values in (4) and (5) are c = 1, m + 1, 2m + 1, . . . , (r − 1)m + 1 , and all other values of b give values greater than or equal to two. To complete our proof, we need to check (4) and (5) for these values of b. We will show that for any such value of b, every viable c less than b will satisfy these equations. Fix b = am + d with 2 ≤ d ≤ m and 0 ≤ a ≤ (r − 1), and fix c = km + 1 with k ≤ a, which ensures that c < b.   (am + d)r = a. To prove this, we use the division algorithm to evaluate this quotient and Claim: rm + 1 obtain amr + dr = a(rm + 1) + (dr − a) . Because of the bounds on a and d we see that 0 ≤ dr − a ≤ mr, thus this is the correct representation using the divisionalgorithm, hence theclaim follows. r[(a − k)m + d − 1] Claim: = a − k. To prove this, we again use the division algorithm to evaluate rm + 1 this quotient and obtain (a − k)mr + (d − 1)r = (a − k)(rm + 1) + ((d − 1)r − (a − k)) . Because of the bounds on a, d, and k, it follows that 1 ≤ (d − 1)r − (a − k) ≤ (m − 1)r , and thus this is the correct representation using the division algorithm. The claim follows. Because k ≤ a < r, it follows that   (km + 1)r = k. rm + 1 Combining this with our two claims above, it follows that       (km + 1)r r[(a − k)m + d − 1] (am + d)r + = k + (a − k) = a = , rm + 1 rm + 1 rm + 1 which completes the proof. Note that this holds for any choice of k ≤ a, hence for all c < b with h(c) = 1. Theorem 3.3 tells us that if q = (1m , (1 + m)x ) for any positive integer x, then ∆(1,q) is reflexive IDP.  To prove our main result, it remains to establish that all IDP reflexive ∆(1,q) supported on two integers are h∗ -unimodal. Theorem 4.2. Let ∆(1,q) be a reflexive simplex supported on two integers. If ∆(1,q) is IDP then it is h∗ -unimodal. Proof. By Theorem 4.1, we know that ∆(1,q) with q = (rm , sx ) is reflexive and IDP if and only if either r 6= 1 with s = 1 + rm and x = r − 1 or r = 1 with s = 1 + m and x arbitrary. In the latter case, ∆(1,q) is the affine free sum ∆(1,q1 ) ∗0 ∆(1,q2 ) , where q1 = (1m ) and q2 = (1x ). By Theorem 3.3 it follows that ∆(1,q) is h∗ -unimodal. Thus, it only remains to prove h∗ -unimodality for the former case. For the sake of clarity, in the following we will let ∆r,m denote the simplex ∆(1,q) , where q = (rm , (rm + r−1 1) ) with r 6= 1. We will also let ⌊ r+m−1 ⌋ r−1 2 X X i (i + 1)z + r pr,m (z) := αi z i i=0 where αi := ( i=1 min{i, m − 1} if m < r, min{i, r} if m ≥ r. 12 BENJAMIN BRAUN, ROBERT DAVIS, AND LIAM SOLUS To prove the desired result, it then suffices to show that h∗∆r,m (z) = pr,m (z) + z r+m−1 pr,m   1 , z and that this sum is unimodal. To begin, recall that the formula given in Theorem 2.2 reduces to r(rm+1)−1 h∗∆r,m (z) = X z ω(b) , b=0 where ω(b) = b − (r − 1)     b b −m . r rm + 1 To compute the desired formula for h∗∆r,m (z) we first consider the sequence   r(rm+1)−1 b W := b − (r − 1) . r b=0 We will now group the terms in this sequence into subsequences of length r, and collect these into blocks of m subsequences as this will help make the desired k more apparent. Following this, we will subtract j formula b in order to produce the desired powers w(b), but from each number in the sequence the value m rm+1 organized in a fashion which reveals the desired polynomial structure. We first note that   r(rm+1)−1 b W = b − (r − 1) r b=0  rm−1 r−1 = (i + j)j=0 i=0  s+m−1 r−1 r−1 = (i + j)j=0 . i=s s=0 In this grouping, each sequence of m sequences of length r  s+m−1 r−1 Ws := (i + j)j=0 i=s k j b from each accounts for rm powers w(b). To compute the powers of h∗∆r,m (z) we must subtract m rm+1 j k b term in W . Since this subtraction causes the term to decrease by m each time rm+1 increases by 1 then for all s = 0, 1, . . . , r − 1, for every i > sm   b r−1 = (i + j − sm)r−1 (i + j)j=0 − m j=0 . rm + 1 Whereas, for i = sm r−1 (i + j)j=0 − m       b s−1 r s−1 r = (sm + j − (s − 1)m)j=0 , (j)j=s = (m + j)j=0 , (j)j=s . rm + 1 So for every s = 0, 1, 2, . . . , r − 1, we have that      m−1  b s−1 r r−1 Ws − m = (m + j)j=0 , (j)j=s , (i + j)j=0 . rm + 1 i=1 Ps−1 s−1 r Notice that the sequence of coefficients (m + j)j=0 gives the polynomial z m j=0 z j , and the sequence (j)j=s Pr yields j=s z j . Varying over s = 0, 1, . . . , r − 1, the sum of these polynomials is (12) r−1 X i=0 (i + 1)z i + z r+m−1 r−1 X i=0 (i + 1)  i 1 . z DETECTING THE INTEGER DECOMPOSITION PROPERTY (1) (2) (3) (4) (5) (6) r1 r1 r1 r1 r1 r1 | r2 | r2 ∤ r2 ∤ r2 ∤ r2 ∤ r2 Case r1 | r3 r1 | r3 r1 | r3 r1 | r3 r1 ∤ r3 r1 ∤ r3 r2 r2 r2 r2 r2 r2 | r3 ∤ r3 | r3 ∤ r3 | r3 ∤ r3 h∗ -unimodal all vectors all vectors all vectors some but not all all vectors ? IDP all vectors all vectors all vectors some but not all all vectors ? 13 known free sum decomposition all vectors all vectors all vectors ? none ? Table 2. IDP and h∗ -unimodality for q-vectors supported on three integers. The results are classified by the six divisibility cases of Lemma 3.2 that capture IDP simplices.  m−1 The sequence of coefficients (i + j)r−1 yields the polynomial j=0 i=1 z r−1 X zj + z2 r−1 X z j + · · · + z m−1 zj , j=0 j=0 j=0 r−1 X or equivalently, (13) ⌊ r+m−1 ⌋ 2 X j=1 i αi z + z r+m−1 ⌊ r+m−1 ⌋ 2 X i=1 αi  i 1 . z Since the sum of one copy of (12) and r copies of (13) is a unimodal polynomial and gives the desired formula for h∗∆r,m (z), this completes the proof.  5. q-Vectors With Support on Three Integers In this section, we study Conjecture 1.3 for the next most general family of ∆(1,q) in regards to Payne’s counterexamples to Conjecture 1.4. Namely, we analyze IDP and h∗ -unimodality for reflexive ∆(1,q) supported on three integers r1 < r2 < r3 . Lemma 3.2 implies that in order for an r-vector r = (r1 , r2 , r3 ) to be a candidate for supporting an IDP and reflexive ∆(1,q) , it must satisfy the following linear system with a positive integer solution:      0 r2 mod r1 r3 mod r1 r1 − 1 x1  r1 0 r3 mod r2   x2  =  r2 − 1  . r1 r2 0 x3 r3 − 1 It is not clear in general how to identify the r-vectors for which a positive integer solution to this system exists, let alone those for which both reflexivity and IDP hold. Thus, we consider our analysis by cases based on divisibility among the ri ’s. Given that it is not possible for r1 | r2 , r2 | r3 , and r1 ∤ r3 , there are seven possible divisibility relations among the ri ’s, leading to seven cases in which to consider the application of Lemma 3.2. Furthermore, we will see in Proposition 5.1 that there are no IDP and reflexive simplices supported on an r-vector satisfying r1 | r2 , r1 ∤ r3 , and r2 ∤ r3 . Thus, there are a total of six cases for which we must check the validity of Conjecture 1.3, and they are numbered (1) to (6) in Table 2. Table 2 further indicates if the simplices satisfying each case exhibit h∗ -unimodality, IDP, and/or a known free sum decomposition. As can be seen from this table, we will prove Conjecture 1.3 for cases (1), (2), (3), and (5). Cases (1), (2), and (3) will be discussed in Subsection 5.1, case (5) will be discussed in Subsection 5.2, and cases (4) and (6) will be discussed in Subsection 5.3. These results suggest that Conjecture 1.3 holds in the next most general family of ∆(1,q) in relation to Payne’s counterexamples to Conjecture 1.4. To begin, our first proposition shows that the linear system from Lemma 3.2 need not have a positive integer solution, thereby eliminating the case r1 | r2 , r1 ∤ r3 , and r2 ∤ r3 (and so it is not presented in Table 2). Proposition 5.1. Suppose the r-vector r = (r1 , r2 , r3 ) satisfies r1 | r2 , r1 ∤ r3 , and r2 ∤ r3 . If r also satisfies Lemma 3.2, then r does not support any q-vectors. 14 BENJAMIN BRAUN, ROBERT DAVIS, AND LIAM SOLUS Proof. If such a q-vector were to exist, it would have the form q = (sx1 , (st)x2 , ((mt + k)s + r)x3 ) for some integers s, m ≥ 1, t ≥ 2, 0 ≤ k < t, 0 < r < s, and x1 , x2 , x3 > 0. To solve for the xi ’s we consider      0 0 r x1 s−1  s 0 ks + r   x2  =  . st − 1 s st 0 x3 (mt + k)s + r − 1 Solving the first row of this system forces rx3 = s − 1, hence there must exist some s − 1 ≥ d ≥ 1 such that rd = s − 1, hence x3 = d. This forces the second row to imply x1 = t − kd − 1; if x1 is to be positive, this implies t − kd − 1 > 0. These values of x2 and x3 lead to the third row having the form mts + ks + r − 1 − st − skd − s . st Since 0 < r < s and we must have that s divides the numerator of this expression, it follows that r = 1, which forces d = s − 1, and thus s ≥ 2. Thus, the above fraction simplifies to x2 = x2 = t(m − 1) + k − (kd + 1) . t If x2 is to be a positive integer, we must have that t divides the numerator, hence t | k−(kd+1) = k(2−s)−1. Since s ≥ 2, we have 2 − s ≤ 0, and hence 0 ≥ k(2 − s) − 1 = k − (kd + 1) ≥ −(kd + 1) < −t . Thus t divides k(2 − s) − 1 only if k(2 − s) = 1, which is impossible since k > 0 and 2 − s ≤ 0. Thus, there is no vector of this type supporting an IDP reflexive.  5.1. Cases (1), (2), and (3) by way of affine free sum decompositions. Propositions 5.2, 5.3, and 5.4 prove Conjecture 1.3 for cases (1), (2), and (3) of Table 2, respectively. The proofs utilize the affine free sum construction, Theorem 3.3, and Proposition 3.4. Proposition 5.2. Suppose the r-vector r = (r1 , r2 , r3 ) satisfies r1 | r2 | r3 . If r also satisfies the Lemma 3.2, then any q-vector supported on r has the following form. For each s, m, x such that s, m, x > 1, set q = (1m−1 , ms−1 , (sm)x ) . Moreover, these simplices are all IDP and have unimodal h∗ -vectors. Proof. Our linear system from Lemma 3.2 is  0 0  r1 0 r1 r2 of the form     0 x1 r1 − 1 0   x2  =  r2 − 1  . 0 x3 r3 − 1 Thus, r1 = 1, hence x1 = r2 − 1 and x3 = x can be arbitrary. Setting (r1 , r2 , r3 ) = (1, m, ms), it follows that q = (1m−1 , ms−1 , (sm)x ) . Using Proposition 3.4, we know that these simplices may be expressed as the affine free sum (∆(1,q1 ) ∗0 ∆(1,q2 ) ) ∗0 ∆(1,q3 ) , where q1 = (1m−1 ), q2 = (1s−1 ), and q3 = (1x ). From Theorem 3.3, we see that these simplices are IDP, reflexive, and h∗ -unimodal.  Proposition 5.3. Consider r-vectors r = (r1 , r2 , r3 ) satisfying r1 | r2 , r1 | r3 , and r2 ∤ r3 . For any k, d > 1 such that kd = r, the vector (1k−1 , rm , (mr + k)d−1 ) corresponds to a reflexive simplex, and these are the only reflexive simplices ∆(1,q) that simultaneously satisfy these divisibility conditions and Lemma 3.2. Moreover, each of these simplices is IDP and h∗ -unimodal. Proof. If r1 | r2 and r1 | r3 , then it follows that r1 = 1 by the first line of Rx = b, so we are in the case where (r1 , r2 , r3 ) = (1, r, mr + k) for 0 < k < r. By Theorem 3.5, there are at most a finite number of IDP reflexives supported by this vector. We first apply Lemma 3.2 to identify potential x1 , x2 , x3 -values; if DETECTING THE INTEGER DECOMPOSITION PROPERTY 15 (1x1 , rx2 , (mr + k)x3 ) corresponds to an IDP reflexive, then there must be a positive integer solution to the linear system      0 0 0 1 − 1/1 x1   1/r 0 k/r   x2  =  1 − 1/r 1/(mk + r) r/(mk + r) 0 x3 1 − 1/(mr + k) Noting that the first row is a zero equation, and scaling the other two rows to clear denominators, we obtain     x1   1 0 k  r−1 x2  = 1 r 0 mr + k − 1 x3 Because any solution to this system is restricted to positive integers, the first equation above forces 0 < x1 < r − 1, which in turn when combined with the second equationforces x2 = m. Hence x1 = k − 1 and x3 = (r − k)/k. Thus, if a positive integer solution is possible, it must be that k | r is a proper divisor of r. Thus, for any k, d > 1 such that kd = r, we see that the vector (1k−1 , rm , (mr + k)d−1 ) corresponds to a reflexive simplex. Note that q1 = (1k−1 )and q2 = (dm , (md + 1)d−1 ) are both valid q-vectors. The former is well-known to have an h∗ -vector of (1, . . . , 1). The latter corresponds to an IDP, reflexive, and h∗ -unimodal simplex by Theorem 4.1 and Theorem 4.2. Thus, (1k−1 , rm , (mr + k)d−1 ) arises as the q-vector of the affine free sum ∆(1,q1 ) ∗0 ∆(1,q2 ) , and from Theorem 3.3 we draw the conclusion that ∆(1,q) is also reflexive, IDP, and h∗ -unimodal.  Proposition 5.4. Suppose the r-vector r = (r1 , r2 , r3 ) satisfies r1 ∤ r2 , r1 | r3 , and r2 | r3 . If r also satisfies Lemma 3.2, then any q-vector supported on r has the following form. For each s, m, x such that s, m ≥ 1 and x ≥ 0, set q = (sm , (ms + 1)s−1 , (s(ms + 1))x ) . Moreover, these simplices are all IDP and have unimodal h∗ -vectors. Proof. Suppose we set r1 = s, r2 = ms + t, and r3 = u(ms + t) for some 0 ≤ t < s and u ≥ 1. The second equation in the linear system from Lemma 3.2 is of the form sx1 = ms + t − 1. This implies s | t − 1, but since t < s and s 6= 1, t must be 1. Therefore, x1 = m, and the first line of the linear system is simply x2 = s − 1. The last equation in the system gives ms + (ms + 1)(s − 1) = u(ms + 1) − 1. Solving for u, we get s = u, so then r3 = s(ms + 1). Since x3 is a free variable, we may have as many copies of s(ms + 1) in q as we like. Note that by Proposition 3.4, q = (sm , (ms + 1)s−1 , (s(ms + 1))x ) arises as the affine free sum (∆(1,q1 ) ∗0 ∆(1,q2 ) ), where q1 = (sm , (ms + 1)s−1 ) and q2 = (1x ). This implies that these simplices are reflexive, IDP, and h∗ -unimodal.  5.2. Proving Conjecture 1.3 for Case (5). Theorems 5.5 and 5.7 together establish that Conjecture 1.3 holds for all ∆(1,q) captured by case (5) of Table 2. Theorem 5.5. Suppose the triple r1 < r2 < r3 satisfies r1 ∤ r2 , r1 ∤ r3 , and r2 | r3 . Every q-vector supported on three integers satisfying these conditions that also satisfies Lemma 3.2 has the following form. For each s, t, r such that s > 1, 0 < t < s with t | s, and r ≥ 1, set q = (sr , (rs + 1)t−1 , (t(rs + 1))s/t−1 ) . For every such q, ∆(1,q) is IDP. Proof. Since r1 does not divide either of r2 , r3 , we see that r1 > 1. Thus, the divisibility conditions imply that (r1 , r2 , r3 ) is of the form (s, rs + k, mrs + ℓs + t) for some s > 1, r ≥ 1, m > 1, where km = ℓs + t with 0 < k < s and 0 < t < s. Lemma 3.2 implies that our necessary linear system is      0 k t x1 s−1  s . 0 0   x2  =  rs + k − 1 s rs + k 0 x3 mrs + ℓs + t − 1 16 BENJAMIN BRAUN, ROBERT DAVIS, AND LIAM SOLUS Gaussian elimination on this system leads to the solutions rs + k − 1 x1 = , s ((m − 1)r + ℓ)s + t − k x2 = , rs + k and (s − 1)(rs + k) − k[((m − 1)r + ℓ)s + t − k] . x3 = t(rs + k) Since x1 must be a positive integer, we see that s | k − 1; by assumption 0 < k − 1 < s − 1, hence k = 1. Thus, x1 = r. Knowing that k = 1, x2 simplifies to m − 1 and x3 = s−m t . Since x3 is a positive integer, we need that t | (s − m), which forces m < s. However, m = ℓs + t and m < s implies m = t. Thus, x2 = t − 1 and x3 = st − 1, which implies that t | s, completing the identification of the necessary x-values. We next verify IDP using Theorem 2.3. Throughout, we use the notation P  X   1 + i6=j qi bqi h(b) := b − . qj qj i6=j Case: qj = s. It is straightforward to verify that h(b) = b − (s/t − 1) and using this formula one can check that  bt s  , h(ks/t) = k . Combining these two observations, it follows that h(b) = 1 only when b = 1 and b = s/t, thus identifying the b-values we are required to check in (3). To verify that (4) always has the desired solution, we consider three cases. If qi = s, the result is trivial. If qi = rs + 1, then we may select c = 1, from which it follows that both sides of (4) are equal to r(b − 1). If qi = t(rs + 1), then we set c = s/t, from which it is straightforward to compute that both sides of (4) are equal to btr − rs − 1 + ⌊bt/s⌋. This completes our first case. Case: qj = rs + 1. It is straightforward to verify that   bs . h(b) = b − r rs + 1 As we are considering values of b ranging from 1 to rs, we see that this analysis is identical to that undertaken in the proof of Theorem 4.1 starting from (11). To verify that (4) always has the desired solution, we consider three cases. If qi = rs + 1, the result is trivial. If qi = s, then the analysis is identical to that given following (11) in our proof of Theorem 4.1. If qi = t(rs + 1), then we can set c = 1 and the result is immediate. This completes our second case. Case: qj = t(rs + 1). Set d = s/t. We first identify those values of b that satisfy (3) and (5). It is straightforward to verify that     bd b h(b) = b − r − (t − 1) . rs + 1 t Writing b = mt + ℓ where 0 ≤ m ≤ rs and 0 ≤ ℓ ≤ t, it follows that   ms + ℓd h(b) = h(mt + ℓ) = m + ℓ − r . rs + 1 We can now further divide into cases: either we have m = rs or we have m = kr + w where 0 ≤ k < s and 0 ≤ w < r, which yields   ws + ℓd − k h(b) = h((kr + w)t + ℓ) = w + ℓ − r . rs + 1 For m j6= rs, observe k that since 0 ≤ ws ≤ r−1 and 0 ≤ ℓd < s, with 0 ≤ k < s, we have that 0 ≤ ws+ℓd < rs. ws+ℓd−k is equal to either 0 or −1. Thus, rs+1 Subcase 1 of 3: Suppose m = rs. Since 0 ≤ ℓd < s, we have   ℓd − s = ℓ + r. h(srt + ℓ) = ℓ − r rs + 1 DETECTING THE INTEGER DECOMPOSITION PROPERTY 17 If this is equal to 1, then it must be that ℓ = 0 and r = 1. Thus, if r = 1, we have that h(st) = 1. Subcase 2 of 3: Suppose now that m 6= rs and that   ws + ℓd − k = −1 . rs + 1 Then since w, ℓ ≥ 0 and r ≥ 1, we have h(b) = w + ℓ + r = 1 which forces w = ℓ = 0 and r = 1. In this case, h(kt) = 1 any time that k > 0. Thus, if r = 1, we have that h(kt) = 1 when 0 < k < s. Subcase 3 of 3: Suppose again that m 6= rs and that   ws + ℓd − k = 0. rs + 1 Then 0 ≤ k ≤ ws + ℓd, which implies that either (A) 0 < w < r with 0 ≤ ℓ < t or (B) w = 0 with k ≤ ℓd. If (A) holds, then h(b) = w + ℓ = 1 forces w = 1 and ℓ = 0 since w > 0, which means that h(b) = 1 when b = (kr + 1)t for 0 ≤ k < s. If (B) holds, then our same equation forces w = 0 and ℓ = 1 when k ≤ d, which means that h(b) = 1 when b = krt + 1 for 0 ≤ k < s and k ≤ d. We summarize the values of b for which h(b) = 1 that were just derived: • If r = 1 and 0 < k ≤ s we have b = kt. • If r ≥ 1 and 0 ≤ k < s, we have b = (kr + 1)t. • If r ≥ 1 and 0 ≤ k < s and k ≤ d, we have b = krt + 1. Our next goal is to establish that (4) is always satisfied; recall that we are in the case where qj = t(rs + 1). If qi = t(rs + 1), then (4) is trivially satisfied. If qi = rs + 1, we write b = mt + ℓ where 0 ≤ m < rs + 1 and 0 ≤ ℓ < t. Substituting this form of b into (4) yields the equation jck ℓ − c − = . t t If b > t, we set c = t and the equation is satisfied. If 2 < b < t, then we set c = 1 and the equation is satisfied. If qi = s, the analysis becomes more complicated. We first define d = s/t and write b = mt + ℓ where 0 ≤ m < d and 0 ≤ ℓ < t. Our argument will proceed by considering r = 1 and r > 1 separately. Suppose r = 1. Then the left-hand-side of (4) is reduced to     cd ℓd − m − s+1 s+1 and the right-hand-side to   ℓd − m − cd . s+1 Since m < d, if ℓd − m < 0 this forces ℓ = 0 and 0 < m, thus b is a multiple of t, and we found earlier that h(mt) = 1. Thus, we need proceed no j k further. If ℓd − m ≥ 0, then since m < d we must have ℓ ≥ 1, and we ℓd−m also have ℓd − m < s. Thus, s+1 = 0, from which it follows that (4) reduces to     ℓd − m − cd cd = . − s+1 s+1 If m = 0, set c = 1 and this equation is solved. If m ≥ 1, set c = mt + 1 which is less than b in this case, and this equation is again satisfied. This completes our proof for r = 1. We next consider when r ≥ 2, maintaining our previous notation of b = mt + ℓ. Write m = f r + g where 0 ≤ f ≤ ⌊d/r⌋ and 0 ≤ g < r except in the case where f = ⌊d/r⌋ in which case g is bounded above by d − d⌊d/r⌋. This leads to the left-hand-side of (4) having the form     cd gs + ℓd − f − f+ rs + 1 rs + 1 while the right-hand-side has the form f+  gs + ℓd − f − cd rs + 1  . 18 BENJAMIN BRAUN, ROBERT DAVIS, AND LIAM SOLUS We thus need to solve the equation       cd gs + ℓd − f − cd gs + ℓd − f − = rs + 1 rs + 1 rs + 1 subject to the constraints 0 ≤ g < r (with the exception mentioned above), 0 ≤ ℓ ≤ t, and ≤ f ≤ ⌊d/r⌋. Note that the first two inequalities imply that 0 ≤ gs + ℓd < rs, and also f ≤ ⌊d/r⌋ ≤ s, hence    gs + ℓd − f 0 if gs + ℓd ≥ f = −1 if gs + ℓd < f rs + 1 Subcase 1 of 2: Suppose gs + ℓd − f ≥ 0. Then (4) reduces to     gs + ℓd − f − cd cd = . − rs + 1 rs + 1 Note that f ≤ ⌊d/r⌋ < d, and thus we can set c = f rt + 1 which is less than b. The left-hand-side of our above equation is given by       f rs + d −f rs + f + d − f (f rt + 1)d − =− =− = −f . rs + 1 rs + 1 rs + 1 Similarly, the right-hand-side of our equation is given by     f − d + gs + ℓd − f −d + gs + ℓd −f + = −f + ,. rs + 1 rs + 1 Since h(b) is assumed to be at least 2, we have that one or both of g and ℓ are non-zero. Combining this observation with gs + ℓd − f ≥ 0 it follows that gs + ℓd > 0. Note that d | (gs + ℓd), and thus rs + 1 > gs + ℓd − d ≥ 0, which forces the right-hand-side of our equation to equal −f , satisfying (4). Subcase 2 of 2: Suppose gs + ℓd − f < 0. Note that since g, s, ℓ, d ≥ 0, it follows that f ≥ 1 and thus b = (f r + g)t + ℓ ≥ t. Set c = t, which is less than b. With these conditions, the left-hand-side of (4) is easily seen to equal −1. The right-hand-side of (4) is given by   gs + ℓd − f − s . rs + 1 Since gs + ℓd − f < 0 and −s < 0, the numerator above is strictly negative. Also, since g, s, ℓ, d ≥ 0, the numerator is minimized by −f − s > −d − s ≥ −2s. But, since we assumed that r ≥ 2, it follows that rs + 1 > 2s and thus the floor function above is equal to −1, satisfying equality for (4). This completes the proof establishing IDP.  Remark 5.6 (IDP and reflexive ∆(1,q) that do not decompose as affine free sums). Note that the previous theorem establishes that, in the case of q-vectors with support on three integers, classes of IDP reflexive simplices arise that do not all decompose as affine free sums. For example, setting s = 6, t = 3, r = 1 in Theorem 5.5 yields q = (6, 7, 7, 21) with h∗ -polynomial 1 + 11z + 18z 2 + 11z 3 + z 4 . Since this h∗ -polynomial is irreducible over the integers, ∆(1,6,7,7,21) does not admit a free sum decomposition [1]. Note, however, the close resemblance between these vectors and those formed by the affine free sum construction. To prove Conjecture 1.3 for the ∆(1,q) satisfying case (5) of Table 2, it only remains to prove that they are h∗ -unimodal. This is the content of the following theorem, for which the utilized proof techniques are essentially the same as those of Theorem 4.2. Theorem 5.7. For each triplet of integers s, t, and r satisfying s > 1, 0 < t < s with t | s, and r ≥ 1 the reflexive lattice simplex ∆(1,q) with q = (sr , (rs + 1)t−1 , (t(rs + 1))s/t−1 ) has a unimodal h∗ -polynomial. Proof. Let s = φt for some integer φ > 1. Notice first that 1+ Pn i=1 qi s(rs+1)−1 h∗ (∆(1,q) ; z) = X b=0 z ω(b) , = s(rs+1). Therefore, by Theorem 2.2 DETECTING THE INTEGER DECOMPOSITION PROPERTY where Consider the sequence 19       b b b ω(b) = b − r − (t − 1) − (φ − 1) . rs + 1 s φ   s(rs+1)−1 b . b − (φ − 1) φ b=0 We first group W into subsequences of length φ, which we then group into sequences of φ-sequences of length t (so each t-sequence of φ-sequences consists of s terms of W ). Since W consists of s(rs + 1) terms, there are exactly sr + 1 of these t-sequences. We group the first rs of these t-sequences into sequences of r t-sequences. This presentation of W looks like   !  ℓt+t−1 kr+r−1 s−1  t−1 φ−1 φ−1 . W = (i + j)j=0 , (srt + i + j)j=0 W := i=ℓt ℓ=kr i=0 k=0   b from each term in W . Since this subtraction causes each term to decrease by We then subtract (t − 1) s   b (t − 1) every time increases by 1, which happens at the start of each new t-sequence then s   !  ℓt+t−1 kr+r−1 s−1  t−1 φ−1 φ−1 , W = (i + j + ℓ(t − 1))j=0 , (srt + i + j − rs(t − 1))j=0 i=ℓt  =  (i + j + φ−1 ℓ)j=0 t−1 kr+r−1 i=0 ℓ=kr ℓ=kr !s−1 k=0  i=0 k=0   t−1 φ−1 . , (i + j + rs)j=0 i=0  b . Thus, we subtract one more copy of r each time Now, from each term in W we must subtract r sr + 1     b b increases by 1. For the first sr t-sequences (containing s elements of W each), the value sr + 1 sr + 1 increases at the (ksr + k)th term of W for k = 1, 2, . . . , s − 1, and therefore is constant on each t-sequence in the r-sequence  t−1 kr+r−1 φ−1 (i + j + ℓ)j=0 i=0 ℓ=kr except for the very first t-sequence, corresponding to ℓ = kr. Thus, for a fixed k ∈ [s − 1], and all ℓ > kr we subtract kr from each term of the t-sequence:  t−1 kr+r−1  t−1 r−1 φ−1 φ−1 (14) (i + j + ℓ − kr)j=0 = (i + j + ℓ)j=0 . i=0 i=0 ℓ=kr+1 ℓ=1 th Similarly, when k = 0, this reduction applies to all t-sequences in the 0 r-sequence:  t−1 kr+r−1  t−1 r−1 φ−1 (15) (i + j + ℓ − kr)φ−1 = (i + j + ℓ) . j=0 j=0 i=0 i=0 ℓ=kr+1 ℓ=0 As for the first t-sequence ofall r-sequences with k = 1, 2, . . . , s − 1, the location of the (ksr + k)th term of W  b increases from k − 1 to k. Notice, we can consider the final t-sequence determines precisely where sr + 1 (i.e. the (sr + 1)st t-sequence) with this collection simply by letting k = 1, 2, . . . , s. So we must determine which φ-sequence of the t-sequence contains this term. Recall that s = φt and k = 1, 2, . . . , s. So when the integer k lies in the interval [(α − 1)φ + 1, αφ] then the  b first α − 1 φ-sequences lose r(k − 1) and the last t − α φ-sequences lose rk when we subtract off . sr + 1 Consider a fixed ℓ = kr for a fixed k = 1, 2, . . . , s. The corresponding t-sequence is  t−1 φ−1 (i + j + kr)j=0 . i=0 20 BENJAMIN BRAUN, ROBERT DAVIS, AND LIAM SOLUS As α ∈ [t] varies, we get the sequences  α−2  α−2 φ−1 φ−1 (i + j + kr − r(k − 1))j=0 = (i + j + r)j=0 , i=0 i=0 th for the φ-sequences before the α φ-sequence and  t−1  t−1 φ−1 φ−1 (i + j + kr − kr)j=0 = (i + j + r)j=0 , i=α i=α for the φ-sequences following the αth φ-sequence. In consideration of all α ∈ [t] this collectively yields the polynomial   t−1 φ−1 α−2 t X X X X φ−1 X z r z i+j  . z i+j + (16) α=1 i=α j=0 i=0 j=0 Finally, we must consider the αth φ-sequence for each fixed k = 1, 2, . . . , s. Here, the index i = α − 1, whenever k lies in the corresponding interval [(α − 1)φ + 1, αφ]. Since k is fixed, then ℓ = kr is also fixed, so the corresponding αth φ-sequence looks like φ−1 (j + kr + α − 1)j=0 . If α = (α − 1)φ + β for β ∈ [φ] then this φ-sequence is φ−1 (j + αφ + β − 1)j=0 . So from the first β − 1 terms we must subtract r(k − 1) and from the last z − β + 1 terms we must subtract kr. Thus, this φ-sequence becomes   β−2 φ−1 (j + r + α − 1)j=0 , (j + α − 1)j=β−1 . In consideration of all α ∈ [t] and β ∈ [φ] this yields the polynomial   ! φ−1 φ−1 φ−1 φ β−2 t t X X  1 i X X X X X z α+r−1 (i + 1)z i + xr+φ−1 z α−1 z i + xα−1 zi = (17) . z α=1 α=1 i=0 i=0 i=0 i=β−1 β=1 Combining the results of equations (14), (15), (16), and (17), we see that h∗ (∆(1,q) ; z) is given by summing equations (18), (19), and (20) presented below. Equation (18) accounts for both equations (14) and (15). (18) 1 + ((r − 1)s) r−1 X t−1 φ−1 X X z i+j+ℓ . ℓ=1 i=0 j=0 t X (19) α=1 (20) t X z  z r α−1 α=1 α−2 X X φ−1 t−1 φ−1 X X z i+j + i=α j=0 i=0 j=0 φ−1 X i r+φ−1 (i + 1)z + x i=0  z i+j  . φ−1 X i=0 1 z i ! . Analogously, to the conclusion of Theorem 4.2 the sum of these three polynomials is unimodal.  5.3. Cases (4) and (6) of Table 2. In the following two remarks, we discuss the remaining divisibility cases, which are significantly more complicated. While it would be interesting to observe that Conjecture 1.3 holds for all ∆(1,q) supported on three integers, it appears that new methods must be developed to deal with these remaining two, and much more unruly, cases. Remark 5.8 (Case (4) in Table 2). Suppose the r-vector r = (r1 , r2 , r3 ) satisfies r1 ∤ r2 , r1 | r3 , and r2 ∤ r3 . Then r1 , r2 , and r3 are of the forms r1 = s for some s > 1, r2 = ms + t for some m ≥ 1 and 0 < t < s, and r3 = u(ms + t) + v for some u ≥ 1 and 0 < v < ms + t. Moreover, since r1 | r3 , we must have ut + v = ks for some k ∈ Z. Every q-vector supported on r1 , r2 , r3 that also satisfies Lemma 3.2 has the following form: q = (smu+k−1− m(s−1) t , (ms + t)(s−1)/t , (u(ms + t) + v)α ), DETECTING THE INTEGER DECOMPOSITION PROPERTY 21 1 where α = tv (ms(t − tu + s − 1) + t(t − 1 − s(k − 1))). To see this, note that, since r3 mod r1 = 0, the linear system of Lemma 3.2 reduces to      0 t 0 s−1 x1 s . 0 v  x2  =  ms + t − 1 s ms + t 0 x3 u(ms + t) + v − 1 The first equation of this system gives us tx2 = s − 1, so x2 = (s − 1)/t. The third equation of the system gives sx1 + (ms + t)x2 = u(ms + t) + v − 1, so solving for x1 , we get x1 = = = = = u(ms + t) + v − 1 − (ms + t) s−1 t s tu(ms + t) + tv − t − (ms + t)(s − 1) st −t − ms(s − 1) − t(s − 1) mu + k + st −st − ms(s − 1) mu + k + st m(s − 1) mu + k − 1 − . t Finally, the second equation of the linear system will give sx1 + vx3 = ms + t − 1, so solving for x3 yields ) ms + t − 1 − s(mu + k − 1 − m(s−1) t v tms + t2 − t − mstu − kst + st + ms(s − 1) = tv ms(t − tu + s − 1) + t(t − 1 − s(k − 1)) . = tv Unlike our previous cases, the above conditions on x1 , x2 , and x3 are necessary for ∆(1,q) to be IDP but are not sufficient. For example, the vector q = (2, 2, 15, 20, 20) examined in the proof of Corollary 2.4 does produce a reflexive simplex ∆(1,q) and does satisfy the conditions of the previous result, yet fails to be IDP. However, it is straightforward to verify that the h∗ -polynomial for ∆(1,2,2,15,20,20) has coefficients (1, 9, 20, 20, 9, 1), hence is unimodal. Other non-IDP q-vectors with unimodal h∗ -polynomials, all found via computer search, include (2, 9, 12, 12), (2, 2, 15, 20, 20), (2, 2, 2, 21, 28, 28), (2, 15, 18, 18, 18, 18), and (3, 16, 16, 36, 36, 36). Even more striking is that the vector q = (3, 20, 24, 24, 24, 24) satisfies these divisibility conditions and Lemma 3.2, yet is non-IDP and has a non-unimodal h∗ -vector, specifically x3 = h∗ (∆(1,q) ; z) = z 6 + 16z 5 + 29z 4 + 28z 3 + 29z 2 + 16z + 1 . It remains a challenge to provide a more clear classification and description of the behavior of the q-vectors satisfying Lemma 3.2 in this case. Remark 5.9 (Case (6) in Table 2). The last case to check is when r1 ∤ r2 , r1 ∤ r3 , and r2 ∤ r3 . If we proceed as usual, using Lemma 3.2, the linear system we produce is rather unwieldy. Computational data suggests that when this is the support of q, the corresponding reflexive simplex is IDP. This has been checked using, in part, the Normaliz interface in Macaulay2 [9] for all possible r1 = s, r2 = ms + t, r3 = u(ms + t) + v with s ≤ 25 and m, u ≤ 10 satisfying r1 ∤ r2 , r1 ∤ r3 , and r2 ∤ r3 , and for every resulting q-vector ∆(1,q) was both IDP and h∗ -unimodal. However, this search only resulted in 33 distinct reflexive simplices, listed in Table 3. The main difficulty in verifying additional cases is due to having more lattice points in the high-dimensional simplices than the software can handle. Although the general status of Conjecture 1.3, remains very open, based on the observations and computational results given in this section, we believe the following conjecture should have a positive answer. The comments given in Remark 5.8 and 5.9 suggest that new methodology must be developed to prove or disprove it. Conjecture 5.10. All reflexive IDP ∆(1,q) supported on three integers are h∗ -unimodal. 22 BENJAMIN BRAUN, ROBERT DAVIS, AND LIAM SOLUS r-vector (65 , 621 , 931 ) (64 , 501 , 751 ) (63 , 381 , 571 ) (62 , 261 , 391 ) (61 , 141 , 211 ) (105 , 1022, 2551 ) (104 , 822 , 2051) (103 , 622 , 1551) (102 , 422 , 1051) (101 , 222 , 551 ) (123 , 1111, 1482 ) (122 , 751 , 1002) (121 , 391 , 522 ) (145 , 1423, 4971 ) (144 , 1143, 3991 ) (143 , 863 , 3011) (142 , 583 , 2031) (141 , 303 , 1051) (185 , 1824, 8191 ) (184 , 1464, 6571 ) (183 , 1104, 4951 ) (182 , 744 , 3331) (181 , 384 , 1711) (202 , 1641, 2053 ) (201 , 841 , 1053) (213 , 1922, 4482 ) (212 , 1292, 3012 ) (211 , 662 , 1542) (225 , 2225, 12211) (224 , 1785, 9791 ) (223 , 1345, 7371 ) (222 , 905 , 4951) (221 , 465 , 2531) h∗ (∆(1,q) ) (1, 20, 36, 36, 36, 36, 20, 1) (1, 20, 36, 36, 36, 20, 1) (1, 20, 36, 36, 20, 1) (1, 20, 36, 20, 1) (1, 20, 20, 1) (1, 28, 76, 100, 100, 100, 76, 28, 1) (1, 28, 76, 100, 100, 76, 28, 1) (1, 28, 76, 100, 76, 28, 1) (1, 28, 76, 76, 28, 1) (1, 28, 52, 28, 1) (1, 40, 109, 144, 109, 40, 1) (1, 40, 109, 109, 40, 1) (1, 40, 74, 40, 1) (1, 36, 100, 164, 196, 196, 164, 100, 36, 1) (1, 36, 100, 164, 196, 164, 100, 36, 1) (1, 36, 100, 164, 164, 100, 36, 1) (1, 36, 100, 132, 100, 36, 1) (1, 36, 68, 68, 36, 1) (1, 44, 124, 204, 284, 324, 284, 204, 124, 44, 1) (1, 44, 124, 204, 284, 284, 204, 124, 44, 1) (1, 44, 124, 204, 244, 204, 124, 44, 1) (1, 44, 124, 164, 164, 124, 44, 1) (1, 44, 84, 84, 84, 44, 1) (1, 72, 203, 268, 203, 72, 1) (1, 72, 137, 137, 72, 1) (1, 61, 225, 385, 385, 225, 61, 1) (1, 61, 225, 329, 225, 61, 1) (1, 61, 169, 169, 61, 1) (1, 52, 148, 244, 340, 436, 436, 340, 244, 148, 52, 1) (1, 52, 148, 244, 340, 388, 340, 244, 148, 52, 1) (1, 52, 148, 244, 292, 292, 244, 148, 52, 1) (1, 52, 148, 196, 196, 196, 148, 52, 1) (1, 52, 100, 100, 100, 100, 52, 1) Table 3. The 33 reflexive simplices described in Remark 5.9 and their h∗ -polynomials. 6. Future Directions The results of this paper set the stage for a detailed investigation of Conjecture 1.3 in the special case of the reflexive simplices of the form ∆(1,q) . As such, these results provide some obvious directions for future work in discrete geometry as outlined by the following questions. First, Remark 5.8 demonstrates that a q-vector satisfying Lemma 3.2 need not be h∗ -unimodal. However, it is not clear in general what role Lemma 3.2 plays regarding h∗ -unimodality and the structure of Hilbert bases for ∆(1,q) , leading to the following question. Question 6.1. Suppose r = (r1 , . . . , rk ) supports a reflexive simplex ∆(1,q) with q-vector satisfying Lemma 3.2.
What constraints on the Hilbert basis of cone ∆(1,q) , if any, are implied by the linear system in Lemma 3.2? Further, for which divisibility patterns in the r-vector, if any, is Lemma 3.2 sufficient to imply h∗ -unimodality for ∆(1,q) ? Which families of q-vectors, if any, both satisfy Lemma 3.2 and fail to be h∗ -unimodal? Second, in Sections 4 and 5, h∗ -unimodality for some but not all of the cases studied follows from an affine free sum decomposition of the simplices. Free sum constructions were used by Payne [16] to produce reflexive polytopes with non-unimodal h∗ -polynomials, and by Braun and Davis [4] to study the IDP condition DETECTING THE INTEGER DECOMPOSITION PROPERTY 23 for reflexive simplices. Given this, it is natural to ask when the free sum construction underlies the h∗ unimodality observed in the context of Section 3. Question 6.2. For each fixed dimension d, what fraction of reflexive IDP simplices of the form ∆(1,q) arise via affine free sums? Third, the reflexive simplices satisfying the necessary condition of Lemma 3.2 correspond to positive integral solutions to multivariate systems of algebraic equations. A full understanding of these simplices may require tools from algebraic geometry and algebraic number theory. Question 6.3. What are the integral solutions to the multivariate system of algebraic equations arising from Lemma 3.2? Fourth, as far as the authors know, in all published proofs that a family of polytopes is both IDP and h∗ -unimodal, IDP and h∗ -unimodality are established separately. Thus, the IDP has only been observed to correlate with h∗ -unimodality; no proofs have shown a causal link between the two properties. For example, in well-known work of Bruns and Römer [6], the IDP condition is a consequence of the existence of a regular unimodular triangulation, while h∗ -unimodality is established by using that triangulation to induce an application of the g-theorem. Thus, there is a need for proofs that demonstrate explicitly how the IDP condition might be used to prove h∗ -unimodality directly. The following question is a step in this direction. Question 6.4. Do there exist q-vectors for which h∗ -unimodality of ∆(1,q) can be established using only the conditions of Theorem 2.3? As a final remark, the results of this paper offer new avenues by which to search for counterexamples to Conjecture 1.3. In the special case of the simplices ∆(1,q) , the expression of the h∗ -polynomial arising from Theorem 2.2 provides a direct link between the structure of a q-vector and the behavior of the associated h∗ polynomial. A deeper study of general polynomials of the form given in Theorem 2.2 might provide insight leading to a counterexample, if one exists. Another possible direction in which to search for counterexamples is suggested by Question 6.1, through the identification of broad families of q-vectors that simultaneously satisfy Lemma 3.2 and fail to be h∗ -unimodal. Even if a counterexample is found to the general conjecture, improved explanations of the broadly-observed h∗ -unimodality for reflexive polytopes, including in the special case of ∆(1,q) , is needed. References [1] Matthias Beck, Pallavi Jayawant, and Tyrrell B. McAllister. Lattice-point generating functions for free sums of convex sets. J. Combin. Theory Ser. A, 120(6):1246–1262, 2013. [2] Matthias Beck and Sinai Robins. Computing the continuous discretely. Undergraduate Texts in Mathematics. Springer, New York, second edition, 2015. Integer-point enumeration in polyhedra, With illustrations by David Austin. [3] Benjamin Braun. Unimodality problems in Ehrhart theory. In Recent trends in combinatorics, volume 159 of IMA Vol. Math. Appl., pages 687–711. Springer, [Cham], 2016. [4] Benjamin Braun and Robert Davis. Ehrhart series, unimodality, and integrally closed reflexive polytopes. Ann. Comb., 20(4):705–717, 2016. [5] Francesco Brenti. Log-concave and unimodal sequences in algebra, combinatorics, and geometry: an update. In Jerusalem combinatorics ’93, volume 178 of Contemp. Math., pages 71–89. Amer. Math. Soc., Providence, RI, 1994. [6] Winfried Bruns and Tim Römer. h-vectors of Gorenstein polytopes. J. Combin. Theory Ser. A, 114(1):65–76, 2007. [7] Heinke Conrads. Weighted projective spaces and reflexive simplices. Manuscripta Math., 107(2):215–227, 2002. [8] Eugène Ehrhart. Sur les polyèdres rationnels homothétiques à n dimensions. C. R. Acad. Sci. Paris, 254:616–618, 1962. [9] Daniel R. Grayson and Michael E. Stillman. Macaulay2, a software system for research in algebraic geometry. Available at http://www.math.uiuc.edu/Macaulay2/. [10] Takayuki Hibi. Flawless O-sequences and Hilbert functions of Cohen-Macaulay integral domains. J. Pure Appl. Algebra, 60(3):245–251, 1989. [11] Takayuki Hibi. Algebraic Combinatorics on Convex Polytopes. Carslaw Publications, Australia, 1992. [12] Takayuki Hibi. Dual polytopes of rational convex polytopes. Combinatorica, 12(2):237–240, 1992. [13] Mircea Mustaţǎ and Sam Payne. Ehrhart polynomials and stringy Betti numbers. Math. Ann., 333(4):787–795, 2005. [14] Benjamin Nill. Volume and lattice points of reflexive simplices. Discrete Comput. Geom., 37(2):301–320, 2007. [15] Hidefumi Ohsugi and Takayuki Hibi. Special simplices and Gorenstein toric rings. J. Combin. Theory Ser. A, 113(4):718– 725, 2006. [16] Sam Payne. Ehrhart series and lattice triangulations. Discrete Comput. Geom., 40(3):365–376, 2008. [17] Richard P. Stanley. Decompositions of rational convex polytopes. Ann. Discrete Math., 6:333–342, 1980. Combinatorial mathematics, optimal designs and their applications (Proc. Sympos. Combin. Math. and Optimal Design, Colorado State Univ., Fort Collins, Colo., 1978). 24 BENJAMIN BRAUN, ROBERT DAVIS, AND LIAM SOLUS [18] Richard P. Stanley. Log-concave and unimodal sequences in algebra, combinatorics, and geometry. In Graph theory and its applications: East and West (Jinan, 1986), volume 576 of Ann. New York Acad. Sci., pages 500–535. New York Acad. Sci., New York, 1989. Department of Mathematics, University of Kentucky, Lexington, KY 40506–0027 E-mail address: benjamin.braun@uky.edu Department of Mathematics, Michigan State University, East Lansing, MI 48824 E-mail address: davisr@math.msu.edu Matematik, KTH, Stockholm, Sweden E-mail address: solus@kth.se 

0math.AC

Data-Driven Approximate Abstraction for Black-Box Piecewise Affine Systems arXiv:1801.09289v2 [cs.SY] 30 Jan 2018 Gang Chen and Zhaodan Kong Abstract— How to effectively and reliably guarantee the correct functioning of safety-critical cyber-physical systems in uncertain conditions is a challenging problem. This paper presents a data-driven algorithm to derive approximate abstractions for piecewise affine systems with unknown dynamics. It advocates a significant shift from the current paradigm of abstraction, which starts from a model with known dynamics. Given a black-box system with unknown dynamics and a linear temporal logic specification, the proposed algorithm is able to obtain an abstraction of the system with an arbitrarily small error and a bounded probability. The algorithm consists of three components, system identification, system abstraction, and active sampling. The effectiveness of the algorithm is demonstrated by a case study with a soft robot. I. I NTRODUCTION The proliferation of cyber-physical systems (CPSs) brings how to effectively and reliably guarantee their correct behaviors to the forefront of problems we as control engineers need to address. One natural choice to attain correct functioning is to consider formal methods techniques, such as model checking [1], [2], which have been successfully used in the formal verification and synthesis of digital circuits and software codes [3]. In recent years, we have seen many efforts of extending formal methods to engineering applications, e.g., automobiles [4], [5] and robotics [6], [7]. One crucial component of formal methods is a precise and potentially concise mathematical model of the system under investigation. However, in reality we rarely have full knowledge of complex CPSs during their design and even testing phase. Thus how to attain formal guarantee for systems with partially or fully unknown dynamics becomes a problem of practical significance. In this paper, we aim to address this problem in the context of abstraction [1]. Given a system model T (which can potentially have infinitely many states) and a formal specification φ written, for instance, in linear temporal logic (LTL), an abstract model of T is a simpler model T 0 ; checking whether the simpler model T 0 satisfies φ suffices to decide whether T satisfies φ [8], [9], [10], [11]. For systems that can be described by discrete state models, abstraction can be achieved by using the concepts of simulation and bisimulation [1], [2]. In control community, recently there have been many successful efforts pertaining to the abstraction of systems of more realistic dynamics, such as those that are piecewise affine [12], [13], [14], [15]. All these studies, as far as we know, assume models of known dynamics, which Authors are with the Department of Mechanical and Aerospace Engineering, University of California, Davis. Z. Kong is the corresponding author (email: zdkong@ucdavis.edu). significantly impedes the application of abstraction in the analysis and design of systems with inherent uncertainties, e.g., those needing to interact with a variety of human users and be deployed in a variety of environments. One principle way of mitigating uncertainties is to utilize machine learning techniques. Actually, the integration of formal methods and machine learning has shown great potential in the formal specification, design, verification, and validation of CPS [16], [17], [18], [19], [20], [21]. In this paper, we will focus on how to combine machine learning techniques, particularly system identification and active learning, and formal methods techniques to generate approximate abstractions for systems with black-box (unknown), piecewise affine (PWA) dynamics. PWA models partition the state space into a finite number of polyhedral regions and consider affine dynamics in each region [22]. It has been shown that PWA models can approximate nonlinear dynamics with arbitrary accuracy [12]. Moreover, there exist efficient techniques for the identification of PWA systems, e.g., optimization-based methods and clustering-based methods [23]. The major contribution of this paper is that it addresses many theoretical and algorithmic issues pertaining to the integration of existing approximate abstraction techniques [24] and system identification techniques [23]. Given a system with unknown PWA dynamics, the paper shows that, by following the algorithm prescribed in the paper, it is possible to extract an abstract model with an arbitrarily small error and a bounded probability (under certain mild assumptions). Even though the paper focuses on PWA systems, the preliminary results obtained in it can potentially pave the way for future developments for systems with more complex dynamics. The remainder of the paper is organized as follows. In Section II, we provide preliminaries and notation used throughout the paper. In Section III, we formally introduce the abstraction problem that will be solved in the paper. Section IV presents our data-driven approximate abstraction algorithm, together with proofs demonstrating the effectiveness of our algorithm. Section V uses a soft robot system as an example to showcase our proposed algorithm. We conclude with final remarks in Section VI. II. P RELIMINARIES AND N OTATION A N dimensional polytope X is defined as the convex hull of at least N + 1 affinely independent vectors in RN . A complete partition of X is a set of open polytopes Xi , i ∈ I (I is a finite index set) in RN such that Xi1 ∩ Xi2 = ∅ for all i1 , i2 ∈ I, i1 6= i2 and cl(X ) = ∪i∈I cl(Xi ), where cl(Xi ) denotes the closure set of Xi . According to the Hrepresentation, each Xi , i ∈ I can be represented as Xi = {x ∈ RN : Hi x ≺ Ki }, where ≺ denotes componentwise inequality. A piecewise affine (PWA) system [22] can be written as follows: xk+1 = f (xk ) + e   A1 x + b1 if x ∈ X1 .. f (x) = .   As x + bs if x ∈ Xs (1) where xk is the state of the system at step k; f : X → RN is a PWA map; e ∈ N (0, σe2 ) is an independently, identically distributed and zero mean Gaussian noise with standard deviation σe ; s is the number of modes; Ai , bi are the parameters of the i-th mode (Ai , i = 1, · · · , s is assumed to be nonsingular in this paper); and all s modes together constitute a complete partition of X . A transition system is a tuple T = (Q, δ, O, o), where Q is the state space; δ : Q → 2Q (2Q is the powerset of Q) is a transition map assigning a state q ∈ Q to its next state q 0 ∈ Q; O is the set of observations; and o : Q → O is an observation map assigning each q ∈ Q an observation o(q) ∈ O [1]. We denote a region of the state space as P ⊂ Q. The embedding transition system of a PWA system S described by Eqn. (1) is a tuple Te = (Qe , δe , Oe , oe ), where Qe = ∪i∈I Xi ; δe : x → x0 iff there exists i ∈ I such that the transition from x to x0 satisfies Eqn. (1); Oe = I; and oe (x) = i iff x ∈ Xi [24]. For a transition system T , including embedding transition systems of PWA systems, the successor of a region P ⊂ Q is define as the set of states that can be reached from the states in P in one step, i.e., P ost(P ) = {q ∈ Q | ∃p ∈ P with p → q}. The predecessor of a region P ⊂ Q can be defined similarly as P re(P ) = {q ∈ Q | ∃q ∈ P with q → p}. A state q ∈ Q is called reachable if there exists a finite execution ending at q. We denote all the reachable states of T by Reach(T ). Given an LTL formula φ over O and a system T , if all the traces originating from a region P ⊂ Q satisfy φ, then we denote the situation as T (P ) |= φ. Let XTφ = {X ⊂ Q, T (X) |= φ} denote the largest region of T from which all the traces satisfy φ [1]. The reachability metric over two transition systems T1 and T2 is defined as [25]: d(T1 , T2 ) = h(Reach(T1 ), Reach(T2 )), where h is the Hausdorff distance. Given two transition systems T1 and T2 with the same observation set O and a reachability metric d defined over them, a relation Sσ ⊆ Q1 × Q2 is called a σ−approximate simulation relation of T1 by T2 [25] if for all (q1 , q2 ) ∈ Sσ : • d(o(q1 ), o(q2 )) ≤ σ, 0 0 • ∀q1 = P ost(q1 ), there exists q2 = P ost(q2 ), such that 0 0 (q1 , q2 ) ∈ Sσ . Moreover, T1 is said to be σ−approximately simulated by T2 , denoted T1 ≺σ T2 . III. P ROBLEM S TATEMENT Formally, in this paper, we wish to solve the following problem: Problem 1: Given a PWA system S with unknown dynamics, an LTL specification φ, and a bound σ > 0, find a finite transition system T̂ such that p(T̂ ≺σ T ) > 1 − δ, where p(.) stands for probability, T is the true abstract transition system of S, and δ is bounded. Remark 1: By unknown dynamics, we mean that the following system characteristics are unknown: (i) the number of modes, s, (ii) the parameters related to the dynamics of each mode, {Ai , bi , i = 1, · · · , s}, (iii) the parameters related to the partitions (regions) of the state space, {Hi , Ki , i = 1, · · · , s}, and (iv) the standard deviation of the Gaussian noise, σe . But we assume that our algorithm, which will be presented in the next section, can use the system as a black-box simulator to generate samples. This is a reasonable assumption since during system design and testing phases, engineers can always get access to a full-scale system model, a scaled system model, or a computer simulation to generate samples [26], [21]. Remark 2: Notice that the requirement p(T̂ ≺σ T ) > 1 − δ is inspired by the concept of probably approximately correct (PAC) models in machine learning [27]. It simply says that the probability that the transition system T̂ (obtained by using our algorithm) is a σ-approximate simulation by the PWA system S is higher than 1 − δ. In other words, given a PWA system with unknown dynamics, we intend to find out its approximate abstract transition system T̂ with a high enough confidence. IV. DATA -D RIVEN A BSTRACTION A LGORITHM Black-Box PWA System S Sample Pairs (xi,yi) System Identification YES Refine? NO LTL Formula φ System Abstraction NO Active Sampling YES Terminated? Abstraction T Fig. 1: Architecture of our data-driven abstraction algorithm. Fig. 1 illustrates the basic architecture of our data-driven abstraction algorithm to solve Problem 1. The inputs of the algorithm are a black-box PWA system S with unknown dynamics and an LTL specification φ; the output of the system is a transition system T . The algorithm can be roughly divided into three components: system identification, system abstraction, and active sampling. The goal of the system identification component is to derive an estimated PWA model Ŝ based on the data sampled from the black-box system S (serving as a simulator); the goal of the system abstraction component is to derive a transition system T given the identified PWA model Ŝ and the specification φ; one important procedure of the system identification component is refinement, which refines the estimated model Ŝ, until no significant improvement can be achieved, based on the currently available data and the current abstraction T ; finally, if no satisfactory abstraction T can be found after the refinement, the active sampling component will be implemented to draw new data points with the help of the black-box simulator S. In the following sub-sections, we will present each of the three components. Proofs regarding the effectiveness of our algorithm will be provided at the end of the section. Algorithm 1: System Identification Initialization Input : A bound σ̂, a set of samples D := {(yk , xk )}, k ∈ {1, · · · , K}, a ratio 0 < r < 1, and a large number J Output: Âi , b̂i , Di , Ĥi , K̂i , i = 1, · · · , s 1: Set l = 0 and I0 = {1, · · · , K}; 2: for |Il+1 | ≥ rK do 3: Randomly generate a set of parameters {Aj , bj }, j ∈ {1, · · · , J}; 4: Construct sets Σj = {k yk − (Aj xk + bj ) k≤ σ̂, k ∈ Il } for each j ∈ {1, · · · , J}; 5: Set l = l + 1 and Σmax = argmaxJj=1 |Σj |; 6: With Σmax , estimate Âl and b̂l by solving |Σmax | (Âl , b̂l ) = argmin A,b A. System Identification Given a black-box PWA system S, or subsequently a set of K samples D := {yk , xk }, k = 1, · · · , K, the system identification component identifies a PWA model, specified by the number of modes s as well as the mode parameters Âi , b̂i , Ĥi and K̂i with i = 1, · · · , s. The problem itself is a well-studied problem. Specifically, we need to find (i) a minimum positive integer, s, (ii) a set of parameter matrices, {Âi }si=1 and {Ĥi }si=1 , and (iii) a set of parameter vector {b̂i }si=1 and {K̂i }si=1 (notice that {Ĥi }si=1 and {b̂i }si=1 together constitute a complete partition {Xi }si of the PWA system’s state space S), such that the estimated parameters are the solution of the following minimization problem: (Âi , b̂i , Ĥi , K̂i ) = K 1 X c(yk − fˆ(xk )) (2) (xk ,yk )∈(D∩Xi ) K argmin k=1 where fˆ(.) is specified by Âi , b̂i , Ĥi and K̂i , i = 1, · · · , s and c is a given penalty function, which is chosen to be c(·) = || · ||2 in this paper. Notice that solving the identification problem involves the simultaneous solving of two subproblems, data classification and parameter estimation. Once the data points have been classified into clusters {Di }si=1 such that (yk , xk ) ∈ Di , i.e., (yk , xk ) is attributed to the ith mode, mode parameters Âi , b̂i , Ĥi and K̂i can be easily estimated by solving Eqn. (2). Our system identification component is modified from the method proposed in [22]. It consists of two main procedures: initialization and refinement. One major difference between our method and the one in [22] is that we utilize the current abstract transition system to guide the refinement. 1) Initialization: The pseudo code of the initialization procedure is shown in Alg. 1. The steps are rather selfexplanatory. Here we would like to provide a few simple comments for clarification. We need to randomly generate a matrix Âj and a vector b̂j (Line 3) in each loop, which is quite inefficient; thus the termination condition |Il+1 | ≥ rK (Line 2) can be set loosely, i.e., with a rather large r. Such a practice is reasonable, given the fact that the initialization procedure is only meant to generate some good X c(yk − (Axk + b)); k=1 Set Dl = {k ∈ Il :k yk − (Âl xk + b̂l ) k≤ σ̂} and Il+1 = Il \ Dl ; 8: Find boundaries specified by {Ĥ}li=1 and {K̂}li=1 between sets {Di }li=1 ; set s = l. 7: enough partitions (modes), which will be further refined in the refinement procedure. As for the boundaries specified by {Ĥ}li=1 and {K̂}li=1 between sets {Di }li=1 (Line 8), standard support vector machines (SVM) regression methods [28] can be used to compute them. 2) Refinement: Given the random nature of the way Âi , b̂i , i = 1, · · · , s are generated in Alg. 1, it is quite unlikely that we are able to identify all the correct modes with the initialization procedure. Potentially there are two main issues: • Undecidable data points: these are the data points that belong to more than one mode, i.e., they satisfy k yk − (Âi xk + b̂i ) k≤ σ̂ for more that one i = 1, · · · , s; • Unfeasible data points: these are the data points that don’t belong to any mode, i.e., there is no i = 1, · · · , s such that k yk − (Âi xk + b̂i ) k≤ σ̂. We refine the identified system model by eliminating these two types of data points as follows: For undecidable points, we can reassign them based on their maximum likelihood with respect to all modes, i.e., we assign each undecidable point (xk , yk ) to the optimal mode according to the following rule: CR(k) = argmax CC(xk , yk )i /CC(xk , yk ), (3) i=1,··· ,s where CC(xk , yk ) is the total number of data points that are within D and inside the hyper ball region centered at (xk , yk ) and with a predefined radius ρ, and CC(xk , yk )i is the number of data points that are in CC(xk , yk ) and belong to mode i. For unfeasible points, given the current abstract transition system T , we discard those points that meet the following simply using the definition of embedding transition system. Third, an observation map oe is created by partitioning the state space of the system Te into uniform grids. Fourth, given the system Te , the observation map oe , and the LTL formula φ (or its corresponding Buchi automaton Bφ ), an initial transition system T0 is constructed by following standard abstraction procedures, such as those prescribed in [1]. Fifth, a product automaton P is constructed as P = T0 × Bφ , which concerns both the initial transition system T0 and the specification φ. The product automaton is a tuple P = (Sp , Sp0 , δp , Fp ), where Sp is the set of states, Sp0 is the set of initial states, δp is the transition map, and Fp is the acceptance condition. Finally, refinement is conducted by solving a deterministic Rabin game. Algorithm 2: System Identification Refinement Input : Current abstract transition system T , parameters Âi , b̂i , Di , Ĥi , K̂i , i = 1, · · · , s obtained in Alg. 1, and a bound σ̂ Output: Âi , b̂i , Ĥi , K̂i , i = 1, · · · , s Set β ≥ 0, µ ≥ 0, κ ≥ 0, 0 < θ < 1, l = 1; while Not terminated do 1: Compute (i∗ , j ∗ ) = argmin1≤i<j≤s βi,j with βi,j =k Âi − Âj k; 2: if βi∗ ,j ∗ ≤ θ t β then 3: Merge modes i∗ and j ∗ ; Set s = s − 1; 4: Recompute Di∗ = {k ∈ {1, · · · , K} : k yk − (Âi∗ xk + b̂i∗ ) k≤ σ̂}; 5: Use CR() and rule Dis() to reassign points; / ∗ CR() and Dis() are defined in the text as Eqn. (3) and Eqn. (4), respectively∗/ 6: Compute i∗ = argmini=1,··· ,s |Di |/|D|; 7: if |Di |/|D| ≤ θ t µ then 8: Discard mode i∗ ; let s = s − 1; 9: Go to Step 4; s 10: Store {Âi }si=1 as {Âold i }i=1 ; 11: Update Âi , b̂i , Ĥi , K̂i with new {Di }si=1 ; 12: if k Âi − Âold k≤ κ then i 13: Terminated. 14: else 15: l = l + 1; end Algorithm 3: System Abstraction Refinement Input : Current abstract transition system T , current product automaton P, an initial state q of T , and a ratio 0 < η < 1 Output: Refined abstract transition system T̂ and refined product automaton P̂ Initialize T̂ = T , P̂ = P, and Su = ∅; while |Su | ≥ η|Q| do 1: 2: 3: 4: 5: 6: condition: Dis(k) = {(xk , yk ) ∈ D|d(yk , o−1 (P ost(o(xk )))) ≥ σ̂}, (4) where d(., .) is the Hausdorff distance, o(xk ) maps a continuous state xk to a discrete state of the transition system T (it would be helpful for the readers to review the definitions in Section II), P ost(o(xk )) maps the region corresponding to o(xk ) to its successor region, and finally o−1 (P ost(o(xk )))) finds the set of continuous, PWA states corresponding to the region P ost(o(xk )). The remaining unfeasible data points are reassigned according to the rule CR(·) (Eqn. (3)). The pseudo code of the whole refinement procedure is shown in Alg. 2. B. System Abstraction Given an estimated PWA model Ŝ (or fˆ), parameterized by Âi , b̂i , Ĥi and K̂i with i = 1, · · · , s, and an LTL formula φ, the goal of the system abstraction component is to generate a good enough abstraction T . We roughly follow the approximate abstraction procedures described in [24] to design and implement the system abstraction component. Here we are just going to provide a rough outline of the abstraction algorithm. Interested readers can refer to [24] for more details. First, a deterministic Buchi automaton Bφ is constructed from the formula φ. Second, the corresponding embedding transition system Te is constructed for Ŝ by 7: 8: 9: 10: 11: Compute S> and S⊥ for P̂; Set Su := Pp \ (S> ∪ S⊥ ); /∗See text for the definitions of S> , S⊥ , and Su ∗ / for all (q, g) ∈ Su do if q ∈ Su then Set q := q; for all (∃qr ∈ q, q 0 ∈ P ost(qr ) and (qr ∩ P re(q 0 )) 6= ∅) do Construct states q1 , q2 such that q1 := qr ∩ P re(q 0 ), q2 := qr \ P re(q 0 ); q := (q \ qr ) ∪ {q1 , q2 }; q := q; Update δ and o for T̂ ; Update P̂ and T̂ . end The pseudo code of the abstraction refinement procedure is shown in Alg. 3. Given the current abstract transition system T and the current product automaton P, a state q ∈ Q of T falls into one of the following three categories: • S> : the set of states from which all traces are accepted by P, • S⊥ : the set of states from which no trace is accepted by P, • Su : the set of states from which some but not all traces are accepted by P. The goal of the refinement is thus to eliminate Su . This leads to the termination condition of the refinement procedure as |Su | < r|Q|, i.e., the refinement will be terminated once the volume of Su , |Su |, is smaller than a fraction r (specified by the user) of the volume of Q (the state space of T ). C. Active Sampling The system identification component and the system abstraction component described in the last two sub-sections are based on a fixed data set D := {(yk , xk )}K k=1 . It is quite obvious that the quality of the system identification and the system abstraction depends on the quality of the data set. To improve the quality of the system identification (in other words, to decrease the number of data points needed for the system identification), we use an active learning algorithm developed by our group [29] to sample high quality (or “informative”) data points for the system identification component after the initial round (see Fig. 1). The strategy to find the next data point to sample for round t + 1 has two steps. In the first step, the best candidate for each mode is identified as follows: 1/2 xi := argmax(Ψi,t (x) + λt %i,t (x)) (5) x∈X̂i where Ψi,t (x) is the Gaussian process regression mean of the prediction error defined over the data points in Di , %i,t (x) is the Gaussian process regression variance of the prediction error defined over the data points in Di , and λt is a regularization factor. In the second step, the active learning algorithm chooses mode i∗ = argmini=1,··· ,s (maxx∈X̂i Ψi,t (x)) and the corresponding best candidate xi∗ to sample. second condition in the definition of approximate relation is satisfied. Therefore, T̂ ≺ε T . Lemma 2: Given an estimation fˆ(.) (or Ŝ) of the PWA system f (.) (or S), if k fˆ(x) − y k≤ η1 holds with probability > 1 − α1 , where α1 ∈ (0, 1), η1 > 0, then the system abstraction component described in Section IVB can derive an abstract transition system T̂ such that it is (η1 + ε)−approximately simulated by the real abstract transition system with a probability greater than 1 − α1 . Proof: Here for the sake of clarify, the notations in this proof that are slightly different from the ones used in the other parts of the paper. Let’s use f and fˆ to denote the real and estimated dynamics of the PWA system, respectively. Moreover, let’s call their true abstract transition systems as Treal and T̂real , respectively. Finally, let’s call the approximate abstract transition systems obtained by using the system abstraction component described in Section IVB as Tapp and T̂app , respectively. According to Lemma 1, we have Tapp ≺ε Treal and T̂app ≺ε T̂true . Then set k eB k=k fˆ(x) − y k≤ η1 , such that y = fˆ(x) + y − fˆ(x) = fˆ(x) + eB , which can be seen a PWA system with bounded noise, following [24], we have T̂real ≺η1 Treal . Since the relationship ≺. is transitive, we have T̂app ≺η1 +ε Treal . The conclusion regarding probability follows easily. Lemma 3: Provided with Assumption 1, k fˆ(x) − y k≤ η2 holds with ≥ 1 − α2 , where α2 = 1 − R η2 probability 2 √ 1 exp(−υ /(2(σ e + C)))dυ. −η2 2π(σe +C) D. Theoretical Results Regarding the Effectiveness of Our Algorithm We add the following assumption regarding the performance of the PWA system identification. Assumption 1: Assume the prediction error of the system identification component described in Section IV-A can be characterized by a zero mean Gaussian with a bounded variance, i.e., f (x) − fˆ(x) ∼ N (0, σp (x)2 ) and σp (x) ≤ C, where f (x) is the real PWA dynamics and fˆ(x) is the estimated PWA dynamics. Remark 3: The system identification of PWA system is still an open problem and has been proven to be NP-hard [30], [31]. In [22], the authors were able to demonstrate that, for a fixed data set, the error is bounded, i.e., |f (x)− fˆ(x)| < σ for any σ > 0. Thus, we believe our assumption here, even though unproven, is still reasonable. We have the following three lemmas regarding the integration of identification and abstraction (without active sampling in the loop). Lemma 1: Given a PWA system S with known dynamics, for any bound ε > 0, Alg.3 can derive an abstract transition system T̂ that is ε−approximately simulated by the real abstract transition system T of S, i.e., T̂ ≺ε T . Proof: Consider a relationship V (q1 , q2 ) = {(q1 , q2 )|d(q1 , q2 ) ≤ ε}, and set r|Q| ≤ ε. As (q1 , q2 ) ∈ V implies d(q1 , q2 ) ≤ ε, the first condition in the definition of approximate relation is satisfied. Then for all d(q1 , q2 ) < ε, the conclusion in [24] shows that the result of Alg.3 can guarantee that d(P ost(q1 ), P ost(q2 )) ≤ |Su | ≤ ε. Then the Proof: Set eT := y− fˆ(x) = f (x)+e− fˆ(x), where e is the true Gaussian noise, e ∼ N (0, σe2 ). Since f (x) − fˆ(x) ∼ N (0, σp (x)2 ), we have eT ∼ N (0, σp2 +σe2 ). The probability that k fˆ(x) − y k≤ η2 can then be computed. When σp = C, the probability reach value, such that α2 = R η2 the minimum 2 1− √ 1 exp(−υ /(2(σ + C)))dυ. e −η2 2π(σe +C) We have the following two lemmas regarding the performance of our active sampling component. Lemma 4: For any ρ ∈ (0, 1), if λt = 2B + 300γt log3 (t/ρ), where γt is the maximum information gain defined in [29], then 1/2 k Ψi,t (x) k≤ λt %i,t (x) (6) holds with probability > 1 − ρ. Proof: The lemma can be proved by following similar steps as the proof of Lemma 2 in [29]. Lemma 5: For any σ > 0, if λt = 2B + 300γt log3 (t/ρ), where t is the number of points in Di , then there exists t ≤ T, 1/2 λt %i,t (x) ≤ σ, ∀x ∈ Di (7) holds with finite number of sample time T . Proof: According to Theorem 5 in [32], for any bounded Xi ⊂ X , the information gain for exponential kernel is γT = O((logT )d+1 ), where d is the of X . Lemma 7.1 in [32] shows that PT dimension 2α −2 2 min{σ % i,t−1 (x), α} ≤ log(1+α) γT , ∀α > 0. As the t=1 information gain γT is convergent, such that the variance %i,t (x) will be convergent to zero. As λt is bounded, 1/2 λt %t (x) will be convergent to zero, then the lemma has been proved. Finally, we can prove that the algorithm presented in this paper can solve Problem 1 with the relatively moderate Assumption 1. Theorem 1: Given a PWA system S with unknown dynamics, for any σ > ε > 0, the algorithm described in this paper (with the assumption that the standard deviation of the prediction error for the system identification component is bounded by C) can obtain an approximate abstract transition system T̂ that is σ-approximately simulated by the true abstract transition system T of S with a bounded probability Rthat is greater than 1 − δ, where δ = 1 − σ−ε √ 1 exp(−υ 2 /(2(σe + C)))dυ. −σ+ε 2π(σe +C) Proof: Set δ = α2 and η2 = σ − ε. According to Lemma 3, Lemma 4, and Lemma 5, the system identification component together with the active sampling component can achieve an estimation k fˆ(x) − y k≤ η2 with probability > 1 − α2 . Then with this estimation result, according to Lemma 1 and Lemma 2, the system abstraction component can generate an abstract transition system T̂ that is (η2 + ε)−approximately simulated by the real abstract transition system T with probability > 1 − δ. V. C ASE S TUDY In this section, we will use a soft robot driven by series pneumatic artificial muscles as an example to demonstrate our algorithm. A. Model and LTL Specification Fig. 2: (a) Geometrical model of the soft robot driven by series pneumatic artificial muscles (sPAM) used in the case study. (b) Simplified planar robot with x = (x(1), x(2))T as the coordinate vector of the moving platform. In this case study, we will focus on a particular type of soft robots, those that are driven by series pneumatic artificial muscles (sPAM) [33]. Fig. 2 shows an example of such a robot. It has two polyethylene tubing sPAMs, each of which is controlled by a corresponding actuator with pressurized air. Even though the dynamics of such a robot is nonlinear, in [34], the authors have shown that its closed-loop behavior can be approximated by piecewise affine dynamics. Here let’s assume that the robot under investigation has dynamics as follows: xk+1 = f (xk ) + e (8) and    1 0   x if 0 ≤ xk (1) ≤ 0.3  0 0.98    f (x) = 0.83 0.12 0.01   x + if xk (1) ≥ 0.3  0.12 0.81 0.03 where x = (x(1), x(2))T is the coordinate of the moving platform and e is a Gaussian noise with a zero mean and a standard deviation of 0.1. Please keep in mind that the model is unknown to our algorithm but the model itself can be used by our algorithm as a simulator to generate samples. The specification that needs to be verified is written as an LTL formula φ := (π1 ∧ ♦π2 ), where “π1 := x(1) is below 0.3", “π2 := x(2) is above 0.6", and  is the temporal operator “Always”, ♦ is the temporal operator “Eventually”. Put together, φ specifies that “it should always be true that x(1) is below 0.3 and eventually x(2) is above 0.6”. B. Implementation Results The algorithm proposed in this paper is implemented as a Matlab tool. The tool takes an LTL formula and a black-box PWA system as inputs and it outputs an abstract transition system. The active learning used in the tool is the Gaussian Process Adaptive Confidence Bound (GP-ACB) algorithm proposed and implemented by our group in [29]. Moreover, a Gaussian kernel function is used in GP-ACB. In order to demonstrate the effectiveness of our algorithm, here we will provide two sets of implementation results, one regarding the system identification component and the other one regarding the entire algorithm. 1) Performance of the System Identification Component: Here we compare the performance of our system identification component with an existing state-of-the-art black-box system identification tool called the Hybrid Identification Tool (HIT) [35]. One thing we would like to point out here is that the evaluation of our system identification component is conducted in conjunction with other components, e.g., abstraction and active sampling, in the loop. It should be expected that, at worst, our component should have the same performance as the HIT. However, it may also be expected that factors such as active-learning based sampling and abstraction-guided refinement can potentially improve the identification performance, which turns out to be the case, at least for this particular case study. We use two metrics to quantify system identification errors: parameter estimation error and region estimation error. Given the parameters of a real PWA model (unknown to the investigated algorithms), the parameter estimation error is the sum of the Euclidean distances between the real parameters and the estimated parameters. The region estimation error is defined the sum of the following Hausdorff distance: e(X̂i , Xi ) = max{ sup inf d(x, y), sup inf d(x, y)}, x∈Xi y∈X̂i y∈X̂i x∈Xi (a) (b) Fig. 3: Comparison results of the system identification component used in this paper (green) and the HIT algorithm (red): (a) Average parameter estimation error with respect to the number of samples; (b) Average region (partition) estimation error with respect to the number of samples. where e(X̂i , Xi ) the region estimation error related to the ith mode with Xi as the real region or partition and X̂i as the estimated one, d(x, y) is the Euclidean distance between x and y, and sup and inf stand for supremum and infimum, respectively. To calculate the Hausdorff distance, we randomly generate 100 samples inside the state space X̂i and the state space Xi . The comparison results based on 5 trials are shown in Fig. 3. It shows that, averagely speaking, the system identification component proposed in this paper has a comparable or better performance than HIT. Particularly, Fig. 3b shows that our algorithm has a faster convergent rate regarding the region estimation error. This is probably due to the fact that, generally speaking, active learning, which is used in our algorithm, out-performs its randomly sampling counterpart, which is used in HIT. 2) Performance of the Entire Algorithm: As the true abstract transition system of the system, described by Eqn. (8), is unknown, here we use the abstract transition system obtained by using the algorithm proposed in [24] as a benchmark. We will call this abstract transition system as T ∗ . In [24], the authors have shown that even though their algorithm cannot attain the true abstract transition system, it still can get an abstract transition system that is arbitrarily close to the real one. Of course, we should point out that, in order to get this abstract transition system T ∗ , the algorithm in [24] should have access to the system model, i.e., T ∗ is generated with a known model. In parallel, we run our algorithm to extract an abstract transition system T̂ without access to the dynamics of the model, i.e., T̂ is generated with a black-box system with unknown dynamics. Then in order to demonstrate the effectiveness of our algorithm, we need to show that T̂ should approach T ∗ . This turns out to be the case, at least for this particular case study. Based on the system dynamics, Eqn. (8), and the LTL specification φ, we implement the abstraction algorithm proposed in [24] and obtain an abstract transition system T ∗ , which will be used as a benchmark. Then the algorithm proposed in this paper is applied to the same system, but with unknown dynamics, and the same LTL specification. The output of the algorithm is another abstract transition system T̂ . Here we use a metric that is inspired by the concept of approximate simulation to quantify the difference between the two transition systems. To be more specific, we set the abstraction error to σ if T ∗ is σ−approximately simulated by T̂ . Moreover, the metric σ is normalized to σ̄ by the volume of the state space, i.e., σ̄ := σ/|X |. TABLE I: Comparison result with respect to different number of samples and a fixed number of refinement steps, which is set to 20. σ̄ Number of samples 20 40 60 0.046 0.042 0.035 TABLE II: Comparison result with respect to different number of refinement steps and a fixed number of samples in active sampling component, which is set to 10. σ̄ Refinement Steps 5 10 20 0.077 0.068 0.050 The comparison results are shown in Table I and Table II. Table I shows the comparison results with respect to different number of samples and a fixed number of refinement steps in the system abstraction component. Table II shows the comparison results with respect to different number of refinement steps and a fixed number of samples in the active sampling component. The tables show the average normalized errors σ̄ based on 5 trials. The results show that the larger the number of refinement steps, the smaller the abstraction error; and the larger the number of samples, the smaller the abstraction error. The results also show that, even with a black-box system, our algorithm can attain an approximate abstract transition system T̂ that is close to the benchmark abstract transition system T ∗ . Fig. 4: State space partitions of the abstract transition system obtained by using our algorithm for the case study. The number of refinement steps and the number of samples in the active sampling component are both set at 20. VI. C ONCLUSIONS In this paper, we proposed a data-driven approximate abstraction algorithm for piecewise affine systems with unknown dynamics. We demonstrated both theoretically and empirically that given a black-box PWA system and an LTL specification, we were able to derive an abstract transition system that is approximately simulated by the true abstract transition system. We demonstrated the effectiveness of our proposed algorithm with a soft robot as a case study. ACKNOWLEDGMENTS This work was partially supported by the Hyundai Motor Company. R EFERENCES [1] C. Baier, J.-P. Katoen, and K. G. Larsen, Principles of model checking. MIT press, 2008. [2] R. Alur, Principles of cyber-physical systems. MIT Press, 2015. [3] D. A. Peled, Software reliability methods. Springer Science & Business Media, 2013. [4] M. Kamali, L. A. Dennis, O. McAree, M. Fisher, and S. M. Veres, “Formal verification of autonomous vehicle platooning,” Science of Computer Programming, 2017. [5] H. Abbas, B. Hoxha, G. Fainekos, and K. Ueda, “Robustness-guided temporal logic testing and verification for stochastic cyber-physical systems,” in Cyber Technology in Automation, Control, and Intelligent Systems (CYBER), 2014 IEEE 4th Annual International Conference on. IEEE, 2014, pp. 1–6. [6] Y. K. Lopes, S. M. Trenkwalder, A. B. Leal, T. J. Dodd, and R. Groß, “Supervisory control theory applied to swarm robotics,” Swarm Intelligence, vol. 10, no. 1, pp. 65–97, 2016. [7] A. Ulusoy, T. Wongpiromsarn, and C. Belta, “Incremental controller synthesis in probabilistic environments with temporal logic constraints,” The International Journal of Robotics Research, vol. 33, no. 8, pp. 1130–1144, 2014. [8] A. Girard, G. Pola, and P. Tabuada, “Approximately bisimilar symbolic models for incrementally stable switched systems,” IEEE Transactions on Automatic Control, vol. 55, no. 1, pp. 116–126, 2010. [9] E. Haghverdi, P. Tabuada, and G. J. Pappas, “Bisimulation relations for dynamical, control, and hybrid systems,” Theoretical Computer Science, vol. 342, no. 2-3, pp. 229–261, 2005. [10] P. Prabhakar, G. Dullerud, and M. Viswanathan, “Stability preserving simulations and bisimulations for hybrid systems,” IEEE Transactions on Automatic Control, vol. 60, no. 12, pp. 3210–3225, 2015. [11] G. Pola, P. Pepe, and M. D. Di Benedetto, “Symbolic models for timevarying time-delay systems via alternating approximate bisimulation,” International Journal of Robust and Nonlinear Control, vol. 25, no. 14, pp. 2328–2347, 2015. [12] B. Yordanov, J. Tumova, I. Cerna, J. Barnat, and C. Belta, “Temporal logic control of discrete-time piecewise affine systems,” IEEE Transactions on Automatic Control, vol. 57, no. 6, pp. 1491–1504, 2012. [13] E. A. Gol, X. Ding, M. Lazar, and C. Belta, “Finite bisimulations for switched linear systems,” IEEE Transactions on Automatic Control, vol. 59, no. 12, pp. 3122–3134, 2014. [14] J. A. DeCastro, V. Raman, and H. Kress-Gazit, “Dynamics-driven adaptive abstraction for reactive high-level mission and motion planning,” in Robotics and Automation (ICRA), 2015 IEEE International Conference on. IEEE, 2015, pp. 369–376. [15] G. Pola and M. D. Di Benedetto, “Symbolic models and control of discrete-time piecewise affine systems: An approximate simulation approach,” IEEE Transactions on Automatic Control, vol. 59, no. 1, pp. 175–180, 2014. [16] T. Dreossi, A. Donzé, and S. A. Seshia, “Compositional falsification of cyber-physical systems with machine learning components,” arXiv preprint arXiv:1703.00978, 2017. [17] K. Urayama, M. C. Fu, and S. I. Marcus, “Simulation-based work load and job release control for semiconductor manufacturing,” in Decision and Control (CDC), 2015 IEEE 54th Annual Conference on. IEEE, 2015, pp. 7329–7334. [18] S. A. Seshia, S. Hu, W. Li, and Q. Zhu, “Design automation of cyberphysical systems: Challenges, advances, and opportunities,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2016. [19] S. Haesaert, A. Abate, and P. M. Van den Hof, “Data-driven and model-based verification: A bayesian identification approach,” in Decision and Control (CDC), 2015 IEEE 54th Annual Conference on. IEEE, 2015, pp. 6830–6835. [20] A. Kozarev, J. Quindlen, J. How, and U. Topcu, “Case studies in data-driven verification of dynamical systems,” in Proceedings of the 19th International Conference on Hybrid Systems: Computation and Control. ACM, 2016, pp. 81–86. [21] Z. Kong, A. Jones, and C. Belta, “Temporal logics for learning and detection of anomalous behavior,” IEEE Transactions on Automatic Control, vol. 62, no. 3, pp. 1210–1222, 2017. [22] A. Bemporad, A. Garulli, S. Paoletti, and A. Vicino, “A bounded-error approach to piecewise affine system identification,” IEEE Transactions on Automatic Control, vol. 50, no. 10, pp. 1567–1580, 2005. [23] A. Garulli, S. Paoletti, and A. Vicino, “A survey on switched and piecewise affine system identification,” IFAC Proceedings Volumes, vol. 45, no. 16, pp. 344–355, 2012. [24] B. Yordanov, J. Tumova, I. Cerna, J. Barnat, and C. Belta, “Formal analysis of piecewise affine systems through formula-guided refinement,” Automatica, vol. 49, no. 1, pp. 261–266, 2013. [25] A. Girard and G. J. Pappas, “Approximation metrics for discrete and continuous systems,” IEEE Transactions on Automatic Control, vol. 52, no. 5, pp. 782–798, 2007. [26] X. Jin, A. Donzé, J. V. Deshmukh, and S. A. Seshia, “Mining requirements from closed-loop control models,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 34, no. 11, pp. 1704–1717, 2015. [27] L. Valiant, Probably Approximately Correct: Nature’s Algorithms for Learning and Prospering in a Complex World. Basic Books (AZ), 2013. [28] C. Cortes and V. Vapnik, “Support-vector networks,” Machine learning, vol. 20, no. 3, pp. 273–297, 1995. [29] G. Chen, Z. Sabato, and Z. Kong, “Active learning based requirement mining for cyber-physical systems,” in Decision and Control (CDC), 2016 IEEE 55th Conference on. IEEE, 2016, pp. 4586–4593. [30] L. Ljung, “Perspectives on system identification,” Annual Reviews in Control, vol. 34, no. 1, pp. 1–12, 2010. [31] F. Lauer, “On the complexity of piecewise affine system identification,” Automatica, vol. 62, pp. 148–153, 2015. [32] N. Srinivas, A. Krause, S. M. Kakade, and M. Seeger, “Gaussian process optimization in the bandit setting: No regret and experimental design,” arXiv preprint arXiv:0912.3995, 2009. [33] F. Daerden and D. Lefeber, “The concept and design of pleated pneumatic artificial muscles,” International Journal of Fluid Power, vol. 2, no. 3, pp. 41–50, 2001. [34] G. Andrikopoulos, G. Nikolakopoulos, I. Arvanitakis, and S. Manesis, “Piecewise affine modeling and constrained optimal control for a pneumatic artificial muscle,” IEEE Transactions on Industrial Electronics, vol. 61, no. 2, pp. 904–916, 2014. [35] G. Ferrari-Trecate, “Hybrid identification Toolbox (HIT),” 2005. 

3cs.SY

arXiv:1512.08452v1 [math.RA] 28 Dec 2015 Typical ranks for 3-tensors, nonsingular bilinear maps and determinantal ideals Toshio Sumi, Mitsuhiro Miyazaki and Toshio Sakata Abstract Let m, n ≥ 3, (m − 1)(n − 1) + 2 ≤ p ≤ mn, and u = mn − p. The set Ru×n×m of all real tensors with size u × n × m is one to one corresponding to the set of bilinear maps Rm × Rn → Ru . We show that Rm×n×p has plural typical ranks p and p + 1 if and only if there exists a nonsingular bilinear map Rm × Rn → Ru . We show that there is a dense open subset O of Ru×n×m such that for any Y ∈ O, the ideal of maximal minors of a matrix defined by Y in a certain way is a prime ideal and the real radical of that is the irrelevant maximal ideal if that is not a real prime ideal. Further, we show that there is a dense open subset T of Rn×p×m and continuous surjective open maps ν : O → Ru×p and σ : T → Ru×p , where Ru×p is the set of u × p matrices with entries in R, such that if ν(Y ) = σ(T ), then rank T = p if and only if the ideal of maximal minors of the matrix defined by Y is a real prime ideal. 1 Introduction For positive integers m, n, and p, we consider an m × n × p tensor which is an element of the tensor product of Rm , Rn , and Rp with standard basis. This tensor can be identified with a 3-way array (aijk ) where 1 ≤ i ≤ m, 1 ≤ j ≤ n and 1 ≤ k ≤ p. We denote by Rm×n×p the set of all m × n × p tensors. This set is a topological space with Euclidean topology. Hitchcock [15] defined the rank of a tensor. An integer r is called a typical rank of Rm×n×p if the set of tensors with rank r is a semi-algebraic set of dimension mnp. In the other words, r is a typical rank of Rm×n×p if the set of tensors with rank r contains an open set of Rm×n×p . In this paper we discuss the typical ranks of 3-tensors and connect between plurality of typical ranks and existence of a nonsingular bilinear map. Let n ≤ p. A typical rank of R1×n×p is equal to an n × p matrix full rank, that is, n. If n ≥ 2, then the set of typical ranks of R2×n×p is equal to {n, n + 1} if n = p and otherwise min{p, 2n} [36]. This is also obtained from the equivalent class: almost all 2 × n × p tensors are equivalent to ((En , On×(p−n) ); (On×(p−n) , En )) which has rank min{p, 2n} if n < p (see [18] or [32]), see Section 2 for notation. Suppose that n ≥ m ≥ 3. The set of typical ranks of Rm×n×p is equal to min{p, mn} if (m − 1)n < p [35]. If p = (m − 1)n then the set of typical ranks of Rm×n×p depends on the existence of a nonsingular bilinear map Rm × Rn → Rn : It is equal to {p} if there is no nonsingular bilinear map Rm × Rn → Rn and {p, p + 1} otherwise [33]. Here, a bilinear map f : Rm × Rn → Rr is called nonsingular if f (x, y) = 0 implies x = 0 or y = 0. Suppose that (m − 1)(n − 1) + 1 ≤ p ≤ (m − 1)n. A typical rank of Rm×n×p is unknown except a few cases. First, p is a minimal typical rank, since p is a generic rank of Cm×n×p [5]. The authors [34] showed that the Hurwitz-Radon function gives a condition that Rm×n×(m−1)n has 1 plural typical ranks. We [24] also showed that Rm×n×p has plural typical ranks for some (m, n, p) by using the concept of absolutely full column rank tensors. We let m#n be the minimal integer r such that there is a nonsingular bilinear map Rm × Rn → Rr . Then m#n ≤ m + n − 1 (see Section 2). The set Rr×m×n of r × m × n tensors is one to one corresponding to the set of bilinear maps Rm × Rn → Rr . By this map the set of absolutely full column rank tensors is one to one corresponding to the set of nonsingular bilinear maps. Theorem 1.1 Let m, n ≥ 3 and (m − 1)(n − 1) + 1 ≤ p ≤ mn. (1) If there exists a nonsingular bilinear map Rm ×Rn → Rmn−p , then Rm×n×p has plural typical ranks. (2) If p ≥ (m − 1)(n − 1) + 2 and Rm×n×p has plural typical ranks, then there exists a nonsingular bilinear map Rm × Rn → Rmn−p . (1) of Theorem 1.1 is an extension of one of [24]. Furthermore, we completely determine the set trank(m, n, p) of typical ranks of Rm×n×p for p≥(m − 1)(n − 1) + 2 by the number m#n. Theorem 1.2 Let m, n ≥ 3, k ≥ 2, and p = (m − 1)(n − 1) + k. The set of typical ranks of Rm×n×p is given as follows.   {p, p + 1}, 2 ≤ k ≤ m + n − 1 − (m#n) trank(m, n, p) = {p}, max{2, (m + n) − (m#n)} ≤ k ≤ m + n − 2   {mn}, k ≥ m + n − 1. Consider the case where p = (m − 1)(n − 1) + 1, Friedland [12] showed that Rn×n×((n−1) has plural typical ranks. We extend this result. 2 +1) Theorem 1.3 Let m, n ≥ 3 and p = (m − 1)(n − 1) + 1. Rm×n×p has plural typical ranks if m − 1 and n − 1 are not bit-disjoint. This article is organized as follows. Sections 2–7 are preparation to show the above theorems. In Section 2, we set notations and discuss the number m#n. In Section 3, we study absolutely full column rank tensors. Since the set of absolutely full column rank tensors is an open set, there exists a special form of an absolutely full column rank tensor if an absolutely full column rank tensor exists. In Section 4, we state the other notions and deal with ideals of minors of matrices. Theorem 4.31 in Section 4 which corresponds with the real radical ideals is quite interesting in its own right. We show that for integers with 0 < t ≤ min{u, n} and m ≥ (u−t+1)(n−t+1)+2, there exist open subsets O1 and O2 of Ru×n×m such that the union of them is dense, I(V(It (M (x, Y )))) = It (M (x, Y )) for Y ∈ O1 and I(V(It (M (x, Y )))) = (x1 , . . . , xm ) for Y ∈PO2 , where It (M (x, Y )) is the ideal generated by all t-minors of the u × n matrix M (x, Y ) = m k=1 xk Yk given by the indeterminates x1 , . . . , xm and Y = (Y1 ; . . . ; Ym ) ∈ Ru×n×m . From this, we can give a subset of m × n × p tensors with rank p for 3 ≤ m ≤ n and (m − 1)(n − 1) + 2 ≤ p ≤ (m − 1)n. In Section 5 we discuss a property for the determinantal ideals by using monomial preorder. This property plays an important role for proving Theorem 1.1. We characterize m × n × p tensors with rank p in Section 6. In Section 7, we show that the existence of an absolutely full column rank tensor with suitable size implies that p + 1 is a typical rank of Rm×n×p . Moreover there exist a nonempty 2 open subset T1 consisting of tensors with rank p and a possibly empty open subset T2 consisting of tensors with rank greater than p, corresponding O1 and O2 respectively, such that the union of them is a dense subset of Rm×n×p (see Theorem 7.14). Finally, in Section 8, we show that p + 2 is not a typical rank of Rm×n×p and complete proofs of the above theorems. 2 Nonsingular bilinear maps We first recall some basic facts and establish terminology. Notation (1) We denote by Rn (resp. R1×n ) the set of n-dimensional column (resp. row) real vectors and by En the n × n identity matrix. Let ej be the j-th column vector of an identity matrix. P (2) For a tensor x ∈ Rn ⊗ Rp ⊗ Rm with x = ijk aijk ei ⊗ ej ⊗ ek , we identify x with T = (aijk )1≤i≤n,1≤j≤p,1≤k≤m and denote it by (A1 ; . . . ; Am ), where Ak = (aijk )1≤i≤n,1≤j≤p for k = 1, . . . , m is an n × p matrix, and call (A1 ; . . . ; Am ) a tensor. (3) We denote the set of n × p × m tensors by Rn×p×m and the set of typical ranks by trank(n, p, m). (4) For an n × p × m tensor T = (T1 ; . . . ; Tm ), an l × n matrix P and an k × p matrix Q, we denote by P T the l × p × m tensor (P T1 ; . . . ; P Tp ) and by T Q⊤ the n × k × m tensor (T1 Q⊤ ; . . . ; Tp Q⊤ ). (5) For n × p matrices A1 , . . . , Am , we denote by (A1 , . . . , Am ) the n × mp matrix obtained by aligning A1 , . . . , Am horizontally.   A1 O   A2   (6) We set Diag(A1 , A2 , . . . , At ) =   for matrices A1 , A2 , . . . , At . ..   . O At (7) For an m × n matrix M , we denote by M≤j (resp. j< M ) the m × j matrix consisting of the first j (resp. last n − j) columns of M . We denote by M ≤i (resp. i< M ) the i × n (resp. (m−i)×n) matrix consisting of the first i (resp. last m−i) rows of M . We put M <i = M ≤i−1 M<i = M≤i−1 , and M =i = i−1< (M ≤i ) which is the i-th row vector of M .   T1   (8) We set fl1 (T ) = (T1 , . . . , Tm ) and fl2 (T ) =  ...  for a tensor T = (T1 ; . . . ; Tm ). Tm Definition 2.1 A bilinear map f : Rm × Rn → Rl is called nonsingular if f (x, y) = 0 implies x = 0 or y = 0. For positive integers m and n, we set m#n := min{l | there exists a nonsingular bilinear map Rm × Rn → Rl }. 3 Let g : R1×u × R1×v → R1×(u#v) be a nonsingular bilinear map. For positive integers m and n, let f : R1×mu × R1×nv → R1×(m+n−1)(u#v) be a map defined by f ((a1 , . . . , am ), (b1 , . . . , bn )) = P (g(a1 , b1 ), g(a1 , b2 ) + g(a2 , b1 ), . . . , i+j=k g(ai , bj ), . . . , g(am , bn )). It is easily verified that f is a nonsingular bilinear map. Thus we have the following: Lemma 2.2 (mu)#(nv) ≤ (m + n − 1)(u#v). By applying this lemma to nonsingular bilinear maps obtained by multiplications of R, C, quaternions and octanions respectively, we have the following: Proposition 2.3 (cf. [30, Proposition 12.12 (3)]) For k = 1, 2, 4 and 8, it holds that km#kn ≤ k(m + n − 1). Let H (r, s, n) be the condition
on the binomial coefficients, called the Stiefel-Hopf criterion, that the binomial coefficient nk is even whenever n − s < k < r. If there exists a continuous, nonsingular, biskew map Rr × Rs → Rn then the Stiefel-Hopf criterion H (r, s, n) holds. Put r ◦ s = min{n | H (r, s, n) holds}. We have max{r, s} ≤ r ◦ s ≤ r#s ≤ r + s − 1. Putting n∗ = ⌈ n2 ⌉ for n ∈ Z, the number r ◦ s is easily obtained by the formula ( 2(r∗ ◦ s∗ ) − 1 if r, s are both odd and r∗ ◦ s∗ = r∗ + s∗ − 1, r◦s= 2(r∗ ◦ s∗ ) otherwise (cf. [30, Proposition 12.9]). P∞ For a positive integer n, we put integers αj (n) = 0, 1, j ≥ 0 such that n = j=0 αj (n)2j is the P∞ dyadic expansion of n and let α(n) := j=0 αj (n) be the number of ones in the dyadic expansion of n. Two integers m and n are bit-disjoint if {j | αj (m) = 1} and {j | αj (n) = 1} are disjoint. For k > h, let τ (k, h) be a nonnegative number defined as τ (k, h) = #{j ≥ 0 | αj (k − h) = 0, αj (k) 6= αj (h)}. Proposition 2.4 r#s = r + s − 1 if and only if r − 1 and s − 1 are bit-disjoint. Proof If r − 1 and s − 1 are bit-disjoint, then r ◦ s = r#s = r + s − 1 (cf. [30, p. 257]). Moreover, τ (k, h) = 0 if and only if h and k − h are bit-disjoint. There is a nonsingular bilinear map Rh+1 × Rk−h+τ (k,h) → Rk for k > h ≥ 0 [20] and thus (h + 1)#(k − h + τ (k, h)) ≤ k. Putting r = h + 1 and k = r + s − 2, we have r#(s − 1 + τ (r + s − 2, r − 1)) ≤ r + s − 2. In particular, if r − 1 and s − 1 are not bit-disjoint then r#s ≤ r + s − 2. Let ρ be the Hurwitz-Radon function defined as ρ(n) = 2b + 8c for nonnegative integers a, b, c such that n = (2a + 1)2b+4c and 0 ≤ b < 4. There is a nonsingular bilinear map Rn × Rρ(n) → Rn [17, 26] and is no nonsingular bilinear map Rn × Rρ(n)+1 → Rn for any n ≥ 1 [1]. Therefore, n#ρ(n) ≤ n and n#(ρ(n) + 1) > n. Corollary 2.5 n#n ≤ 2n − 2. In particular, the equality n#n = 2n − 2 holds for n = 2a + 1. 4 Proof The inequality n#n ≤ 2n − 2 is clear by Proposition 2.4 since n − 1 and n − 1 are not bit-disjoint. There is an immersion RPn → Rn+k if and only if there is a nonsingular biskew map Rn+1 × n+1 R → Rn+1+k (cf. [2, 30]). Note that ρ(2) = 2 and ρ(4) = 4. Then 2#2 = 2 and 3#3 = 4 which follows from 4#4 = 4. Suppose that a ≥ 2. Put m = 2a−1 . Since there is no immersion RP2m →R4m−2 (cf. [21]), we have 4m = (2m + 1)#(2m + 1). Many estimations for m#n are known from immersion problem for manifolds, as projective spaces. For example, the existence of a nonsingular bilinear map Rn+1 × Rn+1 → Rn+1+k implies that RPn immerses in Rn+k [13]. Proposition 2.6 (1) (n + 1)#(n + 1) ≤ 2n − α(n) + 1 [7]. (2) (2n + α(n))#(2n + α(n)) ≥ 4n − 2α(n) + 2 [8]. (3) (8n + 9)#(8n + 9) ≥ 16n + 6 and (16n + 12)#(16n + 12) ≥ 32n + 14 if α(n) = 2 [10, 31]. (4) (8n + 10)#(8n + 10) ≥ 16n + 1 and (8n + 11)#(8n + 11) ≥ 16n + 4 if α(n) = 3 [9, 10]. (5) (n + 1)#(m + 1) ≤ n + m + 1 − (α(n) + α(n − m) + min{k(n), k(m)}) if m, n are odd and n ≥ m, where k(n) is a nonnegative function depending only in the mod 8 residue class of n with k(8a + 1) = 0, k(8a + 3) = k(8a + 5) = 1 and k(8a + 7) = 4 [23]. (6) d(h + 1)#(d(k − h) + τ (k, h)) ≤ dk for k > h ≥ 0 and d = 1, 2, 4, 8 [20]. (7) (n + 1)#(n + τ (2n, n)) ≤ 2n. 3 Absolutely full column rank tensors For a tensor T of Rn ⊗ Rp ⊗ Rm , we define the rank of T , denoted by rank T , the minimal number r so that there exist ai ∈ Rn , bi ∈ Rp , and ci ∈ Rm for i = 1, . . . , r such that T = r X ai ⊗ bi ⊗ ci . i=1 The set Rn×p×m has an action of GL(m) × GL(p) × GL(n) as (A, B, C) · r X ai ⊗ bi ⊗ ci = r X Aai ⊗ Bbi ⊗ Cci . i=1 i=1 For tensors T1 , T2 ∈ Rm×n×p , T1 and T2 are said to be equivalent if T1 = (A, B, C) · T2 for some (A, B, C) ∈ GL(n) × GL(p) × GL(m). The equivalence relation preserves the rank. For a subset U and an open semi-algebraic subset S of Rm×n×p , we say that almost all tensors in S are equivalent to tensors in U if there exists a semi-algebraic subset S0 of S with dim S0 < mnp such that any tensor of S \ S0 is equivalent to a tensor of U . In particular, for a given tensor T0 , if almost all tensors in Rm×n×p are equivalent to {T0 }, then we say that any tensor is generically equivalent to T0 . 5 An integer r is called a typical rank of n × p × m-tensors if there is a nonempty open subset O of Rn×p×m such that rank X = r for X ∈ O. Over the complex number field C, it is known that there is a unique typical rank, called the generic rank, of n × p × m-tensors for any n, p and m. The set of typical ranks of n × p × m-tensors over R is denoted by trank(n, p, m) and the generic rank of n × p × m-tensors over C is denoted by grank(n, p, m). We recall the following facts. Theorem 3.1 ([12, Theorem 7.1]) The space Rm1 ×m2 ×m3 , m1 , m2 , m3 ∈ N, contains a finite number of open connected disjoint semi-algebraic sets O1 , . . . , OM satisfying the following properties. SM (1) Rm1 ×m2 ×m3 \ i=1 Oi is a closed semi-algebraic set Rm1 ×m2 ×m3 of dimension strictly less than m1 m2 m3 . (2) Each T ∈ Oi has rank ri for i = 1, . . . , M . (3) min{r1 , . . . , rM } = grank(m1 , m2 , m3 ). (4) trank(m1 , m2 , m3 ) = {r ∈ Z | min{r1 , . . . , rM } ≤ r ≤ max{r1 , . . . , rM }}. Let T = (A1 ; . . . ; Ap ) be an m × n × p tensor over R. The tensor T is called an absolutely full column rank tensor if p X yj Aj ) = n rank( j=1 for any (y1 , . . . , yp )⊤ ∈ Rp \ {0}. From the definition of the absolutely full column rank property, we see the following fact. Lemma 3.2 Let T be an m × n × p tensor over R and P ∈ GL(m, R). Then T is absolutely full column rank if and only if so is P T . Lemma 3.3 (see Corollary 4.20 or [24, Theorem 3.6]) The set of m × n × p absolutely full column rank tensors is an open subset of Rm×n×p . Let T be an m × n × p-tensor. We define fT : Rn × Rp → Rm as fT (x, y) = p X yj Aj x, j=1 where y = (y1 , . . . , yp )⊤ . Then fT is a bilinear map. This assignment T 7→ fT induces a bijection from Rm×n×p to the set of all bilinear maps Rn ×Rp → Rm . It is easily verified that fT : Rn ×Rp → Rm is nonsingular if and only if T is absolutely full column rank. Therefore Corollary 3.4 There is an m × n × p absolutely full column rank tensor if and only if there is a nonsingular bilinear map Rn × Rp → Rm , i.e., n#p ≤ m. Lemma 3.5 Let n, m, and u be positive integers with u ≤ mn. Set p = mn − u. Then the following conditions are equivalent. 6 (1) n#m ≤ u. (2) There is a u × n × m absolutely full column rank tensor. (3) There is a u × n × m absolutely full column rank tensor Y such that p< fl1 (Y ) = −Eu . Proof (1) ⇔ (2) follows from Corollary 3.4. It is clear that (3) ⇒ (2). (2) ⇒ (3): Let X = (X1 ; . . . ; Xm ) be a u × n × m absolutely full column rank tensor. By Lemma 3.3, we may assume that p< fl1 (X) is nonsingular. Set Y = −p< fl1 (X)X. Then Y satisfies the required conditions. 4 Ideals of minors In this section, we state some results on ideals of minors, which we use in the following of this paper and interesting in its own right. First we recall the definition of normality of a ring. Definition 4.1 (see [22, Section 9]) Let R be a commutative ring. We say that R is normal if RP is an integrally closed integral domain for any prime ideal P of R. Remark 4.2 (1) A Noetherian integral domain is normal if and only if it is integrally closed. (2) If R is a Noetherian normal ring, then R ≃ R/P1 × · · · × R/Pr , where P1 , . . . , Pr are associated prime ideals of R. We recall a criterion of normality in terms of Serre’s condition. Definition 4.3 ([22, page 183]) Let R be a Noetherian ring and i a nonnegative integer. (1) We say that R satisfies (Ri ) if RP is regular for any prime ideal P of R with htP ≤ i. (2) We say that R satisfies (Si ) if depthRP ≥ min{i, htP } for any prime ideal P of R. Lemma 4.4 ([22, Theorem 23.8]) Let R be a Noetherian ring. Then R is normal if and only if R satisfies (R1 )+(S2 ). The condition (R1 )+(S2 ) is restated as follows. Lemma 4.5 Let R be a Noetherian ring. Then R satisfies (R1 )+(S2 ) if and only if the following condition is satisfied: if P is a prime ideal of R with depthRP ≤ 1, then RP is regular. Proof First assume that R satisfies (R1 )+(S2 ). Let P be a prime ideal of R with depthRP ≤ 1. Since R satisfies (S2 ), we see that depthRP ≥ min{htP, 2}. Therefore, htP ≤ 1. Thus by (R1 ), we see that RP is regular. Conversely, assume that RP is regular for any prime ideal P of R with depthRP ≤ 1. First we show that R satisfies (R1 ). If P is a prime ideal with htP ≤ 1, then depthRP ≤ htP ≤ 1. 7 Thus by assumption, we see that RP is regular. Next we show that R satisfies (S2 ). Let P be an arbitrary prime ideal of R. If depthRP ≤ 1, then by assumption, RP is regular. Thus depthRP = htP = min{htP, 2}. If depthRP ≥ 2, then depthRP ≥ min{htP, 2} holds trivially. Next we state notations and definitions used in this section. Definition 4.6 We denote by u, n, m, and t positive integers with t ≤ min{u, n} and set v = (u − t+ 1)(n− t+ 1). Let M = (mij ) be a u × n matrix with entries in a commutative ring A. We denote by It (M )A, or simply It (M ), the ideal of A generated by t-minors of M . For α(1), . . . , α(t) ∈ {1, . . . , u} and β(1), . . . , β(t) ∈ {1, . . . , n}, we set [α(1), . . . , α(t) | β(1), . . . , β(t)]M := det(mα(i)β(j) ), and if u ≥ n and α(1), . . . , α(n) ∈ {1, . . . , u}, we set [α(1), . . . , α(n)] PMm := det(mα(i)j ). For a tensor T = (T1 ; . . . ; Tm ) and a = (a1 , . . . , am ) we set M (a, T ) := i=1 ai Ti and we define Γ(u × n) = {[a1 , . . . , an ] | 1 ≤ a1 < · · · < an ≤ u, ai ∈ Z}. For γ = [a1 , . . . , an ] ∈ Γ(u × n), we set suppγ = {a1 , . . . , an }. If B is a ring, A is a subring of B and T is a tensor (resp. matrix, vector) with entries in B, we denote by A[T ] the subring of B generated by the entries of T over A. If moreover, B is a field, we denote by A(T ) the subfield of B generated by the entries of T over A. If the entries of a tensor (resp. matrix, vector) T are independent indeterminates, we say that T is a tensor (resp. matrix, vector) of indeterminates. Here we note the following fact, which is verified by using [3, Chapter 1 Exercise 2] or [25, (6.13)]. Lemma 4.7 Let A be a commutative ring, X a square matrix of indeterminates. Then det X is a non-zerodivisor of A[X]. Next we recall the following fact. Lemma 4.8 ([16, Theorem 1 and Corollaries 3 and 4] see also [4, (6.3) Theorem]) Let A be a Noetherian ring and X a u × n matrix of indeterminates. (1) ht(It (X)A[X]) = grade(It (X)A[X]) = v. (2) If A is a domain, then It (X)A[X] is a prime ideal of A[X]. (3) If A is a normal domain, then so is A[X]/It (X)A[X]. We also recall the following fact. Lemma 4.9 ([16, Theorem 1 and Corollaries 2 and 4] see also [4, (2.1) Theorem]) Let A be a Noetherian commutative ring and M a u × n matrix with entries in A. If It (M ) 6= A, then htIt (M ) ≤ v. Moreover, if A is Cohen-Macaulay and htIt (M ) = v, then It (M ) is height unmixed. The following Lemma is a generalization of [4, (12.4) Lemma]. Lemma 4.10 Let u, n, m, t and v be as in Definition 4.6, A a commutative Noetherian ring, T = (tijk ) a u × n × m tensor of indeterminates and f1 , . . . , fm elements of A. Suppose that (f1 , . . . , fm ) 6= A. Set g = grade(f1 , . . . , fm )A, f = (f1 , . . . , fm ) and M = M (f , T ) = (mij ). (1) grade It (M )A[T ] = min{g, v}. 8 (2) If g ≥ v + 1 and A is a domain, then It (M )A[T ] is a prime ideal. (3) If g ≥ v + 2 and A is a Cohen-Macaulay normal domain, then A[T ]/It (M )A[T ] is a normal domain. Remark 4.11 If g ≥ v, then grade It (M ) = ht It (M ) = v by Lemma 4.10 (1) and [16, Theorem 1 and Corollary 4]. Proof of Lemma 4.10 Set R = A[T ]. First we prove (1). Set v ′ = min{g, v}. Since It (M ) ⊂ (f1 , . . . , fm )R, we see that gradeIt (M )R ≤ grade(f1 , . . . , fm )R = g. Thus we see by Lemma 4.9, gradeIt (M )R ≤ v ′ . To prove the converse inequality, it is enough to show that if P be a prime ideal of R with P ⊃ It (M ), then depthRP ≥ v ′ . Since if P ⊃ (f1 , . . . , fm )R, then depthRP ≥ g ≥ v ′ , we may assume that P 6⊃ (f1 , . . . , fm )R. Take l with fl 6∈ P . Then M is essentially a matrix of indeterminates over A[fl−1 ][tijk | k 6= l]. Thus grade(It (M )R[fl−1 ]) = v by Lemma 4.8. Since RP is a localization of R[fl−1 ], we see that depthRP ≥ v ≥ v ′ . Next we prove (2). We may assume f1 , . . . , fm 6= 0. Set B = R/It (M )R. Since It (M )R is grade unmixed by (1) and [16, Theorem 1 Corollaries 2 and 4] (see also [27, Corollary of Theorem 1.2] or [22, Exercise 16.3]), we see that every associated prime ideal of It (M )R is of grade v. In particular any associated prime ideal of It (M )R does not contain (f1 , . . . , fm )R, since g > v by assumption. Thus (f¯1 , . . . , f¯m )B has grade at least 1, where f¯k denote the natural image of fk in B for 1 ≤ k ≤ m. Since A[fl−1 ][tijk | k 6= l] is an integral domain and M is essentially a matrix of indeterminates −1 over A[fl−1 ][tijk | k 6= l], we see that B[f¯l ] = R[fl−1 ]/It (M )R[fl−1 ] is an integral domain for any l. Thus we see that f¯l is contained in all associated prime ideals of B but one. We denote this prime ideal by Pl . Since B[(f¯l f¯l′ )−1 ] = R[(fl fl′ )−1 ]/It (M )R[(fl fl′ )−1 ] is not a zero ring by the same reason as above, we see that Pl = Pl′ for any l and l′ . In particular, Pl = P1 for any l with 1 ≤ l ≤ m. Since grade(f¯1 , . . . , f¯m )B ≥ 1 and any associated prime of B other than P1 contains (f¯1 , . . . , f¯m )B, we see that P1 is the only associated prime ideal of B. −1 Therefore, B ⊂ B[f¯1 ] and we see that B is a domain. Finally we prove (3). Assume that P is a prime ideal of B with depthBP ≤ 1. Since B is Cohen-Macaulay by (1) and [16, Theorem 1 and Corollary 4] and ht(f¯1 , . . . , f¯m )B ≥ 2, we see that P 6⊃ (f¯1 , . . . , f¯m )B. Take l with f¯l 6∈ P . Then B[f¯l−1 ] is a normal domain by Lemma 4.8 and the same argument as above. Since BP is a localization of B[f¯l−1 ], we see that BP is regular by Lemmas 4.4 and 4.5. Thus B is normal by Lemmas 4.4 and 4.5. Here we note the following fact, which can be verified by considering the associated prime ideals of I and using [22, Theorems 15.5, 15.6]. Lemma 4.12 Let K be a field, x = (x1 , . . . , xm ) a vector of indeterminates and I a proper ideal of K[x]. Then dim K[x]/I = max{r | ∃i1 , . . . , ir ; x̄i1 , . . . , x̄ir are algebraically independent over K}, where x̄i denote the natural image of xi in K[x]/I. 9 Lemma 4.13 Let K be a field, T = (tijk ) a u × n × m-tensor of indeterminates, and x = (x1 , . . . , xm ) a vector of indeterminates. Set R = K[T ], L = K(T ), M = M (x, T ) and v ′ = min{m, v}. Then L[x1 , . . . , xm−v′ ] ∩ It (M )L[x] = (0), R[x1 , . . . , xm−v′ ] ∩ It (M )R[x] = (0), L[x]/It (M )L[x] is algebraic over the natural image of L[x1 , . . . , xm−v′ ] in L[x]/It (M )L[x] and R[x]/It (M )R[x] is algebraic over the natural image of R[x1 , . . . , xm−v′ ] in R[x]/It (M )R[x]. Proof Since It (M )L[x] is generated by homogeneous polynomials of positive degree with respect to x1 , . . . , xm , we see that L ∩ It (M )L[x] = (0). By Lemma 4.10, we see that It (M ) is an ideal of height v ′ . Thus by Lemma 4.12, we see tr.degL L[x]/It (M )L[x] = m − v ′ . Thus there is a permutation i1 , . . . , in of 1, . . . , n such that x̄i1 , . . . , x̄im−v′ are algebraically independent over L and L[x]/It (M )L[x] is algebraic over L[x̄i1 , . . . , x̄im−v′ ], where x̄i denote the natural image of xi in L[x]/It (M )L[x]. By symmetry, we see that x̄1 , . . . , x̄m−v′ are algebraically independent over L and L[x]/It (M )L[x] is algebraic over L[x̄1 , . . . , x̄m−v′ ]. We also see that R[x]/It (M )R[x] is algebraic over R[x̄1 , . . . , x̄m−v′ ]. Since x̄1 , . . . , x̄m−v′ are algebraically independent over L, we see that L[x1 , . . . , xm−v′ ] ∩ It (M )L[x] = (0) and therefore R[x1 , . . . , xm−v′ ] ∩ It (M )R[x] = (0). Lemma 4.14 Let L/K be a field extension with charK = 0, x = (x1 , . . . , xm ) a vector of indeterminates and Y a u × n × m tensor with entries in L. Set M = M (x, Y ). Suppose that the entries of Y are algebraically independent over K. Then the followings hold. (1) If m ≥ v + 1, then L[x1 , . . . , xm−v ] ∩ It (M )L[x] = (0) and L[x]/It (M )L[x] is algebraic over the natural image of L[x1 , . . . , xm−v ]. (2) If m ≥ v + 2, then L[x]/It (M )L[x] is a normal domain. In particular, It (M )L[x] is a prime ideal of L[x] of height v. Proof Since the entries of Y are algebraically independent over K, we see by Lemma 4.13 that K(Y )[x]/It (M )K(Y )[x] is algebraic over K(Y )[x1 , . . . , xm−v ]. Thus L[x]/It (M )L[x] is algebraic over L[x1 , . . . , xm−v ] since L[x]/It (M )L[x] = (K(Y )[x]/It (M )K(Y )[x]) ⊗K(Y ) L. On the other hand, since tr.degL L[x]/It (M )L[x] = dim L[x]/It (M )L[x] ≥ m − v, by Lemmas 4.9 and 4.12, we see that x̄1 , . . . , x̄m−v are algebraically independent over L, where x̄i denote the natural image of xi in L[x]/It (M )L[x]. Thus L[x1 , . . . , xm−v ] ∩ It (M )L[x] = (0). This proves (1). Next we prove (2). Take a transcendence basis S of L/K(Y ) and put A = C = K(Y )[x]/It (M )K(Y )[x], K(Y )(S)[x]/It (M )K(Y )(S)[x] and B L[x]/It (M )L[x]. = 10 By Lemma 4.10 (3), we see that A is a normal domain. Since K(Y )[S][x]/It (M )K(Y )[S][x] = (K(Y )[x]/It (M )K(Y )[x]) ⊗K(Y ) K(Y )[S] = A[S] is a polynomial ring (with possibly infinitely many variables) over A, it is an integrally closed integral domain. Since C is a localization of the above ring, C is a normal domain. Now let P be a prime ideal of B with depthBP ≤ 1. By Lemmas 4.4 and 4.5, it is enough to show that BP is regular. Set Q = C ∩ P . Then since B = C ⊗K(Y )(S) L is flat over C, we see that depthCQ ≤ 1 by [22, Theorem 23.3 Corollary]. Thus CQ is regular since C is normal. The fiber ring CQ /QCQ ⊗C BP is a localization of CQ /QCQ ⊗C B = CQ /QCQ ⊗K(Y )(S) L which is a 0-dimensional reduced ring, thus regular, since L is separably algebraic over K(Y )(S). (Note that L is an inductive limit of finitely generated algebraic extension fields of K(Y )(S). Or see [25, Theorem 3.2.6 and Theorem 3.2.8 (i)] and note the assumption of the existence of a field containing M and N is not used in the proof of [25, Theorem 3.2.8 (i)].) Thus by [22, Theorem 23.7], we see that BP is regular. Thus, B is a normal ring. Since B is a nonnegatively graded ring whose degree 0 component is a field, B is not a direct product of 2 or more rings. Therefore, B is a domain by Remark 4.2. Moreover, we see by (1), that htIt (M )L[x] = v. Definition 4.15 Let u, n, m, t and v be as in Definition 4.6. We set Atu×n×m := {Y ∈ Ru×n×m | It (M (a, Y )) 6= (0) for any a ∈ R1×m \ {0}}, Ctu×n×m := Ru×n×m \ Atu×n×m , I := {Y ∈ Ru×n×m | the entries of Y are algebraically independent over Q}, and for Y = (Y1 ; . . . ; Ym ) ∈ Ru×n×m and for integers k, k ′ with t ≤ k ≤ u and t ≤ k ′ ≤ n, we set µk,k′ (x, Y ) := [1, . . . , t − 1, k | 1, . . . , t − 1, k ′ ]M(x,Y ) , where x is a vector of indeterminates. We also define ∂(µtt , µt,t+1 , . . . , µtn , µt+1,t , . . . , µt+1,n , . . . , µu,t , . . . , µun ) (x, Y ), Jt (x, Y ) = ∂(xm−v+1 , . . . , xm )  <t 1×m St (Y ) := a ∈ R | det M (a, Y )<t 6= 0, Jt (a, Y ) 6= 0 and It (M (a, Y )) = (0) Pt := {Y ∈ Ru×n×m | St (Y ) 6= ∅}. Remark 4.16 , u×n×m (1) A1u×n×m ⊃ A2u×n×m ⊃ · · · ⊃ Amin{u,n} . (2) Atu×n×m is stabel under the action of GL(u, R) for any t. (3) Ctu×n×m 6= ∅ for any t. (4) Pt is a subset of Ctu×n×m and   det M (a, Y )<t   <t 6= 0, Jt (a, Y ) 6= 0 St (Y ) := a ∈ R1×m and there exist linearly independent vectors bt , . . . , bn ∈ Rn   such that M (a, Y )bj = 0 for any t ≤ j ≤ n 11 Lemma 4.17 Let u, n and t be as in Definition 4.6, A an integral domain and M a u × n matrix <t with entries in A. Suppose that det M<t 6= 0 and [1, . . . , t − 1, k | 1, . . . , t − 1, k ′ ]M = 0 for any ′ integer with t ≤ k ≤ u and t ≤ k ≤ n. Then It (M ) = (0). In particular,   det M (a, Y )<t   <t 6= 0, Jt (a, Y ) 6= 0 St (Y ) := a ∈ R1×m and [1, . . . , t − 1, k | 1, . . . , t − 1, k ′ ]M = 0 for any integer with   t ≤ k ≤ u and t ≤ k ′ ≤ n. Proof Set  (−1)t+1 [1, . . . , t − 1 | 2, . . . , t − 1, l]M  (−1)t+2 [1, . . . , t − 1 | 1, 3, . . . , t − 1, l]M      ..   .   ξl :=  (−1)2t−1 [1, . . . , t − 1 | 1, . . . , t − 2, l]M  ∈ An     0   (−1)2t [1, . . . , t − 1 | 1, . . . , t − 2, t − 1]M  0  for each l with t ≤ l ≤ n ((−1)2t [1, . . . , t − 1 | 1, . . . , t − 2, t − 1]M in the l-th position). Then, since <t [1, . . . , t − 1 | 1 . . . , t − 1]M = det M<t 6= 0, we see that ξt , . . . , ξn are linearly independent over A. Since the k-th entry of M ξl is [1, . . . , t − 1, k | 1, . . . , t − 1, l]M , we see, by assumption, that M ξl = 0 for t ≤ l ≤ n. Thus, rank M < t and we see that It (M ) = (0). It is verified the following fact, since Q is a countable field. Lemma 4.18 I is a dense subset of Ru×n×m . We also see that Atu×n×m is an open subset of Ru×n×m . First note the following fact, which is easily verified. Lemma 4.19 Let X and Y be topological spaces with X compact and f : X×Y → R is a continuous map. Set g : Y → R by g(y) := minx∈X f (x, y). Then, g is a continuous map. Corollary 4.20 Atu×n×m is an open subset of Ru×n×m . Proof Since Atu×n×m is the set consisting of Y ∈ Ru×n×m such that min (the maximum of the absolute values of t-minors of M (a, Y )) > 0, a∈S m−1 we see the result by the previous lemma. Lemma 4.21 If v < m, then Pt is a dense subset of Ctu×n×m . In particular, Pt 6= ∅. Proof Let Y ∈ Ctu×n×m and U an open neighborhood of Y in Ru×n×m . In order to prove the first assertion, it suffices to show that Pt ∩ U 6= ∅. There exist a nonzero vector a ∈ R1×m and linearly independent vectors bt , . . . , bn ∈ Rn such that M (a, Y )bj = 0 for t ≤ j ≤ n. Let g3 ∈ GL(m) and g2 ∈ GL(n) such that the first entry of g3⊤ a is nonzero, t≤ (g2⊤ bt , . . . , g2⊤ bn ) is nonsingular and sufficiently close to Em and En respectively so that (1, g2−1 , g3−1 ) · Y ∈ U . By replacing Y , a and bt , . . . , bn by (1, g2−1 , g3−1 ) · Y , g3⊤ a and g2⊤ bt , 12 . . . , g2⊤ bn respectively, we may assume that the first entry of a is nonzero and t≤ (bt , . . . , bn ) is nonsingular. Let e ∈ R. We take a tensor P (e) = (pijk ) ∈ Ru×n×m as follows. For any i, j, k with j < t or k 6= 1, we put  ij e (k = 1, i < t, j < t, )    e ((i, j, k) = (t + l, t + l′ , m − v + 1 + l + l′ (u − t + 1)), pijk =  0 ≤ l ≤ u − t, 0 ≤ l′ ≤ n − t)    0 (otherwise) and take pij1 for 1 ≤ i ≤ u and t ≤ j ≤ n so that M (a, P (e))bj = 0 for t ≤ j ≤ n. Note that we can take such pij1 since the first entry of a is nonzero and t≤ (bt , . . . , bn ) is nonsingular. Then we have det M (a, Y + P (e))<t <t 6= 0 and Jt (a, Y + P (e)) 6= 0 for e ≫ 0. Therefore, since the entries of P (e) are polynomials of e, we see that for a real number e0 6= 0 which is sufficiently closed to 0, det M (a, Y + P (e0 ))<n <n 6= 0, (4.1) Jt (a, Y + P (e0 )) 6= 0, Y + P (e0 ) ∈ U . (4.2) (4.3) (4.1), (4.2) and the fact M (a, Y + P (e0 ))bj = M (a, Y )bj + M (a, P (e0))bj = 0 for t ≤ j ≤ n imply that a ∈ S(Y + P (e0 )). Thus we have Y + P (e0 ) ∈ Pt and we see that Pt ∩ U 6= ∅. The latter assertion follows from Remark 4.16. Lemma 4.22 Suppose that v < m. Then the set Pt is an open subset of Ru×n×m and for any Y ∈ Pt and a = (a1 , a2 ) ∈ St (Y ), where a1 ∈ R1×(m−v) and a2 ∈ R1×v , there exists an open neighborhood O(a, Y ) of a1 ∈ R1×(m−v) such that for any b1 ∈ O(a, Y ), there exists b2 ∈ R1×v such that (b1 , b2 ) ∈ St (Y ). Proof Assume that Y ∈ Pt and a = (a1 , a2 ) ∈ St (Y ). Then µkk′ (a, Y ) = 0 for any t ≤ k ≤ u and t ≤ k ′ ≤ n. Thus by implicit function theorem, we see that there is an open neighborhood U of (a1 , Y ) in R1×(m−v) × Ru×n×m and a continuous map ν : U → R1×v such that ν(a1 , Y ) = a2 , and µkk′ (b, Z) = 0 for any (b1 , Z) ∈ U and any k, k ′ with t ≤ k ≤ u and t ≤ k ′ ≤ n, where b := (b1 , ν(b1 , Z)). By replacing U by a smaller neighborhood if necessary, we may assume that det M (b, Z)<t <t 6= 0 and Jt (b, Z) 6= 0 for any (b1 , Z) ∈ U . Assume (b1 , Z) ∈ U . Put b = (b1 , ν(b1 , Z)). By Lemma 4.17, we see that b ∈ St (Z). Thus it suffices to set O(a, Y ) := {b1 ∈ R1×(m−v) | (b1 , Y ) ∈ U }. Moreover, since {Z ∈ Ru×n×m | (a1 , Z) ∈ U } is an open subset of Pt containing Y , we see that Pt is an open subset of Ru×n×m . By Corollary 4.20 and Lemmas 4.21 and 4.22, we see the following: Corollary 4.23 If v < m, then Pt ⊂ int Ctu×n×m and P t = int Ctu×n×m = Ctu×n×m . 13 Definition 4.24 We set P̃t := {P Y | P ∈ GL(u, R), Y ∈ Pt }. Lemma 4.25 P̃t is an open subset of Ru×n×m , stable under the action of GL(u, R) and P̃t = Ctu×n×m . S Proof Since P̃t = P ∈GL(u,R) P Pt and P Pt is an open subset of Ru×n×m for any P ∈ GL(u, R) by Lemma 4.22 and the fact that multiplication of a nonsingular matrix is a homeomorphism on Ru×n×m . Therefore, P̃t is an open subset of Ru×n×m . The fact that P̃t is stable under the action of GL(u, R) is clear from the definition of P̃t . Finally, since Atu×n×m is stable under the action of GL(u, R), we see, by Remark 4.16, that P̃t ⊂ Ctu×n×m . Therefore, we see that P̃t = Ctu×n×m by Lemma 4.21. Lemma 4.26 Let L be an infinite field and x = (x1 , . . . , xm ) a vector of indeterminates. Set v ′ = min{m, v} and v ′′ = min{m, (u − t + 2)(n − t + 2)}. Then there is a Zariski dense open subset Q1 of Lu×n×m such that if Y ∈ Q1 , then L[x]/It (M (x, Y ))L[x] is algebraic over the natural image of L[x1 , . . . , xm−v′ ], L[x1 , . . . , xm−v′ ] ∩ It (M (x, Y ))L[x] = (0), htIt (M )L[x] = v ′ , L[x]/It−1 (M (x, Y ))L[x] is algebraic over the natural image of L[x1 , . . . , xm−v′′ ], L[x1 , . . . , xm−v′′ ] ∩ It−1 (M (x, Y ))L[x] = (0) and htIt−1 (M (x, Y ))L[x] = v ′′ . Proof Let T = (tijk ) be the u × n × m tensor of indeterminates. Then by Lemma 4.13, we see that L[T ][x]/It (M (x, T ))L[T ][x] is algebraic over the natural image of L[T ][x1 , . . . , xm−v′ ]. Denote the natural image of xl in L[T ][x]/It (M (x, T ))L[T ][x] by x̄l . Take a nonzero polynomial fl (t) with coefficient in L[T ][x1 , . . . , xm−v′ ] such that fl (x̄l ) = 0 for each l with m − v ′ + 1 ≤ l ≤ m. Let g be the product of all nonzero elements of L[T ] appearing as the coefficient of at least one of fl and set Q1′ = Lu×n×m \ V(g). Then Q1′ is a Zariski dense open subset of Lu×n×m . Suppose that Y ∈ Q1′ . And let x̃i be the natural image of xi in L[x]/It (M (x, Y ))L[x] and f˜l be an element of L[x1 , . . . , xm−v′ ] obtained by substituting Y to T . Then f˜l is a nonzero element of L[x1 , . . . , xm−v′ ] and f˜l (x̃l ) = 0 for m − v ′ + 1 ≤ l ≤ m. Therefore, L[x]/It (M (x, Y ))L[x] is algebraic over the natural image of L[x1 , . . . , xm−v′ ]. Thus htIt (M (x, Y ))L[x] = m − dim L[x]/It (M )L[x] = m − tr.degL L[x]/It (M )L[x] ≥ v ′ . Thus, htIt (M (x, Y ))L[x] = v ′ by Lemma 4.9 and we see that tr.degL L[x]/It (M (x, Y ))L[x] = m − v ′ . Therefore x̃1 , . . . , x̃m−v′ are algebraically independent over L, that is, L[x1 , . . . , xm−v′ ] ∩ It (M (x, Y ))L[x] = (0). We see by the same way that there is a Zariski dense open subset Q1′′ of Lu×n×m such that if Y ∈ Q1′′ , then L[x]/It−1 (M (x, Y ))L[x] is algebraic over the natural image of L[x1 , . . . , xm−v′′ ], L[x1 , . . . , xm−v′′ ] ∩ It−1 (M (x, Y ))L[x] = (0) and htIt−1 (M (x, Y ))L[x] = v ′′ . Thus it is enough to set Q1 = Q1′ ∩ Q1′′ . Let L be a field, x = (x1 , . . . , xm ) a vector of indeterminates and M a u × n matrix with entries p <t −1 <t ) ] is in L[x]. Suppose that htIt (M ) = v and det(M<t ) 6∈ It (M ). Then It (M )L[x][det(M<t <t −1 a proper ideal of L[x][det(M<t ) ] and M is equivalent to the matrix of the following form in <t −1 L[x][(det(M<t ) ].   Et−1 O O ∗ <t −1 ) ]. By symmetry, we see In particular, It (M ) is a complete intersection ideal in L[x][(det(M<t that if htIt−1 (M ) > v, then It (M ) is a generically complete intersection ideal. 14 We use the notation of [11, p. 219]. Let L be a field of characteristic 0, T a u × n × m tensor of indeterminates and x = (x1 , . . . , xm ) a vector of indeterminates. Set M = L(T ). Suppose that m > v. Then It (M (x, T ))M[x] is a prime ideal and htIt−1 (M (x, T ))M[x] > v by Lemma 4.10. Thus It (M (x, T )) : Jm−v (It (M (x, T ))) = It (M (x, T )) by [11, Theorem 2.1] and the argument above. Thus if we set I ′ = It (M (x, T )) + ′ Jm−v (It (M (x, T ))), then htI > v. Therefore the natural images of x1 , . . . , xm−v in M[x]/I ′ are algebraically dependent over M by Lemma 4.12. Take a transcendence basis xi1 , . . . xid of M[x]/I ′ over M. By symmetry, we may assume that ik = k for 1 ≤ k ≤ d. Since htI ′ > v, we see that d < m − v. Take a nonzero polynomial f (t) with coefficients in L[T ][x1 , . . . , xd ] such that f (xm−v ) ∈ I ′ and let g be the product of all nonzero elements of L[T ] which appear in some nonzero coefficient of f . Set Q2 = Lu×n×m \V(g). Then Q2 is a Zariski dense open subset of Lu×n×m and if Y ∈ Q1 ∩Q2 , where Q1 is the one in Lemma 4.26, then ht(It (M (x, Y ))+Jm−v (It (M (x, Y )))) > v since tr.degL L[x]/(It (M (x, Y )) + Jm−v (It (M (x, Y )))) < m − v. Until the end of this section, assume that m ≥ v + 2 and let U be the m × m matrix of indeterminates, T the u × n × m tensor of indeterminates, x = (x1 , . . . , xm ) the vector of indeterminates and L the algebraic closure of R(U ). Set    ′ x1 x1  ..   ..   .  = U  . . x′m xm Then L(T )[x′1 , . . . , x′m ] = L(T )[x] and L(T )[x′1 , . . . , x′m−v+1 ] ∩ It (M (x, T )) is a principal ideal generated by a polynomial F called the ground form of It (M (x, T )), since It (M (x, T )) is a prime ideal therefore is unmixed of height v. See [29, parts 28 and 29]. Since L(T )[x′1 , . . . , x′m−v+1 ] ∩ It (M (x, T )) is the elimination ideal, F is obtained by the Buchberger’s algorithm. Let g3 be the products of all elements of L[T ] which appear as a numerator or a denominator of a nonzero coefficient of at least one polynomial in the process of Buchberger’s algorithm to obtain the reduced Gröbner basis of It (M (x, T )) in L[T ][x]. Set Q3 = Lu×n×m \ V(g3 ). Then Q3 is a Zariski dense open subset of Lu×n×m and if Y ∈ Q3 , then the Buchberger’s algorithm to obtain the reduced Gröbner basis of It (M (x, Y ))L[x] in L[x] is identical with that of It (M (x, T ))L(T )[x] in L(T )[x]. In particular, L[x′1 , . . . , x′m−v+1 ] ∩ It (M (x, Y )) is a principal ideal generated by FY , the polynomial obtained by substituting Y in T in the coefficients of F . Let d = deg F and let PL (d, m − v + 1) (resp. PR (d, m − v + 1)) be the set of homogeneous polynomials with coefficients in L (resp. R) with variables x′1 , . . . , x′m−v+1 and degree d. Since m − v + 1 ≥ 3 and L is an algebraically closed field containing R, we see by [14] that {G ∈ PR (d, m − v + 1) | G is absolutely irreducible} = PR (d, m − v + 1) ∩ {G ∈ PL (d, m − v + 1) | G is irreducible} is a Zariski dense open subset of PR (d, m − v + 1). Definition 4.27 Set Q = {Y ∈ Q1 ∩ Q2 ∩ Q3 ∩ Ru×n×m | FY is absolutely irreducible}, 15 where Q1 is the one in Lemma 4.26 and Q2 , Q3 and FY are the ones defined after the proof of Lemma 4.26. Remark 4.28 Q is a Zariski open subset of Ru×n×m , since the correspondence Y to FY is a rational map whose domain contains Q3 . Moreover, we see the following fact. Lemma 4.29 Q ⊃ I ∩ Q1 ∩ Q2 ∩ Q3 . In particular, Q is not an empty set. Proof Suppose that Y ∈ I ∩ Q1 ∩ Q2 ∩ Q3 . Then we see, by applying Lemma 4.14 to L/Q, that It (M (x, Y ))L[x] is a prime ideal. Thus the elimination ideal is also prime and therefore the generator FY of the elimination ideal is an irreducible polynomial in L[x′1 , . . . , x′m−v+1 ]. Therefore, Y ∈ Q. Since I is a dense subset of Ru×n×m by Lemma 4.18, we see that I ∩ Q1 ∩ Q2 ∩ Q3 6= ∅. Thus, Q 6= ∅. Thus we see that Q is a non-empty Zariski open subset of Ru×n×m . In particular, Q is dense. Lemma 4.30 If Y ∈ Q, then It (M (x, Y ))R[x] is a prime ideal of height v. Proof Since Y ∈ Q, htIt (M (x, Y ))R[x] = v and L[x′1 , . . . , x′m−v+1 ] ∩ It (M (x, Y ))L[x] = (FY )L[x′1 , . . . , x′m−v+1 ]. Thus R(U )[x′1 , . . . , x′m−v+1 ]∩It (M (x, Y ))R(U )[x] = (FY )R(U )[x′1 , . . . , x′m−v+1 ] since L is faithfully flat over R(U ). Thus we see that FY is the ground form of It (M (x, Y ))R[x] [29, part 28]. Since FY is an irreducible polynomial in L[x′1 , . . . , x′m−v+1 ] and therefore in R(U )[x′1 , . . . , x′m−v+1 ], we see by [29, part 31], that It (M (x, Y ))R[x] is a primary ideal. On the other hand, since Y ∈ Q1 ∩Q2 , we see that ht(It (M (x, Y ))+ Jm−v (It (M (x, Y )))) > v. Since It (M (x, Y )) is a primary ideal of height v, we see that It (M (x, Y )) : Jm−v (It (M (x, Y ))) = It (M (x, Y )). Therefore, by [11, Theorem 2.1], we see that It (M (x, Y )) is a radical ideal. Thus It (M (x, Y )) is a prime ideal. Now we show the following result. Theorem 4.31 Suppose that m ≥ v + 2. Set O1 = Q ∩ P̃t and O2 = Q ∩ Atu×n×m . Then the followings hold. (1) O1 and O2 are disjoint open subsets of Ru×n×m and O1 6= ∅. (2) O1 ∪ O2 is a dense subset of Ru×n×m . (3) O 1 = Ctu×n×m = Ru×n×m \ Atu×n×m . (4) If Y ∈ O1 ∪ O2 , then It (M (x, Y ))R[x] is a prime ideal of height v. (5) If Y ∈ O1 , then I(V(It (M (x, Y )))) = It (M (x, Y )). (6) If Y ∈ O2 , then I(V(It (M (x, Y )))) = (x1 , . . . , xm ). 16 Proof The set Atu×n×m is an open subset of Ru×n×m by Corollary 4.20 and P̃t is a nonempty open subset of Ru×n×m with Atu×n×m ∩ P̃t = ∅ by Lemmas 4.21 and 4.25. Further Q is a Zariski open subset. Thus (1) holds. (2): Since Q and Atu×n×m ∪ P̃t are dense open subsets of Ru×n×m by Corollary 4.20 and Lemma 4.25, we see that O1 ∪ O2 ⊃ Q ∩ (Atu×n×m ∪ Pt ) is also a dense open subset of Ru×n×m . (3): Since Q is a dense subset of Ru×n×m , and P̃t is an open set, we see that O 1 = Q ∩ P̃t = P̃t = Ctu×n×m by Lemma 4.25. (4) follows from Lemma 4.30. (5): Assume the contrary and take g ∈ I(V(It (M (x, Y )))) with g 6∈ It (M (x, Y )). Set J = (g)R[x] + It (M (x, Y )). Then J ) It (M (x, Y ))R[x]. Since It (M (x, Y ))R[x] is a prime ideal of height v by Lemma 4.30, we see that htJ > v and therefore R[x1 , . . . , xm−v ] ∩ J 6= (0). Take 0 6= f ∈ J ∩ R[x1 , . . . , xm−v ]. Since Y ∈ P̃t , we can take P ∈ GL(u, R) such that P Y ∈ Pt . Take b ∈ St (P Y ). Since O(b, P Y ) defined in Lemma 4.22 is an open set and f is a nonzero polynomial, we can take (a1 , . . . , am−v ) ∈ O(b, P Y ) with f (a1 , . . . , am−v ) 6= 0. On the other hand, we see that there are am−v+1 , . . . , am ∈ R such that It (M (a, Y )) = It (P M (a, Y )) = It (M (a, P Y )) = (0) by Lemma 4.22, where a = (a1 , . . . , am ). Thus by assumption, we see that g(a) = 0. This contradicts to the fact that f ∈ J = (g)R[x] + It (M (x, Y ))R[x] and f (a1 , . . . , am−v ) 6= 0. Finally, (6) is clear from the definition of Atu×n×m . 5 Monomial preorder In this section, we introduce the notion of monomial preorder and prove a result about ideals of minors by using it. First we recall the notion of preorder. Definition 5.1 Let S be a nonempty set and  a binary relation on S. We say that  is a preorder on S or (S, ) is a preordered set if the following two conditions are satisfied. (1) a  a for any a ∈ S. (2) a  b, b  c ⇒ a  c. If moreover, (3) a  b or b  a for any a, b ∈ S. is satisfied, then we say that (S, ) is a totally preordered set or  is a total preorder. Notation Let (S, ) be a preordered set. We denote by b  a the fact a  b. We denote by a ≺ b or by b ≻ a the fact that a  b and b  a. We also denote by a ∼ b the fact that a  b and b  a. Remark 5.2 The binary relation ∼ defined above is an equivalence relation and if a ∼ a′ and b ∼ b′ , then a  b ⇐⇒ a′  b′ . def In particular, we can define a binary relation ≤ on the quotient set P = S/ ∼ by ā ≤ b̄ ⇐⇒ a  b, where ā is the equivalence class which a belongs to. It is easily verified that (P, ≤) is a partially 17 ordered set and (S, ) is a totally preordered set if and only if (P, ≤) is a totally ordered set. As usual, we denote ā > b̄ the fact ā ≥ b̄ and ā 6= b̄. Definition 5.3 Let x1 , . . . , xr be indeterminates. We denote the set of monomials or power products of x1 , . . . , xr by PP (x1 , . . . , xr ). A monomial preorder on x1 , . . . , xr is a total preorder  on PP (x1 , . . . , xr ) satisfying the following conditions. (1) 1  m for any m ∈ PP (x1 , . . . , xr ). (2) For m1 , m2 , m ∈ PP (x1 , . . . , xr ), m1  m2 ⇐⇒ m1 m  m2 m. Let ∼ be the equivalence relation on PP (x1 , . . . , xr ) defined by the monomial preorder . We denote by P (x1 , . . . , xr ) the quotient set PP (x1 , . . . , xr )/ ∼ and by qdeg m the class of m in P (x1 , . . . , xr ) and call it the quasi-degree of m, where m ∈ PP (x1 , . . . , xr ). Remark 5.4 Our definition of monomial preorder may seem to be different from that of [19], but it is identical except we allow m ∼ 1 for a monomial m 6= 1. Example 5.5 (c.f. [19, Example 3.1]) Let x1 , . . . , xr be indeterminates, W = (w1 , . . . , ws ) an r × s matrix whose entries are real numbers such that the first nonzero entry of each row is positive. If one defines def xa  xb ⇐⇒ (a · w1 , . . . , a · ws ) ≤lex (b · w1 , . . . , b · ws ), where · denotes the inner product and ≤lex denotes the lexicographic order, then  is a monomial preorder. In fact, one can prove by the same way as [28] that every monomial preorder is of this type. Definition 5.6 Let K be a field and x1 , . . . , xr indeterminates. If a monomial preorder on x1 , . . . , xr is defined, we say that K[x1 , . . . , xr ] is a polynomial ring with monomial preorder. Let f be a nonzero element of K[x1 , . . . , xr ]. We say that f is a form if all the monomials appearing in f have the same quasi-degree. We denote by qdeg f the quasi-degree of the monomials appearing in f . Let g be a nonzero element of K[x1 , . . . , xr ]. Then there is a unique expression g = g1 + g2 + · · · + gt of g, where gi is a form for 1 ≤ i ≤ t and qdeg g1 > qdeg g2 > · · · > qdeg gt . We define the leading form of g, denoted lf(g) as g1 . Remark 5.7 Let K[x1 , . . . , xr ] be a polynomial ring with monomial preorder and f , g nonzero elements of K[x1 , . . . , xr ]. Then lf(f g) = lf(f )lf(g). Remark 5.8 It is essential to assume both implications in (2) of Definition 5.3. For example, let x and y be indeterminates. We define total preorder on PP (x, y) by 1 ≺ y ≺ x and m1 ≺ m2 if the total degree of m1 is less than that of m2 . Then it is easily verified that (1) 1 ≺ m for any m ∈ PP (x, y) \ {1}. 18 (2) m1  m2 ⇒ m1 m  m2 m. Let f = x + y. Then lf(f ) = x while lf(f 2 ) = x2 + 2xy + y 2 6= x2 = (lf(f ))2 . Definition 5.9 Let x1 , . . . , xr be indeterminates. Suppose that a total preorder on {x1 , . . . , xr } is defined. Rewrite the set {x1 , . . . , xr } as follows. {x1 , . . . , xr } = {y11 , . . . , y1s1 , y21 , . . . , y2ss , . . . , yt1 , . . . , ytst }, s1 + · · · + st = r, y11 ∼ · · · ∼ y1s1 ≻ y21 ∼ · · · ∼ y2s2 ≻ · · · ≻ yt1 ∼ · · · ∼ ytst . Qt Qsi aij The lexicographic monomial preorder on PP (x1 , . . . , xr ) is defined as follows. i=1 j=1 yij  Qt Qsi bij if and only if one of the following conditions is satisfied. y j=1 ij i=1 Ps1 Ps1 • j=1 a1j < j=1 b1j . Ps1 Ps1 Ps2 Ps2 • j=1 a1j = j=1 b1j and j=1 a2j < j=1 b2j . Ps1 Ps1 Ps2 Ps2 Ps3 Ps3 • j=1 a1j = j=1 b1j , j=1 a2j = j=1 b2j and j=1 a3j < j=1 b3j . .. . Ps1 Ps1 Ps2 Ps2 Pst−2 Pst−2 • a1j = j=1 at−2,j = j=1 bt−2,j and j=1 b1j , j=1 a2j = j=1 b2j , . . ., Pj=1 P st−1 st−1 a < b . t−1,j t−1,j j=1 j=1 Ps1 Ps1 Ps2 Ps2 Pst−1 Pst−1 • a1j = b1j , a2j = b2j , . . ., at−1,j = j=1 j=1 bt−1,j and j=1 j=1 j=1 j=1 Pst Pst j=1 at,j ≤ j=1 bt,j . Remark 5.10 Suppose that x1 , . . . , xr are indeterminates and total preorder  on {x1 , . . . , xr } is defined. Suppose also that x1 ∼ · · · ∼ xm1 ≻ xm1 +1 ∼ · · · ∼ xm2 ≻ · · · ≻ xmt−1 +1 ∼ · · · ∼ xmt , mt = r. Then the lexicographic monomial preorder induced by this preorder on {x1 , . . . , xr } is the one defined as in Example 5.5 by the r × t matrix whose j-th column has 1 in mj−1 + 1 through mj -th position and 0 in others, where we set m0 = 0. Definition 5.11 We set k [a1 , . . . , ai , . . . , an ] := [a1 , . . . , ai−1 , k, ai+1 , . . . , an ] and k l [a1 , . . . , ai , . . . , aj , . . . , an ] := [a1 , . . . , ai−1 , k, ai+1 , . . . , aj−1 , l, aj+1 , . . . , an ]. Lemma 5.12 Let K be a field, K[x1 , . . . , xr ] a polynomial ring with monomial preorder, S a subset of {x1 , . . . , xr } and g1 , . . . , gt ∈ K[x1 , . . . , xr ]. Set L = K[S]. If lf(g1 ), . . . , lf(gt ) are linearly independent over L, then g1 , . . . , gt are linearly independent over L. Proof Assume the contrary and suppose that X ci g i = 0 i 19 is a non-trivial relation where ci ∈ L and ci = 6 0 for any i which appears in the above sum. Then X′ lf(ci )lf(gi ) = 0, P′ where runs through i’s with qdeg lf(ci gi ) are maximal. Since lf(ci ) ∈ L for any i, it contradicts the assumption. Lemma 5.13 (Plücker relations, see e.g. [4, (4,4) Lemma]) For every u × n-matrix M, u ≥ n, with entries in a commutative ring and all indices a1 , . . . , ak , bl , . . . , bn , c1 , . . . , cs ∈ {1, . . . , u} such that s = n − k + l − 1 > n, t = n − k > 0 one has X sgn(i1 , . . . , is )[a1 , . . . , ak , ci1 , . . . , cit ]M [cit+1 , . . . , cis , bl , . . . , bn ]M = 0, i1 <···<it it+1 <···<is {1,...,s}={i1 ,...,is } where sgn(σ) is the signature of a permutation σ and the notations are defined in Definition 4.6. An element a is called a non-zerodivisor if ab = 0 implies b = 0. Lemma 5.14 Let A = A0 ⊕ A1 ⊕ · · · be a graded ring,  X  = (xij ) a u × n matrix with u > n and X 1×n entries in A1 and y = (y1 , . . . , yn ) ∈ A1 . Set X̃ = and Γ = Γ(u × n). Suppose that y • δ = [1, 2, . . . , n]X is a non-zerodivisor of A, • for any k1 and k2 with 1 ≤ k1 < k2 ≤ n, δ X γX (γ ∈ Γ), n+1 [1, . . . , k2 , . . . , n]X γX (γ ∈ Γ \ {δ}) and n+1 [1, . . . , k1 , . . . , n]X γX (γ ∈ Γ, suppγ 6⊃ {1, . . . , k̂2 , . . . , n}) are linearly independent over A0 and • In (X̃) = In (X), where the notations are defined in Definition 4.6. Then, y is an A0 -linear combination of rows of X. X X Proof We denote γX̃ as γ and as for simplicity. Set γ∈Γ γ u+1 [1, . . . , k , . . . , n] = X a(k) γ γ γ and n+1 u+1 [1, . . . , k1 , . . . , k2 , . . . , n] = X γ 20 1 ,k2 ) a(k γ γ (k) (k1 ,k2 ) where aγ , aγ (k1 ,k2 ) (k) ∈ A. By considering the degree, we may assume that aγ , aγ n+1 ∈ A0 . u+1 [1, . . . , k1 , . . . , k2 , . . . , n]X̃Cof(X ≤n ) = = n+1 u+1 det(Cof(X ≤n ))[1, . . . , k1 , . . . , k2 , . . . , n] X 1 ,k2 ) δ n−1 a(k γ, γ γ where Cof(X ≤n ) denotes the matrix of cofactors of X ≤n . On the other hand, since   δ   δ     ..   .     δ    n+1  n+1 ≤n n+1 X̃Cof(X ) = [ 1 , 2 . . . , n] [1, 2 , . . . , n] · · · [1, 2, . . . , n ] ,   n+2  n+2  [ 1 , 2 . . . , n] [1, 2 , . . . , n] · · · [1, 2, . . . , n+2  n ]     .. .. .. ..   . . . .   u+1 u+1 [ 1 , 2 . . . , n] ··· [1, 2 , . . . , n] u+1 [1, 2, . . . , n ] we see that n+1 = u+1 [1, . . . , k1 , . . . , k2 , . . . , n]X̃Cof(X ≤n )   n+1 n+1 [1, . . . , k , . . . , n] [1, . . . , k , . . . , n] 1 2 . δ n−2 det  u+1 u+1 [1, . . . , k1 , . . . , n] [1, . . . , k2 , . . . , n] Since δ is a non-zerodivisor, we see that X 1 ,k2 ) a(k δγ γ γ = = n+1 u+1 γ = n+1 u+1 [1, . . . , k1 , . . . , n][1, . . . , k2 , . . . , n] − [1, . . . , k2 , . . . , n][1, . . . , k1 , . . . , n] X X n+1 n+1 aγ(k2 ) [1, . . . , k1 , . . . , n]γ − aγ(k1 ) [1, . . . , k2 , . . . , n]γ γ (k ) −aδ 1 δ[1, . . . , n+1 k2 , . . . , n] − n+1 X aγ(k1 ) [1, . . . , k2 , . . . , n]γ γ∈Γ\{δ} n+1 (k ) +aδ 2 δ[1, . . . , k1 , . . . , n] X + suppγ∩{1,...,n}={1,...,k̂2 ,...,n} + X n+1 aγ(k2 ) [1, . . . , k1 , . . . , n]γ suppγ6⊃{1,...,k̂2 ,...,n} 21 n+1 aγ(k2 ) [1, . . . , k1 , . . . , n]γ Suppose that suppγ ∩ {1, . . . , n} = {1, . . . , k̂2 , . . . , n}. Then γ = [1, . . . , k̂2 , . . . , n, l] for some l l with n + 1 ≤ l ≤ u. Thus γ = (−1)n−k2 [1, . . . , k2 , . . . , n]. By applying Lemma 5.13 to l n+1 [1, . . . , k2 , . . . , n][1, . . . , k1 , . . . , n], by substituting u, n, k, s, l by u, n, k1 − 1, n + 1, k1 + 1 respectively, we see that l n+1 [1, . . . , k2 , . . . , n][1, . . . , k1 , . . . , n] n+1 l n+1 l n+1 l = δ[1, . . . , k1 . . . , k2 , . . . , n] + [1, . . . , k2 , . . . , n][1, . . . , k1 , . . . , n] = δ[1, . . . , k1 . . . , k2 , . . . , n] + (−1)n−k1 [1, . . . , k2 , . . . , n][1, . . . , k̂1 , . . . , n, l] n+1 Therefore, if we set bγ(k1 ) = ( (k ) (k1 ) 2 − (−1)k2 −k1 a[1,..., k̂ aγ if γ = [1, . . . , k̂1 . . . , n, l] 2 ,...,n,l] (k ) aγ 1 otherwise for γ ∈ Γ \ {δ}, we see that there are bγ ∈ A0 such that X X bγ δγ − γ n+1 bγ(k1 ) [1, . . . , k2 , . . . , n]γ γ∈Γ\{δ} n+1 X + aγ(k2 ) [1, . . . , k1 , . . . , n]γ = 0. suppγ6⊃{1,...,k̂2 ,...,n} Thus we see, by the assumption, that bγ(k1 ) = 0 if γ ∈ Γ \ {δ} and aγ(k2 ) = 0 (k1 ) Therefore, by the definition of bγ if suppγ 6⊃ {1, . . . , k̂2 , . . . , n}. , we see that (k ) (k ) 1 (−1)n−k1 a[1,..., k̂ 1 ,...,n,l] 2 = (−1)n−k2 a[1,..., k̂ 2 ,...,n,l] and aγ(k1 ) = 0 if suppγ 6⊃ {1, . . . , k̂1 , . . . , n}. Since these hold for any k1 , k2 with 1 ≤ k1 < k2 ≤ n, we see that if we set (n) cl = a[1,...,n−1,l] for l with n + 1 ≤ l ≤ u, u+1 (k) [1, . . . , k , . . . , n] = aδ δ + (−1)n−k u X cl [1, . . . , k̂, . . . , n, l] l=n+1 (k) = aδ δ + u X l=n+1 22 l cl [1, . . . , k, . . . , n] for any k with 1 ≤ k ≤ n. Set z=y− n X (s) aδ X =s − s=1 and u X cl X =l l=n+1   X Z= . z Then u+1 [1, . . . , k , . . . , n]Z u+1 = [1, . . . , k , . . . , n]X̃ − n X s (s) aδ [1, . . . k, . . . n]X − s=1 = 0 u X l cl [1, . . . k, . . . n]X l=n+1 for any k with 1 ≤ k ≤ n. Thus zCof(X ≤n ) = 0 and we see that z = 0 since det Cof(X ≤n ) = δ n−1 is a non-zerodivisor. Lemma 5.15 Let L be a field, n, m integers with n ≥ m ≥ 3, x1 , . . . , xm indeterminates and α, β ∈ L with 0 6= α 6= β 6= 0. Then the following polynomials are linearly independent over L. x2n 1 x2n−1 (x2 − βxm ) 1 x2n−1 xs (3 ≤ s ≤ m − 1) 1 xn1 (x2 − αxm )(x2 − βxm )xb1 · · · xbn−2 (1 ≤ b1 ≤ · · · ≤ bn−2 ≤ 2) xn1 (x2 − αxm )xb1 · · · xbn−2 xs (1 ≤ b1 ≤ · · · ≤ bn−2 ≤ 2, 3 ≤ s ≤ m − 1) xn1 xb1 · · · xbn (1 ≤ b1 ≤ · · · ≤ bn ≤ m − 1, bn−1 ≥ 3) x2n−2 (x2 − βxm )2 1 x2n−2 (x2 − βxm )xs (3 ≤ s ≤ m − 1) 1 n−1 x1 (x2 − αxm )(x2 − βxm )2 xb1 · · · xbn−2 x1n−1 (x2 x1n−1 (x2 x1n−2 (x2 x1n−2 (x2 x1n−2 (x2 − αxm )(x2 − βxm )xb1 · · · xbn−2 xs − βxm )xb1 · · · xbn (1 ≤ b1 ≤ · · · ≤ bn−2 ≤ 2) (1 ≤ b1 ≤ · · · ≤ bn−2 ≤ 2, 3 ≤ s ≤ m − 1) (1 ≤ b1 ≤ · · · ≤ bn ≤ m − 1, bn−1 ≥ 3) 2 − αxm ) (x2 − βxm )2 xb1 · · · xbn−2 2 − αxm ) (x2 − βxm )xb1 · · · xbn−2 xs − αxm )(x2 − βxm )xb1 · · · xbn (1 ≤ b1 ≤ · · · ≤ bn−2 ≤ 2) (1 ≤ b1 ≤ · · · ≤ bn−2 ≤ 2, 3 ≤ s ≤ m − 1) (1 ≤ b1 ≤ · · · ≤ bn ≤ m − 1, bn−1 ≥ 3) Proof Set deg x1 = deg x2 = deg xm = 0 and deg x3 = · · · = deg xm−1 = 1. Then the polynomials under consideration are homogeneous. Thus it is enough to show that for each integer d, the polynomials of degree d in the above list are linearly independent over L. First consider the polynomials with degree more than 1. They are xn1 xb1 · · · xbn x1n−1 (x2 x1n−2 (x2 (1 ≤ b1 ≤ · · · ≤ bn ≤ m − 1, bn−1 ≥ 3) − βxm )xb1 · · · xbn (1 ≤ b1 ≤ · · · ≤ bn ≤ m − 1, bn−1 ≥ 3) − αxm )(x2 − βxm )xb1 · · · xbn (1 ≤ b1 ≤ · · · ≤ bn ≤ m − 1, bn−1 ≥ 3). 23 By first substituting β −1 x2 to xm and next by substituting α−1 x2 to xm one sees that these polynomials are linearly independent. Next consider the polynomials with degree 1. They are x2n−1 xs (3 ≤ s ≤ m − 1) 1 xn1 (x2 − αxm )xb1 · · · xbn−2 xs (x2 − βxm )xs x2n−2 1 n−1 x1 (x2 − αxm )(x2 (1 ≤ b1 ≤ · · · ≤ bn−2 ≤ 2, 3 ≤ s ≤ m − 1) (3 ≤ s ≤ m − 1) − βxm )xb1 · · · xbn−2 xs (1 ≤ b1 ≤ · · · ≤ bn−2 ≤ 2, 3 ≤ s ≤ m − 1) x1n−2 (x2 − αxm )2 (x2 − βxm )xb1 · · · xbn−2 xs (1 ≤ b1 ≤ · · · ≤ bn−2 ≤ 2, 3 ≤ s ≤ m − 1). By a similar but more subtle argument as above, one sees that these polynomials are linearly independent. Finally consider the polynomials with degree 0. They are x2n 1 x2n−1 (x2 − βxm ) 1 n x1 (x2 − αxm )(x2 − βxm )xb1 · · · xbn−2 (1 ≤ b1 ≤ · · · ≤ bn−2 ≤ 2) x2n−2 (x2 − βxm )2 1 n−1 x1 (x2 − αxm )(x2 − βxm )2 xb1 · · · xbn−2 x1n−2 (x2 − αxm )2 (x2 − βxm )2 xb1 · · · xbn−2 (1 ≤ b1 ≤ · · · ≤ bn−2 ≤ 2) (1 ≤ b1 ≤ · · · ≤ bn−2 ≤ 2) By a similar but more subtle argument, one sees that these polynomials are linearly independent. Lemma 5.16 Let K be a field, T = (T1 ; · · · ; Tm ) = (tijk ) a u × n × m-tensor of indeterminates with u > n ≥ m ≥ u − n + 2, x = (x1 , . . . , xm ) a vector of indeterminates. Set X = M (x, T ) (c.f. Definition 4.6), Γ = Γ(u × n), δ = δ0 = [1, . . . , n] and A = K[T ][x]. Then δX is a non-zerodivisor of A and for any k1 , k2 with 1 ≤ k1 , k2 ≤ n, k1 6= k2 , δ X γX (γ ∈ Γ) n+1 [1, . . . , k2 , . . . , n]X γX (γ ∈ Γ \ {δ}) n+1 [1, . . . , k1 , . . . , n]X γX (γ ∈ Γ, suppγ 6⊃ {1, . . . , k̂2 , . . . , n}) are linearly independent over K[T ]. Proof The first assertion is clear since A is an integral domain and δX 6= 0. Next we prove the second assertion. By symmetry, we may assume that k1 = n − 1 and k2 = n. Set δ1 = [1, 2, . . . , n − 1, n + 1] and δ2 = [1, 2, . . . , n − 2, n, n + 1]. We introduce the lexicographic monomial preorder induced by the preorder on the indeterminates defined as follows. If one of the following is satisfied, we define tijk ≻ ti′ j ′ k′ . 24 • i < i′ . • i = i′ and j < j ′ . • i = i′ , j = j ′ , k < k ′ and “i < j or i > j + u − n”. In case 0 ≤ i − j ≤ u − n and (i, j) 6= (n, n − 1), (n + 1, n), we define ti,j,i−j+1 ≻ ti,j,i−j+2 ≻ · · · ≻ ti,j,m ≻ ti,j,1 ≻ · · · ≻ ti,j,i−j . In case (i, j) = (n, n − 1) or (n + 1, n), we define ti,j,2 ∼ ti,j,m ≻ ti,j,1 ≻ ti,j,3 ≻ · · · ≻ ti,j,m−1 . And tijk ≻ x1 ∼ x2 ∼ · · · ∼ xm for any i, j and k. Set Γ0 = {[a1 , . . . , an ] ∈ Γ | an−1 = n − 1}, Γ1 = {[a1 , . . . , an ] ∈ Γ | an−1 = n}, Γ2 = {[a1 , . . . , an ] ∈ Γ | an−1 ≥ n + 1}, Γ00 = {[a1 , . . . , an ] ∈ Γ0 | an = n} = {δ0 }, Γ01 = {[a1 , . . . , an ] ∈ Γ0 | an = n + 1} = {δ1 }, Γ02 = {[a1 , . . . , an ] ∈ Γ0 | an ≥ n + 2}, Γ11 = {[a1 , . . . , an ] ∈ Γ1 | an = n + 1} = {δ2 }, Γ12 = {[a1 , . . . , an ] ∈ Γ1 | an ≥ n + 2}. Then Γ = Γ0 ⊔ Γ1 ⊔ Γ2 Γ0 = Γ00 ⊔ Γ01 ⊔ Γ02 Γ1 = Γ11 ⊔ Γ12 . Set α1 = tn,n−1,2 , α2 = tn,n−1,m , β1 = tn+1,n,2 and β2 = tn+1,n,m . Then for γ = [a1 , . . . , an ] ∈ Γ, lf(γX ) is, up to multiplication of nonzero element of K[T ], as follows. xn1 if γ ∈ Γ00 x1n−1 (β1 x2 + β2 xm ) if γ ∈ Γ01 x1n−1 xan −n+1 if γ ∈ Γ02 xa1 xa2 −1 · · · xan−2 −n+3 (α1 x2 + α2 xm )(β1 x2 +β2 xm ) if γ ∈ Γ11 xa1 xa2 −1 · · · xan−2 −n+3 (α1 x2 + α2 xm )xan −n+1 if γ ∈ Γ12 xa1 xa2 −1 · · · xan −n+1 if γ ∈ Γ2 . 25 Therefore, for γ = [a1 , . . . , an ] ∈ Γ, lf(δX γX ) is, up to multiplication of nonzero element of K[T ], x2n 1 if γ ∈ Γ00 x2n−1 (β1 x2 + β2 xm ) if γ ∈ Γ01 1 2n−1 x1 xan −n+1 if γ ∈ Γ02 xn1 xa1 xa2 −1 · · · xan−2 −n+3 (α1 x2 + α2 xm )(β1 x2 +β2 xm ) if γ ∈ Γ11 xn1 xa1 xa2 −1 · · · xan−2 −n+3 (α1 x2 + α2 xm )xan −n+1 if γ ∈ Γ12 xn1 xa1 xa2 −1 · · · xan −n+1 if γ ∈ Γ2 . For γ = [a1 , . . . , an ] ∈ Γ \ {δ}, lf((δ1 )X γX ) is, up to multiplication of nonzero element of K[T ], x2n−2 (β1 x2 + β2 xm )2 1 if γ ∈ Γ01 2n−2 x1 (β1 x2 + β2 xm )xan −n+1 if γ ∈ Γ02 n−1 x1 xa1 xa2 −1 · · · xan−2 −n+3 (α1 x2 + α2 xm )(β1 x2 +β2 xm )2 if γ ∈ x1n−1 xa1 xa2 −1 · · · xan−2 −n+3 (α1 x2 + α2 xm )(β1 x2 +β2 xm )xan −n+1 x1n−1 (β1 x2 +β2 xm )xa1 xa2 −1 · · · xan −n+1 if γ ∈ Γ2 . Γ11 if γ ∈ Γ12 Finally, consider the leading form of (δ2 )X γX , where γ = [a1 , . . . , an ] ∈ Γ and suppγ 6⊃ {1, . . . , n − 1}. It is easily verified that suppγ 6⊃ {1, . . . , n − 1} if and only if γ ∈ Γ1 ⊔ Γ2 . Thus the leading form of (δ2 )X γX is, up to multiplication of nonzero element of K[T ], x1n−2 xa1 xa2 −1 · · · xan−2 −n+3 (α1 x2 + α2 xm )2 (β1 x2 +β2 xm )2 if γ ∈ Γ11 x1n−2 xa1 xa2 −1 · · · xan−2 −n+3 (α1 x2 + α2 xm )2 (β1 x2 +β2 xm )xan −n+1 x1n−2 (α1 x2 + α2 xm )(β1 x2 +β2 xm )xa1 xa2 −1 · · · xan −n+1 if γ ∈ Γ2 . if γ ∈ Γ12 Since 1 ≤ a1 ≤ a2 − 1 ≤ · · · ≤ an − n + 1 ≤ u − n + 1 ≤ m − 1 an−2 − n + 3 ≤ 2 if γ ∈ Γ1 an−1 − n + 2 ≥ 3 if γ ∈ Γ2 and an − n + 1 ≥ 3 if γ ∈ Γ02 ⊔ Γ12 , we see by Lemma 5.15 that lf(δX γX ) (γ ∈ Γ) lf((δ1 )X γX ) (γ ∈ Γ \ {δ}) lf((δ2 )X γX ) (γ ∈ Γ, suppγ 6⊃ {1, . . . , n − 1}) are linearly independent over K[T ]. The assertion follows by Lemma 5.12. Corollary 5.17 Let K be a field, T a u × n × m tensor of indeterminates with u > n ≥ m ≥ u − n + 2, R a commutative ring containing K(T ), x = (x1 , . . . , xm ) a vector of indeterminates. 26 Set M = M (x, T ) (c.f. Definition 4.6) and B = R[x]. Then δM is a non-zerodivisor of B and for any k1 , k2 with 1 ≤ k1 , k2 ≤ n, k1 6= k2 , δ M γM (γ ∈ Γ) n+1 [1, . . . , k2 , . . . , n]M γM (γ ∈ Γ \ {δ}) n+1 [1, . . . , k1 , . . . , n]M γM (γ ∈ Γ, suppγ 6⊃ {1, . . . , k̂2 , . . . , n}) are linearly independent over R. Proof Set A = K[T ][x]. Then B = A ⊗K[T ] R. Since R is flat over K[T ], we see that B is flat over A. By Lemma 5.16, we see that δM is a non-zerodivisor of A and for any k1 , k2 with 1 ≤ k1 , k2 ≤ n, k1 6= k2 , δ M γM (γ ∈ Γ) n+1 [1, . . . , k2 , . . . , n]M γM (γ ∈ Γ \ {δ}) n+1 [1, . . . , k1 , . . . , n]M γM (γ ∈ Γ, suppγ 6⊃ {1, . . . , k̂2 , . . . , n}) are linearly independent over K[T ]. Since R (resp. B) is flat over K[T ] (resp. A), we see that δM is a non-zerodivisor of B and for any k1 , k2 with 1 ≤ k1 , k2 ≤ n, k1 6= k2 , δ M γM (γ ∈ Γ) n+1 [1, . . . , k2 , . . . , n]M γM (γ ∈ Γ \ {δ}) n+1 [1, . . . , k1 , . . . , n]M γM (γ ∈ Γ, suppγ 6⊃ {1, . . . , k̂2 , . . . , n}) are linearly independent over R. Corollary 5.18 Let x = (x1 , . . . , xm ) be a vector of indeterminates. Suppose that u > n and Y ∈ I (c.f. Definition 4.15) and set M = M (x, Y ). Then δM is a non-zerodivisor of R[x] and for any k1 , k2 with 1 ≤ k1 , k2 ≤ n, k1 6= k2 , δ M γM (γ ∈ Γ) n+1 [1, . . . , k2 , . . . , n]M γM (γ ∈ Γ \ {δ}) n+1 [1, . . . , k1 , . . . , n]M γM (γ ∈ Γ, suppγ 6⊃ {1, . . . , k̂2 , . . . , n}) are linearly independent over R. Definition 5.19 Let x = (x1 , . . . , xm ) be a vector of indeterminates. {Y ∈ Ru×n×m | δM 6= 0 and δM γM (γ ∈ Γ), n+1 [1, . . . , k1 , . . . , n]M γM We set Q ′ := n+1 [1, . . . , k2 , . . . , n]M γM (γ ∈ Γ \ {δ}), (γ ∈ Γ, suppγ 6⊃ {1, . . . , k̂2 , . . . , n}) are linearly independent over R for any k1 , k2 with 1 ≤ k1 < k2 ≤ n, where M = M (x, Y )}. 27 Remark 5.20 Q ′ is a Zariski open set of Ru×n×m and by Corollary 5.18, we see that Q ′ ⊃ I . In particular, Q ′ is a Zariski dense open subset of Ru×n×m . By Lemma 5.14, we see the following fact. Proposition 5.21 Let u, n and m be integers with u > n ≥ m ≥ u−n+2 and let x = (x1 , . . . , xm ) be a vector of indeterminates. Suppose that Y ∈ Q ′ and y ∈ R1×n×m . Set   Y Ỹ = . y If In (M (x, Ỹ )) = In (M (x, Y )), then fl1 (y) is an R-linear combination of rows of fl1 (Y ). Proof Set M = M (x, Y ). Since R[x] is a domain and δM 6= 0 by the definition of Q ′ , we see that δM is a non-zerodivisor of R[x]. Moreover, δ M γM (γ ∈ Γ) n+1 [1, . . . , k2 , . . . , n]M γM (γ ∈ Γ \ {δ}) n+1 [1, . . . , k1 , . . . , n]M γM (γ ∈ Γ, suppγ 6⊃ {1, . . . , k̂2 , . . . , n}) are linearly independent over R for any k1 , k2 with 1 ≤ k1 < k2 ≤ n by the definition of Q ′ . Thus by Lemma 5.14, we see that M (x, y) is an R-linear combination of rows of M = M (x, Y ). Since x1 , . . . , xm are indeterminates, we see that fl1 (y) is an R-linear combination of rows of fl1 (Y ). 6 Tensor with rank p Let 3 ≤ m ≤ n, (m − 1)(n − 1) + 1 ≤ p ≤ (m − 1)n and set l = (m − 1)n − p and u = n + l. In the following of this paper, we use the results of the previous sections by setting t = n. See Definition 4.6. Then v = l + 1 and it follows that v < m since l ≤ m − 2. Note also u + p = nm. We make bunch of definitions used in the sequel of this paper. Definition 6.1 We put V = V n×p×m := {T ∈ Rn×p×m | fl2 (T )≤p is nonsingular} and define σ : V → Ru×p be a map defined as σ(T ) = (p< fl2 (T ))(fl2 (T )≤p )−1 . We denote by A u×n×m the set of all u × n × m absolutely full column rank tensors and put C = Ru×n×m \ A u×n×m . Note that A u×n×m = Anu×n×m and C u×n×m = Cnu×n×m in the notation of Definition 4.15. Let M be the subset of Ru×nm consisting of all matrices W = (W1 , . . . , Wm ) satisfying that there are A = (a1 , . . . , ap ) ∈ Rn×p and p × p diagonal matrices D1 , . . . , Dm such that Dk = u×n×m 28 Diag(d1k , . . . , dpk ) for 1 ≤ k ≤ m, (dj1 W1 + · · · + djm Wm )aj = 0 for 1 ≤ j ≤ p, and   AD1   ..   .    ADm−2  A≤n−l Dm−1 (6.1) is nonsingular. Let ι : Ru×p → Ru×nm be a map which sends A to (A, −Eu ). Moreover put C := {W ∈ Ru×nm | W ∈ / fl1 (A u×n×m )} = {W ∈ Ru×n×m | W ∈ fl1 (C u×n×m )}. We define φ : R1×m × Rn → Rp a map defined as   a1 b   a2 b   φ(a, b) =   ∈ Rp , where a = (a1 , . . . , am−1 , am ). ..   . am−1 b≤n−l Recall that the set A u×n×m is an open subset of Ru×n×m by Lemma 3.3 or Corollary 4.20. Proposition 6.2 σ is an open, surjective and continuous map. Proof Clearly σ is surjective and continuous. Let O be an open subset of V and let h : Rnm×p → Rp×p ×Ru×p be a homeomorphism defined as h(M ) = (M ≤p , p< M ). Then h(fl2 (O)) can be written by [ O1,λ × O2,λ λ for some open subsets O1,λ ⊂ R p×p and O2,λ ⊂ Ru×p and thus [ [ σ(O) = O2,λ A−1 . λ A∈O1,λ The set O2,λ A−1 is open and then σ(O) is open. The following fact follows from the definition. Lemma 6.3 M is stable under the action of GL(u, R). Lemma 6.4 M ⊂ C. By observing the first  column of (6.1),  we see that a1 6= 0 and at least one of d11 , d12 , AD1   ..   . . . . , d1,m−1 is nonzero, since   is nonsingular, where Dk = Diag(d1k , . . . , dpk ) for  ADm−2  A≤n−l Dm−1 1 ≤ k ≤ m − 1. Since d11 W1 + · · · + d1m Wm is singular, (W1 ; . . . ; Wm ) is not absolutely full column rank, i.e., M ⊂ C. Proof 29 Theorem 6.5 Let X ∈ V ⊂ Rn×p×m . Put (W1 , . . . , Wm−1 , Wm ) = ι(σ(X)), where W1 , . . . , Wm ∈ Ru×n . The following four statements are equivalent. (1) rank X = p. (2) There are an n × p matrix A, and diagonal p × p matrices D1 , . . . , Dm such that W1 AD1 + W2 AD2 + · · · + Wm−1 ADm−1 + Wm ADm = O and    N =  is nonsingular.  AD1 .. . ADm−2 A≤n−l Dm−1     (6.2) (6.3) (3) ι(σ(X)) ∈ M. Proof It holds that rank X ≥ p, since fl2 (X)≤p has rank p. Put (S1 ; . . . ; Sm ) = X(fl2 (X)≤p )−1 Then rank X = rank(S1 ; . . . ; Sm ) and n−l<  Sm−1 = (W1 , W2 , . . . , Wm−2 , (Wm−1 )≤n−l ). Sm (1) ⇒ (2): Since rank(S1 ; . . . ; Sm ) = p, there are an n × p-matrix A, a p × p-matrix Q, and p × p diagonal matrices D1 , . . . , Dm such that ADk Q = Sk Since  AD1 .. . for k = 1, . . . , m.   S1 .. .          NQ =   = Ep , Q =   Sm−2   ADm−2  (Sm−1 )≤n−l A≤n−l Dm−1 we see that N and Q are nonsingular and Q−1 = N . Since  n−l< Sm−1 = σ(X) = (W1 , W2 , . . . , Wm−2 , (Wm−1 )≤n−l ), Sm and ADk = Sk N for k = m − 1, m, we see that = = = Ou×p n−l< n−l<   Sm−1 ADm−1 N− Sm ADm W1 AD1 + · · · + Wm−2 ADm−2     −El n−l< O ≤n−l +(Wm−1 )≤n−l A Dm−1 + ADm ADm−1 + O −En W1 AD1 + · · · + Wm−2 ADm−2 + Wm−1 ADm−1 + Wm ADm . 30 Therefore the equation (6.2) holds. (2) ⇒ (1): Set Q = N −1 . Then, since N Q = Ep , we see that ADk Q = Sk and 1≤k ≤m−2 ≤n−l A≤n−l Dm−1 Q = Sm−1 . Furthermore, since W1 AD1 + · · · + Wm−2 ADm−2 +(Wm−1 )≤n−l A≤n−l Dm−1 + = = we see that Thus W1 AD1 + · · · + Wm ADm Ou×p ,     −El n−l< O ADm ADm−1 + O −En n−l<      Sm−1 El n−l< O N= ADm . ADm−1 + Sm O En   n−l< Sm−1 ADm−1 Q= Sm ADm n−l< and we see that ADk Q = Sk for k = m − 1, m. Therefore, rank X = rank(S1 ; . . . ; Sm ) ≤ p and we see (1). Finally it is easy to see that (2) ⇔ (3). Since M ⊂ C by Lemma 6.4, we see the following: Proposition 6.6 For X ∈ V , if rank X = p, then ι(σ(X)) 6∈ fl1 (A u×n×m ). 7 Contribution of absolutely full column rank property Let m, n and p be integers with 3 ≤ m ≤ n and (m − 1)(n − 1) + 1 ≤ p ≤ (m − 1)n. We set u = nm − p and t = n and use the results of Sections 4 and 5. Note v = u − n + 1 = (m − 1)n − p + 1 in the notation of Definition 4.6. It is known that the generic rank grank(n, p, m) of n × p × m tensors over C is equal to p ([5, Theorem 3.1] or [6, Theorem 2.4 and Remark 2.5]) and it is also equal to the minimal typical rank of n × p × m tensors over R. Thus if we discuss the plurality of typical ranks, it is enough to consider whether there exists a typical rank that is greater than p or not. Definition 7.1 We set A := ι−1 (fl1 (A u×n×m )) ⊂ Ru×p , where ι and A u×n×m are defined in Definition 6.1. Lemma 7.2 If Y ∈ V n×p×m and σ(Y ) ∈ A, then rank Y > p. Proof This follows from the fact that rank Y ≥ p if Y ∈ V n×p×m and Proposition 6.6. 31 Theorem 7.3 If A u×n×m 6= ∅, then there are plural typical ranks of n × p × m tensors over R. Proof By Lemma 3.5, we see that A 6= ∅. Since A u×n×m is an open subset of Ru×n×m , we see that A is an open subset of Ru×p . Moreover, since σ : V n×p×m → Ru×p is a surjective continuous map, we see that σ −1 (A) is a nonempty open subset of V n×p×m . Thus, there is a typical rank greater than p by Lemma 7.2. Since p is a typical rank of n × p × m tensors over R, we see that there are plural typical ranks of n × p × m tensors over R. From now on until the end of this section, we assume that p ≥ (m − 1)(n − 1) + 2. Thus, m ≥ v + 2. Definition 7.4 Let Y ∈ Ru×n×m and let x = (x1 , . . . , xm ) be a vector of indeterminates. For i1 , . . . , in−1 ∈ {1, . . . , u}, we set   (−1)n+1 [i1 , . . . , in−1 | 2, . . . , n − 1, n]M(x,Y )  (−1)n+2 [i1 , . . . , in−1 | 1, 3, . . . , n − 1, n]M(x,Y )    ψi1 ,...,in−1 (x, Y ) :=   ∈ R[x]n . ..   . (−1)2n [i1 , . . . , in−1 | 1, . . . , n − 2, n − 1]M(x,Y ) We define ψ̂i1 ,...,in−1 : R1×m × For the definition [a1 , . . . , at | b1 , . . . , bt ], see Definition 4.6. Ru×n×m → R[x]p by  x1 ψi1 ,...,in−1 (x, Y ) x2 ψi1 ,...,in−1 (x, Y ) .. . ≤p     ψ̂i1 ,...,in−1 (x, Y ) :=     xm−1 ψi1 ,...,in−1 (x, Y ) ∈ R[x]p . We also define the R-vector space U (Y ) by U (Y ) := hψ̂i1 ,...,in−1 (u, Y ) | u ∈ V(In (M (x, Y ))), i ⊂ Rp . i1 , . . . , in−1 ∈ {1, . . . , u} For c = (c11 , . . . , cn1 , c12 , . . . , cn2 , . . . , c1m , . . . , cnm ) ∈ R 1×nm , we set Zk = 1 ≤ k ≤ m, Z = (Z1 ; . . . ; Zm ) ∈ R(u+1)×n×m and  c1k Yk · · · cnk  for gi1 ,...,in−1 (x, Y, c) = [i1 , . . . , in−1 , u + 1]M(x,Z) for any i1 , . . . , in−1 ∈ {1, . . . , u}. For the definition [i1 , . . . , in−1 , u + 1]M(x,Z) , see Definition 4.6. Lemma 7.5 Suppose that Y ∈ Ru×n×m . Then the following claims are equivalent. (1) dim U (Y ) = p. (2) If c ∈ R1×nm satisfies the following conditions, then c = 0. (∗) p< c = 0 and 32 (∗∗) gi1 ,...in−1 (u, Y, c) = 0 for any u ∈ V(In (M (x, Y ))) and any i1 , . . . , in−1 ∈ {1, . . . , u}. Proof The vector d ∈ Rp is perpendicular to U (Y ) if and only if d is perpendicular to ψ̂i1 ,...,in−1 (u, Y ) for any u ∈ V(In (M (x, Y ))) and any i1 , . . . , in−1 . Since the inner product of ψ̂i1 ,...,in (u, Y ) with d is gi1 ,...,in−1 (u, Y, (d⊤ , 0)), the result follows. Next we show the following result. For the definition of M, see Definition 6.1. Lemma 7.6 If dim U (Y ) = p, then fl1 (Y ) ∈ M. Proof Set Y = (Y1 ; . . . ; Ym ). Suppose that dim U (Y ) = p. Then there are u1 , . . . , up ∈ V(In (M (x, Y ))) and t11 , . . . , t1,n−1 , . . . , tp1 , . . . , tp,n−1 such that ψ̂t11 ,...,t1,n−1 (u1 , Y ), . . . , ψ̂tp1 ,...,tp,n−1 (up , Y ) are linearly independent over R. Set uj = (uj1 , . . . , ujm ) for 1 ≤ j ≤ p, Dk = Diag(u1k , . . . , upk ) for 1 ≤ k ≤ m and A = (ψt11 ,...,t1,n−1 (u1 , Y ), . . . , ψtp1 ,...,tp,n−1 (up , Y )). Then, ≤p AD1     ..     .  =  = (ψ̂t11 ,...,t1,n−1 (u1 , Y ), . . . , ψ̂tp1 ,...,tp,n−1 (up , Y ))  ADm−2  ADm−2  A≤n−l Dm−1 ADm−1  AD1 .. .   is a nonsingular matrix and  [1, tj1 , . . . , tj,n−1 ]M(uj ,Y )  [2, tj1 , . . . , tj,n−1 ]M(uj ,Y )    (uj1 Y1 + · · · + ujm Ym )ψtj1 ,...,tj,n−1 (uj , Y ) =  =0 ..   .  [u, tj1 , . . . , tj,n−1 ]M(uj ,Y ) since In (M (uj , Y )) = (0) for 1 ≤ j ≤ p. Definition 7.7 Set U := {Y ∈ Ru×n×m | p< fl1 (Y ) is nonsingular}, O3 := U ∩ Q ∩ Q ′ ∩ P̃n = O1 ∩ U ∩ Q ′ and O4 := U ∩ Q ∩ Q ′ ∩ A u×n×m = O2 ∩ U ∩ Q ′ , where Q, Q ′ and P̃n are the ones defined in Definitions 4.27, 5.19, and 4.24 and O1 and O2 are the ones in Theorem 4.31 under t = n. Define ν : U → Ru×p as ν(Y ) := −(p< fl1 (Y ))−1 fl1 (Y )≤p for i = 1, 2, where σ is the one defined in Definition 6.1. Set Oi = ν(Oi+2 ) ⊂ Ru×p and Ti = σ −1 (Oi )⊂ V n×p×m for i = 1, 2. The following fact is immediately verified. Lemma 7.8 ι(ν(Y )) = fl1 (−(p< fl1 (Y ))−1 Y ). By the same way as Theorem 4.31 (1), (2) and (3), we see the following fact. Lemma 7.9 Then the followings hold. 33 (1) O3 and O4 are disjoint open subsets of Ru×n×m and O3 is nonempty. (2) O3 ∪ O4 is a dense subset of Ru×n×m . (3) O3 = C u×n×m = Ru×n×m \ A u×n×m . Lemma 7.10 fl1 (O3 ) ⊂ M. Proof Let Y ∈ O3 . By Lemmas 7.6 and 7.5, it is enough to show that if c ∈ R1×nm satisfies (∗) and (∗∗), then c = 0.   Y 1×n ′ 1×n×m Set c = (c1 , . . . , cm ), where cj ∈ R , c = (c1 ; · · · ; cm ) ∈ R and Ỹ = . Then by c′ (∗∗), gi1 ,...,in−1 (x, Y, c) ∈ I(V(In (M (x, Y )))) for any i1 , . . . , in−1 ∈ {1, . . . , u}. Therefore, by the definition of O3 and Theorem 4.31 (5), we see that gi1 ,...,in−1 (x, Y, c) ∈ In (M (x, Y )) for any i1 , . . . , in−1 ∈ {1, . . . , u}. Thus we see that In (M (x, Ỹ )) = In (M (x, Y )). Thus, by Proposition 5.21, we see that fl1 (c) is an R-linear combination of rows of fl1 (Y ). Since p< c = 0 and Y ∈ U , we see that c = 0. By the same way as Proposition 6.2, we see the following: Proposition 7.11 ν is an open, surjective and continuous map. We see the following fact. Lemma 7.12 Then the followings hold. (1) Y ∈ A u×n×m if and only if ν(Y ) ∈ A for Y ∈ U . (2) O1 and O2 are disjoint open subsets of Ru×p and O1 6= ∅. (3) O1 ∪ O2 is a dense subset of Ru×p . (4) O1 = Ru×p \ A. (5) O2 ⊂ A and O2 = A. Proof (1): Suppose that Y ∈ U . Since A u×n×m and U are stable under the action of GL(u, R), we see that Y ∈ A u×n×m if and only if −(p< fl1 (Y ))−1 Y ∈ A u×n×m . Since ι(ν(Y )) = fl1 (−(p< fl1 (Y ))−1 Y ) and fl1 is a bijection, we see (1). We see (2) by the facts that P̃n and A u×n×m are stable under the action of GL(u, R), Lemma 7.9, and Proposition 7.11. (3) also follows from Lemma 7.9 and Proposition 7.11. We see by (1) that if Y ∈ O3 , then ν(Y ) 6∈ A. Thus O1 = ν(O3 ) ⊂ Ru×p \ A. Since O3 = Ru×n×m \ A u×n×m by Lemmas 7.9 (3), we see that O1 ⊃ ν(O3 ∩ U ) = ν((Ru×n×m \ A u×n×m ) ∩ U ) = Ru×p \ A by (1) and the surjectivity of ν. Thus we see (4). Therefore O2 ⊂ A by (2). Further, we see that O 2 ⊃ A by (3) and (4). Thus we see (5). Lemma 7.13 Let X and Y be topological spaces, f : X → Y a mapping and B a subset of Y . (1) If f is continuous, then f −1 (B) ⊃ f −1 (B). 34 (2) If f is an open map, then f −1 (B) ⊂ f −1 (B). Proof (1): Since f −1 (B) is a closed subset of X containing f −1 (B), we see that f −1 (B) ⊃ f −1 (B). (2): Suppose that x ∈ f −1 (B) and let U be an open neighborhood of x. We show that U ∩ f −1 (B) 6= ∅. Since f (x) ∈ B, f (x) ∈ f (U ) and f (U ) is an open subset of Y , we see that f (U ) ∩ B 6= ∅. Take b ∈ f (U ) ∩ B and a ∈ U such that f (a) = b. Then, since f (a) ∈ B, we see that a ∈ f −1 (B). Thus, a ∈ U ∩ f −1 (B) and we see that U ∩ f −1 (B) 6= ∅. Theorem 7.14 Let m, n and p be integers with 3 ≤ m ≤ n and (m − 1)(n − 1) + 2 ≤ p ≤ (m − 1)n. The followings hold. (1) T1 and T2 are disjoint open subsets of V n×p×m and T1 is nonempty. (2) T1 ∪ T2 is a dense subset of Rn×p×m . (3) T1 ∩ V n×p×m = V n×p×m \ σ −1 (A) and T2 ∩ V n×p×m = σ −1 (A) ∩ V n×p×m . (4) If T ∈ T1 , then rank T = p. (5) If T ∈ T2 , then rank T > p. Proof First note that σ −1 (X ) ∩ V n×p×m = σ −1 (X ) for any subset X of Ru×p by Lemma 7.13, since σ is an open continuous map. (1) and (2) follow from Lemma 7.12 and the facts that σ is surjective and V n×p×m is a dense subset of Rn×p×m . (3): We see by Lemma 7.12 that T 1 ∩V n×p×m = σ −1 (O1 ) = σ −1 (Ru×p \A) = V n×p×m \σ −1 (A) and T 2 ∩ V n×p×m = σ −1 (O2 ) = σ −1 (A) ∩ V n×p×m . (4): Suppose that T ∈ T1 . Then σ(T ) ∈ O1 . Thus there exists Y ∈ O3 such that ν(Y ) = σ(T ). By Lemma 7.10, we see that fl1 (Y ) ∈ M. Hence ι(σ(T )) = ι(ν(Y )) = −(p< fl1 (Y ))−1 fl1 (Y ) ∈ M, since M is stable under the action of GL(u, R). Therefore rank T = p by Theorem 6.5. (5): If T ∈ T2 , then σ(T ) ∈ O2 ⊂ A by Lemma 7.12. Thus rank T > p by Lemma 7.2. 8 Upper bound for typical ranks In Lemma 7.2, we see a class of tensors with rank greater than p. To complete the proof of Theorem 1.2, we give an upper bound of the set of typical ranks of Rn×p×m : Theorem 8.1 Let 3 ≤ m ≤ n and (m − 1)(n − 1) + 1 ≤ p ≤ (m − 1)n. Any typical rank of Rn×p×m is less than or equal to p + 1. We prepare the proof. Let 3 ≤ m ≤ n, (m − 1)(n − 1) + 1 ≤ p < (m − 1)n and u = mn − p. Let σ ′ : V n×(p+1)×m → R(u−1)×(p+1) be the counterpart of σ : V n×p×m → Ru×p . Also, let A′ ⊂ R(u−1)×(p+1) and T1′ ⊂ V n×(p+1)×m be the counterparts of A ⊂ Ru×p and T1 ⊂ V n×p×m respectively. Let π : Rn×(p+1)×m → Rn×p×m be a canonical projection defined as π(Y1 ; . . . ; Ym ) = ((Y1 )≤p ; . . . ; (Ym )≤p ). Clearly π is a continuous, surjective and open map. 35 Lemma 8.2 π(T1′ ) is an open dense subset of Rn×p×m . Proof Since T1′ is an open set and π is an open map, π(T1′ ) is an open subset of Rn×p×m . We show that π(T1′ ) is dense. Let X ∈ V n×p×m . Consider the map f : V n×p×m → V n×(p+1)×m defined as f (X1 ; . . . ; Xm−2 ; Xm−1 ; Xm ) = ((X1 , 0); . . . ; (Xm−2 , 0); (Xm−1 , e); (Xm , 0)) where e is the (2n − u + 1)th column vector of the identity matrix En . Since the (p + 1)th column vector of the matrix σ ′ (f (X)) is zero, f (X) ∈ / σ ′−1 (A′ ) holds and by Theorem 7.14 (3), ′ f (X) ∈ T1 . Since π ◦ f is the identity map and π is continuous, X ∈ π(T1′ ) ⊂ π(T1′ ) holds. Therefore V n×p×m ⊂ π(T1′ ) and thus Rn×p×m = π(T1′ ). By Theorem 7.14 (5), and Lemma 8.2, we have immediately the following corollary. Corollary 8.3 Let 3 ≤ m ≤ n and (m − 1)(n − 1) + 1 ≤ p < (m − 1)n. T2 6= ∅ if and only if T2 ∩ π(T1′ ) 6= ∅, and rank T = p + 1 for any T ∈ T2 ∩ π(T1′ ). Note that arbitrary tensor of π(T1′ ) has rank less than or equal to p + 1 by Theorem 7.14 (4). Proof of Theorem 8.1 The assertion for p = (m − 1)n holds by [33]. Suppose that (m − 1)(n − 1) + 1 ≤ p < (m − 1)n. Then rank(T ) ≤ p + 1 for T ∈ π(T1′ ). Since π(T1′ ) is dense, arbitrary integer greater than p + 1 is not a typical rank. Recall that trank(m, n, p) = trank(n, p, m). We are ready to prove main theorems. Proof of Theorem 1.1 (1) follows from Theorem 7.3 and Corollary 3.4. (2): We may assume that 3 ≤ m ≤ n without the loss of generality. Ten Berge [35] showed that Rm×n×p has a unique typical rank for p ≥ (m − 1)n + 1. Therefore, we see that p ≤ (m − 1)n. Set u = mn − p. By Theorem 7.14 (2) and (4), we see that T2 6= ∅. Furthermore, T2 6= ∅ ⇒ O4 6= ∅ by definitions and the surjectivity of σ. Since O4 ⊂ A u×n×m , we see that there exists an absolutely full column rank u × n × m tensor. The result follows from Corollary 3.4. Proof of Theorem 1.2 We may assume that 3 ≤ m ≤ n. Note that trank(m, n, p) = {min{p, mn}} for k ≥ m [35]. Suppose that 2 ≤ k ≤ m−1. By Theorem 8.1, the maximal typical rank of Rm×n×p is less than or equal to p + 1. Since p is the minimal typical rank of Rm×n×p , trank(m, n, p) is {p} or {p, p+ 1}. By Theorem 1.1, Rm×n×p has a unique typical rank if and only if m#n≥ mn − p + 1, equivalently, k≥ m + n − (m#n). This completes the proof. We immediately have Theorem 1.3 by Proposition 2.4 and Theorem 7.3. In the case where p = (m − 1)(n − 1) + 1, we have many examples for having plural typical ranks. Corollary 8.4 Let m, n ≥ 3 and a ≥ 1. If m ≡ 2a−1 + s (mod 2a ) and n ≡ 2a−1 + t (mod 2a ) for some integers s and t with 1 ≤ s, t ≤ 2a−1 then Rm×n×((m−1)(n−1)+1) has plural typical ranks. Proposition 8.5 Let a = 4, 8. If m and n are divisible by a, then for each 1 ≤ k < a, Rm×n×((m−1)(n−1)+k) has plural typical ranks. 36 Proof For a = 4, 8, if m and n are divisible by a, then m#n ≤ m + n − a by Proposition 2.3 and thus m + n − 1 − (m#n) ≥ a − 1. Then the assertion follows by Theorem 1.2. Corollary 8.6 (1) R4×4×k has plural typical rank whenever 10 ≤ k ≤ 12. (2) R8×8×k has plural typical rank whenever 50 ≤ k ≤ 56. Proposition 8.7 Let m, n ≥ 3. If Rm×n×((m−1)(n−1)+1) has a unique typical rank, i.e., m#n = m + n − 1, then trank(m, n, (m − 1)(n − 1) − k) = {(m − 1)(n − 1) + 1} holds whenever 0 ≤ k < (m−1)(n−1) m+n−1 . Proof Let 0 ≤ k < (m−1)(n−1) m+n−1 , q = (m − 1)(n − 1) − k and p = (m − 1)(n − 1) + 1. Suppose m×n×p that R has a unique typical rank. Then trank(n, p, m) = {p}. Since the set of all n × p × m tensors with rank p is a dense subset of Rn×p×m , the image of this set by a canonical projection Rn×p×m → Rn×q×m is also a dense subset of Rn×q×m . Thus any typical rank of Rn×q×m is less than or equal to p. On the other hand, by elementary calculation, we see that (m − 1)(n − 1) < (m−1)(n−1) mnq n×q×m is greater than or equal m+n+q−2 ⇔ k < m+n−1 . Thus the minimal typical rank of R n×q×m to p. Therefore R has a unique typical rank p. Corollary 8.8 Let 3 ≤ m ≤ n. Suppose that Rm×n×((m−1)(n−1)+1) has a unique typical rank, i.e., m#n = m + n − 1. If 0 ≤ k ≤ ⌊ m 2 ⌋ − 1 then trank(m, n, (m − 1)(n − 1) − k) = {(m − 1)(n − 1) + 1}. m m Proof Let 0 ≤ k ≤ ⌊ m 2 ⌋ − 1. Then (m + n − 1)(k + 1) ≤ (m + n − 1) 2 ≤ (n + n − 1) 2 < mn and thus (m + n − 1)k < (m − 1)(n − 1). Therefore the assertion follows from Proposition 8.7. References [1] J. F. Adams. Vector fields on spheres. Ann. of Math. (2), 75:603–632, 1962. [2] J. Adem. Some immersions associated with bilinear maps. Bol. Soc. Mat. Mexicana (2), 13:95–104, 1968. [3] M. F. Atiyah and I. G. Macdonald. Introduction to commutative algebra. Addison-Wesley Publishing Co., Reading, Mass.-London-Don Mills, Ont., 1969. [4] W. Bruns and U. Vetter. Determinantal rings, volume 1327 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1988. [5] M. V. Catalisano, A. V. Geramita, and A. Gimigliano. Ranks of tensors, secant varieties of Segre varieties and fat points. Linear Algebra Appl., 355:263–285, 2002; erratum: 367:347–348, 2003. [6] M. V. Catalisano, A. V. Geramita, and A. Gimigliano. On the ideals of secant varieties to certain rational varieties. J. Algebra, 319(5):1913–1931, 2008. [7] R. L. Cohen. The immersion conjecture for differentiable manifolds. Ann. of Math. (2), 122(2):237–328, 1985. 37 [8] D. M. Davis. A strong non-immersion theorem for real projective spaces. Ann. of Math., 120(3):517–528, 1984. [9] D. M. Davis. Some new nonimmersion results for real projective spaces. Bol. Soc. Mat. Mexicana (3), 17(2):159–166, 2011. [10] D. M. Davis and M. Mahowald. Nonimmersions of RPn implied by tmf, revisited. Homology, Homotopy Appl., 10(3):151–179, 2008. [11] D. Eisenbud, C. Huneke, and W. Vasconcelos. Direct methods for primary decomposition. Invent. Math., 110(2):207–235, 1992. [12] S. Friedland. On the generic and typical ranks of 3-tensors. Linear Algebra Appl., 436(3):478– 497, 2012. [13] M. Ginsburg. Some immersions of projective space in Euclidean space. Topology, 2:69–71, 1963. [14] J. Heintz and M. Sieveking. Absolute primality of polynomials is decidable in random polynomial time in the number of variables. In Automata, languages and programming (Akko, 1981), volume 115 of Lecture Notes in Comput. Sci., pages 16–28. Springer, Berlin-New York, 1981. [15] F. L. Hitchcock. The expression of a tensor or a polyadic as a sum of products. J. Math. Phys., 6(1):164–189, 1927. [16] M. Hochster and J. A. Eagon. Cohen-Macaulay rings, invariant theory, and the generic perfection of determinantal loci. Amer. J. Math., 93:1020–1058, 1971. [17] A. Hurwitz. Über die Komposition der quadratischen Formen. Math. Ann., 88(1-2):1–25, 1922. [18] J. Ja’Ja’. Optimal evaluation of pairs of bilinear forms. SIAM J. Comput., 8(3):443–462, 1979. [19] G. Kemper and N. Viet Trung. Krull dimension and monomial orders. J. Algebra, 399:782–800, 2014. [20] K. Y. Lam. Construction of some nonsingular bilinear maps. Bol. Soc. Mat. Mexicana (2), 13:88–94, 1968. [21] J. Levine. Imbedding and immersion of real projective spaces. Proc. Amer. Math. Soc., 14:801–803, 1963. [22] H. Matsumura. Commutative ring theory, volume 8 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, second edition, 1989. Translated from the Japanese by M. Reid. [23] R. J. Milgram. Immersing projective spaces. Annals of Mathematics, 85(3):473–482, 1967. [24] M. Miyazaki, T. Sumi, and T. Sakata. Typical ranks of certain 3-tensors and absolutely full column rank tensors. preprint, arXiv:1103.0154v2, Dec. 2012. 38 [25] M. Nagata. Theory of commutative fields, volume 125 of Translations of Mathematical Monographs. American Mathematical Society, Providence, RI, 1993. Translated from the 1985 Japanese edition by the author. [26] J. Radon. Lineare scharen orthogonaler matrizen. Abh. Math. Sem. Univ. Hamburg, 1(1):1–14, 1922. [27] D. Rees. The grade of an ideal or module. Proc. Cambridge Philos. Soc., 53:28–42, 1957. [28] L. Robbiano. Term orderings on the polynomial ring. In EUROCAL ’85, Vol. 2 (Linz, 1985), volume 204 of Lecture Notes in Comput. Sci., pages 513–517. Springer, Berlin, 1985. [29] A. Seidenberg. Constructions in algebra. Trans. Amer. Math. Soc., 197:273–313, 1974. [30] D. B. Shapiro. Compositions of quadratic forms, volume 33 of de Gruyter Expositions in Mathematics. Walter de Gruyter & Co., Berlin, 2000. [31] N. Singh. On nonimmersion of real projective spaces. Topology and its Applications, 136:233– 238, 2004. [32] T. Sumi, M. Miyazaki, and T. Sakata. Rank of 3-tensors with 2 slices and Kronecker canonical forms. Linear Algebra Appl., 431(10):1858–1868, 2009. [33] T. Sumi, M. Miyazaki, and T. Sakata. Typical ranks of m × n × (m − 1)n tensors with 3 ≤ m ≤ n over the real number field. Linear Multilinear Algebra, 63(5):940–955, 2015. [34] T. Sumi, T. Sakata, and M. Miyazaki. Typical ranks for m × n × (m − 1)n tensors with m ≤ n. Linear Algebra Appl., 438(2):953–958, 2013. [35] J. M. F. ten Berge. The typical rank of tall three-way arrays. Psychometrika, 65(4):525–532, December 2000. [36] J. M. F. ten Berge and H. A. L. Kiers. Simplicity of core arrays in three-way principal component analysis and the typical rank of p × q × 2 arrays. Linear Algebra Appl., 294(13):169–179, 1999. 39 

0math.AC

Data-driven Workflows for Microservices Larisa Safina∗ , Manuel Mazzara∗ , Fabrizio Montesi† ∗ Innopolis arXiv:1511.02597v1 [cs.PL] 9 Nov 2015 University, Russia {l.safina, m.mazzara}@innopolis.ru † University of Southern Denmark fmontesi@imada.sdu.dk of a system is highly dependent on the structural design of the organization producing it. In the context of microservices the Jolie programming language [5], [6] emerged as a paradigmatic solution tuned at getting the best out of this architectural style. Jolie is comprehensive and capable of implementing both simple services and complex orchestrations. Everything in Jolie is a microservice and all these microservices can be easily reused or composed for obtaining in turn new microservices. This approach supports distributed architecture and guarantees simple managing of components, which reduces maintenance and development costs. This work is devoted to extend the Jolie programming language in order to support data-driven workflow, i.e. moving the control-flow decision-making from process-driven to datadriven. That means that control-flow can be directed at the time of message passing according to the nature of the message strucutre and type, instead of requiring post-reception I. Introduction processing. A typical example of process-driven workflow is The increasing complexity of modern software requires new presented in [10]. In the microservice scenario this opportunity approaches to architectural design and system modeling. Com- opens to novel programming patterns. Whether the processplex systems also show high level of concurrency, i.e. multiple driven or data-driven programming style is more suitable for an intertwined threads of executions, often running on different application really depends on the specific problem domain and, hardware, which need to be synchronized and coordinated, to same extent, to developers preferences and style. Data-driven and which much share information, often through different flows have been realized in Jolie by means of an extension to its paradigms of communication. Therefore improving software type system, which manifests as implementation of choice type. quality and deploying reliable services is the consequence Further extensions can be implemented in order to strengthen of an accurate use of optimal service-based architectural further this possibility. styles and well-established software engineering techniques Two major contributions appear as a result of this work. for requirements elicitation, design, testing and verification, in The first is of scientific nature, i.e., moving from processparticular when it come to concurrent service-based systems. driven to data-driven and therefore open to new programming In order to tackle these issues from the architectural viewpoint, patterns and styles. The second is of purely technical interest, Microservices architecture appeared lately as a new paradigm and stands in the re-engineering of the Jolie interpreter as for programming applications by means of the composition of a consequence of the extended data type. This represents a small services, each running its own process and communi- relevant case study of interpreter re-engineering, and therefore cating via light-weighted mechanisms [8]. This approach has can be valuable experience for the practitioners involved in been built on the concepts of Service-oriented Architectures[14] activities of comparable complexity. brought from crossing-boundaries workflows to the application The paper is structured as follows: in section II a quick level, and into the applications architectures, i.e. it Service- overview of Jolie programming language is given in order to oriented Architecture and Programming from the large to the intorduce the unfamiliar reader with the main concepts and small. syntax. Far from being an exhaustive report, the section is Microservices architecture still shows distinctive character- just a compendium providing the links for further study of istics which blend into something unique and different from the topic. Section III presents a case study on top of which SOA itself. The size is comparatively small versus a typical the major narrative of the paper is built and the contributions service [8], supporting the belief that the architectural design are defined. In particular, two different kind of approaches Abstract—Microservices is an architectural style inspired by service-oriented computing that has recently started gaining popularity. Jolie is a programming language based on the microservices paradigm: the main building block of Jolie systems are services, in contrast to, e.g., functions or objects. The primitives offered by the Jolie language elicit many of the recurring patterns found in microservices, like load balancers and structured processes. However, Jolie still lacks some useful constructs for dealing with message types and data manipulation that are present in service-oriented computing. In this paper, we focus on the possibility of expressing choices at the level of data types, a feature well represented in standards for Web Services, e.g., WSDL. We extend Jolie to support such type choices and show the impact of our implementation on some of the typical scenarios found in microservice systems. This shows how computation can move from a process-driven to a datadriven approach, and leads to the preliminary identification of recurring communication patterns that can be shaped as design patterns. computation are described: process-driven and data-driven, and examples of both are explained in order to understand the differences. Sections IV and V describe the architecture of the Jolie interpreter and the changes that were necessary in order to extend the type system. Finally, section VI wraps up final considerations regarding the contribution of this work, and presents ideas on how future developments may be built up on top of current achievements. II. Jolie language In this section we brielfy recall the Jolie programming language in order to simplify the reading of the remaining part of the paper. Each Jolie program is composed by two parts, a behavioral part and a deployment part. A program can be formally expressed as: Program F Deployment Behavior 1) Deployment part: The deployment part contains directives which help the Jolie program to receive and send messages and be orchestrated among other microservices. The deployment part is separated from the program behavioral part, so that the same behavior may be reused later with a different deployment configuration. Formally the deployment part is expressed as: Deployment F DeploymentInstruction* Where DeploymentInstruction can include • Interfaces: sets of operations equipped with information about their request (and sometimes response) types; • Message types: can be represented as native types, linked types or undefined. Message types are the main subject of this report and shall be discussed later in section II-A. • Communication ports: define how communications with other services are actually performed. 2) Behavioral part: The behavioral part contains microservice implementation of the functionalities, containing both computations and communication expressions. Examples of expanded behavioral part utilization will be provided later in section V-B. Formally, behavioral part can be expressed as: Behavior 3) Communication: Communication of processes can be performed by two possible patterns: one-way (the endpoint receives a message) and request-response (the endpoint receives a message, and sends a response back to the caller): Process F | InputStatement F | OutputStatement F | . . . | InputStatement OutputStatement op(x) op(x)(y) Process op@OPort(x) op@OPort(x)(y) (One-Way) (R-Response) (Notification) (S-Response) One-Way is used to receive a message for operation op in variable x. Request-Response is used to receive a message for operation op in variable x, execute a Process and then send back a response to the caller containing the value of variable y. Notification and Solicit-Response are the dual of the former ones to be used, respectively, for sending a message to a OneWay statement or to a Request-Response one. OPort defines an output port name of desired endpoint. A. Jolie type system Jolie provides a language for describing the types that are allowed to be communicated over a network. Communications are type checked at run-time when a message is received [13]. Message types are introduced in the deployment part of Jolie programs: DeploymentInstruction F . . . | type id : TypeDefinition Where id is an identifier in order to use the message type in other program parts and TypeDefinition can be a native type, native type with subtypes, native type with undefined subnodes, link type or can be undefined (means that variable is null until a value is assigned to it): TypeDefinition F | | | | | NativeType F ... NativeType NativeType { SubTypeList } NativeType { ? } id undefined int | double | string | raw | void | any F BehaviouralBlock* Where {?} represents untyped subnodes and undefined stands main {Process} BehaviouralBlock* a shortcut for any:{?}. BehaviouralBlock F define id {Process} Except commonly-used native types as int, double or string, | init {Process} Jolie also has the following types: Main is a procedure which defines an entry point of execution. • raw (used for transmission of raw data streams as byte Main can be followed or preceded by define procedures with arrays) id identifier. • void (is used for indicating that no value is contained by Init supports special procedures for initializing a service the variable) before it makes its behaviors available. Procedures specified • any (means that any native type with which variable is with define can be used many times, while the one specified initialized will be accepted. with init is executed only once, when the service is started. Untyped subnodes, expressed as {?} construction, indicate Process defines the activities to be performed by the service. that a node may have any kind of subtree. Already defined Processes can be composed in sequences, parallels and (input types can be reused in other types definition as link types by guarded) non-deterministic choices [5]. means of their ids. Type may have any number of subtypes, which bnf-form is the following: • • SubTypeList F SubType | SubType SubTypeList SubType F .id Cardinality : TypeDefinition Each subnode has its cardinality defined as one by default or as following: Cardinality F | | | | [ int , int ] [ int , * ] * ?  (Range) (Lower-bound) (Shortcut for [0,*]) (Shortcut for [0,1]) III. Case Study In this section we show how the extension to the Jolie type system led to an enhanced arsenal at developers’ disposal. We will proceed by examples. The short compendium of the Jolie syntax as presented in the previous section, combined with the code examples presented here, may be sufficient to grasp a general understanding. For a more comprehensive information the reader can refer to [6] and [3]. Let us consider an example implementing a car rental. This consists of three parts: • Server, which provides the rental service; • Client, which wants to use it; • Interface, which declares the operations by means of which client and server can interact with each other. customer: stores customers personal information (name, age, and driving license number) necessary to rent a car; car_return: holds the reference to customer’s profile who has rent this car, the car’s identification number and the state of the car after renting. / / Car r e n t i n t e r f a c e type customer : void { . name : s t r i n g . age : i n t . license : string } type c a r _ r e t u r n : void { . car_state : string . c ?: customer . car_id : string } interface CarRentInterface { RequestResponse : get_car ( customer ) ( s t r i n g ) RequestResponse : return_car ( car_return )( string ) } the server indicates the useage of operations defined in the above interface. This is done by the include directive at the beginning of the source code. The server’s deployment part contains declaration of input port, by means of which it can be accessed with the name and the protocol provided on the defined location. The behavior part of the program contains definition of two operations, get_car and return_car. They are placed inside the square brackets "[ .. ]", which are used in input-guarded choice syntax. This means, that only one of the operations can be executed at the time, while the others will be deactivated. / / S e r v e r with input −guarded o p e r a t i o n s include " carRentInterface . iol " inputPort RentService { Location : " socket : / / l o c a l h o s t :2001" P r o t o c o l : sodep Interfaces : CarRentInterface } execution { concurrent } main { Figure III.1. [ get_car ( request ) ( response ){ r e s p o n s e = "43535" }] Client-service use-case diagram We wil consider two possible approaches for the server to handle clients’ requests: process-driven, implemented in Jolie language by means of input-guarded operations, and data-driven implemented by the newly added choice operator. We will then add some considerations on the usage of both. [ r e t u r n _ c a r ( request ) ( response ){ i f ( r e q u e s t . c a r _ s t a t e == " damaged " ) { r e s p o n s e = " Car i s damaged ! " } else { r e s p o n s e = " Thank you ! " } }] A. Process-driven approach In this example, the Interface contains the definition of two operations, dedicated to renting and returning the car, and several data types: } The deployment part of the client describes how the connection to the Rent Service works: the same interface, protocol and Rent Service location. The behavioral part of the client program executes the following operations: • creating the request information • sending this information to Rent Service to be processed by get_car procedure • Checking the response and printing it out / / Client . ol include " carRentInterface . iol " include " console . i o l " interface CarRentInterface { ... RequestResponse : process ( request )( string ) } process operation takes variable of request type, which itself can be either of customer or car_return types Let’s implement the new server, supporting process operation: include " carRentInterface . iol " outputPort RentService { Location : " socket : / / l o c a l h o s t :2001" P r o t o c o l : sodep Interfaces : CarRentInterface } main { / / sending request for a car c u s t o m e r . name = " J o h n S m i t h " ; customer . age = 32; customer . l i c e n s e = " l23454675 " ; get_car@RentService ( customer ) ( response ) ; println@Console ( " Car r e n t r e q u e s t i s a c c e p t e d . " ) ( ) ; println@Console ( " Car i d i s " + r e s p o n s e ) ( ) / / returing the car return . car_id = response ; r e t u r n . c a r _ s t a t e = " damaged " ; return_car@RentService ( return )( response ) ; println@Console ( " Car i s r e t u r n e d . " + r e s p o n s e ) ( ) } In order to better understand the execution, we show here the results of running the application: Process −driven approach Car r e n t r e q u e s t i s a c c e p t e d . Car i d i s 43535 Car i s r e t u r n e d . Car i s damaged ! The process-driven approach shows how, classically, the action flows is directed via the use of input-guarded choice. This approach is heavily influenced by the heritage of process algebra, as described in [3]. Input-guarded choice directed flow is indeed the basic mechanism in, for example, the untyped π-calculus [11]. Consequently, for this approach to be supported and implementable the langauge does not require to be particularly rich in term of type system and language primitives. Jolie itself originally supported only this mechanism. B. Data-driven approach In this case, the choice operator allows us to use a data-driven approach to computations. Let’s enrich the interface syntax with the following data type and operation: / / Car r e n t i n t e r f a c e ... type r e q u e s t : customer | c a r _ r e t u r n inputPort RentService2 { Location : " socket : / / l o c a l h o s t :2002" P r o t o c o l : sodep Interfaces : CarRentInterface } execution { concurrent } main { process_user_request ( request ) ( response ){ r e q u e s t match { customer { r e s p o n s e = "43535" } ; car_return { i f ( r e q u e s t . c a r _ s t a t e == " damaged " ) { r e s p o n s e = " Car i s damaged ! " } else { r e s p o n s e = " Thank you ! " } } } } Please note that "match" directive is not in the stable version of the project yet, so this example will be not compiled with the current (1.4.1) version of the interpreter. In this case we don’t need to separate execution of operations by means of processes. Here it is done by means of the type of the input request variable. In order to test new approach, let’s change the client program’s code: / / Client . ol include " carRentInterface . iol " include " console . i o l " outputPort RentService { ... } outputPort RentService2 { Location : " socket : / / l o c a l h o s t :2002" P r o t o c o l : sodep Interfaces : CarRentInterface } main { / / sending request for a car println@Console ( " Process −driven approach " ) ( ) ; ... / / returing the car ... / / w o r k i n g w i t h s e r v e r b a s e d on d a t a t y p e p r i n t l n @ C o n s o l e ( " Data − d r i v e n a p p r o a c h " ) ( ) ; process@RentService2 ( customer ) ( response ) ; println@Console ( " Car r e n t r e q u e s t i s a c c e p t e d . " ) ( ) ; println@Console ( " Car i d i s " + r e s p o n s e ) ( ) ; process@RentService2 ( r e t u r n ) ( response ) ; println@Console ( " Car i s r e t u r n e d . " + r e s p o n s e ) ( ) } The result of the execution will be the following: Data − d r i v e n a p p r o a c h Car r e n t r e q u e s t i s a c c e p t e d . Car i d i s 43535 Car i s r e t u r n e d . Car i s damaged ! The results are clearly the same than the process-driven approach, showing how behaviorally the two implementations appear undistinguishable for an external observer. However, as we have seen, the internal computations actually differ, and performances can differ too as long as other quality attributes. It is beyond the scope of this paper to make any quantitative assessment. It is enough to notice how enriched language mechanisms offer alternative programming pattern – and therefore desing patterns – to developers who adopt the service-oriented paradigm. The development of specific design guidelines is left as future work. IV. Architecture of Jolie interpreter In order to explain better the changes that were applied to the interpreter, we have to describe how its basic components are working together. The Jolie interpreter is written in Java, and its architecture is organized into several components [12], the most significant of which regard parsing the source files, establish the communication between components and running them. The Parser part scans and parses the source code, transforms and organizes it as a tree of objects with desired semantics. As a result, parser produces OOIT (Object oriented interpretation tree), which implements the execution of the semantic rules relative to the input program. Runtime environment instantiates other components and execute the OOIT. And the Communication core part is in charge of performing communication between different components abstracting from the communication methods and protocols. This section is dedicated to discussion of interpreter components in more details. A. Parser The parsing process follows the following stages and involve the following components: 1) Scanner reads input and create token objects based on the ones defined in token types enumeration; 2) OLParser is a recursive descent parser, which take created tokens, checks them by the grammar rules and generate the corresponding syntax node in the abstract syntax tree (AST); 3) OLParseTreeOptimizer takes ready AST and optimizes it by reducing the number of nodes or transforming the code to more efficient versions; 4) SemanticVerifier checks whether the code is well-formed and semantically correct. Both OLParserTreeOptimizer and SemanticVerifier use visitor design pattern [2] to access AST (implement OLVisitor class). Types mentioned in section II-A are expressed as nodes of AST. They are implemented by means of descendants of abstract class TypeDefinition: • • • TypeInlineDefinition for expressing native types; TypeDefinitionLink for linked types; TypeDefinitionUndefined for undefined types. B. Runtime environment OOITBuilder reads the AST and produces object tree-like data structure called object-oriented interpretation tree (OOIT), that defines the semantics for program execution. OOIT nodes implement Process interface, so that each node is responsible for the semantics of a single statement. Runtime Environment handles parallel execution by means of native threads. Threads responsible for executing a part of the OOIT can be of two types: 1) Session threads, which are used for handling different sessions and retain a local state for variable values. 2) Parallel threads, which are used for handling parallel composition and refer to their parent session thread for state handling. C. Communication core Figure IV.1. Jolie interpreter architecture Communication Core component is used for performing communications by means of such mechanisms as messages and channels. Channels are used for the sending and receiving of messages, which consist of resource path, name of operation they are dedicated for, message content and in some cases fault name. Channels are in charge of encoding/decoding messages using the right protocol and sending/receiving them by means of the right communication medium. V. Towards data-driven Workflows for Microservices applicable to choice type (like checking if a type has subtypes or its native type), and which might raise null reference exception while trying to access them. This is why these abstract methods were removed. Jolie still experiences lack of some data types and operations, which could enrich its syntax and add extra flexibility, like, for example, regular expressions or choice operator. This work is dedicated to implementing the choice operator as the one 2) Extending parser: Extension of parser requires working which is proved to be useful. on Scanner, OLParser, OLParserTreeOptimizer and SemanThe idea of choice operator was taken from XML[7]. Choice ticVerifier. operator in XML allows to put several elements in choice Scanner. No changes to the scanner are needed, because the element declaration, but to be presented with only one of them. pipe-symbol ("|"), which represents choice operator, is already For example, there is a choice element called "animal" in in use as the parallel operator, so it is already present in the example below, which can be either presented as a dog or as TokenType enum. Parallel operator, placed between operands, a cat: indicates that they are are executed concurrently. <x s : e l e m e n t name=" a n i m a l "> <x s : complexType > <x s : c h o i c e > <x s : e l e m e n t name=" c a t " t y p e =" c a t "/ > <x s : e l e m e n t name=" dog " t y p e =" dog "/ > </ x s : c h o i c e > </ x s : complexType > </ x s : e l e m e n t > statementA | { statementB ; statementC } However, execution relates to the behaviour part of the program, while message types have to be defined in the deployment part, and they are separately processed, so no effort on pipe symbol processing redefinition is needed. OLParser parses message types and their subtypes by means of parseTypes and parseSubTypes methods correspondingly. Despite choice element in XML allows to choose between These methods check that the sequence of tokens, generated several types, in our implementation, this construction provides by Scanner, is correct regarding the type definition grammar, possibility of choosing between two types only. The same as presented in section II and then generate a node in AST of one of the TypeDefinition descendants. Since this grammar has example will look in Jolie language in the following way: been enriched, parseTypes and parseSubTypes methods have t y p e a n i m a l : c a t | dog been changed correspondingly, so that they could expect "|" We use pipe character ("|") for choice operator. Note, that in token inside the type definition. this case cat and dog are linked types and need to be declared As OLParserTreeOptimizer and SemanticVerifier classes explicitly or otherwise Jolie interpreter will raise an exception. works with AST by means of visitors, so visit method for TypeDefinition formal grammar was enriched correspond- TypeChoiceDefinition type was added to OLVisitor class. In ingly: this case, no specific optimization is needed, so this visitor simply does the same actions as the other TypeDefinition class TypeDefinition F . . . descendants. | TypeDefinition | TypeDefinition SemanticVerifier checks whether the types with the same name have been already defined and checks cardinality of A. Architectural changes types. New visit method, added to deal with objects of In this section we will describe the architectural changes TypeChoiceDefinition type, is doing the same, except that necessary to implement the new data type. it also invoke semantic verification on both of its types inside. 1) Adding choice type to AST: In order to support possibility 3) Extending runtime facilities: Extending runtime facilities of storing two types, TypeChoiceDefinition class was created regards mostly building OOIT process and message types (extends TypeDefinition as TypeInlineDefinition and others). themseleves. Additions to building object-oriented interpretation tree process. OOITBuilder generates object oriented interpretation tree, based on AST produced by OLParser. It also uses visitors to access nodes of AST. Since new TypeChoiceDefinition type was added, corresponding visitor was created, which is in charge of creating type object based on current TypeChoiceDefinition object. Type objects are discussed in the next section. Figure V.1. TypeChoiceDefinition class diagram Message types. As messages are being passed by means of type objects, each of the TypeDefinition descendants in AST should have corresponding representation in Type class The TypeChoiceDefinition class contains two attributes left descendants (TypeImpl for TypeInlineDefinition, TypeLink for and right representing two possible types, that can be any of TypeDefinitionLink). TypeDefinition members. The addition of this class requires For representing TypeChoiceDefinition, new Type descena change in the parent TypeDefinition class and its other dant, TypeChoice, has been created. descendants, as they contain several methods, that are not Figure V.2. Jolie type system class diagram B. Examples 1) Working with types and subtypes: Choice operator can work with native types: type numeric : i n t | long Linked types: type linked_type : s t r i n g type linked_choice : linked_type | void Subtypes: It is also possible to use choice operator if you need two support several versions of data structure. For example, it was needed to process data related to Old-Software corporation, later its title has been changed to New-Software, but some customers can still use the old one. t y p e Old − S o f t w a r e −Corp : v o i d { . name : s t r i n g address : string } t y p e New− S o f t w a r e −Corp : v o i d { . name : v o i d { . firstname : string . lastname : s t r i n g address : string phone : i n t } t y p e c o r p o r a t i o n : Old − S o f t w a r e −Corp | New− S o f t w a r e −Corp person_info : void {. id : s t r i n g | i n t } 2) Functions genericity: One of the way of choice operator utilization is for making functions generic. If we declare the following choice type and function in interface type choice : s t r i n g | i n t fun_choice ( choice )( choice ) We can provide the same behaviour without considering which type (string or int) was passed to the function: fun_choice ( request ) ( response ){ response = request } Or imagine that we need to implement different behaviour of the function based on the arguments passed. Let’s have an interface with type person, able to be presented as element with type personSSN or personCCN: t y p e p e r s o n : personSSN | personCCN t y p e personSSN : v o i d { . ssn : i n t } t y p e personCCN : v o i d { . ccn : s t r i n g } And a function pay, which takes input argument of person type and run the corresponding code based on particular type (personSSN or personCCN) of the argument passed. pay ( p e r s o n ) ( r e s p o n s e ) { p e r s o n match { personCCN : . . . | personSSN : . . . } i f ( is_defined ( person . ssn ) ) { / / ask the person r e g i s t r y } else { / / c o n t a c t t h e bank } } VI. Conclusions and Future work Jolie is a comprehensive programming language based on the service-oriented paradigm [5], which emerged in the context of an extended research effort aimed at formalizing ServiceOriented Computing on top of broadly accepted models of concurrency in the EU Project SENSORIA (see, e.g., [3], [9]). Due to its support for the quick prototyping of both simple services and complex service coordination, Jolie has been used in the development of other research projects involving the programming and deployment of services (including [16], [15]). However, it has been identified how the language still lacks of data types and operations able to enrich the syntax and add extra flexibility to program common SOA scenarios. Regular expressions and choice operator are just examples of this deficiency. This work has been devoted to extend the Jolie type system in order to add the choice operator and realize the necessary changes into the interpreter. The major outcomes can be summarized as follows: • Identification of data types able to support common SOA programming scenarios • Addition of choice operator in the syntax and semantics of the language • Analysis and reengineering of Jolie interpreter This work shows how computation can move from a processdriven to a data-driven approach. Control-flow can be now directed at the time of message passing and according to the nature of the message structure and type, instead of requiring post-reception processing, hence leading to a preliminary identification of recurring communication patterns that can be then shaped as design patterns. In the microservice scenario this represent a novel opportunity opening to new programming scenario. Future work leaves space to both theoretical investigation and practical realization. Implementing regular expressions is a natural step in order to further enrich the type system and manage a broader set of programming scenarios. For what concerns theoretical aspects, formalization of the extended type system is considered a priority. Acknowledgements We would like to thank Innopolis University for logistic and financial support. This work was also partially supported by CRC (Choreographies for Reliable and efficient Communication software), grant no. DFF–4005-00304 from the Danish Council for Independent Research. Our gratitude goes to colleagues of the Institute of Technologies and Software Development who participated in the discussion and the seminar, in particular Bertrand Meyer, Victor Rivera, Daniel de Carvalho, Mohamed Elwakil, Leonard Johard, Alexander Naumchev, Alexander Chichigin and Rasul Tumyrkin. References [1] Choice operator, http://www.w3schools.com/schema/el_choice.asp. [2] E. Gamma, R. Helm, R. Johnson, J. Vlissides, Design Patterns: Elements of Reusable Object-Oriented Software, Addison-Wesley Professional; 1 edition, 1994. [3] C. Guidi, R. Lucchi and M. Mazzara, A Formal Framework for Web Services Coordination, Electronic Notes in Theoretical Computer Science, 180:2 pages 55-70, 2007. [4] C. Guidi, R. Lucchi, R. Gorrieri, N. Busi, and G. Zavattaro, SOCK: A Calculus for Service Oriented Computing., In Proceedings of ICSOC 2006, pages 327-338, 2006. [5] F. Montesi, C. Guidi, G. Zavattaro, Service-oriented Programming with Jolie, Book Chapter in Web Services Foundations, pages 81-107, 2014. [6] Jolie Language Website, http://www.jolie-lang.org/. [7] Jolie vision, http://www.italianasoftware.com/vision.html. [8] J. Lewis and M. Fowler, Microservices. Common characteristics of architectural style, http://martinfowler.com/articles/microservices.html. [9] R. Lucchi and M. Mazzara, A pi-calculus based semantics for WS-BPEL, Journal of Logic and Algebraic Programming, 70:1, pages 96-118, 2007. [10] M. Mazzara, F. Abouzaid, N. Dragoni and A. Bhattacharyya Toward Design, Modelling and Analysis of Dynamic Workflow Reconfigurations - A Process Algebra Perspective, Web Services and Formal Methods 8th International Workshop, WS-FM 2011, Clermont-Ferrand, France, September 1-2, 2011, Revised Selected Papers [11] M. Robin,Communicating and Mobile Systems: The pi-calculus, Cambridge University Press, 1999. [12] F. Montesi, Jolie: a Service-oriented Programming Language, Master’s thesis, University of Bologna, Department of Computer Science, 2010. [13] J. M. Nielsen, A Type System for the Jolie Language, Technical University of Denmark Informatics and Mathematical Modelling, 2013. [14] Service-Oriented Architecture (SOA) Definition http://www.servicearchitecture.com/articles/web-services/service-oriented_architecture_soa_definition.html. [15] M. Dalla Preda, S. Giallorenzo, I. Lanese, J. Mauro, M. Gabbrielli, AIOCJ: A Choreographic Framework for Safe Adaptive Distributed Applications, Proc. of SLE, pages 161-170, 2014. [16] M. Gabbrielli, S. Giallorenzo, F. Montesi, Applied Choreographies, arXiv, http://arxiv.org/abs/1510.03637, 2015. [17] Web Services Description Language (WSDL), Version 2.0. http://www.w3.org/TR/wsdl20/. 

6cs.PL

King Saud University College of Computer & Information Sciences General-Purpose Visual Language and Information System with Case-Studies in Developing Business Applications Mahmoud Fayed Riyadh May 11, 2017 Master of Science Thesis ii DEDICATION This Thesis is dictated to my Father; to my greatest Mather, to my lovely family. iii DECLARATION I hereby declare that I am the sole author of this thesis. I authorize King Saud University to lend this thesis to other institutions or individuals for the purpose of scholarly research. iv ACKNOWLEDGMENTS I would like to express my gratitude to my supervisor Dr.Atif M. Alamri, for his inspired comments and his patience to lead this thesis to the best level. Also, I would like to say thanks to all those who helped me in this work. Special Thanks to Dr. Ahmed Zayed Emam and Dr. Djamal Ziani for reviewing this thesis and providing very valuable advices and comments. Thanks to Dr. Muhammad S. Al-Qurishi and Dr. Ahmad A. Al-Daraiseh from our research team for their great contributions. Thanks to our Programming Without Coding Technology Team members:* Arcangelo Molinaro * Stephen France * Sameh Kamel * Amine Debabsia * Aboelsoud Abd Alraouf * Ahmed Omar * Abdulrazak Hussien * Djardi Masoud * Mouaddh Mebarki * Ibrahim Abdelelah * Dale Jones * Rabeh Torchi * Khalid Bentaleb * Reem Mohamed * Mohammed Fathy * Shreen Hassan * Sinan Qahtan * Sozan Samy * Vidlanovic Zlatko Aurel * Waseem Salem * Mohammed Al-Mahdi * Antonio Fidel Perez Rubio * Abobakar Ahmad Jallo * Salama Mohammed Abo Mosslam * Zerourou Bachir * Navid Dezashibi * Mahmoud Mohammed Hassan * Belarir Abdelkader * Mabrook AlRakhami * Nabeel Hassanien v ABSTRACT Learning computer programming has been always challenging. Since the sixties of the last century, many researchers developed Visual Programming Languages (VPLs) to help in this regard. In this thesis, ten VPLs were specifically selected, studied, experimented with, and evaluated. A total of fifteen metrics were used to evaluate the tools. Comparisons, classification, and gap analysis were then presented. A list of requirements for a general-purpose VPL and a guide to help the novice programmer choose the right tool were generated and finally the PWCT (Programming Without Coding Technology, a novel general-purpose visual programming language) is developed and presented. The main objective behind developing this tool was to create not only a general purpose VPL but one that possesses textual language’s capabilities. A language that can be used to develop similar programs to the ones developed in C++ or Java for example. As the name indicates, PWCT requires no coding at all. A person needs only to know basic programming concepts to be able to use the tool. PWCT has many attractive features such as: General Purpose, Open source, easy to use, Exports code to different languages, allows customizing visual components and creating new ones, uses keyboard shortcuts in addition to the mouse, and generates automatic documentation. PWCT has been launched as a Sourceforge project, which currently has more than 230,000 downloads for the language and more than 19,500,000 downloads for samples, tutorials and movies. Many business applications and projects are developed using PWCT, Also we developed the Supernova programming language and the Ring programming language using PWCT to prove that it can be used for advanced and large projects. Feedback from developers and results from the studies indicate that PWCT is a very appealing, competitive, and powerful language. vi ‫الملخص‬ ‫تعلم برمجة الكمبيوتر كان دائما تحدى ومنذ الستينات فى القرن الماضى اتجه الكثير من الباحثين نحو تطوير لغات‬ ‫البرمجة المرئية (الصورية) للمساعدة فى هذا الجانب‪ .‬فى هذه الرسالة تم إختيارعشرة لغات برمجة مرئية ثم تم بعناية‬ ‫دراستهم وتجربتهم وعمل تقييم لهم وقد تم استخدام ‪ 15‬معيار فى عملية التقييم وعمل المقارنات والتصنيف ثم تم عمل‬ ‫تحليل للفجوة الموجودة و تم تقديم قائمة من المتطلبات للغة برمجة مرئية متعددة األغراض باالضافة الى إرشاد لمساعدة‬ ‫المبتدئين فى تعلم البرمجة على إختياراألداة المناسبة ثم تم تطوير و تقديم تقنية البرمجة بدون كود وهى لغة برمجة‬ ‫مرئية متعددة االغراض و مبتكرة‪ .‬الهدف من تطويرها ليس فقط انشاء لغة برمجة مرئية عامة ومتعددة األغراض ولكن‬ ‫إنشاء لغة مرئية يمكن ان تقدم قوة وقدرات اللغات النصية على مستوى القدرة على تطوير التطبيقات والمشاريع التى‬ ‫تطور بلغات مثل السى‪ ++‬والجافا على سبيل المثال‪ .‬كما يشير اإلسم فإن تقنية البرمجة بدون كود ال تتطلب كتابة كود‬ ‫والمستخدم يحتاج فقط لمعرفة أساسيات مفاهيم البرمجة ليتمكن من إستخدامها‪ .‬تقنية البرمجة بدون كود لها الكثير من‬ ‫الخصائص مثل (عامة ومتعددة األغراض – مفتوحة المصدر – سهلة اإلستخدام – تنتج لنا الكود بلغات برمجة نصية‬ ‫مختلفة – قابلة لتخصيص المكونات وإنشاء مكونات جديدة – تتيح استخدام لوحة المفاتيح بجانب الفارة النجاز المهام‬ ‫بمرونة – تنتج توثيق تلقائى لخطوات انشاء التطبيقات(‪ .‬تم نشر تقنية البرمجة بدون كود كمنتج حر مفتوح المصدر‬ ‫على موقع السورس فورج للمشاريع مفتوحة المصدر وتم تحميل تقنية البرمجة بدون كود أكثر من ‪ 230‬الف مرة‬ ‫باالضافة الى تحميل يتخطى ‪ 19‬مليون و ‪ 500‬ألف بالنسبة لالمثلة والدروس التعليمية الخاصة بتعلم التقنية‪ .‬أيضا تم‬ ‫تصميم و تطوير لغة البرمجة سوبرنوفا ولغة البرمجة رينج بإستخدام تقنية البرمجة بدون كود إلثبات أن التقنية يمكن‬ ‫ان تستخدم فى المشاريع المتقدمة والكبيرة‪ .‬نتائج الدراسات وردود الفعل من المطورين تشير الى أن تقنية البرمجة بدون‬ ‫كود لغة برمجة مرئية قوية وقابلة للمنافسة‪.‬‬ ‫‪vii‬‬ TABLE OF CONTENTS Chapter 1 Introduction......................................................................................................................... 15 1.1. Introduction to Visual Programming Languages …………………….………….…………....... 16 1.2. Examples of Visual Programming Languages …...………….………………….….….……….. 17 1.3. Motivation………………………………………………...………………….….……………… 21 1.4. Research Questions………………………………………………………….…….……………. 21 1.5. Research Objectives……………………………………………………….…….…...…………. 22 1.6. Research Publications …………………………………………………….….……………….... 22 Chapter 2 Background......................................................................................................................... 23 2.1 Visual Programming Environment vs. Visual Programming Language…….……….……….… 23 2.2 Visual Programming Dimensions……………………………………….…………….……........ 25 2.3 Visual Representations…………………………………………………….……….…………… 25 2.4 Programming Paradigms……………………………………………………..…………………. 25 2.5 Syntax Directed Editor vs. Free Editor and VPL Compiler……………….….………………… 25 2.6 Visual Programming Languages’ Frameworks………………………….…….……………....... 26 2.7 Visual Programming Languages’ Issues ……………………………….…….………………… 26 Chapter 3 Literature Review............................................................................................................... 27 3.1 Previous Studies ……………………....………………………………..………………………. 27 3.2 Evaluation and Comparative Analysis for Selected Characteristics .………..…………………. 31 3.2.1 Selection Process …………………………………….…….…………..……………………... 31 3.2.1.1 Tools Selection …………………………………………………….….……………….…… 31 3.2.1.2 Metrics Selection ………………………………………………….….…………………...... 32 3.2.2 Comparative Analysis Process …..……………………………………..…………………….. 35 3.2.2.1 Scratch ………………………………….……………………………..……………………. 38 viii 3.2.2.2 Alice …………………………………………….….….…………………………………… 38 3.2.2.3 Prograph…………………………………………..…….…………………………………... 39 3.2.2.4 LabView………………………………………….……..…………………………………... 39 3.2.2.5 Forms/3………………………………………….…….…….………………………………. 40 3.2.2.6 VIPR………………………………………………..…….…………………………………. 40 3.2.2.7 Limnor……………………………………………..………….……………………………. 41 3.2.2.8 Tersus……………………………………………………………………………………….. 41 3.2.2.9 Envision………………………………………………….…...…………………………….. 42 3.2.2.10 Lava……………………………………………………………………………………….. 42 3.2.3 Implications ………………………………………………………...………………………... 42 3.2.3.1 The need for productive general purpose visual programming language ………..………... 43 3.2.3.2 A guide to help the novice programmer choose the right tool …..……………..………...... 45 Chapter 4 Proposed Solution and the Main Design .......................................................................... 46 4.1 The requirements for the new General Purpose Visual Programming Language...…..………... 46 4.2 Graphical Code Replacement ………………………………………………..…………............ 48 4.3 Graphical Code Replacement Method ……………………………………..………………….. 50 4.4 PWCT Architecture …………………………………………………………………………… 52 4.4.1 VPL Layer……………………………………………………………….…………………… 53 4.4.1.1 Goal/Module Designer …………………………………………….…………………......... 54 4.4.1.2 Components Browser…………………………………………….……………………........ 56 4.4.1.3 Interaction Pages…................................................................................................................ 58 4.4.1.4 Form Designer ……………………………………………….…………………………….. 58 4.4.2 PWCT Middleware Layer……………………………………….…………………………… 58 4.4.3 PWCT System Layer…………………………………………….………………………….... 62 Chapter 5 Implementation and validation........................................................................................... 63 ix 5.1 Usability and Capability Evaluation ……………………………..……………………………... 63 5.2 Efficiency Evaluation……………………………………………..……………………………... 65 5.3 Reputation Evaluation…………………………………………..……………………………….. 68 Chapter 6 Summary and Conclusion.................................................................................................... 77 6.1 Conclusion ………………………………………………………………………………………. 77 6.2 Future Direction………………………………………………………………………………….. 78 Bibliography......................................................................................................................................... 79 Appendices………………………………………………………...…………………………………. 88 A Guided Tour………………………………………………………….…..………………………………….. 88 A.1 Hello World Example………………………………………………..……………………………………. 88 A.2 How to maintain programs………………………………………………………………….……………. 97 A.3 GUI Applications……………………………………………………………………….………………… 98 A.4 Search in Visual Source………………………………………………………………….……………… 103 A.5 the Time Dimension…………………………………………………………………….………………. 103 A.6 See the code behind every step………………………………………………….………………………... 105 A.7 Dealing with multiple steps at the same time……………………………………….…………………… 107 A.8 the Time Machine and playing programs as movies…………………………….………………………. 108 B Students Progress In Learning PWCT…………………………………………….……………………….. 109 C Designers in our Visual Programming Languages’ Framework Designers …..….………………………… 118 x Table of Figures Figure 1: The Alice system ………………………………............................................. 17 Figure 2: The Scratch system………………………………………….….……………. 18 Figure 3: Application Modeling in the Tersus Platform…………….….…………….... 19 Figure 4: The Limnor system……………………………………….….…………….… 19 Figure 5: Visual Studio – Writing code using C#............................................................ 23 Figure 6: Netbeans – Using Java to create Mobile Application………….….……….… 24 Figure 7: Qt Creator – Using C++ to create multi-platform GUI Applications……...… 24 Figure 8: Evaluation and Comparative Analysis …………………………..………….. 31 Figure 9: Flowchart to help the novice programmer choose the right tool……..…….... 45 Figure 10: Source Code and the equivalent Steps Tree………………..……………..... 51 Figure 11: PWCT Architecture……………………………………..………………….. 53 Figure 12: A program in PWCT can have many goals………………..……………….. 55 Figure 13: Essential Use Case Model for Goal/Module Designer……..………………. 56 Figure 14: Components browser use case diagram……………………..…………….... 57 Figure 15: Components categories in PWCT Library…………………..…………….... 57 Figure 16: Time in seconds and Memory in MB and Number of user steps…..……….. 66 Figure 17: The difference of application source size between PWCT and VS……..….. 67 Figure 18: PWCT and Visual Studio actual interaction time in hours………..………... 67 Figure 19: Weekly downloads according to www.sourceforge.net.................................. 69 Figure 20: User satisfaction according to www.sourceforge.net...................................... 69 Figure 21: Multimedia applications screen snapshots developed using PWCT..………. 70 Figure 22: RVC Business applications screen snapshots developed using PWCT..….... 71 Figure 23: Multi-user client-server database application developed using PWCT…..….71 Figure 24: Production Follow-up Application developed using the GCR method…...... 72 Figure 25: Courses Management Application developed using the GCR method…....... 73 Figure 26: Critical Nodes Detection Application using the GCR method………..….… 74 Figure 27: The Supernova programming language developed using the GCR method... 75 Figure 28: The Ring programming language developed using the GCR method..…….. 75 Figure 29: The Ring Virtual Machine developed using the GCR method…………...… 76 Figure 30: The Ring Lists Implementation developed using the GCR method….…..… 76 Figure 31: PWCT – Goal/Module Designer………………………………………...…. .88 Figure 32: PWCT - Components Browser…………………………………………..…. 91 Figure 33: PWCT – Interaction Using Transporter window……………………..…...… 92 Figure 34: PWCT – Goal/Module Designer – Generated steps………………………... 93 xi Figure 35: PWCT - Components Browser – Using Keyboard………………………… 94 Figure 36: PWCT – Interaction window………………………………………………. 95 Figure 37: PWCT – Goal/Module Designer – Generated steps……………………….. 96 Figure 38: Application during the runtime…………………………………………….. 96 Figure 39: PWCT – Goal/Module Designer – Steps Tree - GUI application………..... 98 Figure 40: PWCT – Form Designer…………………………………………………… 99 Figure 41: PWCT – Components inside the domain “Controls”……………………… 99 Figure 42: PWCT – Interaction Page for the component “Define new window”……...100 Figure 43: PWCT – Interaction Page for Multi pages component……………………..101 Figure 44: PWCT – Interaction Page for Multi pages component……….…………….103 Figure 45: PWCT – Search for a step inside the Goal/Module designer………….…...104 Figure 46: PWCT – Using the Time Machine to go backward ………………………..104 Figure 47: PWCT – Looking at the code behind a step………………….…………….105 Figure 48: PWCT – Code Extractor Application………………………………………106 Figure 49: PWCT – Goal/Module Viewer……………………………………………..107 Figure 50: PWCT – Goal/Module Designer - Using the Time Machine………………108 Figure 51: Students Level in PWCT Course…………………………………………...116 Figure 52: YouTube Playlist Statistics for Level 1 of PWCT Course…………………117 Figure 53: YouTube Playlist Statistics for Level 2 of PWCT Course…………………117 Figure 54: YouTube Playlist Statistics for Level 3 of PWCT Course…………………117 Figure 55: Interaction Designer………………………………………………………...118 Figure 56: Transporter/Component Designer – Interaction Pages……………………..119 Figure 57: Transporter/Component Designer – Code Mask (Script)…………………..119 Figure 58: Transporter/Component Designer – Matching……………………………..120 Figure 59: Transporter/Component Designer – Rules………………….……………...120 xii List of Tables Table 1: Evaluation Metrics…………………………………………………….………..32 Table 2: Evaluation and a Comparative Analysis………………………………….…….36 Table 3: the average results of a simple program ……………….……………………....44 Table 4: Requirements for General-Purpose VPL ………….….………………………..46 Table 5: shows the replacement process for one line of code…….…………………..… 51 Table 6: Textual Programming Language’s vs PWCT’s terms…….………………....... 55 Table 7: Sample of the variables and values……………………………….…………… 60 Table 8: Sample of the variables of Interaction Page Variable and Code Mask….…...... 61 Table 9: GCR Rules to write a Code Mask………………………………….………...... 61 Table 10: Usability and Capability Evaluation…………………………….…………… 64 Table 11: The average results of a simple program ……………………….…………… 65 Table 12: The results of large scale application …………………………….………...... 67 Table 13: Goal/Module Designer buttons and their description…..……….…………… 89 Table 14: Students progress in learning PWCT………………………………………...109 xiii List of Abbreviation VPL Visual Programming Language GUI Graphical User Interface GPVPL General Purpose Visual Programming Language VPE Visual Programming Environment VR Visual Representation SDE Syntax Directed Editor FE Free Editor VPLC Visual Programming Language Compiler VPLF Visual Programming Language Framework TPL Textual Programming Language PWCT Programming Without Coding Technology GCR Graphical Code Replacement GD Goal/Module Designer CB Components Browser IR Interaction Designer FD Form Designer VS Visual Studio xiv Chapter 1: Introduction 1.1. Introduction to Visual Programming Languages In the age of information technology, software development plays a vital role in responding to companies’ and organizations’ needs for high-quality information systems. Providing highquality, reliable, scalable, efficient, easy to use, and cheap software is a nontrivial task. Meeting the requirements of such software requires proficient programmers and more productive software development tools. And so, there is an increasing need for software development tools that are easy to learn, general purpose, and highly productive [1]. As a result of the complexity in software requirements, many aspects of the software development process evolved and many tools were developed to help the programmers. Integrated development environments (IDEs) like Microsoft Visual Studio, NetBeans and Eclipse are essential for large projects. When using such tools, the programmers need to know not only the general programming paradigms but also the strict syntax of each programming language used. In these tools, the program representation in the source code is limited to text where photos and graphics can’t be part of the source code. Moreover, the more expressive the programming language is the more complicated its syntax becomes and this makes it harder to understand and code with. This challenge made it more appealing to use VPLs, which attract more programmers, and can increase the productivity of developing software [2-5]. A VPL combines the features of integrated development environments and information systems to provide a visual tool for applications’ development [5]. In addition to that, VPLs enable the development of applications and computer programs using more than one dimension, and provide a programming system based on interaction with graphical elements that mix between Text, Shapes, Colors, and Time, instead of typing text source code [6-10]. Currently, there are many VPLs in the market, but most of the successful and widely used ones are educational tools such as: Scratch, Alice & Kodu; or domain-specific such as: Max/MSP (Music and Multimedia) and LabView (Data Acquisition, Instrument Control and Industrial Automation). General-purpose VPLs like Limnor, Tersus, and Envision exist as well, but these languages are not widely used, According to TIOBE Index that measures the popularity of 15 programming languages. The TIOBE Index is updated once a month and uses rating based on courses, skilled engineer and third party vendors. Also it uses the popular search engines to calculate the ratings. [5, 11]. This survey is a step forward towards developing a new General-Purpose VPL that can be used by novice and mainstream-programmers to get more productivity without adding significant limitations that prevent the practical use of the proposed language. This research determines the limitations in previous implementations and defines the requirements of the new general-purpose VPL. In addition to that, a guide to help novice programming choose the right tool for their application is provided. 1.2. Examples of Visual Programming Languages VPLs improve the methods in which programmers convey the display and processing of information. They allow the users to comprehend and then alter the results visually by utilizing graphics, animations, drawings and icons [20]. Experts in computer graphics, programming languages, and human-computer interaction need to collaborate to develop a VPL. The first VPL was put into practice in the 1960s, however, the success of visual programming was not achieved until the middle of the 1980s when graphics hardware became widely used [18, 19, 94]. Since 1960, more than 60 VPLs and visual programming environment have been developed in several domains. These tools can be classified into three main categories, namely, Forms/Reports based, Diagrammatic, and Iconic [8]. Most of the successful VPLs are domain-specific languages. For example, in the science and industrial domains, LabVIEW is an eminent VPL, which is designed for automating the usage of processing and measuring equipment in any laboratory setup [12]. Ladder Logic is also a VPL used to develop software for Programmable Logic Controllers (PLCs) in industrial control applications [13]. In the multimedia domain, OpenMusic is an object-oriented visual programming environment for musical composition based on Common Lisp [14]. In the educational domain, Alice [15] and Scratch [16, 17] are reputable VPLs that allow people of any age to be familiar with the programming world. Besides being domain specific, many of the existing solutions are unable to deal with large and complex programming problems let alone preserving a reasonable level of readability and maintainability. Moreover, many of these VPLs are still struggling to avoid frustrating programmers with strict limitations, unnecessary complexity, low productivity, and problems related to the quality of the final product. 16 Alice is a VPL where objects are manipulated in a 3D world (as shown in figure 1). It was developed at Carnegie Mellon University; it gives students the chance to learn about objectoriented programming concepts without the syntax frustrations imposed by text-based programming languages. With Alice, a programmer using a Graphical User Interface (GUI) environment selects program constructs and methods from lists of available choices [15, 21-22, 95]. Figure 1: the Alice System [15] Scratch is developed at the Massachusetts Institute of Technology Media Laboratory. It is a new VPL and environment that supports the creation of interactive stories, games, animations, music and art projects (as shown in figure 2). Scratch allows educators to reduce the cognitive load that learners experience when they are initially introduced to programming [23-24, 96]. 17 Figure 2: the Scratch System [23] Kodu is an educational tool that was developed by Microsoft Research in 2010 to teach the basics of game development. Kodu is a new icon-based visual language programming made especially for creating games. It was designed to be appealing to children and enjoyable by everyone. The Kodu language is entirely event driven, where each line of programming is in the form of a condition and an action [25, 97]. LabView is a licensed software and it targets scientists and engineers. It is a graphical development environment with built-in functionality for data acquisition, instrument control, measurement analysis, and data presentation (Li et al., 2011). In LabView, the execution is determined by the structure of a graphical block diagram in which the programmer connects different function-nodes by drawing wires [13]. Tersus is an open source VPL developed by Tersus Software Ltd. It is used to build rich web and mobile applications by visually defining user interface, client side behavior, and server side processing (as shown in figure 3). It is a general purpose language that utilizes flow diagrams [93, 98]. 18 Figure 3: Application Modeling in the Tersus Platform [93] Limnor (shown in figure 4) developed by Longflow Enterprises Limited is general-purpose VPL. It is based on visual studio .NET. It can be used to create most applications, interactive multimedia kiosks, sales presentations, database applications with interactive query and search, business management systems, internet payphone kiosks, etc. Non-technically oriented users can use it [26]. Limnor supports code generation only in C#. Figure 4: the Limnor System [26] 19 Prograph is a popular general purpose VPL licensed under Prograph CPX and Marten. The Object-Oriented, Data-flow & multi-paradigm language has long been known to reduce development time through drag and drop of iconic symbols to create programs[38]. Envision is a general purpose VPL that integrates imperative and object oriented programming paradigm in its design. It is an open source language that can be used in small and large scale projects as it can adopt both a High Level and Low Level programming. It partially supports Import/Export of Code from Java & C++. It’s still under development [5]. VIPR is visual programming language that can be leveraged in small to large scale projects depending on the needs of a user. The language depicts both high level and low level functionalities on the basis of TCL and C++/TCL extensions respectively [90]. Lava is a general purpose VPL language but it is not used for production in the sense that it is only used for experimental purposes. As an open source language. The language is designed for Ease of Learning, Use, and Program Comprehension [92]. Forms/3 is a Declarative and spreadsheet-based high-level language that has long been used for general small and large projects. It is one of the language created to escalate development time in project realization[37]. Scratch, Alice, and Kodu are limited to the educational environment. Moreover, Scratch and Kodu were constricted at student level, while Alice can be used also at faculty level and other novice programmers to explore learning concepts of algorithmic thinking. 20 1.3. Motivation The requirements for software applications are increasing because computers are now very large part of our everyday life. Software programs are more complex because of many factors. Nowadays programs run on a wide variety of hardware like high-performance clusters, personal computers, embedded devices and distributed systems. Applications are developed for different fields and the cost can vary significantly as free open-source software compete with proprietary software. Decreasing cost, improving reliability and increasing scalability are some of the requirements facing software developers. In the age of information technology, software development plays a vital role to respond to companies and organizations need of high quality information systems. This leads to the need to more programmers and more productive software development tools to be able to respond quickly to companies need with high quality [5]. After the success of many domain specific VPLs like Scratch, Alice by reaching millions of users worldwide, it’s expected to find more interest in creating new VPLs that help novice programmers to learn programming and help mainstream programmers to create high-quality programs faster. Such languages must be designed carefully to solve current VPLs’ issues without creating new problems and this is an important factor for new VPLs to gain popularity. Also, to increase the usage of these new VPLs, they need to compete and/or integrate with mature development tools that are based on popular programming languages such as C++, Java, C#, Python and Ruby [5-7, 13, 23]. 1.4. Research Questions The questions that we aim to answer during this thesis are: 1. What are the problems that prevent mainstream programmers from using the current visual programming languages? 2. What is the design of a general-purpose visual programming language that can solve mainstream programmers problems and encourage them to use the proposed visual programming language in developing practical software? 21 1.5. Research Objectives The objectives of this research subject are: 1. Creating a Framework for creating visual programming languages to help us in creating the visual programming language components faster. The framework must contains the designers and domain-specific language to quickly create the visual programming language components. The implementation of each component should be done in little minutes to ensure the high-level of productivity. 2. Designing a General-Purpose visual programming language to make programming easier to understand for audiences other than programmers and to help expert programmers to quickly create programs and applications in less time. 3. Implementing the General-Purpose visual programming language through an information system that support programs and applications creation and manipulation, to reduce errors and to help users program faster. 4. Evaluating the system in developing business applications and comparing this to using other traditional programming languages that are text based. 1.6. Research Publications The author had participated in two International Conferences. These conferences provided opportunities for the research to be shared with researchers. This first paper present the PWCT software while the second paper present a novel algorithm that we implemented using PWCT. (1) Mahmoud S. Fayed, Muhammad Al-Qurishi, Atif Alamri, Ahmad A. Al-Daraiseh , PWCT: Visual Language for IoT and Cloud Computing Applications and Systems , ICC '17, March 22 2017, Cambridge, United Kingdom (2) M Imran, MA Alnuem, Mahmoud S. Fayed, Atif Alamri, "Localized algorithm for segregation of critical/non-critical nodes in mobile ad hoc and sensor networks", Procedia Computer Science 19, 1167-1172 22 Chapter 2: Background 2.1 Visual Programming Environment (VPE) vs Visual Programming Language (VPL) It is worth explaining the difference between VPLs and VPEs. Programming languages like Visual Basic, Visual FoxPro, Visual C# and Visual J# are not VPLs. All of these languages are textual languages. The programmer must write textual code using the language’s syntax. Other environments such as Microsoft Visual Studio are Visual Programming Environments (VPE) not VPLs [27]. Visual Studio (shown in figure 5), NetBeans (shown in figure 6) and Qt Creator ( shown in figure 7) provide designers and tools to create some parts of the application using visual components, but the textual code is necessary to complete useful and real applications. On the other hand, VPLs use different visual components instead of writing code. Figure 5: Visual Studio – Writing code using C# [106] 23 Figure 6: Netbeans – Using Visual Programming + Java textual code to create Mobile Application [106] Figure 7: Qt Creator – using C++ and the Qt framework to create multi-platform GUI Applications [106] 24 2.2 Visual Programming Dimensions In textual programming languages like C, C++ and Java, the code is text based and is unidimensional. In VPLs, the graphical code uses more than one dimension. Each graphical object has its place in 2D or 3D worlds. Each object may have its shape, colors and images. There are many relationships that can appear between objects such as: Inside, Outside, Touch, next-to, etc. Some languages also use the time as another dimension in the graphical code [27]. 2.3 Visual Representations A VPL can have one of the following four representations [28]: 1. Diagrammatic: uses components of shapes and text and use links to connect between shapes and represent the control or data flow. 2. Iconic: uses icons from the domain of the problem. 3. Form based: uses forms such as spreadsheet or data-entry forms. 4. Hybrid: uses a mixture of any of the above three. 2.4 Programming Paradigms Many programming paradigms exist. Imperative, Declarative, Procedural, ObjectOriented, Functional, Concurrent, Data-Flow, and Event-Driven Programming paradigms are common. All of these paradigms can be used in VPLs beside the visual programming approach. Some VPLs use one paradigm only, while others are considered multi-paradigm [29]. 2.5 Syntax Directed Editor vs Free Editor and VPL Compiler Some VPLs use a syntax directed editor which is syntax-aware and prevents the programmer from making syntax errors. When used, the user can’t make syntax errors, and after each step, the visual source or the graphical code will be correct, and can be executed directly after compiling without problems. Some VPLs compile the graphical code directly after each step during the design process. Other VPLs provide a free editor where the programmer have freedom to make mistakes during the development process and the VPL compiler can discover the errors during the compiling process. While the syntax directed editor is more suitable for novice programmers, the free editor provides more flexibility to the advanced ones [30]. 25 2.6 Visual Programming Languages’ Frameworks A VPL’s framework is a collection of tools that enable the developers to create new VPLs in less time with less efforts and better quality by utilizing ready to use and well tested tools [31]. In this thesis we created a visual programming languages’ framework to use for developing the general-purpose visual programming language. In the Appendix C. we shows the designers in our visual programming languages’ framework. 2.7 Visual Programming Languages’ Issues In general, four main issues need to be addressed when talking about the limitations of VPLs, namely, space, static representation, domain, and comments. In terms of space, almost all visual representations are physically larger than the text they generate, so the space used to show a program in a VPL is greater when compared with a text based program. Regarding the static representation, many programs created by VPLs look like a maze of wires that are hard to understand. Concerning the domain, a VPL can be classified as General or Specific domain. The most successful VPLs are designed for specific applications because general purpose languages are difficult to design and implement. In regard to comments, many VPLs don’t provide a place for writing comments [32-35]. 26 Chapter 3: Literature Review In a study about LabVIEW [36], the authors designed a survey to capture the cognitive effects of visual programming representations. The research was centered on a well-designed questionnaire to capture users’ perceptions on visual effects of LabVIEW. After collecting and analyzing the data, it was found that most of the respondents felt that LabVIEW is an effective tool. Moreover, it was clearly established that the users found the visual aspects of LabVIEW to be more appealing than the other features rated during the survey. In another study about Forms/3 and the Spreadsheet Paradigm [37], The main goal of the authors in this article is to show that procedural abstraction, graphics, and data abstraction are possible in the spreadsheet model. The authors presented a review based on forms/3 in a bid to show how limitations in a spreadsheets paradigm of programming can be alleviated for better enduser functionality. The author utilizes forms/3 in this research as a prototype from a language design perspective and from a human oriented perspective. It was found that it was actually possible to improve spreadsheets beyond the state they were at the time of documentation of this survey. Some of the features that were seen to have potency for improvement in spreadsheets were graphical types, gestural programming, dynamically-sized grids, generalized abstractions, graphical I/O, time travel, and steering. In a Survey about Data Flow VPLs [18], The Authors did an in-depth review of fifteen data flow VPLs. The tools were first categorized depending on their application domain. Then, they were studied for characteristics and strengths. The authors argue that for a language to be considered powerful it should have predefined functions and programming constructs. Lastly, they gave an overview of unsolved problems in the topic such as, the large interfaces of such VPLs. The article concluded by giving a general overview of the characteristics of data flow VPLs, the strengths and weaknesses at the time of research. It’s worth mentioning here that the authors stated that the creation of visual languages that suit a wide array of application domains can be a good area of research, not to mention that users should also be allowed to choose their level of expertise rather than being presented with rigid functionalities by languages. 27 In a Pleminary Report about PROGRAPH [38], the authors presented a review of PROGRAPH which is a VPL based on data flow computational model. The main goal was to provide the reader with a detailed analysis of the features of PROGRAPH plus its built-in operations. And hence, the article was structured to make it easy for novice PROGRAPH enthusiast to get an understanding of the core concepts of the language. The authors integrated many examples in the discussion in a bid to cover the whole language and to ensure that the reader understands this VPL better. In a study about the Users of Scratch [39], the authors analyzed Scratch to identify the problems faced by skilled programmers when using the software. The authors created four teams of programmers and asked them to provide a comparison between Scratch and traditional integrated development environments. They discussed the usability of the tool at the fundamental programming level as well as at the graphical user interface level. Later, the experience of the programmers on each of the two interfaces was categorized based on the Nielsen usability framework. The authors found that there is a broader field for possible improvements in the operation of Scratch based on the logical errors that were encountered by some of the users. Furthermore, the developers of the tool need to do a better job in explaining some of the tool’s aspects. In a study about LabVIEW and Simulink [40], the authors evaluated both LabVIEW and Simulink and did a comparison between them. Although there are many perceived differences between these two tools, the authors tried to find the relationship between LabVIEW and Simulink based on the criteria of their functionality as well as ease of use. The results showed that the majority of the users preferred using LabVIEW over Simulink due to its capabilities and ease of use. According to the authors, there is need to carry out a proportional study on LabVIEW and Simulink based on other criteria apart from the ease of use and functionality. Furthermore, even though LabVIEW has an advantage over Simulink based on graphical display and DSP hardware integration tools, it is not necessarily the best deal among the two. Extensive research is needed to develop a comprehensive analysis between the two languages. In another study about Teaching Programming Concepts to High School Students with Alice [22], the authors examined the possibility of using Alice to teach programming concepts to high school students. In this study, the authors divided 166 students into two groups and taught 28 C++ to one group and Alice to the other. The results of the achievement test showed that the Alice group scored higher than the C++ group. The results suggest that students who were taught Alice had a better understanding of the concepts of programming and were more interested in exploring the different aspects of programming on their own. In a study about Student Opinions of Alice in CSI [42], using a questioner, the authors studied the responses of eighty-four college students who were taught Alice first and then Java. The authors discovered that the majority (59.5%) of the students felt that their experience with Alice helped them learn Java, and 66.7% of the students felt that the department should keep using Alice in the curricula. In another study about Alice and Pair – Programming [43], the authors used 6 noncomputing major sections with a total of 89 students to determine to which extent the students enjoyed the programming environment provided by Alice. In two sections the students were asked to work separately, however, in the other four, the students worked in pairs. The authors gave the students few assignments and later asked them to write an essay about their experience. The results showed that the student enjoyed working with Alice, became more confident, and had a better understanding of programming concepts. In another study about Children Learning Computer Science Concepts via Alice Game Programming [44], the authors studied 325 middle school students to see if they can gain knowledge on the complex concepts of computer science. The students were asked to develop simple games using Alice. The authors reported that the students were able to learn many concepts such as, abstraction, modeling, control structure, and events’ handling. In a study about the effects of the media to promote the scratch programming capabilities creativity of elementary school students [87], the authors tried to see the effect of using Scratch on elementary school student’s creativity. They developed a simple program using Scratch to provide guidelines to students and used a sample of 60 students in their first grade to use it. The authors reported that using such application had a positive impact on the students’ creativity. In a study about Prograph 2.0 [88], the author expressed his opinion of a specific version of Prograph that includes a compiler. The compiler facilitates the production of executable 29 applications that operate as standalone. The author stated that this tool is well thought out, well integrated with Macintosh environment, and has good documentation. In a Review on Teaching and Learning of Computational Thinking through Programming [89], the authors reviewed 27 articles that focused on introducing computational thinking to K-12 curriculum. The authors found that the majority of the studies focused on introducing computational thinking concepts and few considered practice and perspective. And hence, they encouraged researchers to do more studies focusing on the latter two aspects. In a study about Visual Programming Languages [90], the authors studied five VPL tools namely, Ark, Vipr, Prograpgh, Forms/3, and Cube. They presented different classifications, and touched on the theory of VPLs in general. Later, they showed where the five tools fall. Lastly, they commented on VPLs related issues such as, control flow; and procedural and data abstraction. In a study about programming environments and languages for novice programmers [91], the authors in this article provided a taxonomy of a large number of tools. Some of the tools are VPLs and others are not. They categorized the tools in two major groups, tools to help teaching programming and others to support the use of programming (empowering) to achieve other goals. Then they divided the teaching category into multiple categories based “mechanics of programming”, “social learning” and “providing reason to program”. The empowering category was divided into two groups “mechanics of programming” and “activities enhanced by programming”. Each of the group was later divided into other groups based on different aspects. Each of the tools was classified according to the different categories mentioned above. Lastly, the authors concluded that the majority of the tools focused on the mechanics of programming and there is a need for tools to provide reasons so more people start learning programming. 30 3.2. Evaluation and Comparative Analysis for Selected Characteristics In this section we will compare between different VPLs, We have three stages, In the first stage we will do the selection process and In the second stage we will do the comparative analysis process. And in the third stage we will provide the implications as demonstrated in figure 8. Figure 8. Evaluation and Comparative Analysis 3.2.1. Selection Process The selection process is divided into two stages, In the first stage we will select the tools and in the second stage we will select the metrics. 3.2.1.1 Tools Selection In this chapter, ten different VPLs were studied and analyzed to show their strengths and weaknesses. The main goal of this study is to come up with a set of requirements for a generalpurpose VPL that is easy to learn, fast, can be appreciated by both novice and expert programmers, used to develop small and large applications, and produces output code in different languages. A minor goal of this study was to help a person decide which of the ten tools suits his needs. 31 There exist many visual programming languages. In this research, ten tools that cover different types of visual representations and design goals were selected. The tools were selected for different reasons, for example, Scratch, Alice and LabVIEW were selected because they are the most popular VPLs according to TIOBE index. LabView, Prograph, VIPR and FORMS/3 were selected because they are very popular in the literature. Tersus and Limnor were selected because they are general purpose and used in practical and different types of applications. Finally, Lava and Envision were selected because they look familiar to programmers with knowledge of textual code and paradigms like object-oriented. 3.2.1.2 Metrics Selection To start with, we had to decide on the criteria for assessment/evaluation. A team of researchers backed with few experts in the field [107] suggested fifteen different criteria as shown in Table 1. Most of the criteria are based on the common non-functional requirements in every software [108] and the common criteria of programming languages in general [109]. Criteria Domain & Scope Compatibility (Exporting/ Importing text code) Programming Paradigm Code SubCr Description SubCr Description Cr1 A Domain Specific B General-Purpose C High Level D Low Level Programming Programming E Can Create Large F Creates only Small Programs Programs Cr2 A Export to many B Import from many languages. languages. C Export to one D Import from one languages. languages. E Export not supported F Import not supported G Allows scripts and code. Cr3 A Multi-paradigms B One programming paradigm C Imperative D Declarative Programming Programming 32 E G License & Cost Cr4 Development Time I A C E A Cr5 (Decreased:Dec C Increased: Inc) E G I Size of programs Cr6 A C Readability (Decreased:Dec Increased: Inc) Cr7 Performance Cr8 Event-Driven Programming Data-Flow Programming Spread-sheet based Published Software Free Software Open Source Dec by Drag-AndDrop Dec by allowing writing input such as equations. Dec by Keyboard shortcuts Inc: many ops to do simple tasks such as assignment. Dec by providing tools such as GUI builder to create UI. Small (uses small blocks or icons). Small (hides optional details A Inc: using (text, colors, shapes & icons) C Dec due to the use of many connections between blocks. A Fast enough for applications F H J B D F B D F H J Object-Oriented Programming Procedural Programming Control-Flow Not published Commercial Closed Source Dec by High Level components. Dec by using other dimensions such as the Time dimension Inc: Mouse only is used. Inc: missing editor features like (Search, Replace, etc.). Dec using Liveness B Large (uses large components and many connections). D The size is dynamic B Dec due to missing features like changing the font size D Dec by allowing text similar to textual code. B Inc: compiles to languages like C 33 C Memory Usage Portability Inc: allows C extensions Cr9 A No critical problems found C Needs Large RAM (>1 GB) E Optimized by the programmer Cr10 A Multiplatform(Win, Mac, etc.) C Only MS-Windows E Only Apple Mac OS X Security and Safety Cr11 A No critical issues B Some Critical Problems D Needs small RAM (256MB) F Depends on the application. B One Platform D Only Web Based (Online) F Only Sun/Solaris B Depends on the developed app’ C Scalability Extension The tool is a prototype (issues) Cr12 A Suitable for small applications C Hard to manage large projects (missing OO) E Divides Apps into subroutines G Uses containment or nesting Cr13 A Support Extension D Application dependent C Documentation uses HTTP and JavaScript E The Ch interpreter Cr14 A Reference B Suitable for large applications D OO is used to manage large projects. F App can run across multi-cores H Uses Zoom in and Zoom out B No support for extension D Done by the Tool Developers B Getting Started 34 Support C Tutorials E Research Papers G Documents for Contributors Cr15 A There is support C Provided by a Company E No English Forum Table 1. Evaluation Metrics D Book F No Documentation H Online Documents B No Support D Provided by (Forums) Table 1 shows all fifteen criteria and 92 sub-criteria. The majority of them are selfexplanatory, and hence, there is no need to define them. It is important to know that, generalpurpose here means the language can be used to develop similar programs to those done with textual languages such as C++. To evaluate the tools, each tool was evaluated by three different team members for all 15 criteria. The results of each of the evaluators were then gathered and compared. To claim whether a tool has a specific feature or not, two of the evaluators at least needed to agree. Few randomly chosen points were re-evaluated by the main author to make sure that the evaluation process is accurate and complete. To gather the required information, the evaluators had to download, install, and experiment with the tools. They also visited the tools’ websites, read published articles about the tools, searched the web, and sent questions to developers or to forums. Each of the evaluator did their evaluation independently to be as comprehensive and unbiased as possible. The results of the evaluation are discussed in the next section. 3.2.2. Comparative Analysis Process Table 2 below shows the results of evaluating the ten tools. One of the main findings is that eight of the tools claim to be general purpose. According to our definition above, a VPL is general purpose if it can be in different domains and develop different types of applications. In other words, it has to be comparable to a textual language such as C++. Only the authors of Scratch and Alice state that they designed their VPLs to be used in education while the others claim to have developed general purpose VPLs and not for a particular domain. 35 In practice Lava, VIPR and Forms/3 are just research tools and are known to be used for serious development. Tersus is used in web and mobile database applications only. Prograph is used in desktop database applications and only mac os x. Limnor is used in database applications for windows – desktop and web but not for low level programming. LabView is used by engineers for automating the usage of processing and measuring equipment in any laboratory setup. Evaluating The other criteria doesn’t provide a proof that the above VPLs are general purpose that can replace C++ and java for system programming and applications development. T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T Lava Envision T Tersus T Limnor Cr6 T VIPR Cr5 T Forms/3 Cr4 T L-VIEW Cr3 T ProGraph Cr2 Description Domain Specific General-Purpose High Level Programming Low Level Programming Can Create Large Programs Creates only Small Programs Export to textual code in many languages Import from textual code in many languages. Export to textual code in one languages. Import from textual code in one languages. No support to export to textual code. No support to import from textual code. You can write scripts and code too. Multi-paradigms One programming paradigm Imperative Programming Declarative Programming Event-Driven Programming Object-Oriented Programming Data-Flow Programming Procedural Programming Spread-sheet based Control-Flow Published Software Not published Free Software Commercial Open Source Closed Source Increased using Drag-And-Drop Increased using High Level components that hide the details. Increased because some input can be written directly like equations. Increased because of using other dimensions like the Time dimension Increased because of using the Keyboard shortcuts to quickly create programs. Decreased because of using the Mouse only in the interaction. Decreased because of the need to connect between many components to create small tasks. Decreased due to missing editor features like (Search, Replace & Multi selection). Increased using the right tool for the right task like using GUI builder to create UI Forms. Increased using Liveness Small (for example: using small blocks or text and small icons). Alice Cr1 # A B C D E F A B C D E F G A B C D E F G H I J A B C D E F A B C D E F G H I J A Scratch Cr’ T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T 36 Cr7 Cr8 Cr9 Cr10 Cr11 Cr12 Cr13 Cr14 Cr15 B C D A B C D A B C D A B C D E F A B C D E F A B C A B C D E F G H A B C D E A B C D E F G H A B C D E Large (for example: using large components and many connections). Small due to hiding details that are optional The size is dynamic and can be changed by selecting more options Enhanced using visual representation (text, colors, shapes & icons) Decreased due to some missing features like changing the font size Decreased due to the use of many lines and connections between blocks. The text used in the visual representations looks like the textual code. No critical problems (The speed is enough to run most of the applications Increase Performance by creating Native Applications Can be increased by better programmers (Like adding C extensions ) Depend on the application (can be compared to textual code) No critical problems found that prevent using the software Critical Problems in memory usage Development tool need 1GByte RAM Development tool need 256Mbyte RAM Memory usage can be optimized by the programmer Memory usage depend on the application. Multiplatform(Windows, Linux & Mac OS X) One Platform MS-Windows Web Based (Online) Apple Mac OS X Sun/Solaris No critical problem the prevent using the software Depend on the application under development and the programmer skills The tool is just a prototype (bugs and problems are present) Very suitable for small applications Very suitable for large applications The absence of object-orientation/functional programming makes it difficult to scale The usage of Object-Orientation is good to manage large projects. Dividing programs to subroutines Applications Automatically scales and run across multiple cores Containment or nesting (avoid the graph edge problem) Scalable using Zoom in and Zoom out Support Extension No support for extension Can be done using HTTP requests and JavaScript Done by the Tool Developers The Ch interpreter Reference Getting Started Tutorials Book Research Papers No Documentation Documents for Contributors Online Documents There is support No Support Provided by a Company Provided by open source community (Forums) No English Forum T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T Table 2. Evaluation and a comparative analysis 37 Below we provide a brief description of the findings related to each tool. 3.2.2.1. Scratch The first Language under the microscope is scratch-a domain specific high level language. This language is domain specific in the sense that it is tailored to create Games, Stories & Animations courtesy of an imperative and event driven programming paradigm. Written in Written in Squeak and ActionScript, Scratch has a good number of advantages that makes it attractive for use with one of them being the fact that it is open source under GPL v2. and Scratch Source Code Licenses. The development time is scaled down a lot with this language thanks to the use of components and drag-and-drop way of programming. Most of blocks depend on small number of parameters that can be modified directly while the block size is compact to hide the unnecessary detail. The readability in this platform is also enhanced due to use text, colors and shapes. Its performance in terms of speed is also good enough to run most of applications (presentations, stories & games) created by students. It is also very portable as it is supported on Mac, Linux, Windows and on the Web. It also has free support and documentation that is also quite palatable for both starters and advanced users-this is through Getting Started, Reference & Video Tutorials. On the downside, it should be noted that some features are missing that might decrease the speed of maintaining programs. Some of these features are Search, Replace and multi-selection of block, expand/collapse of inner blocks. It is also suitable for small programs and applications given that the absence of object-orientation and functional programming makes it difficult to manage large projects. There is also No Standard support to export or Import from text based source code using languages like C/C++, C# & Java. However, this can be added by the aid of library for reading/writing MIT's Scratch file format. 3.2.2.2. Alice This is another domain specific high level language that is popular for programming in 3D environment mainly for education purposes. Apart from the fact that this language is open source, it has a wide array of features to put the icing on the cake. One of the most pronounced features is the speed of creating programs using advanced components and drag-and-drop functionalities, with most of blocks in use also depending on small number of parameters that can be modified directly. The size of these blocks is also pretty compact to hide unnecessary details. It can also be 38 used for big projects but it is highly recommended for small projects thanks to its adoption of object orientation. Furthermore, it is highly portable in the sense that it supports Mac, Linux and Windows not to mention that it works well with normal memory specs. In terms of compatibility, the language supports source code generation using Java. The presence of complete documentation and free support is also welcome among the community members. It is however important to note that for better performance, a 2 GB or more RAM, VGA graphics card capable of high (32 bit) color and at least 1024x768 resolution and a sound card are recommended. 3.2.2.3. Prograph This is a popular general purpose VPL licensed under Prograph CPX and Marten. The Object-Oriented, Data-flow & multi-paradigm language has long been known to reduce development time through drag and drop of iconic symbols to create programs. Though the size of the programs may look larger, this platform design alleviates this through the use of text, colors and shapes that increase readability. The ability to create classes or simply the object oriented nature of the language makes it very ideal in dealing with small and large programs. When it comes to portability, the language is supported in both Mac and Windows while it exhibits awesome speeds performance wise. The other colossal aspect is that it comes with books that can be handy not only to beginners but advanced users as well. On a negative dimension is that it requires MacOS X 10.6.8 or better and it does not have an official support platform barring email support. Also, it is not open source and one can only get trial versions for testing purposes. 3.2.2.4. LabView This is yet another Object-Oriented, Data-flow & multi-paradigm language that is popular among engineers and scientists. It is a high level general purpose language that has found great use in Data acquisition, instrument control and industrial automation among other scientific and engineering tasks. It was developed to accelerate development time among users through sophisticated elements coupled by drag and drop functionalities. This time is also reduced given the possibility to divide of programs into VI and the ability to run any written code to on multiple with no changes. Perhaps one of the most advanced feature of LabView is in its extensibility and Compatibility. In the realm of compatibility, the language can be integrated with graphical, textbased, and other programming approaches not to mention that users can write code and scripts too! This is made possible through a variety of possible extensions such as the Ch interpreter that 39 can be added to Labview for scripting. Users are also afforded the possibility of buying add ons for specific applications. The language also boosts a high quality documentation plus lots of tutorials and forums for user support. The only challenge is that large programs can grow complex especially with the presence of large blocks, lines as connections between blocks. The complexity has however been mitigated by enhanced readability through the use of text, colors, shapes and images in the visual representation. 3.2.2.5. Forms/3 This is Declarative, spreadsheet-based high-level language that has long been used for general small and large projects. It is one of the language created to escalate development time in project realization. This has been well orchestrated by a declarative and drag and drop programming capability it affords programmers. A good example of this is the ability to write or demonstrate equations and the ability to use time dimension in cells. It performs quite well and it does not consume much computer memory especially in research projects. This open source language has however been pegged back by security issues from bugs and other problems that is compounded by the lack of proper support and documentation since 2004. This has also played a role in the difficulty of extension development. It is also not supported by common Operating systems as it is only supported by Sun/Solaris and Hewlett-Packard/HP-UXat. 3.2.2.6. VIPR This is yet another powerful VPL language that can be leveraged in small to large scale projects depending on the needs of a user. The language depicts both high level and low level functionalities on the basis of TCL and C++/TCL extensions respectively. This way, the language is known to exhibit imperative, procedural and object oriented programming paradigms tailored for improved development time. Some admirable advantages of these is that it is to execute runtime animation of the code. Code debugging is also made easy through single stepping through program snapshots. The use of containment or nesting otherwise known as “Follow the Ring into the nesting” has also made the language scalable in that eliminates graph edge problem in programs. The language performance highly depends on the nature of program but it is worth saying that reduced latency between command issue and execution has increased performance. The readability of this platform rests on graphical primitives like rings, arrows, and parameter 40 rings. The memory requirements will also depend on the program being executed while its security will also be influenced by the programmer. The disadvantages of this platform is the lack of official support and documentation which has been a drawback has far as extensibility is concerned. It is also limited to Mac platform only. 3.2.2.7. Limnor This is another open source control flow VPL that can be leveraged for a wide array of applications ranging from building Windows, Kiosk and Web Applications. One of the huge advantages with language is the compatibility with common object oriented programming languages like C#, ASP.NET, PHP & JavaScript conglomerated with a control-flow paradigm. The drag and drop mode of programming coupled by use of the mouse for simple tasks-instead of the keyboard, cuts development time by far. Programs are generally large as shapes and arrows are used but readability is enhanced through text and color. The negative aspect of this platform is the lack of documentation and the memory requirementsit needs a CPU speed of over 2GHZ and 1GB RAM. Regardless, most of the support can be found in Limnor forum. The other disadvantage is the lack of third party extensions and plugins to scale functionality. It is also limited to windows platform only. 3.2.2.8. Tersus This is an open source data flow VPL language that is designed for creating mobile and web applications by virtue of visual diagrams. Based on AJAX, the language generates java code from visual programs meaning that it performs pretty well in with normal computer specs. Its extensions can also be done using java code. The other positive aspect of this language is the fact that it is supported by every common computer platform be it Linux, Unix, Mac and Windows. Its interactive design is another strong point of the language as it deploys drag and drop functionality in development. This plus the use of natural language for description and plenty of well blended color makes it quite easy for a user to navigate around hence reducing development time. Starters do not have to hassle a lot to learn as there is a documentation and lots of good tutorials. For advanced users, there is a forum that helps solves problems in the realm of the platform. 41 3.2.2.9. Envision This is yet another general purpose VPL that integrates imperative and object oriented programming paradigm in its design. It is an open source language that can be used in small and large scale projects as it can adopt both a High Level and Low Level programming. It partially supports Import/Export of Code from Java & C++. It actually uses programs like that look like textual code but only this time the development time is reduced as there are no syntax errors. Just like most of the VPL languages discussed, Envision enhances interaction through a blend of color and texts. It also affords the user a chance to zoom in and out of large code bases and millions of objects on screen. It is supported by common operating systems namely: Windows, Linux & Mac OS X. Its security, performance and memory requirements are all dependent on the kind of application being designed. The only glitch with this language is the lack of proper documentation, forums and tutorials but this is bound to change as a forum is under development. 3.2.2.10. Lava This is one other general purpose VPL language but the only one in the bunch that is not used for production in the sense that it is only used for experimental purposes. As an open source language, it is exudes quality in building small and large scale projects courtesy of an object oriented programming paradigm. The development time of a user is boosted with this platform given that it uses point-and-click structure editor, a GUI builder etc. The program size is relatively smaller as it uses Treeview enhanced with natural language text to boost readability. The language is supported on Mac, Linux and in Mac. Its performance, scalability and memory usage will depend on the application being developed. To the negatives now and it can be seen that, the platform cannot be extended by third party developers-it can only be done by the Lava team using C++ and Qt. There are lots of online documentations to learn from but no English forum to get support from other developers. 3.2.3. Implications In this section we will present the implications from the comparative study 42 3.2.3.1 The need for productive general purpose visual programming language From the aforementioned tables, we see that there is a wide array of VPL languages in the programming science to choose from depending on the needs and preferences. This paper presents some of the most common of these languages with an in-depth analysis of these languages based on a variety of metrics. The metrics of interests in this case are Domain & Scope, Compatibility, Programming Paradigm, License & Cost, Development Time, Size of programs, Readability, Performance, Memory Usage, Portability, Security and Safety, Scalability, Extension, Documentation and Support. It can be seen that this VPL vary depending on the metrics of interest. Alice and Scratch are domain specific while the rest are general purpose. In terms of portability, Scratch, Alice, Prograph, LabView, Tersus Envision and Lava cover Mac, Linux and windows platforms. Forms/3 runs on Sun/Solaris and Hewlett-Packard/HP-UXat while Limnor and VIPR will only run on Mac and Windows respectively. Most of these VPL are open source except Prograph and LabView. Turning attention to support and documentation, it is seen that VIPR, Envision and Lava lag behind the rest which have some sort of support. As mentioned earlier, these VPL languages reduce development time but their programs are generally large due to the use of visual graphics. It should however be distinguished that Lava and Envision have relatively small visual representations than the others. LabView is the most extensible of the bunch as it allows a variety of possible extensions such as the Ch interpreter to be added for scripting. Users are also afforded the possibility of buying add ons for specific applications. In this section we conducted an evaluation between well-known general-purpose VPLs (Tersus and Limnor) and two reputable VPE (Visual Studio and Net Beans). One professional programmer for each language was hired to develop programs in that specific language. All programmers were asked to develop 20 simple to intermediate applications. The authors monitored the time and memory needed to complete each task. Table 3 shows the averages of all tasks. We can find that the time required to create a simple application using VPE is the smallest amount of time among the other languages. This criterion is very important because it tackles the productivity. In terms of time, Visual Studio comes in the first place with approximately 117 seconds, then Tersus 140 seconds, Limnor with 144 second and lastly NetBeans 165 seconds. 43 With respect to the number of user steps used to build the application. We can notice that using Visual Studio and NetBeans need 26, and 33 steps respectively, whereas Limnor, and Tersus need 41,56, and 46 steps respectively. Building applications in VPLs involves many steps; however, most of these steps are duplicated and simple, which make them easy to learn for the novice user. On the other hand, Visual Studio and NetBeans require less number of user steps but one of these steps is writing code, which is complex and advanced. Regarding to the professionals, the productivity is very important factor, and as we notice from the table, Visual Studio proof the minimum development time regardless the number of user steps. Finally, memory usage by Tersus is near to 12.12 MB which is the least size for the created application; however, Visual Studio has 58 MB. Criteria\ Alt. Average time required to create and test the application (Seconds) The steps to create and test the application (steps) Average time required for each step (Seconds) Average Time spent using keyboard in user steps (Seconds) Memory used before application design starts (Kbyte) Memory used after the application design (Kbyte) Memory consumed in application representation during design time (Kbyte) General Purpose VPLs Limnor Tersus Visual Programming Environments Visual Studio Netbeans 144 140 115.66 165 54 46 26 33 2.66 3.04 4.45 5 13 18 31.1 30 13460 50268 51276 156052 29292 62392 109420 173352 15832 12124 58144 17300 Table. 3 the average results of a simple program 44 3.2.3.2 A guide to help the novice programmer choose the right tool The next flowchart (figure 9) provides a guide to help the novice programmer choose the right tool based on the purpose (Educational, Development or Research) and age. Start 1 Do you need VPL for educational purpose? Yes Scratch and Alice No 2 Do you need the VPL for developing purpose? Yes Prograph, LabView, Limnor, envision and Tersus No 3 Do you need the VPL for researching purpose? Yes Viper, Lava, Form3 No 4 Do you look for a VPL for children their ages <14 years? Yes Scratch No 5 Do you look for a VPL for students their ages >14 years? Yes Alice No You have to refine the search parameters Figure 9: Flowchart to help the novice programmer choose the right tool 45 Chapter 4: Main Design 4.1 The requirements for the new General Purpose Visual Programming Language We talked about the progress in the visual programming languages, the success in different programming domain, and the features of the most popular visual programming languages. According to our literature review and comparative analysis we see a gap in attracting mainstream programmers to using visual programming languages in each new programming task. The current VPLs are not equivalent to the current textual programming languages at the practical side. Yes, some limitations are a result of missing features that can be added but this doesn’t count only for “what we can do?” but for “how we can do things with respect to the productivity of the programmer during software development and the quality of the final software”. We will provide the requirements of a general purpose visual programming tool that can be used as replacement for the usage of popular programming languages like ( C,C++, C# and Java) + visual programming environments like Visual Studio and Netbeans). We expect that after the implementation and marketing of this new tool, many new software projects can be developed using Visual Programming Language in less time with high quality. Also we expect attracting more users to real programming industry after the existence of this tool. The next table provides the requirements for the new general purpose VPL. These requirements are the result of our comparative study where our research team selected the best features that we need to have together in one visual programming language. Index 1 Metric Domain & Scope 2 Compatibility 3 Programming Paradigm 4 License & Cost • • • • • • • • • Description General Purpose High Level Programming Low Level Programming Creating small and large systems and applications • Generate Code in different programming languages like C, C++ , C#, Java & Python Imperative Procedural Object-Oriented Programming Functional Free Open Source (GPL version 2) 46 5 Development Time 6 Size of programs Readability 7 8 9 10 11 12 13 14 15 Performance Memory Usage Portability Security and Safety Scalability • Exception allows developing commercial applications without releasing the source code • Increased using the keyboard, Intellisense/Autocomplete, Cut/Copy/Paste/Search/Replace • Interaction using Keyboard/Mouse • Colors are useful for interaction • Syntax Directed Editor (prevent errors) • VPL Compiler help in discovering errors when the Syntax directed editor is off. • Small – using TreeView • Better Using Text, Colors, Tree & Froms • Change the font name / size • Time Dimension • Depend on the application • Depend on the application • Windows, Linux and Mac OS X • Depend on the application • Depend on the application • Using Standard GUI controls like TreeView is faster than owner drawn controls. Extension Comes with Designers and a scripting language called creating new components or VPLs Documentation Complete documentation (Tutorials, Movies & Reference) Support Forums Table 4. Requirements for General-Purpose VPL 47 4.2 Graphical Code Replacement (GCR) Many researchers believe that Visual Programming Languages are inevitable tools to attract more people to the programming world. In the past half century, many Visual Programming Languages were developed. A large number of the developed languages are domain specific while only few are general purpose. In this chapter, PWCT (Programming Without Coding Technology, a novel visual programming language) is presented. The main objective behind developing this tool was to create not only a general purpose VPL but one that possesses textual language’s capabilities. A language that can be used to develop similar programs to the ones developed in C++ or Java for example. As the name indicates, PWCT requires no coding at all. A person needs only to know basic programming concepts to be able to use the tool. PWCT has many attractive features such as: General Purpose, Open source, easy to use, Exports code to different languages, allows customizing visual components and creating new ones, uses keyboard shortcuts in addition to the mouse, and generates automatic documentation Most of the successful VPLs are domain-specific languages. For example, LabVIEW (Laboratory Virtual Instrumentation Engineering Workbench), although it is claimed to be general purpose it is only popular in data acquisition and control applications [12]. Ladder logic is a VPL used to develop software for Programmable Logic Controllers (PLCs) in industrial control applications [13]. In the multimedia domain, OpenMusic is an object-oriented visual programming environment for musical composition based on Common Lisp [14]. In the educational domain, Alice [15] and Scratch [16, 17] are reputable VPLs that allow even little kids get familiar with the programming world. Besides being domain specific, developing software with VPLs becomes much harder as the software gets larger and more complex due to their inherent limitations and graphical representation. One main limitation of VPLs is known as Deutsch Limit. It states that “The problem with visual programming is that you can't have more than 50 visual primitives on the screen at the same time”[41, 105]. And so, for any VPL to be comparable to textual programming languages it needs to overcome this limitation at least. In fact, the authors believe that a superior VPL should have the following features to be a rival to text-based languages: • Pure visual language: the language should require no coding at all and no knowledge of syntax. • General purpose: the language could be used to create any software just like text-based language. • Suitable graphical representation: the graphical representation used in the language should be 48 • • • • • • suitable for small and large software development. Ease of use: the language should be easy to use and require short time to get mastered. Open source: an open source VPL encourages people to use it and allows users to tailor it to their needs. Capability of creating new visual components: a developer should be able to create new components easily. Capability of producing (exporting) code in different programming languages such as C# and java. Provides multi-lingual interface. Utilizes both mouse and keyboard interaction. In this chapter, we propose a novel technique called Graphical Code Replacement (GCR) to overcome Deutsch’s limit and to provide the needed scalability and flexibility. GCR was used in building a new VPL called PWCT. The proposed technique facilitates the application construction through a set of sequential interactions with visual components. GCR depends on a generated Tree of Steps (Steps Tree) that describes the program under development. The manipulation of TS is done through the interaction with visual components without the need to deal with text based code. The code generation process is managed in the background by the visual components to hide the complexity from the user. The visual representation of the program is constructed using traditional GUI controls (Tree, labels, textbox, list box ...etc). GCR replaces the common drag-and-drop technique to create an application by the give inputget generated representation technique. The input is accepted by visual components through interaction pages (Forms to get users’ inputs).The drag-and-drop method enable the user to select a component from a toolbox then drag it to be dropped in a design region. Here the visual representation of the component is determined from the beginning and attributes of the object can be changed from options available in the component representation as in scratch and Alice or from properties window as in limnor. But, in the GCR the visual representation is completely dynamic and is generated later after providing the options to the user through interaction pages (forms) and after that we get the visual representation. For example the if—elseif-else-endif component in GCR is one component which provides the optional elseif and else operations. And based on the user selection of these operations in the interaction page (forms) through checkbox the visual representation will be generated to include 49 elseif/else or not. In scratch we have separate components for the “else” operation. The proposed work introduces a new general purpose visual programming language by combining the strengths of the current successful VPLs. PWCT is completely visual. It uses a standard graphical notation for building a program. The source editor is replaced by a TreeView control with an adaptive and semantic representation. PWCT provides benefits for both novice and expert programmers. It pays special attention to how expert programmers can easily adapt to this new visual programming methodology. In addition, PWCT is built based on an open architecture where users can easily add/modify his own components. 4.3 Graphical Code Replacement Method NOTE: that the words "step", "control", and "component" are used interchangeably and they all mean a graphical control also the words "goal" and "module" are used interchangeably to mean a programming module. The basic idea behind the Graphical Code Replacement (GCR) was to find a way to make the visual programming language fully equivalent to a traditional textual programming language and to overcome all the limitations and weaknesses of VPLs. Looking at the structure of code written in textual programming languages, it can be easily seen that the code contains a set of instructions and commands written as statements in a specific syntax. The statements are arranged in nested structures similar to a computer tree with one or more nodes/leaves. A complex statement such as “For Loop” can be represented with a node that can be expanded, while a simple statement such as a “printf()” can be represented with a leaf. For example, the code shown in Figure 10 is written 50 in C. This code can be seen as a tree of statements or steps (a step here is defined as a visual component) arranged in a way to print numbers from 1 to 3, print a message, and then continue Printing 4 to 10. Keep in mind that each instruction in any programming language consists of two parts: the operation, and the required data for the specification of that operation. Figure 10: Source Code (in the right) and the equivalent Steps Tree (left). To better describe the concept, the following equivalences can be used: • Text code structure IS EQUIVALENT TO Nested structures of Instructions (Steps Tree) • Instruction IS EQUIVALENT TO a function F(X), where F is the operation or the function and X is the required data. • Text code structure IS EQUIVALENT TO Nested structures of F(X). Notice that F can be any function and X can be any input data. When the above equivalences are applied to the C code shown in Figure 10, we can have a clear view of the essential concept of the GCR technique. printf( “ This message between number 3 and number 4 \n“) ; Line of code (Source Code) Print text in new line The operation/function (F) “ This message between number 3 and number 4 “ Data = X Print Text “This message between number 3 and number 4” Step’s Description/Label Table 5. Shows the replacement process for one line of code. 51 At this point many questions need to be answered. How the steps will be generated automatically? How will they determine what the program will do? How can the developer/programmer control and manage the program as she does with traditional languages? GCR uses the Nested Tree (Steps Tree) to represent the structure of the program. Instead of using F(X) which is related to the syntax of the programming language, GCR uses natural language (English, Arabic, etc) to describe the steps (i.e. the name and description of the step or control which could be in any language). Therefore, it’s easy to understand and remember. The last row in Table 5. Implies that a step (visual component) was added to the tree and it was labeled “Print Text “This message between number 3 and number 4””. 4.4 PWCT architecture PWCT is composed of three main layers: VPL Abstract Layer; Middleware Layer; and System Layer. Figure 11. Shows each layer and its constituent components within PWCT’s architecture. In the following sections, each layer and its components are described in more detailes. 52 Figure 11: PWCT Architecture. 4.4.1. VPL Layer The VPL layer is the presentation layer. It provides all the functionality needed to allow a user to perform a specific task. It is composed of four important components: the Goal/Module Designer; the Components Browser; the Interaction Pages; and the Form 53 Designer. A user would interact with at least three components to produce a meaningful program. For example, a user starts with the Goal/Module Designer to create a new goal (in other words module). Next, the user might choose a visual component from the Components Browser. Then, the user might work with the Interaction Pages to provide the needed data or to specify some attributes. If the user wants to develop a GUI, they can use the Form Designer to do so. 4.4.1.1. Goal/Module Designer The Goal Designer is the main component in the VPL layer. It can be thought of as an equivalent one to the code editor in a textual programming language. Through the Goal Designer, a developer can manipulate her program using the Steps Tree. The tree structure was chosen to simplify the replacement process. Furthermore, the Treeview control has some special advantages which distinguish it from other controls. These advantages are listed below [99]: • Hierarchical data structure and relationships between elements can be represented and digested more clearly. • All controls and their children (data, control structure, etc.) can be entirely seen together in a single window. • Navigation and paging to focus on a particular part of data are easily manageable with expansion and collapsing functionalities when compared to alternative forms of viewing (i.e. tabular form view, text editor view, or list view). • The developer can arrange the controls in the run-time. She has the ability to move, update, delete and add any kind of program’s control according to her needs. Developers could have more than one goal and each goal could include more than one step as shown in Figure 12. These steps provide a visual representation of the code which is generated in the background. The developer doesn’t write the steps, instead, she selects them from the Components Browser and manipulates them using the Interaction Pages. 54 Module 1 Step1 Module 2 Step1 Step2 Module 3 … Step n Module 4 Step2 Step1 …… Step2 … Module n Step m Figure 12: A program in PWCT can have many goals/modules To make it easier to understand, table 6 shows a comparison between few text-based language’s terms and their PWCT counterparts. TPL’s Terms PWCT’s Terms Problem Goal/Module Statement/ Line of code Step Coding Selecting & interaction with components Code Editor Goal/Module Designer Table 6: Textual Programming Language’s vs Programming Without Coding Technology’s terms In the following, the use case diagram in figure 13 describes the behavior of the Goal Designer when the developer interacts with it. The use case diagram is composed of four main use cases addressing the interaction process. Goal Designer allows developers 55 to efficiently add, delete, or modify both goals and steps. Along with data processing, the developer can arrange the program steps according to his needs, even in the run-time (move-up, move-down, cut, paste, copy, etc.). The developer would need to open the form designer, when he/she defines a new window, to deal with the components and tools and modify their properties according to the requirements. Developer Figure 13: Essential Use Case Model for Goal/Module Designer 4.4.1.2. Components Browser There are several definitions of the word component in the literature. In [100], components are defined as units of independent production, acquisition, and deployment that interact to form a functioning system. The authors of [101] define components as executable units of code that provide a set of services through a specific interface. Lastly, components were defined as self-contained replaceable parts of a system that encapsulate their data and processes [102]. In PWCT, the last definition is adopted. The second component of the VPL layer is the Component Browser. Developer uses the Components Browser to search, select, and navigate amongst components in the components’ library. Figure 14 shows how a programmer might interact with Components Browser. 56 Figure 14: Components Browser Use Case Diagram In PWCT, a large number of components (496) were built and stored in a library. The developer can search the library for a specific component, select the component to be included in her program, or just navigate through the library to see the available components. The developer can use both the mouse and the keyboard to do such operations. Furthermore, the developer can limit her search to a specific category of components by selecting the required category. In addition to that, the developer can create a new component or modify an existing one if needed. In general, the components in PWCT are either controls (such as GUI controls like Windows, Labels, Buttons, Textbox) or software modules (such as Database, Threads and Sockets components) that perform a specific service needed by the programmer. A component could be designed to generate code in any language such as Harbor, C#, Python, etc. To better organize the components, several categories were created. Figure 15. Shows the most important categories. Basics Structure … Datatypes Consol … Windows xBase AdoDB … GUI Database Http Socket … Networks Paradigms OOP Procedural … Figure 15: Components’ Categories in PWCT Library 57 When selected, a component will be inserted as a step/s into the Steps Tree. Simultaneously, a link will be created between the generated step in the goal designer and a file of the selected component indicative of the step. The link information is stored in the data repository independently. Users may change/delete some parameters without affecting the component file. 4.4.1.3. Interaction Pages The interactive pages are acting like a properties panel in traditional languages, but they differ in their simplicity and ease of use. In addition, through the interactive page a user can change and adjust not only a component properties but also its functionality. For example we have the console application component that can be used to do different tasks related to console applications like (Print text, clear screen, draw rectangles, get input, and menus). Also we have the classes’ components where each class is represent by one component contains three pages to (Set Attributes, Get Attributes and Invoke methods). 5.3.1.4. Form Designer When a Graphical User Interface (GUI) is needed, a developer can use the Form Designer to create forms for her application. Similar to those that exist in Visual Studio; Eclipse; and other programing environments, the Form Designer in PWCT allows the user to build a form and add to it any typical component such as: a button, textbox, label, etc. There are components for controls events that determine the procedure/method that will be invoked when an event is fired. Using the events components the user can select and event (click event for example) then determine the procedure that will be called when the user click on the control. The procedure name can be written or selected from the list of names that appears during typing (intellisense). Later the user define the procedure using the define procedure component and add steps to the procedure that will be executed when the procedure is called. 4.4.2. PWCT’s Middleware Layer A Middleware is software that connects applications with the underlying operating system or database. The software consists of a set of services that allows multiple processes running on one or more machines to interact [103]. PWCT’s middleware connects the user’s view (i.e. the four components in the VPL Layer) with the system’s view (i.e. processes in the System Layer). 58 This layer (Middleware Layer) has four main functionalities: interpretation of user’s interactions with components in the VPL Layer; generating the Steps Tree according to user’s interactions; replacing variables in the script that generate the code with the values entered during the user’s interactions with components; and generating the source code. As demonstrated in figure 11, the Middleware layer consists of three units; the Step Generator, a database, and the source code files. The most important part of the middleware layer is the Step Generator. Generating the code in this layer if fairly simple and straight forward, Algorithm 1 illustrates the mechanism to generate the source code and to store it in a data repository. Lib_file ¬ open library file (*.TRF) InteractionPageTable ¬ load interaction pages data from (Lib_file) CodeMaskTable ¬ load code mask data from (Lib_file) CodeMatchTable ¬ load code match data from (Lib_file) if InteractionPageTable is not empty then tmpDataTable ¬ empty for each page in InteractionPageTable, do page_file ¬ open library file(*.IDF) tmpDataTable ¬ load controls name and values from (page_file) close the interaction page file endfor for each item in CodeMatchTable, do Control_Name ¬ load controls name from CodeMatchTable Control_Variable ¬ load controls variable from CodeMatchTable Control_Value ¬ load controls value from tmpDataTable(Control_Name) Codemask_Variable ¬ Control_Value endfor endif for each item in CodeMaskTable, do 59 if the item is an instruction then Execute the instruction with respect to GCR Rules else Add step code endif endfor close the library file Algorithm 1. Code Generation Algorithm To illustrate how this layer works, a simple scenario is presented below: When a developer selects a component using the components browser, an interaction page will open which asks her to provide more details. For example, if the user selects a print component, the interaction page will ask the user to insert the value she wants printed. This interaction will be interpreted and fed to the Step Generator the results of the interaction process will be an input for the process of generating new steps or updating an existing one in the steps tree as describe in algorithm 1. The component may have more than one interaction page. It depends on the type of the component, for examples if the selected component is a class, it needs three interaction pages; set properties, get properties and the data. Ultimately, when the data table is ready with the required variables, the process of generating code starts by matching the variables with values then masking this operation to get the new/updated code line. For example, printing “Hello World” will proceed as shown in the following tables: Variable Value Page1_Text1 "Hello World” Table 7: Sample of the variables and values The Code Mask Simple Script for printing text looks like the following: 60 <RPWI:NEWSTEP> Print Text – New Line – ( <T_V1> ) cout << <T_V1> << “\n” ; The component file include a matching table to match between interaction pages variables and Code Mask Variables. Interaction Page Variable Code Mask Variable Page1_Text1 T_V1 Table 8: Sample of the variables of Interaction Page Variable and Code Mask The Code Mask is written using simple rules which enables controlling and determining the generated code. Table 9 depicts these rules. Rule Description Put code mask variables between '<' and '>' Syntax: <variable name> Example: <Name> <RPWI:POSITIVE> This command used to ensure Generating the code between <RPWI:TEST> and <RPWI:ENDTEST> if the result of the test is true. <RPWI:NEGATIVE> This command used to ensure Generating the code between <RPWI:TEST> AND <RPWI:ENDTEST> if the result of the test is false. <RPWI:VALUE> Determine that value that will be compared with the value of the variable after <RPWI:TEST> , if the two values are typical , the test result is true, if not, the test result is false. Nested <RPWI:TEST> and <RPWI:ENDTEST> <RPWI:INFORMATION> Used to generate output in Information tab of the step details, in goal designer <RPWI:NOTE> or <*> This is not more than comment <RPWI:NEWSTEP> To create step in Goal Designer <RPWI:TEST> <Variable> more instruction <RPWI:ENDTEST> <RPWI:PUTMARK> NUMBER_FROM_1_TO_30 Save the ID of the current step (active step) to special area (Range 1-30) <RPWI:SETMARK> NUMBER_FROM_1_TO_30 Set a step as the active/parent using a saved ID (Range 1-30) <RPWI:IGNORELAST> <ANY CHARACTER> Used for deleting the last occurrence character from the generated code 61 <RPWI:NEWVAR> VARIABLE_NAME Used for creating new variable <RPWI:SETVARVALUE> VARIABLE_VALUE Used for setting the value of the variable (Active Variable) <RPWI:SELECTVAR> VARIABLE_NAME Set the active variable <RPWI:REPLACEVARSWITHVALUES> Replace variables with corresponding values Table 9: GCR Rules to write a Code Mask 4.4.3 PWCT’s System Layer This layer is a low-level layer that deals with the source code generated by the upper layer. PWCT generates and executes code in four text-based programming languages namely C#, Java, Harbour, and Python and hence, the System Layer had to support the different mechanisms in all four languages to produce executable code. Programs that were written by text-based languages are implemented either by an interpreter, a compiler or a combination of both [104]. While Python uses only an interpreter and C# uses a compiler, Scala uses both a compiler and an interpreter. . Consequently, the System Layer support compilers and interpreters at the same time. After generating the source code in any of the supported text-based language (C#, Java, Harbour and Python), executing the code follows the procedures of that specific language. PWCT doesn’t impose any restrictions on the process of building the source code. 62 Chapter 5: Implementation and Validation 5.1 Usability and Capability Evaluation To assess the usability and capability of PWCT, a free online remote learning course was offered. The goals of the course were to get feedback about the tool and to monitor the learning and progress of the participants. A total of 298 participants were accepted. For each registered participant, the team provided a folder where the participant’s work is stored. The course outline was very simple. At the beginning, participants were asked to download the tool and install it based on a short tutorial. Next, Participants were asked to watch videos that show how to develop different programs using PWCT. At first, the programs are very simple and later they become more advanced. After watching the video, the participants were asked to develop the program they just saw in the video. If the participant was successful, he/she was asked to send the files which were saved to the participant’s folder. Those who faced problems during the development of the project sent their questions to the project’s team and got the needed help. The lessons were classified into three levels of difficulty. The first level was an introduction to programming which contained 32 lessons. The second level was an introduction on how to develop database applications which contains 18 lessons. The third level was more advanced and introduced developing database applications using templates which contained 20 lessons. Table 14. In the Appendix B. shows some information about the all the participants such as the number of developed applications as well as the level of expertise of the participant. All of the developed applications can be found at URL: http://sourceforge.net/projects/doublesvsoop/files/PWCT%20For%20MSWindows/PWCT%20Students/ The main results of the course can be summarized as follows: 63 • All students completed their first application without problems. • 41 level-3 (advanced) students were able to develop real world database applications using PWCT. • 25 level-2 (intermediate) students were able to develop simple database applications using PWCT. • 232 level-1 (beginner) students were able to develop simple applications using PWCT. 46 level-1 students created only one application. From the progress of the students. The main results are presented in table 10 below. Ease-of-use Results Students that don’t like PWCT 15% (Students created only one application) Students like to learn PWCT 85% (Students created more than one application) Gained Knowledge Results Students learned programming basics only 78% (Level 1) Students learned basics of database 8% programming (Level 2) Students learned how to create professional 14% database applications (Level 3) Average Learning Curve (Time required for watching the Course) Time required to learn programming basics 4 hours, 41 minutes and 42 according to the course (Level 1) seconds Time required to learn basics of database 2 hours, 39 minutes and 39 programming according to the course (Level seconds 2) Time required to learn how to create 4 hours, 37 minutes and 22 professional database applications according seconds to the course (Level 3) Table 10: results from students 64 It is very clear from the data above that the majority of the participants liked the tool. 78% of the participants were able to learn the basics of programming by developing few applications only. And finally the average learning curve was extremely short even for the advanced lessons. 5.2 Efficiency Evaluation To make sure that PWCT’s performance was comparable to that of other VPL’s and even textual languages, a programming experiment was carried out. In this experiment PWCT competed against two well-known general-purpose VPLs (Tersus and Limnor), and two reputable VPE (Visual Studio and Net Beans). In the evaluation One professional programmer for each language was hired to develop programs in that specific language. All programmers were asked to develop 20 simple to intermediate applications. The authors monitored the time and memory needed to complete each task. Table 7 shows the averages of all tasks. It is extremely obvious that PWCT’s outperformed the other four languages by a considerable margin. Table 11. Shows that PWCT’s time was 8.5% less than Visual Studio (the second best) and 55.6% better than Net Beans. This criterion is very important as it hints at the productivity of the language. Similarly, while PWCT used only 4.5 MB of RAM on average, the second best (Tersus) used 12.1 MB. Figure 16 shows that PWCT outperforms all other languages as far as time and memory are concerned. Criteria\ Alt. General Purpose Visual Programming Languages PWCT Limnor Tersus Visual Programming Environments Visual Netbeans Studio Average time required to 106.66 144 140 115.66 165 create and test the seconds seconds seconds Seconds seconds application The steps to create and test 41 46 26 54 steps 33 steps the application steps steps Steps Average time required for 2.6 2.66 3.04 4.45 5 each step seconds seconds seconds seconds seconds 65 Average Time spent using 106.66 13 18 31.1 keyboard in user steps seconds Seconds seconds seconds Memory used before 13356 13460 50268 51276 application design starts Kbyte Kbyte Kbyte Kbyte Memory used after the 17880 29292 62392 109420 application design Kbyte Kbyte Kbyte Kbyte Memory consumed in 4524 15832 12124 58144 application representation Kbyte Kbyte Kbyte Kbyte during design time 30 seconds 156052 Kbyte 173352 Kbyte 17300 Kbyte Table 11: The average results of a simple program developed three programmers with three different levels Comparison between PWCT and other VPLs and VPEs (Time in seconds and Memory in MB and number of User steps ) 165 144 140 115.66 106.66 Time Memory 41 54 46 26 15.8 4.524 PWCT 12.124 Limnor Steps 58.144 Tersus Visual Studio 33 17.3 NetBeans Figure 16: Time in seconds and Memory in MB and Number of user steps To make sure that PWCT outperforms Visual Studio even on large applications, the authors carried out a second experiment where two programmers were used to develop one large application in both PWCT and VS (The application mange projects flow in the Riyadh Valley Company Occupation). One programmer is used to develop in PWCT and the other one in VS. Table 12 below illustrates the results of all programmers. 66 Visual Studio (CS + ASP.NET) 20606 lines 102 files 32 forms PWCT Lines of source code 20477 lines No of source files 96 files No of forms 26 forms No of Generated steps 7616 NA (Visual Source) No of Interaction 4723 NA Processes Time required for writing source code line 2 * 1.12(Average in visual studio or doing time for user step) 2.07 seconds Interaction process in = 2.24 seconds PWCT Pure 2.24 * 4723 = 2.07 * 20606 = Coding/Development 10579.52 seconds 42654.42 seconds Time - Effort Time = 2.94 hours = 11.85 hours only Table 12: The results of large scale developed by two professional developers Application Source Size 25000 20477 20000 20606 15000 10000 7616 5000 4723 0 PWCT Interaction Processes PWCT PWCT Visual Studio generated Steps generated LOC LOC Figure 17: The differences of Application source size between PWCT and Visual Studio In Figure 17 visual studio LOC are 20606 lines, which have been written during the development time. They also represent the source code managed by the programmers during maintenance and any future development. On the other hand, PWCT LOC got almost the same 67 number 20477, and we expect this result as we developed the same application. Nevertheless, LOC in PWCT generated, managed, and maintained in the background. The generated steps that actually represent the application are 7616 which approximately near to one third of the LOC. Thus, source in PWCT is more efficient to read and maintain. PWCT and Visual Studio actual interaction time in hourse 14 12 10 8 Hours 6 4 2 0 VS PWCT Figure 18: PWCT and Visual Studio actual interaction time in hours Figure 18 depicts the actual average time consumed in the interaction with both VS and PWCT. It is obvious that the average time to develop the application with PWCT was much less than that of VS. in fact, only 25.38% of the time was needed to develop in PWCT. 5.3 Reputation Evaluation PWCT is very popular on Sourceforge. Out 300,000 projects on Sourceforge, PWCT rank is 3 in the science and engineering category as of April 2016. The number of weekly downloads (environment, samples and tutorials) according to Sourceforge is more than 70,000 downloads. 68 Weekly Download Regarding Different Countries 2016-04-03 to 2016-04-09 United States United States 16% Italy… Other 45% Indi… United Kingdom 4% France 4% Indonesia Germany 3% South Africa 3% Algeria Canada 3% Egypt 3% 3% Netherlands 3% 3% Italy India United Kingdom France Germany Algeria Canada Netherlands South Africa Indonesia Figure 19: Weekly downloads according to www.sourceforge.net According to Sourceforge’s information which is depicted in figure 19, USA has the highest download rate of all countries followed by the Italy, India, United Kingdom, France, Germany, Algeria and Canada in the 8th place. Moreover, according to Sourceforge’s user’s satisfaction information, 93% of all users are satisfied with PWCT while only 7% are not as illustrated in Figure 20. User Satisfaction according to Sourceforge statistic 7% like Unlike 93% Figure 20: User satisfaction according to www.sourceforge.net 69 Many applications were developed using PWCT in several domains, such as: multi-media, database, network, business management systems, and Mathematics (for more information please see http://doublesvsoop.sourceforge.net/pwcthelp/main.htm). Figure 21 shows two applications made by PWCT’s users. Both applications are related to multimedia. The first one is a movie player and the second is an audio player. (a) (b) (c) (d) (e) (f) Figure 21: Multimedia applications screen snapshots developed using PWCT Using PWCT, one can create a console application and this feature is not available in the other VPLs even the general-purpose ones as shown in the first screen snapshot in Figure 21. The second snapshot depicts a famous puzzle game application, which was built by one of the users of PWCT. The third application shows a computer-interfacing application to send SMS message using an SMS Gateway connected to the computer through serial communication. The fourth application is chatting software that uses TCP/IP protocol for establishing a connection between 70 two nodes, one works as client and the other works as server. Figure 22: RVC Business applications screen snapshots developed using PWCT PWCT has not ignored the business management systems, where a group of programmers in Riyadh Valley Company, built a business application (Projects Management Application) for the company to follow-up its projects and monitor the performance of its divisions as showing in the dashboard in Figure 22. The screen snapshot in Figure 23. Depicts a multi-user client-server database application for managing car-hiring companies. This software achieved by one of the active users of PWCT in Netherlands. Figure 23: Multi-user client-server database application developed using GCR method 71 The screen snapshot in Figure 24. Depicts a multi-user client-server database application for a factory to manage orders data and their progress (Production Follow-up Application). The application include items data and their amounts in the store. The application is designed based on the requirements of the Al-Jabreen Factory in (Riyadh). Figure 24: Production Follow-up Application developed using the GCR method. The screen snapshot in Figure 25. Depicts a multi-user client-server database application for a training center (Courses Management Application) includes data store & analysis of (Courses, Trainers, Students, Courses review by students and Certifications). The application is designed based on requirements of Izdehar Training Center in (Jeddah). 72 Figure 25: Courses Management Application developed using the GCR method. Each application (Projects Management Application, Production Follow-up Application and Courses Management Application) can work on one machine (for testing) and can be deployed in a Client-Server environment for production. Applications comes with a control panel to define users and set permissions. Also applications support changing the database management system (DBMS) and tests are done using MS-Access, Microsoft SQL Server and MySQL. The applications are real world application based on real requirements of the company that will uses the application and during the design and the development we used the Object-Oriented Programming (OOP) paradigm where each application contains a group classes and components. 73 The screen snapshot in Figure 26. Depicts an Implementation of a new Localized Algorithm for detection of Critical Nodes using PWCT. Figure 26: Critical Nodes Detection Application using the GCR method. The screen snapshot in Figure 27. Depicts an Implementation of Natural Programming Language called Supernova using PWCT based on the Harbour Language through the HarbourPWCT visual programming language. The screen snapshot in Figures 28, 29 and 30. Depicts an Implementation of Scripting Language called Ring using PWCT based on the C Language through the CPWCT visual programming language. 74 Figure 27: The Supernova programming language developed using the GCR method. Figure 28: The Ring programming language developed using the GCR method. 75 Figure 29: The Ring Virtual Machine developed using the GCR method. Figure 30: The Ring Lists Implementation developed using the GCR method. 76 Chapter 6: Summary and Conclusion In this thesis we talked about the progress in the visual programming languages, the success in different programming domain, and the features of the most popular visual programming languages. According to our study we see a gap in attracting mainstream programmers to using visual programming languages in each new programming task. The current VPLs are not equivalent to the current textual programming languages at the practical side. Yes, some limitations are a result of missing features that can be added but this doesn’t count only for “what we can do?” but for “how we can do things with respect to the productivity of the programmer during software development and the quality of the final software”. We provided the requirements of a general purpose visual programming tool that can be used as replacement for the usage of popular programming languages like ( C,C++, C# and Java) + (visual programming environments like Visual Studio and Netbeans). We expect that after the implementation and marketing of this new tool, many new software projects can be developed using Visual Programming Language in less time with high quality. Also we expect attracting more users to real programming industry after the existence of this tool. Also in this thesis, PWCT was presented. PWCT is a novel general purpose VPL that was designed to compete with textual programming languages such as C++ and Java. The novelty of this work comes from the invention of a technique called GCR. GCR uses graphical components to replace textual code in an easy and seamless process. PWCT’s architecture, design, and implementation were covered in this thesis. Three types of evaluations were used to make sure that PWCT is a competitive VPL. Usability and capability evaluation shows that PWCT is extremely easy to use and requires a very short time for learning. Moreover, Efficiency evaluations show that PWCT requires 77 less time and memory to develop applications than the other used languages (VS, NB, Tersus and Limnor). Finally, data from Sourceforge indicates that PWCT is a popular, well accepted, and highly downloadable VPL. The development of many business applications and some programming languages like Supernova and Ring using PWCT demonstrates that the new VPL can be used in serious development for creating real world projects. PWCT include many features such as generating code in multiple languages and providing record/replay functionality. Also PWCT allow the users to create their own VPLs based on GCR. In the future, we will expand PWCT to support more platforms, web and mobile development. 6.2 Future Direction A more advanced version of PWCT is under development. It will support more platforms and will provide code generation in more textual languages. The components will be redesigned to provide more productivity. An interesting feature will be the support of our new textual programming language that we designed for the development of PWCT 2.0. This language called Ring. Using the Ring language we will have behind the Visual Programming Language, a scripting language that uses Natural Language Programming and Declarative Programming Paradigms. This will provide high-level of abstraction when the developer decide to switch from visual programming and look at the generated source code. Also the other textual programming languages will still be supported when the developer select them. 78 Bibliography [1] Takao Shimomura, Visual design and programming for Web applications, Journal of Visual Languages and Computing 16 (2005): p. 213-230. [2] Melisa Koorsse, Charmain Cilliers, Andre Calitz, Programming assistance tools to support the learning of IT programming in South African secondary schools, Computers & Education 82 (2015): p. 162-178. [3] Syahanim Mohd Salleh, Zarina Shukur, Hairulliza Mohamad Judi, Analysis of Research in Programming Teaching Tools: An Initial Review, Procedia – Social and Behavioral Science 103 (2013): p. 127-135. [4] G. Costagliola et al, Assessing the usability of a visual tool for the definition of e-learning processes, Journal of Visual Languages and Computing 19 (2008): p. 721-737. [5] Dimitar Asenov, Design and implementation of Envision - a visual programming system , Master Thesis, Chair of Programming Methodology ETH Zurich, (2011). [6] T. R. G. Green & M. Petre, Usability Analysis of Visual Programming Environments: a ‘cognitive dimensions’ framework, Journal of Visual Languages and Computing (1996): p. 131-174. [7] P. Bottoni et all, Dimensions of Visual Interaction Design. Visual Languages Proceedings, IEEE Symposium on date 13-16 sept. (1999). [8] K. Marriott & B. Myers (Eds.), Visual Language Theory, Springer-Verlag Publishers, Berlin, 1997, Chapter 3, Towards a Human Computer Interaction Perspective, p. 85-127. [9] Clement Leung, Visual Information Systems. Journal of Visual Languages and Computing 8 (1997), p. 261-263. [10] Levent V. Orman, Visual development of database applications. Interacting with Computers Vol 3 no 3 (1991), p. 302-318. 79 [11] T. Software, Tiobe programming community index, http://www.tiobe.com/tpci.htm, (2014). [12] P. Ponce, D. Ramirez, Intelligent Control Systems With LabView™ , Springer, New York (2010). [13] Jeremy Garrett, Thomas Walker, Student attitudes towards the use of graphical programming languages, Proceedings of the 2008 Annual ASEE Conference , Pittsburgh , PA, (2008) : p. 60-69. [14] Jean Bresson, Carlos Agon, Gérard Assayag,Visual Lisp/CLOS programming in OpenMusic, Higher-Order Symb Comput 22 (2009) : p. 81–111. [15] Caitlin Kelleher, Motivating Programming: using storytelling to make computer programming attractive to middle school girls, PhD thesis, in School of Computer Science, Carnegie Mellon University, CMU-CS-06-171, (2006) [16] Mitchel Resnick, et al., Scratch: Programming for All. Communications of the ACM (CACM), New York, NY, USA, November (2009). 52 (11). [17] Andrés Monroy Hernández, A. and M. Resnick, FEATURE: Empowering kids to create and share programmable media, interactions, (2008). 15(2): p. 50-53. [18] DANIEL D. HILS, Visual Languages and Computing Survey: Data Flow VPLs, Journal of Visual Languages and Computing, 3 (1992) : p. 69-101. [19] Brad Myers, Taxonomies of Visual Programming and Program Visualization, Journal of Visual Languages and Computing, (1990), 1: p. 97-123. [20] Burnett, M.M., et al., Scaling up VPLs. Computer, (1995). 28(3): p. 45-54. 80 [21] Caitlin Kelleher and Randy Pausch, Lessons Learned from Designing a Programming System to Support Middle School Girls Creating Animated Stories. in Visual Languages and Human-Centric Computing, 2006. VL/HCC 2006. IEEE Symposium on. (2006). [22] Ting-Chung Wang, et al. Teaching programming concepts to high school students with Alice, 39th ASEE/IEEE Frontiers in Education Conference. (2009). [23] Brendan Tangney, et al., Pedagogy and Processes for a Computer Programming Outreach Workshop - The Bridge to College Model, IEEE Transactions on Education, (2010). 53(1): p. 53-60. [24] Stuart Garner, Learning to Program from Scratch, 2009 Ninth IEEE International Conference on Advanced Learning Technologies, (2009) [25] Stolee, K.T., Kodu Language and http://research.microsoft.com/en-us/projects/kodu/, (2010). Grammar Specifcation. [26] Longflow, Limnor Tutorial http://www.limnor.com/. Longflow Enterprises Ltd., (2003). [27] Margaret M. Burnett, Visual Programming, Encyclopedia of Electrical and Electronics Engineering, New York, (1999). [28] Margaret M. Burnett and Marla J. Baker, A Classification System for VPLs, Journal of Visual Languages and Computing, (1994): p. 287-300. [29] James D. Kiper, Elizabeth Howard and Chuck Ames, Criteria for Evalutation of VPLs, Journal of Visual Languages and Computing (1997) , p. 175-192. [30] Gennaro Costagliola, Vincenzo Deufemia,Giuseppe Polese and Michele Risi, Building syntax-aware editors for visual languages, Journal of Visual Languages and Computing, (2005) : p: 508-540. [31] Luis A. Ortiz Ortiz, A Framework For VPLs, Master thesis in Electrical Engineering, University of PUERTO RICO MAYAGUEZ CAMPUS, (1997). 81 [32] K. N. Whitly, VPLs and the Empirical Evidence for and against, Journal of Visual Languages and Computing, (1997): p.109-142 [33] Stefano Levialdi, Visual Languages: Where Do We Stand? ,Proceedings of the International Workshop on Applications of Graph Transformations with Industrial Relevance ,Springer-Verlag Berlin Heidelberg (2000) : p145-164 [34] RAQUEL NAVARRO-PRIETO, AND JOSE J. CAN3 AS, Are VPLs better? The role of imagery in program comprehension, Int. J. Human-Computer Studies (2001): p 799-829 [35] Stefano Levialdi, Visual Languages: Concepts, Constructs and Claims, Information Technology Interfaces, (2001). [36] K. N. Whitley and Alan F. Backwell, Visual Programming in the Wild: A Survery of LabVIEW Programmers, Journal of Visual Languages and Computing (2001): p. 435-472. [37] Margaret Burnett, John Atwood, Rebecca Walpole Djang, Herkimer Gottfried, James Reichwein and Sherry Yang, Forms/3: A First-Order Visual Language to Explore the Boundaries of the Spreadsheet paradigm, Journal of Functional Programming (2001). [38] S. Matwin and T. Pietrzykowski, PROGRAPH: A PRELIMINARY REPORT, Comput. Lang. Vol. 10. No. 2, (1985): pp. 91-126 [39] Elif Tanrikulu, Bjorn Christian Schaefer, The users who touched the ceiling of scratch, Procedia – Social and Behavioral Sciences 28 (2011): p. 764-769. [40] N. Kehtarnavaz and C. Gope, DSP System Desing using LabVIEW and Simulink: A Comparative Evaluation, IEEE, (2006). [41] https://en.wikipedia.org/wiki/Deutsch_limit - last access 2016.03.26 [42] Daniel C. Cliburn, Student Opinions of Alice in CS1,38th ASEE/IEEE Frontiers in Education Conference, (2008). [43] Elizabeth V. Howard et all, A Qualitative Look at Alice and Pair-Programming, Information Systems Education Journal, Vol. 7 (80), (2009). [44] Linda Werner, Shannon Cample and Jill Denner, Children Learning Computer Science Concepts via Alice Game Programming, Proceedings of the 43 rd ACM Technical Symposium on Computer Science Education, SIGCSE, Raleigh, N. Carolina, USA, (2012). 82 [45] Python library for reading/writing MIT's Scratch file format, Retrieved Dec. (2015) from https://github.com/tjvr/kurt [46] Open source version of the Scratch 2.0 project editor, Retrieved Dec. (2015) from https://github.com/LLK/scratch-flash [47] Scratch Extensions, Retrieved Dec. (2015) from http://wiki.scratch.mit.edu/wiki/Scratch_Extension [48] Scratch Help, Retrieved Dec. (2015) from https://scratch.mit.edu/help/ [49] Discuss Scratch, Retrieved Dec. (2015) from https://scratch.mit.edu/discuss/ [50] The Alice Project, Retrieved Dec. (2015) from http://www.alice.org/index.php [51] Alice Publications, Retrieved Dec. (2015) from http://www.alice.org/index.php?page=publications/publications [52] Alice Software Wikipedia, Retrieved Dec. (2015) from https://en.wikipedia.org/wiki/Alice_%28software%29 [53] Alice 3 EULA, Retrieved Dec. (2015) from http://alice3.pbworks.com/w/page/28830524/Alice%203%20EULA [54] Download Alice 3, Retrieved Dec. (2015) from http://www.alice.org/index.php?page=downloads/download_alice3.1 [55] How to Guide for Alice 3, Retrieved Dec. (2015) from http://www.alice.org/3.1/materials_guide.php [56] Alice Help, Retrieved Dec. (2015) from http://alice3.pbworks.com/w/page/1756703/Alice%20help [57] Alice Community Forums, Retrieved Dec. (2015) from http://www.alice.org/community/ [58] Alice FAQ, Retrieved Dec. (2015) from http://alice3.pbworks.com/w/page/59015092/Alice%203_1%20FAQ 83 [59] Martin, Graphical programming for Mac OS X, Retrieved Dec. (2015) from http://www.andescotia.com/products/marten/ [60] Martin Video Lessons, Retrieved Dec. (2015) from http://www.andescotia.com/lessons/ [61] Martin Tutorials, Retrieved Dec. (2015) from http://www.andescotia.com/tutorials/ [62] Visual Programming With Prograph Book, Retrieved Dec. (2015) from www.andescotia.com/pdfs/Visual_Programming_With_Prograph_CPX_PDFs.zip [63] Applications written in Prograph using the Martin IDE, Retrieved Dec. (2015) from http://www.andescotia.com/examples/ [64] LabVIEW, use the right approach to solve any problem, Retrieved Dec. (2015) from http://www.ni.com/labview/why/programming-approaches/ [65] LabVIEW System Design Software, Retrieved Dec. (2015) from http://www.ni.com/labview/ [66] LabVIEW performance and memory management, Retrieved Dec. (2015) from https://decibel.ni.com/content/docs/DOC-15425 [67] LabVIEW Security, Retrieved Dec. (2015) from http://www.ni.com/support/security/ [68] LabView Powerful Multithreaded Execution, Retrieved Dec. (2015) from http://www.ni.com/labview/why/multicore/ [69] CH interpreter Wikipedia, Retrieved Dec. (2015) from https://en.wikipedia.org/wiki/Ch_%28computer_programming%29 [70] LabVIEW Tookkits, Retrieved Dec. (2015) from http://zone.ni.com/reference/enXX/help/371361M-01/lvhelp/labview_toolkits/ [71] LabVIEW 2015 Help, Retrieved Dec. (2015) from http://digital.ni.com/manuals.nsf/websearch/23F07E74C5B96E7286257E840050FD9E 84 [72] LabVIEW Technical Support, Retrieved Dec. (2015) from http://www.ni.com/support/ [73] LabVIEW Forums, Retrieved Dec. (2015) from http://forums.ni.com/ [74] Forms/3 License, Retrieved Dec. (2015) from http://web.engr.oregonstate.edu/~burnett/Forms3/getForms3.license.html [75] Forms3 Tour, Retrieved Dec. (2015) from http://web.engr.oregonstate.edu/~burnett/Forms3/Tour/tour.html [76] Limnor Download, Retrieved Dec. (2015) from http://limnor.com/LimnorDownload.html [77] Limnor Tutorials, Retrieved Dec. (2015) from http://www.limnor.com/LimnorTutorial.html [78] Limnor Forums, Retrieved Dec. (2015) from http://limnor.com/LimnorForum.html [79] Tersus Documentation, Retrieved Dec. (2015) from http://www.tersus.com/#Id=12 [80] Tersus Forums, Retrieved Dec. (2015) from http://www.tersus.com/#Id=22 [81] Retrieved Dec. (2015) from http://www.pm.inf.ethz.ch/research/envision.html [82] Envision Documentation Retrieved Dec. (2015) from http://dimitarasenov.github.io/Envision/ [83] Envision Source Code, Retrieved Dec. (2015) from https://github.com/dimitarasenov/Envision [84] Envision Videos, Retrieved Dec. (2015) from https://www.youtube.com/user/dimitarasenov/videos [85] Lava FAQ, Retrieved Dec. (2015) from http://lavape.sourceforge.net/doc/index.htm [86] Lava website, Retrieved Dec. (2015) from http://lavape.sourceforge.net/index.htm 85 [87] Worarit Kobsiripat, Effects of the media to promote the scratch programming capabilities creativity of elementary school students, Procedia – Social and Behavioral Sciences (2015) p. 227-232. [88] Eric J. Golin, Tool Review : Prograph 2.0 from TGS Systems, Journal of visual languages and computing, (1992) p. 189-194. [89] Sze Yee Lye and Joyce Hwee Ling Koh, Review on teaching and learning of computational thinking through programming: What is next for K-12? , Computers in Human Behavior (2014), p. 51-61. [90] Marat Boshernitsan and Michael Downes, Visual Programming Language: A Survery, Computer science division, University of California, Report No. UCB/CSD-04-1368, Dec. (2004). [91] Caitlin Kelleher and Randy Pausch, Lowering the Barriers to Programming: A taxonomy programming environments and languages for novice programmers, ACM Computing Surveys (CSUR). (2005). [92] Klaus D. Günther, Irmtraut Günther (GMD-SIT, Darmstadt), Lava - An ObjectOriented RAD Language Designed for Ease of Learning, Use, and Program comprehension , the proceedings of a workshop of Gesellschaft für Informatik, May (2001). [93] Ofer Brandes, Tersus visual programming platform, IEEE Symposium on Visual Languages and Human-Centric Computing (VL/HCC). (2011). [94] Shu, N.C. A visual programming language designed for automatic programming. in System Sciences, 1988. Vol.II. Software Track, Proceedings of the Twenty-First Annual Hawaii International Conference on. (1988). [95] Johnsgard, K. and J. McDonald. Using Alice in Overview Courses to Improve Success Rates in Programming I. in Software Engineering Education and Training, 2008. CSEET '08. IEEE 21st Conference on. (2008). [96] Kaucic, B. and T. Asic. Improving introductory programming with Scratch? in MIPRO, 2011 Proceedings of the 34th International Convention. (2011). 86 [97] MacLaurin, M.B., The design of kodu: a tiny visual programming language for children on the Xbox 360. SIGPLAN Not., (2011). 46(1): p. 241-246. [98] Avraham, L., WebRB: A Different Way to Write Web Applications, T.R. James, Editor (2008). p. 52-61. [99] Cebeci, Z. and Y. Erdoðan, Tree view editing learning object metadata. Interdisciplinary Journal of Knowledge and Learning Objects, (2005). 1(1): p. 99–108. [100] Szyperski, C., Component Software: Beyond Object-Oriented Programming1998, New York, N.Y.: ACM Press and Addison-Wesley. [101] Vitharana, P., Risks and challenges of component-based software development. Commun. ACM, (2003). 46(8): p. 67-72. [102] Crnkovic, I. and M. Larsson, Challenges of component-based development. Journal of Systems and Software, (2002). 61(3): p. 201-212. [103] Wikipedia, http://en.wikipedia.org/wiki/Middleware. Retrieved April. (2016) [104] Proebsting, T.A., et al., Toba: java for applications a way ahead of time (WAT) compiler, in Proceedings of the 3rd conference on USENIX Conference on Object-Oriented Technologies (COOTS) - Volume 31997, USENIX Association: Portland, Oregon. p. 3-3. [105] Andrew Begel, LogoBlocks: A Graphical Programming Language for Interacting with the World, Cambridge, MA: MIT Media Lab., (1996) [106] Sampada K. Satav, S. K. Satpathy, K, J. Satao, A Comparative Study and Critical Analysis of Various Integrated Development Environments of C, C++, and Java Languages for Optimum Development, UNLASCIT, Vol 1(1), (2011), p09-15 [107] PWCT Researchers and Experts, Retrieved Dec. (2015) from http://doublesvsoop.sourceforge.net/pwcthelp/team.htm. [108] Roger S. PRESSMAN, Bruce R. MAXIM, Software Engineering A Practitioner’s Approach, 8th Edition. ISBN-13: 978-0078022128 [109] Robert W. Sebesta, Concepts of programming languages (10th Edition), ISBN-13: 978-0131395312 87 Appendix A: Guided tour A.1 Hello World Example In this section, the concept of GCR is illustrated through a Hello World program using PWCT. The application development process starts from the Goal/Module Designer where the application contains one or more of visual source files and each visual source file contains one or more of goals and each goal contains one or more of steps; the step maybe a visual representation that present a computer instruction that do something or the step maybe a comment written by the programmer to describe the process. From the next figure we see that the active visual programming language is HarbourPWCT and the active goal name is “main” and this goal contains a tree of steps that includes Start Point (NOT STEP) which is not more than a visible root node for the steps tree. Inside the steps tree by default there are a step called (The First Step) which is a step of type (comment) written by the programmer to indicates that this is the first step in our program where we can add more steps the do something (computer instructions). 88 Figure 31: PWCT – Goal/Module Designer The next table describe the function of the most important buttons in the Goal/Module designer window Index 1 2 3 4 5 6 7 8 9 10 Table 13: Goal/Module Designer buttons and their description Button Icon Description Create new step of type comment – the step name is written by the programmer to describe something in the program Edit step name – the step name is modified by the programmer – if the step type is a comment then this feature is used to modify the comment but if the step type is a computer instruction then this feature is used to provide better description for the instruction. Delete step from the steps tree to remove a comment or a computer instruction. Move step up in the steps tree to organise the visual source and/or to control the flow of instructions execution during the runtime Move step down in the steps tree to organise the visual source and/or to control the flow of instructions execution during the runtime Start new interaction process to generate new steps inside the steps tree. These steps will do something during the runtime because it’s from the type (computer instruction) Modify a step. This features is available for steps of type computer instructions to modify something in the interaction process by changing attribute values Open the form designer. This feature is available for steps that are related to GUI programming (Windows, controls,...etc) Ignore/Enable a step. This feature is available for steps of type computer instructions to include/exclude the step from the designed program during runtime Increase the font size of the steps names in the steps tree 89 11 12 13 14 15 Decrease the font size of the steps names in the steps tree Cut a step, remove the step from the steps tree and keeping it in the memeory Copy a step to the memory without removing it from the steps tree Paste a step from the memory to be a child step to the selected step in the steps tree Search in the visual source by step name Toward creating hello world program, after selecting the step (The First Step) which is selected by default Click the interact button which carry the icon to start new interaction process. After clicking the interact button the components browser window will be opened The components are classified into different domains were each domain contains group of components and using the component browser you can select a component to use. From the domain tree select the domain (Print Text) then from the components list select the component (Print Text to Console) Then click the Ok button. 90 Figure 32: PWCT - Components Browser After selecting the component using the components browser and clicking Ok, The interaction using transporter window will be opened to enter the data required for the selected component. The transporter is another name of the component inside PWCT. The component (Print Text To Console) presents simple interaction page (data entry form) that ask the used to enter the text that will be printed on the screen. In our example we will write the text “Hello World”. Then we will click Ok button. 91 Figure 33: PWCT – Interaction Using Transporter window After entering the text and pressing Ok button we will return again to the goal designer window but this time will notice that a new step is added to the steps tree. And the name of the new step is Print text (“Hello World”) to console. The new step name can be changed by the programmer using the edit step button At this stage we can run our program. But running the program at this stage will not give us the chance to see the message because the program will be executed and the execution will done and the console screen will disappear before seeing the message. To see the message we can run the program from the windows command line window or we can add another step to the program to wait for a number of seconds to see the results. To modify the program by adding a step for waiting for number of seconds we will do another interaction process with another component but this time we will control the interaction process using the keyboard instead of using the mouse to clear that the Coding Simulation Method is designed to give us the ability to use the keyboard to get more productivity. 92 Figure 34: PWCT – Goal/Module Designer – Generated step as a result of interaction process To use the keyboard to start new interaction process we can press CTRL+T or we can write the first letter of the component name while we still in the Goal Designer and the first step is selected. In our case we will press the letter “W” because the required component name is (Wait Key/Seconds) After pressing the letter “W” the components browser window will be opened and this letter will be available in the search line where we can continue the search process by typing more letters. After typing the second letter in the component name, in our case it’s the letter “A” we will see that the required component (Wait Key/Seconds) is selected. 93 Figure 35: PWCT - Components Browser – Using Keyboard to quickly find a component Instead of clicking the ok button we will press Enter to use the selected component. 94 Figure 36: PWCT – Interaction using transporter window Inside the interaction using transporter window and using arrows we can move to the checkbox (Wait nSeconds) and this will be done after pressing the down arrow for two times. Then we will press enter to set the checkbox (Wait nSeconds) to On. Then press enter to move the textbox to determine the number of seconds In our example we will type 3. Then press CTRL+W to end the interaction process to return to the Goal Designer. 95 Figure 37: PWCT – Goal/Module Designer – Generated steps as a result of interaction process. After returning to the Goal designer we notice a new step added to the steps tree and this step carry the name Wait (3 Seconds). Now we can run the program by pressing CTRL+R or by click the (i) button from the toolbar. Figure 38: Application during the runtime 96 A.2 How to maintain programs Using the Goal/Module designer buttons we can add, edit and delete steps And using the Modify button after selecting a step the interaction using transporter window will be opened where we can change the values (entered data) and this change will be reflected in the step name inside the goal designer. 97 A.3 GUI Applications The GCR is flexible enough to represent different types of applications and GUI applications can be developed and represented simply using the steps tree and interaction pages. Also we can use a form designer to set controls positions and sizes The next figure presents the steps tree of a simple GUI application This applications contains a window carry the title (Welcome to PWCT) and this window contains a label carry a caption the same as the window title. Also the window contains a button to close the window. When the button is clicked by the user a procedure called (myclose) will be invoked and this procedure will close the window. All of these steps are a result of interaction processes with visual components like (Define new window, Label, Button, Button Events, Define Procedure and Window Class) Figure 39: PWCT – Goal/Module Designer – Steps Tree for simple GUI application 98 Figure 40: PWCT – Form Designer Figure 41: PWCT – Components Browser – Components inside the domain “Controls” 99 Figure 42: PWCT – Interaction using transporter window for the component “Define new window” The interaction page of the component (Define new window) is an example about designing a large component with many features. Note: the same component may contain a group of interaction pages like the (Window Class) component which contain three interaction pages (Set Property page, Get Property Page & Invoke method page). 100 Figure 43: PWCT – Interaction Page for Multi pages component 101 Figure 44: PWCT – Interaction Page for Multi pages component 102 A.4 Search in Visual Source Figure 45: PWCT – Search for a step inside the goal designer The programmer can do search operation in the visual source to quickly find and select a step. The search process can be done in the step name (goal designer) or in the step data (data entered through the interaction using transporter window). A.5 The Time Dimension One of the characteristics of visual programming is programming in more the one dimension. One of the available dimensions for PWCT users is the time dimension where the programmer can move forward/backward during the application construction process and the programmer can run the application at any point on the time dimension to test the programmer before or after adding a step to the steps tree. This can be done easily using a slider control available in the Goal/Module Designer. 103 Figure 46: PWCT – Using the Time Machine to go backward at a time in the past during the design process 104 A.6 See the code behind every step PWCT is a hybrid system where we can design programs visually without writing source code by hand but we still can see the generated textual source code which is generated and managed in the background. We can see the textual source code behind each step in the goal designer and we can get the source code behind all the steps when we run the application or using a utility comes with PWCT and called the Code Extractor application. Figure 47: PWCT – Looking at the code behind a step 105 Figure 48: PWCT – Code Extractor Application 106 A.7 Dealing with multiple steps at the same time PWCT Contains a tool called the Goal/Module Viewer which contains some of the features available through the goal designer but enable the programmer to work with more than one step at the same time where each node in the steps tree contains a checkbox for selecting the step. Using checkboxes more than one step can be selected then we can do different operations of the selected steps like (Cut, Copy, Insert, Delete , Move Up, Move Down and Enable/Disable). Figure 49: PWCT – Goal/Module Viewer 107 A.8 The Time Machine and playing programs as movies When you open any project/application developed using PWCT and instead of looking at the design to see how to create like this program, simply you can see a movie that teach you how to design the program using PWCT step by step To do that from the Goal/Module designer press The Time Machine button then select the item (Play as Movie from the first time frame). Then the steps will disappear from the Goal/Module designer and you will see how every step will be added by selecting the required component then entering the data to the interaction pages. This feature reduces the need for sample documentation because every sample developed using PWCT is a movie that can be played. Also this feature help in program understanding where the program can use this feature and keep watching instead of manually moving through the visual representation Figure 50: PWCT – Goal/Module Designer - Using the Time Machine 108 Appendix B: Students Progress In Learning PWCT In Table 14 we see the students’ progress in learning PWCT through the free remote learning course. The first column in the table represent the student ID (instead of the student name) and the second column contains the number of applications made by the student using PWCT. The third column contains the student level. In figure 50 we see the graph that represent the number of students and the student level and in figures 51, 52 and 53 we see the statistics of the YouTube Playlists for each level in the course. Table 14: Students progress in learning PWCT Student ID 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 No. of Applications 1 7 47 29 2 4 64 24 19 4 6 18 2 1 7 9 4 7 1 42 25 10 46 3 19 4 1 Level 1 1 2 1 1 1 3 1 1 1 1 1 1 1 1 1 1 1 1 2 1 1 2 1 1 1 1 109 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 1 11 51 66 3 5 6 62 64 13 18 35 64 6 3 1 16 7 4 5 27 1 11 3 3 7 34 2 59 6 4 64 5 9 8 1 1 5 11 60 1 43 1 1 3 3 1 1 1 3 3 1 1 2 3 1 1 1 1 1 1 1 1 1 1 1 1 1 2 1 3 1 1 3 1 1 1 1 1 1 1 3 1 2 110 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 61 46 1 24 4 6 7 38 20 5 5 1 60 4 5 45 4 5 1 53 8 1 1 61 7 10 2 1 61 1 65 4 9 64 4 1 3 5 65 23 2 3 3 2 1 1 1 1 1 2 1 1 1 1 3 1 1 2 1 1 1 3 1 1 1 3 1 1 1 1 3 1 3 1 1 3 1 1 1 1 3 1 1 1 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 62 17 64 62 65 49 70 9 4 5 4 27 6 49 22 23 26 48 22 2 1 68 2 5 70 8 10 1 4 4 63 2 4 3 4 4 12 15 47 2 3 64 3 1 3 3 3 2 3 1 1 1 1 1 1 2 1 1 1 2 1 1 1 3 1 1 3 1 1 1 1 1 3 1 1 1 1 1 1 1 2 1 1 3 112 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 41 1 4 8 4 70 7 2 30 65 65 66 20 8 7 4 7 5 6 1 26 7 12 1 24 64 6 2 2 1 6 70 65 1 6 1 1 70 1 6 1 6 2 1 1 1 1 3 1 1 1 3 3 3 1 1 1 1 1 1 1 1 1 1 1 1 1 3 1 1 1 1 1 3 3 1 1 1 1 3 1 1 1 1 113 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 4 65 3 56 1 53 1 3 46 7 67 5 2 4 2 70 4 3 26 37 1 18 19 32 35 20 14 36 44 36 18 6 3 17 14 1 2 36 36 6 28 18 1 3 1 3 1 3 1 1 2 1 3 1 1 1 1 3 1 1 1 2 1 1 1 1 2 1 1 2 2 2 1 1 1 1 1 1 1 2 2 1 1 1 114 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 3 2 18 11 18 1 12 6 31 25 3 2 49 1 9 7 19 11 2 4 1 2 17 53 51 1 2 5 8 2 1 3 1 46 2 1 1 8 30 1 2 12 1 1 1 1 1 1 1 1 1 1 1 1 2 1 1 1 1 1 1 1 1 1 1 3 3 1 1 1 1 1 1 1 1 2 1 1 1 1 1 1 1 1 115 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 25 15 9 1 1 2 19 1 6 11 6 1 22 35 8 24 6 2 2 1 1 1 1 1 1 1 1 1 1 1 1 1 2 1 1 1 1 1 Students Level 3.5 3 2.5 2 1.5 1 0.5 1 12 23 34 45 56 67 78 89 100 111 122 133 144 155 166 177 188 199 210 221 232 243 254 265 276 287 298 0 Figure 51: Students Level in PWCT course. 116 Figure 52: YouTube Playlist Statistics for Level 1 of PWCT Course. Figure 53: YouTube Playlist Statistics for Level 2 of PWCT Course. Figure 54: YouTube Playlist Statistics for Level 3 of PWCT Course. 117 Appendix C: Designers in our Visual Programming Languages’ framework (1) Interaction Designer (Component User Interface Designer) Figure 55: Interaction Designer 118 (2) Transporter Designer (Component Designer) Figure 56: Transporter/Component Designer – Interaction Pages Figure 57: Transporter/Component Designer – Code Mask (Script) 119 Figure 58: Transporter/Component Designer – Matching Figure 59: Transporter/Component Designer – Rules 120 

6cs.PL

Joint Mixability of Elliptical Distributions and arXiv:1706.05499v2 [math.ST] 18 Oct 2017 Related Families Chuancun Yin ∗ School of Statistics, Qufu Normal University Shandong 273165, China e-mail: ccyin@mail.qfnu.edu.cn Dan Zhu School of Statistics, Qufu Normal University Shandong 273165, China October 19, 2017 Abstract In this paper, we further develop the theory of complete mixability and joint mixability for some distribution families. We generalize a result of Rüschendorf and Uckelmann (2002) related to complete mixability of continuous distribution function having a symmetric and unimodal density. Two different proofs to a result of Wang and Wang (2016) which related to the joint mixability of elliptical distributions with the same characteristic generator are present. We solve the Open Problem 7 in Wang (2015) by constructing a bimodal-symmetric distribution. The joint mixability of slash-elliptical distributions and skew-elliptical distributions is studied and the extension to multivariate distributions is also investigated. Keywords: Complete mixability, joint mixability, multivariate dependence, slash/skewelliptical distributions, elliptical distributions AMS subject classifications: 60E05 60E10 60E15 ∗ Corresponding author. 1 1 Introduction In recent years, the problem of studying complete mixability and joint mixability of distributions has received considerable attention. Complete mixability and joint mixability describe whether it is possible to generate random variables (or vectors) from given distributions with constant sum. The formally definition of complete mixability for a distribution was first introduced in Wang and Wang (2011) and then extended to an arbitrary set of distributions in Wang, Peng and Yang (2013), although the concept has been used in variance reduction problems earlier (see Gaffke and Rüschendorf (1981), Knott and Smith (2006), Rüschendorf and Uckelmann (2002)). The properties are particularly of interest in quantitative risk management and optimization problems in the theory of optimal couplings, where dependence between risks is usually unknown or partially und known. Throughout the paper, we write X = Y if the random variables (or vectors) X and Y have the same distribution. For a cumulative distributions function F , we write X ∼ F to denote F (x) = P (X ≤ x). By convention, all vectors will be written in bold and will be considered as column vectors, with the superscript ⊤ for transposition. Next we introduce the concepts of completely mixable and jointly mixable distributions. Definition 1.2 (Wang, Peng and Yang (2013)). Suppose n is a positive integer The univariate distribution functions F1 , · · · , Fn are jointly mixable (JM) if there exist n random variables X1 , · · · , Xn with distribution functions F1 , · · · , Fn , respectively, such that P (X1 + · · · + Xn = C) = 1, (1.1) for some C ∈ R. If (1.1) holds with Fj = F, 1 ≤ j ≤ n, the distribution F is said to be n-completely mixable (n-CM). Any such C is called a joint center of (F1 , · · · , Fn ). For a brief history of the concept of the complete mixability, we refer to the recent papers of Wang (2015) and Wang and Wang (2016). Existing results on complete mixability and joint mixability are summarized in Wang and Wang (2011), Puccetti, Wang and Wang (2012) and Wang and Wang (2016). As pointed out in Puccetti and Wang (2015b), as a full characterization of completely mixable distribution is still out of reach, there are even less results concerning sufficient conditions for joint mixable distributions. The only available ones are given in the recent paper of Wang and Wang (2016). The rest of the paper is organized as follows. In Section 2 we discuss the condi- 2 tions on a result of Rüschendorf and Uckelmann (2002) related to complete mixability of continuous distribution function having a symmetric and unimodal density. Section 3 is dedicated to joint mixability of elliptical distributions and slash/skew-elliptical distributions, respectively. Section 4 extended the result to the class of multivariate elliptical distributions. 2 Symmetric Distributions It would be of interest to characterize the class of completely mixable distributions. Only partial characterizations are known in the literature. One nice result for the complete mixability is given by Rüschendorf and Uckelmann (2002), which is equivalent to state that any continuous distribution function having a symmetric and unimodal density is nCM for any n ≥ 2. Wang (2014) provided a new proof using duality representation. The property was also extended to multivariate distributions by Rüschendorf and Uckelmann (2002). Wang and Wang (2016) based on a different technical approach to study the joint mixability of unimodal-symmetric densities. Lemma 2.1. (Rüschendorf and Uckelmann (2002)). Any continuous distribution function having a symmetric and unimodal density is n-CM for any n ≥ 2. Lemma 2.2. (Wang and Wang (2016)). Suppose that F1 , · · · , Fn are distributions with unimodal-symmetric densities from the same location-scale family with scale parameters θ1 , · · · , θn , respectively. Then F1 , · · · , Fn is JM if and only if the scale inequality n X i=1 is satisfied. θi ≥ 2 max θi (2.1) 1≤i≤n Proof Using Theorem 3.1 in Wang and Wang (2016), we can give a more simple proof d to “if part” of Lemma 2.2. In fact, Xi ∼ Fi can be written as Xi = θi RUi + µ, where µ is a constant, R is a random variable on (−∞, ∞) and Ui is uniformly distributed on (−1, 1) independent of R. The result follows from Theorem 3.1 in Wang and Wang (2016) since θi Ui ∼ U(−θi , θi ), i = 1, · · · , n. This ends the proof. Suppose that Y has a distribution F and that θ has some distribution on (0, ∞) with a distribution function H and, Y and θ are independent. Then the distribution of X = θY 3 is referred to as a scale mixture of F with a scale mixing distribution H. The following corollary is a direct consequence of Lemma 2.1. Corollary 2.1. The scale mixture of an unimodal-symmetric continuous distribution with center µ is n-CM (n ≥ 2) with center µ. In order to give the main results of this section, we first give a lemma which is closely related to the φ-complete mixability and φ-joint mixability for some supermodular functions φ; see Bignozzi and Puccetti (2015). Lemma 2.3. Suppose that F1 , · · · , Fn are n univariate continuous distribution functions, g is a nonconstant continuous function on Rn , f is a nonconstant continuous function on the real line. Then there exist n random variables Xi ∼ Fi and constant C such that P (f (g(X1, · · · , Xn ))) = f (C)) = 1 (2.2) P (g(X1, · · · , Xn ) = C) = 1. (2.3) if and only if Proof. The proof of “if part” is obvious. Now we prove the “only if part”. Since Fi ’s are continuous, for any n random variables Xi ∼ Fi (i = 1, 2, · · · , n), their joint distribution function is continuous, and thus g(X1 , · · · , Xn ) is either an almost surely constant random variable or a continuous random variable. So that f (g(X1, · · · , Xn )) is either an almost surely constant random variable or a continuous random variable. Whenever g(X1 , · · · , Xn ) is a continuous random variable, we have for any constant K, P (f (g(X1, · · · , Xn )) = K) = 0. Consequently, if (2.2) holds then it is necessary that (2.3) should hold, thus the result follows. Theorem 2.1. Suppose F is a continuous distribution with unimodal-symmetric density, f is a nonconstant continuous function on the real line. For any integer n ≥ 2, there exist n random variables Xi ∼ F (i = 1, 2, · · · , n) such that ! ! n X P f Xi = K = 1, (2.4) i=1 where K = f (C). Here C is the joint center of the n-tuple (F, · · · , F ). Moreover, if C is unique, then K is unique. 4 Proof. Taking g(X1 , · · · , Xn ) = Pn i=1 Xi and Fi = F, i = 1, 2, · · · in Lemma 2.3, the result follows from Lemma 2.1. This ends the proof. Similarly, by using Lemmas 2.2 and 2.3 we get the following theorem. Theorem 2.2. Suppose that F is a continuous distribution with unimodal-symmetric density, θ1 , · · · , θn are positive constants, f is a piecewise continuous function on the real line. Then there exist n random variables Xi ∼ F and constant K such that ! ! n X P f θi Xi = K = 1 (2.5) i=1 if and only if the scale inequality (2.1) is satisfied. The complete mixability and joint mixability is a concept of negative dependence (cf. Puccetti and Wang (2015a)) and not all univariate distributions F are n-CM. If the supports of Fi (i = 1, 2, · · · , n) are unbounded from one side, then (F1 , · · · , Fn ) is not JM for any n ≥ 1; see Remark 2.2 in Wang and Wang (2016). Now we list more examples (The proof learned largely from Ruodu Wang). Example 2.1. Assume F1 , · · · , F2n+1 are 2n + 1 univariate distribution functions with n symmetric densities on the same interval [−a, a] (a > 0), if Fi ( n+1 a) ≤ n+1 2n+1 (i = 1, 2, · · · , 2n + 1), then (F1 , · · · , F2n+1 ) is not JM. n Proof For any Xi ∼ Fi (i = 1, 2, · · · , 2n + 1), the conditions Fi ( n+1 a) ≤ 1, 2, · · · , 2n + 1) imply that  P |Xi | >  n 2n a > , i = 1, 2, · · · , 2n + 1. n+1 2n + 1 n+1 2n+1 It follows that    2n+1 X  na na na ≥ P |Xi | > − 2n , · · · , |X2n+1| > P |X1 | > n+1 n+1 n+1 i=1 > (2n + 1) Note that (2n+1 X i=1 Hence Xi 6= 0 ) ⊇  P na na |X1 | > , · · · , |X2n+1 | > n+1 n+1 2n+1 X i=1 Thus (F1 , · · · , F2n+1 ) is not JM. 2n − 2n = 0. 2n + 1 Xi 6= 0 5 ! > 0.  . (i = Corollary 2.2. (Necessary Condition) Assume F1 , · · · , F2n+1 are 2n + 1 univariate dis- tribution functions with symmetric densities on the same interval [−a, a] (a > 0), if (F1 , · · · , F2n+1 ) is JM, then there exists some i (1 ≤ i ≤ 2n + 1) such that   1 n a > . P |Xi | ≤ n+1 2n + 1 The interval [−a, a] in Example 2.1 can be changed as (−∞, ∞). Example 2.2. Assume F1 , · · · , F2n+1 are 2n + 1 univariate distribution functions with symmetric densities on the same interval (−∞, ∞), if there exists a > 0 such that   n n Fi (a) − Fi a ≥ , i = 1, 2, · · · , 2n + 1, n+1 2n + 1 then (F1 , · · · , F2n+1 ) is not JM. The following example tells us the symmetry of F does not implied F is 3-CM and that the unimodality assumption on the density can not removed. Example 2.3 Assume that F has the following bimodal symmetric density ( 2r+1 2r x , if x ∈ [−a, a], 2a2r+1 f (x) = 0, if x ∈ / [−a, a], where r is a positive integer. The distribution is given by   if x < −a,   0, 1 F (x) = (x2r+1 + a2r+1 ) , if − a ≤ x < a, 2a2r+1    1, if x ≥ a. It is easy to see that 1 1 2 + 2r+2 < . 2 2 2 3 Thus (F, F, F ) is not JM. Or, equivalently, F is not 3-CM. F a = Example 2.4 Assume that F has the following bimodal symmetric density ( 2m Cm √x1−x2 , if x ∈ (−1, 1), fm (x) = 0, if x ∈ / (−1, 1), where Cm is a normalizing constant and m ≥ 0 is an integer. When m = 0, fm is the density of Pearson type II distribution, which is convex, bimodal and symmetric, so 6 that Lemma 2.1 can not applicable. Note that f0 is the density of cos θ or sin θ with θ ∼ U(0, 2π) and thus it is n-CM for any integer n ≥ 2; see Puccetti, Wang and Wang (2012) for more details. When m ≥ 1, using Example 2.1 one can checks that F is not (2n + 1)-CM for not very large n. The (2n + 1)-complete mixability of this distribution is not covered by any known theoretical results for large n. Further we consider a distribution F with density f (x) = ∞ X αm fm (x), m=1 where {αm }m≥1 is a sequence of positive values with P∞ m=1 αm = 1. This distribution is not (2n+1)-completely mixable for any n ≥ 1. Thus we give a counterexample to the Open Problem 7 in Wang (2015) which states that are all absolutely continuous distributions on a bounded interval n-CM for large enough n? 3 Elliptical Distributions and Related Families 3.1 Elliptical Distributions Let Ψn be a class of functions ψ : [0, ∞) → R such that function ψ(|t|2 ), t ∈ Rn is an n- dimensional characteristic function. It is clear that Ψn ⊂ Ψn−1 · · · ⊂ Ψ1 . Denote by Ψ∞ the set of characteristic generators that generate an n-dimensional elliptical distribution for arbitrary n ≥ 1. That is Ψ∞ = ∩∞ n=1 Ψn . Defination 3.1 The random vector X = (X1 , X2 , · · · , Xn )⊤ is said to have an elliptical distribution with parameters µ and Σ, written as X ∼ En (µ, Σ, ψ), if its characteristic function can be expressed as

ϕX (t) = exp it⊤ µ ψ t⊤ Σt⊤ , t ∈ Rn , (3.1) for some column-vector µ, n × n positive semidefinite matrix Σ and for some function ψ ∈ Ψn with ψ(0) = 1, which is called the characteristic generator. In general, elliptical distributions can be bounded or unbounded, unimodal or multimodal. When ψ(u) = exp(−u/2), En (µ, Σ, ψ) is the normal distribution Nn (µ, Σ) and when n = 1 the class of elliptical distributions consists of the class of symmetric distributions. It is well known that an n-dimensional random vector X ∼ En (µ, Σ, ψ) if and only if for any vector 7 α ∈ Rn , one has (Cambanis et al. (1981)) α⊤ X ∼ E1 (α⊤ µ, α⊤ Σα, ψ). In particular, P ⊤ Xi ∼ E1 (µi , σi2 , ψ) and ni=1 Xi ∼ E1 (e⊤ n µ, en Σen , ψ). The next result is due to Wang and Wang (2016) in which they proved the existence of the dependence structure. Here we present two another proofs by finding the exact dependence structure and by using Lemma 2.2, respectively. Theorem 3.1. Suppose that Fi ∼ E1 (µi , σi2 , ψ), where µi ∈ R, σi > 0, ψ is a characteristic generator for an n-elliptical distribution. Then (F1 , · · · , Fn ) is JM if the inequality n X i=1 σi ≥ 2 max{σ1 , · · · , σn } (3.2) is satisfied. Proof (First proof). Assume X ∼ En (µ, Σ, ψ), where µ = (µ1 , · · · , µn )⊤ and Σ = (σij )n×n . Here σij = ( σi2 , if i = j, 1 (σk2 (n−1)(n−2) − P l6=k σl2 ), if k 6= i 6= j. It is straightforward to check that Σ is positive semidefinite under condition (3.2) and the summation of all entries in Σ is zero. Each component Xi of X has distribution P E1 (µi , σi2 , ψ), i = 1, 2, · · · , n. The characteristic function of ni=1 Xi can be expressed as ! ! n n X X ϕPni=1 Xi (t) = exp it µi ψ(0) = exp it µi , t ∈ R. (3.3) i=1 i=1 Hence, P n X i=1 and thus (F1 , · · · , Fn ) is JM. Xi = n X i=1 µi ! = 1, Second proof. We remark that if Fi ∼ E1 (µi , σi2 , ψ) has a density and ψ is a char- acteristic generator for an n-elliptical distribution (n ≥ 2), then Fi is unimodal and symmetric. Thus Theorem 3.1 is a direct consequence of Lemma 2.2. The following result is a direct consequence of Theorem 3.1. Corollary 3.1. Suppose that F ∼ E1 (µ, σ 2, ψ) with ψ ∈ Ψ∞ . Then F is n-CM for any n ≥ 2. 8 Remark 3.1. Theorem 2.21 in Fang, Kotz and Ng (1990) shows that ψ ∈ Ψ∞ if and only if F ∼ E1 (µ, σ 2 , ψ) is a mixture of normal distributions. Some such elliptical distributions are normal distribution, the T -distribution, Cauchy distribution, Stable laws distribution and Pearson type VII distribution; see Andrews and Mallows (1974) and Kano (1994). Note that there are continuous, unimodal and symmetric densities do not belong to the class of normal scale mixtures; see West (1987). Thus Corollary 3.1 is a special case of Corollary 2.1. In the sequel, we list more examples. Example 3.1 Consider the generalized logistic distribution with density f (x) = C exp(−αxβ ) , −∞ < x < ∞, (1 + exp(−xβ ))2α (3.4) where C > 0, α > 0, β > 0 are constants. If α = 1 and β = 2, (3.4) is 1-dimensional logistic distribution which is unimodal and symmetric but not a scale mixture of normal densities; see Gómez-Sánchez-Manzano et al. (2006). If α = β = 1, (3.4) is standard logistic distribution which is unimodal and symmetric and can be represented as a scale mixture of normal densities; see Stefanski (1990). Example 3.2 Kotz type distributions with density generator g(r) = Cr N −1 exp(−mr β ), m, β > 0, N > 1 have symmetric and bimodal densities, the (2n + 1)-complete mixability of those Kotz type distributions is not covered by any known theoretical results for n ≥ 1. 3.2 Slash-Elliptical Distributions In this subsection, we investigate joint mixability of slash-elliptical distributions. We say that a random variable X follows a slash elliptical distribution if it can be written as X= Z 1 Uq + µ, (3.1) where Z ∼ E1 (0, σ 2 , ψ) is independent of U ∼ U(0, 1) and q > 0 is the parameter related to the distribution kurtosis. We use the notation X ∼ SE1 (µ, σ 2 , ψ; q). Similarly, we say that a random vector X ∈ Rp has slash-elliptical multivariate distribution with vector 9 location parameter µ, positive semidefinite matrix scale parameter Σ, and tail parameter q > 0, if it can be represented as Z (3.2) 1 + µ, Uq where Z ∼ Ep (0, Σ, ψ) is independent of U ∼ U(0, 1) and kurtosis parameter q > 0. We X= denote this as X ∼ SEp (µ, Σ, ψ; q). Properties of this family are discussed in Gómez, Quintana and Torres (2007) and Bulut and Arslan (2015). Using the representation (3.1), the following theorem is a consequence of Theorem 3.1. Theorem 3.2. Suppose that Fi ∼ SE1 (µi , σi2 , ψ; q), where µi ∈ R, σi > 0, ψ is a charac- teristic generator for an n-variate slash-elliptical distribution. Then (F1 , · · · , Fn ) is JM if the inequality n X i=1 is satisfied. 3.3 σi ≥ 2 max{σ1 , · · · , σn } Skew-Elliptical Distributions A univariate random variable Z has a skew-elliptical distribution if its probability density function (pdf) is 2g(z)π(λz), −∞ < z < ∞, where g is a pdf of univariate elliptical distribution with center 0, λ ∈ R and π is the distribution function of g. We write Z ∼ SE1 (0, g, π, λ). In particular, if g is the pdf of N(0, 1), then Z is called has a skew-normal distribution and write Z ∼ SN(µ, σ 2 , λ). A random variable X follows a skew scale mixture of normal distribution with location parameter µ ∈ R, scale parameter σ 2 and skewness parameter λ ∈ R if its pdf is given by   x−µ g(x) = 2g0 (x)Φ λ , x ∈ R, σ where λ ∈ R, Φ is the distribution function of N(0, 1) and g0 is the pdf of scale mixture of normal distribution defined as g0 (x) = Z ∞ φ(x; µ, v 2 σ 2 )dH(v). 0 Here H is a (unidimensional) probability distribution function such that H(0) = 0. We use the notation X ∼ SSMN(µ, σ 2 , λ, H). For more details see Andrews and Mallows (1974) and Gómez-Sánchez-Manzano et al. (2006). 10 For complete mixability of skew-normal distribution, we have the following result. Theorem 3.3. Suppose that F has the distribution SN(µ, σ 2 , λ), then F is not n-CM (n ≥ 2) for sufficiently large |λ|. d Proof Assume random variable Xλ has the distribution F . Since Xλ = σX +µ, where X ∼ SN(0, 1, λ), we prove the theorem for the case Xλ ∼ SN(0, 1, λ) only. It follows from Henze (1986) that if Xλ ∼ SN(0, 1, λ), then it has the stochastic representation d Xλ = √ λ 1 |U| + √ V, 2 1+λ 1 + λ2 where U and V are independent N(0, 1) random variables. Moreover, r 2 λ . E(Xλ ) = √ 2 π 1+λ Without loss of generality, we assume λ > 0. For any Ui and Vi are independent N(0, 1) random variables, we have P √ λ 1 + λ2 n X i=1 |Ui | + √ n X 1 1 + λ2 n X Vi > nE(Xλ ) i=1 n X ! n \ 1 λ |Ui | + √ Vi ≥ 0} 2 1+λ 1 + λ2 i=1 i=1 i=2 ! n \ λ 1 λ 1 ≥P √ |U1 | + √ V1 > nE(Xλ ), { √ |Ui | + √ Vi ≥ 0} 2 1 + λ2 1 + λ2 1 + λ 1 + λ2 i=2   1 λ |U1 | + √ V1 ≤ nE(Xλ ) ≥ 1−P √ 1 + λ2 1 + λ2   n X λ 1 − P √ |Ui | + √ Vi < 0 > 0, 2 2 1 + λ 1 + λ i=2 ≥P √ λ 1 + λ2 |Ui | + √ 1 1 + λ2 Vi > nE(Xλ ), {√ for sufficiently large λ. This shows that the distribution SN(0, 1, λ) is not n-CM (n ≥ 2) for sufficiently large |λ|. Remark 3.2. For F ∼ SN(µ, σ 2 , λ), we conjecture that there exists an integer n0 (λ) such that F is not n-CM for n ≤ n0 (λ) and, F is n-CM for n > n0 (λ); For an integer n ≥ 2, there exists a λ0 (n) ≥ 0 such that F is n-CM whenever |λ| ∈ [0, λ0 (n)] and, F is not n-CM whenever |λ| ∈ (λ0 (n), ∞). Theorem 3.4. Suppose that F has the distribution SSMN(µ, σ 2 , λ, H), then F is not n-CM (n ≥ 2) for sufficiently large |λ|. 11 ! Proof For any Xi ∼ SSMN(µ, σ 2 , λ, H) (i = 1, 2, · · · , n) and V ∼ H such that V is independent of X1 , · · · , Xn . By the definition of skew scale mixture of normal distribution, we have Xi |V = v ∼ SN(µ, σ 2 v 2 , λv). Then for any constant C and sufficiently large |λ|, using Theorem 3.3, ! Z ! Z ∞ n n ∞ X X P Xi = C = P Xi = C|V = v dH(v) < dH(v) = 1. i=1 0 0 i=1 Thus F is not n-CM for sufficiently large |λ|. 4 Extensions to multivariate distributions In this section we extent some results in Section 3 to the class of n-variate elliptically contoured distributions. We first introduce some notions. The notation vec(A) denotes the ⊤ ⊤ vector (a⊤ 1 , · · · , an ) , where ai denotes the ith column of p × n matrix A, i = 1, 2, · · · , n. we use A ⊗ B to denote the Kronecker product of the matrices A and B; We use tr(A) to denote the trace of the square matrix A and etr(A) to denote exp(tr(A)) if A is a square matrix. We use the following definition given in Gupta, Varga and Bodnar (2013). Definition 4.1. Let X be a random matrix of dimensions p × n. Then, X is said to have a matrix variate elliptically contoured distribution if its characteristic function has the form E(etr(iT⊤ X)) = etr(iT⊤ M)Ψ(tr(T⊤ ΣTΦ)). (4.1) with T : p × n, M : p × n, Σ : p × p, Φ : n × n, Σ ≥ 0 (positive semidefinite), Φ ≥ 0, and Ψ : [0, ∞) → R. This distribution will be denoted by X ∼ Ep,n (M, Σ ⊗ Φ, Ψ). The important special case of matrix variate elliptically contoured distribution is the matrix variate normal distribution (X ∼ Np,n (M, Σ ⊗ Φ)), its characteristic function is   1 ⊤ ⊤ ⊤ (4.2) E(etr(iT X)) = etr iT M − T ΣTΦ . 2 The next lemma shows that linear functions of a random matrix with matrix variate elliptically contoured distribution have elliptically contoured distributions also (see Theorem 2.2 in Gupta, Varga and Bodnar (2013)). 12 Lemma 4.1. Let X ∼ Ep,n (M, Σ ⊗ Φ, Ψ). Assume C : q × m, A : q × p, and B : n × m are constant matrices. Then, AXB + C ∼ Ep,n (AMB + C, AΣA⊤ ⊗ B⊤ ΦB, Ψ). The next lemma gives the marginal distributions of a matrix variate elliptically contoured distribution (see Theorem 2.9 in Gupta, Varga and Bodnar (2013)). Lemma 4.2. Let X ∼ Ep,n (M, Σ ⊗ Φ, Ψ), and partition X, M, and Φ as X = (X1 , X2 ), M = (M1 , M2 ), and Φ= " Φ11 Φ12 Φ21 Φ22 # , where X1 is p × m, M1 is p × m, and Φ11 is m × m, 1 ≤ m < n. Then X1 ∼ Ep,m (M1 , Σ ⊗ Φ11 , Ψ). Theorem 4.1. Suppose that F ∼ Ep (0, Σ, Ψ), where Σ ≥ 0 is a p × p matrix, Ψ is a characteristic generator for a p × n matrix variate elliptically contoured distribution (n ≥ 2). Then there exist n p-dimensional random vectors X1 , · · · , Xn identically distributed as F such that P (X1 + · · · + Xn = 0) = 1. Proof Using Lemma 4.2 we can choose X ∼ Ep,n (0, Σ ⊗ Φ, Ψ) with all marginals Xi ’s (the ith column of X, i = 1, 2, · · · , n) have the same p-elliptical distribution Ep (0, Σ, Ψ), where Φ ≥ 0 is an n × n matrix with diagonal elements are 1. Using Lemma 4.1 one finds that X1 + · · · + Xn = Xen ∼ En (0, Σ ⊗ (etn Φen ), Ψ). Taking Φ = (1 − ρ)En + ρen eTn with 1 ρ = − n−1 , En is n × n identity matrix. It follows that P (X1 + · · · + Xn = 0) = 1. Acknowledgements. We are grateful to Ruodu Wang for insightful comments and valuable suggestions on the first manuscript. The research was supported by the National Natural Science Foundation of China (No. 11571198, No. 11701319). 13 References [1] Andrews, D. R. and Mallows, C. L. (1974). Scale mixtures of normal distributions. Journal of the Royal Statistical Society, Series B 36, 99-102. [2] Bignozzi, V. and Puccetti, G. (2015). Studying mixability with supermodular aggregating fuctions. Statistics and Probability Letters 100, 48-55. [3] Bulut, Y. M. and Arslan, O. (2015). Matrix variate slash distribution. Journal of Multivariate Analysis 137, 173-178. [4] Cambanis, S., Huang, S. and Simons, G. (1981). On the theory of elliptically contoured distributions. Journal of Multivariate Analysis 11, 368-385. [5] Fang, K. T., Kotz, S. and Ng, K. W. (1990). Symmetric Multivariate and Related Distributions. Chapman & Hall, London. [6] Gaffke, N. and Rüscherndorf, L. (1981). On a class of extremal problems in statistics. Mathematische Operationsforschung und Statistik. Series Optimization 12(1), 123135. [7] Gómez-Sánchez-Manzano, E., Gómez-Villegas, M. A. and Marı́n, J. M. (2006). Sequences of elliptical distributions and mixtures of normal distributions. Journal of Multivariate Analysis 97, 295-310. [8] Gómez, H.W., Quintana, F. A. and Torres, F. J. (2007). A new family of slashdistributions with elliptical contours. Statistics and Probability Letters 77(7), 717725. [9] Gupta, A. K., Varga, T. and Bodnar, T. (2013). Elliptically Contoured Models in Statistics and Portfolio Theory (Second Edition), Springer New York. [10] Henze, N. (1986). A probabilistic representation of the ‘skew-Normal’ distribution. Scandinavian Journal of Statistics 13(4), 271-275. [11] Kano, Y. (1994). Consistancy property of elliptical probability density functions. Journal of Multivariate Analysis 51,139-147. [12] Knott, M. and Smith, C. S. (2006). Choosing joint distributions so that the variance of the sum is small. Journal of Multivariate Analysis 97, 1757-1765. 14 [13] Puccetti, G., Wang, B. and Wang, R. (2012). Advances in complete mixability. Journal of Applied Probability 49 (2), 430-440. [14] Puccetti, G., Wang, B. and Wang, R. (2013). Complete mixability and asymptotic equivalence of worst-possible VaR and ES estimates. Insurance: Mathematics and Economics 53(3), 821-828. [15] Puccetti, G. and Wang, R. (2015a). Extremal dependence concepts. Statistical Science 30(4), 485-517. [16] Puccetti, G. and Wang, R. (2015b). Detecting complete and joint mixability. Journal of Computational and Applied Mathmatics 280, 174-187. [17] Rüschendorf, L. and Uckelmann, L. (2002). Variance minimization and random variables with constant sum, in: Cuadras, et al. (Eds.), Distributions with Given Marginals, Kluwer, 211-222. [18] Stefanski, L. A. (1990). A normal scale mixture representation of the logistic distribution. Statistics & Probability Letters 11, 69-70. [19] Wang, R. (2014). Asymptotic bounds for the distribution of the sum of dependent random variables. Journal of Applied Probability 51(3), 780-798. [20] Wang, R. (2015). Current open questions in complete mixability. Probability Surveys 12, 13-32. [21] Wang, R., Peng, L. and Yang, J. (2013). Bounds for the sum of dependent risks and worst value-at-risk with monotone marginal densities. Finance and Stochastics 17 (2), 395-417. [22] Wang, B. and Wang, R. (2011). The complete mixability and convex minimization problems with monotone marginal densities. Journal of Multivariate Analysis 102 (10), 1344-1360. [23] Wang, B. and Wang, R. (2016). Joint mixability. Mathematics of Operations Research 41(3), 808-826. [24] West, M. (1987). On scale mixtures of normal distributions. Biometrika 74(3), 646648. 15 

10math.ST

DISTRIBUTIONS AND STATISTICAL POWER OF OPTIMAL SIGNAL-DETECTION METHODS IN FINITE CASES arXiv:1702.07082v1 [math.ST] 23 Feb 2017 By Hong Zhang∗,‡ , Jiashun Jin§ and Zheyang Wu∗,†,‡ Worcester Polytechnic Institute‡ and Carnegie Mellon University§ In big data analysis for detecting rare and weak signals among n features, some grouping-test methods such as Higher Criticism test (HC), Berk-Jones test (B-J), and φ-divergence test share the similar asymptotical optimality when n → ∞. However, in practical data analysis n is frequently small and moderately large at most. In order to properly apply these optimal tests and wisely choose them for practical studies, it is important to know how to get the p-values and statistical power of them. To address this problem in an even broader context, this paper provides analytical solutions for a general family of goodness-of-fit (GOF) tests, which covers these optimal tests. For any given i.i.d. and continuous distributions of the input test statistics of the n features, both p-value and statistical power of such a GOF test can be calculated. By calculation we compared the finite-sample performances of asymptotically optimal tests under the normal mixture alternative. Results show that HC is the best choice when signals are rare, while B-J is more robust over various signal patterns. In the application to a real genome-wide association study, results illustrate that the p-value calculation works well, and the optimal tests have potentials for detecting novel disease genes with weak genetic effects. The calculations have been implemented in an R package SetTest and published on the CRAN. 1. Introduction. In big data analysis, signals are often buried within a large amount of noises and are thus relatively weak and rare. It is ideal to apply the optimal tests that are capable of detecting the minimal signals required by statistics. Through asymptotics, many theoretical studies have brought exciting results on designing such optimal tests. In particular, under the Asymptotically Rare and Weak (ARW) setting, the Higher Criticism (HC) and its various modifications, the Berk-Jones (B-J) type tests, a spectrum of φ-divergence type tests, etc. were studied and proven asymptotically optimal [1, 2, 3, 4]. These optimal tests are attractive in many scientific researches. For example, in large-scale genetic association studies, a main strategy to find disease-associated genes is to determine whether ∗ † Partially supported by NSF grant DMS-1309960. Corresponding author. 1 2 H. ZHANG ET AL. some of the genetic variants within candidate genes could affect disease outcome. Such genetic effects are often weak and rare, especially relative to the cumulated noise level in big data. However, for practical applications under finite cases the questions remain on 1) how to analytically calculate p-values as well as statistical power, and 2) what are the real performances of those methods that are asymptotically equivalent. First, to obtain the p-value for error control, the Monte-Carlo simulation or permutation tests have significant limitations: (A) they require daunting computation, and (B) empirical p-values are discrete, causing ties among candidates. These issues are especially concerned when very small p-values are demanded to handle a huge number of simultaneous tests. Secondly, for those optimal tests with the same asymptotic property under n → ∞, it is important to understand their relative performance for finite n. In order to solve these problems we need to calculate the distributions of these tests under both H0 and H1 at each given n. Comparing with the literature, this paper gives a complete answer by providing a comprehensive calculation for a broad family of relevant tests. 1.1. Limitations of current methods. In general there are two types of methods for distribution calculation. The first is to calculate the exact distribution. Recursive methods (e.g., Noe’s recursion[5, 6], Bolshev’s recursion [7, 8], Steck’s recursion [9, 10], Ruben’s recursion [11], etc.) were developed to calculate the distribution of Kolmogorov-Smirnov type statistics under H0 . In a similar fashion, Barnett and Lin [12] provided a calculation method specifically for HC. Such recursive methods have heavy computation load, with complexity of O(n3 ). Moscovich, Nadler and Spiegelman [13] reduced the computation to O(n2 ). However, all these methods assume the full domain R = {1 ≤ i ≤ n} in the supremum formula of the relevant tests (see the HC statistic in (3) as an example). However, calculation that allows arbitrary R is important, because many tests improve performance by properly restricting R. For example, a modified version of HC is under R = {1 < i ≤ n/2, p(i) ≥ 1/n}. This is because that including too small p-values p(i) could make HC heavy tailed; while including too big p-values may unnecessarily reduces computational efficiency (c.f. [1]). Moreover, these methods did not give statistical power calculation under H1 yet. The second type methods are to approximate the distribution. KolmogorovSmirnov type statistics have been proven to converge in law to an extremevalue distribution [15, 16]. However, such convergence is too slow to be accurate for even moderately large n [12]. Recently, Li and Siegmund (LS) [14] developed an asymptotic approximation method. One problems is that DISTRIBUTIONS OF SIGNAL-DETECTION METHODS IN FINITE CASES 3 Fig 1: Comparison among different methods for calculating the right-side probability of the modified HC test (MHC) in (3) with R = {1 < i ≤ i.i.d. n/2, p(i) ≥ 1/n}[1] under H0 : Xi ∼ N (0, 1). Black solid curves: by simulation; red dashed curves: by Corollary 1; green dotted curves: by Li and Siegmund’s method[14]. this method cannot approximate the whole distribution. Figure 1 shows that the LS calculation (green dotted curves) fails to match the simulation (black solid curves) at small thresholds, whereas our calculation (e.g., the red dashed curves) reveal the whole distributions. Moreover, comparing with our study setting, LS assumes slightly more restriction for the supremum searching domain R, and covers a narrower range of test types. Again, literature studies did not yet provide satisfiable method for deriving the relevant distributions under H1 for statistical power calculation. 1.2. Our contribution. This paper has two folds of contributions. First, it provides techniques for calculating the distributions of a general family of goodness-of-fit (GOF) tests, which covers the optimal tests described above. We give calculation methods for the exact as well as the approximated 4 H. ZHANG ET AL. Fig 2: The association p-values for genes by exact calculation of four tests. First row: HC 2004 and Berk-Jones; second row: reverse Berk-Jones and HC 2008 . distributions, balancing between accuracy and computation burden. The methods allow (A) arbitrary truncation strategies (e.g., R for HC in (3)), and (B) arbitrary null and alternative hypotheses as long as they are i.i.d. and continuous. With such techniques, p-values and statistical power can be calculated at any testing threshold. Second, based on analytical calculation we systematically compared the power of the asymptotically optimal tests. Simulations were applied for examing these tests under various n values and signals patterns in practice. We also demonstrated the application of these weak-signal-sensitive tests in a real genome-wide association study (GWAS) for detecting genes associated to the Crohn’s disease (see Figure 2). DISTRIBUTIONS OF SIGNAL-DETECTION METHODS IN FINITE CASES 5 In summary, the innovative contribution of this work lies on providing a more mature statistical framework that revolves the problem of calculating distributions for a broader family of GOF tests under both H0 and H1 . Thus, those asymptotically equivalent tests for detecting weak signals can be apply, as well as compared by power, in practical applications. The paper is organized as follows. In section 2 we formulate the problem under a general setup of hypotheses and GOF tests, and brief the essence of our strategy. The analytical results are presented in Section 3 for both exact and approximated calculations. Through simulations Section 4 numerically evidences the calculation accuracy, and provides power comparisons among the asymptotically optimal tests. We show the application of the GOF tests in a real GWAS in Section 5. In Section 6 we discuss relevant theoretical and practical issues. The techniques of relevant proofs and lemmas are given in the Appendix. 2. Problem Formulation and General Strategy. 2.1. Background on Detection of Weak and Sparse Signals. We consider the signal detection problem through testing whether there exist “signals” in noisy background. Under the broadly applicable Gaussian Means Model, signals refer to the nonzero means in the normally distributed data values (or statistics) X1 , X2 , ..., Xn . A test is said of detecting signals if it successfully concludes the alternative when it is true. A typical setting for the null and the alternative is (1) i.i.d. H0 : Xi ∼ N (0, 1), i.i.d. H1 : Xi ∼ N (µ, 1) + (1 − )N (0, 1), where  ∈ (0, 1) denotes the expected proportion of signals. Under the Asymptotic Rare and Weak (ARW) setting, the p proportion of −α signals is n = n , α ∈ (1/2, 1), and the mean is µn = 2r log(n), r ∈ (0, 1). Landmark studies [1, 17, 18] have provided the asymptotic detection boundary in terms of a function curve of the signal strength and sparsity:  α − 1/2 1/2 < α ≤ 3/4 √ (2) r = ρ? (α) = 3/4 < α < 1. (1 − 1 − α)2 When the signals are below the curve, all tests will fail to distinguish H0 and H1 as n → ∞. Whenever signals are above the curve, the so-called asymptotically optimal tests are able to make both the type I and the type II error rates converge to zero. A particular optimal test is the Higher Criticism (HC) test [19, 1]: (3) i/n − p(i) √ HCn,R = sup n q , R p(i) (1 − p(i) ) 6 H. ZHANG ET AL. where p(1) ≤ ... ≤ p(n) are the order statistics of pi = 2(1 − F0 (|Xi |)), i = 1, ..., n, the two-sided p-values of Xi . The supremum domain R regards to the p-value magnitude, or the index i, or both. Note that in literature HC formula could also be written as (e.g., [20, 12] ) P i > t} − 2nΦ̄(t) i {X p (4) HC = sup . ∗ t∈R 2nΦ̄(t)Φ(t) Because the formula is monotone, the supremum domain R on p(i) is equivalent to the supremum domain R∗ on t. Following that, a variety of modified Higher Criticism (HC), the BerkJones (B-J) type tests, a spectrum of φ-divergence type tests, etc. were studied and proven asymptotically optimal [1, 2, 3, 4]. In this paper, we provide calculation for the p-value and statistical power for relevant tests under finite n. 2.2. Study Formulation. In this paper, we consider arbitrary i.i.d. and continuous null and alternative hypotheses: (5) i.i.d. H0 : Xi ∼ F0 , i.i.d. Ha : Xi ∼ F1 , i = 1, ..., n. We consider the goodness-of-fit (GOF) tests that evidence the distinction between the data and any given null distribution. Consider either one-sided p-values pi = 1 − F0 (Xi ) or two-sided p-values pi = 2(1 − F0 (|Xi |)). When F0 is continuous, the null hypothesis is equivalent to i.i.d. H0 : pi ∼ Uniform(0, 1), i = 1, ..., n. The typical idea of GOF is that if any Xi do not follow F0 , its p(i) shall distinct from it expectation, which is roughly i/n. Following this idea we consider a general family of tests, (6) i Sn,R = sup f ( , p(i) ), n R where the supremum domain (7) R = {i : k0 ≤ i ≤ k1 } \ {p(i) : α0 ≤ p(i) ≤ α1 } for given k0 ≤ k1 ∈ {1, ..., n} and α0 ≤ α1 ∈ [0, 1]. We assume that for fixed x = i/n the function f (x, y) is monotonically decreasing in y = p(i) , so that the smaller the input p-values, the larger the statistic and the stronger the evidence against H0 . DISTRIBUTIONS OF SIGNAL-DETECTION METHODS IN FINITE CASES 7 This GOF family contains lots of test statistics widely applicable to practice. For example, the simple one-sided Kolmogorov-Smirnov test statistic (c.f. [21], page 447, denoted KS + here) is a classic GOF test, which directly measures the difference between p(i) and i/n. That is, the f function is defined by fKS + (x, y) = x − y. (8) To improve the test, the difference between p(i) and i/n should be scaled with regard to p(i) or i/n. This is because smaller p(i) values are more important to evidence agains H0 . Such scaled KS tests are related to the Higher Criticism (HC) statistics proposed in 2004 and 2008 [1, 2], respectively, where the f functions are defined as √ x−y np ; y(1 − y) √ x−y . fHC 2008 (x, y) = n p x(1 − x) fHC 2004 (x, y) = (9) Jager and Wellner introduced a collection of φ-divergence statistics [4], which are also based on the supremum of functions: 1 (1 − xs y 1−s − (1 − x)s (1 − y)1−s ), s 6= 0, 1, s(1 − s) x 1−x f1φ (x, y) = x log( ) + (1 − x) log( ), y 1−y y 1−y f0φ (x, y) = y log( ) + (1 − y) log( ). x 1−x fsφ (x, y) = (10) For certain s (e.g., s = 2 or −1) these statistics are two-sided in the sense that switching the values of x = i/n and y = p(i) gives the same statistic. However, this property is not appropriate in the scenario of signal detection. It is the relationship p(i) < i/n, instead of the opposite, indicates signals. Thus for signal detection purpose, one-sided test is more reasonable and more powerful, as is the same idea for defining KS and HC. Thus, here we consider the one-sided version of φ-divergence statistics, which can be achieved by a simple adjustment of the f function to be, for example,  q  2nf φ (x, y) y ≤ x, q s (11) fs (x, y) =  − 2nf φ (x, y) y > x. s Now for all s, fs (x, y) is guaranteed decreasing in y. Such one-sided φdivergence statistics cover HC exactly: f2 = fHC 2004 and f−1 = fHC 2008 . 8 H. ZHANG ET AL. Also, s = 1 and 0 correspond to the Berk-Jones statistic [22, 1, 14] and the reverse Berk-Jones statistic [22], respectively. Note that the input p-values themselves could be two-sided in order to accommodate the consideration of signal directionality in practical problems. 2.3. Calculation Strategy. In the following we first introduce the essential idea of the calculation. Under various settings and assumptions, detailed strategies for obtaining the exact and approximated distributions will be described in Section 3. For any given continuous CDFs in (5), we define a monotone transformation function in (0, 1):  x under H0 : F0 , (12) D(x) = under H1 : F1 6= F0 . 1 − F1 (F0−1 (1 − x)) Note that for any p-value pi , we have D(pi ) ∼ U nif orm(0, 1) under either H0 or H1 . Consider the function f (x, y) of any statistic Sn,R in (6), for each fixed x define its inverse function (13) g(x, ·) = f −1 (x, ·). For example, for the HC statistics defined (9), the g functions are √ √ b2 /n−(b/ n) b2 /n+4x(1−x) 1 gHC 2004 (x, b) = 1+b2 /n [x + ]; 2 (14) √ p gHC 2008 (x, b) = x − (b/ n) x(1 − x). In general if the closed form of g(x, ·) is not available, it can always be found numerically, since f (x, y) is strictly decreasing in y. Now under either H0 or H1 , the CDF function of Sn,R is (15) i P (Sn ≤ b) = P (sup f ( , p(i) ) ≤ b) n R \ i = P ( {p(i) > g( , b)}) n R = P {D(p(i) ) > D(g( ni , b)), all i and p(i) in R}. The key idea is that under either H0 or H1 , U(i) := D(p(i) ) is the ith order statistic of Uniform(0, 1), and the joint distribution of U(i) , i = k0 , ...., k1 , can be studies one way or another for getting the final probability. To simplify the presentation, we list below the notations to be referred later on. DISTRIBUTIONS OF SIGNAL-DETECTION METHODS IN FINITE CASES 9 (N1) uk := D(g( nk , b) ∨ α0 ), following the definitions in (12)–(15), and a potential constant α0 ≥ 0 in (7). (N2) F̄B(α,β) (x) denotes the survival function of Beta(α, β) distribution. (N3) FΓ(α) (x) and F̄Γ(α) (x) denote the CDF and survival function of Gamma(α, 1) distribution, respectively, where the shape parameter is α, the scale parameter is 1. (N4) Based on the notation (N3) define hk (x) := xFΓ(k−1) (kx) − FΓ(k) (kx). (N5) fP (λ) (x) denotes the probability mass function of P oisson(λ) distribution. 3. Analytical Results. 3.1. Exact Calculations. The first theorem provides the exact calculation for the distribution of GOF in (6), where the supremum domain R involves truncation of the index i. For example, the initial HC was defined with R = {1 ≤ i ≤ n/2} [1]. Theorem 3.1. Consider any GOF statistic in (6) with R = {k0 ≤ i ≤ k1 } for given 1 ≤ k0 ≤ k1 ≤ n. Let m = n − k1 + 1. Following the notations (N1), (N2), and ak1 = ak = n! (n−k1 +1)! F̄B(1,m) (uk1 ), and kX 1 −k ujk+j−1 n! F̄B(k1 −k+1,m) (uk1 ) − ak+j , (n − k + 1)! j! k = k1 − 1, ..., 1. j=1 Under either H0 or H1 we have P (Sn ≤ b) = F̄B(k1 ,m) (uk1 ) − kX 1 −1 i=k0 uii ai+1 . i! Another type of truncation is based on the value of p(i) . In the case of HC under the null of N (0, 1), this truncation is equivalent to the truncation on t in (4). The following theorem gives the exact calculation for the general GOF tests in (6) with such truncations. Theorem 3.2. Consider any GOF with statistic in (6) with R = {α0 ≤ p(i) ≤ α1 } for given 0 ≤ α0 < α1 ≤ 1. Following the notations (N1) and 10 H. ZHANG ET AL. (N2), define β0 = D(α0 ), aj (k) = β1 = D(α1 ), β0i−1 (1−β1 )n−j+1 (i−1)!(n−j+1)! j−k l X uk+l−1 uj−1 cij = n! )− β1j−k F̄B(j−k,1) ( (j − k)! β1 aj (j) = 0, l! l=1 aj (k + l), and 1 ≤ i ≤ k1 , i < j ≤ n + 1, k = 1, ..., j − 1 Under either H0 or H1 , we have P (Sn,R ≤ b) = k1 n+1 X X cij aj (i) i=1 j=i+1 The most general R is in (7), which defines the truncation for both the index and the p-values. The following theorem provides the calculation for exact distribution under this general setup, for which no literature has provided solution before. Theorem 3.3. Consider any GOF with statistic in (6)with R = {α0 ≤ p(i) ≤ α1 } ∩ {k0 ≤ i ≤ k1 } for given 1 ≤ k0 ≤ k1 ≤ n and 0 ≤ α0 < α1 ≤ 1. Following the notations in Theorem 3.2 and (N1) and (N2), define ( ( i i ≥ k0 j j ≤ k1 + 1 ĩ = , j̃ = , β̃0 = β0 I{i<k0 } , k0 i < k0 k1 + 1 j > k1 + 1 j̃−k aj (k) = X ulk+l−1 uj̃−1 n! (j−k) β1 )− aj (k + l), and F̄B(j̃−k,j−j̃+1) ( (j − k)! β1 l! l=1 aj (j̃) = 0, 1 ≤ i ≤ k1 , ĩ < j ≤ n + 1, k = 1, ..., j̃ − 1. Under either H0 or H1 , we have P (Sn,R ≤ b) = k1 n+1 X X i=1 j=ĩ+1   j̃−1 j−i k−i+1 X ˜ u − β̃ 0 n!(β − β̃ ) (u − β ) j̃−1 1 0 0 k cij  )− aj (k + 1) . F̄B(j̃−i,j−j̃+1) ( (j − i)! (k − i + 1)! β1 − β̃0 k=ĩ The study setting in Li and Siegmund [14] is a special case of the general truncation with α1 = 1. For example, the “modified HC” is to take R = {1 < i ≤ n/2, p(i) ≥ 1/n}[1]. Li and Siegmund’s calculation provided an approximation for the modified HC under H0 . As shown in Figure 1, the approximation does not give the whole distribution. For this setting, the corollary below gives the exact distribution for the general GOF in (6) under both H0 and H1 . DISTRIBUTIONS OF SIGNAL-DETECTION METHODS IN FINITE CASES 11 Corollary 1. Consider any GOF with statistic in (6) with R = {α0 ≤ p(i) } ∩ {k0 ≤ i ≤ k1 } for given 1 ≤ k0 ≤ k1 ≤ n and α0 > 0. Following the notations in Theorem 3.1, 3.2 and (N1) and (N2), define ci = β0i−1 , (i − 1)! 1 ≤ i ≤ k1 . Under either H0 or H1 , we have   k1 kX 1 −1 n+1−i k+1−i X n!(1 − β̃ ) u − β̃ (u − β̃ ) 0 0 0 k k P (Sn,R ≤ b) = ci  F̄B(k1 +1−i,m) ( 1 )− ak+1  . (n + 1 − i)! (k + 1 − i)! 1 − β̃0 i=1 k=ĩ To sum up, Theorem 3.3 covers Theorem 3.1 by fixing i = 1, j = n + 1 (so that cij = 0) and α0 = 0, α1 = 1. It covers Theorem 3.2 by letting k0 = 1, k1 = n, and covers Corollary 1 by fixing j = n + 1 and α1 = 1. Regarding the computational complexity, the calculation given by Theorem 3.1 is equivalent to solving an upper triangular linear systems using backward substitution, which is O(n2 ), an improvement from O(n3 ) required by other exact distribution calculations [5, 6, 12]. More importantly, our method can handle a more general family of GOFs defined in (6) with a more flexible supremum domain R. When R is more complicated, Corollary 1 is still O(n2 ) because the inner loop ak is not dependent on i. Theorem 3.2 and Theorem 3.3 are O(n3 ). 3.2. Approximate the distributions. In this section, we provide several calculation methods to approximate the distributions of the GOF family. Under more restricted conditions, computation load could be significantly reduced. For simplicity we provide results here for the supremum domain R = {k0 ≤ i ≤ k1 }. More general results for R in (7) can be obtained by following the similar idea of the calculation in Theorem 3.3. The theorem below gives an approximated calculation based on a joint Gamma distribution. Theorem 3.4. Consider any GOF with statistic in (6) with R = {k0 ≤ i ≤ k1 }. Following the notations (N1) and (N3), define dk = (n + 1)D(g( nk , b)), k = k0 , ..., k1 . k−1 j X dk1 −k+j ck = F̄Γ(k) (dk1 ) − ck−j , k = 2, ..., k1 , and j! j=1 c1 = F̄Γ(1) (dk1 ). 12 H. ZHANG ET AL. Under either H0 or H1 , we have P (Sn ≤ b) = (1 + o(1))(F̄Γ(k1 ) (dk1 ) − kX 1 −1 k=k0 dkk ck −k ). d! 1 To further reduce the computation, we can also deduce a one-step formula under stronger assumptions. In particular, if D(g( nk , b)) is a linear function of k, we can provide a closed-form formula for the distribution that gives the same accuracy as the above theorem. Theorem 3.5. Consider any GOF with statistic in (6) with R = {1 ≤ i ≤ k1 }. Assume the function D(g( nk , b)) = a+λk, for some λ ≥ 0. Following the notation (N3) and (N4), under either H0 or H1 we have P (Sn ≤ b) = (1 + o(1))e−a (1 − λ + hk1 (λ)). One example that satisfies the linearity of D(g( nk , b)) is the simple KolmogorovSmirnov (KS + ) test in (8) under H0 , where a = −(n + 1)b and λ = n+1 n . In this case, based on the above theorem we have the following corollary Corollary 2. Consider the test statistic KS + in (8) under H0 . Following the notation (N3) and (N4), for b ≤ n1 , we have P (KS + ≤ b) = (1 + o(1))e(n+1)b (− 1 n+1 + hk1 ( )). n n The requirement of linear D(g( nk , b)) in Theorem 3.5 is stringent. However, if D(g( nk , b)) is close to linear, we can still further simplify the calculation of Theorem 3.4. As one example, Theorem 3.6 below provides such a calculation formula for GOF and the conditions on D(g( nk , b)). The strategy for proving this theorem largely follows the idea in Li and Siegmund [14]. The key difference is that, instead of using the beta distribution in Li and Siegmund, we use the gamma distribution, which has a simpler density function for addressing a wider family of the GOF tests. See Appendix for details. Theorem 3.6. Consider GOF tests in (6) with R = {k0 ≤ i ≤ k1 }. Follow notations (N1) – (N5), and define dk = (n + 1)D(g( nk , b)), d0k = √ d (n + 1) dx D(g( nk , b)), and k ∗ = min{k1 − k, n}. Assume D(g(x, b)) satisfies 1. D(g(x, b)) < 1 is increasing and convex in x for d 2. dx D(g(x, b)) < 1, and k0 n ≤x≤ k1 n, DISTRIBUTIONS OF SIGNAL-DETECTION METHODS IN FINITE CASES 13 3. D(g(k/n, b)) < k n+1 , for k > 1 and large n. Under either H0 or H1 we have P (Sn ≥ b) = (1 + o(1)) k1 X k=k0 (1 − d0k d0 + hk∗ ( k ))fP (dk ) (k). n n Such a close-to-linear property of D(g( nk , b)) can be satisfied by HC 2004 under H0 . In this case, D(g( nk , b)) = g(x, b) is given in (14), and the condi√ tions can be satisfied when b is in the order of O( n). Thus, we can get a computation-easy formula. Corollary 3. Consider the test statistic HC 2004 in (9) with R = {k0 ≤ i ≤ k1 }. Let b0 = √bn be a positive constant > 2x − 1, kn0 < x < kn1 . Define function q 1 2 2 g(x, b0 ) = [x + (b − b 0 b0 + 4x(1 − x))/2], 0 1 + b20 1 b0 (1 − 2x) g 0 (x, b0 ) = [1 − p 2 ]. 2 1 + b0 b0 + 4x(1 − x) Following the notation (N2), under H0 , we have  k1  X 0 k 2004 0 k P (HC ≥ b) = (1+o(1)) 1 − g ( , b0 ) + hk∗ (g ( , b0 )) fP (g( k ,b0 )n) (k). n n n k=k0 The above formula is different from that given in Li and Sigmund [14]. However, for the HC under H0 , both formulas require the threshold b = √ O( n). Thus, both methods do not get the whole distribution. Furthermore, the accuracy depends on the linear approximation of the D(g( nk , b)) function, which could be far from the truth under general H1 . Thus this type of calculation may not be very attractive for calculation statistical power. Meanwhile, numeric results in Section 4 show that both methods are accurate for calculating small p-values under relatively large threshold b. 4. Numerical results. In this section we first evidence the accuracy of our methods by comparing the calculations with the Monte-Carlo simulations under various settings of H0 and H1 . Then, based on calculation we compare the finite-n performance of the asymptotically optimal tests over various signal patterns. The supremum domain is R = {1 ≤ i ≤ n/2} by default, and will be specified otherwise. The simulations run 5,000 repetitions by default. 14 H. ZHANG ET AL. i.i.d. 4.1. Evaluate the accuracy of calculations. Under H0 : Xi ∼ N (0, 1), i = 1, ..., n. Figure 3 shows the right-tail probability of HC over the threshold b. Comparing with simulation (black solid curves), the exact calculation by Theorem 3.1 (cyan dashed curves) has a perfect match. The approximation by Theorem 3.4 is accurate over the whole distribution when n is fairly big. The closed-formula calculation methods by Li and Siegmund [14] (blue dotted curves) and by Corollary 3 (green dashed curves) do not fit the whole distribution curve, but both can provide good approximation for calculation small p-values at large threshold. For distribution calculation under the alternative, we considered the null of i.i.d. N (0, 1) (on which the input p-values are based) and the alternatives: i.i.d. H1 : Xi ∼ (1 − )N (0, 1) + N (1, 1), or i.i.d. H1 : Xi ∼ (1 − )N (0, 1) + tν . Figure 4 shows the right-tail probability of HC under the normal mixture alternative with µ = 1,  = 0.1 (row 1), or under the Student’s t distribution with degrees of freedom ν = 5 (row 2). In both cases the distribution curves by exact calculation (Theorem 3.1, cyan dashed curves) and by approximation (Theorem 3.4, red dot-dashed) are close to simulation (black solid curves). As expected, the approximation is more accurate for larger n. Besides the normal distributions, Theorems 3.1 – 3.4 can handle any given continuous F0 and F1 . Here we test on four settings studied in the initial paper of HC [1]. The first setting regards a Chi-squared model: i.i.d. H0 : Xi ∼ χ2ν (0), vs. i.i.d. H1 : Xi ∼ (1 − )χ2ν (0) + χ2ν (δ). where ν is the degree of freedom, δ is the non-centrality parameter. The second setting is a Student’s t mixture model: i.i.d. H0 : Xi ∼ tν (0), vs. i.i.d. H1 : Xi ∼ (1 − )tν (0) + tν (δ). The third setting is a chi-squared-exponential mixture model, i.i.d. H0 : Xi ∼ exp(ν), vs. i.i.d. H1 : Xi ∼ (1 − ) exp(ν) + χ2ν (δ). The fourth setting concerns a generalized normal distribution (also known as power exponential distribution) model, i.i.d. H0 : Xi ∼ GNp (0, σ), vs. i.i.d. H1 : Xi ∼ (1 − )GNp (0, σ) + GNp (µ, σ), DISTRIBUTIONS OF SIGNAL-DETECTION METHODS IN FINITE CASES 15 Fig 3: Comparison among different calculations for the distributions of HC i.i.d. over various n value under H0 : Xi ∼ N (0, 1). Black solid curves: by simulations; cyan dashed curves: by Theorem 3.1 for exact distribution; red dot-dashed curves: by Theorem 3.4; blue dotted curves: by Li and Siegmund [14]; green dashed curves: by Corollary 3 . 16 H. ZHANG ET AL. i.i.d. Fig 4: The alternative distributions of HC under H0 : Xi ∼ N (0, 1) vs. H1 : Xi ∼ 0.9N (0, 1) + 0.1N (1, 1) (row 1), or H1 : Xi ∼ 0.5N (0, 1) + 0.5t5 (row 2). Column 1: n = 10; column 2: n = 100. Black solid curves: by simulations; cyan dashed curves: by Theorem 3.1; red dot-dashed curves: by Theorem 3.4. DISTRIBUTIONS OF SIGNAL-DETECTION METHODS IN FINITE CASES 17 where the probability density function of GNp (µ, σ) is 1 |x − µ|p exp(− ), Cp pσ p Cp = 2p1/p Γ(1 + 1/p)σ. Notice that GN1 (µ, σ) is the Laplace distribution and GN2 (µ, σ) is N (µ, σ 2 ). Each row of Figure 5 illustrates the alternative distribution of HC under each of the four settings for n = 10 (left column) and 100 (right column). Again the distribution calculation is fairly accurate in all cases. For the closed-form calculation formula given by Theorem 3.5, the boundary is assumed linear: D(g( ni , b)) = a + λk ≥ 0 in (15). One example is the KS + in (8) under H0 . Figure 6 demonstrates the accuracy of the calculation based on either fixed slope λ = 0.5 or fixed intercept a = 0.5. Here k0 = 1, k1 = n = 50. As the boundary a + λk increases, the probabilities from both calculation and simulation decrease and well-matched as expected. 4.2. Compare statistical power of asymptotically optimal tests. All φdivergence statistics with s ∈ [−1, 2] possess the same asymptotic optimality property for detecting weak and sparse signals [4]. It is of interest to know the performance of such statistics under finite n. Here we study s = 2, 1, 0, −1, which correspond to HC 2004 , the Berk-Jones statistic, the reverse Berk-Jones statistic, and HC 2008 , respectively. To show the calculation accuracy, we calculated (by Theorem 3.1) the critical values at the significance levels of 10%, 5% and 1%. Then at these critical values we got the empirical type I error rates through simulation (10,000 repetitions). Table 1 presents the thresholds by calculation and the correspondingly empirical type I error rates. The closeness of these empirical and nominal type I error rates evidences that the calculation for these tests is accurate. Now through the calculation based on Theorem 3.1, we systematically compared the power of these tests under i.i.d. H0 : Xi ∼ N (0, 1), vs. i.i.d. H1 : Xi ∼ (1 − )N (0, 1) + N (µ, 1). With the type I error rate controlled at 5%, Figure 7 provides the statistical power at various µ, n and . There are a few interesting observations. First, HC 2004 performs well when signals are sparse. However, it seems more relevant to the average number of signals, i.e., n, rather than the proportion  itself. For example, at fixed n = 5 (panels in the first column), HC 2004 is always the best, while at fixed  = 0.05 (panels in the diagonal), HC 2004 becomes worse when n increases. Second, Smaller s parameter (s = −1 for 18 H. ZHANG ET AL. Fig 5: The alternative distributions of HC under four non-normal settings for H0 and H1 . Column 1: n = 10; column 2: n = 100. Black solid curves: by simulations; cyan dashed curves: by Theorem 3.1; red dot-dashed curves: by Theorem 3.4. DISTRIBUTIONS OF SIGNAL-DETECTION METHODS IN FINITE CASES 19 Fig 6: Probability in (15) with boundary D(g( ni , b)) = a + λk. Black solid curves: by simulations; red dot-dashed curves: by Theorem 3.5. Table 1 Type I error rates: exact calculation vs. simulation. s 2 1 0 -1 n 10 50 100 10 50 100 10 50 100 10 50 100 10% threshold simulation threshold 3.357 3.507 3.539 2.181 2.408 2.478 1.750 2.040 2.136 1.618 1.909 2.010 4.648 4.714 4.723 2.504 2.716 2.780 1.974 2.301 2.402 1.838 2.165 2.271 0.992 0.102 0.103 0.101 0.098 0.104 0.100 0.101 0.101 0.098 0.099 0.107 5% simulation 0.049 0.050 0.049 0.050 0.048 0.049 0.049 0.047 0.051 0.051 0.049 0.052 threshold 10.088 10.102 10.102 3.110 3.300 3.354 2.390 2.803 2.915 2.227 2.662 2.777 1% simulation 0.010 0.011 0.009 0.011 0.010 0.009 0.011 0.011 0.010 0.009 0.009 0.010 20 H. ZHANG ET AL. HC 2008 and s = 0 for the reverse Berk-Jones) gives better power for denser signals. However, when signals are sparse, they are less powerful because these statistics are insensitive to small p-values as they are weighted by i/n rather than p(i) . Third, Berk-Jones statistic has a more robust performance over various µ, n and , which is consistent with the finding of Li and Siegmund [14]. Moreover, it is interesting to see the performance of these methods along the asymptotic detection boundary. This is because the optimal tests are most valuable for detecting subtle signals around the detection boundary, for which sub-optimal tests will have asymptotically zero power. Here we consider the ARW setting with the detection boundary given in (2). Figure 8 shows the power (calculated by Theorem 3.1) of the four methods over the sparsity parameter α ∈ (1/2, 1) and r calculated accordingly. It shows that the power curves of HC 2004 and B-J have the similar pattern, while the performance of HC 2008 and reverse B-J are alike. In the less sparse case with α ∈ (1/2, 3/4), B-J is preferred; in the sparser case with α ∈ (3/4, 1), HC 2004 outperforms the others. Last but not least, the supremum domain R in (6) is quite relevant in constructing the test statistic. Here we compare HC 2004 under R = {1 ≤ i ≤ n/2} with the modified HC under R = {1 < i ≤ n/2, p(i) ≥ 1/n}. Figure 9 shows that the MHC performs poorly when the number of signals is small, whereas it improves the performance when the number of signals increases. When signals are sparse, MHC is less powerful because it tends to exclude signals by considering only those p-values bigger than 1/n, the latter is a fairly large value when n is small. Thus, in practice when n is not too big, there may be no need to truncate p-values by 1/n. 5. A Genome-wide Association Study for Crohn’s Disease. According to the genetics of complex diseases, disease-associated markers usually have moderate to small effects[23]. In genome-wise studies that tend to screen as many markers as possible, the number of true disease markers often account for a small proportion of the total candidates. Therefore, it is appealing to apply tests that are asymptotically optimal for detecting rare and weak signals. In this section, based on a logistic regression model, we applied optimal tests to a SNP-set association study at the gene level. Specifically, for the ith individual, i = 1, ..., N , let yi = 1 (or 0) indicate the case (or control), let Gi = (Gi1 , ..., Gin ) denote the genotype vector of n SNPs in a given gene, and let Zi = (1, Zi1 , Zi2 ) contain the intercept and the first two principal components of the genotype data, which serve the purpose for controlling potential population structure [24]. The logis- DISTRIBUTIONS OF SIGNAL-DETECTION METHODS IN FINITE CASES 21 Fig 7: Comparison of statistical power (at type I error rate 5%). 22 H. ZHANG ET AL. Fig 8: Statistical power along the ARW detection boundary (at type I error rate 5%). Fig 9: Power comparison for HC vs. MHC (at type I error rate 5%). DISTRIBUTIONS OF SIGNAL-DETECTION METHODS IN FINITE CASES 23 tic regression model is logit(E(yi |Gi , Zi )) = G0i β + Zi0 γ, where β and γ are the coefficient parameters. A classic marginal-test statistic of the jth SNP, adjusted for the non-genetic measures, is [25, 26] Uj = N X Gij (yi − ỹi ), j = 1, ..., n, i=1 where ỹi is the fitted outcome value under H0 : β = 0, i.e., none SNPs in the gene are associated. Under H0 , the vector of statistics U = (U1 , ..., Un ) N (0, Σ), where Σ can be estimated by Σ̂ = G0 W G − G0 W Z(Z 0 W Z)−1 Z 0 W G, where G = (Gij ) and Z = (Zij ) are the corresponding design matrices, and W = diag(ỹi (1 − ỹi )). After de-correlation we get test statistics X = 1 Σ̂− 2 U N (0, In×n ). Thus the i.i.d. condition of H0 in (5) is reasonably satisfied, and our calculation methods can apply to obtain the p-value of a GOF statistic in (6), which measures the association of the given gene. We examined HC 2004 , B-J, reverse B-J, and HC 2008 (again, they corresponding to the φ-divergence statistics with s = 2, 1, 0, −1, respectively). The GWAS data from NIDDK-IBDGC (National Institute of Diabetes, Digestive and Kidney Diseases - Inflammatory Bowel Disease Genetics Consortium) contain 1,145 individuals from non-Jewish population (572 Crohn’s disease cases and 573 controls) [27]. After typical quality control for genotype data, 308,330 somatic SNPs were grouped into 15,857 genes according to their physical locations. Figure 2 gives the QQ plots of the gene-association-indicating p-values calculated by Theorem 3.1. The genomic inflation factors, i.e., the ratios of empirical median of -log(p-values) vs. the expected median under H0 , are all close to 1, evidencing that the genome-wide type I errors were well controlled. Among the four tests, the B-J seemed having higher power because it yielded more genes significantly above the red line of the H0 -expected p-values. Among the top ranked genes, IL23R and CARD15 (also known as NOD2) are well-known Crohn’s disease genes [28, 29, 27]. Gene NPTX2 was top ranked by both HC 2004 and B-J. It hasn’t been reported to be directly associated, but it encodes a member of the family of neuronal petraxins, synaptic proteins that are related to C-reactive protein [30], while C-reactive protein is an indicator for the activity level of Crohn’s disease [31]. Furthermore, NPTX2 has an important paralog gene APCS (www.genecards.org), which is related to arthritis, a disease highly correlated with Crohn’s disease [32]. Gene SLC44A4 is also related to the pathophysiology of Crohn’s. 24 H. ZHANG ET AL. Defects in this gene can cause sialidosis, a lysosomal storage disease [30] that results from a deficiency of the enzyme sialidase, the later is important for various cells to defend against infection [33]. Gene BMP2 was identified by B-J, reversed B-J, and HC 2008 . This gene could also be relevant because it is associated with digestive phenotypes, especially colon cancer [34, 35]. Meanwhile, for those top ranked genes, further studies are needed to validate. 6. Discussion. This paper provides techniques to calculate the exact and approximated null and alternative distributions of a general family of GOF tests. Thus we can calculate both the p-value and the statistical power of these tests. These calculations are important for both practical applications of GOF tests and for performance comparisons under finite n. Comparing with the relevant literature, our methods are novel, accurate, general to broad statistic family and supremum domains. To calculate the exact distribution, the result in Theorem 3.1 brings down the computational complexity from O(n3 ) to O(n2 ) when comparing with corresponding literature methods. In the meanwhile, when n and the search range k1 −k0 are large, the calculation could suffer from the loss of significant digits. In this case, we could truncate the summation to the first 25 - 30 terms, which yields a fairly accurate result and saves computation time too. Moreover, we could also apply proper approximated calculations, for which the accuracy increases together with n. APPENDIX A: PROOFS A.1. Proofs of the Main Theorems. A.1.1. Proof of Theorem 3.1. For k = k1 , ..., 1, define Z xk Z xk+1 Z 1 1 n! ak = (1 − xk1 )n−k1 ... dxk ...dxk1 −1 dxk1 . uk (n − k1 )! uk −1 uk 1 1 n! (n−k1 +1)! F̄B(1,m) (uk1 ), Then obviously ak1 = and for k ≤ k1 − 1, Z 1 Z xk Z xk+2 1 n! n−k1 ak = (1 − xk1 ) ... xk+1 dxk+1 ...dxk1 −1 dxk1 − uk ak+1 uk1 (n − k1 )! uk1 −1 uk+1 Z 1 kX 1 −k ujk+j−1 xkk11 −k n! n−k1 = (1 − xk1 ) dxk1 − ak+j (k1 − k)! j! uk (n − k1 )! j=1 1 = n! F̄ (uk ) − (n − k1 + k)! B(k1 −k+1,m) 1 kX 1 −k ujk+j−1 j=1 j! ak+j DISTRIBUTIONS OF SIGNAL-DETECTION METHODS IN FINITE CASES 25 Now by Lemma 1, P (Sn ≤ b) = P {U(k) > uk , k0 ≤ k ≤ k1 } Z 1 Z xk +1 Z xk xkk00 −1 0 1 n! n−k1 = ... (1 − xk1 ) dxk0 ...dxk1 −1 dxk1 (k0 − 1)! uk0 uk1 (n − k1 )! uk1 −1 Z xk +2 k0 Z xk Z 1 xk0 +1 uk0 0 1 n! n−k1 ... (1 − xk1 ) dxk0 +1 ...dxk1 −1 dxk1 − k0 ak0 +1 = k0 ! k0 ! uk0 +1 uk1 −1 uk1 (n − k1 )! Z 1 kX 1 −1 xk1 −1 n! uii = (1 − xk1 )n−k1 k1 dxk1 − ai+1 (k1 − 1)! i! uk (n − k1 )! i=k0 1 kX 1 −1 =F̄B(k1 ,m) (uk1 ) − i=k0 uii ai+1 . i! A.1.2. Proof of Theorem 3.2. The idea is to use total probability theorem. P (Sn,R ≤ b) = k1 n+1 X X
P {Sn,R ≤ b} ∩ {exactly p(i) ...p(j−1) fall in [α0 , α1 ]} i=1 j=i+1 = k1 n+1 X X Pij i=1 j=i+1 Notice that 1 ≤ i ≤ n, i + 1 ≤ j ≤ n + 1, we have Pij = P (U(i−1) < β0 , U(i) ≥ ui , ..., U(j−1) ≥ uj−1 , U(j−1) ≤ β1 , U(j) > β1 ), where the joint density of U(i−1) , ..., U(j) is f (xi−1 , ..., xj ) = n! xi−2 (1 − xj )n−j , (i − 2)!(n − j)! i−1 0 ≤ xi−1 ≤ ... ≤ xj ≤ 1. Then i−2 xi−1 Pij = n! ··· dxi−1 dxi ...dxj−1 dxj (i − 2)! β1 uj−1 ui 0 Z 1 Z β1 Z xi+1 Z β0 i−2 xi−1 (1 − xj )n−j = n! dxj ··· dxi ...dxj−1 dxi−1 (n − j)! (i − 2)! β1 uj−1 ui 0 Z β1 Z xi+1 = cij n! ··· dxi ...dxj−1 . Z 1 (1 − xj )n−j (n − j)! uj−1 Z β1 Z xi+1 Z β0 ui Direct calculation similar to the proof of Theorem 3.1 gives the final result. 26 H. ZHANG ET AL. A.1.3. Proof of Theorem 3.3. Following the idea in the proof of Theorem 3.2, P (Sn,R ≤ b) = k1 n+1 X X
P {Sn,R ≤ b} ∩ {exactly p(i) ...p(j−1) fall in [α0 , α1 ]} i=1 j=ĩ+1 = k1 n+1 X X Pij . i=1 j=ĩ+1 For each feasible pair of (i, j), direct calculation shows Pij can be concisely written as (β1 − xj̃−1 )j−j̃ β1 Z Pij = cij n! (j − j̃)! uj̃−1  = cij n! Z )j−i (β1 − β̃0 (j − i)! xj̃−1 Z ··· uj̃−2 xĩ+1 (xĩ − β̃0 )ĩ−i dxĩ ...dxj̃−2 dxj̃−1 (ĩ − i)! uĩ F̄B(j̃−i,j−j̃+1) ( uj̃−1 − β̃0 β1 − β̃0 )− j̃−1 X (uk − β̃0 )k−i+1 k=ĩ (k − i + 1)!  aj (k + 1) , where Z β1 aj (k) = n! uj̃−1 (β1 − xj̃−1 )j−j̃ (j − j̃)! Z xj̃−1 uj̃−2 Z xk+1 ··· uk dxk ...dxj̃−2 dxj̃−1 j̃−k−1 X ulk+l−1 uj̃−1 β1j−k = n! F̄ ( )− aj (k + 1). (j − k)! B(j̃−k,j−j̃+1) β1 l! l=1 A.1.4. Proof of Theorem 3.4. The main idea of the proofs of Theorem 3.4 and 3.5 is as follows. Note that U(i) := D(p(i) ) defined in (15) follow P i.i.d. Γi the same distribution of Γn+1 , where Γi = ij=1 εj , εj ∼ Exp(1), so that Γi ∼ Gamma(i, 1). Thus we can approximate k P {U(k) > D(g( , b)), for all k0 ≤ k ≤ k1 }. n k ≈P {Γ(k) > (n + 1)D(g( , b)), for all k0 ≤ k ≤ k1 }. n We take advantage of the joint density of (Γk0 , ..., Γk1 ), which is given by Lemma 2, while Γn+1 can be approximated by n + 1 when n is reasonably large. Accordingly, we apply similar calculation as the proof of Theorem 3.1 except using 5 and Lemma 6 instead. DISTRIBUTIONS OF SIGNAL-DETECTION METHODS IN FINITE CASES 27 A.1.5. Proof of Theorem 3.5. Note that P (Γk > a + λk, 1 ≤ k ≤ k1 ) Z Z zk Z +∞ 1 ... e−zk1 = a+λk1 a+λ(k1 −1) ! z2 dz1 ...dzk1 −1 dzk1 a+λ By lemma 4, Z +∞ (zk1 − a − (k1 − 1)λ) (zk1 − a)k1 −2 = e−zk1 dzk1 (k1 − 1)! a+λk1  Z +∞  k1 −1 k1 −2 −zk1 (zk1 − a) −zk1 (k1 − 1)λ(zk1 − a) = e dzk1 −e (k1 − 1)! (k1 − 1)! a+λk1 Let x = zk1 − a, Z +∞  Z +∞ k1 −1 k1 −2 −a −x x −x x =e e dx − λ e dx (k1 − 1)! (k1 − 2)! λk1 λk1 =e−a [1 − P (Γk1 ≤ λk1 ) − λ (1 − P (Γk1 −1 ≤ λk1 ))] =e−a (1 − λ + λFΓ(k1 −1) (λk1 ) − FΓ(k1 ) (λk1 )) =e−a (1 − λ + hk1 (λ)). Thus Theorem 3.5 is proved by combining this equation and Lemma 5. The idea of the proof of Theorem 3.6 is motivated by [14]. Instead of directly considering the distribution function, we look at the right-tail probability which can be decomposed into the union of disjoint sets. A.1.6. Proof of Theorem 3.6 and Corollary 3. Let event An,k be defined S 1 as in Lemma 7. They are disjoint and {Sn ≥ b} = kk=k An,k . In this proof 0 we mainly focus on approximating P (An,k ) Let dk = (n + 1)D(g( nk , b)), d0k = (n + 1) k dD(g( n ,b)) , dx d0k,max = (n + 1) Notice that D(g(x, b)) is convex in x, so dk+j > dk + and Lemma 7, we have for 1 ≤ k ≤ k1 − 1, 0 dk n j. From Lemma 6 dkk (Γj > dk+j , 1 ≤ j ≤ k1 − k) k! 0 dkk dk ≤ P (Γj > dk + j, 1 ≤ j ≤ k1 − k) k! n 0 0 k dk −dk dk dk = e (1 − + hk1 −k ( )) k! n n 0 0 dk dk = (1 − + hk1 −k ( ))fP (dk ) (k). n n P (An,k ) = dk1 −dk k1 −k . 28 H. ZHANG ET AL. Next we need to find the lower bound for P (An,k ). √ √ We first consider k ≥ k1 − n, here n is chosen for covenience and the proof works for any nγ , 0 < γ < 1. Note that, due to the convexity, dk+j ≤ 0 dk+√n √ dk + n j, 1 ≤ j ≤ n 0 dk+√n √ dkk P (Γj ≥ dk + j, 1 ≤ j ≤ n) k! n 0 √ k d d0k+√n dk −dk k+ n √ ≥ e (1 − + h n( ) k! n n 0 0 d d = (1 + o(1))(1 − k + h√n ( k ))fP (dk ) (k). n n P (An,k ) ≥ The last equation is due to Theorem 3.5 and the continuity of D(g(x, b)). We 0 0 √ d d can see P (An,k ) → (1 − nk + h√n ( nk ))fP (dk ) (k) uniformly in k, k ≥ k1 − n. √ We then consider k ≤ k1 − n. In this case, the proof is slightly more complicated than the first case, however the idea is similar. Let pn,k (y) = P (Γj > dk+j − dk + y, 1 ≤ j ≤ k1 − k), 0 ≤ y ≤ dk , and fΓk Rd be the density function of Γk , then P (An,k ) = 0 k pn,k (y)fΓk (dk − y)dy. Similar to the proof in the first case, pn,k (y) ≥ P (Γj ≥ y + −y ≥e (1 − d0k+√n n d0k+√n n + = (1 + o(1))e−y (1 − ,1 ≤ j ≤ √ n) − P ( k[ 1 −k √ {Γj ≤ y + dk+j − dk }) j= n h√ n( d0k+√n n )) − kX 1 −k √ j= n 0 Γj y dk,max P( ≤ + ) j j n kX 0 1 −k d0k d0 Γj y dk,max P( + h√n ( k )) − ≤ + ). n n j j n √ j= n To prove the residual uniformly in k converges to 0, we need to apply Lemma 3. For y of some lower order, say O(log n) 0 0 y dk,max log n dk,max + ≤ √ + j n n n log n n + 1 dD(g(x, b)) < √ + sup { } n k0 ≤x≤ k1 dx n n ≤δ<1 n when n is large. DISTRIBUTIONS OF SIGNAL-DETECTION METHODS IN FINITE CASES 29 Then, by Lemma 3, we have the residual resk ≤ kX 1 −k √ e −jI(δ) √ j= n e−I(δ) n → 0 exponentially uniformly in k. ≤ 1 − e−I(δ) From (1 + o(1))e−y (1 − d0k d0 d0 d0 + h√n ( k )) − resk ≤ pn,k ≤ e−y (1 − k + h√n ( k )), n n n n we can conclude that d0 d0k + h√n ( k )) uniformly in k, n n √ n −I(δ) e with error bound resk ≤ . 1 − e−I(δ) pn,k (y) → e−y (1 − Let α denote log n for simplicity, α > 1, then, for some constant c > 1, (16) Z min{dk ,cα} d0 d0 e−y (1 − k + h√n ( k ))fΓk (dk − y)dy Iˆn,k = n n 0 √ Z min{dk ,cα} ne−I(δ) n →In,k = pn,k (y)fΓk (dk − y)dy uniformly with error bound ≤ . 1 − e−I(δ) 0 When dk > cα, dk − y < dk < k − 1, fΓk (dk − y) is decreasing in y P (An,k ) − In,k (dk − cα)fΓk (dk − cα) ≤ In,k αfΓk (dk − α) (dk − cα)ecα−dk (dk − cα)k−1 αeα−dk (dk − α)k−1 (dk − cα) (c−1)α (c − 1)α k−1 = e (1 − ) α dk − α (dk − cα) (c−1)α (k−1) log(1− (c−1)α ) dk −α = e e α (c−1)α dk −(k−1)( d −α ) k ≤ eα(c−1) e α dk α(c−1)(1− k−1 ) dk ≤ e α k Let x = , y = D(g(x, b)), then n nx−1 (n + 1)y −(c−1)( (n+1)y −1) = n . α = 30 H. ZHANG ET AL. Notice that nx−1 (n+1)y is c = sup 2 ≤x≤ k1 n n nx−1 > 1, we need (c − 1)( (n+1)y − 1) > 1 for all x, that nx−1 nx−1−(n+1)D(g(x,b)) > 1. Therefore P (An,k )−In,k In,k ≤ y −γ αn , γ > 0. For k = 1, the result follows d1 → 0, (17) In,k → P (An,k ) uniformly in k as n → ∞. Define P̂ (An,k ) = R dk 0 e−y (1 − P̂ (An,k ) − Iˆn,k P̂ (An,k ) d0k n d0 + h√n ( nk ))fΓk (dk − y)dy, (dk −y)k−1 (k−1)! dy R dk (dk −y)k−1 0 (k−1)! dy dk − (dk − cα)k ) ( k dkk k R min{dk ,cα} =1− =1− 0  cα = 1− dk   k ≤ exp −c log n . dk Since c, dkk > 1, we have (18) Iˆn,k → P̂ (An,k ) uniformly in k as n → ∞. k−1 −y) Finally, notice that e−y fΓk (dk − y) = e−dk (dk(k−1)! , then (19) d0 d0 P̂ (An,k ) = (1 − k + h√n ( k ))e−dk n n 0 Z dk 0 (dk − y)k−1 dy (k − 1)! 0 d d = (1 − k + h√n ( k ))fP (dk ) (k). n n Combine lemma 5 and equations (16), (17), (18), and (19), we have P (Sn ≥ b) =(1 + o(1))P (Γk ≤ dk , for some k0 ≤ k ≤ k1 ) =(1 + o(1)) k1 X P̂ (An,k ) k=k0 =(1 + o(1)) k1 X k=k0 √ where k ∗ = min{k1 − k, n}. (1 − d0k d0 + hk∗ ( k ))fP (dk ) (k), n n DISTRIBUTIONS OF SIGNAL-DETECTION METHODS IN FINITE CASES 31 To proof Corollary 3, by Theorem 3.6, we have  k1  X n+1 0 k n+1 0 k P (HC 2004 ≥ b) = (1+o(1)) 1− g ( , b0 ) + hk∗ ( g ( , b0 )) fP ((n+1)g( k ,b0 )) (k) . n n n n n k=k0 A.2. Fundmental Lemmas and Proofs. Lemma 1. Let U(i) be the ith order statistic of i.i.d. samples from Uniform(0,1), 1 ≤ k0 < k1 ≤ n. Then the joint density of (U(k0 ) , ..., U(k1 ) ) is f(U(k 0) ,...,U(k1 ) ) (zk0 , ..., zk1 ) = n! z k0 −1 (1 − zk1 )n−k1 , 0 < zk0 < ... < zk1 < 1. (n − k1 )!(k0 − 1)! k0 Proof. Standard result from formula (2.2.2) of the book Order Statistics[36]. Lemma Pk2. Let εi be i.i.d. exponential random variables with parameter 1. Γk = i=1 εi , 1 ≤ k0 < k1 ≤ n. Then the joint density of (Γk0 , ..., Γk1 ) f (zk0 , ..., zk1 ) = zkk00 −1 (k0 − 1)! e−zk1 , 0 < zk0 < ... < zk1 < ∞. Specially, when k0 = 1, the joint density is f (z1 , ..., zk1 ) = e−zk1 , 0 < z1 < ... < zk1 < ∞. Proof. The proof follows the deduction process shown in Theorem 1.1 by Mathai and Moschopoulos [37]. Note that their theorem only considers the case of k0 = 1, but the deduction idea can be applied to general k0 . P Lemma 3. Let c < 1, I(c) = c − 1 − log(c) > 0, Γn = ni=1 i , where i are i.i.d exponential distributed with parameter 1, then P( Γn ≤ c) ≤ e−nI(c) . n Proof. Let i be exp(1) distributed. The log moment generating function of −i is Λ(t) = − log(1 + t), t > −1. Then the convex rate function is Λ∗ (x) = sup tx − Λ(t) t>−1 = −x − 1 − log(−x). By Cramer’s theorem, let x = −c, we have desired result. 32 H. ZHANG ET AL. Lemma 4. Z (Abel-Goncharov Polynomial) zk 1 Z z2 dz1 ...dzk1 −1 = ... a+λ(k1 −1) a+λ (zk1 − a − λ(k1 − 1)) (zk1 − a)k1 −2 . (k1 − 1)! Proof. We will prove by induction. When k1 = 2, it’s easily shown that Z z2 (z2 − a − λ(2 − 1)) (z2 − a)2−2 dz1 = z2 − a − λ = . (2 − 1)! a+λ Assume it’s true for k1 = k, Z zk Z z2 (zk − a − λ(k − 1)) (zk − a)k−2 . ... dz1 ...dzk−1 = (k − 1)! a+λ(k−1) a+λ Then for k1 = k + 1, Z zk+1 Z zk Z z2 ... a+λk Z zk+1 = a+λk zk+1 Z = a+λk a+λ(k−1) dz1 ...dzk−1 dzk a+λ (zk − a − λ(k − 1)) (zk − a)k−2 dzk (k − 1)! Z zk+1 (zk − a)k−1 λ(zk − a)k−2 dk − dk (k − 1)! (k − 2)! a+λk (λk)k λ(zk+1 − a)k−1 λ(λk)k−1 (zk+1 − a)k − − + = k! k! (k − 1)! (k − 1)! k−1 (zk+1 − a) (zk+1 − a − λk) . = k! This finished the proof. Lemma 5. Let c = (ck0 , ..., ck1 ), 1 ≤ k0 < k1 ≤ n be a finite sequence of increasing numbers. Then given  > 0, as n → ∞ P( Γj > cj , k0 ≤ j ≤ k1 ) = (1 + O())P (Γj > (n + 1)cj , k0 ≤ j ≤ k1 ). Γn+1 Proof. Given  > 0, let L(c; n, ) and R(c; n, ) be the lower bound and DISTRIBUTIONS OF SIGNAL-DETECTION METHODS IN FINITE CASES 33 Γ j > cj , k0 ≤ j ≤ k1 ), upper bound for P ( Γn+1 L(c; n, ) =P (Γj > (1 + )(n + 1)cj , k0 ≤ j ≤ k1 ) Γn+1 −P ( ∈ [1 − , 1 + ]c ), n+1 R(c; n, ) =P (Γj > (1 − )(n + 1)cj , k0 ≤ j ≤ k1 ) Γn+1 ∈ [1 − , 1 + ]c ). +P ( n+1 By Lemma 3, P( Γn+1 1 ∈ [1 − , 1 + ]c ) ≤ 2 e−nΛ(1−) → 0 exponentially. n+1  Because of the continuity of the multivariate gamma distribtution, L(c; n, ) →(1 + O())P (Γj > (n + 1)cj , k0 ≤ j ≤ k1 ), R(c; n, ) →(1 + O())P (Γj > (n + 1)cj , k0 ≤ j ≤ k1 ). Lemma 6. Let (d1 , ..., dk1 ) be a sequence of nondecreasing and nonnegative numbers. Qj (dj+1 , ..., dk1 ) = P {Γk ≥ dk+j , 1 ≤ k ≤ k1 − j}, 0 ≤ j ≤ k1 − 1. F̄Γk (x) is the survival function of Gamma distribution with shape parameter k and scale parameter 1. Then for l = 2, 3, ..., k1 Qk1 −l = F̄Γl (dk1 ) − l−1 j X dk 1 −l+j j=1 j! Qk1 −l+j . with Qk1 −1 = F̄Γ1 (dk1 ), and, for k0 ≥ 1, the joint survival probability P {Γk ≥ dk , k0 ≤ k ≤ k1 } = F̄Γk1 (dk1 ) − kX 1 −1 j=k0 djj j! Qj . 34 H. ZHANG ET AL. Proof. Q0 = P {Γk ≥ dk , 1 ≤ k ≤ k1 } Z z2 Z zk Z +∞ 1 −zk1 dz1 ...dzk1 −1 dzk1 ... e = +∞ = e−zk1 Z +∞ e = −zk1 Z +∞ = e −zk1 Z +∞ = dk1 e−zk1 zk 1 Z z2 dz2 ...dzk1 −1 dzk1 − d1 Q1 d2 Z ... djj j! zk0 +1 dk0 (k1 − 1)! j=1 z3 ... zk 1 kX 1 −1 (z2 − d1 )dz2 ...dzk1 −1 dzk1 d2 zkk11 −1 = F̄Γk1 (dk1 ) − z3 ... dk1 −1 dk1 Z Z dk1 −1 dk1 Z zk 1 dk1 −1 dk1 Z d1 dk1 −1 dk1 Z dzk1 − zkk00 −1 (k0 − 1)! kX 1 −1 j=1 djj j! dzk0 ...dzk1 −1 dzk1 − kX 0 −1 j=1 djj j! Qj Qj Qj . The rest of recursive formulae can be derived similarly and from the last third equation. Also notice that {Γk ≥ dk , k0 ≤ k ≤ k1 } = {Γk ≥ dk , 1 ≤ S k ≤ k1 } {some Γk < dk , 1 ≤ k ≤ k0 − 1 and Γk ≥ dk , k0 ≤ k ≤ k1 }, by Lemma 7 we have P {Γk ≥ dk , k0 ≤ k ≤ k1 } = Q0 + kX 0 −1 j=1 djj j! = F̄Γk1 (dk1 ) − Qj kX 1 −1 j=k0 djj j! Qj . Lemma 7. Let (d1 , ..., dk1 ) be a sequence of increasing numbers, An,k = {Γk ≤ dk , Γk+l > dk+l , 1 ≤ l ≤ k1 − k}, 1 ≤ k ≤ k1 − 1. Then P (An,k ) = dkk Qk . k! DISTRIBUTIONS OF SIGNAL-DETECTION METHODS IN FINITE CASES 35 Proof. P (An,k ) = P {Γk ≤ dk , Γk+l > dk+l for all 1 ≤ l ≤ k1 − k} Z zk+1 Z dk Z zk Z +∞ 1 zkk−1 ... e−zk1 = dzk dzk+1 ...dzk1 −1 dzk1 (k − 1)! 0 dk+1 dk1 −1 dk1 Z zk+1 k Z zk Z +∞ 1 dk −zk1 ... e = dzk+1 ...dzk1 −1 dzk1 dk+1 k! dk −1 dk 1 1 dk = k P {Γl ≥ dk+l , 1 ≤ l ≤ k1 − k} k! dk = k Qk . k! REFERENCES [1] D. Donoho and J. Jin, “Higher criticism for detecting sparse heterogeneous mixtures,” The Annals of Statistics, vol. 32, no. 3, pp. 962–994, 2004. [2] D. Donoho and J. Jin, “Higher criticism thresholding: Optimal feature selection when useful features are rare and weak,” Proceedings of the National Academy of Sciences of the United States of America, vol. 105, pp. 14790–14795, Sep 30 2008. [3] D. Donoho and J. Jin, “Higher criticism for large-scale inference: especially for rare and weak effects,” Statistical Science, vol. 30, no. 1, pp. 1–25 DOI: 10.1214/14– STS506, 2015. [4] L. Jager and J. A. Wellner, “Goodness-of-fit tests via phi-divergences,” The Annals of Statistics, pp. 2018–2053, 2007. [5] M. Noé, “The calculation of distributions of two-sided kolmogorov-smirnov type statistics,” The Annals of Mathematical Statistics, pp. 58–64, 1972. [6] M. Noé and G. Vandewiele, “The calculation of distributions of kolmogorov-smirnov type statistics including a table of significance points for a particular case,” The Annals of Mathematical Statistics, vol. 39, no. 1, pp. 233–241, 1968. [7] V. Kotel’Nikova and E. Chmaladze, “On computing the probability of an empirical process not crossing a curvilinear boundary,” Theory of Probability & Its Applications, vol. 27, no. 3, pp. 640–648, 1983. [8] G. R. Shorack and J. A. Wellner, Empirical processes with applications to statistics, vol. 59. SIAM, 2009. [9] G. Steck, “The smirnov two sample tests as rank tests,” The Annals of Mathematical Statistics, pp. 1449–1466, 1969. [10] M. Breth, “On a recurrence of steck,” Journal of Applied Probability, pp. 823–825, 1976. [11] H. Ruben, “On the evaluation of steck’s determinant for rectangle probabilities of uniform order statistics,” Communications in Statistics-Theory and Methods, vol. 5, no. 6, pp. 535–543, 1976. [12] I. J. Barnett and X. Lin, “Analytical p-value calculation for the higher criticism test in finite-d problems,” Biometrika, vol. 101, no. 4, pp. 964–970, 2014. 36 H. ZHANG ET AL. [13] A. Moscovich, B. Nadler, C. Spiegelman, et al., “On the exact berk-jones statistics and their p-value calculation,” Electronic Journal of Statistics, vol. 10, no. 2, pp. 2329– 2354, 2016. [14] J. Li and D. Siegmund, “Higher criticism: p-values and criticism,” Annals of Statistics, vol. 43, no. 3, pp. 1323–1350, 2015. [15] F. Eicker, “The asymptotic distribution of the suprema of the standardized empirical processes,” The Annals of Statistics, pp. 116–138, 1979. [16] D. Jaeschke, “The asymptotic distribution of the supremum of the standardized empirical distribution function on subintervals,” The Annals of Statistics, vol. 7, no. 1, pp. 108–115, 1979. [17] Y. I. Ingster, “Some problems of hypothesis testing leading to infinitely divisible distributions,” Mathematical Methods of Statistics, vol. 6, no. 1, pp. 47–69, 1997. [18] Y. I. Ingster, “Minimax detection of a signal for in -balls,” Mathematical Methods of Statistics, vol. 7, no. 4, pp. 401–428, 1998. [19] J. Tukey, “The higher criticism.” Course Notes, Statistics 411, Princeton University., 1976. [20] E. Arias-Castro, E. J. Candès, and Y. Plan, “Global testing under sparse alternatives: Anova, multiple comparisons and the higher criticism,” The Annals of Statistics, vol. 39, no. 5, pp. 2533–2556, 2011. [21] J. Shao, Mathematical Statistics. Springer Verlag, 2010. [22] R. H. Berk and D. H. Jones, “Goodness-of-fit test statistics that dominate the kolmogorov statistics,” Probability Theory and Related Fields, vol. 47, no. 1, pp. 47–59, 1979. [23] D. B. Goldstein, “Common genetic variation and human traits,” New England Journal of Medicine, vol. 360, no. 17, pp. 1696–1698, 2009. [24] A. L. Price, N. J. Patterson, R. M. Plenge, M. E. Weinblatt, N. A. Shadick, and D. Reich, “Principal components analysis corrects for stratification in genome-wide association studies,” Nature genetics, vol. 38, no. 8, pp. 904–909, 2006. [25] P. McCullagh and J. A. Nelder, Generalized linear models, vol. 37. CRC press, 1989. [26] D. J. Schaid, C. M. Rowland, D. E. Tines, R. M. Jacobson, and G. A. Poland, “Score tests for association between traits and haplotypes when linkage phase is ambiguous,” The American Journal of Human Genetics, vol. 70, no. 2, pp. 425–434, 2002. [27] R. Duerr, K. Taylor, S. Brant, J. Rioux, M. Silverberg, M. Daly, A. Steinhart, C. Abraham, M. Regueiro, A. Griffiths, et al., “A genome–wide association study identifies il23r as an inflammatory bowel disease gene,” Science Signalling, vol. 314, no. 5804, p. 1461, 2006. [28] J.-P. Hugot, M. Chamaillard, H. Zouali, S. Lesage, J.-P. Cézard, J. Belaiche, S. Almer, C. Tysk, C. A. O’Morain, M. Gassull, et al., “Association of nod2 leucine-rich repeat variants with susceptibility to crohn’s disease,” Nature, vol. 411, no. 6837, pp. 599– 603, 2001. [29] Y. Ogura, D. K. Bonen, N. Inohara, D. L. Nicolae, F. F. Chen, R. Ramos, H. Britton, T. Moran, R. Karaliuskas, R. H. Duerr, et al., “A frameshift mutation in nod2 associated with susceptibility to crohn’s disease,” Nature, vol. 411, no. 6837, pp. 603–606, 2001. [30] D. Maglott, J. Ostell, K. D. Pruitt, and T. Tatusova, “Entrez gene: gene-centered information at ncbi,” Nucleic acids research, vol. 39, no. suppl 1, pp. D52–D57, 2011. [31] P. Chamouard, Z. Richert, N. Meyer, G. Rahmi, and R. Baumann, “Diagnostic value of c-reactive protein for predicting activity level of crohns disease,” Clinical Gastroenterology and Hepatology, vol. 4, no. 7, pp. 882–887, 2006. [32] G. Trikudanathan, P. G. Venkatesh, and U. Navaneethan, “Diagnosis and therapeu- DISTRIBUTIONS OF SIGNAL-DETECTION METHODS IN FINITE CASES 37 [33] [34] [35] [36] [37] tic management of extra-intestinal manifestations of inflammatory bowel disease,” Drugs, vol. 72, no. 18, pp. 2333–2349, 2012. W. D. James, T. Berger, and D. Elston, Andrew’s diseases of the skin: clinical dermatology. Elsevier Health Sciences, 2011. S. Yuvaraj, S. H. Al-Lahham, R. Somasundaram, P. A. Figaroa, M. P. Peppelenbosch, and N. A. Bos, “E. coli-produced bmp-2 as a chemopreventive strategy for colon cancer: a proof-of-concept study,” Gastroenterology research and practice, vol. 2012, 2012. M. L. Slattery, A. Lundgreen, J. S. Herrick, S. Kadlubar, B. J. Caan, J. D. Potter, and R. K. Wolff, “Genetic variation in bone morphogenetic protein and colon and rectal cancer,” International Journal of Cancer, vol. 130, no. 3, pp. 653–664, 2012. H. David and H. Nagaraja, “Order statistics,” 2003. A. M. Mathai and P. G. Moschopoulos, “On a multivariate gamma,” Journal of Multivariate Analysis, vol. 39, no. 1, pp. 135–153, 1991. Department of Mathematical Sciences and Program of Bioinformatics and Computational Biology and Program of Data Science 100 Institute Road Worcester, MA 01609 E-mail: zheyangwu@wpi.edu URL: https://www.wpi.edu/people/faculty/zheyangwu 

10math.ST

1 Testing Optimality of Sequential Decision-Making Meik Dörpinghaus, Member, IEEE, Izaak Neri, Édgar Roldán, arXiv:1801.01574v1 [cs.IT] 4 Jan 2018 Heinrich Meyr, Life Fellow, IEEE, and Frank Jülicher Abstract This paper provides a statistical method to test whether a system that performs a binary sequential hypothesis test is optimal in the sense of minimizing the average decision times while taking decisions with given reliabilities. The proposed method requires samples of the decision times, the decision outcomes, and the true hypotheses, but does not require knowledge on the statistics of the observations or the properties of the decision-making system. The method is based on fluctuation relations for decision time distributions which are proved for sequential probability ratio tests. These relations follow from the martingale property of probability ratios and hold under fairly general conditions. We illustrate these tests with numerical experiments and discuss potential applications. This work has been partly supported by the German Research Foundation (DFG) within the Cluster of Excellence EXC 1056 ’Center for Advancing Electronics Dresden (cfaed)’ and within the CRC 912 ’Highly Adaptive Energy-Efficient Computing (HAEC)’. The material in this paper has been presented in part at the IEEE International Symposium on Information Theory (ISIT), Aachen, Germany, June 2017 [1]. M. Dörpinghaus is with the Vodafone Chair Mobile Communications Systems and with the Center for Advancing Electronics Dresden (cfaed), Technische Universität Dresden, 01062 Dresden, Germany (e-mail: meik.doerpinghaus@tu-dresden.de). Izaak Neri is with the Max-Planck-Institute for the Physics of Complex Systems, Dresden, Germany, with the Max-PlanckInstitute of Molecular Cell Biology and Genetics, Dresden, Germany, and with the Center for Advancing Electronics Dresden (cfaed), Technische Universität Dresden, 01062 Dresden, Germany (e-mail: izaakneri@posteo.net). Édgar Roldán is with the Max-Planck-Institute for the Physics of Complex Systems, Dresden, Germany, with the Center for Advancing Electronics Dresden (cfaed), Technische Universität Dresden, 01062 Dresden, Germany, and with GISC – Grupo Interdisciplinar de Sistemas Complejos, Madrid, Spain (e-mail: edgar@edgarroldan.com). H. Meyr is an emeritus of the Institute for Integrated Signal Processing Systems, RWTH Aachen University, 52056 Aachen, Germany and is now a grand professor of the Center for Advancing Electronics Dresden (cfaed) at Technische Universität Dresden, 01062 Dresden, Germany (e-mail: meyr@iss.rwth-aachen.de). Frank Jülicher is with the Max-Planck-Institute for the Physics of Complex Systems, Dresden and with the Center for Advancing Electronics Dresden (cfaed), Technische Universität Dresden, 01062 Dresden, Germany, Germany (e-mail: julicher@pks.mpg.de). January 8, 2018 DRAFT 2 Index Terms Sequential hypothesis testing, sequential probability ratio test, sequential analysis, decision-making, mutual information. I. I NTRODUCTION In sequential decision-making it is important to make fast and reliable decisions. In this regard, consider, e.g., an autonomous car which has to decide whether an obstacle is present or not on the road. Such decisions are executed by dedicated signal processing algorithms. These algorithms should use the available measurements in an optimal way such that the average time to take a decision is minimized. For practical use it is key to test if the implemented decision algorithms achieve the optimum performance, i.e., if decisions are made as fast as possible with a given reliability. Sequential decision-making has been first mathematically formulated in the seminal work by A. Wald who introduced a sequential probability ratio test [2]. Wald’s test takes binary decisions on two hypotheses based on sequential observations of a stochastic process. For independent and identically distributed (i.i.d.) observations this test yields the minimum mean decision time for decisions with a given probability of error and a given hypothesis [3]. Wald’s test accumulates the likelihood ratio given by the sequence of observations and decides as soon as this cumulative likelihood ratio exceeds or falls below two given thresholds which depend on the required reliability of the decision. A key characteristic of such a sequential test is that its termination time is a random quantity depending on the actual realization of the observation sequence. The Wald test has been applied to non i.i.d. observation processes, nonhomogeneous and correlated continuous-time processes, and has been generalized for multiple hypotheses [4]; general optimality criteria for sequential probability ratio tests have been proved when probabilities of errors tend to zero, see e.g. [4]–[8]. Now we consider the decision-making device as a black box which takes as input the observation process, corresponding to one of two hypotheses, and gives as output a binary decision variable at a random decision time. Can we determine whether this decision-making device is optimal based on the statistics of the output of the device — the decisions and the decision times — and the knowledge of the true hypothesis? Indeed, in the present paper we introduce a test for optimality of sequential decision-making based on necessary conditions for optimality. January 8, 2018 DRAFT 3 Notably, this test does not require knowledge of the realizations or the statistics of the observation processes. We first consider a device which takes as input the realization of a continuous stochastic process corresponding to one of the two hypothesis H1 or H2 , and gives as output a binary decision variable D ∈ {1, 2} (corresponding to the hypotheses H1 and H2 , respectively) at the random decision time T elapsed since the beginning of the observations. We will show that optimality of sequential probability ratio tests — in the sense that the mean decision time is minimized while fulfilling given reliability constraints — requires that the following conditions on the distribution of the decision time T hold pT (t|H = 1, D = 1) = pT (t|H = 2, D = 1) (1) pT (t|H = 1, D = 2) = pT (t|H = 2, D = 2). (2) Here, pT is the probability density of the decision time and H ∈ {1, 2} (corresponding to the hypotheses H1 and H2 ) denotes the random binary hypothesis. The necessary conditions (1) and (2) for optimality imply that the distribution of the decision time T given a certain outcome are independent of the actual hypothesis. Moreover, this implies that the decision time T of the optimal sequential test does not contain any information on which hypothesis is true beyond the decision outcome D. As a consequence, we can quantify the optimality of a given black-box test by measuring the mutual information between the hypothesis H and the decision time T conditioned on the output of the test D, i.e., I(H; T|D). In case the test is optimal it must hold that I(H; T|D) = 0. (3) Based on the following example it can be seen that (3) is a necessary but not a sufficient condition for optimality in the sense of minimizing the mean decision time given a certain reliability. Consider we have an optimal decision device using the Wald test. Now we delay all decisions by a constant time tdelay . Still I(H; T + tdelay |D) = 0 with T being the decision time of the Wald test. Indeed, (3) is not a sufficient condition for the minimal mean decision time, but rather a measure for the optimal usage of information by the decision device. If I(H; T|D) > 0 this means that the decision time T contains additional information on the hypothesis H beyond the actual decision D implying that the decision device does not exploit all the available information. Hence, I(H; T|D) measures the degree of divergence from optimality January 8, 2018 DRAFT 4 in the sense of optimal usage of information. For practical purposes it is easier to test whether the black-box decision device fulfills the optimality condition in (3) rather than testing if the decision device decides with the minimum mean decision time, since the minimum mean decision time is in general not known. Furthermore, consider that in experimental setups we can measure a decision output D at a certain time T + Tdelay , with Tdelay a random delay time, but in general we do not know at which time T the decision has been taken, as the decision device is a black box device and we cannot clearly separate the actual decision making process from nondecision processes. If the decision time Tdelay is statistical independent of H when conditioned on D and T, then I(H; T|D) = 0 implies that I(H; T + Tdelay |D) = 0, and I(H; T + Tdelay |D) can be used as a necessary condition to test optimality of decision devices. If additionally I(H; Tdelay |D, T + Tdelay ) = 0, then I(H; T|D) = I(H; T + Tdelay |D). In this paper we derive the optimality conditions (1) - (3) and generalize them to discrete-time observation processes. We furthermore formulate tests for optimal sequential hypothesis testing (sequential decisionmaking) based on (1) - (3). Finally, we illustrate our results in computer experiments. The optimality conditions (1) - (3) are of interest in different contexts. For example these conditions allow to test optimality of sequential decision-making in engineered devices. An example are decision-making devices based on machine learning such as deep neural networks. Such algorithms have the power to solve very complex tasks and adapt to specific environments by learning. The advantage of machine learning is that not all environmental situations need to be learned at design time, which for many applications like self driving cars is not practical. However, while neural networks exhibit best performance in comparison to other approaches the principles which lead to their successful operation are yet unclear. It would be very useful to quantify if decisions made on such deep learning approaches are close to optimal. In addition to engineering, the tests proposed in this paper could also allow to understand if specific biological systems use all available information optimally to make reliable decisions on the fly. In this regard, consider the example of sequential decision-making by humans in two-choice decision tasks based on perceptual stimuli or biological cells making decisions on their fate based on extracellular cues. We discuss the application of the tests for optimality to these examples in more detail below. Notation: We denote random variables by upper case sans serif letters, e.g., X. All random quantities are defined on the measurable space (Ω, F) and are governed by the probability measure P. The probability density function of a random variable X given Y = y is written January 8, 2018 DRAFT 5 pX (x|Y = y). Moreover, for discrete random variables P (X = x|Y = y) denotes the probability of X = x given Y = y. The restriction of the measure P to a sub-σ algebra G ⊆ F is written as P|G . Finally, log denotes the natural logarithm and log2 is the logarithm w.r.t. base 2. The mutual i h information and the conditional mutual information are defined by I(X; Y) = E log2 ppXX(X|Y) (X) h i pX (X|Y,Z) and I(X; Y|Z) = E log2 pX (X|Z) , respectively, where the mathematical expectation E[·] is taken with respect to the measure P. Organization of the Paper: The paper is organized as follows. After the introduction, we describe the system setup in detail in Section II where we also give a precise problem formulation including definitions of optimality for sequential decision-making. Subsequently, in Section III we derive the main theorems and corollaries describing properties of optimal sequential probability ratio tests for the case of continuous observation processes. In Section IV for certain conditions we extend these theorems and corollaries to the discrete-time scenario. In Section V we formulate statistical tests to decide whether a black-box decision device performs optimal sequential decision-making based on the theorems and corollaries derived in Section III and IV, and we also discuss how to measure the distance of the black-box decision device to optimality. We illustrate the application of these tests based on numerical experiments in Section VI. In Section VII we discuss the applicability and the limitations of the provided tests for optimality. Readers who are mainly interested in the application of the statistical tests for optimality may skip Sections III and IV. II. S YSTEM S ETUP AND P ROBLEM F ORMULATION A. System Setup We consider a sequential binary decision problem based on an observation process Xt with the time index t either discrete, t ∈ Z+ , or continuous, t ∈ R+ . The stochastic process Xt is generated by one of two possible models corresponding to two hypotheses H1 and H2 . To describe the statistics of the process Xt we consider the filtered probability space (Ω, F, {Ft }t≥0 , P) with {Ft }t≥0 the natural filtration generated by the observation process Xt and the hypothesis H. We consider H to be a time independent random variable. The statistics of the observation process under the two hypothesis are described by the conditional probability measures given the hypothesis Pl [Φ] = E [1Φ |H = l] with l ∈ {1, 2} corresponding to the hypothesis H1 and H2 , respectively [9]; here 1Φ (ω) is the indicator function on the set Φ. We also consider the filtered probability spaces (Ω, F, {Ft }t≥0 , Pl ) with l ∈ {1, 2} associated with the two hypotheses. We January 8, 2018 DRAFT 6 consider for continuous-time processes that the filtration {Ft }t≥0 is right-continuous [4], i.e., Ft = ∩s>t Fs for all times t ∈ R+ . A sequential test makes binary decisions based on sequential observations of the process Xt and tries to guess which of the hypotheses H1 and H2 is true. A sequential test δ = (D, T) returns a binary output D at a random time T. The decision time T is a stopping time, which is determined by the time when Xt satisfies for the first time a certain criterion. This stopping rule is non-anticipating in the sense that it depends only on observations of the input sequence up to the current time, i.e., the decision causally depends on the observation process. The decision function is a map from the trajectory {Xt }T0 to {1, 2}, which determines the decision of the test. We now consider the following class of sequential tests with given reliabilities C(α1 , α2 ) = {δ : P (D = 2|H = 1) ≤ α2 , P (D = 1|H = 2) ≤ α1 , E[T|H = i] < ∞, i ∈ {1, 2}} (4) where E[T|H = i] denotes the expected termination time in case hypothesis i is true and where the expectation is taken over the observation sequences Xt . Moreover, α1 and α2 are the maximum allowed error probabilities of the two error types. We assume that α1 , α2 < 0.5. Notice that we restrict ourselves to tests which terminate almost surely. This assumption is fulfilled in many cases like the case of i.i.d. observation processes [10, Th. 6.2-1] and stationary observation processes. Note that the class of sequential tests given by C(α1 , α2 ) does not consider prior knowledge on the statistics of H. We define the following optimality criterion. Definition 1 (Optimality in terms of mean decision times). An optimal test δ ∗ = (D∗ , T∗ ) minimizes the two mean decision times E[T∗ |H = i] corresponding to the hypothesis i = 1 and i = 2 for a given reliability, i.e. E[T∗ |H = i] = inf δ∈C(α1 ,α2 ) E[T|H = i], i = 1, 2. (5) Note that in Definition 1 we assume that there exists a test for which the infimum is attained. If such a test does not exist than we can still find a test for which the two mean decision times are arbitrarily close to their infimum values. Sequential probability ratio tests or Wald-tests are optimal in the sense of Definition 1 [2]. It has been proved that for the case of time-discrete i.i.d. observation processes the Wald test is January 8, 2018 DRAFT 7 optimal in the sense of Definition 1 [3]. Furthermore, under broad conditions for the observation process it has been proved that sequential probability ratio tests are optimal in the sense of Definition 1 in the limit of small error probabilities [4]–[8]. For discrete-time and i.i.d. processes the Wald test collects observations Xt (which can be understood as samples of a corresponding continuous-time process) until the cumulated loglikelihood ratio Sk = k X ∆n = k X n=1 n=1  log pX (Xn |H = 1) pX (Xn |H = 2)  for k ≥ 1 (6) exceeds (falls below) a prescribed threshold L1 (L2 ) for the first time. In (6) ∆n are the increments of the log-likelihood ratio at time instant n. The test decides D = 1 (D = 2), i.e., for H1 (H2 ), when Sk first crosses L1 (L2 ). In (6), pX (·|H) denotes the probability density function of the observations Xk conditioned on the event H. The thresholds L1 and L2 depend on the maximum allowed probabilities for making a wrong decision α1 and α2 . A decision with the given reliability constraints α1 and α2 can be made when the cumulative log-likelihood ratio Sk for the first time crosses one of the thresholds before crossing the opposite one. The thresholds are functions of α1 and α2 . In general, the thresholds L1 and L2 are difficult to obtain. However, the optimal thresholds yielding the minimum mean decision time can be approximated by [10, p. 148] 1 − α2 α1 α2 . L2 = log 1 − α1 L1 = log (7) (8) The choice in (7) and (8) still guarantees that the error constraints in (4) are fulfilled. In summary, the sequential probability ratio test decides at the time TWald = min{k ∈ N : Sk ∈ / (L2 , L1 )} (9) for the decision DWald   1 if S TWald ≥ L1 =  2 if ST ≤ L2 . Wald (10) Analogously, the Wald test for non-i.i.d. observation processes is given by (9) and (10) with the log-likelihood ratio k X k X pXn (Xn |X1n−1 , H = 1) Sk = ∆n = log pXn (Xn |X1n−1 , H = 2) n=1 n=1 for k ≥ 1 (11) where Xn−1 = [X1 , . . . , Xn−1 ]. 1 January 8, 2018 DRAFT 8 The Wald test can also be formulated for continuous-time observation processes. In this case probability densities of the observation trajectories do not always exist. However, the likelihood ratio eSt can be defined in terms of the Radon-Nikodým derivative of the probability space (Ω, F, {Ft }t≥0 , P1 ) with respect to the probability space (Ω, F, {Ft }t≥0 , P2 ): eSt = dP1 |Ft dP2 |Ft (12) with t ≥ 0. Here, the process St is the cumulative log-likelihood ratio and Pi |Ft (i = 1, 2) are restricted measures of Pi w.r.t. the σ-algebra Ft . A decision with the given reliability constraints α1 and α2 can be made when the cumulative log-likelihood ratio St for the first time crosses one of the thresholds before crossing the opposite one. I.e., the test decides D = 1 (D = 2) in case it crosses L1 (L2 ) for the first time before crossing L2 (L1 ) where the thresholds are functions of α1 and α2 . For continuous observation processes St is continuous and the thresholds are exactly given by (7) and (8), see, e.g., [10, p. 148]. Therefore, the Wald test for continuous-time observation processes is defined by Tdec = inf{t ∈ R+ : St ∈ / (L2 , L1 )} (13) with the decision output given by Ddec   1 if S Tdec ≥ L1 =  2 if ST ≤ L2 . dec (14) B. Problem Statement Consider now the black-box decision device as illustrated in Fig. 1 for which the stochastic observation process Xt and the algorithm of the decision device are both unknown. Such a blackbox decision device is a sequential test δ for which the function (D, T) is unknown. We ask now the question: Is it possible to determine whether such a black-box decision device is optimal in the sense of Definition 1 based on many outcomes D and T of the device? Having access to the decision outcomes and decision times it is impossible to verify optimality in terms of Definition 1. In this regard consider that the value of the minimum mean decision time is typically unknown since the observed process Xt and its statistics are often not known. We thus introduce the following alternative definition of optimality, which is based on the idea that optimal sequential decision-making needs to exploit the available information optimally. January 8, 2018 DRAFT 9 Stochastic process Decision device Input Decision output Output Decision time Fig. 1. Black-box model of binary sequential decision-making: a decision device observes sequentially a stochastic process Xt until it takes a decision D = 1 (D = 2) corresponding to the hypothesis H = 1 (H = 2) at a random decision time T. Definition 2 (Optimality in terms of information). An optimal test δ ∗ = (D∗ , T∗ ) minimizes the mutual information I(H; T|D), i.e. I(H; T∗ |D∗ ) = inf I(H; T|D). δ=(D,T)∈C(α1 ,α2 ) (15) Later we will show that for continuous observation processes optimality in the sense of Definition 1 implies optimality in the sense of Definition 2 but not vise versa. In this regard, consider that (15) is invariant w.r.t. time delays Tdelay in the decisions, i.e., I(H; T∗ |D∗ ) = I(H; T∗ + Tdelay |D∗ ), if Tdelay is statistically independent of H conditioned on D and T and if additionally Tdelay satisfies that I(H; Tdelay |D∗ , T∗ + Tdelay ) = 0. Moreover, we will show that for continuous observation processes optimal information usage implies that I(H; T∗ |D∗ ) = 0, because a test achieving I(H; T|D) = 0 always exists and I(H; T|D) is nonnegative. For these reasons Definition 2 will allows us to formulate practical tests for optimality of sequential decisionmaking in black-box decision devices. In general, for the discrete-time setting I(H; T∗ |D∗ ) > 0, as the information on the hypothesis does not arrive continuously but in chunks, which makes it more difficult to test optimality in discrete-time settings. III. O PTIMALITY C ONDITIONS FOR C ONTINUOUS O BSERVATION P ROCESSES To understand the conditions on optimal sequential decision-making we will derive relations between decision time distributions of optimal binary sequential probability ratio tests. In this section we consider optimal sequential probability ratio tests for continuous observation processes, which are given by (Tdec , Ddec ) in (13) - (14). We call these relations decision time fluctuation relations for their reminiscence to stopping time fluctuation relations in non-equilibrium statistical physics, in particular stochastic thermodynamics [11], [12]. In order to derive these relations January 8, 2018 DRAFT 10 we use a key property of the exponential of the cumulative log-likelihood ratio eSt defined in (12), namely that it is a positive and uniformly integrable martingale process with respect to the probability measure P2 and the filtration generated by the observation process [9]. An Ft adapted and integrable process is called a martingale w.r.t. Ft and a measure P if its expected value at time t equals to its value at a previous time t̃, when the expected value is conditioned on observations up to the time t̃. For eSt , Ft , and P2 this implies that h i St E e Ft̃ , H = 2 = eSt̃ (16) P2 -almost surely and with t̃ < t. Integrability of eSt implies that E[eSt |H = 2] = 1 < ∞. A. Decision Time Fluctuation Relation for Optimal Decision Devices Theorem 1. We consider a binary sequential hypothesis testing problem with the hypotheses H ∈ {1, 2}. Let P1 and P2 be two probability measures on the same filtered probability space (Ω, F, {Ft }t≥0 ) corresponding to the hypothesis H = 1 and H = 2, respectively. We assume that {Ft }t≥0 is right continuous. We consider that on F∞ = ∩t≥0 Ft , the probability measure P2 is absolutely continuous with respect to P1 . Furthermore, we consider that the realization of the process eSt = dP1 |Ft dP2 |Ft is P2 almost surely continuous. Let Tdec and Ddec be as in (13) and (14) with E[Tdec |H = i] < ∞ (i = 1, 2). We also assume that Tdec has a density function. Under these assumptions the following holds pTdec (t|H = 1, Ddec = 1) = pTdec (t|H = 2, Ddec = 1) (17) pTdec (t|H = 1, Ddec = 2) = pTdec (t|H = 2, Ddec = 2) (18) where pTdec (t|H, Ddec ) is the decision time distribution conditioned on the hypothesis H and the decision output Ddec . Proof. Let Φ1 (t) = {ω ∈ Ω : Tdec (ω) ≤ t and Ddec (ω) = 1} (19) be the set of trajectories for which the decision time does not exceed t and the test decides for Ddec = 1. The probability of the event Φ1 (t) with respect to the measures P1 or P2 is equal to the cumulative distribution of the decision time Tdec conditioned on the hypothesis H = 1 or January 8, 2018 DRAFT 11 H = 2, respectively, and conditioned on the decision outcome Ddec = 1. We find the following identity between P1 (Φ1 (t)) and P2 (Φ1 (t)): Z P1 (Φ1 (t)) = dP1 |Ft (20) eSt dP2 |Ft (21) eSTdec dP2 |Ft (22) dP2 |Ft (23) ω∈Φ1 (t) Z = ω∈Φ1 (t) Z = ω∈Φ1 (t) L1 Z =e ω∈Φ1 (t) = eL1 P2 (Φ1 (t)) (24) where for (21) we have used the Radon-Nikodým theorem and the definition (12). For equality (22) we have applied Doob’s optional sampling theorem [9], [13] to the uniformly integrable P2 -martingale process eSt . For (23) we have used that eSt is a continuous process and achieves the value eL1 at time Tdec . The probability density functions of Tdec can be expressed in terms of the derivatives of the cumulative distributions Pk (Φ1 (t)) (k = 1, 2) d P1 (Φ1 (t)) dt d = 1|H = 2) = P2 (Φ1 (t)). dt pTdec (t|H = 1, Ddec = 1)P (Ddec = 1|H = 1) = (25) pTdec (t|H = 2, Ddec = 1)P (Ddec (26) The ratio of the decision probabilities is P (Ddec = 1|H = 1) 1 − α2 = = eL1 P (Ddec = 1|H = 2) α1 (27) which follows from P (Ddec = 1|H = 1) = limt→∞ P1 (Φ1 (t)), P (Ddec = 1|H = 2) = limt→∞ P2 (Φ1 (t)), Eq. (24), and from the assumption that the test terminates almost surely. Taking the derivative of the left hand side (LHS) of (20) and the right hand side (RHS) of (24), and using Eqs. (25) to (27), we prove Eq. (17). Analogously, Eq. (18) can be proved. B. Decision Time Fluctuation Relation for Optimal Decision Devices with Unknown Hypotheses In the following, we derive a second fluctuation relation, which we will apply to test optimality of sequential decision-making with less information than required for Theorem 1 (see Section V-B2), but holds only if the maximal allowed error probabilities are symmetric, i.e., January 8, 2018 DRAFT 12 α1 = α2 , and the measures P1 and P2 on (Ω, F, {Ft }t≥0 ) are related by a measurable involution Θ. We consider that P2 = P1 ◦ Θ (28) with Θ : Ω → Ω a measurable involution, i.e., Θ is invertible with inverse Θ−1 = Θ and with Θ(Φ) ∈ F for all Φ ∈ F. Theorem 2. Under the same conditions as in Theorem 1, with the additional assumption that P2 = P1 ◦ Θ with Θ a measurable involution, and with the additional assumption that the maximal allowed error probabilities fulfill α1 = α2 , the following holds pTdec (t|H = 1, Ddec = 1) = pTdec (t|H = 1, Ddec = 2) (29) pTdec (t|H = 2, Ddec = 1) = pTdec (t|H = 2, Ddec = 2). (30) Furthermore, it holds that pTdec (t|Ddec = 1) = pTdec (t|Ddec = 2). (31) The proof of Theorem 2 is given in Appendix A. A special case of the result in Theorem 2 has been found in the context of nonequilibrium statistical physics [11], [12]: the two hypotheses correspond to a forward and a backward direction of the arrow of time, and Θ corresponds to the time-reversal operation. The RadonNikodým derivative St is then the stochastic entropy production, and the decision time Tdec is its two-boundary first-passage time to cross one of two given symmetric values. Moreover, in communication theory such a symmetry has been found to show that the probability of cycle slips to the positive/negative boundary in phase-locked loops used for synchronization is independent of time [14, Eq. (74)]. C. Information Theoretic Implications of Optimal Sequential Decision-Making Theorem 1 and Theorem 2 express statistical dependencies of different random quantities involved in optimal sequential decision-making. Based on Theorem 1 we will now show the following. Corollary 1. Under the same conditions as in Theorem 1, the following equality for mutual information holds I(H; Tdec |Ddec ) = 0 January 8, 2018 (32) DRAFT 13 i.e., I(H; Tdec , Ddec ) = I(H; Ddec ). Proof. By the chain rule for mutual information, I(H; Tdec |Ddec ) can be expressed by I(H; Tdec , Ddec ) = I(H; Ddec ) + I(H; Tdec |Ddec ). (33) The second term on the RHS of (33) is given by    pTdec (Tdec |Ddec , H) I(H; Tdec |Ddec ) = E log2 (34) pTdec (Tdec |Ddec )    pTdec (Tdec |Ddec , H) (35) = E log2 pTdec ,H (Tdec , H = 1|Ddec ) + pTdec ,H (Tdec , H = 2|Ddec )    pTdec (Tdec |Ddec , H) = E log2 pT (Tdec |Ddec , H = 1)P (H = 1|Ddec ) + pTdec (Tdec |Ddec , H = 2)P (H = 2|Ddec )    dec pTdec (Tdec |Ddec , H) = E log2 pTdec (Tdec |Ddec , H = 1)P (H = 1|Ddec ) + pTdec (Tdec |Ddec , H = 1)P (H = 2|Ddec ) (36)    pTdec (Tdec |Ddec , H) = E log2 (37) pTdec (Tdec |Ddec , H = 1) (P (H = 1|Ddec ) + P (H = 2|Ddec ))    pTdec (Tdec |Ddec , H) (38) = E log2 pTdec (Tdec |Ddec , H = 1)       pTdec (Tdec |Ddec , H = 1) pTdec (Tdec |Ddec , H = 2) = P (H = 1)E log2 + P (H = 2)E log2 pTdec (Tdec |Ddec , H = 1) pTdec (Tdec |Ddec , H = 1) =0 (39) where for (36) and for (39) we have used Theorem 1. Corollary 1 states that in case of optimal sequential decision-making the decision time Tdec does not give any additional information on the hypothesis H beyond the decision outcome Ddec . In this regard, consider that the first term on the RHS of (33) is the mutual information the decision outcome of the test Ddec gives about the actual hypothesis H. The second term on the RHS of (33) I(H; Tdec |Ddec ) is the additional information the termination time Tdec gives on the hypothesis H beyond the information given by the decision Ddec . Thus, we have proved that for continuous observation processes optimal sequential decision-making w.r.t. Definition 2 is achievable and that I(H; T∗ |D∗ ) = 0. Note that since sequential probability ratio tests (Ddec , Tdec ) have been shown to be optimal in the sense of Definition 1, Corollary 1 implies that optimality in the sense of Definition 1 also implies optimality in the sense of Definition 2. In case the assumptions of Theorem 2 are satisfied additionally, the following two corollaries hold. January 8, 2018 DRAFT 14 Corollary 2. Under the same conditions as in Theorem 2, the following equality holds I(Ddec ; Tdec ) = 0. (40) Proof. It holds that   pTdec (Tdec |Ddec ) I(Ddec ; Tdec ) = E log (41) pTdec (Tdec )   pTdec (Tdec |Ddec ) = E log pTdec (Tdec |Ddec = 1)P (Ddec = 1) + pTdec (Tdec |Ddec = 2)P (Ddec = 2)   pTdec (Tdec |Ddec ) (42) = E log pTdec (Tdec |Ddec = 1) =0 (43) where we have used (31) in (42) and (43). Corollary 3. Under the same conditions as in Theorem 2, and with the additional assumption that P (H = 1) = P (H = 2), the following equality holds I(H; Tdec ) = 0. (44) The proof of Corollary 3 is given in Appendix B. IV. O PTIMALITY C ONDITIONS FOR D ISCRETE -T IME O BSERVATION P ROCESSES In the following, we extend the analysis on optimal information usage in sequential decisionmaking to the discrete-time setting. In discrete time the optimal test in the sense of Definition 1 is given by TWald and DWald defined in (9) and (10). Extending our results to a discretetime setting is relevant for discrete-time systems. Moreover, in usual experimental setups a continuous-time system is sampled yielding a discrete-time representation. The extension from continuous processes to discrete-time processes is not straightforward, as one key characteristic in the continuous-time setting is the fact that the test terminates with a cumulative log-likelihood ratio exactly hitting one of the thresholds. This property of continuous processes does not hold true in the discrete-time setting, where the mean value of the cumulative log-likelihood ratio at the decision time slightly overshoots the thresholds. The thresholds L1 and L2 depend on the maximum allowed error probabilities α1 and α2 , cf. (4). Due to the fact that in the discrete-time setting the trajectory of the accumulated loglikelihood ratios Sk in (6) does not necessarily hit one of the thresholds the determination of the January 8, 2018 DRAFT 15 optimal thresholds L1 and L2 in terms of α1 and α2 are rather involved, see [2]. L1 and L2 are chosen such that the allowed error probabilities given in (4) are obeyed with equality. In the following, we study the statistical dependencies between the hypothesis H, the decision DWald , and the number of observations TWald the sequential probability ratio test given by (9) and (10) uses to make decisions. The necessary condition for optimal decision devices given in Theorem 1 for the continuoustime setting does not carry over to the discrete-time settings as we will discuss in the following. This can be understood from applying the steps in the proof of Theorem 1 in (20) to (24) to the discrete-time setting. In the discrete-time case with t ∈ Z+ the measure P1 (Φ1 (t)) of the discrete-time version of the set Φ1 (t) in (19) can be expressed by Z P1 (Φ1 (t)) = dP1 |Ft (45) ω∈Φ1 (t) Z = eSt dP2 |Ft (46) eSTWald dP2 |Ft (47) ω∈Φ1 (t) Z = ω∈Φ1 (t) R = ω∈Φ1 (t) eSTWald dP2 |Ft Z R ω∈Φ1 (t) dP2 |Ft dP2 |Ft (48) ω∈Φ1 (t)   = E eSTWald |H = 2, ω ∈ Φ1 (t) P2 (Φ1 (t))   = eL1 E eM1 |H = 2, ω ∈ Φ1 (t) P2 (Φ1 (t)) (49) (50) where M1 > 0 in (50) is the overshoot beyond the threshold L1 . Since in general the distribution of the overshoot M1 = STWald − L1 depends on the time t the fluctuation relations (17) and (18) do not extend to the discrete-time case. Translating (25) and (26) to the discrete-time setting yields pTdec (t|H = 1, Ddec = 1)P (Ddec = 1|H = 1) = P1 (Φ1 (t + 1)) − P1 (Φ1 (t)) (51) pTdec (t|H = 2, Ddec = 1)P (Ddec = 1|H = 2) = P2 (Φ1 (t + 1)) − P2 (Φ1 (t)) (52) for t ≥ 1. Taking the difference between the values of P1 (Φ1 (t)) at two consecutive time instants we get   P1 (Φ1 (t + 1)) − P1 (Φ1 (t)) = eL1 E eM1 |H = 2, ω ∈ Φ1 (t + 1) P2 (Φ1 (t + 1))   − eL1 E eM1 |H = 2, ω ∈ Φ1 (t) P2 (Φ1 (t)). January 8, 2018 (53) DRAFT 16   In case E eM1 |H = 2, ω ∈ Φ1 (t) = K is time independent, we get from (51) and (52) 1 − α2 P (DWald = 1|H = 1) = = eL1 K P (DWald = 1|H = 2) α1 (54) where we have used the assumption that the test terminates almost surely, and we get the fluctuation relations corresponding to Theorem 1 for decision times TWald in the discrete-time case.   The constraint that E eM1 |H = 2, ω ∈ Φ1 (t) is time independent is approximately fulfilled in case the size of the thresholds L1 and L2 is large in comparison to the average increase of the log-likelihood ratio ∆n per observation sample, see (11). This can be seen as taking the continuum limit of the decision making process. In this regard, consider that the distribution of the overshoot M1 is time independent if the distribution of the distance LDWald − STWald −1 , at the time instant before a decision is taken, is time independent, and if the distribution of the increment ∆TWald is independent of time. The distribution of LDWald − STWald −1 is time independent if the initial value of the cumulative log-likelihood has no significant influence anymore on the distribution of STWald −1 when conditioning on termination at time instant TWald . This is satisfied in case TWald is sufficiently large, which holds if the thresholds L1 and L2 are large in comparison to the average of the increments of the log-likelihood ratio ∆n . This is illustrated for an example based on numerical simulations in Section VI-A4. In the following, we assume that the condition   E eM1 |H = 2, ω ∈ Φ1 (t) = K (55) is fulfilled. For many practical applications this condition is approximately fulfilled, see the numerical experiments in Section VI. The results on optimal information usage carry over from continuous time to discrete time given that (55) holds. Theorem 3. We consider a binary sequential hypothesis testing problem with the hypotheses H ∈ {1, 2}. Let {pXk+1 (·|H = 1)} and {pXk+1 (·|H = 2)} be two sequences of probability density 1 1 functions of the sequence of real valued observations {X1 , X2 , . . .} in case hypothesis H = 1 and H = 2 are true, respectively, and with k ∈ Z+ . Let TWald and DWald be as in (9) and (10) January 8, 2018 DRAFT 17 with E[TWald |H = i] < ∞ (i = 1, 2). Under these assumptions and the assumption that (55) is fulfilled it holds that P (TWald = k|H = 2, DWald = 1) = P (TWald = k|H = 1, DWald = 1) (56) P (TWald = k|H = 2, DWald = 2) = P (TWald = k|H = 1, DWald = 2) (57) for all k ∈ Z+ . Theorem 3 implies optimal usage of information with respect to Definition 2 for the discretetime setting yielding the following corollary. Corollary 4. Under the same conditions as in Theorem 3, the following equality for mutual information holds I(H; TWald , DWald ) = I(H; DWald ) (58) I(H; TWald |DWald ) = 0. (59) implying that The proof follows along the same line as the proof of Corollary 1, but this time based on Theorem 3. Analogously, Theorem 2 carries over to the discrete-time case. Theorem 4. Under the same conditions as in Theorem 3, with the additional assumption that pXk+1 (·|H = 2) = pXk+1 ◦Θ (·|H = 1) for k ∈ Z+ , where Θ is a measurable involution, and the 1 1 additional assumption that the maximal allowed error probabilities fulfill α1 = α2 , the following holds P (TWald = k|H = 1, DWald = 1) = P (TWald = k|H = 1, DWald = 2) (60) P (TWald = k|H = 2, DWald = 1) = P (TWald = k|H = 2, DWald = 2) (61) for all k ∈ Z+ . Furthermore, it holds that P (TWald = k|DWald = 1) = P (TWald = k|DWald = 2) for all k ∈ Z+ . (62) Theorem 4 can be proved by carrying over the proof of Theorem 2 to the discrete-time case and additionally using a modification of the application of Doob’s optional sampling theorem similar to (48) to (50) leading to the additional assumption that (55) is fulfilled. January 8, 2018 DRAFT 18 A special case of Theorem 4 was shown in [1] for the case of i.i.d. observation processes and low error probabilities α1 = α2 . Using Theorem 4 also Corollary 2 and Corollary 3 carry over to the discrete-time case. Corollary 5. Under the same conditions as in Theorem 4, the following equality holds I(DWald ; TWald ) = 0. (63) Corollary 6. Under the same conditions as in Theorem 4, and with the additional assumption that P (H = 1) = P (H = 2), the following equality holds I(H; TWald ) = 0. (64) The proofs of Corollary 5 and Corollary 6 follow along the lines of the proves of Corollary 2 and Corollary 3. V. T ESTS FOR O PTIMALITY OF S EQUENTIAL D ECISION -M AKING For the case of continuous observation processes Theorem 1 and Corollary 1 hold for binary sequential probability ratio tests which are optimal in the sense of Definition 1. In case of additional symmetry conditions, Theorem 2, Corollary 2, and Corollary 3 hold as well. Under the reasonable assumption that the joint statistics of (H, T, D) have a unique solution over all tests fulfilling (5), these theorems and corollaries are necessary conditions for optimal sequential decision-making in the sense of Definition 1. Likewise, for discrete-time observation processes which fulfill the condition given by (55) Theorem 3 and Corollary 4 give necessary conditions for optimal sequential decision-making. In case additional symmetry conditions are fulfilled also Theorem 4, Corollary 5, and Corollary 6 hold. Based on these theorems and corollaries we formulate tests to test optimality of sequential decision-making in black-box decision devices and present algorithms to measure the distance to optimality of the decision process in the black-box decision device. A. Continuous-Observation Processes 1) Testing Optimality and Measuring the Distance to Optimality in Case of Known Hypotheses: We sample m independent realizations {hi , ti , di }i=1,...,m of the joint random variables (H, T, D) given by subsequent decisions, where H corresponds to the random variable describing the actual hypothesis and T and D are the outputs, decision time and decision variable, of the January 8, 2018 DRAFT 19 black-box decision device. In case the observation window of the experimentalist is not sufficiently large such that for certain samples the black box has not decided yet, the experimentalist can discard those samples. We first state a statistical test which can reject, with a certain statistical significance, the null hypothesis that the given black-box device is optimal in the sense of Definition 2 and, thus, also in the sense of Definition 1. We create from the whole set of realizations four subsets of decision times Ar,s = {ti , i ∈ {1, 2, ..., m} : (hi , di ) = (r, s)} with (r, s) ∈ {(1, 1), (1, 2), (2, 1), (2, 2)}. Under the null hypothesis, Theorem 1 implies that the subsets A1,1 and A2,1 contain independent realizations of decision times from the same distribution and, analogously, A1,2 and A2,2 contain independent realizations of decision times from the same distribution. Whether two sets of independent realizations are sampled from the same continuous distribution can be tested with a certain significance using the two-sample Kolmogorov-Smirnov test [15, pp. 663-665]. Note that we do not require knowledge on the statistics P1 and P2 of the observation process Xt , which makes our test for optimality of decision devices very useful for practical applications where in many situations such statistics are unknown. Notice that because the observation process Xt may take values in a high-dimensional space it can be difficult to get a good estimate of its statistics. A quantity for the distance of the sequential decision-making process of the black-box decision device with respect to the optimal sequential decision process in the sense of Definition 2 is ˆ T|D) of the mutual information I(H; T|D). The estimate given by the empirical estimate I(H; ˆ T|D) can be gained from empirical estimates of entropy and differential entropy, see [16] I(H; and [17]. Note that with the chain rule for mutual information it holds that I(H; T|D) = I(H; D, T) − I(H; D). (65) The first term on the RHS of (65) is the complete mutual information that the output of the black-box decision device, (D, T), gives on the hypothesis H. The second term on the RHS of (65) is the mutual information between the decision D and the hypothesis H, which in case of optimal sequential decision-making equals the complete mutual information I(H; D, T). Hence, with (65) I(H; T|D) measures the information the black-box device discards in case of nonˆ T|D) provides a measure for how much the decision optimal decision-making. Therefore, I(H; statistics of a certain black-box device diverge from the optimal solution, or less formally stated, how close to optimality a decision device behaves. January 8, 2018 DRAFT 20 2) Testing Optimality in Case of Unknown Hypotheses: If the statistics of the observation process Xt fulfill the involution condition (28) and if the constraints on the error probabilities α1 and α2 are equal implying symmetric thresholds L1 = −L2 = L, then based on Theorem 2 we formulate a test which is able to reject optimality in the sense of Definition 2 and, thus, also in the sense of Definition 1. Different to the test formulated in Section V-A1 we do not require knowledge of the actual realizations of the hypothesis H, which in certain situations is important for practical application. We sample m independent realizations {ti , di }i=1,...,m of the joint random variables (T, D), where T and D are the outputs, decision variable and decision time, of the black-box decision device. As in the case of known hypotheses samples for which the black-box decision device has not terminated yet can be discarded. Under the above assumptions, we state a statistical test which can reject, with a certain statistical significance, the null hypothesis that the given black-box device is optimal. We create from the whole set of realizations two subsets of decision times Ar = {ti , i ∈ {1, 2, ..., m} : di = r} with r ∈ {1, 2}. Under the current null hypothesis, Theorem 2 implies that the subsets A1 and A2 contain independent realizations of decision times from the same distribution. We can again use a two-sample Kolmogorov-Smirnov test [15, pp. 663-665] to reject the null hypothesis with a certain significance. Corollaries 2 and 3 provide alternative means to test optimality of sequential decision-making. Quantifying the degree of optimality using the mutual informations I(D; T) or I(H; T) provides in general no clear interpretation. Hence, in order to quantify the divergence of the black-box device from optimal sequential decision-making we can use I(H; T|D) based on Corollary 1. B. Discrete-Time Observation Processes Analogously to the case of continuous observation processes we formulate tests for optimality of sequential decision-making in discrete time and we also present algorithms to measure the distance to optimality of black-box decision devices. We use Theorems 3, Theorem 4 and Corollary 4. 1) Testing Optimality and Measuring the Distance to Optimality in Case of Known Hypotheses: We sample m independent realizations {hi , ti , di }i=1,...,m of the joint random variables (H, T, D) given by subsequent decisions, where H corresponds to the random variable describing the actual hypothesis and T and D are the outputs, decision time and the decision variable, of January 8, 2018 DRAFT 21 the black-box decision device. As before we discard samples for which the black-box decision device has not decided yet. The algorithm to test optimality of a black-box decision device is analogous to the case of continuous observation processes. We construct from the whole set of realizations {hi , ti , di }i=1,...,m four sets of decision times Ar,s = {ti , i ∈ {1, 2, ..., m} : (hi , di ) = (r, s)} with (r, s) ∈ {(1, 1), (1, 2), (2, 1), (2, 2)}. Theorem 3 implies that the subsets A1,1 and A2,1 contain independent realizations of decision times from the same distribution and, analogously, A1,2 and A2,2 contain independent realizations of decision times from the same distribution. Whether two sets of independent realizations are sampled from the same discrete distribution can be tested with a certain significance using the two-sample χ2 -test [18, p. 253, Problem 3]. A quantity for the distance to optimality of a black-box decision device with respect to the optimal sequential decision process in the sense of Definition 2 is in discrete time given by the ˆ T|D) of the mutual information I(H; T|D), see , cf. Corollary 4. empirical estimate I(H; 2) Testing Optimality in Case of Unknown Hypotheses: As in the continuous case testing optimality of a black-box decision device can be done even in case of unknown hypothesis in case certain additional conditions are fulfilled. Namely, the statistics of the observation process Xt have to fulfill the involution condition (28) and the constraints on the error probabilities α1 and α2 have to be equal, implying symmetric thresholds L1 = −L2 = L. Then based on Theorem 4 we can formulate the following test. We sample m independent realizations {ti , di }i=1,...,m of the joint random variables (T, D), where T and D are the decision time and decision variable of the black-box decision device. We create from the whole set of realizations two subsets of decision times Ak = {ti , i ∈ {1, 2, ..., m} : di = k} with k ∈ {1, 2}. Theorem 4 implies that the subsets A1 and A2 contain independent realizations of decision times from the same distribution in case the black-box decision device is optimal. We can again use a two-sample χ2 -test [18, p. 253, Problem 3] to test whether the two subsets A1 and A2 are sampled from the same distribution. Corollary 5 and Corollary 6 provide alternatives to test optimality of sequential decision-making. VI. T ESTING O PTIMALITY IN N UMERICAL E XPERIMENTS In this section we apply our algorithms to test optimality of binary sequential decision-making of black-box devices and to measure the degree of divergence from optimality. We consider a class of decision devices which for certain parameter values are optimal, and we verify whether our algorithms are able to detect the parameters for which the decision devices are optimal. In January 8, 2018 DRAFT 22 this section we distinguish again continuous and discrete observation processes. However, here we will start with discrete-time processes which allow for simpler numerical study. To distinguish theoretical quantities from empirical estimates we denote by P̂ and Iˆ the empirical estimates of the probabilities P and mutual informations I. Furthermore, we write Ê [·] for the empirical estimate of the expectation E [·]. A. Discrete-time observation processes 1) Testing optimality in case of known hypotheses: We consider an observation sequence X = (X1 , X2 , . . . , Xk ) where the Xn (where n ∈ [1, k]) are i.i.d. random variables drawn from one of two possible probability distributions corresponding to the two hypotheses H ∈ {1, 2}, i.e., k
Y pXk1 xk1 |H = pX (xn |H) . (66) n=1 In our example the densities pX (·|H) are Gaussian with mean µi and variance σi2 with i ∈ {1, 2} corresponding to the two hypotheses H = i. In the special case where µ1 = −µ2 and σ1 = σ2 the involution property (28) holds. We consider a class of decision models representing the black-box decision devices. These decision models use the Wald sequential probability ratio test based on a model of the external world which may be incorrect. Each decision model computes the cumulative log-likelihood ratio of two Gaussian distributions with mean µ̃i and variance σ̃i2 (with i ∈ {1, 2}), i.e.,   k X pX (Xn |H = 1) log S̃k = pX (Xn |H = 2) n=1  k  σ̃2 X (Xn − µ̃2 )2 (Xn − µ̃1 )2 = k log + − . σ̃1 n=1 2σ̃22 2σ̃12 (67) The decision time of the model is T = min{k ∈ N : S̃k ∈ / (L2 , L1 )} (68) and the decision variable is given by   1 if S̃ ≥ L T 1 D=  2 if S̃T ≤ L2 (69) with the two thresholds L1 > 0 and L2 < 0. If µ̃i = µi and σ̃i = σi then the black-box device uses the correct model of the external world and makes optimal sequential decisions in the sense January 8, 2018 DRAFT 23 0.2 0.3 0.25 0.15 0.2 0.1 0.15 0.1 0.05 0 0 0.05 5 10 15 20 25 30 35 40 0 0 5 10 15 20 25 30 35 40 (a) Optimal decision-making: black-box decision device uses (b) Sub-optimal decision-making: black-box decision device the correct model of the external world uses the wrong model of the external world Fig. 2. Illustration of Theorem 3 using the observation model given by (66) and the decision model given by (67) to (69). The distributions of the observation process corresponding to the two hypotheses are Gaussian with parameters (µ1 , σ1 ) = (0, 5) and (µ2 , σ2 ) = (1, 10), respectively. The decision model has threshold values L1 = 4 and L2 = −2 and parameters: (a) (µ̃1 , σ̃1 ) = (µ1 , σ1 ) and (µ̃2 , σ̃2 ) = (µ2 , σ2 ), (b) (µ̃1 , σ̃1 ) = (µ1 , σ1 ) and (µ̃2 , σ̃2 ) = (5, σ2 ). The empirical error probabilities are in (a) α1 = 0.041 and α2 = 0.0133 and (b) α1 = 0.0335 and α2 = 0.0506. Distributions are estimated using 1e + 6 simulation runs. of minimizing the decision time (see Definition 1) since T = TWald and D = DWald . Corollary 4 implies that for these parameter values the black-box device makes also optimal sequential decisions in the sense of information usage (see Definition 2). If additionally L1 = −L2 , then α1 = α2 = α. We now study the decision time distributions to illustrate Theorem 3 using numerical simulations. In Fig. 2 we present the estimated decision time distributions for optimal and suboptimal sequential decision-making. Consistent with Theorem 3 the estimates of the distributions P̂ (T = k|H = 2, D = a) and P̂ (T = k|H = 1, D = a) (a ∈ {1, 2}) overlap if the black-box decision device performs the Wald test and if condition (55) is approximately fulfilled as shown in Fig. 2(a). If the black-box decision device is suboptimal, as is the case in Fig. 2(b), then these two distributions are different. Moreover, since (55) is only approximately fulfilled, the theoretical distributions P (T = k|H = 2, D = a) and P (T = k|H = 1, D = a) (a ∈ {1, 2}) corresponding to the estimates shown in Fig. 2(a) are also different. This example illustrates the value of Theorem 3 to quantify optimality for practical purposes in discrete time. In Fig. 3 we use the statistical test for optimality described in Section V-B1. We plot the January 8, 2018 DRAFT 24 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 -4 -3 -2 -1 0 1 2 3 4 0 -4 (a) Two sample χ2 -test of (56) for D = 1 -3 -2 -1 0 1 2 3 4 (b) Two sample χ2 -test of (57) for D = 2 Fig. 3. Illustration of testing optimality for discrete-time observations and for known hypotheses as described in Section V-B1. Numerical results presented are for the observation model given by (66) and the decision model of the black box given by (67) to (69). The observation process has parameters σ1 = 5, (µ2 , σ2 ) = (1, 10) and the values of µ1 are given in the legend. The decision model of the black-box decision device has threshold values L1 = 4 and L2 = −2 and parameters (µ̃1 , σ̃1 ) = (µ1 , σ1 ), σ̃2 = σ2 and with µ̃2 as given by the abscissa. We plot the p-values for the null hypothesis that (a) pT (T|D = 1, H = 1) = pT (T|D = 1, H = 2) (b) pT (T|D = 2, H = 1) = pT (T|D = 2, H = 2). The estimates of the p-values are average values over 1e + 4 two-sample χ2 -tests; each two-sample χ2 -tests evaluates a p-value over a population of 1e + 5 outcomes of the black-box decision device. We took P (H = 1) = P (H = 2) = 1/2, and a maximum observation window of 10 observations, i.e., all outcomes with more than 10 observations are discarded. 0.15 0.001 0.1 0.0001 0.05 1e-05 -4 -3 -2 -1 0 1 2 3 (a) Divergence to optimality in information usage 4 0 -4 -3 -2 -1 0 1 2 3 4 (b) Divergence to optimality in average decision times Fig. 4. Estimating the divergence to optimality in the sense of Definitions 1 and 2, using the observation model given by (66) and the decision model of the black box given by (67) to (69). The parameters of the observation process and of the decision model are the same as in Fig. 3. Fig. 4(b) shows the estimate of the average decision time of the black-box decision device divided by the estimate of the average decision time of a Wald test (µ2 = µ̃2 ), with the same error probabilities as achieved by the black-box decision device, minus one; therefore each Wald test has different values of the thresholds L1 and L2 depending on µ̃2 . In Fig. 4(a) each sample point is calculated using 1e + 9 simulation runs, and in Fig. 4(b) using 1e + 8 simulation runs. We took P (H = 1) = P (H = 2) = 1/2. January 8, 2018 DRAFT 25 2 2 test test corresponding to, respectively, a two-sample χ2 -test and pχD=2 estimates of the p-values pχD=1 of the subsets A1,1 and A2,1 , see Fig. 3(a), and a two-sample χ2 -test of the subsets A1,2 and A2,2 , see Fig. 3(b). These p-values denote the probability to falsely reject the null hypothesis that the samples in the two data sets are drawn from the same decision time distribution. Therefore, for example in the case of µ1 = −2 and D = 1 we can safely reject the null hypothesis for values of µ˜2 > 3 since the p-value is small. For values of µ˜2 ∈ [−4, 3] we need more data to safely reject the hypothesis that the test is optimal. 2) Measuring divergence to optimality of black-box decision devices: With the same example as in Fig. 3 we illustrate how to use Corollary 4 to estimate the divergence of a black-box decision device to the optimal case given by Definition 2. Note that in the example of Fig. 3 condition (55) is only approximately fulfilled, and therefore we only expect (59) to be approximately fulfilled. In Fig. 4(a) we present the numerical estimates of I (H; T|, D). In accordance with Corollary 4, if the test is optimal, i.e., the black-box decision device uses a Wald test, than the estimate of the mutual information I (H; T|, D) is minimal and approaches zero. In the example of Fig. 4(a) this happens at µ̃2 = 1. For this case also the mean decision time is minimized as shown in Fig. 4(b). Note that Corollary 4 is not a sufficient condition to test optimality with respect to Definition 1, i.e., to test whether the black-box decision device achieves the minimum mean decision time, as is illustrated by Fig. 4(a) where we observe a second minimum for the estimate of I (H; T|, D). However, for this second minimum the black-box decision device is optimal with respect to Definition 2, which is not related to a minimum mean decision time as illustrated in Fig. 4(b). Note that the estimation of I (H; T|, D) in Fig. 4(a) requires only knowledge of the output of the decision device whereas the estimation of the minimum mean decision time E [TWald ] requires knowledge on the statistics of the observation process, which in practical applications is often unavailable. 3) Testing optimality in case of unknown hypotheses: In this section now we consider testing optimality in case of unknown hypotheses based on Theorem 4. However, the example given by (66) to (69) is not suitable to discuss Theorem 4. The reason is that the cumulative log-likelihood ratio process S̃k becomes a drift-diffusion process in the continuous limit independent of the choice of µ̃i and σ̃i , for which it is known that the two-boundary first-passage time distribution with symmetric thresholds satisfies the fluctuation relation [11]. Therefore, the estimates of the distributions P̂ (T = k|H = a, D = 1) and P̂ (T = k|H = a, D = 2) (a ∈ {1, 2}) always overlap (data not shown). January 8, 2018 DRAFT 26 0.02 0.05 0.04 0.015 0.03 0.01 0.02 0.005 0.01 0 0 100 50 150 200 0 0 50 100 150 200 (a) Optimal decision-making: black-box decision device uses (b) Sub-optimal decision-making: black-box decision device the correct model of the external world uses the wrong model of the external world Fig. 5. Illustration of Theorem 4 using the observation model given by (70) and the decision model given by (71), (68), and (69). The distributions of the observation process corresponding to the two hypotheses are with parameters v1 = −v2 = 1, w1 = w2 = −1, σ1 = σ2 = 5, respectively. The decision model has threshold values L1 = 4 and L2 = −4 and uses the parameters: (a) (ṽ1 , w̃1 , σ̃1 ) = (v1 , w1 , σ1 ) and (ṽ2 , w̃2 , σ̃2 ) = (v2 , w2 , σ2 ), (b) (ṽ1 , w̃1 , σ̃1 ) = (v1 , w1 , σ1 ) and (ṽ2 , w̃2 , σ̃2 ) = (v2 , −0.5, σ2 ). The empirical error probabilities are in (a) α1 = 0.014 and α2 = 0.014 and (b) α1 = 0.0092 and α2 = 0.6932. The distributions are estimated using 1e + 7 simulation runs. Therefore, we choose a different example to illustrate the value of Theorem 4. We consider Markovian observation processes X = (X1 , X2 , . . . , Xk ) drawn from one of two probability distributions k
Y pXk1 xk1 |H = h = pX2 (xn |H = h, X1 = xn−1 ) (70) n=1 with X0 = 0 and h ∈ {1, 2}. In our example the densities pX2 (·|H = h, X1 = xn−1 ) are Gaussian with mean µh (X1 ) = vh + (wh + 1) X1 and variance σh2 with h ∈ {1, 2}, corresponding to the two hypotheses H = h. If v1 = −v2 , w1 = w2 and σ1 = σ2 , then the involution property (28) holds, such that Theorem 4 can be applied. We consider again a class of black-box decision devices which use the Wald sequential probability ratio test based on its model of the external world. The black-box decision devices compute the cumulative log-likelihood ratio based on parameters ṽh , w̃h and σ̃h2 (with h ∈ {1, 2}), i.e., S̃k = k X n=1 January 8, 2018  log pX2 |H,X1 (Xn |H = 1, Xn−1 ) pX2 |H,X1 (Xn |H = 2, Xn−1 )  DRAFT 27 k σ̃2 X + = k log σ̃1 n=1  (Xn − Xn−1 − ṽ2 − w̃2 Xn−1 )2 (Xn − Xn−1 − ṽ1 − w̃1 Xn−1 )2 − 2σ̃22 2σ̃12  . (71) The decision time T and the decision variable D are still given by (68) and (69) with the two thresholds L1 > 0 and L2 < 0. We now illustrate Theorem 4 using numerical simulations. In Fig. 5 we illustrate Theorem 4 for optimal and suboptimal sequential decision-making with symmetric thresholds L1 = −L2 = 4 and for v1 = −v2 = 1, w1 = w2 = −1, σ1 = σ2 = 5 such that the involution property (28) holds. Consistent with Theorem 4 the distributions P̂ (T = k|H = h, D = 1) and P̂ (T = k|H = h, D = 2) (h ∈ {1, 2}) overlap if the black-box decision device performs the Wald test, and is thus optimal, and if (55) approximately applies. If the black-box decision device is suboptimal, as is the case in Fig. 5(b), then these two distributions may be different. Note that since Theorem 3 also applies, all distributions in Fig. 5(a) overlap. 4) Overshoot problem: Due to the overshoot problem for discrete-time observation processes in general the condition given by (55) is violated. Therefore, even in the case of the Wald test I(H; TWald |DWald ) is in general larger than zero. In the present section, we discuss how far I(H; TWald |DWald ) deviates from zero in practical situations. We also discuss how far the condition imposed by (55) is fulfilled in our numerical examples. For this purpose, we first estimate I(H; TWald |DWald ) as a function of the threshold values and the number of test runs. In Fig. 6(a) and Fig. 6(b) it can be seen that for the Wald test the estimate of I(H; TWald |DWald ) saturates with an increasing number of test runs at a non-zero value, and therefore I(H; TWald |DWald ) > 0. This is an intrinsic aspect of sequential decision-making with discrete observation processes and cannot be avoided. For small values of λ, which parameterizes the threshold values, we see ripples in the mutual information. The minima occur approximately at integer multiples of the most likely value of the increase of the cumulative log-likelihood ratio Sk . For example, in Fig. 6(b) we illustrate how the estimate of the mutual information converges to its asymptotic value for λ = 0.16 and λ = 0.36, corresponding to the first maximum and the third minimum in Fig. 6(b). For large values of λ, i.e., when the distance of the thresholds to the origin is large with respect to the typical increase of the cumulative log-likelihood ratio, the mutual information I(H; TWald |DWald ) decreases as a function of λ. Even for large values of λ, the estimate of the mutual information I(H; TWald |DWald ) does not converge to zero as a January 8, 2018 DRAFT 28 function of the number of test runs but saturates, as the condition in (55) is not fulfilled. This is illustrated in Fig. 6(b) for the values λ = 1 and λ = 3. The fact that I(H; TWald |DWald ) is larger than zero indicates that here the condition given by   (55) is not fulfilled. To show this, in Fig. 6(c) we plot E eM1 |Φ1 (k), H = 2 as a function of time k. In conclusion, Corollary 4 is applicable to test optimality of the black-box decision device if condition (55) is approximately fulfilled, which is the case when the threshold values of the Wald test are far enough from the origin in comparison to the average increase of the cumulative log-likelihood ratio per observation. B. Continuous observation processes In this section we illustrate Theorem 1 and Corollary 1 for continuous observation processes. The decision model we study here is a drift diffusion process and has been used to describe reaction-time distributions of two-choice decision tasks of human subjects [19], [20]. 1) Observation model and decision model: We consider an observation process Xt which is an Itô-process solving the stochastic differential equation dXt = µi dt + σ dWt (72) where µi is a constant drift, with i ∈ {1, 2} corresponding to the two hypotheses H = i, where σ is a constant noise amplitude, and where X0 = 0. Here Wt is a standard Wiener process. If µ1 = −µ2 then the involution property (28) holds. We consider black-box decision devices which compute the continuous version of the cumulative log-likelihood ratio in the Wald sequential probability ratio test, cf. (67), which is given by µ̃1 − µ̃2 µ̃22 − µ̃21 + Xt . 2 2σ̃ σ̃ 2 (73) T = inf{t ∈ R : S̃t ∈ / (L2 , L1 )} (74) S̃t = t The decision time of the model is and the decision variable is given by   1 if S̃ ≥ L T 1 D=  2 if S̃T ≤ L2 January 8, 2018 (75) DRAFT 29 0.01 1e+6 1e+7 1e+8 1e+9 1e+10 0.0001 0.01 0.001 0.0001 1e-05 1e-06 1e-06 1e-08 0 0.5 1 1.5 2 2.5 3 1e-07 10000 = 0.16 = 0.36 =1 =3 1e+06 1e+08 1e+10 ˆ TWald |DWald ) as a function of the distance (a) Estimate I(H; ˆ TWald |DWald ) over the (b) Convergence of the estimate I(H; parameter λ of the thresholds and for given values of the number of test runs N and for given values of the distance number of test runs N . parameter λ of the thresholds. 2 1.5 1 0.5 0 0 5 10 15 20 (c) Evaluation of the condition (55) for the Wald test based on
plotting E eM1 |Φ1 (k), H = 2 over k (solid lines). Also the corresponding P̂ (TWald = k|H = 2, DWald = 1) is shown
(dashed lines). It can be seen that E eM1 |Φ1 (k), H = 2 is not independent of k such that (55) does not hold. Especially for λ = 0.16, the value corresponding to the first maximum
in Fig. 6(a), E eM1 |Φ1 (k), H = 2 varies in the area with the majority of the probability mass of the termination time yielding a larger I(H; TWald |DWald ) than in the case of λ = 0.36, the value corresponding to the third minimum in
Fig. 6(a), where E eM1 |Φ1 (k), H = 2 varies less over k. Fig. 6. Illustration of the impact of discreteness of the observation process. As in Fig. 2, 3 and 4, we use the observation model given by (66) with parameters (µ1 , σ1 ) = (0, 5) and (µ2 , σ2 ) = (1, 10). The decision model is the Wald test given by (67) to (69) with parameters (µ̃1 , σ̃1 ) = (µ1 , σ1 ), (µ̃2 , σ̃2 ) = (µ2 , σ2 ), L1 = 4λ, L2 = −2λ. January 8, 2018 DRAFT 30 with the two thresholds L1 > 0 and L2 < 0. Note that the cumulative log-likelihood ratio, in the case the hypothesis H = i is true, is the following Itô process dS̃t = ai dt + √ 2b dWt (76) with   µ̃1 − µ̃2 µ̃1 + µ̃2 ai = − + µi , σ̃ 2 2  2 1 µ̃1 − µ̃2 b= σ . 2 σ̃ 2 i ∈ {1, 2} (77) (78) The sequential decision-making device (T, D) has error probabilities α1 = P (D = 1|H = 2) = α2 = P (D = 2|H = 1) = 1−e 1−e e a1 L2 b a2 L2 b (79) a2 (L2 −L1 ) b −e 1−e a1 (L2 −L1 ) b a1 (L2 −L1 ) b . (80) The values of a1 , a2 , b, L1 and L2 are chosen such that α1 , α2 ∈ [0, 1/2]. If µ̃1 = µ1 , µ̃2 = µ2 and σ̃ = σ, then S̃t = St and (T, D) = (Tdec , Ddec ) with error probabilities as given by (7) and (8). Notice that the stochastic differential equation of St is of the form [21]  2 µ1 − µ2 (−1)i+1 µ1 − µ2 dWt , i ∈ {1, 2} dt + dSt = 2 σ σ √ Rt and e−St = − 2b 0 e−St0 dWt0 is a Pi -martingale process. For the special case of (81) µ̃1 + µ̃2 = µ1 + µ2 (82)
2  µ̃1 −µ̃2  we have S̃t = c St with c = σ̃σ and hence (T, D) = (Tdec , Ddec ) with error probabiliµ1 −µ2 ties, α1 = (1 − eL2 /c )/(1 − e(L2 −L1 )/c ) and α2 = (eL2 /c − e(L2 −L1 )/c )/(1 − e(L2 −L1 )/c ). Thus, in case (82) holds the black box decision device is optimal. Note that (82) implies that a1 = −a2 . 2) Illustration of Theorem 1: We consider now the special case of |L2 |  b/|a1 |, |L2 |  b/|a2 | (83) for which the expression of the distribution of decision times simplifies and allows analytical evaluation. January 8, 2018 DRAFT 31 In the following we illustrate Theorem 1. The Laplace transform of the distributions of decision times are known for arbitrary values of L1 and L2 [22]. If the conditions in (83) are fulfilled, we get (|a1 |t−L1 )2 L1 + o(1) pT (t|D = 1, H = 1) = √ e− 4bt 2 πb t3/2 (|a2 |t−L1 )2 L1 pT (t|D = 1, H = 2) = √ e− 4bt + o(1) 2 πb t3/2 (|a |t+L )2 1 1 − 1 4bt 2 √ pT (t|D = 2, H = 1) = e |a1 | 1 − e− b L1 πb t3/2   1 1 −L21 /(bt)−|L2 |L1 /(bt) + o(1) × |L2 | − (L1 + |L2 |)e 2 2 (|a |t+L )2 1 1 − 2 4bt 2 √ pT (t|D = 2, H = 2) = e |a2 | 1 − e− b L1 πb t3/2   1 1 −L21 /(bt)−|L2 |L1 /(bt) × |L2 | − (L1 + |L2 |)e + o(1) 2 2 (84) (85) (86) (87) where o denotes the little-o notation taken with respect to |L2 | going to infinity. The fluctuation relations (17) and (18) hold for a1 = −a2 , and thus for µ̃1 + µ̃2 = µ1 +µ2 . This is consistent with Theorem 1 which states that the fluctuation relation must hold whenever (T, D) = (Tdec , Ddec ). 3) Optimality in mean decision times: With this example we can also verify optimality of sequential hypothesis testing in the sense of Definition 1. The mean decision times are given by
L1 + O |L2 |e|a1 |L2 /b E [T|D = 1, H = 1] = (88) |a1 |
L1 + O |L2 |e|a2 |L2 /b E [T|D = 1, H = 2] = (89) |a2 |  
e−(|a1 |/b)L1 1 |a1 |L2 /b E [T|D = 2, H = 1] = |L2 | − 2L1 + O |L |e (90) 2 |a1 | 1 − e−(|a1 |/b)L1  
1 e−(|a2 |/b)L1 −a2 L2 /b |L2 | − 2L1 + O |L (91) E [T|D = 2, H = 2] = 2 |e −(|a |/b)L 1 |a2 | 1−e 2 where O denotes the big-O notation. The corresponding values of the average decision times of the Wald test yielding the same error probabilities α1 and α2 as (T, D) are  2
σ 1 − α2 E [Tdec |D = 1, H = 1] = E [Tdec |D = 1, H = 2] = 2 log + O |L2 |eL2 . µ1 − µ2 α1 (92) It can be shown that for L2 → −∞ we have E [T|D = 1, H = 1] − E [Tdec |D = 1, H = 1] ≥ 0, which is consistent with the optimality of the sequential probability ratio test in the sense of minimal decision times. January 8, 2018 DRAFT 32 0.05 1 0.04 0.8 0.03 0.6 0.02 0.4 0.01 0.2 0 0 1 0.5 1.5 2 (a) Divergence to optimality in information usage; note that 0 0 0.5 1 2 1.5 (b) Divergence to optimality in average decision times the black solid line and the green dotted line overlap. 1 0.1 0.01 0.001 0.0001 1e-05 1e-06 1000 10000 1e+05 1e+06 1e+07 number of runs 1e+08 1e+09 (c) Empirical estimate of the mutual information as a function of the number of test runs Fig. 7. Measuring the divergence to optimality in the sense of Definitions 1 and 2 using the continuous observation process (72) and the decision-making model (73) - (75) . The parameters defining the observation process are µ1 = 0, µ2 = 1 and σ = 5 q µ̃2 log α1 and the parameters defining the decision-making device are µ̃1 = 0, L1 = 4, α1 = 0.01, σ̃ = σ 2 2µ2µ̃−1 − and µ̃2 µ̃1 −µ̃2 L1 as given on the abscissa. The range of µ̃2 plotted corresponds with σ̃ ≥ 0. (a): Theoretical value of the mutual information I(H; N|D) as given by (94) for given values of the time resolution Tr in the legend. (b): Theoretical values of the mean decision ˆ N|D) as a function of the number of test runs, for given values of time as given by (88) - (92). (c): Empirical estimate I(H; µ̃2 and for Tr = E [Tdec |H = 1]. The corresponding theoretical value of the mutual information is indicated by the horizontal dashed lines. Note that we have generated variates of the random variables (T, H, D) according to the distributions (79) - (80), (84) - (87), and P (H = 1) = P (H = 2) = 1/2; we have generated variates from the inverse Gaussian distribution with the algorithm in [23]. January 8, 2018 DRAFT 33 4) Illustration of Corollary 1: We can also compute the mutual information in the limit L2 → −∞, which is for P (H = 1) = P (H = 2) = 1/2 given by I (H; T|D) = 1 + α1 log2 (1 + α1 ) 2   Z ∞ (|a |t−L )2 (|a |t−L )2 (|a |t−L )2 L1 −3/2 − 1 4bt 1 − 2 4bt 1 + 1 4bt 1 dt t e log2 1 + α1 e − √ 4 πb 0   Z (|a |t−L )2 (|a2 |t−L1 )2 (|a1 |t−L1 )2 α1 L1 ∞ −3/2 − 2 4bt 1 − 4bt dt t e log2 α1 + e 4bt (93) − √ 4 πb 0 with α1 = ea2 L1 /b following from (79) and α1 ∈ [0, 1/2]. If |a2 | = |a1 |, then (93) yields I (H; T|D) = 0, and otherwise I (H; T|D) > 0. In Fig. 7(a) with the black solid line we q log α1 1 −µ̃2 illustrate I (H; T|D) as a function µ̃2 for values of σ̃ = σ 2 2µ2µ̃−µ̃ such that the error 1 −µ̃2 L1 probability is fixed and we choose α1 = 0.01. The mutual information is zero for µ̃2 = µ2 = 1, corresponding to a2 = −a1 and (T, D) = (Tdec , Ddec ). In Fig. 7(b) we can easily observe the optimality of the test for µ̃2 = µ2 = 1 where the decision time takes its minimal value given by E [Tdec |Ddec = 1, H = 1]. Fig. 7(a) and Fig. 7(b) illustrate the advantage of using mutual information as a measure for testing optimality with respect to the average decision time. The mutual information is a useful quantity since I(H; T|D) = 0 at the optimal point, whereas the average decision time E[T|H] = E[Tdec |H], and hence we require knowledge of E[Tdec |H] to test optimality using decision times. 5) Practical implementation of tests for optimality of continuous observation processes: Implementation of our tests for optimality in a computer does not allow to directly treat the absolutely continuous random variable T. Moreover as any practical time measurement device has a finite time resolution, we are only able to retrieve T up to a finite quantization resolution. Thus, we discuss here how far finite resolution of T influences our tests for optimality. Note that we still consider that the decision device operates in continuous time and also that the observation process is continuous. Measuring the decision times T under a finite resolution Tr is equivalent R nTr to discretizing the distributions (84) - (87) such that P (N = n|D, H) = (n−1)T dtpT (t|D, H) r with N ∈ N being a discrete random variable. Corresponding to (93) the mutual information I (H; N|D) can be expressed by 1 + α1 log2 (1 + α1 ) 2  R nTr ! ∞ Z nTr 2 X (n−1)Tr L1 dt (|a1 |t − L1 ) − √ exp − log2 1 + α1 R nT 3/2 r 4bt 4 πb n=1 (n−1)Tr t I (H; N|D) = dt t3/2 dt (n−1)Tr t3/2 January 8, 2018   (|a2 |t−L1 )2 exp − 4bt   (|a1 |t−L1 )2 exp − 4bt DRAFT 34 α1 L1 − √ 4 πb ∞ X n=1 Z nTr (n−1)Tr 2 dt (|a2 |t − L1 ) exp − 3/2 t 4bt !  R nTr dt (n−1)Tr t3/2 log2 α1 + R nT r dt (n−1)Tr t3/2   (|a1 |t−L1 )2 exp − 4bt   . (|a2 |t−L1 )2 exp − 4bt (94) Fig. 7(a) illustrates the impact of the discretization time Tr on I (H; N|D). Note that for a1 = −a2 , corresponding to µ̃2 = 1, the mutual information I (H; N|D) = 0 for any value of Tr , since by the data processing inequality time discretization of T can just discard information [24, Theorem 2.8.1]. However, for a1 6= −a2 , corresponding to µ̃2 6= 1, the mutual information might significantly decrease because of discarding information by time discretization. Fig. 7(a) shows that for Tr ∼ 0.1 E [Tdec |H = 1] the mutual information I (H; N|D) ≈ I (H; T|D) and the effect of finite resolution is negligible. Direct implementation of our tests for optimality also requires to deal with a finite number of runs of the test. In Fig. 7(c) we evaluate the dependency of the estimate Iˆ (H; N|D) on the number of runs of the test for suboptimal tests (µ̃2 = 0.5, µ̃2 = 1.5) and an optimal test (µ̃2 = 1). The estimate of the mutual information decreases with the number of test runs, and for suboptimal tests converges to a theoretical value which is larger than zero. For an optimal sequential decision-making test, the estimate of the mutual information converges to zero. Note that this is contrary to the case of discrete-time observation processes, as illustrated in Fig. 6(b), where the estimate of the mutual information, even in the optimal case, saturates as a function of the number of test runs and converges to a positive value. VII. D ISCUSSION In the present paper we have shown that optimality of black box decision devices can be tested by studying decision time distributions given the knowledge of the actual hypothesis and the decision variable. To obtain these results we have shown that decisions times of binary sequential probability ratio tests of continuous observation processes satisfy fluctuation relations given by Theorem 1 and Theorem 2. Based on these fluctuation relations we have shown that the conditional mutual information I(H, Tdec |Ddec ) between the hypothesis H, the decision time Tdec conditioned on the decision variable Ddec is equal to zero, see Corollary 1. Using several numerical experiments we have illustrated our statistical tests. We have also discussed the limitations of our tests for sequential decision-making based on discrete-time observations. Applying our tests for optimality has several advantageous properties. Testing the necessary conditions given by Theorem 1 and Corollary 1 requires knowledge about three random variables, January 8, 2018 DRAFT 35 namely, the hypothesis H, the decision variable D, and the output time of the decision device T + Tdelay . Note that we do not require direct measurements of the decision time T, but allow for random or deterministic delay Tdelay in the output time, which needs to be statistically independent of H conditioned on T and D. Remarkably, the statistics of the actual observation process and the properties of the decision-making device, such as the allowed error probabilities α1 and α2 , are not required. For these reasons our tests are well applicable under practical experimental conditions. We now discuss a few practical examples. Studies in cognitive psychology have measured the reaction time distributions in experiments of two-choice decision tasks performed by human subjects about simple perceptual and cognitive stimuli, see e.g. [19], [20]. For fast decisions – of the order of one second – distributions of reaction times and error probabilities can be well described with a simple model for sequential decision-making in continuous time [19], [20]. Neural activity associated with the actual decisionmaking process has been identified in experiments with rhesus monkeys trained to perform rapid two-choice decisions in simple visual tasks [25], [26]. Interestingly, it was found that the firing rates of neurons in the lateral intra-parietal area correlate with the cumulative evidence associated with the hypothesis, and that a decision model based on a threshold crossing process describes the decision-time data well [27]. Furthermore, it has been conjectured that the cortex and basal ganglia, two brain regions in vertebrates, perform a multihypothesis sequential probability ratio test [28], [29], which is optimal for small error probabilities [4], [8]. Theorem 1 and Corollary 1 may be used as tools to quantify the closeness to optimality of sequential decision making by human subjects or monkeys in two-choice decision tasks. In this regard, note that experiments of two-choice decision tasks performed by human subjects or monkeys allow to measure reaction times, decision variables, and the actual realizations of the hypothesis, which are known by the construction of the experiment. Cell fate decisions are important changes of cell behavior in response to external signals. Examples are cell division controlled by growth factors, programmed cell death due to signals or the differentiation of pluripotent progenitor cell to a specific cell type as a result of biochemical signals. Cellular signaling events that control cell fate can involve signaling molecules, such as, hormones, growth factors, and cytokines [30], [31]. Because of intrinsic and extrinsic noise, cellular signaling processes have a stochastic component. Cell-fate decisions can be considered as an example of sequential decision-making based on a sequence of noisy input signals. An example of how cells could implement sequential probability ratio tests with simple examples January 8, 2018 DRAFT 36 of protein reaction networks has been given in [32]. Theorem 1 and Corollary 1 could be used to investigate the degree of optimality of cell-fate decisions. The timing of cell-fate decisions could be measured in experiments by monitoring the expression levels of fluorescently labelled molecular markers associated with the cell-fate transition within clonal populations [30], [31]. Following at the same time the input signals could in principle permit to calculate the differences of decision time distributions of correct and incorrect decisions. As already stated with our introductory example on obstacle detection for autonomous cars, sequential binary decision problems arise in many engineered systems. However, different to the assumption made for the Wald test the statistics of the observation processes Pl (l = 1, 2) are often unknown, corresponding to a nonparametric decision problem. One approach to tackle such sequential decision problems is to apply neural networks in combination with reinforcement learning [33]. The approach presented in [33] closely approximates the behavior of the optimal sequential probability ratio test and achieves a similar performance. Alternatively, in [34] an approach for nonparametric binary sequential hypothesis testing is presented, where the binary sequential detector is learned form training samples based on a so-called Wald-Kernel. The aim of these algorithms is to use the available measurements in an optimal way such that the average time to take a decision is minimized. However, the behavior of algorithms like neural networks [35], [36] can hardly be analyzed making them similar to a black-box decision device causing the problem to verify their optimality which nevertheless is crucial for application in safety critical systems like autonomous cars. This gap can be filled by out test for optimality based on Theorem 1 and Corollary 1 allowing to determine the degree of optimality of these decisionmaking devices just requiring the actual hypothesis H, the decision variable D and the decision time T of several test runs. This is especially important to determine, whether the learning process already converged sufficiently. So far our approach is limited to binary sequential probability ratio tests without prior knowledge on the hypotheses. Sequential probability ratio tests have been extended to a Bayesian setting where prior knowledge on the hypothesis H is available [10, Ch. 6.2], and have also been extended to the multihypothesis scenario. The extension of our results to these settings is for further study. January 8, 2018 DRAFT 37 A PPENDIX A P ROOF OF T HEOREM 2 Proof. We first show that the log-likelihood ratio St is odd under the transformation given by the involution Θ. This can be shown as follows eSt (Θ(ω)) = = = = = dP1 |Ft (Θ(ω)) dP2 |Ft dP1 |Ft (Θ(ω)) d(P1 ◦ Θ)|Ft d(P1 ◦ Θ)|Ft (ω) d(P1 ◦ Θ ◦ Θ)|Ft d(P1 ◦ Θ)|Ft (ω) dP1 |Ft dP2 |Ft (ω) dP1 |Ft = e−St (ω) . (95) (96) (97) (98) (99) (100) Let Φ1 (t) = {ω ∈ Ω : Tdec (ω) ≤ t and Ddec (ω) = 1} (101) Φ2 (t) = {ω ∈ Ω : Tdec (ω) ≤ t and Ddec (ω) = 2} (102) be the set of trajectories for which the decision time does not exceed t and the test decides for Ddec = 1 and Ddec = 2, respectively. Since α1 = α2 we have also L1 = −L2 , and because of the property St (Θ(ω)) = −St (ω), it follows that Φ1 (t) = Θ (Φ2 (t)) . (103) P2 (Φ1 (t)) = (P1 ◦ Θ)(Φ1 (t)) (104) Therefore, also January 8, 2018 = P1 (Θ(Φ1 (t))) (105) = P1 (Φ2 (t)). (106) DRAFT 38 Now the following holds Z P1 (Φ1 (t)) = dP1 |Ft (107) eSt dP2 |Ft (108) eSt d(P1 ◦ Θ)|Ft (109) ω∈Φ1 (t) Z = ω∈Φ1 (t) Z = ω∈Φ1 (t) Z eSt (Θ(ω)) dP1 |Ft = (110) ω∈Θ(Φ1 (t)) Z = e−St dP1 |Ft (111) e−STdec dP1 |Ft (112) ω∈Φ2 (t) Z = ω∈Φ2 (t) −L2 Z dP1 |Ft =e (113) ω∈Φ2 (t) = e−L2 P1 (Φ2 (t)) (114) where for (108) we have used the Radon-Nikodým theorem and the definition in (12). For equality (109) we have used the involution relation (28) between the measures. In (110) we have applied a variable transformation in the integral. In (111) we have used the sign reversal of St given by Eqs. (95) and (100) and the involution relation between the sets Φ2 (t) = Θ (Φ1 (t)). In (112) we have applied Doob’s optional sampling theorem to the P1 -martingale e−St . For (113) we have used that e−St is a continuous process and reaches the value e−L2 at the time Tdec . The probability density functions of Tdec can be expressed in terms of the derivatives of the cumulative distributions P(Φk (t)) (k = 1, 2) d P1 (Φ1 (t)) dt d = 2|H = 1) = P1 (Φ2 (t)). dt pTdec (t|H = 1, Ddec = 1)P (Ddec = 1|H = 1) = (115) pTdec (t|H = 1, Ddec = 2)P (Ddec (116) For the ratio of the decision probabilities we find P (Ddec = 1|H = 1) 1 − α2 1 − α1 = = = e−L2 P (Ddec = 2|H = 1) α2 α2 (117) P (Ddec = 1|H = 1) = lim P1 (Φ1 (t)) (118) P (Ddec = 2|H = 1) = lim P1 (Φ2 (t)). (119) which follows from t→∞ t→∞ January 8, 2018 DRAFT 39 Eq. (114), and the assumption that the test terminates almost surely. Notice that we have used symmetric error probabilities for which α1 = α2 . Taking the derivative of the LHS of (107) and the RHS of (114) and using Eq. (117) we prove Eq. (29). Analogously, Eq. (30) can be proved. Equation (31) follows from the identities pTdec (t|Ddec = d) = pTdec (t|Ddec = d, H = 1)P (H = 1|Ddec = d) + pTdec (t|Ddec = d, H = 2)P (H = 2|Ddec = d) = pTdec (t|Ddec = d, H = 1)P (H = 1|Ddec = d) + pTdec (t|Ddec = d, H = 1)P (H = 2|Ddec = d) = pTdec (t|Ddec = d, H = 1) (120) with d ∈ {1, 2} and where we have used Theorem 1. Using (29) and (120) we find (31), which completes the proof. A PPENDIX B P ROOF OF C OROLLARY 3 Proof. The mutual information in (44) is given by    pTdec (Tdec |H) I(H; Tdec ) = E log pTdec (Tdec )    P (Ddec = 1)pTdec (Tdec |H, Ddec = 1) + P (Ddec = 2)pTdec (Tdec |H, Ddec = 2) = E log . pTdec (Tdec ) (121) We find P (Ddec = 1|H = 1)P (H = 1) + P (Ddec = 1|H = 2)P (H = 2) P (Ddec = 1) = P (Ddec = 2) P (Ddec = 2|H = 1)P (H = 1) + P (Ddec = 2|H = 2)P (H = 2) P (Ddec = 1|H = 1) P (H = 1) + = · P (Ddec = 2|H = 1) P (H = 1) + = 1 − α P (H = 1) + · α P (H = 1) + α P (H 1−α 1−α P (H α = 2) = 2) (1 − α)P (H = 1) + αP (H = 2) αP (H = 1) + (1 − α)P (H = 2) =1 = P (Ddec =1|H=2) P (H P (Ddec =1|H=1) P (Ddec =2|H=2) P (H P (Ddec =2|H=1) (122) = 2) = 2) (123) (124) (125) (126) where we have used that the priors on H are identical, and that α1 = α2 = α. For (124) we have used (27) and (117). As Ddec is a binary random variable, it follows that P (Ddec = 1) = P (Ddec = 2) = 12 . January 8, 2018 DRAFT 40 It also holds that pTdec (Tdec ) = X X P (H = h)P (Ddec = d|H = h) pTdec (Tdec |H = h, Ddec = d)(127) h∈{1,2} d∈{1,2} = X X P (H = h)P (Ddec = d|H = h) pTdec (Tdec |H, Ddec = d) (128) h∈{1,2} d∈{1,2} = pTdec (Tdec |H, Ddec ) X X P (H = h)P (Ddec = d|H = h) (129) h∈{1,2} d∈{1,2} = pTdec (Tdec |H, Ddec ) (130) where in (128) we have used Theorem 1 and in (129) we have used Theorem 2. Using (126) and (130) we get for the argument of the log in (121) P (Ddec = 1)pTdec (Tdec |H, Ddec = 1) + P (Ddec = 2)pTdec (Tdec |H, Ddec = 2) pTdec (Tdec ) pT (Tdec |H, Ddec = 1) + pTdec (Tdec |H, Ddec = 2) = dec (131) 2 pTdec (Tdec |H, Ddec ) pT (Tdec |H, Ddec ) = dec (132) pTdec (Tdec |H, Ddec ) =1 (133) where we have applied again Theorem 2. This completes the proof. ACKNOWLEDGEMENT We acknowledge Yannis Kalaidzidis, Mostafa Khalili-Marandi, and Marino Zerial for fruitful discussions. R EFERENCES [1] M. Dörpinghaus, É. Roldán, I. Neri, H. Meyr, and F. Jülicher, “An information theoretic analysis of sequential decisionmaking,” in Proc. of the 2017 IEEE International Symposium on Information Theory (ISIT), Aachen, Germany, June 2017, pp. 3050–3054. [2] A. Wald, “Sequential tests of statistical hypotheses,” Ann. Math. Statist., vol. 16, no. 2, pp. 117–186, June 1945. [3] A. Wald and J. Wolfowitz, “Optimum character of the sequential probability ratio test,” Ann. Math. Statist., pp. 326–339, 1948. [4] A. Tartakovsky, I. Nikiforov, and M. Basseville, Sequential Analysis: Hypothesis Testing and Changepoint Detection, ser. Chapman & Hall/CRC Monographs on Statistics & Applied Probability. CRC Press, 2014. [5] T. L. Lai, “Asymptotic optimality of invariant sequential probability ratio tests,” The Annals of Statistics, pp. 318–333, 1981. [6] A. Tartakovsky, “Asymptotically optimal sequential tests for nonhomogeneous processes,” Sequential analysis, vol. 17, no. 1, pp. 33–61, 1998. January 8, 2018 DRAFT 41 [7] A. G. Tartakovsky, “Asymptotic optimality of certain multihypothesis sequential tests: Non-iid case,” Statistical Inference for Stochastic Processes, vol. 1, no. 3, pp. 265–295, 1998. [8] V. Draglia, A. G. Tartakovsky, and V. V. Veeravalli, “Multihypothesis sequential probability ratio tests. i. asymptotic optimality,” IEEE Transactions on Information Theory, vol. 45, no. 7, pp. 2448–2461, 1999. [9] R. Liptser and A. Shiryaev, Statistics of Random Processes: I. General Theory, ser. Applications of mathematics : stochastic modelling and applied Probability. Springer, 2001. [10] J. L. Melsa, D. L. Cohn et al., Decision and estimation theory. McGraw-Hill, 1978. [11] É. Roldán, I. Neri, M. Dörpinghaus, H. Meyr, and F. Jülicher, “Decision making in the arrow of time,” Physical Review Letters, vol. 115, no. 25, p. 250602, 2015. [12] I. Neri, E. Roldán, and F. Jülicher, “Statistics of infima and stopping times of entropy production and applications to active molecular processes,” Phys. Rev. X, vol. 7, p. 011019, Feb 2017. [13] D. Williams, Probability with martingales. Cambridge university press, 1991. [14] W. C. Lindsey and H. Meyr, “Complete statistical description of the phase-error process generated by correlative tracking systems,” IEEE Trans. Inf. Theory, vol. 23, no. 2, pp. 194–202, 1977. [15] M. H. DeGroot and M. J. Schervish, Probability and statistics, 4th ed. Addison-Wesley„ 2012. [16] J. Beirlant, E. J. Dudewicz, L. Györfi, and E. C. Van der Meulen, “Nonparametric entropy estimation: An overview,” International Journal of Mathematical and Statistical Sciences, vol. 6, no. 1, pp. 17–39, 1997. [17] J. Jiao, K. Venkat, Y. Han, and T. Weissman, “Minimax estimation of functionals of discrete distributions,” IEEE Transactions on Information Theory, vol. 61, no. 5, pp. 2835–2885, 2015. [18] A. W. van der Vaart, Asymptotic Statistics, 1st ed. Cambridge University Press, 1998, vol. 3. [19] R. Ratcliff and P. L. Smith, “A comparison of sequential sampling models for two-choice reaction time.” Psychological review, vol. 111, no. 2, p. 333, 2004. [20] R. Ratcliff and G. McKoon, “The diffusion decision model: theory and data for two-choice decision tasks,” Neural computation, vol. 20, no. 4, pp. 873–922, 2008. [21] S. Pigolotti, I. Neri, E. Roldán, and F. Jülicher, “Generic properties of stochastic entropy production,” Phys. Rev. Lett., vol. 119, p. 140604, Oct 2017. [Online]. Available: https://link.aps.org/doi/10.1103/PhysRevLett.119.140604 [22] S. Redner, A guide to first-passage processes. Cambridge University Press, 2001. [23] J. R. Michael, W. R. Schucany, and R. W. Haas, “Generating random variates using transformations with multiple roots,” The American Statistician, vol. 30, no. 2, pp. 88–90, 1976. [24] T. Cover and J. Thomas, Elements of Information Theory, 2nd edition. New York: Wiley & Sons, 2006. [25] M. N. Shadlen and W. T. Newsome, “Neural basis of a perceptual decision in the parietal cortex (area lip) of the rhesus monkey,” Journal of neurophysiology, vol. 86, no. 4, pp. 1916–1936, 2001. [26] J. D. Roitman and M. N. Shadlen, “Response of neurons in the lateral intraparietal area during a combined visual discrimination reaction time task,” Journal of neuroscience, vol. 22, no. 21, pp. 9475–9489, 2002. [27] S. Kira, T. Yang, and M. N. Shadlen, “A neural implementation of Wald’s sequential probability ratio test,” Neuron, vol. 85, no. 4, pp. 861–873, 2015. [28] R. Bogacz and K. Gurney, “The basal ganglia and cortex implement optimal decision making between alternative actions,” Neural computation, vol. 19, no. 2, pp. 442–477, 2007. [29] R. Bogacz, “Optimal decision-making theories: linking neurobiology with behaviour,” Trends in cognitive sciences, vol. 11, no. 3, pp. 118–125, 2007. [30] R. Losick and C. Desplan, “Stochasticity and cell fate,” science, vol. 320, no. 5872, pp. 65–68, 2008. January 8, 2018 DRAFT 42 [31] A. Raj and A. van Oudenaarden, “Nature, nurture, or chance: stochastic gene expression and its consequences,” Cell, vol. 135, no. 2, pp. 216–226, 2008. [32] E. D. Siggia and M. Vergassola, “Decisions on the fly in cellular sensory systems,” Proceedings of the National Academy of Sciences, vol. 110, no. 39, pp. E3704–E3712, 2013. [33] C. Guo and A. Kuh, “Temporal difference learning applied to sequential detection,” IEEE Transactions on Neural Networks, vol. 8, no. 2, pp. 278–287, Mar 1997. [34] D. Teng and E. Ertin, “Learning to aggregate information for sequential inferences,” arXiv preprint arXiv:1508.07964, 2015. [35] C. M. Bishop, Pattern recognition and machine learning. [36] A. Engel, Statistical mechanics of learning. January 8, 2018 Springer, 2006. Cambridge University Press, 2001. DRAFT 

9cs.NE

Robust Propensity Score Computation Method based on Machine Learning with Label-corrupted Data Chen Wanga , Suzhen Wangb,1 , Fuyan Shib , Zaixiang Wangb arXiv:1801.03132v1 [stat.ME] 9 Jan 2018 a Dept. of Computer Science, University College London, Gower Street, London, England, United Kingdom WC1E 6BT b Dept. of Health Statistics, Weifang Medical College, No.7166 BaoTong West Street, Weifang, Shandong, China 261053 Abstract In biostatistics, propensity score is a common approach to analyze the imbalance of covariate and process confounding covariates to eliminate differences between groups. While there are an abundant amount of methods to compute propensity score, a common issue of them is the corrupted labels in the dataset. For example, the data collected from the patients could contain samples that are treated mistakenly, and the computing methods could incorporate them as a misleading information. In this paper, we propose a Machine Learning-based method to handle the problem. Specifically, we utilize the fact that the majority of sample should be labeled with the correct instance and design an approach to first cluster the data with spectral clustering and then sample a new dataset with a distribution processed from the clustering results. The propensity score is computed by Xgboost, and a mathematical justification of our method is provided in this paper. The experimental results illustrate that xgboost propensity scores computing with the data processed by our method could outperform the same method with original data, and the advantages of our method increases as we add some artificial corruptions to the dataset. Meanwhile, the implementation of xgboost to compute propensity score for multiple treatments is also a pioneering work in the area. Keywords: Biostatistics, Propensity Score, Spectral Clustering, Data ∗ This research is granted by China National Natural Science Foundation, Grant: 81473071 1 Corresponding Author Sampling, Xgboost, Machine Learning 1. Introduction Confounding covariates, or alternatively named as noise features, is a significant problem in studying Biostatistics data. In practice, data in this field is usually collected with detailed information, thus it will contain some irrelevant and redundant features. For example, in a dataset of a certain disease, the information about the marital status might be irrelevant to the cause of the disease, and it will be acting as a confounding (noise) covariate. The confounding covariates could bring negative effects in analyzing the data: it could provide misleading information to reveal a false correlation between variables (features) and labels, and it could introduce unnecessary statistical differences between groups and lead to a incomparable situation. Consequently, recognizing and balancing confounding covariates have become an open task in biostatistics research. A common approach to address this problem is propensity score [1]. The key idea of propensity score is to compute the ’likelihood’(propensity) of a sample to be belonging into a certain group (usually termed as ’treatment’ in biostatistics). And by matching the samples in different groups with similar propensity scores and/or weighting the samples with propensity score, we could reduce or eliminate the bias of confounding features. [2] experimentally shows that for some randomly generated data, propensity score could effectively eliminate the statistical significance between groups. There are multiple methods to compute propensity score, ranging from simple logistic regression to cutting-edge Machine Learning methods. From the perspective of Machine Learning, the task of propensity score computing is essentially to train and utilize a classifier with probability as retrievable outputs (like soft-max output). Consequently, although there are a variety of Machine Learning algorithms which could be plunged to the task, the basic procedures are almost the same. While the state-of-the-art Machine Learning methods are promising in this area, they typically suffer from one problem: the corrupted data. Data is typically collected from medical practice, and there could be some mistreated samples in each group. The conventional approach to solving this problem would be to hire some experts to select the correctly-labeled samples and use them as a subset to train the classifier. However, this procedure will be 2 expensive in both time and money. To tackle the issue with a smarter approach, we propose a novel Machine Learning-based propensity score computation procedure in this paper. We exploit the fact that the majority of samples should be treated (classified) with the correct group (label) and patients (samples) be treated with different groups should have distinct data manifolds. Guided by this premise, we design a system that first clusters the data with spectral clustering method, and then computes the weight of the data by computing the proportion of the number of samples through different classes (treatments). The weights are processed with an interpolation method to perform a ’moderate softmax’ computation and it will lead to a distribution over clusters after being normalized. Based on this new distribution, we will sample a subset for the usage of the training procedure. And finally, the propensity score is computed based on Xgboost proposed by Chen and Guestrin [3]. The rest of the paper is arranged in the following order: the second section will briefly review literatures related to our research; the third part will introduce the proposed system and explain the Machine Learning algorithms applied; the fourth portion will be mathematically analyzing the assumptions and properties of the proposed method; the experimental result for the proposed method based on SEER data is illustrated in the fifth section, and the outcomes are analyzed and compared; and finally, the last section of the paper concludes our research and discusses future research for this topic. 2. Related Work Despite the solid theoretical foundation of conventional Logistic Regression method in computing propensity score, efforts to utilize novel Machine Learning algorithms in computing propensity score has emerged for a long time. Huge research potentials of propensity score could be found in the state-of-the-art algorithms, which could significantly outperform the traditional Logistic Regression. Existed papers related to this topic mainly focus on implementing Machine Learning methods with advanced models (such as SVM and Random Forest) and comparing different approaches with certain datasets. [4] tested different Machine Learning methods in computing propensity score and demonstrated that boosting methods and tree algorithms could outperform Logistic Regression in general. [5] studied the theoretical foundations of implementing Machine Learning methods for dichotomous and multicategory outcome in biostatistics, and analyzed algorithms 3 like kNN, Random Forest and SVM. [6] compared the propensity score results of Classification Tree Analysis (CTA) and Logistic Regression and argued that although Logistic Regression method still has the lowest average standardized difference, CTA could provide the greatest predictive accuracy thus it has strong potentials and could provide an alternative way in propensity score computation. And for the tree-based method specifically, [7] compared different classification and regression and discussed their performances. There are also published literatures concentrating on improving the model complexity to avoid potential model misspecified problem. A typical example of these algorithms is the so-called ’super-leaner’ algorithm, which is roughly equivalent to stacking algorithms in Machine Learning. The method was originally proposed by [8] to process data with complex relationship between input and output through a weighed combination of different learners. [9] shows experimentally that the super learner method could achieve a significantly better performance under the situation of severe model misspecification. [10] proposed an algorithm to use a convex combination of generalized propensity score computed by Super Learner. The authors then takes traumatic brain injury as a study case and discussed the statistical significance between time for the patient to be transferred from emergency to specialist and the treatment effect. [11] further studied the Super Learner method in the application of electronic healthcare data. Whilst there are a large amount of work discussing advanced algorithms in computing propensity score, approaches dealing with propensity score under corrupted data is relatively under-developed. [12] analyzed the problem of propensity score with missing data and suggested to use multiple imputation method. Similar with this paper, [13] analyzed propensity score computation under missing covariate with multiple imputation and specifically solved the issue with general location mixture models. To the best our knowledge, hitherto, there has not been any significant publication in studying propensity score with mislabeled data, although this is common in bioscience. In the area of Machine Learning, this issue could be regarded as corrupted label problem, and there are some related literatures building algorithms for this task [14][15]. However, in this work, we do not follow the idea of advanced algorithmic target functions employed in these methods. Instead, we used a clustering and normalization based method to tackle the problem. Based on our assumption, the method should be valid for biostatistic data and the rationality of the method is shown mathematically in our paper. 4 3. Methodology 3.1. Problem Setup Assuming we have data (X, Y ) that generated from space X × Y, where X is the input space and Y is a discrete output space. We now have m data pairs (X1 , y1 ), (X2 , y2 ), ..., (Xm , ym ) drawn from a certain distribution. Then, there will be an underlying conditional distribution: p(y|X; θ) (1) And to infer the probability of each data point, we need to obtain parameter θ that determines the model. With proper hypothesis space, the conventional MLE of θ should be: ∗ θ = argmax θ m Y p(yi |Xi ; θ) (2) i=1 However, here our situation is the set Y has been replaced by Ŷ , which contains many noisy(incorrect) labels. Then directly learning from (X1 , yˆ1 ), (X2 , yˆ2 ), ..., (Xm , yˆm ) will incorporate some noise. Thus, it demands an approach to solve the problem. Given the research context of medical and biological science, we could assume the probability for a sample to be labeled (treated) correctly is higher than the reverse situation. Mathematically, this assumption could be noted as: p(ŷ = a|y = a) > p(ŷ 6= a|y = a), ∀a ∈ Y (3) And under this assumption, we could try to sample a subset of (X1 , y˜1 ), (X2 , y˜2 ), ... , (Xk , y˜k ), k 6 m to fit the model, and find the parameter with: θ̃ = argmax θ k Y p(ỹi |Xi ; θ) (4) i=1 that could minimize the loss L(θ̃, θ∗ ) between the estimated parameter and the parameter estimated under uncorrupted data. However, since the true data set is not visible for us, it is impossible to directly optimize over L(θ̃, θ∗ ) function. An alternative way would be to sample from a distribution that could produce a subset close to the true dataset, and this is the core idea of our noise-eliminating procedure. The next subsection will be introducing this method and the employed Machine Learning algorithm. 5 3.2. Avoid Noisy Label: Sampling with Spectral Clustering and Normalization Now we need a method to sample a subset of data with the distribution as close as possible to the original correct samples. The method should have the following properties: • The proportion of mis-specified labels should monotonously decrease as the performance of the algorithm increase • The result should not be affected by the proportion of data in each class (treatment) • The results of the algorithm should be interpretable, and the interpretation should have a fixed methodology To satisfy the above principles, we could assume some special properties for the data. Without losing generality, we could assume the data, which is classified (treated) with d different classes, comes from n underlying patterns under condition n > d, and each pattern will have a proportion of pi , i = 1, 2, ..., n to be classified into class j ∈ {1, 2, .., d}. Formally, we could specify the number of true samples under each class j: Ntrue (j) = n X pi N (i) (5) i=1 Where N (·) denotes the number of samples. In practice, we choose the number of n as n = d for a better interpretability, and this setting will be further discussed in section 5 with the example of SEER data. Under this assumption, we will be able to sample the data with their underlying patterns, which could be represented by clusters under the skeleton of unsupervised learning. Intuitively, an effective method would be to sample from the clustering results, and now the question is which kind of technique should we use for clustering. The commonly-used clustering method is K-means and its variations. This branch of algorithms are straightforward to implement and the computational complexities of them are generally acceptable for ordinary computers. However, K-means related algorithms suffer from several problems: it is sensitive with initial clustering centers, and in practice we usually need to run 6 K-mean for several times with different initial clustering centroids to obtain the best performance; it is vulnerable towards outliers, as few outliers will make significant effects on the result; and most importantly, the clustering results of K-means branch algorithms have a strictly spherical curvature, thus it could only deal with simple data manifold, but in bio-medical science the data manifold could be of great complexity. A better alternative would be Spectral Clustering. Instead of measure the Euclidean distance between samples, which will inevitably loss some manifold information, Spectral Clustering method utilize the Graph Distance that could preserve local neighborhood informations. In Spectral Clustering, the affine matrix that denotes the Graph Distance and its corresponding Graph Laplacian will be computed, and then eigenvalue decomposition will be performed to capture the information in the so-called embedded space and the clustering result could be obtained by applying ’ordinary’ clustering methods on this processed matrix. Spectral Clustering works especially well when the number of clusters is not large, and it could deal with complicated data geometries. The core problem of Spectral Clustering is the Neighbor Graph (Similarity Graph), which represents the ’similarity’ between different data points. Common methods to be implemented include -Neighbourhood, knearest Neighbourhood and fully-connected graph with positive semi-definite kernels. One popular similarity graph matrix construction method is Gaussian Kernel method, which computes the similarity between data points xi and xj with: ||xi − xj ||22 ) (6) 2σ 2 Which σ is a parameter that could be specified or computed by the following equation: q Kg (xi , xj ) = exp(− σ= T median(I(||X̂ − X̂ ||2 )) (7) Where X̂ is a m × m matrix with X̂ ij = ||xi ||22 , ∀j ∈ 1, 2, ..., m (duplicated column matrix with each row as the square l2 norm of the data point) and I stands for the operation to remove the diagonal values of the matrix. The method could be computed with high efficiency for high dimension data. However, it could be of great instability if some of the features are discrete values. To improve the clustering method specifically for our data with a mixture of continuous and discrete features, we proposed a feature-wise Similarity Matrix method here. For each pair of samples (xi , xj ), we compute 7 Data: Input data, separated by continuous features and discrete features Result: Clustering result for each data point Compute the affine matrix of continuous features based on equation 6 ; for Each Discrete Feature do Compute the affine matrix of discrete features based on equation 9 and sum them with equation 8 ; end Perform Spectral Clustering with the affine matrix ; Return the clustering label for each data ; Algorithm 1: Spectral Clustering of the Data the sum of the similarity across each feature, which could be denoted by: K(xi , xj ) = P X Kd (xip , xjp ) + Kg (x∗i , x∗j ) (8) p=1 Where P is the number of discrete features(co-variates) and Kg (·, ·) stands for Gaussian Kernel for continuous features computed by equation 6. For the discrete features, we compute each of the similarity matrix with Delta Kernel: ( 1 , x i = xj (9) Kd (xi , xj ) = N (x) 0, otherwise Where N (x) is the number of samples of the specific class. And since all the components of the similarity is positive semi-definite, the resulting matrix would also be positive semi-definite. The Process of the detailed algorithm could be shown in algorithm 1. The practical implement of this method will be discussed in section 5. And after clustering, we could further divide each cluster with sample in different classes(treatments) and get a d × n matrix:   w11 w12 · · · w1n w21 w22 · · · w2n   W = (10) ···  wd1 wd2 · · · wdn Where d is the number of classes and n is the number of underlying patterns (clusters). 8 Intuitively,now we would be able to choose the top k clusters at each class(treatment) and use them as the new training set. However, the number of samples in each cluster, which serves are an indication of underlying patterns, could be imbalance and misguiding in our circumstance. A smarter way would be to normalize the quantity of each cluster across the classes, which could be representation as: wij ∗ wij = Pd i=1 wij (11) This will give us the exact information about ’how large is the portion of this pattern being assign with treatment i’, and it could best reflect and satisfy the assumption denoted by equation 5. And here we need to augment the probability of the cluster with largest probability because we have mis-specified data in each treatment class. Based on the assumption of 3, the augmented probabilities should be belonging to the clusters that should contribute to the treatment. And inspired by [16], we use an interpolation ∗ to get ŵ: method to combine the first and second order value of wij ∗ ŵ = wij + γ ∗2 wij 1−η∗ (12) Parameters γ denotes the trade-off between first and second order interpolation, and  is the proportion of corrupted labels (default as 0 even with some mild corruptions, and we assume  ≤ 0.5) and η is a parameter denoting how will we treat the importance of data corruption. Here we also have a constraint to η for 0 ≤ η ≤ 2. In our program we set γ = 0.7 and η = 2. And after the normalization, we could sample a new dataset with the proportion of ŵ in each row of the matrix. Formally, it could be denoted by p∗ij : wˆij p∗ij = Pn j=1 wˆij (13) Equation 13 is the equation we finally used in computing the probability to sample a subset from each treatment class. In section 4 we will discuss the rationality of it mathematically, but in this section, we will continue on discussing the technique to compute propensity score. 9 3.3. Propensity Score Computing with Xgboost As it is stated in section 2, implementing Machine Learning techniques in computing Propensity Score has been studied for a long time. Propensity Score computing is very close to the probabilistic output in supervised classification problem, thus a large amount of logistic or softmax functionbased approach could be introduced into the Propensity score area. Previous studies generally suggest that boosting method is one of the most promising branches of ML algorithms in computing propensity score. And in our work, we use Xgboost as a novel Machine Learning method (designed in 2016) to compute propensity score. Xgboost is actually not a newly-proposed algorithm. Instead, it is built on the algorithm of Gradient Boost Machine (GBM), and the novelty of this platform is that it considered Structural Learning scores and designed a better interface for Machine Learning considerations. The platform has R, C++, Python, Scalar and Java API, and it could be conveniently modified by simply change some parameters and arguments. It support softmax output for multiple-class probability, but using softmax could be too extreme. The package also provide another mode called ’softprob’ model, which could output the normalized probabilities of each class. And this is also the method we adopted in this work. Key parameters of Xgboost also include l2 regularization parameter, named ’lambda’, l1 regularization coefficient, named ’alpha’, and tree construction and growth method, which is often selected automatically by heuristic search. Meanwhile, since we are asking the Xgboost to output multi-class result, here we use multi-class log loss to denote our training and validation loss. The loss could be expressed by: l(Y , P ) = m X y Ti log(pi ) (14) i=1 Where y i is the n × 1 one-hot vector that has value 1 at the true output position and 0 otherwise; pi is the probability output vector. Y and P are the m × n matrices that denotes all the data. Xgboost is a well-developed platform which is ’efficient, flexible and portable’ (from their Github page). Here, we will not elaborate on the basic principles of the GBM algorithm it utilizes. One who interested in could refer to the original publication of the method. 10 3.4. Data Manifold Analysis with t-SNE Another point we might interest in is the visualization of the data, which could bring significant benefits for result interpretation. Visualizing the projection of the Manifold of data could provide us information about how data in different classes could be overlapped or separated. And sometimes it could even help us determine whether there exists significant difference between different groups (If there is not the experiment is probably unnecessary). Typical data visualization approaches often project the data manifold to a 2-d plane, which could be straightforwardly analyzed by human. Popular methods used to compute data visualization include PCA(and pPCA, kPCA, etc.), Isomap, Sammon Mapping, SNE(tSNE) and auto-encoder. The amjor drawbacks of PCA and mapping methods are that they could not preserve local structures and will thus lose the information of the data manifold in the higher dimension. On the contrary, SNE(tSNE) could manipulate the graph neibourhood structure. It is similar with Spectral Clustering method we mentioned before, the difference is that here we compute the data point in the lower dimension instead of clustering the Graph Laplacians. To begin with, we will start to introduce SNE method. SNE stands Stochastic Neighbour Embedding, and it start with the transition matrix between different data points in the original space, computed by: ( exp(−||(xi )−(xj )||22 /(2σ 2 )) P i 6= j 2 2 j=1,j6=i exp(−||(xi )−(xj )||2 /(2σ )) (15) pi|j = 0 i=j Similarly, we could compute the transition probabilities in the lower-dimension space, which could define: ( exp(−||(x∗i )−(x∗j )||22 /(2σ 2 )) P i 6= j ∗ ∗ 2 2 j=1,j6=i exp(−||(xi )−(xj )−||2 /(2σ )) qj|i = (16) 0 i=j And now our purpose is to fit the model to get the data x∗ that could preserve the structural information. SNE is a method that consider to minimize the KL divergence between pi|· and q·|i , and the target function could be describe as: 11 l((x∗ ), (x)) = X KL(pi|· |q·|i ) i = XX i pi|j log( j pi|j ) qj|i (17) Equation 17 could be solved by gradient descent. And t-SNE is a variate of SNE to modify the ’unsymmetrical’ KL-divergence. It defines a ’joint’ distribution of pi,j with: pi|j + pj|i (18) 2m Where m is the number of samples. The similar method could be used on q, and then we would have a ’symmetric’ target function. In this work, we utilized the tSNE package in sk-learn API and specified certain parameters to compute an optimal solution. Notice that in our situation features are partially discrete and partially continuous, thus we compute the distance between samples explicitly with Euclidean for continuous features and matching numbers for discrete features as the following equation: pi,j = N − Nm (19) N Where N is the number of features(covariates) Nm is the number of features that match in sample xi and xj . Computing continuous and discrete features separately could cause another problem: the imbalance of distance. Since the value of Dd (xi , xj ) has a maximum value of 1, where the distance of continuous data is not constrained, the effect of Dd (xi , xj ) could be under-valued. An approach to deal with this problem is to multiply a correction coefficient τ to Dd with value: Dd (xi , xj ) = τ= Nd max(Dc ) × Nc max(Dd ) (20) Where Nd is the number of discrete features and Nc is the amount of continuous features. max(Dc ) stands for the maximum value of continuous distance and max(Dd ) demotes the maximum value of discrete features. And the overall distance will be computed by: D(xi , xj ) = Dc (xi , xj ) + τ ∗ Dd (xi , xj ) 12 (21) This method is applied in our program, the details will be further illustrated in section 5. 4. Mathematical Analysis of the Proposed Method The novelties of our method mostly lie in the designing of the embedded graph with Gaussian and Delta Kernel (equation 8) and the clustering-based sampling method (equation 13). The rationality of the new graph-neightbour method is easy to prove, as the kernel methods used are symmetric and positive semi-definite. However, one might have doubt of our proposed method and its assumption and rationality. In this section, we will analyze the assumptions and the mathematical properties of the proposed method, and show that the method could have an analytical upper bound of error rate. The basic assumption has been stated in section 3.2 with equation 5. Here to further illustrate the assumption, we could denote the sample number of underlying patterns and the classes with the following matrix:      N1 p11 p12 · · · p1n c1 N2   p21 p22 · · · p2n   c2   =   (22) · · · · · · · · · · · ·  · · · Nd pd1 pd2 · · · pdn cn Where Ni means the sample number of class i = 1, 2, ..., d and cj means the number of samples belonging to underlying pattern j ∈ 1, 2, ..., n. The component pij of the d × n matrix P denotes the proportion of underlying cluster j to be assigned to category (treatment) i. According to equation 10, data after clustering could also be partitioned into different parts with the d × n matrix W . Now we assume that the corrupted data (with mislabeled samples) could be denoted as:  ∗  ∗   N1 p11 p∗12 · · · p∗1n c1 N2∗   p∗21 p∗22 · · · p∗2n   c2   =   (23)  · · ·  · · · · · · · · ·  · · · cn Nd∗ p∗d1 p∗d2 · · · p∗dn And here P ∗ = p∗ij could be regarded as the true distribution proportion matrix with corrupted (misclassified/mistreated) samples. And our clustering method is performed upon this matrix. To satisfy our assumptions, the matrix should have the following property: 13 d X p∗ij = 1 (24) i=1 To make sure that all samples from each cluster will be fully assigned to different classes. And based on our assumption in equation 3, the following property should hold as the number of the samples asymptotically grows to infinity: X X (25) p∗ki ci > p∗kj cj , ∀k ∈ 1, 2, ..., d i∈I j∈I Where I is the set of the pattern symbols that pki > 0 and J is the set with originally pkj = 0, which means that cluster j should not have any member been assigned to class k. For each class k ∈ 1, 2, .., d, the whole number of data is composed of: Nk∗ = X p∗ki Pd i∈I k=1 p∗ki + p∗kj X Pd j∈J k=1 p∗kj (26) In the above equation, the first term of the right hand side comes from the clusters that should be assigned to the class and the second term comes as pure mistakes. The error-specified size of sample could then be calculated as: ε= X |pki − p∗ |ci X p∗kj cj ki + Pd Pd ∗ ∗ p ki k=1 k=1 pkj j∈J i∈I (27) Plugging in the equation 24 and 25 to equation 27, we could get: X X ε= |pki − p∗ki |ci + p∗kj cj i∈I < X j∈J |pki − p∗ki |ci i∈I = X i∈I1 + X p∗ki ci (28) i∈I pki ci + X 2(p∗ki − pki )ci i∈I2 Where I1 stands for the set that p∗ki 6 pki , and I2 denotes the set of i that p∗ki > pki . And by computing the error rate η with ErrorRate = ErrorSamples/(ErrorSamples + CorrectSamples), we could get the following result 14 from equation 27: P P ∗ i∈I1 pki ci + i∈I2 2(pki − pki )ci P η<P ∗ ∗ i∈I1 (pki ci + pki ci ) + i∈I2 2pki ci (29) And now, we could further analyze equation 29 with two special cases. 1. When p∗ki 6 pki , ∀i ∈ 1, 2, ..., n. P Then we will have a number of correct samples as i∈I p∗ki ci . The error rate would therefore be: P i∈I pki ci (30) η<P ∗ i∈I (pki ci + pki ci ) And equation 30 would be the upper bound of the error rate under this situation. As we could see, in this case the function is a decrement function w.r.t p∗ki , and when p∗ki = pki ∀i, which means at this special point, the upper bound of error would have a minimal value of η < 0.5. 2. When p∗ki > pki ∀i ∈ 1, 2, ..., n. P Then we will have a number of correct samples as i∈I pki ci . The error rate would therefore be: P 2(p∗ − pki )ci P ki ∗ (31) η < i∈I i∈I 2pki ci The equation 31 is the upper bound of the second situation and it is an increment function. As we could prove the minimal of this upper bound is also 0.5, and the situation is still p∗ki = pki ∀i. This comes with the property of a continuous upper bound of the method. And in the above method we have shown the assumptions of the method and why it is rational. In practice, if the clustering algorithm is strong enough to capture the majority of correct-specified samples, then p∗ki should be close to pki given the assumption in 3. 5. Practical Settings and Experimental Results In addition to algorithm designing and analyzing, in this work parameter setting problems are also discussed and experimental results are obtained based on SEER data. This section will first introduce the SEER data, and then discuss (hyper)parameter setting problems. And finally, the experimental results with our method will be demonstrated. 15 5.1. SEER Dataset The SEER dataset from 2004 to 2014 was queried for patients who were confirmed pancreatic adenocarcinoma defined according to the International Classification of Disease for Oncology, Third Edition (ICD-O-3) codes for morphology (8140 and 8500) and topography (C25.0, C25.1, C25.2, C25.3, C25.7, C25.8, and C25.9). Data collected for each patient includes patient characteristics (age at diagnosis, sex, race, and marital status), tumor characteristics (tumor location, tumor grade, and AJCC stage), and treatment characteristics (type of surgery, radiotherapy and status of surgery adjuvant radiotherapy). These three treatments are considered as the three classes/categories under the term of Machine Learning in out work. There are some missing data in the dataset, and discussing them will be out of our research scope. Thus, in our program we only consider the features(covariates) without data missing, and the rationality of the selection of features could be supported from the perspective of bioscience and medical science. According to the true meaning of each feature, there only four features which could be regarded as continuous: EOD10 PN, CS SIZE, CS EXT, and CS NODE. Some of the features could imply the treatment (say, NO SURG flag) or should not be treated as feature (say, CASENUM, the index of patient), so we also remove these features. The original data is believed to only have mild mis-labeled samples (mistreated patients). To demonstrate the performance of our method under data with different extent of corruption, we also generate some further-corrupted data for training purpose. We produce data with 10%, 20% and 40% of artificial label corruptions based on the original data by corrupting the labels to other two classes with equal opportunities. Notice that the artificially corrupted data is only utilized for training, and when it comes to evaluation, we will use the original data to test the robustness of our proposed approach. 5.2. Parameter Settings Following the method proposed in section 3, there are some hyper- parameters that should be assigned based on our assumption and practical Requirement. Here we consider three aspects of parameter setting, namely the number of clusters, Spectral Clustering and Xgboost parameter settings, and considerations of tSNE. The first hyper-parameter to determine is the number of clusters we would like to use. This reflects the our basic assumption of the number of underlying patterns of the data. In our program we choose n = d where n is the 16 number of clusters and d as the number of categories (classes/treatments). This setting is for a better interpretation property: if there are no overlaps between each classes (mutually exclusive, say partial gastrectomy and total gastrectomy), then ideally the clustering result should have a distinct distribution between different classes and the correctly labeled samples will be straightforward to discover; otherwise, if there are overlaps between different classes/treatments(say in our SEER data situation, there are operation treatment, radian treatment and operation+radian treatment), then this setting will be convenient for us to analyze the demanded components of each class. In the subsection 5.3 we will demonstrate the interpretation of the SEER data we used in the experiment based on this scheme of cluster number setting. The second topic of parameter setting is our selections for the Spectral Clustering and Xgboost API. In our work, we use the Spectral Cluster method provided by SK-learn package with Python API. In the package, we specify to use Arpack to compute the eigenvalue decomposition. This will lead to a slightly higher time consumption, but the result will be significantly more stable. The final clustering results are obtained by K-mean over the embedded space, and to obtain a close-to-optimal result, the method will compute K-means in the embedded space with 10 different initializations to get the best result. And in our program, the σ value used in equation 6 could either be specified or be computed via equation 7. The parameters of Xgboost has been discussed in section 3.3. In practice, we set tree maximum depth as 7, and η value as 0.5 to control the learning with a relatively conservative scheme. We specified the max training iteration as 10 rounds for original data and 5 rounds for artificially corrupted data, as this setting could provide as an optimal validation error. The third point to be considered in practice is the tSNE visualization technique. In this project, we use the tSNE package provided by SK-learn. The authors of the original paper proposed tSNE suggested that hyper-parameters of tSNE is robust and won’t make too large an effect on the result. However, in our practical experiment, things are much more tricky: the parameter of ’perplexity’, which determines the balance between local manifold and global manifold, could significantly affect the clustering result and the optimal setting of this parameter will vary with the number of samples for visualization. Briefly speaking, the quantity of perplexity should be in the same magnitude of number of samples. For example, if we have 300 samples, then a perplexity of 5-50 will be good for visualization; however, if we have 30000 sample, then 17 it seems the value of perplexity score should be adjusted to a larger value accordingly. In our program, we visualized the project of data in different clusters and treatments(classes) with tSNE respectively. For the cluster illustration, we randomly choose 50 samples from each cluster (150 sample in total) and set perplexity=90, learning rate = 7 and use our precomputed distance with equation 21; And for the classes illustration problem, we choose 100 samples from each treatment group (before and after processed by our sampling method) and set perplexity=180 and learning rate = 7. 5.3. Experimental Result There are in total 8683 valid records, with 3531 being treated with method 1, 3200 with method 2 and 1952 with method 3. And the result of clustering divide the data into a proportion of 2905 in cluster 1, 3817 in cluster 2 and 1961 in cluster 3. Specifically, the treatment-cluster data segment result could be shown in table 1: And the sampling probability computed by equation 13 could be illustrated Cluster 1 New Treatment 1 210 New Treatment 2 2615 New Treatment 3 80 Cluster 2 2159 250 1408 Cluster 3 1162 335 464 Table 1: Treatment-Clustering Result in table 2. As we mentioned in the previous sections, the setting of d = n could helps us Treatment 1 Treatment 2 Treatment 3 Cluster 1 0.0587 0.7921 0.0435 Cluster 2 0.4597 0.0576 0.5827 Cluster 3 0.4816 0.1503 0.3738 Table 2: Probability to sample for new data explain the result. In table 1 and 2, it could be derived that: 1. patient characterized by cluster 1 should be predominately treated with treatment 2 (radian therapy); 2. patients characterized by cluster 3 and cluster 1 prefer to be treated by a treatment involved with operations, but those 18 who belongs to cluster 3 are more likely to reject radian therapy. The illustration of the clustering result in 2-dimension could be shown as figure 1. The result is produced with the settings stated in section 5.2 and we could find out that the three clusters have clear boundaries against each other, implying that the result should be rational. The red points, which represent data from cluster 1, have a more distinct boundary comparing to cluster 2 and 3, and this could be a supporting evidence for patients characterized by cluster 1 to be treated by a separate method. The 2-d projection of training sets before and after process could be il- Figure 1: Clustering Result in 2d. In the figure, red samples are from cluster 1, blue samples are from cluster 2 and green samples are from cluster 3. The group of red samples has a clear boundary with the rest two clusters, comparing the relatively vague boundary between cluster 2 and 3 lustrated in figure 2. The figure could prove the rationality of our process and the overlap between treatment 1 and 3. Meanwhile, we could find out the data manifolds are more compact and have multiple ’clusters’ inside the class, indicating that the dataset after process could be better categorized comparing to those before process. This is more significant in figure 3 as we specifically illustrate the manifold of treatment 1 and treatment 3. To evaluate the quality of the computed propensity score, we need to compute the confounding metric after applying propensity score to the original data. Here we consider propensity score weighting and employ Standardized Bias (SB) for several key features under ATE settings according to [17]. The metric could denote the level of confound between two groups (treatment and control groups) for each feature (covariate) and if the confounding level is lower (smaller in the metric) then it indicate that the data after propensity score weighted is better randomized. The evaluation method 19 Figure 2: The comparison of data projection in 2-d plane of training set before and after the proposed process. left: data manifold before process; right: data manifold before process. In the figure, red samples are from treatment 1, blue samples are from treatment 3 and green samples are from treatment 3. Figure 3: The comparison of data projection inside treatment 1 and 2 in 2-d plane of training set before and after the proposed process. In the figure, red samples are from treatment 1 and green samples are from treatment 3. top: data manifold before process; bottom: data manifold before process. From the figure, we could find out that data manifold after our process tends to have some difference between treatment 1 and 3, although they are still very similar. 20 (a) Comparison of data (b) Comparison of data (c) Comparison of data with 20% corrupted with 10% corrupted with 40% corrupted Figure 4: Data manifold before and after sampling procedure for data that added artificial label corruptions. red samples are from treatment 1, blue samples are from treatment 3 and green samples are from treatment 3. For each sub-figure, left: before precess, right: after process Age Processed Data + 0.165 Xgboost Original Data + 0.138 Xgboost Raw Data with- 0.089 out Propensity Weighting Race CS NODE 0.0025 0.0085 CS SIZE 0.2054 D AJCC S 0.0695 Survive Month 0.3118 0.0241 0.1096 0.0519 0.1377 0.1315 0.0446 0.2346 0.3857 0.7434 0.5806 Table 3: Standardized Bias between treatment group 1 and 2
could only compare two groups (classes) so that here we have 32 = 3 combinations. The comparison of the Standardized Bias could be found in table 3, 4 and 5. Here, the Processed Data was trained with xgboost for 10 iterations and the Original Data is a random subset of 3000 samples from the whole dataset. Final training and testing mlogloss on Processed Data are 0.243960 and 0.552677; Final training and testing mlogloss on Original Data are 0.222134 and 0.500497. To demonstrate the effect of noise-correction for further corrupted data, we also compared the Standardized Bias between treatment 1 and treatment 2 for the data with 10%, 20% and 40% artificial corrupted label in table 6, 7 and 8 respectively. Notice here when computing the Standardized Bias with the whole dataset, we use the authentic data. This will be the common 21 Age Processed Data + 0.052 Xgboost Original Data + 0.134 Xgboost Raw Data with- 0.137 out Propensity Weighting Race CS NODE 0.0085 0.1597 CS SIZE 0.2363 D AJCC S 0.1560 Survive Month 0.0218 0.0051 0.3821 0.4261 0.3287 0.0714 0.0605 0.0364 0.0875 0.0307 0.1113 Table 4: Standardized Bias between treatment group 1 and 3 Age Processed Data + 0.238 Xgboost Original Data + 0.297 Xgboost Raw Data with- 0.246 out Propensity Weighting Race CS NODE 0.0134 0.0981 CS SIZE 0.3440 D AJCC S 0.2042 Survive Month 0.3237 0.0357 0.1130 0.3157 0.4209 0.2432 0.0195 0.1997 0.4235 0.8343 0.8037 Table 5: Standardized Bias between treatment group 2 and 3 setting in real-life situations. Also notice that for the corrupted experiments the xgboost only computed for 5 iteration for the overfitting consideration. From the table, we could find that 1. propensity score computed by processed data with xgboost could usually outperform the scores computed by original sampled data with the same method; 2. the advantages of processed data will increase as the proportion of corrupted data increase 3. xgboost generally has a excellent and robust performance in computing propensity scores, and the propensity score computed could still improve the randomization even with 40% corrupted labels. 6. Conclusion In this paper, a novel approach to deal with the corrupted-label problem in computing propensity score for medical data is proposed and its property 22 Age Processed Data + Xgboost Original Data + Xgboost Race 0.053 CS NODE 0.0269 0.0932 CS SIZE 0.3351 D AJCC S 0.4977 Survive Month 0.3784 0.045 0.0666 0.0811 0.1496 0.4988 0.4201 Table 6: Standardized Bias between treatment group 1 and 2 with 10% labels corrupted Age Processed Data + Xgboost Original Data + Xgboost Race 0.059 CS NODE 0.0553 0.0149 CS SIZE 0.2696 D AJCC S 0.4712 Survive Month 0.4745 0.037 0.0474 0.1238 0.2757 0.5243 0.5577 Table 7: Standardized Bias between treatment group 1 and 2 with 20% labels corrupted is analyzed. The Xgboost approach is utilized to compute the propensity score and the performance of the integrated approach is examined and compared. TSNE method is employed for comprehensive purpose and the data manifold of different clusters and treatments are demonstrated. The paper made following major contributions to the field: Firstly, the paper finds out the mislabeling-data problem in medical science and researched on this topic. Given the fact that the problem has been under-investigated before but the phenomenon is very common, the research would significantly fill the blank of this aspect in bioscience research; Secondly, the paper proposed an effective method for getting a relatively authentic subset of training data, and proved mathematically about its rationality and effectiveness. The rigorous proof makes it possible for this method to be utilized into general Machine Learning problems and solve a broader scope of tasks. Thirdly, the paper employed Xgboost and tSNE methods to this area, which are relatively novel attempts. The results imply a huge research potential in this area. In the future, the author would like to concentrate more on mislabeling data in bioscience area. We intend to try penalization-based label correction methods proposed in the field of Machine Learning, and design new algorithms 23 Age Processed Data + Xgboost Original Data + Xgboost Race CS NODE 0.0250 0.0197 0.0968 CS SIZE 0.3667 D AJCC S 0.5715 Survive Month 0.4491 0.0968 0.0244 0.1230 0.3194 0.5793 0.5826 Table 8: Standardized Bias between treatment group 1 and 2 with 40% labels corrupted specifically in dealing with medical data. Meanwhile, to analyze the confounding covariates, we plan to research on graphical model-based methods in the long run. 24 References [1] P. R. Rosenbaum, D. B. Rubin, The central role of the propensity score in observational studies for causal effects, Biometrika 70 (1983) 41–55. [2] W. Meng, S. Wang, J. Xia, J. Lyu, F. Shi, A simulation study for the treatment effect using propensity score matching, Unpublished (2015). [3] T. Chen, C. Guestrin, Xgboost: A scalable tree boosting system, CoRR abs/1603.02754 (2016). [4] Propensity score estimation: neural networks, support vector machines, decision trees (cart), and meta-classifiers as alternatives to logistic regression, Journal of Clinical Epidemiology 63 (2010) 826 – 833. [5] J. Kruppa, Y. Liu, G. Biau, M. Kohler, I. R. Knig, J. D. Malley, A. Ziegler, Probability estimation with machine learning methods for dichotomous and multicategory outcome: Theory, Biometrical Journal 56 (2014). [6] A. Linden, P. R. Yarnold, Using classification tree analysis to generate propensity score weights, Journal of Evaluation in Clinical Practice 23 (2017) 703–712. JECP-2017-0075. [7] B. K. Lee, J. Lessler, E. A. Stuart, Improving propensity score weighting using machine learning, Statistics in medicine 29 (2010). [8] van der Laan, E. M., Polley, A. Hubbard, Super learner, Statistical Applications in Genetics and Molecular Biology 6 (2007). [9] R. Pirracchio, M. L. Petersen, M. van der Laan, Improving propensity score estimators’ robustness to model misspecification using super learner, American Journal of Epidemiology 181 (2015) 108–119. [10] N. Kreif, R. Grieve, I. Daz, D. Harrison, Evaluation of the effect of a continuous treatment: A machine learning approach with an application to treatment for traumatic brain injury, Health Economics 24 (2015) 1213–1228. [11] C. Ju, M. Combs, S. D. Lendle, J. M. Franklin, R. Wyss, S. Schneeweiss, M. J. van der Laan, Propensity score prediction for electronic healthcare 25 databases using Super Learner and High-dimensional Propensity Score Methods, ArXiv e-prints (2017). [12] C. Leyrat, S. R. Seaman, I. R. White, I. Douglas, L. Smeeth, J. Kim, M. Resche-Rigon, J. R. Carpenter, E. J. Williamson, Propensity score analysis with partially observed confounders: how should multiple imputation be used?, ArXiv e-prints (2016). [13] R. Mitra, J. P. Reiter, Estimating propensity scores with missing covariate data using general location mixture models, Statistics in Medicine (2011) 627–41. [14] N. Natarajan, I. S. Dhillon, P. K. Ravikumar, A. Tewari, Learning with noisy labels, in: C. J. C. Burges, L. Bottou, M. Welling, Z. Ghahramani, K. Q. Weinberger (Eds.), Advances in Neural Information Processing Systems 26, Curran Associates, Inc., 2013, pp. 1196–1204. [15] Y. Li, J. Yang, Y. Song, L. Cao, J. Luo, J. Li, Learning from noisy labels with distillation, CoRR abs/1703.02391 (2017). [16] A. de Brébisson, P. Vincent, An exploration of softmax alternatives belonging to the spherical loss family, CoRR abs/1511.05042 (2015). [17] D. F. McCaffrey, B. Griffin, D. Almirall, M. Slaughter, R. Ramchand, L. Burgette, A tutorial on propensity score estimation for multiple treatments using generalized boosted models, Statistics in medicine 32 (2013) 3388–3414. 26 

2cs.AI

arXiv:1509.01519v3 [math.AC] 4 May 2017 THE SUPPORT OF LOCAL COHOMOLOGY MODULES MORDECHAI KATZMAN AND WENLIANG ZHANG Abstract. We describe the support of F -finite F -modules over polynomial rings R of prime characteristic. Our description yields an algorithm to compute the support of such modules; the complexity of our algorithm is also analyzed. To the best of our knowledge, this is the first algorithm to avoid extensive use of Gröbner bases and hence of substantial practical value. We also use the idea behind this algorithm to prove that the support of HIj (S) is Zariski closed for each ideal I of S where R is noetherian commutative ring of prime characteristic with finitely many isolated singular points and S = R/gR (g ∈ R). 1. Introduction Local cohomology is a powerful tool introduced by Alexander Grothendieck in the 1960’s ([Har67]) and it has since yielded many geometric and algebraic insights. From an algebraic point of view, given an ideal I in a commutative ring R, local cohomology modules HiI (−) (i ≥ 0) arise as right-derived functors of the torsion functor on R-modules given by ΓI (M ) = {a ∈ M | I k a = 0 for some k ≥ 0}. A central question in the theory of local cohomology is to determine for which values of i does the local cohomology module HiI (M ) vanish. This question is both useful and difficult even in the case where R is a regular local ring and M = R, and this case has been studied intensely since the introduction of local cohomology (e.g., cf. [Har68], [PS73] and [Ogu73]). The aim of this paper is to describe the support of local cohomology modules in prime characteristic. Specifically, we first study the support of F -finite F -modules over polynomial rings R and show a computationally feasible method for computing these without the need to compute generating roots. To the best of our knowledge, this is the first computationally feasible algorithm for calculating the support of these modules in prime characteristic. We then apply this to the calculation of supports of local cohomology modules and of iterated local cohomology modules    in i2 i1 thus, for example, giving an effective method HI1 HI2 . . . HIn (R) . . . for determining the vanishing of Lyubeznik numbers. 2010 Mathematics Subject Classification. 13D45, 13A35. Key words and phrases. local cohomology, prime characteristic. M.K. gratefully acknowledges support from EPSRC grant EP/J005436/1. W.Z. is partially supported by NSF grants DMS #1405602/#1606414. 1 2 MORDECHAI KATZMAN AND WENLIANG ZHANG Our methods are interesting both from theoretical and practical points of view. A careful analysis of the algorithms resulting from these methods (see Section 4 below) shows that (a) the degrees of the polynomials appearing in the calculations have a low upper bound, and, furthermore, (b) when the method is applied to the calculation of supports of local cohomology modules, if the input is given by polynomials with integer coefficients, then the calculation of supports modulo different primes p involves polynomials whose degrees can be bounded from above by a constant times p, that constant being independent of p. In [Lyu97]) Gennady Lyubeznik described an algorithm for computing the support of F -finite F -modules. That algorithm requires the calculation for roots of these modules, and this relies on the repeated calculation of Grobner bases; these are often too complex to be computed in practice. Our algorithm consists of an iterative procedure (as described in section 3) which produces a quotient of a finite-rank free module with the same support as the given F -finite F -module. To find the support itself one needs to find a presentation for this submodule as a cokernel of a matrix: the support is then defined by the ideal of maximal minors of that matrix. Crucially, the iterative procedure above does not require the calculation of Gröbner bases, and consists essentially of matrix multiplications together with the listing of terms of polynomials whose degrees are bounded by a constant (independent of p) times p. The final step of the algorithm, finding a presentation of a finitely generated module, requires the calculation of one module of syzygies, hence the calculation of one Gröber basis. It is this that makes our algorithm a practical tool for computing supports of F -finite F -modules.1 The reason why we are able to compute and analyze in characteristic p the support of F -finite F -modules is the existence of the eth iterated Frobenius e endomorphism f e : R → R, taking a ∈ R to ap (e ≥ 0). The usefulness of these lies in the fact that given an R-module M , we may endow it with a new R-module structure via f e : let F∗e M denote the additive Abelian group M denoting its elements {F∗e m | m ∈ M }, and endow F∗e M with the R-module e structure is given by aF∗e m = F∗e ap m for all a ∈ R and m ∈ M . This also allows us to define the eth Frobenius functors from the category of R-modules to itself given by FRe (M ) = F∗e R ⊗R M and viewing this as a R-module via the identification of F∗e R with R: the resulting R-module structure on FRe (M ) satisfies a(F∗e b ⊗ m) = F∗e ab ⊗ m and F∗e ap b ⊗ m = F∗e b ⊗ am for all a, b ∈ R and m ∈ M . We will be interested in this construction mainly for regular rings and henceforth in this paper R will denote a regular ring of characteristic p > 0. 1The various algorithms in this paper have been incorporated in the “FSing” package of Macaulay 2[GS]. THE SUPPORT OF LOCAL COHOMOLOGY MODULES 3 Recall that an F -finite FR -module M is an R-module obtained as a direct limit of a direct limit system of the form U F 2 (U ) F 1 (U ) M− → FR1 (M ) −−R−−→ FR2 (M ) −−R−−→ . . . where M is a finitely generated module and U is an R-linear map (cf. [Lyu97]). The main interest in F -finite F -modules follows from the fact that local cohomology modules are F -finite F -modules, as we now explain. The jth local cohomology module of M with support on an ideal I ⊂ R is defined as (1) e HjI (M ) = lim ExtjR (R/I [p ] , M ) → e where maps in the direct limit system are induced by the surjections R/I [p e R/I [p ] . If we apply this with M = R, we obtain e+1 ] e HjI (R) = lim ExtjR (R/I [p ] , R) → e ∼ ExtjR (FRe (R/I), FRe R) = lim → e   j e ∼ Ext (R/I, R) F = lim R R → e e where we use the facts that FRe (R) ∼ = R, FRe (R/I) ∼ = R/I [p ] , and that, e since R is regular, the Frobenius functor FR (−) is exact and thus commutes with the computation of cohomology. This shows that HjI (R) are F -finite F -modules, and we may apply our F -finite F -module machinery to them. Finally, in section 7 we turn our attention to hypersurfaces and describe the support of their local cohomology modules, which turn out to be closed.2 Given a fixed g ∈ R, one can ask for the locus of primes P ⊆ R for which g → HiI (RP ) is injective and the locus the multiplication by g map HiI (RP ) − of primes for which this is surjective. We show that these two loci are Zariski closed by describing explicitly the defining ideals of these loci, and we use these to describe the defining ideal of the (Zariski closed) support for HiI (R/gR). We also extend the Zariski-closedness of HiI (R/gR) to the case when R has finitely many isolated singular points. The methods used for the various calculations in this paper are described in section 2. 2. Prime characteristic tools Definition 2.1. Let e ≥ 0. Let T be a commutative ring of prime characteristic p. 2The fact that the support is closed was simultaneously and independently also discov- ered by Mel Hochster and Luis Núñez-Betancourt in [HNB] using a different method. → 4 MORDECHAI KATZMAN AND WENLIANG ZHANG e (a) Given any matrix (or vector) A with entries in T , we define A[p ] to be the matrix obtained from A by raising its entries to the pe th power. e (b) Given any submodule K ⊆ T α , we define K [p ] to be the R-submodule e of T α generated by {v [p ] | v ∈ K}. Henceforth in this section, T will denote a regular ring with the property that F∗e T are intersection flat T -modules for all e ≥ 0, i.e., for any family of T -modules {Mλ }λ∈Λ , \ \ F∗e T ⊗T Mλ = F∗e T ⊗T Mλ . λ∈Λ λ∈Λ F∗e T are free T -modules (e. g. , polyThese include rings T for which nomial rings and power series rings with F -finite coefficient rings,) and also all complete regular rings (cf. [Kat08, Proposition 5.3]). These rings have that property that for any collection of submodules {Lλ }λ∈Λ of T α ,
[pe ] T T [pe ] = λ∈Λ Lλ : indeed, the regularity of T implies that for λ∈Λ Lλ e any submodule L ⊆ T α , L[p ] can be identified with FTe (L) and and the intersection-flatness of F∗e T implies \ \ \ \ FTe ( Lλ ) = F∗e T ⊗T Lλ = F∗e T ⊗T Lλ = FTe (Lλ ). λ∈Λ λ∈Λ λ∈Λ λ∈Λ The theorem below extends the Ie (−) operation defined on ideals in [Kat08, Section 5] and in [BMS08, Definition 2.2] (where it is denoted e (−)[1/p ] ) to submodules of free R-modules. Theorem 2.2. Let e ≥ 1. Given a submodule K ⊆ T α there exists a e minimal submodule L ⊆ T α for which K ⊆ L[p ] . We denote this minimal submodule Ie (K). Proof. Let L be the intersection of all submodules M ⊆ T α for which K ⊆ e e M [p ] . The intersection-flatness of T implies that K ⊆ L[p ] and clearly, L is minimal with this property.  When F∗e T is T -free, this is a straightforward generalization of the calculation of Ie for ideals. To do so, fix a free basis B for F∗e T and note that P [pe ] every element v ∈ T α can be expressed uniquely in the form v = b∈B ub b where ub ∈ T α for all b ∈ B. Proposition 2.3. Let e ≥ 1. (a) For any submodules V1 , . . . , Vℓ ⊆ Rn , Ie (V1 + · · · + Vℓ ) = Ie (V1 ) + · · · + Ie (Vℓ ). (b) Let B be a free basis for F∗e T . Let v ∈ Rα and let X [pe] v= ub b b∈B THE SUPPORT OF LOCAL COHOMOLOGY MODULES 5 be the unique expression for v where ub ∈ T α for all b ∈ B. Then Ie (T v) is the submodule W of T α generated by {ub | b ∈ B}. Proof. The proof of this proposition is a straightforward modification of the proofs of propositions 5.2 and 5.6 in [Kat08] and Lemma 2.4 in [BMS08]. Clearly, Ie (V1 +· · ·+Vℓ ) ⊇ Ie (Vi ) for all 1 ≤ i ≤ ℓ, hence Ie (V1 +· · ·+Vℓ ) ⊇ Ie (V1 ) + · · · + Ie (Vℓ ). On the other hand e e e (Ie (V1 ) + · · · + Ie (Vℓ ))[p ] = Ie (V1 )[p ] + · · · + Ie (Vℓ )[p ] ⊇ V1 + · · · + Vℓ and the minimality of Ie (V1 + · · · + Vℓ ) implies that Ie (V1 + · · · + Vℓ ) ⊆ Ie (V1 ) + · · · + Ie (Vℓ ) and (a) follows. e Clearly v ∈ W [p ] , and so Ie (T v) ⊆ W . On the other hand, let W be a P e [pe ] submodule of T α such that v ∈ W [p ] . Write v = si=1 ri wi for ri ∈ T and P pe wi ∈ W for all 1 ≤ i ≤ s, and for each such i write ri = b∈B rbi b where rbi ∈ T for all b ∈ B. Now ! s X [pe] X X e e [p ] p b ub b = v = wi rbi b∈B b∈B i=1 [pe ] and are direct sums, we comparePcoefficients and obtain ub = P sincee these [pe ] p s for all b ∈ B and so ub = ( si=1 rbi wi ) for all b ∈ B hence i=1 rbi wi ub ∈ W for all b ∈ B.  The behavior of the Ie operation under localization and completion will be crucial for obtaining the results of this paper. To investigate this we need the following generalization of [LS01, Lemma 6.6]. Lemma 2.4. Let T be a completion of T at a prime ideal P . Let α ≥ 0 and e e let W be a submodule of T α . For all e ≥ 0, W [p ] ∩ T = (W ∩ T )[p ] . Proof. If T is local with maximal ideal P , the result follows from a straightforward modification of the proof of [LS01, Lemma 6.6]. We now reduce the general case to the previous case which implies that e e W [p ] ∩ TP = (W ∩ TP )[p ] . Intersecting with T now gives e e e e W [p ] ∩ T = (W ∩ TP )[p ] ∩ T = (W ∩ TP ∩ T )[p ] = (W ∩ T )[p ] .  Lemma 2.5 (cf. [Mur13]). Let T be a localization of T or a completion at a prime ideal. For all e ≥ 1, and all submodules V ⊆ T α , Ie (V ⊗T T) exists and equals Ie (V ) ⊗T T. e Proof. Let L ⊆ T α be a submodule, such that L[p ] ⊇ V ⊗T T. We clearly e e have L[p ] ∩ T α = (L ∩ T α )[p ] when T is a localization of T and when T is a completion of T this follows from the previous Lemma. We deduce that (L ∩ T α ) ⊇ Ie (V ⊗T T ∩ T α ) and hence L ⊇ (L ∩ T α ) ⊗T T ⊇ Ie (V ⊗T T ∩ T α ) ⊗T T. 6 MORDECHAI KATZMAN AND WENLIANG ZHANG But since Ie (V ⊗T T ∩ T α ) ⊗T T satisfies e e (Ie (V ⊗T T∩α ) ⊗T T)[p ] = Ie (V ⊗T T∩T α )[p ] ⊗T T ⊇ (V ⊗T T∩T α )⊗T T ⊇ V ⊗T T we deduce that Ie (V ⊗T T ∩ T α ) ⊗T T is the smallest submodule K ⊆ T α e for which K [p ] ⊇ V ⊗T T. We conclude that Ie (V ⊗T T) equals Ie (V ⊗T T ∩ T α ) ⊗T T. We always have Ie (V ⊗T T) = Ie (V ⊗T T ∩ T α ) ⊗T T ⊇ Ie (V ) ⊗T T. On the other hand e e (Ie (V ) ⊗T T)[p ] = Ie (V )[p ] ⊗T T ⊇ V ⊗T T hence Ie (V ⊗T T) ⊆ Ie (V ) ⊗T T and thus Ie (V ⊗T T) = Ie (V ) ⊗T T.  3. Calculation of supports of FR -finite FR -modules We begin by recalling the following result from [Lyu97, Proposition 2.3]. Remark 3.1 (Vanishing of γt ). Let M be an FR -finite FR -module with a generating homomorphsim γ : M → FR (M ). Let γt denote the composition M → FR (M ) → · · · → FRt (M ). We may assume that M has a presentaA → Rβ → M → 0 and write the generating homomorphism as tion: Rα − U Coker(A) − → Coker(A[p] ) where U is a β × β matrix with entries in R. Then t γt is the composition of Coker(A) → · · · → Coker(A[p ] ). Note that M = 0 if and only if there is a t such that γt = 0. We have γt = 0 ⇔ Im(U [p t−1 ] ⇔ It (Im U [p t ◦ · · · ◦ U [p] ◦ U ) ⊆ Im(A[p ] ) t−1 ] ◦ · · · ◦ U [p] ◦ U ) ⊆ Im A ⇔ I1 (U I1 (· · · I1 (U I1 (Im U ))) ⊆ Im A ⇔ I1 (U I1 (· · · I1 (U I1 (Im U ))) + Im A =0 Im A where we made repeated use of the facts that for any submodule M ⊆ Rβ ℓ we have Iℓ+1 (M ) = I1 (Iℓ (M )) and also Iℓ (U [p ] M ) = U Iℓ (M ). Theorem 3.2. (a) If Ie (Im U [p e−1 ] e ◦ · · · ◦ U [p] ◦ U ) = Ie+1 (Im U [p ] ◦ · · · ◦ U [p] ◦ U ) then (2) Ie (Im U [p e−1 ] ◦ · · · ◦ U [p] ◦ U ) = Ie+j (Im U [p e+j−1 ] for all j ≥ 0. (b) There exists an integer e such that (2) holds. ◦ · · · ◦ U [p] ◦ U ) THE SUPPORT OF LOCAL COHOMOLOGY MODULES 7 e−1 Proof. Write Ve = Ie (Im U [p ] ◦ · · · ◦ U [p] ◦ U ). First we claim that if Ve = Ve+1 then Ve = Ve+j for all j ≥ 0; we proceed by induction on j ≥ 0. Using again the facts that for any submodule W ⊆ Rβ we have Iℓ+1 (M ) = ℓ I1 (Iℓ (M )) and that Iℓ (U [p ] M ) = U Iℓ (M ), we deduce that, if j ≥ 1, then e+j−1 ] ◦ · · · ◦ U [p] ◦ U )) = I (U V Ve+j = I1 (Ie+(j−1) (Im U [p 1 e+j−1 ) and this, by the induction hypothesis, equals I1 (U Ve ) = Ve+1 . Next, we wish to show that for each prime ideal p there exists an integer ep such that (3) Vep Rp = Vep +1 Rp . and that for this ep , Vep +j Rp = Vep Rp for all j ≥ 0. After completing at p, we have assume that our ring is a complete regular local ring, and we let E cp . As this ring is complete denote the injective hull of the residue field of R and regular, there is a natural Frobenius map on E which we denote T , which can be extended to a Frobenius map on direct sums of E by letting T act coordinate-wise; we denote these Frobenius maps also with T . We now consider the Frobenius map Θ = U t T on E β ; in [KZ14, Lemma e−1 3.6] it is shown that annE β Ie (Im U [p ] ◦ · · · ◦ U [p] ◦ U Rp )t ⊆ E β consists of all elements killed by Θe . Now (cf. [HS77, Proposition 1.11] and [Lyu97, e−1 Proposition 4.4]) show that there is an integer ep such that Iep (Im U [p ] ◦ ep ep · · · ◦ U [p] ◦ U Rp ) = Iep +1 (Im U [p ] ◦ · · · ◦ U [p] ◦ U Rp ) and Iep (Im U [p −1] ◦ · · · ◦ ep +j−1 U [p] ◦ U Rp ) = Iep +j (Im U [p  ] ◦ · · · ◦ U [p] ◦ U Rp ) forall j ≥ 0. Crucially, e−1 Lemma 2.5 implies that Ie Im U [p ] ◦ · · · ◦ U [p] ◦ U Rp = Ve Rp for all e ≥ ep and so (3) holds. Consider the following subsets of Spec(R): Pt = {p ∈ Spec(R) | Vt Rp = Vt+1 Rp } = Spec R \ Supp Vt . Vt+1 These form an increasing sequence of open subsets of Spec R, and since for each prime ideal p there is an integer tp such that S Vtp Rp = Vtp +1 Rp , we have t Pt = Spec(R). Now the quasicompactness of Spec R, guarantees the existence of an integer e such that Pe = Spec(R); clearly that e satisfies (2).  e−1 e Corollary 3.3. If Ie (Im U [p ] ◦· · ·◦U [p] ◦U ) = Ie+1 (Im U [p ] ◦· · ·◦U [p] ◦U ), then  Im I (U [pe−1 ] ◦ · · · ◦ U [p] ◦ U ) + Im A  e = SuppR (M). SuppR Im A 4. Our algorithm and its complexity Henceforth in this paper R will denote a polynomial ring over a field K of prime characteristic p. 8 MORDECHAI KATZMAN AND WENLIANG ZHANG Let M be an F -finite F -module with a generating morphism M → FR (M ). Let A be an α×β matrix, which gives a presentation of M , i.e. Coker A ∼ = M. U [p] Let U be a β × β matrix, for which the map Coker A − → Coker A is isomorphic to a generating morphism M → FR (M ). We compute the support of M as follows. Initialize L = Rβ . Compute L′ = I1 (U L). Check whether L′ ⊆ L. If this holds, let L = L′ and go to (2). A Compute a presentation for L+Im Im A as the cokernel of a matrix W with entries in R. (5) Compute the ideal J of maximal minors of W . (6) Output J: this is the defining ideal of the support of M. (1) (2) (3) (4) In the rest of this section we discuss the complexity of this algorithm for computing supports of F -finite F -modules. We start with the following observation, relevant to the complexity of steps (2) and (3) above. Remark 4.1. Let δ be the largest degree of an entry in U and for any j ≥ 0 let δj be the largest degree of a polynomial in a generator of Lj . The calculation of I1 (−) as described in Proposition 2.3 implies that δj+1 ≤ (δj + δ)/p hence δe ≤ 1 1 δ δ0 . + δ( + · · · + e ) ≤ pe p p p−1 The iterative calculation I1 (U L) in step (2) involves a matrix multiplication and a I1 (−) operation, which, in view of Proposition 2.3, amounts to collecting terms in polynomials. The complexity of this step depends on (a) the size β of U , which is an input to the algorithm and does not depend on p, (b) the total number of terms occurring in each of the coordinates of a set of generators of L. In the worst case scenario, if the maximal degree of an entry in Up is Cp, the total number of terms in (b) is bounded by     Cp + n − 1 β = O pβ(n−1) . n−1 In practice, the number of terms is much lower than this worst case. Checking the inclusion in step (3) does not require computing Gröbner bases: the discussion above shows that both L and L′ are generated by vectors whose coordinates are polynomials whose degrees are bounded by D := δ/(p − 1), i.e., L and L′ are given by generators in (R≤D )β where R≤D denotes the K-vector space of polynomials of degrees at most D in R. Thus L′ ⊆ L can be checked by checking whether each given generator of L′ is in the sub-vector space of (R≤D )β spanned by the generators of L. THE SUPPORT OF LOCAL COHOMOLOGY MODULES 9 Computing the presentation in step (4) of the algorithm involves computing the syzygies of the generators of L + Im A; this involves computing a Gröbner bases for L + Im A. In order to assess the practical advantage of our algorithm, we computed the support of 100 F -finite F -modules with randomly generated generating morphism C → FR1 (C) where C is a quotient of R2 and R = Z/2Z[x1 , . . . , x5 ]. We denote t1 the time in seconds required by our algorithm to compute the support and t2 the time in seconds required to compute a root using Grobner bases. The following is a plot of log t2 as a function of log t1 log t2 + + 6 + + + + 4 + ++ + + ++ + + + + + + + + ++ + + + + + + + ++ + ++ + +++++ ++ + + + + + +++ ++ + ++ + + ++ ++ + + + + + 2 + + ++ + −4 −2 ++ + + + + ++ + + ++ + + + + 2 log t1 4 + −2 + ++ + + + + + + + −4 This suggests that for this characteristic and rank, t2 is approximately t21 .3 To further illustrate the effectiveness of our algorithm we compute the following example. Example 4.2. Consider three generic degree-2 polynomials in t: F1 (t) = x0 + x1 t + x2 t2 , F2 (t) = y0 + y1 t + y2 t2 , F1 (t) = z0 + z1 t + z2 t2 . For any two polynomials F (t), G(t) let Res(F, G) denote their Sylvester resultant, e. g. ,   x0 x1 x2 0  0 x0 x1 x2   Res(F1 , F2 ) = det   y0 y1 y2 0  0 y0 y1 y2 Let I denote the ideal generated by Res(F1 , F2 ), Res(F1 , F3 ), Res(F2 , F3 ), Res(F1 + F2 , F3 ) in the polynomial ring R over a field k whose variables are the x,y, and zs above. In [Lyu95] it was asked whether H4I (R) = 0 and this was settled in prime characteristic p > 2 (cf. [Kat97]) and in characteristic zero 3The Macaulay2 code used to produce this data and the data itself is available at [KZ]. 10 MORDECHAI KATZMAN AND WENLIANG ZHANG (cf. [Yan99, Theorem 3].) We used an implementation of our algorithm [KZ] with Macaulay2 ([GS]) to settle the remaining case of characteristic 2: a 20-second run calculated the support of H4I (R) to be empty. 5. The complexity of computing the support of local cohomology We can use the fact that local cohomology modules HiJ (R) are F -finite F -modules and apply the results in the previous sections to compute their supports. To do so we would need to exhibit a generating morphism for these; a standard choice of generating morphism is given by the R-linear map Extj (R/J, R) → Extj (R/J [p] , R) induced by the natural surjection R/J [p] → R/J (cf. [Lyu97, Proposition 1.11]). The calculation of this induced map would normally involve finding a free resolution F• for R/J, extending the quotient map R/J [p] → R/J to a map of free resulutions FR1 (F• ) → F• , applying Hom(−, R) to both resolutions, and finally computing the induced map of cohomologies. Each of these steps involves computing multiple Gröbner bases, and, therefore, potentially unfeasible even in simple instances. For example, even when J is generated by one f ∈ R, the map of f p−1 Ext1 (R/f R, R) → Ext1 (R/f p R, R) above is isomorphic to R/f −−−→ R/f p and even when f is, say, a random polynomial of degree 5 in 5 variables, expanding f p−1 is not feasible beyond the first few primes (e.g., the M2 server Habanero crashes after p = 13.) Clearly, if the calculation of a generating morphism is infeasible, the algorithm of the preceding section cannot be used to calculate the support of the given local cohomology module. However, in those cases when a generating morphism can actually be computed, the method above is guaranteed to produce generating morphisms which yield “bounded” generating morphism working over increasing prime characteristics p, as we describe below. Let S = Z[x1 , . . . , xn ] and let J ⊆ S be an ideal. For any prime p let Jp denote the image of J in Rp = Z/pZ[x1 , . . . , xn ]. An interesting and natural question arising in this context is the description of the properties of the local cohomology module HjJp (Rp ) as p ranges over all primes and we now turn our attention to these. For different choices of prime p the matrices A and U above will be different and this could result in different values of δe which are unbounded as p ranges over all primes. We now show that this is not the case. Let Up denote the square matrix that induces the map Extj (Rp /Jp , Rp ) → [p] Extj (Rp /Jp , Rp ) and δp to denote the maximal degree of entries in Up . We also denote L0,p = Rp and Li+1,p = I1 (Up Li,p ) and use δe,p to denote the largest degree of a polynomial in a generator of Le,p . A A A 1 2 s Theorem 5.1. Let 0 → S bs −→ −→ S → S/J → 0 be a → S b1 − · · · −− free resolution of S/J. Let ∆ denote the maximal degree of any entry in THE SUPPORT OF LOCAL COHOMOLOGY MODULES 11 A1 , . . . , As . Let p be a prime integer which is also a regular element on S/J. Then δe,p ≤ 2j∆ for all integers e ≥ 1. Proof. Since p is a regular element on S/J, tensoring the free resolution of S/J with R/pR produces a free resolution of Rp /Jp . Hence the maximal degree of entries in the maps of this free resolution of Rp /Jp is at most ∆. b b Let θj denote the map Rpj → Rpj in the following commutative diagram [p] induced by Rp /Jp → Rp /Jp 0 / Rbs p O As,p / ··· / Rbs p Aj,p / ··· / Rb1 p A1,p [p] As,p / ··· / Rpbj / Rp O / Rp /Jp O /0 / Rp /Jp[p] /0 = θj θs 0 / Rpbj O [p] Aj,p / ··· / Rb1 p [p] A1,p / Rp An easy induction on j shows that the maximal degree of entries in θj is at [p] most jp∆. The map Up : Extj (Rp /Jp , Rp ) → Extj (Rp /Jp , Rp ) is induced by the transpose of θj and hence the maximal degree in an entry of Up is also bounded by jp∆ (cf. Remark 4.1.) Now jp∆ ≤ 2j∆, δe,p ≤ p−1 for all e ≥ 1.  Corollary 5.2. There is an integer N , independent of p, such that δe,p ≤ N for all e and all p. In particular, there is an integer N ′ , independent of p, such that min{e | Le,p = Le+1,p } ≤ N ′ for all p, i.e. for each prime integer p, the number of steps required to compute the stable value Le,p is bounded by N ′ . Proof. The second statement follows immediately from the first since, once the degree is bounded, the number of steps will be bounded by the number of monomials with the bonded degree. To prove the first statement, it suffices to note there are only finitely many associated prime ideals of S/J in S and hence p is a regular element on S/J for almost all p.  6. Iterated local cohomology modules Let f1 , . . . , fm be a L sequence of elements in R and let N be an R-module. We will write Ki := 1≤j1 <···<ji ≤m Nj1 ···ji to denote the i-th term of the Koszul (co)complex K• (M ; f ) (where each Nj1 ···ji = N ), and we will use H i (N ; f ) to denote the i-th Koszul (co)homology. Proposition 6.1. Let M be an FR -finite FR -module with a generating hoϕ momorphism M − → FR (M ) and let I = (f1 , . . . , fm ) be an ideal of R. Then HIi (M) admits a generating homomorphism H i (M ; f ) → FR (H i (M ; f )). 12 MORDECHAI KATZMAN AND WENLIANG ZHANG Proof. Consider the following commutative diagram: .. .O .. .O FR (φ1 ) FR (φi ) / FR (K 1 ) / ··· O FR (δ1 ) 0 φ1 / K1 0 .. .O FR (φm ) / ··· / FR (K i ) O FR (δi ) / FR (K m ) = FR (M ) O φi δ1 / ··· / Ki /0 φm δi / Km = M / ··· / where the bottom row is the Koszul (co)complex of M on f and φi : M 1≤j1 <···<ji ≤m ⊕1≤j p−1 <···<j ≤m ϕ◦(fj1 ···fji ) i Mj1 ···ji −−−−1−−−−− −−−−−−−−−−−→ FR ( M Mj1 ···ji ). 1≤j1 <···<ji ≤m It follows from [Lyu97, 1.10(c)] that the φi are generating morphisms of Mfj1 ···fji . Therefore taking direct limit of each row of the diagram produces the Čech complex Č(M; f ). Since taking direct limits preserves exactness, lim(H i (M ; f ) → FR (H i (M ; f )) → · · · ) = HIi (M). Our conclusion follows. −→  Combining what we have so far in this section, we now have an algorithm to compute the support of HIi11 · · · HIiss (R). For example, the case s = 2 relevant to the calculation of Lyubeznik numbers in handled as follows. Start with a generating morphism Exti2 (R/I2 , R) → FR (Exti2 (R/I2 , R)). Using Proposition 6.1, we know that the Koszul cohomology H i1 (Exti2 (R/I2 , R); f ) (with I1 = (f )) is a generating homomorphism of HIi11 HIi22 (R). We may then apply Corollary 3.3 to compute the support of HIi11 HIi22 (R). 7. The support of local cohomology of hypersurfaces Throughout this section R denotes a regular ring of prime characteristic p, I ⊆ R an ideal, and g ∈ R some fixed element. Following [Lyu97, §2] we write   3φ 2φ FR FR φ 2 i i i 1 i HI (R) = lim ExtR (R/I, R) − → FR ExtR (R/I, R) −−→ FR ExtR (R/I, R) −−→ . . . → φ where FRe (−) denotes the eth Frobenius functor, and φ : ExtiR (R/I, R) − → i i 1 [p] ∼ FR ExtR (R/I, R) = ExtR (R/I , R) is the R-linear map induced by the A i surjection R/I [p] → R/I. For all i ≥ 0 we fix a presentation Rαi −→ Rβi where Ai is a βi × αi matrix with entries in R. We can now find a φ βi × βi matrix Ui with entries in R for which the map φ : ExtiR (R/I, R) − → FR1 ExtiR (R/I, R) is isomorphic to the map Ui : Coker Ai → FR1 (Coker Ai ) = [p] Coker Ai given by multiplication by Ui . THE SUPPORT OF LOCAL COHOMOLOGY MODULES 13 Theorem 7.1. For any i ≥ 0 consider the map g : HiI (RP ) → HiI (RP ) given by multiplication by g. Let Ii denote the set of primes P ⊂ R for which the map g is not injective and let Si denote the set of primes P ⊂ R for which the map g is not surjective. For ℓ, e, j ≥ 0 write [pe+j−1 ] (ℓ) Vej = Uℓ [pe+j−2 ] Uℓ [pe ] · · · Uℓ . Then (i) (a) Ii is closed and equal to Supp (ker V0η :Rβ g) (i) ker V0η (b) Si is closed and equal to Supp S j≥0  for some η > 0, Rβ (i) j gRβ + Im A[p ] :Rβ V0j (c) the support of HiI (R/gR) is closed and equal to Ii ∪ Si .  , and Proof. Fix some i ≥ 0 and write β, A and U for βi , Ai and Ui . The map g : HiI (R) → HiI (R) can be described as a map of direct limit systems (4) Coker A U / Coker A[p]U g  Coker A [p] e−1 ] U [p / / ... Coker A[p g U  [pe ] eU ] / ... . g U [p] / Coker A[p] e−1 ] U [p  / Coker A[peU] / ... [pe ] / ... (i) For any e, j ≥ 0 abbreviate Vej = Vej , and note that it is the matrix e e+j corresponding to the composition map Coker A[p ] → Coker A[p ] in the direct limits in (4). Any element in HiI (RP ) can be represented by an element [pe ] a ∈ Coker AP for some e ≥ 0, and this element represents the zero element [pe+j ] if and only if there exists a j ≥ e for which Vej a ∈ Im AP only if a ∈ (Im A[p e+j ] , i.e., if and :Rβ Vej )P . j Consider the kernels Kj of the maps V0j : Coker A → Coker A[p ] ; these form an ascending chain of submodules of Coker A and hence stabilize for all j e e+j beyond some η ≥ 0. Note that the map Vej : Coker A[p ] → Coker A[p ] is obtained by applying the exact functor FRe (−) to the map V0j , hence the kernels of the maps Vej also stabilize for j ≥ η. To prove (a) we now note that an element in HiI (RP ) represented a ∈ [pe ] Coker AP is multiplied by g to zero  a ∈ (ker Veη :Rβ g)P  if and only if (ker Veη :Rβ g) = 0, i.e., if g is not and so g is injective if and only iff ker Veη P
a zero divisor on Rβ / ker Veη P . But Rβ / ker Veη = FRe (Rβ / ker V0η ) and, since R is regular, FRe (Rβ / ker V0η ) and Rβ / ker V0η have the same associated primes, so we deduce that multiplication by g is injective if and only if g is
not a zero divisor on Rβ / ker V0η P . We deduce that for a prime P ⊂ R, 14 MORDECHAI KATZMAN AND WENLIANG ZHANG multiplication by g on HiI (RP ) is injective if and only if   (ker V0η :Rβ g) =0 ker V0η P so Ii = Supp (ker V0η :Rβ g) . ker V0η To prove (b) we now note that an element in HiI (RP ) represented a ∈ [pe ] Coker AP is in the image  of g if and only if there exists a j ≥ 0 such that [pe+j ] β hence g is surjective if for all e ≥ 0, Vej a ∈ gR + Im AP  [ e+j gRβ + Im A[p ] :Rβ Vej = Rβ P . P j≥0 Furthermore,  [p]   e+1+j ] e+j = gp Rβ + Im A[p :Rβ Ve+1,j gRβ + Im A[p ] :Rβ Vej   e+1+j ] ⊆ gRβ + Im A[p :Rβ Ve+1,j so for for all e ≥ 0, [ gRβ + Im A[p e+j ] :Rβ Vej j≥0 if and only if [ j≥0 j gRβ + Im A[p ] :Rβ V0j   P P = Rβ P = Rβ P . We conclude that g is not surjective if and only if P ∈ Supp Rβ / To prove (c) consider the long exact sequence S j≥0   j gRβ + Im A[p ] :Rβ V0j . g g → Hi+1 · · · → HiI (R) − → HiI (R) → HiI (R/gR) → Hi+1 I (R) → . . . I (R) − g induced by the short exact sequence 0 → R − → R → R/gR  → 0. Note that g i i i HI (R/gR)P = 0 if and only if both HI (R) − → HI (R) is surjective and P   g → Hi+1 Hi+1 and the result follows. I (R) I (R) − P  Question 7.2. Theorem 3.2 gives us an effective method for the calculation of Ii . However, we do not know how to compute Si , hence we ask the following: is there an effective method to bound the value of e for which    [ e j (i) (i) gRβ + Im A[p ] :Rβ V0j = gRβ + Im A[p ] :Rβ V0e ? j≥0 It turns out that part of our Theorem 7.1 can be extended to the case of isolated singular points. THE SUPPORT OF LOCAL COHOMOLOGY MODULES 15 Corollary 7.3. Let R be a noetherian commutative ring of prime characteristic that has finitely many isolated singular points. Let g ∈ R be a nonzerodivisor. Then Supp(HIj (R/gR)) is Zariski-closed for each integer j and ideal I of R. Proof. Let T {m1 , . . . , mt } denotes the set of isolated singular points of R. Set a = ti=1 mi . Let {f1 , . . . , fs } be a set of generators of a. It follows from Theorem 7.1 that SuppRf (HjI (Rfk /gRfk )) is closed, i.e. it has finitely k many minimal associated primes. By the bijection between the set of associated primes of HjI (R/gR) that do not contain fk and Ass(HjI (Rfk /gRfk )), it follows that the minimal associated primes of HjI (R/gR) are contained in the union of {m1 , . . . , mt } and the set of minimal associated primes of  HjI (Rfk /gRfk ) which is a finite set. The proof of Corollary 7.3 can also be used to prove the following result which is of independent interest. Proposition 7.4. Let R be either (1) a noetherian commutative ring of prime characteristic, or (2) of finite type over a field of characteristic 0. Suppose that R has finitely many isolated singular points. Then HjI (R) has only finitely many associated primes for each integer j and each ideal I of R. Proof. Let {m1 , . . . , mt } denotes the set of isolated singular points of R. Set T a = ti=1 mi . Let {f1 , . . . , fs } be a set of generators of a. It follows from our assumptions on R that Rfk is either a noetherian regular ring of prime characteristic or a regular ring of finite type over a field of characteristic 0 (cf. [Lyu93, Corollary 3.6]). Consequently, Ass(HjI (Rfk )) is finite for each generator fk . Since there is a bijection between the set of associated primes of HjI (R) that do not contain fk and Ass(HjI (Rfk )), it follows that Ass(HjI (R)) ⊆ s [ Ass(HjI (Rfk )) k=1 [ {m1 , . . . , mt }.  Acknowledgement We would like to thank the anonymous referee whose suggestions and criticisms improved this paper substantially. References [BMS08] M. Blickle, M. Mustaţă, and K. Smith: Discreteness and rationality of F-thresholds, Michigan Math. J. 57 (2008), 43–61. [GS] D. R. Grayson and M. E. Stillman: Macaulay2, a software system for research in algebraic geometry. 16 [Har67] MORDECHAI KATZMAN AND WENLIANG ZHANG R. Hartshorne: Local cohomology, A seminar given by A. Grothendieck, Harvard University, Fall, vol. 1961, Springer-Verlag, Berlin, 1967. MR0224620 (37 #219) [Har68] R. Hartshorne: Cohomological dimension of algebraic varieties, Ann. of Math. (2) 88 (1968), 403–450. 0232780 (38 #1103) [HS77] R. Hartshorne and R. Speiser: Local cohomological dimension in characteristic p, Ann. of Math. (2) 105 (1977), no. 1, 45–79. MR0441962 (56 #353) [HNB] M. Hochster and L. Núñez-Betancourt: Support of local cohomology modules over hypersurfaces and rings with FFRT, preliminary preprint. [Kat97] M. Katzman: The cohomological dimension of a resultant variety in prime characteristic p, J. Algebra 191 (1997), no. 2, 510–517. 1448806 (98a:13023) [Kat08] M. Katzman: Parameter-test-ideals of Cohen-Macaulay rings, Compos. Math. 144 (2008), no. 4, 933–948. MR2441251 (2009d:13030) [KZ] M. Katzman and W. Zhang: A macaulay2 library for computing supports of F-finite F-modules, Freely available from http://www.katzman.staff.shef.ac.uk/Fsupport/. [KZ14] M. Katzman and W. Zhang: Annihilators of Artinian modules compatible with a Frobenius map, J. Symbolic Comput. 60 (2014), 29–46. 3131377 [Lyu93] G. Lyubeznik: Finiteness properties of local cohomology modules (an application of D-modules to commutative algebra), Invent. Math. 113 (1993), no. 1, 41–55. 1223223 (94e:13032) [Lyu95] G. Lyubeznik: Minimal resultant systems, J. Algebra 177 (1995), no. 2, 612– 616. 1355218 (96i:13013) [Lyu97] G. Lyubeznik: F -modules: applications to local cohomology and D-modules in characteristic p > 0, J. Reine Angew. Math. 491 (1997), 65–130. MR1476089 (99c:13005) [LS01] G. Lyubeznik and K. E. Smith: On the commutation of the test ideal with localization and completion, Trans. Amer. Math. Soc. 353 (2001), no. 8, 3149– 3180 (electronic). MR1828602 (2002f:13010) [Mur13] S. Murru: On the upper semi-continuity of HSL numbers, arXiv:1302.1124. [Ogu73] A. Ogus: Local cohomological dimension of algebraic varieties, Ann. of Math. (2) 98 (1973), 327–365. 0506248 (58 #22059) [PS73] C. Peskine and L. Szpiro: Dimension projective finie et cohomologie locale. Applications à la démonstration de conjectures de M. Auslander, H. Bass et A. Grothendieck, Inst. Hautes Études Sci. Publ. Math. (1973), no. 42, 47–119. MR0374130 (51 #10330) [Yan99] Z. Yan: Minimal resultant systems, J. Algebra 216 (1999), no. 1, 105–123. 1694582 (2000d:14022) Department of Pure Mathematics, University of Sheffield, Hicks Building, Sheffield S3 7RH, United Kingdom E-mail address: M.Katzman@sheffield.ac.uk Department of Mathematics, Statistics, and Computer Science, University of Illinois at Chicago, 851 S. Morgan Street, Chicago, IL 60607-7045 E-mail address: wlzhang@uic.edu 

0math.AC

Picasso, Matisse, or a Fake? Automated Analysis of Drawings at the Stroke Level for Attribution and Authentication∗ arXiv:1711.03536v1 [eess.IV] 8 Nov 2017 Ahmed Elgammal1,2 †, Yan Kang1 , Milko Den Leeuw3 1 Artrendex LLC., NJ, USA‡ 2 Department of Computer Science, Rutgers University, NJ, USA 3 Atelier for Restoration & Research of Paintings (ARRS), The Hague, The Netherlands November 13, 2017 Abstract This paper proposes a computational approach for analysis of strokes in line drawings by artists. We aim at developing an AI methodology that facilitates attribution of drawings of unknown authors in a way that is not easy to be deceived by forged art. The methodology used is based on quantifying the characteristics of individual strokes in drawings. We propose a novel algorithm for segmenting individual strokes. We designed and compared different handcrafted and learned features for the task of quantifying stroke characteristics. We also propose and compare different classification methods at the drawing level. We experimented with a dataset of 300 digitized drawings with over 80 thousands strokes. The collection mainly consisted of drawings of Pablo Picasso, Henry Matisse, and Egon Schiele, besides a small number of representative works of other artists. The experiments shows that the proposed methodology can classify individual strokes with accuracy 70%-90%, and aggregate over drawings with accuracy above 80%, while being robust to be deceived by fakes (with accuracy 100% for detecting fakes in most settings). 1 Introduction Attribution of art works is a very essential task for art experts. Traditionally, stylistic analysis by expert human eye has been a main way to judge the authenticity of artworks. This has been pioneered and made a methodology by Giovanni Morelli (1816-1891) who was a physician and ∗ This paper is an extended version of a paper that will be published on the 32nd AAAI conference on Artificial Intelligence, to be held in New Orleans, USA, February 2-7, 2018 † Corresponding author: Ahmed Elgammal elgammal@artrendex.com ‡ http://www.artrendex.com 1 art collector, in what is known as Morellian analysis. This connoisseurship methodology relies on finding consistent detailed “invariant” stylistic characteristics in the artist’s work that stay away from composition and subject matter. For example Morelli paid great attention to how certain body parts, such as ears and hands are depicted in paintings by different artists, not surprisingly given his medical background. This methodology relies mainly on the human eye and expert knowledge. The work of van Dantzig [1] that we follow in this paper belongs to this methodology. In contrast, technical analysis focuses on analyzing the surface of the painting, the underpainting, and/or the canvas material. There is a wide spectrum of imaging (e.g. infrared spectroscopy and x-ray), chemical analysis (e.g. Chromatography), and radiometric (e.g. carbon dating) techniques that have been developed for this purpose. Mostly, this analysis aims to get insights on the composition of the materials and pigments used in making the different layers of the work and how that relates to what materials, were available at the time of the original artist or what the artist typically used. These techniques are complementary and each of them has limitations to the scope of their applicability. We refer the reader to [2] for comprehensive surveys of these techniques. Analysis using computer vision and image processing techniques has been very sparsely and cautiously investigated in the domains of attribution and forgery detection (e.g. [3, 4, 5, 6]). Image processing has been used as a tool in conjunction with non-visual spectrum imaging, such as analysis of x-ray imaging to determine canvas material and thread count (e.g. [4, 7]). The question we address in this paper, is what role can the computer vision and AI technology plays in this domain given the spectrum of the other available technical analysis techniques, which might seem more conclusive. We argue that developing this technology would complement other technical analysis techniques for three reasons. First, computer vision can uniquely provide a quantifiable scientific way to approach the traditional stylistic analysis, even at the visual spectrum level. Second, it would provide alternative tools for the analysis of art works that lie out of the scope of applicability for the other techniques. For example, this can be very useful for detecting forgery of modern and contemporary art where the forger would have access to pigments and materials similar to what original artist had used). Third, computer vision has the potential to provide a cost-effective solution compared to the cost of other technical analysis methods. For example, in particular related to the topic of this paper, there are large volumes of drawings, prints, and sketches for sale and are relatively cheap (in the order of a few thousand dollars, or even few hundreds) compared to paintings. Performing sophisticated technical analysis in a laboratory would be more expensive than the price of the work itself. This prohibitive cost makes it attractive for forgers to extensively target this market. It is worthy to mention that several papers have addressed art style classification, where style is an art movement (e.g. Impressionism), or the style of a particular artist (e.g. the style of Van Gogh) [8, 9, 10, 11, 12]. Such stylistic analysis does not target authentication. Such works use global features that mainly capture the composition of the painting. In fact, such algorithm will classify a painting done on the style of Van Gogh, for example, as Van Gogh, since it is designed to do so. Methodology The methodology used in this paper is based on quantifying the characteristics of individual strokes in drawings and comparing these characteristics to a large number of strokes by different artists using statistical inference and machine learning techniques. This process is inspired by the Pictology 2 Figure 1: Illustration of van Dantzig methodology on simple strokes. Spontaneous strokes differ in their shape and tone at their beginning, middle and end. Figure from [1] methodology developed by Maurits Michel van Dantzig [1] (1903 - 1960). Van Dantzig suggested several characteristics to distinguish the strokes of an artist, and suggested that such characteristics capture the spontaneity of how original art is being created, in contrast to the inhibitory nature of imitated art. Among the characteristics suggested by van Dantzig to distinguish the strokes of an artist are the shape, tone, and relative length of the beginning, middle and end of each stroke. The characteristics include also the length of the stroke relative to the depiction, direction, pressure, and several others. The list of characteristics suggested by van Danzig is comprehensive and includes, in some cases, over one hundred aspects that are designed for inspection by the human eye. The main motivation is to characterize spontaneous strokes characterizing a certain artist from inhibited strokes, which are copied from original strokes to imitate the artist style. In this paper we do not plan to implement the exact list of characteristics suggested by van Dantzig; instead we developed methods for quantification of strokes that are inspired by his methodology, trying to capture the same concepts in a way that is suitable to be quantified by the machine, is relevant to the digital domain, and facilitates statistical analysis of a large number of strokes by the machine rather than by human eye. We excluded using comparisons based on compositional and subject-matter-related patterns and elements. Most forged art works are based on copying certain compositional and subjectmatter-related elements and patterns. Using such elements might obviously and mistakenly connect a test subject work to figures and composition in an artist known works. In contrast to subject matter and compositional elements, the characteristics of individual strokes carry the artist’s unintentional signature, which is hard to imitate or forge, even if the forger intends to do. 3 Contribution In this paper we propose a computational approach for analysis of strokes in line drawings that is inspired and follow the principles of Pictology, as suggested by van Dantzig. We propose and validate a novel algorithm for segmenting individual strokes. We designed and compared different hand-crafted and learned deep neural network features for the task of quantifying stroke characteristics. We also propose and compare different classification methods at the drawing level. We experimented with a dataset of 300 digitized drawings with over 70 thousands strokes. The collection mainly consisted of drawings of Pablo Picasso, Henry Matisse, and Egon Schiele, besides a small number of representative works of other artists. We extensively experimented on different settings of attributions to validate the proposed methodology. We also experimented with forged art works to validate the robustness of the proposed methodology and its potentials in authentication. Table 1: Dataset collection: technique distribution Technique Pen/brush (ink) Etching Pencil Drypoint Lithograph Crayon Charcoal Unknown Total Picasso 80 38 8 2 2 0 0 0 130 Matisse 45 10 5 2 14 1 0 0 77 Schiele 0 0 10 0 0 5 4 17 36 Modigliani 0 0 9 0 0 8 1 0 18 Others 20 0 0 0 9 4 1 2 36 Total 145 48 32 4 25 18 6 19 297 Strokes 36,533 19,645 9,300 914 6,180 4,648 666 2,204 80,090 Others: Georges Braque, Antoine Bourdelle, Massimo Campigli, Marc Chagall, Marcel Gimond, Alexej Jawlensky, Henri Laurens, Andre Marchand, Albert Marquet, Andr Masson, Andre Dunoyer Dr Segonzac, Louis Toughague 2 2.1 Detailed Methodology Challenges The variability in drawing technique, paper type, size, digitization technology, spatial resolution, impose various challenges in developing techniques to quantify the characteristic of strokes that are invariant to these variability. Here we highlight some these challenges and how we addressed them. Drawings are made using different techniques, materials and tools, including, but not limited to drawings using pencil, pen and ink, brush and ink, crayon, charcoal, chalk, and graphite drawings. Different printing techniques also are used such as etching, lithograph, linocuts, wood cuts, and others. Each of these techniques results in different stroke characteristics. This suggests developing technique-specific models of strokes. However, typically each artist prefers certain techniques over others, which introduce unbalance in the data collection, which need to be addressed. Therefore, in this paper we are testing two hypotheses: technique specific vs. across technique comparisons, to test if we can capture invariant stroke characteristic for each artist that persists across techniques. Drawings are executed on different types of papers, which, along with differences in digitization, imply variations in the tone and color of the background. This introduces a bias in the data. We want to make sure that we identify artists based on their strokes and not based on the color 4 tone of the paper used. Different types of papers along with the type of ink used result in different diffusion of ink at the boundaries of the strokes which, combined with digitization effects, alter the shape of the boundary of the stroke. Drawings are made on different-sized papers, and digitized using different resolutions. The size of the original drawing as well as the digitization resolution are necessary to quantify characteristics related to the width or length of strokes. Therefore, in this paper we quantify the characteristics of the strokes in a metric basis after converting all the measurements to the metric system. 2.2 Data collection A collection of 297 drawings were gathered from different sources to train, optimize, validate, and test the various classification methodologies used in this study. The drawings selected are restricted to line drawings, i.e, it excludes drawings that have heavy shading, hatching and watercolored strokes. The collection included drawings and prints by Picasso (130), Henry Matisse (77), Egon Schiele (36), Amedeo Modigliani (18), and a small representative works of other artists (36), ranging from 1910-1950AD. These artists were chosen since they were prolific in producing line drawings during the first half of the Twentieth century. The collection included a variety of techniques including: pen and ink, pencil, crayon, and graphite drawings as well as etching and lithograph prints. Table 1 shows the number of drawings for each artist and technique. In the domain of drawing analysis it is very hard to obtain a dataset that is uniformly sampling artists and techniques. The collection is biased towards ink drawings, executed mostly with pen, or using brush in a few cases. There is a total of 145 ink drawings in the collection. The collection contains more works by Picasso than other artists. In all the validation and test experiments an equal number of strokes were sampled from each artist to eliminate data bias. The collection included digitized works from books, downloaded digitized images from different sources, and screen captured images for cases where downloading was not permitted. The resolution of the collected images varies depending on the sources. The effective resolution varies from 10 to 173 pixel per cm depending on the actual drawing size and the digitized image resolution. Figure 2 shows the distribution of the digitized images resolution. Given this wide range of resolutions, the algorithms and features used were designed to be invariant to the digitization resolution. Fake drawing dataset: In order to validate the robustness of the proposed approaches against being deceived by forged art, we commissioned five artists to make drawings similar to those of Picasso (24), Matisse (39) and Schiele (20) using the same techniques. We collected a total of 83 drawings (24, 39, 20). None of these fake drawings was used in training the models. We only used them for testing. Figure 3 shows examples of the fake dataset mixed up with real drawings. Because we do not expect the reader to be experts in authentication in art, to be able to judge the quality of the fake drawings in isolation, we deliberately mixed up a collection of the fake drawings with real drawings in Figure 3. If the reader is interested to know which of these images are of fake or real drawings, please refer to the end of the paper. 5 Figure 2: Distribution of digitization resolution (in Pixel per cm units) 3 Stroke Segmentation The stroke segmentation algorithm will be described here. See sample results in Figure 4 and 5. A typical isolated stroke is a line or curve, with a starting point and endpoint. A stroke can have zero endpoints (closed curve) or 1 endpoint, which are special cases that do not need further segmentation. However, strokes typically intersect to form a network of tangled strokes that needs to be untangled. A network of strokes is characterized by having more than 2 endpoints. Since strokes are thin elongated structures; a skeleton representation would preserve their topological structure even in a network configuration [13]. Therefore, the segmentation of strokes is done on such a skeleton representation. There is a large classical literature in computer vision on detecting junctions on edge maps as a way to characterize object boundaries, infer about three-dimensional structure and form representations for recognition. Unlike classical literature which look at natural images, in our case detecting junctions and endpoints is fortunately relatively easy since they persists in a skeleton representation of the network of strokes. On the other hand, the challenge in our case is to use the information of such junctions and endpoints to segment individual strokes. In our case, junctions play crucial role in identifying the intersections between strokes. There are two basic ways strokes intersect: occluder-occluded configuration to form a T-junction or two strokes crossing each other to form an X-junction. A T-junction is a continuation point of the occluding stroke and an endpoint for the occluded stroke. We need to preserve the continuation of the occluding stroke at the T-junction. The stroke segmentation algorithm takes a network of strokes and identifies one occluding stroke at a time and remove it from the network of strokes to form a residual network(s) that is recursively segmented. This is achieved by constructing a fully connected graph whose vertices 6 Figure 3: Examples of images of the fake dataset mixed up with real images of drawings by Matisse, Picasso, and Schiele. See the key at the end of the document to tell which are real and which are fake! 7 are the endpoints in the network and edges are weighted by the cost of reaching between each two endpoints. The cost between two endpoints reflects the bending energy required at the junctions. Let the endpoints in a network of strokes denoted by e1 , · · · , em and let the junction locations denoted by j1 , · · · , jn . The cost of the path between any two end points ei to ej is cumulative curvature along the skeleton path between them, where the curvature is only counted close to junctions. The rational is that it does not matter how much bending a stroke would take as long as it is not at junctions. Let γ(t) : [0 : 1] → R2 be the parametric representation of the skeleton curve connecting ei and ej . The cost is defined as c(ei , ej ) = Z 1 κ(t) · φ(γ(t))dt 0 where κ(·) is the curvature and φ(·) is a junction potential function, which is a function of the proximity to junction locations defined as n 1X 2 φ(x) = e(x−ji ) /σ n i=1 After the graph construction, the minimum cost edge represents a path between two endpoints with minimum bending at the junctions, which corresponding to an occluding stroke. In case of a tie, the path with the longest length is chosen. The optimal stroke is removed from the skeleton representation and from the graph. This involves reconnecting the skeleton at X-junctions (to allow the detection of the crossing strokes) and new endpoints have to be added at T-junctions (to allow the detection of occluded strokes. Removing a stroke from the graph involves removing all edges corresponding to paths that go through the removed stroke. This results in breaking the graph to one or more residual subgraphs, which are processed recursively. 4 4.1 Stroke Analysis Methodology Quantifying Stroke Characteristics This section explains the process of quantifying the characteristics of individual strokes and the extracted features used to represent each stroke. The goal is to construct a joint feature space that captures the correlation between the shape of the stroke, its thickness variation, tone variation, local curvature variation. For this purpose we studied two different types of features and their combination: 1) Hand-crafted features capturing the shape of each stroke and its boundary statistics, 2) Learned-representation features capturing the tone variation as well as local shape characteristics. The next two subsection describe each of these features. 4.1.1 Hand-crafted Features In our study, each stroke is represented by its skeleton, its boundary, and the rib length around the skeleton. The following descriptors are extracted to quantify the characteristics of each stroke. All the descriptors are designed to be invariant to translation, rotation, scaling, and change in digitization resolution. 8 Figure 4: Examples of segmentation results. Top: Picasso lithograph. Bottom: Picasso ink drawing. - best seen in color 9 Figure 5: Examples of segmentation results. Top: Matisse etching. Bottom: Schiele ink drawing. - best seen in color 10 Shape of the boundary: The shape of the stroke boundary is quantified by Fourier descriptors [14]. Fourier descriptors are widely used shape features for a variety of computer vision applications such as character recognition and shape matching. Fourier descriptors provide shape features that are proven to be invariant to translation, scaling, rotation, sampling, and contour starting points [14]. We used 40 amplitude coefficients (first 20 harmonics in each direction) to represent the shape of the boundary of the stroke. Reconstruction error profile: The mean reconstruction error, as a function of the number of harmonics used to approximate the shape of the strokes, is used as a descriptor of the smoothness of the contour and the negative space associated with the stroke. In particular, we compute the mean reconstruction error at each step while incrementally adding more harmonics to approximate the shape of the stroke. The reconstruction error profile is normalized by dividing by the stroke mean width in pixels to obtain a descriptor invariant to digitization resolution. Contour Curvature descriptor: To quantify the curvature of the stroke contours, we use the first and second derivatives of the angular contour representation. The distributions of these derivatives are represented by their histograms. Stroke thickness profile: To quantify the thickness of the stroke, we compute the mean and standard deviation of the rib length around the skeleton of the stroke, as well as a histogram of the rib length. All rib length measurements are mapped to mm units to avoid variations in digitization resolution. Stroke Length: The length of the stroke is quantified as the ratio between the stroke skeleton length to the canvas diagonal length. This measure is invariant to digitization resolution. 4.1.2 Deep Learned Features using RNNs GRU Classification with Truncated Back Propagation Through Time: Other than the traditional feed-forward neural networks specialized at fixed size input, e.g. images, recurrent neural network (RNN) could handle variable length sequence input x = (x1 , · · · , xT ) and either fixed length output or variable length output y = (y1 , · · · , yT ) by utilizing the hidden state within. RNN sequentially takes input xt from the input sequence and update its hidden state by: ht = φθ (ht−1 , xt ) φθ is a nonlinear activation function with θ as parameters. In each time step, a corresponding output could be generated through: ŷt = gθ (ht , xt ) gθ is an arbitrary parametric function that is trained together with the recurrent cell. Recently, it has been widely shown that the more complicated RNN model such as Long ShortTerm Memory (LSTM, [15]) or Gated Recurrent Unit (GRU, [16]) would eliminate the problem of vanishing gradient [17], [18]. LSTM and GRU introduce some gating units that can automatically determine how much the information flow could be used in each time step, by which the vanishing gradient could be avoided. In GRU, two gate units are introduced: reset gate and update gate. Reset gate controls how much former hidden state would be used in calculating the current time step candidate ĥt . rt = σ(Ur ht−1 + Wr xt ) 11 Update gate controls how much current candidate ĥt would be updated and how much old hidden state would be kept. zt = σ(Uz ht−1 + Wz xt ) Then the candidate hidden state ĥt would be: ĥt = tanh(U (rt ht−1 ) + W xt ) And the final hidden state ht at time t is updated though: ht = zt ĥt + (1 − zt ) ht−1 Where U, W, Ur, Wr, Uu, Wu are the learned weight matrices and the biases are omitted here to make the equations more compact. is a pointwise multiplication, and σ is a sigmoid function. Given a stroke, a sequence of patches of fixed size are collected along the skeleton of the stroke and fed to a GRU model as inputs. We tested both fixed size patches or adaptive size patches where the radius of the patch is a function of the average stroke width in the drawing. In both cases the input patches are scaled to 11x11 input matrices. To achieve invariant to the direction of the stroke, each stroke is sampled in both directions as two separate data sequences (at classification, both a stroke and its reverse either appear in training or testing splits ). We normalized the grey scale into range (-1, 1), and flattened the 11 × 11 image into a 121-dimension vector. The activation function we used in experiments is tanh function. Parameters are initialized from normal distribution with mean = 0, standard deviation = 1. After comparing several optimizer functions, we found that the RMSProp optimizer with learning rate 0.001 outperforms others. The gradient is globally clipped to be less than 5 to prevent from gradient exploding,. And to avoid gradient vanishing, we calculated the gradient by the truncated Back Propagation Through Time. Each sequence is unrolled into a fixed size τ steps (τ = 30 in the experiments) at each time to calculate the gradient and to update the network’s parameters. The label of original sequence is assigned to each unrolling. Between each unrolling, the hidden state is passed on to carry former time steps information. And within each unrolling, only the last time step hidden state is used in the final linear transformation and Softmax function to get the predicted score of each class. The loss function we used was the cross entropy. xτ = (xt , ...xt+τ ) ht+τ = GRU (hinitial , xτ ) ŷ = sof tmax(Us ht+τ ) loss = − X ylog(ŷ) hinitial = ht+τ 4.2 Stroke Classification For the case of hand-crafted features, strokes are classified using a support vector machine (SVM) classifier [19]. We evaluated SVM using Radial basis kernels as well as polynomial kernels. The classifier produces posterior distribution over the classes. For the case of learned GRU features, the classification of strokes is directly given by the trained networks. SVM was used to combine hand-crafted features with the learned features in one classification framework. In such case, the activation of the hidden units were used as features, and combined to the hand-crafted features. 12 4.3 Drawing classification: A given drawing is classified by aggregating the outcomes of the classification of its strokes. We used four different strategies for aggregating the stroke classification results, as described below. • Majority Voting: In this strategy each stroke votes for one class. All strokes have equal votes regardless of the certainty of the output of the stroke classifier. • Posterior aggregate: In this strategy each stroke votes with a weight equal to its posterior class probability. This results in reducing the effect of strokes that are not classified with high certainty by the stroke classifier. • k-certain voting: In this strategy, only the strokes with class posterior greater than a threshold k are allowed to vote. This eliminates effect of uncertain strokes. • certainty weighted voting: In this strategy each stroke vote is weighted using a gamma function based on the certainty of the stroke classifier in classifying it. 5 Validation Experiments This section describes the experiments conducted to test and validate the performance of the proposed stroke segmentation, stroke classification, and the drawing classification approaches on the collected dataset. In particular, the experiments are designed to test the ability of the algorithms to determine the attribution of a given art work and test its robustness to forged art. 5.1 Segmentation Validation Validating the segmentation algorithm is quite challenging since there is no available ground truth segmentation and because of the difficulty of collecting such annotation. It is quit a tedious process for a human to trace individual strokes to provide segmentation of them, specially such task requires certain level of expertise. To validate the segmentation algorithm, we collected 14 drawings with medium difficulty (in terms of number of strokes) from the collection and showed the segmentation results to two artists and asked them independently to locate errors in the segmentations. Figure 6 shows an example of a drawing with its two annotations of the results. A closer look highlights that annotators make several mistakes (false positive, and false negatives). Table 2 shows the number of marked errors for each sample image by two evaluators. The overall error per annotator is computed as: Error rate = total marked errors at junctions / total number of strokes; where the total is aggregated over all evaluated images. The average error rate over the two annotators is 12.94% , counting all labeled errors by annotators. The annotation shows large deviations between the two annotators, with mean deviation 24.93 and standard deviation 12%. This highlight the challenge in validating the segmentation results by human annotation. However, most of the marked errors are at small detailed strokes that are hard to segment, even by the human eye, and does not contribute much to the classification of strokes since small strokes are filtered out anyway. 13 Figure 6: Example of two drawings by Picasso, stroke segmentation results, and segmentation errors marked by two artists. 5.2 Stroke Classification Evaluation Methodology: In all experiments the image datasets were split into five 80/20% folds to perform five-fold cross validation. Since strokes from the same drawings might share similar characteristics, we did this splits at the image level and not at the stroke level. For each fold, after splitting the images to train and test sets, equal number of strokes were sampled for each artist class for training and testing to avoid the bias in the data, which is significant in our case. We evaluated different classification settings including pair-wise classification, and one-vs-all classification, and multi-class classification. Extensive ablation studies are also performed to evaluate the different features and their effects, as well as to choose the optimal settings. 14 Table 2: Validation of Stroke Segmentation Sample Number of Strokes 1 596 2 366 3 314 4 216 5 267 6 122 7 136 8 131 9 102 10 71 11 159 12 123 13 103 14 196 Total 2902 Mean Std 5.2.1 Evaluator 1 Marked Errors 34 37 38 24 69 40 28 32 22 17 48 30 25 65 509 Evaluator 2 Marked Errors 75 17 11 11 13 14 10 12 10 6 15 8 10 30 242 Absolute deviation between Evaluators 41 20 27 13 56 26 18 20 12 11 33 22 15 35 24.93 12.72 Stroke Classification Validation - Technique Specific - Pairwise: For testing technique-specific classifiers, we trained pairwise classifiers to discriminate between Picasso and Matisse drawings made using either pen/ink or etching. We chose these two techniques and these two artists since they have the largest representation in our collection. Table 3 shows the stroke classification results. The experiments is done using five-fold cross validation and the mean and standard deviations are reported. The table shows a comparison between the different types of proposed features. Table 3: Validation of Stroke Classifier - Technique specific (Picasso vs Matisse) Ink Drawing (Pen/Brush) (Picasso vs Matisse) Approach Train Test Hand-Crafted - SVM-RBF 87.99% (0.39%) 79.16% (0.26%) Hand-Crafted - SVM-POLY 79.88% (0.14%) 77.17% (0.58%) GRU 84.92 % ( 1.89%) 65.86 (13.58 % ) Etching Prints (Picasso vs Matisse) Approach Train Test Hand-Crafted - SVM-RBF 94.53% (0.22%) 84.18% ( 0.85 %) Hand-Crafted - SVM-POLY 94.27% (0.21%) 93.09% ( 0.88%) GRU 83.74% (4.60 % ) 75.08% (8.11%) 5.2.2 Stroke Classification Validation - One-vs-all: In this experiment a one-vs-all classification settings is used to build classifiers for Picasso-vsNon-Picasso, Matisse-vs-Non-Matisse, Schiele-vs-Non-Schiele. These three artists are chosen 15 since they have enough data for training and testing the classifiers in a five-fold split setting. The classifiers are then evaluated on the fake dataset (see Section 5.3 ). We evaluated the performance of two settings: 1. across-techniques: we evaluated the performance of the stroke classifiers on all techniques combined to evaluate whether the classifier can capture an invariant for the artist regardless of the technique used. 2. Technique-specific: in this setting each classifier is trained and tested using strokes from the same drawing technique. Given the data collection, we tested a) Picasso-vs-Non-Picasso classifier using ink/pen, b) Matisse-vs-Non-Matisse classifier using ink/pen, c) Schiele-vsNon-Schiele using pencil. Table 4 shows the mean and standard deviations of the five folds for the hand-crafter features, the GRU features and the combination. Both types of features have very good stroke classification performance. GRU has better performance over the three artists tested. Combining the features further improved the results and reduced the cross-fold variances, which indicate that both types of features are complementary to each other as we hypothesized. Comparing the performance of stroke classifiers on both the technique-specific and acrosstechnique settings, we notice that in both cases the classifiers performed well. The GRU performed better in the across-technique settings than in the technique-specific setting, which can be justified by the lack of data in the later case. Table 4: Validation of Stroke Classifiers - One-vs-All Across-Techniques - Mean (std) of five folds Hand-crafted + SVM GRU Classifier technique train test train test Picasso vs. all All 72.59% (1.19 %) 67.26% (8.37 %) 81.92% (2.59 %) 75.09% ( 5.09%) Matisse vs. all All 65.83% (1.72 %) 60.61% ( 8.71%) 81.01% (3.41 %) 72.68% (5.58 %) Schiele vs. all All 84.76% ( 0.91%) 81.49% (3.30 %) 85.55% (1.74 %) 78.54% ( 8.77%) Technique-specific - Mean (std) of five folds Hand-crafted + SVM GRU Classifier Technique train test train test Picasso vs. all Pen/Ink 73.20% ( 2.21%) 68.93% ( 7.04%) 84.08% ( 2.20%) 72.24% (1.87 %) Matisse vs. all Pen/Ink 73.35% ( 1.99 %) 70.08% ( 7.94%) 86.88% (1.98 %) 75.03% ( 6.47%) Schiele vs. all Pencil 82.58% ( 2.78%) 75.39% (20.64 %) 94.33% (3.52 %) 69.60% (20.62 %) 5.2.3 Combined train test 86.05% (1.08 %) 78.54% (4.36 %) 87.92% ( 1.73%) 77.08% ( 4.33%) 91.85% (0.87 %) 86.20 % (3.78%) Combined train test 88.40% ( 1.19%) 75.92% ( 4.22%) 91.56% ( 1.03%) 79.10% (6.65 %) 91.30% (4.57 %) 72.93% (19.67 %) Multi-class Stroke Classification Experiment Although for attribution and authentication one-vs-all setting is the most obvious choice, we also tested a multi-class setting for completeness. In this experiment we train and test stroke classifiers built to discriminate between 5 classes: Picasso, Matisse, Schiele, Modigliani, and Others. The challenge in this setting is that training and test data will be bounded to the class that has the least number of samples (since we equalize the number of samples in training and test sets to avoid data bias). In this experiment we compared the performance of the hand-crafted features and the GRU features. For the GRU, the output directly has 5 nodes to encode the classes. For the hand-crafted 16 features we used error-correcting output codes (ECOC) classification setup where binary SVM classifiers were trained for each pair of classes. Table 5 shows the results of five-fold cross-validation experiment. There is a significant difference in performance between the two types of features in this experiment, which is far from the differences in all other experiments. We hypothesize that this is because the ECOC setting limits significantly the number of data samples used for training each binary classifier, while the GRU utilized the samples from all 5 classes in training. In particular, the number of training strokes per classes in this experiment were 1418, 1656, 1551, 1162, 1317 for each of the 5-fold splits respectively, which is a very small number. As a conclusion, we rule out the use of multi-class setting for attribution and authentication due to the hardship in obtaining sizable collections of data set. Instead, for drawing classification and fake detection we relay only on one-vs-all settings. Table 5: Validation of Stroke Classifier - Across Techniques ( 5 Classes: Picasso, Matisse, Schiele, Modigliani, Others) - five-folds mean (std) Approach Train Test Hand-Crafted - SVM 55.01 % (1.41%) 48.97% (5.82%) GRU 87.72% (2.43%) 74.65% (3.41%) 5.3 Drawing Classification and Detection of Fakes Drawing Classification Validation: Given the trained stroke classifiers, their performance is tested on drawing classification settings, also using one-vs-all settings. We used the four aforementioned strategies for aggregating the results from the stroke level to the drawing level. Given that the stroke classifiers are trained on a five-fold cross-validation setting, the drawing classification followed that strategy, i.e. in each fold, each drawing in the test split is classified using the classifier trained on the 80% of the images in the training split, hence there is no standard deviation to report. Table 6 shows the results for the across-technique setting and Table 7 shows the results for the techniquespecific setting. Evaluation on Fake Drawings: The trained stroke classifiers were also tested on the collected fake drawings to evaluate whether the classifiers are really capturing artists’ stroke characteristics and invariants or just statistics that can be easily deceived by forged versions. We used the Picassovs-all stroke classifiers to test the fake drawings that are made to imitate Picasso drawings (We denote them as Picasso fakes). A similar setting is used for Matisse fakes and Schiele fakes. Since the stroke classifiers are trained on a five-fold setting, we have five different classifiers trained per artist, one for each fold. Each test stroke is classified using the five classifiers and the majority vote is computed. The different aggregation methods are used to achieve a final classification for each drawing. Since one-vs-all setting is adapted, classifying a fake Picasso as others in a Picasso-vs-all setting is considered a correct classification, while classifying fake Picasso as Picasso is considered a wrong prediction. The bottom parts of Table 6 and Table 7 shows the classification results for the fake dataset for the across-technique and technique-specific settings respectively. The table shows that the trained one-vs-all stroke classifiers for all the three artists, are robustly rejecting fake drawing with accuracy reaching 100% in the across-technique case. A notable dif17 ference here is that the GRU failed to detect the fake drawings, in particular for the Picasso-vs-all, while the hand-crafted features detected all the fakes. Similar case happens for Schiele-vs-all as well. We hypothesize that this is because of the limited training data in the technique-specific case, which did not allow the GRU to learn an invariant model that generalizes well as in the acrosstechnique case. In contrast the hand-crafted models did not suffer from this limitation. Overall, the hand-crafted features are outperforming in detecting the fakes. Table 6: Validation of Drawing Classifiers - - One-vs-All -Across Techniques Across-Techniques Matisse-vs-All Combined Hand-crafted GRU Combined 82.49% 54.88% 81.14% 80.47% 81.48% 56.90% 81.48% 79.12% 82.83% 38.05% 80.47% 78.79% 82.83% 58.25% 80.81% 80.47% Detection of Fake Drawings Picasso-vs-All Matisse-vs-All Aggregation Hand-crafted GRU Combined Hand-crafted GRU Combined Majority 100% 87.50 % 100% 76.92 % 100% 100% Posterior 100% 87.50 % 100% 76.92 % 100% 100% 85%-certain 100% 87.50 % 100% 76.92 % 100% 100% Certainty-weighted 100% 87.50 % 100% 76.92 % 100% 100% Picasso-vs-All Aggregation Hand-crafted GRU Majority 66.67% 76.77% Posterior 67.68% 77.44% 85%-certain 73.06% 79.80% Certainty-weighted 67.34% 79.80% Schiele-vs-All Hand-crafted GRU Combined 74.41% 82.49% 81.82% 74.75% 83.50% 82.49% 75.42% 83.50% 83.84% 75.42% 85.19% 83.16% Schiele-vs-All Hand-crafted GRU Combined 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% Table 7: Validation of Drawing Classifiers - One-vs-All - Technique-specific Technique-Specific Picasso-vs-All Matisse-vs-All Aggregation Hand-crafted GRU Combined Hand-crafted GRU Majority 72.41% 82.76% 81.38% 65.52% 78.62% Posterior 72.41% 82.76% 81.38% 66.21% 79.31% 85%-certain 72.41% 82.76% 82.76% 69.66% 76.55% Certainty-weighted 71.72% 82.76% 82.07% 69.66% 77.93% Detection of Fake Drawings Picasso-vs-All Matisse-vs-All Aggregation Hand-crafted GRU Combined Hand-crafted GRU majority 100.00% 12.50% 16.67% 94.87% 100.00% Posterior 100.00% 12.50% 16.67% 97.44% 100.00% k-certain 100.00% 12.50% 20.83% 97.44% 100.00% certainity-weighted 100.00% 12.50% 20.83% 97.44% 100.00% 6 Combined 82.76% 80.69% 80.69% 80.00% Schiele-vs-All Hand-crafted GRU Combined 81.25% 78.12% 81.25% 84.38% 78.12% 81.25% 84.38% 78.12% 81.25% 87.50% 78.12% 81.25% Combined 100.00% 100.00% 100.00% 100.00% Schiele-vs-All Hand-crafted GRU Combined 100.00% 45.00% 55.00% 100.00% 45.00% 55.00% 100.00% 45.00% 60.00% 100.00% 45.00% 60.00% Conclusion In this paper we proposed an automated method for quantifying the characteristics of artist stokes in drawings. The approach is inspired by the Pictology methodology proposed by van Dantzig. The approach segments the drawing into individual strokes using a novel segmentation algorithm. 18 The characteristics of each stroke is captured using global and local shape features as well as a deep neural network that captures the local shape and tone variations of each stroke. We compared different types of features and showed results at the stroke classification and drawing classification levels. The main result of this paper is that it shows that we can discriminate between artists at the stroke-level with high accuracy, even using images of drawing of typical off-the-web or scanned books resolutions. We also tested the methodology using a collected data set of fake drawings and the results show that the proposed method is robust to such imitated drawings, which highlights that the method can indeed capture artists’ invariant characteristics that is hard to imitate. References [1] M. M. van Dantzig. Pictology: An Analytical Method for Attribution and Evaluation of Pictures. E. J. Brill, Leiden, Netherlands, 1973. [2] Josef Riederer. The detection of art forgeries with scientific methods. In Forensic science progress, pages 153–168. Springer, 1986. [3] Jinhong Katherine Guo, David Doermann, and A Rosenfield. Off-line skilled forgery detection using stroke and sub-stroke properties. In Pattern Recognition, 2000. Proceedings. 15th International Conference on, volume 2, pages 355–358. IEEE, 2000. [4] C Richard Johnson, Ella Hendriks, Igor J Berezhnoy, Eugene Brevdo, Shannon M Hughes, Ingrid Daubechies, Jia Li, Eric Postma, and James Z Wang. Image processing for artist identification. IEEE Signal Processing Magazine, 25(4), 2008. [5] Güngör Polatkan, Sina Jafarpour, Andrei Brasoveanu, Shannon Hughes, and Ingrid Daubechies. Detection of forgery in paintings using supervised learning. In Image Processing (ICIP), 2009 16th IEEE International Conference on, pages 2921–2924. IEEE, 2009. [6] Jia Li, Lei Yao, Ella Hendriks, and James Z Wang. Rhythmic brushstrokes distinguish van gogh from his contemporaries: findings via automated brushstroke extraction. IEEE transactions on pattern analysis and machine intelligence, 34(6):1159–1176, 2012. [7] Walter Liedtke, C Richard Johnson Jr, and Don H Johnson. Canvas matches in vermeer: A case study in the computer analysis of fabric supports. Metropolitan Museum Journal, 47(1):101–108, 2012. [8] R. Sablatnig, P. Kammerer, and E. Zolda. Hierarchical classification of paintings using face- and brush stroke models. In ICPR, 1998. [9] Maria Vanrell Fahad Shahbaz Khan, Joost van de Weijer. Who painted this painting? 2010. [10] Thomas Edward Lombardi. The classification of style in fine-art painting. ETD Collection for Pace University. Paper AAI3189084., 2005. [11] Ravneet Singh Arora and Ahmed M. Elgammal. Towards automated classification of fine-art painting style: A comparative study. In ICPR, 2012. [12] Babak Saleh, Kanako Abe, Ravneet Singh Arora, and Ahmed Elgammal. Toward automated discovery of artistic influence. Multimedia Tools and Applications, pages 1–27, 2014. 19 [13] Louisa Lam, Seong-Whan Lee, and Ching Y Suen. Thinning methodologies-a comprehensive survey. IEEE Transactions on pattern analysis and machine intelligence, 14(9):869–885, 1992. [14] Wilhelm Burger and Mark J. Burge. Fourier Shape Descriptors, pages 665–711. Springer London, London, 2016. [15] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735–1780, 1997. [16] J. Chung, C. Gulcehre, K. Cho, and Y. Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. NIPS’2014 Deep Learning workshop, 2014. [17] Y. Bengio, P. Simard, and P. Frasconi. Learning long-term dependencies with gradient descent is difficult. IEEE Transactions on Neural Nets, pages 157–166, 1994. [18] S. Hochreiter, Y. Bengio, P. Frasconi, and J. Schmidhuber. Gradient flow in recurrent nets: the difficulty of learning long-term dependencies. Field Guide to Dynamical Recurrent Networks, 2000. [19] Corinna Cortes and Vladimir Vapnik. Support-vector networks. Machine Learning, 20(3):273–297, 1995. 20 Appendix: Ablation studies and other experiments Adaptive GRU Patch Radius Table 9 shows a comparison between choosing the patch size based on an adaptive radius vs. fixed radius for the GRU model. For the fixed radius case, 11x11 patches are used. For the adaptive case, a radius r is computed for each drawing by computing the mean rip length for each stroke and taking the median over all strokes in the drawing. Square patches of size 2 ∗ r + 1 are used and scaled to 11x11 patches. The comparison in the table shows that the adaptive radius does not improve over the fixed radius in most of the cases. The comparison is shown for both the acrosstechniques and technique-specific cases. The adaptive radius showed improvement only in the case of technique-specific Schiele vs. all classification, Table 8: Comparison of GRU Drawing Classifiers using Adaptive vs. Fixed Radius Patches Across-Techniques one-vs-all drawing classification Picasso-vs-All Matisse-vs-All Aggregation Adaptive Radius Fixed Radius Adaptive Radius Fixed Radius Majority 78.11% 76.77% 78.79% 81.14% Posterior 78.11% 77.44% 78.45% 81.48% 85%-certain 79.46% 79.80% 76.77% 80.47% Certainty-weighted 79.46% 79.80% 78.45% 80.81% Technique-Specific one-vs-all drawing classification Picasso-vs-All Matisse-vs-All Aggregation Adaptive Radius Fixed Radius Adaptive Radius Fixed Radius Majority 77.24% 82.76% 75.17% 78.62% Posterior 77.24% 82.76% 74.48% 79.31% 85%-certain 77.24% 82.76% 76.55% 76.55% Certainty-weighted 78.62% 82.76% 77.24% 77.93% Schiele-vs-All Adaptive Radius Fixed Radius 73.40% 82.49% 74.41% 83.50% 76.77% 83.50% 77.78% 85.19% Schiele-vs-All Adaptive Radius Fixed Radius 90.62% 78.12% 90.62% 78.12% 90.62% 78.12% 87.50% 78.12% Ablation study of the Hand-crafted Features In this set of experiments we conduct an ablation study of the elements of the hand-crafted stroke features. These experiment is done on a binary classification setting to discriminate between the strokes of Picasso and Matisse drawn using ink/pen technique. SVM with polynomial kernel of degree 3 is used in all experiments. Five-fold cross validation is performed. 21 Table 9: Ablation study of the Hand-crafted features - five-folds Feature training accuracy mean std Fourier Descriptors (FD) 61.75% 1.11% Reconstruction Error Profile (REP) 64.39% 0.75% Stroke Thickness Profile (STP) 74.11% 1.33% Curvature 66.22% 0.60% Stroke Length (SL) 57.30% 0.28% FD+REP 69.23% 0.57% STP + Curvature 79.63% 0.71% STP + Curvature + SL 80.60% 0.80% FD+REP+STP 77.42% 1.95% FD+REP+STP+Curvature 85.91% 0.88% FD+REP+STP+Curvature+SL 86.70% 1.08% 22 test accuracy mean std 57.55% 5.46% 63.20% 11.09% 63.68% 2.58% 60.17% 2.06% 57.12% 1.86% 64.23% 6.69% 68.93% 4.04% 70.17% 3.81% 68.35% 4.58% 74.13% 6.47% 75.07% 5.67% Figure 3 key: Fake, Fake, Matisse Matisse, Fake, Fake, Matisse Fake, Matisse, Picasso, Fake Fake, Picasso, Picasso, Fake Schiele, Fake, Fake, Schiele, Schiele, Fake About the Authors: Dr. Ahmed Elgammal: is a professor at the Department of Computer Science, Rutgers University. He is the founder and director of the Art and Artificial Intelligence at Rutgers, which focuses on data science in the domain of digital humanities. He is also an Executive Council Faculty at Rutgers University Center for Cognitive Science. Prof. Elgammal has published over 160 peer-reviewed papers, book chapters, and books in the fields of computer vision, machine learning and digital humanities. He is a senior member of the Institute of Electrical and Electronics Engineers (IEEE). He received the National Science Foundation CAREER Award in 2006. Dr. Elgammal recent research on knowledge discovery in art history received wide international media attention, including reports on the Washington Post, New York Times, NBC News, the Daily Telegraph, Science News, and many others. Dr. Elgammal is the founder and CEO of Artrendex, a startup that builds innovative AI technology for the art market. Dr. Elgammal received his M.Sc. and Ph.D. degrees in computer science from the University of Maryland, College Park, in 2000 and 2002, respectively. Yan Kang main research interests are Deep Learning,Computer Vision and Natural Language Processing. Mr. Kang was a senior deep learning researcher at Artrendex. Currently Mr. Kang is a Computer Vision Engineer at JD.com American Technologies Corporation, working on Object Classification & Detection and Neural Network Compression. He obtained his M.S. degree in Computer Science at Rutgers University (New Brunswick) in May 2016. 23 Milko den Leeuw is a painting conservator, specialized in the technical and scientific investigation of paintings. He completed his training in painting conservation and pictology (an analytical method for attribution and evaluation of paintings) at the studio Dora van Dantzig in Amsterdam in 1989. In 1991 he founded the Atelier for Restoration & Research of Paintings (ARRS) with drs Ingeborg de Jongh. ARRS was from 1991-2006 the outdoor conservator atelier of the Rijksmuseum Catharijneconvent Utrecht. ARRS has produced numerous international publications concerning new techniques in authentication research and rediscoveries of paintings. ARRS has been involved in the conservation of important masters in Dutch Fine Art such as Rembrandt van Rijn, Hendrick ter Brugghen and Gerard ter Borch and has contributed to the rediscovery of paintings by Jan Lievens, Capser Netscher and Quinten Matsijs. ARRS has also performed conservation and research on National Art Treasures like the ceiling dating from 1672 by Gerard de Lairesse from the International Court for Peace & Justice in The Hague (Carnegie Foundation), and Italian masters Giovanni Bellini and Canaletto, French masters Manet and Degas, German masters Lucas Cranach and Max Beckmann, Russian masters as Kasimir Malevich and Nathalia Goncharova and American masters Andy Warhol and Mark Rothko. ARRS is a professional equipped lab for conservation treatment and (authentication) research on paintings. ARRS holds a worldwide reputation for bridging art history, conservation technique and material sciences. ARRS is Authentication in Art organizer of the 2012, 2014, 2016 AiA conferences. 24 

1cs.CV

Evaluating and Tuning n-fold Integer Programming Kateřina Altmanová Department of Applied Mathematics, Charles University Prague, Czech Republic kacka@kam.mff.cuni.cz Dušan Knop1 arXiv:1802.09007v1 [cs.DS] 25 Feb 2018 Department of Informatics, University of Bergen Bergen, Norway dusan.knop@uib.no 0000-0003-2588-5709 Martin Koutecký2 Faculty of Industrial Engineering and Management, Technion – Israel Institute of Technology Haifa, Israel koutecky@technion.ac.il 0000-0002-7846-0053 Abstract In recent years, algorithmic breakthroughs in stringology, computational social choice, scheduling, etc., were achieved by applying the theory of so-called n-fold integer programming. An n-fold integer program (IP) has a highly uniform block structured constraint matrix. Hemmecke, Onn, 2 and Romanchuk [Math. Programming, 2013] showed an algorithm with runtime aO(rst+r s) n3 , where a is the largest coefficient, r, s, and t are dimensions of blocks of the constraint matrix and n is the total dimension of the IP; thus, an algorithm efficient if the blocks are of small size and with small coefficients. The algorithm works by iteratively improving a feasible solution with augmenting steps, and n-fold IPs have the special property that augmenting steps are guaranteed to exist in a not-too-large neighborhood. However, this algorithm has never been implemented and evaluated. We have implemented the algorithm and learned the following along the way. The original algorithm is practically unusable, but we discover a series of improvements which make its evaluation possible. Crucially, we observe that a certain constant in the algorithm can be treated as a tuning parameter, which yields an efficient heuristic (essentially searching in a smaller-thanguaranteed neighborhood). Furthermore, the algorithm uses an overly expensive strategy to find a “best” step, while finding only an “approximatelly best” step is much cheaper, yet sufficient for quick convergence. Using this insight, we improve the asymptotic dependence on n from n3 to n2 log n which yields the currently asymptotically fastest algorithm for n-fold IP. Finally, we tested the behavior of the algorithm with various values of the tuning parameter and different strategies of finding improving steps. First, we show that decreasing the tuning parameter initially leads to an increased number of iterations needed for convergence and eventually to getting stuck in local optima, as expected. However, surprisingly small values of the parameter already exhibit good behavior. Second, our new strategy for finding “approximatelly best” steps wildly outperforms the original construction. 2012 ACM Subject Classification Software and its engineering → Software design engineering, Theory of computation → Parameterized complexity and exact algorithms 1 2 Author supported supported by the project NFR MULTIVAL and the project P202/12/G061 of GA ČR. Author supported by Technion postdoctoral fellowship. 23:2 Evaluating and Tuning n-fold Integer Programming Keywords and phrases n-fold integer programming, integer programming, analysis of algorithms, primal heuristic, local search Supplement Material https://github.com/katealtmanova/nfoldexperiment 1 Introduction In this article we consider the general integer linear programming (ILP) problem in standard form, min {wx | Ax = b , l ≤ x ≤ u , x ∈ Zn } . (ILP) with A an integer m × n matrix, b ∈ Z , w ∈ Z , l, u ∈ (Z ∪ {±∞}) . It is well known to be strongly NP-hard, but models many important problems in combinatorial optimization such as planning [27], scheduling [12] and transportation [3] and thus powerful generic solvers have been developed for it [24]. Still, theory is motivated to search for tractable special cases One such special case is when the constraint matrix A has a so-called N -fold structure:   E1 E1 · · · E1  E2 0 · · · 0      A = E (N ) =  0 E2 · · · 0  .  . .. ..  ..  .. . . .  m 0 0 ··· n n E2 Here, r, s, t, N ∈ N, u, l, w ∈ ZN t , b ∈ Zr+N s , E (N ) is an (r + N s) × N t-matrix, E1 ∈ Zr×t is an r ×
t-matrix and E2 ∈ Zs×t is an s × t-matrix. We call E (N ) the N -fold product of (N ) 1 E= E is known as N -fold integer programming (N -fold E2 . Problem (ILP) with A = E IP). Hemmecke, Onn, and Romanchuk [15] prove the following. I Proposition 1 ([15, Theorem 6.2]). There is an algorithm that solves3 (ILP) with A = E (N ) 2 encoded with L bits in time ∆O(trs+t s) · n3 L, where ∆ = 1 + max{kE1 k∞ , kE2 k∞ }. Recently, algorithmic breakthroughs in stringology [19], computational social choice [20], scheduling [4, 17, 21], etc., were achieved by applying this algorithm and its subsequent non-trivial improvements. The algorithm belongs to the larger family of augmentation (primal) algorithms. It starts with an initial feasible solution x0 ∈ ZN t and produces a sequence of increasingly better solutions x1 , . . . , xs (better means wxs < wxs−1 < · · · < wx0 ). It is guaranteed that the algorithm terminates, that xs is an optimal solution, and that the algorithm converges quickly, i.e., s is polynomial in the length of the input. A key property of N -fold IPs is that, if an augmenting step exists, then it can be decomposed into a bounded number of elements of the so-called Graver basis of A, which in turn makes it possible to compute it using dynamic programming [15, Lemma 3.1]. In a sense, this property makes the algorithm a local search algorithm which is always guaranteed to find an improvement in a not-too-large neighborhood. The bound on the number of elements or the size of the neighborhood which needs to be searched is called the Graver complexity of E. This, in turn, implies that, if an augmenting step exists, then there is always one with small `1 -norm; for a matrix A, we denote this bound g1 (A) = maxg∈G(A) kgk1 [23, Theorem 4]. However, the algorithm has never been implemented and evaluated. 3 Given an IP, we say that to solve it is to either (i) declare it infeasible or unbounded or (ii) find a minimizer of it. K. Altmanová, D. Knop, and M. Koutecký 1.1 23:3 Our Contributions We have implemented the algorithm and tested it on two problems for which n-fold formulations were known: makespan minimization on uniformly related machines (Q||Cmax ) and Closest String; we have used randomly generated instances. In the course of implementing the algorithm we learn the following. The algorithm in its initial form is practically unusable due to an a priori construction of the Graver basis G(E2 ) of size exponential in r, s, t and a related (even larger) set Z(E). However, we discover a series of improvements (some building on recent insights [23]) which avoid the construction of these two sets. Moreover, we adjust the algorithm to treat g1 (A) as a tuning parameter g1 , which turns it into a heuristic. We also study the augmentation strategy, which is the way the algorithm chooses an augmenting step among all the possible options. The original algorithm uses an overly expensive strategy to find a “best” step, which means that a large number of possible steps is evaluated in each iteration. We show that finding only an “approximatelly best” step is sufficient to obtain asymptotically equivalent convergence rate, and the work per iteration decreases exponentially. Using this insight, we improve the asymptotic dependence on N from N 3 to N 2 log N . Together with recent improvements, this yields the currently asymptotically fastest algorithm for N -fold IP: I Theorem 2. Problem (ILP) with A = E (N ) can be solved in time ∆r where M = log(wx∗ − wx0 ) for some minimizer x∗ of wx. 2 s+rs2 (N t)2 log(N t)M , Finally, we evaluate the behavior of the algorithm. We ask how is the performance of the algorithm (in terms of number of dynamic programming calls and quality of the returned solution) influenced by 1. the choice of the tuning parameter 1 < g1 ≤ g1 (A)? 2. the choice of the augmentation strategy between “best step”, “approximate best step” and “any step”? As expected, as g1 moves from g1 (A) to 1, we first see an increase in the number of iterations needed for convergence and eventually the algorithm gets stuck in local optima. However, surprisingly small values (e.g. g1 = 5 when g1 (A) > 105 ) of the parameter already exhibit close to optimal behavior. Second, our new strategy for finding “approximatelly best” steps outperforms the original construction by orders of magnitude, while the naive “any step” strategy behaves erratically. We note that at this stage we are not (yet) interested in showing supremacy over existing algorithms; we simply want to understand the practical behavior of an algorithm whose theoretical importance was recently highlighted. For this reasons our experimental focus is on the two aforementioned questions rather than simply measuring the time. Similarly, due to the rigid format of E (N ) we are limited to few problems for which N -fold formulations are known. For Closest String we use the same instances as Chimani et al. [5]; for Makespan Minimization we generate our own data because standard benchmarks are not limited to short jobs or few types of jobs. 1.2 Related Work Our work mainly relates to primal heuristics [2] for MIPs which are used to help reach optimality faster and provide good feasible solutions early in the termination process. Specifically, our algorithm is a neighborhood (or local) search algorithm. The standard paradigm is Large Neighborhood Search (LNS) [26] with specializations such as for example Relaxation Induced 23:4 Evaluating and Tuning n-fold Integer Programming Neighborhood Search (RINS) [6] and Feasibility Pump [1]. Using this paradigm, our proposed algorithm searches in the neighborhood induced by the `1 -distance from the current feasible solution and the search procedure is formulated as an ILP subproblem with the additional constraint kxk1 ≤ g1 . In this sense the closest technique to ours is local branching [10] which also searches in the `1 -neighborhood; however, it was mainly applied to binary programs without any additional structure of the constraint matrix. On the theoretical side, very recently Koutecký et al. [23] have studied parameterized strongly polynomial algorithms for various block-structured ILPs, not just N -fold IP. Eisenbrand et al. [8] independently (and using slightly different techniques) arrive at the same complexity of N -fold IP as our Theorem 2. 2 Preliminaries For positive integers m, n we set [m, n] = {m, . . . , n} and [n] = [1, n]. We write vectors in boldface (e.g., x, y) and their entries in normal font (e.g., the i-th entry of x is xi ). Given the problem (ILP), we say that x is feasible for (ILP) if Ax = b and l ≤ x ≤ u. Graver bases and augmentation. Let us now introduce Graver bases and discuss how they can be used for optimization. We also recall N -fold IPs; for background, we refer to the books of Onn [25] and De Loera et al. [7]. Let x, y be n-dimensional integer vectors. We call x, y sign-compatible if they lie in the same orthant, that is, for each i ∈ [n] the sign of xi and yi is the same. We call P i i i g a sign-compatible sum if all g are pair-wise sign-compatible. Moreover, we write y v x if x and y are sign-compatible and |yi | ≤ |xi | for each i ∈ [n]. Clearly, v imposes a partial order called “conformal order” on n-dimensional vectors. For an integer matrix A ∈ Zm×n , its Graver basis G(A) is the set of v-minimal non-zero elements of the lattice of A, kerZ (A) = {z ∈ Zn | Az = 0}. An important property of G(A) is the following. I Proposition 3 ([25, Lemma 3.4]). Every integer vector x 6= 0 with Ax = 0 is a sign-comPt patible sum x = i=1 αi gi , αi ∈ N, gi ∈ G(A) and t ≤ 2n − 2. Let x be a feasible solution to (ILP). We call g an x-feasible step (or simply feasible if x is clear) if x + g is feasible for (ILP). Further, we call a feasible step g augmenting if w(x + g) < wx; note that g decreases the objective by wg. An augmenting step g and a step length γ ∈ N form an x-feasible step pair with respect to a feasible solution x if l ≤ x + γg ≤ u. A pair (γ, g) ∈ (N × G(A)) is a γ-Graver-best step pair and γg is a γ-Graver-best step if it is feasible and for every feasible step pair (γ, g0 ), g0 ∈ G(A), we have wγg ≤ wγg0 . An augmenting step g and a step length γ ∈ N form a Graver-best step pair if it is γ-Graver-best and it minimizes wγ 0 g0 over all γ 0 ∈ N, where (γ 0 , g0 ) is a γ 0 -Graver-best step pair. We say that γg is a Graver-best step if (γ, g) is a Graver-best step pair. The Graver-best augmentation procedure for (ILP) with a given feasible solution x0 and initial value i = 0 works as follows: 1. If there is no Graver-best step for xi , return it as optimal. 2. If a Graver-best step γg for xi exists, set xi+1 := xi + γg, i := i + 1, and go to 1. I Proposition 4 (Convergence bound [25, Lemma 3.10]). Given a feasible solution x0 for (ILP), the Graver-best augmentation procedure finds an optimum in at most (2n − 2) log M steps, where M = w(x0 − x∗ ) and x∗ is any minimizer of wx. By standard techniques (detecting unboundedness etc.) we can ensure that log M ≤ L. K. Altmanová, D. Knop, and M. Koutecký 23:5 N -fold IP. The structure of E (N ) allows us to divide the N t variables of x into N bricks of size t. We use subscripts to index within a brick and superscripts to denote the index of the brick, i.e., xij is the j-th variable of the i-th brick with j ∈ [t] and i ∈ [N ]. 3 Approximate Graver-best Steps In this section we introduce the notion of a c-approximate Graver-best step (Definition 5), show that such steps exhibit good convergence (Lemma 6), can be easily obtained (Lemma 7), and result in a significant speed-up of the N -fold IP algorithm (Theorem 2). I Definition 5 (c-approximate Graver-best step). Let c ∈ R, c ≥ 1. Given an instance of (ILP) and a feasible solution x, we say that h is a c-approximate Graver-best step for x if, for every x-feasible step pair (γ, g) ∈ (N × G(A)), we have wh ≤ 1c · γwg. Recall the Graver-best augmentation procedure. We call its analogue where we replace a Graver-best step with a c-approximate Graver-best step the c-approximate Graver-best augmentation procedure. I Lemma 6 (c-approximate convergence bound). Given a feasible solution x0 for (ILP), the c-approximate Graver-best augmentation procedure finds an optimum of (ILP) in at most c · (2n − 2) log M steps, where M = w(x0 − x∗ ) and x∗ is any minimizer of wx. Proof. The proof is a straightforward adaptation of the proof of Proposition 4. Let x∗ be P2n−2 a minimizer and let h = x∗ − x0 . Since Ah = 0, by Proposition 3, h = i=1 αi gi for some αi ∈ N, gi ∈ G(A), i ∈ [2n − 2]. Thus, an x-feasible step pair (γ, g) such that γg is 1 M . In other words, any Graver-best step pair a Graver-best step must satisfy wγg ≤ 2n−2 1 improves the objective function by at least a 2n−2 -fraction of the total optimality gap M , and thus (2n − 2) log M steps suffice to reach an optimum. It is straightforward to see that a c c-approximate Graver-best step satisfies wx − w(x + γg) ≤ 2n−2 M , and thus c(2n − 2) log M steps suffice. J I Remark. Lemma 6 extends naturally to separable objectives; see the original proof [25, Lemma 3.10]. I Lemma 7 (Powers of c step lengths). Let c ∈ N, x be a feasible solution of (ILP), and let  Γc-apx = ci | ∃g ∈ G(A) : l ≤ x + ci g ≤ u . Let (γ, g) ∈ (Γc-apx × G(A)) be an x-feasible step pair such that γg ≤ γ 0 g0 for any x-feasible step pair (γ 0 , g0 ) ∈ (Γc-apx × G(A)). Then γg is a c-approximate Graver-best step. Proof. Let (γ, g) satisfy the assumptions, and let (γ̃, g̃) ∈ (N × G(A)) be a Graver-best step pair. Let γ 0 be a nearest smaller power of c from γ̃, and observe that γ 0 g̃ is a c-approximate Graver-best step because γ 0 ≥ γ̃c . On the other hand, since γg is a γ-Graver-best step, we have γg ≤ γ 0 g̃ and thus γg is also a c-approximate Graver-best step. J I Theorem 2. Problem (ILP) with A = E (N ) can be solved in time ∆O(r where M = log(wx∗ − wx0 ) for some minimizer x∗ of wx. 2 s+rs2 ) (N t)2 log(N t)M , Proof. Recall that ∆ = kAk∞ +1. Koutecký et al. [23, Theorem 2] show that a γ-Graver-best 2 2 step can be found in time ∆r s+rs N t. Moreover, Hemmecke et al. [14] prove a proximity 23:6 Evaluating and Tuning n-fold Integer Programming theorem which allows the reduction of an instance of (ILP) to an equivalent instance with new bounds l0 , u0 satisfying ku0 − l0 k∞ ≤ N tg∞ , with g∞ = max kgk∞ ≤ max kgk1 ≤ (∆rs)O(rs) , g∈G(A) g∈G(A) where the last inequality can be found in the proof of [23, Theorem 4]. This bound implies that Γ2-apx from in Lemma 7 satisfies |Γ2-apx | ≤ rs log(∆N trs). By Lemma 7, finding a γ-Graver-best for each γ ∈ Γ2-apx and picking the minimum results in a 2-approximate 2 2 Graver-best step, and can be done in time ∆r s+rs (N t) log(N t). By Lemma 6, (4n−4) log M steps suffice to reach the optimum. J 4 Implementation We first give an overview of the original algorithm, which is our starting point. Then we discuss our specific improvements and mention a few details of the software implementation. 4.1 Overview of the Original Algorithm The key property of the N -fold product E (N ) is that, for any N ∈ N, the number of nonzero bricks of any g ∈ G(E (N ) ) is bounded by some constant g(E) called the Graver complexity of E, and, moreover, that the sum of all non-zero bricks of g can be decomposed into at most g(E) elements of G(E2 ) [15, Lemma 3.1]. This facilitates the following construction. Let ( Z(E) = z ∈ Zt | ∃g1 , . . . , gk ∈ G(E2 ), k ≤ g(E), z = k X ) gi . i=1 Then, every prefix sum of the bricks of g ∈ G(E (N ) ) is contained in Z(E) and a γ-Graver-best step, γ ∈ N, can be found using dynamic programming over the elements of Z(E). To ensure that a Graver-best step is found, a set of step-lengths Γbest is constructed as
follows. Observe that any Graver-best (and thus feasible) step pair (γ, g) ∈ N × G(E (N ) ) , must satisfy that in at least one brick i ∈ [n] it is “tight”, that is, (γ, g) is x-feasible while (γ + 1, g) is not specifically because li ≤ x + γg ≤ ui holds but li ≤ x + (γ + 1)g ≤ ui does not. Thus, for each z ∈ Z(E) and each i ∈ [n], we find all the potentially “tight” step lengths γ and add them to Γbest , which results in a bound of |Γbest | ≤ |Z(E)| · n. 4.2 Replacing Dynamic Programming with ILP We have started off by implementing the algorithm exactly as it is described by Hemmecke et al. [15]. The first obstacle is encountered almost immediately and is contained in the constant g(E). This constant can be computed, but the computation is extremely difficult [9, 13]. Another possibility is to estimate it, in which case it is almost always larger than N and thus is essentially meaningless. Finally, one can take the approach partially suggested in [15, Section 7], where we consider g(E) in the construction of Z(E) to be a tuning parameter and consider the approximate set Zgc (E), gc ∈ N, obtained by taking sums of at most gc elements of G(E2 ). This makes the algorithm more practical, but turns it into a heuristic. In spite of this sacrifice, already for small (r = 3, s = 1, t = 7, n = 10) instances and extremely small value of gc = 3, the dynamic programming based on the Zgc (E) construction was taking an unreasonably long time (over one minute). Admittedly this could be improved; however, already for gc > 5, it becomes infeasible to compute Zgc (E), and for larger instances K. Altmanová, D. Knop, and M. Koutecký 1 2 3 4 5 6 7 8 9 input : matrices E1 , E2 , positive integer N , and vectors b, l, u, w output : optimal solution to (ILP) with A = E (N ) g = GraverComplexity(E1 , E2 ); x0 = FindFeasibleSolution(E, b, l, u, w), i = 0; G(E1 ) = GraverBasis(E1 , g); Z(E) = DynamicProgramStates(G(E1 ), g); do Γ = BuildGamma(xi ); i = i + 1; foreach γ ∈ Γ do gγ = gammaBestStep(γ, g); xi = xi−1 + argmin{g 10 11 12 23:7 } wgγ ; γ |γ∈Γ while xi−1 6= xi ; return xi ; Algorithm 1: Pseudocode of the algorithm of Hemmecke, Onn, and Romanchuk. (r > 5, t > 12) it becomes very difficult to compute even G(E2 ). For these reasons we sought to completely replace the dynamic program involving Z(E). Koutecký et al. [23] show that all instances of (ILP) with the property that the so-called dual treedepth tdD (A) of A is bounded and the largest coefficient kAk∞ is bounded also have the property that g1 (A) = maxg∈G(A) kgk1 is bounded, which implies that augmenting steps can be found efficiently. This class of ILPs contains N -fold IP. The interpretation of the above fact is that, in order to solve (ILP), it is sufficient to repeatedly (for different x and γ) solve an auxiliary instance min {wh | Ah = 0, l ≤ x + γh ≤ u, khk1 ≤ g1 (A)} (AugILP) in order to find good augmenting steps; we note that the constraint khk1 ≤ g1 (A) can be linearized [23, Lemma 25]. The heuristic approach outlined above transfers easily: we replace g1 (A) in (AugILP) with some integer g1 , 1 < g1 ≤ g1 (A); this makes (AugILP) easier to solve at the cost of losing the guarantee that an augmenting step is found if one exists. In theory, solving (AugILP) should be easier than solving the original instance (ILP) due to the special structure of A [23, Lemma 25]. Our approach here is to simply invoke an industrial MILP solver on (AugILP) in order to find a γ-Graver-best step. 4.3 Augmentation Strategy: Step Lengths Logarithmic Γ. The majority of algorithms based on Graver basis augmentation rely on the Graver-best augmentation procedure [4, 7, 15, 19, 21, 25] and thus require finding (exact) Graver-best steps. In the aforementioned algorithms this is always done using the construction of a set Γbest mentioned above, which is of size f (k) · n where k is the relevant 2 parameter (e.g., (ars)O(rst+st ) in the original algorithm for N -fold IP). We replace this construction with Γ2-apx = {1, 2, 4, 8, . . . } which, combined with the proximity technique, is only of size O(log n) (Theorem 2); in particular, independent of the function f (k). Exhausting γ. Moreover, we have noticed that sometimes the algorithm finds a step g for γ = 2k which is not tight in any brick, and then repeatedly applies it for shorter step-lengths γ 0 < γ. In other words, the discovered direction g is not exhausted. Thus, for each γ ∈ N, 23:8 Evaluating and Tuning n-fold Integer Programming 1 2 3 4 5 6 7 8 9 10 input : matrices E1 , E2 , positive integers N , c, and g1 , and vectors b, l, u, w output : a feasible solution to (ILP) with A = E (N ) x0 = FindFeasibleSolution(E, b, l, u, w), i = 0; do Γ = ∅; j = 0, i = i + 1; do γ = cj ; gγ = min{wγg | Ag = 0, l ≤ x + γg ≤ u, kgk1 ≤ g1 , g ∈ ZN t }; γ 0 = ExhaustDirection(gγ ); Γ = Γ ∪ {γ 0 }, j = j + 1; while gγ 6= 0; xi = xi−1 + argmin{g |γ∈Γ} wgγ ; γ while xi−1 6= xi ; 12 return xi Algorithm 2: Pseudocode of our new heuristic algorithm. The algorithm is exact if g1 ≥ g1 (A) = maxg∈G(A) kgk1 . 11 upon finding the γ-Graver-best step g, we replace γ with the largest γ 0 ≥ γ for which (γ 0 , g) is still x-feasible. Early termination. Another observation is that in any given iteration of the algorithm, if γ > 1 then some augmenting step has been found and if the computation is taking too long, we might terminate it and simply apply the best step found so far. 4.4 Software and Hardware We have implemented our solver in the SageMath computer algebra system [30]. This was a convenient choice for several reasons. The SageMath system offers an interactive notebookstyle web-based interface, which allows rapid prototyping and debugging. Data types for vectors and matrices and a unified interface for several MILP solvers are also readily available. We have experimented with the open-source solvers GLPK [29], Coin-OR CBC [28], and the commercial solver Gurobi [11] and have settled for using the latter since it performs the best. The downside of SageMath is that an implementation of the original dynamic program is likely much slower than a similar implementation in C; however this DP is impractical anyway as explained in Section 4.2. For random instance generation and subsequent data evaluation and graphing, we have used the Jupyter notebook environment [18] and Matplotlib library [16]. The computations were performed on a computer with an Intel® Xeon® E5-2630 v3 (2.40GHz) CPU and 128 GB RAM. 5 Evaluation We begin our evaluation with two main questions, specifically, how is the performance of the algorithm (both in terms of the number of iterations and the quality of the returned solution) influenced by: 1. the tuning parameter g1 and 2. the augmentation strategy? K. Altmanová, D. Knop, and M. Koutecký Regarding our first question, theoretically we should see either an increase in the number of iterations, a decrease in the quality of the returned solution, or both. However, the range of the tuning parameter g1 is quite large: any number between 2 and g1 (A) is a valid choice, and in all our scenarios the true value of g1 (A) exceeds 300. Thus, we are interested in the threshold values of g1 when the algorithm no longer finds the true optimum or when its convergence rate drops significantly. Regarding our second question, there are two main candidates for the set of step-lengths Γ. We can either use the “best step” construction Γbest of the original algorithm, which assures that we always make a Graver-best step before moving to the next iteration. Or, we can use the “approximate best step” construction Γ2-apx of Theorem 2, which provides a 2-approximate Graver-best step. To make this comparison more interesting, we also consider Γ5-apx and also the trivial “any step” strategy where we always make the 1-Graver-best step, which corresponds to taking Γany = {1}. Recall that due to the trick of always exhausting the discovered direction, this strategy actually has a chance at quick convergence, unlike if we only made the step with γ = 1. 5.1 Instances We choose two problems for which N -fold IP formulations were shown in the literature, namely the Q||Cmax scheduling problem [21] and the Closest String problem [19]. Uniformly related machines makespan minimization (Q||Cmax ) Input: Set of m machines M , each with a speed si ∈ N. A set of n jobs J, each with a processing time pj ∈ N. Find: Find an assignment of jobs to machines such that the time when last job finishes (the makespan) is minimized; a job j scheduled on a machine i takes time pj /si to execute. Closest String Input: A set of k strings s1 , . . . , sk of length L over an alphabet Σ. Find: Find a string y ∈ ΣL minimizing maxki=1 dH (si , y), where dH is the Hamming distance. For both problems we generate random instances as follows. Scheduling. We view the problem as a decision problem where, given a number B, we ask whether a schedule of makespan at most B exists. This is equivalent to the multi-sized bin packing problem, where we have m bins of various capacities instead of m machines of different speeds, and we adopt this view as it is more convenient. We also view it as a high-multiplicity problem where the items are not given explicitely as a list of item sizes, but succinctly by a vector of item multiplicities. Because the algorithm is primarily an optimization algorithm, we follow the standard approach [15, Lemma 3.8] and turn the feasibility problem into an auxiliary optimization instance where finding a starting feasible solution is easy. Specifically for Q||Cmax this means introducing auxiliary slack variables for “not-yet-scheduled” jobs and minimizing the total “not-yet-scheduled” length. The input parameters of the instance generation are number of bins m, the smallest and largest capacities S and L, respectively, item sizes p1 , . . . , pk and probability weights Pk w1 , . . . , wk , W = i=1 wi , and a slack ratio σ. The instance is then generated as follows. First, we choose m capacities from [S, L] uniformly at random. This determines the total capacity of the bins C. Our goal is to generate items whose total size is roughly σ · C. We do this by repeatedly picking an item length from p1 , . . . , pk , where pj is selected with probability wj /W , until the total size of items picked so far exceeds σ · C, when we terminate and return the generated instance. 23:9 23:10 Evaluating and Tuning n-fold Integer Programming Closest String. As before, we view the problem as a decision problem: given a number d ∈ N, decide whether there is a string y with maxki=1 dH (si , y) ≤ d. The random instance is generated exactly as done by Chimani et al. [5]: first, we generate a random “target” string y ∈ ΣL and create k copies s1 , . . . , sk of it; then, we make α random changes in s1 , . . . , sk . This way, we have an upper bound α on the optimum. The input parameters of the instace generation are thus k, L, Σ and the distance ratio r such α = n/r. Again, we solve an auxiliary instance where we essentially start with a string of “all blanks” and try to fill in all the blanks while staying in the specified distance d; the objective is thus the remaining number of blanks. 5.2 Results Here we demonstrate the overall behavior of the algorithm on three selected instances; we encourage the reader to see the full data (incl. plots) at https://github.com/katealtmanova/ nfoldexperiment. We have determined that the sensible range of values of g1 is between 2 and roughly 100 and beyond that the behavior does not change significantly. For all augmentation strategies there are values of g1 which take much longer to converge than any other values; because of these outliers we clip our figures. To expose the behavior of the algorithm we use two types of figures. Outer loop. In the outer loop figure we focus on the behavior of the algorithm with respect to the loop starting at line 2 of Algorithm 2, i.e., where each iteration corresponds to making an augmenting step. There is a line plot for each tested value of g1 . The x axis shows the iteration, the y axis shows the objective value attained in this iteration, i.e.,wxi for iteration i. We indicate the expensiveness of computing one augmentation by the thickness of the line in a given iteration – the thicker the line, the more times the (AugILP) has been solved in this iteration. The legend indicates the exact number of times (AugILP) has been solved for this value of g1 . Inner loop. In the inner loop figure we focus on the loop starting at line 4 of Algorithm 2, i.e., where iterations correspond to solutions of (AugILP). As before, each color corresponds to a tested value of g1 . There is a line plot displaying the minimum over augmenting steps found in each outer iteration; however now there is also a semiopaque region above this line, indicating the values of all the augmenting steps (including the non-minimal ones) found in this iteration. We chose two instances among the tested once as representative of the overall behavior: Our first instance is Q||Cmax with parameters m = 15, S = 2000, L = 10000, item sizes (2, 3, 13, 35) (so that we have an instance with nontrivial kAk∞ ), and weights (6, 13, 2, 1) and σ = 0.45. The theoretical upper bound on g1 (A) is (rskAk∞ + 1)O(rs) , and here we have r = 4, s = 1 and kAk∞ = 35; thus, without computing g1 (A) exactly, we should consider it to be at least (4 · 36)4 . See Figures 1 and 2. The third instance is Closest String with parameters k = 5, |Σ| = 2, L = 10000 and r = 1. The N -fold model has r = 5, s = 1 and kAk∞ = 1, thus, without computing g1 (A) exactly, we should consider it to be at least (2 · 5)5 . See Figure 3. 5.3 Conclusions Our main takeaway regarding Question #1 is that, while the theoretical upper bounds for g1 (A) are huge, already small values of g1 (g1 > 5 for Closest String and g1 > 20 K. Altmanová, D. Knop, and M. Koutecký 35000 30000 objective 25000 20000 15000 10000 35000 30000 25000 20000 15000 10000 5000 5000 5 10 15 20 iteration 25 30 35 40 g1=2 AugILP: 1 g1=3 AugILP: 1 g1=4 AugILP: 5 g1=5 AugILP: 15 g1=6 AugILP: 15 g1=7 AugILP: 11 g1=8 AugILP: 16 g1=9 AugILP: 16 g1=10 AugILP: 12 g1=12 AugILP: 12 g1=14 AugILP: 5304 g1=16 AugILP: 111 g1=18 AugILP: 127 g1=20 AugILP: 114 g1=23 AugILP: 129 g1=26 AugILP: 107 g1=29 AugILP: 128 g1=33 AugILP: 106 g1=40 AugILP: 88 g1=45 AugILP: 81 g1=50 AugILP: 74 g1=60 AugILP: 58 g1=75 AugILP: 50 g1=90 AugILP: 48 g1=110 AugILP: 38 40000 35000 30000 25000 20000 15000 10000 5000 0 5 10 15 20 iteration 25 30 35 40 g1=2 AugILP: 1 g1=3 AugILP: 1 g1=4 AugILP: 2 g1=5 AugILP: 4 g1=6 AugILP: 4 g1=7 AugILP: 4 g1=8 AugILP: 4 g1=9 AugILP: 6 g1=10 AugILP: 4 g1=12 AugILP: 6 g1=14 AugILP: 5765 g1=16 AugILP: 40 g1=18 AugILP: 127 g1=20 AugILP: 257 g1=23 AugILP: 96 g1=26 AugILP: 119 g1=29 AugILP: 1864 g1=33 AugILP: 99 g1=40 AugILP: 36 g1=45 AugILP: 37 g1=50 AugILP: 30 g1=60 AugILP: 80 g1=75 AugILP: 30 g1=90 AugILP: 44 g1=110 AugILP: 52 40000 35000 30000 25000 objective 0 objective g1=2 AugILP: 1 g1=3 AugILP: 1 g1=4 AugILP: 10 g1=5 AugILP: 25 g1=6 AugILP: 25 g1=7 AugILP: 25 g1=8 AugILP: 25 g1=9 AugILP: 24 g1=10 AugILP: 24 g1=12 AugILP: 24 g1=14 AugILP: 2379 g1=16 AugILP: 233 g1=18 AugILP: 228 g1=20 AugILP: 232 g1=23 AugILP: 237 g1=26 AugILP: 207 g1=29 AugILP: 193 g1=33 AugILP: 149 g1=40 AugILP: 139 g1=45 AugILP: 119 g1=50 AugILP: 127 g1=60 AugILP: 109 g1=75 AugILP: 94 g1=90 AugILP: 77 g1=110 AugILP: 69 40000 objective g1=2 AugILP: 1 g1=3 AugILP: 1 g1=4 AugILP: 8 g1=5 AugILP: 119 g1=6 AugILP: 156 g1=7 AugILP: 195 g1=8 AugILP: 231 g1=9 AugILP: 273 g1=10 AugILP: 307 g1=12 AugILP: 386 g1=14 AugILP: 13161 g1=16 AugILP: 4449 g1=18 AugILP: 5167 g1=20 AugILP: 5580 g1=23 AugILP: 5937 g1=26 AugILP: 6013 g1=29 AugILP: 6290 g1=33 AugILP: 3747 g1=40 AugILP: 3616 g1=45 AugILP: 3210 g1=50 AugILP: 3093 g1=60 AugILP: 2383 g1=75 AugILP: 2186 g1=90 AugILP: 2122 g1=110 AugILP: 1786 40000 23:11 20000 15000 10000 5000 0 5 10 15 20 iteration 25 30 35 40 0 5 10 15 20 iteration 25 30 35 40 Figure 1 Outer loop results for Makespan Minimization when using (left to right) Γbest , Γ2-apx , Γ5-apx , and Γany ; clipped to 40 outer iterations. for Makespan Minimization) are sufficient for convergence. We remark that, in the case of Closest String, this hints at the possibility that the maximum value of any feasible augmenting step g ∈ G(A) is bounded by k O(1) rather than k O(k) , which would imply an algorithm with runtime k O(k) log L while the currently best algorithm runs in time 2 k O(k ) log L [19]. Regarding Question #2, we see that Γ2-apx provides essentially the same convergence rate as Γbest but is orders of magnitude cheaper to compute. The “any step” augmentation strategy Γany usually converges surprisingly quickly, but our results make it clear that its behavior is erratic and unpredictable. The inner loop Figure 2 reveals that a good step (close to a Graver-best step) is usually found for larger step-length γ; this motivates adding the Γ5-apx augmentation strategy to the comparison, as it spends less time on short steplengths than Γ2-apx . Figure 1 shows that using 5-approximate Graver-best steps instead of 2-approximate does not affect the outer loop convergence much, and Figure 2 shows that in terms of the total number of (AugILP) calls it performs better. Furthermore, we observe that solving (AugILP) using a MILP solver such as Gurobi typically takes essentially as much time as solving (ILP) itself; in other words, current MILP solvers are (without tuning) unable to make any use neither of the extra structure of E (N ) , nor the fact that we are seeking a solution with small `1 norm and the right hand side is 0. Moreover, with a growing number of bricks N , the time to solve (AugILP) using a MILP solver grows superlinearly, suggesting that, for large enough N , a specialized dynamic programming algorithm might be competitive with generic MILP solvers. Evaluating and Tuning n-fold Integer Programming 35000 30000 objective 25000 20000 15000 g1=2 AugILP: 1 g1=3 AugILP: 1 g1=4 AugILP: 8 g1=5 AugILP: 119 g1=6 AugILP: 156 g1=7 AugILP: 195 g1=8 AugILP: 231 g1=9 AugILP: 273 g1=10 AugILP: 307 g1=12 AugILP: 386 g1=14 AugILP: 13161 g1=16 AugILP: 4449 g1=18 AugILP: 5167 g1=20 AugILP: 5580 g1=23 AugILP: 5937 g1=26 AugILP: 6013 g1=29 AugILP: 6290 g1=33 AugILP: 3747 g1=40 AugILP: 3616 g1=45 AugILP: 3210 g1=50 AugILP: 3093 g1=60 AugILP: 2383 g1=75 AugILP: 2186 g1=90 AugILP: 2122 g1=110 AugILP: 1786 35000 30000 25000 20000 15000 10000 10000 5000 5000 20 40 60 iteration 80 100 g1=2 AugILP: 1 g1=3 AugILP: 1 g1=4 AugILP: 5 g1=5 AugILP: 15 g1=6 AugILP: 15 g1=7 AugILP: 11 g1=8 AugILP: 16 g1=9 AugILP: 16 g1=10 AugILP: 12 g1=12 AugILP: 12 g1=14 AugILP: 5304 g1=16 AugILP: 111 g1=18 AugILP: 127 g1=20 AugILP: 114 g1=23 AugILP: 129 g1=26 AugILP: 107 g1=29 AugILP: 128 g1=33 AugILP: 106 g1=40 AugILP: 88 g1=45 AugILP: 81 g1=50 AugILP: 74 g1=60 AugILP: 58 g1=75 AugILP: 50 g1=90 AugILP: 48 g1=110 AugILP: 38 40000 35000 30000 25000 20000 15000 0 20 40 iteration 60 80 100 g1=2 AugILP: 1 g1=3 AugILP: 1 g1=4 AugILP: 2 g1=5 AugILP: 4 g1=6 AugILP: 4 g1=7 AugILP: 4 g1=8 AugILP: 4 g1=9 AugILP: 6 g1=10 AugILP: 4 g1=12 AugILP: 6 g1=14 AugILP: 5765 g1=16 AugILP: 40 g1=18 AugILP: 127 g1=20 AugILP: 257 g1=23 AugILP: 96 g1=26 AugILP: 119 g1=29 AugILP: 1864 g1=33 AugILP: 99 g1=40 AugILP: 36 g1=45 AugILP: 37 g1=50 AugILP: 30 g1=60 AugILP: 80 g1=75 AugILP: 30 g1=90 AugILP: 44 g1=110 AugILP: 52 40000 35000 30000 25000 objective 0 objective g1=2 AugILP: 1 g1=3 AugILP: 1 g1=4 AugILP: 10 g1=5 AugILP: 25 g1=6 AugILP: 25 g1=7 AugILP: 25 g1=8 AugILP: 25 g1=9 AugILP: 24 g1=10 AugILP: 24 g1=12 AugILP: 24 g1=14 AugILP: 2379 g1=16 AugILP: 233 g1=18 AugILP: 228 g1=20 AugILP: 232 g1=23 AugILP: 237 g1=26 AugILP: 207 g1=29 AugILP: 193 g1=33 AugILP: 149 g1=40 AugILP: 139 g1=45 AugILP: 119 g1=50 AugILP: 127 g1=60 AugILP: 109 g1=75 AugILP: 94 g1=90 AugILP: 77 g1=110 AugILP: 69 40000 objective 40000 20000 15000 10000 10000 5000 5000 0 20 40 60 iteration 80 100 0 20 40 iteration 60 80 100 Figure 2 Inner loop results for Makespan Minimization when using (left to right) Γbest , Γ2-apx , Γ5-apx , and Γany ; clipped to 100 inner iterations. 1500 2000 1500 2000 1500 1000 1000 1000 500 500 500 0 0 5 10 15 iteration 20 25 30 0 0 5 10 15 iteration 20 25 30 g1=2 AugILP: 1 g1=3 AugILP: 11 g1=4 AugILP: 31 g1=5 AugILP: 31 g1=6 AugILP: 39 g1=7 AugILP: 37 g1=9 AugILP: 45 g1=10 AugILP: 41 g1=12 AugILP: 50 g1=14 AugILP: 44 g1=16 AugILP: 51 g1=18 AugILP: 50 g1=20 AugILP: 47 g1=23 AugILP: 41 g1=26 AugILP: 48 g1=29 AugILP: 49 g1=33 AugILP: 53 g1=40 AugILP: 51 2500 objective 2000 g1=2 AugILP: 1 g1=3 AugILP: 81 g1=4 AugILP: 176 g1=5 AugILP: 176 g1=6 AugILP: 151 g1=7 AugILP: 147 g1=9 AugILP: 115 g1=10 AugILP: 103 g1=12 AugILP: 97 g1=14 AugILP: 90 g1=16 AugILP: 81 g1=18 AugILP: 71 g1=20 AugILP: 65 g1=23 AugILP: 69 g1=26 AugILP: 60 g1=29 AugILP: 55 g1=33 AugILP: 43 g1=40 AugILP: 39 2500 objective g1=2 AugILP: 1 g1=3 AugILP: 378 g1=4 AugILP: 758 g1=5 AugILP: 860 g1=6 AugILP: 634 g1=7 AugILP: 661 g1=9 AugILP: 499 g1=10 AugILP: 515 g1=12 AugILP: 396 g1=14 AugILP: 448 g1=16 AugILP: 378 g1=18 AugILP: 387 g1=20 AugILP: 398 g1=23 AugILP: 410 g1=26 AugILP: 398 g1=29 AugILP: 428 g1=33 AugILP: 278 g1=40 AugILP: 274 2500 objective 23:12 0 0 10 20 iteration 30 40 50 Figure 3 Outer loop results for the Closest String instance when using (left to right) Γbest , Γ2-apx , and Γany . 6 Outlook We have initiated an experimental investigation of certain subclasses of ILP with block structured constraint matrices. Our results show that, as theory suggests, for such ILPs a primal algorithm always augmenting with steps of small `1 norm converges quickly. We close with a few interesting research directions. First, is there a way to tune generic MILP solvers to solve (AugILP) significantly faster than (ILP)? Second, what is the behavior of our algorithm on instances other than N -fold IP? For example, how large must be g1 in order to attain the optimum quickly for standard benchmark instances, e.g. MIPLIB [22]? Third, the approach of Koutecký et al. [23] suggests that a key property for the efficient solvability of (AugILP) is a certain “sparsity” and “shallowness” (formally captured by the graph parameter tree-depth) of graphs related to A; what are “natural” instances with small tree-depth? K. Altmanová, D. Knop, and M. Koutecký References 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 Livio Bertacco, Matteo Fischetti, and Andrea Lodi. A feasibility pump heuristic for general mixed-integer problems. Discrete Optimization, 4(1):63 – 76, 2007. Mixed Integer Programming. doi:https://doi.org/10.1016/j.disopt.2006.10.001. Timo Berthold. Measuring the impact of primal heuristics. Operations Research Letters, 41(6):611–614, 2013. Ralf Borndörfer, Martin Grötschel, and Ulrich Jäger. Planning problems in public transit. In Production Factor Mathematics, pages 95–121. Springer, 2010. Lin Chen and Daniel Marx. Covering a tree with rooted subtrees–parameterized and approximation algorithms. In Proceedings of the Twenty-Ninth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 2801–2820. SIAM, 2018. Markus Chimani, Matthias Woste, and Sebastian Böcker. A closer look at the closest string and closest substring problem. In 2011 Proceedings of the Thirteenth Workshop on Algorithm Engineering and Experiments (ALENEX), pages 13–24. SIAM, 2011. Emilie Danna, Edward Rothberg, and Claude Le Pape. Exploring relaxation induced neighborhoods to improve mip solutions. Mathematical Programming, 102(1):71–90, 2005. Jesus A. De Loera, Raymond Hemmecke, and Matthias Köppe. Algebraic and Geometric Ideas in the Theory of Discrete Optimization, volume 14 of MOS-SIAM Series on Optimization. SIAM, 2013. Friedrich Eisenbrand, Christoph Hunkenschröder, and Kim-Manuel Klein. Faster algorithms for integer programs with block structure. arXiv preprint arXiv:1802.06289, 2018. Elisabeth Finhold and Raymond Hemmecke. Lower bounds on the graver complexity of m-fold matrices. Annals of Combinatorics, 20(1):73–85, 2016. Matteo Fischetti and Andrea Lodi. Local branching. Mathematical programming, 98(13):23–47, 2003. Gurobi Optimization, Inc. Gurobi optimizer reference manual, 2016. URL: http://www. gurobi.com. Stefan Heinz, Wen-Yang Ku, and Christopher J. Beck. Recent improvements using constraint integer programming for resource allocation and scheduling. In International Conference on AI and OR Techniques in Constriant Programming for Combinatorial Optimization Problems, pages 12–27. Springer, 2013. Raymond Hemmecke. Exploiting symmetries in the computation of graver bases. arXiv preprint math/0410334, 2004. Raymond Hemmecke, Matthias Köppe, and Robert Weismantel. Graver basis and proximity techniques for block-structured separable convex integer minimization problems. Math. Program., 145(1-2, Ser. A):1–18, 2014. Raymond Hemmecke, Shmuel Onn, and Lyubov Romanchuk. n-fold integer programming in cubic time. Math. Program., 137(1-2, Ser. A):325–341, 2013. John D. Hunter. Matplotlib: A 2d graphics environment. Computing In Science & Engineering, 9(3):90–95, 2007. doi:10.1109/MCSE.2007.55. Klaus Jansen, Kim-Manuel Klein, Marten Maack, and Malin Rau. Empowering the configuration-ip-new ptas results for scheduling with setups times. arXiv preprint arXiv:1801.06460, 2018. Thomas Kluyver, Benjamin Ragan-Kelley, Fernando Pérez, Brian E Granger, Matthias Bussonnier, Jonathan Frederic, Kyle Kelley, Jessica B Hamrick, Jason Grout, Sylvain Corlay, et al. Jupyter notebooks-a publishing format for reproducible computational workflows. In ELPUB, pages 87–90, 2016. Dušan Knop, Martin Koutecký, and Matthias Mnich. Combinatorial n-fold integer programming and applications. In Proc. ESA 2017, volume 87 of Leibniz Int. Proc. Informatics, pages 54:1–54:14, 2017. 23:13 23:14 Evaluating and Tuning n-fold Integer Programming 20 21 22 23 24 25 26 27 28 29 30 Dušan Knop, Martin Koutecký, and Matthias Mnich. Voting and bribing in singleexponential time. In Proc. STACS 2017, volume 66 of Leibniz Int. Proc. Informatics, pages 46:1–46:14, 2017. Dušan Knop and Martin Koutecký. Scheduling meets n-fold integer programming. Journal of Scheduling, Nov 2017. doi:10.1007/s10951-017-0550-0. Thorsten Koch, Tobias Achterberg, Erling Andersen, Oliver Bastert, Timo Berthold, Robert E Bixby, Emilie Danna, Gerald Gamrath, Ambros M Gleixner, Stefan Heinz, et al. Miplib 2010. Mathematical Programming Computation, 3(2):103, 2011. Martin Koutecký, Asaf Levin, and Shmuel Onn. A parameterized strongly polynomial algorithm for block structured integer programs. arXiv preprint arXiv:1802.05859, 2018. Andrea Lodi. Mixed integer programming computation. In 50 Years of Integer Programming 1958-2008, pages 619–645. Springer, 2010. Shmuel Onn. Nonlinear discrete optimization. Zurich Lectures in Advanced Mathematics, European Mathematical Society, 2010. David Pisinger and Stefan Ropke. Large neighborhood search. In Handbook of metaheuristics, pages 399–419. Springer, 2010. Yves Pochet and Laurence A Wolsey. Production planning by mixed integer programming. Springer Science & Business Media, 2006. Matthew J. Saltzman. Coin-or: an open-source library for optimization. In Programming languages and systems in computational economics and finance, pages 3–32. Springer, 2002. Tommi Sottinen. Operations research with gnu linear programming kit. ORMS, 1020:200, 2009. The Sage Developers. SageMath, the Sage Mathematics Software System (Version 7.6), 2017. http://www.sagemath.org. 

8cs.DS

arXiv:1711.05869v1 [stat.ML] 16 Nov 2017 Predictive Independence Testing, Predictive Conditional Independence Testing, and Predictive Graphical Modelling Samuel Burkart∗ and Franz J. Király† Department of Statistical Science, University College London, Gower Street, London WC1E 6BT, United Kingdom November 17, 2017 Abstract Testing (conditional) independence of multivariate random variables is a task central to statistical inference and modelling in general - though unfortunately one for which to date there does not exist a practicable workflow. State-of-art workflows suffer from the need for heuristic or subjective manual choices, high computational complexity, or strong parametric assumptions. We address these problems by establishing a theoretical link between multivariate/conditional independence testing, and model comparison in the multivariate predictive modelling aka supervised learning task. This link allows advances in the extensively studied supervised learning workflow to be directly transferred to independence testing workflows - including automated tuning of machine learning type which addresses the need for a heuristic choice, the ability to quantitatively trade-off computational demand with accuracy, and the modern black-box philosophy for checking and interfacing. As a practical implementation of this link between the two workflows, we present a python package ’pcit’, which implements our novel multivariate and conditional independence tests, interfacing the supervised learning API of the scikit-learn package. Theory and package also allow for straightforward independence test based learning of graphical model structure. We empirically show that our proposed predictive independence test outperform or are on par to current practice, and the derived graphical model structure learning algorithms asymptotically recover the ’true’ graph. This paper, and the ’pcit’ package accompanying it, thus provide powerful, scalable, generalizable, and easy-to-use methods for multivariate and conditional independence testing, as well as for graphical model structure learning. ∗ samuel.burkart.16@ucl.ac.uk † f.kiraly@ucl.ac.uk 1 Contents 1 Introduction 1.1 Setting: testing independence . . . 1.2 Predictive independence testing . . 1.3 Graphical model structure learning 1.4 Principal contributions . . . . . . . 1.5 Paper overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 Statistical independence testing 2.1 About (in)dependence . . . . . . . . . . . 2.2 Statistical independence testing . . . . . . 2.3 The state-of-art in advanced independence 2.4 Issues in current methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 4 4 5 5 5 . . . . . . . . . . testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 7 7 8 10 . . . . . . . . . . . . . . . . . . . . 3 Predictive independence testing 3.1 Mathematical setting . . . . . . . . . . . . . 3.2 Elicitation by convex losses . . . . . . . . . 3.3 Predictive uninformedness . . . . . . . . . . 3.4 Statistical dependence equals predictability 3.5 Conditional independence . . . . . . . . . . 3.6 Testing prediction error against baseline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 11 12 13 14 18 19 4 Graphical Models 4.1 Informal definition of graphical models . . . 4.2 Types of graphical models . . . . . . . . . . 4.2.1 Bayesian Networks . . . . . . . . . . 4.2.2 Markov Networks . . . . . . . . . . . 4.3 Graphical model structure learning . . . . . 4.3.1 Score-based methods . . . . . . . . . 4.3.2 Independence testing based methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 22 22 22 24 25 25 26 5 Predictive inference algorithms 5.1 Predictive conditional independence testing . . . . . . . . . 5.1.1 False-discovery rate control . . . . . . . . . . . . . . 5.1.2 Improving the prediction functionals . . . . . . . . . 5.1.3 Supervised learning for independence testing . . . . 5.2 Predictive structure learning of undirected graphical models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 27 28 28 29 29 6 pcit package 6.1 Overview . . . . . . . . . . . . . . . . 6.1.1 Use cases . . . . . . . . . . . . 6.1.2 Dependencies . . . . . . . . . . 6.2 API description . . . . . . . . . . . . . 6.3 Function signatures . . . . . . . . . . . 6.4 API design . . . . . . . . . . . . . . . 6.4.1 Sklearn interface . . . . . . . . 6.4.2 Wrapper for Sklearn estimators 6.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 31 31 31 31 32 34 34 34 35 . . . . . . . . . . . . . . . . . . 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 Experiments 7.1 Performance tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.1.1 Performance of conditional independence test . . . . . . . . . . . . 7.1.2 Performance of structure learning algorithm: Error rates . . . . . . 7.1.3 Performance of structure learning algorithm: Variance of estimated 7.2 Experiments on real data sets . . . . . . . . . . . . . . . . . . . . . . . . . 7.2.1 Sklearn data sets: Boston Housing and Iris . . . . . . . . . . . . . 7.2.2 Key short-term economic indicators (UK) . . . . . . . . . . . . . . . . . . . . . . . . . . model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 36 36 37 40 41 41 41 8 Discussion 43 Appendices 45 A Best uninformed predictors: classification A.1 Misclassification loss is a probabilistic loss . . . . . . . . . . . . . . . . . . . . . . . . . A.2 Logarithmic loss is a proper loss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.3 Brier loss is a proper loss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 45 46 46 B Elicited statistics for regression losses B.1 Squared loss elicits the mean . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B.2 Quantile loss elicits the quantile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 47 48 3 1. Introduction 1.1. Setting: testing independence The study of dependence is at the heart of any type of statistical analysis, and independence testing is an important step in many scientific investigations, be it to determine if two things are related, or to assess if an intervention had the desired effect, such as: • When conducting market research, one might be interested in questions such as “Are our advertisement expenditures independent of our profits?” (hopefully not), or the more sophisticated version “conditional on the state of the market, are advertisement expenditures independent of our profits?”, which, if found to be true, would mean we are unlikely to increase profits through an increase in our advertising budget (subject to the usual issue that inferring causality from data requires an intervention or instrument). • When collecting data for a medical study on the occurrence of an outcome Y , one might ask “In the presence of data about attributes A for the subjects, should we still collect data for attributes B ”. If Y is independent of attributes B given attributes A, additionally collecting information about attributes B will not improve the knowledge of the state of Y (subject to the usual issue that this conclusion is valid only for patients sampled in the same way from the same population). The difficulty of the independence testing task crucially relies on whether the following two complications present: • Multivariate independence testing. This concerns the type of values which the involved variables, in the second example the values which attributes A and attributes B may take: if the domain of possible values consists either of a single number (continuous variable), or one class out of many (categorical variable), the hypothesis test is “univariate”, and powerful methodology exists that deals well with most scenarios, subject to some constraints. Otherwise, we are in the “multivariate hypothesis testing” setting. • Conditional independence testing. Whether there are conditioning random variables which are to be controlled for, in the sense of testing independence conditional on a possible third attributes C. If so, we are in the “conditional hypothesis testing” setting. For the task which is neither multivariate nor conditional, well-recognized and universally applicable hypothesis tests (such as the t-test or chi-squared test) are classically known. The multivariate setting and the conditional setting are less well studied, and are lacking approaches which are general and universally accepted, due to difficulties in finding a suitable approach which comes with theoretical guarantees and is free of strong model assumptions. The setting which is both multivariate and conditional is barely studied. The three major state-of-the-art approaches are density estimation, copula and kernel based methods. Most instances are constrained to specific cases or rely on subjective choices that are difficult to validate on real-world data. A more detailed overview of the state-of-art and background literature is given in Section 2. 1.2. Predictive independence testing The methodology outlined in this paper will consider multivariate and conditional independence testing from a new angle. The underlying idea for the test is that if two random variables X, Y are independent, it is impossible to predict Y from X - in fact it will be shown that these two properties are equivalent (in a certain quantitative sense). The same applies to conditional independence tests: two random variables X, Y are conditionally independent given Z, if adding X as predictor variable 4 above Z will not improve the prediction of Y . In both cases, the predictive hypothesis test takes the form of comparing a good prediction strategy against an optimal baseline strategy via a predictive loss function. By determining if losses stemming from two predictions are significantly different, one can then test statistically if a variable adds to the prediction of another variable (potentially in the presence of a conditioning set), which by the equivalence is a test for (conditional) independence. 1.3. Graphical model structure learning Probabilistic graphical models are a concept that heavily relies on independence statements for learning and inference. Most structure learning algorithms to date are, as a result of the lack of a scalable conditional independence tests and additional combinatorial issues, constraint-based, or make heavy assumptions on underlying distributions of a sample. This paper will leverage the predictive independence test into a new routine to estimate the undirected graph for the distribution underlying a sample, based on conditional independence testing, allowing it to make only weak assumptions on the underlying distribution. 1.4. Principal contributions The new approach to multivariate and conditional independence testing outlined in this paper improves concurrent methodology by deriving an algorithm that • features a principled model selection algorithm for independence testing by linking the field of independence testing to the field of predictive modelling, thus filling a gap in state-of-the-art methodology, • additionally allowing independence testing to directly benefit from the well-understood and efficiently implemented theory for model selection and parameter tuning in predictive modelling, • is comparatively fast and scalable on a wide variety of problems and • deals with the multivariate and conditional independence testing task in a straightforward manner. Additionally, an algorithm leveraging the newly derived test into a scalable independence testingbased graphical model structure learning algorithm is outlined, which overcomes issues in the field by offering a test for undirected graph structure learning that • has polynomial time complexity, O(p2 ) in the number of variables, p, • is exact and • offers stringent methodology to control the number of type 1 errors in the estimated graph. 1.5. Paper overview Section 2 will provide an overview of the most important tasks in and approaches to statistical independence testing and outline the issues in current methodology. Section 3 will then propose and derive a novel approach to independence testing, the predictive conditional independence test (PCIT). After, section 4 will introduce the concept of a graphical model and survey the current structure learning algorithms. Section 5 will then state the relevant algorithms of the Python implementation, which is outlined in section 6. Section 7 provides performance statistics and examples for the conditional independence test as well as the graphical model structure learning algorithm. Lastly, section 8 will describe the advantages of using the PCIT for independence testing and outline drawbacks and directions for further research. 5 Authors’ contributions This manuscript is based on SB’s MSc thesis, submitted September 2017 at University College London and written under supervision of FK, as well as on an unpublished manuscript of FK which relates predictive model selection to statistical independence. The present manuscript is a substantial reworking of the thesis manuscript, jointly done by SB and FK. FK provided the ideas for the independence tests (paper) in section 3 and the usage of them for graphical models, SB and FK jointly conceived the ideas for the graphical model structure learning algorithm. Literature overview is due to SB with helpful pointers by FK. Python package and experiments are written and conducted by SB, partially based on discussions with FK. Acknowledgements We thank Fredrik Hallgren and Harald Oberhauser for helpful discussions. FH has briefly worked on a thesis about the same topic under the supervision of FK, before switching to a different topic. While FH did, to our knowledge, not himself make any of the contributions found in this paper, discussions with him insipred a few of them. HO pointed FK in generic discussions about loss functions towards some prior results (elicitation of median, and elicitation as a concept as defined by Gneiting and Raftery [17]) which helped making some of the proven statements more precise (such as the Q-losses being faithful for the univariate real case). 6 2. Statistical independence testing 2.1. About (in)dependence Statistical (in)dependence is a property of a set of random variables central to statistical inference. Intuitively, if a set of random variables are statistically independent, knowing the value of some will not help in inferring the values of any of the others. Mathematically: let X, Y and Z be random variables taking values in X, Y and Z respectively. As usual, we denote for A ∈ X by P (X ∈ A) the probability of X taking a value in A, and for A ∈ X, C ∈ Z by P (X ∈ A|Z ∈ C) := P (X ∈ A, Z ∈ C)/P (Z ∈ C) be the conditional probability of X taking a value in A when Z is known/observed to have taken a value in C. Definition 2.1. X and Y are called marginally independent (of each other) if for all A ⊆ X and B ⊆ Y (where the below expression is defined) it holds that P (X ∈ A, Y ∈ B) = P (X ∈ A)P (Y ∈ B). This formulation allows for X and Y to be defined over sets of random variables that are a mixture of continuous and discrete, as well as being univariate or multivariate (thus implicitly covering the case of multiple univariate random variables as well). Definition 2.2. X and Y are called conditionally independent (of each other) given Z if for all A ⊆ X, B ⊆ Y, and C ⊆ Z (where the below expression is defined) it holds that P (X ∈ A, Y ∈ B|Z ∈ C) = P (X ∈ A|Z ∈ C)P (Y ∈ B|Z ∈ C). For absolutely continuous or discrete X and Y , Definition 2.1 straightforwardly implies that marginal independence is equivalent to the joint distribution or mass function factorizing, i.e., it equals the product of the marginal distributions’ probability or mass function. The analogue result also shows for the conditional case. We would like to note that the mathematical definition independence is symmetric, i.e., having the property from Definition 2.1 is unaffected by interchanging X and Y . In contrast, the intuitive motivation we gave at the beginning however, namely that knowing the value of X does not yield any additional restriction regarding the value of Y , is non-symmetric. This non-symmetric statement of statistical independence is commonly phrased mathematically as P (Y ∈ B|X ∈ A) = P (Y ∈ B) which is directly implied by Definition 2.1 (and that of the conditional). This non-symmetric characterization of statistical independence is also morally at the heart of the predictive characterization which we later give in the supervised learning setting: Y cannot be predicted from X, phrased in a quantitative way that connects checking of statistical independence to estimation of a supervised generalization error difference which eventually allows for a quantitatve testing of the hypothesis of independence. 2.2. Statistical independence testing Testing whether two random variables are (conditionally) independent is one of the most common tasks in practical statistics (perhaps the most common one after summarization) and one of the most important topics in the theoretical study of statistical inference (perhaps the most important one). The most frequently found types of quantitative certificates for statistical independence are phrased in the (frequentist) Neyman-Pearson hypothesis testing paradigm. In terms of the methodological/mathematical idea, the main distinction of independence tests is into parametric (where the data is assumed to stem from a certain model type and/or distribution) and non-parametric (”distribution-free”) tests. The hypothesis tested usually takes one of the three following forms: 7 |= Marginal independence, X Y Conditional independence, X |= Marginal independence of two random variables is the topos of much of classical statistics. Well known classical test statistics for the continuous univariate case are the Pearson correlation coefficient, Kendall’s τ and Spearman’s ρ. For discrete variables, Pearson’s χ-squared statistic can be used. From theoretical results on these statistics’ asymptotic or exact distribution, univariate independence tests may be derived. For X being discrete and Y being continuous, t-family tests or Wilcoxon family tests may be used. Testing with at least one of X or Y being multivariate is much more difficult and less standard, current methodology will be outlined in the next section. Y |Z Testing for conditional independence is an inherently more difficult problem than marginal independence [7]. Hence, few viable options exist to date to test if two sets of variables are conditionally independent given a conditioning set. The most common instances are either based on parametric model assumptions (e.g., linear) or on binning Z and X, then comparing if the distribution of Y conditioned on Z changes, when additionally conditioning on X. Strategies for of conditional independence testing may demand much larger sample sizes than marginal independence tests due to the explicitly modelling sub-domains of the values which Z takes. d Equality of distribution of two unpaired samples, X1 = X2 , d where = indicates equality in underlying distribution. An unpaired two-sample test for equality of distribution tests whether two unpaired i.i.d. samples from X1 and X2 are sampled from the same underlying distributions. The connection to independence testing is not obvious but may be established by pairing each draw from X1 or X2 with a draw from a variable Y taking values in {1, 2}, indicating whether the draw was from X1 or X2 , i.e., taking value i if the draw was from Xi . 2.3. The state-of-art in advanced independence testing The task of independence testing (in various settings) has been tackled from many different angles. As soon as one considers observations that are multivariate, or the conditional task, there is not one universally agreed-upon method, but many different approaches. The most prominent ideas used for multivariate and/or conditional independent testing will be presented in this section together with their advantages and shortcomings. Density estimation The classical approach. In case of existence, the joint probability density function contains all the necessary information about independence structures for a set of random variables in the multivariate case. While many univariate tests are based on or may be interpreted as being based on some sort of density estimation, for the multivariate and conditional case, density estimation is a difficult problem. One example is shown in [12], where density estimation-based information-theoretical measures are used to conduct multivariate marginal independence tests. Copulas An approach that is widely used in finance and risk management. Copulas are multivariate probability distributions with uniform marginals. To use copulas for independence tests, one transforms the marginals of a multivariate distribution into uniform distributions, and the resulting copula contains all information about the independence structure. A simple example using the empirical cdf can be found in [16]. Copulas have mainly been used for marginal multivariate independence tests, such as in 8 [34]. The task of testing for conditional independence has been attempted through the use of partial copulas [6], however strong assumptions (linearity) are made on the type of relationship between the variables that are to be tested for independence and the conditioning set. Two-sample testing has largely remain unaddressed, since its application in the financial sector are less relevant. [33] describes a two-sample test for the estimated copulas, and hence, the independence structures, but the test does not extend to a comparison of the joint distributions as a whole. Kernel methods |= A relatively recent approach using optimization techniques in reproducing kernel Hilbert spaces to answer independence queries based on various test-statistics. Kernel methods were first used for marginal multivariate independence testing via a quantitative measure of dependence, the HilbertSchmidt Independence Criterion (HSIC), which is defined as the squared Hilbert Schmidt norm of the cross-covariance operator (the covariance between two difference spaces), see [19]. [20] further expands on the theoretical foundations of the HSIC’s distribution under the null, resulting in a hypothesis test for the independence of two variables X and Y using the HSIC criterion. [18] outlines a non-parametric test for independence, and demonstrates its performance on samples of random variables with up to three dimensions. The conditional case is tackled in [37], which derives a test statistic based on conditional crosscovariance operators, for which the asymptotic distribution under the null-hypothesis X Y |Z is derived. They state that the manual choice of kernel can affect type II and, more importantly, type I errors, especially when the dimensionality of the conditioning set is large. Additionally, they found that the performance of their method decreases in the dimensions of the conditioning set, constraining the set of problems for which the test is viable for. As for two-sample testing, [3] derive a test based on the Euclidean inter-point distances between the two samples. A different approach to the same test statistics with the additional use of characteristic functions was made by [13]. While the difficulty of density estimation is reduced when using the empirical distributions, the data requirements tend to be much larger, while additionally imposing a sometimes taxing computational complexity. Both of these methods can be related to kernel functions, which was picked up in [22], which proposed kernel-based two-sample test by deriving three multivariate tests for assessing the null-hypothesis that the distributions p and q generating two samples X ∼ p, Y ∼ q are equal, p = q. The criterion introduced is the “Maximum Mean Discrepancy” (MMD) between p and q over a function space F. M M D = sup (EX∼p [f (X)] − EY ∼q [f (Y )]) f ∈F That is, the maximum difference between the expected function values with respect to the function space (in their case, a Reproducing Kernel Hilbert Space). Predictive independence testing: precursors There are a few instances in which independence testing has already been approached from the perspective of predictability: most prominently, Lopez-Paz and Oquab [28] present a two-sample test (on equality of distribution) based on abstract binary classifiers, e.g., random forests. The presented rationale is in parts heuristic and specific to the two-sample setting. Their main idea is as follows: if the null-hypothesis of the two samples being drawn from the same distribution is true, no classifier assigning to a data point the distribution from it was drawn should fare significantly better than chance, on an unseen test set. The work of Lopez-Paz and Oquab [28] may be seen as a special case of our predictive independence test presented later, for Y taking values in a two-element set, and the loss being the misclassification loss. It is probably the first instance in which the abstract inability to predict is explicitly related to independence, albeit somewhat heuristically, and via the detour of two-sample testing for equality in 9 distribution. In earlier work, Sriperumbudur et al. [35] already relate the kernel two-sample test to a specific classifier, the Parzen window classifier, evaluated by the misclassification loss. Thus this ideas of Sriperumbudur et al. [35] may in turn be seen as a precursor of the later work of Lopez-Paz and Oquab [28] which abstracts the idea that any classifier could be used to certify for equality of the two samples. 2.4. Issues in current methodology Density estimation While the probability density function is in some sense optimal for measuring dependence, since it contains all the information about the random variable, its estimation is a difficult task, which requires either strong assumptions or large amounts of data (or both). Furthermore, due to the curse of dimensionality, density estimation for 3 or more dimensions/variables is usually practically intractable, hence its practical usefulness is often limited to testing pairwise independence rather than testing for full (mutual) independence. Copulas Leaving aside issues arising by practitioners misunderstanding the method (which had a strong contribution to the 2007/2008 financial crisis, but no bearance whatsoever on the validity of the method when applied correctly), copula-based independence testing is largely a heuristics driven field, requiring many subjective manual choices for estimation and testing. Above all, copula methods require a user to subjectively choose an appropriate copula from a variety of options, such as for example the survival copula, the multivariate Gaussian or the multivariate Student’s t-copula [9]. Additionally, two-sample testing is to date largely unaddressed in the field. Kernels As for the copula-based methods, kernels require many subjective and manual choices, such as the choice of kernel function, and its hyper-parameters. While in a theoretical setting, these choices are made by using the ground truth (and generating artificial data from it), in practice it is difficult to tune the parameters and make statements about the confidence in results. Additionally, the cost of obtaining the test statistic and its asymptotic distribution may be high for many of the proposed approaches. There are attempts at resolving these issues, [36] and [23] outline heuristics and strategies to minimize the heuristic choices, and [21] and [10] propose more computationally efficient strategies. While all methods have their merits, and there is vast research on specific problems, providing specific solutions to the individual challenges, what is missing is an approach that is scalable and powerful not just for specific cases, but for the general case, that automatically finds good solutions to real problems, where the ground truth is not known and cannot be used for tuning. 10 3. Predictive independence testing This section will explore the relationship between predictability and dependence in a set of variables. It will be shown that one can use supervised learning methodology to conduct marginal and conditional independence tests. This distinction between marginal and conditional dependence is not theoretically necessary (since the marginal case can be achieved by setting the conditioning set to the empty set), but is made to highlight specific properties of the two approaches. First, equivalence of independence statements and a specific supervised learning scenario will be shown. After, a routine leveraging this equivalence to test for conditional independence will be proposed. 3.1. Mathematical setting The independence tests will be based on model comparison of supervised learning routines. Supervised learning is the task where given i.i.d. samples (X1 , Y1 ), . . . , (XN , YN ) ∼ (X, Y ) taking values in X × Y, to find a prediction functional f : X → Y such that f (Xi ) well approximates a target/label variables Yi , where ”well” is defined with respect to a loss function L which is to be minimized in expectation. Definition 3.1. A (point prediction) loss functional is an element of [Y×Y → R], i.e., a function with range Y × Y and image R. By convention, the first argument will be considered the proposed/predicted value, the second the true value to compare to. By convention, a lower loss will be considered more beneficial. A loss functional L : Y × Y → R is called: (i) convex if L(., y) is lower bounded and E [L(Y, y)] ≤ L (E[Y ], y) for any Y-valued random variable Y and any y ∈ Y, and (ii) strictly convex if L(., y) is lower bounded and E [L(Y, y)] L (E[Y ], y) for any non-constant Y-valued random variable Y and any y ∈ Y. More formally, a good prediction functional possesses a small expected generalization error L (f ) := E[L(f (X), Y )]. In usual practice and in our setting f is estimated from training data D = {(Xi , Yi )}N i=1 , hence it is in fact a random object. The generalization error may be estimated from test data T = {(Xi∗ , Yi∗ )}M i=1 ∗ ∗ where (X1∗ , Y1∗ ), . . . , (XM , YM ) ∼ (X, Y ), which we assume i.i.d. and independent of the training data D and f , as b L (f ) = M X [L(f (Xi∗ ), Yi∗ )]. i=1 Independence of training data D and test data T is required to avoid bias of the generalization error estimate (“over-fitting”). In the following sections, when estimating the generalization error, it will be assumed that f is independent of the test data set, and we condition on D, hence treat f as fixed (i.e., conditional on D). In the following, we assume that Y is a vector space, and Y ⊆ Rq , which naturally includes the setting of supervised regression. However, it also includes the setting of supervised classification through a special choice of loss: 11 Remark 3.2. Deterministic classification comprises the binary case of Y = {−1, 1}, or more generally Y being finite. In neither case are additions or expectations defined on Y, as Y is just a discrete set. Supervised classification algorithms in the deterministic case are asked to produce a prediction functional f : X → Y, and are usually evaluated by the misclassification loss or 0/1-loss L : Y × Y → R : (y, y∗ ) 7→ 1[y 6= y∗ ], where 1[y 6= y∗ ] is the indicator function which evaluates to 0 if y = y∗ and to 1 otherwise. Hence, Definition 3.1 of convexity does not directly apply to the misclassification loss as expectations are not defined on Y. However, by allowing all algorithms to make predictions which are Y-valued random variables (instead of constants), one finds that E[L(Y, y)] = L0 (p, y) where L0 : (pY , y) 7→ 1 − pY (y) and pY is the probability mass function of Y . Identifying Y-valued random variables with their probability mass functions, one may replace (i) Y by the corresponding subset Y0 of R#Y−1 which is the set of probability vectors (the so-called probability simplex). For example, Y = {−1, 1} would be replaced by [0, 1]. (ii) the misclassification loss by the probabilistic, convex (but not strictly convex) loss L0 : Y0 ×Y0 → R, where the observations in the second Y0 argument are always pmf describing constant random variables. A further elementary calculation (see appendix A.1 for an explicit derivation) shows that L0 is always minimized by making deterministic predictions: argmin E[L0 (p, Y )] ∩ [Y → {0, 1}] is non-empty, p∈Y0 i.e., the L0 -best classifier may always be chosen to be a deterministic one, i.e., one that always predicts a probability mass functions with probabilities 0 or 1. This exhibits deterministic classification as a special case of probabilistic classification with a special choice of loss function. 3.2. Elicitation by convex losses Loss functionals are canonically paired with summary statistics of distributions: Definition 3.3. Let L : Y × Y → R be a convex loss functional. For a Y-valued random variable Y , we define µL ([Y ]) := argminE [L(y, Y )] y∈Y where [Y ] denotes Y as a full object (i.e., a measurable function), rather than its value. Following Gneiting and Raftery [17], we term the functional which maps Y-valued distributions to sub-sets of Y the eliciting functional associated to L. We call µL ([Y ]) the summary of Y elicited by L. Note that well-definedness, i.e., existence of the minimizer, is ensured by convexity of L (and the implied continuity). If L is strictly convex, there is furthermore a unique minimizer, in which case we will exchangeably consider µL to be a functional with target Y. Well-known examples of elicited summaries are given in the following: Lemma 3.4. The following relations between losses and elicited statistics of real-valued random variables hold: 12 (i) the (strictly convex) squared loss L : (y, y∗ ) 7→ (y−y∗ )2 elicits the mean. That is, µL ([Y ]) = E[Y ] for any Rn -valued random variable Y . (ii) the (convex but not strictly convex) absolute loss L : (y, y∗ ) 7→ |y − y∗ | elicits the median(s). That is, µL ([Y ]) = median[Y ] for any R-valued random variable Y . (iii) the (convex but not strictly convex) quantile-loss (or short: Q-loss) L(y, y∗ ) = α · m(y∗ , y) + (1 − α) · m(y, y∗ ), with m(x, z) = min(x − z, 0), elicits the α-quantile(s). That is, µL ([Y ]) = FY−1 (α) for any R-valued random variable Y , where FY−1 : [0, 1] → P (R) is the set-valued inverse c.d.f. of Y (with the convention that the full set of inverse values is returned at jump discontinuities rather than just an extremum). Proof. (i) after substitution of definition, the claim is equivalent to the statement to the mean being the minimizer of squared distances. A more explicit proof is given in Appendix B.1. (ii) follows, by setting α = 12 , from (iii). (iii) This is carried out in Appendix B.2. In the supervised setting, the best possible prediction functional can be exactly characterized in terms of elicitation: Proposition 3.5. Let L be a (strictly) convex loss. Then, it holds that argmin εL (f ) = [x 7→ µL [Y |X = x]] . f ∈[X→Y] That is, the best possible prediction as measured by L is predicting the statistic which L elicits from the conditional random variable Y |X = x. Proof. The prediction functional ω : x 7→ µL [Y |X = x] is well-defined, hence it suffices to show that εL (f ) ≥ εL (ω) for any prediction functional f : X → Y. Now by definition of µL , it holds that E [L(ω(X), Y )|X] ≤ E [L(f (X), Y )|X] . Taking total expectations yields the claim. Intuitively, the best prediction functional, as measured by a convex loss L, always predicts the statistic elicited by L from the conditional law [Y |X = x]. 3.3. Predictive uninformedness We will now introduce the notion of an uninformed baseline which will act as a point of comparison. Definition 3.6. A prediction functional f : X → Y is called uninformed if it is a constant functional, i.e., if f (x) = f (y) for all x, y ∈ X. We write uα for the uninformed prediction functional uα : x 7→ α. We will show that, for a given loss function, there is one single uninformed baseline that is optimal. Lemma 3.7. Let L be a (strictly) convex loss, let µ := µL ([Y ]) be the/a statistic elicited by L (see Definition 3.3). Then, the following quantities are equal: (i) inf{εL (f ) : f is an uninformed prediction functional} (ii) εL (uµ ) That is, uµ is achieves the lowest possible (L-)loss amongst uninformed prediction functionals and prediction strategies. Proof. Note that ε(uα ) = E[L(α, Y )]. It follows hence by definition of µ that ε(uµ ) ≤ ε(uα ) for any (constant) α ∈ Y. I.e., uµ is the best uninformed prediction functional. 13 Lemma 3.7 motivates the definition of the best uninformed predictor: Definition 3.8. We call uµ , as defined in Lemma 3.7, the (L-)best uninformed predictor (even though it may not be unique, the choice, when possible, will not matter in what follows). We call a prediction functional (L−)better-than-uninformed if its expected generalization loss is strictly smaller than of the (L−)best uninformed predictor. More formally, a prediction functional f is Lbetter-than-uninformed if εL (f ) ≤ εL (uµ ). For convenience, we further introduce some mathematical notation for best predictors: Notation 3.9. Let L be a (strictly) convex loss. We will write: (L) (i) ωY (ii) (L) (L) ωY |X := [x 7→ µL [Y ]] for the/a (L-)best uninformed predictor as defined in Definition 3.8. := [x 7→ µL [Y |X = x]] for the/a (L-)best predictor as considered in Proposition 3.5. (L) ωY and ωY |X are unique when L is strictly convex, as per the discussion after Definition 3.3. When multiple choices are possible for the minimizer, i.e., if L is convex but not strictly convex, an arbitrary choice may be made (not affecting subsequent discussion). The superscript L may be omitted in situations where the loss is clear from the context. An important fact which we will use in testing is that if a better-than-uninformed prediction functional exists, then ωY |X is an example: Proposition 3.10. Fix a (strictly) convex loss. The following are equivalent: (i) ωY |X is L-better-than-uninformed. (ii) ωY |X is not L-uninformed. (iii) There is an L-better-than-uninformed prediction functional. Note that equivalence of (i) and (ii) is not trivial, since there are prediction functionals which are not better-than-uninformed but not uninformed. Proof. The equivalence of (i) and (iii) follows directly from Lemma 3.7, and noting that (i) implies (iii).
(i)⇔(ii): By Proposition 3.5, ε ωY |X ≤ ε(f ) for any f , in particular also any uninformed f . By Lemma 3.7 and the above, the inequality is strict if and only if ωY |X is better-than-uninformed. 3.4. Statistical dependence equals predictability We continue with a result that relates - or more precisely, equates - better-than-uninformed predictability with statistical dependence. As such, it shows that the choice of constant functions as a proxy for uninformedness was canonical. We start with the more intuitive direction of the claimed equivalence: Proposition 3.11. Let X, Y be random variables taking values in X, Y. Let L be any convex loss functional. If X, Y are statistically independent, then: There exists no L-better-than-uninformed prediction functional for predicting Y from X. More precisely, there is no prediction functional f : X → Y and no convex loss L : Y × Y → R such that f is L-better-than-uninformed. Proof. Assume X, Y are statistically independent. Let L be a convex loss function, let f : X → Y be a prediction functional. Then, by convexity of L, E [L(f (X), Y )|Y ] ≥ L(E[f (X)], Y )|Y. 14 Since X is independent of Y (by assumption) and f is not random, it holds that E[f (X)]|Y = E[f (X)], i.e., E[f (X)]|Y = y, as a function of y, is constant. Writing ν := E[f (X)], we hence obtain that L(E[f (X)], Y )|Y = L(ν, Y ) = L(uν , Y ). After taking total expectations, it hence holds by the law of total expectation that E [L(f (X), Y )] ≥ E [L(uν , Y )] , meaning that f is not better-than-uninformed w.r.t. L. Since f and L were arbitrary, the statement holds. Proposition 3.11 states that independence implies unpredictability (as measured per expected loss). A natural question to ask is whether the converse holds, or more generally which converses exactly, since the loss functional L in (ii) and (iii) of Proposition 3.10, was arbitrary, and independence in (i) is a statement which remains unchanged when exchanging X and Y , while predictability is not symmetric in this respect. Hence the weakest possible converse would require unpredictability w.r.t. all convex losses and w.r.t. either direction of prediction, however much stronger converses may be shown. Before stating the mathematically more abstract general result for the converse direction, we first separately present special cases for the three important sub-cases, namely deterministic classification, probabilistic classification, and regression. Theorem 1. As in our setting, consider two random variables X, Y , taking values in X and Y. Assume Y = {−1, 1}. The following are equivalent: (i) X and Y are statistically independent. (ii) There exists no better-than-uninformed prediction functional predicting Y from X. More precisely, there is no prediction functional f : X → Y such that f is L-better-than-uninformed for the misclassification loss L : (p, y) 7→ 1 − p(y). Regarding the specific loss L, see Remark 3.2 for the identification of a class prediction with a probability-1 probabilistic prediction. Proof. By Proposition 3.11, the only direction which remains to be proven is (ii)⇒(i): (i)⇒(ii) follows directly from substituting the specific L into the implication between statements with the same numbers in Proposition 3.11. (ii)⇒ (i): We prove this by contraposition: we assume X and Y are statistically dependent and construct a better-than-uninformed prediction functional. By the equivalence established in Remark 3.2, the best uninformed predictor ωY is predicting the most probable class, w.l.o.g. 1 ∈ Y, and its expected generalization loss is one minus its generative frequency ε(ωY ) = P (Y = −1). Since X and Y are statistically dependent, there is a positive probability X0 ⊆ X (measurable with positive probability measure) such that P (Y = 1|X ∈ X0 ) ≥ P (Y = −1|X ∈ X0 ) (the definition yields 6=, but w.l.o.g. replacing X0 by its complement yields ≥). An elementary computation shows that the prediction functional f : X → Y : 2 · 1[x ∈ X0 ] − 1 is better-than-uninformed. The proof of Theorem 1 shows that for a converse, one does not necessarily need to consider predictability of X from Y as well. However, the deterministic binary classification case is somewhat special, since the misclassification loss is insufficient in the multi-class case, and in general a single loss function will be unable to certify for independence. In order to formulate these negative results, we define shorthand terminology for stating correctness of the converse. 15 Definition 3.12. Fix a (label) domain Y, let L be a set with elements being convex loss functionals in [Y × Y → R]. We call L faithful for Y if for every statistically dependent pair of random variables X, Y taking values in X, Y, there is L ∈ L and a prediction functional f : X → Y such that f is L-better-than-uninformed. If L is a one-element-set, we call its single element faithful for Y. If L is faithful for Y, and L is endowed with a measure µ such that no µ-strict sub-set of L is faithful for Y, then we call L (µ-)strictly faithful. If Y is canonically associated with a prediction task (such as classification for finite Y or classprobabilities Y, regression for continuous Y), the reference to Y may be replaced by a reference to that task. In this terminology, Theorem 1 states that the misclassification loss is faithful for the label domain Y = {−1, 1}, or equivalently that the misclassification loss is faithful for deterministic binary classification (and strictly faithful, since any set smaller than an one-element set, as by the counting measure, is empty and by definition not faithful). We can now state some negative and positive results on deterministic multi-class classification and probabilistic classification: Proposition 3.13. (i) The misclassification loss L : (p, y) 7→ 1 − p(y) is not faithful for deterministic multi-class classification, i.e., for Y being finite and containing 3 or more elements. (ii) The log-loss L : (p, y) 7→ − log(p(y)) is (strictly) faithful for probabilistic classification. (iii) The squared probabilistic classification loss L : (p, y) 7→ (1 − p(y))2 is (strictly) faithful for probabilistic classification. P (iv) The Brier loss L : (p, y) 7→ (1 − p(y))2 + y0 6=y p(y 0 )2 is (strictly) faithful for probabilistic classification. Proof. (i): It suffices to construct a counterexample for each possible Y, i.e., every finite Y with 3 or more elements. Let X = {−1, 1}, and let y ∈ Y be arbitrary. Define X, Y such that P (Y = y|X = 1) = P (Y = y|X = −1) := 0.9 and P (Y = y 0 |X = 1) 6= P (Y = y 0 |X = −1) for some class y 0 6= y. This choice is possible as Y has 3 or more elements. The best uninformed predictor always predicts y, with expected generalization loss 0.1, while it cannot be outperformed, see e.g. the discussion in Remark 3.2. (ii)-(iv): for faithfulness, it suffices to show that ε(ωY |X ) = ε(ωY ) implies statistical independence of X, Y . Explicit computations in each case (see Appendices A.2 and A.3) show that argminp E[L(p, Y )] = pY , where pY is the probability mass function of Y , implying that ωY = [x 7→ [y 7→ P (Y = y)]]. By conditioning the same statement on X, this also implies that argminp E[L(p, Y )|X] = [y 7→ P (Y = y|X)], thus argminf E[L(f (X), Y )] = [y 7→ P (Y = y|X = x)]. In particular, ωY |X = [x 7→ [y 7→ P (Y = y|X = x)]] is the unique minimizer of the expected generalization loss. Thus, ε(ωY |X ) = ε(ωY ) only if both functions are identical, i.e., P (Y = y|X = x) = P (Y = y) for all x, which is one possible definition of X and Y being statistically independent. For regression, the usual loss functions are unable to certify for independence anymore: Proposition 3.14. The following convex loss functions (taken as single-element sets) are not faithful for univariate regression, i.e., for the label domain Y = R (i) the squared loss L(y, y∗ ) = (y − y∗ )2 (ii) the absolute loss L(y, y∗ ) = |y − y∗ | (iii) the distance loss L(y, y∗ ) = d(y, y∗ )2 , for any metric d : Y × Y → R Proof. It suffices to construct, for each convex loss functional L, a counterexample where X, Y are statistically dependent, but no prediction functional predicting Y from X is better-than-L-uninformed. In each case, one may construct two Y-valued random variables Y1 , Y2 with distinct laws such that µL [Y1 ] = µL [Y2 ] - for example, an arbitrary non-constant Y1 and the Y2 being constant µL [Y1 ]. Further 16 setting X = {−1, 1} and defining a non-constant X-valued random variable X, together with an Yvalued random variable such that Yi = Y |X = i for i ∈ X yields an example of a statistically dependent pair X, Y of random variables where the constant prediction of µL [Y1 ] = µL [Y2 ] is not only the L-best uninformed prediction functional, but also the L-best prediction functional. Using equivalence of (i) and (iii) in Proposition 3.10 proves the claim. The previously introduced quantile losses form a strictly faithful set of losses for univariate regression: Theorem 2. The set of Q-losses is strictly faithful for (univariate) regression. More precisely, the set L = {Lα : α ∈ [0, 1]}, where Lα (y, y∗ ) = α · m(y∗ , y) + (1 − α) · m(y, y∗ ), with m(x, z) = min(x − z, 0) (see Lemma 3.4 (iii)), endowed with the Lebesgue measure through canonical identification of Lα with [0, 1], is strictly (Lebesgue-measure-)faithful for Y = R. Proof. For faithfulness, we show that impossibility of (Lα )-better-than-uninformed prediction of a R-valued random variable Y from an X-valued random variable X implies statistical independence of X and Y . Thus, assume that there is no α ∈ R and no prediction functional f such that f is Lα -betterthan-uninformed. By equivalence of (ii) and (iii) in Proposition 3.10 and negation/contraposition, (Lα ) ωY |X is uninformed hence constant. By Lemma 3.4 (iii), µLα [Y |X = x] is the α-quantile of the conditional random variable Y |X = x, which by the previous statement does not depend on x. Since α was arbitrary, none of the quantiles of Y |X = x depends on x, i.e., the cdf and hence the laws of all conditionals Y |X = x agree, which implies (by one common definition/characterization of independence) that X and Y are statistically independent. Strict faithfulness follows by following through the above argument after removing a positive-measure open set U ⊆ [0, 1] from the indices, i.e., Lα for α∈ U from L. As U has positive measure, we may pick u ∈ U and X = {−1, 1} as well as conditional cdf such that P (Y ≤ u|X = 1) 6= P (Y ≤ u|X = −1) while P (Y ≤ x|X = 1) = P (Y ≤ X|X = −1) for all x 6∈ U . By the above argument, predicting the α-quantile of Y is the Lα -best prediction functional from X for any α 6∈ U , and it is furthermore a uninformed prediction strategy, thus L is not faithful after removing Lα , α ∈ U . In the light of Theorem 2, it may be interesting to ask for a theoretical characterization of a strictly faithful set of losses for univariate regression (e.g., does it need to be infinite?), or what may be a set of strictly faithful losses for multivariate regression. A-priori, it is even unclear whether there is a set of (not necessarily strictly) faithful losses for general prediction tasks, which the following result answers positively: Theorem 3. Assume that Y may be identified (including the taking of expectations) with a sub-set of Rn , for some n, e.g., in multivariate regression, or simultaneous prediction of multiple categorical and continuous outputs. Then, the set of convex losses is faithful for Y. Proof. Consider random variables X and Y taking values in X and Y, are statistically dependent. By definition of (in)dependence, this is equivalent to there existing X0 ⊆ X, Y0 ⊆ Y (measurable with positive joint probability measure) such that P (Y ∈ Y0 |X ∈ X0 ) 6= P (Y ∈ Y0 ). By taking (or not taking) the complement of X0 within X, we may assume without loss of generality that P (Y ∈ Y0 |X ∈ X0 ) P (Y ∈ Y0 ). Define g : Y → R ; x 7→ x2 (where we have used the identification of Y with Rn . Define L : Y × Y → R ; (y, y∗ ) 7→ g(y) · 1(y∗ ∈ Y0 ) + g(y − α) · 1(y∗ 6∈ Y0 ), where 1(y∗ ∈ Y0 ) is the indicator function for y∗ ∈ Y0 , and α ∈ Y \ {0} is arbitrary. Define f : X → Y ; x 7→ 0 if (x ∈ X0 ), otherwise α. An elementary computation shows that εL (f ) is better-thanuninformed, hence we have proved non-(ii). 17 For the general case, it seems interesting to ask what would be a strictly faithful set of losses, how such sets may be characterized, or whether they even exist (which seems neither obviously nor directly implied by the existence of a faithful set of losses). Due to the constructive nature of respective proofs (or, more precisely, the semi-constructivity of the proof for multi-variate regression in Theorem 3), model comparison procedures suggested by Theorems 1 and 2 on univariate classification and regression will be used in the testing procedures. For convenience we briefly repeat the main results used in testing as a corollary: Corollary 3.15. Consider two random variables X, Y , taking values in X and Y, where Y is finitely supported pmf (classification) or where Y ⊆ Rq (regression). The following are equivalent: (i) X and Y are statistically dependent. (L) (L) (ii) ε(ωY |X ) εL (ωY ) for L the log-loss/Brier-loss (classification), resp. for some convex loss L (regression). I.e., the L-best predictor is L-better-than-uninformed, for some L. (L) (iii) there exists a prediction functional f : X → Y such that εL (f ) εL (ωY ) for L the logloss/Brier-loss (classification), resp. for some convex loss L (regression). I.e., there exists an L-better-than-uninformed prediction functional, for some L. |= Since Corollary 3.15 (i) is the alternative hypothesis in an independence test, we may use a test for the equivalent hypothesis Corollary 3.15 (iii), i.e., comparing the performance a learnt f with the best uninformed baseline, as an independence test. Any choice of f is sufficient for the methodology outlined in the latter Section 5.1. Since the null-hypothesis is that X Y , choosing a bad f or failing to detect multivariate dependence (a specific case) only decreases the power of the test, while the type 1-error remains unaffected. 3.5. Conditional independence The last section established a theoretical foundation for marginal independence testing, now this foundation will be expanded by adding a framework for testing conditional independence. The statement one would like to make thus connects two random variables, X and Y , that are conditionally independent given a third variable, Z, taking values in Z, with the expected generalization loss from predicting Y from Z and from the set {X, Z}. In slight extension of our setting, we consider prediction functionals in [X × Z → Y], i.e., we separate the features in two parts corresponding to X and the conditioning Z. We generalize the main definitions from Section 3.3 to the conditional case: Definition 3.16. A prediction functional f : X × Z → Y is called conditionally uninformed if it does not depend on the first argument, i.e., if f (x, z) = f (y, z) for all x, y ∈ X and z ∈ Z. By notational convention, a functional g : Z → Y is identified with the conditionally uninformed prediction functional (x, z) 7→ g(z). In the text, it will be implicitly understood that conditioning will happen on the second argument. We define straightforward conditional generalizations of baselines and best predictors: Definition 3.17. We define the following prediction functionals in [X × Z → Y : (L) The best conditionally uninformed prediction ωY |Z : (x, z) 7→ µL ([Y |Z = z]) (L) The best conditional prediction ωY |X,Z : (x, z) 7→ µL ([Y |X = x, Z = z]) (L) (L) It is briefly checked that ωY |Z and ωY |X,Z have the properties their names imply: Lemma 3.18. Let L be a (strictly) convex loss. The following holds: (L) (i) εL (ωY |Z ) = min{εL (f ) : f is a conditionally uninformed prediction functional} 18 (L) (ii) εL (ωY |X,Z ) = min{εL (f ) : f ∈ [X × Z → Y]} Proof. (i) Conditioning on the event Z = z, Lemma 3.7 for the unconditional case implies the equality conditional for the event. Since z was arbitrary, it holds without the conditioning. (ii) This is directly implied by Proposition 3.5, by substituting the joint random variable (X, Z) for the X in Proposition 3.5. With these definitions, the conditional variant of Theorem 3 can be stated: Theorem 4. As in our setting, consider three random variables X, Y, Z, taking values in X, Y, Z, where Y is finitely supported pmf (classification) or where Y ⊆ Rq (regression). The following are equivalent: (i) X and Y are statistically dependent conditional on Z. (L) (L) (ii) εL (ωY |(X,Z) ) εL (ωY |Z ) for L the log-loss/Brier-loss (classification), resp. loss L (regression) for some convex (L) (iii) there exists a prediction functional f : X × Z → Y such that εL (f ) εL (ωY |Z ) for L the log-loss/Brier-loss (classification), resp. for some convex loss L (regression) (iv) there exists a prediction functional f : X × Z → Y such that for all conditionally uninformed prediction functionals g : (X×)Z → Y, one has εL (f ) εL (g) for L the log-loss/Brier-loss (classification), resp. for some convex loss L (regression) The set of losses in which existence is required in (ii) may be restricted to a set of losses which is faithful for the unconditional setting (such as: quantile losses for Y = R as per Theorem 2), as in Section 3.4, without changing the fact that the stated equivalence is correct. Proof. The proof is an analogue of that of Corollary 3.15. It may be obtained from following the whole proof through while in addition conditioning on Z = z, and noting that (i) holds if and only if the conditionals X|Z = z and Y |Z = z are statistically dependent for some z. Our conditional independence test is based on statement (iv) in Theorem 4. Unlike in the uncondi(L) tional case, there is no easy way to estimate ωY |Z directly. Thus, in the algorithm in Section 5.1, the statement will be replaced by a slightly weaker statement, where automatic model selection will (L) (L) determine an f as well as a g, without guarantees that f = ωY |X,Z or that g = ωY |Z are optimal. This is in line with the usual approach to and caveats of supervised learning - if there were a universal (L) way to estimate ωY |Z , that would amount to a universally perfect supervised learning strategy. 3.6. Testing prediction error against baseline Theorems 3 or 4 in the previous sections establish a new basis for (marginal or conditional) independence testing. Namely, the theorems relate testing of (marginal or conditional) dependence between X and Y to testing predictability of Y from X. Namely, Theorems 3 or 4 state the following: If there exists a significantly better-than-(conditionally-)uninformed prediction strategy f , i.e, εL (f ) εL (g), (L) (1) (L) where g is a suitable baseline (an approximation of ωY or ωY |Z ), then we may conclude by Theorem 3 or 4 that X and Y are not (marginally or conditionally) independent. Thus the problem of independence testing is reduced to the problem of testing whether there exists a prediction functional f which outperforms the baseline g as measured by some convex loss function L. We make a few remarks about the strategy and logic of our proposed significance test. 19 • Neither the proposed functional f nor the baseline g are a-priori known in the usual practical setting, and neither is an L which may make a difference apparent. However, both f and g may be seen to approximate a best prediction functional which is unknown, thus as instances of supervised learning. Hence, some choice has to be made, in absence of the “perfect learning strategy”. A bad choice will contribute to a loss of power only if type I error control is achieved. (L) • The baseline g approximates ωY in the marginal case. For frequently used L, the baseline g predicts constants which are the mean or the median or other well-studied statistics of a sample from Y , hence g may have a beneficial asymptotic. • Significance tests to compare prediction strategies are studied in [29], which proposes amongst others a paired sample test of prediction residuals. The values εL (f ), εL (g) may be estimated by respective empirical estimates, by the usual estimator M 1 X Li (f ) εbL (f ) := M i=1 where Li (f ) = L(f (Xi∗ ), Yi∗ ), and similar εbL (g) for g. Since the test data (Xi∗ , Yi∗ ) are independent of each other and of f, g, by the central limit theorem, one has √ d M (b εL (f ) − ε(f )) − → N(0, Var[L(f (X), Y )]), M → ∞, conditional on f and g being fixed. That is, the empirical mean of the loss residuals for prediction strategy f and loss L is asymptotically normal with mean ε(f ) and variance Var[L(f (X), Y )])/N . Instead of directly estimating εL (f ) and εL (g) with confidence intervals and then comparing, one notes (as also in [29]) that the samples of loss residuals Li (f ) and Li (g) are inherently paired, which eventually leads to a more powerful testing procedure. We hence consider the difference in the i-th loss residual, Ri := Li (g) − Li (f ). The Ri are still i.i.d. (conditional on f, g) and also have a normal asymptotic, usually with smaller variance than either of f, g in empiry. An effect size of the difference is obtained with normal confidence intervals, and one may conduct a one-sided paired two-sample test for the null H0 : E(ε(g) − ε(f )) ≤ 0 against HA : E(ε(g) − ε(f )) 0. Note that the test is one-sided since we test for f to outperform the baseline g. To assess this null-hypothesis, two simple tests, one parametric and one non-parametric, are implemented. • T-test for paired samples [29]: A parametric test assuming that the sample mean of loss residual differences Ri is normally distributed with mean 0 and standard deviation σ. Under the assumptions of the test and the null hypothesis, the normalized empirical mean t := m d u/b σ, where m d u and σ b are the sample mean and standard deviation of the Ri , respectively, follows a t-distribution with M degrees of freedom, where we have used the fact that under H0 it holds that E[R] = 0. • Wilcoxon signed-rank test: If the sample mean of loss residual differences Ri is not normally distributed, an increase in power can be achieved by testing if the rank differences are symmetric around the median. Since the Ri in question are differences in loss residuals, there is no ex-ante reason to assume normality, in particular for a “good” non-baseline method one may expect that most are positive, hence the normality assumption may be too strong. Hence the nonparametric Wilcoxon class test should be a better general approach to compare loss residuals (see pages 38-52 of [11] for the Wilcoxon tests). 20 One subtlety to notice is that instead of testing for the alternative hypothesis “there exists f such (L) that εL (f ) ≤ εL (ωY )” which certifies for independence, by the above strategy we are testing for the alternative “there exists f such that for all/the best g it holds that εL (f ) ≤ εL (g)” where g is a (L) fixed estimate of ωY . However, g is dependent on the training data and itself uncertain. Thus, for a (L) conservative estimate of the significance level, uncertainty in g estimating ωY needs to be taken into account. (L) In the unconditional case, ωY will be a constant predictor of an elicited statistic of the training labels, such as the mean, median or one minus the majority class frequency, with known central limit theorems that may be used to adapt the Student or Wilcoxon type tests (e.g., by explicitly increasing the variance in the t-test). In either case, one may also bootstrap the distribution of g and make use of it as follows: Since the training set is independent of the test set, the function g, is as a random variable which is a function of the training data, also independent of the test data. Thus instead of using a fixed g in prediction, one can ensure a conservative test by using a bootstrap sample of pseudo-inputs for the predictions, a different sample of g per test data point that is predicted. Since the exact form of the correction appears an unresolved sub-problem of the predictive model validation and model selection process in general, and since bootstrapping of g is computationally costly, we have implemented the “fixed g” variant (potentially liberal) in the package which can be easily adapted to potential corrections based on explicit asymptotics. We will present an algorithmic implementation in Section 5.1 as an algorithmic conditional testing routine, which is later applied to graphical model structure learning, in Section 5.2 21 4. Graphical Models This section will briefly review graphical models and graphical model structure learning, which we will later address using the conditional independence test outlined in Section 3. 4.1. Informal definition of graphical models A probabilistic graphical model is a graph-based description of properties of a joint probability distribution (possibly but not necessarily a full specification thereof). It usually consists of a directed or undirected graph, together with a fixed convention how to translate the graph structure into a full probability distribution, or more generally, into conditional independence statements about a distribution of not further specified type. Graphs A graph G is an ordered pair (V, E) where V ⊆ E × E. The set V are interpreted as vertices (nodes) of the graph, and the elements in E are interpreted as edges (links) of the graph (ordered for directed graphs, unordered for undirected graphs). G is usually visualized by drawing all vertices V and edges between all pairs of vertices in E. In graphical models, the vertex set is identified with a collection of (possibly associated) random variables X = [X1 , ..., Xn ], and the edges encode some independence or conditional independence information about the components of X. Two popular choices are discussed in Section 4.2 below. Probabilistic graphical models Graphical model theory is usually concenred two main tasks, structure learning (of the graph) and inference (on a given graph). Inference is concerned with estimating parameters or statistics of the parametric distribution assuming the independence structure prescribed by a fixed graph, from a finite sample drawn from X. Structure learning is concerned with inferring the graph from such a sample, thus inferring the encoded independence relations. For an extended introduction to graphical models, the reader might refer to [26] or [2]. Graphical model structure learning is usually considered the more difficult of the two tasks, due to the combinatorial explosion of possible graphs. Manual approaches involve an expert encoding domain knowledge in presence or absence of certain edges; automated approaches usually conduct inference based on parametric distributional assumptions combined with selection heuristics [26]. 4.2. Types of graphical models We review two of the most frequently used types of graphical models. 4.2.1. Bayesian Networks A Bayesian network is a graphical model which states conditional independence assumptions without making parametric distributional assumptions. The conditional independence assumptions are encoded in a directed acyclic graph over a set of random variables. Acyclicity of the graph implies that each node Xi has a (potentially empty) set of descendants, Definition 4.1. A node Xj is a descendant of Xi in the graph G if there is a directed path from Xi to Xj , where a path is any connection of links that lead from Xi and Xj . That is, if there is a path following the arrows in G, going from Xi to Xj . Since the graph is acyclic, no cycles exist, and following the arrows in the graph, the same variable can never be reached twice. Examples of such a networks are shown in Figure 1. 22 Difficulty Intelligence Difficulty Intelligence GPA Grade Grade Grade Letter Letter Letter (a) (b) (c) Figure 1: Example of expert based graphical model structure learning (backtracking), adapted from [26]. (a) The quality of an academic reference letter is determined by the grade (b) The students intelligence and course difficulty determine the grade. (c) Knowing a students GPA gives additional information about the state of intelligence Bayesian Network graphs arise in a natural way when considering the factorization properties of a probability distribution. Assume we are interested in a probability distribution P over the Dif f iculty of a course, a students Intelligence, the Grade a student achieved in a course, and the quality of a reference Letter received from the tutor of the course, P (Dif f iculty, Intelligence, Grade, Letter). Further assume the following is true |= • The quality of the Letter is solely determined by the Grade a student received in the course. That is, Letter {Dif f iculty, Intelligence}|Grade • The Grade of a student depends on the Dif f iculty of the course and his Intelligence • Dif f iculty and Intelligence are not causally influenced by any other variable in the graph This gives a natural way to order the variables for factorization. Dif f iculty and Intelligence are so-called root nodes, hence their order is irrelevant, however Grade depends on both of them and Letter depends on Grade, giving the ordering: {Letter, Grade, Dif f iculty, Intelligence}, which will now be denoted as {L, G, D, I}. An ordering {X1 , ..., Xn } implies that for ∀i, j ∈ {1, ..., n} and i < j, Xj can not be a descendant of Xi . The distribution is then factorized, according to the ordering, P (L, G, D, I) = P (L|G, D, I)P (G, |D, I)P (D|I)P (I), a standard way to factorize distributions. Working in the independence statements gives P (L, G, D, I) = P (L|G)P (G, |D, I)P (D)P (I), |= |= since L {D, I}|G and D I. Returning to Figure 1 (b) shows that this factorized distribution exactly matches the arrows in the graph, which start at the parents (the conditioning variables) and lead to the children (the variable that the conditional distribution is over). To understand the independence statements encoded in a Bayesian Network, one needs to first distinguish between active and passive trails [26, p. 71]. Let a trail be any path between Xi and Xj on G and a v-structure a structure such that Xi−1 → Xi ← Xi+1 , where Xi is a descendant of both Xi−1 and Xi+1 . 23 A C E B D F G Figure 2: Markov Network Definition 4.2. A trail between Xi and Xj is active given a conditioning set Z, if for every v-structure Xi−1 → Xi ← Xi+1 along the trail, Xi or one if it’s descendants is in Z and no other variables on the trail are in Z. In Figure 1 (c) that means, the trail from Letter to GPA is active only if conditioned on either the empty set or Difficulty, and the trail between Intelligence and Difficulty is only active if Grade or Letter is in the conditioning set. |= Definition 4.3. [2] If there is no active trail in the graph G between the nodes X and Y given Z, then X Y |Z in any distribution consistent with the graph G 4.2.2. Markov Networks A Markov Network, or Markov Random Field, is a graphical model which encodes conditional independence statements in an undirected graph over a set of random variables. While the Bayesian Network is defined in terms of probability distributions, the Markov Network is usually specified in terms of factor products which are unnormalized probability distributions. The scopes of these factors determine the complexity of the system, if the scope of a factor covers all the variables in the graph, the graph is unconstrained, whereas any constraint to smaller scopes decreases its complexity. Figure 2 shows the Markov Network for the (normalized) factor product P (a, b, c, d, e, f, g) = 1 φ1 (a, b, c)φ2 (b, c, d)φ3 (c, e)φ4 (e, f )φ4 (f, g)φ5 (e, g), Z with the normalizing constant (partition function) Z. Small letters denote realizations of the capitallettered nodes. If P is a probability mass function, X Z= φ1 (a, b, c)φ2 (b, c, d)φ3 (c, e)φ4 (e, f )φ4 (f, g)φ5 (e, g), a,b,c,d,e,f,g |= |= where φ(s) is such that φ : S → R, ∀s ∈ S. If P is a probability density function over continuous variables, Z would be attained by integrating over the support of the variables in the scopes of P . Since the links are undirected, independence statements arising from Markov Networks are symmetric. A path between Xi and Xj , i 6= j is called active given Z, if no node on the path is in Z. To determine whether Xi and Xj are conditionally independent given Z in all distributions consistent with G, one again considers all paths going from Xi to Xj . It holds that Xi Xj |Z if there is no path between Xi and Xj that is active given Z. So to attain a Markov Network that is consistent with a probability distribution, one has to consider the pairwise conditional independence properties of the variables that the nodes in the graph represent. In general, if in a Markov Network there is no link between a variable Xi and Xj , then Xi Xj |X \ {Xi , Xj } in any distribution P over X consistent with the graph. These are called the pairwise Markov-independencies of the graph [15, ch. 17.2]. The independencies that can be encoded using Bayesian and Markov Networks differ. Bayesian 24 X Y X Y Figure 3: Two DAGs equivalent with respect to independence statements Networks are natural to express (hypothesized) causal relationships, distributions where an ordering can be attained and thus have nice factorization properties, while Markov Networks are natural for expressing a set of pairwise Markov independencies. Additionally, Bayesian Networks can be transformed into an undirected graph by a (potentially forgetful) process called ”moralization” [26, p. 135]. 4.3. Graphical model structure learning There are two dominant approaches to structure learning, independence testing based and score-based methods, however they both suffer to a varying extent from the same underlying problem: combinatorial explosion. Given a probability distribution P (with associated graph G) over a multivariate random variable X = [X1 , ..., Xn ], for undirected networks, between any pair of variables Xi and Xj potential there can be a link or no link in G. Since the links are undirected, there are thus p(p−1) 2 edges in the graph, and an exponential number of possible graphs. If one wants to test if Xi and Xj are independent given Z, Z ⊂ X, one needs to test if any path between the two is active, leading to a potentially very large number of tests. A graph’s complexity can be decreased by upper bounding the number of edges for each node, however, it was shown that, for directed graphs where each node can have d parents, for d > 1, the problem a finding an optimal graph is NP-hard [26, p. 811]. Generally, the space of graphs is reduced by making structural (what type of relationships can be captured in the graph) and parametric (constraints on the distribution) assumptions. An additional problem is posed when there exists more than one optimum to the structure learning method, as a result of the fact that different graphs can be equivalent (and thus not identifiable) in the respective search space. An example of this are the two graphs shown in Figure 3, which arises from the fact that if {X, Y } ∼ P , P can be decomposed either into P (X|Y )P (Y ) or P (Y |X)P (X), which are equivalent. There are different ways to approach these problems, some of which are outlined below. 4.3.1. Score-based methods Score-based methods attempt to find a (usually local) optimum of a score function in the space of possible graphs. Examples of this can be the Kullback-Leibler divergence scoring function for directed graphical models, which can be used to find the optimal Chow-Liu tree graph [2, p.219]. While this is a convex algorithm, finding the global optimum, it does so under the heavy constraint that each node only has one parent at maximum (resulting in a tree-structured graph). For undirected graphs over discrete variables, gradient descent can be used directly on the log likelihood to find the local optimum [2, p. 221], however each gradient evaluation requires calculation of the partition function Z (by summing over all discrete states), which makes this algorithm very expensive. Many of the algorithms for score-based methods are designed for discrete variables, and not applicable when the variables are continuous. Performance is usually evaluated by calculating the likelihood score on a hold-out sample. One other area not mentioned above is Bayesian model averaging (e.g. [14]), which seeks to improve the above methods by averaging over different model structures. It can be viewed as an ensembling method for the score-based methods mentioned above. State-of-the-art score based methods are oftentimes very expensive or make very strong assumptions on the underlying distribution P , such as tree-structure of the graph [26, ch. 18]. 25 4.3.2. Independence testing based methods Unlike the score-based models, independence testing based models conduct independence tests between variables or sets of variables locally and then aggregate the information to produce the graphical model. These sets of independence tests are then used to infer the structure of the graph in a straightforward manner, based on the properties of the respective type of graphical model. An example of this is the PC-Algorithm [2, e.g. p. 214] which is used in attempts at causal discovery. Another algorithm for recovering the undirected skeleton is given by Algorithm 3.3 of [26]. There are probably two main reasons that this latter independence testing based method is not used for graphical modelling: (a) is that it relies on a conditional independence test where the conditioning is on multiple variables, i.e., all except two. As outlined in Section 2, this is a largely unsolved problem expect when there is at most one conditioning variable, i.e., if there are three or less variables in total. (b) It is hard to regulate the bias-variance trade-off, unlike for score-based methods where this may be achieved through the use of constraints such as an upper bound on the number of parents. However, these two issues may be addressed through our novel predictive conditional independence testing framework: (a) The predictive conditional independence test described in the subsequent Section 5.1, based on the theoretical insights in Section 3, allows for efficient conditional independence testing with variables of any dimension, and thus provides a predictive approach to learning graphical models. (b) The link to the supervised learning workflow allows for direct adoption of well-known strategies to trade-off bias and variance and error estimation from supervised learning, such as regularization and cross-validatied tuning, to the graphical modelling workflow.. In addition, our suggested algorithms have further desirable features: • The intimate link to supervised learning also allows for a controlled trade-off between power of the algorithm and time complexity. This trade-off is readily available when using the conditional independence testing routine described in Section 5.1, since the user can choose estimation methods of varying power/complexity in order to influence the total run time needed by the algorithm. • Section 5.1.1 will also introduce a false-discovery-rate control routine to provide a tool that lets a user control the proportion of false-positives (erroneously found edges in the estimated graphical model structure), regardless of the size of the graph. Note on causal inference Graphical models, and specifically Bayesian Networks, are a natural way to express hypothesized causality, since the arrows seem to express causation. However, when actually learning graphical models from data, causality may only be inferred by making strong (and often incorrect) assumptions on the underlying distribution, or by collecting data in a manner that allows for causal interpretation, namely in proper experimental set-ups (e.g., following Pearl’s do-calculus, or standard experimental study design). As generally for graphical model structure learning, all algorithms outlined in this paper are not to be interpreted as expressing causality, but merely as producing a collection of statements about association which certify for causality only in combination with the right experimental set-up. 26 5. Predictive inference algorithms This section will first introduce the proposed predictive conditional independence testing routine (PCIT), which is based on the methodology outlined in Section 3, and important subroutines related to performance and error control of the test. After, an algorithm to leverage the independence test into a graphical model structure learning routine that addresses the issues outlined in 4.3.2 is presented. 5.1. Predictive conditional independence testing Algorithm 1 implements the results from Section 3 to test if a set of variables Y is independent of another set of variables X, given a conditioning set Z (optional, if not provided the algorithm tests for marginal independence). It will later be used as a subroutine in the graphical model structure learning, but can also be used in it’s own right for tasks such as determining if a subset X would add additional information when trying to predict Y from Z. Algorithm 1 Predictive conditional independent test (PCIT) Split data into training and test set for all variables y ∈ Y do on training data: 4: find optimal functional f for predicting y from Z 5: find optimal functional g for predicting y from {X,Z} 1: 2: 3: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: 17: on test data: calculate and store p-value for test that generalization loss of g is lower than f end for if symmetric test needed then exchange X and Y, repeat above process end if p values adjusted ← Apply FDR control to array of all calculated p-values return p values adjusted When the test is used as a marginal independence test, the optimal prediction functional for f is the functional elicited by the loss function, that is, the mean of y for continuous outputs, and the class probabilities of y for the discrete case (Appendix ??). The link to supervised learning further allows/forces the user to distinguish between two cases. Independence statements are symmetric, if X is independent of Y , then Y is independent of X, in both the marginal and conditional setting. The same cannot be said in supervised learning, where adding X to predicting Y from Z might result in a significant improvement, but adding Y does not significantly improve the prediction of X, given Z (as can be seen in the asymmetry of OLS). So if a user is interested in a one-sided statement, the algorithm can be run for a single direction, for example to evaluate if a new set of variables improves a prediction method, and is thus worth collecting for the whole population. If one is interested in making a general statement about independence, the test is “symmetrized” be exchanging X and Y , and thus testing in both directions, and FDR control be applied to the union of the p-values. It is important to distinguish between the two types of tests and null-hypotheses in this test. On one hand (in the symmetric case) for each variable in X and Y , it will be assessed, if adding X to the prediction of y ∈ Y from Z results in an improvement (and vice versa). The null-hypothesis of this “prediction-null” is that no improvement is achieved. After all these p-values are collected, a multiple testing adjustment is applied (Section 5.1.1), after which the original null-hypothesis, that X and 27 Algorithm 2 The Benjamini-Hochberg-Yekuteli Procedure j Require: Set {p(i) }m i=1 s.t. pj : p-value for observing Xj under H0 1: Sort the p-values in ascending order, p(1) ≤ ... ≤ p(m) Pm 2: Let q = α/( i=1 1/i) for chosen confidence level α i 3: Find the k s.t. k = max(i : p(i) ≤ m q) j 4: Reject H0 for j ∈ 1, ..., k Y are conditionally independent, is assessed. We reject this “independence-null”, if any one of the “prediction-nulls” can be rejected after adjustment. The p-value of the “independence-null” hence reduces to the lowest p-value in all the “prediction-nulls”. As such, the null of the independence test is that all the “prediction-nulls” are true. False discovery rate control, the chosen multiple testing adjustment, is appropriate since it controls the family-wise error rate (FWER), the probability of making at least one type 1 error, if all the null-hypotheses are true [4]. 5.1.1. False-discovery rate control To account for the multiple testing problem in Algorithm 1, the Benjamini - Hochberg - Yekutieli procedure for false-discovery rate (FDR) control is implemented [5]. In their paper, they state that traditional multiple testing adjustments, such as the Bonferroni method, focus on preserving the FWER. That is, they aim to preserve the probability of making any one type 1 error at the chosen confidence level. As a result, tests are usually very conservative, since in many multiple-testing scenarios the p-values are not independently distributed (under the null), and thus the power of these tests can be significantly reduced. In their 2001 paper, they propose to control the false discovery rate instead, “[..] the expected proportion of erroneous rejections among all rejections”, as an alternative to the FWER. The FDR allows for more errors (in absolute terms) when many null-hypothesis are rejected, and less errors when few null-hypotheses are rejected. Algorithm 2 shows the procedure outlined in [5]. They showed, that this procedure always controls the FDR at a level that is proportional to the fraction of true hypotheses. As hinted at before, while this algorithm controls the false-discovery rate, in the special case where all null-hypotheses in the multiple testing task are assumed to be true, it controls the FWER, which then coincides with the FDR. This is especially useful, since both scenarios occur in the graphical model structure learning routine described in Section 5.2. For the choice of optimal false discovery-rate for an investigation, even more so than in the classical choice of appropriate type 1 error, there is no simple answer for which rate might serve as a good default, and it is highly application dependent. If the goal of the procedure is to gain some insight into the data (without dire consequences for a false-discovery), a user might choose a FDR as high as 20%, meaning that, in the case of graphical model structure learning, one in five of the discovered links is wrong on average, which might still be justifiable when trying to gain insight into clustering properties of the data. This paper will still set the default rate to 5%, but deviate willingly from this standard whenever deemed appropriate, as should any user of the test. 5.1.2. Improving the prediction functionals In practice, when assessing the individual “prediction-nulls” in Algorithm 1, the power of the test (when holding the loss function constant) depends on the capability of the the prediction method to find a suitable functional g that outperforms the baseline f . That means, a general implementation of the independence test needs to include a routine to automatically determine good prediction functionals for g and f . The implementations in the pcit package presented in Section 6 support two methods for the automatic selection of an optimal prediction functional. Both methods ensemble over a set 28 Stage 1 Regression Classification ElasticNetCV BernoulliNB GradientBoostingRegressor MultinomialNB RandomForestRegressor GaussianNB SVR SGDClassifier RandomForestClassifier SVC Stage 2 LinearRegression LogisticRegression Table 1: Prediction functionals used for Stacking/Multiplexing of estimators, which are shown in Table 1. The prediction functionals refer to the estimator names in sklearn1 . Some are constrained to specific cases (e.g. BernouilliNB, which only applies when the classification problem is binary). Stacking Stacking refers to a two-stage model, where in the first stage, a set of prediction function is fit on a training set. In the second stage, another prediction function is fit on the outputs of the first stage. If the prediction function in the second stage is a linear regression, this can be viewed as a simple weighted average of the outputs in the first stage. In theory, stacking allows the user to fit one single stacking predictor instead of having to compare many potential prediction functions based on the model diagnostics, as in the second stage, better methods get more weight (in the expectation). As an additional effect, improvement in the prediction accuracy through ensembling of predictors can take place (see e.g. Section 19.5 of [1]). The used stacking regressor and classifier can be found in the Python package Mlxtend [32]. Multiplexing When multiplexing over estimators, the training set is first split into a training and validation set, a common procedure to find optimal hyperparameters. After, the predictors are fit individually on the training set, and each predictors expected generalization loss is estimated on the validation set. One then proceeds by choosing the predictor with the lowest estimate for the empirical generalization loss, and refits it using the whole training data (including the former validation set), and then uses the fitted estimator for prediction. 5.1.3. Supervised learning for independence testing Algorithm 1 shows how to leverage supervised learning methodology into a conditional independence test. This has a major advantage over the methods outlined in Section 2, as the supervised prediction workflow is of utmost interest to many areas of science and business, and, as a result, a lot of resources are going into development and improvement of the existing methods. By making a link between predictive modelling and independence testing, the power of independence testing will grow in the continuously increasing power of the predictive modelling algorithms. 5.2. Predictive structure learning of undirected graphical models This section outlines a routine to learn the vertices in a directed graph (the skeleton) for a data set by conducting a range of conditional independence tests with the null hypothesis of conditional independence. Section 4.2 outlines the conditional independence statements of a Markov network. In a directed graph, if variables Xi and Xj have no direct edge between them, they are conditionally independent given all other variables in the network. 1 Details can be found here http://scikit-learn.org/stable/modules/classes.html 29 Algorithm 3 Undirected graph structure estimation for any combination Xi , Xj s.t. i 6= j do X− ← X \ {Xi , Xj } p vali,j ← p-value for test Xi Xj |X− end for p val adj ← Apply FDR control on p val matrix return p val adj |= 1: 2: 3: 4: 5: 6: Algorithm 3 describes the skeleton learning algorithm for an input set X = [X1 , ..., Xn ], by considering all possible pairs of variables in the data set, and testing if they are conditionally independent given all other variables. The output p val adj is a symmetric matrix with entries i,j being the p-value for the hypothesis that in the underlying distribution, Xi and Xj are independent, given all other variables in X, and hence in the graph G describing it, there is no link between the vertices for Xi and Xj . Ultimately, links should be drawn where the adjusted p-values are below a predefined threshold. There are O(n2 ) function evaluations in the for-loop, where n is the number of variables. Section 7 will provide experimental performance statistics for the algorithm and showcase applications on real data sets. 30 6. pcit package 6.1. Overview The Python2 package implementing the findings and algorithms of this paper can be found on https: //github.com/SamBurkart/pcit and is distributed under the name pcit in the Python Package Index. This section will first provide an overview of the structure and important functions of the package. As this implementation can be thought of as a wrapper for scikit-learn (sklearn) estimators, this section will then describe the sklearn interface and how the package interacts with the sklearn estimators. Lastly, simple application-examples are given. 6.1.1. Use cases The package has two main purposes, independence testing and structure learning. While univariate unconditional independence testing is possible, it is not expected to outperform current methodology in the simple tasks, but the difficult ones, such as Multivariate independence tests, such as for hedging purposes in a multivariate test to determine which financial products are independent from the ones already in the portfolio. Conditional independence tests, such as assessing if a data set adds information to existing data in a prediction task. Graphical model structure learning, such as for finding clusters in the data as part of an exploratory data analysis or thoroughly investigating associations in the data (both in section 7.2.2). For these tasks, the PCIT serves as a readily available tool that works without the need for manual choices or hyperparameter tuning, and scales well in the dimensionality of the data. 6.1.2. Dependencies The package has the following dependencies: Scipy [25], for the calculation of p-values Sklearn [31], for its estimators (predictors) Mlxtend [32], for the implementation of stacking 6.2. API description The package introduces three main routines, one for automated prediction (MetaEstimator), conditional independence testing (PCIT), and undirected graph structure learning (find neighbours). The following section gives on overview, the function signatures can be found in Appendix 6.3. MetaEstimator The MetaEstimator class provides a user with a type-independent predictor that automates model selection for given training data, by automatically determining appropriate loss functions and prediction functionals. It is initialized for sensible defaults, which should generally lead to good results, but can be changed for specific tasks (such as the use of more complex base estimators for more powerful, but also more computationally demanding, predictions). The ensembling methods used for the estimator are described in Section 5.1.2. For regression, the square loss is used for training and calculation of the residuals, for classification, the logistic loss serves as the loss function. 2 https://www.python.org/ 31 PCIT PCIT implements the conditional independence test in Algorithm 1 to test if two samples stem from conditionally (or marginally) independent random variables. The MetaEstimator class is used as a prediction functional, and hence the user can trade off between computational complexity and power by adjusting the chosen MetaEstimator. That is, if speed is important, the used MetaEstimator should be a combination of base estimators and ensembling method that is quick in execution, whereas if computational resources are vast and a more powerful test is needed, the used base-estimators should be highly tuned. find neighbours find neighbours implements Algorithm 5.2 to learn the undirected skeleton for an input data set X, using the PCIT. 6.3. Function signatures PCIT Type: function Inputs: Outputs: Name Description (type) Default x Input data set 1 ([n x p] numpy array) y Input data set 2 ([n x q] numpy array) z Conditioning set ([n x r] numpy array) None (empty set) estimator Estimator object to use for test (MetaEstimator) MetaEstimator() parametric Parametric or nonparametric test (bool), section 3.6 False confidence Confidence level for test (float [0,1]) 0.05 symmetric Conducts symmetric test (bool), section 5.1 True p values adj p-values for ”prediction nulls” of each y ∈ Y (list) independent tuple, first value shows if ”independence-null” is rejected (bool) second value is p-value of ”independence-null” (float [0,1]) loss statistics RMSE difference for baseline f and altern. g, loss residuals with standard deviation, for each y ∈ Y . Only applicable if Y continuous Note: The variance of the difference is estimated by assuming 0 covariance between the residuals of baseline f and alternative g, which generally leads to more conservative confidence intervals for error residuals (due to the irreducible error, prediction residuals for different methods are generally positively correlated). 32 MetaEstimator Type: class Methods init Name Description (type) Default Inputs: method Ensembling method (’stacking’, ’multiplexing’ or None) ’stacking’ estimators Estimators to ensemble over (2-tuple of lists of sklearn None (default estim, estimators [regression estim], [classification estim]) section 5.1.2) method type Task definition (’regr’, ’classif’, None) None (auto selection) cutoff Cutoff for automatic selection of method type (integer) 10 Inputs: y Dependent variable ([n x 1] numpy array) Outputs: estimators Appropriate set of estimators (list) x Independent variables ([n x p] numpy array) y Dependent variable ([n x 1] numpy array) fitted Fitted estimator (MetaEstimator) x Independent variables ([n x p] numpy array) y Dependent variable ([n x 1] numpy array) fitted Fitted uninformed baseline estimator (MetaEstimator) get estim fit Inputs: Outputs: fit baseline Inputs: Outputs: predict Requires: MetaEstimator has been fitted Inputs: x Test set independent variables ([n x p] numpy array) Outputs: predictions Predictions for test set ([n x 1] numpy array) x train Training set independent var. ([n x p] numpy array) y train Training set dependent variables ([n x 1] numpy array) x test Test set independent variables ([n x p] numpy array) y test Test set dependent variables ([n x 1] numpy array) baseline Should baseline be fitted (boolean) resid Residuals for prediction strategy ([n x 1] numpy array) get resid Inputs: Outputs: 33 False find neighbours Type: function Inputs: Outputs: Name Description (type) Default X Input data set ([n x p] numpy array) estimator Estimator object to use for test (MetaEstimator) MetaEstimator() confidence False-discovery rate (float [0,1]) 0.05 skeleton Matrix, p-values for each indep test ([p x p] numpy array) skeleton adj Learnt graph after applying FDR control ([p x p] numpy array) 6.4. API design The API is designed as a wrapper for Scikit-learn (sklearn), a package in the Python programming language, that aims to provide a user with a consistent, easy-to-use set of tools to analyze data [31]. It is one of the most-used tools in today’s supervised learning community, which is why it is chosen as the supervised prediction workflow to build on for the predictive conditional independence test. This section will outline the advantages of basing the test on the sklearn package. 6.4.1. Sklearn interface Sklearn is built around estimator objects, which implement a consistent set of methods. Estimators provide a fit and, if applicable, a predict method, in order to be able to fit the estimator to training data and then predict on a test set. Additionally, sklearn provides standardized approaches to model selection and hyperparameter-tuning as well as data transformation and ensembling methods (see [8] for a more thorough discussion). The methods can easily be combined to create more powerful estimators. Defaults are chosen sensibly so that in most cases, an initial fit of a method to data requires little manual parameter specification from the user’s side. While the highly automated and simplified approach of sklearn lowers the bar of entry when aiming to generate knowledge from data, it also comes with a downside. For most statistical applications that exceed fitting and predicting from a data set, such as inference on the parameters and hypotheses about prediction accuracies, the relevant subroutines are missing from the API. Relevant statistics can however be attained by interfacing it with other packages (such as SciPy). 6.4.2. Wrapper for Sklearn estimators As we saw in section 6.2, the conditional independence test described in Algorithm 1 uses the newly defined MetaEstimator class to automate determining the optimal prediction functional for a given task. It does so, by ensembling over a set of base estimators from sklearn. These are either chosen to be the sensible defaults described in Table 1, or can be passed by the user as a tuple of lists of sklearn base estimators. This is required since regression and classification tasks rely on vastly different prediction functionals, and thus need to be specified separately. As a general rule, the passed regressors need to have a fit and a predict method, whereas the classifiers need to have a fit and a predict proba method. Requirements might be more stringent for certain types of data or certain estimators, however specifying an unsuitable estimator class will result in an error message as specified by the respective class, allowing the user to either remove the unsuitable class or proceed with the sensible defaults. As mentioned before, this gives a user a flexible tool to trade off between power and computational complexity. If in need of a fast method, one can use an algorithm that runs in linear time, such as stochastic gradient descent linear regression, whereas if a test with high power is needed, one can pass hyper-tuned estimators to the function, that take longer to run but generalizes better for prediction on unseen data. 34 6.5. Examples This section will provide some simple examples of how the code is used. For the following it is assumed that data sets X, Y and Z, all of the size [number of samples × number of dimensions], are loaded as numpy arrays, and have matching numbers of samples (sample indices in X, Y and Z correspond to the same sample). After installing the pcit package, import the relevant objects: Testing if X Y |Z, using the default values: I n d e p e n d e n c e T e s t . PCIT(X, Y, z = Z ) Testing if X mators: |= 1 |= 1 from p c i t import MetaEstimator , S t r u c t u r e E s t i m a t i o n , I n d e p e n d e n c e T e s t Y |Z, with a custom MetaEstimator, multiplexing over a manually chosen set of esti- 1 from s k l e a r n . l i n e a r m o d e l import RidgeCV , LassoCV , 2 SGDClassifier , LogisticRegression 3 4 r e g r e s s o r s = [ RidgeCV ( ) , LassoCV ( ) ] 5 c l a s s i f i e r s = [ SGDClassifier () , LogisticRegression ( ) ] 6 7 c u s t o m e s t i m = MetaEstimator . MetaEstimator ( method = ’ m u l t i p l e x i n g ’ , 8 estimators = ( regressors , c l a s s i f i e r s )) 9 10 I n d e p e n d e n c e T e s t . PCIT(X, Y, z = Z , 11 estimator = custom estim ) Learning the undirected skeleton of X: 1 S t r u c t u r e E s t i m a t i o n . f i n d n e i g h b o u r s (X) Concrete outputs will be shown in section 7 and can be found on GitHub. 35 7. Experiments This section will first evaluate the performance of the proposed algorithms, and then provide some examples of applications on real world data sets. All performance tests can be found on Github3 . 7.1. Performance tests This section will report on performance tests for the algorithms derived in Section 5. First the power of the the predictive conditional independence routine is bench-marked against current state-of-the-art methodology, then various tests on the directed graph structure learning algorithm are conducted. 7.1.1. Performance of conditional independence test In this section the conditional independence routine will be bench-marked against the previous research, namely the kernel based approach for conditional independence testing, which is is shown to be more powerful than other conditional independence testing algorithms in [37] (on multivariate Gaussian data). The used code for the kernel test is taken from GitHub4 . To conduct a test using data that is drawn from a distribution that more closely resembles real world data, as opposed to the synthetic (Gaussian) data commonly used for performance tests, the UCI Wine Repository data set [27] is used as follows: • The columns ’Alcohol’, ’Malic Acid’ and ’Magnesium’ are randomly permuted (to make them independent) and will serve as X, Y and noise arrays respectively • Vector Z is created by individually sampling vectors X 0 , Y 0 and noise0 of size n with replacement from X, Y and the noise vector, and then calculating p Zi = log(Xi0 ) × exp(Yi0 ) + u ∗ noise0i , i ∈ {1, ..., n}, where u is the sign, uniformly drawn from {−1, 1} |= This results in a scenario, where X Y , but X 6⊥⊥ Y |Z, and the signal to noise ratio is high for small sample sizes. The test will be conducted by increasing the sample size from 100 to 5000, and calculating the run times and power for both approaches. For each sample size, the PCIT is run 500 times, and the KCIT is run 200 times, since the KCIT is more computationally demanding than the PCIT. The only exception is n = 5000 for the KCIT, which is run 25 times, since the time complexity would be too high to draw more samples. Both methods are run for their default values, without additional manual tuning, and at a confidence level of 5%. Each time, {X 0 , Y 0 , Z} is sampled, and then the conditional independence tests are applied and the run time is recorded. If they reject independence at a 5% level, the round counts as a success, 1, otherwise 0. The power and run times are then calculated by averaging over the results, and standard errors for the power are attained by realizing that the standard error of the power for a rerun number of B is the standard error of X B, where X ∼ Bin(B, θ), where θ is the observed power (the sample mean). The results are shown in Table 2. The power at the higher end of the sample sizes seems to be similar (it is important to note that the power of 1 for the KCIT for n = 5000 was achieved on 25 resamples only), where as in the range 500 to 1000 samples, the proposed predictive conditional independence test shows a significantly higher power. For small n, the KCIT seems to fare significantly better, however both approaches have a very low power, and the PCIT especially shows the power levels below the confidence levels, which might indicate a discrepancy between true type 1 error and expected type 1 error. Important to note is the very high computational complexity of the kernel-based approach for a data set of size 5000, with a run time of approximately 80 minutes per test, while the predictive 3 https://github.com/SamBurkart/pcit/tree/master/tests 4 https://github.com/devinjacobson/prediction/tree/master/correlations/samplecode/KCI-test 36 n PCIT 100 Power Time (s) KCIT Power 200 500 1000 2000 5000 0.020 0.046 0.332 0.672 0.832 0.970 (0.006) (0.009) (0.021) (0.021) (0.017) (0.007) 0.32 0.38 0.49 0.624 1.31 4.79 0.050 0.085 0.185 0.325 0.8 1 (0.015) (0.019) (0.027) (0.033) (0.028) (∗) Time (s) 0.57 1.25 9.8 44 383 4758 Stand. difference -1.8 -1.78 4.25 8.85 0.97 * Table 2: Performance statistics for the newly proposed predictive conditional independence test (PCIT) and the kernel based approach (KCIT). The values in brackets show the estimated standard errors. The last row shows the standardized difference between the power estimates, PCIT KCIT . conditional independence test still has a very low run time of 4.8 seconds. This is to be taken with a grain of salt, since the tests were run in different languages (PCIT in Python, KCIT in MATLAB), but it is apparent that PCIT scales much better than KCIT, while the both converge to a power of 1. 7.1.2. Performance of structure learning algorithm: Error rates This test will show that the graphical model learning algorithm is capable of recovering the true graph with high probability as the sample size increases. No comparison with alternative tests is made, as it would lead to infeasible run times. The data used in the performance tests is generated as follows: 1. Sample a random positive definite, sparse, symmetric precision matrix M . The size of the entries of this matrix (relative to the diagonal) determine the signal to noise ratio, and are thus thresholded to 0, if below a certain value. |= 2. Invert M , M 0 = inv(M) and use M 0 as covariance matrix to sample from a multivariate Gaussian distribution. This has the effect, that the zero-entries in M express zero-partial correlations [15, ch. 17.3] between respective variables (and hence, lack of an edge between the two nodes in the graph). That is, for a multivariate Gaussian random variable X, X = [X1 , ..., Xp ], Mi,j = 0 =⇒ Xi Xj |X \ {Xi , Xj }. 3. Sample D, a data set of size n, from the multivariate normal P = N (0, M 0 ). The choice of n will allow to evaluate the algorithms performance for an increasing signal to noise ratio. 4. The undirected graph G consistent with the probability distribution P is given by M, where edges occur between variables Xi and Xj , if Mi,j is non-zero. Then, Algorithm 3 will be tested by getting an estimate Ĝ of the structure of an undirected graph G induced by the distribution P from which the data set D was drawn. The performance evaluation will be guided by three metrics: • False-discovery rate: The FDR is the fraction of type 1 errors (edges in Ĝ that are not in G) over the total number of identified edges in the learned Ĝ • Power: Number of found edges (links in true graph G that were found by the structure learning algorithm) over the total number of links in G 37 FDR Time (sec) No ensembling 3.09% 30 Stacking 3.03% 450 Multiplexing 2.75% 1000 Table 3: False-discovery rates and run times for a data set of 22000 for all used methods • Time: Run time needed to estimate Ĝ For the test, the number of random variables is chosen to be 10. This means, that each node in the graph has up to 9 neighbours, and the total number of possible undirected links (before inducing sparsity) is 45. The sparsity parameter is chosen in a way that generally between 10 and 15 of those links exist in the true graph. The size and sparsity of the graph are chosen to produce estimators of the metrics with reasonably low variance, but are otherwise arbitrary. The sample sizes range from approximately 400 to 20000, increasing in steps of 10%. For each sample size, 10 tests are run to decrease the variance in the estimators. The test is conducted for conditional independence testing using stacking and multiplexing, as well as without using any ensembling method, which, since the data is continuous, results in the usage of Elastic Net regularized regression. If the algorithms work as expected, the FDR is expected to be at or below 5%. High power-levels indicate a better performance of the algorithm, with respect to this specific task. At the very least, the power level is expected to increase in the number of samples, suggesting that asymptotically the routine will find the correct graph. Table 3 shows the average FDR and run times for each of the three methods. The average FDR seems to be similar across all three methods, whereas the computational complexities differ by a large amount. No ensembling PCIT runs very quick, about 15 times faster than stacking, which itself only takes about half as long as multiplexing. This is the case, since multiplexing requires the calculation of performance measures for each used estimator. Figure 4 shows the power and FDR of the algorithm for increasing sample size. The FDR for all 3 methods seem to be slightly higher for small sample sizes than they are for large sample sizes, but they are generally around or below the desired 5% in the expectation (the variances are quite high, as the number of possible reruns is low due to the computational complexity of the multiplexing method). While it might seem surprising that stacking and multiplexing are outperformed by the no-ensembling method, one has to remember that the ensembling is used to choose the optimal prediction functional automatically from a set of estimators. However, the data is multivariate Gaussian, for which ridge regression is optimal in terms of generalization performance. While stacking and multiplexing are tasked to find this property in a large set of estimators, the estimator used in the no ensembling case is Elastic Net regularized linear regression, a generalization of ridge regression, and hence fares better since there is less variance in finding the optimal estimator. For all three methods, the power increases roughly logarithmically in the sample size, implying that for a test that recovers the truth with high probability, a large data set or a more powerful set of estimators (see Section 7.2.2) might be needed for that specific task. However, asymptotically, all three tests seem to recover a graph that is close to the truth, in this specific scenario, unless the power starts to plateau before reaching 1 (which there is no indication for). Since the power for the no ensembling case is biased by the fact that it uses an optimal prediction functional, the power curves for stacking and ensembling provide a better estimate for the performance on an unseen data set of an unknown distribution family. As graphical model structure learning is an inherently hard problem (due to issues such as multiple testing, combinatorial explosion, as outlined in Sections 4), it is promising that the algorithm finds an increasing fraction of the links while keeping the ratio of false-discoveries constant. 38 Figure 4: Power (blue lines) and FDR (red lines) for increasing sample size for all 3 methods, showing that the performance increases as expected for all 3 methods, when the sample size is increased. The transparent blue and red areas denote the 90% confidence intervals, the dashed line shows the expected FDR (0.05) 39 Figure 5: Frequencies of edge occurrence in the learned structure. Green denotes edges that occur in less than 7% (as advocated by FDR) of the models, blue for edges that occur in more than a third of the model, and red everything in between 7.1.3. Performance of structure learning algorithm: Variance of estimated model As a second performance metric, this section will assess the consistency of the learned structures on resamples from a data set D. Assuming that all the observation in D are identically distributed, the structure learning method should arrive at the same conclusions on the resamples, less some variance in the process. The Auto MPG Data Set5 containing various continuous and discrete car performanceattributes, from the UCI machine learning repository [27], is used to conduct the test. The data set contains 398 instances of 8 variables. For the purpose of the experiment, data sets of the same size will be sampled with replacement from the full data set 100 times. On each resample, a graph is learned using the stacking estimator on a 10% FDR level. Each subsample contains about two thirds of the original data points (with some instances repeated). This is a commonly used procedure to estimate the sampling distribution of a statistics, and which will here allow us to assess the variance in the learned graph structure. Figure 5 shows the results. On average, there are 6 links in the learned structure, hence the FDR advocates about 0.6 type 1 errors per learned model. The green lines are connections that are found less times than expected by the FDR and the blue lines are connections that are found in a large fraction of the models (over one third of the resamples). The concerning links are the ones in between, for which the null of independence is rejected more than occasionally, but not in a reasonably large fraction of the learned graphs. There are only 2 links in the model for which this occurs, so overall, the variance across learned graphs seems to be reasonably small and the learning process is hence consistent. 5 https://archive.ics.uci.edu/ml/datasets/auto+mpg 40 Figure 6: Learned graphical model structures for the Boston Housing and the Iris data set 7.2. Experiments on real data sets In this section, the outputs of the graphical model structure learning routine are shown on a selection of real world data sets. It will outline some possibilities of the user to trade off between power and computational complexity. 7.2.1. Sklearn data sets: Boston Housing and Iris Setup: One receives a data set and is interested in initial data exploration. This involves finding related sets of variables, and variables that lack obvious relationships with other variables. Boston Housing and Iris are two standard data sets from sklearn. The housing data set contains 506 samples of 14 variables, originally collected to build a model for predicting house prices (descriptions of the variables can be found online6 ). The Iris data set is a data set for predicting flower species based on petal and sepal measurements with 150 observations in 5 dimensions. Estimation will take place using the default stacking estimator. Since we are interested in initial exploration, and finding interesting groups of variables, a large FDR (20%) was chosen. Note that, unlike for other (mostly score-based) structure learning algorithms, the outcome of one experiment (the presence of an edge in the model) does not influence other experiments, and hence, false discoveries do not compromise the estimated structure additionally. The results are shown in Figure 6 (graphs drawn with NetworkX [24]). For the Boston housing data, seemingly sensible variable groupings occur. In the top left, the variables related to industrialization of a neighbourhood are shown, while on the right, demographic attributes form a cloud of related variables. For the Iris data set, while length and width of both sepal and petal are related, as expected, it seems that petal size has a higher association with class, and, in fact, width and length are independent given the class. Both of these analyses require no parameter tuning (the only non-default chosen for this experiment was the adjusted confidence level) and take very little time (less than 15 seconds). The implementation of algorithm 3 thus provides a quick, ready to use tool for initial exploratory data analysis. 7.2.2. Key short-term economic indicators (UK) Setup: One is interested in finding the relationships within a set of variables and making local conditional independence statements for sets of variables. The focus is on finding associations that we are confident about. The economic indicator data set contains monthly data of key economic indicators between February 6 http://scikit-learn.org/stable/modules/classes.html#module-sklearn.datasets 41 (a) Default Approach (b) Bagged Support Vector Machines Figure 7: Learned graphs for the Economic Indicator data set for the default approach (left) and a more powerful approach using bagged SVM’s (right) 1987 and June 2017 for the UK from the OECD statistics database [30]. The economics indicators are export and import figures, stock price and industrial production levels, overnight and 3 month Interbank interest rates, money supply (M4) and GDP. This rather small data set, around 369 instances of 9 variables, will outline the possibility of a user to trade off between computational complexity and power of the learning algorithm. As for most economic data sets, the signal to noise ratio can be quite low. Figure 7a shows the structure learned by the default approach for confidence (false discovery-rate) level of 5%, with a run time of 15 seconds. While the ground truth is not known (economic theory aside), it is apparent that many links are missed by the fast default approach. This becomes evident when considering a variable like the GDP, which (by definition) includes exports and imports, and hence some connection between the variables should exist. If there is need for a more powerful approach, a more powerful estimator can be chosen. Figure 7b show the learned structure for an estimator that resamples the data 50 times and learns 50 different Support Vector Machines with automatically tuned hyperparameters. The implementation of this is straightforward, since sklearn provides hyperparameter and bagging wrappers, so an estimator can be tuned easily and passed to the MetaEstimator. The graph shows the many edges that were found. While it is impossible to judge the correctness of the graph, it seems that some reasonable groups of variables are found, such as industrial production, exports and imports (in economic theory, countries seek to balance their trade deficit), or the grouping of stock prices and the main drivers of inflation, money supply and interest rates. Additionally, all the variables are connected in some way, which would be expected from a data set of economic indicators. The computational complexity of this approach was rather high, with a run time of about 20 minutes, however this shows how easily the user can trade off between power and complexity,without having a large effect on the false discovery-rate (as shown in Section 7.1). 42 8. Discussion This paper has introduced a novel way for multivariate conditional independence testing based on predictive inference and linked to a supervised learning workflow, which addresses many of the current issues in independence testing: • Few subjective choices: By connecting the classical task of independence testing to supervised learning, and its well-understood hyperparameter tuning capabilities, there is no need for the heuristics and manual tuning choices prevalent in current state-of-the-art conditional independence tests • Low computational complexity: By linking independence testing to one of the most-researched fields in data science, predictive modelling, the high level of power in the state-of-the-artprediction methods can be directly benefited from, and any efficiency gains in the latter directly benefits the former. • High-dimensional problems: By connecting the problem to the task of supervised learning, the method easily deals with the multivariate case, largely removing any issues concerning dimensionality of the data It is important to note, that some choices remain necessary, as is the case in statistical methodology in general. Since it is not possible to test for an infinite amount of loss functions and predictive models, one has to decide on a subset to conduct this test. The larger the subset, the higher the data requirements to arrive at a respective power level. How the outlined methodology differs from the methods reviewed in section 2 is by outlining a principled way to choose from a set of subjective choices, and by using a subset of all-star predictive modelling algorithms to ensemble over as a default, a test that is expected to be able to deal with most of the usual scenarios, given a reasonable sample size. To validate these claims, the test was bench-marked against current state-of-the-art conditional independence tests, and showed a similar or better performance in regions where the power of the tests exceeded 10%. Subsequently, the PCIT was applied in a method for learning the structure of an undirected graph best describing the independence structure of an input data set. The combination of the new conditional independence test and the structure learning algorithm address some of the current issues in graphical model structure learning: • Low computational complexity: While current exact algorithms, such as the PC-algorithm, often require a number of tests that is exponential in the number of variables, the proposed algorithm runs in quadratic time in the size of the graph. Additionally, a straightforward power-complexity trade off is provided • Exact algorithm: Unlike many scored-based methods, the algorithm does not make any strong constraints on the underlying distribution and is not subject to local optima • False-discovery rate control: FDR control is added to provide a powerful tool to control the fraction of type 1 errors when the number of variables is increasing Performance tests showed that the false-discovery rate is as advertised and the power of the test increases constantly in the number of samples. Additionally, consistency under perturbations in the data set was demonstrated. The algorithms have been distributed in the pcit package, providing users with an easy-to-use implementation to test for multivariate and conditional independence, as well as to perform graphical model structure learning with little need for manual parameter tuning. The implementations are particularly interesting for users • assessing the value of collecting additional variables for a prediction task 43 • in need of a conditional independence test for multivariate data • performing exploratory data analysis • looking for a visual way to communicate the structure in their data set There are a few ways in which the current implementation can be generalized to make for a more powerful test. The power of the conditional independence test is directly linked to the power of the supervised learning algorithm to find the optimal prediction functional and the correct loss. Currently, only the necessary minimum of 2 loss functions is implemented, one for regression tasks and one for classification tasks, but this can easily be generalized by including more losses, and checking if the baseline can be beaten with statistical significance in any of them. This would also strengthen the argument when reasoning about variables being independent, when the null hypothesis cannot be rejected. What’s more, the current methodology connected single-output prediction with FDR control to make statements about the behaviour between the joint distributions. While this procedure results in a logically sound test, the feasibility of multi-output predictors, predicting several outputs at once, should be assessed, for appropriate multivariate loss functions. On the other side, some extensions of the tests need to be conducted before a definitive conclusion can be made as to its power. In terms of performance evaluation, the power of the proposed routine was assessed for two specific cases only, for multivariate Gaussian data, and a simple conditional independence test. To get a better idea of the general feasibility of the algorithm, more scenarios need to be analyzed and performance tests conducted. Additionally, the power of the test in the context of alternative graphical model structure learning algorithms should be evaluated. 44 A. Best uninformed predictors: classification This appendix collects proofs of number of elementary computations to obtain some of the explicit statements about classification losses found in the main corpus. A.1. Misclassification loss is a probabilistic loss In this sub-section, we consider the alternative misclassification loss L : Y0 × Y → [0, 1]; (p, y) 7→ 1 − p(y), with Y being a discrete set, and Y0 being the probability simplex of probability mass functions on Y, as considered in Remark 3.2. In said remark, it is claimed that rounding a probabilistic prediction to a deterministic one never increases the generalization loss. We first prove this for the unconditional case which is equivalent to constant predictions: Lemma A.1. Let Y be an Y-valued random variable with probability mass function pY . There is always a deterministic minimizer of the expected generalization loss, i.e., there exists a pmf p0 : Y → {0, 1} such that p0 = argmin E[L(p, Y )]. p∈Y0 Proof. Let p : Y → [0, 1] be any pmf. For its expected generalization loss, one has, by substituting definitions, X X E[L(p, Y )] = p(y)(1 − pY (y)) = 1 − p(y)pY (y). y∈Y y∈Y Let y0 := argmaxpY (y) (if there are multiple maximizers, choose arbitrary). By construction, pY (y0 ) ≥ y∈Y P y∈Y p(y)pY (y). Thus, for p0 : y 7→ 1 if y = y0 ; 0 otherwise, one has E[L(p0 , Y )] = 1 − pY (y0 ) ≤ 1 − X p(y)pY (y) = E[L(p, Y )]. y∈Y Since p was arbitrary, this proves the claim. Note that Lemma A.1 does not exclude that there are other minimizers of the expected generalization loss (in general there will be an infinity), it only states that a deterministic minimizer may be found. Proposition A.2. Let X, Y, Z be a random variables taking values in X, Y, Z. Then one can make choices which always predict a deterministic class, for the following prediction functionals as considered in Section 3: (L) (i) the best uninformed predictor ωY : X → Y0 (L) (ii) the best predictor ωY |X : X → Y0 (L) (iii) the best conditionally uninformed predictor ωY |Z : X × Z → Y0 (L) (iv) the best conditional predictor ωY |X,Z : X × Z → Y0 That is, in all four cases, choices may be made where the image is always a deterministic pmf p : Y → {0, 1}. 45 Proof. (i) follows directly from Lemma A.1 which directly implies that the function y 7→ p0 is the best constant predictor and hence the best uninformed predictor, where p0 is defined as in the statement (L) (or more constructively in the proof) of Lemma A.1, thus ωY : x 7→ p0 is a possible choice. (ii) follows from noting that the Lemma A.1 and its proof remain still valid when considering the conditional under X, i.e., defining a conditional p0 : X → Y0 ; x 7→ argminE[L(p, Y )|X = x] p∈Y0 (L) and thus ωY |X : x 7→ p0 (x). (iii) and (iv) follow analogously by additional conditioning on Z in the same way. A.2. Logarithmic loss is a proper loss In this sub-section, we consider the logarithmic loss (or cross-entropy loss) L : Y0 × Y → R+ ; (p, y) 7→ − log p(y), with Y being a discrete set, and Y0 being the probability simplex of probability mass functions on Y, as considered in Remark 3.2. Proposition A.3. The expected generalization log-loss is minimized by the true distribution. I.e., let Y be random variable taking values in Y, with probability mass function pY . Then, pY = argmin E[L(y, Y )]. y∈Y0 Proof. Let p ∈ Y be arbitrary. Substituting definitions, it holds that X E[L(p, Y )] = − pY (y) log p(y) y∈Y ≥− X p(y) log p(y) y∈Y = E[L(pY , Y )], where the inequality in the middle is Gibbs’ inequality. A.3. Brier loss is a proper loss In this sub-section, we consider the Brier loss (or squared classification loss) X L : Y0 × Y → R+ ; (p, y) 7→ (1 − p(y))2 + p(y 0 )2 , y 0 6=y with Y being a discrete set, and Y0 being the probability simplex of probability mass functions on Y, as considered in Remark 3.2. Proposition A.4. The expectedBrier loss is minimized by the true distribution. I.e., let Y be random variable taking values in Y, with probability mass function pY . Then, pY = argmin E[L(y, Y )]. y∈Y0 46 Proof. Let p ∈ Y be arbitrary. By applying the binomial rule, observe that X p(y 0 )2 L(p, y) = 1 − 2p(y) + y 0 ∈Y Substituting definitions, it holds that   X X pY (y) − 2pY (y)p(y) + pY (y) p(y 0 )2  E[L(p, Y )] = y 0 ∈Y y∈Y =1−2 X pY (y)p(y) + =1− 2 pY (y) + y∈Y =1− X p(y)2 y∈Y y∈Y X X X y∈Y 2 pY (y) + y∈Y X X pY (y)2 − 2 pY (y)p(y) + y∈Y X p(y)2 y∈Y 2 (p(y) − pY (y)) . y∈Y Note that the first two terms (the 1 and the sum over pY (y)) do not depend on p, while the last term is non-negative and minimized with a value of zero if and only if p = pY . Since p was arbitrary, this proves the claim. B. Elicited statistics for regression losses This appendix collects explicit proofs of the elicited statistics for squared loss, and Q-loss (and the absolute loss which is a special case). B.1. Squared loss elicits the mean In this sub-section, we consider the univariate regression case Y = R, and the squared loss L : Y × Y → R; (y, y ∗ ) 7→ (y − y ∗ )2 . Proposition B.1. The squared loss elicits the mean. I.e., let Y be random variable taking values in Y. Then, E[Y ] = argmin E[L(y, Y )]. y∈Y Proof. Substituting definitions, it holds that E[L(y, Y )] = E[(y − Y )2 ] = E[y 2 − 2yY + Y 2 ] = y 2 − 2yE[Y ] + E[Y 2 ] = y 2 − 2yE[Y ] + E[Y 2 ] − E[Y 2 ] + E[Y 2 ] = (E[Y ] − y)2 + Var(Y ), which is the well-known derivation of the bias-variance decomposition. The first term, (E[Y ] − y)2 , is minimized whenever E[Y ] = y, and the second term, Var(Y ), does not depend on y. Thus, the sum of both terms is minimized (in y while fixing Y ) for the choice y = E[Y ]. 47 B.2. Quantile loss elicits the quantile In this sub-section, we consider the univariate regression case Y = R, and the Q-loss (or quantile loss) L : Y × Y → R; (y, y∗ ) 7→ α · m(y∗ , y) + (1 − α) · m(y, y∗ ), with m(x, z) = min(x − z, 0). Proposition B.2. The Q-loss elicits the α-quantile. I.e., let Y be random variable taking values in Y with cdf F : R → [0, 1]. Then, FY−1 (α) = argmin E[L(y, Y )]. y∈Y Proof. We first assume that Y is absolutely continuous, i.e., Y has a probability density function p : R → R+ and F is bijective. One then computes EY [Lα (y ∗ , Y )] = Z y∗ (1 − α)(y ∗ − y)p(y)dy − Z (α)(y ∗ − y)p(y)dy y∗ −∞ ∗ ∞ Z ∗ y∗ = y (F (y ) − α) + αE[Y ] − yp(y)dy −∞ ∂EY [Lα (y ∗ , Y )] ! = y ∗ p(y ∗ ) + F (y ∗ ) − α − y ∗ p(y ∗ ) = F (y ∗ ) − α = 0 ∗ ∂y =⇒ P (y ∗ ) = α ∂ 2 EY [Lα (y ∗ , Y )] = p(y ∗ ) ≥ 0 ∂(y ∗ )2 Hence, the first order condition is a minimum, minimized by the α-quantile of Y , and thus the quantile loss elicits the quantile. For general Y , note that F always exists, and thus when p appears inside integrals, the integrals welldefined. The partial derivatives may not always be defined but is the same as the sign of sufficiently small finite differences, thus the proof logic follows through for the general case. In case of jump discontinuities of F , any monotone inverse F −1 may be chosen for the statement. 48 References [1] C.C. Aggarwal. Data Classification: Algorithms and Applications. Chapman & Hall/CRC, 1st edition, 2014. [2] D. Barber. Bayesian reasoning and machine learning. Cambridge University Press, 2012. [3] L. Baringhaus and C. Franz. On a new multivariate two-sample test. Journal of Multivariate Analysis, 88:190–206, 2004. [4] Y. Benjamini and Y. Hochberg. Controlling the false discovery rate: A practical and powerful approach to multiple testing. Journal of the Royal Statistical Society. Series B (Methodological), 57:289–300, 1995. [5] Y. Benjamini and D. Yekutieli. The control of the false discovery rate in multiple testing under dependency. The Annals of Statistics, 29:1165–1188, 2001. [6] W. Bergsma. Nonparametric testing of conditional independence by means of the partial copula. ArXiv e-prints: 1101.4607, 2011. [7] W.P. Bergsma. Testing conditional independence for continuous random variables. Eurandom, 2004. [8] L. Buitinck, G. Louppe, M. Blondel, F. Pedregosa, A. Mueller, O. Grisel, V. Niculae, P. Prettenhofer, A. Gramfort, J. Grobler, et al. API design for machine learning software: experiences from the scikit-learn project. arXiv preprint arXiv:1309.0238, 2013. [9] U. Cherubini, E. Luciano, and W. Vecchiato. Copula methods in finance. John Wiley & Sons, 2004. [10] K.P. Chwialkowski, A. Ramdas, D. Sejdinovic, and A. Gretton. Fast two-sample testing with analytic representations of probability measures. In Advances in Neural Information Processing Systems 28, pages 1981–1989. Nips, 2015. [11] G.W. Corder and D.I. Foreman. Nonparametric Statistics for Non-Statisticians: A Step-by-Step Approach. John Wiley & Sons, Inc., 2009. [12] A. Dionisio, R Menezes, and D.A. Mendes. Entropy-based independence test. Nonlinear Dynamics, 44:351–357, 2006. [13] V.A. Fernández, M.D.J. Gamero, and J.M. Garcı́a. A test for the two-sample problem based on empirical characteristic functions. Computational Statistics & Data Analysis, 52:3730–3748, 2008. [14] T.M. Fragoso and F.L. Neto. Bayesian model averaging: A systematic review and conceptual classification. arXiv preprint arXiv:1509.08864, 2015. [15] J. Friedman, T. Hastie, and R. Tibshirani. The elements of statistical learning, volume 1. Springer series in statistics New York, 2001. [16] C. Genest and B. Rémillard. Test of independence and randomness based on the empirical copula process. Test, 13:335–369, 2004. [17] T. Gneiting and A.E. Raftery. Strictly proper scoring rules, prediction, and estimation. Journal of the American Statistical Association, 102:359–378, 2007. [18] A. Gretton and L. GyĂśrfi. Consistent nonparametric tests of independence. Journal of Machine Learning Research, 11:1391–1423, 2010. 49 [19] A. Gretton, O. Bousquet, A. Smola, and B. Scholkopf. Measuring statistical dependence with hilbert-schmidt norms. In ALT, volume 16, pages 63–78. Springer, 2005. [20] A. Gretton, K. Fukumizu, C.H. Teo, L. Song, B. Schölkopf, and A. Smola. A kernel statistical test of independence. In Advances in neural information processing systems, pages 585–592, 2008. [21] A. Gretton, K. Fukumizu, Z. Harchaoui, and B.K. Sriperumbudur. A fast, consistent kernel two-sample test. Advances in Neural Information Processing Systems 22, pages 673–681, 2009. [22] A Gretton, K.M. Borgwardt, M.J. Rasch, B. Schölkopf, and A. Smola. A kernel two-sample test. Journal of Machine Learning Research, 13:723–773, 2012. [23] A. Gretton, D. Sejdinovic, H. Strathmann, S. Balakrishnan, M. Pontil, K. Fukumizu, and B.K. Sriperumbudur. Optimal kernel choice for large-scale two-sample tests. In Advances in Neural Information Processing Systems 25, pages 1205–1213. Nips, 2012. [24] A.A. Hagberg, D.A. Schult, and P.J. Swart. Exploring network structure, dynamics, and function using NetworkX. In Proceedings of the 7th Python in Science Conference (SciPy2008), pages 11– 15, 2008. [25] E. Jones, T. Oliphant, and P. Peterson. SciPy: Open source scientific tools for Python, 2001. URL http://www.scipy.org/. [26] D. Koller and N. Friedman. Probabilistic graphical models: principles and techniques. MIT press, 2009. [27] M. Lichman. UCI machine learning repository, 2013. URL http://archive.ics.uci.edu/ml. [28] D. Lopez-Paz and M. Oquab. arXiv:1610.06545, 2016. Revisiting classifier two-sample tests. arXiv preprint [29] Y. Nadeau, C.and Bengio. Inference for the generalization error. Machine Learning, 52:239–281, 2003. [30] OECD. OECD statistics: Key short-term economic indicators. http://stats.oecd.org/, 2017. Accessed: 2017-08-06. [31] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825–2830, 2011. [32] S. Raschka. Mlxtend, 2016. URL http://dx.doi.org/10.5281/zenodo.594432. [33] B. Rémillard and O. Scaillet. Testing for equality between two copulas. Journal of Multivariate Analysis, 100:377–386, 2009. [34] B. Schweizer and E. F. Wolff. On nonparametric measures of dependence for random variables. The Annals of Statistics, 9:879–885, 1981. [35] Bharath K Sriperumbudur, Kenji Fukumizu, Arthur Gretton, Gert RG Lanckriet, and Bernhard Schölkopf. Kernel choice and classifiability for RKHS embeddings of probability distributions. In NIPS, pages 1750–1758, 2009. [36] W. Zaremba, A. Gretton, and M. Blaschko. B-test: A non-parametric, low variance kernel twosample test. Advances in Neural Information Processing Systems 26, pages 755–763, 2013. [37] K. Zhang, J. Peters, D. Janzing, and B. Schölkopf. Kernel-based conditional independence test and application in causal discovery. arXiv preprint arXiv:1202.3775, 2012. 50 

10math.ST

Weakly-supervised Semantic Parsing with Abstract Examples Omer Goldman∗, Veronica Latcinnik∗, Udi Naveh∗, Amir Globerson, Jonathan Berant Tel-Aviv University {omergoldman@mail,veronical@mail,ehudnave@mail,gamir@post,joberant@cs}.tau.ac.il arXiv:1711.05240v1 [cs.CL] 14 Nov 2017 Abstract Semantic parsers translate language utterances to programs, but are often trained from utterance-denotation pairs only. Consequently, parsers must overcome the problem of spuriousness at training time, where an incorrect program found at search time accidentally leads to a correct denotation. We propose that in small well-typed domains, we can semi-automatically generate an abstract representation for examples that facilitates information sharing across examples. This alleviates spuriousness, as the probability of randomly obtaining a correct answer from a program decreases across multiple examples. We test our approach on CNLVR, a challenging visual reasoning dataset, where spuriousness is central because denotations are either T RUE or FALSE, and thus random programs have high probability of leading to a correct denotation. We develop the first semantic parser for this task and reach 83.5% accuracy, a 15.7% absolute accuracy improvement compared to the best reported accuracy so far. 1 Introduction The goal of semantic parsing is to map language utterances to executable programs. Early work on statistical learning of semantic parsers utilized supervised learning, where training examples included pairs of language utterances and programs. (Zelle and Mooney, 1996; Kate et al., 2005; Zettlemoyer and Collins, 2005, 2007). However, collecting such training examples at scale has quickly turned out to be difficult, because expert annotators that are familiar with formal languages are required. This has led to a significant body of work on weakly-supervised semantic parsing (Clarke et al., 2010; Liang et al., 2011; Krishnamurthy and Mitchell, 2012; Kwiatkowski et al., 2013; Berant et al., 2013; Cai and Yates, 2013; Artzi and ∗ Authors equally contributed to this work. I: k :[[{y loc: ..., color: ’Black’, type: ’square’, x loc: ... size: 20}, ...}]] x :There is a small yellow item not touching any wall y :True z : Exist(Filter(ALL ITEMS, λx.And(And(IsYellow(x), IsSmall(x)), Not(IsTouchingWall(x, Side.Any)))))) Figure 1: Overview of our setup for visual reasoning. Given an image I that was rendered from a KB k and an utterance x, our goal is to parse x to the correct program z that results in the denotation y. Our training data includes (x, k, y) triplets. Zettlemoyer, 2013), where training examples correspond to utterance-denotation pairs. A denotation is the result of executing a program against the target knowledge-base (KB) (see Figure 1). Naturally, collecting denotations is much easier, because it can be performed by non-experts. Training semantic parsers from denotations raises two fundamental learning challenges: (a) Search: When training from denotations, the algorithm must learn to search through the huge space of programs, to find ones that execute to the correct denotations. This is a difficult search problem due to the combinatorial nature of the search space. (b) Spuriousness: Incorrect programs can easily lead to the right denotation, and thus the learner can go astray based on these programs. Of the two mentioned problems, spuriousness has attracted relatively less attention (Pasupat and Liang, 2016; Guu et al., 2017). Recently, the Cornell Natural Language for Visual Reasoning corpus (CNLVR) was released (Suhr et al., 2017), and presented a difficult problem of spuriousness. In this task, an image con- taining three boxes is shown, and in each box there are several objects of various shapes, colors and sizes. Each image is paired with a complex natural language utterance, and the goal is to determine whether the statement is true or false given the image (see Figure 1). The task comes in two flavors, where in one the input is the image itself (pixels), and in the other the input is the KB from which the image was synthesized. Because the images are fairly simple, the vision problem of translating the image to the structured KB is relatively simple. Given the input KB, it is easy to view CNLVR as a semantic parsing problem: our goal is to translate language utterances into programs that will be executed against the KB to determine their correctness (Johnson et al., 2017b; Hu et al., 2017). Because there are only two return values, it is easy to generate incorrect programs that execute to the right denotation, and thus the problem of spuriousness is significant compared to previous datasets. In this paper, we present the first semantic parser for the CNLVR dataset. We develop a formal language for visual reasoning, inspired by Johnson et al. (2017b), and a semantic parser based on recent work by Guu et al. (2017). Our main insight is that in a small and well-typed domain, we can manually cluster a small number of lexical items and generate abstract representations for utterances and programs. These representations alleviate the challenges of search and spuriousness inherent to weakly-supervised semantic parsing. By creating abstract examples, we can identify similar examples and reward programs that are correct in multiple examples, thereby substantially reducing the problem of spuriousness. This can also be done for partial programs at search time, which helps combat the search challenge. Moreover, we can generate thousands of utterance-program pairs automatically from a small number of manually created abstract examples, and train a supervised parser that will provide a good starting point for the search process in the weakly-supervised setup. Our method allows us to train a semantic parser from weak supervision, and obtain an accuracy of 83.5%, a 15.7% absolute accuracy improvement compared to state-of-the-art. All our code is publicly available at https://github.com/ udiNaveh/nlvr_tau_nlp_final_proj. 2 Setup Problem Statement Our task is as follows: given a training set of N examples {(xi , ki , yi )}N i=1 , where xi is a language utterance, ki is a KB describing objects in an image and yi ∈ {T RUE, FALSE} denotes whether the utterance is true or false in the context of the KB, our goal is to learn a semantic parser that will map new utterances-KB pairs (x, k) to a program z that is executed to the correct denotation y (see Figure 1) Programming language The original KBs in the CNLVR dataset describe an image as three sets of objects, where each object has a color, shape, size and location in absolute coordinates. We define a programming language over the KB that is more amenable to spatial reasoning, inspired by work on the CLEVR dataset (Johnson et al., 2017b). This programming language provides access to functions that allow us to check the size, shape, and color of an object, to check whether it is touching a wall, to obtain sets of items that are above and below a certain set of items, etc.1 More formally, a program is a sequence of tokens describing a possibly recursive sequence of function applications in prefix notation. Each token is either a function with fixed arity (all functions have either one or two arguments), a constant, a variable or a λ term used to define Boolean functions. Functions, constants and variables have one the following atomic types: Int, Bool, Item, Size, Shape, Color, Side (sides of a box in the image); or a composite type Set(?), and Func(?,?). Valid programs have the type Bool. Table 1 provides two examples for utterances and their correct program in the context of an image. Note that we provide commas and parenthesis in the table for readability, but they are not part of the language. We provide a full description of all program tokens, their arguments and return types in Appendix A. Unlike CLEVR, CNLVR requires substantial set-theoretic reasoning (utterances refer to various aspects of sets of items in one of the three boxes in the image), which required extending the language described by Johnson et al. (2017b) to include set operators and lambda abstraction. We manually 1 We leave the problem of learning the programming language functions from the original KB for future work. x: “There is a small yellow item not touching any wall.” z: Exist(Filter(ALL ITEMS, λx.And(And(IsYellow(x), IsSmall(x)), Not(IsTouchingWall(x, Side.Any))))) x: “One tower has a yellow base.” z: ≤(1, Count(Filter(ALL ITEMS, λx.And(IsYellow(x), IsBottom(x))))) Table 1: Examples for utterance-program pairs. Commas and parenthesis are provided for readability only. sampled 100 training examples from the training data and estimate that roughly 95% of the utterances in the training data can be expressed with this programming language. 3 Model We base our model on the semantic parser presented by Guu et al. (2017). In their work, they use an standard encoder-decoder architecture (Sutskever et al., 2014) to define a distribution pθ (z | x, k). The utterance x is encoded with a bidirectional LSTM (Hochreiter and Schmidhuber, 1997) that creates a contextualized representation hi for every utterance token xi , and the decoder is a feed-forward network combined with attention mechanism over the encoder outputs (Bahdanau et al., 2015). The feed-forward decoder takes as input the last K tokens that were decoded. More formally the probability of a program is the product of the probability of its tokens Q given the history and KB: pθ (z | x, k) = t pθ (zt | x, z1:t−1 , k) (we will drop the KB k when it is clear from context), and the probability of a decoded token is computed as follows. First, an utterance representation x̂ is computed by the BiLSTM encoder: hFi = LSTM(hFi−1 , φx (xi )), B hB i = LSTM(hi+1 , φx (xi )), hi = [hFi ; hB i ], x̂ = [hF|x| ; hB 1 ], where φx (·) provides utterance embeddings and ’;’ corresponds to concatenation. Then decoding produces the program token by token: qt = relu(Wq [x̂; zt−K−1:t−1 ]), αt ∝ exp(qt> Wα hi ), X ct = αi · hi , i > pθ (zt | x, z1:t−1 ) ∝ exp(φz (zt ) Ws [qt ; ct ]), where φz (·) provides program embeddings, zi:j = (zi , . . . , zj ), and the matrices Wq , Wα , Ws are learned parameters (along with the LSTM parameters). Search As explained, searching through the large space of programs is a fundamental challenge in weakly-supervised semantic parsing. To combat this challenge we apply several techniques. First, we use beam search at decoding time and when training from weak supervision (see Section 4), as was done in prior work (Liang et al., 2017; Guu et al., 2017). At each decoding step we maintain a beam of B program prefixes of length n, expand them exhaustively to programs of length n + 1 and keep the top B program prefixes with highest model probability. Second, we utilize our semantic typing system to only construct programs that are syntactically valid, and substantially prune the program search space (similar to the type constraints introduced by Krishnamurthy et al. (2017)). We maintain a stack that keeps track of the expected semantic type at each decoding step. The stack is initialized with the type Bool. Then, at each decoding step, only tokens that return the semantic type at the top of the stack are allowed, the stack is popped, and if the decoded token is a function, the semantic types of its arguments are pushed to the stack. This dramatically reduces the search space and guarantees that only syntactically valid programs will be produced. Figure 2 illustrates the state of the stack when decoding a program for an input utterance. Given the constrains on valid programs, our model is defined as: Y p0θ (z | x) = p0θ (zt | x, z1:t−1 ), t pθ (zt | x, z1:t−1 ) · 1(zt | z1:t−1 ) p0θ (zt | x, z1:t−1 ) = P , 0 0 z 0 pθ (z | x, z1:t−1 ) · 1(z | z1:t−1 ) where 1(zt | z1:t−1 ) indicates whether a certain program token is valid given the program prefix. Program re-ranking Our model is a locally normalized model that provides a distribution for every token emitted by the decoder. We noticed that often programs generated on the final beam do not cover well all the words that occur in the x :One tower has a yellow base. z: EqualInt 1 Count Filter ALL ITEMS λx And IsYellow x IsBottom x s: Int Set Bool Item Bool Int Int Set BoolFunc BoolFunc Bool Bool Bool Bool Item Figure 2: An example for the state of the type stack s while decoding a program z for an utterance x. Utterance “yellow” “big” “square” “3” “exactly” “top” “above” Program IsYellow IsBig IsSquare 3 Equal IsTop GetAbove Cluster C-Color C-Size C-Shape C-Num C-QuantMod C-Location C-SpaceRel Total: # 3 3 4 9 5 2 9 35 Table 2: Example mappings from utterance tokens to program tokens for the seven clusters used in the abstract representation. The rightmost column counts the number of mapping in each cluster, resulting in a total of 26 mappings. utterance. Therefore, we optionally apply a reranker that re-ranks the beam programs according to the following heuristic. For every program, we count how many of the utterance tokens the program covers, based on a small lexicon that maps partial programs to utterance phrases. Then, we return the program that has the best coverage. In Section 5 we show this results in a small boost to the final performance of the parser. In the future we plan to train a globally normalized re-ranker that will re-rank the programs in the beam and replace this heuristic. 4 Learning from Abstract Examples The main premise of this work is that in small, well-typed domains such as visual reasoning, the main challenge is handling language compositionality, as utterances can be complex. Conversely, the problem of mapping lexical items to functions and constants in the programming language is limited and can be substantially reduced using a handful of manually defined rules. Thus, if we abstract away from the actual utterance into a more general representation we can more easily generalize across examples in datasets with a small number of examples. Consider the utterances: 1. “There are exactly 3 yellow squares touching the wall.” 2. “There are at least 2 blue circles touching the wall.” While the surface forms are different in these two utterances, at an abstract level they are quite similar and we it would be useful to leverage this similarity at training time. We therefore define an abstract representation for utterances and logical forms that is suitable for spatial reasoning. We define seven abstract clusters (see Table 2) that correspond to the main spatial concepts in our domain. Then, we associate each cluster with a small lexicon that contains language-program token pairs associated with this cluster. Table 2 shows the seven clusters we use, with an example for an utterance-program token pair from the cluster, and the number of mappings in each cluster. In total, 35 mappings are used to define abstract representations. We now show how abstract examples can be employed to develop a rule-based semantic parser, a supervised semantic parser, and to combat the spuriousness and search challenges in a weaklysupervised semantic parser. Rule-based semantic parser To examine the diversity in the dataset, we created abstract representations for all 3,962 utterances in the training examples by mapping utterance tokens to their cluster label. We counted how many distinct abstract utterances exist, and found that 200 abstract utterances cover roughly half of the training examples in the original training set. To create a rule-based parser, we manually annotated 106 abstract utterances with their corresponding abstract program (including alignment between abstract tokens in the utterance and program). For example, the aforementioned utterances are both mapped to the abstract program C-QuantMod(C-Num, Count(Filter(ALL ITEMS, λx. And (And (IsC-Color(x), IsC-Shape(x), IsTouchingWall(x)))))). This defines a rule-based semantic parser: Given a new utterance x, we create its abstract representation x̄, and if it exactly matches one of our manually annotated utterances, we map it to its corresponding abstract program z̄, replace the abstract program tokens with the aligned corresponding program tokens, and obtain a final program z. Supervised semantic parser through data augmentation While the rule-based semantic parser has high precision and gauges the amount of structural variance in the data, it cannot generalize beyond observed examples. Because our clusters are small and semantically coherent, we can generate correct examples from abstract examples by sampling pairs of utteranceprogram tokens for each cluster. This is equivalent to a synchronous context-free grammar (Chiang, 2005) that has a rule for generating each manuallyannotated abstract utterance-program pair, and rules for generating synchronously utterance and program tokens from the seven clusters. We generated 4,962 (x, z) examples using this method and trained a standard supervised semantic parser by maximizing log p0θ (z | x) in the model described above. Our goal is to examine whether by training on this dataset we can generalize beyond the rule-based semantic parser. 4.1 Weakly-supervised semantic parser To train a weakly-supervised semantic parser we follow Guu et al. (2017) and treat the program z as a latent variable that is approximately marginalized. Our objective is to maximize: p(y | x) = X p0θ (z | x) · p(y | z, x) z∈Z = X p0θ (z | x) · R(z, y) z∈Z ≈ X p0θ (z | x) · R(z, y), z∈Ẑ where Z is the space of all programs and Ẑ is a subset of all programs that is found by the process of beam search. Importantly, the probability of obtaining the denotation y given x and z is defined by a binary reward function R(z, y) that is 1 iff z is “correct”. In most semantic parsers z is defined to be correct if when executing it it results in the correct denotation y. This goes a long way because often the probability of generating the correct denotation y by a randomly generated z is low. However, in CNLVR, we observe binary denotations, and so spuriousness is a central problem as previously explained. To alleviate the spuriousness problem we utilize an interesting property of the data: the same utterance appears 4 times with 4 different images. If a program is spurious it is likely that it will yield the wrong denotation in one of those 4 images. Thus, we can re-define each training example to be (x, {(kj , yj )}4j=1 ), where each utterance x is paired with 4 different KBs and the denotations of the utterance with respect to these KBs. Then, we maximize p({yj }4j=1 | x, {kj }4j=1 ) by maximizing the objective above, except that R(z, {yj }4j=1 ) = 1 iff the denotation of z is correct for all four KBs. This dramatically reduces the problem of spuriousness, as the chance of randomly obtaining a correct deno1 tation goes down from 21 to 16 . This approach is reminiscent of the method proposed by Pasupat and Liang (2016), where random permutations of Wikipedia tables are shown to crowdsourcing workers to eliminate spurious programs over those tables. 4.2 Tackling spuriousness by caching abstract examples Our approach for tackling spuriousness described above hinges on the fact that the same utterance appears in the dataset multiple times. However, as explained above, many examples are similar to one another at an abstract level even if they are not identical. Thus, a natural idea is to combat spuriousness by sharing information across utterances that have the same abstract representation. We construct a cache that maps abstract utterances to abstract programs. After obtaining a beam of final programs for every training example utterance x, we add to the cache every abstract utterance-program pair (x̄, z̄), and record whether it obtained positive reward. Thus, for every (x̄, z̄) pair we can estimate a probability psp for whether it obtains positive reward. To construct an abstract example from from an utterance-program pair in the beam we perform the following procedure. First, we create an abstract utterance by replacing utterance tokens with their cluster label, as in the rule-based semantic parser. Then, we go over every program token, and replace it with an abstract cluster if the utterance contains a token that is mapped to this program token according to one of the 35 mappings from Table 2. This also provides an alignment from ab- stract program tokens to abstract utterance tokens that is necessary when utilizing the cache. We propose two variants for taking advantage of the constructed cache: 1. Full program retrieval: For every utterance x, we construct an abstract utterance x̄, retrieve from the cache the top-D abstract programs z̄ based on psp , compute the programs z using the alignments from program tokens to utterance tokens, and add those D programs to the final beam. 2. Program prefix retrieval: At every decoding step, let Zt be the beam of decoded programs at step t and Zt+1 be the beam of decoded programs computed at step t + 1. For every utterance x, we find the top-D cashed programs Z̄ as in the former variant. Then at each step t, for every abstract program z̄ ∈ Z̄ we add z̄1:t to Zt , and exhaustively search over their possible continuations as well when constructing Zt+1 . This allows the parser to potentially construct new programs that are not in the cache already in subsequent steps. This combats spuriousness but also the search challenge, because we add to the beam promising program prefixes that might have fallen off of it in early decoding steps. Liang et al. (2017) also suggested to use a cache for adding correct programs to the beam, but they did not use any example abstraction, and only stored full programs as a way to stabilize learning. 5 Experimental Evaluation Experimental details Our encoder is a BiLSTM where the hidden state dimension for each LSTM is 30. The decoder is a feed-forward network with one hidden layer of dimension 50, that obtains the last 4 decoded tokens as input (in addition to encoder information). Utterance tokens (words) embeddings are of dimension 12, beam size B = 40 and we retrieve D = 10 programs from the cache. We initialize word embeddings by running the CBOW algorithm (Mikolov et al., 2013) on the training data and then optimize them end-to-end. In the weakly-supervised parser we encourage exploration at training time by using meritocratic gradient updates with β = 0.5 (Guu et al., 2017), where the weight of a program with correct denotation is an interpolation of the model probability distribution and a uniform distribution. Using standard policy gradient (β = 1) resulted in slightly lower performance. Models are implemented in TensorFlow. In the supervised parser, the parameters other than word embeddings are randomly initialized (Glorot and Bengio, 2010), while in the weakly-supervised parser we initialize parameters using the learned parameters of the supervised model to warm-start the model. We use the Adam optimizer (Kingma and Ba, 2014), a learning rate of 0.001, a mini-batch size of 8 examples, and trained for 7 epochs for the supervised model and 15 epochs for the weakly-supervised model. Pre-processing Because the number of utterances is relatively small for training a neural model, we take the following steps to reduce sparsity. We lowercase all utterance tokens, and also use their lemmatized form. We also use spelling correction to replace words that contain typos. Last, we replace 8 domain-specific words with their more frequent synonym according to the training set (e.g., “object” is replaced with “item”, and “wall” is replaced with “edge”). After pre-processing we replace every word that occurs less than 5 times with an UNK symbol. Evaluation We evaluate on the public development set and test set of CNLVR as well as on the hidden test set by submitting our code to the authors of CNLVR. The standard evaluation metric is accuracy, that is, how many examples are correctly classified. In addition, we report a consistency measure, which is the proportion of utterances for which the resulting program has the correct denotation for all 4 images/KBs. We believe this is an important measure as it captures better whether a model can consistently produce a correct answer for an utterance. Baselines We compare our models to the M A JORITY baseline that always picks the majority class (T RUE in our case). We also compare to the best results reported by Suhr et al. (2017) when taking the KB as input, which is a maximum entropy classifier (M AX E NT). We compare these previous results to the following models: • RULE: The rule-based parser described in Section 4. We back-off to the M AJORITY baseline when the parser cannot parse the utterance. Model M AJORITY M AX E NT RULE S UP. S UP.+R ERANK W EAK S UP. W.+R ERANK Dev Acc. Con. 55.3 68.0 66.0 29.2 72.8 43.1 76.5 53.9 79.5 57.3 83.5 63.7 Test-P Acc. Con. 56.2 67.7 66.3 32.7 69.5 40.2 74.4 50.4 77.8 52.6 80.4 59.0 Test-H Acc. Con. 55.4 67.8 83.5 64.7 Table 3: Results of baselines and models on the development set, public test set (Test-P), and hidden test set (Test-H). For each of our models we report both accuracy and consistency. • S UP.: A supervised semantic parser trained on 3,721 examples automatically generated from abstract examples (and validated on the remaining 1,241 examples). • W EAK S UP.: A weakly-supervised semantic parser described in Section 4. • +R ERANK: For both the supervised parser and the weakly-supervised parser, we add the heuristic global re-ranking describe in Section 3. Main results Table 5 describes our main results. Our weakly-supervised semantic parser combined with global re-ranking of the beam (W.+R ERANK) obtains 83.5 accuracy and 64.7 consistency on the test set, improving accuracy by 15.7 points compared to state-of-the-art. This is the only model we evaluated on the hidden test set according to the rules of CNLVR evaluation.2 The rule-based parser, RULE, is able to handle a non-negligible part of the data and its accuracy is less than two points lower than the M AX E NT baseline on then development set and public test set. Training a supervised parser improves accuracy to 69.5 on the public test set, showing that generalizing from generated examples is better than memorizing the manually-defined patterns. Our weakly-supervised parser significantly improves over S UP., reaching an accuracy of 77.8 on the test before re-ranking. For both S UP. and W EAK S UP. re-ranking substantially improves both accuracy and consistency. To conclude, we develop a semantic parsing approach for the CNLVR dataset. We show that a supervised parser trained by generating examples 2 https://github.com/clic-lab/nlvr# running-on-the-held-out-test-set Model -A BSTRACTION -B EAM C ACHE -E VERY S TEP B EAM C ACHE Dev. Acc. Con. 56.7 9.0 72.8 46.4 81.1 61.4 Table 4: Results of ablations of our main models on the development set. Explanation for the nature of the models is in the body of the paper. from abstract examples already improves over approaches that do not use a programming language, and that we are able to overcome the challenges of search and spuriousness, and develop a weakly supervised semantic parser that substantially improves state-of-the-art. 5.1 Analysis We analyze our results by running several ablations on our best model W.+R ERANK on the development set. To examine the importance of abstract examples, we run our weakly-supervised parser from scratch without pre-training a supervised parser on abstract examples and without using a cache for abstract examples (-A BSTRACTION). We find this parser is unable to overcome the challenges of training from weak supervision. Training and development accuracies remain very low throughout training, and thus we stopped training after 9 epochs, seeing that performance is close to the M AJORITY baseline. To further examine the importance of the cache itself, we train the weakly-supervised parser by starting from the initialized supervised parser, but without the cache of abstract examples (B EAM C ACHE). We find that this weaklysupervised parser performs worse than the supervised parser it was initialized from by a few points. that is, we are still unable to improve learning by training from denotations. Last, we use a beam cache with full program retrieval only (Section 4), that is, programs are added to the beam only after decoding terminates, rather than at every decoding step (-E VERY S TEP B EAM C ACHE). We see that this already results in good performance that is substantially higher than S UP., but is still 2.4 points worse than our best performing model on the development set. 6 Related Work Training semantic parsers from denotations has been one of the most popular training schemes for scaling semantic parsers since the beginning of the decade. Early work focused on traditional loglinear models (Clarke et al., 2010; Liang et al., 2011; Kwiatkowski et al., 2013; Berant et al., 2013), but recently denotations have been used to train neural semantic parsers as well (Liang et al., 2017; Krishnamurthy et al., 2017; Rabinovich et al., 2017; Cheng et al., 2017). Visual reasoning has attracted considerable attention, with the release of datasets such as VQA (Antol et al., 2015) and CLEVR (Johnson et al., 2017a). The advantage of CNLVR is that the language utterances are both natural and compositional. Treating the problem of visual reasoning as an end-to-end semantic parsing problem has been previously done on CLEVR (Hu et al., 2017; Johnson et al., 2017b). Our method for generating training examples resembles ideas for data re-combination suggested recently (Jia and Liang, 2016), where examples are generated automatically by replacing entities with their categories. While spuriousness is central to semantic parsing when denotations are not very informative, there has been relatively little work on explicitly tackling it. Pasupat and Liang (2015) used manual rules to prune unlikely programs on the W IK I TABLE Q UESTIONS dataset, and then later utilized crowdsourcing (Pasupat and Liang, 2016) to eliminate spurious programs. Guu et al. (2017) proposed R ANDOMER, a method for increasing exploration and handling spuriousness by adding randomness to beam search and a proposing a “meritocratic” weighting scheme for gradients. In our work we found that random exploration during beam search did not improve results while meritocratic updates slightly improved performance. 7 Conclusion In this work we presented the first semantic parser for the CNLVR dataset, working on structured representations as input. Our main insight is that in small, well-typed domains we can generate abstract examples that can help combat the difficulties of training a parser from delayed supervision. First, we use abstract examples to semiautomatically generate utterance-program pairs that help warm-start our parameters, thereby reducing the difficult search challenge of finding correct programs with random parameters. Second, we focus on an abstract representation of ex- amples, which allows us to tackle spuriousness and alleviate search, by sharing information about promising programs between different examples. Our semantic parser substantially improves results compared to previous state-of-the-art on the CNLVR dataset. In future work, we plan to investigate training a globally-normalized re-ranker over the beam of programs generated by the our locally-normalized sequence-to-sequence model, and also train a parser over a simpler programming language that is closer to the representation in the target KB. References S. Antol, A. Agrawal, J. Lu, M. Mitchell, D. Batra, C. L. Zitnick, and D. Parikh. 2015. Vqa: Visual question answering. In International Conference on Computer Vision (ICCV). pages 2425–2433. Y. Artzi and L. Zettlemoyer. 2013. Weakly supervised learning of semantic parsers for mapping instructions to actions. Transactions of the Association for Computational Linguistics (TACL) 1:49–62. D. Bahdanau, K. Cho, and Y. Bengio. 2015. Neural machine translation by jointly learning to align and translate. In International Conference on Learning Representations (ICLR). J. Berant, A. Chou, R. Frostig, and P. Liang. 2013. Semantic parsing on Freebase from question-answer pairs. In Empirical Methods in Natural Language Processing (EMNLP). Q. Cai and A. Yates. 2013. Large-scale semantic parsing via schema matching and lexicon extension. In Association for Computational Linguistics (ACL). J. Cheng, S. Reddy, V. Saraswat, and M. Lapata. 2017. Learning structured natural language representations for semantic parsing. In Association for Computational Linguistics (ACL). D. Chiang. 2005. A hierarchical phrase-based model for statistical machine translation. In Association for Computational Linguistics (ACL). pages 263–270. J. Clarke, D. Goldwasser, M. Chang, and D. Roth. 2010. Driving semantic parsing from the world’s response. In Computational Natural Language Learning (CoNLL). pages 18–27. X. Glorot and Y. Bengio. 2010. Understanding the difficulty of training deep feedforward neural networks. In International Conference on Artificial Intelligence and Statistics. K. Guu, P. Pasupat, E. Z. Liu, and P. Liang. 2017. From language to programs: Bridging reinforcement learning and maximum marginal likelihood. In Association for Computational Linguistics (ACL). S. Hochreiter and J. Schmidhuber. 1997. Long shortterm memory. Neural Computation 9(8):1735– 1780. P. Pasupat and P. Liang. 2015. Compositional semantic parsing on semi-structured tables. In Association for Computational Linguistics (ACL). R. Hu, J. Andreas, M. Rohrbach, T. Darrell, and K. Saenko. 2017. Learning to reason: End-toend module networks for visual question answering. In International Conference on Computer Vision (ICCV). P. Pasupat and P. Liang. 2016. Inferring logical forms from denotations. In Association for Computational Linguistics (ACL). R. Jia and P. Liang. 2016. Data recombination for neural semantic parsing. In Association for Computational Linguistics (ACL). J. Johnson, B. Hariharan, L. van der Maaten, L. Fei-Fei, C. L. Zitnick, and R. Girshick. 2017a. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Computer Vision and Pattern Recognition (CVPR). J. Johnson, B. Hariharan, L. van der Maaten, J. Hoffman, L. Fei-Fei, C. L. Zitnick, and R. Girshick. 2017b. Inferring and executing programs for visual reasoning. In International Conference on Computer Vision (ICCV). R. J. Kate, Y. W. Wong, and R. J. Mooney. 2005. Learning to transform natural to formal languages. In Association for the Advancement of Artificial Intelligence (AAAI). pages 1062–1068. D. Kingma and J. Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 . J. Krishnamurthy, P. Dasigi, and M. Gardner. 2017. Neural semantic parsing with type constraints for semi-structured tables. In Empirical Methods in Natural Language Processing (EMNLP). J. Krishnamurthy and T. Mitchell. 2012. Weakly supervised training of semantic parsers. In Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP/CoNLL). pages 754–765. T. Kwiatkowski, E. Choi, Y. Artzi, and L. Zettlemoyer. 2013. Scaling semantic parsers with on-the-fly ontology matching. In Empirical Methods in Natural Language Processing (EMNLP). C. Liang, J. Berant, Q. Le, and K. D. F. N. Lao. 2017. Neural symbolic machines: Learning semantic parsers on Freebase with weak supervision. In Association for Computational Linguistics (ACL). P. Liang, M. I. Jordan, and D. Klein. 2011. Learning dependency-based compositional semantics. In Association for Computational Linguistics (ACL). pages 590–599. T. Mikolov, K. Chen, G. Corrado, and Jeffrey. 2013. Efficient estimation of word representations in vector space. arXiv . M. Rabinovich, M. Stern, and D. Klein. 2017. Abstract syntax networks for code generation and semantic parsing. In Association for Computational Linguistics (ACL). A. Suhr, M. Lewis, J. Yeh, and Y. Artzi. 2017. A corpus of natural language for visual reasoning. In Association for Computational Linguistics (ACL). I. Sutskever, O. Vinyals, and Q. V. Le. 2014. Sequence to sequence learning with neural networks. In Advances in Neural Information Processing Systems (NIPS). pages 3104–3112. M. Zelle and R. J. Mooney. 1996. Learning to parse database queries using inductive logic programming. In Association for the Advancement of Artificial Intelligence (AAAI). pages 1050–1055. L. S. Zettlemoyer and M. Collins. 2005. Learning to map sentences to logical form: Structured classification with probabilistic categorial grammars. In Uncertainty in Artificial Intelligence (UAI). pages 658– 666. L. S. Zettlemoyer and M. Collins. 2007. Online learning of relaxed CCG grammars for parsing to logical form. In Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP/CoNLL). pages 678–687. Appendix A: Programming Language Logical Function Exist Filter Count And, Or Not GreaterThan, LessThan, GreaterEqual, LessEqual Equal QueryColor QueryShape QuerySize GetAbove, GetBelow, GetTouching IsYellow, IsBlack, IsBlue, IsCircle, IsTriangle, IsSquare, IsBig, IsMedium, IsSmall, IsTop, IsBottom IsTouchingWall All, Any Union Select AllSame Return Type Bool Set(?) Int Bool Bool Bool Argument(s) Type(s) Set(?) Set(?),BoolFunc(?) Set(?) Bool,Bool Bool Int,Int Bool Set(Color) Set(Shape) Set(Size) Set(Item) (?,?) Set(Item) Set(Item) Set(Item) Set(Item) Bool Item Bool Bool Set(?) Set(?),Int Bool Item,Side Set(?),BoolFunc(?) Set(Set(?)) Set(Set(?) Set(?) Constants ALL ITEMS ALL BOXES Color.Blue, Color.Yellow, Color.Black Shape.Circle, Shape.Triangle, Shape.Square Side.Top, Side.Bottom, Side.Left, Side.Right, Side.Any Type Set(Item) Set(Set(Item)) Color Shape Side 

2cs.AI

Metric Perturbation Resilience arXiv:1607.06442v1 [cs.DS] 21 Jul 2016 Konstantin Makarychev Microsoft Research Yury Makarychev∗ TTIC Abstract We study the notion of perturbation resilience introduced by Bilu and Linial (2010) and Awasthi, Blum, and Sheffet (2012). A clustering problem is α-perturbation resilient if the optimal clustering does not change when we perturb all distances by a factor of at most α. We consider a class of clustering problems with center-based objectives, which includes such problems as k-means, k-median, and k-center, and give an exact algorithm for clustering 2perturbation resilient instances. Our result improves upon the result of Balcan and Liang √ (2016), who gave an algorithm for clustering 1 + 2 ≈ 2.41 perturbation resilient instances. Our result is tight in the sense that no polynomial-time algorithm can solve (2 − ε)-perturbation resilient instances unless N P = RP , as was shown by Balcan, Haghtalab, and White (2016). We show that the algorithm works on instances satisfying a slightly weaker and more natural condition than perturbation resilience, which we call metric perturbation resilience. ∗ Supported by NSF CAREER award CCF-1150062 and NSF award IIS-1302662. 1 Introduction In this paper, we present an exact algorithm for solving 2-perturbation resilient instances of clustering problems with natural center-based objectives. The notion of perturbation resilience was proposed by Bilu and Linial [5] and Awasthi, Blum, and Sheffet [1]. Informally, an instance is perturbation resilient if the optimal solution does not change when we perturb the instance. The definition was introduced in the context of beyond-the-worst-case analysis and aims to capture a wide class of real-life instances that are computationally easier than worst-case instances. As several authors argue, in instances arising in practice, the optimal solution is often significantly better than all other solutions and thus does not change if we slightly perturb the instance [5, 4]. It was shown that perturbation resilient instances of such problems as k-center, k-means, k-median, clustering problems with center-based and min-sum objectives, Max Cut, Minimum Multiway Cut, and TSP (with a sufficiently large value of the resilience parameter) can be solved exactly in polynomial-time [1, 2, 3, 4, 5, 6, 7]; while the worst-case instances of these problems are NP-hard. In a clustering problem, we are given a metric space (X, d) and an integer parameter k; our goal is to partition X into k clusters C1 , . . . , Ck so as to minimize the objective function H(C1 , . . . , Ck ; d) (which depends on the problem at hand). The most well-studied and, perhaps, most interesting clustering objectives are k-means, k-median, and k-center. These objectives are defined as follows. Given a clustering C1 , . . . , Ck , the objective is equal to the minimum over all choices of centers c1 ∈ C1 ,. . . ,ck ∈ Ck of the following functions: Hmeans (C1 , . . . , Ck ; d) = k X X d(u, ci )2 ; Hmedian (C1 , . . . , Ck ; d) = k X X d(u, ci ); Hcenter (C1 , . . . , Ck ; d) = i=1 u∈Ci i=1 u∈Ci max max d(u, ci ). i∈{1,...,k} u∈Ci Note that in the optimal solution each cluster Ci consists of those vertices u that are closer to ci than to other centers cj ; i.e. (C1 , . . . , Ck ) is the Voronoi partition of X with centers c1 , . . . , ck . We refer to objectives satisfying this property as center-based objectives. We study two closely related classes of center-based objectives – separable center-based objectives and natural centerbased objectives (which we discuss below and formally define in Section 2). We note that k-means, k-median, and k-center are separable and natural center-based objectives. Now we formally define perturbation resilience. Definition 1.1. Consider an instance I = ((X, d), H, k) of a clustering problem on a metric space (X, d) with objective H. An instance ((X, d′ ), H, k) is an α-perturbation of I if d(u, v) ≤ d′ (u, v) ≤ αd(u, v) for every u, v ∈ X; here, d′ does not have to be a metric. An instance I is α-perturbation resilient if every αperturbation of I has the same optimal clustering as I. We will refer to α as the perturbation resilience parameter. This definition does not require that the perturbation d′ is a metric (d′ does not have to satisfy the triangle inequality). It is more natural to consider only metric perturbations of I — those perturbations in which d′ is a metric. In this paper, we give the definition of metric perturbation resilience, in which we do require that d′ is a metric (see Definition 2.4). (Every α-perturbation resilient instance is also α-metric perturbation resilient.) 1 Known results. Awasthi, Blum, and Sheffet [1] initiated the study of perturbation resilient instances of clustering problems. They offered the definition of a separable center-based objective (s.c.b.o.) and introduced an important center proximity property (see Definition 2.5). They presented an exact algorithm for solving 3-perturbation resilient instances of clustering problems √ with s.c.b.o.; they also gave an algorithm for (2 + 3)-perturbation resilient instances of clustering problems with s.c.b.o. that have Steiner points. Additionally, they showed that α-perturbation resilient instances of k-median with Steiner points are N P -hard when α < 3. Later, Balcan and √ Liang [3] designed an exact algorithm for (1 + 2)-perturbation resilient instances of problems with s.c.b.o., improving the result of Awasthi, Blum, and Sheffet. Balcan and Liang also studied clustering with the min-sum objective and (α, ε)-perturbation resilience (a weaker notion of perturbation resilience, which we do not discuss in this paper). Recently, Balcan, Haghtalab, and White [2] designed an algorithm for 2-perturbation resilient instances of symmetric and asymmetric k-center and showed that there is no polynomial-time algorithm for (2 − ε) perturbation resilience instances unless N P = RP . They also gave an algorithm for 2-perturbation resilient instances of problems with s.c.b.o. satisfying a strong additional condition of cluster verifiability. √ To summarize, the best known algorithm for arbitrary s.c.b.o. requires that the instance be 1 + 2 ≈ 2.4142 perturbation resilient [3]; the best known algorithm for k-center requires that the instance be 2-perturbation resilient, and this result cannot be improved [2]. We refer the reader to [6] for an overview of known results for stable instances of combinatorial optimization problems. Our results. We define a class of clustering problems with natural center-based objectives; an objective is a natural center-based objective if it is representable in the following form. For some functions fc and gu (r) (fc is a function of c, gu (r) is a function of u and r; intuitively, fci is the cost of having a center at ci and gu (r) is the cost of connecting u to a center at distance r from u), we have k   X X fci + gu (d(u, ci )) , H(C1 , . . . , Ck ; d) = min ci ∈Ci or H(C1 , . . . , Ck ; d) = min max ci ∈Ci i=1  u∈Ci  max fci , max gu (d(u, ci )) . i∈{1,...,k} i∈{1,...,k} u∈Ci This class includes such problems as k-center, k-means, k-median, and metric facility location (for the first three problems: fc = 0 and gu (r) equals r 2 , r, and r respectively). We present a polynomial-time algorithm for 2-metric perturbation resilient instances of clustering problems with natural center based objectives;√thus, we improve the known requirement on the perturbation resilience parameter α from α ≥ 1 + 2 ≈ 2.4142 to α ≥ 2 and relax the condition on instances from a stronger α-perturbation resilience condition to a weaker and more natural α-metric perturbation resilience condition. Our result is optimal, since even (2 − ε)-perturbation resilient instances of k-center cannot be solved in polynomial time unless N P = RP √ [2]. In particular, our result improves the requirement for k-median and k-means from α ≥ 1 + 2 to α ≥ 2. Theorem 1.2. There exists a polynomial-time algorithm that given any 2-metric perturbation resilient instance ((X, d), H, k) of a clustering problem with natural center-based objective, returns the (exact) optimal clustering of X. Our algorithm is quite simple. It first runs the single-linkage algorithm to construct the minimum spanning tree on points of X and then partitions the minimum spanning tree into k clusters 2 using dynamic programming. We note that Awasthi, Blum, and Sheffet [1] also used the singlelinkage algorithm together with dynamic programming to cluster 3-perturbation resilient instances. However, their approach is substantially different from ours: They first find a hierarchical clustering of X using the single-linkage algorithm and then pick k optimal clusters from this hierarchical clustering. This approach fails for α-perturbation resilient instances with α < 3 (see [1]). That is why, we do not use the single-linkage hierarchical clustering in our algorithm, and, instead, partition the minimum spanning tree. We note that the definitions of separable and natural center-based objectives are different. However, in Section 5 we consider a slightly strengthened definition of s.c.b.o. and show that every s.c.b.o., under this new definition, is also a natural center-based objective; thus, our result applies to it. We are not aware of any non-pathological objective that satisfies the definition of s.c.b.o. but is not a natural center-based objective. Finally, we consider clustering with s.c.b.o. and show that the optimal solution for every αmetric perturbation resilient instance satisfies the α-center proximity property; previously, that was only known for α-perturbation resilient instances [1]. Our result implies that the algorithms by Balcan and Liang [3] and Balcan, Haghtalab, and White [2] for clustering with s.c.b.o. and k-center, respectively, apply not only to α-perturbation resilient but also α-metric perturbation resilient instances. Overview. In Section 2, we formally define key notions used in this paper. Then, in Section 3, we prove that the optimal solution for every α-metric perturbation resilient instance satisfies the α-center proximity property. We use this result later in the analysis of our algorithm; also, as we noted above, it is of independent interest and implies that previously known algorithms from [2, 3] work under the metric permutation resilience assumption. In Section 4, we present our algorithm for solving 2-perturbation resilient instances of problems with natural center-based objectives. The algorithm consists of two steps. In the first step, we find a minimum spanning tree in the metric space (X, d). In the second step, we partition the tree into k subtrees using dynamic programming and then output the corresponding clustering of X. Finally, in Section 5, we prove that if a s.c.b.o. satisfies some additional properties, then it is a natural center-based objective. 2 Preliminaries In this section, we formally define key notions used in this paper: a clustering problem, α-metric perturbation resilience, separable center-based and natural center-based objectives. Definition 2.1. An instance of a clustering problem is the tuple ((X, d), H, k) of a metric space (X, d), objective function H, and integer number k > 1. The objective H is a function that given a partition of X into k sets C1 , . . . , Ck and a metric d on X returns a nonnegative real number, which we call the cost of the partition. Given an instance of a clustering problem ((X, d), H, k), our goal is to partition X into disjoint sets C1 , . . . , Ck , so as to minimize H(C1 , . . . , Ck ; d). Now we define center-based and separable center-based objectives. We note that the definitions of Awasthi et al. [1] make several implicit assumptions that we make explicit here. Definition 2.2. We say that H is a center-based objective function if the following properties hold. 1. Given a subset S ⊂ X and distance dS on S, we can find the optimal center c ∈ S for S, or, if there is more than one choice of an optimal center, a set of optimal centers center(S, dS ) ⊂ S. (In the former case, center(S, dS ) = {c}). 3 2. For every metric d on X, there exists an optimal clustering C1 , . . . , Ck of X (i.e., a clustering that minimizes H(C1 , . . . , Ck ; d)) such that for every two distinct clusters Ci , Cj , every point p in Ci and centers ci ∈ center(Ci , d|Ci ) and cj ∈ center(Cj , d|Cj ) we have d(p, ci ) ≤ d(p, cj ). Furthermore, if the optimal clustering C1 , . . . , Ck is unique, then d(p, ci ) < d(p, cj ). The objective is separable if, additionally, we can define individual cluster scores so that the following holds. • The cost of the clustering is either the sum (for separable sum-objectives) or maximum (for separable max-objectives) of the cluster scores. • The score H(C, d|C ) of each cluster C depends only on C and d|C , and can be computed in polynomial time. In this paper, we consider a slightly narrower class of natural center-based objectives (which we described in the introduction). The class contains most important center-based objectives: kcenter, k-means, and k-median, as well as the facility location objective. We are not aware of any reasonable center-based objective that is not a natural center-based objective. Now, we formally define natural center-based objectives. Definition 2.3. We say that H is a natural center-based objective function for a ground set X, if there exist functions f : X → R and g : X × R → R such that H(C1 , . . . , Ck ; d) = min ci ∈Ci or H(C1 , . . . , Ck ; d) = min max ci ∈Ci k   X X gu (d(u, ci )) , fci + i=1  u∈Ci  max fci , max gu (d(u, ci )) . i∈{1,...,k} (1) i∈{1,...,k} u∈Ci (2) We require that the functions f and g be computable in polynomial time, and that gu be nondecreasing for every u ∈ X. We call points ci the centers of the clustering. P We could have also defined an objective with an ℓp -aggregate function H = ( ki=1 fcpi +
P p 1/p , but it is equivalent to (1) with f ′ (c) = f p (c) and g ′ (u, r) = g p (u, r). u∈Ci gu (d(u, ci )) Now, we formally define a metric perturbation and metric perturbation resilience. Since we do not require that the objective H is homogeneous as a function of the metric d, we introduce two perturbation resilience parameters α1 and α2 in the definition, which specify by how much the distances can be contracted and expanded, respectively, in the perturbed instances. Definition 2.4. Consider a metric space (X, d). We say that a metric d′ is an (α1 , α2 )-metric per′ turbation of (X, d) if α−1 1 d(u, v) ≤ d (u, v) ≤ α2 d(u, v) for every u, v ∈ X. An instance ((X, d), H, k) is (α1 , α2 )-metric perturbation resilient if for every (α1 , α2 )-metric perturbation d′ of d, the unique optimal clustering for ((X, d′ ), H, k) is the same as for ((X, d′ ), H, k). Note that the centers of clusters in the optimal solutions for ((X, d), H, k) and ((X, d′ ), H, k) may differ. We say that an instance (H, (X, d), k) is α-metric perturbation resilient if it is (α, 1)metric perturbation resilient. Observe that if instance ((X, d), H, k) is (α1 , α2 )-metric perturbation resilient, then ((X, λd), H, k) is (λα1 , λα2 )-metric perturbation resilient for λ ∈ [α−1 1 , α2 ]. Particularly, if ((X, d), H, k) is (α1 , α2 )metric perturbation resilient, then ((X, α2 d), H, k) is (α, 1)-metric perturbation resilient and the 4 optimal solution for ((X, α2 d), H, k) is the same as for ((X, d), H, k). Thus, to solve an (α1 , α2 )metric perturbation resilient instance ((X, λd), H, k), it suffices to solve α = (α1 α2 ) metric perturbation resilient instance ((X, α2 d), H, k). Consequently, we will only consider α-metric perturbation resilient instances in this paper. Finally, we recall the definition of the α-center proximity property, introduced in [1]. Definition 2.5. We say that a clustering C1 , . . . , Ck of X with centers c1 , . . . , ck satisfies the α-center proximity property if for all i 6= j and p ∈ Ci , we have d(p, cj ) > αd(p, ci ). 3 Center Proximity for Metric Perturbation Resilience In this section, we prove that the (unique) optimal solution to an α-metric perturbation resilient clustering problem satisfies the α-center proximity property. Our proof is similar to the proof of Awasthi, Blum, and Sheffet, who showed that the optimal solution to a (non-metric) α-perturbation resilient clustering problem satisfies the α-center proximity property. Theorem 3.1. Consider an α-metric perturbation resilient clustering problem ((X, d), H, k) with a center-based objective. Let C1 , . . . , Ck be the unique optimal solution; and let c1 , . . . , ck be a set of centers of C1 , . . . , Ck (that is, each ci ∈ center(Ci , d|Ci )). Then, the following α-center proximity property holds: for all i 6= j and p ∈ Ci , we have d(p, cj ) > αd(p, ci ). Proof. Suppose that d(p, cj ) ≤ αd(p, ci ). Let r ∗ = d(p, ci ). Define a new metric d′ as follows. Consider the complete graph on X. Assign length len(u, v) = d(u, v) to each edge (u, v) other than (p, cj ). Assign length len(p, cj ) = r ∗ to the edge (p, cj ). Let metric d′ (u, v) be the shortest path metric on the complete graph on X with edge lengths len(u, v). Note that d(p, cj ) ≥ d(p, ci ) = r ∗ since p ∈ Ci and C1 , . . . , Ck is an optimal clustering. Hence, for every (u, v): len(u, v) ≤ d(u, v) and d′ (u, v) ≤ d(u, v). It is easy to see that d′ (u, v) = min(d(u, v), d(u, p) + r ∗ + d(cj , v), d(v, p) + r ∗ + d(cj , u)). Observe that since the ratio d(u, v)/len(u, v) is at most d(p, cj )/r ∗ ≤ α for all edges (u, v), we have d(u, v)/d′ (u, v) ≤ α for all u and v. Hence, d(u, v) ≤ αd′ (u, v) ≤ αd(u, v), and consequently, d′ is an (α, 1)-metric perturbation of d. We now show that d′ is equal to d within the cluster Ci and within the cluster Cj . Lemma 3.2. For all u, v ∈ Ci , we have d(u, v) = d′ (u, v), and for all u, v ∈ Cj , we have d(u, v) = d′ (u, v). Proof. I. Consider two points u, v in Ci . We need to show that d(u, v) = d′ (u, v). It suffices to prove that d(u, v) ≤ min(d(u, p) + r ∗ + d(cj , v), d(v, p) + r ∗ + d(cj , u)). Assume without loss of generality that d(u, p) + r ∗ + d(cj , v) ≤ d(v, p) + r ∗ + d(cj , u). We have d(u, p) + r ∗ + d(cj , v) = d(u, p) + d(p, ci ) + d(cj , v) ≥ d(u, ci ) + d(cj , v). Since v ∈ Ci , we have d(v, ci ) ≤ d(v, cj ), and thus d(u, p) + r ∗ + d(cj , v) ≥ d(u, ci ) + d(ci , v) ≥ d(u, v). 5 II. Consider two points u, v in Cj . Similarly to the previous case, we need to show that d(u, v) ≤ d(u, p) + r ∗ + d(cj , v). Since now u ∈ Cj , we have d(u, cj ) ≤ d(u, ci ). Thus,
d(u, p)+r ∗ +d(cj , v) = d(u, p)+d(p, ci ) +d(cj , v) ≥ d(u, ci )+d(cj , v) ≥ d(u, cj )+d(cj , v) ≥ d(u, v). By the definition of α-metric perturbation stability, the optimal clusterings for metrics d and d′ are the same. By Lemma 3.2, the distance functions d and d′ are equal within the clusters Ci and Cj . Hence, the centers of Ci and Cj w.r.t. metric d′ are also points ci and cj , respectively (see Definition 2.2, item 1). Thus, by the definition of center-based objective, d′ (ci , p) < d′ (cj , p), and, consequently, d(ci , p) = d′ (ci , p) < d′ (cj , p) = r ∗ = d(ci , p). We get a contradiction, which finishes the proof. Corollary 3.3. Consider a 2-metric perturbation resilient instance. Let C1 , . . . , Ck be an optimal clustering with centers c1 , . . . , ck . Then each point u in Ci is closer to ci than to any point v outside of Ci . Proof. Suppose that v ∈ Cj for some j 6= i. By the triangle inequality, d(u, v) ≥ d(u, cj ) − d(cj , v). By Lemma 3.1, d(u, cj ) > 2d(u, ci ) and d(cj , v) < d(ci , v)/2. Thus, d(u, v) ≥ d(u, cj ) − d(cj , v) > 2d(u, ci ) − d(ci , u) + d(u, v) d(ci , v) ≥ 2d(u, ci ) − . 2 2 In the last inequality, we used the triangle inequality d(ci , v) ≤ d(ci , u) + d(u, v). Rearranging the terms, we get d(u, v) > d(ci , u). 4 Clustering Algorithm In this section, we present our algorithm for solving 2-metric perturbation resilient instances of clustering problems with natural center-based objectives. Our algorithm is based on single–linkage clustering: first, we find a minimum spanning tree on the metric space X (e.g., using Kruskal’s algorithm) and then run a dynamic programming algorithm on the spanning tree to find the clusters. We describe the two steps of the algorithm in Sections 4.1 and 4.2. 4.1 Minimum Spanning Tree At the first phase of the algorithm, we construct a minimum spanning tree on the points of the metric space using Kruskal’s algorithm. Kruskal’s algorithm maintains a collection of trees. Initially, each tree is a singleton point. At every step, the algorithm finds two points closest to each other that belong to different trees and adds an edges between them. The algorithm terminates when all points belong to the same tree. We denote the obtained spanning tree by T . The key observation is that each cluster Ci forms a subtree of the spanning tree T . Lemma 4.1. Each cluster Ci forms a subtree of the spanning tree T . In other words, the unique path between every two vertices u, v ∈ Ci does not leave the cluster Ci . 6 Proof. Let ci be the center of Ci . We show that the (unique) path p from u to ci lies in Ci , and, therefore, the statement of Lemma 4.1 holds. Let u′ be the next vertex after u on the path p. Consider the step at which Kruskal’s algorithm added the edge (u, u′ ). At that step, u and ci were in distinct connected components (as p is the only path connecting u and ci ). Thus, d(u, u′ ) ≤ d(u, ci ) as otherwise the algorithm would have added the edge (u, ci ) instead of (u, u′ ). By Corollary 3.3, the inequality d(u, u′ ) ≤ d(u, ci ) implies that u′ belongs to Ci . Proceeding by induction we conclude that all vertices on the path p belong to Ci . 4.2 Dynamic Programming Algorithm At the second phase, we use dynamic programming to compute the optimal clustering. We root the tree T at an arbitrary vertex. We denote its root by root and the subtree rooted at u by Tu . We first assume that the tree is binary. Later, we explain how to transform any tree into a binary tree by adding dummy vertices. The algorithm partitions the tree into (non-empty) subtrees P1 , . . . , Pk and assigns a center ci ∈ Pi to all vertices in the subtree Pi so as to minimize the objective: k X fci + i=1 k X X gu (d(u, ci )). (3) i=1 u∈Pi Lemma 4.1 implies that the optimal partitioning of X is the solution to this problem. Let costu (k′ , c) be the minimum cost (3) of partitioning the subtree Tu into k′ subtrees P1 , . . . , Pk and choosing k centers c1 , . . . , ck so that the following conditions hold: 1. u ∈ P1 and c1 = c (we denote the tree that contains u by P1 and require that its center be c). 2. if c1 ∈ Tu , then c1 ∈ P1 (if the center c1 of P1 lies in Tu , then it must be in P1 ); 3. ci ∈ Pi for i > 1 (the center ci of every other tree Pi lies in Pi ). That is, we assume that u belongs to the first subtree P1 and that c is the center for P1 . Every center ci must belong to the corresponding set Pi except for c1 . However, if c1 ∈ Tu , then c1 ∈ P1 . Denote the children of vertex u by lu and ru (recall that we assume that the tree is binary). The cost costu (k′ , c) is computed using the following recursive formulas: if c ∈ / Tlu ∪ Tru , then costu (j, c) =fc + gu (d(c, u)) + min (4) ′ ′ ′′ ′ ′ ′′ ′′ ′ ′′ ′ ′′ min{costlu (j , c ) + costru (j , c ) : j + j = j − 1, c ∈ Tlu , c ∈ Tru }, ′′ ′ (6) ′ ′′ ′′ (7) min{costlu (j , c ) + costru (j , c) − fc : j + j = j, c ∈ Tlu }, ′ ′′ ′′ (5) ′ min{costlu (j , c) + costru (j , c ) − fc : j + j = j, c ∈ Tru },
min{costlu (j ′ , c) + costru (j ′′ , c) − 2fc : j ′ + j ′′ = j + 1} . (8) If c ∈ Tlu , then we remove lines (5) and (6) from the formula. If c ∈ Tru , then we remove lines (5) and (7) from the formula. The first term fc + gu (d(c, u)) is the cost of opening a center in c and assigning u to c. The lines (5– 8) correspond to the following cases: (5) neither lu nor ru is in P1 ; they are assigned to (trees Pi and Pj with centers) c′ and c′′ , (6) lu is not in P1 , but ru is in P1 ; they are assigned to c′ and c, (7) lu is in P1 , but ru is not in P1 ; they are assigned to c and c′′ , 7 (8) both lu and ru are in P1 ; they are assigned to c. For leaves, we set costu (1, c) = fc + gu (d(u, c)) and costu (j, c) = ∞ for j > 1. Note that if we want to find a partitioning of T into at most k subtrees, we can use a slightly simpler dynamic program. It is easy to verify that the formulas above hold. The cost of the optimal partitioning of T into k subtrees equals minc∈X costroot (k, c). We now explain how to transform the tree T into a binary tree. If a vertex u has more than two children, we add new dummy vertices between u and its children by repeating the following procedure: take a vertex u having more than two children; pick any two children of u: u1 and u2 ; create a new child v of u and rehang subtrees Tu1 and Tu2 to the vertex v. We forbid opening centers in dummy vertices v by setting the opening cost fv to be infinity. We set the assignment costs gv to be 0. Note that Lemma 4.1 still holds for the new tree if we place every dummy vertex in the same part Pi as its parent. 5 Universal Separable Center-Based Clustering Objectives In this section, we show that every separable center-based objective, satisfying some additional properties, is a natural center-based objective. To this end, we define a universal center-based objective and show that every universal center-based objective is a natural center-based objective. Loosely speaking, a universal center-based objective is a center-based objective that satisfies two properties, which we will discuss now: • An arbitrary center-based objective is defined on a specific set of points X and can be used to compute the cost of clustering only of the set X (given a metric d on X). In contrast, a universal objective can be used to compute the cost of clustering of any ground set X. • Recall that in every optimal clustering with a center-based objective each point u ∈ X is closer to the center of its own cluster than to the center of any other cluster. If a partition is not optimal, some points might be closer to the centers of clusters that do not contain them than to the centers of their own clusters. Then, we can move such points to other clusters so as to minimize their distance to the cluster centers. In fact, one of the two steps of Lloyd’s algorithm does exactly this; hence, we call such a transformation a Lloyd’s improvement. We slightly strengthen the definition of a center-based objective by requiring that a universal objective not increase when we make a Lloyd’s improvement of any – not necessarily optimal – clustering. Now we give a few auxiliary definitions and then formally define universal center-based objectives. Since data sets used in applications are usually labeled, we will consider “labeled metric spaces”. We will assume that the cost of clustering of X may depend on the distances between the points in X and point labels (but not on the identities of points). Definition 5.1 (Labeled Metric Space). A metric space labeled with a set of labels L is a pair ((X, d), l), where (X, d) is a metric space, and l : X → L is a function that assigns a label to each point in X. Definition 5.2 (Isomorphic Labeled Metric Spaces). We say that two metric spaces ((X ′ , d′ ), l′ ) and ((X ′′ , d′′ ), l′′ ) labeled with the same set L are isomorphic if there exists an isometry ϕ : X ′ → X ′′ (i.e., ϕ is a bijection that preserves distances: d′ (u, v) = d′′ (ϕ(u), ϕ(v)) for all u, v ∈ X ′ ) that preserves labels; i.e. l′ (u) = l′′ (ϕ(u)) for all u ∈ X ′ . 8 We note that in the definition above the set L may be infinite. We denote the restriction of ((X, d), l) to a non-empty subset C ⊂ X by ((X, d), l)|C : ((X, d), l)|C = ((C, d|C ), l|C ). Note that the restriction of a metric set labeled with L to a cluster C is also a metric set labeled with L. Definition 5.3. Consider a clustering problem ((X, d), H, k) with a center-based objective. We say that a clustering C1′ , . . . , Ck′ is a Lloyd’s improvement of a clustering C1 , . . . , Ck if there exists a set of centers c1 , . . . , ck of C1 , . . . , Ck (i.e., each ci ∈ center(Ci , d)) such that • ci ∈ Ci′ (a Lloyd’s improvement does not move the centers to other clusters) • for every x ∈ X: if x ∈ Ci and x ∈ Cj′ , then d(x, cj ) ≤ d(x, ci ) (a Lloyd’s improvement may move point x from Ci to Cj′ only if d(c, cj ) ≤ d(x, ci )). Definition 5.4 (Universal Objective). We say that H is a universal center-based clustering objective for a label set L, if for every metric space ((X, d), l) labeled with L the problem ((X, d), Hl , k) is a clustering problem with a separable center-based objective and the following two conditions hold. 1. Cluster scores Hl are universal (“can be used on any metric space”): Given any finite metric space ((C, d), l) labeled with L, the function Hl (C, d) returns a real number – the cost of C; and Hl′ (C ′ , d′ ) = Hl′′ (C ′′ , d′′ ) for any two isomorphic labeled metric spaces ((C ′ , d′ ), l′ ) and ((C ′′ , d′′ ), l′′ ). 2. If C1′ , . . . , Ck′ is a Lloyd’s improvement of C1 , . . . , Ck , then H(C1′ , . . . , Ck′ ; d) ≤ H(C1 , . . . , Ck ; d). Note that every natural center-based objective is a universal objective. The label set is the set of pairs (f, g), where f ∈ R is a real number; g is a nondecreasing function from R≥0 to R. Every point x ∈ X is assigned the label l(x) = (fx , hx ). The score of a cluster C equals   X gx (d(x, c)) ; (9) Hl (C, d) = min fc + c∈C x∈C
 Hl (C, d) = min max fc , max gx (d(x, c) . c∈C  x∈C (10) It is easy to see that Lloyd’s improvements may only decrease the cost of a clustering, since the functions gx are non-decreasing. We now show that every clustering problem with universal separable center-based objectives is a problem with natural center-based objectives. Theorem 5.5. I. Let ((X, d), H, k) be a clustering problem with a universal center-based separable sum-objective. Then, the scoring function Hl can be represented as (9) for some nondecreasing functions fPand g such that the minimum is attained when c is a center of C; and thus H(C1 , . . . , Ck ; d) = ki=1 Hl (Ci ; d|Ci ) is a natural center-based objective. II. Let ((X, d), H, k) be a clustering problem with a universal center-based separable max-objective. If the cost of any singleton cluster {x} equals 0, then the scoring function Hl can be represented as (10) for some nondecreasing functions f and g such that the minimum is attained when c is a center of C; ; and thus H(C1 , . . . , Ck ; d) = maxi∈{1,...,k} Hl (Ci ; d|Ci ) is a natural center-based objective. Proof. I. For every label a ∈ L define two special labeled metric spaces Ma and Ma,a,r . The metric space Ma is the metric space on a single point labeled with a; the metric space Ma,a,r is the metric space on two points at distance r that are both labeled with a. Consider a point x ∈ X labeled with a. Let fx = H(Ma ) and gx (r) = H(Ma,a,r ) − fx . We claim that the cost of any cluster C equals X
min fc + gx (d(x, c)) , c∈C x∈C\{c} 9 and the minimum is attained when c is a center of C. Consequently, equation (9) holds for functions f˜c = fc − gc (0) and gx . To prove the claim, we construct a new metric space Cp′ for every p ∈ C. The metric space Cp′ consists of all points of C and new points x′ , which are “duplicates” of points x ∈ C \ {p}. We do not change distances between points in C and labels on the points from C. We copy the label of each x to x′ ; i.e., we let l(x′ ) = l(x). We place x′ at the same location as the point p. That is, we let d(x′ , y) = d(p, y) for all x, y ∈ C. Consider two clusterings of Cp′ . The first clustering partitions Cp′ into the cluster C and singleton clusters {x′ } (there is a singleton cluster for each x ∈ C \ {p}). The cost of each singleton cluster {x′ } equals fx , since {x′ } is isomorphic to Ml(x) . Hence, the cost of the first clustering equals X Hl (C, d|C ) + fx . (11) x∈C\{p} The second clustering partitions C ′ into pairs of points {x, x′ } for each x ∈ C \{p} and the singleton cluster {p}. The cost of each cluster {x, x′ } equals H(Ma,a,d(x,x′ ) ) = gx (d(x, x′ )) + fx = gx (d(x, p)) + fx , since the cluster {x, x′ } is isomorphic to Ma,a,d(x,x′ ) . The cost of {p} equals fp . Hence, the cost of the second clustering equals X
fx + gx (d(x, c)) . (12) fp + x∈C\{p} x′ Note that the points x and in each cluster {x, x′ } are symmetric, so we may assume that x′ is the center of the cluster {x, x′ }. Observe that the first clustering is a Lloyd’s improvement of the second clustering: To get the first clustering from the second clustering we move x from each cluster {x, x′ } to the singleton cluster {p}. We are allowed to do that since d(x, p) = d(x, x′ ). Hence, the cost of the second clustering is upper bounded by the cost of the first clustering. We have from (11) and (12), X X X
fx + gx (d(x, c)) − fx = fp + gx (d(x, c)). (13) Hl (C, d|C ) ≤ fp + x∈C\{p} x∈C\{p} x∈C\{p} On the other hand, if p is the center of C, then the second clustering is a Lloyd’s improvement of the first one. Hence, (13) is an equality when p is the center of C. It remains to show that functions gx are non-decreasing. We need to prove that for every r1 ≥ r2 ≥ 0, gx (r1 ) ≥ gx (r2 ). We create a metric space on three points x′ , x′′ , x′′′ each having the same label as x. Let d(x′ , x′′ ) = r1 , d(x′′ , x′′′ ) = r2 , d(x′ , x′′′ ) = r1 + r2 . The cost of clustering {x′ , x′′ }, {x′′′ } equals 2fx + gx (r1 ); the cost of clustering {x′ }, {x′′ , x′′′ } equals 2fx + gx (r2 ). Since r1 ≥ r2 , the second clustering is a Lloyd’s improvement of the first one. Hence, 2fx + gx (r2 ) ≤ 2fx + gx (r1 ) and gx (r2 ) ≤ gx (r1 ). This completes the proof of part I. The proof of part II is similar. References [1] Pranjal Awasthi, Avrim Blum, and Or Sheffet. Center-based clustering under perturbation stability. Information Processing Letters, 112(1):49–54, 2012. [2] Maria-Florina Balcan, Nika Haghtalab, and Colin White. Symmetric and asymmetric k-center clustering under stability. In Proceedings of the International Colloquium on Automata, Languages, and Programming, 2016. 10 [3] Maria-Florina Balcan and Yingyu Liang. Clustering under perturbation resilience. SIAM Journal on Computing, 2016. [4] Yonatan Bilu, Amit Daniely, Nati Linial, and Michael Saks. On the practically interesting instances of maxcut. In Proceedings of the Symposium on Theoretical Aspects of Computer Science, pages 526–537, 2013. [5] Yonatan Bilu and Nathan Linial. Are stable instances easy? Science, pages 332–341, 2010. In Innovations in Computer [6] Konstantin Makarychev, Yury Makarychev, and Aravindan Vijayaraghavan. Bilu—Linial stable instances of Max Cut and Minimum Multiway Cut. In Proceedings of the Symposium on Discrete Algorithms, pages 890–906, 2014. [7] Matúš Mihalák, Marcel Schöngens, Rastislav Šrámek, and Peter Widmayer. On the complexity of the metric tsp under stability considerations. In SOFSEM 2011: Theory and Practice of Computer Science, pages 382–393. Springer, 2011. 11 

8cs.DS

(Preprint) Control of a Bucket-Wheel for Surface Mining of Asteroids and Small-Bodies Ravi teja Nallapu,* Erik Asphaug,† and Jekan Thangavelautham‡ Near Earth Asteroids (NEAs) are thought to contain a wealth of resources, including water, iron, titanium, nickel, platinum and silicates. Future space missions that can exploit these resources by performing In-Situ Resource Utilization (ISRU) gain substantial benefit in terms of range, payload capacity and mission flexibility. Compared to the Moon or Mars, the milligravity on some asteroids demands a fraction of the energy for digging and accessing hydrated regolith just below the surface. However, asteroids and small-bodies, because of their low gravity present a major challenge in landing, surface excavation and resource capture. These challenges have resulted in adoption of a “touch and go techniques”, like the upcoming Osiris-rex sample-return mission. Previous asteroid excavation efforts have focused on discrete capture events (an extension of sampling technology) or whole-asteroid capture and processing. This paper analyzes the control of a bucket-wheel design for asteroid or smallbody excavation. Our study focuses on system design of two counter rotating bucket-wheels that are attached to a hovering spacecraft. Regolith is excavated and heated to 1000oC to extract water. The water in turn is electrolyzed to produce hydrogen and oxygen for rocket fuel. We analyze control techniques to maximize traction of the bucket-wheels on the asteroid surface and minimize lift-off the surface, together with methods to dig deeper into the asteroid surface. Our studies combine analytical models, with simulation and hardware testing. For initial evaluation of material-spacecraft dynamics and mechanics, we assume lunar-like regolith for bulk density, particle size and cohesion. Our early studies point towards a promising pathway towards refinement of this technology for demonstration aboard a future space mission. INTRODUCTION Asteroids are known to contain large reserves of water, iron, titanium, nickel, platinum, palladium and silicates1. In-Situ Resource Utilization (ISRU)2 aims at extracting these resources onsite, which if successful, can improve our current space exploration capabilities. While there has been progress reported in ISRU research for applications on the lunar surfaces3 and Mars4, the low-gravity environments on asteroids, on the other hand, presents additional challenges in performing ISRU5. One of the major challenges includes lift-off due to excavator thrusting off the small body. * PhD Student, Space and Terrestrial Robotic Exploration Laboratory, Arizona State University, 781 E. Terrace Mall, Tempe, AZ. † Professor and Ronald Greeley Chair of Planetary Science, Space and Terrestrial Robotic Exploration Laboratory, Arizona State University, 781 E. Terrace Mall, Tempe, AZ ‡ Assistant Professor, Space and Terrestrial Robotic Exploration Laboratory, Arizona State University, 781 E. Terrace Mall, Tempe, AZ 1 In this paper we focus on use of bucket-wheels to perform excavation. Bucket-wheels have been identified as one of the best tools for excavation in low-gravity environments, beating out bulldozers and front-loaders10,13. This paper addresses the problem of excavator lift-off which is the predominant concern on very low-gravity environments; by considering a dual bucket wheel excavator on the surface of Martian moon Phobos. Phobos as shown in Figure 1 has an escape velocity of about 11.1 m/s. Hence if any of the excavation mechanisms experience a force that can impart this speed, the mechanism would be thrown off the surface of Phobos. Figure 1. Martian moon Phobos is a promising target for ISRU. In this paper, we present a bucketwheel excavator system that has two counter rotating bucket wheels, like RASSOR developed by NASA KSC6, and can apply the necessary vertical plunge force. The counter rotation ensures that no horizontal velocities are imparted, and the vertical plungers ensure the compensation for lift-off due to any reactive-excavation forces. The paper starts with introduction to the bucket-wheel system design followed by optimization of its size, operating parameters and power source 7,8. We then present dynamical and control models of the system and a simulation of the proposed mechanism, followed by preliminary results and conclusions. BUCKET WHEEL SYSTEM DESIGN The design of a bucketwheel is a multi-disciplinary problem. Several forces, geometry, and performance requirements are considered. To understand the role of an excavator, a diagram of the full ISRU system is shown in Figure 2. The excavation system is powered by solar insolation. The solar panels generate the power based on the incident radiation, which is used to charge an onboard system battery that powers various subsystems. The ISRU unit is tasked with excavating regolith, extracting water by heating regolith and electrolyzing hydrogen and oxygen from the water. The hydrogen and oxygen would be used as propellant for rockets on their journey to earth or to power fuel cells power supplies during eclipse14,15,16. A design of a bucket-wheel and ISRU system was provided by Nallapu et al.8 (Figure 3) where the geometric parameters of the bucket wheel, including number of buckets and their dimensions were obtained by constraining the total operating power to 10 kW. 2 Figure 2. Systems diagram for the proposed ISRU system Figure 3. Proposed excavator with counter rotating bucket wheels This proposed unit consists of 2 counter rotating bucket-wheels that cancel out horizontal motion. The unit is also capable of compensating for any lift-off vertical forces by exerting a radial thrust as shown in Figure 3. SYSTEM DYNAMICS Excavation dynamics Consider a bucket wheel excavator, plunging the bucket into the regolith as shown in Figure 4. Let the wheel be applied torque, τ, and be spinning with a speed, ω. During the excavation, the wheel experiences the resistive forces from the regolith, Freg, normal to the cutting surface. Let Fsup, a force applied to counter-act any lift-off experienced. 3 Figure 4. Excavation modelling The resistive regolith forces are modeled based on a Luth-Wismer model9. The Luth-Wismer model separates the regolith forces as a non-cohesive force (sand) and pure cohesive (clay) forces. Skonieczny et al.10 formulates these forces as: (1) and, (2) The force acts normal to the cutting surface, and the total force is given by: (3) Equations of motion This method of separating the forces along the horizontal and vertical axis is similar in procedure used by Park11 and Xiao et al.12. The equations of motion for the system shown in Figure 3 are therefore given by: (4) (5) (6) (7) (8) (9) where F1 and F2 are the forces applied to counteract any lift-off forces on wheels 1 and 2 respectively. The forces Fr1 and Fr2 are the resistive forces on regolith acting on wheels 1 and 2 respec- 4 tively, and whose magnitude is calculated by Equation 3. Additionally, a random disturbance force in the range [0, Freg/2] was added to the magnitudes of Fr1 and Fr2 to account for any unmodeled dynamics. Finally, τ1 and τ2 represent the control torques exerted on the wheels 1 and 2 respectively. SIMULATION The design of the wheel as mentioned earlier was to maintain the total operational power under 10 kW. This required that the wheels’ spin at about 3 RPM8, which is the value of ωdes in Equations 11 and 12. Additionally, the translator motion needs to be confined vertically downwards, and no horizontal motion was desired. To realize this, a proportional-derivative (PD) controller was applied to the vertical forces, and wheel torques as given by Equations 10-12. The control law given by Equation 10 is only applied when y>0. The gains Ky, Kvy, K1, K2, and Kx are chosen by repetitively running simulations and picking the values that give satisfactory performance. (10) (11) (12) Equations 4 through 5 were propagated with an ODE4513 solver in MATLAB with the following parameters shown in Table 1. Table 1. Simulation Parameters. Simulation Parameter # Wheels Value 2 # Buckets per wheel 24 Mass of wheel (kg) 5 D (m) 0.622 W (m) 6.31E-02 β1 (degrees) 10 β2(degrees) -10 V (m/s) 0.12 3 1880 p (kg/m ) 2 g (m/s ) 0.0057 c (Pa) 147 Specific heat of the material (J/[Kg-Celsius]) 1,430 Surface temperature of the material (Celsius) 200 5 Water extraction temperature (Celsius) 1,000 Water content in the regolith (%) 10 Kx 1 Ky 0.9 Kvy 90,000 K1 4,000 K2 4,000 RESULTS This section presents the results of simulations described above. These simulations were run for a time length of 100 seconds with a step length of 0.1 seconds. Excavator Displacement The displacement response of the excavator is shown in Figure 5. As seen here, the displacement is confined to the negative y direction, while the counter rotating wheels ensure that there is no horizontal displacement. Figure 5. Target position of bucket-wheel Angular Velocity Response Angular velocity response of the 2 bucket wheels are shown in Figure 6. As seen here the 2 wheels reach the desired speed of 3.3 RPM almost instantly while rotating in opposite directions. 6 Figure 6. Angular velocity response Wheel Motion The rotations of the two bucket wheels are shown in Figure 7. As expected, their magnitudes are equal in opposite directions because of the counter rotations of the wheel. Figure 7. Wheel rotations CONCLUSIONS In this paper, we present the significance of ISRU in extracting resources from small-bodies such as Martian moon Phobos. Phobos is strategic location and a potential base-camp for a future human mission to Mars. This paper identifies one of the principle challenges of ISRU, namely lift-off of the excavation system in low-gravity environments such as Phobos. A dynamics and control model of a 10 kW excavator and ISRU system is presented, and its equations of motion are then propagated with computer simulations. The simulations show the proposed bucket-wheel excavator can indeed be stable on an asteroid surface, and is not susceptible to liftoff forces. Further work will be required in testing algorithms in a relevant environment and in identifying the required sensors and actuators to maintain a steady-pace of excavation on the surface of the small-body. 7 REFERENCES 1 W. F. Bottke, Asteroids III, University of Arizona Press, Tucson, Arizona 2002. 2 K. Zacny, P. Chu, G. Paulsen, et al. 'Mobile in-Situ Water Extractor (MISWE) for Mars, Moon, and Asteroids in Situ Resource Utilization', AIAA SPACE Conference, 2012. 3 P. Mueller, and R. H. King. 'Trade Study of Excavation Tools and Equipment for Lunar Outpost Development and ISRU', Space Technology and Applications International Forum (STAIF 2008), vol. 969, 2008, pp. 237-244. 4 G. B. Sanders, A. Paz, L. Oryshchyn, et al. 'Mars ISRU for Production of Mission Critical Consumables – Options, Recent Studies, and Current State of the Art', AIAA SPACE 2015 Conference and Exposition, 2015. 5 A. Probst, G. Peytavi, B. Eissfeller, et al. 'Mission Concept Selection for an Asteroid Mining Mission', Aircraft Engineering and Aerospace Technology, vol. 88, no. 3, 2016, pp. 458-470. 6 R. Mueller, R. E. Cox, T. Ebert, et al. 'Regolith Advanced Surface Systems Operations Robot (RASSOR)', 2013 IEEE Aerospace Conference, 2013, pp. 1-12. 7 T. Nakamura, and B. K. Smith. 'Solar Thermal System for Lunar ISRU Applications: Development and Field Operation at Mauna Kea, HI', Proceedings of SPIE, vol. 8124, no. 1, 2011. 8 R. T. Nallapu, A. Thoesen, L. Garvie, E. Asphaug, J. Thangavelautham. “Optimized Bucket Wheel Design for Asteroid Excavation,” International Astronautic Congress, 2016, Guadalajara, Mexico, IAF. 9 H.J. Luth and R.D. Wismer, 'Performance of Plane Soil Cutting Blades in Sand’. Journal of Terramechanics, vol. 9, no. 4, 1973, pp. 67-67. 10 K. Skonieczny,. 'Lightweight Robotic Excavation', PhD Thesis, Carnegie Mellon University, 2013. 11 B. Park. 'Development of a Virtual Reality Excavator Simulator: A Mathematical Model of Excavator Digging and a Calculation Methodology', 2002. 12 Thangavelautham, J., Abu El Samid, N., Grouchy, P., Earon E., Fu, T., Nagrani, N., D'Eleuterio, G.M.T., “Evolving Multirobot Excavation Controllers and Choice of Platforms Using Artificial Neural Tissue Controllers,” Proceedings of the IEEE Symposium on Computational Intelligence for Robotics and Automation, 2009, DOI: 10.1109/CIRA.2009.542319 13 J. Thangavelautham, K. Law, T. Fu, N. Abu El Samid, A. Smith, G. M.T. D’Eleuterio, “Autonomous Multirobot Excavation for Lunar Applications,” Robotica, pp. 1-39, 2017. 14 J. Thangavelautham and S. Dubowsky, “On the Catalytic Degradation in Fuel Cell Power Supplies for Long-Life Mobile Field Sensors.” Journal of Fuel Cells: Fundamental to Systems, pp. 181-195, 2013. 15 P. Iora, J. Thangavelautham, “Design of a mobile PEM power backup system through detailed dynamic and control analysis,” International Journal of Hydrogen Energy, vol. 37, no. 22, pp. 17191–17202, 2012. 16 J. Thangavelautham, D. Strawser, M. Cheung, S. Dubowsky, “Lithium Hydride Powered PEM Fuel Cells for LongDuration Small Mobile Robotic Missions,” IEEE International Conference on Robotics and Automation (ICRA), St. Paul, Minnesota, 2012. 8 

3cs.SY

Opinion dynamics in social networks with stubborn agents: An issue-based perspective ? Ye Tian a , Long Wang b,∗ arXiv:1609.03465v2 [cs.SY] 13 Aug 2017 a Center for Complex Systems, School of Mechano-electronic Engineering, Xidian University, Xi’an 710071, P. R. China b Center for Systems and Control, College of Engineering, Peking University, Beijing 100871, P. R. China Abstract Classic models on opinion dynamics usually focus on a group of agents forming their opinions interactively over a single issue. Yet generally agreement can not be achieved over a single issue when agents are not completely open to interpersonal influence. In this paper, opinion formation in social networks with stubborn agents is considered over issue sequences. The social network with stubborn agents is described by the Friedkin-Johnsen (F-J) model where agents are stubborn to their initial opinions. Firstly, we propose a sufficient and necessary condition in terms of network topology for convergence of the F-J model over a single issue. Secondly, opinion formation of the F-J model is investigated over issue sequences. Our analysis establishes connections between the interpersonal influence network and the network describing the relationship of agents’ initial opinions for successive issues. Taking advantage of these connections, we derive the sufficient and necessary condition for the F-J model to achieve opinion consensus and form clusters over issue sequences, respectively. Finally, we consider a more general scenario where each agent has bounded confidence in forming its initial opinion. By analyzing the evolution of agents’ ultimate opinions for each issue over issue sequences, we prove that the connectivity of the state-dependent network is preserved in this setting. Then the conditions for agents to achieve opinion consensus over issue sequences are established. Simulation examples are provided to illustrate the effectiveness of our theoretical results. Key words: Opinion dynamics, issue sequences, path-dependence, convergence, confidence bound 1 Introduction Recently, opinion dynamics has attracted much attention of researchers from various disciplines, such as applied mathematics, economics, social psychology, control theory, etc., due to its broad applications in modeling and explaining complex phenomena in social and artificial networks (Degroot, 1974; Lamport et al., 1982; Hegselmann & Krause, 2002; Acemoglu et al., 2010; Frasca et al., 2015; Ravazzi et al., 2015). In a social network, agents form opinions on various political, economic and social issues according to the information they received from neighbors determined by the network topology (Jadbabaie et al., 2003; Olfati-Saber & Murray, 2004; Ren & Beard, 2005) or the confidence/influence bound (Blondel et al., 2009; Mirtabatabaei & Bullo, 2012; Jing et al., 2016). A fundamental question in opinion dynamics is: how do the ? This work was supported by National Science Foundation of China (Grant Nos. 61533001 and 61375120) and the Fundamental Research Funds for the Central University (Grant No. JBZ170401). ∗ Corresponding author : Long Wang Email addresses: tinybeta7.1@gmail.com (Ye Tian), longwang@pku.edu.cn (Long Wang). Preprint submitted to Automatica network structure and opinions’ initial distribution influence the diffusion and aggregation of scattered opinions in the process of opinion formation? In Degroot (1974), a model is presented to characterize the process of a group of agents reaching opinion consensus on a common issue by pooling their subjective opinions, which is known as the Degroot model. The interactions between agents are described by a stochastic matrix which can be regarded as the one-step transition probability matrix of a Markov chain and some sufficient conditions for achieving opinion consensus are provided. To further investigate how the interpersonal influence contributes to the opinion formation, the work in Friedkin & Johnsen (1999) extends the Degroot model by introducing a diagonal matrix which represents agents’ susceptibilities to interpersonal influence. In the Degroot model and the Friedkin-Johnsen (F-J) model, interactions between agents are specified by given networks. Different from these linear models, in Hegselmann & Krause (2002), the authors present a nonlinear model in which agents have bounded confidence for others. In the Hegselmann-Krause (H-K) model, two agents are said to be connected if and only if the difference between their opinions is smaller than a given confidence bound, which means that the network topology of the H-K model is state-dependent. 15 August 2017 about convergence of the F-J model over a single issue usually require that the interpersonal influence network is undirected and connected or its adjacent matrix is strictly row-substochastic (Ghaderi & Srikant, 2014; Mirtabatabaei et al., 2014). We propose a milder sufficient and necessary condition in terms of network topology to guarantee that opinions of agents converge to constant values. Several properties of agents’ ultimate opinions for a single issue are provided. Secondly, we study opinion formation of the F-J model over a sequence of interdependent issues inspired by the path-dependence theory (North, 1990; Page, 2006; MacKay et al., 2006; Egidi & Narduzzo, 1997). By virtue of the inherent coherence between basic assumption of the F-J model and the path-dependence theory, we assume that the factor of each agent’s cognitive inertia over two interdependent issues equals to its stubborn factor. Then, each agent will form its initial opinion for the next issue by making a tradeoff between its initial opinion for the last interdependent issue and other agents’ initial opinions for the next issue. The connections between the interpersonal influence network and the network which characterizes the relationship of agents’ initial opinions for successive issues are established, and the sufficient and necessary condition for the F-J model to achieve opinion consensus or form clusters over issue sequences is proposed. Finally, we consider the more general case in which an information assimilation mechanism is employed to weaken agents’ interpersonal influence. We assume that each agent maintains a confidence bound in forming its initial opinions. This assumption is consistent with the reported echo-chamber effect (Sunstein, 2012; Dandekar et al., 2013), i.e., people usually assimilate information in a selective way: they tend to give considerable weight to the information supporting their initial opinions, and dismiss the undermining information meanwhile. Connectivity preservation of the modified F-J model is analyzed. Then we derive the conditions for achieving opinion consensus. The main difficulties for our analysis are twofold. On one hand, the evolution of agents’ initial opinions over issue sequences is not directly determined by the interpersonal influence network and agents’ stubborn extent, which increases complexity of our analysis from the graphical perspective. On the other hand, due to the fact that the evolution of agents’ initial opinions over issue sequences is influenced by the evolution of their opinions over time sequences, connectivity preservation of the modified F-J model is more complex compared with connectivity preservation of the H-K model. As has been widely reported (Ghaderi & Srikant, 2014; Frasca et al., 2015; Friedkin, 2015), a social network with stubborn agents does not generally achieve consensus over a single issue. Our investigation shows that opinions of a group of individuals can achieve consensus over each path-dependent issue sequence 1 , even if agents are stubborn to their initial opinions. Simply put, repeatedly arising or interdependent As the research in multi-agent systems continues, the tools that are available within opinion dynamics have been enriched considerably. The Degroot model is further investigated both on continuous-time dynamics and switched topologies (Jadbabaie et al., 2003; Olfati-Saber & Murray, 2004; Ren & Beard, 2005). In addition, many complex scenarios are also considered, including asynchronous consensus (Xiao & Wang, 2008), time delays (Wang & Xiao, 2007), finite-time consensus (Wang & Xiao, 2010), leaderfollowing framework (Liu et al., 2012; Ma et al., 2016), heterogeneous model (Zheng & Wang, 2014) and antagonistic interactions (Altafini, 2013), to name but a few. Based on the gossip algorithm (Boyd et al., 2006), randomized opinion dynamics is investigated in Acemoglu et al. (2010), Ravazzi et al. (2015), Frasca et al. (2015) and Li et al. (2013), respectively. In Ghaderi & Srikant (2014), the equilibrium and convergence rate of the F-J model are investigated. Under the assumption that the adjacent matrix of the interpersonal influence network is an irreducible sub-stochastic matrix with the row sum of at least one row strictly smaller than one, the authors transform the F-J model to a random walk, and the form of equilibrium is proposed based on the first hitting probabilities of the random walk. In Bindel et al. (2011), the F-J model is interpreted as a best-response game, and the ratio between the optimal solution and the Nash equilibrium solution, which is defined as the price of anarchy, is discussed under both undirected and directed networks. Most of the available literature on the F-J model focuses on opinion evolving over a single issue. Yet the F-J model does not in general converge to consensus over a single issue due to the presence of stubborn agents. Actually, the empirical evidence shows that consensus may be reached over a sequence of issues (Joshi et al., 2010). In practice, associations of individuals (such as small groups within firms, deliberative bodies of government, etc.) are usually constituted to deal with issues within particular issue domains which consist of deeply interdependent issues, especially repeatedly arising issues. In this scenario, individuals’ opinions for interdependent issues are always correlated. Thus, extending the existing theories of opinion dynamics to issue sequences is necessary and can uncover the underlying mechanism of opinion formation in the real world. Mirtabatabaei et al. (2014) and Jia et al. (2015) modify the F-J model and the Degroot model to investigate the evolution of agents’ selfappraisals over an issue sequence, respectively. In Parsegov et al. (2017), the authors present a multidimensional extension of the F-J model in which agents’ opinions for several interdependent issues evolve over time sequences. Different from the works which focus on agents’ selfappraisal dynamics (Mirtabatabaei et al., 2014; Jia et al., 2015) over sequences of independent issues, or evolution of agents’ opinions for multiple interdependent but unordered issues over time sequences (Parsegov et al., 2017), we consider the opinion formation of the F-J model which evolves over both interdependent issue sequences and time sequences in this paper. Firstly, convergence of the F-J model is studied over a single issue. The existing results 1 In the sequel, we say that an issue sequence is path-dependent or has the property of path-dependence if individual’s opinions over it have the path-dependent property. A detailed explanation of path dependence will be presented below. 2 ei j ∈ E means w ji , 0, otherwise, w ji = 0. A directed path of length m+1 from i to j is a finite ordered sequence of distinct vertices of G with the form i, k1 , k2 , · · · , km , j. Particularly, if i = j, the directed path is called a cycle of i( j). The directed graph G is said to be strongly connected if between every pair of distinct vertices i, j, there exists a directed path that begins at i and ends at j. A maximal subgraph of G that is strongly connected forms a strongly connected component (SCC). In an SCC of digraph G, if there exists a vertex which has parents belonging to other SCCs, we say that the in-degree of this SCC is nonzero, otherwise, we call it an independent strongly connected component (ISCC). The directed graph G is said to have a spanning tree if there exists a vertex that is called the root which has a directed path to every other vertex. The directed graph G is said to contain a star subgraph if there exists a vertex that is called the center vertex which has directed edges to any other vertices. An undirected graph is a digraph G(W) which satisfies that wi j = w ji for any i, j ∈ V. For more details about algebraic graph theory, we refer to Godsil & Royal (2001). issues lead to consensus. Our study also provides a theoretical explanation for the cohort effect (Joshi et al., 2010). Moreover, the evolution of opinions over issue sequences enhances the connectivity of the social network. As long as one of the partially stubborn agent’s opinions can scatter to all the rest partially stubborn agents through partially stubborn or non-stubborn agents, opinion consensus can be reached, and all the partially stubborn agents form a star subgraph in the network which characterizes the relationship of agents’ initial opinions for successive issues. The differences between opinion dynamics over issue sequences and a single issue give us a deeper understanding of opinion formation in social networks. The rest of this paper is organized as follows. In Section 2, we introduce some notions on graph theory and the F-J model. In Section 3, convergence of the F-J model is studied over a single issue, and some properties of ultimate opinions are proposed. In Section 4, opinion formation of the F-J model is investigated over issue sequences. In Section 5, we consider opinion formation of the F-J model over issue sequences with bounded confidence. In Section 6, numerical simulations are given to illustrate the effectiveness of theoretical results. Some conclusions are drawn in Section 7. 2.2 The F-J model presented in Friedkin & Johnsen (1999) sheds light on the relationship between interpersonal influence and opinion change in social networks. The F-J model is a generalization of the Degroot model (Degroot, 1974). It introduces a positive diagonal matrix to quantify the extent of each member of the group open to the interpersonal influence. Notation: Denote the set of real numbers, natural numbers, n-dimensional real vector space and n × n real square matrix space by R, N, Rn and Rn×n , respectively. 1n is a vector of size n having all the elements equal 1. In is the n × n identity matrix. For a given vector or matrix A, AT denotes its transpose, kAk∞ denotes its maximum row sum norm. 0 denotes an all-zero vector or matrix with compatible dimension. ρ(A) denotes the spectral radius for a matrix A ∈ Rn×n . By | · | we denote the absolute value, modulus and cardinality of real number, complex number and set, respectively. Given a vector ζ ∈ Rn and a matrix sequence {Ak }m k=1 with Ak ∈ Rn×n , diag(ζ) denotes an n×n diagonal matrix whose diagonal elements are the elements of ζ, diag(A1 , A2 , . . . , Am ) denotes an nm × nm block diagonal matrix with Ak being its block on diagonal. 2 Consider n agents forming opinions on a certain issue in a network formulated by a weighted digraph G(W) = (V, E, W). Each agent holds an initial opinion xi (0), i ∈ V. The classic F-J model is: x(k + 1) = ΞW x(k) + (I − Ξ)x(0), (1) where x(k) = (x1 (k), x2 (k), . . . , xn (k))T ∈ Rn is a vector representing the stack of opinions with k ∈ N being the scale of time, W is a stochastic weighted adjacency matrix characterizing the structure of the interpersonal influence network, Ξ = diag(ξ) is the diagonal matrix mentioned above with ξ = (ξ1 , ξ2 , . . . , ξn )T ∈ Rn satisfying ξi ∈ [0, 1]. ξi represents the susceptibility of agent i to interpersonal influence while 1 − ξi quantifies the extent to which agent i is stubborn to its initial opinion. Here, xi (0) can be regarded as the internal opinions of agent i for the issue (Bindel et al., 2011). Note that if ξ = 1n , the F-J model collapses to the Degroot model. However, different from the Degroot model in which disagreement is mainly the consequence of lack of communication, it is more likely the result of the presence of stubborn agents anchored to their own initial opinions in the F-J model. This is coincident with the empirical evidence that failing in reaching agreement usually owes to the stubbornness of individuals in practice, rather than the insufficiency of persistent contracts and interactions (Friedkin & Johnsen, 1999; Friedkin, 2011; Frasca et al., 2015). Preliminaries In this section, some basic concepts on graph theory are introduced. Then, a brief introduction of the F-J model is presented. 2.1 Introduction of the F-J model Basic concepts on graph theory A weighted directed graph (digraph) G of order n is a triple G(W) = (V, E, W) which consists of a vertex set V = {1, 2, . . . , n}, an edge set E = {ei j : i, j ∈ V} and a weighted adjacency matrix W = [wi j ]n×n with entries wi j . An edge ei j is an ordered pair (i, j) which means that node j can obtain information from node i, but not necessarily vice versa. i is called the parent of j if (i, j) ∈ E. The adjacency elements associated with the edges of the graph are nonzero, i.e., 3 As mentioned above, if Ξ = I in equation (1), the F-J model is transformed into the Degroot model: model is convergent if and only if 1 is the only maximummodulus eigenvalue of W. x(k + 1) = W x(k), For two positive integers, we say they are coprime if and only if their greatest common divisor is 1. For agent i, if ξi < 1, we say it is a stubborn agent, otherwise, we say it is a non-stubborn agent. In Golub & Jackson (2010), the authors prove that the Degroot model (2) is convergent if and only if each ISCC of G(W) is aperiodic from the perspective of Markov chain corresponding to W. In Friedkin (2015), convergence of the F-J model requires that the lengths of all cycles of G(ΞW) are coprime. In the next theorem, we propose a graph-theoretical condition for the F-J model to achieve convergence by virtue of properties of irreducible and primitive matrix’s eigenvalue. Different from the Degroot model (2), convergence of the F-J model only depends on the ISCC composed of non-stubborn agents in G(W). Before proposing our convergence condition, we need the following lemmas in Horn & Johnson (2012). (2) where W is a stochastic matrix. On the other hand, consider the augmented system of the F-J model (1). Let x̂(k) = [x(0)T x(k)T ]T , we have x̂(k + 1) = Ŵ x̂(k), (3) where   In×n Ŵ =  (I − Ξ)  0   . ΞW From equation (3), the F-J model is equivalent to the leaderfollowing structure (Liu et al., 2012) where agents’ initial opinion are virtualized as stationary leaders. 3 Lemma 2 Let A ∈ Rn×n be irreducible and suppose that λ ∈ R is not in the interior of any Ger s̆gorin disc of A. If some Ger s̆gorin circle of A does not pass through λ, then λ is not an eigenvalue of A. Convergence of the F-J model over a single issue A fundamental problem with respect to the F-J model is its limiting behavior when the time scale k tends to infinity. In this subsection, convergence of the F-J model is considered over a single issue. A sufficient and necessary condition on the regularity of ΞW for convergence of the F-J model is presented in Parsegov et al. (2017). Here we give a sufficient and necessary condition from the perspective of the interpersonal influence network G(W) which ensure that x(k) converges to a constant vector as k → ∞. Firstly, we propose the definition of the convergence of dynamic systems. Lemma 3 Let A ∈ Rn×n be irreducible and nonnegative, and let P1 , P2 , . . . , Pn be the vertices of the directed graph G(A). Let Li = {l1i , l2i , . . .} be the set of lengths of all cycles of Pi in G(A), Let γi be the greatest common divisor of all the lengths in Li . Then A is primitive if and only if γ1 = γ2 = · · · = γn = 1. Remark 1 Note that the condition in Lemma 3 is consistent with the definition of aperiodicity (Definition 2 in Golub & Jackson (2010)). For a nonnegative and irreducible matrix, aperiodicity ensures that it does not have any periodic eigenvalues, that is , it is primitive. Thus, aperiodicity is a graph-theoretical criterion for primitivity. By the virtue of aperiodicity, we provide a sufficient and necessary condition concerning on the interpersonal influence network G(W) for the F-J model to achieve convergence in the next theorem. Definition 1 For a dynamic system with state vector `(k) ∈ Rn , k ∈ N, if for any initial state `(0), there exists a constant vector `ˆ ∈ Rn , such taht ˆ lim `(k) = `, k→∞ we say the system is convergent. Lemma 1 The F-J model (1) is convergent if and only if 1 is the only maximum-modulus eigenvalue of ΞW. Assumption 1 For each ISCC of G(W) which only consists of more than one non-stubborn agent, there exists at least one agent having two cycles which are coprime in length. Lemma 1 naturally follows from Corollary 2 and Lemma 5 in Parsegov et al. (2017). In Lemma 1, that 1 is the only maximum-modulus eigenvalue of ΞW means ΞW has no other maximum-modulus eigenvalues except 1, here the algebraic multiplicity of eigenvalue 1 may be larger than one. In other words, suppose λ is an eigenvalue of ΞW, then |λ| = 1 implies λ = 1. Otherwise, let a + bi , 1 be an eigenb value of ΞW with |a+bi| = 1. We have a+bi = ei arctan a = eiθ , where θ ∈ [−π, π] is the principal value of the argument of a + bi. Thus, (a + bi)k = eiθk , which shows that (a + bi)k is periodic. Suppose that J is the Jordan canonical form of ΞW, then J k is periodic, which indicates that system (1) is nonconvergent. Let Ξ = I, one can readily get that the Degroot Assumption 1 means that all ISCCs of G(W) consisting of only non-stubborn agents are aperiodic. This condition is milder than condition for convergence of the Degoot model (2) which requires that all ISCC of G(W) are aperiodic (Golub & Jackson, 2010). Theorem 1 The F-J model (1) achieves convergence if and only if Assumption 1 holds. Proof. Suppose that G(W) has α SCCs with α ≤ n. Because W is reducible, then there exists a permutation matrix P such 4 Since system (1) is convergent under Assumption 1, the limit of Ψ(k) exists. Let Ψ = limk→∞ Ψ(k). that   W1 × · · · ×      W · · · × 2   T  W̃ = P WP =  . . ..  ,  . .      Wα Property 1 Ψ is a stochastic matrix. 0 Proof. Firstly, we show that Ψ(k) is stochastic for any k ∈ N inductively. Note that ΞW and I − Ξ are both nonnegative, thus Ψ(k) is nonnegative. From equation (5), we have Ψ(0) = I and Ψ(1) = ΞW + I − Ξ are both stochastic. Now suppose that Ψ(l) is a stochastic matrix, we have where Wi , i ∈ {1, 2, . . . , α} is the irreducible adjacent matrix corresponding to each SCC of G(W) (Brualdi & Ryser, 1991). Let Ξ̃ = PT ΞP = diag(Ξ1 , Ξ2 , . . . , Ξα ), then we have ΞW = PΞ̃W̃PT . If there exists ξ j = 0 which is a diagonal element of Ξi , then G(Wi ) is separated into a vertex j which corresponds to a row of Ξi Wi with all elements being zero and several SCCs with nonzero in-degrees in G(Ξi Wi ). Otherwise, if G(Wi ) has nonzero in-degree, or G(Wi ) is an ISCC but contains at least one stubborn agent, Ξi Wi is a irreducible matrix which satisfies kΞi Wi k∞ ≤ 1 and contains at least one row whose row sum is strictly smaller than 1. From Lemma 2, we obtain ρ(Ξi Wi ) < 1. If G(Wi ) is an ISCC and contains no stubborn agent, i.e., Ξi Wi = Wi is an irreducible stochastic matrix, then 1 is its only maximum-modulus eigenvalue if and only if Wi is primitive. In this case, if G(Wi ) contains only one non-stubborn agent i, then the eigenvalue of Wi is 1. Otherwise, by Lemma 3, 1 is the only eigenvalue of Wi that has maximum modulus if and only if there exists at least one agent in G(Wi ) which has at least two cycles with coprime length. In conclusion, 1 is the only eigenvalue of ΞW with maximum-modulus if and only if Assumption 1 holds. By Lemma 1, the F-J model (1) achieves convergence if and only if Assumption 1 holds.  Ψ(l + 1)1n = ((ΞW)l+1 + = ((ΞW)l ΞW + = ((ΞW)l (ΞW + I − Ξ) + l−1 X (ΞW)t (I − Ξ))1n t=0 l−1 X = (ΞW)l Ψ(1)1n + (ΞW)t (I − Ξ)1n . t=0 l l Since Ψ(1) is stochastic, we have (ΞW) Pl−1Ψ(1)1nt = (ΞW) 1n . l Therefore, Ψ(l + 1)1n = ((ΞW) ΞW + t=0 (ΞW) (I − Ξ))1n = Ψ(l)1n = 1n , which implies that Ψ(k) is stochastic for any k ∈ N. Moreover, since Ψ1n = limk→∞ Ψ(k)1n = 1n , we obtain that Ψ is a stochastic matrix.  Property 1 shows that each agent’s opinion converges to a convex combination of all agents’ initial opinions. Let x∗ = limk→∞ x(k). Denote the set of fully stubborn agents by V f = {i ∈ V|ξi = 0}, the set of partially stubborn agents by V p = {i ∈ V|0 < ξi < 1} and the set of non-stubborn agents by Vn = {i ∈ V|ξi = 1}. Without loss of generality, let V f = {1, 2, . . . , r1 }, V p = {r1 + 1, r1 + 2, . . . , r1 + r2 }, Vn = {r1 + r2 + 1, r1 + r2 + 2, . . . , n}. Let Ψ = [ψ1 , ψ2 , . . . , ψn ], where ψi ∈ Rn . The following property reveals the influence of the interpersonal influence network G(W) and the levels of stubbornness of agents to the ultimate opinion x∗ . Property 2 Under Assumption 1, for any i ∈ Vn , if there S exists j ∈ V f V p which has a directed path to agent i in graph G(W), then ψi = 0. Proof. Following equation (5), we obtain Ψ(k + 1) − Ψ(k) k k−1 X X k+1 t k =(ΞW) + (ΞW) (I − Ξ) − (ΞW) + (ΞW)t (I − Ξ) t=0 t=0 (4) =(ΞW) k+1 − (ΞW) + (ΞW) (I − Ξ) k k =(ΞW)k Ξ(W − I). where k−1 X (ΞW)t (I − Ξ) + (ΞW)l (I − Ξ))1n t=0 Next, we shall further analyze the ultimate opinion of system (1). Employing equation (1) iteratively, we have Ψ(k) = (ΞW)k + (ΞW)t (I − Ξ))1n t=0 l−1 X Remark 2 In the available literature, convergence of the F-J model over a single issue usually requires that G(W) is undirected connected or every ISCC of G(W) contains at least one stubborn agent (Mirtabatabaei et al., 2014; Ghaderi & Srikant, 2014). Theorem 1 shows that the convergence is achieved under the milder condition which ensures condition in Lemma 1. The proof of Theorem 1 also implies that each ISCC consisting of non-stubborn agents achieves consensus as the F-J model converges. Note that from Theorem 1, that each ISCC consisting of non-stubborn agents of G(W) has at least one agent containing self-loop is a sufficient but not necessary condition for convergence of the F-J model. However, if G(W) is undirected, the F-J model achieve convergence if and only if each ISCC consisting of non-stubborn agents has at least one agent containing selfloop in G(W). The proof directly follows from Theorem 1. x(k) = Ψ(k)x(0), l X (ΞW)t (I − Ξ). (5) From the proof of Theorem 1, limk→∞ (ΞW)k exists. More- t=0 5 over, since limk→∞ Ψ(k) = Ψ , we have lim (ΞW)k Ξ(W − I) = lim (ΞW)k (I − Ξ) = 0. k→∞ k→∞ represents the proportion of the initial opinion of agent i in each agent’s ultimate opinion. Property 2 suggests that if a non-stubborn agent wants to determine the group’s ultimate opinion, i.e., to be a leader, it must ensure that it is not influenced by stubborn agents. If we regard initial opinions of stubborn agents as external information, then this result is consistent with the experimental results in Dyer et al. (2009) and Couzin et al. (2005) that the leaders emerging spontaneously in groups are usually the individuals who have more power in communication or external information. Furthermore, for any individual who has no external information, it can never be a leader unless it does not receive information of the individuals who have external information. (6) Let limk→∞ (ΞW)k = Q = [q1 , q2 , . . . , qn ] = [qi j ]n×n with qi ∈ Rn , from equation (6) we obtain ( QΞ = Q QW = Q. S Because QΞ = Q, we have qi = 0 for any i ∈ VS V p . It f follows from QW = Q that for any i ∈ V, j ∈ V V , one f p P has nt=r1 +r2 +1 qit wt j = 0. Recalling that Q and W are both nonnegative, we have that for any i ∈ Vn , if there exists S j ∈ V f V p such that wi j , 0, then qli = 0 for any l ∈ V, i.e., qi = 0. Furthermore, for aforementioned i ∈ Vn , l ∈ V and qli = 0, if there exists u ∈ Vn satisfying wui , 0, then qlu wui = qli = 0. Thus, qu = 0. This indicates that for any agent i ∈ Vn , as long as it receives information from any stubborn agents directly or indirectly, there holds qi = 0. 4 Opinion formation of the F-J model over issue sequences In this section, we study opinion formation of the F-J model over sequences consisting of interdependent issues. On one hand, most of the prior investigations on opinion dynamics are considered over a single issue. In practice, however, associations of individuals are constituted to deal with sequences of issues within particular domains, and successive issues in specific issue domain are always correlated. On the other hand, for individuals, related issues always arise repeatedly in practice. Empirical evidences show that a group of individuals who share a common set of experiences will hold highly similar beliefs or characteristics (Joshi et al., 2010). Hence, it is significant to consider the formation of opinions over issue sequences. P t Let Q̃ = limk→∞ k−1 t=0 (ΞW) (I − Ξ) = [q̃1 , q̃2 , . . . , q̃n ], then there holds Ψ = Q + Q̃ = [ψ1 , ψ2 , . . . , ψn ] = [q1 + q̃1 , q2 + q̃2 , . . . , qn + q̃n ]. Since ξi = 1 for any i ∈ Vn , we obtain q̃i = 0 for any i ∈ Vn , namely, qi = ψi for anySi ∈ Vn . In conclusion, for any i ∈ Vn , if there exists j ∈ V f V p which has a directed path to agent i in graph G(W), then ψi = 0.  In Ghaderi & Srikant (2014), the authors study the equilibrium of system (1). Under the assumption that the network topology is undirected and connected, they present the form of equilibrium taking advantage of the appropriately defined hitting probabilities of a random walk over the network. However, since their conditions for convergence require that ρ(ΞW) < 1, the equilibrium is only related to the initial opinions of stubborn agents. A similar assumption for directed graph is presented in Mirtabatabaei et al. (2014). In Property 2, we show that under the condition of Assumption 1, if a non-stubborn agent can obtain information directly or indirectly from any stubborn agents, including fully stubborn and partially stubborn agents, then the ultimate opinion x∗ does not include the information of this non-stubborn agent’s initial opinion. Otherwise, according to the proof of Theorem 1, the ultimate opinion is also related to the initial opinions of non-stubborn agents which form ISCCs with no stubborn agents in it. Next, we shall briefly introduce the path-dependence theory. The path-dependence theory is originally studied in economic and institutional change (North, 1990; Page, 2006), and subsequently developed in cognitive psychology (Egidi & Narduzzo, 1997; MacKay et al., 2006). In Egidi & Narduzzo (1997), path-dependence is observed in the behavioral experiments of human groups. The main point of pathdependence is “history matters”. That is, the decision one faces for any given circumstance are constrained by the decisions one has made in the past, even though past circumstance may no longer be relevant. This phenomenon is ubiquitous in technological innovation, institutional change and cognitive process of humans, etc. In the cognitive and decision-making process of humans, people always have cognitive inertia with which they are unwilling to change their beliefs or decisions over interdependent issues. For example, one may wonder that why the U.S. standard railroad gauge is 4 feet 8.5 inches, which is an exceedingly odd number. A well known explanation for this question is that it is resulted from path-dependence (North, 1990). Firstly, the U.S. railroads were built by English expatriates and that is the guage they built them in England. Further, the first rail lines of England were built by the same people who built the pre-railroad tramways, where that gauge was used. Moreover, the tramways were built by the same people who built wagons using the same jigs and tools. In fact, that gauge is the wheel spacing of the wagon. In this example, peo- Remark 3 Property 1 and Property 2 also provide an interesting perspective on leadership in social networks (Dyer et al., 2009) and animal groups (Couzin et al., 2005). In Wang & Xiao (2007), a generalized definition of “leader” and “follower” is presented. A leader refers to the agent who determines the ultimate state of the system, and the rest are followers. This definition includes the notion of “leader” and “follower” in which a leader is the agent who sends information to others but never receives information. From Property 1, Ψ is a stochastic matrix, thus the element of ψi 6 issue 0 issue 1 ... issue 2 issue s where the first term of Ci (x) depicts the cognitive inertia of agent i to change its initial opinions (i.e., internal opinions) for interdependent issues, and the second term represents the willingness of agent i to escape being unsocial. As a consequence, the initial opinion of agent i for issue s + 1 satisfies issue s+1 time 0 k0 k1 k2 ks-1 ks Fig. 1. An issue sequence. ple’s decisions over the interdependent issue sequence “The gauge of the U.S. railroad”, “The gauge of the English railroad”, “The gauge of the tramway” and “The wheel spacing of the wagon” remain unaltered. This phenomenon is named cognitive freezing (MacKay et al., 2006). Since in the FJ model, agents’ initial opinions can be regarded as their internal opinions, we will establish a connection between agents’ initial opinions for interdependent issues based on the path-dependence theory. xi (s+1, k s+1 ) = ζi xi (s, k s )+(1−ζi ) n X wi j x j (s+1, k s+1 ), (9) j=1 i.e., x(s + 1, k s+1 ) = Λx(s, k s ) + (I − Λ)W x(s + 1, k s+1 ), (10) where Λ = diag(ζ1 , ζ2 , . . . , ζn )T . Let s = 0, 1, 2, . . . denote a sequence of interdependent or repeatedly arising issues over which path-dependence works 2 , k s be a constant denoting the time at which issue s arises, x(s, k) denote agents’ opinions for issue s at time k. Suppose that issue 0 arises at time 0, agents stop dealing with issue s − 1 and start to deal with issue s at time k s (Fig. 1). Note that different from the assumptions in Mirtabatabaei et al. (2014) and Jia et al. (2015) which require the time interval k s+1 − k s tends to infinity, here k s+1 − k s can be any positive integer. For a given issue s, system (1) is modified as: The hypothesis in equation (10) can be naturally observed from reality. In practice, an individual always has cognitive inertia over sequences of path-dependent issues, especially when an issue arises repeatedly. In extreme cases, cognitive inertia may reinforce into cognitive freezing with which individuals keep their decisions unchangeable, even they are exposed to more efficient options (Egidi & Narduzzo, 1997; MacKay et al., 2006). Next we propose an assumption about G(W) which ensures that I − ΞW is nonsingular. x(s, k + 1) = ΞW x(s, k) + (I − Ξ)x(s, k s ), (7) Assumption 2 For each ISCC of G(W), there exists at least one stubborn agent in it. where x(s, k s ) represents initial opinions, i.e., internal opinions, of agents for issue s, and k ∈ [k s , k s+1 ]. Lemma 4 I − ΞW is nonsingular if and only if Assumption 2 holds. Let ζi denote the factor of cognitive inertia of agent i for two successive issues in an path-dependent issue sequence. Note that the basic assumption of the F-J model is that agents are anchored to their initial opinions over a given issue, this is coincident with the path-dependence theory which emphasizes that individual has the cognitive inertia over a sequence of interdependent issues. Thus, we assume that for each agent i, its factor of cognitive inertia ζi equals to its stubborn factor, i.e., ζi = 1 − ξi . Then, basing on the game theory formulation of the F-J model in Bindel et al. (2011) and the self-reinforcing process of path-dependence (North, 1990), we assume that in forming its initial opinion for issue s + 1, agent i attempts to myopically minimize the following cost function: Ci (x) = ζi (xi − xi (s, k s ))2 + (1 − ζi ) n X wi j (xi − x j )2 , Proof. Sufficiency. Since Assumption 2 holds, the proof of Theorem 1 shows that ρ(ΞW) = ρ(Ξ̃W̃) < 1, which implies that I − ΞW is nonsingular. Necessity. If Assumption 2 does not hold, there exists at least one eigenvalue of ΞW equals to 1. Namely, I − ΞW is singular.  From equation (10) we have (I − W + ΛW)x(s + 1, k s+1 ) = Λx(s, k s ). If Assumption 2 holds, then I − ΞW is nonsingular. Since Λ = I − Ξ, we obtain (8) x(s + 1, k s+1 ) = (I − ΞW)−1 (I − Ξ)x(s, k s ). j=1 2 Note that the issue sequence is not necessarily consecutive over time, some other uncorrelated issues may interpolate in it. For example, Let S 1 = S 10 , S 11 , S 12 , . . . denote a issue sequence which consists of interdependent issues, and S 2 denote another. The issues in S 1 and S 2 may be irrelevant and arise alternately. Without loss of generality, we focus on each path-dependent subsequence in this paper. (11) Since ΞW + I − Ξ is a stochastic matrix, we have that for any nontrivial Ξ and W, x(s, k s+1 ) = χ1n if and only if x(s, k s ) = χ1n , where χ ∈ R is a constant. Thus, we focus on consensus of agents’ initial opinions for each issue in the sequel. Now, we shall propose a definition of opinion consensus for system (7), (10). 7 1 for any i ∈ V. From (14) we obtain Definition 2 System (7), (10) is said to achieve consensus if and only if for any initial opinion x(0, 0), there exists a constant χ ∈ R, such that lim x(s, k s ) = χ1n . 1 + ξi (12) s→∞ ψ̂ii = Remark 4 Suppose that Assumption 2 holds. If k s+1 − k s is large enough for each s, such that ∞ P then, according to Lemma 4 and equation (7), we obtain x(s + 1, k s+1 ) = x(s, ∞), (13) which implies that agents’ initial opinions for issue s + 1 equal to their limiting opinions for issue s, namely, cognitive freezing happens. This is consistent with our discussion above that cognitive inertia can be reinforced into cognitive freezing in extreme cases. For example, as more people bought QWERTY keyboard, buying other keyboards with different key configuration requires learning to type anew. Thus, people’s opinions for the sequence of issues related to keyboard will be locked in QWERTY keyboard, even though there are many other more efficient types of keyboard (Page, 2006). Other examples of cognitive freezing are also reported, such as the fall of the Dundee jute industry (MacKay et al., 2006), the persistence of the narrow-gauge rail (North, 1990), etc. In Page (2006), the author attributes cognitive freezing to the externalities of other people. Note that in equation (8), we assume that each agent’s initial opinion for the next issue is influenced by other agents, i.e., the interpersonal influence creates externalities under Assumption 2, which lead to cognitive freezing in equation (13). S 1 > 0. 1 − ξi wii V p , we have ψii = ψ̂ii (1 − (ΞW)t . Let (ΞW)t = [ψ̂ti j ]n×n . Note that ψ̂ti j > 0 repre- Lemma 5 connects G(Ψ) with G(W), such that we can address property of G(Ψ) by analyzing G(W). From the proof of Lemma 5, we have  Ir1 ×r1 0  Ψ =  Ψ p f Ψ pp  Ψn f Ψnp Let Ψ = [ψi j ]n×n = (I − ΞW)−1 (I − Ξ). The next lemma reveals the connections between G(Ψ) and G(W).  0  0 ,  0 where Ψ pp ∈ Rr2 ×r2 . Property 3 Any root of subgraph G(Ψ pp ) is a center vertex which has directed edges to any other vertices of G(Ψ pp ). Lemma 5 Under Assumption 2, the following statements hold: S (i) ψii > 0 for any i S ∈ V f Vp. S (ii) For any i ∈ V p Vn , j ∈ V f V p , i , j, ψi j > 0 if and only if there exists a directed path ( j, j1 ), ( j1 , j2 ), . . . , ( jm , i) from S vertex j to vertex i in G(W), where j1 , j2 , . . . , jm ∈ V p Vn . Proof. Suppose that i is a root of G(Ψ pp ). Let (i, j) and ( j, l) be edges of G(Ψ pp ) for j, l ∈ V p . By Lemma S 5, there exist directed paths consisting of vertices in V p Vn from i to j and j to l in G(W), respectively. Thus, there exists at least aSdirected path consisting of edges between vertices in V p Vn from i to l in G(W). From Lemma 5 we have ψli > 0, namely, there exists a directed edge from i to l in G(Ψ pp ). Following the same lines as above, one has that any root of subgraph G(Ψ pp ) is a center vertex which has directed edges to any other vertices of G(Ψ pp ).  Proof. Let Ψ̂ = [ψ̂i j ]n×n = (I − ΞW)−1 . (i) Since (I − ΞW)Ψ̂ = I, we have i ∈ V. ≥ sents that there exist paths consisting of t edges (maybe appear repeatedly or contain self-loops ) in G(ΞW) start from vertex j to vertex i. Moreover, if we take no account of the weights, G(ΞW) and G(W) have no difference except the edges to fully stubborn agents. Thus, if there exists a directed path ( j, j1 ), ( j1 , j2 ), . . . , ( jm , i) S from vertex j to vertex i in G(W), and j1 , j2 , . . . , jm ∈ V p Vn , we have ψ̂m i j > 0. Therefore, ψ̂i j > 0. Otherwise, we have ψ̂ti j = 0 for any t ∈ N. Thus, ψ̂i j = 0. Since ξ j < 1 for S j ∈ V f S V p , ψi j = ψ̂Si j (1 − ξ j ), we have that ψi j > 0 for any i ∈ V p Vn , j ∈ V f V p , i , j if and only if there exists a directed path ( j, j1 ), ( j1 , j2 ), . . . , ( jm , i) from S vertex j to vertex i in G(W), where j1 , j2 , . . . , jm ∈ V p Vn .  Thus, wit ψ̂ti , 1 − ξi wii (ii) Since Assumption 2 holds, one has (I − ΞW)−1 = t=0 x(s, ∞) = (I − ΞW)−1 (I − Ξ)x(s, k s ). n X wit ψ̂ti t=1 t,i Since 1 − ξi > 0 for any S i ∈ Vf ξi ) > 0 for any i ∈ V f V p . x(s, k s+1 ) = x(s, ∞), ψ̂ii (1 − ξi wii ) = 1 + ξi n P (14) t=1 t,i Lemma 5 and Property 3 show that if there exists a directed path from a stubborn agent i to a partially stubborn or non-stubborn agent j in G(W), then (i, j) is an edge of G(Ψ). Moreover, if G(Ψ pp ) contains a spanning tree, then it Assumption 2 implies that if wii = 1, ξi < 1. Thus, ξi wii < 8 also contains a star subgraph. In other words, the connectivity of the social network is enhanced over issue sequences. The block form of Ψ shows that fully stubborn agents never change their opinions in the process of opinion formation. Thus, we consider the scenario in which there are no fully stubborn agents in the social network. form of Ψ indicates that when there exists more than one fully stubborn agent in the network, consensus can not be achieved for arbitrary initial opinion unless the influence of fully stubborn agents to others is identical (Friedkin, 2015). Since the condition for achieving opinion consensus in Theorem 2 is sufficient and necessary, it also gives the condition for opinions of agents to form clusters. Theorem 2 Suppose that V f = ∅ and Assumption 2 holds. System (7), (10) achieves opinion consensus if and only if there exists a partially stubborn agent which has directed paths to any other partially stubborn agents in G(W). Corollary 1 Suppose that V f = ∅ and Assumption 2 holds. System (7), (10) forms opinion clusters if and only if there exists more than one ISCC in G(W). Proof. Sufficiency. Since V f = ∅, i.e., r1 = 0, we have 5    Ψ pp W pp W pn     , Ψ =  W =  Ψnp Wnp Wnn  0  , 0 As discussed in Section 4, the cognitive inertia of agent can be reinforced into cognitive freezing under Assumption 2, i.e., each agent’s initial opinion for issue s + 1 equals to its limiting opinion for issue s if its opinion for s converges. It is mainly because of the externalities created by agents’ interpersonal influence under which agents treat opinions of others with no difference. However, in practice, people usually assimilate information in a selective way: they tend to give considerable weight to the information supporting their initial opinions, and dismiss the undermining information meanwhile, which is named echo-chamber effect (Sunstein, 2012; Dandekar et al., 2013). Similar observations in animal groups are reported in Couzin et al. (2005). In this section, we consider a more general scenario that each agent maintains a confidence bound in forming its initial opinions. where W pp ∈ Rr2 ×r2 , Wnn ∈ Rn−r2 ×n−r2 , Ψ pp ∈ Rr2 ×r2 . From Property 1 we have that Ψ pp and Ψnp are both stochastic matrices. Lemma 5. (i) indicates that Ψ pp has positive diagonal elements. Moreover, since there exists a partially stubborn agent which has S directed paths consisting of edges between vertices in V p Vn to any other partially stubborn agents in G(W), Lemma 5. (ii) implies that G(Ψ pp ) has a spanning tree. By Corollary 3.5 and Lemma 3.7 in Ren & Beard (2005), we have that Ψ pp is stochastic, indecomposable and aperiodic (SIA). That is, there exists a nonnegative vector ν ∈ Rr2 such that lim Ψ spp = 1r2 νT , where ν satisfies ΨTpp ν = ν and s→∞ 1T ν = 1. Note that   Ψ spp lim Ψ = lim  s→∞ s→∞ Ψ Ψ s−1 np pp s Opinion formation of the F-J model over issue sequences with bounded confidence For simplicity, we employ the assumption used in Mirtabatabaei et al. (2014) and Jia et al. (2015) that the scale of time k increases faster than the scale of issue s in this section. In other words, the convergence of opinion x(s, k) can be achieved with k increasing before the next issue s+1 arising, i.e., k s = ∞ for issue s − 1 and k s = 0 for issue s. From Remark 4, we have x(s + 1, 0) = x(s, ∞) under Assumption 2, which means that agents take their ultimate opinions for the last issue as their initial opinions for the next issue. Inspired by the echo-chamber effect, we suppose that agent i forms its initial opinion xi (s + 1, 0) by taking a weighted average of other agents’ ultimate opinions x j (s, ∞), j ∈ V which satisfy |xi (s, ∞) − x j (s, ∞)| < d, where d ∈ R is the given confidence bound.      0  1r2 νT 0  1r2 νT 0  =   =   , 0 Ψnp 1r2 νT 0 1n−r2 νT 0 which indicates that Ψ spp converges to a matrix of rank one as s → ∞. Since Assumption 2 holds, we have lim s→∞ x(s, k s ) = Ψ s x(0, 0) = [νT 0Tn−r2 ]x(0, 0)1n , namely, system (7), (10) achieves opinion consensus as s → ∞. Necessity. If there exists no partially stubborn S agent which has directed paths consisting of vertices in V p Vn to any other partially stubborn agents in G(W), Lemma 5. (ii) suggests that G(Ψ pp ) does not contain any spanning trees. This implies that the eigenvalue 1 of Ψ pp has algebraic multiplicity larger than one, that is, Ψ has more than one eigenvalue equal to 1. In conclusion, opinion consensus cannot be achieved.  Let Ni (s) = { j ∈ V : |xi (s, ∞) − x j (s, ∞)| < d} denote the set of agents whose ultimate opinions for issue s are within the confidence interval of agent i, NiΨ = { j ∈ V : ψi j > 0} denote the set of neighbors of agent i in G(Ψ). Here we employ a symmetric modification of the H-K model (Jing et al., 2016; Yang et al., 2014): X xi (s + 1, 0) = xi (s, ∞) + h (x j (s, ∞) − xi (s, ∞)), (15) As is often pointed out, the F-J model does not generally achieve opinion consensus over a single issue even if the network topology is completely connected (Ghaderi & Srikant, 2014; Frasca et al., 2015). However, Theorem 2 shows that in the absence of fully stubborn agents, as long as there exists at least one partially stubborn agent whose opinion can scatter to other partially stubborn agents through partially stubborn agents or non-stubborn agents, opinion consensus can be achieved over a sequence of issues. The block j∈Ni (s) where h ∈ R satisfies h > 0 and 1 − (n − 1)h > 0. Note that if the differences of agents’ initial opinions are large enough 9 such that Ni (s) = {i} for any i ∈ V and s ∈ N, equation (15) will collapse to equation (11). Combining equation (15) and Lemma 4, we have G(Φ0 (s)) has a spanning tree, we have G(Φ1 (s)) contains a spanning tree. Following from Corollary 3.5 and Lemma 3.7 in Ren & Beard (2005), Φ1 (s) is SIA.  x(s, ∞) = Ψx(s, 0), Lemma 6 implies that as long as the connectivity of G(Φ0 (s)) can be preserved, Φ1 (s) is SIA. This is on account of enhanced connectivity of the interpersonal influence network over issue sequences, which is shown in Property 3. Next, we give the condition for H(0) to preserve connectivity. and x(s + 1, 0) = H(s)x(s, ∞), where Ψ = (I − ΞW)−1 (I − Ξ), H(s) = [hi j (s)]n×n , and     hii (s) = 1 − h(|Ni (s)| − 1) i ∈ V,  hi j (s) = h j , i, j ∈ Ni (s),     hi j (s) = 0 otherwise. (16) Let y(s) = Ψx(s), then x(s + 1) = H(s)y(s). Denote the set of common neighbors of agents i and j for issue s by For simplicity, we omit the time scale k in the sequel. Let Φ(s) = [ϕi j (s)]n×n = H(s)Ψ, then we have x(s + 1) = Φ(s)x(s). According to the partition of V, we have  Φ1 (s) Φ(s) =  Φ2 (s) Ni j (s) = {l ∈ V : |yi (s) − yl (s)| < d, |y j (s) − yl (s)| < d}.  0  , 0 From equation (15), we have that G(H(s)) is undirected. The following assumption proposes the conditions for connectivity preservation of G(H(s)). where Φ1 (s) ∈ R(r1 +r2 )×(r1 +r2 ) . Rewrite H(s) as a same block form of W:   H f f (s) H f p (s) H f n (s)   H(s) = H p f (s) H pp (s) H pn (s) ,   Hn f (s) Hnp (s) Hnn (s) S T Assumption 3 (i) For any i ∈ V f , j ∈ (V p Vn ) Ni (0) S and l ∈ N Ψ , l ∈ Ni (0). For any i ∈ V p Vn , j ∈ T S T j , l ∈ N j (0) Ni (0). (V p Vn ) Ni (0) and l ∈ N Ψ Tj (ii) For any i ∈ V f , j ∈ V f Ni (0), |Ni j (0)| > n2 . For any S T 1 . i ∈ V p Vn , j ∈ V Ni (0), |Ni j (0)| > n2 + 4h where H f f (s) ∈ Rr1 ×r1 , H pp (s) ∈ Rr2 ×r2 and Hnn (s) ∈ R(n−r1 −r2 )×(n−r1 −r2 ) . Let   H f f (s) H f p (s)  Φ0 (s) =   Wp f W pp . As mentioned above, only if the difference between two agents’ opinions is small enough, agents will take account of other agents’ opinions. In Assumption 3, we give some conditions about the distribution of agents’ ultimate opinions for the first issue to ensure that G(H(0)) is adequately connected and thus the connectivity of G(H(s)) can be preserved for any s ∈ N. Note that given a social network, the interpersonal influence network G(W) and agents’ stubborn extent Ξ are determined, and y(0) = (I −ΞW)−1 (I −Ξ)x(0, 0). Therefore, the conditions on y(0) in Assumption 3 are actually conditions on x(0, 0). From equation (15), we have 1 . Moreover, since |Ni j (0)| < n, from Assumption 3 h < n−1 1 1 1 we have h > 2n . Thus, 2n < h < n−1 . (r1 +r2 )×(r1 +r2 ) Lemma 6 For any s ∈ N, if G(Φ0 (s)) has a spanning tree, then Φ1 (s) is SIA. Proof. From equation (16), H(s) is a nonnegative and stochastic matrix with positive diagonal elements. By S Lemma 5, we have ψii > 0 for any i ∈ V f V p . Thus, Φ(s) and Φ1 (s) are both stochastic matrices and Φ1 (s) has positive diagonal elements. Since Φ(s) = H(s)Ψ, we have  n P    ϕi j (s) = hi j (s) + hit (s)ψt j    t=r1 +1  n  P    hit (s)ψt j ϕi j (s) = t=r1 +1 i ∈ Vf S Vp, j ∈ V f , i ∈ Vf S Vp, j ∈ Vp. Lemma 7 Under Assumption 3, no edge in G(H(0)) will be lost in G(H(s)) for any s ∈ N. (17) S Hence, for any i ∈ V f , j ∈ V f V p , hi j (s) > 0 indicates ϕi j (s) > 0.SMoreover, Lemma 5. (ii) shows that for i ∈ V p , j ∈ V f V pS , wi j > 0 means ψi j > 0. Thus, for any i ∈ V p , j ∈ V f V p , wi j > 0 implies ϕi j > 0. Because Proof. Firstly, we shall show that if Assumption 3 holds for issue s, then it holds for issue s + 1. Suppose that for any |yi (s) − y j (s)| < d, i, j ∈ V, Assumption 3 holds for issue s. 10 For i, j ∈ V f , we have N j (s) T Ni (s) and |Nil (0)| > X |yi (s + 1) − y j (s + 1)| =|xi (s + 1) − x j (s + 1)| X X = yi (s) + h (yl (s) − yi (s)) − y j (s) − h (yl (s) − y j (s)) l∈Ni (s) = yi (s) − y j (s) + h X + (y j (s) − yi (s))+ h X (yl (s) − yi (s)) − h ≤ (yl (s) − y j (s)) X ≤(1 − h|Ni j (s)|)|yi (s) − y j (s)| + h X < l∈N j (s)\Ni j (s) ψ jl (xi (s + 1) − xl (s + 1)) . l∈N Ψ j Noting that for any l ∈ N Ψ j , one has l ∈ Ni (s) and |Nil (0)| > n . Following the same lines as above, we obtain 2 |yi (s + 1) − y j (s + 1) |< X ψ jl (xi (s + 1) − xl (s + 1)) X ψil d X d ψ jl = d. + 2 2 Ψ l∈N j The connectivity preservation is a crucial and challenging problem in research for the consensus problem of the H-K model since the evolution of agents’ opinions is only associated with the initial states and their distribution. In Jing et al. (2016), a potential function is presented to measure the total difference of agents’ initial opinions such that the connectivity is preserved by restricting the distribution of agents’ initial opinions. Following a different line, in Lemma 7 we ensure that the links existing in G(H(0)) will never be lost as s increases by proposing a lower bound for the number of agents’ common neighbors in G(H(0)). In addition, different from the modified H-K model considered in Yang et al. (2014), in system (7), (15), the evolution of opinions over time scale increases the complexity of our analysis. Vn , we have l∈N Ψ j = ψil (xl (s + 1) − xi (s + 1)) That is to say, if Assumption 3 holds for issue s, then no edge contained in G(H(s)) will be lost, which implies that Assumption 3 holds for issue s+1. Thus, no edge in G(H(0)) will be lost in G(H(s)) for any s ∈ N under Assumption 3.  |yi (s + 1) − y j (s + 1)| X = xi (s + 1) − ψ jl xl (s + 1) X ψ jl (xi (s + 1) − xl (s + 1)) X l∈NiΨ <(1 − h|Ni j (s)|)d + h(|Ni (s)| + |N j (s)| − 2|Ni j (s)|)d ≤(1 − h|Ni j (s)|)d + h(n − |Ni j (s)|)d ≤(1 + hn − 2h|Ni j (s)|)d <d. S X l∈N Ψ j |(yl (s) − y j (s))| For i ∈ V f , j ∈ V p ψil (xl (s + 1) − xi (s + 1)) X + |(yl (s) − yi (s))| l∈Ni (s)\Ni j (s) +h ψ jl xl (s + 1) l∈N Ψ j l∈NiΨ l∈N j (s)\Ni j (s) l∈Ni (s)\Ni j (s) X we obtain l∈N Ψ j l∈Ni j (s) X 1 4h , l∈NiΨ l∈N j (s) X + ψil xl (s + 1) − l∈NiΨ = n 2 Lemma 8 (Lemma 3.1, Ren & Beard (2005)) Let P1 , P2 , ..., Pm ∈ Rn×n be m nonnegative matrices with positive diagonal elements, then ψ jl d = d. P1 P2 ...Pm ≥ η(P1 + P2 + ... + Pm ), l∈N Ψ j where η can be specified from matrices Pi , i = 1, 2, ..., m. ψ jl xl (s + 1) . By the proof of Lemma 6 and Lemma 7, that the subgraph of G(H(0)) corresponding to stubborn agents is connected also implies that Φ1 (s) is SIA. In the next theorem, we employ a more feasible condition which reduces the requirements for the distribution of agents’ initial opinions. Since for any l ∈ N Ψ j , Assumption 3 ensures l ∈ Theorem 3 Under Assumption 2 and 3, system (7), (15) achieves opinion consensus if the following conditions hold: (i) there exists a partially stubborn Sagent which has directed paths consisting of vertices in V p Vn to any other partially For i, j ∈ (V p S Vn ), we get |yi (s + 1) − y j (s + 1)| = X l∈NiΨ ψil xl (s + 1) − X l∈N Ψ j 11 stubborn agents in G(W); (ii) G(H f f (0)) is connected; (iii) there exists an edge in G(H(0)) between a partially stubborn agent and a fully stubborn agent. Theorem 3 show that consensus can be achieved by the F-J model over a sequence of issues. Yet in the same context, it generally fails to achieve consensus over a single issue, even under completely connected networks. In practice, as the number of the issues that individuals collaborate to deal with increases, initial opinions of individuals for each issue become more and more similar and reach agreement ultimately. This phenomenon is named cohort effect (Joshi et al., 2010), which describes a high degree of similarity in characteristics or beliefs of a group of individuals who share a common set of experiences. Moreover, this mechanism is usually used to train groups for specific tasks (to name but a few, athlete team, special action force or fire brigade) to make the opinions or decisions of the group members grow similar. Proof. From Assumption 3 and Lemma 7, we have that G(Φ0 (s)) contains a spanning tree for any s ∈ N. Hence, Lemma 6 implies that Φ1 (s) is SIA and G(Φ1 (s)) contains a spanning tree for any s ∈ N. Let ∆(s) = Φ(s − 1)Φ(s − 2) . . . Φ(1)Φ(0), then x(s) = ∆(s)x(0). From the block form of Φ(s), we have  Φ1 (s − 1)Φ1 (s − 2) . . . Φ1 (0) ∆(s) =  Φ2 (s − 1)Φ1 (s − 2) . . . Φ1 (0)  0  . 0 Since the set of all possible matrices H(s) is finite under equation (16), Φ(s) = H(s)Ψ, one has that the set of all possible matrices Φ(s) is finite, which indicates that the set of all possible matrices Φ1 (s) is finite. Moreover, because G(Φ1 (s)) contains a spanning tree for any s ∈ N, from Lemma 8, we have Φ1 (s − 1)Φ1 (s − 2) . . . Φ1 (0) is SIA for any s ∈ N. From Wolfowitz Theorem (Wolfowitz, 1963), there exists a nonnegative vector ν ∈ Rr1 +r2 such that 6 In this section, numerical simulations are provided to illustrate the effectiveness of the theoretical results proposed above. Example 1 (Convergence of the F-J model over a single issue) Consider 10 agents labeled from 1 to 10 interacting with each other in digraph G(W 1 ) (Fig. 2a). W 1 is the stochastic adjacent matrix of G(W 1 ). Let ξ1 = (0, 0, 0, 0.2, 0.5, 0.7, 1, 1, 1, 1)T , namely, V f = {1, 2, 3}, V p = {4, 5, 6} and Vn = {7, 8, 9, 10}. Initialize x(k) with x(0) = (−1, 0, 1, 1, −2, 0, −1, −2, 1, 2)T , the trajectory of x(k) under G(W 1 ) with k increasing is shown in Fig. 2b. lim Φ1 (s − 1)Φ1 (s − 2) . . . Φ1 (0) = 1r1 +r2 νT . s→∞ Thus,   Φ1 (s − 1)Φ1 (s − 2) . . . Φ1 (0) 0  lim  s→∞ Φ (s − 1)Φ (s − 2) . . . Φ (0) 0 2 1 1   T   1 ν 0 r +r 1 2  . =  lim s→∞ Φ2 (s − 1)1r1 +r2 νT 0 Note that in Fig. 2b, the opinions of fully stubborn agents are fixed, while the opinions of partially stubborn agents and non-stubborn agents tend to be constant as k increases. The network G(W 1 ) has four SCCs, i.e., {1, 4, 5}, {3, 6, 7}, {8, 9, 10}, {2}. The ISCC consisting of non-stubborn agents is {8, 9, 10}, in which agent 9 and agent 10 both have two cycles whose lengths are 2 and 3, respectively. Thus, Assumption 1 is satisfied. By computing we find√ that the maximum-modulus √ eigenvalues of W 1 are 1, − 21 + 23 i, and − 12 − 23 i, while the maximum-modulus eigenvalue of Ξ1 W 1 is 1, which accords with the algebraic condition in Lemma 1. In Fig. 2b, the trajectory of x(k) is almost fixed after k = 20. When k = 20, one can find that the ISCC {8, 9, 10} achieve consensus, while the other non-stubborn agent 7 has different ultimate opinion with them. Moreover, we have ψ7 = 0, ψ8 , ψ9 and ψ10 are nonzero, which is consistent with Property 2 and the discussion following it. Since Φ2 is stochastic, we obtain   1r +r νT lim ∆(s) =  1 2 s→∞ 1n−r1 −r2 νT  0  . 0 Therefore, lim lim x(s, k) = lim x(s) = [νT 0Tn−r1 −r2 ]x(0)1n , s→∞ k→∞ Simulations s→∞ which implies that system (7), (15) achieves consensus as k → ∞, s → ∞.  Example 2 (Consensus of the F-J model over issue sequences) Consider 10 agents labeled from 1 to 10 in which agents 1 to 5 are partially stubborn agents, others are non-stubborn agents. The network topology is characterized by the digraph G(W 2 ) (Fig. 3a). Let ξ2 = (0.3, 0.6, 0.2, 0.8, 0.7, 1, 1, 1, 1, 1)T . x(0, 0) = (−1, 0, 1, 1, −2, 0, −1, −2, 1, 2)T is the initial opinion vector at s = 0 and k = 0. The trajectory of x(k) is showed in Fig. 4. As we discussed in Rmark 4, cognitive freezing emerges in model (7), (10) under Assumption 2. This is attributed to the externality resulted from the interpersonal influence. Thus, in model (7), (10), fully stubborn agents will keep their opinions unchangeable for all issues. Theorem 3 shows that the cognitive freezing is weakened by the bounded confidence of agents so that consensus can be achieved even in the presence of fully stubborn agents. Theorem 2 and 12 2 2 partially stubborn agent fully stubborn agent 6 partially stubborn agent 7 non-stubborn agent 6 non-stubborn agent 7 1 5 5 3 3 4 4 10 1 9 9 2 10 8 6 7 6 7 8 a . G(W 1 ) a . G(W 2 ) 2.5 fully stubborn agents partially stubborn agents non−stubborn agents 2 partially stubborn agent 6 1.5 non-stubborn agent 7 1 x(k) 0.5 0 2 5 −0.5 3 −1 4 −1.5 1 −2 9 −2.5 0 10 4 8 12 16 20 k 7 b . x(k) 6 8 b . G(Ψ2 ) Fig. 2. The network topology G(W 1 ) and the trajectory of x(k). Fig. 3. The network topologies G(W 2 ) and G(Ψ2 ). Fig. 4 shows the trajectories of x(s, k) over issues s = 0, 1, 2, 5, 10, 20, respectively. Note that over each issue s, x(s, k) tends to a steady opinion vector as k increases, and all steady opinions tend to be a common value with s increasing. When s = 30, Fig. 4 shows that consensus is achieved by system (7), (10). Moreover, since G(W 2 ) satisfies the condition in Assumption 2, we have Ψ3 = (I − Ξ2 W 2 )−1 (I − Ξ2 ). From Lemma 5 and Property 3, the existence of a partially stubborn agents in G(W 2 ) which has a directed path to other partially stubborn agents implies that the subgraph corresponding to partially stubborn agents in G(Ψ2 ) not only has a spanning tree, but also contains a star topology, and selfloops are contained by partially stubborn agents in G(Ψ2 ), which are shown in Fig. 3b. We can find visually from Fig. 3b that partially stubborn agents do not receive information from non-stubborn agents, which indicates that ψi = 0 for any i ∈ {5, 6, . . . , 10}. This is consistent with the theoretical result in Property 2. Example 3 (Consensus of the F-J model over issue sequences with bounded confidence) Consider system (7), (15) with 10 agents labeled from 1 to 10. The network topology G(W 3 ) is showed in Fig. 5a. Let ξ3 = (0, 0, 0, 0.8, 0.3, 0.2, 1, 1, 1, 1)T , namely, V f = {1, 2, 3}, V p = {4, 5, 6} and others are non-stubborn agents. Let x(0, 0) = (−0.7, 0.2, 0, 0.2, 2, −2.5, −1.5, 1, 1.5, −1)T . 1 1 The confidence bound is set as d = 1. Since 2n < h < n−1 , let h = 0.1. Fig. 6a and Fig. 6b show the trajectories of x(s, k) and x(s, 0), respectively. Because that in Fig. 5a, G(W 3 ) satisfies the condition of Assumption 2, Ψ3 exists. G(Ψ3 ) is depicted in Fig. 5b. Moreover, G(Ψ3 ) and x(0, 0) satisfy the conditions in Assumption 3 and Theorem 3, thus x(s, 0) reaches consensus as s → ∞ (Fig. 6b), which implies x(s, k) reaches consensus (Fig. 6a). The connectivity of G(H(0)) is related with the percentage of the fully stubborn agents. In this exam- 13 s=0 s=1 2 fully stubborn agents partially stubborn agents non−stubborn agents 2 x(1,k) 0 −2 0 10 20 s=2 30 0 1 −2  40 k 0 −2  s=5    k 0   s=10      s=20       k −1 −3 10 9  k 2 0 0 −2  k x(20,k) x(10,k)  −2  2 −2   2 x(5,k) x(2,k) 2 x(s,k) x(0,k) 2 8 7 6 k 0 −2   partially stubborn agents     5 4 3 2 1 0 0 25 20 15 10 5 39 35 30 s k non−stubborn agents a . The trajectories of x(s, k). Fig. 4. The trajectories of x(s, k) at s = 0, 1, 2, 3, 4, 30, respectively. 2 fully stubborn agents partially stubbornagents non−stubborn agents 1.5 fully stubborn agent 2 1 partially stubborn agent 2 0.5 6 x(s,0) non-stubborn agent 7 10 1 5 6 0 −0.5 0.1 −1 0 −1.5 −0.1 −2 −0.2 0.5 −2.5 0 2 4 6 8 10 12 14 1 16 1.5 18 20 2 22 2.5 24 3 26 28 30 32 34 36 3839 s 4 9 b . The trajectory of x(s, 0). 8 7 3 Fig. 6. The trajectories of x(s, k) and x(s, 0). a . G(W 3 ) 10 ple, Assumption 3 requires that each fully stubborn agent has at least 8 neighbors in G(H(0)). By computing we have y(0) = (−0.7, 0.2, 0, 0.2, 1.46, −1.9876, 0.0082, 0.0274, 0.2, −0.25)T , thus agents 1, 2, 3, 4, 7, 8, 9, 10 form a completely connected subgraph in G(H(0)), and agents 5 and 6 are isolated with others, respectively. fully stubborn agent 2 partially stubborn agent 6 non-stubborn agent 1 7 2 7 6 In this paper, opinion dynamics in social networks with stubborn agents has been investigated over issue sequences. The social network with stubborn agents is described by the F-J model. In order to investigate opinion formation of the F-J model over issue sequences, we studied convergence of the F-J model over a single issue and properties of the ultimate opinion. Then taking advantage of the underlying connections between the network of interpersonal influence and the network describing the relationship of agents’ initial opinion for successive issues, opinion consensus and cluster of the F-J model was investigated over issue sequences based on the path-dependence theory. Moreover, we further 5 3 4 9 8 Conclusions 7 b . G(Ψ3 ) Fig. 5. The network topologies G(W 3 ) and G(Ψ3 ). 14 the community cleavage problem, IEEE Control Systems, 35(3), 40-51. Friedkin, N. E., & Johnsen, E. C. (1999). Social influence networks and opinion change, Advances in Group Processes, 16(1), 1-29. Friedkin, N. E. (2011). A formal theory of reflected appraisals in the evolution of power, Administrative Science Quarterly, 56(4), 501-529. Frasca, P., Ishii, H., Ravazzi, C., & Tempo, R. (2015). Distributed randomized algorithms for opinion formation, centrality computation and power systems estimation: A tutorial overview, European journal of control, 24(1), 213. Golub, B., & Jackson, M. O. (2010). Naive learning in social networks and the wisdom of crowds, American Economic Journal: Microeconomics, 2(1), 112-149. Godsil, C., & Royal, G. (2001). Algebraic graph theory, New York: Springer-Verlag. Ghaderi, J., & Srikant, R. (2014). Opinion dynamics in social networks with stubborn agents: Equilibrium and convergence rate, Automatica, 50(12), 3209-3215. Horn, R. A., & Johnson, C. R. (2012). Matrix analysis, New York: Cambridge University Press. Hegselmann, R., & Krause, U. (2002). Opinion dynamics and bounded confidence models, analysis, and simulation, Journal of Artificial Societies and Social Simulation, 5(3). Jia, P., Mirtabatabaei, A., Friedkin, N. E., & Bullo, F. (2015). Opinion dynamics and the evolution of social power in influence networks, SIAM Review, 57(3), 367-397. Jing, G., Zheng, Y., & Wang, L. (2016). Consensus of multiagent systems with distance-dependent communication networks, IEEE Transactions on Neural Networks and Learning Systems, DOI: 10.1109/TNNLS.2016.2598355. Jadbabaie, A., Lin, J., & Morse, A. S. (2003). Coordination of groups of mobile autonomous agents using nearest neighbor rules, IEEE Transactions on Automatic Control, 48(6), 988-1001. Joshi, A., Dencker, J. C., Franz, G., & Martocchio, J. J. (2010). Unpacking generational identities in organizations, Academy of Management Review, 35(3), 392C414. Liu, H., Xie, G., & Wang, L. (2012). Necessary and sufficient conditions for containment control of networked multiagent systems, Automatica, 48(7), 1415-1422. Li, L., Scaglione, A., Swami, A., & Zhao, Q. (2013). Consensus, polarization and clustering of opinions in social networks, IEEE Journal on Selected Areas in Communications, 31(6), 1072-1083. Lamport, L., Shostak, R., & Pease, M. (1982). The Byzantine generals problem, ACM Transactions on Programming Languages and Systems, 4(3), 382-401. Mirtabatabaei, A., & Bullo, F. (2012). Opinion dynamics in heterogeneous networks: convergence conjectures and theorems, SIAM Journal on Control and Optimization, 50(5), 2763-2785. Mirtabatabaei, A., Jia, P., Friedkin, N. E., & Bullo, F. (2014). On the reflected appraisals dynamics of influence networks with stubborn agents, American Control Conference, IEEE, 3978-3983. Ma, J., Zheng, Y., Wu, B., & Wang, L. (2016). Equilibrium considered the more general scenario where agents have bounded confidence to others when they form their initial opinions. Mathematical conditions for opinions achieving consensus or forming clusters were established, respectively. We showed that the F-J model may achieve opinion consensus over each path-dependent issue sequence, while it generally does not achieve consensus over a single issue. The connections between network characterizing interpersonal influence and the one describing the relationship of agents’ initial opinions for successive issues uncover the difference between opinion formation over issue sequences and over a single issue, such as the enhanced network connectivity in the former. Our investigation may provide some theoretical explanations for many observations in the real world such as decision making in human society or animal kingdom. The future work will focus on the evolution of agents’ extent of stubborn over issue sequences. References Acemoglu, D., Ozdaglar, A., & ParandehGheibi, A. (2010). Spread of (mis) information in social networks, Games and Economic Behavior, 70(2), 194-227. Altafini, C. (2013). Consensus problems on networks with antagonistic interactions, IEEE Transactions on Automatic Control, 58(4), 935-946. Brualdi, R. A., & Ryser, H. J. (1991). Combinatorial matrix theory, New York: Cambridge University Press. Boyd, S., Ghosh, A., Prabhakar, B., & Shah, D. (2006). Randomized gossip algorithms, IEEE Transactions on Information Theory, 52(6), 2508-2530. Blondel, V. D., Hendrickx, J. M., & Tsitsiklis, J. N. (2009). On Krause’s multi-agent consensus model with statedependent connectivity, IEEE Transactions on Automatic Control, 54(11), 2586-2597. Bindel, D., Kleinberg, J., & Oren, S. (2011). How bad is forming your own opinion?, Proceedings of the 52nd Annual IEEE Symposium on Foundations of Computer Science, 57-66. Couzin, I. D., Krause, J., Franks, N. R., & Levin, S. A. (2005). Effective leadership and decision-making in animal groups on the move, Nature, 433(7025), 513-516. Degroot, M. H. (1974). Reaching a consensus, Journal of the American Statistical Association, 69(345), 118-121. Dyer, J. R. G., Johansson, A., Helbing, D., Couzin, I. D., & Krause, J. (2009). Leadership, consensus decision making and collective behaviour in humans, Philosophical Transactions of the Royal Society of London B: Biological Sciences, 364(1518), 781-789. Dandekar, P., Goel, A., & Lee, D. T. (2013). Biased assimilation, homophily, and the dynamics of polarization, Proceedings of the National Academy of Sciences, 110(15), 5791-5796. Egidi, M., & Narduzzo, A. (1997). The emergence of pathdependent behaviors in cooperative contexts. International Journal of Industrial Organization, 15(6), 677-709. Friedkin, N. E. (2015). The problem of social control and coordination of complex systems in sociology: A look at 15 topology of multi-agent systems with two leaders: A zerosum game perspective, Automatica, 73, 200-206. MacKay, R., Masrani, S., & McKiernan, P. (2006). Strategy options and cognitive freezing: The case of the Dundee jute industry in Scotland, Futures, 38(8), 925-941. North, D. C. (1990). Institutions, institutional change and economic performance, Cambridge university press. Olfati-Saber, R., & Murray, R. M. (2004). Consensus problems in networks of agents with switching topology and time-delays, IEEE Transactions on Automatic Control, 49(9), 1520-1533. Parsegov, S. E., Proskurnikov, A. V., Tempo, R., & Friedkin, N. E. (2017). Novel Multidimensional Models of Opinion Dynamics in Social Networks, IEEE Transactions on Automatic Control, 62(5), 2270-2285. Page, S. E. (2006). Path dependence, Quarterly Journal of Political Science, 1(1), 87-115. Ren, W., & Beard, R. W. (2005). Consensus seeking in multiagent systems under dynamically changing interaction topologies, IEEE Transactions on Automatic Control, 50(5), 655-661. Ravazzi, C., Frasca, P., Tempo, R., & Ishii, H. (2015). Ergodic randomized algorithms and dynamics over networks. IEEE transactions on control of network systems, 2(1), 78-87. Sunstein, C. (2012). Breaking up the echo, NY Times, A25. Wang, L., & Xiao, F. (2007). A new approach to consensus problems in discrete-time multiagent systems with timedelays, Science in China Series F: Information Sciences, 50(4), 625-635. Wang, L., & Xiao, F. (2010). Finite-time consensus problems for networks of dynamic agents, IEEE Transactions on Automatic Control, 55(4), 950-955. Wang, L., Jiang, F., Xie, G., & Ji, Z. (2009). Controllability of multi-agent systems based on agreement protocols, Science in China Series F: Information Sciences, 52(11), 2074-2088. Wang, L., Shi, H., Chu, T., Zhang, W., & Zhang, L. (2004). Aggregation of foraging swarms, In Australasian Joint Conference on Artificial Intelligence (pp. 766-777). Springer, Berlin, Heidelberg. Wolfowitz, J. (1963). Products of indecomposable, aperiodic, stochastic matrices, Proceedings of the American Mathematical Society, 14(5), 733-736. Xiao, F., & Wang, L. (2008). Asynchronous consensus in continuous-time multi-agent systems with switching topology and time-varying delays, IEEE Transactions on Automatic Control, 53(8), 1804-1816. Yang, Y., Dimarogonas, D. V., & Hu, X. (2014). Opinion consensus of modified Hegselmann-Krause models, Automatica, 50(2), 622-627. Zheng, Y., & Wang, L. (2014). Containment control of heterogeneous multi-agent systems, International Journal of Control, 87(1), 1-8. Zheng, Y., Ma, J., & Wang, L. (2017). Consensus of Hybrid Multi-agent Systems, IEEE Transactions on Neural Networks and Learning Systems, DOI:10.1109/TNNLS.2017. 2651402. 16 

3cs.SY

1 The Stochastic Geometry Analyses of Cellular Networks with α-Stable Self-Similarity arXiv:1709.05733v1 [cs.IT] 18 Sep 2017 Rongpeng Li, Zhifeng Zhao, Yi Zhong, Chen Qi, and Honggang Zhang Abstract—To understand the spatial deployment of base stations (BSs) is the first step to facilitate the performance analyses of cellular networks, as well as the design of efficient networking protocols. Poisson point process (PPP) has been widely adopted to characterize the deployment of BSs and established the reputation to give tractable results in the stochastic geometry analyses. However, given the total randomness assumption in PPP, its accuracy has been recently questioned. On the other hand, the actual deployment of BSs during the past long evolution is highly correlated with heavy-tailed human activities. The αstable distribution, one kind of heavy-tailed distributions, has demonstrated superior accuracy to statistically model the spatial density of BSs. In this paper, we start with the new findings on α-stable distributed BSs deployment and investigate the intrinsic feature (i.e., the spatial self-similarity) embedded in the BSs. Based on these findings, we perform the coverage probability analyses and provide the relevant theoretical results. In particular, we show that for some special cases, our work could reduce to the fundamental work by J. G. Andrews in [1]. We also examine the network performance under extensive simulation settings and validate that the simulations results are consistent with our theoretical derivations. Index Terms—Stochastic geometry, coverage probability, cellular networks, Poisson point process (PPP), α-stable distributions, self-similarity I. I NTRODUCTION As a key enabler in the information and communications technology (ICT) industry, cellular networks play a decisive role in delivering communication messages and entertainment content [2]. In order to meet the increasing traffic demand, cellular network operators gradually deploy different types of necessary infrastructure including lots of base stations (BSs). To understand the spatial deployment of BSs is the first step to facilitate the performance analyses of cellular networks, as well as the design of efficient networking protocols. In the earliest stages, a two-dimensional hexagonal grid model was used, implying that BSs were deployed regularly, which deviates from the real scenarios. In the recent stages, Poisson point process (PPP) [1], [3]–[5] was assumed, which could roughly model the randomness in the realistic deployment of BSs in cellular networks, meanwhile leading to tractable results. However, given the total randomness assumption in PPP, its practical accuracy has been recently questioned [6], [7]. Consequently, in order to reduce the modeling gap between the PPP model and the realistic spatial deployment of R. Li, Z. Zhao, C. Qi and H. Zhang are with Zhejiang University, Hangzhou 310027, China (email:{lirongpeng, zhaozf, qichen7c, honggangzhang}@zju.edu.cn). Y. Zhong is with Huazhong University of Science and Technology, Wuhan 430073, China (email: yzhong@hust.edu.cn). BSs, other kinds of point processes have been exploited to obtain more precise analytical results [8]. In particular, Poisson cluster processes [9], [10] and Gauss-Poisson process [11] are exploited to model the attractive behavior between BSs, while Matérn hard core point process (MHCP) [12], Ginibre point process (GPP) [8], [13], [14], and determinantal point process [15] are studied to characterize the repulsive feature between BSs [7], [16]–[18]. On the other hand, the actual deployment of BSs during the past long evolution is highly correlated with human activities [19]. Humans tend to live together with societal relationship, and their social behaviors lead to mobile traffic hotspots, thus causing BSs to be more densely deployed in certain areas. Interestingly, according to an observation named “preferential attachment” [20], Barabási et al. argued that many large networks should grow to be heavy-tailed. As its names implies, a heavy-tailed distribution has non-exponential bounded tail. Mathematically, for a heavy-tailed random variable X, the probability Pr(X > x) satisfies lim eκx Pr(X > x) = ∞, x→∞ for all κ > 0. Many well-known statistical distributions including power-law distribution (also named as generalized Pareto distribution), Weibull distribution, log-normal distribution, and α-stable distribution [6], [21], [22] belong to the heavy-tailed family. Therefore, heavy-tailed distributions appear to be more suitable to capture the societal feature and clustering property for the practical BSs. Based on the practical BS deployment information, we have validated that the α-stable distributions demonstrate superior accuracy to statistically model the spatial density of BSs and also outperform other aforementioned heavy-tailed distributions [7], [19], [23]–[25]. But the spatial density alone cannot reach some intuitive conclusions. Fortunately, the α-stable distributions, which have been widely adopted to characterize the distribution of aggregated traffic at the BS level in cellular networks and at the switch level in wired broadband networks, often imply the self-similarity of traffic [22], [26]. As a popular property in complex networks [27], the self-similarity of traffic in the temporal domain means that the distribution of traffic remains invariant under different temporal scales. Similar to the case of traffic, as the practically deployed BSs are αstable distributed, it is natural to ask whether the spatial selfsimilarity holds in cellular networks. Furthermore, compared to other point process models, the inhomogeneous PPP with αstable distributed density becomes applicable with significant merit. The system model based on multiple tiers of PPPs may lead to some tractable results, but it lacks the reasonable explanation to divide the gradually deployed and increasingly 2 denser heterogeneous cells [28] into different tiers. On the contrary, α-stable distributions contribute to understanding the spatial self-similarity and bridging the gap between cellular networks and other social behavior-based complex networks. A. Related Works Stochastic geometry tools have successfully established the reputation to model and analyze wireless networks, as they are able to capture the topological randomness in the network geometry and lead to tractable analytical results [1], [3]– [5], [9]–[18]. There is no doubt that the PPP belongs to the most popular point process used in the literature because of its tractability and has led to a lot of meaningful research results for cellular networks with different kinds of cutting-edge techniques. The baseline operations in singletier and/or multiple-tier downlink cases are examined in [1], [29]–[32] (and references therein) while their counterparts in uplink cases are considered in references like [33]–[38]. User association and load balancing are examined in [39]– [42]. Cognitive, cooperative and intelligent cellular networks are taken into account in [18], [43]–[47]. Energy efficiency, energy harvesting, and BS sleeping for green cellular networks are investigated in [48]–[52]. Besides, multiple-input multipleoutput (MIMO) antenna and in-band full-duplex (FD) systems are studied in [53]–[62]. Also, the PPP-related works have been extended to mobile ad hoc networks [63]. Along with the continuous progress in techniques for physical-layer and network management in wireless networks, the PPP model offers a useful theoretical platform for the initial performance calibration, so as to save economic expenditure and avoid timeconsuming large-scale setup for realistic tests. For example, FD communication, which benefits from the advance in signal processing and self-interference cancellation techniques, is optimistically promoted to double the spectral efficiency for wireless networks [64]. However, given the more complicated interference in FD networks, it is essential to conduct a careful re-examination. Consistent with the system-level simulation results in [64], [65] has taken advantage of the PPP model and demonstrated the potential negative effect that FD communications might impose on the uplink transmission. Although the PPP model provides many useful performance trends, the concern about total independence between nodes (e.g., BSs) has never stopped. Hence, in order to reduce the modeling gap between the PPP model and the practical BS deployment, point processes with either attractive or repulsive spatial correlations have been explored as well. • For attractive point processes, [66] has shown that Poisson cluster process is more suitable to model the spatial distribution of BSs in urban areas. The general clustering nature of deployed BSs in highly populated urban areas clearly reflects the aggregation property of ever-growing traffic demands in cellular networks [7]. [10] has further verified the aggregated interference when the transmitting nodes are modeled by Poisson cluster process and compared the corresponding results with that in the classical PPP model. • On the contrary, [29] and [67] have argued that BSs, in particular macro BSs, in cellular networks tend to be deployed systematically, such that any two BSs are not too close. Thus, a spatial model based on a point process with the repulsive nature seems to be more desirable [29], [67]. Furthermore, the MHCP model [12] possesses limited tractability and leaves many open challenges to be addressed. The GPP model, one typical example of determinantal point processes on the complex plane, has been widely adopted and shows a promising tradeoff between accuracy and tractability [8], [13], [14]. In particular, Deng et al. [8] have shown that the GPP leads to the same trend curve of coverage probability1 as the PPP. Hence, we can come to the following conclusion that the spatial BS density of large-scale cellular networks simultaneously possesses two conflicting features, that is, the density is very large in some clustering regions while being significantly smaller in others. The conclusion is also consistent with our common sense. Furthermore, the heavy-tailed distributions, which generate small values with the high probability but still allow comparatively larger values, fit well to model the spatial BS density. Our previous works [6], [7], [23]–[25] have shown that the α-stable distribution, one kind of heavy-tailed distributions, could be used to perfectly model the spatial BS density in China and Italy, especially in urban areas. In fact, α-stable distributions are not totally new to researchers in the communications field. On one hand, a distribution is called stable if a linear combination of two independent identically distributed random variables has the same distribution [21]. In particular, the Gaussian distribution, which is one specific kind of α-stable distributions with α = 2, has been widely adopted to model the distribution of channel noise. On the other hand, α-stable distributions have been observed for the distribution of wired broadband network traffic [68] since two decades ago, and have recently been found in modeling the cellular network traffic [22], [26]. αstable distributions also owe their importance in both theory and practice to the generalization of the central limit theorem [21] and the accompanying self-similarity [27], [69]–[71] of the stable family. Furthermore, the well-known self-similarity lays the very foundation for the traffic predictability, thus making it possible to perform some actions beforehand. For example, [70] has highlighted the practicability to perform resource reservation schemes for the broadband network traffic with self-similarity. Hence, as the spatial BS density could be better modeled by the α-stable distributions, it is essential to theoretically examine the spatial self-similarity and understand its impact on coverage probability. In other words, we will take the very first step to investigate the spatial self-similarity in the BS deployment and perform stochastic geometry analyses of cellular networks when the BS density is α-stable distributed. B. Contributions Different from the previous works, this paper aims at explaining the implications to statistically model the spatial 1 The coverage probability indicates the probability that the signal-tointerference-plus-noise ratio (SINR) for a mobile user achieves a target threshold. 3 BS density by α-stable distributions and to build the theory to derive the coverage probability mathematically. Belonging to one of the precursor works, we take advantage of a large amount of practical BS deployment records and provide the following key insights: • Firstly, this paper visits and confirms the self-similarity in the spatial deployment of BSs. To our best knowledge, it is the first work in the literature to validate the spatial selfsimilarity among BSs in cellular networks. Therefore, this paper lays the foundation to apply the advance in complex network theory to examine cellular networks. • Secondly, this paper provides the approach to characterize the coverage probability and derive the computable representation when the spatial BS deployment satisfies the self-similarity and the density of the practical BSs obeys α-stable distributions. • Thirdly, this paper theoretically proves that for some special cases, our result could be reduced to the fundamental work by J. G. Andrews in [1]. Besides, this paper studies the coverage probability performance under extensive simulation settings and validates that the simulations results are consistent with our theoretical derivations. The paper also evaluates the impact of self-similarity on the coverage probability performance. The remainder of the paper is organized as follows. In Section II, we present some necessary mathematical background, introduce the realistic dataset of spatial BS deployment and validate the spatial self-similarity. In Section III, we provide the analyses of coverage probability. Section IV evaluates the computable representation obtained by the theoretical analysis and compares it with the classical PPP model. Finally, we conclude this paper in Section V. Our previous works [6], [23] have validated that β = 1 holds for the fitting results of actual spatial BS density in both Hangzhou, China and Rome, Italy. For simplicity of representation, we use the operator b ∼ c to denote that b and c have the same distribution and further have λ ∼ S(α, σ, µ) to indicate that the spatial BS density λ follows the α-stable distributions with β = 1. 2) The self-similarity: The self-similarity has been mostly applied to model time series [69]. Given zero-mean, stationary time series Y = (Yt , t = 1, 2, 3, · · · ), Y is called as self(m) similar process if its m-aggregated series Y (m) = (Yt , t = tm P (m) Yk satisfy 1, 2, 3, · · · ) with each element Yt = k=(t−1)m+1 (m) Yt H ∼ m Yt for all m > 0, where H ∈ (0, 1) is the Hurst parameter indicating the decay rate of statistical dependence of two points with increasing time interval. In other words, the self-similarity implies that the time series is exactly or approximately similar to a part of itself. The concept of self-similarity is also applicable to spatial scenarios. For example, H. ElSawy et al. [16] have shown the aggregated interference from distributed nodes to a receiver obey α-stable distributions, while X. Ge et al. [72] have verified that the statistical characteristics of the wireless cellular coverage boundary possess the self-similarity or fractal property. However, no exact self-similar phenomenon exists in the real world. Most of the self-similar phenomena observed in the real world only have the statistical characteristic. Moreover, the self-similarity of random processes is usually evaluated by the Hurst parameter, which can be estimated using two typical methods and should be in the interval (0, 1). • The rescaled adjusted range statistic (R/S) method: For the random process Yt , the partial sum is dePn noted by Z(n) = Y , and then the sample varit II. M ATHEMATICAL BACKGROUND AND S TATISTICAL t=1 Pn Z(n) 2 t=1 (Yt − n ) M ODELING . ance could be denoted by S 2 (n) = n R(n) A. Mathematical Background The R/S statistic could be formulated as S(n) = i i max (0,Z(i)− n Z(n))− min (0,Z(i)− n Z(n)) Beforehand, Table I summarizes the most used notations in 0≤i≤n 0≤i≤n . The log-log S(n) this paper. relationship between the R/S statistic and n should also 1) α-Stable distributions: A random variable x is said to be linear with a slope of H. obey the α-stable distribution f (x) if there are parameters • The variance-time (V-T) method: This involves the aggre0 < α ≤ 2, σ ≥ 0, −1 ≤ β ≤ 1, and µ ∈ R such that its (m) gated series Yt and the blocking size m. The log-log characteristic function is of the following form: (m) plot of the variance of Yt with respect to m should Φ(ω) = E (exp(jωx)) = follow a straight line with the slope ι if the data are   πα  n o  statistical self-similar. In this case, H = 1 − 2ι . α α  + jµω , α 6= 1;  exp −σ |ω| 1 − jβsgn(ω) tan 2 Moreover, a larger H ∈ (0.5, 1) corresponds to stronger self    2β  similarity.  exp −σ|ω| 1 + j sgn(ω) ln |ω| + jµω , α = 1. π (1) The function E(·) represents the expectation operation with B. Statistical Modeling respect to a random variable. α is called the characteristic As validated in [6], [23], [25], when we divide the region exponent and indicates the index of stability, while β is iden- into several parts and calculate the corresponding probability tified as the skewness parameter. α and β together determine density function (PDF) of the spatial BS density in each part, the shape of the distribution. Moreover, σ and µ are called we observe that the PDF follows the α-stable distributions. scale and location shift parameters, respectively. In particular, For example, Fig. 1(a) depicts that a region is divided into if α = 2, the α-stable distribution reduce to the Gaussian four parts. The spatial BS density in each part λ1 , λ2 , λ3 and λ4 are mutually dependent and obey the α-stable distributions distribution. 4 TABLE I A Λ(s, b, c) Ξ(s, b, c) The coverage probability α-stable distributions with stability α, skewness β = 1, scale σ and shift µ Pathloss exponent The Rayleigh fading gain for the signal link The fading exponent for the interfering links Rayleigh fading factor Noise factor The Hurst parameter The self-similarity zooming parameter The predefined radius of inner circle in Fig. 1 The spatial density of BSs The number of BSs within an r-radius R circle The standard Gamma function Γ(d) = 0∞ tRd−1 e−t dt The incomplete Gamma function Γ(d, x) = x∞ td−1 e−t dt
def Θ(s, b, c) = (sg)2/δ Γ(− 2δ + 1, sgb−δ ) − Γ(− δ2 + 1, sgc−δ ) 
 
 def Λ(s, b, c) = c2 1− exp −sgc−δ − b2 1 − exp −sgb−δ def Ξ(s, b, c) = πEg Λ(s, b, c) − Θ(s, b, c) – α = 0.6, σ = 0.25 and µ = 0.25 δ=4 – Assumed to follow Rayleigh fading. ζ=1 N0 = 1 (i.e., SNR = 0 dB) H = 0.9 a=2 R = 40 – – – – (a) (b) l1 – (c) N 3 N 2 N1 3R l3 lS l1 R l4 – 2R l2 – aR Θ(s, b, c) Default Value in Sec. IV R pc S(α, σ, µ) δ h g ζ N0 H a R λ N (r) Γ(d) Γ(d, x) Meaning r Symbol LIST OF THE MAIN SYMBOLS AND FUNCTIONS IN THE PAPER . Fig. 1. An illustration of the self-similarity for the spatial BS deployment. with the same values α, σ and µ. The correlation between BS density in different parts makes it challenging to directly derive the coverage probability. Instead, it is essential to take the statistical modeling of the correlation into consideration. In that regards, self-similarity emerges as a promising technique to characterize the correlation of the spatial density, as it has manifested its importance and effectiveness by modeling the correlation in different scales. In order to show the self-similarity in spatial BS deployment, we take advantage of the collected BS deployment records in Hangzhou, China and Rome, Italy2 . Fig. 2(a) and 2(d) illustrate the corresponding deployment situation in both cities. Similar to the common validation process in temporal dimension [69], we verify the accuracy of self-similarity in spatial BS deployment as follows. As depicted in Fig. 1(b), we randomly select a point as the starting “origin point” and thus get some concentric circles with increasingly larger radii (e.g., from R to 2R, 3R, · · · ). Then, we could get a series, each value corresponding to the number of BSs in one circle. Afterwards, we could apply the aforementioned R/S method and V-T method in Section II-A2. Fig. 2(b)(c) and Fig. 2(e)(f) 2 Interested readers could also refer to [25] to check our preliminary results on the spatial self-similarity in a more systematic manner. provide the corresponding log-log plots for both cities. From these figures, it can be observed that all the estimated Hurst parameters are very close to 1. Therefore, we could boldly argue that the spatially deployed BSs possess the spatial selfsimilarity. Then, according to the definition of self-similarity, the number of BSs N (aR) within a circle of radius aR should satisfy that N (aR) ∼ aH N (R), ∀R, a > 0, 0 ≤ H < 1, where a indicates the self-similarity zooming parameter and H denotes the Hurst parameter. Remark 1. For cellular networks, the number of BSs exhibit the spatial self-similarity. III. T HE C OVERAGE P ROBABILITY A NALYSES In this part, we derive the coverage probability in a downlink cellular network. Specifically, the coverage probability is defined as the probability that the SINR for a mobile user achieves a target threshold. Mathematically, the coverage probability is formulated as   hr−δ P(SINR > T ) = P >T , (2) N0 + Ir where δ denotes the pathloss exponent factor of the standard propagation channel. N0 is the additive white Gaussian noise 5 (a) BS Deployment in Hangzhou, China 2.5 2 200 1.5 150 0.5 50 0 0 50 150 100 150 1.94 1.92 1.9 -0.5 0 log(Var) vs. log(m) fitting line 1.96 1 100 (c) V-T Method H=0.97553 1.98 log(R/S) vs. log(n) fitting line log(Var) 250 log(R/S) km 300 (b) R/S Method H=0.96645 1.88 0 0.5 1 1.5 2 0 0.5 1 km log(n) log(m) (d) BS Deployment in Rome, Italy (e) R/S Method H=0.84158 (f)V-T Method H=0.86876 2 1.2 log(R/S) vs. log(n) fitting line 1.5 1.5 log(Var) vs. log(m) fitting line 1.1 1 log(Var) km log(R/S) 100 0.5 1 0.9 50 0 0.8 -0.5 0 0 50 100 150 200 250 300 0.7 0 0.5 km 1 1.5 2 0 0.5 log(n) 1 1.5 log(m) Fig. 2. The self-similarity modeling results for the spatial BS deployment in Hangzhou and Rome. (AWGN) factor and T is the target threshold. Besides, assuming that the fading follows the Rayleigh fading, we have h ∼ exp(ζ). Without loss of generality, we assume the mobile user is located at the “origin point”. When the distance from a mobile user to its serving BS b0 is r, Ir denotes the cumulative interference from all other BSs i (except BS b0 ) to the mobile user. A. Impact of α-Stable Distributions and Self-similarity Before delving into the coverage probability, we first talk about how the α-stable distributions and the self-similarity will mathematically shape the BS deployment? As depicted in Fig. 1(c), we assume that there exists a specific R (i.e., R > r) to divide the whole region of interest into two parts. Besides, the spatial BS density λ for the region within the radius R satisfies λ ∼ S(α, σ, µ). Moreover, we can have the following lemma to characterize the spatial BS density λS for the region outside the radius R. Lemma 1. The spatial BS density λ has a Laplace transform    σα α    exp − cos πα s − µs , α 6= 1; 2   Ψ(s) = E (exp(−sλ)) =  2σ   exp s ln s − µs , α = 1. π (3) while the spatial BS density λS will follow α-stable distributions (i.e., λS ∼ λaH−2 ) and the corresponding Laplace transform could be formulated as ΨS (s) = E (exp(−sλS ))    σ α aα(H−2) α  H−2  exp − s − sa µ , α 6= 1;    cos πα  2   = 2σaH−2 H−2  exp s [ln s + (H − 2) ln a] − sa µ ,   π    α = 1. (4) Proof. Following the definition of self-similarity, the number of BSs in the interfering region is coupled with the selfsimilarity zooming parameter a and could be formulated as N (aR) = λS πa2 R2 ∼ aH N (R) = λπaH R2 . Therefore, λS ∼ λaH−2 . (5) When x ∼ S(α, σ, 0), we get the corresponding Laplace transform (Proposition 1.2.12, [21]) as    σα α    exp − cos πα s , α 6= 1; 2   Ψ(s) = E (exp(−sx)) =  2   exp σ s ln s , α = 1. π (6) 6     α (πσ)α 2α−1   exp −µπr2 − (πσ) r2α 2µπr + 2α , α 6= 1; r cos πα cos πα f (r) = 2 2   exp −µπr2 + 2σr2 ln(πr2 ) 2µπr + 4σr ln(πr2 ) + 4σr
, α = 1. (     σ α aα(H−2) h    Λ(s, R, aR) πE Λ(s, r, R) − Θ(s, r, R) − exp − λπE g g   cos πα  2   )     iα   H−2   − Θ(s, R, aR) −a µπEg Λ(s, R, aR) − Θ(s, R, aR) , α 6= 1;   ( LIr (s) =  o  i  n h     exp 2σaH−2 ln πEg Λ(s, R, aR) − Θ(s, R, aR) + (H − 2) ln a − πaH−2 µ      )           · Eg Λ(s, R, aR) − Θ(s, R, aR) − λπEg Λ(s, r, R) − Θ(s, r, R) , α = 1.  ΨS (s) By merging Eq. (6) and Eq. (7), we obtain the result. (12) 0, ∀λ < 0. The equation (b) is originated from the Laplace transform for a variable k ∼ S(α, σ, 0) in Eq. (6). So, the PDF of r Then, the Laplace transform of λS could be derived as
= E exp(−saH−2 (x + µ))  = Ψ(saH−2 ) · exp −saH−2 µ . (8) (7) f (r) =  B. The PDF of the Distance from Serving BS to UE In this paper, we adopt the minimum distance as the user association metric. Therefore, when the distance from a mobile user to its serving BS b0 is r, all distances from the interference BSs to the target mobile user must be larger than r. The derivation methodology of the PDF of r could basically follow the well-established lines in [1]. However, as the BS density varies in space and obeys the α-stable distributions, the derivations should be re-considered. Therefore, we give the following the theorem. Theorem 1. The PDF of the distance from the closest BS r to the target mobile user in a cellular network with α-stable distributed BS density is expressed in Eq. (8). Proof. Firstly, we take advantage of the simple fact that the null probability of a 2-D Poisson process in an area A with the given density λ is exp(−λA). Since λ ∼ S(α, σ, µ), we have P [r > rT ] = P(No BS closer than rT ) Z ∞ 2 e−λπ(rT ) f (λ)dλ = Z0 ∞ Z ∞ 2 (a) −λπ(rT )2 = e f (λ)dλ = e−(k+µ)π(rT ) f (k)dk (9) −∞ −∞   α (πσ)  2α  exp −µπ(rT )2 − , α 6= 1; (rT ) (b) cos πα = 2    exp −µπ(r )2 + 2σ(r )2 ln(π(r )2 ) , α = 1, T T T where the equation (a) is based on the fact that the spatial density in physical worlds should be non-negative, or f (λ) = dF (r) d (1 − P [r > rT ]) = dr drT (10) Then, merging Eq. (9) with Eq. (10) leads to the result.  Corollary 1. For homogeneous PPP with static density λ, the PDF of r could reduce to f (r) = exp(−λπr2 )2πλr. Similarly, for varying distribution λ, the conditional probability P(r|λ) = exp(−λπr2 )2πλr. This corollary could be achieved by directly applying Theorem 1 with a static density λ = µ and thus σ = 0. It is also consistent with the conclusions in [1]. C. Main Result We now state our main result for the coverage probability analysis. Generally, the coverage probability pc is mainly determined by the SINR threshold T , the spatial BS density distribution S(α, σ, µ), and the pathloss exponent δ. In some sense, pc (T, S(α, σ, µ), δ) is tightly coupled with the cumulative interference Ir for the α-stable distributed and selfsimilar BS deployment. The following lemma characterizes the relationship between pc (T, S(α, σ, µ), δ) and Ir . Lemma 2. The coverage probability pc (T, S(α, σ, µ), δ) could be obtained from the following formula, that is, pc (T, S(α, σ, µ), δ) ZZ δ e−ζT r N0 LIr (ζT rδ )P(r|λ)f (λ)drdλ, = (11) r>0,λ>0 where LIr (s) is the Laplace transformation of random variable Ir evaluated at s conditioned on the distance to the closest BS from the origin. The Rayleigh fading coefficient is assumed to satisfy h ∼ exp(ζ). We leave the proof of Lemma 2 in Appendix A, as the proof could be simply obtained by calculating the condition probability on the distance r from the serving BS to the mobile user. 7   exp −λS L∆S (s) = (1 − f (x))dx f (λS )dλS λS =0 ∆S Z α   Z  σ α aα(H−2) H−2   (1 − f (x))dx (1 − f (x))dx , α 6= 1; −a µ   exp − cos πα ∆S ∆S (a) 2 ( (16) =   Z    Z ) H−2  2σa  H−2  ln (1 − f (x))dx + (H − 2) ln a − a µ (1 − f (x))dx , α = 1.  exp π ∆S ∆S  ( !) !#α "  σ α aα(H−2)  H−2  exp − , α 6= 1; −a µπEg Λ(s, R, aR) − Θ(s, R, aR) πEg Λ(s, R, aR) − Θ(s, R, aR)    cos πα  2   !) ! (" (   (b) H−2 = + (H − 2) ln a exp 2σa ln πEg Λ(s, R, aR) − Θ(s, R, aR)     !) #     H−2  , α = 1. − πa µ · Eg Λ(s, R, aR) − Θ(s, R, aR)   Z ∞ Z Next, we focus on how to calculate the Laplace transform of the cumulative interference LIr (s), when the BS deployment obey α-stable distributions and self-similarity. In Section III-A, we divide the region of interest into two concentric circles. For the inner circle ∆ with a radius R, the impact of the spatial BS density λ is similar to the impact of r and explicitly represented by the latter’s PDF f (r). Meanwhile, the outer circle ∆S with the radius spanning from R to aR has a spatial BS density λS ∼ λaH−2 and understanding the impact of the interference from BSs in the outer circle is one of the core contributions Pof this paper. ForPsimplicity of representation, we let Ir = i∈∆/b0 gi R̂i−δ + i∈∆S gi R̂i−δ to denote the cumulative interference from all the other BSs i (except the serving BS b0 ) to the mobile user with the distance R̂i and pathloss gi . We could obtain the following lemma, Lemma 3. The Laplace transform of the cumulative interference Ir for a cellular network with α-stable distributed def BSs could be formulated as Eq. (12), where
Θ(s, b, c) = 2 −δ 2/δ ) − Γ(− 2δ + 1, sgc−δ ) . Γ(d, x) = (sg) δ + 1, sgb R ∞ d−1 Γ(− −t t e dt Rdenotes the incomplete Gamma function, x ∞ while Γ(d) = 0 td−1 e−t dt denotes the standard Gamma 
 def function. Besides, Λ(s, b, c) = c2 1 − exp −sgc−δ − 
 b2 1 − exp −sgb−δ . The equation (a) comes from the identical, independent distribution of gi and its further independence from the point process. From Eq. (13), it can be observed that the Laplace transform his composedh of two parts ii (i.e., Q −δ and exp −sg R̂ E L∆ (s) = E∆ g i ii  hQ i∈∆/b0 h −δ ), which L∆S (s) = E∆S i∈∆S Eg exp −sg R̂i is also consistent with our intuition. Next, we derive the representation of these two parts separately. • For L∆ (s), we have a fixed spatial density λ. Then, from  Q the Campbell theorem [12], x∈∆ f (x) =
we have E R exp −λ ∆ (1 − fh(x))dx i  . Since f (x) = Eg exp −sg R̂−δ , we have Z (1 − f (x))dx ∆ Z R
   vdv 1 − Eg exp −sgv −δ = 2π r( ) (a) = πEg Λ(s, r, R) − Θ(s, r, R) We leave the derivation of the equation (a) in Appendix B. Accordingly, we have !# " Proof. According to the definition of Laplace transform, we have LIr (s)   = EIr e−sIr   = E∆  · E∆S " Y i∈∆/b0 Y i∈∆S L∆ (s) = exp −λπEg Λ(s, r, R) − Θ(s, r, R) (15) For L∆S (s), as λS ∼ λaH−2 , we have Eq. (16), where the equation (a) comes from the Laplace transform of λS in Lemma 1. RThe equation (b) is achieved by applying the result ∆S (1 − f (x))dx = πEg {Λ(s, R, aR) − Θ(s, R, aR)}, which can be obtained by adopting a similar methodology as the derivation of Eq. (14). Combining Eq. (15) and Eq. (16), we obtain the result.  •  = E∆+∆s ,gi exp −s (a) (14) X i∈(∆+∆s )/b0 gi R̂i−δ   i  h −δ  Eg exp −sg R̂i # i h  −δ Eg exp −sg R̂i  (13) With the results above, we could obtain the final result in the following theorem. 8 pc (T, S(α, σ, µ), δ) Z n σ α aα(H−2)  α = 2πr exp − Ξ(ζT rδ , R, aR) − aH−2 µΞ(ζT rδ , R, aR) − µΞ(ζT rδ , r, R) πα cos 2 r>0 o h σα α i σα δ 2 α 2 δ δ 2 α−1 · (Ξ(ζT r , r, R) + πr ) − µπr − ζT r N (Ξ(ζT r , r, R) + πr ) + µ dr − 0 cos πα cos πα 2 2 pc (T, S(α, σ, µ), δ) ( Z 2πr exp = r>0 δ h 2σaH−2 i 
ln Ξ(ζT rδ , R, aR) + (H − 2) ln a − aH−2 µ Ξ(ζT rδ , R, aR) π 2σΞ(ζT r , r, R) + 2σπr2 ln(Ξ(ζT rδ , r, R) + πr2 ) − µΞ(ζT rδ , r, R) − µπr2 ) π  i h 2σ  ln(Ξ(ζT rδ , r, R) + πr2 ) + 1 + µ dr − ζT rδ N0 · − π + Υ(r) = Z λ>0 ( " δ δ ! exp −λπ Eg Λ(ζT r , r, R) − Θ(ζT r , r, R) +r 2 #) Proof. From Lemma 2, the coverage probability could be calculated as o σα δ 2 α 2 δ πα (Ξ(ζT r , r, R) + πr ) − µπr − ζT r N0 cos 2 h σα α i δ 2 α−1 · (Ξ(ζT r , r, R) + πr ) + µ dr (21) cos πα 2 2) if α = 1 r>0 where the equation (a) is the direct result of Corollary 1 and Lemma 3 and Υ(r) couldRbe formulated as Eq. (20), where the equation (b) is due to x e−sx xf (x)dx = − dΦ(s) ds . Merging Eq. (20), Eq. (19), and Lemma 3, we have the result.  When we consider the whole region (i.e., the outer circle spanning from R to ∞, or a → 0), we get the following corollary. Corollary 2. When a → ∞, the coverage probability could be reduced to 1) if α 6= 1 Z n 2πr exp − µΞ(ζT rδ , r, R) pc (T, S(α, σ, µ), δ) = r>0 Z r>0 (19) (20) − pc (T, S(α, σ, µ), δ) = pc (T, S(α, σ, µ), δ) ZZ δ e−ζT r N0 LIr (ζT rδ )P(r|λ)f (λ)drdλ = Z r>0,λ>0 δ (a) 2πre−ζT r N0 L∆s (ζT rδ )Υ(r)dr = (18) λf (λ)dλ  n o σα   exp − (Ξ(ζT rδ , r, R) + πr2 )α − µΞ(ζT rδ , r, R) − µπr2  πα  cos 2    h σα α i   δ 2 α−1  · (Ξ(ζT r , r, R) + πr ) + µ , α 6= 1  (b) cos πα 2 = n 2σΞ(ζT rδ , r, R) + 2σπr2     ln(Ξ(ζT rδ , r, R) + πr2 ) − µΞ(ζT rδ , r, R) exp   π   oh 2σ   i    − µπr2 − ln(Ξ(ζT rδ , r, R) + πr2 ) + 1 + µ , α = 1 π Theorem 2. The coverage probability in a cellular networks with α-stable distributed BS density and the self-similarity is given by Eq. (17) and Eq. (18) for α 6= 1 and α = 1, respectively, where Ξ(s, b, c) = πEg (Λ(s, b, c) − Θ(s, b, c)). (17) 2πr exp ( − µΞ(ζT rδ , r, R) − µπr2 2σΞ(ζT rδ , r, R) + 2σπr2 − ζT rδ N0 + ln(Ξ(ζT rδ , r, R)+ π ) i h 2σ  ln(Ξ(ζT rδ , r, R) + πr2 ) + 1 dr (22) πr2 ) · µ − π Proof. Firstly, we have h  r δ i lim a2 R2 1 − exp −ζT g( ) a→∞ aR
  r δ R2 1 − exp −ζT g( aR ) = lim 2 a→∞ 1/a
  r δ δ ) k R2 1 − exp −ζT g( R = lim+ (23) k2 k→0 δ   −ζT g(δ − 1) Rrδ−2 r (a) exp −ζT g( )δ k δ k δ−2 = lim+ 2 R k→0 =0 where the equation (a) comes from the l’Hôpital’s Rule and applies the fact that the pathloss exponent δ ≥ 2. 9 #  α−1  2(ζT g)2/δ r2  2 2  σα πEg Γ(− , ζT g) − Γ(− ) +µ · pc (T, S(α, σ, µ), δ) = 2πr cos πα δ δ δ r>0 2 !α (  2(ζT g)2/δ r2  2 2  σα α Γ(− , ζT g) − Γ(− ) − ζT rδ N0 exp − πEg cos πα δ δ δ 2   !) 2(ζT g)2/δ r2 2 2 − µπEg Γ(− , ζT g) − Γ(− ) dr δ δ δ ! Z  2(ζT g)2/δ r2   i h 2 2σ 2  ln πEg +1 · Γ(− , ζT g) − Γ(− ) 2πr µ − pc (T, S(α, σ, µ), δ) = π δ δ δ r>0 (  2(ζT g)2/δ r2  2(ζT g)2/δ r2  2 2  Γ(− , ζT g) − Γ(− ) ln πEg exp − ζT rδ N0 2σEg δ δ δ δ ! )    2/δ 2  2(ζT g) r  2 2  2  2 − µπEg dr Γ(− , ζT g) − Γ(− ) Γ(− , ζT g) − Γ(− ) δ δ δ δ δ " Z Hence, as a → ∞, we have aH−2 → 0 for H ∈ (0, 1). So, a H−2
  lim R2 1 − exp −ζT g( Rr )δ R→∞ ! = aH−2 πEg Λ(s, b, c) − Θ(s, b, c) h  r i =a πEg − R 1 − exp −ζT g( )δ R ! r 2 − (ζT g)2/δ r2 Γ(− + 1, ζT g( )δ ) δ R = πEg (24) − r2 [1 − exp (−ζT g)] − (ζT g)2/δ r2 · ! h  r i 2 Γ(− + 1, ζT g) + πEg R2 1 − exp −ζT g( )δ δ R ! r
2 (27) + (ζT g)2/δ r2 Γ − + 1, ζT g( )δ δ R 2 − r2 [1 − exp (−ζT g)] + (ζT g)2/δ r2 Γ(− + 1) δ ! 2 − (ζT g)2/δ r2 Γ(− + 1, ζT g) δ  ! 2 2 2(ζT g)2/δ r2 (a) 2 Γ(− , ζT g) − Γ(− ) = −πr + πEg δ δ δ → πEg →0 From Theorem 2, it can be observed that when → ∞, only the term aH−2 Ξ(ζT rδ! , R, aR) = will be af- fected. We have the conclusion.  Based on Corollary 2, we have the following remark. Remark 2. When we expand the interference to the whole region, the cumulative interference from the outer circle could be negligible and the coverage probability is mainly contingent on the SINR within the inner circle. When R → ∞, we can further simplify the results to get more interesting insight. Theorem 3. When R → ∞, the coverage probability could be reduced to Eq. (25) and Eq. (26) for α 6= 1 and α = 1, respectively. Proof. Similar to  the lines to prove
Corollary 2, we r δ ) = 0 and have lim a2 R2 1 − exp −ζT g( aR R→∞ 2. Ξ(ζT rδ , r, R) = πEg Λ(s, b, c) − Θ(s, b, c) 2 aH−2 πEg Λ(ζT rδ , R, aR) − Θ(ζT rδ , R, aR) ≥ δ ! h  r δ i + aH−2 πEg a2 R2 1 − exp −ζT g( ) aR ! r δ 2 2/δ 2 ) ) + (ζT g) r Γ(− + 1, ζT g( δ aR a 0 for all δ (26) Then lim Ξ(ζT r , R, aR) = 0. On the other hand, as R→∞ R → ∞, δ Ξ(ζT r , R, aR) H−2 = (25) where the Equation (a) comes from Γ(s, x) = (s − 1)Γ(s − 1, x) + xs−1 e−x and Γ(s) = (s − 1)Γ(s − 1). There, we obtain the result in the theorem after simple mathematical manipulations.  Next, we state the coverage probability when extra constraints are imposed, that is, the spatial density in the inner circle is fixed (i.e., σ = 0 and λ = µ in Theorem 3). We get the following corollary. Corollary 3. The coverage probability in cellular networks with fixed spatial BS density is pc (T, S(α, 0, λ), δ) ( Z 2(ζT g)2/δ r2 = 2πµr exp − ζT rδ N0 − µπEg δ r>0 10 1 0.9 -Stable theory, SNR = 20 dB -Stable theory, SNR = 0 dB PPP theory, SNR = 20 dB PPP theory, SNR = 0 dB -Stable simulation, SNR = 20 dB -Stable simulation, SNR = 0 dB PPP simulation, SNR = 20 dB PPP simulation, SNR = 0 dB Coverage Probability 0.8 0.7 0.6 0.7 0.5 0.4 0.3 0.6 0.5 0.4 0.3 0.2 0.2 0.1 0.1 0 -10 -Stable, = 0.25 -Stable, = 25 -Stable, = 100 PPP, = 0.25 PPP, = 25 PPP, = 100 0.8 Coverage Probability 0.9 -8 -6 -4 -2 0 2 4 6 8 0 -10 10 -8 -6 -4 SINR Threshold (dB) Fig. 3. The coverage probability comparison under different SNR environment. !) dr 0 2 4 6 8 10 Fig. 4. The coverage probability comparison under various (average) spatial BS densities. 0.9 (28) Remark 3. When the spatial density is fixed and equals λ, the self-similarity patterns no longer take effect and our result could be reduced to the well-recognized conclusions obtained by J. G. Andrews et al. in [1]. IV. N UMERICAL A NALYSES In this part, we provide numerical evaluations through which we compare the coverage probability of the cellular network with α-stable distributed BS density and the classical PPP model. Our simulation parameters are configured according to Table I. In particular, the default spatial BS density for the PPP model is consistent with that in [1]. As [21] states that when α ∈ (0, 1), the mean value of S(α, σ, µ) equals µ. Therefore, it is fair to compare the network with α-stable distributed BS density and the classical PPP model. Firstly, we leverage the theoretical analysis results and give the coverage probability comparison under different AWGN environment in Fig. 3. From the figure, the network with αstable distributed BS density exhibits significant performance improvement than the PPP model in low SNR environment while the performance gap becomes gradually narrow along with the increase of SNR. On the other hand, the performance gap between SNR = 0 dB and SNR = 20 dB is only 4 dB and 6 dB for the α-stable model and the PPP, respectively, which demonstrates that the cellular network is interferencelimited rather than noise-limited. Fig. 3 also depicts the results by Monte Carlo simulations (see the dashed or dotted curves) and each curve is averaged under 15000 iterations. It can be observed that the results obtained by Monte Carlo simulations match well with those obtained by the theoretical analyses, which further verifies the correctness of our theoretical derivations. Fig. 4 illustrates interesting coverage probability comparison under various (average) spatial BS densities. It shows that µ = 0.25 has lead to remarkable coverage probability and -Stable, -Stable, -Stable, PPP 0.8 = 0.5 = 0.7 = 0.9 0.7 Coverage Probability h 2 i 2 Γ(− , ζT g) − Γ(− ) δ δ -2 SINR Threshold (dB) 0.6 0.5 0.4 0.3 0.2 0.1 -10 -8 -6 -4 -2 0 2 4 6 8 10 SINR Threshold (dB) Fig. 5. The impact of α on the coverage probability comparison. also made the networks to be interference-limited. Therefore, further increasing the spatial density would add more to the aggregated interference and impose negative impact on the coverage probability. It also implies that the influence of the spatial BS density on the coverage probability is rather sophisticated. Fig. 5 and Fig. 6 further examine the impact of α and σ. They show that when α becomes larger or σ gets smaller, the network coverage probability increases. In some cases, the network with α-stable self-similarity could produce interior performance than the PPP model. Next, we focus on the impact of the spatial self-similarity on the coverage probability. Recalling the statements in Section II-A2, the spatial self-similarity is examplified by the relationship N (aR) ∼ aH N (R) or λS ∼ aH−2 λ. Therefore, we examine the coverage probability when a and H differs and provide the corresponding results in Table II. From the table, we observe increasing H or decreasing a will result in a slight reduction of the coverage probability. The trivial impact of a and H on the network performance can be explained as that a variation of H or a will make the spatial BS density 11 TABLE II A SUMMARY OF THE NETWORK COVERAGE PROBABILITY UNDER DIFFERENT a AND SNR -10 dB 0 dB 10 dB a=2 H = 0.9 a = 20 H = 0.9 α-Stable a = 200 H = 0.9 a=2 H = 0.1 a=2 H = 0.5 PPP 0.8015 0.4437 0.1355 0.8019 0.4441 0.1357 0.8020 0.4443 0.1357 0.8016 0.4439 0.1356 0.8016 0.4438 0.1355 0.7531 0.3580 0.1153 0.9 0.9 -Stable, -Stable, -Stable, -Stable, PPP 0.8 = 0.25 = 2.5 = 25 = 250 -Stable, R = 40 -Stable, R = 1 -Stable, R = 0.5 PPP 0.8 0.7 Coverage Probability Coverage Probability 0.7 0.6 0.5 0.4 0.3 0.6 0.5 0.4 0.3 0.2 0.2 0.1 0 -10 H. -8 -6 -4 -2 0 2 4 6 8 0.1 -10 10 -8 -6 -4 SINR Threshold (dB) -2 0 2 4 6 8 10 SINR Threshold (dB) (a) The impact of R Fig. 6. The impact of σ on the coverage probability comparison. 0.9 -Stable, = 2.5 -Stable, = 3 -Stable, = 4 PPP, = 2.5 PPP, = 3 PPP, = 4 0.8 0.7 Coverage Probability λS of the outer circle larger from a probabilistic sense, thus making BSs over-crowded and generating huge interference. This phenomenon that stronger self-similarity incurs negative impact is also consistent with its counterpart of self-similar Ethernet traffic. As stated in [73], stronger self-similarity in Ethernet traffic will make heavier traffic prone to arrive in a sequel and congest the network. We continue the performance analyses in Fig. 7(a) and Fig. 7(b). Fig. 7(a) shows reducing R will cause a reduction of coverage probability for low SINR threshold but an increment for high SINR threshold. On the other hand, consistent with our intuition, Fig. 7(b) depicts that a smaller pathloss exponent will bring a lower coverage probability. However, the network with α-stable distributed BSs always demonstrates relatively larger coverage probability than the PPP model. 0.6 0.5 0.4 0.3 0.2 0.1 0 -10 -8 -6 -4 -2 0 2 4 6 8 10 SINR Threshold (dB) (b) The impact of δ Fig. 7. Performance sensitivity analyses under different R and δ. V. C ONCLUSION In this paper, we have performed the stochastic geometry analyses in cellular networks with α-stable self-similarly deployed BSs. Based on the practical BS deployment records in China and Europe (Italy), we have validated the spatial self-similarity among BSs. Afterwards, we have theoretically derived the coverage probability for cellular networks with self-similarity and provided a tractable result. In particular, we have demonstrated that our analytical results could be reduced to the works achieved by J. G. Andrews [1]. We have also simulated the coverage probability performance under extensive parameter settings and verified the consistence between the theoretical and simulation results. Our results have shown that compared to the classical PPP model, the modeling with α-stable self-similarity could contribute to analyze the performance under more sophisticated network configurations, thus making it easier to understand the actual network variations. There still exist some open questions to be addressed. For example, our work has shown that instead of improving the coverage performance, simple yet stubborn deployment of BSs incurs significant interference and degrades the coverage performance. Therefore, frequency reuse has been applied in practical cellular communication. In this case, it is still mean- 12 ingful to combine α-stable self-similarity with more realistic network configurations, so as to produce more valuable results. A PPENDIX A. Proof of Lemma 2 Proof. According to the definition of the coverage probability, we have pc (T, S(α, σ, µ), δ) = Er [P(SINR > T |r)] ZZ P(SINR > T |r)P(r|λ)f (λ)drdλ = r>0,λ>0   ZZ hr−δ > T |r P(r|λ)f (λ)drdλ P = N0 + Ir r>0,λ>0 (29) Next, recalling that theh pathloss in Rayleigh fading satisfies i hr −δ h ∼ exp(ζ), the term P N0 +Ir > T |r will satisfy   hr−δ > T |r N0 + Ir   = P h > rδ T (N0 + Ir )|r    = EIr P h > rδ T (N0 + Ir )|r, Ir   = EIr exp(−ζrδ T (N0 + Ir ))|r, Ir P = e−ζT r δ N0 (30) LIr (ζT rδ ) Combining Eq. (29) and Eq. (30), we get the result.  B. Derivations of Eq. (14) In this part, we provide the details about the equation (a) in the Eq. (14). Z Z R
  (1 − f (x))dx = 2π vdv 1 − Eg exp −sgv −δ ∆ r Z ∞Z R 
1 − exp −sgv −δ vdvf (g)dg (31) = 2π =π Z 0 0 ∞ Z |r r R−δ −δ 2 {1 − exp (−sgy)} dy − δ f (g)dg {z } (>) n o = πEg Λ(s, r, R) − Θ(s, r, R) The last equation holds, since for the inside integral (>), we have 2 R−δ (>) = y − δ [1 − exp (−sgy)] −δ r Z R−δ 2 2 (32) (sgy)− δ exp (−sgy) dy − (sg) δ r −δ
 
  = R2 1 − exp −sgR−δ − r2 1 − exp −sgr−δ   2 2 2 −δ −δ δ − (sg) Γ(− + 1, sgr ) − Γ(− + 1, sgR ) δ δ R EFERENCES [1] J. G. Andrews, F. Baccelli, and R. Ganti, “A tractable approach to coverage and rate in cellular networks,” IEEE Trans. Wireless Commun., vol. 59, no. 11, pp. 3122–3134, Nov. 2011. [2] R. Li, Z. Zhao, X. Zhou, G. Ding, Y. Chen, Z. Wang, and H. Zhang, “Intelligent 5G: When cellular networks meet artificial intelligence,” IEEE Wireless Commun., Mar. 2017, (early access). [3] J. G. Andrews, H. Claussen, M. Dohler, S. Rangan, and M. C. Reed, “Femtocells: Past, present, and future,” IEEE J. Sel. Area. Comm., vol. 30, no. 3, pp. 497–508, Apr. 2012. [4] J. G. Andrews, S. Singh, Q. Ye, X. Lin, and H. Dhillon, “An overview of load balancing in HetNets: Old myths and open problems,” arXiv:1307.7779 [cs, math], Jul. 2013. [Online]. Available: http://arxiv.org/abs/1307.7779 [5] M. Haenggi, J. G. Andrews, F. Baccelli, O. Dousse, and M. Franceschetti, “Stochastic geometry and random graphs for the analysis and design of wireless networks,” IEEE J. Sel. Area. Comm., vol. 27, no. 7, pp. 1029–1046, Sep. 2009. [6] Y. Zhou, R. Li, Z. Zhao, X. Zhou, and H. Zhang, “On the α–stable distribution of base stations in cellular networks,” IEEE Commun. Lett., vol. 19, no. 10, pp. 1750–1753, Oct. 2015. [7] Y. Zhou, Z. Zhao, Y. Louet, Q. Ying, R. Li, X. Zhou, X. Chen, and H. Zhang, “Large-scale spatial distribution identification of base stations in cellular networks,” IEEE Access, vol. 3, pp. 2987 – 2999, Dec. 2015. [8] N. Deng, W. Zhou, and M. Haenggi, “The Ginibre point process as a model for wireless networks with repulsion,” IEEE Trans. Wireless Commun., vol. 14, no. 1, pp. 107–121, Jan. 2015. [9] M. Haenggi and R. K. Ganti, “Interference in large wireless networks,” Foundations and Trends in Networking, vol. 3, no. 2, pp. 127–248, Nov. 2009. [Online]. Available: http://www.nowpublishers.com/article/Details/NET-015 [10] Y. J. Chun, M. O. Hasna, and A. Ghrayeb, “Modeling heterogeneous cellular networks interference using poisson cluster processes,” IEEE J. Sel. Area. Comm., vol. 33, no. 10, pp. 2182 – 2195, May 2015. [11] A. Guo, Y. Zhong, W. Zhang, and M. Haenggi, “The Gauss-Poisson process for wireless networks and the benefits of cooperation,” IEEE Trans. Commun., vol. 64, no. 5, pp. 1916–1929, May 2016. [12] S. N. Chiu, D. Stoyan, W. S. Kendall, and J. Mecke, Stochastic geometry and its applications,, 3rd ed. Wiley, Jul. 2013. [Online]. Available: http://www.wiley.com/WileyCDA/WileyTitle/productCd-0470664819.html [13] N. Miyoshi and T. Shirai, “A cellular network model with Ginibre configured base stations,” Adv. in Appl. Probab., vol. 46, no. 3, pp. 832–845, Sep. 2014. [14] T. Kobayashi and N. Miyoshi, “Uplink cellular network models with Ginibre deployed base stations,” in Proc. IEEE ITC 2014, Karlskrona, Sweden, Sep. 2014. [15] Y. Li, F. Baccelli, H. S. Dhillon, and J. G. Andrews, “Fitting determinantal point processes to macro base station deployments,” in Proc. IEEE Globecom 2014, Austin, TX, USA, Dec. 2014. [16] H. ElSawy, A. Sultan-Salem, M. S. Alouini, and M. Z. Win, “Modeling and analysis of cellular networks using stochastic geometry: A tutorial,” IEEE Commun. Surveys Tuts., vol. 19, no. 1, pp. 167–203, 2017. [17] I. Flint, X. Lu, N. Privault, D. Niyato, and P. Wang, “Performance analysis of ambient RF energy harvesting with repulsive point process modeling,” IEEE Trans. Commun., vol. 14, no. 10, pp. 5402–5416, Oct. 2015. [18] H. ElSawy and E. Hossain, “Two-tier HetNets with cognitive femtocells: Downlink performance modeling and analysis in a multichannel environment,” IEEE Trans. Mob. Comput., vol. 13, no. 3, pp. 649–663, Mar. 2014. [19] X. Zhou, Z. Zhao, R. Li, Y. Zhou, J. Palicot, and H. Zhang, “Human mobility patterns in cellular networks,” IEEE Commun. Lett., vol. 17, no. 10, pp. 1877–1880, Oct. 2013. [20] A.-L. Barabsi and R. Albert, “Emergence of scaling in random networks,” Science, vol. 286, no. 5439, pp. 509–512, Oct. 1999. [21] G. Samorodnitsky, Stable non-gaussian random processes: Stochastic models with infinite variance. New York: Chapman and Hall/CRC, Jun. 1994. [Online]. Available: http://www.amazon.com/Stable-Non-Gaussian-Random-Processes-Stochastic/dp/041205 [22] R. Li, Z. Zhao, J. Zheng, Y. Chen, C. Mei, Y. Cai, and H. Zhang, “The learning and prediction of application-level traffic data in cellular networks,” IEEE Trans. Wireless Commun., vol. 16, no. 6, pp. 3899 – 3912, Jun. 2017. [23] L. Chiaraviglio, F. Cuomo, M. Maisto, A. Gigli, J. Lorincz, Y. Zhou, Z. Zhao, C. Qi, and H. Zhang, “What is the best spatial distribution 13 [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35] [36] [37] [38] [39] [40] [41] [42] [43] [44] [45] [46] to model base station density? A deep dive into two European mobile networks,” IEEE Access, vol. 4, pp. 1434–1443, Apr. 2016. L. Chiaraviglio, F. Cuomo, A. Gigli, M. Maisto, Y. Zhou, Z. Zhao, and H. Zhang, “A reality check of base station spatial distribution in mobile networks,” in Proc. IEEE INFOCOM 2016, San Francisco, CA, USA, Apr. 2016. Z. Zhao, M. Li, and R. Li, “Temporal-spatial distribution nature of traffic and base stations in cellular networks,” IET Commun., Aug. 2017, accepted. R. Li, Z. Zhao, C. Qi, X. Zhou, Y. Zhou, and H. Zhang, “Understanding the traffic nature of mobile instantaneous messaging in cellular networks: A revisiting to α-stable models,” IEEE Access, vol. 3, pp. 1416–1422, Sep. 2015. C. Song, S. Havlin, and H. A. Makse, “Self-similarity of complex networks,” Nature, vol. 433, no. 7024, pp. 392–395, Jan. 2005. X. Ge, S. Tu, G. Mao, C. X. Wang, and T. Han, “5G ultra-dense cellular networks,” IEEE Wireless Communications, vol. 23, no. 1, pp. 72–79, Feb. 2016. A. Guo and M. Haenggi, “Spatial stochastic models and metrics for the structure of base stations in cellular networks,” IEEE Trans. Wireless Commun., vol. 12, no. 11, pp. 5800–5812, Nov. 2013. W. C. Cheung, T. Q. S. Quek, and M. Kountouris, “Throughput optimization, spectrum allocation, and access control in two-tier femtocell networks,” IEEE J. Sel. Area. Comm., vol. 30, no. 3, pp. 561–574, Apr. 2012. S. Mukherjee, “Distribution of downlink SINR in heterogeneous cellular networks,” IEEE J. Sel. Area. Comm., vol. 30, no. 3, pp. 575–585, Apr. 2012. B. Yu, L. Yang, H. Ishii, and S. Mukherjee, “Dynamic TDD support in macrocell-assisted small cell architecture,” IEEE J. Sel. Area. Comm., vol. 33, no. 6, pp. 1201–1213, Jun. 2015. M. D. Renzo and P. Guan, “Stochastic geometry modeling and systemlevel analysis of uplink heterogeneous cellular networks with multiantenna base stations,” IEEE Trans. Commun., vol. 64, no. 6, pp. 2453– 2476, Jun. 2016. F. J. Martin-Vega, G. Gomez, M. C. Aguayo-Torres, and M. D. Renzo, “Analytical modeling of interference aware power control for the uplink of heterogeneous cellular networks,” IEEE Trans. Wireless Commun., vol. 15, no. 10, pp. 6742–6757, Oct. 2016. T. D. Novlan, H. S. Dhillon, and J. G. Andrews, “Analytical modeling of uplink cellular networks,” IEEE Trans. Wireless Commun., vol. 12, no. 6, pp. 2669–2679, Jun. 2013. S. Singh, X. Zhang, and J. G. Andrews, “Joint rate and SINR coverage analysis for decoupled uplink-downlink biased cell associations in hetnets,” IEEE Trans. Wireless Commun., vol. 14, no. 10, pp. 5360– 5373, Oct. 2015. H. Y. Lee, Y. J. Sang, and K. S. Kim, “On the uplink SIR distributions in heterogeneous cellular networks,” IEEE Communications Letters, vol. 18, no. 12, pp. 2145–2148, Dec. 2014. Z. Zeinalpour-Yazdi and S. Jalali, “Outage analysis of uplink two-tier networks,” IEEE Trans. Commun., vol. 62, no. 9, pp. 3351–3362, Sep. 2014. M. Mirahsan, R. Schoenen, and H. Yanikomeroglu, “HetHetNets: Heterogeneous traffic distribution in heterogeneous wireless cellular networks,” IEEE J. Sel. Area. Comm., vol. 33, no. 10, pp. 2252–2265, Oct. 2015. Y. Lin, W. Bao, W. Yu, and B. Liang, “Optimizing user association and spectrum allocation in hetnets: A utility perspective,” IEEE J. Sel. Area. Comm., vol. 33, no. 6, pp. 1025–1039, Jun. 2015. H. S. Dhillon, R. K. Ganti, and J. G. Andrews, “Load-aware modeling and analysis of heterogeneous cellular networks,” IEEE Trans. Wireless Commun., vol. 12, no. 4, pp. 1666–1677, Apr. 2013. S. Singh, H. Dhillon, and J. G. Andrews, “Offloading in heterogeneous networks: Modeling, analysis, and design insights,” IEEE Trans. Wireless Commun., vol. 12, no. 5, pp. 2484–2497, May 2013. H. ElSawy, E. Hossain, and D. I. Kim, “HetNets with cognitive small cells: User offloading and distributed channel access techniques,” IEEE Commun. Mag., vol. 51, no. 6, pp. 28–36, Jun. 2013. Y. S. Soh, T. Q. S. Quek, M. Kountouris, and G. Caire, “Cognitive hybrid division duplex for two-tier femtocell networks,” IEEE Trans. Wireless Commun., vol. 12, no. 10, pp. 4852–4865, Oct. 2013. K. Huang and J. G. Andrews, “An analytical framework for multicell cooperation via stochastic geometry and large deviations,” IEEE Trans. Inf. Theory, vol. 59, no. 4, pp. 2501–2516, Apr. 2013. C. H. M. de Lima, M. Bennis, and M. Latva-aho, “Statistical analysis of self-organizing networks with biased cell association and interference [47] [48] [49] [50] [51] [52] [53] [54] [55] [56] [57] [58] [59] [60] [61] [62] [63] [64] [65] [66] [67] [68] [69] avoidance,” IEEE Trans. Veh. Tech., vol. 62, no. 5, pp. 1950–1961, Jun. 2013. P. Semasinghe, E. Hossain, and K. Zhu, “An evolutionary game for distributed resource allocation in self-organizing small cells,” IEEE Trans. Mob. Comput., vol. 14, no. 2, pp. 274–287, Feb. 2015. Y. Deng, L. Wang, M. Elkashlan, M. D. Renzo, and J. Yuan, “Modeling and analysis of wireless power transfer in heterogeneous cellular networks,” IEEE Trans. Commun., vol. 64, no. 12, pp. 5290–5303, Dec. 2016. C. Li, J. Zhang, and K. B. Letaief, “Throughput and energy efficiency analysis of small cell networks with multi-antenna base stations,” IEEE Trans. Wireless Commun., vol. 13, no. 5, pp. 2505–2517, May 2014. D. Cao, S. Zhou, and Z. Niu, “Optimal combination of base station densities for energy-efficient two-tier heterogeneous cellular networks,” IEEE Trans. Wireless Commun., vol. 12, no. 9, pp. 4350–4362, Sep. 2013. ——, “Improving the energy efficiency of two-tier heterogeneous cellular networks through partial spectrum reuse,” IEEE Trans. Wireless Commun., vol. 12, no. 8, pp. 4129–4141, Aug. 2013. Y. S. Soh, T. Q. Quek, M. Kountouris, and H. Shin, “Energy efficient heterogeneous cellular networks,” IEEE J. Sel. Area. Comm., vol. 31, no. 5, pp. 840–850, May 2013. M. D. Renzo and P. Guan, “A mathematical framework to the computation of the error probability of downlink MIMO cellular networks by using stochastic geometry,” IEEE Trans. Commun., vol. 62, no. 8, pp. 2860–2879, Aug. 2014. N. Lee, D. Morales-Jimenez, A. Lozano, and R. W. Heath, “Spectral efficiency of dynamic coordinated beamforming: A stochastic geometry approach,” IEEE Trans. Wireless Commun., vol. 14, no. 1, pp. 230–241, Jan. 2015. A. K. Gupta, H. S. Dhillon, S. Vishwanath, and J. G. Andrews, “Downlink multi-antenna heterogeneous cellular network with load balancing,” IEEE Trans. Commun., vol. 62, no. 11, pp. 4052–4067, Nov. 2014. L. H. Afify, H. ElSawy, T. Y. Al-Naffouri, and M. S. Alouini, “A unified stochastic geometry model for MIMO cellular networks with retransmissions,” IEEE Trans. Wireless Commun., vol. 15, no. 12, pp. 8595–8609, Dec. 2016. T. Bai and R. W. Heath, “Analyzing uplink SINR and rate in massive MIMO systems using stochastic geometry,” IEEE Trans. Commun., vol. 64, no. 11, pp. 4592–4606, Nov. 2016. S. Singh, M. N. Kulkarni, A. Ghosh, and J. G. Andrews, “Tractable model for rate in self-backhauled millimeter wave cellular networks,” IEEE J. Sel. Area. Comm., vol. 33, no. 10, pp. 2196–2211, Oct. 2015. F. Muhammad, Z. H. Abbas, and F. Y. Li, “Cell association with load balancing in nonuniform heterogeneous cellular networks: Coverage probability and rate analysis,” IEEE Trans. Veh.Tech., vol. 66, no. 6, pp. 5241–5255, Jun. 2017. I. Atzeni and M. Kountouris, “Full-duplex mimo small-cell networks: Performance analysis,” in Proc. IEEE Globecom 2015, San Diego, CA, USA, Dec. 2015. J. Lee and T. Q. S. Quek, “Hybrid full-/half-duplex system analysis in heterogeneous wireless networks,” IEEE Trans. Wireless Commun., vol. 14, no. 5, pp. 2883–2895, May 2015. H. Tabassum, A. H. Sakr, and E. Hossain, “Analysis of massive MIMOenabled downlink wireless backhauling for full-duplex small cells,” IEEE Trans. Commun., vol. 64, no. 6, pp. 2354–2369, Jun. 2016. K. Huang, J. G. Andrews, D. Guo, R. W. Heath, and R. A. Berry, “Spatial interference cancellation for multiantenna mobile ad hoc networks,” IEEE Trans. Inf. Theory, vol. 58, no. 3, pp. 1660–1676, Mar. 2012. R. Li, Y. Chen, G. Y. Li, and G. Liu, “Full-duplex cellular networks,” IEEE Commun. Mag., vol. 56, no. 4, pp. 184 – 191, Apr. 2017. A. AlAmmouri, H. ElSawy, O. Amin, and M. S. Alouini, “In-band αduplex scheme for cellular networks: A stochastic geometry approach,” IEEE Trans. Wireless Commun., vol. 15, no. 10, pp. 6797–6812, Oct. 2016. C.-H. Lee, C.-Y. Shih, and Y.-S. Chen, “Stochastic geometry based models for modeling cellular networks in urban areas,” Wireless Net., vol. 19, no. 6, pp. 1063–1072, Aug. 2013. N. Miyoshi and T. Shirai, “Spatial modeling and analysis of cellular networks using the Ginibre point process: A tutorial,” IEICE Transactions on Communications, vol. E99.B, no. 11, pp. 2247–2255, Nov. 2016. A. Karasaridis and D. Hatzinakos, “Network heavy traffic modeling using α-stable self-similar processes,” IEEE Trans. Commun., vol. 49, no. 7, pp. 1203–1214, Jul. 2001. M. Crovella and A. Bestavros, “Self-similarity in world wide web traffic: Evidence and possible causes,” IEEE/ACM Trans. Netw., vol. 5, no. 6, pp. 835–846, Dec. 1997. 14 [70] W. Song and W. Zhuang, “Resource reservation for self-similar data traffic in cellular/WLAN integrated mobile hotspots,” in Proc. IEEE ICC 2010, Cape Town, South Africa, May 2010. [71] C. Qi, Z. Zhao, R. Li, and H. Zhang, “Characterizing and modeling social mobile data traffic in cellular networks,” in Proc. IEEE VTC 2016Spring, Nanjing, China, May 2016. [72] X. Ge, Y. Qiu, J. Chen, M. Huang, H. Xu, J. Xu, W. Zhang, Y. Yang, C. X. Wang, and J. Thompson, “Wireless fractal cellular networks,” IEEE Wireless Commun., vol. 23, no. 5, pp. 110–119, Oct. 2016. [73] K. Park, G. Kim, and M. Crovella, “On the effect of traffic self-similarity on network performance,” in Proc. SPIE ICPCNS 1997, Dallas, TX, USA, Oct. 1997. 

7cs.IT

Ref: IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Workshop on Registration of Very Large Images, pp. 323—328, Columbus, OH, June 2014. Image Registration of Very Large Images via Genetic Programming Sarit Chicotay Dept. of Computer Science Bar-Ilan University Ramat-Gan, Israel Eli (Omid) David Dept. of Computer Science Bar-Ilan University Ramat-Gan, Israel Nathan S. Netanyahu Dept. of Computer Science Bar-Ilan University Ramat-Gan, Israel saritc@gmail.com mail@elidavid.com nathan@cs.biu.ac.il that it does not make any prior assumptions about the transformation model. Therefore, incorporating various building blocks into the GP framework could yield potentially a large pool of transformations. The advantage of this approach is that it offers, in principle, much greater flexibility in the registration of very large images which give rise to various, relatively complex transformation types. We present good results on some real datasets on small images and compare them to those obtained due to a recent method assuming a simple transformation model. We also present initial results on some real datasets containing larger images on simple transformations. From a pure GP perspective, other transformations, possibly more complex ones, could be searched for and discovered, as long as the GP is supported by the proper building blocks. The paper is organized as follows. Section 2 reviews IR components that are essential for our solution. Section 3 presents a brief GP background and provides the motivation for our GP-based IR approach. Section 4 presents an overview of related evolutionary-based solutions for IR. Section 5 provides a detailed description of our suggested GP-based approach. In Section 6 we present our initial empirical results. Section 7 makes concluding remarks. Abstract Image registration (IR) is a fundamental task in image processing for matching two or more images of the same scene taken at different times, from different viewpoints and/or by different sensors. Due to the enormous diversity of IR applications, automatic IR remains a challenging problem to this day. A wide range of techniques has been developed for various data types and problems. However, they might not handle effectively very large images, which give rise usually to more complex transformations, e.g., deformations and various other distortions. In this paper we present a genetic programming (GP)based approach for IR, which could offer a significant advantage in dealing with very large images, as it does not make any prior assumptions about the transformation model. Thus, by incorporating certain generic building blocks into the proposed GP framework, we hope to realize a large set of specialized transformations that should yield accurate registration of very large images. 1. Introduction Image registration (IR) is an important, significant component in many practical problems in diverse fields where multiple data sources are integrated/fused, in order to extract high-level information as to the contents of the given scene. A wide range of registration techniques has been developed over time, where typically, specific domain knowledge is taken into account and certain a priori assumptions are made, e.g., with respect to the transformation model used, specific bounds on its parameter values, etc. In contrast, in this paper we present a genetic programming (GP)-based approach for IR. GP is part of a family of evolutionary algorithms (EAs) which are stochastic optimization methods whose goal is to find an "optimal" solution or a set of solutions with respect to certain objective(s). Exploiting this strength of GP for search and optimization problems [8-9], we use GP and image processing techniques to search efficiently for an "optimal transformation" with respect to a given similarity measure. The novel aspect of our proposed GP algorithm is 2. Image Registration Image registration involves searching for a transformation that generates a maximal match in the overlap between the reference image and the transformed sensed image. It uses a similarity measure to assess the quality of a specific transformation. In this work, we focus on one of the most common measures, mutual information (MI) [14],[18], inspired by information theory [17]. MI is a measure of statistical dependency between two datasets that has been applied in a robust and efficient manner to IR. Analogous to the Kullback-Leibler measure [15] for the distance between two distributions, MI of two images measures the degree of dependence between the gray values in the area of overlap, defined as, ( , ) (1) ( , )= ( , ) ( ) ( ) , 1 The assumption is that maximal dependence between the gray values of the images achieved when they are correctly aligned, while misregistration results in a decrease of this measure. We considered close gray levels as having the same gray level for the purpose of calculating the MI. The set of transformations (e.g., translation, rotation, scaling) that are available during the registration process define the search space. The search for an optimal transformation in terms of the similarity measure is carried out using a search strategy which governs how the search space is explored. The search strategy has a great impact on the efficiency of the image registration process. In this paper we present an alternative GP-based search strategy. Since GP is known to be robust for a variety of search and optimization problems, it could also prove useful in the above scenario. 4. 3.1. Select a pair of chromosomes from the current population based on their fitness value (a chromosome can be selected more than once). 3.2. Produce offspring using crossover between the two selected chromosomes and perform random mutation on those offspring. Replace the population with the newly created one. 3. Genetic Programming As noted, GP is part of evolutionary algorithms family inspired by the evolution in nature. It uses the principles of evolution and natural selection for optimization and search for the global min/max of a given objective function. GP maintains a population of candidate solutions called chromosomes. The fundamental elements of an individual chromosome are its genes, which come together to form the tree-like program representing the chromosome consist of: (1) Functions, which in tree terminology, are nodes with children (the function's arguments) and (2) terminals which are leaves (i.e., nodes without branches). The functions can be basic mathematical functions (e.g., +, −,∗) and possibly more complex ones. Terminals consist of the variables and constants of the program. See Figure 1. Figure 2: Basic flow chart of GP The process of creating the offspring is called recombination and is combined of two steps: (1) Crossover, where the selected candidates are combined by swapping portions to produce hopefully better candidates with higher fitness for the next generation, and (2) mutation, which is performed on each of the new offspring with some low probability (mutation rate). Mutation generates variations in the population and helps avoid local minima. GP is designed to create and preserve a legal program structure during these steps. The process continues iteratively until it satisfies certain termination criteria, such as finding an acceptable approximate solution, reaching a specific number of generations, or until the solution converges. 4. Related Work Figure 1: Example of a GP chromosome and the program it represents Evolutionary algorithms (EAs) are stochastic optimization methods whose goal is to find an "optimal" solution or a set of solutions with respect to certain objective(s). EAs are considered robust methods, as one of their main advantages is to avoid local optima. These algorithms have been successful during the past decades in solving a variety of search and optimization problems including IR-related problems. Most of the earliest evolutionary approaches for IR [19] used genetic algorithms (GAs) with proportionate selection and a binary representation. This approach has several A basic GP algorithm consists of four iterative steps: (See Figure 2.) 1. Start with a randomly generated population of chromosomes. 2. Calculate the fitness value of each candidate in the population using a fitness function. 3. Repeat until offspring are created: 2 limitations when solving optimization problems in the continuous domain due to the binary representation that is limited for each decision variable and selection method that have several drawbacks when compared to ordinal-based selection methods such as ranking or tournament selection [5]. Another limitation of the early approach that it only deals with translation and rotation, ignoring scaling and shearing. Current EA studies for IR represent a GA-based approach [3],[11-12], which makes a prior assumption about the transformation model (e.g., similarity transformation). The optimal transformation is searched for in the space of transformation parameters. Santamaria et al. [10] present a more detailed overview of evolutionary methods for image registration and some state-of-the-art methods. In this work, we present a GP-based algorithm for IR. Unlike a GA approach, GP is not limited to a predefined number of transformation parameters. Therefore, incorporating various generic building blocks into the GP framework could yield potentially a diverse pool of transformation types, which should prove effective in the registration of very large images. A detailed description of this approach is given in the following section. 1.1.1 5. GP-Based Image Registration = , is a function of , and its three child nodes, defined as follows: ( , , ℎ )= (6) , ℎ , ℎ = {+, −,∗,/, , , , ( , ), ( , ), ( , ), ( , )} where , are the coordinates on which the transformation is performed. {+, −,∗,/, } – have arity of two, i.e., 2 child nodes. , , , - 1 child. { , } – 3 child nodes. , are functions of , and a child node; they describe a 2D rotation of an image around its center, i.e., ( , , ℎ )=( − (3) ℎ/2) cos( ℎ ) −( − ℎ ℎ /2)sin( ℎ ) + ℎ/2 ( , , ℎ ) = ( − ℎ ℎ /2) (4) cos( ℎ ) +( − ℎ/2)sin( ℎ ) + ℎ ℎ /2 = , is a function of and its three child nodes, defined as follows: ( , , ℎ )= (5) , ℎ , ℎ ( − ℎ This section describes our GP-based approach to optimize the IR similarity measures by utilizing tools common for IR. We set the initial population to be twice as large as a predefined size in order to maintain enough diversity. For the next generation, elitism is used to retain the top 2% chromosomes from the current generation and the rest are taken from the new pool of chromosomes. ( − ℎ For modeling an IR problem, the GP tree-like chromosome should represent a possible transformation. A transformation on a 2D image maps each pixel = ( , ) of the sensed image to a new pixel = ( , ) in the coordinate system of the referenced image, i,e., ( ) = ′. More specifically, we obtain: = ( , ), = ) +( − ℎ ) +( − ℎ , )+ ℎ )+ ℎ Although we have included some special functions as prospective building blocks of the chromosome trees, the GP algorithm may not necessarily use them "as-is". Despite this, we include these specific elements in order to assist in the convergence of the GP algorithm, should their explicit use become relevant. 5.1. Chromosome Representation (2) Function set 1.1.2 Terminal set = { , , } A is some random number selected uniformly { between ± ℎ, ℎ ℎ }, where ℎ, ℎ ℎ are the dimensions of the sensed image. A can be one of the coordinates, or . is the base of the natural logarithm ≈ 2.71828 ( , ) Therefore, each chromosome in our solution will be comprised of two such GP trees, an -tree and a -tree, each representing the transformation on one of the coordinates. We incorporate into our GP framework a list of functions and terminals that are essential for composing a diverse set of global transformations. Inspired by the discussion on transformation models in [1-2], we define, in particular, the following "primitives". 5.2. Crossover Operator Crossover is carried out separately on the -tree and tree of the two chromosomes. We used a simple one-point crossover, i.e., we select a crossover point in each of the two matching trees and then swap the corresponding subtrees. We also allow crossover to take place between an tree and a -tree with a low probability of ~0.2. This enhances the possibility of maintaining consistent 3 components by both transformations (e.g., rotation, scale). The resulting chromosomes are added to the new pool. generations. The chromosome with rank 1 at this stage is chosen as the suggested solution by the GP. 5.3. Mutation Operator 6. Empirical Results We have used the following two mutation variants: 1. Mutation of a randomly selected node in the tree, by replacing its sub-tree with a random generated sub-tree (with limited depth.) 2. Mutation of each node with some low probability by replacing it with a random node (function/terminal) with the same arity (i.e., same number of branches.) The chromosome created by recombination is mutated with probability 0.3, i.e., each of the above two variants is invoked with probability of 0.15. When the algorithm reaches a steady fitness over a predefined number of generations, the mutation rate is incremented by 0.02. If the fitness continues to improve the mutation rate is reset to its original value. 5.4. The following parameter values were used consistently by our GP-based IR algorithm for all evaluations: Population size Crossover probability Mutation rate Elitism Initial tree max. height Mutation max. height Table 1: Parameter values of the GP algorithm The GP approach was initially tested on the 256x256 images described in [13]. We used several sub-images as reference images ("chips") and registered to them overlapping sub-images sensed at different times. Several examples are shown in Figure 3. In order to assess the correctness of the final transformation, we selected manually points (typically about 10 points) in the reference image {( , )} ,…, and their corresponding points in the input image {( ′ , ′ )} ,…, , respectively. We consider these sets as the ground truth (i.e., most accurate transformation). Next we apply the transformation to each point in {( ′ , ′ )} ,…, , getting {( , )} ,…, , and compute the root mean square error (RMSE) defined as: Fitness Function The IR goal is to optimize the similarity measure(s) between the transformed sensed image and the referenced image. We used MI as mentioned in Section 2. In order for this function to be relevant for the registration of large images, we used the following sampling strategy. For each fitness evaluation, we select randomly a predefined number of pixels in the sensed image (typically 0.5% of the pixels.) The fitness calculation is performed on these pixels only. The overlap percentage between the images is also calculated only with respect to these randomly selected pixels, a chromosome associated with an overlap that is smaller than a predefined threshold is discarded. This strategy speeds up computation and has been proven useful in our initial tests. 5.5. (8) Selection Operator 5.6. ( − ∑ = ∑ ( − ) +( − ) We achieved an RMSE value of ~1 pixel on the examples shown in Figure 3, which is considered a successful registration. We compared also our results to another method tested on the same datasets [4]. The results obtained by both methods are comparable. Nevertheless, the quality of the results due to our GP based method should not be underestimated as no assumption was made about the transformation model in order to initiate or bound the registration process. Table 2 presents some of our results. Notice that the resulting transformations do not necessarily fit a familiar structure of a transformation combined of translation, rotation, and scaling. The GP scheme produces various transformations as part of evolving improved transformations, in an attempt to meet the optimal similarity measure. As with most IR applications, the objective is to align the images so that additional high-level operations could be performed. Thus, the characteristics of the transform are not that important, as We used rank-based selection which is considered more robust than other selection methods [7],[16]. The chromosomes are ordered by their fitness values due to the MI measure. Then, each chromosome is given a rank according to its position (the lower the better). Two chromosomes are selected for recombination based on their calculated ranking. The probability for an individual position to be selected within a population of size is given by: (7) 150 0.9 0.3 3/150 6 3 +1) Stopping Criteria As with most stochastic optimization algorithms, we have no clear way of knowing when to stop the search and accept the current best solution as the (near) optimal solution. We stop the algorithm when the solution is not "significantly" updated for a predefined number of 4 (a1) (b1) (c1) (d1) (e1) (a2) (b2) (c2) (d2) (e2) (a3) (b3) (c3) (d3) (e3) Figure 3: (a1), (b1), (c1), (d1), (e1): 256x256 reference images over the Washington, DC area, (a2), (b2), (c2), (d2), (e2): corresponding 256x256 sensed images, and (a3), (b3), (c3), (d3), (e3): registration results. long as the alignment is achieved with high level of accuracy. Motivated by the results on the set of 256x256 images, we have been exploring the validity of our approach on several, much larger space/airborne images. Our initial tests were performed on a Landsat 7 scene over Colorado (from NASA's IMAGESEER database). The scene is of size 5351x5051, (~27M pixels) and contains ~5% cloud coverage. We created two semi-synthetic image pairs. The first pair was created by applying only a translation with ( , ): − 50, the ground truth transformation: ( , ): − 100. We obtained, in this case, a successful registration with ~1 pixel RMSE. The second pair was created by applying both a rotation and translation with the SIFT-based IR [4] Image pair (a1)-(a2) s 1.001 [deg] -0.24 [pix] -5.76 [pix] -46.39 RMSE [pix] 0.74 (b1)-(b2) 1.006 -0.27 -4.75 -47.61 0.86 (c1)-(c2) 1.003 -0.47 -4.55 -48.73 0.63 (d1)-(d2) (e1)-(e2) 1.038 1.002 2.38 0.006 -6.36 -5.4 -46.18 -47.7 0.76 0.6 ground truth transformation: (angles are in radians): ( , ): (0.2618) + 15, ( , ): (0.2618) – 165. Figure 4 shows the registration result obtained with an RMSE of ~6 pixels. The transformation obtained by the GP scheme in this case (angles are in radians) is: ( , ): sin(3.714) + (12.24 + rotate ((((((−1953.79) + 1995.74) ∗ sin(−6.05)) − (−5048.61)))) ( , ): (( (cos(550.67 − (−420.77) + (−69.61)))/ 69.58) + ((( (((((9.78 + 550.67) + ((cos(−94.34) + (−98.93)) + (−98.93)))/(((−340.51)/1937.69) ∗ (−13.18) )))))/(9.78 + (((((−1953.79) + 1995.74) ∗ sin(−81.45) ) − (−5048.61))) ) + (( (cos(((((((2185.03/(−89.18)) − (−1055.97)) − 2301.13) + (cos(926.18) + (−98.93))) − (−98.93)))) ) + (−98.93)) + (−98.93)))) GP-Based IR ( , ) RMSE ( , ) [pix] ( - cos(358.26))-4.23 ( 2.7+( + sin(128.33)) -48.99) 0.85 ((cos(202.16)+cos(202.16)) +(sin(242.67)+(( +cos(183.68)) +sin(278.16)))) ((sin(310.19)*4.17)+ ((0.47-5.11)+ ) ((sin(319.21)+(sin(319.21) +cos(188.9)))+(sin(309.02)+ (( +sin(208.03))+sin(319.21)))) (((( -cos(8.74))-cos(66.27)) -((43.03-sin(209.9)) -sin(209.9)))+sin(254.8)) ((sin(2.32)+cos(267.27)) +((8.64+ ) +(59.4*sin(284.74)))) (cos(24.8)+( -45.05)) ((( -sin(39.8))-46.41) -cos(92.43)) 1.04 1.3 1.3 1.22 Table 2: Comparison of registration results due to our GP-based algorithm and a SIFT-based method [4] (angles are in degrees) 5 (a) (b) (c) (d) Figure 4: (a)-(b) Two Landsat 7 images to be aligned: (a) reference image (5351x5051 pixels), (b) semi-synthetic sensed image (6151x5151 pixels), (c) registration of sensed image to reference image, and (d) difference between reference image (a) and transformed sensed image (c) in overlap area. Tranactions on Pattern Analysis and Machine Intelligence, (15):850–863, 1993 7. Conclusions [7] D. Whitley. “The GENITOR algorithm and selection We presented a GP-based image registration algorithm, pressure: Why rank-based allocation of reproductive trials is best”. Proceedings of the Third International Conference on whose main novelty over existing IR techniques is that it Genetic Algorithms. 116-121, 1989. attempts to provide a robust and automatic solution without [8] J.R. Koza. "Genetic Programming: On The Programming assuming any transformation model. The initial of Computers by Means of Natural Selection". The MIT encouraging results demonstrate the potential of the Press, 1992. powerful, evolutionary GP approach, as far as its [9] J.R. Koza."Genetic Programming II". The MIT Press, 1994. application to IR. The accuracy of the results was assessed [10] J. Santamaría, O. Cordón, S. Damas. "A comparative study using RMSE, and was found comparable to another IR of state-of-the-art evolutionary image registration methods method presented recently. for 3D modeling". Computer Vision and Image The initial tests suggest that this approach can be Understanding, (115):1340–1354, 2010. [11] L. Zhang, W. Xu, and C. Chang. "Genetic algorithm for potentially useful when considering very large images that affine point pattern matching". Pattern Recognition Letters, give rise to more complex transformations. This would be (24):9–19, 2003. further investigated as part of future research. [12] M. Bazargani, A. Anjos , F. G. Lobo , A. Mollahosseini and We intend to test extensively this scheme on additional, H. R. Shahbazkia . "Affine Image Registration challenging datasets containing deformations and other Transformation Estimation Using a Real Coded Genetic distortions. Also, we will incorporate more of the common Algorithm with SBX". Proceedings of the fourteenth similarity metrics used in IR in addition to the MI measure international conference on Genetic and evolutionary used in this work (e.g., PHD, cross-correlation, sum of computation conference companion. 1459-1460, 2012. squared intensity) and consider additional functional [13] N. S. Netanyahu, J. Le Moigne, and J. G. Masek. "Georegistration of Landsat data via robust matching of building blocks for our GP-based scheme. multiresolution features", IEEE Transactions on Geoscience In view of our initial results, we believe there is indeed and Remote Sensing, (42):1586-1600, 2004. promise for such future directions. [14] P. Thevenaz and M. Unser. “Optimization of mutual information for multiresolution image registration”. IEEE References Transactions on Image Processing, (9), 2000. [15] S. Kullback and R.A Leibler. "On information and [1] A. Ardeshir Goshtasby."2-D and 3-D Image registration". sufficiency". The Annals of Mathematical Statistics, Wiley Pres, 2005. (22):79-86, 1951. [2] A. Ardeshir Goshtasby."Image Registration: Principles, [16] T. Back and F.Hoffmeister. “Extended selection Tools and Methods". Advances in Computer Vision and mechanisms in genetic algorithms”, Proceedings of the Pattern Recognition. Springer, 2012. Fourth International Conference on Genetic Algorithms, [3] A. Valsecchi and S. Damas. "An Image Registration 92-99, 1991. Approach using Genetic Algorithms". IEEE Congress on [17] T. M. Cover and J. A. Thomas. "Elements of Information Evolutionary Computation, 2012. Theory". John Wiley & Sons, Inc, 1991. [4] B. Kupfer, N. S. Netanyahu, and I. Shimshoni. "A SIFT[18] W. M. Wells III, P. Viola, H. Atsumi, S. Nakajima, and R. based image registration algorithm for remotely sensed Kikinis. “Multi-modal volume registration by maximization data", Proceedings of the IEEE International Geoscience of mutual information". Medical Image Analysis, (1):35–51, and Remote Sensing Symposium, 2013 1996. [5] D. E. Goldberg and K. Deb. "A comparative analysis of selection schemes used in genetic algorithms". Foundations [19] W. Peter and M. Tsang. "A Genetic algorithm for affine of Genetic Algorithms, (1):69–93, 1991. invariant object shape recognition". Proceedings of the First [6] D. P. Huttenlocher, G.A. Klanderman, and W. J. Rucklidge. IEEE International Symposium on Genetic Algorithms in “Comparing images using the Hausdorff distance” IEEE Engineering Systems. 293–298, 1995. 6 

9cs.NE

Subspace arrangements and motives of classifying stacks of reflection groups arXiv:1507.03486v2 [math.AG] 8 Sep 2017 Emanuele Delucchi, Ivan Martino September 12, 2017 Abstract We study some combinatorial aspects of a subspace arrangement associated to every faithful representation of finite groups, with a view on the triviality problem for motivic classes of classifying stacks. In particular, for finite reflection groups we reduce this problem to a conjecture which would generalize a theorem by Aluffi. In this paper we unfold the combinatorics underlying the computation of the motivic class {BG} of the classifying stack of a finite group G in the Grothendieck ring of algebraic stacks K0 (Stackk ). In the case of complex reflection groups, we reduce it to the following conjecture. Conjecture. Let W ⊂ GL(V ) be a finite complex reflection group and let AW be the arrangement of hyperplanes associated to W (see Definition 1.3). Call Ve the open complement of this arrangement and call Ve/W its quotient under the action of W . Then, the motivic class {Ve/W } is a polynomial in L, the class of the affine line. More precisely, it has the following form: {Ve/W } = χAW /W (L), where χAW /W (L) denote the characteristic polynomial of the quotient arrangement. This conjecture generalizes a theorem of Aluffi [1] asserting that {Ve } = χAW (L). Before we go any further, let us put this in context and explain our motivation. The Noether Problem. Let k be a any field and G be a finite group. A century ago, Emmy Noether [11] studied the rationality of the field extension k(V )G /k, where k(V )G denotes the invariants of the field of rational functions k(V ) over the regular representation V of G. This rationality question became known as the Noether problem and no counterexamples to the rationality were found for more than 50 years. In fact, the first such counterexample came in 1969 with work of Swan [16] Z proving that the extension Q(V ) /47Z /Q is not rational. In the Eighties, several other non rational extensions were found over the complex field by Saltman [13] and Bogomolov [2] for certain p-groups of order p9 and of order p6 , respectively. More recently Hoshi, Kang and Kunyavskii [7] studied the order p5 cases with non rational extension C(V )G /C. 1 The triviality of the motivic class of BG. In 2009, Ekedahl [4] showed that the Noether problem relates to the computation of a certain second degree cohomology of the motivic class {BG}. Each group of this cohomology is called an Ekedahl invariant of G in [10, 9], to which we refer for an overview of the topic. A motivic class is said to be trivial if it coincides with the class of a point {Spec(k)}, that we denote for simplicity by 1 because it is the neutral element with respect to the product in K0 (Stackk ). In all above-mentioned nonrational instances of the Noether Problem, {BG} is not trivial in K0 (Stackk ). A further point of interest in this triviality problem comes from recent work of Totaro [17] which relates it with other five properties of finite groups: stable rationality of quotient varieties V /G, triviality for the birational motive of the quotient varieties V /G, the weak Chow Künneth property of BG, the Chow Künneth property of BG and the mixed Tate property of BG. The known instances of triviality for {BG} are – the finite subgroups G ⊂ GL1 (k) (see [4, Proposition 3.2]); – the symmetric groups Sn (see [4, Theorem 4.3]); – all finite subgroups of the group of affine transformations of A1k , assuming k algebraically closed (see [4, Example ii) on page 8]). – the finite subgroups G ⊆ GL3 (k), if char(k) = 0 [9, Theorem 2.4]. Our results. The computation of the motivic class of {BG} relates to an orbit stratification of the quotient stack [V /G] for any faithful representation ρ of G in GL(V ). In this paper we focus on the combinatorics of such stratification. We say that a subgroup H of G is a stabilizer subgroup if there exists v ∈ V stabilized only by H, that is Stabρ (v) = H (see Definition 1.4). The set of stabilizer subgroups Pρ with reverse inclusion has a natural poset structure (see Definition 1.4). Moreover the group G acts on Pρ by conjugation and so we denote by P̃ρ := Pρ/G the set of conjugacy classes of stabilizer subgroups. It will be useful to denote by P̃ρ′ the poset of all all conjugacy classes of stabilizer subgroups. Under certain group theoretic and geometric conditions, we reduce the computation of {BG} to the knowledge of the combinatorics of P̃ρ′ . In order to state our first result, let us denote by ∆ρ the order complex of P̃(G) (this is the set of its ordered chains, that is the set of flags f as H0 < H1 < · · · < Hk ), and call NG (f ) the intersection of all normalizer subgroups NG (Hi ) with Hi stabilizer subgroup in the flag. Theorem A. Let ρ : G → GL(V ) be a faithful representation of a finite group G such that (1) {BNG (f )} = 1 for all f ∈ ∆ρ and (2) {V /NG (H)} = Ldim V H H for all nontrivial stabilizer subgroups H ∈ P̃ ′ (G). Then, {BG}ϕρ (L) = ϕρ (L), where ϕρ is the characteristic polynomial of ρ. (see Definition 1.6). 2 Notice that, ϕρ (t) is a polynomial defined only using the combinatorial information of P̃(G). As a token of motivation let us emphasize the following fact: under the hypothesis of Theorem A, if ϕρ (L) is invertible in K0 (Stackk ), then {BG} = 1. Finite reflection groups are a distinguished family of finite subgroups of the general linear group, which contains the symmetric groups and all cyclic groups. This fact already suggests finite reflection groups as sensible candidates for the study of {BG}. Moreover, the wealth of combinatorial, geometric and representation-theoretic results about finite reflection groups provides many useful tools. In particular, to every finite reflection group W is associated a reflection arrangement, i.e., the arrangement of hyperplanes defined by the fix-point sets of the reflections in W . Reflection arrangements lie at the roots of the general theory of hyperplane arrangements, for an introduction to which we point to the textbook [12]. For this class of groups, we are able to identify a combinatorial property of the poset of stabilizer subgroups (Definition 1.4) which allows us to explicitly compute the polynomial ϕρ (t) for the standard linear representation of finite reflection groups. Theorem B. If W is a d-dimensional irreducible finite reflection group in GL(V ) that fulfills the hypothesis of Theorem A, then ϕW (t) = td − 1. Using the fact that Ld − 1 is invertible in K0 (Stackk ), these computations allow us to pave the way for the proof of the triviality of the motivic class for the reflection groups. The missing part is precisely the proof of the conjecture, affirming that {Ve/W } = χAW /W (L). We stress that χAW /W (L) = χP̃(W ) (L). Theorem C. If the Conjecture holds, then {BW } = 1 in K0 (Stackk ). This type of rationality problems have also been studied in literature by different authors. For instance, Looijenga [8, Lemma 5.1] has shown that {V /G} = Ln for every abelian group acting linearly on a n-dimensional kvector space, where k is a field of characteristic zero containing a root of unity of degree the order of G. However, this condition fails for the counterexamples to the Noether problem given by Saltman [13], Bogomolov [2] and Hoshi, Kang and Kunyavskii [7]. Usm ing [4, Proposition 3.1], one shows that for m large enough {V /G} = {BG}Lmn in the Kontsevich value ring of algebraic varieties. Since {BG} is not trivial then m {V /G} is not Lmn in the K0 (VarC ). This rationality problem has also been studied by Esnault and Viehweg (see Lemma 1.1 and Proposition 1.3 in [6]). To conclude, we do not know whether the conjecture might be true for a class of group larger than that of the reflection groups. Indeed, the hypothesis of Theorem A seem very restrictive and may well characterize finite reflection groups. The structure of the paper is the following. First, in Section 1 we review some basics on subspace arrangements, partially ordered sets, Motivic classes 3 and reflection groups. We present the new Conjecture in Section 2 and then in Section 3 we prove Theorem A. In section 4, we study the poset of conjugacy classes of stabilizer subgroups for a complex reflection group and we show Theorem B. Finally in Section 4.1, we prove that the conjecture implies that the class of the classifying stack of W is trivial, that is Theorem C. Acknowledgements. We thank Emanuele Ventura for pointing out the reference [1]. We are very thankful to an anonymous referee who pointed out a mistake in a older version of this manuscript. Both authors have been partially supported by the Swiss National Science Foundation Professorship grant PP00P2 150552/1. The second author have been partially supported by the Zelevinsky Research Instructor Fund, by the Knut and Alice Wallenberg Fundation and by the Royal Swedish Academy of Science. 1 1.1 Preliminaries and generalities Combinatorics of subspace arrangements The main combinatorial structure of interest in this paper is that of a partially ordered set or, for short, poset. In this section we briefly review some basic terminology and the results we will need later. For more background, a standard reference is [15, Chapter 3]. A poset is a pair (P, ≤) consisting of a set P with a partial order relation ≤. In what follows the set P will always be finite and the order relation will be understood (so we will just say ”the poset P ” referring to (P, ≤)). If p, q ∈ P are such that p ≤ q, the interval determined by p and q is the subset [p, q] := {r ∈ P | p ≤ r ≤ q}. The set of all intervals of P is denoted I(P ). A lattice is a poset where every pair of elements has a unique minimal upper bound as well as a unique maximal lower bound. For example, the poset LG of all subgroups of a given finite group G, ordered by inclusion, is a lattice. The Möbius function of the poset P is the function µP : I(P ) → Z defined recursively as follows: X µ(p, p) = 1 for all p ∈ P ; µ(p, r) 0 for all p < q in P. = r∈[p,q] If the poset P has a unique minimal element (i.e., an element 0̂ with p ≤ 0̂ for all p ∈ P \ {0̂}), for every p ∈ P we will write µP (p) as a shorthand for µP (0̂, p). To every poset P is associated the poset ∆(P ) of all chains of P . Precisely, this is the set ∆(P ) := {{p1 , p2 , . . . , pk } ⊆ P | p1 < p2 < . . . < pk } ordered by inclusion. If the poset P has a unique minimal element 0̂ and a b ) := ∆(P \ {0̂, 1̂}). unique maximal element 1̂, we define also ∆(P Since every subset of a totally ordered set is again totally ordered, ∆(P ) and ˆ ) are naturally abstract simplicial complexes. Their topology is related to ∆(P the Möbius function as follows. 4 Theorem 1.1. For every p ≤ q in a poset P , µP (p, q) is the reduced Euler P b characteristic of ∆([p, q]), that is µP (p, q) = ω∈∆([p,q]) (−1)|ω|−1 . b We will be concerned with posets arising from subspace arrangements. Let k be a field, V a k-vector space of dimension d. An arrangement of subspaces in V is a finite set A of linear subspaces of V ; its poset of intersections is the set \ L(A) := { W | X ⊆ A} (1) W ∈X ordered by reverse inclusion: for x, y ∈ L(A), x ≤ y if x ⊇ y. These posets are always lattices, and in particular they have a unique minimal element 0̂ (corresponding T to the intersection of the empty family) and a unique maximal element 1̂ = W ∈A W . The arrangement is called essential if dim ∩A = 0. The characteristic polynomial of the arrangement A is X µ(x)tdim x . χA (t) = x∈L(A) 1.2 The subspace arrangement of a representation Fix a field k. Let G be a finite group and let ρ : G → GL(V ) be a krepresentation of G. For every g ∈ G and v ∈ V , we write the action as g · v = ρ(g)v for every g ∈ G and v ∈ V , where gv is the usual matrix-vector product. Following [4], for any v ∈ V and for any subgroup H ⊆ G we define the following objects. NG (H) Stabρ (v) VH VH := {g ∈ G | gH = Hg} ⊆ G := {g ∈ G | g · v = v} ⊆ G := {v ∈ V | H = Stabρ (v)} ⊆ V := {v ∈ V | H ⊆ Stabρ (v)} ⊆ V. In other words, NG (H) is the normalizer of the subgroup H in G; V H is the k-vector subspace of V of all points fixed (at least) by all elements of H; VH is the set of points that are fixed exactly by the elements of H. Definition 1.2. We say that a subgroup H is a stabilizer subgroup with respect to the representation ρ if VH is not empty. Definition 1.3. The arrangement of subspaces of the representation ρ is Aρ := {V H | H 6= {e} and H is a stabilizer subgroup with respect to ρ}. The set Aρ is ordered by reverse inclusion similar to L(A) in (1). It is useful to observe that V e = V and so the poset Aρ ∪ {V e } as a minimal unique element. Notice that every VH is the complement of the union of a set of subspaces inside V H. Definition 1.4. The poset of stabilizer subgroups of the representation ρ is Pρ := {H | H is a stabilizer subgroup with respect to ρ}, ordered by inclusion H1 ≤ H2 if H1 ⊆ H2 . 5 Then one has: Aρ ∪ {V e } ≃ L(Aρ ) ≃ Pρ The first isomorphism is proven essentially by definition and by observing that V e = V is the only missing subspace in Aρ . For the second isomorphism, consider the function Pρ → L(Aρ ) H 7→ V H . This is order-preserving and bijective. Indeed, if V H1 = V H2 for two stabilizer subgroups, then, by definition, they contain each other. Moreover, every intersection S in L(Aρ ) is of the form ∩H∈X V H , where X is a finite set of subgroups. ′ Thus, there exists a maximal H ′ containing all H in X, with S = V H . Since both posets are finite, then it is an isomorphism of posets. In particular, if ρ is a representation of a reflection group, then Pρ is a lattice. The group G acts on Pρ by conjugation. Definition 1.5. Define P̃ρ := Pρ/G, i.e., the set of all orbits under the G-action, ordered according to G·H1 ≤ G·H2 if H1 ≤ g · H2 in Pρ for some g ∈ G. It will be useful to denote by P̃ρ′ the poset of all all nontrivial stabilizer subgroups H. We close this section by recalling a definition due to Ekedahl [4]. Definition 1.6. Let Ñρ denote the poset of normal stabilizer subgroups of G ordered by reverse inclusion. The following is the characteristic polynomial of ρ: X X H (−1)dim f (tdim V − 1). ϕρ (t) = H∈Ñρ f ∈∆(P̃ρ ), max f =H Assumption. From now on we will assume that the representation ρ is faithful, i.e. Ker ρ = {e}. Moreover, notice that, given such a representation of G, we can always reduce to the case where dim VG = 0 and, when this happens, we say that this representations is essential (because Aρ is). If the representation is faithful and clear from the context, we write AG instead of Aρ and, similarly, for AG , PG , and P̃G . 1.3 Motivic classes of classifying stacks, Ekedahl’s formula The aim of this section is to briefly introduce the motivic classes of the classifying stack of a group and to present and to collect the necessary tools for this paper. As a general reference for a more comprehensive introduction we point to [10]. We begin by recalling that the Grothendieck ring of algebraic varieties K0 (Vark ) is the group generated by the isomorphism classes {X} of algebraic k-varieties X, under the relation {X} = {Y } + {X \ Y }, for all closed subvarieties Y of X. The ring structure of K0 (Vark ) is given by {X} · {Z} = {X × Z}. 6 Like algebraic varieties, algebraic stacks also have their own Grothendieck ring. In [5, Theorem 4.1] it is proved that K0 (Stackk ) = K0 (Vark )[L−1 , (Ln − 1)−1 , ∀n ∈ N], where the symbol L stands for the class of the affine line A1k . The classifying stack of the group G is usually defined as the stack quotient of the point Spec(k) by the group G, i.e. BG := [Spec(k)/G]. Its motivic class is denoted by {BG} ∈ K0 (Stackk ). Ekedahl in [4] proposed a combinatorial way to approach the computation of {BG}. Here we present his theorem in purely combinatorial terms, after a preparatory definition. Definition 1.7. Given a representation ρ : G → GL(V ) of a finite group G, we write b P̃ρ ) ∆ρ := ∆( and, for any f = {H1 ⊆ H2 ⊆ . . . ⊆ Hm } ∈ ∆ρ , dim f := |f | − 1, df := dim V Hm , and NG (f ) = ∩m i=1 NG (Hi ). Theorem 1.8 ([4, Theorem 3.4]). Let G be a finite group and let ρ : G → GL(V ) be a faithful and essential representation of dimension d. Then X {BG}(Ld − 1) = {Ve/G} + (−1)dim(f ) {BNG (f )}(Ldf − 1). f ∈∆ρ Remark 1.9. We point out that in the statement of Theorem 3.4 in [4] Ekedahl uses the term stabilizer flag instead of our chain for f . 1.4 Characteristic classes of subspace arrangements An arrangement of hyperplanes in kd is an arrangement of subspaces all whose elements have codimension 1. In this section we recall a result of Aluffi [1] linking the characteristic polynomial of a hyperplane arrangement to the characteristic class of its complement M (A). We reproduce its proof in order to show that it immediately generalizes to the case of subspace arrangements. Theorem 1.10 ([1, Theorem 2.1]). Let L be the class of the affine line in the Grothendieck ring of algebraic k-varieties, K0 (Vark ). Let A be an arrangement of hyperplanes in kn . Then, we have the following equality in K0 (Vark ): {M (A)} = χA (L). Proof. Consider the intersection lattice L(A) defined in (1 and pick an element x ∈ L(A). We set xo = x \ ∪y>x y, that is the complement in x of the union of all intersections inside x. Trivially {x} = Ldim x . Every P element y in L(A) is the disjoint union of all xo with Px > y, that is {y} = x≥y {xo }. Now, by Möbius inversion we have {y o } = x≥y µ(y, x){x}, and therefore we can compute X X µ(x)Ldim x = χA (L). µ(0̂, x){x} = {M (A)} = {0o } = x∈L(A) x≥0̂ 7 It is now apparent that Aluffi’s argument also proves the following stronger statement. Theorem 1.11. Let B denote a subspace arrangement in kd . Then, with notations as above, {M (B)} = χB (L) in K0 (Vark ).  2 Towards the conjecture We now focus on the case of finite groups which admit representations whose associated subspace arrangement is the set of intersections of an arrangement of hyperplanes. These are the so-called finite reflection groups. A detailed treatment of geometric and combinatorial aspects of hyperplane arrangements associated to reflection groups is [12, Chapter 6]. Here we only sketch some basics we’ll have use for. Let V be a k-vector space. Any g ∈ GL(V ) is called a reflection if it has finite order and it fixes a subspace of codimension 1 (which is then called the reflecting hyperplane of g). Following [12, Section 6.2], we call reflection group any finite group W ⊆ GL(V ) whose order is not divisible by char k and which is generated by reflections. The Chevalley-Shephard-Todd theorem [3, 14] shows that W is a finite reflection group exactly if and only if V /W is isomorphic to affine space. From now on when we talk about a finite reflection group W , we consider the obvious representation ρ given by the inclusion in GL(V ) and we write AW omitting reference to the representation. In particular, we are interested in the quotient arrangement in V ≃ V /W : AW /W := {X/W , X ∈ AW }. Even if V /W ≃ V , this is not anymore an arrangement of subspaces in V , but an arrangement of quotient varieties. We can now state a conjecture that aims at generalizing Theorem 1.11. Conjecture. If W is a finite complex reflection group, then, {Ve/W } = χAW /W (L). The statement of the conjecture would for instance follow from the existence of a faithful representation of the group W such that one of the two following conditions are satisfied for all nontrivial stabilizer subgroups H: A {V /NG (H)} = Ldim V H B V H/NG (H) ≃ Akdim V H H for all H ∈ P̃ ′ (G); for all H ∈ P̃ ′ (G). In what follows we are going to show that the proof of this conjecture will have immediate application in the study of the motivic class of the classifying stack of a finite group. It is worth to note that any of the above conditions will imply that the hypothesis of Theorem A are satisfied. 8 3 An inductive criterion In this section we are going to prove Theorem A. For this we need two technical results: the first one deals with the motivic class of the open complement Ve/G and the second one relates this result to the class of BG. Here and in what follows, µ denotes the Möbius function of P̃(G). Lemma 3.1. Let ρ : G → GL(V ) be a d-dimensional faithful representation of a finite group G. Then X H µ(H){V /NG (H)}. {Ve/G} = H∈P̃ρ Proof. We follow the same step of the proof in Theorem 1.10 and we have that X X H H µ(H){V /NG (H)}. µ({e}, H){V /NG (H)} = {Ve/G} = H≥{e} H∈P̃ρ Let us apply this to the computation of the class of BG. Proposition 3.2. Let ρ : G → GL(V ) be a d-dimensional faithful representation of a finite group G. Then X X X H H (−1)dim f {BNG (f )}(Ldim V −1). µ(H){V /NG (H)}+ {BG}(Ld −1) = H∈P̃ρ′ H∈P̃ρ f ∈∆ρ max f =H Proof. We use Theorem 1.8 and Proposition 3.1 and we obtain the following. X X H (−1)dim(f ) {BNG (f )}(Ldf − 1), µ(H){V /NG (H)} + {BG}(Ld − 1) = f ∈∆ρ H∈P̃ρ which readily implies the claim. The next theorem is an inductive triviality criterion. Together with the statement of the Conjecture presented in Section 2, it opens an avenue for inductive proofs of triviality of the motivic class of BG. We will explain more about this strategy in Section 4.1. Theorem A. Let ρ : G → GL(V ) be a d-dimensional faithful and essential representation of a finite group G such that (1) {BNG (f )} = 1 for all f ∈ ∆ρ and (2) {V /NG (H)} = Ldim V H H for all nontrivial stabilizer subgroups H ∈ P̃ ′ (G). Then, {BG}ϕρ (L) = ϕρ (L), where ϕρ is the characteristic polynomial of ρ (see Definition 1.6). 9 Proof. We start our computation by expanding the formula of Proposition 3.2: {BG}(Ld − 1) = Ld + µ(G) X H µ(H)Ldim V + H∈P̃ρ′ X + H∈P̃ρ′ X − H∈P̃ρ′ X f ∈∆ρ max f =H  X 
H (−1)dim f {BNG (f )}Ldim V
(−1)dim f {BNG (f )}. f ∈∆ρ max f =H Now isolate on the right hand side all the terms containing the motivic class of BG: {BG}(Ld − 1) = + Ld + µ(G) X H µ(H)Ldim V H∈P̃ρ′ +  X   {BG}Ldim V H   H∈P̃ ′ ρ  +  X     ′ H∈P̃ ρ − X f ∈∆ρ max f =H NG (f )6=G X {BG} X X H∈P̃ρ′ − H∈P̃ρ′ 
 (−1)dim f    X f ∈∆ρ max f =H NG (f )=G  
H (−1)dim f {BNG (f )}Ldim V    X (−1)dim f f ∈∆ρ max f =H NG (f )=G f ∈∆ρ max f =H NG (f )6=G 

(−1)dim f {BNG (f )}. Using assumption (1) we can continue the computation as follows. {BG}(Ld − 1) = + Ld + µ(G) X H µ(H)Ldim V H∈P̃ρ′ + X  dim V H {BG}L H∈P̃ρ′ 10      X f ∈∆ρ max f =H NG (f )=G  dim f (−1) 
    X +  X     f ∈∆ρ dim V H L H∈P̃ρ′ − X {BG} X X H∈P̃ρ′ Now we add and subtract X (−1)dim f (−1)dim f f ∈∆ρ max f =H NG (f )6=G X X X (−1) f ∈∆ρ max f =H NG (f )=G H∈P̃ρ′ max f =H, NG (f )=G and dim f max f =H NG (f )6=G H∈P̃ρ′ −   X H∈P̃ρ′ max f =H, NG (f )=G 
   


(−1)dim f ,
H (−1)dim f Ldim V . Remark also that, for all x 6∈ {{e}, G}, by Theorem 1.1. X X (−1)dim f +1 = −µ(x). (−1)dim f = f ∈∆ρ max f =x ′ f ∈P̃ρ,<x Using this identity, we can rewrite as follows. {BG}(Ld − 1) = + Ld + µ(G) X H µ(H)Ldim V H∈P̃ρ′ +  X ({BG} − 1)L X µ(H)Ldim V X ({BG} − 1) X µ(H). dim V H H∈P̃ρ′ − H∈P̃ρ′ + + X dim f (−1) f ∈∆ρ max f =H NG (f )=G 
    H X f ∈∆ρ max f =H NG (f )=G H∈P̃ρ′       (−1)dim f
H∈P̃ρ′ The defining identity of the Möbius function implies µ(G) + µ(e) = 0. Thereby, we simplify as follows: {BG}(Ld − 1) = Ld − 1 11 P H∈P̃ρ′ µ(H) + +  X ({BG} − 1)L X ({BG} − 1) dim V H H∈P̃ρ′ + X f ∈∆ρ max f =H NG (f )=G H∈P̃ρ′      X f ∈∆ρ max f =H NG (f )=G  dim f (−1) 
   
(−1)dim f . The claim now follows by regrouping terms and noticing that NG (f ) = G implies that all elements of f are normal subgroups. 4 The conjugacy classes stabilizer poset for W Let us now focus on finite reflection groups. The following is a useful fact in order to study P̃W . Proposition 4.1. If W is a finite reflection group in GL(V ), then N is a normal stabilizer subgroup if and only if W/N is a reflection group. Proof. Let N be a normal subgroup of W . The subspace V N fixed by N is an element of the lattice of intersections of AW , thus V N = ∩ni=1 Ri where Ri are the reflecting hyperplanes of the reflections ri ∈ N . Call t1 , . . . , ts the reflections that are not in N and Tj the reflecting hyperplane of tj . If N is a normal stabilizer subgroup, then V N is a faithful representation of W/N by Lemma 3.3 in [4]. Moreover, W/N is generated by τj = tj + N . Each τj fixes V N ∩ Tj , a subspace of codimension one in V N . For the reverse implication assume that W/N is generated by the reflections γ1 , . . . , γk but N is not a normal stabilizer subgroup. Call N ′ the stabilizer ′ ′ subgroup N ⊆ N ′ such that V N = V N . The G-orbit of V N consists of the H ′ ′ subspaces V with H in the same conjugacy class of N . If N is not a normal subgroup then this orbit consists of at least two distinct subspaces. But this ′ goes against the fact that V N = V N and the orbit of V N consists only of the N points of V because N is normal. Thus N ′ needs to be normal. Since N ′ is ′ also a stabilizer subgroup, then (by the proof of the first implication) V N is a faithful representation of the reflection group W/N ′ ≃ W/N/(N ′/N ). We know that V N = ∩ni=1 Ri because W/N is generated by reflections, and similarly deduce ′ that V N is the intersection of ∩ni=1 Ri with certain Tj s (at least one), and this is not possible. Therefore N ′ = N . So we can compute the poset of conjugacy classes for a complex reflection group. Proposition 4.2. If W is an irreducible finite reflection group in GL(V ), then ÑW = P̃W ∩ NW = {e, W }. Proof. In order to prove the statement, we observe that if N is a non trivial normal subgroup of W , then because of Proposition 4.1, the representation decomposes in the sum of two representations V = V N ⊕ (V N )⊥ . This is against the fact that W is an irreducible reflection group. 12 Hence, under the hypothesis of Theorem A, one can compute the main ingredient for the computation of BW . Theorem B. If W is a d-dimensional irreducible finite reflection group in GL(V ) that fulfills the hypothesis of Theorem A, then ϕW (t) = td − 1. Proof. The proof of the results comes directly by combining Proposition 4.2 and Theorem A. 4.1 The reduction to the conjecture We are ready to show that if the conjecture holds, the motivic class of the classifying stack of BW is trivial in the Grothendieck group of algebraic stacks. Theorem C. If W is a finite reflection group such that Conjecture holds, then {BW } = 1 in K0 (Stackk ). Proof. Every reflection group is product of irreducible ones and we observe that B(W1 × W2 ) = BW1 × BW2 and so {B(W1 × W2 )} = {BW1 }{BW2 }. Hence we assume without loss of generality that W is an irreducible complex reflection group. One then easily proves the statement for small (inclusion-minimal) reflection groups and proceeds by using the induction step as in the previous computations. Then {BW }pW (L) = pW (L) and so by Proposition 4.2 {BW }(Ld − 1) = (Ld − 1). The claim now readily follows since the polynomial Ld − 1 is clearly invertible in K0 (Stackk ) = K0 (Vark )[L−1 , (Ln − 1)−1 ]. References [1] P. Aluffi, Grothendieck classes and Chern classes of hyperplane arrangements, Int. Math. Res. Not. IMRN, (2013), pp. 1873–1900. [2] F. A. Bogomolov, The Brauer group of quotient spaces of linear representations, Izv. Akad. Nauk SSSR Ser. Mat., 51 (1987), pp. 485–516, 688. [3] C. Chevalley, Invariants of finite groups generated by reflections, Amer. J. Math., 77 (1955), pp. 778–782. [4] T. Ekedahl, A geometric invariant of a finite group. arXiv:0903.3148v1, 2009. [5] , The Grothendieck group of algebraic stacks. arXiv:0903.3143v2, 2009. [6] H. Esnault and E. Viehweg, On a rationality question in the Grothendieck ring of varieties, Acta Math. Vietnam., 35 (2010), pp. 31– 41. [7] A. Hoshi, M.-C. Kang, and B. E. Kunyavskii, Noether’s problem and unramified Brauer groups, Asian J. Math., 17 (2013), pp. 689–713. 13 [8] E. Looijenga, Motivic measures, Astérisque, (2002), pp. 267–297. Séminaire Bourbaki, Vol. 1999/2000. [9] I. Martino, The Ekedahl invariants for finite groups, J. Pure Appl. Algebra, 220 (2016), pp. 1294–1309. [10] , Introduction to the Ekedahl invariants, Math. Scand., 120 (2017), pp. 211–224. [11] E. Noether, Gleichungen mit vorgeschriebener Gruppe, Math. Ann., 78 (1917), pp. 221–229. [12] P. Orlik and H. Terao, Arrangements of hyperplanes, vol. 300 of Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences], Springer-Verlag, Berlin, 1992. [13] D. J. Saltman, Noether’s problem over an algebraically closed field, Invent. Math., 77 (1984), pp. 71–84. [14] G. C. Shephard and J. A. Todd, Finite unitary reflection groups, Canadian J. Math., 6 (1954), pp. 274–304. [15] R. P. Stanley, Enumerative combinatorics. Volume 1, vol. 49 of Cambridge Studies in Advanced Mathematics, Cambridge University Press, Cambridge, second ed., 2012. [16] R. G. Swan, Invariant rational functions and a problem of Steenrod, Invent. Math., 7 (1969), pp. 148–158. [17] B. Totaro, The motive of a classifying space. arXiv:1407.1366, 2014. Emanuele Delucchi Department of Mathematics, University of Fribourg, 1700 Fribourg, Switzerland. E-mail address: emanuele.delucchi@unifr.ch Ivan Martino Department of Mathematics, Northeastern University, Boston, MA 02115, USA. E-mail address: i.martino@northeastern.edu 14 

4math.GR

Gradient Flows in Uncertainty Propagation and Filtering of Linear Gaussian Systems arXiv:1704.00102v4 [math.OC] 29 Sep 2017 Abhishek Halder, and Tryphon T. Georgiou University of California, Irvine Abstract— The purpose of this work is mostly expository and aims to elucidate the Jordan-Kinderlehrer-Otto (JKO) scheme for uncertainty propagation, and a variant, the LaugesenMehta-Meyn-Raginsky (LMMR) scheme for filtering. We point out that these variational schemes can be understood as proximal operators in the space of density functions, realizing gradient flows. These schemes hold the promise of leading to efficient ways for solving the Fokker-Planck equation as well as the equations of non-linear filtering. Our aim in this paper is to develop in detail the underlying ideas in the setting of linear stochastic systems with Gaussian noise and recover known results. I. I NTRODUCTION n Consider the gradient flow dx dt = −∇ψ(x) in R , where ∇ is the gradient (w.r.t. the Euclidean metric) of a function ψ(x), and consider the discretization xk = xk−1 − h∇ψ(xk−1 ), for k ∈ N. was to show that the minimizer of (2) approximates the solution ρ(x, t) of the heat equation ∂ρ(x, t) = ∆ρ(x, t), with ρ(x, 0) = ρ0 (x), ∂t in the sense that ̺k (x, h) → ρ(x, t = kh), as h ↓ 0. Thus, (2) establishes the remarkable result that the heat equation is the gradient descent flow of the (negative) entropy integral with respect to the Wasserstein metric. An analogous JKO-like scheme was introduced recently in Laugesen et al. [9] for the measurement update-step in continuous-time filtering. More specifically, let us consider the general system of Itô stochastic differential equations (SDE’s) p (4a) dx(t) = −∇U (x) dt + 2β −1 dw(t), dz(t) = c(x(t), t) dt + dv(t), (4b) (1) where x ∈ Rn , z ∈ Rm , β > 0, U (·) is a potential, the process and measurement noise processes w(t) and v(t) are Wiener and satisfy E [dwi dwj ] = Qij dt ∀ i, j = 1, . . . , n and E [dvi dvj ] = Rij dt ∀ i, j = 1, . . . , m, with Q, R ≻ 0, respectively. Then x(t) and z(t) represent state and sensor measurements at time t. Further, as usual, v(t) is assumed to be independent of w(t) and independent of the initial state x(0). Given the history of noise corrupted sensor data up to time t, the filtering problem requires computing the posterior probability distribution that obeys the Kushner-Stratonovich stochastic PDE [19]–[21]. the solution, which depends on the choice of the step size h, satisfies xk (h) → x(t = kh), as h → 0. For the special case of trivial state dynamics, i.e., dx = 0, and R the identity, Laugesen et al. [9] introduced
− ̺+ k (x, h) = arg inf {DKL ̺k̺k + hΦ(̺)}, k ∈ N, (5) As is well known in finite-dimensional optimization, 1 xk = arg min{ kx − (xk−1 − h∇ψ(xk−1 ))k2 } 2 x 1 = arg min kx − xk−1 k2 + hψ(x) + o(h). 2 x By recursively evaluating the proximal operator [1], [2] xk k·k = proxhψ (xk−1 ) 1 = arg min{ kx − xk−1 k2 + hψ(x)}, 2 x The Jordan-Kinderlehrer-Otto (JKO) scheme, introduced in [3], is a similar recursion in the infinite-dimensional space of density functions with respect to the Wasserstein geometry [4], namely, ̺k (x, h) = arg min ̺ 1 2 W (̺, ̺k−1 ) + hS(̺), k ∈ N, (2) 2 2 where W2 (·, ·) denotes the Wasserstein-2 distance between two (probability) density functions, Z ̺(x) log(̺(x))dx (3) S(̺) := Rn is the negative differential entropy functional, and dx is the volume element. In other words, (2) can be viewed as the 2 proximal operation proxW hS (̺k−1 ). The main result in [3] ̺∈D2 with 1 E̺ {(yk − c(x))⊤ R−1 (yk − c(x))}, (6) 2 where yk is the noisy measurement in discrete-time defined via yk := h1 ∆zk , ∆zk := zk − zk−1 , and {zk−1 }k∈N the sequence of samples of z(t) at {tk−1 }k∈N for tk−1 := (k − 1)h. Laugesen et al. [9] proved that the LMMR equation (5) approximates the solution of h
⊤ dρ+ (x(t), t) = c(x(t), t) − Eρ+ {c(x(t), t)} R−1
 dz(t) − Eρ+ {c(x(t), t)}dt ρ+ (x(t), t), (7) Φ(̺) := i.e., of the Kushner-Stratonovich PDE corresponding to + dx = 0, in the sense that ̺+ k (x, h) ⇀ ρ (x(t), t) over t ∈ [(k − 1)h, kh), as h ↓ 0. Thus, they showed that in this special case, the Kushner-Stratonovich PDE is the gradient descent of functional Φ(·) with respect to DKL , i.e., − KL computed by proxD hΦ (̺k ). The purpose of the present paper is to develop this circle of ideas, namely, that both uncertainty propagation and filtering can be viewed as gradient flows in the special case of linear stochastic systems with Gaussian noise. In fact, we consider the general case of the linear stochastic system dx(t) = Ax(t) dt + B dw(t), (8) where w(t) is a Wiener process as before, though possibly not of the same dimension as x. We suppose that the uncertain initial condition x(0) has a known Gaussian PDF, the matrix A is Hurwitz, and that the diffusion matrix B is such that (A, B) is a controllable pair. For this, we recover the well-known propagation equations (see for example [6, Ch. 3.6]) for the mean and covariance of the state x(t) out of the JKO-scheme via a two-step optimization. The applicability of the JKO-scheme to (8) is not immediately obvious since the development in [3] requires the state dynamics to be in the canonical form (4a) with the drift being a gradient and the diffusion coefficient being a positive scalar. We further show that this two-step optimization procedure that we introduce, can be used to derive the Kalman-Bucy filter from a generalized version of the LMMR equation (5). We remark that variational schemes for estimator/observer design based on gradient flows can also be seen as regularized dynamic inversion in the spirit of [23]. Notation Throughout we use bold-faced upper-case letters for matrices, and bold-faced lower case letters for vectors. The notation I stands for identity matrix of appropriate dimension, we use tr(·) and det(·) to respectively denote the trace and determinant of a matrix, and the symbols ∇ and △ denote the gradient and Laplacian operators, respectively. We denote the space of probability R density functions (PDFs) on Rn by D := {ρ : ρ ≥ 0, Rn ρ = 1}, by D2 := {ρ ∈ R D | Rn x⊤ x ρ(x)dx < ∞} the space of PDFs with finite second moments, by Dµ,P denote the space of PDFs which share the R same mean vector µ and same covariance matrix P := Rn (x − µ)(x − µ)⊤ ρ(x)dx. Likewise, let Dµ,τ denote the space of PDFs which have the same mean µ and same trace of covariance τ := tr(P ) > 0. Clearly, Dµ,P ⊂ Dµ,τ ⊂ D2 ⊂ D. We use the symbol N (µ, P ) to denote a multivariate Gaussian PDF with mean µ, and covariance P . The notation x ∼ ρ means that the random vector x has PDF ρ; and E {·} denotes the expectation operator while, when the probability density is to be specified, Eρ {·} := R (·)ρ(x)dx. n R II. JKO S CHEME IN G ENERAL We now discuss in some detail the JKO scheme for the case of the diffusion process in (4a), and the corresponding Fokker-Planck equation [7] ∂ρ = ∇ · (∇U (x)ρ) + β −1 △ρ, ρ(x, 0) = ρ0 (x). (9) ∂t To this end we first introduce the Wasserstein metric, the free energy, and the Kullback-Leibler divergence. The Wasserstein-2 distance W2 (ρ1 , ρ2 ) between a pair of PDFs ρ1 (x), ρ2 (y) ∈ D (or, even between probability measures, in general), supported on X , Y ⊆ Rn , is   12 Z 2 (10) W2 (ρ1 , ρ2 ) := inf kx − yk2 dσ(x, y) dσ∈Π(ρ1 ,ρ2 ) X ×Y where Π (ρ1 , ρ2 ) is a probability measure on the product space X × Y having finite second moments and marginals ρ1 , ρ2 , respectively. It is well known that W2 : D × D 7→ [0, ∞) is a metric [4, p. 208]. Further, its square W22 (ρ1 , ρ2 ) represents the smallest amount of “work” needed to “morph” ρ1 into ρ2 [5]. The infimum is achieved over a space of measures, and under mild assumptions, the minimizing dσ has support on the graph of the optimal “transportation map” T : X 7→ Y that pushes ρ1 to ρ2 . Alternatively, one may view the optimization problem in (10) as seeking the joint distribution of two random vectors x and y, distributed according to  ρ1 and ρ2 respectively, that minimizes the variance E kx − yk22 . Another important notion of distance that enters into our discussion, which however is not a metric, is the KullbackLeibler divergence (also known as relative entropy) between PDFs or positive Rmeasures in general. This is given dρ1 dρ1 1 by DKL (dρ1 kdρ2 ) := ( dρ dρ2 ) log( dρ2 )dρ2 where dρ2 denotes the Radon-Nikodym derivative. When1 dρi = ρi (x)dx, i ∈ {1, 2}, are absolutely continuous with respect to the Lebesgue measure, then Z ρ1 (x) ρ1 (x) log DKL (dρ1 kdρ2 ) = dx. ρ 2 (x) Rn Gradient flow requires an energy functional, which we R denote by E (ρ) := U (x)ρ(x)dx, where U (·) is the potential energy. Then, a stochastically driven gradient flow is modeled by the Itô SDE (4a) and the Fokker-Planck equation (9) for the corresponding PDF as before. The stationary solution of (9) is the Gibbs distribution ρ∞R(x) = Z1 e−βU(x) , where the normalization constant Z := Rn e−βU(x) dx is known as the partition function. The distance to equilibrium which, in a way, quantifies the amount of work that the system can deliver, is captured by the so-called free energy functional F (ρ), defined as the sum of the energy functional E (ρ) and the negative differential entropy S (ρ) given in (3), that is, F (ρ) := E (ρ) + β −1 S (ρ)   = β −1 DKL ρke−βU(x) . (11a) (11b) 1 Here we use a slight abuse of notation in that we denote both, the measure and the density with the same symbol. For the case of (4a), the JKO scheme becomes parameterized subset of D2 in such a way that when we optimize the functionals 21 W 2 (̺, ̺0 ) and hF (̺) individually 1 ̺k (x, h) = arg inf { W22 (̺, ̺k−1 ) + h F (̺)}, k ∈ N, (12) over this chosen subspace, the arginfs (which are achieved) 2 ̺∈D2 of the two individual optimization problems match. Hence, for step-size h > 0, and initialized by a given ̺0 (satisfying the sum of the two has the same arginf over the chosen F (̺0 ) < ∞). For U (x) ≡ 0, (12) reduces to (2). Solving subspace. In the second step, we optimize over the subspace (12) results in a sequence of PDFs {̺k (x, h)}k∈N in D2 . parameters. Our choice for the parameterized set of densities It can be shown following [3] that ̺k (x, h) ⇀ ρ (x(t), t) is Dµ,P ⊂ D2 , i.e., the PDFs with given mean-covariance weakly in L1 (Rn ) for t ∈ [(k − 1)h, kh), k ∈ N, as h ↓ 0. pair (µ, P ); the choice of the optimal pair is to be decided in the second optimization step. III. JKO S CHEME FOR LINEAR G AUSSIAN SYSTEMS The development below requires several technical lemmas that are collected in the Appendix. We now develop and solve the JKO scheme for the linear Gaussian system in (8) with ρ0 = N (µ0 , P0 ) and Q ≡ I, without loss of generality. Therefore, we are concerned with the linear Fokker-Planck (Kolmogorov’s forward) PDE n
∂ρ ∂2 1 X = −∇ · (ρAx) + ρBQB ⊤ ij . (13) ∂t 2 i,j=1 ∂xi ∂xj Under the stated assumptions, it is well-known that (13) admits a steady-state, which is Gaussian with mean zero and covariance P∞ ≻ 0 that uniquely solves the algebraic Lyapunov equation AP∞ + P∞ A⊤ + BQB ⊤ = 0. Also, starting from ρ0 (x) = N (µ0 , P0 ), the transient is ρ(x(t), t) = N (µ(t), P (t)) where the µ(t) and P (t) satisfy the following ordinary differential equations (ODEs) [6, Ch. 3.6] µ̇(t) = Aµ(t), µ(0) = µ0 , (14a) Ṗ (t) = AP (t) + P (t)A⊤ + BQB ⊤ , P (0) = P0 . (14b) Below, we recover these equations using the JKO scheme. First, in Section III-A, wep explain how this is done when A is symmetric and B ≡ 2β −1 I, β > 0, in which case, Ax = −∇U (x) for a suitable potential. The general case, in Section III-B, is more involved and requires to view the drift as the gradient of a time-varying potential. p A. The case where A is symmetric and B ≡ 2β −1 I p 2β −1 I, the pair (A, B) is controllable. Since B ≡ Further, since A is Hurwitz and symmetric, Γ := −A ≻ 0, and utilizing the potential 1 ⊤ x Γx ≥ 0, 2 we can cast (8) in the canonical form (4a). Then,
1 ⊤ E(̺) := E[U (x)] = µ Γµ + tr (ΓP ) , 2 where P is the covariance of x. Notice that E(·) depends on the PDF of x only via its mean and covariance. U (x) := To carry out the optimization (12) over D2 , we adopt a two-step strategy. Our approach is motivated by the observation that the objective function in (12) is a sum of two functionals. In the first step, we choose a suitable 1) Optimizing over Dµ,P : Given ̺0 ≡ ρ0 = N (µ0 , P0 ), and a µ and P ≻ 0, we first determine 1 ̺1 = arg inf { W22 (̺, N (µ0 , P0 )) + h F (̺)}. (15) ̺∈Dµ,P 2 1 From Lemma 2 we see that arg inf W22 (̺, N (µ0 , P0 )) 2 ̺∈Dµ,P is achieved by ̺ = N (µ, P ) (uniquely). From Lemma 3, since U (x) = 21 x⊤ Γx, we also know that arg inf h F (̺) is ̺∈Dµ,P achieved by ̺ = N (µ, P ) (uniquely). Thus, ̺1 = N (µ, P ). The infimal value in (15) is now the sum of the two infima, "  1 !#  2 1 1 h 1 2 2 2 + k µ − µ0 k2 + tr P + P0 − 2 P0 P P0 2 2β h i −n − n log(2π) − log det(P ) + βµ⊤ Γµ + β tr (ΓP ) . (16) 2) Optimizing over (µ, P ): Equating the gradient of (16) w.r.t. µ to zero, results µ = φ(µ0 ) := (I + hΓ)−1 µ0 . The recursion µk = φ(µk−1 ), up to first order in h, becomes µk = (I − hΓ) µk−1 + O(h2 ). (17) We see that this recursion coincides with the solution of (14a) in the “small h” limit. Specifically, µ(t) = eAt µ0 ⇒
k µk := µ(t = kh) = eAh µ0 ⇒ µk = eAh µk−1 = (I + hA) µk−1 + O(h2 ), which is same as (17) since Γ := −A. Thus, we have recovered (14a) using discrete timestepping via JKO scheme in the small step-size limit. Setting the gradient of (16) w.r.t. P to zero (using Lemma 4), we obtain  1 1 1 h −1 −1 2 P02 − P −1 + hΓ = 0. (18) I − P02 P0 2 P −1 P0 2 β − 12 By pre and post multiplying both sides of (18) with P0 1  −1 2 −1 =: Z, we arrive at and letting P0 2 P −1 P0 2 , β β −1 −1 Z − P0 2 (I + hΓ)P0 2 = 0, h h which admits the unique closed-form solution [16, p. 304]   21 ! h − 12 β − 12 −I + I + 4 P0 (I + hΓ) P0 . (19) Z= 2h β Z2 + Expanding (19), we obtain   β 1 h −I + I + 4 P0 (I + hΓ) P0 + 2h 2 β
# 1 1 2 1 − 1 16h −1 − −1 3 2 2 2 2 (I + hΓ) P0 (I + hΓ) P0 + O(h ) P 2! β2 0   h −1 −1 = P0 2 I + hΓ − P0−1 P0 2 + O(h2 ). (20) β Z= −1 2 −1 2 1  −1 2 −1 back into (20), squarSubstituting Z = P0 2 P −1 P0 2 ing, and rearranging, we get that   −1  −1  1 1 P = I + h Γ − P0−1 P0 I + h Γ − P0−1 +O(h2 ) β β       1 1 P0 I − h Γ − P0−1 + O(h2 ) = I − h Γ − P0−1 β β = Ψ(P0 ) + O(h2 ),
where Ψ(P0 ) := P0 + h −ΓP0 − P0 Γ + 2β −1 I . Set the matrix-valued recursion Pk = Ψ (Pk−1 ), where
Ψ (Pk−1 ) := Pk−1 + h −ΓP0 − P0 Γ + 2β −1 I + O(h2 ). (21) To show that (21) indeed recovers (14b), p first notice that substituting A = A⊤ = −Γ and B = 2β −1 I in (14b) results the Lyapunov differential equation Ṗ (t) = −ΓP (t) − P (t)Γ + 2β −1 I subject to P (0) = P0 , which can be solved via the method of integrating factor as
1 (22) P (t) = Γ−1 I − e−2Γt + e−Γt P0 e−Γt . β Thus, for t = kh, (22) gives
Pk := P (kh) = β −1 Γ−1 I − e−2Γkh + e−Γkh P0 e−Γkh 1) Equipartition of energy coordinate transformation: Consider the stationary covariance P∞ associated with √ (A, 2B) that satisfies AP∞ + P∞ A⊤ + 2BB ⊤ = 0. For a system at a stationary distribution, we define the thermodynamic temperature θ as the average amount of “energy” per degree of freedom, that is, 1 θ := tr(P∞ ), n and, thereby, β := θ−1 the inverse temperature. By pre and −1 post multiplying (23) with P∞ 2 , and rescaling by θ so as to preserve the temperature in the new coordinates, we get √ √ Aep θI + θIA⊤ 2θBep ( 2θBep )⊤ = 0, (24) ep + Pk − Pk−1 = 2β −1 hI − hΓP0 − hP0 Γ + O(h2 ), which is same as (21) derived from JKO scheme. Thus, we have recovered the covariance evolution through FokkerPlanck dynamics using the time-stepping procedure via JKO scheme in the small step-size limit. −1 1 −1 2 , Bep := P∞ 2 B, while where Aep := P∞ 2 AP∞ the stationary covariance θI reflects equipartition of energy.√The equipartition√of energy co-ordinate transformation (A, 2B) 7→ (Aep , 2θBep ), corresponds to the state√ −1 transformation x 7→ xep := θP∞ 2 x, leading to √ dxep (t) = Aep xep (t) dt + 2θBep dw(t). (25) This settles how β is to be defined and interpreted in the context of JKO scheme (11) and (12). On the other hand, Aep being similar to A, is guaranteed to be Hurwitz but not symmetric, unless A was symmetric to begin with. Thus, it remains for us to “symmetrize” Aep and define a suitable potential energy U (·) as needed in (12). We do this next. 2) Symmetrization transformation: We introduce the time-varying transformation = 2β −1 khI + (P0 − khΓP0 − khP0 Γ) + O(h2 ). Replacing k with k−1 in the latter yields a similar expression for Pk−1 . Then, subtracting these expressions for Pk−1 from Pk we obtain that skew where Askew ep xep 7→ xsym := e−Aep := 1 2 (Aep skew t t xep A⊤ ep ). − This results in √ √ (Aep , 2θBep ) 7→ (F (t), 2θG(t)), with F (t) := e−Aep skew Aep Asym ep e where, similarly, Asym := ep satisfies t skew , and G(t) := e−Aep 1 2 (Aep dxsym (t) = F (t)xsym (t) dt + B. The case of Hurwitz A and controllable (A, B) √ We scale B into 2B without loss of generality, and take as initial PDF ̺0 ≡ ρ0 = N (µ0 , P0 ). Since we allow any Hurwitz√(not necessarily symmetric) A, and any B that makes (A, 2B) a controllable pair, it is not apparent if and how one can express (8) in the canonical form (4a). The main impediment in doing so, is twofold: (1) how to define the potential energy U (x), and (2) how to interpret and define the parameter β in the generic case. In the following, we show that by two successive co-ordinate transformations, system (8) can indeed be put in the form (4a). Similar transformations have been mentioned in [17, p. 1464], [18] in a different context. (23) t Bep , + A⊤ ep ). Thus, xsym (t) √ 2θG(t) dw(t). (26) Notice that F (t) is symmetric for all t. Furthermore, observe that the new coordinates xsym is simply obtained by a (time-varying) orthogonal transformation of the equipartition of energy coordinates xep . Hence the stationary covariance of xsym is identical to that of xep , which is θI (from Section III-B.1). What happens is that the covariance of xsym (t) tends to the same steady state value as t → ∞ in spite of the fact that (26) has time varying coefficients. To see this ⊤ in different way, we can rewrite (24) as Bep Bep = −Asym ep , and deduce that skew G(t)G(t)⊤ = e−Aep skew ⊤ Aep t Bep Bep e = −F (t), √ √ ⇒ F (t)θI + θIF (t) + 2θG(t)( 2θG(t))⊤ = 0. t (27) The symmetrization xep 7→ xsym leaves the stationary covariance θI invariant. This guarantees that the definition of temperature θ stays intact. The coordinate transformations described above are summarized in Table I. 3) Recovery of the Fokker-Planck solution: We are now ready to apply the JKO scheme to√ the generic stochastic linear system dx(t) = Ax(t) dt + 2B dw(t), with initial PDF ρ(x(0), 0) = N (µ0 , P0 ). To this end, we carry out a computation akin to the two steps in Section III-A, for the transformed SDE (26) in the symmetrized coordinate xsym . From there on, we recover the Fokker-Planck solution in the original coordinate x. Since x 7→ xsym is a linear transformation, it follows that xsym ∼ N (µsym , Psym ) whenever x ∼ N (µ, P ). Thus, carrying out the first step of the optimization in xsym coordinate, we get an expression similar to (16) wherein (µ, P ) is to be replaced by (µsym , Psym ), and (µ0 , P0 ) is to be replaced by (µsym0 , Psym0 ). To carry out the second ⊤ step of optimization, notice that Asym = −Bep Bep  0, ep skew skew −Aep t sym Aep t and consequently F (t) = e Aep e  0. Thus, considering the time-varying potential 1 U (xsym ) := − x⊤ F (t)xsym ≥ 0, 2 sym and setting the partial derivative of the infimal value from first stage of the optimization w.r.t. µsym to zero, results the recursion µsymk = (I − hF (kh))−1 µsymk−1 . Recalling skew √ −1 that xsym = e−Aep t θP∞ 2 x, we arrive at a recursion in original coordinate: 1 skew 2 Aep µk = P∞ e kh skew {(I − hF (kh))−1 eAep e −Askew kh ep −1 P∞ 2 h } µk−1 . (28) By series expansion and collecting linear terms in h, one can verify the following: 2 (I − hF (kh))−1 = I + hAsym ep + O(h ), −1 Askew h ep (I − hF (kh)) e Askew kh ep e where C ∈ Rm×n , and ρ0 = N (µ0 , P0 ). The conditional PDF ρ+ (x(t), t) = N (µ+ (t), P + (t)), given measurements up to time t, is well-known and given by the Kalman-Bucy filter [22]
dµ+ (t) = Aµ+ (t)dt + K(t) dz(t) − Cµ+ (t)dt , (31a) Ṗ +(t) = AP +(t)+P +(t)A⊤+2BB ⊤−K(t)RK(t)⊤ (31b) that specifies a vector SDE and a matrix ODE, respectively, for the conditional mean µ+ (t) and covariance P + (t). The initial conditions are µ+ (0) = µ0 , P + (0) = P0 , and K(t) := P + (t)C ⊤ R−1 is the so-called Kalman gain. In the sequel, we demonstrate that by applying the twostep optimization strategy we used before in Section III, we can recover the Kalman-Bucy filter from LMMR-equation (5) for the linear Gaussian case as the h ↓ 0 limit. A. LMMR gradient descent scheme Once again we proceed with carrying out the following two optimization steps. First, we optimize (5) over Dµ,P , and then optimize the minimum value over the choice of parameters (µ, P ). − − 1) Optimizing over Dµ,P : Consider ̺− k = N (µk , Pk ) to be our prior for the state PDF at time t = kh. Observe that inf ̺∈Dµ,P −
DKL ̺kN (µ− k , Pk ) = inf ̺∈Dµ,P −E̺ {log e F ILTERING dz(t) = Cx(t) dt + dv(t), = I + hAep + O(h ), e FOR In this section, we focus on the linear Gaussian filtering problem, with process model and measurement models √ dx(t) = Ax(t)dt + 2Bdw(t), 2 −1 Askew h −Askew kh ep ep (I − hF (kh)) IV. JKO- LIKE S CHEMES Z ̺(x) log ̺(x)dx Rn  − N (µ− k , Pk )} , (32) and that 2 = I + hAep + O(h ).
−1 1h ⊤ (µ − µ− Pk− (µ − µ− k) k ) 2
 1
+ tr P (Pk− )−1 − log (2π)n det(Pk− ) 2 − E̺ {log N (µ− k , Pk )} = − Hence (28) yields   1 −1 2 Aep P∞ 2 µk−1 + O(h2 ) µk = I + hP∞ = (I + hA) µk−1 + O(h2 ), − 21 (29) 1 2 where the last equality follows from Aep := P∞ AP∞ . hA 2 Since µ̇ = Aµ and e = I + hA + O(h ), in the small h limit, equation (29) thus recovers (14a), as in Section III-A. A similar straightforward but tedious computation leads to the matrix recursion Pk − Pk−1 = h(APk−1 + Pk−1 A⊤ + 2BB ⊤ ) + O(h2 ), (30) which in the limit h ↓ 0, is indeed a first-order approximation of the Lyapunov equation for the original system. We omit the details for brevity. remains invariant for all ̺ ∈ Dµ,P . Therefore, the arginf in (32) is achieved by the Gaussian PDF N (µ, P ) (i.e., the maximum entropy PDF with given mean-covariance), and − the infimal value is precisely DKL (N (µ, P )kN (µ− k , Pk )). On the other hand, notice that 1h 1 (yk − Cµ)⊤ E̺ {(yk − Cx)⊤ R−1 (yk − Cx)} = 2 2  i R−1 (yk − Cµ) + tr C ⊤ R−1 CP = constant (33) inf ̺∈Dµ,P ❤❤❤ → ❤❤❤Coordinate ❤❤❤ Attribute ↓ ❤❤ Original Equipartition of energy Symmetrization State vector x xep xsym (A, System matrices Stationary covariance √ 2B) P∞ (Aep , √ 2θBep ) θI (F (t), √ 2θG(t)) θI TABLE I: Summary of the coordinate transformations for Section III-B. as well over Dµ,P . Hence
 − arg inf DKL ̺kN (µ− k , Pk ) B. Alternative JKO-like schemes for filtering ̺∈Dµ,P +  h E̺ {(yk − Cx)⊤ R−1 (yk − Cx)} = N (µ, P ), 2 and the corresponding infimum value is
1 − −1 ⊤ tr (Pk− )−1 P + (µ− (µ− k − µ) (Pk ) k − µ) − n− 2
 h (yk − Cµ)⊤ R−1 (yk − Cµ) log det (Pk− )−1 P + 2
 + tr C ⊤ R−1 CP . (34) 2) Optimizing over (µ, P ): Equating the partial derivative of (34) w.r.t. µ to zero, and setting µ ≡ µ+ k in the resulting algebraic equation, we get

+ ⊤ −1 (Pk− )−1 µ− yk − Cµ+ k − µk + hC R k = 0,
− − ⊤ −1 (35) yk − Cµ+ ⇒µ+ k . k = µk + hPk C R On the other hand, equating the partial derivative of (34) w.r.t. P to zero, and then setting P ≡ Pk+ in the resulting algebraic equation, we get  (Pk+ )−1 = (Pk− )−1 + hC ⊤ R−1 C ⇒ Pk+ = I + hPk− C ⊤
−1 − R−1 C Pk = Pk− − hPk− C ⊤ R−1 CPk− + O(h2 ). (36) With ∆zk = yk h, as in Section I, 2 µ+ (t)dt = µ+ k h + O(h ), and from (29) that + 2 µ− k = (I + hA)µk−1 + O(h ). These, together with (36), allow us to simplify (35) as
2 + + + ⊤ −1 µ+ ∆zk − hCµ+ k −µk−1 = hAµk−1 + Pk C R k + O(h ), which in the limit h ↓ 0, leads to (31a). Substituting (30) into (36) we arrive at + + −hPk−1 C ⊤ R−1 CPk−1 + O(h2 ). In the limit h ↓ 0, (37) recovers (31b). ̺+ k (x, h) = arg inf ̺∈D2
1 2 W2 ̺, ̺− + hΦ(̺), k 2 k ∈ N, (38) where the functional Φ(·) is as in (6). The template of the two-step optimization again applies and, specializing to the linear Gaussian case, the solution of (38) in the h ↓ 0 limit, is N (µ+ (t), P + (t)), given by
dµ+ (t) = Aµ+ (t)dt + L dz(t) − Cµ+ (t)dt , (39a) Ṗ + (t)=(A − LC)P + (t)+P + (t)(A − LC)⊤+2BB ⊤(39b) where L := C ⊤ R−1 , and µ+ (0) = µ0 , P + (0) = P0 . This follows by noticing from Sections III-A.1 and IV-A.1 that  
1 2 − − W ̺, N (µk , Pk + hΦ(̺) = N (µ, P ), arg inf 2 2 ̺∈Dµ,P where the infimum value is    1   1 − 21 2 − − 12 2 P (P ) k µ−µ− k + tr P + P −2 (P ) k k 2 k k 2
 h + (yk − Cµ)⊤ R−1 (yk − Cµ)+tr C ⊤ R−1 CP .(40) 2 dz(t) = ∆zk + O(h2 ), + + + Pk+ − Pk−1 = h(APk−1 + Pk−1 A⊤ + 2BB ⊤ ) The ideas in the LMMR-scheme suggest the possibility of alternative variational schemes to approximate stochastic estimators. Such a viewpoint has been put forth in [23], promoting the notion of regularized dynamic inversion. As an example, one may consider a gradient descent with respect to the Wasserstein distance 12 W22 , instead of KL-divergence DKL in (5). In that case, the posterior may be constructed according to (37) Equating the partial derivative of (40) w.r.t. µ to zero, then + + setting µ ≡ µ+ k , and using (29), we find (µk −µk−1 ) equals
⊤ −1 2 ∆zk − hCµ+ hAµ+ k + O(h ), k−1 + C R (41) which in the limit h ↓ 0, results the SDE (39a). Similarly, using Lemma 4, we equate the partial derivative of (40) w.r.t. P to zero, and then setting P ≡ Pk+ , we get

(Pk+ )−1 = I + hC ⊤ R−1 C (Pk− )−1 I + hC ⊤ R−1 C ⇒
Pk+ = Pk− − h Pk− C ⊤ R−1 C + C ⊤ R−1 CPk− + O(h2 ), + + which combined with the recursion Pk− = Pk−1 +h(APk−1 + ⊤ ⊤ 2 + Pk−1 A + 2BB ) + O(h ) from Section III-B.3, yields 
+ + + Pk+ = Pk−1 +h A − C ⊤ R−1 C Pk−1 + Pk−1 (A− i
⊤ 2 ⊤ −1 ⊤ C R C + 2BB + O(h ). (42) In the limit h ↓ 0, recursion (42) gives Lyapunov ODE (39b). It is instructive to compare the SDE-ODE system (39) with that in (31). In the case of (39), the estimator is of a Luenberger type with a static gain matrix L which is decoupled from the covariance, unlike (31). The estimator (39) is obviously not optimal in the minimum mean-square error sense. It is only presented here as a guideline to explore other variational schemes with desirable properties. V. C ONCLUDING REMARKS Reformulating uncertainty propagation and the filtering equations as gradient flows [8] is potentially transformative [3] [9]. The full power of this viewpoint is yet to be uncovered. Moreover, casting the iterative approximation steps in the language of proximal operators on the space of density functions may provide theoretical insights and computational benefits. A specific direction of future work would be developing proximal algorithms [2] to numerically solve the nonlinear filtering problem by recursively solving convex optimization problems, and to quantify computational performance of the same with respect to existing sequential Monte Carlo algorithms like the particle filter. The purpose of the present paper has been to highlight and elucidate the ideas in [3] and [9] in the context of linear Gaussian systems. We hope that this study will help to motivate further exploration of this topic. A PPENDIX In this Appendix, we collect some lemmas that are used in Sections III and IV. In addition, we will show in Corollary 1 below that applying Lemma 1 and 2 together enables us to provide an alternative proof of a Theorem in [14], which might be of independent interest. Lemma 1: If X and Y are symmetric positive definite  1  21 p 1 matrices, then tr X 2 Y X 2 ≤ tr (X) tr (Y ). Proof: From Uhlmann’s variational formula (see [10], also Theorem 6.1 in [11]) , given any G ≻ 0, we have   21  p 1 1 ≤ tr (XG) tr (Y G−1 ), (43) tr X2Y X2 where the equality in (43) is achieved for the specific choice  1 − 21 1 1 1 1 1 Gopt = Y 2 X 2 Y X 2 X 2 Y 2 X − 2 . Specializing (43) for G = Y , and noting that tr (XY ) ≤ tr (X) tr (Y ), the statement follows. Lemma 2: Given a PDF ̺0 (x) ∈ D2 with mean µ0 ∈ Rn , and n × n covariance matrix P0 ≻ 0. Then inf W22 (̺, ̺0 ) equals ̺∈Dµ,P kµ− µ0 k22   1  21  1 2 2 , + tr P + P0 − 2 P0 P P0 (44) and is achieved by push-forward of ̺0 (x) via an affine trans 1 − 12 1 1 1 port map M x + m, where M := P 2 P 2 P0 P 2 P 2, and m := µ− µ0 , that is, the arginf for (44) is ̺(x)  = q  1  12 1 1 1 det(P0 ) −2 P 2 P0 P 2 P − 2 (x − µ) + µ0 . In det(P ) ̺0 P particular, if ̺0 = N (µ0 , P0 ), then ̺ = N (µ, P ). Proof: Let ̺0 be as given, and choose any ̺ ∈ Dµ,P . Let ̺0 and ̺ be obtained by translating ̺0 and ̺ respectively, such that both ̺0 and ̺ have zero mean. Using (10), we can directly verify [12, p. 236] that W22 (̺, ̺0 ) =k µ − µ0 k22 +W22 (̺, ̺0 ). On the other hand, it is known [13, p. 11, Proposition 1.1.6] that   1  21  1 2 2 2 W2 (̺, ̺0 ) ≥ tr P + P0 − 2 P0 P P0 ⇒ W22 (̺, ̺0 ) ≥ right hand side of (44). (45) Now consider a candidate transport map M x + m where M and m are functions of P , P0 , µ, µ0 as in the statement. It suffices to prove that our candidate transport map indeed achieves the equality in (45). To this end, directly substituting the expressions for M and m, notice that the push-forward has mean M µ0 + m = µ, and covariance M P0 M ⊤ = P . So our candidate transport map (M , m) is feasible. To show optimality, from (10) notice that W22 (̺, ̺0 ) = inf tr(P + P0 − 2C), where C := M P0 solves P0 − C∈Rd×d CP −1 C ⊤  0, which has known optimal solution Copt :=  1 − 21 1 1 1 Mopt P0 = P0 P 2 P 2 P0 P 2 P 2 . Since our candi1   − 2 1 1 1 1 P 2 satisfies tr (M P0 ) = date M := P 2 P 2 P0 P 2    12 1 1 , the statement follows. tr (Mopt P0 ) = tr P02 P P02 In the Corollary below, combining Lemma 1 and 2, we recover a result in [14, Theorem 3.1]. Corollary 1: Given d-dimensional joint PDF ̺0 with mean µ0 , covariance P0 ≻ 0, suppose tr(P0 ) = τ0 . For fixed µ and τ > 0, √
2 √ inf W22 (̺, ̺0 ) = τ − τ 0 + k µ − µ0 k22 , (46) ̺∈Dµ,τ and is achieved by ̺(x) = τ0 τ
d2 ̺0 τ0 τ
(x − µ) + µ0 . Proof: Let us choose P := ττ0 P0 , and from (44) √ 2 √ observe that inf W22 (ξ, ̺0 ) = ( τ − τ 0 ) + k µ − ξ∈Dµ,P µ0 k22 . On the other hand, for any ̺ ∈ Dµ,τ , we know from (45) that  1  12 1 W22 (̺, ̺0 ) ≥ τ + τ0 − 2 tr P02 SP02 + k µ − µ0 k22 , where S is the covariance of ̺. Using Lemma 1, we get  21  1 1 √ ≤ τ τ0 ⇒ W22 (̺, ̺0 ) ≥ tr P02 SP02 √ √
2 inf W22 (ξ, ̺0 ) = τ − τ 0 + k µ − µ0 k22 , (47) [7] H. Risken, The Fokker-Planck Equation: Methods of Solution and Applications. Springer Series in Synergetics, Vol. 18, First ed., Springer; 1989. [8] L. Ambrosio, N. Gigli, and G. Savaré, Gradient Flows: in Metric Spaces and in the Space of Probability Measures. Lectures in Mathematics, ETH Zürich, Second ed., Birkhäuser; 2008. [9] R.S. Laugesen, P.G. Mehta, S.P. Meyn, and M. Raginsky, “Poisson’s Equation in Nonlinear Filtering”. SIAM Journal on Control and Optimization. Vol. 53, No. 1, pp. 501–525, 2015. ξ∈Dµ,P and that the equality is achieved when S = P = ττ0 P0 . In  d that case, det(S) = ττ0 det(P0 ), and hence Lemma 2 yields the arg inf ̺(x) for (46) as s    1  12 1 det(P0 ) − 12 − 12 2 2 ̺0 S S P0 S S (x − µ) + µ0 det(S)   τ  d2  τ 0 0 ̺0 (x − µ) + µ0 . = τ τ Lemma 3: If E(·) depends on ̺ only via the mean and covariance of ̺, then inf F (̺) is achieved by N (µ, P ). ̺∈Dµ,P Proof: As E(̺) ≡ E(µ, P ), henceR from (11) we get ̺ log ̺ dx. Since inf F (̺) = E(µ, P ) + β −1 inf ̺∈Dµ,P ̺∈Dµ,P [10] A. Uhlmann, “The “Transition Probability” in the State Space of a *-Algebra”. Reports on Mathematical Physics, Vol. 9, No. 2, pp. 273– 279, 1976. [11] D. Petz, Quantum Information Theory and Quantum Statistics. Theoretical and Mathematical Physics, First ed., Springer; 2008. [12] C.R. Givens, and R.M. Shortt, “A Class of Wasserstein Metrics for Probability Distributions”. The Michigan Mathematical Journal, Vol. 31, No. 2, pp. 231–240, 1984. [13] S.T. Rachev, and L. Rüschendorf, Mass Transportation Problems. Volume I: Theory. First ed., Springer; 1998. [14] E.A. Carlen, and W. Gangbo, “Constrained Steepest Descent in the 2-Wasserstein Metric”. Annals of Mathematics, pp. 807–846, 2003. [15] A. Halder, and E.D.B. Wendel, “Finite Horizon Linear Quadratic Gaussian Density Regulator with Wasserstein Terminal Cost”. Proceedings of the 2016 American Control Conference, pp. 7249–7254, 2016. N (µ, P ) is the maximum entropy PDF under prescribed mean µ and covariance P , hence the statement. [16] N.J. Higham, and H.M. Kim, “Solving A Quadratic Matrix Equation by Newton’s Method with Exact Line Searches”. SIAM Journal on Matrix Analysis and Applications, Vol. 23, No. 2, pp. 303–316, 2001. Lemma 4: For P , P0 ≻ 0,  21 1 1  1 1 ∂ 1 1  −1 −1 2 tr P02 P P02 = P02 P0 2 P −1 P0 2 P02 . ∂P 2 Proof: We refer the readers to Appendix B in [15]. [17] D. Liberzon, and R.W. Brockett, “Spectral Analysis of Fokker–Planck and Related Operators Arising from Linear Stochastic Differential Equations”. SIAM Journal on Control and Optimization, Vol. 38, No. 5, pp. 1453–1467, 2000. R EFERENCES [1] H.H. Bauschke, and P.L. Combettes, Convex Analysis and Monotone Operator Theory in Hilbert Spaces. CMS Books in Mathematics, Springer; 2011. [18] R.W. Brockett, and J.C. Willems, “Stochastic Control and the Second Law of Thermodynamics”. Proceedings of the 1978 IEEE Conference on Decision and Control including the 17th Symposium on Adaptive Processes, pp. 1007–1011, 1978. [19] R.L. Stratonovich, “Application of the Theory of Markov Processes for Optimum Filtration of Signals”. Radio Eng. Electron. Phys. (USSR), Vol. 1, pp. 1–19, 1960. [2] N. Parikh, and S. Boyd, “Proximal Algorithms”. Foundations and Trends in Optimization. Vol. 1, No. 3, pp. 127–239, 2014. [20] H.J. Kushner, “On the Differential Equations Satisfied by Conditional Densities of Markov Processes, with Applications”. Journal of the SIAM Series A Control, Vol 2, No. 1, pp. 106–119, 1964. [3] R. Jordan, D. Kinderlehrer, and F. Otto, “The Variational Formulation of the Fokker–Planck Equation”. SIAM Journal on Mathematical Analysis. Vol. 29, No. 1, pp. 1–17, 1998. [21] M. Fujisaki, G. Kallianpur, and H. Kunita, “Stochastic Differential Equations for the Non Linear Filtering Problem”. Osaka Journal of Mathematics, Vol. 9, No. 1, pp. 19–40, 1972. [4] C. Villani, Topics in Optimal Transportation. Graduate Studies in Mathematics, Vol. 58, First ed., American Mathematical Society; 2003. [22] R.E. Kalman, and R.S. Bucy, “New Results in Linear Filtering and Prediction Theory”. Journal of Basic Engineering, Vol. 83, No. 3, pp. 95–108, 1961. [5] J.-D. Benamou, and Y. Brenier, “A Computational Fluid Mechanics Solution to the Monge–Kantorovich Mass Transfer Problem”. Numerische Mathematik. Vol. 84, No. 3, pp. 375–393, 2000. [6] K.J. Åström, Introduction to Stochastic Control Theory. Academic Press; 1970. [23] A. Yezzi, and E.I. Verriest, “Nonlinear Observers via Regularized Dynamic Inversion”. Proceedings of the 2007 American Control Conference, pp. 1693–1698, 2007. 

3cs.SY

Leveraging Joint Interactions for Credibility Analysis in News Communities Subhabrata Mukherjee Gerhard Weikum Max Planck Institute for Informatics {smukherjee, weikum}@mpi-inf.mpg.de arXiv:1705.02667v2 [cs.AI] 9 May 2017 Abstract Media seems to have become more partisan, often providing a biased coverage of news catering to the interest of specific groups. It is therefore essential to identify credible information content that provides an objective narrative of an event. News communities such as digg, reddit, or newstrust offer recommendations, reviews, quality ratings, and further insights on journalistic works. However, there is a complex interaction between different factors in such online communities: fairness and style of reporting, language clarity and objectivity, topical perspectives (like political viewpoint), expertise and bias of community members, and more. This paper presents a model to systematically analyze the different interactions in a news community between users, news, and sources. We develop a probabilistic graphical model that leverages this joint interaction to identify 1) highly credible news articles, 2) trustworthy news sources, and 3) expert users who perform the role of “citizen journalists” in the community. Our method extends CRF models to incorporate real-valued ratings, as some communities have very fine-grained scales that cannot be easily discretized without losing information. To the best of our knowledge, this paper is the first full-fledged analysis of credibility, trust, and expertise in news communities. Categories and Subject Descriptors H.3.3 [Information Storage and Retrieval]: Information Search and Retrieval - Information Filtering; I.2.7 [Computing Methodologies]: Artificial Intelligence - Natural Language Processing Keywords Credibility; News Community; Probabilistic Graphical Models 1. INTRODUCTION Motivation: Media plays a crucial role in the public dissemination of information about events. Many people find online information and blogs as useful as TV or magazines. At the same time, however, people also believe that there is substantial media bias in news coverage [24, 8], especially in view of inter-dependencies and crossownerships of media companies and other industries (like energy). Several factors affect the coverage and presentation of news in media incorporating potentially biased information induced via the Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from Permissions@acm.org. CIKM’15, October 19-23, 2015, Melbourne, VIC, Australia Copyright 2015 ACM. ISBN 978-1-4503-3794-6/15/10 ...$15.00. DOI: http://dx.doi.org/10.1145/2806416.2806537 . fairness and style of reporting. News are often presented in a polarized way depending on the political viewpoint of the media source (newspapers, TV stations, etc.). In addition, other source-specific properties like viewpoint, expertise, and format of news may also be indicators of information credibility. In this paper, we embark on an in-depth study and formal modeling of these factors and inter-dependencies within news communities for credibility analysis. A news community is a news aggregator site (e.g., reddit.com, digg.com, newstrust.net) where users can give explicit feedback (e.g., rate, review, share) on the quality of news and can interact (e.g., comment, vote) with each other. Users can rate and review news, point out differences, bias in perspectives, unverified claims etc. However, this adds user subjectivity to the evaluation process, as users incorporate their own bias and perspectives in the framework. Controversial topics create polarization among users which influence their ratings. [30, 6] state that online ratings are one of the most trusted sources of user feedback; however they are systematically biased and easily manipulated. Problem Statement: Given a set of news sources generating news articles, and users reviewing those articles on different qualitative aspects with mutual interactions — our objective is to jointly rank the sources, articles, and users based on their trustworthiness, credibility, and expertise respectively. In this process, we want to analyze the influence of various factors like the writing style of a news article, its topic distribution, type of media and format of news, political viewpoint and expertise, and other user traits on the credibility analysis of the community. Our Approach: To analyze the factors and inter-dependencies in a news community, we have developed a sophisticated probabilistic graphical model, specifically a Continuous Conditional Random Field (CCRF) model, which exploits several moderate signals of interaction jointly between the following factors to derive a strong signal for information credibility (refer to Figures 1a and 1b). In particular, the model captures the following factors. • Language and credibility of a news article: objectivity, rationality, and general quality of language in the news article. Objectivity is the quality of the news to be free from emotion, bias and prejudice of the author. The credibility of a news article refers to presenting an unbiased, informative and balanced narrative of an event. • Properties and trustworthiness of a news source: trustworthiness of a news source in the sense of generating credible articles based on source properties like viewpoint, expertise and format of news. • Expertise of users and review ratings: expertise of a user, in the news community, in properly judging the credibility of news articles. Expert users should provide objective evaluations – by reviews and/or ratings – of news articles, corroborating with the evaluations of other expert users. This can be used to identify potential “citizen journalists” [17] in the community. We show that the CCRF performs better than sophisticated col- 2. MODELING NEWS COMMUNITIES Our approach exploits the rich interaction taking place between the different factors in a news community. We propose a probabilistic graphical model that leverages the interplay between news credibility, language objectivity, source trustworthiness, and user expertise. Refer to Figure 1 for the following discussion. Consider a set of news sources hsi (e.g., s1 in Figure 1c) generating articles hdi which are reviewed and analyzed by users hui for their credibility. Consider rij to be the review by user uj on article di . The overall article rating of di is given by yi . In our model, each news source, news article, user and her rating or review, and overall article rating is associated with a continuous random variable r.v. ∈ [1 . . . 5], that indicates its trustworthiness, objectivity, expertise, and credibility, respectively. 5 indicates the best quality that an item can obtain, and 1 is the worst. Discrete ratings, being a special case of this setting, can be easily handled. Each node is associated with a set of observed features that are extracted from the news community. For example, a news source has properties like topic specific expertise, viewpoint and format of news; a news article has features like topics, and style of writing from the usage of discourse markers and subjective words in the article. For users we extract their topical perspectives and expertise, engagement features (like the number of questions, replies, Joint Probabilistic Inference Source Trustworthiness News Article Language Objectivity News Source Users Topics Reviews User Expertise …, Engagement, Expertise V Ex iew pe p rt oin is t e, , ... Observable Features Emotionality, Discourse,... laborative filtering approaches based on latent factor models, and regression methods that do not consider all these interactions. Although this work is focused on news communities, the framework can also be used for instance, in health communities (e.g. healthboards.com) where users write posts on drug usage — the objective being to jointly rank posts, drug side-effects, and users based on their quality, credibility, and trustworthiness respectively. In this work, the attributes credibility and trustworthiness are always associated with a news article and a news source, respectively. The joint interaction between several factors also captures that a source garners trustworthiness by generating credible news articles, which are highly rated by expert users. Similarly, the likelihood of a news article being credible increases if it is generated by a trustworthy source. Some communities offer users fine-grained scales for rating different aspects of news articles and news sources. For example, the newstrust.net community analyzes an article on 15 aspects like insightful, fairness, style and factual. These are aggregated into an overall real-valued rating after weighing the aspects based on their importance, expertise of the user, feedback from the community, and more. This setting cannot be easily discretized without blowup or risking to lose information. Therefore, we model ratings as real-valued variables in our CCRF. Contributions: The paper introduces the following novel elements: • A continuous CRF that captures the mutual dependencies between credibility of articles, trustworthiness of sources, expertise of users, and expresses real-valued ratings. • An inference method for the CCRF that allows us to jointly (a) predict ratings; and (b) rank articles, sources, and users by their credibility, trustworthiness, and expertise, respectively. • A large experimental study with data from newstrust.net, one of the most sophisticated news communities with a focus on quality journalism. The rest of the paper is organized as follows. Section 2 presents how we model news communities, and which factors we include in the model. Section 3 develops the CCRF that captures the interaction between all the factors. Section 4 introduces the dataset that we use for experimental evaluation and further studies. Section 5 presents our experimental results followed by discussion. Article Credibility Ratings (a) Interactions between source trustworthiness, article credibility, language objectivity, and user expertise. (b) Sample instantiation. (c) Clique representation. Figure 1: Graphical model representation. reviews posted) and various interactions with other users (like upvotes/downvotes) and news sources in the community. The objective of our model is to predict credibility ratings hyi of news articles hdi by exploiting the mutual interactions between different variables. The following edges between the variables capture their interplay: • Each news article is connected to the news source from where it is extracted (e.g., s1 − d1 , s1 − d2 ) • Each news article is connected to its review or rating by a user (e.g., d1 − r11 , d1 − r12 , d2 − r22 ) • Each user is connected to all her reviews (e.g., u1 − r11 , u2 − r12 , u2 − r22 ) • Each user is connected to all news articles rated by her (e.g., u1 − d1 , u2 − d1 , u2 − d2 ) • Each source is connected to all the users who rated its articles (e.g., s1 − u1 , s1 − u2 ) • Each source is connected to all the reviews of its articles (e.g., s1 − r11 , s1 − r12 , s1 − r22 ) • For each article, all the users and all their reviews on the article are inter-connected (e.g., u1 − r12 , u2 − r11 , u1 − u2 ). This captures user-user interactions (e.g., u1 upvoting/downvoting u2 ’s rating on d1 ) influencing the overall article rating. Therefore, a clique (e.g., C1 ) is formed between a news article, its source, users and their reviews on the article. Multiple such cliques (e.g., C1 and C2 ) share information via their common news sources (e.g., s1 ) and users (e.g., u2 ). News topics play a significant role on information credibility. Individual users in community (and news sources) have their own perspectives and expertise on various topics (e.g., environmental politics). Modeling user-specific topical perspectives explicitly captures credibility judgment better than a user-independent model. However, many articles do not have explicit topic tags. Hence we use Latent Dirichlet Allocation (LDA) [1] in conjunction with Support Vector Regression (SVR) [4] to learn words associated to each (latent) topic, and user (and source) perspectives for the topics. Documents are assumed to have a distribution over topics as latent variables, with words as observables. Inference is by Gibbs sampling. This LDA model is a component of the overall model, discussed next. We use a probabilistic graphical model, specifically a Conditional Random Field (CRF), to model all factors jointly. The modeling approach is related to the prior work of [23]. However, unlike that work and traditional CRF models, our problem setting requires a continuous version of the CRF (CCRF) to deal with real-valued ratings instead of discrete labels. In this work, we follow an approach similar to [26, 27, 32] in learning the parameters of the CCRF. We use Support Vector Regression [4] to learn the elements of the feature vector for the CCRF. The inference is centered around cliques of the form h source, article, h users i, h reviews ii. An example is the two cliques C1 : s1 − d1 − hu1 , u2 i − hr11 , r12 i and C2 : s1 − d2 − u2 − r22 in the instance graph of Figure 1c. This captures the “cross-talk” between different cliques sharing nodes. A news source garners trustworthiness by generating multiple credible articles. Users attain expertise by correctly identifying credible articles that corroborate with other expert users. Inability to do so brings down their expertise. Similarly, an article attains credibility if it is generated by a trustworthy source and highly rated by an expert user. The inference algorithm for the CCRF is discussed in detail in Section 3. In the following subsections, we discuss the various feature groups that are considered in our CCRF model. 2.1 Articles and their Stylistic Features The style in which news is presented to the reader plays a pivotal role in understanding its credibility. The desired property for news is to be objective and unbiased. In this section, we examine the different stylistic indicators of news credibility. All the lexicons used in this section are compiled from [28, 23]. Assertives: Assertive verbs (e.g., “claim”) complement and modify a proposition in a sentence. They capture the degree of certainty to which a proposition holds. Factives: Factive verbs (e.g., “indicate”) pre-suppose the truth of a proposition in a sentence. Hedges: These are mitigating words (e.g., “may”) to soften the degree of commitment to a proposition. Implicatives: These words trigger pre-supposition in an utterance. For example, usage of the word complicit indicates participation in an activity in an unlawful way. Report verbs: These verbs (e.g., “argue”) are used to indicate the attitude towards the source, or report what someone said more accurately, rather than using just say and tell. Discourse markers: These capture the degree of confidence, perspective, and certainty in the set of propositions made. For instance, strong modals (e.g., “could”), probabilistic adverbs (e.g., “maybe”), and conditionals (e.g., “if”) depict a high degree of uncertainty and hypothetical situations, whereas weak modals (e.g., “should”) and inferential conjunctions (e.g., “therefore”) depict certainty. Subjectivity and bias: News is supposed to be objective: writers should not convey their own opinions, feelings or prejudices in their stories. For example, a news titled “Why do conservatives hate your children?” is not considered objective journalism. We use a subjectivity lexicon1 , a list of positive and negative opinionated words2 , and an affective lexicon3 to detect subjective clues in arti1 http://mpqa.cs.pitt.edu/lexicons/subj_lexicon/ http://www.cs.uic.edu/ liub/FBS/opinion-lexicon-English.rar 3 http://wndomains.fbk.eu/wnaffect.html 2 Latent Topics Topic Words Obama admin. Citizen journ. US military AmyGoodman Alternet Climate obama, republican, party, election, president, senate, gop, vote cjr, jouralism, writers, cjrs, marx, hutchins, reporting, liberty, guides iraq, war, military, iran, china, nuclear, obama, russia, weapons democracy, military, civil, activist, protests, killing, navajo, amanda media, politics, world news, activism, world, civil, visions, economy energy, climate, power, water, change, global, nuclear, fuel, warming Table 1: Latent topics (with illustrative labels) and their words. cles. The affective features capture the state of mind (like attitude and emotions) of the writer while writing an article or post (e.g., anxiousness, confidence, depression, favor, malice, sympathy etc.). We additionally harness a lexicon of bias-inducing words extracted from the Wikipedia edit history from [28] exploiting its Neutral Point of View Policy to keep its articles “fairly, proportionately, and as far as possible without bias, all significant views that have been published by reliable sources on a topic”. Feature vector construction: For each stylistic feature type fi and each news article dj , we compute the relative frequency of words of type fi occurring in dj , thus constructing a feature vector F L (dj ) = hf reqij = #(words in fi ) / length(dj )i. Consider the review rj,k written by user uk on the article dj . For each such review, analogous to the per-article stylistic feature vector hF L (dj )i, we construct a per-review feature vector hF L (rj,k )i. 2.2 Articles and their Topics Topic tags for news articles play an important role in user-perceived prominence, bias and credibility, in accordance to the ProminenceInterpretation theory [7]. For example, the tag Politics is often viewed as an indicator of potential bias and individual differences; whereas tags like Energy or Environment are perceived as more neutral news and therefore invoke higher agreement in the community on the associated articles’ credibility. Obviously, this can be misleading as there is a significant influence of Politics on all topics in all format of news. Certain users have topic-specific expertise that make them rate articles on those topics better than others. News sources also have expertise on specific topics and provide a better coverage of news on those topics than others. For example, National Geographic provides a good coverage of news related to environment, whereas The Wall Street Journal provides a good coverage on economic policies. However, many news articles do not have any explicit topic tag. In order to automatically identify the underlying theme of the article, we use Latent Dirichlet Allocation (LDA) [1] to learn the latent topic distribution in the corpus. LDA assumes a document to have a distribution over a set of topics, and each topic to have a distribution over words. Table 1 shows an excerpt of the top topic words in each topic, where we manually added illustrative labels for the topics. The latent topics also capture some subtle themes not detected by the explicit tags. For example, Amy Goodman is an American broadcast journalist, syndicated columnist and investigative reporter who is considered highly credible in the community. Also, associated with that topic cluster is Amanda Blackhorse, a Navajo activist and plaintiff in the Washington Redskins case. Feature vector construction: For each document dj and each of its review rj,k , we create feature vectors hF T (dj )i and hF T (rj,k )i respectively, using the learned latent topic distributions, as well as the explicit topic tags. Section 3.1 discusses our method to learn the topic distributions. 2.3 News Sources A news source is considered trustworthy if it generates highly credible articles. We examine the effect of different features of a news source on its trustworthiness based on user assigned ratings Category Elements Variables Type Description Media Format Scope Viewpoint Top Topics Expertise on Topics newspaper, blog, radio, magazine, online editorial, investigative report, news, research local, state, regional, national, international far left, left, center, right, neutral politics, weather, war, science„ U.S. military U.S. congress, Middle East, crime, presidential election, Bush administration, global warming dj s u rj,k Vector Vector Vector Vector yj,k z SVRuk , SVRsi SVRL , SVRT Ψ = f (hψj i) Real Number Vector Real Number ∈ [1 . . . 5] Real Number λ= hαu , βs , γ1 , γ2 i yn×1 Xn×m Qn×n bn×1 Σn×n µn×1 Vector Document with sequence of words hwi Sources Users Review by user uk on document dj with sequence of words hwi Rating of rj,k Sequence of topic assignments for hwi SVR prediction for users, sources, language, and topics Clique potential with ψj = hyj , si , dj , huk i, hrj,k ii for clique of dj Combination weights for users hui, sources hsi, language and topic models Credibility rating of documents hdi Feature matrix with m = |U | + |S| + 2 f (λ) f (λ, X) f (λ) f (λ, X) Table 2: Features for source trustworthiness. in the community. We consider the following source features (summarized in Table 2): the type of media (e.g., online, newspaper, tv, blog), format of news (e.g., news analysis, opinion, special report, news report, investigative report), (political) viewpoint (e.g., left, center, right), scope (e.g., international, national, local), the top topics covered by the source, and their topic-specific expertise. Feature vector construction: For each news source sl , we create a feature vector hF S (sl )i using features in Table 2. Each element fiS (sl ) is 1 or 0 indicating presence or absence of a feature. Note that above features include the top (explicit) topics covered by any source, and its topic-specific expertise for a subset of those topics. 2.4 Users, Ratings and Interactions A user’s expertise in judging news credibility depends on many factors. [5] discusses the following traits for recognizing an expert. Community Engagement of the user is an obvious measure for judging the user authority in the community. We capture this with different features: number of answers, ratings given, comments, ratings received, disagreement and number of raters. Inter-User Agreement: Expert users typically agree on what constitutes a credible article. This is inherently captured in the proposed graphical model, where a user gains expertise by assigning credibility ratings to articles that corroborate with other expert users. Topical Perspective and Expertise: The potential for harvesting user preference and expertise in topics for rating prediction of reviews has been demonstrated in [22, 21]. For credibility analysis the model needs to capture the user’s perspective and bias towards certain topics based on their political inclination that bias their ratings, and their topic-specific expertise that allows them to evaluate articles on certain topics better as “Subject Matter Experts”. These are captured as per-user feature weights for the stylistic indicators and topic words in the language of user-contributed reviews. Interactions: In a community, users can upvote (digg, like, rate) the ratings of users that they appreciate, and downvote the ones they do not agree with. High review ratings from expert users increase the value of a user; whereas low ratings bring down her expertise. Similar to this user-user interaction, there can be user-article, usersource and source-article interactions which are captured as edges in our graphical model (by construction). Consider the following anecdotal example in the community showing an expert in nuclear energy downvoting another user’s rating on nuclear radiation: “Nonexpert: Interesting opinion about health risks of nuclear radiation, from a physicist at Oxford University. He makes some reasonable points ... Low rating by expert to above review: Is it fair to assume that you have no background in biology or anything medical? While this story is definitely very important, it contains enough inaccurate and/or misleading statements...” Feature vector construction: For each user uk , we create an engagement feature vector hF E (uk )i. In order to capture user subjectivity, in terms of different stylistic indicators of credibility, we consider the per-review language feature vector hF L (rj,k )i of user uk (refer to Section 2.1). To capture user perspective and expertise on different topics, we consider the per-review topic feature vector hF T (rj,k )i of each user uk . Vector Matrix Diagonal Matrix Vector CovarianceMatrix Mean Vector Table 3: Symbol table. 3. JOINT PROBABILISTIC INFERENCE In this section we incorporate the discussed features and insights into a joint probabilistic graphical model. The task is to identify credible news articles, trustworthy news sources, and expert users jointly in a news community. Table 3 summarizes the important notations used in this section. 3.1 Topic Model Consider an article d consisting of a sequence of {Nd } words denoted by w1 , w2 , ...wNd . Each word is drawn from a vocabulary V having unique words indexed by 1, 2, ...V . Consider a set of topic assignments z = {z1 , z2 , ...zK } for d, where each topic zi can be from a set of K possible topics. LDA [1] assumes each document d to be associated with a multinomial distribution θd over topics Z with a symmetric dirichlet prior ρ. θd (z) denotes the probability of occurrence of topic z in document d. Topics have a multinomial distribution φz over words drawn from a vocabulary V with a symmetric dirichlet prior ζ. φz (w) denotes the probability of the word w belonging to the topic z. Exact inference is not possible due to intractable coupling between Θ and Φ. We use Gibbs sampling for approximate inference. Let n(d, z, w) denote the count of the word w occurring in document d belonging to the topic z. In the following equation, (.) at any position in the above count indicates marginalization, i.e., summing up the counts over all values for the corresponding position in n(d, z, w). The conditional distribution for the latent variable z (with components z1 to zK ) is given by: P (zi = k|wi = w, z−i , w−i ) ∝ n(., k, w) + ζ n(d, k, .) + ρ P ×P k n(d, k, .) + Kρ w n(., k, w) + V ζ (1) Let hT E i and hT L i be the set of explicit topic tags and latent topic dimensions, respectively. The topic feature vector hF T i for an article or review combines both explicit tags and latent topics and is constructed as follows:  if T E0 = FtT  #f req(w, d), t T Ft (d) = #f req(w, d) × φT L0 (w), if TtL0 = FtT and φT L0 (w) > δ t  t  0 otherwise So for any word in the document matching an explicit topic tag, the corresponding element in the feature vector hF T i is set to its occurrence count in the document. If the word belongs to any latent topic with probability greater than threshold δ, the probability of the word belonging to that topic (φt (w)) is added to the corresponding element in the feature vector, and set to 0 otherwise. 3.2 Support Vector Regression We use Support Vector Regression (SVR) [4] to combine the different features discussed in Section 2. SVR is an extension of the max-margin framework for SVM classification to the regression problem. It solves the following optimization problem to learn weights w for features F : N min w X 1 T (max(0, |yd − wT F | − ))2 w w+C× 2 d=1 (2) Article Stylistic Model: We learn a stylistic regression model SVRL using the per-article stylistic feature vector hF L (dj )i for article dj (or, hF L (rj,k )i for review rj,k ), with the overall article rating yj (or, yj,k ) as the response variable. Article Topic Model: Similarly, we learn a topic regression model SVRT using the per-article topic feature vector hF T (dj )i for article dj (or, hF T (rj,k )i for review rj,k ), with the overall article rating yj (or, yj,k ) as the response variable. Source Model: We learn a source regression model SVRsi using the per-source feature vector hF S (si )i for source si , with the overall source rating as the response variable . User Model: For each user uk , we learn a user regression model SVRuk with her per-review stylistic and topic feature vectors hF L (rj,k )∪F T (rj,k )i for review rj,k for article dj , with her overall review rating yj,k as the response variable. Note that we use overall article rating to train article stylistic and topic models. For the user model, however, we take user assigned article ratings and per-user features. This model captures user subjectivity and topic perspective. The source models are trained on news-source specific meta-data and its ground-truth ratings. 3.3 Continuous Conditional Random Field We model our learning task as a Conditional Random Field (CRF), where the random variables are the ratings of news articles hdj i, news sources hsi i, users huk i, and reviews hrj,k i. The objective is to predict the credibility ratings hyj i of the articles hdj i. The cliques in the CRF consist of an article dj , its source si , set of users huk i reviewing it, and the corresponding user reviews hrj,k i — where rj,k denotes the review by user uk on article dj . Different cliques are connected via the common news sources, and users. There are as many cliques as the number of news articles. Let ψj (yj , si , dj , huk i, hrj,k i) be a potential function for clique j. Each clique has a set of associated vertex feature functions. In our problem setting, we associate features to each vertex.The features constituted by the stylistic, topic, source and user features explained in Section 2 are: F L (dj )∪F T (dj )∪F S (si )∪k (F E (uk )∪ F L (rj,k ) ∪ F T (rj,k )). A traditional CRF model allows us to have a binary decision if a news article is credible (yj = 1) or not (yj = 0), by estimating the conditional distribution with the probability mass function of the discrete random variable y: Qn j=1 exp(ψj (yj , si , dj , huk i, hrj,k i)) P r(y|D, S, U, R) = P Qn y j=1 exp(ψj (yj , si , dj , huk i, hrj,k i)) Given a news article dj , its source id si , and a set of user ids huk i who reviewed the article, the regression models SVRL (dj ), SVRT (dj ), SVRsi , hSVRuk (dj )i (discussed in Section 3.2) predict rating of dj . For notational brevity, hereafter, we drop the argument dj from the SVR function. These SVR predictors are for separate feature groups and independent of each other. Now we combine the different SVR models to capture mutual interactions, such that the weight for each SVR model reflects our confidence on its quality. Errors by an SVR are penalized by the squared loss between the predicted article rating and the ground-truth rating. There is an additional constraint that for any clique only the regression models corresponding to the news-source and users present in it should be activated. This can be thought of as partitioning the input feature space into subsets, with the features inside a clique capturing local interactions, and the global weights capture the overall quality of the random variables via the shared information between the cliques (in terms of common sources, users, topics and language features) — an ideal setting for using a CRF. Equation 5 shows one such linear combination. Energy function of an individual clique is given by: ψj (yj , si , dj , huk i, hrj,k i) = − P r(y|D, S, U, R) = n j=1 exp(ψj (yj , si , dj , huk i, hrj,k i)) R∞ Q n j=1 exp(ψj (yj , si , dj , huk i, hrj,k i))dy −∞ (4) αk Iuk (dj )(yj − SVRuk )2 k=1 − i=S X βi Isi (dj )(yj −SVRsi )2 −γ1 (yj −SVRL )2 −γ2 (yj −SVRT )2 i=1 (5) Indicator functions Iuk (dj ) and Isi (dj ) are 1 if uk is a reviewer and si is the source of article dj respectively, and are 0 otherwise. As the output of the SVR is used as an input to the CCRF in Equation 5, each element of the input feature vector is already predicting the output variable. The learned parameters λ = hα, β, γ1 , γ2 i (with dimension(λ) = |U | + |S| + 2) of the linear combination of the above features depict how much to trust individual predictors. Large λk on a particular predictor places large penalty on the mistakes committed by it, and therefore depicts a higher quality for that predictor. αu corresponding to user u can be taken as a proxy for that user’s expertise, allowing us to obtain a ranked list of expert users. Similarly, βs corresponding to news source s can be taken as a proxy for that source’s trustworthiness, allowing us to obtain a ranked list of trustworthy news sources. Overall energy function of all cliques is given by: Ψ= n X ψj (yj , si , dj , huk i, hrj,k i) j=1 (Substituting ψj from Equation 5 and re-organizing terms) Ψ= n X (− j=1 − i=S X k=U X αk Iuk (dj )(yj − SVRuk )2 k=1 βi Isi (dj )(yj − SVRsi )2 − γ1 (yj − SVRL )2 − γ2 (yj − SVRT )2 ) i=1 (3) But in our problem setting, we want to estimate the credibility rating of an article. Therefore, we need to estimate the conditional distribution with the probability density function of the continuous random variable y: Q k=U X =− n X yj2 [ j=1 + n X j=1 2yj [ k=U X αk Iuk (dj ) + k=1 k=U X βi Isi (dj ) + γ1 + γ2 ] i=1 αk Iuk (dj )SVRuk + k=1 + γ2 SVRT ] − i=S X i=S X βi Isi (dj )SVRsi + γ1 SVRL i=1 n k=U i=S X X X [ αk Iuk (dj )SVR2uk + βi Isi (dj )SVR2si j=1 k=1 i=1 + γ1 SVR2L + γ2 SVR2T ] Organizing the bracketed terms into variables as follows: (P k=U k=1 Qi,j = bi = 2[ αk Iuk (di ) + Pl=S l=1 βl Isl (di ) + γ1 + γ2 0 k=U X l=S X αk Iuk (di )SVRuk + ∂Σ ∂Σ−1 = −Σ Σ ∂λk ∂λk = −2ΣΣ i=j i 6= j ∂ ∂ (µT Σ−1 µ) = (bT Σb) ∂λk ∂λk βl Isl (di )SVRsl +γ1 SVRL +γ2 SVRT ] l=1 k=1 = bT i=S n k=U X X X βi Isi (dj )SVR2si +γ1 SVR2L +γ2 SVR2T ] αk Iuk (dj )SVR2uk + c= [ i=1 j=1 k=1 = bT (Σ We can derive: Ψ = −y T Qy + y T b − c Qn −∞ = 4X(.),k µ − 2µT µ ∂log|Σ−1 | 1 ∂Σ−1 = Trace(|Σ−1 |Σ ) ∂λk |Σ−1 | ∂λk exp(ψj ) j=1 exp(ψj )dy = 2Trace(Σ) exp(Ψ) −∞ exp(Ψ)dy = R∞ We can derive the gradient vector: exp(−y T Qy + y T b) T T −∞ exp(−y Qy + y b)dy (7) ∂logP (y|X) T = −y T y+2y T X(.),k −2X(.),k µ+µT µ+ Trace(Σ) ∂λk (11) = R∞ = R∞ exp(− 21 y T Σ−1 y + y T Σ−1 µ) −∞ exp(− 21 y T Σ−1 y + y T Σ−1 µ)dy Let η denote the learning rate. The update equation is given by: (Substituting Q = 12 Σ−1 , b = Σ−1 µ) Equation 7 can be transformed R ∞ into a multivariate Gaussian distribution after substituting −∞ exp(− 12 y T Σ−1 y + y T Σ−1 µ)dy = (2π)n/2 1 |Σ−1 | 2 exp( 12 µT Σ−1 µ). Therefore obtaining, P (y|X) = 1 n ∂Σ ∂bT ∂b + b) + Σb ∂λk ∂λk ∂λk = 4X(.),k Σb − 2bT ΣΣb (6) Substituting Ψ in Equation 4: j=1 P (y|X) = R ∞ Qn ∂Σb ∂bT + Σb ∂λk ∂λk 1 (2π) 2 |Σ| 2 1 exp(− (y − µ)T Σ−1 (y − µ)) 2 (8) Q represents the contribution of λ to the covariance matrix Σ. Each row of the vector b and matrix Q corresponds to one training instance, representing the active contribution of features present in it. To ensure Equation 8 represents a valid Gaussian distribution, the covariance matrix Σ needs to be positive definite for its inverse to exist. For that the diagonal matrix Q needs to be a positive semidefinite matrix. This can be ensured by making all the diagonal elements in Q greater than 0, by constraining λk > 0. Since this is a constrained optimization problem, gradient ascent cannot be directly used. We follow the approach similar to [27] and maximize log-likelihood with respect to log λk , instead of λk as in standard gradient ascent, making the optimization problem unconstrained as: ∂logP (y|X) ∂logP (y|X) = αk ( ) ∂logλk ∂λk (9) Taking partial derivative of the log of Equation 8 w.r.t λk : ∂logP (y|X) 1 ∂ = (−y T Σ−1 y + 2y T Σ−1 µ − µT Σ−1 µ ∂λk 2 ∂λk + log|Σ−1 | + Constant) (10) Substituting the following in the above equation: ∂Q ∂Σ−1 =2 ∂λk ∂λk = 2I ∂Σ−1 µ ∂b = [∵ µ = Σb] ∂λk ∂λk = 2X(.),k where, X(.),k indicates the kth column of the feature matrix X. logλnew = logλold k k +η ∂logP (y|X) ∂logλk (12) Once the model parameters are learned using gradient ascent, the inference for the prediction y of the article credibility rating is straightforward. As we assume the distribution to be Gaussian, the prediction is the expected value of the function, given by the mean of the distribution: y0 = argmaxy P (y|X) = µ = Σb. Note that Σ and b are both a function of λ = hα, β, γ1 , γ2 i which represents the combination weights of various factors to capture mutual interactions. The optimization problem determines the optimal λ for reducing the error in prediction. 4. USE CASE: NEWSTRUST We performed experiments with data from a typical news community: newstrust.net4 . This community is similar to digg.com and reddit.com, but has more refined ratings and interactions. We chose NewsTrust because of the availability of ground-truth ratings for credibility analysis of news articles; such ground-truth is not available for the other communities. We collected stories from NewsTrust from May, 2006 to May, 2014. Each such story features a news article from a source (E.g. BBC, CNN, Wall Street Journal) that is posted by a member, and reviewed by other members, many of whom are professional journalists and content experts5 . We crawled all the stories with their explicit topic tags and other associated meta-data. We crawled all the news articles from their original sources that were featured in any NewsTrust story. The earliest story dates back to May 1, 1939 and the latest one is in May 9, 2014. We collected all member profiles containing information about the demographics, occupation and expertise of the members along with their activity in the community in terms of the posts, reviews and ratings; as well as interaction with other members. The members in the community can also rate each others’ ratings. The earliest story rating by a member dates back to May, 2006 and the most recent one 4 Code and data available at http://www.mpi-inf.mpg.de/impact/credibilityanalysis/ 5 http://www.newstrust.net/help#about_newstrust Factors Count Unique news articles reviewed in NewsTrust NewsTrust stories on news articles NewsTrust stories with ≥ 1 reviews NewsTrust stories with ≥ 3 reviews NewsTrust member reviews of news articles 62,064 84,704 43,107 18,521 134,407 News articles extracted from original sources NewsTrust stories on extracted news articles 47,565 52,579 News sources Journalists who wrote news articles Timestamps (month and year) of posted news articles 5,658 19,236 3,122 NewsTrust members who reviewed news articles NewsTrust members who posted news articles 668 Explicit topic tags Latent topics extracted 456 300 Nodes Sources Members News articles Reviews Edges Total triangles Count 181,364 1,704 6,906 42,204 130,550 602,239 521,630 Factors No. of weakly connected components Diameter Average path length Average degree Average clustering coefficient Modularity Latent Factor Models (LFM) Simple LFM [15] Experience-based LFM [21] Text-based LFM [20] Our Model: User SVR 0.95 0.85 0.78 0.60 Model Table 4: Dataset statistics. Factors MSE Table 6: MSE comparison of models for predicting users’ credibility rating behavior with 10-fold cross-validation. Improvements are statistically significant with P-value < 0.0001. 7,114 1,580 News sources reviewed by NewsTrust members Model Count 12 8 47 6.641 0.884 0.516 Table 5: Graph statistics. is in Feb, 2014. In addition, we collected information on member evaluation of news sources, and other information (e.g., type of media, scope, viewpoint, topic specific expertise) about source from its meta data. Crawled dataset: Table 4 shows the dataset statistics. In total 62K unique news articles were reviewed in NewsTrust in the given period, out of which we were able to extract 47K full articles from the original sources like New York Times, TruthDig, ScientificAmerican etc — a total of 5.6K distinct sources. The remaining articles were not available for crawling. There are 84.7K stories featured in NewsTrust for all the above articles, out of which 52.5K stories refer to the news articles we managed to extract from their original sources. The average number of reviews per story is 1.59. For general analysis we use the entire dataset. For experimental evaluation of the CCRF and hypotheses testing, we use only those stories ( 18.5K) with a minimum of 3 reviews that refer to the news articles we were able to extract from original sources. Generated graph: Table 5 shows the statistics of the graph constructed by the method of Section 2. Ground-Truth for evaluation: The members in the community can rate the credibility of a news article on a scale from 1 to 5 regarding 15 qualitative aspects like facts, fairness, writing style and insight, and popularity aspects like recommendation, credibility and views. Members give an overall recommendation for the article explained to them as: “... Is this quality journalism? Would you recommend this story to a friend or colleague? ... This question is similar to the up and down arrows of popular social news sites like Digg and Reddit, but with a focus on quality journalism." Each article’s aspect ratings by different members are weighted (and aggregated) by NewsTrust based on findings of [16], and the member expertise and member level (described below). This overall article rating is taken as the ground-truth for the article credibility rating in our work. A user’s member level is calculated by NewsTrust based on her community engagement, experience, other users’ feedback on her ratings, profile transparency and validation by NewsTrust staff. This member level is taken as the proxy for user expertise in our work. Members rate news sources while reviewing an article. These ratings are aggregated for each source, and taken Only Title MSE Title & Text MSE Language Model: SVR Language (Bias and Subjectivity) Explicit Topics Explicit + Latent Topics All Topics (Explicit + Latent) + Language 3.89 1.74 1.68 1.57 0.72 1.74 1.01 0.61 News Source Features and Language Model: SVR News Source News Source + All Topics + Language 1.69 0.91 1.69 0.46 Aggregated Model: SVR Users + All Topics + Language + News Source 0.43 0.41 Our Model: CCRF+SVR User + All Topics + Language + News Source 0.36 0.33 Table 7: MSE comparison of models for predicting aggregated article credibility rating with 10-fold cross-validation. Improvements are statistically significant with P-value < 0.0001. as a proxy for the source trustworthiness in our work. Training data: We perform 10-fold cross-validation on the news articles. During training on any 9-folds of the data, the algorithm learns the user, source, language and topic models from user-assigned ratings to articles and sources present in the train split. We combine sources with less than 5 articles and users with less than 5 reviews into background models for sources and users, respectively. This is to avoid modeling from sparse observations, and to reduce dimensionality of the feature space. However, while testing on the remaining blind 1-fold we use only the ids of sources and users reviewing the article; we do not use any user-assigned ratings of sources or articles. For a new user and a new source, we draw parameters from the user or source background model. The results are averaged by 10-fold cross-validation, and presented in the next section. Experimental settings: In the first two experiments we want to find the power of the CCRF in predicting user rating behavior, and credibility rating of articles. Therefore, the evaluation measure is taken as the Mean Squared Error (MSE) between the prediction and the actual ground-rating in the community. For the latter experiments in finding expert users (and, trustworthy sources) there is no absolute measure for predicting user (and, source) quality; it only makes sense to find the relative ranking of users (and, sources) in terms of their expertise (and, trustworthiness). Therefore, the evaluation measure is taken as the Normalized Discounted Cumulative Gain (NDCG) [13] between the ranked list of users (and, sources) obtained from CCRF and their actual ranking in the community. 5. 5.1 EXPERIMENTS Predicting User Ratings of Articles First we evaluate how good our model can predict the credibility ratings that users assign to news articles using the Mean Squared Error (MSE) between the prediction and the actual rating. Model Experience LFM [21] PageRank CCRF NDCG 0.80 0.83 0.86 Table 8: NDCG scores for ranking trustworthy sources. Model Experience LFM [21] Member Ratings CCRF NDCG 0.81 0.85 0.91 Table 9: NDCG scores for ranking expert users. Baselines: We consider the following baselines for comparison: 1. Latent Factor Recommendation Model (LFM) [15]: LFM considers the tuple huserId, itemId, ratingi, and models each user and item as a vector of latent factors which are learned by minimizing the MSE between the rating and the product of the user-item latent factors. In our setting, each news article is considered an item and rating refers to the credibility rating assigned by a user to an article. 2. Experience-based LFM [21]: This model incorporates experience of a user in rating an item in the LFM. The model builds on the hypothesis that users at similar levels of experience have similar rating behaviors which evolve with time. The model has an extra dimension: the time of rating an item which is not used in our SVR model. Note the analogy between the experience of a user in this model, and the notion of user expertise in the SVR model. However, these models ignore the text of the reviews. 3. Text-based LFM [20]: This model incorporates text in the LFM by combining the latent factors associated to items in LFM with latent topics in text from topic models like LDA. 4. Support Vector Regression (SVR) [4]: We train an SVR model SVRuk for each user uk (refer to Section 3.2) based on her reviews hrj,k i with language and topic features hF L (rj,k )∪F T (rj,k )i, with the user’s article ratings hyj,k i as the response variable. We also incorporate the article language features and the topic features, as well as source-specific features to train the user model for this task. The other models ignore the stylistic features, and other fine-grained user-item interactions in the community. Table 6 shows the MSE comparison between the different methods. Our model (User SVR) achieved the lowest MSE and thus performed best. 5.2 Finding Credible Articles As a second part of the evaluation, we investigate the predictive power of different models in order to find credible news articles based on the aggregated ratings from all users. The above LFM models, unaware of the user cliques, cannot be used directly for this task, as each news article has multiple reviews from different users which need to be aggregated. We find the Mean Squared Error (MSE) between the estimated overall article rating, and the groundtruth article rating. We consider stories with at least 3 ratings about a news article. We compare the CCRF against the following baselines: 1. Support Vector Regression (SVR) [4]: We consider an SVR model with features on language (bag-of-all-words, subjectivity, bias etc.), topics (explicit tags as well as latent dimensions), and news-sourcespecific features. The language model uses all the lexicons derived and used in [28, 23]. The source model also includes topic features in terms of the top topics covered by the source, and its topic-specific expertise for a subset of the topics. 2. Aggregated Model (SVR) [4]: As explained earlier, the user features cannot be directly used in the baseline model, which is agnostic of the user cliques. Therefore, we adopt a simple aggregation approach by taking the average rating of all the user ratings SVRuk (dj ) for an article dj as a feature. Note that, in contrast to this |uk | simple average used here, our CCRF model learns the weights hαu i per-user to combine their overall ratings for an article. Table 7 shows the MSE comparison of the different models. MSE Comparison: The first two models in Table 6 ignore the textual content of news articles, and reviews, and perform worse than the ones that incorporate full text. The text-based LFM considers title and text, and performs better than its predecessors. However, the User SVR model considers richer features and interactions, and attains 23% MSE reduction over the best performing LFM baselines. The baselines in Table 7 show the model performance after incorporating different features in two different settings: 1) with news article titles only as text, and 2) with titles and the first few paragraphs of an article. The language model, especially the bias and subjectivity features, is less effective using only the article titles due to sparseness. On the other hand, using the entire article text may lead to very noisy features. So including the first few paragraphs of an article is the “sweet spot”. For this, we made an ad-hoc decision and included the first 1000 characters of each article. With this setting, the language features made a substantial contribution to reducing the MSE. The aggregated SVR model further brings in the user features, and achieves the lowest MSE among the baselines. This shows that a user-aware credibility model performs better than user-independent ones. Our CCRF model combines all features in a more sophisticated manner, which results in 19.5% MSE reduction over the most competitive baseline (aggregated SVR). This is empirical evidence that the joint interactions between the different factors in a news community are indeed important to consider for identifying highly credible articles. 5.3 Finding Trustworthy Sources We shift the focus to two use cases: 1) identifying the most trustworthy sources, and 2) identifying expert users in the community who can play the role of “citizen journalists”. Using the model of Section 3, we rank all news sources in the community according to the learned hβsi i in Equation 5. The baseline is taken as the PageRank scores of news sources in the Web graph. In the experience-based LFM we can consider the sources to be users, and articles generated by them to be items. This allows us to obtain a ranking of the sources based on their overall authority. This is the second baseline against which we compare the CCRF. We measure the quality of the ranked lists in terms of NDCG using the actual ranking of the news sources in the community as ground-truth. NDCG gives geometrically decreasing weights to predictions at the various positions of the ranked list: N DCGp = DCGp IDCGp where DCGp = rel1 + Pp reli i=2 log2 i Table 8 shows the NDCG scores for the different methods. 5.4 Finding Expert Users Similar to news sources, we rank users according to the learned hαuk i in Equation 5. The baseline is the average rating received by a user from other members in the community. We compute the NDCG score for the ranked lists of users by our method. We also compare against the ranked list of users from the experience-aware LFM [21]. Table 9 shows the NDCG scores for different methods. 5.5 Discussion Hypothesis Testing: We test various hypotheses under the influence of the feature groups using explicit labels, and ratings available in the NewsTrust community. A summary of the tests is presented in Table 10 showing a moderate correlation between various factors which are put together in the CCRF to have a strong indicator for information credibility. Language: The stylistic features (factor (a) in Table 10) like assertives, hedges, implicatives, factives, discourse and affective play a significant role in the credibility detection of news, in conjunction with other language features like topics. Factors Corr. a) Stylistic Indicators Vs. Article Credibility Rating Insightful (Is it well reasoned? thoughtful?) Fairness (Is it impartial? or biased?) Style (Is this story clear? concise? well-written?) Responsibility (Are claims valid, ethical, unbiased?) Balance (Does this story represent diverse viewpoints?) 0.77 0.75 0.65 0.72 0.49 b) Influence of Politics Vs. Disagreement c) Expertise (Moderate, High) Vs. Disagreement Interactions d) User Expertise Vs. User-User Rating e) Source Trustworthiness Vs. Article Credibility Rating f) User Expertise Vs. MSE in Article Rating Prediction 0.11 -0.10, -0.31 0.40 0.47 -0.29 Table 10: Pearson’s product-moment correlation between various factors (with P-value < 0.0001 for each test). Money - Politics War in Iraq Media - Politics Green Technology Most Trusted rollingstone.com nybooks.com consortiumnews discovermagazine.com truthdig.com consortiumnews thenation.com nature.com democracynow.org truthout.org thedailyshow.com scientificamerican.com Least Trusted firedoglake.com suntimes.com trueslant.com crooksandliars rushlimbaugh.com timesonline rightwingnews.com suntimes.com foxnews.com Table 11: Most and least trusted sources on sample topics. Topics: Topics are an important indicator for news credibility. We measured the influence of the Politics tag on other topics by their co-occurrence frequency in the explicit tag sets over all the news articles. We found significant influence of Politics on all topics, with an average measure of association of 54% to any topic, and 62% for the overall news article. The community gets polarized due to different perspectives on topical aspects of news. A moderate correlation (factor (b) in Table 10) indicates a weak trend of disagreement, measured by the standard deviation in article credibility rating among users, increasing with its political content. In general, we find that community disagreement for different viewpoints are as follows: Right (0.80) > Left(0.78) > Center(0.65) > Neutral (0.63). Users: User engagement features are strong indicators of expertise. Although credibility is ultimately subjective, experts show moderate agreement (factor (c) in Table 10) on highly credible news. There is a moderate correlation (factor (d) in Table 10) between feedback received by a user on his ratings from community, and his expertise. Sources: Various traits of a news source like viewpoint, format and topic expertise are strong indicators of trustworthiness. In general, science and technology websites (e.g., discovermagazine.com, nature.com, scientificamerican.com), investigative reporting and non-partisan sources (e.g., truthout.org, truthdig.com, cfr.org), book sites (e.g., nybooks.com, editorandpublisher.com), encyclopedia (e.g., Wikipedia) and fact checking sites (e.g., factcheck.org) rank among the top trusted sources. Table 11 shows the most and least trusted sources on four sample topics. Overall, news sources are considered trustworthy with an average rating of 3.46 and variance of 0.15. Tables 12 and 13 show the most and least trusted sources on different viewpoints and media types respectively. Contents from blogs are most likely to be posted followed by newspaper, magazine and other online sources. Contents from wire service, TV and radio are deemed the most trustworthy, although they have the least subscription, followed by magazines. Interactions: In principle, there is a moderate correlation between trustworthy sources generating credible articles (factor (e) in Table 10) identified by expert users (factor (f) in Table 10). A negative sign of correlation indicates decrease in disagreement or MSE with Left Right Most Trusted democracynow, truthdig.com, rollingstone.com courant.com, opinionjournal.com, townhall.com Least Trusted crooksandliars, suntimes.com, washingtonmonthly.com rightwingnews, foxnews.com, weeklystandard.com Center Neutral armedforcesjournal.com, bostonreview.net spiegel.de,cfr.org, editorandpublisher.com sltrib.com, examiner.com, spectator.org msnbc.msn.com, online.wsj.com, techcrunch.com Table 12: Most and least trusted sources with different viewpoints. Magazine Online Newspaper Blog Most Trusted Sources rollingstone.com nybooks.com thenation.com truthdig.com cfr.org consortiumnews nytimes.com nola.com seattletimes juancole.com dailykos.com huffingtonpost investigativevoice northbaltimore hosted.ap.org suntimes.com nydailynews.com dailymail.co.uk rightwingnews firedoglake.com crooksandliars Least Trusted Sources weeklystandard.com commentarymagazine nationalreview.com Table 13: Most and least trusted sources on different types of media. increase in expertise. In a news community, we can observe moderate signals of interaction between various factors that characterize users, articles, and sources. Our CCRF model brings all these features together to build a strong signal for news credibility. 6. RELATED WORK Rating prediction in online communities: Collaborative filtering based approaches [15] for rating prediction exploit user and item similarities by latent factors. [21] further studies the temporal evolution of users and their rating behavior in this framework. Recent works [20, 22] also tap into user review texts to generate userspecific ratings of reviews. Other papers have studied temporal issues for anomaly detection [10].Prior work that tapped user review texts focused on other issues. Sentiment analysis over reviews aimed to learn latent topics [18], latent aspects and their ratings [35], and user-user interactions [36]. Our model unifies several dimensions to jointly study the role of language, users, topics, and interactions for information credibility. Information credibility in social media: [3] analyzes micro-blog postings in Twitter related to trending topics, and classifies them as credible or not based on features from user posting and re-posting behavior. [14] focuses on credibility of users, harnessing the dynamics of information flow in the underlying social graph and tweet content. [2] analyzes both topical content of information sources and social network structure to find credible information sources in social networks. Information credibility in tweets has been studied in [11]. [33] conducts a user study to analyze various factors like contrasting viewpoints and expertise affecting the truthfulness of controversial claims. However, none of these prior works analyze the interplay between sources, language, topics, and users. The works closest to our problem and approach are [34, 23]. [34] presents an algorithm for propagating trust scores in a heterogeneous network of claims, sources, and documents. [23] proposes a method to jointly learn user trustworthiness, statement credibility, and language objectivity in online health communities. However, these works do not analyze the role of topics, language bias, user perspective, expertise, and fine-grained interactions in community. Bias in social communities and media: The use of biased language in Wikipedia and similar collaborative communities has been studied in [9, 28]. Even more broadly, the task of characterizing subjective language has been addressed, among others, in [37, 19]. The influence of different kinds of bias in online user ratings has been studied in [30, 6]. [6] proposes an approach to handle users who might be subjectively different or strategically dishonest. Citizen journalism: [29] defines citizen journalism as “the act of a citizen or group of citizens playing an active role in the process of collecting, reporting, analyzing and dissemination of news and information to provide independent, reliable, accurate, wide-ranging and relevant information that a democracy requires.” [31] focuses on user activities like blogging in community news websites. Although the potential of citizen journalism is greatly highlighted in the recent Arab Spring [12], misinformation can be quite dangerous when relying on users as news sources (e.g., the reporting of the Boston Bombings in 2013 [25]). 7. CONCLUSIONS In this work, we analyzed the effect of different factors like language, topics and perspectives on the credibility rating of articles in a news community. These factors and their mutual interactions are the features of a novel model for jointly capturing credibility of news articles, trustworthiness of news sources and expertise of users. From an application perspective, we demonstrated that our method can reliably identify credible articles, trustworthy sources and expert users in the community. As future work, we plan to model and analyze the temporal evolution of the factors associated with each of the components in our model. We have a strong intuition that time has a significant influence on the trustworthiness of sources and credibility of news. 8. REFERENCES [1] D. M. Blei, A. Y. Ng, and M. I. Jordan. Latent dirichlet allocation. J. Mach. Learn. Res., 3, 2003. [2] K. R. Canini, B. Suh, and P. Pirolli. Finding credible information sources in social networks based on content and social structure. In PASSAT, 2011. [3] C. Castillo, M. Mendoza, and B. Poblete. Information credibility on twitter. In WWW, 2011. [4] H. Drucker, C. J. C. Burges, L. Kaufman, A. J. Smola, and V. Vapnik. Support vector regression machines. In NIPS, 1996. [5] H. J. Einhorn, R. M. Hogarth, and E. Klempner. Quality of group judgment. Psychological Bulletin, 1977. [6] H. Fang, J. Zhang, and N. Magnenat Thalmann. Subjectivity grouping: Learning from users’ rating behavior. In AAMAS, 2014. [7] B. J. Fogg. Prominence-interpretation theory: explaining how people assess credibility online. In CHI, 2003. [8] Gallup.com. Americans’ confidence in newspapers continues to erode. http://www.gallup.com/poll/163097/americans-confidencenewspapers-continues-erode.aspx. Accessed: 2015-05-07. [9] S. Greene and P. Resnik. More than words: Syntactic packaging and implicit sentiment. In NAACL, 2009. [10] S. Günnemann, N. Günnemann, and C. Faloutsos. Detecting anomalies in dynamic rating data: A robust probabilistic model for rating evolution. KDD, 2014. [11] A. Gupta and P. Kumaraguru. Credibility ranking of tweets during high impact events. In PSOSM, 2012. [12] P. N. Howard, A. Duffy, D. Freelon, M. Hussain, W. Mari, and M. Mazaid. Opening closed regimes: What was the role of social media during the arab spring? 2011. [13] K. Järvelin and J. Kekäläinen. Cumulated gain-based evaluation of IR techniques. ACM Trans. Inf. Syst., 20(4), 2002. [14] B. Kang, J. O’Donovan, and T. Höllerer. Modeling topic specific credibility on twitter. In IUI, 2012. [15] Y. Koren. Factorization meets the neighborhood: A multifaceted collaborative filtering model. KDD, 2008. [16] C. Lampe and R. K. Garrett. It’s all news to me: The effect of instruments on ratings provision. In HICSS, 2007. [17] S. C. Lewis, K. Kaufhold, and D. L. Lasorsa. Thinking about citizen journalism: The philosophical and practical challenges of user-generated content for community newspapers. Journalism Practice, 4(2), 2010. [18] C. Lin and Y. He. Joint sentiment/topic model for sentiment analysis. CIKM, 2009. [19] C. Lin, Y. He, and R. Everson. Sentence subjectivity detection with weakly-supervised learning. In IJCNLP, 2011. [20] J. McAuley and J. Leskovec. Hidden factors and hidden topics: Understanding rating dimensions with review text. RecSys, 2013. [21] J. J. McAuley and J. Leskovec. From amateurs to connoisseurs: modeling the evolution of user expertise through online reviews. In WWW, 2013. [22] S. Mukherjee, G. Basu, and S. Joshi. Joint author sentiment topic model. In SDM, 2014. [23] S. Mukherjee, G. Weikum, and C. Danescu-Niculescu-Mizil. People on drugs: Credibility of user statements in health communities. KDD, 2014. [24] Nber.org. Media bias and voting. http://www.nber.org/digest/oct06/w12169.html. Accessed: 2015-05-07. [25] Nytimes.com. Should reddit be blamed for the spreading of a smear? http://www.nytimes.com/2013/07/28/magazine/should-reddit-beblamed-for-the-spreading-of-a-smear.html. Accessed: 2015-05-07. [26] T. Qin, T. Liu, X. Zhang, D. Wang, and H. Li. Global ranking using continuous conditional random fields. NIPS, 2008. [27] V. Radosavljevic, S. Vucetic, and Z. Obradovic. Continuous conditional random fields for regression in remote sensing. ECAI, 2010. [28] M. Recasens, C. Danescu-Niculescu-Mizil, and D. Jurafsky. Linguistic models for analyzing and detecting biased language. In ACL, 2013. [29] B. Shayne and W. Chris. We media: How audiences are shaping the future of news and information. 2003. [30] Sloanreview.mit.edu. The problem with online ratings. http://sloanreview.mit.edu/article/the-problem-with-online-ratings-2/ . Accessed: 2015-05-07. [31] A. Stuart. Citizen journalism and the rise of ‘mass self-communication’: Reporting the london bombings. Global Media, 1(1), 2007. [32] B. Tadas, P. Robinson, and L. Morency. Continuous conditional neural fields for structured regression. In ECCV, 2014. [33] V. Vydiswaran et al. BiasTrust: Teaching biased users about controversial topics. CIKM, 2012. [34] V. V. Vydiswaran, C. Zhai, and D. Roth. Content-driven trust propagation framework. In KDD, 2011. [35] H. Wang et al. Latent aspect rating analysis without aspect keyword supervision. KDD, 2011. [36] R. West, H. S. Paskov, J. Leskovec, and C. Potts. Exploiting social network structure for person-to-person sentiment analysis. TACL, 2(2), 2014. [37] J. Wiebe and E. Riloff. Creating subjective and objective sentence classifiers from unannotated texts. In CICLing, 2005. 

2cs.AI

Recognizing Abnormal Heart Sounds Using Deep Learning arXiv:1707.04642v2 [cs.SD] 19 Oct 2017 Jonathan Rubin1 , Rui Abreu2 , Anurag Ganguli2 , Saigopal Nelaturi2 , Ion Matei2 , Kumar Sricharan2 1 Philips Research North America, 2 PARC, A Xerox Company jonathan.rubin@philips.com, rui@computer.org, {anurag.ganguli, saigopal.nelaturi, ion.matei, sricharan.kumar}@parc.com Abstract The work presented here applies deep learning to the task of automated cardiac auscultation, i.e. recognizing abnormalities in heart sounds. We describe an automated heart sound classification algorithm that combines the use of time-frequency heat map representations with a deep convolutional neural network (CNN). Given the cost-sensitive nature of misclassification, our CNN architecture is trained using a modified loss function that directly optimizes the trade-off between sensitivity and specificity. We evaluated our algorithm at the 2016 PhysioNet Computing in Cardiology challenge where the objective was to accurately classify normal and abnormal heart sounds from single, short, potentially noisy recordings. Our entry to the challenge achieved a final specificity of 0.95, sensitivity of 0.73 and overall score of 0.84. We achieved the greatest specificity score out of all challenge entries and, using just a single CNN, our algorithm differed in overall score by only 0.02 compared to the top place finisher, which used an ensemble approach. 1 PhysioNet [Goldberger et al., 2000] has held a Computing in Cardiology Challenge since 2000 that requires participants to automatically analyze physiologic time series data. The 2016 challenge [Clifford et al., 2016] asked participants to perform automated analysis of phonocardiogram (PCG) waveforms, i.e. heart sound data collected using digital stethoscopes. The objective of the challenge was to accurately classify normal and abnormal heart sounds. Recordings were collected from both healthy individuals, as well as those with heart disease, including heart valve disease and coronary artery disease. A PCG plot showing the recording of the (normal) sounds made by the heart is given in Figure 1. Introduction Advances in deep learning [LeCun et al., 2015] are being made at a rapid pace, in part due to challenges such as ILSVRC – the ImageNet Large-Scale Visual Recognition Challenge [Russakovsky et al., 2015]. Successive improvements in deep neural network architectures have resulted in computer vision systems that are better able to recognize and classify objects in images [Lin et al., 2013; Szegedy et al., 2015] and winning ILSVRC entries [Szegedy et al., 2014; He et al., 2015]. While a large focus of deep learning has been on automated analysis of image and text data, advances are also increasingly being seen in areas that require processing other input modalities. One such area is the medical domain where inputs into a deep learning system could be physiologic time series data. An increasing number of large scale challenges in the medical domain, such as [Kaggle, 2014] and [Kaggle, 2015] have also resulted in improvements to deep learning architectures [Liang and Hu, 2015]. Figure 1: A phonocardiogram showing the recording of normal heart sounds, together with corresponding electrocardiogram tracing. S1 is the first heart sound and marks the beginning of systole. Source [Springer et al., 2016]. Heart disease remains the leading cause of death globally, resulting in more people dying every year due to cardiovascular disease compared to any other cause of death [World Health Organization, 2017]. Successful automated PCG analysis can serve as a useful diagnostic tool to help determine whether an individual should be referred on for expert diagnosis, particularly in areas where access to clinicians and medical care is limited. In this work, we present an algorithm that accepts PCG waveforms as input and uses a deep convolutional neural network architecture to classify inputs as either normal or abnor- mal using the following steps: 1. Segmentation of time series A logistic regression hidden semi-Markov model is used to segment incoming heart sound instances into shorter segments beginning at the start of each heartbeat, i.e. the S1 heart sound. 2. Transformation of segments into heat maps Using Mel-frequency cepstral coefficients, one dimensional time series input segments are converted into twodimensional spectrograms (heat maps) that capture the time-frequency distribution of signal energy. 3. Classification of heat maps using a deep neural network A convolutional neural network is trained to perform automatic feature extraction and distinguish between normal and abnormal heat maps. The contributions of this work are as follows: 1. We introduce a deep convolutional neural network architecture designed to automatically analyze physiologic time series data for the purposes of identifying abnormalities in heart sounds. 2. Given the cost-sensitive nature of misclassification, we describe a novel loss function used to train the above network that directly optimizes the sensitivity and specificity trade-off. 3. We present results from the 2016 PhysioNet Computing in Cardiology Challenge where we evaluated our algorithm and achieved a Top 10 place finish out of 48 teams who submitted a total of 348 entries. The remainder of this paper is organized as follows. In Section 2, we discuss related works, including historical approaches to automated heart sound analysis and deep learning approaches that process physiologic time series input data. Section 3 introduces our approach and details each step of the algorithm. Section 4 further describes the modified costsensitive loss function used to trade-off the sensitivity and specificity of the network’s predictions, followed by Section 5, which details the network training decisions and parameters. Section 6 presents results from the 2016 PhysioNet Computing in Cardiology Challenge and in Section 7 we provide a final discussion and end with conclusions in Section 8. 2 Related Work Before the 2016 PhysioNet Computing in Cardiology Challenge there were no existing approaches (to the authors’ knowledge) that applied the tools and techniques of “deep learning” to the automated analysis of heart sounds [Liu et al., 2016]. Previous approaches relied upon a combination of feature extraction routines input into classic supervised machine learning classifiers. Features extracted from heart cycles in the time and frequency domains, as well as wavelet features, time-frequency and complexity-based features were input into artificial neural networks [De Vos and Blanckenberg, 2007; Uğuz, 2012a; Uğuz, 2012b; Sepehri et al., 2008; Bhatikar et al., 2005] and support vector machines [Maglogiannis et al., 2009; Ari et al., 2010; Zheng et al., 2015] for classification. Previous works have also employed Hidden Markov Models for both segmenting PCG signals into the fundamental heart sounds [Springer et al., 2014; Springer et al., 2016], as well as classifying normal and abnormal instances [Wang et al., 2007; Saraçoglu, 2012]. While there have been many previous efforts applied to automated heart sound analysis, gauging the success of historical approaches has been somewhat difficult, due to differences in dataset quality, number of recordings available for training and testing algorithms, recorded signal lengths and the environment in which data was collected (e.g. clinical vs. non-clinical settings). Moreover, some existing works have not performed appropriate train-test data splits and have reported results on training or validation data, which is highly likely to produce optimistic results due to overfitting [Liu et al., 2016]. In this work, we report results from the 2016 PhysioNet Computing in Cardiology Challenge, which evaluated entries on a large hidden test-set that was not made publicly available. To reduce overfitting, no recordings from the same subject were included in both the training and the test set and a variety of both clean and noisy PCG recordings, which exhibited very poor signal quality, were included to encourage the development of accurate and robust algorithms. The work presented in this paper, is one of the first attempts at applying deep learning to the task of heart sound data analysis. However, there have been recent efforts to apply deep learning approaches to other types of physiological time series analysis tasks. An early work that applied deep learning to the domain of psychophysiology is described in [Martı́nez et al., 2013]. They advocate the use of preference deep learning for recognizing affect from physiological inputs such as skin conductance and blood volume pulse within a game-based user study. The authors argue against the use of manual ad-hoc feature extraction and selection in affective modeling, as this limits the creativity of attribute design to the researcher. One difference between the work of [Martı́nez et al., 2013] and ours is that they perform an initial unsupervised pre-training step using stacked convolutional denoising auto-encoders, whereas our network does not require this step and is instead trained in a supervised fashion end-to-end. Similar deep learning efforts that process physiologic time series have also been applied to the problems of epileptic seizure prediction [Mirowski et al., 2008] and human activity recognition [Hammerla et al., 2016]. 3 Approach Recall from Section 1 that our approach consists of three general steps: segmentation, transformation and classification. Each is described in detail below. 3.1 Segmentation of time series The main goal of segmentation is to ensure that incoming time series inputs are appropriately aligned before attempting to perform classification. We first segment the incoming heart sound instances into shorter segments and locate the beginning of each heartbeat, i.e. the S1 heart sound. A logistic regression hidden semi-Markov model [Springer et al., 2016] is used to predict the most likely sequence of heart sound states Figure 2: MFCC heat map visualization of a 3-second segment of heart sound data. Sliding windows, i, are represented on the horizontal axis and filterbank frequencies, j, are stacked along the inverted y-axis. MFCC energy information, ci,j is represented by pixel color in the spectrograms. Also shown are the original one-dimensional PCG waveforms that produced each heat map. (S1 → Systole → S2 → Diastole) by incorporating information about expected state durations. Once the S1 heart sound has been identified, a time segment of length, T , is extracted. Segment extraction can either be overlapping or non-overlapping. Our final model used a segment length of, T = 3 seconds, and we chose to use overlapping segments as this led to performance improvements during initial training and validation. 3.2 Transformation of segments into heat maps Each segment is transformed from a one-dimensional time series signal into a two-dimensional heat map that captures the time-frequency distribution of signal energy. We chose to use Mel Frequency Cepstral Coefficents [Davis and Mermelstein, 1980] to perform this transformation, as MFCCs capture features from audio data that more closely resembles how human beings perceive loudness and pitch. MFCCs are commonly used as a feature type in automatic speech recognition [Godino-Llorente and Gomez-Vilda, 2004]. We apply the following steps to achieve the transformation: log transformation as sound volume is not perceived on a linear scale. c∗i,j = log( ci,j = k (1) n=1 where k ∈ [1, K], K is the length of the DFT and h(n) is a hamming window of length N . The power spectral estimate for window, i, is then given by (2). Pi (k) = 1 |Si (k)|2 N (2) 3. Apply a filterbank of, j ∈ [1, J], triangular band-pass filters, dj,1...K , to the power spectral estimates, Pi (k), and sum the energies in each filter together. Include a J X j=1 c∗i,j cos h k(2i − 1)π i ,k = 1...J 2J (4) The result is a collection ofcepstral coefficients, ci,j for  T window, i. For i = 1 . . . ∆ , ci,j can be stacked together to give a time-frequency heat map that captures changes in signal energy over heart sound segments. Figure 2 illustrates two example heat maps (one derived from a normal heart sound input and the other from an abnormal input), where ci,j is the MFCC value (represented by color) at location, i, on the horizontal axis and, j, on the (inverted) vertical axis. 3.3 si (n)h(n)ei2πn `ν (3) We used a filterbank consisting of J = 26 filters, where frequency ranges were derived using the Mel scale that maps actual measured frequencies, f , to values that better match how humans perceive pitch, M (f ) = f 1125 ln(1 + 700 ). 4. Finally, apply a Discrete Cosine Transform to decorrelate the log filterbank energies, which are correlated due to overlapping windows in the Mel filterbank. 2. Compute the Discrete Fourier transform for each window. Si (k) = dj,k Pi (k)) k=1 1. Given an input segment of length, T , and sampling rate, ν, select a window length, `, and step size, ∆, and extract overlapping sliding windows, si (n),  from the input time T series segment, where i ∈ [1, ∆ ] is the window index and n ∈ [1, `ν] is the sample index. We chose a window length of 0.025 seconds and a step size of 0.01 seconds. `ν X K X Classification of heat maps using a deep neural network The result of transforming the original one-dimensional timeseries into a two-dimensional time-frequency representation is that now each heart sound segment can be processed as an image, where energy values over time can be visualized as a heat map (see Figure 2). Convolutional neural networks are a natural choice for training image classifiers, given their ability to automatically learn appropriate convolutional filters. Therefore, we chose to train a convolutional neural network architecture using heat maps as inputs. Decisions about the number of filters to apply and their sizes, as well as how many layers and their types to include in the network were made by a combination of initial manual Figure 3: Convolutional neural network architecture for predicting normal versus abnormal heart sounds using MFCC heat maps as input exploration by the authors, followed by employing a random search over a limited range of network architecture parameters. Figure 3 depicts the network architecture of a convolutional neural network that accepts as input a single channel 6x300 MFCC heat map and outputs a binary classification, predicting whether the input segment represents a normal or abnormal heart sound. The first convolutional layer learns 64 2x20 kernels, using same-padding. This is followed by applying a 1x20 maxpooling filter, using a horizontal stride of 5, which has the effect of reducing each of the 64 feature maps to a dimension of 6x60. A second convolutional layer applies 64 2x10 kernels over the previous layer, once again using same padding. This is again followed by a max-pooling operation using a filter size of 1x4 and a stride of 2, further reducing each feature map to a dimension of 6x30. At this stage in the architecture a flattening operation is applied that unrolls each of the 64 6x30 feature maps into a single dimensional vector of size 11,520. This feature vector is fed into a first fully connected layer consisting of 1024 hidden units, followed by a second layer of 512 hidden units and finally a binary classification output. 4 Sensitivity-Specificity Loss Trade-off The loss function of the network was altered from a standard softmax cross entropy loss function to instead directly tradeoff between sensitivity and specificity. Given unnormalized log-probabilities, y = W x + b, from a classifier consisting of weight matrix, W , and bias b. The softmax function: s(y (1) ) s(y (2) ) .. .    Y =  e s(yi ) = P eyj (5) j gives probability predictions P (yi |x; W, b) for the class at index, i, for input x. Consider, y ∗(1) y ∗(2) .. .     ∗  ,Y =    y ∗(n) s(y (n) )      (j) where s(yi ), refers to the ith entry of row j and Y ∗ is the corresponding one hot encoded matrix of actual class labels. For the binary class labels of normal (y0∗ ) and abnormal ∗ (y1 ), we define the mask matrices, YN n and YAa , where entries within each matrix are softmax prediction values extracted ∀s(y(j) )∈Y , as follows: YN n  (j) ∗(j) s(y0 ), where y0 = 1 and     arg max {s(y (j) )} = arg max{y ∗(j) } =     0, otherwise YAa  (j) ∗(j)  s(y1 ), where y1 = 1 and   arg max {s(y (j) )} = arg max{y ∗(j) } =     0, otherwise We then define softmax sensitivity, Se , and specificity, Sp , as follows: Se = yi  X Y (j) Aa , ∗(j) j YAa Sp = X Y (j) Nn ∗(j) j (6) YN n The final loss function we wish to minimize is given in (7). LSeSp = −(Se + Sp ) + λR(W ) (7) where λR(W ) is a regularization parameter and routine, respectively. Hyper-parameters Learning rate Beta Dropout Network parameters Regularization Type Batch Size Weight Update Value 0.00015822 0.000076253698849 0.85565561 Value L2 256 Adam Optimization Table 1: Listing of hyper-parameters and selected network parameters. Hyper-parameters were learned over the network architecture described in Section 3.3, using random search over a restricted parameter space. 5 Network Training L2 regularization was computed for each of the fully connected layers’ weight and bias matrices and applied to the loss function. Dropout was applied within both fully connected layers. Table 1 shows the values of hyper-parameters chosen by performing a random search through parameter space, as well as a list of other network training choices, including weight updates and use of regularization. Adam optimization [Kingma and Ba, 2014] was used to perform weight updates. Models were trained on a single NVIDIA GPU with between 4 – 6 GB of memory. A mini-batch size of 256 was selected to satisfy the memory constraints of the GPU. 5.1 Given that each model was trained on 3-second MFCC heat map segments, it was necessary to stitch together a collection of predictions to classify a single full instance. The simple strategy of averaging each class’s prediction probability was employed and the class with the greatest probability was selected as the final prediction. 6 Equations (8) and (9) show the modified sensitivity and specificity scoring metrics that were used to assess the submitted entries to the 2016 PhysioNet Computing in Cardiology Challenge [Clifford et al., 2016]. Uppercase symbols reflect the true class label, which could either be (A)bnormal, or (N )ormal. Lowercase symbols refer to a classifier’s predicted output where, once again, a is abnormal, n is normal and q is a prediction of unsure. A subscript of 1 (e.g. Aa1 , N a1 ) refers to heart sound instances that were considered good signal quality by the challenge organizers and a subscript of 2 (e.g. An2 , N n2 ) refers to heart sound instances that were considered poor signal quality by challenge organizers. Finally, the weights used to calculate sensitivity, wa1 and wa2 , capture the percentages of good signal quality and poor signal quality recordings in all abnormal recordings. Correspondingly for specificity, the weights wn1 and wn2 are the proportion of good signal quality and poor signal quality recordings in all normal recordings. Overall, scores are given by Se+Sp . 2 Training/Validation/Test Datasets The overall dataset used within the PhysioNet Computing in Cardiology Challenge was provided by the challenge organizers and consisted of eight heart sound databases collected from seven countries over a period of more than a decade [Clifford et al., 2016]. In total 4,430 recordings were taken from 1,072 subjects, resulting in 30 hours of heart sound recordings. From this total dataset, 1,277 heart sound recordings from 308 subjects were removed to be used as heldout test data for evaluating challenge submissions. The test dataset was not made publicly available and challengers were only allowed to make 15 submissions, in total, to the challenge server to evaluate their models on a small 20% subset of the hidden dataset, before final results were computed. The number of allowed submissions was limited to avoid the issue of participants implicitly overfitting their models on the hidden test dataset. From the 3153 publicly available PCG waveforms supplied by the challenge organizers, the authors set aside a further 301 instances to be used as a local held-out test-set to gauge model performance before making a submission to the challenge server. The remaining instances were used to train initial models. Models were trained on the overlapping 3-second MFCC segments extracted from the remaining 2852 PCG waveforms. This resulted in approximately 90,000 MFCC heat maps, which were split into a training (∼ 75, 000 instances) and validation set (∼ 15, 000 instances). This training and validation set was unbalanced, consisting of approximately 80% normal segments and 20% abnormal segments. Training was performed on the unbalanced dataset and no attempt was made to compensate for this class imbalance. Results Se = Sp = wa1 · Aa1 wa2 · (Aa2 + Aq2 ) + , Aa1 + Aq1 + An1 Aa2 + Aq2 + An2 (8) wn1 · N n1 wn2 · (N n2 + N q2 ) + N a1 + N q1 + N n1 N a2 + N q2 + N n2 (9) Table 2 shows a selected subset of the results for the 2016 PhysioNet Computing in Cardiology Challenge. For each selected entry, sensitivity, specificity and overall scores are shown, as well as the entry’s final ranking and a brief description of its approach. In total, 348 entries were submitted by 48 teams. Our entry, as described by the algorithm presented in this paper, was ranked 8th with a sensitivity of 0.7278 and specificity of 0.9521, giving an overall score of 0.8399. The top entry to the competition achieved sensitivity of 0.9424, specificity of 0.7781 for an overall score of 0.8602. Also included in Table 2 is the result of a benchmark entry that was supplied by the challenge organizers, which ranked 43rd overall, with a sensitivity of 0.6545 and specificity of 0.7569, for an overall score of 0.7057. 7 Discussion Table 2 shows that the overall scores for the top entries to the PhysioNet Computing in Cardiology challenge were very close. In particular, our entry, which achieved an 8th place ranking, had a difference in score of only 0.02, compared to the top place finisher. For our entry, the overall score of 0.8399 was achieved using a single convolutional neural network, whereas other top place finishers achieved strong classification accuracies using an ensemble of classifiers. Improvements in performance have often been witnessed using Rank 1 2 3 4 5 6 7 8 9 10 43 48 Sensitivity 0.9424 0.8691 0.8743 0.8639 0.8848 0.8063 0.7696 0.7278 0.8691 0.7696 0.6545 0.8063 Specificity 0.7781 0.849 0.8297 0.8269 0.8048 0.8766 0.9125 0.9521 0.7873 0.8831 0.7569 0.2643 Overall 0.8602 0.859 0.852 0.8454 0.8448 0.8415 0.8411 0.8399 0.8282 0.8263 0.7057 0.5353 Description AdaBoost & CNN Ensemble of SVMs Regularized Neural Networks MFCCs, Wavelets, Tensors & kNN Random Forest + LogitBoost Unofficial entry Probability-distribution based Our Approach (see Section 3) Approach Unknown Approach Unknown Provided Benchmark Entry Approach Unknown Table 2: Selected results from the 2016 PhysioNet Computing in Cardiology Challenge an ensemble of networks or separate classifiers and we leave this for future work/improvement. For practical purposes, a diagnostic tool that relies on only a single network, as opposed to a large ensemble, has the advantage of limiting the amount of computational resources required for classification. Deployment of such a diagnostic tool on platforms that impose restricted computational budgets, e.g mobile-based, could perhaps benefit from such a trade-off between accuracy and computational cost. Another point of interest is that our entry to the PhysioNet Computing in Cardiology challenge achieved the greatest specificity score (0.9521) out of all challenge entries. However, the network architecture produced a lower sensitivity score (0.7278). Once again, considering the practical result of deploying a diagnostic tool that relied upon our algorithm, this would likely result in a system with few false positives, but at the expense of misclassifying some abnormal instances. Final decisions about the trade-off between sensitivity and specificity would require further consideration of the exact conditions and context of the deployment environment. A final point of discussion and area of future improvement is that the approach presented was limited to binary decision outputs, i.e. either normal or abnormal heart sounds. An architecture that also considered signal quality as an output would likely result in performance improvement. 8 Conclusion The work presented here is one of the first to apply deep convolutional neural networks to the task of automated heart sound classification for recognizing normal and abnormal heart sounds. We have presented a novel algorithm that combines a CNN architecture with MFCC heat maps that capture the time-frequency distribution of signal energy. The network was trained to automatically distinguish between normal and abnormal heat map inputs and it was designed to optimize a loss function that directly considers the trade-off between sensitivity and specificity. We evaluated the approach by submitting our algorithm as an entry to the 2016 PhysioNet Computing in Cardiology Challenge. The challenge required the creation of accurate and robust algorithms that could deal with heart sounds that exhibit very poor signal quality. Overall, our entry to the challenge achieved a Top-10 place finish out of 48 teams who submitted 348 entries. Moreover, using just a single CNN, our algorithm differed by a score of at most 0.02 compared to other top place finishers, all of which used an ensemble approach of some kind. References [Ari et al., 2010] Samit Ari, Koushik Hembram, and Goutam Saha. Detection of cardiac abnormality from pcg signal using lms based least square svm classifier. Expert Systems with Applications, 37(12):8019–8026, 2010. [Bhatikar et al., 2005] Sanjay R Bhatikar, Curt DeGroff, and Roop L Mahajan. A classifier based on the artificial neural network approach for cardiologic auscultation in pediatrics. Artificial intelligence in medicine, 33(3):251–260, 2005. [Clifford et al., 2016] Gari D Clifford, CY Liu, Benjamin Moody, David Springer, Ikaro Silva, Qiao Li, and Roger G Mark. Classification of normal/abnormal heart sound recordings: the physionet/computing in cardiology challenge 2016. Computing in Cardiology, pages 609–12, 2016. [Davis and Mermelstein, 1980] Steven Davis and Paul Mermelstein. Comparison of parametric representations for monosyllabic word recognition in continuously spoken sentences. IEEE transactions on acoustics, speech, and signal processing, 28(4):357–366, 1980. [De Vos and Blanckenberg, 2007] Jacques P De Vos and Mike M Blanckenberg. Automated pediatric cardiac auscultation. IEEE Transactions on Biomedical Engineering, 54(2):244–252, 2007. [Godino-Llorente and Gomez-Vilda, 2004] Juan Ignacio Godino-Llorente and P Gomez-Vilda. Automatic detection of voice impairments by means of short-term cepstral parameters and neural network based detectors. IEEE Transactions on Biomedical Engineering, 51(2):380–384, 2004. [Goldberger et al., 2000] A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. Ch. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, and H. E. Stanley. PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals. Circulation, 101(23):e215–e220, 2000. [Hammerla et al., 2016] Nils Y. Hammerla, Shane Halloran, and Thomas Plötz. Deep, convolutional, and recurrent models for human activity recognition using wearables. In Subbarao Kambhampati, editor, Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence, IJCAI 2016, New York, NY, USA, 9-15 July 2016, pages 1533–1540. IJCAI/AAAI Press, 2016. [He et al., 2015] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CoRR, abs/1512.03385, 2015. [Kaggle, 2014] Kaggle. American epilepsy society seizure prediction challenge. https://www.kaggle.com/ c/seizure-prediction, 2014. [Online; accessed 22-December-2016]. [Kaggle, 2015] Kaggle. Grasp-and-Lift EEG Detection. https://www.kaggle.com/c/ grasp-and-lift-eeg-detection, 2015. [Online; accessed 22-December-2016]. [Kingma and Ba, 2014] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. [LeCun et al., 2015] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436– 444, 2015. [Liang and Hu, 2015] Ming Liang and Xiaolin Hu. Recurrent convolutional neural network for object recognition. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, pages 3367–3375. IEEE Computer Society, 2015. [Lin et al., 2013] Min Lin, Qiang Chen, and Shuicheng Yan. Network in network. CoRR, abs/1312.4400, 2013. [Liu et al., 2016] Chengyu Liu, David Springer, Qiao Li, Benjamin Moody, Ricardo Abad Juan, Francisco J Chorro, Francisco Castells, José Millet Roig, Ikaro Silva, Alistair EW Johnson, et al. An open access database for the evaluation of heart sound algorithms. Physiological Measurement, 37(12):2181, 2016. [Maglogiannis et al., 2009] Ilias Maglogiannis, Euripidis Loukis, Elias Zafiropoulos, and Antonis Stasis. Support vectors machine-based identification of heart valve diseases using heart sounds. Computer methods and programs in biomedicine, 95(1):47–61, 2009. [Martı́nez et al., 2013] Héctor Perez Martı́nez, Yoshua Bengio, and Georgios N. Yannakakis. Learning deep physiological models of affect. IEEE Comp. Int. Mag., 8(2):20– 33, 2013. [Mirowski et al., 2008] Piotr W Mirowski, Yann LeCun, Deepak Madhavan, and Ruben Kuzniecky. Comparing svm and convolutional networks for epileptic seizure prediction from intracranial eeg. In 2008 IEEE Workshop on Machine Learning for Signal Processing, pages 244–249. IEEE, 2008. [Russakovsky et al., 2015] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115(3):211– 252, 2015. [Saraçoglu, 2012] Ridvan Saraçoglu. Hidden markov modelbased classification of heart valve disease with PCA for dimension reduction. Eng. Appl. of AI, 25(7):1523–1528, 2012. [Sepehri et al., 2008] Amir A Sepehri, Joel Hancq, Thierry Dutoit, Arash Gharehbaghi, Armen Kocharian, and A Kiani. Computerized screening of children congenital heart diseases. Computer methods and programs in biomedicine, 92(2):186–192, 2008. [Springer et al., 2014] David B Springer, Lionel Tarassenko, and Gari D Clifford. Support vector machine hidden semimarkov model-based heart sound segmentation. In Computing in Cardiology 2014, pages 625–628. IEEE, 2014. [Springer et al., 2016] David B Springer, Lionel Tarassenko, and Gari D Clifford. Logistic regression-hsmm-based heart sound segmentation. IEEE Transactions on Biomedical Engineering, 63(4):822–832, 2016. [Szegedy et al., 2014] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. CoRR, abs/1409.4842, 2014. [Szegedy et al., 2015] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. CoRR, abs/1512.00567, 2015. [Uğuz, 2012a] Harun Uğuz. Adaptive neuro-fuzzy inference system for diagnosis of the heart valve diseases using wavelet transform with entropy. Neural Computing and applications, 21(7):1617–1628, 2012. [Uğuz, 2012b] Harun Uğuz. A biomedical system based on artificial neural network and principal component analysis for diagnosis of the heart valve diseases. Journal of medical systems, 36(1):61–72, 2012. [Wang et al., 2007] Ping Wang, Chu Sing Lim, Sunita Chauhan, Jong Yong A Foo, and Venkataraman Anantharaman. Phonocardiographic signal analysis method using a modified hidden markov model. Annals of Biomedical Engineering, 35(3):367–374, 2007. [World Health Organization, 2017] World Health Organization. Cardiovascular diseases (cvds). http://who. int/mediacentre/factsheets/fs317/en/, 2017. [Online; accessed 01-Febuary-2017]. [Zheng et al., 2015] Yineng Zheng, Xingming Guo, and Xiaorong Ding. A novel hybrid energy fraction and entropybased approach for systolic heart murmurs identification. Expert Systems with Applications, 42(5):2710–2721, 2015. 

1cs.CV

Shaping Influence and Influencing Shaping: A Computational Red Teaming Trust-based Swarm Intelligence Model Jiangjun Tang1 , Eleni Petraki2 , and Hussein Abbass1 arXiv:1802.09647v1 [cs.AI] 26 Feb 2018 1 University of New South Wales, School of Engineering and Information Technology, Canberra, ACT 2600, Australia?? . 2 Faculty of Science, Technology, Education, and Mathematics, University of Canberra Canberra, Australia. j.tang@adfa.edu.au,Eleni.Petraki@canberra.edu.au,h.abbass@adfa.edu.au Abstract. Sociotechnical systems are complex systems, where nonlinear interaction among different players can obscure causal relationships. The absence of mechanisms to help us understand how to create a change in the system makes it hard to manage these systems. Influencing and shaping are social operators acting on sociotechnical systems to design a change. However, the two operators are usually discussed in an ad-hoc manner, without proper guiding models and metrics which assist in adopting these models successfully. Moreover, both social operators rely on accurate understanding of the concept of trust. Without such understanding, neither of these operators can create the required level to create a change in a desirable direction. In this paper, we define these concepts in a concise manner suitable for modelling the concepts and understanding their dynamics. We then introduce a model for influencing and shaping and use Computational Red Teaming principles to design and demonstrate how this model operates. We validate the results computationally through a simulation environment to show social influencing and shaping in an artificial society. Keywords: Influence, Shaping, Trust, Boids 1 Introduction Recently, computational social scientists are attracted to studying means for measuring the concepts of influence and shaping. For influence to work is to exert a form of social power. Servi and Elson [1] introduce a new definition of influence which they apply to online contexts as ‘the capacity to shift the patterns of emotion levels expressed by social media users’. They propose that measuring influence entails identifying shifts in users’ emotion levels followed by the examination of the extent to which these shifts can be connected with a user. ?? Portions of this work was funded by the Australian Research Council Discovery Grant number DP140102590. However, if the process of influence creates shifts in patterns of emotions which can be detected in the short-term, can a persistent application of influencing operators create a long-term shift (i.e. shaping)? Shmueli et.al. [2] discuss computational tools to measure processes for shaping and affecting human behaviour in real life scenarios. Trust was identified as a means to influence humans in a social system. Moreover, trust was found to have a significant impact on social persuasion. Trust is a complex psychological and social concept. A review of the concept can be found in [3]. Larson et.al. [4] are among a few to imply a distinction between influence and shaping, whereby shaping is perceived to be a change to the organization or the environment, while influence fosters attitudes, behaviours or decisions of individuals or groups. However, the majority of the literature follows a tendency to assume that social influencing would lead to shaping. In this paper, we aim to distil subtle differences to distinguish between the two concepts. This distinction is very important for a number of reasons. First, it examines the validity of the implicit assumption that influencing is a sufficient condition for shaping. Second, it is important when studying social sciences using computational models (computational social sciences) to create models that are not ambiguous about the social and psychological phenomena under investigation. Third, it is vital to make it explicit that social influencing and shaping work on different time scales; that is, social influencing is effective in the short run, while shaping requires time and is more effective in the long run. We will use a computational red teaming (CRT) model, whereby a red agent acts on a blue team to influence, shape and sometimes distract the blue team. The idea of the model should not be seen from a competition or conflict perspective. The model is general, were the red agent can be an agent that promotes a positive attitude within a team (a servant leader) or a social worker correcting the social attitude of a gang. 2 Influence, Shaping, and Trust Influence will be defined in this paper as: an operation which causes a short-term effect in the attitude or behaviour of an individual, group or an organization. Shaping, on the other hand, is defined as: an operation which causes a long-term effect in the attitude or behaviour of an individual, group or an organization. We use the more accurate term “effect” rather than the term “change” because sometimes social influence and shaping need to operate to maintain the status quo. If agent A is attempting to influence agent B by changing B’s behaviour, agent C can attempt to counteract agent A’s influence by influencing B to maintain its behaviour. Therefore, influence does not necessarily require a change to occur. In a strict mathematical sense, social influence would change the parameters of a model, while social shaping would alter the constraint system. To illustrate the difference, we will use a model, whereby a group of blue agents attempts to follow a blue leader. A red agent has a self-interest to influence or shape the blue team. All agents are connected through a network. Each agent, excluding the blue leader and the red agent, attempts to align their behaviour with their neighbours (the other agents it is connected to). The blue leader attempts to reach the blue goal (a position in space). When all agents fully trust each other, and in the absence of the red agent’s effect, it is expected that the intention of the blue leader will propagate throughout the network. Over time, the blue agents will also move towards the blue goal. The red agent has a different goal. It aligns with the agents it is connected to, but it also attempts to influence and/or shape them towards its own goal (or away from the blue’s goal). Social influence by red is represented through changing red movements; thus, affecting the movements of its neighbours. Social shaping by red is represented through a network re-wiring mechanism. Connections in a network are the constraints on the network’s topology. By rewiring the network, the red agent changes the constraints system. We abstract trust to a scale between -1 and 1, whereby “1” implies maximum trust, while “-1” implies maximum distrust. We do not differentiate in this paper between distrust and mistrust. A “0” value is a neutral indicator that is equivalent to not knowing a person. 3 The Model An agent-based Boids [5] model is proposed in this paper. All agents are randomly initialized with random headings and locations. Agents are connected through a network structure that allows information exchange among the agents. In this setup, the neighborhood is mostly defined by the hamming distance between two agents in the network, while sometimes it will be defined by the proximity of one agent to another in the physical space. This latter definition is the classic and default one used in the Boids model. Three Boids rules: cohesion, alignment, and separation, are still adopted here. However, the first two vectors are sensed by network connections while the separation vector is perceived through the Euclidean distance in the physical space. Each agent has a trust factor value which decides how much this agent trusts the information perceived from others. The first two vectors are scaled using the trust factor before an agent’s velocity gets updated. An agent 100% trusts the cohesion and alignment information from its linked neighbours when it has a trust factor of 1. When the trust factor is -1, the agent totally believe that the information is deliberately altered to the opposite value, and therefore, the agent reverses the information it receives. In the model, there are three types of agents: blue leader (AB ), red agent (AR ), and blue agent. The blue leader always moves towards a specific location/goal, and attempts to make the other blue agents follow it. The blue agent is an agent that senses its neighbours through network links for both cohesion and alignment but by Euclidean distance for separation, and then makes decisions on its new velocity. The red agent is a special agent in the model who controls the level of noise (η) in the velocity and network connections for influ- encing and shaping. Many blue agents can exist but there is only a single blue leader and a single red agent. Agents form a set A and live in a space (S) defined by a given width (spaceW ) and a given length (spaceL). All agents are connected by a random network. To establish network connections, a probability (p) is defined. If we have n agents including one blue leader, one red agent, and n − 2 blue agents, the network can be denoted as G(n, p). A Goal (G) is a 2-D position that sits at one of the corners of S. Blue leader always aims to move towards G. The area surrounding of G is denoted by δ. Once the blue leader enters this area, the position of G changes. An agent has the following common attributes: – Position (p), p ∈ S, is a 2-D coordinate. – Velocity (v) is a 2-D vector representing the agent’s movement (heading and speed) in a time unit. – Cohesion Velocity (cohesionV ) of an agent is the velocity calculated based on the mass of all agents that are connected to this agent. – Alignment Velocity (alignmentV ) of an agent is the velocity calculated based on the average velocity of all agents that are connected to this agent. – Separation Velocity (separationV ) of an agent is the velocity that forces this agent to keep a certain small distance from its neighbors and is based on the Euclidean distance. – Velocity weights: • Cohesion weight (wc ): a scaler for the cohesion velocity. • Alignment weight (wa ): a scaler for the alignment velocity. • Separation weight (ws ): a scaler for the separation velocity. – Trust factor (τ ) defines how much an agent trusts its connected neighbours. It has an impact on both the cohesion velocity and alignment velocity but not on the separation velocity. All agents except the blue leader attempt moving towards the neighbours’ location guided with the cohesion vector. The cohesion vector,cohesionVi , of an agent Ai is: P|N | j=0 pj − pi (1) cohesionVi = |N | where, |N | is the cardinality of the neighbourhood N . The alignment velocity of an agent with its linked neighbours is: P|N | alignmentVi = j=0 |N | vj − vi (2) The separation velocity of an agent is calculated using neighbours Nd in the spatial proximity of other agents as follows: |Nd | separationVi = − X j=0 (pj − pi ) (3) The trust factor of a blue agent is updated by the average trust factors of all its connected neighbours (N ) as below: P|N | τi = 0.5 × (τi + j=0 τj |N | ) (4) The blue leader and red agent’s trust factors are not updated. The velocity of the blue leader always aims at the goal G at each step and it is not affected by any other factor. The velocities at time t of all agents except the blue leader are updated by Equation 5. v = v + τ × (wc × cohesionV + wa × alignmentV ) + ws × separationV (5) where, cohesionV , AlignmentV , and separationV are normalized vectors. The position at time t of each agent can be updated by: p = p + vt (6) If an agent’s new position is outside the bounds of S, the reflection rule is applied. According to Equation 5, an agent adjusts its own velocity in compliance with both cohesionV and alignmentV when it has a positive trust value, and follows the opposite direction as suggested by cohesionV and alignmentV when its trust factor is negative. If τ = 0, only the separation vector takes effect on this agent so that this agent doesn’t anyone. The red agent introduces heading noise, changes its network structure, or does both at each time step. The heading noise can be propagated to blue agents through the connections of the network to cause a deviation in some blue agents’ moving directions. Changes in the network structure may result in long term effects on blue agents. At each time step, Red agent updates its own velocity (vRedAgent ) by Equation 5 and then Equation 7 uses a normal distribution (N (0, η)) to generate noise and add it to Red’s velocity. vRedAgent = vRedAgent + N (0, η) (7) Equation 6 is used to update the red agent’s position. Furthermore, the red agent has the ability to re-configure network connections by using the noise level η as a probability that governs the eventuality of the following steps: 1. 2. 3. 4. Randomly pick up a blue agent (Ai ) who is connected with the red agent. Randomly pick up another blue agent (Aj ) who is connected with Ai . Break the connection between Ai and Aj . Connect the red agent with a randomly chosen blue agent Aj . In this way, the connection between the red agent and blue agents changes but the number of edges of the whole network remains as before. The long term effects of these topological updates are expected because the path along with information propagates changes and some blue agents may not get consistent updates from their neighbours. The blue leader attempts to lead other blue agents towards a given destination, and the red agent attempts to disorient through influence (deliberate changes in heading) and/or shaping (deliberate re-configuration of network structure). Therefore, the “effect” from our model can be derived as how well the blue agents follow the blue leader given the influence/shaping by the red agent. A straightforward measure of this effect within our model is the average distance between blue agents and the goal when the blue leader reaches the goal. If this distance is small, blue agents followed the blue leader. If it is large, red agent distracted the blue agent. During a single simulation run, the blue leader is tasked to reach the goal multiple times. Each time it reaches the goal (an iteration), the location of the goal changes. The effect is measured at the end of each iteration. The overall effect of a simulation run is the average of all iterations except the first iteration, which is excluded to eliminate the warm-up period in the system resultant from the random initialisation of agents. In summary, the effect is defined by Equation 8. ! n M X 1X 1 ¯ dm,i (8) d= M m=1 n i=1 where, M is the number of iterations except the first one, n is the number of blue agents, and dm,i is the distance between agent i and the goal location at the m’th iteration. 4 Experimental Design Our aim is to evaluate the “effects” of the short term inference and long term shaping caused by the red agent. Two stages are used for the experiments. The first stage focuses on the noise of red agent where the effect from the trust factor is minimised. The second stage investigates both the trust factor and the red agent’s noise. The number of blue agents is 25, so there are a total of 27 agents including a blue leader and a red agent. All agents’ initial locations are uniformly initialised at random within S, where S is a square space with the dimension of 500×500. All agents’ headings are uniformly initialised at random with constant speed of 1. All agents except the blue leader have the same velocity weights: wc = 0.4 for cohesion, wa = 0.4 for alignment, and ws = 0.2 for separation. The initial trust factor values of all blue agents are uniformly assigned at random within the range of [−1, 1]. Connections among agents are created by a random network G(n, 0.1), where n = 27. In all experiments, two levels of noise (η − = 0.1 and η + = 0.9) are used. In the first stage, to reduce the effect of the trust factor, it is assumed constant with a value of 1 for all agents; that is, all blue agents trust any perceived information, including the information arriving from red. In the second stage, the blue leader has two trust levels: τB− = 0.2 and τB+ = 1, and the red agent has two levels of trust: τR− = −0.2 and τR+ = −1. Three scenarios are designed for investigating the red agent’s impact in our experiments. In Scenario 1, the red agent introduces noise to its heading at each time step thus this noise can immediately affect direct neighbours and can be propagated through the network. In Scenario 2, the red agent changes the network structure at each time step thus shaping the environment of the blue agents. In Scenario 3, the red agent introduces noises to its heading and changes network structures at each time step, so that both influence and shaping take place in our model. Using 2k factorial design [6], a total of 2 factor combinations is available at the first stage and 8 at the second stage. Each combination has three individual scenarios to study. Moreover, the randomness exists in the initialisation phase, therefore 10 runs for each factor combination and impact are desired in order to obtain meaningful results. In summary, there are 60 (3 × 2 × 10) simulation runs in the first stage and 240 (3 × 8 × 10) runs in the second stage. The results and analysis are provided in the next section. 5 Results and Discussion The results from the first stage experiments are presented in Table 1. The dis¯ is listed from the tance between blue agents and goal location of each run (d) second column to the eleventh column. And the last three columns of the table are the averages of 10 runs, the standard deviations and the confidence intervals that are obtained at α = 0.05. Table 1: Results of red agent’s noise impact when τB = 1 and τR = 1 R1 R2 η = 0.1 47.64 46.78 η = 0.9 145.90 155.75 eη 98.27 108.97 η = 0.1 45.71 59.28 η = 0.9 61.23 57.63 eη 15.52 -1.65 η = 0.1 45.34 47.09 η = 0.9 213.49 168.69 eη 168.15 121.59 R3 R4 R5 R6 R7 R8 R9 R10 Scenario1: Velocity Noise 39.26 56.63 47.09 67.29 60.65 38.76 42.99 44.86 168.04 199.94 171.94 243.61 162.15 144.08 103.82 117.94 128.78 143.31 124.85 176.33 101.50 105.33 60.83 73.08 Scenario 2: Network Changes 47.39 54.31 58.14 69.65 50.27 44.35 43.90 48.83 56.30 81.25 53.65 74.69 55.76 40.86 47.74 52.03 8.91 26.94 -4.49 5.04 5.49 -3.49 3.85 3.20 Scenario 3: Velocity Noise and Network Changes 65.90 54.05 51.93 84.91 54.66 41.11 43.88 52.21 197.52 188.80 171.62 236.93 174.46 183.98 84.95 122.82 131.62 134.75 119.69 152.02 119.80 142.87 41.07 70.61 Avg STD Conf. 49.19 8.90 6.36 161.32 37.65 26.93 112.12 31.70 22.68 52.18 7.78 58.11 11.36 5.93 9.00 5.56 8.13 6.44 54.11 12.23 8.75 174.32 41.20 29.47 120.22 35.90 25.68 The results show that the more noise the red agent has in its velocity, the more deviation from the goal observed by the blue agents. Changes in the network structure can lower blue agents performance, although the magnitude of this decrease may not be significant. This is expected since the shaping operates work on a smaller timescale than the influence operator. When both influence and shaping work together, the effect is more profound than any of the individual cases in isolation. Table 2: Results of effects from red agent’s noise impact and trust factors. The confidence level is at 0.05. Effect R1 R2 eτ B eτ R eN -170.40 -146.54 165.32 160.08 1.78 15.84 eτ B eτ R eN -122.99 -165.55 142.72 177.17 42.90 1.75 eτ B eτ R eN -151.65 -140.59 157.63 152.23 2.27 22.60 R3 R4 R5 R6 R7 R8 R9 R10 Scenario 1: Velocity Noise -83.92 -55.00 -131.83 -6.05 -128.09 -110.66 -184.81 -152.86 95.83 71.21 149.41 8.11 133.20 111.97 167.79 160.23 -9.18 6.22 6.74 3.40 -13.50 0.35 -14.64 -17.58 Scenario 2: Network Changes -144.34 -64.94 -168.15 -8.09 -154.27 -170.61 -189.59 -187.66 154.10 77.45 186.62 24.03 164.25 172.21 171.13 172.02 12.81 19.08 8.50 -1.08 -0.29 -12.17 25.41 -15.03 Scenario 3: Velocity Noise and Network Changes -132.44 -35.97 -166.98 -7.63 -159.37 -171.89 -194.38 -163.85 147.97 38.06 176.90 16.03 175.96 163.48 171.95 174.82 15.16 14.70 21.23 4.64 5.73 10.22 20.50 8.15 Avg STD Conf. -117.02 55.01 39.35 122.31 51.75 37.02 -2.06 11.04 7.90 -137.62 58.31 41.71 144.17 52.25 37.38 8.19 17.62 12.61 -132.47 61.12 43.72 137.50 59.31 42.43 12.52 7.38 5.28 The results from the second stage are summarised in Table 2. Interestingly, the trust factors of the blue leader and red agent become critical but the red agent’s noise is not critical. The model responses to the trust factor as expected. When the blue leader has a higher level of trust, all blue agents follow it better (smaller effect values can be observed from eτB ). On the other hand, the blue agents demonstrate disorder behaviours (larger value of eτR ) if red agents have small negative trust. These situations are found in all three scenarios and the effects of trust factors are all statistically significant. Although the red agent’s noise has some effects on blue agents, it is very little and can be ignored when compared to the effect of the trust factor. Negative trust values taken by the blue agents counteract the influence generated by both blue and red agents. The red agent’s noise can have impact on blue agents’ behaviours through short term influence (velocity) and long term shaping (network structures) if the effects of trust are low. When the trust factors are high, the situation changes. Trust has a significant impact on blue agents’ behaviours. Figure 1 illustrates the agents’ footprints when the red agent impacts velocity, network structure or both but with minimum trust effects. These footprints are obtained from the first runs of all three scenarios in the first stage and the results are listed in the column “R1” of Table 1. Figure 1a, 1b, and 1c show that the blue leader leads other agents towards the goal well as being demonstrated by a few congested trajectories. When noise increases, blue agents’ trajectories are disturbed as shown in Figure 1d. Figure 1e shows that changes in the network structure seem to not generate much effects on blue agents’ behaviours. However, the blue agents behaviours are more random when red affects both velocity and network structure. This manifests disorderliness as more scattered blue agents’ footprints can be observed in the last figure. Figure 2 shows two examples of agents’ footprints that are affected by trust with small noise values (η = 0.1). The footprints presented in Figure 2a are extracted from the first run of the third scenario in the second stage with τB = 0.2 and τR = −1. When the red agent’s trust is -1, the negative effect on blue 500 45 500 55 500 450 40 450 50 450 400 35 45 400 80 70 400 60 40 350 350 350 30 35 300 300 25 250 30 250 40 250 25 20 200 50 300 200 200 30 20 15 150 150 150 15 10 100 5 50 100 200 300 400 500 100 5 100 200 300 400 50 500 100 (b) Scenario 2: η = 0.1 25 450 20 400 20 100 10 0 (a) Scenario 1: η = 0.1 500 10 50 200 300 400 500 0 (c) Scenario 3: η = 0.1 500 55 500 20 450 50 450 18 400 16 350 14 300 12 250 10 200 8 150 6 100 4 45 400 40 350 350 35 15 300 300 30 250 250 25 10 200 200 20 150 150 15 5 100 50 100 10 50 100 200 300 400 500 (d) Scenario 1: η = 0.9 5 100 200 300 400 500 (e) Scenario 2: η = 0.9 2 50 100 200 300 400 500 0 (f) Scenario 3: η = 0.9 Fig. 1: Agents’ footprints under Red agent’s noise (η) impacts on velocity and network with minimum trust effects (τB = 1 and τR = 1). agents’ trust is continuously broadcasted throughout the network. Eventually, all blue agents will have a negative trust value that is close to -1 since the blue leader doesn’t have much power (τB = 0.2) to compete against the red agent. This results in all blue agents distrusting each other. In this case, the blue agents spread out to the boundaries. However, the reflection rule forces them back into the given space, causing the blue agents to move around the corners after several time steps as shown in Figure 2a. The right side of Figure 2 depicts agents’ footprints extracted from the third scenario in the second stage with τB = 1, τR = −0.2, and η = 0.1. Some trajectory patterns can be observed from Figure 2b. In this case, the blue leader has enough power to beat the red agent in terms of trust. All blue agents will have positive trust that are passed from the blue leader. Although the red agent has influence on their velocity and connections, the blue agents are still capable to follow the blue leader to reach the goal locations (corners) as the trajectory patterns show. From the above examples and previous results, it can be concluded that trust has a more significant impact on blue agents behaviours than the effect of noise caused by the red agent. 6 Conclusion In this paper, we presented a CRT trust-based model which is an extension of the classic Boids. The network topologies for situation awareness and a trust 500 500 500 10 450 450 450 9 400 400 400 8 350 350 350 7 300 300 300 6 250 250 250 5 200 200 200 4 150 150 150 3 100 100 100 2 50 1 50 50 100 200 300 400 500 0 (a) τB = 0.2, τR = −1, and η = 0.1 100 200 300 400 500 0 (b) τB = 1, τR = 0.2, and η = 0.1 Fig. 2: Trust effects on agents behaviours with red agent noise level at 0.1 in Scenario 3. factor on perceived information are introduced into our model. They provide the necessary tools to investigate influence and shaping using CRT. A number of experiments are designed and conducted in order to differentiate the potential impact from influence and shaping on a system. As the results of the first experimental stage suggest, short term influence can have an immediate effect on the system which is easily observed. The long term shaping effects may not be easily observable although it has effect on the system, especially when it interacts with influence. However, trust among agents plays a critical role in the model. Based on our findings in the second experiment, trust dominates the agents’ behaviours regardless of noise. Acknowledgement This is a pre-print of an article published in International Conference in Swarm Intelligence, Springer, Cham, 2016. The final publication is available at https: //doi.org/10.1007/978-3-319-41000-5_2. References 1. Les Servi and Sara Beth Elson. A mathematical approach to gauging influence by identifying shifts in the emotions of social media users. Computational Social Systems, IEEE Transactions on, 1(4):180–190, 2014. 2. Erez Shmueli, Vivek K Singh, Bruno Lepri, and Alex Pentland. Sensing, understanding, and shaping social behavior. Computational Social Systems, IEEE Transactions on, 1(1):22–34, 2014. 3. Eleni Petraki and Hussein Abbass. On trust and influence: A computational red teaming game theoretic perspective. In Computational Intelligence for Security and Defense Applications (CISDA), 2014 Seventh IEEE Symposium on, pages 1–7. IEEE, 2014. 4. Eric V Larson, Richard E Darilek, Daniel Gibran, Brian Nichiporuk, Amy Richardson, Lowell H Schwartz, and Cathryn Q Thurston. Foundations of effective influence operations: A framework for enhancing army capabilities. Technical report, DTIC Document, 2009. 5. Craig W Reynolds. Flocks, herds and schools: A distributed behavioral model. In ACM SIGGRAPH computer graphics, volume 21, pages 25–34. ACM, 1987. 6. Douglas C Montgomery. Design and analysis of experiments. John Wiley & Sons, 2008. 

2cs.AI

arXiv:1705.03510v1 [math.PR] 9 May 2017 The middle-scale asymptotics of Wishart matrices Didier Chételat and Martin T. Wells Department of Applied Mathematics and Industrial Engineering École Polytechnique Montréal, Québec H3T 1J4 Canada e-mail: didier.chetelat@polymtl.ca Department of Statistical Science Cornell University 1198 Comstock Hall Ithaca, NY 14853-3801 USA e-mail: mtw1@cornell.edu Abstract: We study the behavior of a real p-dimensional Wishart random matrix with n degrees of freedom when n, p Ñ 8 but p{n Ñ 0. We establish the existence of phase transitions when p grows at the order npK`1q{pK`3q for every k P N, and derive expressions for approximating densities between every two phase transitions. To do this, we make use of a novel tool we call the G-transform of a distribution, which is closely related to the characteristic function. We also derive an extension of the t-distribution to the real symmetric matrices, which naturally appears as the conjugate distribution to the Wishart under a G-transformation, and show its empirical spectral distribution obeys a semicircle law when p{n Ñ 0. Finally, we discuss how the phase transitions of the Wishart distribution might originate from changes in rates of convergence of symmetric t statistics. MSC 2010 subject classifications: Primary 60B20, 60B10; secondary 60E10. 1. Introduction The roots of random matrix theory lies in statistics, with the work of Wishart [1928] and Bartlett [1933], and in numerical analysis, with the work of Von Neumann and Goldstine [1947]. In this early period, many well-known matrix distributions were introduced. This includes the real Gaussian matrix ensemble Gpp, qq, a p ˆ q matrix with independent standard Gaussian entries, the Gaussian orthogonal ensemble GOEppq, the distribution of a symmetric matrix pX ` ? X t q{ 2 with X „ Gpp, pq, and the Wishart (also known as Laguerre) distribution Wp pn, Ip {nq, the distribution of a symmetric matrix XX t {n with X „ Gpp, nq. During that time, the main concern was to derive properties of these distributions for a fixed dimension. Some asymptotics of the Wishart distribution were considered, but only as n Ñ 8 for fixed p. 1 Chételat and Wells/Mid-scale Wishart asymptotics 2 Starting with the pioneering work of Wigner [1951, 1955, 1957], Porter and Rosenzweig [1960], Gaudin [1961] and Mehta [1960a,b], researchers began investigating the asymptotics of Gaussian ensembles as their dimension grew to infinity. As a result of decades of work, the behavior of a GOEppq matrix is now well understood both in the classical setting where p is fixed, and in the setting where p Ñ 8. However, the situation asymptotics of the Wishart distribution is more complicated, as it depends on two parameters, n and p, and initial progress was slow. The work of Marchenko and Pastur [1967] clearly established that the analogue of a Gaussian orthogonal ensemble matrix whose dimension p grows to infinity is a Wishart matrix whose degrees of freedom n and dimension p jointly grow to infinity in such a way that p{n Ñ c P p0, 1q. Since then, we gained a very good understanding of the behavior of Wishart matrices in this regime. But this body of work left open the question as to what happens to a Wishart matrix when n, p Ñ 8 with p{n Ñ 0. Since such asymptotics are middle-scale between the classical regime where p is fixed as n Ñ 8 and the high-dimensional regime where p{n Ñ c P p0, 1q, we might refer to them as middle-scale regimes. Hence, we might ask: what is the asymptotic behavior of a Wishart matrix Wp pn, Ip {nq in the middle-scale regimes? This question is addressed this article. To gain some intuition, it is instructive to look at the eigenvalues λ1 ą ¨ ¨ ¨ ą λp ą 0 of a Wp pn, Ip {nq Wishart matrix. In the classical regime where p is fixed as n Ñ 8, the eigenvalues must all almost surely tend to 1 by the strong law of large numbers. In constrast, in the high-dimensional regime where both n, p Ñ 8 with p{n Ñ c P p0, 1q, the Marchenko-Pastur law states that for any bounded, continuous f , a ż c` n pc` ´ lqpc´ ´ lq 1 ÿ dl a.s., f pλi q Ñ f plq n i“1 2πcl c´ ? where c˘ “ p1 ˘ cq2 . Thus the eigenvalues do not all tend to 1, but rather distribute themselves in the shape of a Marchenko-Pastur law with parameter c. What happens between these two extremes? When c Ñ 0, the MarchenkoPastur law converges weakly to a Dirac measure with mass at 1. This suggests that whenever n, p Ñ 8 with p{n Ñ 0 n 1 ÿ f pλi q Ñ f p1q n i“1 a.s., or in other words that the eigenvalues converge almost surely to 1, as in the classical case. This motivates a binary view of Wishart asymptotics. It appears that the behavior of a Wishart matrix in the middle-scale regimes is the same as in the classical regime, and therefore that there really are only two regimes: lowdimensional where p{n Ñ 0, and high-dimensional where p{n Ñ c P p0, 1q. This binary view has very concrete repercussions. For example, in statistics, many covariance matrix estimators have been developed that leverage high- Chételat and Wells/Mid-scale Wishart asymptotics 3 dimensional Wishart asymptotics (see Pourahmadi [2013] for a review). When faced with a problem where p is large with respect to n, it has been argued that the high-dimensional asymptotics, rather than the classical, constitute the correct model. The binary view provides a useful rule of thumb: small p’s call for classical covariance estimators, while large p’s call for high-dimensional covariance estimators. Unfortunately, recent results establish that this binary view is incorrect. In the classical regime where p is fixed, the central limit theorem implies that ı ? ” n Wp pn, Ip {nq ´ Ip ñ GOEppq, as n Ñ 8, where the arrow stands for weak convergence. In fact, something better is known: recent work has extended this result to the case where p tends to infinity. Recall that the total variation distance between two absolutely continuous distributions şF1 and F2 with densities f and g is given by dTV pF1 , F2 q “ dTV pf1 , f2 q “ |f1 pxq ´ f2 pxq|dx. With different approaches, Jiang and Li [2015] and Bubeck et al. [2016] independently established that ˙ ˆ ” ı ? n Wp pn, Ip {nq ´ Ip , GOEppq Ñ 0 dTV whenever p3 {n Ñ 0. Thus, when p3 {n Ñ 0, the same asymptotics hold as in the p fixed case, and we might regard these regimes as rightfully belonging to the classical setting. The surprising part is that the converse is true! When p3 {n Û 0, results of Bubeck et al. [2016] and Rácz and Richey [2016] show that ˆ ” ˙ ı ? dTV n Wp pn, Ip {nq ´ Ip , GOEppq Û 0. ? Thus a phase transition occurs when p is of the order 3 n. This begs the?question: if a normal approximation fails to hold when p grows faster than 3 n, what asymptotics hold? Is there a uniform asymptotic behavior that holds whenever p{n Ñ 0 with p3 {n Û 0, or are there further phase transitions as the growth rate of p is increased? The results of this paper offers a mostly complete answer to this question. Namely, we establish that when p3 {n Û 0 but p2 {n Ñ 0, ˆ ” ˙ ı ? dTV n Wp pn, Ip {nq ´ Ip , F1 Ñ 0, where F1 is a continuous distribution on the space of real symmetric matrices whose density is given when n ě 3p ´ 3 by ? ˇ „ " ˇ 2i i i 1 3 4 f1 pXq 9 ˇˇE exp ? tr XZ ´ ? tr Z ` tr Z ` 3{2 tr Z 5 4n 5n 8 3 2n ? *ˇ2 ˇ 1 ipp ` 1q p`1 4 2ipp ` 1q 6 2 3 ˇ , ´ 2 tr Z ` ? tr Z ´ tr Z ´ tr Z (1.1) ˇ 3{2 3n 4n 3n 2 2n Chételat and Wells/Mid-scale Wishart asymptotics 4 Fig 1. Correct picture of Wishart asymptotics. This contrasts with the binary view, where no phase transitions occur between p held constant and p growing like n. ? for a Z „ GOEppq. When p grows like n, another phase transition occurs. Namely, we establish that when p2 {n Û 0 but p5{3 {n Ñ 0, ˙ ˆ ” ı ? n Wp pn, Ip {nq ´ Ip , F2 Ñ 0, dTV where F2 is a continuous distribution on the space of real symmetric matrices whose density is given when n ě 3p ´ 3 by ? ˇ „ " ˇ i i 1 2i 3 4 ˇ f2 pXq 9 ˇE exp ? tr XZ ´ ? tr Z ` tr Z ` 3{2 tr Z 5 4n 5n 8 3 2n ? ? ipp`1q 1 2 2i p`1 4 2ipp`1q 7 6 2 tr Z 3 ´ 2 tr Z ´ 5{2 tr Z ` ? tr Z ´ tr Z ´ 3n 4n 7n 3n3{2 2 2n *ˇ2 ˇ p`1 1024pp`1q 512pp`1q 5 4 6 ˇ , ` tr Z ´ tr Z ` i tr Z (1.2) ˇ 4n2 3n3 5n5{2 again for a Z „ GOEppq. In general, for every K P N we find a continuous distribution FK on the space of real symmetric matrices, with density given when n ě 3p ´ 3 by 2K`3` 1rK odds ˇ « # ÿ ˇ ´ 2 ¯ k2 tr Z k i n ˇ fK pXq 9 ˇE exp ? trpXZq ` ik ˇ 4 n k 8 k“3 2K`2´ 1rK odds p `1 ` 4 ÿ k i +ffˇ2 ˇ ˇ ˇ ˇ k ´ 2 ¯ k2 tr Z k n (1.3) k“1 for a Z „ GOEppq, which approximates the normalized Wishart distribution in some (but not all) middle-scale regimes. Namely, we prove the following, which can be regarded as the main result of this paper. Theorem 1.For any K P N,?the total variation distance between the the normalized Wishart distribution nrWp pn, Ip {nq ´ Ip s and the K th degree density fK satisfies ˆ ” ˙ ı ? dTV n Wp pn, Ip {nq ´ Ip , FK Ñ 0 Chételat and Wells/Mid-scale Wishart asymptotics 5 as n Ñ 8 with pK`3 {nK`1 Ñ 0. The definition of fK and proof of Theorem 1 are found in Section 6, and follow from definitions and results from Sections 3, 4 and 5 that constitute the bulk of this paper. The main consequence of this theorem is the existence of an infinite countable number of phase transitions, occurring when p grows like npK`1q{pK`3q for K P N. A diagram is provided at Figure 1. This naturally groups“ the middle-scale ˘ log p K K`1 regimes satisfying lim log n ă 1 by which semi-open interval K`2 , K`3 their nÑ8 log p nÑ8 log n limit lim belongs to. We might refer to this grouping as the degree of the log p regime. In other words, we will say an middle-scale regime satisfying lim log n ă nÑ8 “ K K`1 ˘ log p 1 has degree K when lim log n P K`2 , K`3 . nÑ8 The main result of this paper, Theorem 1, may then be summarized as saying that the normalized Wishart distribution can be approximated by the distribution with density fK in every middle-scale regime of degree K or less. The 0th degree case corresponds to the classical setting, while the higher degrees correspond to previously unknown behavior. In fact, we show that our 0th degree approximation F0 is asymptotically equivalent to the Gaussian orthogonal ensemble. The results of this paper can therefore be regarded as a wide generalization of the Wishart asymptotics results of Jiang and Li [2015], Bubeck et al. [2016], Bubeck and Ganguly [2016] and Rácz and Richey [2016]. Our approach relies on a novel technical tool we call the G-transform. It turns out that to understand middle-scale regime behavior of Wishart matrices, densities are less clear than characteristic functions (that is, Fourier transforms of densities). Unfortunately, characteristic functions are difficult to relate to metrics like the total variation distance. To remedy this problem, we develop the G-transform and some associated theory in Section 3. An interesting aspect of G-transform theory is that to every distribution we can associate a closely related distribution called its G-conjugate. In fact, the G-conjugate of a Wishart matrix is essentially a generalization of the t distribution to real symmetric matrices. In Section 4, we define and derive several results concerning this new distribution, including a semicircle law. From these results, we derive in Section 5 approximations to the Wishart distribution for middle-scale regimes of every degree. Since these approximations are given using the language of G-transforms, we derive in Section 6 density approximations, from which Theorem 1 follows. We briefly discuss what concrete effects the phase transitions might have on Wishart asymptotics in Section 7. Finally, we compile auxiliary results in Section 8, while we discuss in Section 9 open questions that arise from these results. Although the results of this paper explain a large part of the behavior of Wishart matrices when p{n Ñ 0, there exists regimes for which p{n Ñ 0 yet log p p R OpnpK`1q{pK`3q q for all K P N, or in other words for which lim log n “ 1. nÑ8 ? An example is when p grows at the order n1´1{ log n . Although the results of our paper characterize almost all middle-scale regimes in the sense that among log p log p those regimes satisfying lim log n ď 1, those such that lim log n “ 1 represent nÑ8 nÑ8 Chételat and Wells/Mid-scale Wishart asymptotics 6 a negligible set, they nonetheless exist. One might regard regimes such as those as having infinite degree. Beyond this, however, it is difficult to say anything about the behavior of Wishart matrices in these regimes. More work in that direction is clearly needed. 2. Notation and definitions The transpose of a matrix is denoted t , and the identity matrix of dimension p is Ip . As is standard, we take the trace operator to have lower priority than the power operator: thus for a matrix X, tr X k means the trace of X k . We will write trk X when we mean the k th power of the trace of X. The Kronecker delta is the symbol δkl “ 1rk “ ls. The space of all real-valued symmetric matrices is denoted Sp pRq “ tX P Mp pRq|X “ X t u. For a symmetric matrix X, we define the symmetric differentiation operator Bs {Bs Xkl by # 1 B for k ­“ l 1 ` δkl B Bs kl “ “ 2 BX B Bs Xkl 2 BXkl for k “ l. BXkk This operator has the elegant property that BSBXS kl trpXY q “ Ykl for any two symmetric matrices X, Y . The space of symmetric matrices Sp pRq can be assimilated to Rppp`1q{2 by mapping a symmetric matrix to its upper triangle. By integration over Sp pRq, we mean integration with respect to the pullback Lebesgue measure under this isomorphism, that is, ż ż f pXq dX “ Sp pRq p ` ˘ź f X dXij . Rppp`1q{2 iďj We say a real symmetric matrix follows the Gaussian orthogonal ensemble GOEppq distribution if Xkl , k ď l are all independent, with diagonal elements Xkk „ Np0, 2q and off-diagonal elements Xkl „ Np0, 1q. Let X be a n ˆ p matrix of i.i.d. Np0, 1q random variables, and let Σ be a p ˆ p positive-definite matrix. The Wishart distribution Wp pn, Σq is the distri1 1 bution of the random matrix Σ 2 X t XΣ 2 . This is a special case of the matrix gamma distribution. Following Gupta and Nagar [1999, Section 3.6], we say a positive-definite matrix X has a matrix gamma distribution Gp pα, Σq with shape parameter α ą pp ´ 1q{2 and scale parameter Σ if it has density over Sp pRq given by ! ) ˇ ˇα´ p`1 1 2 ˇX ˇ exp ´ trpΣ´1 Xq 1rX ą 0s , f pXq “ α |Σ| Γp pαq where Γp is the multivariate gamma function. With this definition, the Wishart distribution Wp pn, Σq is a matrix gamma with shape n2 and scale 2Σ. While studying the Wishart distribution, the expression n ´ p ´ 1 comes up so often that it makes sense to give it its own symbol. We will therefore write m “ n ´ p ´ 1. Chételat and Wells/Mid-scale Wishart asymptotics 7 The Hellinger distance is metric between absolutely continuous probability measures. For two distributions F1 and F2 with densities f1 and f2 , their Hellinger distance is defined as ˇ2 ¯ 12 ´ż ˇ ˇ ˇ 1{2 1{2 HpF1 , F2 q “ Hpf1 , f2 q “ ˇf1 pxq ´ f2 pxqˇ dx . The Hellinger distance is closely related to the total variation distance by the inequalities 1 1{2 dTV pf1 , f2 q ď Hpf1 , f2 q ď dTV pf1 , f2 q. (2.1) 2 In particular, Hpf1 , f2 q Ñ 0 if and only if dTV pf1 , f2 q Ñ 0. Thus they can be seen as inducing the same topology on absolutely continuous probability measures, called the strong topology, in contrast to the topology induced by weak convergence of measures called the weak topology. One can show that if a sequence of measures converges in the strong sense (i.e. in the dTV or H metrics), then it converges weakly. 3. G-transforms Our analysis of Wishart matrices relies heavily on a tool we call the G-transform of a probability measure. To do so, we first need to define the Fourier transform over symmetric matrices. In Section 2, we clarified what we meant by integration over Sp pRq. For a function f : Sp pRq Ñ C in L1 pSp pRqq, we define its Fourier transform to be ż Ftf upT q “ 1 p 22 π ppp`1q 4 e´i trpT Xq f pXq dX. (3.1) Sp pRq It is more common to define the Fourier transform on symmetric(matrices with ( ř the integrand exp ´ i kďl Tkl Xkl , but choosing exp i trpT Xq considerably simplifies our computations. We extend this definition to f P Lr pSp pRqq, 1 ă r ď 2 in the usual manner. Because of the specific normalization chosen, this definition obeys a simple version of Plancherel’s theorem, namely ż ż f pXqs g pXq dX “ Sp pRq Ę q dT. Ftf upT qFtgupT Sp pRq We now define the G-transform. In itself, the definition has nothing to do with symmetric matrices and could have been perfectly well defined on any other space endowed with a Fourier transform. Definition 1 (G-transform of a density). Let f be an integrable function Sp pRq Ñ C. Its G-transform is the complex-valued function Gtf u : Sp pRq Ñ C defined by Gtf u “ Ftf 1{2 u2 , where z 1{2 stands for the principal branch of the complex logarithm. (3.2) Chételat and Wells/Mid-scale Wishart asymptotics 8 In the same way that the Fourier transform maps L2 pSp pRqq to itself, the G-transform maps L1 pSp pRqq to itself. By extension, for an absolutely continuous distribution on Sp pRq with density f , we will define its G-transform to be the G-transform of its density. (This usage mirrors other transforms, such as the Stietjes transform.) We will usually denote the G-transform of f by ψ. Since a density is integrable, this is always well-defined. Moreover, f “ F ´1 tψ 1{2 u2 , so the density can be recovered from the G-transform, and therefore to understand a distribution it is equivalent to study its density or its G-transform. Two comments are in order. First, for many densities, f 1{2 P L1 pSp pRqq. In this case, the G-transform can be written explicitly as ż ˆ ˙2 1 ´i trpT Xq 1{2 (3.3) e f pXq dX . ψpT q “ Gtf upT q “ ppp`1q 2p π 2 Sp pRq Second, throughout this article we will often talk about “the” square root of a G-transform. To be clear, by ψ 1{2 we will always mean Ftf 1{2 u. Now, in many ways, the G-transform behaves similarly to the characteristic function (Fourier transform of a density), but it has unique features. First, Plancherel’s theorem yields that ż ż |ψ 1{2 pT q|2 dT |ψpT q| dT “ Sp pRq żS pRq p ż |f 1{2 pXq|2 dX “ “ Sp pRq |f pXq| dX “ 1. (3.4) Sp pRq Thus |ψ| is itself a density, which we will call the G-conjugate of f . (In particular, ψ 1{2 is much like a quantum-mechanical wavefunction.) We will also use an asterisk notation, so that the G-conjugate of a Np0, 1q distribution will be denoted Np0, 1q˚ . For example, straightforward computations yield that 2 2 ?1 Np0, 1q˚ “ Np0, 1{8q, χ2˚ n “ 8n tn{2 (where χν and tν are the univariate χ and ˚ ´1 ˚ t distributions with ν degrees of freedom, respectively) and paF ` bq “ a F for any distribution F and scalars a ­“ 0, b P R. Studying the G-conjugate of the Wishart distribution will play a key part in deriving results about the Wishart distribution itself. We should note that, in general, the double G-conjugate F ˚˚ is not the same as F . For example, χ2˚˚ is a density involving modified Bessel n functions of the first kind, not a χ2n . A second feature that distinguishes G-transforms from characteristic functions is that they are easy to relate to the Hellinger distance between probability measures. Consider two densities f1 , f2 with G-transforms ψ1 , ψ2 . By analogy, we could define the “total variation” and “Hellinger” distances of ψ1 and ψ2 by ż dTV pψ1 , ψ2 q “ |ψ1 pT q ´ ψ2 pT q| dT, Sp pRq (3.5) Chételat and Wells/Mid-scale Wishart asymptotics ż ´ and 1{2 1{2 |ψ1 pT q ´ ψ2 pT q|2 dT Hpψ1 , ψ2 q “ ¯ 12 . 9 (3.6) Sp pRq Since the modulus of the G-transforms integrate to one, their total variation and Hellinger distances are related to each other in the same way as in Equation 2.1 for densities, namely 1 1{2 dTV pψ1 , ψ2 q ď Hpψ1 , ψ2 q ď dTV pψ1 , ψ2 q. (3.7) 2 Thus dTV pψ1 , ψ2 q Ñ 0 if and only if Hpψ1 , ψ2 q Ñ 0. But the Hellinger distance between G-transforms is much more useful. Indeed, by the Plancherel theorem, for any two densities f1 , f2 with G-transforms ψ1 , ψ2 , their Hellinger distance satisfies ż H2 pf1 , f2 q “ 1{2 1{2 |f1 pXq ´ f2 pXq|2 dX żS pRq p “ 1{2 1{2 |ψ1 pT q ´ ψ2 pT q|2 dT “ H2 pψ1 , ψ2 q. (3.8) Sp pRq Thus to compute the Hellinger distance H2 pf1 , f2 q between two densities, we can instead compute the Hellinger distance H2 pψ1 , ψ2 q of their G-transforms. In contrast, there is no explicit way to express the Hellinger distance in terms of characteristic functions. And no such connection exists between the total variation distances of densities and G-transforms. The G-transform does have some disadvantages compared to the Fourier transform. It is a non-linear transformation (and therefore not a true transform), and it does not behave well with respect to convolution. For our purposes, however, the advantages listed above outweigh these problems. In practice, it is not aways easy to control the Hellinger distance directly, and one often focuses on the Kullback-Leibler divergence instead. The two quantities are related through the well known inequality  „ f1 pXq for X „ f1 . H2 pf1 , f2 q ď E log f2 pXq For G-transforms, the following analog holds, which clarifies our interest in Gconjugates: ˇ2 „  „ˇ ˇ ψ1 pT q ψ1 pT q ˇˇ H2 pψ1 , ψ2 q ď E < Log ` 2E ˇˇ= Log for T „ F1˚ , ψ2 pT q ψ2 pT q ˇ where Log stands for the principal branch of the complex logarithm. In fact, in this article we will need a further generalization, where ψ2 does not need to be a G-transform of a density. Proposition 1 (Kullback-Leibler inequality for G-transforms). Let ψ1 be the G-transform of an absolutely continuous distribution F1 on Sp pRq, and let ψ2 Chételat and Wells/Mid-scale Wishart asymptotics 10 be an integrable function Sp pRq Ñ C. Then ż  „  „ ψ1 pT q H2 pψ1 , ψ2 q ď |ψ2 |pT q dT ´ 1 ` E < Log ψ2 pT q Sp pRq ż ˇ 1 „ˇ ˇ 1 ψ1 pT q ˇˇ 2 ˇ 2 `2 |ψ2 |pT q dT ¨ E ˇ= Log ψ2 pT q ˇ Sp pRq for T „ F1˚ , where Log stands for the principal branch of the complex logarithm. Proof. We can write ż 1{2 H2 pψ1 , ψ2 q “ Sp pRq „ 1{2 1{2 ´ ψ1 ψs2 dT ż ż «  |ψ2 |pT q| ´ 1 ` 2 ´ “ 1{2 |ψ1 |pT q ` |ψ2 |pT q| ´ ψs1 ψ2 1{2 ψ1 pT q Sp pRq 1{2 1{2 ψ2 pT q ` ψ̄2 pT q 1{2 ψ̄1 pT q ff |ψ1 |pT q dT Sp pRq ż „ ż „ |ψ2 |pT q| ´ 1 ` 2 1 ´ #  “ żS pRq żS pRq p „ < p 1{2 ψ2 pT q 1{2 ψ1 pT q +  |ψ1 |pT q dT  „ |ψ2 |pT q| ´ 1 ` 2 1 ´ " * 1 ψ1 pT q exp ´ < Log 2 ψ2 pT q Sp pRq Sp pRq “ ˙  ψ1 pT q 1 = Log |ψ1 |pT q dT . 2 ψ2 pT q a Now using the inequality ´ cospxq ď ´1 ` 2|x| that holds for any x P R. The last quantity is bounded as ˆ ¨ cos ż „ ż  „ |ψ2 |pT q| ´ 1 ` 2 1 ´ " *  1 ψ1 pT q exp ´ < Log |ψ1 |pT q dT 2 ψ2 pT q Sp pRq Sp pRq ď ż `2 ˇ " *dˇ ˇ 1 ψ1 pT q ψ1 pT q ˇˇ ˇ exp ´ < Log ˇ= Log ψ2 pT q ˇ|ψ1 |pT q dT. 2 ψ2 pT q Sp pRq In the second term, use 1 ´ x ď ´ logpxq for x ě 0, while in the third term, use the Cauchy-Schwarz inequality to obtain ż ż „ ď  |ψ2 |pT q| ´ 1 ´ 2 log " * ψ1 pT q 1 |ψ1 |pT q dT exp ´ < Log 2 ψ2 pT q Sp pRq Sp pRq Chételat and Wells/Mid-scale Wishart asymptotics ż `2 11 ż " * 1 ψ1 pT q exp ´< Log |ψ1 |pT q dT 2 ψ2 pT q ˇ 1 ψ1 pT q ˇˇ ˇ ˇ= Log ˇ|ψ1 |pT q dT 2 . ψ2 pT q Sp pRq Sp pRq Now use Jensen’s inequality in the second term and the algebraic identity expt´< Log ψ1 pT q{ψ2 pT qu “ |ψ2 |pT q{|ψ1 |pT q in the third term to obtain ż ż „ ď  |ψ2 |pT q| ´ 1 ` < Log Sp pRq Sp pRq ψ1 pT q |ψ1 |pT q dT ψ2 pT q ż ż |ψ2 |pT q dT 2 ˇ 1 ψ1 pT q ˇˇ ˇ ˇ= Log ˇ|ψ1 |pT q dT 2 , ψ2 pT q Sp pRq Sp pRq 1 `2 as desired. Let us now compute the G-transform of the Gaussian Orthogonal Ensemble and the normalized Wishart distribution, which will be needed in our proofs. The density of a GOEppq matrix over Sp pRq is ! 1 ) 1 fGOE pXq “ ppp`3q ppp`1q exp ´ tr X 2 . (3.9) 4 2 4 π 4 To compute its G-transform, we will make use of the fact that the elements of a GOEppq matrix are independent to reduce the expression to a product of characteristic functions. Proposition 2. The G-transform of the Gaussian Orthogonal Ensemble density on Sp pRq is pp3p`1q ! ) 2 4 ψGOE pT q “ ppp`1q exp ´ 4 tr T 2 . π 4 ? 1{2 Proof. From Equation (3.9), fGOE is proportional to the density of the 2 GOEppq distribution, so it is integrable. Therefore, we can apply Equation (3.3) to find that ż 1 1{2 ψGOE pT q “ p 2 2 π ppp`1q 4 ! ) 1{2 exp ´ i trpT Xq fGOE pXq dX żS pRq p 1 “ 2 ppp`7q 8 π 3ppp`1q 8 ! ) 1 exp ´ i trpT Xq ´ tr X 2 dX 8 żS pRq p 1 “ 2 ppp`7q 8 π 3ppp`1q 8 p p ! ÿ 1ÿ 2 exp ´ 2i Tkl Xkl ´ X 4 kăl kl kăl Rppp`1q{2 ´i p ÿ k“1 Tkk Xkk ´ p p 1 ÿ 2 )ź Xkk dXkl 8 k“1 kďl Chételat and Wells/Mid-scale Wishart asymptotics 12 ż “ 2 pp3p`1q 8 π ppp`1q 8 p ź kăl ) exp ´ 1 X 2 ( ? 4 kl dXkl exp ´ 2iTkl Xkl 4π ! R ż p ź ) exp ´ 1 X 2 ( ? 8 kk dXkk exp ´ iTkk Xkk 8π ! ¨ k“1 “ 2 pp3p`1q 8 π ppp`1q 8 R p p ” ! )ı ” ! ? )ı ź ź E exp ´ 2iTkk Z E exp ´ 8iTkl Z kăl k“1 for Z „ Np0, 1q. The characteristic function of a Np0, 1q is expp´t2 {2q, so pp3p`1q p p ! )ź ! ) ź 2 8 2 2 “ ppp`1q exp ´ 4Tkl exp ´ 2Tkk π 8 kăl k“1 pp3p`1q ! ) 2 8 “ ppp`1q exp ´ 2 tr T 2 . π 8 Squaring this result yields the desired expression for ψGOE . In particular, we see that |ψGOE | is the density of a GOEppq{4 distribution, or in other words that GOEppq˚ “ GOEppq{4. In particular the Gaussian orthogonal ensemble is its own G-conjugate, up to a constant factor. Let us now compute the G-transform of the normalized Wishart distribution. Unlike the GOEppq case, the elements of the matrix are not independent, but the elements of its Cholesky decomposition are. By being careful about complex changes of variables, we can reduce the computation of the G-transform to the computation of characteristic functions of the Cholesky elements. Proposition?3. Let n ě p´2. Then the G-transform of the normalized Wishart distribution nrWp pn, Ip {nq ´ Ip s density on Sp pRq is given by ˇ´ n`p`1 " *ˇ 2 ˇ ˇ ? 4T ψNW pT q “ Cn,p exp 2i n tr T ˇˇIp ` i ? ˇˇ n with ˘ ` ppn`2pq Γ2p n`p`1 2 2 4˘ ` Cn,p “ ppp`1q ppp`1q . Γp n2 π 2 n 4 Proof. Recall the notation m “ n ´ p ´ 1 used throughout this article. The density of a Y „ Wp pn, Ip {nq distribution is np ! n ) m n 2 1rY ą 0s ` n ˘ exp ´ tr Y |Y | 2 . fW pY q “ np 2 2 2 Γp 2 ? ? If we do a change of variables X “ npY ´ Ip q, so that Y “ Ip ` X{ n and p p ź ź 1 dYij “ ppp`1q dXij , n 4 iďj iďj ? we see that the normalized Wishart distribution nrWp pn, Ip {nq ´ Ip s has den- Chételat and Wells/Mid-scale Wishart asymptotics 13 sity „  ppn`mq n 4 X ` ˘ ? 1 I ` ą 0 np p n 2 2 Γp n2 fNW pXq “ ! n ” X ˇˇ m2 X ı)ˇˇ ¨ exp ´ tr Ip `? ˇIp `? ˇ . (3.10) 2 n n ` 4 ´1 ˘ p`1 n`p`1 1{2 Notice that fW is proportional to expt´ tr r n Ip s Y u|Y | 4 ´ 2 , so it ` ˘ must be proportional to the density of a matrix gamma distribution Gp n`p`1 , n4 Ip 4 ą p´1 when n`p`1 4 2 , i.e. n ě p ´ 2. In particular, it must be integrable. As fNW 1{2 was obtained by a linear change of variables from fW , fNW must be integrable too, that is ż 1{2 fNW pXqdX ă 8. (3.11) Sp pRq Therefore, we can apply Equation (3.3) to obtain ż 1 1{2 ψN W pT q “ “ p 22 π 1 p 22 π ppp`1q 4 ppn´2q 4 1 ppp`1q 4 ! ) 1{2 exp ´ i trpT Xq fNW pXq dX Sp pRq ” ! ) ı ´1{2 E exp ´ i trpT Xq fNW pXq `n˘ „ " *  n ” X ı ˇˇ X ˇˇ´ m4 ? ? E exp ´ i trpT Xq ` tr I ` ` . ˇI ˇ p p ppp`1q ppn`mq 4 n n π 4 n 8 ? If we rewrite the expectation in terms of Y “ Ip ` X{ n, this last expression equals 1 ` ˘ ppn´2q *  „ "´ `? ˘ ? 2 4 Γp2 n2 n ¯ ´m 4 . (3.12) “ ppp`1q ppn`mq exp i n tr T E etr ´i nT ` Ip Y |Y | 4 π 4 n 8 Since T is real symmetric, there must be a spectral decomposition T “ ODOt with O real orthogonal and D real diagonal. As Ot Y O has the same distribution as Y , namely Wp pn, Ip {nq, we can rewrite Equation (3.12) as 1 ` ˘ ppn´2q „ "´ *  ˘ `? ? 2 4 Γp2 n2 n ¯ ´m 4 “ ppp`1q ppn`mq exp i n tr T E etr ´i nD ` Ip Y |Y | . (3.13) 4 π 4 n 8 Now, since Y is positive-definite it has a Cholesky decomposition Y “ U t U with U upper-triangular. According to Bartlett’s theorem [see Muirhead, 1982, Theorem 3.2.14], all the elements of U are independent, the diagonal elements 2 have the distribution Ukk „ χ2n´k`1 {n and the upper diagonal elements have Ukl „ Np0, 1{nq for k ă l. Since „´  p ´ ÿ ? ? Ip ¯ n ¯ 1 tr ´ i nD ` Y “ ´ i nD ` Ip Ukl Ulj 4 4 jk j,k,l “ 2 Γp2 2 Chételat and Wells/Mid-scale Wishart asymptotics p ´ ÿ “ lďk 14 ? n¯ 2 ´ i nDkk ` U 4 lk śp 2 k“1 Ukk , we have by independence and Equation (3.13) that # + p „ 1 ` ˘ ppn´2q * "´ ź ? ? 2 4 Γp2 n2 n¯ 2 “ ppp`1q ppn`mq exp i n tr T U E exp ´ i nDkk ` 4 lk π 4 n 8 lăk *  „ "´ p ź ? n¯ 2 2 ´m E exp ´ i nDkk ` ¨ Ukk pUkk q 4 . (3.14) 4 k“1 and |Y | “ We will now compute these expected values in several steps. For a given 1 ď k ď p, let „ "´  * ? n¯ 2 ´m A “ E exp ´ i nDkk ` (3.15) Tkk Tkk 2 . 4 2 Since Tkk „ χ2n´k`1 {n and m “ n ´ p ´ 1, we have ż8 " * n´k`1 ´ 2 ? m n n¯ ¯ exp ´ i nDkk ` x x´ 4 A “ n´k`1 ´ n´k`1 4 2 2 Γ 0 2 ! n ) n´k`1 ¨ x 2 ´1 exp ´ x dx 2 ż8 " * n´k`1 ´ ¯ n´2k`p`3 n 2 n ? ´1 4 ¯ “ n´k`1 ´ dx (3.16) exp ´ ` nDkk i x x n´k`1 4 2 Γ 2 0 2 Consider the truncated integrands " ´ ¯ * n´2k`p`3 n ? ´1 4 hM pxq “ exp ´ ` nDkk i x x 1r0 ă x ă M s . 4 Clearly this sequence is dominated by the integrable positive function h, " * n´2k`p`3 n ´1 4 , |hM pxq| ď hpxq “ exp ´ x x 4 ˆ ˙ n´2k`p`3 ˙ ˆ ż8 4 4 n ´ 2k ` p ` 3 hpxqdx “ ă 8. Γ n 4 0 Therefore, by the Dominated Convergence Theorem and Equation (3.16), żM " n´k`1 ´n ? ¯ * n´2k`p`3 2 n ´1 4 ¯ lim A “ n´k`1 ´ exp ´ ` nDkk i x x dx. M Ñ8 n´k`1 4 2 2 Γ 0 2 `n ? ˘ By the change of variables z “ 4 ` nDkk i x, this can be rewritten “ n n´k`1 2 ? ˘´ n´2k`p`3 4 nDkk i 4 ` ´ ¯ lim n´k`1 M Ñ8 2 2 Γ n´k`1 2 `n ż nM 4 ? ` nDkk M i e´z z n´2k`p`3 ´1 4 dz. (3.17) 0 To compute this integral, we use a contour argument. Consider the closed path nM C “ C1 ` C2 ` C3 given by C1 a path from 0 to nM 4 , C2 a path from 4 to Chételat and Wells/Mid-scale Wishart asymptotics y nM 4 15 ? ` nDkk M i C3 C2 C1 0 nM 4 x Fig 2. Contour C “ C1 ` C2 ` C3 when Dkk ě 0. The diagram is mirrored around the x axis when Dkk ă 0. ? ? ` nDkk M i and finally C3 a path from nM nDkk M i to 0. A diagram 4 ` is provided as Figure 2. n´2k`p`3 ´1 4 As k ď p, z ÞÑ e´z z is entire and its integral over C must be zero. Therefore ? ˇ ˇ nM ˇ ˇ 4 ` nDkk M i ˇ ˙ˇ ˆ ˇ ˇ n´2k`p`3 n ´ 2k ` p ` 3 ´1 ˇ lim ˇ 4 e´z z dz ´ Γ ˇM Ñ8 ˇ 4 ˇ ˇ ˇ ˇ 0 ? ˇ ˇ nM nM ˇ ˇ 4 ` nDkk M i 4 ˇ ˇ ˇ ˇ n´2k`p`3 n´2k`p`3 ´1 ´1 4 4 “ lim ˇˇ e´z z dz ´ e´x x dxˇˇ M Ñ8 ˇ ˇ ˇ 0 ˇ 0 ? ˇ ˇ ˇ ˇ nM ˇ ˇ ˇ ˇ 4 ` nDkk M i ˇ ˇ ˇ ˇ ˇ ˇ ˇ ˇ n´2k`p`3 n´2k`p`3 ´z ´z ´1 ´1 ˇ ˇ ˇ 4 4 e z “ lim ˇ e z dz ˇ “ lim ˇ dz ˇˇ . M Ñ8 ˇ M Ñ8 ˇ ˇ ˇ ˇ C2 ˇ ˇ nM ˇ 4 nM ? z´ 4 ? Do a change of variables z “ nM 4 ` y nDkk M i, so that y “ nDkk M i is real on the path. It yields ˇ ˇ 1 ˇ ˇ ˇ „  n´2k`p`3 ´1 ˇ 4 ? ˇ ? ? M ˇˇ nM “ n|Dkk | lim nM ˇ e´y nDkk M i `y nDkk M i dy ˇˇ M Ñ8 e 4 ˇ 4 ˇ ˇ 0 ˇ nM 4 ż ż ż ż ż ż ż ď ? 1 ˇ ˇ n´2k`p`3 n´2k`p`3 ´1 4 ˇn ˇ 4 ? M ˇ ` y nDkk iˇ dy ¨ lim . nM ˇ4 ˇ M Ñ8 e 4 n|Dkk | 0 This last integral is finite, since it is continuous on a bounded interval. Therefore Chételat and Wells/Mid-scale Wishart asymptotics the limit is zero and by Equation (3.17) and the previous expression, ´ ¯ n´k`1 ´n ? ¯´ n´2k`p`3 n 2 Γ n´2k`p`3 4 4 ´ ¯ A “ . ` nDkk i n´k`1 n´k`1 4 2 2 Γ 2 16 (3.18) Going back to (3.14), let us now consider the expectations in the second products. For fixed 1 ď l ă k ď p, let „ "´ * ? n¯ 2 ´ i nDkk ` B “ E exp T . (3.19) 4 lk 2 Since Tlk „ χ21 {n, ż8 c B “ "´ n 2π exp * n ? n ¯ e´ 2 x ´ i nDkk ` x ? dx 4 x 0 ż8 c “ " ´ ¯ * 1 ? n exp ´ ` i nDkk x ? dx. 4 x n 2π (3.20) 0 Consider the truncated integrands  " ´ ¯ * 1 „ 1 ? n ` i nDkk x ? 1 ăxăM . (3.21) hM pxq “ exp ´ 4 x M We see that they are dominated by a positive, integrable function hpxq, c ż8 n e´ 4 x π |hM pxq| ď hpxq “ ? , ă 8. hpxqdx “ 2 x n 0 Therefore, by the Dominated Convergence Theorem and Equation (3.20), we conclude that M c " ´ ¯ * 1 ? n n B “ lim exp ´ ` i nDkk x ? dx. 2π M Ñ8 4 x 1{M ` ˘ ? A complex change of variables z “ n4 ` i nDkk x yields ż ż c “ ¯´ 12 ? n ´n ` i nDkk lim M Ñ8 2π 4 nM 4 ? ` nDkk M i ? e´z { z dz. ? nDkk n 4M ` M (3.22) i Let’s compute this integral again using a contour integration argument. Consider n the contour C “ C1 ` C2 ` C3 ` C4 given by C1 a line from 4M to nM 2 a line 4 , C? ? ? nM nM nM n kk from 4 to 4 ` nDkk M i, C3 a line from 4 ` nDkk M i to 4M ` nD i M ? nD n n kk and C4 a line from?4M ` M i to 4M . A diagram is provided as Figure 3. Since z ÞÑ e´z { z is holomorphic away from zero, ? ? ˇ ˇ ˇ ˇ nM nM nM ˇ ˇ ˇ ˇ 4 ` nDkk M i 4 ` nDkk M i 4 ˇ ˇ ˇ ˇ ˇ ˇ ˇ ˇ ? ? ? ? ´z ´z ´x ˇ lim ˇ “ lim ˇ ˇ e { z dz ´ π e { z dz ´ e { x dx ˇM Ñ8 ˇ M Ñ8 ˇ ˇ ˇ ˇ ˇ ˇ ? ? nDkk nDkk ˇ ˇ ˇ ˇ n n n ` i ` i ż 4M ż ż M 4M M 4M Chételat and Wells/Mid-scale Wishart asymptotics y nM 4 17 ? ` nDkk M i C3 n ` 4M ? nDkk M C2 i C4 0 n 4M C1 nM 4 x Fig 3. Contour C “ C1 ` C2 ` C3 ` C4 when Dkk ě 0. The diagram is mirrored around the x axis when Dkk ă 0. ˇ ˇ ˇ ˇ “ lim ˇ M Ñ8 ˇ ˇ ˇ ˇ ˇ ˇ ď lim ˇˇ M Ñ8 ˇ ˇ ż ż ˇ ˇ ˇ ? ? ˇ e´z { z dz ` e´z { z dz ˇ ˇ ˇ C2 C4 ? ˇ ˇ ˇ nM n ˇ ˇ ˇ 4 ` nDkk M i 4M ˇ ˇ ˇ ˇ ˇ ˇ ? ? e´z { z dz ˇˇ ` ˇˇ e´z { z dz ˇˇ . ˇ ˇ ˇ ˇ ˇ n ` ?nDkk i ˇ nM 4 4M M ? ? nDkk n By changes of variables z “ nM i on the 4 ` y nDkk M i and z “ 4M ` y M two respective integrals, we get ˇ 1 ˇ ˇ ˇ ? ˇ ˇ nD M i y kk ? M ˇ e ˇ “ n|Dkk | lim nM ˇ 1 dy ˇ ı ” M Ñ8 e 4 ˇ ˇ ? 2 ˇ 0 nM ˇ 4 ` y nDkk M i ˇ 1 ˇ ? ˇ ˇ nDkk ˇ ˇ i y ? 1 ˇ e M ˇ ` n|Dkk | lim dy ˇ ˇ n 1 ” ı ? M Ñ8 M e 4M ˇ ˇ 2 n kk ˇ 0 4M ˇ ` y nD i M ż ż ż ż ż1 ď ? n|Dkk | 0 ? 1 M dy ¨ lim nM ˇ ˇ 12 M Ñ8 ? ˇn ˇ e 4 ˇ 4 ` y nDkk iˇ ż1 1 1 ? n . ˇ ˇ 12 dy ¨ Mlim Ñ8 ? M e 4M ˇn ˇ 0 ˇ 4 ` y nDkk iˇ 1 ˇ ˇ ´ 2 ¯´ 14 ? ˇ ˇ´ 2 2 “ n16 ` y 2 nDkk is continuous on r0, 1s, a bounded Since ˇ n4 ` y nDkk iˇ interval, we conclude that the integrals are finite and that the limits are zero. ` ? n|Dkk | Chételat and Wells/Mid-scale Wishart asymptotics Therefore, by Equation (3.22), c ´ ¯´ 12 ? n n B “ ` i nDkk . 2 4 18 (3.23) Recall the definitions of A and B at Equations (3.15) and (3.19). Combining 1{2 both Equations (3.18) and (3.23) into the expression for ψNW at Equation (3.14) provides 1 ` ˘ ppn´2q " *ź p c ´ ¯´ 12 ? ? 2 4 Γp2 n2 n n 1{2 ψNW pT q “ ppp`1q ppn`mq exp i n tr T ` i nDkk 2 4 π 4 n 8 lăk ¯ ´ n´k`1 p n 2 ¯´ n´2k`p`3 ´n ? Γ n´2k`p`3 ź 4 4 ´ ¯ ¨ nD i ` kk n´k`1 n´k`1 4 Γ k“1 2 2 2 ´ ¯ 1 ` ˘ ppn´2q " * n´k`1 k´1 p Γ n´2k`p`3 p n 2 ` ź ź 4 2 ? 2 Γp 2 4 n 2 ` ˘ “ ppp`1q ppn`mq exp i n tr T n´k`1 ` k´1 Γ n´k`1 2 2 π 4 n 8 2 k“1 k“1 2 p ´ n´2k`p`3 ¯ ź ´ ´ k´1 n ? 4 2 ´ nDkk i ¨ 4 k“1 ´ ¯ n´2k`p`3 ˇ n`p`1 " *ˇ pp2n`p`1q p Γ ź ˇn 8 ? ? ˇ´ 4 4 n ` n´k`1 ˘ exp i n tr T ˇˇ Ip ` i nT ˇˇ “ ppn`2q ppp`1q . 4 2 4 π 4 k“1 Γ 2 But by Muirhead [1982, Theorem 2.1.12], p ` ˘ ´ ¯ ppp´1q ś ˘ ` π 4 Γ n`p`1 ` 1´k p Γ n´2k`p`3 4 2 ź Γp n`p`1 4 k“1 4 ` ˘ , ` ˘ “ p ˘ “ Γp n ` ppp´1q ś Γ n´k`1 2 2 k“1 π 4 Γ n2 ` 1´k 2 k“1 so by taking a n{4 factor out of the determinant, we find that ` ˘ ˇ´ n`p`1 " *ˇ ppn`2pq 4 ˇ ? Γp n`p`1 2 4 4T ˇˇ 1{2 4 ˇ ? ψNW pT q “ ppp`1q ppp`1q exp i n tr T I ` i . ` ˘ p ˇ 1{2 nˇ Γp n2 π 4 n 8 Squaring this result yields the desired expression for ψNW . By Proposition 3, when n ě p ´ 2 the G-conjugate of a normalized Wishart distribution must have a density on Sp pRq given by ` ˘ˇ ˇ´ n`p`1 ppn`2pq 4 ˇ Γ2p n`p`1 2 2 16T 2 ˇˇ 4 ` n ˘ ˇˇIp ` |ψNW |pT q “ ppp`1q ppp`1q . (3.24) ˇ n Γp 2 π 2 n 4 As mentioned in the paragraph following Equation (3.4), the G-conjugate of a χ2n {n distribution is a scaled tn{2 . Thus, by analogy, Equation (3.24) should be represent some kind of generalization of the t distribution to the real symmetric matrices. Matrix-variate generalizations of the t distribution have been investigated in the past, but not for symmetric matrices. Hence it appears the concept is new. Chételat and Wells/Mid-scale Wishart asymptotics 19 This motivates us to propose in Section 4 a candidate for a symmetric matrix variate t distribution. Using that definition, the G-conjugate to the normalized Wishart could then be regarded as the t distribution with n{2 degrees of freedom and scale matrix Ip {8, which we denote Tn{2 pIp {8q. But regardless of its name, this distribution will play a key role in our results about the middle-scale regime asymptotics of Wishart matrices, and will be investigated in depth in Section 4. 4. The symmetric matrix variate t distribution In Section 3, Equation (3.24), we proved ? that when n ě p ´ 2, the G-conjugate of the normalized Wishart distribution nrWp pn, Ip {nq ´ Ip s has density on Sp pRq given by ` ˘ˇ ˇ n`p`1 ppn`2pq 2 ˇ´ 4 ˇ Γ2p n`p`1 16T 2 2 4 ˇ ` ˘ ˇˇIp ` |ψNW |pT q “ ppp`1q ppp`1q . (4.1) n ˇ Γp n2 π 2 n 4 Two remarks are in order. First, we are unaware of any matrix calculus tools that could let us integrate this expression directly. Thus, the mere fact that this expression integrates to unity, a consequence of being the G-conjugate of another distribution, seems remarkable.? Second, when p “ 1, this is the tn{2 { 8 distribution. Thus, as we mentioned while discussing Equation (3.24), it is natural to interpret this distribution as the parametrization of some generalization of the t distribution to Sp pRq, the space of real-valued symmetric matrices. The purpose of this section is to propose a candidate definition for such generalization, as well as prove several results concerning the normalized Wishart G-conjugate. To the best of our knowledge, no extension of the t distribution to symmetric matrices has ever been proposed. However, a non-symmetric matrix variate t distribution has been thoroughly investigated in the literature – see Gupta and Nagar [1999, Chapter 4] for a thorough summary. Several definitions exist. For our purposes, we say that a p ˆ q real-valued random matrix T has the matrix variate t distribution with ν degrees of freedom and q ˆ q positive-definite scale matrix Ω if it has density ¯ ´ ˇ ˇ´ ν`p`q´1 Γp ν`p`q´1 2 pˇ 2 T Ω´1 T t ˇˇ ´ ˇ 2 ´ ¯ |Ω| I ` . p ˇ ˇ pq pq ν ν 2 π 2 Γ ν`p´1 p 2 It is not exactly clear what should be the proper analog of this distribution for symmetric matrices. But it would be elegant if the degrees of freedom of Equation (4.1) were to be exactly n{2, as in the univariate case. Thus, the following definition seems natural. Definition 2 (Symmetric matrix variate t distribution). We say a real symmetric p ˆ p matrix T has the symmetric matrix variate t distribution with ν ě p{2 ´ 1 degrees of freedom and p ˆ p positive-definite scale matrix Ω, Chételat and Wells/Mid-scale Wishart asymptotics 20 denoted Tν pΩq, if it has density ˇ´ ν`pp`1q{2 ˇ 2 ˇ T Ω´1 T ˇˇ ˇ . fTn pΩq pT q 9 ˇIp ` ˇ ν With this definition, the G-conjugate to the normalized Wishart distribution, whose density is given by Equation (4.1), is the Tn{2 pIp {8q distribution on Sp pRq. In fact, since Equation (4.1) integrates to one, we can deduce the normalization constant of Definition 2. For an arbitrary degrees of freedom parameter ν, imagine the density |ψNW | of the G-conjugate of a normalized Wishart distribution with n “ 2ν ě p ´ 2. By virtue of being a G-conjugate, it must integrate to 1 ? 1 unity. Then from the change of variables T “ Ω´ 4 SΩ´ 4 { 8 which has Jacobian ppp`1q p`1 dT “ 8´ 4 |Ω|´ 4 dS, we see that ż 1“ ż |ψNW |pT q dT “ Sp pRq 2ppν´1q Γ2p “ π ppp`1q 2 ν ´ 1 2 3ppp`1q 4 ν`pp`1q{2 2 ppp`1q 4 |Ω| p`1 4 ˆ ´1 1 ˙ Ω 4 SΩ´ 4 ? dS |ψNW | 8 Sp pRq ż ¯ |Ω| ´ p`1 4 Γp pνq ˇ ˇ ν`pp`1q{2 ´1 ˇ´ 2 ˇ ˇIp ` SΩ S ˇ dS. ˇ ˇ ν Sp pRq Thus we must have fTn pΩq pT q “ 2ppν´1q Γ2p ´ ν`pp`1q{2 2 ¯ ˇ ˇ´ ν`pp`1q{2 2 ˇ T Ω´1 T ˇˇ I ` . p ˇ ˇ ν ´ p`1 4 ˇ |Ω| (4.2) ppp`1q ppp`1q π 2 ν 4 Γp pνq It would be interesting to see if this distribution satisfies the properties we would expect of a t distribution, to ensure our guess is the “correct” one. However, this would take us too far away from the topic of this article. Instead, we will focus in the rest of this section on proving results about Tn{2 pIp {8q, the G-conjugate to the normalized Wishart distribution. Our first result will concern the asymptotic expansion of its normalization constant. We mention that this constant is the same as the Cn,p term appearing in the expression of the G-transform of the normalized Wishart in Proposition 3. Lemma 1. The normalization constant of the Tn{2 pIp {8q distribution ` ˘ ppn`2pq Γ2p n`p`1 2 2 4˘ ` Cn,p “ ppp`1q ppp`1q Γp n2 π 2 n 4 has, for every K P N, the asymptotic expansion " K`1 pp3p`1q 2 4 1 ÿ 1rk evens pk`2 Cn,p “ ppp`1q exp ´ 2 k“1 kpk`1qpk`2q nk π 4 (4.3) K`1 ´ pK`3 ¯* 1 ÿ 1`21rk evens pk`1 ´ ` o K`1 . 4 k“1 kpk`1q nk n as n Ñ 8 with p{n Ñ 0. Chételat and Wells/Mid-scale Wishart asymptotics 21 Proof. By Stirling’s approximation applied to log Γ, as well as Muirhead [1982, Theorem 2.1.12], we find that ˙ ˆ p ÿ ppp ´ 1q i´1 log Γp pxq “ log π ` log Γ x ´ 4 2 i“1 „ˆ ˙ ˆ ˙ p ÿ ppp ´ 1q i´1 1 i´1 “ log π ` x´ ´ log x ´ 4 2 2 2 i“1 ˙ ˆ ˙ ˆ 1 1 i´1 ` log 2π ` O ´ x´ 2 2 x ppp ` 1q p ppp ´ 1q “ log π ` log 2 ´ px ` 4 2 4 ˙ ˆ ˙ p ˆ ´p¯ ÿ i i´1 ` log x ´ `O x´ 2 2 x i“1 as x Ñ 8. Thus ˙ ˆ ´ n ¯ ppp ` 1q p n`p`1 ´ log Γp “ log π ` log 2 2 log Γp 4 2 4 2 ˆ ˙ ˆ ˙ p ppp ` 3q ÿ n`p`1 n`p`1 i´1 ´ ` ´ i log ´ 4 2 4 2 i“1 ˙ ˆ ˙ ˆ p ÿ n i n i´1 ´ log ´ ` op1q ´ 2 2 2 2 i“1 “ ppp ` 1q pp2n ` p ´ 1q ppp ` 1q ppp ` 3q log π ´ log 2 ` log n ´ 4 4 4 4 ˆ ˙ p ¯ 1 ÿ´ 2i´p´3 ` n ´ r2i´p´1s log 1 ´ 2 i“1 n ˆ ˙ p i´1 1ÿ pn ´ iq log 1 ´ ` op1q, ´ 2 i“1 n as n Ñ 8 with p{n Ñ 0, and so by Equation (4.3), pp3p ` 1q ppp ` 1q ppp ` 3q log Cn,p “ log 2 ´ log π ´ 4 4 4 ˙ p ¯ ˆ 1 ÿ´ 2i´p´3 ` n ´ r2i´p´1s log 1 ´ 2 i“1 n ˆ ˙ p 1ÿ i´1 ´ pn ´ iq log 1 ´ ` op1q. (4.4) 2 i“1 n Let us now focus on the two sums in this expression. Recall that for any k ě 1, x2 x3 xk ´ logp1 ´ xq “ x ` ` ` ¨¨¨ ` ` Opxk`1 q as x Ñ 0, (4.5) 2 3 k even for negative x. Therefore, ˆ ˙ p 1ÿ i´1 ´ pn ´ iq log 1 ´ 2 i“1 n Chételat and Wells/Mid-scale Wishart asymptotics 22 ˆ K`4 ˙ p p K`2 K`1 1 ÿ ÿ pi´1qk 1 ÿ ÿ pi´1qk p ´ ` O n i 2 i“1 k“1 k nk 2 i“1 k“1 k nk nK`2 ˆ K`3 ˙ ˙ p p K`1 ˆ p p 1ÿ 1 ÿ ÿ i´1 i pi´1qk `O “ pi´1q ` ´ k K`1 2 i“1 2 i“1 k“1 k´1 k n n n ˆ ˙ ˙ ˆ p K`1 pK`3 ppp ´ 1q 1 ÿ ÿ 1 pi´1qk i´1 `o “ ´ ` k 4 2 i“1 k“1 kpk`1q k n nK`1 ˆ K`3 ˙ p´1 p´1 K`1 K`1 ÿ ppp ´ 1q 1 ÿ 1 1 ÿ 1 ÿ k p k`1 “ ´ i i . ´ ` o k k 4 2 k“1 kpk`1qn i“1 2 k“1 k n i“1 nK`1 “ Now let Bk denote the Bernoulli numbers, with the convention B1 “ 12 . Faulhaber’s formula provides K`1 k`2 ppp ´ 1q 1 ÿ 1 1 ÿ “ ´ ¨ Bk`2´l pp´1ql 4 2 k“1 kpk`1qnk k`2 l“1 ˆ K`3 ˙ K`1 k`1 1 ÿ p 1 ÿ 1 l ¨ Bk`1´l pp´1q ` o . ´ 2 k“1 k nk k`1 l“1 nK`1 But by the binomial theorem, pk`1 n ` op1q and pp´1ql “ op1q nk K`1 ÿ „ˆ pp´1qk`2 nk “ pk`2 nk k`1 ´ pk ` 2q pnk ` op1q, pp´1qk`1 nk “ for any 1 ď l ď k. Thus ˙ k`2 B0 pk`2 ´ pk`2qpk`1 ppp ´ 1q 1 ´ “ 4 2 k“1 k`2 kpk`1qpk`2q nk ˆ ˙  k`2 B1 pk`1 ` ` op1q k`1 kpk`1qpk`2q nk ˙  ˆ K`3 ˙ K`1 „ˆ 1 ÿ k`1 B0 pk`1 p ` op1q ` o . ´ 2 k“1 k`1 kpk`1q nk nK`1 Using that B0 “ 1 and B1 “ 12 , we obtain “ K`1 ppp ´ 1q 1 ÿ 1 pk`2 ´ 4 2 k“1 kpk`1qpk`2q nk ˆ K`3 ˙ K`1 1 ÿ 1 pk`1 p ´ `o . 4 k“1 kpk`1q nk nK`1 (4.6) The analysis of the other sum of Equation (4.4) is similar but more involved, as we must distinguish the cases where p is even and where p is odd. We find, from Equation (4.5) again, that ˙ p ¯ ˆ 1 ÿ´ 2i´p´3 n ´ r2i´p´1s log 1 ´ 2 i“1 n ˆ ˙ ˆ ˙ p p 1ÿ 1ÿ 2i´p´3 2i´p´3 “´ ` p2i´p´1q log 1 ´ n log 1 ´ 2 i“1 n 2 i“1 n Chételat and Wells/Mid-scale Wishart asymptotics 23 ˆ K`4 ˙ p p K`2 K`1 ÿ p2i´p´3qk 1ÿ 1 ÿ ÿ p2i´p´3qk p ´ ` O p2i´p´1q n k k K`2 2 i“1 k n 2 k n n i“1 k“1 k“1 ˙ p p K`1ˆ 1ÿ 1 ÿ ÿ 2i´p´1 2i´p´3 p2i´p´3qk “ p2i´p´3q ` ´ 2 i“1 2 k“1 k“1 k k`1 nk ˆ K`3 ˙ p p `O K`1 n n ˆ K`3 ˙ ˆ ˙ p K`1 k ÿ ÿ p 2i´p´3 2 p2i´p´3q 1 `o ` “p` 2 i“1 k“1 kpk`1q k nk nK`1 “ “p` p K`1 ÿ 1 ÿ 1 p2i´p´3qk`1 2 k“1 kpk`1qnk i“1 ˆ K`3 ˙ p 1 ÿ p k ` p2i´p´3q ` o k K`1 kn n i“1 k“1 „  p ÿ k`1 k`1 k`1 p´p´1q ` p´p`1q ` p2i´p´3q k K`1 ÿ K`1 1 ÿ 1 2 k“1 kpk`1qn i“3  ˆ K`3 ˙ p K`1 ÿ 1 „ ÿ p k k k ` p´p´1q ` p´p`1q ` p2i´p´3q ` o . k kn nK`1 i“3 k“1 “p` (4.7) At this point, it is simpler to analyze the cases where p is even and odd separately. If p is odd, define q “ pp´3q{2 and observe that by Faulhaber’s formula, p q ” ıÿ ÿ p2i´p´3ql “ 1`p´1ql p2iql i“3 i“1 “ 1rl evens l`1 By the binomial theorem, op1q and p2qql “ nk k`1 p´p´1q nk l`1 ÿ ˆ s“1 p2qqk`2 nk “ ˙ l`1 Bl`1´s 2l`1´s p2qqs . s pk`2 nk k`1 ´ 3pk`2q pnk ` op1q, p2qqk`1 nk “ pk`1 nk ` op1q for 1 ď l ď k. Moreover, “ p´p`1qk`1 p´pqk`1 “ ` op1q k n nk (4.8) and p´p´1qk p´p`1qk “ “ op1q. (4.9) nk nk Thus, for odd p, Equation 4.7 equals „ ˆ ˙ K`1 1 ÿ 1 k`2 B0 pk`2 ´ 3pk`2qpk`1 “p` ` 1rk odds 2 k“1 kpk`1q k`2 k`2 nk ˆ ˙  k`2 2B1 pk`1 p´pqk`1 ` 1rk odds `2 ` op1q nk k`1 k`2 nk Chételat and Wells/Mid-scale Wishart asymptotics ` 24  ˆ K`3 ˙ ˆ ˙ K`1 „ 1 ÿ 1 p k`1 B0 pk`1 ` op1q ` o 1rk evens . 2 k“1 k k`1 k`1 nk nK`1 Moreover, 2p´pqk`1 ´ 31rk odds pk`1 ` 1rk odds pk`1 ` 1rk evens pk`1 “ ´1rk evens pk`1 . Thus, ˆ K`3 ˙ K`1 K`1 p 1 ÿ 1rk odds pk`2 1 ÿ 1rk evens pk`1 ´ `o K`1 . “p` 2 k“1 kpk`1qpk`2q nk 2 k“1 kpk`1q nk n (4.10) When p is even, let q “ pp ´ 2q{2 and observe that by Faulhaber’s formula, ˆÿ ˙ p q 2q q ” ıÿ ÿ ÿ l l l l l p2i´p´3q “ 1`p´1q p2i ´ 1q “ 21rl evens i ´ p2iq i“3 i“1 “ 1rl evens l`1 But by the binomial theorem, p2qql nk pk`1 nk l`1 ÿ s“1 ˆ i“1 i“1 ˙ l`1 Bl`1´s p2 ´ 2l`1´s q p2qqs . s p2qqk`2 nk “ pk`2 nk k`1 ´ 2pk `2q pnk ` op1q, p2qqk`1 nk “ ` op1q and “ op1q for 1 ď l ď k. If we apply Equations (4.8)–(4.9), then Equation (4.7) becomes „ ˆ ˙ K`1 1 k`2 p2 ´ 1qB0 pk`2 ´ 2pk ` 2qpk`1 1 ÿ 1rk odds “p` 2 k“1 kpk`1q k`2 k`2 nk ˆ ˙  k`2 p2 ´ 2qB1 pk`1 ` 1rk odds ` op1q k`1 k`2 nk „ ˆ ˙  ˆ K`3 ˙ K k`1 p2 ´ 1qB0 pk`1 p 1 ÿ 1 1rk evens ` op1q `o K`1 . ` k 2 k“1 k k`1 k`1 n n Moreover, 2p´pqk`1 ´ 21rk odds pk`1 ` 1rk evens pk`1 “ ´1rk evens pk`1 . Thus again, “p` ˆ K`3 ˙ K`1 K`1 1 ÿ 1rk odds pk`2 1 ÿ 1rk evens pk`1 p ´ `o . k k 2 k“1 kpk`1qpk`2q n 2 k“1 kpk`1q n nK`1 (4.11) which is the exact same result as in the odd p case (see Equation 4.10). Plugging Equations (4.6) and (4.10)–(4.11) in Equation (4.4), we obtain K`1 ppp ` 1q 1 ÿ pp3p ` 1q 1rk evens pk`2 log 2 ´ log π ´ log Cn,p “ 4 2 2 k“1 kpk`1qpk`2q nk ˆ K`3 ˙ K`1 1 ÿ 1 ` 21rk evens pk`1 p ´ `o K`1 , 4 k“1 kpk`1q nk n as desired. Chételat and Wells/Mid-scale Wishart asymptotics 25 Thus the constant Cn,p is closely related to the normalization constant of the GOEppq distribution 2pp3p`1q{4 {π ppp`1q{2 . We now turn our attention to the study of the asymptotic moments of a Tn{2 pIp {8q distribution. We first remind the reader of some classic results. For a Gaussian Orthogonal Ensemble matrix Z „ GOEppq, a moment-based approach to Wigner’s theorem states that for any k P N, its k th moment satisfy  „ 1 ´ Z ¯k “ Ck{2 1rk evens , lim E tr ? pÑ8 p p `2k˘ 1 where Ck “ k`1 is the k th Catalan number. In fact, Anderson et al. k [2010, section 2.1.4 on‰ p.17] show that the variance of the k th moment satisfies “1 ? lim Var p trpZ{ pqk “ 0, so we really have pÑ8 1 ´ Z ¯k L2 tr ? ÝÑ Ck{2 1rk evens p p as p Ñ 8. “ Now,‰ what do we know about the moments of Tn{2 pIp {8q? By symmetry, E tr T k “ 0 for odd k, but it is much less clear what happens for even k. It turns out that in many ways, if T „ Tn{2 pIp {8q then 4T „ Tn{2 p2Ip q mimics the Gaussian Orthogonal Ensemble results outlined above, especially when p{n Ñ 0 as n Ñ 8. We have the following result. Theorem 2. Let k P N and T „ T“n{2 pIp {8q. ‰ If p{n Ñ c P r0,“1q, the‰moments of T satisfy the asymptotic bounds E tr T 2k “ Oppk`1 q and E tr2 T k “ Oppk`2 q as n Ñ 8. In fact, for any k P N, 1 ´ 4T ¯k L2 tr ? ÝÑ Ck{2 1rk evens p p `2k˘ 1 th Catalan number. as n, p Ñ 8 with p{n Ñ 0, where Ck “ k`1 k is the k Although our proof will rely on the close relationship between the Wishart and the t distribution, it is worthwhile to step back and think why a Tn{2 p2Ip q should behave like a GOEppq when p{n Ñ 0. One good reason might be the classic result that as n Ñ 8, the density of a t distribution converges pointwise to a standard normal density. Thus, we might think that as long as p does not grow too fast, in some aspects the symmetric t distribution should behave like a GOEppq. In the context of the proof, it will prove useful to use the notion of power sum symmetric polynomials. For any integer partition κ “ pκ1 , . . . , κq q in decreasing order κ1 ě ¨ ¨ ¨ ě κq ą 0, define its associated power sum polynomial to be q ź rκ pZq “ tr Z κi . (4.12) i“1 The norm of the partition κ is |κ| “ κ1 ` ¨ ¨ ¨ ` κq ą 0, which should not be confused with its length qpκq “ q (number of elements). By convention, we will assume there also exists an empty partition ∅ “ pq with length qp∅q “ 0, norm |∅| “ 0 and power sum polynomial r∅ pZq “ 1. Chételat and Wells/Mid-scale Wishart asymptotics 26 Let’s now turn to the proof of the theorem. The odd moments of the Tn{2 pIp {8q moments sense to focus on the even moments “ ‰ are zero by symmetry, so it makes “ ‰ E tr T 2k and the square moments E tr2 T k . Our first step in the proof is to express these in terms of expectations of power sum polynomials of an inverse Wishart Y ´1 „ W´1 p pn, Ip {nq, where by power sum polynomials we mean expressions like at Equation (4.12). Recall the useful shorthand m “ n ´ p ´ 1. Lemma 2. Let T „ Tn{2 pIp {8q. Then for any k P N, whenever n is large enough so that n ě p ` 16k ` 6, we can compute the 2k th moment of T by ż np “ ‰ p´1qk E tr T 2k “ nk p ÿ ¨ i1 ,...,i2k ! n )ˇ ˇ m n2 ` n ˘ exp ´ tr Y ˇY ˇ 4 np 4 2 2 Γp 2 Y ą0 ˇ )ˇ ! n ˇ m4 ˇ Bs Bs Bs ˇ ˇ ... dY exp ´ trpY `Xq Y `X ˇˇ Bs Xi1 i2k Bs Xi3 i2 Bs Xi2 i1 4 X“0 “ “ ‰ p´1qk ÿ p1q bκ pn, m, pq E rκ pY ´1 q k n (4.13) |κ|ď2k and its squared k th moment by ż np “ ‰ p´1qk E tr2 T k “ nk ! n )ˇ ˇ m n2 ` n ˘ exp ´ tr Y ˇY ˇ 4 4 2 Γp 2 np 2 p ÿ Y ą0 Bs Bs Bs Bs Bs Bs ... ... B X B X B X B X B X B X s j3 j2 s j2 j1 s i1 ik s i3 i2 s i2 i1 i1 ,...,ik s j1 jk j1 ,...,jk ˇ )ˇ ! n ˇ m4 ˇ ˇ ˇ dY exp ´ trpY `Xq Y `X ˇˇ 4 X“0 k ÿ “ ‰ p´1q “ bκp2q pn, m, pq E rκ pY ´1 q , (4.14) k n ¨ |κ|ď2k`1 for Y ´1 „ W´1 p pn, Ip {nq p1q p2q p1q p2q and some bκ , bκ . These bκ , bκ are polynomials in p1q n, m, p, indexed by integer partitions κ, whose degrees satisfy deg bκ ď 2k ` 1 ´ p2q qpκq and deg bκ ď 2k ` 2 ´ qpκq. The sums are taken over all partitions of the integers κ satisfying |κ| ď 2k and |κ| ď 2k ` 1 respectively, including the empty partition. Proof. Let fNW and ψNW ? stand for the density and the G-transform of a normalized Wishart matrix nrWp pn, Ip {nq ´ Ip s. In the proof of Proposition 3, we 1{2 concluded at Equation (3.11) that fNW had to be integrable when n ě p ´ 2, as its integral was proportional to a multivariate gamma function. Let RpXq “ ´X 1{2 1{2 be the flip operator. Since fNW is integrable, fNW ˝R must be integrable as well, Chételat and Wells/Mid-scale Wishart asymptotics 27 ` 1{2 ˘ 1{2 and so their convolution fNW ‹ fNW ˝ R is well-defined and integrable. At Equation (3.1) at the start of Section 3, we defined our notion of Fourier transform for integrable functions on Sp pRq. Define the map ι : Sp pRq Ñ Rppp`1q{2 that maps a symmetric matrix to its vectorized upper triangle, and let τ : Sp pRq Ñ Sp pRq be the map # 2Xij if i ‰ j τ pXqij “ Xjj if i “ 1. Then in terms of the usual Fourier transform on Rppp`1q{2 , (` ˘ ppp´1q Ftf upT q “ 2 4 F f ˝ ι´1 ι ˝ τ pT q . This close relationship transfer properties to our Fourier transform on Sp pRq. We will need three. ( 1. For any integrable function f , we have F f ˝ R “ Ftf u. 2. (Convolution) For any two integrable functions f1 and f2 , we have Ftf1 ‹ ppp`1q p f2 u “ 2 2 π 4 Ftf1 uFtf2 u. 3. (Fourier inversion) For any continuous integrable f with integrable Fourier transform φ, we have ż f pXq “ 1 p 2 2 π ppp`1q 4 ei trpT Xq φpT q dT “ Ftφup´Xq, Sp pRq for all X P Sp pRq. These properties are important for the following. Since fNW is real-valued, properties 1 and 2 provide ! ) ` 1{2 ˘) ( ppp`1q p 1{2 1{2 1{2 2´ 2 π ´ 4 F fNW ‹ fNW ˝ R “ F fNW F fNW ˝ R ˇ 1{2 ˇ 1{2 ( 1{2 1{2 1{2 ( “ F fNW F fNW “ ψNW ψNW “ ˇψNW ˇ. But then, since |ψNW | is integrable (in fact, to unity), the Fourier inversion formula yields that ż ` 1{2 ˘ 1{2 fNW ‹ fNW ˝ R pXq “ ˇ ˇ ei trpT Xq ˇψNW pT qˇ dT. (4.15) Sp pRq Thus we might say the characteristic function of the Tn{2 pIp {8q distribution is ˘ 1{2 ` 1{2 given by fNW ‹ fNW ˝ R . It is well known that the derivatives of the characteristic function of a distribution evaluated at zero provide its moments, up to a ˘ 1{2 ` 1{2 constant. This suggests should try“ to repeatedly differentiate fNW ‹ fNW ˝R “ we2k ‰ ‰ at zero to compute E tr T and E tr2 T k , our ultimate goal. Unfortunately, the convolution is given by an integral whose domain makes it difficult to directly interchange the differentiation and integration symbols. Because the integrand is orthogonally invariant, we found it easier to compute the derivatives at zero by taking a limit over a sequence of decreasing positive- Chételat and Wells/Mid-scale Wishart asymptotics 28 definite matrices at both sides instead. In this spirit, define on the open set t0 ă X ă Ip u Ă Sp pRq the real-valued functions p´1qk H1 pXq “ nk p ÿ ˘? Bs Bs Bs 1{2 ` 1{2 ... fNW ‹ fNW ˝R p nXq Bs Xi1 i2k Bs Xi3 i2 Bs Xi2 i1 i1 ,...,i2k and p´1qk H2 pXq “ nk p ÿ Bs Bs Bs Bs Bs Bs ... ... Bs Xj1 jk Bs Xj3 j2 Bs Xj2 j1 Bs Xi1 ik Bs Xi3 i2 Bs Xi2 i1 i1 ,...,ik j1 ,...,jk for fixed k, p and n. Here ˘? 1{2 ` 1{2 fNW ‹ fNW ˝R p nXq Bs Bs Xij stands for the symmetric differentiation operator ? “ as defined in Section 2. The n scaling in the argument helps link the convolution to an expectation with respect to an inverse Wishart distribution. Let’s first relate these functions to the moments of the Tn{2 pIp {8q distribution. The symmetric differentation operator has the pleasant property that Bs Bs Xij trpXT q “ Tij for any two symmetric matrices X, T . Thus, for any 1 ď l ď 2k and indices 1 ď i1 , . . . , i2l ď p, we find that ˇ ˇ ? ˇ ˇ ˇ ˇ Bs Bs Bs i n trpT Xq ˇ ˇ ˇ ... ψNW pT qˇˇ e ˇ Bs Xi i Bs Xi4 i3 Bs Xi2 i1 2l 2l´1 ˇ ˇˇ ˇ ˇ ˇ “ nl ˇTi2l i2l´1 ¨ ¨ ¨ Ti4 i3 Ti2 i1 ˇˇψNW ˇpT q (4.16) 1`δij B 2 BXij , Bs Bs Xij for all X P Sp pRq. We now show that the right hand side (4.15) is integrable. This is not a mere formality: when p “ 1, asking if this expression is integrable is the same as asking if the t distribution with n{2 degrees of freedom has an lth moment, and it is well-known that the t distribution only possesses moments of order smaller than its degrees of freedom. So the answer is most likely to be positive, but only for n large enough. Let us see why. For any symmetric matrix T , g f p ´ ¯ bˇ a fź ˇ 2 |Tij | ď λ1 pT q ď e 1 ` λi pT 2 q “ ˇIp ` T 2 ˇ, i“1 where λ1 pT 2 q ě ¨ ¨ ¨ ě λp pT 2 q ě 0 are the ordered eigenvalues of the positivedefinite matrix T 2 . Thus ż ˇ ˇˇ ˇ ˇ ˇ nl ˇTi2l i2l´1 ¨ ¨ ¨ Ti2 i1 ˇˇψNW ˇpT q dT Sp pRq ż 3l 2 n “ l Cn,p 4 ˇ´ n`p`1 ˇ ˇ 4T 4 4Ti2 i1 ˇˇˇˇ 16T 2 ˇˇ ˇ i2l i2l´1 ¨ ¨ ¨ ? ˇˇIp ` dT ˇ ? ˇ n n n Sp pRq Chételat and Wells/Mid-scale Wishart asymptotics ż ˇ ˇ pn´2lq`p`1 2 ˇ´ 4 ˇ ˇIp ` 16T ˇ dT ˇ n ˇ 3l 2 n ď l Cn,p 4 Sp pRq 3l 2 ˆ n Cn,p ď l 4 Cn´2l,p 29 ż n n ´ 2l ˙ppp`1q 4 ˇ ˇ´ pn´2lq`p`1 4 ˇ 16T 2 ˇˇ ˇ Cn´2l,p ˇIp ` dT. n ´ 2l ˇ (4.17) Sp pRq When n ´ 2l ě p ´ 2, the last integrand is the density of a Tn{2´l pIp {8q distribution, so integrates to unity. Thus, when n ě p ` 4k ´ 2, the right hand side of Equation (4.16) is an integrable function for all 1 ď l ď 2k and 1 ď i1 , . . . , i2l ď p. By Equation (4.15), and repeated differentiation under the integral sign justified by the integrability bounds given by Equations (4.16) and (4.17), we find that k H1 pXq “ p´1q nk ż p ÿ ? Bs Bs Bs ... Bs Xi1 i2k Bs Xi3 i2 Bs Xi2 i1 ei i1 ,...,i2k ˇ n trpT Xq ˇ ˇ ψNW pT qˇdT Sp pRq ż tr T 2k ei “ ? ˇ n trpT Xq ˇ ˇ ψNW pT qˇdT (4.18) Sp pRq and p´1qk H2 pXq “ nk p ÿ Bs Bs ... Bs Xj1 jk Bs Xj2 j1 i1 ,...,ik j1 ,...,jk ż Bs Bs Bs ... Bs Xi1 ik Bs Xi3 i2 Bs Xi2 i1 ei ˇ ψNW pT qˇdT Sp pRq ż “ ? ˇ n trpT Xq ˇ tr2 T k ei ? ˇ n trpT Xq ˇ ˇ ψNW pT qˇdT. (4.19) Sp pRq for any X P Sp pRq and any n ě p ` 4k ´ 2. ˘ 1{2 ` 1{2 Now let’s relate H1 and H2 to the definition of fNW‹ fNW˝R as a convolution. This is where restricting H1 and H2 to small positive-definite matrices becomes useful. By Equation (3.10), the expression is ż ˘? 1{2 ` 1{2 fNW ‹ fNW ˝R p nXq “ ż np “ n2 ` ˘ np 2 2 Γp n2 1{2 1{2 fNW pZqfNW pZ ´ ? nXqdZ Sp pRq ! n )ˇ ! n )ˇ ˇ m ˇm exp ´ trpY `Xq ˇY `X ˇ 4 exp ´ tr Y ˇY ˇ 4 dY 4 4 Y `Xą0, Y ą0 ? ppp`1q using the change of variables Y “ Ip ` Z{ n ´ X with dZ “ n 4 dY . For Chételat and Wells/Mid-scale Wishart asymptotics 30 X ą 0, we have 1rY ` X ą 0, Y ą 0s “ 1rY ą 0s, and thus H1 , H2 satisfy H1 pXq “ p ÿ p´1qk nk ż Bs Bs Bs ... Bs Xi1 i2k Bs Xi3 i2 Bs Xi2 i1 i1 ,...,i2k and p´1qk H2 pXq “ nk p ÿ Y ą0 np ˇm ˇ ¨ˇY `X ˇ 4 ! n ) exp ´ trpY `Xq 4 ! n )ˇ ˇ m n2 ` n ˘ exp ´ tr Y ˇY ˇ 4 dY np 4 2 2 Γp 2 (4.20) Bs Bs Bs ... Bs Xj1 jk Bs Xj3 j2 Bs Xj2 j1 i1 ,...,ik j1 ,...,jk ż Bs Bs Bs ... Bs Xi1 ik Bs Xi3 i2 Bs Xi2 i1 ! n )ˇ ˇm exp ´ trpY `Xq ˇY `X ˇ 4 4 Y ą0 np 2 ! n )ˇ ˇ m n ` n ˘ exp ´ tr Y ˇY ˇ 4 dY. (4.21) 4 2 Γp 2 We would now like to interchange the integral and differentiation signs. )To do m so, we must understand what the repeated derivatives of expt´ n4 tr Y |Y | 4 look like. Differentiating once, we see that: ! n )ˇ ˇm Bs exp ´ trpY `Xq ˇY `X ˇ 4 Bs Xi2 i1 4  „ ! n )ˇ ˇm n m ´1 pY `Xqi2 i1 ´ pIp qi2 i1 exp ´ trpY `Xq ˇY `X ˇ 4 . “ 4 4 4 Differentiating twice, we see that: ! n )ˇ ˇm Bs Bs exp ´ trpY `Xq ˇY `X ˇ 4 Bs Xi4 i3 Bs Xi2 i1 4 „ m m ´1 ´1 “ ´ pY `Xq´1 pY `Xq´1 i2 i4 pY `Xqi3 i1 ´ i2 i3 pY `Xqi4 i1 8 8 mn m2 ´1 ´1 ` pY `Xq´1 pY `Xq´1 i4 i3 pY `Xqi2 i1 ´ i4 i3 pIp qi2 i1 16 16  ! n )ˇ ˇm n2 mn ´1 ´1 ´1 ´ pIp q´1 pY `Xq ` pI q pI q trpY `Xq ˇY `X ˇ 4. p i4 i3 p i2 i1 exp ´ i4 i3 i2 i1 16 16 4 So in general, it is clear that the repeated derivatives are) given by some polym nomial in entries of pY ` Xq´1 , times expt´ n4 trpY ` Xq |Y ` X| 4 . We won’t investigate further the nature of these polynomials beyond remarking that for any indices 1 ď l ď 2k and 1 ď i1 , . . . , i2l ď p, and any symmetric matrices X, Y P Sp pRq, we must have some crude bound ˇ ˇ * " ˇ ˇ ˇ m4 ˇ Bs Bs Bs n ˇ ˇ ˇ ... exp ´ trpY `Xq Y `X ˇˇ ˇ Bs Xi i Bs Xi i Bs Xi i 4 ¨ 2l 2l´1 4 3 2 1 np 2 Chételat and Wells/Mid-scale Wishart asymptotics l ÿ ÿ ď 31 s l ˇ ź ˇ ˇ ˇ ˇ ˇź ˇ ˇ ˇpY `Xq´1 ˇ ˇ ˇ pI q pn, mq ˇ ˇaJ,s p j2t j2t´1 j2t j2t´1 "t“1 * ˇ ˇm n ¨ exp ´ trpY `Xq ˇY `X ˇ 4 4 for some polynomials aJ,s that do not depend on X or Y . We relegate a proof of this result as Lemma 3 in Section 8. This can be uniformly bounded for all 0 ď X ď Ip by l ! n )“ ÿ ‰ mp ď C1 pn, m, pq trs pY ´1 q exp ´ tr Y 1 ` tr Y 4 (4.22) 4 s“0 s“0 t“s`1 JP t1,...,pu2l for some constant C1 pn, m, pq that does not depend on X or Y . But for any n ě p ´ 2 and l ě 0, ż l ÿ )“ ! n ‰ mp trs pY ´1 q exp ´ tr Y 1 ` tr Y 4 4 s“0 np ! n )ˇ ˇ m n2 ¨ np ` n ˘ exp ´ tr Y ˇY ˇ 4 dY 4 2 2 Γp 2 C1 pn, m, pq Y ą0 ż “ n np 2 l C1 pn, m, pq ÿ `n˘ np 2 2 Γp 2 s“0 “ 1`tr Y ‰ mp 4 trs pY ´1 q Y ą0 ! n )ˇ ˇ n`p`1 p`1 ´ 2 exp ´ tr Y ˇY ˇ 4 dY 2 ` ˘ ” ı ´ n ¯ mp Γp n`p`1 mp 4 4˘ ` E p1 ` tr Y q 4 trs pY ´1 q “ n 2 Γp 2 ` ˘ for a Y with a matrix gamma distribution Gp n`p`1 , n2 Ip . The Cauchy-Schwarz 4 inequality then entails the bound ˘ ” ` ı 12 ´ n ¯ mp ı 21 ” Γp n`p`1 mp 4 4 ` n ˘ E p1 ` tr Y q 2 E tr2s pY ´1 q . (4.23) ď 2 Γp 2 The first expectation is always finite when n ě p ´ 2. Since tr2s pY ´1 q can be written as a sum of zonal polynomials indexed by partitions of the integer 2s, the results of Muirhead [1982, Theorem 7.2.13] imply that the second expectation is finite whenever n`p`1 ą 2s ` p´1 4 2 ô n ě p ` 8s ´ 2. Thus, in Equation (4.20) with l ď k and (4.21) with l ď 2k, whenever n ě p ` 16k ´ 2 we are justified in repeatedly differentiating under the integral sign by the integrability bounds given by Equations (4.22) and (4.23), and obtain in that case ż p´1qk H1 pXq “ nk p ÿ Y ą0 i1 ,...,i2k Bs Bs Bs ... Bs Xi1 i2k Bs Xi3 i2 Bs Xi2 i1 (4.24) Chételat and Wells/Mid-scale Wishart asymptotics 32 np ! n )ˇ ! n )ˇ ˇ m ˇ m4 n2 ˇ ˇ ` n ˘ exp ´ tr Y ˇY ˇ 4 dY exp ´ trpY `Xq Y `X np 4 4 2 2 Γp 2 (4.25) and ż p ÿ p´1qk H2 pXq “ nk i ,...,ik Y ą0 1 Bs Bs Bs ... Bs Xj1 jk Bs Xj3 j2 Bs Xj2 j1 j1 ,...,jk ! n )ˇ ˇm Bs Bs Bs ... exp ´ trpY `Xq ˇY `X ˇ 4 Bs Xi1 ik Bs Xi3 i2 Bs Xi2 i1 4 np ! n )ˇ ˇ m n2 ¨ np ` n ˘ exp ´ tr Y ˇY ˇ 4 dY. 4 2 2 Γp 2 (4.26) Let us now look at how H1 pXq and H2 pXq behave as X Ñ 0. On one hand, a ? for any symmetric matrix T we have | tr T k | ď p tr T 2k ď p|Ip ` T 2 |k{2 , so we must have the bounds ˇ ˇ ˇ ˇ pn´4kq`p`1 k 2 ˇ´ 4 ? ˇ ˇ ˇˇ ˇ n C 16T n,p 2k i n trpT Xq ˇ tr T e ˇψNW pT qˇˇ ď ˇ ˇ ˇ 16k Cn´4k,p Cn´4k,p ˇIp ` n ˇ ˇ and ˇ ˇ ˇ pn´4kq`p`1 ˇ k 2 ˇ´ 4 ˇˇ ˇ 2 k i?n trpT Xq ˇ ˇ pn C 16T n,p ˇψNW pT qˇˇ ď ˇ tr T e ˇ ˇ . ˇ 16k Cn´4k,p Cn´4k,p ˇIp ` n ˇ ˇ holding uniformly in X. When n ´ 4k ě p ´ 2 ô n ě p ` 4k ´ 2, the right hand sides are proportional to the density of the G-conjugates of the normalized Wishart distributions with n ´ 4k degrees of freedom, so are integrable. Thus, by the dominated convergence theorem and Equations (4.18) and (4.19), ż lim XÑ0 H1 pXq 0ăXăIp “ ˇ ˇ “ ‰ lim ? tr T 2k XÑ0 ei n trpT Xq ˇψNW pT qˇdT “ E tr T 2k (4.27) 0ăXăIp Sp pRq and lim XÑ0 H2 pXq 0ăXăIp ż “ lim ei?n trpT Xq ˇˇψ pT qˇˇdT “ E“tr2 T k ‰ tr2 T k XÑ0 NW (4.28) 0ăXăIp Sp pRq for T „ Tn{2 pI{ 8q. On the other hand, the integrands at Equations (4.25) and (4.26) take a particularly simple form. Lemma 4 establishes by induction that there must be p1q p2q p1q polynomials bκ and bκ in n, m and p with degrees deg bκ ď 2k ` 1 ´ qpκq Chételat and Wells/Mid-scale Wishart asymptotics 33 p2q and deg bκ ď 2k ` 2 ´ qpκq such that ż p´1qk H1 pXq “ nk ÿ bp1q κ pn, m, pqrκ prY ` Xs |κ|ď2k Y ą0 " * n q exp ´ trpY ` Xq 4 np ˇ ˇm ¨ˇY ` X ˇ 4 and ´1 ! n )ˇ ˇ m n2 ` n ˘ exp ´ tr Y ˇY ˇ 4 dY 4 2 Γp 2 np 2 (4.29) ż p´1qk H2 pXq “ nk ÿ bp2q κ pn, m, pqrκ prY |κ|ď2k`1 ´1 ` Xs " * n q exp ´ trpY ` Xq 4 Y ą0 np ! n )ˇ ˇ m n2 ` n ˘ exp ´ tr Y ˇY ˇ 4 dY (4.30) 4 2 Γp 2 for any 0 ă X ă Ip and n ě p ` 16k ´ 2. The sums are taken over all partitions of the integers κ satisfying |κ| ď 2k and |κ| ď 2k ` 1 respectively, including the empty partition. But for any integer partition κ, the bound ˇm ˇ “ ‰ mp n n rκ prY ` Xs´1 qe´ 4 trpY `Xq ˇY ` X ˇ 4 ď tr|κ| pY ´1 qe´ 4 tr Y 1 ` tr Y 4 ˇ ˇm ¨ˇY ` X ˇ 4 np 2 holds uniformly in 0 ď X ď Ip . Thus for |κ| ď 2k ` 1, the right hand side is integrable for n ě p ` 16k ` 6, by the same argument as for Equation (4.23). Thus for such n, by the dominated convergence theorem and Equations (4.29) and (4.30), we obtain that ż lim XÑ0 H1 pXq 0ăXăIp “ p´1qk nk lim XÑ0 0ăXăIp Y ą0 ÿ ´1 bp1q q κ pn, m, pqrκ prY ` Xs |κ|ď2k " * np ! n )ˇ ˇ m ˇ ˇm n n2 ¨ exp ´ trpY ` Xq ˇY ` X ˇ 4 np ` n ˘ exp ´ tr Y ˇY ˇ 4 dY 4 4 2 2 Γp 2 k ÿ “ ‰ p´1q ´1 “ bp1q q (4.31) κ pn, m, pq E rκ pY nk |κ|ď2k and lim ż XÑ0 H2 pXq 0ăXăIp “ p´1qk nk lim XÑ0 0ăXăIp Y ą0 ÿ ´1 bp2q q κ pn, m, pqrκ prY ` Xs |κ|ď2k`1 " * np )ˇ ˇ m ! n ˇ ˇm n2 n ¨ exp ´ trpY ` Xq ˇY ` X ˇ 4 np ` n ˘ exp ´ tr Y ˇY ˇ 4 dY 4 4 2 2 Γp 2 k ÿ “ ‰ p´1q ´1 “ bp2q q , (4.32) κ pn, m, pq E rκ pY k n |κ|ď2k`1 Chételat and Wells/Mid-scale Wishart asymptotics 34 where Y follows a Wp pn, Ip {nq distribution. Combining Equations (4.27)–(4.28) with Equations (4.31)–(4.32) and Lemma 4 concludes the proof. Something remarkable about Lemma 2 is that it provides us with an algorithm to compute the moments of a symmetric t distribution in terms of the moments of an inverse Wishart matrix. For example, when k “ 1, repeated differentiation yields that p ! n )ˇ ÿ ˇm Bs Bs exp ´ trpY `Xq ˇY `X ˇ 4 BX BX 4 i1 ,i2 s i1 i2 s i2 i1 „ mpm ´ 2q mn n2 p “ trpY ` Xq´2 ´ trpY ` Xq´1 ` 16 8 16  ! n )ˇ ˇm m 2 ´ tr pY ` Xq´1 exp ´ trpY `Xq ˇY `X ˇ 4 . (4.33) 8 4 We can recognize trpY ` Xq´2 and tr2 pY ` Xq as the power sum polynomials rp2q prY ` Xs´1 q and rp1,1q prY ` Xs´1 q in the sense of Equation (4.12), so we p1q p1q p1q p1q must have bp2q “ mpm ´ 2q{16, bp1,1q “ ´m{8, bp1q “ mn{8 and b∅ “ n2 p{16 in the “ result ‰ of Lemma 4. Hence Lemma 2 really tells us that whenever n ě p ` 22, E tr T 2 for T „ Tn{2 pIp {8q can be expressed as “ ‰ m “ 2 ´1 ‰ ‰ mpm ´ 2q “ E tr Y ´2 ` E tr Y E tr T 2 “ ´ 16n 8n ‰ np m “ E tr Y ´1 ´ (4.34) ` 8 16 where Y „ Wp pn, Ip {nq. Of course, this also works with square moments and higher k. For example, the same strategy for, “ ‰ say, square moments with k “ 2 yields that whenever n ě p ` 38, E tr2 T 2 for T „ Tn{2 pIp {8q can be expressed as ` ˘ “ 2 2‰ ‰ m pm ´ 2q “ ‰ m m2 ´ 5m ` 10 “ E tr T “ ´ E tr Y ´4 ` E tr Y ´3 tr Y ´1 2 2 16n 4n ` ˘ m m3 ´ 4m2 ` 20m ´ 32 “ 2 ´2 ‰ m2 “ 4 ´1 ‰ E tr Y E tr Y ` ` 2 256n˘ 64n2 ` 2 ‰ m pm ´ 2q “ ‰ m m ´ 2m ` 16 “ E tr Y ´2 tr2 Y ´1 ` ´ E tr Y ´3 2 64n 8n ` ˘ ‰ m m2 ´ 2m ` 16 “ m2 “ 3 ´1 ‰ ´ E tr Y ´2 tr Y ´1 ` E tr Y 64n 32n ‰ m p´m ` pq “ 2 ´1 ‰ m pmp ´ 2p ´ 8q “ ` E tr Y ´2 ´ E tr Y 128 64 ‰ n2 p 2 mnp “ ´ E tr Y ´1 ` (4.35) 64 256 again where Y „ Wp pn, Ip {nq. Unfortunately, as we consider larger orders, the repeated differentiation of expt´ n4 tr Zu|Z|m{4 quickly becomes too cumbersome to perform by hand. But Chételat and Wells/Mid-scale Wishart asymptotics 35 at“ least ‰in theory, “ we can ‰ compute expressions like (4.34) and (4.35) for any E tr T 2k and E tr2 T k , and Lemma 2 summarize that fact. That is, using the Fourier inversion theorem we have reduced the problem of computing moments of the t-distribution Tn{2 pIp {8q to that of computing expected power sum polynomials of the inverse Wishart distribution W´1 p pn, Ip {nq, for large enough n. How can we compute expected power sum polynomials of an inverse Wishart? There are two approaches in the literature. Letac and Massam [2004] found an expression in terms of a different basis, the zonal polynomials, which behave particularly nicely with respect to the inverse Wishart distribution, and whose expectations have a simple closed form. From this, they provided an algorithm for computing expected power sum polynomials to arbitrary order. Matsumoto [2012] found expressions of coordinate-wise moments in terms of modified Wiengarten orthogonal functions, from which expectations of power sum polynomials can be computed. We follow the idea of Letac and Massam [2004] in our asymptotic analysis. For any integer partition κ, there exist coefficients cκ,λ (which depend solely on κ and λ) such that ÿ rκ pY ´1 q “ cλ Cλ pY ´1 q |λ|“|κ| for Cλ the so-called zonal polynomials. For an overview of the topic with a focus on random matrix theory, see Muirhead [1982, Chapter 7]. The coefficients cκ,λ are explicitly computable. If we follow the normalization of zonal polynomials of Muirhead [1982], for example, we find that “ ‰ “ ‰  „  „ „ r∅ “ C∅ , Cp2q rp2q 1 ´ 21 and . “ (4.36) “ ‰ “ ‰ Cp1,1q rp1,1q 1 1 rp1q “ Cp1q , As mentioned, expectations of zonal polynomials with respect to a Wishart or inverse Wishart distribution take a particularly simple form. From Muirhead [1982, Theorem 7.2.13 and Equation (18) on p.237], the expected zonal polynomials for Y ´1 „ W´1 p pn, Ip {nq are “ ‰ n|λ| E Cλ pY ´1 q “ Cλ pIp q ś qpλq 2|λ| i“1 m´i`1 2 śqpλq i ´1 ź λź 2|λ| |λ|! iăj p2λi ´ 2λj ´ i ` jq |λ| qpλq p ` p1 ´ i ` 2lq n “ (4.37) śqpλq m ´ p1 ´ i ` 2lq i“1 l“0 i“1 p2λi ` qpλq ´ iq! “ ‰ for λ ‰ ∅, and E C∅ pY ´1 q “ 1. For example, the first few expected zonal polynomials are “ ‰ “ ‰ np , E C∅ pY ´1 q “ 1, E Cp1q pY ´1 q “ m “ ‰ 2n2 p pp ´ 1q “ ‰ n2 p pp ` 2q E Cp1,1q pY ´1 q “ , E Cp2q pY ´1 q “ . 3m pm ` 1q 3m pm ´ 2q ‰ “ ‰ “ ‰ “ From this, we can exactly compute E rκ pY ´1 q and thus E tr T 2k and E tr2 T k , Chételat and Wells/Mid-scale Wishart asymptotics 36 as a function of p and n (or m). For example, by Equation (4.36) we find that “ ‰ “ ‰ np E r∅ pY ´1 q “ 1, E rp1q pY ´1 q “ , m “ ‰ “ ‰ n2 p pm ` pq n2 p pmp ´ p ` 2q , E rp2q pY ´1 q “ . E rp1,1q pY ´1 q “ m pm ´ 2q pm ` 1q m pm ´ 2q pm ` 1q and thus, by Equation (4.34), whenever n ě p ` 38 ‰ “ np pmp ` m ` 2q E tr T 2 “ . 16 pm ´ 2q pm ` 1q (4.38) In a similar way, we can compute the expected zonal polynomials and hence, the expected power sum polynomials of Y ´1 for |κ| “ 3, 4. So from Equation (4.35), we obtain for n ě p ` 38 that ´ “ ‰ n2 p m3 p3 ` 2m3 p2 E tr2 T 2 “ 256 pm ´ 6q pm ´ 2q pm ´ 1q pm ` 1q pm ` 3q ` 5m3 p ` 4m3 ´ 3m2 p3 ` 6m2 p2 ` 9m2 p ` 24m2 ´ 12mp3 ¯ ´ 36mp2 ` 36mp ´ 36p . (4.39) Of course, this reasoning also works for other k’s. In particular, we essentially derived a (potentially inefficient) algorithm to compute the moments of a Tn{2 pIp {8q distribution to arbitrary order on our path to proving this theorem. At this point, it is worthwhile to realize that Equations (4.38) and (4.39) are already enough to prove the theorem for small moments. For example, when n Ñ 8 such that p{n Ñ c P r0, 1q, then m „ p1 ´ cqn and ˆ 2 ˙ “ ‰ m2 np np n 1 ` ` „ p2 , E tr T 2 “ pm ´ 2q pm ` 1q 16m 16m 8m2 16p1 ´ cq “ ‰ which proves that E tr T 2 “ Opp2 q. In fact,  ˆ ˙ „ m2 1 p 2 3 2 2 1 ´ 4T ¯2 ´1“ ` ` ` ` 2` 2 E tr ? pm´2q pm`1q p m m mp m p p m p 1 p Ñ 0. (4.40) „ ` p m Moreover, when n, p Ñ 8 such that p{n Ñ 0, then m „ n and „ˆ ˙  2 ‰ 32 “ ‰2 1 ´ 4T ¯2 256 “ E tr ? ´1 “ 4 E tr2 T 2 ´ 2 E tr T 2 ` 1 p p p p ˆ 5 m 5 4 2 22 “ ` 3` ` pm´6q pm´2q pm´1q pm`1q pm`3q p2 p m mp 27 32 p2 6p 33 60 115 52 ` ` ` ` 2` 2` 2 ` 2 2` 2 3 2 3 2 mp mp m m m m p m p m p 24 12p2 3p2 6p 3 118 93 96p ´ 3 ` 3` 3` 3 ` 3 2` 3 3´ ´ 4 m m m m p m p m p m4 m Chételat and Wells/Mid-scale Wishart asymptotics 48 84 144 144 36 36 ´ 4´ 4 ´ 4 2´ 5 ´ 5 ´ 5 2 m m p m p m m p m p 2 p2 5 ` 2 Ñ 0. „ 2` p m m 37 ˙ (4.41) Thus 1 ´ 4T ¯2 L2 tr ? ÝÑ 1 “ C1 p p and the theorem is proven for the second moment. In theory, we could proceed in the same way for any moment of interest, but naturally we could never conclude that the theorem holds for all moments that way. Nonetheless, the calculations give us some hints about how to argue in the general case. The idea is to express the moments of the symmetric t distribution as polynomials of p and p{m. There are two regimes where random matrix theory is well understood: the classical regime where p is held fixed as n Ñ 8, and the linear, high-dimensional regime where p grows linearly with n. From this, we can therefore conclude a few facts regarding the behavior of symmetric t moments in these regimes. But these moments are polynomials, and a polynomial is a very rigid object: results from the two extreme cases where p is fixed and p grows linearly will be enough to prove results for every regime in between, yielding the first part of the theorem. Proving the second part will then be the simple matter of applying the GOE approximation of Jiang and Li [2015] and Bubeck et al. [2016] to the specific shape found for the symmetric t moments while proving the first part, namely Equations (4.54) and (4.55). Proof of Theorem 2. Recall the expected zonal polynomial of an inverse Wishart W´1 p pn, Ip {nq is given by Equation (4.37). Based on the previous calculations, it is tempting to define śqpλq qpλq i ´1 ź λź 2|λ| |λ|! iăj p2λi ´ 2λj ´ i ` jq m 1 cλ “ , Rλ pmq “ śqpλq m ´ p1 ´ i ` 2lq i“1 l“0 i“1 p2λi ` qpλq ´ iq! ˆ ˙ qpλq i ´1 ź λź p 1 ´ i ` 2l and Pλ pm, pq “ ` (4.42) m m i“1 l“0 so that “ ‰ E Cλ pY ´1 q “ c1λ n|λ| Rλ pmqPλ pm, pq. With these expressions the expected power sum polynomials can be written as ˜ ¸ p ÿ ź ź “ ‰ ´1 1 |κ| ´1 E rκ pY q “ cκ,λ cλ n Rµ pmq Rµ pmq Pλ pm, pq |λ|“|κ| “ |µ|“|κ| |µ|“|κ| µ­“λ ź ÿ ź n|κ| ¨ Rµ pmq ¨ m|κ| cκ,λ c1λ Rµ´1 pmqPλ pm, pq. |κ| m |µ|“|κ| |λ|“|κ| |µ|“|κ| µ­“λ Chételat and Wells/Mid-scale Wishart asymptotics 38 In other words, if we define ź 1 R|µ| “ Rµ pmq, |µ|“|κ| 1 P|λ| pm, pq “ m|κ| ÿ cκ,λ c1λ |λ|“|κ| ź Rµ´1 pmqPλ pm, pq (4.43) |µ|“|κ| µ­“λ then “ ‰ n|κ| 1 1 E rκ pY ´1 q “ |κ| R|κ| pmqP|κ| pm, pq. (4.44) m ` ˘ ś ś λi ´1 qpλq 1´ 1´i`2l But Rµ´1 pmq “ i“1 l“0 is a polynomial in 1{m, while Pλ pm, pq “ m ˘ śqpλq śλi ´1 ` p 1´i`2l is a polynomial in p{m and 1{m, both of degree at i“1 i“1 m ` m most |µ| “ |λ| “ |κ|. Thus ÿ ź 1 cκ,λ c1λ Rµ´1 pmqPλ pm, pq P|κ| pm, pq ” m|κ| |λ|“|κ| “ m|κ| |κ| ÿ |κ| ÿ i“0 j“0 |µ|“|κ| µ­“λ bij ´ p ¯i 1 m mj (4.45) for some coefficients bij that don’t depend on m, p (or n). Define the polynomials ř|κ| fj pαq “ i“0 bij αi , so that |κ| ´ ÿ p ¯ ´j 1 P|κ| pm, pq “ m|κ| f m . (4.46) m j“0 Let us show that for all 0 ď`j ă |κ|´qpκq, the polynomial fj must be identically ˘ zero over the interval α P 0, 1{ maxp|κ| ´ 2, 0q . Indeed, say this was not the case, and let 0 ď j0 ă |κ| ˘ smallest j with the property that ` ´ qpκq1 be the . As fj0 is a polynomial, by continuity fj0 pα0 q ‰ 0 for some α0 P 0, maxp|κ|´2,0q it must be non-zero in a neighborhood of α0 , so we may as well assume α‰0 is “ rational without loss of generality. Now look at what happens to E rκ pY ´1 q as α0 p grows to infinity at the very specific linear rate p “ t 1`α pn ´ 1qu. Since α0 is 0 rational, there must be a subsequence nl such that pl is exactly an integer (for example, if α0 “ a{b with a, b integers, we can take nl “ pa ` bql ` 1). Then for α0 pl “ 1`α pnl ´ 1q, we have exactly pl “ α0 ml . 0 ` α ˘´1 1 0 Since α0 ă maxp|κ|´2,0q , then |κ| ă 1 ` 1`α . Thus by Hölder’s inequality 0 and Lemma 5, ı “ ‰ 1 1 1 ” lim |κ|´qpκq´j ¨ qpκq E rκ pY ´1 q ď 0 ¨ lim E tr Y ´|κ| “ 0. 0 lÑ8 m lÑ8 pl p l l On the other hand, by Equations (4.44) and (4.46), the definition of j0 and the fact that R|κ| pmq Ñ 1 as m Ñ 8, “ ‰ 1 1 lim ¨ qpκq E rκ pY ´1 q lÑ8 m|κ|´qpκq´j0 pl l Chételat and Wells/Mid-scale Wishart asymptotics ˆ “ lim lÑ8 “ p1 ` nl ml ˙|κ| ˆ 1 R|κ| pml q ml pl ˙qpκq ÿ |κ| 39 fj pα0 qmjl 0 ´j j“j0 ´qpκq α0 q|κ| α0 fj0 pα0 q. As α0 ą 0, fj0 pα0 q must therefore equal zero, a contradiction. Hence, as claimed, the ˘ fj pαq for 0 ď j ă |κ| ´ qpκq all vanish over the interval ` polynomials 1 0, maxp|κ|´2,0q . But a polynomial can have an infinite number of zeros only if all its coefficients are zero, so we conclude that bij “ 0 for 0 ď j ă |κ| ´ qpκq. Thus, from Equations (4.44) and (4.45) we have “ ‰ ´ n ¯|κ| qpκq 1 R|κ| pmqPκ2 pm, pq m E rκ pY ´1 q “ m where |κ| |κ| ´ p ¯i ÿ ÿ 1 Pκ2 pm, pq “ bij . j´|κ|`qpκq m m i“0 j“|κ|´qpκq Going back to equations (4.13) and (4.14) and plugging in the above yields that as long as n ě p ` 16k ` 6, “ ‰ m2k`1 1 R2k pmqQp1q (4.47) E tr T 2k “ κ pm, pq, nk “ ‰ m2k`2 1 E tr2 T k “ R2k`1 pmqQp2q (4.48) κ pm, pq nk where 1 p1q ÿ ´ 1 ¯|κ| R|κ| pmq bκ pn, m, pq 2 p k ` Qp1q pm, pq “ p´1q 1` P pm, pq, κ 1 pmq m m R2k m2k`1´qpκq κ |κ|ď2k k Qp2q κ pm, pq “ p´1q 1 p2q ÿ ´ 1 ¯|κ| R|κ| pmq bκ pn, m, pq 2 p P pm, pq. 1` ` 1 m m R2k`1 pmq m2k`2´qpκq κ |κ|ď2k`1 Now, for any a ď b, we can associate a partition µ of norm |µ| “ a with the partition µ˚ “ pµ1 ` b ´ a, µ` 2, . . . , µqpµq q of norm |µ˚ | “ b, which satisfies ˚ ˚ ˆ qpµ i ´1 ź q µź i“1 j“0 1´i`2j 1´ m ˆ qpµq i ´1 ź µź ˙ “ i“1 j“0 1´i`2j 1´ m µ1 `b ˙ ´a´1 ź ˆ j“µ1 ˙ 2j 1´ . m By definition for the Rµ pmq’s at Equation (4.42), this means that every factor that appears in Rµ´1 pmq appears in Rµ´1 ˚ pmq, so by definition of the R|µ| pmq’s p1q 1 at Equation (4.43), Ra pmqR´1 pmq is a polynomial in m . Moreover, as bκ and p2q bκ are polynomials of degrees d1 pκq ” 2k ` 1 ´ qpκq and d2 pκq ” 2k ` 2 ´ qpκq Chételat and Wells/Mid-scale Wishart asymptotics p1q 40 p2q respectively, there exists coefficients cijl and cijl such that p1q bκ pn, m, pq m2k`1´qpκq “ 1 md1 pκq d1 pκq d1 pκq dÿ ´i ´i´j 1 pκq ÿ ÿ i“0 j“0 d1 pκq d1 pκq dÿ ´i ´i´j 1 pκq ÿ ÿ “ i“0 j“0 p1q cijl mi nj pl l“0 1 p1q cijl l“0 md1 pκq´i´j´l ´ p 1 ¯j ´ p ¯l 1` ` m m m and p2q bκ pn, m, pq m2k`2´qpκq “ 1 md2 pκq d2 pκq d2 pκq dÿ ´i ´i´j 2 pκq ÿ ÿ i“0 j“0 d2 pκq d2 pκq dÿ ´i ´i´j 2 pκq ÿ ÿ “ i“0 j“0 l“0 p2q cijl p2q cijl mi nj pl l“0 ´ 1 ¯j ´ p ¯l p ` . 1` m m m md2 pκq´i´j´l 1 As d1 p´i ´ j ´ l ě 0q, j, l ě 0 in the first case and d2 pκq ´ i ´ j ´ l, j, l ě 0 in the second case, we conclude that these two expressions are polynomials p 1 in m and m . Therefore, looking back at (4.47) and (4.48), we conclude that p2q p1q p 1 the Qκ pm, pq’s and Qκ pm, pq’s are polynomials in m and m . Therefore, if p1q p2q n ě p ` 16k ` 6 there must be coefficients aij , aij and large enough integers D1 , D2 such that D1 ÿ i j ÿ “ ‰ m2k`1 p1q p E tr T 2k “ R pmq a 2k ij nk mi i“0 j“0 D1 ´ m ¯k ÿ p1q “ R2k pmq gi ppqmk`1´i (4.49) n i“0 D2 ÿ i j ÿ “ ‰ m2k`2 p2q p E tr2 T k “ R pmq a 2k`1 ij nk mi i“0 j“0 D2 ´ m ¯k ÿ p2q “ R2k`1 pmq gi ppqmk`2´i (4.50) n i“0 p1q for polynomials gi ppq “ i ř j“0 p1q p2q p1q i ř p2q aij pj . j“0 p2q gi must vanish aij pj and gi ppq “ We will now proceed to show that gi and on N for 0 ď i0 ă k ` 1 and 0 ď i0 ă k ` 2 respectively. Our argument relies on the analysis of the asymptotic behavior of the moments of T in the classical regime where p is held fixed while n grows to‰infinity.“ “ ‰ Observe first that E tr T 2k and E tr2 T k must have a finite limit as n Ñ 8 with p held fixed. Indeed, since 16T 2 {n is positive definite, |Ip ` 16T 2 {n| is Chételat and Wells/Mid-scale Wishart asymptotics 41 greater than one and so we have the bound ż ‰ “ E tr T 2k “ Cn,p tr T 2k ˇ ˇ n`p`1 2 ˇ´ 4 ˇ ˇIp ` 16T ˇ dT ˇ n ˇ Sp pRq ż ď Cn,p ˇ ˇ´ n ˇ 16T 2 ˇˇ 4 tr T 2k ˇˇIp ` dT. n ˇ Sp pRq pp3p`1q 4 ppp`1q 4 When p is held fixed, lim Cn,p “ 2 {π by Lemma 1. Moreover, nÑ8 n n ˇ ˇ ˆ ˙ p ´ ´ 2ˇ 4 4 ź ˇ λi p4T 2 q ˇIp ` 16T ˇ “ 1` ˇ ˇ n n{4 i“1 for λ1 p4T 2 q ě ¨ ¨ ¨ ě λp p4T 2 q ě 0 the eigenvalues of 4T 2 , and p1 ` x{nq´n is monotone decreasing towards exppxq. Therefore, for a fixed dimension p we can apply the monotone convergence theorem to obtain that ż pp3p`1q “ ‰ 2 4 lim E tr T 2k ď ppp`1q nÑ8 π 4 “ ‰ 2 tr T 2k e´4 tr T dT “ E tr Z 2k ă 8 (4.51) Sp pRq for Z „ GOEppq{4. Repeating the argument with tr2 T k yields similarly ż pp3p`1q 4 “ ‰ 2 lim E tr2 T k ď ppp`1q π 4 nÑ8 “ ‰ 2 tr2 T k e´4 tr T dT “ E tr2 Z k ă 8. (4.52) Sp pRq “ ‰ “ ‰ Thus indeed E tr T 2k and E tr2 T k have finite limits when p is held fixed. p1q p2q We can use this to show that gi and gi must vanish on N for 0 ď i0 ă k `1 and 0 ď i0 ă k ` 2 as follows. Say the first statement wasn’t true, and let p1q 0 ď i0 ă k ` 1 be the smallest i such that for some p0 P N, gi0 pp0 q ‰ 0. Then “ ‰ by Equation (4.49) and the definition of i0 , the limit of E tr T 2k as n Ñ 8 with p fixed at p0 satisfies “ ‰ D1 ÿ E tr T 2k p1q p1q k “ 1 ¨ 1 ¨ lim gi pp0 qmi0 ´i “ gi0 pp0 q. lim nÑ8 nÑ8 mk`1´i0 i“i 0 But m “ n´p´1 tends to infinity as n tends to infinity, and since k `1´i0 ą 0, “ ‰ p1q Equation (4.51) means that E tr T 2k {mk`1´i0 must tend to zero. Thus gi0 pp0 q has to equal zero, which contradicts our assumption. Thus for every 0 ď i ă k`1, p1q the polynomial gi must vanish on N. p2q Similarly, for 0 ď i ă k ` 2, the polynomial gi must vanish on N, because if it wasn’t the case, we could take 0 ď i0 ă k ` 2 as the smallest i with the p2q property that for some p0 P N, gi0 pp0 q ‰ 0, and then by Equation (4.50) with Chételat and Wells/Mid-scale Wishart asymptotics 42 p fixed at p0 as n Ñ 8 we would get “ ‰ D2 ÿ E tr2 T k p2q p2q k “ 1 ¨ 1 ¨ lim gi pp0 qmi0 ´i “ gi0 pp0 q. lim nÑ8 nÑ8 mk`2´i0 i“i 0 But then by Equation (4.52), as m tends to infinity and k ` 2 ´ i0 ě 1 the “ ‰ p2q ratio E tr2 T k {mk`2´i0 must tend to zero. Thus we must have gi0 pp0 q “ 0, a p2q contradiction. Hence indeed for 0 ď i ă k ` 2, the polynomial gi must vanish on N. But of course, a polynomial can only have an infinite number of zeroes if its coefficients are all zero, so we must have p1q for 0 ď i ă k ` 1, p2q for 0 ď i ă k ` 2. aij “ 0 aij “ 0 (4.53) Now say that p varies with n in such a way that limnÑ8 p{n “ α ă 1. Then for large enough n, n ě p ` 16k ` 6 so by Equations (4.49) and (4.53), D1 ÿ i ´ m ¯k j´pk`1q ÿ “ ‰ 1 p1q p R2k pmq aij i´pk`1q lim k`1 E tr T 2k “ lim nÑ8 n nÑ8 p m i“k`1 j“0 « p1q D1 ÿ k ÿ aij “ p1 ´ αqk ¨ 1 ¨ lim i´pk`1q ppk`1q´j nÑ8 m i“k`1 j“0 ff D1 i ´ p ¯j´pk`1q 1 ÿ ÿ p1q ` aij m mi´j i“k`1 j“k`1 ff « ˙ ˆ p1q D1 i´pk`1q k ÿ ÿ ak`1 α p1q (4.54) “ p1 ´ αqk ` a ` ˘pk`1q´j ii 1´α j“0 lim p i“k`1 pÑ8 and by Equations (4.50) and (4.53), D2 ÿ i ´ m ¯k j´pk`2q ÿ “ ‰ 1 p2q p lim k`2 E tr2 T k “ lim R2k`1 pmq aij i´pk`2q nÑ8 p nÑ8 n m i“k`2 j“0 « p2q D k`1 ÿ2 ÿ aij “ p1 ´ αqk ¨ 1 ¨ lim i´pk`2q nÑ8 m ppk`2q´j i“k`2 j“0 ff D1 i ´ p ¯j´pk`2q 1 ÿ ÿ p2q ` aij m mi´j i“k`2 j“k`2 « ff ˆ ˙ p2q D2 i´pk`2q k`1 ÿ ÿ ak`2 α p2q “ p1 ´ αqk aii . (4.55) ` ˘pk`2q´j ` 1´α j“0 lim p i“k`2 pÑ8 Although we might not know what aij coefficients are, this shows at least that the limits are finite. In‰ particular, from “Equations (4.54) and (4.55) we can “ ‰ conclude that E tr T 2k “ Oppk`1 q and E tr2 T k “ Oppk`2 q, which shows the first claim of the theorem. For the second claim, let n, p Ñ 8 with p{n Ñ α “ 0. Then Equations (4.54) Chételat and Wells/Mid-scale Wishart asymptotics 43 and (4.55) specialize to lim nÑ8 1 pk`1 1 lim nÑ8 pk`2 “ ‰ p1q E tr T 2k “ apk`1qpk`1q , “ ‰ p2q E tr2 T k “ apk`2qpk`2q . (4.56) What is interesting about this result is that these limits must be the same regardless of the way p grows! As long as p Ñ 8 with p{n Ñ 0, the limits are ? p1q p2q apk`1qpk`1q and apk`2qpk`2q , regardless of whether p „ log n or p „ n or some other growth. Now, Bubeck et al. [2016, Theorem 7] and Jiang and Li [2015, Theorem 1] have shown that when p Ñ 8?with p3 {n Ñ 0, the total variation distance between a normalized Wishart npWp pn, Ip {nq ´ Ip q matrix and a Gaussian Orthogonal Ensemble GOEppq matrix tends to zero as n Ñ 8. Therefore, the Hellinger distance satisfies also H 2 pψNW , ψGOE q “ H 2 pfNW , fGOE q Ñ 0 as n Ñ 8. But convergence in Hellinger distance has strong implications for real-valued statistics. Indeed, for T1 „ Tn{2 pIp {8q “ |ψNW |, T2 „ GOEppq{4 “ |ψGOE | and any function g : Sp pRq Ñ R such that gpT1 q, gpT2 q are square-integrable, ˇ ˇ ˇˇ ” ı ˇˇ ˇ ˇ gpT q |ψNW |pT q ´ |ψGOE |pT q dT ˇˇ ˇ ErgpT1 qs ´ ErgpT2 qs ˇ “ ˇˇ ż Sp pRq ż ˇ ˇ ď ˇˇ ” ı ˇˇ 1{2 1{2 1{2 gpT qψNW pT q ψNW pT q ´ ψGOE pT q dT ˇˇ Sp pRq ˇ ˇ ` ˇˇ ż ” ı ˇˇ 1{2 1{2 1{2 gpT qψGOE pT q ψNW pT q ´ ψGOE pT q dT ˇˇ Sp pRq ż ż „  1 gpT q2 |ψNW |pT qdT 2 ` ď Sp pRq ż ¨ 1 gpT q2 |ψGOE |pT qdT 2 Sp pRq ˇ ˇ2 1 ˇ 1{2 ˇ 1{2 ˇψNW pT q ´ ψGOE pT qˇ dT 2 S pRq p 1 1 “ ‰ “ ‰ “ E gpT1 q2 2 ` E gpT2 q2 2 HpψNW , ψGOE q „ (4.57) by the Cauchy-Schwarz inequality. Let’s consider applying this result to gpT q “ tr T 2k {pk`1 and gpT q “ tr2 T k {pk`2 . What do we know about these statistics? In the case where T2 „ GOEppq{4, results of Anderson et al. [2010, Lemma 2.1.6 and the equation above Equation (2.1.21)] provide that „  tr T22k Ck lim E k`1 “ 2k nÑ8 p 4 Chételat and Wells/Mid-scale Wishart asymptotics 44 „ 2 k ˙2 ´ „ tr T k  ¯2 ˆ C tr T k{2 2 lim E k`22 “ 1 rk evens “ lim E k{2`1 nÑ8 nÑ8 p 4k p „´  „  ˆ ˙2 ´ tr T22k ¯2 tr T22k ¯2 Ck lim E “ lim E k`1 ă8 “ nÑ8 nÑ8 pk`1 p 42k „ „´ 2 k ¯   tr T24k tr T2 2 C2k ď lim E lim E “ 4k ă 8 nÑ8 nÑ8 pk`2 p2k`1 4 because these expressions only depend on p, and since p Ñ 8 as n Ñ 8, taking a limit as n Ñ 8 is the same as taking a limit as p Ñ 8. Moreover, in the T1 „ Tn{2 pIp {8q case, using Jensen’s inequality and Equations (4.54)–(4.55) we can at least see that „´  „  tr T12k ¯2 tr T14k lim E ď lim E 2k`1 ă 8 nÑ8 nÑ8 pk`1 p „ 2 2k  „´ 2 k ¯  2 tr T1 tr T1 ď lim E ă 8. lim E nÑ8 nÑ8 pk`2 pk`2 Therefore, using Equation (4.57) with gpT q “ tr T 2k {pk`1 and gpT q “ tr2 T k {pk`2 we find that when n, p Ñ 8 with p3 {n Ñ 0, ˇ ˇ „  ˆ „´ 1 ˙ 2k ˇ Ck ˇˇ tr T12k ¯2 2 C2k ˇ lim E tr T1 ´ 2k ˇ ď lim E ` 2k ¨ 0 “ 0, ˇ nÑ8 nÑ8 pk`1 4 pk`1 4 ˇ ˇ „ 2 k ´ ¯ 2ˇ ˇ Ck{2 ˇ lim E tr T1 ´ 1rk evens ˇˇ ˇ nÑ8 k`2 k p 4 ˆ „´ 2 k ¯  12 1{2 ˙ C2k tr T1 2 ď lim E ` 2k ¨ 0 “ 0. nÑ8 pk`2 4 Since p3 {n Ñ 0 implies p{n Ñ 0, we conclude from Equation (4.56) that 2 Ck{2 Ck p1q p2q 2 apk`1qpk`1q “ 2k , apk`2qpk`2q “ 2k 1rk evens . 4 4 But then, by that same equation, we conclude that when n, p Ñ 8, not only when p3 {n Ñ 0 but for all p such that p{n Ñ 0, we have “ ‰ Ck 1 lim E tr T 2k “ 2k , nÑ8 pk`1 4 (4.58) 2 “ 2 k ‰ Ck{2 1 2 lim k`2 E tr T “ 2k 1rk evens nÑ8 p 4 for “ T „‰Tn{2 pIp {8q. To finish the proof, use Equation (4.58) with the fact that E tr T k “ 0 for odd k to find that „  „ 2 k „ 2 tr T k tr T tr T k lim Var k{2`1 “ lim E k`2 ´ E k{2`1 nÑ8 nÑ8 p p p 2 ´ ¯2 Ck{2 Ck{2 2 “ 2k 1rk evens ´ 1 rk evens “ 0. 4 4k L2 Thus tr T k {pk`1 ÝÑ Ck {42k , as desired. This proves the second claim and concludes the proof. Chételat and Wells/Mid-scale Wishart asymptotics 45 A pleasant consequence of this result is that when n, p Ñ 8 with p{n Ñ 0, we can conclude a version of the semicircle law holds for the Tn{2 p2Ip q distribution. This is interesting because the Tn{2 p2Ip q distribution has dependent entries with heavy tails, whose distribution varies with n, p. ? ? ? ? Let 4T { p „ 4Tn{2 pIp {8q{ p, with eigenvalues λ1 p4T { pq ě ¨ ¨ ¨ ě λp p4T { pq. Then define its empirical spectral measure to be p ? 1ÿ L4T {?p pAq “ 1rλi p4T { pq P As . p i“1 Since L4T {?p depends on the random matrix T , it is a random measure on R. Corollary 1 (Semicircle law for the t distribution). The empirical spectral ? measure L4T {?p of a 4Tn{2 pIp {8q{ p random matrix converges weakly, in square mean, to the semicircle distribution ż ? 4 ´ x2 1r|x| ď 2s dx. LpAq “ 2π A Proof. Let f be any continuous function R Ñ R that vanishes at infinity. By the Stone-Weierstrass theorem, there exists a sequence f1 , f2 , . . . of polynomials such that for any  ą 0, supxPR |f pxq ´ fl pxq| ă . To fix some notation, write fl pxq “ degf ÿl alk xk . k“1 Then since L4T {?p and L are both probability measures, «ˆ ż ˙2 ff 12 ż f pxqdL4T {?p pxq ´ f pxqdLpxq E R R «ˆ ż ˙2 ff 21 “ ‰ f pxq ´ fl pxq dL4T {?p pxq ď E R «ˆ ż ˙2 ff 21 ż fl pxqdL4T {?p pxq ´ fl pxqdLpxq `E R R «ˆ ż ˙2 ff 21 “ ‰ `E fl pxq ´ f pxq dLpxq R ď ` degf ÿl k“1 «ˆ ˙2 ff 21 1 ´ 4T ¯k |alk | E tr ? ´ Ck{2 1rk evens ` . p p By Theorem 2, the expectation tends to zero as n, p Ñ 8 with p{n Ñ 0. Thus «ˆ ż ˙2 ff 21 ż lim E f pxqdL4T {?p pxq ´ f pxqdLpxq ď 2. nÑ8 R R But this is true for every  ą 0, so the limit must ş be zero. Hence for every continuous f that vanishes at infinity, the integral f dL4T {?p converges in square Chételat and Wells/Mid-scale Wishart asymptotics 46 ş mean to f dL. By Chung [2001, Theoremş 4.4.1 and 4.4.2], this implies that for every bounded continuous f , the integral f dL4T {?p converges in square mean ş to f dL. Thus the empirical spectral distribution L4T {?p converges weakly, in square mean, to the semicircle distribution L, as desired. 5. Wishart asymptotics: the G-transform point-of-view We now turn our attention to the main objective of this paper, namely studying the behavior of Wishart matrices in the various middle-scale regimes. To do this, we exploit the close connection between the Wishart and the symmetric t distributions and make use of the results found Section 4. The main result of this section, Theorem 3, states that we can approximate for every middlescale regime the G-transform ψNW of a normalized Wishart by a degree-specific function ψK . This can be seen as an analogue of Theorem 1 in the G-transform domain. The reasoning behind the approximations is as follows. We could imagine writing ψNW from Proposition 3 in exponential form, and expanding the terms as a Taylor series would yield ˇ ˇ* " ˇ ? 4T ˇˇ n`p`1 ˇ log ˇIp ` i ? ˇ ψNW pT q “ Cn,p exp 2i n tr T ´ 2 n " ˆ ˙k * 8 ´ ? n ` p ` 1 ÿ p´iqk p ¯ k2 4T tr ? . “ Cn,p exp 2i n tr T ` 2 k n p k“1 Now imagine that the T ’s appearing in the expression follow a Tn{2 pIp {8q dis` 4T ˘k tribution. By Theorem 2, we know that tr ? “ Θppq when k is even, in p ` 4T ˘k 2 an L sense. When k is odd, the theorem merely proves that tr ? “ oppq, p k but for a GOEppq matrix Z, we know that tr ?Zp is asymptotically normal for odd k by Anderson et al. [2010, Theorem 2.1.31]. This would suggest that ` 4T ˘k tr ? “ Θp1q when k is odd. Thus we would have, in some sense, p "ÿ ˆ k{2`1 ˙ ÿ ˆ k{2 ˙ * 8 8 p p ψNW pT q “ Cn,p exp Θ ` Θ k{2´1 k{2´1 n n k“2 k“2 even “ Cn,p exp "ÿ 8 k“0 odd ˆ Θ k`2 p nk 8 ÿ ˙ ` k“1 ˜c Θ pp2k´1q`2 n2k´1 ¸* . In other words, terms in the power series would be associated with some degree K, such that they would be non-negligible in any middle-scale regime of degree up to K, and negligible in higher degrees. In fact, a similar phenomenon occurs with Cn,p , by Lemma 1. This suggests we should try truncating these power series to derive degree-specific approximations. Definition 3 (G-transform approximations). For any K P N, define the K th Chételat and Wells/Mid-scale Wishart asymptotics 47 degree approximation ψK : Sp pRq Ñ C as $ 2K`3` , 2K`2´ ’ / odds 1 rK odds & 1rK . ÿ ÿ ¯ ¯ ´ ´ k k k k pKq p `1 4i tr T 4i tr T ψK pT q “ Cn,p exp n ? ? ` ’ / %2 n k 2 n k k“2 with pKq Cn,p “ 2 pp3p`1q 4 π ppp`1q 4 k“1 " K`1 1 ÿ 1rk evens pk`2 exp ´ 2 k“1 kpk`1qpk`2q nk * K`1 1 ÿ 1`21rk evens pk`1 ´ . 4 k“1 kpk`1q nk Just like the G-transform of a normalized Wishart matrix, these functions implicitly depend on n. The first three are " * pp3p`1q 32 p`1 p`1 2 4 tr T 2 , ψ0 pT q “ ppp`1q exp ´4 tr T 2 ´i ? tr T 3 `2i ? tr T ´4 3 n n n π 4 " 2 pp3p`1q 2 4 32 32 512 p ψ1 pT q “ ppp`1q exp ´ ´4 tr T 2 ´i ? tr T 3 ` tr T 4 `i 3{2 tr T 5 8n 3 n n 5n π 4 * p ` 1 p ` 1 32pp ` 1q 1024 6 2 3 tr T `2i ? tr T ´4 tr T ´i tr T ´ 3n2 n n 3n3{2 and " pp3p`1q 2 4 p4 32 32 p2 ψ2 pT q “ ppp`1q exp ´ ´4 tr T 2 ´i ? tr T 3 ` tr T 4 ´ 2 48n 8n 3 n n 4 π 512 1024 8192 p`1 `i 3{2 tr T 5 ´ tr T 6 ´i 5{2 tr T 7 `2i ? tr T 2 3n n 5n 7n p`1 32pp ` 1q p ` 1 ´4 tr T 2 ´i tr T 3 `32 2 tr T 4 3{2 n n 3n * 512pp ` 1q 1024pp ` 1q 5 6 `i tr T ´ tr T . 3n3 5n5{2 These functions have the pleasant property that their modulus is bounded, up to a constant, by the G-conjugate density |ψK |. Indeed, on one hand we can rewrite Definition 3 into K`1` $ 1rK odds ’ ? 2k ÿ & n tr p4T { nq pKq ψK pT q “ exp log Cn,p ` p´1qk ’ 2 2k % k“1 K`1´ 1rK odds p`1 ` 2 ÿ p´1q k“1 ? k tr p4T { 2k 2k nq K`1 n ´i 2 ÿ k“1 p´1qk ? 2k`1 tr p4T { nq 2k ` 1 Chételat and Wells/Mid-scale Wishart asymptotics 48 , ? 2k`1 / . tr p4T { nq ´i p´1qk . (5.1) / 2 2k ` 1 k“1 ? ` ˘ On the other hand, for any x P R, we can write 1 ` ix “ 1 ` x2 exp i atanpxq with the arctangent function taking values in p´π{2, π{2q. Thus by Proposition 3 we can rewrite ψNW as ˇ ˇ " ˇ 16T 2 ˇˇ n`p`1 ˇ log ˇIp ` ψNW pT q “ exp log Cn,p ´ 4 n ˇ ˆ ˙ * ? n`p`1 4T ´i tr atan ? ` 2i n tr T , (5.2) 2 n with the understanding that the matrix-variate arctangent function operates on eigenvalues by functional calculus. Now, for any x P R and odd integer L, there is an elementary inequality L ÿ 1 x2l ď ´ logp1 ` x2 q. p´1ql 2l 2 l“1 K ÿ p`1 Notice that K ` 1 ˘ 1rK odds is always an odd integer. Thus, from the above inequality and Equations (5.1) and (5.2), we can derive the bound ˇ ! ´n p ` 1¯ ˇ ˇ 16T 2 ˇˇ) ˇ pKq ˇψK ˇpT q ď Cn,p exp ´ ` log ˇIp ` ˇ 4 4 n pKq ˇ Cn,p ˇˇ “ ψNW ˇpT q. (5.3) Cn,p In particular, since ψNW is integrable whenever n ě p ´ 2 by Proposition 3, Equation (5.3) implies that every ψK must also be integrable whenever n ě p´2. In particular, for large enough n it makes sense to talk about the asymptotic total variation or Hellinger distance between ψNW and ψK . We now state the main result, which is that each function ψK approximates the G-transform of a normalized Wishart for all middle-scale regimes of degree K or lower, but no other. log p Theorem 3.Let lim log n ă 1 as n Ñ 8. For any K P N, the total varianÑ8 tion distance between the G-transform of the normalized Wishart distribution ? nrWp pn, Ip {nq ´ Ip s and the K th approximating function ψK satisfies ż ` ˘ dTV ψNW , ψK “ ˇ ˇ ˇψNW pT q ´ ψK pT qˇ dT Ñ 0 as n Ñ 8 Sp pRq if and only if pK`3 {nK`1 Ñ 0. Proof. If statement. For the first part of the theorem, remark that by Equation (3.7) it is equivalent to show that the Hellinger distance tends to zero, i.e. that ż ` ˘ H2 ψNW , ψK “ ˇ ˇ2 ˇ 1{2 ˇ 1{2 ˇψNW pT q ´ ψK pT qˇ dT Ñ 0 Sp pRq as n Ñ 8 Chételat and Wells/Mid-scale Wishart asymptotics 49 when pK`3 {nK`1 Ñ 0. To control this quantity, we use the Kullback-Leibler inequality for G-transforms. Notice that for any x P R and L P N, ˇ ˇ L´1 ˇ 1 2l ˇ ÿ x2L ˇ 2 lx ˇ p´1q , (5.4) ˇ ď ˇ´ logp1 ` x q ´ ˇ 2 2l ˇ 2L l“1 ˇ ˇ L´1 ˇ 2l`1 ˇ ÿ x2L x ˇ ˇ p´1ql . (5.5) ˇ ď ˇatanpxq ´ ˇ 2l ` 1 ˇ 2L l“1 Let Log stand for the principal branch of the complex logarithm, and let us study Log ψNW {ψK . Its real part can be bounded by ˇ ˇ ˇ ˇˇ ˇ ˇ 2ˇ ˇ ˇ ˇ pKq ˇ< Log ψNW pT q ˇ “ ˇˇ log Cn,p ´ n ` p ` 1 log ˇIp ` 16T ˇ ´ log Cn,p ˇ ˇ ψK pT q ˇ ˇ 4 n ˇ ˇ K`1` K`1´ k“1 k“1 1rK odds 1rK odds ? 2k ? 2k ÿ ÿ n p`1 tr p4T { nq k tr p4T { nq ´ ´ p´1q p´1qk 2 2k 2 2k ˇ ˇ ˇ ˇ ˇ ˇ ˇ K`1` ˇ ˇ 1rK odds ˇ ˇ ˇ ˇ ˇ ˇ ? ÿ 2k ˇ ˇ 2ˇ ˇ ˇ nˇ 1 ˇ 16T tr p4T { nq ˇ pKq k ˇ´ ď ˇˇ log Cn,p ´log Cn,p ˇˇ ` ˇ´ log ˇˇIp ` p´1q ˇ ˇ 2ˇ 2 n ˇ 2k ˇ ˇ k“1 K`1´ ˇ ˇ odds ˇ ˇ ˇ 1rK ? 2k ˇˇ ÿ ˇ 2 ˇ 16T ˇˇ p`1ˇ 1 tr p4T { nq ˇ ´ ` p´1qk ˇ´ log ˇˇIp ` ˇ. ˇ ˇ 2 ˇ 2 n 2k ˇ ˇ k“1 By Equation (5.4), this can be bounded by ˇ ˇ ? ˇ ˇ n trp4T { nq2K`4`21rK odds pKq ˇ ˇ ď ˇ log Cn,p ´log Cn,p ˇ ` 2 2K ` 4 ` 21rK odds ? p ` 1 trp4T { nq2K`4´21rK odds ` 2 2K ` 4 ´ 21rK odds ˇ ˇ 2K`3`21rK odds ˇ ˇ 4 tr T 2K`4`21rK odds pKq ˇ “ ˇˇ log Cn,p ´log Cn,p ˇ ` K ` 2 ` 1rK odds nK`1`1rK odds 42K`3´21rK odds pp ` 1q tr T 2K`4´21rK odds . (5.6) K ` 2 ´ 1rK odds nK`2´1rK odds We can bound the imaginary part of Log ψNW {ψK in a similar way. Define x Pp´π,πs : R Ñ p´π, πs to be the projection mapping Pp´π,πs x “ x ´ 2πr 2π ´ 12 s. A plot is given as Figure 4. It satisfies = Log z “ Pp´π,πs = log z for all branches of log z, as well as the inequality |Pp´π,πs x| ď |x|. Using this mapping, we can see that the imaginary ` Chételat and Wells/Mid-scale Wishart asymptotics 50 y π -5π -3π -π π 3π 5π x -π Fig 4. Plot of Pp´π,πq on p´5π, 5πs. part of Log ψNW {ψK can be bounded as ˇ » ˇ ˇˇ ˇ ˆ ˙ ˇ ˇ ? n`p`1 4T ˇ= Log ψNW pT q ˇ “ ˇˇ Pp´π,πs— tr atan ? ` 2 n tr T –´ ˇ ˇ ˇ ψK pT q 2 n ˇ fiˇ K ? 2k`1 ? 2k`1 ˇˇ ÿ p`1 tr p4T { nq tr p4T { nq ffiˇ ` ` p´1qk p´1qk flˇ ˇ 2 2k ` 1 2 2k ` 1 ˇ k“1 k“1 ˇ ˇ ˇ K`1 ˆ ˙ ? ? 2k`1 ˇˇ ÿ 4T 2 n tr T n ˇˇ ˇ k tr p4T { nq ´ ´ ď ˇ tr atan ? p´1q ˇ ˇ 2ˇ n n{2 2k ` 1 ˇ ˇ k“1 ˇ ˇ ˇ K ˙ ÿ ˆ ? 2k`1 ˇˇ ˇ p`1ˇ 4T tr p4T { nq ˇ ` ´ p´1qk ˇ tr atan ? ˇ. ˇ 2 ˇ n 2k ` 1 ˇ ˇ k“1 K`1 ÿ n By Equation (5.5), this can be bounded by ? ? p ` 1 trp4T { nq2K`2 n trp4T { nq2K`4 ` ď 2 2K ` 4 2 2K ` 2 2K`3 2K`4 2K`1 4 tr T 4 pp ` 1q tr T 2K`4 ď ` . (5.7) K`1 K `2 n K `1 nK`1 Recall that the G-conjugate of the normalized Wishart distribution is the t distribution with n{2 degrees of freedom and scale matrix Ip {8, denoted Tn{2 pIp {8q – see Equation (3.24) and Section 4 for details. Let us bound the expectations of these absolute real and imaginary parts under this distribution. By Equations (5.10), (5.6), (5.7) and Theorem 2, we find that for T „ |ψNW | “ Tn{2 pIp {8q, ‰ “ ˇ „ˇ 2K`3`21rK odds E tr T 2pK`2`1rK oddsq ˇ ˇ ψ pT q 4 NW ˇ ď E ˇˇ< Log ψK pT q ˇ K ` 2 ` 1rK odds nK`1`1rK odds “ ‰ ´ pK`3 ¯ 42K`3´21rK odds pp ` 1q E tr T 2pK`2´1rK oddsq ` ` o K ` 2 ´ 1rK odds nK`1 nK`2´1rK odds ˆ K`2`1rK odds`1 ˙ ˆ ˙ ´ K`2´1rK odds`1 K`3 ¯ p p¨p p ď O `O ` o K`1 K`1` 1 odds K`2´ 1 odds rK rK n n n Chételat and Wells/Mid-scale Wishart asymptotics “ O 51 ´ pK`3 ¯ (5.8) nK`1 and “ ‰ “ ‰ ˇ „ˇ 2K`3 E tr T 2pK`2q 2K`1 pp`1qE tr T 2pK`2q ˇ ˇ ψ pT q 4 4 NW ˇ ď E ˇˇ= Log ` ψK pT q ˇ K `2 nK`1 K `1 nK`1 ˆ K`2`1 ˙ ˆ ˙ ´ pK`3 ¯ p p ¨ pK`2`1 ď O (5.9) `O “ O K`1 K`1 K`2 n n n as n Ñ 8 with p{n Ñ 0. Moreover, from Lemma 1, we see that " ˆ K`3 ˙* p p pKq Cn,p “ Cn,p exp o as n Ñ 8 with Ñ 0. (5.10) K`1 n n Thus, from Equations (5.3) and (5.10), we see that when pK`3 {nK`1 Ñ 0, the asymptotic L1 norm of ψK is bounded by ż ż ! lim nÑ8 ˇ ˇ ˇψK pT qˇ dT ď lim exp ´o nÑ8 ´ pK`3 ¯) nK`1 Sp pRq ˇ ˇ ˇψNW pT qˇ dT “ 1. (5.11) Sp pRq In fact, at the end of this proof we will see that this bound is sharp and the limit must be exactly 1. Using Equations (5.8), (5.9) and (5.11) with Proposition (1) implies that when pK`3 {nK`1 Ñ 0, ´ ¯ 0 ď lim H2 ψNW , ψK nÑ8 ż ż „ ď lim nÑ8  |ψK |pT q dT ´ 1 ` 0 ` 2 lim |ψK |pT q dT 2 ¨ 0 2 Sp pRq Sp pRq nÑ8 1 2 1 2 1 1 ď r1 ´ 1s ` 0 ` 2 ¨ 1 ¨ 0 “ 0. ` ˘ Thus` H2 ψNW˘, ψK Ñ 0, hence by Equation (3.7) we must have the limit dTV ψNW , ψK Ñ 0, as desired. Only if statement. For the second part of the theorem, assume that˘ the total ` variation distance satisfies dTV pψNW , ψK q Ñ 0, hence H ψNW , ψK Ñ 0 by Equation (3.7), as n Ñ 8. We will show by contradiction this implies that pK`3 {nK`1 Ñ 0. log p Assume this wasn’t the case. Since limnÑ8 log n ă 1, there must be an L P N such that pL`3 {nL`1 Ñ 0, and since pK`3 {nK`1 Û 0, we must have K ă L. By Equation (5.8), we must have for T „ |ψNW | “ Tn{2 pIp {8q that ˇ „ˇ ´ pL`3 ¯ ˇ ψNW pT q ˇˇ ˇ lim E ˇ< Log ď lim O L`1 “ 0, ˇ nÑ8 nÑ8 ψL pT q n so 1 ψNW pT q L1 < Log ÝÑ 0. (5.12) 2 ψL pT q Chételat and Wells/Mid-scale Wishart asymptotics 52 Now write, by Equation (5.1) and Definition 3, L`1` RpT q ” 1 ψL pT q 1 < Log “ log 2 ψK pT q 2 1rLÿ odds ˆ ˙2k n p´1qk 4T ` tr ? 4 2k n pLq Cn,p pKq Cn,p K`2` k“1rK odds L`1´ 1rLÿ odds ˆ ˙2k p`1 p´1qk 4T tr ? ` 4 2k n K`2´ k“1rK odds “ ´ L`1 ÿ 1`21rk evens pk`1 1 ÿ 1rk evens pk`2 1 L`1 ´ 4 k“K`2 kpk`1qpk`2q nk 8 k“K`2 kpk`1q nk L`1` ` 1rLÿ odds n 4 L`1´ p´1qk ´ p ¯k tr 2k n ˆ 4T ? p ˙2k ` p`1 4 K`2` k“1rK odds 1rLÿ odds ˆ ˙2k 4T p´1qk ´ p ¯k tr ? . 2k n p K`2´ k“1rK odds But as pL`3 {nL`1 , we must have p{n Ñ 0, so by Theorem 2, we have 1 p L2 4T 2k trp ? Ñ pq Ck . Moreover, as we assumed that pK`3 {nK`1 Û 0, we must have p Ñ 8. Thus L`1 1 ÿ 1rk evens pk´K´1 nK`1 RpT q “ ´ pK`3 4 k“K`2 kpk`1qpk`2q nk´K´1 L`1` 1 1`21rk evens 1 pk´K´1 1 ` 8 k“K`2 kpk`1q p nk´K´1 4 L`1 ÿ ´ 1rLÿ odds p´1qk pk´K´2 1 tr 2k nk´K´2 p ˆ 4T ? p ˙2k ˆ 4T ? p ˙2k K`2` k“1rK odds L`1´ 1 ` 4 L2 ÝÑ 0 ` 0 ` “ 1rK evens 8pK ` 2q ˆ 1 1` p odds ˙ 1rL ÿ p´1qk pk´K´1 1 tr 2k nk´K´1 p K`2´ k“1rK odds CK`2 ` 1rK odds 8pK ` 1q CK`1 CK`1`1rK evens ą 0. 8pK ` 1 ` 1rK evensq Then by the reverse triangle inequality, ż ` ˘ 0 “ lim H2 ψNW , ψK “ lim nÑ8 nÑ8 ˇ ˇ2 ˇ 1{2 ˇ 1{2 ˇψNW pT q ´ ψK pT qˇ dT Sp pRq (5.13) Chételat and Wells/Mid-scale Wishart asymptotics 53 ż ď lim nÑ8 ˇ2 ˇ ˇ ˇ ˇ|ψNW |1{2 pT q ´ |ψK |1{2 pT qˇ dT Sp pRq «ˇ ˇ2 ff " * ˇ ˇ 1 ψ pT q K “ lim E ˇˇ exp < Log ´ 1ˇˇ nÑ8 2 ψNW pT q for a T „ |ψNW | “ Tn{2 pIp {8q, that is " * 1 ψK pT q L2 exp < Log ÝÑ 1. 2 ψNW pT q Since Lp convergence implies convergence in probability, by the continuous mapping theorem we must have ψK pT q P 1 < Log ÝÑ 0 2 ψNW pT q as n Ñ 8, so by Equation (5.12) ψK pT q 1 ψNW pT q P 1 ´ < Log ÝÑ ´0 ´ 0 “ 0. RpT q “ ´ < Log 2 ψNW pT q 2 ψL pT q But then, from Equation (5.13) and Slutsky’s lemma [van der Vaart, 2000, Lemma 2.8 (iii)], ˆ K`1 ˙´1 pK`3 n P 8pK ` 1 ` 1rK evensq ¨ 0 “ 0. “ RpT q RpT q ÝÑ K`1 K`3 n p CK`1`1rK evens K`3 as n Ñ 8. As pK`3 {nK`1 is deterministic, this {nK`1 Ñ 0 ` implies ˘that p 2 as n Ñ 8, a contradiction. Thus whenever H ψNW , ψK Ñ 0 as n Ñ 8 with lim log p ă 1, we must have pK`3 {nK`1 Ñ 0, as desired. This concludes the nÑ8 log n proof. Although Theorem 3 states that the functions ψK approximate ψNW , there is no guarantee that they are G-transforms of a probability density. In other words, nothing guarantees that their inverse G-transforms f˜K “ G ´1 tψK u are real-valued, non-negative and integrate to unity. However, the reverse triangle inequality applied to the L2 -norm provides that ż ˇ ˇ ˇ ż |ψNW pT q| dT 2 ´ ˇ ` ˘ 1ˇ |ψK pT q| dT 2 ˇ ď H ψNW , ψK , Sp pRq Sp pRq 1 so Theorem 3 and the Plancherel theorem implies that ż lim nÑ8 ż |f˜K pT q| dT “ lim |ψK pXq| dX “ 1 Sp pRq Sp pRq nÑ8 (5.14) when pK`3 {nK`1 Ñ 0. That is, the theorem at least guarantees that |f˜K | is asymptotically a density in its associated regime. We discuss this further in Section 6. Chételat and Wells/Mid-scale Wishart asymptotics 54 We independently know, by the results of Jiang and Li [2015] and Bubeck et al. [2016], that a Gaussian orthogonal ensemble approximation holds in the classical regime. Although ψ0 is not the G-transform of a GOEppq, a simple Kullback-Leibler argument is sufficient to prove that it approximates ψGOE for 0th degree regimes. Proposition 4. The total variation distance between the 0th degree G-transform approximation ψ0 and the Gaussian orthogonal ensemble G-transform ψGOE satisfies dTV pψ0 , ψGOE q Ñ 0 as n Ñ 8 with p3 {n Ñ 0. Proof. We use a similar strategy to Theorem 3: namely, by Equation (3.7) is it equivalent to prove that Hpψ0 , ψGOE q Ñ 0 as n Ñ 8 with p3 {n Ñ 0, and to control that quantity we can use the Kullback-Leibler inequality for G-transforms. Let T „ |ψGOE | “ GOEppq{4. Since the Gaussian orthogonal ensemble has been extensively studied, we understand well its empirical moments. “ ‰For example, according to Anderson et al. [2010, Lemma 2.2.2], we have E tr T 2 “ Opp2 q, while (2.1.45) of the same book we have Er| tr T |s “ Opp1{2 q and “ from ‰ Equation 3 3{2 E | tr T | “ Opp q. Then from Definition 3 and Proposition 2, and using the projection map Pp´π,πq as in the proof of Theorem 3, we find that  ˆ 3˙ „ ‰ p`1 “ p ψGOE pT q 2 “ ´4 E tr T “ O E < Log ψ0 pT q n n and ˇ „ˇ „ „ˇ ˇ ˇ ˇ ˇ ψGOE pT q ˇˇ p`1 3 ˇPp´π,πq ´ 32 ˇ ? ? E ˇˇ= Log “ E tr T ` tr T ˇ ˇ ˇ ψ0 pT q 3 n n ˆc 3 ˙ ‰ p`1 p 32 “ . ď ? E | tr T 3 | `2i ? Er| tr T |s “ O 3 n n n ş Since Sp pRq |ψ0 |pT qdT Ñ 0 when n Ñ 8 with p3 {n Ñ 0 by Equation (5.14), if we apply Proposition 1 we find thatt 1 1 lim H2 pψ0 , ψGOE q ď 0 ` 0 ` 2 ¨ 1 2 ¨ 0 2 “ 0 nÑ8 for p3 {n Ñ 0. By Equation (3.7), this concludes the proof. As a consequence, HpfNW , fGOE q “ HpψNW , ψGOE q ď HpψNW , ψ0 q`Hpψ0 , ψGOE q Ñ 0 when n Ñ 8 with p3 {n Ñ 0 by Theorem 3 and Proposition 4. Hence dTV pfNW , fGOE q Ñ 0 by Equation (2.1) in the classical setting. This provides an alternative proof of the results of Jiang and Li [2015] and Bubeck et al. [2016]. 6. Wishart asymptotics: the density point-of-view In Section 5,?we studied the asymptotic behavior of the normalized Wishart distribution nrWp pn, Ip {nq ´ Ip s using its G-transform ψNW . In particular, we derived an approximation to ψNW for every middle-scale regime of a given degree. But although it is equivalent to study a probability distribution from a density or a G-transform point of view, it is still natural to wonder if we can find Chételat and Wells/Mid-scale Wishart asymptotics 55 approximations to the density of a normalized Wishart for every middle-scale regime of a given degree. Recall from Theorem 3 that dTV pψNW , ψK q Ñ 0 when pK`3 {nK`1 Ñ 0. Define f˜K “ G ´1 tψK u. In general, there is no guarantee that these should be real-valued. On the other hand, we know from Equation (5.3) that whenever n ě p ´ 2, ψK must be integrable, and since the G-transform maps integrable functions to integrable functions, f˜K must also be integrable. In fact, according to Equation (5.14), we know |f˜K | must be asymptotically a density when pK`1 {nK`1 Ñ 0. This suggests we define the following densities. Definition 4 (Density approximations). For any K P N and n ě p ´ 2, define the K th degree density approximation as |f˜K |pXq , fK pXq “ ş |f˜K pY q|dY Sp pRq where f˜K “ G ´1 tψK u and ψK is as in Definition 3. The distribution on the real symmetric matrices with density fK will be denoted FK . The main interest is that we can asymptotically approximate the density fNW of a normalized Wishart by the bona fide densities fK . This was the content of Theorem 1 from Section 1, which we now prove as a simple corollary of its G-transform analogue Theorem 3 from Section 5. Proof of Theorem 1. As in the rest ? of this paper, we write the density of the normalized Wishart distribution nrWp pn, Ip {nq ´ Ip s by fNW , and by Definition 4 the density of FK is fK . Notice that by Equation (2.1), to prove dTV pfNW , fK q Ñ 0 it is equivalent to prove that HpfNW , fK q Ñ 0. From the triangle inequality, the reverse triangle inequality, Theorem 3 and Equation (5.14), ´ ¯ ´ ¯ ´ ¯ lim H fNW , fK ď lim H fNW , |f˜K | ` lim H |f˜K |, fK nÑ8 nÑ8 nÑ8 ż ˇ ˇ2 1 ˇ 1{2 ˇ 1{2 ˇfNW pXq ´ |f˜K |pXqˇ dX 2 “ lim nÑ8 Sp pRq ż ˇ ˇ 1{2 ˇ|f˜ |pXq ´ ş ˇ K ` lim nÑ8 Sp pRq ˇ2 1{2 ˇ |f˜K |pXq ˇ dX 12 ˇ 1{2 ˜ |fK |pY qdY Sp pRq ż ď lim nÑ8 ˇ ˇ2 1 ˇ 1{2 ˇ 1{2 ˇfNW pXq ´ f˜K pXqˇ dX 2 Sp pRq ż ˇ ˇ ` lim ˇˇ1 ´ ş nÑ8 ˇ ˇ 1 ˇ ˇ 1{2 ˜ | f |pY qdY K Sp pRq 1 |f˜K |pXqdX 2 Sp pRq Chételat and Wells/Mid-scale Wishart asymptotics 56 ˇ 1 ˇˇ ˇ “ 01{2 ` ˇ1 ´ 1{2 ˇ ¨ 11{2 “ 0. 1 when pK`3 {nK`1 Ñ 0. Thus HpfNW , fK q, hence dTV pfNW , fK q, tends to zero when n Ñ 8 with pK`3 {nK`1 Ñ 0, as desired. We defined fK in terms of the inverse G-transform of the ψK functions given by Definition 3. How can we express this explicitely? By Equation (5.3), we see that |ψK |pT q is asymptotically bounded by the Tn{2 pIp {8q density |ψNW |pT q, which is integrable whenever n ´ p ` 2 ě 0. But |ψNW |1{2 pT q is proportional n Ip q density in the sense of Definition 2, which is integrable for to a Tm{4 p 4m m{4 ě p{2 ´ 1, that is whenever n ´ 3p ` 3 ě 0. Thus |ψNW |1{2 pT q and therefore |ψK |1{2 pT q is integrable whenever n ´ 3p ` 3 ě 0. Hence we can use the Fourer inversion theorem to conclude that fK is proportional to the integral ˇ ˇ ˇ ˇ fK pXq 9 ˇG ´1 tψK uˇpXq ˇ ˇ ˇ 9 ˇ ˇ ż 2K`3` 1rK odds ÿ ´ 4i ¯k tr T k n ? exp i trpXT q ` 4 n k # k“2 Sp pRq 2K`2´ 1rK odds + ˇ2 ÿ ˇ ´ 4i ¯k tr T k p `1 ˇ ? ` dT ˇ ˇ 4 n k (6.1) k“1 ? whenever n ´ 3p ` 3 ě 0. In particular, if we do a change of variables Z “ 8T , we obtain Equation (1.3) from Section 1 whenever n ě 3p ´ 3, from which we can derive Equations (1.1) and (1.2). It would be quite pleasant if there was a way to solve the integral in Equation (6.1) or (1.3) and obtain a (potentially quite complicated) closed form expression for fK up to its normalization constant. So far, our efforts have been unfruitful. We close our discussion with a final remark. At the end of Section 5, we showed that ψ0 approximates ψGOE in 0th degree middle-scale regimes, from which the classical asymptotic normality follows. It is natural to wonder if f0 approximates fGOE in the same context. An argument similar to that of Theorem 1 shows this is the case. Proposition 5. The total variation distance between the 0th degree density approximation f0 and the Gaussian orthogonal ensemble G-transform fGOE satisfies dTV pf0 , fGOE q Ñ 0 as n Ñ 8 with p3 {n Ñ 0. Proof. The Hellinger distance between f0 and fGOE satisfies ´ ¯ ´ ¯ ´ ¯ lim H f0 , fGOE ď lim H f0 , |f˜0 | ` lim H |f˜0 |, fGOE nÑ8 nÑ8 nÑ8 ż ˇ ˇ ď lim ˇˇ1 ´ ş nÑ8 Sp pRq ˇ ˇ 1 ˇ ˇ 1{2 ˜ |f |pY qdY ´ ¯ 1 |f˜0 |pXqdX 2 ` lim H f˜0 , fGOE nÑ8 0 Sp pRq Chételat and Wells/Mid-scale Wishart asymptotics 57 ˇ ´ ¯ 1 ˇˇ ˇ “ ˇ1 ´ 1{2 ˇ ¨ 11{2 ` lim H ψ0 , ψGOE “ 0. nÑ8 1 By Equation (2.1), the result follows. Of course, we could conclude from this that HpfNW , fGOE q ď HpfNW , f0 q ` Hpf0 , fGOE q Ñ 0 when n Ñ 8 with p3 {n Ñ 0, offering yet again another proof that a Gaussian orthogonal ensemble approximation holds in the classical setting. 7. The effect of phase transitions Although we have established the existence of phase transitions, it does not shed much light on how the behavior of a normalized Wishart distribution might differ across phase transitions. To do this, it can be very illuminating to study the asymptotics of some of its statistics. For example, we could study its empirical moments. ? For a normalized Wishart matrix X „ nrWp pn, Ip {8q ´ Ip s, a direct computation yields „ˆ ˙  ˆ ˙ 2 1 ´ X ¯2 5 4 8 20 20 5 1 E tr ? ´1 “ 2` 3` ` 2 ` 3 “ 2 `o p p p p np np np p p2 so in every middle-scale regime, that is whenever n, p Ñ 8 with p{n Ñ 0, ? › › ´ ¯ › › 1 ´ X ¯2 5 all middle-scale › › tr ? „ ´ 1 . › 2 ›p regimes p p L Thus we have L2 convergence of the second empirical moment to 1, but otherwise nothing very interesting. There doesn’t seem to be any change of behavior across the different middle-scale regimes. In contrast, the situation with the symmetric t distribution is striking, and illustrates yet again that middle-scale regime behavior becomes clearer under a G-transform. Indeed, we know from 4T 2 Theorem 2 that for a T „ Tn{2 pIp {8q, the quantity p1 trp ? p q also converges to 1, but we know more. At Equation (4.41), we computed the exact L2 distance 4T 2 between p1 trp ? p q and 1, and found that ˙  „ˆ 2 m5 1 ´ 4T ¯2 E tr ? ´1 “ p p pm´6q pm´2q pm´1q pm`1q pm`3q ˙ „ ˆ 2 p2 2 p2 5 1 ¨ 2` ` 2 `o ` ` . p m m p2 m m2 Thus the L2 distance must have middle-scale asymptotics › › › 1 ´ 4T ¯2 › › tr ? ´ 1›› ›p p L2 Chételat and Wells/Mid-scale Wishart asymptotics ? $ 5 ’ ’ ’ ’ ’ p ’ ’ ’ ’ ’ ’ ? 1 ’ 2 ’ ’ & 5 ` 2α ` α p ? „ 1 ’ “ 5α´1 ` 2 ` α ? ’ ’ n ’ ’ ? ’ ’ ´2 ´1 ’ “ 5α ` 2α ` 1 p ’ ’ n ’ ’ ’ p ’ % n 58 for p2 Ñ0 n for p2 Ñα n for ¯ ´ p2 second or higher Ñ8 degree n ´ ` classical or first degree second degree ¯ ˘ (7.1) 4T 2 as n, p Ñ 8 with p{n Ñ 0. Thus there is a sharp change in behavior of p1 trp ? pq ? when p grows like n, and despite the symmetric t distribution satisfying a semicircle law according to Corollary 1, it must ultimately behave differently than a Gaussian orthogonal ensemble matrix. The first-order asymptotics look the same: it is rather in the rate of this convergence that they differ. This matters for both the symmetric t and the Wishart distribution because rates of convergence can be distinguished in the strong topology. As a simple example, consider the sequence of one-dimensional distributions ` ˘ ` ˘ Fp “ N 0, 1{p , and Gp “ N 0, 1{p2 . In the weak topology, these are asymptotically the same, since they converge to the same distribution – namely Fp , Gp ñ δ0 as p Ñ 8, for δ0 the Dirac measure at 0. In other words, in a metric that induces the weak topology such as the Lévy metric, dLévy pFp , Gp q Ñ 0. Yet, by a direct computation of the Hellinger distance, which induces the strong topology, d ´ ¯1{4 ? 4p dHellinger pFp , Gp q “ HpFp , Gp q “ 2 1 ´ 2 p ` 2p ` 1 ? Ñ 2ą0 as p Ñ 8. Thus it is clear that the strong topology captures rates of convergence in a way that the weak topology can’t. But then, we should expect ? a phase transition when p grows like n for the Tn{2 pIp {8q distribution. And since the symmetric t is the G-conjugate of the Wishart, this should imply a ? phase transition when p grows like n for the Wishart distribution as well. This is consistent with Theorem 3, and provides an alternative explanation for the existence of the second phase transition. A natural question then is to ask whether we can find symmetric t statistics that exemplify all the middle-scale regime phase transitions. It is tempting to look at the L2 error of the other empirical moments of the symmetric t distribution, because we can use the methodology developed in Section 4 to compute their asymptotics to arbitrary order. As a reference, we compiled a table of the Chételat and Wells/Mid-scale Wishart asymptotics Normalized empirical moment 1 ´ 4T ¯ tr ? p p 1 ´ 4T ¯2 tr ? p p 1 ´ 4T ¯3 tr ? p p 1 ´ 4T ¯4 tr ? p p Limit Asymptotics of its squared L2 error 0 2 p2 C1 “ 1 5 2 p2 ` ` 2 2 p m m 0 24 p2 C2 “ 2 50 25p2 97 ` ` 2 p m m2 59 Table 1 Asymptotics of small normalized empirical moments of T „ Tn{2 pIp {8q. few first few moments as Table 1. As can be seen from the table, the odd moments seem to have uniform behavior across all middle-scale regimes. In contrast, the even ? moments seem to all change asymptotics at the second phase transition p “ Θp nq, but nowhere else. Hence finding statistics that “flag” the other phase transitions remain an open question. 8. Auxiliary results This section compiles several lemmas used elsewhere in the article. Lemma 3 (First derivatives lemma).For any indices 1 ď i1 , . . . , i2l ď p and real symmetric matrix Z, there exist polynomials aJ,s pn, mq in n and m “ n ´ p ´ 1, indexed by 0 ď s ď l and J “ pj1 , . . . , j2l q, such that " * ˇ ˇm Bs Bs Bs n ... exp ´ tr Z ˇZ ˇ 4 Bs Zi2l i2l´1 Bs Zi4 i3 Bs Zi2 i1 4 " * s l l ź ÿ ÿ ź ˇ ˇm n ´1 Zj2t j2t´1 exp ´ tr Z ˇZ ˇ 4 . “ aJ,s pn, mq pIp qj2t j2t´1 4 t“1 s“0 t“s`1 JP t1,...,pu2l Proof. To simplify notation, let * " l s ź ź ˇ ˇm n ´1 MJ,s pZq “ pIp qj2t j2t´1 Zj2t j2t´1 exp ´ tr Z ˇZ ˇ 4 , 4 t“s`1 t“1 and let Ml “ tMJ,s | J P t1, . . . , pu2l , s ď lu be the set of all such terms “on 2l indices”. Let xMl y denote the linear span of Ml , that is, the space of all linear combinations of elements of Ml , with as coefficients real polynomials in n and m. Then we are really claiming that " * ˇ ˇm Bs n Bs Bs ... exp ´ tr Z ˇZ ˇ 4 P xMl y. (8.1) Bs Zi2l i2l´1 Bs Zi4 i3 Bs Zi2 i1 4 Chételat and Wells/Mid-scale Wishart asymptotics 60 To see this, let J “ pj1 , . . . , j2l´2 q P t1, . . . , pu2l´2 and define the extension q Ja,b “ pj1 , . . . , jq´1 , a, b, jq`1 , . . . , j2l´2 q P t1, . . . , pu2l to be J with indices a, b inserted (in this order) at the q th position. Then using that ı 1 ” ´1 ´1 Bs ´1 ´1 Zab “ ´ Zai Zi2l´1 b ` Zai Zi´1 2l 2l´1 2l b Bs Zi2l i2l´1 2 and " * ˇ ˇm n exp ´ tr Z ˇZ ˇ 4 Bs Zi2l i2l´1 4 " * ”m ı ˇ ˇm n n “ Zi2l i2l´1 ´ pIp qi2l i2l´1 exp ´ tr Z ˇZ ˇ 4 , 4 4 4 we conclude that s s Bs 1 ÿ 1 ÿ MJi2r i ´ M 2r MJ,s pZq “ ´ ,s`1 2l 2l´1 Bs Zi2l i2l´1 2 r“1 2 r“1 Ji2l´1 i2l ,s`1 m n ` MJ 2s`1 ,s`1 ´ MJ 2s`1 ,s P xMl y. i i i2l i2l´1 4 4 2l 2l´1 Thus, by linearity, Bs {Bs Zi2l i2l´1 maps xMl´1 y to xMl y. But naturally we have expt´ n4 tr Zu|Z|m{4 P xM0 y, so by induction Equation (8.1) must then hold, as desired. Bs Lemma 4 (Second derivatives lemma).For any k P N and any Z P Sp pRq, p ÿ i1 ,...,i2k ˇ ˇm n Bs Bs Bs ... e´ 4 tr Z ˇZ ˇ 4 Bs Zi1 i2k B s Z i3 i2 B s Z i2 i1 n “ e´ 4 ˇ ˇ m4 ÿ p1q Zˇ bκ pn, m, pqrκ pZ ´1 q tr Z ˇ |κ|ď2k and p ÿ i1 ,...,ik j1 ,...,jk ˇ ˇm n Bs Bs Bs Bs Bs Bs e´ 4 tr Z ˇZ ˇ 4 ... ... Bs Zj1 jk Bs Zj3 j2 Bs Zj2 j1 Bs Zi1 i2k Bs Zi3 i2 Bs Zi2 i1 n “ e´ 4 ˇ ˇ m4 ÿ p2q Zˇ bκ pn, m, pqrκ pZ ´1 q tr Z ˇ |κ|ď2k`1 p1q bκ pn, m, pq p2q bκ pn, m, pq p1q with degrees deg bκ ď 2k ` and for some polynomials p2q 1 ´ qpκq and deg bκ ď 2k ` 2 ´ qpκq. The sums at the right hand sides are taken over all integer partitions κ of norm at most 2k and 2k ` 1, including the empty partition. Proof. We give a spectral proof. Let OLOt be the spectral decomposition of Z, with eigenvalues λ1 ě ¨ ¨ ¨ ě λp , and notice that p t t Bs Ohl 1 ÿ Oha Oai 1 ÿ Oha Oaj “ Ojl ` Oil , Bs Zij 2 a‰l λl ´ λa 2 a‰l λl ´ λa Bs λ h “ Oih Ojh Bs Zij for any 1 ď i, j, h, l ď p. As a consequence, for any differentiable real-valued Chételat and Wells/Mid-scale Wishart asymptotics 61 functions F1 pLq, . . . , Fp pLq, we have ˆÿ ˙ p p p p ÿ ÿ Fb ´ Fa t Bs 1 ÿ BFa t t Oha Oja Fa Oai “ Oai ` Oha O . B Z 2 λ ´ λ Bλa ai s hj b a a“1 a“1 j“1 a,b b‰a This suggests we define a new operator DL that would map the space of diagonal matrices F pLq “ diagpF1 pLq, . . . , Fp pLqq that differentially depends on L, to itself, by p p ÿ BFa 1 ÿ Fb ´ Fa Bs t t ` so that OF Oji “ OF Oki . DL tF ua “ 2 b‰a λb ´ λa Bλa B Z j“1 s hj In particular, p ÿ i1 ,...,i2k ˇ ˇm n Bs Bs Bs ... e´ 4 tr Z ˇZ ˇ 4 Bs Zi1 i2k Bs Zi3 i2 Bs Zi2 i1 p ÿ “ i1 ,...,i2k Bs ” ´ n tr Z ˇˇ ˇˇ m4 ı Bs Bs Z Ip e 4 ... Bs Zi0 i2k Bs Zi3 i2 Bs Zi2 i1 i1 i0 ! n ˇ ˇm ) 2k “ tr DL e´ 4 tr Z ˇZ ˇ 4 Ip , (8.2) and similarly p ÿ i1 ,...,ik j1 ,...,jk ˇ ˇm n Bs Bs Bs Bs Bs Bs e´ 4 tr Z ˇZ ˇ 4 ... ... Bs Zj1 jk Bs Zj3 j2 Bs Zj2 j1 Bs Zi1 ik B s Z i3 i2 B s Z i2 i1 ! ! n ˇ ˇ m )) k k “ tr DL tr DL e´ 4 tr Z ˇZ ˇ 4 Ip . (8.3) Let us look more closely at this operator DL . It satisfies the following. (i) DL is linear, in the sense that for diagonals F pLq, GpLq and constants a, b with respect to L, DL taF ` bGu “ aDL tF u ` bDL tGu. (ii) DL satisfies a restricted product rule, in the sense that for a diagonal F pLq of the form F pLq “ f pLqIp for some function f pLq, and any diagonal GpLq, DL tF Gu “ DL tF uG ` F DL tGu. Moreover, from the definition of DL , ( n n n DL e´ 4 tr L Ip “ ´ e´ 4 tr L Ip , 4 ( DL trpL´s qIp “ ´sL´ps`1q , and ( m m m DL |L| 4 Ip “ |L| 4 Ip , 4 s ( s 1ÿ trpL´rs`1´ts qL´t . DL L´s “ ´ L´ps`1q ´ 2 2 t“1 Chételat and Wells/Mid-scale Wishart asymptotics 62 Now define the spaces # Ml “ ´n 4 bpn, m, pqe ˇ + ˇ bpn, m, pq is a polynomial with m ˇ ˇ ˇ tr L ˇ ˇ 4 L rκ pL´1 qL´s ˇ degree at most l ´ qpκq, and κ ˇ and s satisfy |κ| ď l ´ s. for l “ 1, . . . , 2k, and let xMl y denote the linear span of Ml , i.e. the space of all real linear combinations of elements of Ml . Moreover, for a partition κ, let κ ˘ i denote κ with the integer i added or removed, respectively. For example, p3, 1, 1, 1q`2 “ p3, 2, 1, 1, 1q and p3, 2, 1, 1, 1q´1 “ p3, 2, 1, 1q. Note that |κ˘i| “ |κ| ˘ i. Then, for any F P Ml , ) ! ˇ ˇm n DL tF u “ DL bpn, m, pqe´ 4 tr L ˇLˇ 4 rκ pL´1 qL´s ! n )ˇ ˇ m “ bpn, m, pqDL e´ 4 tr L Ip ˇLˇ 4 rκ pL´1 qL´s !ˇ ˇ m ) n ` bpn, m, pqe´ 4 tr L DL ˇLˇ 4 Ip rκ pL´1 qL´s ! ) ˇ ˇm n ` bpn, m, pqe´ 4 tr L ˇLˇ 4 DL rκ pL´1 qIp L´s ! ) ˇ ˇm n ` bpn, m, pqe´ 4 tr L ˇLˇ 4 rκ pL´1 qDL L´s ı n ” n ˇ ˇm “ ´ bpn, m, pq e´ 4 tr L ˇLˇ 4 rκ pL´1 qL´s ” m4 ı n ˇ ˇm ` bpn, m, pq e´ 4 tr L ˇLˇ 4 rκ pL´1 qL´s 4 qpκq ı n ÿ” ˇ ˇm ` ´ κi bpn, m, pq e´ 4 tr L ˇLˇ 4 rκ´κi pL´1 qL´s i“1 ı n ˇ ˇm s ` ´ bpn, m, pq e´ 4 tr L ˇLˇ 4 rκ pL´1 qL´ps`1q 2 s ” ı n ÿ ˇ ˇm 1 ` ´ bpn, m, pq e´ 4 tr L ˇLˇ 4 rκ`ps`1´tq pL´1 qL´t 2 t“1 ” Thus DL tF u P xMl`1 y. It follows by linearity that DL maps xMl y to xMl`1 y. ˇ ˇm ˇ ˇm n 2k ´ n te 4 tr L ˇLˇ 4 Ip u P xM2k y. Now, e´ 4 tr L ˇLˇ 4 Ip P M0 , so by induction DL p1q Hence, for some polynomials bκ,s pn, m, pq of degree at most 2k ´ qpκq, ! n ÿ ˇ ˇm ) ˇ ˇm 2k ´n 4 tr L ˇLˇ 4 r pL´1 q trpL´s q tr DL e´ 4 tr L ˇLˇ 4 Ip “ bp1q κ κ,s pn, m, pqe |κ|`sď2k ÿ “ p1q n bκ1 pn, m, pqe´ 4 ˇ ˇ m4 Lˇ rκ1 pL´1 q tr L ˇ (8.4) |κ1 |ď2k p1q p1q p1q p1q for κ1 “ κ ` s, bκ1 “ bκ,s when s ‰ 0, while κ1 “ κ, bκ1 “ pbκ,s when s “ 0. Notice that when s ‰ 0, the degree of the bκ1 ’s is at most 2k ´ qpκq “ 2k ´ pqpκ1 q ´ 1q, while when s “ 0 it is at most 2k ´ qpκq ` 1 “ 2k ´ qpκ1 q ` 1. p1q Thus in both cases, deg bκ1 ď 2k ´ qpκ1 q ` 1, which by Equation (8.2) shows the first statement of the lemma. For the second statement of the lemma, by an argument analoguous to Equa- Chételat and Wells/Mid-scale Wishart asymptotics 63 ˇ ˇ m4 Lˇ Ip u P xMk`1 y. Thus by induction ˇ ˇm n k k again, we must have DL ttr DL te´ 4 tr L ˇLˇ 4 Ip uu P xM2k`1 y. Hence for some n k tion (8.4) we find that tr DL te´ 4 tr L ˇ p2q polynomials bκ,s pn, m, pq of degree at most 2k ` 1 ´ qpκq, ! ! n ˇ ˇ m )) k k DL tr DL e´ 4 tr L ˇLˇ 4 Ip ÿ ˇ ˇm ´n 4 tr L ˇLˇ 4 r pL´1 q trpL´s q “ bp2q κ κ,s pn, m, pqe |κ|`sď2k`1 ÿ “ n p2q bκ1 pn, m, pqe´ 4 ˇ ˇ m4 Lˇ rκ1 pL´1 q tr L ˇ |κ1 |ď2k`1 p2q p2q p2q p2q for again κ1 “ κ ` s, bκ1 “ bκ,s when s ‰ 0, while κ1 “ κ, bκ1 “ pbκ,s when p2q s “ 0. By the same argument as before, deg bκ1 ď 2k ´ qpκ1 q ` 2, which by Equation (8.3) shows the second statement of the lemma. This concludes the proof. We will also need in our proof a result about the asymptotics of inverse moments of the Wishart distribution. Because we couldn’t find anything like it in the literature, we think it is worthwhile to provide some context. Let f : p0, 4q Ñ R be the restriction to the positive reals of a complex function analytic in a neighborhood of p0, 4q. We are often interested in the linear spectral statistic p1 tr f pY q for Y „ Wp pn, Ip {nq. Much is known about its distributional properties in the high-dimensional regime where p Ñ 8 such that lim np “ α ă 1. For example, if 0 ă α ă 1, there must be an  ą 0 such nÑ8 that p{n P r, 1 ´ s for all n large enough, so Bai and Silverstein [2010, Theorem 9.10] and the dominated convergence theorem yield that ż rp{ns` a prp{ns` ´ tqpt ´ rp{ns´ q f ptq dt 2πrp{nst 1 P tr f pY q ÝÑ lim nÑ8 p ż α rp{ns ´ ` “ a pα` ´ tqpt ´ α´ q dt f ptq 2παt (8.5) α´ ? P as n Ñ 8. Here, Ñ stands for convergence in probability and x˘ for p1 ˘ xq2 . In fact, the theorem states more, namely a central limit theorem, but what we want to draw to attention is the class of functions for which this result was proven. This is sometimes enough, but often we would like to understand the expectation of this linear spectral statistic. If f is bounded, then Equation (8.5) implies that α` a ı pα` ´ tqpt ´ α´ q 1 ” f ptq dt. (8.6) lim E tr f pY q “ nÑ8 p 2παt ż α´ This is nice for a function f pzq like ez or sin z that happens to be bounded Chételat and Wells/Mid-scale Wishart asymptotics 64 on a neighborhood of p0, 4q, but it unfortunately excludes many interesting unbounded functions, such as log z or 1{z. In fact, for unbounded f , it is in general not even clear if lim p1 Ertr f pY qs will be finite! nÑ8 The following result shows that, at least in the case f pzq “ 1{z s for s P N, we can use Stein’s lemma to obtain Equation (8.6) and its α “ 0 analogue. Lemma 5. Let for Y „ Wp pn, Ip {nq and s be any integer s ě 1. Then as long as n ě p ` 4s ` 2, the sth inverse moment satisfies the recursive bound ˆ ˙ ” ı ” ı pp ` 1qs 1´ E tr Y ´s ď E tr Y ´ps´1q . n In particular, as p Ñ 8 such that lim np “ α ă 1, if s ă α´1 ´ 1 then nÑ8 $ α` a ’ ’ pα` ´ tqpt ´ α´ q & ı ’ 1 ” dt if 0 ă α ă 1, ´s lim E tr Y “ 2παts`1 nÑ8 p ’ α ´ ’ ’ % 1 if α “ 0. ? 2 for α˘ “ p1 ˘ αq . ż Proof. The classical lemma that for any differentiable function f : ˇstates “ˇ`Stein’s ˘ ‰ 2 B R Ñ R such that E ˇ BZ ´ Z f pZqˇ ă 8 for Z „ Np0, 1q and lim f pzqe´z {2 “ zÑ˘8 0, we must have  „´ ¯ B ´ Z f pZq “ 0. E BZ Let Z „ Nnˆp p0, In b Ip q be an n ˆ p matrix of i.i.d. standard normal random variables, and let Y “ n1 Z t Z „ Wp pn, Ip {nq. For any 1 ď α ď n and 1 ď β, i, j ď p, ´s p s ı Bs Yβj B 2 ÿ 1 ÿ ” ´l ´ps´l`1q Bs ´l ´ps´l`1q Yβi Yβj ` Yββ “ Zαi and “´ Yij , BZαβ n i“1 Bs Yiβ Bs Yiβ 2 l“1 so for δ the Kronecker delta, ˙ ˆ ` ˘ B ´ Zαβ ZY ´s αβ BZαβ  p „ p ÿ 2 ÿ Bs ´s ´s ´s “ δβj Yβj ` Zαj Zαi Y ´ Zαβ Zαj Yβj n i“1 Bs Yiβ βj j“1 ´s “ Yββ ´ s s ˘ ` ˘ ˘ 1 ÿ` 1 ÿ ´l ` ZY l αβ ZY ´ps´l`1q αβ ´ Yββ ZY ´ps´l`1q Z t αα n l“1 n l“1 ` ˘ ´Zαβ ZY ´s αβ (8.7) Let us first show that this expression is integrable. For any matrix X, |Xij | ď 1{2 }X}2 “ }X t X}2 . Thus by Equation (8.7), ˇ „ˇˆ ˙ ˇ ` ˘ ˇ B ´s ˇ ˇ E ˇ ´ Zαβ ZY αβ ˇ BZαβ Chételat and Wells/Mid-scale Wishart asymptotics « ď E }Y ´s }2 ` s ÿ 1 65 1 }Y 2l`1 }22 }Y 2s`2l´1 }22 l“1 ff s ÿ }Y ` ´2l 1 2 }2 }Y ´2s`2l 1 2 1 2 }2 ` n}Y }2 }Y ´2s`1 1 2 }2 l“1 As Y is positive definite, }Y ˘a }2 ď tr Y ˘a for any a P N, so by the CauchySchwarz inequality, „  ÿ s ” ı 12 ” ı 12 E tr Y ´2l`1 E tr Y ´2s`2l´1 ď E tr Y ´s ` l“1 s ” ı 12 ” ı 12 ” ı 12 ” ı 12 ÿ E tr Y ´2l E tr Y ´2s`2l ` n E tr Y E tr Y ´2s`1 , ` l“1 which is finite for n ě p ` 4s ` 2. Moreover, pZY ´s qαβ can be expressed using minors and determinants as a rational function of the entries of Z, so lim Zαβ Ñ˘8 2 pZY ´s qαβ e´Zαβ {2 “ 0. So all conditions are fulfilled to apply Stein’s lemma to Equation 8.7 and obtain ff « ˙ p ˆ n ` ˘ B 1 ÿ ÿ ´s ´ Zαβ ZY 0“E αβ n α“1 β“1 BZαβ « ff s s 1 ÿ ´s ´s ´l ´ps´lq ´ps´1q “ E tr Y ´ tr Y ´ trpY q trpY q ´ tr Y n n l“1 As trpY ´l q trpY ´ps´lq q ď p tr Y ´s for any 1 ď l ď s, and every term is integrable as n ě p ` 4s ` 2, this means that ˆ ˙ ” ı ” ı pp ` 1qs 1´ E tr Y ´s ď E tr Y ´ps´1q . (8.8) n This shows the first part of the proof. For the second part, let S P N. If we let n Ñ 8 such that lim np “ α ă 1, nÑ8 then any S ă α´1 we will have n ě p ` 4S ` 2 and n ě pp ` 1qS for n large enough. So by repeatedly applying Equation (8.8) for s “ S, . . . , 1 and dividing by p, we obtain ˙ S ˆ ı ı ź pp ` 1ql 1 ” 1 ” 1´ ¨ E tr Y ´S ď E tr Y ´0 “ 1. n p p l“1 Taking a limit in the above yields S ı ź 1 ” p1 ´ αlq lim E tr Y ´S ď 1. nÑ8 p l“1 Thus for any S ă α´1 , we have S ı ź 1 1 ” lim E tr Y ´S ď ă 8. nÑ8 p 1 ´ αl l“1 (8.9) Chételat and Wells/Mid-scale Wishart asymptotics 66 In the case 0 ă α ă 1, if s`1 ă α´1 then by Jensen’s inequality and Equation (8.9) applied to S “ s ` 1, we have ı ˘1` 1s ı 1 ” 1 ”´ ď lim E tr Y ´ps`1q ă 8. lim E tr Y ´s nÑ8 p nÑ8 p 1 ´s Thus p tr Y is uniformly integrable, and by Equation (8.5), żα ` a ” ı pα` ´ tqpt ´ α´ q 1 dt lim E tr T ´s “ nÑ8 p 2παts`1 α´ ? for α˘ “ p1 ˘ αq2 . In contrast, by applying Jensen’s inequality twice, „´ „ ´s „  ¯´s  1 1 1 ďE 1 “ E tr Y tr Y ď E tr Y ´s p p p so when α “ 0, by applying Equation (8.9) with S “ s, we obtain that lim p1 Ertr Y ´s s “ 1, as desired. nÑ8 9. Conclusion The results of this paper raise more questions than they answer. We enumerate some that we found particularly interesting. (1) The univariate t distribution with ν degrees of freedom is often defined as ? the distribution of Z{ s, for Z „ Np0, 1q and s „ χ2ν {ν independent. In the real symmetric matrix case, we could imagine studying the distribution of S 1{4 ZS 1{4 , for Z „ GOEppq and S „ Wp pν, Ip {νq independent. Is this the Tν p2Ip q distribution in the sense of Section 4? (2) By Theorem 2 and Corollary 1, it is clear the empirical moments of a symmetric t distribution are quite similar to those of a Gaussian orthogonal ensemble matrix, except perhaps in their rates of convergence. From Anderson et al. [2010, Theorem 2.1.31], we know the empirical moments of a Gaussian orthogonal ensemble are asymptotically normal. Are the empirical moments of the symmetric t distribution also asymptotically normal? (3) In Section 4, we showed that the rate of convergence of the even normalized empirical moments of a symmetric t distribution change when p grows like ? n. Can we find analogue symmetric t statistics that change their rates of convergence when p grows like npK`1q{pK`3q for every K P N? This would establish phase transitions for the symmetric t distribution. If so, can we find approximating densities between every two transitions, just like in the Wishart case? (4) As a counterpart of Theorem 1, could we prove that dTV pfNW , fK q Û 0 whenever pK`3 {nK`1 Û 0 as n Ñ 8? This is delicate because we have no guarantee that the L1 norm of ψK is asymptotically bounded for regimes of degree K ` 1 or higher. Chételat and Wells/Mid-scale Wishart asymptotics 67 (5) Can we find the normalization constant or, better, solve the expectation of Equation (1.3) in closed form? (6) What asymptotics hold for the symmetric t or the Wishart distribution in a middle-scale regime of infinite degree? How do these asymptotics differ from the other middle-scale regimes, or the high-dimensional regime? (7) The symmetric t distribution was discovered as the G-conjugate of the Wishart distribution. What other distributions can be realized as the Gconjugate of some well-known distribution? (8) In Lemma 2, we expressed the characteristic function of the G-conjugate F ˚ of a distribution F as f 1{2 ‹ pf 1{2 ˝ Rq, for f the density of F and R the flip operator. To obtain the moments, we then repeatedly differentiated under the convolution integral at zero, and obtained an expression of the moments as an expectation with respect to f . The argument worked when F ˚ was the symmetric t distribution. Can this argument be generalized to other F ˚ ? If F ˚ is a well-known distribution, does this give rise to novel and nontrivial expressions for its moments? (9) The G-transform of a distribution encodes all the information relative to that distribution. However, taking a modulus removes some information, and so in some sense the G-conjugate distribution is “less informative” than the original distribution. What happens when we repeatedly apply the G-conjugation operator, destroying information every time? For example, is there an attractor distribution G that is the limit of this process regardless of the initial distribution? (10) Can we find distinct random operators which can be regarded, in some sense, as the total variation limit of a normalized Wishart matrix between every two phase transitions? It appears to us that some of these questions might be very difficult to answer. We would be pleased if future work were able to shed light on any of them. References Greg W. Anderson, Alice Guionnet, and Ofer Zeitouni. An Introduction to Random Matrices. Cambridge University Press, 2010. Zhidong Bai and Jack W. Silverstein. Spectral Analysis of Large Dimensional Random Matrices, volume 20. Springer, 2010. Maurice S. Bartlett. On the theory of statistical regression. Proceedings of the Royal Society of Edinburgh, 53:260–283, 1933. Sébastien Bubeck and Shirshendu Ganguly. Entropic CLT and phase transition in high-dimensional Wishart matrices. International Mathematics Research Notices, pages 243–258, 2016. Sébastien Bubeck, Jian Ding, Ronen Eldan, and Miklós Z. Rácz. Testing for high-dimensional geometry in random graphs. Random Structures & Algorithms, 49:503–532, 2016. Kai-Lai Chung. A Course in Probability Theory. Academic Press, 2001. Chételat and Wells/Mid-scale Wishart asymptotics 68 Michel Gaudin. Sur la loi limite de l’espacement des valeurs propres d’une matrice alé atoire. Nuclear Physics, 25:447–458, 1961. Arjun K. Gupta and Daya K. Nagar. Matrix Variate Distributions. CRC Press, 1999. Tiefeng Jiang and Danning Li. Approximation of rectangular beta-Laguerre ensembles and large deviations. Journal of Theoretical Probability, 28:804– 847, 2015. Gérard Letac and Hélène Massam. All invariant moments of the wishart distribution. Scandinavian Journal of Statistics, 31:295–318, 2004. Vladimir A. Marchenko and Leonid A. Pastur. Distribution of eigenvalues for some sets of random matrices. Matematicheskii Sbornik, 114:507–536, 1967. Sho Matsumoto. General moments of the inverse real wishart distribution and orthogonal weingarten functions. Journal of Theoretical Probability, 25:798– 822, 2012. Madan L. Mehta. On the statistical properties of the level-spacings in nuclear spectra. Nuclear Physics, 18:395–419, 1960a. Madan L. Mehta. Lemploi des polynômes orthogonaux pour calculer certain déterminants. Rapport S PH (Saclay), 658, 1960b. Robb J. Muirhead. Aspects of Multivariate Statistical Theory. Wiley Series in Probability and Mathematical Statistics. Wiley, 1982. Charles E. Porter and Norbert Rosenzweig. Statistical properties of atomic and nuclear spectra. Technical report, Univ. of Minnesota, Minneapolis, 1960. Mohsen Pourahmadi. High-dimensional Covariance Estimation: with HighDimensional Data. John Wiley & Sons, 2013. Miklós Z. Rácz and Jacob Richey. A smooth transition from Wishart to GOE. arXiv:1611.05838, 2016. Aad W. van der Vaart. Asymptotic statistics, volume 3. Cambridge University Press, 2000. John Von Neumann and Herman H. Goldstine. Numerical inverting of matrices of high order. Bulletin of the American Mathematical Society, 53:1021–1099, 1947. Eugene P. Wigner. On the statistical distribution of the widths and spacings of nuclear resonance levels. In Mathematical Proceedings of the Cambridge Philosophical Society, volume 47, pages 790–798, 1951. Eugene P. Wigner. Characteristic vectors of bordered matrices with infinite dimensions. Annals of Mathematics, 62:548–564, 1955. Eugene P. Wigner. Statistical properties of real symmetric matrices with many dimensions. In Proceedings of the Fourth Canadian Mathematical Congress. University of Toronto Press, 1957. John Wishart. The generalised product moment distribution in samples from a normal multivariate population. Biometrika, 20A:32–52, 1928. 

10math.ST

1 Time-Varying Block Codes for Synchronization Errors: MAP Decoder and Practical Issues arXiv:1802.08595v1 [cs.IT] 23 Feb 2018 Johann A. Briffa, Victor Buttigieg, Member, IEEE, and Stephan Wesemeyer Abstract—In this paper we consider Time-Varying Block (TVB) codes, which generalize a number of previous synchronization error-correcting codes. We also consider various practical issues related to MAP decoding of these codes. Specifically, we give an expression for the expected distribution of drift between transmitter and receiver due to synchronization errors. We determine an appropriate choice for state space limits based on the drift probability distribution. In turn, we obtain an expression for the decoder complexity under given channel conditions in terms of the state space limits used. For a given state space, we also give a number of optimizations that reduce the algorithm complexity with no further loss of decoder performance. We also show how the MAP decoder can be used in the absence of known frame boundaries, and demonstrate that an appropriate choice of decoder parameters allows the decoder to approach the performance when frame boundaries are known, at the expense of some increase in complexity. Finally, we express some existing constructions as TVB codes, comparing performance with published results, and showing that improved performance is possible by taking advantage of the flexibility of TVB codes. Index Terms—Insertion-Deletion Correction, MAP Decoder, Forward-Backward Algorithm I. I NTRODUCTION Most error-control systems are designed to detect and/or correct substitution errors, where individual symbols of the received sequence have been substituted while maintaining synchronization with the transmitted sequence. Some channels, however, also experience synchronization errors, where symbols may additionally be deleted from or inserted into the received sequence. It has long been recognized that codes can be designed specifically for synchronization error correction [1], [2]. Except for the less-known work by Gallager [3], for a long time only some short block codes were known. This changed when Davey and MacKay [4] proposed a concatenated scheme combining an outer LDPC code with good error-correction capability with an inner code whose aim is to correct synchronization errors. Ratzer [5] took a different approach, using short marker sequences inserted in binary LDPC codewords; a similar approach was used by Wang et al. [6]. Yet another approach extends the state space of convolutional codes to allow correction of synchronization errors [7]–[9]. The problem of convolutional code design for Manuscript submitted February 27, 2014; accepted May 29, 2014 for publication in the IET Journal of Engineering. This research has been carried out using computational facilities procured through the European Regional Development Fund, Project ERDF-080. J. A. Briffa and S. Wesemeyer are with the Dept. of Computing, University of Surrey, Guildford GU2 7XH, England. Email: j.briffa@surrey.ac.uk; s.wesemeyer@surrey.ac.uk V. Buttigieg is with the Dept. of Comm. & Computer Engineering, University of Malta, Msida MSD 2080, Malta. Email: victor.buttigieg@um.edu.mt synchronization error channels has been considered in [10]. More recently, this approach has been applied successfully to turbo codes [11]. The renewed increase in interest is mainly due to new applications requiring such codes. A recent survey can be found in [12]. We have in previous papers extended the work of DaveyMacKay, proposing a maximum a posteriori (MAP) decoder [13], improved code designs [14], [15], as well as a parallel implementation of the MAP decoder resulting in speedups of up to two orders of magnitude [16]. However, these papers were restricted to the case where the frame boundaries were known by the decoder. While Davey and MacKay showed that the frame boundaries could be accurately determined for their bit-level decoder and code construction [4], it has not been shown whether this property extends to our MAP decoder and improved constructions. In this paper we define Time-Varying Block (TVB) codes in terms of the encoding used in [16], and show that TVB codes represent a new class of codes which generalizes a number of previous synchronization error-correcting codes. We use the MAP decoder of [16] for these codes, showing how it can be used in an iterative scheme with an outer code. We also consider a number of important issues related to any practical implementation of the MAP decoder. Specifically, we give an expression for the expected distribution of drift between transmitter and receiver due to synchronization errors. We determine an appropriate choice for state space limits based on the drift probability distribution. In turn, we obtain an expression for the decoder complexity under given channel conditions in terms of the state space limits used. For a given state space, we also give a number of optimizations that reduce the algorithm complexity with no further loss of decoder performance. We also show how the MAP decoder can be used for stream decoding, where the boundaries of the received frames are not known a priori. In doing so we demonstrate how an appropriate choice of decoder parameters allows stream decoding to approach the performance when frame boundaries are known, at the expense of some increase in complexity. We express some previously published codes as TVB codes, comparing performance with published results, and showing that the greater flexibility of TVB codes permits the creation of improved codes. In the following, we start with definitions in Section II and summaries of results from earlier work. The applicable design criteria for TVB codes are considered in Section III, together with the representation of previously published codes as TVB codes. The appropriate choice for state space limits is given in Section IV, followed by expressions for the decoder 2 complexity in Section V. MAP decoder optimizations are given in Section VI and the changes necessary for stream decoding in Section VII. Finally, practical results are given in Section VIII. II. BACKGROUND A. TVB Codes Consider the encoding defined in [16], used there to simplify the representation of the inner code of [4]. We observe that this encoding generalizes a number of additional previous schemes, including the inner codes of [5], [6], [15] (c.f. Section III-B). We define a TVB code in terms of this encoding by the sequence C = (C0 , . . . , CN −1 ), which consists of the constituent encodings Ci : Fq ,→ Fn2 for i = 0, . . . , N − 1, where n, q, N ∈ N, 2n ≥ q, and ,→ denotes an injective mapping. Two constituent encodings Ci , Cj are said to be equal if Ci (D) ∈ Cj ∀D. For a given TVB code, the set of unique constituent encodings is that set where no two constituent encodings are equal; the cardinality of this set, denoted by M ≤ N , is called the order of the code. Note that unique constituent encodings may still have some common codewords. We denoted a TVB code by the tuple (n, q, M ). We restrict ourselves to binary TVB codes, where codewords are sequences of bits; the extension to the non-binary case is trivial. For any sequence z, denote arbitrary subsequences as zba = (za , . . . , zb−1 ), where zaa = () is an empty sequence. Given a message DN = (D0 , . . . , DN −1 ), each 0 Ci maps the q-ary message symbol Di ∈ Fq to codeword nN Ci (Di ) of length n. That is, DN = 0 is encoded as X0 C0 (D0 )k · · · kCN −1 (DN −1 ), where ykz is the juxtaposition of y and z. Each q-ary symbol is encoded independently of previous inputs, and different codebooks may be used for each input symbol. This time-variation offers no advantage on a fully synchronized channel. However, in the presence of synchronization errors, the differences between neighbouring codebooks provide useful information to the decoder to recover synchronization. In practice a TVB code is suitable as an inner code to correct synchronization errors in a serially concatenated construction. A conventional outer code corrects residual substitution errors. In such a scheme, the inner code’s MAP decoder a posteriori probabilities (APPs) are used to initialize the outer decoder. The concatenated code can be iteratively decoded, in which case the prior symbol probabilities of the inner decoder are set using extrinsic information from the previous pass of the outer decoder. B. Channel Model We consider the Binary Substitution, Insertion, and Deletion (BSID) channel, an abstract random channel with unbounded synchronization and substitution errors, originally presented in [17] and more recently used in [4], [5], [13]–[15] and others. At time t, one bit enters the channel, and one of three events may happen: insertion with probability Pi where a random bit is output; deletion with probability Pd where the input is discarded; or transmission with probability Pt = 1 − Pi − Pd . A substitution occurs in a transmitted bit with probability Ps . After an insertion the channel remains at time t and is subject to the same events again, otherwise it proceeds to time t + 1, ready for another input bit. We define the drift St at time t as the difference between the number of received bits and the number of transmitted bits before the events of time t are considered. As in [4], the channel can be seen as a Markov process with the state being the drift St . It is helpful to see the sequence of states as a trellis diagram, observing that there may be more than one way to achieve each state transition. Also note that the state space is unlimited for positive drifts but limited for negative drifts. Specifically, St may take any positive value for t > 0, though with decreasing probability as the value increases. On the other hand, St ≥ −t where the lower limit corresponds to receiving the null sequence. C. The MAP Decoder We summarize here the MAP decoder of [16]; this is the same as the MAP decoder of [13] with a trivial modification to work with the notation of TVB codes. The decoder uses the standard forward-backward algorithm for hidden Markov models. We assume a message sequence DN 0 , encoded using a (n, q, M ) TVB code to the sequence Xτ0 , where τ = nN . The sequence Xτ0 is transmitted over the BSID channel, resulting in the received sequence Y0ρ , where in general ρ is not equal to τ . To avoid ambiguity, we refer to the message sequence as a block of size N and the encoded sequence as a frame of size τ . We calculate the APP Li (D) of having encoded symbol D ∈ Fq in position i for 0 ≤ i < N , given the entire received sequence, using X 1 σi (m0 , m, D), (1) Li (D) = λN (ρ − τ ) 0 m ,m where λi (m) = αi (m)βi (m), (2) σi (m0 , m, D) = αi (m0 )γi (m0 , m, D)βi+1 (m), (3) and αi (m), βi (m), and γi (m0 , m, D) are the forward, backward, and state transition metrics respectively. Note that strictly, the above metrics depend on Y0ρ , but for brevity we do not indicate this dependence in the notation. The summation in (1) is taken over the combination of m0 , m, being respectively the drift before and after the symbol at index i. The forward and backward metrics are obtained recursively using X αi (m) = αi−1 (m0 )γi−1 (m0 , m, D), (4) m0 ,D and βi (m) = X βi+1 (m0 )γi (m, m0 , D). (5) m0 ,D Initial conditions for known frame boundaries are given by ( ( 1 if m = 0 1 if m = ρ − τ α0 (m) = and βN (m) = 0 otherwise, 0 otherwise. Finally, the state transition metric is defined as n(i+1)+m γi (m0 , m, D) = Pr {Di = D} R(Yni+m0 | Ci (D)) (6) where Ci (D) is the n-bit sequence encoding D and R(ẏ|x) is the probability of receiving a sequence ẏ given that x was sent 3 through the channel (we refer to this as the receiver metric). The a priori probability Pr {Di = D} is determined by the source statistics, which we generally assume to be equiprobable so that Pr {Di = D} = 1/q. In iterative decoding, the prior probabilities are set using extrinsic information from the previous pass of an outer decoder, as explained in Section II-A. The receiver metric is obtained by calculating the forward recursion X
t+m α̇t−1 (m0 ) · Q ẏt−1+m α̇t (m) = (7) 0 |xt−1 , m0 where for brevity we do not show the dependence on ẏ and x, and Q(y|x) can be directly computed from y, x and the channel parameters:  if µ = 0   Pd

Pi µ−1 1 Q(y|x) = Pt P s + 2 Pi P d if µ > 0, yµ−1 6= x 2 
 Pi
µ−1 1 Pt P̄s + 2 Pi Pd if µ > 0, yµ−1 = x, 2 where µ is the length of y and P̄s = 1 − Ps . The required value of the receiver metric is given by R(ẏ|x) = α̇n (µ̇ − n), where µ̇ is the length of ẏ, n is the length of x. As in [16], the α, β, and α̇ metrics are normalized as they are computed to avoid exceeding the limits of floatingpoint representation. We also assume that (7) is computed at single precision1 , while the remaining equations use double precision. III. TVB C ODE D ESIGN A. Construction Criteria In any error-correcting scheme, the decoder’s objective is to minimize the probability of decoding error (at the bit or codeword level depending on the application). If the channel does not introduce synchronization errors, this optimization may be performed independently of previous or subsequent codewords. Hence the performance of the code depends exclusively on its distance properties. In particular, the performance of the code at low channel error rate is dominated by the code’s minimum Hamming distance. At any channel error rate the performance is determined by the code’s distance spectrum [18], [19]. Thus when designing codes for substitution error channels, either the minimum Hamming distance or the more complete distance spectrum needs to be optimized for the given code parameters. In the case of the BSID channel, and other channels that allow synchronization errors, a similar behaviour is observed if the codeword boundaries are known, only this time the Levenshtein distance [2] replaces the Hamming one. Recall that the Levenshtein distance gives the minimum number of edits (insertions, deletions, or substitutions) that will change one codeword into another. For the BSID channel an upper bound for the probability of decoding a codeword in error was given in [15], assuming codeword boundaries are known. For Pi , Pd , Ps  1, the bound of [15, (9)] is dominated by the number of correctable 1 We refer to 32-bit floating point as single precision, and 64-bit floating point as double precision. errors, t. Now, for a code with jminimumk Levenshtein distance d −1 [2]. Hence designing dlmin , it can be shown that t = lmin 2 TVB codes with constituent encodings having large dlmin will result in the greatest improvement to the code’s performance at low channel error rates. However, in general the codeword boundaries are not known and need to be estimated by the decoder. Therefore decoding a given codeword on synchronization error channels depends not only on the current received word, but also on previous and subsequent ones. This means that the performance of a TVB code depends not only on the distance properties of constituent encodings considered separately, but also on the relationship between constituent encodings. This effect becomes more significant under poorer channel conditions, where the drift can easily exceed the length of a codeword. Unfortunately, the required relationship between constituent encodings for optimal performance over the BSID channel is still an open problem. What is known is that the diversity created by a sequence of different encodings helps the decoder estimate the drift within a codeword length, improving performance at higher channel error rates [15]. B. Representation of Previous Schemes as TVB Codes TVB codes generalize a number of existing synchronization error-correcting codes. The flexibility of the generalization allows the creation of improved codes at the same size and rate, as we shall show. Consider first the sparse inner codes with a distributed marker sequence2 of the Davey-MacKay construction [4]. It is clear that the sparse code is a fixed encoding C 0 : Fq ,→ Fn2 ; these codewords are then added to a distributed marker sequence wi of length n, specific for each codeword index i. Thus we can write Ci (Di ) = C 0 (Di ) + wi to represent the inner codes of [4] as TVB codes. The equivalence of this mapping to the inner code of [4] has also been shown in [16]. The distributed marker serves the same function as the use of different encodings in TVB codes. The decoder of [4] tracks the marker sequence directly, treating the additive encoded message sequence as substitution errors. Therefore, to corrupt the marker sequence as little as possible, the inner code used is sparse. The sparseness results in a low dlmin , making it harder for the decoder to distinguish between the various codewords, and leads to relatively poor performance at low channel error rates. The codes of [15] can similarly be represented as TVB codes, with C 0 corresponding to the Synchronization and Error Correcting (SEC) code and wi corresponding to the Allowed Modification Vectors (AMVs). SEC codes are designed with a large dlmin for good performance at low channel error rates. For such channels this code can perform much better than the sparse code of [4]. AMVs are chosen such that when added to the SEC code the resulting code’s dlmin does not change. Clearly, the AMVs serve the same function as the use of different encodings in TVB codes. In contrast to a random distributed marker sequence, the use of AMVs does 2 This was originally referred to as a watermark sequence. 4 Table I A (7, 8, 4) TVB CODE C = (C0 , . . . , C3 ) WITH dlmin = 3. 100 C1 0000000 0000111 0011110 0110101 1001001 1100110 1111000 1111111 C2 0000011 0001100 0011111 0101010 1011001 1100000 1100111 1111110 C3 0000000 0001111 0101001 0110110 1000011 1001100 1110000 1111111 10−1 Symbol Error Rate C0 0000000 0000111 0011001 0110110 1001010 1100001 1111000 1111111 Comparison of Inner Codes with (n,q) =(7,8) and N =666 (7,8) sparse + random marker 3 data bits + 4 marker bits (7,8,3) SEC code with 8 AMVs (7,8,4) TVB code 10−2 10−3 not compromise the performance of the underlying SEC code at low channel error rates. In general, however, the Levenshtein distance spectrum is altered. The separate constituent encodings in TVB codes give greater design freedom than SEC codes with AMVs and also allows the design of constituent encodings that maintain the required optimized Levenshtein distance spectrum. The marker codes given by Ratzer [5] can also be cast as TVB codes by letting each possible sequence of data bits (between markers) be represented by a q-ary symbol. For example, consider a marker code with 3 marker bits inserted after every 9 data bits, where the 3-bit marker is randomly chosen between the sequences 001 and 110. This can be represented as a (12, 512, 2) TVB code, where encoding C0 consists of all possible 9-bit sequences appended with 001, and C1 consists of all possible 9-bit sequences appended with 110. Like the sparse codes of [4], these marker codes suffer from a low dlmin , leading to relatively poor performance at low channel error rates. On the other hand, the fixed marker bits improve the determination of codeword boundaries, and the random use of different marker bits creates the necessary diversity to improve performance in poorer channel conditions. To illustrate the difference in performance between the various designs, consider encodings of size (n, q) = (7, 8) with N = 666 (same size as codes C and H in [4]). A (7, 8, 4) TVB code where each constituent code has the best possible Levenshtein distance spectrum with dlmin = 3, found through an exhaustive search, is given in Table I. In Fig. 1 we compare this TVB code with earlier constructions from the literature at the same size. Consider first the SEC code of the same size and dlmin from [15], used with 8 AMVs in a random sequence. As expected, the TVB code performs better due to its improved Levenshtein distance spectrum, even though both TVB and SEC codes have the same dlmin . The performance of the sparse code with random distributed marker from [4] is considerably worse, particularly at low channel error rates. Similarly a code with three data bits and four marker bits (randomly chosen between 0011/1100), similar to [5], also performs poorly at low channel error rates. IV. A PPROPRIATE L IMITS ON S TATE S PACE The equations in Section II-C assume that summations can be taken over the set of all possible states. For a channel such as the one considered, the state space is unbounded for positive drifts. A practical implementation will have to take sums over a finite subset of states. In [4] the state space was limited 10−4 −5 10 10−4 10−3 10−2 Channel Error Event Probability (Pd = Pi; Ps =0) 10−1 3 ×10−1 Figure 1. Comparison of inner code designs of size (n, q) = (7, 8) with N = 666: a sparse code with random distributed marker from [4], a marker code with 0011/1100 marker bits similar to [5], a SEC code with randomlysequenced AMVs from [15], and the TVB code of Table I. to a drift |St | ≤ xmax , where xmax was chosen to be ‘several times larger’ than the standard deviation of the synchronization drift over one block length, assuming this takes a Gaussian distribution. No recommendation was given for the value that should be used. Limiting the state space is by definition sub-optimal. However, we can arbitrarily lower the number of cases where the sub-optimal solution is worse than the optimal one, by ensuring that only the least likely states are omitted. The choice of summation limits also involves a trade-off with complexity, which has a polynomial relationship with the size of the state space (c.f. Section V). Therefore, an appropriate choice of summation limits will result in the smallest state space such that the probability of the drift being outside that range is as low as required. The first step to identify good summation limits is to derive an accurate probability distribution of the state space, avoiding the Gaussian approximation of [4]. A. Drift Probability Distribution The drift ST after transmission of T bits was stated in [4] (and shown in [20]) to be normally distributed with zero mean and a variance equal to T p/(1 − p) for the special case where p := Pi = Pd . This distribution is asymptotically valid as T → ∞. For cases where Pi 6= Pd or where T is not large enough, this distribution cannot be used. This is particularly relevant for determining the summation limits of (7) where the sequence length n is not large. An exact expression for the probability distribution of ST is given by ΦT (m) = Pr {ST = m}  j T   X T T +m+j−1 Pi Pd T m = Pt Pi , j m+j Pt j=j (8) 0 where j0 = max(−m, 0). Observe that for a drift of m bits, we need m insertion events more than we have deletion events. Over a sequence of T bits, for j deletion events, this means 5 m + j insertion events and T − j transmission events. The j probability of this is Pim+j Pdj PtT −j = PtT Pim [Pi Pd /Pt ] . We get (8) by adding all different combinations of these events, and summing over j, noting that we cannot have fewer than 0 events of any type. Specifically, the number of combinations
for j deletions in T transmitted bits is given by Tj . The number of
combinations for m + j insertions is given by T +m+j−1 , as the m + j insertion events create an additional m+j m + j opportunities for insertion. B. Avoiding Numerical Issues In a practical implementation, computing the drift probability (8) requires a few special considerations. Practical codes from the literature have codeword size n in the range 5–12 bits and number of codewords N up to 1000, for a frame length nN of about 4000–6000 bits. These codes are designed to operate under channel conditions Pi , Pd from 10−3 to above 10−1 . Evaluating (8) under these conditions, one encounters very large values for the two binomial coefficients and very small values for the power term. For example, consider evaluating (8) at m = 0 for T = 6000 and Pi = Pd = 10−3 . The two binomial coefficients have a range of up to 1.56 × 101804 (at j = 3000) and 8.34 × 103609 (at j = 6000) respectively. The power term has a range of down to 1.65 × 10−35995 (at j = 6000). This range is far beyond that representable even in double-precision floating point. A direct implementation of (8) will therefore result in numerical overflow and underflow (in computing the binomial coefficients and power term respectively) for typical frame sizes and channel conditions, even though the summation term itself is representable. The above numerical range problem can be avoided by combining the computation of all terms in the summation as follows. Observe that (8) can be rewritten as ΦT (m) = T X δj , (9) j=j0 j    T +m+j−1 Pi Pd T m T (10) where δj = Pt Pi Pt j m+j and j0 = max(−m, 0) as before. In this expression, note that the summation is empty if j0 > T , resulting in zero probability. Also, since j ≥ 0, the first binomial coefficient is always non-zero, while the second binomial coefficient is non-zero if T > 0. Expanding the binomial coefficients using the factorial formula, we can express the summation term recursively as δj = δj−1 · Pi P d T + m + j − 1 T − j + 1 · · , Pt m+j j (11) allowing successive factors to be determined easily from previous ones. The initial factor required is the one at j0 , and can be determined from (10) by expanding the binomial coefficients using the multiplicative formula:  j j0 m+j0 Y T − j0 − i Y T − 1 − i Pi Pd 0 δj0 = PtT Pim . i i Pt i=1 i=1 (12) Consider the earlier example, now evaluating (9) at m = 0 for T = 6000 and Pi = Pd = 10−3 . In this case, the initial value δj0 = 6.07 × 10−6 , and the multiplier δj /δj−1 in the recursive expression (11) has its smallest value of 3.34×10−10 at j = 6000. Both values are easily representable as floating point numbers. Using (9), numerical range issues remain when computing δj0 for larger values of Pi , Pd , and consequently also for storing successive values of δj . For example, consider evaluating (9) at m = 0 for T = 6000 and Pi = Pd = 10−1 . In this case δj0 = 3.47 × 10−582 , and one needs to accumulate a number of δj values in this range to obtain the required result Φ6000 (0) = 0.0109. Again, the intermediate values are beyond the range of double-precision floating point numbers although the final result is representable. These numerical range issues can be avoided by computing (11) and (12) using logarithms. For the earlier example with m = 0, T = 6000, and Pi = Pd = 10−1 , we now get log δj0 = −1.34 × 103 and the smallest value of log δj is −1.93 × 104 at j = 6000. Finally, the required drift probability is obtained by accumulating the exponential of the log δj values using (9). However, the individual values of δj are still beyond the range of double-precision floating point. In practice, we have found that the use of extended-precision (80-bit) floating point provides sufficient range. Alternatively, the accumulation in (9) may be computed in logarithmic domain using the property
log(A + B) = log A + log 1 + elog B−log A . Note that expression (8) is valid for any Pi ≥ 0, Pd ≥ 0, and Pi + Pd < 1. However, the computation using logarithms cannot be applied directly when either or both of Pi and Pd are zero. These degenerate cases have to be handled as special cases, by first reducing (8) and then implementing the simplified equations using logarithms. C. Probability of Drift Outside Range + We want to choose lower and upper limits m− T , mT such that the drift after transmitting a sequence of T bits is outside + the range {m− T . . . mT } with an arbitrarily low probability Pr :   + Pr ST < m− < Pr , T + Pr ST > mT (13) m+ T or equivalently: 1 − X ΦT (m) < Pr . (14) m=m− T An appropriate choice of limits can be obtained iteratively as follows. Observe that for the BSID channel ΦT (m) is monotonically decreasing with increasing |m|. A first estimate for the limits is given by: −(1) mT +(1) and mT Pr 2 Pr = min m ΦT (m + 1) < , 2 = max m ΦT (m − 1) < (15) (16) where the number in superscript parentheses indicates the iteration count. If these estimates satisfy (14), we use them 6 103 That is, we extend the range by one in the direction of greatest gain. The iterative process is repeated until (14) is satisfied. − The size of the state space is given by MT = m+ T − mT + 1. E. Example Overestimating the required state space increases computational complexity, while underestimating the state space often results in poor decoding performance. Accurate limits are particularly important for restricting the drifts considered across each codeword. It is therefore useful to illustrate the discrepancy between the approximate distribution of [4] and the exact expression for the distribution of the drift. Consider a system with typical block and codeword sizes N = 500, n = Mτ (appr.) Mτ (exact) Mn (appr.) Mn (exact) M1 (appr.) M1 (exact) 102 101 100 −4 10 D. Choice of Summation Limits 10−3 10−2 10−1 Channel Error Event Probability (Pd = Pi; Ps =0) 100 (a) Probability of Event Outside Chosen Limits When considering the whole frame, T = τ , so that the overall size of the state space is given by Mτ . Now the final output of the MAP decoder is calculated using (1), which sums 0 + over all Mτ prior states m− τ ≤ m ≤ mτ . For each prior state, however, only the drifts introduced by the transmission of n bits need to be considered, corresponding to a subset Mn of + states m− n ≤ m ≤ mn . Similarly, the computation of (4) + and (5) is required for all Mτ states m− τ ≤ m ≤ mτ , each involving a summation over Mn prior or posterior states m− n ≤ respectively. Finally, the state transition metric is m0 ≤ m+ n obtained using the forward pass of (7); this is computed over + a sequence of n bits for each of Mn states m− n ≤ m ≤ mn . Each recursion consists of a summation over prior states m0 ; in this case only the drifts introduced by the transmission of one bit need to be considered, corresponding to a subset M1 + 0 of prior states m− 1 ≤ m ≤ m1 . Now consider that we want to limit the probability of any of these summations not covering an actual channel event over a whole frame to, say, no more than Pe . When computing the limits over the whole frame, m± τ , we simply need to set Pr = Pe . However, when computing limits over an n-bit sequence, m± repeated for each of N such n , since this summation is √ sequences, we set Pr = 1 − N 1 − Pe ≈ PNe for small Pe . Similarly, for limits over a 1-bit sequence, m± 1 , we use Pr = √ τ 1 − 1 − Pe ≈ Pτe for small Pe . Except in the case of stream decoding (c.f. Section VII), the state space limits only need to be determined once and remain valid as long as the channel conditions do not change. In any case, the required values of ΦT (m) depend only on the code parameters and channel conditions, so that a table may be pre-computed. This makes the average complexity of determining the state space limits negligible. N =500, n =10, Pe =10−10 104 Number of States as our lower and upper limits. Otherwise these estimates are updated iteratively as follows: ( −(i) +(i) −(i) mT − 1 if ΦT (mT + 1) ≤ ΦT (mT − 1) −(i+1) mT = −(i) mT otherwise, (17) ( +(i) +(i) −(i) mT + 1 if ΦT (mT + 1) > ΦT (mT − 1) +(i+1) mT = +(i) mT otherwise. (18) 100 10−1 10−2 10−3 10−4 10−5 10−6 10−7 10−8 10−9 10−10 10−11 10−12 10−13 −4 10 N =500, n =10, Pe =10−10 Mτ (appr.) Mτ (exact) Mn (appr.) Mn (exact) M1 (appr.) M1 (exact) 10−3 10−2 10−1 Channel Error Event Probability (Pd = Pi; Ps =0) 100 (b) Figure 2. (a) Number of states within summation limits, and (b) probability of encountering a channel event outside the chosen summation limits over a single frame, using the approximation of [4] and our exact computation. 10. We plot in Fig. 2a the number of states within summation limits using the approximate and exact expressions, in each case for Pe = 10−10 . For T = 1, [4, Section VII.A] assumes a maximum of two successive insertions; this is equivalent to setting m+ 1 = 2, so that M1 = 4. It is immediately apparent that while the approximation is very close for large T and high Pi = Pd , it quickly starts to underestimate the required range at lower channel error rates. As expected, the discrepancy is particularly large when considering shorter sequences. For T = 1 it is not surprising that there is a large discrepancy for channels with high error rate. Next, we determine the probability of encountering a channel event outside the chosen limits over a single frame, shown in Fig. 2b for the same limits used in Fig. 2a. For the exact distribution, this probability is always below the chosen threshold Pe = 10−10 , as expected. For the approximation, however, the probability of exceeding the chosen limits is higher than the threshold throughout the range considered. 7 At lower channel error rates, the discrepancy is significant (several orders of magnitude) even for large T . For small T , the probability of exceeding the chosen limits is high enough to make the approximation useless. For T = 1, the artificial limit of two successive insertions of [4] means that channels with high error rate will exceed this limit with high probability. V. A LGORITHM C OMPLEXITY A. Complexity of the MAP Decoder As a first step towards determining the overall complexity of the decoder, consider first the calculation of the state transition metric in (6). This is recursively computed using the forward pass of (7) over a sequence of n bits, for each of Mn states m. Each recursion consists of a summation over M1 prior states m0 as argued in Section IV-D. The bit-level probability Q can be obtained by a look-up table. Thus the complexity for calculating a single state transition metric is Θ(nMn M1 ). The final output of the algorithm consists of q probabilities for each of N symbols, calculated using (1). This equation sums over all Mτ prior states m0 and Mn states m, defining the domain for σi (m0 , m, D). It follows from (3) that the domain for γi (m0 , m, D) is the same. Now the computation of (3) is dominated by the evaluation of γi (m0 , m, D) in (6), whose complexity is Θ(nMn M1 ) as shown. Considering the number of times the γ metric is computed, the MAP decoder has an asymptotic complexity of Θ(N nqMτ Mn2 M1 ). B. Complexity of the Davey-MacKay Decoder It would initially appear that the MAP decoder complexity is significantly higher than that of the Davey-MacKay decoder, given as O(N nMτ M1 ) in [4] for a direct implementation (using our notation). However, the expression of the DaveyMacKay decoder seems to consider only the complexity of the initial forward and backward passes, ignoring the additional small forward passes needed to compute the final decoder output. The final output of the Davey-MacKay algorithm also consists of q probabilities for each of N symbols. Each of these is computed using [4, (4)], which sums over all possible prior and posterior states. In a direct implementation all possible prior states need to be considered; using our notation the number of states is Mτ . While not stated in [4], the number of posterior states that need to be considered is Mn , as argued for the MAP decoder. The computation within the summation of [4, (4)] is dominated by the conditional probability, which is computed using a separate forward pass. This forward pass is effectively identical to (7) whose complexity is Θ(nMn M1 ). Considering the number of times the forward pass is computed, it follows that the Davey-MacKay decoder has an overall asymptotic complexity of Θ(N nqMτ Mn2 M1 ). C. Comments on Algorithm Complexity Comparing the complexity expressions for the MAP decoder for TVB codes and the Davey-MacKay decoder for sparse codes with a distributed marker sequence, it follows that the asymptotic complexity for both decoders is the same. This is consistent with experimental running times for both decoders in [13]. In the complexity expression note that N , n, and q depend only on the code parameters while Mτ , Mn , and M1 also depend on the channel conditions. For M1 , it was argued in [4, Section VII.A] that it is sufficient to consider a maximum of two successive insertions, at a minimal cost to decoding performance. This is equivalent to setting m+ 1 = 2, so that M1 = 4; these limits were also used in [13], [14]. However, this artificially low limit is insufficient for more advanced code constructions, as shown in [15]. It was also argued in [4] that useful speedups can be obtained by only following paths through the trellis that pass through nodes with probabilities above a certain threshold. However, the choice of this threshold was not analyzed. It is also likely that this choice would depend on the properties of the inner code being used. VI. S PEEDING T HINGS U P A. Batch Computation of Receiver Metric In a naïve implementation, each γ computation (6) requires the computation of the receiver metric as a separate forward pass using (7). However it can be observed that for a given starting state m0 and symbol D, the γ metric will be computed for each end state m within the limits considered (c.f. equations (1), (3), (4), and (5), where the γ computations are used). In turn, this means that for a given x, the receiver metric will need to be determined for all subsequences ẏ within the drift limit considered. It is therefore sufficient to compute the forward pass (7) once, with the longest subsequence ẏ required. In doing so, the values of the receiver metric for shorter subsequences are obtained for free. We call this approach batch computation. This effectively reduces the complexity of computing the collection of γ metrics by a factor of Mn . The asymptotic complexity of the MAP decoder is therefore reduced to Θ(N nqMτ Mn M1 ). B. Lattice Implementation of Receiver Metric To compute the receiver metric, an alternative to the trellis of (7) is to define a recursion over a lattice as in [17]. For the computation of R(ẏ|x), the required lattice has n + 1 rows and µ̇ + 1 columns. Each horizontal path represents an insertion with probability 21 Pi , each vertical path is a deletion with probability Pd , while each diagonal path is a transmission with probability Pt Ps if the corresponding elements from x and ẏ are different or Pt P̄s if they are the same. Let Fi,j represent the lattice node in row i, column j. Then the lattice computation in the general case is defined by the recursion Fi,j = 1 Pi Fi,j−1 + Pd Fi−1,j + Q̇(ẏj |xi )Fi−1,j−1 , 2 (19) which is valid for i < n, and where Q̇(y|x) can be directly computed from y, x and the channel parameters: ( Pt Ps if y 6= x Q̇(y|x) = (20) Pt P̄s if y = x. 8 Table II C OMPLEXITY EXPRESSIONS FOR THE MAP DECODER FOR VARIOUS COMPUTATION MODES OF THE RECEIVER METRIC . Initial conditions are given by ( 1 if i = 0, j = 0 Fi,j = 0 if i < 0 or j < 0. (21) The last row is computed differently as the channel model does not allow the last event to be an insertion. In this case, when i = n, the lattice computation is defined by Fn,j = Pd Fn−1,j + Q̇(ẏj |xn )Fn−1,j−1 . C. Optimizing the Lattice Implementation In the lattice implementation of the receiver metric, it can be readily seen that the horizontal distance of a lattice node from the main diagonal is equivalent to the channel drift for the corresponding transmitted bit. It should therefore be clear that the likelihood of a path passing through a lattice node decreases as the distance to the main diagonal increases. We can take advantage of the above observation by limiting the lattice computation to paths within a fixed corridor around the main diagonal. Specifically, the arguments of Section IV can be applied directly, resulting in a corridor of width Mn in general for the transmitted sequence of n bits considered. Exceptions to this width occur in the first few rows with index + i < −m− n and the last few rows with index i > µ̇ − mn , where part of the corridor falls outside the lattice rectangle. The number of nodes within this corridor is given by where κUL = κLR = − ∆(−m− n ) − ∆(−mn − n), + ∆(n + m+ n − µ̇) − ∆(mn − ( and ∆(k) = k2 +k 2 0 if k > 0 otherwise. (23) (24) µ̇), Algorithm Original Batch computation Lattice receiver D Corridor constraint Complexity Θ(N nqMτ Mn2 M1 ) Θ(N nqMτ Mn M1 ) + Θ(N  nqMτ h[n + mn ]) h ii − 2 Θ N qMτ nMn − 21 (m− n ) − mn (22) Finally, the required receiver metric is obtained from this computation as R(ẏ|x) = Fn,µ̇ . The calculation of a single run through the lattice requires a number of computations proportional to the number of nodes in the lattice. Now for the transmitted sequence of n bits considered, the number of rows will always be n while the number of columns is at most n + m+ n . The complexity of a direct implementation of this algorithm is therefore Θ(n[n + m+ n ]). It has been argued in Section VI-A that for a given x, the receiver metric R(ẏ|x) needs to be determined for all subsequences ẏ within the drift limit considered. Observe that the same argument applies equally when the receiver metric is computed using the lattice implementation (19). Therefore, when the lattice implementation is used in batch mode, the MAP decoder has an asymptotic complexity of Θ(N nqMτ [n + m+ n ]). κ = nMn − κUL − κLR , A B C (25) (26) The complexity of the corridor-limited lattice algorithm is therefore Θ(nMn − κUL − κLR ). Some simplification of this expression is possible when the corridor-limited lattice algorithm is used in the MAP decoder with batch computation for the channel considered. When batch computation is used, µ̇ = n + m+ n by definition, so that κLR = 0. Furthermore, for the BSID channel,  − 2  1 − −n ≤ m− ≤ 0, so that κ = (m ) − m . Under these UL n n n 2 conditions,the MAP decoder has an asymptotic complexity of  
2 − Θ N qMτ nMn − 12 (m− ) − m . n n D. Comparing Complexity A summary of the complexity expressions for the MAP decoder for various computation modes of the receiver metric is given in Table II. Comparing the expressions in rows A and B of Table II we can immediately see that the batch computation of the receiver metric reduces complexity by a factor equal to Mn . Unfortunately, the remaining complexity expressions contain terms that depend on the code parameters and channel conditions in a rather opaque way, making it harder to understand the benefits of these improvements. In the first instance, we can simplify the expressions further to facilitate comparison. Consider the expression in row C of Table II, when the lattice implementation is used. It can be shown that n + m+ n → Mn − 1 as channel conditions get worse; we can therefore simplify the complexity expression to O(N nqMτ Mn ). Comparing this to the expression in row B of Table II we can see that the use of the lattice implementation reduces complexity by a factor of at least M1 . Finally, consider the expression in row D of Table II, when the corridor − constraint  −is 2applied  to the lattice algorithm. Since mn ≤ 0, 1 − the 2 (mn ) − mn term is strictly positive. The reduction in complexity offered by the corridor constraint is therefore equal 2nMn to 2nM −(m − , and becomes significant as channel con− 2 n n ) +mn ditions improve. As channel conditions get worse, m− n → −n, so that the expression is dominated by the nMn term. Under these conditions, the complexity of the corridor-constrained lattice implementation becomes approximately equal to that of the unconstrained lattice implementation. We can also illustrate the effect of the proposed speedups by considering a rate- 21 TVB code with typical block and codeword sizes N = 500, n = 10, and q = 32. We compute the MAP decoder complexity for this code under a range of channel conditions, using the original algorithm of Section II-C and the improvements described above. We plot these in Fig. 3 using the same summation limits as in Section IV-E. Note that for a fairer comparison between the lattice and trellis modes, we include a constant factor of three in the lattice computation. This follows the observation that each lattice node computation (19) requires three multiplications while each trellis computation (7) requires only one. A few general observations can be made on this graph: a) The batch computation of the receiver metric results in a considerable reduction of complexity throughout, but is even more significant under poor channel conditions. b) The lattice implementation is considerably less complex than the 9 N =500, n =10, q =32 Complexity (Number of Arithmetic Operations) 1014 1013 1012 Trellis Trellis (batch) Lattice (batch) Lattice corridor (batch) 1011 1010 109 108 107 −4 10 10−3 10−2 10−1 Channel Error Event Probability (Pd = Pi; Ps =0) 100 Figure 3. MAP decoder complexity (in number of arithmetic operations) under a range of channel conditions, for various computation modes of the receiver metric. trellis implementation at high channel error rates. c) The lattice corridor constraint extends this improvement to the low channel error rate range. In conclusion, the proposed speedups result in a considerable reduction in complexity of almost two orders of magnitude for typical code sizes and channel conditions. We have observed a similar trend under a range of typical code sizes, so this result can be taken as representative. Note that in general this distribution βN (m) has a wider spread than Φτ (m). As discussed in Sections IV-C and IV-D, the choice of state space limits depends on the expected distribution of drift. For limits involving the whole frame, the distribution used is Φτ (m), which assumes that the initial drift is zero. The assumption does not hold under stream decoding conditions, where the initial drift is not known a priori, although its distribution can be estimated. The uncertainty in locating the startof-frame position increases the uncertainty in locating the endof-frame position, resulting in a wider prior distribution for the end-state boundary condition βN (m). Therefore, any limits on state space determined using Φτ (m) will be underestimated. The severity of this error depends on the difference between βN (m) and Φτ (m), which increases as channel conditions get worse. For stream decoding, therefore, it is sensible to recompute the state space limit Mτ at the onset of decoding a given frame, using βN (m) in lieu of Φτ (m). Doing so avoids underestimating the required state space, and implies that for stream decoding, the state space size will change depending on how well-determined the frame boundaries are. After decoding the current frame, we obtain the posterior probability distribution for the drift at end-of-frame, given by: n o n o τ +m+ τ +m+ Pr Sτ = m Ym− τ = λN (m)/ Pr Ym− τ τ τ λN (m) . =P 0 m0 λN (m ) (28) The most likely drift at end-of-frame can be found by: VII. S TREAM D ECODING We have so far considered the case where frame boundaries are known exactly. While there are practical cases involving single-frame transmission where this is true, exact frame boundaries are often unknown. The MAP decoder can handle such cases by changing the initial conditions for (4) and (5) and choosing appropriate state space limits. This obviates the need for explicit frame-synchronization markers as used in conventional communication systems, and can therefore reduce this overhead. The approach presented here is in principle similar to that used in [4] for ‘sliding window’ decoding. However, there are some critical differences which we explore further in Section VII-C. A. Choosing End-of-Frame Priors Consider first the common case where a sequence of frames is transmitted in a stream. The usual practice in communication systems is for the receiver to decode one frame at a time, starting the decoding process as soon as all the data related to the current frame is obtained from the channel. In this case, τ +m+ the current received frame is considered to be Ym− τ , which τ may include some bits from the end of the previous frame and start of the next frame. The end-state boundary condition for (5) can be obtained by convolving the expected end-of-frame drift probability distribution with the start-state distribution: X βN (m) = α0 (m0 )Φτ (m − m0 ). (27) m0 λN (m) = arg max λN (m). 0 m m0 λN (m ) Ŝτ = arg max P m (29) As in [4], we determine the nominal start position of the next frame by shifting the received stream by τ + Ŝτ positions. The initial condition for the forward metric for the next frame, α̂0 (m), is set to: λN (m + Ŝτ ) α̂0 (m) = P , 0 m0 λN (m ) (30) replacing the initial condition for (4) reflecting a known frame boundary. B. Stream Look-Ahead Taking advantage of the different constituent encodings in TVB codes, the MAP decoder can make use of information from the following frame to improve the determination of the end-of-frame position. We augment the current block of N symbols with the first ν symbols from the following block (or blocks, when ν > N ), for an augmented block size N 0 = N +ν. The MAP decoder is applied to the corresponding augmented frame. After decoding, only the posteriors for the initial N symbols are kept; the start of the next frame is determined from the drift posteriors at the end of the first N symbols, and the process is repeated. Consider the latency of the MAP decoder to be the time from when the first bit of a frame enters the channel to when the decoded frame is available. The cost of look-ahead is an increase in decoding complexity and latency corresponding 10 to the change in block size from N to N 0 . The effect on complexity is seen by using terms corresponding to the augmented block size in the expressions of Table II. The latency is equal to the time it takes to receive the complete frame and decode it. Look-ahead increases the time to receive the augmented frame linearly with ν and the decoding time according to the increase in complexity. The required look-ahead ν depends on the channel conditions and the code construction. In general, a larger value is required as the channel error rate increases. We show how an appropriate value for ν can be chosen for a given code under specific channel conditions in Section VIII-A. Typical values for ν are small (ν < 10) for good to moderate channels (Pi , Pd < 10−2 ). The required look-ahead increases significantly for poor channels: the example in Section VIII-A requires ν = 1000 at Pi = Pd = 2 × 10−1 . C. Comparison with Davey-MacKay Decoder A key feature of the Davey-MacKay construction is the presence of a known distributed marker sequence that is independent of the encoded message. This allows the decoder, in principle, to compute the forward and backward passes over the complete stream. However, to reduce decoding delay, the decoder of [4] performs frame-by-frame decoding using a ‘sliding window’ mechanism. The ‘sliding window’ mechanism seems intended to approximate the computation of the forward and backward passes over all received data at once. This approach is similar in principle to ours when stream lookahead is used; however, there are some critical differences which we discuss below. In [4], the starting index for a given frame is taken to be the most likely end position of the previous frame, as determined by the Markov model posteriors. This is the same as the approach we use in Section VII-A. However, in [4], the initial conditions of the forward pass are simply copied from the final values of the forward pass for the previous frame. This is consistent with the view that the ‘sliding window’ mechanism approximates the computation over all received data at once, but contrasts with our method. In Section VII-A the initial conditions of the forward pass are determined from the posterior probabilities of the drift at the end of the previous frame. These drift posteriors include information from the look-ahead region and from the priors at the end of the augmented frame, which were determined analytically from the channel parameters. Observe that in the ‘sliding window’ mechanism of [4], the backward pass values cannot be computed exactly as for the complete stream. Instead, the decoder of [4] computes the forward pass for some distance beyond the expected end of frame position, and initializes the backward pass from that point. The suggested distance by which to exceed the expected end of frame position is ‘several (e.g. five) multiples of xmax ’, where xmax is the largest drift considered. The concept is the same as the stream look-ahead of Section VII-B. However, we recommend choosing the look-ahead quantity ν based on empirical evidence (c.f. Section VIII-A). It is claimed in [4] that the backward pass is initialized from the final forward pass values; the reasoning behind this is unclear, and does not seem to have a theoretical justification. We initialize the backward pass with the prior probabilities for the drift at the end of frame, as explained in Section VII-A. D. Initial Synchronization The only remaining problem is to determine start-of-frame synchronization at the onset of decoding a stream. This can be obtained by choosing state space limits Mτ large enough to encompass the initial desynchronization and by setting equiprobable initial conditions: α0 (m) = βN (m) = 1 ∀m. Previous experimental results [4] have assumed a Mτ known start for the first frame, with the decoder responsible for maintaining synchronization from that point onwards. We adopt the same strategy in the following. VIII. R ESULTS Practical results are given in this section. We show how an appropriate choice of decoder parameters allows stream decoding to perform as well as when frame boundaries are known. Results are also given for existing constructions which can be expressed as TVB codes, showing how the symbol-level MAP decoder improves on the original decoder (in the case of [4]) or is equivalent (in the case of [5]). We also demonstrate some improved constructions allowed by the flexibility of TVB codes. These are achieved by using simulated annealing to find TVB codes of a required order with a good Levenshtein distance spectrum. Specifically, we seek to find constituent codes with the highest possible minimum Levenshtein distance and the lowest multiplicity at small distances. For all codes so designed, M < N ; we construct our TVB codes using a random sampling with replacement of the unique constituent codes, and use this as our inner code. Construction parameters for all codes used in this section are given in Table III. To facilitate reproduction of these results, the TVB codebooks used are available for download from the first author’s web site3 . A. Stream Decoding with MAP Decoder The results of [13] assumed known frame boundaries; under these conditions, the decoder is arguably at an advantage when comparing with the results of [4], [5]. It is also not clear whether the MAP decoder can keep track of frame boundaries with the non-sparse constructions of [13], [15] and the TVB codes introduced here, especially in the absence of a known marker sequence. In the following we investigate the performance of the MAP decoder under stream decoding conditions, and consider the choice of look-ahead required. As in [4] we assume that the start of the first frame is known, while the decoder is responsible for keeping synchronization from that point onwards. We use the limits specified in Section IV. We start by investigating the effect of stream decoding on the ability of the MAP decoder to track codeword boundaries. We consider a (6, 8, 12) TVB code, which is the inner code for the concatenated system N2b of Table III. We simulate 3 Available at http://jabriffa.wordpress.com/publications/data-sets/. 11 Table III C ONSTRUCTION PARAMETERS OF CODES USED IN SIMULATIONS . Labela P1 N1a N1b P2 N2a N2b P3 N3 P4 N4 P5 N5 Inner code (5, 16) sparse (5, 16) sparse (10, 256, 3) TVB (6, 8) sparse (6, 8) sparse (6, 8, 12) TVB 9 bits, uncoded 9 bits, uncoded not applicable (7, 8, 8) TVB not applicable (7, 4, 32) TVB Marker random, distributed random, distributed none random, distributed random, distributed none 001/110, appended 001/110, appended not applicable none not applicable none Outer code LDPC (999, 888) F16 LDPC (999, 888) F16 LDPC (499, 444) F256 LDPC (1000, 100) F8 LDPC (1000, 100) F8 LDPC (1000, 100) F8 LDPC (3001, 2000) F2 LDPC (2997, 1998) F2 b 3 turbo code F2 Rate 14 LDPC (666, 333) F8 1 turbo code F2 Rate 10 LDPC (855, 300) F4 Comment Published in [4, Fig. 8, Code D] Identical construction to P1, symbol-level MAP decoder Same overall rate and block size as P1 Published in [4, Fig. 8, Code I] Identical construction to P2, symbol-level MAP decoder Same overall rate, block size, and outer code as P2 Published in [5, Fig. 7, Code D] Identical inner code to P3, marginally smaller outer code Published in [11, Fig. 4, Code T2] Same overall rate as P4 Published in [11, Fig. 4, Code T4] Same overall rate as P5 a Labels starting with P indicate previously published results, while labels starting with N indicate new simulation results. by truncating the LDPC (3001, 2000) F2 of P3. This truncation is necessary so that the outer-encoded sequence can be expressed by an integral number of inner codewords. b Obtained Fidelity at codeword boundaries, (p = Pd = Pi; Ps =0) 100 Frame Error Rate Fidelity 10−2 10−3 10−1 10−2 (6,8,12) TVB, p =0.22 10−4 0 500 1000 1500 Index for codeword following boundary 2000 Figure 4. The fraction of correctly resynchronized codeword boundaries (fidelity) as a function of codeword index, for code N2b of Table III. this inner code with a block size N = 2000 under the channel conditions at the onset of convergence for the concatenated system, that is at Pi = Pd = 0.22, assuming only the start position of the first frame is known. At each codeword boundary we plot the fraction of correctly determined drifts (fidelity) in Fig. 4. As expected, the fidelity drops at the end of the frame, where the actual drift is unknown to the decoder. However, it can be observed that the fidelity reaches a steady high value within about 1000 codewords from the end of frame. It could therefore be supposed that a look-ahead of ν = 1000 would be sufficient for this code under these channel conditions. The dip at the start of the frame is caused by the uncertainty in the frame start position, due to the very low fidelity at the end of the previous frame. To test this hypothesis, we concatenate this inner code with the (1000, 100) LDPC code over F8 of [4, Code I]. We simulate this system under the following conditions: 1) known frame start and end (frame decoding); 2) known start for the first frame, unknown frame ends (stream decoding), no Comparison of frame and stream decoding for code N2b 10−3 −1 10 Frame [1 iter] Frame [5 iter] Stream, ν =0 [1 iter] Stream, ν =0 [5 iter] Stream, ν =1000 [1 iter] Stream, ν =1000 [5 iter] Channel Error Event Probability (Pd = Pi; Ps =0) 100 Figure 5. Demonstration of the effect of look-ahead on the MAP decoder’s performance under stream decoding conditions, and comparison with frame decoding. look-ahead; 3) stream decoding with look-ahead ν = 1000 codewords. Results are shown in Fig. 5. We give results after the first and fifth iterations. As anticipated, performance under stream decoding conditions is poorer than frame decoding if there is no look-ahead. However, an appropriate look-ahead quantity allows the decoder to perform as well under stream decoding as under frame decoding. It is important to highlight that this result is dependent on the inner code structure, and that therefore the generalization to other constructions is not obvious. However, we have repeated the same test with other constructions, including those of [4], [5], [13]–[15] and the new constructions in this paper, and under different channel conditions. In all cases we have found that the result is repeatable, in that it is possible to approach the performance of frame decoding with stream decoding, as long as an appropriate look-ahead quantity is chosen. The only cost of stream decoding is therefore the need for a fidelity analysis to determine a suitable look-ahead value 12 100 Comparison with Davey & MacKay 100 10−1 Frame Error Rate Frame Error Rate 10−1 10−2 10−3 P1 N1a [1 iter] N1a [5 iter] N1b [1 iter] N1b [5 iter] 10−4 10−5 −3 3 ×10 Comparison with Ratzer and Mansour & Tewfik 10−2 P2 N2a [1 iter] N2a [5 iter] N2b [1 iter] N2b [5 iter] 10−1 Channel Error Event Probability (Pd = Pi; Ps =0) 3 ×10−1 Figure 6. Comparison with Davey-MacKay: improving the performance of [4, Code D] (left) and [4, Code I] (right) using our MAP decoder, iterative decoding, and an inner code with better Levenshtein distance spectrum. and the increased decoding latency and complexity caused by the augmented block size. Since the code performance is undiminished, to simplify our analysis from this point onwards we assume known frame start and end positions. B. Comparison with Prior Art We have already shown in [13] that the (symbol-level) MAP decoder allows us to obtain better performance from the codes of [4]. Further improvement can be obtained with iterative decoding, as we show here. Additionally, the flexibility of TVB codes allows us to obtain codes that perform better at the same size and/or rate. In the following, we simulate channel conditions Pi = Pd ; Ps = 0 in order to compare with published results. For low channel error rates, consider [4, Code D], listed as P1 in Table III. We compare the previously published result with a MAP decoding of the same code (N1a) in Fig. 6. As shown in [13], the MAP decoder improves the performance of this code even after the first iteration; additional iterations improve the result further. At the same overall code rate and block size we can improve the performance further by designing an inner TVB code with a better Levenshtein distance spectrum (N1b). We repeat the process at higher channel error rates for [4, Code I], listed as P2 in Table III. Again, compared to the published result, a MAP decoding of the same code (N2a) improves performance even after the first iteration, also in Fig. 6. Additional iterations improve the result, but the difference in this case is less pronounced. Replacing the inner code with one of the same size but a better Levenshtein distance spectrum (N2b) improves performance further. As we have already discussed in Section II-A, the marker codes given by Ratzer [5] can also be cast as TVB codes. In [5] binary outer LDPC codes were used. To use a binary outer code with our MAP decoder, the bitwise APPs can be obtained from the q-ary symbol APPs by marginalizing over the other bits [21, p. 326]; these are then passed to the decoder for the binary outer code. In this case, for a binary 10−2 P3 [serial] P3 [iterative] N3 [1 iter] N3 [20 iter] P4 [10 iter] N4 [5 iter] P5 [10 iter] N5 [20 iter] 10−3 10−4 10−5 −2 10 10−1 Channel Error Event Probability (Pd = Pi; Ps =0) 100 Figure 7. Comparison with Ratzer and with Mansour & Tewfik: demonstrating the equivalence of our MAP decoder on [5, Code D], with and without iterative decoding, and improving performance on [11, Codes T2, T4] at the same overall code rate, using inner codes with improved Levenshtein distance in concatenation with LDPC outer codes. outer code we expect the performance of the concatenated code to be identical, whether the inner code is decoded with the bit-level (MAP) decoder of [5] or with our symbol-level decoder. We show this in Fig. 7 for [5, Code D], listed as P3 in Table III, in comparison with an almost-identical code (N3) using our MAP decoder. It is important to highlight that decoding marker codes as TVB codes provides no material advantage; in fact, a cost is paid in complexity for doing so. We do not propose or expect that marker codes will be decoded as TVB codes. However, there is value in showing that marker codes can be decoded as TVB codes with no loss in performance. Specifically, this allows us to compare the structure of marker codes with other constructions, within the same context. Finally, in Fig. 7 we compare with the more recent turbo codes of [11, Codes T2, T4], respectively listed as P4 and P5 in Table III, with concatenated systems of the same overall rate, using a TVB inner code and an LDPC outer code. It can be seen that our concatenated systems outperform the codes of [11], significantly at lower channel error rates, and somewhat less so at higher channel error rates. IX. C ONCLUSIONS In this paper we have considered TVB codes, which generalize a number of previous codes for synchronization errors. We discussed the applicable design criteria for TVB codes, expressing some previously published codes as TVB codes and showing that the greater flexibility of TVB codes allows improved constructions. For example, our (7, 8, 4) TVB code achieves a SER of 10−4 at a Pi , Pd that is almost two orders of magnitude higher than a marker or distributed marker code of the same size, and slightly better than our earlier SEC codes. We also considered a number of important issues related to a practical implementation of the corresponding MAP decoder. Specifically, we have given an expression for the 13 expected distribution of drift between transmitter and receiver due to synchronization errors, with consideration for practical concerns when evaluating this expression. We have shown how to determine an appropriate choice for state space limits based on the drift probability distribution. The decoder complexity under given channel conditions is then expressed as a function of the state space limits used. For a given state space, we have also given a number of optimizations that reduce the algorithm complexity with no further loss of decoder performance. The proposed speedups, which are independent of the TVB code construction, result in a considerable reduction in complexity of almost two orders of magnitude for typical code sizes and channel conditions. For code constructions with appropriate mathematical structure we expect to be able to replace the receiver metric, which considers each possible transmitted codeword, with a faster soft-output algorithm. Next, we have considered the practical problem of stream decoding, where there is no prior knowledge of the received frame boundary positions. In doing so we have also shown how an appropriate choice of decoder parameters allows stream decoding to approach the performance when frame boundaries are known, at the expense of some increase in complexity. Finally, practical comparisons of TVB codes with earlier constructions were given, showing that TVB code designs can in fact achieve improved performance. Even compared to the state of the art codes of [11], the TVB codes presented here 1 achieve a FER of 10−3 at 24% higher Pi , Pd for a rate- 10 3 code, and at 84% higher Pi , Pd for a rate- 14 code. We expect further improvements to the codes shown here to be possible, particularly by co-designing optimized outer codes. However, a detailed treatment of the design process is beyond the scope of this paper, and will be the subject of further work. R EFERENCES [1] F. F. Sellers, “Bit loss and gain correction code,” IRE Trans. Inf. Theory, vol. IT-8, pp. 35–38, 1962. [2] V. I. Levenshtein, “Binary codes capable of correcting deletions, insertions and reversals,” Sov. Phys. Doklady, vol. 10, no. 8, pp. 707–710, 1966. [3] R. G. Gallager, “Sequential decoding for binary channels with noise and synchronization errors,” Massachusetts Inst. of Tech. Lexington Lincoln Lab, Tech. Rep. 2502, Oct. 27th, 1961. [4] M. C. Davey and D. J. C. MacKay, “Reliable communication over channels with insertions, deletions, and substitutions,” IEEE Trans. Inf. Theory, vol. 47, no. 2, pp. 687–698, 2001. [5] E. A. Ratzer, “Marker codes for channels with insertions and deletions,” Annals of Telecommunications, vol. 60, pp. 29–44, Jan. 2005. [6] F. Wang, D. Fertonani, and T. M. Duman, “Symbol-level synchronization and LDPC code design for insertion/deletion channels,” IEEE Trans. Commun., vol. 59, no. 5, pp. 1287–1297, May 2011. [7] T. Swart, H. Ferreira, and M. dos Santos, “Using parallel-interconnected Viterbi decoders to correct insertion/deletion errors,” in 7th AFRICON Conference in Africa, vol. 1, Sep. 2004, pp. 341–344. [8] M. Schlauweg, D. Profrock, and E. Muller, “Correction of insertions and deletions in selective watermarking,” in IEEE Intern. Conf. on Signal Image Technology and Internet Based Systems (SITIS), 2008, Nov. 30th– Dec. 3rd, 2008, pp. 277–284. [9] M. Mansour and A. Tewfik, “Convolutional decoding in the presence of synchronization errors,” IEEE J. Sel. Areas Commun., vol. 28, no. 2, pp. 218–227, Feb. 2010. [10] M. F. Mansour and A. H. Tewfik, “Convolutional decoding in the presence of synchronization errors,” IEEE J. Selected Areas in Commun., vol. 28, no. 2, pp. 218–227, Feb. 2010. [11] ——, “A turbo coding scheme for channels with synchronization errors,” IEEE Trans. Commun., vol. 60, no. 8, pp. 2091–2100, Aug. 2012. [12] H. Mercier, V. Bhargava, and V. Tarokh, “A survey of error-correcting codes for channels with symbol synchronization errors,” IEEE Commun. Surveys Tuts., vol. 12, no. 1, pp. 87–96, First Quarter 2010. [13] J. A. Briffa, H. G. Schaathun, and S. Wesemeyer, “An improved decoding algorithm for the Davey-MacKay construction,” in Proc. IEEE Intern. Conf. on Commun., Cape Town, South Africa, May 23–27, 2010. [14] J. A. Briffa and H. G. Schaathun, “Improvement of the Davey-MacKay construction,” in Proc. IEEE Intern. Symp. on Inform. Theory and its Applications, Auckland, New Zealand, Dec. 7–10, 2008, pp. 235–238. [15] V. Buttigieg and J. A. Briffa, “Codebook and marker sequence design for synchronization-correcting codes,” in Proc. IEEE Intern. Symp. Inform. Theory, St. Petersburg, Russia, Jul. 31–Aug. 5, 2011. [16] J. A. Briffa, “A GPU implementation of a MAP decoder for synchronization error correcting codes,” IEEE Commun. Lett., vol. 17, no. 5, pp. 996–999, May 27th 2013. [17] L. R. Bahl and F. Jelinek, “Decoding for channels with insertions, deletions, and substitutions with applications to speech recognition,” IEEE Trans. Inf. Theory, vol. 21, no. 4, pp. 404–411, Jul. 4, 1975. [18] L. C. Perez, J. Seghers, and D. J. Costello, Jr., “A distance spectrum interpretation of turbo codes,” IEEE Trans. Inf. Theory, vol. 42, no. 6, pp. 1698–1709, Nov. 1996. [19] G. Ferrari and K. M. Chugg, “Linear programming-based optimization of the distance spectrum of linear block codes,” IEEE Trans. Inf. Theory, vol. 49, no. 7, pp. 1794–1800, 2003. [20] M. C. Davey, “Error-correction using low-density parity-check codes,” Ph.D. dissertation, University of Cambridge, 1999. [21] D. J. C. MacKay, Information Theory, Inference, and Learning Algorithms. Cambridge University Press, 2003. 

7cs.IT

JOURNAL OF LATEX CLASS FILES, VOL. 11, NO. 4, DECEMBER 2012 1 A Dissipative Systems Theory for FDTD with Application to Stability Analysis and Subgridding arXiv:1606.08761v1 [cs.CE] 28 Jun 2016 Fadime Bekmambetova, Xinyue Zhang, and Piero Triverio, Senior Member, IEEE Abstract—This paper establishes a far-reaching connection between the Finite-Difference Time-Domain method (FDTD) and the theory of dissipative systems. The FDTD equations for a rectangular region are written as a dynamical system having the magnetic and electric fields on the boundary as inputs and outputs. Suitable expressions for the energy stored in the region and the energy absorbed from the boundaries are introduced, and used to show that the FDTD system is dissipative under a generalized Courant-Friedrichs-Lewy condition. Based on the concept of dissipation, a powerful theoretical framework to investigate the stability of FDTD methods is devised. The new method makes FDTD stability proofs simpler, more intuitive, and modular. Stability conditions can indeed be given on the individual components (e.g. boundary conditions, meshes, embedded models) instead of the whole coupled setup. As an example of application, we derive a new subgridding method with material traverse, arbitrary grid refinement, and guaranteed stability. The method is easy to implement and has a straightforward stability proof. Numerical results confirm its stability, low reflections, and ability to handle material traverse. Index Terms—finite-difference time-domain, stability, energy, dissipation, subgridding. I. I NTRODUCTION The Finite-Difference Time-Domain (FDTD) method is widely used to solve Maxwell’s equations numerically in microwave and antenna engineering, photonics and physics [1], [2]. FDTD is versatile, easy to implement, and has a low computational cost per time step. Through explicit update equations, FDTD recursively computes the electric and magnetic field in the region of interest without requiring the solution of linear systems. This update process is stable if time step ∆t satisfies the Courant-Friedrichs-Lewy (CFL) stability limit [2] 1 ∆t < q q , (1) 1 1 1 1 εµ ∆x2 + ∆y 2 + ∆z 2 where ∆x, ∆y, ∆z denote cell size, ε denotes permittivity and µ denotes permeability. While many breakthroughs have been achieved in FDTD since Yee’s original algorithm [3], the efficiency of FDTD for multiscale problems remains an open problem. The simultaneous presence of large and small geometrical features can dramatically reduce FDTD efficiency Manuscript received ...; revised ... This work was supported in part by the Natural Sciences and Engineering Research Council of Canada (Discovery grant program) and in part by the Canada Research Chairs program. F. Bekmambetova, X. Zhang and P. Triverio are with the Edward S. Rogers Sr. Department of Electrical and Computer Engineering, University of Toronto, Toronto, M5S 3G4 Canada (email: fadime.bekmambetova@mail.utoronto.ca, xinyuezhang.zhang@mail.utoronto.ca, piero.triverio@utoronto.ca). because of two factors. First, the mesh has to be refined, at least locally, to properly resolve small features, which increases the number of unknowns and the cost per iteration. Second, because of (1), a mesh refinement imposes a smaller time step, which further increases computational cost. For example, a 3X refinement of the entire FDTD mesh increases computational cost by 81 times. These issues are unfortunate since multiscale problems abound in practice. Numerous solutions have been proposed to mitigate this issue, including sophisticated boundary conditions to mimic open spaces [2], local grid refinement [4], [5], [6] (commonly known as subgridding), thin wire models [1], lumped elements [7], [8], and hybridizations with model order reduction [9], [10], [11], finite elements [12], integral equations [13], ray tracing [14] and implicit schemes like ADI-FDTD [15], [16]. From a system theory viewpoint, many of these methods consist of the interconnection of subsystems, such as models, algorithms, boundary conditions. For example, in [13], an FDTD model, used to describe an inhomogeneous scatterer, is coupled to the time-domain method of moments, used to model a thin-wire antenna. In this way, one can leverage the respective strengths of different algorithms. However, while the stability properties of the individual algorithms may be well understood, ensuring the stability of their combination can be a formidable task. Ensuring stability is not trivial even in the relatively simple case of FDTD subgridding, where one just couples a coarse and a fine FDTD grid through an interpolation rule to relate fields at the grid transition. The lack of a systematic approach to ensure the stability of advanced FDTD methods is a major issue, that limits the development of new methods and their adoption by industry, where guaranteed stability is mandatory. This issue is the main motivation for this paper, that proposes a new theoretical framework to investigate and ensure the stability of both simple and advanced FDTD methods. Several techniques are available to analyze the stability of FDTD-like methods. Von Neumann analysis [1] is the simplest, but is only applicable to uniform meshes and homogeneous materials. The iteration method [2] checks if the eigenvalues of the matrix that relates the current and next field solution are all below one in magnitude. In the energy method [8] instead, one writes an expression for the total energy stored in the simulation domain, and then checks if the algorithm satisfies a discrete equivalent of the principle of energy conservation. Since energy conservation prevents an unphysical growth of the solution, this implies stability [17]. The iteration and energy methods are general, but they can lead to long derivations, since they require the analysis of the whole JOURNAL OF LATEX CLASS FILES, VOL. 11, NO. 4, DECEMBER 2012 West boundary North boundary (Nx + 1, Ny + 1) ∆x Ey |1, 3 2 Hz |1, 3 2 (1, 1) ∆y Hz | 3 , 3 East boundary coupled scheme, which may consist of many subsystems such as FDTD meshes, different boundary conditions, reduced order models, and so on. For example, in a subgridding scenario, one must derive the iteration matrix or energy function of the whole scheme, taking simultaneously into account coarse and fine meshes, interpolation rules and boundary conditions. This issue makes stability analysis quite involved. Moreover, it does not yield stability conditions on the individual subsystems. Consequently, if a subsystem is changed (for example, a different boundary condition is introduced), the iteration matrix or energy function of the whole problem must be derived again. In this paper, we propose a new stability framework for FDTD. The framework is based on the theory of dissipative systems [18], and generalizes the energy method. First, starting from FDTD update equations, we develop a self-contained mathematical model for a region with arbitrary permittivity, permeability and conductivity. The model is in the form of a discrete time dynamical system. The magnetic field tangential to the boundary is taken as input, while the electric field tangential to the boundary is taken as output. Through this form, we reveal that an FDTD model can be interpreted as a dynamical system which is dissipative when ∆t satisfies a generalized CFL condition. If this condition is not met, the system can generate energy on its own, leading to an unstable simulation. We thus establish a connection between FDTD and the elegant theory of dissipative systems, which is a novel result. We believe that this connection will greatly benefit the FDTD community, since the theory of dissipative systems has been extremely successful in control theory for ensuring the stability of interconnected systems. This key result sets the basis for a powerful FDTD stability theory, where each part of a given FDTD setup (standard meshes, boundary conditions, interpolation schemes, ...) is interpreted as a subsystem, and is required to be dissipative. Since the connection of dissipative systems is dissipative [18], [17], this will ensure the stability of the FDTD algorithm resulting from the connection of the subsystems. The proposed theory has numerous advantages. It simplifies stability proofs, since conditions can be imposed on each subsystem individually, rather than on the whole coupled algorithm. Stability proofs are thus made modular and “reusable”: once a given FDTD model (e.g., an advanced boundary condition) has been deemed to be dissipative, it can be combined to any other dissipative FDTD subsystem with guaranteed stability. The proposed approach also naturally provides the CFL stability limit of the resulting scheme, which will be the most restrictive CFL limit of the individual subsystems. Finally, the theory is intuitive, since it is based on the concept of energy, familiar to most scientists. The proposed theory is presented in 2D, for the sake of clarity. An extension to 3D is feasible and is currently under development. As an example of application, the proposed theory is used to derive the subgridding algorithm which is stable by construction, and has several desirable features. Material traverse is supported for both dielectrics and highly conductive materials, which is a limitation of other stable subgridding methods [19]. The proposed method has low reflections, and avoids non-rectangular cells [6], finite element concepts [20] and Withney forms [21], [22]. Corners are natively supported 2 y z x 2 2 Hz | 3 ,1 Ex | 3 ,1 2 2 South boundary Figure 1: Graphical representation of the 2D region considered in Sec. II. The hanging variables introduced on the four boundaries are denoted in green. without any special treatment [5] nor L-shaped cells [6]. Ultimately, the proposed algorithm just consists of a compact FDTD-like update equation for the edges between coarse and fine mesh, and is thus easy to implement. This update equation is provided explicitly for an arbitrary integer refinement ratio r, while several previous works [5], [6] provide the update weights only for specific refinements (typically r = 2 or 3), leaving the derivation of other cases to the Reader. The paper is organized as follows. In Sec. II, we cast the FDTD update equations for a rectangular region into the form of a dynamical system with suitable inputs and outputs. In Sec. III, we show that FDTD equations can be interpreted as a dissipative system, and propose the new stability theory. The theory is applied to derive a stable subgridding algorithm in Sec. IV, followed by numerical results in Sec. V. II. D ISCRETE T IME DYNAMICAL M ODEL FOR A 2D FDTD R EGION The goal of this section is to cast the FDTD equations for a 2D region into the form of a discrete time dynamical model. The model shall be self-contained, involving only field samples from the nodes belonging to the region. This goal will be achieved by introducing suitable magnetic field samples at the boundaries. We consider the 2D rectangular region shown in Fig. 1, operating in a TE mode with components Ex , Ey , and Hz . The region is discretized with a uniform rectangular grid with Nx × Ny cells, of width ∆x and height ∆y. In addition to the field samples used by standard FDTD, we also sample the H field on the four boundaries of the region. These additional samples will be referred to as hanging variables [22], and will allow us to: 1) develop a self-contained model for the region, which does not involve field samples beyond its boundaries; 2) derive an expression for the energy absorbed from each boundary; 3) connect the FDTD grid to other subsystems while maintaining stability. To keep the notation compact, we collect all Ex and Ey samples into column vectors Enx and Eny , of size NEx = Nx (Ny + 1) and NEy = (Nx + 1)Ny , respectively. The Hz JOURNAL OF LATEX CLASS FILES, VOL. 11, NO. 4, DECEMBER 2012 3 samples at the internal nodes are collected into column vector n+ 1 Hz 2 of size NHz = Nx Ny . The hanging variables on the South boundary of the region are collected into the Nx × 1 vector h iT n+ 1 n+ 1 n+ 1 (2) HS 2 = Hz |1+ 12,1 . . . Hz |N +2 1 ,1 . x 2 A. State Equation for Each Node The dynamical model for the region consists of an update equation for each E and H sample, excluding the hanging variables. Those variables will indeed be eliminated once the region is connected to the surrounding subsystems or to some boundary conditions. 1) Internal Hz Nodes: For these nodes, we use a standard FDTD update [2] µ µ n+ 1 n− 1 Hz |i+ 12,j+ 1 = ∆x∆y Hz |i+ 12,j+ 1 2 2 2 2 ∆t ∆t − ∆xEx |ni+ 1 ,j + ∆xEx |ni+ 1 ,j+1 2 2 2 (3) 2 for i = 1, . . . , Nx and j = 1, . . . , Ny . In (3), µ denotes n+ 1 the average permittivity on the edge where Hz |i+ 12,j+ 1 is 2 2 sampled. Subscripts are omitted from µ in order to simplify the notation. Equations (3) can be written in matrix form as Dµ n+ 12 Dµ n− 12 Hz = DA Hz + Gy Dlx Enx − Gx Dly Eny , ∆t ∆t (4) where DA is a diagonal matrix containing the area of the primary cells, Dµ is a diagonal matrix containing the average permittivity on each edge of the secondary grid. Diagonal matrices DA Dly = ∆yINEy (5) contain the length of the x- and y-directed edges of the primary grid, respectively. With Im , we denote the m × m identity matrix. Matrix Gx is the discrete derivative operator along x, which can be written as Gx = INy ⊗ WNx , (6) where ⊗ is the Kronecker’s product [23] and WNx is the Nx × (Nx + 1) matrix [24]   −1 +1   −1 +1   Wn =  (7) . .. ..   . . −1 +1 Similarly, Gy = WNy ⊗ INx . 2 2 2 for i = 1, . . . , Nx and j = 2, . . . , Ny . In (9), εx and σx denote, respectively, the average permittivity and average conductivity on the corresponding edge. While the common factor ∆x could be eliminated from (9), it is kept as it will be useful later. For the Ex nodes on the South boundary, the standard FDTD update equation involves a Hz sample outside the region. In order to avoid this, we apply the finite difference approximation to the
half step of the
secondary grid between nodes i + 12 , 1 + 12 and i + 12 , 1 , making use of the hanging variables. In this way, we obtain an update equation that involves only field samples from the considered region σx  ∆y  εx + Ex |n+1 = ∆x i+ 21 ,1 2 ∆t 2  σx  ∆y εx − Ex |ni+ 1 ,1 ∆x 2 2 ∆t 2 n+ 12 n+ 1 + ∆xHz |i+ 1 ,1+ 1 − ∆xHz |i+ 12,1 , (10) 2 + ∆yEy |ni,j+ 1 − ∆yEy |ni+1,j+ 1 Dlx = ∆xINEx 2 2 Similarly, the hanging variables on the North, East and West n+ 1 n+ 1 n+ 1 sides are cast into column vectors HN 2 , HE 2 and HW 2 , respectively. ∆x∆y 2) Ex Nodes: For the Ex nodes that fall strictly inside the region, we use a standard FDTD update [2] ε ε σx  σx  x x ∆x∆y Ex |n+1 = ∆x∆y Ex |ni+ 1 ,j + − 1 i+ 2 ,j 2 ∆t 2 ∆t 2 n+ 21 n+ 21 + ∆xHz |i+ 1 ,j+ 1 − ∆xHz |i+ 1 ,j− 1 (9) (8) 2 2 for i = 1, . . . , Nx . In (10), εx and σx denote the material
properties
of the half cell between nodes i + 12 , 1 + 12 and i + 12 , 1 . The update equation for the Ex nodes on the North boundary is obtained similarly, and reads ∆y  εx σx  ∆x = + Ex |n+1 i+ 21 ,Ny +1 2 ∆t 2   ∆y εx σx ∆x − Ex |ni+ 1 ,Ny +1 2 2 ∆t 2 n+ 21 n+ 1 + ∆xHz |i+ 1 ,N +1 − ∆xHz |i+ 12,N + 1 , (11) 2 y 2 y 2 for i = 1, . . . , Nx . Relations (9), (10) and (11) can be compactly written as   D εx Dσx Dlx Dly0 + En+1 = x ∆t 2   Dσx Dεx n+ 1 − Enx − Dlx GTy Hz 2 Dlx Dly0 ∆t 2 " 1# 2   Hn+ S + Dlx BS Dlx BN , (12) n+ 1 HN 2 where • Dl0 is an NEx × NEx diagonal matrix containing the y length of the y-directed edges of the secondary grid, including the half-edges of length ∆y/2 that intersect the North and South boundaries; • Dεx and Dσx are diagonal matrices storing the permittivity and conductivity on each x-directed primary edge, respectively; • BS has all entries set to zero, except for a −1 at the intersection of each row associated with a South boundary edge and the column of the corresponding n+ 1 hanging variable in HS 2 ; JOURNAL OF LATEX CLASS FILES, VOL. 11, NO. 4, DECEMBER 2012 4 BN is defined similarly to BS , but has a +1 on each entry associated with an edge of the North boundary. 3) Ey Nodes: The update equations for the Ey nodes are derived with a similar procedure. A standard FDTD update equation is written for the Ey nodes that fall strictly inside the region. For the Ey nodes on the West and East boundaries, instead, finite differences are applied on a half grid step, similarly to (10) and (11). This process leads to   Dσy D εy + En+1 = Dly Dlx0 y ∆t 2   Dσy D εy n+ 1 Dly Dlx0 − Eny + Dly GTx Hz 2 ∆t 2 # " n+ 21   HW , (13) + Dly BW Dly BE n+ 1 HE 2 • where • Dl0 is an NEy × NEy diagonal matrix containing the x length of the x-directed edges of the secondary grid, including the half-edges of length ∆x/2 that intersect the East and West boundaries; • Dεy and Dσy are diagonal permittivity and conductivity matrices associated with the y-directed edges of the primary grid, respectively; • BW has all entries set to zero, except for a +1 at the intersection of each row associated with a West boundary edge and the column of the corresponding n+ 1 hanging variable in HW 2 ; • BE is analogous to BW , but has a −1 on each entry associated with an East boundary edge. B. Descriptor System Formulation The update equations derived in the previous sections form a complete dynamical model for the rectangular region (R + F)xn+1 y n = 1 (R − F)xn + Bun+ 2 , T n = L x . (14a) equation (14b) extracts these values from the state vector xn . The coefficients matrices in (14a) and (14b) read   Dεx 1 T Dlx Dly0 ∆t 0 2 Dlx Gy   Dε R= 0 Dly Dlx0 ∆ty − 12 Dly GTx  , (17) D 1 − 12 Gx Dly DA ∆tµ 2 Gy Dlx   D 1 T Dlx Dly0 2σx 0 D G l y x 2   Dσ F= 0 Dly Dlx0 2 y − 21 Dly GTx  , (18) 1 − 12 Gy Dlx 0 2 Gx Dly   Dlx BS Dlx BN 0 0 0 Dly BW Dly BE  , B= 0 (19) 0 0 0 0   −BS BN 0 0 0 BW −BE  . L= 0 (20) 0 0 0 0 Equations (14a)-(14b) define an FDTD-like model for a 2D lossy, source-free region, with possibly non-uniform permittivity and permeability. From a control perspective, this model is a discrete time descriptor system [25], also known as generalized state-space. From an electrical standpoint, the model is an impedance-type description of the region, because it gives the electric field tangential to the boundary in terms of the tangential magnetic field. In [24], a special case of the proposed representation was derived for a lossy region enclosed by PEC boundaries. The proposed developments generalize [24] in two directions: 1) the assumption of a PEC termination is removed. Suitable input-output variables are introduced in such a way that model (14a)-(14b) can be connected to other subsystems, such as different FDTD grids, boundary conditions, or reduced models; 2) the proposed formulation will allow us to prove, in the next section, that FDTD equations can be interpreted as a dissipative dynamical system. (14b) III. FDTD AS A D ISSIPATIVE D ISCRETE T IME S YSTEM The first equation (14a) is formed by update equations (4), (12) and (13), and updates the state vector  n  Ex n xn =  Ey  , (15) n− 12 Hz which consists of all electric and magnetic field samples in the 1 region, excluding hanging variables. The input vector un+ 2 n and output vector y are given by  n+ 1   n H 2 ES  Sn+ 21  n     1 H E N  yn =  nN  . un+ 2 =  (16) 1  n+ 2  EW  HW  n EE n+ 1 HE 2 The input vector contains all hanging variables, i.e. all magnetic field samples on the region boundaries. The output vector is made by the E samples at the same nodes, which are collected into vectors EnS , EnN , EnW and EnE . Output In this section, we define suitable expressions for the energy stored in the 2D region and the energy absorbed from its four boundaries. These expressions are then used to derive the conditions under which system (14a)-(14b) is dissipative. A. Dissipation Inequality A discrete time system like (14a)-(14b) is dissipative when it satisfies the following condition [26]. Definition 1. Dynamic system (14a)-(14b) is said to be 1 dissipative with supply rate s(yn , un+ 2 ) if there exists a n nonnegative function E(x ) with E(0) = 0, called storage function, such that 1 E(xn+1 ) − E(xn ) ≤ s(yn , un+ 2 ) n+ 12 for all u (21) and all n. The storage function E(xn ) can be interpreted as the energy stored in the system at time n, while the supply rate JOURNAL OF LATEX CLASS FILES, VOL. 11, NO. 4, DECEMBER 2012 5 1 s(yn , un+ 2 ) is the energy absorbed by the system from its boundaries between time n and n+1. Clearly, only dissipative systems can satisfy the dissipation inequality (21). Indeed, their stored energy can increase between time n and n + 1 by at most the energy absorbed from the outside world. If this limit is violated, the system is considered active, since it can generate energy on its own. For most practical systems, inequality (21) is satisfied strictly because of the presence of losses. B. Storage Function and Supply Rate For system (14a)-(14b), we choose as candidate storage function ∆t n T (x ) Rxn , (22) E(xn ) = 2 and as supply rate n+ 12 n s(y , u yn + yn+1 ) = ∆t 2
T 1 LT Bun+ 2 . Substituting (27) into (23), we obtain (EnS + En+1 )T n+ 12 S HS 2 n+1 T n n T (E + EN ) (E + En+1 n+ 1 n+ 1 W ) +∆t∆x N HN 2 +∆t∆y W HW 2 2 2 T (EnE + En+1 n+ 1 E ) HE 2 , (28) − ∆t∆y 2 and see that the supply rate is the sum of the energy absorbed by the region from each boundary between time n and n + 1. Signs in (28) are consistent with the direction of the Poynting vector on each boundary. 1 s(yn , un+ 2 ) = −∆t∆x C. Dissipativity Conditions Using the proposed storage function (22) and supply rate (23), we can derive simple dissipativity conditions on the coefficients matrices R, F, B and L in (14a)-(14b). Theorem 1. If (23) R = RT > 0 , (29a) Before showing that these functions satisfy (21), we investigate their physical meaning. Substituting (17) into (22), we get F + FT ≥ 0 , (29b) B = LLT B , (29c) 1 n T (E ) Dlx Dly0 Dεx Enx 2 x
T 1 Eny Dly Dlx0 Dεy Eny + 2   ∆t Dµ n− 21 n n Hz + Gy Dlx Ex − Gx Dly Ey + DA . (24) 2 ∆t E(xn ) = Since the term between square brackets is the right hand side of (4), we can reduce (24) to 1 T E(x ) = (Enx ) Dlx Dly0 Dεx Enx 2
T 1 Eny Dly Dlx0 Dεy Eny + 2 1  n− 12 T n+ 1 Hz DA Dµ Hz 2 . (25) + 2 n Z  A  1 2 1 1 2 2 εE (t) + εEy (t) + µHz (t) dA , 2 x 2 2 (26) the continuous expression for the electromagnetic energy per unit height of a 2D TE mode in a region of area A. The first term of (25) is the energy stored in the x component of the electric field in the region. Indeed, the diagonal matrix Dlx Dly0 is the area of the half cells around the Ex edges. Similarly, the second and third term in (25) represent the energy stored in the Ey and Hz components, respectively. We now discuss the physical meaning of supply rate (23). Direct inspection reveals that   −∆xINx 0 0 0   0 +∆xINx 0 0 . LT B =    0 0 +∆yINy 0 0 0 0 −∆yINy (27) then system (14a)-(14b) is dissipative according to Definition 1, with (22) as storage function and (23) as supply rate. Proof. Condition (29a) makes storage function (22) nonnegative for all xn , as required by Definition 1. Next, we show that if (29b) and (29c) hold, the dissipation inequality (21) will hold as well. Substituting (22) and (23) into (21), we obtain
T
T 1 T xn+1 Rxn+1 −(xn ) Rxn − yn + yn+1 LT Bun+ 2 ≤ 0 . This inequality can be rewritten as
T

T xn+1 R xn+1 − xn + xn+1 − xn Rxn
T 1 − yn + yn+1 LT Bun+ 2 ≤ 0 (30) From (14a), we have that

1 R xn+1 − xn = −F xn+1 + xn − Bun+ 2 (31)

T Using (31), the terms R xn+1 − xn and xn+1 − xn R in (30) can be expressed in terms of F
T

T − xn+1 F xn+1 + xn − xn+1 + xn FT xn  T
T 1 1 + xn+1 Bun+ 2 + un+ 2 BT xn
T 1 1 T − (xn ) LLT Bun+ 2 − xn+1 LLT Bun+ 2 ≤ 0 (32) Under (29c), we can finally rewrite (32) as
T

xn+1 + xn F + FT xn+1 + xn ≥ 0 (33) which is clearly satisfied if (29b) holds. We now investigate the physical meaning of the three dissipativity conditions, starting from (29c), which can be directly verified by combining (27) and (20). This condition holds because the region inputs and outputs in (16) are sampled at the same nodes, and is reminiscent of a similar JOURNAL OF LATEX CLASS FILES, VOL. 11, NO. 4, DECEMBER 2012 6 relation which holds for linear circuits under the impedance representation [27]. Condition (29b) is related to losses, and reads   Dlx Dly0 Dσx 0 0 F + FT =  0 Dly Dlx0 Dσy 0 ≥ 0 . (34) 0 0 0 Since Dlx , Dly , Dlx0 and Dly0 are diagonal with positive elements only, (34) boils down to D σx ≥ 0, (35) Dσy ≥ 0. (36) As one may expect, the average conductivity on each primary edge must be non-negative in a dissipative system. Condition (29a) is the most interesting, and is a generalized CFL criterion. From (17), we see that R = RT by construction. With the Schur’s complement [28], we can transform (29a) into Dlx Dly0 Dεx > 0 , (37) Dly Dlx0 Dεy > 0 , 4 IN , SST < ∆t2 Hz (38) (39) where h 1 −1 −1 −1 −1 S = DA 2 Dµ 2 Gy Dl2x Dl0 2 Dεx2 i 1 −1 −1 −Gx Dl2y Dl0 2 Dεy2 . y x (40) Inequalities (37) and (38) are clearly satisfied by construction. If we denote with sk the singular values of (40), inequality (39) will hold if 2 ∆t < ∀k . (41) sk This condition sets an upper bound on the FDTD time step, and is analogous to the generalized CFL constraint derived in [24] for an FDTD grid terminated on PEC boundaries. From (40), we indeed see that the time step upper limit depends on grid size, permittivity and permeability. Using a time step which violates (41) has two consequences. First, the discretization of Maxwell’s equations leads to an active discrete time model, even if the real physical system has positive losses everywhere, and should therefore be dissipative. The numerical model is thus inconsistent with the actual physics. Second, being able to generate energy on its own, model (14a)(14b) can lead to divergent transient simulations. Even when such model is connected to other dissipative subsystems, its ability to generate energy on its own can destabilize the whole simulation. The proposed theory provides a new, deeper understanding on the root causes of FDTD instability, and is able to pinpoint which part of an FDTD setup is responsible for it. It is hoped that this new explanation will facilitate the development of powerful FDTD schemes with guaranteed stability. D. Relation to CFL Stability Limit The relation between (41) and the CFL limit can be further understood if we apply (41) to a single cell. The purpose of (41) is to make storage function (22) non-negative. A sufficient condition for this is to require the energy stored in each primary cell to be positive. This condition can be derived by applying (39) to a single primary cell. We consider the primary cell extending from node (i, j) to node (i + 1, j + 1). The coefficient matrices for this special case can be obtained from the formulas in Sec. II by setting Nx = Ny = 1 Dlx = ∆xI2 ∆x I2 Dlx0 = 2  Gx = Gy = −1 1 Dly = ∆yI2 ∆y Dly0 = I2 2 DA = ∆x∆y  (42) (43) (44) The permeability matrix Dµ = µ|i+ 12 ,j+ 12 is the average permeability on the secondary edge, while the permittivity matrices     0 0 εx |i+ 12 ,j εy |i,j+ 12 Dεx = Dεy = 0 εx |i+ 21 ,j+1 0 εy |i+1,j+ 21 (45) contain the average permittivity on the four primary edges. Substituting (42)-(45) into (39), we obtain " ! 1 1 1 ∆t < + 2∆x2 µ|i+ 12 ,j+ 12 εy |i,j+ 12 εy |i+1,j+ 12 ! #− 21 1 1 1 , (46) + + 2∆y 2 µ|i+ 21 ,j+ 12 εx |i+ 12 ,j εx |i+ 21 ,j+1 which is a generalized CFL condition. For a uniform medium with permittivity ε and permeability µ, (46) reduces to ∆t < q 1 1 εµ q 1 ∆x2 + 1 ∆y 2 , (47) the CFL limit of 2D FDTD. This derivation confirms that (29a) is a generalized CFL condition, here reinterpreted in the context of dissipation. E. Application to Stability Analysis The proposed dissipation theory can be effectively used to investigate and enforce the stability of FDTD algorithms. Most FDTD setups consist of an interconnection of FDTD subsystems. In the simplest scenario, a uniform FDTD grid is connected to some boundary conditions. In most advanced scenarios, one may want to couple a main FDTD grid to refined grids, reduced models, lumped elements, or models from other numerical techniques, such as finite elements, integral equations or ray tracing. Ensuring stability of these hybrid schemes can be very challenging, since stability is a property of the overall scheme, rather than of its individual subsystems. By invoking the concept of dissipation, we can instead achieve stability in an easy and modular way. Each part is seen as an FDTD subsystem and required to satisfy dissipativity conditions (29a)-(29c). Since the connection of dissipative systems is dissipative by construction [17], the overall method will be guaranteed to be stable. The proposed theoretical framework generalizes the socalled energy method [8], and has numerous advantages over the state of the art: JOURNAL OF LATEX CLASS FILES, VOL. 11, NO. 4, DECEMBER 2012 1) non-uniform problems can be handled, unlike in the von Neumann analysis [1]; 2) stability conditions can be given on each subsystem separately, unlike in the iteration method [2], which requires the analysis of the iteration matrix of the whole scheme. This makes stability analysis modular and thus simpler; 3) once some given FDTD models have been proven dissipative, they can be arbitrarily interconnected without having to carry out further stability proofs. With the iteration method, when a single part of a coupled scheme changes, the whole proof must be revised; 4) the CFL limit of the resulting scheme can be easily determined by applying (37), (38) or (39) to each subsystem and taking the most restrictive CFL limit; 5) the stability framework is intuitive, since it is based on the fundamental physical concept of energy dissipation. 7 ∆x r Ĥz |ı̂+ 12 ,̂+ 12 Ĥz |ı̂+ 32 ,̂+ 12 ∆y r (ı̂, ̂) Ĥz |ı̂+ 21 ,̂ Êx |ı̂+ 12 ,̂ Ĥz |ı̂+ 32 ,̂ Êx |ı̂+ 32 ,̂ Hz |i+ 1 ,j (i, j) Ex |i+ 1 ,j 2 y 2 z Hz |i+ 1 ,j− 1 2 x 2 Figure 2: Subgridding scenario considered in Sec. IV for the case of r = 2. For clarity, a virtual gap has been inserted between the two grids. This virtual gap is closed when the two grids are connected. Fine mesh Ĥz |ı̂+ 1 ,̂ 2 Êx |ı̂+ 3 ,̂ Ĥz |ı̂+ 3 ,̂ 2 2 Êx |ı̂+ 1 ,̂ 2 IV. A PPLICATION : S TABLE FDTD S UBGRIDDING We demonstrate the proposed theory by deriving a subgridding algorithm which is stable by construction, easy to implement and supports an arbitrary grid refinement ratio. The goal is to derive stable update equations for a setup where one or more fine grids are embedded in a main coarse grid. Without loss of generality, we consider the case where a coarse grid with cell size ∆x × ∆y hosts a single fine grid with cell size ∆x/r × ∆y/r, where r is an arbitrary integer. The algorithm will ultimately consist of conventional FDTD equations to update the fields that fall strictly inside the two grids, and a special update equation to update the fields on the edges at the grid transition. To derive the method, it is sufficient to consider the interface between a single coarse cell and the corresponding fine cells, as shown in Fig. 2. In the figure a virtual gap has been opened between the two grids for clarity. Without loss of generality, we consider a refinement in the positive x direction. The other three cases can be derived in the same way. The coarse cell under consideration is centered at node (i+ 12 , j − 12 ) and the corresponding fine cells at nodes (ı̂ + 12 , ̂ + 21 ), ... , (ı̂ + r − 12 , ̂ + 21 ), where coordinates (ı̂, ̂) correspond to the same physical location as (i, j) in the coarse grid. Superscript “ˆ” denotes variables related to the fine grid. A subgridding algorithm can be interpreted as a the result of the connection of the three subsystems, as shown in Fig. 3. Two of those subsystems correspond to the coarse and fine grids to be coupled. The third subsystem represents the interpolation rule which is used to relate the fields on the boundaries of the two grids, that are sampled with different resolution. A. State Equations for the Coarse and Fine Cells on the Boundary For the coarse and fine grids, we adopt the formulation of Sec. II, introducing hanging variables on the two boundaries to be connected. The purpose of the hanging variables is to facilitate the coupling of the two meshes and the proof of stability. These extra variables will be eliminated when deriving the update equation for the fields at the interface. Interpolation rule Hz |i+ 1 ,j 2 Ex |i+ 1 ,j 2 Coarse mesh Figure 3: Interpretation of the subgridding method as the connection of three dynamical systems, representing the coarse grid, the fine grid, and the interpolation rule. On the North boundary of the coarse cell, we introduce the hanging variable n+ 12 HN n+ 1 = Hz |i+ 12,j . (48) 2 Similarly, on the South fine cell boundaries we introduce the hanging variables h iT n+ 1 n+ 1 n+ 12 ĤS 2 = Ĥz |ı̂+ 12,̂ . . . Ĥz |ı̂+r− , (49) 1 ,̂ 2 2 as shown in Fig. 2. From (11), we obtain the following state equation for the coarse E-field sample at the interface and the coarse hanging variable σx  n+1 ∆y  εx + EN = 2 ∆t 2  ∆y εx σx  n n+ 1 n+ 1 − EN + HN 2 − Hj− 12 , (50) 2 2 ∆t 2 where n+ 1 n+ 1 Hj− 12 = Hz |i+ 12,j− 1 , 2 2 2 n EN = Ex |ni+ 1 ,j , (51) 2 and εx and σx are the permittivity and conductivity on the primary edge of the coarse cell below the interface. Similarly, the state equation for the r fine cells can be written as ! ∆y D̂εx D̂σx + Ên+1 = S 2r ∆t 2 ! ∆y D̂εx D̂σx n+ 1 n+ 1 − ÊnS + Ĥ̂+ 12 − ĤS 2 , (52) 2 2r ∆t 2 JOURNAL OF LATEX CLASS FILES, VOL. 11, NO. 4, DECEMBER 2012 8 where n+ 1 Ĥ̂+ 12 2 h iT n+ 1 n+ 21 = Ĥz |ı̂+ 12,̂+ 1 . . . Ĥz |ı̂+r− , 1 1 2 2 2 ,̂+ 2 iT h ÊnS = Êx |nı̂+ 1 ,̂ . . . Êx |nı̂+r− 1 ,̂ , 2 2 (53) (54) and D̂εx and D̂σx are r × r diagonal matrices containing the values of permittivity and conductivity above the interface for the South edges where the fine Êx fields are sampled. B. Interpolation Rule The interpolation rule [29] relates the field sampled on the coarse and fine grids. From the boundary condition for tangential electric fields, we have that n ÊnS = EN T ∀n , (55) where T is an r × 1 matrix of ones. Condition (55) forces the coarsely- and finely-sampled E fields to be equal at all times. On the magnetic fields at the boundary, we impose a constraint reciprocal to (55) n+ 12 n+ 1 TT ĤS 2 ∀n . (56) r We will see in Sec. IV-D that the reciprocity between the E and H interpolation rules is required to ensure stability. HN = C. Explicit Update Equation for the Interface Interpolation conditions (55) and (56) can now be used to combine (50) and (52) in order to derive an explicit update equation for the fields at the coarse-fine interface, and eliminate hanging variables. Substituting (55) into (52), and multiplying the obtained equation by TT /r on the left yields ! ∆y TT D̂εx T TT D̂σx T n+1 + EN = 2r r∆t 2r ! ∆y TT D̂εx T TT D̂σx T n − EN 2r r∆t 2r n+ 1 TT Ĥ̂+ 12 n+ 1 TT ĤS 2 + − . (57) r r For the simplicity of notation we define symbols for the average permittivity and conductivity of the r South boundary fine cells 2 TT D̂σx T TT D̂εx T , σ̂x = . r r Equation (57) can now be added to (50), yielding ! σx + σ̂rx ∆y εx + ε̂rx n+1 + EN = 2 ∆t 2 ! σx + σ̂rx ∆y εx + ε̂rx n − EN 2 ∆t 2 ε̂x = n+ 12 + HN n+ 1 − Hj− 12 + 2 n+ 1 TT Ĥ̂+ 12 2 r (58) TT ĤS r Rearranging (60), we get the following explicit update equation for EN in terms of the neighboring magnetic fields !−1 ! σ̂x ε̂x σ̂x ε̂x σ + ε + σ + ε + x x x x n+1 n r r r r EN + − EN = ∆t 2 ∆t 2  !−1  T n+ 21 T Ĥ̂+ 1 1 εx + ε̂rx σx + σ̂rx 2 n+ 2  + + − Hj− 12  . 2 ∆y ∆t 2 r (61) The fine interface electric fields are then updated using (55). It should be noted that, when r = 1, equation (61) reduces to the standard FDTD update equation. The overall subgridding algorithm can be summarized as follows: 1) Calculate the magnetic field samples everywhere at time n + 12 using conventional FDTD update equations. 2) Use standard FDTD update equations to compute the E fields at time n + 1 on the edges that are strictly inside the coarse and fine grids. n+1 3) Compute EN , the coarsely-sampled electric field at the interface, using (61). 4) Update the finely-sampled Êx fields at the interface using (55). The computational overhead of this scheme is minimal, since the coefficients in (61) can be pre-computed before the update iterations. The proposed method is thus simple to implement, since it consists of conventional FDTD update equations inside the two meshes and a modified update equation for the edges at the interface. In comparison to previous subgridding methods, we avoid non-rectangular cells [6], finite element concepts [20] and Withney forms [21], [22]. The proposed update equation can be also used at corners with no modifications, unlike in previous works that require special treatment [5] or L-shaped cells [6]. Finally, we remark that in Sec. II, the FDTD update equations have been given in matrix form in order to reveal the dissipative nature of FDTD systems. This form, however, does not have to be used in the practical implementation, which can use conventional for loops or, in languages like MATLAB, vectorized operations. D. Proof of Stability n+ 21 − With the interpolation rule (56), we cancel the hanging variables to obtain ! σx + σ̂rx ∆y εx + ε̂rx n+1 + EN = 2 ∆t 2 ! n+ 12 T T Ĥ 1 σx + σ̂rx ∆y εx + ε̂rx ̂+ 12 n+ n EN − Hj− 12 + − . 2 2 ∆t 2 r (60) . (59) The proposed dissipation theory makes it straightforward to prove that the subgridding algorithm is stable under the CFL limit of the fine grid. For stability, all three subsystems in Fig. 3 need to be dissipative, which requires one to use the JOURNAL OF LATEX CLASS FILES, VOL. 11, NO. 4, DECEMBER 2012 9 more restrictive fine grid time step. With the time step chosen correctly, in order to guarantee the overall stability we need to only ensure dissipativity of the interpolation rule. Analogously to (23) the supply rate for the interpolation subsystem is defined as n+1 EN + n+ 1 ) = −∆t∆x HN 2 s(yn , u 2 )T n+ 21 ∆x (ÊnS + Ên+1 S + ∆t ĤS . (62) r 2 Substituting (55) and (56) into (62), we have n+ 12 20 mm ∆x = 1 mm ∆y = 2 mm 40 mm n EN PEC Probe y ∆x̂ = ∆x/4 ∆ŷ = ∆y/4 z x 40 mm Source 60 mm Figure 4: Layout of the PEC cavity considered in Sec. V-A. 1 s(yn , un+ 2 ) = n+ 12 TT ĤS r − n+ 1 HN 2 ! = 0 . (63) Therefore, the proposed interpolation rule is a lossless system that does not dissipate nor absorb any energy. Physically, this result makes sense, since the connection system corresponds to an infinitely thin region where no energy dissipation can take place. In conclusion, since the proposed subgridding method can be seen as the connection of three dissipative systems, it is overall dissipative, and thus stable. 0.5 Hz (A/m) n+1 T (E n + EN ) ∆t∆x N 2 0 −0.5 0 2 4 6 Timestep 8 10 5 x 10 Figure 5: Magnetic field at the probe for the the empty cavity with subgridding of Sec. V-A, computed for 106 time steps. V. N UMERICAL E XAMPLES The following sections provide the results of FDTD simulations that were done to verify the proposed theory. The subgridding algorithm was implemented in Matlab and tests were performed in order to check its stability, ability to handle material traverse, its accuracy and speedup capability. A. Stability Verification Stability was verified by simulating an empty cavity with perfect electric conductor (PEC) walls with a centrally placed subgridding region for 106 time steps. The layout of the simulation is shown in Fig. 4. The cavity was excited using a modulated Gaussian magnetic current source with central frequency of 3.75 GHz and half-width at half-maximum of 0.74 GHz. Magnetic field was recorded at a probe placed inside the cavity. The time step was set 1% below the CFL limit of the fine grid. The resulting waveform in Fig. 5 shows that no instability occurred after 106 time steps. Stable behavior after such a large number of time steps verifies the correctness of the proposed stability enforcement technique, especially since no lossy materials were present to dissipate any spurious energy artificially created by the algorithm. B. Material Traverse The ability of the proposed method to produce meaningful results when objects traverse the subgridding interface was tested using the setup in Fig. 6. A 16 × 16 mm slab of material was simulated for three different placements of the subgridding region: enclosing, traversing and away from the slab. The test was done for copper and a lossy dielectric with conductivity of 5 S/m and relative permittivity of 2. As a reference, uniformly discretized all-coarse and all-fine simulations were performed at the fine time step, in addition to the subgridding simulations. Coarse and fine meshes in uniformly discretized and subgridding runs were chosen as 1 mm and 0.2 mm respectively. 15 mm-thick perfectly matched layer (PML) terminated the simulation region. Modulated Gaussian magnetic current excitation was used at 15.0 GHz central frequency with 8.82 GHz half-width at half-maximum bandwidth. The time step was chosen as 0.467 ps in all test cases. The magnetic field waveforms at the probe recorded in the different subgridding scenarios are shown in Fig. 7, and are in excellent agreement among each other. This result confirms that the proposed subgridding method can properly handle material traverse, for both very good conductors and for lossy dielectrics. C. Application to Simulating Scatterer Reflections In order to investigate the accuracy of the proposed scheme, we looked at waveguide reflections from the scatterer shown in Fig. 8, which consisted of four copper rods with 1 mm radius. We have also investigated the reflections from the subgridding interface, in order to assess the quality of the subgridding scheme. The coarse cell was chosen to be 1 mm, which was 1 10 of the minimum wavelength of interest that corresponded to 30.0 GHz. The fields of the incident wave were computed by running a reference simulation without the scatterer and without subgridding. Reflected wave fields were found by subtracting the incident wave fields from the total fields in simulations with the scatterer. 15 mm-thick PML boundary was chosen to terminate the two sides of the waveguide. The JOURNAL OF LATEX CLASS FILES, VOL. 11, NO. 4, DECEMBER 2012 10 PEC 40 mm x z Source subgrids y Jy current Probes 8×8 mm y Probe PML PML x z PML 20 mm 2 mm ∆x = 1 mm ∆y = 1 mm Figure 6: Layout used for the material traverse test of Sec. V-B, and the three different placements of the subgridding region. 2 mm Scatterer 66 mm Figure 8: Layout of the four-rod reflection simulation discussed in Sec. V-C. The dashed line shows the location of the subgrid in the subgridding run. Lossy dielectric −5 With scatterer −10 0 11 |S | (dB) Hz (A/m) 0.02 −0.02 −15 −20 −25 −30 −35 Copper Sub−grid interface only −40 0 |S | (dB) Hz (A/m) 0.02 11 −0.02 0.2 0.25 Time (ns) 0.3 0.35 Figure 7: Time-domain magnetic field at the probe recorded for the three locations of the subgrid in Sec. V-B: enclosed ( ), outside ( ) and traversing the object ( ). Waveforms from the uniformly discretized all-coarse (  ) and allfine ( ) simulations are also shown. time steps in the subgridding runs were set 1% below the CFL limit of the refined region. Uniformly discretized simulations were run 1% below the CFL limits of the respective grids. The resulting reflections are shown in Fig. 9 for the subgridding case with different refinement ratios, as well as the reference run with full refinement by a factor of 6. The simulation times are shown in Table I. It can be seen that the local refinement of the grid around the scatterer with the proposed subgridding method can improve the accuracy substantially compared to the coarse grid run. The larger choice of refinement ratio makes the solution very close to the reference all-fine solution. Moreover, very good speedup - almost by a factor of 11 - is achieved even for the grid refinement of 6 when the redundant high resolution of the grid in air is eliminated. The reflections from subgridding interface were significantly lower than those from the scatterer, further demonstrating the accuracy of the proposed method. −60 −80 −100 0 5 10 15 20 Frequency (GHz) 25 Figure 9: Reflected power with respect to the incident for the example of Sec. V-C. Top panel: reflections from the four-rod scatterer for different global discretization: all-coarse (  ) and all-fine with r = 6 ( ); and for the subgridding runs: r=2( ), r = 4 ( ) and r = 6 ( ). Bottom panel: reflections from the subgrid interface only. D. Application to Exposure Studies We show the possibility of applying the method for multiscale human exposure simulations. The chosen setup is shown in Fig. 10. A transverse cross-section of the head of the model of Ella from IT’IS Virtual Population V1.x [30] was used to assign the material properties [31] to the FDTD cells. A 900 MHz source was placed approximately 3 meters away from the human head. The simulation region was terminated with 20 cm PML. The reference (fine) resolution in the tissues was set to 2 mm, based on the mesh size chosen by [32] in a radiation exposure study at that frequency. Specific absorption rate, or SAR, was evaluated according to the formula in [33], which was used as follows for sinusoidal excitation SAR = σ(Ex2p + Ey2p ) 2ρ , (64) JOURNAL OF LATEX CLASS FILES, VOL. 11, NO. 4, DECEMBER 2012 Table I: Simulation times for different meshes in four-rod scatterer simulations in Sec. V-C. Simulation times were recorded for the total field run only - not for the reference run. Method All-fine (r = 6) All-coarse Subgridding (r = 2) Subgridding (r = 4) Subgridding (r = 6) Simulation time (s) 427.4 3.2 10.7 23.3 39.5 3.05m Head slice Method All-fine All-coarse Subgridding All-fine z Error in SAR integral Not applicable 59.5% -3.1% Subgridding Simulation time (s) 1596.4 6.9 34.5 All-coarse x PML 4.01m Figure 10: Layout of the simulation in Sec. V-D with human head cross-section placed approximately 3 meters away from a point source. where subscript “p” denotes the peak absolute value of a field component and ρ corresponds to tissue density. SAR was calculated for each of the primary cells. The values of the electric field components at the primary cell centers were found by averaging the nearest known samples at the cell edges. The peak values of the fields were found for the time interval from 25.6 ns to 26.8 ns, which gave the wave sufficient time to reach the head and penetrate inside it. In the subgridding run, the empty space was coarsened to 1 1 cm, which corresponded to 33.3 of the wavelength. The reference all-fine run was performed at 2 mm resolution, along with the all-coarse run where the entire simulation region was discretized at 1 cm. The volumetric integral of SAR over the tissues was used as an accuracy metric XX SAR|i+ 12 ,j+ 12 ∆x∆y , (65) i Table II: Error in the integral of SAR and simulation times in the simulations discussed in Sec. V-D. y subgrid Source 11 j where ∆x and ∆y are FDTD cell dimensions in the tissues. Time step of 4.67 ps was chosen for the all-fine simulation and for the subgridding simulation. The coarse grid case was run at 23.11 ps. The resulting SAR maps are shown in Fig. 11. Table II shows simulation times and percent error in the total SAR with respect to the all-fine simulation. The results show that with subgridding a speedup of 46X can be obtained with only 3.1% loss in accuracy. No noticeable difference can be seen on the SAR maps in Fig. 11 between the subgridding and all-fine simulations. When, instead, the coarse resolution was chosen for the entire grid, the head tissues were not sufficiently resolved and the integral SAR differed from the reference by 59.5%, showing the necessity of local grid refinement. VI. C ONCLUSION This paper proposed a dissipative systems theory for FDTD, recognizing that FDTD equations can be seen as a dynamical Figure 11: SAR maps obtained in Sec. V-D with all-fine (2 mm), all-coarse (1 cm) discretization and with subgridding (2 mm in the tissues and 1 cm in the air). system which is dissipative under a generalized CourantFriedrichs-Lewy condition. The theory provides a new, powerful framework to make FDTD stability analysis simpler and modular. Stability conditions can indeed be given on the individual components (e.g. boundary conditions, meshes, thin-wire models) rather than on the whole coupled FDTD setup. The theory is intuitive since rooted on the familiar concept of energy dissipation, and sheds new light on the root mechanisms behind FDTD instability. As an example of application, a simple yet effective subgridding algorithm is derived, with straightforward stability proof. The proposed algorithm allows material traverse, is simpler to implement than existing solutions, and supports an arbitrary grid refinement ratio. Numerical results confirm its stability and accuracy. Speedups of up to 46X were observed with only 3.1% error with respect to standard FDTD. R EFERENCES [1] A. Taflove and S. C. Hagness, Computational electrodynamics. Artech house, 2005. [2] S. D. Gedney, Introduction to the Finite-Difference Time-Domain (FDTD) Method for Electromagnetics, 1st ed. San Rafael, CA: Morgan & Claypool Publishers, 2011. [3] K. Yee, “Numerical solution of initial boundary value problems involving Maxwell’s equations in isotropic media,” IEEE Trans. Antennas Propag., vol. 14, no. 3, pp. 302–307, 1966. [4] M. Okoniewski, E. Okoniewska, and M. Stuchly, “Three-dimensional subgridding algorithm for FDTD,” IEEE Trans. Antennas Propag., vol. 45, no. 3, pp. 422–429, 1997. [5] P. Thoma and T. Weiland, “A consistent subgridding scheme for the finite difference time domain method,” Int. J. Numer. Model El., vol. 9, no. 5, pp. 359–374, 1996. [6] K. Xiao, D. J. Pommerenke, and J. L. Drewniak, “A three-dimensional FDTD subgridding algorithm with separated temporal and spatial interfaces and related stability analysis,” IEEE Trans. Antennas Propag., vol. 55, no. 7, pp. 1981–1990, 2007. [7] W. Sui, D. A. Christensen, and C. H. Durney, “Extending the twodimensional FDTD method to hybrid electromagnetic systems with active and passive lumped elements,” IEEE Trans. Microw. Theory Techn., vol. 40, no. 4, pp. 724–730, 1992. [8] F. Edelvik, R. Schuhmann, and T. Weiland, “A general stability analysis of FIT/FDTD applied to lossy dielectrics and lumped elements,” International Journal of Numerical Modelling: Electronic Networks, Devices and Fields, vol. 17, no. 4, pp. 407–419, 2004. JOURNAL OF LATEX CLASS FILES, VOL. 11, NO. 4, DECEMBER 2012 [9] B. Denecker, F. Olyslager, L. Knockaert, and D. De Zutter, “Generation of FDTD subcell equations by means of reduced order modeling,” IEEE Trans. Antennas Propag., vol. 51, no. 8, pp. 1806–1817, 2003. [10] X. Li, C. D. Sarris, and P. Triverio, “Structure-Preserving Reduction of Finite-Difference Time-Domain Equations with Controllable Stability Beyond the CFL Limit,” IEEE Trans. Microw. Theory Techn., vol. 62, no. 12, pp. 3228–3238, 2014. [11] X. Li and P. Triverio, “Stable FDTD Simulations with Subgridding at the Time Step of the Coarse Grid: a Model Order Reduction Approach,” in IEEE MTT-S Int. Conf. on Numerical Electromagnetic and Multiphysics Modeling and Optimization, Ottawa, Canada, August 11–14 2015. [12] J.-F. Lee, R. Lee, and A. Cangellaris, “Time-domain finite-element methods,” IEEE Trans. Antennas Propag., vol. 45, no. 3, pp. 430–442, 1997. [13] A. R. Bretones, R. Mittra, and R. G. Martı́n, “A hybrid technique combining the method of moments in the time domain and FDTD,” IEEE Microw. Guided Wave Lett., vol. 8, no. 8, pp. 281–283, 1998. [14] Y. Wang, S. Safavi-Naeini, and S. K. Chaudhuri, “A hybrid technique based on combining ray tracing and FDTD methods for site-specific modeling of indoor radio wave propagation,” IEEE Trans. Antennas Propag., vol. 48, no. 5, pp. 743–754, 2000. [15] T. Namiki, “A new FDTD algorithm based on alternating-direction implicit method,” IEEE Microw. Wireless Compon. Lett., vol. 47, no. 10, pp. 2003–2007, 1999. [16] F. Zheng, Z. Chen, and J. Zhang, “A Finite-Difference time-domain method without the Courant stability conditions,” IEEE Microwave Guided Wave Lett., vol. 9, no. 11, pp. 441–443, Nov 1999. [17] P. Triverio, S. Grivet-Talocia, M. S. Nakhla, F. Canavero, R. Achar, “Stability, causality, and passivity in electrical interconnect models,” IEEE Trans. Adv. Packag., vol. 30, no. 4, pp. 795–808, 2007. [18] J. C. Willems, “Dissipative dynamical systems part i: General theory,” Archive for rational mechanics and analysis, vol. 45, no. 5, pp. 321–351, 1972. [19] Y. Wang, S. Langdon, and C. Penney, “Analysis of accuracy and stability of FDTD subgridding schemes,” in 2010 European Microwave Conf. IEEE, 2010, pp. 1297–1300. [20] F. Collino, T. Fouquet, and P. Joly, “Conservative space-time mesh refinement methods for the FDTD solution of Maxwell’s equations,” J. Comput. Phys., vol. 211, no. 1, pp. 9–35, 2006. [21] R. A. Chilton and R. Lee, “Conservative and provably stable FDTD subgridding,” IEEE Trans. Antennas Propag., vol. 55, no. 9, pp. 2537– 2549, 2007. [22] N. V. Venkatarayalu, R. Lee, Y.-B. Gan, and L.-W. Li, “A stable FDTD subgridding method based on finite element formulation with hanging variables,” IEEE Trans. Antennas Propag., vol. 55, no. 3, pp. 907–915, 2007. [23] J. Brewer, “Kronecker products and matrix calculus in system theory,” IEEE Trans. Circuits Syst., vol. 25, no. 9, pp. 772–781, 1978. [24] B. Denecker, F. Olyslager, L. Knockaert, and D. De Zutter, “A new statespace-based algorithm to assess the stability of the Finite-Difference time-domain method for 3D finite inhomogeneous problems,” AEU-Int. J. Electron. C, vol. 58, no. 5, pp. 339 – 348, 2004. [25] L. Dai, “Singular control systems,” 1989. [26] C. Byrnes and W. Lin, “Losslessness, feedback equivalence, and the global stabilization of discrete-time nonlinear systems,” IEEE Trans. Autom. Control, vol. 39, no. 1, pp. 83–98, 1994. [27] S. Grivet-Talocia and B. Gustavsen, Passive Macromodeling: Theory and Applications. Hoboken, NJ: Wiley, 2015. [28] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan, Linear Matrix Inequalities in System and Control Theory, ser. Studies in Applied Mathematics. SIAM, 1994, vol. 15. [29] L. Kulas and M. Mrozowski, “Reciprocity principle for stable subgridding in the finite difference time domain method,” in EUROCON, 2007. The International Conference on “Computer as a Tool”, Sept 2007, pp. 106–111. [30] A. Christ et al., “The Virtual Family - Development of surface-based anatomical models of two adults and two children for dosimetric simulations,” Physics in Medicine and Biology, vol. 55, no. 2, pp. N23– N38, 2010. [31] IT’IS Foundation, “Overview - Database of Tissue Properties.” [Online]. Available: http://www.itis.ethz.ch/virtual-population/tissueproperties/overview/ [Aug. 19, 2015]. [32] S. Kühn et al., “MMF/GSMA phase 2: scientific basis for base station exposure compliance standards,” IT’IS Foundation, Tech. Rep. 21, June 2009. [33] D. A. Sánchez-Hernández, High Frequency Electromagnetic Dosimetry. Boston, MA: Artech House, 2009. 12 

5cs.CE

Approximation algorithms for the maximum weight internal spanning tree problem∗ arXiv:1608.03299v2 [cs.DS] 26 May 2017 Zhi-Zhong Chen† Guohui Lin‡§ Lusheng Wang¶ Yong Chenk‡ Dan Wang¶ May 30, 2017 Abstract Given a vertex-weighted connected graph G = (V, E), the maximum weight internal spanning tree (MwIST for short) problem asks for a spanning tree T of G such that the total weight of the internal vertices in T is maximized. The unweighted variant, denoted as MIST, is NPhard and APX-hard, and the currently best approximation algorithm has a proven performance ratio 13/17. The currently best approximation algorithm for MwIST only has a performance ratio 1/3 − ǫ, for any ǫ > 0. In this paper, we present a simple algorithm based on a novel relationship between MwIST and the maximum weight matching, and show that it achieves a better approximation ratio of 1/2. When restricted to claw-free graphs, a special case been previously studied, we design a 7/12-approximation algorithm. Keywords: Maximum weight internal spanning tree; maximum weight matching; approximation algorithm; performance analysis ∗ An extended abstract appears in the Proceedings of COCOON 2017. Division of Information System Design, Tokyo Denki University. Hatoyama, Saitama 350-0394, Japan. Email: zzchen@mail.dendai.ac.jp ‡ Department of Computing Science, University of Alberta. Edmonton, Alberta T6G 2E8, Canada. Email: guohui@ualberta.ca § Correspondence author. ¶ Department of Computer Science, City University of Hong Kong. Tat Chee Avenue, Kowloon, Hong Kong, China. Email: cswangl@cityu.edu.hk k Institute of Operational Research and Cybernetics, Hangzhou Dianzi University. Hangzhou, Zhejiang 310018, China. Email: chenyong@hdu.edu.cn † 1 2 1 Chen et al. /v:May 30, 2017 Introduction In the maximum weight internal spanning tree (MwIST for short) problem, we are given a vertexweighted connected graph G = (V, E), where each vertex v of V has a nonnegative weight w(v), with the objective to compute a spanning tree T of G such that the total weight of the internal vertices in T , denoted as w(T ), is maximized. MwIST has applications in the network design for cost-efficient communication [18] and water supply [1]. When the vertex weights are uniform, or simply vertex-unweighted, the problem is referred to as the maximum internal spanning tree (MIST for short) problem. MIST is clearly NP-hard because it includes the NP-hard Hamiltonian-path [6] problem as a special case. Furthermore, MIST has been proven APX-hard [12], suggesting that it does not admit a polynomial-time approximation scheme (PTAS). In the literature, much research is done on designing (polynomial-time, if not specified) approximation algorithms for MIST to achieve the worst-case performance ratio as close to 1 as possible. The probably first approximation for MIST is a local search algorithm, which achieves a ratio of 1/2 and is due to Prieto and Sliper [14]. Salamon and Wiener [18] later modified slightly Prieto and Sliper’s algorithm to make it run faster (in linear-time) while achieving the same ratio of 1/2. Besides, two special cases of MIST were considered by Salamon and Wiener [18]: when restricted to claw-free graphs, they designed a 2/3-approximation algorithm; when restricted to cubic graphs, they designed a 5/6-approximation algorithm. Later, Salamon [17] proved that the 1/2-approximation algorithm in [18] actually achieves a performance ratio of 3/(r + 1) for the MIST problem on r-regular graphs (r ≥ 3). Based on local optimization, Salamon [16] presented an O(n4 )-time 4/7-approximation algorithm for MIST restricted to graphs without leaves. The algorithm was subsequently simplified and re-analyzed by Knauer and Spoerhase [7] to run faster (in cubic time), and it becomes the first improved 3/5-approximation for the general MIST. Via a deeper local search strategy than those in [7] and [16], Li et al. [8] presented a further improved approximation algorithm for MIST with ratio 2/3. At the same time, Li and Zhu [12] presented another 2/3-approximation algorithm for MIST. Unlike the other previously known approximation algorithms for MIST, the 2/3-approximation by Li and Zhu [12] is based on a simple but crucial observation that the maximum number of internal vertices in a spanning tree of a graph G can be upper bounded by the maximum number of edges in a triangle-free 2-matching (a.k.a. path-cycle cover) of G. The time complexity of this approximation algorithm is dominated by computing the maximum triangle-free 2-matching, O(nm1.5 log n), where n and m are the numbers of vertices and edges in G, respectively. Li and Zhu [11] claimed that they are able to further improve their design to achieve a 3/4-approximation algorithm for MIST, of the same time complexity. Recently, Chen et al. [2] gave another 3/4approximation algorithm for MIST, which is simpler than the one in [11]; and they showed that by applying three more new ideas, the algorithm can be refined into a 13/17-approximation algorithm for MIST of the same time complexity. This is currently the best approximation algorithm for MIST. The parameterized MIST by the number of internal vertices k, and its special cases and variants, have also been extensively studied in the literature [14, 13, 15, 3, 5, 1, 4, 9, 10]. The best known kernel for the general problem has a size 2k, which leads to the fastest known algorithm with running time O(4k nO(1) ) [9]. For the vertex-weighted version, MwIST, Salamon [16] designed the first O(n4 )-time 1/(2∆−3)approximation algorithm, based on local search, where ∆ is the maximum degree of a vertex in the input graph. For MwIST on claw-free graphs without leaves, Salamon [16] also designed an O(n4 )-time 1/2-approximation algorithm. Subsequently, Knauer and Spoerhase [7] proposed the Approximating the maximum weight internal spanning tree 3 first constant-ratio 1/(3 + ǫ)-approximation algorithm for the general MwIST, for any constant ǫ > 0. The algorithm is based on a new pseudo-polynomial time local search algorithm, that starts with a depth-first-search tree and applies six rules to reach a local optimum. It yields a 1/3-approximation for MwIST and then is extended to a polynomial time 1/(3 + ǫ)-approximation scheme. The authors also showed that the ratio of 1/3 is asymptotically tight. In this paper, we deal with the MwIST problem. We first prove a novel relationship between the total weight of the internal vertices in a spanning tree of the given vertex-weighted graph and the maximum weight matching of an edge-weighted graph, that is constructed out of the given vertexweighted graph. Based on this relationship, we present a simple 1/2-approximation algorithm for MwIST; this ratio 1/2 significantly improves upon the previous known ratio of 1/3. When restricted to claw-free graphs, a special case previously studied in [18, 16], we design a 7/12-approximation algorithm, improving the previous best ratio of 1/2. 2 The 1/2-approximation algorithm Recall that in the MwIST problem, we are given a connected graph G = (V, E), where each vertex v of V has a nonnegative weight w(v), with the objective to compute a spanning tree T of G such that the total weight of the internal vertices in T , denoted as w(T ), is maximized. We note that for such an objective function, we may assume without loss of generality that every leaf in the given graph G has weight 0. We construct an edge-weighted graph based on G = (V, E). In fact, the structure of the new graph is identical to that of G: the vertex set is still V , but instead the vertices have no weights; the edge set is still E, where the weight of each edge e = {u, v} is w(e) = w(u) + w(v), i.e., the weight of an edge is the total weight of its two ending vertices in the original graph. Since there is no ambiguity when we discuss the edge weights or the vertex weights, the new edge-weighted graph is still referred to as G. The weight of an edge subset refers to the total weight of the edges therein; while the weight of an acyclic subgraph refers to the total weight of the internal (and those surely will become internal) vertices therein. Let M ∗ denote the maximum weight matching of (the edge-weighted graph) G, which can be computed in O(n min{m log n, n2 })-time, where n = |V | and m = |E|. Lemma 1 Given a spanning tree T of G, we can construct a matching M of G such that w(T ) ≤ w(M ). Proof. We construct M iteratively. Firstly, we root the tree T at an internal vertex r, and all the edges of T are unmarked; then in every iteration we include into M an unmarked edge e = {u, v} of T such that 1) both u and v are internal and 2) e is the closest to the root r measured by the number of edges on the path from r to e, followed by marking all the edges incident at u or v. This way, the total weight of the two internal vertices u and v in the tree T is transferred to M by adding the edge e to M . At the time this iterative procedure stops, there is no unmarked edge of T connecting two internal vertices, and thus every internal vertex whose weight has not been transferred to M must be adjacent to at least a leaf each via an unmarked edge. Next, we iteratively include into M a remaining unmarked edge e = {u, v} of T , followed by marking all the edges incident at u or v. This way, the total weight of u and v, which is greater than or equal to the weight of the internal vertex between u and v, is transferred to M by adding the edge e to M . At the end of this procedure, T contains no more unmarked edges. Since leaves in the tree T count nothing towards w(T ), we conclude that w(T ) ≤ w(M ). This proves the lemma. ✷ 4 Chen et al. /v:May 30, 2017 The following corollary directly follows from Lemma 1, stating an upper bound on the total weight of an optimal solution to the MwIST problem. Corollary 1 Let T ∗ denote an optimal (maximum weight internal) spanning tree of G. Then, w(T ∗ ) ≤ w(M ∗ ). We next start with M ∗ to construct a spanning tree T . Let the edges of M ∗ be e1 , e2 , . . . , ek ; let ej = {aj , bj }, such that w(aj ) ≥ w(bj ), for all j = 1, 2, . . . , k. Note that there could be vertices of degree 0 in the spanning subgraph G[V, M ∗ ] with the edge set M ∗ , and there could be edges of weight 0 in M ∗ ; let X denote the set of such degree-0 vertices and the end-vertices of such weight-0 edges. Essentially we do not worry about the degree of any vertex of X in our final tree T , since their weights (if any) are not counted towards w(M ∗ ). This way, we assume without loss of generality that w(aj ) > 0 for each edge ej of M ∗ , and consequently the degree of aj is dG (aj ) ≥ 2, that is, aj is adjacent to at least one other vertex than bj in the graph G. Let A = {aj | j = 1, 2, . . . , k}, and B = {bj | j = 1, 2, . . . , k}; note that V = A ∪ B ∪ X. Let E aa = E(A, A ∪ X), i.e., the set of edges each connecting a vertex of A and a vertex of A ∪ X, and E ab = E(A, B), i.e., the set of edges each connecting a vertex of A and a vertex of B. Our construction algorithm first computes a maximal acyclic subgraph of G, denoted as H0 , by adding a subset of edges of E aa to M ∗ . This subset of edges is a maximum weight spanning forest on A ∪ X, and it can be computed in O(|E aa | log n)-time via a linear scan. In the achieved subgraph H0 , if one connected component C contains more than one edge, then the vertex aj of each edge ej = {aj , bj } in C has degree at least 2, i.e. is internal. Therefore, the total weight of the internal vertices in the component C is at least half of w(C ∩ M ∗ ), and C is called settled and left alone by the algorithm. Our algorithm next considers an arbitrary edge of M ∗ that is not yet in any settled component, say ej = {aj , bj }. In other words, the edge ej is an isolated component in the subgraph H0 . This implies that the vertex aj is not incident to any edge of E aa , and thus it has to be adjacent to some vertex in B − {bj }. If aj is adjacent to some vertex bi in a settled component, then this edge (aj , bi ) is added to the subgraph H0 (the edge ej is said merged into a settled component) and the iteration ends. The updated component remains settled, as w(aj ) ≥ w(ej )/2 is saved towards the weight of the final tree T . In the other case, the vertex aj is adjacent to a vertex bi , such that the edge ei = {ai , bi } is also an isolated component in the current subgraph. After adding the edge (aj , bi ) to the subgraph, the algorithm works with the vertex ai exactly the same as with aj at the beginning. That is, if ai is adjacent to some vertex bℓ in a settled component, then this edge (ai , bℓ ) is added to the subgraph (the component that ai belongs to is merged into a settled component) and the iteration ends; if ai is adjacent to a vertex bℓ , such that the edge eℓ = {aℓ , bℓ } is also an isolated component in the current subgraph, then the edge (ai , bℓ ) is added to the subgraph, the algorithm works with the vertex aℓ exactly the same as with aj at the beginning; in the last case, ai is adjacent to a vertex bℓ inside the current component that ai belongs to, then the edge (ai , bℓ ) is added to the current component to create a cycle, subsequently the lightest edge of M ∗ in the cycle is removed, the iteration ends, and the current component becomes settled. We note that in the above last case, the formed cycle in the current component contains at least 2 edges of M ∗ ; breaking the cycle by removing the lightest edge ensures that at least half of the total weight of the edges of M ∗ in this cycle (and thus in this component) is saved towards the weight of the final tree T . Therefore, when the iteration ends, the resulting component is settled. When the second step of the algorithm terminates, there is no isolated edge of M ∗ in the current subgraph, denoted as H1 , and each component is acyclic and settled. In the last step, the algorithm Approximating the maximum weight internal spanning tree 5 connects the components of H1 into a spanning tree using any possible edges of E. We denote the entire algorithm as Approx. Lemma 2 At the end of the second step of the algorithm Approx, every component C of the achieved subgraph H1 is acyclic and settled (i.e., w(C) ≥ w(C ∩ M ∗ )/2). Proof. Let C denote a component; C ∩ M ∗ is the subset of M ∗ , each edge of which has both end-vertices in C. If C is obtained at the end of the first step, then C is acyclic and for every edge ej ∈ C ∩ M ∗ , the vertex aj has degree at least 2, and thus w(C) ≥ w(C ∩ M ∗ )/2. If a subgraph of C is obtained at the end of the first step but C is finalized in the second step, then C is also acyclic and for every edge ej ∈ C ∩ M ∗ , the vertex aj has degree at least 2, and thus w(C) ≥ w(C ∩ M ∗ )/2. If C is newly formed and finalized in the second step, then at the time C was formed, there was a cycle containing at least 2 edges of M ∗ of which the lightest one is removed to ensure the acyclicity, and thus the total weight of the internal vertices on this path is at least half of the total weight of the edges of M ∗ on this cycle. Also, the vertex aj of every edge ej not on the cycle has degree at least 2. Thus, w(C) ≥ w(C ∩ M ∗ )/2. ✷ Theorem 1 The algorithm Approx is a 1/2-approximation for the MwIST problem. Proof. One clearly sees that Approx runs in polynomial time, and in fact the running time is dominated by computing the maximum weight matching M ∗ . From Lemma 2, at the end of the second step of the algorithm Approx, every component C of the achieved subgraph H1 is acyclic and satisfies w(C) ≥ w(C ∩ M ∗ )/2. Since there is no edge of M ∗ connecting different components of the subgraph H1 , the total weight of the internal vertices in H1 is already at least w(M ∗ )/2, i.e. w(H1 ) ≥ w(M ∗ )/2. The last step of the algorithm may only increase the total weight. This proves that the total weight of the internal vertices of the tree T produced by Approx is w(T ) ≥ w(H1 ) ≥ w(M ∗ )/2 ≥ w(T ∗ )/2, where the last inequality is by Corollary 1, which states that w(M ∗ ) is an upper bound on the optimum. Thus, Approx is a 1/2-approximation for the MwIST problem. ✷ 3 A 7/12-approximation algorithm for claw-free graphs We present a better approximation algorithm for the MwIST problem on claw-free graphs. A graph G = (V, E) is called claw-free if, for every vertex, at least two of its arbitrary three neighbors are adjacent. We again assume without loss of generality that every leaf in the graph G has weight 0. Besides, we also assume that |V | ≥ 5. We first present a reduction rule, which is a subcase of Operation 4 in [2], that excludes certain induced subgraphs of the given graph G from consideration. Operation 1 If G has a cut-vertex v such that one connected component C of G − v has two, three or four vertices, then obtain G1 from G − V (C) by adding a new vertex u of weight 0 and a new edge {v, u}. 6 Chen et al. /v:May 30, 2017 Let tw(C) denote the maximum total weight of the internal vertices in a spanning tree of the subgraph induced on V (C) ∪ v, in which w(v) is revised to 0. Then there is an optimal spanning tree T1 of G1 of weight w(T1 ) if and only if there is an optimal spanning tree T of G of weight w(T ) = w(T1 ) + tw(C). Proof. Let Gc denote the subgraph induced on V (C) ∪ v, that is, Gc = G[V (C) ∪ v]; and let Tc denote the spanning tree of Gc achieving the maximum total weight of the internal vertices, that is, w(Tc ) = tw(C) (Tc can be computed in O(1)-time). Note that in T1 , the leaf u must be adjacent to v and thus w(v) is counted towards w(T1 ). We can remove the edge {v, u} and u from T1 while attach the tree Tc to T1 by collapsing the two copies of v. This way, we obtain a spanning tree T of G, of weight w(T ) = w(T1 ) + w(Tc ) since w(v) is not counted towards w(Tc ). Conversely, for any spanning tree T of G, the vertex v is internal due to the existence of C. We may duplicate v and separate out a subtree Tc on the set of vertices V (C) ∪ v, in which the weight of v is revised to 0. This subtree Tc is thus a spanning tree of Gc , and every vertex of V (C) is internal in T if and only if it is internal in Tc . We attach the 0-weight vertex u to the vertex v in the remainder tree via the edge {v, u}, which is denoted as T1 and becomes a spanning tree of G1 ; note that the vertex v is internal in T1 . It follows that w(T ) = w(Tc ) + w(T1 ). ✷ See Figure 1 for an illustration of the local configurations specified in Operation 1. When |V (C)| = 2 and Gc is a triangle, we refer the configuration as a hanging triangle; when |V (C)| = 3 and Gc contains a length-4 cycle, we refer the configuration as a hanging diamond; when |V (C)| = 4 and Gc contains a length-5 cycle, we refer the configuration as a hanging pentagon. Applying Operation 1, we assume in the sequel that there is no hanging triangle, or hanging diamond, or hanging pentagon in the given graph G. v v v (a) A hanging triangle. (b) A hanging diamond. (c) A hanging pentagon. Figure 1: Local configurations of a hanging triangle, a hanging diamond, and a hanging pentagon, specified in Operation 1, where the dotted edges could be in or not in the graph. Let M ∗ denote a maximum weight matching of G, which is computed in O(n min{m log n, n2 })time, where n = |V | and m = |E|. Let the edges of M ∗ be e1 , e2 , . . . , ek ; let ej = {aj , bj }, such that w(aj ) ≥ w(bj ), for all j = 1, 2, . . . , k. For convenience, aj and bj are referred to as the head and the tail vertices of the edge ej , respectively. The same as in the last section, we assume without loss of generality that w(aj ) > 0 for each j, and consequently the degree of aj is dG (aj ) ≥ 2, that is, aj is adjacent to at least one vertex other than bj in the graph G. Let A = {aj | j = 1, 2, . . . , k}, B = {bj | j = 1, 2, . . . , k}, and X = V − (A ∪ B). Let E aa = E(A, A), i.e., the set of edges each connecting two vertices of A, E ax = E(A, X), i.e., the set of edges each connecting a vertex of A and a vertex of X, and E ab = E(A, B), i.e., the set of edges each connecting a vertex of A and a vertex of B, respectively. Approximating the maximum weight internal spanning tree 7 Let M aa ⊆ E aa be a maximum cardinality matching within the edge set E aa . We next prove a structure property of the spanning subgraph G[V, M ∗ ∪ M aa ], which has the edge set M ∗ ∪ M aa . For an edge ej = {aj , bj } of M ∗ , if aj is not incident to any edge of M aa , then ej is called isolated in G[V, M ∗ ∪ M aa ]. Lemma 3 Assume that two edges ej1 = {aj1 , bj1 } and ej2 = {aj2 , bj2 } of M ∗ are connected by the edge {aj1 , aj2 } ∈ M aa in G[V, M ∗ ∪ M aa ]. Then there is at most one isolated edge ej3 = {aj3 , bj3 } whose head aj3 can be adjacent to aj1 or aj2 . Proof. By contradiction, assume that there are two isolated edges ej3 = {aj3 , bj3 } and ej4 = {aj4 , bj4 } such that both the vertices aj3 and aj4 are adjacent to aj1 or aj2 . Then from the maximum cardinality of M aa , aj3 and aj4 must be both adjacent to aj1 or both adjacent to aj2 . Suppose they are both adjacent to aj1 ; from the claw-free property, at least two of aj2 , aj3 and aj4 are adjacent, which contradicts the maximum cardinality of M aa . This proves the lemma. ✷ For an isolated edge ej3 = {aj3 , bj3 } whose head is adjacent to an edge {aj1 , aj2 } ∈ M aa (i.e., satisfying Lemma 3), and assuming that {aj2 , aj3 } ∈ E aa , we add the edge {aj2 , aj3 } to G[V, M ∗ ∪ M aa ]; consequently the edge ej3 is no longer isolated. Let N aa denote the set of such added edges associated with M aa . At the end, the achieved subgraph is denoted as H0 = G[V, M ∗ ∪ M aa ∪ N aa ]. Lemma 4 In the subgraph H0 = G[V, M ∗ ∪ M aa ∪ N aa ], • every connected component containing more than one edge has either two or three edges from M ∗ , with their head vertices connected (by the edges of M aa ∪ N aa ) into a path; it is called a type-I component (see Figure 2a) and a type-II component (see Figure 2b), respectively; • for every isolated edge ej = {aj , bj }, the head vertex is incident with at least one edge of E ax ∪ E ab , but with no edge of E aa . Proof. The proof directly follows the definition of the subgraph H0 and Lemma 3. aj2 aj1 aj3 bj2 bj1 bj3 (a) A type-I component. aj2 bj2 ✷ aj1 bj1 (b) A type-II component. Figure 2: The configurations of a type-I component and a type-II component. The following lemma is analogous to Lemma 3. Lemma 5 Any vertex of X can be adjacent to the head vertices of at most two isolated edges in the subgraph H0 = G[V, M ∗ ∪ M aa ∪ N aa ]. 8 Chen et al. /v:May 30, 2017 Proof. By contradiction, assume that x ∈ X and there are three isolated edges ejk = {ajk , bjk }, k = 1, 2, 3, in the subgraph H0 = G[V, M ∗ ∪ M aa ∪ N aa ], such that the edge {ajk , x} ∈ E ax . From the claw-free property, at least two of aj1 , aj2 and aj3 are adjacent, which contradicts Lemma 4. This proves the lemma. ✷ For an isolated edge ej = {aj , bj } in the subgraph H0 = G[V, M ∗ ∪ M aa ∪ N aa ] whose head is adjacent to a vertex x ∈ X (i.e., satisfying Lemma 5), we add the edge {aj , x} to H0 ; consequently the edge ej is no longer isolated. Let N ax denote the set of such added edges associated with X. At the end, the achieved subgraph is denoted as H1 = G[V, M ∗ ∪ M aa ∪ N aa ∪ N ax ]. Lemma 6 In the subgraph H1 = G[V, M ∗ ∪ M aa ∪ N aa ∪ N ax ], • every connected component of H0 containing more than one edge remains unchanged in H1 ; • every connected component containing a vertex x of X and some other vertex has either one or two edges from M ∗ , with their head vertices connected (by the edges of N ax ) to the vertex x; it is called a type-III component (see Figure 3a) and a type-IV component (see Figure 3b), respectively; • for every isolated edge ej = {aj , bj }, the head vertex is incident with at least one edge of E ab , but with no edge of E aa ∪ E ax . Proof. The proof directly follows the definition of the subgraph H1 and Lemmas 4 and 5. aj1 aj1 aj2 x ✷ x bj1 (a) A type-III component. bj2 bj1 (b) A type-IV component. Figure 3: The configurations of a type-III component and a type-IV component. Let E0ab denote the subset of E ab , to include all the edges {aj , bℓ } where both the edges ej = {aj , bj } and eℓ = {aℓ , bℓ } are isolated in the subgraph H1 = G[V, M ∗ ∪ M aa ∪ N aa ∪ N ax ]. Let M ab ⊆ E0ab be a maximum cardinality matching within the edge set E0ab . Let H2 = G[V, M ∗ ∪ M aa ∪ N aa ∪ N ax ∪ M ab ] be the subgraph obtained from H1 by adding all the edges of M ab . One clearly sees that all the isolated edges in the subgraph H1 are connected by the edges of M ab into disjoint paths and cycles; while a path may contain any number of isolated edges, a cycle contains at least two isolated edges. Such a path and a cycle component are called a type-V component (see Figure 4a) and a type-VI component (see Figure 4b), respectively. Note that in a type-V component, there is exactly one head vertex of degree 1 and there is exactly one tail vertex of degree 1. We assume that for the tail vertex in a type-V component, it is not adjacent to the head of any other edge (via an edge of E ab ) in the same component; otherwise, through an edge exchange, the component is decomposed into a smaller type-V component and a new type-VI component. Approximating the maximum weight internal spanning tree ...... ...... (a) A type-V component. 9 (b) A type-VI component. Figure 4: The configurations of a type-V component and a type-VI component. Lemma 7 In the subgraph H2 = G[V, M ∗ ∪ M aa ∪ N aa ∪ N ax ∪ M ab ], for every type-V component, the degree-1 head vertex is adjacent (via an edge of E ab ) to the tail vertex of an edge in a type-I, -II, -III, or -IV component; on the other hand, the tail vertex of every edge in a type-I, -II, -III, or -IV component is adjacent to at most one such head vertex. Proof. We first show that the degree-1 head vertex in a type-V component C, denoted as aj , cannot be adjacent to the tail of any edge in another type-V or a type-VI component C ′ . By contradiction, assume {aj , bℓ } ∈ E ab and eℓ is in C ′ . If the tail bℓ is already incident to some edge of M ab , say {ai , bℓ }, then by the claw-free property at least two of ai , aj , aℓ must be adjacent, contradicting the fact that they are all isolated in the subgraph H1 . In the other case, the tail bℓ is the tail vertex of C ′ (which is a type-V component too), then it violates the maximum cardinality of M ab since {aj , bℓ } ∈ E ab can be added to increase the size of M ab . This proves the first half of the lemma. The second half can be proven by a simple contradiction using the claw-free property of the graph. ✷ Subsequently, every type-V component C is connected to a type-I, -II, -III, or -IV component C ′ , via the edge between the degree-1 head vertex of C and the tail vertex of an edge in C ′ ∩ M ∗ . This way, the degree-1 tail vertex of C takes up the role of “the tail vertex” of the edge in C ′ ∩ M ∗ , to become a tail vertex in the newly formed bigger component. For simplicity, the type of the component C ′ is passed to the newly formed bigger component. Denote this set of newly added edges as N ab , which is a subset of E ab − M ab . The achieved subgraph is denoted as H3 = G[V, M ∗ ∪ M aa ∪ N aa ∪ N ax ∪ M ab ∪ N ab ]. Lemma 8 In the subgraph H3 = G[V, M ∗ ∪ M aa ∪ N aa ∪ N ax ∪ M ab ∪ N ab ], • there is no isolated type-V component; • the head vertex of every edge of M ∗ has degree at least 2. Proof. The first half of the lemma follows from Lemma 7; the second half holds since there is no more isolated type-V component, which is the only type of component containing a degree-1 head vertex. ✷ We next create a set F of edges that are used to interconnect the components in the subgraph H3 . F is initialized to be empty. By Lemma 8, for every type-I, -II, -III, or -IV component C in the subgraph H3 , of weight w(C ∩M ∗ ), it is a tree and the total weight of the internal vertices therein is at least 12 w(C ∩ M ∗ ); for every type-VI component C, which is a cycle, by deleting the lightest edge 10 Chen et al. /v:May 30, 2017 of C ∩ M ∗ from C we obtain a path and the total weight of the internal vertices in this path is also at least 12 w(C ∩ M ∗ ). In the next five lemmas, we show that every component C in the subgraph H3 can be converted into a tree on the same set of vertices, possibly with one edge specified for connecting a leaf of this tree outwards, such that the total weight of the internal vertices (and the leaf, if specified) in the tree is at least 32 w(C ∩ M ∗ ). The specified edge for the interconnection purpose, is added to F . At the end of the process, the component C is called settled. A settled component C can be expressed in multiple equivalent ways, for example, that the total weight of the internal vertices (and the leaf, if specified) in the resulting tree is at least 32 w(C ∩ M ∗ ), or that the total weight of the internal (and the leaf, if specified) vertices in the resulting tree is at least twice the total weight of the leaves (excluding the specified leaf, if any). In the sequel, we abuse the vertex notation to also denote its weight in math formulae; this simplifies the presentation and the meaning of the notation is easily distinguishable. For estimating the total weight of the internal vertices in a tree in the sequel, we frequently use the following inequality: ∀w1 , w2 , w3 ∈ R, w1 + w2 + w3 − min{w1 , w2 , w3 } ≥ 2 min{w1 , w2 , w3 }. Lemma 9 A type-I component in the subgraph H3 can be settled. Proof. Consider a type-I component C in the subgraph H3 . Recall from Lemma 8 and the paragraph right above it, that a general type-I component is an original type-I component (shown in Figure 2a) augmented with zero to two type-V components. Let the two original edges of M ∗ in C be ej1 and ej2 and the two tail vertices be b1 and b2 (which replace bj1 and bj2 to be the tail vertices, respectively) with w(b1 ) ≥ w(b2 ). The corresponding two head vertices to the tails b1 and b2 are denoted as a1 and a2 , respectively. See Figure 5 for the general configuration of such a component. We assume that w(b1 ) > 0, since otherwise C is settled automatically. a2 b2 aj2 bj2 aj1 a1 bj1 b1 Figure 5: The general configuration of a type-I component, with a type-V component adjacent to each of the two tail vertices. Case 1. If b1 is adjacent to a vertex v outside C, then we add the edge {b1 , v} to F ; this settles C, since the total weight of the internal vertices in C is at least a1 + b1 + a2 ≥ 3b2 (recall that a vertex notation here represents the weight of the vertex). We next consider the case where b1 is not adjacent to any vertex outside C, and thus it has to be adjacent to some vertex inside C. Note that C is a path with b1 and b2 being its two ending vertices. Let v denote the vertex adjacent to b1 that is the farthest to b1 on C. We distinguish this distance dC (b1 , v) ≥ 2 and where v locates. Approximating the maximum weight internal spanning tree 11 Case 2. If b1 6= bj1 and v is in the type-V component containing b1 , then by the construction of a type-V component we know that v must be a tail of an edge of M ∗ (Figures 6a and 6b), and thus dC (b1 , v) is even. Denote this edge as ej3 = {aj3 , bj3 }, that is v = bj3 . a2 b2 aj2 bj2 aj1 a1 a2 bj1 v b1 b2 (a) dC (b1 , v) ≥ 4. aj2 bj2 a1 aj1 bj1 v b1 (b) dC (b1 , v) = 2. Figure 6: Local configurations corresponding to Case 2, where v is inside the type-V component adjacent to bj1 . Case 2.1. If dC (b1 , v) ≥ 4, then denote the head vertex other than aj3 that bj3 is adjacent to as aj4 (see Figure 6a). We conclude from the claw-free property that there must be at least an edge among aj3 , aj4 , b1 , which contradicts the identity of the type-V component. Therefore, it is impossible to have dC (b1 , v) ≥ 4. Case 2.2. If dC (b1 , v) = 2, then we conclude that dG (b1 ) = 2 and thus dG (a1 ) ≥ 3 by Operation 1 (see Figure 6b), i.e. there is at least another edge incident at a1 besides {a1 , bj3 } and {a1 , b1 }. Denote this neighbor of a1 as u. If u is inside C, then u = b2 ; in this case, add the edges {b1 , bj3 } and {a1 , b2 } to C while delete the edge {a1 , bj3 } and the lightest among the edges of C ∩ M ∗ from C. This way, the component becomes a tree and thus C is settled. If u is outside C, then we add the edge {b1 , bj3 } to C while delete the edge {a1 , bj3 } from C, and add the edge {a1 , u} to F ; this way, the component becomes a tree and thus C is settled. Case 3. If b1 6= bj1 and v = bj1 , we consider the size of the type-V component containing b1 (see Figure 7). a2 b2 aj2 bj2 aj1 a1 bj1 b1 Figure 7: Local configurations corresponding to Case 3, where v = bj1 . Case 3.1. If this type-V component contains more than one edge of M ∗ , then by the claw-free property aj1 must be adjacent to b1 , which violates the definition of v being the farthest and thus it is impossible. Case 3.2. If the type-V component containing b1 has only one edge of M ∗ , which is {a1 , b1 }, then we have dG (b1 ) = 2, and thus dG (a1 ) ≥ 3 by Operation 1, i.e. there is at least another edge incident at a1 besides {a1 , bj1 } and {a1 , b1 }. (The following is the same as in Case 2.2.) Denote this neighbor of a1 as u. If u is inside C, then u = b2 ; in this case, add the edges {b1 , bj1 } and 12 Chen et al. /v:May 30, 2017 {a1 , b2 } to C while delete the edge {a1 , bj1 } and the lightest among the edges of C ∩ M ∗ from C. This way, the component becomes a tree and thus C is settled. If u is outside C, then we add the edge {b1 , bj1 } to C while delete the edge {a1 , bj1 } from C, and add the edge {a1 , u} to F ; this way, the component becomes a tree and thus C is settled. Case 4. If b1 6= bj1 and v = aj1 (see Figure 8), then we leave C as it is when w(b1 ) ≤ w(bj1 ), or we add the edge {aj1 , b1 } to C while delete the edge {aj1 , bj1 } from C. The total weight of the internal vertices in the resulting path is at least a2 + aj1 + a1 + max{b1 , bj1 } ≥ 2(b2 + min{b1 , bj1 }), and thus C is settled. a2 b2 aj2 bj2 aj1 a1 bj1 b1 Figure 8: Local configurations corresponding to Case 4, where v = aj1 . Case 5. If b1 6= bj1 and v = aj2 (see Figure 9), then by the claw-free property there is at least an edge among bj2 , aj1 , b1 . Note that bj2 and b1 cannot be adjacent due to the definition of the vertex v (being the farthest). If aj1 and b1 are adjacent, then it has been proven in Case 4 that C can be settled. If bj2 and aj1 are adjacent, then similarly as in Case 4 we either leave C as it is when w(b1 ) ≤ w(bj1 ), or add the edges {aj2 , b1 } and {aj1 , bj2 } to C while delete the edges {aj1 , bj1 } and {aj2 , bj2 } from C; the total weight of the internal vertices in the resulting path is at least a2 + aj1 + a1 + max{b1 , bj1 } ≥ 2(b2 + min{b1 , bj1 }), and thus C is settled. a2 b2 aj2 bj2 aj1 a1 bj1 b1 Figure 9: Local configurations corresponding to Case 5, where v = aj2 . Case 6. If b1 6= bj1 and v = bj2 (see Figure 10), then by the claw-free property aj2 and b1 must be adjacent. Case 6.1. If b2 6= bj2 , then we have three ways to convert C into a path ending at b2 : 1) doing nothing to leave b1 as a leaf; 2) adding the edge {aj2 , b1 } to C while deleting the edge {aj2 , aj1 } to leave aj1 as a leaf; 3) adding the edge {bj2 , b1 } to C while deleting the edge {aj2 , bj2 } to leave aj2 as a leaf. Then, the maximum total weight of the internal vertices among these three paths is at least a2 + a1 + aj1 + aj2 + b1 − min{b1 , aj1 , aj2 } ≥ 2(b2 + min{b1 , aj1 , aj2 }). Thus, C is settled. Approximating the maximum weight internal spanning tree a2 aj2 b2 bj2 aj1 a1 bj1 b1 13 Figure 10: Local configurations corresponding to Case 6, where v = bj2 . Case 6.2. If b2 = bj2 , that is, there is no type-V component adjacent to bj2 , then we add the edge {bj2 , b1 } to C while delete the lightest edge of C ∩ M ∗ from C to settle C, because C ∩ M ∗ contains at least three edges. Case 7. If b1 6= bj1 , b2 6= bj2 , and v is in the type-V component containing b2 , we distinguish whether v is a head or a tail (see Figure 11). a2 2 a v aj2 aj1 b bj2 bj1 (a) v is a head. aj1 a1 a2 bj1 b1 b2 aj2 aj1 a1 bj1 b1 a b2 2 aj2 1 b bj2 v bj2 1 (b) v = b2 . (c) v is a tail. Figure 11: Local configurations corresponding to Case 7, where v is inside the type-V component adjacent to bj2 . Case 7.1. If v is a head, say aj3 of the edge ej3 (see Figure 11a), and assume that aj3 is also adjacent to bj4 , then we do nothing to C to leave b1 as a leaf when w(b1 ) ≤ w(bj4 ), or add the edge {aj3 , b1 } to C while delete the edge {aj3 , bj4 } from C to leave bj4 as a leaf. Then, the total weight of the internal vertices in the resulting path is at least a2 +a1 +aj4 +max{b1 , bj4 } ≥ 2(b2 +min{b1 , bj4 }). Thus, C is settled. Case 7.2. If v is a tail, say bj3 of the edge ej3 (see Figures 11b and 11c). If b2 = bj3 (see Figure 11b), then we add the edge {b1 , b2 } to C makes it a cycle while delete the lightest edge of C ∩ M ∗ to settle C since there are at least four edges in C ∩ M ∗ ; if b2 6= bj3 (see Figure 11c), then by the claw-free property and the definition of v we conclude that b1 is adjacent to aj3 too, and thus the argument in Case 7.1 applies to settle C. In summary, Cases 2-7 together prove that when b1 6= bj1 , the component C can be settled. We next consider the situation where b1 = bj1 , that is, there is no type-V component adjacent to bj1 (see Figure 12). Case 8. If b1 = bj1 and v = aj2 , then we conclude that dG (b1 ) = 2 and thus dG (a1 ) ≥ 3 by Operation 1 (see Figure 13). Case 8.1. If there is a vertex u outside C that is adjacent to a1 , then we add the edge {a1 , u} to F , add the edge {aj2 , b1 } to C while delete the edge {a1 , aj2 } from C; this way, C is settled. Case 8.2. Otherwise by the claw-free property a1 must be adjacent either to b2 or to bj2 . 14 Chen et al. /v:May 30, 2017 a2 a1 aj2 b2 b1 bj2 Figure 12: A special configuration without a type-V component adjacent to bj1 , i.e. bj1 = b1 . a2 a1 aj2 b2 b1 bj2 Figure 13: Local configurations corresponding to Case 8, where v = aj2 . Case 8.2.1. Assuming b2 6= bj2 , in the former case, we add the edges {aj2 , b1 } and {a1 , b2 } to C while delete the edge {a1 , aj2 } and the lightest edge of C ∩ M ∗ from C; this way, C is settled since there were at least three edges in C ∩ M ∗ . In the latter case, we conclude that dG (aj2 ) ≥ 4 by Operation 1. Recursively, if there is a vertex u outside C that is adjacent to aj2 , then we add the edge {aj2 , u} to F , add the edges {aj2 , b1 } and {a1 , bj2 } to C while delete the edges {a1 , aj2 } and {aj2 , bj2 } from C; this way, C is settled. Otherwise by the claw-free property aj2 must be adjacent to b2 ; we add three edges {aj2 , b1 }, {a1 , bj2 } and {aj2 , b2 } to C while delete the edges {a1 , aj2 } and {aj2 , bj2 }, and the lightest edge of C ∩ M ∗ from C; this way, C is settled. Case 8.2.2. Assuming b2 = bj2 , that is, there is no type-V component adjacent to bj2 . We conclude that dG (a2 ) ≥ 4 and there is a vertex u outside C that is adjacent to a2 by Operation 1. Thus, we add the edge {a2 , u} to F , add the edges {a2 , b1 } and {a1 , b2 } to C while delete the edges {a1 , a2 } and {a2 , b2 } from C; this way, C is settled. Case 9. If b1 = bj1 and v = bj2 (see Figure 14), we consider two possible scenarios. a2 aj2 a1 a2 a1 b2 bj2 b1 b2 b1 (a) b2 6= bj2 . (b) b2 = bj2 . Figure 14: Local configurations corresponding to Case 9, where v = bj2 . Approximating the maximum weight internal spanning tree 15 Case 9.1. If b2 6= bj2 (see Figure 14a), then we conclude from the claw-free property and the definition of v that b1 is also adjacent to aj2 . By Operation 1, at least one of a1 and aj2 must be adjacent to another vertex u. Case 9.1.1. If there is a vertex u outside C that is adjacent to a1 (aj2 , respectively) then we add the edge {a1 , u} ({aj2 , u}, respectively) to F , add the edges {bj2 , b1 } and {aj2 , b1 } ({bj2 , b1 }, respectively) to C while delete the edges {a1 , b1 } and {aj2 , bj2 } ({aj2 , bj2 }, respectively) from C; this way, C is settled. Case 9.1.2. Otherwise by the claw-free property u ∈ {b2 , bj2 }. If u = b2 , then we can settle C by converting C into a cycle on the same set of vertices, followed by deleting the lightest edge of C ∩ M ∗ from C. In the other case, u = bj2 and thus a1 is adjacent to bj2 , which by Operation 1 is impossible. Case 9.2. If b2 = bj2 , that is, there is no type-V component adjacent to bj2 (see Figure 14b). We conclude that either there is a vertex u outside C that is adjacent to a2 , or a2 is adjacent to b1 and there is a vertex u outside C that is adjacent to a1 . In either case, we add the edge {·, u} to F , and convert C into a path with b2 and · as two ending vertices; this way, C is settled. Case 10. If b1 = bj1 , b2 6= bj2 , and v is in the type-V component containing b2 , we distinguish whether v is a head or a tail (see Figure 15). a2 a2 b2 v aj2 a1 bj2 b1 (a) v is a head. b2 aj2 bj2 a1 a2 b1 b2 (b) v = b2 . aj2 v bj2 a1 b1 (c) v is a tail. Figure 15: Local configurations corresponding to Case 10, where v is inside the type-V component adjacent to bj2 . Case 10.1. If v is a head, say aj3 of the edge ej3 (see Figure 15a), and assume that aj3 is also adjacent to bj4 , then we do nothing to C to leave b1 as a leaf when w(b1 ) ≤ w(bj4 ), or add the edge {aj3 , b1 } to C while delete the edge {aj3 , bj4 } from C to leave bj4 as a leaf. Then, the total weight of the internal vertices in the resulting path is at least a2 +a1 +aj4 +max{b1 , bj4 } ≥ 2(b2 +min{b1 , bj4 }). Thus, C is settled. Case 10.2. If v is a tail, say bj3 of the edge ej3 (see Figures 15b and 15c). If b2 = bj3 (see Figure 15b), then we add the edge {b1 , b2 } to C makes it a cycle while delete the lightest edge of C ∩ M ∗ to settle C since there are at least three edges in C ∩ M ∗ ; if b2 6= bj3 (see Figure 15c), then by the claw-free property and the definition of v we conclude that b1 is adjacent to aj3 too, and thus the argument in Case 10.1 applies to settle C. All possible cases have be discussed in the above. The lemma is proven. ✷ Lemma 10 A type-II component in the subgraph H3 can be settled. Proof. Consider a type-II component C in the subgraph H3 . Recall from Lemma 8 and the paragraph right above it, that a general type-II component is an original type-II component (shown 16 Chen et al. /v:May 30, 2017 in Figure 2b) augmented with zero to three type-V components. Let the three original edges of M ∗ in C be ej1 , ej2 , ej3 , with ej2 in the middle (see Figure 2b), and the three tail vertices be b1 , b2 , b3 (which replace bj1 , bj2 , bj3 , respectively). We consider the situation where b1 is the heaviest among the three tail vertices (the other situation is where b2 is the heaviest, and can be similarly discussed). In the following, discussion for most cases is similar to the cases in the proof of Lemma 9, and thus not all details are presented (neither the illustration figures). Case 1. If b1 is adjacent to a vertex v outside C, then we add the edge {b1 , v} to F , which settles C, since the total weight of the internal vertices in C is at least a1 + b1 + a2 + a3 ≥ 2(b2 + b3 ) (recall that a vertex notation here represents the weight of the vertex). In the sequel we assume b1 is not adjacent to any vertex outside C, and thus it has to be adjacent to some vertex inside C. Let v denote the vertex adjacent to b1 that is the farthest to b1 on C (tie breaks arbitrarily). We distinguish this distance dC (b1 , v) ≥ 2 and where v locates. Case 2. If b1 6= bj1 and v is in the type-V component containing b1 , then v must be the tail of an edge of M ∗ and thus dC (b1 , v) is even. Denote this edge as ej4 = {aj4 , bj4 }, that is v = bj4 . Case 2.1. If dC (b1 , v) ≥ 4, then denote the head vertex other than aj4 that bj4 is also adjacent to as aj5 . We conclude from the claw-free property that there must be at least an edge among aj4 , aj5 , b1 , which contradicts the identity of the type-V component. Thus it is impossible to have dC (b1 , v) ≥ 4. Case 2.2. If dC (b1 , v) = 2, then we conclude that dG (b1 ) = 2 and thus dG (a1 ) ≥ 3 by Operation 1, i.e. there is at least another edge incident at a1 besides {a1 , bj4 } and {a1 , b1 }. Denote this neighbor of a1 as u. If u is inside C, then u ∈ {b2 , b3 }; in the case of u = b2 (the argument for u = b3 is identical), we know that the branch incident at aj1 contains at least three edges of M ∗ (ej1 , ej4 , e1 ), and thus we may add the edges {bj4 , b1 } and {a1 , b2 } to C while delete the edge {a1 , bj4 } and the lightest among the three edges ej1 , ej4 , e2 (which are on the created cycle), denoted as ex = {ax , bx }, from C. This way, the component becomes a tree with leaves b3 , ax , bx . The total weight of the internal vertices in the tree is at least a3 + (a1 + b1 ) + 2(ax + bx ) ≥ 2(b3 + ax + bx ); that is, C is settled. If u is outside C, then we add the edge {bj4 , b1 } to C while delete the edge {a1 , bj4 } from C, and add the edge {a1 , u} to F ; this way, the component becomes a tree and thus C is settled. Case 3. If b1 6= bj1 and v = bj1 , we consider the size of the type-V component containing b1 . Case 3.1. If this type-V component contains more than one edge of M ∗ , then by the claw-free property aj1 must be adjacent to b1 , which violates the definition of v being the farthest and thus it is impossible. Case 3.2. If the type-V component containing b1 has only one edge of M ∗ , which is {a1 , b1 }; in this case, we have dG (b1 ) = 2, and thus dG (a1 ) ≥ 3 by Operation 1, i.e. there is at least another edge incident at a1 besides {a1 , bj1 } and {a1 , b1 }. Denote this neighbor of a1 as u. If u is inside C, then u ∈ {b2 , b3 }. In the case of u = b2 (the argument for u = b3 is identical), we may add the edges {b1 , bj1 } and {a1 , b2 } to C while delete the edge {a1 , bj1 } to form a cycle. When bj2 = b2 , we either delete the edge {aj2 , bj2 } to leave bj2 as a leaf, or delete the edge {aj2 , aj1 } to leave aj1 as a leaf. The maximum total weight of the internal vertices between the two trees is at least a3 + a1 + b1 + a2 + max{b2 , aj1 } ≥ 2(b3 + min{b2 , aj1 }), and thus C is settled. When bj2 6= b2 , we delete the lightest among the three edges ej1 , ej2 , e2 (which are on the created cycle), denoted as ex = {ax , bx }, from C. This way, the component becomes a tree with leaves b3 , ax , bx . The total Approximating the maximum weight internal spanning tree 17 weight of the internal vertices in the tree is at least a3 + (a1 + b1 ) + 2(ax + bx ) ≥ 2(b3 + ax + bx ); that is, C is settled. If u is outside C, then we add the edge {b1 , bj1 } to C while delete the edge {a1 , bj1 } from C, and add the edge {a1 , u} to F ; this way, the component becomes a tree and thus C is settled. Case 4. If b1 6= bj1 and v = aj1 , we consider the size of the type-V component containing b1 . Case 4.1. If this type-V component contains more than one edge of M ∗ , then denote one edge other than {a1 , b1 } as ej4 = {aj4 , bj4 }. We add the edge {aj1 , b1 } to C while either delete the edge {aj1 , bj1 } to have a tree with leaves bj1 , b2 , b3 , or delete the edge {aj4 , bj4 } to have a tree with leaves aj4 , bj4 , b2 , b3 . It follows that the maximum total weight of the internal vertices between the two trees is at least a2 + a3 + a1 + b1 + aj1 + max{bj1 , aj4 + bj4 } ≥ 2(b2 + b3 + min{bj1 , aj4 + bj4 }); therefore, C is settled. Case 4.2. If this type-V component contains only one edge of M ∗ , which is {a1 , b1 }, then from the claw-free property and the definition of v being the farthest we conclude that bj1 is adjacent to at least one of aj2 and b1 . Case 4.2.1. Assume bj1 and aj2 are adjacent. By treating aj2 as the cut-vertex in Operation 1, we conclude that at least one of the three vertices aj1 , bj1 , a1 is adjacent to a vertex u outside of the set {aj2 , aj1 , bj1 , a1 , b1 }. Case 4.2.1.1. If aj1 is adjacent to a vertex u outside the component C, then we add the edges {aj2 , bj1 } and {aj1 , b1 } to C while delete the edges {aj2 , aj1 } and {aj1 , bj1 } from C to obtain a tree with leaves b3 , b2 , aj1 ; thus, adding the edge {aj1 , u} to F settles C. If bj1 is adjacent to a vertex u outside the component C, then we add the edge {aj1 , b1 } to C while delete the edge {aj1 , bj1 } from C to obtain a tree with leaves b3 , b2 , bj1 ; thus, adding the edge {bj1 , u} to F settles C. If a1 is adjacent to a vertex u outside the component C, then we add the edges {aj2 , bj1 } and {aj1 , b1 } to C while delete the edges {aj2 , aj1 } and {bj1 , a1 } from C to obtain a tree with leaves b3 , b2 , a1 ; thus, adding the edge {a1 , u} to F settles C. Case 4.2.1.2. If aj1 is not adjacent to any vertex outside the component C, but to some vertices inside C, then let u denote the farthest neighbor on C (tie breaks arbitrarily). There are only four possibilities. When u = aj3 (u = bj3 and bj3 6= b3 implying that aj3 and aj1 are adjacent), we can add the edges {aj3 , aj1 }, {aj2 , bj1 } and {aj1 , b1 } to C while delete the edges {aj3 , aj2 }, {aj2 , aj1 } and {aj1 , bj1 } from C to obtain a path with leaves b3 , b2 ; thus, C is settled. When u = bj2 , we can add the edges {bj2 , aj1 }, {aj2 , bj1 } and {aj1 , b1 } to C while delete the edges {aj2 , aj1 }, {aj2 , bj2 } and {aj1 , bj1 } from C to obtain a path with leaves b3 , b2 ; thus, C is settled. When u = b3 (whether b3 = bj3 or not), we can add the edge {b3 , aj1 } to C while delete the edge {aj2 , aj1 } from C to obtain a path with leaves b2 , b1 ; we also can add the edges {b3 , aj1 } and {aj1 , b1 } to C while delete the edges {aj2 , aj1 } and {aj1 , bj1 } from C to obtain a path with leaves b2 , bj1 . It follows that the maximum total weight of the internal vertices between the two paths is at least a2 + aj1 + a1 + max{bj1 , b1 } ≥ 2(b2 + min{bj1 , b1 }); therefore, C is settled. When u = b2 , it can be shown the same as in the last paragraph by replacing b3 with b2 , that C can be settled. Case 4.2.1.3. If a1 is not adjacent to any vertex outside the component C, but to some vertices inside C, then let u denote the farthest neighbor on C (tie breaks arbitrarily). Note that u cannot be any head vertex, as otherwise it violates the algorithm; u cannot be any tail vertex either, unless it is b2 or b3 . Therefore there are only two possibilities. 18 Chen et al. /v:May 30, 2017 When u = b3 (whether b3 = bj3 or not), we can add the edges {b3 , a1 } and {aj1 , b1 } to C while delete the edge {aj2 , aj1 }. Then, we either delete {a1 , b1 } to obtain a path with leaves b2 , b1 , or delete {a1 , bj1 } to obtain a path with leaves b2 , bj1 . It follows that the maximum total weight of the internal vertices between the two paths is at least a2 + aj1 + a1 + max{bj1 , b1 } ≥ 2(b2 + min{bj1 , b1 }); therefore, C is settled. When u = b2 , it can be shown the same as in the last paragraph by replacing b3 with b2 , that C can be settled. Case 4.2.1.4. If bj1 is not adjacent to any vertex outside the component C, but to some vertices inside C, then let u denote the farthest neighbor on C (tie breaks arbitrarily). When u ∈ {aj3 , bj3 , bj2 } (u = bj3 and bj3 6= b3 implying that aj3 and bj1 are adjacent), similarly as in Case 4.2.1.2, we can convert C into a path with leaves b3 , b2 to settle C. In the remaining case, u is inside one of the attached type-V components. When u is a head, say aj4 of the edge ej4 in the type-V component attached to bj2 (or bj3 ), and assume that aj4 is adjacent to bj5 besides bj4 , then we can do nothing to have a tree with leaves b3 , b2 , b1 ; or we can add the edge {aj1 , b1 } to C while delete the edge {aj1 , bj1 } to obtain a tree with leaves b3 , b2 , bj1 ; or we can add the edges {aj4 , bj1 } and {aj1 , b1 } to C while delete the edges {aj4 , bj5 } and {aj1 , bj1 } to obtain a tree with leaves b3 , b2 , bj5 . The maximum total weight of the internal vertices among the three trees is at least a3 + a2 + aj1 + bj1 + aj5 + bj5 + a1 + b1 − min{b1 , bj1 , bj5 } ≥ 2(b3 + b2 + min{b1 , bj1 , bj5 }), which settles C. When u is a tail, say bj4 of the edge ej4 in the type-V component attached to bj2 (or bj3 ), but bj 4 ∈ / {b2 , b3 }, then from the claw-free property and the definition of u we conclude that bj1 is also adjacent to aj4 . Thus the argument in the last paragraph applies to settle C. Lastly, when u = b2 (or b3 , which can be shown in the same way), we conclude that b2 is adjacent either to aj1 , which is settled in Case 4.2.1.2, or to a1 , which is settled in Case 4.2.1.3. Case 4.2.2. Assume bj1 is not adjacent to aj2 but to b1 . By treating aj1 as the cut-vertex in Operation 1, we conclude that at least one of the two vertices bj1 , a1 is adjacent to a vertex u outside the set {aj1 , bj1 , a1 , b1 }. Case 4.2.2.1. If bj1 is adjacent to a vertex u outside the component C, then we add the edge {aj1 , b1 } to C while delete the edge {aj1 , bj1 } from C to obtain a tree with leaves b3 , b2 , bj1 ; thus, adding the edge {bj1 , u} to F settles C. If a1 is adjacent to a vertex u outside the component C, then we add the edges {bj1 , b1 } and {aj1 , b1 } to C while delete the edges {aj1 , bj1 } and {a1 , b1 } from C to obtain a tree with leaves b3 , b2 , a1 ; thus, adding the edge {a1 , u} to F settles C. Case 4.2.2.2. If a1 is not adjacent to any vertex outside the component C, but to some vertices inside C, then let u denote the farthest neighbor on C (tie breaks arbitrarily). Note that u cannot be any head vertex, as otherwise it violates the algorithm; u cannot be any tail vertex either, unless it is b2 or b3 . Therefore there are only two possibilities. (This is very similar to Case 4.2.1.3.) When u = b3 (whether b3 = bj3 or not), we can add the edges {b3 , a1 } and {aj1 , b1 } to C while delete the edge {aj2 , aj1 }. Then, we can delete {a1 , b1 } to obtain a path with leaves b2 , b1 , or we can delete {a1 , bj1 } to obtain a path with leaves b2 , bj1 . It follows that the maximum total weight of the internal vertices between the two paths is at least a2 + aj1 + a1 + max{bj1 , b1 } ≥ 2(b2 + min{bj1 , b1 }); therefore, C is settled. When u = b2 , it can be shown the same as in the last paragraph by replacing b3 with b2 , that C can be settled. Case 4.2.2.3. If bj1 is not adjacent to any vertex outside the component C, but to some vertices inside C, then let u denote the farthest neighbor on C (tie breaks arbitrarily). Note that u 6= aj2 , which is Case 4.2.1. When u ∈ {aj3 , bj3 , bj2 }, similarly as in Case 4.2.1.2, we can convert Approximating the maximum weight internal spanning tree 19 C into a path with leaves b3 , b2 to settle C. In the remaining case, u is inside one of the attached type-V components. (This is very similar to Case 4.2.1.4.) When u is a head, say aj4 of the edge ej4 in the type-V component attached to bj2 (or bj3 ), and assume that aj4 is adjacent to bj5 besides bj4 , then we can do nothing to have a tree with leaves b3 , b2 , b1 ; or we can add the edge {aj1 , b1 } to C while delete the edge {aj1 , bj1 } to obtain a tree with leaves b3 , b2 , bj1 ; or we can add the edges {aj4 , bj1 } and {aj1 , b1 } to C while delete the edges {aj4 , bj5 } and {aj1 , bj1 } to obtain a tree with leaves b3 , b2 , bj5 . The maximum total weight of the internal vertices among the three trees is at least a3 + a2 + aj1 + bj1 + aj5 + bj5 + a1 + b1 − min{b1 , bj1 , bj5 } ≥ 2(b3 + b2 + min{b1 , bj1 , bj5 }), which settles C. When u is a tail, say bj4 of the edge ej4 in the type-V component attached to bj2 (or bj3 ), but bj 4 ∈ / {b2 , b3 }, then from the claw-free property and the definition of u we conclude that bj1 is also adjacent to aj4 . Thus the argument in the last paragraph applies to settle C. Lastly, when u = b2 (or b3 , which can be shown in the same way), we conclude that b2 is adjacent either to aj1 or to a1 , the latter of which is settled in Case 4.2.2.3. In the remaining case where b2 is adjacent to aj1 , we can add the edge {b2 , aj1 } to C while delete the edge {aj1 , aj2 } to obtain a path with leaves b3 , b1 ; or we can add the edges {b2 , aj1 } and {b1 , aj1 } to C while delete the edges {aj1 , aj2 } and {aj1 , bj1 } to obtain a path with leaves b3 , bj1 . The maximum total weight of the internal vertices between the two trees is at least a3 + a1 + aj1 + max{b1 , bj1 } ≥ 2(b3 + min{b1 , bj1 }), which settles C. Case 5. If b1 6= bj1 and v = aj2 , then by the claw-free property there is at least an edge among bj2 , aj1 , b1 . Note that bj2 and b1 cannot be adjacent due to the definition of the vertex v being the farthest. If aj1 and b1 are adjacent, then it has been proven in the above Case 4 that C can be settled. If bj2 and aj1 are adjacent, then we add the edges {aj2 , b1 } and {aj1 , bj2 } to C while delete the edges {aj2 , aj1 } and {aj2 , bj2 } from C; this way we obtain a path with two leaves b2 and b3 , and thus it settles C. Case 6. If b1 6= bj1 and v = bj2 , then we add the edge {bj2 , b1 } to C while delete the edge {aj2 , bj2 } from C; this way we obtain a path with two leaves b2 and b3 , and thus it settles C. If b1 6= bj1 and v = aj3 , then we add the edge {aj3 , b1 } to C while delete the edge {aj3 , aj2 } from C; this way we obtain a path with two leaves b2 and b3 , and thus it settles C. If b1 6= bj1 and v = bj3 , then there are two possible scenarios. When b3 6= bj3 , by the clawfree property and the definition of v we conclude that aj3 and b1 must be adjacent, and the last paragraph shows that C is settled. When b3 = bj3 , we add the edge {b3 , b1 } to C while either delete the edge {a3 , aj2 } from C to achieve a path with two leaves a3 , b2 , or delete the edge {aj2 , aj1 } from C to achieve a path with leaves aj1 , b2 ; we may also do nothing to C which is a tree with leaves b1 , b2 , b3 . Among these three trees, the maximum total weight of the internal vertices is at least a3 + b3 + a2 + aj1 + a1 + b1 − min{a3 , aj1 , b1 + b3 } ≥ 2(b2 + min{a3 , aj1 , b1 + b3 }); thus, C is settled. Case 7. If b1 6= bj1 , b2 6= bj2 , and v is in the type-V component containing b2 , we distinguish whether v is a head or a tail. (Note that the same argument applies to b3 6= bj3 and v is in the type-V component containing b3 .) Case 7.1. If v is a head, say aj4 of the edge ej4 , and assume that aj4 is adjacent to bj5 besides bj4 , then we consider two distinct scenarios. Case 7.1.1. When bj5 6= bj2 , besides two leaves b2 and b3 , we either do nothing to C to leave 1 b as a leaf, or add the edge {aj4 , b1 } to C while delete the edge {aj4 , bj5 } from C to leave bj5 as a leaf, or add the edge {aj4 , b1 } to C while delete the edge {aj2 , bj2 } from C to leave bj2 as 20 Chen et al. /v:May 30, 2017 a leaf. Among these three trees, the maximum total weight of the internal vertices is at least a3 + a2 + a1 + b1 + aj2 + bj2 + aj5 + bj5 − min{b1 , bj2 , bj5 } ≥ 2(b3 + b2 + min{b1 , bj2 , bj5 }); thus, C is settled. Case 7.1.2. When bj5 = bj2 , by the claw-free property and the definition of the vertex v we conclude that bj2 is adjacent to at least one of bj4 and b1 . If bj2 and bj4 are adjacent, then we add the edges {aj4 , b1 } and {bj2 , bj4 } to C while delete the edges {aj2 , bj2 } and {aj4 , bj4 } from C to obtain a path with leaves b2 and b3 ; if bj2 and b1 are adjacent, then we add the edge {bj2 , b1 } to C while delete the edge {aj2 , bj2 } from C to obtain a path with leaves b2 and b3 . Thus, C is settled. Case 7.2. In the other case v is a tail, say bj4 of the edge ej4 . Case 7.2.1. If b2 = bj4 , then besides the leaf b3 , we either do nothing to C to leave b1 , b2 as leaves, or add the edge {b1 , b2 } to C and delete the edge {aj2 , aj1 } to leave aj1 as a leaf, or add the edge {b1 , b2 } to C and delete the edge {aj2 , bj2 } to leave bj2 as a leaf. Among these three trees, the maximum total weight of the internal vertices is at least a3 + a1 + aj1 + bj2 + b1 + b2 − min{b1 + b2 , aj1 , bj2 } ≥ 2(b3 + min{b1 + b2 , aj1 , bj2 }); thus, C is settled. (In this paragraph, we deal with the case where b3 takes the role of b2 and prove our claim at the beginning of Case 7 “that the same argument applies to b3 6= bj3 and v is in the type-V component containing b3 ”. The complete assumption of Case 7.2.1 is thus b3 6= bj3 and v = b3 . Then, besides the leaf b2 , we either do nothing to C to leave b1 , b3 as leaves, or add the edge {b1 , b3 } to C and delete the edge {aj2 , aj1 } to leave aj1 as a leaf, or add the edge {b1 , b3 } to C and delete the edge {aj2 , aj3 } to leave aj3 as a leaf. Among these three trees, the maximum total weight of the internal vertices is at least a2 + a1 + aj1 + aj3 + b1 + b3 − min{b1 + b3 , aj1 , aj3 } ≥ 2(b2 + min{b1 + b3 , aj1 , aj3 }); thus, C is settled. In summary, here the vertex aj3 takes up the role of bj2 correspondingly.) Case 7.2.2. If b2 6= bj4 , then by the claw-free property and the definition of v we conclude that b1 is adjacent to aj4 too. Assume that aj4 is adjacent to bj5 besides bj4 . When bj5 6= bj2 , the argument in Case 7.1.1 can be applied to settle C; when bj5 = bj2 (the argument in Case 7.1.2 does not applied to settle C due to the deferent v), besides the leaves b2 and b3 , we either do nothing to C to leave b1 as a leaf, or add the edge {b1 , aj4 } to C and delete the edge {aj2 , bj2 } to leave bj2 as a leaf, or add the edge {b1 , aj4 } to C and delete the edge {aj2 , aj1 } to leave aj1 as a leaf. Among these three trees, the maximum total weight of the internal vertices is at least a3 + a2 + aj1 + aj2 + bj2 + a1 + b1 − min{b1 , aj1 , bj2 } ≥ 2(b3 + b2 + min{b1 , aj1 , bj2 }); thus, C is settled. This finishes the discussion on Case 7. In summary, Cases 2-7 together prove that when b1 6= bj1 , the component C can be settled. We next consider the situation where b1 = bj1 , that is, there is no type-V component attached to bj1 . Case 8. If b1 = bj1 and v = aj2 , then dG (b1 ) = 2 and thus dG (a1 ) ≥ 3 by Operation 1. Case 8.1. If there is a vertex u outside C that is adjacent to a1 , then we add the edge {a1 , u} to F , add the edge {aj2 , b1 } to C while delete the edge {aj2 , a1 } from C; this way, C is settled. Case 8.2. Note that if a1 and bj3 (6= b3 ) are adjacent, then a1 and aj3 are adjacent too. By the claw-free property, we conclude that a1 must be adjacent to a vertex u ∈ {aj3 , b3 , bj2 , b2 }. Case 8.2.1. When u = aj3 (or u = bj2 ), we add the edges {a1 , u} and {b1 , aj2 } to C while delete the edges {u, aj2 } and {aj2 , a1 } from C; this way, we obtain a path with two leaves b3 and b2 , and thus settle C. Case 8.2.2. Otherwise, aj3 and bj2 are adjacent. When u = b2 (6= bj2 ), we add the edge {a1 , b2 } to C while delete the edge {aj2 , a1 } from C, to obtain a path with two leaves b3 and b1 ; we may also add the edges {a1 , b2 } and {aj2 , b1 } to C while delete the edges {aj2 , a1 } and {aj2 , bj2 } from C to obtain a path with two leaves b3 and bj2 . Between these two paths, the maximum total weight of the internal vertices is at least a3 + aj2 + bj2 + a1 + b1 − min{b1 , bj2 } ≥ 2(b3 + min{b1 , bj2 }); thus, Approximating the maximum weight internal spanning tree 21 C is settled. When u = b3 (6= bj3 ), the same argument applies to settle C. Case 9. If b1 = bj1 and v = bj2 , we add the edge {b1 , bj2 } to C while delete the edge {aj2 , bj2 } from C; this way, we obtain a path with two leaves b3 and b2 , and thus C is settled. If b1 = bj1 and v = aj3 , we add the edge {b1 , aj3 } to C while delete the edge {aj3 , aj2 } from C; this way, we obtain a path with two leaves b3 and b2 , and thus C is settled. If b1 = bj1 and v = bj3 , then we distinguish whether bj3 = b3 or not. When bj3 6= b3 , then b1 and aj3 must be adjacent in G and thus C can be settled as in the last paragraph. When bj3 = b3 , if aj3 is adjacent to one of a1 and bj2 , then we can obtain a path with two leaves b3 and b2 to settle C; if a1 and bj2 are adjacent and bj2 6= b2 , then we can obtain a path with two leaves being b2 and the lightest among a3 , aj2 , a1 , and thus the total weight of the internal vertices of this path is at least a2 + a3 + aj2 + a1 + b1 − min{a3 , aj2 , a1 } ≥ 2(b2 + min{a3 , aj2 , a1 }), which settles C; otherwise bj2 = b2 , and then we add the edge {a1 , b2 } to C while delete the edge {a2 , a1 } from C to obtain a cycle, followed by deleting the lightest edge of C ∩ M ∗ = {{a1 , b1 }, {a2 , b2 }, {a3 , b3 }} to settle C. Case 10. If b1 = bj1 , b2 6= bj2 , and v is in the type-V component containing b2 , we distinguish whether v is a head or a tail. Note that the case where b3 6= bj3 and v is in the type-V component containing b3 can be argued in exactly the same way. Case 10.1. If v is a head, say aj4 of the edge ej4 , assume that aj4 is adjacent to bj5 besides bj4 . Exactly the same argument as in Case 7.1 applies to settle C, since it does not matter whether b1 = bj1 or not. Case 10.2. If v is a tail, say bj4 of the edge ej4 . Note that bj4 6= bj2 , which has been dealt in Case 9. Case 10.2.1. If b2 = bj4 , then besides the leaf b3 , we either do nothing to C to leave b1 , b2 as leaves, or add the edge {b1 , b2 } to C and delete the edge {aj2 , a1 } to leave a1 as a leaf, or add the edge {b1 , b2 } to C and delete the edge {aj2 , bj2 } to leave bj2 as a leaf. Among these three trees, the maximum total weight of the internal vertices is at least a3 + a1 + aj2 + bj2 + b1 + b2 − min{b1 + b2 , a1 , bj2 } ≥ 2(b3 + min{b1 + b2 , a1 , bj2 }); thus, C is settled. Case 10.2.2. If b2 6= bj4 , then by the claw-free property and the definition of v we conclude that b1 is adjacent to aj4 too. Assume that aj4 is adjacent to bj5 besides bj4 . When bj5 6= bj2 , the argument in Case 7.1.1 can be applied to settle C; when bj5 = bj2 , if aj3 and bj2 are adjacent, then we add the edges {aj3 , bj2 } and {b1 , aj4 } to C and delete the edges {aj3 , aj2 } and {bj2 , aj4 } to achieve a path with leaves b3 and b2 ; if aj3 and a1 are adjacent, then we add the edges {aj3 , a1 }, {b1 , aj4 } and {b1 , bj4 } to C and delete the edges {aj3 , aj2 }, {a1 , b1 } and {aj4 , bj4 } to achieve a path with leaves b3 and b2 ; if bj2 and a1 are adjacent, then we add the edges {bj2 , a1 }, {b1 , aj4 } and {b1 , bj4 } to C and delete the edges {aj2 , bj2 }, {a1 , b1 } and {aj4 , bj4 } to achieve a path with leaves b3 and b2 . Thus, C is settled. All possible cases have be discussed in the above. The lemma is proven. ✷ Lemma 11 A type-III component in the subgraph H3 can be settled. Proof. Recall that a type-III component C in its original form contains only one edge ej1 = {aj1 , bj1 } of M ∗ , with another edge {aj1 , x} where x ∈ X (see Figure 3a); there could be a type-V component attached to bj1 , with its tail b1 replacing the role of bj1 . 22 Chen et al. /v:May 30, 2017 Case 1. If b1 is adjacent to a vertex v outside C, then we add the edge {b1 , v} to F , which settles C, since the total weight of the internal vertices in C is w(C ∩ M ∗ ). We next consider the case where b1 is not adjacent to any vertex outside C, and thus it has to be adjacent to some vertex inside C. Note that C is a path with b1 and x being its two ending vertices. Let v denote the vertex adjacent to b1 that is the farthest to b1 on C. We distinguish this distance dC (b1 , v) ≥ 2 and where v locates. Case 2. If b1 6= bj1 and v is in the type-V component containing b1 , then v must be a tail of an edge of M ∗ and thus dC (b1 , v) is even. Denote this edge as ej2 = {aj2 , bj2 }, that is v = bj2 . Case 2.1. If dC (b1 , v) ≥ 4, then denote the head vertex bj2 is adjacent to in the type-V component as aj3 , besides aj2 . We conclude from the claw-free property that there must be at least an edge among aj2 , aj3 , b1 , which contradicts the identity of the type-V component. Therefore, it is impossible to have dC (b1 , v) ≥ 4. Case 2.2. If dC (b1 , v) = 2, then we conclude that dG (b1 ) = 2 and thus dG (a1 ) ≥ 3 by Operation 1, i.e. there is at least another edge incident at a1 besides {a1 , bj2 } and {a1 , b1 }. Denote this neighbor of a1 as u, which is impossible to be inside C by our construction algorithm. Thus, we add the edge {b1 , bj2 } to C while delete the edge {a1 , bj2 } from C, and add the edge {a1 , u} to F ; this way, the component becomes a path and thus C is settled. Case 3. If b1 6= bj1 and v = bj1 , we consider the size of the type-V component containing b1 . Case 3.1. If this type-V component contains more than one edge of M ∗ , then by the claw-free property aj1 must be adjacent to b1 , which violates the definition of v being the farthest and thus it is impossible. Case 3.2. If the type-V component containing b1 has only one edge of M ∗ , that is {a1 , b1 }, then exactly the same argument in Case 2.2 settles C. Case 4. If b1 6= bj1 and v = aj1 , then we leave C as it is if w(b1 ) ≤ w(bj1 ), or we add the edge {aj1 , b1 } to C while delete the edge {aj1 , bj1 } from C. In either case, the total weight of the internal vertices in the resulting path is at least aj1 + a1 + max{bj1 , b1 } ≥ 3 min{bj1 , b1 }, and thus it settles C. Case 5. If b1 6= bj1 and v = x, then we leave C as it is if w(b1 ) ≤ w(aj1 ), or we add the edge {x, b1 } to C while delete the edge {aj1 , x} from C. In either case, the total weight of the internal vertices in the resulting path is at least a1 + max{aj1 , b1 } ≥ 2 min{aj1 , b1 }, and thus it settles C. In summary, Cases 2-5 together prove that when b1 6= bj1 , the component C can be settled. We next consider the situation where b1 = bj1 , that is, there is no type-V component adjacent to bj1 . Case 6. If b1 = bj1 , then v = x, and we conclude that dG (b1 ) = 2 and thus dG (a1 ) ≥ 3 by Operation 1, that is, there is a vertex u outside C that is adjacent to a1 . Thus we add the edge {a1 , u} to F , add the edge {x, b1 } to C while delete the edge {a1 , x} from C; this way, C is settled. All possible cases have be discussed in the above. The lemma is proven. ✷ Lemma 12 A type-IV component in the subgraph H3 can be settled. Proof. Denote the two edges of M ∗ in the type-IV component C in its original form as ej1 and ej2 . Note that both aj1 and aj2 are adjacent to a vertex x ∈ X (see Figure 3b), and there could be Approximating the maximum weight internal spanning tree 23 a type-V component attached to bj1 and bj2 , respectively, with the tails b1 , b2 replacing the roles of bj1 , bj2 . We assume w.l.o.g. that w(b1 ) ≥ w(b2 ). Case 1. If b1 is adjacent to a vertex v outside C, then we add the edge {b1 , v} to F , which settles C, since the total weight of the internal vertices in C is at least a1 + b1 + a2 ≥ 3b2 . We next consider the case where b1 is not adjacent to any vertex outside C, and thus it has to be adjacent to some vertex inside C. Note that C is a path with b1 and b2 being its two ending vertices. Let v denote the vertex adjacent to b1 that is the farthest to b1 on C. We distinguish this distance dC (b1 , v) ≥ 2 and where v locates. Case 2. If b1 6= bj1 and v is in the type-V component containing b1 , then v must be a tail of an edge of M ∗ and thus dC (b1 , v) is even. Denote this edge as ej3 = {aj3 , bj3 }, that is v = bj3 . Case 2.1. If dC (b1 , v) ≥ 4, then denote the head vertex bj3 is adjacent to in the type-V component as aj4 , besides aj3 . We conclude from the claw-free property that there must be at least an edge among aj3 , aj4 , b1 , which contradicts the identity of the type-V component. Therefore, it is impossible to have dC (b1 , v) ≥ 4. Case 2.2. If dC (b1 , v) = 2, then we conclude that dG (b1 ) = 2 and thus dG (a1 ) ≥ 3 by Operation 1, i.e. there is at least another edge incident at a1 besides {a1 , bj3 } and {a1 , b1 }. Denote this neighbor of a1 as u. If u is inside C, then u = b2 by our construction algorithm and the claw-free property. We add the edges {a1 , b2 } and {bj3 , b1 } to C while delete the edge {bj3 , a1 } from C to obtain a cycle, followed by deleting the lightest edge of C ∩ M ∗ ; this settles C since C ∩ M ∗ has at least three edges. If u is outside C, then we add the edge {b1 , bj3 } to C while delete the edge {a1 , bj3 } from C, and add the edge {a1 , u} to F ; this way, the component becomes a path and thus C is settled. Case 3. If b1 6= bj1 and v = bj1 , we consider the size of the type-V component containing b1 . Case 3.1. If this type-V component contains more than one edge of M ∗ , then by the claw-free property aj1 must be adjacent to b1 , which violates the definition of v being the farthest and thus it is impossible. Case 3.2. If the type-V component containing b1 has only one edge of M ∗ , that is {a1 , b1 }, then from the definition of v we have dG (b1 ) = 2, and thus dG (a1 ) ≥ 3 by Operation 1, i.e. there is at least another edge incident at a1 besides {a1 , bj1 } and {a1 , b1 }. The same argument as in Case 2.2, with j3 replaced by j1 , applies to settle C. Case 4. If b1 6= bj1 and v = aj1 , then we leave C as it is when w(b1 ) ≤ w(bj1 ), or we add the edge {aj1 , b1 } to C while delete the edge {aj1 , bj1 } from C. In either case, the total weight of the internal vertices in the resulting path is at least a2 + aj1 + a1 + max{bj1 , b1 } ≥ 2(b2 + min{bj1 , b1 }), and thus it settles C. Case 5. If b1 6= bj1 and v = x, then by the definition of v and the claw-free property aj1 and b1 are adjacent. We thus settle C as in Case 4. We note that this is simpler than Case 5 in the proof of Lemma 10 because here aj1 and aj2 cannot be adjacent. Case 6. If b1 6= bj1 and v = aj2 , then we add the edge {aj2 , b1 } to C while delete the edge {aj2 , x} from C. This gives a path with leaves b2 and x, and thus it settles C. If b1 6= bj1 and v = bj2 , then we conclude that aj2 and b1 are adjacent when b2 6= bj2 , and thus we settle C as in the last paragraph; when b2 = bj2 , we add the edge {b2 , b1 } to C to obtain a cycle, 24 Chen et al. /v:May 30, 2017 followed by deleting the lightest edge of C ∩ M ∗ from C. Since C ∩ M ∗ has at least three edges, this settles C. Case 7. If b1 6= bj1 , b2 6= bj2 , and v is in the type-V component containing b2 , we distinguish whether v is a head or a tail. Case 7.1. If v is a head, say aj3 of the edge ej3 , and assume that aj3 is adjacent to bj4 besides bj3 , then we do nothing to C to leave b1 as a leaf when w(b1 ) ≤ w(bj4 ), or otherwise add the edge {aj3 , b1 } to C while delete the edge {aj3 , bj4 } from C to leave bj4 as a leaf. In either way, the total weight of the internal vertices of the resulting path is at least a2 +aj4 +a1 +max{bj4 , b1 } ≥ 2(b2 +min{bj4 , b1 }). Therefore, in either case C can be settled. Case 7.2. If v is a tail, say bj3 of the edge ej3 . If b2 = bj3 , then we add the edge {b1 , b2 } to C to obtain a cycle, followed by deleting the lightest edge of C ∩ M ∗ from C. Since C ∩ M ∗ has at least three edges, this settles C. If b2 6= bj3 , then by the claw-free property and the definition of v we conclude that b1 is adjacent to aj3 too, and thus the argument in Case 7.1 applies to settle C. This finishes the discussion on Case 7. In summary, Cases 2-7 together prove that when b1 6= bj1 , the component C can be settled. We next consider the situation where b1 = bj1 , that is, there is no type-V component adjacent to bj1 . Case 8. If b1 = bj1 and v = x, and we conclude that dG (b1 ) = 2 and thus dG (a1 ) ≥ 3 by Operation 1, that is there is a vertex u adjacent to a1 other than x and b1 . Case 8.1. If u is inside C, then u = b2 by our construction algorithm and the claw-free property. 2 If b 6= bj2 , then we add the edges {a1 , b2 } and {x, b1 } to C while delete the edge {x, a1 } from C to obtain a cycle, followed by deleting the lightest edge of C ∩ M ∗ ; this settles C since C ∩ M ∗ has at least three edges. If b2 = bj2 , then we add the edges {a1 , b2 } and {x, b1 } to C while delete the edges {x, a1 } and {x, a2 } from C to obtain a path with leaves x and a2 when w(a2 ) ≤ w(b1 ), or otherwise we add the edge {a1 , b2 } to C while delete the edge {x, a1 } from C to obtain a path with leaves x and b1 . In either way, the total weight of the internal vertices of the resulting path is at least a1 + max{a2 , b1 } ≥ 2 min{a2 , b1 }. Therefore, C can be settled. Case 8.2. If u is outside C, then we add the edge {b1 , x} to C while delete the edge {a1 , x} from C, and add the edge {a1 , u} to F ; this way, the component becomes a path and thus C is settled. Case 9. If b1 = bj1 and v = aj2 , then we add the edge {aj2 , b1 } to C while delete the edge {aj2 , x} from C to obtain a path with leaves x and b2 . Therefore, C can be settled. If b1 = bj1 and v = bj2 , when b2 6= bj2 then we conclude from the claw-free property and the definition of v that aj2 and b1 are also adjacent, and thus C can be settled as in the last paragraph; when b2 = bj2 , we can either do nothing to C to leave b1 and b2 as leaves, or add the edge {b2 , b1 } to C to obtain a cycle, followed by either deleting the edge {a2 , x} to obtain a path with leaves a2 and x, or deleting the edge {a1 , x} to obtain a path with leaves a1 and x. This way, the maximum total weight of the internal vertices among the three resulting paths is at least a1 + a2 + b1 + b2 − min{a1 , a2 , b1 + b2 } ≥ 2 min{a1 , a2 , b1 + b2 }. Thus, C can be settled. Case 10. If b1 = bj1 , b2 6= bj2 , and v is in the type-V component containing b2 , we distinguish whether v is a head or a tail. Case 10.1. If v is a head, say aj3 of the edge ej3 , and assume that aj3 is adjacent to bj4 besides bj3 , then we do nothing to C to leave b1 as a leaf if w(b1 ) ≤ w(bj4 ), or otherwise add the edge {aj3 , b1 } to C while delete the edge {aj3 , bj4 } from C to leave bj4 as a leaf. This way, the total weight of the Approximating the maximum weight internal spanning tree 25 internal vertices of the resulting path is at least a1 + a2 + aj4 + max{b1 , bj4 } ≥ 2(b2 + min{b1 , bj4 }). Thus, C can be settled. Case 10.2. If v is a tail, say bj3 of the edge ej3 . If b2 = bj3 , then we add the edge {b1 , b2 } to C to make it a cycle followed by deleting the lightest edge of C ∩ M ∗ ; this settles C since there are at least three edges in C ∩ M ∗ . If b2 6= bj3 , then from the claw-free property and the definition of v, we conclude that aj3 must be adjacent to b1 too. Therefore, the same as in Case 10.1, C can be settled. All possible cases have be discussed in the above. The lemma is proven. ✷ Lemma 13 A type-VI component in the subgraph H3 can be settled. Proof. Recall that a type-VI component C is a cycle containing two or more edges of M ∗ , where the head of one edge of M ∗ is adjacent to the tail of another edge of M ∗ (see Figure 4b). Clearly, if there are three or more edges of M ∗ in C, we simply delete the lightest one to settle C. In the sequel we deal with the case where C is a length-4 cycle. Denote the two edges of C ∩M ∗ as {a1 , b1 } and {a2 , b2 }, and assume that w(b1 ) ≥ w(b2 ). If a1 (a2 , respectively) is adjacent to a vertex v outside C, then we add the edge {a1 , v} ({a2 , v}, respectively) to F and delete the edge {a1 , b2 } ({a2 , b2 }, respectively) from C; this way, the total weight of the internal vertices is at least a1 + a2 + b1 ≥ 3b2 , and thus C is settled. If neither a1 nor a2 is adjacent to any vertex v outside C, then we conclude from the construction algorithm that dG (a1 ) = dG (a2 ) = 2 since a1 and a2 are not adjacent to each other. It follows from the claw-free property that neither b1 nor b2 can be adjacent to any vertex u outside C. This implies |V | = 4, a contradiction to our assumption that |V | ≥ 5. The lemma is proved. ✷ Theorem 2 The MwIST problem on claw-free graphs admits a 12/7-approximation algorithm. Proof. The above Lemmas 9–13 state that every component of the subgraph H3 = G[V, M ∗ ∪ M aa ∪ N aa ∪ N ax ∪ M ab ∪ N ab ] can be settled, without affecting any other components. Also, such a settling process for a component takes only linear time, by scanning once the edges in the subgraph induced on the set of vertices of the component. By settling, essentially the component is converted into a tree, possibly with one edge of F specified for connecting a leaf of the tree outwards. In the next step of the algorithm, it iteratively processes the heaviest component C, i.e. with the largest w(C ∩ M ∗ ). If the component C has been associated with an edge e of F , and using the edge e to connect a leaf of the resulting tree for C outwards does not create a cycle, then the algorithm does this and C is processed. This guarantees the total weight of the internal vertices in V (C) is at least 2w(C ∩ M ∗ )/3. If using the edge e to connect a leaf of the resulting tree for C outwards would create a cycle, the algorithm processes C by replacing C with another tree that guarantees the total weight of the internal vertices in V (C) at least 21 w(C ∩ M ∗ ). Notice that the latter case happens only because of (at least) one edge of F in an earlier iteration where a distinct component C ′ was processed, which connects a vertex of C ′ into a vertex of C. Therefore, every such C is associated with a distinct component C ′ processed by the algorithm in an earlier iteration, and thus w(C ′ ) ≥ w(C). On the other hand, every such component C ′ is associated to one C only, due to its edge in F connecting a leaf outwards into a vertex of C. It follows that for this pair of components C and C ′ , the total weight of the internal vertices in V (C) ∪ V (C ′ ) is at least w(C)/2 + 2w(C ′ )/3 ≥ 7(w(C) + w(C ′ ))/12. 26 Chen et al. /v:May 30, 2017 After all components of H3 are processed, we obtain a forest for which the total weight of the internal vertices therein is at least 7w(M ∗ )/12. The algorithm lastly uses any other available edges of E to interconnect the forest into a final tree, denoted as T ; clearly w(T ) ≥ 7w(M ∗ )/12. The time for the interconnecting purpose is at most O(m log n). Therefore, by Corollary 1 we have a 7/12-approximation algorithm for the MwIST problem on claw-free graphs. ✷ 4 Concluding remarks We have presented an improved approximation algorithm for the vertex weighted MIST problem, denoted MwIST, which achieves the worst-case performance ratio 1/2, beating the previous best ratio of 1/(3 + ǫ), designed by Knauer and Spoerhase in 2009 [7]. The key ingredient in the design and analysis of our algorithm is a novel relationship between MwIST and the maximum weight matching, which we uncovered and it is inspired by the work [12, 11, 2]. A step further, for the problem restricted to claw-free graphs, we presented a 7/12-approximation algorithm, improving the previous best ratio of 1/2 designed by Salamon in 2009 for claw-free graphs without leaves. It would be interesting to see whether this newly uncovered relationship, possibly combined with other new ideas, can be explored further to design better approximation algorithms for MwIST, or special cases of MwIST including claw-free graphs and cubic graphs. Acknowledgement The authors are grateful to two reviewers for their insightful comments on the COCOON 2017 submission and for their suggested changes that improve the presentation greatly. ZZC was supported in part by the Grant-in-Aid for Scientific Research of the Ministry of Education, Culture, Sports, Science and Technology of Japan, under Grant No. 24500023. GL was supported by the NSERC Canada and the NSFC Grant No. 61672323; most of his work was done while visiting ZZC at the Tokyo Denki University at Hatoyama. LW was supported by the Hong Kong GRF Grants CityU 114012 and CityU 123013. YC was supported in part by the NSERC Canada, the NSFC Grants No. 11401149, 11571252 and 11571087, and the China Scholarship Council Grant No. 201508330054. References [1] D. Binkele-Raible, H. Fernau, S. Gaspers, and M. Liedloff. Exact and parameterized algorithms for max internal spanning tree. Algorithmica, 65:95–128, 2013. [2] Z.-Z. Chen, Y. Harada, and L. Wang. An approximation algorithm for maximum internal spanning tree. CoRR, abs/1608.00196, 2016. [3] N. Coben, F. V. Fomin, G. Gutin, E. J. Kim, S. Saurabh, and A. Yeo. Algorithm for finding kvertex out-trees and its application to k-internal out-branching problem. Journal of Computer and Systems Science, 76:650–662, 2010. [4] F. V. Fomin, S. Gaspers, S. Saurabh, and S. Thomasse. A linear vertex kernel for maximum internal spanning tree. Journal of Computer and Systems Science, 79:1–6, 2013. Approximating the maximum weight internal spanning tree 27 [5] F. V. Fomin, D. Lokshtanov, F. Grandoni, and S. Saurabh. Sharp separation and applications to exact and parameterized algorithms. Algorithmica, 63:692–706, 2012. [6] M. R. Garey and D. S. Johnson. Computers and Intractability: A Guide to the Theory of NP-completeness. W. H. Freeman and Company, San Francisco, 1979. [7] M. Knauer and J. Spoerhase. Better approximation algorithms for the maximum internal spanning tree problem. In Proceedings of WADS 2009, LNCS 5664, pages 489–470, 2009. [8] W. Li, J. Chen, and J. Wang. Deeper local search for better approximation on maximum internal spanning tree. In Proceedings of ESA 2014, LNCS 8737, pages 642–653, 2014. [9] W. Li, J. Wang, J. Chen, and Y. Cao. A 2k-vertex kernel for maximum internal spanning tree. In Proceedings of WADS 2015, LNCS 9214, pages 495–505, 2015. [10] X. Li, H. Jiang, and H. Feng. Polynomial time for finding a spanning tree with maximum number of internal vertices on interval graphs. In Proceedings of FAW 2016, LNCS 9711, pages 92–101, 2016. [11] X. Li and D. Zhu. A 4/3-approximation algorithm for the maximum internal spanning tree problem. CoRR, abs/1409.3700, 2014. [12] X. Li and D. Zhu. Approximating the maximum internal spanning tree problem via a maximum path-cycle cover. In Proceedings of ISAAC 2014, LNCS 8889, pages 467–478, 2014. [13] E. Prieto. Systematic kernelization in FPT algorithm design. PhD thesis, The University of Newcastle, Australia, 2005. [14] E. Prieto and C. Sliper. Either/or: Using vertex cover structure in designing fpt-algorithms – the case of k-internal spanning tree. In Proceedings of WADS 2003, LNCS 2748, pages 474–483, 2003. [15] E. Prieto and C. Sloper. Reducing to independent set structure – the case of k-internal spanning tree. Nordic Journal of Computing, 12:308–318, 2005. [16] G. Salamon. Approximating the maximum internal spanning tree problem. Theoretical Computer Science, 410:5273–5284, 2009. [17] G. Salamon. Degree-Based Spanning Tree Optimization. PhD thesis, Budapest University of Technology and Economics, Hungary, 2009. [18] G. Salamon and G. Wiener. On finding spanning trees with few leaves. Information Processing Letters, 105:164–169, 2008. 

8cs.DS

arXiv:1610.03376v1 [math.GR] 9 Oct 2016 CUBULATING RANDOM GROUPS IN THE SQUARE MODEL TOMASZ ODRZYGÓŹDŹ 3 Abstract. Our main result is that for densities < 10 a random group in the square model has the Haagerup property and is residually finite. Moreover, we generalize the Isoperimetric Inequality, to some class of non-planar diagrams and, using this, we introduce a system of modified hypergraphs providing the structure of a space with walls on the Cayley complex of a random group. Then we show that the natural action of a random group on this space with walls is proper, which gives the proper action of a random group on a CAT(0) cube complex. 1. Introduction In [Odr16] we introduced the square model for random groups, where we draw at random relations of length four. The motivation was that the Cayley complex of such a group has a natural structure of a square complex, which should be easier to analyze than the polygonal Cayley complexes for groups in the Gromov model. Definition 1.1. [Square model, [Odr16, Definition 1.3]] Consider the set An = {a1 , . . . , an }, which we will refer to as an alphabet. Let Wn be the set of all cyclically reduced words of length 4 over An . Note that |Wn | = (2n − 1)4 up to a multiplicative constant. By Fn we will denote the free group generated by the elements of An . By relators we will understand words over generators and by relations the corresponding equalities holding in the group. For d ∈ (0, 1) let us choose randomly, with uniform distribution, a subset Rn ⊂ Wn such that |Rn | = ⌊(2n − 1)4d ⌋. Quotienting Fn by the normal closure of the set Rn , we obtain a random group in the square model at density d. We say that property P occurs in the square model at density d with overwhelming probability (w.o.p.) if the probability that a random group has property P converges to 1 as n → ∞. In [Odr16] we showed that a random group in the square model at density d w.o.p.: • is trivial for d > 21 , • is infinite, hyperbolic, torsion-free of geometric dimension 2 for d < 12 , • is free for d < 41 • does not have Property (T) for d < 31 In this paper we prove 3 a random group in the square model w.o.p. Theorem 1.2. For densities d < 10 acts properly and cocompactly on a CAT(0) cube complex. 3 Corollary 1.3. For densities d < 10 a random groups in the square model w.o.p. has the Haagerup property and is residually finite. 1 2 TOMASZ ODRZYGÓŹDŹ Proof. The residual finiteness results from [Ago13, Corollary 1.2] and the Haagerup Property by a folklore remark (see for example [CMV04]).  To obtain Theorem 1.2 we needed to generalize the isoperimetric inequality ([Oll07, Theorem 2]) to a class of non-planar diagrams (see Theorem 2.5). This generalization has already proved to be useful in other work on random groups (see for example [MP15]). Theorem 2.5 is stated in the regime of the square model, but the reasoning can be easily repeated in the Gromov model (see [Odr14] for some version of it). The main idea of the proof of Theorem 1.2 is based on introducing modified hypergraps, which provide a structure of a space with walls on the Cayley complex of a random group, that has desired metric properties. This modified walls allowed us to construct a proper action of a random group on a space with walls, which was impossible to obtain using standard hypergraphs (introduced in [OW11] for the Gromov model and in [Odr16] for the square model). Organization. The paper is organized as follows: first we prove a generalization of the isoperimetric inequality, then using this we introduce “corrected” hypergraphs that provide the structure of a space with walls on the Cayley complex of a random group. Finally we show that a random group acts properly on this space with walls. The idea of correcting hypergraphs was inspired by [MP15]. Acknowledgements. The author would like to thank Piotr Przytycki for many discussions and Piotr Nowak for general advice on writing. 2. Non-planar isoperimetric inequality The goal of this section is to generalize the isoperimetric inequality to some class of non-planar complexes, but having “disc-like” structure, meaning that they are unions of a large disc component and some bounded non-planar pieces. Firstly, let us recall the isoperimetric inequality: Theorem 2.1 ([Oll07, Theorem 2]). For any ε > 0, in the Gromov model at density d < 21 , with overwhelming probability all reduced van Kampen diagrams associated to the group presentation satisfy (2.1) |∂D| ≥ l(1 − 2d − ε)|D|. Here ∂D denotes the set of boundary edges of the diagram D and |D| denotes the number of 2-cells of D. One of the corollaries of Theorem 2.1 is that in the Gromov density model at densities < 21 a random group is w.o.p. hyperbolic (see [Oll04, Theorem 1] and [Gro93]). In [Odr16, Theorem 3.15] the author proved that Equation (2.1) is satisfied w.o.p. as well for the square model of random groups. Let us now introduce some definitions. Definition 2.2. Suppose Y is a finite 2–complex, not necessarily a disc diagram. ˜ |, is • The generalized boundary length of Y , denoted |∂Y X ˜ | := (2 − deg(e)), |∂Y e∈Y (1) CUBULATING RANDOM GROUPS IN THE SQUARE MODEL • and the cancellation in Y is Cancel(Y ) := X e∈Y 3 (deg(e) − 1), (1) where deg(e) is the number of times that e appears as the image of an edge of the attaching map of a 2–cell of Y . ˜ | = |∂Y |. Moreover, Cancel(Y ) = Note, that for every planar diagram Y we have |∂Y | − |∂Y |). 1 2 (4|Y Definition 2.3. We say that a finite 2-complex Y is fulfilled by a set of relators e R if there is a combinatorial map from Y to the presentation complex X/G that is locally injective around edges (but not necessarily around vertices). e is fulfilled by R. In particular, any subcomplex of the Cayley complex X Definition 2.4 (Diagram with K-small legs). Let G = hS|Ri be a finite group presentation where R consists of cyclically reduced words of length 4 over S. For K > 0 let Y be a 2–complex that is a union of a reduced Van Kampen diagram Z and a family Zi of connected complexes, called legs of Y , such that |Zi | ≤ K, for each i the leg Zi contains an external vertex of Z and that every edge of Y belongs to maximally two legs. We call such Y a diagram with K-small legs (with respect to the presentation hS|Ri). The disc diagram Z is called the disc basis of Y . Until the end of this section G will denote the random group in the square model e — the Cayley complex of G with respect to this with the presentation hS|Ri and X presentation. Theorem 2.5 (Generalized Isoperimetric Inequality). In the square model at density d ≤ 21 the following statement holds w.o.p.: for each K and ε > 0 there is no diagram Y with K-small legs fullfilable by R and satisfying: (2.2) Cancel(Y ) > 4(d + ε)|Y |, or equivalently (2.3) ˜ | > 4(1 − 2d − ε)|Y |. |∂Y Our strategy to prove Theorem 2.5 is to first prove a “local” version of it, that is with the additional limit on the number of 2-cells in a diagram, and then to show that this locality assmuption can be ommited. Analogous results for the Gromov model was obtained by the author in [Odr14] and used in [MP15] to construct balanced walls. We thank Piotr Przytycki for suggesting the statement of Theorem 2.5 involving the notion of cancellation. 2.1. Local version of the generalized isoperimetric inequality. Lemma 2.6. In the square model at density d ≤ 21 , for each K, ε > 0 w.o.p. there is no 2-complex Y with |Y | ≤ K fulfilled by R and satisfying (2.4) Cancel(Y ) > 4(d + ε)|Y |. Our proof of Lemma 2.6 will be only a slight modification of Olliver’s proof of Theorem 2.1. We postpone it to to introduce some necessary tools first. 4 TOMASZ ODRZYGÓŹDŹ Definition 2.7. Let Y be a 2-complex. Suppose that the following additional information is given: (1) Every face of Y is labeled by a number i ∈ {1, 2, . . . , n}. The labels can repeat and for each number 1 ≤ i ≤ n there is at least one face labeled with i. We call n the number of distinct relators. (2) For each face there is a distinguished edge, called the starting edge and an orientation. Such Y will be referred to as an abstract 2-complex. We say that Y is fulfilled by a sequence of relators (r1 , r2 , . . . , rn ) if it is fulfilled by {r1 , r2 , . . . , rn } as 2complex, in such a way that each face f labelled by i ∈ {1, 2, . . . , n} (according to (1) in Definition 2.7) is sent under the combinatorial map to the face of X/G corresponding to ri , so that the image in X (1) /G of the attaching path of f , starting at the oriented starting edge (according to (2) in Definition 2.7) reads off the word ri . If f is labeled by i we say that f bears the relator i. Recall that S denotes the cardinality of the generating set of a random group in the square model. Let m denote the cardinality of the set S. Proposition 2.8. Let R be a random set of relators at density d and at length 4 on m generators. Let Y be an abstract 2-complex. Then either Cancel(Y ) < 4(d + 2ε)|Y | or the probability that there exists a tuple of relators in R fulfilling Y is less than (2m − 1)−4ε . Before we prove this proposition we introduce some additional notation. Let n be the number of distinct relators in an abstract 2-complex Y . For 1 ≤ i ≤ n let mi be the number of times relator i appears in Y . Up to reordering the relators we can suppose that m1 ≥ m2 ≥ · · · ≥ mn . For 1 ≤ i1 , i2 ≤ n and 1 ≤ k1 , k2 ≤ l we say that (i1 , k1 ) > (i2 , k2 ) if i1 > i2 or i1 = i2 but k1 > k2 (lexicographic order). Suppose that for some s ≥ 2 an edge e of Y is adjacent to faces f1 , f2 , . . . , fs labeled by i1 , i2 , . . . , is accordingly. Suppose moreover that for 1 ≤ j ≤ s the edge e is the kj –th edge of the face fj . Since Y → X/G is locally injective around e, the pairs (ij , kj ) are distinct. Choose j = jmin for which (ij , kj ) is minimal. We say that edge e belongs to faces fj for j ∈ {1, 2, . . . s} \ {jmin }. Let δ(f ) be the number of edges belonging to a face f . For 1 ≤ i ≤ n let κi = max{δ(f ) : f is a face labeled by relator i} Note that (2.5) Cancel(Y ) = X f ∈Y (2) δ(f ) ≤ X mi κ i 1≤i≤n Definition 2.9. Let Y be an abstract 2-complex with n distinct relators. For 1 ≤ k ≤ n let R′ = (w1 , w2 , . . . , wk ) be a sequence of relators. We say that Y is partially fulfilled by R′ if the abstract 2-complex Y ′ ⊂ Y that is the closure of the faces of Y labeled by the numbers i ∈ {1, 2, ..., k} is fulfilled by R′ . Lemma 2.10. For 1 ≤ i ≤ n let pi be the probability that i randomly chosen words w1 , w2 , . . . , wi partially fulfill Y and let p0 = 1. Then pi ≤ (2m − 1)−κi . (2.6) pi−1 CUBULATING RANDOM GROUPS IN THE SQUARE MODEL 5 Proof. Suppose that first i − 1 words w1 , . . . , wi−1 partially fulfilling Y are given. We will successively analyze what is the choice for the consecutive letters of the word wi , so that Y is fulfilled. Let k ≤ 4 and suppose that the first k − 1 letters of wi are chosen. Let f be the face realizing δ(f ) = κi and let e be the k-th edge of the face f . If e belongs to f this means that there is another face f ′ meeting e which bears relator i′ < i or bears i too, but e appears in f ′ as a k ′ < k-th edge. In both cases the letter on the edge e is imposed by some letter already chosen so drawing it at 1 up to some small error. In fact, this estimate random has probability ≤ (2m−1) would be valid only if the words were only required to be reduced; since they are cyclically reduced a negligible error appears, which we ignore. Combining all these observations we get that the probability to choose at random  the correct word wi is at most pi−1 (2m − 1)−κi . Now we can provide the proof of Proposition 2.8 Proof of Proposition 2.8. For 1 ≤ i ≤ n let Pi be the probability that there exists an i-tuple of words partially fulfilling Y in the random set of relators R. We trivially have: Pi ≤ |R|i pi = (2m − 1)4id pi (2.7) Combining equations (2.5) and (2.6) we get n X Cancel(Y ) ≤ mi (log2m−1 pi−1 − log2m−1 pi ) = i=1 = n−1 X (mi+1 − mi ) log2m−1 pi − mn log2m−1 pn + m1 log2m−1 p0 . i=1 Now p0 = 1 so log2m−1 p0 = 0 and we have Cancel(Y ) ≤ n−1 X (mi+1 − mi ) log2m−1 pi − mn log2m−1 pn . i=1 From (2.7) and the fact that mi+1 − mi ≤ 0 we obtain Cancel(Y ) ≤ n−1 X (mi+1 − mi )(log2m−1 Pi − 4id) − mn log2m−1 (Pn − 4nd) i=1 Observe that Pn−1 i=1 (mi − mi+1 )i + mn n = Cancel(Y ) ≤ 4|Y |d + n−1 X Pn i=1 mi = |Y |. Hence (mi+1 − mi ) log2m−1 Pi − mn log2m−1 Pn i=1 Setting P = mini Pi and using the fact that mi+1 − mi ≤ 0 we get Cancel(Y ) ≤ 4|Y |d + (log2m−1 P ) n−1 X (mi+1 − mi ) − mn log2m−1 P = i=1 = 4|Y |d − m1 log2m−1 P ≤ |Y |(4d − log2m−1 P ), 6 TOMASZ ODRZYGÓŹDŹ since m1 ≤ |Y |. It is clear that a complex is fulfillable if it is partially fulfillable for any i ≤ n and so: Probability(Y is fullfillable by relators of R) ≤ P ≤ (2m − 1) 4|Y |d−Cancel(Y ) |Y | which was to be proven. ,  Proof of Lemma 2.6. Denote by C(K, m) the number of abstract square complexes with at most K 2-cells. Observe that there are finitely many square complexes with at most K faces. There are also finitely many ways to decide which faces would bear the same relator, and also finitely many ways to choose the starting point of each relator. Therefore, the values {C(K, m)}m∈N with fixed K have a uniform bound M . We know by Proposition 2.8 that for any abstract 2-complex with at most K faces violating the inequality Cancel(Y ) < 4(d + ε)|Y | the probability that it is fulfilled by a random set of relators is ≤ (2m − 1)−4ε . So the probability that there exists a fulfilled 2-complex with at most K faces, violating the inequality is ≤ C(K, m)(2m − 1)−4εl ≤ M (2m − 1)−4εl , so it converges to 0 as m → ∞.  2.2. From the local version to the global. We start this subsection by reformulating [Oll07, Lemma 11] by replacing the length of relator l by 4. Lemma 2.11. Let G = hS|Ri be a finite presentation in which all elements of R have length 4. Suppose that for some constant C ′ > 0 every van Kampen diagram E of this presentation satisfies: |∂E| ≥ 4C ′ |E|. Then every van Kampen diagram D can be partitioned into two diagrams D′ , D′′ by with endpoints on the boundary cutting it along a path of length at most 4 + 8 log(|D|) C′ of D such that each of D′ and D′′ contains at least one quarter of the boundary of D. We will state now prove two propositions „approximating” Theorem 2.5. Proposition 2.12. Let G = hS|Ri be a finite presentation such that all elements of R are reduced words of length 4. Suppose that for some constant C ′ all van Kampen diagrams D with respect to this presentation satisfy |∂D| ≥ C ′ 4|D|. 7A Choose any K, ε > 0. Take A large enough to satisfy εA > 2(1 + C2′ log( 6C ′ ) + 2K). Suppose that for some C > 0 all diagrams with K–small legs Y having the disc basis of boundary at most 4A satisfy: ˜ | ≥ C4|Y |. |∂Y Then all diagrams with K–small legs Y having the disc basis of boundary at most 14 3 A satisfy: ˜ | ≥ (C − ε)4|Y | |∂Y CUBULATING RANDOM GROUPS IN THE SQUARE MODEL 7 Proof. Let Y be the diagram with K–small legs such that its disc basis Z has the boundary length between 4A and 14 3 A. By Lemma 2.11 we can perform a partition of Z into two reduced disc diagrams Z ′ and Z ′′ such that: |∂Z ′ |, |∂Z ′′ | ≥ 41 |∂Z| and |∂Z ′ ∩ ∂Z ′′ | ≤ 4 + C8′ log(|Z|). The number of 1–cells that belong both to Z and one of the legs Zi is bounded by 4K, since there are no more than 4K 1–cells in Zi . We define Y ′ to be the union of Z ′ and all legs of Y that are adjacent to Z ′ , and Y ′′ to be the union of Z ′′ and all legs of Y adjacent to it. Hence we can perform a partition of Y into two diagrams Y ′ and Y ′′ with K–small legs such that ˜ | ≥ |∂Y ˜ ′ | + |∂Y ˜ ′′ | − 2(4 + 8′ log(|Z|) + 8K). |∂Y C Note that |∂Z ′ |, |∂Z ′′ | < 4A so by our assumption we know ˜ ′ | ≥ C4|Y ′ | |∂Y ˜ ′′ | ≥ C4|Y ′′ |. |∂Y Moreover by the assumption on van Kampen diagrams we obtain that 7A 1 C ′ log( 6C ′ ). Hence 1 C′ log(|Z|) < ˜ | ≥ |∂Y ˜ ′ | + |∂Y ˜ ′′ | − 2(4 + 8 log(|Z|) + 8K) |∂Y C′ 7A 8 ≥ C4(|Y ′ | + |Y ′′ |) − 2(4 + ′ log( ′ ) + 8K). C 6C 7A We have chosen A large enough so that 2(1 + 2 C1′ log( 6C ′ ) + 2K) < εA so we can continue estimation ˜ | ≥ Cl|Y | − 4εA ≥ 4(C − ε)|Y | |∂Y since 4|Y | ≥ |∂Z| ≥ 4A.  The last approximation to Theorem 2.5 is the following Proposition 2.13. Let G = hS|Ri be a finite presentation such that all elements of R are reduced words of length 4. Suppose that for some constant C ′ all van Kampen diagrams D with respect to the presentation satisfy (2.8) |∂D| ≥ C ′ l|D|. 7A Choose any K, ε > 0. Take A large enough to satisfy εA > 2(1+2 C1′ log( 6C ′ )+2K). Suppose that for some C > 0 all diagrams Y with K–small legs and the disc basis with boundary length at most 4A satisfy: (2.9) ˜ | ≥ 4C|Y |. |∂Y Then all diagrams with K–small legs Y satisfy: ˜ | ≥ 4(C − ε)|Y | |∂Y Proof. The assumptions of this proposition and Proposition 2.12 are the same. Hence by the statement of Proposition 2.12 we can conclude that the assumptions 1 of Proposition 2.12 are fulfilled with the new parameters: A1 = 67 A, ε1 = ε( 67 ) 2 and C1 = C − ε instead of A, ε, C and with the same C ′ (these new parameters 1 indeed satisfy ε1 A1 > 2(1 + C2′ log( 7A 6C ′ ) + 2K)). By induction every diagram with 8 TOMASZ ODRZYGÓŹDŹ K − small legs such that its disc basis has boundary of length at most 4A satisfies ˜ |≥ |∂Y C−ε k−1 X i=0 and we conclude by the inequality P∞ i=0 6 7 6 7
2i  2i !
7 k 6 4|Y | < 14.  Proof of Theorem 2.5. First, by Theorem 2.1 we know that all Van Kampen diagrams satisfy Equation (2.8) for C = (1 − 2d − ε′ ), for arbitrary small ε′ . By Lemma 2.6 we know that for any K, ε all diagrams with K-small legs satisfy Equation (2.9). Hence the assumptions of the Proposition 2.13 are satisfied, which gives the statement.  3. Colored hypergraphs In this section we we will introduce colored hypergraphs, which can be effectively used in investigating random groups in the square model. We start with e be the Cayley complex of a random group in the square Proposition 3.1. Let X 1 model at density d < 3 . Then w.o.p.: e having three common edges. (1) There is no pair of 2-cells in X e such that |∂D ∩ ∂E| = 2, then there is no 2-cell (2) If D and E are 2-cells in X e F in X such that |(∂D ∪ ∂E) ∩ ∂F | > 1. e having three common edges forms a van Kampen Proof. (1) A pair of 2-cells in X diagram D satisfying |∂D| = 2 and |D| = 2. Therefore |∂D| ≤ 43 |D|, so D violates Theorem 2.5, which implies that w.o.p. there is no such a pair of 2-cells. (2) Suppose, on the contrary, that there exist such 2-cells D, E and F . Consider a diagram D consisting of 2-cells D and E. Adding a new 2-cell to D by gluing it along two edges does not change the generalized boundary length of D. Hence ˜ | = 4, |Y | = 3. the diagram Y that is a union of D and the 2-cell F satisfies: |∂Y ˜ Therefore, Y violates Theorem 2.5, which states that |∂Y | > 4(1 − 2d)|Y |, so w.o.p. there is no such a triple of 2-cells.  e be the Cayley complex of a random group in the square Definition 3.2. Let X 1 e have two common edges, we say they are model at density ≤ 3 . If two 2-cells in X strongly adjacent. If a 2-cell D is strongly adjacent to some other 2-cell we say that D is a distinguished 2-cell. The common edges of two strongly adjacent 2-cells we call distinguished edges. A 2-cell which is not a distinguished 2-cell we call a regular 2-cell. Proposition 3.3. Let D and E be strongly adjacent 2-cells. Then there is no e element of a random group G sending D to E under the natural action of G on X. e In other words, D and E are in different orbits of the action of G on X.  (2n−1)4 −6(2n)2 (2n−1)4d  (2n−1)4 (2n−1)4d   , CUBULATING RANDOM GROUPS IN THE SQUARE MODEL 9 Proof. Suppose that there exists g ∈ G sending D to E. By Lemma 3.1 we know that a 2-cell can be strongly adjacent to at most one other 2-cell, co g preserves the pair {D, E}. The diagram Y consisting of 2-cells D and E has exactly one internal vertex (that is a vertex isolated from the boundary). If the diagram {D, E} is preserved by g the internal vertex must be also preserved, which contradicts the fact that the action of G on its Cayley graph is free.  Definition 3.4 (Procedure of painting the Cayley complex). By Proposition 3.1 we know that distinguished 2-cells form a set S of disjoint unordered pairs of 2-cells. By Proposition 3.3 we know that none of these pairs lies in one orbit of G-action e We will now perform a procedure of painting distinguished 2-cells in two on X. colors: red and blue. In the first step we consider one of the pairs {D, E} ∈ S and choose arbitrarily one 2-cell from this pair. Without loss of generality we can suppose that D is chosen. We say that D is a red 2-cell and E is a blue cell. Moreover we say that each element in the orbit of D is red and each element in the orbit of E is blue. In the next step we choose another pair of strongly adjacent 2-cells {D1 , D2 }, that is not in the orbit of the pair {D, E}. Then we paint all elements in the orbit of D1 in red, and all elements in the orbit of E1 in blue. We repeat this step until all distinguished 2-cells are painted. The Cayley complex is cofinite, so there are finitely many orbits of 2-cells and therefore this painting procedure must end after a finite number of steps. The Cayley complex with painted distinguished 2-cells we call painted Cayley complex. From now we will always assume that the Cayley complex of a random group is e painted, and we will denote it by X. We now recall the definition of a hypergraph from [OW11, Definition 2.1]. Definition 3.5 (standard hypergraph). Let X be a connected square complex. We define a graph Γ as follows: The set of vertices of Γ is the set of 1-cells of X. There is an edge in Γ between two vertices if there is some 2-cell R of X such that these vertices correspond to opposite 1-cells in the boundary of R (if there are several such 2-cells we put as many edges in Γ). The 2-cell R is the 2-cell of X containing the edge. There is a natural map ϕ from Γ to X, which sends each vertex of Γ to the midpoint of the corresponding 1-cell of X and each edge of Γ to a segment joining two opposite points in the 2-cell R. Note that the images of two edges contained in the same 2-cell R always intersect, so that in general ϕ is not an embedding. A standard hypergraph in X is a connected component of Γ. The 1-cells of X through which a hypergraph passes are dual to it. The hypergraph Λ embeds if ϕ is an embedding from Λ to X, that is, if no two distinct edges of Λ are mapped to the same 2-cell of X. We call the subdiagram of X consisting of all open faces containing edges of the hypergraph Λ the carrier of Λ and denote it by Car(Λ). The hypergraph segment in X is a finite path in a hypergraph immersed into X. The carrier of a segment Λ′ denoted Car(Λ′ ) is the diagram consisting of all open 2-cells containing edges of the segment Λ′ . By [Odr16, Lemma 5.16] we know that standard hypergraphs are embedded trees. Moreover, [OW11, Lemma 2.3] states that for each standard hypergraph Γ e − Γ has exactly two connected components. However the system of the space X 10 TOMASZ ODRZYGÓŹDŹ standard hypergraphs is not sufficient to construct a proper action of a random group on a CAT(0) cube complex, which we need to prove Theorem 1.3. Our strategy of constructing such an action is based on showing that the wall space metric on the Cayley complex is equivalent to the edge path metric. This means in particular that for every number N if we have two enough distant points in the Cayley complex they are separated by at least N walls, which are hypergraphs in our case. In Figure 1 we present an example of a geodesic edge-path γ of an arbitary length such that no standard hypergraph separates its ends x and y. Situation presented in Figure 1 cannot be excluded for densities ≥ 14 in the square model, therefore we need a better system of walls, based on more systems of hypergraphs. x γ ... y Figure 1. Arbitrary long geodesic edge-path joining points x and y that are not separated by a hypergraph. We now introduce a new type of hypergraph - red hypegraph, which is modified on red 2-cells. Definition 3.6 (red and blue hypergraph). Let X be a connected square complex e We define a graph Γ as that is a subcomplex of the painted Cayley complex X. follows: The set of vertices of Γ is the set of 1-cells of X. We put an edge in Γ between two vertices if one the following holds: (1) there is some regular or blue 2-cell R of X such that these vertices correspond to opposite 1-cells in the boundary of R (if there are several such 2-cells we put as many edges in Γ (2) there is some red 2-cell R of X such that one of this vertices corresponds to distinguished edge e of R and the second corresponds to non distinguished edge of R which is not opposite to e. The 2-cell R is the 2-cell of X containing the edge. There is a natural map ϕ from Γ to X, which sends each vertex of Γ to the midpoint of the corresponding 1-cell of X and each edge of Γ to a segment joining two points in the boundary 2-cell R. A red hypergraph in X is a connected component of Γ. The 1-cells of X through which a hypergraph passes are dual to it. The red hypergraph Λ embeds if ϕ is an embedding from Λ to X. We call the subdiagram of X consisting of all open 2-cells containing edges of the red hypergraph Λ the carrier of Λ and denote it by Car(Λ). The hypergraph segment in X is a finite path in a hypergraph immersed into X. The carrier of a segment Λ′ denoted Car(Λ′ ) is the diagram consisting of all open 2-cells containing edges of the segment Λ′ . CUBULATING RANDOM GROUPS IN THE SQUARE MODEL 11 If in the above definition we replace every occurrence of the word “red” by the word “blue” we obtain the definition of the blue hypergraph. The comparison of the hypergraphs introduced above is presented in Figure 2. Λst blue 2-cell Λst blue 2-cell Λred red 2-cell Λblue red 2-cell a) b) Figure 2. a) The comparison of a red hypergraph Λred and a standard hypergraph Λst , b) The comparison of a blue hypergraph Λblue and a standard hypergraph Λst . 3.1. Hypergraphs are embedded trees. Now we are going to prove Theorem 3.7. In the square model at density d < e are embedded trees. hypergraphs in X 1 3 w.o.p. all red and blue To investigate standard and red hypergraphs we need a notion of a collared diagram. Our definition is a slight modification of the one introduced by Ollivier and Wise in [OW11, Definition 3.2]. Definition 3.8. (collared diagram) Let D be a Van Kampen with the following properties: (1) there is an external 2-cell C called a corner of D e of a hypergraph (standard, red or blue) of (2) there is a segment λ → D → X length at least 2 (3) the first and the last edge of λ lie in C, and no other edge lies in C (4) λ passes through every other external 2-cell of D exactly once (5) λ does not pass through any internal 2-cell of D We call D a collared diagram and λ a collaring segment. We also say that D is cornerless if the first and the last edge of λ coincide in C (in which case the hypergraph cycles). Now we will use the additional color-structure of a Painted Cayley complex to perform the following procedure of adding horns to a diagram. e Definition 3.9 (adding horns procedure). Let D be a van Kampen diagram in X. Let ∂2 (D) be the set of all 2-cells in D that contain a boundary edge of D. Let ∂2s (D) be the set of all distinguished 2-cells c in 2 − ∂(D) such that a 2-cell c′ , that is strongly adjacent to c, does not belong to D. 12 TOMASZ ODRZYGÓŹDŹ We define a diagram with horns D in the following way: for each c ∈ ∂2 (D) we glue to a diagram D a 2-cell c′ that is strongly adjacent to c (gluing is in a way that e these 2-cells are glued in X). If that will be clear we will call diagrams with horns just diagrams, for shortness. ˜ Lemma 3.10. Let D be a diagram with horns. Then |∂(D)| (the generalized boundary length) is an even number. Proof. The statement results from the fact that every 2-cell of D has even number of edges and every gluing of 2-cells along an edge decreases the generalized boundary length by 2.  Remark 3.11. A diagram with horns is a union of a disc diagram D and a set Z of 2-cells, each of them being adjacent to the boundary of D along at least one vertex and such that no two elements of Z share an edge. Proof. The only thing that we need to prove is that no two elements of Z share an edge. This results by Proposition 3.1.  Corollary 3.12 (of Remark 3.11). Collared diagrams with horns w.o.p. satisfy the assumptions of Theorem 2.5. The following theorem shows the relation between collared diagrams and hypergraphs Theorem 3.13 (reformulation of [OW11, Theorem 3.5]). Let Λ be a hypergraph in e The following conditions are equivalent: X. (1) Λ is an embedded tree. (2) There is no collared diagram collared by a segment of Λ. There is a definition of a diagram collared by a hypergraph in [OW11, Definition 3.11]. However, in the paper [OW11] this definition was stated for a general l-gon complex and for the standard hypergraphs that join antipodal midpoints of faces they are passing through. Definition 4.3 is a modification of [OW11, Definition 3.11]: we use it for l = 4 and for hypergraphs defined in a different way, but the e is not necessary fact the hypergraph joins the antipodal midpoints of edges in X to provide the proof of Theorem 3.13, in fact the proof can be repeated, as well, in case of colored hypergraphs. The next theorem is motivated by Theorem 3.13 and shows the relation between collared diagrams with horns and colored hypergraphs. e If Theorem 3.14. Let Λ be a red or blue hypergraph in X. Λ is not an embedded tree, then there is a collared diagram with horns collared by a segment of Λ. Proof. By Theorem 3.13 we know that there exists collared diagram collared by a segment of Λ. We obtain the statement of Theorem 3.14 by applying the procedure of adding horns to a diagram described in Definiton 3.9  Proof of Theorem 3.7. Suppose, on the contrary, that there is a red or blue hypergraph that is not an embedded tree. By Theorem 3.14 we know that there is a ˜ collared diagram with horns E collared by some segment λ. Let n = |∂E|. For
1 ε < 2 3 − d by Theorem 2.5 we have w.o.p. CUBULATING RANDOM GROUPS IN THE SQUARE MODEL 13 ˜ ≥ 4|E|(1 − 2d − ε) > 4 |E| n = |∂E| 3 There are two possibilities: either E is cornerless or not. First consider the case where E is cornerless. Then every external 2-cell, which is not red or blue has exactly one external edge. Each pair of strongly adjacent 2-cells gives a contribution ˜ equal 2, independently of the way the blue or red 2-cell is glued to the to |∂E| diagram E. Hence |E| ≥ n. By (3.1) we know that with overwhelming probability all collared cornerless diagrams with horns satisfy: (3.1) 4 n, 3 which is a contradiction. Therefore, with overwhelming probability there are no such diagrams. Let us now consider the case where the diagram E is not cornerless. The only difference between the previous case is that a corner can have two external edges. Therefore |E| ≥ n − 1. We have two possibilities |E| ≥ n or |E| = n − 1. If |E| ≥ n we again obtain (3.2), which is a contradiction. Hence, the only remaining case is ˜ = n and |E| = n − 1. Again we use (3.1) to obtain: where |∂E| (3.2) n> 4 (n − 1). 3 It can be easily seen that for n > 3 this is not possible. So we only have to exclude ˜ = 3. But there are no diagrams with odd generalized the diagram |E| = 2, |∂E| boundary length, according to Lemma 3.10.  n> Lemma 3.15. Suppose that a red hypergraph or a blue hypergraph Λ is an embedded e Then X e − Λ consists of two connected components. tree in X. Proof. The proof is analogous to the proof of [OW11, Lemma 2.3].  4. Metrically proper action on a space with walls In this section our goal is to prove the following 3 . Let Theorem 4.1. Let G be a random group in the square model at density d < 10 e e X be the Cayley complex of G. For x, y ∈ X we denote by dwall (x, y) the number of hypergraphs Λ (standard, red and blue) such that x and y lie in the different e − Λ. Then w.o.p. connected components of X 1 d e (1) (x, y)⌋, 15 X e denotes the edge metric on the 1-skeleton of X. dwall (x, y) ≥ ⌊ where dXe (1) The meaning of the foregoing theorem comes from the following Proof of Theorem 1.2. By Lemma 3.15 hypergraphs are embeddes trees, and by Lemma 3.15 and [OW11, Lemma 2.3] they separate the Cayley complex into two connected components, so they provide a structure of a space with walls on it. Every discrete group acts properly on its Cayley complex, and by Theorem 4.1 the wall metric and edge metric on the Cayley complex are equivalent. Therefore a 14 TOMASZ ODRZYGÓŹDŹ group acts properly on a space with walls. Then by [CN05] there is an action of a random group on a CAT(0) cube complex.  Before we provide the proof of Theorem 4.1 we need several facts about the geometry of hypergraphs. For simplicity, until the end of this section we will denote e the painted Cayley complex of a random group in the square model at density by X 3 e (1) the 1-skeleton of X e (i.e. the Cayley graph of a random group). and by X d < 10 e let E be a 2-cell having exactly Lemma 4.2. Let D be a distinguished 2-cell in X, one common edge with D and let F be a 2-cell different than D and having at least one common edge with E. Then w.o.p. E and F are regular 2-cells. Proof. Suppose, on the contrary, that E is a distinguished 2-cell. Then let D′ and E ′ be the 2-cells strongly adjacent to D and E respectively. By the assumption we know that D′ 6= E and E ′ 6= D (otherwise D and E would be strongly adjacent). By Proposition 3.1 we know that D′ 6= E ′ . Therefore there are two pairs of distinguished 2-cells {D, D′ } and {E, E ′ } which have at least one common edge. Consider the diagram D consisting of 2-cells D, D′ , E, E ′ , glued in the way they ˜ ≤ 6. Note, that every diagram A consiste We will show that |∂D| are glued in X. ˜ = 4. Gluing two ing of two strongly adjacent 2-cells satisfies: |A| = 4 and |∂A| such diagrams, consisting of disjoint sets of 2-cells, along one edge decreases the ˜ ≤ 6. Therefore, the diagram D violates generalized boundary length by 2, so |∂D| ˜ ≥ |D|4(1 − 2d), so by Theorem 2.5 w.o.p. there is no such a the inequality |∂D| diagram D. Therefore E is a regular 2-cell. Now suppose, on the contrary, that F is a distinguished 2-cell. Let F ′ be the 2-cell strongly adjacent to F . Again by Propistion 3.1, shows that F ′ 6= D, D′ 6= F and F ′ 6= D′ . Moreover, from the previous part of the proof we know that with w.o.p. E is a regular 2-cell, so F ′ 6= E. Hence, there is a diagram E, which is a union of a pair {D, D′ } of strongly adjacent 2-cells, 2-cell E, and a second pair of strongly adjacent 2-cells {F, F ′ }. Analogously as above, we can prove that E ˜ ≤ 8. Therefore the diagram E violates the inequality satisfies: |E| = 5 and |∂E| ˜ |∂E| ≥ |E|4(1 − 2d), so by Theorem 2.5 w.o.p. there is no such a diagram E. Hence F is a regular 2-cell.  Now, we introduce the notion of a diagram collared by hypergraph and path. Definition 4.3. Let λ be a segment of a hypergraph (red, blue or standard) and e (1) . Let D be a van Kampen diagram with the following let γ be an edge-path in X additional properties: (1) λ passes only through external 2-cells of D (2) every external edge of D belongs to the boundary of the carrier of λ or belongs to γ. We call D a diagram collared by hypergraph and path. After applying the procedure of adding horns(see Definition 3.9) to D we obtain a diagram with horns collared by hypergraph and path. We call λ a collaring hypergraph segment and γ a collaring edge-path. There is a definition of a diagram collared by hypergraphs and paths in [OW11, Definition 3.11]. In the paper [OW11] this definition was stated for a general polygonal complex and for the hypergraphs that join antipodal midpoints of faces they are passing through. Definition 4.3 is a modification of [OW11, Definition 3.11]: we CUBULATING RANDOM GROUPS IN THE SQUARE MODEL 15 use it for l = 4, one path and one hypergraph segment, and for hypergraphs defined in a different way, but the fact the hypergraph joins the antipodal midpoints is not necessary to provide the proofs of upcoming lemmas. The proofs of lemmas from [OW11] that we will use can be as well repeated for red and blue hypergraphs. The following lemma is then a slight modification of [OW11, Lemma 3.17]. e and γ be an edge path in X e (1) . Suppose, Lemma 4.4. Let Λ be a hypergraph in X that γ intersects Λ in two consecutive points x, y. Then there is a diagram with horns collared by Λ and a segment of γ bounded by points x and y (here γ is an edge path which starts and ends at“midedge vertices” corresponding to vertices of Λ. Proof. First we prove that there exists a diagram collared collared by Λ and a segment of γ bounded by points x and y. The proof of this fact is analogous to the proof of [OW11, Lemma 3.17]. Then we apply a procedure of adding horns (Definition 3.9) to that diagram.  e (1) and let Λ be a hypergraph in X. e Lemma 4.5. Let γ be a geodesic edge-path in X Suppose that there is a finite, nonempty geodesic segment γ ′ ⊂ γ, such that the first and the last vertex of γ ′ lie in Car(Λ) but no other vertices of γ ′ lie in Car(Λ) and no edge of γ ′ lie in Car(Λ). Then w.o.p. γ ′ has length 2, and there is the following e diagram D consisting of three 2-cells immersed in X: γ′ Λ′ Figure 3. The diagram D. Moreover, every 2-cell of D is a regular 2-cell. Definition 4.6. A diagram D having the same shape as the one presented in Figure 4.5 will be called a house diagram. The geodesic segment γ ′ we be called a roof (of a house diagram). Proof. Let e0 , e1 be some edges dual to Λ with endpoints in the first and the last vertex of γ ′ . By Lemma 4.4 we know that there is a diagram D with horns collared by the edge-path {e0 , γ ′ , e1 } and a segment Λ′ of Λ. Suppose that the length of this hypergraph segment equals k and the length of γ ′ equals l. Let k ′ denote the number of horns in D. Note that ˜ ≤ k + k ′ + l + 2. (4.1) |∂D| Note that every edge of γ ′ lies in the boundary of some 2-cell in D that is not in the carrier of Λ, since otherwise there would be an inner vertex of γ ′ lying in Car(Λ). Moreover, more than 2 edges of γ ′ cannot lie in the boundary of one 2-cell in D, 16 TOMASZ ODRZYGÓŹDŹ since otherwise γ ′ would not be a geodesic. Therefore, there are at least in D not lying in Car(Λ). Hence   l ′ . (4.2) |D| ≥ k + k + 2 l 2 2-cells Moreover, l ≥ 2, because there is at least one vertex of γ ′ not lying in the carrier of Λ and l ≤ k because γ is a geodesic. By Lemma 3.10 the number k + k ′ + l must be even. Combining this with Theorem 2.5 gives the following system of conditions, which is w.o.p. satisfied by k, k ′ and l: (4.3)   l 
8 ′ ′  k + k + l + 2 > 5 k + k + 2 l ≤ k, l ≥ 2, k ≥ 1 ⇐⇒ 2 ≤ l ≤ k   k + k ′ + l is an even number Now we will find the possible values of l (see 4.3). If l is even then we can rewrite the first inequality in (4.3) in form k + k ′ + l + 2 > 8 l ′ ′ 5 (k + k + 2 ). This is equivalent to 3(k + k ) < 10 + l. Combining this with the second inequality in (4.3), that is l ≤ k, we obtain k < 5. Therefore l ≤ 4, since l is even. Hence for even l possible solutions are: l = 4 and l = 2. If l is odd then the first inequality in (4.3) becomes k +k ′ +l+2 > 85 (k +k ′ + l+1 2 ), which is equivalent to 3(k + k ′ ) < l + 6. Combining this with l ≤ k we get k < 3. Therefore l ≤ 1 since l is odd. This contradicts l ≥ 2. Hence, there are no solutions with odd l. First, consider the case, where l = 2. Then 8 k + k ′ + 4 > (k + k ′ + 1), 5 ′ so k + k < 4. Combining this with the parity of k + k ′ + l we obtain that only possible solution for l = 2 is k + k ′ = 2. Hence, there are two subcases to consider: k ′ = 0 and k ′ = 1. If k ′ = 1 then the segment Λ′ consists of a distinguished 2-cell c, which is strongly adjacent to some 2-cell c′ : c′ A c B Λ γ′ D C Figure 4. The diagram D with one of possible hypergraphs Λ and geodesic segments γ ′ . The geodesic segment γ ′ joins vertices A and C or B and D, since otherwise γ would not be geodesic. It can be easily seen that the 2-cell c′ also belongs to the carrier of Λ. Therefore γ ′ cannot run along the boundary path of c′ , since it has at least one vertex that lies outside the carrier of Λ. Hence, there must be at least one ˜ = 4 and |D| ≥ 3, which contradicts Theorem more 2-cell in D. This gives us |∂D| 2.5. CUBULATING RANDOM GROUPS IN THE SQUARE MODEL 17 The second subcase (of the case l = 2) is k ′ = 0 and k = 2. In that subcase Car(Λ′ ) consists of two regular 2-cells (see Figure 5 a) ). γ′ A B C A B Λ′ F E C Λ′ D F a) E D b) Figure 5. The diagram D. Note, that γ ′ can join only vertices A with C or D with F , since in all other cases γ ′ is either not a geodesic or intersects Λ, according to Lemma 3.15. After relabelling vertices, we can always obtain a diagram D presented in Figure 5 b), as desired. Moreover, every 2-cell in D is a regular 2-cell, since D does not have any horns. There are exactly three 2-cells in D since by 2.5 there are no diagrams with the boundary equal 6 consisting of more than three 2-cells. Now, consider the case where l = 4. Then 8 k + k ′ + 6 ≥ (k + k ′ + 2), 5 so k + k ′ ≤ 4. We start with the subcase, where k + k ′ = 4. We know that ˜ |∂D| ≤ k + k ′ + l = 10 and |D| ≥ k + k ′ + 2l = 6. By Theorem 2.5 and the fact ˜ ≤ 10 we obtain that |D| = 6 and |∂D| ˜ = 10. Let C = D ∩ Car(Λ). The that |∂D| diagram D − C consists of two 2-cells, call them A and B. Note that all edges of γ ′ belong to A ∪ B. Therefore 2-cells A and B have no common edges, since otherwise there will be no two points in A ∪ B within distance 4. This means, that there are three vertices of γ ′ belonging to C. This contradicts our assumption on γ ′ so there is no diagram with k + k ′ = 4 satisfying the assumptions of the Lemma 4.5. Now consider the case, where k + k ′ = 2 and l = 4. If k + k ′ = 2 then the carrier of Λ′ consists of maximally two 2-cells, so every pair of vertices in this carrier lies in the distance dXe (1) equal at most 3. Therefore γ ′ cannot join any vertices of the carrier of Λ′ . Hence, there is no diagram D with k + k ′ = 2 and l = 4. There are no other possibilities, because of the parity of k + k ′ + l.  Definition 4.7. Let λ be a segment of a hypergraph. A diagram consisting of two strongly adjacent 2-cells, such that λ passes through this pair will be called a divided tile (of λ) we. A 2-cell in the carrier of λ that is not a regular 2-cell but is not an element of a divided tile of λ will be called an isolated 2-cell. A tile is a divided tile or a regular 2-cell. Lemma 4.8 (hypergraphs do not osculate). Let Λ′ be a hypergraph segment in e (1) . Suppose that one of the following conditions is satisfied X 18 TOMASZ ODRZYGÓŹDŹ a) Λ′ is a segment of a red hypergraph and there are no isolated red 2-cells of Λ′ , b) Λ′ is a segment of a blue hypergraph and there are no isolated blue 2-cells of Λ′ , c) Λ′ is a segment of a standard hypergraph and there is no divided tile of Λ′ . Then w.o.p. holds (1) For every vertex v of Car(Λ′ ) there is only one edge in Car(Λ′ ) ending in v, dual to Λ′ that is not an internal edge of a divided tile (the situation presented in Figure 6 a) does not occur). In other words: if there is a vertex in Car(Λ′ ) being the endpoint of two edges dual to Λ′ then one of these edges must be an internal edge of a divided tile. e (1) such that its vertices lie in Car(Λ′ ) but e does (2) There is no edge e in X ′ not lie in Car(Λ ), see Figure 6 b). c1 v c2 e a) b) Figure 6. Two types of hypergraphs osculations. Proof. (1) Suppose, on the contrary, that such a point v exists. Then let e1 , e2 be the two edges, ending in v that are dual to the hypergraph segment Λ′ and that are none of them is an internal edge of a divided tile. Let D be a diagram collared by ˜ ≤ |D| + 2. Therefore, by Λ′ and an edge-path {e1 , e2 }. Such diagram satisfies |∂D| ˜ Theorem 2.5 we obtain |D| ≤ 3. Hence |∂D| ≤ 5, and combining this with Lemma ˜ ≤ 4. This by Theorem 2.5 gives us that that |D| ≤ 2. A 3.10 we get that |∂D| diagram consisting of two 2-cells and having the generalized boundary length at most 4 must be a pair of strongly adjacent 2-cells. This ends the proof in the case c). If Λ′ is red or blue hypergraph we need to provide further argument. Note, that both red and blue hypergraph join the antipodal midpoints of bundary edges of every tile. Hence, there are no two edges that are dual do Λ′ , have common end and non of them is an internal edge of a divided tile. This ends the proof in cases a) and b). Therefore there is a blue 2-cell and a red 2-cell in the carrier of Λ′ , which contradicts assumptions a), b) and c). (2) Let v1 and v2 be the ends of the edge e. For i = 1, 2 let ei be the edge dual to Λ′ ending in vi that is not an internal edge of a divided tile. Let E be a CUBULATING RANDOM GROUPS IN THE SQUARE MODEL 19 diagram with horns collared by Λ′ and the edge-path {e1 , e, e2 }. As previously, E satisfies |∂E| ≤ |E| + 2. Therefore by 2.5 |E| ≤ 3, so |∂E| ≤ 5. By Lemma 3.10 we obtain that in fact |∂E| ≤ 4, which, by previous observation, means that |E| ≤ 2. A diagram with two 2-cells and the boundary length equal 4 must be a pair of strongly adjacent 2-cells. In a diagram which is a pair of strongly adjacent 2-cells there is no edge which does not lie in Car(Λ) but its endpoints do. This end the proof in cases a), b) and c).  e (1) Lemma 4.9. Let Λ1 be a hypergraph segment and γ be a geodesic segment in X that intersects Λ1 in two points, which are the endpoints of γ. Let Λ2 6= Λ1 be a e intersecting γ in point x. hypergraph in X (1) Λ1 and Λ2 pass through a common 2-cell cx , that is Car(Λ1 )∩Car(Λ2 ) ⊇ cx . (2) If moreover Λ2 intersects γ in two different points x and y that do not lie in the boundary of the same 2-cell of Car(Λ1 ) then there are two different 2-cells in Car(Λ1 ) ∩ Car(Λ2 ). Proof. Part (1) Let e be an edge of γ dual to Λ2 . There are two possibilities • The edge e lies in the boundary of a 2-cell c ∈ Car(Λ1 ). In that case Λ2 is dual to the boundary edge of c, so cx := c belongs to Car(Λ2 ). See Figure 7 a). • The edge e does not lie in the boundary edge of a 2-cell c ∈ Car(Λ1 ). Since γ is a geodesic edge-path, by Lemma 4.5 combined with Lemma 4.8 we know that e lies in the boundary of a house diagram presented in Figure 7 b). It is clear from the Figure 7 b) that cx ∈ Car(Λ1 ) ∩ Car(Λ2 ). Part (2). If x and y lie in the boundaries of distinct 2-cells cx and cy belonging to Car(Λ) then it is clear that cx and cy both lie in Car(Λ1 ) ∩ Car(Λ2 ) (see Figure 7. Consider the case where x lies in the boundary of a 2-cell c1 in Car(Λ1 ) and y do not. Considering a hypergraph Λ1 dual to γ passing through y we will find (as in part (1) ) a regular 2-cell c2 that lies in Car(Λ1 ) ∩ Car(Λ2 ) and that there is no edge of γ in the boundary of c2 . Hence c2 6= c1 and c1 and c2 both lie in Car(Λ1 ) ∩ Car(Λ2 ). The last case to consider is when neither x nor y lie in the boundary of Car(Λ1 ). First, note that x and y cannot lie in the roof of the same house diagram since then there would be a self intersection of Λ2 (see Figure 7 b) ) which contardicts Theorem 3.7. Hence, x and y lie in the roofs of different house diagrams, which means that one 2-cell from each of those diagrams belongs to Car(Λ1 ) ∩ Car(Λ2 ). This means that there are at least two different 2-cells in Car(Λ1 ) ∩ Car(Λ2 ).  e Let γ be a geodesic edge-path in the Lemma 4.10. Let Λ be a hypergraph in X. e Suppose that γ intersects Λ in two consecutive points x and y. Let 1-skeleton of X. Λxy be the segment of Λ joining x and y. Then: (1) if Λ is a red hypergraph, there is an isolated red 2-cell c of Λxy , (2) if Λ is a blue hypergraph, there is an isolated blue 2-cell c of Λxy , (3) if Λ is a standard hypergraph, there is a divided tile of Λxy , such that the none of its internal edges belong to γ. Proof. The proof relies on the idea, that if a geodesic crosses a hypergraph two times it must be worthwile in terms of distance. If the hypergraph does not “turn” 20 TOMASZ ODRZYGÓŹDŹ cx cx Λ1 γ e Λ2 γ Λ1 x y x e a) b) Figure 7. Intersections of hypergraphs Λ1 and Λ2 . there is no reason for a geodesic to cross it two times. A colored hypergraph can “turn” only if there is an isolated 2-cell of the segment of hypergraph and a standard hypergraph “effectively turns” when meeting a pair of strongly adjacent 2-cells. Let now state a precise argument. First, consider the case where Λ is a red hypergraph. Suppose, on the contrary, that there is no isolated red 2-cell of Λxy . The the carrier of Λxy consists only of tiles (regular 2-cells or divided tiles) and every edge of Λxy joins an antipodal midpoints of boundary edges of every tile in Car(Λ′ ) i.e. the carrier of Λ′ has the shape illustrated in Figure 8. γ x′ x x′′ Λ′ ... y′ y y ′′ k Figure 8. The diagram collared by Λxy and γ. e containing x, and suppose that Let x′ and x′′ be the two ends of the edge of X ′ ′′ e x lies on γ between x and y. Let y and y be the two ends of the edge of X ′ ′ containing y, and suppose that y lies on γ between x and y. Let γ0 be the subpath of γ joining x′ and y ′ (see Figure 8). Let V be the set of vertices of γ ′ . Define ′ V ′ = {v ∈ V |there is at least one edge of γ ′ ending in v and not belonging to Car(Λ′ )}. For each pair s, t of two consecutive vertices in V ′ consider the subsegment of ′ ′ γ joining s and t and denote it by γst . By Lemma 4.8 part (2) we obtain that γst either (1) lies in the boundary of the carrier of Λ, or (2) has a length at least 2. ′ In the latter case, we know that γst is a roof of a house diagram. Therefore ′ without change of the total length of γ ′ we can replace γst with an edge-path of length 2, that lies in the boundary of Car(Λxy ) and does not cross Λxy . Hence, we can modify γ ′ to a path having the same length such that all its edges lie in the boundary of Car(Λxy ). From now γ ′ denotes modified path. For every ′ CUBULATING RANDOM GROUPS IN THE SQUARE MODEL 21 vertex v of γ ′ denote by ev the edge dual to Λ′ ending in v that is not an internal edge of a divided tile (this definition is unambiguous according to Lemma 4.8 part (1) ). We will prove now that the length d of γ ′ equals the number k of tiles of Λxy . Every edge of γ ′ lies in the boundary of exactly one tile so k ≥ d. Suppose, on the contrary, that k > d. This means that there must be a vertex of γ ′ that is a common vertex of more than two tiles. But in that case there are at least two edges dual to Λxy that are not internal edges of a divided tile. This contradicts e joining x′′ Lemma 4.8 part 1). Hence k = d. Consider now the edge-path γ ′′ in X ′′ and y and goes along the boundary of Car(Λxy ). The length of this edge-path is not greater then k since every tile of Λxy shares at most one boundary edge with γ ′′ . Hence there exists an edge-path joining x′′ and y ′′ of length k. Note that the segment of γ joining x′′ and y ′′ has length k + 2. This contradicts the fact that γ is a geodesic. If Λ′ is a blue hypergraph the proof is analogous. Consider now the case, where Λ is a standard hypergraph. In that case we can provide an analogous argument to prove that there must be a divided tile of Λxy . We only need to prove that none of the internal edges of a divided tile belong to γ. Suppose, on the contrary, that this is the case. By Lemma 4.4 there exists a planar diagram D collared by Λxy and γ. By definiton of a diagram collared by a hypergraph and path γ cannot contain any internal edge of D which ends the proof.  The definition of a 2-collared diagram in case of standard hypergraphs can be found in [OW11, Subsection 3.3]. In a nutshell this is a diagram collared by two hypergraphs (see Definition [OW11, Lemma 3.17]). Let us state this definition for a case of general hypergraphs. Definition 4.11 (2-collared diagram). Let λ1 and λ2 be two hypergraph segments. Let D be a van Kampen diagram with the following additional properties: (1) (2) (3) (4) There are two external 2-cells of D called corners λ1 and λ2 pass only through external 2-cells of D every external edge of D belongs to Car(λ1 ) ∪ Car(λ2 ) corners are the only 2-cells of D containing the edges of both hypergraph segments λ1 and λ2 . We call D 2-collared diagram (collared by λ1 and λ2 ). Lemma 4.12 (reformulation of [OW11, Theorem 3.12]). Suppose, Λ1 and Λ2 are e and that they intersect at least distinct hypergraphs that are embedded trees in X, twice. Let c be a common 2-cell of Car(Λ1 ) and Car(Λ2 ) Then there exists a 2collared diagram collared by segments Λ1 and Λ2 . Moreover we can choose c to be one of its corners. There is a definition of a 2-collared diagram in the beginning of [OW11, Section 3.11]. However this is the special case of the general definition of a diagram collared by hypergraphs and paths that was stated for a general polygonal complex and for the standard hypergraphs. Definition 4.11 is a modification of it: we use it for l = 4 and for general hypergraphs, but the fact that the hypergraph joins the antipodal e is not necessary to provide the proof of Lemma 4.12, in midpoints of edges in X fact the proof can be provided, as well, in the case of colored hypergraphs. 22 TOMASZ ODRZYGÓŹDŹ Figure 9. 2-collared diagram. Collaring segments are indicated with thick lines. e is a pair of strongly adjacent Lemma 4.13. W.o.p. every 2-collared diagram in X 2-cells. Proof. Let D be a 2-collared diagram (with horns) collared by hypergraphs Λ1 and Λ2 . Let k1 be the length of the collaring segment of Λ1 and k2 be the length of the collaring segment of Λ2 . Let h denote the number of horns in D. Each pair of strongly adjacent 2-cells gives a contribution to the generalized boundary length ˜ ≤ k1 + k2 + h. There are two 2-cells in Car(Λ1 ) ∩ Car(Λ2 ), so equal 2, hence |∂D| 3 |D| ≥ k1 + k2 + h − 2. Hence, by Theorem 2.5 for the density d < 10 we obtain that w.o.p.: ˜ > 8 |D| > 8 (k1 + k2 + h − 2). k1 + k2 + h ≥ |∂D| 5 5 ˜ is even. Therefore Hence k1 + k2 + h ≤ 5. By Lemma 3.10 we know that |∂D| ˜ |∂D| ≤ 4 and by (4.4) we obtain |D| ≤ 2. Hence, D must be a diagram consisting of two strongly adjacent 2-cells.  (4.4) Lemma 4.14. The following statement holds w.o.p.: Let Λ1 and Λ2 be two hypere Suppose that Λ1 and Λ2 intersect transversally in a 2-cell c, meaning graphs in X. that Car(Λ1 ) ∩ Car(Λ2 ) contains a regular 2-cell c and that Λ1 ∩ c 6= Λ2 ∩ c. Then Λ and Λ′ intersect only once. Moreover c is the only one 2-cell in Car(Λ1 ) ∩ Car(Λ2 ). Proof. Suppose, on the contrary, that there is another point of intersection. Then by Lemma 4.12 there is a 2-collared diagram with horns D collared by Λ1 and Λ2 and containing c. By Lemma 4.13 we know that every 2-cell of D is distinguished, so this is a contradiction with the assumption that c is regular. Now suppose, on the contrary, that there is a 2-cell d ∈ Car(Λ1 ) ∩ Car(Λ2 ) such that d 6= c. We know that there is only one point of intersection of Λ1 and Λ2 , so the only possibility is that d is a distinguished 2-cell of some color, and Λ1 and Λ2 are both hypergraphs of this color (see Figure 10). But then Λ1 and Λ2 intersect in the middle of the 2-cell strongly adjacent to d, so there are two points of intersection of Λ1 and Λ2 , which is a contradiction.  CUBULATING RANDOM GROUPS IN THE SQUARE MODEL Λ1 23 c Λ2 d Figure 10. Two 2-cells in Car(Λ1 ) ∩ Car(Λ2 ). e (1) . Let e be an edge in γ, and Lemma 4.15. Let γ be a geodesic edge-path in X denote by s its midpoint. Then either: (a) There is a hypergraph intersecting γ only in the point s. (b) There is a hypergraph Λ such that its segment Λsx starting in s intersects γ in two consecutive points: s and x such that there are at least two not sharing an edge distinguished 2-cells in Λsx and Λ does not intersect γ anywhere else between points s and x. Proof. If one of the hypergraphs (red, blue or standard) dual to the edge e intersects γ only once, the statement is satisfied. Hence, consider the case where all hypergraphs dual to the edge e intersect γ at least twice. Denote by x, y and z the e (1) ), in which three hypergraphs nearest points to s (in the edge-path metric in X x y z Λ , Λ and Λ dual to the edge e intersect γ. A priori, it may happen that some of these points are equal. Without loss of generality, we can assume that x and y lie in the same connected component of γ − s and that x is at least as far from s as y. Now, we will prove that y 6= x. Suppose, on the contrary, that x = y. This means that every time segments Λsx and Λsy split, they also rejoin at some point, since they end in the same point. Every situation where Λsx and Λsy split an rejoin gives us a 2-collared diagram collared by a subsegments of Λsx and Λsy . By Lemma 4.14 every 2-collared diagram is a pair of strongly adjacent 2-cells. This means that if Λsx and Λsy split they rejoin in the next 2-cell of the common carrier, see Figure 11. This means that Car(Λsx ) = Car(Λsy ) and that distinguished 2-cells Car(Λsx ) occur only in strongly adjacent pairs. blue 2-cell red 2-cell Λred Λblue Figure 11. The only possible way of splitting and rejoining of Λx and Λy . One of the hypergraphs Λx and Λy is red or blue, so by Lemma 4.10 in the carrier of Λsx exists an isolated distinguished 2-cell, so this is a contradiction. Therefore y 6= x. 24 TOMASZ ODRZYGÓŹDŹ Therefore, y lies on a geodesic edge-path that intersects Λsx twice. Hence, by Lemma 4.5 we know that y lies in the boundary of the carrier of Λsx (see Figure 12 a) or y lies the boundary of the house diagram F presented in Figure 12 b). c c1 Λy c2 Λx γsx Λx y γsx y F a) b) Figure 12. Two possible locations of y. The hypergraphs Λx and Λy cannot intersect transversely in a regular 2-cell c1 , since otherwise by Lemma 4.14 they would intersect only once in the middle of c1 , and we know that they intersect also in s. Hence y lies in the boundary of a distinguished 2-cell cy ∈ Car(Λsx ). Therefore the carriers of the segments Λsx and Λsy share a common distinguished 2-cell cy . Without loss of generality, we can suppose that cy is red (if it is blue the argument is analogous). If Λx or Λy is blue then, according to Lemma 4.10, its carrier contains an isolated blue 2-cell cb , so cb does not share an edge with cy . In that case the carrier of Λx or Λy contains a red 2-cell cy and some non-adjacent by an edge to it blue 2-cell cb , so we are done. The only cases left to consider are: (1) Λy is a red hypergraph and Λx is a standard hypergraph (2) Λx is a red hypergraph and Λy is a standard hypergraph. First consider the situation 1). If there are at least two red 2-cells in the carrier of Λsx , we are done. Hence, we can suppose that cy is the only isolated red 2-cell in the carrier of Λsx . In that case Car(Λsx ) has the shape presented in Figure 13. We know that y lies in the boundary of cy and y does not belong to Λsx . Hence y = y1 or y = y2 . Recall that points s, x and y lie on γ. If y = y2 then there is a segment of standard hypergraph bounded by s and y intersecting γ twice at its ends, such that there is only one distinguished 2-cell in its carrier. This contradicts Lemma 4.10 (3). If y = y2 we obtain analogous contradiction. Hence there must be at least one more isolated red 2-cell in the carrier of Λsx , so we are done, since two red 2-cells cannot share an edge. Now consider situation 2), that is Λx is a standard hypergraph and there is only one divided tile of Λsx and y belongs to this divided tile. In that case Car(Λsx ) has the shape presented in Figure 14. We know that y lies in the boundary of cy but y does not belong to Λsx . Moreover by 4.10 part (3) we know that y does not lie on an internal edge of a divided tile. Hence y = y3 or y = y4 . If y = y3 then there is a segment of a red hypegraph bounded by x and y intersecting γ twice with no isolated red 2-cell in its carrier. This contradicts Lemma 4.10 part (1). If y = y4 CUBULATING RANDOM GROUPS IN THE SQUARE MODEL 25 y1 s Λsx ... y2 x Figure 13. The carrier of Λsx in case where Λx is red. y3 s Λsx ... y4 x Figure 14. The carrier of Λsx in case where Λx is standard hypergraph. we obtain the contradiction analogously. Hence, there must be at least two not sharing an edge 2-cells in the carrier of Λsx .  e of length at Theorem 4.16. Let γ be a geodesic edge-path in the 1-skeleton of X least 21. Let E := {e−7 , e−6 , . . . , e0 , . . . , e7 } be the set of 15 consecutive edges of e intersecting γ exactly once and such that this γ. Then there is a hypergraph in X intersection is the midpoint of one of the edges in E. Proof. Denote the midpoint of the edge e0 by s. If one of the hypergraphs dual to the edge e0 intersects γ only once, we are done. If not, by Lemma 4.10 we know that there exists a hypergraph Λ, such that its segment Λsx intersects γ in two consecutive points of γ: s and x and that the segment Λsx contains at least two not sharing an edge distinguished 2-cells: ⌈1 and ⌈2 . We will consider two cases: (1) There are at least seven 2-cells in Car(Λsx ) (2) There are at most six 2-cells in Car(Λsx ). We will use slightly different methods to deal with each of these two cases. 4.1. Case 1: There are at least seven 2-cells in Car(Λsx ). Let c1 , c2 , c3 be the 3-th, 4-th and 5-th 2-cells in Λsx respectively counted from s. According to Lemma 4.2 at least one of them must be a regular 2-cell, call it c. Denote by Λc 26 TOMASZ ODRZYGÓŹDŹ the standard hypergraph passing through 2-cell c and dual to Λ. By Lemma 4.14 Λc and Λ intersect only once. We will prove two assertions: (a) Λc intersects some edge in E. (b) Λc intersects γ exactly once. We start with the proof of assertion (a). Let us resolve where is the point of intersection of γ and Λc . If there is an edge ec of γ lying in the boundary of c then Λc and γ intersect in the midpoint of ec . In this case denote by xc the midpoint of e (1) that goes along the ec , see Figure 15 a). In that case there is an edge-path in X boundary of Car(Λ) of length at most 6, because there are at most five 2-cells in Car(Λsx ) between s and c. Hence dγ (s, xc ) ≤ 6. If there is no edge of γ lying in the boundary of c then, by Lemma 4.5 we know that the 2-cell c lies in the house diagram presented in Figure 15 b) so Λc intersects the edge ec from Figure 15 b). In that case we denote by the xc the midpoint of e (1) that goes along the boundary of ec . In this situation, there is an edge-path in X Car(Λ) and a boundary of a house diagram containing c of length at most 7, because there are at most five 2-cells in Car(Λsx ) between s and c. Hence dγ (s, xc ) ≤ 7. This ends the proof of assertion (a). c c Λc Λc Λ γ Λ ec xc γ xc ec a) b) Figure 15. Two possible ways of intersection of Λc and γ Now we will prove assertion (b). By Lemma 3.15 there are exactly two cone − Λ. Hence, we can split Λc into two parts: nected components of X e − Λ containing • Λ+ which is the intersection of Λc with the component of X xc e − Λ not con• Λ− which is the intersection of Λc with the component of X taining xc . We will show that Λ− does not intersect γ and that Λ+ intersects γ only once (in the point xc ). Firstly, consider Λ+ . Suppose, on the contrary, that Λ+ intersects γ in some point yc 6= xc . Suppose that yc is the nearest among what point to xc in the metric dγ . Points xc and yc cannot both lie in the boundary of c since there is at most one common edge of c and γ. We will consider three cases: (1) yc belongs to the segment γsx of γ bounded by s and x, , see Figure 16 (1) (2) yc does not belong to γsx and points xc and yc lie in the different connected e − Λ, , see Figure 16 (2) components of X CUBULATING RANDOM GROUPS IN THE SQUARE MODEL 27 (3) yc does not belong to γsx and points xc and yc lie in the same connected e − Λ. , see Figure 16 (3) component of X Λ− Λ (1) γ s c yc xc x Λ+ Λ− Λ (2) γ c s xc x yc Λ+ Λ− Λ (3) γ s c z2 xc z1 yc Λ+ Figure 16. Intersections of Λ+ and γ. 1)If yc belongs to the segment γsx of γ bounded by s and x. Then by Lemma 4.9 applied to Λ1 = Λ and Λ2 = Λc we obtain that there are two common 2-cells of the carriers of Λ and Λc , which is a contradiction with Lemma 4.14. e − Λ then 2) If points xc and yc lie in the different connected components of X there is an intersection of Λ+ which contradicts the fact that Λ and Λc intersect only once ( Lemma 4.14). e − Λ then Λ 3) If points xc and yc lie in the same connected component of X intersects γsx at least twice in some points z1 , z2 . If this two points do not lie in the boundary of the same 2-cell c′ in Car(Λc ) then ′ c 6= c. By Lemma 4.9 applied for Λ1 = Λc and Λ2 = Λ we obtain that there are two different 2-cells in Car(Λ) ∩ Car(Λc ), which contradicts Lemma 4.14. If z1 and z2 lie in the boundary of the same 2-cell c′ in Car(Λc ) then we know, that c′ 6= c, since c shares at most one edge with γ. Hence, c′ , c ∈ Car(Λc ) ∩ Car(Λ), which again contardicts 4.14. Hence, we proved that Λ+ does not intersect γ. Let us now consider Λ− . Suppose, on the contrary, that Λ− intersects γ. Let tc be the point of the intersection of Λ− and γ, which is the nearest to the point xc 28 TOMASZ ODRZYGÓŹDŹ in the edge-path metric on Λ− (see Figure 17). Note, that tc lies in the connected e − Λ not containing xc , since otherwise there will be a second component of X intersection of Λ and Λc which is not allowed, according to Lemma 4.14. Λ− Λ γ tc c s xc x Λ+ Figure 17. Intersection of Λ− and γ Let Λ′− be the hypergraph segment of Λc bounded by the points xc and tc . One of the points {x, s} lies on γ between points xc and tc . Suppose first that this point is s. Since s lies on the geodesic segment, that intersects Λc twice, by Lemma 4.5 we have two possibilities: s lies on the boundary of the 2-cell c (see Figure 18 a) ) or s lies in the diagram presented in Figure 18 b). c c Λ Λ Λc γ Λc x γ a) x b) Figure 18. Two possible locations of x Note, that in both cases the segment of Λ joining the point s and the 2-cell c consists of at most one 2-cell, which, if occurs, is a regular 2-cell. Remember that between s and x in Car(Λ) was at least seven 2-cells and c is 3-rd, 4-th or 5-th of them. Hence there must be a least two 2-cells between c and s. This is a contradiction, so Λ− does not intersect γ. If x lies on γ between xc and tc the reasoning is analogous. This ends the proof in Case 1. CUBULATING RANDOM GROUPS IN THE SQUARE MODEL 29 4.2. Case 2: There are at most six 2-cells in Car(Λsx ). Now let us consider the case, where there are at most six 2-cells in the carrier of Λsx . Let c be any regular 2-cell in the carrier of Λsx that lies between ⌈1 and ⌈2 (recall that ⌈1 and ⌈2 were two not sharing an edge distinguished 2-cells in the segment Λsx ). Again, we denote by Λc the hypergraph passing through c and dual to Λsx . Analogously we state two assertions: (a) Λc intersects some edge of the set E. (b) Λc intersects γ exactly once. To prove assertions (a) and (b) we first define the point xc as in Case 1. Proof of assertion (a) is analogous as in Case 1. Now, consider assertion (b). Again, we split Λc into two connected components: e − Λ containing • Λ+ which is the intersection of Λc with the component of X xc e − Λ not con• Λ− which is the intersection of Λc with the component of X taining xc . To prove that Λ+ intersects γ only once we can repeat the argument from Case 1. Consider now Λ− . Suppose, on the contrary that Λ− intersects γ in a point e − Λ not containing xc , so tc . Note that tc lies in the connected component of X the geodesic segment of γ bounded by xc and tc contains x or s. Without loss of generality we can assume that this geodesic segment contains x. Let Λ′− be the hypergraph segment of Λc bounded by the points xc and tc . The same reasoning, as in previous case, shows that there is at most one 2-cell in Car(Λ) between c and x, which, if occurs, is a regular 2-cell. This is a contraction since c lies in Car(Λ) between two distinguished 2-cells, so there must be a distinguished 2-cell in Car(Λ) between c and x.  Thorem 4.1 is an immediate consequence of Theorem 4.16. References [Ago13] Ian Agol, The virtual Haken conjecture, Doc. Math. 18 (2013), 1045–1087. With an appendix by Agol, Daniel Groves, and Jason Manning. [KK13] Marcin Kotowski and Michał Kotowski, Random groups and property (T ): Żuk’s theorem revisited, J. Lond. Math. Soc. (2) 88 (2013), no. 2, 396–416. [CMV04] Pierre-Alain Cherix, Florian Martin, and Alain Valette, Spaces with measured walls, the Haagerup property and property (T), Ergodic Theory Dynam. Systems 24 (2004), no. 6, 1895–1908. [CN05] Indira Chatterji and Graham Niblo, From wall spaces to CAT(0) cube complexes, Internat. J. Algebra Comput. 15 (2005), no. 5-6, 875–885. [Oll05] Yann Ollivier, A January 2005 invitation to random groups, Ensaios Matemáticos [Mathematical Surveys], vol. 10, Sociedade Brasileira de Matemática, Rio de Janeiro, 2005. , Some small cancellation properties of random groups, Internat. J. Algebra [Oll07] Comput. 17 (2007), no. 1, 37–51. [Gro93] M. Gromov, Asymptotic invariants of infinite groups, Geometric group theory, Vol. 2 (Sussex, 1991), London Math. Soc. Lecture Note Ser., vol. 182, Cambridge Univ. Press, Cambridge, 1993, pp. 1–295. [Oll07] Yann Ollivier, Some small cancellation properties of random groups, Internat. J. Algebra Comput. 17 (2007), no. 1, 37–51. 30 TOMASZ ODRZYGÓŹDŹ [Oll04] Y. Ollivier, Sharp phase transition theorems for hyperbolicity of random groups, Geom. Funct. Anal. 14 (2004), no. 3, 595–679. [OW11] Yann Ollivier and Daniel T. Wise, Cubulating random groups at density less than 1/6, Trans. Amer. Math. Soc. 363 (2011), no. 9, 4701–4733. [MP15] John M. Mackay and Piotr Przytycki, Balanced walls for random groups, Michigan Math. J. 64 (2015), no. 2, 397–419. [Odr16] Tomasz Odrzygóźdź, The square model for random groups, Colloq. Math. 142 (2016). , Nonplanar isoperimetric inequality for random groups, Note, aviable at: [Odr14] http://students.mimuw.edu.pl/∼to277393/web/files/nonplanar.pdf (2014). Institute of Mathematics, Polish Academy of Sciences, Warsaw, Śniadeckich 8 E-mail address: tomaszo@impan.pl 

4math.GR

Average Age of Information with Hybrid ARQ under a Resource Constraint Elif Tuğçe Ceran, Deniz Gündüz, and András György arXiv:1710.04971v1 [cs.IT] 1 Oct 2017 Department of Electrical and Electronic Engineering Imperial College London Email: {e.ceran14, d.gunduz, a.gyorgy}@imperial.ac.uk Abstract—Scheduling of the transmission of status updates over an error-prone communication channel is studied in order to minimize the long-term average age of information (AoI) at the destination, under an average resource constraint at the source node, which limits the average number of transmissions. After each transmission, the source receives an instantaneous ACK/NACK feedback, and decides on the next update, without a priori knowledge on the success of the future transmissions. The optimal scheduling policy is studied under different feedback mechanisms; in particular, standard automatic repeat request (ARQ) and hybrid ARQ (HARQ) protocols are considered. Average-cost reinforcement learning algorithms are proposed when the error probabilities for the HARQ system are unknown. I. I NTRODUCTION We consider a source node which continually communicates the most up-to-date status packets to the destination. In particular, we are interested in the age of information (AoI) [1], [2] at the destination, for a system in which the source node samples an underlying time-varying process and sends the sample values over an imperfect link which introduces delays. Our goal is to minimize the average AoI taking into account packet retransmissions. Most of the prior work on AoI either do not consider retransmissions, or consider a standard automatic repeat request (ARQ) protocol, where, if a packet cannot be decoded, it is retransmitted until correct decoding, up to a maximum number of retransmissions (e.g., [3], [4]). Here, we analyze the AoI for both the standard ARQ and hybrid ARQ (HARQ) protocols. In the latter, the receiver combines information from all previous transmission attempts of a packet in order to increase the success probability of decoding after each retransmission [5]. To address the trade-off between the success probability and the freshness of the status update to be transmitted, we develop scheduling policies to minimize the expected average AoI under an average transmission-rate constraint. This constraint is motivated by the fact that sensors sending status updates have usually limited energy supplies (e.g., are powered via energy harvesting); hence, they cannot afford to send an unlimited number of updates, or increase the signal-to-noise-ratio in the transmission. First, we assume that the success probability before each transmission attempt is known (which depends on the number of previous unsuccessful transmissions); hence, the source can judiciously decide when to retransmit, and when to discard failed information and send a fresh update. Then, we consider transmitting status updates over an unknown channel, Process H(t) ACK/ NACK feedback Sample of the process Source Error-prone channel Destination Figure 1. System model of a status update system over an error-prone pointto-point link in the presence of ACK/NACK feedback from the destination. in which the success probabilities of transmission attempts are not known a priori, and must be learned in an online fashion. The main contributions of this paper are outlined as follows: • An long-term average resource constraint is imposed on the transmitter, which limits the average number of transmissions. • Both retransmissions and pre-emption following a failed transmission are considered, corresponding, respectively, to the HARQ and ARQ protocols. • The optimal preemptive transmission policy for the standard ARQ protocol is shown to be threshold-type, and the optimal threshold value is derived in closed-form. • An average-cost reinforcement learning (RL) algorithm is proposed to learn the optimal scheduling decisions when the transmission success probabilities are unknown. • Extensive simulations are conducted in order to show the effect of feedback, resource constraint and ARQ/ HARQ error correction algorithms on the data freshness. A. Related Work Earlier work on AoI considered queue-based models, in which the status updates arrive at the source node randomly as a memoryless Poisson process, and are stored in a buffer before being transmitted to the destination [1], [2], [6]. Instead, in this paper, we follow the so-called generate-at-will model that has been recently considered in [4] and [7], in which the status updates of the underlying process of interest can be sampled and generated at any time by the source. A constant packet failure probability for status update systems is investigated for the first time in [3], but this work focuses on an M/M/1 queuing model, and no feedback is considered for retransmissions. The paper [4] considers broadcasting of status updates to multiple clients over an unreliable broadcast channel. However, this paper only considers index policies, which update the information at every time slot, since no constraint is imposed on the number of updates. Optimizing the scheduling decisions in an AoI system is also investigated in [7], focusing on multiple receivers, which leads to a different set of constraints than ours, as we limit the average number of transmissions to a single receiver. The AoI in the presence of HARQ is modeled through an M/G/1 queue in [8]; however, no resource constraint is taken into account, and the status update arrivals are assumed to be memoryless and random, in contrast to our work, which considers general and controlled status update generation. The rest of the paper is organized as follows. In Section II, the system model is presented, and the problem of minimizing the average AoI with HARQ under a resource constraint is formulated as a constrained Markov decision process (CMDP). In Section III, a primal-dual algorithm is proposed to solve this CMDP. Then, minimizing the AoI under the standard ARQ protocol is investigated in Section IV, and a computationally efficient solution is proposed. Section V introduces a reinforcement learning algorithm to minimize AoI in an unknown environment. Simulation results are presented in Section VI, and the paper is concluded in Section VII. II. S YSTEM M ODEL AND P ROBLEM F ORMULATION We consider a slotted status update system over an errorprone wireless link. The source monitors an underlying timevarying process, for which it is able to generate a status update at the beginning of each time slot; this is known as the generate-at-will model [4]. A transmission attempt of a status update takes constant time, which is assumed to be equal to the duration of one time slot. We assume that the channel state changes randomly from one time slot to the next in an independent and identically distributed fashion. We further assume the availability of errorand delay-free single-bit feedback from the destination to the source node. Successful reception of a status update is acknowledged by an ACK signal, while a NACK signal is sent in case of a failure. In the classical ARQ protocol, a packet is retransmitted after each NACK feedback, until it is successfully decoded, and the received signal is discarded after each failed transmission attempt. However, in the AoI framework there is no point in retransmitting a failed out-ofdate status packet if it has the same error probability with a fresh status update. Hence, the source always removes a failed status signal, and transmits a fresh status update (preemption). On the other hand, in the HARQ protocol, signals from all previous transmission attempts are combined for decoding; and therefore, the probability of error decreases with every retransmission. In general, the error probability of each retransmission attempt depends on the particular combining technique used, as well as the channel conditions [5]. The AoI is defined as the time elapsed since the most upto-date packet at the destination is generated. Assume that the most up-to-date packet at the destination at time t has a time stamp of generation U (t), then the AoI at time t, denoted by aging due to erronous transmission aging due to resource constraint minimum AoI level t at i i n x x n x i n i i i n δt rt 1 2 3 4 5 3 4 2 3 1 1 2 3 0 0 0 1 2 0 1 0 0 0 0 0 0 Feedback NACK NACK ACK NACK ACK ACK Figure 2. Illustration of the AoI in a slotted status update system with HARQ. δt , is defined as: δt , t−U (t). Therefore, the AoI increases by one when a transmission fails, while it decreases to one (or, to the number of retransmissions in the case of HARQ) when a status update is successfully decoded. The system model and the evolution of AoI for a slotted status update system are illustrated in Figures 1 and 2, respectively. The probability of error after r retransmissions, denoted by g(r), depends only on r, and the particular HARQ scheme used for combining multiple transmission attempts. An empirical method to estimate g(r) is presented in [9]. As in any reasonable HARQ strategy, g(r) is non increasing in the number of retransmissions r, i.e., g(r1 ) ≥ g(r2 ) for all r1 ≤ r2 . To simplify the analysis, we assume that there is a maximum number of retransmissions rmax , and a packet is assumed to be successfully decoded after rmax transmissions, i.e., g(r) = 0, for all r ≥ rmax . Let δt ∈ Z+ denote the AoI at time slot t, and rt ∈ {0, . . . , rmax } denote the transmission attempt number. We define the state vector at time slot the t as st , (δt , rt ). At each time slot, the source node takes one of the three possible actions, denoted by a ∈ A, where A = {i, n, x}. It can i) remain idle (a = i); ii) generate and transmit a new status update packet (a = n); or, iii) retransmit the previously failed packet (a = x). The state transitions can be summarized as follows: The next state (δt+1 , rt+1 ) is determined stochastically depending on action at and the current state (δt , rt ): Pr(δt+1 = δt + 1, rt+1 = rt |at = i, δt , rt ) = 1 (1a) Pr(δt+1 = δt + 1, rt+1 = 1|at = n, δt , rt ) = g(0) (1b) Pr(δt+1 = 1, rt+1 = 0|at = n, δt , rt ) = 1−g(0) (1c) Pr(δt+1 = δt + 1, rt+1 = rt + 1|at = x, δt , rt ) = g(rt ) (1d) Pr(δt+1 = rt + 1, rt+1 = 0|at = x, δt , rt ) = 1−g(rt ) (1e) and P r(δt+1 , rt+1 |at , δt , rt ) is zero elsewhere. Note that if there is no resource constraint, remaining idle for a time slot is not a favorable action since it increases the AoI by 1. However, continuous transmission is typically not possible in practice due to energy or interference constraints. Thus, we impose a constraint on the average number of transmissions denoted by Cmax . This leads to the CMDP formulation, defined by the 5tuple {S, A, P, c, d} [10]. A countable set of states (δ, r) ∈ S and a finite set of actions A = {i, n, x} are already defined. P contains the transition probabilities outlined in (1). The instantaneous cost ct at time slot t is equal to age δt , and we use dt as the transmission count, that is, dt = 0 if at = i, or dt = 1 otherwise. Policies are represented by π : S × A → [0, 1], π ∈ Π, where at maps states st ∈ S into actions at ∈ A with some probability π(at |st ). We will use sπt = (δtπ , rtπ ) and aπt to denote the states and actions, respectively, induced by policy π. Let J π (s0 ) denote the infinite horizon average age, and C π (s0 ) denote the expected average number of transmissions when policy π is employed with initial state s0 . We restrict our attention to stationary policies, and denote the set of feasible policies by ΠF , such that C π ≤ Cmax , ∀π ∈ ΠF . We can state the CMDP optimization problem as follows, where Eπ [·] represents the expectation with respect to policy π and error probabilities g(r): Problem 1. " T # 1 π X π (2) δt s0 , Minimize J (s0 ) , lim sup E T →∞ T t=1 # " T 1 π X π π s.t. C (s0 ) = lim sup E 1[at 6= i] s0 ≤ Cmax . T →∞ T t=1 π ∗ We define a policy π ∗ to be optimal in ΠF if J ∗ , J π ≤ π J for all π ∈ ΠF . A policy is said to be deterministic if it chooses an action with certainty at each decision epoch, and for convenience we denote a deterministic policy simply by π [10]. Also, without loss of generalization, problem horizon is assumed to begin when the information at the destination is up-to-date, i.e. s0 = (1, 0); and s0 will be removed from the notation for simplicity. We assume throughout this paper that the MDP is unichain [10], [11], which is intuitive since any reasonable policy for Problem 1 results in an ergodic Markov chain. A. Structure of the Optimal Policy Note that a deterministic stationary policy which is optimal for countable-state average-cost MDPs exist and can be found under certain conditions defined in [11], [12]. However, the optimal policies for constrained MDPs are no longer limited to the set of deterministic policies [10], [13], and randomized stationary policies should be considered. Since the CMDP defined in Problem 1 has a single global constraint, Corollary 1 below follows immediately from Theorem 4.4 of [10]. Corollary 1. An optimal stationary policy for the CMDP in Problem 1 exists, and it is a randomized mixture of two deterministic policies. The next proposition formalizes the simple observation that retransmitting a packet immediately after a failed attempt is better than waiting some slots and then retransmitting (since the waiting time just increases the age in case of a successful retransmission, and the age will be the same in case of a failure). The proof of the proposition is trivial, and hence omitted. Proposition 1. There exists an optimal policy for Problem 1 that takes a retransmission action only after a failed transmission event, that is P r(a∗t+1 = x|a∗t = i) = 0. Based on Proposition 1, the transition for rt in (1) can be simplified when at = n, and we have rt+1 = 0 instead of rt+1 = rt . In the next section, the methodology to solve Problem 1 will be described. III. P RIMAL -D UAL A LGORITHM TO M INIMIZE AO I To solve the average cost CMDP in Problem 1, we adopt the Lagrangian primal-dual method [10], [13]. Similar methods have been adopted in other wireless network problems including [14], [15]. A. Relaxed Unconstrained MDP Lagrangian relaxation of the constraint with non-negative multiplier η can be written as: ! " T # T 1 π X π 1 X π π 1[at 6= i] , Jη = lim E δt − η Cmax − T →∞ T T t=1 t=1 (3) and the optimal Jη∗ for a given η is defined as Jη∗ , minπ Jηπ . Following the methodology proposed in [10], for a constant η, the above minimization is equivalent to an unconstrained average-cost MDP, in which the instantaneous overall cost becomes δt + η 1[aπt ∈ {n, x}]. Then, the problem is to minimize Jη which is the averagecost for the unconstrained problem for a given η. There exits an optimal stationary deterministic policy, and thus, the optimality equations can be written. Let h(δ, r) denote the differential cost (bias) function, and ρ represent the gain, which is an approximation of the average overall cost; i.e., Jη∗ → ρ [11]. The differential cost function h(δ, r) : S → R is defined as the difference of the value function V (δ, r) with respect to an arbitrary but fixed reference state (δ ref , rref ) [11], i.e. h(δ ref , rref ) = 0, and ∀(δ, r) ∈ S,
h(δ, r) + ρ = min δ + η · 1[a 6= i] + Eπ [h(δ ′ , r′ )] , a∈{i,n,x} (4) ′ ′ where (δ , r ) denotes the next state when decision a is taken at state (δ, r) with a transition matrix given in (1. Equations in (4) are also known as the Bellman optimality equations [11]. For countable S and finite A, there exists a differential cost function h(δ, r), ∀(δ, r) ∈ S, and a unique gain ρ, which satisfy the optimality equations in (4). Moreover, for a given η value the optimal stationary deterministic policy π ∗ (δ, r) is given by:
πη∗ (δ, r) ∈ arg min δ + η · 1[a 6= i] + Eπ [h(δ ′ , r′ )] . (5) a∈{i,n,x} Note that the state space of possible status age values is possibly countably infinite. In order to write a practical algorithm that solves the Bellman equations in (4), we truncate to the state space S N = {1, 2, . . . , N } × {0, 1, . . . , rmax } which approaches to the original countable state CMDP as N → ∞ [16], [7]. Under any reasonable policy for Problem 1, probability of visiting a state with a large age value decreases exponentially whereas the cost incurred due to a large age is polynomial, so it is intuitive to truncate to a large N . The state transition function when at = i, or at ∈ {n, x} but the transmission fails can be written as δt+1 = ⌊δt + 1⌋N , where y = ⌊x⌋N implies y = x if x ≤ N , and y = N otherwise; which provides the finite state space approximation similarly to [16], [7]. B. Relative Value Iteration Value iteration (VI) algorithm can be used to solve countable-state finite-action MDPs [11]. However, for averagecost problems with Bellman equations defined as in (4), value functions in VI generally do not converge and differences between successive value iterations approximate to the true value of Jη∗ . Due to possible large values, it may result in numerical instability [11]. Therefore, the relative value iteration (RVI) algorithm [11] is adopted to solve (4). Starting with an arbitrary but a fixed reference state (δref , rref ), a single value iteration for the RVI algorithm [11] is given as follows: Qn+1 (δ, r, a) ← δ + η · 1[aπ 6= i] + Eπ [hn (δ ′ , r′ )] , (6) Vn+1 (δ, r) ← min(Qn+1 (δ, r, a)), (7) hn+1 (δ, r) ← Vn+1 (δ, r) − Vn+1 (δ ref , rref ), (8) a where Qn (δ, r, a), Vn (δ, r) and hn (δ ′ , r′ ) denote the state action value function, value function and differential value function for iteration n, respectively. C. Finding the Approximate Lagrange Parameter η ∗ and Randomization We start with an initial parameter η0 for m = 0, and at each iteration, run the RVI algorithm and compute ηm+1 as follows: ηm+1 ← ηm + α(Cηm − Cmax ). (9) We continue iteration until convergence, and find the optimal parameter η ∗ . Note that, it is not straightforward to obtain two different η values for two different deterministic policies, such that the randomization between the two results in an optimal stochastic policy. Thus, a heuristic method will be adopted: after obtaining η ∗ , it can be perturbed by small ξ, and the optimal policies for Lagrange parameters η ∗ + ξ and η ∗ − ξ are computed and randomized with probabilities µ and 1 − µ, respectively. The randomization probability µ can be computed using the average power consumptions of Cη∗ +ξ and Cη∗ −ξ , such that µCη∗ +ξ + (1 − µ)Cη∗ −ξ = Cmax . The average transmission bη∗ +ξ , C bη∗ −ξ can be computed by empirical estimation, rates C and randomization probability µ is given by: µ= bη∗ +ξ Cmax − C . bη∗ −ξ − C bη∗ +ξ C (10) IV. AO I WITH C LASSICAL ARQ P ROTOCOL Now, assume that the system adopts the classical ARQ protocol; that is, failed transmissions are discarded at the destination. In this case, there is no point in re-transmitting an undecoded packet since the successful transmission probabilities are the same for a retransmission and the transmission of a new update. The state space reduces to δ ∈ {1, 2, . . . , N } as rt = 0, ∀t, and the action space to A ∈ {n, x}. The probability of error in a status update is p , g(0). State transitions in (1), Bellman optimality equations in (4) and the RVI algorithm can all be simplified accordingly. Thanks to these simplifications, we are able to provide a closedform solution to the corresponding unconstrained MDP with Lagrange relaxation. Lemma 1. The optimal policy that minimizes Jη with the standard ARQ protocol admits a deterministic threshold structure characterized as follows for some threshold value ∆: ( n if δ ≥ ∆, π(δ) = (11) i if δ < ∆. Proof. Proof is not included here due to space limitations, but will be provided in the extended version of the paper. Theorem 1. Optimal threshold value for the standard ARQ protocol is given by: p 2η(1 − p) + p − p ∗ ∆ = . (12) 1−p Proof. Proof will be provided in the extended version of the paper. Note that the optimal threshold value in (12) depends on η, which stands for the trade-off between the average AoI and the constraint. We can compute the average transmission rate for a given η using the optimal threshold policy in (12), denoted by Cη . An approximate solution for η ∗ can be obtained by solving Cη = Cmax . Together with (12) we can obtain the threshold value for the CMDP as follows: p 1 − . (13) ∆∗ = Cmax (1 − p) 1 − p The threshold policy defined in (13) is the optimal deterministic policy for the standard ARQ case when randomization is not considered, which is not sufficient for the CMDP. Corollary 2. The optimal policy for the constrained average AoI problem with standard ARQ is a randomized mixture of two different threshold policies defined in Theorem 1. Since we have the threshold in closed-form for the optimal deterministic policy, which corresponds to η ∗ ; we can compute ∆1 , ∆2 directly by using ∆∗ , instead of finding η1 , η2 Algorithm 1: Average-cost modified SARSA with softmax 1 2 3 4 Input : Lagrange parameter η n ← 0 /* time iteration τ ← 1 /* softmax temperature parameter ×3 QN×M ← 0 , ρ ← 0 /* initialization η */ */ */ foreach n do foreach a ∈ A do 5 exp(−Qη (sn , a)/τ ) π(a|sn ) = X exp(−Qη (sn , a′ )/τ ) 6 a′ ∈A end Sample an from π(a|Sn ), observe next state sn+1 and cost cn = δn + η1{an =1,2} foreach a ∈ A do exp(−Qη (sn+1 , an+1 )/τ ) π(a|sn+1 ) = X exp(−Qη (sn+1 , a′n+1 )/τ ) 7 8 9 10 a′n+1 ∈A 11 12 13 14 15 16 17 18 end end Sample an+1 from π(an+1 |sn+1 ) Update Q√ η (sn , an ) as: αn ← 1/ n Qη (sn , an ) ← Qη (sn , an ) + αn [cn − ρ + Qη (sn+1 , an+1 ) − Qη (sn , an )] /* update ρ at every step */ ρ ← ρ + 1/n[cn − ρ] n← n+1 first, and then computing ∆1 , ∆2 accordingly. We propose a heuristic policy (optimal if (∆, C ∆ ) pairs form a convex set) in order to minimize the average AoI with standard ARQ by randomizing between the two policies corresponding to the nearest integers to ∆∗ in (13), that is ∆1 , ⌊∆∗ ⌋, ∆2 , ⌈∆∗ ⌉. Randomization coefficient µ ∈ [0, 1] can be computed by using the power consumptions C ∆j of the deterministic threshold policies corresponding to ∆j , j = 1, 2. From (13) we have: C ∆j , (1 − p)−1 (∆j + p/(1 − p))−1 . (14) Therefore, randomization factor µ of stochastic policy can be found using ∆j , j = 1, 2 as in 10, and the resultant algorithm is a threshold policy that uses ∆1 on the threshold with with probability µ and ∆2 with probability (1 − µ) at each slot. V. L EARNING TO MINIMIZE AO I IN AN UNKNOWN ENVIRONMENT Here we consider a practical scenario, in which the source node may not have a-priori information about the channel or the decoding probabilities at the destination, or these may change over time. Therefore, an online learning algorithm is employed to learn g(r) over time without degrading the performance significantly. The literature for average-cost reinforcement learning (RL) is quite limited compared to discounted cost problems. R-learning [17] is an average-cost RL algorithm which is an off-policy method similar to Q-learning [18], and initially proposed with a uniform exploration strategy (ǫgreedy ). For the average AoI minimization problem in Problem 1, a modified version of the R-learning algorithm is employed. In Figure 3. Deterministic policies πη−ξ (top) and πη+ξ (bottom) when Cmax = 0.4, p0 = 0.3, and rmax = 9, for a time horizon T = 1000. The results are averaged over 1000 trials. (Blue circles, red stars and green diamonds represent actions πη (δ, r) = i, n and x, respectively.) Algorithm 1 we use the softmax exploration strategy instead of ǫgreedy for the current and the next states [18]. Then, Algorithm 1 becomes an on-policy RL algorithm. Moreover, we update the gain ρ at every time slot based on the empirical average instead of updating at non-explored time slots and losing information. The resultant algorithm is called averagecost SARSA with softmax. VI. N UMERICAL R ESULTS Decoding error probability is assumed to be given by g(r) , p0 21−r , where p0 denotes the failure probability of the first transmission, and r is the retransmission count. The exponential behavior of the error probability follows from previous research on HARQ [5], [9]. In Figure 3, we illustrate the deterministic policies obtained by RVI and η search for given Cmax , rmax and p0 values. Final policies are generated by randomizing between πη−ξ and πη+ξ . As it can be seen from the figure, the designed policy transmits less as η increases, and vice versa. Figure 4 illustrates the performance of the proposed randomized HARQ policy with respect to Cmax for different p0 values. We also include the performance of deterministic and randomized threshold policies with ARQ. As expected, average AoI can be reduced by randomization between two deterministic threshold policies. We observe that average AoI decreases exponentially with respect to Cmax . The average AoI also decreases with decreasing p0 and with increasing rmax as expected. Figure 5 shows the evolution of the average AoI over time when the proposed average-cost SARSA learning algorithm is employed. It can be observed that the average AoI achieved by Algorithm 1 converges to the one obtained from the RVI algorithm which has a priori knowledge of g(r). Average AoI achieved by the proposed online learning algorithm is 12 12 p =0.5, ARQ, deterministic RL, p 0 =0.5, r max =3 0 p =0.5, ARQ, randomized RVI, p 0 =0.5, r max =3 p0=0.5, rmax=3, HARQ 0 8 p =0.1, r 0 max max RVI, p 0 =0.2,r max =9 =3, HARQ =9, HARQ 6 Average AoI p =0.3, r Average AoI 10 0 10 8 RL, p 0 =0.2, r max =9 6 4 4 2 2 0 0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Average number of transmissions, C max Figure 4. Expected average AoI J π with respect to Cmax for ARQ and HARQ protocols for different p0 and rmax values. Time horizon is set to T = 1000, and the results are averaged over 1000 trials. 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Average number of transmissions, C max Figure 6. Performance of the RL (average-cost SARSA) with respect to RVI Algorithm for n = 1000 iterations, and values are averaged over 1000 trials for different p0 and rmax values (both the mean and the variance are shown). R EFERENCES Figure 5. Performance of the average-cost SARSA for rmax = 9, p0 = 0.2, η = 10 and n ≤ 10000, averaged over 1000 trials (both the mean and the variance are shown). presented in Figure 6 as a function of Cmax , which shows that the performance is close to that of RVI. VII. C ONCLUSIONS We have considered a wireless system transmitting timesensitive data over an imperfect channel with the average AoI as the performance measure, which quantifies the timeliness of the data available at the destination. Considering both the classical ARQ and the HARQ protocols, preemptive scheduling policies have been proposed by taking into account retransmissions under a resource constraint. In addition to identifying a randomized threshold structure for the optimal policy when the error probabilities are known, an efficient RL algorithm is also proposed for practical applications, when the system characteristics may not be known in advance. The algorithms designed in this paper are relevant to other systems concerning the timeliness of information, and the proposed methodology can be used in other CMDP problems. As a future work, the problem will be extended to time-correlated channel statistics in a multi-user setting. [1] S. Kaul, M. Gruteser, V. Rai, and J. Kenney, “Minimizing age of information in vehicular networks,” in IEEE Coms. Society Conf. on Sensor, Mesh and Ad Hoc Coms. and Nets., June 2011, pp. 350–358. [2] S. Kaul, R. Yates, and M. Gruteser, “Real-time status: How often should one update?” in INFOCOM, Proc. IEEE, March 2012, pp. 2731–2735. [3] K. Chen and L. Huang, “Age-of-information in the presence of error,” in IEEE Int’l Symp. on Inf. Theory (ISIT), July 2016, pp. 2579–2583. [4] I. Kadota, E. Uysal-Biyikoglu, R. Singh, and E. Modiano, “Minimizing age of information in broadcast wireless networks,” in Annual Allerton Conf. On on Communication, Control, and Computing, September 2016. [5] P. Frenger, S. Parkvall, and E. Dahlman, “Performance comparison of harq with chase combining and incremental redundancy for hsdpa,” in IEEE Vehicular Technology Conf. Proc., vol. 3, 2001, pp. 1829–1833. [6] M. Kashef and A. Ephremides, “Optimal packet scheduling for energy harvesting sources on time varying wireless channels,” Journal of Communications and Networks, vol. 14, no. 2, pp. 121–129, April 2012. [7] Y. P. Hsu, E. Modiano, and L. Duan, “Age of information: Design and analysis of optimal scheduling algorithms,” in 2017 IEEE International Symposium on Information Theory (ISIT), June 2017, pp. 561–565. [8] E. Najm, R. D. Yates, and E. Soljanin, “Status updates through M/G/1/1 queues with HARQ,” CoRR, vol. abs/1704.03937, 2017. [9] V. Tripathi, E. Visotsky, R. Peterson, and M. Honig, “Reliability-based type ii hybrid arq schemes,” in Communications, 2003. ICC ’03. IEEE International Conference on, vol. 4, May 2003, pp. 2899–2903 vol.4. [10] E. Altman, Constrained Markov decision processes, ser. Stochastic modeling. Boca Raton, London: Chapman & Hall/CRC, 1999. [11] M. L. Puterman, Markov Decision Processes: Discrete Stochastic Dynamic Programming. New York, NY, USA: John Wiley & Sons, 1994. [12] L. I. Sennott, “Average cost optimal stationary policies in infinite state markov decision processes with unbounded costs,” Operations Research, vol. 37, no. 4, pp. 626–633, 1989. [13] ——, “Constrained average cost Markov decision chains,” Probability in Eng. and Informational Sciences, vol. 7, no. 1, p. 6983, 1993. [14] M. H. Ngo and V. Krishnamurthy, “Monotonicity of constrained optimal transmission policies in correlated fading channels with arq,” IEEE Trans. on Signal Processing, vol. 58, no. 1, pp. 438–451, Jan 2010. [15] A. Roy and A. Karandikar, “Optimal radio access technology selection policy for lte-wifi network,” in Int’l Symp. Modeling and Opt. in Mobile, Ad Hoc, and Wireless Nets. (WiOpt), May 2015, pp. 291–298. [16] E. Altman, R. E. Azouzi, D. S. Menasché, and Y. Xu, “Forever young: Aging control in dtns,” CoRR, vol. abs/1009.4733, 2010. [17] S. Mahadevan, “Average reward reinforcement learning: Foundations, algorithms, and empirical results,” Machine Learning, vol. 22, no. 1, pp. 159–195, 1996. [18] R. S. Sutton and A. G. Barto, Introduction to Reinforcement Learning, 1st ed. Cambridge, MA, USA: MIT Press, 1998. 

7cs.IT

Class-Splitting Generative Adversarial Networks Grinblat, G.L.1 , Uzal, L.C.1 and Granitto, P.M.1 arXiv:1709.07359v1 [stat.ML] 21 Sep 2017 1 CIFASIS, French Argentine International Center for Information and Systems Sciences, UNR-CONICET, Argentina Figure 1: Samples generated with our Splitting GAN method with supervised training on CIFAR-10 dataset. Each line has samples of one of the original classes. Each side has samples corresponding to one of the two clusters generated for each class. We use ResNet-B architecture (see text for details). Abstract when no class information is available, i.e. in the unsupervised setup. Our generated samples reach Generative Adversarial Networks (GANs) produce state-of-the-art Inception scores for CIFAR-10 and systematically better quality samples when class STL-10 datasets in both supervised and unsuperlabel information is provided., i.e. in the con- vised setup. ditional GAN setup. This is still observed for the recently proposed Wasserstein GAN formulaIntroduction tion which stabilized adversarial training and al- 1 lows considering high capacity network architectures such as ResNet. In this work we show how The irruption of Generative Adversarial Nets to boost conditional GAN by augmenting available (GAN) [5] produced a great leap for image data class labels. The new classes come from cluster- generation. Samples were generated simply by aping in the representation space learned by the same plying a neural network transform to an input ranGAN model. The proposed strategy is also feasible dom vector sampled from a uniform distribution. 1 There is no need for any Markov chains or unrolled approximate inference networks during either training or generation of samples [5]. GAN based generative models did not take long to reach impressive image quality [13, 14, 17, 10] at least for some specific datasets. However, current GAN models cannot produce convincing samples when trained on datasets of images with high variability, even for relatively low resolution images. On the other hand, it is observed that sample quality improves when class information is taken into account in a conditional GAN setup [11, 12]. These findings suggest that it is hard to learn a multimodal distribution from a smooth transform of a uniform (or Gaussian) distribution and that providing categorical class information to the generator alleviates this problem. Our proposal is inspired in two observations. First, as mentioned above, conditioning generation with categorical class labels with high level of abstraction improves image quality. Second, as early observed in [5, 13], the adversarial network pair learns an useful hierarchy of representations in an unsupervised setup. We propose to exploit the same representation space learned by the GAN model in order to generate new class labels with a high level of abstraction. This is done by applying a simple clustering method in this representation space. By conditioning generation with this new class labels the model is able to generate better samples. This can be done either when prior class information is available or not. The main contributions of the present paper are1 : Figure 2: Class-conditioned samples generated with WGAN-GP method [7] over the ResNet-B architecture for CIFAR-10 dataset. Each row has samples of a given class. art Inception scores for CIFAR-10 and STL-10 datasets in both tasks. 2 Background In the original GAN formulation [5], the generator is a neural network that transform noise input z ∼ p(z) into fake samples and the discriminator D is a neural network with a single scalar output with a sigmoid activation. This output is interpreted as the model probability for its input being a real image from the dataset distribution against being a fake image sampled by generator G. The discriminator D is trained using the standard binary classification formulation by minimizing the binary cross-entropy between fake and real distributions. On the other hand, the generator G is simultaneously trained to mislead the discriminator. This is accomplished in practice by updating G parameters minimising the same loss but with fake samples tagged with a ‘true’ label [5]. In other words, the discriminator is updated by • We propose a method for increasing the number of class labels during conditional GAN training based on clustering in the representation space learned by the same GAN model. We base our implementation on the more stable Wasserstein GAN formulation [2, 7]. • We show an effectively way of adapting networks architecture to handle this increasing number of classes. • We show that this splitting GAN strategy improves samples quality both in the supervised and unsupervised tasks, reaching state-of-the1 The source code will be available at https://github. com/CIFASIS 2 Figure 3: Real CIFAR-10 samples corresponding to the 20 clusters found by our method. Each line is divided in the same way as in Figure 1. gradient descent over a negative log likelihood loss of the generated samples by adding more structure to the GAN latent space and an auxiliary classiLD = − E [log(D(x̃))]− E [log(1−D(G(z)))], fier. This approach requires the dataset to include x∼Pr z∼p(z) class labels, i.e. to work in a supervised setting. (1) The generator receives the noise vector z and also while the generator minimizes the selected label c so that the generated sample LG = − E [log(D(G(z)))]. (2) is a function of both. Furthermore, the discrimiz∼p(z) nator has, in addition to the usual objective, the task of correctly classifying the real and generated The main issue in the original GAN formula- samples (through an auxiliary classifier). The gention was the instability of the training process that erator is optimized not only to deceive the discrimmade very hard to improve architectures and to inator, but also to generate fake samples that minscale up to bigger images. In [13] a deep convo- imize the auxiliary classifier error, i.e. to produce lutional architecture was proposed for both gen- well class-defined samples. erator and discriminator which presents some degree of stability for adversarial training. This work was the first one producing convincing image sam- 2.2 WGAN ples for datasets with low variability (Bedrooms and Faces) and relatively low resolution (64x64). In order to address the problem of instability in However, standard GAN formulation fails to gen- GAN training, Arjovsky et al. in a series of works erate globally consistent samples when trained on [1, 2] proposed a reformulation of the function to be optimized. They argue that the original loss funcdatasets with high variability like ImageNet. tion presents discontinuities and vanishing gradients with respect to generator parameters. Instead, 2.1 AC-GAN they proposed a distance for distributions known In order to tackle datasets with high variability, as Earth-Mover distance or Wasserstein-1, which Odena et al. [12] proposed to improve the quality captures the cost of transporting mass in order to 3 transform one distribution into the other. From this distance they derive the WGAN loss function for the minimax objective min max E [D(x)] − G D∈D x∼Pr E [D(G(z))] (3) z∼p(z) where D (called critic in WGAN formulation) is not anymore a binary classifier and is restricted to be in the set D of 1-Lipschitz functions. Again, z is a noise vector sampled from a simple distribution (uniform or Gaussian distribution). The Lipschitzness of D was imposed by weight clipping in this first version of WGAN. The importance of Arjovsky’s contribution lies on a gain in the robustness of the adversarial training process and a reduction in the frequency of the mode collapse phenomenon. Furthermore, the proposed loss function correlates well with the observed sample quality as opposed to the original GAN loss which gives little information about training progress. Figure 4: Samples generated with the ResNet-B architecture trained with Splitting GAN over CIFAR10 without class labels (unsupervised). is the most recurrent benchmark for image generation in GAN literature. In the unsupervised setting (without using any class information) they reach 2.2.1 WGAN-GP the state-of-the-art, while in the supervised case An improved version of WGAN was recently pro- (following the strategy of AC-GAN and without posed by Gulrajani et al. [7]. They found that the tuning any hyperparameter nor architecture) they 2 weight clipping can cause convergence problems in reach the second place behind SGAN [8] . All this some settings and propose to enforce the Lipschitz advantageous features made –to our knowledge– constraint on the critic D by penalizing its gradi- WGAN the current standard formulation for adversarial training. ent’s norm. The penalty term is computed over a set of random points x̂ uniformly sampled from straight lines Our Method: Splitting between real and fake sample points. Naming as Px̂ 3 the latter distribution, the new loss can be written GAN as The main idea is to generate new artificial classes L = E [D(x̃)] − E [D(x)] based on the representation learned by the last hidx∼Pr x̃∼Pg den layer of the critic after enough training itera  +λ E (k∇x̂ D(x̂)k2 − 1)2 (4) tions. This is done by applying k-means to each x̂∼Px̂ class set in this representation space. We divide each set in two clusters only when the class has where the penalty coefficient is set to λ = 10. This improved WGAN formulation exhibits high more samples than a certain threshold. After that, robustness against changing model architecture. training is resumed replacing the old labels with The authors tested six different network designs the new ones for the entire dataset. With this procedure we need to make two mifor both G and D, which typically break standard GAN training but show stable WGAN training for nor modifications to the model architecture before all cases. Furthermore, WGAN formulation helps resuming learning: achieving better quality samples. Quantitative re2 By simply duplicating the number of feature maps of sults are reported by the authors in terms of the Gulrajani’s networks we found WGAN outperforms SGAN Inception score [14] over CIFAR-10 dataset, which score. See Sec. 4 4 1. The auxiliary classifier needs to predict a different number of classes, so we extend the last layer of this classifier adding a copy of the weights of the parent class for each child class. Method Improved GAN (-L+HA) [14] EGAN-Ent-VI [4] DFM [15] Splitting GAN ResNet-B (ours) WGAN-GP ResNet-B WGAN-GP ResNet-A [7] Splitting GAN ResNet-A (ours) Inception Score 6.86 ± 0.06 7.07 ± 0.10 7.72 ± 0.13 7.80 ± 0.08 7.81 ± 0.10 7.86 ± 0.07 7.90 ± 0.09 2. In the conditional WGAN-GP implementation [6], the class labels are injected in each batch normalization layer of the generative network by setting a specific gain and bias parameters (γ and b) for each class. We follow this strategy in our proposal and, for the class splitting, we set the new pair (γ, b) for each child class as Table 1: Unsupervised Inception scores on CIFARγchild = γf ather +∆γ and bchild = bf ather +∆b, 10 with initialization ∆γ = 0 and ∆b = 0 when the new classes are created. This formulation Method Inception Score implies that child classes start both with the father class params and they eventually beImproved GAN [14] 8.09 ± 0.07 come different. AC-GAN [12] 8.25 ± 0.07 WGAN-GP ResNet-A [7] 8.42 ± 0.10 SGAN [8] 8.59 ± 0.12 4 Results WGAN-GP ResNet-B 8.67 ± 0.14 Splitting GAN ResNet-A (ours) 8.73 ± 0.08 To demonstrate the effectiveness of our proposal, Splitting GAN ResNet-B (ours) 8.87 ± 0.09 we conducted several experiments with CIFAR-10 [9], a dataset containing 50000 32x32 images corresponding to 10 different classes, and the unlabeled Table 2: Supervised Inception scores on CIFAR-10 set of STL-10, containing 100000 larger and more diverse images [3]. We based our model on the improved WGAN algorithm proposed by Gulrajani et al. [7, 6]. In all cases, during training, we sample ture with the original WGAN-GP algorithm. The 50000 images from the current model to select the results are detailed in Table 1 and Table 2. best one so far based on the Inception score. FiAlso for comparison, samples obtained with the nally, we sample another 50000 with the best model WGAN-GP supervised model (Figure 2) and obin order to calculate the reported score, following tained with the proposed method (Figure 1) are [12]. shown. Figure 3 has real samples of CIFAR-10 corresponding to each of the 20 clusters found. 4.1 CIFAR-10 Figures 4 and 5 show generated images and clusters found in the unsupervised test. With CIFAR-10, an unsupervised test was performed starting from all the examples considered as a single class and dividing them into two clusMethod Inception Score ters every approximately 10 epochs. This was done with the Gulrajani’s ResNet architecture without Original Dataset [15] 26.08 ± 0.26 changes (named ResNet-A) and a modified version DFM [15] 8.51 ± 0.13 (ResNet-B) doubling the number of maps in each WGAN-GP ResNet-A 9.05 ± 0.12 convolutional layer. A supervised test was also conSplitting GAN ResNet-A (ours) 9.50 ± 0.13 ducted with these two architectures, starting from the original 10 classes of CIFAR and dividing them into two at approximately 20 training epochs. For comparison we also trained the ResNet-B architec- Table 3: Unsupervised Inception scores for STL-10 5 Figure 6: Images generated by the model trained on STL-10. 4.2 STL-10 We treat STL-10 in the same way as [15]. That is, we downsample each dimension by 2, resulting in 48x48 RGB images. We tested our algorithm with the ResNet-A architecture, with the minimum changes necessary for the model to generate 48x48 images. Table 3 shows the resulting Inception score. Figures 6 and 7 show the generated images and the clusters found by the method. 5 Discussion Several things can be observed from the results presented in the previous section. First, regarding the obtained clusterization of the real samples (Figure 3 for the supervised case and Figure 5 for the unsupervised one), we can visually find rules that define the vast majority of samples, for at least several clusters. As an example, in the supervised case (Figure 3) we can see in the left side of the fourth row cats looking forward and in the left side of the eighth row horse side views. Compare with cats and horses in several positions corresponding to the Figure 5: The 25 clusters found in the unsupervised clusters in the right side. In the unsupervised case case (real CIFAR-10 samples). Each line has two (Figure 5) we can see a tendency to generate clusdifferent clusters. ters for cars, planes or ships, but in general they 6 are much more mixed. Regarding the generated samples in the supervised case (Figure 1 for our method and Figure 2 for WGAN-GP), we can see that the class splits allows the model to generate better samples. Not only for the more uniform clusters such as the horse side views or the cats looking forward, but for the whole original class. Compare for example the fourth row (cats) or the eighth row (horses) in Figure 1 with those rows in Figure 2, corresponding to the same model trained with WGAN-GP. Note that these samples do not differ too much from those shown in [7]. Even in classes where the clustering step does not seem to have found an obvious separation rule, such as cars (second row), a better sample quality can be observed than in the original WGAN-GP. In the unsupervised case with CIFAR-10 (Figure 4), although the Inception score is similar than the one obtained by the state-of-the-art so far, the samples generated seem to be of a higher quality. Nevertheless, they do not reach the quality of the generated images in a supervised setting. It is always advisable to start the division into clusters from the predefined classes, if this information is available. In the case of STL-10 (Figure 6), there is a noticeable difference in the Inception score. The reason of this may be that STL-10 is a much more diverse dataset, so a division into a large number of classes can be beneficial. It should be noted that in this case the state-of-the-art is much further from the actual dataset score than in the case of CIFAR-10. The success of our Splitting GAN method suggests that reinjecting high level information from critic to the generative model improves sampling quality. This breaks the strictly adversarial training and allows some degree of information sharing between both networks. We believe that this simFigure 7: The 15 clusters found by the model (real ple (but successful) strategy could inspire a new STL-10 samples). and better adversarial training formulation where a small amount of high level information directly flows from critic’s last layers to generator input. 6 Conclusions Work and Future In this work we showed that our Splitting GAN method allows generating better images. This can 7 be seen in the results on CIFAR-10 and STL-10 for which clearly better images were obtained. This is supported by an Inception score well above the previous state-of-the-art for both datasets. A future direction of research to improve the current Splitting GAN version is oriented to understand how a given model architecture or dataset determines the optimal number of clusters (or classes). Also, clusterization could be enhanced during adversarial training with the addition of an extra loss term like in [16]. We are also currently working in a generalization of the Splitting GAN ideas following two paths: First, making the high level information from the critic’s representation flows continually to the generator, avoiding special purpose steps at predefined times in the training process. Second, avoiding the hard threshold clustering step and replacing it with some sort of continuous representation capturing the same amount of information. A. Courville, and Y. Bengio. Generative adversarial nets. In Advances in neural information processing systems, pages 2672–2680, 2014. [6] I. Gulrajani. Improved Training of Wasserstein GANs. https://github.com/igul222/ improved_wgan_training, 2017. [Online; accessed 18-Sep-2017]. [7] I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, and A. Courville. Improved training of wasserstein gans. arXiv preprint arXiv:1704.00028, 2017. [8] X. Huang, Y. Li, O. Poursaeed, J. Hopcroft, and S. Belongie. Stacked generative adversarial networks. arXiv preprint arXiv:1612.04357, 2016. [9] A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images. 2009. [10] X. Mao, Q. Li, H. Xie, R. Y. Lau, Z. Wang, and S. P. Smolley. Least squares generative adversarial networks. arXiv preprint The authors acknowledge grant support from ANArXiv:1611.04076, 2016. PCyT PICT-2016-4374. [11] M. Mirza and S. Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. References 7 Acknowledgements [1] M. Arjovsky and L. Bottou. Towards princi- [12] A. Odena, C. Olah, and J. Shlens. Conditional image synthesis with auxiliary classifier pled methods for training generative adversarGANs. In D. Precup and Y. W. Teh, editors, ial networks. In International Conference on Proceedings of the 34th International ConferLearning Representations 2017, 2017. ence on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages [2] M. Arjovsky, S. Chintala, and L. Bot2642–2651, International Convention Centre, tou. Wasserstein gan. arXiv preprint Sydney, Australia, 06–11 Aug 2017. PMLR. arXiv:1701.07875, 2017. [3] A. Coates, A. Ng, and H. Lee. An analysis of [13] A. Radford, L. Metz, and S. Chintala. Unsupervised representation learning with deep single-layer networks in unsupervised feature convolutional generative adversarial networks. learning. In Proceedings of the fourteenth inarXiv preprint arXiv:1511.06434, 2015. ternational conference on artificial intelligence and statistics, pages 215–223, 2011. [14] T. Salimans, I. Goodfellow, W. Zaremba, V. Cheung, A. Radford, X. Chen, and [4] Z. Dai, A. Almahairi, P. Bachman, E. Hovy, X. Chen. Improved techniques for training and A. Courville. Calibrating energygans. In D. D. Lee, M. Sugiyama, U. V. based generative adversarial networks. arXiv Luxburg, I. Guyon, and R. Garnett, editors, preprint arXiv:1702.01691, 2017. Advances in Neural Information Processing [5] I. Goodfellow, J. Pouget-Abadie, M. Mirza, Systems 29, pages 2234–2242. Curran AssoB. Xu, D. Warde-Farley, S. Ozair, ciates, Inc., 2016. 8 [15] D. Warde-Farley and Y. Bengio. Improving generative adversarial networks with denoising feature matching. In International Conference on Learning Representations 2017, 2017. [16] Y. Wen, K. Zhang, Z. Li, and Y. Qiao. A discriminative feature learning approach for deep face recognition. In European Conference on Computer Vision, pages 499–515. Springer, 2016. [17] J. Zhao, M. Mathieu, and Y. LeCun. Energybased generative adversarial network. arXiv preprint arXiv:1609.03126, 2016. 9 

1cs.CV

i F WLBP: A Scale Invariant Descriptor for Texture Classification arXiv:1801.03228v1 [cs.CV] 10 Jan 2018 Swalpa Kumar Roy, Nilavra Bhattacharya, Bhabatosh Chanda, Bidyut B. Chaudhuri, Life Fellow, IEEE, and Dipak Kumar Ghosh Abstract—In this paper we propose a novel texture recognition feature called Fractal Weighted Local Binary Pattern (FWLBP). It has been observed that fractal dimension (FD) measure is relatively invariant to scale-changes, and presents a good correlation with human perception of surface roughness. We have utilized this property to construct a scale-invariant descriptor. We have sampled the input image using an augmented form of the local binary pattern (LBP), and then used an indexing operation to assign FD weights to the collected samples. The final histogram of the descriptor has its features calculated using LBP, and its weights computed from the FD image. The proposed descriptor is scale, rotation and reflection invariant, and is also partially tolerant to noise and illumination changes. In addition, it is also shown that the local fractal dimension is relatively insensitive to the bi-Lipschitz transformations, whereas its extension is able to correctly discriminate between fundamental texture primitives. Experimental results show the proposed descriptor has better classification rates compared to the state-of-the-art descriptors on standard texture databases. Index Terms—Fractal Dimension (FD), Fractal weighted local binary pattern (FWLBP), Scale invariance, Scale space, Texture classification C I. I NTRODUCTION ONSTRUCTING effective texture feature is a major challenges in computer vision [1]. It has received a lot of attention in the past decades due to its value in understanding how human beings recognize textures. A crucial issue of texture analysis is constructing an effective texture representation. There are primarily five methodologies of texture representation: statistical, geometrical, structural, model-based, and signal processing based [2], [3]. In recent literature, two categories of texture classification approaches dominate current research, namely texture descriptor based methods [4], [5] and deep learning based methods. In the second category, a convolutional neural network (C NN) is trained to classify the texture images [6], [7]. The deep learning based methods offer good classification performance, however it has following limitations: it requires a large amount of data and hence is computationally expensive to trains. S. K. Roy and B. B. Chaudhuri are with the Optical Character Recognition Lab, Computer Vision and Pattern Recognition Unit at Indian Statistical Institute, Kolkata, India (email: swalpa@ieee.org; bbc@isical.ac.in). Nilavra Bhattacharya is with the School of Information, The University of Texas, Austin, USA (email: nilavra@ieee.org). B. Chanda is with the Image Processing Lab, Electronics and Communication Sciences Unit at Indian Statistical Institute, Kolkata, India (email: chanda@isical.ac.in). D. K. Ghosh is with the Department of Electronics & Communication Engineering, National Institute of Technology Rourkela, Orissa, India (email: dipak@ieee.org) Basically the complex models take weeks to train using several machines equipped with expensive GPUs. Also, at present, no strong theoretical foundation on topology/training method/flavor/hyper-parameters for deep learning exist in the literature. On the other hand, the descriptor based methods have the advantage of ease of use, data independence, and robustness to real-life challenges such as illumination and scale differences. In this category of methods the design of effective texture descriptor is regarded as extremely important for texture recognition. A picture may be captured under geometric and photometric varying conditions. An ideal model for describing and recognizing textures should be able to capture essential perceptual properties of the texture. It should be robust enough against changes, like view-point alteration, luminance variation, image rotation, reflection, scale change, and geometry of the underlying surface. Attention has been focussed on the design of local texture descriptors capable of achieving local invariance [8]–[11]. The search for invariant descriptors started in the 1990s [12]. Kashyap and Khotanzad first proposed a circular autoregressive dense model [13] to study rotation invariant texture classification. Many other models including multiresolution [14], hidden Markov model [15], and Gaussian Markov model [16] were explored to study rotation invariance. More recently Varma and Zisserman [10] proposed to learn a rotation invariant texton dictionary from a training set, and then classify the texture image based on its texton distribution. They also proposed another texton based algorithm [11] by using the local image patch to represent the feature directly. In addition some works have been done on scale and affine invariant feature extraction in texture classification. Among them Varma and Garg [17] extracted a local fractal vector for each pixel, and computed a statistical histogram. Liu and Fieguth [18] applied random projection for densely sampled image patches, and extracted the histogram signature. Yao and Sun [19] normalized statistical edge feature distribution to resist the variation in scale. Lazebinik et al. [20] and Zhang et al. [8] detected Harris and Laplacian key points for extracting texture signatures. Recently, global scale invariant feature extraction methods drew more attention because local scale normalization is slow because of pixel by pixel operations are needed there. Xu et al. [21] and Quan et al. [5] have classified the image pixels into multiple point sets by gray intensities or local feature descriptors. Instead of extracting scale invariant features, pyramid histograms with shifted matching scheme was proposed by some researchers [9], [22]. Roy et al. [23] introduced a complete dual-cross pattern (C DCP) to address ii the scale and rotational effects in unconstrained texture classification. Generally, a texture can be characterized by geometric multi-scale self-similar macro-structure (like fractal dimension and L BP). Geometric structure of object in the image vary with scale changes while illumination change does not change the object structure. Several studies have been done using fractal based texture classification [17], [24]–[26] but the drawback of these techniques are that the value of fractal dimension is continuous, so a quantization stage is required to compute the histogram. However, an accurate quantization depends upon a excessive training with large number of samples. In this paper we approach the quantization problem by proposing a new descriptor called Fractal Weighted Local Binary Pattern (F WLBP) based on a commonly used method of combining fractal dimension proposed by Chaudhuri and Sarkar [27]. This method is based on differential boxcounting (D BC) algorithm [28]. Where instead of directly measuring a texture surface, the measures at different scales are obtained by means of counting the number of boxes of different size, which can cover the whole surface. In this paper, our main contributions are summarized below: • We propose a simple, effective, yet robust fractaldimension based texture descriptor called Fractal Weighted Local Binary Pattern (Sec. III-A) for texture classification. • The F WLBP descriptor achieves sufficient invariance to address the challenges of scale, translation, and rotation (or reflection); and is also partially tolerant to noise and illumination for texture classification. • The proposed algorithm takes a reasonable amount of time in the feature extraction stage (Sec. III-B). • We experimented and observed that our proposed descriptor outperforms the traditional L BP based and other stateof-the-art methods on different texture databases, such as KTH-TIPS and CUReT (Sec. IV). The remaining of this the paper is organized as follows. Sec. II deals with the mathematical background of fractal dimension, and our approach to convert the input image to its FD form. The proposed F WLBP feature extraction scheme is presented in Sec. III. In Sec. IV the performance of texture classification is compared with the state-of-the-art methods. Finally, concluding remarks are drawn in Sec. V. II. F RACTAL D IMENSION A fractal is a geometrical set whose Hausdorff - Besicovitch dimension is strictly greater than its topological dimension [30]. Mandelbrot introduced the term fractal to describe nonEuclidean, self-similar structures. The Fractal Dimension (FD) of a structure gives an idea of its texture complexity, which can be used to measure, analyse and classify shape and texture. Fractal surfaces show the property of self-similarity. In the Euclidian n-space, a set S is said to be self-similar, if S is the union of Nr (s) distinct (non-overlapping) copies of itself, each of which is scaled down by a ratio r > 0. Let D(s) (i.e. an exponential function of r), where Nr (s) = k( 1r ) D(s) is a density function, and k is a constant. Then 1 log Nr (s) = log k + D(s) · log (1) r Then, the local Fractal Dimension (FD) of s becomes log(Nr (s)) D(s) = lim . (2) r→0 log(1/r) It is difficult to compute the local Fractal dimension D using Eqn. (2) directly. In our implementation (Fig. 1), we transform the texture images into FD images using the differential boxcounting (D BC) algorithm [28]. To achieve scale-invariance, we construct a Gaussian scale space [31] using Eqn. (3): Z Γx,y,r = hGr (x, y)|I(x, y)i = Gr (x, y)I(x, y)dxdy. x,y∈R (3) In Eqn (3) a convolution formalism may be used when convolution across the entire image is required. However, here an inner product formalism (h.|.i) is more convenient where Gr is a Gaussian smoothing kernel with variance r: −kx2 k 1 Gr (x) = √ exp 2σ2 r 2π As the 2D Gaussian function is separable [32], [33], its convolution with the input image can be efficiently computed by applying two passes of the 1D Gaussian function in the horizontal and vertical directions: Gr (x, y, r) = Gr (x)Gr (y). The scaling operation to obtain Ir = Gr ∗ I (where ∗ represents convolution operator) can be computed efficiently and stably because of the properly localization of the Gaussian both in space and frequency [34] even if the input image (I) is the result of physical measurement, so called directly sampled [35]. After constructing the Gaussian scale space, we use the D BC algorithm on each layer of the scale space to generate a set of intermediate images. We combine the intermediate images into the final FD image using a linear-regression technique adapted from [29]. The input texture image Ax,y (X × Y ) is used to generate a Gaussian scale space Γx,y,r (X ×Y ×L) having L layers, with scaling factor r varying from rmin , to rmax , and total number of layers L = rmax −rmin +1. Each layer is considered as a 3− D space, with (x, y) denoting the 2 − D position, and the third coordinate, g, denoting the gray level. The D BC algorithm is implemented by applying a varying-size non-linear kernel bi,j (m × n), which operates on the image surface and finds the difference between the maximum (gmax ) and the minimum (gmin ) gray values of the region enclosed within the kernel (Fig. 2). Non-negative integers α and β are used to center the kernel b(i, j) on the pixel gx,y in the image layer. The kernel is calculated as   β α X X gmax − gmin b(i, j) = floor +1 (4) r i=−α j=−β where r is the scaling factor of the layer l, and     m−1 n−1 α = ceil , and β = ceil . (5) 2 2 The kernel is applied as  2 β α X X L Ω(x, y, r) = b(i, j)Γ(x + α, y + β, r) (6) r i=−α j=−β iii   rmax rmax  rmax Box n Ax , y ( X  Y )   rmin rmin  rmin Image intensity surface Box 2 (a) Fx , y ( X  Y ) Box 1  x , y ,r ( X  Y  L) (b) (e)  x , y ,r ( X  Y  L) (d) (c) Fig. 1. Illustrates the Differential Box Counting (DBC) technique [28] for computing Fractal Dimension. (a) Input texture image. (b) Gaussian Scale Space with six layers, generated from the input image with box sizes from rmin × rmin , . . . , rmax × rmax . (c) DBC is used to calculate the FD for each image layer of the scale space. (d) Intermediate images. (e) Final FD image obtained using technique adapted from [29]. The slope can be determined by computing the sums of squares: P 2 X ( r) (9) Ψ1 = r2 − L P 2 P 2 X ( r) ( Θ) Ψ2 = rΘ − . (10) L Finally, the FD image Fx,y (X × Y ) is calculated as: g max g min F (x, y) = λxy =    =  Fig. 2. Image intensity surface enclosed within a box, showing the difference between maximum and minimum gray values. where Ωx,y,r (X × Y × L) is a matrix of intermediate images, such that the first layer denotes the original image after it has been filtered by the kernel with scale r = rmin , and the highest layer denoting the image filtered by the kernel with scale r = rmax . Thus, we define   g11r g12r · · · g1Y r  g21r g22r . . . g2Y r    Ω(x, y, r) =  . (7)  .. .. ..  ..  . . . gX1r gX2r ... gXY r The slope λx,y of the least square linear regression line between Ω(x, y, r) and r will denote the pixel value F (x, y) of the FD image Fx,y (X × Y ). For this purpose, a columnvector Θ is defined, such that the first gray values of all the layers in Ω(x, y, r) comprise vector θ1 , all the second gray values comprise vector θ2 etc. as shown in Eqn. (8).    Θ  θ1 θ2 .. . θX×Y       =   g111 g121 .. . g112 g122 .. . ··· ... .. . g11L g12L .. . gXY 1 gXY 2 ... gXY L      (8) X X Y X Ψ2 x=1 y=1 Ψ1 g11r g21r .. . g12r g22r .. . ··· ... .. . g1Y r g2Y r .. . gX1r gX2r ... gXY r    .  (11) Algorithm 1: Algorithm to compute FD image of input texture image. 1 2 3 4 5 6 Data: Input texture image, Ax,y (X × Y ) Result: Matrix Fx,y (X × Y ) containing FD values of each pixel of A initialize (rmin = 2, rmax = 7, L = rmax − rmin + 1) ; while (r ≥ rmin and r ≤ rmax ) do Gr = Gaussian Kernel([r, r], r/2) ; Γ(x,y,r) = hGr (x, y)|I(x, y)i; /* Update D BC kernel b */ h i β α P P min b(i, j) = floor gmax −g +1 ; r i=−α j=−β 7 8 /* Compute Ω(x, y, d) Matrix of intermediate FD values */  2 β α P P Ω(x, y, d) = b(i, j)Γ(x + α, y + β, l) L ; r i=−α j=−β 9 end while 10 Convert the 3-D matrix Ωx,y,r (X × Y × L) to a column vector Θ, as per Eqn. (8). ; 11 /* Compute sum of square */ 12 Ψ1 = 13 Ψ2 = P P P ( r)2 ; L P P ( r)2 ( Θ)2 − L r2 − rΘ 14 Compute FD image F (x, y) = ; X P Y P x=1 y=1 Ψ2 Ψ1 ; A. Effects of Fractal Transform An interesting property of the Fractal Transform is its invariance under bi-Lipschitz transform. A transform t : R2 → R2 iv is called a bi-Lipschitz transform, if there exist two constants 0 < c1 ≤ c2 ≤ ∞ such that for any two points x1 , x2 ∈ R2 , c1 kx1 − x2 k < kt(x1 ) − t(x2 )k < c2 kx1 − x2 k (12) where kp1 − p2 k is the Euclidean metric between points p1 and p2 . So, any smooth transform (like translation, rotation, projective transformation, texture warping of a regular surface, change in viewpoint, and non-rigid deformation) is a biLipschitz transform [36], [37]. Up-sampling or down-sampling by a factor of n are also special cases of the bi-Lipschitz transform, where c1 = c2 = n. Then, we can drive the following theorem: Theorem 1. Let us consider an image as f (x), x ∈ R2 . Let the fractal dimension and length of point x be D and L, respectively. For a sampling function t(x) = nx, where n > 1 means up-sampling, and n < 1 means down-sampling, the fractal dimension and length at point t(x) become Dt = D, Lt = L + D · log n (13) where L = log k. Proof. For a point x, we have log Nr (x) = D · log( 1r ) + L (from Eqn. (1)). By the definition of t(x) we know that t is invertible and kt(x1 ) − t(x2 )k = n kx1 − x2 k. The biLipschitz constant ensures that Nr (t(x)) = N nr (x). Hence we can write, n log N nr (x) = D · log + L r 1 ⇒ log Nr (t(x)) = D · log + D log n + L r (14) 1 ⇒ log Nr (t(x)) = Dt · log + Lt r ⇒ Dt = D, Lt = L + D · log n Hence, from the above theorem, the fractal dimension of an image is invariant to scale changes, or in other words, invariant to local bi-Lipschitz transforms, but not the fractal length of the image. Theoretically, the fractal dimension of an image remains unchanged when it is sampled-up with a factor of n and vice-versa. This is our main motivation to take advantage of the fractal dimension in our proposed descriptor. III. P ROPOSED T EXTURE D ESCRIPTOR Scale variations can create a dramatic impact on the appearance of a texture. There is no locally invariant texture descriptor that can handle scale changes of such image intensity surface. Our goal in this paper is to build texture descriptors which are invariant to scale changes in the intensity surface of natural texture. Even local fractal dimension is a powerful measurement of surface “roughness” of a natural image, though the fractal dimension alone is not sufficient to characterize natural texture. However, the FD yields continuous values and so it needs to be quantized. The quantization process can be done by calculating the feature distribution from all the training samples to know about the distribution of feature space. Finally, some threshold value is calculated to divided the feature space into fixed number of bins and the quantization of test FD image is done according to the training threshold values. Fig. 3 illustrates quantization process of the feature space using four bins. There are the following important limitations due to the quantization process, which need to be addressed carefully. • To determine the threshold values for each bin, it require a pre-training stage. • Texture images are captured under different geometric and photometric varying conditions. So, the quantization process fully depends on the training samples. • The choice of number of bins is a challenging task although there are some techniques [38] to select it. It is always a trade-off between optimal number of bins for the best of accuracy and feature size. Since the distribution has a finite number of entries, choosing a few bins may fail to provide sufficient discriminative power while large number of bins may lead to sparse and unreliable results. Hence, feature matching becomes computationally inefficient. The method is explained in the following sections. Fig. 3. Quantization procedure of the feature space [39] A. Fractal Weighted Local Binary Pattern (F WLBP) A good texture descriptor should have four powerful characteristics: reliability, independence, discrimination, and small size. Hence, we propose a simple, efficient, yet robust texturedescriptor called the Fractal Weighted Local Binary Pattern (F WLBP) and offer a solution to the above problem based on the concept of local image patterns exhibiting self-similarity. The discrimination power of the proposed descriptor is enhanced by using the fractal dimensions as weights of the histogram where the sampling method of indexing is an augmentation of the native L BP [38]. We generated multi-scale indexing images at different scales by varying the sampling radius (R ∈ [Rmin , Rmax ]). However, to keep reasonable complexity, we have kept number of sampling points (N ) constant across all the scales. Since fractal dimension is a logarithmic function (Eqn. (2)), it helps to reduce effects of illumination, because the log function expands the values of darker pixels and compresses the brighter pixels in the image. As a result, the FD values are spread more uniformly [40]. Thus, the proposed descriptor is insensitive to scale, translation, rotation or reflection, illumination; and is also noise-tolerant. After generating the scale-invariant FD image (Fig. 4b) of the original texture using the D BC technique (Sec. II), we proceed to compute the proposed descriptor – Fractal v the FD values of these n locations in the FD image are F (x1 , y1 ), F (x2 , y2 ), . . . , F (xn , yn ), then the weight w of the L BP value in the feature histogram becomes (16) w = F (x1 , y1 ) + F (x2 , y2 ) + . . . + F (xn , yn ). (b) (a) I Mathematically, it can be expressed as: I F WLBPR,N (p) = I(L BPR,N x,y , Fx,y , p), Rmin O Rmax where (e) I (c) (d) (f) Fig. 4. Proposed F WLBP algorithm. (a) input texture image (b) computation of FD image using Gaussian Scale Space and D BC algorithm (c) taking L BPR,N samples from input image using R ∈ [Rmin , Rmax ] and constant N (d) L BP images for varying R (e) indexing output FD image with the different L BP images to calculate fractal weights (f) final feature vector Weighted Local Binary Pattern (F WLBP). We generate multiresolution L BP images (Fig. 4(d)) from the input texture image by varying the sampling radius R ∈ [Rmin , Rmax ] (Fig. 4(c)). The L BP for a given pixel at location (x, y) in the input image Ax,y (X × Y ) is computed by comparing its gray value ax,y with a set of N local circularly and equally spaced neighbours R,N Px,y placed at a radius R ( R ∈ [Rmin , Rmax ]) around the pixel. L BP is computed for only those pixels whose all N local neighbours lie within the image, and not for pixels along the image boundary. The final L BP value of a given pixel is calculated as LBPR,N x,y = N X n=1 R,N 2n−1 × sign(Px,y − ax,y ) p ∈ [0, 2N −1 ] (17) (15) where sign is a unit step function to denote whether a given input is positive or not, and is defined as ( 1, z ≥ 0 sign(z) = . 0, z < 0 The range of L BPR,N x,y depends on the number of neighbouring sampling points (N ) around the pixel (x, y) at radius R to form the pattern, and its value lies in between 0 to 2N −1 . In other words, the range of L BP is [0, 2N −1 ]. Fig. 4(d) shows the computed local binary patterns (i.e. L BPR,N x,y |R = [Rmin , Rmax ]) for a candidate texture image (Fig. 4(a)). Finally, we use an indexing operation I to combine each of the L BP images with the previously generated FD image, and then compute the feature histogram. The indexing operation works as follows: in order to compute the histogram frequency R,N N −1 ]) of (weight) of a given pixel value (pR,N x,y | px,y ∈ [0, 2 R,N the L BP image L BPx,y , we refer to the FD image Fx,y (X × Y ). For all the pixel locations in the L BP image which have the same value as pR,N x,y , we find the sum of the FD values in all the same pixel locations in the FD image. This sum is the weight of the value pR,N x,y in the histogram. For instance, if the L BP value pR,N x,y occurs at n locations in the L BP image, namely (x1 , y1 ), (x2 , y2 ), . . . , (xn , yn ), and I(L BPR,N x,y , Fx,y , p), = (P ∴ L BPR,N x,y == p otherwise Fx,y , 0, where F WLBP(p) gives the weight of the value p in the histogram, L BPR,N x,y is the L BP image matrix constructed with sampling radius R and N sampling points, Fx,y is the FD image as computed in Sec.( II), and I is the indexing operation. The final F WLBP feature vector (Fig. 4(f)) is constructed by concatenating all F WLBP(p) values. Since the weights of the L BP histogram are decided by the value of the fractal dimensions, we named the descriptor as Fractal Weighted Local Binary Pattern. Given a m texture image, we have created the feature matrix Mm×n where each row corresponds to the F WLBP descriptors of a texture. The principal component analysis (P CA) feature is computed as following. Initially, we compute the co-variance matrix Σ of feature matrix as, Pn (M(., i) − M(., i))(M(., i) − M(., i))T Σ(i, j) = i=1 n−1 (18) where Mm×n is the feature matrix, n is the number of feature, M(., i) represents the ith column of M, and M(., i) is the mean of respective column. We calculate eigen-vector e of Σ if it satisfies Σe = λe. Where λ is an eigen-value of Σ. The eigen-values of Σ are sorted in descending order as λ1 ≥ λ2 ≥ . . . ≥ λn and corresponding eigenvectors e1 , e2 , . . . , en are used as columns of linear transform matrix U. Finally, M is multiplied by UT . Thus, the P CA features are provided by D = UT M. The scatter plot of proposed descriptor without and with P CA are shown in Fig. 5. 20 200 0.01 15 0.008 150 0.006 10 0.004 100 5 0.002 0 50 140 120 100 80 60 40 20 50 100 150 200 250 300 0 1500 1000 500 (a) 1000 2000 3000 4000 0.035 0.03 0.025 0.02 0.015 0.01 0.005 (b) 0.04 0.06 0.08 0.1 (c) 0.15 300 2000 200 0.1 1000 100 0.05 0 0 0 −1000 −100 −2000 −200 −3000 −300 −4000 200 100 200 0 −100 0 −200 (d) −200 6000 4000 2000 −0.05 −0.1 −0.15 −0.2 0.2 5000 0 −2000 0 −5000 (e) 0.1 0.2 0.1 0 0 −0.1 −0.1 −0.2 −0.2 (f) Fig. 5. Top row (a)-(c) represent scatter plot of F WLBP feature, whereas bottom row (d)-(f) represent F WLBP feature after P CA transform, are extracted from Brodatz, KTH-TIPS and Outex TC10 databases . vi Algorithm 2: Algorithm to extract FWLBPx,y R,N descriptor, where N is the number of circular samples taken at radius R. 1 2 3 4 5 6 7 8 9 10 11 Data: Input texture image, Ax,y (X × Y ), FD image F (x, y) Result: Normalized Fractal Weighted Local Binary Pattern (F WLBP) Descriptor initialize (Rmin = 1, Rmax = 3, N = 8, i = R + 1, j = R + 1, k = [0, K]); while (R ≤ Rmax ) do /* Compute L BP image with sampling radius R, from a sample of R,N N local circularly spaced neighbours Px,y around a central pixel A(x, y) */ PN R,N L BP(x,y)R,N = n=1 2(n−1) × sign(Px,y − A(x, y); /* Compute weights of pixel values for histogram by indexing operation */ R,N F WLBP(p) = I(L (PBPx,y , Fx,y , p) ; Bx,y , f or all Bx,y where Ax,y == k ; I(A, B, k) = 0, otherwise /* Generate histrograms, concatenate, and normalize */ HR,N = H IST(F WLBPR,N ); H FWLBP = concatNormalizeF WLBP(HR,N ); end while B. Complexity of FWLBP Assuming the image is of size M × M and the kernel is of size r × r, the algorithmic complexity of the proposed descriptors is computed as follows. To create the FD image ×M of an input texture image with the number of blocks Mr×r using Algorithm 1, the following operations are performed: 2M 2 comparisons, 4(M/r)2 addition, 5(M/r)2 subtractions, 4(M/r)2 multiplication and 2(M/r)2 divisions. Therefore overall complexity of Algorithm 1 is O(M 2 ). In Algorithm 2, assuming that neighboring pixels are considered with respect to each center pixel for the evaluation of the L BP for FD image of size M × M . The time complexity to evaluate Step 4 of Algorithm 2 is O(p × M 2 ) [41] because p circular shifting are required for each pixel, and there are M 2 pixels in total. Step 6 of Algorithm 2 takes O(M 2 ) for Indexing operation and Step 9 takes O(M 2 ) for building histogram. The overall time complexity to compute the proposed F WLBP descriptor using the combined Algorithms 1 and 2 is O(M 2 ). C. Comparing Distribution of FWLBP After computing the F WLBP descriptors, as elaborated in the previous section, we need to represent their distributions in the training (model) and test (sample) images. This is done since the dissimilarity of sample and model histograms is a test of goodness-of-fit, which can be measured with a nonparametric statistical test. Examples of metrics for evaluating the fit between two histograms are histogram intersection, log-likelihood ratio, and chi-square (χ2 ) statistic [38]. In this paper, the classification is performed via a non-parametric classifier called nearest subspace classifier (N SC) [42]. Here we have chosen N SC over popular parametric classifier like S VM due to the following reasons: the N SC finds an estimation of the underlying subspace within each class and assigns data points to the class that corresponds to its nearest subspace, which does not required any parameter tuning, while S VM has limitations in speed during both training and testing phase, and there is always a trade-off between the selection of kernel function, tuning of its hyper-parameters and classification performance. In order to avoid the over emphasizing patterns with large frequency, a pre-processing step is applied to the proposed feature before using N SC, similar to that in [43]: p (19) X¯m = Xm , m = 1, 2, ..., M where M is the number of bins, and Xm is the original frequency of the F WLBP at mth bin. The nearest subspace classifier (N SC) first calculates the distance from the test sample y to the ith class and measures the projection residual ri from y to the orthogonal principle subspace Bi ∈ RM ×m of the training sets is spanned by the principal P Xi , which T th eigenvectors of = X X class, given as i i for the i i follows, ri = k(I − PBi )yk2 = k(I − Bi BT i )yk2 (20) where P = I ∈ RM ×M is a identity matrix where M rows are selected uniformly at random. The test sample y is then assigned to the one of the C classes with the smallest residual among all classes, i.e. i∗ = arg min ri (21) i=1,...,C IV. E XPERIMENTAL E VALUATION A. Texture Databases We have evaluated the performance of our proposed descriptor on five standard texture databases: Outex TC-00010 (Outex TC10) [4], Outex TC-00012 (Outex TC12) [4], Brodatz album [44], KTH-TIPS [45], and CUReT [46] texture database. We have compared the performance of the proposed F WLBP descriptor with L BPR,N [38], L BPu2 R,N , D LBP R,N [41], su2 riu2 riu2 [47], multiscale C LBP S L BPsri R,N R,N /MR,N /C(1, 8 + 3, 16 + 5, 24) [48], B IF [9] and other state-of-the-art descriptors. We have tested the classification accuracy using a nonparametric classifiers – N SC, and have reported the classification accuracy using k-fold cross-validation test. In k-fold cross-validation test, the feature set is randomly partitioned into k equal sized subsets (k = 10). Of the k subsets, a single subset is retained as the validation data for testing the classifier, and the remaining (k − 1) subsets are used as training data. The average of the classification accuracies over k rounds give us a final cross-validation accuracy. We have normalized each input image to have an average intensity of 128 and a standard deviation of 20 [38]. In VZ-MR8 and VZPatch methods, the texture samples are normalized to have an average intensity of 0 and a standard deviation of 1 [10], [11], [17]. This is done to remove global intensity and contrast. The details of two experimental setups are given as follows: TABLE I S UMMARY OF T EXTURE DATABASE USED IN E XPERIMENT #2 Texture Database KTH-TIPS Brodatz CUReT Image Rotation X X X Illumination Variation X X Scale Variation X X Texture Classes 10 32 61 Sample Size (pixels) 200 x 200 64 x 64 200 x 200 Samples per Class 81 64 92 Total Samples 810 2048 5612 E XPERIMENT #1: Brodatz [44] album is chosen to allow a direct comparison with the state-of-the-art results [41]. There vii are 32 homogeneous texture classes 1 . Each image is partitioned into 25 non-overlapping sub-images of size 128 × 128, and each sub-image is down-sampled to 64 × 64 pixels. Fig. 6. Ten texture images randomly taken from each class of KTH-TIPS database. For the CUReT database [46], we have used the same subset of images as in [8], [10], [18], which contain 61 texture classes with 92 images per class. It is designed to contain large intra-class variation and is widely used to assess the classification performance. The images are captured under different illumination and viewing directions with constant scale. All 92 images of 61 texture classes are cropped into 200 × 200 region and converted to gray scale [10]. The KTH-TIPS database [45] is extended by imaging new samples of ten CUReT textures as shown in Fig. 6. It contains texture images with 3 different poses, 4 illuminations, and 9 different scales of size 200×200 and hence each class contains 81 samples. The KTH-TIPS, Brodatz, and CUReT, databases are summarized in Table I. TABLE II S UMMARY OF T EXTURE DATABASE USED IN E XPERIMENT #2 Texture Database Outex TC10 Outex TC12 Image Rotation X X Illumination Variation X Scale Variation Texture Classes 24 24 Sample Size (pixels) 128 x 128 128 x 128 Samples per Class 180 200 Total Samples 4320 4800 E XPERIMENT #2: 24 different homogeneous texture classes are selected from the Outex texture databases [4], each having the size of 128 × 128 pixels. Outex TC 00010 (Outex TC10) contains texture with illuminant “inca”, and Outex TC 00012 (Outex TC12) contains textures with illuminants “inca”, “horizon”, and “tl84”. Both of the Outex test suites are collected under 9 different rotation angles ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ (0 , 5 , 10 , 15 , 30 , 45 , 60 , 75 , and 90 ) in each texture class. The test suites Outex TC10, and Outex TC12 are summarized in Table II. Running experiments on the whole Fig. 7. 24 texture images randomly taken from each class of Outex TC10 and Outex TC12 database database is challenging due to a large number of texture classes, small number of samples per class, and lack of intraclass variation. 1 The 32 Brodatz textures are Bark, Beachsand, Beans, Burlap, D10, D11, D4, D51, D52, D5, D6, D95, Fieldstone, Grass, Ice, Image09, Image15, Image17, Image19, Paper, Peb54, Pigskin, Pressedcl, Raffia2, Raffia, Reptile, Ricepaper, Seafan, Straw2, Tree, Water, Woodgrain; TABLE III C OMPARISON WITH OTHER VARIANTS OF L BP Methods Classifier L BPV [49] B RINT [50] L BPriu2 1,8 [38] D LBP3,24 [41] sri L BP1,8 su2 [47] N NC N NC N NC S VM N NC N NC N NC N NC S VM N NC N NC N NC N SC L BP(1,8+2,16+3,24) [38] C LBP SMC [48] S SLBP [43] P RI C O L BPg [51] L BPHF S [52] L BPHF S M [53] C OA L BP [54] Proposed F WLBP Classification Accuracy (%) KTH-TIPS [45] Brodatz [44] CUReT [46] 95.50 93.80 94.00 97.75 99.22 97.06 82.67 82.16 80.63 86.99 99.16 84.93 89.73 69.50 85.00 95.17 91.60 95.84 97.19 94.80 97.40 97.80 98.55 98.40 96.90 98.40 97.00 94.60 95.90 97.00 94.60 95.90 97.00 94.20 98.00 99.75 99.62 99.10 B. Results of Experiment #1 To test the scale invariance performance of the proposed descriptor, we have used the KTH-TIPS database, which contains relatively small variations in scales (i.e, 9 scales, continually from 0.5 to 2) [45]. In literature, for texture classification task usually L BP variants were used, in this section, we compared F WLBP descriptor with some powerful variant of L BPs, which includes su2 C LBP [48], L BPV [49], LTP [55], Dlbp [41], L BPsri [47], R,N P RI C OLBP [51], B RINT [50], and C OA L BP [54]. The classification performance is evaluated on three well known benchmark texture databases (KTH-TIPS, Brodatz, and CUReT) and results are shown in Table III. The observations noted su2 from Table III are as follows: The scale-invariant L BPsri R,N riu2 descriptor performs better than L BPR,N . However, the perriu2 formance is worse than multi-resolution L BPR,N descriptor, riu2 riu2 and C LBP SR,N /MR,N /C descriptor; and much worse than the proposed descriptor. This is because extracting consistent and accurate scale for each pixel is difficult. However, the su2 L BPsri R(i,j),8 provides good performance in controlled environment [47], but it fails over more complex databases. The D LBP, when combined with Gabor features, attains a higher classification rate than the conventional L BP with N NC. However, its performance is quite less than the proposed F WLBP, as it does not consider changes in scale. F WLBP achieves remarkably better classification performance than C LBP, L BPV, L BPHF S and L BPHF S M on KTH-TIPS, Brodatz and CUReT, texture datasets and yields comparable performance with D LBP. Note that L BP, L BPHFS , L BPHF S M , C LBP, and L BPV generally encode multi-scale information whereas the proposed F WLBP descriptor encodes the spatial fractal dimension to characterize the local scale information into the 1-D L BP histogram. Table IV shows performance of texture classification performance other than L BP variants. Lazebinik et al. [20] proposed to detect interest regions using Harris-affine corner and Laplacian-affine blobs and then extracted S PIN and R IFT as texture signatures after normalizing these regions. Finally, texture classification is performed using nearest neighbor classifier (N NC). Caputo et al. [56] used S VM kernel instead of N NC and reveal that the S VM classifier could achieve reasonably better performance. Zhang et al. [8] proposed viii 0.8 FWLBP LBP 0.6 1 S 3 S 4 S 5 S 6 S 7 S 8 S 9 S (d) S 2 0.4 0.2 0 (a) (b) 1 2 3 4 5 Scales 6 7 8 (c) Fig. 8. (a)-(c) represent the individual histograms (HR,N of F WLBP, where R = 1, 2, 3 and N = 8) for a texture sample taken from the KTH-TIPS database [45], having 9 different scales. The abscissa and ordinate represent number of bins and feature probability distribution, respectively. As the histograms of individual F WLBP for the 9 different scales are overlapping, it implies that the F WLBP descriptor is scale invariant. (d) represents the Chi-square (χ2 ) distance of proposed F WLBP and L BP descriptor extracted from original texture image (S 1 ) and 8 texture image with different scale (S 2 − S 9 ). Methods Classifier VZ-MR8 [10] VZ-Patch [11] Lazebnik et al. [20] Zhang et al. [8] Liu et al. [18] M FS [24] Capato et al. [56] P FS [5] B IF [9] Proposed F WLBP N NC N NC N NC S VM S VM N NC S VM S VM Shift N NC N SC Classification Accuracy (%) KTH-TIPS [45] Brodatz [44] CUReT [46] 94.50 94.62 97.43 92.40 87.10 98.03 91.30 88.20 72.50 96.10 95.90 95.30 94.20 98.50 81.62 94.80 95.00 98.50 97.35 98.50 98.47 98.60 99.75 99.62 99.10 It is interesting to note that even for the database having large number of variations and a few number of samples, this property can make the proposed descriptor suitable for other applications where the number of features plays an important role. 100 99 98 Accuracy (%) TABLE IV T EXTURE C LASSIFICATION R ESULTS ON KTH-TIPS, B RODATZ , AND CUR E T 97 96 95 Outex_TC10 KTH−TIPS Brodatz 94 93 object and texture classification by analyzing different texture features and kernels. Recently, global scale invariant feature extraction methods drew attention because local scale normalization is usually slow due to pixel by pixel operations. Xu et al. [21] and Quan et al. [5] have classified the image pixels into multiple point sets by gray intensities or local feature descriptors. Multi-scale B IF [9] at scales σ, 2σ, 4σ, and 8σ gives better performance than the proposed F WLBP with N NC classifier. This is mainly because B IF uses pyramid histogram with time inefficient shift matching scheme. Also, the feature dimension of B IF descriptor [9], C LBP [48], S SLBP [43] and P RI C OLBP [51] are larger (64 = 1296, 2200, 480 × 5 = 1000 and 590 × 2 = 1180) than the dimension of proposed F WLBP descriptor (256×3 = 768). The performance of B IF is reduced when scale shifting scheme is not considered [9]. Apart from bag-of-words model, the variant of L BP based descriptors, sri su2 e.g., C LBP 2 , L BPV, LTP, D LBP, L BPR,N , P RI C OLBP, B RINT, C OA L BP, and S S L BP also perform reasonably well on texture classification task. The observations noted from Table IV are as follows: F WLBP also exceeds the classification performance of several bag-of-words methods, [8], [11], [18], [20], [54], [56]. Fig. 9 shows the graph of success rates against the number of selected feature on Brodatz, KTHTIPS and Outex TC10 texture databases. The classification rate of proposed descriptor rapidly increases due to P CA transform, which carries the most relevant information with large variance among the first component of the feature matrix. 2 We used the best C LBP settings i.e., C LBP S/M/C in this paper 92 50 100 150 200 Feature Dimension 250 300 Fig. 9. Success rate against the number of descriptors on Brodatz, KTH-TIPS and Outex TC10 texture databases Although the images are captured under scale, rotation, and illumination variations, the proposed F WLBP gives sound performance as given by results in Table IV, and achieves comparable, and sometimes even better performance, than the state-of-the-art methods. It validates that the encoding strategy preserves the relative information about local scale change along with orientation angle are effective. An illustration of the scale invariance property of the proposed descriptor is given in Fig. 8, where S1 -S9 show nine KTH-TIPS texture images of “corduroy” class at different scales from 0.5 to 2, and (a)-(c) depict the histograms (HR,N |R ∈ [1, 2, 3]) of individual F WLBPR,N descriptor with R = 1, 2, 3, and N = 8. It is evident from Fig. 8(a)-(c) that the histograms of individual F WLBP for the 9 different scales are approximately identical. In addition Fig. 8(d) shows the Chi-square (χ) distance [11] between extracted F WLBP and L BP descriptor from original texture image (S 1 ) and 8 texture images with different scale (S 2 − S 9 ), its shown that proposed descriptor has less χ-distance compared to L BP. Therefore, from both the observations it implies that the F WLBP descriptor is not sensitive to small scale variations. The proposed descriptor achieves scale invariance by utilizing the properties of Fractal Dimension and Gaussian Scale Space, as explained in ix Theorem 1. Fig. 10(a)-(b) represent images of two texture classes taken form Brodatz database and Fig. 10(c) shows P CA transformed F WLBP descriptor extracted from three images per texture class and respective feature probability distribution functions (P DF). Since both texture classes contain very small intra-class variation even though their P DF provides a strong indication about the discriminative power of the proposed descriptor. 3 class 1 class 2 2.8 2.6 PDF 2.4 (a) 2.2 2 1.8 1.6 1.4 1.2 1 2 4 6 8 10 Descriptors (b) (c) Fig. 10. Discriminative power of the proposed F WLBP descriptors are visualized using P CA. We have taken 6 images from two classes of Brodatz textures and extracted the F WLBP. Notice that the classes are substantially separated by the curves. TABLE V T HE C LASSIFICATION ACCURACIES OF PROPOSED F WLBP AND F WLBPu2 DESCRIPTORS USING DIFFERENT SCALING FACTORS OF D BC ALGORITHM ON BRODATZ TEXTURE DATABASE . Methods Proposed FWLBP Std. Error Proposed FWLBPu2 Std. Error Classification Rates With Different box sizes ranges rmin to rmax 2×2 2×2 2×2 2×2 2×2 to to to to to 2×2 3×3 4×4 5×5 6×6 7×7 99.24 99.25 99.27 99.27 99.31 99.62 ±0.9216 0.0052 ±0.9688 0.0052 ±0.7547 0.0052 ±0.6701 0.0104 ±0.6484 0.0104 ±0.3257 0.0002 98.05 98.16 98.30 98.45 98.53 98.57 ±0.6717 0.0156 ±0.8426 0.0208 ±0.8590 0.0208 ±0.8058 0.0156 ±0.7514 0.0104 ±0.7243 0.0260 C. Analysis of parameter rmax The proposed fractal based descriptor has only one hyperparameter rmax , which is the maximum value of the scaling factor of the D BC algorithm. The rmax represents how much a specific intrinsic structure is self-similar [36] to its surrounding pixels. The different values of rmax influence the deviation of classification accuracies. Table V shows recognition accuracy for Brodatz textures using proposed F WLBP and F WLBPu23 , where rmax values are ranging from 2 to 7. Its shows clearly that both the descriptors have similar trends towards rmax and recognition rate increases till it reaches rmax to 7. Table V also indicates that for different scaling factors the proposed F WLBP descriptor significantly outperforms F WLBPu2 on well-known Brodatz texture database (includes images of scale and rotation or view-point variations). To compute fractal dimension of an image of size less than let say, 200 × 200 fast, Faraji and Qi [40] recommend rmax to be used in between 7 and 10 to 3 Where ‘u2’ indicates uniform pattern encoding; compromise between good texture recognition accuracy and computational overheads. In our experimental setup, we use 7 as rmax value for all the texture databases. D. Results of Experiment #2 The results of Experiment #2, carried out on Outex TC10 and Outex TC12 are tabulated in Table VI. It shows the average classification accuracy obtained by k-fold cross-validation test (k = 10) for proposed descriptor, and the comparative summary of the results for variants of L BP e.g., L BP(R,N ) [38], VZ-MR8 [10], VZ-Patch [11] and recent state-of-theart bag-of-word methods. The observations made from the results of Experiment #2 are stated as follows. L BP and L BPV have similar feature dimensions, but the later one incorporates additional contrast measures to the pattern histogram, it produces a significant performance improvement compared to the conventional L BP. L BPriu2 R,N /VAR R,N provides better performance compared to L BPVriu2 R,N , because L BP and local variance of a texture image are complementary; and hence the joint distribution of L BP and local variance gives better results than any one alone. C LBP S/M/C is created by fusing C LBP S and C LBP M/C. It provides better performance compared to other variants of C LBP as it has complementary features of sign and magnitude, in addition to the center pixel representing the gray value of the local patch. C DCP [23] achieves improved performance over C LBP since in C DCP patterns are extracted from component and holistic levels. The state-of-the-art bag-of-words statistical algorithms, VZTABLE VI AVERAGE CLASSIFICATION ACCURACY (%) ON O UTEX TC10 AND O UTEX TC12 USING STATE - OF - THE - ART SCHEMES Method Classifier Outex TC10 VZ-MR8 [10] VZ-Patch [11] LTP [55] VAR [57] L BP [38] L BPriu2 R,N L BP/VAR riu2 L BPVR,N [49] C LBP S/M/C [48] T L BPN R,N [58] D LBPR=3,N =24 [41] B RINT C S C M [50] P TP [59] C DCP [23] Proposed F WLBP N SC N SC N NC N NC S VM N NC N NC N NC N NC N NC S VM N NC N NC N SC N SC 93.59 92.00 76.06 90.00 97.60 84.89 96.56 91.56 98.93 99.24 98.10 99.35 99.56 99.76 99.97 Outex TC12 horizon tl84 92.82 92.55 92.06 91.41 63.42 62.56 64.35 62.93 85.30 91.30 63.75 65.30 78.08 79.31 77.01 76.62 92.29 90.30 96.18 94.28 87.40 91.60 97.69 98.56 98.08 97.94 99.82 99.62 99.98 99.93 Average 92.99 91.82 67.34 72.42 91.40 71.31 84.65 81.73 93.84 96.56 92.36 98.12 98.52 99.72 99.96 MR8 and VZ-Patch, take dense response from multiple filters. However, their performance is poor compared to the proposed F WLBP. Also, feature extraction and matching complexity of these two techniques is quite high [11], when compared to the proposed F WLBP due to the MR8 needs to find 8 maximum responses after 38 filters convolving with the image and compares every 8-dimension vector in an image with all the textons to build histograms using clustering technique. D LBP + N GF utilizes the top ranking 80% of L BP pattern to improve the recognition, as compared with the results obtained using original L BPu2 R,N . But like VAR R,N , it neglects local spatial structure which is important for texture discrimination. Also, D LBP needs pre-training stage, and the dimensionality of the D BLP varies with the training samples. x 0.7 FWLBP 0.6 5 10 15 30 45 60 75 LBP 0.5 90 (d) 0 0.4 0.3 0.2 0.1 0 (a) (b) 1 2 3 4 5 Angles 6 7 8 (c) Fig. 11. (a)-(c) represent individual F WLBP histograms (F W LBPR,N |R ∈ [1, 2, 3]) of a texture sample taken from the Outex database, with 9 different orientations. The abscissa and ordinate represent the number of bins and feature probability distribution, respectively. The histograms show that the feature distribution for the different texture orientations are approximately overlapping, which signify the rotation invariance property of the proposed F WLBP descriptor. ◦ ◦ ◦ (d) represents the Chi-square (χ) distance between original texture image (0 ) and 8 texture image with different angles (5 − 90 ) Though the trend is clear from the performance Table VI, we have further analysed the performance using one way statistical analysis of variance (A NOVA) test [60]. A NOVA is a collection of statistical tests used to analyze the differences among group means and their associated procedures. In other word, one-way A NOVA is used to test the equality of two or more means at one time using the variances. The null hypothesis H0 for the test is, there is no significant difference among group means. We have taken the significance level α = 0.05 for this A NOVA test. We can reject H0 if the p-value for an experiment is less than the selected significant level, which implies that the at least one group mean is significantly different from the others. To understand why the performance of proposed F WLBP descriptor was significantly different from well-known descriptors such as VZ-Patch, VZ-MR8, B RINT, D LBP, C LBP, and L BPriu2 , we conduct one way A NOVA test with significance level α = 0.05. The test results are shown in Table VII where the p-value (1.6460e−07 ) is much lesser than the pre-select significance level α = 0.05. This indicates that the performance of proposed descriptor significantly better than other descriptors and hence we can reject the hypothesis H0 . In addition, the box plot corresponding to aforementioned A NOVA test is shown in Fig. 12, which also clearly indicates that the mean performance of F WLBP descriptor is significantly better than the well-known descriptors such as VZ-MR8 [10], VZ-Patch [11], B RINT [50], D LBP [41], C LBP [48], and L BPriu2 [38]. 100 Recognition rates (%) For comparison, we have taken the best results of D LBP with T R = 3 and N = 24 in Table. VI. The L BPN R,N [58] based methods and B RINT [50] give better performance compared to other state-of-the-art L BP methods. However, compared to the proposed F WLBP, their accuracies are lower. This is T because L BPN R,N extracts features by locally rotation invariant riu2 L BPR,N approach, which produces only 10 bins. Such small features cannot represent each class well. On the other hand, B RINT extracts a large number of features from multiple resolution (R = 1, 2, 3, 4) by utilizing rotation invariant L BPri R,N approach, but it loses some global image information. From Table VI, it is evident that our proposed descriptor provides better classification performance compared to other state-of-the-art methods. The better performance is due to generation of a six-layer Gaussian Scale Space to provide basic scale-invariance (Fig. 1), transforming the scale space to fractal dimension (FD) images, which also provides additional levels of scale invariance and some degree of illumination invariance (Sec. III), and finally using the FD as weights for the values in the feature vector histograms. When all these are put together we achieve high degree of invariance to scale, rotation, and reflection, and are able to capture the micro structure in the different illumination controlled environments (for images like “inca”, “horizon” and “tl84”). To visualize the rotation invariant characteristic of our proposed descriptor, an example of the F WLBP feature distribution for a texture sample is taken from the Outex TC10 database (with 9 orientations – ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ 0 , 5 , 10 , 15 , 30 , 45 , 60 , 75 , and 90 ) and is shown in Fig. 11. Fig 11(a)-(c) represent the histograms (HR,N |R ∈ [1, 2, 3]) of individual (F WLBPR,N |R ∈ [1, 2, 3]). The histograms show that the F WLBP feature distribution for the different texture orientations are approximately overlapping, which signify the rotation invariance property of the F WLBP descriptor. In addition, Fig. 11(c) represents the histograms of individual F WLBP for the 9 different rotational angles, which are approximately identical. In addition, Fig. 8(d) shows the Chi-square (χ) distance [11] between extracted F WLBP and ◦ L BP features from original texture image (0 ) and 8 texture ◦ ◦ images with different rotational view-points (5 − 90 ), it is shown that the proposed descriptor has less χ-distance compared to L BP. 95 90 85 80 75 70 65 FWLBP VZ−Patch VZ−MR8 BRINT DLBP CLBP LBP_riu2 State−of−the−art descriptors Fig. 12. The box plot (Descriptor vs. Accuracy) corresponding to one way statistical A NOVA test for proposed F WLBP and state-of-the-art descriptors on Outex, KTH-TIPS, Brodatz, CUReT databases. xi TABLE VII O NE WAY STATISTICAL A NOVA TEST RESULTS FOR O UTEX , KTH-TIPS, B RODATZ , AND CUR E T DATABASES , WHERE THE SIGNIFICANCE LEVEL α IS 0.05. Source Groups Error Total SS 1923.34 0621.82 2515.67 df 06 28 34 MS 325.557 022.208 F 14.36 Prob (p) > F 1.6460e−07 where the convolution filters are pre-defined as wavelet and no learning process is needed. Inspired by ScatNet, a simple deep learning network, P CA N ET is proposed by chan et al. [64] which is based on cascading of multistage P CA, binary hashing and histogram pooling. E. Robustness to Noise S NR = 5 24.06 99.74 51.22 57.37 50.88 88.31 99.93 Table VIII demonstrates the noise robustness of different methods on Outex TC10 database based on classification rates of various noise levels (measured using S NR i.e Signal to Noise Ratio). The proposed descriptor achieves state-of-theart results in term of accuracy mean and standard deviations 100 ± 0%, 100 ± 0%, 99.97 ± 0.0732%, 99.95 ± 0.0976%, and 99.93 ± 0.1118% on S NR = 100 dB, 30 dB, 15 dB, 10 dB, and 5 dB, respectively. It can be observed from Table VIII that the proposed F WLBP descriptor is very modestly improved as compared with other L BP based descriptors. The proposed descriptor has inherited this property from the calculation of fractal dimension in Gaussian scale space representation in contrast to the L BP and its variants. F. Comparison of Proposed F WLBP with state-of-art deep learning based methods A key characteristic of deep convolutional neural networks (C NN)(see Fig. 13) is a hierarchical representation which is universal and directly generated from the data, used to perform image classification task. Deep C NNs have universal shown their power of pattern recognition. However, the robustness for recognition is limited due to the lack of geometric invariance of global C NN activations. An effective texture descriptor FVCNN has been introduced by Cimpoi et al. [7], where at first C NN features are extracted at multiple scales. Then an orderless Fisher Vector pooling operation is performed. Despite significant progress of deep C NN models, not much analytical insights into its internal operation and behaviour is available. Mallat et al. [62], [63] have introduced scale and rotation invariant wavelet convolution scattering network (ScatNet) TABLE IX T HE TEXTURE CLASSIFICATION PERFORMANCE AND FEATURE DIMENSIONALITY OF THE PROPOSED L MP AND STATE - OF - ART DEEP LEARNING BASED METHODS Classification Accuracy (%) Methods FV-VGGDM (S VM) [7], [65] FV-AlexNet (S VM) [65], [67] S CAT N ET (P CA) [62], [63] S CAT N ET (N NC) [62], [63] P CA N ET (N NC) [64] P CA N ETriu2 (N NC) [64] R AND N ET (N NC) [64] R AND N ETriu2 (N NC) [64] Proposed F WLBP (N NC) Proposed F WLBP (N RS) 80.00 67.30 99.69 98.59 39.87 35.36 47.43 43.54 99.79 99.86 82.30 72.30 99.06 98.10 45.53 40.88 52.45 45.70 99.41 99.73 88.20 77.90 68.92 63.66 59.43 52.15 60.67 56.90 98.37 99.87 99.80 99.10 96.15 88.64 57.70 49.80 56.57 48.20 97.12 99.90 Feature C LBP SMC [48] LTPriu2 R=3,N =24 [55] N RLBPriu2 R,N [61] B RINT [50] Proposed F WLBP Classification Accuracy (%) S NR = 30 S NR = 15 S NR = 10 86.93 67.24 49.79 99.79 99.76 99.76 98.12 94.58 86.07 98.31 93.44 84.32 81.16 77.52 70.16 96.48 95.47 92.97 100 99.97 99.95 Dimention N NC N NC N NC N NC N NC N NC N NC S NR = 100 95.03 99.30 99.45 84.49 97.76 100 UIUC Classifier L BPriu2 R,N [38] T L BPN R,N,k [58] KTH-TIPS Methods A simple variation of P CA N ET, named R AND N ET, in which the cascaded filters are randomly selected but not learned, was also proposed by Chan et al. [64]. One of the major trends in deep C NN research community is to use more and more complex networks to improve the classification performances. However, it needs the powerful and large memory computer and GPUs to train very deep and computationally expensive networks. The comparative results of texture classification performance with feature dimensionality of the proposed F WLBP and state-of-art deep learning based methods such as F V-C NN [7], [65], S CAT N ET [62], [63], P CA N ET, P CA N ETriu2 , and R AND N ET [64], are tabulated in Table IX. It is observed from Table IX that the proposed F WLBP provides comparable or better classification performance compared to state-of-art deep learning based methods. In addition, the feature dimension of proposed F WLBP descriptor is less than the deep learning based models4 . Outex TC12 TABLE VIII C LASSIFICATION ACCURACY (%) OF P ROPOSED M ETHOD AND D IFFERENT STATE - OF - THE - ART M ETHODS ON O UTEX TC10 WITH D IFFERENT N OISE L EVELS IN TERM OF D B . Fig. 13. A Deep C NN architecture where the input is processed in a feed forward manner through the stage of convolutions and sub-sampling and finally classified with a linear or non-linear classifier. Outex TC10 To evaluate the performance of the proposed method in noisy environment, the experiments are carried out on Outex TC10 texture database (discussed in subsec. IV-A) by adding white Gaussian noise, resulting in different Signal to Noise Ratio (S NR). The training and testing set-up are the same as in noise-free situation. 65536 32768 596 596 2048 80 2048 80 354 354 V. C ONCLUSIONS In this paper, we have proposed a simple, efficient yet robust descriptor, called the Fractal Weighted Local Binary Pattern (F WLBP) for texture classification. At first, a Gaussian Scale Space representation of the texture image is generated, and 4 The results of deep learning based methods are taken from Liu et al. [66] xii then Differential Box Counting (D BC) algorithm is used to transform the images into fractal dimension. An augmented form of Local Binary Pattern (L BP), with fixed sample size and varying radius is used, and the weights for the samples are calculated using an indexing function. Finally, the normalized feature vector is formed by concatenating the histogram of F WLBP for all elements of L BP. Experimental results show that our proposed F WLBP descriptor provides promising performance under scale, rotation, reflection, and illumination variation. Comparative study indicates that our descriptor also performs better, as compared to other state-of-the-art methods. In addition, the dimensionality of the extracted feature vector is significantly less the state-of-the-art method, making it computationally efficient. As it offers reasonably high speed to extract the features, this descriptor may be applied to real time scenarios. Our current work is focused on texture classification. In future, this work can be utilized in texture classification in multi-texture scenario, also in other domains like face and object recognition. ACKNOWLEDGEMENT The authors sincerely thank the Machine Vision Group (MVG), University of Oulu, Finland, Visual Geometry Group (VGG), University of Oxford, UK, Department of Computing, The Hong Kong Polytechnic University, China, and Department of Computer Science, University College London, UK for sharing the source codes of L BP, VZ-MR8, C LBP, and B IF descriptors. The authors would also like to thank Manik Varma, Timor Kadir and Andrew Zisserman for providing preprocessed CUReT datasets. R EFERENCES [1] X. Xie and M. Mirmehdi, “A Galaxy of Texture Features,” in Handbook Of Texture Analysis. World Scientific, 2008, pp. 375–406. [2] R. M. Haralick, K. Shanmugam et al., “Textural features for image classification,” IEEE Transactions on systems, man, and cybernetics, no. 6, pp. 610–621, 1973. [3] T. Randen and J. H. Husoy, “Filtering for texture classification: A comparative study,” IEEE Transactions on pattern analysis and machine intelligence, vol. 21, no. 4, pp. 291–310, 1999. [4] T. Ojala, T. Maenpaa, M. Pietikainen, J. Viertola, J. Kyllonen, and S. Huovinen, “Outex-new framework for empirical evaluation of texture analysis algorithms,” in Pattern Recognition, 2002. Proceedings. 16th International Conference on, vol. 1. IEEE, 2002, pp. 701–706. [5] Y. Quan, Y. Xu, and Y. Sun, “A distinct and compact texture descriptor,” Image and Vision Computing, vol. 32, no. 4, pp. 250–259, 2014. [6] V. Andrearczyk and P. F. Whelan, “Using filter banks in convolutional neural networks for texture classification,” Pattern Recognition Letters, vol. 84, pp. 63–69, 2016. [7] M. Cimpoi, S. Maji, I. Kokkinos, and A. Vedaldi, “Deep filter banks for texture recognition, description, and segmentation,” International Journal of Computer Vision, vol. 118, no. 1, pp. 65–94, 2016. [8] J. Zhang, M. Marszałek, S. Lazebnik, and C. Schmid, “Local features and kernels for classification of texture and object categories: A comprehensive study,” International journal of computer vision, vol. 73, no. 2, pp. 213–238, 2007. [9] M. Crosier and L. D. Griffin, “Using basic image features for texture classification,” International Journal of Computer Vision, vol. 88, no. 3, pp. 447–460, 2010. [10] M. Varma and A. Zisserman, “A statistical approach to texture classification from single images,” International Journal of Computer Vision, vol. 62, no. 1-2, pp. 61–81, 2005. [11] ——, “A statistical approach to material classification using image patch exemplars,” IEEE transactions on pattern analysis and machine intelligence, vol. 31, no. 11, pp. 2032–2047, 2009. [12] I. WEISS, “Geometric invariants and object recognition,” International journal of computer vision, vol. 10, no. 3, pp. 207–231, 1993. [13] R. L. Kashyap and A. Khotanzad, “A model-based method for rotation invariant texture classification,” IEEE Transactions on Pattern Analysis and Machine Intelligence, no. 4, pp. 472–481, 1986. [14] J. Mao and A. K. Jain, “Texture classification and segmentation using multiresolution simultaneous autoregressive models,” Pattern recognition, vol. 25, no. 2, pp. 173–188, 1992. [15] J.-L. Chen and A. Kundu, “Rotation and gray scale transform invariant texture identification using wavelet decomposition and hidden Markov model,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 16, no. 2, pp. 208–214, 1994. [16] H. Deng and D. A. Clausi, “Gaussian MRF rotation-invariant features for image classification,” IEEE transactions on pattern analysis and machine intelligence, vol. 26, no. 7, pp. 951–955, 2004. [17] M. Varma and R. Garg, “Locally invariant fractal features for statistical texture classification,” in Computer Vision, 2007. ICCV 2007. IEEE 11th International Conference on. IEEE, 2007, pp. 1–8. [18] L. Liu and P. Fieguth, “Texture classification from random features,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 34, no. 3, pp. 574–586, 2012. [19] C.-H. Yao and S.-Y. Chen, “Retrieval of translated, rotated and scaled color textures,” Pattern Recognition, vol. 36, no. 4, pp. 913–929, 2003. [20] S. Lazebnik, C. Schmid, and J. Ponce, “A sparse texture representation using local affine regions,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 27, no. 8, pp. 1265–1278, 2005. [21] Y. Xu, H. Ji, and C. Fermüller, “Viewpoint invariant texture description using fractal analysis,” International Journal of Computer Vision, vol. 83, no. 1, pp. 85–100, 2009. [22] J. Zhang, J. Liang, and H. Zhao, “Local energy pattern for texture classification using self-adaptive quantization thresholds,” IEEE transactions on image processing, vol. 22, no. 1, pp. 31–42, 2013. [23] S. K. Roy, B. Chanda, B. B. Chaudhuri, D. K. Ghosh, and S. R. Dubey, “A complete dual-cross pattern for unconstrained texture classification.” [24] Y. Xu, H. Ji, and C. Fermuller, “A projective invariant for textures,” in Computer Vision and Pattern Recognition, 2006 IEEE Computer Society Conference on, vol. 2. IEEE, 2006, pp. 1932–1939. [25] J. B. Florindo, A. R. Backes, M. de Castro, and O. M. Bruno, “A comparative study on multiscale fractal dimension descriptors,” Pattern Recognition Letters, vol. 33, no. 6, pp. 798–806, 2012. [26] Y. Quan, Y. Xu, Y. Sun, and Y. Luo, “Lacunarity analysis on image patterns for texture classification,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2014, pp. 160–167. [27] B. B. Chaudhuri and N. Sarkar, “Texture segmentation using fractal dimension,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 17, no. 1, pp. 72–77, 1995. [28] N. Sarkar and B. B. Chaudhuri, “An efficient differential box-counting approach to compute fractal dimension of image,” IEEE Transactions on Systems, Man and Cybernetics, vol. 24, no. 1, pp. 115–120, 1994. [29] O. S. Al-Kadi, D. Watson et al., “Texture analysis of aggressive and nonaggressive lung tumor ce ct images,” IEEE transactions on biomedical engineering, vol. 55, no. 7, pp. 1822–1830, 2008. [30] B. B. Mandelbrot and R. Pignoni, The fractal geometry of nature. WH freeman New York, 1983, vol. 173. [31] T. Lindeberg, “Scale-space theory: A basic tool for analyzing structures at different scales,” Journal of applied statistics, vol. 21, no. 1-2, pp. 225–270, 1994. [32] D. G. Lowe, “Distinctive image features from scale-invariant keypoints,” International journal of computer vision, vol. 60, no. 2, pp. 91–110, 2004. [33] L. D. Griffin and M. Lillholm, “Symmetry sensitivities of derivativeof-gaussian filters,” IEEE transactions on pattern analysis and machine intelligence, vol. 32, no. 6, pp. 1072–1083, 2010. [34] J. D. Victor and B. W. Knight, “Simultaneously band and space limited functions in two dimensions, and receptive fields of visual neurons,” in Perspectives and Problems in Nolinear Science. Springer, 1994, pp. 375–419. [35] J. J. Koenderink, “The structure of images,” Biological cybernetics, vol. 50, no. 5, pp. 363–370, 1984. [36] K. Falconer, Fractal geometry: mathematical foundations and applications. John Wiley & Sons, 2004. [37] R. Van Den Boomgaard and A. Smeulders, “The morphological structure of images: The differential equations of morphological scale-space,” IEEE transactions on pattern analysis and machine intelligence, vol. 16, no. 11, pp. 1101–1113, 1994. xiii [38] T. Ojala, M. Pietikainen, and T. Maenpaa, “Multiresolution gray-scale and rotation invariant texture classification with local binary patterns,” IEEE Transactions on pattern analysis and machine intelligence, vol. 24, no. 7, pp. 971–987, 2002. [39] M. Pietikäinen, A. Hadid, G. Zhao, and T. Ahonen, Computer Vision Using Local Binary Patterns. Springer Science & Business Media, 2011, vol. 40. [40] M. R. Faraji and X. Qi, “Face recognition under varying illumination with logarithmic fractal analysis,” IEEE Signal Processing Letters, vol. 21, no. 12, pp. 1457–1461, 2014. [41] S. Liao, M. W. Law, and A. C. Chung, “Dominant local binary patterns for texture classification,” IEEE transactions on image processing, vol. 18, no. 5, pp. 1107–1118, 2009. [42] J. Wright, A. Y. Yang, A. Ganesh, S. S. Sastry, and Y. Ma, “Robust face recognition via sparse representation,” IEEE transactions on pattern analysis and machine intelligence, vol. 31, no. 2, pp. 210–227, 2009. [43] Z. Guo, X. Wang, J. Zhou, and J. You, “Robust texture image representation by scale selective local binary patterns,” IEEE Transactions on Image Processing, vol. 25, no. 2, pp. 687–699, 2016. [44] P. Brodatz, Textures: a photographic album for artists and designers. New York, NY, USA: Dover Pubns, 1966. [45] E. Hayman, B. Caputo, M. Fritz, and J.-O. Eklundh, “On the significance of real-world conditions for material classification,” in European conference on computer vision. Springer, 2004, pp. 253–266. [46] K. J. Dana, B. Van Ginneken, S. K. Nayar, and J. J. Koenderink, “Reflectance and texture of real-world surfaces,” ACM Transactions On Graphics (TOG), vol. 18, no. 1, pp. 1–34, 1999. [47] Z. Li, G. Liu, Y. Yang, and J. You, “Scale-and rotation-invariant local binary pattern using scale-adaptive texton and subuniform-based circular shift,” IEEE Transactions on Image Processing, vol. 21, no. 4, pp. 2130– 2140, 2012. [48] Z. Guo, L. Zhang, and D. Zhang, “A completed modeling of local binary pattern operator for texture classification,” IEEE Transactions on Image Processing, vol. 19, no. 6, pp. 1657–1663, 2010. [49] ——, “Rotation invariant texture classification using LBP variance (LBPV) with global matching,” Pattern recognition, vol. 43, no. 3, pp. 706–719, 2010. [50] L. Liu, Y. Long, P. W. Fieguth, S. Lao, and G. Zhao, “Brint: binary rotation invariant and noise tolerant texture classification,” IEEE Transactions on Image Processing, vol. 23, no. 7, pp. 3071–3084, 2014. [51] X. Qi, R. Xiao, C.-G. Li, Y. Qiao, J. Guo, and X. Tang, “Pairwise rotation invariant co-occurrence local binary pattern,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 36, no. 11, pp. 2199– 2213, 2014. [52] T. Ahonen, J. Matas, C. He, and M. Pietikäinen, “Rotation invariant image description with local binary pattern histogram fourier features,” Image analysis, pp. 61–70, 2009. [53] G. Zhao, T. Ahonen, J. Matas, and M. Pietikainen, “Rotation-invariant image and video description with local binary pattern features,” IEEE Transactions on Image Processing, vol. 21, no. 4, pp. 1465–1477, 2012. [54] R. Nosaka, Y. Ohkawa, and K. Fukui, “Feature extraction based on cooccurrence of adjacent local binary patterns,” in Pacific-Rim Symposium on Image and Video Technology. Springer, 2011, pp. 82–91. [55] X. Tan and B. Triggs, “Enhanced local texture feature sets for face recognition under difficult lighting conditions,” in International Workshop on Analysis and Modeling of Faces and Gestures. Springer, 2007, pp. 168–182. [56] B. Caputo, E. Hayman, M. Fritz, and J.-O. Eklundh, “Classifying materials in the real world,” Image and Vision Computing, vol. 28, no. 1, pp. 150–163, 2010. [57] T. Ojala, M. Pietikäinen, and D. Harwood, “A comparative study of texture measures with classification based on featured distributions,” Pattern recognition, vol. 29, no. 1, pp. 51–59, 1996. [58] A. Fathi and A. R. Naghsh-Nilchi, “Noise tolerant local binary pattern operator for efficient texture analysis,” Pattern Recognition Letters, vol. 33, no. 9, pp. 1093–1100, 2012. [59] K. Wang, C.-E. Bichot, C. Zhu, and B. Li, “Pixel to patch sampling structure and local neighboring intensity relationship patterns for texture classification,” IEEE Signal Processing Letters, vol. 20, no. 9, pp. 853– 856, 2013. [60] G. R. Iversen and H. Norpoth, Analysis of variance. Sage, 1987, no. 1. [61] J. Ren, X. Jiang, and J. Yuan, “Noise-resistant local binary pattern with an embedded error-correction mechanism,” IEEE Transactions on Image Processing, vol. 22, no. 10, pp. 4049–4060, 2013. [62] J. Bruna and S. Mallat, “Invariant scattering convolution networks,” IEEE transactions on pattern analysis and machine intelligence, vol. 35, no. 8, pp. 1872–1886, 2013. [63] L. Sifre and S. Mallat, “Rotation, scaling and deformation invariant scattering for texture discrimination,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2013, pp. 1233– 1240. [64] T.-H. Chan, K. Jia, S. Gao, J. Lu, Z. Zeng, and Y. Ma, “Pcanet: A simple deep learning baseline for image classification?” IEEE Transactions on Image Processing, vol. 24, no. 12, pp. 5017–5032, 2015. [65] M. Cimpoi, S. Maji, and A. Vedaldi, “Deep filter banks for texture recognition and segmentation,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015, pp. 3828–3836. [66] L. Liu, P. Fieguth, X. Wang, M. Pietikäinen, and D. Hu, “Evaluation of lbp and deep texture descriptors with a new robustness benchmark,” in European Conference on Computer Vision. Springer, 2016, pp. 69–86. [67] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” in Advances in neural information processing systems, 2012, pp. 1097–1105. 

1cs.CV

On the stable Andreadakis problem Jacques Darné March 2, 2018 arXiv:1711.05991v2 [math.AT] 1 Mar 2018 Abstract Let Fn be the free group on n generators. Consider the group IAn of automorphisms of Fn acting trivially on its abelianization. There are two canonical filtrations on IAn : the first one is its lower central series Γ∗ ; the second one is the Andreadakis filtration A∗ , defined from the action on Fn . In this paper, we establish that the canonical morphism between the associated graded Lie rings L(Γ∗ ) and L(A∗ ) is stably surjective. We then investigate a p-restricted version of the Andreadakis problem. A calculation of the Lie algebra of the classical congruence group is also included. Introduction Automorphisms of free groups have been widely studied over the years, from many different points of view. They are linked to the mapping class groups of surfaces and braid groups [FM12a]; they also act on a moduli space of graphs, called the outer space, introduced in [CV86], which is still actively studied nowadays (see, for instance, [BBM07], or [FM12b]). Recently, several results have also been obtained regarding the stable homology of these groups [Gal11, RWW17, DV15, Dja16a]. One way to try and understand the structure of these automorphism groups is to cut them into pieces, by considering a family of subgroups and studying how these interact with each other. Such families of subgroups can arise from the action on the free group Fn and related geometric objects, as is the case with the automorphisms with boundaries (see for instance [JW04] or [DP12]), and for the Andreadakis subgroups, which we now focus on. The first Andreadakis subgroup of Aut(Fn ) is the IA-group. Precisely, we can first look at how automorphisms act on Fnab ∼ = Zn . That is, we can consider the projection from Aut(Fn ) onto GLn (Z). We then put aside this linear part by considering only IAn , the subgroup of automorphisms acting trivially on Zn , which is an algebraic analogue of the Torelli subgroup of the mapping class group. An explicit finite set of generators of IAn has been known for a long time [Nie24] – see also [BBM07, 5.6]. Nevertheless, the structure of IAn remains largely mysterious. For instance, IA3 is not finitely presented [KM97], and it is not known if IAn is finitely presented for n > 3. Recent results about the IA-groups include the finite L-presentation of IAn given in [DP16], or finiteness results on the lower central series of IAn obtained in [CP17]. The IA-group is the first step of the Andreadakis filtration IAn = A1 ⊇ A2 ⊇ · · · , in which Aj is the group of automorphisms acting trivially on Fn /Γj+1 (Fn ), where Fn is filtered by its lower central series Fn = Γ1 (Fn ) ⊇ Γ2 (Fn ) ⊇ · · · . The Andreadakis filtration is an N -series. As such, it contains the minimal N -series on IAn , its lower 1 central series: for all k, Ak ⊇ Γk (IAn ). We are thus led to the problem of comparing these filtrations, that we call the Andreadakis problem. Since the two filtrations are N -series, the associated graded objects are graded Lie rings (that is, Lie algebras over Z), the Lie bracket being induced by the commutator map (x, y) 7→ [x, y] = xyx−1 y −1 . The inclusion i : Γ∗ (IAn ) ⊆ A∗ induces a morphism of Lie rings: M M i∗ : L(Γ∗ (IAn )) = Γj (IAn )/Γj+1 (IAn ) −→ L(A∗ ) = Aj /Aj+1 . (0.0.1) j>1 j>1 Thus, the Andreadakis problem translates into the following question: Problem 1 (Andreadakis). How close is the morphism (0.0.1) to be an isomorphism? Andreadakis conjectured that the filtrations were the same [And65, p. 253]. In [Bar13], Bartholdi disproved the conjecture, using computer calculations. He then tried to prove that the two filtrations were the same up to finite index, but in the erratum [Bar16], he showed that even this weaker statement cannot be true. His proof uses the L-presentation of IAn given in [DP16], to which he applies algorithmic methods described in [BEH08] to calculate (using the software GAP) the first degrees of the Lie algebra associated to each filtration. In this paper, we are interested in the difference between Ak (Fn ) and Γk (IAn ) for n  k, that is, in the stable range. We thus ask the following question: Problem 2 (Andreadakis - stable version). How close is the morphism i∗ : Lk (Γ∗ (IAn )) = Γk (IAn )/Γk+1 (IAn ) −→ Lk (A∗ ) = Ak /Ak+1 (0.0.2) to be an isomorphism when n  k? Our main goal here is to show the following partial answer to this question. Theorem 2.38 (Stable surjectivity). When n > k + 2, the morphism (0.0.2) is surjective. A (weaker) rational version of this theorem has been obtained independently by Massuyeau and Sakasai [MS17, th. 5.1]. Like them, we prove it by building on results from [Sat12], but using quite different methods. These methods include a description of Andreadakis-like filtrations via a categorical framework, allowing us to state and study a p-restricted version of the problem. We answer the questions asked in [HM17] about this problem, and use our answers to study the stable p-restricted Andreadakis problem. Also, we solve the stable q-torsion Andreadakis problem for Zn , getting a complete calculation of the Lie ring of the congruence group GLn (p Z) for n > 5. Let us now describe in more detail the methods we use and the results contained in the present paper. In section 1, we set up a general framework for understanding N -series and their associated Lie algebras. We introduce a category SCF of N -series. We remark that the categorical definition of an action of an object on another makes sense in this category. This allows us to interpret an old construction of Kaloujnine (see theorem 1.16) as the construction of universal actions in SCF. This category is thus action-representative, a situation studied in [BJK05, BB07, Bou08]. Using this language, we are able to recover and generalize several classical constructions: 2 • Taking the graded rings associated to N -series gives a functor L from SCF to the category of Lie rings. This functor preserves actions, and the Johnson morphism admits a nice generalisation as the classifying morphism associated to an action between Lie rings obtained from an action in SCF. • Lazard’s classical construction of N -series from algebra filtrations described in [Laz54] is recovered as a particular case of Kaloujnine’s construction. • We also obtain the filtrations on congruence groups studied in [Lop14]. In particular, we show that the filtration given by the last construction on the classical congruence group GLn (q Z) coincides with its lower central series when n > 5. As a consequence, we get an explicit calculation of this group’s Lie ring (generalizing [LS76, Th. 1.1], which is the degree-one part): Corollary 1.46. For all n > 5 and all q > 3, there is a canonical isomorphism of graded Lie rings (in degrees at least one): L(GLn (q Z)) ∼ = sln (Z /q)[t], where the degree of t is 1, and the Lie bracket of M ti and N tj is [M, N ]ti+j . Section 2 deals with the proof of our stable surjectivity result (Theorem 2.38). The proof relies on the constructions of the first section, applied to Fox’s free differential calculus. The Jacobian matrix map D : f 7→ Df turns out to be a derivation from Aut(Fn ) to GLn (Z Fn ), sending the Andreadakis filtration to the congruence filtration GLn ((IFn )∗ ) (the group algebra Z Fn being filtered by the powers of its augmentation ideal IFn ). We then study such derivations, and the maps they induce on the graded Lie rings associated to N -series they preserve. We thus show that the trace map defined by Tr(f ) = Tr(Df − 1n ) induces a well-defined map: Tr : L(A∗ ) −→ gr(Z Fn ). The graded algebra gr(Z Fn ) is in fact the tensor algebra T V over V = Fnab . A result from [BLGM90] implies that this trace map takes values in [T V, T V ]. Studying the links between free differential calculus and differential calculus in T V , we show that this trace map is exactly the one introduced by Morita [Mor93, Def. 6.4], getting the explicit description in terms of contraction maps notably used by Satoh in [Sat12]. Denoting the Johnson morphisms by τ and τ 0 , we get a commutative diagram of graded linear maps: L(IAn ) τ0 i∗ L(A∗ ) τ Der(LV ) TrM T V, where TrM ◦τ = Tr This gives the following inclusions of subspaces of Der(LV ): Im(τ 0 ) ⊆ L(A∗ ) ⊆ Tr−1 M ([T V, T V ]). We observe that calculations from [Sat12] work over Z. From this , we deduce that the subspaces Im(τ 0 ) and Tr−1 M ([T V, T V ]) are stably the same, so these inclusions are equalities in the stable range, and i∗ must be stably surjective. We close the section by 3 investigating some of the consequences of this result for automorphisms of free nilpotent groups. In section 3, we turn to the p-restricted version of the Andreadakis problem. Precisely, we can do the same construction as above, replacing the lower central series [p] Γ∗ (Fn ) by the mod-p lower central series Γ∗ (Fn ), which is an Np -series: [p] [p] (Γi )p ⊆ Γpi . [p] Kaloujnine’s construction gives an associated Andreadakis filtration A∗ on the group IA[p] of automorphisms of Fn acting trivially on Fnab ⊗ Fp . This filtration was shown in [p] [HM17] to be an Np -series. It then contains the minimal Np -series Γ∗ (IA[p] ). Whence the natural question: Problem 3 (Andreadakis – p-restricted version). What is the difference between the [p] [p] [p] Np -series A∗ and Γ∗ (IAn ) ? Answering to [HM17, rk. 8.6], we show that these two filtrations fit in the same kind of nice machinery as their classical counterparts, but they turn out to always differ. The paper ends on a quantification of the lack of stable surjectivity in this p-restricted case (see Proposition 3.20). Acknowledgements: This work is part of the author’s PhD thesis. The author is indebted to his advisors, Antoine Touzé and Aurélien Djament, for their constant support, countless helpful discussions, and numerous remarks and comments on earlier versions of the present paper. He also thanks Gwenaël Massuyeau and Takao Satoh, who kindly agreed to be the reviewers of his thesis, for their useful observations and comments. Contents Introduction 1 1 Generalities on strongly central series 1.1 Notations and reminders . . . . . . . . . . . . . . . . . 1.2 Strongly central filtrations and Lie algebras . . . . . . 1.3 Actions in the category of strongly central filtrations . 1.4 Johnson’s morphisms . . . . . . . . . . . . . . . . . . . 1.5 The Andreadakis problem . . . . . . . . . . . . . . . . 1.6 Lazard’s theorem . . . . . . . . . . . . . . . . . . . . . 1.7 Congruence groups . . . . . . . . . . . . . . . . . . . . 1.8 Comparison between filtrations obtained from an action 2 Traces and stable surjectivity 2.1 Free differential calculus . . . . . . . . . . . . 2.2 Derivations and strongly central filtrations . . 2.3 Algebras, actions and derivations . . . . . . . 2.4 Traces . . . . . . . . . . . . . . . . . . . . . . 2.4.1 The induced map between Lie algebras 2.4.2 Introducing the contraction map . . . . 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 5 6 7 11 12 14 15 17 . . . . . . 19 19 21 22 25 26 27 2.5 Stable surjectivity . . . . . . . . . . . . . . . . . . 2.5.1 Vanishing of the trace map . . . . . . . . . 2.5.2 Linear algebra . . . . . . . . . . . . . . . . 2.5.3 Stable cokernel of τ 0 and stable surjectivity Automorphisms of free nilpotent groups . . . . . 2.6 3 The 3.1 3.2 3.3 3.4 1 case of positive caracteristic Dark’s theorem . . . . . . . . . . . . . . . . p-Restricted strongly central series . . . . . The p-restricted Andreadakis problem . . . . The stable p-restricted Andreadakis problem 3.4.1 Vanishing of the trace map . . . . . . 3.4.2 Linear algebra . . . . . . . . . . . . . 3.4.3 Stable cokernel of i∗ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 28 30 30 31 . . . . . . . 33 33 34 35 37 37 38 39 Generalities on strongly central series 1.1 Notations and reminders Throughout the paper, G will denote an arbitrary group, and k a commutative unitary ring. The left and right action of G on itself by conjugation are denoted respectively by xy = y −1 xy and yx = yxy −1 . The commutator of two elements x and y in G is denoted by: [x, y] := xyx−1 y −1 . If A and B are subsets of G, we denote by [A, B] the subgroup generated by the commutators [a, b] with (a, b) ∈ A×B. If A and B are stable by conjugation by elements of G (resp. by all automorphisms of G), then [A, B] is a normal (resp. characteristic) subgroup of G. For instance, [G, G] is a characteristic subgroup of G, called the derived subgroup of G. The quotient Gab := G/[G, G] is the abelianization of G, its bigger abelian quotient. The derived subgroup is the second step of a filtration of G by characteristic subgroups: Definition 1.1. The lower central series of G, denoted by Γ∗ (G), or shortly Γ∗ , is the filtration of G defined by: ( Γ1 := G, Γk+1 := [G, Γk ]. Definition 1.2. A group G is said to be nilpotent if its lower central series stops. The least integer c such that Γc+1 (G) = {1} is then G’s nilpotency class. More generally, G is said to be residually nilpotent if its lower central series is separated, i.e. if: \ Γi (G) = {1}. i One can easily check the following formulas: Proposition 1.3. For all x, y, z ∈ G, • [x, x] = 1, • [x, y]−1 = [y, x], • [x, yz] = [x, y] (y [x, z]) , 5 • [[x, y], yz] · [[y, z], zx] · [[z, x], xy] = 1, x z y • [[x, y −1 ], z −1 ] · [[z, x−1 ], y −1 ] · [[y, z −1 ], x−1 ] = 1. The last ones are two versions of the Witt-Hall identity, which implies the following: Lemma 1.4 (3-subgroups lemma). Let A, B and C be three subgroups of a group G. If two of the three following subgroups are trivial, then so is the third: [A, [B, C]], [B, [C, A]], [C, [A, B]]. Equivalently, one of them is contained in the normal closure of the two others. 1.2 Strongly central filtrations and Lie algebras The theory of strongly central series has notably been studied by M. Lazard [Laz54]. Definition 1.5. Let G be a group. A strongly central filtration of G (also called strongly central series or N -series) is a filtration G = G1 ⊇ · · · ⊇ Gi ⊇ · · · of G by subgroups, satisfying: ∀i, j > 1, [Gi , Gj ] ⊆ Gi+j . Remark that indexation has to begin from G = G1 . In particular, [G, Ni ] ⊆ Ni+1 ⊆ Ni , which means exactly that the Ni are normal subgroups of G. Proposition 1.6. Let G be a group. The lower central series Γ∗ (G) is a strongly central series on G, and it is the minimal one. Proof. The strong centrality is shown by induction, using the 3-subgroup lemma 1.4. Given a strongly central filtration G∗ of G = G1 , a straightforward induction then gives: Gi ⊇ Γi (G) for any i > 1. Let G = G1 ⊇ · · · ⊇ Gk ⊇ · · · be any strongly central filtration of a group G. The quotients Li (N∗ ) := Ni /Ni+1 are abelian (for any i > 1), since [Ni , Ni ] ⊆ N2i ⊆ Ni+1 . The graded abelian group M L(N∗ ) := Li (N∗ ), i>1 is endowed with a bracket induced by the commutator map (x, y) 7→ [x, y] of G. Using the formulas 1.3, one easily checks that this defines a Lie bracket: L(G∗ ) is a Lie ring (i.e. a Lie algebra over Z). Notation 1.7. We denote L(Γ∗ (G)) by L(G) (that is, if we do not specify a strongly central filtration on a group, it is understood to be filtered by its lower central series). Example 1.8. If G is a free group, then L(G) is the free Lie algebra over the Z-module Gab [Laz54, th. 4.2]. As products of commutators become sums of brackets inside the Lie algebra, the following fundamental property follows from the definition of the lower central series: Proposition 1.9. The Lie ring L(G) is generated in degree 1. Precisely, it is generated (as a Lie ring) by L1 (G) = Gab . As a consequence, if G is of finite type, then each Ln (G) is too. 6 1.3 Actions in the category of strongly central filtrations Let SCF be the category whose objects are the strongly central filtrations, where morphisms between G∗ and H∗ are the group morphisms from G1 to H1 preserving filtrations. There is a forgetful functor ω1 : SCF −→ Grps defined by G∗ 7→ G1 . This functor admits a left adjoint Γ : G 7→ Γ∗ (G) (see Proposition 1.6). It also admits a right adjoint G 7→ (G, G, ...). Proposition 1.10. The category SCF is complete and cocomplete, and is homological (but not semi-abelian). A general reference on homological categories is [BB04]. The reader can also consult [HL11] for a simple version of the axioms defining homological cocomplete and semiabelian categories. Proof of Proposition 1.10. The forgetful functor ω1 admits both a left and a right adjoint, so it has to commute to limits and colimits. It does not create either of them (in the sense of [ML98], V.1), but it almost does. Precisely, let F : D −→ SCF be a diagram. The colimit G∞ of the group diagram ω1 F is in general endowed with several strongly central filtrations making ω1 F (d) −→ G∞ into filtration-preserving morphisms (for instance the trivial one). One checks easily that the minimal such filtration (which is the intersection of all those) is the colimit of F . Similarly, the limit G∞ of the group diagram ω1 F is endowed with several strongly central filtrations making ϕd : G∞ −→ ω1 F (d) into filtration-preserving morphisms (for instance its lower central series). However, the maximal such filtration is the limit of F . It is explicitly described as: \ G∞ = ϕ−1 ∗ d (F (d)). d To check that SCF is homological, one can check the axioms given in [HL11]. It is not semi-abelian, because there are equivalence relation R∗ ⊆ G2∗ for which R∗ is not the induced filtration on R1 (like in topological groups – or more generally in the categories of topological algebras considered in [BC05] – where an equivalence relation does not have to be endowed with the induced topology). In a homological category, we need to distinguish between usual epimorphisms (resp. monomorphisms) and regular ones, that is, the ones obtained as coequalizers (resp. equalizers). In SCF, the former are the u such that u1 = ω1 (u) is an epimorphism (resp. a monomorphism), whereas the latter are surjections (resp. injections): Definition 1.11. Let u : G∗ −→ H∗ be a morphism in SCF. It is called an injection (resp. a surjection) when u1 is injective (resp. surjective) and u−1 (Hi ) = Gi (resp. u(Gi ) = Hi ) for all i. Examples of homological categories include abelian categories, the category Grps of groups, or the category Lie of Lie algebras. The usual lemmas of homological algebra (the nine lemma, the snake lemma, the five lemma...) are true in these categories. Homological categories differ from abelian ones notably by the fact that in general, two split extensions between the same objects are not isomorphic (by an isomorphism preserving the splittings). This allows us to define an action of an object on another. 7 Definition 1.12. Let C be a homological category. If X and Z are two objects of C, we define an action of Z on X as a split extension (with a given splitting): X Y Z. When such an action is given, we will say that Z acts on X, and write: Z  X. This definition (which needs only the weaker setting of pointed protomodular categories to make sense) is motivated by the situation in Grps, where an action of a group K on a group G is encoded by a semi-direct product structure G o K. Remark 1.13. The choice of splitting is crucial here. For instance, the canonical extension:   X   0 1 1   0 X ×X  0  1 X   1 1 can be split by , or by the diagonal . The first choice gives the trivial action, whereas the second one gives the adjoint action, which is highly non-trivial: in Lie, this gives the adjoint representation; in Grps, we get the action of a group on itself by conjugation. The set Act(Z, X) of actions of Z on X is a contravariant functor in Z: the restriction of an action along a morphism is defined via a pullback. In Grps, as in Lie, this functor is representable, for any X. Indeed, an action of a group K on a group G is given by a morphism K −→ Aut(G). Similarly, an action of a Lie algebra k on a Lie algebra g is given by a morphism k −→ Der(g), where Der(g) is the Lie algebra of derivations from g to itself. Recall that a derivation ∂ is a linear map satisfying: ∂([x, y]) = [∂x, y] + [x, ∂y]. The situation when actions are representable has notably been studied in [BJK05], and in several subsequent papers [BB07, Bou08]. The following terminology was introduced in [BB07, Def. 1.1]: Definition 1.14. A homological category C is said to be action-representative when the functor Act(−, X) is representable, for any object X ∈ C. Our goal for the rest of this section is to construct universal actions in SCF, getting in particular the following result: Proposition 1.15. The category SCF is action-representative. A representative for Act(−, X) is a universal action on X. Explicitly, it is an action of an object A(X) on X such that any action Z  X is obtained by restriction along a unique morphism Z → A(X). For instance, in Grps, the universal action on G is: G G o Aut(G) Aut(G), where the group G o Aut(G) is the holomorph of G. Its underlying set is G × Aut(G), endowed with the product defined by (g, σ) · (h, τ ) := (gσ(h), στ ). The construction of universal actions in SCF is given by a theorem of Kaloujnine [Kal50a, Kal50b], quoted by Lazard in [Laz54, p. 117]: 8 Theorem 1.16 (Kaloujnine). Let G∗ be a strongly central series. Let j > 1 be an integer. Let Aj (G∗ ) ⊆ Aut(G∗ ) be the subgroup of automorphisms acting trivially on every quotient Gi /Gi+j . Then A∗ (G∗ ) is a strongly central series. Notation 1.17. As for Lie rings (Notation 1.7) we abbreviate A∗ (Γ∗ (G)) to A∗ (G). We can rewrite the definition of Aj (G∗ ) given in the theorem as: ! Y Aj (G∗ ) = ker Aut(G∗ ) −→ Aut (Gi /Gi+j ) . i Identifying G and Aut(G) to the subgroups G × 1 and 1 × Aut(G) of the holomorph G o Aut(G), we can define the commutator of an automorphism with an element of G: [σ, g] = σ(g)g −1 . Note that [Aut(G), G] ⊆ G. Using this point of view, we can rephrase the previous definition: Aj (G∗ ) = { σ ∈ Aut(G∗ ) | ∀i > 1, [σ, Gi ] ⊆ Gi+j } ⊆ G1 o Aut(G∗ ). (1.17.1) Proof of Theorem 1.16. We abbreviate Aj (G∗ ) to Aj . Obviously, Aj+1 ⊆ Aj . We show the strong centrality using the 3-subgroup lemma (Lemma 1.4). Precisely, let α, β > 1 be two integers. For all i > 1, the group Gi+α+β is normal in G o Aut(G∗ ) (it is normal in G and Aut(G∗ )-stable). Lemma 1.4 thus implies: [[Aα , Aβ ], Gi ] ⊆ Gi+α+β . This says exactly that [Aα , Aβ ] ⊆ Aα+β , which is the desired conclusion. Remark 1.18. The group A1 (G∗ ) is the group of automorphisms of G1 preserving G∗ and acting trivially on L(G∗ ). Example 1.19. Let G∗ = Γ∗ (G). Then L(G∗ ) is generated in degree one as a Lie algebra. As a consequence A1 (G∗ ) is the subgroup of automorphisms acting trivially on the abelianization Gab = L1 (G), denoted by IAG . In order to show that the filtration A∗ (G∗ ) acts universally on G∗ , we need to investigate actions in SCF. Proposition 1.20. An action K∗  G∗ in SFC consists of a group action of K = K1 on G = G1 such that : ∀i, j, [Ki , Gj ] ⊆ Gi+j . Proof. Let an action of K∗ on G∗ be given: G∗ H∗ K∗ . The forgetful functor ω1 : G∗ 7→ G1 from SCF to groups has a left adjoint G 7→ Γ∗ G. Hence, it commutes with kernels. Applying this functor to the given action, we get a split extension of groups: G1 H1 9 K1 . The group H1 thus decomposes as a semi-direct product G1 o K1 of G1 by K1 . In fact, there are other forgetful functors ωi : G∗ 7→ Gi , each one with a left adjoint G 7→ Γd ∗i e G (where d−e is the usual ceiling function). We then get split extensions of groups: Gi Hi Ki . The groups Hi thus decomposes as semi-direct products Gi o Ki of Gi by Ki . As H∗ is a strongly central filtration, we can apply Lemma 1.21 below to get the desired relation. Conversely, let a group action be given as in the statement of the proposition. Using the same lemma, we see that H∗ = G∗ o K∗ is a strongly central filtration on H = K o G, and the corresponding split sequence in SFC is exact. Lemma 1.21. Let K  G be an action in Grps, encoded in a semi-direct product structure H = G o K. Let G∗ and K∗ be given filtrations on G = G1 and K = K1 respectively. Then the Hi := Gi o Ki are subgroups of H defining a strongly central filtration of H if and only if:   K∗ is a strongly central series on K, G∗ is a strongly central series on G,  ∀i, j, [Kj , Gi ] ⊆ Gi+j . Proof. Suppose first that (Gi o Ki )i is strongly central. Then its projection K∗ on K also is. So is its intersection G∗ with G. Hence, the conclusion follows from: [Kj , Gi ] ⊆ (Gi+j o Ki+j ) ∩ G = Gi+j . Conversely, under the hypothesis listed above, Gi is stable under the action of Ki , so the Hi = Gi o Ki are subgroups of H. We then use the formulas 1.3 to compute [kg, k 0 g 0 ] with k, g, k 0 , g 0 in Ki , Gi , Kj and Gj respectively: 0 [kg, k 0 g 0 ] = [kg, k 0 ] ·g [kg, g 0 ] 0 0 = k[g, k 0 ] · [k, k 0 ] · k k[g, g 0 ] · k [k, g 0 ] ∈ Gi+j · Ki+j · Gi+j · Gi+j = Gi+j o Ki+j . This completes the proof. We are now ready to show the result announced in Proposition 1.15: Proposition 1.22. Let G∗ be a strongly central series. The strongly central series A∗ (G∗ ) acts canonically on G∗ , and this action is universal. Proof. That A∗ (G∗ ) acts on G∗ follows from the formula (1.17.1), Theorem 1.16 and Proposition 1.20. Given an action of a strongly central series K∗ on G∗ , the underlying group action is described by a unique morphism from K1 to Aut(G1 ). From Proposition 1.20, we deduce that this morphism sends Kj into Aj (G∗ ). Conversely, any morphism from K∗ to A∗ (G∗ ) in SCF gives a group action lifting to an action in SCF by Proposition 1.20. Remark 1.23. If a group K acts on a group G, and G∗ is a strongly central filtration on G = G1 , we can pull back the canonical filtration A∗ (G∗ ) by the associated morphism: K −→ Aut(G). 10 This gives a strongly central filtration A∗ (K, G∗ ), maximal amongst strongly central filtrations on subgroups of K which act on G∗ via the given action K  G. It can be described explicitly as: Aj (K, G∗ ) = { k ∈ K | ∀i > 1, [k, Gi ] ⊆ Gi+j } ⊆ K. 1.4 Johnson’s morphisms The construction of the Johnson morphism associated with an action in SCF relies on the following: Proposition 1.24. The Lie functor L : SCF −→ LieZ is exact, i.e. it preserves short exact sequences [BB04, Def 4.1.5]. Proof. The exactness of L is equivalent to the exactness of each Li : SCF −→ Ab. Consider the forgetful functors ωi : G∗ 7→ Gi from SCF to Grps. Each ωi has a left adjoint G 7→ Γd ∗i e G, so these functors preserve kernels. Moreover, they preserve regular epimorphisms, which are surjections in SCF (cf. Definition 1.11). Hence, they are exact. Since Li is the cokernel of the injection ωi+1 ,→ ωi , its exactness follows from the nine lemma in Grpes. Precisely, if E is any short exact sequence in SCF, apply the nine lemma to the diagram ωi+1 (E) ,→ ωi (E)  Li (E) to get the exactness of Li (E) from that of ωi (E) and ωi+1 (E). As a consequence of Proposition 1.24, the functor L preserves actions. Precisely, from an action in SCF: G∗ H∗ K∗ , we get an action in the category of graded Lie rings: L(G∗ ) L(H∗ ) L(K∗ ). Such an action is given by a morphism of graded Lie rings: L(K∗ ) −→ Der∗ (L(G∗ )). (1.24.1) The target is the (graded) Lie algebra of graded derivations: a derivation is of degree k when it raises degrees of homogeneous elements by k. Definition 1.25. The morphism (1.24.1) is called the Johnson morphism associated to the given action K∗  G∗ . We can give an explicit description of this morphism: for k ∈ K, the derivation associated to k̄ is induced by [k̄, −] inside L(G∗ oK∗ ) = L(G∗ )oL(K∗ ), so it is induced by [k, −] inside G∗ o K∗ . Example 1.26. The Johnson morphism associated to the universal action A∗ (G∗ )  G∗ is the Lie morphism: τ : L(A∗ (G∗ )) −→ Der∗ (L(G∗ )) induced by σ 7→ (x 7→ σ(x)x−1 ). 11 If K1 is in fact a normal subgroup of a group K0 , such that each Ki is normal in K0 , and such that the action of K1 on G∗ can be extended to an action of K0 , then K0 acts on L(G∗ o K∗ ) and this action factorizes through K0 /K1 . Moreover, as this action is by automorphisms of the Lie ring: Lemma 1.27. Let K0 .K1 be given as above. Then the Johnson morphism τ is K0 /K1 equivariant. The action of K0 on derivations is by conjugation. Precisely: τ (k · x) = [k · x, −] = k · [x, k −1 · −] = k ◦ τ (x) ◦ k −1 . Lemma 1.28. Let K∗  H∗ be an action in SCF. The associated Johnson morphism τ : L(K∗ ) −→ Der(L(G∗ )) is injective if and only if K∗ is the filtration A∗ (K, G∗ ) defined in Remark 1.23. Proof. Every non-trivial element in ker(τj ) lifts to an element in Aj (K, G∗ ) − Kj . Conversely, an element σ in Aj (K, G∗ ) − Kj is in Kk − Kk+1 for some k < j. Then σ ∈ Kk /Kk+1 − {0} is a non-trivial element in ker(τ ). Remark 1.29. The definition of Aj (K, G∗ ) makes sense for j = 0, giving a subgroup K0 = A0 (K, G∗ ) of K acting as above. The morphism τ is then A0 (K, G∗ )-equivariant. In fact, τ can then be extended to a morphism of extended Lie algebras, in the sense of [HM17]. Ideed, their construction of an algebra of extended derivations is exactly a construction of universal actions in the category of extended Lie algebras, and their version of the Johnson morphism is exactly the one we find if we replace N -series and Lie algebras by their extended version in the constructions above. 1.5 The Andreadakis problem Let G be a group. To study the structure of Aut(G), we can consider first how automorphisms act on Gab . Then we can put aside this linear part by considering the kernel of the projection from Aut(G) to GL(Gab ). This kernel IAG is (residually) nilpotent when G is, and in endowed with two strongly central filtrations: its lower central series, and the Andreadakis filtration A∗ (G). We are thus led to the problem of comparing these filtrations, which we call the Andreadakis problem (Problem 1). Recall from Theorem 1.16 and Example 1.19 that A∗ (G) is a strongly central filtration on A1 (G) = IAG . Since A1 (G) is the kernel of the canonical action of Aut(G) on L(G), we get an induced faithful action of Aut(G)/IAG (6 GL(Gab )) on L(G). The next lemma gives a similar concrete description of all the Aj (G): Lemma 1.30. The group Aj (G) is the subgroup of automorphisms acting trivially on G/Γj+1 (G), i.e. Aj (G) = { σ ∈ Aut(G) | [σ, G] ⊆ Γj+1 (G) } . Proof. Let us denote by Kj the right-hand side of the equality. We only need to show the inclusion Kj ⊆ Aj (G). The case j = 1 is given in Example 1.19. Suppose it true for j − 1, and let σ ∈ Kj . In particular, σ is inside Kj−1 = Aj−1 (G), so it induces a derivation [σ, −] of degree j − 1 of L(G), via the Johnson morphism. By definition, [σ, G] ⊆ Γj+1 . This says exactly that the derivation [σ, −] is trivial on L1 (G). Since L(G) is generated in degree one (Proposition 1.9), [σ, −] has to be trivial on all of L(G): for all i, [σ, Γi ] ⊆ Γi+j , which is the desired conclusion. 12 The filtration A∗ (G) is strongly central on A1 (G) = IAG . As a consequence, Ak (G) contains Γk (IAG ). A first consequence of this inclusion is that the (residual) nilpotency of G implies the (residual) nilpotency of IAG . Precisely, let G be a c-nilpotent group. Then Ac (G) = {1G }. Accordingly, Γc (IAG ) = {1G }, so IAG is (c − 1)-nilpotent. In a similar fashion, one can check that IAG has to be residually nilpotent when G is. The following question is crucial for trying to understand the structure of automorphism groups of residually nilpotent groups, in particular for trying to understand the structure of Aut(Fn ): Problem 1 (Andreadakis). What is the difference between A∗ (G) and Γ∗ (IAG ) ? Example 1.31. Consider the alternating group An . When n 6= 2, 6, Aut(An ) = Σn (acting by conjugation), as is easily deduced from [Rot95, cor. 7.5]. But if n > 5, then An is perfect: [An , An ] = An . On the one hand, the Andreadakis filtration is thus constant equal to IA(An ) = Aut(An ) = Σn . On the other hand, the lower central series of Σn is Σn , An , An , ..., so the two filtrations differ in this case. Recall from the Introduction that we are interested in a stable form of the problem for G = Fn : Problem 2 (Andreadakis - stable version). What is the difference between Ak (Fn ) and Γk (IAn ) for n  k ? The Johnson morphism turns out to be a powerful tool in the study of the Andreadakis filtration (which is the analogous of Johnson’s filtration on the mapping class group, defined by Johnson in [Joh83]). Recall the A∗ (G) acts on Γ∗ (G) (see Proposition 1.22). The associated Johnson morphism τ is described in Example 1.26. The Γi (G) are characteristic subgroups of G, so A0 (G) = Aut(G). Lemmas 1.27 and 1.28 then give: Lemma 1.32. The morphism τ is injective and Aut(G)/IAG -equivariant. The filtration Γ∗ (IAG ) also acts on Γ∗ (G), by pulling back the universal action along the inclusion i : Γ∗ (IAG ) ⊆ A∗ (Fn ). The associated Johnson morphism is denoted by: τ 0 = τ ◦ i∗ . The morphism τ 0 is still equivariant, but is injective if and only if G satisfies Andreadakis’ conjecture (i.e. if i∗ = 1 ). Example 1.33. If G is a free group, then L(Fn ) is the free Lie algebra LV on V = Gab . If g  h is an action in Lie, then g also acts on ho , which is obtained by considering the module h as an abelian Lie algebra. Derivations from g to h then identify with sections of the projection ho o g  g. Hence, the free Lie algebra is also free with respect to derivations. In particular: Derk (LV ) ∼ = Homk (V, Lk V ) ∼ = V ∗ ⊗ Lk V. The following result is well-known [Kaw06, th. 6.1]: Proposition 1.34. In degree one, the Johnson morphism τ 0 is a GLn (Z)-equivariant isomorphism: 2 ∼ ∗ τ10 : IAab n = V ⊗ Λ V. Proof. The group IAn is generated by the following elements [Nie24] – see also [BBM07, 5.6]: ( ( xj xi x−1 if t = i [xj , xk ]xi if t = i j and Kijk : xt 7−→ (1.34.1) Kij : xt 7−→ xt else xt else. One can check by a direct calculation that these generators are sent to a basis of the free abelian group V ∗ ⊗ Λ2 V . 13 1.6 Lazard’s theorem Definition 1.35. A filtered algebra A∗ is an associative k-algebra A0 endowed with a filtration by ideals: A = A0 ⊇ A1 ⊇ ... such that: ∀i, j, Ai Aj ⊆ Ai+j . We denote by f Alg the category of filtered algebras (and filtration-preserving morphisms). Example 1.36. Let kG be the group algebra of G with coefficients in the (commutative) ring k. We denote by ε : kG → k its canonical augmentation and by Ik G := ker(ε) its augmentation ideal (we will sometimes write IG, or even I, for short). Then kG is filtered by the powers I ∗ G of its augmentation ideal. If G is a free group, then gr(kG) is the tensor algebra over Gab ⊗ k [Pas79, th. 6.2]. From Theorem 1.16, we can deduce the useful corollary [Laz54, th. 3.1]: Theorem 1.37 (Lazard). Let A = A0 ⊃ A1 ⊇ · · · be a filtered algebra. Then A× ∗ := × A× ∩ (1 + A∗ ) is a strongly central filtration on A× ⊆ A , and (−) − 1 induces an 1 embedding of graded Lie algebras: L(A× ∗ ) ,→ gr∗ (A1 ). α Remark 1.38. Fix a morphism G −→ A× . We can pull back the filtration given by the theorem to get a strongly central filtration α−1 (1 + A∗ ) on G1 = α−1 (1 + A1 ). Proof of Theorem 1.37. The filtration A = A0 ⊃ A1 ⊇ · · · can be seen as a filtration of the abelian group A. As A is abelian, it has to be strongly central. Consider the action by left multiplication: ρ : A× −→ Aut(A, +). Let a ∈ A× . One can easily check that ρ(a) ∈ Aj (A∗ ) (where Aj (A∗ ) is the filtration defined in Kaloujnine’s theorem 1.16) if and only if a ∈ 1 + Aj . Thus, A× ∩ (1 + Ai ) = ρ−1 (Aj (A∗ )) is a strongly central filtration, as announced. It remains to show that ∂ = α − 1 induces a morphism of Lie ring (necessarily injective). This can be checked directly from the formula: [g, h] − 1 = [g − 1, h − 1]g −1 h−1 . We will give a slightly different proof later on, using the concept of derivations (see Paragraph 2.2). Example 1.39. Applying Remark 1.38 to the inclusion of G in kG filtered by the powers of the augmentation algebra, we get the dimension series of G: D∗k G = G ∩ (1 + Ik∗ G). It is a strongly central series on G, so it contains Γ∗ G. The question of the equality of D∗Z G and Γ∗ G was known as the dimension subgroup problem during a long time, until an example of a group for which the two filtration differ was given in [Rip72]. See [MP09, chap. 2] for more on this subject. If G is a free group, then L(D∗Z G) is the sub-Lie ring generated in degree one in the tensor algebra gr(Z G) ∼ = T V . Hence (by the PBW theorem), it is the free Lie ring. It then has to coincide with L(G), so D∗Z G = Γ∗ G. Thus, free groups have the dimension property. 14 Lazard’s theorem gives a construction of a strongly central filtration from a filtered algebra. Conversely, we can define a filtered algebra from a strongly central filtration G∗ on G = G1 . Indeed, let kG be filtered by: Fi := kG · (Ni − 1) = ker(kG −→ k(G/Hi )). This filtration does not make kG into a filtered algebra, but it generates a filtration which does: X akj (N∗ ) := kG · (Ni1 − 1) · · · (Nin − 1). i1 +···+in >j One can easily check that these constructions are universal: Proposition 1.40. The above constructions define an adjunction: SCF 1.7 ak∗ (−)× ∗ f Algk . Congruence groups If I is an (associative) ring without unit, recall that its congruence group GLn (I) is defined as: GLn (I) := ker(GLn (A) −→ GLn (A/I)), where A is any unitary (associative) ring containing I as a (two-sided) ideal, (e.g. A = I o Z). This group depends only on I, as it is exactly (1 + Mn (I))× . If A = A0 ⊇ A1 ⊇ · · · is a filtered algebra (see Definition 1.35), then so is the matrix algebra Mn (A), endowed with the filtration Mn (A∗ ). Theorem 1.37 gives us a strongly central filtration of the congruence group: GLn (A1 ) = GLn (A) ∩ (1 + Mn (A1 )) = ker(GLn (A) −→ GLn (A/A1 )). by congruence subgroups: GLn (Aj ) = GLn (A) ∩ (1 + Mn (Aj )) = ker(GLn (A) −→ GLn (A/Aj )), and an embedding of the associated Lie ring into a matrix algebra: L(GLn (A∗ )) ,→ gr∗ (Mn (A∗ )) ∼ = Mn (gr∗ (A∗ )). As in the proof of Theorem 1.37, this filtration can be interpreted as: GLn (A∗ ) = A∗ (GLn (A), An∗ ). We also can recover Lazard’s theorem as the case n = 1 of this construction. Suppose that A∗ is commutative. Then the usual determinant defines a filtrationpreserving morphism: det : GLn (A∗ ) −→ GL1 (A∗ ) = A× ∗. Indeed, if M ∈ Mn (Aj ), then det(1 +M ) ∈ 1+Aj . The following proposition determines the associated graded morphism: 15 Proposition 1.41. The following square commutes: L(GLn (A∗ )) (−)−1 Mn (gr(A∗ )) det L(A× ∗) Tr (−)−1 gr(A∗ ). Moreover, it is Cartesian, that is: L(GLn (A∗ )) ∼ = Tr−1 (L(A× ∗ ) − 1). The kernels of det and Tr then coincide. We thus recover (an generalize slightly) a result of [Lop14]: Corollary 1.42. Let SLn (A∗ ) be the kernel of the determinant. Then: L(SLn (A∗ )) ∼ = ker(Tr) = sln (gr(A∗ )). Remark 1.43. If L(A× ∗ ) = 0, then L(GLn (A∗ )) = L(SLn (A∗ )), and this Lie ring identifies to sln (gr(A∗ )). This happens for example when GL1 (A1 ) = {1} (implying SLn (A∗ ) = GLn (A∗ )), which is verified for A∗ = q ∗ Z (if q > 2), or A∗ = t∗ k[t]. Proof of Proposition 1.41. Let M ∈ Mn (Aj ). Then: det(1 + M ) ≡ 1 + Tr(M ) (mod A2j ). When j > 1, then A2j ⊆ Aj+1 , so this formula gives the commutativity of the above square. The module Tr−1 (Lj (A× ∗ ) − 1) is additively generated by the matrices: āeαβ , ā(e11 − eαα ) and b̄e11 , for α 6= β, a ∈ Aj and 1 + b ∈ A× j . Replacing ā(e11 − eαα ) by ā(e11 + e1α − eα1 − eαα ), we can lift these to GLn (Aj ) as follows:  lifts āeαβ ,  1 + aeαβ 1 + a(e11 + e1α − eα1 − eαα ) lifts ā(e11 + e1α − eα1 − eαα ),  1 + be11 lifts b̄e11 . This completes the proof. Let n > 3. If A is a ”non-totally-imaginary Dedekind ring of arithmetic type” and q is an ideal of A, then we have [BMS67, cor. 4.3 (b)] that SLn (q) is normally generated in SLn (A) (in fact in En (A)) by the shear mappings 1 + teαβ with α 6= β and t ∈ q. This applies for instance to A = Z and q = (q). From this we deduce: Proposition 1.44. If n > 5, under the above hypothesis: Γ∗ (SLn (q)) = SLn (q∗ ). Proof. The filtration SLn (q∗ ) is strongly central on SLn (q), so it contains its lower central series. Conversely, using that:  1 + (a + b)eαβ = (1 + aeαβ )(1 + beαβ ) if α 6= β, 1 + abeαβ = [1 + aeαγ , 1 + beγβ ] if α, β and γ are pairwise distinct, one can easily check that for any t in qk and any α 6= β, if n > 5: 1 + teαβ ∈ Γk (SLn (q)). Using the result from [BMS67], we see that these generate SLn (qk ) as a normal subgroup of SLn (A). Hence SLn (qk ) ⊆ Γk (SLn (q)), as required. 16 Remark 1.45 (On the q-torsion Andreadakis problem for Zn ). Fix an integer q. A qtorsion strongly central filtration is a strongly central filtration G∗ such that Gqi ⊆ Gi+1 for every i > 1 (this means exactly that q L(G∗ ) = 0). On any group G, there is (q) a minimal q-torsion strongly central filtration Γ∗ (G). Moreover, A∗ (G∗ ) is q-torsion whenever G∗ is, because L(A∗ (G∗ )) embeds into a q-torsion Lie algebra via the Johnson (q) (q) (q) morphism. Let us denote A∗ (Γ∗ (G)) by A∗ (G) and A1 (G) by IA(q) (G) (it is the group of automorphisms acting trivially on Gab ⊗ (Z /q)). Thus we get an inclusion of (q) (q) Γ∗ (IA(q) ) into A∗ and a corresponding q-torsion Andreadakis problem. (q) Apply this with G = Zn and q > 3. Then Γ∗ (G) = q ∗ Z, Aut(G) = GLn (Z), (q) and A∗ (G) = GLn (q ∗ Z) = SLn (q ∗ Z) Proposition 1.44 then gives an answer to the (q) q-torsion Andreadakis problem for Zn in the stable range (n > 5) : A∗ (Zn ) is the lower central series of IA(q) (Zn ) = SLn (q Z). When A = Z and q = (q), the graded ring gr(q ∗ Z) is (Z /q)[t]. Moreover, GLn (q Z) = SLn (q Z) (see Remark 1.43). Proposition 1.44 and Corollary 1.42 thus give: Corollary 1.46. For all n > 5 and all q > 3, there is a canonical isomorphism of graded Lie rings (in degrees at least one): L(GLn (q Z)) ∼ = sln (Z /q)[t], where the degree of t is 1, and the Lie bracket of M ti and N tj is [M, N ]ti+j . Remark 1.47. This generalizes [LS76, Th. 1.1], which is the degree-one part. Remark 1.48. For n = 2 and q > 5 a prime number, the group SL2 (q Z) is free on 1 + q(q 2 − 1)/12 generators [Gro52, Fra33]. Its Lie ring is then a free Lie ring on the same number of generators. The author does not know a complete calculation for n = 3 or 4 : the above calculus does give the abelianization, but it fails to determine the whole lower central series. 1.8 Comparison between filtrations obtained from an action Let G be a group. Suppose that G acts on two strongly central series H∗ et K∗ (by automorphisms in SCF). We then can ask what link exists between A∗ (G, H∗ ) and A∗ (G, K∗ ) (as defined in Remark 1.23), depending on the links between H∗ an K∗ . The next proposition describes the behaviour of the construction A∗ (G, −) with respect to injections, surjections (see Definition 1.11) and semi-direct products in the category G−SCF of strongly central series endowed with a G-action (where morphisms respect this action): Proposition 1.49. Let G be a group acting on strongly central series N∗ , H∗ and K∗ . Let u : N∗ −→ H∗ and v : H∗ −→ K∗ be G-equivariant morphisms. If u : N∗ −→ H∗ is an injection, then A∗ (G, H∗ ) ⊆ A∗ (G, N∗ ). If v : H∗ −→ K∗ is a surjection, then A∗ (G, H∗ ) ⊆ A∗ (G, K∗ ). If N∗ H∗ K∗ is a split exact sequence in G − SCF, then: A∗ (G, H∗ ) = A∗ (G, K∗ ) ∩ A∗ (G, N∗ ). 17 Proof. In the first case, we identify N = N1 to a subgroup of H = H1 . Let g ∈ Aj (G, H∗ ). We write: [g, Ni ] = [g, N ∩ Hi ] ⊆ [g, N ] ∩ [g, Hi ] ⊆ N ∩ Hi+j = Ni+j . Similarly, to show the second assertion, let us take g ∈ Aj (G, H∗ ). We write: [g, Ki ] = [g, v(Hi )] = ϕ([g, Hi ]) ⊆ v(Hi+j ) = Ki+j . The third assertion’s hypothesis comes down to require that H∗ decompose as a semidirect product K∗ o N∗ , the action of G on H∗ being factor-wise. We then get Gequivariant isomorphisms: Hi /Hi+j ∼ = Ki /Ki+j o Ni /Ni+j . An element g of G acts trivially on the left hand side if and only if it does on the right hand side. Whence the result. Remark 1.50. Let 1 N∗ u H∗ v K∗ 1 be a non-split short exact sequence in G − SCF, the first part of Proposition 1.49 gives: A∗ (G, H∗ ) ⊆ A∗ (G, K∗ ) ∩ A∗ (G, N∗ ). Nevertheless, equality is not true in general. Indeed, the sequences: 1 Ni /Ni+j u Hi /Hi+j v Ki /Ki+j 1 are exact, but g ∈ G can act trivially on the kernel and quotient without acting trivially on the middle term. For instance, H 1 (G) = Ext1G (Ztriv , Ztriv ) is non-trivial in general. We can get a little more about semi-direct products: H K be a split exact sequence in G − Grps. SupProposition 1.51. Let N pose that N is filtered by a strongly central series N∗ . Then: A∗ (G, N∗ ) ⊆ A∗ (G, A∗ (K, N∗ )) . Proof. Let us denote by K∗ the filtration A∗ (K, N∗ ) and by G∗ the filtration A∗ (G, N∗ ). A straightforward application of the 3-subgroup lemma (Lemma 1.4) in (N o K) o G provides the inclusion: [[Gα , Kβ ], Nγ ] ⊆ Nα+β+γ . Thus [Gα , Kβ ] ⊆ Aα+β (K, N∗ ) = Kα+β , which means that Gα ⊆ Aα (G, K∗ ). Corollary 1.52. Let G be a group acting on a filtered algebra A∗ by automorphisms of filtered algebras. Then: A∗ (G, A∗ ) ⊆ A∗ (G, A× ∗ ). Proof. Apply Proposition 1.51 to the G-equivariant split exact sequence: A A o A× A× and the given filtration on A. Here is an interesting case when the filtrations of Corollary 1.52 are equal: 18 Proposition 1.53. Let G be a group, and k a commutative ring. Then:
A∗ Aut(G), D∗k G = A∗ (Aut(G), Ik∗ G) . Proof. The algebra Z G is filtered by Ik∗ G, the powers of its augmentation ideal. The group Aut(G) acts on Z G by automorphisms of filtered algebras. As A∗ (G, Ik∗ G) = D∗k G, Proposition 1.51, applied to the action of Aut(G) on kG o G, gives an inclusion: A∗ (Aut(G), Ik∗ G) ⊆ A∗ (Aut(G), D∗k G). To show that it is in fact an equality, take ϕ ∈ A∗ (Aut(G), D∗k G). Then: ∀g ∈ G = D1 G, [ϕ, g] = (ϕ(g)g −1 − 1)g ∈ (Dj+1 − 1)G ⊆ I j+1 . We then show that [ϕ, I i ] ⊆ I i+j by induction on i > 1, using the formula: [ϕ, uv] = [ϕ, u]ϕ(v) + u[ϕ, v] (= (ϕ(u) − u) · ϕ(v) + u · (ϕ(v) − v)) . Let us remark that, in the language of paragraph 2.3, the last formula states that [ϕ, −] is a (1 , ϕ)-derivation, so we have in fact used lemma 2.18. 2 Traces and stable surjectivity 2.1 Free differential calculus We recall some basic concepts of free differential calculus. A detailed account can be found in [Fox53]. Definition 2.1. Let G be a group, and M a kG-module. A derivation from G to M is a map ∂ : G −→ M such that: ∀g, h ∈ G, ∂(gh) = ∂g + g · ∂h. It can be extended to a linear map ∂ : kG −→ M , which verifies: ∀u, v ∈ kG, ∂(uv) = ∂(u)ε(v) + u · ∂(v). We denote by Der(G, M ) or Der(kG, M ) the space of derivations from G to M . We will often write Der(kG) for Der(kG, kG). Remark 2.2. Let G = FS be the free group over a set S. As derivations identify with sections of M o G  G, we get: Der(kFS , M ) ∼ = M S for any kG-module M . Definition 2.3. Let S = (xi )i be a chosen basis of a free group F . The following requirement defines a derivation of kF : ( 1 if t = i, ∂ : xt 7−→ ∂xi 0 else. Let us give a first version of the chainrule: Proposition 2.4. Let λ : FY −→ G a group morphism, where FY is the free group on a set Y = {yj }. Then, for u in kY and ∂ in Der(kG): X  ∂u  ∂(λu) = λ ∂(λyj ). ∂y j j 19 Remark 2.5. The sums involved here are finite, because only a finite number of letters appear in a given element u. Proof of proposition 2.4. One can check that each member of this equality defines a derivation from kFY to kG, where kFY acts on kG by y · g = λ(y)g. As these formulas give the same result when evaluated at elements of the basis, the corresponding derivations are equal. If G = FX = hxi i is free too, we can apply proposition 2.4 with ∂ =   ∂u ∂(λyj ) ∂(λu) X = λ . ∂xi ∂y ∂x j i j ∂ ∂xi to get: For X = Y and λ = 1FX , this gives a change-of-base formula similar to the usual one. The following definition keeps on with our analogy with classical differential calculus: Definition 2.6. Let FY = hyj i and FX = hxi i be free groups as above. Let f be a morphism from FY to FX . We define its Jacobian matrix (with respect to the chosen basis) by:   ∂f (yj ) D(f ) := ∈ MY X (kX). ∂xi ji Remark 2.7. The morphism f is determined by Df . Indeed, proposition 2.4, applied with λ = 1, ∂ : v 7−→ v − ε(v) and u = f (xi ), gives: f (xi ) − 1 = X ∂f (xi ) ∂xj j (xj − 1). g f Let FZ = hzk i, FY = hyj i and FX = hxi i be free groups, and let FZ −→ FY −→ FX be morphisms between them. We can use proposition 2.4 to get a chainrule for Jacobian matrices:   ∂g(zk ) ∂f (yj ) ∂f (g(zk )) X = f . ∂xi ∂yj ∂xi j This can be restated as: g f Corollary 2.8. Let FZ −→ FY −→ FX be morphisms between free groups with fixed basis, as above. Then: D(f g) = f (Dg)D(f ). Remark 2.9. The reader may have noticed that this formula seems to come ”in the wrong way”. This can be explained as follows: a morphism f : FY −→ FX is in fact a r-tuple of monomials f (yj ) = fj (xi ) ∈ kFX , and should as such be considered as a ”polynomial function from FX to FY ”, whose coordinates would be the fj s. From this point of view, f g would be a ”polynomial function from FX to FZ ”, whose coordinates would be given by f g(zk ) = gk (fj (xi )): it looks more like ”g ◦ f ” ! This seems to be completely analogous to the classical setting of algebraic geometry. To get more accurate statements, one would have to interpret kFX as an algebra of functions over a geometric object associated to FX . Which should look like a one-point object with some local structure (as Df determines f ). 20 2.2 Derivations and strongly central filtrations We introduce the notion of a derivation from a group G to a group H on which it acts. Our aim here is to describe a general framework which will be useful to study Jacobian matrices and their interactions with Lie brackets: the chainrule formula 2.8 tells us that D is a derivation. We will also get back to Lazard’s theorem 1.37 in this framework. Definition 2.10. Let H be a group, on which another group G acts. A map ∂ : G −→ H is a derivation if: ∀ x, y ∈ G, ∂(xy) = ∂x · x∂y. Remark 2.11. If H = M is an abelian group, i.e. a representation of G, then we recover the usual definition of a derivation from G to M (see Definition 2.1). To give a derivation ∂ : G −→ H is exactly the same as giving a section σ = (∂, 1G ) of the canonical projection: σ H oG p G. Keeping this in mind, the following lemma follows immediately: Lemma 2.12. Let ∂ be a derivation from G to H. Then ∂ −1 sends G-stable subgroups of H on subgroups of G. Let H∗ be strongly central filtration on a subgroup H1 of H. Let G∗ be a strongly central filtration on a subgroup G1 of G, which acts on H∗ through the given action of G on H (see Proposition 1.20). A derivation ∂ from G to H being given, we can use the morphism σ = (∂, 1G ) to pull back the filtration H∗ o G∗ . We thus get a strongly central filtration on G1 ∩ ∂ −1 (H1 ): σ −1 (Hi o Gi ) = Gi ∩ ∂ −1 (Hi ). Remark 2.13. For instance, if H∗ is given, we can let G∗ be A∗ (G, H∗ ), the maximal filtration acting on H∗ , as described in Remark 1.23. The above construction then gives a strongly central filtration on A1 ∩ ∂ −1 (H1 ). This subgroup is all of G if and only if:   G stabilises H∗ , G acts trivially on L(H∗ ), (2.13.1)  ∂(G) ⊆ H1 . Under these conditions, A∗ ∩ ∂ −1 (H∗ ) is a strongly central series on G. In particular, it then contains Γ∗ (G). Keeping the above notations, the morphism σ = (∂, 1G ) induces a Lie ring morphism (which is injective by definition of the filtration on the domain): σ̄ : L(G∗ ∩ ∂ −1 (H∗ )) ,−→ L(H∗ ) o L(G∗ ). This ensures that ∂ induces a well-defined linear map ∂¯ between the Lie algebras. ¯ 1̄ ) : x 7−→ ∂x ¯ + x preserves Lie brackets, hence: Moreover, the map σ̄ = (∂, ¯ ¯ y] + [x, ∂y] ¯ + [∂x, ¯ ∂y]. ¯ ∂([x, y]) = [∂x, If L(H∗ ) is an abelian Lie algebra, then the last term is zero, and ∂¯ is a Lie derivation. This happens in particular when H is an abelian group. 21 Back to the proof of Lazard’s theorem 1.37. Take the filtered (abelian) group (A, +) as H, and the group A× as G acting by left multiplication ρ. We already know that: × Aj (A× , A∗ ) = A× ∗ = A ∩ (1 + Aj ). Let ∂ be the derivation from A× to A defined by g 7−→ g −1. Obviously, ∂ −1 (Aj ) = A× j , × × −1 which is exactly Aj (A , A∗ ). It is then equal to Aj (A , A∗ ) ∩ ∂ (Aj ). The Lie ring L(A∗ ) is abelian (because A is an abelian group). Hence, the induced map ∂¯ is a derivation (with respect to the canonical action of L(A× ∗ ) on L(A∗ )): ∂¯ : L(A× ∗ ) −→ L(A∗ ). Let us remark that the Lie algebra L(A∗ ) is quite different from gr(A∗ ): the associative structure of A has been completely forgotten. Nevertheless, some part of this structures is encoded by the action of L(A× ∗ ), which is inherited from left multiplication. The ¯ map ∂ is a derivation with respect to this action, that is: ¯ ¯ y] + [x, ∂y]. ¯ ∂([x, y]) = [∂x, These brackets are described through the action of L(A× ∗ ) on L(A∗ ), as induced by commutators in A o A× : ∀g ∈ A× 1 , ∀x ∈ A, [g, x] = gx − x. As a consequence: ¯ ∂([x, y]) = −(y(x − 1) − (x − 1)) + (x(y − 1) − (y − 1)) = xy − yx, so ∂¯ is in fact a Lie morphism to gr(A∗ ). 2.3 Algebras, actions and derivations We now turn to studying derivations of algebras. In particular, we get a precise link between free differential calculus and differential calculus in the tensor algebra (see Proposition 2.25). We will use this in Paragraph 2.4.2 to get an explicit description of the trace map. Let Alg− be the category of associative non-unitary algebras over a fixed commutative ring k. This category is pointed (by 0) and protomodular. We can define actions there, as in paragraph 1.3. Actions in Alg− turn out to be representable. Precisely, for any algebra I, let Endr (I) (resp. Endl (I)) be the algebra of right (resp. left) I-linear endomorphisms of I, i.e. k-linear maps u from I to I satisfying: ∀x, y ∈ I, u(xy) = u(x)y (resp. u(xy) = xu(y)). Define Endr,l (I) as the kernel of :  Endr (I) × Endl (I) −→ Endk (I) α: (u, v) 7−→ u ◦ v − v ◦ u. Proposition 2.14. An action A  I in Alg− can be represented by a (unique) morphism: (λ,ρ) A −→ Endr,l (I). 22 Proof. If an action I B A is given, λ(a) and ρ(a) are obtained from left and right multiplications by a in B. Conversely, a morphism (λ, ρ) as above can be used to define an (associative) algebra structure on I × A defining an action of A on I. Let us remark that if I 2 = 0 (that is, I is endowed with a trivial algebra structure), then an action of A on I is just a A-bimodule structure. Remark 2.15. The same construction works in the category of (non-unitary) filtered algebras f Alg− , where we also get a representation of actions. The algebras Endr (I) and Endl (I) are then filtered by the usual requirement: a morphism u is of degree at least j if u(Ii ) ⊆ Ii+j for all i. The same requirement will be used to define a filtration on any module of morphisms between filtered modules or algebras. Definition 2.16. Let A act on I as above. A derivation from A to I is a k-linear map ∂ : A −→ I satisfying: ∂(ab) = ∂a · b + a · ∂b. The k-module of derivations from A to I is denoted by Der(A, I). Remark 2.17. The relation defining derivations depends only on the A-bimodule structure on I. We are thus led to consider I o , the algebra obtained by taking the same underlying k-module as I, endowed with the trivial product. The action of A on I induces an action of A on I o in the obvious manner, and a derivation from A to I is then the same as a section of the projection: Io o A  A in the category of algebras. When we work in the category of filtered algebras, Der(A, I) is a filtered module, a derivation ∂ being of degree at least j if ∂(Ai ) ⊆ Ii+j for all i. If A is filtered by its powers Ai , we just have to check this in degree one: Lemma 2.18. Let A be an algebra, filtered by its powers Ai := Ai , acting on a filtered algebra I∗ . Let ∂ ∈ Der(A, I). Then ∂ is of degree at least j if and only if: ∂(A) ⊆ Ij+1 . Proof. An action of A∗ on I∗ is given by a left and a right multiplication which are filtered, meaning that Ai Ij ⊆ Ii+j and Ij Ai ⊆ Ii+j . Use the formula: X ∂(a1 · · · ai ) = a1 · · · ak−1 ∂(ak )ak+1 · · · ai k to get the desired result. We can get examples of actions from algebras acting on themselves. Precisely, the adjoint action of A on itself is just the obvious A-bimodule structure on A. Derivations are then the usual ones. Given an action of A on I represented by (ρ, λ), we can twist it by choosing endomorphisms ϕ and ψ of A and letting A act on I through (ρ ◦ ϕ, λ ◦ ψ). This means that we let a ∈ A act on I by ϕ(a) · − on the left, and by − · ψ(a) on the right. We give a name to derivations from A to the the twisted A-bimodule I. 23 Definition 2.19. Let ϕ and ψ be endomorphisms of A. A (ϕ, ψ)-derivation is a linear map ∂ : A −→ I satisfying: ∂(ab) = ∂a · ψb + ϕa · ∂b. We denote by Der(ϕ,ψ) (A, I) the k-module of such derivations. Example 2.20. Let A be a group algebra kG and M a kG-module. We can make M into a bimodule by making kG act trivially on the right (that is, through ε). Then, Der(kG, M ) is exactly the usual module of derivations (see Definition 2.1). If M = kG, it is already a bimodule, but the above structure can be obtained through twisting the right action by ηε : g 7→ ε(g) · 1. Then Der(kG) (defined in Definition 2.1) is exactly Der(id,ηε) (kG, kG). We can apply Lemma 2.18 to A = I = IG, to get: Corollary 2.21. Let ∂ be a derivation of Z G such that ∂(Z G) ⊆ (IG)l+1 (which is always true for l = −1). For all integer k, we have: ∂((IG)k ) ⊂ (IG)k+l . Remark 2.22. Let us stress that the proof given here is fairly direct. In fact, it gets even shorter in this case, the result following from: ∀v ∈ I, ∂(uv) = u · ∂v. Remark 2.23. The classical inclusion of Γk − 1 into IGk can be shown by a direct induction, or follows from Lazard’s theorem (as D∗ (G) is a strongly central series, it contains Γ∗ G). Under the hypothesis of Proposition 2.21, it implies ∂(Γk ) ⊆ (IG)k+l , using that ∂(1) = 0. Remark 2.24. Some sets of derivations obtained from actions of A on itself can have more structure than just a module structure. Precisely, if we twist the adjoint action of A by some (ϕ, ψ) as above, and if we add the requirement that ϕ and ψ are idempotents, then the set of (ϕ, ψ)-derivations from A to A commuting to ϕ and ψ is a sub-Lie algebra of Endk (A): [∂, ∂ 0 ](ab) = ∂∂ 0 a · ψ 2 b + ϕ∂ 0 a · ∂ψb + ∂ϕa · ψ∂ 0 b + ϕ2 a · ∂∂ 0 b − ∂ 0 ∂a · ψ 2 b − ϕ∂a · ∂ 0 ψb − ∂ 0 ϕa · ψ∂b − ϕ2 a · ∂ 0 ∂b = [∂, ∂ 0 ](a) · ψ(b) + ϕ(a) · [∂, ∂ 0 ](b). In particular, if (ϕ, ψ) = (1 , 1 ), we get that Der(A) is a sub-Lie algebra of Endk (A). Another example is given by A = kG and (ϕ, ψ) = (1 , ηε). But more is true is this last case. Let kG be filtered by the powers of the augmentation ideal. If ∂, ∂ 0 ∈ Der(kG) are such that ∂ 0 has degree at least 0, then ∂ ◦ ∂ 0 ∈ Der(kG), because ε(∂ 0 v) = 0 for any v. Let A∗  I∗ be an action of filtered algebras. Since the functor gr : f Alg −→ grAlg from filtered algebras to graded ones is exact (the same proof as that of Proposition 1.24 works), this action is sent to an action gr(A∗ )  gr(I∗ ) of graded algebras. Moreover, gr also commutes with (−)o (the definition of (−)o being extended to graded algebras in the obvious way), so we get a morphism: gr(Der(A∗ , I∗ )) ,→ Der∗ (gr(A∗ ), gr(I∗ )), 24 where the target is the graded module of graded derivations. This morphism is obviously injective. It is in fact a restriction of the natural injection: gr(Homk (M∗ , N∗ )) ,→ Hom∗ (gr(M∗ ), gr(N∗ )) between bifunctors on graded modules. As such, it preserves all algebraic structure inherited from the additive bifunctor structure (see Remark 2.24). When A∗ is kG, filtered by the powers of IG, acting on itself, we thus get a morphism preserving the structure induced by composition: gr(Der(kG)) ,→ Der∗ (gr(kG)). Proposition 2.25. If G is a free group, and M∗ is a filtered kG-module (considered as a bimodule with trivial right action), the canonical map: gr(Der(kG, M∗ )) ,→ Der∗ (gr(kG), gr(M∗ )) is an isomorphism. Here, by derivations, we mean (id, ε)-ones. Proof. Let S be a free set of generators for G. Then V = Gab is free abelian on S, and gr(kG) ∼ = T V is the tensor algebra. Identifying derivations with sections as above (see Remark 2.17), we see that a derivation is completely determined by the choice of its values on S: Der∗ (T V, N∗ ) = F∗ (S, N∗ ), for any graded T V -bimodule N∗ , where F∗ (S, N∗ ) is the set of graded maps from S (concentrated in degree 0) to N∗ . The same is true for the other side. Indeed, a derivation from kG to M is a section of the projection M o G  G, so is determined by a map S → M : Der(kG, M∗ ) = Der(G, M∗ ) ∼ = M∗S . The second member is the set of maps from S to M , with the filtration inherited from the one on M . The desired isomorphism is then exactly: gr(M∗S ) ∼ = F∗ (S, gr(M∗ )). Remark 2.26. If M is a G-module, we can endow it with the universal kG-filtration (IG)∗ · M . Remark 2.27. The isomorphism gr(Der(kG)) ∼ = Der∗ (T V ) thus obtained preserves the algebraic structure obtained from the composition of derivations. 2.4 Traces In [Bar13], Bartholdi defines the trace of an automorphism ϕ of Fn by: Tr(ϕ) := Tr(Dϕ − 1 ) ∈ Z Fn , where Dϕ denotes ϕ’s Jacobian matrix. We will show that Tr induces a well-defined map between the graded Lie algebras, which we still call Tr: Tr : L(A∗ (Fn )) −→ gr(Z Fn ) ∼ = T V. The aim of this paragraph is to show that this map is indeed well-defined, to investigate its behaviour with respect to Lie structures, and to get Morita’s algebraic description [Mor93], used by Satoh in [Sat12]. 25 2.4.1 The induced map between Lie algebras Let ϕ ∈ Ak (Fn ). By definition, ϕi := x−1 i ϕ(xi ) ∈ Γk+1 . The Jacobian matrix of ϕ can be described explicitly: ∂(xi ϕi ) ∂xi ∂(ϕi ) = . +xi ∂xj ∂xj ∂xj |{z} (Dϕ)ij = δij Hence: Dϕ − 1 =  ∂ϕi xi ∂xj  . ij Using Remark 2.23, we see that this matrix is in fact in Mn (I k ) (to shorten notations, we write I for IFn in the sequel). Moreover, xi acts trivially on I k /I k+1 . We thus get an explicit formula for the trace map: Tr(Dϕ − 1 ) = X i xi ∂ϕi X ∂ϕi ≡ ∂xi ∂xi i (mod I k+1 ). Let G be any group. We can apply the construction of Paragraph 1.7 to A = kG, filtered by the powers of the augmentation ideal. This gives a strongly central filtration GLn (I ∗ G) on GLn (IG), which comes with an embedding of Lie algebras: L(GLn (I ∗ G)) ,→ gr(Mn (kG)) ∼ = Mn (gr(kG)). The next proposition replaces a formula from [Bar13, section 6]: Proposition 2.28. The Jacobian matrix D induces a morphism between graded modules: D : L(A∗ (Fn )) −→ L(GLn (I ∗ Fn )), satisfying: D([f, g]) = [g, Df ] + [Dg, f ] + [Dg, Df ] . In order to address the issue raised in Remark 2.9, let us introduce some notations before proving the proposition. If G is any group, we denote by Gop the opposite group, where multiplication is defined by: g ·op h = hg. Let G∗ be a strongly central filtration op on G. Then Gop and one easily checks that: ∗ is such a filtration on G op L(Gop ∗ ) = L(G∗ ) , where the bracket in L(G∗ )op is L(G∗ )’s additive inverse: [x, y]op = [y, x]. Proof of Proposition 2.28. Corollary 2.8 states exactly that D is a derivation from Aut(Fn ) to GLn (I)op ⊂ Mn (Z Fn )op , where Mn (Z Fn ) is endowed with the obvious Aut(Fn )-action. We thus can apply the results from 2.2 with ∂ = D, G = Aut(Fn ), H = GLn (I)op , and H∗ = GLn (I ∗ )op . The strongly central filtration A∗ (G, H∗ ) is in fact the Andreadakis filtration A∗ = A∗ (Fn ) on A1 = IAn ⊂ Aut(Fn ). Indeed, there is a series of inclusions: A∗ (G, D∗ Fn ) ⊇ A∗ (G, GLn (I ∗ )) ⊇ A∗ (G, Mn (Z Fn )) = A∗ (G, Z Fn ) . 26 The first one comes from 1.49 applied to the injection of D∗ Fn into GLn (I ∗ ) defined by w 7−→ w · 1 . The second one is a particular case of 1.52. The last equality comes from the fact that G acts component-wise on matrices: [g, (mij )] = g · (mij ) − (mij ) = ([g, mij ]). According to proposition 1.53, these inclusions are in fact equalities. Moreover, we have seen at the beginning of the present paragraph that D sends A∗ to GLn (I ∗ )op . The filtration A∗ ∩ ∂ −1 (H∗ ) is thus only A∗ . The work already done in 2.2 allows us to get the desired result. The map given by Proposition 2.28 can be composed with the morphism: (−)−1 L(GLn (I ∗ )) −→ gr(Mn (Z Fn )) ∼ = Mn (gr(Z Fn )) = Mn (T V ). Thus, for ϕ an element of Ak /Ak+1 , Dϕ − 1 is well-defined modulo Mn (I k+1 ). Composing with the usual trace, we get the announced well-defined linear map induced by ϕ 7−→ Tr(Dϕ − 1 ): Tr : L(A) −→ T V. Remark 2.29. That the map D− 1 (hence Tr) induces a well-defined map between the Lie algebras can be seen through explicit calculation, but the behaviour with respect to the Lie bracket is much less obvious from this point of view. Indeed, let ϕ ∈ Ak . If ϕ = ψχ, with χ ∈ Ak+1 , then: ψ(xi ) = ϕ(xi χi ) = xi ϕi ϕ(χi ). As χi stands inside Γk+2 , its image by ϕ does too. Thus: ∂ϕi ∂ϕ(χi ) ∂ϕi ∂ψi = + ϕi ≡ ∂xj ∂xj ∂x ∂xj | {zj } (mod I k+1 ). ∈I k+1 2.4.2 Introducing the contraction map Consider the evaluation map: ev : Der(1 ,ε) (T V ) ⊗ T V −→ T V. Using the universal property of T V , as in the proof of 2.25, we get a linear isomorphism: Der(1 ,ε) (T V ) ∼ = Hom(V, T V ) ∼ = V ∗ ⊗ T V. The evaluation map then is:  ∗ V ⊗ TV ⊗ TV ω⊗u⊗v −→ T V 7−→ Φ(ω ⊗ v)u, where Φ is the contraction map:  ∗ V ⊗ V ⊗k+1 −→ V ⊗k Φ: α ⊗ Xi1 · · · Xik+1 7−→ Xi1 · · · Xik α(Xik+1 ), extended by zero on k·1. This follows from the fact that any (1 , ε)-derivation ∂ verifies: ∂(uv) = u · ∂v, when the degree of v is at least 1 (that is, when ε(v) = 0), and ∂(1) = 0. We sum this up in the following: 27 Proposition 2.30. Let ∂ ∈ Der(1 ,ε) (T V ). Then: ∂ = Φ(∂|V ⊗ −). Let us consider the derivation ∂x∂ i of Z Fn . It induces a (1 , ε)-derivation of degree −1 of T V , denoted by ∂i (any derivation of kG is of degree at least −1, by Corollary 2.21). As ∂i |V = Xi∗ , we get the following: Corollary 2.31. The (1 , ε)-derivation of T V induced by ∂x∂ i ∈ Der(Z Fn ) is represented as: ∂ = ∂i = Φ(Xi∗ ⊗ −) : T V −→ T V. ∂xi We can use these results to interpret the trace map in a way more suited to explicit calculations: Proposition 2.32. The trace map can be described as: Tr = Φ ◦ ι ◦ τ, where τ is the Johnson morphism (see Definition 1.26), ι denotes the inclusion of Derk (LV ) ∼ = V ∗ ⊗ Lk+1 V into Derk (T V ) ∼ = V ∗ ⊗ V ⊗k+1 , and Φ is the contraction map. Proof. Let ϕ ∈ Ak . Then τ (ϕ) is defined by: ∼ τ (ϕ)(Xi ) = x−1 i ϕ(xi ) = ϕi ∈ Γk+1 /Γk+2 = Lk+1 V. We have seen at the beginning of Paragraph 2.4.1 that the trace map is given by: Tr(Dϕ − 1 ) = X ∂ϕi i ∂xi . The formula of the proposition is then equivalent to: ! X X ∂ϕi ∗ Φ . Xi ⊗ ϕi = ∂xi i i To get this formula, we evaluate the equality given by Corollary 2.31 to the elements ϕi − 1 (keeping in mind that the inclusion of Lk+1 V into Tk+1 V is given by w 7→ w − 1). 2.5 2.5.1 Stable surjectivity Vanishing of the trace map Here, we show that the trace map takes values in brackets inside T V . This result can also be found in [MS17, Prop. 5.3], where rational methods are used to get it. Proposition 2.33 ([BLGM90, Th. 2.1], quoted in [Bar13, Th. 6.2]). Let k > 2, and let J ∈ GLm (IZk Fn ). Denote by V the abelianization V = Fnab ∼ = Zn . Then: Tr (J − 1 ) ∈ [T V, T V ]k ⊂ V ⊗k ∼ = I k /I k+1 . This result relies on the following criterion: 28 Proposition 2.34 ([BLGM90, prop. 2.2]). Let f (X1 , ..., Xn ) ∈ V ⊗k . Let C ⊂ Mk (Z) be the sub-Z-module generated by the ei,i+1 . Suppose: ∀Ci ∈ C, Tr(f (C1 , ..., Cn )) = 0. Then f ∈ [T V, T V ]k . The proof can be found in [BLGM90]. The reader is also referred to the proof of Proposition 3.16, which is the same proof, adapted to the case of positive carateristic. Proof of Proposition 2.33. The main idea is to use evaluations into commutative algebras to be able to use Proposition 1.41, and to then get back to the non-commutative setting by using the above criterion. Let 1 + tAi ∈ GLk (tk[t]). There is an evaluation morphism xi 7→ 1 + tAi from Fn to GLk (t Z[t]), extending to a morphism from kFn to Mk (tk[t]) sending I ∗ to t∗ Mk (k[t]). Taking congruence groups, we get an evaluation morphism: ev1 +tAi : GLm (I ∗ Fn ) −→ GLm (t∗ Mk (k[t])) = GLmk (t∗ k[t]). There is a commutative diagram: L(GLm (I ∗ Fn )) (−)−1 Mm (T V ) ev1 +tAi L(GLmk (t∗ k[t])) (−)−1 Mmk (k[t]) TV evtAi evtAi det L(k[t]× ∗) = 0 Tr T rm Mm (k[t]) Tr Tr (−)−1 k[t] Here, we identify gr(I ∗ Fn ) with T V by xi 7→ 1 + Xi . We also identify gr(t∗ k[t]) with k[t]. The evaluation xi 7→ 1 + tAi thus induces Xi 7→ tAi between the associated graded. The bottom-left square is just the one in Proposition 1.41. The map T rm is the usual trace when the base algebra is Mm (k[t]). Let k = Z and f = Tr(J − 1 ) ∈ V ⊗k . The commutativity of the above diagram gives: 0 = Tr(f (tAi )) = tk Tr(f (Ai )). As a consequence, Tr(f (Ai )) = 0, for any 1 + tAi ∈ GLk (t Z[t]). We can then evaluate this at t = 0 to get: Tr(f (πAi )) = 0. This evaluation π is the map:
π : L1 (GLk (t∗ Z[t]) ,→ Mk t Z[t]/t2 Z[t] ∼ = Mk (Z). Using Proposition 1.42 and Remark 1.43, we see that its image is exactly sln (Z), so the conclusion follows from the above criterion (2.34). Because of Proposition 2.33, we will consider the trace map as taking values in the abelianization T V ab = T V /[T V, T V ]. As [T V, T V ]k is generated by the elements: [Xi1 ⊗ · · · ⊗ Xip , Xip+1 ⊗ · · · ⊗ Xik ] = Xi1 ⊗ · ⊗ Xik − tp · Xi1 ⊗ · ⊗ Xik , where t = 1̄ ∈ Z /k, the module T V ab is the module of cyclic powers C∗ V : (T V ab )k = Ck V := V ⊗k /(Z /k). The conclusion of Proposition 2.33 becomes, in this context: Tr (J − 1 ) = 0 ∈ C∗ V. 29 2.5.2 Linear algebra Consider the Johnson morphism τ 0 : L (ΓIAn ) −→ Der∗ (L(Fn )) (cf. 1.5). The morphism τ10 is an isomorphism (see Proposition 1.34). Moreover, L (ΓIAn ) is generated in degree 1 (cf. 1.9). As a consequence, the image of τ 0 is exactly the sub-Lie ring generated in degree 1 inside Der (L(Fn )). As L(Fn ) is the free Lie ring LV , the study of coker(τ 0 ) is solely a problem of linear algebra. Recall from Proposition 2.32 that the trace map can be seen as the composite of the Johnson morphism τ : Lk (A∗ (Fn )) → Derk (LV ) ∼ = V ∗ ⊗ Lk+1 V with: ι Φ π TrM : V ∗ ⊗ Lk+1 V −→ V ∗ ⊗ V ⊗k+1 −→ V ⊗k −→ Ck V := V ⊗k /(Z /(k)), where ι and π denote the canonical maps. All these morphisms are obviously GLn (Z)equivariant (with respect to the canonical actions). Notation 2.35. Let I denote the image of τ 0 , which is the sub-Lie ring generated in degree 1 inside Der(LV ). The following proposition can be seen as a consequence of Proposition 2.33. Precisely, I = Im(τ 0 ) ⊆ Im(τ ), and Proposition 2.33 implies that TrM ◦τ = Tr vanishes. Proposition 2.36. For every k > 2, TrM (Ik ) = {0}. 2.5.3 Stable cokernel of τ 0 and stable surjectivity Let k > 2. Using Proposition 2.36, we get a commutative diagram with exact rows: Ik V ∗ ⊗ Lk+1 V φ coker(τk0 ) Φ V ⊗k [T V, T V ]k Φ π Ck V. In [Sat12], Satoh shows:   For n > k + 1, Φ is surjective, (lemma 3.2) For n > k + 2, φ is surjective, (prop. 3.2)  For n > k + 2, ker Φ ⊆ I. (Prop. 3.3) His Theorem 3.1 is still true over Z: Proposition 2.37. Let k > 2 and n > k + 2 be integers. Then Φ is a GLn (Z)equivariant isomorphism: coker(τk0 ) ∼ = Ck V. Proof. Let us denote by K (resp. L) the kernel of Φ (resp. its cokernel). There is a commutative diagram in GLn (Z) − ModZ : ker Φ Ik φ 0 V ∗ ⊗ Lk+1 V coker(τ 0 ) Φ [T V, T V ]k V ⊗k 0 0 30 K Φ π Ck V L. The snake lemma ensures that K and L are zero: Φ is an isomorphism. We can now state our main result: Theorem 2.38. Let k + 2 6 n. Then the canonical morphism Lk (IAn ) −→ Lk (A∗ (Fn )) is surjective, and τ induces an isomorphism: Lk (A∗ (Fn )) ∼ = Ik . Remark 2.39. Basis being chosen, there is an injection of Fn in Fn+1 ∼ = Fn ∗ Z. An automorphism ϕ of Fn can be extended to an automorphism ϕ∗ 1 of Fn+1 . This induces injections IAn ,→ IAn+1 which in turn induce morphisms L(A∗ (Fn )) → L(A∗ (Fn+1 )). Taking the colimit over n, we can define a Lie ring Lst (A∗ ). In the same way, we can ab define injections from Der(L(Fnab )) into Der(L(Fn+1 )) and take the colimit Ist of the sub-algebras generated in degree one. With this point of view, the isomorphisms of Theorem 2.38 give an isomorphism between graded Lie algebras: τ st : Lst (A∗ ) ∼ = Ist , meaning exactly that Lst (A) is generated in degree one. In fact, all the constructions appearing here are functors on the category denoted by S(Z) in [Dja16b, section 7], where it is shown (using methods similar to the ones of [CEFN14]) that these functors are finitely supported. This implies the equivalence between τkst being an isomorphism and τk being one for n big enough. Proof of theorem 2.38. Consider the commutative diagram: i∗ L(IAn ) L(A∗ ) τ τ0 Der(LV ). The image I of τ 0 is the sub-Lie ring generated in degree one inside Der(LV ). Using the results quoted in Paragraph 2.5.3, we see that in degrees k 6 n − 2, it also is the kernel of the trace map. Proposition (2.33) tells us exactly that Tr ◦τ = 0, so that Im τk ⊆ ker Trk = Im τk0 when k > n − 2. As a consequence, Im τ = Im τ 0 . As τ is injective (1.32), it is an isomorphism onto its image, hence the result. 2.6 Automorphisms of free nilpotent groups Automorphisms of nilpotent groups are easy to deal with, due to the following classical fact: Lemma 2.40. Let G be a finite-type nilpotent group. An endomorphism ϕ ∈ End(G) is an automorphism if and only if the induced morphism ϕab ∈ End(Gab ) is. Proof. If ϕ is an automorphism then ϕab has to be, with (ϕab )−1 = (ϕ−1 )ab . Conversely, suppose that ϕab is an automorphism. This means that L1 (ϕ) is. Since L(G) is generated in degree one, L(ϕ) is surjective. But each Lk (G) is abelian of finite type, so each Lk (ϕ), being surjective, has to be bijective (it is obviously the case on the torsion 31 part, which is finite, and it also is on the free abelian part, for reasons of rank). The lemma then follows by induction from the five-lemma applied to: Lk (G) Lk (ϕ) Lk (G) G/Γk+1 G G/Γk G ϕ̄ G/Γk+1 G ϕ̄ G/Γk G. This induction process stops since there is a c such that G = G/Γc+1 G. b := lim (G/Γk G) . Definition 2.41. The pro-nilpotent completion of a group G is: G ←− b is canonically filtered by the Γ̄j G := lim (Γj G/Γk G). This filtraThe completion G ←− tion is its closed lower central series, defined as the closure of the lower central series. b An endomorphism of G b It is minimal amongst closed strongly central filtrations on G. is continuous if and only if it preserve this filtration. b is an automorLemma 2.42. Let G be a group. A continuous endomorphism ϕ of G ab ab phism if and only if the induced morphism ϕ ∈ End(G ) is. Proof. Such an endomorphism is an automorphism if and only if the associated morphism between projective system is. These are the induced endomorphisms of the G/Γk G, which are nilpotent groups. This condition amounts to ϕ inducing an isomorphism on Ĝ/Γ̄2 = Gab , by Lemma 2.40. b In fact, we can readily deduce the following explicit description of the group AutC 0 (G) b of continuous automorphisms of G: Proposition 2.43. The canonical map is an isomorphism: b ∼ AutC 0 (G) (Aut(G/Γk G)) . = lim ←− Let G = Fn be a free group of finite type. It is residually nilpotent, so it embeds into its completion Fbn . A continuous endomorphism of Fbn is uniquely determined by its (arbitrary) values on the topological generators xi . Let us denote by Fn,c the free c-nilpotent group Fn /Γc+1 Fn . Using Lemmas 2.40 and 2.42, we can show two surjectivity results for automorphisms of free nilpotent groups: Proposition 2.44. The canonical morphisms Aut(Fn,c ) → Aut(Fn,c−1 ) are surjective. Proof. Let (xi ) be a free basis of Fn . Let ϕ ∈ Aut(Fn /Γc (Fn )). Lift the elements ϕ(x̄i ) to elements ti of Fn /Γc+1 (Fn ). Then define the endomorphism ϕ̃ of Fn /Γc+1 (Fn ) by x̄i 7→ ti . Since ϕ and ϕ̃ induce the same endomorphism of Fnab , Lemma 2.40 implies that ϕ̃ is an isomorphism. Proposition 2.45 ([Bar13, th. 5.1]). The Johnson morphism associated to the universal action on Γ̄∗ (Fn ) is an isomorphism: τ : L(A∗ (Γ̄∗ Fn )) ∼ = Der(LV ). Proof. Because of Lemma 1.28, we know that τ is injective. We need to show that it is surjective. Let us first remark that L(Γ̄∗ (Fn )) = L(Γ∗ (Fn )) ∼ = LV . Let ∂ ∈ Derk (LV ). ∼ Lift each ∂(Xi ) ∈ Γk /Γk+1 = Γ̄k /Γ̄k+1 to an element ti ∈ Γ̄k . We can define a continuous endomorphism of Fbn by ϕ : xi 7→ ti xi . Then ϕ acts trivially on Fnab , so it is an isomorphism by Lemma 2.42. As a consequence, ϕ ∈ Ak (Γ̄∗ (Fn )) satisfies τ (ϕ̄) = ∂. This concludes the proof. 32 We can translate this last result into a statement about automorphisms of free nilpotent groups : because of Propositions 2.43 and 2.44, Aut(Fn,c ) is a quotient of AutC 0 (Fbn ). Moreover, the kernel of the canonical surjection is Ac (Γ̄∗ Fn ), by definition. Thus, this projection induces the c-truncations at the level of the associated graded objects : Corollary 2.46. The Johnson morphism associated to the universal action on Γ∗ (Fn,c ) is an isomorphism: τ : L(A∗ (Fn,c )) ∼ = Der(L6c V ). Consider the Johnson morphism τ 0 : L(Γ∗ (IAFn,c )) → Der(L6c V ). Using the identification of corollary 2.46, we see that τ 0 identifies with the Andreadakis morphism i∗ : L(Γ∗ (IAFn,c )) → L(A∗ (Fn,c )). Its image is the subalgebra generated in degree one inside Der(L6c V ), as was the case in paragraph 2.5.2. This subalgebra is exactly the truncation I<c , so is inside (and stably equal to) the kernel of the trace map. As a consequence, the Andreadakis equality never holds for free nilpotent groups. Moreover, in this context, our stable surjectivity result translates as : Corollary 2.47. The following sequence always is a complex, and is exact for n > c+1: i∗ Tr L(Γ∗ (IAFn,c )) −→ L(A∗ (Fn,c )) ∼ = Der(L6c V ) −→ C<c V −→ 0. Remark 2.48 (Non-tame automorphisms). The canonical morphism p : Aut(Fn ) → Aut(Fn,c ) is in general not surjective: some basis of the free nilpotent group do not lift to basis of the free group via Fn  Fn,c . Automorphisms of Fn,c induced by automorphisms of Fn are called tame. This was the original motivation of [BLGM90] for considering the trace map. In this regard, our stable surjectivity result could be re-stated as follows : in the stable range, the trace is the only obstruction for an automorphism to be tame. 3 The case of positive caracteristic 3.1 Dark’s theorem Let w be a word in a free group FS . Let w(r) be the word obtained from w by replacing each generator s ∈ S by some power srs . Dark’s theorem [Dar68] describes how to decompose w(r) as a product of commutators. This gives very useful universal formulas, that can then be evaluated in any group. The reader is referred to [Pas79, chap. IV, Th. 1.11] for a precise statement and a proof of the theorem. Here we recall two corollaries, obtained by taking w = [x, y] and w = xy in F{x,y} . The case w = [x, y] is [Pas79, IV, cor. 1.16]: Corollary 3.1. There exists a unique map θ : (N∗ )2 −→ F2 = hx, yi verifying: Y α β ∀α, β ∈ N, [xα , y β ] = θ(r, s)( r )( s ) . r,s>1 Each θ(r, s) is a product of {x±1 , y ±1 }-commutators such that x±1 appears at least r times and y ±1 at least s times in each factor. The case w = xy has been known for a long time (quoted in [Pas85, chap. 11, Th. 1.14], it already appears for instance in [Hal34]): 33 Corollary 3.2. There exists a unique map θ : N −→ F2 = hx, yi such that: Y α ∀α ∈ N, xα y α = θ(r)( r ) , r>0 Each θ(r) is a product of {x±1 , y ±1 }-commutators of length at least r. Remark 3.3. In both cases, uniqueness of the map θ is obvious: it can be defined by induction. 3.2 p-Restricted strongly central series Definition 3.4. Let p be a prime number. A strongly central series G∗ is said to be p-restricted if: ∀i, Gpi ⊆ Gip . Let G∗ be a p-restricted strongly central series. Using [Pas79, Th. III.1.7], we see that the morphism: L(G∗ ) −→ gr(aF∗ p (G∗ )) induced by g 7−→ g − 1 (see Theorem 1.37 and Proposition 1.40) is injective. We can identify L(G∗ ) with its image, which is stable by the p-th power operation in the F associative Fp -algebra gr(a∗ p (G∗ )), since (g − 1)p = g p − 1. From this we deduce that L(G∗ ) is a p-restricted Lie algebra with p-th power operation induced by g 7−→ g p in G1 . The reader is referred to the classical [Jac41] for a discussion of p-restricted Lie algebras. We also can deduce from [Pas79, Th. III.1.7] that any p-restricted strongly central series F F has to contain the dimension series D∗ p G defined in Example 1.39 (because a∗ p (G∗ ) contains IF∗p G): F Proposition 3.5. The filtration D∗ p G is the minimal p-restricted strongly central series, on any group G. [p] The filtration DFp , also denoted by Γ∗ also admits a description by induction, or the more explicit description: Y j [p] Γk G = (Γi G)p . ipj >n These can be found in [Laz54, Th. 5.6] or [Pas79, Th. IV.1.9]. We also refer to [CE16] for a nice discussion of filtrations defined by induction. This description has a nice [p] consequence, similar to Proposition 1.9 (we abbreviate L(Γ∗ G) to L[p] (G)): Proposition 3.6. The p-restricted Lie ring L[p] (G) is generated in degree 1. Precisely, [p] it is generated (as a p-restricted Lie ring) by L1 (G) = Gab ⊗ Fp . Example 3.7. If G is a free group, then L[p] (G) = L(D∗ p G) is the p-restricted Lie algebra generated by the degree-one part inside gr(Fp G) ∼ = TFp (Gab ) so, using PBW over Fp [Jac41, Th. 1] it is the free p-restricted Lie algebra over the Fp -module Gab ⊗ Fp [Laz54, Th. 6.5]. F [p] Remark 3.8. The filtration Γ∗ G (already defined in [Zas39]) is to be distinguished (p) from Stallings’ filtration Γ∗ G, defined in [Sta65]. The latter is the minimal p-torsion strongly central filtration on a group G, where a strongly central filtration is p-torsion when Gpi ⊆ Gi+1 (for all i). See remark 3.14 for more on q-torsion strongly central series. 34 3.3 The p-restricted Andreadakis problem [p] [p] [p] [p] Let us denote A∗ (Γ∗ G) by A∗ (G). Remark that A1 (G) is the group IAG of auto[p] [p] morphisms acting trivially on L1 (Γ∗ G) = Gab ⊗Fp , hence on all of L(Γ∗ G) (because of [p] Proposition 3.6). If we show that A∗ (G) is p-restricted (and we will – see Proposition 3.10) then we get an inclusion:   [p] [p] Γ∗ IAG ⊆ A∗[p] (G). We are thus led to consider a p-restricted version of the Andreadakis problem: Problem 3 (Andreadakis – p-restricted version). What is the difference between the [p] [p] [p] p-restricted strongly central series A∗ (G) and Γ∗ (IAG ) ? [p] Remark 3.9. The group IAG contains IAG as a normal subgroup. Moreover, the [p] quotient IAG /IAG is a subgroup of Aut(G)/IAG , hence of GL(Gab ). In fact, by [p] definition of IAG , it is contained in the congruence group: GL(pGab ) = ker(GL(Gab ) → GL(Gab ⊗ Fp )). When G = Fn is a free group of finite type, then Aut(G)/IAG ∼ = GLn (Z), and [p] IAG /IAG is exactly GLn (p Z). Proposition 3.10 ([HM17], Prop. 8.5). Let G∗ be a p-restricted strongly central filtration, and K be a group acting on G∗ . Then A∗ (K, G∗ ) is a p-restricted strongly central filtration. Proof. Let κ ∈ Aj (K, G∗ ) and g ∈ Gi . Using Corollary 3.1, we get: p [κ , g] = p Y p θ(k, 1)(k) , k=1 p with θ(k, 1) ∈ Gi+kj , for all k. If k < p, then θ(k, 1)(k) ∈ Gp(i+kj) ⊆ Gi+pj+1 . As θ(p, 1) also is in Gi+pj , we have: [κp , g] ∈ Gi+pj . This is true for every g ∈ Gi , for all i. Hence κp ∈ Apj (K, G∗ ), which completes the proof. Proposition 3.10 can be refined: Proposition 3.11. Under the same hypothesis as Proposition 3.10, A∗ (K, G∗ ) n G∗ is a p-restricted strongly central series. Proof. Denote A∗ (K, G∗ ) n G∗ by K∗ . An element of Kj = Aj n Gj is a product κ · g, with κ ∈ Aj and g ∈ Gj . Using Corollary 3.2, we get: p p κg = p Y p p θ(k)(k) = (κg)p · θ(2)(2) · · · θ(p − 1)p · θ(p), (3.11.1) k=1 with θ(1) = κg and θ(k) ∈ Kkj for any k, as K∗ is strongly central. We use this formula to show, by induction on d 6 p, the following result: ∀j, Kjp ⊆ Kdj . 35 This is true for d = 1, obviously. Let us assume that it holds for d − 1. Let κg ∈ Kj . For the sake of clarity, let us rewrite the formula (3.11.1): p p p −1 (κg) = κ g · θ(p) · 2 Y p θ(k)−(k) . k=p−1 Using, respectively, that A∗ is p-restricted (Proposition 3.10), that G∗ is (by definition) and that K∗ = A∗ n G∗ is strongly central, we get: κp , g p , θ(p) ∈ Kpj ⊆ Kdj , where the inclusion comes from the inequality d 6 p. If 2 6 k < p, then θ(k) ∈ Kkj .
As p divides kp , the induction hypothesis implies: p p ⊆ K(d−1)kj ⊆ Kdj , θ(k)(k) ∈ Kkj because (d − 1)kj > dj. Finally, we get what we were looking for: (κg)p ∈ Kdj , which completes the induction step, and the proof of the proposition. Let SCF p be the full subcategory of SCF given by p-restricted strongly central series. As a consequence of Propositions 3.10 and 3.11, we get: Corollary 3.12. The category SCF p is action-representative, the universal action on G∗ being A∗ (G∗ )  G∗ . This allows us to answer [HM17, rk. 8.6]. Indeed, the Lie functor restricts to a functor L : SCF p −→ pLie with values in the category pLie of p-restricted Lie algebras (over Fp ). Actions in pLie are represented by p-restricted derivations, in the sense of Jacobson [Jac41]. As in Paragraph 1.4, an action K∗  G∗ in SCF p induces, by exactness of the Lie functor, an action L(K∗ )  L(G∗ ) in pLie, which is encoded by a morphism between p-restricted Lie algebras: τ : L(K∗ ) −→ Der[p] (L(G∗ )), where Der[p] ⊆ Der is the p-restricted sub-algebra of p-restricted derivations, i.e. derivations ∂ satisfying: ∂(ap ) = adp−1 a (∂a). Let us stress that for g ∈ pLie, the Lie algebra Der(g) is indeed a p-restricted subalgebra of EndFp (g), but is does not act on g in pLie: the Lie algebra g o Der(g) bears no natural p-restricted structure. Remark 3.13. Using Proposition 3.6 instead of 1.9, and replacing derivations by p[p] restricted ones in the proof, we can get an analogous of Lemma 1.30 for A∗ (G): n o [p] [p] Aj (G) = σ ∈ Aut(G) [σ, G] ⊆ Γj+1 (G) . [p] [p] In other words A∗ (G) in the subgroup of automorphisms acting trivially on G/Γj+1 (G). This is exactly the definition used by Cooper [Coo15, def 3.2]. 36 Remark 3.14 (On the q-torsion case). The same statements are true when considering q-torsion strongly central filtrations (q does not have to be a prime number here), except that they are easier to show, because the condition Gqi ⊆ Gi+1 is equivalent to the fact that L(G∗ ) is q-torsion. Precisely, if L(G∗ ) is q-torsion, then Der(L(G∗ )) is too, and the injectivity of the Johnson morphism L(A∗ (G∗ )) ,→ Der(L(G∗ )) implies that L(A∗ (G∗ )) also is. Hence A∗ (G∗ ), and A∗ (G∗ ) n G∗ are q-torsion, so that these give a universal action on G∗ in the category of q-torsion strongly central series. Moreover, L(G∗ ) also gets some kind of q-th power operation, induced by q-th (q) powers in G = G1 . If G∗ = Γ∗ G is Stallings’ filtration on G, then these operations, together with the Lie algebra structure, generate L(G∗ ) from its degree one part, which (q) allow us to get an analogue of Lemma 1.30: Aj (Γ∗ G) is the subgroup of automor(p) phisms acting trivially on G/Γj+1 (G). Using only this definition, Cooper managed to get the above results on the q-torsion case [Coo15]. However, his claim that the p-restricted case worked similarly [Coo15, Lem. 3.7] seems flawed, and we do not see how to get it without the technical work done above (Proposition 3.10). The minimality of Stallings’ filtration also gives an inclusion:  
[q] (q) G , Γ∗ IAG ⊆ A∗ Γ(q) ∗ and a corresponding Andreadakis problem. Nevertheless, our methods in studying the Andreadakis problems so far rely heavily on algebraic structures associated to the F [p] dimension subgroups D∗Z (Fn ) = Γ∗ (Fn ) and D∗ p (Fn ) = Γ∗ (Fn ), so they are not suited to the study of this particular problem. 3.4 3.4.1 The stable p-restricted Andreadakis problem Vanishing of the trace map In the p-restricted context, Proposition 2.33 is replaced by: Proposition 3.15. Let k > 2, and let J ∈ GLm (IFkp Fn ). Then: Tr (J − 1 ) ∈ [T V, T V ]k + (T V )p ⊂ V ⊗k ∼ = I k /I k+1 , Where V = Fnab ⊗ Fp ∼ = Fnp . The proof is exactly the same as the proof of proposition 2.33, over Fp instead of Z, Proposition 2.34 being replaced by: Proposition 3.16. Let f (X1 , ..., Xn ) ∈ V ⊗k . Let C ⊂ Mk (Fp ) be the sub-Z-module generated by the ei,i+1 . Suppose: ∀Ci ∈ C, Tr(f (C1 , ..., Cn )) = 0. Then f ∈ [T V, T V ]k + (T V )p . Proof. We say that two elements u and u0 of V ⊗k are cyclically equivalent, and we write u ∼ u0 if they are conjugate under the action of Z /p. We want to show that f is cyclically equivalent to a p-th power. Let us decompose f , up toP cyclical equivalence, as a sum of pairwise non-cyclically equivalent monomials: f ∼ µg g, where each g is of the form g = Xi1 · · · Xik . If g is such that the iα are pairwise distinct, evaluate each Xiα as Ciα = eα,α+1 , and all Xi not appearing in g as Ci = 0. Then µg = 37 Tr(f (C1 , ..., Cn )) = 0. As a consequence, no such g can appear in our decomposition of f . P Let λ be the algebra morphism from T V to T (V ⊗ V ) sending Xi to j Xij (where Xij = Xi ⊗ Xj ). Take a monomial g = Xi1 · · · Xik as above. Its image is λ(g) = P k j Xi1 j1 · · · Xik jk , the sum being taken over every j = (j1 , ..., jk ) ∈ {1, ..., k} . Let r be the number of monomials cyclically equivalent to h = Xi1 1 · · · Xik k in this sum. Then r is exactly the number of elements of Z /k stabilizing g. It is a multiple of p if and only if g is a p-th power. If we decompose λ(f ) up to cyclic equivalence, as we did earlier for f , the only occurrences of h must come from λ(g), hence the coefficient of h must be rµg . Note that λ(f ) satisfies the same hypothesis as f , because C is stable under addition. Since the Xiα α are pairwise distinct, we can apply the above argument and find that rµg = 0. Thus, µg = 0 or g is a p-th power. Whence the result. Remark 3.17. Any bracket and any p-th power satisfies the condition of Proposition 3.16. For brackets, it follows from the fact that the P trace of a bracket is itself a sum of brackets. For p-th power, remark that if M Q = mi ei,i+1 ∈ Mn (R), where R is an p associative ring of caracteristic p, then M = ( mi ) · 1p , and Tr(1p ) = p · 1 = 0 in k. If k = pl and f = (Xi1 · · · Xil )p , apply this to M = Ci1 · · · Cil (where the Ciα are in C), seen as a p × p-matrix with coefficients in Ml (k). Because of Proposition 3.15, in characteristic p, we will consider the trace map [p] as taking values in C∗ V = T V /([T V, T V ] + (T V )p ), which is the quotient of the cyclic power C∗ V by p-th powers. The conclusion of Proposition 3.15 then becomes: [p] Tr (J − 1 ) = 0 ∈ C∗ V. 3.4.2 Linear algebra Consider the Johnson morphism  
[p] [p] ∼ (τ [p] )0 : L[p] IA[p] −→ Der L (F ) = V ∗ ⊗ L[p] V, n n ∗ the last isomorphism being obtained as in Example 1.33, using Example 3.7 instead of Example 1.8, and replacing derivations by p-restricted ones. When p 6= 2, the morphism [p] (τ1 )0 is surjective. Indeed, the free Fp -Lie algebra LV is a sub-algebra of the free prestricted algebra L[p] V , and this inclusion is an isomorphism in degrees prime to p, in particular in degree 2; thus we can lift the generators of V ∗ ⊗ Λ2 V by the generators of IAn used in the proof of Proposition 1.34. Moreover, L[p] (IAn ) is generated in degree 1 as a p-restricted Lie algebra (cf. 3.6). As a consequence, the image of τ 0 is exactly [p] the p-restricted Lie algebra generated in degree 1 inside Der[p] ∗ (L V ). The reader can easily check that the obvious p-restricted version of Proposition 2.32 does hold: the trace map obtained from free differential calculus can be seen as the [p] [p] [p] composite of the Johnson morphism τ : Lk (A∗ (Fn )) → Derk (L[p] V ) ∼ = V ∗ ⊗ Lk+1 V with: Φ π ι [p] [p] TrM : V ∗ ⊗ Lk+1 V −→ V ∗ ⊗ V ⊗k+1 −→ V ⊗k −→ Ck V, where ι and π again denote the canonical maps. Notation 3.18. Let I[p] denote the image of (τ [p] )0 , the p-restricted Lie algebra generated in degree 1 inside Der(LV ). The following proposition can be seen as a direct consequence of Proposition 3.15. 38 [p] Proposition 3.19. For every k > 2, TrM (Ik ) = {0}. Consider be the subspace of p-restricted derivations stabilizing the free Fp -Lie algebra LV ⊂ L[p] V . It is a p-restricted Lie sub-algebra of Der[p] (L[p] V ). Since each derivation of LV extends to a unique p-restricted derivation of L[p] V , this sub-algebra is isomorphic to Der(LV ). Under the identification with the graded module V ∗ ⊗L[p] V , it corresponds exactly to V ∗ ⊗ LV . As a consequence, if p 6= 2, the degree one part is the same. Hence: [p] I[p] ⊆ DerLV (L[p] V ). This implies that there is no stable surjectivity here: we can easily easily produce examples of automorphisms whose associated derivation does not preserve LV . For instance, take any world in Γk (Fn ) not containing any occurrence of x1 . Then the [p] automorphism ϕ defined by x1 7→ wp x1 and xi 7→ xi when i 6= 1 is obviously in Apk−1 , but τ (ϕ) = X1∗ ⊗ (w − 1)p sends X1 outside of LV . 3.4.3 Stable cokernel of i∗ We close the present paper with a quantification of the lack of stable surjectivity in the p-restricted case. Let k > 2. Like in paragraph 2.5.3, we get a commutative diagram with exact rows: Īk V ∗ ⊗ Lk+1 V φ [T V, T V ]k Xk Φ Φ π V ⊗k Ck V. Here, V denotes Fnab ⊗ Fp , and Ī∗ = I∗ ⊗ Fp is the sub-Lie algebra generated in degree one inside Der(LV ). The space Xk is just the quotient of V ∗ ⊗ Lk+1 V by Īk . We will be interested in a slightly different diagram, though: [p] Ik V ∗ ⊗ Lk+1 V Xk0 Φ φ Φ [p] π V ⊗k [T V, T V ]k + (T V )p Ck V. We can still apply the calculations from [Sat12] to show that if n > k + 2, then Φ is [p] surjective and ker Φ ⊆ Īk ⊆ Ik . Only, now φ could have a cokernel. From [Sat12], we only get that brackets are in its image, so this cokernel can only come from p-th powers. In particular, it is concentrated in degrees divisible by p. We denote it by K. The same application of the snake lemma as in the proof of Proposition 2.37 gives that K is also the kernel of Φ, and that Φ is surjective. Now consider the diagram: [p] Ik V ∗ ⊗ Lk+1 V Xk0 ι ker(TrM )k Φ [p] V ∗ ⊗ Lk+1 V 39 TrM [p] Ck V. [p] Denote by L the cokernel of the middle inclusion, then L = V ∗ ⊗ (Lk+1 V /Lk+1 V ) is concentrated in degrees k = pl − 1 (with l > 1). The snake lemma gives a short exact sequence: 0 K coker(ι) L 0. Since the trace map TrM ◦τ vanishes, we have a commutative diagram: i∗ [p] L(IAn ) τ0 [p] L(A∗ ) τ ker(TrM ). This implies that the cokernel of i∗ injects into the cokernel of ι. Thus, we have proved: Proposition 3.20. Fix n an integer, and consider only degrees k 6 n−2. The cokernel of the canonical morphism i∗ : L(IAn[p] ) → L(A∗[p] (Fn )) is concentrated in degrees k = pl − 1 and k = pl (for l > 1). If k = pl − 1, then [p] coker((i∗ )k ) injects into V ∗ ⊗ (Lk+1 V /Lk+1 V ). If k = pl, it is a sub-quotient of V ⊗l . Remark 3.21. The tensor power V ⊗l appearing in the proposition is in fact the Frobenius twist of V ⊗l . This has no consequence here, as the Frobenius map is trivial on Fp , but it should be kept in mind for any functorial study of this situation. References [And65] Stylianos Andreadakis. On the automorphisms of free groups and free nilpotent groups. Proc. London Math. Soc. (3), 15:239–268, 1965. 2 [Bar13] Laurent Bartholdi. Automorphisms of free groups. I. New York J. Math., 19:395–421, 2013. 2, 25, 26, 28, 32 [Bar16] Laurent Bartholdi. Automorphisms of free groups. I—erratum [ MR3084710]. New York J. Math., 22:1135–1137, 2016. 2 [BB04] Francis Borceux and Dominique Bourn. Mal’cev, protomodular, homological and semi-abelian categories, volume 566 of Mathematics and its Applications. Kluwer Academic Publishers, Dordrecht, 2004. 7, 11 [BB07] Francis Borceux and Dominique Bourn. Split extension classifier and centrality. In Categories in algebra, geometry and mathematical physics, volume 431 of Contemp. Math., pages 85–104. Amer. Math. Soc., Providence, RI, 2007. 2, 8 [BBM07] Mladen Bestvina, Kai-Uwe Bux, and Dan Margalit. Dimension of the Torelli group for Out(Fn ). Invent. Math., 170(1):1–32, 2007. 1, 13 [BC05] F. Borceux and Maria Manuel Clementino. Topological semi-abelian algebras. Adv. Math., 190(2):425–453, 2005. 7 40 [BEH08] Laurent Bartholdi, Bettina Eick, and René Hartung. A nilpotent quotient algorithm for certain infinitely presented groups and its applications. Internat. J. Algebra Comput., 18(8):1321–1344, 2008. 2 [BJK05] Francis Borceux, George Janelidze, and Gregory M. Kelly. On the representability of actions in a semi-abelian category. Theory Appl. Categ., 14:No. 11, 244–286, 2005. 2, 8 [BLGM90] Roger M. Bryant, Frank Levin, Chander K. Gupta, and Horace Y. Mochizuki. Nontame automorphisms of free nilpotent groups. Comm. Algebra, 18(11):3619–3631, 1990. 3, 28, 29, 33 [BMS67] Hyman Bass, John Milnor, and Jean-Pierre Serre. Solution of the congruence subgroup problem for SLn (n > 3) and Sp2n (n > 2). Inst. Hautes Études Sci. Publ. Math., 33:59–137, 1967. 16 [Bou08] Dominique Bourn. Commutator theory, action groupoids, and an intrinsic Schreier-Mac Lane extension theorem. Adv. Math., 217(6):2700–2735, 2008. 2, 8 [CE16] Michael Chapman and Ido Efrat. Filtrations of free groups arising from the lower central series. J. Group Theory, 19(3):405–433, 2016. 34 [CEFN14] Thomas Church, Jordan S. Ellenberg, Benson Farb, and Rohit Nagpal. FI-modules over Noetherian rings. Geom. Topol., 18(5):2951–2984, 2014. 31 [Coo15] James Cooper. Two mod-p Johnson filtrations. J. Topol. Anal., 7(2):309– 343, 2015. 36, 37 [CP17] Thomas Church and Andrew Putamn. Generating the johnson filtration ii: finite generation. arXiv:1704.01529, 2017. 1 [CV86] Marc Culler and Karen Vogtmann. Moduli of graphs and automorphisms of free groups. Invent. Math., 84(1):91–119, 1986. 1 [Dar68] Rex S. Dark. On nilpotent products of groups of prime order. PhD thesis, University of Cambridge, 1968. 33 [Dja16a] Aurélien Djament. Décomposition de hodge pour l’homologie stable des groupes d’automorphismes des groupes libres. arXiv:1510.03546, 2016. 1 [Dja16b] Aurélien Djament. Des propriétés de finitude des foncteurs polynomiaux. Fund. Math., 233(3):197–256, 2016. 31 [DP12] Matthew Day and Andrew Putman. A Birman exact sequence for Aut(Fn ). Adv. Math., 231(1):243–275, 2012. 1 [DP16] Matthew B. Day and Andrew Putamn. On the second homology group of the torelli subgroup of aut(fn ). arXiv:1408.6242v3, 2016. 1, 2 [DV15] Aurélien Djament and Christine Vespa. Sur l’homologie des groupes d’automorphismes des groupes libres à coefficients polynomiaux. Comment. Math. Helv., 90(1):33–58, 2015. 1 41 [FM12a] Benson Farb and Dan Margalit. A primer on mapping class groups, volume 49 of Princeton Mathematical Series. Princeton University Press, Princeton, NJ, 2012. 1 [FM12b] Stefano Francaviglia and Armando Martino. The isometry group of outer space. Adv. Math., 231(3-4):1940–1973, 2012. 1 [Fox53] Ralph H. Fox. Free differential calculus. I. Derivation in the free group ring. Ann. of Math. (2), 57:547–560, 1953. 19 [Fra33] Hermann Frasch. Die Erzeugenden der Hauptkongruenzgruppen für Primzahlstufen. Math. Ann., 108(1):229–252, 1933. 17 [Gal11] Søren Galatius. Stable homology of automorphism groups of free groups. Ann. of Math. (2), 173(2):705–768, 2011. 1 [Gro52] Emil Grosswald. On the parabolic generators of the principal congruence subgroups of the modular group. Amer. J. Math., 74:435–443, 1952. 17 [Hal34] Philip Hall. A contribution to the theory of groups of prime-power order. Proc. London Math. Soc. (2), 36:29–95, 1934. 33 [HL11] Manfred Hartl and Bruno Loiseau. A characterization of finite cocomplete homological and of semi-abelian categories. Cah. Topol. Géom. Différ. Catég., 52(1):77–80, 2011. 7 [HM17] Kazuo Habiro and Gwenael Massuyeau. Generalized johnson homomorphisms for extended n-series. arXiv:1707.07428, 2017. 2, 4, 12, 35, 36 [Jac41] Nathan Jacobson. Restricted Lie algebras of characteristic p. Trans. Amer. Math. Soc., 50:15–25, 1941. 34, 36 [Joh83] Dennis Johnson. A survey of the Torelli group. In Low-dimensional topology (San Francisco, Calif., 1981), volume 20 of Contemp. Math., pages 165– 179. Amer. Math. Soc., Providence, RI, 1983. 13 [JW04] Craig A. Jensen and Nathalie Wahl. Automorphisms of free groups with boundaries. Algebr. Geom. Topol., 4:543–569, 2004. 1 [Kal50a] Léo Kaloujnine. Sur quelques propriétés des groupes d’automorphismes d’un groupe abstrait. C. R. Acad. Sci. Paris, 230:2067–2069, 1950. 8 [Kal50b] Léo Kaloujnine. Sur quelques propriétés des groupes d’automorphismes d’un groupe abstrait. (Généralisation d’un theorème de M. Ph. Hall). C. R. Acad. Sci. Paris, 231:400–402, 1950. 8 [Kaw06] Nariya Kawazumi. Cohomological aspects of magnus expansions. arXiv:math/0505497, 2006. 13 [KM97] Sava Krstić and James McCool. The non-finite presentability of IA(F3 ) and GL2 (Z[t, t−1 ]). Invent. Math., 129(3):595–606, 1997. 1 [Laz54] Michel Lazard. Sur les groupes nilpotents et les anneaux de Lie. Ann. Sci. Ecole Norm. Sup. (3), 71:101–190, 1954. 3, 6, 8, 14, 34 42 [Lop14] Jonathan Lopez. Lie algebras and cohomology of congruence subgroups for SLn (R). J. Pure Appl. Algebra, 218(2):256–268, 2014. 3, 16 [LS76] Ronnie Lee and Robert H. Szczarba. On the homology and cohomology of congruence subgroups. Invent. Math., 33(1):15–53, 1976. 3, 17 [ML98] Saunders Mac Lane. Categories for the working mathematician, volume 5 of Graduate Texts in Mathematics. Springer-Verlag, New York, second edition, 1998. 7 [Mor93] Shigeyuki Morita. Abelian quotients of subgroups of the mapping class group of surfaces. Duke Math. J., 70(3):699–726, 1993. 3, 25 [MP09] Roman Mikhailov and Inder Bir Singh Passi. Lower central and dimension series of groups, volume 1952 of Lecture Notes in Mathematics. SpringerVerlag, Berlin, 2009. 14 [MS17] Gwenael Massuyeau and Takuya Sakasai. Morita’s trace maps on the group of homology cobordisms. arXiv:1606.08244, 2017. 2, 28 [Nie24] Jakob Nielsen. Die Isomorphismengruppe der freien Gruppen. Math. Ann., 91(3-4):169–209, 1924. 1, 13 [Pas79] Inder Bir S. Passi. Group rings and their augmentation ideals, volume 715 of Lecture Notes in Mathematics. Springer, Berlin, 1979. 14, 33, 34 [Pas85] Donald S. Passman. The algebraic structure of group rings. Robert E. Krieger Publishing Co., Inc., Melbourne, FL, 1985. Reprint of the 1977 original. 33 [Rip72] Eliyahu Rips. On the fourth integer dimension subgroup. Israel J. Math., 12:342–346, 1972. 14 [Rot95] Joseph J. Rotman. An introduction to the theory of groups, volume 148 of Graduate Texts in Mathematics. Springer-Verlag, New York, fourth edition, 1995. 13 [RWW17] Oscar Randal-Williams and Nathalie Wahl. Homological stability for automorphism groups. Adv. Math., 318:534–626, 2017. 1 [Sat12] Takao Satoh. On the lower central series of the IA-automorphism group of a free group. J. Pure Appl. Algebra, 216(3):709–717, 2012. 2, 3, 25, 30, 39 [Sta65] John Stallings. Homology and central series of groups. J. Algebra, 2:170– 181, 1965. 34 [Zas39] Hans Zassenhaus. Ein Verfahren, jeder endlichen p-Gruppe einen Lie-Ring mit der Charakteristik p zuzuordnen. Abh. Math. Sem. Univ. Hamburg, 13(1):200–207, 1939. 34 43 

4math.GR

Programmable Agents arXiv:1706.06383v1 [cs.AI] 20 Jun 2017 Misha Denil Sergio Gómez Colmenarejo Serkan Cabi David Saxton Nando de Freitas DeepMind {mdenil,sergomez,cabi,saxton,nandodefreitas}@google.com Abstract We build deep RL agents that execute declarative programs expressed in formal language. The agents learn to ground the terms in this language in their environment, and can generalize their behavior at test time to execute new programs that refer to objects that were not referenced during training. The agents develop disentangled interpretable representations that allow them to generalize to a wide variety of zero-shot semantic tasks. 1 Introduction This paper shows how to build agents that can execute declarative programs expressed in a simple formal language. The agents learn to ground the terms of the language in their environment through experience. The learned groundings are disentangled and compositional; at test time we can ask the agents to perform tasks that involve novel combinations of properties and they will do so successfully. The agents learn to distinguish distinct properties that are referenced together during training; when trained on tasks that always reference objects through a conjunction of shape and color the agents can generalize at test time to tasks that reference objects through either property in isolation. Completely novel object properties can be referenced through the principle of exclusion (i.e. the object whose color you have not seen before), and our agents are able to successfully complete tasks that reference novel objects in this way. This works even when the agents have never encountered programs involving this type of reference during training. Referring to objects that possess multiple novel properties is also successful, as is referring to objects through combinations of known and unknown properties. Our agents are robust to catastrophic forgetting. If we train on a subset of tasks then not only can the agents accomplish novel tasks zero-shot, but if we switch to training only on the novel tasks then performance on the original tasks does not degrade. Our agents are implemented as deep neural networks, and trained end to end with reinforcement learning. The agents learn how programs refer to properties of objects and how properties are assigned to objects in the world entirely through their experience interacting with their environment. Natural and interpretable assignments of properties to objects emerge without any direct supervision of how these properties should be assigned. No auxiliary prediction or control tasks are required for good performance (although exploring their effects in this setting is a nice possibility for future work). The representations learned by our agents are extremely interpretable. There is an explicit mapping between activation maps in the programmable layers of our agents and terms in the programs they execute. Visualizing these activation maps provides an immediately interpretable picture of how the agents assign properties to objects in their world. The extremely powerful generalization our agents achieve is made possible by our novel “Programmable Network” architecture. Programmable Networks impose a sophisticated bottleneck on the agent’s representations whose structure ensures that their representations will generalize. Submitted to 31st Conference on Neural Information Processing Systems (NIPS 2017). 2 Related work The work in this paper builds on two recent lines of work in deep learning. One is the resurgence of modular networks that can be decomposed and recomposed to compute different functions. Neural Module Networks [2, 1], which use modularity in visual question answering, have been particularly influential on our thinking. This type of approach has also recently seen great success in answering questions that involve relational reasoning [15, 16]. The second recent trend we follow is the rise of relational neural network architectures, particularly Interaction Networks [4] and the Neural Physics Engine [8]. At a high level these works on relational architectures are subsumed by the various neural approaches to graph processing. A nice overview of recent techniques can be found in Gilmer et al. [12]. Compositionality has become a topic of great interest in machine learning, robotics and cognitive science [20, 30, 19, 9, 35, 26], and it has been central to research in language [32, 25, 33, 36]. Compositional neural architectures paired with different forms of attention have led to impressive results in natural language interfaces for database tables, and language acquisition in 2D navigation environments [28, 37]. Yu et al. [37] present a different deep RL approach to grounding symbols on perception, and introduce a visual question answering auxiliary task to improve zero-shot generalization. Use of programs to specify hard deterministic gating is also reminiscent of PAQ8 [24, 18]. 3 Tasks as declarative programs Building neural networks that infer or execute computer programs has been a popular topic of recent research. Different approaches to this problem focus on networks that infer programs [3, 10, 22], on networks that execute programs [39, 30, 5] and on networks that jointly do both [14, 17, 38, 29, 13]; although there is much overlap between many of these approaches. The majority of these works model their networks after an imperative programming style, although researchers have also begun to explore structuring their networks as functional (recursive) programs as well [7, 30]. In this work we depart from previous work by considering networks that execute a simple declarative language. Paradigmatic examples of declarative languages are PROLOG [6, 5] and SQL. The declarative paradigm provides an appealing and flexible way to describe tasks for agents [23]. Our general framework is as follows: A goal is specified as a state of the world that satisfies a relation between two objects. Objects are associated with sets of properties (e.g. their color and shape). The vocabulary of properties gives rise to a system of base sets which are the sets of objects that share each named property (e.g. RED is the set of red objects, etc). The full universe of discourse is then the Boolean algebra generated by these base sets. We require two things for each program. The verifier has access to the true state of the environment, and can inspect this state to determine if it satisfies the program. We also need a search procedure which inspects the program as well as some summary of the environment state and decides how to modify the environment to bring the program closer to satisfaction. These components correspond directly to components of the standard RL setup. Notably, the verifier is a reward function (which has access to privileged information about the environment state) and the search procedure is an agent (which may have a more restrictive observation space). There are several advantages to thinking in this way. The first is that building semantic tasks becomes straightforward, we need only specify a new program to obtain a new reward function that depends on semantic properties of objects in the environment. Consequently, we can easily specifiy diverse, combinatorial tasks. Another advantage is that this framing places the emphasis squarely on generalization to new tasks. A program interpreter is not very useful if you must enumerate all programs you might want to run up front. Our goal is not only to perform combinatorial tasks, but to be able to specify new behaviors at test time, and for them to be accomplished successfully without additional training. This type of generalization is quite difficult to achieve with deep RL. In this paper we focus on reaching for blocks, but the simplicity of the individual tasks belies the complexity of what we achieve. The challenge here is to have agents ground the terms of the programming language in their environment and, at test time, to be able to execute new programs that use these terms in novel ways. 2 Figure 1: The programmable reaching environment consists of a square arena with a robot arm in the center. Blocks are scattered randomly and the task in each episode is to reach towards a specific block, identified by a combination of shape and color. The leftmost frame shows a training episode, and the remaining frames show various types of generalization that agents trained in this setting are capable of, including operating on single properties, varying numbers of objects, and novel shapes. 4 Programmable reaching environment Figure 1 shows several visualizations of our programmable reaching environment, which consists of a mechanical arm in the center of a large table. The arm is a simplified version of the Jaco arm, where the body has been stereotyped to basic geoms (rigid body building components [34]), and the finger actuators have been disabled. In each episode a fixed number of blocks appear at random locations on the table. Each block has both a shape and a color, and the combination of both are guaranteed to uniquely identify each block within the episode. The programmable reaching environment is implemented with the MuJoCo [34] physics engine, and hence the objects are subject to friction, contact forces, gravity, etc. Each task in the reaching environment is to put the “hand” of the arm (the large white geom) near the target block, which changes in each episode. The task is communicated to the agent with two integers specifying the target color and shape, respectively. The complexity of the environment can be varied by changing the number, colors and shapes that blocks can take. In this work we consider 2x2 (two colors and two shapes) and 3x3 variants. We can also control the number of blocks that appear on the table in each episode, and we fix it to four blocks during training to study generalization to other numbers. When there are more possible blocks than are allowed on the table the episode generator ensures that the reaching task is always achievable (i.e. the agent is never asked to reach for a block that is not present). The arm has 6 actuated rotating joints, which results in 6 continuous actions in the range [0, 1]. The observable features of the arm are the positions of the 6 joints, along with their angular velocities. The joint positions are represented as the sin and cos of the angle of the joint in joint coordinates. This results in a total of 18 (6 × 2 + 6) body features describing the state of the arm. Objects are represented using their 3d position as well as a 4d quaternion representing their orientation, both represented in the coordinate frame of the hand. Each block also has a 1-hot encoding of its shape (4d) and its color (5d), for a total of 16 object features per block. We provide object features for all of the blocks on the table as well as the hand, but not for the other bodies that compose the arm. We can write a reaching program in this environment as NEAR(HAND, AND(RED, CUBE)) (1) which specifies that the hand should be near the red cube. This condition can be checked automatically by the verifier to produce a reward function that takes on the values ±1. The verifier interprets NEAR by thresholding the distance between its arguments, but this is exposed to the agent only through the values the reward function takes in different states. To show zero-shot generalization we partition the set of possible target blocks into train and test conditions. We train an agent by choosing a target randomly from the train conditions in each episode, and evaluate the agent on its performance reaching for blocks from the test conditions, which were never targets during training. We call an assignment of targets to train and test conditions a design. The left panel in Figure 2 shows the designs we consider for the 2x2 and 3x3 variants of the reaching environment. Rows and columns of each matrix correspond to different shapes and colors, respectively and each cell of each matrix corresponds to a different task. The cells are color coded so that yellow 3 Objects 2-L 2-C Program Red Properties 3-D Blue White ∧ Cube Sphere ⌦ Hand ∨ Relevant Objects Vector (p) Figure 2: Left: Different designs for 2x2 and 3x3 environments. Center: Relationship between Ω and Φ. The features in Ω have detectors applied to them to produce the disentangled representation Φ. Right: Example of how the detectors interact with the program. The property detector outputs Φ are shown as binary to ease interpretation, but in practice they are learned continuous values. indicates train conditions, and magenta indicates conditions only seen at test time. The magenta tasks are used to evaluate zero-shot generalization after the agent is trained. 5 Programmable networks Due to space constraints we describe only the main components of the Programmable Network architecture in this section. Full details of how everything comes together can be found in Appendix A. 5.1 Executing programs We begin by explaining how the program operates when given a mapping between objects and properties, denoted by Φ and explained below. In later sections we expand on this to explain how the assignment of properties to objects can be learned, and how the assignment process can be made differentiable to enable end to end training from a behavioural objective; however, for clarity of presentation it is easiest to begin under the simplifying assumption that the assignments of properties to objects are crisp (i.e. 0-1) and known. The input to the program is a matrix Φ whose columns are objects and rows are properties. The elements of this matrix are in {0, 1} (this will be relaxed later) where Φij = 1 indicates that the object j has property i. Figure 2 shows how the set of relevant objects are identified for the example reaching program in Equation 1. In this example the environment contains four blocks and the robot hand, for a total of five objects. Each object has two properties, a color and a shape, which are together enough to uniquely identify it. The set of relevant objects for this program can be expressed as OR(HAND, AND(RED, CUBE)) (2) and an indicator function for this set can be obtained by combining the rows of Φ, as explained below. Each row of Φ corresponds to a particular property that can be referenced in a program, and the values in the rows serve as indicator functions over subsets of objects that have the corresponding property. These can be used to select new groups of objects by standard Boolean operations, which can be implemented by applying elementwise operations to the rows of Φ. Figure 2 shows how the set in Equation 2 is obtained. The functions AND and OR in the specification (shown as ∧ and ∨ in the figure) correspond to the set operations of intersection and union, respectively. The result is a vector whose elements form an indicator function over objects. The set corresponding to the indicator function contains both the robot hand and the red cube and excludes the remaining objects. We call the result of this operation the relevant objects vector and denote it with p (for “presence” in the set of relevant objects). This vector will play a role in the down stream reasoning process of our agents. The order of rows and columns of Φ is arbitrary. We take advantage of this to assign indexes to named properties in an arbitrary (but fixed) order. This is the same type of assignment that is done 4 for language models when words in the model vocabulary are assigned to indexes in an embedding matrix, and imposes no loss of generality beyond restricting our programs to a fixed “vocabulary” of properties. Note that none of the operations described in this section depend on the number of objects. We will take advantage of this flexibility in the experiments to show that the behaviors we learn can generalize to different numbers of objects in the environment. 5.2 Differentiating through program execution The program execution described in the previous section makes use of set operations on indicator functions, which are uniquely defined when the sets are crisp; however, this uniqueness is lost if the sets are soft. We would like to apply programs to soft sets so that the assignment of properties to objects can be learned by backprop. This requires not only that our set operations apply to soft sets, but also that they be differentiable. There are many ways to meet both of these requirements; for our purposes it is convenient to choose the following assignment: not(x) = 1 − x and(x, y) = xy or(x, y) = x + y − xy (3) It can be verified that these operations are self-consistent (e.g. x = not(not(x))), and reduce to standard Boolean operations when x, y ∈ {0, 1}. This particular assignment is convenient because each operation always gives non-zero derivatives to all arguments. 5.3 Learned grounding The discussion in Section 5.1 assumed that the assignment of properties to objects is given; however, it is much more interesting if our agents can learn to create the matrix Φ rather than having it provided by the environment. In our architecture learning to populate the elements of Φ is the role of detectors. A detector operates on a matrix of features Ω. Similar to Φ, the columns of Ω correspond to objects but the rows of Ω are opaque vectors, populated by whatever information the environment provides about objects. We create one detector for each property in our vocabulary. Each detector is a small neural network that maps columns ωj of Ω to a value in [0, 1]. Detectors are applied independently to each column of the matrix Ω, and each detector populates a single row of Φ. Groups of detectors corresponding to sets of mutually exclusive properties (e.g. blocks can only have one color in our experiments) have their outputs coupled by a softmax. For example, when learning the matrix in Figure 2 each column is the output of two softmaxes, one over colors and one over shapes. These columns of Ω are filled with whatever features the environment provides, position, orientation, etc. These features must have enough information to identify the properties in our vocabulary, but this information is permitted to be entangled with other features in Ω, while it must be disentangled in Φ. This relationship is diagrammed in Figure 2. It would be simple to pre-train the detectors for each property and provide those to the agent. However, one of the contributions of this paper is to show that with our network architecture we do not need to do this. Our agents learn to identify meaningful properties of objects and to reason about sets of objects formed by combinations of these properties in a completely end to end way. 5.4 Relational reasoning Up to this point we have described completely separate processing for each object. The agent receives a matrix Ω whose rows are features and columns are objects. We apply a battery of detectors to each column Ω to create the matrix Φ where rows are properties and columns are again objects. The program then applies elementwise operations to the rows of Φ to create the relevant objects vector p. In order to allow reasoning over relationships between objects we introduce a message passing scheme to exchange information between the objects selected by the relevant objects vector. Our message passing operation closely resembles an Interaction Network [4] with some additional features. 5 1.0 0.5 0.0 0.5 1.0 Canonical 2-L Train 2-L Eval 2-C Programmable 2-L 1.0 0.5 0.5 0.0 0.0 2-L Train 2-L Eval 0.5 0 2 4 6 8 Steps (millions) 10 1.0 1.0 0 0.5 1.0 1.5 Steps (millions) Programmable 3-D 0.5 0.0 3-D Train 3-D Eval 0.5 1.0 0.0 2.0 Distilled 3-D 1.0 1.0 2.0 3.0 4.0 Steps (millions) 3-D Train 3-D Eval 0.5 5.0 1.0 0.0 1.0 2.0 3.0 4.0 Steps (millions) 5.0 Figure 3: Left: Performance of the canonical and programmable architectures on the 2-C and 2-L environment variants. Center: Training and generalization curves for programmable agents with learned detectors trained on the 2-L and 3-D reaching variants. Learning is somewhat unstable in these cases, but successful agents learn to generalize perfectly. Right: Training and generalization curves for distilled agents trained on the 3-D variant (results for 2-L are similar). Exploiting ground truth property information at training time substantially stabilizes learning. Using ωi and ωj to represent columns of Ω, we can write a single round of message passing as X ωi0 = f (ωi ) + αij r(ωi , ωj ) (4) j ωi0 where is the resulting transformed features of object i. This operation is applied to each column of Ω, and the resulting vectors are aggregated into the columns of a new matrix Ω0 . The function f (ωi ) produces a local transformation of the features of a single object, and r(ωi , ωj ) provides a message from object j → i. We implement the functions f and r with small MLPs, and structurally the message passing operation is similar to that of Interaction Networks [4] and the Neural Physics Engine [8]. Messages between objects are mediated by edge weights αij which are determined using a modified version of the neighborhood attention operation of Duan et al. [11], pj exp α̃ij ci = Linear(ωi ) qi = Linear(ωi ) α̃ij = wT tanh(qi + cj ) αij = P k pk exp α̃ik Recall that p is the relevant objects vector (which is a result of the program, see Figure 2), and that its elements lie in the interval [0, 1]. To understand why they are included in this way, consider what happens if pj = 0, which means that object j is not a relevant object for the current task. In this case the resulting αij = 0 also, and the effect is that the message from j → i in Equation 4 does not contribute to ωi0 . In other words, task-irrelevant objects do not pass messages to task-relevant objects during relational reasoning. If we were to consider different types of relations in our programs (beyond NEAR), we would do so by parameterizing the message passing operation based on the relation to be achieved. Since we consider only reaching tasks, this extra complexity is not needed in the present work. After message passing an output can be produced by computing a = MLP(tanh(hΩ0 , pi)). For full details see Appendix A. 6 Experiments In this section we look at various ways that agents implemented as Programmable Networks can generalize. Throughout this section we refer to different variants of the programmable reaching environment using the names indicated in the left panel of Figure 2. All of our agents are trained using Deterministic Policy Gradient [31, 21], where both the actor and the critic are Programmable Networks. See Appendix A for full details of our training setup. 6.1 Standard methods fail to generalize We start by showing that the programmable reaching environment poses a non-trivial challenge. This is perhaps not immediately obvious, because the individual tasks (i.e. reaching for a single block) are very simple. The 2-C design, where all possible target blocks are seen during training, is readily solved by standard deep architectures. Where standard architectures fail (and do so catastrophically) is in generalization to unseen combinations of properties. 6 1.0 0.5 0.0 0.5 1.0 train eval only only not not three five nine ignore new new new random color shape color shape blocks blocks blocks new color shape color agent and shape Figure 4: Generalization experiments. Each column shows the performance distribution of a single agent on zero-shot tasks under different conditions. See text for a full description. For the standard architecture we compare against an architecture similar to the one of Lillicrap et al. [21], modified for feature based observations. We call this the “canonical” architecture as it has been shown to achieve good performance on many continuous control tasks. Performance of the canonical architecture on the 2-C and 2-L designs is shown in the left plot of Figure 3. Reward per step of ~0.6 is effectively perfect performance in these environments. What we see from these curves is that the canonical architecture achieves good performance on all of the tasks it was trained on, but in the 2-L case it completely fails to generalize its behavior to reaching for the held out target. Good performance on 2-C verifies that this is not merely a capacity problem; when the network is trained on all four tasks it is able to achieve all four. 6.2 Zero-shot tasks The center plots in Figure 3 show performance curves for the Programmable Agent on the training and evaluation conditions of the 2-L and 3-D designs. Learning is not always stable, but is nonetheless successful in many cases. We emphasize that the difficulty here is stability rather than a performance; agents that succeed in training perform very well in generalization. Stability of training can be improved by taking advantage of ground truth property information in the critic at training time. In this setting the actor still learns detectors, but the critic does not. We call this the “distilled” model, and learning curves for distilled agents on the 3-D design are shown in the right of Figure 3. 6.3 Strong generalization Because our architecture partitions the objects into sets that have different properties we can achieve very strong generalization. Figure 4 shows zero-shot performance of one of our agents with fully learned detectors (i.e. no distillation) trained on the 3-D design. Each column summarizes the reward per step achieved by the agent over 100 testing episodes under different generalization conditions. Figure 4 also shows two baseline conditions. The train condition corresponds to the agent being tested on new episodes of the tasks it was trained on, and the random condition corresponds to the performance of the same agent with random weights. We only show results for the Programmable Agent because we were not successful in training agents with the canonical architecture on any of 3x3 design. Even when all tasks are seen during training (the most informative possible 3x3 setting) the canonical architectures were no better than random. The zero-shot conditions in Figure 4 are as follows:1 • • • • eval shows performance on the zero-shot tasks of the 3-D design, as shown in Figure 2. only color/shape shows performance when only the color or shape is specified. not color/shape shows performance when the color or shape to not reach for is specified. # blocks show performance of zero-shot tasks with different numbers of blocks on the table. All training episodes had exactly four blocks. • ignore new blocks shows the performance of zero-shot tasks in the presence of new block shapes and colors. 1 Videos for each condition can be found at: https://goo.gl/UkvWYE 7 3-D Train Conditions 1.0 0.5 0.5 0.0 0.0 Train Eval 0.5 1.0 3-D Eval Conditions 1.0 0 1 2 3 4 Steps (millions) Train Eval 0.5 5 1.0 0 1 2 3 4 Steps (millions) 5 Figure 5: Left: performance on the off-diagonal and diagonal tasks for the 3x3 environment. Training begins on the off-diagonal tasks and switches to the diagonal tasks at 2.5m steps (marked by the solid black line). The dashed black line shows the point where all transitions from off-diagonal task have been pushed out of the replay buffer. Right: The learned property matrix Φ (actually its transpose) produced from the adjacent environment state. See Appendix C for additional visualizations. • reach new shape shows the performance when reaching for blocks with a novel shape and known color. • reach new color shows the performance when reaching for blocks with a novel color and known shape. • reach new shape and color shows the performance when reaching for a block when both the color and shape are novel. We introduce new blocks by adding an additional shape and color to the set of possible blocks that can appear on the table (for a total of 16 possible combinations). In the ignore new blocks condition blocks with the new properties can appear on the table as distractors, but are never the target of a reaching program. In the only shape and only color conditions we modify the environment generation process to ensure that the target is uniquely identified by the specified color or shape. Apart from the train and random conditions, every episode of every task in Figure 4 is zero-shot. Every program specifies the target in a way the agent did not see at training time. When showing generalization to different numbers of blocks (in the # blocks conditions) and the ability to ignore novel distractor blocks (in ignore new blocks) we are showing the agent reaching for a zero-shot target under these conditions. The right panel of Figure 5 shows the learned detector outputs in a successful agent. This figure shows that the agent can correctly identify properties of objects referenced by the programs it executes. 6.4 No catastrophic forgetting Our agents are robust against catastrophic forgetting. To show this we trained a collection of agents on the 3x3 environment for 5m steps. For the first 2.5m steps we follow the train condition of the 3-D design, and for the remaining 2.5m steps we switch to training on the test conditions. The plots in Figure 5 show performance curves for these agents on both the training and test conditions throughout the full 5m steps of training. Performance on the train condition does not degrade even millions of steps after the switching to training on new tasks. 7 Conclusion Our Programmable Network architecture enables us to build agents that execute declarative programs expressed in formal language. Our agents learn to ground the terms of the programs in their environment and can leverage these grounded terms to generalize beyond the tasks they were trained on. Our agents achieve nearly perfect generalization on a variety of zero-shot tasks where standard deep RL architectures completely fail. All of this is achieved with only small concessions from the Deep Learning zeitgeist. We assume that the boundaries between objects in the agent observations are known, and the object properties our agents can reason over must come from a predefined vocabulary, which is similar to the restriction one has in word based language modelling. We also use a fixed mapping from programs to architecture. Future work will focus on improving the robustness of training, and on scaling the method to deal with more types of properties (e.g. mass, size, texture, etc) and relations. We are also interested in extending this method to work from vision. 8 References [1] J. Andreas, M. Rohrbach, T. Darrell, and D. Klein. Learning to compose neural networks for question answering. In North American Chapter of the Association for Computational Linguistics, 2016. [2] J. Andreas, M. Rohrbach, T. Darrell, and D. Klein. Neural module networks. In Computer Vision and Pattern Recognition, 2016. [3] M. Balog, A. L. Gaunt, M. Brockschmidt, S. Nowozin, and D. Tarlow. DeepCoder: Learning to write programs. In International Conference on Learning Representations, 2017. [4] P. W. Battaglia, R. Pascanu, M. Lai, D. Rezende, and K. Kavukcuoglu. Interaction networks for learning about objects, relations and physics. In Neural Information Processing Systems, 2016. [5] M. Bošnjak, T. Rocktäschel, J. Naradowsky, and S. Riedel. Programming with a differentiable forth interpreter. Technical report, University College London, 2016. [6] I. Bratko. Prolog programming for artificial intelligence. Pearson education, 2001. [7] J. Cai, R. Shin, and D. Song. Making neural programming architectures generalize via recursion. In International Conference on Learning Representations, 2017. [8] M. B. Chang, T. Ullman, A. Torralba, and J. B. Tenenbaum. A compositional object-based approach to learning physical dynamics. In International Conference on Learning Representations, 2017. [9] C. Devin, A. Gupta, T. Darrell, P. Abbeel, and S. Levine. Learning modular neural network policies for multi-task and multi-robot transfer. In IEEE International Conference on Robotics and Automation, 2016. [10] J. Devlin, J. Uesato, S. Bhupatiraju, R. Singh, A. Mohamed, and P. Kohli. RobustFill: Neural program learning under noisy I/O. In International Conference on Machine Learning, 2017. [11] Y. Duan, M. Andrychowicz, B. C. Stadie, J. Ho, J. Schneider, I. Sutskever, P. Abbeel, and W. Zaremba. One-shot imitation learning. Technical report, OpenAI, 2017. [12] J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl. Neural message passing for quantum chemistry. Technical report, Google Brain, 2017. [13] A. Graves, G. Wayne, M. Reynolds, T. Harley, I. Danihelka, A. Grabska-Barwińska, S. G. Colmenarejo, E. Grefenstette, T. Ramalho, J. Agapiou, et al. Hybrid computing using a neural network with dynamic external memory. Nature, 538(7626):471–476, 2016. [14] E. Grefenstette, K. M. Hermann, M. Suleyman, and P. Blunsom. Learning to transduce with unbounded memory. In Advances in Neural Information Processing Systems, pages 1828–1836, 2015. [15] R. Hu, J. Andreas, M. Rohrbach, T. Darrell, and K. Saenko. Learning to reason: End-to-end module networks for visual question answering. Technical report, University of California, Berkeley, 2017. [16] J. Johnson, B. Hariharan, L. van der Maaten, J. Hoffman, L. Fei-Fei, C. L. Zitnick, and R. Girshick. Inferring and executing programs for visual reasoning. Technical report, Stanford, 2017. [17] Ł. Kaiser and I. Sutskever. Neural GPUs learn algorithms. In International Conference on Learning Representations, 2015. [18] B. Knoll and N. de Freitas. A machine learning perspective on predictive coding with PAQ8. In Data Compression Conference (DCC), 2012, pages 377–386. IEEE, 2012. [19] T. D. Kulkarni, K. Narasimhan, A. Saeedi, and J. Tenenbaum. Hierarchical deep reinforcement learning: Integrating temporal abstraction and intrinsic motivation. In Advances in Neural Information Processing Systems, pages 3675–3683, 2016. [20] B. M. Lake, T. D. Ullman, J. B. Tenenbaum, , and S. J. Gershman. Building machines that learn and think like people. Behavioral and Brain Sciences, in press. [21] T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y. Tassa, D. Silver, and D. Wierstra. Continuous control with deep reinforcement learning. In International Conference on Learning Representations, 2015. [22] W. Ling, D. Yogatama, C. Dyer, and P. Blunsom. Program induction by rationale generation: Learning to solve and explain algebraic word problems. arXiv preprint 1705.04146, 2017. [23] M. L. Littman, U. Topcu, J. Fu, C. Isbell, M. Wen, and J. MacGlashan. Environment-independent task specifications via GLTL. Technical report, Brown University, 2017. [24] M. V. Mahoney. Adaptive weighing of context models for lossless data compression. Technical report, CS-2005-16, 2005. [25] T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, and J. Dean. Distributed representations of words and phrases and their compositionality. In Advances in neural information processing systems, pages 3111–3119, 2013. 9 [26] I. Misra, A. Gupta, and M. Hebert. From Red Wine to Red Tomato: Composition with Context. In Computer Vision and Pattern Recognition, 2017. [27] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski, et al. Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 2015. [28] A. Neelakantan, Q. V. Le, M. Abadi, A. McCallum, and D. Amodei. Learning a natural language interface with neural programmer. In International Conference on Learning Representations, 2017. [29] A. Neelakantan, Q. V. Le, and I. Sutskever. Neural programmer: Inducing latent programs with gradient descent. In International Conference on Learning Representations, 2016. [30] S. Reed and N. De Freitas. Neural programmer-interpreters. International Conference on Learning Representations, 2016. [31] D. Silver, G. Lever, N. Heess, T. Degris, D. Wierstra, and M. Riedmiller. Deterministic policy gradient algorithms. In ICML, 2014. [32] R. Socher, B. Huval, C. D. Manning, and A. Y. Ng. Semantic compositionality through recursive matrixvector spaces. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 1201–1211, 2012. [33] R. Socher, A. Perelygin, J. Y. Wu, J. Chuang, C. D. Manning, A. Y. Ng, C. Potts, et al. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the conference on empirical methods in natural language processing, 2013. [34] E. Todorov, T. Erez, and Y. Tassa. Mujoco: A physics engine for model-based control. In International Conference on Intelligent Robots and Systems, 2012. [35] D. Tran, M. D. Hoffman, R. A. Saurous, E. Brevdo, K. Murphy, and D. M. Blei. Deep probabilistic programming. arXiv preprint 1701.03757, 2017. [36] D. Yogatama, P. Blunsom, C. Dyer, E. Grefenstette, and W. Ling. Learning to compose words into sentences with reinforcement learning. arXiv preprint 1611.09100, 2016. [37] H. Yu, H. Zhang, and W. Xu. A deep compositional framework for human-like language acquisition in virtual environment. Technical report, Baidu Research, 2017. [38] W. Zaremba, T. Mikolov, A. Joulin, and R. Fergus. Learning simple algorithms from examples. In International Conference on Machine Learning, pages 421–429, 2016. [39] W. Zaremba and I. Sutskever. Learning to execute. arXiv preprint 1410.4615, 2014. 10 A Model and training details Our agents are trained using DPG [31, 21], where both the actor and critic are programmable networks. The full architecture of the actor and critic are shown in Figure A. The actor and critic share the same programmable structure (including the vocabulary of properties), but they do not share weights. ^ _ p ⌦0 ⌦ h + a a Q The relationship between Ω, Ω0 , Φ is explained in the main text, as is the derivation of the relevant objects vector p. In both the actor and critic the vector h is produced by taking a weighted sum over the columns of Ω0 . Using ωi0 to denote these columns, we can write h as X h= pi ωi0 i The motivation for weighting the columns by p here is the same as for incorporating p into the message passing weights in Equation 4; we want h to include only information about relevant objects, and the role of p is precisely to identify these objects. Reducing over the columns of Ω0 fixes the size of h to be independent of the number of objects. From this point the architectures of the actor and critic diverge. Recall that there are two networks here that do not share weights, so there are in fact two different h vectors to consider. We distinguish between the activations at h in the actor and critic by using ha to denote h produced in the actor and hc to denote h produced in the critic. The actor produces an action from ha using a single linear layer, followed by a tanh to bound the range of the actions a = tanh(Linear(tanh(ha ))) . The computation in the critic is slightly more complex. Although hc contains information about the observation, it does not contain any information about the action, which the critic requires. We combine the action with hc by passing it through a single linear layer which is then added to hc Q(Ω, a) = Linear(tanh(hc + Linear(a))) . No final activation function is applied to the critic in order to allow its outputs to take unbounded values. 11 A.1 Building the observation matrix The observations consumed by our agent are collected into the columns of Ω. The matrix Ω has one column for each object in the environment, where objects include all of the blocks on the table and also the hand of the robot arm. As discussed in Section 4, each object is described by its 3d position and 4d orientation, represented in the coordinate frame of the hand. Each block also has a shape and a color which are represented to the agent using 1-hot vectors. There are 4 possible colors and 5 possible shapes for a total of 9 property features (and 16 total features) for each object. We also provide the joint positions of the arm as observations (encoded as sin and cos of the joint angle). The arm has 6 joints, for a total of 18 features to represent the arm. The full observation matrix Ω is built by appending the arm positions to each (object) column, which effectively represents each object in a “body pose relative” way. This means that in an environment with four blocks we have an observation of shape 34 × 5 (16 + 18 features per object for each of the 4 + 1 objects). In addition to the above, we also provide the agent with the index of the hand in Ω, so it does not need to learn to detect its own body. A.2 Training parameters Each reaching episode lasts 10 seconds and controls are issued at 0.1 second intervals, for a total of 101 observations per episode (including the initial observation at t = 0). Our agents are trained for either 2m or 5m frames (in the 2x2 or 3x3 environments, respectively). We collect experience into a replay buffer with a capacity of 1m frames, which we use for training. After each action we sample a batch of 64 transitions from the replay buffer which is used to update both the actor and the critic. We use a target network [27] for both the actor and the critic to stabilize training, and we copy the latest parameters to the target networks after each episode (we do not use the soft updates of Lillicrap et al. [21]). All agents are trained with a discount factor of γ = 0.99, and when computing updates for the actor we clip each element of the gradient from the critic to the range [−1, 1]. During training we use exploration noise that follows a Ornstein–Uhlenbeck process with standard deviation 0.3 and a damping factor of 1.0. The programmable networks use a hidden layer size of 150 units in both the actor and the critic. The neighborhood attention operation using a context and query size (ci and qi ) with 64 units. Detectors are implemented as logistic regressions on object features. action Q The structure of the canonical actor and critic networks is shown in Figure 6. In our experiments these networks have 400 units in each hidden layer and use tanh acaction tivations throughout. Their weights were initialized following the strategy of Lillicrap et al. [21]. When training the canonical networks we used the same environobservation observation ment parameters as for the programmable networks. Settings for discounting, target networks and exploration noise were also Figure 6: Structure of the canonical actor and critic netthe same. The canonical networks accept works. Arrows represent linear layers; tanh activations vectors of observations (rather than the ma- are used throughout both networks. trices we feed to the Programmable Networks). We obtain observations for the canonical networks by flattening the Ω matrix into a vector, ensuring that objects with the same properties always appear at the same locations in the flattened vector. 12 B Referencing objects by exclusion Once you eliminate the impossible, whatever remains, no matter how improbable, must be the truth. — Sherlock Holmes Referencing objects by properties they do not have (e.g. “the cube that is not red”) works by exclusion. To reach for an object without a property we can simply write a program that expresses this. The program NEAR(HAND, AND(NOT(RED), CUBE)), directs the agent to reach for the cube that is not red. This method of referencing the absence of properties highlights yet another type of generalization that is taking place. We specified three logical operations in Equation 3, and, or and not; however, training programs are all of the form NEAR(HAND, AND(shape, color)), which do not make use of the not operation. Nonetheless, agents are still capable of executing programs that contain negations. The reason this works is that the operations in Equation 3 are compatible. DeMorgan’s laws require that negation interact with and and or in a particular way, and the rules of classical logic require that these laws hold. The definition of not we have chosen satisfies these relations, and this is enough for the not operation to acquire negation semantics in our networks. Referencing novel colors and shapes works in a similar way. For example let us say we have a vocabulary of five colors (the procedure for referencing novel shapes is completely analogous), but only three of them have appeared in the training data. We can label the colors [RED, GREEN, BLUE, A, B], where we use generic names like a A and B to stand in for color terms that have never been used (and so, in principle, could be anything). In this case we can express the concept of “novel color” in two ways. The first is an exclusive expression, NOT(OR(RED, BLUE, GREEN)), which says “not any of the colors that have appeared,” and the second is an inclusive expression, OR(A, B), which says “any of the colors that have not appeared.” In practice we have found that combining both methods OR(NOT(OR(RED, BLUE, GREEN)), OR(A, B)) (5) to give the best performance, since it leverages our model’s assumption that every object has exactly one color (i.e. the soft membership values for all color sets must sum to 1). Using the technique of Equation 5 we can write a program to reach for the block with a new shape and a new color as NEAR( HAND, AND( OR(NOT(OR(RED, BLUE, GREEN)), OR(A, B)), OR(NOT(OR(CUBE, SPHERE, CYLINDER)), C) ) ) The programs referencing novel objects are much more complex than the training programs, but agents perform much better than chance in these cases as well. 13 C Detector output visualizations The figures show an environment state along with the corresponding Φ (transposed, so objects are in rows). Columns corresponding to different properties have been annotated with their corresponding color and shape terms. Generic names (A, B, C) indicate un-grounded concepts. C.1 Known properties Property identification is not always perfect. 14 C.2 Behavior with previously unseen properties The agent has never seen capsules, or any magenta object. 15 

9cs.NE

HAND SEGMENTATION FOR HAND-OBJECT INTERACTION FROM DEPTH MAP Byeongkeun Kang? Kar-Han Tan† Nan Jiang? Hung-Shuo Tai† Daniel Tretter‡ Truong Nguyen? arXiv:1603.02345v3 [cs.CV] 10 Jan 2018 ? Department of Electrical and Computer Engineering, UC San Diego, La Jolla, CA 92093 USA † NovuMind Inc., Santa Clara, CA 95054 USA ‡ Hewlett-Packard, Inc., Palo Alto, CA 94304 USA ABSTRACT Hand segmentation for hand-object interaction is a necessary preprocessing step in many applications such as augmented reality, medical application, and human-robot interaction. However, typical methods are based on color information which is not robust to objects with skin color, skin pigment difference, and light condition variations. Thus, we propose hand segmentation method for hand-object interaction using only a depth map. It is challenging because of the small depth difference between a hand and objects during an interaction. To overcome this challenge, we propose the two-stage random decision forest (RDF) method consisting of detecting hands and segmenting hands. To validate the proposed method, we demonstrate results on the publicly available dataset of hand segmentation for hand-object interaction. The proposed method achieves high accuracy in short processing time comparing to the other state-of-the-art methods. Index Terms— Hand segmentation, human-machine interaction, random decision forest, depth map 1. INTRODUCTION Recently, with the expansion of virtual reality (VR), augmented reality (AR), robotics, and intelligent vehicles, the development of new interaction technologies has become unavoidable since these applications require more natural interaction methods rather than input devices. For these applications, many researches have been conducted such as gesture recognition and hand pose estimation. However, most technologies focus on understanding interactions which do not involve touching or handling any real world objects although understanding interactions with objects is important in many applications. We believe that this is because hand segmentation is much more difficult in hand-object interaction. Thus, we present a framework of hand segmentation for hand-object interaction. This work is supported in part by NSF grant IIS-1522125. 1.1. Related work Hand segmentation has been studied for many applications such as hand pose estimation [1–6], hand tracking [7–9], and gesture/sign/grasp recognition [10, 11]. In color imagebased methods, skin color-based method has been popular [10, 12–16]. For hand-object interaction, Oikonomidis et al. and Romero et al. segmented hands by thresholding skin color in HSV space [4, 5, 7, 8]. Wang et al. processed hand segmentation using a learned probabilistic model where the model is constructed from the color histogram of the first frame [6]. Tzionas et al. applied skin color-based segmentation using the Gaussian mixture model [17]. However, skin color-based segmentation has limitations in interacting with objects in skin color, segmenting from other body parts, skin pigment difference, and light condition variations. An alternative method is wearing a specific color glove [18]. For depth map-based methods, popular methods are using a wrist band [3, 9, 11] or using random decision forest (RDF) [1,2,19]. Although the method using a black wristband is uncomplicated and effective, it is inconvenient. Moreover, the method cannot segment hands from objects during handobject interaction since it processes segmentation by finding connected components. Tompson et al. [1] and Sharp et al. [2] proposed the RDF-based methods based on [19]. Although the purposes of the methods are slightly different comparing to the proposed method, the methods are the most relevant methods. In this paper, we propose the hand segmentation method for hand-object interaction using only a depth map to avoid the limitations of skin color-based methods. We present the two-stage RDF method to achieve high accuracy efficiently. 2. METHOD We propose two-stage RDF for hand segmentation for handobject interaction. In our two-stage RDF, the first RDF detects hands by processing the RDF on an entire depth map. Then, the second RDF segments hands in pixel-level by applying the RDF in the detected region. This cascaded architecture is designed for the second RDF to focus on the segmentation of hands from objects and close body parts such as an arm. Fig. 1. Random decision forest. Red, black, and green circles represent root nodes, split nodes, and leaf nodes, respectively. RDF consists of a collection of decision trees as shown in Fig. 1. Each decision tree is composed of a root node, splitting nodes, and leaf nodes. Given an input data at the root node, it is classified to child nodes based on the split function at each splitting node until it reaches a leaf node. In this paper, the input data is the location of each pixel on a depth map. The split function uses the feature of the depth difference between two relative points on the depth map in [19]. At a leaf node, a conditional probability distribution is learned in a training stage, and the learned probability is used in a testing stage. For more details about RDF, we refer the readers to [20–22]. Fig. 2. Detection of hands using the RDF in the first stage. designed to evalutate the separation using the learned offset vectors and criteria as follows: X X |xc | L(x, D, u, v, θ) = − p(h|c) log p(h|c) |x| c∈{l,r} h∈{0,1} (3) where c and h are indexes for child nodes {l, r} and for classes, respectively; |xc | denotes the number of data points in the c child node; p(h|c) is the estimated probability of being the class h at the child node c. To learn offsets and a criteria, the algorithm randomly generates possible candidates and selects the candidate with a minimum loss L(·) as follows: (u, v, θ) = argmin L(x, D, u, v, θ). 2.1. Training (4) (u,v,θ) Given a training dataset D, the algorithm randomly selects a set Di of depth maps D and then randomly samples a set of data points x in the region of interest (ROI) on the selected depth maps D. The ROI is the entire region of the depth maps in the first stage. It is the detected regions using the first RDF in the second stage (see Fig. 2). The sampled set of data points x and the corresponding depth maps D are inputs to the training of a decision tree. Using the inputs (x, D), the algorithm learns a split function at each splitting node and a conditional probability distribution at each leaf node. First, learning the split function includes learning a feature f (·) and a criteria θ. We use the feature f (·) of the depth difference between two relative points {x + u/D x , x + v/D x } in [19] as follows: f (x, D, u, v) = Dx+u/Dx − Dx+v/Dx (1) where D x denotes the depth at a pixel x on a depth map D; u ∈ R2 and v ∈ R2 represent offset vectors for each relative point. Then, the criteria θ decides to split the data x to the left child or the right child. f (x, D, u, v) ≶ θ. (2) Thus, the algorithm learns two offset vectors (u, v) and a criteria θ at each splitting node. Since the goal is separating the data points x of different classes to different child nodes, the objective function is Learning a split function at each splitting node is repeated until the node satisfies the condition for a leaf node. The condition is based on (1) the maximum depth of the tree, (2) the probability distribution p(h|c), and (3) the amount of training data |x| at the node. Specifically, it avoids too many splitting nodes by limiting the maximum depth of the tree and by terminating if the child node has a high probability for a class or if the amount of remaining training data is too small. At each leaf node, the algorithm stores the conditional probability p(h|l) (probability of being each class h given reaching the node l) for the prediction in a testing stage. 2.2. Testing Using the learned RDF, the algorithm predicts the probability of being a class for a new data x. The new data is classified to child nodes using the learned split function at each splitting node until it reaches a leaf node. At the leaf node l, the learned conditional probability pT (h|l) is loaded. These steps are repeated for entire trees T in the forest T . Then, the probabilities are averaged to predict the probability p(h|x) of being a class h for the new data x. p(h|x) = 1 X pT (h|l) |T | T ∈T where |T | is the number of trees in the learned forest T . (5) (a) (b) Fig. 3. Scores depending on the decision boundary on the validation dataset. (a) Score of the RDF in the first stage. (b) Score of the two-stage RDF with filtering in Section 2.3. In the first stage, the first RDF is applied on an entire depth map to compute a probability map. Then, the probability map is used to detect hands as shown in Fig. 2. In the second stage, the second RDF processes the data points in the detected regions to predict the probability of being each class. The proposed two-stage RDF improves both accuracy and efficiency by focusing on each task in each stage. Decision boundaries are exhaustively searched with the step size of 0.01 using the predicted probability maps of the validation dataset as shown in Fig. 3. Although the most typical boundary is 0.5 for a probability map, we found that it is not the best parameter. The selected boundaries are shown in Table 1. Fig. 4. Analysis of accuracy and efficiency. 3. EXPERIMENTAL EVALUATIONS 3.1. Dataset We collected a new dataset1 using Microsoft Kinect v2 [26]. The newly collected dataset consists of 27,525 pairs of depth maps and ground truth labels from 6 people (3 males and 3 females) interacting with 21 different objects. Also, the dataset includes the cases of one hand and both hands in a scene. The dataset is separated into 19,470 pairs for training, 2,706 pairs for validation, and 5,349 pairs for testing, respectively. 2.3. Modified bilateral filter 3.2. Results Before classifying a data x to a class h, modified bilateral filter is applied to the predicted probability p(h|x) to make the probability more robust. Since the probability p(h|x) is predicted for each pixel independently, the probability is stabilized by averaging the probabilities of the data points in close distance and similar intensity on the depth map. Unlike typical bilateral filter whose weights are based on the input image (in this case, the probability map) [23], the weights in the modified bilateral filter are based on a separate image, the depth map. The filtering is defined as follows: The proposed method is analyzed by demonstrating the results on the dataset in Section 3.1. For the quantitative comparison of accuracy, we measure F1 score, precision, and recall as follows: tp tp precision = , recall = tp + fp tp + fn (7) precision × recall F1 = 2 × precision + recall 1 X pe(h|x) = gr (|D xi −D x |)gs (kxi −xk)p(h|xi ) (6) w xi ∈Ω where Ω is the set of pixels within the filter’s radius and the pre-defined depth difference; w is the normalization term, P w = xi ∈Ω gr (|D xi − D x |)gs (kxi − xk); gr (·) and gs (·) are the Gaussian functions for the depth difference and for the spatial distance from the data point x, respectively. gr (r) = s2 r2 exp(− 2σ 2 ); gs (s) = exp(− 2σ 2 ). The parameters in the filr s ter were selected based on the experiments using validation dataset. The selected parameters are as follows: the maximum depth difference to be considered is 400mm. Both standard deviations (σr and σs ) are 100. where tp, fp, and fn represent true positive, false positive, and false negative, respectively. For the comparison of efficiency, we measure the processing time using a machine with Intel i7-4790K CPU and Nvidia GeForce GTX 770. The proposed method is compared with the RDF-based method in [1, 19] and the fully convolutional networks (FCN) in [24, 25] using only a depth map. The proposed method is not compared with color-based methods since the characteristics of depth sensors and color imaging sensors are quite different. For example, a captured depth map using a depth sensor does not vary depending on light condition. However, a captured color image varies a lot depending on light condition. Thus, choosing the capturing environment affects the 1 https://github.com/byeongkeun-kang/HOI-dataset Table 1. Quantitative comparison. The two boundaries for the proposed method are for each stage. Method Score Processing time Method Boundary (ms) F1 score Filter Precision Recall RDF [1, 19] RDF [1, 19] + Proposed in Sec. 2.2 FCN-32s [24, 25] FCN-16s [24, 25] FCN-8s [24, 25] Proposed method (a) 0.50 0.78 0.50, 0.50 0.50, 0.52 0.50, 0.52 11 × 11 (b) (c) 38.1 54.5 70.0 68.0 70.4 59.2 60.8 62.9 91.2 72.7 68.6 72.2 74.4 77.4 75.1 75.6 53.7 62.3 69.3 70.1 72.3 67.1 67.2 68.7 (d) 6.7 6.7 376 376 377 8.9 8.9 10.7 (e) Fig. 5. Visual comparison. (a) Ground truth label. (b) Result using RDF [1, 19]. (c) Result using RDF [1, 19] with the proposed method in Section 2.2. (d) Result using FCN-8s [24, 25]. (e) Result using the proposed method. The results and ground truth label are visualized using different color channels for better visualization. comparison of results using depth maps and color images. Hence, we only compare the proposed method with the stateof-the-art methods which can process using only depth maps. Table 1 and Fig. 5 show quantitative results and visual results. The scores in Table 1 are scaled by a factor of 100. The quantitative results show that the proposed method achieves about 25% and 8% relative improvements in F1 score comparing to the RDF-based methods [1, 19] and its combination with the proposed method in Section 2.2, respectively. Comparing to the deep learning-based methods [24, 25], the proposed method achieves about 7% lower accuracy, but processes in about 42 times shorter processing time. Thus, deep learning-based methods can not be used in real-time applica- tions. Fig. 4 shows the comparison of methods in accuracy and efficiency. The proposed method achieves high accuracy in short processing time. 4. CONCLUSION In this paper, we present two-stage RDF method for hand segmentation for hand-object interaction using only a depth map. The two stages consist of detecting the region of interest and segmenting hands. The proposed method achieves high accuracy in short processing time comparing to the state-of-the-art methods. 5. REFERENCES [1] J. Tompson, M. Stein, Y. Lecun, and K. Perlin, “Realtime continuous pose recovery of human hands using convolutional networks,” ACM Trans. Graph., 2014. [2] T. Sharp, C. Keskin, D. Robertson, J. Taylor, J. Shotton, D. Kim, C. Rhemann, I. Leichter, A. Vinnikov, Y. Wei, D. Freedman, P. Kohli, E. Krupka, A. Fitzgibbon, and S. Izadi, “Accurate, robust, and flexible real-time hand tracking,” in Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems, 2015. [3] C. Qian, X. Sun, Y. Wei, X. Tang, and J. Sun, “Realtime and robust hand tracking from depth,” in Computer Vision and Pattern Recognition (CVPR), 2014 IEEE Conference on, June 2014. [4] J. Romero, H. Kjellstrom, and D. Kragic, “Hands in action: real-time 3d reconstruction of hands in interaction with objects,” in Robotics and Automation (ICRA), 2010 IEEE International Conference on, May 2010. [5] J. Romero, H. Kjellstrom, C. H. Ek, and D. Kragic, “Non-parametric hand pose estimation with object context,” Image Vision Comput., vol. 31, no. 8, Aug. 2013. [6] Y. Wang, J. Min, J. Zhang, Y. Liu, F. Xu, Q. Dai, and J. Chai, “Video-based hand manipulation capture through composite motion control,” ACM Trans. Graph., vol. 32, no. 4, July 2013. [7] I. Oikonomidis, N. Kyriazis, and A.A. Argyros, “Full dof tracking of a hand interacting with an object by modeling occlusions and physical constraints,” in Computer Vision (ICCV), 2011 IEEE International Conference on, 2011. [8] A. A. Argyros and M. I. A. Lourakis, “Real-time tracking of multiple skin-colored objects with a possibly moving camera,” in Computer Vision - ECCV, 2004. [9] B. Kang, Y. Lee, and T. Nguyen, “Efficient hand articulations tracking using adaptive hand model and depth map,” in Advances in Visual Computing, Dec. 2015. [10] M. Cai, K.M. Kitani, and Y. Sato, “A scalable approach for understanding the visual structures of hand grasps,” in Robotics and Automation (ICRA), 2015 IEEE International Conference on, May 2015. [11] B. Kang, S. Tripathi, and T. Nguyen, “Real-time sign language fingerspelling recognition using convolutional neural networks from depth map,” in Pattern Recognition, 2015 3rd IAPR Asian Conference on, Nov. 2015. [12] M. J. Jones and J. M. Rehg, “Statistical color models with application to skin detection,” International Journal of Computer Vision, vol. 46, Jan. 2002. [13] R. Khan, A. Hanbury, J. Stttinger, and A. Bais, “Color based skin classification,” Pattern Recognition Letters, vol. 33, Jan. 2012. [14] C. Li and K.M. Kitani, “Pixel-level hand detection in ego-centric videos,” in Computer Vision and Pattern Recognition, 2013 IEEE Conference on, June 2013. [15] S. L. Phung, A. Bouzerdoum, and D. Chai, “Skin segmentation using color pixel classification: analysis and comparison,” Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 27, 2005. [16] P. Kakumanu, S. Makrogiannis, and N. Bourbakis, “A survey of skin-color modeling and detection methods,” Pattern Recognition, vol. 40, Mar. 2007. [17] D. Tzionas and J. Gall, “3d object reconstruction from hand-object interactions,” in International Conference on Computer Vision (ICCV), Dec. 2015. [18] R. Y. Wang and J. Popović, “Real-time hand-tracking with a color glove,” ACM Trans. Graph., vol. 28, 2009. [19] J. Shotton, A. Fitzgibbon, M. Cook, T. Sharp, M. Finocchio, R. Moore, A. Kipman, and A. Blake, “Real-time human pose recognition in parts from single depth images,” in Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, June 2011. [20] A. Criminisi and J. Shotton, “Decision forests for computer vision and medical image analysis,” Advances in Computer Vision and Pattern Recognition, 2013. [21] L. Breiman, “Random forests,” Machine Learning, vol. 45, no. 1, 2001. [22] T. Ho, “Random decision forests,” in Proceedings of the Third International Conference on Document Analysis and Recognition, 1995. [23] C. Tomasi and R. Manduchi, “Bilateral filtering for gray and color images,” in Computer Vision, Sixth International Conference on, Jan. 1998. [24] J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks for semantic segmentation,” in The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2015. [25] E. Shelhamer, J. Long, and T. Darrell, “Fully convolutional networks for semantic segmentation,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2016. [26] B. Kang, Y. Lee, and T. Q. Nguyen, “Depth adaptive deep neural network for semantic segmentation,” IEEE Transactions on Multimedia, 2018. 

1cs.CV

A Spatio-Temporal Representation for the Orienteering Problem with Time-Varying Profits arXiv:1611.08037v2 [cs.AI] 2 Jul 2017 Zhibei Ma, Kai Yin, Lantao Liu, Gaurav S. Sukhatme Abstract— We consider an orienteering problem (OP) where an agent needs to visit a series (possibly a subset) of depots, from which the maximal accumulated profits are desired within given limited time budget. Different from most existing works where the profits are assumed to be static, in this work we investigate a variant that has arbitrary time-dependent profits. Specifically, the profits to be collected change over time and they follow different (e.g., independent) time-varying functions. The problem is of inherent nonlinearity and difficult to solve by existing methods. To tackle the challenge, we present a simple and effective framework that incorporates time-variations into the fundamental planning process. Specifically, we propose a deterministic spatio-temporal representation where both spatial description and temporal logic are unified into one routing topology. By employing existing basic sorting and searching algorithms, the routing solutions can be computed in an extremely efficient way. The proposed method is easy to implement and extensive numerical results show that our approach is time efficient and generates near-optimal solutions. I. I NTRODUCTION The rapid progress of smart vehicle technologies allows us to envision that, in the future autonomous vehicles are able to carry out various tasks with little or even no human effort. We are interested in designing an efficient routing method to navigate a vehicle (agent) among a number of known and fixed depots, where each depot has some profit (e.g., score, benefit, utility, load) to be collected. If the travel time is limited, it is likely that the agent is not able to traverse all depots due to the limited time budget. This variant of routing problems is called the orienteering problem (OP) [11]. In a nutshell, an OP aims to find a tour traversing a subset of depots so that the accumulated profit collected from those traversed depots are maximized. Different from many existing vehicle routing problems which focus on analyzing path properties in the static context (i.e., unchanging environment or topology with stationary cost metrics) [17], [12], in this work we are intrigued to investigate a time-varying variant of OPs, i.e., each depot has a time-varying profit. Here are a few motivational examples: • As illustrated in Fig. 1, an autonomous truck needs to pick up goods from a number of fixed depots where manufacturing factories are located. The goods are produced consistently and accumulated as time goes by — so they are time-varying. The growth rate of goods at different factories may be nonidentical: larger factories Z. Ma, L. Liu and G.S. Sukhatme are with the Department of Computer Science at the University of Southern California, Los Angeles, CA 90089, USA. {zhibeima, lantao.liu, gaurav}@usc.edu. K. Yin is with HomeAway, Inc., yinkai1000@gmail.com. Fig. 1. An autonomous vehicle needs to route through depots with timevarying profits (loads). The goal is to maximize overall collected profits within given limited time. The task can be persistent. grow faster and smaller ones grow slower. Assume the truck has sufficiently large capacity and it empties all goods from a depot when it arrives there. The objective is to find a route so that the truck will load the most goods in a given time window. • In environmental monitoring, autonomous robots are deployed to collect environment data in order to estimate an underlying environment state. However, the environment attributes at different locations can be timevarying (e.g., dissolved chemical compounds and algae blooms in the water vary both spatially and temporally.) An important objective of environment monitoring is to plan so-called “informative paths” [5] that navigate the robots to acquire data from those most information-rich spots which best help estimate the environment. In this paper, we present our initial study for the timevarying OP. We start our analysis from the single agent planning case, and attempt to establish a new framework that is different from traditional modeling and solving routines. We will show that, a big challenge of the time-varying OP (even for single agent case) lies in the requirement of a special treatment in an extra dimension — the time dimension, which inevitably introduces extra complexity as we need to model, predict, and integrate future dynamics. II. R ELATED W ORKS The routing problems have been well studied in many research domains including operations research, theoretical computer science, and transportation systems [3], [4], [7]. Typical routing problems involve incorporation of constraints expressed from the nature of the target problems, which also narrow the space of searching for solutions [19]. We are interested in a variant of routing problems called the orienteering problems (OPs) [11], [22]. An OP considers both travel cost (e.g., travel time) and scores collected along the travel. The goal of the OP is to determine which subset of vertices to visit and in which order so that the collected scores within a given period is maximized. The OPs integrate characteristics of both knapsack problems (KPs) [15] and travelling salesman problems (TSPs) [19], and OPs are NPhard as well. In contrast to the TSP, not all vertices of an OP need to be visited due to the limited time budget. During the past few decades, several variants of OPs have been studied, such as time-dependent OPs, Team OPs, (Team) OPs with Time Windows and OPs with stochastic profits. Recent survey papers [22], [12] have profoundly discussed state-of-the-art techniques of these variants as well as their applications. While OPs can be formulated as mixed integer programs (MIPs), the problem size typically is too large to directly use commercial solvers. A wide range of decomposition methods such as branch-and-price algorithms have been developed so that a large-scale MIP can be decomposed into smaller problems (e.g., a master problem and a series of sub-problems) which can then be iteratively solved by commercial solvers [13], [16]. In order to reduce the heavy computational burden in decomposition methods, heuristics and metaheuristics have been extensively studied, typically including tabu-based or neighborhood search based procedures [6], [21], [24]. Although in general OPs have been well researched, the time-dependent OPs (TOPs) have received relatively less attention comparing with other variants [10], [12]. Even so, most of existing TOPs discuss the time-varying properties that are associated with the real travel time between pairwise nodes, and assume that travel time between two nodes depends on the departure time at the first (or an earlier) node [18], [9], [8], [23]. Very rarely we could find the works that discuss about time-varying scores of OPs. One work that share certain similarity with this proposed problem is [2], where multiple vehicles need to serve a number of clients and the profit of each client follows a decreasing function of time. The work analyzed a lower bound and upper bound based on a classic MIP formulation. Instead of employing conventional techniques such as the column generation approach used in [2], in this work we present our first study that models and tackles the problem from a different perspective: we start from establishing a representation built from the spatial and temporal constraints, so that the time dependence attribute is transferred from the bulk MIP to a separate and intuitive representation. With that, fast approximate OP solutions can be found by employing and extending order-sensitive topological algorithms. III. P ROBLEM D ESCRIPTION AND F ORMULATION A routing problem can be represented with a graph G = (V, E), where V is the set of vertices and E denotes the set of edges. Let us denote the number of vertices as |V | = n+1, and suppose every edge takes time to traverse. Although the travel time, in many situations, depends on the states and properties of the network such as congestion and capacity, we assume that the travel time τij between two vertices vi and vj is time-invariant for simplicity∗ . We associate each ∗ It is relatively straightforward to incorporate time-dependent travel times in our proposed model described in the following section. vertex vi with a time-varying value, called profit, which is denoted by fi (t) ≥ 0 at time t ≥ 0. Here fi (·) is of arbitrary nonlinear function form. We assume that function fi (·) is known or can be predicted or approximated. Suppose that the agent (vehicle) starting from a dummy node v0 at time 0 travels across a subset of vertices on the graph G. When the agent visits vertex vi at time ti , it will collect the profit fi (ti ). The remaining profit at the vertex vi right after the agent leaves becomes 0 and accumulates again. Additionally, the profit at v0 is assumed to be 0, i.e., f0 (t) = 0. Assuming the agent visits each vertex once, the objective is to determine the order of a subset of vertices to visit so that total profits collected by the agent is maximized within a given planning period T . Note that we do not assign a specific destination to the agent, as the problem with a fixed end vertex is a special case for our problem. Let xij = 1 if the agent travels from vi to vj , and 0 otherwise. Let t0 = 0. We also introduce extra variables ui with u0 = 0 to eliminate the subtours. Then the problem in this study can be formulated as the following mixed-integer program. max xij n X i=1 n X j=1 n X fi (ti ) · x0j = xik = i=0 tj 6 T · X j n X i=1 n X j=0 n X xji , (1) xi0 = 1, (2) xkj 6 1, xij , ∀k 6= 0, ∀j, (3) (4) i=0 ti + xij (τij + T ) 6 T + tj , 1 6 ui 6 n, ui − uj + 1 6 n(1 − xij ), xij ∈ {0, 1}, ∀i, j, ∀i 6= 0, (5) (6) ∀i 6= 0, j 6= 0, ∀i, j. (7) (8) Eq. (1) is to maximize the total collected profits. Constraint (2) guarantees that the path starts and ends at dummy node v0 . Note that our setup allows a path to end at any node by discarding the inward edge. Constraints (3) ensure the connectivity of the path and guarantee that every vertex is visited at most once. The limited time budget is ensured by Constraint (4), while Constraint (5) determines the visiting time ti along the path. Constraints (6) and (7), so called Miller-Tucker-Zemlin (MTZ) formulation in the traveling sales problem[19], are used to prevent subtours. IV. A S PATIO -T EMPORAL R EPRESENTATION It appears a daunting task to solve the problem (1)-(8) due to the complexity of objective function. Most of existing literature in OPs either lumps all constraints together and solves it by conventional solvers of MIPs [20], [22], or decouples the constraints into master-subproblem modules such as column generation [2], or use certain heuristics such as center-of-gravity heuristic [11]. In contrast, we (a) (b) (c) (d) Fig. 2. Illustrations of spatio-temporal representation. (a) Graph G = (V, E) constructed on a x-y plane. Numbers on vertices are the vertex IDs, and numbers on edges denote required travel time costs; (b) Spatio-temporal graph of (a). In this example each vertex has 5 time units along the (z-axis); (c)(d) Another example with 7 vertices and 8 time windows. desire to develop a framework starting from a constraintincluded representation that is intuitive to understand, easy to implement from scratch, and flexible to modify and extend. In this section, we present a means for embedding constraints into a spatio-temporal representation built on which the original routing problem can be tackled by efficient methods (though it is still a NP-hard problem). As the problem essentially aims to determine the visiting time and order of vertices, it inspires us to incorporate a time dimension to extend the 2-dimensional graph on a spatial plane to a 3-dimensional graph (topology). To make the model implementable, the time range [0, T ] is discretized into a sequence of time intervals ∆t of equal length, and the interval ∆t is used as a time unit that specifies discrete time resolution. Thus, the travel time τij can be expressed as multiple times of the time unit. For example, τij = nij ∆t and T = nT ∆t , where nij and nT are integers. Figure 2 shows two examples that describe the basic idea. Intuitively, one can imagine that each spatial graph vertex (Fig. 2(a)) is extended to a sequence of vertices expanding in the time dimension, where each vertex is on an unique time layer corresponding to a future time moment (Fig. 2(b)). Then edges are added by concatenating vertices of different time layers, constrained by the real travel time. Formally, two vertices on different time layers form an edge if and only if the two vertices are spatially traversable and the time difference between the two time layers is exactly equal to the anticipated real travel time between the two vertices. Each vertex contains a profit and the profit is time-varying. In other words, the profits at different time layers are not the same. Therefore, the problem is equivalent to finding a path from the given start vertex at time 0 such that the path transits other vertices within time T and the total collected profits are maximized. It is worth mentioning that, the spatiotempral edges are “directed” since each edge must start from a vertex at an earlier time layer and ends at one at a later moment; Because the time is uni-directional and cannot travel backwards, it is impossible to form a loop or cycle on the spatio-temporal graph. As a result, such a spatio-temporal representation is equivalent to a vertex-weighted Directed Acyclic Graph (DAG) [18]. This allows us to conveniently develop our own routing algorithm built on many existing efficient DAG algorithms. More formally, the spatio-temporal representation is a vertex-weighted DAG, denoted by G0 = (V 0 , E 0 , W 0 ), where V 0 = V × T , T = {0, ∆t , ..., nT ∆t }; eijus is a directed 0 0 edge from vertex viu = (vi , u) to vjs = (vj , s), and 0 eijus ∈ G if and only if the edge e(vi , vj ) ∈ E in G and 0 0 ∈ W 0 for vertex viu τij = ns ∆t − nu ∆t ; the weight wiu 0 has value wiu = fi (u). Keep in mind that u = nu ∆t and a similar equation holds for s. As an example, suppose that we have a complete graph G with 3 vertices, ∆t = 1, and T = 5. In the spatiotemporal representation G0 shown in Fig. 2(b), there are 18 0 0 0 vertices v0,0 , v0,1 , ..., v2,5 , each vertex in the original graph G is duplicated at the discrete time in T = {0, 1, 2, 3, 4, 5}. If 0 0 τ01 = 3 and τ21 = 5, then vertex v0,0 is directed to v1,3 and 0 0 v2,0 is directed to v1,5 . In this example, we have 9 directed edges in total. With graph G = (V, E) and given time limit T , the spatiotemporal graph G0 = (V 0 , E 0 , W 0 ) can be constructed by Alg. 1. Assume there are n vertices in G, the time complexity of Alg. 1 is O(n2 T /∆t ) because there are three for loops each of which is associated with either n or T . Also there are n(T /∆t + 1) vertices in G0 , so the space complexity is O(nT /∆t ) . Algorithm 1: SpatioTemporalGraph (G, T, v0 , ∆t ) Input: 1: Graph G = (V, E), time limit T , start vertex v0 , time interval ∆t Output: 2: for each vertex vi in graph G do 3: for each t in time period [0, T ], time step is ∆t do 0 4: vit .id = i 0 0 5: vit .weight = fi (t), put wit into the weight set W 0 0 6: vit .position = vi .position 0 7: vit .prof it = fi (t) 0 8: vit .sum = −∞ 0 9: vit .parent = −1 10: for each vi ’s neighbor vj do 11: if t + τij is equal to t0 AND t0 <= T then 12: put edge eijtt0 into the edge set E 0 0 0 13: push vjt 0 into vector vit0 .successors 0 14: update indegree of vit . 0 15: put vit into the vertex set V 0 0 16: set the v0,0 .sum to 0 17: return graph G0 = (V 0 , E 0 , W 0 ) 0 Note: fi (t) is the time-varying profit function of vertex vi at time t. v0,0 is the start vertex in G0 , and .sum is used to store accumulated profit from prior traversal along a path. The label .parent points to the predecessor vertex. V. ROUTING A LGORITHM An important advantage of the spatio-temporal graph lies in that, the time related constraints have been incorporated into this spatio-temporal representation, so that temporal related constraints can be decoupled and eliminated from the process of optimizing the profits. Thus, the problem (1)-(8) is 0 equivalent to finding a path P from P vertex v00,0 = (v0 , 0) to 0 0 vks on G where s ≤ T such that (v0 u)∈P wiu is maximum. i Since the spatio-temporal graph is essentially a DAG, we develop our routing solution via extending classic DAG algorithms. Specifically, we found that the profit maximization can be transformed to a longest path problem by accumulating the profits collected from vertices instead of summing up edge lengths along a path. We manipulate the DAG so that vertices are sorted in a topological order along the temporal dimension, and then employ a dynamic programming paradigm to compute the maximal profit path. It is also noteworthy that, while developing a solution to the time-varying OP, we take two concerns that are related to applications into account. • Specification of a Routing Destination: Many routing problems, including classic OPs, require to specify a routing destination. The destination can be the original depot where the agent departures (e.g., a mail truck needs to return a central processing office); the destination can also be an arbitrary depot located somewhere else (e.g., a freight truck needs to pick up goods and unload them into some specified processing location that is different from the starting depot). • Persistent Task: Many long-term missions need repetitive and persistent routing, for which specifying a routing destination is not necessary or even inappropriate. For instance, in the persistent environmental monitoring task, we do not need the robot to stop at some specified location, as the robot will need to resume to next round of routing after the completion of current one. Thus, the routing destination should be computed on the fly based on the profit optimization constrained by time T , instead of being manually specified. We will show that the proposed framework works for both specified and unspecified routing destinations. A. Topological Sorting of DAG in Temporal Dimension The main purpose of topological sorting of the DAG is that, the vertices are “placed” onto different “stages” according to their temporal constraints, so that a dynamic programming structure (discussed in the following subsection) can be applied. Formally, a topological sort of a directed graph is a linear ordering of its vertices such that for every directed edge euv from vertex u to vertex v, u comes before v in ordering. A topological sort of a graph requires that the graph must be a DAG. We employ a well-known algorithm developed by Kahn [14] to sort our spatial-temporal graph G0 , with main steps shown in Alg. 2. Briefly, we first find a list of vertices that have no incoming edges (with deg − (v) = 0), and insert them into a set S. Note, at least one such vertex must exist in a non-empty graph. Then we traverse the set S. Each time we remove a vertex v from S, and add it to the tail of the list L. After removing v, the indegree of its successors should be decreased by 1. Then we insert those vertices with updated indegree equal to 0 in the set S. To analyze the time complexity, assume there are n vertices in G, so there will be n(T /∆t + 1) vertices in G0 . 0 For each vertex vit , there will be at most n−1 directed edges. 0 Therefore |E | = O(n2 T /∆t ), and the time complexity of Alg. 2 is O(|V 0 | + |E 0 |) = O(n2 T /∆t ). Algorithm 2: TopologicalSort (G0 ) 1: 2: 3: 4: 5: 6: 7: 8: 9: 10: L: an empty list that will contain the sorted elements S: a set of all nodes with no incoming edges while S is not empty do remove a vertex v from S add v to the end of L for each successor u of v do deg − (u) = deg − (u) − 1 if deg − (u) equals to 0 then insert u into S return L (a topologically sorted order) Note: G0 is a DAG, the indegree of v is denoted as deg − (v). B. Computing Maximal Profit Path We transfer the time-varying OP to a longest path problem in a DAG. The classic longest path problem is the problem of finding a simple path of maximum length in a given graph. We employ a dynamic programming structure to memorize incumbent maximal accumulated profit at each vertex of topologically sorted stages. Note that, in our problem, we need to optimize profits collected from vertices, instead of adding up length of edges. Therefore, instead of using the longest path update function between two successive stages: l[w] = l[v] + τvw , if l[v] + τvw > l[w], (9) where l[w] is the largest distance from start vertex to w, we utilize an update function: w.sum = v.sum+w.prof it, if v.sum+w.prof it > w.sum. (10) The computation of the maximal profit path is described in Alg. 3. Briefly, the topologically sorted vertices V 0 of the spatio-temporal graph G0 are used as an input. Then vertices from different stages form dynamic programming subproblems and they are updated with accumulated profits 0 0 recursively, starting from v0,0 . Here we use vit .path to store vertices along the path from start v00 to vi0 (Note, only spatial information is recorded). To prevent from forming routing subtours, we check and discard those already visited vertices before each value update. After the completion of dynamic programming, each vertex contains information of the maximal profit path that 0 routes from v0,0 to it. Since the graph G0 has incorporated the time limit T , every vertex is feasible to the time constraint. To find a maximal profit path to a specified vertex (destination) vi ∈ G, one simply needs to enumerate 0 all states vij ∈ G0 , ∀j of vertex vi and retrieve the path with the largest value. If a destination is not specified, one needs to enumerate all states of all vertices and find out the maximal one among them. The time complexity of Alg. 3 is O(n2 T /∆t ) due to its two for loops. Algorithm 3: M aximalP rof itP ath(L) vit .sum = 0, i = 0, t = 0 0 for each vertex vit in topologically sorted order L do 0 0 for each vertex vjt 0 in vit .successors do 0 if vertex vj is not in the vit .path then 0 0 0 if vit .sum + vjt 0 .prof it > vjt0 .sum then 0 0 0 vjt0 .sum = vit .sum + vjt0 .prof it 0 0 update the vjt 0 .parent to vit 0 0 vjt0 .path = vit .path + vj 0 for each vertex vit in G0 do find the maximal sum retrieve a path P by backtracking from end to start 12: return P 1: 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: Note: The label .path is a vector that stores all visited vertices from start to vi C. Main Routing Algorithm With the components descried above, the main algorithm is shown in Alg. 4. According to the analysis of each part, the overall time complexity of our algorithm is n2 T /∆t , and the space complexity is n2 T /∆t as well. Algorithm 4: Time-Varying OP Input: 1: 2D graph G = (V, E), start vertex v0 , time limit T , time interval ∆t Output: 2: construct a spatio-temporal graph G0 = SpatioT emporalGraph(G, T, v0 , ∆t ) 3: L = T opologicalSort(G0 ) 4: path = M aximalP rof itP ath(L) 5: return path Our algorithm has a polynomial time complexity given a fixed ∆t , but cannot guarantee to find the optimal solution to this NP-hard problem. The main reason for possibly reaching at sub-optimality lies in that, we added a non-subtour constraint in order to (1) prevent the path from traversing back and forth among a small subset of adjacent vertices and (2) comply with the structure of dynamic programming. Such a constraint eliminates certain feasible searching space which possibly contains the optimal solution. One may regard this as a trade-off between solution quality and practical runtime. Nevertheless, our evaluation results with extensive trials show that on average our algorithm produces near-optimal solutions. D. Discussion: Effects of Discretization The time interval ∆t plays a critical role in our framework as the total profits are actually evaluated at travel time steps equal to multiple ∆t . Yet it may be impractical to select a small enough time interval because it would result in an extremely large size of DAG. One possible implementation is to round the travel time to a value closet to some integer times of ∆t . Hence, it is important to analyze the impact of errors due to such rounding procedure. Assume that the profit function fi (t) satisfies the following Lipschitz condition: |fi (t) − fi (s)| ≤ K|t − s|, ∀t, s ∈ [0, T ], ∀i ∈ {0, ..., n}, (11) where K is a constant independent of i. Suppose that the optimal objective value for the problem (1)-(8) is z, and the optimal objective value by the proposed framework is z 0 . Then we have the result as follows. Proposition 1: Under the condition (11), we have the upper bound for the difference in two objective values: |z − z 0 | ≤ [n(n + 1)/2] · K · ∆t . (12) proof: Let P be the optimal path in the graph G for the problem (1)-(8) and let P 0 be the optimal path in the proposed spatio-temporal DAG G0 . By the construction of G0 , P must correspond to a path P̃ in the graph G0 . Let z̃ be the total profits (i.e. the value of objective (1) on this path) for the path P̃ . Since the difference between the discrete time in our framework and the original real-valued time is within ∆t , the fact that any path contains no more than n vertices and the condition (11) lead to |z − z̃| ≤ [n(n + 1)/2]K∆t . As P 0 rather than P̃ is the optimal path in G0 , we have z̃ ≤ z 0 . By combining these results, we should have z − z 0 ≤ [n(n + 1)/2]K∆t . Conversely, by construction any feasible path on the graph G0 must correspond to a feasible path on G. Using similar arguments, we have z 0 −z ≤ [n(n+1)/2]K∆t . Therefore, the result (12) holds. It is worthwhile noting that the constraint that the path in G or G0 contains no subtours is critical to the result (12). VI. R ESULTS We conducted numerical evaluations to validate the proposed algorithm. We compared our algorithm with other algorithms, all of which were implemented in C++. The experiments were performed on a system with an Intel i5 2.2GHz processor and 4GB RAM. In our experiments, vertices are randomly generated in the area [−50, 50] × [−50, 50] on the x-y plane, and each vertex has a profit (weight) following certain time-varying functions. The distance between two vertices is Euclidean distance. A. Routing Feature of “Orienteering” First, we demonstrate that our solution has the basic orienteering property: given limited time, the path always goes through those vertices with large profits. Fig. 3(a) shows an example. We purposely manipulate the profit distribution among the vertices, such that in the upper right corner, there are more vertices with larger profits. For better visualization, in the figure the size of a vertex is proportional to the profit on it. Additionally, we purposely put the starting vertex on the left and do not specify a destination. From Fig. 3(a) we (a) (b) Fig. 3. (a) Route feature of orienteering. ∆t is 1, vertex number is 200, T is 200 and the start vertex is at (-49,0); (b) Route behavior under timevarying profit function, ∆t is 1, vertex number is 200, T is 200. The start vertex v0 is at (-49,0). In all of our experimental graphs, the size of a vertex represents the weight, which is proportional to its profit. can see that, in order to obtain more profits, the route goes to the upper-right corner almost directly from the start vertex. In this way, a big part of time is used among traversing those vertices with large profits. B. Routing Properties under Temporal Variations To examine the effects caused by temporal variations, we manipulate the profit functions. Specifically, we divide the space into four regions (quadrants) I,II,III and IV, as shown in Fig. 3(b). The profits are the same if the vertices are in the same region, but different if not. For example, in region 0 I, the profit function for each vit is ( 5wi if t 6 T /2 fi (t)I = 0 if t > T /2 where wi is the weight of vertex vi in graph G. (Keep in 0 mind that, wi is different from wit which is time-varying. Here the weight wi refers to a fixed parameter for profit function.) Similarly, the profit functions in regions II, III and IV are ( 10wi if t 6 T /2 fi (t)II = 0 if t > T /2 ( 0 if t 6 T /2 fi (t)III = 5wi if t > T /2 ( 0 if t 6 T /2 fi (t)IV = 10wi if t > T /2 Figure 3(b) reveals that, the path first transits the vertices in region II, because during the first half T , vertices in region II have larger profits than those in region III. After the path enters region I and after the time passes T /2, vertices in region IV contain larger profits. Such variations attract the path to go through vertices in region IV. This example indicates that our algorithm is sensitive to the time-varying profit functions. C. Comparison with Optimal Solution We compare the result of our algorithm with the optimal solution that is obtained by enumerating all solutions in a brute-force way. Because of the prohibitive time complexity O(n!) for searching for the optimal solution, the practical runtime for 13 vertices requires more than 10 minutes. Thus, we tested up to 12 vertices to compare with the optimal solutions. We investigated three representative profit functions in the form of linear, quadratic and logarithmic, respectively. The results are shown in Fig. 4. We can see that the results of our algorithm are very close to those of the optimal solutions, for all the three functions. Figure 5 shows a group of paths produced from our algorithm and the optimal solution. We use a linear profit fi (t) = wi t/T for this example. In many cases, our algorithm get the same results as those of the optimal solution, as shown in Fig. 5(a) and Fig. 5(b). Sometimes the paths are different, such as Fig. 5(c) and Fig. 5(d), but the collected profits are quite similar (in this example, ours obtains 183.5 whereas the optimal is 185.0). Fig. 5(e) and Fig. 5(f) show another example. D. Comparison with Classic OP Algorithm We also compared our algorithm with the classic OP algorithms. We implemented a well-known heuristic called center-of-gravity [11], and tested with both time-invariant and time-variant profits. Fig. 6 demonstrates the differences between the results of our approach and those of the OP. Fig. 7 show statistics of numerical evaluations with 200 vertices. We can see that, our algorithm outperforms the OP for both the time-invariant and time-varying profits. Particularly, the difference margin is larger for the timevarying profits. Actually, the classic OP cannot handle timevarying profits and the total profit curve quickly converges (with fixed given time T , the overall profit of OP stops growing along with the increment of the number of vertices). The running time is compared in Fig. 7(c), from which we can observe that, in practice our algorithm costs less time than that of the OP using the center-of-gravity heuristic. E. Path Quality under Different Time Intervals Time interval ∆t determines time discretization resolution, and therefore affacts the optimality as well. Table I shows statistics of the route’s quality under different time intervals. In the table, numbers in the first row are interval values, and numbers in the first column are the number of vertices. The remaining values are the collected profits from our algorithm. We can see that, in general the route’s quality is better when the time interval is smaller. However, obviously a smaller time interval will inevitably lead to a larger graph and thus require a larger running time. F. Demonstration with New York Taxi Data We demonstrate that the proposed algorithm can be used for (autonomous) taxi routing. Fig. 8 shows real data of taxi calls in New York city. The data is obtained from NYC Taxi and Limousine Commission [1]. We use the kmeans clustering algorithm to cluster the whole New York taxi calls into 50 local regions, as shown in Fig. 9(a). The clustering centers are shown in Fig. 9(b). We assume that each pick-up and the subsequent drop-off occur in the same local region, and each pick-up and drop-off is counted as (a) f (t) = wi t/T (b) f (t) = −wi (t2 +tT +T 2 )/T 2 (c) f (t) = wi log(t + 1) Fig. 4. Profit statistics on (a) Linear function; (b) Quadratic function; (c) Logarithmic function. (a) (b) (c) (d) (e) (f) Fig. 5. Examples with linear profit functions. The paths in red are from our algorithm while the paths in green are found with brute-force search. (a)(b) Our algorithm obtains optimal solution 281; (c)(d) Our solution is 183.5 whereas the optimal solution is 185.0; (e)(f) Our solution is 186.6 whereas the optimal solution is 195.1. (a) (b) (c) (d) Fig. 6. Path comparisons. The paths in red are computed from our algorithm while the paths in green are from the classic OPs using the center-of-gravity heuristic. (a)(b) Time-invariant profits fi (t) = wi ; (c)(d) Time-varying profits fi (t) = wi t/T TABLE I S TATISTICS OF SOLUTION QUALITY WITH VARIOUS TIME INTERVALS ∆t Intervals (a) (b) 0.1 0.5 1 2 5 10 50 541.1 530.3 529.2 522.1 515.1 498.3 100 788.2 783.4 767.9 735.6 727.5 619.2 150 950.2 919.0 909.7 892.9 823.0 686.8 200 1129.5 1117.6 1107.1 1094.3 935.5 696.0 Note: T = 150, graph within [−50, 50] × [−50, 50] on x-y plane, profit fi (t) = wi t/T. 50,100,150 and 200 are the numbers of vertices. (c) Fig. 7. (a) Performance on time-invariant profits fi (t) = wi ; (b) Performance on time-varying profits fi (t) = wi t/T ; (c) Statistics of practical running time. other taxis too). Therefore, this problem can be formulated as an orienteering problem with time-varying profits, where we need to treat those clustered regions as super routing depots that offer different profits at different times of a day. Fig. 11 demonstrates a result generated from running our algorithm, from which we can observe that the taxi is routed across those regions with the most taxi calls, even though we set the starting point at a distant location with sparse requests. a completion of one service. Then a taxi driver aims at maximizing the profit by providing more services during a fixed period. In each region, we calculate the total number of taxi calls in every predefined time unit (e.g., minute). An example is shown in Fig. 10. We regard the total number of taxi calls in a region as a time-varying profit, and more calls indicate higher chance of getting profits (note, there are We presented a framework for addressing the orienteering problem with time-varying profits. Instead of following traditional mixed integer program solution routines, we develop an intuitive and effective framework that incorporates timevariations into the fundamental planning process. Specifically, we first construct a deterministic spatio-temporal VII. C ONCLUSION AND F UTURE W ORK Fig. 8. Data of taxi calls in New York city Fig. 11. The taxi routing path generated from our algorithm. The big square is the starting location. (a) (b) Fig. 9. (a) Taxi calls are clustered through k-means; (b) Clustering centers represent high-level routing depots. representation where both spatial description and temporal logic are unified into one routing topology, and then we extend existing sorting and searching algorithms, so that the routing solutions can be computed in an extremely efficient way. Finally, we validated our algorithm with numerical evaluations and the results show that our framework produces near-optimal solutions in a very efficient way. R EFERENCES [1] NYC Taxi and Limousine Commission. http://www.nyc.gov/ html/tlc/html/about/trip_record_data.shtml. Accessed: 2017-02-25. [2] H. M. Afsar and N. Labadie. Team orienteering problem with decreasing profits. Electronic Notes in Discrete Mathematics, 41:285– 293, 2013. [3] R. Bellman. On a routing problem. Quarterly of applied mathematics, pages 87–90, 1958. [4] M. Bellmore and G. L. Nemhauser. The Traveling Salesman Problem: A Survey. Operations Research, 16(3):538–558, May 1968. [5] J. Binney, A. Krause, and G. S. Sukhatme. Optimizing waypoints for monitoring spatiotemporal phenomena. The International Journal of Robotics Research, 32(8):873–888, 2013. [6] I.-M. Chao, B. L. Golden, and E. A. Wasil. The team orienteering problem. European journal of operational research, 88(3):464–474, 1996. [7] N. Christofides. The vehicle routing problem. Revue française d’automatique, d’informatique et de recherche opérationnelle. Recherche opérationnelle, 10(1):55–70, 1976. (a) (b) Fig. 10. For a particular region (region 1 in this example), the number of taxi calls is varying as time elapses. (a) Time unit is minute; (b) Time unit is hour. [8] H. C. B. de Oliveira and G. C. Vasconcelos. A hybrid search method for the vehicle routing problem with time windows. Annals of Operations Research, 180(1):125–144, 2010. [9] F. V. Fomin and A. Lingas. Approximation algorithms for timedependent orienteering. Information Processing Letters, 83(2):57–62, 2002. [10] A. Garcia, O. Arbelaitz, P. Vansteenwegen, W. Souffriau, and M. T. Linaza. Hybrid approach for the public transportation time dependent orienteering problem with time windows. In International Conference on Hybrid Artificial Intelligence Systems, pages 151–158, 2010. [11] B. L. Golden, L. Levy, and R. Vohra. The orienteering problem. Naval research logistics, 34(3):307–318, 1987. [12] A. Gunawan, H. C. Lau, and P. Vansteenwegen. Orienteering problem: A survey of recent variants, solution approaches and applications. European Journal of Operational Research, 2016. [13] G. Gutiérrez-Jarpa, G. Desaulniers, G. Laporte, and V. Marianov. A branch-and-price algorithm for the vehicle routing problem with deliveries, selective pickups and time windows. European Journal of Operational Research, 206(2):341–349, 2010. [14] A. B. Kahn. Topological sorting of large networks. Commun. ACM, 5(11):558–562, Nov. 1962. [15] H. Kellerer, U. Pferschy, and D. Pisinger. Introduction to NPCompleteness of knapsack problems. Springer, 2004. [16] M. Keshtkaran, K. Ziarati, A. Bettinelli, and D. Vigo. Enhanced exact solution methods for the team orienteering problem. International Journal of Production Research, 54(2):591–601, 2016. [17] G. Laporte. The traveling salesman problem: An overview of exact and approximate algorithms. European Journal of Operational Research, 59(2):231–247, 1992. [18] M. Mahmoudi and X. Zhou. Finding optimal solutions for vehicle routing problem with pickup and delivery services with time windows: A dynamic programming approach based on state–space–time network representations. Transportation Research Part B: Methodological, 89:19–42, 2016. [19] C. E. Miller, A. W. Tucker, and R. A. Zemlin. Integer programming formulation of traveling salesman problems. Journal of the ACM (JACM), 7(4):326–329, 1960. [20] G. L. Nemhauser and L. A. Wolsey. Integer and Combinatorial Optimization. Wiley-Interscience, New York, NY, USA, 1988. [21] H. Tang and E. Miller-Hooks. A tabu search heuristic for the team orienteering problem. Computers & Operations Research, 32(6):1379– 1407, 2005. [22] P. Vansteenwegen, W. Souffriau, and D. Van Oudheusden. The orienteering problem: A survey. European Journal of Operational Research, 209(1):1–10, 2011. [23] C. Verbeeck, K. Sörensen, E.-H. Aghezzaf, and P. Vansteenwegen. A fast solution method for the time-dependent orienteering problem. European Journal of Operational Research, 236(2):419–432, 2014. [24] T. Vidal, N. Maculan, L. S. Ochi, and P. H. Vaz Penna. Large neighborhoods with implicit customer selection for vehicle routing problems with profits. Transportation Science, 50(2):720–734, 2015. 

2cs.AI

Synchronization Detection and Recovery of Steganographic Messages with Adversarial Learning Haichao Shi1,2 and Xiao-Yu Zhang1 1 2 Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China Abstract. As a means for secret communication, steganography aims at concealing a message within a medium such that the presence of the hidden message can hardly be detected. In computer vision tasks, adversarial training has become a competitive learning method to generate images. However, the generative tasks are confronted with great challenge on synthesizing images. This paper studies the mechanism of applying adversarial learning for discriminative tasks to learn the steganographic algorithm. We show that through unsupervised adversarial training, the adversarial model can produce robust steganographic solutions, which act like an encryption. Through supervised adversarial training, we can also train a robust steganalyzer, which is utilized to discriminate whether an image contains secret information. Our model is composed of three modules, i.e. a generator, a discriminator and a steganalyzer. All the three members are trained simultaneously. To formulate the algorithm, we use a game to represent the communication between the three parties. In this game, the generator and discriminator attempt to communicate with each other with secret messages hidden in an image. While the steganalyzer attempts to analyze whether there is a transmission of confidential information. Experimental results demonstrate the effectiveness of the proposed method on steganography tasks. Keywords: Steganography, Steganalysis, Adversarial learning. 1 Introduction Steganography has become a heated research topic in information security, which is widely used in various applications. It is used to conceal secret information (i.e. a message, an image or a sound) within a medium such that both the secret message as well as its content will remain invisible. Therefore, the ultimate goal is to conceal a payload into a cover object without affecting the sharpness of the cover object. With the emergence of all sorts of media, steganography, especially image-based steganography, has drawn worldwide attention. The design of robust steganographic algorithms which are used in secret communications is of vital importance. Steganographic algorithms are designed to hide the secret information within a cover message such that the cover message appears unaltered to an external adversary. Recent researchers have made great efforts to design steganographic algorithms to minimize 2 the perturbations within the cover object when embedded into the secret message. Obviously, concealing both the content of the carrier as well as the presence of a secret message is of vital importance for privacy sensitive communication. The steganography aims to hide the presence of a secret message, while the steganalysis algorithms are utilized to find out and recover the secret information of the carrier object. Similar with adversarial training, the steganalysis serves as the adversarial to steganography. For the game design of the steganographic techniques, we leverage the ideas from the adversarial learning. Adversarial learning is based on game theory, and is combined with unsupervised way to jointly train the model. In our previous work [19], we proposed a novel strategy of secure steganograpy based on generative adversarial networks to generate suitable and secure covers for steganography, which was not considered the decoding process. This paper is different from that, we not only encode the secret messages into the images, but decode it utilizing the network. Then we utilize a steganalysis network to detect the presence of hidden messages. Through unsupervised training, the generator plays the role of a sender, which is utilized to generate steganographic images as real as possible. As the discriminator plays the role of a receiver, it not only differentiates real images and generated images, but also extracts the secret messages. And the steganalysis network, as a listener of the whole process, incorporates supervised learning with adversarial training to compete against state-of-the-art steganalysis methods. In summary, this paper makes the following contributions: (1) We incorporate the generative adversarial network into steganography, which proves to be a good way of encryption. (2) We integrate the unsupervised learning method and supervised learning method to train the generator and steganalyzer respectively, and received robust steganographic techniques in unsupervised manner. (3) We also utilize the discriminative network to extract the secret information. Experiments are conducted on widely used datasets, which demonstrate the advantages of the proposed method. The rest of the paper is structured as follows. In Section 2, we discuss the related work of steganography, adversarial networks. In Section 3, we elaborate the proposed method. In Section 4, experiments are conducted to demonstrate the effectiveness of the proposed method. In Section 5, we draw conclusions. 2 Related work 2.1 Steganography The image-based steganography algorithm can be split into two categories. The one is based on the spatial domain, the other is based on the DCT domain. In our work, we focus on the spatial domain steganography. The state-of-the-art steganographic schemes concentrate on embedding secret information within a medium while minimizing the perturbations within that medium. On the contrary, steganalysis is to figure out whether there is secret information or not in the medium. 3 Least Significant Bit (LSB) [12] is one of the most popular embedding methods in spatial domain steganography. If LSB is adopted as the steganography method, the statistical features of the image are destroyed. And it is easy to detect by the steganalyzer. For convenience and simple implementation, the LSB algorithm hides the secret to the least significant bits in the given image’s channel of each pixel. Mostly, the modification of the LSB algorithm is called ±1-embedding. It randomly adds or subtracts 1 from the channel pixel, so the last bits would match the ones needed. Besides the LSB algorithm, some sophisticated steganographic schemes choose to use a distortion function which is used for selecting the embedding localization of the image. This type of steganography is called the content-adaptive steganography. The minimization of the distortion function between the cover image 𝐶 and the steganographic image 𝑆 is usually required. These algorithms are the most popular and the most secure image steganography in spatial domain, such as HUGO (Highly Undetectable steGO) [4], WOW (Wavelet Obtained Weights) [2], S-UNIWARD (spatial universal wavelet relative distortion) [3], etc. 𝑑 𝐶, 𝑆 = 𝑓 𝐶, 𝑆 ∗ |𝐶 − 𝑆| (1) where 𝑓 𝐶, 𝑆 is the cost of modifying a pixel, which is variable in different steganographic algorithms. HUGO is a steganographic scheme that defines a distortion function domain by assigning costs to pixels based on the effect of embedding some information within a pixel. It uses a weighted norm function to represent the feature space. WOW is another content-adaptive steganographic method that embeds information into a cover image according to textural complexity of regions. In WOW shows that the more complex the image region is, the more pixel values will be modified in this region. S-UNIWARD introduces a universal distortion function that is independent of the embedded domain. Despite the diverse implementation details, the ultimate goals are identical, i.e. they are all devoted to minimize this distortion function, to embed the information into the noise area or complex texture, and to avoid the smooth image coverage area. 2.2 Adversarial Learning In recent years, GANs have been successfully applied to image generation tasks. The method that generative adversarial networks generate images can be classified in two manners in general. The first is mainly exploring image synthesis tasks in an unconditioned manner that generates synthetic images without any supervised learning schemes. Goodfellow et al. [7] propose a theoretical framework of GANs and utilize GANs to generate images without any supervised information. However, the early GANs has somewhat noisy and blurry results and sometimes the gradient will be vanished when training the networks. Later, Radford et al. [15] propose a deep convolutional generative adversarial networks (DCGANs) for unsupervised representation. To solve the situation of gradient vanishing, WGAN [14] is proposed using the Wasserstein distance instead of the Jensen-Shannon divergence, to make the data set distribution compared with the learning distribution from G. 4 Another direction of image synthesis with GANs is to synthesize images by conditioning on supervised information, such as text or class labels. The Conditional GAN [16] is one of the works that develop a conditional version of GANs by additionally feeding class labels into both generator and discriminator of GANs. Info-GAN [17] introduces a new concept, which divides the input noise z into two parts, one is the continuous noise signal that cannot be explained, and the other is called C. Where C represents a potential attribute that can be interpreted as a facial expression, such as the color of eyes, whether with glasses or not, etc. in the facial tasks. Recently, Reed et al. [18] utilize GANs for image synthesis using given text descriptions, enabling translation from character level to pixel level. Adversarial learning has been applied to steganographic and cryptographic problems. In Abadi’s [1] work, they apply two neural networks to adversarial game to encrypt a secret message. In which game the discriminator can be deceived. In this paper, we are devoted to training a model that can learn a steganographic technique by itself making full use of the discriminator. 3 Adversarial Steganography This section discusses our proposed adversarial steganographic scheme, we present the experiments supporting our claims after laying the theoretical groundwork. 3.1 Network Architecture Generator. The generator receives the cover images and secret messages. Before learning the distribution of the images, the secret messages are first embedded into the cover image. The generated steganographic images are then input into the generator. As is shown in Fig.1, we use a fully connected layer, which can cover anywhere of the image not merely a fixed region. Then four fractionally-strided convolution layers, and finally a Hyperbolic tangent function layer. The architecture is shown as follows: Fig. 1. The architecture of the generator Discriminator. The discriminator is mainly responsible for extracting the secret messages. Besides, it can also help to optimize the visual quality of the generated images. 5 We use four convolutional layers, and then a fully connected layer. In detail, we additionally add a decoding function at the end of the network to extract the secret messages. Which acts as an interface, accepting the steganographic images as input. The decoding function analyzes the modification of pixels in images, and recovers the message. As is shown as follows: Fig. 2. The architecture of the discriminator Steganalyzer. The steganalyzer’s inputs are both covers and steganographic images. The steganalyzer’s architecture is similar to the discriminator, as is illustrated in Fig. 3. With distinct, we first use a predefined high-pass filter to make a filtering operation, which is mainly for steganalysis. And then four convolutional layers. Finally, we use a classification layer, which includes several fully connected layers. Fig. 3. The architecture of the steganalyzer 3.2 Optimization Objective Our training scheme includes three parties: generator, discriminator and steganalyzer. In this game, the generator conveys a secret medium to the discriminator, which the steganalyzer attempts to eavesdrop the conveying process and find out whether there are some special information containing in the process. Generally speaking, the generator generates a steganographic image within the cover image embedded with a secret message. Then generator passes it to the next network which we called discriminator, concentrating on decoding and recovering the message. The steganalyzer figures out a 6 confidence probability of how much likely the image is hidden with a secret message when given an image. For the whole process of information conveying is transparent to the discriminator, so the decoding can be easily achieved. In the game of the three parties, the generator is trained to learn to produce a steganographic image such that the discriminator can decode and recover the secret message, and such that the steganalyzer can output a better probability whether it is a cover or a steganographic image. Fig. 4. The whole architecture of the training game. 𝑀 and 𝑀′ represents the secret message and extracted message from the steganographic image respectively, 𝐶 represents the cover image, and 𝑆𝐶 represents the steganographic image. 𝑝 is the probability of classifying the co- vers and steganographic images. 𝑝′ is the probability of classifying the real images and generated images. The whole architecture can be depicted in Fig.4, the generator receives a cover image, 𝐶, and a secret message, 𝑀. Combined with the above two inputs, it outputs a steganographic image, 𝑆- , which is simultaneously given to the discriminator and steganalyzer. Once receiving the steganographic image, the discriminator decodes it and attempts to recover the message. For the game is under generative adversarial networks, the discriminator can also improve the visual quality of the generated images. In addition to the steganographic images, the cover images are also input into the steganalyzer, whose output is a probability to classify the steganographic images and covers. We let 𝜔/ , 𝜔0 , 𝜔1 denote the parameters of generator, discriminator and steganalyzer respectively. We utilize 𝐺(𝜔/ , 𝑀, 𝐶) represents generator’s output given an image 𝐶 and secret message 𝑀 , 𝐷(𝜔0 , 𝑆- ) for discriminator’s output on steganographic images 𝑆- , and 𝑆(𝜔1 , 𝐶, 𝑆- ) for steganalyzer’s output on covers and steganographic images. Then we use 𝐿/ , 𝐿0 , 𝐿1 represent the loss of three networks respectively. And we can calculate as follows: 𝐷 𝜔0 , 𝑆- = 𝐷(𝜔0 , 𝐺(𝜔/ , 𝑀, 𝐶)) (2) 𝑆 𝜔1 , 𝐶, 𝑆- = 𝑆(𝜔1 , 𝐶, 𝐺(𝜔/ , 𝑀, 𝐶)) (3) 7 where the 𝐷 𝜔0 , 𝑆- is output of the discriminator, 𝑆 𝜔1 , 𝐶, 𝑆- is the output of the steganalyzer. For the generator’s loss, we correspond it with the other two networks. And we add constraints to normalize the loss. 𝐿/ 𝜔/ , 𝑀, 𝐶 = 𝜆/ ∙ 𝑑 𝐶, 𝑆- + 𝜆0 ∙ 𝐿0 (𝜔0 , 𝑆- ) + 𝜆1 ∙ 𝐿1 (𝜔1 , 𝐶, 𝑆- ) (4) where 𝑑 𝐶, 𝑆- is the Euclidean distance between the covers and the steganographic images. And 𝜆/ , 𝜆0 , 𝜆1 are the calculation loss weights of the generator, discriminator and steganalyzer respectively. These variables are varying from [0,1]. Also, we set discriminator’s loss to be the norm form. Here, we use the ℓ@ norm. 𝐿0 𝜔/ , 𝜔0 , 𝑀, 𝐶 = 𝑀, 𝐷 𝜔0 , 𝑆- @ = ||𝑀, 𝐷(𝜔0 , 𝐺(𝜔/ , 𝑀, 𝐶))||@ = ||𝑀, 𝑀′||@ (5) where the form ||𝑀, 𝑀′||@ is the distance between 𝑀 and 𝑀′. And we set steganalyzer’s loss to be binary cross-entropy of logistic regression. 𝐿1 𝜔/ , 𝜔1 , 𝐶, 𝑆- = − A B F 𝑦 ln 𝑆 𝜔1 , 𝐶, 𝑆- + (1 − 𝑦) ln(1 − 𝑆 𝜔1 , 𝐶, 𝑆- ) (6) where 𝑦 = 1 if 𝑥 = 𝐶, which means classifying the images as covers, 𝑦 = 0 if 𝑥 = 𝑆- , which means the steganographic images are identified. We apply adversarial learning techniques to discriminative tasks to learn the steganographic algorithm. Under unsupervised adversarial training, the model can produce steganographic techniques, which act like an encryption. 4 Experiments As a proof of concept, we conducted our experiments on two image datasets: celebrity faces in the wild (CelebA) [8] consisting of 202,599 images and BossBase, which is a dataset of 10000 grayscale images of various scenes. Our datasets are formed of 64 × 64 pixel images. For the choice of 𝑀, we concatenate a random paragraph message to generalize our model to random message, with each sample of each dataset. And we consider different kinds of embedding rates, which vary from 0.1bpp (bits per pixel) to 0.4bpp. Generally, most steganographic algorithms can successfully hide secret message approximately at 0.4bpp. All experiments were performed in TensorFlow[9], on a workstation with a Titan X GPU. 4.1 CelebA dataset For the CelebA dataset, which contains 202,599 images of celebrity faces from 10,177 unique identities. We train the model using RMSProp optimization algorithm with the learning rate of 2×10JK . At each batch we train either the generator, discriminator, or steganalyzer alternatively. From the results, we can see that there is no noticeable image quality decrease between 0.1bpp and 0.4bpp. As is shown in Fig. 5. And we also calculate the PSNR (Peak Signal to Noise Ratio) to evaluate the steganographic effect. 8 Fig. 5. The above are the covers and steganographic images of CelebA dataset. (a) represents the covers, (b) represents the steganographic images with the embedding rates of 0.1bpp, (c) represents the steganographic images with the embedding rates of 0.4bpp. The bottom is the contrast of PSNR of embedding rates at 0.1bpp and 0.4bpp. 4.2 BossBase dataset In addition to the experiments on the CelebA dataset, we also trained our model on the BossBase dataset, which is a standard steganography dataset. For the images from this dataset do not come from a single distribution, so the dataset can perform worse than the experiments on the CelebA dataset. 9 Fig. 6. Similarly, the above are the covers and steganographic images of BossBase dataset. (a) represents the covers, (b) represents the steganographic images with the embedding rates of 0.1bpp, (c) represents the steganographic images with the embedding rates of 0.4bpp. The bottom is also the contrast of PSNR of embedding rates at 0.1bpp and 0.4bpp. We can conclude from the two groups of experiments that for the datasets with distinct distribution, different embedding rates can influence the similarity between the original images and the embedded images. 4.3 Comparison with related work We train in such a way that the steganalyzer can detect the covers and steganographic images with low rates. Thus, on the basis of this experiment, we additionally conduct another experiment compared with the state-of-the-art steganographic algorithms. Which also shows competitive results. For both datasets, we compare our scheme against steganographic algorithms HUGO and S-UNIWARD. As is shown in Table 1, our steganalyzer performs competitively. By training steganalyzer in a supervised fashion, the experiments show that it has the capacity to become a strong steganalyzer, competing against established techniques like GNCNN, and so is a good choice for the steganalyzer. Furthermore, for both the CelebA 10 and BossBase datasets, we use 10,000 samples and split them in half, creating a training set and a test set. And we then trained the generator on the training set. Following this, we then created an additional 5,000 steganographic images utilizing each steganography algorithm. Thus, we can make the cover set and steganographic images set consist of 5,000 images respectively. Table 1. Accuracy of distinguishing between cover and steganographic images for the GNCNN and Steganalyzer at an embedding rate of 0.4bpp. Steganographic Algorithm HUGO S-UNIWARD Generator CelebA Steganalyzer GNCNN 0.93 0.91 0.94 Steganalyzer 0.91 0.89 0.92 BossBase Steganalyzer GNCNN 0.72 0.75 0.83 Steganalyzer 0.68 0.71 0.82 As demonstrated in the experiments, the steganalyzer performs competitively against the GNCNN [10], and the generator also performs well against other steganographic techniques. The experimental data shows that on the one hand, the generated images are good at visual quality. On the other hand, the generated images are harder to detect, which shows the security of the generated images as covers. 5 Conclusion In this paper, we have proposed an adversarial steganography architecture with generative adversarial networks. On the basis of generative adversarial networks, we leverage the adversarial structure to form an effective steganographic method. Encouraging results are received from experiments conducted on the widely used datasets in comparison with several state-of-the-art techniques. We think it could be used for adaptive steganographic algorithm for social network in the future. We believe that, by exploring more steganography properties, better performance can be achieved. References 1. Abadi M, Andersen D G. Learning to Protect Communications with Adversarial Neural Cryptography[J]. 2016:15. 2. Holub V, Fridrich J. Designing steganographic distortion using directional filters[C]// IEEE International Workshop on Information Forensics and Security. IEEE, 2012:234-239. 3. Holub V, Fridrich J, Denemark T. Universal distortion function for steganography in an arbitrary domain[J]. Eurasip Journal on Information Security, 2014, 2014(1):1. 4. Pevný T, Filler T, Bas P. Using High-Dimensional Image Models to Perform Highly Undetectable Steganography[M]// Information Hiding. Springer Berlin Heidelberg, 2010:161177. 5. Xu G, Wu H Z, Shi Y Q. Structural Design of Convolutional Neural Networks for Steganalysis[J]. IEEE Signal Processing Letters, 23(5):708-712, 2016. 11 6. Xu G, Wu H Z, Shi Y Q. Ensemble of CNNs for Steganalysis: An Empirical Study[C]// ACM Workshop on Information Hiding and Multimedia Security. ACM, 2016:103-107. 7. Goodfellow I J, Pouget-Abadie J, Mirza M, et al. Generative Adversarial Networks[J]. Advances in Neural Information Processing Systems, 3:2672-2680, 2014. 8. Liu Z, Luo P, Wang X, et al. Deep Learning Face Attributes in the Wild[J]. pp. 3730-3738, 2014. 9. Abadi M, Agarwal A, Barham P, et al. TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems[J]. 2016. 10. Yinlong Qian, Jing Dong, WeiWang, and Tieniu Tan. Deep learning for steganalysis via convolutional neural networks. In IS&T/SPIE Electronic Imaging, pp. 94090J–94090J. International Society for Optics and Photonics, 2015a. 11. Browne M, Ghidary S S, Mayer N M. Convolutional Neural Networks for Image Processing with Applications in Mobile Robotics[J]. Studies in Computational Intelligence, 2008, 83:327-349. 12. Mielikainen, Jarno. Lsb matching revisited. IEEE signal processing letters, 13(5):285–287, 2006. 13. Andrew D Ker. Resampling and the detection of lsb matching in color bitmaps. In Eletronic Imaging 2005, pp. 1–15. International Society for Optics and Photonics, 2005. 14. Arjovsky M, Chintala S, Bottou L. Wasserstein GAN[J]. 2017. 15. Radford, Alec, Metz, Luke, and Chintala, Soumith. Unsupervised representation learning with deep convolutional generative adversarial networks. CoRR, abs/1511.06434, 2015. URL http://arxiv.org/abs/1511.06434. 16. Odena A, Olah C, Shlens J. Conditional Image Synthesis With Auxiliary Classifier GANs[J]. 2016. 17. Chen X, Duan Y, Houthooft R, et al. InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets[J]. 2016. 18. Reed S, Akata Z, Yan X, et al. Generative Adversarial Text to Image Synthesis[J]. 2016:1060-1069. 19. Shi H, Dong J, Wang W, et al. SSGAN: Secure Steganography Based on Generative Adversarial Networks[C]. 2017. 

1cs.CV

A NEW ROBUST FREQUENCY DOMAIN ECHO CANCELLER WITH CLOSED-LOOP LEARNING RATE ADAPTATION Jean-Marc Valin and Iain B. Collings arXiv:1602.08609v1 [cs.SD] 27 Feb 2016 CSIRO ICT Centre, Sydney, Australia (jmvalin@jmvalin.ca) ABSTRACT One of the main difficulties in echo cancellation is the fact that the learning rate needs to vary according to conditions such as double-talk and echo path change. Several methods have been proposed to vary the learning. In this paper we propose a new closed-loop method where the learning rate is proportional to a misalignment parameter, which is in turn estimated based on a gradient adaptive approach. The method is presented in the context of a multidelay block frequency domain (MDF) echo canceller. We demonstrate that the proposed algorithm outperforms current popular double-talk detection techniques by up to 6 dB. Fig. 1. Block diagram of echo cancellation system. 1. INTRODUCTION In any echo cancellation system, the presence of near end speech (double-talk) tends to make the adaptive filter diverge. To counter the effect, robust echo cancellers require adjustment of the learning rate to account for the presence of doubletalk in the signal. Most echo cancellation algorithms attempt to explicitly detect double-talk [1] conditions and then react by freezing the coefficients of the adaptive filter (setting the learning rate to zero). Reliable double-talk detection is a difficult problem and sometimes it is not clear what should be considered as double-talk, especially in an acoustic echo cancellation context with stationary background noise. In previous work [2], we proposed a new approach to make echo cancellation more robust to double-talk. Instead of attempting to explicitly detect double-talk conditions, a continuous learning rate was used. The learning rate depends on a misalignment estimate, which is obtained through a linear regression. While the technique gives good results, the estimation of the misalignment remains a difficult problem. In this paper, we propose a new approach where the misalignment is estimated in closed-loop based on a gradient adaptive approach. This closed-loop technique is applied to the block frequency domain (MDF) adaptive filter [3] and shows a significant improvement over previous approaches. c 2007 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. In Section 2, we discuss the optimal learning rate in presence of noise. Section 3 introduces the multidelay frequency domain (MDF) adaptive filter. In Section 4, we propose a gradient adaptive technique for adjusting the learning rate of the MDF algorithm. Experimental results and a discussion are presented in Section 5 and Section 6 concludes this paper. 2. OPTIMAL LEARNING RATE IN THE PRESENCE OF DOUBLE-TALK In the acoustic echo cancellation context illustrated in Fig. 1, the speech signal x (n) received from the far end is played into a loudspeaker. The microphone signal d (n) captures both the convoluted far end speech y (n) and the near end noisy speech v (n). The adaptive filter attempts to estimate the impulse response ĥ (n) to be as close as possible to the real impulse response h (n). The estimated echo ŷ (n) is subtracted from d (n), so the output signal e (n) contains both double-talk v (n) and residual echo r (n) = y (n) − ŷ (n). The conventional approach to double-talk robustness consists of setting the learning rate to zero when double-talk is detected. Double-talk detectors [1] are thus an important aspect of the approach. Unfortunately, they are sometimes unreliable, especially in acoustic echo cancellation context when background noise is present. In this paper, we investigate continuous learning rates that do not depend on a binary doubletalk decision. Whenever an adaptive filter is not perfectly adjusted, its residual signal r (n) can be used to gain better information about the exact (time-varying) filter weights h (n). However, the amount of information about h (n) present in e (n) decreases with the amount of noise and near end speech v(n). In the case of the normalised least mean square (NLMS) filter, it means that the stochastic gradient becomes less reliable when the noise increases or when the filter misalignment decreases (as the filter converges). The theoretical optimal learning rate is approximately equal to the residual-to-error ratio [2]:  E r2 (n) µopt (n) ≈ (1) E {e2 (n)} where r (n) = y (n) − ŷ (n) is the (unknown) residual echo and e (n) is the error signal. One possible method to vary the learning rate µ (n) would be to use the generalized normalized gradient descent (GNGD) algorithm [4], which includes the NLMS learning rate:  µ0 E x2 (n) (2) µ (n) = E {x2 (n)} + ǫ (n) where ǫ (n) is adapted based on the NLMS stochastic gradient behaviour. To examine ǫ (n) more closely, it is reasonable to surmise that (2) eventually converges to the optimal learning rate defined by (1). Assuming steady state behaviour (ǫ (n) is stable) and µ0 = 1, we find that:   E r2 (n) E r2 (n) ≈ (3) E {r2 (n)} + γ (n) ǫ (n) E {e2 (n)}  2  2 where  2 γ = E r (n)2 /E x (n)  . Knowing that we have E e (n) = E r (n) + E v 2 (n) , we find the relation ǫ (n) ≈ E v 2 (n) /γ (n). In other words, the gradientadaptive parameter ǫ (n) is approximately proportional to the variance of the near-field signal and independent of the farfield signal. Because ǫ (n) can only be adapted slowly  over time, it is clear that (2) implicitly assumes that E v 2 (n) also varies slowly. While this is a reasonable assumption in applications where the background noise is constant or slowly varying, the assumption does not hold for acoustic echo cancellation, where double-talk can start or stop at any time. In previous work [2], we proposed to use (1) directly to  adapt the learning rate. While E e2 (n) can  easily be estimated, the estimation of the residual echo E r2 (n) is difficult because one does not have access to the real filter coefficients. One reasonable assumption we make is that:    E r2 (n) = η (n) E y 2 (n) ≈ η (n) E ŷ 2 (n) (4) where η (n) is the normalised filter misalignment (or the inverse of the echo return loss enancement) and is easier to estimate because it is assumed to vary slowly as a function of time. Although direct estimation of η (n) through linear regression can lead to good results, estimating η (n) remains a difficult problem. In this paper we propose to apply a gradient adaptive approach to the problem of estimating η (n). 3. THE MDF ALGORITHM In this paper, we consider the special case of the multidelay block frequency domain (MDF) adaptive filter [3]. The MDF algorithm in matrix form is detailed here for the sake of clarity. Let N be the MDF block size, K be the number of blocks and F denote the 2N × 2N Fourier transform matrix, we denote the frequency-domain signals for frame ℓ as: T (5) e (ℓ) = F [01×N , e (ℓN ) , . . . , e (ℓN + N − 1)] o n T xk(ℓ) = diag F[x((ℓ−k−1)N ),. . . ,x((ℓ−k+1)N −1)] (6)  X (ℓ) = x0 , x1 , . . . , xK−1  (7) d (ℓ) = F [01×N , d (ℓN ) , . . . , d (ℓN + N − 1)] T (8) The MDF algorithm is then expressed in matrix form as: e (ℓ) = d (ℓ) − y (ℓ) (9) ŷ (ℓ) = G1 X (ℓ) ĥ (ℓ) (10) ĥ (ℓ + 1) = ĥ (ℓ) + G2 µ (ℓ) ∇ĥ (ℓ) (11) H ∇ĥ (ℓ) = Φ−1 xx (ℓ) X (ℓ) e (ℓ) (12) where Φxx (ℓ) is the diagonal normalisation matrix as computed in [3], G1 and G2 are the constraint matrices:   0N ×N 0N ×N F−1 (13) G1 = F 0N ×N IN ×N   IN ×N 0N ×N F−1 (14) G̃2 = F 0N ×N 0N ×N o n (15) G2 = diag G̃2 , G̃2 , . . . , G̃2 and µ (ℓ) is the 2KN × 2KN diagonal learning rate matrix: n T o (16) µ (ℓ) = diag µT (ℓ) , µT (ℓ) , . . . , µT (ℓ) µ (ℓ) = [µ0 (ℓ) , µ1 (ℓ) , . . . , µ2N −1 (ℓ)] T (17) If µk (ℓ) = µo , we have the standard MDF algorithm. 4. GRADIENT-ADAPTIVE LEARNING RATE In [2], we proposed to use the frequency-dependent learning rate:   2 (ℓ) ŷ k   (18) µk (ℓ) = min η (ℓ) 2 , µ0  |ek (ℓ)| where η(ℓ) is a direct estimation of the normalised misalignment exploiting the non-stationarity of the signals and using linear regression between the power spectra of the estimated echo and the output signal. The motivation behind this formulation is that it factors the residual echo estimation into a slowly-evolving (but unfortunately difficult to estimate) normalised misalignment η(ℓ) and a rapidly-evolving (but easy to 2 estimate) term far-end term yk (ℓ) . The learning rate can thus react quickly to double-talk even if the estimation of the residual echo (through the misalignment estimate) requires a longer time period. A remaining problem with that approach is that η(ℓ) is difficult to estimate and the algorithm does not know whether its estimate of η(ℓ) is too low or too high. In this sense, the η(ℓ) update in [2] is an open-loop estimate. 4.1. Adaptation algorithm In this paper we bypass the difficulty of estimating η(ℓ) directly and instead propose a closed-loop gradient adaptive estimation of η(ℓ). The parameter η(ℓ) is no longer an estimate of the normalised misalignment, but is instead adapted in closed-loop in such a way as to maximise convergence of the adaptive filter. As with other gradient-adaptive methods [5, 4] we compute the derivative of the mean square error E (ℓ) = eH (ℓ) e (ℓ), this time with respect to η (ℓ − 1), using the chain derivation rule: ∂E (ℓ) ∂E (ℓ) ∂e (ℓ) ∂ŷ (ℓ) ∂ ĥ (ℓ) ∂µ (ℓ − 1) = ∂η(ℓ−1) ∂e (ℓ) ∂ŷ (ℓ) ∂ ĥ (ℓ) ∂µ (ℓ − 1) ∂η (ℓ − 1) (19) = −e ≈− H H (ℓ)G1 X(ℓ)G2 Φ−1 xx (ℓ)X (ℓ) e(ℓ) ŷk (ℓ) ŷk (ℓ) 2 2 |ek (ℓ)| (20) 2 |ek (ℓ)|2 H ∇ĥ (ℓ) Φxx ∇ĥ (ℓ) (21) We propose to use a filtered version of the gradient with a multiplicative update, similar to the general approch in [6], and to drop the normalisation term Φxx . This results in the new gradient-based learning rate adaptation rule:  n o  2 H (ℓ) ℜ ψ (ℓ) ∇ ĥ (ℓ) ŷ k  n o  η (ℓ + 1) = η (ℓ) expρ  2P ∗ |ek (ℓ)| k ℜ ψk (ℓ) ∇ĥk (ℓ) (22) ψ (ℓ + 1) = αψ (ℓ) + ∇ĥ (ℓ) (23) where ρ is the step size and α controls the gradient smoothing (typically ρ = 1, α = 0.9). An intuitive interpretation for the update equation in (22) is that when the learning rate is too high (because η (ℓ)is too high), the gradient oscillates, causing η (ℓ) to decrease. On the other hand, when the learning rate is too low (perhaps because the echo path has changed) the gradient keeps pointing in the same direction and η (ℓ) increases. Also, because of |ŷ (ℓ)|2 the k 2 factor, we can apply (22) at each frame regardless |ek (ℓ)| of double-talk or even when no far-end speech is present (in 2 which case ŷk (ℓ) = 0 anyway). The last aspect that needs to be addressed is the initial condition. When the filter is initialised, all the weights are set to zero, which means that ŷ (ℓ) = 0 and no adaptation can take place in (18) and (22). In order to bootstrap the adaptation process, the learning rate µk (ℓ) is set to a fixed constant (we use µk (ℓ) = 0.25) for a short time equal to twice the filter length (only non-zero portions of signal x(n) are taken into account). This procedure is only necessary when the filter is initialised and is not required in case of echo path change. 4.2. Algorithm behaviour Here we show that the adaptive learning rate described above is able to deal with both double-talk and echo path change without explicit modelling. From (18) we can see that when 2 double-talk occurs, the denominator |ek (ℓ)| rapidly increases, causing an instantaneous decrease in the learning rate that lasts only as long as the double-talk period lasts. In the case of background noise, the learning rate depends on both the presence of an echo signal as well as the misalignment estimate. As the filter misalignment becomes smaller, the learning rate also becomes smaller. One major difficulty involved in double-talk detection is the need to distinguish between double-talk and echo path change since both cause a sudden increase in the filter error signal. This distinction is made implicitly in the gradientbased adaptation of η(ℓ). When the echo path changes, ∇ĥ (ℓ) starts pointing steadily in the same direction, thus significantly increasing η(ℓ), which is a clear sign that the filter is no longer properly adapted. In gradient adaptive methods [5, 6], the implicit assumption is that both the near-end and the far-end signals are nearly stationary. We have shown that the GNGD algorithm [4] only requires the near-end signal to be nearly stationary. In our new proposed method, both signals can be non-stationary, which is a requirement for double-talk robustness. 5. RESULTS AND DISCUSSION The proposed system is evaluated in an acoustic echo cancellation context with background noise, double-talk and a change in the echo path (16 seconds into a 32-second recording). The two impulse responses used are 1024-sample long and measured from real recordings in a small office room with both the microphone and the loudspeaker resting on a desk. The proposed algorithm is compared to our previous work [2], to the normalised cross-correlation (NCC) method [1] and to a baseline with no double-talk detection (no DTD). The optimal threshold found for the NCC algorithm was 0.35 with a learning rate µ = 0.25. It was found that choosing µmax = 0.75 as the upper bound on the learning rate gave good results for the proposed algorithm. In practise, µmax has little impact 25 4.5 proposed previous [2] NCC [1] no DTD 20 4 less double−talk more double−talk PESQ LQO-MOS ERLE (dB) 3.5 15 10 3 2.5 2 proposed previous [2] NCC [1] no DTD no cancellation 5 1.5 0 -20 less double−talk -15 -10 more double−talk -5 0 5 10 Near-end-to-far-end ratio (dB) 15 20 1 -20 -15 -10 -5 0 5 10 Near-end-to-far-end ratio (dB) 15 20 Fig. 2. Steady-state ERLE (first two seconds of adaptation not considered) as a function of the near-end to far-end ratio. Fig. 3. PESQ objective listening quality measure (LQOMOS) as a function of the near-end to far-end ratio. on the algorithm because the gradient-based adaptation would compensate for a higher value of µmax by reducing η (ℓ). The GNGD algorithm [4] is not included in the evaluation because it is not intended for applications where the near-field signal is highly non-stationary and it was not possible to obtain better results than the baseline. Fig. 2 shows the average steady-state (the first 2 seconds of adaptation are not considered) ERLE for the test data with different ratios of near-end signal and echo. Clearly, the proposed algorithm performs better than both our previous work (2 dB average improvement) and the NCC algorithm (6 dB average improvement). The perceptual quality of the output speech signal is also evaluated by comparing it to the near field signal v(n) using the Perceptual Evaluation of Speech Quality (PESQ) ITU-T recommendations P.862 [7]. The perceptual quality of the speech shown in Fig. 3 is evaluated based on the entire file, including the adaptation time. It is again clear that the proposed algorithm performs better than all other algorithms. It is worth noting that the reason why the results in Fig. 2 improve with double-talk (unlike in Fig. 2) is that the signal of interest is the double-talk v(n), so the higher the double-talk the less (relative) echo in the input signal. algorithm, we believe future work could apply it to other adaptive filtering algorithms, including the NLMS algorithm. 7. REFERENCES [1] J. Benesty, D.R. Morgan, and J.H. Cho, “A new class of doubletalk detectors based on cross-correlation,” IEEE Trans. on Speech and Audio Processing, vol. 8, no. 2, pp. 168–172, 2000. [2] J.-M. Valin, “On adjusting the learning rate in frequency domain echo cancellation with double-talk,” IEEE Trans. on Audio, Speech and Language Processing, 2007. [3] J.-S. Soo and K.K. Pang, “Multidelay block frequency domain adaptive filter,” IEEE Trans. ASSP, vol. 38, no. 2, pp. 373–376, 1990. [4] D.P. Mandic, “A generalized normalized gradient descent algorithm,” IEEE Signal Processing Letters, vol. 11, no. 2, pp. 115–118, 2004. [5] V.J. Mathews and Z. Xie, “A stochastic gradient adaptive filter with gradient adaptive step size,” IEEE Trans. on Signal Processing, vol. 41, pp. 2075–2087, 1993. 6. CONCLUSION We have demonstrated a novel method for adjusting the learning rate of frequency-domain adaptive filters based on a gradient adaptive estimation of the current misalignment. The proposed method performs better than a double-talk detector and previous work using direct estimation of the misalignment. In addition, the closed-loop gradient adaptive estimation of η (ℓ) makes the algorithm conceptually simple and means that there are very few important parameters to be tuned. Although the proposed algorithm is presented in the context of the MDF [6] W.-P. Ang and B. Farhang-Boroujeny, “A new class of gradient adaptive step-size LMS algorithms,” IEEE Trans. on Signal Processing, vol. 49, no. 4, pp. 805–810, 2001. [7] ITU-T, P.862: Perceptual evaluation of speech quality (PESQ): An objective method for end-to-end speech quality assessment of narrow-band telephone networks and speech codecs, 2001. 

3cs.SY

Reconfiguring spanning and induced subgraphs∗ Tesshu Hanaka† Takehiro Ito‡ Naomi Nishimura§ Haruka Mizuta‡ Vijay Subramanya§ Benjamin Moore§ Akira Suzuki‡ arXiv:1803.06074v1 [cs.DS] 16 Mar 2018 Krishna Vaidyanathan§ March 19, 2018 Abstract Subgraph reconfiguration is a family of problems focusing on the reachability of the solution space in which feasible solutions are subgraphs, represented either as sets of vertices or sets of edges, satisfying a prescribed graph structure property. Although there has been previous work that can be categorized as subgraph reconfiguration, most of the related results appear under the name of the property under consideration; for example, independent set, clique, and matching. In this paper, we systematically clarify the complexity status of subgraph reconfiguration with respect to graph structure properties. 1 Introduction Combinatorial reconfiguration [6, 5, 11] studies the reachability/connectivity of the solution space formed by feasible solutions of an instance of a search problem. More specifically, consider a graph such that each node in the graph represents a feasible solution to an instance of a search problem P , and there is an edge between nodes representing any two feasible solutions that are “adjacent,” according to a prescribed reconfiguration rule A; such a graph is called the reconfiguration graph for P and A. In the reachability problem for P and A, we are given source and target solutions to P , and the goal is to determine whether or not there is a path between the two corresponding nodes in the reconfiguration graph for P and A. We call a desired path a reconfiguration sequence between source and target solutions, where a reconfiguration step from one solution to another corresponds to an edge in the path. 1.1 Subgraph reconfiguration In this paper, we use the term subgraph reconfiguration to describe a family of reachability problems that take subgraphs (more accurately, vertex subsets or edge subsets of a given graph) as feasible solutions. Each of the individual problems in the family can be defined by specifying the node set and the edge set of a reconfiguration graph, as follows. (We use the terms node for reconfiguration graphs and vertex for input graphs.) Nodes of a reconfiguration graph. The set of feasible solutions (i.e., subgraphs) can be defined in terms of a specified graph structure property Π which subgraphs must satisfy; for example, “a graph is a tree,” “a graph is edgeless (an independent set),” and so on. By the choice of how ∗ This work is partially supported by JST ERATO Grant Number JPMJER1201, JST CREST Grant Number JPMJCR1402, and JSPS KAKENHI Grant Numbers JP16K00004 and JP17K12636, Japan. Research by Canadian authors is supported by the Natural Science and Engineering Research Council of Canada. † Kyushu University, Fukuoka, Japan, 3EC15004S@s.kyushu-u.ac.jp ‡ Tohoku University, Sendai, Japan, takehiro@ecei.tohoku.ac.jp, haruka.mizuta.s4@dc.tohoku.ac.jp, a.suzuki@ecei.tohoku.ac.jp § University of Waterloo, Waterloo, Canada, {brmoore, nishi, v7subram, kvaidyan}@uwaterloo.ca 1 Table 1: Subgraph representations and variants Subgraph representations Variant names Known reachability problems edge subset edge spanning tree [6] matching [6, 10], and b-matching [10] clique [7] independent set [6, 8] vertex subset induced induced forest [9] induced bipartite [9] induced tree [12] spanning clique [7] Es = E0 E1 E2 E3 = Et Figure 1: A reconfiguration sequence hE0 , E1 , E2 , E3 i in the edge variant under the TJ rule (also under the TS rule) with the property “a graph is a path,” where the edges forming solutions are depicted by thick lines. to represent subgraphs, each specific problem in the family can be categorized into one of three variants. (See also Table 1.) If a subgraph is represented as an edge subset, which we will call the edge variant, then the subgraph formed (induced) by the edge subset must satisfy Π. For example, Figure 1 illustrates four subgraphs represented as edge subsets, where Π is “a graph is a path.” On the other hand, if a subgraph is represented as a vertex subset, we can opt either to require that the subgraph induced by the vertex subset satisfies Π or that the subgraph induced by the vertex subset contains at least one spanning subgraph that satisfies Π; we will refer to these as the induced variant and spanning variant, respectively. For example, if Π is “a graph is a path,” then in the induced variant, the vertex subset must induce a path, whereas in the spanning variant, the vertex subset is feasible if its induced subgraph contains at least one Hamiltonian path. Figure 2 illustrates feasible vertex subsets of the induced variant and spanning variant. In the figure, the vertex subset V1′ is feasible in the spanning variant, but is not feasible in the induced variant, because it contains a spanning path but does not induce a path. As can be seen by this simple example, in the spanning variant, we need to pay attention to the additional complexity of finding a spanning subgraph and the complications resulting from the fact that the subgraph induced by the vertex subset may contain more than one spanning subgraph which satisfies Π. Edges of a reconfiguration graph. Since we represent a feasible solution by a set of vertices (or edges) in any variant, we can consider that tokens are placed on each vertex (resp., edge) in the feasible solution. Then, in this paper, we mainly deal with the two well-known reconfiguration rules, called the token-jumping (TJ) [8] and token-sliding (TS) rules [2, 4, 8]. In the former, a token can move to any other vertex (edge) in a given graph, whereas in the latter it can move only to an adjacent vertex (adjacent edge, that is sharing a common vertex.) For example, Figure 1 and Figure 2 illustrate reconfiguration sequences under the TJ rule for each variant. Note that the sequence in Figure 1 can also be considered as a sequence under the TS rule. In the reconfiguration graph, two nodes are adjacent if and only if one of the two corresponding solutions can be obtained from the other one by a single move of one token that follows the specified reconfiguration rule. Therefore, all nodes in a connected component of the reconfiguration graph represent subgraphs having the same number of vertices (edges). We note in passing that since in most cases we wish to retain the same number of vertices and/or edges, we rarely use the token-addition-and-removal (TAR) rule [6, 8], where we can add or remove a single token at a time, for subgraph reconfiguration problems. 2 Vs = V0 V1 V2 V3 V4 = Vt (a) Induced variant Vs = V0’ V1’ V2’ = Vt (b) Spanning variant Figure 2: Reconfiguration sequences hV0 , V1 , V2 , V3 , V4 i in the induced variant under the TJ rule and hV0′ , V1′ , V2′ i in the spanning variant under the TJ rule with the property “a graph is a path,” where the vertices forming solutions are depicted by colored circles, and the subgraphs satisfying the property by thick lines. 1.2 Previous work Although there has been previous work that can be categorized as subgraph reconfiguration, most of the related results appear under the name of the property Π under consideration. Accordingly, we can view reconfiguration of independent sets [6, 8] as the induced variant of subgraph reconfiguration such that the property Π is “a graph is edgeless.” Other examples can be found in Table 1. We here explain only known results which are directly related to our contributions. Reconfiguration of cliques can be seen as both the spanning and the induced variant; the problem is PSPACE-complete under any rule, even when restricted to perfect graphs [7]. Indeed, for this problem, the rules TAR, TJ, and TS have all been shown to be equivalent from the viewpoint of polynomial-time solvability. It is also known that reconfiguration of cliques can be solved in polynomial time for several well-known graph classes [7]. Wasa et al. [12] considered the induced variant under the TJ and TS rules with the property Π being “a graph is a tree.” They showed that this variant under each of the TJ and TS rules is PSPACE-complete, and is W[1]-hard when parameterized by both the size of a solution and the length of a reconfiguration sequence. They also gave a fixed-parameter algorithm when parameterized by both the size of a solution and the maximum degree of an input graph, under both the TJ and TS rules. In closely related work, Mouawad et al. [9] considered the induced variants of subgraph reconfiguration under the TAR rule with the properties Π being either “a graph is a forest” or “a graph is bipartite.” They showed that these variants are W[1]-hard when parameterized by the size of a solution plus the length of a reconfiguration sequence. 1.3 Our contributions In this paper, we study the complexity of subgraph reconfiguration under the TJ and TS rules. (Our results are summarized in Table 2, together with known results, where an (i, j)-biclique is a complete bipartite graph with the bipartition of i vertices and j vertices.) As mentioned above, because we consider the TJ and TS rules, it suffices to deal with subgraphs having the same number of vertices or edges. Subgraphs of the same size may be isomorphic for certain properties Π, such as “a graph is a path” and “a graph is a clique,” because there is only one choice of a path or a clique of a particular size. On the other hand, for the property “a graph is a tree,” there are several choices of trees of a particular size. (We will show an example in Section 3 with Figure 4.) As shown in Table 2, we systematically clarify the complexity of subgraph reconfiguration for several fundamental graph properties. In particular, we show that the edge variant under the TJ rule is computationally intractable for the property “a graph is a path” but tractable for the 3 Property Π path cycle tree Table 2: Previous and new results Edge Induced variant variant NP-hard (TJ) PSPACE-c. (TJ, TS) [Theorem 2] [Theorems 7, 9] P (TJ, TS) PSPACE-c. (TJ, TS) [Theorem 3] [Theorems 8, 9] P (TJ) PSPACE-c. (TJ, TS) [Theorem 6] [12] (i, j)-biclique P (TJ, TS) [Theorem 5] clique P (TJ, TS) [Theorem 4] diameter two any property XP for solution size (TJ, TS) [Theorem 1] PSPACE-c. for i = j (TJ) PSPACE-c. for fixed i (TJ) [Corollary 1, Theorem 12] PSPACE-c. (TJ, TS) [7] PSPACE-c. (TS) [Theorem 15] XP for solution size (TJ, TS) [Theorem 1] Spanning variant PSPACE-c. (TJ, TS) [Theorems 7, 9] PSPACE-c. (TJ, TS) [Theorems 8, 9] P (TJ) PSPACE-c. (TS) [Theorems 11, 10] NP-hard for i = j (TJ) P for fixed i (TJ) [Theorems 13, 14] PSPACE-c. (TJ, TS) [7] PSPACE-c. (TS) [Theorem 15] XP for solution size (TJ, TS) [Theorem 1] property “a graph is a tree.” This implies that the computational (in)tractability does not follow directly from the inclusion relationship of graph classes required as the properties Π; one possible explanation is that the path property implies a specific graph, whereas the tree property allows several choices of trees, making the problem easier. 1.4 Preliminaries Although we assume throughout the paper that an input graph G is simple, all our algorithms can be easily extended to graphs having multiple edges. We denote by (G, Vs , Vt ) an instance of a spanning variant or an induced variant whose input graph is G and source and target solutions are vertex subsets Vs and Vt of G. Similarly, we denote by (G, Es , Et ) an instance of the edge variant. We may assume without loss of generality that |Vs | = |Vt | holds for the spanning and induced variants, and |Es | = |Et | holds for the edge variant; otherwise, the answer is clearly no since under both the TJ and TS rules, all solutions must be of the same size. 2 General algorithm In this section, we give a general XP algorithm when the size of a solution (that is, the size of a vertex or edge subset that represents a subgraph) is taken as the parameter. For notational convenience, we simply use element to represent a vertex (or an edge) for the spanning and induced variants (resp., the edge variant), and candidate to represent a set of elements (which does not necessarily satisfy the property Π.) Furthermore, we define the size of a given graph as the number of elements in the graph. Theorem 1. Let Π be any graph structure property, and let f (k) denote the time to check if a candidate of size k satisfies Π. Then, all of the spanning, induced, and edge variants under the TJ or TS rules can be solved in time O(nk+1 k + nk f (k)), where n is the size of a given graph and k is the size of a source (and target) solution. Furthermore, a shortest reconfiguration sequence between source and target solutions can be found in the same time bound, if it exists. Proof. Our claim is that the reconfiguration graph can be constructed in the stated time. Since a given source solution is of size k, it suffices to deal only with candidates of size exactly k. For 4 Ps Pt e1s s1 ens s2 sn e1t t1 G sn+1 tn e1 x t en+1 ent t2 s en+1 e2 v e3 s t w tn+1 Figure 3: Reduction to the edge variant under the TJ rule for the property “a graph is a path.” a given graph, the total number of possible candidates of size k is O(nk ). For each candidate, we can check in time f (k) whether it satisfies Π. Therefore, we can construct the node set of the reconfiguration graph in time O(nk f (k)). We then obtain the edge set of the reconfiguration graph. Notice that each node in the reconfiguration graph has O(nk) adjacent nodes, because we can replace only a single element at a time. Therefore, we can find all pairs of adjacent nodes in time O(nk+1 k). In this way, we can construct the reconfiguration graph in time O(nk+1 k + nk f (k)) in total. The reconfiguration graph consists of O(nk ) nodes and O(nk+1 k) edges. Therefore, by breadthfirst search starting from the node representing a given source solution, we can determine in time O(nk+1 k) whether or not there exists a reconfiguration sequence between two nodes representing the source and target solutions. Notice that if a desired reconfiguration sequence exists, then the breadth-first search finds a shortest one. 3 Edge variants In this section, we study the edge variant of subgraph reconfiguration for the properties of being paths, cycles, cliques, bicliques, and trees. We first consider the property “a graph is a path” under the TJ rule. Theorem 2. The edge variant of subgraph reconfiguration under the TJ rule is NP-hard for the property “a graph is a path.” Proof. We give a polynomial-time reduction from the Hamiltonian path problem. Recall that a Hamiltonian path in a graph G is a path that visits each vertex of G exactly once. Given a graph G and two vertices s, t ∈ V (G) of G, the NP-complete problem Hamiltonian path is to determine whether or not G has a Hamiltonian path which starts from s and ends in t [3]. For an instance (G, s, t) of Hamiltonian path, we construct a corresponding instance (G′ , Es , Et ) of our problem, as follows. (See also Figure 3.) Let n = |V (G)|. We first add two new vertices v and x to G with two new edges e1 = xv and e2 = vs. We then add two paths Ps = hs1 , s2 , . . . , sn+1 , xi and Pt = ht1 , t2 , . . . , tn+1 , xi, where s1 , s2 , . . . , sn+1 and t1 , t2 , . . . , tn+1 are distinct new vertices. Each of Ps and Pt consists of n+1 edges; we denote by es1 , es2 , . . . , esn+1 the edges s1 s2 , s2 s3 , . . . , sn+1 x in Ps , respectively, and by et1 , et2 , . . . , etn+1 the edges t1 t2 , t2 t3 , . . . , tn+1 x in Pt , respectively. We finally add a new vertex w with an edge e3 = tw, completing the construction of G′ . We then set Es = {es1 , es2 , . . . , esn+1 , e1 , e2 } and Et = {et1 , et2 , . . . , etn+1 , e1 , e2 }; these edge subsets clearly form paths in G′ . We have thus constructed our corresponding instance (G′ , Es , Et ) in polynomial time. We now prove that an instance (G, s, t) of Hamiltonian path is a yes-instance if and only if the corresponding instance (G′ , Es , Et ) is a yes-instance. To prove the only-if direction, we first suppose that G has a Hamiltonian path P starting from s and ending in t. Then, we construct an actual reconfiguration sequence from Es to Et using the edges in P . Notice that P consists of n − 1 edges. Thus, we first move the n − 1 edges es1 , es2 , . . . , esn−1 in Es to the edges in P one by one, and then move esn to e3 . Next, we move esn+1 to etn+1 , and then move the edges in E(P ) ∪ {e3 } to etn , etn−1 , . . . , et1 one by one. By the construction of G′ , we know that each of the intermediate edge subsets forms a path in G′ , as required. 5 We now prove the if direction by supposing that there exists a reconfiguration sequence hEs = E0 , E1 , . . . , Eℓ = Et i. Let Eq be the first edge subset in the sequence such that E(Pt ) ∩ Eq 6= ∅; we claim that Eq contains a Hamiltonian path in G. First, notice that the edge in E(Pt ) ∩ Eq is etn+1 ; otherwise the subgraph formed by Eq is disconnected. Since |Eq | = |Es | = n + 3 and |E(Ps )| = n + 1, we can observe that Eq contains no edge in Ps ; otherwise the degree of x would be three, or Eq would form a disconnected subgraph. Therefore, the n + 2 edges in Eq \ {etn+1 } must be chosen from E(G) ∪ {e1 , e2 , e3 }. Since |V (G)| = n and Eq must form a path in G′ , we know that Eq \ {etn+1 } consists of e1 , e2 , e3 and n − 1 edges in G. Thus, Eq \ {etn+1, e1 , e2 , e3 } forms a Hamiltonian path in G starting from s and ending in t, as required. We now consider the property “a graph is a cycle,” as follows. Theorem 3. The edge variant of subgraph reconfiguration under each of the TJ and TS rules can be solved in linear time for the property “a graph is a cycle.” Proof. Let (G, Es , Et ) be a given instance. We claim that the reconfiguration graph is edgeless, in other words, no feasible solution can be transformed at all. Then, the answer is yes if and only if Es = Et holds; this condition can be checked in linear time. Let E ′ be any feasible solution of G, and consider a replacement of an edge e− ∈ E ′ with an edge e+ other than e− . Let u, v be the endpoints of e− . When we remove e− from E ′ , the resulting edge subset E ′ \{e} forms a path whose ends are u and v. Then, to ensure that the candidate forms a cycle, we can choose only e− = uv as e+ . This contradicts the assumption that e+ 6= e− . The same arguments hold for the property “a graph is a clique,” and we obtain the following theorem. We note that, for this property, both induced and spanning variants (i.e., when solutions are represented by vertex subsets) are PSPACE-complete under any rule [7]. Theorem 4. The edge variant of subgraph reconfiguration under each of the TJ and TS rules can be solved in linear time for the property “a graph is a clique.” We next consider the property “a graph is an (i, j)-biclique,” as follows. Theorem 5. The edge variant of subgraph reconfiguration under each of the TJ and TS rules can be solved in polynomial time for the property “a graph is an (i, j)-biclique” for any pair of positive integers i and j. Proof. We may assume without loss of generality that i ≤ j holds. We prove the theorem in the following three cases: Case 1: i = 1 and j ≤ 2; Case 2: i, j ≥ 2; and Case 3: i = 1 and j ≥ 3. We first consider Case 1, which is the easiest case. In this case, any (1, j)-biclique has at most two edges. Therefore, by Theorem 1 we can conclude that this case is solvable in polynomial time. We then consider Case 2. We show that (G, Es , Et ) is a yes-instance if and only if Es = Et holds. To do so, we claim that the reconfiguration graph is edgeless, in other words, no feasible solution can be transformed at all. To see this, because i, j ≥ 2, notice that the removal of any edge e in an (i, j)-biclique results in a bipartite graph with the same bipartition of i vertices and j vertices. Therefore, to obtain an (i, j)-biclique by adding a single edge, we must add back the same edge e. We finally deal with Case 3. Notice that a (1, j)-biclique is a star with j leaves, and its center vertex is of degree j ≥ 3. Then, we claim that (G, Es , Et ) is a yes-instance if and only if the center vertices of stars represented by Es and Et are the same. The if direction clearly holds, because we can always move edges in Es \ Et into ones in Et \ Es one by one. We thus prove the only-if direction; indeed, we prove the contrapositive, that is, the answer is no if the center vertices of stars represented by Es and Et are different. Consider such a star Ts formed by Es . Since Ts has j ≥ 3 leaves, the removal of any edge in Es results in a star having j − 1 ≥ 2 leaves. Therefore, to ensure that each intermediate solution is a star with j leaves, we can add only an edge of G which is incident to the center of Ts . Thus, we cannot change the center vertex. 6 Es = E0 E1 E2 = Et Figure 4: Reconfiguration sequence hE0 , E1 , E2 i in the edge variant under the TJ rule with the property “a graph is a tree.” In this way, we have proved Theorem 5. Although Case 1 takes non-linear time, our algorithm can be easily improved (without using Theorem 1) so that it runs in linear time. We finally consider the property “a graph is a tree” under the TJ rule. As we have mentioned in the introduction, for this property, there are several choices of trees even of a particular size, and a reconfiguration sequence does not necessarily consist of isomorphic trees (see Figure 4). This “flexibility” of subgraphs may yield the contrast between Theorem 2 for the path property and the following theorem for the tree property. Theorem 6. The edge variant of subgraph reconfiguration under the TJ rule can be solved in linear time for the property “a graph is a tree.” Proof. Suppose that (G, Es , Et ) is a given instance. We may assume without loss of generality that |Es | = |Et | ≥ 2; otherwise |Es | = |Et | ≤ 1 holds, and hence the instance is trivially a yes-instance. We will prove below that any instance with |Es | = |Et | ≥ 2 is a yes-instance if and only if all the edges in Es and Et are contained in the same connected component of G. Note that this condition can be checked in linear time. We first prove the only-if direction of our claim. Since |Es | = |Et | ≥ 2 and subgraphs always must retain a tree structure (more specifically, they must be connected graphs), observe that we can exchange edges only in the same connected component of G. Thus, the only-if direction follows. To complete the proof, it suffices to prove the if direction of our claim. For notational convenience, for any feasible solution Ei we denote by Ti the tree represented by Ei , and by Vi the vertex set of Ti . In this direction, we consider the following two cases: (a) Vs ∩ Vt = ∅, and (b) Vs ∩ Vt 6= ∅. We consider case (a), that is, Vs ∩ Vt = ∅. Since Ts and Tt are contained in one connected component of G, there exists a path hv0 , v1 , . . . , vℓ i in G such that v0 ∈ Vs , vℓ ∈ Vt and vi ∈ / Vs ∪ Vt for all i ∈ {1, 2, . . . , ℓ − 1}. Since Ts is a tree, it has at least two degree-one vertices. Let vs be any degree-one vertex in Vs \ {v0 }, and let es be the leaf edge of Ts incident to vs . Then, we can exchange es with v0 v1 , and obtain another tree represented by the resulting edge subset (Es ∪ {v0 v1 }) \ {es }. By repeatedly applying this operation along the path hv1 , v2 , . . . , vℓ i, we can obtain a solution Ek such that Vk ∩ Vt = {vℓ } 6= ∅; this case will be considered below. We finally consider case (b), that is, Vs ∩ Vt 6= ∅. Consider the graph (Vs ∩ Vt , Es ∩ Et ). Then, (Vs ∩ Vt , Es ∩ Et ) is a forest, and let G′ = (V ′ , E ′ ) be a connected component (i.e., a tree) of (Vs ∩ Vt , Es ∩ Et ) whose edge set is of maximum size. We now prove that there is a reconfiguration sequence between Es and Et by induction on k = |Es \ E ′ | = |Et \ E ′ |. If k = 0, then Es = E ′ = Et and hence the claim holds. We thus consider the case where k > 0 holds. Since G′ is a proper subtree of Tt , there exists at least one edge et in Et \ E ′ such that one endpoint of et is contained in V ′ and the other is not. We claim that there exists an edge es in Es \ E ′ which can be moved into et , that is, the subgraph represented by the resulting edge subset (Es ∪ {et }) \ {es } forms a tree. If both endpoints of et are contained in Vs (not just V ′ ), Es ∪ {et } contains a cycle; let C ⊆ Es ∪ {et } be the edge set of the cycle. Since the subgraph Tt has no cycle, there exists at least one edge in C \ Et , and we choose one of them as es . On the other hand, if just one endpoint of et is contained in Vs , then we choose a leaf edge of Ts in Es \ E ′ as es . Note that there exists such a leaf edge since G′ is a proper subtree of Ts . From the choice of es and et , we know the subgraph represented by the resulting edge subset (Es ∪ {et }) \ {es } forms a tree; let Ek = (Es ∪ {et }) \ {es }. Furthermore, since Ek ∩ Et includes E ′ ∪ {et } and the subgraph formed by E ′ ∪ {et } is connected, the subgraph 7 G G s2 s1 s t t1 s1 t2 (a) s t t1 (b) Figure 5: Reductions for the properties (a) “a graph is a path,” and (b) “a graph is a cycle.” formed by Ek ∩ Et has a connected component whose edge set has size at least |E ′ | + 1. Therefore, we can conclude that Ek is reconfigurable into Et by the induction hypothesis. 4 Induced and spanning variants In this section, we deal with the induced and spanning variants where subgraphs are represented as vertex subsets. Most of our results for these variants are hardness results, except for Theorems 11 and 14. 4.1 Path and cycle In this subsection, we show that both induced and spanning variants under the TJ or TS rules are PSPACE-complete for the properties “a graph is a path” and “a graph is a cycle.” All proofs in this subsection make use of reductions that employ almost identical constructions. Therefore, we describe the detailed proof for only one case, and give proof sketches for the other cases. We give polynomial-time reductions from the shortest path reconfiguration problem, which can be seen as a subgraph reconfiguration problem, defined as follows [1]. For a simple, unweighted, and undirected graph G and two distinct vertices s, t of G, shortest path reconfiguration is the induced (or spanning) variant of subgraph reconfiguration under the TJ rule for the property “a graph is a shortest st-path.” Notice that there is no difference between the induced variant and the spanning variant for this property, because any shortest path in a simple graph forms an induced subgraph. This problem is known to be PSPACE-complete [1]. Let d be the (shortest) distance from s to t in G. For each i ∈ {0, 1, . . . , d}, we denote by Li ⊆ V (G) the set of vertices that lie on a shortest st-path at distance i from s. Therefore, we have L0 = {s} and Ld = {t}. We call each Li a layer. Observe that any shortest st-path contains exactly one vertex from each layer, and we can assume without loss of generality that G has no vertex which does not belong to any layer. We first give the following theorem. Theorem 7. For the property “a graph is a path,” the induced and spanning variants of subgraph reconfiguration under the TJ rule are both PSPACE-complete on bipartite graphs. Proof. Observe that these variants are in PSPACE. Therefore, we construct a polynomial-time reduction from shortest path reconfiguration. Let (G, Vs , Vt ) be an instance of shortest path reconfiguration. Since any shortest stpath contains exactly one vertex from each layer, we can assume without loss of generality that G has no edge joining two vertices in the same layer, that is, each layer Li forms an independent set in G. Then, G is a bipartite graph. From (G, Vs , Vt ), we construct a corresponding instance (G′ , Vs′ , Vt′ ) for the induced and spanning variants; note that we use the same reduction for both variants. Let G′ be the graph obtained from G by adding four new vertices s1 , s2 , t1 , t2 which are connected with four new edges s2 s1 , s1 s, tt1 , t1 t2 . (See Figure 5(a).) Note that G′ is also bipartite. 8 We then set Vs′ = Vs ∪ {s1 , s2 , t1 , t2 } and Vt′ = Vt ∪ {s1 , s2 , t1 , t2 }. Since each of Vs and Vt induces a shortest st-path in G, each of Vs′ and Vt′ is a feasible solution to both variants. This completes the polynomial-time construction of the corresponding instance. We now give the key lemma for proving the correctness of our reduction. Lemma 1. Let V ′ ⊆ V (G′ ) be any solution for the induced or spanning variant which is reachable by a reconfiguration sequence from Vs′ (or Vt′ ) under the TJ rule. Then, V ′ satisfies the following two conditions: (a) s2 , s1 , s, t, t1 , t2 ∈ V ′ ; and (b) V ′ contains exactly one vertex from each layer of G. Proof. We first prove that condition (a) is satisfied. Notice that both Vs′ and Vt′ satisfy this condition. Because V ′ is reconfigurable from Vs′ or Vt′ , it suffices to show that we cannot move any of s2 , s1 , s, t, t1 , t2 in a reconfiguration sequence starting from Vs′ or Vt′ . Suppose for sake of contradiction that we can move at least one of s2 , s1 , s, t, t1 , t2 . Then, the first removed vertex v ∈ {s2 , s1 , s, t, t1 , t2 } must be either s2 or t2 ; otherwise the resulting subgraph would be disconnected. Let w be the vertex with which we exchanged v. Then, w ∈ V (G′ ) \ {s2 , s1 , s, t, t1 , t2 } = V (G) \ {s, t}. Therefore, the resulting vertex subset cannot induce a path, and hence it cannot be a solution for the induced variant. Similarly, the induced subgraph cannot contain a spanning path, and hence it cannot be a solution for the spanning variant. We next show that condition (b) is satisfied. Recall that d denotes the number of edges in a shortest st-path in G. Then, we have |V ′ | = |Vs′ | = |Vt′ | = d+ 5. By condition (a), we know that V ′ contains s2 , s1 , t1 , t2 , and hence in both induced and spanning variants, s and t must be connected by a path formed by d + 1 vertices in V ′ \ {s2 , s1 , t1 , t2 }. Since the length of this st-path is d, this path is shortest and hence V ′ must contain exactly one vertex from each of d + 1 layers of G. Consider any vertex subset V ′′ ⊆ V (G′ ) which satisfies conditions (a) and (b) of Lemma 1; note that V ′′ is not necessarily a feasible solution. Then, these conditions ensure that V ′′ \ {s2 , s1 , t1 , t2 } forms a shortest st-path in G if and only if the subgraph represented by V ′′ induces a path in G′ . Thus, an instance (G, Vs , Vt ) of shortest path reconfiguration is a yes-instance if and only if the corresponding instance (G′ , Vs′ , Vt′ ) of the induced or spanning variant is a yes-instance. This completes the proof of Theorem 7. Similar arguments give the following theorem. Theorem 8. Both the induced and spanning variants of subgraph reconfiguration under the TJ rule are PSPACE-complete for the property “a graph is a cycle.” Proof. Our reduction is the same as in the proof of Theorem 7 except for the following point (see also Figure 5(b)): instead of adding four new vertices, we connect s and t by a path of length three with two new vertices s1 and t1 . Then, the same arguments hold as the proof of Theorem 7. 4.2 Path and cycle under the TS rule We now consider the TS rule. Notice that, in the proofs of Theorems 7 and 8, we exchange only vertices contained in the same layer. Since any shortest st-path in a graph G contains exactly one vertex from each layer, we can assume without loss of generality that each layer Li of G forms a clique. Then, the same reductions work for the TS rule, and we obtain the following theorem. Theorem 9. Both the induced and spanning variants of subgraph reconfiguration under the TS rule are PSPACE-complete for the properties “a graph is a path” and “a graph is a cycle.” 4.3 Tree Wasa et al. [12] showed that the induced variant under the TJ and TS rules is PSPACE-complete for the property “a graph is a tree.” In this subsection, we show that the spanning variant for this 9 property is also PSPACE-complete under the TS rule, while it is linear-time solvable under the TJ rule. We first note that our proof of Theorem 9 yields the following theorem. Theorem 10. The spanning variant of subgraph reconfiguration under the TS rule is PSPACE-complete for the property “a graph is a tree.” Proof. We claim that the same reduction as in Theorem 9 applies. Let V ′ ⊆ V (G′ ) be any solution which is reachable by a reconfiguration sequence from Vs′ (or Vt′ ) under the TS rule, where (G′ , Vs′ , Vt′ ) is the corresponding instance for the spanning variant, as in the reduction. Then, the TS rule ensures that s2 , s1 , s, t, t1 , t2 ∈ V ′ holds, and V ′ contains exactly one vertex from each layer of G. Therefore, any solution forms a path even for the property “a graph is a tree,” and hence the theorem follows. In contrast to Theorem 10, the spanning variant under the TJ rule is solvable in linear time. We note that the reduction in Theorem 10 does not work under the TJ rule, because the tokens on s2 and t2 can move (jump) and hence there is no guarantee that a solution forms a path for the property “a graph is a tree.” Theorem 11. The spanning variant of subgraph reconfiguration under the TJ rule can be solved in linear time for the property “a graph is a tree.” Suppose that (G, Vs , Vt ) is a given instance. We assume that |Vs | = |Vt | ≥ 2 holds; otherwise it is a trivial instance. Then, Theorem 11 can be obtained from the following lemma. Lemma 2. (G, Vs , Vt ) with |Vs | = |Vt | ≥ 2 is a yes-instance if and only if Vs and Vt are contained in the same connected component of G. Proof. We first prove the only-if direction. Since |Vs | = |Vt | ≥ 2 and the property requires subgraphs to be connected, the subgraph induced by any feasible solution contains at least one edge. Because we can exchange only one vertex at a time and the resulting subgraph must retain connected, Vs and Vt are contained in the same connected component of G if (G, Vs , Vt ) is a yesinstance. Next, we prove the if direction for the case where |Vs | = |Vt | = 2. In this case, G[Vs ] and G[Vt ] consist of single edges, say es and et , respectively. Since Vs and Vt are contained in the same connected component of G, there is a path in G between es and et . Thus, we can exchange vertices along the path, and obtain a reconfiguration sequence from Vs to Vt . In this way, the if direction holds for this case. Finally, we prove the if direction for the remaining case, that is, |Vs | = |Vt | ≥ 3. Consider any spanning trees Ts of G[Vs ] and Tt of G[Vt ]. Since |Vs | = |Vt | ≥ 3, each of Ts and Tt has at least two edges. Then, if we regard (G, E(Ts ), E(Tt )) as an instance of the edge variant under the TJ rule for the property “a graph is a tree,” we know from the proof of Theorem 6 that it is a yes-instance. Thus, there exists a reconfiguration sequence E = hE(Ts ) = E0 , E1 , . . . , Eℓ = E(Tt )i of edge subsets under the TJ rule. We below show that, based on E, we can construct a reconfiguration sequence between Vs and Vt for the spanning variant under the TJ rule. For each Ei in E, let Vi be the vertex set of the tree represented by Ei . Notice that Vi is a feasible solution for the spanning variant, and that V0 = Vs and Vℓ = Vt hold. We claim that the sequence hVs = V0 , V1 , . . . , Vℓ = Vt i of vertex subsets is a reconfiguration sequence for the spanning variant under the TJ rule (after removing redundant vertex subsets if needed). To show this, it suffices to prove |Vi \ Vi−1 | = |Vi−1 \ Vi | ≤ 1 for all i ∈ {1, 2, . . . , ℓ}. Suppose for the sake of contradiction that there exists Vi such that |Vi \Vi−1 | ≥ 2 holds. (See also Figure 6.) Since |Ei | ≥ 2 and |Ei \ Ei−1 | = 1 hold, we have Ei ∩ Ei−1 6= ∅ and hence Vi ∩ Vi−1 6= ∅. Then there is at least one edge e = uv in Ei \ Ei−1 joining a vertex u ∈ Vi \ Vi−1 and v ∈ Vi ∩ Vi−1 , because Ei must form a connected subgraph. Since |Vi \ Vi−1 | ≥ 2, there is another vertex u′ 6= u in Vi \ Vi−1 , and there is an edge e′ incident to u′ . Note that e 6= e′ . Furthermore, we know that e′ ∈ Ei \ Ei−1 because u′ ∈ Vi \ Vi−1 . Therefore, we have e, e′ ∈ Ei \ Ei−1 , which contradicts the fact that |Ei \ Ei−1 | = 1 holds. 10 Vi-1 Vi v e u e’ u’ Figure 6: Illustration for the proof of Lemma 2. L R G ... Figure 7: Reduction for the property “a graph is an (i, j)-biclique” for any fixed i ≥ 1. 4.4 Biclique For the property “a graph is an (i, j)-biclique,” we show that the induced variant under the TJ rule is PSPACE-complete even if i = j holds, or i is fixed. On the other hand, the spanning variant under the TJ rule is NP-hard even if i = j holds, while it is polynomial-time solvable when i is fixed. We first give the following theorem for a fixed i ≥ 1. Theorem 12. For the property “a graph is an (i, j)-biclique,” the induced variant of subgraph reconfiguration under the TJ rule is PSPACE-complete even for any fixed integer i ≥ 1. Proof. We give a polynomial-time reduction from the maximum independent set reconfiguration problem [13], which can be seen as a subgraph reconfiguration problem. The maximum independent set reconfiguration problem is the induced variant for the property “a graph is edgeless” such that two given independent sets are maximum. Note that, because we are given maximum independent sets, there is no difference between the TJ and TS rules for this problem. This problem is known to be PSPACE-complete [13]. Suppose that (G, Vs , Vt ) is an instance of maximum independent set reconfiguration. We now construct a corresponding instance (G′ , Vs′ , Vt′ ) of the induced variant under the TJ rule for the property “a graph is an (i, j)-biclique,” where i is any fixed positive integer. (See also Figure 7.) Let L and R be distinct sets of new vertices such that |L| = i and |R| = 1. The vertex set of G′ is defined as V (G′ ) = V (G) ∪ L ∪ R, and the edge set of G′ as E(G′ ) = E(G) ∪ {uv | u ∈ V (G), v ∈ L} ∪ {vw | v ∈ L, w ∈ R}, that is, new edges are added so that there are edges between each vertex of L and each vertex of V (G) ∪ R. Let Vs′ = Vs ∪ L ∪ R and Vt′ = Vt ∪ L ∪ R. Since L, R, Vs and Vt are all independent sets in G′ , both Vs′ and Vt′ form (i, j)-bicliques, where i = |L| and j = |Vs ∪ R| = |Vt ∪ R|. We have now completed the construction of our corresponding instance, which can be accomplished in polynomial time. Because each vertex in L is connected to all vertices in V (G), a vertex subset V ′′ ⊆ V (G′ ) cannot form a bipartite graph (and hence an (i, j)-biclique) if V ′′ ∩ L 6= ∅ or if V ′′ contains two vertices joined by an edge in G. In addition, we cannot move any token placed on L ∪ R onto a vertex in V (G) because both Vs and Vt are maximum independent sets of G. Note that, in the case of i = 1, there may exist a vertex u in V (G) which is adjacent to all vertices in Vs or in Vt . However, u is not adjacent to the vertex in R, and hence the token placed on the vertex in L cannot be moved to u in this case, either. Therefore, for any feasible solution V ′ ⊆ V (G′ ) which is reconfigurable from Vs′ or Vt′ under the TJ rule, the vertex subset V ′ ∩ V (G) forms a maximum independent set of G. Thus, an instance (G, Vs , Vt ) of maximum independent set reconfiguration is a 11 G G1 G2 ... ... ... ... B1 ... ... A1 A2 B2 B A Figure 8: Reduction for the property “a graph is a (k, k)-biclique.” yes-instance if and only if our corresponding instance (G′ , Vs′ , Vt′ ) is a yes-instance. The corresponding instance (G′ , Vs′ , Vt′ ) constructed in the proof of Theorem 12 satisfies i = j if we set i = |Vs | + 1 = |Vt | + 1. Therefore, we can obtain the following corollary. Corollary 1. For the property “a graph is an (i, j)-biclique,” the induced variant of subgraph reconfiguration under the TJ rule is PSPACE-complete even if i = j holds. We next give the following theorem. Theorem 13. For the property “a graph is an (i, j)-biclique,” the spanning variant of subgraph reconfiguration under the TJ rule is NP-hard even if i = j holds. Proof. We give a polynomial-time reduction from the balanced complete bipartite subgraph problem, defined as follows [3]. Given a bipartite graph G and a positive integer k, the balanced complete bipartite subgraph problem is to determine whether or not G contains a (k, k)biclique as a subgraph; this problem is known to be NP-hard [3]. Suppose that (G, k) is an instance of balanced complete bipartite subgraph, where G is a bipartite graph with bipartition (A, B). Then, we construct a corresponding instance (G′ , Vs , Vt ) of the spanning variant under the TJ rule for the property “a graph is a (k, k)-biclique.” We first construct a graph G′ . (See Figure 8.) We add to G two new (k, k)-bicliques G1 and G2 ; let (A1 , B1 ) be the bipartition of G1 , and (A2 , B2 ) be that of G2 . We then add edges between any two vertices x ∈ B1 and y ∈ A, and between any two vertices x ∈ B and y ∈ A2 . Therefore, G′ [B1 ∪ A] and G′ [B ∪ A2 ] are bicliques in G′ . This completes the construction of G′ . We then set Vs = V (G1 ) and Vt = V (G2 ). Then, Vs and Vt are solutions, since G[Vs ] and G[Vt ] contain (k, k)-bicliques G1 and G2 , respectively. In this way, the corresponding instance can be constructed in polynomial time. By the construction of G′ , any reconfiguration sequence between Vs = V (G1 ) and Vt = V (G2 ) must pass through a (k, k)-biclique of G. Therefore, the theorem follows. We now give a polynomial-time algorithm solving the spanning variant for a fixed constant i ≥ 1. Theorem 14. For the property “a graph is an (i, j)-biclique,” the spanning variant of subgraph reconfiguration under the TJ rule is solvable in polynomial time when i ≥ 1 is a fixed constant. Proof. We give such an algorithm. We assume without loss of generality that j > i; otherwise both i and j are fixed constants, and hence such a case can be solved in polynomial time by Theorem 1. We will refer to the i vertices in the bounded-size part of the biclique as hubs, and the j vertices in the other part as terminals. Let H ⊆ V (G) be an arbitrary vertex subset such that |H| = i. We denote by C(H) the set of all common neighbors of H in G. Notice that, if |C(H)| ≥ j, then the subgraph represented by H ∪ C(H) contains at least one (i, j)-biclique whose hub set is H. We denote by S(H) the set of all solutions that contain (i, j)-bicliques with the hub set H. We know that S(H) = ∅ if |C(H)| < j; otherwise H ∪ T is in S(H) for any subset T ⊆ C(H) such that |T | = j. It should be noted that a solution in the spanning variant is simply a vertex subset 12 V ′ of V (G), and there is no restriction on how to choose a hub set from V ′ . (For example, if a solution V ′ induces a clique of size five, then there are ten ways to choose a hub set from V ′ for (2, 3)-bicliques.) Therefore, S(H) ∩ S(H ′ ) 6= ∅ may hold for distinct hub sets H, H ′ . We describe two key observations in the following. The first one is that for a hub set H, any two solutions Va , Vb ∈ S(H) are reconfigurable because we can always move vertices in Va \ Vb into ones in Vb \ Va one by one. The second one is that for any two distinct hub sets Ha and Hb , if there exist Va ∈ S(Ha ) and Vb ∈ S(Hb ) such that |Va \ Vb | ≤ 1 and |Vb \ Va | ≤ 1 (this means that Va and Vb are reconfigurable by one reconfiguration step, or Va = Vb ), then all pairs of solutions in S(Ha ) ∪ S(Hb ) are reconfigurable. Based on these observations, we construct an auxiliary graph A for a given instance (G, Vs , Vt ), as follows. Each node in A corresponds to a set H of i vertices (hubs) in the input graph G such that |C(H)| ≥ j; we represent the node in A simply by the corresponding hub set H. Two nodes Ha and Hb are adjacent in A if there exist Va ∈ S(Ha ) and Vb ∈ S(Hb ) such that |Va \ Vb | ≤ 1 and |Vb \ Va | ≤ 1. Let Hs and Ht be any two nodes such that Vs ∈ S(Hs ) and Vt ∈ S(Ht ), respectively. Then, we claim that there is a reconfiguration sequence between Vs and Vt if and only if there is a path in A between Hs and Ht . We first suppose that there is a path P in A between Hs and Ht . We know that any two consecutive nodes Ha and Hb in P are adjacent in A. Then, as we mentioned above, all pairs of solutions in S(Ha ) ∪ S(Hb ) are reconfigurable. Since Vs ∈ S(Hs ) and Vt ∈ S(Ht ), we can thus conclude that there is a reconfiguration sequence between Vs and Vt . We now suppose that there exists a reconfiguration sequence R between Vs and Vt . For each solution Va in R except for Vs and Vt , we choose an arbitrary node Ha in A which satisfies Va ∈ S(Ha ). Consider any two consecutive solutions Va and Vb in R. Then, by the construction of A, the chosen nodes Ha and Hb are adjacent in A (or sometimes Ha = Hb ) because |Va \ Vb | = 1 and |Vb \ Va | = 1. In this way, we can ensure the existence of a desired path in A. The running time of the algorithm depends on the size i of a hub set. Let n be the number of vertices in G. The size of the node set of A is in O(ni ). For any two nodes Ha and Hb in A, we can determine in O(n) time whether there is an edge between them by checking that all of the following four conditions hold or not: (a) |(Ha ∪ C(Ha )) ∩ (Hb ∪ C(Hb ))| ≥ i + j − 1; (b) |Ha \ (Hb ∪ C(Hb ))| ≤ 1; (c) |Hb \ (Ha ∪ C(Ha ))| ≤ 1; and (d) |Ha ∪ Hb | ≤ i + j + 1. Note that, since we have assumed that i < j, condition (d) is always satisfied. Therefore, we take O(n2i+1 ) time to construct A and to check whether the nodes corresponding to Vs and Vt are connected. 4.5 Diameter-two graph In this subsection, we consider the property “a graph has diameter at most two.” Note that the induced and spanning variants are the same for this property. Theorem 15. Both induced and spanning variants of subgraph reconfiguration under the TS rule are PSPACE-complete for the property “a graph has diameter at most two.” Proof. Since the induced variant and the spanning variant are the same for this property, it suffices to show the PSPACE-hardness only for the induced variant. We give a polynomial-time reduction from the clique reconfiguration problem, which is the induced variant (also the spanning variant) of subgraph reconfiguration for the property “a graph is a clique.” This problem is known to be PSPACE-complete under both the TJ and TS rules [7], and we give a reduction from the problem under the TS rule. Suppose that (G, Vs , Vt ) is an instance of clique reconfiguration under the TS rule such that |Vs | = |Vt | ≥ 2; otherwise it is a trivial instance. Then, we construct a corresponding instance (G′ , Vs′ , Vt′ ) of the induced variant under the TS rule. Let V (G) = {v1 , v2 , . . . , vn }, where n = |V (G)|. We form G′ by making two copies of G and adding edges between corresponding 13 l2 v2 v1 l1 l4 v4 l6 v6 r5 l5 v5 r4 r3 l3 v3 r2 = rs r1 (a) G r6 (b) G’ Figure 9: Reduction for the property “a graph has diameter at most two.” The vertices of Vs in G and of Vs′ in G′ are depicted by gray vertices, where rs = r2 . vertices of the two graphs. (See Figure 9.) More formally, the vertex set V (G′ ) is defined as V (G′ ) = L ∪ R, where L = {li | vi ∈ V (G)} and R = {ri | vi ∈ V (G)}, and the edge set E(G′ ) is defined as E(G′ ) = El ∪ Er ∪ Ec , where El = {li lj | vi vj ∈ E(G)}, Er = {ri rj | vi vj ∈ E(G)} and Ec = {li ri | vi ∈ V (G)}. For each i ∈ {1, 2, . . . , n}, we call li and ri corresponding vertices, and an edge joining corresponding vertices a connecting edge. We then construct Vs′ and Vt′ . We say that a vertex v ′ in a vertex subset V ′ ⊆ V (G′ ) is exposed in V ′ if the corresponding vertex of v ′ does not belong to V ′ . We construct Vs′ and Vt′ so that they each have exactly one exposed vertex. Let vs ∈ Vs be an arbitrary vertex in Vs and vt ∈ Vt be an arbitrary vertex in Vt . Then, we let Vs′ = {li , ri | vi ∈ Vs } \ {rs } and Vt′ = {li , ri | vi ∈ Vt } \ {rt }. Note that ls and lt are the unique exposed vertices in Vs′ and Vt′ , respectively. Since Vs and Vt form cliques in G, G′ [Vs′ ] and G′ [Vt′ ] have diameter at most two. We have thus constructed our corresponding instance in polynomial time. Let V ′ be any subset of V (G′ ), and let VL′ = V ′ ∩ L and VR′ = V ′ ∩ R. The key observation is that G[V ′ ] has diameter more than two if both VL′ and VR′ contain exposed vertices in V ′ . We below prove that (G, Vs , Vt ) of clique reconfiguration is a yes-instance if and only if the corresponding instance (G′ , Vs′ , Vt′ ) of the induced variant under the TS rule is a yes-instance. We first prove the only-if direction, supposing that there exists a reconfiguration sequence hVs = V0 , V1 , . . . , Vℓ = Vt i of cliques in G. Then, we show that Vs′ is reconfigurable into Vt′ by induction on ℓ. If ℓ = 0 and hence Vs = Vt , then we can obtain Vt′ from Vs′ by exchanging rt in Vs′ with rs (or Vs′ = Vt′ already holds). We then consider the case where ℓ ≥ 1. Let {vi } = V0 \ V1 and {vj } = V1 \ V0 ; note that vi and vj are adjacent in G since we consider the TS rule. Then, we consider two vertex subsets V0′′ = (Vs′ ∪ {rs }) \ {ri } and V1′ = (V0′′ ∪ {lj }) \ {li }; note that rs and ri are adjacent in G′ since rs , ri ∈ V0 = Vs , and that li and lj are adjacent in G′ since vi and vj are adjacent in G. Notice that V ′′ and V1′ have distinct exposed vertices li and lj , respectively. By the construction of G′ , since V0 and V1 are cliques in G, both G′ [V0′′ ] and G′ [V1′ ] have diameter at most two. Then, the sequence hVs′ , V0′′ , V1′ i is a reconfiguration sequence between Vs′ and V1′ . Therefore, by applying the induction hypothesis to (G′ , V1′ , Vt′ ) and (G, V1 , Vt ), we obtain a reconfiguration sequence between V1′ and Vt′ . Thus, we can conclude that (G′ , Vs′ , Vt′ ) is a yes-instance. To prove the if direction, we now suppose that there exists a reconfiguration sequence V ′ = ′ hVs = V0′ , V1′ , . . . , Vℓ′ = Vt′ i. Consider the case where Vi′ ∈ V ′ has the exposed vertex in the side L, ′ say li ∈ Vi′ , and hence ri 6∈ Vi′ ; the other case is symmetric. Because G′ [Vi+1 ] must have diameter ′ at most two (and hence it must have only one exposed vertex), we know that Vi+1 is obtained from Vi′ by one of the following three moves (1) a token on a vertex rk ∈ Vi′ is moved to ri ; (2) the token on li is moved to its corresponding vertex ri ; or (3) the token on li is moved to a vertex lj ∈ L which is not in Vi′ and is adjacent to all vertices Vi′ ∩ L. Notice that the other moves increase the number of exposed vertices or make the resulting graph have diameter more than two. Then, each Vi′ ∈ V ′ induces a clique of size |Vs | = |Vt | in either L or R, and we can obtain a desired sequence of cliques between Vs and Vt . We note that the TS rule is critical in the reduction of Theorem 15. Under the TJ rule, there is no guarantee that we maintain a clique (and we cannot even guarantee that the resulting clique gets bigger). 14 5 Conclusions and future work The work in this paper initiates a systematic study of subgraph reconfiguration. Although we have identified graph structure properties which are harder for the induced variant than the spanning variant, it remains to be seen whether this pattern holds in general. For the general case, questions of the roles of diameter and the number of subgraphs satisfying the property are worthy of further investigation. Another obvious direction for further research is an investigation into the fixed-parameter complexity of subgraph reconfiguration. A natural extension of subgraph reconfiguration is the extension from isomorphism of graph structure properties to other mappings, such as topological minors. References [1] P. Bonsma. The complexity of rerouting shortest paths. Theoretical Computer Science, 510:1– 12, 2013. [2] P. Bonsma and L. Cereceda. Finding paths between graph colourings: PSPACE-completeness and superpolynomial distances. Theoretical Computer Science, 410(50):5215–5226, 2009. [3] M. R. Garey and D. S. Johnson. Computers and Intractability: A Guide to the Theory of NP-Completeness. Freeman, 1979. [4] R. A. Hearn and E. D. Demaine. PSPACE-completeness of sliding-block puzzles and other problems through the nondeterministic constraint logic model of computation. Theoretical Computer Science, 343(1–2):72–96, 2005. [5] J. van den Heuvel. The complexity of change. Surveys in Combinatorics 2013, 409:127–160, 2013. [6] T. Ito, E. D. Demaine, N. J. A. Harvey, C. H. Papadimitriou, M. Sideri, R. Uehara, and Y. Uno. On the complexity of reconfiguration problems. Theoretical Computer Science, 412(12–14):1054–1065, 2011. [7] T. Ito, H. Ono, and Y. Otachi. Reconfiguration of cliques in a graph. In Proceedings of the 12th Annual Conference on Theory and Applications of Models of Computation, pages 212–223, 2015. [8] M. Kamiński, P. Medvedev, and M. Milanič. Complexity of independent set reconfigurability problems. Theoretical Computer Science, 439:9–15, 2012. [9] A. E. Mouawad, N. Nishimura, V. Raman, N. Simjour, and A. Suzuki. On the parameterized complexity of reconfiguration problems. Algorithmica, 78(1):274–297, 2017. [10] M. Mühlenthaler. Degree-constrained subgraph reconfiguration is in P. In Proceedings of the 40th International Symposium on Mathematical Foundations of Computer Science, pages 505–516, 2015. [11] N. Nishimura. Introduction to reconfiguration. Preprints 2017090055, 2017. [12] K. Wasa, K. Yamanaka, and H. Arimura. The complexity of induced tree reconfiguration problems. In Proceedings of the 10th International Conference of Language and Automata Theory and Applications, pages 330–342, 2016. [13] M. Wrochna. Reconfiguration in bounded bandwidth and tree-depth. Journal of Computer and System Sciences, 93:1–10, 2018. 15 

8cs.DS

arXiv:1504.06729v4 [cs.DS] 12 Jul 2016 Optimal Principal Component Analysis in Distributed and Streaming Models Christos Boutsidis New York, New York christos.boutsidis@gmail.com David P. Woodruff IBM Research Almaden, California dpwoodru@us.ibm.com Peilin Zhong Institute for Interdisciplinary Information Sciences Tsinghua University, Beijing, China zpl12@mails.tsinghua.edu.cn Abstract We study the Principal Component Analysis (PCA) problem in the distributed and streaming models of computation. Given a matrix A ∈ Rm×n , a rank parameter k < rank(A), and an accuracy parameter 0 < ε < 1, we want to output an m × k orthonormal matrix U for which kA − UUT Ak2F ≤ (1 + ε) · kA − Ak k2F , where Ak ∈ Rm×n is the best rank-k approximation to A. We show the following. 1. In the arbitrary partition Psmodel of Kannan et al. (COLT 2014), each of s machines holds a matrix Ai and A = i=1 Ai . Each machine should output U. Kannan et al. achieve O(skm/ε) + poly(sk/ε) words (of O(log(nm)) bits) communication. We obtain the improved bound of O(skm) + poly(sk/ε) words, and show an optimal Ω(skm) lower bound up to low order terms. This resolves an open question for high precision PCA. A poly(ε−1 ) dependence is known to be required, but we separate this dependence from m. 2. We bypass the above lower bound when A is φ-sparse in each column and each server receives a subset of columns. Here we obtain an O(skφ/ε) + poly(sk/ε) word protocol. Our communication is independent of the matrix dimensions, and achieves the guarantee that each server, in addition to outputting U, outputs a subset of O(k/ε) columns of A containing a U in its span (that is, we solve distributed column subset selection). We show a matching Ω(skφ/ε) lower bound for distributed column subset selection. Achieving our communication bound when A is sparse but not sparse in each column, is impossible. 3. In the streaming model in which the columns arrive one at a time, an algorithm of Liberty (KDD, 2013) with an improved analysis by Ghashami and Phillips (SODA, 2014) shows O(km/ε) “real numbers” of space is achievable in a single pass, which we first improve to an O(km/ε)+poly(k/ε) word space upper bound. This almost matches a known Ω(km/ε) bit lower bound of Woodruff (NIPS, 2014). We show with two passes one can achieve O(km) + poly(k/ε) words of space and (up to the poly(k/ε) term and the distinction between words versus bits) this is optimal for any constant number of passes. 4. In turnstile streams, in which we receive entries of A one at a time in an arbitrary order, we show how to obtain a factorization of a (1 + ε)-approximate rank-k matrix using O((m+n)kε−1 ) words of space. This improves the O((m+nε−2 )kε−2 ) bound of Clarkson and Woodruff (STOC 2009), and matches their Ω((m + n)kε−1 ) word lower bound. Notably, our results do not depend on the condition number of A. 1 Introduction In distributed-memory computing systems, such as, Hadoop [1] or Spark [3], Principal Component Analysis (PCA) and the related Singular Value Decomposition (SVD) of large matrices is becoming very challenging. Machine learning libraries implemented on top of such systems, for example mahout [2] or mllib [4], provide distributed PCA implementations since PCA is often used as a building block for a learning algorithm. PCA is useful for dimension reduction, noise removal, visualization, etc. In all of these implementations, the bottleneck is the communication; hence, the focus has been on minimizing the communication cost of the related algorithms, and not the computational cost, which is the bottleneck in more traditional batch systems. The data matrix corresponding to the dataset in hand, e.g., a term-document matrix representing a text collection, or the Netflix matrix representing user’s ratings for different movies, could be distributed in many different ways [46]. In this paper, we focus on the following so-called arbitrary partition and column partition models. In the arbitrary partition model, a matrix A ∈ Rm×n Ps is arbitrarily distributed among s machines. Specifically, A = i=1 Ai where Ai ∈ Rm×n is held by machine i. Unless otherwise stated, we always assume m ≤ n. In the column partition model,
A A . . . A a matrix A ∈ Rm×n is distributed column-wise among s < n machines: A = ; 1 2 s P here, for i = 1 : s, Ai is an m × wi column submatrix of A with i wi = n. Note that the column partition model is a special case of the arbitrary partition model. Thus, it is desirable to prove upper bounds in the arbitrary partition model, and lower bounds in the column partition model. Both models have been adapted by traditional numerical linear algebra software libraries for distributed memory matrix computations [12]. A recent body of work [29, 10, 41, 31, 37, 13] (see Section 3 for a comparison) has focused on designing algorithms which minimize the communication needed for each machine to output an m × k matrix U for which an m × k orthonormal matrix U for which kA − UUT Ak2F ≤ (1 + ε) · kA − Ak k2F , where Ak ∈ Rm×n is the best rank-k approximation to A. Each machine should output the same matrix U which can then be used for downstream applications, such as clustering, where one first projects the data to a lower dimensional subspace (see, e.g., [10]). The protocol should succeed with large constant probability. The model is such that each of the s machines communicates with one machine, called the “coordinator” (which we sometimes refer to as Server), but does not communicate with other machines. This is known as the coordinator model and one can simulate arbitrary point-to-point communication with a constant factor overhead in communication together with an additive O(log(mn)) bit overhead per message (see, e.g., [45]). 1.1 Our Results The best upper bound in the arbitrary partition model is O((skmε−1 ) + s · poly(kε−1 )) words (of O(log(mn)) bits) communication [37], while the only known lower bound for these problems is in the arbitrary partition model and is Ω(skm) [37]. In high precision applications one may want to set ε to be as small as possible and the leading order term of O(skmε−1 ) is undesirable. A natural question is whether there is an O((skm) + s · poly(kε−1 )) word protocol. We note that there is a fairly simple known Ω(ε−2 ) bit lower bound [54] (we present this in Lemma 75 for completeness), so although one cannot achieve an O(log(1/ε)) dependence in the communication as one would maybe expect given iterative algorithms for regression with running time O(log(1/ε)) (see section 2.6 of [55] for an overview), an O((skm) + s · poly(kε−1 )) would at least separate the dependence of ε and m and allowing for much smaller ε with the same amount of communication. There are many existing protocols using O(skmε−1 ) words of communication [29, 10, 41, 31, 37, 13], and the protocols work in very different ways: that of [29, 10] is based on coresets, while 1 that of [41, 31] is based on adapting a streaming algorithm to the communication setting, while that of [37] is based on sketching, and that of [13] is based on alternating minimization and is useful only under some assumptions on the condition number (see Section 3 for a more detailed discussion of these protocols). It was thus natural to assume there should be a lower bound of Ω(skmε−1 ). Instead, we obtain a new upper bound in the arbitrary partition model and a matching lower bound (up to lower order terms) in the column partition model. Theorem 1 (Restatement of Theorem 43 and Theorem 82). Suppose matrix A ∈ Rm×n is partitioned in the arbitrary-partition model (See Definition 14). For any 1 ≥ ε > 0, there is an algorithm which on termination leaves the same orthonormal matrix U ∈ Rm×k on each machine such that kA − UUT Ak2F ≤ (1 + ε) · kA − Ak k2F holds with arbitrarily large constant probability. Further, the algorithm runs in polynomial time with total communication complexity O(skm + s · poly(kε−1 )) words each containing O(log(smnε−1 )) bits. If matrix A ∈ Rm×n is distributed in the column-partition model (See Definition 15), then for any positive constant C ≤ O(poly(skm)), the algorithm which can leave a orthonormal matrix U ∈ Rm×k on each machine such that kA − UUT Ak2F ≤ C · kA − Ak k2F holds with constant probability has at least Ω(skm log(skm)) bits of communication. In some applications even an O(skm) word protocol may be too costly, as m could be very large. One could hope to do for communication what recent work on input sparsity algorithms [22, 43, 44, 16, 25] has done for computation, i.e., obtain protocols sensitive to the number of non-zero entries of A. Many matrices are sparse, e.g., Netflix provides a training data set of 100, 480, 507 ratings that 480, 189 users give to 17, 770 movies. If users correspond to columns in the matrix, then the average column sparsity is ≈ 200 non-zero elements (out of the 17, 770 possible coordinates). Our second contribution is the first protocol depending on the number of non-zero entries of A in the column partition model. Our communication is independent of the matrix dimensions. Denote by φ the maximum number of non-zero elements of a column in A. When we say that A is φ-sparse, we mean that every column of A has at most φ non-zero elements and nnz(A) ≤ φ · n. Our protocol has the additional feature of leaving on each machine the same subset C of O(k/ε) columns of A for which there exists an m × k orthonormal matrix U in the column span of C for which kA − CC† Ak2F ≤ kA − UUT Ak2F ≤ (1 + ε) · kA − Ak k2F . This is known as the column subset selection problem, which is useful since C may be sparse if A is, and also it can lead to better data intepretability. To partially complement our protocol, we show our protocol is optimal for any protocol solving the column subset selection problem. We summarize these results as follows. Theorem 2 (Restatement of Theorem 63). Suppose a φ-sparse matrix A ∈ Rm×n is partitioned in the column-partition model (See Definition 15). For any 1 ≥ ε > 0, there is an algorithm which on termination leaves C ∈ Rm×c with c = O(k/ε) columns of A and an orthonormal matrix U on each machine such that kA − CC† Ak2F ≤ kA − UUT Ak2F ≤ (1 + ε) · kA − Ak k2F holds with arbitrarily large constant probability. Further, the algorithm runs in polynomial time with total communication complexity
O skφε−1 + sk2 ε−4 words each containing of O(log(smnε−1 )) bits. If a φ-sparse matrix A ∈ Rm×n is distributed in the column-partition model (See Definition 15), then for any positive constant C ≤ O(poly(skφ)), the algorithm which can leave a orthonormal matrix U ∈ Rm×k on each machine such that kA − UUT Ak2F ≤ C · kA − Ak k2F holds with constant probability has at least Ω(skφ log(skφ)) bits of communication. Remark 3. Our protocol has no dependence on condition number and works for arbitrary matrices. We 2 Upper bounds Lower bounds Definition 16 (arbitrary partition model) O(skm + s · poly(k/ε)) (Theorem 43) Ω(skm) (Theorem 1.2 in [37]) Definition 17 (column partition model) O(skm + s · poly(k/ε)) (Theorem 43) Ω(skm) (Theorem 82) Definition 17 with sparsity φ = o(ε · m) O(skφε−1 + s · poly(k/ε)) (Theorem 63) Ω(skφ) (Corollary 83 ) Table 1: Communication upper/lower bounds for the Distributed PCA problems. make a serious effort as shown in Section 6 to achieve this. Note that even if the entries of A are specified using O(log(mn)) bits, the condition number can be 2−Ω(mn log(mn)) [7]. Remark 4. We did not discuss the running time of the algorithms, but we are interested in the fastest possible algorithms with the minimal communication cost, and provide fast algorithms as well. Remark 5. The hard instance in Section 10.2.2 implies an Ω(skm) lower bound even if the input matrix is sparse overall but has skm non-zero positions in arbitrarily locations. Therefore, to obtain our bounds we discuss the sparsity on each column instead of the overall sparsity of the matrix. A model closely related to the distributed model of computation is the streaming model of computation. The model we focused on is the so-called turnstile streaming model. In this model, there is a stream of update operations and each operation indicates that the corresponding entry of A should be incremented by a specific number. We present novel PCA algorithms in this model. Our first one-pass algorithm improves upon the best existing streaming PCA algorithm [41, 31] in two respects. First, the space of [41, 31] is described in “real numbers” while our space bound (O(mk/ε + poly(k/ε)) - see Theorem 52) is in terms of words (we also bound the word size). This matches an Ω(km/ε) bit lower bound for one-pass algorithms in [53], up to the distinction between words versus bits and a low order term poly(k/ε). Second, our algorithm can be applied in the turnstile streaming model which is stronger than column update streaming model in [41, 31]. Theorem 6 (Restatement of Theorem 52). Suppose A ∈ Rm×n is given by a stream of update operations in the turnstile streaming model (See Definition 20). For any 1 ≥ ε > 0, there is an algorithm which uses a single pass over the stream and on termination outputs an orthonormal matrix U ∈ Rm×k such that kA − UUT Ak2F ≤ (1 + ε) · kA − Ak k2F holds with arbitrarily large constant
probability. Further, the algorithm runs in polynomial time with space of total O mk/ε + poly(kε−1 ) words each containing O(log(smnε−1 )) bits. A slight modification of the previous algorithm leads to a one-pass algorithm which can compute a factorization of a (1 + ε)-approximate rank-k matrix. The modified algorithm only needs O((n + m)k/ε + poly(k/ε)) words of space which improves the O((m + nε−2 )kε−2 ) upper bound in [21] and matches the Ω((n + m)k/ε) bit lower bound given by [21], up to the low order term poly(k/ε). Theorem 7 (Restatement of Theorem 53). Suppose A ∈ Rm×n is given by a stream of update operations in the turnstile streaming model (See Definition 20). For any 1 ≥ ε > 0, there is an algorithm which uses a single pass over the stream and on termination outputs a matrix A∗k ∈ Rm×n with rank(A∗k ) ≤ k such that kA − A∗k k2F ≤ (1 + ε) · kA − Ak k2F holds with arbitrarily large constant probability. Further, the algorithm runs in polynomial time with space of total O((m + n)k/ε + poly(kε−1 )) words each containing O(log(smnε−1 )) bits. 3 Upper bounds One-pass turnstile, Def. 21 One-pass turnstile, factorization, Def. 22 Two-pass turnstile, Def. 21 O(mkε−1 + poly(k, ε−1 )) Lower bounds (Theorem 52) O((n + m)kε−1 + poly(k, ε−1 )) (Theorem 53) O(mk + poly(k, ε−1 )) (Theorem 55) Ω(mkε−1 ) bits [53] Ω((n + m)kε−1 ) words [21] Ω(mk) bits [53] Table 2: Space upper/lower bounds for the Streaming PCA problem. We also show a two-pass algorithm which is an implementation of our distributed PCA protocol. It uses O((km)+poly(k/ε)) words of space, which up to the poly(k/ε) term and the distinction between words versus bits, is optimal for any constant number of passes. A “next natural goal” in [53] was to improve the lower bound of Ω(km) to a bound closer to the 1-pass Ω(km/ε) lower bound established in that paper; our upper bound shows this is not possible. Theorem 8 (Restatement of Theorem 55). Suppose A ∈ Rm×n is given by a stream of update operations in the turnstile streaming model (See Definition 20). For any 1 ≥ ε > 0, there is an algorithm which uses two passes over the stream and on termination outputs an orthonormal matrix U ∈ Rm×k such that kA − UUT Ak2F ≤ (1 + ε) · kA − Ak k2F holds with arbitrarily large constant probability. Further, the algorithm
runs in polynomial time with total space O mk + poly(kε−1 ) words each containing O(log(smnε−1 )) bits. 1.2 Technical Overview 1.2.1 Upper Bounds The Arbitrary Partition Model. Recall that the inputP matrix A ∈ Rm×n is arbitrarily partitioned m×n into s matrices Ai ∈ R , i.e., for i = 1, 2, . . . , s: A = si=1 Ai . By sketching on the left and right 2 2 by two small random sign matrices S ∈ RO(k/ε )×m and T ∈ Rn×O(k/ε ) , with O(sk 2 /ε4 ) words of communication, the server can learn à = SAT which is a “sketch” of A. It suffices to compute T the best rank-k approximation Ãk to Ã. Suppose the SVD of Ãk = UÃk ΣÃk Và . Then the server k can learn ATVÃk by the second round of communication which needs O(skm) words. We then prove (see Lemma 28): kA − UUT Ak2F ≤ (1 + ε)kA − Ak k2F where U is an orthonormal basis of the column space of ATVÃk . Notice that rank(U) ≤ k. Thus U is exactly what we want. Section 5.3 shows the details. A technical obstacle in the analysis above is that it may require a large number of machine words to specify the entries of VÃk , even if each entry in A is only a single word. In fact, one can show (we omit the details) that the singular values of Ãk can be exponentially large in k/ε, which means one would need to round the entries of VÃk to an additive exponentially small precision, which would translate to an undesirable sm · poly(k/ε) bits of communication. To counter this, we use the smoothed analysis of Tao and Vu [52] to argue that if we add small random Bernoulli noise to A, then its minimum singular value becomes inverse polynomial in n. Moreover, if the rank of A is at least 2k and under the assumption that the entries of A are representable with a single machine word (so we can identify them with integers in magnitude at most poly(nms/ε) by scaling), then we can show the additional noise preserves relative error approximation. To our knowledge, this is the first application of this smoothing technique to distributed linear algebra algorithms. Section 6 discusses the details of this approach. 4 On the other hand, if the rank of A is smaller than 2k, the smoothing technique need not preserve relative error. In this case though, we can learn a basis C ∈ Rm×O(k) for the column span of A by multiplying by a pseudorandom matrix based on Vandermonde matrices. Since C has at most 2k columns, we can efficiently communicate it to all servers using O(skm) communication. At this point we set up the optimization problem minrank(X)≤k kCXCT A−AkF . The server cannot learn CT A and A directly, as this would be Ω(nm) words, but we can choose additional “sketching matrices” Tlef t and Tright to instead solve minrank(X)≤k kTlef t (CXCT A − A)Tright kF , which is a generalization of the subspace embedding technique to “affine embeddings” [53]. the server learns Tlef t CXCT ATright and Tlef t ATright , which are small matrices, and can be sent to each machine. Each machine locally solves for the rank-k solution X∗ minimizing kTlef t CXCT ATright − Tlef t ATright kF . Now each machine can find the same m × k orthonormal basis and output it without further communication. Sparse Matrices in the Column Partition Model. Our algorithm for sparse matrices is technically simpler than that in the arbitrary partition model. Section 8.2 presents a distributed PCA algorithm that has O((φksε−1 ) + poly(sk/ε)) words of communication cost. Our idea in order to take advantage of the sparsity in the input matrix A is to select and transfer to the coordinator a small set of “good” columns from each sub-matrix Ai . Specifically, we design an algorithm that first computes, in a distributed way, a matrix C̃ ∈ Rm×c with c = O(kε−1 ) columns of A, and then finds U ∈ span(C̃) using a distributed, communication-efficient algorithm developed in [37]. The matrix C̃ is constructed using optimal algorithms for column sampling [17, 26, 24], extended properly to the distributed case. As mentioned, our algorithm actually solves the distributed column subset selection problem. It builds on results for column subset selection in the batch setting [17, 26, 24], but gives a novel analysis showing that in the distributed setting one can find columns as good as in the batch setting. To the best of our knowledge, only heuristics for distributed column subset selection were known [28]. We also optimize the time complexity of our algorithm, see Table 4. Turnstile streaming model. Our one-pass streaming PCA algorithm starts by generating two random sign matrices S ∈ RO(k/ε)×n and R ∈ Rm×O(k/ε) . As shown in Lemma 51, min kARXSA − Ak2F ≤ (1 + ε) · kA − Ak k2F rank(X)≤k T The intuition of the above is as the following. Suppose the SVD of Ak is UAk ΣAk VA where k Ak is the best rank-k approximation matrix to A. Since X can be chosen as ΣAk VAk , we have minrank(X)≤k kUAk X − Ak2F = kAk − Ak2F . As shown in [21], S can provide a sketch for the regression problem. Thus, kUAk X̃ − Ak2F ≤ (1 + ε) · minrank(X)≤k kUAk X − Ak2F where X̃ = arg min rank(X)≤k kSUAk X − SAk2F Notice that X̃ is in the row space of span(SA). Then we focus on the regression problem: min kXSA − Ak2F . rank(X)≤k Similarly, we can use R to sketch the problem. Thus finally we only need to optimize min kARXSA − Ak2F . rank(X)≤k 5 Upper bounds Definition 18 Definition 19 O(skφε−1 ) Lower bounds (Theorem 74) Ω(skφε−1 ) (Theorem 90)   O skφε−1 + s · poly(k/ε) (Theorem 74) Ω(skφε−1 ) (Corollary 91) Table 3: Communication upper/lower bounds for the CSSP problems of Definitions 18 and 19. However, it is too expensive to store the entire matrix A. Thus, we sketch on the left and right by Tlef t and Tright . Since all of Tlef t AR, SATright and Tlef t ATright can be maintained in small space, X∗ = arg minrank(X)≤k kTlef t (ARXSA−A)Tright k2F can be constructed after one pass over the stream. Additionally, if AR is also maintained in the algorithm, we can get U ∈ Rm×k of which columns are an orthonormal basis of span(ARX∗ ) satisfying kA − UUT Ak2F ≤ (1+ ε)·kA − Ak k2F . T , Furthermore, if SA is also maintained, since it suffices to compute the SVD of X∗ = UX∗ ΣX∗ VX ∗ T T = Tlef t ARUX∗ and K = VX SAT can be constructed in O((n + m)k) words of space. right ∗ Therefore, the algorithm can compute A∗k = TΣX∗ K = ARX∗ SA in O((n + m)k) words of space. The algorithm then can output A∗k with total space O((n + m)k/ε + poly(k/ε)) words (see Theorem 53). Our two-pass streaming PCA algorithm is just an implemententation of our distributed PCA algorithm in the streaming model. 1.2.2 Lower Bounds Table 3 summarizes the matching communication lower bounds that we obtain for distributed column-based matrix reconstruction. Theorem 90 proves a lower bound for the problem in Definition 18; then, a lower bound for the problem of Definition 19 follows immediately since this is a harder problem (see Corollalry 91). Distributed Column Subset Selection. Our most involved lower bound argument is in showing the tightness for the distributed column subset selection problem, and is given in Theorem 90. To illustrate the ideas, suppose k = 1. We start with the canonical hard matrix for column subset selection, namely, an m × m matrix A whose first row is all ones, and remaining rows are a subset of the identity matrix. Intuitively the best rank-1 approximation is very aligned with the first row of A. However, given only o(1/ε) columns of the matrix, there is not a vector in the span which puts a (1 − ε)-fraction of its mass on the first coordinate, so Ω(1/ε) columns are needed. Such a matrix has φ = 2. Obtaining a lower bound for general φ involves creating a large net of such instances. Suppose we set m = φ and consider L̃A, where L̃ is formed by choosing a random φ × φ orthonormal matrix, and then rounding each entry to the nearest integer multiple of 1/poly(n). We can show that o(1/ε) columns of L̃A do not span a rank-1 approximation to L̃A. We also need a lemma which shows that if we take two independently random orthonormal matrices, and round their entries to integer multiples of 1/poly(n), then these matrices are extremely unlikely to share a constant fraction of columns. The idea then is that if one server holds L̃A, and every other server holds the zero matrix, then every server needs to output the same subset of columns of L̃A. By construction of A, each column of L̃A is the sum of the first column of L̃ and an arbitrary column of L̃, and since we can assume all servers know the first column of L̃ (which involves a negligible 6 O(sφ) amount of communication), this implies that each server must learn Ω(1/ε) columns of L̃. The probability that random discretized orthonormal matrices share Ω(1/ε) columns is sufficiently small that we can choose a large enough net for these columns to identify L̃ in that net, requiring Ω(φ/ε) bits of communication per server, or Ω(sφ/ε) words in total. The argument for general k can be viewed as a “direct sum theorem”, giving Ω(skφ/ε) words of communication in total. Dense Matrices in the Column Partition Model. Our optimal lower bound for dense matrices is technically simpler. It gives an Ω(skm) word communication lower bound in column partition model for dense matrices. Theorem 82 argues that there exists an m × n matrix A such that for this A, any k ≤ 0.99m, and any error parameter C with 1 < C < poly(skm), if there exists a protocol to construct an m × k matrix U satisfying kA − UUT Ak2F ≤ C · kA − Ak k2F , with constant probability, then this protocol has communication cost at least Ω(skm) words. This lower bound is proven for a matrix A that has k fully dense columns. We should note that the lower bound holds even if each machine only outputs the projection of Ai onto U, i.e., UUT Ai , rather than U itself. Note that this problem is potentially easier, since given U, a machine could find UUT Ai . The intuition of the proof is that machine 1 has a random matrix A1 ∈ Rm×k with orthonormal columns (rounded to integer multiples of 1/poly(n)), while all other machines have a very small multiple of the identity matrix. When concatenating the columns of these matrices, the best kdimensional subspace to project the columns onto must be very close to A1 in spectral norm. Since machine i, i > 1, has tiny identity matrices, after projection it obtains a projection matrix very close to the projection onto the column span of A1 . By choosing a net of m × k orthonormal matrices, all of which pairwise have high distance in spectral norm, each machine can reconstruct a random element in this net, which requires lots of information, and hence communication. Our lower bound of Theorem 82 implies a lower bound of Ω(skφ) words for matrices in which the column sparsity is φ (see Corollary 83) as a simple corollary. 1.3 Road Map In the first ten pages, we have one additional section, Section 2, which gives an overview of our algorithm in the arbitrary partition model. In the Appendix, we then present more detail on prior results on distributed PCA algorithms in Section 3. Section 4 introduces the notation and basic results from linear algebra that we use in our algorithms. Section 5 presents our distributed PCA algorithm for arbitrary matrices for Definition 16. The communication of the algorithm in this section can be stated only in terms of “real numbers”. We resolve this issue in Section 6 where a modified algorithm has communication cost bounded in terms of machine words. Section 7 discusses space-optimal PCA methods in the streaming model of computation. Section 8 presents a distributed PCA algorithm for sparse matrices in column partition model, while Section 9 extends this algorithm to a faster distributed PCA algorithm for sparse matrices. Section 10 presents communication lower bounds. 2 Outline of our result in the arbitrary partition model Here we give a brief outline of our result in the arbitrary patition model, deferring the full details to the appendix. We describe a fast distributed PCA algorithm with total communication O(msk) words plus low order terms, which is optimal in the arbitrary partition model in the sense that an Ω̃(msk) bit 7 lower bound was given by [37]. The algorithm employs, in a novel way, the notion of projectioncost preserving sketches from [24]. In particular, whereas all previous [48, 23] dimension-reductionbased SVD methods reduce one dimension of the input matrix to compute some approximation to the SVD, our method reduces both dimensions and computes an approximation to the SVD from a small almost square matrix (we note that other work, such as work on estimating eigenvalues [8], uses sketches in both dimensions, but it is not clear how to compute singular vectors using such sketches). Unlike [37] which reduces only one dimension in the first communication round, we do the reduction on both dimensions in the same round. We first present the batch version of the algorithm which offers a new low-rank matrix approximation technique; a specific implementation of this algorithm offers a communication-optimal distributed PCA algorithm (we also discuss in Section 7 a variant of this algorithm that offers a two-pass space-optimal PCA method in the turnstile streaming model). Before presenting all these new algorithms in detail, we present the relevant results from the previous literature that we employ in the analysis. 2.1 Projection-cost preserving sketching matrices In this section, we recap a notion of sketching matrices which we call “projection-cost preserving sketching matrices”. A sketching matrix from this family is a linear matrix transformation and it has the property that for all projections it preserves, up to some error, the difference between the matrix in hand and its projection in Frobenius norm. Definition 9 (Projection-cost preserving sketching matrices). We say that W ∈ Rn×ξ is an (ε, k)projection-cost preserving sketching matrix of A ∈ Rm×n , if for all rank-k orthogonal projection matrices P ∈ Rm×m , it satisfies (1 − ε)kA − PAk2F ≤ kAW − PAWk2F + c ≤ (1 + ε)kA − PAk2F where c is a non-negative constant which only depends on A and W. We also call AW an (ε, k)-projectioncost preserving sketch of A. Due to the following lemma, we know that a good rank-k approximation projection matrix of (ε, k)-projection-cost preserving sketch AW also provides a good rank-k approximation to A. Lemma 10 (PCA via Projection-Cost Preserving Sketches - Lemma 3 in [24]). Suppose W ∈ Rn×ξ is ∗ an (ε, k)-projection-cost preserving sketching matrix of A ∈ Rm×n . Let P̂ = arg minrank(P)≤k kAW − ∗ PAWk2F . For all P̂, ε′ satisfying rank(P̂) ≤ k, ε′ ≥ 0, if kAW− P̂AWk2F ≤ (1+ε′ )kAW− P̂ AWk2F , kA − P̂Ak2F ≤ 1+ε · (1 + ε′ )kA − Ak k2F 1−ε [24] also provides several ways to construct projection-cost preserving sketching matrices. Because we mainly consider the communication, we just choose one which can reduce the dimension as much as possible. Furthermore, it is also an oblivious projection-cost preserving sketching matrix. Lemma 11 (Dense Johnson-Lindenstrauss matrix - part of Theorem 12 in [24]). √ For ε < √ 1, suppose n×ξ each entry of W ∈ R is chosen O(log(k))-wise independently and uniformly in {1/ ξ, −1/ ξ} where ξ = O(kε−2 ) [21]. For any A ∈ Rm×n , with probability at least 0.99, W is an (ε, k)-projection-cost preserving sketching matrix of A. 8 2.2 A batch algorithm for the fast low rank approximation of matrices In this section, we describe a new method for quickly computing a low-rank approximation to a given matrix. This method does not offer any specific advantages over previous such techniques [48, 23, 18]; however, this new algorithm can be implemented efficiently in the distributed setting (see Section 5.3) and in the streaming model of computation (see Section 7); in fact we are able to obtain communication-optimal and space-optimal results, respectively. For completeness as well as ease of presentation, we first present and analyze the simple batch version of the algorithm. The algorithm uses the dense Johnson-Lindenstrauss matrix of Lemma 27 in order to reduce both dimensions of A, before computing some sort of SVD to a poly(k/ε) × poly(k/ε) matrix (see Step 2 in the algorithm below). Consider the usual inputs: a matrix A ∈ Rm×n , a rank parameter k < rank(A), and an accuracy parameter 0 < ε < 1. The algorithm below returns an orthonormal matrix U ∈ Rm×k such that kA − UUT Ak2F ≤ (1 + ε) · kA − Ak k2F . Algorithm 1. Construct two dense Johnson-Lindenstrauss matrices S ∈ Rξ1 ×m , T ∈ Rn×ξ2 with ξ1 = O(kε−2 ), ξ2 = O(kε−2 ) (see Lemma 27). 2. Construct à = SAT. T 3. Compute the SVD of Ãk = UÃk ΣÃk Và (UÃk ∈ Rξ1 ×k , ΣÃk ∈ Rk×k , VÃk ∈ Rξ2 ×k ). k 4. Construct X = ATVÃk 5. Compute an orthonormal basis U ∈ Rm×k for span(X) (notice that rank(X) ≤ k). Theorem 29 later in this section analyzes the approximation error and the running time of the previous algorithm. First, we prove the accuracy of the algorithm. Lemma 12. The matrix U ∈ Rm×k with k orthonormal columns satisfies with probability at least 0.98: kA − UUT Ak2F ≤ (1 + ε) · kA − Ak k2F . Proof. T T T kà − Ãk k2F = kà − ÃVÃk Và k2F = kSAT − SATVÃk Và k2F = kTT AT ST − VÃk Và TT AT ST k2F k k k T . The second equality is by the The first equality follows by the SVD of Ãk = UÃk ΣÃk Và k construction of à = SAT. The third equality is due to ∀M, kMk2F = kM T k2F . Due to Lemma 27, with probability at least 0.99, ST is an (ε, k)-projection-cost preserving sketch matrix of TT AT . According to Lemma 26, T T TT AT k2F ≤ k2F = kTT AT − VÃk Và kAT − ATVÃk Và k k 1+ε · kAT − (AT)k k2F 1−ε Observe that T T kUUT AT − ATk2F = kXX† AT − ATk2F ≤ kXVà − ATk2F = kATVÃk Và − ATk2F k which is at most 1+ε 1−ε · kAT − (AT)k k2F . 9 k (1) The first equality uses the fact that U is an orthonormal basis of span(X). The first inequality is followed by ∀X, M, N, kXX† M − Mk2F ≤ kXN − Mk2F . The second equality uses the construction that X = ATVÃk . The second inequality follows by Eqn (2). Due to Lemma 27, with probability at least 0.99, T is an (ε, k)-projection-cost preserving sketch matrix of A. Due to Lemma 26, kUUT A − Ak2F ≤ (1 + ε)2 · kA − Ak k2F (1 − ε)2 Due to union bound, the probability that ST is an (ε, k)-projection-cost preserving sketch matrix of TT AT and T is an (ε, k)-projection-cost preserving sketch matrix of A is at least 0.98. Note that (1+ε)2 is 1 + O(ε) when ε is small enough, so we can adjust ε here by a constant factor to show the (1−ε)2 statement. Next, we present the main theorem. Theorem 13. The matrix U ∈ Rm×k with k orthonormal columns satisfies with probability at least 0.98: kA − UUT Ak2F ≤ (1 + ε) · kA − Ak k2F . The running time of the algorithm is
O nmkε−2 + mk2 ε−4 + poly(kε−1 ) . Proof. The correctness is shown by Lemma 28. The running time is analyzed in Section 5. 2.3 The distributed PCA algorithm Ps Recall that the input matrix A ∈ Rm×n is partitioned arbitrarily as: A = i Ai for i = 1 : s, Ai ∈ Rm×n . The idea in the algorithm below is to implement the algorithm in Section 5.2 in the distributed setting. Input: 1. A ∈ Rm×n arbitrarily partitioned A = 2. rank parameter k < rank(A) Ps i Ai for i = 1 : s, Ai ∈ Rm×n . 3. accuracy parameter ε > 0 Algorithm 1. Machines agree upon two dense Johnson-Lindenstrauss matrices S ∈ Rξ1 ×m , T ∈ Rn×ξ2 with ξ1 = O(kε−2 ), ξ2 = O(kε−2 ) (see Lemma 27). P 2. Each machine locally computes Ãi = SAi T and sends Ãi to the server. Server constructs à = i Ãi . T 3. Server computes the SVD of Ãk = UÃk ΣÃk Và (UÃk ∈ Rξ1 ×k , ΣÃk ∈ Rk×k , VÃk ∈ Rξ2 ×k ). k 4. Server sends VÃk to all machines. 5. Each machine construct Xi = Ai TVÃk and sends Xi to the server. Server constructs X = 6. Server computes an orthonormal basis U ∈ Rm×k for span(X) (notice that rank(X) ≤ k). P i Xi . 7. Server sends U to each machine. Notice that in the first step, S and T can be described using a random seed that is O(log(k))wise independent due to Lemma 27. The major remaining challenge here is to obtain small bit complexity, since this protocol could have very large bit complexity. See Section 5 in the Appendix for the full details of our algorithm. 10 References [1] http://hadoop.apache.org/. [2] https://mahout.apache.org. [3] https://spark.apache.org/. [4] https://spark.apache.org/mllib/. [5] D. Achlioptas. Database-friendly random projections: Johnson-lindenstrauss with binary coins. J. Comput. Syst. Sci., 66(4):671–687, 2003. [6] N. Ailon and E. Liberty. Fast dimension reduction using rademacher series on dual bch codes. In ACM-SIAM Symposium on Discrete Algorithms (SODA), 2008. [7] N. Alon and V. H. Vu. Anti-hadamard matrices, coin weighing, threshold gates, and indecomposable hypergraphs. J. Comb. Theory, Ser. A, 79(1):133–160, 1997. [8] A. Andoni and H. L. Nguyen. Eigenvalues of a matrix in the streaming model. In Proceedings of the Twenty-Fourth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2013, New Orleans, Louisiana, USA, January 6-8, 2013, pages 1729–1737, 2013. [9] Z.-J. Bai, R. H. Chan, and F. T. Luk. Principal component analysis for distributed data sets with updating. In Advanced Parallel Processing Technologies, pages 471–483. Springer, 2005. [10] M.-F. Balcan, V. Kanchanapally, Y. Liang, and D. Woodruff. Improved distributed principal component analysis. arXiv preprint arXiv:1408.5823, to appear in NIPS, 2014. [11] J. Batson, D. Spielman, and N. Srivastava. Twice-ramanujan sparsifiers. In Proceedings of the 41st annual ACM symposium on Theory of computing, pages 255–262. ACM, 2009. [12] A. R. Benson, D. F. Gleich, and J. Demmel. Direct qr factorizations for tall-and-skinny matrices in mapreduce architectures. In Big Data, 2013 IEEE International Conference on, pages 264–272. IEEE, 2013. [13] S. Bhojanapalli, P. Jain, and S. Sanghavi. Tighter low-rank approximation via sampling the leveraged element. http://uts.cc.utexas.edu/˜bsrinadh/main.pdf, to appear in SODA, 2015. [14] L. S. Blackford, J. Choi, A. Cleary, E. D’Azevedo, J. Demmel, I. Dhillon, J. Dongarra, S. Hammarling, G. Henry, A. Petitet, et al. ScaLAPACK users’ guide, volume 4. siam, 1997. [15] Y.-A. L. Borgne, S. Raybaud, and G. Bontempi. Distributed principal component analysis for wireless sensor networks. Sensors, 2008. [16] J. Bourgain, S. Dirksen, and J. Nelson. Toward a unified theory of sparse dimensionality reduction in euclidean space. In Proceedings of the Forty-Seventh Annual ACM on Symposium on Theory of Computing, STOC 2015, Portland, OR, USA, June 14-17, 2015, pages 499–508, 2015. [17] C. Boutsidis, P. Drineas, and M. Magdon-Ismail. Near optimal column based matrix reconstruction. SIAM Journal on Computing (SICOMP), 2013. 11 [18] C. Boutsidis and A. Gittens. Improved matrix algorithms via the subsampled randomized hadamard transform. SIAM Journal on Matrix Analysis and Applications, 34(3):1301–1340, 2013. [19] C. Boutsidis and D. P. Woodruff. Optimal cur matrix decompositions. STOC, 2014. [20] A. Chakrabarti and O. Regev. An optimal lower bound on the communication complexity of gap-hamming-distance. SIAM Journal on Computing, 41(5):1299–1317, 2012. [21] K. Clarkson and D. Woodruff. Numerical linear algebra in the streaming model. In Proceedings of the 41st annual ACM symposium on Theory of computing (STOC), 2009. [22] K. L. Clarkson and D. P. Woodruff. Low rank approximation and regression in input sparsity time. In In STOC, 2013. [23] K. L. Clarkson and D. P. Woodruff. Low rank approximation and regression in input sparsity time. ArxiV report: http://arxiv.org/pdf/1207.6365v4.pdf, 2013. [24] M. Cohen, S. Elder, C. Musco, C. Musco, and M. Persu. Dimensionality reduction for k-means clustering and low rank approximation. arXiv preprint arXiv:1410.6801, 2014. [25] M. B. Cohen. Simpler and tighter analysis of sparse oblivious subspace embeddings. In SODA, 2016. [26] A. Deshpande, L. Rademacher, S. Vempala, and G. Wang. Matrix approximation and projective clustering via volume sampling. Theory of Computing, 2(12):225–247, 2006. [27] A. Deshpande and S. Vempala. Adaptive sampling and fast low-rank matrix approximation. In RANDOM - APPROX, 2006. [28] A. K. Farahat, A. Elgohary, A. Ghodsi, and M. S. Kamel. Distributed column subset selection on mapreduce. In Data Mining (ICDM), 2013 IEEE 13th International Conference on, pages 171– 180. IEEE, 2013. [29] D. Feldman, M. Schmidt, and C. Sohler. Turning big data into tiny data: Constant-size coresets for k-means, pca and projective clustering. In SODA, pages 1434–1453. SIAM, 2013. [30] S. Friedland and A. Torokhti. Generalized rank-constrained matrix approximations. SIAM Journal on Matrix Analysis and Applications, 29(2):656–659, 2007. [31] M. Ghashami and J. Phillips. Relative errors for deterministic low-rank matrix approximations. In SODA, 2013. [32] G. H. Golub and C. F. Van Loan. Matrix computations, volume 3. JHU Press, 2012. [33] M. Gu and S. C. Eisenstat. A divide-and-conquer algorithm for the bidiagonal svd. SIAM Journal on Matrix Analysis and Applications, 16(1):79–92, 1995. [34] R. Horn and C. Johnson. Matrix Analysis. Cambridge University Press, New York, 1985. [35] E. Jessup and D. Sorensen. A parallel algorithm for computing the singular value decomposition of a matrix. Siam Journal on Matrix Analysis and Applications, 15(2):530–548, 1994. [36] R. Kannan, S. Vempala, and D. P. Woodruff. Nimble algorithms for cloud computing. arXiv preprint arXiv:1304.3162, v3, 2013. 12 [37] R. Kannan, S. S. Vempala, and D. P. Woodruff. Principal component analysis and higher correlations for distributed data. In Proceedings of The 27th Conference on Learning Theory, pages 1040–1057, 2014. [38] M. Kapralov and K. Talwar. On differentially private low rank approximation. In Proceedings of the Twenty-Fourth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2013, New Orleans, Louisiana, USA, January 6-8, 2013, pages 1395–1414, 2013. [39] D. Kempe and F. McSherry. A decentralized algorithm for spectral analysis. In Proceedings of the thirty-sixth annual ACM symposium on Theory of computing, pages 561–568. ACM, 2004. [40] Y. Li, H. L. Nguyên, and D. P. Woodruff. On sketching matrix norms and the top singular vector. In Proceedings of the Twenty-Fifth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 1562–1581. SIAM, 2014. [41] E. Liberty. Simple and deterministic matrix sketching. In KDD, pages 581–588. ACM, 2013. [42] S. V. Macua, P. Belanovic, and S. Zazo. Consensus-based distributed principal component analysis in wireless sensor networks. In Signal Processing Advances in Wireless Communications (SPAWC), 2010 IEEE Eleventh International Workshop on, pages 1–5. IEEE, 2010. [43] X. Meng and M. W. Mahoney. Low-distortion subspace embeddings in input-sparsity time and applications to robust linear regression. In Proceedings of the forty-fifth annual ACM symposium on Theory of computing, pages 91–100. ACM, 2013. [44] J. Nelson and H. L. Nguyên. Osnap: Faster numerical linear algebra algorithms via sparser subspace embeddings. In Foundations of Computer Science (FOCS), 2013 IEEE 54th Annual Symposium on, pages 117–126. IEEE, 2013. [45] J. M. Phillips, E. Verbin, and Q. Zhang. Lower bounds for number-in-hand multiparty communication complexity, made easy. In Proceedings of the Twenty-Third Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2012, Kyoto, Japan, January 17-19, 2012, pages 486– 501, 2012. [46] J. Poulson, B. Marker, R. A. van de Geijn, J. R. Hammond, and N. A. Romero. Elemental: A new framework for distributed memory dense matrix computations. ACM Transactions on Mathematical Software (TOMS), 39(2):13, 2013. [47] Y. Qu, G. Ostrouchov, N. Samatova, and A. Geist. Principal component analysis for dimension reduction in massive distributed data sets. In Proceedings of IEEE International Conference on Data Mining (ICDM), 2002. [48] T. Sarlos. Improved approximation algorithms for large matrices via random projections. In IEEE Symposium on Foundations of Computer Science (FOCS), 2006. [49] K. C. Sou and A. Ranzer. On generalized matrix approximation problem in the spectral norm. Linear Algebra and its Applications, 436(7):2331–2341, 2012. [50] F. Tisseur and J. Dongarra. A parallel divide and conquer algorithm for the symmetric eigenvalue problem on distributed memory architectures. SIAM Journal on Scientific Computing, 20(6):2223–2236, 1999. 13 [51] M. D. Vose. A linear algorithm for generating random numbers with a given distribution. Software Engineering, IEEE Transactions on, 17(9):972–975, 1991. [52] V. H. Vu and T. Tao. The condition number of a randomly perturbed matrix. In Proceedings of the 39th Annual ACM Symposium on Theory of Computing, San Diego, California, USA, June 11-13, 2007, pages 248–255, 2007. [53] D. Woodruff. Low rank approximation lower bounds in row-update streams. In Advances in Neural Information Processing Systems, pages 1781–1789, 2014. [54] D. Woodruff and P. Zhong. Distributed low rank approximation of implicit functions of a matrix. In manuscript, 2015. [55] D. P. Woodruff. Sketching as a tool for numerical linear algebra. Foundations and Trends in Theoretical Computer Science, 10(1-2):1–157, 2014. 14 3 Related Work Distributed PCA (or distributed SVD) algorithms have been investigated for a long time. One line of work, developed primarily within the numerical linear algebra literature, studies such algorithms from the perspective of parallelizing existing standard SVD algorithms without sacrificing accuracy. This approach aims at high accuracy implementations with the least possible communication cost. The distributed models of computation go typically beyond the column-partition model and arbitrary-partition model that we study in this paper [46]. An extensive survey of this line of work is out of the scope of our paper; we refer the reader to [35, 50, 33] and references therein for more details as well as to popular software for distributed SVD such ScaLAPACK [14] and Elemental [46]. Another line of work for distributed PCA algorithms has emerged within the machine learning and datamining communities. Such algorithms have been motivated by the need to apply SVD or PCA to extremely large matrices encoding enormous amounts of data. The algorithms in this line of research are typically heuristic approaches that work well in practice but come with no rigorous theoretical analysis. We refer the reader to [47, 42, 9, 15] for more details. Finally, distributed PCA algorithms in the column-partition model have been recently studied within the theoretical computer science community. Perhaps the more intuitive algorithm for PCA in this model appeared in [29, 10]: first, a number of left singular vectors and singular values are computed in each machine; then, the server collects those singular vectors and concatenates them column-wise in a new matrix and then it computes the top k left singular vectors of this “aggregate” matrix. It is shown in [29, 10] that if the number of singular vectors and singular values in the first step is O(kε−1 ), then, the approximation error in Frobenius norm is at most (1 + ε) times the optimal Frobenius norm error; the communication cost is O(skmε−1 ) real numbers because each of the s machines sends O(kε−1 ) singular vectors of dimension m; unfortunately, it is unclear how one can obtain a communication cost in terms of words/bits. A different algorithm with the same communication cost (only in terms of real numbers since a word/bit communication bound remained unexplored) is implicit in [41, 31] (see Theorem 3.1 in [31] and the discussion in Section 2.2 in [41]). Kannan, Vempala and Woodruff proposed arbitrary-partition model in [37]. They developed a (1 + ε) Frobenius norm error algorithm with communication cost
−1 2 −4 O skmε + sk ε words. Bhojanapalli, Jain, and Sanghavi in [13] developed an algorithm that provides a bound with respect to the spectral norm. Their algorithm is based on sampling elements from the input matrix, but the communication cost is prohibitive if n is large. The cost also depends on the condition number of A. Moreover, to implement the algorithm, one needs to know kA − Ak kF . Finally, [39] discussed a distributed implementation of the “orthogonal iteration” to compute eigenvectors of graphs. The model they consider is different, but perhaps their protocol could be extended to our model. 4 Preliminaries Definitions. We formally define the problems and the models here. Definition 14 (Arbitrary-partition model). P An m×n matrix A is arbitrarily partitioned into s matrices Ai ∈ Rm×n , i.e., for i = 1, 2, . . . , s: A = si=1 Ai . There are s machines, and the i-th machine has Ai as input. There is also another machine, to which we refer to as the “server”, which acts as the central coordinator. The model only allows communication between the machines and the server. The communication cost of an algorithm in this model is defined as the total number of words transferred between the machines and the server, where we assume each word is O(log(nms/ε)) bits. 15 Reference kA − UUT Ak2 ≤ kA − UUT Ak2 F ≤ δ Communication cost Implicit in [29] - (1 + ε)kA − Ak k2 F 0 - Theorem 2 in [10] - (1 + ε)kA − Ak k2 F 0 - Theorem 6 in [10] - (1 + ε)kA − Ak k2 F >0 - Implicit in [41, 31] - (1 + ε)kA − Ak k2 F 0 O(mnkε−2 ) Thm 1.1 in [37]* - (1 + ε)kA − Ak k2 F O(1) Thm 5.1 in [13] kA − Ak k2 + Γ - >0 Remark p. 11 [13] kA − Ak k2 + Γ - >0 Theorem 43* - (1 + ε)kA − Ak k2 F O(1) -   O skmε−1 + sk2 ε−4   5 −2 O sm + nk ε ∆   O sm + nk3 ε−2 ∆ Theorem 63 - (1 + ε)kA − Ak k2 F O(1) O(skφε−1 + sk2 ε−4 ) Theorem 74 - (1 + ε)kA − Ak k2 F O(1) O(skφε−1 + sk3 ε−5 ) O (mns · poly(k, 1/ε))  O(skm + sk3 ε−5 ) Total  number of arithmetic operations  O mn min{m, n} + mskε−1 min{m, skε−1 }   O mn min{m, n} + mskε−1 min{m, skε−1 }        2 2 3 log sk log m O nnz(A) + s m 4k + k m 6 ε dε ε ε O(poly(m, n, k, s, ε−1 ))   −2 2 O nnz(A) + δ −1 nk5 ε−2 σ1 (A)σk (A)   −2 2 O nnz(A) + δ −1 nk5 ε−2 σ1 (A)σk (A)   O mnkε−2 + mk2 ε−4 + poly(kε−1 ) .  O nnz(A) · log2 ( ns ) + (m + n) · s · poly( k log( ns ) . δ ε δ Table 4: Distributed PCA Algorithms in the column-partition model with s machines (see Definition 15). ∗ indicates that the algorithm can also be applied in the arbitrary partition model (see Definition 14). A ∈ Rm×n has rank ρ, U ∈ Rm×k with k < ρ is orthonormal, 0 < ε < 1, and each column of A contains at most φ ≤ m  non-zero elements. δ is the failure probability. Finally, for −1 . notational convenience let Γ := εkA − Ak kF , ∆ := σ12 (A)σk−2 (A) log2 kAk2 kA − Ak k−1 F ε Definition 15 (Column-partition model). An m × n matrix A is partitioned arbitrarily column-wise
P m×w i into s blocks Ai ∈ R , i.e., for i = 1, 2, . . . , s: A = A1 A2 . . . As . Here, wi = n. There are s machines, and the i-th machine has Ai as input. There is also another machine, to which we refer to as the “server”, which acts as the central coordinator. The model only allows communication between the machines and the server. The communication cost of an algorithm in this model is defined as the total number of words transferred between the machines and the server, where we assume each word is O(log(nms/ε)) bits. Definition 16 (The Distributed Principal Component Analysis Problem in arbitrary partition Psmodel). m×n Give an m×n matrix A arbitrarily partitioned into s matrices Ai ∈ R (i : 1, 2, . . . , s): A = i=1 Ai , a rank parameter k < rank(A), and an accuracy parameter 0 < ε < 1, design an algorithm in the model of Definition 14 which, upon termination, leaves on each machine a matrix U ∈ Rm×k with k orthonormal columns such that kA − UUT Ak2F ≤ (1 + ε) · kA − Ak k2F , and the communication cost of the algorithm is as small as possible. Definition 17 (The Distributed Principal Component Analysis Problem in column partition model). Given an m × n matrix A
partitioned column-wise into s arbitrary blocks Ai ∈ Rm×wi (i : 1, 2, . . . , s): A = A1 A2 . . . As , a rank parameter k < rank(A), and an accuracy parameter 0 < ε < 1, design an algorithm in the model of Definition 15 which, upon termination, leaves on each machine a matrix U ∈ Rm×k with k orthonormal columns such that kA − UUT Ak2F ≤ (1 + ε) · kA − Ak k2F , and the communication cost of the algorithm is as small as possible. Definition 18 (The Distributed Column Subset Selection Problem). Given an m × n matrix A par
titioned column-wise into s arbitrary blocks Ai ∈ Rm×wi (i : 1, 2, . . . , s): A = A1 A2 . . . As , a rank parameter k < rank(A), and an accuracy parameter 0 < ε < 1, design an algorithm in the model of Definition 15 that, upon termination, leaves on each machine a matrix C ∈ Rm×c with c < n columns of A such that kA − CC† Ak2F ≤ (1 + ε) · kA − Ak k2F , and 1. The number of selected columns c is as small as possible. 2. The communication cost of the algorithm is as small as possible. Definition 19 (The Distributed Column Subset Selection Problem - rank k subspace version). Given an m × n matrix A
partitioned column-wise into s arbitrary blocks Ai ∈ Rm×wi (i : 1, 2, . . . , s): A = A1 A2 . . . As , a rank parameter k < rank(A), and an accuracy parameter 0 < ε < 1, design an algorithm in the model of Definition 15 that, upon termination, leaves on each machine a matrix C ∈ Rm×c with c < n columns of A and a matrix U ∈ Rm×k with k orthonormal columns with U ∈ span(C), such that kA − CC† Ak2F ≤ kA − UUT Ak2F ≤ (1 + ε) · kA − Ak k2F , and 16 1. The number of selected columns c is as small as possible. 2. The communication cost of the algorithm is as small as possible. Definition 20 (Streaming model for Principal Component Analysis). Let all the entries in A ∈ Rm×n initially be zeros. In the streaming model of computation, there is a stream of update operations that the q th operation has form (iq , jq , xq ) which indicates that Aiq ,jq should be incremented by xq where iq ∈ {1, ..., m}, jq ∈ {1, ..., n}, xq ∈ R. An algorithm is allowed a single pass over the stream. At the end of the stream the algorithm stores some information regarding A which we call a “sketch” of A. The space complexity of an algorithm in this model is defined as the total number of words required to describe the information the algorithm stores during the stream including the sketch. Each word is O(log(nms/ε)) bits. Definition 21 (The Streaming Principal Component Analysis Problem). Given an m × n matrix A, a rank parameter k < rank(A), and an accuracy parameter 0 < ε < 1, design an algorithm that, using as little space as possible, first finds a sketch of A in the streaming model (see Definition 20) and then, using only this sketch outputs U ∈ Rm×k with k orthonormal columns such that kA − UUT Ak2F ≤ (1 + ε) · kA − Ak k2F . Definition 22 (The Streaming Principal Component Analysis Problem (factorization)). Given an m × n matrix A, a rank parameter k < rank(A), and an accuracy parameter 0 < ε < 1, design an algorithm that, using as little space as possible, first finds a sketch of A in the streaming model (see Definition 20) and then, using only this sketch outputs A∗k ∈ Rm×n with rank(A∗k ) ≤ k such that kA−A∗k k2F ≤ (1 + ε) · kA − Ak k2F . Notation. A, B, . . . are matrices; a, b, . . . are column vectors. In is the n × n identity matrix; 0m×n is the m × n matrix of zeros; 1n is the n × 1 vector of ones; ei is the standard basis (whose dimensionality will be clear from the context): the ith element of ei is one and the rest are zeros. A(i) and A(j) denotes the ith column and jth row of A, respectively. Aij is the (i, j)th entry in A. Sampling Matrices. Let A = [A(1) , . . . , A(n) ] ∈ Rm×n and let C = [A(i1 ) , . . . , A(ic ) ] ∈ Rm×c consist of c < n columns of A. Note that we can write C = AΩ, where the sampling matrix is Ω = [ei1 , . . . , eic ] ∈ Rn×c (here ei are the standard basis vectors in Rn ). If D ∈ Rc×c is a diagonal matrix, then AΩD contains c columns of A rescaled with the corresponding elements in D. We abbreviate S := ΩD, hence the matrix S ∈ Rn×c “samples” and “rescales” c columns from A. P Matrix norms. We use the Frobenius and the spectral matrix-norms: kAk2F = i,j A2ij ; kAk2 = maxkxk2 =1 kAxk2 . kAkξ is used if a result holds for both norms ξ = 2 and ξ = F. The standard submultiplicativity property for matrix norms implies that for any A and B: kABkξ ≤ kAkξ ·kBkξ . The triangle inequality for matrix norms implies that kA + Bkξ ≤ kAkξ + kBkξ . A version of the triangle inequality for the norms squared is: kA + Bk2ξ ≤ 2 · kAk2ξ + 2 · kBk2ξ . A version of the matrix pythagorean theorem is: if AT B is the all-zeros matrix, then kA + Bk2F = kAk2F + kBk2F . If V has orthonormal columns, then kAV T kξ = kAkξ , for any A. If P is a symmetric projection matrix (i.e., P = PT and P2 = P) then, kPAkξ ≤ kAkξ , for any A. 17 Singular Value Decomposition (SVD) and Moore-Penrose Pseudo-inverse. Rm×n with rank(A) = ρ is   
Σk VkT 0 A = Uk Uρ−k , T 0 Σρ−k Vρ−k {z } | | {z } | {z } UA ∈Rm×ρ ΣA ∈Rρ×ρ The SVD of A ∈ T VA ∈Rρ×n with singular values σ1 ≥ . . . σk ≥ σk+1 ≥ . . . ≥ σρ > 0. Here, k < ρ. We will use σi (A) to denote the i-th singular value of A when the matrix is not clear from the context. The matrices Uk ∈ Rm×k and Uρ−k ∈ Rm×(ρ−k) contain the left singular vectors of A, and, similarly, the matrices Vk ∈ Rn×k and Vρ−k ∈ Rn×(ρ−k) contain the right singular vectors of A. It is well-known that Ak = Uk Σk VkT = AVk VkT = Uk UkT A minimizes kA − Xkξ over P all matrices X ∈ Rm×n of rank 2 2 2 at most k. Specifically, kA − Ak k2 = σk+1 (A) and kA − Ak kF = ρi=k+1 σi2 (A) (see [32]). A† = T n×m denotes the so-called Moore-Penrose pseudo-inverse of A ∈ Rm×n (here Σ−1 VA Σ−1 A A UA ∈ R is the inverse of ΣA ). By the SVD of A and A† , for all i = 1, . . . , ρ = rank(A) = rank(A† ): σi (A† ) = 1/σρ−i+1 (A). 4.1 The best rank k matrix ΠξV,k (A) within a subspace V Our analysis uses theory involving computing the best rank k approximation of a matrix A within a given column subspace V. Let A ∈ Rm×n , let k < n be an integer, and let V ∈ Rm×c with m×n is the best rank k approximation to A in the column span of V. k < c < n. ΠF V,k (A) ∈ R Equivalently, we can write ΠF V,k (A) = VXopt , where Xopt = X∈R argmin kA − VXk2F . :rank(X)≤k c×n In order to compute ΠF V,k (A) given A, V, and k, one can use the following algorithm: V = YΨ is a qr decomposition of V with Y ∈ Rm×c and Ψ ∈ Rc×c. This step requires O(mc2 ) arithmetic operations. 2: Ξ = Y T A ∈ Rc×n . This step requires O(mnc) arithmetic operations. 1: T Ξk = ∆Σ̃k Ṽk ∈ Rc×n is a rank k SVD of Ξ with ∆ ∈ Rc×k , Σ̃k ∈ Rk×k , and Ṽk ∈ Rn×k . This step requires O(nc2 ) arithmetic operations. 4: Return Y∆∆T Y T A ∈ Rm×n of rank at most k. 3: Notice that Y∆∆T Y T A ∈ Rm×n is a rank k matrix that lies in the column span of V. The next lemma is a simple corollary of Lemma 4.3 in [21]. T Lemma 23. Given A ∈ Rm×n , V ∈ Rm×c and an integer k, Y∆∆T Y T and QŨk Σ̃k Ṽk satisfy: T 2 kA − Y∆∆T Y T Ak2F ≤ kA − Y∆Σ̃k Ṽk k2F = kA − ΠF V,k (A)kF The above algorithm requires O(mnc + nc2 ) arithmetic operations to construct Y, Ψ, and ∆. We will denote the above procedure as [Y, Ψ, ∆] = BestColumnSubspaceSV D(A, V, k). Proof. The equality was proven in Lemma 4.3 in [21]. To prove the inequality, notice that for any matrix X : kA − Y∆(Y∆)† Ak2F ≤ kA − Y∆Xk2F . Also, (Y∆)† = ∆† Y† = ∆T Y T , because both matrices are orthonormal. 18 Next, we state a version of the above lemma for the transpose of A, equivalently for the best rank k approximation within the row space of a given subspace R. Let A ∈ Rm×n , let k < n be an m×n is the best rank k approximation to A integer, and let R ∈ Rc×n with k < c < m. ΠF R,k (A) ∈ R in the row span of R. Equivalently, we can write ΠF R,k (A) = Xopt R, where Xopt = argmin kA − XRk2F . X∈Rm×c :rank(X)≤k In order to compute ΠF R,k (A) given A, R, and k, one can use the following algorithm: RT = YZ is a qr decomposition of RT with Y ∈ Rn×c and Z ∈ Rc×c . This step requires O(nc2 ) arithmetic operations. 2: Ξ = AY ∈ Rm×c . This step requires O(mnc) arithmetic operations. 3: Ξk = Ũk Σ̃k ∆T ∈ Rm×c is a rank k SVD of Ξ with Ũk ∈ Rm×k , Σ̃k ∈ Rk×k , and ∆ ∈ Rc×k . This step requires O(mc2 ) arithmetic operations. 4: Return AY∆∆T Y T ∈ Rm×n of rank at most k. 1: Notice that AY∆∆T Y T ∈ Rm×n is a rank k matrix that lies in the row span of R. The next lemma is a corollary of Lemma 23 when applied with A := AT and V := RT . Lemma 24. Given A ∈ Rm×n , R ∈ Rc×n and an integer k, Y∆∆T Y T and Ũk Σ̃k ∆T satisfy: 2 kA − AY∆∆T Y T k2F ≤ kA − Ũk Σ̃k ∆T Rk2F = kA − ΠF R,k (A)kF . The above algorithm requires O(mnc + mc2 ) arithmetic operations to construct Y, Z, and ∆. We will denote the above procedure as [Y, Z, ∆] = BestRowSubspaceSV D(A, R, k). 5 Distributed PCA in the arbitrary partition model This section describes a fast distributed PCA algorithm with total communication O(msk) words plus low order terms, which is optimal in the arbitrary partition model in the sense that an Ω̃(msk) bits lower bound was given by [37]. The algorithm employs, in a novel way, the notion of projection-cost preserving sketches from [24]. In particular, whereas all previous [48, 23] dimensionreduction-based SVD methods reduce one dimension of the input matrix to compute some approximation to the SVD, our method reduces both dimensions and computes an approximation to the SVD from a small almost square matrix. Unlike [37] which reduces only one dimension in the first communication round, we do the reduction on both dimensions in the same round. We first present the batch version of the algorithm which offers a new low-rank matrix approximation technique; a specific implementation of this algorithm offers a communication-optimal distributed PCA algorithm (we also discuss in Section 7 a variant of this algorithm that offers a two-pass space-optimal PCA method in the turnstile streaming model). Before presenting all these new algorithms in detail, we present the relevant results from the previous literature that we employ in the analysis. 19 5.1 Projection-cost preserving sketching matrices In this section, we recap a notion of sketching matrices which we call “projection-cost preserving sketching matrices”. A sketching matrix from this family is a linear matrix transformation and it has the property that for all projections it preserves, up to some error, the difference between the matrix in hand and its projection in Frobenius norm. Definition 25 (Projection-cost preserving sketching matrices). We say that W ∈ Rn×ξ is an (ε, k)projection-cost preserving sketching matrix of A ∈ Rm×n , if for all rank-k orthogonal projection matrices P ∈ Rm×m , it satisfies (1 − ε)kA − PAk2F ≤ kAW − PAWk2F + c ≤ (1 + ε)kA − PAk2F where c is a non-negative constant which only depends on A and W. We also call AW an (ε, k)-projectioncost preserving sketch of A. Due to the following lemma, we know that a good rank-k approximation projection matrix of (ε, k)-projection-cost preserving sketch AW also provides a good rank-k approximation to A. Lemma 26 (PCA via Projection-Cost Preserving Sketches - Lemma 3 in [24]). Suppose W ∈ Rn×ξ is ∗ an (ε, k)-projection-cost preserving sketching matrix of A ∈ Rm×n . Let P̂ = arg minrank(P)≤k kAW − ∗ PAWk2F . For all P̂, ε′ satisfying rank(P̂) ≤ k, ε′ ≥ 0, if kAW− P̂AWk2F ≤ (1+ε′ )kAW− P̂ AWk2F , kA − P̂Ak2F ≤ 1+ε · (1 + ε′ )kA − Ak k2F 1−ε [24] also provides several ways to construct projection-cost preserving sketching matrices. Because we mainly consider the communication, we just choose one which can reduce the dimension as much as possible. Furthermore, it is also an oblivious projection-cost preserving sketching matrix. Lemma 27 (Dense Johnson-Lindenstrauss matrix - part of Theorem 12 in [24]). √ For ε < √ 1, suppose n×ξ each entry of W ∈ R is chosen O(log(k))-wise independently and uniformly in {1/ ξ, −1/ ξ} where ξ = O(kε−2 ) [21]. For any A ∈ Rm×n , with probability at least 0.99, W is an (ε, k)-projection-cost preserving sketching matrix of A. 5.2 A batch algorithm for the fast low rank approximation of matrices In this section, we describe a new method for quickly computing a low-rank approximation to a given matrix. This method does not offer any specific advantages over previous such techniques [48, 23, 18]; however, this new algorithm can be implemented efficiently in the distributed setting (see Section 5.3) and in the streaming model of computation (see Section 7); in fact we are able to obtain communication-optimal and space-optimal results, respectively. For completeness as well as ease of presentation, we first present and analyze the simple batch version of the algorithm. The algorithm uses the dense Johnson-Lindenstrauss matrix of Lemma 27 in order to reduce both dimensions of A, before computing some sort of SVD to a poly(k/ε) × poly(k/ε) matrix (see Step 2 in the algorithm below). Consider the usual inputs: a matrix A ∈ Rm×n , a rank parameter k < rank(A), and an accuracy parameter 0 < ε < 1. The algorithm below returns an orthonormal matrix U ∈ Rm×k such that kA − UUT Ak2F ≤ (1 + ε) · kA − Ak k2F . Algorithm 20 1. Construct two dense Johnson-Lindenstrauss matrices S ∈ Rξ1 ×m , T ∈ Rn×ξ2 with ξ1 = O(kε−2 ), ξ2 = O(kε−2 ) (see Lemma 27). 2. Construct à = SAT. T 3. Compute the SVD of Ãk = UÃk ΣÃk Và (UÃk ∈ Rξ1 ×k , ΣÃk ∈ Rk×k , VÃk ∈ Rξ2 ×k ). k 4. Construct X = ATVÃk 5. Compute an orthonormal basis U ∈ Rm×k for span(X) (notice that rank(X) ≤ k). Theorem 29 later in this section analyzes the approximation error and the running time of the previous algorithm. First, we prove the accuracy of the algorithm. Lemma 28. The matrix U ∈ Rm×k with k orthonormal columns satisfies with probability at least 0.98: kA − UUT Ak2F ≤ (1 + ε) · kA − Ak k2F . Proof. T T T k2F = kSAT − SATVÃk Và k2F = kTT AT ST − VÃk Và TT AT ST k2F kà − Ãk k2F = kà − ÃVÃk Và k k k T The first equality follows by the SVD of Ãk = UÃk ΣÃk Và . The second equality is by the k construction of à = SAT. The third equality is due to ∀M, kMk2F = kM T k2F . Due to Lemma 27, with probability at least 0.99, ST is an (ε, k)-projection-cost preserving sketch matrix of TT AT . According to Lemma 26, T T kAT − ATVÃk Và k2F = kTT AT − VÃk Và TT AT k2F ≤ k k 1+ε · kAT − (AT)k k2F 1−ε (2) Observe that kUUT AT − ATk2F = kXX† AT − ATk2F T ≤ kXVà − ATk2F k T = kATVÃk Và − ATk2F k ≤ 1+ε · kAT − (AT)k k2F 1−ε The first equality uses the fact that U is an orthonormal basis of span(X). The first inequality is followed by ∀X, M, N, kXX† M − Mk2F ≤ kXN − Mk2F . The second equality uses the construction that X = ATVÃk . The second inequality follows by Eqn (2). Due to Lemma 27, with probability at least 0.99, T is an (ε, k)-projection-cost preserving sketch matrix of A. Due to Lemma 26, kUUT A − Ak2F ≤ (1 + ε)2 · kA − Ak k2F (1 − ε)2 Due to union bound, the probability that ST is an (ε, k)-projection-cost preserving sketch matrix of TT AT and T is an (ε, k)-projection-cost preserving sketch matrix of A is at least 0.98. Note that (1+ε)2 is 1 + O(ε) when ε is small enough, so we can adjust ε here by a constant factor to show the (1−ε)2 statement. Next, we present the main theorem. 21 Theorem 29. The matrix U ∈ Rm×k with k orthonormal columns satisfies with probability at least 0.98: kA − UUT Ak2F ≤ (1 + ε) · kA − Ak k2F . The running time of the algorithm is
O nmkε−2 + mk2 ε−4 + poly(kε−1 ) . Proof. The correctness is shown by Lemma 28. Running time. Next, we analyze the running time of the algorithm: 1. There are a total of (ξ1 × m + ξ2 × n) entries of S and T. It is enough to generate them in O((n + m)kε−2 ) operations. 2. We first compute AT with O(mnξ2 ) arithmetic operations. Then, we compute SAT with O(ξ1 mξ2 ) arithmetic operations. 3. This step requires O(poly(k/ε)) operations since we compute the SVD of a O(poly(k/ε)) × O(poly(k/ε)) matrix Ã. 4. We already have AT from the second step. Hence, O(mξ2 k) additional arithmetic operations suffice to compute ATVÃk 5. O(mk2 ) operations suffice to compute an orthonormal basis for X, e.g., with a QR factorization. 5.3 The distributed PCA algorithm Ps Recall that the input matrix A ∈ Rm×n is partitioned arbitrarily as: A = i Ai for i = 1 : s, m×n Ai ∈ R . The idea in the algorithm below is to implement the algorithm in Section 5.2 in the distributed setting. Input: 1. A ∈ Rm×n arbitrarily partitioned A = 2. rank parameter k < rank(A) Ps i Ai for i = 1 : s, Ai ∈ Rm×n . 3. accuracy parameter ε > 0 Algorithm 1. Machines agree upon two dense Johnson-Lindenstrauss matrices S ∈ Rξ1 ×m , T ∈ Rn×ξ2 with ξ1 = O(kε−2 ), ξ2 = O(kε−2 ) (see Lemma 27). P 2. Each machine locally computes Ãi = SAi T and sends Ãi to the server. Server constructs à = i Ãi . T 3. Server computes the SVD of Ãk = UÃk ΣÃk Và (UÃk ∈ Rξ1 ×k , ΣÃk ∈ Rk×k , VÃk ∈ Rξ2 ×k ). k 4. Server sends VÃk to all machines. 5. Each machine construct Xi = Ai TVÃk and sends Xi to the server. Server constructs X = 6. Server computes an orthonormal basis U ∈ R m×k 7. Server sends U to each machine. for span(X) (notice that rank(X) ≤ k). P i Xi . Notice that in the first step, S and T can be described using a random seed that is O(log(k))wise independent due to Lemma 27. 22 5.3.1 Main result The theorem below analyzes the approximation error, the communication complexity, and the running time of the previous algorithm. Notice that the communication cost of this algorithm is only given in terms of “real numbers”. The only step where we can not bound the length of a machine word is when the server communicates VÃk to all machines; and this is because the entries of VÃk could be unbounded (see the discussion regarding the upper bounds in Section 1.2). We resolve this issue in the following section. Theorem 30. The matrix U ∈ Rm×k with k orthonormal columns satisfies w.p. 0.98: kA − UUT Ak2F ≤ (1 + ε) · kA − Ak k2F . (3) The communication cost of the algorithm is O(msk + s · poly(k/ε)) “real numbers” and the running time is of the order
O nmkε−2 + mk2 ε−4 + poly(kε−1 ) . Proof. The matrix U - up to the randomness in the algorithm - is exactly the same matrix as in the batch algorithm in Section 5.2, hence Theorem 29 proves Eqn. 3. The algorithm communicates O(msk + s · poly(k/ε)) real numbers in total: O(s · poly(k/ε)) in steps 2 and 4, and O(smk) in steps 5 and 7. The operations in the algorithm are effectively the same operations as in the batch algorithm in Section 5.2, hence the analysis of the running time in Theorem 29 shows the claim. 6 Obtaining bit complexity for the distributed PCA algorithm For the algorithm in the previous section, we were only able to provide a communication upper bound in terms of “real numbers”. In this section, we describe how to obtain a communication upper bound in terms of words for the above protocol, where each word is O(log(mnsk/ε)) bits. The basic idea is that we have a case analysis depending on the rank of the matrix A. If the rank of A is less than or equal to 2k, we follow one distributed protocol and if the rank is at least 2k we follow a different protocol. In Section 6.1, Section 6.2, and Section 6.3 we describe the algorithm that tests the rank of a distributed matrix, and the two PCA protocols, respectively. Then, in Section 6.4 we give the details of the overall algorithm and in Section 6.5 we give its analysis. 6.1 Testing the rank of a distributed matrix Lemma 31. Given A ∈ Rm×n and a rank parameter k < rank(A), there exists a distributed protocol in the arbitrary partition model to test if the rank of A is less than or equal to 2k using O(sk2 ) words of communication and succeeding with probability 1 − δ for an arbitrarily small constant δ > 0. Proof. This is an immediate implementation of a streaming algorithm due to [21] for testing if an n × n matrix A has rank at least 2k in the streaming model, using O(k2 ) words of space. In that algorithm, there is a fixed 6nk/δ × n matrix H whose entries are integers of magnitude at most poly(n), where δ > 0 is an arbitrarily small constant. The algorithm simply chooses 4k random 23 rows from H. Letting H′ be the 2k × n matrix of the first 2k random rows, and H′′ be the n × 2k matrix whose columns are the next 2k randomly chosen rows, the algorithm just declares that A has rank at least k iff the rank of H′ AH′′ is 2k. The above streaming algorithm can be implemented in the distributed setting by having the coordinator choose 4k random rows of the fixed, known matrix, and send the row identities to each of the machines. This only takes O(sk) words of communication. Then machine i computes H′ Ai H′′ , and returns this to the coordinator. The coordinator can then add these up to compute H′ AH′′ and compute its rank. The total communication is O(sk 2 ) words and the protocol succeeds with probability at least 1 − δ for an arbitrarily small constant δ > 0. Note that we can assume our input matrix A, which is m × n, is a square n × n matrix by padding with all-zeros rows. Those rows of course, will never get communicated in the implementation described above. 6.2 Distributed PCA protocol when rank(A) ≤ 2k 6.2.1 Subsampled Randomized Hadamard Transform and Affine Embeddings Our algorithms use the following tool, known as “Subsampled Randomized Hadamard Transform” or SRHT for short, to implement efficiently fast dimension reduction in large matrices. Definition 32 (Normalized Walsh–Hadamard Matrix). Fix an integer m = 2p , for p = 1, 2, 3, .... The (non-normalized) m × m matrix of the Walsh–Hadamard transform is defined recursively as,     Hm/2 Hm/2 +1 +1 . , with H2 = Hn = +1 −1 Hm/2 −Hm/2 1 The m × m normalized matrix of the Walsh–Hadamard transform is equal to H = m− 2 Hm ∈ Rm×m . Definition 33 (Subsampled Randomized Hadamard Transform (SRHT) matrix). Fix integers ξ and m = 2p with ξ < m and p = 1, 2, 3, .... An SRHT matrix is an ξ × m matrix of the form r m T= · RHD; ξ • D ∈ Rm×m is a random diagonal matrix whose entries are independent random signs, i.e. random variables uniformly distributed on {±1}. • H ∈ Rm×m is a normalized Walsh–Hadamard matrix (see Definition 32). • R ∈ Rξ×m is a subset or r rows from the n × n identity matrix, where the rows are chosen uniformly at random and without replacement. The next lemma argues that an SRHT matrix is a so-called “affine embedding matrix”. The SRHT is one of the possible choices of Lemma 32 in [23] that will satisfy the lemma. Lemma 34 (Affine embeddings - Theorem 39 in [23]). Suppose G and H are matrices with m rows, and G has rank at most r. Suppose T is a ξ × m SRHT matrix (see Definition 33) with ξ = O(r/ε2 ). Then, with probability 0.99, for all X simultaneously: (1 − ε) · kGX − Hk2F ≤ kT(GX − H)k2F ≤ (1 + ε) · kGX − Hk2F . Finally, we note that matrix-vector multiplications with SRHT’s are fast. Lemma 35 (Fast Matrix-Vector Multiplication, Theorem 2.1 in [6]). Given x ∈ Rm and ξ < m, one can construct T ∈ Rξ×m and compute Tx in at most 2m log2 (ξ + 1)) operations. 24 6.2.2 Generalized rank-constrained matrix approximations Let M ∈ Rm×n , N ∈ Rm×c , L ∈ Rr×n , and k ≤ c, r be an integer. Consider the following optimization problem, Xopt ∈ argmin kM − NXLk2F . X∈Rc×r ,rank(X)≤k Then, the solution Xopt ∈ Rc×r with rank(Xopt ) ≤ k that has the minimum kXopt kF out of all possible feasible solutions is given via the following formula,   T T L† . MVL VL Xopt = N† UN UN k
T T T T MVL VL ∈ UN UN MVL VL k ∈ Rm×n of rank at most k denotes the best rank k matrix to UN UN m×n R . This result was proven in [30] (see also [49] for the spectral norm version of the problem). 6.2.3 The PCA protocol Lemma 36. Suppose the rank ρ of A ∈ Rm×n satisfies ρ ≤ 2k, for some rank parameter k. Then, there is a protocol for the Distributed Principal Component Analysis Problem in the arbitrary partition model using O(smk + sk 2 /ε2 ) words of communication and succeeding with probability 1 − δ for an arbitrarily small constant δ > 0. Proof. The n×2k matrix H′′ chosen in the protocol of Lemma 31 satisfies that with probability 1−δ, for an arbitrarily small constant δ > 0, the rank of AH′′ is equal to the rank of A if ρ < 2k. Indeed, if this were not true, the algorithm could not be correct, as the rank of H′ AH′′ is at most the rank of AH′′ (and the same algorithm can be used for any ρ < 2k). It follows that with probability 1 − δ, the column span of AH′′ is equal to the column span of A. Hence, as in the protocol of Lemma 31, the coordinator learns the column space of A, which can be described with O(km) words. The coordinator thus communicates this to all machines, using O(skm) total words of communication, assuming the entries of A are integers of magnitude at most poly(mns/ε). Let C = AH′′ , which is m × 2k. We can set up the optimization problem: min kCXCT A − AkF . rank(X)≤k (4) Because the size of C is only m × 2k, every machine can know C by sending a total of O(smk) words. Since the rank of C and A are small, we can sketch on the left and right using affine embeddings Tlef t and Tright . Then machine i sends CT Ai Tright to the coordinator, together with Tlef t Ai Tright . The coordinator computes CT ATright and Tlef t ATright by adding up the sketches, and sends these back to all the machines. Each machine can then solve the optimization problem min kTlef t CXCT ATright − Tlef t ATright kF , rank(X)≤k (5) each obtaining the same X∗ which is the optimal solution to Eqn (5). Due to Lemma 34, X∗ is a (1 + O(ε))-approximation to the best solution to Eqn (4). Finally, every machine outputs the same orthonormal basis U ∈ Rm×k for CX∗ . We can construct affine embedding matrices Tlef t ∈ Rξ1 ×m and Tright ∈ Rn×ξ2 with ξ1 = O(k/ε2 ), ξ2 = O(k/ε2 ) (see Definition 33). The total communication of this protocol is O(skm + sk 2 /ε2 ) words and the success probability can be made 1−δ for an arbitrarily small constant δ > 0. 25 6.3 Distributed PCA protocol when rank(A) > 2k The idea here is more involved than in the previous subsection and in order to describe the algorithm we need several intermediate results. 6.3.1 Lower bounds on singular values of matrices with integers entries The first lemma gives a lower bound on the singular values of a matrix with integer entries with bounded magnitude. Lemma 37. (Lemma 4.1 of [21], restated) If an m × n matrix A has integer entries bounded in magnitude by γ, and has rank ρ = rank(A), then the k-th largest singular value A satisfies σk ≥ (mnγ 2 )−k/(2(ρ−k)) . Proof. In the proof of Lemma 4.1 of [21], equation (10), it is shown that if λk is the k-th largest eigenvalue of AT A, then λk ≥ (mnγ 2 )−k/(ρ−k) . This implies the k-th singular value σk of A satisfies σk ≥ (mnγ 2 )−k/(2(ρ−k)) . Next, we state two immediate corollaries of this lemma for future reference. Corollary 38. If an m × n matrix A has integer entries bounded in magnitude by poly(mns/ε) and kA − Ak kF > 0, then kA − Ak kF > (mns/ε)−O(k) . Proof. Since kA − Ak kF > 0, the rank ρ of A is at least k + 1, so by the preceding lemma, kA − Ak kF ≥ σk+1 ≥ (poly(mns/ε))−k/2 , as desired. Corollary 39. If an m × n matrix A has integer entries bounded in magnitude by poly(mns/ε) and rank(A) ≥ 2k, then kA − Ak kF > 1/poly(mns/ε). Proof. This follows by plugging ρ = 2k into Lemma 37. 6.3.2 Lower bounds on singular values of integer-perturbed matrices In this section, we describe a perturbation technique for matrices and provide lower bounds for the smallest singular value of the perturbed matrix. We start with a theorem of Tao and Vu. Theorem 40. (Theorem 2.5 of [52]) Let M be an n×n matrix with integer entries bounded in magnitude by nC for a constant C > 0. Let Nn be a matrix with independent entries each chosen to be 1 with probability 1/2, and −1 with probability 1/2. Then, there is a constant B > 0 depending on C, for which Pr[k(M + Nn )−1 k2 ≥ nB ] ≤ 1/n. In words, the result indicates that the spectral norm of the inverse of the perturbed matrix is bounded from below with high probability. We now describe a simple corollary of this result. 26 Corollary 41. Let M be an n × n matrix with integer entries bounded in magnitude by nC for a constant C > 0. Let Nn be a matrix with independent entries each chosen to be 1/nD with probability 1/2, and −1/nD with probability 1/2, where D > 0 is a constant. Then, there is a constant B > 0 depending on C and D for which Pr[k(M + Nn )−1 k2 ≥ nB ] ≤ 1/n. Proof. This follows by Theorem 40 after replacing the constat C in that theorem with C + D, and scaling by nD . We need to generalize Corollary 41 to rectangular matrices since we will eventually apply this perturbation technique to the matrix A to which we would like to compute a distributed PCA. Lemma 42. Let M be an m × n matrix with integer entries bounded in magnitude by nC for a constant C > 0, and suppose m ≤ n. Let Nm,n be a matrix with independent entries each chosen to be 1/nD with probability 1/2 and −1/nD with probability 1/2, where D > 0 is a constant. Then, there is a constant B > 0 depending on C and D for which Pr[σm (M + Nm,n ) < 1/nB ] ≤ 1/n, where σm (M + Nm,n ) denotes the smallest singular value of M + Nm,n . Proof. Suppose we were to pad M with n − m zero rows, obtaining a square matrix M. Now consider the n × n matrix Nn,n with independent entries each chosen to be 1/nD with probability 1/2 and −1/nD with probability 1/2. By Corollary 41, all singular values of M + Nn,n are at least 1/nB . Now consider a unit vector x ∈ Rn which is zero on all but its top m coordinates. Let y ∈ Rm be the unit vector which agrees with x on its top m coordinates. Then, 1/nB ≤ kx(M + Nn,n )k2 = ky(M + Nm,n )k2 , where the inequality uses the lower bound on the singular values of M + Nn,n , which occurs with probability at least 1 − 1/n, and the equality follows by definition of the matrices and vectors we have defined. As y ∈ Rm can be chosen to be an arbitrary unit vector, it follows that σm (M + Nm,n ) ≥ 1/nB , which completes the proof. 6.4 Description of algorithm Using the above results, we are now ready to describe a distributed PCA algorithm whose communication cost can be bounded in terms of machine words and not just in terms of “real numbers”. As in the algorithm in Section 5.3, we denote with Bi ∈ Rm×n the matrix that B1 arises from A1 after applying the Bernoulli perturbation technique P discussed above in Lemma 42 and ∀i > 1, Bi is equal to Ai . Using this notation, we have B := i Bi ∈ Rm×n . Notice that B exactly arises from A after applying the such Bernoulli perturbation technique. Input: 1. A ∈ Rm×n arbitrarily partitioned A = 2. rank parameter k < rank(A) Ps i Ai for i = 1 : s, Ai ∈ Rm×n . 3. accuracy parameter ε > 0 Algorithm 27 1. Use the protocol of Lemma 31 with δ = 0.01 to test if the rank of A is less than 2k. 2. If rank(A) ≤ 2k, use the protocol of Lemma 36 to find some orthonormal U ∈ Rm×k . 3. If rank(A) > 2k, (a) machine 1 locally and independently adds 1/nD with probability 1/2, and −1/nD with probability 1/2, to each of the entries of A1 , where D is the constant of Lemma 42. Note that this effectively adds the matrix Nm,n of Lemma 42 to the entire matrix A. For notational convenience let B = A + Nm,n , B1 ∈ Rm×n be the local perturbed matrix of machine 1 and ∀i > 1, Bi is just equal to Ai . (b) Machines agree upon two dense Johnson-Lindenstrauss matrices S ∈ Rξ1 ×m , T ∈ Rn×ξ2 with ξ1 = O(kε−2 ), ξ2 = O(kε−2 ) (see Lemma 27). (c) Each machine locally computes B̃i = SBi T and sends B̃i to the server. Server constructs B̃ = P i B̃i . T (d) Server computes the SVD of B̃k = UB̃k ΣB̃k VB̃ (UB̃k ∈ Rξ1 ×k , ΣB̃k ∈ Rk×k , VB̃k ∈ Rξ2 ×k ). k (e) Now server rounds each of the entries in VB̃k to the nearest integer multiple of 1/nγ for a sufficiently large constant γ > 0. Let the matrix after the rounding be V̂B̃k (f) Server sends V̂B̃k to all machines. (g) Each machine construct X̂i = Bi TV̂B̃k and sends X̂i to the server. Server constructs X̂ = P i X̂i . (h) Server computes an orthonormal basis U ∈ Rm×k for span(X̂) (notice that rank(X̂) ≤ k), e.g. with a QR factorization. (i) Server sends U to each machine. 6.5 Main result The theorem below analyzes the approximation error, the communication complexity, and the running time of the previous algorithm. Notice that the communication cost of this algorithm is given in terms of machine words. Theorem 43. The matrix U ∈ Rm×k with k orthonormal columns satisfies with arbitrarily large constant probability: (6) kA − UUT Ak2F ≤ (1 + ε) · kA − Ak k2F . The communication cost of the algorithm is O(msk + s · poly(k/ε)) words and the running time is
O nmkε−2 + mk2 ε−4 + poly(kε−1 ) . 6.6 Proof of Theorem 43 Proof of Eqn. 6 If rank(A) ≤ 2k, then Lemma 36 shows the claim. If rank(A) > 2k, then the situation is more involved and we prove the approximation bound in the following subsection. Communication Complexity Step 1 requires O(sk 2 ) words (Lemma 31). Step 2 requires O(smk+ sk 2 /ε2 ) words (Lemma 36). Step 3 requires O(msk + s · poly(k/ε)) words and this follows from the analysis in Theorem 30. The only difference with Theorem 30 is that now all matrices communicated in the algorithm have real numbers which can be represented efficiently with one machine word of at most O(log(mns/ε)) bits. To see this, note that each entry of V̂B̃k is bounded in magnitude by nO(B+C)/γ . Therefore, the entries of V̂B̃k and the entries of Xi = Bi TV̂B̃k can each be described using O(log n) bits, i.e., a constant number of machine words. 28 Running time The operations in the third step in algorithm are effectively the same operations as in the batch algorithm in Section 5.2 (plus some extra operations whose running time is not asymptotically larger), hence the analysis of the running time in Theorem 29 shows the claim (the number of operations in the first two steps of the algorithm is asymptotically less than the number of operations in the third step). 6.7 Proof of Eqn. 6 if rank(A) > 2k ξ×p be a matrix of which Lemma 44 (Result in [21]). Suppose M ∈ Rp×q . For ε > 0, let√P ∈ R√ entries are O(q)-wise independently and uniformly chosen from {−1/ ξ, +1/ ξ} where ξ = O(q/ε2 ). With arbitrarily large constant probability, P is an ε-subspace embedding matrix of column space of M. Specifically, ∀x, it has (1 − ε)kMxk2 ≤ kPMxk2 ≤ (1 + ε)kMxk2 Lemma 45 (Lemma C.2 in [40]). Suppose P is an ε-subspace embedding matrix of column space of M ∈ Rp×q (q ≤ p). With arbitrarily large constant probability, ∀1 ≤ i ≤ q (1 − ε)σi (PM) ≤ σi (M) ≤ (1 + ε)σi (PM) where σi means the ith singular value. It is suffice to show the following: Lemma 46. With arbitrarily large constant probability, k(BTVB̃k )† k2 ≤ 8nB Proof. Notice that BTVB̃k is an m × k matrix. With high probability, S is a ε-subspace embedding matrix of column space of BTVB̃k due to Lemma 44. According to Lemma 45, σmin (BTVB̃k ) ≥ (1 − ε)σmin (SBTVB̃k ) where σmin means the minimum singular value. Without loss of generality, we assume ε < 1/2. Then σmin (BTVB̃k ) ≥ 12 σmin (SBTVB̃k ). Because VB̃k are the top k right singular vectors of SBT, σmin (SBTVB̃k ) ≥ σmin (SBT). Applying Lemma 44 again, TT is a constant error subspace embedding matrix of column space of BT ST . Combining with Lemma 45, we can make σmin (SBT) ≥ 21 σmin (SB). Notice that Lemma 44 also implies that q ξ1 T ξ1 which means ∀x ∈ Rξ1 , we can m S is a constant error subspace embedding matrix of R q q make kx ξm1 Sk2 > 12 kxk2 . Thus, ∀x ∈ Rξ1 , kx ξm1 SBk2 > 12 σmin (B)kxk2 . Due to Lemma 42, q q σmin (B) > 1/nB which implies σmin (SB) > 12 · ξm1 σmin (B) > 12 ξm1 n1B > 12 n1B . To conclude k(BTVB̃k )† k2 ≤ 1/σmin (BTVB̃k ) ≤ 8nB . Let E = VB̃k − V̂B̃k , we have kA − UUT AkF = kB − Nm,n − UUT (B − Nm,n )kF ≤ kB − UUT BkF + kNm,n kF † = kB − (BTV̂B̃k )(BTV̂B̃k ) BkF + kNm,n kF ≤ kB − (BT(VB̃k + E))(BTVB̃k )† BkF + kNm,n kF ≤ kB − (BTVB̃k )(BTVB̃k )† BkF + kBTE(BTVB̃k )† BkF + kNm,n kF ≤ (1 + O(ε))kB − Bk kF + kBTE(BTVB̃k )† BkF ≤ (1 + O(ε))kB − Bk kF 29 The first equality follows by using the relation B = A + Nm,n . The first inequality uses the triangle inequality for the Frobenius norm and the fact that I − UUT is a projector matrix and can be dropped without increasing the Frobenius norm. The second equality uses the fact that both † UUT and (BTV̂B̃k )(BTV̂B̃k ) are the same projector matrices. The second inequality follows by ∀M, C, X, kM − CC† MkF ≤ kM − CXkF and the relation E = VB̃k − V̂B̃k . The third inequality uses the triangle inequality. The fourth inequality uses that kB − Bk kF ≥ 1/poly(mns/ε) (follows from Lemma 42) while kNm,n kF can be made 1/nt for an arbitrarily large integer t, and the fact that kB − (BTVB̃k )(BTVB̃k )† BkF ≤ (1 + O(ε))kB − Bk kF (implied by replacing all A with B in Lemma 28). The last inequality uses the fact kBTE(BTVB̃k )† BkF ≤ kBkF kTkF kEkF k(BTVB̃k )† kF kBkF where kBkF , kTkF , k(BTVB̃k )† kF are poly(nms/ε) (Lemma 46 implies the bound of k(BTVB̃k )† kF ) and we can make |E|F be 1/np for an arbitrarily large p, and kB − Bk kF ≥ 1/poly(mns/ε). Overall, after rescaling ε, we have kA − UUT AkF ≤ (1 + ε)kB − Bk kF . (7) Finally, we need to relate kB − Bk kF to kA − Ak kF , which we do in the following lemma. Lemma 47. Let A ∈ Rm×n be any matrix, Nm,n be the random matrix of Lemma 42, and B = A + Nm,n . Let k < rank(A), ε > 0. Then, for arbitrarily large constant probability, kB − Bk kF ≤ (1 + ε)kA − Ak kF . Proof. Since B = A + Nm,n , the proof idea is to relate the singular values of B to the singular values of A using Weyl’s inequality. Specifically, for i : 1 : m (recall we assume m ≤ n) we have |σi (B) − σi (A)| ≤ kNm,n k2 ≤ 1/nD . Rearranging terms in this inequality, and taking squares of the resulting relation we obtain (for i = 1 : m),
2 σi2 (B) ≤ σi (A) + 1/nD . (8) We manipulate the term kB − Bk kF as follows: v u X u m σi2 (B) kB − Bk kF = t i=k+1 v u X u m (σi (A) + 1/nD )2 ≤ t i=k+1 v u X u m = t σi2 (A) + 1/n2D + 2σi (A)/nD i=k+1
v u X q q u m 2 2D t σi (A) + 1/n + 2σi (A)/nD ≤ i=k+1 ≤ kA − Ak kF + ε · kA − Ak kF + ε · kA − Ak kF ≤ (1 + 2ε)kA − Ak kF . 30 The first inequality uses Eqn. (8). In the third inequality, we used q 1/n2D ≤ εkA − Ak kF , which follows from Corollary 39 for a sufficiently large constant D. Also, we used q q q 2σi (A)/nD ≤ 2σ1 (A)/nD ≤ poly(nms/ε)/nD ≤ εkA − Ak kF , where the second inequality follows because kAk2 ≤ poly(nms/ε) and the last inequality uses again Corollary 39 for a sufficiently large constant D. Completing the proof. Using Lemma 47 in Eqn (7), taking squares in the resulting inequality, and rescaling ε concludes the proof. 7 Streaming Principal Component Analysis In this section, we are interested in computing a PCA of a matrix in turnstile streaming model. Specifically, there is a stream of update operations that the q th operation has form (iq , jq , xq ) which indicates that the (iq , jq )th entry of A ∈ Rm×n should be incremented by xq where iq ∈ {1, ..., m}, jq ∈ {1, ..., n}, xq ∈ R. Initially, A is zero. In the streaming setting of computation, we are allowed only one pass over the update operations, i.e., the algorithm “sees” each update operation one by one and only once. Upon termination, the algorithm should return a matrix U with k orthonormal columns which is a “good” basis for span(A) . In Section 7.1.1 below, we describe an algorithm which gives a space-optimal streaming algorithm. Further more, we provide a variation of this algorithm in Section 7.1.2 which can output A∗k satisfying kA − A∗k k2F ≤ (1 + ε) · kA − Ak k2F . It meets the space lower bound shown in [21]. In Section 7.2, we relax the problem and we describe a two-pass streaming algorithm which is a version of the algorithm of Section 5.2; unfortunately, the space complexity of this algorithm can be bounded only in terms of “real numbers”. We fix this in Section 7.3 where we describe an optimal two-pass streaming algorithm. Inputs to the algorithms in Section 7.1, Section 7.2 and Section 7.3 are a stream of update operations (i1 , j1 , x1 ), (i2 , j2 , x2 ), ..., (il , jl , xl ), a rank parameter k < rank(A), and an accuracy parameter 0 < ε < 1. 7.1 One-pass streaming PCA 7.1.1 The algorithm which outputs U In the following algorithm, the output should be U with orthogonal unit columns and satisfying kA − UUT Ak2F ≤ (1 + ε) · kA − Ak k2F Our algorithm uses the following property of random sign matrices: Lemma 48 (Sketch for regression - Theorem 3.1 in [21]). Suppose both of A and B have m rows and rank(A) ≤ k. Further more, if each entry of S ∈ Rξ×m is O(k)-wise independently chosen from {−1, +1} where ξ = O(k/ε) and X̃ = arg min kSAX − SBk2F X , with probability at least 0.99, kAX̃ − Bk2F ≤ (1 + ε) · min kAX − Bk2F X 31 Algorithm 1. Construct random sign sketching matrices S ∈ Rξ1 ×m with ξ1 = O(kε−1 ) and R ∈ Rn×ξ2 with ξ2 = O(kε−1 ) (see Lemma 48) 2. Construct affine embedding matrices Tlef t ∈ Rξ3 ×m and Tright ∈ Rn×ξ4 with ξ3 = O(kε−3 ), ξ4 = O(k/ε−3 ) (see Definition 33). 3. Initialize all-zeros matrices: M ∈ Rξ3 ×ξ4 , L ∈ Rξ1 ×ξ4 , N ∈ Rξ3 ×ξ2 , D ∈ Rm×ξ2 . We will maintain M, L, N, D such that M = Tlef t ATright , L = SATright , N = Tlef t AR and D = AR. 4. For (iq , jq , xq ) := (i1 , j1 , x1 ), ..., (il , jl , xl ) (one pass over the stream of update operations) (a) For all i = 1, ..., ξ3 , j = 1, ..., ξ4 , let Mi,j = Mi,j + (Tlef t )i,iq · xq · (Tright )jq ,j . (b) For all i = 1, ..., ξ1 , j = 1, ..., ξ4 , let Li,j = Li,j + Si,iq · xq · (Tright )jq ,j . (c) For all i = 1, ..., ξ3 , j = 1, ..., ξ2 , let Ni,j = Ni,j + (Tlef t )i,iq · xq · Rjq ,j . (d) For all j = 1, ..., ξ2 , let Diq ,j = Miq ,j + xq · Rjq ,j 5. end 6. Construct X∗ = argmin kN·X·L−Mk2F . (Notice that X∗ = argmin kTlef t (ARXSA − A) Tright k2F .) rank(X)≤k rank(X)≤k T (UX∗ ∈ Rξ×k , ΣX∗ ∈ Rk×k VX∗ ∈ Rξ×k ); then, compute 7. Compute the SVD of X∗ = UX∗ ΣX∗ VX ∗ T = DUX∗ . 8. Compute an orthonormal basis U ∈ Rm×k for span(T). Theorem 52 later in this section analyzes the approximation error, the space complexity, and the running time of the previous algorithm. First, we prove a few intermediate results. Lemma 49. For all matrices X ∈ Rξ2 ×ξ1 and with probability at least 0.98: (1 − ε)2 kARXSA − Ak2F ≤ kTlef t (ARXSA − A) Tright k2F ≤ (1 + ε)2 kARXSA − Ak2F Proof. Notice that rank(AR) ≤ ξ2 = O(kε−1 ). Then from Lemma 34 (G := AR and H := A), with probability at least 0.99 for all Y ∈ Rξ2 ×n : (1 − ε)kARY − Ak2F ≤ kTlef t (ARY − A)k2F ≤ (1 + ε)kARY − Ak2F Replacing Y = XSA, for an arbitrary X ∈ Rξ2 ×ξ1 , we obtain that with probability at least 0.99 and for all X ∈ Rξ2 ×ξ1 simultaneously: (1 − ε)kARXSA − Ak2F ≤ kTlef t (ARXSA − A) k2F ≤ (1 + ε)kARXSA − Ak2F Now notice that rank(SA) ≤ ξ1 = O(kε−1 ). Then from Lemma 34 (G := AT ST and H := T m×ξ2 : AT Tlef t ), with probability at least 0.99 for all Z ∈ R T 2 T T T T T T 2 T T T T T 2 (1 − ε)kA T ST ZT − AT Tlef t kF ≤ kTright (A S Z − A Tlef t )kF ≤ (1 + ε)kA S Z − A Tlef t kF T ξ2 ×ξ1 , we obtain that with probability at least Replacing Z = XT RT AT Tlef t for an arbitrary X ∈ R 0.99 and for all X simultaneously: (1 − ε)2 k(ARXSA − A)T k2F ≤ k(Tlef t (ARXSA − A) Tright )T k2F ≤ (1 + ε)2 k(ARXSA − A)T k2F 32 Lemma 50. Let Xopt = arg minrank(X)≤k kARXSA−Ak2F with Xopt ∈ Rξ2 ×ξ1 . Then, with probability at least 0.98 (1 + ε)2 min kTlef t (ARXSA − A) Tright k2F ≤ kARX∗ SA − Ak2F ≤ kARXopt SA − Ak2F 2 (1 − ε) rank(X)≤k Proof. From Lemma 49, we have that for all matrices X ∈ Rξ2 ×ξ1 and with probability at least 0.98: (1 − ε)2 · kARXSA − Ak2F ≤ kTlef t (ARXSA − A) Tright k2F ≤ (1 + ε)2 · kARXSA − Ak2F Applying this for X := Xopt , we obtain: (e) (1 − ε)2 · kARXopt SA − Ak2F ≤ kTlef t (ARXopt SA − A) Tright k2F ≤ (1 + ε)2 · kARXopt SA − Ak2F Applying this for X := X∗ , we obtain: (f ) (1 − ε)2 · kARX∗ SA − Ak2F ≤ kTlef t (ARX∗ SA − A) Tright k2F ≤ (1 + ε)2 · kARXopt SA − Ak2F Combining (f ), (e) along with the optimality of X∗ , formally using the relation: kTlef t (ARX∗ SA − A) Tright k2F ≤ kTlef t (ARXopt SA − A) Tright k2F , shows the claim. In words, the lemma indicates that in order to (1 + ε)-approximate the optimization problem minrank(X)≤k kARXSA − Ak2F , it suffices to “sketch” the matrix ARXSA − A from left and right with the matrices Tlef t and Tright . Recall that X∗ = argmin kTlef t (ARXSA − A) Tright k2F . rank(X)≤k Lemma 51. Let Xopt = arg minrank(X)≤k kARXSA − Ak2F . Then, with probability at least 0.98 kARXopt SA − Ak2F ≤ (1 + ε) · kA − Ak k2F , T T Proof. Suppose the SVD of A = UA ΣA VA and Ak = UAk ΣAk VA . Consider about the following k regression problem: min kUAk X − Ak2F rank(X)≤k Since we can choose X to be T UA A, k we have min rank(X)≤k kUAk X − Ak2F ≤ kA − Ak k2F Let X̃ = arg minrank(X)≤k kSUAk X − SAk2F . According to Lemma 48, with probability 0.99 kUAk X̃ − Ak2F ≤ (1 + ε) min rank(X)≤k kUAk X − Ak2F ≤ (1 + ε)kA − Ak k2F Since X̃ minimizes kSUAk X − SAk2F , the row space of X̃ is in the row space of SA (Otherwise, we can project SUAk X into the row space of SA to get a better solution). We denote X̃ = WSA where rank(W) ≤ k. Now, consider about another regression problem: min rank(X)≤k k(WSA)T XT − AT k2F 33 If we choose X to be UAk , we have min rank(X)≤k k(WSA)T XT − AT k2F ≤ kUAk X̃ − Ak2F ≤ (1 + ε)kA − Ak k2F Let X̂ = arg minrank(X)≤k kRT (WSA)T XT − RT AT k2F . According to Lemma 48, with probability 0.99 T kRT (WSA)T X̂ − RT AT k2F ≤ (1 + ε) min rank(X)≤k k(WSA)T XT − AT k2F ≤ (1 + ε)2 kA − Ak k2F Since X̂ minimizes kXWSAR − ARk2F , the column space of X̂ is in the column space of AR. We denote X̂ = ARG where rank(G) ≤ k. Thus, we have kARXopt SA − Ak2F ≤ kARGWSA − Ak2F ≤ (1 + ε)2 kA − Ak k2F We scale the ε with a constant factor. The statement is shown by applying union bound. Theorem 52. The matrix U ∈ Rm×k with k orthonormal columns satisfies w.p. 0.96: kA − UUT Ak2F ≤ (1 + O (ε)) · kA − Ak k2F .
The space complexity of the algorithm is O mk/ε + poly(kε−1 ) words, and the running time for each
update operation is O(poly(kε−1 )) and the total running time is of the order O l · poly(kε−1 ) + mk2 ε−1 where l is total number of updates. Proof. Because UUT projects the columns of A into the column space of Tlef t ARX∗ , kA − UUT Ak2F ≤ kTlef t (ARXSA − A) Tright k2F According to Lemma 50 and Lemma 51, kA − UUT Ak2F ≤ (1 + O (ε)) · kA − Ak k2F To see the space complexity of the algorithm, observe that it suffices to maintain the matrices in the fourth step of the algorithm. Furthermore, observe that by the end of the stream: M = Tlef t ATright ∈ Rξ3 ×ξ4 , L = SATright ∈ Rξ1 ×ξ4 , N = Tlef t AR ∈ Rξ3 ×ξ2 and D = AR ∈ Rm×ξ2 . Those matrices form the so called “sketch” of the algorithm. Since all of ξ1 , ξ2 , ξ3 , ξ4 are O(kε−1 ), the running time for each update operation in the fourth step is O(poly(kε−1 )). We can do the sixth step by using the result of Section 6.2.2. The computation takes poly(kε−1 ) running time. In the seventh step, computing SVD needs poly(kε−1 ), and computing T needs O(mk2 ε−1 ). We can use O(mk2 ) to compute U in the final step, e.g. QR decomposition. 7.1.2 A variation which outputs A∗k We just slightly modify the previous algorithm in Section 7.1.1 to get the following one which can output a matrix A∗k ∈ Rm×n satisfying kA − A∗k k2F ≤ (1 + ε) · kA − Ak k2F Algorithm 34 1. Construct random sign sketching matrices S ∈ Rξ1 ×m with ξ1 = O(kε−1 ) and R ∈ Rn×ξ2 with ξ2 = O(kε−1 ) (see Lemma 48) 2. Construct affine embedding matrices Tlef t ∈ Rξ3 ×m and Tright ∈ Rn×ξ4 with ξ3 = O(kε−3 ), ξ4 = O(k/ε−3 ) (see Definition 33). 3. Initialize all-zeros matrices: M ∈ Rξ3 ×ξ4 , L ∈ Rξ1 ×ξ4 , N ∈ Rξ3 ×ξ2 , D ∈ Rm×ξ2 and C ∈ Rξ1 ×n . We will maintain M, L, N, D, C such that M = Tlef t ATright , L = SATright , N = Tlef t AR, D = AR and C = SA. 4. For (iq , jq , xq ) := (i1 , j1 , x1 ), ..., (il , jl , xl ) (one pass over the stream of update operations) (a) For all i = 1, ..., ξ3 , j = 1, ..., ξ4 , let Mi,j = Mi,j + (Tlef t )i,iq · xq · (Tright )jq ,j . (b) For all i = 1, ..., ξ1 , j = 1, ..., ξ4 , let Li,j = Li,j + Si,iq · xq · (Tright )jq ,j . (c) For all i = 1, ..., ξ3 , j = 1, ..., ξ2 , let Ni,j = Ni,j + (Tlef t )i,iq · xq · Rjq ,j . (d) For all j = 1, ..., ξ2 , let Diq ,j = Miq ,j + xq · Rjq ,j (e) For all i = 1, ..., ξ1 , let Ci,jq = Ci,jq + Si,iq · xq 5. end 6. Construct X∗ = argmin kN·X·L−Mk2F . (Notice that X∗ = argmin kTlef t (ARXSA − A) Tright k2F .) rank(X)≤k rank(X)≤k T (UX∗ ∈ Rξ×k , ΣX∗ ∈ Rk×k VX∗ ∈ Rξ×k ); then, compute 7. Compute the SVD of X∗ = UX∗ ΣX∗ VX ∗ T = DUX∗ T C K = VX ∗ 8. Output A∗k = TΣX∗ K Theorem 53. With probability at least 0.96: kA − A∗k k2F ≤ (1 + O (ε)) · kA − Ak k2F .
The space complexity of the algorithm is O (m + n)k/ε + poly(kε−1 ) words, and the running time for each update operation is O(poly(kε−1 )) and the total running time is of the order
O l · poly(kε−1 ) + (m + n)k2 ε−1 + mnk where l is total number of updates. Proof. Notice that A∗k = ARX∗ SA. According to Lemma 50 and Lemma 51, we have kA − A∗k k2F ≤ (1 + O (ε)) · kA − Ak k2F . The total space of “sketch” matrices M, L, N, D is the same as the algorithm in Section 7.1.1. The maintenance of C needs additional O(nkε−1 ) space. T The running time is almost the same as before. Computation of K = VX C needs additional ∗ O(nk2 ε−1 ), and computation of A∗k = TΣX∗ K needs additional O(mnk). 35 7.2 Two-pass streaming PCA with “real numbers” space complexity A simple modification of the batch algorithm in Section 5.2 leads to a two-pass streaming algorithm with O(mk) + poly(k/ε) “real numbers” space complexity: Algorithm 1. Construct two dense Johnson-Lindenstrauss matrices S ∈ Rξ1 ×m , T ∈ Rn×ξ2 with ξ1 = O(kε−2 ), ξ2 = O(kε−2 ) (see Lemma 27). 2. Initialize all-zero matrices: Ã ∈ Rξ1 ×ξ2 , X ∈ Rm×k . 3. For (iq , jq , xq ) := (i1 , j1 , x1 ), ..., (il , jl , xl ) (first pass over the stream) 4. For all i = 1, ..., ξ1 , j = 1, ..., ξ2 , let Ãi,j = Ãi,j + Si,iq · xq · Tjq ,j . 5. end T 6. Compute the SVD of Ãk = UÃk ΣÃk VÃ (UÃk ∈ Rξ1 ×k , ΣÃk ∈ Rk×k , VÃk ∈ Rξ2 ×k ). k 7. For (iq , jq , xq ) := (i1 , j1 , x1 ), ..., (il , jl , xl ) (second pass over the stream) 8. For all i = 1, ..., ξ2 , j = 1, ..., k, let Xiq ,j = Xiq ,j + xq · Tjq ,i · (VÃk )i,j . 9. end 10. Compute an orthonormal basis U ∈ Rm×k for span(X) (notice that rank(X) ≤ k). The theorem below analyzes the approximation error, the space complexity, and the running time of the previous algorithm. Notice that the space complexity of this algorithm is only given in terms of “real numbers” - we resolve this issue in the following section. Theorem 54. The matrix U ∈ Rm×k with k orthonormal columns satisfies w.p. 0.98: kA − UUT Ak2F ≤ (1 + O (ε)) · kA − Ak k2F . The space complexity of the algorithm is
O mk + poly(kε−1 ) “real numbers”, the running time of each update operation is O(poly(kε−1 )), and the total running time is of the order
O l · poly(kε−1 ) + mk2 where l is the total number of update operations. Proof. The matrix U - up to the randomness in the algorithms - is exactly the same matrix as in the algorithm in Section 5.2, hence Theorem 29 shows the claim. To see the space complexity of the algorithm, observe that it suffices to maintain the matrices in the fourth and eighth steps of the algorithm. Furthermore, observe that by the end of the stream: Ã = SAT, and X = ATVÃk Since ξ1 , ξ2 are poly(kε−1 ), the running time for each update operation is poly(kε−1 ). Computing SVD in the sixth step needs poly(kε−1 ). Computing U in the final step needs O(mk 2 ). 36 7.3 Two-pass streaming PCA with bit space complexity The previous algorithm could suffer from large space complexity in case we need a lot of machine words to write down the entries of VÃk . To fix this issue we use the same idea as in the case of the distributed PCA algorithm in Section 6. This leads to a two-pass streaming algorithm for PCA. Again, as in the distributed case, we need to test if the rank of A is less than 2k, and then we use one approach if rank(A) < 2k and another approach if rank(A) ≥ 2k. Both of these approaches can be implemented with two passes. In the overall streaming PCA algorithm that we would like to design, we can not wait for the algorithm that tests the rank to finish and then start running one of the two PCA protocols, because this will lead to a three-pass algorithm (one pass to test the rank and two passes for the actual PCA protocol). To keep the number of passes to two, we just start running the PCA protocols in parallel with the protocol that tests the rank of the input matrix. In the end of the first pass over the stream of update operations, we already know which protocol to follow, and we just do this, disregarding the other protocol. We already discussed the streaming version of the algorithm that tests the rank of the matrix in Lemma 31. Below, we describe separately the two streaming PCA protocols. 7.3.1 Streaming PCA protocol when rank(A) ≤ 2k The idea here is to implement a streaming version of the distributed PCA protocol in Lemma 36. Algorithm 1. Construct an n × 2k matrix H′′ as in Lemma 36. 2. Construct affine embedding matrices Tlef t ∈ Rξ1 ×m and Tright ∈ Rn×ξ2 with ξ1 = O(k/ε2 ), ξ2 = O(k/ε2 ) (see Definition 33). 3. Initialize all-zeros matrices: C ∈ Rm×2k , M ∈ Rξ1 ×ξ2 , L ∈ Rξ×ξ2 , N ∈ Rξ1 ×2k . 4. (iq , jq , xq ) := (i1 , j1 , x1 ), ..., (il , jl , xl ) (first pass over the stream) (a) For all j = 1, ..., 2k, let Ciq ,j = Ciq ,j + xq · H′′jq ,j 5. end 6. For (iq , jq , xq ) := (i1 , j1 , x1 ), ..., (il , jl , xl ) (second pass over the stream) (a) For all i = 1, ..., ξ1 , j = 1, ..., ξ2 , let Mi,j = Mi,j + (Tlef t )i,iq · xq · (Tright )jq ,j . (b) For all i = 1, ..., 2k, j = 1, ..., ξ2 , let Li,j = Li,j + CT i,iq · xq · (Tright )jq ,j . (c) For all i = 1, ..., ξ1 , j = 1, ..., 2k, let Ni,j = Ni,j + (Tlef t )i,iq · xq · H′′jq ,j . 7. end 8. Construct X∗ = argmin kN · X · L − Mk2F := argmin kTlef t AH′′ X(H′′ )T AT ATright − Tlef t ATright k2F . rank(X)≤k rank(X)≤k 9. Compute an orthonormal basis U ∈ Rm×k for span(CX∗ ). 37 7.3.2 Streaming PCA protocol when rank(A) > 2k The idea here is to implement a streaming version of step 3 of the algorithm in Section 6.4. Algorithm 1. Construct two dense Johnson-Lindenstrauss matrices S ∈ Rξ1 ×m , T ∈ Rn×ξ2 with ξ1 = O(kε−2 ), ξ2 = O(kε−2 ) (see Lemma 27). 2. Initialize all-zero matrices: B̃ ∈ Rξ1 ×ξ2 , X̂ ∈ Rm×k . 3. For p := 1, ..., ξ1 , q := 1, ..., ξ2 , i := 1, ..., m, j := 1, ..., n 4. Let ri,j = −1/nD with probability 1/2 and ri,j = 1/nD with probability 1/2. 5. B̃p,q = B̃p,q + Sp,i · ri,j · Tj,q 6. end 7. For (iq , jq , xq ) := (i1 , j1 , x1 ), ..., (il , jl , xl ) (first pass over the stream) 8. For all i = 1, ..., ξ1 , j = 1, ..., ξ2 , let B̃i,j = B̃i,j + Si,iq · xq · Tjq ,j . 9. end T 10. Compute the SVD of B̃k = UB̃k ΣB̃k VB̃ (UB̃k ∈ Rξ1 ×k , ΣB̃k ∈ Rk×k , VB̃k ∈ Rξ2 ×k ). Then, we k round each of the entries of VB̃k to the nearest integer multiple of 1/nγ for a sufficient large constant γ > 0. Let the matrix after the rounding be V̂B̃k . 11. For (iq , jq , xq ) := (i1 , j1 , x1 ), ..., (il , jl , xl ) (second pass over the stream) 12. For all i = 1, ..., ξ2 , j = 1, ..., k, let X̂iq ,j = X̂iq ,j + xq · Tjq ,i · (V̂B̃k )i,j . 13. end 14. Compute an orthonormal basis U ∈ Rm×k for span(X̂) (notice that rank(X̂) ≤ k). 7.3.3 Main result The theorem below analyzes the approximation error, the space complexity, and the running time of the previous algorithm. Notice that the space complexity of this algorithm is given in terms of machine words. Theorem 55. The matrix U ∈ Rm×k with k orthonormal columns satisfies with arbitrarily large constant probability: kA − UUT Ak2F ≤ (1 + ε) · kA − Ak k2F . The space complexity of the algorithm is
O mk + poly(kε−1 ) machine words, the running time of each update operation is O(poly(kε−1 )), and the total running time is of the order
O l · poly(kε−1 ) + mk2 where l is the total number of update operations. Proof. The matrix U - up to the randomness in the algorithms - is exactly the same matrix as in the algorithm in Theorem 43, hence the approximation bound in that theorem shows the claim. To see the space complexity of the algorithm, observe that it suffices to maintain all the “sketch” matrices in the both two protocols above. Since ξ1 , ξ2 in both protocols are poly(kε−1 ), the running time for each update operation is O(poly(kε−1 )). Additional O(mk 2 ) running time is caused by computing U in the final step. 38 8 Distributed PCA for sparse matrices in column-partition model Recall that in the problem of Definition 15 we are given
P 1) an m × n matrix A partitioned columnwise as A = A1 A2 . . . As , with Ai ∈ Rm×wi ( wi = n); 2) a rank parameter k < rank(A); 3) an accuracy parameter ε > 0. We would like to design an algorithm that finds an m × k matrix U with UT U = Ik and, upon termination, leaves this matrix U in each machine of the network. The high level idea of our algorithm is to find, in a distributed way, a small set - O(kε−1 ) - of columns from A and then find U in the span of those columns. To choose those O(kε−1 ) columns of A in a distributed way we implement the following three-stage sampling procedure: 1. Local sampling: Each machine samples O(k) columns using the sampling algorithm from [24]. 2. Global sampling: The server collects the columns from each machine and down-samples them to O(k) columns using the deterministic algorithm from [17]. 3. Adaptive sampling: the server sends back to each machine those O(k) columns; then, an extra of O(kε−1 ) columns are selected randomly from the entire matrix A using [26]. We argue that if C̃ contains the columns selected with this three-stage approach, then, w.p. 0.99, † (A)k2F ≤ (1 + O(ε)) · kA − Ak k2F . kA − C̃C̃ Ak2F ≤ kA − ΠF C̃,k (A) to be the rank k matrix that approximates A, we are not Though we could have used ΠF C̃,k (A). To address this issue, using familiar with any communication-efficient computation of ΠF C̃,k (A)k2F ; an idea from [37], we compute U ∈ span(C̃) such that kA−UUT Ak2F ≤ (1+O(ε))kA−ΠF C̃,k this U can be calculated with small communication cost and it is sufficient for our purposes. Before presenting the new algorithm in detail, we discuss results from previous literature that we employ in the analysis. 8.1 Background material 8.1.1 Column sampling algorithms Our distributed PCA algorithm in Section 8 samples columns from the input matrix in three stages: 1. Local sampling: O(k) columns are selected locally in each machine. 2. Global sampling: O(k) columns are selected in the server. 3. Adaptive sampling: an extra O(kε−1 ) columns are selected from the entire matrix A. In the first stage, we use a sampling algorithm mentioned in [24]. Actually, the sampling algorithm is the same as the Batson, Spielman, and Strivastava (BSS) spectral sparsification algorithm [11]. But Cohen et al. showed that a small set of columns sampled by BSS sampling algorithm can provide a projection-cost preserving sketch. In the second stage, we use a deterministic algorithm developed in [17], which itself extends the Batson, Spielman, and Strivastava (BSS) spectral sparsification algorithm [11]. For the actual algorithm in the first stage we defer the reader to Theorem 15 in [24]. Lemma 56 states the result. And for the actual algorithm in the second stage we defer the reader to Lemma 3.6 in [17]. Lemma 57 and Lemma 58 below present the relevant results. In the third sampling stage, we use an adaptive sampling algorithm from [26]. 39 Lemma 56 (BSS Sampling for Projection-Cost Preserving Sketch. Theorem 15 in [24]). Let A ∈ Rm×w be an arbitrary matrix. For any integer 0 < k < m, real numbers 0 < ε < 1, 0 < δ, there is a randomized algorithm that runs in O(nnz(A) log(m/δ) + m · poly(k, ε, log(1/δ))) time, and constructs a “sampling/rescaling” w × O(k/ε2 ) matrix S with probability at least 1 − δ such that S is an (ε, k)-projection-cost preserving sketching matrix of A. We denote this sampling procedure as S = BssSamplingI(A, k, ε, δ). In words, there exists a randomized algorithm that runs in near input-sparsity running time can construct a projection-cost preserving sketch of a given matrix A by sampling and rescaling a small subset of columns of A. Lemma 57 (Dual Set Spectral-Frobenius Sparsification. Lemma 3.6 in [17]). Let V ∈ Rw×k be a matrix with w > k and V T V = Ik . Let E ∈ Rm×w be an arbitrary matrix. Then,
given an integer ℓ such that k < ℓ ≤ w, there exists a deterministic algorithm that runs in O ℓwk2 + mw time, and constructs a “sampling/rescaling” w × ℓ matrix S such that    p 2 σk2 V T S ≥ 1 − k/ℓ , kESk2F ≤ kEk2F . Specifically, rank(V T S) = k. We denote this sampling procedure as S = BssSamplingII(V, E, ℓ). In words, given V and E, there exists an algorithm to select (and rescale) a small number of columns from E and rows from V such that: 1) the Frobenius norm squared of the sampled E is less than the Frobenius norm squared of E; 2) the sampled V has rank equal to the rank of V; and  p 2 3) the smallest singular value squared of the sampled V is bounded from below by 1 − k/ℓ . Lemma 58 (Constant factor column-based matrix reconstruction; Theorem 5 in [17]). Given matrix G ∈ Rm×α of rank ρ
and a target rank k 1 , there exists a deterministic algorithm that runs in O αm min{α, m} + αck2 time and selects c > k columns of G to form a matrix C ∈ Rm×c with † kG − CC Gk2F ≤   rank  X(G) p −2 1 + 1 − k/c · σi2 (G). i=k+1 The algorithm in this theorem finds C as C = GS, where S = BssSamplingII(V, G − GVV T , c) and V ∈ Rα×k contains the top k right singular vectors of G. We denote this sampling procedure as C = DeterministicCssF robenius(G, k, c). In words, there exists a deterministic algorithm, running in polynomial time, to select any number of c > k columns from the given matrix G, such that the residual error, in Frobenius norm, from projecting G onto the span of the sampled columns is bounded from above with respect to the residual error of the best rank k matrix from the SVD of G. Notice that rank X(G) i=k+1 σi2 (G) = kG − Gk k2F , where Gk ∈ Rm×α has rank at most k and is computed via the SVD of G. 1 The original Theorem 5 in [17] has the assumption that k < ρ, but this assumption can be dropped having the result unchanged. The only reason the assumption k < ρ exists is because otherwise column subset selection is trivial. 40 Before presenting the adaptive sampling theorem from [26] we introduce some useful notation m×n is the used in the lemma. Let A ∈ Rm×n , k < n, and V ∈ Rm×c with k < c < n. ΠF V,k (A) ∈ R best rank k approximation to A - wrt the Frobenius norm - in the column span of V. Equivalently, ΠF V,k (A) = VXopt , where Xopt = argmin kA − VXk2F . X∈Rc×n ,rank(X)≤k Lemma 59 (Adaptive sampling; Theorem 2.1 of [26]). Given A ∈ Rm×n and V ∈ Rm×c1 (with c1 ≤ n, m), define the residual matrix Ψ = A − VV† A ∈ Rm×n . For j = 1, . . . , n, let pj be a probability distribution such that pj ≥ βkΨ(j) k22 /kΨk2F , for some 1 > β > 0, where Ψ(j) is the j-th column of the matrix Ψ. Sample c2 columns from A in c2 i.i.d. trials, where in each trial the j-th column is chosen with probability pj . Let C2 ∈ Rm×c2 contain the c2 sampled columns and let C = [V C2 ] ∈ Rm×(c1 +c2 ) contain the columns of V and C2 . Then, for any integer k > 0, rank i h X(A)   2 F † 2 E kA − CC AkF ≤ E kA − ΠC,k (A)kF ≤ σi2 (A) + i=k+1 k · kA − VV† Ak2F . β · c2 Given A and C, this method requires O(c1 mn) time to compute Ψ, another O(mn) time to compute the probabilities pj ’s and another O(n + c2 ) time for the sampling step - using the method in [51]. In total, the method requires O(c1 mn + c2 ) time to compute C2 . We denote this sampling procedure as C2 = AdaptiveCols(A, V, c2 , β). In words, given the matrix A and the subspace V, there exists a randomized algorithm to sample an additional of c2 columns from A, based on probabilities from the residual error matrix A − VV† A, such that residual error, in Frobenius norm, from projecting A onto the span of the union of the columns of V and the sampled columns is bounded from above with respect to the best rank k approximation to A, the residual A − VV† A, and the number of sampled columns c2 . 8.1.2 Distributed adaptive sampling In our distributed PCA algorithm below, we also need to use a distributed version of the previous adaptive sampling procedure. We provide some preliminary results for that task in this section. Lemma 60. Suppose that the columns of an m × n matrix A are partitioned arbitrarily across the machines into matrices A1 , . . . , As . Let C be an arbitrary m × r matrix. Consider the distribution p on n columns in which kaj − CC† aj k2F , pj = kA − CC† Ak2F where aj is the j-th column of A (j = 1 : n here). For each i ∈ [s], let some value βi satisfies kAi − CC† Ai k2F ≤ β i ≤ 2kAi − CC† Ai k2F . For each j ∈ [n], if column aj is held on the i-th server (denoted aij ), then let q j = Ps βi i′ =1 βi′ Then for each j ∈ [n], · kaij − CC† aij k22 kAi − CC† Ai k2F pj /2 ≤ qj ≤ 2pj . 41 . Proof. By definition of the β i , we have that Hence, kAi − CC† Ai k2F 2kAi − CC† Ai k2F βi P ≤ . ≤ s 2kA − CC† Ak2F kA − CC† Ak2F i′ =1 βi′ pj /2 ≤ qj ≤ 2pj . Lemma 61. Suppose the coordinator has an m × r matrix C of columns of an m × n matrix A, where r = O(k). Suppose the entries of A are integers bounded by poly(mns/ε) in magnitude, and let the columns of A be partitioned arbitrarily across the servers into matrices A1 , . . . , As . There is a protocol using O(skm) machine words of communication for the coordinator to learn values i β so that for all i ∈ [s], kAi − CC† Ai k2F ≤ β i ≤ 2kAi − CC† Ai k2F . Proof. The coordinator first sends C to all machines. The i-th server locally computes kAi − CC† Ai k2F . If this number is 0, he/she sends 0 to the coordinator, and in this case β i satisfies the requirement in the statement of the lemma. Otherwise, consider the matrix Bi which is formed by concatenating the columns of Ai with those of C. Then kBi − CC† Bi k2F = kAi − CC† Ai k2F > 0, (9) since the columns of Bi in C contribute a cost of 0. However, since Bi contains the columns of C and its cost is non-zero, it implies the rank of Bi is at least r + 1. By Corollary 38, this implies kBi − CC† Bi k2F > (mns/ε)−O(k) , which by (9) gives the same lower bound on kAi − CC† Ai k2F . Note also that kAi − CC† Ai k2F ≤ poly(mns/ε), since kAk2F ≤ poly(mns/ε). This implies if the i-th machine sends β i to the coordinator, where β i is the nearest power of 2 to kAi − CC† Ai k2F , there will only be O(k log(mns/ε)) possible values of β i , and hence each can be specified using O(log k + log log(mns/ε)) bits, i.e., a single machine word. This completes the proof. 8.1.3 Low-rank matrix approximations within a subspace The final stage of our distributed PCA algorithm below finds U ∈ span(C̃) such that the error of the residual matrix A − UUT A is “small”. To implement this step, we employ an algorithm developed in [37]. Lemma 62. Let A ∈ Rm×n be the input matrix and V ∈ Rm×c be the input subspace. We further assume that for some rank parameter k < c and accuracy parameter 0 < ε < 1 : 2 2 kA − ΠF V,k (A)kF ≤ (1 + O(ǫ))kA − Ak kF . Let V = YΨ be a qr decomposition of V with Y ∈ Rm×c and Ψ ∈ Rc×c . Let Ξ = Y T AW T ∈ Rc×ξ , √ √ where W T ∈ Rn×ξ with ξ = O(c/ε2 ), each element of which is chosen i.i.d. to be {+1/ n, −1/ n} with 42 probability 1/2. Let ∆ ∈ Rc×k contain the top k left singular vectors of Ξ. Then, with probability at least 1 − e−c , kA − Y∆∆T Y T Ak2F ≤ (1 + ε)kA − Ak k2F . Y, and ∆ can be computed in O(mnξ) time. We denote this procedure as [Y, ∆] = ApproxSubspaceSV D(A, V, k, ε). Proof. This result was essentially proven inside the proof of Theorem 1.1 in [36]. Specifically, the error bound proven in [36] is for the transpose of A (also Y, ∆ are denoted with U, V in [36]). As for the running time, given A, V, and k, one can compute (i) Y in O(mc2 ) time; (ii) Ξ in O(mnξ + mcξ) time; and (iii) ∆ in O(c2 ξ) time. In words, the lemma indicates that given the matrix A and the subspace V such that 2 2 kA − ΠF V,k (A)kF ≤ (1 + O(ǫ))kA − Ak kF , for some small ε > 0, there exists a randomized algorithm to compute matrices Y ∈ span(V) and ∆ such that the residual error, in Frobenius norm, from projecting A onto the span of Y∆ is also bounded by (1 + O(ǫ))kA − Ak k2F . Notice that ∆ contains the top k left singular vectors of a “sketched” version of Y T A, a property which makes the computation particularly effective in terms of running time. 8.2 Detailed description of the algorithm Input: 1. A ∈ Rm×n partitioned column-wise A = A1 A2 2. rank parameter k < rank(A)
P . . . As ; for i = 1 : s, Ai ∈ Rm×wi ; i wi = n. 3. accuracy parameter ε > 0 Algorithm 1. Local Column Sampling (a) For each sub-matrix Ai ∈ Rm×wi , compute Ci ∈ Rm×ℓ containing ℓ = O(k) columns from Ai as follows: Ci = Ai Si . Here, Si has dimensions wi × ℓ and is constructed as follows: Si = BssSamplingI(Ai , 4k, 1/2, 0.01/s) (see Lemma 56). (b) Machine i sends Ci to the server. 2. Global Column Sampling (a) Server constructs m × (s
· ℓ) matrix G containing (s · ℓ) actual columns from A as follows: G = C1 C2 . . . Cs . Then, server constructs C ∈ Rm×c1 via choosing c1 = 4k columns from G as follows: C = DeterministicCssF robenius(G, k, c1 ) (see Lemma 58). (b) Server sends C to all the machines. 3. Adaptive Column Sampling (a) Machine i computes Ψi = Ai − CC† Ai ∈ Rm×wi and then computes βi as it was described in Lemma 61. Machine i sends βi to server. (b) Server computes probability distribution gi = Pβiβi . Server samples i.i.d. with replacement i ⌈ 50k/ε ⌉ samples (machines) from gi . Then, server determines numbers ti (i = 1, 2, . . . , s), where ti is the number of times the ith machine was sampled. It sends the ti ’s to the machines. 43 (c) Machine i computes probabilities qji = kxk22 /kΨi k2F (j = 1 : wi ), where x is the jth column of Ψi . And now machine i samples ti samples fromP it’s local probability distribution and sends the corresponding columns to the server. Let c2 = i ti = ⌈ 50k/ε ⌉. (d) Server collects the columns and assigns them to Ĉ ∈ Rm×c2 . Server constructs C̃ to be the
m × (c1 + c2 ) matrix: C̃ = C; Ĉ . Let c = c1 + c2 = 4k + ⌈ 50k/ε ⌉. 4. Rank-k matrix in the span of C̃ (a) Server sends C̃ to all the machines and each machine computes (the same) qr factorization of C̃: C̃ = YR where Y ∈ Rm×c has orthonormal columns and R ∈ Rc×c is upper triangular. 1/2. Implicitly (b) Machine i generates W̃i ∈ Rξ×wi with ξ = O(c/ε2 ) to be i.i.d.
{+1, −1} w.p.ξ×n , with W̃ ∈ R . Machine i all machines together generate W̃ = W̃1 W̃2 . . . W̃s T computes Hi = C̃ (Ai WiT ) ∈ Rc×ξ . Machine i sends Hi to the server. P (c) Server computes Ξ = si=1 Hi ∈ Rc×ξ and sends this back to all the machines. Now machines
√ compute Ξ := R−1 · Ξ · 1/ n(:= Y T AWT ), where ∈ Rξ×n is W √= W1 W2 . . . Ws √ random matrix each element of which is {+1/ n, −1/ n} w.p. 1/2, and then they compute ∆ ∈ Rc×k to be the top k left singular vectors of Ξ. Each machine computes U = Y · ∆ ∈ Rm×k . Discussion. A few remarks are necessary for the last two stages of the algorithm. The third stage (adaptive column sampling), implements the adaptive sampling method of Lemma 59. To see this, note that each column in A (specifically the jth column in Ai ) is sampled with probability gi · qji ≥ 1 ||xij ||22 , · 2 kΨk2F where Ψ = A − CC† A. This follows from Lemma 60. Overall, this stage effectively constructs Ĉ such that (see Lemma 59) Ĉ = AdaptiveCols(A, C, c2 , 1/2). The last stage in the algorithm implements the algorithm in Lemma 62. To see this, note that W satisfies the properties in the lemma. Hence, [Y, ∆] = ApproxSubspaceSV D(A, C̃, k, ε). 8.3 Main Result The theorem below analyzes the approximation error, the communication complexity, and the running time of the previous algorithm. Theorem 63. The matrix C̃ ∈ Rm×c with c = O(k/ε) columns of A satisfies w.p. at least 0.98, † (A)k2F ≤ (1 + O(ε)) · kA − Uk UkT Ak2F . kA − C̃C̃ Ak2F ≤ kA − ΠF C̃,k (10) The matrix U ∈ Rm×k with k orthonormal columns satisfies with probability at least 0.97, kA − UUT Ak2F ≤ (1 + O(ε)) · kA − Uk UkT Ak2F . (11) Let each column of A
has at most φ non-zero elements. Then, the communication cost of the algorithm is O skφε−1 + sk2 ε−4 and the running time is O (mns · poly(k, 1/ε)) . 44 8.4 Proof of Theorem 63 8.4.1 Proof of Eqn. 10 From Lemma 59: h † E kA − C̃C̃ Ak2F i h ≤ E kA − (A)k2F ΠF C̃,k i ≤ rank X(A) σi2 (A) + i=k+1 2ε · kA − CC† Ak2F . 50 (12) Therefore, we need to bound kA − CC† Ak2F now. Lemma 64. With probability at least 0.99, matrix G is a (4k, 1/2)-projection-cost preserving sketch of A. Proof. For each i, Ci = Ai Si is a (4k, 1/2)-projection-cost preserving sketch of Ai with probability at least 1 − 0.01/s. Due to union bound, with probability at least 1 − 0.01, ∀i, Ci is a (4k, 1/2)projection-cost preserving sketch of Ai . Since Ci is a (4k, 1/2)-projection-cost preserving sketch of Ai , we can assume ci is a constant which is independent from projection matrix P which has rank at most 4k such that 3 1 kAi − PAi k2F ≤ kCi − PCi k2F + ci ≤ kAi − PAi k2F 2 2 Then, for any projection matrix P which has rank at most 4k we have: s 1 1X kAi − PAi k2F kA − PAk2F = 2 2 i=1 ≤ s X i=1 kCi − PCi k2F + ci s ≤ 3X kAi − PAi k2F 2
i=1 3 ≤ kA − PAk2F 2 Consider the right hand side of the first inequality, it is equal to kG − Let constant c = Ps i=1 ci , PGk2F + s X ci i=1 we have: 1 3 kA − PAk2F ≤ kG − PGk2F + c ≤ kA − PAk2F 2 2 Therefore, G is a (4k, 1/2)-projection-cost preserving sketch of A. For convenience, we use PG to denote a rank-k projection matrix which provides the best rank-k approximation Gk = PG G to matrix G. We also use PA to denote a rank-k projection 45 matrix which provides the best rank-k approximation Ak = PA A to matrix A. Thus, 1 kA − CC† Ak2F ≤ kG − CC† Gk2F + c 2 ≤ 5kG − PG Gk2F + c ≤ 5kG − PA Gk2F + c
≤ 5 kG − PA Gk2F + c 15 kA − PA Ak2F ≤ 2 The first inequality is due to Lemma 64. G is a (4k, 1/2)-projection cost preserving sketch of A and C has rank at most 4k. The second inequality is based on Lemma 58. The third inequality is because PG provides the best rank-k approximation to G. The fourth inequality used the fact c is non-negative. The fifth inequality held because G is a (4k, 1/2)-projection cost preserving sketch of A. Thus, kA − CC† Ak2F ≤ 15kA − Ak k2F Combining with equation 12, we have: h E kA − C̃C̃ † Ak2F i h ≤ E kA − (A)k2F ΠF C̃,k i 30 ≤ (1+ ε) 50 rank X(A) i=k+1 σi2 (A) = (1+O(ε))kA−Ak k2F (13) Concluding the proof. Consider about rank i rank i h h X(A) X(A) † 2 2 ≤ (A) + O(ε) σ σi2 (A). (A)k E kA − C̃C̃ Ak2F ≤ E kA − ΠF i F C̃,k i=k+1 (14) i=k+1 Prank(A) The expectation is taken w.r.t. the randomness in constructing C̃; hence, the term i=k+1 σi2 (A) is a constant w.r.t. the expectation operation, which means that Eqn. 14 implies the bound:   rank rank X(A) X(A) F 2 2   E kA − ΠC̃,k (A)kF − σi (A) ≤ O(ε) · σi2 (A). i=k+1 (A)k2F − Let Y := kA − ΠF C̃,k Prank(A) i=k+1 i=k+1 σi2 (A). Y is a random variable with Y ≥ 0, because (A)) ≤ k. Markov’s inequality on Y implies that with arbitrary constant probability, rank(ΠF C̃,k kA − ΠF (A)k2F C̃,k kA − (A)k2F ΠF C̃,k − rank X(A) σi2 (A) ≤ rank X(A) σi2 (A) + i=k+1 ≤ O(ε) · rank X(A) σi2 (A), rank X(A) σi2 (A). i=k+1 equivalently, † i=k+1 O(ε) · (A)k2F concludes the proof. Using kA − C̃C̃ Ak2F ≤ kA − ΠF C̃,k 46 i=k+1 8.4.2 Proof of Eqn. 11 We would like to apply Lemma 62 for the matrix U ∈ Rm×k in the algorithm. Note that we have already proved that the matrix C̃ in the algorithm satisfies with probability at least 0.99: (A)k2F ≤ (1 + O(ε))kA − Ak k2F . Lemma 62 and a simple union bound conclude the kA − ΠF C̃,k proof. 8.4.3 Running time Next, we give an analysis of the arithmetic operations required in various steps in the algorithm. 1. Local Column Sampling: O(nnz(A) log(ms) + m · poly(k, ε, log(s))) arithmetic operations in total. (a) O(nnz(Ai ) log(ms) + m · poly(k, ε, log(s))) for each Si (b) - 2. Global Column Sampling: O(sk 3 + skm) arithmetic operations in total. (a) O(sk 3 + skm) to construct C. (b) 3. Adaptive Sampling O(sk 2 m + mnk + k/ε) arithmetic operations in total. (a) O(k2 m) for each C† locally; O(wi mk) for each Ψi and O(mwi ) for each ψi . (b) O(s) in total. (c) O(k/ε) in total using the method of [51] locally. (d) O(k/ε). 4. Rank-k matrix in span(C̃): O(mnk/ε3 + smk 2 /ε4 + sk 3 /ε5 ) arithmetic operations in total. (a) O(smk 2 /ε2 ) in total T (b) O(mwi k/ε3 ) for each Ai WiT and another O(mk2 ε−4 ) for each C̃ (Ai WiT ) (c) O(sk 2 ε−3 ) to find Ξ in the server; then another O(sk 3 ε−3 + sk 3 ε−5 )) to update Ξ locally in each machine and another O(sk 3 ε−5 ) to find ∆ locally in each machine; and O(smk 2 /ε) to find U locally in each machine. 8.4.4 Communication Complexity Assume that we can represent each element in the input matrix A with at most b bits. Also, we assume that one word has length b bits and b = O(log(mns/ε)). 1. Local Column Sampling: O(skφ) words. (a) (b) O(skφ) elements of A. 2. Global Column Sampling:O(skφ) words. (a) 47 (b) O(skφ) elements of A. 3. Adaptive Sampling: O(s + φk/ε) words. (a) s integers each of which is representable with O(log k+log log(mns/ε)) (from Lemma 61). (b) s integers (the ti ’s) each with magnitude at most n, hence representable with b bits. (c) O(φk/ε) elements of A. (d) 4. Best rank-k matrix in the span of C̃: O(sφk/ε + sk 2 ε−4 ) words. (a) O(skφε−1 ) elements of A. (b) O(sk 2 ε−4 ) numbers each of which can be represented with b bits. (c) O(sk 2 ε−4 ) numbers each of which can be represented with b bits. In total the communication complexity is O(skφ/ε + sk 2 /ε4 ) words. 9 Faster Distributed PCA for sparse matrices We now explain how to modify certain steps of the distributed PCA algorithm in Section 8 in order to improve the total running time spent to compute the matrix U; this, at the cost of increasing the communication cost slightly. Specifically, we replace the parts 2-(a), 3-(a), and 4-(b) with similar procedures, which are almost as accurate as the original procedures but run in time proportional to the number of the non-zero elements of the underlying matrices. Before presenting the new algorithm in detail, we discuss results from previous literature that we employ in the analysis. We remark that in the analysis below we have not attempted to optimize the constants. 9.1 Background material 9.1.1 Constant probability sparse subspace embeddings and sparse SVD First, we present the so-called sparse subspace embedding matrices of Clarskon and Woodruff [23]. Definition 65. [Sparse Subspace Embedding [23]] We call W ∈ Rξ×n a sparse subspace embedding of dimension ξ < n if it is constructed as follows, W = Ψ · Y, where • h : [n] → [ξ] is a random map so that for each i ∈ [n], h(i) = ξ ′ , for ξ ′ ∈ [ξ] w.p. 1/ξ. • Ψ ∈ Rξ×n is a binary matrix with Ψh(i),i = 1, and all remaining entries 0. • Y ∈ Rn×n is a random diagonal matrix, with each diagonal entry independently chosen to be +1 or −1, with equal probability. Such sparse subspace embedding matrices have favorable properties which we summarize below: Lemma 66 ([23]). Let AT ∈ Rn×m have rank ρ and let W ∈ Rξ×n be a randomly chosen sparse subspace embedding with dimension ξ = Ω(ρ2 ε−2 ), for some 0 < ε < 1. Then, 1) computing AW T requires O(nnz(A)) time; 2) nnz(AW T ) ≤ nnz(A); and 3) with probability at least 0.99, and for all vectors y ∈ Rm simultaneously, (1 − ε)kA T yk22 ≤ kWA T yk22 ≤ (1 + ε)kA T yk22 . 48 Notice that the third property in the lemma fails with constant probability. Based on the above sparse subspace embeddings, Clarkson and Woodruff [23] described a low-rank matrix approximation algorithm for A that runs in time nnz(A) plus low-order terms. Lemma 67 (Theorem 47 in [23]; for the exact statement see Lemma 3.4 in [19]). Given A ∈ Rm×n of rank ρ, a target rank 1 ≤ k < ρ, and 0 < ǫ ≤ 1, there exists a randomized algorithm that computes Z ∈ Rn×k with ZT Z = Ik and with probability at least 0.99, kA − AZZT k2F ≤ (1 + ǫ) kA − Ak k2F .
The proposed algorithm requires O nnz(A) + (m + n) · poly(k, ε−1 ) arithmetic operations. We denote this procedure as Z = SparseSV D(A, k, ε). Notice that the algorithm in the lemma fails with constant probability. 9.1.2 High probability sparse SVD Lemma 67 above presents a randomized algorithm to compute quickly an SVD of a sparse matrix; however, this lemma fails with constant probability. Since in our fast distributed PCA algorithm we will use the lemma multiple times (specifically, s times, since we need such a fast SVD locally in each machine), the overall failure probability would be of the order O(s) (by a union bound). To boost this failure probability, we develop a simple scheme to obtain a high probability result. Before we proceed, we quickly review the Johnson Lindestrauss transform, which we need in the analysis. Lemma 68. [Theorem 1 in [5] for fixed ε = 1/2] Let B ∈ Rm×n . Given β > 0, let r = (1/2)4+2β 2 −(1/2)3 log n. √ r×m Construct a matrix S ∈ R , each element of which is a random variable which takes values ±1/ r with equal probability. Let B̃ = SB. Then, if bi and b̃i denote the ith column of B and B̃, respectively, with probability at least 1− n−β , and for all i = 1, ...n, (1− 21 )kbi k22 ≤ kb̃i k22 ≤ (1+ 12 )kbi k22 . Given B, it takes O(nnz(B) log n) arithmetic operations to construct B̃. We will denote this procedure as B̃ = JLT (B, β). Now, we are fully equipped to design the “high-probability” analog of Lemma 67: Lemma 69. Given A ∈ Rm×n of rank ρ, a target rank 1 ≤ k < ρ, and 0 < δ, ǫ ≤ 1, there exists a randomized algorithm that computes Z ∈ Rn×k with ZT Z = Ik and with probability at least 1 − δ
− 1/n, kA − AZZT k2F ≤ 3 (1 + ǫ) kA − Ak k2F . The proposed algorithm requires O nnz(A) log2 ( nδ ) + n · poly(k, ε, log( nδ )) arithmetic operations. We denote this procedure as Z = SparseSV DBoosting(A, k, ε, δ). Proof. The idea is to use the algorithm in Lemma 67 and generate i.i.d. r = O(log( 1δ )) matrices Z1 , Z2 , . . . , Zr , where Zi = SparseSV D(A, k, ε), and kA − AZi ZiT k2F ≤ (1 + ε)kA − Ak k2F , with probability at least 0.99 for a single Zi . Now, we could have chosen Z := Zi which minimizes kA − AZi ZiT k2F , and this would have implied that kA − AZZT k2F ≤ (1 + ε)kA − Ak k2F , with probability 1 − δ (via a simple chernoff argument), but evaluating the errors kA − AZi ZiT k2F is computationally intensive for the overall running time that we would like to obtain. Selection of Z. Instead, we use the Johnson Lindestrauss transform to speedup this computation (evaluating the errors). Hence, for every Zi , we (implicitly) set Bi := A − AZi ZiT and then we compute B̃i = JLT (Bi , 1) (see Lemma 68). Now, we choose the Zi that minimizes kB̃i k2F . 49 Correctness. For a fixed Zi , from Lemma 68 it follows that with probability 1 − 1/n it is: 21 kA − AZi ZiT k2F ≤ kB̃i k2F ≤ 23 kA−AZi ZiT k2F , where the failure probability is over the randomness of S in Lemma 68. Now, for some Zi with probability at least 0.99 it is kA− AZi ZiT k2F ≤ (1+ ε)kA − Ak k2F , where the failure probability is over the randomness in constructing Zi . Overall, given that n > 1, with probability at least 0.49 (over the randomness of both S and Zi ; the failure probability follows via a union bound) for each single Zi it is: kB̃i k2F ≤ 23 (1 + ε)kA − Ak k2F . Hence, since r = O(log 1δ ) and we have picked Z as the Zi that minimizes kB̃i k2F , it follows that with probability at least 1− δ, kB̃i k2F ≤ 23 (1+ε)kA −Ak k2F . Combining with 21 kA−AZi ZiT k2F ≤ kB̃i k2F , and using a union bound, it follows that with probability 1 − 1/n − δ: kA − AZi ZiT k2F ≤ 62 (1 + ε)kA − Ak k2F .
Running time. The following costs occur O(log( 1δ )) times: (i) O nnz(A) + (m + n) · poly(k, ε−1 ) to find Zi (via Lemma 67); (ii) O(nnz(A) log n + nk log(n)) to evaluate each cost, i.e., find kB̃i k2F . 9.1.3 Fast column sampling techniques Section 8.1.1 presented the column sampling algorithms that we used in our distributed PCA algorithm in Section 8. Next, we present the “fast” analogs of those algorithms. To design such fast analogs we employ the sparse subspace embedding matrices in the previous section. All the results in this section developed in [19]. There is a small difference in Lemma 70 below, hence we present a short proof for completeness. Specifically, the original result in [19] has a constant failure probability; here, via using standard arguments, we extended this result to a high probability bound. Lemma 70 (Input-Sparsity-Time Dual-Set Spectral-Frobenius Sparsification [19].). Let V ∈ Rw×k be a matrix with w > k and V T V = Ik . Let E ∈ Rm×w be an arbitrary matrix. Let δ be a failure probability parameter. For i = 1, 2, . . . , r = O(log 1δ ), let Bi = EWiT ∈ Rm×ξ , where Wi ∈ Rξ×w is a randomly chosen sparse subspace embedding with ξ = O(k2 /ε2 ) < w, for some 0 < ε < 1 (see Definition 65 and Lemma 66), and run the algorithm of Lemma 57 with V, Bi , and some ℓ with k < ℓ ≤ w, to construct Si ∈ Rw×ℓ . Choose some S from the Si ’s - see the proof for the details. 2   p 2
1+ε · kEk2F . Then, with probability at least 1 − δ, σk2 V T S ≥ 1 − k/ℓ and kESk2F ≤ 1−ε


The matrix S can be computed in O log 1δ · nnz(A) + ℓwk2 + mk 2 /ε2 + mℓ time. We denote this procedure as S = BssSamplingSparse(V, A, r, ε, δ). Proof. The algorithm constructs the Si ’s as follows, Si = BssSampling(V, Bi , r) - see Lemma 57.  p 2
From that lemma, we also have that σk2 V T Si ≥ 1 − k/ℓ and kBiT Sk2F ≤ kBiT k2F , i.e., kWi AT Si k2F ≤ kWi AT k2F . Since Wi is a subspace embedding, from Lemma 66 we have that with probability at least 0.99 and for all vectors y ∈ Rn simultaneously, (1 − ε) kAT yk22 ≤ kWi AT yk22 . Apply this r times for y ∈ Rn being columns from Si ∈ Rn×r and take a sum on the resulting inequalities: (1 − ε) kAT Si k2F ≤ kWAT Si k2F ; also, apply this n times for the basis vectors in Rn and take a sum on the resulting inequalities: kWA T k2F ≤ (1 + ε) kA T k2F . Combining all these · kA T k2 . inequalities together, we conclude that with probability at least 0.99, kAT Si k2F ≤ 1+ε 1−ε p F2
To summarize, for each Si and with probability at least 0.99, σk2 V T Si ≥ 1 − k/ℓ and 2  · kEk2F . kESi k2F ≤ 1+ε 1−ε 50 Selection of S. Now that we have r matrices Si each of which satisfying the
above bounds with probability 0.99 we select one of them as follows: 1) we compute all σk2 V T Si and sort them from the largest to the smallest value; 2) we compute all kAT Si k2F and sort them from the smallest to the largest value; 3) we choose an Si which occurs in the top 2/3 fraction of each of the lists.
Correctness. Let Xi be an indicator random variable: Xi = 1 if the ith matrix Si satisfies σk2 V T Si ≥  p 2 P 1 − k/ℓ , and Xi = 0, otherwise. Then, Pr[Xi ] ≥ 0.99. Let X = ri Xi . The expected value of X is µX = E [X] = 0.99r. A standard Chernoff bound applied on X gives: Pr[X ≤ (1 − α)µX ] ≤ µX α e− 2 . Let (1 − α)0.99r = 0.75r, i.e., α = 0.14/0.99. Applying this to the Chernoff bound, we get: Pr[X ≤ 0.75r] ≤ e−O(r) . Replacing r = O(log(1/δ)): Pr[X ≤ 0.75r] ≤ δ/2. This means precisely p 2
 that 3/4 of the fraction of the Si ’s satisfy with probability 1 − δ/2 : σk2 V T Si ≥ 1 − k/ℓ . Similarly, let Yi be an indicator random variable: Yi = 1 if the ith matrix Si satisfies kESi k2F ≤ 2 P 1+ε · kEk2F , and Yi = 0, otherwise. Then, Pr[Yi ] ≥ 0.99. Let Y = ri Xi . The expected value of 1−ε Y is µY = E [Y ] = 0.99r. A standard Chernoff bound applied on Y gives: Pr[Y ≤ (1 − α)µY ] ≤ µY α e− 2 . Let (1 − α)0.99r = 0.75r, i.e., α = 0.14/0.99. Applying this to the Chernoff bound, we get: Pr[Y ≤ 0.75r] ≤ e−O(r) . Replacing r = O(log(1/δ)): Pr[Y ≤ 0.75r] ≤ δ/2. This means precisely 2  1+ε · kEk2F . that 3/4 of the fraction of the Si ’s satisfy w.p. 1 − δ/2 : kESi k2F ≤ 1−ε Now a simple union bound indicates that with probability at least 1 − δ, the following two events happen simultaneously: X > 0.75r and Y > 0.75r. Since we select an element Si on top of the two sorted lists, it follows that this element should have Xi = Yi = 1.  Running time. The following costs occur O(log( 1δ )) times: (i) O(nnz(A)) to find Bi . (i) O(ℓwk2 +
mξ) to find Si (via Lemma 57); and (iii) O(k 2 ℓ + mℓ) to evaluate each cost and, i.e., find σk2 V T Si and kESi k2F . Additionally, O(log( 1δ ) log log( 1δ )) time in total is spent in sorting. The following Lemma is the “fast” analog of Lemma 58. The failure probability in the lemma is constant, but this is sufficient for our purposes since we apply this lemma only once in the global column sampling step of the distributed PCA algorithm in Section 9.2. Lemma 71 (Input-Sparsity-Time constant factor column-based matrix reconstruction; Lemma 6.3 2 in [19]). Given matrix G ∈ Rm×α of rank ρ and a target
rank k , there exists a randomized algorithm −1 that runs in O nnz(A) · log α + m · poly(log α, k, ε ) time and selects c = 4k columns of G to form a matrix C ∈ Rm×c such that with probability at least 0.69: kG − CC† Gk2F ≤ 4820 · rank X(G) σi2 (G). i=k+1 We denote this procedure as C = DeterministicCssF robeniusSparse(G, k, c). Finally, the lemma below presents the “fast” analog of Lemma 59. The failure probability in the lemma is, again, constant; we have not attempted to obtain a high probability bound since we employ this lemma only once in the algorithm in Section 9.2. 2 The original Lemma 6.3 in [19] has the assumption that k < ρ, but this assumption can be dropped having the result unchanged. The only reason the assumption k < ρ exists is because otherwise column subset selection is trivial. 51 Lemma 72 (Input-sparsity-time Adaptive Sampling [19]). Given A ∈ Rm×n and V ∈ Rm×c1 (with c1 ≤ n, m), define the residual matrix Ψ = A − VV† A ∈ Rm×n . Let Ψ̃ = JLT (Ψ, 1). For j = 1, . . . , n, let pj be a probability distribution such that, for some positive constant β ≤ 1, pj ≥ βkΨ̃ (j) (j) 2 k2 /kΨ̃k2F , where Ψ̃ is the j-th column of the matrix Ψ̃. Sample c2 columns from A in c2 i.i.d. trials, where in each trial the j-th column is chosen with probability pi . Let C2 ∈ Rm×c2 contain the c2 sampled columns and let C = [V C2 ] ∈ Rm×(c1 +c2 ) contain the columns of V and C2 . Then, for any integer k > 0, and with probability 0.9 − n1 2 2 kA − ΠF C,k (A)kF ≤ kA − Ak kF + 30 · k kA − VV† Ak2F . β · c2 Given A and V, the algorithm takes O(nnz(A) log n + mc1 log n + mc21 ) time to find C2 . We denote this sampling procedure as C2 = AdaptiveColsSparse(A, V, c2 , β). 9.1.4 Fast Low-rank matrix approximations within a subspace Finally, we present the fast analog of the result in Section 8.1.3. The failure probability in the lemma is, again, constant; we have not attempted to obtain a high probability bound since we employ this lemma only once in the algorithm in Section 9.2. Lemma 73. Let A ∈ Rm×n be the input matrix and V ∈ Rm×c be the input subspace. We further assume that for some rank parameter k < c and accuracy parameter 0 < ε < 1 : 2 2 kA − ΠF V,k (A)kF ≤ (1 + O(ε))kA − Ak kF . Let V = YΨ be a qr decomposition of V with Y ∈ Rm×c and Ψ ∈ Rc×c . Let Ξ = Y T AW T ∈ Rc×ξ , where W T ∈ Rn×ξ with ξ = O(c/ε2 ), is a sparse subspace embedding matrix (see Definition 65 and Lemma 66). Let ∆ ∈ Rc×k contain the top k left singular vectors of Ξ. Then, with probability at least 0.99, kA − Y∆∆T Y T Ak2F ≤ (1 + ε)kA − Ak k2F . Y and ∆ can be computed in O(nnz(A) + mcξ) time. We denote this procedure as [Y, ∆] = ApproxSubspaceSV DSparse(A, V, k, ε). Proof. This result was proven inside the proof of Theorem 1.5 in [36]. Specifically, the error bound proven in [36] is for the transpose of A (also Y, ∆ are denoted with U, V in [36]). The only requirement for the embedding matrix W (denoted with P in the proof of Theorem 1.5 in [36]) is to be a subspace embedding for Y T A, in the sense that W is a subspace embedding for A in Lemma 66. Since our choice of W with ξ = O(c2 /ε2 ) satisfies this requirement we omit the details of the proof. The running time is O(nnz(A)+mcξ): one can compute (i) Y in O(mc2 ); (ii) Ξ in O(nnz(A)+mcξ); and (iii) ∆ in O(cξ min{c, ξ}). The failure probability 0.01 from Lemma 66. 9.2 Detailed description of the algorithm This algorithm is very similar to the algorithm in Section 8.2; we only replace the parts 2-(a), 3-(a), and 4-(b) with faster procedures. Input: 1. A ∈ Rm×n partitioned column-wise A = A1 A2 2. rank parameter k < rank(A) 52
P . . . As ; for i = 1 : s, Ai ∈ Rm×wi ; i wi = n. 3. accuracy parameter ε > 0 4. failure probability δ Algorithm 1. Local Column Sampling (a) For each sub-matrix Ai ∈ Rm×wi , compute Ci ∈ Rm×ℓ containing ℓ = O(k) columns from Ai as follows: Ci = Ai Si . Here, Si has dimensions wi × ℓ and is constructed as follows: Si = BssSamplingI(Ai , 4k, 1/2, 0.01/s) (see Lemma 56). (b) Machine i sends Ci to the server. 2. Global Column Sampling (a) Server constructs m × (s
· ℓ) matrix G containing (s · ℓ) actual columns from A as follows: G = C1 C2 . . . Cs . Then, server constructs C ∈ Rm×c1 via choosing c1 = 4k columns from G as follows: C = DeterministicCssF robeniusSparse(G, k, c1 ) (see Lemma 71). (b) Server sends C to all the machines. 3. Adaptive Column Sampling (a) Server initializes the random seed and communicates it to all the machines. Each machine constructs the same S ∈ Rr×m , with r = (1/2)4+2 2 −(1/2)3 log n, each element of which is a random vari√ able which takes values ±1/ r with equal probability. Machine i finds Ψ̃i = SAi − SCC† Ai ∈ Rm×wi and then computes βi as it was described in Lemma 61. Machine i sends βi to server. (b) Server computes probability distribution gi = Pβiβi . Server samples i.i.d. with replacement i ⌈ 50k/ε ⌉ samples (machines) from gi . Then, server determines numbers ti (i = 1, 2, . . . , s), where ti is the number of times the ith machine was sampled. It sends the ti ’s to the machines. (c) Machine i computes probabilities qji = kxk22 /kΨ̃i k2F (j = 1 : wi ), where x is the jth column of Ψ̃i . And now machine i samples ti samples fromP it’s local probability distribution and sends the corresponding columns to the server. Let c2 = i ti = ⌈ 50k/ε ⌉. (d) Server collects the columns and assigns them to Ĉ ∈ Rm×c2 . Server constructs C̃ to be the
m × (c1 + c2 ) matrix: C̃ = C; Ĉ . Let c = c1 + c2 = 4k + ⌈ 50k/ε ⌉. 4. Rank-k matrix in the span of C̃ (a) Server sends C̃ to all the machines and each machine computes (the same) qr factorization of C̃: C̃ = YR where Y ∈ Rm×c has orthonormal columns and R ∈ Rc×c is upper triangular. (b) Server initializes the random seed and sends this to each machine, such that each machine generates the same matrix Ψ ∈ Rξ×n , for ξ = O(c2 /ε2 ), (see
Definition 65). Machine i generates Di ∈ Rn×wi such that implicitly D = D1 D2 . . . Ds is an n × n matrix each element of ξ×wi . which is ±1, with probability 1/2. Each machine constructs implicitly i ∈ R
Wi = Ψ · Dξ×n and W Implicitly all machines together generate W = W1 W2 . . . Ws , with W ∈ R T is a subspace embedding as in Definition 65. Machine i computes Hi = C̃ (Ai WiT ) ∈ Rc×ξ . Machine i sends Hi to the server. Ps (c) Server computes Ξ = i=1 Hi ∈ Rc×ξ and sends this back to all the machines. Now machines compute Ξ := R−1 · Ξ(:= Y T Ai WiT ), and then they compute ∆ ∈ Rc×k to be the top k left singular vectors of Ξ. Each machine computes U = Y · ∆ ∈ Rm×k . 53 Discussion. A few remarks are necessary for the last two stages of the algorithm. The third stage (adaptive column sampling), implements the adaptive sampling method of Lemma 72. To see this note that each column in A is sampled with probability qji · gi ≥ 1 · ||x||22 /kΨ̃k2F , 2 where x is this column in Ψ̃. This follows from Lemma 60. Overall, this stage constructs Ĉ such that Ĉ = AdaptiveColsSparse(A, C, c2 , 1/2). The last stage in the algorithm implements the algorithm in Lemma 73. To see this, note that W satisfies the properties in the lemma. Hence, [Y, ∆] = ApproxSubspaceSV DSparse(A, C̃, k, ε). 9.3 Main result The theorem below analyzes the approximation error, the communication complexity, and the running time of the previous algorithm. Theorem 74. The matrix C̃ ∈ Rm×c with c = O(k/ε) columns of A satisfies w.p. 0.59 − s+1 n   rank X(A) (A)k2F ≤ (1 + O(ε)) ·  kA − C̃C̃ Ak2F ≤ kA − ΠF σi2 (A) . C̃,k † − 2δ: (15) i=k+1 The matrix U ∈ Rm×k with k orthonormal columns satisfies w.p. 0.58 −  kA − UUT Ak2F ≤ (1 + O(ε)) ·  rank X(A) i=k+1 s+1 n − 2δ:  σi2 (A) . (16) Let each column of A
has at most φ non-zero elements. Then, the communication cost of the algorithm is O skφε−1 + sk3 ε−5 and the running time is  ns    ns  O nnz(A) · log2 + (m + n)s · poly(k, ε−1 , log . δ δ 9.4 Proof of Theorem 74 9.4.1 Proof of Eqn. 15 From Lemma 72, we have that with probability at least 0.9 − n1 : † (A)k2F ≤ kA − C̃C̃ Ak2F ≤ kA − ΠF C̃,k rank X(A) i=k+1 σi2 (A) + 60ε · kA − CC† Ak2F . 50 (17) According to Lemma 64, G is a (4k, 1/2)-projection-cost preserving sketch of A. It means that there exists a constant c which is independent from projection matrix P and we have the following: 3 1 kA − PAk2F ≤ kG − PGk2F + c ≤ kA − PAk2F 2 2 54 For convenience, we use PG to denote a rank-k projection matrix which provides the best rank-k approximation Gk = PG G to matrix G. We also use PA to denote a rank-k projection matrix which provides the best rank-k approximation Ak = PA A to matrix A. With probability at least 0.68, the following will be held, 1 kA − CC† Ak2F ≤ kG − CC† Gk2F + c 2 ≤ 4820kG − PG Gk2F + c ≤ 4820kG − PA Gk2F + c
≤ 4820 kG − PA Gk2F + c 4820 ≤ kA − PA Ak2F 2 The first inequality is due to Lemma 64. G is a (4k, 1/2)-projection cost preserving sketch of A and C has rank at most 4k. The second inequality is based on Lemma 71. The third inequality is because PG provides the best rank-k approximation to G. The fourth inequality used the fact c is non-negative. The fifth inequality held because G is a (4k, 1/2)-projection cost preserving sketch of A. Thus, kA − CC† Ak2F ≤ 4820kA − Ak k2F Concluding the proof. Combining with Equation 17, we have: † kA − C̃C̃ Ak2F ≤ kA − ΠF (A)k2F ≤ C̃,k 9.4.2 rank X(A) i=k+1 σi2 (A) + O(ε) · rank X(A) σi2 (A). (18) i=k+1 Proof of Eqn. 16 We would like to apply Lemma 73 for the matrix U ∈ Rm×k in the algorithm. Note that we already proved that the matrix C̃ in the algorithm satisfies with probability 0.59 − s+1 n − 2δ: kA − ΠF (A)k2F ≤ (1 + O(ε))kA − Ak k2F . C̃,k Lemma 73 and a simple union bound conclude the proof. 9.4.3 Running time 1. Local Column Sampling: O(nnz(A) log(ms) + m · poly(k, ε, log(s))) arithmetic operations in total. (a) O(nnz(Ai ) log(ms) + m · poly(k, ε, log(s))) for each Si (b) -
2. Global Column Sampling: O nnz(A) · log k + m · poly(k, ε−1 ) .
(a) O nnz(A) · log k + m · poly(k, ε−1 ) to construct C - from Lemma 71. (b) - 55 3. Adaptive Sampling: O(nnz(A) · log n + m · s · poly(k, ε−1 , log n)). (a) First, we analyze the cost to compute the matrix Ψ̃i . O(k2 m) for each C† locally; then, we compute (SAi ) − (((SCi )C†i )Ai ). The costs are: SAi := D takes O(nnz(Ai )) time, SCi := G takes O(nnz(Ai ) log(n)) time, GC†i := H takes O(m log(n)k/ε2 ) time, HA := L takes O(nnz(Ai ) log(n)) time, and D − L takes O(n log n) time. So, the total time to compute one Ψ̃i is O(nnz(Ai ) log n + m · poly(k, ε−1 , log n)). Also, there is a cost of computing ψi , which is O(log n · wi ) arithmetic operations. (b) O(s) in total. (c) O(k/ε) in total using the method of [51] locally. (d) O(k/ε). 4. Rank-k matrix in span(C̃): O(nnz(A) + m · s · poly(k, ε−1 )) arithmetic operations in total. (a) O(smk 2 /ε2 ) in total. (b) O(sn) in total to generate s times the same Ψ. Then another O(n) in total to generate the Di ’s. For all Ai WiT we need O(nnz(A)) operations and then another O(mk 3 ε−7 ) for each Y T · (Ai WiT ) (c) O(sk 3 ε−5 ) to find Ξ; then another O(sk3 ε−3 + sk4 ε−6 )) to update Ξ locally in each machine and another O(sk 4 ε−6 ) to find ∆; and O(smk 2 /ε) to find U (s times). 9.4.4 Communication Complexity Assume that we can represent each element in the input matrix A with at most b bits. Also, we assume that one word has length b bits and b = O(log(mns/ε)). 1. Local Column Sampling: O(skφ) words. (a) (b) O(skφ) elements of A. 2. Global Column Sampling:O(skφ) words. (a) (b) O(skφ) elements of A. 3. Adaptive Sampling: O(s + φk/ε) words. (a) s integers each of which is representable with O(log k+log log(mns/ε)) (from Lemma 61). (b) s integers (the ti ’s) each with magnitude at most n, hence representable with b bits. (c) O(φk/ε) elements of A. (d) 4. Best rank-k matrix in the span of C̃: O(sφk/ε + sk 3 ε−5 ) words. (a) O(skφε−1 ) elements of A. (b) O(sk 3 ε−5 ) numbers each of which can be represented with b bits. (c) O(sk 3 ε−5 ) numbers each of which can be represented with b bits. In total the communication complexity is O(skφ/ε + sk 3 /ε5 ) words. 56 10 Lower bounds Only in Section 10.1, we discuss the result in arbitrary partition model. In Section 10.2, Section 10.3 and Section 10.4, we discuss results in column partition model which provides stronger lower bounds than in arbitrary partition model. 10.1 Dependence on Ω(ε−2 ) bits for distributed PCA in arbitrary partition model Lemma 75 (Theorem 8 in [54]). There are two machines, and each of them hold a matrix Ai ∈ Rm×n . Let A = A1 + A2 . Suppose ε2 ≤ 2n, the communication of computing a rank-k matrix U with constant probability such that kA − UUT Ak2F ≤ (1 + ε) · kA − Ak k2F needs Ω(1/ε2 ) bits. Proof. (from [54]). The reduction is from GHD problem. Lemma 76 (Theorem 1.1 in [20]). Each of two machines has an n-bit vector. There is a promise: either √ √ the inputs are at Hamming distance less than n/2 − c n or greater than n/2 + c n. If they want to distinguish these two cases with constant probability, the communication required Ω(n) bits. Without loss of generality, we assume 1/ε2 is an integer, and machine 1 and machine 2 have 2 x, y ∈ {−1, 1}1/ε respectively. There is a promise that either xT y < −2/ε or xT y > 2/ε holds. Consider about the following protocol: 1. Machine 1 constructs A1 ∈ R(k+1)×(1/ε  xT ε √0  0 2   A1 =  0  0  ... ...  0 0 0 √ 0 2 +k) , and machine 2 constructs A2 ∈ R(k+1)×(1/ε ... ... 0 0 ... ... ... 0 ... 0 ... √ 2(1+ε) ε  2(1+ε) ε           A2 =    y T ε 0 0 ... 0 0 0 0 ... 0 0 0 0 0 0 ... ... ... ... ... 0 0 0 ... 0 2 +k) :       2. The server computes U such that kA − UUT Ak2F ≤ (1 + ε) · kA − Ak k2F and sends U to both of two machines, where A = A1 + A2 . 3. Let P = Ik+1 − UUT . Each machine construct v = P(1) /kP(1) k2 4. Each machine checks whether v12 < 21 (1 + ε). If yes, return the case xT y > 2/ε. Otherwise return the case xT y < 2/ε Suppose the server successfully computes U. Since the rank of A is at most k + 1: kA − UUT Ak2F = kPAk2F = kvAk22 = vAAT v T 57 An observation is that     AAT =    kx + yk22 ε2 0 0 2 0 0 ... ... 0 0 0 0 2(1+ε) ε2 ... 0 ... ... ... ... ... 0 0 0 ... 2(1+ε) ε2 We have kA − Ak k2F ≤ 2. Then, k+1 2(1 + ε) X 2 vi = ε2 i=3        2(1 + ε) (1 − v12 − v22 ) ε2 < 2(1 + ε) We have v12 + v22 > 1 − ε2 . When xT y > 2/ε, kx + yk22 ε2 ≥ 2 + 4ε (1 + ε)kA − Ak k2F = 2(1 + ε) > v12 (2 + 4ε) + 2v22 > 2 − 2ε2 + 4εv12 So, v12 < 12 (1 + ε). When xT y < −2/ε, kx + yk22 ε2 ≤ 2 − 4ε (1 + ε)kA − Ak k2F = kx + yk22 ε2 (1 + ε) > v12 kx + yk22 ε2 + 2v22 = 2(v12 + v22 ) − (2 − kx + yk22 ε2 )v12 > 2(1 − ε2 ) − (2 − kx + yk22 ε2 )v12 So, v12 > 1 2(1 − ε2 ) − kx + yk22 ε2 (1 + ε) ≥ (1 + ε) 2 2 2 2 − kx + yk2 ε Therefore, machines can distinguish these two cases. The only communication is computing U. This cost shoulde be the same as Ω(1/ε2 ) bits lower bound of the gap hamming distance problem. 10.2 Lower bounds for distributed PCA on dense matrices This section provides a communication cost lower bound for the Distributed PCA problem of Definition 17. Specifically, we describe the construction of an m × n matrix A (hard instance), and formally argue that for this A, any k ≤ 0.99m, and any error parameter C with 1 < C < poly(skm), if there exists some algorithm to construct an m × k matrix U such that, with constant probability, kA − UUT Ak2F ≤ C · kA − Ak k2F , then this algorithm has communication cost Ω(skm) words. 10.2.1 Preliminaries We use the notation Gk,m to denote the set of k-dimensional subspaces of Rm , which we identify with corresponding projector matrices Q ∈ Rm×m , with rank(Q) = k ≤ m, onto the subspaces, i.e., Gk,m = {Q ∈ Rm×m : Q2 = Q, rank(Q) = k ≤ m.} 58 We also need a description of a subset from Gk,m: Cδk (P) = {Q ∈ Gk,m : kP − Qk2 ≤ δ}. This is the set of those projectors Q ∈ Gk,m which differ in operator norm by at most δ from another fixed projector P ∈ Gk,m . In our analysis below we also need a result from [38]: Theorem 77. (Net Bound - Corollary 5.1 of [38]) For any m, k, δ > 0, there is a family N = {P1 , . . . , PN }, N = 2Ω(k(m−k) log(1/δ)) , where Pi ∈ Gk,m and Cδk (Pi ) ∩ Cδk (Pj ) = ∅, for all i 6= j. Theorem 77 proves the existence of a large, but finite, set N of projection matrices, such that if one considers the ball of matrices of operator norm at most δ centered at each matrix in N , then these balls are disjoint. Our hard instance matrix A is constructed from some member in N . 10.2.2 Hard instance construction Recall that in the column-partition model of Definition 15 there are s servers holding matrices A1 , . . . , As , respectively, P where Ai has m rows and some subset of wi 3columns of some m × n matrix A. Notice that wi = n. First of all, for arbitrary m, s we assume that n ≥ sm. Below, we describe a specific construction for a matrix A. First of all, we set δ = 1/(skm), the parameter to be used in Theorem 77. Next, fix a family N of subspaces of Gk,m with the guarantees of Theorem 77. Let PR = RRT be a uniformly random member of N , where R ∈ Rm×k has orthonormal columns (any projector can be expressed as such a matrix, where the columns of R span the subspace that P projects onto). The entries of A1 are equal to the entries of R, each rounded to the nearest integer multiple of 1/B, where B = poly(skm) is a large enough parameter specified later. We denote this rounded matrix with k R̃ ∈ Rm×k . So, if αij is the (i, j)th entry in A1 , then αij = kmin B , with kmin = argmin |Rij − B |. Each k∈Z Ai for i = 2, 3, . . . , s − 1 is 1/B times the m × m identity matrix. Finally, As is the m × t matrix of all zeros with t = n − (s − 1)m − k. I.e.,
A = R̃ B1 Im B1 Im . . . B1 Im 0m×t . 10.2.3 Intermediate results First, we give a bound regarding the best rank k approximation for the above matrix A. Lemma 78. For the matrix A ∈ Rm×n in Section 10.2.2, and any k ≤ 0.99m : kA − Ak k2F < sm . B2 Proof. We have, kA − Ak k2F ≤ kA − A1 A†1 Ak2F X = k(I − A1 A†1 )Ai k2F i>1 ≤ X i>1 kAi k2F m = (s − 2) · 2 B sm < , B2 As, otherwise we can choose a value s′ < s so that n − m ≤ s′ m ≤ n and apply the argument of Theorem 82 with s replaced with s′ . Then, the lower bound in Theorem 82 will then be Ω(s′ mk), which assuming n ≥ 2d, is an Ω(n) communication lower bound 3 59 where the first inequality uses the fact that A1 A†1 is a matrix of rank at most k, the first equality uses the fact that (I − A1 A†1 )Ai is the all-zeros matrix, the second inequality uses that a projector cannot increase a unitarily invariant norm, and the third inequality follows by construction. Next, we bound, in the operator norm, the difference of A1 from the matrix R. The lemma follows from the fact that A1 ∈ Rm×k is obtained by R ∈ Rm×k by rounding each entry to the nearest integer multiple of 1/B, and then summing the squared differences across all km entries. Lemma 79. (Precision Lemma) kA1 − Rk2 ≤ Proof. kA1 − Rk22 ≤ kA1 − Rk2F ≤ √ km B . km . B2 Next, we prove a pure linear algebraic result. The following lemma captures the fact that if some P ∈ Gk,m is close to QP (in Frobenius norm) for Q ∈ Gk,m , then P is also close to Q (in Frobenius norm). We will use this lemma later for P = PR . Lemma 80. (Error Measure) Let P, Q ∈ Gk,m with kP − QPk2F ≤ ∆. Then kP − Qk2F ≤ 2∆. Proof. By the matrix Pythagorean theorem, k = kQk2F = kQPk2F + kQ(I − P)k2F . (19) Also by the matrix Pythagorean theorem, k = kPk2F = kQPk2F + k(I − Q)Pk2F , and so using the premise of the lemma, kQPk2F ≥ k − ∆. Combining with (19), kQ(I − P)k2F ≤ ∆. (20) Hence, kP − Qk2F = k(P − Q)Pk2F + k(P − Q)(I − P)k2F = kP − QPk2F + kQ − QPk2F ≤ ∆ + kQ(I − P)k2F ≤ 2∆, where the first equality follows by the matrix Pythagorean theorem, the second equality uses P2 = P (since P is a projector), the third equality uses the bound in the premise of the lemma, and the fourth inequality uses (20). 10.2.4 Main Argument Before presenting the main theorem, we give an intermediate technical lemma. Lemma 81. (Implication of Correctness) Let A ∈ Rm×n be the hard instance matrix in Section 10.2.2 and let A be distributed in s machines in the way it was described in Section 10.2.2:
A = R̃ B1 Im B1 Im . . . B1 Im 0m×t . Suppose n = Ω(sm), k < rank(A), and 1 < C < poly(skm). A is revealed to the machines by just describing the entries in A and without specifying any other detail regarding the construction. Given this A and assuming further that each machine knows some projector matrix Q ∈ Rm×m with rank at most k such that kA − QAkF ≤ CkA − Ak kF , there is a deterministic algorithm (protocol) which, upon termination, leaves on each machine the matrix R ∈ Rm×k which was used to construct A1 . 60 1 Proof. First, we prove that kRRT − Qk2 < 2skm . Then, using this bound, we describe a protocol that deterministically reveals the matrix R which was used to construct A1 . , which in Using Lemma 78 and the premise in the lemma, it follows that: kA − QAk2F ≤ C sm B2 particular implies that kA1 − QA1 k2F ≤ C sm . B2 (21) We further manipulate the term kA1 − QA1 kF as follows: kA1 − QA1 kF = k(I − Q)A1 kF = k(I − Q)RRT + (I − Q)(A1 − RRT )kF ≥ k(I − Q)RRT kF − k(I − Q)(A1 − RRT )kF ≥ k(I − Q)RRT kF − kI − QkF k(A1 − RRT )k2 √ √ km √ T ≥ k(I − Q)RR kF − ( m + k), B where the first inequality is the triangle inequality, the second inequality uses sub-multiplicativity, and the third inequality uses the triangle inequality, i.e., that kI − QkF ≤ kIkF + kQk , and Lemma √ F √ T Csm 2m k 79. Combining this bound with (21), it follows that k(I − Q)RR kF ≤ B + B . At this point we would like to apply Lemma 80 to conclude that kRRT − QkF is small. To do so, we need RRT and Q to be of rank k. While R has rank k by construction, Q may not. However, increasing the rank of Q cannot increase the error kA−QAkF . Hence, if rank(Q) < k, given Q, the protocol in the premise of the lemma allows each server to locally add standard basis vectors to the column space of Q until the column space of Q becomes k. This involves no communication and each server ends up with the same new setting of Q, which for simplicity we still denote with Q. Therefore, rank(Q) = k. Applying Lemma 80, it now follows that √ ! √ √ Csm 2m k T T + . kRR − Qk2 ≤ kRR − QkF ≤ 2 B B 1 . By setting B to be a sufficiently large poly(skm), we have kRRT − Qk2 < 2skm i Therefore, given Q, by Theorem 77 there is a unique matrix P in N for which Q ∈ Cδk (Pi ), 1 1 with δ = skm . Indeed, since kR − Qk2 ≤ skm < δ/2, there cannot be two matrices Pi and Pj for i 6= j for which Q ∈ Cδk (Pi ) and Q ∈ Cδk (Pj ) since then by the triangle inequality kPi − Pj k2 < δ, contradicting the construction of N . Hence, it follows that each machine can deterministically identify this Pi , by enumeration over N , and therefore compute R, via, for example, an SVD. We are now fully equipped to present the main argument about the communication cost lower bound for distributed PCA. Theorem 82. (Main) Let A ∈ Rm×n be the hard instance matrix in Section 10.2.2 and let A be distributed in s machines in the way it was described in Section 10.2.2:
A = R̃ B1 Im B1 Im . . . B1 Im 0m×t . Suppose n = Ω(sm), k ≤ .99m, and 1 < C < poly(skm). Assume that there is an algorithm (protocol) which succeeds with probability at least 2/3 in having the i-th server output QAi , for all i, where Q ∈ Rm×m is a projector matrix with rank at most k such that kA − QAkF ≤ CkA − Ak kF . Then, this algorithm requires Ω(skm log(skm)) bits of communication. 61 Further, the bound holds even if the input matrices Ai to the servers have all entries which are integer multiples of 1/B, for a value B = poly(skm), and bounded in magnitude by 1, and therefore all entries can be specified with O(log(skm)) bits. Note that assuming a word size of Θ(log(skm)) bits, we obtain an Ω(skm) word lower bound. Proof. If the protocol succeeds, then, since each server outputs QAi , and Ai is 1/B times the identity matrix for all i = 2, 3, ...s − 1, each of those servers can compute Q. One of those servers can now send this Q to the first and the last server. It follows by Lemma 81, that each server can identify PR and R. Letting E be the event that the protocol succeeds, and letting Πi be the ordered sequence of all incoming and outgoing messages at the i-th server, it follows that I(Πi ; Q | E) = Ω(log |N |) = Ω(km log(smk)), where I(X; Y | F) = H(X | F)−H(X | Y, F) is the mutual information between random variables X and Y conditioned on F. To see the first equality in the above, observe that I(Πi ; Q | E) = H(Q | E) − H(Q | Πi , E), and conditioned on E, there is a uniquely determined matrix PR each server identifies, which is uniform over a set of size |N |, and so H(Q | E) ≥ H(PR ) = Ω(log2P |N |). Here, H(X) is the Shannon entropyP of a random variable X, given by H(X) = x Pr[X = x] log(1/ Pr[X = x]), and H(X | Y ) = y Pr[Y = y]H(X | Y = y) is the conditional Shannon entropy. Hence, if Z is an indicator random variable which is 1 if and only if E occurs, then using that I(X; Y ) ≥ I(X; Y | W ) − H(W ) for any random variables X, Y, and W , and that I(X; Y | W ) = P w Pr[W = w]I(X; Y | W = w), we have I(Πi ; Q) ≥ I(Πi ; Q | Z) − 1 ≥ I(Πi ; Q | Z = 1) Pr[Z = 1] − 1 = Ω(km log(smk)), mapping X to Πi , Y to Q, and W to Z in the mutual information bound above. It follows that Ω(km log(smk)) ≤ I(Πi ; Q) ≤ H(Πi ) ≤ EPR ,rand [|Πi |], where |Πi | denotes the length, in bits, of the sequence Πi , and rand is the concatenation of the private randomness of all s players Note that the entropy of Πi is a lower bound on the expected encoding length of |Πi |, by the Shannon coding theorem. By linearity of expectation, P kΠi | = Ω(skm log(smk)), which implies there exists an R and setting of rand for which Pi ER,rand i i kΠ | = Ω(skm log(smk)). It follows that the total communication is Ω(smk log(smk)) bits. 10.3 Lower bounds for distributed PCA on sparse matrices Applying the previous theorem with m = φ gives a communication lower bound for sparse matrices. Corollary 83. Let A ∈ Rφ×n be the hard instance matrix in Section 10.2.2 (with m replaced with φ in noting the number of rows in A):
A = R̃ B1 Iφ B1 Iφ . . . B1 Iφ 0φ×t . Suppose n = Ω(sφ), k ≤ .99φ, and 1 < C < poly(skφ). Now, for arbitrary m consider the matrix  ∈ Rm×n which has A in the top part and the all-zeros (m − φ) × n matrix in the bottom part.  admits the same column partition as A after padding with zeros: 62  =  R̃ 1 B Iφ 1 B Iφ 0(m−φ)×k 0(m−φ)×φ 0(m−φ)×φ  1 ... 0φ×t B Iφ . . . . 0(m−φ)×φ 0(m−φ)×t We denote the new sub-matrices with Âi ∈ Rm×wi . Assume that there is an algorithm (protocol) which succeeds with probability at least 2/3 in having the i-th server output Q̂Âi , for all i, where Q̂ ∈ Rm×m is a projector matrix with rank at most k such that k − Q̂ÂkF ≤ Ck − Âk kF . Then, this algorithm requires Ω(skφ log(skφ)) bits of communication. Note that assuming a word size of Θ(log(skφ)) bits, we obtain an Ω(skφ) word lower bound. 10.4 Lower bounds for distributed column-based matrix reconstruction In this section we develop a communication lower bound for the Distributed Column Subset Selection Problem of Definition 19. 10.4.1 A Net of discretized Deshpande-Vempala hard instances We start by describing the construction of a special sparse matrix from the work of Desphande and Vempala [27], which was further studied by [17]. This matrix gives a lower bound (on the number of columns need to be selected in order to achieve a certain accuracy) for the standard column subset selection problem (see the following Fact). Suppose φ ≥ 2k/ǫ. Consider a (φ + 1) × φ matrix B for which the i-th column is e1 + ei+1 , where ei is the i-th standard basis vector in Rφ+1 . Now define an ((φ + 1)k) × (φk) matrix A with k blocks along the diagonal, where each block is equal to the matrix B. Let n = (φ + 1)k. Fact 84. (Proposition 4 in [27], extended in Theorem 37 of [17]) Let A be as above. If P is an n×n projection operator onto any subset of at most k/(2ǫ) columns of A, then kA − PAk2F > (1 + ǫ)kA − Ak k2F . To prove a communication lower bound for distributed column subset selection, we will need to transform the above matrix A by multiplying by a rounded orthonormal matrix, as follows. We also need to prove a similar lower bound as in the previous fact for this transformed matrix. Lemma 85. Let A, k, φ, 1/ε, and n be as above. For any n × n orthonormal matrix L, let L̃ denote the matrix formed by rounding each of the entries of L to the nearest integer multiple of M , where M ≤ 1/((φ + 1)k)c , for a sufficiently large constant c > 0. If P is an n × n projection operator onto any subset of at most k/(6ǫ) columns of L̃A, then kL̃A − PL̃Ak2F > (1 + ǫ)kL̃A − [L̃A]k k2F . Proof. We will show that, assuming 1/ε < n/3, if P is an n × n projection operator onto any subset of at most k/(2ǫ) columns of L̃A, then kL̃A − PL̃Ak2F > (1 + ǫ/3)kL̃A − [L̃A]k k2F . The lemma will then follow by replacing ε with 3ε. Suppose S is a subset of k/(2ǫ) columns of L̃A for which kL̃A − PL̃Ak2F ≤ (1 + ǫ/3)kL̃A − [L̃A]k k2F , (22) where P is the projection onto the columns in S and [L̃A]k is the best rank-k approximation to L̃A The proof strategy is to relate the left hand side of (22) to kA − QAk2F where Q is the projection onto the columns of A indexed by the same index set S, and simultaneously to relate the right hand side of (22) to kA − Ak k2F . We start by looking at kL̃A − [L̃A]k kF . Notice that L̃A = LA + EA, where E is a matrix whose entries are all bounded in magnitude by M . Therefore, L̃A = LA + F, where F = EA and
1/2
1/2 kFk2 ≤ kEkF kAkF ≤ n2 M 2 · (2(n − 1))1/2 ≤ 2n3 M 2 , 63 where we have used kAkF ≤ (2(n − 1))1/2 . By Weyl’s inequality (Corollary 7.3.8 of [34]), for all i ≤ φk,
1/2 |σi (LA) − σi (LA + F)| ≤ kFk2 ≤ 2n3 M 2 , where σi (X) denotes the i-th singular value of a matrix X. Since L is orthonormal, σi (LA) = σi (A) for all i. It follows that X kL̃A − [L̃A]k k2F = σi2 (L̃A) i>k X
1/2 2 = σi (LA) ± 2n3 M 2 i>k X
1/2 2 = σi (A) ± 2n3 M 2 i>k = kA − Ak k2F ± O(nkAk2 n3 M 2 = kA − Ak k2F ± O(1/n2 ),
1/2 + n4 M 2 ) (23) where the final line follows for a sufficiently small choice of M ≤ 1/nc , using that kAk2 ≤ kAkF = √ O( n). We now look at kL̃A − PL̃AkF . By the triangle inequality, kL̃A − PL̃AkF ≥ kLA − PLAkF − kEAkF − kPEAkF ≥ kLA − PLAkF − kEkF kAkF − kPk2 kEkF kAkF = kA − LT PLAkF − O(1/n3 ), (24) where the equality uses that multiplying by LT preserves norms, as well as that kEkF ≤ 1/poly(n) for an arbitrarily small poly(n) by making M ≤ 1/nc sufficiently small, whereas kPk2 ≤ 1 and √ kAkF = O( n). We claim LT PL = Q, where Q is the projection onto the columns in A indexed by the set S. To see this, let U be an orthonormal basis for the columns in A spanned by the index set S, so that P = LUUT LT . Then LT PL = UUT = Q. Hence, using that kA − QAkF ≤ kAkF = √ O( n) and squaring (24), we have, kL̃A − PL̃Ak2F ≥ kA − QAk2F − O(1/n2 ). (25) Combining (22), (23), and (25), kA − QAk2F ≤ (1 + ǫ/3)kA − Ak k2F + O(1/n2 ). (26) Note that kA − Ak kF ≥ 1 for any k < n, since if we remove the top row of A, creating a matrix B, then kB − Bk kF ≤ kA − Ak kF , yet B is the identity matrix, and so kB − Bk kF ≥ 1. It follows that the additive O(1/n2 ) error in (26) translates into a relative error, provided k < n and for 1/ε < n/3. Hence, kA − QAk2F ≤ (1 + ǫ)kA − Ak k2F , and therefore by Fact 84, we have |S| > k/(2ǫ). 10.4.2 Column sharing of discretized orthonormal matrices We need the following technical lemma about sampling random orthonormal matrices, concerning the number of columns they have in common after rounding. 64 Lemma 86. Let 1 ≤ r ≤ n, and let M ≤ 1/nc for a sufficiently large absolute constant c > 0. Suppose we choose two independently random orthogonal matrices A ∈ Rn×n and B ∈ Rn×n from the Haar measure (so they each have orthonormal columns and orthonormal rows). We then round each of the entries in A to the nearest integer multiple of M , obtaining a matrix Ã, and similarly obtain B̃. Let E denote the event that there exist r distinct indices i1 , . . . , ir ∈ [n] and r distinct indices j1 , . . . , jr ∈ [n] for which Ã∗,iℓ = B̃∗,jℓ for all ℓ = 1, 2, . . . , r, where Ã∗,iℓ denotes the iℓ -th column of Ã. Then, Pr[E] ≤ e−Θ(nr log M ) . Proof. We fix a subset S of r distinct indices i1 , . . . , ir ∈ [n] and a subset T of r distinct indices j1 , . . . , jr ∈ [n] and show that the probability Ã∗,iℓ = B̃∗,jℓ simultaneously for all ℓ ∈ [r] is at most e−Θ(nr log M ) . It then follows by a union bound that      ne 2r n n −Θ(nr log M ) e−Θ(nr log M ) ≤ e−Θ(nr log M ) , Pr[E] ≤ · ·e ≤ r r r as desired. Since A and B are independent and their distribution is permutation-invariant, we can assume iℓ = jℓ = ℓ for ℓ ∈ [r]. Let F be the event that Ã∗,iℓ = B̃∗,jℓ occurs for all ℓ ∈ [r]. If F occurs, then √ we need kA∗,iℓ − B∗,jℓ k∞ ≤ M for all ℓ ∈ [r], which implies kA∗,iℓ − B∗,jℓ k2 ≤ nM for all ℓ ∈ [r]. Let Ar be the leftmost r columns of A, and Br the leftmost r columns of B. Then, if F occurs r kA − Br k22 r ≤ kA − Br k2F = r X ℓ=1 kA∗,ℓ − B∗,ℓ k22 ≤ rnM 2 . Note that Ar and Br have orthonormal columns, and so PAr = (Ar )(Ar )T and PBr = (Br )(Br )T are projection matrices. Then, kAr − Br k2 ≥ k(Ar )(Ar )T − Br (Ar )T k2 /2 + k(Ar )(Br )T − (Br )(Br )T k2 /2 = k(Ar )(Ar )T − (Ar )(Br )T k2 /2 + k(Ar )(Br )T − (Br )(Br )T k2 /2 ≥ k(Ar )(Ar )T − (Br )(Br )T k2 /2, (27) where the first inequality follows since multiplying by a matrix on the right with orthonormal rows cannot increase the spectral norm, the equality follows by taking transposes, and the second inequality follows by the triangle inequality. We need another result from [38]. Theorem 87. (Probability Bound - Claim 5.2 of [38]) Suppose we choose two independent random subspaces Y and S of Rn each of dimension r from the Haar measure, where we denote the projection operators onto the subspaces by YY T and SST , respectively, where Y and S have r orthonormal columns. Then for any δ ∈ (0, 1), Pr[kYY T − SST k2 ≤ δ] ≤ e−Θ(r(n−r) log(1/δ)) . Combining (27) with Theorem 87, we have that for r ≤ n/2, Pr[F] ≤ e−Θ(rn log(M )) , since by √ taking M ≤ 1/nc sufficiently small, we can ensure kAr − Br k2 ≤ rnM ≤ M 1/2 , and so 2M 1/2 upper bounds k(Ar )(Ar )T − (Br )(Br )T k2 . Note that if r > n/2, then in particular we still have that Ã∗,iℓ = B̃∗,jℓ for all ℓ ∈ [r/2], and therefore in this case we also have Pr[F] ≤ e−Θ(rn log M ) . Hence, Pr[E] ≤ e−Θ(rn log M ) , which completes the proof. We now extend Lemma 86 to a k-fold version. 65 Lemma 88. Let M ≤ 1/nc for a sufficiently large absolute constant c > 0. Suppose we independently choose k pairs of matrices Az , Bz ∈ Rn×n , z ∈ [k], each with orthonormal columns (and hence also orthonormal rows). We then round each of the entries in each Az and Bz to the nearest integer multiple of M , obtaining matrices Ãz , and B̃z . Let r = (r1 , . . . , rk ) ∈ (Z≥0 )k and let E be the event that for each z ∈ [k], there exist distinct indices i1 , . . . , irz ∈ [n] and j1 , . . . , jrz ∈ [n] for which Ãz ∗,iℓ = B̃z ∗,jℓ for all Pk ℓ = 1, 2, . . . , rz . Then, Pr[E] ≤ e−Θ(n z=1 rz log M ) . Proof. This follows by Lemma 86, and the fact that the matrices A1 , . . . , Ak , B1 , . . . , Bk are jointly independent. Here we also use a union bound. Corollary 89. Suppose 1 ≤ k < c0 n for an absolute constant c0 > 0. Let r = (r1 , . . . , rk ) ∈ (Z≥0 )k . Let M ≤ 1/nc for a sufficiently large constant c > 0. There exists a set N of eΘ(nt log M ) k-tuples (Ã1 , . . . , Ãk ) of matrices each of dimensions n × n with entries that are integer multiples of M and bounded in magnitude P by 1, such that for every vector r = (r1 , . . . , rk ) ∈ (Z≥0 )k with kz=1 rz ≤ t, it holds that for all distinct ktuples (Ã1 , . . . , Ãk ), (B̃1 , . . . , B̃k ), there exists a z ∈ [k] for which there are no sets S = {i1 , . . . , irz }, T = {j1 , . . . , jz } ⊂ [n] for which Ã∗,iℓ = B̃∗,jℓ for all ℓ = 1, 2, . . . , rz . Proof. For a fixed vector r and a random choice of (A˜1 , . . . , A˜k ), (B˜1 , . . . , B˜k ), by the guarantee of Lemma 88, the complement of the event in the statement of the corollary happens with probability P at most e−Θ(nt log M ) . The number of vectors r with kz=1 rz ≤ t is at most tk , and tk e−Θ(nt log M ) ≤ e−Θ(nt log M ) provided k < c0 n for an absolute constant c0 > 0, so we can apply a union bound over all pairs in a random choice of eΘ(nt log M ) k-tuples. 10.4.3 Hard instance construction We have s servers holding matrices A1 , . . . , A Ps , respectively, where Ai has m rows and a subset of wi columns of an m × n matrix A, where wi = n. For our lower bound we assume 2/ε ≤ φ and kφ ≤ m. Note that for ε less than a sufficiently small constant, this implies that k < c0 kφ, and so the assumption in Corollary 89 is valid (since the n of that corollary is equal to (k + 1)φ We set M = 1/(mn)c for a sufficiently large constant c > 0. We will have A1 being an m × k(φ − 1) matrix for which all but the first kφ rows are zero. We assume 2/ε ≤ φ. On the first kφ rows, A1 is block diagonal containing k blocks, each block being a φ × (φ − 1) matrix. To specify A1 , let the parameters t and n of Corollary 89 equal k/(2ε) and φ, respectively. That corollary gives us a net N of eΘ(kφ(log(mn))/ε) k-tuples of matrices each of dimensions φ × φ with entries that are integer multiple of M and bounded in magnitude by 1. For a random k-tuple (B1 , . . . , Bk ) in N , we create an m × m block diagonal matrix B, which is 0 on all but its first kφ rows and first kφ columns. On its first kφ rows, it is a block diagonal matrix B whose blocks along the diagonal are B1 , . . . , Bk , respectively. Our matrix A1 is then the matrix product of B and an matrix D, where D is m × k(φ − 1) is formed by taking the kφ × k(φ − 1) matrix of Fact 84 and padding it with m − kφ rows which are all zeros. By Lemma 85, no k/(2ǫ) columns of A1 contain a k-dimensional subspace in their span which is a (1 + ǫ)-approximation to the best rank-k approximationt to A1 . Each Ai for i = 2, 3, . . . , s − 1 is the m × m matrix of all zeros. Finally, As is the m × t matrix
of all zeros with t = n − (s − 1)m − k(φ − 1), i.e., A = A1 0m×m 0m×m . . . 0m×m 0m×t . By construction, each column of A has at most φ non-zero entries, since the only non-zero columns are those in A1 , and each column in A1 has the form Bz∗,1 + Bz∗,i for a z ∈ [k] and an i ∈ [φ], so it has at most φ non-zero entries. 66 10.4.4 Main theorem In the Distributed Column Subset Selection Problem in Definition 18, a correct protocol should have each of the s machines simultaneously output a matrix C ∈ Rm×c with c < n columns of A such that kA − CC† Ak2F ≤ (1 + ε)kA − Ak k2F . Theorem 90. Assume 2/ε ≤ φ and kφ ≤ min(m, n). Then, any, possibly randomized, protocol Π which succeeds with probability at least 2/3 in solving the Distributed Column Subset Selection Problem of Definition 18, under the promise that each column of A has at most φ non-zero entries which are integer multiples of 1/(mn)c and bounded in magnitude by 1, for a constant c > 0, requires Ω(sφk(log(mn))/ε) bits of communication. If the word size is Θ(log(mn)), this implies the communication cost of the porticol is Ω(sφk/ε) words. Proof. By construction of A1 , whenever Π succeeds, each of the s machines outputs the same set C of at least t ≥ k/(2ǫ) columns of A1 . Let Πi denote the transcript (sequence of all messages) between the coordinator and the i-th player. We look at the information Πi reveals about A1 , given the event E that Π succeeds, that is, I(Πi ; A1 | E). A critical observation is that any column C∗,i of C, is equal to Bz∗,1 + Bz∗,i for some z ∈ [k], where Bz is as defined above (i.e., A1 = B · D). Hence, given Bz∗,1 and C∗,i , one can reconstruct Bz∗,i . Now, for any random variable W , I(Πi ; A1 | E) ≥ I(Πi ; A1 | E, W ) − H(W ). (28) We let W = (B1∗,1 , . . . , Bk∗,1 ), and observe that H(W ) ≤ log2 ((2M )kφ ) since there are 2M choices for each coordinate of each Bz∗,1 . Hence, H(W ) = O(kφ log(mn)). Note that I(Πi ; A1 | E, W ) = H(A1 ) = log2 (|N |), since conditioned on the protocol succeeding, and given W , by Corollary 89 we can reconstruct A1 . By (28), it follows that (Πi ; A1 | E) ≥ Θ(φk(log(mn))/ǫ) − O(kφ log(mn)) = Ω(φk(log(mn))/ǫ). Note that if Z is an indicator random variable that indicates that E occurs, we have H(Z) ≤ 1, and so I(Πi ; A1 ) ≥ I(Πi ; A1 | E) Pr[E] − 1 = Ω(φk(log(mn))/ǫ), since Pr[E] ≥ 2/3. It follows that Ω(φk(log(mn))/ǫ) ≤ I(Πi ; A1 ) ≤ H(Πi ) ≤ EA1 ,rand [|Πi |], where |Πi | denotes the length, in bits, of the sequence Πi of messages, and rand is the concatenation of the private randomness of all s machines. Note that the entropy of Πi is a lower bound on the expected encoding length of |Πi |, by the Shannon coding theorem. By linearity of expectation, X EA1 ,rand |Πi | = Ω(φk(log(mn))/ǫ), i which implies there exists an A1 and setting of rand for which X |Πi | = Ω(sφk(log(mn))/ǫ). i It follows that the total communication is Ω(sφk(log(mn))/ǫ) bits. 67 Corollary 91. Assume 2/ε ≤ φ and kφ ≤ min(m, n). Then, any, possibly randomized, protocol Π which succeeds with probability at least 2/3 in solving the Distributed Column Subset Selection Problem - rank k subspace version (see Definition 19), promise that each column of A has at most φ non-zero entries which are integer multiples of 1/(mn)c and bounded in magnitude by 1, for a constant c > 0, requires Ω(sφk(log(mn))/ε) bits of communication. If the word size is Θ(log(mn)), this implies the total communication across all s machines is Ω(sφk/ε) words. Proof. As any such protocol is also a protocol for the Distributed Column Subset Selection Problem of Definition 18 (the problem in Definition 19 is more general/difficult than the problem of Definition 18), the proof follows immediately from Theorem 90. 68 

8cs.DS

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 1 Transmission Line Inspires A New Distributed Algorithm to Solve the Linear System of Circuit Fei Wei, Huazhong Yang, Senior Member, IEEE Abstract—Transmission line, or wire, is always troublesome to integrated circuits designers, but it could be helpful to parallel computing researchers. This paper proposes the Virtual Transmission Method (VTM), which is a new distributed and stationary iterative algorithm to solve the linear system extracted from circuit. It tears the circuit by virtual transmission lines to achieve distributed computing. For the symmetric positive definite (SPD) linear system, VTM is proved to be convergent. For the unsymmetrical linear system, numerical experiments show that VTM is possible to achieve better convergence property than the traditional stationary algorithms. VTM could be accelerated by some preconditioning techniques, and the convergence speed of VTM is fast when its preconditioner is properly chosen. Index Terms—Circuit simulation, distributed computing, numerical analysis, transmission line, wire. T I. INTRODUCTION RANSMISSION line, or interconnect, or wire, is one of the main determinants of chip performance, because of the long interconnect latency and large power dissipation [37]. However, the circuit could not work without transmission lines, because the transmission line plays an irreplaceable role to make the distributed circuit stable and scalable [13]. In this paper, transmission lines are used to assist the distributed computation of sparse linear system of circuit. The basic problem of the distributed circuit simulation is to distributedly solve the sparse linear system extracted from circuit [34, 41, 46]. Generally speaking, the linear system of circuit is unsymmetrical, because of the existence of controlled sources. Restricted to the resistor-capacitor (RC) network, the linear system would be symmetrical [39]. Many numerical algorithms are employed to solve the linear system from circuit. KLU is an optimized LU factorization algorithm for circuit, and it is broadly used in commercial and non-commercial circuit simulators [49]. KLU is sequential. SuperLU is an outstanding distributed algorithm to solve the general unsymmetrical sparse linear system on a large number of processors, and it has been widely adopted for circuit simulation [15]. Xyce is a parallel circuit simulator designed for supercomputers, and it uses the Block Jacobi preconditioned GMRES method [50, 4, 3]. Schur complement method is a non-overlapping domain decomposition method making use of the master-slave model. It is simple to be comprehended and implemented, and was frequently used [29, 65, 20]. [7] presents a parallel direct/iterative mixed approach, based on the Schur complement method and preconditioned GMRES method. [41] adopted the preconditioned overlapping domain decomposition method, which is supported by PETSc, a distributed scientific computing package [1]. The highlight of [41] is that it improves the matrix property by exploring the characters of transistor device models. The symmetric linear system of circuit is mainly extracted from the RC networks. Solving RC network is the key for the power grid analysis and thermal analysis [30, 26]. Meanwhile, the sparse linear system generated by the finite element method from elliptic partial differential equations (PDE) is equivalent to a resistor network. The simulation of power grid has been intensively researched in the past 10 years, and a variety of numerical algorithms had been employed to solve the RC network, such as stationary iterative methods [67], Krylov-subspace iterative method [12], algebraic multigrid (AMG) [31, 69], coupled iterative/direct method [32] and FFT method [48]. [42] proposed a stochastic method, and this method was further expanded as a preconditioning method for the diagonal dominant matrix [43]. To compute the RC network on the parallel computing platform, [68] makes use of the domain decomposition method (DDM). [11] adopts a distributed matrix inversion algorithm. The potential of GPU to solve the RC network was illustrated in [6, 16, 17, 48]. VTM is a new distributed and stationary iterative algorithm to solve the sparse linear systems [59]. VTM is similar to Block Jacobi or Overlapped Block Jacobi method [56, 66]. VTM is inspired by the behavior of transmission line [40, 13]. It inserts virtual transmission lines into the circuit to achieve distributed computing [58, 59]. The idea of VTM is derived from the pseudo transient method [57]. Pseudo transient method inserts pseudo capacitors or inductors into the circuit to approve the convergence property of DC analysis. Similarly, VTM inserts pseudo transmission lines into the circuit to approve the convergence property of distributed simulation. Inserting virtual or pseudo electrical element to assist the circuit simulation is a common approach, and the advantage is that the physical property of circuit could be utilized during numerical computation. VTM draws a parallel between distributed computing of linear system and distributed simulation of linear circuit, and it interprets the relationship between numerical analysis and circuit analysis. To solve a sparse linear system in parallel is equalized to simulate a linear circuit in parallel, and vice versa, > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < we might comprehend the distributed numerical algorithm from the viewpoint of circuit theory and microwave network. To partition a circuit, node tearing and branch tearing are two important methods [64, 47]. This paper proposes the third way, which is called wire tearing. Tearing the circuit by wires is not a new idea, since all the subcircuits are connected by wires, but we first apply this concept into the distributed numerical algorithms [58, 59]. Wire tearing can be considered as the insertion of transmission line to connect the torn nodes [61]. The advantage of wire tearing over the branch tearing and node tearing is that it does not bring in any extra energy, because all the wires are passive [27]. As a distributed numerical algorithm, VTM is similar to overlapped Block Jacobi or Block SOR method [2, 45]. We have proved that VTM is convergent to solve arbitrarily-large SPD linear system on arbitrary number of processors [59, 62]. VTM could also be considered as a new type of algebraic domain decomposition method [55]. VTM is similar to the additive Schwarz method with Robin condition, i.e. Schwarz-Robin method [35, 21, 22, 36]. Additive Schwarz method is mainly used to solve the sparse linear system from elliptic partial differential equations (PDE), and VTM is used to solve the general sparse linear system. The partitioning method for VTM, i.e. wire tearing, is different from the traditional partitioning methods for algebraic additive Schwarz method [22, 36]. Furthermore, the preconditioning technique for VTM is more flexible than Optimized Schwarz Method [21]. This paper is organized as follows. Section 2 presents a brief introduction to the linear system of circuit. Section 3 introduces the mathematical description of transmission line. Section 4 proposes the physical background of VTM. Section 5 describes how to partition the circuit by wire tearing. Section 6 details the algorithm of VTM. Section 7 presents the convergence theory. Section 8 provides a simple example. Section 9 discusses the preconditioning techniques for VTM. Numerical experiments are shown in Section 10. We conclude this work in Section 11. In the appendix, we present a proof for the convergence theory. II. LINEAR SYSTEM OF CIRCUIT In virtue of the nodal analysis, one circuit could be described by a sparse linear system [39]. Ax = b (1) Here x represents the nodes’ voltages, and b is the current sources flowing into the nodes. A is the coefficient matrix, which might be symmetrical, or unsymmetrical, depending on the property of circuit. If the circuit contains no voltage sources but only current sources, we call it current-driven circuit. To construct the linear system of current-driven circuit, we only need to use the nodal analysis technique, and do not need the modified nodal analysis technique [39]. If there were voltage sources in the circuit, they might be equalized into current sources, according to the Norton equivalent theory [18]. 2 Fig. 1. Transmission line. (A) The circuit diagram of the transmission line. (B) The equivalent circuit of the lossless transmission line. Lemma 1: For a current-driven linear resistor network, its coefficient matrix, A, is weakly diagonally dominant, and thus symmetrical-non-negative-definite (SNND). Further, if there is at least one resistor connecting to the ground, A is strictly diagonally dominant, and thus symmetric-positive-definite (SPD). The physical insight of this lemma is that the energy of a resistor network is always non-negative. III. TRANSMISSION LINE Transmission line is an important element in electrical and microwave engineering. It is also called cable, wire or interconnect in different contexts. The circuit diagram of the transmission line is illustrated in Fig. 1. The analytical description of the lossless or ideal transmission line is the wave equation [40, 13]. ∂ 2 u ( x, t ) ∂ 2 u ( x, t ) LC (2) = ∂2 x ∂ 2t Here L is the inductance per unit length, C is the capacitance per unit length. The time domain mathematical description of the lossless transmission line is called Transmission Delay Equations, as shown in (3) [40, 8, 59]. u1 (t ) + Z ⋅ i1 (t ) = u2 (t − τ ) − Z ⋅ i2 (t − τ ) u2 (t ) + Z ⋅ i2 (t ) = u1 (t − τ ) − Z ⋅ i1 (t − τ ) (3) Here u1(t) and u2(t) represent the port voltages. i1(t) and i2(t) represent port inflow currents. t is the time variable. τ is the propagation delay. Z is the characteristic impedance, which is positive. > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 3 Z = L/C The propagation delay of the transmission line is: τ = l LC Where l is the length of the transmission line. The iterative formula of Transmission Delay Equations is similar to the Robin boundary condition in Partial Differential Equations and Schwarz-Robin method [35, 21]. There are two main differences: First, the normal derivative in the robin condition is replaced by a current; Second, (3) expresses the time delay explicitly. For a number of coupled transmission lines, (3) could be re-expressed in the matrix-vector form: u1 (t) + Z ⋅ i1 (t) = u 2 (t - τ) - Z ⋅ i 2 (t - τ) u 2 (t) + Z ⋅ i 2 (t) = u1 (t - τ) - Z ⋅ i1 (t - τ) (4) Here Z is the characteristic impedance matrix of the transmission lines. If there is no coupling among transmission lines, Z is diagonal; if there is coupling, Z is a sparse matrix. (4) could be re-expressed as (5): i1 (t) + W ⋅ u1 (t) = −i 2 (t - τ) + W ⋅ u 2 (t - τ) i 2 (t) + W ⋅ u 2 (t) = −i1 (t - τ) + W ⋅ u1 (t - τ) (5) Here W is the characteristic admittance matrix of the wires. W = Z -1 Associated with the principle of parallel computer, we come to realize that, there is conceptual similarity between the distributed computer and the distributed circuit, as shown in Fig. 2 [61]. First, different circuits are connected by wires, which is similar to the situation where different processors are connected by the digital data link. Second, the wire has transmission delay, and there is also communication delay between the processors. As the result, if we assimilate each circuit to a processor, and consider the wire as the data link between processors, then the distributed circuit would be kind of a distributed computer. Fig. 2. Similarity between the distribute circuit and distributed computer. (a) Distributed circuit, two subcircuits are connected by a wire. (b) Distributed computer, two processors are connected by a digital data link. IV. PHYSICAL BACKGROUND In this section, we present an observation from circuit and microwave network. This lemma presents the stability property of the resistor network with wires. Fig. 3 presents an illustration for this lemma. Lemma 2 could be validated in SPICE [38, 44]. Lemma 2: A linear resistor network with wires inside never oscillates unendingly, and it will finally settle down to the steady state, which is the steady state of the resistor network eliminating all the wires. The physical insight of this lemma is that, the energy of a resistor network is limited. Divergence means infinite energy, which is impossible for the resistor network. Lemma 2 is also valid for the resistor-capacitor network. But for the general circuit with wires, it is not always valid. A distributed circuit with wires inside might go unconvergent. The stability is depending on the property of circuit and the characteristic impedance of the wires. Fig. 3. Illustration of the stability theory of resistor network with wires. (a) The resistor network with wires inside. (b) The resistor network without wires. (c) The voltage of node 2a and node 2b in (a) woud converge to the steady state, which is the voltage of node 2 in (b). > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < V. WIRE TEARING To solve the sparse linear system of a circuit, originally there is no transmission line in it, but we might figure out some way to insert transmission lines into the circuit by splitting the nodes. This partitioning technique is called wire tearing. By means of wire tearing, the original circuit is turned into a distributed circuit. It is partitioned into a number of separate sub-circuits by virtual transmission line. Later, we locate each sub-circuit into one processor, and use the digital data link to imitate the behavior of the transmission line. As the result, the distributed circuit is simulated in a distributed way. According to Lemma 2, the resistor network with wires would finally go to the steady state, which is exactly the steady state of the resistor network without wires. This lemma explains the reason that VTM could always be convergent to solve the resistor network. For the sparse linear system of a general circuit, VTM might be unconvergent. There are 4 steps to perform wire tearing for the circuit G . Γ denotes the set of all the nodes in G . Step 1. Set the splitting interface Γ interface ⊆ Γ . V is called interfacial node if V ∈ Γinterface ; otherwise, V is called inner node. Step 2. Split each interfacial node into a pair of nodes, which are called twin nodes. Step 3. Split the resistors and current sources connected to each interfacial node. Step 4. Connect each pair of twin nodes by a length of lossless wire. Add inflow currents to the twin nodes. Example 1. Fig. 4 illustrates an example, in which a simple resistor network is split into two sub-circuits. 4 First we define the set of inner nodes in Sub-circuit 1 as Γ1,inner , the set of inner nodes in Sub-circuit 2 as Γ 2,inner . By reordering the rows, the sparse linear system of circuit (1) could be re-formatted as below by row reordering: C   F1 F2 E1 D1 0 E2   u   f  0   y 1  = g1  D2   y 2  g 2  (6) Here u is the voltage vector of Γ interface . y1 and y2 are the voltage vector of Γ1,inner and Γ 2,inner , respectively. After wire tearing, the interfacial nodes are split, and Γinterface is split into two sets, the twin nodes in Sub-circuit 1, Γ1,twin , and the twin nodes in Sub-circuit 2, Γ 2,twin . The system of sub-circuit 1 is: C1 F  1 E1  u1   f1  i1  (7) = + D1   y 1  g1   0  u1 represents the voltage vector of Γ1,twin . i1 represents the currents flowing into Γ1,twin . After wire tearing, the system of sub-circuit 2 is: C 2 F  2 E 2  u 2   f 2   i 2  = + D2   y 2  g 2   0  (8) Here u 2 represents the voltages vector of Γ 2,twin . i 2 represents the currents flowing into each node of Γ 2,twin . Also we have, Fig. 4. Illustration of the wire tearing. (a) The original resistor network, and node 2 is set to be the interfacial node. (b) Split node 2 into a pair of twin nodes 2a and 2b, and insert a length of transmission line between them. 1/R2=1/R2a+1/R2b. I2=I2a+I2b. (c) Replace the virtual transmission line with its equivalent circuit, and the original resistor network is partitioned into two sub-networks. > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < C1 + C2 = C f1 + f 2 = f u1 , i1 , u 2 , i 2 are called boundary variables, and they satisfy the Transmission Delay Equations (5). i1 (t) + W ⋅ u1 (t) = −i 2 (t - τ) + W ⋅ u 2 (t - τ) If we set the boundary condition as below: u1 (t) = u 2 (t) = u(t) i1 (t) = -i 2 (t) = i(t) Then (6) could be re-obtained from (7) and (8). The above splitting technique is called level-one wire tearing, since the interfacial node is split once. If the interfacial node was split for more than one times, this splitting technique is called multilevel wire tearing [59]. VI. VIRTUAL TRANSMISSION METHOD After partitioning of the circuit, there are 3 steps to perform the parallel computing by Virtual Transmission Method. Step 1. Replace the lossless transmission line with its equivalent circuit. Step 2. Load each sub-circuit into a processor. Step 3. Use the data link to transfer the previous interfacial variables from one processor to another by message passing, and perform the distributed computing. Example 2. Continuing with Example 1, the analytic expression of sub-circuit 1 is obtained by merging (5) and (7): E1  u1 (t)   f1  i1 (t)  = + D1   y 1 (t)  g1   0  C2 + W E 2  u 2 (t)  =  F D2   y 2 (t)  2  f 2 + W ⋅ u1 (t - τ) - i1 (t - τ)    (13) g2   i 2 (t) = − W ⋅ u 2 (t) + W ⋅ u1 (t - τ) - i1 (t - τ) If all the delays of lines are 1, we obtain: C2 + W E 2  u k2   =  F D2   y k2  2  f 2 + W ⋅ u1k −1 - i1k −1    g2   k k k −1 k −1 i 2 = − W ⋅ u 2 + W ⋅ u1 - i1 (14) Later, we locate (11) on Processor 1, and locate (14) on Processor 2. After that, we set the initial values for boundary variables: u10 = u 02 = 0 i10 = i 02 = 0 At last, we do the distributed computing. (9) VII. CONVERGENCE THEORY i1 (t) + W ⋅ u1 (t) = −i 2 (t - τ) + W ⋅ u 2 (t - τ) Eliminate i 1 (t) , and we obtain: C1 + W E1  u1 (t)  =  F D1   y 1 (t)  1  f1 + W ⋅ u 2 (t - τ) - i 2 (t - τ)    g1   (12) Eliminate i 2 (t) , and we obtain: i 2 (t) + W ⋅ u 2 (t) = −i1 (t - τ) + W ⋅ u1 (t - τ) C1 F  1 C2 E 2  u 2 (t)   f 2  i 2 (t)   F D   y (t)  = g  +  0   2 2  2   2  i 2 (t) + Z ⋅ u 2 (t) = −i1 (t - τ) + W ⋅ u1 (t - τ) 5 Theorem 1 (Convergence Theory): Assume a linear resistor network is partitioned into N sub-circuits by wire tearing, if the characteristic impedance matrix of the virtual transmission lines is symmetrical-positive-definite (SPD), then VTM converges at the answer to the resistor network. (10) i1 (t) = − W ⋅ u1 (t) + W ⋅ u 2 (t - τ) - i 2 (t - τ) Set all the delays of transmission lines to be 1, then, C1 + W E1  u1k  f1 + W ⋅ u 2k −1 - i 2k −1   =   F D1   y1k   g1  1  (11) i1k = − W ⋅ u1k + W ⋅ u 2k −1 - i 2k −1 (11) is an SPD sparse linear system, and it could be solved by any sequential algorithms, such as cholesky factorization, algebraic multigrid method (AMG) and conjugate gradient method (CG). Similarly, the analytic expression of sub-circuit 2 is obtained by merging (5) and (8): This conclusion is supported by Lemma 2, and the mathematical proof is given in the appendix. According to the proof of convergence theory, we know that the convergence speed of VTM is depending on the choice of the characteristic admittance matrix W. So W is also called the preconditioner for VTM. VIII. PRECONDITIONING In this section, we discuss the preconditioning techniques for VTM. We first define the input admittance matrix of a circuit or microwave network [18, 13]. From the view of numerical analysis, input admittance matrix is an alias of Schur complement matrix associated with the interface variables u [45, 68]. Example 3. Refer to (7), for Sub-circuit 1, the Schur complement matrix associated with the interface variables u1 is: > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < S1 = C1 - E1D1-1F1 5. S1 is also the input admittance matrix associated with the interface variables u1 in Sub-circuit 1. Similarly, in Sub-circuit 2, the Schur complement matrix associated with the interface variables u2 is: S 2 = C2 - E2 D-12 F2 T1 = ( W − S1 )( W + S1 ) W  S 1 or S 2 . S 1 or S 2 are called the partial input admittance matrix of the large circuit, which are defined as below:  D  -1  S 1 = C1 - E 1 1 F1  D  -1F S = C - E 2 S2 is also the input admittance matrix associated with the interface variables u2 in Sub-circuit 2. Then, the voltage reflection matrix on the interface of Sub-circuit 1 is defined as: −1 The voltage reflection matrix on the interface of Sub-circuit 2 is: T2 = ( W − S 2 )( W + S 2 ) −1 6 2 2 2 2   E 1  is a submatrix of   1 D1    C 2 E C 2 2 is a submatrix of      F2  F2 D2  C Here  1 F C1 E1   F D  . Similarly,  1 1 E2  . D2  This preconditioning technique is called Schur complement  or S is dense approximation method (SCA). Usually S 1 2 matrix, and we need to drop the small elements inside the matrix. After that, the fill-in pattern of W is similar to C1 or C2. If the circuit is a general circuit with controlled sources, i.e. the linear system of circuit is unsymmetrical, the above 5 CF = ρ (T1T2 ) preconditioning techniques could also be used, but VTM might Here ρ ( A ) is the spectral radius of the square matrix A. be unconvergent. Consequently, we know that the convergence speed of VTM is We have to say that, it is still an open problem to choose the depending on the choice of the characteristic admittance matrix preconditioner for the unsymmetrical linear system. In the next W. This viewpoint is identical to the impedance matching section, we illustrate a simple example to show the potential of technique in microwave engineering [13]. VTM to solve the unsymmetrical system. In the ideal case, if we set the preconditioner, W = S1 or S 2 IX. EXAMPLE According to the convergence proof in the appendix, the convergence factor of VTM is: Then, For the symmetrical linear system, a simple example could be found in [59]. Here we illustrate another example to solve As the result, only 1 iteration is needed to achieve convergence. the unsymmetrical linear system of the circuit. Fig. 5 proposes a Based on the knowledge of microwave network, we know that, if tightly-coupled circuit with 4 operational amplifiers connected one port is precisely matched, there would be no energy reflection end-to-end. The linear system of this circuit is: on this port and the network would settle down after 1 reflection. 0   u1   I1   g1 g 21 0 In practice, it is impossible to obtain the accurate input 0 g 0 g 42  u2   I 2  2 admittance matrix of a large circuit, because the dimension of D1  (15) = ρ (T1T2 ) = 0 −1 or D2 might be very large, and it is impossible to obtain D1 and D−21 . So we need to figure out some practical way to approximate W to make W ≈ S1 or S 2 . If the circuit is a resistor network, first we need to assure that W is SPD. Continue with Example 2, there are several ways to choose W. 1. W = α ⋅ I, α > 0 . I is the identity matrix. 2. W = diag ( C1 ) or diag ( C2 ) . We call it diagonal  g13  0 0 g3 0 g34 0   u3    g 4  u4  Here we set: g1 = g 2 = g3 = g 4 = 1 g13 = g34 = g 42 = − g 21 = 10 The loop gain of this circuit is: Gain = g13 g34 g 42 g 21 g1−1 g3−1 g 4−1 g 2−1 = −104 Then we partition this linear system by wire tearing, preconditioner. 3. W = C1 or C2 . This is called overlapped block preconditioner. It is similar to the overlapped block Jacobi method. 4.  I3     I4  W = α ⋅ C1 or α ⋅ C2 , α > 0 . This is called weighted overlapped block preconditioner (WOB). As the result, we obtain 2 sub-systems: > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <  g1 0   g13  g 2b  0   0 Here: g 21 g2a 0 0 g 3b g 34 0   u1   I1  0  u2 a  =  I 2 a + i2 a  g3a  u3a   I 3a + i2b  g 42  u2b   I 2b + i2b  0   u3b  =  I 3b + i3b  g 4   u4   I 4  7 g 2 a + g 2 b = g 2 , g 3 a + g 3b = g 3 (16) I 2 a + I 2 b = I 2 , I 3 a + I 3b = I 3 For simplicity, we set: g 2 a = g 2b = g 3a = g3b = 0.5 The transmission delay equations of T2 are: (17) i2 a (t ) + W2 ⋅ u2 a (t ) = −i2b (t − τ ) + W2 ⋅ u2b (t − τ ) i2b (t ) + W2 ⋅ u2b (t ) = −i2 a (t − τ ) + W2 ⋅ u2 a (t − τ ) (18) The transmission delay equations of T3 are: Fig. 5. Solving the unsymmetric linear system by VTM . (a) The original circuit with operational amplifiers. (b) Equivalent circuit of (a), with voltage controlled current sources (VCCS). (c) Partition the circuit by wire tearing. Node 2 and 3 is set to be the interfacial node. (d) Partition the equivalent circuit by wire tearing. 1/R2=1/R2a+1/R2b. I2=I2a+I2b. 1/R3=1/R3a+1/R3b. I3=I3a+I3b. (e) Replace the virtual transmission line with its equivalent circuit, and thus the original circuit is partitioned into two sub-circuits. > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < i3a (t ) + W3 ⋅ u3a (t ) = −i3b (t − τ ) + W3 ⋅ u3b (t − τ ) 8 TABLE I SPARSE MATRICES PROPERTY i3b (t ) + W3 ⋅ u3b (t ) = −i3a (t − τ ) + W3 ⋅ u3a (t − τ ) (19) We set: W2 = 0.5 , W3 = 20000 Merge (16), (18) and (19), we obtain the linear system for Sub-circuit 1: 0  u2 a (t )   I 2 a + i2 a (t )  g2a  =  g g −1 g     −1  13 1 21 g3a   u3a (t )   I 3a − g13 g1 I1 + i3a (t )  i2 a (t ) + W2 ⋅ u2 a (t ) = −i2b (t − τ ) + W2 ⋅ u2b (t − τ ) Name Interface nodes Total nodes Property Source Grid2d 100 10000 SPD MESHPART Grid3d 900 27000 SPD H. Qian Fv2 129 9801 SPD UFpack Wang 870 26064 Unsymm. UFpack Pchip 13 2298 Unsymm. MCNC’90 TABLE II ITERATION NUMBER TO ACHIEVE CONVERGENCE Method Grid2d Grid3d Fv2 Wang Jacobi 27759 5238 212 50773 -- GS 18373 2950 109 27423 -- Similarly, we get the linear system for Sub-circuit 2: SGS 11773 1571 79 26842 --  g 2b g 42 g g34  u2b (t )   I 2b − g 42 g I + i2b (t )     = g 3b   u3b (t )   I 3b + i3b (t )  0  i2b (t ) + W2 ⋅ u2b (t ) = −i2 a (t − τ ) + W2 ⋅ u2 a (t − τ ) SOR 14253 2049 66 16289 -- SSOR 8839 1073 74 29135 -- BJ 966 227 50 563 952 i3a (t ) + W3 ⋅ u3a (t ) = −i3b (t − τ ) + W3 ⋅ u3b (t − τ ) (20) −1 4 −1 4 4 i3b (t ) + W3 ⋅ u3b (t ) = −i3a (t − τ ) + W3 ⋅ u3a (t − τ ) (21) Finally we solve this example on 2 processors by VTM, and the computing result is compared with a number of well-known stationary iterative algorithms, e.g. Jacobi, Gauss Seidel (GS), Successive Overrelaxation Method (SOR), Symmetric SOR (SSOR), Block Jacobi (BJ) [45]. Here the global iterative index k increases to k+1 until all the variables are updated. As shown in Fig. 6, the traditional stationary algorithms are Pchip OBJ 490 114 23 287 162 VTM_Diag 963 335 42 854 160 VTM_OB 491 115 23 287 161 VTM_WOB 296 68 22 175 -- VTM_SCE 275 75 14 186 -- However, for the large unsymmetrical linear system, we have not yet figure out an effective method to choose W. The convergence property of VTM is depending on the property of linear circuit, as well as the preconditioner W. X. NUMERICAL EXPERIMENTS Fig. 6. Comparison of the convergence speed of iterative algorithms for resistor network. unconvergent for this circuit, but VTM is convergent. The key is to choose proper characteristic admittances for the virtual transmission lines. According to this example, we conclude that there exists a parallel between the stability of VTM and the stability of distributed linear circuit (or microwave network). The traditional techniques to stabilize a distributed circuit might be transplanted to stabilize VTM. In this section we compare VTM with the traditional stationary algorithms for both the symmetrical matrix and unsymmetrical matrix. We use MATLAB and SIMULINK as our test environment [53, 54], and use METIS and MESHPART as our partitioning methods [28, 51, 19]. The test matrices are described in Table 1. Grid2d and Grid3d are structured resistor networks [51, 43]. Fv2 is a 2D mesh constructed by finite element method (FEM) [14]. Pchip is extracted from a MOS integrated circuit of MCNC’90 testbench [52]. Wang is a sparse matrix from semiconductor device simulation [14]. Table 2 illustrates the computing results of the stationery iterative algorithms. BJ, OBJ and VTM are tested on 2 processors. Jacobi, GS, Symmetrical GS (SGS), SOR and SSOR are tested on 1 processor. VTM_Diag means VTM with diagonal preconditioner. VTM_OB represents VTM with overlapped block preconditioner. VTM_WOB is VTM with weighted overlapped block preconditioner. VTM_SCA means VTM with Schur complement approximation preconditioner. All these preconditioning methods were previously described in Section 8. According to Table 2, we conclude that VTM is convergent to solve the SPD system, and its convergence speed is fast when using the SCA or WOB preconditioner. The convergence speed > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < of OBJ and VTM_OB are same. For the unsymmetrical linear systems, VTM might be unconvergent. At last, we test VTM on a various number of processors, as shown in Fig. 7. In this case, we use Grid3d as the testbench. The test result indicates that the convergence speed of VTM is slightly relative to the number of processors. 9 If the delays of virtual transmission lines are different and stochastic, VTM would turn to be a fully asynchronous and chaotic numerical algorithm [58]. If we replace each variable by a piece of waveform, VTM would become a waveform-relaxation algorithm [60]. To simulate the post-layout integrated circuit in parallel, we have figured out another method to distributedly solve the nonlinear delay differential equations extracted from the distributed circuits [61]. APPENDIX Here we prove that, if the graph of the resistor network is 2-partete, VTM converges. Lemma 3: Assume a linear resistor network is partitioned into 2 sub-circuits by level-one wire tearing, if the characteristic impedance matrix Z of the virtual transmission lines is SPD, VTM converges at the answer to the resistor network. Proof: Fig. 7. Test VTM on a number of processors. First, we consider the resistor network with inner nodes, and eliminate these inner nodes by Schur complement technique. Then, we prove Lemma 3 for the resistor network without inner nodes. The sparse linear system of the resistor network is: C   F1 F2 XI. CONCLUSION AND FUTURE WORK In this paper, we introduce VTM to solve the sparse linear system of circuit in parallel. VTM is a distributed and stationary iterative numerical algorithm, and it is efficient to solve the resistor networks. VTM proves many insightful conclusions linking numerics and electrics. The physical background of VTM is electric circuit, and this makes this algorithm different from the traditional numerical algorithms. As presented in this paper, we have borrowed many concepts from circuit theory to describe this algorithm. According to the simple example presented in Section 9, we conclude that there exists similarity between the stability of VTM and the stability of distributed linear circuit (or microwave network). The convergence speed of VTM is highly depending on the characteristic admittance matrix of the virtual transmission lines, i.e. its preconditioner W. A number of preconditioning techniques are implemented to accelerate VTM. Experiments indicate that, if W is properly chosen, the performance of VTM would be appreciable; if not, VTM would be plain. Nevertheless, the computation cost to design the preconditioner W should be aware of [5]. For the SPD linear system, we have proved that VTM is convergent; for the large unsymmetrical linear system, we have not yet figure out an effective way to choose the characteristic admittance matrix to guarantee convergence. This is an open problem. VTM could also be used to solve the nonlinear system, as we pointed out in [59]. However, the convergence analysis of nonlinear system is more complicated than the linear system. This problem should be further studied. E1 D1 0 E2   u   f  0   y 1  = g1  D2   y 2  g 2  Eliminate the inner variables y1 and y2, we obtain: (C - E D 1 F - E2 D-12 F2 ) u -1 1 1 = f - E1 D1-1g1 - E2 D-12 g 2 y1 = -D1-1F1u + D1-1g1 y 2 = -D-12 F2u + D-12 g 2 Set, S = C - E1D1-1F1 - E 2 D-12 F2 , r = f - E1D1-1g1 - E2 D-12 g 2 , then, (22) S ⋅u = r S is called the Schur complement matrix associated with the interface variables u. C E   is SPD, then the Schur  F D −1 complement matrix S = C - ED F is SPD. Lemma 4: If A =  If the circuit is split into two sub-circuits by level-one wire tearing, the system of Sub-circuit 1 is (7): C1 F  1 E1  u1   f1  i1  = + D1   y 1  g1   0  Eliminate the inner variables y1: > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < (C - E1D1-1F1 ) u1 = f1 - E1D1-1g1 + i1 1 u k2 = ( W + S 2 ) ( W − S1 ) u1k −1 −1 + ( W + S2 ) r y 1 = -D1-1F1u1 + D1-1g1 -1 -1 Set S1 = C1 - E1 D1 F1 , r1 = f 1- E1 D1 g1 , then, S1 ⋅ u1 = r1 + i1 (23) u1k = ( W + S1 ) ( W − S2 ) −1 × ( W + S 2 ) ( W − S1 ) u1k − 2 −1 −1 + ( W + S1 ) ( W − S 2 ) ( W + S 2 ) r −1 +( W + S1 ) r Eliminate the inner variables y2: 2 - E 2 D-12 F2 ) u 2 = f 2 - E2 D-12 g 2 + i 2 -1 2 2 2 -1 2 2 2, T1 = ( W − S1 )( W + S1 ) -1 2 2 S 2 = C2 - E D F r2 = f 2- E2 D-12 g 2 , then, S 2 ⋅ u 2 = r2 + i 2 (24) −1 + ( W − S1 ) T1T2 ⋅ r −1 (28) ( W − S1 ) u1k = T1T2 ( W − S1 ) u1k −2 (29) +T1T2 ⋅ r + T1 ⋅ r If we make use a new variable: = f - E1D1-1g1 - E2 D-12 g 2 a1k = ( W − S1 ) u1k =r Until now, we have eliminated all the inner nodes in the resistor network by Schur complement technique, and all the nodes are interfacial nodes. The linear system of (22) is partitioned into two sub-systems of (23) and (24) by wire tearing. The iterative formula is Transmission Delay Equations (5), i1k + W ⋅ u1k = −i 2k −1 + W ⋅ u 2k −1 i k2 + W ⋅ u 2k = −i1k −1 + W ⋅ u1k −1 then, (29) could be expressed as (30): a1k = T1T2 ⋅ a1k − 2 + h1 where h1 = T1 ⋅ T2 ⋅ r + T1 ⋅ r . (30) In the following text, we will prove that (30) is convergent. Here we use Z-1 instead of W. T1 = ( Z −1 − S1 )( Z −1 + S1 ) −1 = Z −1 ( I − Z ⋅ S1 )( I + Z ⋅ S1 ) Z −1 Here W is the characteristic admittance matrix of the Z = W −1 . i1k and i k2 by combine (23), (24) and (5), we find, W ⋅ u1k + ( S1u1k - r1 ) = W ⋅ u 2k −1 - ( S 2u 2k −1 - r2 ) T2 = ( Z −1 − S 2 )( Z −1 + S 2 ) =Z −1 (31) −1 ( I − Z ⋅ S 2 )( I + Z ⋅ S 2 ) −1 (32) Z Lemma 5: If Z is an SPD matrix of dimension n, then there ( W + S1 ) u1k − r1 = ( W − S 2 ) u 2k −1 + r2 exists a matrix Thus, Similarly, we obtain: u1k = ( W − S1 ) T1T2 ( W − S1 ) u1k − 2 Reformat (28) as (29): =S r1 + r2 = f 1- E1D1-1g1 + f 2 - E 2 D-12 g 2 ( W − S 2 ) u 2k −1 −1 + ( W + S1 ) r Then, −1 = C - E1D1-1F1 − E2 D-12 F2 u1k = ( W + S1 ) −1 +( W − S1 ) T1 ⋅ r S1 + S 2 = C1 - E1D1-1F1 + C2 - E2 D-12 F2 Eliminate (27) −1 T2 = ( W − S 2 )( W + S 2 ) S2 is called the Schur complement matrix associated with the interface variables u2 in Sub-circuit 2. Consequently, we have: transmission lines. W is SPD. −1 Set: y 2 = -D F u + D g Set (26) u1k , E 2  u 2   f 2   i 2  = + D2   y 2  g 2   0  (C −1 Merge (25) and (26), we come to know the iterative formula for S1 is called the Schur complement matrix associated with the interface variables u1 in Sub-circuit 1. The system of Sub-circuit 2 is (8): C 2 F  2 10 −1 (25) Z , which satisfies: Z T Z = Z. Proof: Because Z is SPD, Z = Q T ΛQ = Q T Λ ΛQ = Z T Z > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < T1 = Z −1 ( I − Z ⋅ S1 )( I + Z ⋅ S1 ) Z −1 T where Q Q = I . I is the identity matrix. Λ = diag (α1 , α 2 , , α n ) , α i ∈  + . )( )  ( ) ( )     ( Z Q )( Z Q )  ×  Z + Z Q Λ  ( ) ( Z Q )   = Z ( Z Q )(I − Λ )( Z Q ) ×( Z Q ) (I + Λ ) ( Z Q ) Z = Z ( Z Q ) ( I − Λ )( I + Λ ) ( Z Q ) Z  1− λ 1− λ  = Z ( Z Q ) ⋅ diag  , , ⋅ Z Q ) 1+ λ  ( 1+ λ ( −1 T T  Z Q1 Z Q1  = Z −1  T  − Z TQ Λ Z Q1  1 1  Λ = diag ( α1 , α 2 , , α n ) . Z = ΛQ . End of proof [25]. T Lemma 6: If both Z and S are SPD matrix of dimension n, then Z ⋅ S has the same eigenvalues as T Z ⋅S ⋅ Z . Proof: Because S and Z are SPD, Z ⋅ S is positive-definite, and T Z S Z is SPD. T Z ⋅ S ⋅ Z = QTQ T , where QQ T = I , T = diag (t1 , t 2 , , t n ) , t i > 0, i = 1,2, , n . Then, Assume (Z) = Z ⋅ QTQ ⋅ ( Z ) = ( Z Q) ⋅ T ⋅ ( Z Q) T T -1 T Z⋅S = Z ⋅ Z ⋅S ⋅ Z ⋅ T T -1 T Similarly, we obtain, λi > 0 , i = 1, , n , ) ( T Z Q1 ) -1 . T Z ⋅ S 2 ⋅ Z = Q 2 Λ 2Q 2 T where Λ 2 = diag ( μ1 , μ 2 , , μ n ) , μi > 0 , i = 1, , n , Q 2Q 2 T = I . Z ⋅ S 2 = ( T ) ( Z Q2 Λ 2 Then, (31) could be decomposed as below: T -1 T 1 1 1 -1 T −1 1 1 1 T −1 = Z −1 Z Q1 Λ 1 T T = Z −1 T ( 1 1 T 1 -1 T −1 1 1 1 1 n 1 n T -1 1 Z −1 Z ⋅ S1 ⋅ Z = Q1 Λ1Q1T Q1Q1T = I . Z ⋅ S1 = 1 T2 = Z −1 ( I − Z ⋅ S 2 )( I + Z ⋅ S 2 ) Z According to Lemma 5 and 6, we come to know that, T -1 T 1 Similarly, (32) could be decomposed as: End of proof. where Λ1 = diag (λ1 , λ2 , , λn ) , −1 1 T −1 -1 −1 T 1 −1 T -1 T 11 Z Q2 ) -1 . ( ( ) ( )Z  1− μ 1− μ  , , Z Q ) ⋅ diag  ⋅ Z Q ) 1+ μ  ( 1+ μ T Z Q 2 ( I − Λ 2 )( I + Λ 2 ) T 2 −1 T Z Q2 1 n 1 n -1 T 2 The spectral radius of T1T2 is defined as: ρ (T1T2 ) = lim ( T1T2 ) k 1/ k k →∞ So we first calculate ( T1T2 ) norm of the square matrix A. k . Here A is the spectral -1 Z > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < ( T1T2 ) ( ) ( ) -1 T T −1  −1  Z Z Q I Λ I Λ Z Q Z  − + ( )( ) 1 1 1 1  =   -1 ×Z −1 Z TQ ( I − Λ )( I + Λ )−1 Z TQ Z 2 2 2 2   Z ≤ Z × = ) (  Q1 ( I − Λ1 )( I + Λ1 )−1 Q1T   Z   ×Q 2 ( I − Λ 2 )( I + Λ 2 )−1 Q T2    ≤ ≤ ( −1 −1 Q1 ( I − Λ1 )( I + Λ1 ) Q1T −1 Z Z Z −1 −1 k   ⋅    1 − λ1 1 − λn  ⋅ diag  , ,  1 + λn   1 + λ1 k Z ≤ Z  1 − λ1 1 − λn  Z ⋅ max  , ,   1+ λ 1 + λn  1   1 − μ1 1 − μn  × max k  , , ⋅ Z 1 + μn   1 + μ1 −1 k As the result, ρ (T1 T2 ) = lim ( T1 T2 ) k 1/ k  1 − λ1 1 − λn ρ (T1 ) = max  , , 1 + λn  1 + λ1 Because μi > 0 , i = 1, , n , ( W − S2 ) −1 × ( W + S 2 ) ( W − S1 ) u1k − 2 −1 −1 + ( W + S1 ) ( W − S 2 ) ( W + S 2 ) r −1 +( W + S1 ) r k −1 −1 −1 −1 u1∞ = I − ( W + S1 ) ( W − S 2 )( W + S 2 ) ( W − S1 )    −1 −1 −1 × ( W + S1 ) + ( W + S1 ) ( W − S 2 ) ( W + S 2 )  ⋅ r   = ( W + S1 ) − ( W − S 2 )( W + S 2 )  −1 ( W − S1 ) −1 −1 × I + ( W − S 2 ) ( W + S 2 )  ⋅ r   ( ( ) ) ( ) −1 −1  I + W − S W + S −1 −1  ( 2 )( 2)   =  r −1 × I − ( W − S 2 )( W + S 2 ) W + S1  ( ) −1 ( W + S ) ( ( W + S ) + ( W − S ) )−1  2 2 2  r = × ( ( W + S ) − ( W − S ) ) ( W + S )−1 W + S  2 2 2 1  = ( W + S 2 )( 2 W )  k →∞  1 − λ1 1 − λn  ≤ max  , ,  1 + λn   1 + λ1  1 − μ1 1 − μn , , × max  1 + μn  1 + μ1 = ρ (T1 ) ρ (T2 ) Because λi > 0 , i = 1, , n , So we conclude that VTM converges for 2-partite resistor network [45]. According to (27), when this algorithm is convergent,  I − ( W − S )( W + S )−1 W +  2 2  = −1   I + ( W − S 2 )( W + S 2 ) S1    −1 × I + ( W − S 2 ) ( W + S 2 )  ⋅ r   k  1 − μ1 1 − μn  × diag  , ,  ⋅ 1 + μn   1 + μ1 ρ (T1T2 ) < 1 Z  Q ⋅ ( I − Λ )( I + Λ )−1 ⋅ Q T  1 1 1 1  ⋅ −1  T   × Q 2 ⋅ ( I − Λ 2 )( I + Λ 2 ) ⋅ Q 2   ( I − Λ )( I + Λ )−1 ⋅ 1 1 ⋅ −1  × ( I − Λ 2 )( I + Λ 2 )  Consequently, u1k = ( W + S1 ) ⋅ ×Q 2 ( I − Λ 2 )( I + Λ 2 ) Q T2 k   < 1 .  ρ (T2 ) = max  u1k = u1k − 2 = u1∞ , then, Z k −1 ⋅ ) k −1 =  1 − μ1 1 − μn , , 1 + μn  1 + μ1 k 12 −1 ( 2S 2 )( W + S 2 ) −1 −1 W + S1  r  −1 −1 = ( W + S 2 ) W −1S 2 ( W + S 2 ) W + S1  r   −1    −1 = ( I + S 2 W −1 ) S 2 ( I + W −1S 2 ) + S1  r   −1 −1 = ( S 2 + S 2 W −1S 2 )( I + W −1S 2 ) + S1  r   −1 −1 = S 2 ( I + W −1S 2 )( I + W −1S 2 ) + S1  r     < 1 ;  = [ S 2 + S1 ] r −1 = S −1 ⋅ r Until now, we have proved that, VTM is convergent for 2-partite resistor network when using level-one wire tearing. > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < For the more general cases, the resistor network is k-partite, and the wire tearing might be level-one or multilevel. A general proof was presented in [62]. ACKNOWLEDGMENT Discussions with Hao Zhang, Yi Su, Bin Niu, Yu Wang, Wei Xue, Peng Zhang and Chun Xia were very helpful. Thanks are due to Qi Wei, Bo Zhao, Xia Wei, Xiaojian Mao, Yongpan Liu, Fei Qiao and Rong Luo for encouragement and support. We are very grateful for the public domain softwares provided by John R. Gilbert, Tim A. Davis, Robert Bridson, Haifeng Qian and et al. [17] [18] [19] [20] [21] REFERENCES [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] S. Balay, K. Buschelman, V. Eijkhout, W. D. Gropp, D. Kaushik, M. G. Knepley, L. Curfman McInnes, B. F. Smith and H. Zhang, PETSc Users Manual, ANL-95/11 - Revision 2.1.5, Argonne National Laboratory. R. Barrett, M. Berry, T. Chan, J. Demmel, J. Donato, J. Dongarra, V. Eijkhout, R. Pozo, C. Romine and H. Van der Vorst. Templates for the solution of Linear Systems: Building Blocks for Iterative Methods, 2nd Edition, SIAM, 1994. A. Basermann, U. Jaekel, and K. Hachiy. Preconditioning parallel sparse iterative solvers for circuit simulation. In Proceedings of the 8th SIAM Proceedings on Applied Linear Algebra, Williamsburg VA, 2003. A. Basermann, U. Jaekel, and M. Nordhausen. Parallel iterative solvers for sparse linear systems in circuit simulation. Fut. Gen. Fut. Gen.. Comput. Sys., 21(8):1275–1284, 2005. M. Benzi. Preconditioning techniques for large linear systems: A survey. Journal of Computational Physics, 182:418–477, 2002. J. Bolz, I. Farmer, E. Grinspun, and P. Schröoder. Sparse matrix solvers on the GPU: conjugate gradients and multigrid. In ACM SIGGRAPH 2003 (San Diego, California, July 27 - 31, 2003). C. W. Bomhof and H. A. Van der Vorst, A parallel linear system solver for circuit simulation problems, Numerical Linear Algebra with Applications, 7 (2000). F. H. Branin. Transient analysis of lossless transmission lines, Proceedings of the IEEE, vol. 55, pp. 2012–2013, Nov. 1967. R. Bridson. A MATLAB CMEX interface to the Metis library. Available at http://www.stanford.edu/~rbridson/download/metismex.c R. Bridson and W.-P. Tang. Refining an approximate inverse. Journal on Computational and Applied Math, 123 (2000), Numerical Analysis 2000 vol. III: Linear Algebra, pp. 293-306. S. Cauley, V. Balakrishnan, Cheng-Kok Koh, A parallel direct solver for the simulation of large-scale power/ground networks, IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, v.29 n.4, p.636-641, April 2010. T. H. Chen and C. C.-P Chen. Efficient large-scale power grid analysis based on preconditioned Krylov-subspace iterative methods. In Proc. IEEE/ACM DAC, pages 559--562, 2001. R. E. Collin. Foundations for microwave engineering, 2nd edition, Wiley-IEEE Press, 2000. T. A. Davis and Y. F. Hu. The University of Florida Sparse Matrix Collection. Submitted to ACM Transactions on Mathematical Software. Available at: http://www.cise.ufl.edu/research/sparse/matrices/ J. Demmel, J. Gilbert, and X. Li. An asynchronous parallel supernodal algorithm for sparse gaussian elimination. SIAM J. Matrix Analysis and Applications, 20(4): 915–952, 1999. Zhuo Feng, Peng Li. Multigrid on GPU: tackling power grid analysis on parallel SIMT platforms, Proceedings of the 2008 IEEE/ACM [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35] [36] 13 International Conference on Computer-Aided Design, November 10-13, 2008, San Jose, California. Zhuo Feng, Zhiyu Zeng. Parallel multigrid preconditioning on graphics processing units (GPUs) for robust power grid analysis. In Proceedings of the 47th Design Automation Conference (Anaheim, California, June 13 18, 2010). DAC '10. ACM, New York, NY, 661-666. T. L. Floyd. Principles of electric circuits, 6th edition, Prentice Hall, 1999. David Fritzsche, Andreas Frommer, and Daniel B. Szyld, Overlapping blocks by growing a partition with applications to preconditioning, Research Report 10-07-26, Department of Mathematics, Temple University, July 2010. N. Frohlich, B. M. Riess, U. A. Wever, and Q. Zheng A New Approach for Parallel Simulation of VLSI Circuits on a Transsitor Level. TCAD, June 1998. M. J. Gander, L. Halpern, and F. Nataf. Optimized Schwarz Methods. In T. Chan, T. Kako, H. Kawarada, O. Pironneau (eds.), Proceedings of the Twelveth International Conference on Domain Decomposition, DDM press, 2001, pp. 15–27. M. J. Gander, Schwarz Methods in the Course of Time, Electronic Transactions on Numerical Analysis, 31:228–255, 2008. A. George and J. W. Liu. Computer Solution of Large Sparse Positive Definite Systems. Prentice-Hall, Englewood Cliffs, New Jersey, 1981. John R. Gilbert, Gary L. Miller, and Shang-Hua Teng. Geometric mesh partitioning: Implementation and experiments. SIAM J. Scientific Computing 19:2091-2110, 1998. G. H. Golub and C. F. Van Loan, Matrix computations. Johns Hopkins University Press, 1989. Wei Huang. HotSpot: A chip and package compact thermal modeling methodology for VLSI design. Ph.D. Dissertation, University of Virginia, 2007. Russell Kao. Piecewise Linear Models for Switch-Level Simulation. Chapter 5.6.1, Node and Branch Tearing. Technical Report, CSL-TR-92-532, Stanford University, 1992. G. Karypis and V. Kumar. Metis 4.0: Unstructured graph partitioning and sparse matrix ordering system, Technical report, Department of Computer Science, University of Minnesota, 1998. Available at http://www.cs.umn.edu/metis D. P. Koester. Parallel Block-Diagonal-Bordered Sparse Linear Solvers for Power Systems Applications. Ph.D dissertation, Syracuse University. 1995. J. N. Kozhaya and S. R. Nassif. Fast Power Grid Simulation. In Proceedings of the 37th Design Automation Conference, 2000. J. N. Kozhaya, S. R. Nassif, and F. N. Najm, A multigrid-like technique for power grid analysis, IEEE Trans. Computer-aided Design, vol. 21, pp. 1148--1160, Oct. 2002. Zhao Li, C.-J. Richard Shi. A coupled iterative/direct method for efficient time-domain simulation of nonlinear circuits with power/ground networks. ISCAS (5) 2004: 165-168. Xiaoye S. Li. An overview of SuperLU: Algorithms, implementation, and user interface. ACM Trans. Math. Softw. 2005. Peng Li. What Is Parallel Circuit Simulation? ACM/SIGDA E-NEWSLETTER, vol. 40, No. 4, April 1, 2010 P. L. Lions, On the Schwarz alternating method III: a variant for nonoverlapping subdomains, Third International Symposium on Domain Decomposition Methods for Partial Differential Equations, 1989, Houston, Texas. Sébastien Loisel and Daniel B. Szyld, On the convergence of Optimized Schwarz Methods by way of Matrix Analysis , Domain Decomposition Methods in Science and Engineering XVIII, Michel Bercovier, Martin Gander, Ralf Kornhuber, and Olof B. Widlund, editors. Lecture Notes in Computational Science and Engineering, Vol. 70, Springer, 2009, pages 363-370. > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < [37] J. D. Meindl. Beyond Moore's Law: The Interconnect Era. Computing in Science and Engineering. 2003. [38] L. Nagel. SPICE2: A Computer Program to Simulate Semiconductor Circuits, Electronics Research Laboratory Report No. ERL-M520. University of California, Berkeley, 1975. [39] Jan Ogrodzki. Circuit simulation methods and algorithms. CRC Press, 1994. [40] H. J. Pain. The physics of vibrations and waves, Wiley, 1976. [41] He Peng, Chung-Kuan Cheng. Parallel transistor level full-chip circuit simulation. DATE 2009: 304-307. [42] H. Qian, S. R. Nassif, and S. S. Sapatnekar, Power grid analysis using random walks, IEEE Trans. Computer-aided Design, vol. 24, pp. 1204--1224, Aug. 2005. [43] H. Qian and S. S. Sapatnekar, Stochastic Preconditioning for Diagonally Dominant Matrices, SIAM Journal on Scientific Computing, Vol. 30, No. 3, pp. 1178 – 1204, March, 2008. [44] T. L. Quarles. Analysis of Performance and Convergence Issues for Circuit Simulation. ERL Memo No. UCB/ERL M89/42 April 1989. [45] Y. Saad. Iterative Methods for Sparse Linear Systems. The PWS Publishing Company, Boston, 1996. Second edition, SIAM, Philadelphia, 2003. [46] R. A. Saleh, K. A. Gallivan, M. C. Chang, et al. Parallel circuit simulation on supercomputers. Proceedings of the IEEE, 1989. [47] Alberto Sangiovanni-Vincentelli, Li-Kuan Chen, and Leon O. Chua. A new tearing approach – node-tearing nodal analysis. In IEEE International Symposium on Circuits and Systems, 1977. [48] Jin Shi, Yici Cai, Wenting Hou, Liwei Ma, Sheldon X.-D. Tan, Pei-Hsin Ho, Xiaoyi Wang. GPU friendly fast Poisson solver for structured power grid network analysis, Proceedings of the 46th Annual Design Automation Conference, July 26-31, 2009, San Francisco, California. [49] Ken Stanley, T. A. Davis. KLU: a "Clark Kent" sparse LU factorization algorithm for circuit matrices. 2004 SIAM Conference on Parallel Processing for Scientific Computing (PP04). Originally appeared in NA Digest, 1997. [50] Heidi K. Thornquist, Eric R. Keiter, Robert J. Hoekstra, David M. Day, Erik G. Boman, A parallel preconditioning strategy for efficient transistor-level circuit simulation, Proceedings of the 2009 International Conference on Computer-Aided Design, November 02-05, 2009, San Jose, California. [51] John R. Gilbert. Meshpart, a public domain matlab toolbox for sparse matrix partitioning. Available at http://www.cerfacs.fr/algor/Softs/MESHPART/ [52] CircuitSim90, 1990 Circuit Simulation and Modeling Workshop at MCNC, Available at http://www.cbl.ncsu.edu/CBL_Docs/csim90.html [53] Mathworks Corparation. Matlab User Manual. R13. 2002. [54] Mathworks Corparation. Simulink User Manual. R13. 2002. [55] A. Toselli and O. Widlund. Domain Decomposition Methods – Algorithms and Theory. Springer Series in Computational, Mathematics 34, Springer, Berlin, Heidelberg, 2005. [56] R. S. Varga. Matrix Iterative Analysis. Prentice-Hall, Englewood Cliffs, New Jersey, 1962. Second Edition, Springer Series in Computational Mathematics 27, Springer, Berlin, Heidelberg, New York, 2000. [57] W. T. Weeks, A. J. Jimenez, G. W. Mahoney, and D. Mehta. Algorithms for ASTAP - A Network-Analysis Program. IEEE Trans.Circuit Theory, CT-20(4):628-634, 1973. [58] Fei Wei, Huazhong Yang. Directed Transmission Method, A Fully Asynchronous Approach to Solve Sparse Linear Systems in Parallel. In ACM Proceedings of the 20th Symposium on Parallelism in Algorithms and Architectures (Munich, Germany, June 14 - 16, 2008). SPAA 2008. [59] Fei Wei, Huazhong Yang, Virtual Transmission Method, A New Distributed Algorithm to Solve Sparse Linear Systems. The Fourth International Conference on Networked Computing and Advanced Information Management, vol. 1, pp.703-709, 2008. 14 [60] Fei Wei, Huazhong Yang. Waveform Transmission Method, A New Waveform Relaxation Based Algorithm to Solve Ordinary Differential Equations. Preprinted. 2009. [61] Fei Wei, Huazhong Yang. Transmission Line Inspires a New Distributed Algorithm to Solve the Nonlinear Dynamical System of Physical Circuits. The 5th International Conference on Computer Sciences and Convergence Information Technology, Nov. 30, Seoul, Korea, 2010. [62] Fei Wei, Huazhong Yang. Virtual Transmission Method, Algorithm and Theory. Preprinted. 2009. [63] J. White, A. Sangiovanni-Vincentelli, Relaxation Techniques for the simulation of VLSI circuits. Kluwer Academic Publishers, 1986. [64] Felix F. Wu. Solution of large-scale networks by tearing. IEEE Transactions on Circuits and Systems, 1976. [65] Wei Xue, Jiwu Shu, Yongwei Wu, Weimin Zheng. Parallel Algorithm and Implementation for Realtime Dynamical Simulation of Power System. ICPP 2005: 137-144. [66] D. M. Young. Iterative Solution of Large Linear Systems. Academic Press, New York, 1971. [67] Yu Zhong, M. D. F. Wong, Fast algorithms for IR drop analysis in large power grid, Proceedings of the 2005 IEEE/ACM International conference on Computer-aided design, p.351-357, November 06-10, 2005. [68] Quming Zhou, Kai Sun, Kartik Mohanram, Danny C. Sorensen. Large power grid analysis using domain decomposition, Proceedings of the conference on Design, automation and test in Europe: Proceedings, March 06-10, 2006, Munich, Germany. [69] C. Zhuo, J. Hu, M. Zhao, and K. Chen. Power grid analysis and optimization using algebraic multigrid. IEEE Transactions on Computer-aided Design, 27:738–751, April 2008. Fei Wei received the B.S. degree in electronic engineering from Tsinghua University, Beijing, China in 2004. He is currently a Ph.D student in Department of Electronic Engineering, Tsinghua University. His research interests are distributed numerical algorithms and transistor-level circuit simulation. Huazhong Yang (M’97–SM’00) received the B.S. degree in microelectronics and the M.S. and Ph.D. degrees in circuits and systems from Tsinghua University, Beijing, China, in 1989, 1993, and 1998, respectively. Since 1993, he has been with the Department of Electronic Engineering, Tsinghua University, where he has been a Full Professor since 1998. His research interests include CMOS radio-frequency integrated circuits, VLSI system structure for digital communications and media processing, low-voltage and low power integrated circuits, and computer-aided design methodologies for system integration. He has authored or coauthored 6 books and more than 180 journal and conference papers in this area and holds 9 China patents. He is also a coeditor of the research monograph High-speed Optical Transceivers-Integrated Circuits Designs and Optical Devices Techniques (World Scientific, 2006). Dr. Yang was a recipient of the fund for Distinguished Young Scholars from NSFC in 2000, the outstanding researcher award of the National Keystone Basic Research Program of China in 2004, and the Special Government Allowance from the State Council of China in 2006.served as a TPC member of the Asia-Pacific Conference on Circuits and Systems, the International Conference on Communications, Circuits and Systems, and the Asia and South Pacific Design Automation Conference. He is an Associate Editor of the International Journal of Electronics. 

5cs.CE

arXiv:1708.08062v2 [cs.CV] 18 Oct 2017 Cross-view Asymmetric Metric Learning for Unsupervised Person Re-identification Hong-Xing Yu, Ancong Wu, Wei-Shi Zheng Code is available at the project page: https://github.com/KovenYu/CAMEL For reference of this work, please cite: Hong-Xing Yu, Ancong Wu, Wei-Shi Zheng. “Cross-view Asymmetric Metric Learning for Unsupervised Person Re-identification.” Proceedings of the IEEE International Conference on Computer Vision. 2017. Bib: @inproceedings{yu2017cross, title={Cross-view Asymmetric Metric Learning for Unsupervised Person Re-identification}, author={Yu, Hong-Xing and Wu, Ancong and Zheng, Wei-Shi}, booktitle={Proceedings of the IEEE International Conference on Computer Vision}, year={2017} } Cross-view Asymmetric Metric Learning for Unsupervised Person Re-identification Hong-Xing Yu1,5 , Ancong Wu2 , and Wei-Shi Zheng1,3,4∗ School of Data and Computer Science, Sun Yat-sen University, China 2 School of Electronics and Information Technology, Sun Yat-sen University, China 3 Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China 4 Collaborative Innovation Center of High Performance Computing, NUDT, China 5 Guangdong Key Laboratory of Big Data Analysis and Processing, Guangzhou, China 1 xKoven@gmail.com, wuancong@mail2.sysu.edu.cn, wszheng@ieee.org Abstract While metric learning is important for Person reidentification (RE-ID), a significant problem in visual surveillance for cross-view pedestrian matching, existing metric models for RE-ID are mostly based on supervised learning that requires quantities of labeled samples in all pairs of camera views for training. However, this limits their scalabilities to realistic applications, in which a large amount of data over multiple disjoint camera views is available but not labelled. To overcome the problem, we propose unsupervised asymmetric metric learning for unsupervised RE-ID. Our model aims to learn an asymmetric metric, i.e., specific projection for each view, based on asymmetric clustering on cross-view person images. Our model finds a shared space where view-specific bias is alleviated and thus better matching performance can be achieved. Extensive experiments have been conducted on a baseline and five large-scale RE-ID datasets to demonstrate the effectiveness of the proposed model. Through the comparison, we show that our model works much more suitable for unsupervised RE-ID compared to classical unsupervised metric learning models. We also compare with existing unsupervised REID methods, and our model outperforms them with notable margins. Specifically, we report the results on large-scale unlabelled RE-ID dataset, which is important but unfortunately less concerned in literatures. 1. Introduction Person re-identification (RE-ID) is a challenging problem focusing on pedestrian matching and ranking across non-overlapping camera views. It remains an open problem although it has received considerable exploration recently, in consideration of its potential significance in security ap- plications, especially in the case of video surveillance. It has not been solved yet principally because of the dramatic intra-class variation and the high inter-class similarity. Existing attempts mainly focus on learning to extract robust and discriminative representations [33, 23, 19], and learning matching functions or metrics [38, 14, 18, 22, 19, 20, 26] in a supervised manner. Recently, deep learning has been adopted to RE-ID community [1, 32, 28, 27] and has gained promising results. However, supervised strategies are intrinsically limited due to the requirement of manually labeled cross-view training data, which is very expensive [31]. In the context of RE-ID, the limitation is even pronounced because (1) manually labeling may not be reliable with a huge number of images to be checked across multiple camera views, and more importantly (2) the astronomical cost of time and money is prohibitive to label the overwhelming amount of data across disjoint camera views. Therefore, in reality supervised methods would be restricted when applied to a new scenario with a huge number of unlabeled data. To directly make full use of the cheap and valuable unlabeled data, some existing efforts on exploring unsupervised strategies [8, 35, 29, 13, 21, 24, 30, 12] have been reported, but they are still not very satisfactory. One of the main reasons is that without the help of labeled data, it is rather difficult to model the dramatic variances across camera views, such as the variances of illumination and occlusion conditions. Such variances lead to view-specific interference/bias which can be very disturbing in finding what is more distinguishable in matching people across views (see Figure 1). In particular, existing unsupervised models treat the samples from different views in the same manner, and thus the effects of view-specific bias could be overlooked. In order to better address the problems caused by camera view changes in unsupervised RE-ID scenarios, we pro- Projected by U1 Camera-1 Clustering Projected by U2 Camera-2 Original Space Shared Space Figure 1. Illustration of view-specific interference/bias and our idea. Images from different cameras suffer from view-specific interference, such as occlusions in Camera-1, dull illumination in Camera-2, and the change of viewpoints between them. These factors introduce bias in the original feature space, and therefore unsupervised re-identification is extremely challenging. Our model structures data by clustering and learns view-specific projections jointly, and thus finds a shared space where view-specific bias is alleviated and better performance can be achieved. (Best viewed in color) pose a novel unsupervised RE-ID model named Clusteringbased Asymmetric1 MEtric Learning (CAMEL). The ideas behind are on the two following considerations. First, although conditions can vary among camera views, we assume that there should be some shared space where the data representations are less affected by view-specific bias. By projecting original data into the shared space, the distance between any pair of samples xi and xj is computed as: d(xi , xj ) = kU T xi − U T xj k2 = p (xi − xj )T M (xi − xj ), (1) where U is the transformation matrix and M = U U T . However, it can be hard for a universal transformation to implicitly model the view-specific feature distortion from different camera views, especially when we lack label information to guide it. This motivates us to explicitly model the view-specific bias. Inspired by the supervised asymmetric distance model [4], we propose to embed the asymmetric metric learning to our unsupervised RE-ID modelling, and thus modify the symmetric form in Eq. (1) to an asymmetric one: d(xpi , xqj ) = kU pT xpi − U qT xqj k2 , (2) where p and q are indices of camera views. An asymmetric metric is more acceptable for unsupervised RE-ID scenarios as it explicitly models the variances among views by treating each view differently. By such an explicit means, we are able to better alleviate the disturbances of view-specific bias. The other consideration is that since we are not clear about how to separate similar persons in lack of labeled data, it is reasonable to pay more attention to better separating dissimilar ones. Such consideration motivates us to structure our data by clustering. Therefore, we develop 1 “Asymmetric” means specific transformations for each camera view. asymmetric metric clustering that clusters cross-view person images. By clustering together with asymmetric modelling, the data can be better characterized in the shared space, contributing to better matching performance (see Figure 1). In summary, the proposed CAMEL aims to learn viewspecific projection for each camera view by jointly learning the asymmetric metric and seeking optimal cluster separations. In this way, the data from different views is projected into a shared space where view-specific bias is aligned to an extent, and thus better performance of cross-view matching can be achieved. So far in literatures, the unsupervised RE-ID models have only been evaluated on small datasets which contain only hundreds or a few thousands of images. However, in more realistic scenarios we need evaluations of unsupervised methods on much larger datasets, say, consisting of hundreds of thousands of samples, to validate their scalabilities. In our experiments, we have conducted extensive comparison on datasets with their scales ranging widely. In particular, we combined two existing RE-ID datasets [37, 36] to obtain a larger one which contains over 230,000 samples. Experiments on this dataset (see Sec. 4.4) show empirically that our model is more scalable to problems of larger scales, which is more realistic and more meaningful for unsupervised RE-ID models, while some existing unsupervised RE-ID models are not scalable due to the expensive cost in either storage or computation. 2. Related Work At present, most existing RE-ID models are in a supervised manner. They are mainly based on learning distance metrics or subspace [38, 14, 18, 22, 19, 20, 26], learning view-invariant and discriminative features [33, 23, 19], and deep learning frameworks [1, 32, 28, 27]. However, all these models rely on substantial labeled training data, which is typically required to be pair-wise for each pair of camera views. Their performance depends highly on the quality and quantity of labeled training data. In contrast, our model does not require any labeled data and thus is free from prohibitively high cost of manually labeling and the risk of incorrect labeling. To directly utilize unlabeled data for RE-ID, several unsupervised RE-ID models [35, 29, 21, 13, 24] have been proposed. All these models differ from ours in two aspects. On the one hand, these models do not explicitly exploit the information on view-specific bias, i.e., they treat feature transformation/quantization in every distinct camera view in the same manner when modelling. In contrast, our model tries to learn specific transformation for each camera view, aiming to find a shared space where view-specific interference can be alleviated and thus better performance can be achieved. On the other hand, as for the means to learn a metric or a transformation, existing unsupervised methods for RE-ID rarely consider clustering while we introduce an asymmetric metric clustering to characterize data in the learned space. While the methods proposed in [4, 2, 3] could also learn view-specific mappings, they are supervised methods and more importantly cannot be generalized to handle unsupervised RE-ID. Apart from our model, there have been some clusteringbased metric learning models [34, 25]. However, to our best knowledge, there is no such attempt in RE-ID community before. This is potentially because clustering is more susceptible to view-specific interference and thus data points from the same view are more inclined to be clustered together, instead of those of a specific person across views. Fortunately, by formulating asymmetric learning and further limiting the discrepancy between view-specific transforms, this problem can be alleviated in our model. Therefore, our model is essentially different from these models not only in formulation but also in that our model is able to better deal with cross-view matching problem by treating each view asymmetrically. We will discuss the differences between our model and the existing ones in detail in Sec. 4.3. acceptable to relax the original idea: we focus on gathering similar person images together, and hence separating relatively dissimilar ones. Such goal can be modelled by minimizing an objective function like that of k-means clustering [10]: U 3.1. Problem Formulation Under a conventional RE-ID setting, suppose we have a surveillance camera network that consists of V camera views, from each of which we have collected Np (p = 1, · · · , V ) images and thus there are N = N1 + · · · + NV images in total as training samples. Let X = [x11 , · · · , x1N1 , · · · , xV1 , · · · , xVNV ] ∈ RM ×N denote the training set, with each column xpi (i = 1, · · · , Np ; p = 1, · · · , V ) corresponding to an M dimensional representation of the i-th image from the pth camera view. Our goal is to learn V mappings i.e., U 1 , · · · , U V , where U p ∈ RM ×T (p = 1, · · · , V ), corresponding to each camera view, and thus we can project the original representation xpi from the original space RM into a shared space RT in order to alleviate the view-specific interference. 3.2. Modelling Now we are looking for some transformations to map our data into a shared space where we can better separate the images of one person from those of different persons. Naturally, this goal can be achieved by narrowing intraclass discrepancy and meanwhile pulling the centers of all classes away from each other. In an unsupervised scenario, however, we have no labeled data to tell our model how it can exactly distinguish one person from another who has a confusingly similar appearance with him. Therefore, it is kU T xi − ck k2 , (3) k=1 i∈Ck where K is the number of clusters, ck denotes the centroid of the k-th cluster and Ck = {i|U T xi ∈ k-th cluster}. However, clustering results may be affected extremely by view-specific bias when applied in cross-view problems. In the context of RE-ID, the feature distortion could be view-sensitive due to view-specific interference like different lighting conditions and occlusions [4]. Such interference might be disturbing or even dominating in searching the similar person images across views during clustering procedure. To address this cross-view problem, we learn specific projection for each view rather than a universal one to explicitly model the effect of view-specific interference and to alleviate it. Therefore, the idea can be further formulated by minimizing an objective function below: min 3. Methodology K X X min Fintra = U 1 ,··· ,U V s.t. Fintra = K X X kU pT xpi − ck k2 k=1 i∈Ck U pT p p Σ U =I (4) (p = 1, · · · , V ), where the notation is similar to Eq. (3), with p denotes the view index, Σp = X p X pT /Np + αI and I represents the identity matrix which avoids singularity of the covariance matrix. The transformation U p that corresponds to each instance xpi is determined by the camera view which xpi comes from. The quasi-orthogonal constraints on U p ensure that the model will not simply give zero matrices. By combining the asymmetric metric learning, we actually realize an asymmetric metric clustering on RE-ID data across camera views. Intuitively, if we minimize this objective function directly, U p will largely depend on the data distribution from the p-th view. Now that there is specific bias on each view, any U p and U q could be arbitrarily different. This result is very natural, but large inconsistencies among the learned transformations are not what we exactly expect, because the transformations are with respect to person images from different views: they are inherently correlated and homogeneous. More critically, largely different projection basis pairs would fail to capture the discriminative nature of cross-view images, producing an even worse matching result. Hence, to strike a balance between the ability to capture discriminative nature and the capability to alleviate viewspecific bias, we embed a cross-view consistency regularization term into our objective function. And then, in consideration of better tractability, we divide the intra-class 0.8 1 0.5 View1, ID1 View2, ID1 View1, ID2 View2, ID2 0.6 0.4 where 1 View1, ID1 View2, ID1 View1, ID2 View2, ID2 0.5 0.2 0 0   H = h1 , ..., hK , 0 hT k hl = -0.2 -0.5 -0.4 View1, ID1 View2, ID1 View1, ID2 View2, ID2 -0.5 -0.6 -1 -1.5 -1 -0.5 0 0.5 -0.8 1.5 -1.5 1 (a) Original -1 -0.5 0 0.5 1 -1 -0.8 -0.6 (b) Symmetric -0.4 -0.2 0 0.2 0.4 0.6 Figure 2. Illustration of how symmetric and asymmetric metric clustering structure data using our method for the unsupervised RE-ID problem. The samples are from the SYSU dataset [4]. We performed PCA for visualization. One shape (triangle or circle) stands for samples from one view, while one color indicates samples of one person. (a) Original distribution (b) distribution in the common space learned by symmetric metric clustering (c) distribution in the shared space learned by asymmetric metric clustering. (Best viewed in color) term by its scale N , so that the regulating parameter would not be sensitive to the number of training samples. Thus, our optimization task becomes min U 1 ,··· ,U V Fobj =  1 x1 0  f=  . X  .. 0 s.t. U p p Σ U =I k=1 k=1 i∈Ck     (9) xVNV (10) (11) (12) k As for the second term, we can also rewrite it as follow: For convenience, we denote yi = U pT xpi . Then we have Y ∈ RT ×N , where each column yi corresponds to the projected new representation of that from X. For optimization, we rewrite our objective function in a more compact form. The first term can be rewritten as follow [6]: kyi − ck k2 =  1 fT U eU e T X) f − 1 Tr(H T X fT U eU e T XH). f = Tr(X N N (p = 1, · · · , V ), 3.3. Optimization X 0 0 .. . K 1 XX kyi − ck k2 N i∈C where λ is the cross-view regularizer and k·kF denotes the Frobenius norm of a matrix. We call the above model the Clustering-based Asymmetric MEtric Learning (CAMEL). To illustrate the differences between symmetric and asymmetric metric clustering in structuring data in the REID problem, we further show the data distributions in Figure 2. We can observe from Figure 2 that the view-specific bias is obvious in the original space: triangles in the upper left and circles in the lower right. In the common space learned by symmetric metric clustering, the bias is still obvious. In contrast, in the shared space learned by asymmetric metric clustering, the bias is alleviated and thus the data is better characterized according to the identities of the persons, i.e., samples of one person (one color) gather together into a cluster. 1 N ··· ··· .. . ··· and thus Eq. (6) becomes (5) K X x1N1 0 · · · 0 0 x21 · · · x2N2 .. .. .. .. . . . . 0 0 ··· 0   e = U 1T , · · · , U V T T , U e T X, f Y =U p6=q k ··· ··· .. . ··· so that K X p 1 XX = kU pT xpi − ck k2 + λ kU − U q k2F N i∈C pT (8) is an indicator vector with the i-th entry corresponding to the instance yi , indicating that yi is in the k-th cluster if the corresponding entry does not equal zero. Then we construct 1 Fintra + λFconsistency N k=1 (7)  T √ hk = 0, · · · , 0, 1, · · · , 1, 0, · · · , 0, 1, · · · / nk 0.8 (c) Asymmetric ( 0 k 6= l 1 k=l 1 [Tr(Y T Y ) − Tr(H T Y T Y H)], N (6) λ X p e T DU e ), kU − U q k2F = λTr(U (13) p6=q where  (V − 1)I  −I  D= ..  . −I −I (V − 1)I .. . −I −I −I .. . −I ··· ··· .. . ···  −I −I    . (14) ..  . (V − 1)I Then, it is reasonable to relax the constraints U pT Σp U p = I (p = 1, · · · , V ) (15) e TΣ eU e = V I, U pT Σp U p = U (16) to V X p=1 e = diag(Σ1 , · · · , ΣV ) because what we expect where Σ is to prevent each U p from shrinking to a zero matrix. The relaxed version of constraints is able to satisfy our need, and it bypasses trivial computations. By now we can rewrite our optimization task as follow: 1 fT U eU e T X) f + λTr(U e T DU e) Tr(X N 1 fT U eU e T XH) f − Tr(H T X N e TΣ eU e = V I. s.t. U min Fobj = e U (17) It is easy to realize from Eq. (5) that our objective function is highly non-linear and non-convex. Fortunately, in the form of Eq. (17) we can find that once H is fixed, Lagrange’s method can be applied to our optimization task. And again from Eq. (5), it is exactly the objective of ke is fixed [10]. Thus, we can adopt means clustering once U an alternating algorithm to solve the optimization problem. e . Now we see how we optimize Fix H and optimize U e U . After fixing H and applying the method of Lagrange multiplier, our optimization task (17) is transformed into an eigen-decomposition problem as follow: Gu = γu, (18) Algorithm 1: CAMEL 1 2 3 e K,  = 10−8 Input : X, e Output: U e to initialize Conduct k-means clustering with respect to each column of X H according to Eq. (7) and (8). Fix H and solve the eigen-decomposition problem described by Eq. (18) and e. (19) to construct U while decrement of Fobj >  & maximum iteration unreached do • Construct Y according to Eq. (11). e and conduct k-means clustering with respect to each column • Fix U of Y to update H according to Eq. (7) and (8). • Fix H and solve the eigen-decomposition problem described by e. Eq. (18) and (19) to update U 4 end where γ is the Lagrange multiplier (and also is the eigenvalue here) and T fT fX fT − 1 XHH f e −1 (λD + 1 X X ). G=Σ N N (19) e can be obtained by solving this eigenThen, U decomposition problem. e and optimize H. As for the optimization of H, Fix U e and conduct k-means clustering in the we can simply fix U learned space. Each column of H, hk , is thus constructed according to the clustering result. Based on the analysis above, we can now propose the main algorithm of CAMEL in Algorithm 1. We set maxie , we decompose it mum iteration to 100. After obtaining U 1 V back into {U , · · · , U }. The algorithm is guaranteed to convergence, as given in the following proposition: Proposition 1. In Algorithm 1, Fobj is guaranteed to convergence. e is fixed, if H is the loProof. In each iteration, when U cal minimizer, k-means remains H unchanged, otherwise e has a it seeks the local minimizer. When H is fixed, U closed-form solution which is the global minimizer. Therefore, the Fobj decreases step by step. As Fobj ≥ 0 has a lower bound 0, it is guaranteed to convergence. 4. Experiments 4.1. Datasets Since unsupervised models are more meaningful when the scale of problem is larger, our experiments were conducted on relatively big datasets except VIPeR [9] which is small but widely used. Various degrees of view-specific bias can be observed in all these datasets (see Figure 3). The VIPeR dataset contains 632 identities, with two images captured from two camera views of each identity. The CUHK01 dataset [16] contains 3,884 images of 971 identities captured from two disjoint views. There are two images of every identity from each view. (a) (b) (c) (d) (e) (f) Figure 3. Samples of the datasets. Every two images in a column are from one identity across two disjoint camera views. (a) VIPeR (b) CUHK01 (c) CUHK03 (d) SYSU (e) Market (f) ExMarket. (Best viewed in color) Dataset VIPeR CUHK01 CUHK03 SYSU Market ExMarket # Samples # Views 1,264 2 3,884 2 24,448 2 32,668 6 236,696 6 13,164 6 Table 1. Overview of dataset scales. “#” means “the number of”. The CUHK03 dataset [17] contains 13,164 images of 1,360 pedestrians captured from six surveillance camera views. Besides hand-cropped images, samples detected by a state-of-the-art pedestrian detector are provided. The SYSU dataset [4] includes 24,448 RGB images of 502 persons under two surveillance cameras. One camera view mainly captured the frontal or back views of persons, while the other observed mostly the side views. The Market-1501 dataset [37] (Market) contains 32,668 images of 1,501 pedestrians, each of which was captured by at most six cameras. All of the images were cropped by a pedestrian detector. There are some bad-detected samples in this datasets as distractors as well. The ExMarket dataset2 . In order to evaluate unsupervised RE-ID methods on even larger scale, which is more realistic, we further combined the MARS dataset [36] with Market. MARS is a video-based RE-ID dataset which con2 Demo code for the model and the ExMarket dataset can be found on https://github.com/KovenYu/CAMEL. tains 20,715 tracklets of 1,261 pedestrians. All the identities from MARS are of a subset of those from Market. We then took 20% frames (each one in every five successive frames) from the tracklets and combined them with Market to obtain an extended version of Market (ExMarket). The imbalance between the numbers of samples from the 1,261 persons and other 240 persons makes this dataset more challenging and realistic. There are 236,696 images in ExMarket in total, and 112,351 images of them are of training set. A brief overview of the dataset scales can be found in Table 1. 4.2. Settings Experimental protocols: A widely adopted protocol was followed on VIPeR in our experiments [19], i.e., randomly dividing the 632 pairs of images into two halves, one of which was used as training set and the other as testing set. This procedure was repeated 10 times to offer average performance. Only single-shot experiments were conducted. The experimental protocol for CUHK01 was the same as that in [19]. We randomly selected 485 persons as training set and the other 486 ones as testing set. The evaluating procedure was repeated 10 times. Both multi-shot and single-shot settings were conducted. The CUHK03 dataset was provided together with its recommended evaluating protocol [17]. We followed the provided protocol, where images of 1,160 persons were chosen as training set, images of another 100 persons as validation set and the remainders as testing set. This procedure was repeated 20 times. In our experiments, detected samples were adopted since they are closer to real-world settings. Both multi-shot and single-shot experiments were conducted. As for the SYSU dataset, we randomly picked 251 pedestrians’ images as training set and the others as testing set. In the testing stage, we basically followed the protocol as in [4]. That is, we randomly chose one and three images of each pedestrian as gallery for single-shot and multi-shot experiments, respectively. We repeated the testing procedure by 10 times. Market is somewhat different from others. The evaluation protocol was also provided along with the data [37]. Since the images of one person came from at most six views, single-shot experiments were not suitable. Instead, multi-shot experiments were conducted and both cumulative matching characteristic (CMC) and mean average precision (MAP) were adopted for evaluation [37]. The protocol of ExMarket was identical to that of Market since the identities were completely the same as we mentioned above. Data representation: In our experiments we used the deeplearning-based JSTL feature proposed in [32]. We implemented it using the 56-layer ResNet [11], which produced 64-D features. The original JSTL was adopted to our implementation to extract features on SYSU, Market and ExMarket. Note that the training set of the original JSTL contained VIPeR, CUHK01 and CUHK03, violating the unsupervised setting. So we trained a new JSTL model without VIPeR in its training set to extract features on VIPeR. The similar procedures were done for CUHK01 and CUHK03. Parameters: We set λ, the cross-view consistency regularizer, to 0.01. We also evaluated the situation when λ goes to infinite, i.e., the symmetric version of our model in Sec. 4.4, to show how important the asymmetric modelling is. Regarding the parameter T which is the feature dimension after the transformation learned by CAMEL, we set T equal to original feature dimension i.e., 64, for simplicity. In our experiments, we found that CAMEL can align data distributions across camera views even without performing any further dimension reduction. This may be due to the fact that, unlike conventional subspace learning models, the transformations learned by CAMEL are view-specific for different camera views and always non-orthogonal. Hence, the learned view-specific transformations can already reduce the discrepancy between the data distributions of different camera views. As for K, we found that our model was not sensitive to K when N  K and K was not too small (see Sec. 4.4), so we set K = 500. These parameters were fixed for all datasets. 4.3. Comparison Unsupervised models are more significant when applied on larger datasets. In order to make comprehensive and fair comparisons, in this section we compare CAMEL with the most comparable unsupervised models on six datasets with their scale orders varying from hundreds to hundreds of thousands. We show the comparative results measured by the rank-1 accuracies of CMC and MAP (%) in Table 2. Comparison to Related Unsupervised RE-ID Models. In this subsection we compare CAMEL with the sparse dictionary learning model (denoted as Dic) [13], sparse representation learning model ISR [21], kernel subspace learning model RKSL [30] and sparse auto-encoder (SAE) [15, 5]. We tried several sets of parameters for them, and report the best ones. We also adopt the Euclidean distance which is adopted in the original JSTL paper [32] as a baseline (denoted as JSTL). From Table 2 we can observe that CAMEL outperforms other models on all the datasets on both settings. In addition, we can further see from Figure 4 that CAMEL outperforms other models at any rank. One of the main reasons is that the view-specific interference is noticeable in these datasets. For example, we can see in Figure 3(b) that on CUHK01, the changes of illumination are extremely severe and even human beings may have difficulties in recognizing the identities in those images across views. This impedes other symmetric models from achieving higher accuracies, because they potentially hold an assumption that the invari- SS Dic [13] ISR [21] RKSL [30] SAE [15] JSTL [32] 29.9 27.5 25.8 20.7 25.7 CUHK01 CUHK03 SYSU Market ExMarket SS/MS SS/MS MS MS 49.3/52.9 53.2/55.7 45.4/50.1 45.3/49.9 46.3/50.6 SS/MS 27.4/36.5 31.1/38.5 25.8/34.8 21.2/30.5 24.7/33.2 21.3/28.6 23.2/33.8 17.6/23.0 18.0/24.2 19.9/25.6 50.2(22.7) 52.2(21.2) 40.3(14.3) 34.0(11.0) 42.4(16.2) 44.0(15.1) 44.7(18.4) 46.4(16.7) 95 70 90 85 60 Matching Accuracy (%) VIPeR Setting Matching Accuracy (%) Dataset DIC ISR UsNCA AML SAE RKSL JSTL CAMEL 50 40 30 46.8/51.1 22.2/31.4 20.9/26.4 44.7(18.4) 46.2(16.2) 47.0/51.7 19.8/29.6 21.1/27.2 45.2(18.9) - CAMEL 30.9 57.3/61.9 31.9/39.4 30.8/36.8 54.5(26.3) 55.9(23.9) Table 2. Comparative results of unsupervised models on the six datasets, measured by rank-1 accuracies and MAP (%). “-” means prohibitive time consumption due to time complexities of the models. “SS” represents single-shot setting and “MS” represents multi-shot setting. For Market and ExMarket, MAP is also provided in the parentheses due to the requirement in the protocol [37]. Such a format is also applied in the following tables. 15 70 65 60 50 45 20 5 (a) VIPeR 15 20 80 70 70 60 60 DIC ISR UsNCA AML SAE RKSL JSTL CAMEL 50 40 30 DIC ISR UsNCA AML SAE RKSL JSTL CAMEL 50 40 30 20 20 5 10 Rank 15 5 20 (c) CUHK03 10 Rank 15 20 (d) SYSU 85 80 80 75 Matching Accuracy (%) 75 Matching Accuracy (%) ant and discriminative information can be retained and exploited through a universal transformation for all views. But CAMEL relaxes this assumption by learning an asymmetric metric and then can outperform other models significantly. In Sec. 4.4 we will see the performance of CAMEL would drop much when it degrades to a symmetric model. Comparison to Clustering-based Metric Learning Models. In this subsection we compare CAMEL with a typical model AML [34] and a recently proposed model UsNCA [25]. We can see from Fig. 4 and Table 2 that compared to them, CAMEL achieves noticeable improvements on all the six datasets. One of the major reasons is that they do not consider the view-specific bias which can be very disturbing in clustering, making them unsuitable for RE-ID problem. In comparison, CAMEL alleviates such disturbances by asymmetrically modelling. This factor contributes to the much better performance of CAMEL. Comparison to the State-of-the-Art. In the last subsections, we compared with existing unsupervised RE-ID methods using the same features. In this part, we also compare with the results reported in literatures. Note that most existing unsupervised RE-ID methods have not been evaluated on large datasets like CUHK03, SYSU, or Market, so Table 3 only reports the comparative results on VIPeR and CUHK01. We additionally compared existing unsupervised RE-ID models, including the hand-craft-feature-based SDALF [8] and CPS [7], the transfer-learning-based UDML [24], graph-learning-based model (denoted as GL) [12], and local-salience-learning-based GTS [29] and SDC [35]. We can observe from Table 3 that our model CAMEL can outperform the state-of-the-art by large margins on CUHK01. Comparison to Supervised Models. Finally, in order to see how well CAMEL can approximate the performance of supervised RE-ID, we additionally compare CAMEL with its supervised version (denoted as CAMELs ) which is easily 10 Rank (b) CUHK01 Matching Accuracy (%) 23.1 24.3 Matching Accuracy (%) AML [34] UsNCA [25] 10 Rank DIC ISR UsNCA AML SAE RKSL JSTL CAMEL 75 55 20 5 80 70 DIC ISR UsNCA AML SAE RKSL JSTL CAMEL 65 60 55 50 45 70 65 60 DIC AML SAE JSTL CAMEL 55 50 45 40 5 10 Rank 15 5 20 (e) Market 10 Rank 15 20 (f) ExMarket Figure 4. CMC curves. For CUHK01, CUHK03 and SYSU, we take the results under single-shot setting as examples. Similar patterns can be observed on multi-shot setting. Model SDALF [8] CPS [7] UDML [24] GL [12] GTS [29] SDC [35] CAMEL VIPeR CUHK01 19.9 9.9 22.0 - 31.5 27.1 33.5 41.0 25.2 - 25.8 26.6 30.9 57.3 Table 3. Results compared to the state-of-the-art reported in literatures, measured by rank-1 accuracies (%). “-” means no reported result. derived by substituting the clustering results by true labels, and three standard supervised models, including the widely used KISSME [14], XQDA [19], the asymmetric distance model CVDCA [4]. The results are shown in Table 4. We can see that CAMELs outperforms CAMEL by various degrees, indicating that label information can further improve CAMEL’s performance. Also from Table 4, we notice that CAMEL can be comparable to other standard supervised models on some datasets like CUHK01, and even outperform some of them. It is probably because the used JSTL model had not been fine-tuned on the target datasets: this was for a fair comparison with unsupervised models which work on completely unlabelled training data. Nevertheless, this suggests that the performance of CAMEL may not be Dataset SYSU Market ExMarket Dataset Setting VIPeR CUHK01 CUHK03 SS SS/MS SS/MS SS/MS MS MS Setting VIPeR CUHK01 CUHK03 SS KISSME [14] XQDA [19] CVDCA [4] CAMELs 28.4 28.9 37.6 33.7 53.0/57.1 54.3/58.2 57.1/60.9 58.5/62.7 37.8/45.4 36.7/43.7 37.0/44.6 45.1/53.5 24.7/31.8 25.2/31.7 31.1/38.9 31.6/37.6 51.1(24.5) 50.8(24.4) 52.6(25.3) 55.0(27.1) 48.0(18.3) 47.4(18.1) 51.5(22.6) 56.1(24.1) CMEL CAMEL 27.5 30.9 CAMEL 30.9 57.3/61.9 31.9/39.4 30.8/36.8 54.5(26.3) 55.9(23.9) SS/MS SS/MS SYSU Market ExMarket SS/MS MS MS 52.5/54.9 29.8/37.5 25.4/30.9 47.6(21.5) 48.7(20.0) 57.3/61.9 31.9/39.4 30.8/36.8 54.5(26.3) 55.9(23.9) Table 5. Performances of CAMEL compared to its symmetric version, denoted as CMEL. Table 4. Results compared to supervised models using the same JSTL features. far below the standard supervised RE-ID models. 4.4. Further Evaluations The Role of Asymmetric Modeling. We show what is going to happen if CAMEL degrades to a common symmetric model in Table 5. Apparently, without asymmetrically modelling each camera view, our model would be worsen largely, indicating that the asymmetric modeling for clustering is rather important for addressing the cross-view matching problem in RE-ID as well as in our model. Sensitivity to the Number of Clustering Centroids. We take CUHK01, Market and ExMarket datasets as examples of different scales (see Table 1) for this evaluation. Table 6 shows how the performance varies with different numbers of clustering centroids, K. It is obvious that the performance only fluctuates mildly when N  K and K is not too small. Therefore CAMEL is not very sensitive to K especially when applied to large-scale problems. To further explore the reason behind, we show in Table 7 the rate of clusters which contain more than one persons, in the initial stage and convergence stage in Algorithm 1. We can see that (1) in spite of that K is varying, there is always a number of clusters containing more than one persons in both the initial stage and convergence stage. This indicates that our model works without the requirement of perfect clustering results. And (2), although the number is various, in the convergence stage the number is consistently decreased compared to initialization stage. This shows that the cluster results are improved consistently. These two observations suggests that the clustering should be a mean to learn the asymmetric metric, rather than an ultimate objective. Adaptation Ability to Different Features. At last, we show that CAMEL can be effective not only when adopting deep-learning-based JSTL features. We additionally adopted the hand-crafted LOMO feature proposed in [19]. We performed PCA to produce 512-D LOMO features, and the results are shown in Table 8. Among all the models, the results of Dic and ISR are the most comparable (Dic and ISR take all second places). So for clarity, we only compare CAMEL with them and L2 distance as baseline. From the table we can see that CAMEL can outperform them. K 250 500 750 1000 1250 CUHK01 Market ExMarket 56.59 54.48 55.49 57.35 54.45 55.87 56.26 54.54 56.17 55.12 54.48 55.93 52.75 54.48 55.67 Table 6. Performances of CAMEL when the number of clusters, K, varies. Measured by single-shot rank-1 accuracies (%) for CUHK01 and multi-shot for Market and ExMarket. K 250 500 750 1000 1250 Initial Stage Convergence Stage 77.6% 55.8% 57.0% 34.3% 26.3% 18.2% 11.6% 7.2% 6.0% 4.8% Table 7. Rate of clusters containing similar persons on CUHK01. Similar trend can be observed on other datasets. Dataset VIPeR Setting SS Dic [13] ISR [21] L2 15.8 20.8 11.6 19.6/23.6 8.6/13.4 14.2/24.4 32.8(12.2) 33.8(12.2) 22.2/27.1 16.7/20.7 11.7/21.6 29.7(11.0) 14.0/18.6 7.6/11.6 10.8/18.9 27.4(8.3) 27.7(8.0) 26.4 30.0/36.2 17.3/23.4 23.6/35.6 41.4(14.1) 42.2(13.7) CAMEL CUHK01 CUHK03 SYSU Market ExMarket SS/MS SS/MS MS MS SS/MS Table 8. Results using 512-D LOMO features. 5. Conclusion In this work, we have shown that metric learning can be effective for unsupervised RE-ID by proposing clusteringbased asymmetric metric learning called CAMEL. CAMEL learns view-specific projections to deal with view-specific interference, and this is based on existing clustering (e.g., the k-means model demonstrated in this work) on RE-ID unlabelled data, resulting in an asymmetric metric clustering. Extensive experiments show that our model can outperform existing ones in general, especially on large-scale unlabelled RE-ID datasets. Acknowledgement This work was supported partially by the National Key Research and Development Program of China (2016YFB1001002), NSFC(61522115, 61472456, 61573387, 61661130157, U1611461), the Royal Society Newton Advanced Fellowship (NA150459), Guangdong Province Science and Technology Innovation Leading Talents (2016TX03X157). References [1] E. Ahmed, M. Jones, and T. K. Marks. An improved deep learning architecture for person re-identification. In CVPR, 2015. [2] L. An, M. Kafai, S. Yang, and B. Bhanu. Reference-based person re-identification. In AVSS, 2013. [3] L. An, M. Kafai, S. Yang, and B. Bhanu. Person reidentification with reference descriptor. TCSVT, 2015. [4] Y.-C. Chen, W.-S. Zheng, J.-H. Lai, and P. Yuen. An asymmetric distance model for cross-view feature mapping in person re-identification. TCSVT, 2015. [5] A. Coates, H. Lee, and A. Y. Ng. An analysis of single-layer networks in unsupervised feature learning. Ann Arbor, 2010. [6] C. H. Q. Ding and X. He. On the equivalence of nonnegative matrix factorization and spectral clustering. In ICDM, 2005. [7] S. C. Dong, M. Cristani, M. Stoppa, L. Bazzani, and V. Murino. Custom pictorial structures for re-identification. In BMVC, 2011. [8] M. Farenzena, L. Bazzani, A. Perina, V. Murino, and M. Cristani. Person re-identification by symmetry-driven accumulation of local features. In CVPR, 2010. [9] D. Gray, S. Brennan, and H. Tao. Evaluating appearance models for recognition, reacquisition, and tracking. In PETS, 2007. [10] J. A. Hartigan. Clustering algorithms. John Wiley & Sons Inc, 1975. [11] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, 2016. [12] E. Kodirov, T. Xiang, Z. Fu, and S. Gong. Person reidentification by unsupervised\ ell 1 graph learning. In ECCV, 2016. [13] E. Kodirov, T. Xiang, and S. Gong. Dictionary learning with iterative laplacian regularisation for unsupervised person reidentification. In BMVC, 2015. [14] M. Köstinger, M. Hirzer, P. Wohlhart, P. M. Roth, and H. Bischof. Large scale metric learning from equivalence constraints. In CVPR, 2012. [15] H. Lee, C. Ekanadham, and A. Y. Ng. Sparse deep belief net model for visual area v2. In NIPS, 2008. [16] W. Li, R. Zhao, and X. Wang. Human reidentification with transferred metric learning. In ACCV, 2012. [17] W. Li, R. Zhao, T. Xiao, and X. Wang. Deepreid: Deep filter pairing neural network for person re-identification. In CVPR, 2014. [18] Z. Li, S. Chang, F. Liang, T. S. Huang, L. Cao, and J. R. Smith. Learning locally-adaptive decision functions for person verification. In CVPR, 2013. [19] S. Liao, Y. Hu, X. Zhu, and S. Z. Li. Person re-identification by local maximal occurrence representation and metric learning. In CVPR, 2015. [20] S. Liao and S. Z. Li. Efficient psd constrained asymmetric metric learning for person re-identification. In ICCV, 2015. [21] G. Lisanti, I. Masi, A. D. Bagdanov, and A. Del Bimbo. Person re-identification by iterative re-weighted sparse ranking. TPAMI, 2015. [22] G. Lisanti, I. Masi, and A. Del Bimbo. Matching people across camera views using kernel canonical correlation analysis. In ICDSC, 2014. [23] B. Ma, Y. Su, and F. Jurie. Covariance descriptor based on bio-inspired features for person re-identification and face verification. IVC, 2014. [24] P. Peng, T. Xiang, Y. Wang, M. Pontil, S. Gong, T. Huang, and Y. Tian. Unsupervised cross-dataset transfer learning for person re-identification. In CVPR, 2016. [25] C. Qin, S. Song, G. Huang, and L. Zhu. Unsupervised neighborhood component analysis for clustering. Neurocomputing, 2015. [26] Y. Shen, W. Lin, J. Yan, M. Xu, J. Wu, and J. Wang. Person re-identification with correspondence structure learning. In ICCV, 2015. [27] R. R. Varior, M. Haloi, and G. Wang. Gated siamese convolutional neural network architecture for human reidentification. In ECCV, 2016. [28] F. Wang, W. Zuo, L. Lin, D. Zhang, and L. Zhang. Joint learning of single-image and cross-image representations for person re-identification. In CVPR, 2016. [29] H. Wang, S. Gong, and T. Xiang. Unsupervised learning of generative topic saliency for person re-identification. In BMVC, 2014. [30] H. Wang, X. Zhu, T. Xiang, and S. Gong. Towards unsupervised open-set person re-identification. In ICIP, 2016. [31] X. Wang, W. S. Zheng, X. Li, and J. Zhang. Cross-scenario transfer person reidentification. TCSVT, 2015. [32] T. Xiao, H. Li, W. Ouyang, and X. Wang. Learning deep feature representations with domain guided dropout for person re-identification. In CVPR, 2016. [33] Y. Yang, J. Yang, J. Yan, S. Liao, D. Yi, and S. Z. Li. Salient color names for person re-identification. In ECCV, 2014. [34] J. Ye, Z. Zhao, and H. Liu. Adaptive distance metric learning for clustering. In CVPR, 2007. [35] R. Zhao, W. Ouyang, and X. Wang. Person re-identification by saliency learning. TPAMI, 2016. [36] L. Zheng, Z. Bie, Y. Sun, J. Wang, C. Su, S. Wang, and Q. Tian. Mars: A video benchmark for large-scale person re-identification. In ECCV, 2016. [37] L. Zheng, L. Shen, L. Tian, S. Wang, J. Wang, and Q. Tian. Scalable person re-identification: A benchmark. In ICCV, 2015. [38] W.-S. Zheng, S. Gong, and T. Xiang. Person re-identification by probabilistic relative distance comparison. In CVPR, 2011. 

1cs.CV

PRT (Personal Rapid Transit) Network Simulation CHOROMAŃSKI, Włodzimierz; DASZCZUK, Wiktor; DYDUCH, Jarosław; MACIEJEWSKI, Mariusz; BRACH, Paweł; GRABSKI, Waldemar PRT (PERSONAL RAPID TRANSIT) NETWORK SIMULATION Włodzimierz Choromański, Warsaw University of Technology, Faculty of Transport, prof.wch@gmail.com Wiktor B. Daszczuk, Warsaw University of Technology, Institute of Computer Science, wbd@ii.pw.edu.pl Jarosław Dyduch, Warsaw University of Technology, Faculty of Transport, slavypl@gmail.com Mariusz Maciejewski, Warsaw University of Technology, Faculty of Transport, Mariusz.Maciejewski@kontron.com.pl Paweł Brach, Warsaw University of Technology, Faculty of Transport, braszek@gmail.com Waldemar Grabski, Warsaw University of Technology, Institute of Computer Science, wgr@ii.pw.edu.pl ABSTRACT Transportation problems of large urban conurbations inspire search for new transportation systems, that meet high environmental standards, are relatively cheap and user friendly. The latter element also includes the needs of disabled and elderly people. This article concerns a new transportation system PRT - Personal Rapid Transit. In this article the attention is focused on the analysis of the efficiency of the PRT transport network. The simulator of vehicle movement in PRT network as well as algorithms for traffic management and control will be presented. The proposal of its physical implementation will be also included. Keywords: PRT, simulation, traffic experiments INTRODUCTION In Warsaw University of Technology, a project on Personal Rapid Transit (PRT) (Irving at al. 1978, Andreasson 2010) is under development (Choromański et al. 2011a and 2011b, Daszczuk at al. 2011). Personal rapid transit (PRT), is a public transportation mode featuring small automated vehicles operating on a network of specially-built guide ways. PRT is a type of th 13 WCTR, July 15-18, 2013 – Rio de Janeiro, Brazil PRT (Personal Rapid Transit) Network Simulation CHOROMAŃSKI, Włodzimierz; DASZCZUK, Wiktor; DYDUCH, Jarosław; MACIEJEWSKI, Mariusz; BRACH, Paweł; GRABSKI, Waldemar automated guideway transit (AGT), a class of system which also includes larger vehicles all the way to small driverless subway systems. The whole system is electrically supplied. In PRT designs, vehicles are sized for individual or small group travel, typically carrying no more than 3 to 4 passengers per vehicle. Guide ways are arranged in a network topology, with all stations located on sidings, and with frequent merge/diverge points. This approach allows for nonstop, point-to-point travel, bypassing all intermediate stations. The point-to-point service has been compared to a taxi. There are three kinds of nodes in the PRT network: stations, capacitors and intersections. Stations are places where passengers book their trips and board the vehicles, or wait for vehicles in a queue if there are no empty vehicles on a station. A capacitor is a source of vehicles (and sometimes may serve as a parking place). Intersections are threefold: “fork” (diverge), “join” (merge) and “junction” (such intersections are for technical purposes only). It is assumed that a vehicle has its own control unit, which is linked via radio network with control units of other vehicles and nodes (capacitors, stations and “join” intersections). Radio connections are established to vehicles and nodes that are closest to the vehicle, i.e. not farer that specified distance. This makes a subnet “visible” to the vehicle, the edge of which is called “a horizon”. The horizon distance should be chosen carefully, too large causes too much information to be transmitted (and routing problems), too small results in reduced safety of the traffic. A vehicle gets information about current parameters of movement of preceding vehicles: their positions, velocities and mode of operation (acceleration/ constant velocity/ deceleration/ friction braking). From an intersection controller a vehicle receives the decision on priority of crossing the intersection. Among other mechanical, electrical and transportation research goals, simulation of PRT network is performed, on two abstraction levels:  Coordination level is “behavioural simulation”. On this level, algorithms for following a route, keeping up, coordination on “join” intersections, joining the traffic and similar are tested for effectiveness.  Management level is “statistical simulation”. Simulation experiments identify the impact of various parameters of management algorithms (mainly of empty vehicle management and dynamic routing) on the passenger comfort (trip time and queue size). There are several PRT simulators available (Castangia and Guala 2011, Zheng, Jeffery and McDonald 2009, Andreasson 2010, Hermes, Beamways, RUF), yet the authors have decided to build the project’s own simulators, for two reasons:  Available simulators have fixed, inaccessible traffic algorithms (on both levels: coordination and management); the user can only observe the operation of included algorithms and he/she has no influence on their structure and parameters.  Simulators produce a number of synthetic output parameters; there is no access to individual events (witch should be reported as a kind of “event log”); this lack does not allow the researcher to build his/her own detailed characteristics of the traffic. It seems that the own simulators build are more advanced. They permit the analysis of various algorithms of coordination and management, and they give access to wide and detailed set of traffic parameters and event logs. The purpose of simulators is: th 13 WCTR, July 15-18, 2013 – Rio de Janeiro, Brazil PRT (Personal Rapid Transit) Network Simulation CHOROMAŃSKI, Włodzimierz; DASZCZUK, Wiktor; DYDUCH, Jarosław; MACIEJEWSKI, Mariusz; BRACH, Paweł; GRABSKI, Waldemar Entry buffer Exit buffer Debarking berth Other berths Boarding berth a) in-line Entry buffer Exit buffer Stubberths b) stub-berths Figure 1 – station types      analysis of various network topologies of PRT, identification of capacity and ridership of various networks, identification of saturation points (when performance is degraded to inaccessible level) research on dynamic algorithms (especially empty vehicles management and dynamic routing), and identification the operational cost of performance improvement (number of empty trips, distance of empty movement etc) sensitivity of the system to variation of different traffic parameters values. THE STRUCTURE OF THE PRT NETWORK A structure of PRT network is a graph (Andreasson 2010). Nodes are capacitors, stations and intersections. Edges are segments. There are two types of stations: in-line (with the FIFO rule applied) and with stubberths. Vehicles must have ability to drive backwards to operate inside a berth. A structure of both types of a station is shown in Figure 1. From topography point of view, nodes are simply points of size zero. Yet the nodes have internal structure necessary to model their roles in a network. The stations are most complicated: they have parking places (berths), entry and exit buffers, passenger queues th 13 WCTR, July 15-18, 2013 – Rio de Janeiro, Brazil PRT (Personal Rapid Transit) Network Simulation CHOROMAŃSKI, Włodzimierz; DASZCZUK, Wiktor; DYDUCH, Jarosław; MACIEJEWSKI, Mariusz; BRACH, Paweł; GRABSKI, Waldemar etc. Several parameters characterize a station, including its type (in-line/stub-berths), number of berths, sizes of entry and exit buffer. A capacitor is much simpler, because it is not related to passengers. It has only one characteristic – a number of parking places. Capacitor, station and “junction” intersection have one entry and one exit. “Fork” intersection has two exits (to diverge) while “join” intersection has two entries (to merge). Segments are edges of PRT graph. Segments are unidirectional and they have starting and ending points in nodes (capacitors, stations and intersections). A segment is characterized by its length and maximum velocity. Nodes and edges conform a static structure of PRT network. Two types of dynamic objects define a behaviour of the network: vehicles, which originate from capacitors, and passengers appearing in stations. A Passenger group has a specific destination station. Passengers wait for the vehicle (or take an empty one, standing in berth), then a trip is realized, and after completing the journey, they ‘’disappear” from the system. The basic goal of PRT network is to realize trips between stations for passengers. A vehicle is characterized by: capacity, maximum velocity, maximum acceleration and deceleration, maximum friction deceleration (emergency brake) and minimum (static) separation between vehicles. Passengers are organized in groups for a trip to common destination (station). A group is characterized by its cardinality and the target station. THE MODEL The control algorithm of the PRT network is specified in two levels: lower - coordination level, and upper - management level:  Two coordination algorithms are defined: keeping up, and coordination on “join” intersections (joining the traffic from a capacitor/station is realized as a special case of coordination on “join” intersection).  Several management algorithms, including some empty vehicle management algorithms and dynamic routing algorithm. The coordination algorithms use traffic parameters like maximum velocities, maximum acceleration and deceleration, separation between vehicles etc. Some basic rules of behaviour are fixed in the simulator, including movement inside stations, and movement along the track and keeping up rules, are fixed in the simulator. The routing is based on Dijkstra’s algorithm, although edge weights in the algorithm are parameterized by several values depending on static structure (segment lengths), traffic parameters (maximum velocities) and dynamic traffic situation (traffic jams). Various characteristics of the model have different variability:  The structure and topography of the network, type of individual stations (inline/stub berths), capacity of nodes, number of vehicles, boarding and debarking times as well as distribution of passenger group cardinality are constant.  Traffic parameters stay constant during single simulation.  Mean input of passenger groups and origin-destination matrix may be defined to be valid in specific periods.  Maximal velocity of a segment may be changed in specified while. th 13 WCTR, July 15-18, 2013 – Rio de Janeiro, Brazil PRT (Personal Rapid Transit) Network Simulation CHOROMAŃSKI, Włodzimierz; DASZCZUK, Wiktor; DYDUCH, Jarosław; MACIEJEWSKI, Mariusz; BRACH, Paweł; GRABSKI, Waldemar TWO APPROACHES TOWARDS PRT SIMULATION Two independent simulators have been built: one event-driven and second based on cellular automata. Both simulators use the same model of PRT network and the same base parameters. This allows to compare the results and verify correctness of the two simulators. The simulators have different usability properties: the former simulator is more accurate, while the latter one is much faster. Both simulators use micro-simulation (Daszczuk at al. 2011, Fox 1999, Barcelo 1999, Gabard 1999, Nagel and Schreckenberg 1992). In the event-driven simulator, segments are divided into sectors and passing a sector is a basic simulation step. The second simulator is based on cellular automata, where every automaton corresponds to a singular node of the network. The basic step of simulation is 1 second interval, in which all automata are updated. The design of the model typically follows a hierarchy of design steps, from most general to most detailed:  network structure and topography;  characteristics of network elements: station types, numbers of berths, maximal velocities of segments etc.;  acceleration and deceleration, general velocity limits, separation, boarding and debarking times, origin-destination matrix;  parameters of empty vehicle management, priority rules and dynamic routing;  number of vehicles and their individual characteristics (velocity limits for individual vehicles);  passenger input distribution. The simulation experiments may be used to compare throughput characteristics of various network structures, various station and segment parameters, or particular network in which empty vehicles management parameters or dynamic routing parameters vary. Ability to perform the same experiments on two different simulators assures results verification that will be model independent. Some research is described later. EVENT-DRIVEN SIMULATION ENVIRONMENT The movement of vehicles is performed in the event-driven manner, which is a typical solution (Anderson 1998, Lopez et al. 2008). Every segment is divided into sectors. The decisions on the behaviour of vehicles are made in the connections of sectors. Also, starting some actions in capacitors/stations – passenger group occurring, coupling of passenger group with a vehicle, beginning of boarding or debarking, start form parking place etc. are simulation events. The simulation environment is called Feniks 3.0. The main program of the simulator (called “UI”) is responsible for building a model, setting its parameters and to organizing simulation experiments. The external library (called “external simulator”) implements control algorithms (on both layers). The whole simulator is written in C#. The interface is defined between the simulator parts, which allows to define user’s own control algorithms instead of supplied with external simulator. th 13 WCTR, July 15-18, 2013 – Rio de Janeiro, Brazil PRT (Personal Rapid Transit) Network Simulation CHOROMAŃSKI, Włodzimierz; DASZCZUK, Wiktor; DYDUCH, Jarosław; MACIEJEWSKI, Mariusz; BRACH, Paweł; GRABSKI, Waldemar Every event (for example start of movement, time of passing the sector) is served by calling external simulator and performing adequate acts regarding obtained decision. The main program (UI) is a simulation engine, an animation engine and a statistics collector. Also, in UI are included procedures for defining a model structure and topography. External simulator decides on every change of state of every vehicle, for instance:  coupling of passenger group with a vehicle;  beginning of boarding or debarking, passing a connection of sectors, stopping at connection of sectors;  decision on the time of passing the sector; the time is calculated by the coordination algorithm that guarantees safe distance between vehicles on a track. The Simulator plans the vehicle movement analytically and therefore precisely (not approximate). Although the simulation is event-driven, the result of simulation is similar to the analogue simulator. The only inaccuracy comes from the fact that a vehicle may stop only at the connection of sectors (rather than inside a sector). Coordination - movement along a track Vehicles move along a track with maximal possible velocity (which is a minimum from maximum velocities of a model, a segment, a sector and a vehicle). A “horizon” is defined to be a distance “observed” by the vehicle. The nodes and other vehicles closer than the horizon influence the behaviour of the vehicle. The velocity on an empty track (in a distance closer than the horizon) is limited by the acceleration of the vehicle and by deceleration of the vehicle together with velocity limit on the next sectors. If there is a preceding vehicle on a track, or a node (station/capacitor/”join” intersection) which cannot be crossed because of traffic condition, a distance called “separation” is kept. Static separation defines the minimum distance. The coordination algorithm guarantees stopping after the preceding vehicle in static separation distance if preceding vehicles starts to decelerate of brake at the moment. If there is a “fork” intersection closer than the horizon, only the chosen outgoing track (left or right) is analyzed. If a node (capacitor or station) is closer than the horizon, and the node is the target of the trip, the vehicle must reach the zero velocity in a point of diverging from the track towards the node. The exception is when all the parking places and all entry buffers in the node are occupied – in such a case the vehicle must stop in separation distance before the node. If a vehicles wants to “drive through” a node, it runs with maximal allowed velocity, provided that no vehicle starting from the node is in conflict (this case will be described in the next subsection). Coordination - behaviour on “join” intersections A “join” intersection is allocated to one of approaching vehicles by the intersection controller. Only the vehicles closer then the horizon distance are considered. And only closest vehicles on both segments are taken into account, vehicles following them follow normal keeping up. From a vehicle’s point of view, if there is a possible conflict: th 13 WCTR, July 15-18, 2013 – Rio de Janeiro, Brazil PRT (Personal Rapid Transit) Network Simulation CHOROMAŃSKI, Włodzimierz; DASZCZUK, Wiktor; DYDUCH, Jarosław; MACIEJEWSKI, Mariusz; BRACH, Paweł; GRABSKI, Waldemar  if the intersection is allocated to a vehicle on the other track leading to the intersection, the considered vehicle is planned to stop prior to the intersection in static separation distance;  otherwise the vehicle ask the intersection controller which vehicle has the priority; if the other one – the considered vehicle stops as above. Joining the traffic (from a capacitor or a station) is very similar, just the point of the capacitor on the track (or the point of the station) is treated as intersection, and it is being allocated to a vehicle on the move or to a vehicle joining the traffic, depending on priority rules (parameters of coordination algorithm). CELLULAR AUTOMATA SIMULATION ENVIRONMENT Cellular automata is a structure defined by a matrix of cells and their states, transitions and the rules of those transitions (Nagel and Schreckenberg 1992). Automata in such form are mathematical models that construct an environment for a bigger, discrete classes of models, because all the structures describing them are discrete. Each simple cellular automaton consists of n-dimensional, discrete matrix of cells. Each cell is the same (is a copy of the previous cell) and the whole space of the matrix must be filled with the cells put next to each other. Each cell has exactly one state from the finite number of available states. Transition of each cell takes place based on the same, precisely defined local rules (homogeneity), that depend only on the previous state of the cell and states of finite number of neighbouring cells. Transition is discrete and happens at the same time for all cells (parallelism). In the cellular automata, a cell is finite automaton. In order to model the PRT network and traffic a more elaborated adaptation of cellular automata has been chosen – directed graph that represents the infrastructure. The computational model is a directed graph, in which the nodes are the hubs and the edges between the nodes are the segments. Each node and edge has all the parameters that describe a given element (length of the segment, direction of movement, maximum allowed speed, etc.). Each edge is tied to a discrete model of a segment, which is represented by 1-dimensional array. One cell represents one unit of segment and is a parameter of the model. Each junction is represented as one cell. In a given unit of time, in a given cell, there can be only one vehicle. The cell can have one of two available states – empty or occupied by a vehicle. Each vehicle in the model moves with a velocity from the range 0…V-max. In the simulator a topographic model has been implemented, that consists of 2-dimensional, regular and discrete cell matrix. This model is a layer of abstraction over the directed graph. In the graph the nodes are the elements of the segments and the edges define the direction of movement between the nodes. Each node represents exactly one cell in the 2-dimensional matrix. The configuration describing the infrastructure and the initial state of the environment – location of the vehicles, stations and capacitors and the passengers is one of the parameters of the model. The simulator uses and optimal-path algorithm to define a route to the target station. Such an approach allows for a dynamic control of the vehicles during the movement and is an excellent template of the real life movement of the PRT vehicles. th 13 WCTR, July 15-18, 2013 – Rio de Janeiro, Brazil PRT (Personal Rapid Transit) Network Simulation CHOROMAŃSKI, Włodzimierz; DASZCZUK, Wiktor; DYDUCH, Jarosław; MACIEJEWSKI, Mariusz; BRACH, Paweł; GRABSKI, Waldemar Movement in the network After defining all the elements of the cellular automata, the rules of transition can be applied on the cell matrix. The transition process can be split into several parts (Schadschneidert and Schreckenbergt 1993). Initial state, as mentioned previously, is a the definition of the initial conditions of the cell. Usually those are neutral states, that do not cause any conflicts. Update of the automata matrix is a walkthrough sequence of steps for each cell according to the instructions below: 1. Verification of the transition rules – in this step a current state of the cell is verified along with the states of the neighbouring cells and other parameters of automaton 2. Neighbours verification – during this step a verification of any conflict states of the neighbouring cell takes place. If there are any conflicts they have to be resolved according the predefined rules 3. Verification of boundary conditions – verification of the cell at the borders of the matrix. They can be removed (absorbing closed neighbourhood) or new ones can be created (periodic neighbourhood) 4. Verification of number of iterations- if this is a finite automaton, with predefined lifecycle than in this step it is check whether the transition should stop. In the case of the implemented PRT network prototype, updated model of the classic cellular automaton has been implemented (Benjaminy et al. 1996, Li and Wu 2001). The model consists of the following steps (each is performed in parallel for all vehicles in the environment): 1. Acceleration: if the speed of the vehicle V is smaller than the maximum, allowed speed (for a vehicle or road segment) and if the distance to the next vehicle is bigger than V+1, than the velocity is increased by 1, e.g. V:= V + 1 2. Deceleration: if the vehicle at the position I, with velocity V, sees a vehicle at the position i+j, for j smaller or equal to v, then the speed is reduced to j-1, e.g. V:= j-1 3. Randomization (optional): with a probability p1, the velocity of a vehicle is reduced by 1 (if bigger than 0), e.g. V:=V-1 4. Randomization (optional): with a probability p2, vehicle breaks down for a time period J, e.g. the velocity of vehicle is 0 for J units of time 5. If in the next unit of time vehicle will drive through a junction, following conditions are verified: a. If there is no conflict on the junction, e.g. there is no other vehicle coming to junction in the same time do nothing. th 13 WCTR, July 15-18, 2013 – Rio de Janeiro, Brazil PRT (Personal Rapid Transit) Network Simulation CHOROMAŃSKI, Włodzimierz; DASZCZUK, Wiktor; DYDUCH, Jarosław; MACIEJEWSKI, Mariusz; BRACH, Paweł; GRABSKI, Waldemar b. Otherwise define the order of the vehicles (using weights-based function). The priority vehicle does nothing (drives through the junction) and the other one slows down, letting the priority vehicle to drive through. The order of passing vehicles on the junction is defined by the weight of each vehicle - vehicle with higher weight goes first on the crossroad. The weight of the vehicle is determined based on the following function: W ( p)  Wt  t  Wd  d  W p  p  W pas  pas (1) where: W(p) – weight of the vehicle Wt - weight of waiting time t – waiting time Wd - weight of the priority of the segment d – priority of the segment Wp - weight of the priority of the vehicle p – priority of the vehicle W pas - weight of the number of passengers pas – number of passengers in the vehicle 6. Movement: move vehicles V cells in the direction of movement SIMULATION RESEARCH As an example of research work, a simulation experiment with empty trips in PRT network is presented. A PRT network performs people transport on their demand. Yet, to achieve the main goal, some trips without passengers (called empty trips) must be executed. The examples are:  When a passenger group (typically 1-4 persons) occurs at a station, and there are no empty vehicles available in the station, an empty trip must be organized to supply a vehicle for the passenger group. It is the calling mechanism.  When a vehicle approaches a station, and there is no empty berth for it, one of empty vehicles standing in the station must be expelled from the station.  Whether there are many empty vehicles in a given station, and there are no (or little) empty vehicles in other station, an empty trip may be organized from the station to the latter one (target station), especially if average passenger input at the target station posits to expect new group soon. The mechanism is called balancing.  If an empty vehicle stays in the station for a long time, and there are no passengers waiting or expected at the station, the vehicle may be withdrawn to a capacitor for safety reasons (especially at night). th 13 WCTR, July 15-18, 2013 – Rio de Janeiro, Brazil PRT (Personal Rapid Transit) Network Simulation CHOROMAŃSKI, Włodzimierz; DASZCZUK, Wiktor; DYDUCH, Jarosław; MACIEJEWSKI, Mariusz; BRACH, Paweł; GRABSKI, Waldemar The former two mechanisms (calling and expelling) are necessary in PRT system, while the latter two (balancing and withdrawing) are optional. Figure 2 – The first PRT model (city) In a simulation experiment, an impact of a passenger input rate on a number of empty trips is examined. As a PRT network is complicated and the influence of many design parameters in its behaviour is unknown, the research results may be obtained by simulation only. To show that a found relation is general, various network structures and various behaviour parameters should be applied. For the research on the problem mentioned above, two PRT network structures were used, presented in Figures. 2 (city) and 3 (seashore). The circles in the figures represent stations while solid circles represent capacitors. Two numbers of vehicles were used. The main parameters of the network are:  total tracks length 6064,5m (first model) or 5584m (second model),  12 or 24 vehicles,  4 passengers in every trip,  all stations of in-line type,  4 berths in each station.  maximal velocity 14m/s,  maximal acceleration and deceleration 2m/s2,  boarding and debarking times 10s.  static separation 4m (event-driven simulator) and 2m (cellular automata simulator),  passenger input rate differs in specific simulations,  trip destination chosen randomly th 13 WCTR, July 15-18, 2013 – Rio de Janeiro, Brazil PRT (Personal Rapid Transit) Network Simulation CHOROMAŃSKI, Włodzimierz; DASZCZUK, Wiktor; DYDUCH, Jarosław; MACIEJEWSKI, Mariusz; BRACH, Paweł; GRABSKI, Waldemar  empty vehicle management algorithm: calling, expelling and withdrawing mechanism used; Figure 3 - The second PRT model (seashore)   calling principle: nearest empty vehicle is called, if any; expelling principle: expelled empty vehicle is moved to a station on which there are empty berths and w distance to which is close (weighted expelling rule);  withdrawing principle: if a vehicle stays in a station for longer then 120s, it is moved to the capacitor (timeout 120s), provided that there is an empty berth in the capacitor. During this time it may be called to other station (in which a passenger waits) or expelled (if other vehicle approaches the station and there is no empty berth for it). The models (as with every model) have their saturation points, i.e. maximum numbers of trips per hour (depending on the model structure, number of vehicles etc.). The saturation point defines maximal throughput of the network: maximal number of passengers that may be moved to their destinations. For the research purposes, passenger groups input values were applied at saturation point and at several (5) input values less than maximum (passenger group input greater than in saturation point is useless because passenger queues rise to infinity). The results obtained from event-driven simulator are collected in Figure. 4, for the two PRT structures and for two numbers of vehicles. The results from cellular automata simulator are collected in Figure 5. Experiments in both simulators give similar results, where in every plot there is a maximum of empty trips (though the input rates of the maximum and the maximum values differ). The explanation of the observed feature is simple, although it is hard to expect it without simulation experiment. The reason of the plot shape is as follows: th 13 WCTR, July 15-18, 2013 – Rio de Janeiro, Brazil PRT (Personal Rapid Transit) Network Simulation CHOROMAŃSKI, Włodzimierz; DASZCZUK, Wiktor; DYDUCH, Jarosław; MACIEJEWSKI, Mariusz; BRACH, Paweł; GRABSKI, Waldemar  for input rates near saturation point every vehicle freed by passengers is taken immediately by successive passenger group, therefore empty trips seldom are obeyed; for low passenger input rates vehicles are seldom called (at most as many times as new passenger group occur at the stations); situations requiring expelling occur rarely (because of low number of trips with passengers); withdrawing may occur but at most  Model 1 - number of empty trips /12h Model 1 - number of empty trips /12h 3500 6000 3000 5000 2500 4000 2000 12 vehicles 1500 24 vehicles 3000 2000 1000 1000 500 0 0 12,9 25,8 38,7 51,7 64,6 25,1 77,5 50,2 75,3 100,3 125,4 150,5 Passenger groups input /h Passenger groups input /h Model 2 - number of empty trips /12h Model 2 - number of empty trips /12h 3000 5000 2500 4000 2000 3000 12 vehicles 1500 1000 24 vehicles 2000 1000 500 0 0 10,1 20,3 30,4 40,5 50,6 60,8 Passenger groups input /h 20,2 40,3 60,5 80,6 100,8 120,9 Passenger groups input /h Figure 4 – Simulation results for both models and two numbers of vehicles (event-driven simulator)  as frequent as trips with passengers; if passenger input grows then the numbers of calling situations and expelling situations grow as well, while the number of withdrawing situations falls; the growth of empty trips is limited by coincidences of passenger occurring with empty vehicle standing in the same station; the coincidences lastly dominate and the total number of empty trips begins to fall. th 13 WCTR, July 15-18, 2013 – Rio de Janeiro, Brazil PRT (Personal Rapid Transit) Network Simulation CHOROMAŃSKI, Włodzimierz; DASZCZUK, Wiktor; DYDUCH, Jarosław; MACIEJEWSKI, Mariusz; BRACH, Paweł; GRABSKI, Waldemar Figure 5– Simulation results for both models and two numbers of vehicles (cellular automata simulator) It is obvious that the described relation takes place in every reasonable model (perhaps excluding pathological ones, for example if calling a vehicle to other station precedes taking empty vehicle standing in the same station as passengers occur). IMPLEMENTING PRT STEERING AND MANAGING ALGORITHMS IN PHYSICAL LABORATORY ENVIRONMENT In parallel with developing logical model of the PRT network, a physical one is being built on the Warsaw University of Technology. It implements the algorithms described in the former chapters. The security and safety of PRT system is governed by a central computer control and dispatch system. The system consists of layers containing following subsystems (Figure. 6):  dispatcher system – DS.  central control system – CCS  area control system – ACS  radio communication system – RCS  vehicle control system – VCS The main task of the dispatcher system is the facilitation of the monitoring of the whole system by the maintenance personnel. The PRT system is fully automatic and during regular operation no human intervention is required. In case of emergency the personnel can switch to manual control. Dispatchers monitor the current state of the traffic, power system, data transmission and proper operability of computer systems. Dispatchers can mark any component as faulty and plan maintenance tasks. The steering and controlling system is equipped with surveillance video. The video cameras are mounted on the stations and PRT vehicles. Both the stations and vehicles have systems enabling instant contact with the th 13 WCTR, July 15-18, 2013 – Rio de Janeiro, Brazil PRT (Personal Rapid Transit) Network Simulation CHOROMAŃSKI, Włodzimierz; DASZCZUK, Wiktor; DYDUCH, Jarosław; MACIEJEWSKI, Mariusz; BRACH, Paweł; GRABSKI, Waldemar dispatchers In case of emergency dispatchers have direct live feedback from each vehicle. The video and voice transmissions are recorded and archived. At the infrastructure level the system is redundant, therefore provides the continuous operability in case of failure of single components of the system. The transmission system Figure 6 - Structure of the steering system of PRT also provided redundant connection in case of failure. Central control system is of distributed type - different functions are deployed on different steering layers. Central CCS system manages the vehicles. It sets the routes, assigns vehicles to the orders and manages empty vehicles. Performance and optimization of that part of the system is analyzed in the simulation environments. Optimization aims at the defining most efficient ways of movement in PRT network in terms of travel and waiting time as well as the energy efficiency. System react to emergency and faults situation by dynamically reassigning routes. Each change in the traffic situation is archived with a frequency of 100ms. The archived materials are available for analysis, elimination of bottleneck and errors in the system. Area control systems have the tasks of controlling single moving vehicles, controlling the traffic on the stations and assigning right-of-way on crossroads. The whole system is divided into areas that are managed by dedicated computers. The central system collects information about location of each vehicle and distributes it to all area computers. The rightof-way algorithms are parameterized and can be adjusted according to the traffic state in specific areas. ACS does not assign or manipulate routes. In case of emergency of CCS, the ACS directs the vehicles to the nearest station or evacuation points according to predefined scenario. If the connection with specific vehicle is lost, the ACS marks corresponding area as unavailable in order to maximize the safety of all participants. At the infrastructure level ACS is redundant. Additionally the information about movements and state of a vehicle is distributed to corresponding area computer as well as to th 13 WCTR, July 15-18, 2013 – Rio de Janeiro, Brazil PRT (Personal Rapid Transit) Network Simulation CHOROMAŃSKI, Włodzimierz; DASZCZUK, Wiktor; DYDUCH, Jarosław; MACIEJEWSKI, Mariusz; BRACH, Paweł; GRABSKI, Waldemar neighbouring ones. Such setup assures that in case of emergency one of the neighbouring computers can take over the control of the area. The main tasks of vehicle control system is communication with area control system, marking the position of the vehicle, assuring the safe distance to closest vehicle , maintaining a safe velocity, monitoring of the door controlling system, air conditioning, lightning, etc. VCS facilitates the voice and video communication between vehicles and dispatchers. Each vehicle is equipped with displays showing the current state and location of vehicle. Passenger can alter the target station at any moment and a new route will be assigned. During the trip the main display can present commercials and news information. VCS monitors the operation of the engine, breaking and power systems. Communication between PRT vehicles and ACS is facilitated by digital transmission system. The vehicle is in range of two independent base radio stations in any point of the network. The transmission system is implemented in accordance with European standard EN-50159. Data transmission is encoded and encrypted to disable access to the systems by undesirable individuals. Implementation of the European standards EN-50126, EN-50128, EN-50129 ensure highest safety, reliability and maintainability of the whole system. All the critical technical components are configured in fail-safe technique, using 2 out of 2 model CONCLUSIONS AND FURTHER WORK The results from simulation experiments show that the simulation environment is useful for observing and identification of effects occurring in PRT system. Also, it helps to compare various conditions of network operation, especially to find optimal conditions or to avoid traffic jams. In the future, the comparison of various network structures, management algorithms and other features may be performed from various points of view, for example from passenger, network administrator or network maintenance engineer point of view. The very wide set of output parameters measured give the possibility to observe the behaviour of a PRT network from that outlined or other points of view. A detailed log of events may be also obtained, which allows to build simulation statistics viewed from other points of view. Using the simulators, various design aspects may be viewed as complex optimization tasks. For this purpose, “unit costs” of all PRT elements (length of the track, number of vehicles, number of berths in a station, number of intersections of every type etc.) and PRT services (time and distance of a travel, time of vehicle technical maintenance, cost of amortisement, etc.) must be specified. Then, optimal usage parameters (passenger waiting time, effective travel velocity, delays in comparison to optimal conditions) or maintenance parameters (daily distance travelled, full/empty travel ratio etc.) may be identified in simulation experiments. Also, management algorithms build on other principles will be applied, i.e. decentralized th 13 WCTR, July 15-18, 2013 – Rio de Janeiro, Brazil PRT (Personal Rapid Transit) Network Simulation CHOROMAŃSKI, Włodzimierz; DASZCZUK, Wiktor; DYDUCH, Jarosław; MACIEJEWSKI, Mariusz; BRACH, Paweł; GRABSKI, Waldemar ACKNOWLEDGMENT The research work is carried out within the Eco‑Mobility Project that is being cofinanced by the European Regional Development Fund – Operational Program for Innovative Economy. UND-POIG.01.03.01-14-154 – Project Co-ordinator Prof. W. Choromański. REFERENCES Anderson J. E. (1998). Simulation of the operation of personal rapid transit systems, in Computers in Railways VI, WIT Press, Boston Southampton, Computational Mechanics Publications, 1998, 523-532 Andreasson I. J. (2010). Staged introduction of PRT with Mass Transit, in Proc. PRT@LHR 2010 Conference, September 21023, 2010, London Heathrow Airport, Barceló, J., Casas, J., J.L. Ferrer, J.L., García D. (1999). Modelling Advanced Transport Telematic Applications with Microscopic Simulators: the Case of AIMSUN2, Proc. Micro-simulation workshop, Institute for Transport Studies, University of Leeds , July 1999, 11-24 Beamways, http://www.beamways.com/ Benjamin, S. C., Johnson, N. F., Hui, P. M. (1996) Cellular automata models of traffic flow along a highway containing a junction, J. Phys. A: Math. Gen. 29 3119 Castangia, M., Guala, L. (2011). Modelling and simulation of a PRT network, in Proc . 17th International Conference on Urban Transport and the Environment, June 6-8, 2011, Pisa, Italy, pp 459-472 Choromański W., Dyduch J., Brach P. (2011a). Simulation research on innovative PRT transportation system traffic, Logistyka 6/2011 November/October 2011 (in Polish) Choromański W., Kowara J. (2011b). PRT–Modeling and Dynamic Simulation of Track and Vehicle, in Proc. 13th International Conference on Automated People Movers and Transit Systems, Paris, France, May 22-25, 2011 H. Leder W. H. and Sproule W. J. (eds), ISBN (print): 978-0-7844-1193-3, 294-306 Daszczuk, W.B., Mieścicki, J., Grabski, W. (2011). Building simulation models and simulation experiments in FeniksOnArena 1.0, WUT Research Report (in Polish) Fox K. (1999). Introduction to the SMARTEST Project, in Proc. Micro-simulation workshop, Institute for Transport Studies, University of Leeds, July 1999, 2-10 Gabard J.F., Breheret L. (1999). The SITRA-B+ Microscopic Traffic Simulation Model Examples of Use and Future Developments, in Proc. Micro-simulation workshop, Institute for Transport Studies, University of Leeds , July 1999, 49-58 Hermes, http://students.ceid.upatras.gr/~xithalis/simulation_en.html Irwing, J. et al. (1978). Fundamentals of Personal Rapid Transit, Lexington books, Lexington MA, 1978, ISBN is 0-669-02520-8 Li, X., Wu, Q. (2001) Cellular automata model considering the velocity effect of a car on the successive car, Physical Review E, Volume 64, 066128 Lopes C. V., Popov A., Kan L., Morla R. (2008). PRT simulation in an immersive virtual world, in Proc. of the 1st international conference on Simulation tools and techniques th 13 WCTR, July 15-18, 2013 – Rio de Janeiro, Brazil PRT (Personal Rapid Transit) Network Simulation CHOROMAŃSKI, Włodzimierz; DASZCZUK, Wiktor; DYDUCH, Jarosław; MACIEJEWSKI, Mariusz; BRACH, Paweł; GRABSKI, Waldemar for communications, networks and systems, SIMTOOLS ’08, Marseille, France, 57:157:7 Nagel K., Schreckenberg M. (1992) Cellular automata model for freeway traffic. J. Phys. I 2, 2221-2229 RUF, http://www.ruf.dk/ Schadschneidert A. and Schreckenbergt M. (1993) Cellular automaton models and traffic flow, 1993 J. Phys. A: Math. Gen. 26 L679 Zheng, P., Jeffery, D. and McDonald, M. (2009). Development and evaluation of traffic management strategies for personal rapid transit, in Proc. Industrial Simulation Conference 2009, Loughborough, UK, Jun 1-3, 2009, 191-195. th 13 WCTR, July 15-18, 2013 – Rio de Janeiro, Brazil 

5cs.CE

Fast Algorithms for Robust PCA via Gradient Descent Xinyang Yi∗ Dohyung Park∗ Yudong Chen† The University of Texas at Austin∗ arXiv:1605.07784v2 [cs.IT] 19 Sep 2016 {yixy,dhpark,constantine}@utexas.edu∗ Constantine Caramanis∗ Cornell University† yudong.chen@cornell.edu† Abstract We consider the problem of Robust PCA in the fully and partially observed settings. Without corruptions, this is the well-known matrix completion problem. From a statistical standpoint this problem has been recently well-studied, and conditions on when recovery is possible (how many observations do we need, how many corruptions can we tolerate) via polynomial-time algorithms is by now understood. This paper presents and analyzes a non-convex optimization approach that greatly reduces the computational complexity of the above problems, compared to the best available algorithms. In particular, in the fully observed case, with r denoting rank and d dimension, we reduce the complexity from O(r2 d2 log(1/ε)) to O(rd2 log(1/ε)) – a big savings when the rank is big. For the partially observed case, we show the complexity of our algorithm is no more than O(r4 d log d log(1/ε)). Not only is this the best-known run-time for a provable algorithm under partial observation, but in the setting where r is small compared to d, it also allows for near-linear-in-d run-time that can be exploited in the fully-observed case as well, by simply running our algorithm on a subset of the observations. 1 Introduction Principal component analysis (PCA) aims to find a low rank subspace that best-approximates a data matrix Y ∈ Rd1 ×d2 . The simple and standard method of PCA by singular value decomposition (SVD) fails in many modern data problems due to missing and corrupted entries, as well as sheer scale of the problem. Indeed, SVD is highly sensitive to outliers by virtue of the squared-error criterion it minimizes. Moreover, its running time scales as O(rd2 ) to recover a rank r approximation of a d-by-d matrix. While there have been recent results developing provably robust algorithms for PCA (e.g., [8, 28]), the running times range from O(r2 d2 ) to O(d3 )1 and hence are significantly worse than SVD. Meanwhile, the literature developing sub-quadratic algorithms for PCA (e.g., [16, 15, 3]) seems unable to guarantee robustness to outliers or missing data. Our contribution lies precisely in this area: provably robust algorithms for PCA with improved run-time. Specifically, we provide an efficient algorithm with running time that matches SVD while nearly matching the best-known robustness guarantees. In the case where rank is small compared to dimension, we develop an algorithm with running time that is nearly linear in the dimension. This last algorithm works by subsampling the data, and therefore we also show that our algorithm solves the Robust PCA problem with partial observations (a generalization of matrix completion and Robust PCA). 1 For precise dependence on error and other factors, please see details below. 1 1.1 The Model and Related Work We consider the following setting for robust PCA. Suppose we are given a matrix Y ∈ Rd1 ×d2 that has decomposition Y = M ∗ + S ∗ , where M ∗ is a rank r matrix and S ∗ is a sparse corruption matrix containing entries with arbitrary magnitude. The goal is to recover M ∗ and S ∗ from Y . To ease notation, we let d1 = d2 = d in the remainder of this section. Provable solutions for this model are first provided in the works of [9] and [8]. They propose to solve this problem by convex relaxation: min |||M |||nuc + λkSk1 , s.t. Y = M + S, M,S (1) where |||M |||nuc denotes the nuclear norm of M . Despite analyzing the same method, the corruption models in [8] and [9] differ. In [8], the authors consider the setting where the entries of M ∗ are corrupted at random with probability α. They show their method succeeds in exact recovery with α as large as 0.1, which indicates they can tolerate a constant fraction of corruptions. Work in [9] considers a deterministic corruption model, where nonzero entries of S ∗ can have arbitrary position, but the sparsity of each row √ and column does not exceed αd. They prove that for exact recovery, it can allow α = O(1/(µr d)). This was subsequently further improved to α = O(1/(µr)), which is in fact optimal [12, 19]. Here, µ represents the incoherence of M ∗ (see Section 2 for details). In this paper, we follow this latter line and focus on the deterministic corruption model. The state-of-the-art solver [21] for (1) has time complexity O(d3 /ε) to achieve error ε, and is thus much slower than SVD, and prohibitive for even modest values of d. Work in [22] considers the deterministic corruption model, and improves this running time without sacrificing the robustness guarantee on α. They propose an alternating projection (AltProj) method to estimate the low rank and sparse structures iteratively and simultaneously, and show their algorithm has complexity O(r2 d2 log(1/ε)), which is faster than the convex approach but still slower than SVD. Non-convex approaches have recently seen numerous developments for applications in low-rank estimation, including alternating minimization (see e.g. [20, 18, 17]) and gradient descent (see e.g. [4, 11, 25, 26, 31, 32]). These works have fast running times, yet do not provide robustness guarantees. One exception is [11], where the authors analyze a row-wise `1 projection method for recovering S ∗ . Their analysis hinges on positive semidefinite M ∗ , and the algorithm requires prior knowledge of the `1 norm of every row of S ∗ and is thus prohibitive in practice. Another exception is work [17], which analyzes alternating minimization plus an overall sparse projection. Their algorithm is shown to tolerate at most a fraction of α = O(1/(µ2/3 r2/3 d)) corruptions. As we discuss in Section 1.2, we can allow S ∗ to have much higher sparsity α = O(1/(µr1.5 )), which is close to optimal. After the initial post of our paper on arxiv, Cherapanamjeri et al. [14] posted their paper on solving robust PCA with partial observations by a method modified from AltProj. Their approach is shown to have optimal robustness. The sample complexity they established is O(µ2 r2 d log2 d log2 (1/ε)), which depends on the estimation error ε, since the analysis of AltProj under partial observations requires sampling splitting. In contrast, the sample complexity O(µ2 r2 d log d) of our approach (see Corollary 2) does not have such dependence. Moreover, their method requires computing rank-r SVD of the observed matrices in every round of iterations. Our algorithm only needs to compute rank-r SVD (approximately) once in the initialization step. It is worth mentioning other works that obtain provable guarantees of non-convex algorithms or problems including phase retrieval [5, 13, 30], EM algorithms [2, 27, 29], tensor decompositions [1] and second order method [24]. It might be interesting to bring robust considerations to these works. 2 1.2 Our Contributions In this paper, we develop efficient non-convex algorithms for robust PCA. We propose a novel algorithm based on the projected gradient method on the factorized space. We also extend it to solve robust PCA in the setting with partial observations, i.e., in addition to gross corruptions, the data matrix has a large number of missing values. Our main contributions are summarized as follows.2 1. We propose a novel sparse estimator for the setting of deterministic corruptions. For the lowrank structure to be identifiable, it is natural to assume that deterministic corruptions are “spread out” (no more than some number in each row/column). We leverage this information in a simple but critical algorithmic idea, that is tied to the ultimate complexity advantages our algorithm delivers. 2. Based on the proposed sparse estimator, we propose a projected gradient method on the matrix factorized space. While non-convex, the algorithm is shown to enjoy linear convergence under proper initialization. Along with a new initialization method, we show that robust PCA can be solved within complexity O(rd2 log(1/ε)) while ensuring robustness α = O(1/(µr1.5 )). Our algorithm is thus faster than the best previous known algorithm by a factor of r, and enjoys superior empirical performance as well. 3. Algorithms for Robust PCA with partial observations still rely on a computationally expensive convex approach, as apparently this problem has evaded treatment by non-convex methods. We consider precisely this problem. In a nutshell, we show that our gradient method succeeds (it is guaranteed to produce the subspace of M ∗ ) even when run on no more than O(µ2 r2 d log d) random entries of Y . The computational cost is O(µ3 r4 d log d log(1/ε)). When rank r is small compared to the dimension d, in fact this dramatically improves on our bound above, as our cost becomes nearly linear in d. We show, moreover, that this savings and robustness to erasures comes at no cost in the robustness guarantee for the deterministic (gross) corruptions. While this demonstrates our algorithm is robust to both outliers and erasures, it also provides a way to reduce computational costs even in the fully observed setting, when r is small. 4. An immediate corollary of the above result provides a guarantee for exact matrix completion, with general rectangular matrices, using O(µ2 r2 d log d) observed entries and O(µ3 r4 d log d log(1/ε)) time, thereby improving on existing results in [11, 25]. 1.3 Organization and Notation The remainder of this paper is organized as follows. In Section 2, we formally describe our problem and assumptions. In Section 3, we present and describe our algorithms for fully (Algorithm 1) and partially (Algorithm 2) observed settings. In Section 4.1, we establish theoretical guarantees of Algorithm 1. The theory for partially observed setting are presented in Section 4.2. The numerical results are collected in Section 5. Sections 6, 7 and Appendix A contain all the proofs and technical lemmas. To ease presentation, the discussion here assumes M ∗ has constant condition number, whereas our results below show the dependence on condition number explicitly. 2 3   For any index set Ω ⊆ [d1 ]×[d2 ], we let Ω(i,·) := (i, j) ∈ Ω j ∈ [d2 ] , Ω(·,j) := (i, j) ∈ Ω i ∈ [d1 ] . For any matrix A ∈ Rd1 ×d2 , we denote its projector onto support Ω by ΠΩ (A), i.e., the (i, j)-th entry of ΠΩ (A) is equal to A if (i, j) ∈ Ω and zero otherwise. The i-th row and j-th column of A are denoted by A(i,·) and A(·,j) . The (i, j)-th entry is denoted as A(i,j) . Operator norm of A is |||A|||op . Frobenius norm of A is |||A|||F . The `a /`b norm of A is denoted by |||A|||b,a , i.e., the `a norm of the vector formed by the `b norm of every row. For instance, |||A|||2,∞ stands for maxi∈[d1 ] kA(i,·) k2 . 2 Problem Setup We consider the problem where we observe a matrix Y ∈ Rd1 ×d2 that satisfies Y = M ∗ + S ∗ , where M ∗ has rank r, and S ∗ is corruption matrix with sparse support. Our goal is to recover M ∗ and S ∗ . In the partially observed setting, in addition to sparse corruptions, we have erasures. We assume that each entry of M ∗ + S ∗ is revealed independently with probability p ∈ (0, 1). In particular, for any (i, j) ∈ [d1 ] × [d2 ], we consider the Bernoulli model where ( (M ∗ + S ∗ )(i,j) , with probability p; Y(i,j) = (2) ∗, otherwise. We denote the support of Y by Φ = {(i, j) | Y(i,j) 6= ∗}. Note that we assume S ∗ is not adaptive to Φ. As is well-understood thanks to work in matrix completion, this task is impossible in general – we need to guarantee that M ∗ is not both low-rank and sparse. To avoid such identifiability issues, we make the following standard assumptions on M ∗ and S ∗ : (i) M ∗ is not near-sparse or “spiky.” We impose this by requiring M ∗ to be µ-incoherent – given a singular value decomposition (SVD)3 M ∗ = L∗ Σ∗ R∗> , we assume that r r µr µr ∗ ∗ |||L |||2,∞ ≤ , |||R |||2,∞ ≤ . d1 d2 (ii) The entries of S ∗ are “spread out” – for α ∈ [0, 1), we assume S ∗ ∈ Sα , where n o Sα := A ∈ Rd1 ×d2 kA(i,·) k0 ≤ αd2 for all i ∈ [d1 ] ; kA(·,j) k0 ≤ αd1 for all j ∈ [d2 ] . (3) In other words, S ∗ contains at most α-fraction nonzero entries per row and column. 3 Algorithms For both the full and partial observation settings, our method proceeds in two phases. In the first phase, we use a new sorting-based sparse estimator to produce a rough estimate Sinit for S ∗ based on the observed matrix Y , and then find a rank r matrix factorized as U0 V0> that is a rough estimate of M ∗ by performing SVD on (Y − Sinit ). In the second phase, given (U0 , V0 ), we perform an iterative method to produce series {(Ut , Vt )}∞ t=0 . In each step t, we first apply our sparse estimator to produce a sparse matrix St based on (Ut , Vt ), and then perform a projected gradient descent step on the low-rank factorized space to produce (Ut+1 , Vt+1 ). This flow is the same for full and partial observations, though a few details differ. Algorithm 1 gives the full observation algorithm, and Algorithm 2 gives the partial observation algorithm. We now describe the key details of each algorithm. 3 Throughout this paper, we refer to SVD of rank r matrix by form LΣR> where Σ ∈ Rr×r . 4 Sparse Estimation. A natural idea is to keep those entries of residual matrix Y − M that have large magnitude. At the same time, we need to make use of the dispersed property of Sα that every column and row contain at most α-fraction of nonzero entries. Motivated by these two principles, we introduce the following sparsification operator: For any matrix A ∈ Rd1 ×d2 : for all (i, j) ∈ [d1 ]×[d2 ], we let ( (αd ) (αd ) A(i,j) , if |A(i,j) | ≥ |A(i,·)2 | and |A(i,j) | ≥ |A(·,j)1 | Tα [A] := , (4) 0, otherwise (k) (k) where A(i,·) and A(·,j) denote the elements of A(i,·) and A(·,j) that have the k-th largest magnitude respectively. In other words, we choose to keep those elements that are simultaneously among the largest α-fraction entries in the corresponding row and column. In the case of entries having identical magnitude, we break ties arbitrarily. It is thus guaranteed that Tα [A] ∈ Sα . Algorithm 1 Fast RPCA INPUT: Observed matrix Y with rank r and corruption fraction α; parameters γ, η; number of iterations T . // Phase I: Initialization. 1: Sinit ← Tα [Y ] // see (4) for the definition of Tα [·]. 2: [L, Σ, R] ← SVDr [Y − Sinit ] 4 3: U0 ← LΣ1/2 , V0 ← RΣ1/2 . Let U, V be defined according to (7). // Phase II: Gradient based iterations. 4: U0 ← ΠU (U0 ), V0 ← ΠV (V0 ) 5: for t = 0, 1, . . . , T − 1 do   6: St ← Tγα Y − Ut Vt>
7: Ut+1 ← ΠU Ut − η∇U L(Ut , Vt ; St ) − 12 ηUt (Ut> Ut − Vt> Vt )
8: Vt+1 ← ΠV Vt − η∇V L(Ut , Vt ; St ) − 21 ηVt (Vt> Vt − Ut> Ut ) 9: end for OUTPUT: (UT , VT ) Initialization. In the fully observed setting, we compute Sinit based on Y as Sinit = Tα [Y ]. In the partially observed setting with sampling rate p, we let Sinit = T2pα [Y ]. In both cases, we then set U0 = LΣ1/2 and V0 = RΣ1/2 , where LΣR> is an SVD of the best rank r approximation of Y − Sinit . Gradient Method on Factorized Space. After initialization, we proceed by projected gradient descent. To do this, we define loss functions explicitly in the factored space, i.e., in terms of U, V and S: L(U, V ; S) := e V ; S) := L(U, 1 |||U V > + S − Y |||2F , (fully observed) 2   1 |||ΠΦ U V > + S − Y |||2F . (partially observed) 2p (5) (6) 4 SVDr [A] stands for computing a rank-r SVD of matrix A, i.e., finding the top r singular values and vectors of A. Note that we only need to compute rank-r SVD approximately (see the initialization error requirement in Theorem 1) so that we can leverage fast iterative approaches such as block power method and Krylov subspace methods [23]. 5 Recall that our goal is to recover M ∗ that satisfies the µ-incoherent condition. Given an SVD M ∗ = L∗ ΣR∗> , we expect that the solution (U, V ) is close to (L∗ Σ1/2 , R∗ Σ1/2 ) up to some rotation. In order to serve such µ-incoherent structure, it is natural to put constraints on the row norms of U, V based on |||M ∗ |||op . As |||M ∗ |||op is unavailable, given U0 , V0 computed in the first phase, we rely on the sets U, V defined as r r     2µr 2µr d2 ×r d1 ×r |||A|||2,∞ ≤ U := A ∈ R |||A|||2,∞ ≤ |||U0 |||op , V := A ∈ R |||V0 |||op . (7) d1 d2 Now we consider the following optimization problems with constraints: min U ∈U ,V ∈V,S∈Sα min U ∈U ,V ∈V,S∈Spα 1 L(U, V ; S) + |||U > U − V > V |||2F , 8 e V ; S) + 1 |||U > U − V > V |||2 . L(U, F 64 (fully observed) (partially observed) (8) (9) The regularization term in the objectives above is used to encourage that U and V have the same scale. Given (U0 , V0 ), we propose the following iterative method to produce series {(Ut , Vt )}∞ t=0 and {St }∞ . We give the details for the fully observed case – the partially observed case is similar. For t=0   > t = 0, 1, . . ., we update St using the sparse estimator St = Tγα Y − Ut Vt , followed by a projected gradient update on Ut and Vt   1 > > Ut+1 = ΠU Ut − η∇U L(Ut , Vt ; St ) − ηUt (Ut Ut − Vt Vt ) , 2   1 > > Vt+1 = ΠV Vt − η∇V L(Ut , Vt ; St ) − ηVt (Vt Vt − Ut Ut ) . 2 Here α is the model parameter that characterizes the corruption fraction, γ and η are algorithmic tunning parameters, which we specify in our analysis. Essentially, the above algorithm corresponds to applying projected gradient method to optimize (8), where S is replaced by the aforementioned sparse estimator in each step. Algorithm 2 Fast RPCA with partial observations INPUT: Observed matrix Y with support Φ; parameters τ, γ, η; number of iterations T . // Phase I: Initialization. 1: Sinit ← T2pα [ΠΦ (Y )] 2: [L, Σ, R] ← SVDr [ p1 (Y − Sinit )] 3: U0 ← LΣ1/2 , V0 ← RΣ1/2 . Let U, V be defined according to (7). // Phase II: Gradient based iterations. 4: U0 ← ΠU (U0 ), V0 ← ΠV (V0 ) 5: for t = 0, 1, . . . , T − 1 do 
 > 6: St ← Tγpα Π Y − U V t Φ t   e t , Vt ; St ) − 1 ηUt (U > Ut − V > Vt ) 7: Ut+1 ← ΠU Ut − η∇U L(U t t 16   1 e 8: Vt+1 ← ΠV Vt − η∇V L(Ut , Vt ; St ) − 16 ηVt (Vt> Vt − Ut> Ut ) 9: end for OUTPUT: (UT , VT ) 6 4 Main Results In this section, we establish theoretical guarantees for Algorithm 1 in the fully observed setting and for Algorithm 2 in the partially observed setting. 4.1 Analysis of Algorithm 1 We begin with some definitions and notation. It is important to define a proper error metric because the optimal solution corresponds to a manifold and there are many distinguished pairs (U, V ) that minimize (8). Given the SVD of the true low-rank matrix M ∗ = L∗ Σ∗ R∗> , we let U ∗ := L∗ Σ∗1/2 and V ∗ := R∗ Σ∗1/2 . We also let σ1∗ ≥ σ2∗ ≥ . . . ≥ σr∗ be sorted nonzero singular values of M ∗ , and denote the condition number of M ∗ by κ, i.e., κ := σ1∗ /σr∗ . We define estimation error d(U, V ; U ∗ , V ∗ ) as the minimal Frobenius norm between (U, V ) and (U ∗ , V ∗ ) with respect to the optimal rotation, namely p d(U, V ; U ∗ , V ∗ ) := min |||U − U ∗ Q|||2F + |||V − V ∗ Q|||2F , (10) Q∈Qr for Qr the set of r-by-r orthonormal matrices. This metric controls reconstruction error, as p |||U V > − M ∗ |||F . σ1∗ d(U, V ; U ∗ , V ∗ ), (11) p when d(U, V ; U ∗ , V ∗ ) ≤ σ1∗ . We denote the local region around the optimum (U ∗ , V ∗ ) with radius ω as o n B2 (ω) := (U, V ) ∈ Rd1 ×r × Rd2 ×r d(U, V ; U ∗ , V ∗ ) ≤ ω . The next two theorems provide guarantees for the initialization phase and gradient iterations, respectively, of Algorithm 1. The proofs are given in Sections 6.1 and 6.2. Theorem 1 (Initialization). Consider the paired (U0 , V0 ) produced in the first phase of Algorithm 1. If α ≤ 1/(16κµr), we have √ √ p d(U0 , V0 ; U ∗ , V ∗ ) ≤ 28 καµr r σ1∗ . Theorem 2 (Convergence). Consider the second phase of Algorithm 1. Suppose we choose γ = 2 and η = c/σ1∗ for any c ≤ 1/36. There exist constants c1 , c2 such that when α ≤ c1 /(κ2 µr), given  p any (U0 , V0 ) ∈ B2 c2 σr∗ /κ , the iterates {(Ut , Vt )}∞ t=0 satisfy  c t 2 d2 (Ut , Vt ; U ∗ , V ∗ ) ≤ 1 − d (U0 , V0 ; U ∗ , V ∗ ). 8κ Therefore, using proper initialization and step size, the gradient iteration converges at a linear rate with a constant contraction factor 1 − O(1/κ). To obtain relative precision ε compared to the initial error, it suffices to perform √ O(κ log(1/ε)) iterations. Note that the step size is chosen according to 1/σ1∗ . When α . 1/(µ κr3 ), Theorem 1 and the inequality (11) together imply that |||U0 V0> − M ∗ |||op ≤ 12 σ1∗ . Hence we can set the step size as η = O(1/σ1 (U0 V0> )) using being the top singular value σ1 (U0 V0> ) of the matrix U0 V0> Combining Theorems 1 and 2 implies the following result, proved in Section 6.3, that provides an overall guarantee for Algorithm 1. 7 Corollary 1. Suppose that ( α ≤ c min 1 1 √ 3, 2 µ κr µκ r ) for some constant c. Then for any ε ∈ (0, 1), Algorithm 1 with T = O(κ log(1/ε)) outputs a pair (UT , VT ) that satisfies |||UT VT> − M ∗ |||F ≤ ε · σr∗ . (12) Remark 1 (Time Complexity). For simplicity we assume d1 = d2 = d. Our sparse estimator (4) can be implemented by finding the top αd elements of each row and column via partial quick sort, which has running time O(d2 log(αd)). Performing rank-r SVD in the first phase and computing the gradient in each iteration both have complexity O(rd2 ).5 Algorithm 1 thus has total running time O(κrd2 log(1/ε)) for achieving an  accuracy as in (12). We note that when κ = O(1), our algorithm is orderwise faster than the AltProj algorithm in [22], which has running time O(r2 d2 log(1/ε)). Moreover, our algorithm only requires computing one singular value decomposition. Remark 2 (Robustness). Assuming κ = O(1), our algorithm can tolerate corruption at a sparsity √ √ level up to α = O(1/(µr r)). This is worse by a factor r compared to the optimal statistical guarantee 1/(µr) obtained in [12, 19, 22]. This looseness is a consequence of the condition for (U0 , V0 ) in Theorem 2. Nevertheless, when µr = O(1), our algorithm can tolerate a constant α fraction of corruptions. Notably, we show that gradient descent works in the case of α = O(1/(µr)) if initialization is sufficiently close. Accordingly, to provide an algorithm with optimal robustness, it is straightforward to use a more complicated initial method such as AltProj that can tolerate 1/(µr) fraction of corruptions while satisfying our initial condition. As gradient descent provides a simple and efficient way to successively refine the estimation, such combination still gives a better running time than prior arts. 4.2 Analysis of Algorithm 2 We now move to the guarantees of Algorithm 2. We show here that not only can we handle partial observations, but in fact subsampling the data in the fully observed case can significantly reduce the time complexity from the guarantees given in the previous section without sacrificing robustness. In particular, for smaller values of r, the complexity of Algorithm 2 has near linear dependence on the dimension d, instead of quadratic. In the following discussion, we let d := max{d1 , d2 }. The next two results, proved in Sections 6.4 and 6.5, control the quality of the initialization step, and then the gradient iterations. Theorem 3 (Initialization, partial observations). Suppose the observed indices Φ follow the Bernoulli model given in (2). Consider the pair (U0 , V0 ) produced in the first phase of Algorithm 2. There √ exist constants {ci }3i=1 such that for any  ∈ (0, r/(8c1 κ)), if  2  1 µr 1 log d α≤ , p ≥ c2 + , (13) 2 64κµr  α d1 ∧ d2 then we have 5 p √ √ p d(U0 , V0 ; U ∗ , V ∗ ) ≤ 51 καµr r σ1∗ + 7c1  κσ1∗ , In fact, it suffices to compute the best rank-r approximation with running time independent of the eigen gap. 8 with probability at least 1 − c3 d−1 . Theorem 4 (Convergence, partial observations). Suppose the observed indices Φ follow the Bernoulli model given in (2). Consider the second phase of Algorithm 2. Suppose we choose γ = 3, and η = c/(µrσ1∗ ) for a sufficiently small constant c. There exist constants {ci }4i=1 such that if α≤ κ4 µ2 r2 log d c1 , and p ≥ c 2 κ2 µr d1 ∧ d2 (14) then with probability at least 1 − c3 d−1 , the iterates {(Ut , Vt )}∞ t=0 satisfy 2 ∗ ∗ d (Ut , Vt ; U , V ) ≤  c 1− 64µrκ t d2 (U0 , V0 ; U ∗ , V ∗ )   p for all (U0 , V0 ) ∈ B2 c4 σr∗ /κ . The above result ensures linear convergence to (U ∗ , V ∗ ) (up to rotation) even when the gradient iterations are computed using partial observations. Note that setting p = 1 recovers Theorem 2 up to an additional factor µr in the contraction factor. For achieving ε relative accuracy, now we need O(µrκ log(1/ε)) iterations. Putting Theorems 3 and 4 together, we have the following overall guarantee, proved in Section 6.6, for Algorithm 2. Corollary 2. Suppose that ( α ≤ c min 1 1 √ 3, 2 µ κr µκ r ) , p ≥ c0 κ4 µ2 r2 log d , d1 ∧ d2 for some constants c, c0 . With probability at least 1 − O(d−1 ), for any ε ∈ (0, 1), Algorithm 2 with T = O(µrκ log(1/ε)) outputs a pair (UT , VT ) that satisfies |||UT VT> − M ∗ |||F ≤ ε · σr∗ . (15) This result shows that partial observations do not compromise robustness to sparse corruptions: as long as the observation probability p satisfies the condition in Corollary 2, Algorithm 2 enjoys the same robustness guarantees as the method using all entries. Below we provide two remarks on the sample and time complexity. For simplicity, we assume d1 = d2 = d, κ = O(1). Remark 3 (Sample complexity and matrix completion). Using the lower bound on p, it is sufficient to have O(µ2 r2 d log d) observed entries. In the special case S ∗ = 0, our partial observation model is equivalent to the model of exact matrix completion (see, e.g., [7]). We note that our sample complexity (i.e., observations needed) matches that of completing a positive semidefinite (PSD) matrix by gradient descent as shown in [11], and is better than the non-convex matrix completion algorithms in [20] and [25]. Accordingly, our result reveals the important fact that we can obtain robustness in matrix completion without deterioration of our statistical guarantees. It is known that that any algorithm for solving exact matrix completion must have sample size Ω(µrd log d) [7], and a nearly tight upper bound O(µrd log2 d) is obtained in [10] by convex relaxation. While sub-optimal by a factor µr, our algorithm is much faster than convex relaxation as shown below. 9 d(U, V ; U ∗ , V ∗ ) GD p = 1 GD p = 0.5 GD p = 0.2 10 3 Time(secs) d(U, V ; U ∗ , V ∗ ) 10 1 10 0 10 -1 10 2 10 1 10 -2 10 GD p = 1 GD p = 0.5 GD p = 0.2 GD p = 0.1 AltProj IALM 0 10 -1 10 -2 10 -3 0 2 4 6 8 10 Iteration count (a) 10 3 10 4 10 5 Dimension d (b) 0 20 40 60 80 100 Time(secs) (c) Figure 1. Results on synthetic data. (a) Plot of log estimation error versus number of iterations when using gradient descent (GD) with varying sub-sampling rate p. It is conducted using d = 5000, r = 10, α = 0.1. (b) Plot of running time of GD versus dimension d with r = 10, α = 0.1, p = 0.15r2 log d/d. The low-rank matrix is recovered in all instances, and the line has slope approximately one. (c) Plot of log estimation error versus running time for different algorithms in problem with d = 5000, r = 10, α = 0.1. Remark 4 (Time complexity). Our sparse estimator on the sparse matrix with support Φ can be implemented via partial quick sort with running time O(pd2 log(αpd)). Computing the gradient in each step involves the two terms in the objective function (9). Computing the gradient of the first term Le takes time O(r|Φ|), whereas the second term takes time O(r2 d). In the initialization phase, performing rank-r SVD on a sparse matrix with support Φ can be done in time O(r|Φ|). We conclude that when |Φ| = O(µ2 r2 d log d), Algorithm 2 achieves the error bound (15) with running time O(µ3 r4 d log d log(1/ε)). Therefore, in the small rank setting with r  d1/3 , even when full observations are given, it is better to use Algorithm 2 by subsampling the entries of Y . 5 Numerical Results In this section, we provide numerical results and compare the proposed algorithms with existing methods, including the inexact augmented lagrange multiplier (IALM) approach [21] for solving the convex relaxation (1) and the alternating projection (AltProj) algorithm proposed in [22]. All algorithms are implemented in MATLAB 6 , and the codes for existing algorithms are obtained from their authors. SVD computation in all algorithms uses the PROPACK library.7 We ran all simulations on a machine with Intel 32-core Xeon (E5-2699) 2.3GHz with 240GB RAM. 5.1 Synthetic Datasets We generate a squared data matrix Y = M ∗ + S ∗ ∈ Rd×d as follows. The low-rank part M ∗ is given by M ∗ = AB > , where A, B ∈ Rd×r have entries drawn independently from a zero mean Gaussian distribution with variance 1/d. For a given sparsity parameter α, each entry of S ∗ is set to be nonzero with probability α, and the values of the nonzero entries are sampled uniformly from [−5r/d, 5r/d]. The results are summarized in Figure 1. Figure 1a shows the convergence of our algorithms for different random instances with different sub-sampling rate p (note that p = 1 corresponds to 6 7 Our code is available at https://www.yixinyang.org/code/RPCA_GD.zip. http://sun.stanford.edu/~rmunk/PROPACK/ 10 Original GD (49.8s) GD, 20% sample (18.1s) AltProj (101.5s) IALM (434.6s) Original GD (87.3s) GD, 20% sample (43.4s) AltProj (283.0s) IALM (801.4s) Figure 2. Foreground-background separation in Restaurant and ShoppingMall videos. In each line, the leftmost image is an original frame, and the other four are the separated background obtained from our algorithms with p = 1, p = 0.2, AltProj, and IALM. The running time required by each algorithm is shown in the title. the fully observed setting). As predicted by Theorems 2 and 4, our gradient method converges geometrically with a contraction factor nearly independent of p. Figure 1b shows the running time of our algorithm with partially observed data. We see that the running time scales linearly with d, again consistent with the theory. We note that our algorithm is memory-efficient: in the large scale setting with d = 2×105 , using approximately 0.1% entries is sufficient for the successful recovery. In contrast, AltProj and IALM are designed to manipulate the entire matrix with d2 = 4×1010 entries, which is prohibitive on a single machine. Figure 1c compares our algorithms with AltProj and IALM by showing reconstruction error versus real running time. Our algorithm requires significantly less computation to achieve the same accuracy level, and using only a subset of the entries provides additional speed-up. 5.2 Foreground-background Separation We apply our method to the task of foreground-background (FB) separation in a video. We use two public benchmarks, the Restaurant and ShoppingMall datasets.8 Each dataset contains a video with static background. By vectorizing and stacking the frames as columns of a matrix Y , the FB separation problem can be cast as RPCA, where the static background corresponds to a low rank matrix M ∗ with identical columns, and the moving objects in the video can be modeled as sparse corruptions S ∗ . Figure 2 shows the output of different algorithms on two frames from the dataset. Our algorithms require significantly less running time than both AltProj and IALM. Moreover, even with 20% sub-sampling, our methods still appear to achieve better separation quality (note that in each of the frames our algorithms remove a person that is not identified by the other algorithms). Figure 3 shows recovery results for several more frames. Again, our algorithms enjoy better running time and outperform AltProj and IALM in separating persons from the background images. In Appendix B, we describe the detailed parameter settings for our algorithm. 8 http://perception.i2r.a-star.edu.sg/bk_model/bk_index.html 11 Original GD (49.8s) GD, 20% sample (18.1s) AltProj (101.5s) IALM (434.6s) Original GD (49.8s) GD, 20% sample (18.1s) AltProj (101.5s) IALM (434.6s) Original GD (87.3s) GD, 20% sample (43.4s) AltProj (283.0s) IALM (801.4s) Original GD (87.3s) GD, 20% sample (43.4s) AltProj (283.0s) IALM (801.4s) Figure 3. More results of FB separation in Restaurant and ShoppingMall videos. The leftmost images are original frames. The right four columns show the results from our algorithms with p = 1, p = 0.2, AltProj [22], and IALM [21]. The runtime of each algorithm is written in the title. 12 6 Proofs In this section we provide the proofs for our main theoretical results in Theorems 1–4 and Corollaries 1–2. 6.1 Proof of Theorem 1 Let Y := Y − Sinit . As Y = M ∗ + S ∗ , we have Y − M ∗ = S ∗ − Sinit . We obtain Y − M ∗ ∈ S2α because S ∗ , Sinit ∈ Sα . We claim that kY −M ∗ k∞ ≤ 2kM ∗ k∞ . Denote the support of S ∗ , Sinit by Ω∗ and Ω respectively. Since Y − M ∗ is supported on Ω ∪ Ω∗ , to prove the claim it suffices to consider the following three cases. • For (i, j) ∈ Ω∗ ∩ Ω, due to rule of sparse estimation, we have (S ∗ − Sinit )(i,j) = 0. ∗ | ≤ 2kM ∗ k . Otherwise, we have |Y ∗ • For (i, j) ∈ Ω∗ \ Ω, we must have |S(i,j) ∞ (i,j) | = |(S + M ∗ )(i,j) | > kM ∗ k∞ . So |Y(i,j) | is larger than any uncorrupted entries in its row and column. Since there are at most α fraction corruptions per row and column, we have Y(i,j) ∈ Ω, which violates the prior condition (i, j) ∈ Ω∗ \ Ω. ∗ , trivially we have |(S • For the last case (i, j) ∈ Ω \ Ω∗ , since (Sinit )(i,j) = M(i,j) init )(i,j) | ≤ ∗ kM k∞ . The following result, proved in Section 7.1, relates the operator norm of Y − M ∗ to its infinite norm. Lemma 1. For any matrix A ∈ Rd1 ×d2 that belongs to Sα given in (3), we have p |||A|||op ≤ α d1 d2 kAk∞ . We thus obtain |||Y − M ∗ |||op ≤ 2α p p d1 d2 kY − M ∗ k∞ ≤ 4α d1 d2 kM ∗ k∞ = 4αµrσ1∗ . In the last step, we use the fact that M ∗ satisfies the µ-incoherent condition, which leads to µr kM ∗ k∞ ≤ |||M ∗ |||op |||L∗ |||2,∞ |||R∗ |||2,∞ ≤ √ |||M ∗ |||op . d1 d2 (16) (17) We denote the i-th largest singular value of Y by σi . By Weyl’s theorem, we have |σi∗ − σi | ≤ ∗ |||Y − M ∗ |||op for all i ∈ [d1 ∧ d2 ]. Since σr+1 = 0, we have σr+1 ≤ |||Y − M ∗ |||op . Recall that U0 V0> is the best rank r approximation of Y . Accordingly, we have |||U0 V0> − M ∗ |||op ≤ |||U0 V0> − Y |||op + |||Y − M ∗ |||op = σr+1 + |||Y − M ∗ |||op ≤ 2|||Y − M ∗ |||op ≤ 8αµrσ1∗ . 1 Under condition αµr ≤ 16κ , we obtain |||U0 V0> − M ∗ |||op ≤ 21 σr∗ . Applying Lemma 5.14 in [26] (we provide it as Lemma 15 for the sake of completeness), we obtain d2 (U0 , V0 ; U ∗ , V ∗ ) ≤ √ 10r|||U0 V0> − M ∗ |||2op 2 |||U0 V0> − M ∗ |||2F ≤ . σr∗ σr∗ 2−1 Plugging the upper bound of |||U0 V0> − M ∗ |||op into the above inequality completes the proof. 13 6.2 Proof of Theorem 2 We essentially follow the general framework developed in [11] for analyzing the behaviors of gradient descent in factorized low-rank optimization. But it is worth to note that [11] only studies the symmetric and positive semidefinite setting, while we avoid such constraint on M ∗ . The techniques for analyzing general asymmetric matrix in factorized space is inspired by the recent work [26] on solving low-rank matrix equations. In our setting, the technical challenge is to verify the local descent condition of the loss function (8), which not only has a bilinear dependence on U and V , but also involves our sparse estimator (4). We begin with some notations. Define the equivalent set of optimal solution as o n (18) E(M ∗ ) := (A, B) ∈ Rd1 ×r × Rd2 ×r A = L∗ Σ∗1/2 Q, B = R∗ Σ∗1/2 Q, where Q ∈ Qr .  p  Given (U0 , V0 ) ∈ B2 c2 σr∗ /κ , by (11), we have |||U0 V0> − M ∗ |||op ≤ 21 σr∗ when c2 is sufficiently small. By Weyl’s theorem We thus have q q q q ∗ ∗ ∗ σ1 /2 ≤ |||U0 |||op ≤ 3σ1 /2, and σ1 /2 ≤ |||V0 |||op ≤ 3σ1∗ /2. As a result, for U, V constructed according to (7), we have E(M ∗ ) ⊆ U × V, and U ⊆ U, V ⊆ V, (19) where ( U := A ∈ Rd1 ×r r |||A|||2,∞ ≤ 3µrσ1∗ d1 ( ) , V := A ∈ Rd2 ×r r |||A|||2,∞ ≤ 3µrσ1∗ d2 ) . We let 1 (20) G(U, V ) := |||U > U − V > V |||2F . 8 For L(U, V ; S), we denote the gradient with respect to M by ∇M L(U, V ; S), i.e. ∇M L(U, V ; S) = UV > + S − Y . The local descent property is implied by combining the following two results, which are proved in Section 6.7 and 6.8 respectively. Lemma 2 (Local descent property of  L). Suppose U, V satisfy (19). For any (U, V ) ∈ (U × V) ∩ p B2 ( σ1∗ ), we let S = Tγα Y − U V > , where we choose γ = 2. Then we have that for (Uπ∗ , Vπ∗ ) ∈ argmin(A,B)∈E(M ∗ ) |||U − A|||2F + |||V − B|||2F and β > 0, > > 2 ∗ hh∇M L(U, V ; S), U V > − Uπ∗ Vπ>∗ + ∆U ∆> V ii ≥ |||U V − Uπ ∗ Vπ ∗ |||F − νσ1 δ − 3 q σ1∗ δ 3 . Here ∆U := U − Uπ∗ , ∆V := V − Vπ∗ , δ := |||∆U |||2F + |||∆V |||2F , and ν := 9(β + 6)αµr + 5β −1 . √ Lemma 3 (Local descent property of G). For any (U, V ) ∈ B2 ( σr∗ ) and (Uπ∗ , Vπ∗ ) ∈ arg min (A,B)∈E(M ∗ ) 14 |||U − A|||2F + |||V − B|||2F , (21) we have hh∇U G(U, V ), U − Uπ∗ ii + hh∇V G(U, V ), V − Vπ∗ ii r σ1∗ δ 3 1 1 > 1 ∗ > 2 ≥ |||U U − V V |||F + σr δ − − |||U V > − Uπ∗ Vπ>∗ |||2F , 8 8 2 2 where δ is defined according to Lemma 2. As another key ingredient, we establish the following smoothness condition, proved in Section 6.9, which indicates that the Frobenius norm of gradient decreases as (U, V ) approaches the optimal manifold.   p Lemma 4 (Smoothness). For any (U, V ) ∈ B2 ( σ1∗ ), we let S = Tγα Y − U V > , where we choose γ = 2. We have that |||∇M L(U, V ; S)|||2F ≤ 6|||U V > − M ∗ |||2F , (22) and |||∇U G(U, V )|||2F + |||∇V G(U, V )|||2F ≤ 2σ1∗ |||U > U − V > V |||2F . (23) With the above results in hand, we are ready to prove Theorem 2. Proof of Theorem 2. We use shorthands δt := d2 (Ut , Vt ; U ∗ , V ∗ ), Lt := L(Ut , Vt ; St ), and Gt := G(Ut , Vt ). For (Ut , Vt ), let (Uπt ∗ , Vπt∗ ) := argmin(A,B)∈E(M ∗ ) |||Ut − A|||2F + |||Vt − B|||2F . Define ∆tU := Ut − Uπt ∗ , ∆tV := Vt − Vπt∗ . We prove Theorem 2 by induction. It is sufficient  topconsider  one step of the iteration. For any t ≥ 0, under the induction hypothesis (Ut , Vt ) ∈ B2 c2 σr∗ /κ . We find that δt+1 ≤ |||Ut+1 − Uπt ∗ |||2F + |||Vt+1 − Vπt∗ |||2F ≤ |||Ut − η∇U Lt − η∇U Gt − Uπt ∗ |||2F + |||Vt − η∇V Lt − η∇V Gt − Vπt∗ |||2F ≤ δt − 2η hh∇U Lt + ∇U Gt , Ut − Uπt ∗ ii −2η hh∇V Lt + ∇V Gt , Vt − Vπt∗ ii | {z } | {z } W1 2 W2 2 2 2 + η |||∇U Lt + ∇U Gt |||F +η |||∇V Lt + ∇V Gt |||F , | | {z } {z } W3 (24) W4 where the second step follows from the non-expansion property of projection onto U, V, which is implied by E(M ∗ ) ⊆ U × V shown in (19). Since ∇U Lt = [∇M Lt ] V and ∇V Lt = [∇M Lt ]> U , we have t t> hh∇U Lt , Ut − Uπt ∗ ii + hh∇V Lt , Vt − Vπt∗ ii = hh∇M Lt , Ut Vt> − Uπt ∗ Vπt> ∗ + ∆U ∆V ii. Combining Lemma 2 and 3, under condition δt < σr∗ , we have that 1 1 1 W1 + W2 ≥ |||Ut Vt> − M ∗ |||2F + |||Ut> Ut − Vt> Vt |||2F + σr∗ δt − νσ1∗ δt − 4 2 8 8 15 q σ1∗ δt3 . On the other hand, we have W3 + W4 ≤ 2|||∇U Lt |||2F + 2|||∇U Gt |||2F + 2|||∇V Lt |||2F + 2|||∇V Gt |||2F ≤ 2(|||Ut |||2op + |||Vt |||2op )|||∇M Lt |||2F + 2|||∇U Gt |||2F + 2|||∇V Gt |||2F ≤ 36σ1∗ |||Ut Vt> − M ∗ |||2F + 4σ1∗ |||Ut> Ut − Vt> Vt |||2F ,   p where the last step is implied by Lemma 4 and the assumption (Ut , Vt ) ∈ B2 c2 σr∗ /κ that leads p p to |||Ut |||op ≤ 3σ1∗ /2, |||Vt |||op ≤ 3σ1∗ /2. By the assumption η = c/σ1∗ for any constant c ≤ 1/36, we thus have q 1 ∗ 2 ∗ −2η(W1 + W2 ) + η (W3 + W4 ) ≤ − ησr δt + 2ηνσ1 δt + 8η σ1∗ δt3 . 4 In Lemma 2, choosingpβ = 320κ and assuming α . 1/(κ2 µr), we can have ν ≤ 1/(32κ). Assuming 1 ∗ δt . σr∗ /κ leads to 14 σ1∗ δt3 ≤ 16 σr δt . We thus obtain   ησr∗ δt+1 ≤ 1 − δt . (25) 8 Under initial condition δ0 . σr∗ /κ, we obtain that such condition holds for all t since estimation error decays geometrically after each iteration. Then applying (25) for all iterations, we conclude that for all t = 0, 1, . . .,   ησ ∗ t δt ≤ 1 − r δ0 . 8 6.3 Proof of Corollary 1 We need α . κ21µr due to the condition of Theorem 2. In order to ensure the linear convergence happens, it suffices to let the initial error shown in Theorem 1 be less than the corresponding condition in Theorem 2. Accordingly, we need p √ √ p 28 καµr r σ1∗ . σr∗ /κ, which leads to α . 1 √ 3. µ rκ Using the conclusion that gradient descent has linear convergence, choosing T = O(κ log(1/ε)), we have σ∗ d2 (UT , VT ; U ∗ , V ∗ ) ≤ ε2 d2 (U0 , V0 ; U ∗ , V ∗ ) . ε2 r . κ ∗ ∗ > Finally, applying the relationship between d(UT , VT ; U , V ) and |||UT VT − M ∗ |||F shown in (11), we complete the proof. 6.4 Proof of Theorem 3 Let Y := p1 (Y − Sinit ). Similar to the proof of Theorem 1, we first establish an upper bound on |||Y − M ∗ |||op . We have that 1 1 |||Y − M ∗ |||op ≤ |||Y − ΠΦ M ∗ |||op + ||| ΠΦ (M ∗ ) − M ∗ |||op . p p 16 (26) For the first term, we have Y − p1 ΠΦ M ∗ = p1 (ΠΦ (S ∗ ) − Sinit ) because Y = ΠΦ (M ∗ + S ∗ ). Lemma d 10 shows that under condition p & α(dlog , there are at most 32 pα-fraction nonzero entries in each 1 ∧d2 ) row and column of ΠΦ (S ∗ ) with high probability. Since Sinit ∈ S2pα , we have ΠΦ (S ∗ ) − Sinit ∈ S4pα . (27) kΠΦ (S ∗ ) − Sinit k∞ ≤ 2kM ∗ k∞ . (28) In addition, we prove below that Denote the support of ΠΦ (S ∗ ) and Sinit by Ω∗o and Ω. For (i, j) ∈ Ω∗o ∩ Ω and (i, j) ∈ Ω \ Ω∗o , we ∗ , respectively. To prove the claim, have (ΠΦ (S ∗ ) − Sinit )(i,j) = 0 and (ΠΦ (S ∗ ) − Sinit )(i,j) = −M(i,j) ∗ | < 2kM ∗ k . If this is not true, then we must have it remains to show that for (i, j) ∈ Ω∗o \ Ω, |S(i,j) ∞ ∗ |Y(i,j) | > kM k∞ . Accordingly, |Y(i,j) | is larger than the magnitude of any uncorrupted entries in its row and column. Note that on the support Φ, there are at most 32 pα corruptions per row and column, we have (i, j) ∈ Ω, which violates our prior condition (i, j) ∈ Ω∗o \ Ω. Using these two properties (27), (28) and applying Lemma 1, we have p p 1 |||Y − ΠΦ M ∗ |||op ≤ 4α d1 d2 kΠΦ (S ∗ ) − Sinit k∞ ≤ 8α d1 d2 kM ∗ k∞ ≤ 8αµrσ1∗ , (29) p where the last step follow from (17). For the second term in (26), we use the following lemma proved in [10]. Lemma 5 (Lemma 2 in [10]). Suppose A ∈ Rd1 ×d2 is a fixed matrix. We let d := max{d1 , d2 }. There exists a constant c such that with probability at least 1 − O(d−1 ), s ! n o 1 log d log d > ||| ΠΦ (A) − A|||op ≤ c kAk∞ + max |||A|||2,∞ , |||A |||2,∞ . p p p Given the SVD M ∗ = L∗ ΣR∗> , for any i ∈ [d1 ], we have ∗ kM(i,·) k2 = kL∗(i,·) ΣR∗> k2 ≤ σ1∗ kL∗(i,·) k2 ≤ σ1∗ r µr . d1 We can bound |||M ∗> |||2,∞ similarly. Lemma 5 leads to 1 log d µr √ ||| ΠΦ (M ∗ ) − M ∗ |||op ≤ c0 σ ∗ + c0 p p d1 d2 1 s log d p r √ µr σ1∗ ≤ c0 σ1∗ / r d1 ∧ d2 (30) 2 log d under condition p ≥ 4µr 2 (d ∧d ) . 1 2 Putting (29) and (30) together, we obtain √ |||Y − M ∗ |||op ≤ 8αµrσ1∗ + c0 σ1∗ / r. Then using the fact that U0 V0> is the best rank r approximation of Y and applying Wely’s theorem (see the proof of Theorem 1 for a detailed argument), we have |||U0 V0> − M ∗ |||op ≤ |||U0 V0> − Y |||op + |||Y − M ∗ |||op √ ≤ 2|||Y − M ∗ |||op ≤ 16αµrσ1∗ + 2c0 σ1∗ / r 17 √ Under our assumptions, we have 16αµrσ1∗ + 2c0 σ1∗ / r ≤ 12 σr∗ . Accordingly, Lemma 15 gives d2 (U0 , V0 ; U ∗ , V ∗ ) ≤ √ 10r|||U0 V0> − M ∗ |||2op 2 |||U0 V0> − M ∗ |||2F ≤ . σr∗ σr∗ 2−1 We complete the proof by combining the above two inequalities. 6.5 Proof of Theorem 4 In this section, we turn to prove Theorem 4. Similar to the proof of Theorem 2, we rely on establishing the local descent and smoothness conditions. Compared to the full observation setting, e V ) := 1 |||U > U − V > V |||2 merely we replace L by Le given in (6), while the regularization term G(U, F 64 differs from G(U, V ) given in (20) by a constant factor. It is thus sufficient to analyze the properties e of L. Define E(M ∗ ) according to (18). Under the initial condition, we still have E(M ∗ ) ⊆ U × V, and U ⊆ U, V ⊆ V. (31) We prove the next two lemmas in Section 6.10 and 6.11 respectively. In both lemmas, for any (U, V ) ∈ U × V, we use shorthands (Uπ∗ , Vπ∗ ) = arg min (A,B)∈E(M ∗ ) |||U − A|||2F + |||V − B|||2F , ∆U := U − Uπ∗ , ∆V := V − Vπ∗ , and δ := |||∆U |||2F + |||∆V |||2F . Recall that d := max{d1 , d2 }. e Suppose U, V satisfy (31). Suppose we let Lemma 6 (Local descent property of L). i h  S = Tγpα ΠΦ Y − U V > , √ where we choose γ = 3. For any β > 0 and  ∈ (0, 41 ), we define ν := (14β +81)αµr +26 +18β −1 . There exist constants {ci }2i=1 such that if  2 2  µ r 1 log d p ≥ c1 + +1 , (32) 2  α d1 ∧ d2 then with probability at least 1 − c2 d−1 , q 3 > > 2 ∗ e V ; S), U V > − Uπ∗ Vπ>∗ + ∆U ∆> ∗ V ∗ ||| − νσ δ − 10 hh∇M L(U, ii ≥ |||U V − U σ1∗ δ 3 − 2δ 2 (33) π V π F 1 16 p ∗
for all (U, V ) ∈ (U × V) ∩ B2 σ1 . 
 e Suppose U, V satisfy (31). Suppose we let S = Tγαp ΠΦ Y − U V > Lemma 7 (Smoothness of L). for γ = 3. There exist constants {ci }3i=1 such that for any  ∈ (0, 41 ), when p satisfies condition p (32), with probability at least 1 − c2 d−1 , we have that for all (U, V ) ∈ (U × V) ∩ B2 ( σ1∗ ), i h e V ; S)|||2 + |||∇V L(U, e V ; S)|||2 ≤ c3 µrσ ∗ |||U V > − Uπ∗ V >∗ |||2 + µrσ ∗ δ(δ + σ ∗ ) . (34) |||∇U L(U, F 1 π F 1 1 F 18 In the remainder of this section, we condition on the events in Lemma 6 and 7. Now we are ready to prove Theorem 4. Proof of Theorem 4. We essentially follow the process for proving Theorem 2. Let the following shorthands be defined in the same fashion: δt , (Uπt ∗ , Vπt∗ ), (∆tU , ∆tV ), Let , Get . Here we show error decays in one step of iteration. The induction process is the same as the proof of Theorem 2, and is thus omitted. For any t ≥ 0, similar to (24) we have that δt+1 ≤ δt − 2η hh∇U Let + ∇U Get , Ut − Uπt ∗ ii −2η hh∇V Let + ∇V Get , Vt − Vπt∗ ii {z } | {z } | W1 W2 + η |||∇U Let + ∇U Get |||2F +η 2 |||∇V Let + ∇V Get |||2F . | | {z } {z } 2 W3 W4 We also have t t> hh∇U Let , Ut − Uπt ∗ ii + hh∇V Let , Vt − Vπt∗ ii = hh∇M Let , Ut Vt> − Uπt ∗ Vπt> ∗ + ∆U ∆V ii, which can be lower bounded by Lemma 6. Note that Ge differs from G by a constant, we can still leverage Lemma 3. Hence, we obtain that q 1 1 1 W1 + W2 ≥ |||Ut Vt> − M ∗ |||2F + |||Ut> Ut − Vt> Vt |||2F + σr∗ δt − νσ1∗ δt − 11 σ1∗ δt3 − 2δt2 . 8 64 64 On the other hand, we have W3 + W4 ≤ 2|||∇U Let |||2F + 2|||∇U Get |||2F + 2|||∇V Let |||2F + 2|||∇V Get |||2F h i ≤ c µrσ1∗ |||Ut Vt> − M ∗ |||2F + µrσ1∗ δt (δt + σ1∗ ) + σ1∗ |||Ut> Ut − Vt> Vt |||2F , where c is a constant, and the last step is implied by Lemma 4 and Lemma 7. By the assumption η = c0 /[µrσ1∗ ] for sufficiently small constant c0 , we thus have q 1 −2η(W1 + W2 ) + η 2 (W3 + W4 ) ≤ − ησr∗ δt + 2ηνσ1∗ δt + 22η σ1∗ δt3 + 4ηδt2 . 32 √ Recall that ν := (14β + 81)αµr + 26  + 18β −1 . By letting β = c1 κ,  = c2 /κ2 and assuming α ≤ c3 /(µrκ2 ) and δt ≤ c4 σr∗ /κ for some sufficiently small constants {ci }4i=1 , we can have −2η(W1 + 1 W2 ) + η 2 (W3 + W4 ) ≤ − 64 ησr∗ δt , which implies that   ησr∗ δt+1 ≤ 1 − δt , 64 and thus completes the proof. 6.6 Proof of Corollary 2 We need α . µκ12 r due to the condition of Theorem 4. Letting the initial error provided in Theorem 3 be less than the corresponding condition in Theorem 4, we have p p √ √ p 51 καµr r σ1∗ + 7c1  κσ1∗ . σr∗ /κ, 19 which leads to 1 1 α. √ , . √ . µ r3 κ3 κ3 Plugging the above two upper bounds into the second term in (13), it suffices to have p& κ3 µr2 log d . d1 ∧ d2 Comparing the above bound with the second term in (14) completes the proof. 6.7 Proof of Lemma 2 Let M := U V > . We observe that ∇M L(U, V ; S) = M + S − M ∗ − S ∗ . Plugging it back into the left hand side of (21), we obtain ∗ ∗ ∗ > hh∇M L(U, V ; S), U V > − Uπ∗ Vπ>∗ + ∆U ∆> V ii = hhM + S − M − S , M − M + ∆U ∆V ii ≥ |||M − M ∗ |||2F − |hhS − S ∗ , M − M ∗ ii| − |hhM + S − M ∗ − S ∗ , ∆U ∆> V ii| . {z } | {z } | T1 (35) T2 Next we derive upper bounds of T1 and T2 respectively. Upper bound of T1 . We denote the support of S, S ∗ by Ω and Ω∗ respectively. Since S − S ∗ is supported on Ω∗ ∪ Ω, we have T1 ≤ |hhΠΩ (S − S ∗ ), M − M ∗ ii| + |hhΠΩ∗ \Ω (S − S ∗ ), M − M ∗ ii| . | {z } | {z } W1 W2 Recall that for any (i, j) ∈ Ω, we have S(i,j) = (M ∗ + S ∗ − M )(i,j) . Accordingly, we have W1 = |||ΠΩ (M − M ∗ )|||2F . (36) Now we turn to bound W2 . Since S(i,j) = 0 for any (i, j) ∈ Ω∗ \ Ω, we have W2 = |hhΠΩ∗ \Ω S ∗ , M − M ∗ ii|. Let ui be the i-th row of M − M ∗ , and vj be the j-th column of M − M ∗ . For any k ∈ [d2 ], we let (k) ui denote the element of ui that has the k-th largest magnitude. Similarly, for any k ∈ [d1 ], we (k) let vj denote the element of vj that has the k-th largest magnitude. From the design of sparse estimator (4), we have that for any (i, j) ∈ Ω∗ \ Ω, |(M ∗ + S ∗ − M )(i,j) | is either smaller than the γαd2 -th largest entry of the i-th row of M ∗ + S ∗ − M or smaller than the γαd1 -th largest entry of the j-th column of M ∗ + S ∗ − M . Note that S ∗ only contains at most α-fraction nonzero entries per row and column. As a result, |(M ∗ + S ∗ − M )(i,j) | has to be less than (γαd2 −αd2 ) the magnitude of ui (γαd1 −αd1 ) . Formally, we have for (i, j) ∈ Ω∗ \ Ω, n o (γαd −αd2 ) (γαd −αd1 ) |, |vj 1 | . |(M ∗ + S ∗ − M )(i,j) | ≤ max |ui 2 | {z } or vj bij 20 (37) Furthermore, we obtain (γαd2 −αd2 ) 2 b2ij ≤ |ui (γαd1 −αd1 ) 2 | + |vj | ≤ kvj k22 kui k22 + . (γ − 1)αd2 (γ − 1)αd1 (38) Meanwhile, for any (i, j) ∈ Ω∗ \ Ω, we have ∗ · (M − M ∗ )(i,j) | = |(M ∗ + S ∗ − M − M ∗ + M )(i,j) · (M − M ∗ )(i,j) | |S(i,j) ≤ |(M − M ∗ )(i,j) |2 + |(M ∗ + S ∗ − M )(i,j) | · |(M − M ∗ )(i,j) | ≤ |(M − M ∗ )(i,j) |2 + bij · |(M − M ∗ )(i,j) |   b2ij β |(M − M ∗ )(i,j) |2 + , ≤ 1+ 2 2β (39) where β in the last step can be any positive number. Combining (38) and (39) leads to X ∗ · (M − M ∗ )(i,j) | W2 ≤ |S(i,j) (i,j)∈Ω∗ \Ω   X b2ij β ≤ 1+ |||ΠΩ∗ \Ω (M − M ∗ )|||2F + 2 2β (i,j)∈Ω∗ \Ω    X  kui k2 kvj k22 β 1 ∗ 2 2 ≤ 1+ |||ΠΩ∗ \Ω (M − M )|||F + + 2 2β (γ − 1)αd2 (γ − 1)αd1 ∗ (i,j)∈Ω \Ω   β 1 ≤ 1+ |||ΠΩ∗ \Ω (M − M ∗ )|||2F + |||M − M ∗ |||2F . 2 β(γ − 1) (40) In the last step, we use  X (i,j)∈Ω∗ \Ω ≤ 1 1 kui k22 + kvj k22 d2 d1 X X i∈[d] j∈Ω∗(i,·) ≤α X i∈[d] 1 kui k22 + d2 kui k22 + α X  ≤  X 1 1 2 2 kui k2 + kvj k2 d2 d1 ∗ (i,j)∈Ω X X j∈[d] i∈Ω∗(·,j) 1 kvj k22 d1 kvj k22 = 2α|||M − M ∗ |||2F . (41) j∈[d] We introduce shorthand δ := |||∆U |||2F + |||∆V |||2F . We prove the following inequality in the end of this section. p (42) |||M − M ∗ |||F ≤ 5σ1∗ δ. Combining (36), (40) and (42) leads to   β 5σ1∗ δ T1 ≤ |||ΠΩ (M − M ∗ )|||2F + 1 + |||ΠΩ∗ \Ω (M − M ∗ )|||2F + 2 β(γ − 1) ∗ 5σ1 δ ≤ 9(2γ + β + 2)αµrσ1∗ δ + , β(γ − 1) (43) where the last step follows from Lemma 14 by noticing that ΠΩ (M − M ∗ ) has at most γα-fraction nonzero entries per row and column. 21 Upper bound of T2 . To ease notation, we let C := M + S − M ∗ − S ∗ . We observe that C is supported on Ωc , we have > T2 ≤ |hhΠΩ∗c ∩Ωc (M − M ∗ ), ∆U ∆> V ii| + |hhΠΩ∗ ∩Ωc C, ∆U ∆V ii| . | {z } | {z } W3 W4 By Cauchy-Swartz inequality, we have ∗ W3 ≤ |||ΠΩ∗c ∩Ωc (M − M ∗ )|||F |||∆U ∆> V |||F ≤ |||M − M |||F |||∆U |||F |||∆V |||F ≤ q 5σ1∗ δ 3 /2, where the last step follows from (42) and |||∆U |||F |||∆V |||F ≤ δ/2. It remains to bound W4 . By Cauchy-Swartz inequality, we have ∗ ∗ > W4 ≤ |||ΠΩ∗ ∩Ωc C|||F |||∆U ∆> V |||F ≤ |||ΠΩ∗ ∩Ωc (M + S − M )|||F |||∆U ∆V |||F  1/2 s X 2 2 X (a) (b) kvj k2  kui k2 ≤ b2ij |||∆U |||F |||∆V |||F ≤  + |||∆U |||F |||∆V |||F . (γ − 1)αd2 (γ − 1)αd1 (i,j)∈Ω∗ \Ω (i,j)∈Ω∗ \Ω s r (c) 2 5σ1∗ δ 3 ∗ |||M − M |||F |||∆U |||F |||∆V |||F ≤ , ≤ γ−1 2(γ − 1) where step (a) is from (37), step (b) follows from (38), and step (c) follows from (41). Combining the upper bounds of W3 and W4 , we obtain s q 5σ1∗ δ 3 T2 ≤ 5σ1∗ δ 3 /2 + . (44) 2(γ − 1) Combining pieces. Now we choose γ = 2. Then inequality (43) implies that T1 ≤ [9(β + 6)αµr + 5β −1 ]σ1∗ δ. Inequality (44) then implies that q T2 ≤ 3 σ1∗ δ 3 . Plugging the above two inequalities into (35) completes the proof. Proof of inequality (42). We find that hp i2 σ1∗ (|||∆V |||F + |||∆U |||F ) + |||∆U |||F |||∆V |||F  2 p 1p ∗ 1p ∗ ∗ ≤ σ1 (|||∆V |||F + |||∆U |||F ) + σ1 |||∆U |||F + σ1 |||∆V |||F 2 2 |||M − M ∗ |||2F ≤ ≤ 5σ1∗ (|||∆U |||2F + |||∆V |||2F ), where the first step follows from the upper bound of |||M p − M ∗ |||F shown in Lemma 12, and the second step follows from the assumption |||∆U |||F , |||∆V |||F ≤ σ1∗ . 22 6.8 Proof of Lemma 3 We first observe that 1 1 ∇U G(U, V ) = U (U > U − V > V ), ∇V G(U, V ) = V (V > V − U > U ), 2 2 Therefore, we obtain hh∇U G(U, V ), U − Uπ∗ ii + hh∇V G(U, V ), V − Vπ∗ ii 1 = hhU > U − V > V , U > U − V > V − U > Uπ∗ + V > Vπ∗ ii 2 1 1 > = |||U U − V > V |||2F + hhU > U − V > V , U > U − V > V − 2U > Uπ∗ + 2V > Vπ∗ ii 4 4 1 > 1 > 2 > = |||U U − V V |||F + hhU > U − V > V , U > U − V > V − 2∆> U Uπ ∗ + 2∆V Vπ ∗ ii, 4 4 (45) > > > > > where the last step follows from ∆> U Uπ ∗ − ∆V Vπ ∗ = U Uπ ∗ − V Vπ ∗ since Uπ ∗ Uπ ∗ = Vπ ∗ Vπ ∗ . Note that U > U − V > V = (Uπ∗ + ∆U )> (Uπ∗ + ∆U ) − (Vπ∗ + ∆V )> (Vπ∗ + ∆V ) > > > > = Uπ>∗ ∆U + ∆> U Uπ ∗ + ∆U ∆U − Vπ ∗ ∆V − ∆V Vπ ∗ − ∆V ∆V , where we use Uπ>∗ Uπ∗ = Vπ>∗ Vπ∗ again in the last step. Furthermore, since U > U −V > V is symmetric, we have > > hhU > U − V > V , Uπ>∗ ∆U + ∆> U Uπ ∗ − Vπ ∗ ∆V − ∆V Vπ ∗ ii > = hhU > U − V > V , 2∆> U Uπ ∗ − 2∆V Vπ ∗ ii. Using these arguments, for the second term in (45), denoted by T2 , we have 1 > T2 = hhU > U − V > V , ∆> U ∆U − ∆V ∆V ii. 4 Furthermore, we have > > > 2 2 4T2 ≤ |hhU > U − V > V , ∆> U ∆U − ∆V ∆V ii| ≤ |||U U − V V |||F |||∆U |||F + |||∆V |||F   ≤ |||U > U − Uπ>∗ Uπ∗ |||F + |||V > V − Vπ>∗ Vπ∗ |||F δ q ≤ 2 (|||Uπ∗ |||op |||∆U |||F + |||Vπ∗ |||op |||∆V |||F ) δ ≤ 2 2σ1∗ δ 3 .
(46) It remains to find a lower bound of |||U > U − V > V |||F . The following inequality, which we turn to prove later, is true: |||U > U − V > V |||2F ≥ |||U U > − Uπ∗ Uπ>∗ |||2F + |||V V > − Vπ∗ Vπ>∗ |||2F − 2|||U V > − Uπ∗ Vπ>∗ |||2F . (47) Proceeding with the first term in (45) by using (47), we get 1 > 1 1 |||U U − V > V |||2F = |||U > U − V > V |||2F + |||U > U − V > V |||2F 4 8 8 1 > 1 1 1 ≥ |||U U − V > V |||2F + |||U U > − Uπ∗ Uπ>∗ |||2F + |||V V > − Vπ∗ Vπ>∗ |||2F − |||U V > − Uπ∗ Vπ>∗ |||2F 8 8 8 4 1 > 1 1 > 2 > > 2 > > 2 = |||U U − V V |||F + |||F F − Fπ∗ Fπ∗ |||F − |||U V − Uπ∗ Vπ∗ |||F , (48) 8 8 2 23 where we let     U Uπ∗ ∗ F := , Fπ := . V Vπ ∗ Introduce ∆F := F − Fπ∗ . Recall that δ := |||∆U |||2F + |||∆V |||2F . Equivalently δ = |||∆F |||2F . We have > > |||F F > − Fπ∗ Fπ>∗ |||F = |||∆F Fπ∗ + Fπ∗ ∆> F + ∆F ∆F |||F > 2 > > ≥ |||∆F Fπ∗ + Fπ∗ ∆> F |||F − |||∆F |||F = |||∆F Fπ∗ + Fπ ∗ ∆F |||F − δ. For the first term, we have > 2 > 2 > > |||∆F Fπ∗ + Fπ∗ ∆> F |||F = 2|||∆F Fπ∗ |||F + hh∆F Fπ∗ , Fπ ∗ ∆F ii > ∗ 2 > > ≥ 2σr (Fπ∗ )2 |||∆F |||2F + hh∆F Fπ∗ , F π ∗ ∆> F ii = 4σr |||∆F |||F + hh∆F Fπ∗ , Fπ ∗ ∆F ii. For the cross term, by the following result, proved in [11] (we also provide a proof in Section 7.5 for > , F ∗ ∆> ii ≥ 0. the sake of completeness), we have hh∆F Fπ∗ π F √ ∗ Lemma 8. When |||F − Fπ∗ |||op < 2σr , we have that ∆> F Fπ ∗ is symmetric. p p Accordingly, we have |||F F > −Fπ∗ Fπ>∗ |||F ≥ 2 σr∗ δ −δ ≥ σr∗ δ under condition δ ≤ σr∗ . Plugging this lower bound into (48), we obtain 1 > 1 1 1 |||U U − V > V |||2F ≥ |||U > U − V > V |||2F + σr∗ δ − |||U V > − Uπ∗ Vπ>∗ |||2F . 4 8 8 2 Putting (45), (46) and the above inequality together completes the proof. Proof of inequality (47). For the term on the left hand side of (47), it is easy to check that |||U > U − V > V |||2F = |||U U > |||2F + |||V V > |||2F − 2|||U V > |||2F . (49) The property Uπ>∗ Uπ∗ = Vπ>∗ Vπ∗ implies that |||Uπ∗ Uπ>∗ |||F = |||Vπ∗ Vπ>∗ |||F = |||Uπ∗ Vπ>∗ |||F . Therefore, expanding those quadratic terms on the right hand side of (47), one can show that it is equal to |||U U > |||2F + |||V V > |||2F − 2|||Uπ>∗ U |||2F − 2|||Vπ>∗ V |||2F + 4hhUπ>∗ U , Vπ>∗ V ii − 2|||U V > |||2F . Comparing inequalities (49) and (50), it thus remains to show that −2|||Uπ>∗ U |||2F − 2|||Vπ>∗ V |||2F + 4hhUπ>∗ U , Vπ>∗ V ii ≤ 0. Equivalently, we always have |||Uπ>∗ U − Vπ>∗ V |||2F ≥ 0, and thus prove (47). 6.9 Proof of Lemma 4 First, we turn to prove (23). As 1 1 ∇U G(U, V ) = U (U > U − V > V ), ∇V G(U, V ) = V (V > V − U > U ), 2 2 we have |||∇U G(U, V )|||2F + |||∇V G(U, V )|||2F ≤
1 |||U |||2op + |||V |||2op |||U > U − V > V |||2F . 4 24 (50) p p As (U, V ) p ∈ B2 ( σ1∗ ), we thus have |||U |||op ≤ |||Uπ∗ |||op + |||Uπ∗ − U |||op ≤ 2 σ1∗ , and similarly |||V |||op ≤ 2 σ1∗ . We obtain |||∇U G(U, V )|||2F + |||∇V G(U, V )|||2F ≤ 2σ1∗ |||U > U − V > V |||2F . Now we turn to prove (22). We observe that ∇M L(U, V ; S) = M + S − M ∗ − S ∗ , where we let M := U V > . We denote the support of S, S ∗ by Ω and Ω∗ respectively. Based on the sparse estimator (4) for computing S, ∇M L(U, V ; S) is only supported on Ωc . We thus have |||∇M L(U, V ; S)|||F ≤ |||ΠΩc \Ω∗ (M − M ∗ )|||F + |||ΠΩc ∩Ω∗ (M − M ∗ − S ∗ )|||F ≤ |||M − M ∗ |||F + |||ΠΩc ∩Ω∗ (M − M ∗ − S ∗ )|||F . It remains to upper bound the second term on the right hand side. Following (37) and (38), we have |||ΠΩc ∩Ω∗ (M − M ∗ − S ∗ )|||2F ≤ X (i,j)∈Ωc ∩Ω∗ kvj k22 2 kui k22 + ≤ |||M − M ∗ |||2F , (γ − 1)αd2 (γ − 1)αd1 γ−1 where the last step is proved in (41). By choosing γ = 2, we thus conclude that √ |||∇M L(U, V ; S)|||F ≤ (1 + 2)|||M − M ∗ |||F . 6.10 Proof of Lemma 6 We denote the support of ΠΦ (S ∗ ), S by Ω∗o and Ω. We always have Ω∗o ⊆ Φ and Ω ⊆ Φ. In the sequel, we establish several results that characterize the properties of Φ. The first result, proved in Section 7.2, shows that the Frobenius norm of any incoherent matrix whose row (or column) space are equal to L∗ (or R∗ ) is well preserved under partial observations supported on Φ. Lemma 9. Suppose M ∗ ∈ Rd1 ×d2 is a rank r and µ-incoherent matrix that has SVD M ∗ = log d L∗ Σ∗ R∗> . Then there exists an absolute constant c such that for any  ∈ (0, 1), if p ≥ c 2µr , (d1 ∧d2 ) then with probability at least 1 − 2d−3 , we have that for all A ∈ Rd2 ×r , B ∈ Rd1 ×r ,   (1 − )|||L∗ A> + BR∗> |||2F ≤ p−1 |||ΠΦ L∗ A> + BR∗> |||2F ≤ (1 + )|||L∗ A> + BR∗> |||2F . We need the next result, proved in Section 7.3, to control the number of nonzero entries per row and column in Ω∗o and Φ. Lemma 10. If p ≥ 56 log d 3 α(d1 ∧d2 ) , 1 |Φ(i,·) | − pd2 ≤ pd2 , 2 then with probability at least 1 − 6d−1 , we have 1 3 3 |Φ(·,j) | − pd1 ≤ pd1 , |Ω∗o(i,·) | ≤ αpd2 , |Ω∗o(·,j) | ≤ αpd1 , 2 2 2 for all i ∈ [d1 ] and j ∈ [d2 ]. 25 The next lemma, proved in Section 7.4, can be used to control the projection of small matrices to Φ. 2 2 log d Lemma 11. There exists constant c such that for any  ∈ (0, 1), if p ≥ c µ2 (dr 1 ∧d , then with 2) probability p at least 1 − O(d−1 ),pfor all matrices Z ∈ Rd1 ×d2 , U ∈ Rd1 ×r and V ∈ Rd2 ×r that satisfy |||U |||2,∞ ≤ µr/d1 ,|||V |||2,∞ ≤ µr/d2 , we have p−1 |||ΠΦ (U V > )|||2F ≤ |||U |||2F |||V |||2F + |||U |||F |||V |||F ; (51) p−1 |||ΠΦ (Z)V |||2F ≤ 2µr|||ΠΦ (Z)|||2F ; (52) p−1 |||U > ΠΦ (Z)|||2F ≤ 2µr|||ΠΦ (Z)|||2F . (53) In the remainder of this section, we condition on the events in Lemmas 9, 10 and 11. Now we are ready to prove Lemma 6. Proof of Lemma 6. Using shorthand M := U V > , we have e V ; S) = p−1 ΠΦ (M + S − M ∗ − S ∗ ) . ∇M L(U, Plugging it back into the left hand side of (21), we obtain e V ; S), U V > − Uπ∗ Vπ>∗ + ∆U ∆> hh∇M L(U, V ii 1 = hhΠΦ (M + S − M ∗ − S ∗ ), M − M ∗ + ∆U ∆> V ii p 1 1 1 ≥ |||ΠΦ (M − M ∗ ) |||2F − |hhΠΦ (S − S ∗ ), M − M ∗ ii| − |hhΠΦ (M + S − M ∗ − S ∗ ), ∆U ∆> V ii| . p p p | {z } | {z } | {z } T1 T2 T3 (54) Next we derive lower bounds of T1 , upper bounds of T2 and T3 respectively. > > Lower bound of T1 . We observe that M −M ∗ = Uπ∗∗ ∆> V +∆U Vπ ∗ +∆U ∆V . By triangle inequality, we have > > |||ΠΦ (M − M ∗ )|||F ≥ |||ΠΦ (Uπ∗ ∆> V + ∆U Vπ ∗ )|||F − |||ΠΦ (∆U ∆V )|||F . Note that when c ≥ a − b for a, b ≥ 0, we always have c2 ≥ 21 a2 − b2 . We thus have T1 ≥ ≥ ≥ ≥ ≥ 1 1 > 2 > 2 |||ΠΦ (Uπ∗ ∆> V + ∆U Vπ ∗ )|||F − |||ΠΦ (∆U ∆V )|||F 2p p 1 1 > 2 > 2 (1 − )|||Uπ∗ ∆> V + ∆U Vπ ∗ |||F − |||ΠΦ (∆U ∆V )|||F 2 p 1 2 2 2 ∗ (1 − )|||M − M ∗ − ∆U ∆> V |||F − |||∆U |||F |||∆V |||F − 9σ1 |||∆U |||F |||∆V |||F 2 1 1 (1 − )|||M − M ∗ |||2F − (1 − )|||∆U ∆V |||2F − |||∆U |||2F |||∆V |||2F − 9σ1∗ |||∆U |||F |||∆V |||F 4 2 1 ∗ 2 (1 − )|||M − M |||F − 2δ 2 − 5σ1∗ δ. 4 where the second step is implied by Lemma 9, the third p step follows from (51) in Lemma 11 by p ∗ noticing that |||∆U |||2,∞ ≤ 3 µrσ1 /d1 and |||∆V |||2,∞ ≤ 3 µrσ1∗ /d1 , which is further implied by (31). 26 Upper bound of T2 . Since S − S ∗ is supported on Ω∗0 ∪ Ω, we have pT2 ≤ |hhΠΩ∗o \Ω (S ∗ ), ΠΩ∗o \Ω (M − M ∗ )ii| + |hhΠΩ (S − S ∗ ), ΠΩ (M − M ∗ )ii|. (55) For any (i, j) ∈ Ω, we have (S − S ∗ )(i,j) = (M ∗ − M )(i,j) . Therefore, for the second term on the right hand side, we have |hhΠΩ (S − S ∗ ), ΠΩ (M − M ∗ )ii| ≤ |||ΠΩ (M − M ∗ )|||2F ≤ 18γpαµrσ1∗ δ, (56) where the last inequality follows from Lemma 14 and the fact that |Ω(i,·) | ≤ γpαd2 , |Ω(·,j) | ≤ γpαd1 for all i ∈ [d1 ], j ∈ [d2 ]. We denote the i-th row of ΠΦ (M − M ∗ ) by ui , and we denote the j-th column of ΠΦ (M − M ∗ ) (k) (k) by vj . We let ui denote the element of ui that has the k-th largest magnitude. We let vj denote the element of vj that has the k-th largest magnitude. For the first term on the right hand side of (55), we first observe that for (i, j) ∈ Ω∗o \ Ω, |(M ∗ + ∗ S − M )(i,j) | is either less than the γpαd2 -th largest element in the i-th row of ΠΦ (M ∗ + S ∗ − M ), or less than γpαd1 -th largest element in the j-th row of ΠΦ (M ∗ + S ∗ − M ). Based on Lemma 10, ΠΦ (S ∗ ) has at most 3pαd2 /2 nonzero entries per row and at most 3pαd1 /2 nonzero entries per column. Therefore, we have n o ((γ−1.5)pαd2 ) ((γ−1.5)pαd1 ) |(M ∗ + S ∗ − M )(i,j) | ≤ max |ui |, |vj | . (57) In addition, we observe that |hhΠΩ∗o \Ω (S ∗ ), ΠΩ∗o \Ω (M − M ∗ )ii| X ≤ |(M ∗ + S ∗ − M )(i,j) ||(M ∗ − M )(i,j) | + |(M ∗ − M )(i,j) |2 (i,j)∈Ω∗o \Ω  ≤ β 1+ 2  |||ΠΩ∗o (M ∗ − M )|||2F + ≤ (27 + 14β)pαµrσ1∗ δ + 1 2β 1 2β X |(M ∗ + S ∗ − M )(i,j) |2 , (i,j)∈Ω∗o \Ω X |(M ∗ + S ∗ − M )(i,j) |2 , (58) (i,j)∈Ω∗o \Ω where the second step holds for any β > 0 and the last step follows from Lemma 14 under the size constraints of Ω∗o shown in Lemma 10. For the second term in (58), using (57), we have X X ((γ−1.5)pαd2 ) 2 ((γ−1.5)pαd1 ) 2 |(M ∗ + S ∗ − M )(i,j) |2 ≤ |ui | + |vj | (i,j)∈Ω∗o \Ω = X (i,j)∈Ω∗o X ((γ−1.5)pαd2 ) 2 |ui | i∈[d1 ] j∈Ω∗o(i,·) ≤ X i∈[d1 ] + X X ((γ−1.5)pαd1 ) 2 |vj | j∈[d2 ] i∈Ω∗o(·,j) X 1.5 1.5 3 kui k22 + kvj k22 ≤ |||ΠΦ (M − M ∗ )|||2F . γ − 1.5 γ − 1.5 γ − 1.5 (59) j∈[d2 ] Moreover, we have > 2 > 2 |||ΠΦ (M − M ∗ )|||2F ≤ 2|||ΠΦ (Uπ∗ ∆> V + ∆U Vπ ∗ )|||F + 2|||ΠΦ (∆U ∆V )|||F > 2 2 2 ∗ ≤ 2(1 + )p|||Uπ∗ ∆> V + ∆U Vπ ∗ |||F + 2p|||∆U |||F |||∆V |||F + 18pσ1 |||∆U |||F |||∆V |||F
≤ 4(1 + )p |||Uπ∗ |||2op |||∆V |||2F + |||Vπ∗ |||2op |||∆U |||2F + 2p|||∆U |||2F |||∆V |||2F + 18pσ1∗ |||∆U |||F |||∆V |||F ≤ (4 + 13)pσ1∗ δ + 2pδ 2 , (60) 27 where the second step follows from Lemma 9 and inequality (51) in Lemma 11. Putting (55)-(60) together, we obtain T2 ≤ (18γ + 14β + 27)αµrσ1∗ δ + 3[(2 + 7)σ1∗ δ + δ 2 ] . β(γ − 1.5) Upper bound of T3 . By Cauchy-Schwarz inequality, we have pT3 ≤ |||ΠΦ (M − M ∗ + S − S ∗ )|||F |||ΠΦ (∆U ∆> V )|||F q ≤ |||ΠΦ (M − M ∗ + S − S ∗ )|||F p|||∆U |||2F |||∆V |||2F + 9pσ1∗ |||∆U |||F |||∆V |||F q ≤ |||ΠΦ (M − M ∗ + S − S ∗ )|||F pδ 2 + 5pσ1∗ δ. where we use (51) in Lemma 11 in the second step. We observe that ΠΦ (M − M ∗ + S − S ∗ ) is supported on Φ \ Ω. Therefore, we have |||ΠΦ (M − M ∗ + S − S ∗ )|||F ≤ |||ΠΦ∩Ωc ∩Φ∗c (M − M ∗ )|||F + |||ΠΦ∩Ωc ∩Φ∗ (M − M ∗ − S ∗ )|||F ≤ |||ΠΦ (M − M ∗ )|||F + |||ΠΩc ∩Φ∗ (M − M ∗ − S ∗ )|||F r 3 ∗ ≤ |||ΠΦ (M − M )|||F + |||ΠΦ (M − M ∗ )|||F γ − 1.5 r  q 3 (4 + 13)pσ1∗ δ + 2pδ 2 , ≤ 1+ γ − 1.5 where the third step follows from (59), and the last step is from (60). Under assumptions γ = 3,  ≤ 1/4 and δ ≤ σ1∗ , we have q q q √ T3 ≤ 3 9σ1∗ δ + 2δ 2 δ 2 + 5σ1∗ δ ≤ 10 σ1∗ δ 3 + 23 σ1∗ δ. Combining pieces. Under the aforementioned assumptions, putting all pieces together leads to e V ; S), U V > − Uπ∗ V >∗ + ∆U ∆> ii hh∇M L(U, π V   q √ 3 18 ∗ 2 ∗ ≥ |||M − M |||F − (14β + 81)αµrσ1 δ − 26  + σ1∗ δ − 10 σ1∗ δ 3 − 2δ 2 . 16 β 6.11 Proof of Lemma 7 Let M := U V > . We find that e V ; S) = p−1 ΠΦ (M + S − M ∗ − S ∗ ) V, ∇U L(U, e V ; S) = p−1 ΠΦ (M + S − M ∗ − S ∗ )> U. ∇V L(U, Conditioning on the event in Lemma 11, since (U, V ) ∈ U × V, inequalities (52) and (53) imply that e V ; S)|||2F + |||∇V L(U, e V ; S)|||2F ≤ 12 µrσ1∗ |||ΠΦ (M + S − M ∗ − S ∗ ) |||2F . |||∇U L(U, p 28 It remains to bound the term |||ΠΦ (M + S − M ∗ − S ∗ ) |||2F . Let Ω∗o and Ω be the support of ΠΦ (S ∗ ) and S respectively. We observe that |||ΠΦ (M + S − M ∗ − S ∗ ) |||2F = |||ΠΩ∗o \Ω (M − M ∗ − S ∗ ) |||2F + |||ΠΦ∗c ∩Ωc ∩Φ (M − M ∗ ) |||2F ≤ |||ΠΩ∗o \Ω (M − M ∗ − S ∗ ) |||2F + |||ΠΦ (M − M ∗ ) |||2F . In the proof of Lemma 6, it is shown in (59) that |||ΠΩ∗o \Ω (M − M ∗ − S ∗ ) |||2F ≤ 3 |||ΠΦ (M − M ∗ )|||2F . γ − 1.5 Moreover, following (60), we have that > 2 2 2 ∗ |||ΠΦ (M − M ∗ )|||2F ≤ 2(1 + )p|||Uπ∗ ∆> V + ∆U Vπ ∗ |||F + 2p|||∆U |||F |||∆V |||F + 18pσ1 |||∆U |||F |||∆V |||F ≤ 4(1 + )p|||M − M ∗ |||2F + (6 + 4)p|||∆U |||2F |||∆V |||2F + 18pσ1∗ |||∆U |||F |||∆V |||F ≤ 4(1 + )p|||M − M ∗ |||2F + (6 + 4)pδ 2 + 9pσ1∗ δ. We thus finish proving our conclusion by combining all pieces and noticing that γ = 3 and  ≤ 1/4. 7 Proofs for Technical Lemmas In this section, we prove several technical lemmas that are used in the proofs of our main theorems. 7.1 Proof of Lemma 1 We observe that |||A|||op = sup sup x> Ay. x∈Sd1 −1 y∈Sd2 −1 We denote the support of A by Ω. For any x ∈ Rd1 , y ∈ Rd2 and β > 0, we have X 1 kAk∞ (β −1 x2i + βyj2 ) 2 (i,j)∈Ω (i,j)∈Ω   X X X X 1 = kAk∞  β −1 x2i + βyj2  2 x> Ay = X xi A(i,j) yj ≤ i j∈Ω(i,·) j i∈Ω(·,j)
1 ≤ kAk∞ αd2 β −1 kxk22 + αd1 βkyk22 . 2 It is thus implied that |||A|||op ≤ 21 α(β −1 d2 + βd1 )kAk∞ . Choosing β = 7.2 p d2 /d1 completes the proof. Proof of Lemma 9 We define a subspace K ⊆ Rd1 ×d2 as n o K := X X = L∗ A> + BR∗> for some A ∈ Rd2 ×r , B ∈ Rd1 ×r . 29 Let ΠK be Euclidean projection onto K. Then according to Theorem 4.1 in [6], under our assumptions, for all matrices X ∈ Rd1 ×d2 , inequality p−1 ||| (ΠK ΠΦ ΠK − pΠK ) X|||F ≤ |||X|||F (61) holds with probability at least 1 − 2d−3 . In our setting, by restricting X = L∗ A> + BR∗> , we have ΠK X = X. Therefore, (61) implies that |||ΠK ΠΦ X − pX|||F ≤ p|||X|||F . For |||ΠΦ X|||2F , we have |||ΠΦ X|||2F = hhΠΦ X, ΠΦ Xii = hhΠΦ X, Xii = hhΠK ΠΦ X, Xii ≤ |||ΠK ΠΦ X|||F |||X|||F ≤ p(1 + )|||X|||2F . On the other hand, we have |||ΠΦ X|||2F = hhΠK ΠΦ X, Xii = hhΠK ΠΦ X − pX + pX, Xii = p|||X|||2F − hhX, −ΠK ΠΦ X + pXii ≥ p|||X|||2F − |||X|||F |||ΠK ΠΦ X − pX|||F ≥ p(1 − )|||X|||2F . Combining the above two inequalities, we complete the proof. 7.3 Proof of Lemma 10 We observe that |Φ(i,·) | is a summation of d2 i.i.d. binary random variables with mean p and variance p(1 − p). By Bernstein’s inequality, for any i ∈ [d1 ], !     − 21 (pd2 /2)2 1 3 Pr |Φ(i,·) | − pd2 ≥ pd2 ≤ 2 exp − ≤ 2 exp − pd 2 . 2 28 d2 p(1 − p) + 13 (pd2 /2) By probabilistic union bound, we have " Pr sup |Φ(i,·) | − pd2 i∈[d1 ] #   1 3 ≥ pd2 ≤ 2d1 exp − pd2 ≤ 2d−1 , 2 28 log d where the last inequality holds by assuming p ≥ 56 3 d2 . The term |Ω∗o(i,·) | is a summation of at most αd2 i.i.d. binary random variables with mean p and variance p(1 − p). Again, applying Bernstein’s inequality leads to     h i 1 3 ∗ ∗ Pr |Ωo(i,·) | − E |Ωo(i,·) | ≥ pαd2 ≤ exp − pαd2 . 2 28 Accordingly, by the assumption p ≥ " 56 log d 3 αd2 , we obtain #   3 1 ∗ Pr sup |Ωo(i,·) | − pk ≥ pk ≤ d1 exp − pαd2 ≤ d−1 . 2 28 i∈[d1 ] The proofs for |Φ(·,j) | and |Ω∗o(·,j) | follow the same idea. 30 7.4 Proof of Lemma 11 d d1 ×d2 , we According to Lemma 3.2 in [8], under condition p ≥ c1 µd1log ∧d2 , for any fixed matrix A ∈ R have s d log d |||A − p−1 ΠΦ A|||op ≤ c2 kAk∞ , p holds with probability at least 1 − O(d−3 ). Letting A be all-ones matrix, then we have that for all u ∈ Rd1 , v ∈ Rd2 , X p ui vj ≤ pkuk1 kvk1 + c2 pd log dkuk2 kvk2 . (i,j)∈Φ We find that |||ΠΦ (U V > )|||2F ≤ X kU(i,·) k22 kV(j,·) k22 (i,j)∈Φ sX sX p 4 ≤ p|||U |||F |||V |||F + c2 pd log d kU(i,·) k2 kV(j,·) k42 2 2 i∈[d1 ] j∈[d2 ] p ≤ p|||U |||2F |||V |||2F + c2 pd log d|||U |||F |||V |||F |||U |||2,∞ |||V |||2,∞ s pµ2 r2 d log d ≤ p|||U |||2F |||V |||2F + c2 |||U |||F |||V |||F . d1 d2 By the assumption p & µ2 r2 log d , 2 (d1 ∧d2 ) we finish proving (51). d According to the proof of Lemma 10, if p ≥ c dlog , with probability at least 1 − O(d−1 ), we 1 ∧d2 3 3 have |Φ(i,·) | ≤ 2 pd2 and |Φ(·,j) | ≤ 2 pd1 for all i ∈ [d1 ] and j ∈ [d2 ]. Conditioning on this event, we have X X |||ΠΦ (Z)V |||2F = h(ΠΦ (Z))(i,·) , H(·,k) i2 i∈[d1 ] k∈[r] ≤ X X k(ΠΦ (Z))(i,·) k22 2 V(j,k) j∈Ω(i,·) i∈[d1 ] k∈[r] = |||ΠΦ Z|||2F X X kV(i,·) k22 j∈Ω(i,·) 3 ≤ |||ΠΦ Z|||2F pd2 · |||V |||22,∞ ≤ 2µrp|||ΠΦ Z|||2F . 2 We thus finish proving (52). Inequality (53) can be proved in the same way. 7.5 Proof of Lemma 8 Recall that we let F := [U ; V ] and Fπ∗ := [U ∗ ; V ∗ ]Q for some matrix Q ∈ Qr , which minimizes the following function |||F − [U ∗ ; V ∗ ]Q|||2F . (62) Let F ∗ := [U ∗ ; V ∗ ]. Expanding the above term, we find that Q is the maximizer of hhF, F ∗ Qii = Tr(F > F ∗ Q). Suppose F > F ∗ has SVD with form Q1 ΛQ> 2 for Q1 , Q2 ∈ Qr . When the minimum 31 diagonal term of Λ is positive, we conclude that the minimizer of (62) is unique and Q = Q2 Q> 1. To prove this argument, we note that X Λ(i,i) hpi , qi i, Tr(F > F ∗ Q) = i∈[r] where pi is the i-th column of Q1 and qi is the i-th column of Q> Q2 . Hence, Tr(F > F ∗ Q) ≤ P i∈[r] Λ(i,i) and the equality holds if and only if pi = qi for all i ∈ [r] since every Λ(i,i) > 0. We have Q1 = Q> Q2 and thus finish proving the argument. √ Under our assumption |||F − Fπ∗ |||op < 2σr∗ , for any nonzero vector u ∈ Rr , we have p kF > Fπ∗ uk2 ≥ kFπ>∗ Fπ∗ uk2 − k(Fπ∗ − F )> Fπ∗ uk2 ≥ ( 2σr∗ − |||Fπ∗ − F |||op )|||Fπ∗ u|||F > 0. In √ the second step, we use the fact that the singular values of Fπ∗ are equal to the diagonal terms of 2Σ∗1/2 . Hence, F > Fπ∗ has full rank. Furthermore, it implies that F > F ∗ has full rank and only contains positive singular values. Proceeding with the proved argument, we have > > F > Fπ∗ = Q1 ΛQ> 2 Q2 Q1 = Q1 ΛQ1 , which implies that F > Fπ∗ is symmetric. Accordingly, we have (F − Fπ∗ )> Fπ∗ is also symmetric. Acknowledgment Y. Chen acknowledges support from the School of Operations Research and Information Engineering, Cornell University. References [1] Animashree Anandkumar, Rong Ge, Daniel Hsu, Sham M. Kakade, and Matus Telgarsky. “Tensor decompositions for learning latent variable models”. In The Journal of Machine Learning Research 15.1 (2014), pp. 2773–2832. [2] Sivaraman Balakrishnan, Martin J. Wainwright, and Bin Yu. “Statistical guarantees for the EM algorithm: From population to sample-based analysis”. In arXiv preprint arXiv:1408.2156 (2014). [3] Srinadh Bhojanapalli, Prateek Jain, and Sujay Sanghavi. “Tighter low-rank approximation via sampling the leveraged element”. In Proceedings of the Twenty-Sixth Annual ACM-SIAM Symposium on Discrete Algorithms. SIAM. 2015, pp. 902–920. [4] Srinadh Bhojanapalli, Anastasios Kyrillidis, and Sujay Sanghavi. “Dropping convexity for faster semi-definite optimization”. In arXiv preprint arXiv:1509.03917 (2015). [5] Emmanuel J. Candès, Xiaodong Li, and Mahdi Soltanolkotabi. “Phase retrieval via Wirtinger flow: Theory and algorithms”. In IEEE Transactions on Information Theory 61.4 (2015), pp. 1985–2007. [6] Emmanuel J. Candès and Benjamin Recht. “Exact matrix completion via convex optimization”. In Foundations of Computational mathematics 9.6 (2009), pp. 717–772. 32 [7] Emmanuel J. Candès and Terence Tao. “The power of convex relaxation: Near-optimal matrix completion”. In IEEE Transactions on Information Theory 56.5 (2010), pp. 2053–2080. [8] Emmanuel J. Candès, Xiaodong Li, Yi Ma, and John Wright. “Robust principal component analysis?” In Journal of the ACM (JACM) 58.3 (2011), p. 11. [9] Venkat Chandrasekaran, Sujay Sanghavi, Pablo A. Parrilo, and Alan S. Willsky. “Ranksparsity incoherence for matrix decomposition”. In SIAM Journal on Optimization 21.2 (2011), pp. 572–596. [10] Yudong Chen. “Incoherence-Optimal Matrix Completion”. In IEEE Transactions on Information Theory 61.5 (2015), pp. 2909–2923. [11] Yudong Chen and Martin J. Wainwright. “Fast low-rank estimation by projected gradient descent: General statistical and algorithmic guarantees”. In arXiv preprint arXiv:1509.03025 (2015). [12] Yudong Chen, Ali Jalali, Sujay Sanghavi, and Constantine Caramanis. “Low-rank Matrix Recovery from Errors and Erasures”. In IEEE Transactions on Information Theory 59.7 (2013), pp. 4324–4337. [13] Yuxin Chen and Emmanuel J. Candès. “Solving random quadratic systems of equations is nearly as easy as solving linear systems”. In Advances in Neural Information Processing Systems. 2015, pp. 739–747. [14] Yeshwanth Cherapanamjeri, Kartik Gupta, and Prateek Jain. “Nearly-optimal Robust Matrix Completion”. In arXiv preprint arXiv:1606.07315 (2016). [15] Kenneth L. Clarkson and David P. Woodruff. “Low rank approximation and regression in input sparsity time”. In Proceedings of the forty-fifth annual ACM symposium on Theory of computing. ACM. 2013, pp. 81–90. [16] Alan Frieze, Ravi Kannan, and Santosh Vempala. “Fast Monte-Carlo algorithms for finding low-rank approximations”. In Journal of the ACM (JACM) 51.6 (2004), pp. 1025–1041. [17] Quanquan Gu, Zhaoran Wang, and Han Liu. “Low-Rank and Sparse Structure Pursuit via Alternating Minimization”. In Proceedings of the 19th International Conference on Artificial Intelligence and Statistics. 2016, pp. 600–609. [18] Moritz Hardt. “Understanding alternating minimization for matrix completion”. In 2014 IEEE 55th Annual Symposium on Foundations of Computer Science (FOCS). IEEE. 2014, pp. 651– 660. [19] Daniel Hsu, Sham M. Kakade, and Tong Zhang. “Robust matrix decomposition with sparse corruptions”. In IEEE Transactions on Information Theory 57.11 (2011), pp. 7221–7234. [20] Prateek Jain, Praneeth Netrapalli, and Sujay Sanghavi. “Low-rank matrix completion using alternating minimization”. In Proceedings of the forty-fifth annual ACM symposium on Theory of computing. ACM. 2013, pp. 665–674. [21] Zhouchen Lin, Minming Chen, and Yi Ma. “The Augmented Lagrange Multiplier Method for Exact Recovery of Corrupted Low-Rank Matrices”. In Arxiv preprint arxiv:1009.5055v3 (2013). 33 [22] Praneeth Netrapalli, UN Niranjan, Sujay Sanghavi, Animashree Anandkumar, and Prateek Jain. “Non-convex robust PCA”. In Advances in Neural Information Processing Systems. 2014, pp. 1107–1115. [23] Yousef Saad. Numerical Methods for Large Eigenvalue Problems: Revised Edition. Vol. 66. Siam, 2011. [24] Ju Sun, Qing Qu, and John Wright. “When Are Nonconvex Problems Not Scary?” In arXiv preprint arXiv:1510.06096 (2015). [25] Ruoyu Sun and Zhi-Quan Luo. “Guaranteed matrix completion via nonconvex factorization”. In 2015 IEEE 56th Annual Symposium on Foundations of Computer Science (FOCS). IEEE. 2015, pp. 270–289. [26] Stephen Tu, Ross Boczar, Mahdi Soltanolkotabi, and Benjamin Recht. “Low-rank solutions of linear matrix equations via procrustes flow”. In arXiv preprint arXiv:1507.03566 (2015). [27] Zhaoran Wang, Quanquan Gu, Yang Ning, and Han Liu. “High Dimensional EM Algorithm: Statistical Optimization and Asymptotic Normality”. In Advances in Neural Information Processing Systems. 2015, pp. 2512–2520. [28] Huan Xu, Constantine Caramanis, and Sujay Sanghavi. “Robust PCA via Outlier Pursuit”. In IEEE Transactions on Information Theory 58.5 (2012), pp. 3047–3064. issn: 0018-9448. [29] Xinyang Yi and Constantine Caramanis. “Regularized EM Algorithms: A Unified Framework and Statistical Guarantees”. In Advances in Neural Information Processing Systems. 2015, pp. 1567–1575. [30] Huishuai Zhang, Yuejie Chi, and Yingbin Liang. “Provable Non-convex Phase Retrieval with Outliers: Median Truncated Wirtinger Flow”. In arXiv preprint arXiv:1603.03805 (2016). [31] Tuo Zhao, Zhaoran Wang, and Han Liu. “A Nonconvex Optimization Framework for Low Rank Matrix Estimation”. In Advances in Neural Information Processing Systems. 2015, pp. 559– 567. [32] Qinqing Zheng and John Lafferty. “A convergent gradient descent algorithm for rank minimization and semidefinite programming from random linear measurements”. In Advances in Neural Information Processing Systems. 2015, pp. 109–117. 34 Appendices A Supporting Lemmas In this section, we provide several technical lemmas used for proving our main results. Lemma 12. For any (U ∗ , V ∗ ) ∈ E(M ∗ ), U ∈ Rd1 ×r and V ∈ Rd2 ×r , we have p |||U V > − U ∗ V ∗> |||F ≤ σ1∗ (|||∆V |||F + |||∆U |||F ) + |||∆U |||F |||∆V |||F , where ∆U := U − U ∗ , ∆V := V − V ∗ . ∗> + ∆ ∆> . Hence, Proof. We observe that U V > − U ∗ V ∗> = U ∗ ∆> U V V + ∆U V ∗> |||U V > − U ∗ V ∗> |||F ≤ |||U ∗ ∆> |||F + |||∆U ∆> V |||F + |||∆U V V |||F ≤ |||U ∗ |||op |||∆V |||F + |||V ∗ |||op |||∆U |||F + |||∆U |||F |||∆V |||F . Furthermore, assuming (U, V ) ∈ U × V, where U and V satisfy the conditions in (19), we have the next result. Lemma 13. For any (i, j) ∈ [d1 ] × [d2 ], we have r r µrσ1∗ µrσ1∗ > ∗ ∗> |(U V − U V )(i,j) | ≤ 3 k∆V (j,·) k2 + 3 k∆U (i,·) k2 d1 d2 (63) Proof. We observe that ∗ ∗ |(U V > − M ∗ )(i,j) | ≤ |hU(i,·) , ∆V (j,·) i| + |hV(j,·) , ∆U (i,·) i| + |h∆U (i,·) , ∆V (j,·) i| r r µrσ1∗ µrσ1∗ 1 1 ≤ k∆V (j,·) k2 + k∆U (i,·) k2 + |||∆U |||2,∞ k∆V (j,·) k2 + |||∆V |||2,∞ k∆U (i,·) k2 . d1 d2 2 2 By noticing that r ∗ |||∆U |||2,∞ ≤ |||U |||2,∞ + |||U |||2,∞ ≤ 3 r µrσ1∗ µrσ1∗ ∗ , |||∆V |||2,∞ ≤ |||V |||2,∞ + |||V |||2,∞ ≤ 3 , d1 d2 we complete the proof. Lemma 13 can be used to prove the following result. Lemma 14. For any α ∈ [0, 1], suppose Ω ⊆ [d1 ] × [d2 ] satisfies |Ω(i,·) | ≤ αd2 for all i ∈ [d1 ] and |Ω(·,j) | ≤ αd1 for all j ∈ [d2 ]. Then we have |||ΠΩ (U V > − U ∗ V ∗> )|||2F ≤ 18αµrσ1∗ (|||∆V |||2F + |||∆U |||2F ). 35 Proof. Using Lemma 13 for bounding each entry of U V > − U ∗ V ∗> , we have that X |(U V > − U ∗ V ∗> )(i,j) |2 |||ΠΩ (U V > − U ∗ V ∗> )|||2F ≤ (i,j)∈Ω X 18µrσ ∗ 18µrσ1∗ 1 ≤ k∆V (j,·) k22 + k∆U (i,·) k22 d1 d2 (i,j)∈Ω X X 18µrσ ∗ X X 18µrσ ∗ 1 1 ≤ k∆V (j,·) k22 + k∆U (i,·) k22 d1 d2 j ≤ i i∈Ω(·,j) 18αµrσ1∗ (|||∆V |||2F j∈Ω(i,·) 2 + |||∆U |||F ). Denote the i-th largest singular value of matrix M by σi (M ). Lemma 15 (Lemma 5.14 in [26]). Let M1 , M2 ∈ Rd1 ×d2 be two rank r matrices. Suppose they have SVDs M1 = L1 Σ1 R1> and M2 = L2 Σ2 R2> . Suppose |||M1 − M2 |||op ≤ 21 σr (M1 ). Then we have 1/2 1/2 1/2 1/2 d2 (L2 Σ2 , R2 Σ2 ; L1 Σ1 , R1 Σ1 ) ≤ √ B 2 |||M2 − M1 |||2F . 2 − 1 σr (M1 ) Parameter Settings for FB Separation Experiments We approximate the FB separation problem by the RPCA framework with r = 10, α = 0.2, µ = 10. Our algorithmic parameters are set as γ = 1, η = 1/(2σ̂1∗ ), where σ̂1∗ is an estimate of σ1∗ obtained from the initial SVD. The parameters of AltProj are kept as provided in the default setting. For √ IALM, we use the tradeoff paramter λ = 1/ d1 , where d1 is the number of pixels in each frame (the number of rows in Y ). Note that both IALM and AltProj use the stopping criterion |||Y − Mt − St |||F /|||Y |||F ≤ 10−3 . Our algorithm for the partial observation setting never explicitly forms the d1 -by-d2 matrix Mt = Ut Vt> , which is favored in large scale problems, but also renders the above criterion inapplicable. Instead, we use the following stopping criterion |||Ut+1 − Ut |||2F + |||Vt+1 − Vt |||2F ≤ 4 × 10−4 . |||Ut |||2F + |||Vt |||2F This rule checks whether the iterates corresponding to low-rank factors become stable. In fact, our stopping criterion seems more natural and practical because in most real applications, matrix Y cannot be strictly decomposed into low-rank M and sparse S that satisfy Y = M + S. Instead of forcing M + S to be close to Y , our rule relies on seeking a robust subspace that captures the most variance of Y . 36 

10math.ST

Square-free class sizes in products of groups arXiv:1703.08363v1 [math.GR] 24 Mar 2017 M. J. Felipe · A. Martı́nez-Pastor · V. M. Ortiz-Sotomayor ∗ Abstract We obtain some structural properties of a factorised group G = AB, given that the conjugacy class sizes of certain elements in A ∪ B are not divisible by p2 , for some prime p. The case when G = AB is a mutually permutable product is especially considered. Keywords Finite groups · Soluble groups · Products of subgroups · Conjugacy classes 2010 MSC 20D10 · 20D40 · 20E45 1 Introduction All groups considered thorogouth this paper are finite. Over the last years, many authors have investigated the influence of conjugacy class sizes on the structure of finite groups. In the meantime, numerous studies in the framework of group theory have focused in factorised groups. In this setting, a central question is how the structure of the factors affects the structure of the whole group, in particular when they are connected by certain permutability properties. The purpose of this paper is to show new achievements which combine both current perspectives in finite groups. More precisely, our aim is to get some information about a factorised group, provided that the conjugacy class sizes of some elements of its factors are square-free. The earlier starting point of our investigation can be traced back to the paper of Chillag and Herzog ([5]), where the structure of a group in which all elements have square-free conjugacy class sizes was first analysed. Next, in [6], Cossey and Wang localised one of the main theorems in [5] for a fixed prime p, that is, they considered conjugacy class sizes not divisible by p2 , for certain prime p. Later on, this study was ∗ The first author is supported by Proyecto Prometeo II/2015/011, Generalitat Valenciana (Spain), and the second author is supported by Proyecto MTM2014-54707-C3-1-P, Ministerio de Economı́a, Industria y Competitividad (Spain). The results in this paper are part of the third author’s Ph.D. thesis, and he acknowledges the predoctoral grant ACIF/2016/170, Generalitat Valenciana (Spain). The authors wish to thank John Cossey for helpful conversations during his last visit to Valencia. Instituto Universitario de Matemática Pura y Aplicada (IUMPA), Universitat Politècnica de València, Camino de Vera, s/n, 46022, Valencia, Spain. B: mfelipe@mat.upv.es, anamarti@mat.upv.es, vicorso@doctor.upv.es 1 improved by Li in [12], and by Liu, Wang, and Wei in [13], by replacing conditions on all conjugacy classes by those referring only to conjugacy classes of either p-regular elements or prime power order elements, using the classification theorem of finite simple groups (CFSG). These authors also first obtained some preliminary results in factorised groups. This research was extended in 2012 by Ballester-Bolinches, Cossey and Li in [2], through mutually permutable products. More recently, in 2014, Qian and Wang ([14]) have gone a step further by considering just conjugacy class sizes of p-regular elements of prime power order (although not in factorised groups). In the context of factorised groups, and aiming to obtain criteria for products of supersoluble subgroups to be supersoluble, several authors have considered products in which certain subgroups of the factors permute (see [3] for a detailed account). In this scene, we are interested in mutually permutable products, factorised groups G = AB such that the subgroups A and B are mutually permutable, i.e., A permutes with every subgroup of B and B permutes with every subgroup of A (see also [4]). Obviously, if A and B are normal in G, then they are mutually permutable. We recall that, for a group G, the set xG = {g −1 xg : g ∈ G} is the conjugacy class of the element x ∈ G, and xG denotes the conjugacy class size of x. If p is a prime number, we say that x ∈ G is a p-regular element if its order is not divisible by p, and that it is a p-element if its order is a power of p. Moreover, if n is an integer, let np denote the highest power of p dividing n. The mth group of order n in the SmallGroups library [8] of GAP will be identified by n#m. The remainder notation is standard and is taken mainly from [7]. We also refer to this book for details about classes of groups. In this paper, motivated by the above development, at first we focus on the case of p-groups, extending for factorised groups the well-known Knoche’s theorem (see [11]). Theorem A. Let p be a prime number and let P = AB be a p-group such that p2 does not divide xP for all x ∈ A ∪ B. Then P ′ 6 Φ(P ) 6 Z(P ), P ′ is elementary abelian and |P ′ | ≤ p2 . Our next goal in the paper is to prove the following theorem, regarding mutually permutable products. Theorem B. Let G = AB be the mutually permutable product of the subgroups A and B, and let p be a prime such that gcd(p − 1, |G|) = 1. If p2 does not divide xG for any p-regular element x ∈ A ∪ B of prime power order, then: (1) G is soluble. (2) G is p-nilpotent. (3) The Sylow p-subgroups of G/ Op (G) are elementary abelian. In the particular case when G = A = B, we recover [14, Theorem A] (see Section 3, Corollary 3.5). We remark that both results use the CFSG. Moreover, in relation to the third assertion, we have found a gap in one of the statements in [6, Theorem 1], as it is reported in Remark 3.7 (a). On the other hand, we point out that it is possible to find examples of groups factorised as a product of two (mutually permutable) subgroups which satisfy the hypothe2 ses of Theorem B for some fixed prime p and, however, there exist elements x ∈ A ∪ B such that p2 divides either xA or xB (see Remark 3.7 (b)). The next theorem generalises the last assertion of [2, Theorem 1.3] regarding p-soluble groups, by considering only prime power order elements: Theorem C. Let G = AB be the mutually permutable product of the subgroups A and B, and let p be a prime. Suppose that for every prime power order p-regular element x ∈ A ∪ B, xG is not divisible by p2 . If G is p-soluble, then G is p-supersoluble. In the line of [5, Theorem 1] and [6, Theorem 2], if we consider all prime numbers, then we obtain some information about the structure of the derived subgroup of a factorised group G. Theorem D. Let G = AB be the product of the subgroups A and B, and assume that G is supersoluble. Suppose that every prime power order element x ∈ A ∪ B has square-free conjugacy class size. Then: (1) G′ is abelian. (2) The Sylow subgroups of G′ are elementary abelian. (3) F(G)′ has Sylow p-subgroups of order at most p2 , for every prime p. If we limit our conditions only to p-regular elements, as a consequence of Theorems B and C, we obtain the following result which extends [2, Corollary 1.5] (see Corollary 4.3) for prime power order elements, and also a theorem of [12], for products of groups. Theorem E. Let G = AB be the mutually permutable product of the subgroups A and B. Suppose that for every prime p and every prime power order p-regular element x ∈ A∪B, xG is not divisible by p2 . Then G is supersoluble, and G/ F(G) has elementary abelian Sylow subgroups. We remark that the first statement in Theorem D is not further true under the weaker hypotheses of the above theorem, even for arbitrary groups not necessarily factorised, as pointed out in [14]. Indeed, as a result of Theorem E, the supersolubility condition in Theorem D can be exchanged by the mutual permutability of the factors. On the other hand, with the stronger assumption that all p-regular elements of the factors (not only those of prime power order) have conjugacy class sizes not divisible by p2 , we get extra information about the orders of the Sylow p-subgroups of G/F (G), extending partially [6, Theorem 2]. Theorem F. Let G = AB be the mutually permutable product of the subgroups A and B. Suppose that for every prime p and every p-regular element x ∈ A ∪ B, xG is not divisible by p2 . Then the order of a Sylow p-subgroup of G/ F(G) is at most p2 . In summary, when dealing with mutually permutable products, the next corollary follows directly from the above theorems. Corollary G. Let G = AB be the mutually permutable product of the subgroups A and B. Suppose that xG is square-free for each element x ∈ A ∪ B. Then G is supersoluble, 3 and both G/ F(G) and G′ have elementary abelian Sylow subgroups. Moreover, G′ is abelian, and both groups G/ F(G) and F(G)′ have Sylow p-subgroups of order at most p2 , for each prime p. In Section 3 we prove Theorems A, B and C, which refer to class sizes not divisible by p2 , for a fixed prime p. Theorems E, D and F, which consider square-free conjugacy class sizes (for all primes), are proved in Section 4. In both cases we will illustrate the scope of the results presented with some examples. 2 Preliminary results We use the following elementary properties frequently, sometimes without further reference. Lemma 2.1. Let N be a normal subgroup of a group G, and let p be a prime. Then: (a) xN divides xG , for any x ∈ N . (b) (xN )G/N divides xG , for any x ∈ G. (c) If xN is a p-element of G/N , then there exists a p-element x1 ∈ G such that xN = x1 N . We need specifically the following fact about Hall subgroups of factorised groups. It is a reformulation of [1, 1.3.2] which is convenient for our purposes. We recall that a group is a Dπ -group, for a set of primes π, if every π-subgroup is contained in a Hall π-subgroup, and any two Hall π-subgroups are conjugate. In particular, all π-separable groups are Dπ -groups for any set of primes π, and all groups are Dπ -groups when π consists of a single prime. Lemma 2.2. Let G = AB be the product of the subgroups A and B. Asume that A, B, and G are Dπ -groups for a set of primes π. Then there exists a Hall π-subgroup H of G such that H = (H ∩ A)(H ∩ B), with H ∩ A a Hall π-subgroup of A and H ∩ B a Hall π-subgroup of B. We collect here some results on mutually permutable products, which will be very useful along the paper. Lemma 2.3. Let the group G = AB be the product of the mutually permutable subgroups A and B. Then we have: (a) ([3, 4.1.10]) G/N is the product of the mutually permutable subgroups AN/N and BN/N . (b) ([3, 4.1.21]) If U is a subgroup of G, then (U ∩ A)(U ∩ B) is a subgroup, and U ∩ A and U ∩ B are mutually permutable. Moreover, if N is a normal subgroup of G, then (N ∩ A)(N ∩ B) is also normal in G. Theorem 2.4. ([4, Theorem 1]) Let the non-trivial group G = AB be the product of the mutually permutable subgroups A and B. Then AG BG is not trivial. 4 The following lemma will be essential in the proofs of our theorems. Lemma 2.5. ([2, Lemma 2.4]) Let p be a prime, and Q be a p′ -group acting faithfully on an elementary abelian p-group N with |[x, N ]| = p, for all 1 6= x ∈ Q. Then Q is cyclic. The next result is the first assertion of Theorem A in [14], which uses the CFSG. Theorem 2.6. Let G be a group. For a fixed prime p with gcd(p − 1, |G|) = 1, if p2 does not divide xG for any p-regular element x ∈ G of prime power order, then G is soluble. Finally, the later lemma, which is a nice result due to Isaacs, will be very useful in the proof of Theorem D. Lemma 2.7. ([10, 4.17]) Let K be an abelian normal subgroup of a finite group G, and let x ∈ G be non-central. Then |CG (x)| < |CG (y)|, where y = [k, x] and k ∈ K is arbitrary. 3 Class sizes not divisible by p2 , for a fixed prime p The well-known Knoche’s theorem (see [11]) asserts that if P is a p-group, p a prime, then the conjugacy class sizes of P are square-free if, and only if, |P ′ | ≤ p. We begin this section by proving Theorem A, which clearly extends it for factorised groups. Proof of Theorem A. Since |P : CP (x)| ≤ p for each x ∈ A ∪ B, it follows Φ(P ) 6 CP (x). Therefore, Φ(P ) commutes with both A and B, so P ′ 6 Φ(P ) 6 Z(P ). Hence P/ Z(P ) is elementary abelian, and xp ∈ Z(P ) for all x ∈ P . Since P ′ 6 Z(P ) and [x, y]p = [xp , y] = 1 (see [7, A - 7.3(a)]) for any x, y ∈ P , it follows that P ′ is elementary abelian. Now it remains to prove that |P ′ | ≤ p2 . Let [x, y] be a generator of P ′ . Since P ′ 6 Z(P ) and y = ya yb with ya ∈ A and yb ∈ B, then [x, y] = [x, ya yb ] = [x, yb ][x, ya ]yb = [x, yb ][x, ya ] ∈ [P, B][P, A] . Thus P ′ = [P, B][P, A]. Clearly, [P, B] is elementary abelian. Suppose [P, B] 6= 1, and let 1 6= [x, z] and 1 6= [x′ , z ′ ] be two generators of [P, B], with x, x′ ∈ P and z, z ′ ∈ B. We distinguish three cases in order to prove that h[x, z]i = h[x′ , z ′ ]i: i) Suppose first z, z ′ ∈ B \ Z(B). Let b ∈ B \ (CP (z) ∪ CP (z ′ )). Since |P : CP (z)| = p then P = CP (z)hxi. Moreover, b ∈ / CP (z) implies that 1 6= [b, z] = [xi t, z] = [xi , z] = i i [x, z] , where b = tx with t ∈ CP (z). On the other hand, P = CP (b)hzi so z ′ = z j k, with k ∈ CP (b). Hence 1 6= [b, z ′ ] = [b, z j k] = [b, z]j = [x, z]i+j . Finally, as b ∈ P = CP (z ′ )hx′ i, then b = (x′ )m s with s ∈ CP (z ′ ). Therefore 1 6= [x, z]i+j = [b, z ′ ] = [(x′ )m s, z ′ ] = [x′ , z ′ ]m , and recall that [x, z] and [x′ , z ′ ] both have order p. Thus, h[x, z]i = h[x′ , z ′ ]i. ii) Now suppose z, z ′ ∈ Z(B). Then B 6 CP (z) ∩ CP (z ′ ). There exists w ∈ P \ (CP (z) ∪ CP (z ′ )). Therefore, w = wa wb with wa ∈ A and wb ∈ B 6 CP (z) ∩ CP (z ′ ), so wa ∈ A \ (CP (z) ∪ CP (z ′ )). Arguing analogously as in case i) with wa instead of b, we conclude that h[x, z]i = h[x′ , z ′ ]i too. 5 iii) Finally, suppose z ∈ B \ Z(B) but z ′ ∈ Z(B). Let z ′′ = zz ′ ∈ B \ Z(B). Therefore, we have [x′ , z ′′ ] = [x′ , z ′ ][x′ , z]. If [x′ , z ′′ ] = 1 then 1 6= [x′ , z ′ ]−1 = [x′ , z], and applying case i) to both [x, z] and [x′ , z] we conclude that they generate the same cyclic group of order p. On the other hand, if [x′ , z ′′ ] 6= 1 and [x′ , z] = 1, then we apply again the first case. Finally, if both [x′ , z ′′ ] 6= 1 6= [x′ , z] then they generate the same cyclic group by case i) again. Thus 1 6= h[x′ , z ′ ]i = h[x′ , z ′′ ][x′ , z]−1 i 6 h[x′ , z ′′ ]i. Since the last one has order p, it follows h[x′ , z ′ ]i = h[x′ , z ′′ ]i. So we have h[x′ , z ′ ]i = h[x′ , z ′′ ]i, which is equal to h[x, z]i by i) again. In conclusion, if [P, B] 6= 1, then it has order p. Analogously with [P, A]. Hence |P ′ | = |[P, B][P, A]| ≤ p2 and this establishes the result. Example 3.1. The converse of the above result is not true in general, in contrast to Knoche’s theorem. Let P be the group of the Small groups library of GAP with identification number 32#35, which is the product of a cyclic group of order 4 and a quaternion group of order 8. Then its derived group is P ′ = C2 × C2 , and P ′ = Φ(P ) = Z(P ). Nevertheless, there are elements in the quaternion group with conjugacy class size equal to 4. Example 3.2. Let G = Q8 × D8 be the direct product of a quaternion group and a dihedral group of order 8. Then every element contained in each factor has conjugacy class size equal to either 1 or 2, so Theorem A applies. However, there are elements in G with conjugacy class size divisible by 4, and Knoche’s result cannot be applied. Now we proceed with a key result in the sequel. Proposition 3.3. Let G be a group, and let p be a prime. Suppose that N is an abelian minimal normal subgroup of G, which is a p′ -group. Then: (1) If G is p-nilpotent, and xG is not divisible by p2 for every element x ∈ N , then |Op (G/N Op (G))| ≤ p. (2) If K/N Op (G) = Op (G/N Op (G)) has order p, and P is a Sylow p-subgroup of K, then CN (P ) = 1. Proof. (1) Suppose that the result is not true, and let G be a counterexample of minimal order. Since the hypotheses are inherited by quotients, we may assume by standard arguments that Op (G) = 1, and then also Φ(G) = 1. Since N is abelian, by a Gaschütz’s result ([9, 4.4]) N is complemented, that is, G = N S with N ∩ S = 1. We may assume that G/N is not a p′ -group, so Op (G/N ) ∼ = Op (S) 6= 1 by the minimality of G. Let P be a Sylow p-subgroup of S (so P is a Sylow p-subgroup of G). Hence it follows Op (S) ∩ Z(P ) 6= 1. Let Z be a minimal normal subgroup of Op (S) ∩ Z(P ). Since S is p-nilpotent, we get S = P L, where L is normal in S and P ∩ L = 1. It follows that [L, Z] 6 [L, Op (S)] 6 L ∩ Op (S) = 1, so Z 6 Z(S). Note that CN (Z) is normal in G = SN . Consequently, by the minimality of N , we have either CN (Z) = 1 or CN (Z) = N . If CN (Z) = N , then Z 6 Z(G), which implies that Z 6 Op (G) = 1, a contradiction. So we may affirm CN (Z) = 1, for every minimal normal subgroup Z of Op (S) ∩ Z(P ). 6 Now let 1 6= x ∈ N such that a Sylow p-subgroup of CG (x), say P0 , is contained in P , so P0 = CP (x). By the hypotheses, xG p = |G : CG (x)|p = |P : P0 | is not divisible by p2 , so it follows either |P : P0 | = 1 or |P : P0 | = p. The first case yields P = CP (x) and then x ∈ CN (Z) = 1, a contradiction. Therefore, we may assume that |P : P0 | = p, and so P0 is normal in P . In addition, since Op (S) ∩ Z(P ) is abelian, by the minimality of Z, we have either P0 ∩ Z = 1 or P0 ∩ Z = Z. The last case gives Z 6 CP (x), a contradiction again. Hence, P0 ∩ Z = 1 and it follows that P = P0 × Z and |Z| = |P : P0 | = p. We only need to see that Z = Op (S) to finish the proof. Note that Z(P ) = Z(P )∩P0 Z = Z(Z(P )∩P0 ), so it follows Z(P )∩Op (S) = Z(Z(P )∩ P0 ) ∩ Op (S) = Z(Z(P ) ∩ P0 ∩ Op (S)). If Z(P ) ∩ P0 ∩ Op (S) 6= 1, since it is normal in Op (S) ∩ Z(P ), we can choose a minimal normal subgroup Z1 of Op (S) ∩ Z(P ) such that Z1 6 Z(P )∩P0 ∩Op (S). But then Z1 6 P0 = CP (x), so x ∈ CN (Z1 ) = 1, a contradiction. Therefore, we may assume that Z(P ) ∩ P0 ∩ Op (S) = 1. On the other hand, we have Op (S) = Op (S)∩ZP0 = Z(P0 ∩Op (S)). If P0 ∩Op (S) is a non-trivial subgroup of P , since it is normal in P , we have a contradiction with Z(P ) ∩ P0 ∩ Op (S) = 1. Consequently we get the final contradiction Z = Op (S). The first assertion is then established. (2) Let K/N Op (G) = Op (G/N Op (G)), which has order p, and let P be a Sylow psubgroup of K. Then K = P N . Moreover, [K, N ] is normal in G and [K, N ] = [P, N ] 6 N so, by the minimality of N , we have either [P, N ] = 1 or [P, N ] = N . The first case leads to K = P × N , and then P 6 Op (G), a contradiction. Thus we have [P, N ] = N , and by coprime action it follows CN (P ) = 1. Note that every dihedral group of order 2q (for q an odd prime) verifies the hypotheses of the above proposition (take p = 2). Theorem B (3) is indeed an immediate consequence of the next more general result. Theorem 3.4. Let G = AB be a soluble group, which is the mutually permutable product of the subgroups A and B. Assume that G is p-nilpotent for a prime p. If p2 does not divide xG for any p-regular element x ∈ A ∪ B of prime power order, then G/ Op (G) has elementary abelian Sylow p-subgroups. Proof. Suppose that the result is false and let G be a minimal counterexample. We may assume by the minimality of G that Op (G) = 1, and therefore Φ(G) = 1 too. By Theorem 2.4, we can assume that there exists a minimal normal subgroup N of G such that N 6 A. Moreover, N is q-elementary abelian, for some prime q 6= p. Furthermore, since N ∩ Φ(G) = 1, by Gaschütz’s lemma we may write G = SN , with S ∩ N = 1. Let P be a Sylow p-subgroup of S (so it is a Sylow p-subgroup of G). Let T = Op (S). By the minimality of G we have T ∼ = Op (G/N ) 6= 1, and by Proposition 3.3 (1) it holds |T | = p. We may choose 1 6= x ∈ N such that P0 = CP (x) is a Sylow p-subgroup of CG (x). Since CN (T ) = 1 by Proposition 3.3 (2), it holds that P0 6= P , |P : P0 | = p, and P0 ∩ T = 1. Hence P = P0 × T . Finally, since P0 ∼ = (P N/N )/ Op (G/N ), which = P/T ∼ is elementary abelian by the minimality of G, it follows that P so is, and this leads to the final contradiction. Proof of Theorem B. Note that the quotients of G satisfy the hypotheses. Moreover, if N is a normal subgroup of G such that N = (N ∩ A)(N ∩ B), then N 7 also inherits the hypotheses. (Observe that this occurs, for instance, if either N 6 A or N 6 B.) (1) We first see that G is soluble by induction over |G|. Since every group of odd order is soluble, we may affirm that p = 2 because gcd(p − 1, |G|) = 1. By Theorem 2.4, we can assume that there exists a normal subgroup M of G such that 1 6= M 6 A. If M < G, then M is soluble by minimality. Analogously G/M is also soluble, and then so is G. If M = G, we apply Theorem 2.6. (2) Suppose that the result is false and let G be a counterexample of minimal order. Since the quotients of G inherit the hypotheses, the class of p-nilpotent groups is a saturated formation, and G is soluble, we may assume that G possesses a unique minimal normal subgroup N , with N = CG (N ) = F(G). If N is a p′ -group, since G/N is pnilpotent by the minimality of G, it follows that G is p-nilpotent, which is a contradiction. Thus, we may assume that N = Op (G). By Theorem 2.4, we can assume without loss of generality that N 6 A, and that there exists a minimal normal subgroup E/N of G/N such that either E/N ≤ A/N or E/N 6 BN/N . In the first case, we have E 6 A. In the second case, it follows E = E ∩ BN = N (E ∩ B) 6 (E ∩ A)(E ∩ B) 6 E. Therefore, we have E = (E ∩ A)(E ∩ B), where the factors are mutually permutable by Lemma 2.3 (b). In both cases, E is normal in G and E verifies the hypotheses. Hence, if E < G, then E is p-nilpotent by the minimality of G. Since N = Op (G), we get that E/N is q-elementary abelian for some prime q 6= p, so it follows that E = QN , with Q the normal Sylow q-subgroup of E. Hence, Q is normal in G which implies that E = N , a contradiction. Therefore, we can assume that E = G. So we have G = E = N Q, where Q is an abelian Sylow q-subgroup of G. By Lemma 2.2, we may assume that Q = (Q ∩ A)(Q ∩ B), with either Q ∩ A 6= 1 or Q ∩ B 6= 1. Suppose first that Q ∩ B 6= 1, and take 1 6= x ∈ Q ∩ B. Let E1 = hxiN , which is normal in QN = G. Hence, we have E1 = hxiN 6 (E1 ∩ B)(E1 ∩ A) 6 E1 . If Q ∩ B = 1, then G = A and we can choose 1 6= x ∈ Q, so that E1 = hxiN is normal in QN = G. Thus, in both cases, we have that E1 inherits the hypotheses and, if E1 < G, it follows that it is p-nilpotent. Therefore hxi is a normal Sylow q-subgroup of E1 , which is again a contradiction. Consequently, we may assume that G = E1 = hxiN , for some q-element x. Note that CN (x) is normal in G = hxiN , since N is abelian. By the minimality of N , it follows that either CN (x) = 1 or CN (x) = N . The second case leads to x ∈ CG (N ) = N , a contradiction. Hence, it follows that CG (x) = CG (x) ∩ N hxi = hxi CN (x) = hxi. Then xG = |G : CG (x)| = |N hxi : hxi| = |N | , and so |N | = p, by the hypotheses. Now, we get that hxi ∼ = G/N = NG (N )/ CG (N ) is isomorphic to a subgroup of Aut(N ) ∼ C , the cyclic group of order p − 1. Hence, |hxi| divides both p − 1 and = p−1 |G|, which contradicts the fact that gcd(p − 1, |G|) = 1. This finishes the proof of the p-nilpotency of G. (3) It follows from Theorem 3.4. In the particular case when G = A = B we recover: Corollary 3.5. ([14, Theorem A]) Let G be a group. For a fixed prime p with gcd(p − 1, |G|) = 1, if p2 does not divide xG for any p-regular element x ∈ G of prime power 8 order, then G is soluble, p-nilpotent and G/ Op (G) has elementary abelian Sylow psubgroups. Note that if G is the direct product of two symmetric groups of degree 3, then G verifies the hypotheses of Theorem B for p = 2, but not those of Corollary 3.5. Moreover, the asumption that gcd(p − 1, |G|) = 1 is necessary, which can be seen by considering G = A5 , the alternating group of degree 5, and the prime p = 3. We include here a theorem due to Cossey and Wang [6], which was the initial motivation for our results, to notify a gap that we have found in one of the statements. Theorem 3.6. ([6, Theorem 1]) Let G be a finite group, and p be a prime divisor of |G| such that if q is any prime divisor of |G|, then q does not divide p − 1. Suppose that no conjugacy class size of G is divisible by p2 . Then G is a soluble p-nilpotent group, and G/ Op (G) has a Sylow p-subgroup of order at most p. Further, if P is a Sylow p-subgroup of G, then P ′ has order at most p, and if P 6= Op (G), then Op (G) is abelian. Remark 3.7. (a) The statement “G/ Op (G) has a Sylow p-subgroup of order at most p” in the above theorem (and so the corresponding one in [13, Theorem 6]) is not true. To see this, consider the semidirect product G = [C5 × C5 ](Sym(3) × C2 ) (where Sym(3) is a symmetric group of degree 3), which is the group of the Small groups library of GAP with identification number 300#25, and the prime p = 2. Then G verifies the hypotheses of Theorem 3.6 but O2 (G) = 1 and |G|2 = 4. We reveal that this example has been communicated to us by John Cossey. (b) The same example shows that the hypotheses in Theorem B for the conjugacy class sizes of the elements x ∈ A ∪ B are not necessarily inherited by the factors, unless they are (sub)normal in G. The above group G can be factorised as the mutually permutable product of A = D10 × D10 and B = [C5 × C5 ]C3 (we checked this using GAP). It is clear that G = AB satisfies the hypotheses of Theorem B for p = 2, but there are elements x ∈ A with xA divisible by 4. Remark 3.8. A natural question is how to extend the last assertion of Theorem 3.6 for (mutually permutable) products. Concerning this, we show the following example: Let A = D8 be a dihedral group of order 8 and B = [C5 ]C4 = ha, b | a5 = b4 = 1, ab = a4 i, and consider the prime p = 2. Then G = A × B is a mutually permutable product of A and B, and G is 2-nilpotent. Moreover, 4 does not divide any conjugacy class size of elements in A ∪ B. However, O2 (G) = (O2 (G) ∩ A)(O2 (G) ∩ B) = D8 × C2 is not abelian. Regarding the claim “P ′ has order at most p” in Theorem 3.6, we get the next extension for factorised groups, as an immediate consequence of Theorem A: Corollary 3.9. Let G = AB be the product of the subgroups A and B. Assume that G is p-nilpotent, and that for all p-elements in the factors, p2 does not divide xG . If P is a Sylow p-subgroup of G, then P ′ 6 Φ(P ) 6 Z(P ), with P ′ elementary abelian of order at most p2 . 9 In particular, from this fact and Theorem B, we get [13, Theorem 7] as a corollary, taking G = A = B. Finally, we prove Theorem C, which is motivated by [2, Theorem 1.3]. Proof of Theorem C. Suppose that the result is false and let G be a counterexample of minimal order. Note that G cannot be simple. Since the class of p-supersoluble groups is a saturated formation, we may assume that there exists a unique minimal normal subgroup N of G, and that Φ(G) = 1. By the minimality of G, we get that G/N is p-supersoluble. Since G is p-soluble, it follows that N is either a p-group or a p′ -group. In the second case, since G/N verifies the thesis by minimality, we get a contradiction. Consequently, we may assume that N is p-elementary abelian and we must show that |N | = p. As Φ(G) = 1 and G is p-soluble with Op′ (G) = 1, by [7, A - 10.6] it follows that F(G) = Soc(G) = N = CG (N ), and also N = Op (G). Applying Theorem 2.4, we may assume that there exists a minimal normal subgroup Z/N of G/N such that Z/N 6 AN/N , so Z = Z ∩ AN = N (Z ∩ A). Since G/N is psoluble, it follows that Z/N is either a p-group or a p′ -group. The first case leads to Z/N 6 Op (G/N ) = Op (G/ Op (G)) = 1, a contradiction. Hence, we may assume that Z/N is a p′ -group. Let Q be a Sylow q-subgroup of Z ∩ A, where q 6= p is a prime (so Q is a Sylow q-subgroup of Z). Therefore Q ∼ = QN/N is a Sylow q-subgroup of Z/N , which acts faithfully on N . If 1 6= a ∈ Q 6 A, then N = [N, a]×CN (a). By the hypotheses, since p2 does not divide aG = |G : CG (a)|, then neither divides |N : CN (a)| = |[N, a]|, so either |[N, a]| = 1 or |[N, a]| = p. The first case leads to a ∈ CG (N ) = N , a contradiction. Thus, Lemma 2.5 yields QN/N is cyclic. Since this is valid for all primes q 6= p, we get by [10, 5.15] that Z/N is soluble. By the minimality of Z/N , it follows that (Z/N )′ = 1 and Z/N is abelian with cyclic Sylow subgroups. Consequently Z/N = hxN i, where x∈ / N and the order of xN is q, for some prime q 6= p. We may assume that x ∈ Z ∩ A and Z = N hxi. Hence CN (x) = CN (Z). By the minimality of N , we have either CN (Z) = N or CN (Z) = 1. The first case leads to x ∈ Z 6 CG (N ) = N , a contradiction. Therefore, since N = [N, x] × CN (x), it follows |N | = |[N, x]| = p, and this final contradiction establishes the theorem. Example 3.10. Let G be the symmetric group of degree 4. Then G = AB is a mutually permutable product, where A denotes the alternating group of degree 4 and B is a Sylow 2-subgroup of G, which verifies the hypotheses of Theorem C, for p = 3. 4 Square-free class sizes We begin this section with the proof of Theorem D. Proof of Theorem D. (1) Suppose that the result is false and let G be a counterexample of least possible order. Since G supersoluble, G/ F(G) is abelian, and so G′ 6 F(G). Moreover, the quotients of G inherit the hypotheses and the class of metabelian groups is a formation, so we may assume that there exists a unique minimal normal subgroup N of G with |N | = p, for some prime divisor p of |G|. 10 Hence, F(G) = Op (G) 6 P , a Sylow p-subgroup of G. Since G/ F(G) is abelian, P = Op (G) = F(G). Hence P = (P ∩ A)(P ∩ B), where P ∩ A and P ∩ B are Sylow p-subgroups of A and B respectively, by Lemma 2.2. Applying Theorem A, we have P ′ 6 Φ(P ) 6 Z(P ), and P ′ is elementary abelian of order at most p2 . Note that P ′ 6= 1, because G′ ≤ P . By Lemma 2.2, we may consider H a Hall p′ -subgroup of G, such that H = (H ∩ A)(H ∩ B), where H ∩ A and H ∩ B are Hall p′ -subgroups of A and B respectively. Moreover, H ∼ = G/ F(G) is abelian. Let x ∈ H ∩ A be a prime power order element. Since H 6 CG (x) 6 G, it follows by the hypotheses that xG ≤ p, and so Φ(G) 6 CG (x). Thus Φ(G) 6 CG (H ∩ A), and analogously for H ∩ B. Consequently we get P ′ 6 Φ(P ) 6 Φ(G) 6 CG (H). Since P ′ 6 Φ(P ) 6 Z(P ), it follows P ′ 6 Φ(P ) 6 Z(G). In particular, Φ(G) 6= 1 6= Z(G). If A, B 6 P , then G = P and G′ = P ′ 6 Z(G), a contradiction. Hence we have either H ∩ A 6= 1 or H ∩ B 6= 1. Assume H ∩ A 6= 1. Let QA be a Sylow q-subgroup of H ∩ A, for some prime q 6= p. Note that QA 66 Z(G), because Z(G) is a p-group. Let QA = QA Z(G)/ Z(G), which acts on P = P/ Z(G), which is elementary abelian because Φ(P ) 6 Z(G) 6 F(G) = P . Suppose w = w Z(G) ∈ CQA (P ). Then [w, y] = 1 for all y = y Z(G) ∈ P , so [w, y] ∈ Z(G) 6 P . Let k = o(w) denote the order of w. Thus [w, y]k = [wk , y] = 1. It follows [w, y] = 1 for all y ∈ P , so w ∈ CQA (P ) = CQA (F(G)) = 1. Then CQA (P ) = 1 and the action is faithful. Let 1 6= αq = αq Z(G) ∈ QA . Therefore P = [P , αq ] × CP (αq ), with [P , αq ] 6= 1. Moreover, [P , αq ] = P : CP (αq ) = P : CP (αq ) = |P : CP (αq )| , which divides αG q , because P is normal in G. Since H 6 CG (αq ) and αq is a non-central prime power order element in A, it follows αG q = p, and so [P , αq ] = p. ∼ Applying Lemma 2.5, we get that QA = QA is cyclic. Since this is valid for each prime divisor q of |H ∩ A|, we deduce that H ∩ A has cyclic Sylow subgroups, but it is abelian, so H ∩ A is cyclic. Analogously, if H ∩ B 6= 1, then it is cyclic. Let H ∩ A = hαi. Assume first that 1 6= α is a q-element, for some prime q, and that G = P hαi. By the above argument, hαi Z(G)/ Z(G) acts faithfully on P , and [P , α] = |P : CP (α)| = p. Let y ∈ P \ CP (α). If [y, α] = 1, then [y, α] ∈ Z(G) 6 P . Hence [y, α]o(α) = [y, αo(α) ] = 1, so [y, α] = 1 and y ∈ CP (α), a contradiction. Then [y, α] 6= 1, and since [P , α] = p, it follows [P , α] = h[y, α]i. Therefore we have [P, α] = [P , α] = h[y, α]i = h[y, α]i, so [P, α] 6 [P, α] Z(G) = h[y, α]i Z(G), and then G′ = P ′ [P, α] 6 h[y, α]i Z(G) which is abelian, a contradiction. Hence, we may assume that, for every prime q, if αq is the q-part of α, then P hαq i < G. Note that P hαq i is normal in G, and P hαq i = (P hαq i ∩ A)(P hαq i ∩ B). Therefore, by the minimality of G, it follows that (P hαq i)′ is abelian. Notice that (P hαq i)′ = P ′ [P, αq ], since P is normal in G. Let K = P ′ [P, αq ], which is an abelian normal subgroup of G, and let t ∈ [P, αq ] 6 K. Then Lemma 2.7 leads to |CG (αq )| < |CG ([t, αq ])|. If p divides [t, αq ]G we get a contradiction, because [t, αq ]G < αG q = p. Hence, P 6 CG ([t, αq ]) and [t, αq ] ∈ Z(P ), for each t ∈ [P, αq ]. By coprime action, P = [P, αq ] CP (αq ). Thus [P, αq ] = [[P, αq ] CP (αq ), αq ] = [P, αq , αq ]. If k is a generator of [P, αq , αq ], then k = [t, αq ] ∈ Z(P ) with t ∈ [P, αq ], so [P, αq ] = [P, αq , αq ] 6 Z(P ). Since this is valid for each 11 prime divisor q of the order of H ∩ A = hαi, we get: [P, H ∩ A] = [P, hαq1 i × · · · × hαqt i] = [P, αq1 ] · · · [P, αqt ] 6 Z(P ). Analogously, if H ∩ B 6= 1, then [P, H ∩ B] ≤ Z(P ). Since G′ = P ′ [P, H] = P ′ [P, H ∩ A][P, H ∩ B], we get G′ ≤ Z(P ). This final contradiction establishes statement (1). (2) Suppose that the second assertion is not true and let G be a counterexample of minimal order. We point out that the hypotheses are inherited by every quotient group of G and, by (1), G′ is abelian. There exists a prime divisor p of |G′ | such that G′ does not have any elementary abelian Sylow p-subgroup. By the minimality of G, we may consider that Op′ (G) = 1. Moreover, since G is supersoluble, then G/ F(G) is abelian, and F(G) = Op (G) = P is a normal Sylow p-subgroup of G such that G′ 6 P . Using Lemma 2.2 and Theorem A, we obtain respectively that P = (P ∩ A)(P ∩ B), and that P ′ is elementary abelian with P ′ 6 Φ(P ) 6 Z(P ). Let GN be the nilpotent residual of G. Note that GN 6= 1; in other case, G is a p-group and then G′ = P ′ , a contradiction. Since GN 6 G′ , it follows that GN is abelian. By using [7, III - 4.6, IV - 5.18], we have that GN is complemented in G, and its complements are precisely the Carter subgroups of G. Accordingly, G = GN H with H = NG (H) a nilpotent subgroup of G and GN ∩ H = 1. These facts yield G′ = GN × (H ∩ G′ ), and CGN (H) = 1. On the other hand, the minimality of G implies that G′ /GN ∼ = (H ∩ G′ ) is N N N elementary abelian, and thus G is not so. If CH (G ) 6= 1, since CH (G ) is normal in G, by the minimality of G we have (G/ CH (GN ))′ is an elementary abelian group, but (G/ CH (GN ))′ = G′ CH (GN )/ CH (GN ) = GN CH (GN )/ CH (GN ) ∼ = GN , which is a contradiction. Hence, CH (GN ) = 1. In particular, we deduce that Z(G) = 1. By Lemma 2.2, there exists a Hall p′ -subgroup H0 of G such that H0 = (H0 ∩ A)(H0 ∩ B). Let x ∈ H0 ∩ A be a non-trivial element of prime power order. Since H0 6 CG (x), it follows that xG ≤ p, and so Φ(G) 6 CG (x). Thus Φ(G) 6 CG (H0 ∩ A), and analogously for H0 ∩ B. Consequently, we get P ′ 6 Φ(P ) 6 Φ(G) 6 CG (H0 ). Since P ′ 6 Φ(P ) 6 Z(P ) and G = P H0 , it follows P ′ 6 Φ(P ) 6 Z(G) = 1, which implies that P is elementary abelian, the final contradiction. (3) Assume that the result is false and take G a counterexample of minimal order. Consider a prime p such that | F(G)′ |p > p2 . By minimality, we can affirm that Op′ (F(G)) = 1. Since G is supersoluble, we get that F(G) = Op (G) = P is a normal Sylow p-subgroup of G. Then, we apply both Lemma 2.2 and Theorem A to get the final contradiction. Example 4.1. Let G = A × B be the direct product of two symmetric groups of degree 3. Then G is supersoluble, and every element contained in each factor (not only those of prime power order) has square-free conjugacy class size, but neither the derived subgroup G′ nor G/ F(G) are cyclic, in contrast to [6, Theorem 2]. Example 4.2. In view of [6, Theorem 2], it is natural to wonder if we can affirm in the above result that the Sylow p-subgroups of G′ have order at most p2 . This fact is not further true, as we show: 12 Let G = A×B, where A = D14 is a dihedral group of order 14, and B = D14 ×[C7 ]C3 is the direct product of such a dihedral group and a semidirect product of a cyclic group of order 7 and a cyclic group of order 3 (B has identification number 294#9 in the Small groups library of GAP). Then G is supersoluble, and satisfies that all prime power order elements contained in each factor have square-free conjugacy class size, but G′ has order 73 . Now we proceed with the proof of Theorem E. Proof of Theorem E. Considering the smallest prime divisor of |G| and Theorem B, we conclude that G is soluble. Hence, it is p-soluble for each prime divisor p of |G|. Applying Theorem C, we get that G is p-supersoluble for each prime that divides |G|, so it is supersoluble. Now we prove the second assertion by induction on |G|. Let p be an arbitrary prime, and P be a Sylow p-subgroup of G. We want to show that P F(G)/ F(G) ∼ = P/ Op (G) is elementary abelian. Since G is supersoluble, we have that G/ F(G) is abelian. Moreover, we may assume by induction that Op (G) = 1. Therefore, we have that F(G) 6 H 6 G, where H is a Hall p′ -subgroup of G. Consequently, H is normal in G and G is p-nilpotent. Finally, by Theorem 3.4 the result is established. When considering in the above theorem all p-regular elements in the factors, we get as a corollary: Corollary 4.3. ([2, Corollary 1.5]) Let the group G = AB be the mutually permutable product of the subgroups A and B. Suppose that for every prime p and every p-regular element x ∈ A ∪ B, xG is not divisible by p2 . Then G is supersoluble. Example 4.4. Consider G = A × B, where A = Sym(3) is a symmetric group of degree 3, and B = Sym(3) × D10 is the direct product of such a symmetric group and a dihedral group of order 10. Then G satisfies the hypotheses of Theorem E. However there exists some 2-regular element in B, not of prime power order, such that 4 divides its conjugacy class size, so Corollary 4.3 cannot be applied. In the particular case when A and B are normal in G, we obtain [13, Proposition 9]. Corollary 4.5. Let A and B be normal subgroups of G such that G = AB. Suppose that xG is square-free for every element x of prime power order of A ∪ B. Then G is supersoluble. This development has its origins in the contributions of Chillag and Herzog [5, Theorem 1], and Cossey and Wang [6, Theorem 2]. Our next result Theorem F and Theorem D can be considered somehow extensions of the ones above for (mutually permutable) products. In fact, Theorem F provides further information on the Sylow subgroups of G/ F(G). Proof of Theorem F. Suppose the result is not true and let G be a counterexample of least order possible. Then if P is a Sylow p-subgroup of G, we have |P/ Op (G)| ≥ 13 p3 . We can assume by the minimality of G that Op (G) = 1 = Φ(G) = Z(G), so |P | ≥ p3 . By Lemma 2.2 we can choose P = (P ∩ A)(P ∩ B), with P ∩ A and P ∩ B Sylow p-subgroups of A and B respectively. By Theorem E, we have that G is supersoluble and G/ F(G) has elementary abelian Sylow subgroups. In particular, P F(G) is normal in G. Hence Lemma 2.3 (b) asserts that L = (P F(G) ∩ A)(P F(G) ∩ B) is normal in G, and it is a mutually permutable product. Moreover, P = (P ∩ A)(P ∩ B) 6 (P F(G)∩A)(P F(G)∩B) = L 6 P F(G). If we suppose L < G, by the minimality of G it follows |P/ Op (L)| = |P | ≤ p2 , a contradiction. Thus, we may assume L = G = P F(G), and so G is p-nilpotent. Let N be a minimal normal subgroup of G. We can assume without loss of generality that it is contained in A by Theorem 2.4. Note |N | = q 6= p. By Proposition 3.3 (1), it follows |Op (G/N )| ≤ p, and by the minimality of G we have |(P N/N )/ Op (G/N )| ≤ p2 . Since P ∼ = P N/N , we may assume |P | = p3 . As Φ(G) = 1, [[7], A - 10.6 Theorem] leads to F(G) = Soc(G). If N is the unique minimal normal subgroup of G, then P ∼ = NP (N )/ CP (N ) which is isomorphic to a subgroup of Aut(Cq ) ∼ = Cq−1 , so P is cyclic and elementary abelian, which implies that its order is p, a contradiction. Now we denote by T the product of all minimal normal subgroups of G distinct of N , so T 6= 1 and T ∩ N = 1. It follows F(G) = Soc(G) = N × T . We denote Q1 = Op (P N ) and Q2 = Op (P T ). Since P N ∼ = G/T (and P T ∼ = G/N ), by the minimality of G we may affirm Q1 6= 1 6= Q2 . On the other hand, since [Op (P N ), N ] 6 Op (P N ) ∩ N = 1, we have Q1 6 C1 = CP (N ) (analogously Q2 6 C2 = CP (T )). In addition, it follows C1 ∩ C2 6 CP (F(G)) 6 CG (F(G)) 6 F(G), so C1 ∩ C2 = 1. Let P0 be a Sylow psubgroup of CG (N ) such that P0 6 P . Hence P0 = C1 = CP (N ). In addition, since N = hxi where x is a q-element contained in A, by the hypotheses it follows that p2 does not divide xG p = |G : CG (x)|p = |P : C1 |. Moreover, since |P | = |P : C1 | · |C1 | = p3 , we may assume |C1 | ≥ p2 , and since 1 6= Q2 6 C2 , we have |C2 | ≥ p. Accordingly |C1 C2 | = |C1 | · |C2 | ≥ p3 , and since P is abelian, we have necessarily P = C1 × C2 . This leads to G = P F(G) = C1 C2 N T = (C1 T ) × (C2 N ). Suppose T ∩ A 6= 1, and let 1 6= y ∈ T ∩ A 6 C1 T ∩ A. Let 1 6= x ∈ N 6 A. Then since xy ∈ F(G) ∩ A and F(G) is abelian, we have that xy is a p-regular element, so by the hypotheses p2 does not divide (xy)G . As G is a direct product, we have (xy)G = xG y G . In addition, (N × C1 T ) 6 CG (x) 6 G, and therefore xG divides |G : (C1 T × N )| = |C2 | which is a p-number, so xG = p (recall that Z(G) = 1), and analogously y G = p, a contradiction. We conclude T ∩ A = 1. If F(G) = K × M with M normal in G and K a minimal normal subgroup of G contained in B, by similar arguments we can deduce M ∩ B = 1. This means, in particular, that neither A nor B can contain two distinct minimal normal subgroups of G. On the other hand, since F(G) is the unique p′ -Hall subgroup of G, Lemma 2.2 leads to F(G) = (F(G) ∩ A)(F(G) ∩ B). Moreover, since F(G) ∩ A = N T ∩ A = N (T ∩ A) = N , it follows F(G) = N (F(G) ∩ B). Note that (F(G) ∩ B) ∩ N 6 N with |N | = q so we distinguish two cases: either (F(G)∩ B)∩ N = N or (F(G)∩ B)∩ N = 1. In the first case 14 F(G) = F(G) ∩ B 6 B. Thus there exists another minimal normal subgroup contained in B and distinct of N , a contradiction. Hence we conclude (F(G) ∩ B) ∩ N = 1 so F(G) = N × (F(G) ∩ B). Now suppose that F(G) is a q-group. Then, since F(G) = Soc(G), it follows that F(G) is q-elementary abelian. In addition, P ∩ A acts faithfully over F(G). Let 1 6= x ∈ P ∩ A 6 A, then F(G) = [F(G), x] × CF(G) (x), with CF(G) (x) < F(G) since Z(G) = 1 and P is abelian. By the hypotheses, q 2 does not divide xG , and therefore it does not divide F(G) : CF(G) (x) . Thus we may affirm |[F(G), x]| = q. By Lemma 2.5 we conclude that P ∩ A is cyclic, and analogously P ∩ B is cyclic too. So they are both cyclic and elementary abelian, that is, they both have order p. Thus |P | = |(P ∩ A)(P ∩ B)| ≤ p2 , a contradiction. Hence we may suppose that there exists a prime r 6= q such that r divides |F(G)|. Let 1 6= R be a Sylow r-subgroup of F(G) ∩ B (so it is a Sylow r-subgroup of F(G)). Then 1 6= R = Or (G) 6 B, and since F(G) = Soc(G), necessarily we have that Or (G) is the product of the minimal normal subgroups of G with order r. Let M 6 B be one of those minimal normal subgroups. Arguing exactly in the same way as with N , it follows F(G) = M ×(A∩F(G)). But A∩F(G) = N so F(G) = N ×M with both minimal normal subgroups of G, N 6 A and M 6 B. Let P1 be a Sylow p-subgroup of CG (N ) = CG (x) such that P1 6 P . Then by the hypotheses we have xG p = |P : P1 | ≤ p. Since |P | = p3 , it follows |P1 | ≥ p2 . However, P1 is normal in P N so P1 6 Op (P N ) ∼ = Op (G/M ), and by Proposition 3.3 (1) we have |Op (G/M )| ≤ p. This final contradiction establishes the theorem. Example 4.6. Under the hypotheses of Theorem E (even under those of Theorem D), it is not possible to assure that G/ F(G) has Sylow p-subgroups of order at most p2 , as the following example shows: Let {p1 , p2 , . . . , pn } be a finite set of pairwise distinct odd primes, and let G = D2p1 × D2p2 × · · · × D2pn be the direct product of dihedral groups of order 2pi , 1 ≤ i ≤ n. Then G = A×B is a mutually permutable product of A = D2p1 and B = D2p2 ×· · ·×D2pn , and each prime power order element contained in the direct factors has square-free conjugacy class size. However, G/ F(G) has order 2n . References [1] B. Amberg, S. Franciosi and F. de Giovanni: University Press, New York, 1992. Products of groups. Oxford [2] A. Ballester-Bolinches, J. Cossey and Y. Li: Mutually permutable products and conjugacy classes. Monatsh. Math. 170 (2012) 305–310. [3] A. Ballester-Bolinches, R. Esteban-Romero and M. Asaad: Products of finite groups. Vol. 53 of de Gruyter Expositions in Mathematics, Berlin, 2010. [4] J. C. Beidleman and H. Heineken: Group classes and mutually permutable products. J. Algebra 297 (2006) 409–416. 15 [5] D. Chillag and M. Herzog: On the length of the conjugacy classes of finite groups. J. Algebra 131 (1990) 110–125. [6] J. Cossey and Y. Wang: Remarks on the length of conjugacy classes of finite groups. Commun. Algebra 27 (1999) 4347–4353. [7] K. Doerk and T. Hawkes: Finite Soluble Groups. Vol. 4 of de Gruyter Expositions in Mathematics, Berlin, 1992. [8] The GAP Group: GAP – Groups, Algorithms, and Programming. http://www.gap-system.org, Version 4.7.7, 2015. [9] B. Huppert: Endliche Gruppen I. Springer-Verlag, Berlin, 1967. [10] I. M. Isaacs: Finite Group Theory. American Mathematical Society, USA, 2008. [11] H. G. Knoche: Über den Frobeniusschen Klassenbegriff in nilpotenten Gruppen. Math. Z. 55 (1951) 71–83. [12] S. Li: On the class length of elements of prime power order in finite groups. Guangxi Sci. 6(1) (1999) 12–13. [13] X. Liu, Y. Wang and H. Wei: Notes on the length of conjugacy classes of finite groups. J. Pure Appl. Algebra 196 (2005) 111–117. [14] G. Qian and Y. Wang: On conjugacy class sizes and character degrees of finite groups. J. Algebra Appl. 13 (2014) 1350100-1–9. 16 

4math.GR

A deep learning model integrating FCNNs and CRFs for brain tumor segmentation Xiaomei Zhao1,2, Yihong Wu1, Guidong Song3, Zhenye Li4, Yazhuo Zhang,3,4,5,6 , and Yong Fan7 National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences 1 2 3 4 Beijing Neurosurgical Institute, Capital Medical University, Beijing, China Department of Neurosurgery, Beijing Tiantan Hospital, Capital Medical University, Beijing, China 5 6 7 University of Chinese Academy of Sciences, Beijing, China Beijing Institute for Brain Disorders Brain Tumor Center, Beijing, China China National Clinical Research Center for Neurological Diseases, Beijing, China Department of Radiology, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA *Correspondence to: Dr. Yihong Wu National Laboratory of Pattern Recognition Institute of Automation Chinese Academy of Sciences Beijing, 100190, China Email: yhwu@nlpr.ia.ac.cn Or Dr. Yong Fan Department of Radiology Perelman School of Medicine University of Pennsylvania Philadelphia, PA 19104, USA Email: yong.fan@uphs.upenn.edu; yong.fan@ieee.org Tel: +1 215-746-4065 1 1 This article was accepted in the journal Medical Image Analysis. Please cite this article as: X. Zhao et al., A deep learning model integrating FCNNs and CRFs for brain tumor segmentation, Medical Image Analysis, Volume 43, January 2018, Pages 98-111. http://doi.org/10.1016/j.media.2017.10.002 1 ABSTRACT Accurate and reliable brain tumor segmentation is a critical component in cancer diagnosis, treatment planning, and treatment outcome evaluation. Build upon successful deep learning techniques, a novel brain tumor segmentation method is developed by integrating fully convolutional neural networks (FCNNs) and Conditional Random Fields (CRFs) in a unified framework to obtain segmentation results with appearance and spatial consistency. We train a deep learning based segmentation model using 2D image patches and image slices in following steps: 1) training FCNNs using image patches; 2) training CRFs as Recurrent Neural Networks (CRF-RNN) using image slices with parameters of FCNNs fixed; and 3) fine-tuning the FCNNs and the CRF-RNN using image slices. Particularly, we train 3 segmentation models using 2D image patches and slices obtained in axial, coronal and sagittal views respectively, and combine them to segment brain tumors using a voting based fusion strategy. Our method could segment brain images slice-by-slice, much faster than those based on image patches. We have evaluated our method based on imaging data provided by the Multimodal Brain Tumor Image Segmentation Challenge (BRATS) 2013, BRATS 2015 and BRATS 2016. The experimental results have demonstrated that our method could build a segmentation model with Flair, T1c, and T2 scans and achieve competitive performance as those built with Flair, T1, T1c, and T2 scans. Index Terms— Brain Tumor Segmentation, Fully Convolutional Neural Networks, Conditional Random Fields, Deep learning 1. Introduction Accurate brain tumor segmentation is of great importance in cancer diagnosis, treatment planning, and treatment outcome evaluation. Since manual segmentation of brain tumors is laborious [1], an enormous effort has devoted to the development of semi-automatic or automatic brain tumor segmentation methods. Most of the existing brain tumor segmentation studies are focusing on gliomas that are the most common brain tumors in adults and can be measured by Magnetic Resonance Imaging (MRI) with multiple sequences, such as T2-weighted fluid attenuated inversion recovery (Flair), T1-weighted (T1), T1-weighted contrast-enhanced (T1c), and T2-weighted (T2). The segmentation of gliomas based on MRI data is challenging for following reasons: 1) gliomas may have the same appearance as gliosis and stroke in MRI data [2]; 2) gliomas may appear in any position of the brain with varied shape, appearance and size; 3) gliomas invade the surrounding brain tissues rather than displacing them, causing fuzzy boundaries [2]; and 4) intensity inhomogeneity of MRI data further increases the difficulty. The existing automatic and semi-automatic brain tumor segmentation methods can be broadly categorized as either generative model based or discriminative model based methods [3]. The generative model based brain tumor segmentation methods typically require prior information, which could be gained through probabilistic image atlases [4-6]. Based on probabilistic image atlases, the brain tumor segmentation problem can be modeled 2 as an outlier detection problem [7]. On the other hand, the discriminative model based methods solve the tumor segmentation problem in a pattern classification setting, i.e., classifying image voxels as tumor or normal tissues based on image features. The performance of discriminative model based segmentation methods are hinged on the image features and classification algorithms. A variety of image features have been adopted in tumor segmentation studies, including local histograms [8], image textures [9], structure tensor eigenvalues [10], and so on. The most commonly adopted pattern classification algorithms in brain tumor segmentation studies are support vector machines (SVMs) [11-13] and random forests [8-10, 14]. More recently, deep learning techniques have been adopted in brain tumor segmentation studies following their success in general image analysis fields, such as images classification [15], objects detection[16], and semantic segmentation [17-19]. Particularly, Convolutional Neural Networks (CNNs) were adopted for brain tumor image segmentation in the Multimodal Brain Tumor Image Segmentation Challenge (BRATS) 2014 [20-22]. More deep learning based brain tumor segmentation methods were presented in the BRATS 2015 and different deep learning models were adopted, including CNNs [23-25], convolutional restricted Boltzman machines [26], and Stacked Denoising Autoencoders [27]. Among the deep learning based tumor segmentation methods, the methods built upon CNNs have achieved better performance. Particularly, both 3D-CNNs [22, 28, 29] and 2D-CNNs [20, 21, 23-25, 30, 31] models were adopted to build tumor segmentation methods. Although 3D-CNNs can potentially take full advantage of 3D information of the MRI data, the network size and computational cost are increased too. Therefore, 2D-CNNs have been widely adopted in the brain tumor segmentation methods. Davy et al. proposed a deep learning method with two pathways of CNNs, including a convolutional pathway and a fully-connected pathway [21]. Dvorak et al. modeled the multi-class brain tumor segmentation task as 3 binary segmentation sub-tasks and each sub-task was solved using CNNs [23]. Very deep CNNs [32] were adopted to segment tumors by Pereira et al. [25]. Most of these brain tumor segmentation methods train CNNs using image patches, i.e., local regions in MR images. These methods classify each image patch into different classes, such as healthy tissue, necrosis, edema, non-enhancing core, and enhancing core. The classification result of each image patch is used to label its center voxel for achieving the tumor segmentation. Most of the above CNN brain tumor segmentation methods assumed that each voxel’s label is independent, and they didn’t take the appearance and spatial consistency into consideration. To take the local dependencies of labels into account, Havaei et al. constructed a cascaded architecture by taking the pixel-wise probability segmentation results obtained by CNNs trained at early stages as additional input to their following CNNs [24, 30]. To take into consideration appearance and spatial consistency of the segmentation results, Markov Random Fields (MRFs), particularly Conditional Random Fields (CRFs), have been integrated with deep learning techniques in image segmentation studies, either used as a post-process step of CNNs [28, 33] or formulated as neural networks [18, 19]. In the latter setting, both CNNs and MRFs/CRFs can be trained with 3 back-propagation algorithms, tending to achieve better segmentation performance. Multiple 2D CNNs could be integrated for segmenting 3D medical images. In particular, Prasoon et al. proposed a triplanar CNN [34] for knee cartilage segmentation. The triplanar network used 3 CNNs to deal with patches extracted from xy, yz and zx planes and fused them using a softmax classifier layer. Fritscher et al. proposed a pseudo 3D patch-based approach [35], consisting of 3 convolutional pathways for image patches in axial, coronal, and sagittal views respectively and fully connected layers for merging them. Setio et al. used multi-view convolutional networks for pulmonary nodule detection [36]. Their proposed network architecture composed multiple streams of 2D CNNs, each of which was used to deal with patches extracted in a specific angle of the nodule candidates. The outputs of the multiple streams of 2D CNNs were finally combined to detect pulmonary nodules. However, all these methods built CNNs upon image patches, not readily extendable for building FCNNs. Preprocessing of MRI data plays an important role in the discriminative model based tumor segmentation methods that assume different MRI scans of the same modality have comparable image intensity information. The intensities of different MRI scans can be normalized by subtracting their specific mean values and dividing by their specific standard deviation values or by matching histograms [10, 22]. However, the mean values of intensities of different MRI scans do not necessarily correspond to the same brain tissue, and the histogram matching might not work well for tumor segmentation studies [8]. A robust intensity normalization has been adopted in tumor segmentation studies by subtracting the gray-value of the highest histogram bin and normalizing the standard deviation to be 1 [8]. Inspired by the success of deep learning techniques in medical image segmentation, we propose a new brain tumor segmentation method by integrating Fully Convolutional Neural Networks (FCNNs) and CRFs in a unified framework. Particularly, we formulate the CRFs as Recurrent Neural Networks [18], referred to as CRF-RNN. The integrative model of FCNNs and CRF-RNN is trained in 3 steps: 1) training FCNNs using image patches; 2) training CRF-RNN using image slices with parameters of FCNNs fixed; and 3) fine-tuning the whole network using image slices. To make use of 3D information provided by 3D medical images, we train 3 segmentation models using 2D image patches and slices obtained in axial, coronal and sagittal views respectively, and combine them to segment brain tumors using a voting based fusion strategy. The proposed method is able to segment brain images slice-by-slice, which is much faster than the image patch based segmentation methods. Our method could achieve competitive segmentation performance based on 3 MR imaging modalities (Flair, T1c, T2), rather than 4 modalities (Flair, T1, T1c, T2) [3, 8-10, 14, 20-31], which could help reduce the cost of data acquisition and storage. We have evaluated our method based on imaging data provided by the Multimodal Brain Tumor Image Segmentation Challenge (BRATS) 2013, the BRATS 2015, and the BRATS 2016. The experimental results have demonstrated that our method could achieve promising brain tumor segmentation performance. Preliminary results have been reported in a conference proceeding paper of the BRATS 2016 [37]. 4 2. Methods and materials 2.1 Imaging data All the imaging data used in this study were obtained from the BRATS 20132, the BRATS 20153 and the BRATS 20164. The BRATS 2013 provided clinical imaging data of 65 glioma patients, including 14 patients with low-grade gliomas (LGG) and 51 patients with high-grade gliomas (HGG). The patients were scanned with MRI scanners from different vendors at 4 different centers, including Bern University, Debrecen University, Heidelberg University, and Massachusetts General Hospital. Each patient had multi-parametric MRI scans, including T2-weighted fluid attenuated inversion recovery (Flair), T1-weighted (T1), T1-weighted contrast-enhanced (T1c), and T2-weighted (T2). All the MRI scans of the same patient were rigidly co-registered to their T1c scan and resampled at 1 mm isotropic resolution in a standardized axial orientation with a linear interpolator [3]. All images were skull stripped. Ground truths were produced by manual annotations. The cases were split into training and testing sets. The training set consists of 20 HGG and 10 LGG cases. The testing set consists of Challenge and Leaderboard subsets. The Challenge dataset has 10 HGG cases and the Leaderboard dataset contains 21 HGG and 4 LGG. The imaging dataset provided by BRATS 2015 contains imaging data obtained from the BRATS 2012, 2013, and the NIH Cancer Imaging Archive (TCIA). Each case has Flair, T1, T1c, and T2 scans aligned onto the same anatomical template space and interpolated at 1 mm3 voxel resolution. The testing dataset consists of 110 cases with unknown grades, and the training dataset consists of 220 HGG and 54 LGG cases. In the testing dataset, the ground truth of each case was produced by manual annotation. In the training dataset, all the cases from the BRATS 2012 and 2013 were labeled manually, and the cases from the TCIA were annotated by fusing segmentation results obtained using top-ranked methods of the BRATS 2012 and 2013. The annotations were inspected visually and approved by experienced raters. The tumor labels of the training cases are provided along with their imaging scans, while only imaging data are provided for the testing cases for blind evaluation of the segmentation results. BRATS 2016 shares the same training dataset with BRATS 2015, which consists of 220 HGG and 54 LGG. Its testing dataset consists of 191 cases with unknown grades. The ground truth of each testing case was produced by manual annotation, not released to the competition participants. 2.2 Brain tumor segmentation methods based on FCNNs trained using image patches Deep learning techniques, particularly CNNs, have been successfully adopted in image segmentation studies. A deep learning model of CNNs usually has millions or even billions of parameters. To train the deep CNNs with sufficient training samples, image patch-based techniques are adopted [20-25, 28, 30, 31, 38-40]. With the image 2 3 4 https://www.virtualskeleton.ch/BRATS/Start2013 https://www.virtualskeleton.ch/BRATS/Start2015 https://www.virtualskeleton.ch/BRATS/Start2016 5 patch based representation, the image segmentation problem can be solved as a classification problem of image patches. An image patch is a local region extracted from an image to characterize its central pixel/voxel in 2D/3D, and has the same label as its center pixel/voxel’s label in the classification problem. In the training phase, a large number of image patches can be extracted to train the CNNs. In the testing phase, image patches extracted from a testing image are classified one by one by the trained CNNs. Then, the classification results of all image patches make up a segmentation result of the testing image. However, FCNNs can segment a testing image slice by slice with improved computational efficiency [30], even though the model is trained using image patches. Since the number and location of training image patches for each class can be easily controlled by changing the image patch sampling scheme, image patch-based deep learning segmentation methods can avoid the training sample imbalance problem. However, a limitation of image patch-based segmentation methods is that relationship among image patches is typically lost. Integrating CRF-RNN with FCNNs tends to overcome such a limitation in tumor segmentation. 2.3 The proposed brain tumor segmentation method The proposed brain tumor segmentation method consists of 4 main steps: pre-processing, segmenting image slices using deep learning models with integrated FCNNs and CRF-RNN from axial, coronal and sagittal views respectively, fusing segmentation results obtained in the three different views, and post-processing. 2.3.1 Pre-processing of the imaging data Since MRI scans typically have varied intensity ranges and are affected by bias fields differently, we adopted a robust intensity normalization method to make MRI scans of different patients comparable, besides correcting the bias field of MRI data using N4ITK [41]. Our normalization method is built upon the image mode based method [8], which normalizes image intensity by subtracting the image mode (e.g. the gray-value of the highest histogram bin) and normalizing the standard deviation to be 1. As almost half of the brain is the whiter matter [42], the gray-value of the highest histogram bin typically corresponds to the gray-value of the white matter, and therefore matching intensity values of the white matter across MRI scans and normalizing the intensity distributions accordingly would largely make different MRI scans comparable. However, the standard deviation calculated based on intensity mean value does not necessarily have a fixed tissue meaning. Therefore, in our study a robust intensity deviation is adopted to replace the standard deviation used in [8]. The robust deviation is computed based on the gray-value of the highest histogram bin, representing the discreteness of intensity to the gray-value of white matter. Besides, the intensity mean is more sensitive to noise than the gray value of the highest histogram bin. Thus the standard deviation calculated based on intensity mean is more sensitive to noise than the robust deviation. Given an MRI scan with voxels { , ,⋯, } , and each voxel has intensity , V , ,⋯, , the 6 robust deviation V V , where denotes the gray-value of the highest histogram bin. Our intensity normalization procedure is following: Step 1. Transform the intensity range to 0-255 linearly. Step 2. Calculate the intensity histogram, with 256 bins. Step 3. Subtract the gray-value of the highest histogram bin and divide the robust deviation . Step 4. Multiply each voxel’s intensity by a constant σ and plus a constant are below 0 or above 255 to 0 and 255 respectively. In the present study, we set . Then, set the intensities that and equal to the gray-value of the highest histogram bin and robust deviation of NO. 0001 HGG clinical training image data of BRATS 2013, which has been pre-processed by N4ITK and the step 1. For the Flair, T1c, and T2 scans, =30, 31, 37 and =75, 99, 55 respectively. The image intensity normalization effect is illustrated with T2 scans in Fig. 1. Particularly, we randomly selected 3 subjects from the BRATS 2013 and 3 subjects from the BRATS 2015. The results shown in Fig. 1 clearly demonstrated that the image intention normalization could improve comparability of different scans. The improvement is further confirmed by image intensity histograms of 30 subjects from the BRATS 2013 training dataset, as shown in Fig.2. Fig.1. T2 scans before (top row) and after (bottom row) the proposed intensity normalization. (a-1)-(a-3) and (b-1)-(b-3) are randomly selected subjects from the BRATS 2013, and (a-4)-(a-6) and (b-4)-(b-6) are randomly selected subjects from the BRATS 2015. (a-1)-(a-6): before normalization; (b-1)-(b-6): after normalization. All the scans were preprocessed by N4ITK and the proposed normalization step 1. Fig.2. Image intensity histograms of T2 scans of 30 subjects from the BRATS 2013 training dataset before (left) and after (right) the intensity normalization. All the scans were preprocessed by N4ITK and the proposed normalization step 1. 2.3.2 A deep learning model integrating FCNNs and CRFs The proposed deep learning model for brain tumor segmentation integrates Fully Convolutional Neural Networks 7 (FCNNs) and Conditional Random Fields (CRFs), as illustrated by Fig. 3. We formulated CRFs as Recurrent Neural Networks (RNNs), referred to as CRF-RNN [18]. The proposed method could segment brain images slice by slice. Fig.3. Flowchart of the proposed deep learning model integrating FCNNs and CRFs for brain tumor segmentation. (1) FCNNs The structure of our proposed FCNNs is illustrated by Fig. 4. Similar to the network architectures proposed in [28, 30], the inputs to our network are also in 2 different sizes. Passing through a series of convolutional and pooling layers, the larger inputs turn into feature maps with the same size of smaller inputs. These feature maps and smaller inputs are sent into following networks together. In this way, both local image information and context information in a larger scale can be taken into consideration for classifying image patches. Different from the cascaded architecture proposed in [30], the two branches in our FCNNs are trained simultaneously, rather than trained in different steps. Furthermore, our model has more convolutional layers. Our deep FCNNs are trained using image patches, which are extracted from slices of the axial view, coronal view or sagittal views randomly. Equal numbers of training samples for different classes are extracted to avoid data imbalance problem. There are 5 classes in total, including healthy tissue, necrosis, edema, non-enhancing core, and enhancing core. As shown in Fig. 4, in our deep FCNNs, the kernel size of each max pooling layer is set to , and the size of image patches used to train FCNNs is proportional to the kernel size. Different settings of the kernel size or equivalently the image patch size may affect the tumor segmentation performance. The max pooling layers of our FCNNs are used to capture image information in large scales with a relatively small number of network parameters. We set the stride of each layer to be 1. Therefore, in the testing stage, our model can segment brain images slice by slice. 8 Fig.4. The network structure of our deep FCNNs. (2) CRF-RNN CRF-RNN formulates 2D fully connected Conditional Random Fields as Recurrent Neural Networks [18]. Given a 2D image , comprising a set of pixels { V ,…,u} , the image segmentation problem is solved as an optimization problem using fully connected CRFs by minimizing an energy function [43] : u V V where ,h, is a certain label assignment to , ,h denotes the assignment of label to pixel h, measures the cost of assigning label assigning label and jointly to and to pixel V ,h, h denotes the probability of assigning label Differentiating Eqn. (2) with respect to , h (1) denotes the assignment of label to pixel , and the pairwise term to [19], minimizing to pixel h,h h measures the cost of , h where V V th V is the number of Gaussian kernel; coordinates of h and ( and h, , h , h , , V (3) }, h h , and h h is defined , is a Gaussian kernel, denote the intensity of and (2) , which is the variable that we aim to estimate. as h h equals to minimizing an energy , h , can be obtained from the FCNNs, and the pairwise potential , , and setting the differentiation result equal to 0, we have th { The unary term , V { , ,⋯, } are segmentation labels, the unary term , h . According function: where { ,…,u}, h (4) V exp respectively, are parameters of the Gaussian kernels); h and h h and denote spatial is a weight for the 9 Gaussian kernel intensity ( , h) ; and denote image feature vectors of h and spatial position information ( , h ); Substituting (4) into (3), we get: , , th { V and indicates the compatibility of labels , h Fully connected CRF predicts the probability of assigning label respectively, encoding their h to pixel h and . (5) }. h according to Eqn. (5), and can be calculated using a mean field iteration algorithm formulated as Recurrent Neural Networks so that CNNs and the fully connected CRF are integrated as one deep network and can be trained using a back-propagation algorithm [18]. Fig. 5 shows the network structure of CRF-RNN. G and G in Fig. 5 are two gating functions: V where output V{ , ,…,u , of one mean-field iteration; output of FCNNs, and and t th th Vt V t t V , t , , V denotes the input }, h t , t , V , , h provided by Flair, T1c and T2 slices with t denotes the final prediction results of CRF-RNN. denotes the that after softmax operation. represents the V t， V t while t denotes the denotes the mean-field iteration, is the output of FCNNs, are computed based on pixel features V h ， (6) (7) of one mean-field iteration; is the total number of mean-field iterations. In our study, the unary term and the pairwise potential , and and h with information are learned in the training phase [18]. By integrating FCNNs and CRF-RNN in one deep network, we are able to train the network end-to-end with a typical back-propagation algorithm [18]. Fig.5. The network structure of CRF-RNN. (3) The integration of FCNNs and CRF-RNN The proposed brain tumor segmentation network consists of FCNNs and CRF-RNN. The FCNNs predict the 10 probability of assigning segmentation labels to each pixel, and the CRF-RNN takes the prediction results and image information as its input to globally optimize appearance and spatial consistency of the segmentation results according to each pixel’s intensity and position information. The proposed deep learning network of FCNNs and CRF-RNN is trained in 3 steps: 1) training FCNNs using image patches; 2) training CRF-RNN using image slices with parameters of FCNNs fixed; and 3) fine-tuning the whole network using image slices. Once the fine-tune of deep learning based segmentation model is done, the model can be applied to image slices one by one for segmenting tumors. Given an t image slice with 3 channels, i.e., pre-processed Flair, T1c, and T2 scans respectively, we first pad the image slice with zeros to create 2 larger images with sizes of t t t t and tG h t tG h t respectively. Using these 2 larger images as inputs of the FCNNs, we obtain 5 label predication images , , , ,h , V , ,t,G,橔, V {h ,h , , ,th}, with the same size of the original image slices. h ,h represents one pixel’s predicted probability of brain tissue labels, such as healthy tissue, necrosis, edema, non-enhancing core or enhancing core. Then, these label predication images V{ V , ,t,G,橔} along with the image slice V { h, h, } are used as inputs to the CRF-RNN. Finally, the CRF-RNN obtains a globally optimized segmentation result of the original image slice. Fig. 3 shows the flowchart of the proposed deep learning model integrating FCNNs and CRF-RNN for brain tumor segmentation. In the training steps 2 and 3, we first calculate softmax loss according to the current segmentation results and the ground truth, and then the loss information is back-propagated to adjust network parameters of the integrated FCNNs and CRF-RNN. In the training step 2, we fix FCNNs and adjust the parameters in CRF-RNN. In the training step 3, we set a small learning rate and fine-tune the parameters of the whole network. In our experiments, the initial learning rate was set to 橔 and the learning rate was divided by 10 after each 20 epoches in the training step 1, and the learning rate was set to ㌳ and respectively in the training steps 2 and 3. 2.3.3 Fusing segmentation results obtained in axial, coronal and sagittal views We train 3 segmentation models using patches and slices of axial, coronal and sagittal views respectively. During testing, we use these 3 models to segment brain images slice by slice in 3 different views, yielding 3 segmentation results. A majority voting strategy is adopted to fuse the segmentation results. Let h , hh , and h denote the segmentation results of one voxel gotten in axial, coronal and sagittal views respectively, let h denote the segmentation result after fusion, let , , ,t,G denote a voxel labeled as healthy tissue, necrosis, edema, non-enhancing core, and enhancing core respectively, the fused segmentation result is obtained by following voting procedure: Step 1. If two or more than two of h , hh, and h are above 0, let h V . Step 2. If two or more than two of h , hh, and h equal to 1, let h V . 11 Step 3. If two or more than two of h , hh, and h equal to 3, let h V t. Step 4. If two or more than two of h , hh, and h equal to 4, let h V G. 2.3.4 Post-processing To further improve the brain tumor segmentation performance, a post-processing method is proposed. Hereinafter, h, h, denote pre-processed Flair, T1c, T2 MR images respectively, result obtained by our integrated deep learning model, denote the value of voxel t, , the voxel h h, h, , and , t, , h respectively, t, , , t, , , and t, , V , , ,t,G indicates that t, , is labeled as healthy tissue, necrosis, edema, non-enhancing core, and enhancing core t, , respectively, u in in h denotes the segmentation h and and u denote the average intensity of the whole tumor region indicated by scans. For a segmentation result t denote the average intensity of the with 3D connected tumor regions, 3D connected tumor area in h The post-processing method consists of following step: Step 1. If and h t to be 0 so that the t , set all voxels in the 3D connected tumor region is removed from and h and respectively. 3D connected tumor area in , taking into consideration that isolated local areas with super high intensities are usually caused by imaging noise rather than tumors. In the present study, V 橔 . V Step 2. If a voxel t, , ① set t, , h t, , V . satisfies the following conditions at the same time: h ,② u h , ③ t, , t, , , ④ u t t, , G, In general, tumor tissues have high signal in at least one modality of Flair, T1c, and T2. Voxels with low signal in Flair, T1c, and T2 at the same time are generally not tumor tissues. Thus, this step removes those segmented tumor regions whose intensities in Flair, T1c, T2 are below 3 thresholds respectively. However, enhancing core is an exception. In the present study, Step 3. Let t denote the volume of the V h㌳, t volume of the maximum 3D connected tumor area in connected segmented tumor region in Step 5. If h t, , G . In the present study, and t, , 橔, t V h . 3D connected tumor area in . If t Step 4. Fill the holes in Re s with necrosis. Holes in V t V h . t . t t t , remove the t is the t 3D are very likely to be necrosis. V G , set t, , V . Our model may mistakenly label necrosis areas as enhancing core. This step corrects this potential mistake through a threshold in T1c. In the present study, Step 6. Let G V t the whole tumor. If . denote the volume of enhancing core represented in t t h , h t, , h , and t, , , and V , set t denote the volume of t, , V t . Our tumor segmentation model is not sensitive to non-enhancing core. In our model, non-enhancing regions might be 12 mistakenly labeled as edema, especially when the enhancing core region is very small. In the present study, h 橔, h V ㌳橔. h V The parameters were set based on the BRATS 2013 dataset. Since the number of training cases of the BRATS 2013 is small, we did not cross-validate the parameters, therefore they are not necessarily optimal. We used the same parameters in all of our experiments, including our experiments on BRATS 2013, 2015 and 2016. In addition to the aforementioned post-processing steps, we could also directly use CRF as a post-processing step of FCNNs as did in a recent study [28]. 3. Experiments Our experiments were carried out based on imaging data provided by the BRATS 2013, 2015 and 2016 on a computing server with multiple Tesla K80 GPUs and Intel E5-2620 CPUs. However, only one GPU and one CPU were useable at the same time for our experiments. Our deep learning models were built upon Caffe [44]. Based on the BRATS 2013 data, a series of experiments were carried out to evaluate how different implementation of the proposed method affect tumor segmentation results with respect to CRF, post-processing, image patch size, number of training image patches, pre-processing, and imaging scans used. We also present segmentation results obtained for the BRATS 2013. The segmentation model was built upon the training data and then evaluated based on the testing data. Since no ground truth segmentation result for the testing data was provided, all the segmentation results were evaluated by the BRATS evaluation website. The tumor segmentation performance was evaluated using the BRATS segmentation evaluation metrics for complete tumor, core region, and enhancing region, including Dice, Positive Predictive Value (PPV), and Sensitivity. Particularly, the complete tumor includes necrosis, edema, non-enhancing core, and enhancing core; the core region includes necrosis, non-enhancing core, and enhancing core; and the enhancing region only includes the enhancing core. The tumor segmentation evaluation metrics are defined as follows: where h V , indicates complete, core or enhancing region, segmented region, and , , V , , V , denotes the manually labeled region, denotes the overlap area between and , and and denotes the denote the areas of respectively. 3.1. Experiments on BRATS 2013 dataset The BRATS 2013 training dataset contains 10 LGG and 20 HGG. Its testing dataset consists of two subsets, namely Challenge and Leaderboard. The Challenge dataset has 10 HGG cases and the Leaderboard dataset contains 21 HGG and 4 LGG. A number of experiments were carried out based on the BRATS 2013 dataset, including 1) comparing the segmentation performance of FCNNs with and without post-processing, and the performance of the proposed deep learning network integrating FCNNs and CRF-RNN (hereinafter referred to as FCNN+CRF) with and 13 without post-processing, in order to validate the effectiveness of CRFs and post-processing; 2) evaluating the segmentation performance of FCNN+CRF with 5 post-processing steps (6 post-processing steps in total), in order to test the effectiveness of each post-processing step; 3) evaluating the segmentation performance of FCNNs trained using different sizes of patches; 4) evaluating the segmentation performance of FCNNs trained using different numbers of patches; 5) comparing the segmentation performance of segmentation models built upon scans of 4 imaging sequences (Flair, T1, T1c, and T2) and 3 imaging sequences (Flair, T1c, and T2); and 6) evaluating how the image preprocessing step affect the segmentation performance. All the above experiments were performed in axial view. Apart from these experiments described above, we show the effectiveness of fusing segmentation results of three views in Section 3.1.7 and summarize comparison results with other methods in Section 3.1.8. 3.1.1. Evaluating the effectiveness of CRFs and post-processing Table 1 shows the evaluation results of FCNNs with and without post-processing, and FCNN+CRF (our integrated network of FCNNs and CRF-RNN) with and without post-processing on the BRATS 2013 Challenge dataset and Leaderboard dataset. These results demonstrated that CRFs improved the segmentation accuracy and so did the post-processing. With respect to both Dice and PPV, FCNN+post-process and FCNN+CRF improved the segmentation performance in all complete tumor, core region, and enhancing region. However, CRFs and post-process reduced Sensitivity. It is worth noting that CRFs improved Sensitivity of the enhancing region. In summary, CRFs improved both the Dice and PPV and decreased the Sensitivity on the complete and core regions, FCNN+CRF+post-process obtained the best performance with respect to Dice and PPV, but degraded the performance with respect to the Sensitivity, especially on the complete tumor region. We also adopted a 3D CRF based post-processing step as did in a recent study [28]. Particularly, the parameters of the 3D CRF were optimized by grid searching based on the training dataset of BRATS 2013. Table 1 summarizes segmentation scores obtained by our method with different settings. These results indicated that 3D CRF as a post-processing step could improve the segmentation performance as 3D information was taken into consideration. However, our proposed post-processing procedure could further improve the segmentation performance. Fig. 6 shows representative segmentation results on the BRATS 2013 Challenge dataset. These segmentation results demonstrated that FCNN+CRF could improve the spatial and appearance consistence of segmentation results, and FCNN+CRF+post-process could reduce false positives. Table 1. Evaluation results of FCNNs with and without post-processing, FCNN+CRF with and without post-processing, and FCNN+3D-CRF with and without post-processing. (The sizes of image patches used to train FCNNs were 33*33*3 and 65*65*3 respectively, n=5, and the number of patches used to train FCNNs was 5000*5*20. FCNN+CRF is short for the integrated network of FCNNs and CRF-RNN) Dataset Methods Challenge FCNNs FCNN+post-process Dice complete core 0.74 0.81 0.72 0.75 PPV enhancing complete 0.67 0.73 0.62 0.73 core 0.63 0.70 Sensitivity enhancing complete 0.60 0.73 0.94 0.94 core enhancing 0.86 0.85 0.77 0.74 14 Leaderboard FCNN+CRF FCNN+CRF+post-process FCNN+3D-CRF FCNN+3D-CRF +post-process FCNNs FCNN+post-process FCNN+CRF FCNN+CRF+post-process FCNN+3D-CRF FCNN+3D-CRF +post-process 0.85 0.87 0.85 0.80 0.83 0.80 0.70 0.76 0.73 0.87 0.92 0.84 0.80 0.87 0.78 0.63 0.77 0.69 0.84 0.83 0.88 0.81 0.81 0.83 0.80 0.77 0.80 0.87 0.83 0.78 0.89 0.85 0.78 0.86 0.83 0.79 0.70 0.81 0.83 0.86 0.84 0.61 0.65 0.66 0.73 0.65 0.54 0.61 0.57 0.62 0.61 0.58 0.74 0.85 0.89 0.81 0.57 0.63 0.71 0.76 0.71 0.49 0.62 0.50 0.64 0.57 0.96 0.94 0.85 0.84 0.90 0.74 0.78 0.69 0.78 0.71 0.67 0.66 0.71 0.68 0.71 0.87 0.71 0.63 0.88 0.74 0.63 0.88 0.79 0.70 Fig.6. Example segmentation results on the BRATS 2013 Challenge dataset. The first and second rows show the segmentation results of the 50th and 80th slice of the axial view of Subject 0301. The third and fourth rows show the segmentation results of the 40th and 70th slice of the axial view of Subject 0308. From left to right: Flair, T1c, T2, segmentation results of FCNNs, segmentation results of FCNN+CRF, and segmentation results of FCNN+CRF+post-process. In the segmentation results, each gray level represents a tumor class, from low to high: necrosis, edema, non-enhancing core, and enhancing core. 3.1.2. Evaluating the effectiveness of each post-processing step To investigate the effectiveness of each post-processing step, we obtained segmentation results of FCNN+CRF+post-x, x=1, 2, …, 6. In particular, FCNN+CRF+post-x indicates FCNN+CRF with all other post-processing steps except the step x. As described in section 2.3.4, our post-processing consists of 6 steps in total. All the evaluation results are summarized in Table 2. These results indicated that the post-processing step 3 played the most important role in the tumor segmentation, although all these post-processing steps might contribute the segmentation. Table 2. The evaluation results of FCNN+CRF with 5 of 6 post-processing steps (FCNN+CRF+post-x indicates FCNN+CRF with all other post-processing steps except the step x, the sizes of patches used to train FCNNs were 33*33*3 and 65*65*3 respectively, n=5, and the number of patches used to train FCNNs was 5000*5*20.) Dataset Methods Challenge FCNN+CRF+post-process FCNN+CRF+post-1 Dice complete core 0.87 0.87 0.83 0.83 PPV enhancing complete 0.76 0.76 0.92 0.92 core 0.87 0.87 Sensitivity enhancing complete core 0.77 0.77 0.83 0.83 0.81 0.81 enhancing 0.77 0.77 15 FCNN+CRF+post-2 FCNN+CRF+post-3 FCNN+CRF+post-4 FCNN+CRF+post-5 FCNN+CRF+post-6 FCNN+CRF+post-process FCNN+CRF+post-1 FCNN+CRF+post-2 Leaderboard FCNN+CRF+post-3 FCNN+CRF+post-4 FCNN+CRF+post-5 FCNN+CRF+post-6 0.86 0.85 0.86 0.87 0.87 0.86 0.84 0.86 0.85 0.85 0.86 0.86 0.83 0.80 0.82 0.83 0.83 0.73 0.71 0.73 0.70 0.71 0.72 0.70 0.76 0.72 0.76 0.74 0.76 0.62 0.62 0.62 0.59 0.62 0.59 0.62 0.90 0.88 0.92 0.92 0.92 0.89 0.88 0.88 0.87 0.89 0.89 0.89 0.87 0.80 0.88 0.87 0.87 0.76 0.75 0.76 0.73 0.77 0.76 0.76 0.77 0.69 0.77 0.70 0.77 0.64 0.64 0.64 0.57 0.64 0.57 0.64 0.84 0.83 0.82 0.83 0.83 0.84 0.82 0.86 0.84 0.82 0.84 0.84 0.81 0.81 0.78 0.81 0.81 0.78 0.77 0.77 0.75 0.74 0.76 0.71 0.77 0.77 0.77 0.79 0.77 0.68 0.68 0.68 0.68 0.68 0.71 0.68 3.1.3. Evaluating the impact of image patch size We used different kernel sizes in all the pooling layers to train different FCNNs. The training image patch size changed with the kernel size, as shown in Fig. 4, while the number of parameters in FCNNs was unchanged. We evaluated the segmentation performance of our segmentation models with n=1,3,5, as summarized in Table 3. When n=1,3,5, the corresponding sizes of training patches are 21*21*3 (small input patch) and 41*41*3 (large input patch), 27*27*3 and 53*53*3, 33*33*3 and 65*65*3. Bar plots of the Dice of complete regions on the Challenge dataset with different training patch sizes are shown in Fig. 7. These segmentation results indicated that 1) a bigger patch provided more information and helped improve FCNNs’ performance; 2) the CRF-RNN could reduce the performance differences caused by patch size as CRF-RNN could optimize the segmentation results according to the information in a whole image slice; and 3) the post-processing could further reduce the performance difference caused by patch size. Fig.7. Bar plots for the Dice of complete regions on the Challenge dataset with different training patch sizes. Table 3. Evaluation results of our segmentation model with n=1,2,3 (The number of patches used to train FCNNs was 5000*5*20.) Dataset n 1 Challenge 3 5 Leaderboard 1 Methods FCNNs FCNN+CRF FCNN+CRF+post-process FCNNs FCNN+CRF FCNN+CRF+post-process FCNNs FCNN+CRF FCNN+CRF+post-process FCNNs Dice complete core 0.70 0.83 0.86 0.71 0.83 0.86 0.74 0.85 0.87 0.64 0.62 0.78 0.83 0.67 0.77 0.82 0.72 0.80 0.83 0.54 PPV enhancing complete 0.64 0.69 0.76 0.65 0.69 0.76 0.67 0.70 0.76 0.49 0.57 0.85 0.92 0.58 0.83 0.91 0.62 0.87 0.92 0.50 core 0.50 0.79 0.88 0.56 0.74 0.84 0.63 0.80 0.87 0.45 Sensitivity enhancing complete core enhancing 0.55 0.63 0.77 0.57 0.63 0.78 0.60 0.63 0.77 0.42 0.94 0.83 0.82 0.95 0.85 0.84 0.94 0.84 0.83 0.96 0.87 0.79 0.80 0.86 0.82 0.82 0.86 0.81 0.81 0.76 0.79 0.79 0.77 0.78 0.79 0.76 0.77 0.80 0.77 0.71 16 3 5 FCNN+CRF FCNN+CRF+post-process FCNNs FCNN+CRF FCNN+CRF+post-process FCNNs FCNN+CRF FCNN+CRF+post-process 0.82 0.85 0.67 0.83 0.86 0.70 0.83 0.86 0.65 0.72 0.57 0.66 0.72 0.61 0.66 0.73 0.57 0.61 0.52 0.57 0.62 0.54 0.57 0.62 0.83 0.88 0.54 0.82 0.88 0.58 0.85 0.89 0.70 0.76 0.50 0.67 0.74 0.57 0.71 0.76 0.51 0.59 0.45 0.51 0.60 0.49 0.50 0.64 0.85 0.84 0.96 0.87 0.86 0.96 0.85 0.84 0.69 0.76 0.75 0.72 0.79 0.74 0.69 0.78 0.72 0.70 0.70 0.71 0.69 0.67 0.71 0.68 3.1.4. Evaluating the impact of the number of image patches used to train FCNNs The BRATS 2013 training dataset contains 10 LGG and 20 HGG. In our experiments, we trained our segmentation model using the 20 HGG cases, and our model worked well for segmenting LGG cases. To investigate the impact of the number of training patches on the segmentation performance, we trained FCNNs with varied numbers of image patches. In particular, we sampled training imaging patches randomly from each subject and kept the number of training samples for different classes equal (5 classes in total, including normal tissue, necrosis, edema, non-enhancing core, and enhancing core). We generated 3 sets of image patches by sampling, consisting of 1000*5*20, 3000*5*20, and 5000*5*20 patches respectively, and used them to train different segmentation models. The evaluation results are summarized in in Table 4. Bar plots of the Dice of complete regions on the Challenge dataset with different numbers of training patches are shown in Fig. 8. The results shown in Table 4 and Fig.8 indicated that the brain tumor segmentation accuracy of FCNNs increased with the increasing of the number of training patches. However, both CRFs and post-processing could reduce the performance difference. Fig.8. Bar plots for the Dice of complete regions on the Challenge dataset with different numbers of training image patches. Table.4 Evaluation results of the segmentation models trained using different numbers of training image patches (the sizes of image patches used to train FCNNs were 33*33*3 and 65*65*3 respectively, n=5) Dataset Challenge No. of patches Methods complete FCNNs 0.71 1000*5*20 FCNN+CRF 0.84 FCNN+CRF+post-process 0.87 FCNNs 0.73 3000*5*20 FCNN+CRF 0.84 FCNN+CRF+post-process 0.86 FCNNs 0.74 5000*5*20 FCNN+CRF 0.85 FCNN+CRF+post-process 0.87 Dice core enhancing complete 0.68 0.62 0.57 0.78 0.68 0.84 0.82 0.76 0.91 0.70 0.66 0.60 0.80 0.69 0.86 0.83 0.77 0.92 0.72 0.67 0.62 0.80 0.70 0.87 0.83 0.76 0.92 PPV Sensitivity core enhancing complete core enhancing 0.58 0.54 0.95 0.86 0.76 0.78 0.61 0.85 0.81 0.79 0.86 0.77 0.84 0.81 0.77 0.61 0.59 0.94 0.86 0.76 0.79 0.62 0.84 0.82 0.80 0.86 0.77 0.83 0.82 0.77 0.63 0.60 0.94 0.86 0.77 0.80 0.63 0.84 0.81 0.80 0.87 0.77 0.83 0.81 0.77 17 FCNNs FCNN+CRF FCNN+CRF+post-process FCNNs 3000*5*20 FCNN+CRF Learderboard FCNN+CRF+post-process FCNNs 5000*5*20 FCNN+CRF FCNN+CRF+post-process 1000*5*20 0.66 0.83 0.86 0.69 0.83 0.85 0.70 0.83 0.86 0.58 0.65 0.72 0.60 0.65 0.72 0.61 0.66 0.73 0.51 0.56 0.61 0.52 0.56 0.61 0.54 0.57 0.62 0.53 0.82 0.88 0.57 0.84 0.89 0.58 0.85 0.89 0.53 0.71 0.76 0.56 0.70 0.75 0.57 0.71 0.76 0.45 0.51 0.63 0.47 0.49 0.59 0.49 0.50 0.64 0.96 0.86 0.85 0.96 0.85 0.84 0.96 0.85 0.84 0.74 0.69 0.77 0.74 0.69 0.77 0.74 0.69 0.78 0.67 0.70 0.68 0.67 0.71 0.68 0.67 0.71 0.68 In summary, all the experimental results demonstrated that both CRFs and the post-processing method can narrow the performance difference caused by training patch sizes and training patch numbers, indicating that CRFs and the post-processing method might be able to narrow the performance difference caused by other training tricks. We will confirm this inference in our future work. 3.1.5. Performance Comparison between segmentation models built upon 4 and 3 imaging modalities We also built a segmentation model using all available 4 imaging modalities, i.e., Flair, T1, T1c, and T2, and compared its segmentation performance with that of the segmentation model built upon 3 imaging modalities, i.e., Flair, T1c, and T2. The segmentation results of these segmentation models are summarized in Table 5. These results demonstrated that these two segmentation models achieved similar performance, indicating that a segmentation model built upon Flair, T1c, and T2 could achieve competitive performance as the model built upon 4 imaging modalities. Table.5 Performance comparison of segmentation models built upon scans of 4 imaging modalities and 3 imaging modalities (the sizes of patches used to train FCNNs were 33*33*4 and 65*65*4, or 33*33*3 and 65*65*3 respectively, n=5, and the number of patches used to train FCNNs was 5000*5*20) Dataset No. of modalities 4 Challenge 3 4 Learderboard 3 Methods FCNNs FCNN+CRF FCNN+CRF+post-process FCNNs FCNN+CRF FCNN+CRF+post-process FCNNs FCNN+CRF FCNN+CRF+post-process FCNNs FCNN+CRF FCNN+CRF+post-process complete 0.74 0.85 0.87 0.74 0.85 0.87 0.69 0.82 0.85 0.70 0.83 0.86 Dice PPV Sensitivity core enhancing complete core enhancing complete core enhancing 0.72 0.67 0.61 0.64 0.60 0.95 0.86 0.77 0.80 0.71 0.85 0.79 0.67 0.87 0.82 0.78 0.83 0.76 0.91 0.85 0.78 0.84 0.82 0.76 0.72 0.67 0.62 0.63 0.60 0.94 0.86 0.77 0.80 0.70 0.87 0.80 0.63 0.84 0.81 0.80 0.83 0.76 0.92 0.87 0.77 0.83 0.81 0.77 0.59 0.55 0.57 0.56 0.49 0.95 0.72 0.68 0.64 0.58 0.82 0.70 0.55 0.85 0.67 0.68 0.74 0.62 0.89 0.75 0.62 0.83 0.79 0.66 0.61 0.54 0.58 0.57 0.49 0.96 0.74 0.67 0.66 0.57 0.85 0.71 0.50 0.85 0.69 0.71 0.73 0.62 0.89 0.76 0.64 0.84 0.78 0.68 3.1.6. Evaluation of different pre-processing strategies on the tumor segmentation We preprocessed the imaging data using our robust deviation based intensity normalization and the standard deviation based intensity normalization [8], and then evaluated segmentation models built on them separately. As the results shown in Table 6 indicated, the robust deviation based intensity normalization could slightly improve the segmentation performance. Table.6 Performance comparison of segmentation models built upon images normalized by the robust deviation and the standard deviation (the sizes of patches used to train FCNNs were 33*33*3 and 65*65*3 respectively, n=5, and the number of patches used to train FCNNs was 5000*5*20) Dataset Deviation Methods standard FCNNs FCNN+CRF Dice PPV Sensitivity complete core enhancing complete core enhancing complete core enhancing 0.73 0.69 0.67 0.61 0.60 0.61 0.94 0.86 0.75 0.84 0.80 0.71 0.87 0.79 0.68 0.82 0.82 0.75 18 Challenge FCNN+CRF+post-process FCNNs robust FCNN+CRF FCNN+CRF+post-process FCNNs standard FCNN+CRF FCNN+CRF+post-process Learderboard FCNNs robust FCNN+CRF FCNN+CRF+post-process 0.86 0.74 0.85 0.87 0.69 0.83 0.86 0.70 0.83 0.86 0.83 0.72 0.80 0.83 0.60 0.66 0.73 0.61 0.66 0.73 0.76 0.67 0.70 0.76 0.54 0.58 0.61 0.54 0.57 0.62 0.93 0.62 0.87 0.92 0.57 0.85 0.89 0.58 0.85 0.89 0.86 0.63 0.80 0.87 0.55 0.71 0.75 0.57 0.71 0.76 0.80 0.60 0.63 0.77 0.50 0.56 0.66 0.49 0.50 0.64 0.81 0.94 0.84 0.83 0.97 0.85 0.84 0.96 0.85 0.84 0.81 0.86 0.81 0.81 0.75 0.70 0.78 0.74 0.69 0.78 0.74 0.77 0.80 0.77 0.67 0.67 0.66 0.67 0.71 0.68 3.1.7. Evaluating the effectiveness of fusing the segmentation results gotten in three views We trained 3 segmentation models using patches and slices obtained in axial, coronal and sagittal views respectively. During testing, we used these three models to segment brain images from 3 views and got three segmentation results. The results of different views were fused and the evaluation results are shown in Table 7. Evaluation results in Table 7 indicated that, for both Challenge and Leaderboard datasets, fusing the segmentation results typically led to better segmentation performance without the post-processing procedure. However, the improvement became insignificant after the post-processing procedure was applied to the segmentation results. Table.7 Evaluations of segmentation results obtained in axial, coronal, sagittal views before and after post-processing, and evaluations of fusion results before and after post-processing (the sizes of patches used to train FCNNs were 33*33*3 and 65*65*3 respectively, n=5, and the number of patches used to train FCNNs was 5000*5*20) Dataset 2013 Methods FCNN+CRF Challenge fusing (FCNN+CRF) axial coronal sagittal fusing (FCNN+CRF)+post-process axial FCNN+CRF coronal sagittal fusing (FCNN+CRF) axial FCNN+CRF+post-process coronal sagittal fusing (FCNN+CRF)+post-process FCNN+CRF+post-process Learderboard axial coronal sagittal complete 0.85 0.82 0.83 0.85 0.87 0.87 0.87 0.88 0.83 0.80 0.80 0.84 0.86 0.86 0.84 0.86 Dice PPV Sensitivity core enhancing complete core enhancing complete core enhancing 0.80 0.70 0.87 0.80 0.63 0.84 0.81 0.80 0.80 0.70 0.77 0.78 0.62 0.90 0.83 0.82 0.80 0.71 0.80 0.79 0.64 0.87 0.83 0.82 0.83 0.74 0.85 0.84 0.68 0.88 0.82 0.82 0.83 0.76 0.92 0.87 0.77 0.83 0.81 0.77 0.82 0.76 0.86 0.83 0.74 0.89 0.83 0.79 0.82 0.77 0.88 0.83 0.75 0.86 0.82 0.79 0.84 0.77 0.90 0.87 0.76 0.86 0.82 0.80 0.66 0.57 0.85 0.71 0.50 0.85 0.69 0.71 0.65 0.56 0.75 0.69 0.50 0.90 0.69 0.70 0.66 0.55 0.78 0.69 0.49 0.86 0.69 0.70 0.67 0.60 0.84 0.74 0.54 0.87 0.68 0.72 0.73 0.62 0.89 0.76 0.64 0.84 0.78 0.68 0.73 0.62 0.86 0.73 0.60 0.89 0.79 0.67 0.71 0.60 0.85 0.74 0.59 0.84 0.74 0.67 0.73 0.62 0.89 0.77 0.60 0.85 0.77 0.69 3.1.8. Comparison with other methods Comparison results with other methods are summarized in Table 8. In particular, evaluation results of the top ranked methods participated in the BRATS 2013, shown on the BRATS 2013 website, are summarized in Table 8, along with the results of our method and other two state of art methods. Particularly, the method proposed by Sergio Pereira et al [31] ranked first on the Challenge dataset and second on the Leaderboard dataset right now, while our method ranked second on the Challenge dataset and first on the Leaderboard dataset right now. In general, it took 2-4 min for one of the three views of our method to segment one subject’s imaging data. We do 19 not have an accurate estimation of the training time for our segmentation models since we used a shared GPU server. On the shared GPU server, it took ~12 days to train our segmentation models. Table. 8 Comparisons with other methods on BRATS 2013 dataset Dataset Challenge Leaderboard Methods Nick Tustison Raphael Meier Syed Reza Mohammad Havaei[30] Sergio Pereira[31] Our method(axial) Our method(coronal) Our method(sagittal) Our method(fusing) Nick Tustison Liang Zhao Raphael Meier Mohammad Havaei[30] Sergio Pereira[31] Our method(axial) Our method(coronal) Our method(sagittal) Our method(fusing) complete 0.87 0.82 0.83 0.88 0.88 0.87 0.87 0.87 0.88 0.79 0.79 0.72 0.84 0.84 0.86 0.86 0.84 0.86 Dice core 0.78 0.73 0.72 0.79 0.83 0.83 0.82 0.82 0.84 0.65 0.59 0.60 0.71 0.72 0.73 0.73 0.71 0.73 Positive Predictive Value enhancing complete core enhancing 0.74 0.85 0.74 0.69 0.69 0.76 0.78 0.71 0.72 0.82 0.81 0.70 0.73 0.89 0.79 0.68 0.77 0.88 0.87 0.74 0.76 0.92 0.87 0.77 0.76 0.86 0.83 0.74 0.77 0.88 0.83 0.75 0.77 0.90 0.87 0.76 0.53 0.83 0.70 0.51 0.47 0.77 0.55 0.50 0.53 0.65 0.62 0.48 0.57 0.88 0.79 0.54 0.62 0.85 0.82 0.60 0.62 0.89 0.76 0.64 0.62 0.86 0.73 0.60 0.60 0.85 0.74 0.59 0.62 0.89 0.77 0.60 complete 0.89 0.92 0.86 0.87 0.89 0.83 0.89 0.86 0.86 0.81 0.85 0.88 0.84 0.86 0.84 0.89 0.84 0.85 Sensitivity core enhancing 0.88 0.83 0.72 0.73 0.69 0.76 0.79 0.80 0.83 0.81 0.81 0.77 0.83 0.79 0.82 0.79 0.82 0.80 0.73 0.66 0.77 0.53 0.69 0.64 0.72 0.68 0.76 0.68 0.78 0.68 0.79 0.67 0.74 0.67 0.77 0.69 3.2. Segmentation performance on the BRATS 2015 The BRATS 2015 training dataset contains 54 LGG and 220 HGG, and its testing dataset contains 110 cases with unknown grades. All the training cases in BRATS 2013 are reused in BRATS 2015 as a subset of its training dataset. As what we have done in our experiments with BRATS 2013 dataset, we just used HGG training cases to train our segmentation models in this section. We extracted 1000*5 patches from each of 220 HGG to train FCNNs and initialed the CRF-RNN by the CRF-RNN trained by BRATS 2013 dataset. The whole network was fine-tuned using slices in BRATS 2013 training dataset, which is a subset of BRATS 2015 training dataset. The evaluation results of BRATS 2015 testing dataset are shown in Table 9, including evaluations of segmentation results that were segmented by the models trained by BRATS 2013 training dataset and evaluations of segmentation results that were segmented by the models trained in this section. The results shown in Table. 9 indicated that fusing the segmentation results of multi-views could improve the segmentation accuracy. These results also indicated that a larger training dataset might improve the segmentation performance. Table.9 Evaluation results of 110 testing cases in BRATS 2015 testing dataset (the sizes of image patches used to train FCNNs were 33*33*3 and 65*65*3 respectively, n=5) Methods FCNN+CRF axial coronal sagittal Models trained based fusing (FCNN+CRF) on the BRATS axial 2013 training FCNN+CRF+post-process coronal dataset sagittal fusing (FCNN+CRF)+post-process complete 0.77 0.74 0.74 0.79 0.78 0.78 0.77 0.81 Dice PPV Sensitivity core enhancing complete core enhancing complete core enhancing 0.56 0.52 0.85 0.71 0.46 0.74 0.53 0.67 0.53 0.48 0.70 0.69 0.44 0.84 0.50 0.65 0.54 0.49 0.74 0.67 0.43 0.78 0.52 0.66 0.56 0.56 0.82 0.78 0.52 0.79 0.50 0.68 0.64 0.58 0.87 0.75 0.57 0.73 0.61 0.65 0.61 0.57 0.77 0.68 0.58 0.84 0.63 0.63 0.62 0.57 0.80 0.71 0.56 0.78 0.60 0.63 0.65 0.60 0.87 0.79 0.60 0.78 0.61 0.65 20 FCNN+CRF axial coronal sagittal Models trained based fusing (FCNN+CRF) on the BRATS axial 2015 training FCNN+CRF+post-process coronal dataset sagittal fusing (FCNN+CRF)+post-process 0.78 0.77 0.76 0.80 0.80 0.80 0.78 0.82 0.64 0.66 0.63 0.66 0.68 0.69 0.69 0.72 0.54 0.56 0.47 0.57 0.61 0.61 0.57 0.62 0.78 0.73 0.75 0.81 0.82 0.77 0.80 0.84 0.76 0.73 0.71 0.79 0.75 0.70 0.74 0.78 0.48 0.52 0.38 0.50 0.59 0.61 0.51 0.60 0.81 0.86 0.80 0.83 0.81 0.87 0.80 0.83 0.62 0.67 0.63 0.64 0.71 0.76 0.71 0.73 0.71 0.67 0.75 0.72 0.68 0.65 0.72 0.69 There were only 53 testing cases available during the BRATS 2015, but now there are 110 testing cases. Therefore, we are not able to directly compare our method with the methods that participated in the BRATS 2015. We are aware that K. Kamnitses et al [28] have published their evaluation results with 110 BRATS 2015 testing cases. The comparisons with K. Kamnitses et al’s method are shown in Table 10. Table. 10 Comparisons with other methods on BRATS 2015 dataset Dataset Methods BRATS 2015 testing dataset DeepMedic+CRF[28] Our method(axial) Our method(coronal) Our method(sagittal) Our method(fusing) Our FCNN(axial)+3D CRF Our FCNN(coronal)+3D CRF Our FCNN(sagittal)+3D CRF Our FCNN+3D CRF (fusing) complete 0.847 0.80 0.80 0.78 0.82 0.84 0.83 0.82 0.84 Dice core 0.67 0.68 0.69 0.69 0.72 0.72 0.72 0.72 0.73 Positive Predictive Value enhancing complete core enhancing 0.629 0.85 0.848 0.634 0.61 0.82 0.75 0.59 0.61 0.77 0.70 0.61 0.57 0.80 0.74 0.51 0.62 0.84 0.78 0.60 0.62 0.88 0.75 0.62 0.62 0.88 0.75 0.62 0.60 0.88 0.75 0.59 0.62 0.89 0.76 0.63 complete 0.876 0.81 0.87 0.80 0.83 0.82 0.82 0.81 0.82 Sensitivity core enhancing 0.607 0.662 0.71 0.68 0.76 0.65 0.71 0.72 0.73 0.69 0.76 0.67 0.75 0.66 0.76 0.67 0.76 0.67 3.3. Segmentation performance on the BRATS 2016 We also participated in the BRATS 2016. However, during the competition we just segmented brain images in axial view. Since the BRATS 2016 shares the same training dataset with BRATS 2015, we used the same segmentation models trained based on the BRATS 2015 training dataset. The BRATS 2015 training dataset has been pre-processed with rigid registration, bias field correction and skull stripping. However, the BRATS 2016 test dataset contains a number of unprocessed or partially pre-processed images [45], as shown in Fig. 9. To reduce false positives caused by incomplete pre-processing, apart from the post-processing steps described in Section 2.3.4, we manually placed rectangular bounding boxes around tumors in images and applied the post-processing step 2 to the segmentation results. Among 19 teams participated in the BRATS 2016, we ranked first on the multi-temporal evaluation. The ranking details of our method are shown in Table.11. Fig.9 An example of partial skull stripping case in BRATS 2016 testing dataset. From left to right: Flair, T1c, T2 21 Table. 11 The ranking details of our method on different items on BRAST 2016 (including tie) Items Ranking complete 4 Dice core 3 Tumor Segmentation enhancing 1 complete 7 Hausdorff core enhancing 6 2 Multi-temporal evaluation 1 4. Discussions and Conclusion In this study, we proposed a novel deep learning based brain tumor segmentation method by integrating Fully Convolutional Neural Networks (FCNNs) and Conditional Random Fields (CRFs) in a unified framework. This integrated model was designed to obtain tumor segmentation results with appearance and spatial consistency. In our method, we used CRF-RNN to implement CRFs [18], facilitating easy training of both FCNNs and CRFs as one deep network, rather than using CRFs as a post-processing step of FCNNs. Our integrated deep learning model was trained in 3 steps, using image patches and slices respectively. In the first step, image patches were used to train FCNNs. These image patches were randomly sampled from the training dataset and the same number of image patches for each class was used as training image patches, in order to avoid the data imbalance problem. In the second step, image slices were used to train the following CRF-RNN, with parameters of FCNNs fixed. In the third step, image slices were used to fine-tune the whole network. Particularly, we train 3 segmentation models using 2D image patches and slices obtained in axial, coronal and sagittal views respectively, and combine them to segment brain tumors using a voting based fusion strategy. Our experimental results also indicated that the integration of FCNNs and CRF-RNN could improve the segmentation robustness to parameters involved in the model training, such as image patch size and the number of training image patches. Our experimental results also demonstrated that a tumor segmentation model built upon Flair, T1c, and T2 scans achieved competitive performance as those built upon Flair, T1, T1c, and T2 scans. We also proposed a simple pre-processing strategy and a simple post-processing strategy. We pre-processed each MR scan using N4ITK and intensity normalization, which normalized each MR image’s intensity mainly by subtracting the gray-value of the highest frequency and dividing the robust deviation. The results shown in Fig. 1 and Fig. 2 demonstrated that the proposed intensity normalization method could make different MRI scans comparable, i.e., similar intensity values characterize similar brain tissues across scans. We post-processed the segmentation results by removing small 3D-connected regions and correcting false labels by simple thresholding method. Our experimental results have demonstrated that these strategies could improve the tumor segmentation performance. Our method has achieved promising performance on the BRATS 2013 and BRATS 2015 testing dataset. Different from other top ranked methods, our method could achieve competitive performance with only 3 imaging modalities (Flair, T1c, and T2), rather than 4 (Flair, T1, T1c, and T2). We also participated in the BRATS 2016 and our method ranked first on its multi-temporal evaluation. 22 Our method is built upon 2D FCNNs and CRF-RNN to achieve computational efficiency. For training CRF-RNN and fine-tuning the integrated FCNNs and CRF-RNN, we use image slices as training data. However, in image slices, the numbers of pixels for different classes are different, which may worsen the segmentation performance of the trained network. To partially overcome the imbalanced training data problem, we trained CRF-RNN with the parameters of FCNNs were fixed so that the CRF-RNN are trained to optimize the appearance and spatial consistency of segmentation results. Such a strategy in conjunct with a fine-tuning of the whole network with a small learning rate improved the tumor segmentation performance. However, 2D CNNs are not equipped to take full advantage of 3D information of the MRI data [28, 29]. Our experimental results have demonstrated adopting 3D CRF as a post-processing step could improve the tumor segmentation performance. Our ongoing study is to build a fully 3D network to further improve the tumor segmentation performance. Acknowledgements This work was supported in part by the National High Technology Research and Development Program of China (2015AA020504), the National Natural Science Foundation of China under Grant Nos. 61572499, 61421004, 61473296, and NIH grants EB022573, CA189523. References [1] S. Bauer, R. Wiest, L.-P. Nolte, and M. Reyes, "A survey of MRI-based medical image analysis for brain tumor studies," Physics in medicine and biology, vol. 58, pp. 97-129, 2013. [2] M. Goetz, C. Weber, F. Binczyk, J. Polanska, R. Tarnawski, B. Bobek-Billewicz, et al., "DALSA: domain adaptation for supervised learning from sparsely annotated MR images," IEEE transactions on medical imaging, vol. 35, pp. 184-196, 2016. [3] B. H. Menze, A. Jakab, S. Bauer, J. Kalpathy-Cramer, K. Farahani, J. Kirby, et al., "The multimodal brain tumor image segmentation benchmark (BRATS)," IEEE Transactions on Medical Imaging, vol. 34, pp. 1993-2024, 2015. [4] A. Gooya, K. M. Pohl, M. Bilello, L. Cirillo, G. Biros, E. R. Melhem, et al., "GLISTR: glioma image segmentation and registration," IEEE transactions on medical imaging, vol. 31, pp. 1941-1954, 2012. [5] M. B. Cuadra, C. Pollo, A. Bardera, O. Cuisenaire, J.-G. Villemure, and J.-P. Thiran, "Atlas-based segmentation of pathological MR brain images using a model of lesion growth," IEEE transactions on medical imaging, vol. 23, pp. 1301-1314, 2004. [6] B. H. Menze, K. Van Leemput, D. Lashkari, M.-A. Weber, N. Ayache, and P. Golland, "A generative model for brain tumor segmentation in multi-modal images," in International Conference on Medical Image Computing and Computer-Assisted Intervention, 2010, pp. 151-159. [7] M. Prastawa, E. Bullitt, S. Ho, and G. Gerig, "A brain tumor segmentation framework based on outlier detection," Medical image analysis, vol. 8, pp. 275-283, 2004. [8] M. Goetz, C. Weber, J. Bloecher, B. Stieltjes, H.-P. Meinzer, and K. Maier-Hein, "Extremely randomized trees based brain tumor segmentation," Proceedings MICCAI BraTS (Brain Tumor Segmentation Challenge), pp. 6-11, 2014. [9] S. M. S. Reza and K. M. Iftekharuddin, "Improved brain tumor tissue segmentation using texture features," Proceedings MICCAI BraTS (Brain Tumor Segmentation Challenge), pp. 27-30, 2014. [10] J. Kleesiek, A. Biller, G. Urban, U. Kothe, M. Bendszus, and F. Hamprecht, "Ilastik for multi-modal brain tumor segmentation," Proceedings MICCAI BraTS (Brain Tumor Segmentation Challenge), pp. 12-17, 2014. [11] S. Ruan, S. Lebonvallet, A. Merabet, and J.-M. Constans, "Tumor segmentation from a multispectral MRI images by using support vector machine classification," in 2007 4th IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2007, pp. 1236-1239. [12] H. Li and Y. Fan, "Label propagation with robust initialization for brain tumor segmentation," in 2012 9th IEEE International Symposium on Biomedical Imaging (ISBI), 2012, pp. 1715-1718. 23 [13] H. Li, M. Song, and Y. Fan, "Segmentation of brain tumors in multi-parametric MR images via robust statistic information propagation," in Asian Conference on Computer Vision, 2010, pp. 606-617. [14] R. Meier, S. Bauer, J. Slotboom, R. Wiest, and M. Reyes, "Appearance- and context-sensitive features for brain tumor segmentation," Proceedings MICCAI BraTS (Brain Tumor Segmentation Challenge), pp. 20-26, 2014. [15] A. Krizhevsky, I. Sutskever, and G. E. Hinton, "Imagenet classification with deep convolutional neural networks," in Advances in neural information processing systems, 2012, pp. 1097-1105. [16] R. Girshick, J. Donahue, T. Darrell, and J. Malik, "Rich feature hierarchies for accurate object detection and semantic segmentation," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2014, pp. 580-587. [17] J. Long, E. Shelhamer, and T. Darrell, "Fully convolutional networks for semantic segmentation," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015, pp. 3431-3440. [18] S. Zheng, S. Jayasumana, B. Romera-Paredes, V. Vineet, Z. Su, D. Du, et al., "Conditional random fields as recurrent neural networks," in Proceedings of the IEEE International Conference on Computer Vision, 2015, pp. 1529-1537. [19] Z. Liu, X. Li, P. Luo, C.-C. Loy, and X. Tang, "Semantic image segmentation via deep parsing network," in Proceedings of the IEEE International Conference on Computer Vision, 2015, pp. 1377-1385. [20] D. Zikic, Y. Ioannou, M. Brown, and A. Criminisi, "Segmentation of brain tumor tissues with convolutional neural networks," Proceedings MICCAI BraTS (Brain Tumor Segmentation Challenge), pp. 36-39, 2014. [21] A. Davy, M. Havaei, D. Warde-farley, A. Biard, L. Tran, P.-M. Jodoin, et al., "Brain tumor segmentation with deep neural networks," Proceedings MICCAI BraTS (Brain Tumor Segmentation Challenge), pp. 1-5, 2014. [22] G. Urban, M. Bendszus, F. Hamprecht, and J. Kleesiek, "Multi-modal brain tumor segmentatioin using deep convolutional neural networks," Proceedings MICCAI BraTS (Brain Tumor Segmentation Challenge), pp. 31-35, 2014. [23] P. Dvorak and B. H. Menze, "Structured prediction with convolutional neural networks for multimodal brain tumor segmentation," Proceedings MICCAI BraTS (Brain Tumor Segmentation Challenge), pp. 13-24, 2015. [24] M. Havaei, F. Dutil, C. Pal, H. Larochelle, and P.-M. Jodoin, "A convolutional neural network approach to brain tumor segmentation," Proceedings MICCAI BraTS (Brain Tumor Segmentation Challenge), pp. 29-33, 2015. [25] S. Pereira, A. Pinto, V. Alves, and C. A. Silva, "Deep convolutional neural networks for the segmentatin of gliomas in multi-sequence MRI," Proceedings MICCAI BraTS (Brain Tumor Segmentation Challenge), pp. 52-55, 2015. [26] M. Agn, O. Puonti, I. Law, P. M. a. Rosenschold, and K. V. Leemput, "Brain tumor segmentation by a generative model with a prior on tumor shape," Proceedings MICCAI BraTS (Brain Tumor Segmentation Challenge), pp. 1-4, 2015. [27] K. Vaidhya, R. Santhosh, S. Thirunavukkarasu, V. Alex, and G. Krishnamurthi, "Multi-modal brain tumor segmentation using stacked denoising autoencoders," in Proceedings MICCAI BraTS (Brain Tumor Segmentation Challenge), 2015, pp. 60-64. [28] K. Kamnitsas, C. Ledig, V. F. Newcombe, J. P. Simpson, A. D. Kane, D. K. Menon, et al., "Efficient multi-scale 3D CNN with fully connected CRF for accurate brain lesion segmentation," Medical Image Analysis, vol. 36, pp. 61-78, 2017. [29] D. Yi, M. Zhou, Z. Chen, and O. Gevaert, "3-D Convolutional Neural Networks for Glioblastoma Segmentation," arXiv preprint arXiv:1611.04534, 2016. [30] M. Havaei, A. Davy, D. Warde-Farley, A. Biard, A. Courville, Y. Bengio, et al., "Brain tumor segmentation with deep neural networks," Medical image analysis, vol. 35, pp. 18-31, 2017. [31] S. Pereira, A. Pinto, V. Alves, and C. A. Silva, "Brain tumor segmentation using convolutional neural networks in MRI images," IEEE transactions on medical imaging, vol. 35, pp. 1240-1251, 2016. [32] K. Simonyan and A. Zisserman, "Very deep convolutional networks for large-scale image recognition," arXiv preprint arXiv:1409.1556, 2014. [33] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille, "Semantic image segmentation with deep convolutional nets and fully connected crfs," arXiv preprint arXiv:1412.7062, 2014. [34] A. Prasoon, K. Petersen, C. Igel, F. Lauze, E. Dam, and M. Nielsen, "Deep feature learning for knee cartilage segmentation using a triplanar convolutional neural network," in International conference on medical image computing and computer-assisted intervention, 2013, pp. 246-253. [35] K. Fritscher, P. Raudaschl, P. Zaffino, M. F. Spadea, G. C. Sharp, and R. Schubert, "Deep Neural Networks for Fast Segmentation of 3D Medical Images," in International Conference on Medical Image Computing and Computer-Assisted Intervention, 2016, pp. 158-165. [36] A. A. A. Setio, F. Ciompi, G. Litjens, P. Gerke, C. Jacobs, S. J. van Riel, et al., "Pulmonary nodule detection in CT images: false positive reduction using multi-view convolutional networks," IEEE transactions on medical imaging, vol. 35, pp. 1160-1169, 2016. [37] X. Zhao, Y. Wu, G. Song, Z. Li, Y. Fan, and Y. Zhang, "Brain Tumor Segmentation Using a Fully Convolutional Neural Network with Conditional Random Fields," in International Workshop on Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries, 2016, pp. 75-87. 24 [38] W. Zhang, R. Li, H. Deng, L. Wang, W. Lin, S. Ji, et al., "Deep convolutional neural networks for multi-modality isointense infant brain image segmentation," NeuroImage, vol. 108, pp. 214-224, 2015. [39] P. Moeskops, M. A. Viergever, A. M. Mendrik, L. S. de Vries, M. J. Benders, and I. Išgum, "Automatic segmentation of MR brain images with a convolutional neural network," IEEE transactions on medical imaging, vol. 35, pp. 1252-1261, 2016. [40] A. de Brebisson and G. Montana, "Deep neural networks for anatomical brain segmentation," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2015, pp. 20-28. [41] N. J. Tustison, B. B. Avants, P. A. Cook, Y. Zheng, A. Egan, P. A. Yushkevich, et al., "N4ITK: improved N3 bias correction," IEEE transactions on medical imaging, vol. 29, pp. 1310-1320, 2010. [42] R. D. Fields, "Change in the brain's white matter," Science, vol. 330, pp. 768-769, 2010. [43] P. Krahenbuhl and V. Koltun, "Efficient inference in fully connected crfs with gaussian edge potentials," NIPS, 2011. [44] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, et al., "Caffe: Convolutional architecture for fast feature embedding," in Proceedings of the 22nd ACM international conference on Multimedia, 2014, pp. 675-678. [45] L. Le Folgoc, A. V. Nori, S. Ancha, and A. Criminisi, "Lifted Auto-Context Forests for Brain Tumour Segmentation," in International Workshop on Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries, 2016, pp. 171-183. 25 

1cs.CV

A COPULA APPROACH FOR DEPENDENCE MODELING IN MULTIVARIATE NONPARAMETRIC TIME SERIES NATALIE NEUMEYER1 , MAREK OMELKA2 , ŠÁRKA HUDECOVÁ2 arXiv:1705.07605v1 [math.ST] 22 May 2017 1 Department of Mathematics, University of Hamburg, Bundesstrasse 55, 20146 Hamburg, Germany 2 Department of Probability and Statistics, Faculty of Mathematics and Physics, Charles University, Sokolovská 83, 186 75 Praha 8, Czech Republic May 23, 2017 Abstract. This paper is concerned with modeling the dependence structure of two (or more) time-series in the presence of a (possible multivariate) covariate which may include past values of the time series. We assume that the covariate influences only the conditional mean and the conditional variance of each of the time series but the distribution of the standardized innovations is not influenced by the covariate and is stable in time. The joint distribution of the time series is then determined by the conditional means, the conditional variances and the marginal distributions of the innovations, which we estimate nonparametrically, and the copula of the innovations, which represents the dependency structure. We consider a nonparametric as well as a semi-parametric estimator based on the estimated residuals. We show that under suitable assumptions these copula estimators are asymptotically equivalent to estimators that would be based on the unobserved innovations. The theoretical results are illustrated by simulations and a real data example. Keywords and phrases: Asymptotic representation; CHARN model; empirical copula process; goodness-of-fit testing; nonparametric AR-ARCH model; nonparametric SCOMDY model; weak convergence. 1. Introduction Modeling the dependency of k observed time series can be of utmost importance for applications, e. g. in risk management (for instance to model the dependence between several exchange rates). We will take the approach to model k dependent nonparametric AR-ARCH time series Yji = mj (X i ) + σj (X i ) εji , i = 1, . . . , n, j = 1, . . . , k, where the covariate X i may include past values of the process, Yj i−1 , Yj i−2 , . . . (j = 1, . . . , k), or other exogenous variables. Further the innovations (ε1i , . . . , εki ), i ∈ Z, are assumed to be independent and identically distributed random vectors and (ε1i , . . . , εki ) is independent ∗ The second and the third authors gratefully acknowledge support from the grant GACR 15-04774Y. 1 2 NATALIE NEUMEYER1 , MAREK OMELKA2 , ŠÁRKA HUDECOVÁ2 of the past and present covariates X ` , ` ≤ i, ∀i. For identifiability we assume E εji = 0, var(εji ) = 1 (j = 1, . . . , k), such that the functions mj and σj represent the conditional mean and volatility function in the jth time series. Such models are also called multivariate nonparametric CHARN (conditional heteroscedastic autoregressive nonlinear) models and have gained much attention over the last decades, see Fan and Yao (2005) and Gao (2007) for extensive overviews. Note that due to the structure of the model and Sklar’s theorem (see e.g., Nelsen, 2006), for zj = (yj − mj (x))/σj (x) (j = 1, . . . , k) one has P(Y1i ≤ y1 , . . . , Yki ≤ yk | X i = x) = P(ε1i ≤ z1 , . . . , εki ≤ zk ) = C(F1ε (z1 ), . . . , Fkε (zk )), where Fjε (j = 1, . . . , k) denote the marginal distributions of the innovations and C their copula. Thus the joint conditional distribution of the observations, given the covariate, is completely specified by the individual conditional mean and variance functions, the marginal distributions of the innovations, and their copula. The copula C describes the dependence structure of the k time series, conditional on the covariates, after removing influences of the conditional means and variances as well as marginal distributions. We will model the conditional mean and variance function nonparametrically like Härdle et al. (1998), among others. Semiparametric estimation, e. g. with additive structure for mj and multiplicative structure for σj2 as in Yang et al. (1999) can be considered as well and all presented results remain valid under appropriate changes for the estimators and assumptions. Further we will model the marginal distributions of the innovations nonparametrically, whereas we will take two different approaches to estimate the copula C: nonparametrically and parametrically. As the innovations are not observable, both estimators will be based on estimated residuals. We will show that the asymptotic distribution is not affected by the necessary pre-estimation of the mean and variance functions. This remarkable result is intrinsic for copula estimation, while the asymptotic distribution of empirical distribution functions is typically influenced by pre-estimation of mean and variance functions. Moreover, comparison of the nonparametric and parametric copula estimator gives us the possibility to test goodness-of-fit of a parametric class of copulas. Our approach extends the following parametric and semiparametric approaches in time series contexts. Chen and Fan (2006) introduced SCOMDY (semiparametric copula-based multivariate dynamic) models, which are very similar to the model considered here. However, the conditional mean and variance functions are modeled parametrically, while the marginal distributions of innovations are estimated nonparametrically and a parametric copula model is applied to model the dependence. See also Kim et al. (2007) for similar methods for some parametric time series models including nonlinear GARCH models, Rémillard et al. (2012), Kim et al. (2008) and the review by Patton (2012). Chan et al. (2009) give a goodness-of fit test for the innovation copula in the GARCH context. Further, in an i.i.d. setting Gijbels et al. A COPULA APPROACH FOR DEPENDENCE MODELING 3 (2015) show that in nonparametric location-scale models the asymptotic distribution of the empirical copula is not influenced by pre-estimation of the mean and variance function. This results was further generalized by Portier and Segers (2015) to a completely nonparametric model for the marginals. The remainder of the paper is organized as follows. In Section 2 we define the estimators and state some regularity assumptions. In Subsection 2.1 we show weak convergence of the copula process, while in Subsection 2.2 we show asymptotic normality of a parameter estimator when considering a parametric class of copulas. Subsection 2.3 is devoted to goodness-of-fit testing. In Section 3 we present simulation results and in Section 4 a real data example. All proofs are given in the Appendix. 2. Main results For the ease of presentation we will focus on the case of two time series, i. e. k = 2, but all results can be extended to general k ≥ 2 in an obvious manner. Suppose we have observed  for i = 1, . . . , n a section of the stationary stochastic process Y1i , Y2i , X i i∈Z that satisfies (1) Y1i = m1 (X i ) + σ1 (X i ) ε1i , Y2i = m2 (X i ) + σ2 (X i ) ε2i , i = 1, . . . , n,  where X i = (Xi1 , . . . , Xid )T is a d-dimensional covariate and the innovations (ε1i , ε2i ) i∈Z are independent identically distributed random vectors. Further (ε1i , ε2i ) is independent of the past and present covariates X k , k ≤ i, ∀i, and E ε1i = E ε2i = 0, var(ε1i ) = var(ε2i ) = 1. If the marginal distribution functions F1ε and F2ε of the innovations are continuous, then the copula function C of the innovations is unique and can be expressed as (2)
−1 −1 (u2 ) , C(u1 , u2 ) = Fε F1ε (u1 ), F2ε (u1 , u2 ) ∈ [0, 1]2 . As the innovations (ε1i , ε2i ) are unobserved, the inference about the copula function C is based on the estimated residuals (3) εbji = Yji − m b j (X i ) , σ bj (X i ) i = 1, . . . , n, j = 1, 2, where m b j and σ bj are the estimates of the unknown functions mj and σj . In what follows we will consider the local polynomial estimators of order p; see Fan and Gijbels (1996) or Masry (1996), among others. I.e. m b j (x) is for a given x = (x1 , . . . , xd )T defined as βb0 , the b with multi-index 0 = (0, . . . , 0), where β b is the solution to the minimization component of β problem (4) min β=(βi )i∈I n h X X
i 2 Yj` − βi ψi,hn X ` − x Khn (X ` − x). `=1 i∈I 4 NATALIE NEUMEYER1 , MAREK OMELKA2 , ŠÁRKA HUDECOVÁ2 Here I = I(d, p) denotes the set of multi-indices i = (i1 , . . . , id ) with i. = i1 + · · · + id ≤ p and
ik 1 Q k ψi,hn (x) = dk=1 x(k) ik ! . Further hn Khn (X ` − x) = d Y (k) k=1 with k being a kernel function and hn = Further σj2 (x) is estimated as 1 hn k  X`k −xk (k) hn (1) (d)
hn , . . . , hn  , the smoothing parameter. σ bj2 (x) = sbj (x) − m b 2j (x), where sbj (x) is obtained in the same way as m b j (x) but with Yj` replaced with Yj`2 . For any function f defined on J, interval in Rd , define for ` ∈ N, δ ∈ [0, 1), kf k`+δ = max sup |Di f (x)| + max sup i∈I(d,`) x,x0 ∈J i.=` x6=x0 i∈I(d,`) x∈J where Di = ∂ i. i , i ∂x11 ...∂xdd |Di f (x) − Di f (x0 )| , kx − x0 kδ `+δ (J) the set of and k · k is the Euclidean norm on Rd . Denote by CM e `+δ (J) the subset `-times differentiable functions f on J, such that kf k`+δ ≤ M . Denote by C 2 of C2`+δ (J) of the functions that satisfy inf x∈J f (x) ≥ 21 . In what follows we are going to prove that under appropriate regularity assumptions using the estimated residuals (3) instead of the (true) unobserved innovations εji affects neither the asymptotic distribution of the empirical copula estimator nor the parametric estimator of a copula. 2.1. Empirical copula estimation. Mimicking (2) the copula function C can be estimated nonparametrically as   en (u1 , u2 ) = Fbεb Fb−1 (u1 ), Fb−1 (u2 ) , (5) C 2b ε 1b ε where (6) n  1 X Fbεb(y1 , y2 ) = wni 1 εb1i ≤ y1 , εb2i ≤ y2 , Wn Wn = i=1 n X wnj j=1 is the estimate of the joint distribution function Fε (y1 , y2 ) and n  1 X b Fjbε (y) = wni 1 εbji ≤ y , Wn j = 1, 2, i=1 the corresponding marginal empirical cumulative distribution functions. Here we make use of a weight function wn (x) = 1{x ∈ Jn } and put wni = wn (X i ). For some real positive sequence cn → ∞ we set Jn = [−cn , cn ]d . (or) Now let Cn be the ‘oracle’ estimator based on the unobserved innovations, i.e.   −1 −1 (7) Cn(or) (u1 , u2 ) = Fbε Fb1ε (u1 ), Fb2ε (u2 ) , A COPULA APPROACH FOR DEPENDENCE MODELING 5  P where Fbε (z1 , z2 ) = n1 ni=1 1 ε1i ≤ z1 , ε2i ≤ z2 is the estimator of Fε (z1 , z2 ) based on the unobserved innovations and Fbjε (j = 1, 2) the corresponding marginal empirical cumulative distribution functions. Regularity assumptions. (β) The process (X i , Y1i , Y2i )i∈Z is strictly stationary and absolutely regular (β-mixing) with the mixing coefficient βi that satisfies βi = O(i−b ) with 1 + (s − 1)(1 + d) s−2 where s is specified in assumption (M) below. (1,1) (1,2) (2,2) (Fε ) The second-order partial derivatives Fε , Fε and Fε of the joint cumula(j,k) tive distribution function Fε (y1 , y2 ) = P(ε1 ≤ y1 , ε2 ≤ y2 ), with Fε (y1 , y2 ) = 2 ∂ Fε (y1 ,y2 ) ∂yj ∂yk , satisfy b> max sup j,k∈{1,2} y1 ,y2 ∈R2 Fε(j,k) (y1 , y2 )(1 + yj )(1 + yk ) < ∞. Further the innovation density fjε (j = 1, 2) satisfies



−1 −1 −1 −1 lim 1 + Fjε (u) fjε Fjε (u) = 0 and lim 1 + Fjε (u) fjε Fjε (u) = 0. u→0+ u→1− (FX ) The observations X i (i ∈ Z) have density fX that is bounded and differentiable with bounded uniformly continuous first order partial derivatives. We assume the
existence of some qf > 0 and some positive sequence cn = O (log n)1/d such that (1) (j) (j) d for αn  (log n)−qf and J+ n = ×j=1 [−cn − hn , cn + hn ], inf fX (x) ≥ αn(1) . x∈J+ n (M) For some s > 2, for j = 1, 2, E |εj0 |2s < ∞, the functions σj2s fX and |mj σj |s fX are bounded and there are some i∗ ∈ N, B > 0 such that for all i ≥ i∗ , sup σj2 (x0 )σj2 (xi )fX 0 ,X i (x0 , xi ) ≤ B, x0 ,xi sup mj (x0 )mj (xi ) σj (x0 )σj (xi )fX 0 ,X i (x0 , xi ) ≤ B, x0 ,xi where fX 0 ,X i denotes the joint density of (X 0 , X i ) and is bounded (for i ≥ i∗ ). p+γ (mσ) Let, for j = 1, 2 and for each n ∈ N, mj and σj be elements of CM (Jn ) for n q some γ ∈ (0, 1] and a sequence Mn = O((log n) m ) for some qm ≥ 0. Further, (2) E[σj4 (X 1 )] < ∞ and inf x∈Jn σj (x) ≥ αn for some positive sequence that fulfils (2) αn  (log n)−qσ for some qσ ≥ 0. (k) (Bw) There exists a sequence hn such that hhnn → ak , where ak ∈ (0, ∞), k = 1, . . . , d. Further, for γ from assumption (mσ) and for some δ > 0, nh2p+2γ (log n)D = o(1), n nh3d+2δ (log n)−D → ∞ n NATALIE NEUMEYER1 , MAREK OMELKA2 , ŠÁRKA HUDECOVÁ2 6 for all D > 0, and b−1−d− log n = o(1) for θ = θ d n hn b+3−d− 1+b s−1 1+b s−1 with b from assumption (β) and s from assumption (M). (k) k : R → R is a symmetric (d+2)-times continuously differentiable probability density function supported on [−1, 1].
Remark 1. Using Fε (y1 , y2 ) = C F1ε (y1 ), F2ε (y2 ) assumption (Fε ) requires that n

−1 −1 sup C (j,k) (u1 , u2 ) fjε Fjε (uj ) fkε Fkε (uk ) u1 ,u2 ∈[0,1]2 o  
−1 −1 −1 0 + C (j) (u1 , u2 ) fjε Fjε (uj ) 1{j = k} 1 + Fjε (uj ) 1 + Fkε (uk ) < ∞, 2 C(u1 ,u2 ) 1 ,u2 ) and C (j,k) (u1 , u2 ) = ∂ ∂u stand for the first and second where C (j) (u1 , u2 ) = ∂C(u ∂uj j ∂uk order partial derivatives of the copula function. Thus provided that for some δ > 0   C (j,k) (u1 , u2 ) = O u2 δ (1−u )2 δ1u2 δ (1−u )2 δ , j k j k
−1 −1 then we need that the functions fjε Fjε (u) (1 + Fjε (u) ) are of order O(uδ (1 − u)δ ). Remark 2. Parts of our assumptions are reproduced from Hansen (2008) because we apply his results about uniform rates of convergence for kernel estimators several times in our proofs. Note that in his Theorem 2 we set q = ∞. Theorem 1. Suppose that assumptions (β), (Fε ), (FX ), (Bw), (M), (k), (Jn ) and (mσ) are satisfied. Then  √  en (u1 , u2 ) − C (or) (u1 , u2 ) = oP (1). sup n C n (u1 ,u2 )∈[0,1]2 (or) Note that Theorem 1 together with the weak convergence of Cn (u1 , u2 ) (see e.g., Propo  e n = √n C en − C weakly converges in sition 3.1 of Segers, 2012) implies that that process C the space of bounded functions `∞ ([0, 1]2 ) to a centred Gaussian process GC , which can be written as GC (u1 , u2 ) = BC (u1 , u2 ) − C (1) (u1 , u2 ) BC (u1 , 1) − C (2) (u1 , u2 ) BC (1, u2 ) , where BC is a Brownian bridge on [0, 1]2 with covariance function   E BC (u1 , u2 )BC (u01 , u02 ) = C(u1 ∧ u01 , u2 ∧ u02 ) − C(u1 , u2 ) C(u01 , u02 ) . P Nevertheless in applications we recommend to use rather Wn = ni=1 wni instead of n as the sample size. A COPULA APPROACH FOR DEPENDENCE MODELING 7 2.2. Semiparametric copula estimation. The copula C describes the dependency between the two time series of interest, given the covariate. For applications modeling this dependency structure parametrically is advantageous because a parametric model often gives easier access to interpretations. Goodness-of-fit testing will be considered in the next section. Suppose that the joint distribution of (ε1i , ε2i ) is given by the copula function C(u1 , u2 ; θ), where θ = (θ1 , . . . , θp )T is an unknown parameter that belongs to a parametric space Θ ⊂ Rp . In copula settings we are often interested in semi-parametric estimation of the parameter θ, i.e. estimation of θ without making any parametric assumption on the marginal distributions F1ε and F2ε . The methods of semi-parametric estimation for i.i.d. settings are summarized in Tsukahara (2005). The question of interest is what happens if we use the estimated residuals (3) instead of the unobserved innovations εji . Generally speaking, thanks to Theorem 1 the answer is that using εbji instead of εji does not change the asymptotic distribution provided that the parameter of interest can be written as a Hadamard differentiable functional of a copula. 2.2.1. Method-of-Moments using rank correlation. This method is in a general way described for instance in Embrechts et al. (2005, Section 5.5.1). To illustrate the application of Theorem 1 for this method consider that the parameter θ is one-dimensional. Then the method of the inversion of Kendall’s tau is a very popular method of estimating the unknown parameter. For this method the estimator of θ is given by θbn(ik) = τ −1 (b τn ), where Z 1Z 1 C(u1 , u2 ; θ) dC(u1 , u2 ; θ) − 1 τ (θ) = 4 0 0 is the theoretical Kendall’s tau and τbn is an estimate of Kendall’s tau. In our settings the Kendall’s tau would be computed from the estimated residuals (b ε1i , εb2i ) for which wni > 0. By Theorem 1 and Hadamard differentiability of Kendall’s tau proved in Veraverbeke et al. (2011, Lemma 1), the estimators of Kendall’s tau based on εbji or on εji are asymptotically equivalent. Thus provided that τ 0 (θ) 6= 0 one gets that   n o
d √ 2 τ n θbn(ik) − θ −−−→ N 0, [τ 0σ(θ)] where στ2 = var 8 C(U11 , U21 ; θ) − 4 U11 − 4 U21 , 2 , n→∞ and (8)

U11 , U21 = F1ε (ε11 ), F2ε (ε21 ) . Analogously one can show that working with residuals has asymptotically negligible effects also for the method of moments introduced in Brahimi and Necir (2012). NATALIE NEUMEYER1 , MAREK OMELKA2 , ŠÁRKA HUDECOVÁ2 8 2.2.2. Minimum distance estimation. Here one can follow for instance Tsukahara (2005, Section 3.2). Note that thanks to Theorem 1 the proof of Theorem 3 of Tsukahara (2005) does not (or) en . Thus provided assumptions (B.1)-(B.5) of Tsukahara change when Cn is replaced with C 1 ,u2 ;θ) (2005) are satisfied with δ(u1 , u2 ; θ) = ∂C(u∂θ , then the estimator defined as ZZ
(md) en (u1 , u2 ) − C(u1 , u2 ; t) 2 du1 du2 θbn = arg min C [0,1]2 t∈Θ is asymptotically normal and satisfies    √  (md) d n θbn − θ −−−→ N 0p , Σ(md) , n→∞ where (md) Σ  ZZ = var [0,1]2 h γ(u1 , u2 ; θ) 1{U11 ≤ u1 , U21 ≤ u2 } − 2 X C (j) (u1 , u2 ; θ) 1{Uj1 ) i ≤ uj } du1 du2 , j=1 with #−1 " ZZ δ(v1 , v2 ; θ) δ T (v1 , v2 ; θ) dv1 dv2 γ(u1 , u2 ; θ) = δ(u1 , u2 ; θ). [0,1]2 2.2.3. M-estimator, rank approximate Z-estimators. To define a general M -estimator let us introduce 


e1i , U e2i = Wn Fb1bε εb1i , Fb2bε εb2i U (9) Wn +1 n that can be viewed as estimates of the unobserved (U1i , U2i ). Note that the multiplier WWn +1
e1i , U e2i bounded away is introduced in order to have both of the coordinates of the vector U from zero and one. The M -estimator of the parameter θ is now defined as θbn = arg min t∈Θ n X e1i , U e2i ; t wni ρ U
i=1 where ρ(u1 , u2 ; θ) is a given loss function. This class of estimators includes among others the (pl) pseudo-maximum likelihood estimators (θb ), for which ρ(u1 , u2 ; θ) = − log c(u1 , u2 ; θ), with n c(·) being the copula density function. Note that the estimator θbn is usually searched for as a solution to the estimating equations (10) n X
e1i , U e2i ; θbn = 0p , wni φ U i=1 where φ(u1 , u2 ; θ) = ∂ρ(u1 , u2 ; θ)/∂θ. In Tsukahara (2005) the estimator defined as the solution of (10) is called a rank approximate Z-estimator. A COPULA APPROACH FOR DEPENDENCE MODELING 9 In what follows we give general assumptions under which there exists a consistent root (θbn ) of the estimating equations (10) that is asymptotically equivalent to the consistent root (or) (θb ) of the ‘oracle’ estimating equations given by n n   X b1i , U b2i ; θb(or) = 0p , φ U n (11) i=1 where
b1i , U b2i = U (12) n n+1
Fb1ε (ε1i ), Fb2ε (ε2i ) are the standard pseudo-observations calculated from the unobserved innovations and their marginal empirical distribution functions Fbjε (y). Unfortunately, these general assumptions exclude some useful models (e.g. pseudo-maximum likelihood estimator in the Clayton family of copulas) for which the function φ(u1 , u2 ; θ) viewed as a function of (u1 , u2 ) is unbounded. The reason is that for empirical distribution functions calculated from estimated residuals εbji we lack some of the sophisticated results for an empirical distribution function calculated from (true) innovations εji . For such copula families one can use for instance the Method-of-Moments using rank correlation (see Section 2.2.1) to stay on the safe side. Nevertheless the simulation study in Section 3 suggests that the pseudo-maximum likelihood estimation can be used also for the Clayton copula (and probably also for other families of copulas with non-zero tail dependence) provided that the dependence is not very strong. Regularity assumptions. In what follows let θ stand for the true value of the parameter and V (θ) for an open neighbourhood of θ. (Id) θ is a unique minimizer of the function r(t) = E ρ(U1i , U2i ; t) and θ is an inner point of Θ. (φ) There exists V (θ) such that for each l1 , l2 ∈ {1, . . . , p} the functions φl1 (u1 , u2 ; t) = ∂φ(u1 ,u2 ;t) 1 ,u2 ;t) and φl1 ,l2 (u1 , u2 ; t) = ∂φ(u ∂tl1 ∂tl1 ∂tl2 are for each t ∈ V (θ) continuous in (u1 , u2 ) and of uniformly bounded Hardy-Kraus variation (see e.g., Berghaus et al., 2017). (j) (φ ) There exists V (θ) and a function h(u1 , u2 ) such that for each t ∈ V (θ) (j) (j) max max φl (u1 , u2 ; t) ≤ h(u1 , u2 ), where φl (u1 , u2 ; t) = j=1,2 l=1,...,p ∂φl (u1 ,u2 ;t) ∂uj and E h(U11 , U21 ) < ∞. 2 ;t) (Γ) Each element of the (matrix) function Γ(t) = E ∂φ(U∂t1 ,U is a continuous function T and the matrix Γ = Γ(θ) is positively definite. Theorem 2. Suppose that the assumptions of Theorem 1 are satisfied and that also (Id), (φ), (φ(j) ), and (Γ) hold. Then with probability going to one there exists a consistent root θbn 10 NATALIE NEUMEYER1 , MAREK OMELKA2 , ŠÁRKA HUDECOVÁ2 of the estimating equations (10), which satisfies 
√  d (13) n θbn − θ −−−→ Np 0p , Γ−1 Σ Γ−1 , n→∞ where ( ZZ
Σ = var φ U11 , U21 ; θ + ZZ + 1{U11 ≤ v1 } − v1  1{U21 ≤ v2 } − v2   ∂φ(v1 ,v2 ;θ) ∂v1 dC(v1 , v2 ; θ) )  ∂φ(v1 ,v2 ;θ) ∂v2 dC(v1 , v2 ; θ) . The proof of Theorem 2 is given in Appendix B. Note that the asymptotic distribution of the estimator θbn coincides with the distribution given in Section 4 of Genest et al. (1995) (or) that corresponds to the consistent root θb of the estimating equations (11). Thus using n the residuals instead of the true innovations has asymptotically negligible effect on the (first(or) order) asymptotic properties. In fact, it can be even shown that both θbn and θb have the n same asymptotic representations and thus √ (or)
n θbn − θbn = oP (1). 2.3. Goodness-of-fit testing. When modeling multivariate data using copulas parametrically one needs to choose a suitable family of copulas. When choosing the copula family tests of goodness-of-fit are often a useful tool. Thus we are interested in testing H0 : C ∈ C0 , where C0 = {Cθ , θ ∈ Θ} is a given parametric family of copulas. Many testing methods have been proposed (see e.g. Genest et al., 2009; Kojadinovic and Holmes, 2009, and the references therein). The most standard ones are based on the comparison of nonparametric and parametric estimators of a copula. For instance the Cramér-von Mises statistics is given by ZZ   en (u1 , u2 ) − C(u1 , u2 ; θbn ) 2 dC en (u1 , u2 ), (14) Sn = C where θbn is an estimate of the unknown parameter θ. As the asymptotic distributions of en (u1 , u2 ) and θbn are the same as the asymptotic distribution of C en(or) (u1 , u2 ) and θb(or) C n we suggest that the significance of the test statistic can be assessed in the same way as in i.i.d. settings. Thus one can use for instance the parametric bootstrap by simply generating independent and identically distributed observations from the copula function C(u1 , u2 ; θbn ). The test statistic is then simply recalculated from this observations in the same way as if we directly observed the innovations. The only difference is that instead of generating n observations we recommend to generate only Wn observations. A COPULA APPROACH FOR DEPENDENCE MODELING 11 Similar remarks hold when testing other hypotheses about the copula such as symmetry, for instance. Note that testing H0 : C(u1 , u2 ) ≡ u1 u2 provides a test for conditional independence of the two time series, given the covariate. 3. Simulation study A small Monte Carlo study was conducted in order to compare the semiparametric estimators based on the residuals with the ‘oracle’ estimators based on (unobserved) innovations. The inversion of Kendall’s tau (IK) method and the maximum pseudo-likelihood (MPL) method were considered for the following five copula families: Clayton, Frank, Gumbel, normal, and Student with 4 degrees of freedom. The values of the parameters are chosen so that they correspond to the Kendall’s tau τ = 0.25, 0.50 and 0.75. The data were simulated from the following four models: (Model 1) Y1i = 0.5(Xi + 1)2 + e1+0.1Xi ε1i , Y2i = 0.8(Xi − 1)2 + e1+0.2Xi ε2i , (Model 2) Y1i = 0.7Y1,i−1 + ε1i , Y2i = −0.5Y2,i−1 + ε2i , (Model 3) (Model 4) zY1i = 0.5 Y1,i−1 + ε1i , 2 1 + 0.1Y1,i−1 Y1i = σ1i ε1i , 2 2 σ1i = 1 + 0.3Y1,i−1 , Y2i = −0.4Y2,i−1 + ε2i , Y2i = σ2i ε2i , 2 2 σ2i = 5 + 0.2Y2,i−1 , where the innovations εji , j = 1, 2, follow marginally the standard normal distribution, and Xi is an exogenous variable following the ARMA model Xt = 0.5Xt−1 + ξt + 0.4ξt−1 with ξi being i.i.d. from a standard normal distribution. The simulations were conducted also for innovations εji , j = 1, 2 with Student marginals with 5 degrees of freedom but the results are similar. For brevity of the paper we do not present them here. The nonparametric estimates m b j and σ bj are constructed as local polynomial estimators of order p = 1 with K being the triweight kernel and the bandwidth hn chosen by the crossvalidation method for each estimate separately. The bandwidth is chosen from the interval (0, H), where H is the range of the corresponding explanatory variable. Finally, cn = log n. For each setting, we compute the estimate of the copula parameter θ from the true (but (ik,or) unobserved) innovations using the inversion of Kendall’s tau method (θbn ) and the max(pl,or) imum pseudo-likelihood method (θbn ). These oracle estimators are compared with their (ik) (pl) counterparts computed from the residuals (θbn ) and (θbn ). The performance of the estimators is measured by the bias, variance (VAR) and the mean square error (MSE), which are estimated from the 1 000 random samples. Finally, the sample sizes n = 200, 500 and 1000 are considered. The obtained results are listed for the five copulas separately in Tables 1,2,3,4 and (ik) (pl) 5. As θbn and θbn are natural competitors, the bigger of the two corresponding performance measures (bias, VAR, MSE) is stressed by the bold font. NATALIE NEUMEYER1 , MAREK OMELKA2 , ŠÁRKA HUDECOVÁ2 12 n = 200 Known innovations Model estim bias VAR MSE bias VAR MSE bias VAR 0.25 (ik,or) θbn (pl,or) θbn 0.008 0.019 0.019 0.008 0.008 0.009 -0.001 0.005 0.005 0.026 0.015 0.016 0.019 0.007 0.008 0.007 0.004 0.004 (ik,or) θbn (pl,or) θbn 0.022 0.066 0.067 0.019 0.027 0.028 0.003 0.015 0.016 0.027 0.054 0.054 0.017 0.024 0.024 0.002 0.014 0.014 (ik,or) θbn (pl,or) θbn 0.052 0.383 0.386 0.055 0.162 0.165 0.014 0.084 0.085 -0.065 0.287 0.291 -0.029 0.126 0.126 -0.036 0.073 0.075 (ik) θbn (pl) θbn 0.014 0.029 0.029 0.005 0.011 0.011 0.002 0.005 0.005 0.027 0.023 0.023 0.010 0.009 0.009 0.002 0.004 0.004 (ik) θbn (pl) θbn 0.007 0.104 0.104 -0.004 0.039 0.039 -0.006 0.019 0.019 -0.058 0.080 0.084 -0.057 0.031 0.034 -0.047 0.016 0.018 -0.158 0.569 0.594 -0.123 0.223 0.238 -0.098 0.107 0.117 -0.723 0.509 1.032 -0.566 0.237 0.557 -0.443 0.151 0.347 0.005 0.029 0.029 0.004 0.010 0.010 -0.002 0.005 0.005 0.25 0.50 0.50 0.75 0.75 0.25 0.50 0.50 0.75 0.75 0.25 0.25 2 0.50 0.50 0.75 0.75 0.25 0.25 3 0.50 0.50 0.75 0.75 0.25 0.25 4 n = 1000 τ 0.25 1 n = 500 0.50 0.50 0.75 0.75 (ik) θbn (pl) θbn (ik) θbn (pl) θbn 0.018 0.023 0.023 0.013 0.008 0.009 0.003 0.004 0.004 -0.034 0.096 0.097 -0.006 0.033 0.033 -0.010 0.017 0.017 -0.092 0.077 0.086 -0.038 0.028 0.030 -0.028 0.015 0.016 -0.485 0.522 0.757 -0.194 0.205 0.243 -0.120 0.099 0.113 -1.166 0.481 1.842 -0.592 0.209 0.559 -0.368 0.105 0.240 0.003 0.028 0.028 0.002 0.010 0.010 -0.004 0.005 0.005 0.016 0.023 0.023 0.011 0.008 0.009 0.002 0.004 0.004 0.095 0.096 -0.012 0.033 0.033 -0.016 0.017 0.017 -0.093 0.077 0.086 -0.047 0.028 0.030 -0.036 0.015 0.016 -0.488 0.495 0.733 -0.242 0.189 0.247 -0.168 0.092 0.120 -1.128 0.427 1.699 -0.648 0.177 0.597 -0.432 0.094 0.281 0.005 0.030 0.030 0.004 0.010 0.010 -0.003 0.005 0.005 (ik) θbn (pl) θbn (ik) θbn (pl) θbn (ik) θbn (pl) θbn -0.037 (ik) θbn (pl) θbn (ik) θbn (pl) θbn (ik) θbn (pl) θbn (ik) θbn (pl) θbn (ik) θbn (pl) θbn MSE 0.016 0.024 0.024 0.010 0.009 0.009 0.000 0.004 0.004 -0.018 0.099 0.099 -0.004 0.033 0.033 -0.011 0.016 0.017 -0.083 0.079 0.086 -0.050 0.028 0.030 -0.046 0.015 0.017 -0.233 0.542 0.596 -0.098 0.200 0.209 -0.088 0.091 0.099 -0.837 0.421 1.122 -0.517 0.185 0.452 -0.402 0.102 0.264 Table 1. Estimation for Clayton copula. In agreement with the results of Genest et al. (1995) and Tsukahara (2005) the results for the (oracle) estimates based on (unobserved) innovations are in favour of MPL method. This continues to hold also when working with estimated residuals provided that the dependence is not very strong (i.e. τ = 0.25 or τ = 0.50). But if the dependence is strong (i.e. τ = 0.75) then one should consider using the IK method. This seems to be true in particular for the Clayton copula and to some extent also to the Frank copula and the Gumbel copula. A closer inspection of the results reveals that while the variance of MPL method is almost always slightly smaller than the variance of the IK method, the bias can be substantially larger. On A COPULA APPROACH FOR DEPENDENCE MODELING n = 200 Known innovations Model estim bias VAR MSE bias VAR MSE bias VAR MSE 0.25 (ik,or) θbn (pl,or) θbn 0.026 0.142 0.143 -0.006 0.071 0.071 0.012 0.037 0.037 0.032 0.142 0.143 -0.002 0.071 0.071 0.013 0.037 0.037 (ik,or) θbn (pl,or) θbn 0.044 0.237 0.239 -0.001 0.117 0.117 0.018 0.062 0.063 0.044 0.234 0.236 0.000 0.116 0.116 0.017 0.061 0.062 (ik,or) θbn (pl,or) θbn 0.073 0.774 0.780 0.009 0.388 0.388 0.038 0.199 0.201 0.25 0.50 0.50 0.75 0.75 0.25 0.50 0.50 0.75 0.75 0.25 0.25 2 0.50 0.50 0.75 0.75 0.25 0.25 3 0.50 0.50 0.75 0.75 0.25 0.25 4 n = 1000 τ 0.25 1 n = 500 13 0.50 0.50 0.75 0.75 -0.031 0.758 0.759 -0.040 0.377 0.379 0.006 0.197 0.197 (ik) θbn (pl) θbn -0.017 0.218 0.218 -0.010 0.091 0.091 -0.009 0.043 0.043 -0.005 0.218 0.218 -0.005 0.091 0.091 -0.005 0.043 0.043 (ik) θbn (pl) θbn -0.059 0.358 0.362 -0.043 0.151 0.153 -0.024 0.072 0.073 -0.055 0.349 0.352 -0.040 0.149 0.150 -0.023 0.072 0.072 (ik) θbn (pl) θbn -0.424 1.225 1.405 -0.287 0.509 0.591 -0.168 0.244 0.272 -0.555 1.154 1.462 -0.336 0.499 0.611 -0.192 0.242 0.278 (ik) θbn (pl) θbn -0.018 0.221 0.222 -0.020 0.085 0.085 0.002 0.041 0.041 -0.010 0.222 0.222 -0.016 0.085 0.085 0.004 0.041 0.041 (ik) θbn (pl) θbn -0.093 0.366 0.375 -0.054 0.140 0.143 -0.014 0.070 0.070 -0.093 0.361 0.369 -0.055 0.138 0.141 -0.016 0.069 0.069 (ik) θbn (pl) θbn -0.899 1.390 2.198 -0.415 0.513 0.685 -0.195 0.252 0.290 -1.048 1.325 2.424 -0.481 0.501 0.732 -0.234 0.250 0.305 0.041 (ik) θbn (pl) θbn -0.021 0.222 0.222 -0.025 0.085 0.085 -0.001 0.041 -0.012 0.223 0.223 -0.020 0.085 0.085 0.000 0.041 0.041 (ik) θbn (pl) θbn -0.107 0.359 0.371 -0.070 0.139 0.143 -0.028 0.069 0.070 -0.108 0.354 0.366 -0.071 0.137 0.142 -0.030 0.068 0.069 -0.930 1.298 2.163 -0.511 0.492 0.753 -0.291 0.238 0.323 (ik) θbn (pl) θbn -1.076 1.219 2.376 -0.575 0.479 0.810 -0.331 0.236 0.345 (ik) θbn (pl) θbn -0.015 0.230 0.230 -0.022 0.084 0.085 0.000 0.041 0.041 -0.006 0.230 0.230 -0.017 0.084 0.085 0.002 0.041 0.041 (ik) θbn (pl) θbn -0.059 0.381 0.385 -0.053 0.142 0.144 -0.017 0.070 0.070 -0.058 0.375 0.379 -0.053 0.140 0.143 -0.018 0.069 0.069 -0.483 1.353 1.587 -0.259 0.511 0.578 -0.144 0.238 0.258 -0.632 1.267 1.667 -0.317 0.493 0.594 -0.176 0.233 0.264 (ik) θbn (pl) θbn Table 2. Estimation for Frank copula. the other hand the results suggest that for the normal and the Student copula one can stay with MPL method even in case of a strong dependence. Finally note that for large sample sizes the performance of the estimates based on residuals is usually almost as good as of the oracle estimates based on (unobserved) innovations. But there is still some price to pay even for the sample size n = 1000 and this price relatively increases with the level of dependence. The question for possible further research is how to explain the bad performance of PML method when based on residuals for the Clayton copula with a strong dependence. NATALIE NEUMEYER1 , MAREK OMELKA2 , ŠÁRKA HUDECOVÁ2 14 n = 200 Known innovations Model estim bias VAR MSE bias VAR MSE bias VAR MSE 0.25 (ik,or) θbn (pl,or) θbn 0.057 0.047 0.047 0.007 0.022 0.022 0.011 0.013 0.013 0.142 0.047 0.049 0.062 0.021 0.022 0.040 0.012 0.012 (ik,or) θbn (pl,or) θbn 0.103 0.153 0.154 0.034 0.069 0.069 0.024 0.040 0.040 0.207 0.148 0.152 0.109 0.066 0.067 0.057 0.037 0.037 (ik,or) θbn (pl,or) θbn 0.211 0.799 0.804 0.098 0.359 0.360 0.040 0.201 0.201 0.026 0.713 0.713 0.055 0.321 0.321 -0.034 0.173 0.174 0.25 0.50 0.50 0.75 0.75 0.25 0.50 0.50 0.75 0.75 0.25 0.25 2 0.50 0.50 0.75 0.75 0.25 0.25 3 0.50 0.50 0.75 0.75 0.25 0.25 4 n = 1000 tau 0.25 1 n = 500 0.50 0.50 0.75 0.75 (ik) θbn (pl) θbn 0.080 0.074 0.074 0.026 0.028 0.028 -0.023 0.014 0.014 0.173 0.067 0.070 0.062 0.026 0.027 0.003 0.013 0.013 (ik) θbn (pl) θbn 0.025 0.243 0.243 -0.006 0.089 0.089 -0.070 0.045 0.046 0.042 0.218 0.219 -0.038 0.082 0.082 -0.106 0.043 0.044 -0.595 1.227 1.262 -0.386 0.442 0.457 -0.394 0.235 0.250 -1.607 1.038 1.296 -1.228 0.411 0.562 -1.041 0.259 0.367 0.013 0.068 0.068 -0.018 0.027 0.027 -0.002 0.014 0.014 (ik) θbn (pl) θbn (ik) θbn (pl) θbn (ik) θbn (pl) θbn 0.099 0.066 0.067 0.034 0.026 0.026 0.026 0.013 0.013 -0.158 0.227 0.230 -0.095 0.083 0.084 -0.041 0.044 0.044 -0.152 0.211 0.213 -0.070 0.079 0.079 -0.037 0.041 0.041 (ik) θbn (pl) θbn -2.175 1.182 1.655 -0.986 0.451 0.548 -0.532 0.236 0.265 -3.347 1.019 2.139 -1.537 0.425 0.661 -0.918 0.216 0.300 (ik) θbn (pl) θbn 0.003 0.067 0.067 -0.023 0.027 0.027 -0.007 0.014 0.014 0.086 0.066 0.066 0.029 0.026 0.026 0.018 0.013 0.013 (ik) θbn (pl) θbn -0.173 0.227 0.230 -0.119 0.082 0.083 -0.064 0.044 0.044 -0.165 0.210 0.213 -0.095 0.078 0.079 -0.063 0.040 0.041 (ik) θbn (pl) θbn -2.118 1.150 1.598 -1.154 0.444 0.577 -0.733 0.223 0.277 -3.150 0.988 1.980 -1.667 0.406 0.684 -1.105 0.200 0.322 0.015 0.069 0.069 -0.017 0.027 0.027 -0.006 0.014 0.014 (ik) θbn (pl) θbn (ik) θbn (pl) θbn (ik) θbn (pl) θbn 0.093 0.068 0.068 0.030 0.026 0.026 0.012 0.013 0.013 -0.061 0.234 0.234 -0.069 0.085 0.085 -0.042 0.044 0.045 -0.081 0.210 0.211 -0.064 0.079 0.080 -0.073 0.040 0.041 -0.713 1.268 1.319 -0.429 0.453 0.471 -0.291 0.223 0.231 -1.736 1.087 1.389 -0.965 0.399 0.492 -0.740 0.198 0.253 Table 3. Estimation for Gumbel copula (10 multiples of BIAS, VAR and MSE). 4. Application To illustrate the proposed methods let us consider daily log returns of USD/CZK (US Dollar/Czech Koruna) and GBP/CZK (British Pound/Czech Koruna) exchange rates from 4th January 2010 to 31st December 2012. Note that we take only data till the end of 2012 (total of 758 observations for each series), because in November 2013 the Czech National Bank started its intervention aimed at CZK/EUR exchange rate, and this might violate the assumption of stationarity and independence for more recent data. A COPULA APPROACH FOR DEPENDENCE MODELING n = 200 Known innovations Model tau estim bias VAR 0.25 (ik,or) θbn (pl,or) θbn -0.051 0.710 (ik,or) θbn (pl,or) θbn -0.052 0.418 (ik,or) θbn (pl,or) θbn (ik) θbn (pl) θbn 0.25 0.50 0.50 0.75 0.75 0.25 0.25 1 0.50 0.50 0.75 0.75 0.25 0.25 2 0.50 0.50 0.75 0.75 0.25 0.25 3 0.50 0.50 0.75 0.75 0.25 0.25 4 0.50 0.50 0.75 0.75 n = 500 MSE bias VAR 0.290 0.290 -0.072 0.262 0.267 0.388 0.103 0.103 -0.052 0.087 0.089 0.242 -0.031 0.009 0.009 -0.049 0.008 0.008 -0.580 0.439 0.273 0.402 (ik) θbn (pl) θbn -0.644 -0.264 (ik) θbn (pl) θbn -0.406 (ik) θbn (pl) θbn (ik) θbn (pl) θbn (ik) θbn (pl) θbn 15 n = 1000 MSE bias VAR MSE 0.141 0.141 -0.041 0.071 0.071 0.127 0.129 0.224 0.061 0.062 0.050 0.050 -0.029 0.026 0.026 0.043 0.043 0.139 0.021 0.021 -0.022 0.005 0.005 -0.010 0.002 0.002 -0.021 0.004 0.004 -0.011 0.002 0.002 0.443 -0.130 0.171 0.172 0.026 0.087 0.087 0.403 0.183 0.155 0.156 0.144 0.077 0.077 0.163 0.167 -0.234 0.064 0.064 -0.084 0.031 0.031 0.142 0.142 -0.168 0.055 0.055 -0.116 0.027 0.027 0.016 0.018 -0.209 0.006 0.007 -0.115 0.003 0.003 -0.676 0.016 0.020 -0.483 0.007 0.009 -0.336 0.003 0.004 -0.476 0.452 0.454 -0.287 0.161 0.162 -0.179 0.079 0.079 0.530 0.413 0.416 0.200 0.149 0.150 0.098 0.068 0.068 -0.912 0.169 0.177 -0.402 0.060 0.062 -0.228 0.028 0.029 -0.349 0.142 0.143 -0.122 0.052 0.052 -0.074 0.023 0.023 -1.061 0.022 0.033 -0.441 0.007 0.009 -0.224 0.003 0.003 -1.246 0.021 0.036 -0.526 0.006 0.009 -0.267 0.002 0.003 (ik) θbn (pl) θbn -0.530 0.448 0.451 -0.343 0.162 0.163 -0.216 0.079 0.079 0.469 0.409 0.412 0.137 0.150 0.150 0.060 0.068 0.068 (ik) θbn (pl) θbn -0.967 0.168 0.177 -0.489 0.059 0.062 -0.291 0.028 0.029 -0.401 0.138 0.140 -0.219 0.051 0.052 -0.133 0.023 0.024 (ik) θbn (pl) θbn -1.013 0.020 0.030 -0.504 0.006 0.009 -0.290 0.003 0.004 -1.179 0.019 0.032 -0.580 0.006 0.009 -0.328 0.002 0.003 -0.425 0.478 0.480 -0.352 0.166 0.168 -0.200 0.080 0.080 (ik) θbn (pl) θbn (ik) θbn (pl) θbn (ik) θbn (pl) θbn 0.528 0.438 0.440 0.092 0.154 0.154 0.035 0.069 0.069 -0.648 0.180 0.185 -0.387 0.061 0.062 -0.238 0.029 0.029 -0.125 0.150 0.151 -0.174 0.053 0.054 -0.138 0.024 0.024 -0.407 0.019 0.021 -0.206 0.006 0.006 -0.131 0.003 0.003 -0.565 0.017 0.020 -0.301 0.005 0.006 -0.197 0.002 0.003 Table 4. Estimation for normal copula (100 multiples of BIAS, VAR and MSE). Daily foreign exchange rates have been successfully modelled using the nonparametric autoregression, e.g., in Härdle et al. (1998) and Yang et al. (1999). Here, we apply a simple model of two separate nonparametric autoregressions of order 1 and search for a feasible copula for the innovations. The conditional means and variances are modelled using local polynomials with degree p = 1. The smoothing parameters are chosen using the crossvalidation (CV) method as 0.038, 0.061, 0.025, 0.056 for mU SD , mGBP , σU SD , and σGBP respectively. The fitted conditional means and standard deviations are plotted together with NATALIE NEUMEYER1 , MAREK OMELKA2 , ŠÁRKA HUDECOVÁ2 16 n = 200 Known innovations Model tau estim bias VAR 0.25 (ik,or) θbn (pl,or) θbn -0.276 0.309 (ik,or) θbn (pl,or) θbn -0.184 0.185 (ik,or) θbn (pl,or) θbn (ik) θbn (pl) θbn -0.303 0.294 (ik) θbn (pl) θbn -0.491 -0.268 (ik) θbn (pl) θbn 0.25 0.50 0.50 0.75 0.75 0.25 0.25 1 0.50 0.50 0.75 0.75 0.25 0.25 2 0.50 0.50 0.75 0.75 0.25 0.25 3 0.50 0.50 0.75 0.75 0.25 0.25 4 0.50 0.50 0.75 0.75 n = 500 MSE bias VAR 0.350 0.351 -0.051 0.351 0.352 0.267 0.130 0.130 -0.032 0.119 0.120 0.143 -0.066 0.013 0.013 -0.098 0.011 n = 1000 MSE bias VAR MSE 0.179 0.179 -0.131 0.093 0.093 0.176 0.176 0.064 0.092 0.092 0.067 0.067 -0.084 0.035 0.035 0.061 0.061 0.033 0.032 0.032 -0.015 0.006 0.006 -0.028 0.003 0.003 0.011 -0.046 0.005 0.005 -0.041 0.003 0.003 0.513 0.514 -0.139 0.210 0.210 -0.239 0.113 0.114 0.514 0.514 0.084 0.207 0.208 -0.081 0.112 0.112 0.204 0.206 -0.256 0.079 0.080 -0.249 0.042 0.043 0.183 0.183 -0.231 0.072 0.073 -0.236 0.040 0.040 -0.396 0.022 0.024 -0.230 0.008 0.009 -0.178 0.004 0.005 -0.641 0.020 0.024 -0.435 0.007 0.009 -0.327 0.004 0.005 (ik) θbn (pl) θbn -0.995 0.520 0.529 -0.230 0.209 0.210 -0.254 0.103 0.104 -0.423 0.526 0.528 0.070 0.209 0.209 -0.094 0.102 0.102 (ik) θbn (pl) θbn -1.235 0.201 0.216 -0.381 0.081 0.082 -0.276 0.039 0.040 -1.067 0.189 0.201 -0.317 0.074 0.075 -0.232 0.036 0.037 -1.214 0.028 0.043 -0.449 0.009 0.011 -0.251 0.004 0.005 -1.594 0.027 0.052 -0.651 0.009 0.013 -0.356 0.004 0.005 0.104 (ik) θbn (pl) θbn (ik) θbn (pl) θbn -1.061 0.515 0.526 -0.305 0.210 0.210 -0.308 0.104 -0.469 0.521 0.524 -0.028 0.209 0.209 -0.163 0.102 0.102 (ik) θbn (pl) θbn -1.302 0.202 0.219 -0.466 0.081 0.083 -0.361 0.040 0.041 -1.113 0.190 0.203 -0.416 0.075 0.076 -0.331 0.037 0.038 (ik) θbn (pl) θbn -1.162 0.026 0.039 -0.508 0.009 0.012 -0.328 0.004 0.005 -1.475 0.024 0.046 -0.686 0.008 0.013 -0.434 0.004 0.005 (ik) θbn (pl) θbn -0.881 0.525 0.533 -0.175 0.215 0.216 -0.278 0.103 0.104 -0.243 0.530 0.531 0.071 0.215 0.215 -0.157 0.101 0.101 (ik) θbn (pl) θbn -0.902 0.202 0.210 -0.291 0.082 0.083 -0.277 0.039 0.040 -0.658 0.191 0.196 -0.254 0.077 0.078 -0.282 0.036 0.037 -0.503 0.023 0.025 -0.197 0.009 0.009 -0.153 0.004 0.004 -0.734 0.021 0.026 -0.359 0.008 0.009 -0.268 0.003 0.004 (ik) θbn (pl) θbn Table 5. Estimation for Student copula (100 multiples of BIAS, VAR and MSE). the data in Figure 1. It is visible that the conditional mean functions are rather flat and range around zero. We use the goodness-of-fit test proposed in Section 2.3 in order to decide which copula should be used for modelling the innovations from the two autoregressions. The significance of the test statistics is assessed with the help of the bootstrap test based on B = 999 bootstrap samples. We test Clayton, Frank, Gumbel, normal and Student copula with 4 degrees of freedom respectively and obtain p-values 0.001, 0.001, 0.002, 0.062, 0.289. Hence, we conclude 0.02 0.00 −0.01 0.00 GBP 0.01 0.01 0.02 0.03 data fitted cond. mean fitted st.dev. −0.02 −0.03 −0.02 −0.01 USD 17 0.03 A COPULA APPROACH FOR DEPENDENCE MODELING −0.03 data fitted cond. mean fitted st.dev. 0 200 400 600 0 200 Time 400 600 Time Figure 1. Fitted conditional mean and variance for the analyzed log returns. 1.0 that the Student copula seems to be the best choice for the innovations. The normal copula is also not rejected on the 5% level, but the corresponding p-value is rather borderline, so the Student copula seems to provide a better fit. The maximum likelihood method estimates 4.693 degrees of freedom and parameter ρ = 0.782. Figure 2 shows plot of pseudo-observations e1i , U e2i ) given by (9), together with contours of the fitted Student copula. (U ● ● ●●● ● ●● ●● ●● ● ●● ● ● ● ●● ● ● ● ●●●● ● ● ●● ● ●● ● ●● ●● ● ● ●● ● ●●● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ●● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ●● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ● ●● ●● ● ●● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●●● ●● ● ● ● ● ●● ● ● ● ● ● ● ●●● ● ● ● ● ●●●● ●● ● ● ● ● ● ●● ●● ● ●● ● ● ● ● ● ●●● ● ● ● ● ● ● ●● ● ● ● ● ●●●●● ● ● ●● ● ● ● ● ● ●● ●●● ● ● ● ● ● ● ● 0.0 0.2 0.4 0.6 0.8 ● 0.0 0.2 0.4 ● 0.6 ● ● ●● 0.8 1.0 e1i , U e2i ) given by (9) together with contours Figure 2. Pseudo-observations (U of the fitted Student copula (black curves). NATALIE NEUMEYER1 , MAREK OMELKA2 , ŠÁRKA HUDECOVÁ2 18 Appendix A - Proof of Theorem 1 P Recall the definition Wn = nj=1 wnj . Introduce n  1 X −1 −1 wni 1 εb1i ≤ F1ε (u1 ), εb2i ≤ F2ε (u2 ) Wn i=1
−1 −1 = Fbεb F1ε (u1 ), F2ε (u2 ) b n (u1 , u2 ) = G and note that
bn G b −1 (u1 ), G b −1 (u2 ) , en (u1 , u2 ) = G C 1n 2n b 1n and G b 2n denote the marginals of G b n . Further G b n is a distribution function on [0, 1]2 where G b 1n (0) = G b 2n (0) = 0. Thus one can make use of the with the marginals cdfs satisfying G −1 Hadamard differentiability of the ‘copula mapping’ Φ : G 7→ G(G−1 1 , G2 ) proved in Theorem 2.4 of Bücher and Volgushev (2013) provided we show that the process
√ b n (u1 , u2 ) = n G b n (u1 , u2 ) − C(u1 , u2 ) , (u1 , u2 ) ∈ [0, 1]2 , (A1) G converges in distribution in the space `∞ ([0, 1]2 ) to a process G with continuous trajectories such that G(u, 0) = G(0, u) = G(1, 1) = 0 for each u ∈ [0, 1]. b n . Denote A1: Decomposition and weak convergence of G G(or) n (u1 , u2 ) = n 1X  −1 −1 (u2 ) , (u1 ), ε2i ≤ F2ε 1 ε1i ≤ F1ε n i=1 n X e (or) (u1 , u2 ) = 1 G n Wn  −1 −1 (u2 ) . (u1 ), ε2i ≤ F2ε wni 1 ε1i ≤ F1ε i=1 b n as Now one can decompose the process G
√
√
√ bn = n G en + G e (or) + G(or) , bn − G e (or) + n G e (or) − G(or) + n G(or) − C = G (A2) G n n n n n n (or) (or) e n, G e n and Gn stand for the first, second and third term respectively on the where G right-hand side of the first equation in (A2). In Section A2 it will be shown that the first term on the right-hand side of (A2) satisfies uniformly in (u1 , u2 ) ∈ [0, 1]2 ,
n h −1 i (1) (u , u ) f X C F (u ) 1 2 1ε 1 F −1 (u ) 1ε e n (u1 , u2 ) = √ G ε1i + 1ε 2 1 (ε21i − 1) n i=1
n −1 i (2) C (u1 , u2 ) f2ε F2ε (u2 ) X h F −1 (u ) √ + ε2i + 2ε 2 2 (ε22i − 1) + oP (1) (A3) n i=1 where (in agreement with the last two conditions in (Fε )) for u1 ∈ {0, 1} the first term on the right-hand side of (A3) is defined as zero and analogously for u2 ∈ {0, 1}. A COPULA APPROACH FOR DEPENDENCE MODELING 19 In Section A3, we will show the asymptotic negligibility of the second term on the righthand side of (A2), i.e.
√ e (or) (u1 , u2 ) = e (or) (u1 , u2 ) − G(or) (u1 , u2 ) = oP (1). (A4) sup G sup n G n n n (u1 ,u2 )∈[0,1]2 (u1 ,u2 )∈[0,1]2 Now combining (A2) with (A3) and (A4) yields that uniformly in (u1 , u2 ) (A5) b n (u1 , u2 ) = An (u1 , u2 ) + Bn (u1 , u2 ) + oP (1), G where (A6) n  1 X √ An (u1 , u2 ) = 1{U1i ≤ u1 , U2i ≤ u2 } − C(u1 , u2 ) , n i=1 n 2
h 1 X X (j) −1 Bn (u1 , u2 ) = √ C (u1 , u2 ) fjε Fjε (uj ) εji + n −1 Fjε (uj ) 2 (εji 2 i − 1) . i=1 j=1 The asymptotic representation (A5) together with standard techniques yields the weak conb n. vergence of the process G Now thanks to Hadamard differentiability of the copula functional and Theorem 3.9.4 of van der Vaart and Wellner (1996),
  √  √  en (u1 , u2 ) − C(u1 , u2 ) = n G bn G b −1 (u2 ) − C(u1 , u2 ) b −1 (u1 ), G n C 2n 1n (A7) b n (u1 , u2 ) − C (1) (u1 , u2 )G b n (u1 , 1) − C (2) (u1 , u2 )G b n (1, u2 ) + oP (1). =G Note that for each (u1 , u2 ) ∈ [0, 1]2 (A8) Bn (u1 , u2 ) − C (1) (u1 , u2 )Bn (u1 , 1) − C (2) (u1 , u2 )Bn (1, u2 ) = 0. Further combining (A7) with (A5), (A6) and (A8) gives  √  en (u1 , u2 ) − C(u1 , u2 ) n C (A9) = An (u1 , u2 ) − C (1) (u1 , u2 )An (u1 , 1) − C (2) (u1 , u2 )An (1, u2 ) + oP (1). Now the right-hand side of (A9) coincides with the asymptotic representation of the ‘oracle’  √  (or) copula process n Cn − C , which implies the statement of Theorem 1. A2: Showing (A3). Let us introduce the process n 1 X Zn (f ) = √ f (X i , ε1i , ε2i ), n i=1 that is indexed be the following set of functions n   F = (x, y1 , y2 ) 7→ 1 x ∈ [−c, c]d 1 y1 ≤ z1 b1 (x) + a1 (x), y2 ≤ z2 b2 (x) + a2 (x) , o c ∈ R+ , z1 , z2 ∈ R, a1 , a2 ∈ G, b1 , b2 ∈ Ge , NATALIE NEUMEYER1 , MAREK OMELKA2 , ŠÁRKA HUDECOVÁ2 20 where, for δ from assumption (Bw) and some ν > d(1 + dδ ), 
(A10) G = a : Rd → R | a ∈ C1d+δ Rd , sup kxkν |a(x)| ≤ 1 , x (A11) Ge =  d b:R →R|b∈ e d+δ C 2 d R , sup kxkν |b(x) − 1| ≤ 1 .
x Denote the centred process as (A12) Z̄n (f ) = Zn (f ) − E Zn (f ), f ∈ F, and note that f may be formally identified by (c, z1 , z2 , a1 , b1 , a2 , b2 ). We will use the notation f =(c, ˆ z1 , z2 , a1 , b1 , a2 , b2 ). Consider the semi-norm given by Z 1 kf k2,β = β −1 (u) Q−1 f (u) du, 0 where  β −1 (u) = inf x > 0 : βbxc ≤ u}, 
Q−1 f (u) = inf x > 0 : P f (ε11 , ε21 , X 1 ) > x ≤ u . Then completely analogously as in the proof of Lemma 1 in Dette et al. (2009) one can show that the logarithm of the bracketing number N[ ] (, F, k · k2,β ) can be bounded by
    

d d e k · k2 (A13) log O(−2d )N[ ] 2 , G, k · k2 N[ ] 2 , G, = O log() + O −2 ν + d+δ , where the rate follows from our Lemma 2 in Appendix C. Further one bracket is sufficient for  ≥ 1. Thus by (A13) and because ν > d(1 + dδ ), Z ∞q log N[ ] (, F, k · k2,β ) d < ∞. 0 From Dedecker and Louhichi (2002), Section 4.3, it follows that the centred process Z̄n given by (A12) is asymptotically k.k2,β -equicontinuous. To apply this result in order to prove (A3) note that √ X n  n n b 1 (Xi )−m1 (Xi ) −1 i) e Gn (u1 , u2 ) = wni 1 ε1i ≤ F1ε (u1 ) σσb11 (X +m , (X ) σ1 (Xi ) i Wn i=1 o m b 2 (Xi )−m2 (Xi ) −1 i) ε2i ≤ F2ε (u2 ) σσb22 (X + (Xi ) σ2 (Xi )   −1 −1 −1 ε1i ≤ F1ε (u1 ), ε2i ≤ F2ε (u2 ) and introduce the process √ X n  n o n −1 −1 wni 1 ε1i ≤ F1ε (u1 )bb1 (Xi ) + b a1 (Xi ), ε2i ≤ F2ε (u2 )bb2 (Xi ) + b a2 (Xi ) Ǧn (u1 , u2 ) = Wn i=1 n o −1 −1 − 1 ε1i ≤ F1ε (u1 ), ε2i ≤ F2ε (u2 ) A COPULA APPROACH FOR DEPENDENCE MODELING 21 with b aj , bbj , j = 1, 2, from Lemma 1 and Remark 3 in Appendix C. Then one obtains by monotonicity arguments applying Lemma 1(i) that, on an event with probability converging to one


Wn e Zn fn` − Zn gn ) ≤ Gn (u1 , u2 ) − Ǧn (u1 , u2 ) ≤ Zn fnu − Zn gn ) n for some deterministic positive sequence γn = o(n−1/2 ). Here,
−1 −1 −1 −1 fn` = ˆ cn , F1ε (u1 ), F2ε (u2 ), b a1 − γn , bb1 − γn sign(F1ε (u1 )), b a2 − γn , bb2 − γn sign(F2ε (u2 )) ,
−1 −1 −1 −1 fnu = ˆ cn , F1ε (u1 ), F2ε (u2 ), b a1 + γn , bb1 + γn sign(F1ε (u1 )), b a2 + γn , bb2 + γn sign(F2ε (u2 )) ,
−1 −1 gn = ˆ cn , F1ε (u1 ), F2ε (u2 ), b a1 , bb1 , b a2 , bb2 . We only consider the upper bound, the lower one can be handled completely analogous. First

note that Zn fnu − Zn gn ) = Z̄n fnu − Z̄n gn ) + Rn , where with probability converging to one, √ (A14) |Rn | ≤ 2 n max sup |Fjε (uv + w) − Fjε (u(v + sγn ) + w + γn )| = o(1), j=1,2 u∈R,s∈{−1,1} v∈(1/2,1),w∈(−1,1) where the last equality follows by a Taylor expansion, assumption (Fε ) and γn = o(n−1/2 ). Now introducing the notation (j = 1, 2)   −1 −1 −1 (A15) Fjε (u, x, γ) = Fjε (u) bbj (x) + γ sign(Fjε (u)) + b aj (x) + γ one can show that for a sufficiently large M similarly as in the proof of Lemma 1 in Dette et al. (2009)  n o −1 −1 kfnu − gn k2,β ≤ M P 1 X 1 ∈ Jn , ε11 ≤ F1ε (u1 , X 1 , γn ), ε21 ≤ F2ε (u2 , X 1 , γn ) o  n −1 −1 (u2 , X 1 , 0) > 0 (u1 , X 1 , 0), ε21 ≤ F2ε − 1 X 1 ∈ Jn , ε11 ≤ F1ε   −1 −1 (u1 , X 1 , γn ) ≤ M P X 1 ∈ Jn , F1ε (u1 , X 1 , 0) ≤ ε11 ≤ F1ε   −1 −1 + M P X 1 ∈ Jn , F2ε (u2 , X 1 , 0) ≤ ε21 ≤ F2ε (u2 , X 1 , γn ) and this can be bounded by M n−1/2 times the bound on the right hand side of (A14) and thus converges to zero in probability uniformly in u1 , u2 . From k.k2,β -equicontinuity of Z̄n one obtains Z̄n (fnu ) − Z̄n (gn ) = oP (1) uniformly with respect to u1 , u2 and in combination with (A14) and analogous considerations
for the lower bound Zn fn` − Zn gn ) we obtain e n (u1 , u2 ) − Ǧn (u1 , u2 ) = oP (1). sup G u1 ,u2 NATALIE NEUMEYER1 , MAREK OMELKA2 , ŠÁRKA HUDECOVÁ2 22 The remaining proof of (A3) is divided into two parts. First we prove that sup Ǧn (u1 , u2 ) − E∗ [Ǧn (u1 , u2 )] = oP (1) (A16) u1 ,u2 and then we calculate E∗ [Ǧn (u1 , u2 )]. Here, with slight abuse of notation, E∗ denotes expectation, considering the functions b aj , bbj as deterministic. Showing (A16). Note that we have Ǧn (u1 , u2 ) − E∗ [Ǧn (u1 , u2 )] = Z̄n (fn ) − Z̄n (gn ), (A17) where
−1 −1 (u2 ), b a1 , bb1 , b a2 , bb2 , (u1 ), F2ε fn = cn , F1ε
−1 −1 (u2 ), 0, 1, 0, 1 , (u1 ), F2ε gn = cn , F1ε with 0 and 1 standing for functions that are constantly equal to zero and one respectively. Similarly to before one can show that for a sufficiently large M i  h
 −1 (u1 , X 1 , 0) − u1 1 X 1 ∈ Jn kfn − gn k2,β ≤ M E F1ε F1ε i h
 −1 (u2 , X 1 , 0) − u2 1 X 1 ∈ Jn + E F2ε F2ε using notation (A15). Now note that with Lemma 1 (iii) in Appendix C we obtain kfn − gn k2,β = oP (1) uniformly in u1 , u2 . Finally with the help of (A17) and the asymptotic k.k2,β -equicontinuity of the process Z̄n one can conclude (A16). Calculating E∗ [Ǧn (u1 , u2 )]. With the help of a second-order Taylor series expansion of the right-hand side one gets 

 −1 −1 −1 −1 E∗ [Ǧn (u1 , u2 )] = Fε F1ε (u1 , X, 0), F2ε (u2 , X, 0) − Fε F1ε (u1 ), F2ε (u2 ) = 2 
√ X −1 −1 n E∗ wn (X) Fε(j) F1ε (u1 ), F2ε (u2 ) YjX (uj ) j=1 √ (A18) + n 2 2 X 2 X 
−1 −1 E∗ wn (X) Fε(j,k) F1ε (u1X ), F2ε (u2X ) YjX (uj ) YkX (uk ) , j=1 k=1 where −1 −1 −1 Yjx (u) = Fjε (u, x, 0) − Fjε (u) = b aj (x) + Fjε (u)(bbj (x) − 1), j = 1, 2, A COPULA APPROACH FOR DEPENDENCE MODELING 23
−1 and the point ujx lies between the points Fjε Fjε (uj , x, 0) and uj . Now using Lemma 1 (iv) in Appendix C for j = 1, 2 √ 
−1 −1 n E∗ wn (X) Fε(j) F1ε (u1 ), F2ε (u2 ) YjX (uj )
h    i √ −1 −1 −1 ∗ b + = n Fε(j) F1ε (u1 ), F2ε (u2 ) E∗ b aj (X)1{X ∈ J+ } +F (u )E ( b (X) − 1) 1 {X ∈ J } j j n n jε n h i
1 X F −1 (uj ) √ εji + jε 2 (ε2ji − 1) + oP (1) n i=1
−1 n i Fjε (uj ) X h F −1 (uj ) εji + jε 2 (ε2ji − 1) + oP (1) −1 −1 F1ε (u1 ), F2ε (u2 ) = Fε(j) = C (j) (u1 , u2 ) fjε √ n i=1 uniformly in (u1 , u2 ). To conclude the proof of (A3) we need to show that ‘the second order terms’ in (A18) are asymptotically negligible. To show that note that by assumption (Fε ) and Lemma 1 (iii) there exists a finite constant M such that with probability going to one

−1 −1 (uk ) (uj ) 1 + Fkε 1 + Fjε


−1 −1 −1 −1 (uk ) (uj ) 1 + Fkε (u2x ) 1 + Fjε (u1x ), F2ε ≤ Fε(j,k) F1ε

−1 −1 (ukx ) (ujx ) 1 + Fkε 1 + Fjε

−1 −1 1 + Fjε (ujx ) 1 + Fkε (ukx )

−1 −1 (uk ) (uj ) 1 + Fkε 1 + Fjε

≤M −1 −1 1 + Fjε (ujx ) 1 + Fkε (ukx )

−1 −1 (uk ) (uj ) 1 + Fkε M 1 + Fjε ≤

1
1
1 1 −1 −1 (uk ) 1 + op (n− 4 ) + op (n− 4 ) (uj ) 1 + op (n− 4 ) + op (n− 4 ) 1 + Fkε 1 + Fjε −1 −1 (u2x ) (u1x ), F2ε Fε(j,k) F1ε
≤ 2M uniformly in (u1 , u2 ) ∈ [0, 1]2 and x ∈ Jn . Thus to prove 
−1 −1 EX wn (X) Fε(j,k) F1ε (u1X ), F2ε (u2X ) YjX (uj ) YkX (uk ) = oP (1), it is sufficient to use once more Lemma 1 (iii). A3: Showing (A4). Note that with the notation Wn = Pn i=1 wni one can decompose
√ (or) e (or) n G n (u1 , u2 ) − Gn (u1 , u2 )
√
√ (or) e (or) = n G n (u1 , u2 ) − C(u1 , u2 ) − n Gn (u1 , u2 ) − C(u1 , u2 ) NATALIE NEUMEYER1 , MAREK OMELKA2 , ŠÁRKA HUDECOVÁ2 24 (A19) =
n −1 X    −1 −1 √ wni 1 ε1i ≤ F1ε (u1 ), ε2i ≤ F2ε (u2 ) − C(u1 , u2 ) n n Wn i=1 n    1 X −1 −1 +√ (wni − 1) 1 ε1i ≤ F1ε (u1 ), ε2i ≤ F2ε (u2 ) − C(u1 , u2 ) n i=1
1 = Wn − 1 Bn1 (u1 , u2 ) + Bn2 (u1 , u2 ), where Bn1 (u1 , u2 ) stands for the first term on the right-hand side of the equation (A19) (except for the factor Wnn − 1) and Bn2 (u1 , u2 ) for the second term. Further using standard techniques one can show that both Bn1 (u1 , u2 ) and Bn2 (u1 , u2 ) viewed as processes on [0, 1]2 are asymptotically equi-continuous. Moreover as
n − 1 = o (1), and E w (X ) − 1 = −P(X 1 6∈ Jn ) = o(1), n 1 P Wn
one can conclude that both processes W1n − 1 Bn1 (u1 , u2 ) and Bn2 (u1 , u2 ) are uniformly asymptotically negligible in probability, which together with (A19) implies (A4). Appendix B - Proof of Theorem 2 Thanks to assumption (φ), the estimator θbn is a solution to the estimating equations (10). In what follows, first we prove the existence of a consistent root of the estimating equations (10) and then we derive that this root satisfies (B1) √ n

1 X b b n θbn − θ = Γ−1 √ φ U1i , U2i ; θ + oP (1), n i=1 b1i , U b2i ) are introduced in (12). The statement of the theorem now follows for p = 1 where (U by Proposition A 1(ii) of Genest et al. (1995) and for p > 1 by Theorem 1 of Gijbels et al. (2017).  en0 (u1 , u2 ) = 1 Pn wni 1 U e1i ≤ u1 , U e2i ≤ u2 , where pseudoProving consistency. Put C i=1 Wn e1i , U e2i ) are defined in (9). Note that observations (U (B2) en (u1 , u2 ) − C e 0 (u1 , u2 ) = OP C n sup (u1 ,u2 )∈[0,1]2 1 Wn
= OP 1 n
. Fix l ∈ {1, . . . , p}. By Corollary A.7 of Berghaus et al. (2017) one gets Z 1Z 1 n
1 X e e en0 (v1 , v2 ) wni φl U1i , U2i ; t = φl (v1 , v2 ; t) dC Wn 0 0 i=1 Z 1Z 1
en0 (v1 , v2 ) dφl (v1 , v2 ; t) + φl (1, 1; t = C 0 Z (B3) − 0 0 1 en0 (v1 , 1) dφl (v1 , 1; t) − C Z 0 1 e 0 (1, v2 ) dφl (1, v2 ; t). C n A COPULA APPROACH FOR DEPENDENCE MODELING 25 Note that thanks to assumption (φ(j) ) (uniformly in t ∈ V (θ)) Z 1 Z 1 n 1 X 0 e 1{Ue1i ≤ v1 } dφl (v1 , 1; t) Cn (v1 , 1) dφl (v1 , 1; t) = wni Wn 0 0 i=1 1 = Wn Wn X i=1 Z 1 i Wn +1 Z 1 = φl (1, 1; t) − (B4) Wn 1 X φl ( Wni+1 , 1; t) dφl (v1 , 1; t) = φl (1, 1; t) − Wn i=1 φl (v1 , 1; t) dv1 + OP 0 and analogously also Z 1 Z en0 (1, v2 ) dφl (1, v2 ; t) = φl (1, 1; t) − (B5) C 0 1 Wn
1 φl (1, v2 ; t) dv2 + OP 0 1 Wn
. Now combining (B3), (B4) and (B5) yields Z 1Z 1 n
1 X e e e 0 (v1 , v2 ) dφl (v1 , v2 ; t) + Al (t) + OP (B6) wni φl U1i , U2i ; t = C n Wn 0 0 1 n
, i=1 where (B7)
Al (t) = −φl (1, 1; t + Z 1 Z φl (v1 , 1; t) dv1 + 0 1 φl (1, v2 ; t) dv2 . 0 Analogously one gets (B8)
E φl U11 , U21 ; t = Z 1Z 1 C(v1 , v2 ) dφl (v1 , v2 ; t) + Al (t). 0 0 Now using (B2), (B6), (B8) and assumption (φ) gives that uniformly in t ∈ V (θ) n
1 X e1i , U e2i ; t − E φl (U11 , U21 ; t) wni φl U Wn i=1 Z 1Z 1  0  e (v1 , v2 ) − C(v1 , v2 ) dφl (v1 , v2 ; t) + OP = C n 0 0 1 n
= oP (1), where we have used Theorem 1 and assumption (φ). The existence of a consistent root of estimating equations (10) now follows by assumptions (Id) and (Γ). Analogously one can show the existence of a consistent root of estimating equations (11). Showing (B1). Let θbn be a consistent root of the estimating equations (10). Then by the mean value theorem applied to each coordinate of the vector-valued function n
1 X e1i , U e2i ; t wni φ U Ψn (t) = Wn i=1 one gets n   1 X e1i , U e2i ; θbn 0p = wni φ U Wn i=1 NATALIE NEUMEYER1 , MAREK OMELKA2 , ŠÁRKA HUDECOVÁ2 26 n n    
1 X 1 X ∗ e e e e = wni φ U1i , U2i ; θ + wni Dφ U1i , U2i ; θ n θbn − θ , Wn Wn i=1 i=1 1 ,u2 ;t) and θ ∗n is between θbn and θ. Note that as the mean value where Dφ stands for ∂φ(u∂t ∗,p theorem is applied to a vector valued function there are in fact p different points θ ∗,1 n , . . . , θn for each coordinate of the function Ψn (t) but all of them are consistent so for simplicity of notation we do not distinguish them. Thus to finish the proof of (B1) it is sufficient to show that n
1 X e1i , U e2i ; θ ∗ = Γ + oP (1) wni Dφ U n Wn (B9) i=1 and √ X n n X

n e1i , U e2i ; θ = √1 b1i , U b2i ; θ + oP (1). wni φ U φ U Wn n (B10) i=1 i=1 When proving (B9) one can mimic the proof of consistency of θbn and show that there exists V (θ) (an open neighbourhood of θ such that) sup t∈V (θ) n
1X e1i , U e2i ; t − E Dφ (U11 , U21 ; t) = oP (1). wni Dφ U n i=1 Using the consistency of θbn and assumption (Γ) yields (B9). 0(or) Thus one can concentrate on proving (B10). Put Cn b1i , U b2i ) are defined in (12). Note that u2 , where (U (B11) sup (u1 ,u2 )∈[0,1]2 (u1 , u2 ) = 1 n Cn(or) (u1 , u2 ) − Cn0(or) (u1 , u2 ) = OP Pn i=1 1 n
1 Ub1i ≤ u1 , Ub2i ≤  . Analogously as (B6) one can also show that for l = 1, . . . , p Z 1Z 1 n
1X b b φl U1i , U2i ; θ = φl (v1 , v2 ; θ) dCn0(or) (v1 , v2 ) n 0 0 i=1 Z 1Z 1 (B12) = Cn0(or) (v1 , v2 ) dφl (v1 , v2 ; θ) + Al (θ) + OP 0 0 where Al (θ) is given in (B7). Now using (B2), (B6), (B11), (B12), Theorem 1 and (φ) one gets √ X n n

n 1 X e e b1i , U b2i ; θ wni φl U1i , U2i ; θ − √ φl U Wn n i=1 i=1 Z 1Z 1  √  0 e (u1 , u2 ) − C 0(or) (u1 , u2 ) dφl (u1 , u2 ; θ) + oP = n C n n 0 0 which verifies (B10) and finishes the proof of (B1). √1 n
1 n
, = oP (1), A COPULA APPROACH FOR DEPENDENCE MODELING 27 Appendix C - Auxiliary results Lemma 1. Assume that (β), (Fε ), (M), (FX ), (Bw), (k), (Jn ) and (mσ) are satisfied. Then there exist random functions b aj and bbj on Jn such that for j = 1, 2 (i) sup x∈Jn

m b j (x) − mj (x) σ bj (x) b −b aj (x) = oP n−1/2 , sup − bj (x) = oP n−1/2 , σj (x) x∈Jn σj (x) (ii) kb aj kd+δ = oP (1), aj (x) = oP (iii) sup b kbbj − 1kd+δ = oP (1) for δ > 0 from assumption (Bw),

n−1/4 , sup bbj (x) − 1 = oP n−1/4 , x∈Jn x∈Jn Z (iv) b aj (x)fX (x) dx = Jn 1 n n X
εji + oP n−1/2 , i=1 n   X

bbj (x) − 1 fX (x) dx = 1 ε2ji − 1 + oP n−1/2 . 2n Jn Z i=1
Proof. For ease of presentation we set j = 1 and assume hn = hn , . . . , hn . We will first prove the assertions for m b 1 . The proof basically goes along the lines of the proof of Lemma 1 by Müller et al. (2009), but changes are necessary due to the dependency of observations in our model and because our covariate density is not assumed to be bounded away from zero on its support. Recall that I(d, p) denotes the set of multi-indices i = (i1 , . . . , id ) with i. = i1 + · · · + id ≤ p and we set I = I(d, p), where p is the order of the polynomials used in the local polynomial estimation. b denote the solution of the minimizaProof of assertion (i) for m b 1 . Fix some x ∈ Jn and let β b satisfies the normal equations tion problem (4). Then β X b = 0 ∀i ∈ I, b ik (x)β Ai (x) + Bi (x) − Q k k∈I where n Ai (x) = 1X σ1 (X ` ) ε1` ψi,hn (X ` − x) Khn (X ` − x), n Bi (x) = n 1X m1 (X ` )ψi,hn (X ` − x)Khn (X ` − x), n `=1 b ik (x) = Q 1 n `=1 n X ψi,hn (X ` − x) ψk,hn (X ` − x) Khn (X ` − x). `=1
1/2 From Theorem 2 in Hansen (2008) we obtain for %n = log n/(nhdn ) , (C1) b ik (x) − Qik (x) = OP (%n ), sup Q x∈Jn NATALIE NEUMEYER1 , MAREK OMELKA2 , ŠÁRKA HUDECOVÁ2 28   b ik (x) , i, k ∈ I. Note that where we define Qik (x) = E Q Z Qik (x) = ψi,(1,..1) (u) ψk,(1,..1) (u) fX (x + hn u) K(u) du, and for x ∈ Jn , consider the matrices Q(x) with entries Qik (x), i, k ∈ I. Analogously put b b ik (x), i, k ∈ I. Q(x) for the matrix with entries Q It follows from assumption (FX ) that 0 < λn ≤ aT Q(x) a ≤ Λ < ∞ for all vectors a of unit Euclidean length, where λn is a sequence of positive real numbers of the same rate as (1) αn in assumption (FX ). Thus Q(x) has eigenvalues in the interval [λn , Λ], and on the event   λn b En = sup Q(x) − Q(x) ≤ 2 x∈Jn b b one has aT Q(x) a ≥ λn /2 for all a of unit Euclidean length, such that the matrix Q(x) is invertible as well. Here and throughout kQk denotes the spectral norm of a matrix Q. Note (1)
that P(En ) → 1 by (C1) and %n = o αn , which holds under assumption (Bw). For the remainder of the proof we assume that the event En takes place because its complement does not matter for the assertions of the lemma. It follows from the normal equations that for x ∈ Jn ,
b −1 (x) A(x) + B(x) , m b 1 (x) = eT Q 1 where e1 = (1, 0, . . . , 0)T and A(x) and B(x) denote the vectors with components Ai (x) and Bi (x), i ∈ I, respectively. Now define −1 b a1 (x) = eT 1 Q (x)A(x) (C2) 1 , σ1 (x) x ∈ Jn , then we have the decomposition m b 1 (x) − m1 (x) −b a1 (x) = r1 (x) + r2 (x) σ1 (x) (C3) with remainder terms 1 , σ1 (x)
1 b −1 b r2 (x) = eT , 1 Q (x) B(x) − Q(x)β(x) σ1 (x)
−1 b −1 r1 (x) = eT 1 Q (x) − Q (x) A(x) where β(x) is the vector with components β i (x) = hi.n Di m1 (x), i ∈ I. From Theorem 2 in Hansen (2008) we obtain (C4) sup Ai (x) = OP (%n ), i ∈ I. x∈Jn For the treatment of the inverse matrices in r1 (x) we use Cramer’s rule and write b −1 (x) − Q−1 (x) = Q
T
1 1 b
C(x) T −
C(x) b det Q(x) det Q(x) A COPULA APPROACH FOR DEPENDENCE MODELING 29

b
T
T det Q(x) − det Q(x) 1 b b
C(x) − C(x) , +

C(x) b det Q(x) det Q(x) det Q(x) = b b where C(x) and C(x) denote the cofactor matrices of Q(x) and Q(x), respectively. Due to b the boundedness of the functions Qik each element of C(x) − C(x) can be absolutely bounded b by OP (%n ) by (C1) and the same rate is obtained for det(Q(x)) − det(Q(x)) , uniformly in |I| x. Using the lower bound λn for the determinant of Q(x), and assumption (mσ) to bound 1/σ1 gives the rate  
(%n )2 (C5) sup r1 (x) = OP = oP n−1/2 (1) 2|I| (2) (αn ) x∈Jn αn by assumption (Bw). In order to show negligibility of r2 (x) first note that the spectral b −1 (x) is given by the reciprocal of the square root of the smallest eigenvalue of norm of Q T Q(x). b b Q(x) With
1/2 λn Tb b b b ≥ (aT Q(x)T Q(x)a)1/2 − Q(x) − Q(x) ≥ aT Q(x) Q(x)a = Q(x)a 2 b −1 (on En ) for all a with ||a|| = 1, we obtain the rate O(λ−1 n ) for Q (x) . Further, by Taylor expansion of m1 (X ` ) in the definition of Bi (x) and using assumption (mσ) we have n 1X ≤ Mn kX ` − xkp+γ kψi,hn (X ` − x)k Khn (X ` − x) n `=1
= O Mn hp+γ fbX (x), n
b Bi (x) − Q(x)β(x) i P where the kernel density estimator fbX (x) = n1 n`=1 Khn (X ` − x) converges to fX (x) uniformly in x ∈ Jn , see Theorem 6 by Hansen (2008). Altogether we have  
p+γ n hn = oP n−1/2 (C6) sup r2 (x) = OP M(1) (2) αn αn x∈Jn using assumption (Bw). Now assertion (i) for m b 1 follows from (C2), (C3), (C5) and (C6). Proof of assertion (ii) for b a1 . kb a1 kd+δ = max sup Di b a1 (x) i∈I(d,d) x∈Jn + max max i∈I(d,d) i.=d    sup x,x0 ∈Jn kx−x0 k≤hn |Di b a1 (x) kx − Dib a1 (x0 )| , − x0 kδ sup x,x0 ∈Jn kx−x0 k>hn ≤ max sup Di b a1 (x) i∈I(d,d) x∈Jn + max i∈I(d,d+1) i.=d+1 sup Di b a1 (x) h1−δ + 2 max sup Di b a1 (x) h−δ n n x∈Jn i∈I(d,d) i.=d x∈Jn |Di b a1 (x) kx  − Dib a1 (x0 )|   − x0 kδ NATALIE NEUMEYER1 , MAREK OMELKA2 , ŠÁRKA HUDECOVÁ2 30 by the mean value theorem. Again by Theorem 2 of Hansen (2008) we have sup hi.n Di Ak (x) = OP (%n ), i ∈ I(d, d + 1). x∈Jn Further note that   ∂ ∂ −1 −1 Q (x) = Q (x) Q(x) Q−1 (x) ∂xk ∂xk (1)
and that the spectral norm of Q−1 (x) can be bounded by O 1/αn before. We apply the product rule for derivatives to obtain kb a1 kd+δ ≤ max j ` X X `=1,...,d +2 OP  j=0 k=0 j d X X OP  j=0 k=0 = OP  %n h−d n (1) (2) αn αn = OP  %n (1) hd+δ α n n αn  `−j %n h−k n Mn (2) `−j+1 (1) j−k+1 (αn ) (αn ) d−j %n h−k n Mn (1) j−k+1 (2) d−j+1 (αn ) (αn )   + j d+1 X X OP  j=0 k=0 with considerations as d+1−j %n h−k n Mn (2) (1) j−k+1 (αn )d−j+2 (αn )  hn1−δ h−δ n  −d−1    hn %n h−d 1−δ −δ n + OP %n(1) h + O P (2) (1) (2) hn n αn αn αn αn  = oP (1) (2) by assumption (Bw). Assertion (ii) for b a1 follows. Proof of assertion (iii) for b a1 . From the definition of b a1 and (C4) we obtain that  
a1 (x) = OP (1)%n (2) = oP n−1/4 sup b αn αn x∈Jn and thus (iii) follows for b a1 . Proof of assertion (iv) for b a1 . To prove (iv) note that Z n 1X ε1i ∆n (X i ) b a1 (x)fX (x) dx = n Jn i=1 with Z 1 eT Q−1 (x) ψ hn (X i − x) Khn (X i − x) fX (x) dx σ1 (x) 1 Z fX (X i − uhn ) = σ1 (X i )  eT Q−1 (X i − uhn ) ψ(u) K(u) du.  1 X i −cn X i +cn σ1 (X i − uhn ) , ∆n (X i ) = σ1 (X i ) hn hn From the support properties of the kernel function it follows that ∆n (X i )1{X i 6∈ J+ n } = 0. − d Further, for Jn = [−cn + hn , cn − hn ] note that n (C7)
1X − −1/2 ε1i ∆n (X i ) 1{X i ∈ J+ n \ Jn } = oP n n i=1 because the expectation is zero and the variance is bounded by    1 1 1 − 2 sup sup fX (x) E 1 X 1 ∈ J+ n \ Jn σ1 (X 1 ) n x∈Jn σ1 (x) λn x∈Rd A COPULA APPROACH FOR DEPENDENCE MODELING ≤ 31
1  Mn hn 
1  1  \ J− = O (1) (2) = o n−1 . O (1) (2) Mn P X 1 ∈ J+ n n αn αn αn αn n n It remains to consider n 1X ε1i ∆n (X i )1{X i ∈ J− n }, n i=1 (1) (2) with ∆n (X i ) = ∆n (X i ) + ∆n (X i ), where Z −1 (1) eT ∆n (X i ) = 1 Q (X i − uhn ) ψ(u) K(u) fX (X i − uhn ) du, [−1,1]d Z ∆(2) n (X i ) = [−1,1]d   σ1 (X i ) −1 eT 1 Q (X i − uhn ) ψ(u) K(u) fX (X i − uhn ) σ1 (X i −uhn ) − 1 du. (2) Now, by applying the mean value theorem for σ1 , for X i ∈ J− n , ∆n (X i ) can be bounded (1) (2)
absolutely by O Mn hn /(αn αn ) = o(1). Thus analogously as when showing (C7) one can use Markov’s inequality to get Z n n

1X 1X − −1/2 b a1 (x)fX (x) dx − ε1i = ε1i ∆(1) . n (X i ) − 1 1{X i ∈ Jn } + oP n n n Jn i=1 i=1  (1)
2  To obtain the desired negligibility it remains to show E ∆n (X i ) − 1 1{X i ∈ J− n } → 0. To this end we write Z −1 1= eT 1 Q∗ (x − hn u)ψ(u)K(u)fX (x − hn u) du, [−1,1]d where for x ∈ J+ n the matrix Q∗ (x) has entries Z Q∗ik (x) = fX (x) ψi (u)ψk (u)K(u) du, i, k ∈ I. Note that Q∗ (x) has the smallest eigenvalue of order λn . Thus we obtain the bound 
2  − E ∆(1) n (X i ) − 1 1{X i ∈ Jn }  Z
−1 −1 eT =E 1 Q (X i − uhn ) − Q∗ (X i − uhn ) [−1,1]d ψ(u)K(u)fX (X i − uhn ) du Z ≤ O(1) Z J− n [−1,1]d 2 Q−1 (x − uhn ) − Q−1 ∗ (x − uhn ) 1{X i ∈ 2  J− n} du fX (x) dx 2 ≤ O(1) sup kQ−1 (x)k2 sup kQ−1 ∗ (x)k x∈J+ n x∈J+ n Z J− n Z [−1,1]d kQ(x − uhn ) − Q∗ (x − uhn )k2 du fX (x) dx. NATALIE NEUMEYER1 , MAREK OMELKA2 , ŠÁRKA HUDECOVÁ2 32 Now with bounds for the matrix norms similar to before, and inserting the definitions of Q and Q∗ we obtain   2 − E (∆(1) n (X i ) − 1) 1{X i ∈ Jn } Z  Z Z 1 |fX (x − uhn + hn v) − fX (x − uhn )|2 K(v) dv du fX (x) dx ≤O (1) 4 (αn ) =O  J− n  h2n (1) (αn )4 [−1,1]d [−1,1]d = o(1) by the mean value theorem and assumptions (FX ) and (k). Proof of assertions (i)–(iv) for σ b1 . Recall the definition σ b12 = sb1 − m b 21 , where sb1 is the local polynomial estimator based on (X i , Y1i2 ), i = 1, . . . , n. With the notation s1 (x) = E[Y1i2 | X i = x] = σ12 (x) + m21 (x) we obtain σ b1 (x) σ1 (x) = 1+ sb1 (x) − s1 (x) m b 1 (x) − m1 (x) m1 (x) − + r(x), σ1 (x) σ1 (x) 2σ12 (x) where r(x) = − 1 (m (b σ12 (x) − σ12 (x))2 b 1 (x) − m1 (x))2 − . 2 σ12 (x) 2σ12 (x)(b σ1 (x) + σ1 (x))2 Put −1 e b c1 (x) = eT 1 Q (x)A(x) 1 , 2 σ12 (x) x ∈ Jn , e where A(x) denotes the vector with components n
 1 X e Ai (x) = 2 m1 (X ` )σ1 (X ` ) ε1` + σ12 (X ` ) ε21` − 1 ψi,hn (X ` − x) Khn (X ` − x), n i ∈ I. `=1 Along the lines of the proof of (i) and (ii) for m b 1 one can prove that sup x∈Jn    
sb1 (x) − s1 (x) %2n Mn hp+γ n − b c (x) = O + O = oP n−1/2 1 P P (1) |I| (2) 2 (1) (2) 2 2 (αn ) (αn ) αn (αn ) 2σ1 (x) and sup |b c1 (x)| = OP x∈Jn Now noticing that the rate σ b12 (x)−σ12 (x)  %n (1) (2) αn (αn )2
= oP n−1/4 . = sb1 (x)−s1 (x)−(m b 1 (x)−m1 (x))(m b 1 (x)+m1 (x)) we obtain 
sup |r(x)| = oP n−1/2 + OP x∈Jn  %2n (1) (2) (αn )2 (αn )4  = oP n−1/2
and (i) follows for σ b1 . If we define bb1 (x) = 1 + b c1 (x) − b a1 (x)m1 (x)/σ1 (x), then (ii) and (iii) follow analogously to before. The only difference is an additional factor σ1 (x) in the denominator that needs to be considered. A COPULA APPROACH FOR DEPENDENCE MODELING 33 To show validity of (iv) note that the regression model Y1i2 = s1 (X i ) + ηi holds with error term ηi = σi2 (X i )(ε21i − 1) + 2m1 (X i )σ1 (X i )ε1i . From this one obtains analogously to the derivation of (iv) for b a1 that Z n n
sb1 (x) 1 X 2 1 X m1 (X i ) f (x) dx = (ε − 1) + ε1i + oP n−1/2 . X 1i 2 2n n σ (X i ) Jn 2σ1 (x) i=1 i=1 1 R But the second sum is also the dominating term in Jn b a1 (x)m1 (x)/σ1 (x)fX (x) dx, which is again shown analogously to the proof of (iv) for b a1 . Thus (iv) follows for bb1 .  Remark 3. Note that due to property (iii) of Lemma 1 and assumption (FX ) we have for x ∈ Jn = [−cn , cn ]d , kxkν |b a1 (x)| ≤ O((log n)ν/d )o(n−1/4 ) = o(1) for every ν > 0. In the proof of Lemma 1, b a1 (x) was only defined for x ∈ Jn . Now we define d+δ d b a1 on R in a way that if b a1 ∈ C1 (Jn ) and kxkν |b a1 (x)| ≤ 1, then b a1 ∈ G defined in (A10).

d Then P b a1 ∈ G → 1 by Lemma 1. Analogously bb1 is defined on R such that P bb1 ∈ Ge → 1 for Ge from (A11). Lemma 2. Let H = G or Ge denote one of the function classes defined in (A10) and (A11), then we have with ν > d(1 + d/δ) and δ from assumption (Bw),
  d d log N (, H, k · k∞ ) = O − ν + d+δ for  & 0, and thus the same bound holds for log N[ ] (, H, k · k2 ). e and let  > 0. Choose D = D() = −1/ν . Proof. Let H = G (the proof is similar for G) Let B denote the ball of radius D around the origin. Let a1 , . . . , am : B → R denote the centers of -balls with respect to the supremum norm that cover C1d+δ (B), that is m = N (, C1d+δ (B), k · k∞ ). Then for each a ∈ G we have a|B ∈ C1d+δ (B) and thus there exists j0 ∈ {1, . . . , m} such that supx∈B |a(x) − aj0 (x)| ≤ . Now define aj (x) = 0 for x ∈ Rd \ B, j = 1, . . . , m. Then n o sup |a(x) − aj0 (x)| ≤ max , sup |a(x)| ≤  x∈Rd kxkν |a(x)| kxk≥D because ≤ 1 by definition of G. We obtain N (, G, k · k∞ ) = m and due to van der Vaart and Wellner (1996), Theorem 2.7.1, we have for some universal K
 

d d d − νd + d+δ d 1 − d+δ d − d+δ log m ≤ K λ B  = O (D + 2)  =O  ,  where B 1 = x : kx − Bk < 1 Thus the first assertion follows. The second assertion follows by van der Vaart and Wellner (1996), proof of Cor. 2.7.2.  34 NATALIE NEUMEYER1 , MAREK OMELKA2 , ŠÁRKA HUDECOVÁ2 References Berghaus, B., Bücher, A., and Volgushev, S. (2017). Weak convergence of the empirical copula process with respect to weighted metrics. Bernoulli, 23(1):743–772. Brahimi, B. and Necir, A. (2012). A semiparametric estimation of copula models based on the method of moments. Stat. Methodol., 9(4):467–477. Bücher, A. and Volgushev, S. (2013). Empirical and sequential empirical copula processes under serial dependence. J. Multivariate Anal., 119:61–70. Chan, N.-H., Chen, J., Chen, X., Fan, Y., and Peng, L. (2009). Statistical inference for multivariate residual copula of GARCH models. Statist. Sinica, 19:53–70. Chen, X. and Fan, Y. (2006). Estimation and model selection of semiparametric copula-based multivariate dynamic models under copula misspecification. J. Econometrics, 135:125–154. Dedecker, J. and Louhichi, S. (2002). Maximal inequalities and empirical central limit theorems. In Dehling, H., Mikosch, T., and Sorensen, M., editors, Empirical process techniques for dependent data, pages 137–160. Birkhäuser Boston. Dette, H., Pardo-Fernández, J. C., and Keilegom, I. V. (2009). Goodness-of-fit tests for multiplicative models with dependent data. Scand. J. Statist., 36(4):782–799. Embrechts, P., Frey, R., and McNeil, A. (2005). Quantitative Risk Management: Concepts, Techniques, and Tools. Princeton Series in Finance, Princeton. Fan, J. and Gijbels, I. (1996). Local Polynomial Modelling and Its Applications. Chapman & Hall/CRC, London. Fan, J. and Yao, Q. (2005). Nonlinear Time Series : Nonparametric and Parametric Methods. Springer series in statistics. Springer, New York. Gao, J. (2007). Nonlinear Time Series: Semiparametric and Nonparametric Methods. Chapman & Hall/CRC, Boca Raton. Genest, C., Ghoudi, K., and Rivest, L.-P. (1995). A semiparametric estimation procedure of dependence parameters in multivariate families of distributions. Biometrika, 82:543–552. Genest, C., Rémillard, B., and Beaudoin, D. (2009). Goodness-of-fit tests for copulas: A review and a power study. Insurance Math. Econom., 44:199–213. Gijbels, I., Omelka, M., Pešta, M., and Veraverbeke, N. (2017). Score tests for covariate effects in conditional copulas. J. Multivariate Anal. Available online, doi: 10.1016/j.jmva.2017.05.001. Gijbels, I., Omelka, M., and Veraverbeke, N. (2015). Estimation of a copula when a covariate affects only marginal distributions. Scand. J. Statist., 42(4):1109–1126. Hansen, B. E. (2008). Uniform convergence rates for kernel estimation with dependent data. Econometric Theory, 24:726–748. Härdle, W., Tsybakov, A., and Yang, L. (1998). Nonparametric vector autoregression. J. Statist. Plann. Inference, 68(2):221–245. A COPULA APPROACH FOR DEPENDENCE MODELING 35 Kim, G., Silvapulle, M. J., and Silvapulle, P. (2007). Semiparametric estimation of the error distribution in multivariate regression using copulas. Aust. N. Z. J. Stat., 49(3):321–336. Kim, G., Silvapulle, M. J., and Silvapulle, P. (2008). Estimating the error distribution in multivariate heteroscedastic time-series models. J. Statist. Plann. Inference, 138(5):1442– 1458. Kojadinovic, I. and Holmes, M. (2009). Tests of independence among continuous random vectors based on Cramér–von Mises functionals of the empirical copula process. J. Multivariate Anal., 100(6):1137–1154. Masry, E. (1996). Multivariate local polynomial regression for time series: uniform strong consistency and rates. J. Time Series Anal., 17(6):571–599. Müller, U. U., Schick, A., and Wefelmeyer, W. (2009). Estimating the error distribution function in nonparametric regression. Statist. Probab. Lett., 79:957–964. Nelsen, R. B. (2006). An Introduction to Copulas. Springer, New York. Second Edition. Patton, A. J. (2012). A review of copula models for economic time series. J. Multivariate Anal., 110:4–18. Portier, F. and Segers, J. (2015). On the weak convergence of the empirical conditional copula under a simplifying assumption. arXiv preprint arXiv:1511.06544. Rémillard, B., Papageorgiou, N., and Soustra, F. (2012). Copula-based semiparametric models for multivariate time series. J. Multivariate Anal., 110:30–42. Segers, J. (2012). Weak convergence of empirical copula processes under nonrestrictive smoothness assumptions. Bernoulli, 18:764–782. Tsukahara, H. (2005). Semiparametric estimation in copula models. Canad. J. of Statist., 33:357–375. van der Vaart, A. W. and Wellner, J. A. (1996). Weak Convergence and Empirical Processes. Springer, New York. Veraverbeke, N., Omelka, M., and Gijbels, I. (2011). Estimation of a conditional copula and association measures. Scand. J. Statist., 38:766–780. Yang, L., Härdle, W., and Nielsen, J. (1999). Nonparametric autoregression with multiplicative volatility and additive mean. J. Time Series Anal., 20(5):579–604. 

10math.ST

arXiv:1611.00507v1 [cs.DS] 2 Nov 2016 Worst Case Competitive Analysis of Online Algorithms for Conic Optimization Reza Eghbali Department of Electrical Engineering University of Washington Seattle, WA 98195 eghbali@uw.edu Maryam Fazel Department of Electrical Engineering University of Washington Seattle, WA 98195 mfazel@uw.edu Abstract Online optimization covers problems such as online resource allocation, online bipartite matching, adwords (a central problem in e-commerce and advertising), and adwords with separable concave returns. We analyze the worst case competitive ratio of two primal-dual algorithms for a class of online convex (conic) optimization problems that contains the previous examples as special cases defined on the positive orthant. We derive a sufficient condition on the objective function that guarantees a constant worst case competitive ratio (greater than or equal to 12 ) for monotone objective functions. We provide new examples of online problems on the positive orthant and the positive semidefinite cone that satisfy the sufficient condition. We show how smoothing can improve the competitive ratio of these algorithms, and in particular for separable functions, we show that the optimal smoothing can be derived by solving a convex optimization problem. This result allows us to directly optimize the competitive ratio bound over a class of smoothing functions, and hence design effective smoothing customized for a given cost function. 1 Introduction Given a proper convex cone K ⊂ Rn , let ψ : K 7→ R be an upper semi-continuous concave function. Consider the optimization problem ! m X maximize ψ At x t (1) t=1 subject to xt ∈ Ft , ∀t ∈ [m], where for all t ∈ [m] := {1, 2, . . . , m}, xt ∈ Rk are the optimization variables and Ft are compact convex constraint sets. We assume At ∈ Rn×k maps Ft to K; for example, when K = Rn+ and Ft ⊂ Rk+ , this assumption is satisfied if At has nonnegative entries. We consider problem (1) in the online setting, where it can be viewed as a sequential game between a player (online algorithm) and an adversary. At each step t, the adversary reveals At , Ft and the algorithm chooses x̂t ∈ Ft . The performance of the algorithm is measured by its competitive ratio, i.e., the ratio of objective value at x̂1 , . . . , x̂m to the offline optimum. Problem (1) covers (convex relaxations of) various online combinatorial problems including online bipartite matching [25], the “adwords” problem [31], and the secretary problem [27]. More generally, it covers online linear programming (LP) [9], online packing/covering with convex cost [4, 7, 11], and generalization of adwords [13]. Online LP is an important example on the positive orthant: maximize subject to m X cTt xt t=1 m X Bt xt ≤ b t=1 T 1 xt ≤ 1, xt ∈ Rk+ , ∀t ∈ [m]. Here Ft = {x ∈ Rk+ | 1T x ≤ 1}, ATt = [ct , BtT ] and ψ(u, v) = u + I{·≤b} (v) where I{·≤b} (v) is the concave indicator function of the set {v ≤ b}. The competitive performance of online algorithms has been studied mainly under the worst-case model (e.g., in [31]) or stochastic models (e.g., in [27]). In the worst-case model one is interested in lower bounds on the competitive ratio that hold for any (A1 , F1 ), . . . , (Am , Fm ). In stochastic models, adversary chooses a probability distribution from a family of distributions to generate (A1 , F1 ), . . . , (Am , Fm ), and the competitive ratio is calculated using the expected value of the algorithm’s objective value. The two most studied stochastic models are random permutation and i.i.d. models. In the random permutation model the adversary is limited to choose distributions that are invariant under random permutation while in the i.i.d. model (A1 , F1 ), . . . , (Am , Fm ) are required to be independent and identically distributed. Note that the worst case model can also be viewed as an stochastic model if one allows distributions with singleton support. Online bipartite matching, and its generalization the “adwords” problem, are two main examples that have been studied under the worst case model. The greedy algorithm achieves a competitive ratio of 1/2 while the optimal algorithm achieves a competitive ratio of 1 − 1/e as bid to budget ratio goes to zero [31, 8, 25, 23]. A more general version of adwords in which each agent (advertiser) has a concave cost has been studied in [13]. On the other hand, secretary problem and its generalizations are stated for random permutation model since the worst case competitive ratio of any online algorithm can be arbitrary small [5, 27]. The convex relaxation of the online secretary problem is a simple linear program (LP) with one constraint. Therefore, the competitive ratio of online algorithms for LP has either been analyzed under random permutation or i.i.d models [3, 16, 14, 20, 26] or under the worst case model with further restrictions on the problem data [9]. Several authors have also proposed algorithms for adwords and bipartite matching problems under stochastic models which have a better competitive ratio than the competitive ratio under the worst case model [29, 24, 17, 30, 21, 12]. The majority of algorithms proposed for the problems mentioned above rely on a primal-dual framework [8, 9, 4, 13, 7]. The differentiating point among the algorithms is the method of updating the dual variable at each step, since once the dual variable is updated the primal variable can be assigned using a simple assignment rule based on complementary slackness condition. A simple and efficient method of updating the dual variable is through a first order online learning step. For example, the algorithm stated in [14] for online linear programming uses mirror descent with entropy regularization (multiplicative weight updates algorithm) once written in the primal dual language. Recently, the work in [14] was independently extended to the random permutation model in [19, 15, 2]. In [2], the authors provide competitive difference bound for online convex optimization under random permutation model as a function of the regret bound for the online learning algorithm applied to the dual. In this paper, we consider two versions of the greedy algorithm for problem (1) The first algorithm, Algorithm 1, updates the primal and dual variables sequentially. The second algorithm, Algorithm 2, provides a direct saddle-point representation of what has been described informally in the literature as “continuous updates” of primal and dual variables. This saddle point representation allows us to generalize this type of updates to non-smooth function. In section 2, we bound the competitive ratios of the two algorithms. A sufficient condition on the objective function that guarantees a non-trivial worst case competitive ratio is introduced. We show that the competitive ratio is at least 12 for a monotone non-decreasing objective function. Examples that satisfy the sufficient condition (on the positive orthant and the positive semidefinite cone) are given. In section 3, we derive optimal algorithms, as variants of greedy algorithm applied to a smoothed version of ψ. For example, Nesterov smoothing provides optimal algorithm for online LP. The main contribution of this paper is to show how one can derive the optimal smoothing function (or from the dual point of view the optimal regularization function) for separable ψ on positive orthant by solving a convex optimization 2 problem. This gives a implementable algorithm that achieves the optimal competitive ratio derived in [13]. We also show how this convex optimization can be modified for the design of the smoothing function specifically for the sequential algorithm. In contrast, [13] only considers continuous updates algorithm. Notation. We denote the transpose of a matrix A by AT . The inner product on Rn is denoted by h·, ·i. Given a function ψ : Rn 7→ R, ψ ∗ denotes the concave conjugate of ψ defined as ψ ∗ (y) = inf hy, ui − ψ(u), u for all y ∈ Rn . For a concave function ψ, ∂ψ(u) denotes the set of supergradients of ψ at u, i.e., the set of all y ∈ Rn such that ∀u0 ∈ Rn : ψ(u0 ) ≤ hy, u0 − ui + ψ(u). The set ∂ψ is related to the concave conjugate function ψ ∗ as follows. For an upper semi-continuous concave function ψ we have ∂ψ(u) = argmin hy, ui − ψ ∗ (y). y Under this condition, (ψ ∗ )∗ = ψ. A differentiable function ψ has a Lipschitz continuous gradient with respect to k·k with continuity parameter µ1 > 0 if for all u, u0 ∈ Rn , 1 ∗ k∇ψ(u0 ) − ∇ψ(u)k ≤ ku − u0 k , µ ∗ where k·k is the dual norm to k·k. For an upper semi-continuous concave function ψ, this is ∗ equivalent to ψ ∗ being µ-strongly concave with respect to k·k (see, for example, [36, chapter 12]). The dual cone K ∗ of a cone K ⊂ Rn is defined as K ∗ = {y | hy, ui ≥ 0 ∀u ∈ K}. Two examples of self-dual cones are the positive orthant Rn+ and the cone of n × n positive semidefinite matrices n S+ . A proper cone (pointed convex cone with nonempty interior) K induces a partial ordering on Rn which is denoted by ≤K and is defined as x ≤K y ⇔ y − x ∈ K. For two sets F, G ⊂ Rn , we write F ≤K G when u ≤K v for all u ∈ F, v ∈ G. 1.1 Two greedy algorithms The (Fenchel) dual problem for problem (1) is given by m X minimize σt (ATt y) − ψ ∗ (y), (2) t=1 where the optimization variable is y ∈ Rn , and σt denotes the support function for the set Ft defined as σt (z) = supx∈Ft hx, zi. We denote the optimal dual objective with D? . A pair (x∗ , y ∗ ) ∈ (F1 × . . . Fm ) × K ∗ is an optimal primal-dual pair if and only if x∗t ∈ argmax x, ATt y ∗ ∀t ∈ [m], x∈Ft m X y ∗ ∈ ∂ψ( At x∗t ). (3) (4) t=1 Based on these optimality conditions, we consider two algorithms. Algorithm 1 updates the primal and dual variables sequentially, by maintaining a dual variable ŷt and using it to assign x̂t ∈ argmaxx∈Ft x, ATt ŷt . The algorithm then updates the dual variable based on the second optimality condition (4) 1 . By the assignment rule, we have At x̂t ∈ ∂σt (ŷt ), and the dual variable update can be viewed as * t + X ŷt+1 ∈ argmin As x̂s , y − ψ ∗ (y). y s=1 Therefore, the dual update is the same as the update in dual averaging [34] or Follow The Regularized Leader (FTRL) [39, 1][37, section 2.3] algorithm with regularization −ψ ∗ (y). 1 we assume that ∂ψ nonempty on the boundary of K, otherwise, we can modify the algorithm to start from a point in the interior of the cone very close to zero. 3 Algorithm 1 Sequential Update Initialize ŷ1 ∈ ∂ψ(0) for t ← 1 to m do Receive At , Ft x̂t ∈ argmaxx∈Ft x, ATt ŷt Pt ŷt+1 ∈ ∂ψ( s=1 As x̂s ) end for Algorithm 2 updates the primal and dual variables simultaneously, ensuring that x̃t ∈ argmax x∈Ft x, ATt ỹt , ỹt ∈ ∂ψ( t X As x̃s ). s=1 This algorithm is inherently more complicated than algorithm 1, since finding x̃t involves solving a saddle-point problem. This can be solved by a first order method like the mirror descent algorithm for saddle point problems. In contrast, the primal and dual updates in algorithm 1 solve two separate maximization and minimization problems 2 Algorithm 2 Simultaneous Update for t ← 1 to m do Receive At , Ft (ỹt , x̃t ) ∈ arg miny maxx∈Ft D E Pt−1 y, At x + s=1 As x̃s − ψ ∗ (y) end for For a reader more accustomed to optimization algorithms, we would like to point to alternative views on these algorithms. If 0 ∈ Ft for all t, then an alternative view on Algorithm 2 is coordinate minimization. Initially, all the coordinates are set to zero, then at step t ! t X x̃t ∈ argmax ψ At x̃s + At x . (5) x∈Ft s=1 If in addition ψ is differentiable, Algorithm 1, at each time step t, applies one iteration of Frank-Wolfe algorithm [18] for solving (5). 2 Competitive ratio bounds and examples for ψ In this section, we derive bounds on the competitive ratios of Algorithms P1mand 2 by bounding their respective duality gaps. Let ỹm+1 to be the minimum element in ∂ψ( t=1 At x̃t ) with respect to ordering ≤K ∗ (such an element exists in the superdifferential by Assumption (1), which appears Pm later in this section). We also choose ŷm+1 to be the minimum element in ∂ψ( t=1 At x̂t ) with respect to ≤K ∗ . Note that ŷm+1 is used for analysis and does not play a role in the assignments of Algorithm 1. Let Pseq and Psim denote the primal objective values for the primal solution produced by the algorithms 1 and 2, and Dseq and Dsim denote the corresponding dual objective values, ! ! m m X X Pseq = ψ At x̂t , Psim = ψ At x̃t , t=1 Dseq = m X t=1 σt (ATt ŷt ) − ψ ∗ (ŷm+1 ), Dsim = t=1 m X σt (ATt ỹt ) − ψ ∗ (ỹm+1 ). t=1 2 Also if the original problem is a convex relaxation of an integer program, meaning that each Ft = convFt where Ft ⊂ Zl , then x̂t can always be chosen to be integral while integrality may not hold for the solution of the second algorithm. 4 The next lemma provides a lower bound on the duality gaps of both algorithms. Lemma 1 The duality gap for Algorithm 2 is lower bounded as Psim − Dsim ≥ ψ ∗ (ỹm+1 ) + ψ(0), and the duality gap for Algorithm 1 is lower bounded as Pseq − Dseq ≥ ψ ∗ (ŷm+1 ) + ψ(0) + m X hAt x̂t , ŷt+1 − ŷt i . t=1 Furthermore, if ψ has a Lipschitz continuous gradient with parameter 1/µ with respect to k·k, then Pseq − Dseq ≥ ψ ∗ (ŷm+1 ) + ψ(0) − m X 1 2 kAt x̂t k . 2µ t=1 (6) Note that right hand side of (6) is exactly the regret bound of the FTRL algorithm (with a negative sign) [38]. The proof is given in Appendix A. To simplify the notation in the rest of the paper, we assume ψ(0) = 0. Now we state a sufficient condition on ψ that leads to non-trivial competitive ratios, and we assume this condition holds in the rest of the paper. One can interpret this assumption as having “diminishing returns” with respect to the ordering induced by a cone. Examples of functions that satisfy this assumption will appear later in this section. Assumption 1 Whenever u ≥K v, there exists y ∈ ∂ψ(u) that satisfies y ≤K ∗ z, for all z ∈ ∂ψ(v). When ψ is differentiable, Assumption 1 simplifies to u ≥K v ⇒ ∇ψ(u) ≤K ∗ ∇ψ(v). That is, the gradient, as a map from Rn (equipped with ≤K ) to Rn (equipped with ≤K ∗ ), is orderreversing (also known as antitone). If ψ satisfies Assumption 1, then so does ψ ◦ A, when A is a linear map such that hy, Avi ≥ 0, for all v ∈ K and y ∈ K ∗ . When ψ is twice differentiable, Assumption 1 is equivalent to w, ∇2 ψ(u)v ≤ 0, for all u, v, w ∈ K. For example, this is equivalent to Hessian being element-wise non-positive when K = Rn+ . Competitive ratio for non-decreasing ψ 2.1 To quantify the competitive ratio of the algorithms, we define αψ : K 7→ R as αψ (u) = sup {c | ψ ∗ (y) ≥ cψ(u), ∀y ∈ ∂ψ(u)}, (7) which for u 6= 0 can also be written as αψ (u) = ψ ∗ (y) . y∈∂ψ(u) ψ(u) inf Since ψ ∗ (y) + ψ(u) = hy, ui for all y ∈ ∂ψ(u), αψ is equivalent to3 αψ (u) = sup{c | hy, ui ≥ (c + 1)ψ(u), ∀y ∈ ∂ψ(u)} (8) = sup{c | ψ 0 (u; u) ≥ (c + 1)ψ(u)}, where ψ 0 (u, v) is the directional derivative of ψ at u in the direction of v. Note that for any u ∈ K, we have −1 ≤ αψ (u) ≤ 0 since for any y ∈ ∂ψ(u), by concavity of ψ and the fact that y ∈ K ∗ , we have 0 ≤ hy, ui ≤ ψ(u) − ψ(0). If ψ is a linear function then αψ = 0, while if 0 ∈ ∂ψ(u) for some u ∈ K, then αψ (u) = −1. Figure 1 shows a differentiable function ψ : R+ 7→ R. The value of α(u0 ) is the ratio of ψ ∗ (ψ 0 (u0 )) to ψ(u0 ) as shown. The next theorem provides lower bounds on the competitive ratio of the two algorithms. 3 If ψ(0) 6= 0, then the definition of αψ should be modified to αψ (u) = sup{c | hy, ui ≥ (c + 1)(ψ(u) − ψ(0)), y ∈ ∂ψ(u)}. 5 ψ(u0 ) −ψ ∗ (ψ 0 (u0 )) u0 u Figure 1: ψ 0 is the derivative of ψ and −ψ ∗ (ψ 0 (u0 )) is the y-intercept of the tangent to the function graph at u0 . αψ (u0 ) is the ratio of ψ ∗ (ψ 0 (u0 )) to ψ(u0 ). Theorem 1 If Assumption 1 holds, then for the simultaneous update algorithm we have 4 : Psim ≥ 1 D? . 1 − ᾱψ where D∗ is the dual optimal objective and ᾱψ = inf{αψ (u) | u ∈ K}. (9) For the sequential update algorithm, Pseq ≥ m X 1 (D? + hAt x̂t , ŷt+1 − ŷt i). 1 − ᾱψ t=1 Further, if ψ is differentiable with gradient Lipschitz continuity parameter Pseq ≥ Proof: 1 µ with respect to k·k, m X 1 1 2 kAt x̂t k ). (D? − 1 − ᾱψ 2µ t=1 (10) We first show that Assumption 1 implies that m X σt (ATt ỹt ) ≥ t=1 m X m X σt (At ŷt ) ≥ σt (ATt ỹm+1 ), t=1 m X t=1 (11) σt (At ŷm+1 ), t=1 To do so, we prove that for all t, σt (ATt ỹt ) ≥ σt (ATt ỹm+1 ). The proof for σt (ATt ŷt ) ≥ σt (ATt ŷm+1 ) follows the same steps. Pt Pm For any t, we have for all s. Since ỹt ∈ s=1 As x̃s ≤K s=1 As x̃s since As Fs ⊂ KP Pt m ∂ψ( s=1 As x̃s ) and ỹm+1 was chosen to be the minimum element in ∂ψ( s=1 As x̃s ) with respect to ≤K ∗ , by the Assumption 1, we have ỹt ≥K ∗ ỹm+1 . Since At x ∈ K for all x ∈ Ft , we get hAt x, yt i ≥ hAt x, ym i . Therefore, σt (ATt ỹt ) ≥ σt (ATt ỹm+1 ). Now by (11) Dsim = m X σt (ATt ỹt ) ∗ − ψ (ỹm+1 ) ≥ m X t=1 σt (ATt ỹm+1 ) − ψ ∗ (ỹm+1 ) ≥ D? . t=1 ? Similarly, Dseq ≥ D . Lemma 1 and definition of ᾱψ give the desired result. 4 If ψ(0) 6= 0, then the conclusion becomes Psim − ψ(0) ≥ 6 1 1−αψ (D? − ψ(0)).  Competitive ratio for non-monotone ψ 2.2 If ψ is not non-decreasing with respect to K, the algorithms are not guaranteed to increase the objective at each step; therefore, Pseq and Psim are not guaranteed to be non-negative. However, if we add the assumption that 0 ∈ Ft for all t, then Algorithm 2 will not decrease the objective at any step, t t−1 X X ψ( As x̃s ) − ψ( As x̃s ) ≥ x̃t , ATt ỹt = max x, ATt ỹt ≥ 0, s=1 x∈Ft s=1 (12) which follows from concavity of ψ. In other words, the algorithm simply assigns 0 to xt if any other feasible point in Ft reduces the objective value. Define m X ũ = At x̃t , t=1 and note that under the assumption 0 ∈ Ft , we have ψ(ũ) ≥ 0. Further, we have: 1 Psim ≥ D? . 1 − αψ (ũ) (13) We provide examples of non-monotone ψ and their competitive ratio analysis in this section and section 3.1. Given appropriate conditions on Ft and At , in order to find a lower bound on the competitive ratio independent of ũ, we only need to lower bound αψ over a subset of K. Note that when ψ is not non-decreasing, then there exists a supergradient that is not in K ∗ . Therefore, αψ (ũ) for general ψ can be less than −1. In this case, the lower bound for the competitive ratio of Algorithm (2) is less than 12 . We now consider examples of ψ that satisfy Assumption 1 and derive lower bound on αψ for those examples. 2.3 Examples on the positive orthant. Let K = Rn+ and note that K ∗ = K. To simplify the notation we use ≤ instead of ≤Rn+ . When ψ has continuous partial second derivatives with respect to all the variables, Assumption 1 is equivalent to the Hessian being element-wise non-positive over Rn+ . Assumption 1 is closely related to submodularity. In fact, on the positive orthant this assumption is sufficient for submodularity of ψ on the lattice defined by ≤. However, this assumption is not necessary for submodularity. When ψ has continuous partial second derivatives with respect to all the variables, the necessary and sufficient condition only requires the off-diagonal elements of the Hessian to be non-positive [28]. If ψ is separable, i.e., ψ(u) = n X ψi (ui ), (14) i=1 Assumption 1 is satisfied since by concavity for each ψi we have ∂ψi (ui ) ≤ ∂ψi (vi ) when ui ≤ vi . If ψ satisfies the Assumption 1, then so does ψ(Au), where A is an element-wise non-negative matrix. Adwords In the basic adwords problem, for all t, Ft = {x ∈ Rl+ | 1T x ≤ 1}, At is a diagonal matrix with non-negative entries, and n n X X ψ(u) = ui − (ui − 1)+ , (15) i=1 ∗ i=1 T where (·)+ = max{·, 0}. In this problem, ψ (y) = 1 (y − 1). Since 0 ∈ ∂ψ(1) we have αψ = −1 by (8); therefore, the competitive ratio of algorithm 2 is 12 . Let r = maxt,i,j At,ij , then we have Pm 1 t=1 hAt x̂t , ŷt+1 − ŷt i ≤ nr. Therefore, the competitive ratio of algorithm 1 goes to 2 as r (bid to budget ratio) goes to zero. In adwords with concave returns studied in [13], At is diagonal for all t and ψ is separable 5 . 5 Note that in this case one can remove the assumption that ∂ψi ⊂ R+ since if ỹt,i = 0 for some t and i, then x̃s,i = 0 for all s ≥ t. 7 Online linear program and non-separable budget Recall that online LP is given by ! m m X X T maximize ct xt + I{·≤1} Bt xt t=1 t=1 x t ∈ Ft , ∀t ∈ [m]. where Ft = {x ∈ Rk+ | 1T x ≤ 1} is the simplex. If a lower bound on the optimal dual variable mini yi∗ ≥ −l is given, then the LP can be written in the exact penalty form [6]: ! m m X X T maximizext ∈Ft ct x t + G Bt xt . (16) t=1 t=1 where G is an l-Lipschitz continuous function with respect to l1 norm given by n X G(u) = −l (ui − 1)+ . i=1 −l is a lower bound on dual variable if   ct,j l > max | Bt,ij > 0, j ∈ [k], i ∈ [n] . Bt,ij (17) To prove this, by way of contradiction, we assume that yi∗ ≤ −l for some i ∈ [n]. Now by the definition of l, we have ct,j + (BtT y ∗ )j < 0 for all j such that Bt,ij > 0. On the other hand by the optimality condition (3), ct + BtT y ∗ ∈ NFt (x∗t ) where NFt (x∗t ) is the normal cone of the simplex at x∗t . Therefore, we should have x∗t,j = 0 for all j such that Bt,ij > 0. This results in (Bt x∗t )i = 0 Pm which yield ( t=1 Bt x∗t )i = 0. This means that the corresponding variable yi∗ = 0 which is a contradiction. With this choice of l, Algorithm 2 always maintains a feasible solution for the original problem when applied to (16). This follows from the fact that if ỹt,i = −l for some t and i, then (Bt x̃t )i = 0. Note that in the adwords problem Bt = diag(ct ) and l = 1. For any p ≥ 1 let Bp be Pnthe lp -norm ball. In order to provide examples of non-separable G, we rewrite the function i=1 (ui − 1)+ using the distance from B∞ . For any set C ⊂ Rn+ , let d1 (u, C) = inf ū∈C ku − ūk1 and note that d1 (u, C) is 1-Lipschitz continuous with respect to l1 norm. We have: n X (1 − ui )+ = d1 (u, B∞ ). (18) i=1 Pm Consider a problem with constraint k t=1 Bt xt kp ≤ 1. Given the bound on dual variable in (17), this problem can be equivalently written in the form of (16) with the exact penalty [10, Theorem 5.5] G(u) = −ld1 (u, Bp ). When p = ∞, we get back (18). For p ≥ 1, although not separable, the function G(u) = −d1 (u, Bp ) satisfies Assumption 1. For B1 , that follows from the fact that d1 (u, B1 ) = (1T u − 1)+ ; therefore, ∂d1 (u, B1 ) = 1∂f (1T u), where f (x) = (x − 1)+ . The proof for p > 1 is given in Appendix B. When ψ is given by 16 with G(u) = −ld1 (u, Bp ), we have αψ (ũ) ≥ − θl , where θ = mint minx∈Ft 2.4 cT t x . 1T Bt x The derivation of this bound is also given in Appendix B. Examples on the positive semidefinite cone. n Let K = S+ and note that K ∗ = K. An interesting example that satisfies Assumption 1 is p ψ(U ) = trU with p ∈ (0, 1), where ∇ψ(U ) = pU p−1 and ᾱψ = p − 1. This objective function is used in pth mean optimal experiment design. Another example is ψ(U ) = log det(U + A0 ), where ∇ψ(U ) = (U + A0 )−1 and ᾱψ = −1 since h∇ψ(U ), U i n − tr(A0 + U )−1 A0 = → 0 as trU → ∞. ψ(U ) − ψ(0) log det((A0 + U )A−1 0 ) 8 Maximizing log det arises in several offline applications including D-optimal experiment design [35], maximizing the Kirchhoff complexity of a graph [41], Optimal sensor selection [22, 40]. An example of applications for maximizing the logdet of a projected Laplacian, as a function of graph edge weights, appears in connectivity control of mobile networks [41]. We derive the competitive ratio of the greedy algorithm with smoothing for online experiment design and Kirchhoff complexity maximization of a graph in section 3. 3 Smoothing of ψ for improved competitive ratio The technique of “smoothing” an objective function, or equivalently adding a strongly convex regularization term to its conjugate function, has been used in several areas. In convex optimization, a general version of this is due to Nesterov [32], and has led to faster convergence rates of first order methods for non-smooth problems. In this section, we study how replacing ψ with a appropriately smoothed function ψS helps improve the performance of the two algorithms discussed in section 1.1, and show that it provides optimal competitive ratio for two of the problems mentioned in section 2, adwords and online LP. We then show how to maximize the competitive ratio of Algorithm 2 for a separable ψ and compute the optimal smoothing by solving a convex optimization problem. This allows us to design the most effective smoothing customized for a given ψ: we maximize the bound on the competitive ratio over the set of smooth functions.(see subsection 3.2 for details). Let ψS denote an upper semi-continuous concave function (a smoothed version of ψ), and suppose ψS satisfies Assumption 1. The algorithms we consider in this section are the same as Algorithms 1 and 2, but with ψ replacing ψS . Note that the competitive ratio is computed with respect to the original problem, that is the offline primal and dual optimal values are still the same P ? and D? as before. If we replace ψ with ψS in algorithms 1 and 2, the dual updates are modified to * t + X ŷt+1 ∈ argmin As x̂s , y − ψS ∗ (y), y s=1 * (x̃t , ỹt ) ∈ arg min max y x∈Ft y, At x + t−1 X + As x s − ψS ∗ (y). s=1 From Lemma 1, we have that Dsim ≤ ψS Dseq ≤ ψS m X t=1 m X ! At x̃t − ψ ∗ (ỹm+1 ) ! At x̂t − ψ ∗ (ŷm+1 ) − t=1 (19) m X hAt x̂t , ŷt+1 − ŷt i . (20) t=1 Similar to our assumption on ψ(0), to simplify the notation, by replacing ψS (·) with ψS (·) − ψS (0), we assume ψS (0) = 0. Define αψ,ψS (u) = sup{c |ψ ∗ (y) ≥ ψS (u) + (c − 1)ψ(u), ∀y ∈ ∂ψS (u)}, and ᾱψ,ψS = inf{αψ,ψS (u) | u ∈ K}. Now the conclusion of Theorem 1 holds with ᾱψ replaced by ᾱψ,ψS . Similarly, when ψS is nonmonotone, inequality (13) holds with αψ replaced by αψ,ψS . 3.1 Nesterov Smoothing We first consider Nesterov smoothing [32], and apply it to examples on non-negative orthant. Given a proper upper semi-continuous concave function φ : Rn 7→ R ∪ {−∞}, let ψS = (ψ ∗ + φ∗ )∗ . Note that ψS is the hypo-sum (sup-convolution) of ψ and φ. ψS = ψφ(u) = sup ψ(v) + φ(u − v). v 9 This is called hypo-sum of ψ and φ since the hypo-graph of ψS is the Minkowski sum of hypo-graphs of ψ and φ. If ψ and φ are separable, then ψφ satisfies Assumption 1 for K = Rn+ . Here we provide example of Nesterov smoothing for functions on non-negative orthant. Adwords and Online LP: Consider the problem (16) with G(u) = −l problem we smooth G with: Pn i=1 (ui − 1)+ . For this m 1 X (e − 1) θ φ∗ (y) = ( ((yi − ) log(1 − yi )) − (1 + γ)1T y), γ i=1 (e − 1) θ where γ = log(1 + l(e−1) ), θ cTt x T x∈Ft 1 Bt x θ = min min t and l is defined as in (17). In this case, we have; 1 )γ. e−1 αψ,ψφ (ũ) ≥ 1 − (1 + (see the Appendix C for the derivation). This gives the competitive ratio of γ1 (1 − 1/e). In the case of adwords where θ = l, this yield the optimal competitive ratio of 1 − e−1 and the smoothed function coincides with the one derived in the previous paragraph. For a general LP, this approach provides the optimal competitive ratio, which is known to be O(γ −1 ) [9]. Online experiment design and online graph formation: As pointed in section 2.4, maximizing the determinant arises in graph formation, sensor selection and optimal experiment design. In the offline form, these problems can be cast into maximizing log det subject to linear constraints as follows: m X maximize log det(A0 + at aTt xt ) (21) t=1 subject to m X xt ≤ b, t=1 xt ∈ [0, 1], ∀t, where A0  0 and at ∈ Rn . We describe the online version of this problem for graph formation and experiment design case. Consider the following online problem on a graph: given an underlying connected graph with the Laplacian matrix L0 , at round t, the online algorithm is presented with an edge and should decide whether to pick or drop theP edge by choosing xt ∈ {0, 1}. There is a n bound on the number of the edges that can be picked, t=1 xt ≤ b. The goal is to maximize the Kirchhoff complexity of the graph. The convex relaxation of the optimization problem is given by 21 where A0 = L0 + 11T and at is the incidence vector of the edge presented at round t. The online experiment design or sensor selection is the same as (21). In that problem the vectors at are experiment or measurement vectors. The dual to 21 is as follows maximize m X (aTt Y at + y)+ + n − log det(Y ) + trA0 Y − by t=1 subject to y ≤ 0, Y ≥ 0. The hard budget constraint in problem 21 can be replaced by the exact penalty function G(u) = −l(u − b)+ when −l is a given lower bound on the optimal dual variable corresponding to the constraint. −l is a lower bound on the optimal dual variable corresponding to the constraint if l> 2 . λmin (A0 ) 10 (22) Pn This follows from the fact that by the optimality condition (4), Y ∗ = (A0 + t=1 at aTt x∗t )−1 . If y ∗ ≤ −l, then aTt Y ∗ at + y ∗ < 0 for all t. ThisP combined with the the primal optimality condition m (3) forces x∗t = 0 for all t. Therefore, we get t=1 x∗t = 0 and y ∗ = 0, which is a contradiction. Similar to the previous problems, we smooth G with: b θ e−1 ((y − ) log(1 − y) − (1 + γ)y). γ e−1 θ φ∗ (y) = Here γ = log(1 + θl ) and θ = log(1 + 1/n). For this problem we have: αψ,ψφ (ũ) ≥ 1 − (1 + 1 1 )γ + ᾱlog det ≥ −(1 + )γ. e−1 e−1 This results in the following competitive ratio bound 1 Psim − log det(A0 ) ≥  1+ 1+ 1 e−1  (D? − log det(A0 )). γ The proof is included in the Appendix C. In the graph formation problem, λmin (A0 ) = λ2 (L0 ), that is the second smallest eigenvalue of L0 . For a connected graph λ2 (L0 ) ∈ [1/n, n] with lower bound achieved by the path graph and the upper bound achieved by the complete graph. Therefore, the competitive ratio in this case is Ω( log1 n ). 3.2 Computing optimal smoothing for separable functions on Rn+ We now tackle the problem of finding the optimal smoothing for separable functions on the positive orthant, which as we show in an example at the endP of this section is not necessarily Pn given by Nesterov n smoothing. Given a separable monotone ψ(u) = i=1 ψi (ui ) and ψS (u) = i=1 ψS i (ui ) on Rn+ we have that ᾱψ,ψS ≥ mini ᾱψi ,ψS i . To simplify the notation, drop the index i and consider ψ : R+ 7→ R. We formulate the problem of finding ψS to maximize αψ,ψS as an optimization problem. In section 4 we discuss the relation between this optimization method and the optimal algorithm presented in [13]. We set ψM (u) = Ru y(s)ds with y a continuous function, and state the infinite dimensional convex optimization 0 problem with y as a variable, minimize β Z subject to (23) u y(s)ds − ψ ∗ (y(u)) ≤ βψ(u), ∀u ∈ [0, ∞) 0 y ∈ C[0, ∞). where β = 1− ᾱψ,ψS (theorem 1 describes the dependence of the competitive ratios on this parameter). Note that we have not imposed any condition on y to be non-increasing (i.e., the corresponding ψS to be concave). The next lemma establishes that every feasible solution to the problem (23) can be turned into a non-increasing solution. Lemma 2 Let (y, β) be a feasible solution for problem (23) and define ȳ(u) = inf s≤u y(s). Then (ȳ, β) is also a feasible solution for problem (23). In particular if (y, β) is an optimal solution, then so is (ȳ, β). The proof is given in the supplement.  Re- visiting the adwords problem, we observe that the optimal solution is given by y(u) = e−exp(u) e−1 + , which is the derivative of the smooth function we derived using Nesterov smoothing in section 3.1. The optimality of this y can be established by providing a dual certificate, a measure ν corresponding to the inequality constraint, that together with y satisfies the optimality condition. If we set dν = f (u) du with f (u) = exp (1 − u)/(e − 1), the optimality conditions are satisfied with 11 β = (1 − 1/e)−1 . Z ∞ f (u)ψ(u) du = 1, f  0, Zu=0 ∞ f (s) ds ∈ f (u)∂ψ ∗ (y(u)), ∀u ≥ 0, Zs=u u y(s)ds − ψ ∗ (y(u)) ≤ βψ(u), ∀u ≥ 0, 0 Z u f (u)( y(s)ds − ψ ∗ (y(u)) − βψ(u)) = 0, ∀u ≥ 0. 0 Also note that if ψ plateaus (e.g., as in the adwords objective), then one can replace problem (23) with a problem over a finite horizon. Theorem 2 Suppose ψ(u) = c on [u0 , ∞). Then problem (23) is equivalent to the following problem, minimize β (24) Z u subject to y(s)ds − ψ ∗ (y(u)) ≤ βψ(u), ∀u ∈ [0, u0 ] 0 y(u0 ) = 0, y ∈ C[0, u0 ]. So for a function ψ with a plateau, one can discretize problem (24) to get a finite dimensional problem, minimize β subject to h (25) t X y[s] − ψ ∗ (y[t]) ≤ βψ(ht), ∀t ∈ [d] s=1 y[d] = 0, where h = u0 /d is the discretization step. Figure 2a shows the optimal smoothing for the piecewise linear function ψ(u) = min(.75, u, .5u + .25) by solving problem (25). We point out that the optimal smoothing for this function is not given by Nesterov’s smoothing (even though the optimal smoothing can be derived by Nesterov’s smoothing for a piecewise linear function with only two pieces, like the adwords cost function). Figure 2d shows the difference between the conjugate of the optimal smoothing function and ψ ∗ for the piecewise linear function, which we can see is not concave. Pm In cases where a bound umax on t=1 At Ft is known, we can restrict t to [0, umax ] and discretize problem (23) over this interval. However, the conclusion of Lemma 2 does not hold for a finite horizon and we need to impose additional linear constraints y[t] ≤ y[t − 1] to ensure the monotonicity of y. We find the optimal √ smoothing for two examples of this kind: ψ(u) = log(1 + u) over [0, 100] (Figure 2b), and ψ(u) = u over [0, 100] (Figure 2c). In Figure 2e, we show the competitive ratio achieved with the optimal smoothing of ψ(u) √ = log(1 + u) over [0, umax ] as a function of umax . Figure 2f depicts this quantity for ψ(u) = u. 3.3 Bounds for the sequential algorithm In this section we provide a lower-bound on the competitive ratio of the sequential algorithm (Algorithm 1). Based on this competitive ratio bound we modify Problem (23) for designing the smoothing function for the sequential algorithm. Theorem 3 Suppose ψS is differentiable on an open set containing K and satisfies Assumption 1. In addition suppose there exists c ∈ K is such that At Ft ≤K c for all t, then 1 Pseq ≥ D? , 1 − ᾱψ,ψS + κc,ψ,ψS where κ is given by κc,ψ,ψS = inf{r | hc, ∇ψS (0) − ∇ψS (u)i ≤ rψ(u), u ∈ K} (26) 12 0.8 0.6 5 25 4 20 3 15 2 10 0.4 0.2 ψS ψ 1 ψS ψ 0 0 0 0 0.5 1 1.5 2 0 20 40 u 60 80 0 100 0.9 0.2 0.1 0.4 0.6 0.8 1 0.85 0.8 0.9 0.85 0.8 0.75 0 1 100 0.95 0.7 -0.1 80 (c) 0.75 0 60 u comp. ratio 0.3 comp. ratio 0.95 0.2 40 (b) 0.4 0 20 u (a) ψS∗ (y) − ψ∗ (y) ψS ψ 5 500 0.7 1000 0 200 400 umax y (e) (d) 600 800 1000 umax (f) Figure 2: Optimal√smoothing for ψ(u) = min(.75, u, .5u+.25) (a), ψ(u) = log(1+u) over [0, 100] (b), and ψ(u) = u over [0, 100] (c). The competitive ratio √ achieved by the optimal smoothing as a function of umax for ψ(u) = log(1 + u) (e) and ψ(u) = u (f). ψS ∗ − ψ ∗ for the piecewise linear function (d). Proof: Since ψS satisfies Assumption 1, we have ŷt+1 ≤K ∗ ŷt . Therefore, we can write: m X hAt x̂t , ŷt − ŷt+1 i ≤ t=1 m X hc, ŷt − ŷt+1 i t=1 = hc, ŷ0 − ŷm+1 i (27) Now by combining 20 with 27, we get Dseq ≤ ψS m X ! At x̂t * + c, ∇ψS (0) − ∇ψS t=1 m X !+ At x̂t . t=1 The conclusion of the theorem follows from the definition of ᾱψ,ψS , κc,ψ,ψS and the fact that D̂ ≥ D? .  Based on the result of the previous theorem we can modify the optimization problem set up in Section 3.2 for separable functions on Rn+ to maximize the lower bound on the competitive ratio of the sequential algorithm. Note that in this case we have κc,ψ,ψS ≤ maxi κci ,ψi ,ψS i . Similar to the previous section to simplify the notation we drop the index i and assume ψ is a function of a scalar variable. The optimization problem for finding ψS that minimizes κc,ψ,ψS − ᾱψ,ψS is as follows: minimize β Z subject to (28) u y(s)ds + c(ψ 0 (0) − y(u)) − ψ ∗ (y(u)) ≤ βψ(u), ∀u ∈ [0, ∞) 0 y ∈ C[0, ∞). In the case of Adwords, the optimal solution is given by    u−1 1 β= y(u) = β 1 − exp , −1 , 1+c 1 − exp( c+1 ) +   −1 which gives a competitive ratio of 1 − exp c+1 . In Figure 3b we have plotted the competitive ratio achieved by solving problem 28 for ψ(u) = log det(1 + u) with umax = 100 as a function 13 0.8 Competitive ratio Competitive ratio 0.7 0.6 0.5 0.4 0.3 0.6 0.4 0.2 0 0 0.2 0.4 0.6 0.8 1 0 20 40 60 c c (a) (b) 80 100 Figure 3: The competitive ratio achieved by the optimal smoothing for the sequential algorithm as a function of c for ψ(u) = min(.75, u, .5u + .25) (a) and ψ(u) = log(1 + u) with umax = 100 (b). of c. Figure 3a shows the competitive ratio as a function of c for the piecewise linear function ψ(u) = min(.75, u, .5u + .25). 4 Discussion and related work We discuss results and papers from two communities, computer science theory and machine learning, related to this work. Online convex optimization. In [13], the authors proposed an optimal algorithm for adwords with differentiable concave returns (see examples in section 2). Here, “optimal” means that they construct an instance of the problem for which competitive ratio bound cannot be improved, hence showing the bound is tight. The algorithm is stated and analyzed for a twice differentiable, separable ψ(u). The assignment rule for primal variables in their proposed algorithm is explained as a continuous process. A closer look reveals that this algorithm falls in the framework of algorithm 2, with the only difference being that at each step, (x̃t , ỹt ) are chosen such that where vi : R+ 7→ R+ is an increasing differentiable function given as a solution of a nonlinear differential equation that involves ψi and may not necessarily have a closed form. The competitive ratio is also given based on the differential equation. They prove that this gives the optimal competitive ratio for the instances where ψ1 = ψ2 = . . . = ψn . Note that this is equivalent of setting ψS i (ui ) = ψ(vi (ui ))). Since vi is nondecreasing ψS i is a concave function. On the other hand, given a concave function ψS i (R+ ) ⊂ ψi (R+ ), we can set vi : R+ 7→ R+ as vi (u) = inf{z | ψi (z) ≥ ψS i (u)}. Our formulation in section 3.2 provides a constructive way of finding the optimal smoothing. It also applies to non-smooth ψ. Recently, authors in [4, 7, 11] have provided a primal-dual onlineP algorithm for the Pmdual probm lemP (2) that corresponds to the non-monotone primal objective ψ ( t=1 At xt ) = t=1 cTt xt + m G ( t=1 Bt xt ).The primal and dual updates in their algorithm are presented as a continuous update based on a differential equation. They assume that G∗ is differentiable and that ∇G∗ is monotone on Rn+ , i.e., If y ≥ y 0 , then ∇G∗ (y) ≤ ∇G∗ (y 0 ). In contrast, our assumption written in terms of G∗ for a differentiable function will become: If ∇G∗ (y) < ∇G∗ (y 0 ), then y ≥ y 0 , which is not equivalent to the assumption in [7]. When G is separable the two assumptions coincide and this algorithm is similar Pn Rtoy algorithm 2 applied to the smooth function GM whose conjugate is given by G∗M (y) = γ1 i=1 0 i G∗i 0 (z) log(1 − z/θ) dz. This smoothing coincides with Nesterov smoothing in the case of LP. Online learning. As mentioned before, the dual update in Algorithm 1 is the same as in Follow-theRegularized-Leader (FTRL) algorithm with −ψ ∗ as the regularization. This primal dual perspective has been used in [39] for design and analysis of online learning algorithms. In the online learning literature, the goal is to derive a bound on regret that optimally depends on the horizon, m. The goal in the current paper is to provide competitive ratio for the algorithm that depends on the function ψ. Regret provides a bound on the duality gap, and in order to get a competitive ratio the regularization 14 function should be crafted based on ψ. A general choice of regularization which yields an optimal regret bound in terms of m is not enough for a competitive ratio argument, therefore existing results in online learning do not address our aim. Acknowledgments The authors would like to thank James Saunderson, Ting Kei Pong, Palma London, and Amin Jalali for their helpful comments and discussions. References [1] Jacob Abernethy, Elad Hazan, and Alexander Rakhlin. Competing in the dark: An efficient algorithm for bandit linear optimization. In COLT, pages 263–274, 2008. [2] Shipra Agrawal and Nikhil R Devanur. Fast algorithms for online stochastic convex programming. arXiv preprint arXiv:1410.7596, 2014. [3] Shipra Agrawal, Zizhuo Wang, and Yinyu Ye. A dynamic near-optimal algorithm for online linear programming. arXiv preprint arXiv:0911.2974, 2009. [4] Yossi Azar, Ilan Reuven Cohen, and Debmalya Panigrahi. Online covering with convex objectives and applications. arXiv preprint arXiv:1412.3507, 2014. [5] Moshe Babaioff, Nicole Immorlica, David Kempe, and Robert Kleinberg. Online auctions and generalized secretary problems. SIGecom Exch., 7(2):7:1–7:11, June 2008. [6] Dimitri P Bertsekas. Necessary and sufficient conditions for a penalty method to be exact. Mathematical programming, 9(1):87–99, 1975. [7] Niv Buchbinder, Shahar Chen, Anupam Gupta, Viswanath Nagarajan, et al. Online packing and covering framework with convex objectives. arXiv preprint arXiv:1412.8347, 2014. [8] Niv Buchbinder, Kamal Jain, and Joseph Seffi Naor. Online primal-dual algorithms for maximizing ad-auctions revenue. In Algorithms–ESA 2007, pages 253–264. Springer, 2007. [9] Niv Buchbinder and Joseph Naor. Online primal-dual algorithms for covering and packing. Mathematics of Operations Research, 34(2):270–286, 2009. [10] James V Burke. An exact penalization viewpoint of constrained optimization. SIAM Journal on control and optimization, 29(4):968–998, 1991. [11] TH Chan, Zhiyi Huang, and Ning Kang. Online convex covering and packing problems. arXiv preprint arXiv:1502.01802, 2015. [12] Nikhil R. Devanur and Thomas P. Hayes. The adwords problem: Online keyword matching with budgeted bidders under random permutations. In Proceedings of the 10th ACM Conference on Electronic Commerce, EC ’09, pages 71–78, New York, NY, USA, 2009. ACM. [13] Nikhil R Devanur and Kamal Jain. Online matching with concave returns. In Proceedings of the forty-fourth annual ACM symposium on Theory of computing, pages 137–144. ACM, 2012. [14] Nikhil R Devanur, Kamal Jain, Balasubramanian Sivan, and Christopher A Wilkens. Near optimal online algorithms and fast approximation algorithms for resource allocation problems. In Proceedings of the 12th ACM conference on Electronic commerce, pages 29–38. ACM, 2011. [15] Reza Eghbali, Jon Swenson, and Maryam Fazel. Exponentiated subgradient algorithm for online optimization under the random permutation model. arXiv preprint arXiv:1410.7171, 2014. [16] Jon Feldman, Monika Henzinger, Nitish Korula, Vahab S. Mirrokni, and Cliff Stein. Online stochastic packing applied to display ad allocation. In Proceedings of the 18th Annual European Conference on Algorithms: Part I, ESA’10, pages 182–194, Berlin, Heidelberg, 2010. SpringerVerlag. [17] Jon Feldman, Aranyak Mehta, Vahab Mirrokni, and S Muthukrishnan. Online stochastic matching: Beating 1-1/e. In Foundations of Computer Science, 2009. FOCS’09. 50th Annual IEEE Symposium on, pages 117–126. IEEE, 2009. 15 [18] Marguerite Frank and Philip Wolfe. An algorithm for quadratic programming. Naval research logistics quarterly, 3(1-2):95–110, 1956. [19] Anupam Gupta and Marco Molinaro. How the experts algorithm can help solve lps online. arXiv preprint arXiv:1407.5298, 2014. [20] Patrick Jaillet and Xin Lu. Near-optimal online algorithms for dynamic resource allocation problems. arXiv preprint arXiv:1208.2596, 2012. [21] Patrick Jaillet and Xin Lu. Online stochastic matching: New algorithms with better bounds. Mathematics of Operations Research, 2013. [22] Siddharth Joshi and Stephen Boyd. Sensor Selection via Convex Optimization. IEEE TRANSACTIONS ON SIGNAL PROCESSING, 57(2), 2009. [23] Bala Kalyanasundaram and Kirk R Pruhs. An optimal deterministic algorithm for online b-matching. Theoretical Computer Science, 233(1):319–325, 2000. [24] Chinmay Karande, Aranyak Mehta, and Pushkar Tripathi. Online bipartite matching with unknown distributions. In Proceedings of the forty-third annual ACM symposium on Theory of computing, pages 587–596. ACM, 2011. [25] Richard M Karp, Umesh V Vazirani, and Vijay V Vazirani. An optimal algorithm for on-line bipartite matching. In Proceedings of the twenty-second annual ACM symposium on Theory of computing, pages 352–358. ACM, 1990. [26] Thomas Kesselheim, Klaus Radke, Andreas Tönnis, and Berthold Vöcking. Primal beats dual on online packing lps in the random-order model. In Proceedings of the 46th Annual ACM Symposium on Theory of Computing, STOC ’14, pages 303–312, New York, NY, USA, 2014. ACM. [27] Robert Kleinberg. A multiple-choice secretary algorithm with applications to online auctions. In Proceedings of the sixteenth annual ACM-SIAM symposium on Discrete algorithms, pages 630–631. Society for Industrial and Applied Mathematics, 2005. [28] GG Lorentz. An inequality for rearrangements. The American Mathematical Monthly, 60(3):176– 179, 1953. [29] Mohammad Mahdian and Qiqi Yan. Online bipartite matching with random arrivals: an approach based on strongly factor-revealing lps. In Proceedings of the forty-third annual ACM symposium on Theory of computing, pages 597–606. ACM, 2011. [30] Vahideh H Manshadi, Shayan Oveis Gharan, and Amin Saberi. Online stochastic matching: Online actions based on offline statistics. Mathematics of Operations Research, 37(4):559–573, 2012. [31] Aranyak Mehta, Amin Saberi, Umesh Vazirani, and Vijay Vazirani. Adwords and generalized online matching. Journal of the ACM (JACM), 54(5):22, 2007. [32] Yu Nesterov. Smooth minimization of non-smooth functions. Mathematical programming, 103(1):127–152, 2005. [33] Yurii Nesterov. Introductory lectures on convex optimization, volume 87. Springer Science & Business Media, 2004. [34] Yurii Nesterov. Primal-dual subgradient methods for convex problems. Mathematical programming, 120(1):221–259, 2009. [35] Friedrich Pukelsheim. Optimal design of experiments, volume 50. siam, 1993. [36] R Tyrrell Rockafellar, Roger J-B Wets, and Maria Wets. Variational analysis, volume 317. Springer, 1998. [37] Shai Shalev-Shwartz. Online learning and online convex optimization. Foundations and Trends in Machine Learning, 4(2):107–194, 2011. [38] Shai Shalev-Shwartz and Yoram Singer. Online learning: Theory, algorithms, and applications. 2007. [39] Shai Shalev-Shwartz and Yoram Singer. A primal-dual perspective of online learning algorithms. Machine Learning, 69(2-3):115–142, 2007. 16 [40] Manohar Shamaiah, Siddhartha Banerjee, and Haris Vikalo. Greedy sensor selection: Leveraging submodularity. In 49th IEEE Conference on Decision and Control (CDC), pages 2572–2577. IEEE, dec 2010. [41] Michael M Zavlanos and George J Pappas. Distributed connectivity control of mobile networks. Robotics, IEEE Transactions on, 24(6):1416–1428, 2008. Appendix A Proofs Proof of Lemma 1: Using the definition of Dsim , we can write: Dsim = m X σt (ATt ỹt ) − ψ ∗ (ỹm+1 ) t=1 = m X hAt x̃t , ỹt i − ψ ∗ (ỹm+1 ) t=1 ≤ m X t t−1 X X (ψ( As x̃s ) − ψ( As x̃s )) − ψ ∗ (ỹm+1 ) t=1 = ψ( m X s=1 s=1 As x̃s ) − ψ(0) − ψ ∗ (ỹm+1 ), s=1 where in the inequality follows from concavity of ψ, and the last line results from the sum telescoping. Similarly, we can bound Dseq : Dseq = = m X t=1 m X σt (ATt ŷt ) − ψ ∗ (ŷm+1 ) (29) hAt x̂t , ŷt i − ψ ∗ (ŷm+1 ) t=1 = m X hAt x̂t , −ŷt+1 + ŷt i + t=1 ≤ m X t=1 = m X t=1 m X hAt x̂t , ŷt+1 i − ψ ∗ (ŷm+1 ) t=1 hAt x̂t , −ŷt+1 + ŷt i + m X t t−1 X X (ψ( As x̂s ) − ψ( As x̂s )) − ψ ∗ (ŷm+1 ) t=1 s=1 s=1 m X hAt x̂t , −ŷt+1 + ŷt i + ψ( As x̂s ) − ψ(0) − ψ ∗ (ŷm+1 ). s=1 To provide intuition about the above inequalities Pm we have plotted the derivative of a concave function ψ defined on R+ in Figure 4. The quantity t=1 hAt x̃t , ỹt i is a right Riemann Pm sum approximation of the integral of ψ 0 and lower bounds the integral (Figure 4a). The quantity t=1 hAt x̂t , ŷt i is a left Riemann sum approximation of the integral and upper bounds the integral (Figure 4b). The area of hatched rectangle bounds the error of the left Riemann sum and is equal to hAt x̂t , −ŷt+1 + ŷt i. When ψ is differentiable with Lipschitz gradient, we can use the following inequality that is equivalent to Lipschitz continuity of the gradient: ψ(u0 ) ≥ ψ(u) + h∇ψ(u), u0 − ui − 17 1 2 ku − u0 k 2µ u, u0 ∈ K ψ 0 (u) ψ 0 (u) ŷ1 ỹ1 ŷt ỹ2 ŷ3 ỹ3 A1 x̃1 A2 x̃2 u A3 x̃3 A1 x̂1 (a) Figure 4: Interpretation of At x̂t At x̂3 u (b) Pm t=1 σt (ATt ỹt ) and Pm t=1 σt (ATt ŷt ) as Riemann sums. (see, for example, [33, section 2.1.1]) to get Dseq = = m X t=1 m X σt (ATt ŷt ) − ψ ∗ (ŷm+1 ) (30) hAt x̂t , ŷt i − ψ ∗ (ŷm+1 ) t=1 ≤ t−1 t m m X X X X 1 2 As x̂s )) − ψ ∗ (ŷm+1 ) As x̂s ) − ψ( (ψ( kAt x̂t k + 2µ s=1 s=1 t=1 t=1 = m m X X 1 2 kAt x̂t k + ψ( As x̂s ) − ψ(0) − ψ ∗ (ŷm+1 ). 2µ t=1 s=1  Proof of Lemma 2: Let (y, β) be a feasible solution for problem (23). Note that y ≥ 0 since dom ψ ∗ ⊂ R+ by the fact that ψ is non-decreasing. Let ȳ(u) = inf s≤u y(s). Note that ȳ is continuous. Define Ru Ru y(s) ds − ψ ∗ (y(u)) ȳ(s) ds − ψ ∗ (ȳ(u)) s=0 β(u) = , β̄(u) = s=0 , ψ(u) ψ(u) with the definition modified with the right limit at u = 0. For any u such that ȳ(u) = y(u), we have: Ru Ru y(s) ds − ψ ∗ (y(u)) ȳ(s) ds − ψ ∗ (ȳ(u)) s=0 β(u) = ≥ s=0 = β̄(u). ψ(u) ψ(u) Now, we consider the set {u | ȳ(u) 6= y(u)}. By the definition of ȳ, we have ȳ(0) = y(0). Since both functions are continuous, the set {u | ȳ(u) 6= y(u)} is an open subset of R and hence can be written as a countable union of disjoint open intervals. Specifically, we can define the end points of the intervals as: a0 = b0 = 0, ai = inf{u > bi−1 | y(u) > ȳ(u)}, ∀i ∈ {1, 2, . . .} bi = inf{u > ai | y(u) = ȳ(u)}, ∀i ∈ {1, 2, . . .}. S then {u | ȳ(u) 6= y(u)} = i∈{1,2,...} (ai , bi ).(See Figure 5) For any i ∈ {1, 2, . . .}, we show that β(u) ≥ β̄(u) on (ai , bi ). If ai = ∞, then bi = ∞, so we assume that ai < ∞. By the definition of ai and bi , ȳ(u) is constant on (ai , bi ). Also, we have y(ai ) = ȳ(ai ). Similarly, we have y(bi ) = ȳ(bi ) whenever bi < ∞. 18 a1 b1 a2 b2 u Figure 5: An example of y(u) (solid blue) and ȳ(u) (dashed red). Since ȳ(u) ≤ y(u) for all u and y(ai ) = ȳ(ai ), we have β(ai ) ≥ β̄(ai ). (31) If bi < ∞, similarly by the fact that y(bi ) = ȳ(ai ) = ȳ(bi ), we have β(bi ) ≥ β̄(bi ). (32) Now we consider the case where bi = ∞. In this case we have ȳ(u) = ȳ(ai ) on (ai , ∞). We consider two cases based on the asymptotic behavior of ψ. If limu→∞ ψ(u) = +∞ (ψ is unbounded), then we have Ru Ru y(s) ds ȳ(ai ) ds s=0 lim sup β(u) = lim sup ≥ lim sup s=0 = lim β̄(u). (33) u→∞ ψ(u) ψ(u) u→∞ u→∞ u→∞ Here we used the fact that −ψ ∗ (y(u)) is bounded. This follows from the fact ψ ∗ is monotone thus: −ψ ∗ (y(u)) ≤ −ψ ∗ (ȳ(ai )), and −ψ ∗ (ȳ(ai )) < ∞ because if −ψ ∗ (ȳ(ai )) = ∞, then β(ai ) ≥ β̄(ai ) = ∞ which contradicts the feasibility of (y, β). Now consider the case when limu→∞ ψ(u) = M for some positive constant M . In this case, −ψ ∗ ≤ M . We claim that y(ai ) = 0 and lim inf t→∞ y(u) = 0. Suppose lim inf u→∞ y(u) > 0, then lim supu→∞ β(u) = ∞ since the numerator in the definition of β tends to infinity while the denominator is bounded. But this contradicts feasibility of (y, β). On the other hand, by the definition of ai and bi we should have y(ai ) = ȳ(ai ) ≤ lim inf u→∞ y(u). Combining this with the fact that ȳ(ai ) ∈ dom ψ ∗ ⊂ R+ , we conclude that y(ai ) = 0. Using that y(ai ) = 0 and lim inf u→∞ y(u) = 0, we get: Ru y(s) ds − ψ ∗ (y(u)) lim sup β(u) = lim sup s=0 ψ(u) u→∞ u→∞ Ru y(s) ds − ψ ∗ (0) ≥ lim s=0 u→∞ M R ai ȳ(s) ds − ψ ∗ (0) ≥ s=0 = lim β̄(u), (34) u→∞ M where in the last inequality we used the fact that ȳ(u) = 0 for u ≥ ai . Let ψ 0 be the right derivative of ψ. Since ψ is concave, ψ 0 is non-increasing. Therefore, the interval (ai , bi ) can be written as (ai , u0 ] ∪ [u0 , bi ) such that ψ 0 (u) ≥ ȳ(ai ) on (ai , u0 ] and ψ 0 (u) ≤ ȳ(ai ) on [u0 , bi ). Since ψ 0 (u) ≥ ȳ(ai ) on (ai , u0 ] we have: Z u Z u ȳ(s) ds = ȳ(ai ) ds ai ai Z u ≤ ψ 0 (s) ds = ψ(u) − ψ(ai ), ai 19 for all u ∈ (ai , u0 ]. This yields: R ai β̄(ai ) = ≥ ȳ(s) ds − ψ ∗ (ȳ(ai )) ψ(u) Ru R ai ȳ(s) ds + ai ȳ(s) ds − ψ ∗ (ȳ(ai )) s=0 s=0 ψ(ai ) + ψ(u) − ψ(ai ) = β̄(u). for all u ∈ (ai , u0 ]. Here we used the fact that if c1 ≥ c2 > 0 and d2 ≥ d1 ≥ 0, then c1 c1 + d1 ≥ . c2 c2 + d2 Similarly, we have β̄(bi ) ≥ β̄(u) for any u ∈ [u0 , bi ). Combining this with (31),(32),(33), and (34), we get: sup β̄(u) = max(β̄(ai ), β̄(bi )) ai ≤u≤bi ≤ max(β(ai ), β(bi )) ≤ sup β(u). ai ≤u≤bi We conclude that β̄(u) ≤ β(u) for all t ≥ 0 hence (ȳ, β) is a feasible solution for the problem. Proof of Theorem 2: Let (y, β) be a feasible solution for problem (23). By Lemma 2, we can assume that y is non-increasing. First, note that y ≥ 0 since dom ψ ∗ = [0, ∞). Define ȳ(u) = y(u) for u ≤ u0 and ȳ(u) = 0 for u > u0 . We show that (ȳ, β) is also a feasible solution for (23) modulo the continuity condition. Define Ru Ru y(s) ds − ψ ∗ (y(u)) ȳ(s) ds − ψ ∗ (ȳ(u)) s=0 β(u) = , β̄(u) = s=0 . ψ(u) ψ(u) By the definition of ȳ, for all u, we have: Z u Z y(s) ds ≥ 0 u ȳ(s) ds, (35) 0 and β(u) = β̄(u) for u ∈ [0, u0 ]. Since y(u) is non-increasing and y(u) ≥ 0, limu→∞ y(u) exists. We claim that limu→∞ y(u) = 0. To see this note that if limu→∞ y(u) > 0, then Z u y(s) ds = ∞, lim u→∞ s=0 which contradicts the fact that β(u) ≤ β for all u. For all u ≥ u0 , now we have: Ru limu→∞ s=0 y(s) ds − ψ ∗ (0) sup β(u) ≥ lim β(u) = u→∞ ψ(u0 ) u≥u0 Ru limu→∞ s=0 ȳ(s) ds − ψ ∗ (0) = β̄(u0 ), ≥ ψ(u0 ) where the equality follows from the fact that limu→∞ y(u) = 0, and in the last inequality, we used (35). Since ȳ(u) = 0 for u > u0 , β(u) is constant on [u0 ∞). Therefore, supu≥u0 β̄(u) = β̄(u0 ). Combining this with the previous inequality we get: sup β(u) ≥ sup β̄(u0 ). u≥u0 u≥u0 Therefore, we conclude that β̄(u) ≤ β for all u. Thus (ȳ, β) is also a feasible solution for (23) modulo the continuity condition. Note that ȳ(u) may not be continuous at u0 . However, we can find a sequence of continuous functions z (j) that converge pointwise to y and z (i) (u) = 0 for all i and u ≥ u0 . To do so we consider a sequence of real number i → 0. We define z (i) (u) = ȳ(u) for u ∈ [0, u0 − i ) ∪ [u0 , ∞). On [u0 − i , u0 ] we define z (i) (u) to be a linear function that take values y(u0 − ) and 0 on the endpoints. Define R u (i) z (s) ds − ψ ∗ (z (i) (u)) βz(i) = sup s=0 . ψ(u) u>0 20 By upper semi-continuity of ψ ∗ , βz(i) converges to β̄. Let β ∗ be the optimal solution for problem (23). By the definition, there exits a feasible sequence (y (j) , β (j) ) such that β (j) converges to β ∗ . Let ȳ (j) (u) = y (j) (u) for t ≤ u0 and ȳ (j) (u) = 0 for t > u0 . Note that ȳ (j) (u) may not be continuous at u0 . However, we can find a sequence of continuous functions (z (ji) , βz(ji) ) as in above. Now βz(jj) converges to β ∗ .  Distance from lp norm ball B In this section we prove that the function: G(u) = −d1 (u, Bp ) satisfies Assumption 1 and find a lower bound on ᾱψ when ψ is given by 16 with G(u) = −ld1 (u, Bp ). For any u ∈ Rn+ , there exists ū ∈ Bp such that d1 (u, Bp ) = ku − ūk1 . the subdifferential of distance function is6 : ∂d1 (u, Bp ) = ∂ ku − ūk1 ∩ NBp (ū), where NBp (u) = {ξ | hξ, v − ui ≥ 0, ∀v ∈ Bp } is the normal cone of Bp at u. In fact d1 (u, Bp ) = ku − ūk1 if and only if ∂ ku − ūk1 ∩ NBp (ū) 6= ∅. When u ∈ intBp , ū = u and ∂d1 (u, Bp ) = {0}. In order to find ∂d1 (u, Bp ) when u ∈ / intBp , we first find ū in this case. For any r ≥ 0, define u ∧ r ∈ Rn+ to be: (u ∧ r)i = min(ui , r) ∀i. Note that ku ∧ 0kp = 0 and ku ∧ (maxi ui )kp = kukp ≥ 1. Since ku ∧ rkp is a continuous function of r, by the intermediate value theorem, there exists ru ∈ (0, maxi ui ] such that ku ∧ ru kp = 1. Now ū = u ∧ r. To see this note that:  ∂ ku − ūk1 ∩ NBp (ū) =  ∂ ku − ūk1 ∩ NBp (ū) = 1 ◦(p−1)  (u ∧ ru ) for ru < max ui ; i rup−1  z ◦(p−1) |0≤z≤1 for ru = max ui ; p−1 (u ∧ ru ) i ru (36) (37) where ◦(p−1) denotes element-wise exponentiation. Now if u0 ≤ u, then ru ≤ ru0 since ku ∧ rkp ≥ ku0 ∧ rkp for all r. Thus by (36) and (37), there exists y ∈ ∂d1 (u, Bp ) such that y ≥ ∂d1 (u0 , Bp ). Now bound αψ when ψ is given by 16 with G(u) = −ld1 (u, Bp ). Let (v, u) = Pm we canPlower m ( t=1 ct x̃t , t=1 Bt x̃t ). Note that by (36) when y ∈ ∂ψ(u) with u ∈ / Bp , then mini yi = −l. Now by the definition of l in (17) and the explanation that followed it, we must have u ∈ Bp . If u ∈ intBp , then G is differentiable at u and ∇G(u) = 0 which yields αψ (v, u) = 0. Now suppose u ∈ bdBp . In this case we have: − kykq − kykq G∗ (y) Pm αψ (v, u) = min = min Pn ≥ min . y∈∂G(u) y∈∂G(u) ψ ( y∈∂G(u) θ1T u t=1 At xt ) t=1 ct x̄t Recall that θ = mint minx∈Ft cT t x . 1T Bt x By (37), we have: min − kykq = − y∈∂G(u) l . (maxi ui )p−1 Therefore, αψ (v, u) ≥ − l 1 . θ (1T u)(maxi ui )p−1 As u varies on the bdBp , the right hand side is lower bounded by 6 For convex function we use ∂ to denote subdifferential. 21 −l θ . This yields αψ ≥ − θl . C Derivation of lower bounds on αψ,ψφ We first derive a general inequality which will be specialized to different examples for bounding αψ,ψφ . Let K = K1 × K2 , with K1 and K2 two proper cones. Suppose ψ(v, u) = H(v) + G(u), where H : K1 7→ R is a non-decreasing, and G : K P2m7→ R is non-increasing and l Lipschitz continuous. We assume H(v) ≥ θu for all (v, u) ∈ t=1 At Ft . Note that ψ ∗ (z, y) = H ∗ (z) + G∗ (y). We set      m X θ (e − 1) 1 ∗ yi − log 1 − yi − (1 + γ)yi , φ (y) = γ (e − 1) θ i=1 Pm where γ = log(1 + l(e−1) ). We let ψφ(v, u) = H(v) + Gφ(u). Let (v, u) ∈ t=1 At x̃t . Since θ ψφ(0) = 0, and the simultaneous algorithm does not decrease the objective, H(v) + Gφ(u) ≥ ψφ(0) = 0. (38) Let (z, y) ∈ ∂ψφ(u), then we have: ˜ i G∗ (y) ui = ∇i φ∗ (y) + ∇ 1 (e − 1)yi ˜ i G∗ (y), = log(1 − )−1+∇ γ θ (39) ˜ ∗ (y) ∈ ∂G∗ (y). Using the previous identity, we can derive the following upper bound for some ∇G for Gφ(u): Gφ(u) = hy, ui − G∗ (y) − φ∗ (y) D E ˜ ∗ (y) − G∗ (y) + hy, ∇φ∗ (y)i − φ∗ (y) = y, ∇G ˜ ∗ (y)) + φ(∇φ∗ (y)) = G(∇G m X θ ˜ ∗ (y)) + ˜ i G∗ (y) + 1) + 1 1T y = G(∇G (ui − ∇ (e − 1) i=1 γ m X 1 θ ˜ i G∗ (y)) + 1 1T y. H(v) + (1 − ∇ (40) (e − 1) (e − 1) i=1 γ Pm Now we specialize the bound to the case where G : Rm + 7→ R and G(u) = −l i=1 (ui − 1)+ . We ∗ ˜ lower bound αψ,ψφ (v, u) when u ≤ 1. In that case, (39) is satisfied with ∇i G (y) = 1. Thus from (40) simplifies to: 1 1 Gφ(u) ≤ H(v) + 1T y. (41) (e − 1) γ ˜ ∗ (y)) + ≤ G(∇G Combining this with (38), we get: H(v) + G(u) = H(v) ≥ (1/e − 1) T 1 y. γ (42) In the view of definition of αψ,ψφ , by using (41) and the fact that G∗ (y) = 1T y, we derive the following inequality: 1 1 H(v) + Gφ(u) − G∗ (y) ≤ (1 + )H(v) + ( − 1)1T y. (43) e−1 γ Combining (42) and (43), we get the following lower bound on αψ,ψφ : αψ,ψφ (v, u) ≥ 1 − (1 + 1 )γ + ᾱH . e−1 (44) Online LP: In this problem, ψ m X t=1 ! At x t = m X cTt xt t=1 +G m X t=1 22 ! Bt xt , (45) Pn with G(u) = −l i=1 (ui − 1)+ . In this problem H(v) = v is the identity function. Pm Pm Let (ṽ, ũ) = ( t=1 cTt x̃t , t=1 Bt x̃t ). If ỹt,i ≤ −l for some t and i, then by the definition of l in (17) and the explanation that followed it, we have (Bt x̃t )i = 0. On the other hand, we have ∇i Gφ(u) ≤ −l when ũi ≥ 1. Therefore, we conclude that ũ ≤ 1. Also, by the definition of θ, we have H(ṽ) ≥ θ1T ũ. Since H is a linear function, αH = 0. Thus (44) yields αψ,ψφ (ṽ, ũ) ≥ 1 − (1 + 1 )γ. e−1 (46) Online graph formation and online Experiment design : In this problem, H(U ) = log det(U + A0 ) − log det(A0 ), G(u) = −l(b − u)+ and At x = (aTt at x, x). We can use the following identity for determinant of rank one update of a matrix det(A + vv T ) = det(A)(1 + v T A−1 v T ), to derive a lower bound on log det(U + A0 ) − log det(A0 ), H(U ) = log det(U + A0 ) − log det(A0 ) = m X log det(A0 + s=1 = m X s X at aTt x̃t ) − log det(A0 + t=1 s−1 X at aTt x̃t ) t=1 log(1 + aTt Ỹt at x̃t ) t=1 ≥ log(1 + m 1 X ) x̃t = θu. n t=1 Pm Pm Let (Ũ , ũ) = ( t=1 at aTt x̃t , t=1 x̃t ). If ỹt ≤ −l for some t, then by the definition of l in (22), aTt Ỹt at + ỹt < 0 which results in x̃t = 0. Given the fact that Gφ0 (u) = −l when u ≥ b, we conclude that ũ ≤ b. Therefore, by (44), we get αψ,ψφ (Ũ , ũ) ≥ −(1 + 23 1 )γ. e−1 

8cs.DS

arXiv:1605.03017v3 [math.ST] 27 Jul 2016 Marginalized Particle Filtering and Related Filtering Techniques as Message Passing July 29, 2016 Abstract In this manuscript a factor graph approach is employed to investigate the recursive filtering problem for mixed linear/nonlinear state-space models. Our approach allows us to show that: a) the factor graph characterizing the considered filtering problem is not cycle free; b) in the case of conditionally linear Gaussian systems, applying the sum-product rule, together with different scheduling procedures for message passing, to this graph results in both known and novel filtering techniques. In particular, it is proved that, on the one hand, adopting a specific message scheduling for forward only message passing leads to marginalized particle filtering in a natural fashion; on the other hand, if iterative strategies for message passing are employed, a novel filtering method, dubbed turbo filter for its conceptual resemblance to the turbo decoding methods devised for concatenated channel codes, can be developed. Giorgio M. Vitetta, Emilio Sirignano, Francesco Montorsi and Matteo Sola University of Modena and Reggio Emilia Department of Engineering ”Enzo Ferrari” Via P. Vivarelli 10/1, 41125 Modena - Italy email: giorgio.vitetta@unimore.it, emilio.sirignano@unimore.it, francesco.montorsi@gmail.com, sola.matteo87@gmail.com Keywords: State Space Representation, Hidden Markov Model, Particle Filter, Belief Propagation, Turbo Processing. 1 G. M. Vitetta et al., Marginalized Particle Filtering . . . 1 2 Introduction The nonlinear filtering problem consists of inferring the posterior distribution of the hidden state of a nonlinear dynamic system from a set of past and present measurements [1]. It is well known that, if a nonlinear dynamic system can be described by a state-space model (SSM), a general sequential procedure, based on the Bayes’ rule and known as Bayesian filtering, can be easily derived for recursively computing the posterior distribution of the system current state [1]. Unluckily, Bayesian filtering is analytically tractable in few cases for the following two reasons [2]: a) one of the two steps it consists of requires multidimensional integration which, in most cases, does not admit a closed form solution; b) the functional form of the required posterior distribution may not be preserved over successive recursions. For this reason, sequential techniques employed in practice are based on various analytical approximations and, consequently, generate a functional approximation of the desired distribution. Such techniques are commonly divided into local and global methods on the basis of the way posterior distributions are approximated [3, 4, 5]. Local methods, like extended Kalman filtering [6] and unscented filtering [7], are computationally efficient, but may suffer from the problem of error accumulation over time. On the contrary, global methods, like sequential Monte Carlo methods [8, 9] (also known as particle filtering, PF, methods [10, 11, 12]) and point mass filtering [5, 13] may achieve high accuracy at the price, however, of an unmanageable complexity and numerical problems in the presence of a large dimension of system state [14]. These considerations have motivated various research activities focused on the development of novel Bayesian filters able to achieve high accuracy under given computational constraints. Significant results in this research area concern the use of the new representations for complex distributions, like belief condensation filtering [3], and the development of novel filtering techniques combining local and global methods, like marginalized particle filtering (MPF) [15, 16], and other methods originating from it [4, 17, 18]. Note that the last class of methods applies to mixed nonlinear/nonlinear models [19], that is to models whose state can be partitioned in a conditionally linear portion (usually called linear state variable) and in a nonlinear portion (representing the remaining part of system state and called nonlinear state variable). This partitioning of system state allows to combine a global method (e.g., particle filtering) operating on the nonlinear state variable with a local technique (e.g., Kalman filtering) involving the linear state variable only. In this manuscript the factor graph (FG) approach illustrated by Loeliger et G. M. Vitetta et al., Marginalized Particle Filtering . . . 3 al. in [20] is employed to revisit the problem of recursive Bayesian filtering for mixed linear/nonlinear models from a perspective substantially different from that adopted in MPF [15]. This allows us to shed new light on the problem of filtering for mixed linear/nonlinear models, providing a new interpretation of MPF and paving the way for the development of new filtering techniques. In particular, based on this approach, we are able to show that: a) the considered filtering problem can be formulated as a message massing problem over a specific FG, which, unluckily, is not cycle free; b) in the case of a conditionally linear Gaussian (CLG) SSM [19], MPF results from the application of the sum-product algorithm (SPA) [20, 21], together with a specific scheduling procedure for forward only message passing, to this graph; c) our graphical representation leads, in a natural fashion, to the development of novel filtering methods simplifying and/or generalising it. As far as the last point is concerned, in our work specific attention is paid to the development of a novel iterative filtering technique that exploits the exchange of probabilistic (i.e., soft) messages to progressively refine the posteriors of the linear and nonlinear state variables within each recursion and is dubbed turbo filtering (TF) for its conceptual resemblance to the iterative (i.e., turbo) decoding of concatenated channel codes. It is important to point that our approach has been inspired by various ideas and results already available in the technical literature concerning different research areas; here, we limit to mention the following relevant facts: • A mixed linear/nonlinear Markov system can be represented as the concatenation of two interacting subsystems, one governed by linear dynamics, the other one accounting for a nonlinear behavior; conceptually related (finite state) Markov models can be found in data communications and, in particular, in concatenated channel coding (e.g., turbo coding [22]) and in coded transmissions over inter-symbol interference channels for which turbo decoding methods [22, 23] and turbo equalization techniques [24] have been developed, respectively1 . • Factor graphs play an essential role in the derivation and interpretation of turbo decoding and equalization [20, 26] (for instance, turbo decoding techniques emerge in a natural fashion from graphical models of codes [27]). • Both Kalman filtering and particle methods can be viewed as message 1 Note that these classes of algorithms can be seen as specific applications of the so called turbo principle [25], [35, Par. 10.5.1] 4 G. M. Vitetta et al., Marginalized Particle Filtering . . . passing procedures on factor graphs, as shown in [20, 21] and in [28], respectively. • Various methods to progressively refine distributional approximations through multiple iterations have been developed in the field of Bayesian inference on dynamic systems (even if implementations substantially different from that we devise have been proposed), and, in particular, in expectation propagation in Bayesian networks [29, 30] and in variational Bayesian filtering [4]. Consequently, various links to previous work on Bayesian inference on graphical models and variational Bayes methods [31] can be also established. The remaining part of this manuscript is organized as follows. The mathematical model of the considered class of mixed linear/nonlinear systems is illustrated in Section 2, whereas a representation of the filtering problem for these systems through a proper FG is provided in Section 3. Then, it is shown that applying the SPA and proper message scheduling strategies to this FG leads to MPF in Section 4. This approach paves the way, in a natural fashion, for the development of simplifications and generalizations of MPF, that are devised in Sections 5 and 6, respectively. The novel filtering methods proposed in this manuscript are compared, in terms of accuracy and computational effort, with MPF in Section 7. Finally, some conclusions are offered in Section 8. Notations: The probability density function (pdf) of a random vector R evaluated at point r is denoted f (r); N (r; ηr , Cr ) represents the pdf of a Gaussian random vector R characterized by the mean ηr and covariance matrix Cr evaluated at point r; the precision (or weight ) matrix associated with the covariance matrix Cr is denoted Wr , whereas the transformed mean vector Wr ηr is denoted wr ; xi denotes the i-th element of the vector x. 2 System Model In the following we focus on a discrete-time mixed linear/nonlinear SSM [15], whose hidden state in the l-th interval is represented by a D-dimensional real vector xl , [x0,l , x1,l , ..., xD−1,l ]T . We assume that this vector can be partitioned as  T   T T (L) (N ) xl = xl , xl , (1) (L) (L) (L) (L) (N ) (N ) (N ) (L) where xl , [x0,l , x1,l , ..., xDL −1,l ]T (xl , [x0,l , x1,l , ..., xDN −1,l ]T ) is the so called linear (nonlinear ) component of xl (1), with DL < D (DN = D − DL ). 5 G. M. Vitetta et al., Marginalized Particle Filtering . . . (L) This partitioning of xl is accomplished as follows. First, xl is identified as that portion of xl (1) characterized by the following two properties: 1. Conditionally linear dynamics - This means that its update equation, con(N ) ditioned on xl , is linear, so that     (L) (L) (N ) (L) (L) (N ) (L) xl+1 = Al xl xl + fl xl + wl , (2) (L) (L) (N ) where fl (x) is a time-varying DL -dimensional real function, Al (xl ) (L) is a time-varying DL × DL real matrix and wl is the l-th element of (L) the process noise sequence {wk }, which consists of DL - dimensional independent and identically distributed (iid) noise vectors. 2. Conditionally linear (or almost linear ) dependence of all the available measurements on it - In other words, these quantities, conditioned on (N ) (L) xl , exhibit a linear dependence on xl (additional details about this feature are provided below). (N ) Then, xl is generated by putting together all the components of xl that (L) do not belong to xl . For this reason, generally speaking, this vector is characterized by at least one of the following two properties: a) Nonlinear dynamics - The update equation     (N ) (N ) (N ) (N ) (N ) (L) (N ) xl+1 = fl xl + Al xl xl + wl (3) is assumed in the following for the nonlinear component of system state, where (N ) (N ) (N ) Al (xl ) is a time-varying DN × DL real matrix, fl (x) is a time-varying (N ) DN -dimensional real function and wl is the l-th element of the process noise (N ) sequence {wk }, which consists of DN -dimensional iid noise vectors and is (L) statistically independent of {wk }. b) A nonlinear dependence of all the available measurements on it (further details are provided below). In the following Section we focus on the so-called filtering problem, which concerns the evaluation of the posterior pdf f (xt |y1:t ) at an instant t > 1, given a) the initial pdf f (x1 ) and b) the t · P -dimensional measurement vector T  y1:t = y1T , y2T , ..., ytT , (4) where yl , [y0,l , y1,l , ..., yP −1,l ]T denotes the P -dimensional real vector collecting all the noisy measurements available at time l. As already mentioned G. M. Vitetta et al., Marginalized Particle Filtering . . . 6 above, the measurement vector yl exhibits a linear (nonlinear ) dependence on (L) (N ) xl (xl ), so that the model [18]     (N ) (L) (N ) xl + el (5) + Bl xl yl = hl xl (N ) (N ) can be adopted, where Bl (xl ) is a time-varying P × DL real matrix, hl (xl ) is a time-varying P -dimensional real function and el the l-th element of the measurement noise sequence {ek } consisting of P -dimensional iid noise vectors (N ) (L) and independent of both {wk } and {wk }. 3 Representation of the Filtering Problem via Factor Graphs Generally speaking, the filtering problem for a SSM described by the Markov model f (xl+1 |xl ) and the observation model f (yl |xl ) for any l concerns the computation of the posterior pdf f (xt |y1:t ) for t ≥ 1 by means of a recursive procedure [1]. It is well known that, if the pdf f (x1 ) is known, a general Bayesian recursive procedure, consisting of a measurement update step followed by a time update step, can be employed. In practice, in the first step of the l-th recursion (with l = 1, 2, ..., t) the conditional pdf
f (xl |y1:l ) = f xl y1:(l−1) f (yl |xl ) 1
f yl y1:(l−1) (6) is computed on the basis of pdf f (xl |y1:(l−1) ) (evaluated in the last step of the previous recursion2 ), the present measurement vector yl and the pdf Z

f yl y1:(l−1) = f (yl |xl ) f xl y1:(l−1) dxl . (7) In the second step f (xl |y1:l ) (6) is exploited to compute the pdf Z f (xl+1 |y1:l ) = f (xl+1 |xl ) f (xl |y1:l ) dxl , (8) which represents a prediction about the future state xl+1 . It is important to point out that: 1) the term 1/f (yl |y1:(l−1) ) appearing in the right hand side (RHS) of (6) represents a normalization factor ; 2) both (7) and (8) require integration with respect to xl and this may represent a formidable task when the dimensionality of xl is large and/or the pdfs appearing in the integrands are 2 Note that in the first recursion (i.e., for l = 1) f (x |y l 1:(l−1) ) = f (x1 |y1:0 ) = f (x1 ) and f (yl |y1:(l−1) ) = f (y1 |y1:0 ) = f (y1 ), so that f (x1 |y1 ) = f (x1 )f (y1 |x1 )/f (y1 ). G. M. Vitetta et al., Marginalized Particle Filtering . . . 7 not Gaussian; 3) this recursive procedure lends itself to be efficiently represented by a message passing algorithm over a proper FG3 [28], in which each factor of a product of functions is represented by a distinct node (a rectangle in our diagrams), whereas each variable is associated with a specific (and, usually, unoriented) edge or half edge. As far as the last point is concerned, we also note that the derivation of this FG relies on the fact that the a posteriori pdf f (xt |y1:t ) has the same FG as the joint pdf f (xt , y1:t ) (see [20, Sec. II, p. 1297]) and the last pdf can be computed recursively through a procedure similar to that illustrated above, but in which the measurement update (6) and the time update (8) are replaced by and
f (xl , y1:l ) = f xl , y1:(l−1) f (yl |xl ) , f (xl+1 , y1:l ) = Z f (xl+1 |xl ) f (xl , y1:l ) dxl , (9) (10) respectively, so that the evaluation of the above mentioned normalization factor is no more required. In fact, eqs. (9) and (10) involve only products of pdfs and a sum (i.e., integration) of products, so that they can be represented by means of the FG shown in Fig. 1 (where, following [20], a simplified notation is used for the involved pdfs and the equality constraint node, which represents an equality constraint “function”, that is a Dirac delta function). Since this FG is cycle free, the pdf f (xl , y1:l ) can be evaluated applying the well known sum-product rule 4 (i.e., the SPA) to it, i.e. developing a proper mechanism for passing probabilistic messages along this FG (the flow of messages is indicated by red arrows in Fig. 1). In fact, if the input message m ~ in (xl ) = f (xl , y1:(l−1) ) enters this FG, the message going out of the equality node is given by m ~ e (xl ) = m ~ in (xl ) f (yl |xl ) , (11) so that m ~ e (xl ) = f (xl , y1:l ) (see (9)); then, the message emerging from the function node referring to the pdf f (xl+1 |xl ) is expressed by Z m ~ out (xl+1 ) = f (xl+1 |xl ) m ~ e (xl ) dxl , (12) so that m ~ out (xl+1 ) = f (xl+1 , y1:l ) = m ~ in (xl+1 ) (see (10)). From this result it can be easily inferred that the pdf f (xt , y1:t ) (and, up to a scale factor, the pdf 3 Forney-style factor graphs are always considered in the following [20]. a Forney-style FG, such a rule can be formulated as follows [20]: the message emerging from a node f along some edge x is formed as the product of f and all the incoming messages along all the edges that enter the node f except x, summed over all the involved variables except x. 4 In G. M. Vitetta et al., Marginalized Particle Filtering . . . 8 xl min ( xl ) = f y l / xl f xl +1 / xl yl me ( xl ) x l +1 mout ( xl +1 ) = min ( xl +1 ) Figure 1: Factor graph representing (9) and (10). The SPA message flow characterizing the l-th recursion of Bayesian filtering is indicated by red arrows. f (xt |y1:t )) results from the application of the SPA to the overall FG originating from the ordered concatenation of multiple subgraphs, each structured like the one shown in Fig. 1 and associated with l = 1, 2, ..., t. In this graph the flow of messages produced by the SPA proceeds from left to right, i.e. the pdf f (xt , y1:t ) is generated by a forward only message passing. Note also that, in principle, the desired pdf f (xt , y1:t ) is computed as the product between two messages, one for each direction, reaching the rightmost half edge of the overall FG, but one of − (x ) = 1. the two incoming messages for that edge is the constant function ← m he t Unluckily, as the size D of xl (1) gets large, the computational burden associated with (6)-(8) (or, equivalently, (9) and (10)) becomes unmanageable. In principle, a substantial complexity reduction can be achieved decoupling 5 the (L) (N ) (L) filtering problem for xl from that for xl , i.e. the evaluation of f (xl |y1:l ) (N ) from that of f (xl |y1:l ). In fact, this approach potentially provides the following two benefits: a) a given filtering problem is turned into a couple of filtering problems of smaller dimensionality and b) some form of computationally efficient standard filtering (e.g., Kalman or extended Kalman filtering) can be hopefully (L) exploited for the linear portion xl of the state vector xl (1). As a matter of fact, these principles have been exploited in devising MPF [10, 15] and, as it will become clearer in the following, they must be always kept into account in the derivation of our FG representation. Before illustrating this derivation, however, the measurement and state models on which such a representation relies need to be clearly defined; for this reason, these models are analysed in detail in the following part of this Section. To begin, let us concentrate on the models (L) (L) involved in the filtering problem for xl , i.e. on the evaluation of f (xl |y1:l ), 5 Note that, generally speaking, in the considered problem the coupling of the filtering (L) (N) problem for xl with that for xl is due not only to the structure of the update equations (2) and (3), but also to the measurement vector yl (5), since this exhibits a mixed dependence on the two components of the state vector xl (1). 9 G. M. Vitetta et al., Marginalized Particle Filtering . . . (N ) under the assumption that the nonlinear portion xl of the system state is known for any l. In this case, the evaluation of this pdf can benefit not only from the knowledge of yl (5), but also from that of the quantity (see (3))     (L) (N ) (N ) (N ) (N ) (N ) (L) (N ) zl , xl+1 − fl xl = Al xl xl + wl , (13) which can be interpreted as a pseudo-measurement [15], since it does not originate from real measurements, but from the constraints expressed by the state equation (3). This leads to considering the overall observation model   (L) (L) (N ) f yl , zl xl , xl     (14) (L) (N ) (L) (L) (N ) = f yl xl , xl f zl xl , xl (L) for xl , where   (L) (N ) f yl xl , xl = f (el )|e =y −B x(N ) x(L) −h x(N )  , l l l l l l (15) l and     (L) (L) (N ) (N ) f zl xl , xl = f wl (N ) wl (L) =zl (N ) −Al   (L) (N ) xl xl . (16) If the observation model (14) and the state model (see (2))   (L) (L) (N ) f xl+1 xl , xl       (L) (L) (N ) (L) (N ) (L) xl − Al xl xl = fw(L) xl+1 − fl (17) (L) are adopted for xl , the graph identified by the blue lines and rectangles appearing in Fig. 1 can be drawn. Then, in principle, if the sum-product rule is (N ) (N ) applied to it under the assumption that the couple (xl , xl+1 ) is known for any l, the expressions of the messages flowing in the overall graph for the evaluation (L) (L) of f (xt , y1:t , z1:t ) can be easily derived. It is important to point out that: • This graph contains a node which does not refer to the above mentioned density factorizations6 , but represents the transformation from the couple (N ) (N ) (L) (xl , xl+1 ) to zl (see (13)); this feature of the graph has to be carefully kept into account when deriving message passing algorithms. (L) (L) (N ) • Generally speaking, the evaluation of the conditional pdf f (zl |xl , xl ) (N ) (N ) requires the knowledge of the joint pdf of xl+1 and xl conditioned on (L) xl (see (13)). 6 This peculiarity is also evidenced by the presence of an arrow on all the edges connected to such a node. 10 G. M. Vitetta et al., Marginalized Particle Filtering . . . The same line of reasoning can be followed for the filtering problem concern(N ) (L) ing xl . Consequently, in this case the linear portion xl of the system state is assumed to be known for any l and the pseudo-measurement (see (2))     (N ) (L) (L) (N ) (L) (L) (N ) (L) zl , xl+1 − Al xl xl = fl xl + wl (18) is defined. This leads to the overall observation model   (N ) (N ) (L) f yl , zl xl , xl =     (N ) (L) (N ) (N ) f yl xl , xl f zl xl (N ) (N ) (N ) (19) (L) (L) (N ) for xl , where f (zl |xl ) can be expressed similarly as f (zl |xl , xl ) (see (16)). Then, if the observation model (19) and the state model   (N ) (N ) (L) f xl+1 xl , xl       (20) (N ) (N ) (N ) (N ) (N ) (L) xl − Al xl xl = fwN xl+1 − fl (N ) are adopted for xl , the red graph of Fig. 2 can be drawn and exploited in (N ) (N ) a similar way as the blue graph for the evaluation of f (xt , y1:t , z1:t ), under (L) (L) the assumption that the couple (xl , xl+1 ) is known for any l. Note also that, similarly to what has been mentioned earlier about the conditional pdf (L) (L) (N ) (N ) (N ) f (zl |xl , xl ), the evaluation of f (zl |xl ) requires the knowledge of the (L) (L) (N ) joint pdf of xl+1 and xl conditioned on xl . Finally, merging the blue graph with the red one (i.e., adding four equality (L) (N ) (L) (N ) constraint nodes for the variables xl , xl , xl+1 and xl+1 shared by the red graph and the blue one) produces the overall FG illustrated in Fig. 2. Given this FG, we would like to follow the same line of reasoning as that adopted for the FG (L) (L) of Fig. 1. In other words, given the input messages m ~ in (xl ) = f (xl , y1:(l−1) ) (N ) (N ) and m ~ in (xl ) = f (xl , y1:(l−1) ) (entering the FG along the half edges asso(L) (N ) ciated with xl and xl , respectively), we would like to derive a forward only message passing algorithm based on this FG and generating the output (L) (L) (N ) (N ) messages m ~ out (xl+1 ) = f (xl+1 , y1:l ) and m ~ out (xl+1 ) = f (xl+1 , y1:l ) (emerging (L) (N ) from the FG along the half edges associated with xl+1 and xl+1 , respectively) on the basis of the available a priori information and the noisy measurement yl . Unluckily, the new FG, unlike that shown in Fig. 1, is not cycle-free, so that any application of the SPA to it unavoidably leads to approximate results [21], whatever message scheduling procedure [21, 27] is adopted. This consideration must be carefully kept into account in both the derivation of MPF as a message passing algorithm and in the development of possible modifications and generalizations of this technique, as it will become clearer in Sections 4-6. 11 G. M. Vitetta et al., Marginalized Particle Filtering . . .
mout ( xl(+L1) ) xl( L )
min ( xl( L ) ) ( L) xl+ 1 f x( L ) / x( L ) , x ( N ) l +1 = = l l = = f z ( L ) / x ( L ) ,x ( N ) l z l l ( L) l xl( N ) , xl(+N1) → z l( L ) = f y / x ( L ) ,x ( N ) l l l yl = xl( L ) , xl(+L1) → z l( N ) z l( N ) f z ( N ) / x( N ) l x (N ) l =
min ( xl( N ) ) = l = = f x( N ) / x( N ) , x( L ) l +1 l l
mout ( xl(+N1) ) (N ) xl+ 1 Figure 2: Overall factor graph resulting from the merge of two subgraphs, one (L) (N ) referring to filtering for xl (in blue), the other one to that for xl (in red). The equality constraint nodes introduced to connect these subgraphs are identi(L) (N ) fied by black lines. The flow of the messages along the half edges xl and xl (L) (N ) (input) and that of the messages along the half edges xl+1 and xl+1 (output) are indicated by green arrows. G. M. Vitetta et al., Marginalized Particle Filtering . . . 4 12 Message Passing in Marginalized Particle Filtering In the following Section we show how the equations describing the l-th recursion of MPF result from the application of the SPA to the FG shown in Fig. 2. However, before illustrating the detailed derivation of such equations, it is important to discuss the following relevant issues. First of all, the MPF technique has been developed for the specific class of GLG SSMs [15, 19], to which we always refer in the following discussion. In particular, in the following we assume (L) (N ) that: a) {wk } ({wk }) is a Gaussian random process and all its elements (L) (N ) (L) have zero mean and covariance Cw (Cw ) for any l; b) {ek } is a Gaussian random process having zero mean and covariance matrix Ce for any l; c) all the above mentioned Gaussian processes are statistically independent. Under these (L) (N (L) (L) (L) (L) (N ) assumptions, the pdfs f (yl |xl , xl ), f (zl |xl ) and f (xl+1 |xl , xl ) (see (N ) (L) (N ) (15)-(17)) are Gaussian with mean (covariance matrix) Bl (xl )xl +hl (xl ), (N ) (N ) (L) (L) (N ) (L) (N ) (L) (N ) Al (xl ) xl and fl (xl ) + Al (xl ) xl , respectively (Ce , Cw and (N ) (N ) (N ) (N ) (L) (L) Cw , respectively). Similarly, the pdfs f (zl |xl ) and f (xl+1 |xl , xl ) (L) (N ) (see (19) and (20)) are Gaussian with mean (covariance matrix) fl (xl ) and (N ) (N ) (N ) (N ) (N ) (L) (L) fl (xl ) + Al (xl ) xl , respectively (Cw and Cw , respectively). Secondly, as explained below in detail, the MPF can be interpreted as a forward only message passing algorithm operating over the FG shown in Fig. 2. The scheduling procedure adopted for MPF unavoidably leads to ignoring (N ) the evaluation of the pseudo-measurement zl (18). For this reason, in the following we refer to the simplified FG shown in Fig. 3, which has been obtained from that illustrated in Fig. 2 removing the block representing the transfor(L) (L) (N ) mation from (xl , xl+1 ) to zl and the edges referring to the evaluation of the last vector. Note that: a) in the new graph the block referring to the pdf (L) (N f (yl |xl , xl ) appears twice, since this pdf is involved in the two subgraphs shown in Fig. 2; b) some brown edges and equality nodes have been added to (N ) (L) feed such blocks with min (xl ) and min (xl ), since these represents the only a (N ) (L) priori information available about xl and xl , respectively, at the beginning of the l-th recursion. Thirdly, in MPF a particle-based model and a Gaussian model are adopted (N ) (L) for the input and the output messages referring to xl and xl , respectively, and the functional structure of the generated messages is preserved in each recursion. More specifically, on the one hand, the a priori information avail(N ) able about xl at the beginning of the l-th recursion is represented by a set (N ) (N ) of Np particles Sl/(l−1) = {xl/(l−1),j , j = 0, 1, ..., Np − 1} and their weights G. M. Vitetta et al., Marginalized Particle Filtering . . . 13 {wl/(l−1),j , j = 0, 1, ..., Np − 1)}; following [15], we assume that such weights are uniform (in other words, wl/(l−1),j = 1/Np for j = 0, 1, ..., Np − 1), so that they can be ignored in the following derivation. On the other hand, (L) the a priori information available about xl is represented by a set of Gaussian pdfs, each associated with a specific particle; in particular, the Gaussian (L) (L) (L) model N (xl ; ηl/(l.−1),j , Cl/(l−1),j ) is associated with the j-th particle (with j = 0, 1, ..., Np − 1) at the beginning of the same recursion. From the last point it can be inferred that, in developing a message passing algorithm that represents the MPF technique, we can focus on: a) a single particle contained in the (N ) (N ) input message min (xl ) and, in particular, on the j-th particle xl/(l−1),j ; b) (L) (L) (L) on the Gaussian model N (xl ; ηl/(l.−1),j , Cl/(l−1),j ) associated with that particle. For this reason, we assume that, at the beginning of the l-th recursion, (L) (N ) our knowledge about xl and xl is condensed in the message     (L) (L) (L) (L) (21) = N xl ; ηl/(l.−1),j , Cl/(l−1),j m ~ in,j xl and in the message     (N ) (N ) (N ) = δ xl − xl/(l−1),j , m ~ in,j xl (22) respectively, with j = 0, 1, ..., Np − 1; these are processed to generate the corre(L) (N ) sponding output messages m ~ out,j (xl+1 ) and m ~ out,j (xl+1 ), which are required to (L) (N ) have the same functional form as m ~ in,j (xl ) (21) and m ~ in,j (xl ) (22), respec(N ) tively. For this reason, the algorithm for computing m ~ out,j (xl+1 ) is expected (N ) to generate a new particle x(l+1)/l,j with a (uniform) weight w(l+1)/l,j = 1/Np ; (L) similarly, that for evaluating m ~ out,j (xl+1 ) is expected to produce a new Gaus(N ) (L) (L) (L) sian pdf N (xl+1 ; η(l+1)/l,j , C(l+1)/l,j ) associated with the particle x(l+1)/l,j (note that, in deriving this pdf, possible scale factors are unrelevant and, consequently, can be dropped). (L) (N ) Given m ~ in,j (xl ) (21) and m ~ in,j (xl ) (22), if the SPA is applied to the considered graph and the message scheduling illustrated in Fig. 3 (and, as a matter of fact, adopted in MPF) is employed, the steps described below are (L) (N ) carried out to evaluate m ~ out,j (xl+1 ) and m ~ out,j (xl+1 ) in the l-th recursion of MPF7 . (N ) 1. Measurement update for xl - This step aims at updating the weight (N ) of the j-th particle xl/(l−1),j on the basis of the new measurements yl (this 7 In the following derivations some mathematical results about Gassian random variables (e.g., see [32, Par. 2.3.3]) and Gaussian message passing in linear models (e.g., see [20, Table 2, p. 1303]) are exploited. As far as the MPF formulation is concerned, we always refer to that given by algorithm 1 in [15, Sec. II]. 14 G. M. Vitetta et al., Marginalized Particle Filtering . . .
min , j ( x (l L ) ) x (l L ) yl
min , j ( x l( L ) )
m2, j ( x (l L ) )
m2, j ( x l( L ) )
m4, j ( x l( L ) ) = f x ( L ) / x ( L ) ,x ( N ) l +1 = fy
m1, j ( x l( L ) ) l ,/ x l( L ) ,x l( N ) =
m3, j ( x (l L ) ) l
m j ( z l( L ) ) z l x (l +L1)
m5, j ( x l(+L1) )
= mout , j ( x (l +L1) ) f z( L ) / x ( L ) ,x ( N ) l
min , j ( x l( N ) ) l = l
m2, j ( x l( N ) ) ( L) l x (l N ) , x l(+N1) → z l( L )
m5, j ( x (l +N1) )
m2, j ( x l( N ) ) =
m2, j ( x (l L ) )
m2, j ( x (l L ) )
min , j ( x l( L ) ) fy l = x (l N )
m2, j ( x l( N ) )
m1, j ( x l( N ) ) ,/ x l( L ) ,x l( N ) =
min , j ( x l( N ) )
m2, j ( x (l N ) ) yl
m2, j ( x (l N ) )
m5, j ( x (l +N1) ) = = f x ( N ) / x ( N ) ,x ( L ) l +1 l l
mout , j ( x l(+N1) ) x (l +N1) Figure 3: Overall factor graph for the representation of MPF processing; this graph is obtained from the one shown in Fig. 3 removing the part referring to (N ) the evaluation of zl and inserting two new (brown) equality constraints and (L) (N ) some (brown) edges referring to xl and xl . The message flow characterizing MPF and referring to the j-th particle is also shown. 15 G. M. Vitetta et al., Marginalized Particle Filtering . . . corresponds to step 2) of algorithm 1 in [15, Sec. II]). It involves the computation of the messages    Z    (N ) (L) (N ) (L) (L) m ~ 1,j xl m ~ in,j xl = f yl xl , xl dxl (23) and       (N ) (N ) (N ) , m ~ 1,j xl =m ~ in,j xl m ~ 2,j xl (24)       (N ) (N ) (L) (N ) (N ) ηl/(l.−1),j + hl xl η1,l,j xl , Bl xl (26) which provides the new importance weight for the considered particle (see Fig. (N ) (L) 3). Substituting the expression of f (yl xl , xl ) (see (15)) and (21) in (23) produces, after some manipulation (see the Appendix)        (N ) (N ) (N ) (N ) (N ) = N yl ; η1,l,j xl , C1,l,j xl , (25) m ~ 1,j xl where and T      (N ) (N ) (L) (N ) (N ) + Ce . Cl/(l−1),j Bl xl C1,l,j xl , Bl xl (N ) (27) (N ) Then, substituting m ~ in,j (xl ) (22) and m ~ 1,j (xl ) (25) in (24) yields     (N ) (N ) (N ) = wl,j δ xl − xl/(l−1),j , m ~ 2,j xl (28) where8   (N ) (N ) wl,j , N yl ; η1,l,j , C1,l,j (29) (N ) is the new particle weight combining the a priori information about xl with the information provided by the new measurements; here (see (26) and (27))   (L) (N ) (N ) (N ) (30) η1,l,j , η1,l,j xl/(l−1),j = Bl,j ηl/(l.−1),j + hl,j and   (L) (N ) (N ) (N ) C1,l,j , C1,l,j xl/(l−1),j = Bl,j Cl/(l−1),j BTl,j + Ce , (31) (N ) (N ) with hl,j , hl (xl/(l−1),j ) and Bl,j , Bl (xl/(l−1),j ). In MPF, after normal(w) ization9 of the particle weights {wl,j } (i.e., after dividing them by Pl (N) , 8 In evaluating this weight, the factor [det(C −P /2 appearing in the expression of the 1,l,j )] involved Gaussian pdf is usually neglected, since this entails a negligible loss in estimation accuracy. 9 Note that normalization requires the knowledge of all the weights {w } (29) and that, l,j unlike it, all the previous and following tasks can be carried out in parallel (i.e., on a particleby-particle basis). 16 G. M. Vitetta et al., Marginalized Particle Filtering . . . Np −1 X wl,j ), particle resampling with replacement is accomplished (this corre- j=0 sponds to step 3) of algorithm 1 in [15, Sec. II]). Note that, even if resampling does not emerge from the application of SPA to the considered graph, its use, as it will become clearer at the end of this Section, plays an important role in (N ) the generation of the new particles for xl+1 . Moreover, it can be easily incorporated in our message passing; in fact, resampling simply entails that Np parti(N ) cles {xl/(l−1),j } and their associated weights {wl,j } (29) are replaced by the new (N ) (N ) particles {xl/l,j } (forming the new set Sl/l ) and their weights {wl/l,j = 1/Np }, (N ) respectively. Consequently, m ~ 2 (xl ) (28) is replaced by     (N ) (N ) (N ) = δ xl − xl/l,j , m ~ 2,j xl (32) since the particle weight does not depend on the index j. It is also worth (L) mentioning that, after resampling, the set of Gaussian messages {m ~ in,j (xl )} (21) needs to be properly reordered and that the messages associated with all the discarded particles are not propagated to the next steps. (L) 2. Measurement update for xl - This step aims at updating our statistical (L) knowlege about xl on the basis of the new measurement yl (and corresponds to step 3-a) of algorithm 1 in [15, Sec. II]). It involves the computation of the messages     Z   (N ) (N ) (L) (N ) (L) dxl (33) m ~ in,j xl = f yl xl , xl m ~ 1,j xl and       (L) (L) (L) , m ~ 1,j xl =m ~ in,j xl m ~ 2,j xl (34) (L) which represents the output of the measurement update for xl (see Fig. 3). Substituting (15) and (22) in (33) produces (see [32, Par. 2.3.3, eq. (2.115)])     (L) (L) (35) = N yl ; Bl,j xl + hl,j , Ce m ~ 1,j xl which, after some manipulation (in which unrelevant scale factors are dropped), can be put in the Gaussian form     (L) (L) (L) (L) = N xl ; η1,l,j , C1,l,j , (36) m ~ 1,j xl with (L) (L) (L) w1,l,j , W1,l,j η1,l,j = BTl,j We (yl − hl,j ) ,  −1 (L) (L) W1,l,j , C1,l,j = BTl,j We Bl,j (37) (38) 17 G. M. Vitetta et al., Marginalized Particle Filtering . . . and We , C−1 e . Then, substituting (21) and (36) in (34) yields     (L) (L) (L) (L) = N xl ; ηl/(l−1),j , Cl/(l−1),j m ~ 2,j xl   (L) (L) (L) ·N xl ; η1,l,j , C1,l,j , which can be reformulated as     (L) (L) (L) (L) = N xl ; η2,l,j , C2,l,j , m ~ 2,j xl (39) (40) if scale factors are ignored; here, (L) (L) (L) (L) (41)  −1 (L) (L) (L) (L) W2,l,j , C2,l,j = Wl/(l−1),j + W1,l,j , (42) (L) (L) (L) (L) (L) (L) w2,l,j , W2,l,j η2,l,j = wl/(l−1),j + w1,l,j , Wl/(l−1),j , (Cl/(l−1),j )−1 and wl/(l−1),j , Wl/(l−1),j ηl/(l−1),j . (N ) 3. Time update for xl - This step aims at generating the j-th particle for and its associated weight (this corresponds to step 3-b) of algorithm 1 in (N ) [15, Sec. II]); these information are conveyed by the message m ~ 5,j (xl+1 ), which can be expressed as (see Fig. 3)   RR   (N ) (N ) (L) (N ) f xl+1 xl , xl m ~ 5,j xl+1 =     (43) (N ) (L) (L) (N ) ·m ~ 2,j xl m ~ 2,j xl dxl dxl . (N ) xl+1 The double integral appearing in the RHS of the last equation can be evaluated (N ) as follows. First of all, substituting m ~ 2,j (xl ) (32) in (43) yields      Z  (L) (L) (N ) (N ) (L) (N ) dxl . (44) ~ 2,j xl m ~ 5,j xl+1 = f xl+1 xl , xl/l,j m (N ) (N ) (L) (L) Then, substituting the expression of f (xl+1 |xl , xl ) (see (20)) and m ~ 2,j (xl ) (40) in (44) yields, after some manipulation, the Gaussian message (see [32, Par. 2.3.3, eq. (2.115)])     (N ) (N ) (N ) (N ) (45) m ~ 5,j xl+1 = N xl+1 ; η5,l,j , C5,l,j , where (N ) (N ) C5,l,j (N ) (N ) Al,j , Al (N ) (N ) (L) (N ) η5,l,j , Al,j η2,l,j + fl,j ,  T (N ) (L) (N ) ) , C(N + A C A , w l,j 2,l,j l,j (N ) (N ) (46) (47) (N ) (xl/l,j ) and fl,j , fl (xl/l,j ). Note that, in principle,     (N ) (N ) ~ 5,j xl+1 , m ~ out,j xl+1 = m (48) 18 G. M. Vitetta et al., Marginalized Particle Filtering . . . as it can be easily inferred from Fig. 3. However, as already mentioned above, (j) (N ) in MPF the output message m ~ out (xl+1 ) is required to have the same functional (j) (N ) form as m ~ in (xl ) (22). This result can be achieved a) sampling the Gaussian (N ) (N ) (N ) (N ) function N (xl+1 ; η5,l,j , C5,l,j ) (see (45)), that is drawing a sample x(l+1)/l,j from (N ) it and b) assigning to the sample x(l+1)/l,j a probability w(l+1)/l,j equal to the weight wl/l,j = 1/Np (originating from resampling). It is worth pointing out (N ) that: 1) the particles {x(l+1)/l,j } form the new set S(l+1)/l ; 2) in accomplishing step a) of this procedure, it may be useful to introduce artificial noise (this can be simply done adding the same positive quantity to the diagonal elements of (N ) the matrix Cw appearing in the RHS of (47)) in order to mitigate the so called (N ) degeneracy problem [1, 34]. If this approach is adopted, the message m ~ 5,j (xl+1 ) (45) is replaced by     (N ) (N ) (N ) (49) m ~ 5,j xl+1 = δ xl+1 − x(l+1)/l,j , (N ) which emerges from the graph as m ~ out,j (xl+1 ). This message is also used in the (L) time update for xl , as illustrated in the next step. (L) 4. Time update for xl - This step aims at generating a new Gaussian (L) (N ) ~ 5,j (xl+1 ) = pdf associated with the j-th particle x(l+1)/l,j and conveyed by m (L) m ~ out,j (xl+1 ) (this corresponds to step 3-c) of algorithm 1 in [15, Sec. II]). However, before doing that, a further measurement update is accomplished on (L) the basis of the pseudo-measurement zl (13). This involves the evaluation of (L) the messages m ~ j (zl ),      Z  (L) (L) (L) (N ) (L) (L) f zl xl , xl dzl (50) = m ~ j zl m ~ 3,j xl and       (L) (L) (L) m ~ 4,j xl =m ~ 2,j xl m ~ 3,j xl , (51) (L) as shown in Fig. 3. Generally speaking, the message m ~ j (zl ) can be expressed as   RR   (L) (L) (N ) (N ) = f zl xl , xl+1 m ~ j zl     (52) (N ) (N ) (N ) (N ) (N ) ·f xl+1 xl f xl dxl dxl+1 . (N ) (N ) However, since in this case f (xl ) = δ(xl (N ) x(l+1)/l,j ) (N ) m ~ 2,j (xl ) (N ) (N ) (L) (N +1) −xl/l,j ), f (xl+1 |xl ) = δ(xl (N ) m ~ 5,j (xl+1 ) − can be assumed (see (32) and (49), respectively), eq. (52) easily leads to the expression       (L) (L) (N ) (L) (L) (N ) (53) = f zl xl/l,j , x(l+1)/l,j = δ zl − zl,j , m ~ j zl 19 G. M. Vitetta et al., Marginalized Particle Filtering . . . where (L) (N ) (N ) zl,j , x(l+1)/l,j − fl,j . (54) (L) (L) (N ) Then, substituting (53) and the expression of f (zl |xl , xl (50) yields     (L) (L) (N ) (L) ) . = N zl,j ; Al,j xl , C(N m ~ 3,j xl w ) (see (16)) in Finally, substituting the last expression and (40) in (51) produces     (L) (L) (L) (L) = N xl ; η2,l,j , C2,l,j m ~ 4,j xl   (L) (N ) (L) (N ) ·N zl,j ; Al,j xl , Cw , (55) (56) which, after some manipulation (in which unrelevant scale factors are dropped), can be rewritten as     (L) (L) (L) (L) = N xl ; η4,l,j , C4,l,j , (57) m ~ 4,j xl where  T (L) (L) (L) (L) (N ) (N ) (L) w4,l,j , W4,l,j η4,l,j = w2,l,j + Al,j Ww zl,j , (N )  T  −1 (L) (N ) (L) (L) (N ) (N ) Ww Al,j = W2,l,j + Al,j W4,l,j , C4,l,j (58) (59) (N ) and Ww , [Cw ]−1 . (L) The last part of the time update step for xl requires the evaluation of the output message   RR   (L) (L) (L) (N ) f xl+1 xl , xl m ~ 5,j xl+1 =     (60) (N ) (L) (N ) (L) dxl dxl , m ~ 2,j xl ·m ~ 4,j xl (N ) which, similarly as m ~ 5,j (xl+1 ) (43), requires double integration. Substituting (N ) m ~ 2,j (xl ) (32) in the RHS of the last expression yields      Z  (L) (L) (L) (L) (L) (N ) dxl . (61) ~ 4,j xl f xl+1 xl , xl/l,j m m ~ 5,j xl+1 = (L) (L) (N ) Then, substituting the expression of f (xl+1 |xl , xl/(l−1),j ) (see (17)) and (57) in the last equation gives (see [32, Par. 2.3.3, eq. (2.115)])     (L) (L) (L) (L) m ~ 5,j xl+1 = N xl+1 ; η5,l,j , C5,l,j   (62) (L) =m ~ out,j xl+1 , where (L) (L) (L) (L) (L) η5,l,j , Al,j η4,l,j + fl,j = η(l+1)/l,j , (63) 20 G. M. Vitetta et al., Marginalized Particle Filtering . . . (L) (L)  T (L) (L) (L) (L) (L) C5,l,j , C(L) + A C A = C(l+1)/l,j , w l,j 4,l,j l,j (N ) (L) (L) (64) (L) (N ) ~ out,j (xl+1 ) (62) fl,j , fl (xl/l,j ) and Al,j , Al (xl/l,j ). The evaluation of m concludes the MPF message passing procedure, which needs to be carried out for each of the Np particles available at the beginning of the l-th recursion. Note that this procedure needs a proper inizialization (this corresponds to step 1) of algorithm 1 in [15, Sec. II]). In practice, before starting the first recursion (N ) (N ) (corresponding to l = 1), the set S1/0 = {x1/0,j , j = 0, 1, ..., Np − 1}, consisting (N ) of Np particles, is generated for x1 sampling the pdf  Z  (L) (N ) = f (x1 ) dx1 , f x1 (65) (L) (L) (L) and the same weight w1/0 = 1/Np and Gaussian model N (x1 ; η1/0 , C1/0 ) for (L) are assigned to each of them. Finally, it is worth pointing out that: 1) the processing accomplished in the (L) measurement and time update for xl can be interpreted as a form of Kalman filtering, in which both the real measurement yl and the pseudo-measurement (L) (N ) (L) zl are processed [15]; 2) in the l-th recursion estimates of xl and xl can P Np −1 (w) (L) (N ) (N ) (see (28)) and as x̂l = be evaluated as x̂l = j=0 wl,j xl/(l−1),j /Pl PNp −1 (L) j=0 η4,l,j /Np (see (58)), respectively; 3) the result expressed by eq. (45) shows that, generally speaking, the statistical representation generated by the (N ) SPA for the state xl+1 is a Gaussian mixture (GM), whose Np components have the same weight (equal to 1/Np ) because resampling is always used in step 1. The last point leads to the conclusion that, if resampling was not accomplished in the l-th recursion, the weight of the j-th component of this GM would be proportional by wl,j (29); this would unavoidably raise the problem of sampling (N ) a GM with unequally weighted components in generating the particle set Sl/(l+1) x1 (L) and that of properly handling the resulting pseudo-measurements {zl,j }. These considerations motivate the use of resampling in each recursion, indipendently (N ) of the effective sample size [1] characterizing the particle set Sl/(l−1) . 5 Simplifying Marginalized Particle Filtering The MPF derivation illustrated in the last two Sections unveils the real nature of MPF and its limitations, and shows the inner structure of the processing accomplished within each step. For these reasons, it paves the way for the development of new filtering methods related to MPF. In this Section we exploit our FG-based representation of Bayesian filtering to develop reduced complexity G. M. Vitetta et al., Marginalized Particle Filtering . . . 21 alternatives to MPF by simplifying the message passing derived in the previous Section. It is worth mentioning that some methods for reducing MPF computational complexity [16] have been already proposed in the technical literature [4], [17], [18]. In particular, the method proposed in [4] and [17] is based on rep(N ) resenting the particle set for xl as a single particle (that corresponds to the center of mass of the set itself), so that a single Kalman filter is employed in up(L) dating the statistics of the linear component xl ; consequently, the statistical (L) knowledge about xl is condensed in the single message     (L) (L) (L) (L) (66) = N xl ; ηl/(l.−1) , Cl/(l−1) , m ~ in xl (L) instead of the Np messages {m ~ in,j (xl )} (21). Unluckily, this simplified MPF (N ) algorithm works well only if the posterior distribution of xl is unimodal. Its (N ) generalization to the case in which the posterior distribution of xl is multimodal has been illustrated later in [18]. In the proposed technique the particles are partitioned into Kl groups or clusters (the parameter Kl is required to equal (N ) the number of modes of the posterior density of xl ) and each group is represented by a single particle that corresponds to its center of mass; this allows to reduce the overall number of Kalman filters from Np to Kl . The implementation of this technique requires, however, solving the following two specific problems: a) identifying the number of modes of the posterior distribution of (N ) xl ; b) partitioning the particles into clusters according to a grouping method in each recursion. Unluckily, practical solutions for suche problems have not been proposed in [18]. Our derivation of simplified algorithms has been only partly inspired by the manuscripts cited above. In fact, first of all, it relies on the following specific methods: a) a set of Np equal weight particles {xj ; j = 0, 1, ..., Np − 1} is represented through its center of mass x̄ , Np −1 1 X xj , Np j=0 (67) as already suggested in [17] and [18], when the computation of a message refer(L) (N ) ring to xl involves the particle-based representation of xl ; b) a set of Np Gaussian messages {N (x; ηj , Cj ); j = 0, 1, ..., Np − 1}, that refer to a set of Np equal weight particles, is represented as the Np −component GM Np −1 1 X N (x; ηj , Cj ) , fGM (x) , Np j=0 (68) 22 G. M. Vitetta et al., Marginalized Particle Filtering . . . and this GM is approximated through its projection onto the Gaussian pdf fG (x) = N (x; ηG , CG ), where ηG and CG are selected as ηG , Np −1 1 X ηj Np j=0 (69) and PNp −1 CG = (1/Np ) j=0 Cj PNp −1 T T −ηG (ηG ) + (1/Np ) j=0 ηj (ηj ) (70) respectively, so that the mean and covariance of fGM (x) (68) are preserved (L) (e.g., see [33, Sec. IV]). Secondly, as far as the messages {m ~ in,j (xl )} (21) are concerned, we do not adopt the approximations proposed in [4], [17] and [18]. In fact, we focus on the following two cases: case #1 - the messages (L) {m ~ in,j (xl )} are all different but, when needed in message passing, are con(L) (L) densed in the single message (66), where ηl/(l.−1) and Cl/(l−1) are given by (69) (L) (L) and (70), respectively, with ηj = ηl/(l.−1),j and Cj = Cl/(l−1),j ; b) case #2 (L) the messages {m ~ in,j (xl )} have different means, but their covariance matrices (L) (L) {Cl/(l−1),j } are all equal (their common value is denoted C̃l/(l−1) in the following). In both cases our simplifications aim at minimizing the overall number of a) Cholesky decompositions for the generation of the new particle set S(l+1)/l (L) (such decompositions are required for the Np matrices {C5,l,j } (47)) and b) matrix inversions; such inversions are needed to compute: a) the Np matrices (N ) (N ) {W1,l,j , (C1,l,j )−1 } (required in the evaluation of the weights {wl,j } on the (L) basis of (29)); b) the Np matrices {Wl/(l−1),j } (required to evaluate the vec(L) (L) (L) tors {w2,l,j } (41) and the matrices {W2,l,j } (42)); c) the Np matrices {C2,l,j } (L) (employed in (47)); c) the Np matrices {C4,l,j } (employed in (64)). Based on the methods illustrated above, our simplified versions of MPF are (N ) derived as follows. First of all, in the measurement update for xl , we use a single covariance matrix in the Gaussian pdf appearing in the RHS of (29); in other words, the j-th weigth wl,j is computed as   (N ) (N ) wl,j , N yl ; η1,l,j , C1,l , (71) (N ) (N ) where C1,l is evaluated on the basis of (70) (see also (69)), setting ηj = η1,l,j (N ) (30) and Cj = C1,l,j (31) for any j. (L) (N ) Second, in the measurement update for xl , the particle set Sl/(l−1) is con(N ) densed in its center of mass x̄l/(l−1) (see (67)). Consequently, the message 23 G. M. Vitetta et al., Marginalized Particle Filtering . . . (L) m ~ 1,j (xl ) (35) is replaced by its particle-independent form     (L) (L) = N yl ; B̄l xl + h̄l , Ce m ~ 1 xl (72) (N ) (N ) where B̄l , Bl (x̄l/(l−1) ) and h̄l , hl (x̄l/(l−1) ). This message, similarly as (35), can be put in the Gaussian form (see (36)-(38))     (L) (L) (L) (L) = N xl ; η1,l , C1,l , (73) m ~ 1 xl where (L) (L) (L) w1,l , W1,l η1,l = B̄Tl We yl − h̄l and
(74)  −1 (L) (L) W1,l , C1,l = B̄Tl We B̄. (75) (L) This allows us to replace the message m ~ 2,j (xl ) (40) with its particle-independent counterpart     (L) (L) (L) (L) = N xl ; η2,l , C2,l , (76) m ~ 2 xl (L) (L) (L) (L) (L) (L) (L) where w2,l , W2,l η2,l and W2,l , (C2,l )−1 are easily obtained from (41) and (L) (L) (42) replacing a) w1,l,j and W1,l,j with w1,l (74) and W1,l (75), respectively; (L) (L) (L) (L) (L) (L) b) wl/(l−1),j and Wl/(l−1),j with wl/(l−1) , Wl/(l−1) ηl/(l−1) and Wl/(l−1) , (L) (L) (L) (Cl/(l−1) )−1 (ηl/(l.−1) and Cl/(l−1) are given by (69) and (70), respectively, with (L) (L) (L) ηj = ηl/(l.−1),j and Cj = Cl/(l−1),j ). Note that, since the precision matrix W2,l (L) is particle-independent, a single matrix C2,l has to be computed for the next step. (N ) (L) Thirdly, in the time update for xl , the message m ~ 2 (xl ) (76) can be used (L) (N ) in place of m ~ 2,j (xl ) (40) in the evaluation of m ~ 5,j (xl+1 ) (see (44) and Fig. 3). This leads to the particle-dependent message     (N ) (N ) (N ) (N ) m ~ 5,j xl+1 = N xl+1 ; η5,l,j , C5,l,j , (77) (N ) (N ) where η5,l,j and C5,l,j are obtained from (46) and (47), respectively, replac(L) (L) (L) (L) ing η2,l,j and C2,l,j with η2,l and C2,l , respectively. Then, to simplify the (N ) (N ) generation of the new particle set S(l+1)/l , the covariance matrices {C5,l,j } are (N ) (N ) condensed in a single matrix C5,l using (70) (see also (69)) with ηj = η5,l,j = (N ) (L) (N ) Al,j η2,l + fl,j (N ) (N ) and Cj = C5,l,j = Cw (N ) (N ) (L) (N ) + Al,j C2,l (Al,j )T ; consequently, the particle generation mechanism for xl+1 requires computing the Cholesky de(N ) composition of a single matrix (namely, C5,l ), since it is based on the modified message     (N ) (N ) (N ) (N ) (78) m ~ 5 xl+1 = N xl+1 ; η5,l,j , C5,l , 24 G. M. Vitetta et al., Marginalized Particle Filtering . . . (N ) which depends on the particle index j through the mean vector η5,l,j only. (L) Finally, in the time update for xl , the pdf     (L) (L) (L) (N ) (L) ) f zl xl = N zl ; Āl xl , C(N w (L) (79) (N ) is employed in the evaluation of m ~ 3,j (xl ) through (50), where Āl (N ) = Al (N ) (x̄l/l ) (N ) (N ) and x̄l/l denotes the center of mass of the particle set Sl/l (see (67)). Then, (L) (L) the messages m ~ 3,j (xl ) (55) and m ~ 4,j (xl ) (57) can be replaced by     (L) (L) (N ) (L) ) = N zl,j ; Āl xl , C(N m ~ 3,j xl w (80) and     (L) (L) (L) (L) = N xl ; η4,l,j , C4,l , m ~ 4,j xl (81) respectively, with (see (58) and (59)) and  T (L) (L) (L) (L) (N ) (N ) (L) w4,l,j , W4,l η4,l,j = w2,l + Āl Ww zl,j (82)  −1  T (L) (L) (L) (N ) (N ) (N ) W4,l , C4,l = W2,l + Āl Ww Āl . (83) (L) (L) (N ) Substituting now m ~ 4,j (xl ) (81) (in place of m ~ 4,j (xl ) (57)) and m ~ 2,j (xl ) (32) in (60), produces, after some manipulation     (L) (L) (L) (L) (84) m ~ 5,j xl+1 = N xl+1 ; η5,l,j , C5,l,j , where (L) (L) (L) (L) η5,l,j , Al,j η4,l,j + fl,j and (L) (L) (L) C5,l,j , C(L) w + Al,j C4,l (L)  (L) Al,j (85) T . (86) Note that: a) since the precision matrix W4,l (83) is particle-independent, a (L) single matrix inversion is needed to evaluate the matrix C4,l appearing in (86); (L) (L) b) in case # 2 the matrix set {C5,l,j } is condensed in a single matrix C5,l (this (L) (L) represents the common value C̃(l+1)/l taken on by all the matrices {C(l+1)/l,j } (L) processed in the next recursion); c) the matrix C5,l is evaluated on the basis of (L) (L) (69) and (70), setting ηj = η5,l,j and Cj = C5,l,j for any j. The new filtering techniques, based on MPF and on the simplified messages derived above, are called simplified MPF (SMPF) in the following; in particular the acronyms SMPF1 and SMPF2 are used to refer to case #1 and case #2, respectively. 25 G. M. Vitetta et al., Marginalized Particle Filtering . . . 6 Message Passing in Iterative Filtering Techniques Inspired by Marginalized Particle Filtering As already mentioned above, the suboptimality of MPF can related to the fact that the FG underlying the considered filtering problem is not cycle free. It is well known that the SPA can also be applied to a factor graph with cycles simply by following the same message propagation rules; however, generally speaking, this leads to an “iterative” algorithm with no natural termination (and known as loopy belief propagation), since its messages are passed multiple times on a given edge [20], [21]. Despite this, some of the most relevant applications of the SPA have been developed for systems in which the underlying FG does have cycles, like the one shown in Fig. 1. In the following we show how a novel iterative technique can be developed for our filtering problem following this approach. To begin, we note that our interest in iterative methods is also (N ) motivated by the possibility of exploiting the pseudo-measurement zl (18); in (N ) fact, the message mj (zl ) referring to this random vector cannot be computed in MPF because of the adopted scheduling, but can certainly provide additional information to refine our statistical knowledge about the nonlinear component (N ) (L) xl . This message, similarly as the one referring to10 zl , can be put in a Gaussian form, that is     (N ) (N ) (N ) (N ) = N zl ; ηz,l,j , Cz,l,j , (87) m ~ j zl (L) (L) (N ) since xl+1 and xl , conditioned on xl , are modelled as jointly Gaussian random vectors. Let us show now how this message can be computed in an iterative filtering procedure generalising MPF and how it can exploited in such (L) a procedure. First of all, we assume that the message m ~ 5,j (xl+1 ) (62), repre(L) (N ) (N ) senting the pdf of xl+1 conditioned on xl = xl/l,j , is already available when (L) the time update for xl (i.e., the first step of MPF) is accomplished. Then, (L) (L) (N ) given m ~ 2,j (xl ) (40) and m ~ 5,j (xl+1 ) (62), the mean and covariance of zl can be evaluated as (see (18)) (N ) (L) (L) (L) ηz,l,j = η5,l,j − Al,j η2,l,j 10 If (N) xl (N) (L) = xl/l,j , the random vector zl (N) (N) (88) (N) (13) becomes xl+1 − fl (N) (xl/l,j ); consequently, (L) adopting the Gaussian model (45) for xl+1 results in a Gaussian model for zl too. 26 G. M. Vitetta et al., Marginalized Particle Filtering . . . and  T (N ) (L) (L) (L) (L) Cz,l,j = C5,l,j + Al,j C2,l,j Al,j  T  T (L) (L) (L) (L) , −Al,j Cx,l,j − Cx,l,j Al,j (89) (L) respectively, where Cx,l,j denotes the cross covariance matrix for the vectors (L) xl (L) (N ) and xl+1 (conditioned on xl (L) (L) (L) (N ) ~ 2,j (xl ) (40) and the = xl/l,j ). Given m (L) (N ) (L) (L) (L) (L) conditional pdf f (xl+1 |xl , xl/l,j ) = N (xl+1 ; fl,j + Al,j xl , Cw ) (see (17)), (L) (L) (L) it is easy to show that Cx,l,j = C2,l,j (Al,j )T (e.g., see [32, Par. 2.3.3, eq. (2.104)]); consequently, eq. (89) can be rewritten as  T (N ) (L) (L) (L) (L) Cz,l,j = C5,l,j − Al,j C2,l,j Al,j . (90) Equations (88) and (90) represent the desired result, since they provide a com(N ) plete statistical characterization of the message m ~ j (zl ) (87). In principle, (L) this message could be exploited in a similar way as that adopted for m ~ j (zl ) (N ) (53); this approach would lead to draw a set of Np samples {zl,j } from the Gaussian function appearing in the RHS of (87) and to process the resulting pseudo-measurements to generate a new weight for each particle of the set Sl/l . However, our computer simulations have shown that this approach is outperformed by more refined method illustrated in the following. In practice, in the (N ) iterative filtering method we propose the message m ~ j (zl ) (87), similarly as (L) m ~ j (zl ) (53), is employed to evaluate the new message11       Z (N ) (N ) (N ) (N ) (N ) (91) f zl xl/l,j dzl , = m ~ j zl m ~ 3,j xl (N ) (L) which represents for xl the counterpart of the message m ~ 3,j (xl ) (50). Sub(N ) (N ) (N ) (N ) stituting (87) and the expression of f (zl |xl ) (given xl = xl/l,j ) in the RHS of the last expression gives12 (see the Appendix)   (N ) m ~ 3,j xl   T (N ) (N ) (N ) (N ) = D3,l,j · exp 21 η3,l,j W3,l,j η3,l,j (92)   T  T (N ) (N ) (N ) (L) (L) (L) − ηz,l,j Wz,l,j ηz,l,j − fl,j Ww fl,j , pl,j , 11 Note (N) that the following message represents the correlation between the pdf m ~ j (zl evaluated on the basis of the definition of (N) zl ) (18) and the pdf originating from the fact that (L) (N) (L) this quantity is expected to equal the random vector fl (xl/l,j ) + wl . For this reason, it expresses the degree of similarity between these two functions. 12 In our computer simulations the factor D (N) appearing in this weight has been always 3,l,j neglected, since it negligibly influences estimation accuracy. 27 G. M. Vitetta et al., Marginalized Particle Filtering . . . where  −1 (N ) (N ) (N ) (L) W3,l,j , C3,l,j = Wz,l,j + Ww , (N ) (N ) (N ) (N ) (93) (L) (L) w3,l,j , W3,l,j η3,l,j = wz,l,j + Ww fl,j , (N ) (N ) (N ) (N ) (N ) (L) (94) (L) Wz,l,j , (Cz,l,j )−1 , wz,l,j , Wz,l,j ηz,l,j , Ww , [Cw ]−1 , (N ) (N ) (L) h  i−DL /2 (N ) (N ) D3,l,j , det C̃l,j (95) (N ) and C̃l,j , Cz,l,j + Cw . Then, the new message m ~ 3,j (xl similarly as ) (92) is exploited, (L) m ~ 3,j (xl ) (N ) (55), to generate the message (see (51))       (N ) (N ) (N ) , m ~ 3,j xl =m ~ 2,j xl m ~ 4,j xl (96) where m ~ 2,j (xl ) is expressed by (28) (i.e., it is the message emerging from the (N ) measurement update for xl in the absence of resampling). This produces (see (92))     (N ) (N ) (N ) (97) = Wl,j δ xl − xl/(l−1),j , m ~ 4,j xl where Wl,j , wl,j · pl,j (98) represents the new weight for the j-th particle; such a weight accounts for both (N ) the (real) measurement yl and the pseudo-measurement zl,j . Resampling with (N ) (N ) replacement can now be accomplished for the set Sl/(l−1) , {xl/(l−1),j } on the basis of the more refined weights {Wl,j } (98); if this is done, the message (N ) (N ) m ~ 4,j (xl ) takes on the same form as m ~ 2,j (xl ) (32), i.e. it can be expressed as     (N ) (N ) (N ) (99) = δ xl − x̃l/l,j , m ~ 4,j xl (N ) with j = 0, 1, ..., Np − 1. Note that, generally speaking, the particle set S̃l/l , (N ) (N ) (N ) {x̃l/l,j } produced by resampling in this case is different from Sl/l , {xl/l,j } (i.e., from the one obtained with MPF), even if both of them originate from the (N ) same set Sl/(l−1) ; this is due to the fact that the weights {Wl,j } (98) may be substantially different from the MPF weights {wl,j } because of the factor pl,j . (N ) (N ) Finally, the new message m ~ 4,j (xl ) (99) is used in place of m ~ 2,j (xl ) in the (L) RHS of (43) for the evaluation of the message m ~ 5,j (xl+1 ). As already state above, our previous derivations rely on the assumption that (L) (L) the message m ~ 5,j (xl+1 ) (62) is available when the time update for xl is carried out; unluckily, this message becomes available only in the last step of MPF. However, if MPF is generalised in a way that Nit > 1 iterations (i.e., message 28 G. M. Vitetta et al., Marginalized Particle Filtering . . . passes) are carried out within the same recursion, in the k-th iteration (with k = (N ) 2, 3, ..., Nit ) the message mj (zl ) (87) can be really evaluated exploiting the (L) message m ~ 5,j (xl+1 ) computed in the previous iteration. These considerations lead, in a natural fashion, to the development of the message passing illustrated in Fig. 4, which describes the message flow occurring in the k-th iteration of a new filtering technique, generalising MPF and called turbo filtering (TF) in the following; note that the superscripts (k) or (k − 1) have been added to all the messages flowing in the considered graph to identify the iteration in which they are generated and that the grey circle appearing in the figure represents a unit delay cell. The processing tasks accomplished by TF can be summarized as follows. The first part of this technique can be considered as a (N ) (L) form of initialization, in which the messages {m ~ 2,j (xl )} and {m ~ 2,j (xl )} are computed; however, unlike MPF, resampling is not accomplished in the last part (N ) (N ) of the time update for xl (so that the messages {m ~ 2,j (xl )} are expressed (N ) by (28) instead of (32) and refer to the particle set Sl/(l−1) ). Moreover, as (N ) (L) shown in Fig. 4, the messages {m ~ 2,j (xl )} and {m ~ 2,j (xl )} emerging from the first part of TF remain unchanged within the l-th recursion, since they (N ) (L) represent the a priori information available about xl and xl , respectively; consequently, like in any turbo processing method, these information are made available to all the iterations carried out within each recursion. In the second part of TF Nit -iterations are accomplished with the aim of progressively refining (N ) the set S(l+1)/l (the version of this set generated in the k-th iteration is denoted (k) (N ) (L) ~ 5,j (xl+1 )}. To achieve S(l+1)/l [k]) and the associated Gaussian messages {m these result, in the k-th iteration (with k = 1, 2, ..., Nit ) the ordered computation (k) (N ) (k) (N ) ~ 3,j (xl ) (92) of the following messages is accomplished: m ~ j (zl ) (87), m (k) (N ) (conveying the weight pl,j [k]), m ~ 4,j (xl (k) (N ) (L) (k) (L) ) (97) (conveying the weight Wl,j [k]), (k) (L) (k) (L) ~ 5,j (xl+1 ) ~ 4,j (xl ) (57) and m ~ j (zl ) (53), m ~ 3,j (xl ) (55), m m ~ 5,j (xl+1 ) (45), m 13 (62). Moreover, in the k-th iteration resampling is accomplished on the basis of the particle weights {Wl,j [k]}; generally speaking, this results in a new particle (N ) (N ) set denoted Sl/l [k] = {xl/(l−1),j [k], j = 0, 1, ..., Np − 1} (which is always a (N ) subset of Sl/(l−1) ). It is also important to point out that: (0) (N ) • In the first iteration (i.e, for k = 1) the messages {m ~ 5,j (xl+1 )} are un(L) 13 Note that, after carrying out resampling, the set of messages {m ~ 2,j (xl )} needs to be properly reordered, since the messages associated with the discarded particles are not preserved. This modifies the set of particles available in the next iteration and, consequently, the set of weights {wl,j } associated with them (these weights need to be renormalized after any change). In the following the notation {wl,j [k]} is adopted to denote the set of weights employed in the k-th iteration for the evaluation of the overall weights Wl,j [k] according to (98). 29 G. M. Vitetta et al., Marginalized Particle Filtering . . . (1) (N ) (1) (N ) ~ 3,j (xl ) = defined, so that {m ~ j (zl ) = 1} (and, consequently, {m 1}) must be assumed; moreover, resampling is not accomplished (i.e., (N ) (N ) (N ) Sl/l [1] = Sl/l = Sl/(l−1) ), since the weights pl,j appearing in the overall weight Wl,j (98) become available in the following iterations. − (x(L) ) = ← − (x(N ) ) = 1 is assumed for • In each iteration the equality ← m m he l+1 he l+1 (L) the two messages entering the FG along the half edges associated with xl+1 (N ) and xl+1 (see Fig. 4), since no information comes from the next recursion. For this reason, at the end of the last iteration (i.e., for k = Nit ), the output messages (i.e., the input messages feeding the (l + 1)-th recursion) are evaluated as (see (48)-(49) and (62))       (N ) (N ) (Nit ) (N ) (N ) − (x(N ) ) = m x (100) m ~ ~ 5,jit xl+1 ← m ~ out,j xl+1 = m he l+1 5,j l+1 and       (L) (L) (N ) (L) (N ) − (x(L) ) = m m ~ 5,jit xl+1 , ~ 5,jit xl+1 ← m ~ out,j xl+1 = m he l+1 (101) with j = 0, 1, ..., Np − 1. Another relevant issue concerns the interpretation of the processing tasks accomplished in the TF technique. In fact, our derivations show that, at the end of the k-th iteration, the a posteriori statistical information about the j-th (k) (N ) (N ) (N ) ~ 4,j (xl ) (97), which particle xl/l,j [k] of Sl/l [k] is provided by the message m conveys the weight (see (98)) (a) Wl,j [k] = wl,j [k] · pl,j [k] · wl,j , (102) (a) where wl,j denotes the a priori information available at the beginning of the l(a) th recursion for the j-th particle ( in our derivation wl,j = 1 has been assumed, (a) in place of wl,j = 1/Np , to simplify the notation; see (22)), pl,j [k] is the weight (N ) (k) (N ) originating from m ~ j (zl ) (87) and conveyed by m ~ 3,j (xl ) (92), and wl,j [k] is the weight computed on the basis of the available measurement yl . Taking the natural logarithm of both sides of (102) produces (a) (y) (z) Ll,j [k] = Ll,j + Ll,j [k] + Ll,j [k], (y) (z) (103) (a) where Ll,j [k] , ln(Wl,j [k]), Ll,j [k] , ln(wl,j [k]), Ll,j , ln(pl,j [k]) and Ll,j , (a) ln(wl,j ). The last equation has exactly the same structure as the well known formula (see [35, Sec. 10.5, p. 450, eq. (19.15)] or [36, Par. II.C, p. 432, eq. (20)]) L (uj |y) = L (uj ) + Lc (yj ) + Le (uj ) (104) 30 G. M. Vitetta et al., Marginalized Particle Filtering . . .
min , j ( x (l L ) ) x (l L ) yl
m1, j ( x (l L ) ) f y ,/ x( L ) ,x( N ) l +1 = = l

m2, j ( x l( L ) ) = l m4,( k j) ( x (l L ) )
mhe ( xl(+L1) ) = 1 f x ( L ) / x ( L ) ,x ( N )
m2,( k j) ( x (l L ) )
min , j ( x (l L ) ) = l
m2, j ( x l( L ) ) m5,( kj) ( x (l +L1) )
f z( L ) / x ( L )
m (jk ) ( z l( L ) )
mout , j ( x l(+L1) ) l =
m3,( kj) ( x (l L ) ) l
min , j ( x (l N ) ) l x (l +L1) l
m2, j ( x l( N ) ) z (l L ) x (l N ) , x l( +N1) → z l( L ) m5,( kj) ( x (l +N1) )
m2, j ( x (l N ) )
m2, j ( x l( N ) ) =
m2, j ( x l( L ) ) = m2, j ( x  (N ) l )
m (jk ) ( z l( N ) ) f y ,/ x( L ) ,x( N ) l
m3,( kj) ( x l( N ) ) l = =
min , j ( x (l N ) ) =  ( k −1) ( L ) z (l N ) m ( xl +1 ) 5, j f z( N ) / x ( N ) l l yl
min , j ( x (l N ) )
m2, j ( x l( N ) )
m2,( k j) ( x l( L ) )
m5,( kj) ( x l(+N1) ) = =
m2, j ( x l( N ) ) x (l N )
m5,( kj) ( x (l +L1) ) x (l L ) , x (l +L1) → z (l N )
min , j ( x (l L ) ) l
m2, j ( x l( L ) )
m4,( k j) ( x l( N ) ) f x ( N ) / x ( N ) ,x ( L ) l +1 l l
mhe ( x l(+N1) ) = 1
mout , j ( x (l +N1) ) x (N ) l +1 Figure 4: Message passing over the FG of Fig. 1 for the proposed TF technique. All the quantities appearing in this figure refer to the k-th iteration of the l-th recursion; the messages available at the beginning of the considered iteration are indicated by red arrows, those entering the graph by blue arrows, those leaving it at the end of the last iteration by black arrows and those computed within the considered iteration by green arrows. 31 G. M. Vitetta et al., Marginalized Particle Filtering . . . expressing of the log-likelihood ratio (LLR) available for the j-th information bit uj at the output of a soft-input soft-output channel decoder operating over an additive white Gaussian noise (AWGN) channel and fed by: a) the channel output vector y (whose j-th element yj is generated by the communications channel in response to a channel symbol conveying uj and is processed to produce the so-called channel LLR Lc (yj )); b) the a priori LLR L (uj ) about uj ; c) the extrinsic LLR Le (uj ), i.e. a form of soft information available about uj , but intrinsically not influenced by such a bit (in turbo decoding of concatenated channel codes extrinsic infomation is generated by another channel decoder with which soft information is exchanged with the aim of progressively refining data estimates). This correspondence is not only formal, since in eqs. (103) and (104) terms playing similar roles can be easily identified. For instance, (y) (a) the term Ll,j [k] (Ll,j ) in (103) provides the same kind of information as Lc (yj ) (L (uj )), since these are both related to the noisy data (a priori information) available about the quantities to be estimated (the system state in one case, an (z) information bit the in the other one). What about the term Ll,j [k] appearing in the RHS of (103)? The link we have established between (103) and (104) unavoidably leads to the conclusion that such a term should represent the counterpart of the quantity Le (uj ) appearing in (104), i.e. the so called extrinsic (N ) information (in other words, that part of the information available about xl (N ) and not intrinsically influenced by xl itself). This interpretation is confirmed (z) by the fact that Ll,j [k] is computed on the basis of the statistical knowledge (L) (L) available about xl and xl+1 (see (88) and (90)), which, thanks to (2), does (N ) provide useful information about xl . The theory of turbo decoding of channel codes shows that, generally speaking, extrinsic information originates from code constraints. In our scenario, a similar interpretation can be also provided for (z) (L) Ll,j [k], since xl+1 can be seen as the noisy output of a communication channel, (L) (L) affected by the bias fl,j and the additive noise wl , and over which the code(L) (L) word Al,j xl of a rate-1 block code is transmitted in response to the message (L) xl . (z) These considerations show that, in evaluating Ll,j [k], we are actually exploiting a sort of ‘code’ constraints, which are mathematically expressed by (2). The reader can easily verify that a similar interpretation can be provided (L) for m ~ j (zl ) (53), which represents the extrinsic information component14 con(L) (L) tained in m ~ 4,j (xl ) (57) (conveying our a posteriori information about xl ); (L) the other two components are represented by the message m ~ 2,j (xl ) (40) (repre14 In (L) practice, the mechanism employed to generate zl,j is based on the state update equation (3). G. M. Vitetta et al., Marginalized Particle Filtering . . . (L) 32 (L) senting the measurement information about xl ) and the message m ~ in,j (xl ) (L) (21) (corresponding to our a priori information about xl ). Consequently, TF can be seen, in the domain of Bayesian filtering techniques, as the counterpart of turbo decoding of concatenated codes; this parallelism can be exploited to provide further insights into iterative filtering techniques. For instance, it is well known that, in turbo decoding of concatenated channel codes, the extrinsic information generated by soft decoders become more and more correlated as iterations evolve; this entails that diminishing benefits are provided by additional iterations. This phenomenon should be observed in TF too for similar reasons (N ) (N ) and can be motivated by rewriting ηz,l,j (88) and Cz,l,j (90) as (N ) (L) (L) ηz,l,j = fl,j + Al,j and (N ) (L) Cz,l,j = C(L) w + Al,j h h i (L) (L) η4,l,j − η2,l,j (L) (L) C4,l,j − C2,l,j i T (L) , Al,j (105) (106) respectively (thanks to (85) and (86), respectively). In fact, the last two equa(N ) (N ) tions show that the vector ηz,l,j and the matrix Cz,l,j are influenced by the difference between the statistical information (expressed by a mean vector and (L) a covariance matrix) available about xl before processing the pseudomeasure(L) ment zl and those available after this task has been carried out. In other (L) (N ) words, the extrinsic information provided by zl influences zl and viceversa. Finally, it is important to point out that the proposed analogy between turbo filtering and turbo decoding suggests the potential limits of the TF technique (and of any other iterative filtering method relying on the developed FG). In fact, it is well known that turbo decoding methods do not provide real benefits below a certain signal-to-noise ratio, i.e. when the quality of the received signal is so poor that the transmitted coded sequence cannot be recovered. A similar phenomenon is expected occur with TF too; consequently, this filtering method could not outperform MPF in the presence of strong measurement noise and/or fast dynamics affecting the considered SSM. 7 Numerical Results In this Section MPF and the related filtering methods developed in this manuscript are compared in terms of accuracy and computational load for a specific CLG system, characterized by DL = 3, DN = 1 (so that D = 4) and P = 2. The structure of the considered system has been partly inspired by the example G. M. Vitetta et al., Marginalized Particle Filtering . . . proposed by Schön  0.8 (L) xl+1 =  0 0 and (L) in [37] and is characterized by: a) the state models    (N ) cos(xl ) 0.2 0   (L) (L) ) 0.7 −0.2  xl +  − sin(x(N )  + wl l (N ) 0.2 0.7 0.5 sin(2xl )   (N ) (N ) (L) (N ) xl+1 = arctan xl + (0.9 0 0) xl + wl (L) (N ) 33 (107) (108) (N ) ∼ N (0, (σw )2 I3 ), wl ∼ N (0, (σw )2 ; b) the measurement model     !   (N ) 2 (N ) 0 0 0 0.1 xl · sgn xl (L) yl = + xl + el (109) 1 −1 1 0 with wl with el ∼ N (0, (σe )2 I2 ). Note that the state equation (107), unlike itscounter (N ) (L) (N ) part proposed in [37], depends on xl (i.e., it contains a function fl xl+1 6= (N ) 03 in its RHS), so that TF, which relies on the availability of the vector zl (18), can be employed for this system. In our computer simulations the root mean square error (RMSE) has been evaluated to compare the accuracy of the state estimates generated by different filtering techniques. More specifically, for each technique two RMSEs have been computed, one (denoted RM SEL (alg), where ‘alg’ denotes the algorithm this parameter refers to) representing the square root of the average mean square (L) error (MSE) evaluated for the three elements of xl , the other one (denoted (N ) RM SEN (alg)) referring to the (monodimensional) nonlinear component xl ; this distinction is important since, as shown by our simulation results, the es(L) timation accuracy for xl can be quite different from (and is usually smaller (N ) than) that referring to xl . As far the assessment of the computational requirements of the investigated filtering techniques is concerned, MPF (for which an accurate analysis of its computational complexity is available in [16]) has been taken as a baseline. For this reason, our comparisons between the considered filtering techniques are based on the evaluation of a single parameter, denoted ∆c (alg) and representing the percentage variation in computation time of the considered algorithm (denoted ‘alg’) with respect to MPF (operating with the same parameters and, in particular, with the same Np as alg). Moreover, in our computer simulations, the following choices have been (N ) (L) made: a) σw = σw = 5 · 10−3 has been selected for the standard devia(L) (N ) tions of the process noises {wk } and {wk }, unless differently stated; b) the so called jittering technique [34] has been employed to mitigate the so called depletion problem in the generation of new particles. 34 G. M. Vitetta et al., Marginalized Particle Filtering . . . Some results illustrating the dependence of RM SEL and RM SEN on the number of particles (Np ) for MPF and SMPF#1 are illustrated in Fig. 5 (in this and in the following figures simulation results are identified by markers, whereas continuous lines are drawn to ease reading); σe = 10−2 and Np ∈ [50, 300] have been selected in this case. From these results the following conclusions can be easily inferred for the considered system: 1. A negligible improvement in the estimation accuracy of both MPF and SMPF#1 is achieved if the value of Np exceeds 200. 2. A significant gap between RM SEL (MPF) and RM SEN (MPF) exists; this (L) is motivated by the fact that, in MPF, the estimation of xl relies on both (L) the real measurement yl (109) and the pseudo-measurements {zl,j } (54), (N ) whereas the estimation of xl benefits from yl only. 3. The gap between RM SEL (SMPF#1) and RM SEN (SMPF#1) is much smaller than that observed for MPF, even if the pseudo-measurements (L) {zl,j } are also exploited by SMPF#1. This reduction in the RMSE gap (L) can be related to the degradation in the estimation accuracy of xl ; for instance, for Np = 200, RM SEL (SMPF#1) is about twice RM SEL (MPF) (on the contrary, RM SEN (SMPF#1) ∼ = 1.09· RM SEN (MPF)). Our numerical results have also evidenced that: a) SMPF#1 requires a substantially smaller computational effort than MPF, since ∆c (SMPF#1) approximately ranges in the interval [−62%, −58%] for the considered values of Np ; b) despite the above mentioned degradation in estimation accuracy (see point 3.), SMPF#1 does not suffer from tracking losses in the considered scenario; c) SMPF#2 accuracy is almost indentical that of SMPF#1; d) ∆c (SMPF#2) is approximately lower than ∆c (SMPF#1) by 6% in the considered range for Np and, consequently, achieves a better complexity-performance tradeoff than SMPF#1. All this suggests that simplified MPF techniques can be really developed without incurring the serious technical problems that affect the techniques proposed in [17] (tracking losses and poor RMSE performance) and in [18] (partitioning and update of the particle set into groups within each recursion; see Section 5). The dependence of RM SEL and RM SEN on σe (i.e., on the intensity of the noise affecting the available measurements) has been also analysed for MPF and SMPF#1. Some results are shown in Fig. 6; Np = 500 and σe ∈ [10−3 , 5 · 10−2 ] have been selected in this case. These results show that the gap between RM SEL (MPF) and RM SEL (SMPF#1) slightly increases as σe becomes 35 G. M. Vitetta et al., Marginalized Particle Filtering . . . 5 4 SMPF1 RMSE 3 -2 2x10 RMSEL RMSEN MPF 0.01 50 100 150 200 250 300 Np Figure 5: RMSE performance versus Np for the linear component (RM SEL ) and the nonlinear component (RM SEN ) of the state xl for the system described (L) by eqs. (107)-(109). MPF and SMPF#1 are considered; in both cases σw = (N ) σw = 5 · 10−3 and σe = 10−2 have been selected. 36 G. M. Vitetta et al., Marginalized Particle Filtering . . . 6 SMPF1 5 4 3 RMSE 2 MPF 0.01 9 8 7 RMSEL RMSEN 6 5 10 20 30 40 -3 50x10 σe Figure 6: RMSE performance versus σe for the linear component (RM SEL ) and the nonlinear component (RM SEN ) of the state xl for the system described by (L) eqs. (107)-(109). MPF and SMPF#1 are considered; in both cases σw = (N ) σw = 5 · 10−3 and Np = 200 have been selected. smaller; the opposite occurs for RM SEN (MPF) and RM SEN (SMPF#1). These (L) results can be related again to the fact that the pseudo-measurements {zl,j } (L) (54) really play a role in the MPF estimation of xl and that the quality of the information conveyed by these pseudo-measurements indirectly improves as the real measurements become less noisy. Some results illustrating the dependence of RM SEL and RM SEN on the number of particles (Np ) for MPF and TF are illustrated in Fig. 7; σe = 10−2 , Np ∈ [1, 300] and Nit = 2 for TF, and Np ∈ [30, 300] for MPF have been chosen in this case. From these results it is easily inferred that: 1. TF outperforms MPF in tems of both RM SEL and RM SEN ; for instance, RM SEL (MPF) ∼ = 1.71· RM SEL (TF) and RM SEN (MPF) ∼ = 2.86· RM SEN (TF) if Np = 200 is selected. G. M. Vitetta et al., Marginalized Particle Filtering . . . 37 2. The gap between RM SEL (MPF) and RM SEN (MPF) is substantially larger than the corresponding gap for TF (in particular, RM SEN (MPF)− RM SEL (MPF) ∼ = 13.4·[RM SEN (TF)−RM SEL (TF)]; this can be easily (L) related to the fact that, unlike MPF, in TF the estimation of both xl (N ) and xl benefits from the availability of pseudo-measurements. 3. The performance gap between TF and MPF increases as Np gets smaller. Moreover, TF accuracy starts quickly degrading when Np drops below 11, whereas the same phenomenon starts for Np ∼ = 30 with MPF. Note, however, that different phenomena occur with MPF and TF when the particle set is small. In fact, our computer simulations have shown that MPF suffers from frequent tracking losses for Np < 30 (since it is unable (N ) to generate a reliable representation of xl when a limited number of particles is available); on the contrary, TF does not suffer from the same problem even if a very small particle set is used. We believe that last result is really important and can be motivated as follows. The TF technique, through its feedback mechanism, makes a substantially more efficient use of the available particles than MPF; this results in an appreciable improvement of both stability and accuracy of state estimation. Our simulation results have also shown that: a) in the considered scenario a negligible improvement in estimation accuracy is obtained if Nit > 2 is selected; b) ∆c (TF) approximately ranges in the interval [70%, 80%] for the considered values of Np and, consequently, requires a substantially larger computational effort than MPF if these algorithms operate with the same number of particles. In practice, however, as evidenced by the results shown in Fig. 7, TF can reliably operate with a very small particle set and, consequently, it outperforms MPF in terms of both performance and complexity if Np is properly selected. For instance, in the considered scenario TF with Np = 20 particles achieves a substally better accuracy than MPF with Np = 40 particles, even if, as evidenced by our computer simulations, they approximately require the same computation time. It is not difficult to show that similar considerations hold if SMPF#1 and SMPF#2 are considered in place of MPF. Therefore, our results suggest that the real key to the complexity reduction of the filtering techniques relying on the FG shown in Fig. 2 is not provided by the approximations adopted for the MPF processing tasks in Section 5, but by the exploitation of (N ) the new pseudo-measurement zl (18). We should never forget, however, that TF cannot   be adopted for all CLG systems; in fact, it cannot be employed if (L) (N ) fl xl+1 = 0DL in the RHS of (2). 38 G. M. Vitetta et al., Marginalized Particle Filtering . . . 5 4 RMSEL RMSEN MPF 3 RMSE 2 0.01 9 8 7 TF 6 5 0 50 100 150 Np 200 250 300 Figure 7: RMSE performance versus Np for the linear component (RM SEL ) and the nonlinear component (RM SEN ) of the state xl for the system described (L) (N ) by eqs. (107)-(109). MPF and TF are considered; in both cases σw = σw = −3 −2 5 · 10 and σe = 10 have been selected. 39 G. M. Vitetta et al., Marginalized Particle Filtering . . . 8 7 6 5 RMSEL RMSEN σw = 5 10-3 σw = 10-3 MPF 4 3 RMSE 2 0.01 8 7 6 5 TF 4 3 -2 3 2x10 4 σe Figure 8: RMSE performance versus σe for the linear component (RM SEL ) and the nonlinear component (RM SEN ) of the state xl for the system described by eqs. (107)-(109). MPF and TF are considered; σe = 10−2 and Np = 200 are (L) assumed in both cases. As far system noise is concerned, the cases σw = (N ) (L) (N ) σw = 5 · 10−3 and σw = σw = 10−1 are taken into consideration. Finally, the dependence of RM SEL and RM SEN on σe has been assessed for TF and compared with that characterizing MPF. Some numerical are illustrated in Fig. 8. In this case, we have selected σe ∈ [1.5 · 10−2 , 5 · 10−2 ] , Np = 200 (L) (N ) and Nit = 2 for TF, and Np = 200 for MPF; moreover, σw = σw = 5 · 10−3 (L) (N ) and σw = σw = 10−3 have been considered to analyse the dependence of the performance gap between TF and MPF on the intensity of process noise (and, consequently, on system dynamics). These results show that: a) the performance gap between TF and MPF undergoes small changes if σe varies in the considered interval; b) on the contrary, a substantial change in this gap (L) (N ) is obtained if σw and σw are reduced from 5 · 10−3 to 10−3 . This suggests that measurement noise and process noise can have different impacts on the performance gain provided by TF over MPF. 5 G. M. Vitetta et al., Marginalized Particle Filtering . . . 8 40 Conclusions In this manuscript a FG approach has been employed to analyse the filtering problem for mixed linear/nolinear models. This has allowed us to: a) prove that this problem involves a FG which is not cycle free; b) provide a new interpretation of MPF as a forward only message passing algorithm over a specific FG; c) develop novel filtering algorithms for simplifying or generalising it. In particular, an important iterative filtering technique, dubbed turbo filtering, has been devised and its relation with the turbo decoding techniques for concatenated channel codes has been analysed in detail. All the considered filtering techniques have been compared in terms of both accuracy and computational requirements for a specific CLG system. The most interesting result emerging from our computer simulations is represented by the clear superiority of turbo filtering over marginalized particle filtering. In fact, the former technique, through the exploitation of new pseudo-measurements, can achieve a better accuracy than the latter one at an appreciably smaller computational load. Our ongoing research activities in this area include the development of other related filtering techniques and the application of turbo filtering to specific state estimation problems. A Appendix Given the pdfs f1 (y) , N (y; η1 , C1 ) and f2 (y) , N (y; η1 , C2 ) for the N dimensional vector y, we are interested in evaluating the correlation between these two functions, i.e. the quantity Z c1,2 , f1 (y) · f2 (y) dy (110) Substituting the expressions of f1 (y) and f2 (y) in the RHS of the last equation produces, after some manipulation,  
1 T T T (111) η Wη − η1 W1 η1 − η2 W2 η2 c1,2 = D exp 2 −1 where W1 , C−1 1 , W 2 , C2 , W = W1 + W2 , (112) Wη = W1 η1 + W2 η2 (113) D = (2π det [C1 + C2 ])−N/2 . (114) and G. M. Vitetta et al., Marginalized Particle Filtering . . . 41 References [1] M. S. Arulampalam, S. Maskell, N. Gordon and T. Clapp, “A Tutorial on Particle Filters for Online Nonlinear/Non-Gaussian Bayesian Tracking”, IEEE Trans. Sig. Proc., vol. 50, no. 2, pp. 174-188, Feb. 2002. [2] F. E. Daum, “Exact Finite-Dimensional Nonlinear Filters”, IEEE Tran. Aut. Contr., vol. 31, no. 7, pp. 616-622, July 1986. [3] S. Mazuelas, Y. Shen and M. Z. Win, “Belief Condensation Filtering”, IEEE Trans. Sig. Proc., vol. 61, no. 18, pp. 4403-4415, Sept. 2013. [4] V. Smidl and A. Quinn, “Variational Bayesian Filtering”, IEEE Trans. Sig. Proc., vol. 56, no. 10, pp. 5020-5030, Oct. 2008. [5] M. S̆imandl, J. Královeca and T. Söderströmc, “Advanced Point-Mass Method for Nonlinear State Estimation”, Automatica, vol. 42, pp. 11331145, 2006. [6] B. Anderson and J. Moore, Optimal Filtering, Englewood Cliffs, NJ, Prentice-Hall, 1979. [7] S. J. Julier and J. K. Uhlmann, “Unscented Filtering and Nonlinear Estimation”, IEEE Proc., vol. 92, no. 3, pp. 401-422, Mar. 2004. [8] A. Doucet, J. F. G. de Freitas and N. J. Gordon, “An Introduction to Sequential Monte Carlo methods,” in Sequential Monte Carlo Methods in Practice, A. Doucet, J. F. G. de Freitas, and N. J. Gordon, Eds. New York: Springer-Verlag, 2001. [9] A. Doucet, S. Godsill and C. Andrieu, “On Sequential Monte Carlo Sampling Methods for Bayesian Filtering”, Statist. Comput., vol. 10, no. 3, pp. 197-208, 2000. [10] F. Gustafsson, “Particle Filter Theory and Practice with Positioning Applications”, IEEE Aerosp. and Electr. Syst. Mag., vol. 25, no. 7, pp. 53-82, July 2010. [11] F. Gustafsson, F. Gunnarsson, N. Bergman, U. Forssell, J. Jansson, R. Karlsson and P. Nordlund, “Particle Filters for Positioning, Navigation, and Tracking”, IEEE Trans. Sig. Proc., vol. 50, 425-435, 2002. G. M. Vitetta et al., Marginalized Particle Filtering . . . 42 [12] P. J. Nordlund and F. Gustafsson, “Marginalized Particle Filter for Accurate and Reliable Terrain-Aided Navigation”, IEEE Trans. on Aerosp. and Elec. Syst., vol. 45, no. 4, pp. 1385-1399, Oct. 2009. [13] R. Bucy and K. Senne, “Digital Synthesis of Non-Linear Filters”, Automatica, vol. 7, no. 3, pp. 287-298, 1971. [14] F. Daum and J. Huang, “Curse of Dimensionality and Particle Filters”, Proc. IEEE Aerosp. Conf., vol. 4, pp. 1979-1993, March 2003. [15] T. Schön, F. Gustafsson, P.-J. Nordlund, “Marginalized Particle Filters for Mixed Linear/Nonlinear State-Space Models”, IEEE Trans. Sig. Proc., vol. 53, no. 7, pp. 2279-2289, July 2005. [16] R. Karlsson, T. Schön, F. Gustafsson, “Complexity Analysis of the Marginalized Particle Filter”, IEEE Trans. Sig. Proc., vol. 53, no. 11, pp. 4408-4411, Nov. 2005. [17] F. Mustiere, M. Bolic and M. Bouchard, “A Modified Rao-Blackwellised Particle Filter”, Proc. of the 2006 IEEE Int. Conf. on Ac., Sp. and Sig. Proc. (ICASSP 2006), vol. 3, 14-19 May 2006. [18] T. Lu, M. F. Bugallo and P. M. Djuric, “Simplified Marginalized Particle Filtering for Tracking Multimodal Posteriors”, Proc. IEEE/SP 14th Workshop on Stat. Sig. Proc. (SSP ’07), pp. 269-273, Madison, WI (USA), 2007. [19] F. Lindsten, P. Bunch, S. Särkkä, T. B. Schön and S. J. Godsill, “RaoBlackwellized Particle Smoothers for Conditionally Linear Gaussian Models”, IEEE J. Sel. Topics in Sig. Proc., vol. 10, no. 2, pp. 353-365, March 2016. [20] H.-A. Loeliger, J. Dauwels, Junli Hu, S. Korl, Li Ping, F. R. Kschischang, “The Factor Graph Approach to Model-Based Signal Processing”, IEEE Proc., vol. 95, no. 6, pp. 1295-1322, June 2007. [21] F. R. Kschischang, B. Frey, and H. Loeliger, “Factor Graphs and the SumProduct Algorithm”, IEEE Trans. Inf. Theory, vol. 41, no. 2, pp. 498-519, Feb. 2001. [22] C. Berrou and A. Glavieux, “Near Optimum Error Correcting Coding and Decoding: Turbo-Codes”, IEEE Trans. Commun., vol. 44, no. 10, pp. 1261 - 1271, Oct. 1996. G. M. Vitetta et al., Marginalized Particle Filtering . . . 43 [23] S. Benedetto, D. Divsalar, G. Montorsi and F. Pollara, “Serial Concatenation of Interleaved Codes: Performance Analysis, Design, and Iterative Decoding”, IEEE Trans. Inf. Theory, vol. 44, no. 3, pp. 909-926, May 1998. [24] R. Koetter, A. C. Singer and M. Tüchler, “Turbo Equalization”, IEEE Sig. Proc. Mag., vol. 21, no. 1, pp. 67-80, Jan. 2004. [25] J. Hagenauer, “The Turbo Principle: Tutorial Introduction & State of the Art”, Proc. Int. Symp. Turbo Codes & Related Topics, Brest, France, Sep. 1997, pp. 1-11. [26] A. P. Worthen and W. E. Stark, “Unified Design of Iterative Receivers using Factor Graphs”, IEEE Trans. Inf. Theory, vol. 47, no. 2, pp. 843–849, Feb. 2001. [27] F. R. Kschischang and B. J. Frey, “Iterative Decoding of Compound Codes by Probability Propagation in Graphical Models”, IEEE J. Sel. Areas Commun., vol. 16, no. 2, pp. 219-230, Feb. 1998. [28] J. Dauwels, S. Korl and H.-A. Loeliger, “Particle Methods as Message Passing”, Proc. 2006 IEEE Int. Symp. on Inf. Theory, pp. 2052-2056, 9-14 July 2006. [29] T. P. Minka, “Expectation Propagation for Approximate Bayesian Inference”, Proc. 17th Annual Conf. Uncertainty in Artif. Intell., pp. 362-369, Seattle, WA (USA), Aug. 2001. [30] O. Zoeter and T. Heskes, “Deterministic Approximate Inference Techniques for Conditionally Gaussian State Space Models”, Statistics and Computing, vol. 16, no. 3, pp. 279-292, Sep. 2006. [31] V. Smı́dl and A. Quinn, The Variational Bayes Method in Signal Processing, Berlin, Germany, Springer, 2005. [32] C. M. Bishop, Pattern Recognition and Machine Learning, Springer, 2006. [33] A. R. Runnalls, “Kullback-Leibler Approach to Gaussian Mixture Reduction”, IEEE Trans. on Aer. and Elec. Syst., vol. 43, no. 3, pp. 989-999, July 2007. [34] T. Li, M. Bolic, P. Djuric, “Resampling Methods for Particle Filtering: Classification, Implementation, and Strategies”, IEEE Sig. Proc. Mag., vol. 32, no. 3, pp.70-86, May 2015. G. M. Vitetta et al., Marginalized Particle Filtering . . . 44 [35] G. M. Vitetta, D. P. Taylor, G. Colavolpe, F. Pancaldi and P. A. Martin, Wireless Communications: Algorithmic Techniques, John Wiley & Sons, 2013. [36] J. Hagenauer, E. Offer and L. Papke, “Iterative decoding of binary block and convolutional codes”, IEEE Trans. Inf. Theory, vol. 42, no. 2, pp. 429-445, Mar 1996. [37] T. Schön, “Example Used in Exemplifying the Marginalized (Rao-Blackwellized) Particle Filter”, Nov. 2010 (available at http://users.isy.liu.se/en/rt/schon/Code/RBPF/Document/MPFexample.pdf). 

10math.ST

ASYMPTOTIC GROWTH OF ASSOCIATED PRIMES OF CERTAIN GRAPH IDEALS arXiv:1204.3315v2 [math.AC] 5 Dec 2012 SARAH WOLFF Abstract. We specify a class of graphs, Ht , and characterize the irreducible decompositions of all powers of the cover ideals. This gives insight into the structure and stabilization of the corresponding associated primes; specifically, providing an answer to the question “For each integer t ≥ 0, does there exist a (hyper) graph Ht such that stabilization of associated primes occurs at n ≥ (χ(Ht ) − 1) + t?” [4]. For each t, Ht has chromatic number 3 and associated primes that stabilize at n = 2 + t. 1. Introduction We work at the intersection of graph theory and commutative algebra by relating powers of square-free monomial ideals in polynomial rings to vertex covers of finite simple graphs. Ideal constructions give two correspondences, both of which associate to each such graph a monomial ideal. These correspondences have been established and studied in, for example, [9], [8], [3], and [4]. Motivated by the Strong Perfect Graph Theorem, Francisco, Hà, and Van Tuyl [3] expand the link between graph theory and algebra in order to provide an algorithm for determining if a graph is perfect— neither it nor its complementary graph has an induced odd cycle of length five or greater. Their algorithm results from relating odd cycles to the associated primes of powers of monomial ideals. Francisco, Hà, and Van Tuyl continue to investigate this connection in [4], establishing a lower bound on the stabilization of these associated primes that is derived from the chromatic number of the corresponding graphs. They find that stabilization does not occur when n < χ(G) − 1, for χ(G) the chromatic number of the graph and n the power of the ideal [4, Corollary 4.9]. They further show that this bound is not optimal, leading to the question: “For each integer t ≥ 0, does there exist a (hyper)graph Ht such that the stabilization of associated primes occurs Date: March 24, 2018. 2000 Mathematics Subject Classification. 05C25, 05E40, 05C38, 13F20. 1 2 SARAH WOLFF at n ≥ (χ(Ht ) − 1) + t?” [4, Question 4.10]. In this paper we answer this question in the affirmative. More precisely, let G be a graph with vertex set VG = {v1 , . . . , vm } and edge set EG consisting of unordered pairs of distinct vertices of G. Let K be a field. We identify the vertices of G with the variables in the polynomial ring RG = K[v1 , . . . , vm ]. Further, we associate two square-free monomial ideals to G: the edge ideal, IG , generated by {vi vj | {vi , vj } ∈ EG }, and the cover ideal, JG , generated by {vi1 · · · vij | {vi1 , . . . , vij } a vertex cover of G}. These ideal constructions were first introduced in [9] and give the correspondences mentioned above. Francisco, Hà, and Van Tuyl study the associated primes of RG /(JG )2 , denoted Ass(RG /(JG )2 ), and find that P is in Ass(RG /(JG )2 ) if and only if either P = (vi , vj ) where {vi , vj } is an edge of G, or P = (vi1 , . . . , vis ) where the subgraph induced by {vi1 , . . . , vis } is an odd cycle of G [3, Corollary 3.4]. These results are derived by determining the irreducible decomposition of (JG )2 , i.e., writing (JG )2 as an intersection of irreducible monomial ideals. They then extend these ideas in [4] by studying Ass(RG /(JG )n ) for n ≥ 2. Brodmann [2] in 1979 proved that the set of associated primes of powers of ideals stabilizes, i.e., for a ring R and an ideal I of R, there exists some positive integer s such that Ass(R/I s ) = Ass(R/I n ) for all n ≥ s. Francisco, Hà, and Van Tuyl [4] find a lower bound on the stabilization of Ass(RG /(JG )n ), proving for a graph G that stabilization does not occur when n < χ(G) − 1 [4, Corollary 4.9]. They then pose the question mentioned above [4, Question 4.10], which we answer in the affirmative by providing a family of graphs, Ht , each with chromatic number three such that for each t, Ass(RHt /(JHt )n ) stabilizes at n = (χ(Ht ) − 1) + t = 2 + t. The main result of this paper is: Theorem 4.1. For t ∈ Z+ , let Rt := RHt be the polynomial ring associated to Ht and let Jt := JHt denote the cover ideal of Ht . Then the associated primes of Rt /(Jt )n stabilize at n = (χ(Ht )−1)+t = 2+t. As in [3], we characterize the associated primes of Rt /(Jt )n by studying the monomial ideal powers (Jt )n . Just as the cover ideal is generated by vertex covers of the graph, we find that the generators of the ideals in the irreducible decompositions of (Jt )n are closely related to vertex covers. In Section 2 we develop the necessary tools from algebra and graph theory for our results. In Section 3 we introduce the family of graphs and explore the structure of these graphs with several key lemmas. In Section 4 we use these lemmas to characterize the irreducible ASSOCIATED PRIMES OF CERTAIN GRAPH IDEALS 3 decompositions of (Jt )n (Theorem 4.11). Theorem 4.1 is then a direct consequence of Theorem 4.11. 2. Graph Theory and Associated Primes In this section we briefly introduce the important definitions from graph theory and algebra needed for our results. We use the basic conventions, notation, and definitions from [1], [3], and [7]. For a graph G we enumerate its vertex set VG = {v1 , . . . , vm }. We assume that G is a simple graph so that its edge set, EG , consists of unordered pairs {vi , vj } of distinct vertices. Two vertices are adjacent if they are joined by an edge. A coloring of G is an assignment of colors to the vertices of G such that no two adjacent vertices are assigned the same color. The chromatic number of G is the minimal number of colors in a coloring of G. A subset A of VG is a vertex cover if every edge of G is incident to at least one vertex of A. The subset is a minimum vertex cover if it is a vertex cover of smallest cardinality. In this paper we consider minimum vertex covers rather than minimal vertex covers as this distinction is necessary for the proof of Theorem 4.11. For a subset A of VG , the neighbors of A, denoted N (A), are the vertices of G that are adjacent to vertices of A, but do not lie in A. The subgraph of G induced by A is the subgraph with vertex set A and edge set consisting of those edges of G with both vertices in A. A graph with vertex set {w1 , . . . , wk } is a k-cycle if its edge set consists precisely of edges connecting wk to w1 and wi to wi+1 , 1 ≤ i < k. We call the cycle odd (respectively, even), if k is odd (respectively, even). Let K be a field. We identify the vertices of G with variables in the polynomial ring RG = K[v1 , . . . , vm ]. Let IG denote the edge ideal of G, and let JG denote the cover ideal of G, generated by {vi1 · · · vij | {vi1 , . . . , vij } a vertex cover of G}. We note that JG is often equivalently defined as the Alexander dual of IG (see, e.g., [3], [4]). For an explanation of Alexander duality, see [7]. A typical monomial element of the cover ideal JG is of the form a1 a2 am v1 v2 · · · vm for nonnegative integers ai . Just as the variables v1 , . . . , vm correspond to vertices of the graph G, the vector (a1 , . . . , am ) also has a graph theoretic interpretation. We refer to any such a = (a1 , . . . , am ) (corresponding to the vertices v1 , . . . , vm of G) as a degree vector and say a nonzero degree vector a is a k-cover if ai + aj ≥ k whenever {vi , vj } ∈ EG . 4 SARAH WOLFF The concept of k-cover was first introduced in [6] as a generalization of a vertex cover. Note that if a subset A of VG is a minimum vertex cover then a degree vector a with ai = 1 if vi ∈ A and ai = 0 otherwise is a one-cover of G. We call such a one-cover a minimum one-cover. am ) An irreducible monomial ideal is an ideal of the form V a = (v1a1 , . . . , vm ai n for a ∈ N , where if ai = 0, we omit vi from the set of generators. Every monomial ideal I can be written as the finite intersection of irreducible monomial ideals: I = V a1 ∩ · · · ∩ V as , called an irreducible decomposition of I [7]. The decomposition is called irredundant if no V ai can be omitted. Let R be a commutative ring and M an R-module. The set of associated primes of M , denoted Ass(M ), consists of prime ideals P such that P annihilates some m in M . T For a graph G, the cover ideal JG has irreducible decomposition JG = {vi ,vj }∈EG (vi , vj ) (see e.g., [3]); hence, Ass(R/JG ) = {(vi , vj ) | {vi , vj } ∈ EG }. Generalizing to (JG )n , we compute the irreducible decompositions of (JG )n for our family of graphs in order to deduce the set of associated primes of R/(JG )n . We refer to this set as the associated primes of G. 3. A Family of Graphs We proceed to define a family of graphs, Ht , to give an affirmative answer to the question posed in [4, Question 4.10]: “For each integer t ≥ 0, does there exist a (hyper)graph Ht such that the stabilization of associated primes occurs at n ≥ (χ(Ht ) − 1) + t?” Note by [3, Remark 3.3, Corollary 3.4] that the associated primes of an odd cycle, G, stabilize at n = χ(G) − 1 + 0 = 2. We thus answer the above question for t ≥ 1. Definition 3.1. The graph H1 has vertex set {x1 , . . . , x5 , y1 } and edge set such that the subgraph induced by {x1 , . . . , x5 } is a 5-cycle and the neighbors of y1 are precisely {x1 , x2 , x3 }. For t > 1, Ht has vertex set {x1 , . . . , x4t−1 , y1 , . . . , yt } and edge set such that the subgraph induced by {x1 , . . . , x4t−1 } is a (4t − 1)-cycle, the neighbors of y1 are precisely {x1 , x2 , x3 }, and for 1 < i ≤ t the neighbors of yi are precisely {x4i−4 , x4i−3 , x4i−2 , x4i−1 } (Figure 1). We first note that the graphs Ht have chromatic number three for all t ≥ 1. To see this, assign color 1 to the vertices {x1 , x3 } ∪ {yi | i > 1}, color 2 to the vertices {x2 } ∪ {x2k+1 | k > 1}, and color 3 to the remaining vertices. Each Ht has at least one odd cycle; thus, this 3-coloring is a minimum coloring. ASSOCIATED PRIMES OF CERTAIN GRAPH IDEALS H1 H2 x2 H3 x1 x1 x1 y1 x2 y1 x2 x3 x3 x3 x4 x4 x4 x5 x5 x6 x7 5 y2 x5 x6 y1 y2 ... x7 x8 x9 x10 y3 x11 Figure 1. Ht Let Jt := JHt be the cover ideal of Ht for each t ≥ 1 and let Rt := RHt be the polynomial ring associated to Ht . We show that Ass(Rt /(Jt )n ) stabilizes at n = 2 + t. This result will follow from a characterization of the irreducible decompositions of (Jt )n for all t ≥ 1 and n ≥ 1. For this characterization, we use the structure of the graphs together with the following technical lemmas, all of which follow easily from the definitions: Lemma 3.2. Let G be a graph with vertex set {v1 , . . . , vm }. Let RG be the polynomial ring associated to G, and let JG be the cover ideal of G. If M is a minimal generator of (JG )n , then \ M∈ (vin , vj ) ∩ (vin−1 , vj2 ) ∩ · · · ∩ (vi2 , vjn−1 ) ∩ (vi , vjn ). {vi ,vj }∈EG Lemma 3.3. Let G be a graph with vertex set {v1 , . . . , vm } and cover am ideal JG . Then the mononomial v1a1 · · · vm is an element of (JG )n if and only if a = (a1 , . . . , am ) is an n-cover of G that can be written as the sum of n one-covers, ai , of G: a = a1 + · · · + an . We note that a degree vector, a, for a graph G can also be interpreted as a degree vector for a subgraph of G induced by A ⊆ VG by considering only the ai corresponding to the vertices of A. We denote this subgraph degree vector by a|A . Lemma 3.4. For a graph G, let a be a degree vector for G. Let G1 , . . . , Gl be induced subgraphs of G such that EG ⊆ EG1 ∪ · · · ∪ EGl . 6 SARAH WOLFF Then a is a one-cover of G iff a|Gi is a one-cover of Gi for all i, 1 ≤ i ≤ l. Our interest is in studying (Jt )n . By Lemma 3.3, elements of (Jt )n correspond to n-covers of Ht . These n-covers can be written as the sum of n one-covers of Ht , which in turn arise from one-covers of induced subgraphs, by Lemma 3.4. Since minimum vertex covers give rise to one-covers, it will be informative to look at the structure of minimum vertex covers of induced subgraphs of our family Ht . Lemma 3.5. A minimum vertex cover of an odd k-cycle consists of k+1 vertices. 2 Lemma 3.6. For the graph Ht , a minimum vertex cover of the induced subgraph on {y1 } ∪ N (y1 ) consists of the vertices {x2 , y1 }. For i > 1, a minimum vertex cover of the induced subgraph on {yi } ∪ N (yi ) consists of exactly 3 vertices, one of which is yi . 4. Stability of Associated Primes We restate our main theorem. Recall that Jt is the cover ideal of Ht and Rt the polynomial ring associated to Ht . Theorem 4.1. Let t ∈ Z+ . The associated primes of Rt /(Jt )n stabilize at n = (χ(Ht ) − 1) + t = 2 + t. This result will follow from the determination of the irreducible decompositions of (Jt )n in Theorem 4.11. Each component of the irreducible decomposition of (Jt )n is an irreducible monomial ideal of the form (via11 , . . . , viakk ), for vij vertices of Ht (either x or y vertices). The powers of the variables are in fact determined by the structure of the subgraph induced by {vi1 , . . . , vik }. Thus, we abuse terminology in the following way: for an ideal (via11 , . . . , viakk ) ⊆ Rt , we refer to each variable vij as a vertex, each power aj as a degree, and a = (a1 , . . . , ak ) as a degree vector. For any simple graph G with vertex set {v1 , . . . , vm }, let JG be the cover ideal. Then by [3, Theorem 3.2], the irredundant irreducible decomposition of (JG )2 is: \ \ \ (JG )2 = [(vi2 , vj ) ∩ (vi , vj2 )] (vi21 , . . . , vi2k ), {vi ,vj }∈EG {vi1 ,...,vik } an odd cycle where the last intersection is over all subsets {vi1 , . . . , vik } such that the subgraph induced by {vi1 , . . . , vik } is an odd cycle. For our family of graphs, we generalize the above irreducible decomposition of (Jt )2 to one for (Jt )n , n > 2, by inductively adding ASSOCIATED PRIMES OF CERTAIN GRAPH IDEALS 7 a minimum one-cover’s worth of degrees to each degree vector appearing in the irreducible decomposition of (Jt )n−1 . In other words, if V c = (vic11 , . . . , vickk ) is an ideal in the irreducible decomposition of (Jt )n−1 , and b is a minimum one-cover of the subgraph induced by {vi1 , . . . , vik }, we show that V c+b is an ideal in the irreducible decomposition of (Jt )n . To be more explicit, if the induced subgraph on {vi1 , vi2 , vi3 } ⊆ VHt is a 3-cycle, then by [3, Theorem 3.2], (vi21 , vi22 , vi23 ) is a component of the irreducible decomposition of (Jt )2 . We lift this to (Jt )n by showing that (via11 , via22 , via33 ) is a component of the irreducible decomposition of (Jt )n whenever a = (a1 , a2 , a3 ) is a degree vector for the induced subgraph on {vi1 , vi2 , vi3 } obtained by adding (n − 2) minimum one-covers of this subgraph to the degree vector c = (2, 2, 2). This results in many different possibilities for a: by Lemma 3.5, a minimum one-cover of the subgraph is of form (0, 1, 1),(1, 0, 1),
or (1, 1, 0). Thus, adding (n − 2) minimum one-covers to c leads to n2 possibilities for a. Rather than writing out all these possibilities, we introduce below the terminology n-admissible and n̂-admissible to distinguish degree vectors of specific induced subgraphs of Ht . As the irreducible decompositions of (Jt )n are known for n ≤ 2, we work with n > 2 in the following definitions. Consider an arbitrary subset of vertices, V = {vi1 , . . . , vik } ⊆ VHt . If the subgraph induced by V is an odd cycle, we call V an induced odd cycle. Consider the corresponding degree vector a = (a1 , . . . , ak ). Definition 4.2. A degree vector a corresponding to an induced odd cycle V is n-admissible if a = (2, . . . , 2) + b1 + · · · + bn−2 where each bi is a minimum one-cover of the cycle. Remark 4.3. Note by Lemma 3.5 that for V an induced odd k-cycle, if the corresponding degree vector a is n-admissible, then k X i=1 ai = 2k + (n − 2) k + 1
. 2 Now fix a positive integer r ≤ t and consider a subgraph induced by vertices V ∪Yr = {vi1 , . . . , vik , yj1 , . . . , yjr } where V ⊆ VHt is an induced odd cycle, Yr ⊆ {y1 , . . . , yt } (the specified y-vertices in Definition 3.1), and N (Yr ) ⊆ V . We call such a subgraph an r-cluster, or an r-cluster induced by V ∪ Yr , when we want to specify the vertices. Note that in the definition of r-cluster, V can be an induced odd cycle of any length on any of the vertices of Ht , so long as N (Yr ) ⊆ V . On the other hand, Yr must contain exactly r vertices, and they must be y-vertices. 8 SARAH WOLFF Recall that we assume n > 2 for the following definition. Definition 4.4. A degree vector c corresponding to an r-cluster induced by V ∪ Yr , is n̂-admissible if c = d + e + f1 + · · · + fn−3 , where:  2 vji ∈ V \ N (Yr ) (1) di = , 3 otherwise (2) e|Yr ∪N (Yr ) = (0, . . . , 0), (3) e|V \N (Yr ) is a minimum one-cover of the subgraph induced by V \ N (Yr ), (4) for all 1 ≤ i ≤ n − 3, fi |V \N (Yr ) is a minimum one-cover of the subgraph induced by V \ N (Yr ), (5) for all 1 ≤ i ≤ n − 3 and yjl ∈ Yr , fi |{yjl }∪N (yjl ) is a minimum one-cover of the subgraph induced by {yjl } ∪ N (yjl ). Example 4.5. Let n = 5. Consider the following 2-cluster of H4 : the subgraph induced by V ∪ Y2 = {x1 , x2 , x3 , x4 , y2 , x7 , x8 , y3 , x11 , x12 , x13 , x14 , x15 , y1 , y4 }. Then d = (3, 3, 3, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3), one of the two possibilities for e is (0, 0, 0, 0, 1, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0) and two of the six possibilities for each fi , i ∈ {1, 2}, are (0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 1) and (0, 1, 0, 0, 1, 0, 1, 0, 1, 0, 1, 1, 0, 1, 1). Thus, the degree vector c = (3, 5, 3, 3, 4, 3, 4, 3, 4, 4, 4, 5, 3, 5, 5) is 5̂-admissible. Remark 4.6. Consider the subgraph induced by {x1 , x2 , x3 , y1 }. By Lemma 3.6, the minimum vertex cover of this subgraph is {x2 , y1 }. Thus, for an arbitrary r-cluster, induced by V ∪ Yr , if y1 ∈ Yr , the corresponding degrees for x2 and y1 would be 1 in each fj . By the same lemma, for i > 1, a minimum vertex cover of the subgraph induced by {yi } ∪ N (yi ) consists of yi and two vertices in N (yi ); thus, if yi ∈ Yr , then in each fj the corresponding degrees for yi and two vertices in N (yi ) would be 1. The following lemma is key to the proof of Theorem 4.11. Lemma 4.7. Let r ≥ 1 and n > r + 1. Consider an arbitrary r-cluster of Ht , induced by V ∪ Yr . Let c = (c1 , . . . , ck+r ) be a degree vector for
Pk+r the r-cluster. If c is n̂-admissible then i=1 ci < r + k + n k+1 +r . 2 Proof. Consider an arbitrary r-cluster of Ht , induced by V ∪ Yr , and suppose the corresponding degree vector, c, is n̂-admissible. We compute the degree sum imposed by the definition of n̂-admissible. This degree sum depends on whether y1 ∈ Yr , or not. ASSOCIATED PRIMES OF CERTAIN GRAPH IDEALS 9 First suppose y1 ∈ Yr . If c is n̂-admissible, then c = d + e + f1 + · · · + fn−3 , with each degree vector as specified in Definition 4.4. We first note that k+r X di = 12 + 15(r − 1) + 2(k − 4r + 1). i=1 Next, we claim k+r X ei ≤ i=1 k − 4r + 1 . 2 To see this, we find a bound on the number of vertices in a minimum vertex cover of the subgraph induced by V \ N (Yr ). First note that k − 4r + 1 vertices comprise V \ N (Yr ). Further, the subgraph of Ht induced by V \ N (Yr ) consists of disjoint subgraphs of V induced by one or more subsets V1 , . . . , Vl ⊂ V . Let vα be the number of vertices in Vα , 1 ≤ α ≤ l. If vα is even, a minimum vertex cover of the subgaph induced by Vα consists of v2α vertices, while if vα is odd, a minimum vertex cover consists of vα2−1 vertices. Hence, k+r X i=1 l X vα − 1 X vα 1X k − 4r + 1 ei = + ≤ vα = , 2 2 2 2 v even α=1 v odd α α as claimed. Finally, by Lemma 3.6 and the above claim we see that for each fj , 1 ≤ j ≤ n − 3, k+r X fji ≤ 2 + 3(r − 1) + i=1 Since the degree sum of fj , k+r X i=1 ci = k+r X Pk+r i=1 di + i=1 k − 4r + 1 . 2 fji , is independent of j, we have: k+r X ei + (n − 3) i=1 k+r X f ji , i=1 for any j ∈ {1, . . . , n − 3}. The inequality of the lemma then reduces to showing that r + 1 < n, which is true by assumption. Next suppose y1 ∈ / Yr . Here, k+r X di = 15r + 2(k − 4r), i=1 and we claim k+r X i=1 ei ≤ k − 4r − 1 . 2 10 SARAH WOLFF To see this, note that in this case k − 4r vertices comprise V \ N (Yr ). As above, the subgraph of Ht induced by V \ N (Yr ) consists of disjoint subgraphs of V induced by subsets V1 , . . . , Vl ⊂ V . Further, because V is an induced odd cycle, k − 4r is odd, so at least one of the subsets, without loss V1 , has an odd number of vertices, v1 . Hence, as above, k+r X i=1 l X vα − 1 X vα X
k − 4r − 1 1 ei = + ≤ v1 − 1 + vα = , 2 2 2 2 v even α=2 v odd α α as claimed. Finally, by Lemma 3.6 and the above claim we see that for each fj , 1 ≤ j ≤ n − 3, k+r X k − 4r − 1 fji ≤ 3r + . 2 i=1 Pk+r Again, the degree sum of fj , i=1 fji , is independent of j, and the inequality of the lemma reduces to showing r + 1 < n.  We now define several ideals to simplify the statement of Theorem 4.11. Fix a graph Ht and a positive integer n. For an arbitrary set of vertices V = {vi1 , . . . , vik } of Ht with corresponding degree vector a = (a1 , . . . , ak ), let V a denote the ideal (via11 , . . . , viakk ). Again, we a remind the reader of our convention to omit vijj when aj = 0. Definition 4.8. \ At,n = (vin , vj ) ∩ (vin−1 , vj2 ) ∩ · · · ∩ (vi2 , vjn−1 ) ∩ (vi , vjn ). {vi ,vj }∈EHt \ Definition 4.9. Bt,n = V a. subgraph of Ht induced by V an odd cycle, a n-admissible r Definition 4.10. Dt,n = \ (V ∪ Yr )c . subgraph of Ht induced by V ∪Yr an r-cluster, c n̂-admissible We now give the irreducible decompositions of (Jt )n : Theorem 4.11. Let t ∈ Z+ and let Jt be the cover ideal of Ht . For n > 2, (Jt )n has the following irredundant irreducible decomposition: n (Jt ) = At,n ∩ Bt,n ∩ n−2 \ r=1 r Dt,n ASSOCIATED PRIMES OF CERTAIN GRAPH IDEALS 11 T r Proof. For t ∈ N and n > 2, let L be the ideal At,n ∩ Bt,n ∩ n−2 r=1 Dt,n . Let M be a minimal generator of (Jt )n . We show M is in L by showing Tn−2 r separately. it is in each of the ideals At,n , Bt,n , and r=1 Dt,n First suppose M ∈ / Bt,n . Since M is a minimal generator of the am , for {v1 , . . . , vm } the vertices of monomial ideal (Jt )n , M = v1a1 · · · vm Ht . By Lemma 3.3 the powers, a = (a1 , . . . , am ), when viewed as a degree vector of Ht form an n-cover that can be written as the sum of n one-covers, bi , of Ht : a = b1 + · · · + bn . Let V = {vj1 , . . . , vjk } be an arbitrary induced odd cycle of Ht . Since a is an n-cover of Ht , a|V is an n-cover of the odd cycle. Further, for 1 ≤ i ≤ n, since bi is a one-cover of Ht , bi |V is a one-cover of the cycle. Then a|V = b1 |V + · · · + bn |V with each bi |V a one-cover of the subgraph induced by V . Thus, a|V = (aj1 , . . . , ajk ) is an n-cover of the cycle that can be written as the sum
P of n one-covers. Then by Lemma 3.5, ki=1 aji ≥ n k+1 . However, 2 M ∈ / Bt,n , so M ∈ / V u for some n-admissible u. Thus, in M each variable vij ∈ V has degree less than uj , so k X i=1 aj i ≤ k X (ui − 1) = k + (n − 2) i=1 k + 1
, 2

k+1 by Remark 4.3. But k + (n − 2) k+1 < n , so a|V is not an n2 2 cover of the cycle that can be written as the sum of n one-covers, a contradiction. Thus, M ∈ Bt,n . T a1 r am Now suppose M ∈ / n−2 r=1 Dt,n . Again, M = v1 · · · vm is a minimal generator of (Jt )n so the powers, a = (a1 , . . . , am ), when viewed as a degree vector of Ht form an n-cover that can be written as the sum of n one-covers, bi , of Ht : a = b1 + · · · + bn . For r ≤ n − 2, consider an arbitrary r-cluster of Ht , induced by V ∪ Yr . As above, a|V ∪Yr is an n-cover of the r-cluster and for 1 ≤ i ≤ n, bi |V ∪Yr is a one-cover of the r-cluster, so a|V ∪Yr = (al1 , . . . , alk+r ) is an n-cover of the r-cluster that as
the sum of n one-covers. Tn−2 r Then by Lemma 3.6, Pk+rcan be written k+1 + r . However, M ∈ / r=1 Dt,n , so M ∈ / (V ∪ Yr )c , i=1 ali ≥ n 2 for some n̂-admissible c. Thus, as in the previous case, k+r X i=1 ali ≤ k+r X (ci − 1). i=1 But since n ≥ r + 2, Lemma 4.7 shows k+r X i=1 (ci − 1) < n
k+1 +r , 2 12 SARAH WOLFF so a|V ∪Yr is not an n-cover of the r-cluster that can be written as the Tn−2 r sum of n one-covers, a contradiction. Thus, M ∈ r=1 Dt,n . We finally note by Lemma 3.2 that M ∈ At,n . Hence, (Jt )n ⊆ L. Now let P be a minimal generator of L. The ideal L is an intersection of monomial ideals, so P = via11 · · · viamm . By definition, P ∈ At,n ; therefore, the powers a = (a1 , . . . , am ), of the variables of P when viewed as a degree vector of Ht form an n-cover. By Lemma 3.3, it suffices to show a can be written as the sum of n one-covers of Ht . Consider an arbitrary one-cluster of Ht , induced by V ∪ Y1 . Note that Ht can be written as a union of one-clusters. Thus, by Lemma 3.4, it suffices to show that a|V ∪Y1 can be written as a sum of n one-covers of the one-cluster. Applying Lemma 3.4 again to the subgraph induced by V ∪ Y1 = {vi1 , . . . , vik , yj1 }, it suffices to show a|V and a|{yj1 }∪N (yj1 ) can be written as a sum of n one-covers of the subgraphs induced by V and {yj1 } ∪ N (yj1 ), respectively. Let V be an induced odd k-cycle of Ht . Since P ∈ Bt,n , P ∈ V u for some n-admissible u = (u1 , . . . , uk ). Thus, we need to show that u can be written as the sum of n one-covers of V . By Definition 4.2, u = (2, . . . , 2) + b1 + · · · + bn−2 with each bi a one-cover of V . Thus, we need only show that (2, . . . , 2) can be written as a sum of two onecovers of V , but this is clear because (2, . . . , 2) = (1, . . . , 1) + (1, . . . , 1), which are trivially one-covers of V . Now consider the subgraph induced by {yj1 } ∪ N (yj1 ). Since P ∈ Tn−2 r c r=1 Dt,n , P ∈ (V ∪ Y1 ) , for some n̂-admissible c. Thus, we need to show c|{yj1 }∪N (yj1 ) can be written as a sum of n one-covers of the subgraph induced by {yj1 } ∪ N (yj1 ). By Definition 4.4, c = d + e + f1 + · · · + fn−3 with each fi a one-cover of the subgraph induced by {yj1 } ∪ N (yj1 ). Further, di = 3 for each vertex in {yj1 } ∪ N (yj1 ), so we need only show that (3, . . . , 3) can be written as a sum of 3 one-covers of the subgraph induced by {yj1 } ∪ N (yj1 ), but this is clear by Lemma 3.6. Thus, for an arbitrary one-cluster, induced by V ∪ Y1 , a|V ∪Y1 can be written as the sum of n one-covers of the subgraph induced by V ∪ Y1 . This proves that a can be written as the sum of n one-covers of Ht . Hence, L ⊆ (Jt )n .  Our main theorem now follows as a corollary to Theorem 4.11: Proof of Theorem 4.1. For t ∈ Z+ consider the graph Ht and label the vertices VHt = {v1 , . . . , vm } = X ∪ Y , where X (respectively, Y ) represents the set of x-vertices (respectively, y-vertices) as in Definition ASSOCIATED PRIMES OF CERTAIN GRAPH IDEALS 13 3.1 . By Theorem 4.11, for n > 2 it follows that a prime P is in Ass(Rt /(Jt )n ) if and only if: (a) P = (vi , vj ) where {vi , vj } ∈ EHt , (b) P = (vi1 , . . . , vik ), where the subgraph induced by {vi1 , . . . , vik } is an odd cycle, or, (c) For 1 ≤ r ≤ n − 2, P = (vi1 , . . . , vik , yj1 , . . . , yjr ), where the subgraph induced by {vi1 , . . . , vik , yj1 , . . . , yjr } is an r-cluster. By definition, Ht has exactly t y-vertices, so Ht has a t-cluster, namely the graph Ht itself. Thus, by (c) above, the ideal (v1 , . . . , vm ) is an element of Ass(Rt /(Jt )n ) for n ≥ t + 2, but not for n < t + 2. Thus, the associated primes do not stabilize until at least t+2. Further, note that the ideals of (a) and (b) are elements of Ass(Rt /(Jt )n ) for all n ≥ 2, so the only prime ideals not in Ass(Rt /(Jt )2+t ) that could appear in Ass(Rt /(Jt )n ) for n > 2 + t would be ideals that fall into category (c). However, for 1 ≤ r ≤ t, ideals whose generators correspond to the vertices of r-clusters are elements of Ass(Rt /(Jt )2+t ), so the only new ideals in Ass(Rt /(Jt )n ) for n > 2 + t would be those whose generators correspond to the vertices of r-clusters for r > t. The largest r for which Ht has an r-cluster is r = t; thus, no such new ideals exist in Ass(Rt /(Jt )n ) for n > 2 + t. Thus, the associated primes of Ht stabilize at 2 + t.  Remark 4.12. We note that our family of graphs can be generalized slightly. The proof of Theorem 4.11 used the definition n̂-admissible in two ways— to show that the degree vectors can be written as n onecovers, and to use the degree argument of Lemma 4.7. Both facts remain unchanged if the induced (4t − 1)-cycle is instead an odd cycle of any length greater than 4t − 1 (or in the case of H1 , if the 5-cycle is any odd cycle of length greater than 5). Further, for i > 1, yi could have more than 4 neighbors, as long as these neighbors are distinct from N (yj ), j 6= i, and the degree argument of Lemma 4.7 is still satisfied. 5. Acknowledgements Many of the results of this paper started as computer experiments using the program Macaulay 2 [5]. The author would like to thank Amelia Taylor for introducing her to the subject and for her invaluable support and guidance throughout the research and writing processes, Thomas Shemanske for his help in the editing process, and Adam Van Tuyl for a helpful discussion of Lemma 3.3. We also thank an anonymous referee for their insights and suggestions for improvement. 14 SARAH WOLFF References 1. Béla Bollobás, Modern graph theory, Graduate Texts in Mathematics, vol. 184, Springer-Verlag, New York, 1998. MR 1633290 (99h:05001) 2. M. Brodmann, Asymptotic stability of Ass(M/I n M ), Proc. Amer. Math. Soc. 74 (1979), no. 1, 16–18. MR 521865 (80c:13012) 3. Christopher A. Francisco, Huy Tài Hà, and Adam Van Tuyl, Associated primes of monomial ideals and odd holes in graphs, J. Algebraic Combin. 32 (2010), no. 2, 287–301. MR 2661419 (2011e:05133) , Colorings of hypergraphs, perfect graphs, and associated primes of 4. powers of monomial ideals, J. Algebra 331 (2011), 224–242. MR 2774655 (2012b:13054) 5. Daniel R. Grayson and Michael E. Stillman, Macaulay2, a software system for research in algebraic geometry, Available at http://www.math.uiuc.edu/Macaulay2/. 6. Jürgen Herzog, Takayuki Hibi, and Ngô Viêt Trung, Symbolic powers of monomial ideals and vertex cover algebras, Adv. Math. 210 (2007), no. 1, 304–322. MR 2298826 (2007m:13005) 7. Ezra Miller and Bernd Sturmfels, Combinatorial commutative algebra, Graduate Texts in Mathematics, vol. 227, Springer-Verlag, New York, 2005. MR 2110098 (2006d:13001) 8. Aron Simis, Wolmer V. Vasconcelos, and Rafael H. Villarreal, On the ideal theory of graphs, J. Algebra 167 (1994), no. 2, 389–416. MR 1283294 (95e:13002) 9. Rafael H. Villarreal, Cohen-Macaulay graphs, Manuscripta Math. 66 (1990), no. 3, 277–293. MR 1031197 (91b:13031) 

0math.AC

arXiv:1705.04282v1 [cs.CV] 5 May 2017 Learning to see people like people Amanda Song University of California, San Diego 9500 Gilman Dr, La Jolla, CA 92093 Linjie Li Purdue University 610 Purdue Mall, West Lafayette, IN 47907 feijuejuanling@gmail.com li2477@purdue.edu Chad Atalla University of California, San Diego 9500 Gilman Dr, La Jolla, CA 92093 Garrison Cottrell University of California, San Diego 9500 Gilman Dr, La Jolla, CA 92093 catalla@ucsd.edu gary@ucsd.edu Abstract alignment[39], and face detection[25]. However, humans not only read objective properties from a face, such as gender, expression, race, age and identity, but also form subjective impressions of the social aspects of a face[31, 32], such as facial attractiveness[28], friendliness, trustworthiness[29], sociability, dominance[18], and typicality. Despite the relative less attention received by the social perception of faces, social judgment is an important part of people’s daily interactions, and it has significant impact on social outcomes, ranging from electoral success to sentencing decisions[19, 35]. Whereas current computer vision techniques exceed human abilities at recognizing a face and identifying the objective properties of a face [27, 25], awareness of human subjective judgments is important for social robotics theory-of-mind inferences. Accurate predictions of social aspects of faces can help robots better understand how humans interact with and perceive each other, and can make a robot aware of inherent human biases, as these judgments rarely correspond to reality (except, perhaps, attractiveness) [32]. The human perceptual system can make complex inferences on faces, ranging from the objective evaluations regarding gender, ethnicity, expression, age, identity, etc. to subjective judgments on facial attractiveness, trustworthiness, sociability, friendliness, etc. Whereas the objective aspects have been extensively studied, less attention has been paid to modeling the subjective perception of faces. Here, we adapt 6 state-of-the-art neural networks pretrained on various image tasks (object classification, face identification, face localization) to predict human ratings on 40 social judgments of faces in the 10k US Adult Face Database. Supervised ridge regression on PCA of the conv5 2 layer in VGG-16 network gives best predictions on the average human ratings. Human group agreement was evaluated by repeatedly randomly splitting the raters into two halves for each face, and calculating the Pearson correlation between the two sets of averaged ratings. Due to this methodology, the models correlations with the average human ratings can exceed this score. We find that 1) model performance grows as the consensus on a face trait increases, and 2) model correlations are always higher than human correlations with each other. These results illustrate the learnability of the subjective perception of faces, especially when there is consensus, and the striking versatility and transferability of representations learned for object recognition. This work has strong applications to social robotics, allowing robots to infer human judgments of each other. In this paper, we teach a machine to infer social impressions, that match human judgments, from faces. We examine a list of 20 pairs of social features that are typically studied by social psychologists, and that are relevant to social interactions between people [33, 32, 19]. Examples are attractiveness [7, 28, 5, 10, 8], trustworthiness [6, 29], sociability, aggressiveness [18], friendliness, kindness, happiness, familiarity [20], and memorability [2, 11]). Although social perceptions of faces are subjective, there is often a consensus among human raters in how they perceive facial attractiveness, trustworthiness and dominance[6, 5]. This indicates that faces contain high-level visual cues for social interactions, and therefore it is possible to model this process with machine learning techniques. We take advantage of the state-of-the-art neural network models trained for ob- 1. Introduction Recent advances in deep convolutional networks have driven tremendous progress in a variety of challenging face processing tasks including face recognition[27], face 1 ject recognition and face recognition tasks and use their internal representations for social perception learning. In all 40 social dimensions, our model correlates with human averaged ratings better than the humans correlate with each other. The contributions of the paper are summarized below: • To the best of our knowledge, this work is the first attempt to systematically examine the consistency of human social perceptions of faces, to explore the landscape of social feature semantic space, and to predict human judgments of 40 social attributes of faces; • We adapted 6 state of the art neural network algorithms trained for various visual tasks to make social judgment predictions on faces and achieve high correlations with human ratings in all 40 dimensions; • We evaluate the tuning properties of nodes in the best network and visualize the patterns that maximally ignite the perceptions for each specific social dimensions to facilitate a better understanding of the neural networks’ behavior in face processing. The rest of the paper is organized as follows. In Section 2, we review related work on social perception modeling. Section 3 and 4 summarize the methodology and the experimental framework. The experimental results and visualizations are presented in Section 5 and 6. Section 7 concludes the paper. 2. Related work The focus of our paper is to infer as much social judgment information as possible from a face image and to predict the subjective impression of faces by learning from human group data. We review related work in terms of the visual features they use, the dataset they choose, the evaluation metric they adopt, and the social attributes they examine. Visual features Since the early 1990s, psychologists have identified that high level visual features, such as the averageness of a face[14, 21] and the symmetry of the face [23] can explain why certain faces look more attractive. Machine learning researchers have developed various computer vision features and models to predict social perceptions of faces, especially facial attractiveness. Yael et al.[5] used geometric ratios and distances between facial features based on facial landmarks to build an attractiveness predictor (0.65 correlation with human raters, face database size=184). End-to-end neural networks were applied to predict facial attractiveness in 2010[8] (correlation 0.458, face database size=2056, young female faces only). Amit Kagian and his colleagues have used a combination of landmark-derived features along with global features to obtain a high correlation with human group averages on facial attractiveness [10](0.82 Pearson correlation, face database size=91). Traditional computer vision features such as SIFT, HoG, Gabor filters have been blended to predict the relative ranking of facial attractiveness in [1](rank order correlation 0.63, face database size=200). Rothe et al. incorporate collaborative filtering techniques with visual features extracted from pretrained VGG networks[24] to achieve individual-level prediction of facial attractiveness[22](correlation 0.671 on female face queries, database size = 13,000). McCurrie et al. [17] build a model based on a pretrained VGG network to predict trustworthiness, dominance and IQ in faces (R2 values on trustworthiness, dominance and IQ are 0.5687, 0.4601, 0.3548 respectively, face database size=6000). Previous papers have achieved correlations with human performance between 0.458 to 0.82 in attractiveness predictions, depending on the dataset and method used. However, to date, there is no standard dataset that has been used to compare these approaches. Dataset Earlier studies employ datasets with relatively small numbers of faces (a few hundred) and most face datasets use young Caucasian faces only, as pointed out by [15]. In contrast, the MIT dataset[2] we use contains 2,222 high quality color images that vary in ethnicity, gender, age and expression, with ratings on 40 attributes. This dataset is smaller than two of the ones mentioned above. The first is collected from howhot.io, an online dating website[22] and contains 13,000 face images, but that work focused on personalized prediction of facial attractiveness, rather than average ratings. There are only binary choices (like or dislike) indicating implicit preference of facial attractiveness. The second one is collected from testmybrain.com, contains 6,000 grayscale face images [17], and includes just three social features: dominance, IQ and trustworthiness. Evaluation metric Social perceptions of faces are collected from human participants in various ways. The most common way is to ask for a discrete rating, say from 1-9 [2], or 1-7 [5] from a number of raters, and then use the group average as the score for a face in the specified feature dimension (e.g. attractiveness). The consistency of ratings between humans is checked by repeatedly randomly splitting human participants into two subgroups and then computing the correlation between the two groups’ mean ratings. To compare model predictions with human ratings, Pearson correlation[10, 3], Spearman rank correlation[2] and R-squared values[17] are used, depending on the nature of the data. Another method is to present a pair of faces or multiple faces and ask for a relative ranking in a particular dimension (e.g., attractiveness). Prediction accuracy is measured using Kendall’s Tau and the Gamma Test [1]. In Rothe et al. [22], a person indicates his/ her preference by choosing to like or dislike another user’s face photo. In this paper, since our goal is to predict a continuous score of human average ratings, and our raters do not all rate the same faces, we also use Pearson correlation with average human ratings on a per-face basis. Social attributes Although social perceptions are a subjective judgment, and may not reflect a person’s actual traits or mental states, humans tend to share consensus on their first impressions. Kiapour et al.[12] and Wang et al.[34] find that the social styles of people (bikers vs. hipsters, for example) can be identified and classified from image features. Dhar et al. (2011) show that the interestingness of an image can be quantified and predicted [4]. Bainbridge et al. (2013) prove that the memorability of a face image can be predicted and modified to make it more memorable [2]. Todorov et al. [30, 31, 32] used synthesized faces to study the perception of competence, dominance, extroversion, likeability, threat, trustworthiness and attractiveness in faces [29]. However their face photos lack realism compared to real-world photos and therefore cannot predict human’s social perceptions of real faces in a more natural environment. McCurrie et al. [17] have worked toward removing this limitation by using real human faces to make predictions of trustworthiness and dominance ratings. 12 From the literature, we can observe two trends: (i) Besides McCurrie et al. [17] and Todorov. et al[29], most machine learning work on social perception of faces focuses on attractiveness prediction, leaving the prediction of other social perceptions largely unstudied. We aim to bridge this gap in our paper. (ii) As summarized by Laurentini et al[15] usually small datasets are used, with few variations on expression, gender, ethnicity and age. The dataset we chose overcomes the above limitations and has comprehensive coverage of a list of 40 social feature ratings. The papers closest to ours are McCurrie et al. [17] and Todorov. et al[29]. Our paper differs from theirs in three major ways: (1) Todorov. et al’s work is on synthesized faces whereas ours is on realistic photos; (2) McCurrie et al. [17] predict three social features, dominance, trustworthiness, and IQ, whereas we look at 40 social features including trustworthiness, aggressiveness (a term close to their dominance), and intelligence (close to their IQ term), so our feature set can be considered to be a superset of theirs; and (3) we compared various feature extraction methods, including traditional geometric features and 6 neural networks pretrained for various tasks (face identification, face localization, object recognition). We also examine the effect of fine-tuning the network compared with directly applying ridge regression on extracted features from higher layer of the networks. 3. Method In this section we first describe the dataset used in our experiments. Next, we introduce our method for predict- ing social perceptions of faces. Finally, we explain how we visualize the features that contribute most to social trait predictions. 3.1. Dataset To predict how human evaluate social traits of a face at a glance, we use the dataset collected by Aude Oliva’s group [2]. The dataset consists of 2,222 images of faces sampled from the 10k US Adult Face Database and annotated for 20 pairs of social attributes. Each attribute is rated on a scale of 1-9 (1 means not at all, 9 means extremely) and each image is rated by 15 subjects. We take the average rating across all raters as a collective estimation of the social feature for every face. The 20 pairs of social traits are: (attractive, unattractive), (happy, unhappy), (friendly, unfriendly), (sociable, introverted), (kind, mean), (caring, cold), (calm, aggressive), (trustworthy, untrustworthy), (responsible, irresponsible), (confident, uncertain), (humble, egotistical),(emotionally stable, emotionally unstable), (normal, weird), (intelligent, unintelligent), (interesting, boring), (emotional, unemotional), (memorable, forgettable), (typical, atypical), (familiar, unfamiliar) and (common, uncommon). Clearly, some of these traits will be highly correlated, and are predictable from the others. We compute the Spearman’s rank correlation between every pair of social features and show their correlations in a heatmap (see the left figure in Figure 1). We put features together in the map based on similarity and positive/negativeness. From the figure, we can see that negative social features such as untrustworthy, aggressiveness, cold, introverted, irresponsible form a correlated block, while most positive features such as attractive, sociable, caring, friendly, happy, intelligent, interesting, confident are highly correlated with each other. Although we chose 20 pairs of opposite features, they are not completely complementary and redundant. Principal component analysis shows that it takes 24 principal components to cover 95% of the variance. 3.2. Regression Model for Social Attributes After we average human ratings on each face, each face receives a continuous score from 1 to 9 in all social dimensions. We model these social scores with a regression model. Our proposed algorithm is a ridge regression model on features extracted from deep convolutional neural networks (CNN). Since CNN features are usually highdimensional, we first perform Principal Component Analysis (PCA) on the extracted features of the training set to reduce the dimensionality. The PCA dimensionality is chosen by cross-validation on a validation set, separately for each trait. The PCA weights are saved and further used in fine-tuning our CNN-regression model. Figure 1: Correlation heatmaps among social features. Left: human. Right: network. 3.3. Feature Visualization 4.1. Baseline I: Human Correlations Attempting to understand what features are most helpful in social attribute prediction, we visualize the features extracted from the CNN. Two different methods were proposed in past feature visualization studies: dataset-centric methods [36, 38], and a network-centric method[36, 37]. Since these social attributes are all subjective perceptions rated by people, it is informative to examine to what extent people agree with each other upon those social judgments. We performed the following procedure 50 times for each attribute and then averaged the results: The dataset-centric method we employed is to display image patches from the training set that cause high activation for the feature units and use the deconvolution method to highlight the portions of the image that are responsible for firing the important feature neurons [36, 38]. 1. For each face, we randomly split the 15 raters evenly into two groups of 7 and 8. (Note: the raters for each face will, in general, be different sets). The network-centric approach is usually used in classification networks. This method produces an image that is based on adapting the input by maximizing the output category activation using gradient ascent, i.e., it is mainly a function of the network. [36, 37]. The key idea is to optimize the input image so that the target neuron can be highly activated. We apply this idea to the output (regression) neuron as well as the top nine neurons that influence that output individually. 4. Experimental framework In this section, we report our experimental framework using 6 CNN-based regression models with respect to two baselines, human correlation between groups of raters, and a baseline model using the geometric features. 2. We calculate the two group’s average ratings for each face, obtaining two vectors of length 2,222 (there are 2,222 faces in the dataset). 3. We calculate the correlation between the two vectors. The results are shown in the second column of Table 1. For every social attribute, the averaged correlation between human subgroups serves as an index of the rating consistency. 4.2. Baseline II: Regression on geometric features Past studies on facial attractiveness have found that attractiveness can be inferred from the geometric ratios and configurations of a face[5, 10]. We suggest that other social attributes can also be inferred from geometric features. We compute 29 geometric features based on definitions described in [16] and further extract a ”smoothness” feature and skin color features according to the procedure in [5, 10]. The ”smoothness” of a face was evaluated by applying a Canny edge detector to windows from the cheek/forehead area [5]. The more edges detected by edge detectors within the window, the less smooth the skin is. The regions we chose to compute smoothness and skin color are highlighted in the right subplot of Figure 2). The ”skin color” feature is extracted from the same window as ”smoothness”, converted from RGB to HSV. Regressing on these handcrafted features alone are not enough to capture the richness of geometric details about a face, we therefore use a computer vision library (dlib, C++) to automatically label 68 face landmarks (see Figure 2) for each face and compute distances and slopes between any two landmarks. Combining 29 handcrafted geometric features, smoothness, color and the distance-slope features, we obtain 4592 features in total. Since the features are highly correlated, we apply PCA to reduce dimensionality. Again, the PCA dimensionality is chosen by cross-validating on the hold out set separately for each facial attribute. Then a ridge regression model is applied to predict social attribute ratings of a face. The hyperparameter of ridge regression is selected by leave-one-out validation within the training set. Figure 2: 68 face landmarks labeled by dlib software automatically. The gray regions are the locations used for computing smoothness and skin color. 4.3. CNN-based Regression Model We initially compared six neural network architectures: (1) VGG16, (2) VGG-Face from the Oxford Visual Geometry Groups VGG networks[24], (3) AlexNet (the publicly available CaffeNet reference model) [13] (4) Inception from Google [26] (5) a shallow face identification Siamese neural network that we trained from scratch: Face-SNN and (6) a state of the art VGG-derived network trained for the face landmark localization task: Face-LandmarkNN. These comparisons were performed with the Caffe deep learning framework [9]. To find the best CNN to predict social attribute ratings among all six networks, we first find the best-performing feature layers of each network (with the ridge regression model), and then we compare the results among the networks to select the best network. For each layer of each network, before the ridge regression, we performed PCA and picked the PCA dimension that gave the best results on the validation set. 5. Results Surprisingly, we found that features from conv5 2 layer of VGG16 trained for object classification slightly outperformed the AlexNet and Inception networks, while the three networks trained solely on faces, VGG-Face, FaceLandmarkNN and Face-SNN did not achieve performance as competitive as the other three for most of the social attributes. The best performing VGG16 layer was conv5 2. We speculate that the reason for the relatively poorer performance of the face recognition networks is that they are optimized either to learn differences between faces which define identity or to learn the face landmark configurations, whereas for this task at hand, we are looking for commonalities behind certain social features which go beyond identity. The landmark network presumably should give results similar to the geometric features, but did not learn features corresponding to all of the features we used in that model. These speculations need to be checked, of course, for example by trying to predict all of our measured features, using the landmark network, but we did not do that here. We tried fine-tuning the model as follows. We used backpropagation to fine tune the weights into the conv5 2 layer, the weights to the PCA layer from conv5 2 (initialized by the PCA weights), and the weights from the PCA layer to the output regression unit. However, this fine-tuning did not improve performance, so the results reported in Table 1 are without fine-tuning. We evaluate the performance on 50 random train / validation / test splits of the data with a 64/16/20 percent split for training, cross-validation and testing, respectively. The prediction performance of our model is evaluated using Pearson’s correlation with the human ratings on the test set. For each social attribute, we report its human consistency as described in Section 4.1. Table 1 summarizes the prediction performance of our model for all the social attributes compared to Baseline I and II. The table is organized in a descending order of human agreement on the putative positive attribute of the paired attributes. The three attributes where there is greater agreement among humans for the negative component of the pair are bolded. Among all the social attributes, human subjects agree most with each other about ”happy” and disagree most about ”unfamiliar.” For both regression models (Baseline II and our model), model performance grows as the consensus on a social trait increases and human correlations with each other are consistently lower than the models’ correlations with the average human ratings. Normally, one might consider the human correlations to be an upper bound on performance, but here they are different kinds of correlations. Since the change in expression would produce a change in landmark locations, it is not surprising that landmark- based geometric features (Baseline II) achieve comparable or slightly higher correlation as our model for predicting those social attributes that are highly related to expressions (such as ”happy”, ”unhappy”, ”cold” and ”friendly” etc.). While for other social attributes, our model slightly outperforms landmark-based geometric features by about 0.04 correlation on average and significantly outperforms human correlation by about 0.12 correlation on average. This implies that CNN features encode much more information than just landmark-based features. It is essential to visualize those features and understand what features extracted from CNN make our model powerful enough to predict social attributes. To quantitatively compare the face social features perceived by humans and those predicted by our best performing model, we take the model predictions on all social features, and compute the Spearman correlation between every pair in the set (see the right figure in Figure 1). Not surprisingly, this has very similar patterns compared the heatmap generated from human ratings (see the right panel in Figure 1). Pearson Correlation between the upper triangle of the two similarity matrices (human and model prediction) is 0.9836. However, note that each predictor was trained independently. 6. Feature Visualization In this section, we visualize the features that are of importance to social perceptions. We choose facial attractiveness as an example. The same method can be applied to the other social features. We employ the two methods described in section 3.3 to visualize features learned by our model. 6.1. Data-centric Visualization To identify visual features that ignite attractiveness perception, we find the top 9 units of highest influence on attractiveness at conv5 2 as follows. First, we compute a product of three terms: (1) A unit’s activation from conv5 2, (2) that unit’s weight to the following fc pca layer, (3) the fc pca unit’s weight to the output unit. We then sort all conv5 2 units’ average products of the three terms and identify the top 9 neurons as the ones that contribute most to the output neuron for the corresponding social feature. Then we employ the method described in [36, 38] to find top-9 input images that cause high activations in each of the top-9 conv5 2 neurons. Also we further produce the deconvolutional images by projecting each activation separately down to pixel space. Figure 3 captures the features that are important to predict the attractiveness of a face. The feature importance descends from left to right and top to bottom. The important features identified by our model are related to eyes, hair Social Attributes happy unhappy friendly unfriendly sociable introverted attractive unattractive kind mean caring cold trustworthy untrustworthy responsible irresponsible confident uncertain humble egotistic emotionally stable emotionally unstable normal weird intelligent unintelligent interesting boring calm aggressive emotional unemotional memorable forgettable typical atypical common uncommon familiar unfamiliar Baseline I 0.84 0.75 0.78 0.72 0.74 0.50 0.72 0.62 0.72 0.69 0.72 0.71 0.62 0.60 0.58 0.55 0.55 0.45 0.55 0.52 0.53 0.50 0.49 0.52 0.49 0.43 0.42 0.39 0.41 0.65 0.33 0.56 0.30 0.27 0.28 0.24 0.25 0.27 0.24 0.18 Baseline II 0.86 0.81 0.83 0.80 0.78 0.64 0.66 0.62 0.79 0.75 0.78 0.81 0.72 0.69 0.65 0.64 0.55 0.62 0.64 0.62 0.64 0.62 0.58 0.50 0.53 0.53 0.64 0.54 0.47 0.72 0.60 0.76 0.38 0.40 0.41 0.40 0.37 0.38 0.42 0.40 Our Model 0.84 0.80 0.82 0.79 0.78 0.65 0.75 0.70 0.79 0.73 0.79 0.79 0.73 0.70 0.70 0.67 0.61 0.63 0.63 0.62 0.67 0.64 0.61 0.56 0.62 0.58 0.67 0.60 0.50 0.72 0.60 0.75 0.48 0.48 0.43 0.43 0.40 0.40 0.44 0.44 Table 1: Prediction performance of all the social attributes. The reported performance is averaged on 50 random train/validation/test splits of the data. with bangs, high nose-bridge, high cheeks, dark eyebrows, strong commanding jawline, chin and red lips. Note that among the 9 cropped input image patches, not all the faces are perceived as attractive or rated as attractive. An attractive face needs to activate more than one feature in order to be considered attractive. This observation agrees with our intuition that attractiveness is a kind of holistic judgment, requiring a combination of multiple features. It also seems to be the case that several of the features include relationships between the parts. For example, while the first feature in the upper left of the figure emphasizes the eye, it also includes the nose. This is also true of the upper right feature. Smiling is also important in order to be perceived attractive, as emphasized by the feature in the lower left of the figure. 6.2. Network-centric Visualization In section 6.1, we have identified the top-9 units and their feature maps from the con5 2 layer that maximally activates the attractiveness neuron. Here, we use the gradient-ascent method to optimize the input image that would highly activate a specific neuron of the network. This method is also performed on the pretrained-VGG16 regression model, which is trained to predict attractiveness. Figure 4a shows the optimized image corresponding to the output neuron from a random input image. Optimizing the input image for the output neuron of a regression model does not result in a particularly interpretable figure, although it does appear to emphasize the eyes. Our second approach is to optimize the input image with respect to the top-9 contributing neurons from conv5 2 layer that have been identified in section 6.1. Figure 4b presents 9 optimized images with respect to the corresponding top-9 feature maps of the top-9 neurons from conv5 2 layer. Since we use a pretrained-VGG16 network for visualization, it is not surprising that the corresponding top-9 feature maps at conv5 2 layer are not particularly encoding facial patterns. We also present the optimized image initialize with a face image, along with the original face image for comparison in Figure 5. The optimized image tends to highlight the eyes, nose, cheeks and the contour of the face, which is consistent with the features identified by data-centric method. 7. Conclusion We have shown that a deep network can be used to predict human social judgments with high correlation with the average human ratings. As far as we know, this is the widest exploration of social judgment predictions, showing human-like perceptions on 40 social dimensions. Unsurprisingly, given previous work recognizing facial expressions, where happiness is the easiest to recognize, our highest correlation is on the happy feature. However, previous work in this area tended to classify a face as happy or not, rather than the degree of rated happiness. We find that for attributes that correspond to elements of the face that require muscle movement, or a lack of it (such as happy, unhappy, cold, aggressive, unemotional) a simple regression model based on the placement of facial landmarks works well. For ones that don’t appear to suggest emotions, such as friendly, note that friendly and happy are highly correlated (see Figure 1, and the red block indexed by happy and friendly). Similarly, aggressive and mean are highly correlated, which presumably requires not smiling. Perhaps of more significance are the correlations with judgments of traits, such as trustworthiness, responsibleness, confidence, and intelligence, which would correspond to more static features of the face. In this area, the deep network, which responds to facial textures as well as shape, has superior performance. While these judgments do not correspond to a notion of ”ground truth,” they are things for which humans have a fair amount of agreement, suggesting that there is a signal to be recognized. Of further note is that we have shown, yet again, that a machine can recognize attractiveness, presumably without any hormonal influences. For this dataset, our deep network correlates with human ratings at 0.75. This provides a benchmark for this dataset. Finally, it is of note that we can see that some of the traits considered to be ”opposite” in this list are not simply the reverse of one another. For example, there is a large difference in human agreements on ”sociable” (0.74) versus ”introverted” (0.50), suggesting they are not opposites. These results are significant for the field of social robotics. While a robot should not judge a human based completely on their appearance, it can be useful knowledge that humans might judge a person to be trustworthy, while the robot can be more objective. Similarly, a robot need not treat an attractive and unattractive person differently, but this knowledge could affect how the robot interacts with the unattractive person, knowing in advance that this person may have had many negative experiences interacting with people. In this paper, we train each social feature separately, due to their varied consistency and reliability. In the future, it is worth trying to train one single convnet to learn multiple tasks simultaneously and evaluate whether shared representation may further improve the model performance. In summary, we have provided the first machine learning system to learn subjective human judgments of a wide spectrum of traits. We found that the more humans agree on such subjective judgments, the more the system could pick up on the features driving those judgments. It will be of interest to investigate further what those features are, beyond the attractiveness features we displayed here. One step further from predicting the value in a certain social feature is to move faces on the social manifold and to increase a face’s elicited social perceptions in positive ways (e.g. to make a face look more sociable/ trustworthy/ attractive). Although the images generated by our current visualization method are still far away from being photorealistic, it may be a fruitful area in the future to develop generative models that can achieve this goal. Figure 3: Visualization of features in the pretrained-VGG16 regression network. For conv5 2 layer, we show the top 9 activations of the top 9 neurons that maximally activate the attractiveness neuron across the training data, projected down to pixel space using the deconvolutional network approach [38] and their corresponding cropped image patches. Best viewed in electronic form, and zoomed in. References [1] H. Altwaijry and S. Belongie. Relative ranking of facial attractiveness. In Applications of Computer Vision (WACV), 2013 IEEE Workshop on, pages 117–124. IEEE, 2013. 2 [2] W. A. Bainbridge, P. Isola, and A. Oliva. The intrinsic memorability of face photographs. Journal of Experimental Psychology: General, 142(4):1323, 2013. 1, 2, 3 [3] A. Dantcheva and J. Dugelay. Female facial aesthetics based on soft biometrics and photo-quality. In Proceedings of ICME, 2011. 2 [4] S. Dhar, V. Ordonez, and T. L. Berg. High level describable attributes for predicting aesthetics and interestingness. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, pages 1657–1664. IEEE, 2011. 3 [5] Y. Eisenthal, G. Dror, and E. Ruppin. Facial attractiveness: Beauty and the machine. Neural Computation, 18(1):119– 142, 2006. 1, 2, 4, 5 [6] V. Falvello, M. Vinson, C. Ferrari, and A. Todorov. The robustness of learning about the trustworthiness of other people. Social Cognition, 33(5):368, 2015. 1 [7] K. Grammer and R. Thornhill. Human (homo sapiens) facial attractiveness and sexual selection: the role of symmetry and averageness. Journal of comparative psychology, 108(3):233, 1994. 1 [8] D. Gray, K. Yu, W. Xu, and Y. Gong. Predicting facial beauty without landmarks. In Computer Vision–ECCV 2010, pages 434–447. Springer, 2010. 1, 2 [9] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional architecture for fast feature embedding. arXiv preprint arXiv:1408.5093, 2014. 5 [10] A. Kagian, G. Dror, T. Leyvand, I. Meilijson, D. Cohen-Or, and E. Ruppin. A machine learning predictor of facial attractiveness revealing human-like psychophysical biases. Vision research, 48(2):235–243, 2008. 1, 2, 4 [11] A. Khosla, W. A. Bainbridge, A. Torralba, and A. Oliva. Modifying the memorability of face photographs. In Computer Vision (ICCV), 2013 IEEE International Conference on, pages 3200–3207. IEEE, 2013. 1 [12] M. H. Kiapour, K. Yamaguchi, A. C. Berg, and T. L. Berg. Hipster wars: Discovering elements of fashion styles. In European conference on computer vision, pages 472–488. Springer, 2014. 3 [13] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pages 1097–1105, 2012. 5 [14] J. H. Langlois and L. A. Roggman. Attractive faces are only average. Psychological science, 1(2):115–121, 1990. 2 [15] A. Laurentini and A. Bottino. Computer analysis of face beauty: A survey. Computer Vision and Image Understanding, 125:184–199, 2014. 2, 3 [16] D. S. Ma, J. Correll, and B. Wittenbrink. The chicago face database: A free stimulus set of faces and norming data. Behavior research methods, 47(4):1122–1135, 2015. 4 (a) Original input image (b) Optimized image Figure 5: Visualization of the optimized image with a input face image: Figure 5a is the original face image before optimization. Figure 5b is produced by performing optimization with respect to the output unit. (a) Optimized input image with respect to the output unit [20] [21] [22] [23] [24] [25] (b) Optimized input images with respect to top-9 neurons from conv5 2 layer [26] Figure 4: Visualization of features using network-centric method. To produce Fig 4a, we use gradient ascent to optimize the output neuron. Fig 4b shows 9 optimized images for the feature maps corresponding to top-9 contributor neurons from conv5 2 layer. [27] [28] [17] M. McCurrie, F. Beletti, L. Parzianello, A. Westendorp, S. Anthony, and W. Scheirer. Predicting first impressions with deep learning. arXiv preprint arXiv:1610.08119, 2016. 2, 3 [18] A. Mignault and A. Chaudhuri. The many faces of a neutral face: Head tilt and perception of dominance and emotion. Journal of nonverbal behavior, 27(2):111–132, 2003. 1 [19] N. N. Oosterhof and A. Todorov. The functional basis of [29] [30] face evaluation. Proceedings of the National Academy of Sciences, 105(32):11087–11092, 2008. 1 M. Peskin and F. N. Newell. Familiarity breeds attraction: Effects of exposure on the attractiveness of typical and distinctive faces. PERCEPTION-LONDON-, 33(2):147–158, 2004. 1 G. Rhodes, L. Jeffery, T. L. Watson, C. W. Clifford, and K. Nakayama. Fitting the mind to the world face adaptation and attractiveness aftereffects. Psychological science, 14(6):558–566, 2003. 2 R. Rothe, R. Timofte, and L. Van Gool. Some like it hotvisual guidance for preference prediction. arXiv preprint arXiv:1510.07867, 2015. 2 J. E. Scheib, S. W. Gangestad, and R. Thornhill. Facial attractiveness, symmetry and cues of good genes. Proceedings of the Royal Society of London. Series B: Biological Sciences, 266(1431):1913–1917, 1999. 2 K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2, 5 R. Stewart, M. Andriluka, and A. Y. Ng. End-to-end people detection in crowded scenes. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 1 C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1–9, 2015. 5 Y. Taigman, M. Yang, M. Ranzato, and L. Wolf. Deepface: Closing the gap to human-level performance in face verification. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1701–1708, 2014. 1 R. Thornhill and S. W. Gangestad. Facial attractiveness. Trends in cognitive sciences, 3(12):452–460, 1999. 1 A. Todorov, S. G. Baron, and N. N. Oosterhof. Evaluating face trustworthiness: a model based approach. Social cognitive and affective neuroscience, 3(2):119–127, 2008. 1, 3 A. Todorov, R. Dotsch, J. M. Porter, N. N. Oosterhof, and V. B. Falvello. Validation of data-driven computational models of social perception of faces. Emotion, 13(4):724, 2013. 3 [31] A. Todorov, P. Mende-Siedlecki, and R. Dotsch. Social judgments from faces. Current opinion in neurobiology, 23(3):373–380, 2013. 1, 3 [32] A. Todorov, C. Y. Olivola, R. Dotsch, and P. MendeSiedlecki. Social attributions from faces: Determinants, consequences, accuracy, and functional significance. Psychology, 66(1):519, 2015. 1, 3 [33] A. Todorov, C. P. Said, A. D. Engell, and N. N. Oosterhof. Understanding evaluation of faces on social dimensions. Trends in cognitive sciences, 12(12):455–460, 2008. 1 [34] Y. Wang and G. W. Cottrell. Bikers are like tobacco shops, formal dressers are like suits: Recognizing urban tribes with caffe. In 2015 IEEE Winter Conference on Applications of Computer Vision, pages 876–883. IEEE, 2015. 3 [35] J. Willis and A. Todorov. First impressions making up your mind after a 100-ms exposure to a face. Psychological science, 17(7):592–598, 2006. 1 [36] J. Yosinski, J. Clune, A. Nguyen, T. Fuchs, and H. Lipson. Understanding neural networks through deep visualization. arXiv preprint arXiv:1506.06579, 2015. 4, 6 [37] S. Yu and K. Zipser. A deep neural net trained for person categorization develops both detailed local features and broad contextual specificities. Journal of Vision, 16(12):411–411, 2016. 4 [38] M. D. Zeiler and R. Fergus. Visualizing and understanding convolutional networks. In European Conference on Computer Vision, pages 818–833. Springer, 2014. 4, 6, 8 [39] X. Zhu, Z. Lei, X. Liu, H. Shi, and S. Z. Li. Face alignment across large poses: A 3d solution. arXiv preprint arXiv:1511.07212, 2015. 1 

2cs.AI

arXiv:1607.00743v1 [math.ST] 4 Jul 2016 A Residual Bootstrap for High-Dimensional Regression with Near Low-Rank Designs Miles E. Lopes Department of Statistics University of California, Berkeley Berkeley, CA 94720 mlopes@stat.berkeley.edu Abstract We study the residual bootstrap (RB) method in the context of high-dimensional linear regression. Specifically, we analyze the distributional approximation of linear contrasts c⊤ (βbρ − β), where βbρ is a ridge-regression estimator. When regression coefficients are estimated via least squares, classical results show that RB consistently approximates the laws of contrasts, provided that p ≪ n, where the design matrix is of size n × p. Up to now, relatively little work has considered how additional structure in the linear model may extend the validity of RB to the setting where p/n ≍ 1. In this setting, we propose a version of RB that resamples residuals obtained from ridge regression. Our main structural assumption on the design matrix is that it is nearly low rank — in the sense that its singular values decay according to a power-law profile. Under a few extra technical assumptions, we derive a simple criterion for ensuring that RB consistently approximates the law of a given contrast. We then specialize this result to study confidence intervals for mean response values Xi⊤ β, where Xi⊤ is the ith row of the design. More precisely, we show that conditionally on a Gaussian design with near low-rank structure, RB simultaneously approximates all of the laws Xi⊤ (βbρ − β), i = 1, . . . , n. This result is also notable as it imposes no sparsity assumptions on β. Furthermore, since our consistency results are formulated in terms of the Mallows (Kantorovich) metric, the existence of a limiting distribution is not required. 1 Introduction Until recently, much of the emphasis in the theory of high-dimensional statistics has been on “first order” problems, such as estimation and prediction. As the understanding of these problems has become more complete, attention has begun to shift increasingly towards “second order” problems, dealing with hypothesis tests, confidence intervals, and uncertainty quantification [1, 2, 3, 4, 5, 6]. In this direction, much less is understood about the effects of structure, regularization, and dimensionality — leaving many questions open. One collection of such questions that has attracted growing interest deals with the operating characteristics of the bootstrap in high dimensions [7, 8, 9]. Due to the fact that bootstrap is among the most widely used tools for approximating the sampling distributions of test statistics and estimators, there is much practical value in understanding what factors allow for the bootstrap to succeed in the high-dimensional regime. The regression model and linear contrasts. In this paper, we focus our attention on highdimensional linear regression, and our aim is to know when the residual bootstrap (RB) method consistently approximates the laws of linear contrasts. (A review of RB is given in Section 2.) 1 To specify the model, suppose that we observe a response vector Y ∈ Rn , generated according to Y = Xβ + ε, n×p (1) p where X ∈ R is the observed design matrix, β ∈ R is an unknown vector of coefficients, and the error variables ε = (ε1 , . . . , εn ) are drawn i.i.d. from an unknown distribution F0 , with mean 0 and unknown variance σ 2 < ∞. As is conventional in high-dimensional statistics, we assume the model (1) is embedded in a sequence of models indexed by n. Hence, we allow X, β, and p to vary implicitly with n. We will leave p/n unconstrained until Section 3.3, where we will assume p/n ≍ 1 in Theorem 3, and then in Section 3.4, we will assume further that p/n is bounded strictly between 0 and 1. The distribution F0 is fixed with respect to n, and none of our results require F0 to have more than four moments. Although we are primarily interested in cases where the design matrix X is deterministic, we will also study the performance of the bootstrap conditionally on a Gaussian design. For this reason, we will use the symbol E[. . . |X] even when the design is non-random so that confusion does not arise in relating different sections of the paper. Likewise, the symbol E[. . . ] refers to unconditional expectation over all sources of randomness. Whenever the design is random, we will assume X ⊥ ⊥ ε, denoting the distribution of X by PX , and the distribution of ε by Pε . Within the context of the regression, we will be focused on linear contrasts c⊤ (βb − β), where c ∈ Rp is a fixed vector and βb ∈ Rp is an estimate of β. The importance of contrasts arises from the fact that they unify many questions about a linear model. For instance, testing the significance of the ith coefficient βi may be addressed by choosing c to be the standard basis vector c⊤ = e⊤ i . Another important problem is quantifying the uncertainty of point predictions, which may be addressed by choosing c⊤ = Xi⊤ , i.e. the ith row of the design matrix. In this case, an approximation to the law of the contrast leads to a confidence interval for the mean response value E[Yi ] = Xi⊤ β. Further applications of contrasts occur in the broad topic of ANOVA [10]. Intuition for structure and regularization in RB. The following two paragraphs explain the core conceptual aspects of the paper. To understand the role of regularization in applying RB to highdimensional regression, it is helpful to think of RB in terms of two ideas. First, if βbLS denotes the ordinary least squares estimator, then it is a simple but important fact that contrasts can be written as c⊤ (βbLS − β) = a⊤ ε where a⊤:= c⊤ (X ⊤ X)−1 X ⊤ . Hence, if it were possible to sample directly from F0 , then the law of any such contrast could be easily determined. Since F0 is unknown, the second key idea is to use the residuals of some estimator βb as a proxy for samples from F0 . When p ≪ n, the least-squares residuals are a good proxy [11, 12]. However, it is well-known that leastsquares tends to overfit when p/n ≍ 1. When βbLS fits “too well”, this means that its residuals are b “too small”, and hence they give a poor proxy for F0 . Therefore, by using a regularized estimator β, b overfitting can be avoided, and the residuals of β may offer a better way of obtaining “approximate samples” from F0 . The form of regularized regression we will focus on is ridge regression: βbρ := (X ⊤ X + ρIp×p )−1 X ⊤ Y, (2) where ρ > 0 is a user-specificed regularization parameter. As will be seen in Sections 3.2 and 3.3, the residuals obtained from ridge regression lead to a particularly good approximation of F0 when the design matrix X is nearly low-rank, in the sense that most of its singular values are close to 0. In essence, this condition is a form of sparsity, since it implies that the rows of X nearly lie in a low-dimensional subspace of Rp . However, this type of structural condition has a significant advantage over the the more well-studied assumption that β is sparse. Namely, the assumption that X is nearly low-rank can be inspected directly in practice — whereas sparsity in β is typically unverifiable. In fact, our results will impose no conditions on β, other than that kβk2 remains bounded as (n, p) → ∞. Finally, it is worth noting that the occurrence of near low-rank design matrices is actually very common in applications, and is often referred to as collinearity [13, ch. 17]. Contributions and outline. The primary contribution of this paper is a complement to the work of Bickel and Freedman [12] (hereafter B&F 1983) — who showed that in general, the RB method fails 2 to approximate the laws of least-squares contrasts c⊤ (βbLS − β) when p/n ≍ 1. Instead, we develop an alternative set of results, proving that even when p/n ≍ 1, RB can successfully approximate the laws of “ridged contrasts” c⊤ (βbρ − β) for many choices of c ∈ Rp , provided that the design matrix X is nearly low rank. A particularly interesting consequence of our work is that RB successfully approximates the law c⊤ (βbρ − β) for a certain choice of c that was shown in B&F 1983 to “break” RB when applied to least-squares. Specifically, such a c can be chosen as one of the rows of X with a high leverage score (see Section 4). This example corresponds to the practical problem of setting confidence intervals for mean response values E[Yi ] = Xi⊤ β. (See [12, p. 41], as well as Lemma 2 and Theorem 4 in Section 3.4). Lastly, from a technical point of view, a third notable aspect of our results is that they are formulated in terms of the Mallows-ℓ2 metric, which frees us from having to impose conditions that force a limiting distribution to exist. Apart from B&F 1983, the most closely related works we are aware of are the recent papers [7] and [8], which also consider RB in the high-dimensional setting. However, these works focus on role of sparsity in β and do not make use of low-rank structure in the design, whereas our work deals only with structure in the design and imposes no sparsity assumptions on β. The remainder of the paper is organized as follows. In Section 2, we formulate the problem of approximating the laws of contrasts, and describe our proposed methodology for RB based on ridge regression. Then, in Section 3 we state several results that lay the groundwork for Theorem 4, which shows that that RB can successfully approximate all of the laws L(Xi⊤ (βbρ − β)|X), i = 1, . . . , n, conditionally on a Gaussian design. Due to space constraints, all proofs are deferred to material that will appear in a separate work. Notation and conventions. If U and V are random variables, then L(U |V ) denotes the law of U , conditionally on V . If an and bn are two sequences of real numbers, then the notation an . bn means that there is an absolute constant κ0 > 0 and an integer n0 ≥ 1 such that an ≤ κ0 bn for all n ≥ n0 . The notation an ≍ bn means that an . bn and bn . an . For a square matrix A ∈ Rk×k whose eigenvalues are real, we denote them by λmin (A) = λk (A) ≤ · · · ≤ λ1 (A) = λmax (A). 2 Problem setup and methodology Problem setup. For any c ∈ Rp , it is clear that conditionally on X, the law of c⊤ (βbρ − β) is completely determined by F0 , and hence it makes sense to use the notation  
Ψρ (F0 ; c) := L c⊤ (βbρ − β) (3) X . The problem we aim to solve is to approximate the distribution Ψρ (F0 ; c) for suitable choices of c. Review of the residual bootstrap (RB) procedure. We briefly explain the steps involved in the residual bootstrap procedure, applied to the ridge estimator βbρ of β. To proceed somewhat indirectly, consider the following “bias-variance” decomposition of Ψρ (F0 ; c), conditionally on X,


Ψρ (F0 ; c) = L c⊤ βbρ − E[βbρ |X]  (4) X + c⊤ E[βbρ |X] − β . | {z } {z } | =: Φρ (F0 ;c) =: bias(Φρ (F0 ;c)) Note that the distribution Φ(F0 ; c) has mean zero, and so that the second term on the right side is the bias of Φρ (F0 ; c) as an estimator of Ψρ (F0 ; c). Furthermore, the distribution Φρ (F0 ; c) may be viewed as the “variance component” of Ψρ (F0 ; c). We will be interested in situations where the regularization parameter ρ may be chosen small enough so that the bias component is small. In this case, one has Ψρ (F0 ; c) ≈ Φρ (F0 ; c), and then it is enough to find an approximation to the law Φρ (F0 ; c), which is unknown. To this end, a simple manipulation of c⊤ (βbρ − E[βbρ ]) leads to   Φρ (F0 ; c) = L(c⊤ (X ⊤ X + ρIp×p )−1 X ⊤ ε (5) X). Now, to approximate Φρ (F0 ; c), let Fb be any centered estimate of F0 . (Typically, Fb is obtained by using the centered residuals of some estimator of β, but this is not necessary in general.) Also, let ε∗ = (ε∗1 , . . . , ε∗n ) ∈ Rn be an i.i.d. sample from Fb . Then, replacing ε with ε∗ in line (5) yields   Φρ (Fb ; c) = L(c⊤ (X ⊤ X + ρIp×p )−1 X ⊤ ε∗  (6) X). 3 At this point, we define the (random) measure Φρ (Fb ; c) to be the RB approximation to Φρ (F0 ; c). Hence, it is clear that the RB approximation is simply a “plug-in rule”. A two-stage approach. An important feature of the procedure just described is that we are free to use any centered estimator Fb of F0 . This fact offers substantial flexibility in approximating Ψρ (F0 ; c). One way of exploiting this flexibility is to consider a two-stage approach, where a “pilot” ridge estimator βb̺ is used to first compute residuals whose centered empirical distribution function is Fb̺ , say. Then, in the second stage, the distribution Fb̺ is used to approximate Φρ (F0 ; c) via the relation (6). To be more detailed, if (b e1 (̺), . . . , ebn (̺)) = eb(̺) := Y − X βb̺ are the residuals of βb̺ , b then we define F̺ to be the distribution that places mass 1/n at each of the values ebi (̺) − ē(̺) with P ē(̺) := n1 ni=1 ebi (̺). Here, it is important to note that the value ̺ is chosen to optimize Fb̺ as an approximation to F0 . By contrast, the choice of ρ depends on the relative importance of width and coverage probability for confidence intervals based on Φρ (Fb̺ ; c). Theorems 1, 3, and 4 will offer some guidance in selecting ̺ and ρ. Resampling algorithm. To summarize the discussion above, if B is user-specified number of bootstrap replicates, our proposed method for approximating Ψρ (F0 ; c) is given below. 1. Select ρ and ̺, and compute the residuals eb(̺) = Y − X βb̺ . 2. Compute the centered distribution function Fb̺ , putting mass 1/n at each ebi (̺) − ē(̺). 3. For j = 1, . . . , B: • Draw a vector ε∗ ∈ Rn of n i.i.d. samples from Fb̺ . • Compute zj := c⊤ (X ⊤ X + ρIp×p )−1 X ⊤ ε∗ . 4. Return the empirical distribution of z1 , . . . , zB . Clearly, as B → ∞, the empirical distribution of z1 , . . . , zB converges weakly to Φρ (Fb̺ ; c), with probability 1. As is conventional, our theoretical analysis in the next section will ignore Monte Carlo issues, and address only the performance of Φρ (Fb̺ ; c) as an approximation to Ψρ (F0 ; c). 3 Main results The following metric will be central to our theoretical results, and has been a standard tool in the analysis of the bootstrap, beginning with the work of Bickel and Freedman [14]. The Mallows (Kantorovich) metric. For two random vectors U and V in a Euclidean space, the Mallows-ℓ2 metric is defined by n h i o (7) d22 (L(U ), L(V )) := inf E kU − V k22 : (U, V ) ∼ π π∈Π where the infimum is over the class Π of joint distributions π whose marginals are L(U ) and L(V ). It is worth noting that convergence in d2 is strictly stronger than weak convergence, since it also requires convergence of second moments. Additional details may be found in the paper [14]. 3.1 A bias-variance decomposition for bootstrap approximation To give some notation for analyzing the bias-variance decomposition of Ψρ (F0 ; c) in line (4), we define the following quantities based upon the ridge estimator βbρ . Namely, the variance is vρ = vρ (X; c) := var(Ψρ (F0 ; c)|X) = σ 2 kc⊤ (X ⊤ X + ρIp×p )−1 X ⊤ k22 . To express the bias of Φρ (F0 ; c), we define the vector δ(X) ∈ Rp according to   δ(X) := β − E[βbρ ] = Ip×p − (X ⊤ X + ρIp×p )−1 X ⊤ X β, 4 (8) and then put b2ρ = b2ρ (X; c) := bias2 (Φρ (F0 ; c)) = (c⊤ δ(X))2 . (9) We will sometimes omit the arguments of vρ and b2ρ to lighten notation. Note that vρ (X; c) does not depend on β, and b2ρ (X; c) only depends on β through δ(X). The following result gives a regularized and high-dimensional extension of some lemmas in Freedman’s early work [11] on RB for least squares. The result does not require any structural conditions on the design matrix, or on the true parameter β. Theorem 1 (consistency criterion). Suppose X ∈ Rn×p is fixed. Let Fb be any estimator of F0 , and let c ∈ Rp be any vector such that vρ = vρ (X; c) 6= 0. Then with Pε -probability 1, the following inequality holds for every n ≥ 1, and every ρ > 0,   b2 (10) d22 √1vρ Ψρ (F0 ; c), √1vρ Φρ (Fb ; c) ≤ σ12 d22 (F0 , Fb ) + vρρ . √ Remarks. Observe that the normalization 1/ vρ ensures that the bound is non-trivial, since the √ distribution Ψρ (F0 ; c)/ vρ has variance equal to 1 for all n (and hence does not become degenerate for large n). To consider the choice of ρ, it is simple to verify that the ratio b2ρ /vρ decreases monotonically as ρ decreases. Note also that as ρ becomes small, the variance vρ becomes large, and likewise, confidence intervals based on Φρ (Fb; c) become wider. In other words, there is a trade-off between the width of the confidence interval and the size of the bound (10). Sufficient conditions for consistency of RB. An important practical aspect of Theorem 1 is that for any given contrast c, the variance vρ (X; c) can be easily estimated, since it only requires an estimate of σ 2 , which can be obtained from Fb. Consequently, whenever theoretical bounds on d22 (F0 , Fb) and b2ρ (X; c) are available, the right side of line (10) can be controlled. In this way, Theorem 1 offers a simple route for guaranteeing that RB is consistent. In Sections 3.2 and 3.3 to follow, we derive a bound on E[d22 (F0 , Fb)|X] in the case where Fb is chosen to be Fb̺ . Later on in Section 3.4, we study RB consistency in the context of prediction with a Gaussian design, and there we derive high probability bounds on both vρ (X; c) and b2ρ (X; c) where c is a particular row of X. 3.2 A link between bootstrap consistency and MSPE If βb is an estimator of β, its mean-squared prediction error (MSPE), conditionally on X, is defined as     mspe(βb |X) := n1 E kX(βb − β)k22  (11) X . The previous subsection showed that in-law approximation of contrasts is closely tied to the approximation of F0 . We now take a second step of showing that if the centered residuals of an estimator βb are used to approximate F0 , then the quality of this approximation can be bounded naturally in terms of mspe(βb |X). This result applies to any estimator βb computed from the observations (1). Theorem 2. Suppose X ∈ Rn×p is fixed. Let βb be any estimator of β, and let Fb be the empirical b Also, let Fn denote the empirical distribution of n i.i.d. distribution of the centered residuals of β. samples from F0 . Then for every n ≥ 1,    2   E d22 (Fb, F0 )X ≤ 2 mspe(βb |X) + 2 E[d22 (Fn , F0 )] + 2σn . (12) Remarks. As we will see in the next section, the MSPE of ridge regression can be bounded in a sharp way when the design matrix is approximately low rank, and there we will analyze mspe(βb̺ |X) for the pilot estimator. Consequently, when near low-rank structure is available, the only remaining issue in controlling the right side of line (12) is to bound the quantity E[d22 (Fn , F0 )|X]. The very recent work of Bobkov and Ledoux [15] provides an in-depth study of this question, and they derive a variety bounds under different tail conditions on F0 . We summarize one of their results below. Lemma 1 (Bobkov and Ledoux, 2014). If F0 has a finite fourth moment, then E[d22 (Fn , F0 )] . log(n)n−1/2 . 5 (13) Remarks. The fact that the squared distance is bounded at the rate of log(n)n−1/2 is an indication that d2 is a rather strong metric on distributions. For a detailed discussion of this result, see Corollaries 7.17 and 7.18 in the paper [15]. Although it is possible to obtain faster rates when more b stringent tail conditions are placed on F0 , we will only need a fourth moment, since the mspe(β|X) term in Theorem 2 will often have a slower rate than log(n)n−1/2 , as discussed in the next section. 3.3 Consistency of ridge regression in MSPE for near low rank designs In this subsection, we show that when the tuning parameter ̺ is set at a suitable rate, the pilot ridge estimator βb̺ is consistent in MSPE when the design matrix is near low-rank — even when p/n is large, and without any sparsity constraints on β. We now state some assumptions. A1. There is a number ν > 0, and absolute constants κ1 , κ2 > 0, such that b ≤ κ2 i−ν κ1 i−ν ≤ λi (Σ) for all i = 1, . . . , n ∧ p. A2. There are absolute constants θ, γ > 0, such that for every n ≥ 1, ̺ n = n−θ and ρ n = n−γ . A3. The vector β ∈ Rp satisfies kβk2 . 1. Due to Theorem 2, the following bound shows that the residuals of βb̺ may be used to extract a consistent approximation to F0 . Two other notable features of the bound are that it is non-asymptotic and dimension-free. Theorem 3. Suppose that X ∈ Rn×p is fixed and that Assumptions 1–3 hold, with p/n ≍ 1. Assume 1 ν 1 further that θ is chosen as θ = 2ν 3 when ν ∈ (0, 2 ), and θ = ν+1 when ν > 2 . Then, ( 2ν if ν ∈ (0, 21 ), n− 3 ν (14) mspe(βb̺ |X) . − ν+1 n if ν > 12 . Also, both bounds in (14) are tight in the sense that β can be chosen so that βb̺ attains either rate. b are observable, they may be used to estimate ν and guide Remarks. Since the eigenvalues λi (Σ) the selection of ̺/n = n−θ . However, from a practical point of view, we found it easier to select ̺ via cross-validation in numerical experiments, rather than via an estimate of ν. A link with Pinsker’s Theorem. In the particular case when F0 is a centered Gaussian distribution, the “prediction problem” of estimating Xβ is very similar to estimating the mean parameters of a Gaussian sequence model, with error measured in the ℓ2 norm. In the alternative sequence-model format, the decay condition on the eigenvalues of n1 X ⊤ X translates into an ellipsoid constraint on the mean parameter sequence [16, 17]. For this reason, Theorem 3 may be viewed as “regression version” of ℓ2 error bounds for the sequence model under an ellipsoid constraint (cf. Pinsker’s Theorem, [16, 17]). Due to the fact that the latter problem has a very well developed literature, there may be various “neighboring results” elsewhere. Nevertheless, we could not find a direct reference for our stated MSPE bound in the current setup. For the purposes of our work in this paper, the more important point to take away from Theorem 3 is that it can be coupled with Theorem 2 for proving consistency of RB. 3.4 Confidence intervals for mean responses, conditionally on a Gaussian design In this section, we consider the situation where the design matrix X has rows Xi⊤ ∈ Rp drawn i.i.d. from a multivariate normal distribution N (0, Σ), with X ⊥ ⊥ ε. (The covariance matrix Σ may vary with n.) Conditionally on a realization of X, we analyze the RB approximation of the laws Ψρ (F0 ; Xi ) = L(Xi⊤ (βbρ − β)|X). As discussed in Section 1, this corresponds to the problem of setting confidence intervals for the mean responses E[Yi ] = Xi⊤ β. Assuming that the population eigenvalues λi (Σ) obey a decay condition, we show below in Theorem 4 that RB succeeds with high PX -probability. Moreover, this consistency statement holds for all of the laws Ψρ (F0 ; Xi ) simultaneously. That is, among the n distinct laws Ψρ (F0 ; Xi ), i = 1, . . . , n, even the worst bootstrap approximation is still consistent. We now state some population-level assumptions. 6 A4. The operator norm of Σ ∈ Rp×p satisfies kΣkop . 1. Next, we impose a decay condition on the eigenvalues of Σ. This condition also ensures that Σ is invertible for each fixed p — even though the bottom eigenvalue may become arbitrarily small as p becomes large. It is important to notice that we now use η for the decay exponent of the population eigenvalues, whereas we used ν when describing the sample eigenvalues in the previous section. A5. There is a number η > 0, and absolute constants k1 , k2 > 0, such that for all i = 1, . . . , p, k1 i−η ≤ λi (Σ) ≤ k2 i−η . A6. There are absolute constants k3 , k4 ∈ (0, 1) such that for all n ≥ 3, we have the bounds k3 ≤ np ≤ k4 and p ≤ n − 2. The following lemma collects most of the effort needed in proving our final result in Theorem 4. Here it is also helpful to recall the notation ρ/n = n−γ and ̺/n = n−θ from Assumption 2. Lemma 2. Suppose that the matrix X ∈ Rn×p has rows Xi⊤ drawn i.i.d. from N (0, Σ), and that Assumptions 2–6 hold. Furthermore, assume that γ chosen so that 0 < γ < min{η, 1}. Then, the statements below are true. (i) (bias inequality) Fix any τ > 0. Then, there is an absolute constant κ0 > 0, such that for all large n, the following event holds with PX -probability at least 1 − n−τ − ne−n/16 , max b2ρ (X; Xi ) ≤ κ0 · n−γ · (τ + 1) log(n + 2). 1≤i≤n (15) (ii) (variance inequality) There are absolute constants κ1 , κ2 > 0 such that for all large n, the following event holds with γ PX -probability at least 1 − 4n exp(−κ1 n η ), 1 max 1≤i≤n vρ (X;Xi ) γ ≤ κ2 n1− η . (16) (iii) (mspe inequalities) η when Suppose that θ is chosen as θ = 2η/3 when η ∈ (0, 21 ), and that θ is chosen as θ = 1+η 1 η > 2 . Then, there are absolute constants κ3 , κ4 , κ5 , κ6 > 0 such that for all large n, ( 2η with PX -probability at least 1 − exp(−κ3 n2−4η/3 ), if η ∈ (0, 21 ) κ4 n− 3 η mspe(βb̺ |X) ≤ 2 − η+1 κ6 n with PX -probability at least 1 − exp(−κ5 n 1+η ), if η > 21 . Remarks. Note that the two rates in part (iii) coincide as η approaches 1/2. At a conceptual level, the entire lemma may be explained in relatively simple terms. Viewing the quantities mspe(βb̺ |X), b2ρ (X; Xi ) and vρ (X; Xi ) as functionals of a Gaussian matrix, the proof involves deriving concentration bounds for each of them. Indeed, this is plausible given that these quantities are smooth functionals of X. However, the difficulty of the proof arises from the fact that they are also highly non-linear functionals of X. We now combine Lemmas 1 and 2 with Theorems 1 and 2 to show that all of the laws Ψρ (F0 ; Xi ) can be simultaneously approximated via our two-stage RB method. Theorem 4. Suppose that F0 has a finite fourth moment, Assumptions 2–6 hold, and γ is chosen η so that 1+η < γ < min{η, 1}. Also suppose that θ is chosen as θ = 2η/3 when η ∈ (0, 12 ), and η when η > 12 . Then, there is a sequence of positive numbers δn with limn→∞ δn = 0, such θ = η+1 that the event  h  i   E max d22 √1vρ Ψρ (F0 ; Xi ), √1vρ Φρ (Fb̺ ; Xi )  (17)  X ≤ δn  1≤i≤n has PX -probability tending to 1 as n → ∞. Remark. Lemma 2 gives explicit bounds on the numbers δn , as well as the probabilities of the corresponding events, but we have stated the result in this way for the sake of readability. 7 4 Simulations In four different settings of n, p, and the decay parameter η, we compared the nominal 90% confidence intervals (CIs) of four methods: “oracle”, “ridge”, “normal”, and “OLS”, to be described below. In each setting, we generated N1 := 100 random designs X with i.i.d. rows drawn from N (0, Σ), where λj (Σ) = j −η , j = 1, . . . , p, and the eigenvectors of Σ were drawn randomly by setting them to be the Q factor in a QR decomposition of a standard p × p Gaussian matrix. Then, for each realization of X, we generated N2 := 1000 realizations of Y according to the model (1), where β = 1/k1k2 ∈ Rp , and F0 is the centered t distribution on 5 degrees of freedom, rescaled to have standard deviation σ = 0.1. For each X, and each corresponding Y , we considered the problem of setting a 90% CI for the mean response value Xi⊤⋆ β, where Xi⊤⋆ is the row with the highest leverage score, i.e. i⋆ = argmax1≤i≤n Hii and H := X(X ⊤ X)−1 X ⊤ . This problem was shown in B&F 1983 to be a case where the standard RB method based on least-squares fails when p/n ≍ 1. Below, we refer to this method as “OLS”. To describe the other three methods, “ridge” refers to the interval [Xi⊤⋆ βbρ − qb0.95 , Xi⊤⋆ βbρ − qb0.05 ], where qbα is the α% quantile of the numbers z1 , . . . , zB computed in the proposed algorithm in Section 2, with B = 1000 and c⊤ = Xi⊤⋆ . To choose the parameters ρ and ̺ for a given X and Y , we first computed rb as the value that optimized the MSPE error of a ridge estimator βbr with respect to 5-fold cross validation; i.e. cross validation was performed for every distinct pair (X, Y ). We then put ̺ = 5b r and ρ = 0.1b r, as we found the prefactors 5 and 0.1 to work adequately across various settings. (Optimizing ̺ with respect to MSPE is motivated by Theorems 1, 2, and 3. Also, choosing ρ to be somewhat smaller than ̺ conforms with the constraints on θ and γ in Theorem 4.) The method “normal” refers to the CI based on the normal approximation L(Xi⊤⋆ (βbρ −β)|X) ≈ N (0, τb2 ), where r, and σ b2 is the usual unbiased estimate of σ 2 based τb2 = σ b2 kXi⊤⋆ (X ⊤ X +ρIp×p )−1 X ⊤ k22 , ρ = 0.1b on OLS residuals. The “oracle” method refers to the interval [Xi⊤⋆ βbρ − q̃0.95 , Xi⊤⋆ βbρ − q̃0.05 ], with ρ = 0.1b r, and q̃α being the empirical α% quantile of Xi⊤ (βbρ − β) over all 1000 realizations of Y based on a given X. (This accounts for the randomness in ρ = 0.1b r.) Within a given setting of the triplet (n, p, η), we refer to the “coverage” of a method as the fraction of the N1 ×N2 = 105 instances where the method’s CI contained the parameter Xi⊤⋆ β. Also, we refer to “width” as the average width of a method’s intervals over all of the 105 instances. The four settings of (n, p, η) correspond to moderate/high dimension and moderate/fast decay of the eigenvalues λi (Σ). Even in the moderate case of p/n = 0.45, the results show that the OLS intervals are too narrow and have coverage noticeably less than 90%. As expected, this effect becomes more pronounced when p/n = 0.95. The ridge and normal intervals perform reasonably well across settings, with both performing much better than OLS. However, it should be emphasized that our study of RB is motivated by the desire to gain insight into the behavior of the bootstrap in high dimensions — rather than trying to outperform particular methods. In future work, we plan to investigate the relative merits of the ridge and normal intervals in greater detail. Table 1: Comparison of nominal 90% confidence intervals setting 1 n = 100, p = 45, η = 0.5 setting 2 n = 100, p = 95, η = 0.5 setting 3 n = 100, p = 45, η = 1 setting 4 n = 100, p = 95, η = 1 width coverage width coverage width coverage width coverage oracle 0.21 0.90 0.22 0.90 0.20 0.90 0.21 0.90 ridge 0.20 0.87 0.26 0.88 0.21 0.90 0.26 0.92 normal 0.23 0.91 0.26 0.88 0.22 0.91 0.23 0.87 OLS 0.16 0.81 0.06 0.42 0.16 0.81 0.06 0.42 Acknowledgements. MEL thanks Prof. Peter J. Bickel for many helpful discussions, and gratefully acknowledges the DOE CSGF under grant DE-FG02-97ER25308, as well as the NSF-GRFP. 8 A Proof of Theorem 1 Proof. Due to line (4) and Lemma 8.8 in B&F 1981,
d22 (Ψρ (F0 ; c), Φρ (Fb ; c)) = d22 Φρ (F0 ; c), Φρ (Fb ; c) + (c⊤ δ(X))2 . (18) If ε∗ ∈ Rn is a random vector whose entries are drawn i.i.d. from Fb , then the definition of Φρ gives the matching relations   Φρ (F0 ; c) = L(c⊤ (X ⊤ X + ρ)−1 X ⊤ ε X) (19)   ⊤ ⊤ −1 ⊤ ∗ Φρ (Fb; c) = L(c (X X + ρ) X ε  X). To make use of these relations, we apply Lemma 8.9 in B&F 1981, which implies that if w ∈ Rn is a generic deterministic vector, and if U = (U1 , . . . , Un ) and V = (V1 , . . . , Vn ) are random vectors with i.i.d. entries, then d22 (w⊤ U, w⊤ V ) ≤ kwk22 · d2 (U1 , V1 )2 . Therefore,
d22 Φρ (F0 ), Φρ (Fb) ≤ kc⊤ (X ⊤ X + ρ)−1 X ⊤ k22 · d22 (ε1 , ε∗1 ) = 1 σ2 (20) · vρ (X; c) · d22 (F0 , Fb ). Combining this with line (18) and dividing through by vρ proves the claim. B Proof of Theorem 2 Proof. By the triangle inequality, d22 (Fb , F0 ) ≤ 2 d22 (Fb, Fn ) + 2 d22 (Fn , F0 ). (21) b which places mass 1/n at Let F̃n be the (uncentered) empirical distribution of the residuals eb of β, each value ebi , for i = 1, . . . , n. The proofs of Lemmas 2.1 and 2.2 in Freedman 1981, show that     Pn
2       E d2 (Fb, Fn )2  + E d2 (F̃n , Fn )2  X ≤ E n1 i=1 εi X     (22) ≤ n1 σ 2 + n1 E kb e − εk22  X     b 2 = n1 σ 2 + n1 E kX(β − β)k 2 X , b which holds for any estimator β. b This where we have used the algebraic identity eb − ε = X(β − β), completes the proof. C Proof of Theorem 3 Proof. We begin with a simple bias-variance decomposition, i h  
2     X + mspe(βb̺ |X) = n1 E X βb̺ − E βb̺ |X   2 1 n   X(E βb̺ |X − β) 2 . 2 (23) We will handle the bias and variance terms separately. To consider the bias term, note that E[βbρ |X]− β = Bβ, where B = (X ⊤ X + ̺Ip×p )−1 X ⊤ X − Ip×p . Hence, 1 b n kX(E[β̺ |X] − β)k22 = 1 2 n kXBβk2 = β ⊤ B( n1 X ⊤ X)Bβ. 9 (24) If we let li = λi ( n1 X ⊤ X), then the eigenvalues of B( n1 X ⊤ X)B are of the form µi := In particular, it is simple to check1 that maxi µi ≍ ̺/n whenever θ ≤ ν, and so li (̺/n)2 (li +̺/n)2 . β ⊤ B( n1 X ⊤ X)Bβ . n̺ kβk22 = n−θ kβk22 . (25) Note that this bound is tight, since it is achieved whenever β is parallel to the top eigenvector of B( n1 X ⊤ X)B. To consider the variance term, note that βb̺ − E[βb̺ |X] = (X ⊤ X + ̺Ip×p )−1 X ⊤ ε, and so  h   i
2  1 b̺ − E[βb̺ |X]) 2  X = 1 tr X ⊤ X X ⊤ X + ̺Ip×p −1 E X( β  n n 2
2 Pn∧p li . = n1 i=1 li +̺/n It is natural to decompose the sum in terms of the index set I(n) := {i ∈ {1, . . . , n ∧ p} : li ≥ ̺/n}, (26) (27) θ ν which satisfies |I(n)| ≍ n . We will bound the variance term in two complementary cases; either ν > 1/2 or ν ≤ 1/2. First assume ν > 1/2. Then, 1 n p X i=1
2 li li +̺/n = 1 n X i∈I(n) . 1 n
2 li li +̺/n |I(n)| + 1 n Z + n∧p |I(n)| 1 n X i6∈I(n)
2 li li +̺/n (28) x−2ν (̺/n)2 dx (29)   θ using ν > . n−1 n ν + n2θ · (|I(n)|)1−2ν  θ  θ ≍ n−1 n ν + n2θ · (n ν )(1−2ν) 1 2 (30) (31) θ−ν ν = 2n . (32) To see that this upper bound is tight, note that in line (28), we can use the term-wise lower bounds ( 1 if i ∈ I(n)
2 4 li (33) ≥ l2i 1 li +̺/n if i 6∈ I(n), 4 (̺/n)2 and then apply an integral approximation from below (which leads to the same rate). Combining the bias and variance pieces, we have shown that θ−ν 1 if ν > 1 . EkX(βb̺ − β)k2 . n ν + n−θ 2 n 2 The bound is optimized when the two terms on the right side have the same rate, which leads to the ν . choice θ = ν+1 In the case where ν ∈ (0, 21 ), the calculation proceeds in the same way up to line (30), where we obtain the bound n∧p  θ  X
2 li 1 . n−1 n ν + n2θ · n1−2ν (34) n li +̺/n i=1 θ−ν (35) = n ν + n2(θ−ν) . This bound is also tight due to the same reasoning as above. Note that in order for the bound to tend to 0 as n → ∞, we must choose θ < ν. Furthermore, since we are working under the assumption ν ∈ (0, 12 ), it follows that the right side of line (35) has rate equal to n2(θ−ν) . Combining the rates for the bias and variance shows that 1 2 2(θ−ν) b + n−θ if ν ∈ (0, 21 ). n EkX(β̺ − β)k2 . n The bound is optimized when the two terms on the right side have the same rate, which leads to the choice θ = 2ν 3 . 2 t(̺/n) Note that if t ∈ R and f (t) := (t+̺/n) 2 , then f is maximized at t = ̺/n. Also, if θ ≤ ν, then there at least one li that scales at the rate of ̺/n. 1 10 D Proof of Lemma 2 The proof is split up into three pieces, corresponding to parts (i), (ii), and (iii) in the statement of the result. D.1 The bias inequality (15) We prove inequality (15) by combining Lemmas 3 and 5 below. Lemma 3. Assume the conditions of Lemma 2. For each i ∈ {1, . . . , n}, there are independent random vectors ui (X), w(X) ∈ Rp such that the random variable Xi⊤ δ(X) can be represented algebraically as bρ (X; Xi ) = Xi⊤ δ(X) = ui (X)⊤ w(X). Here, the vectors ui (X) can be represented in law as L ui (X) = 1 kzk2 Πp (z), (36) where z ∈ Rn is a standard Gaussian vector, and Πp (z) := (z1 , . . . , zp ). Also, the vector w(X) satisfies the bound kw(X)k22 ≤ ρ4 kβk22 almost surely. Proof. To fix notation, we write X ⊤ = Σ1/2 Z ⊤ where Z ⊤ ∈ Rp×n is a standard Gaussian matrix. Recall that δ(X) = Bβ, where B = Ip×p − (X ⊤ X + ρIp×p )−1 X ⊤ X. Let Z = HLG⊤ be a “signed s.v.d.” for Z, as defined in Appendix E.2, where H ∈ Rn×p , L ∈ Rp×p , and G ∈ Rp×p . Now define ui (X) and w(X) according to ⊤ 1/2 Xi⊤ δ(X) = e⊤ Bβ = i XBβ = ei ZΣ e⊤ H LG⊤ Σ1/2 Bβ . {z } |i{z } | =: ui (X)⊤ =: w(X) From Lemma 17 in Appendix E.2, the rows e⊤ i H can be represented in distribution as The same lemma also shows that the three matrices H, L, and G are independent. (37) 1 kzk2 Πp (z). Hence, to show that ui (X) and w(X) are independent, it suffices to show that w(X) is a function only of G and L. In turn, it is enough to show that B is a function only of G and L. But this is simple, because B is a function only of the matrix X ⊤ X, which may be written as X ⊤ X = Σ1/2 Z ⊤ ZΣ1/2 = Σ1/2 GL2 G⊤ Σ1/2 . (38) It remains to show that kw(X)k22 ≤ ρ4 kβk22 almost surely. Combining the definition of w(X) with line (38) gives
(39) kw(X)k22 = β ⊤ BX ⊤ XB β. 2 li 1 ⊤ The eigenvalues of BX ⊤ XB are of the form µi := n (l(ρ/n) 2 where li = λi ( n X X), and it is i +ρ/n) ρ simple to check that the inequality maxi µi ≤ 4 holds for every realization of X. Before proceeding to the second portion of the proof of inequality (15), we record some well-known tail bounds for Gaussian quadratic forms due to Laurent and Massart [18], which will be useful at various points later on. Lemma 4 (Laurent & Massart, 2001). Let A ∈ Rn×n be a fixed symmetric matrix, and let z ∈ Rn be a standard Gaussian vector. Then, for every t > 0, h i √ P z ⊤ Az ≥ tr(A) + 2 |||A|||F t + 2 |||A|||op t ≤ exp(−t) (40) and h √i P z ⊤ Az ≤ tr(A) − 2 |||A|||F t ≤ exp(−t). 11 (41) The next lemma completes the proof of inequality (15). Lemma 5. Assume the conditions of Lemma 2, and let τ > 0 be a constant. Then for every n ≥ 1, the following event holds with probability at least 1 − n−τ − ne−n/16 , max b2ρ (X; Xi ) ≤ 5kβk22 · n−γ · (τ + 1) log(n + 2). 1≤i≤n (42) Proof. Applying the representation for bρ (X; Xi ) given in Lemma 3, there is a standard Gaussian L vector z ∈ Rn , such that ui (X) = Πp (z)/kzk2. Consequently,   L 1 ⊤ b2ρ (X; Xi ) = kzk Πp (z), 2 · Πp (z) w(X)w(X) (43) The same lemma also implies that for all t′ ∈ (0, 14 ), i h 1 1 √ ≤ exp(−nt′ ). P kzk 2 ≥ (1−2 t′ )n (45) 2 where we may take z and w(X) to be independent by the same lemma. Using Lemma 4 on Gaussian quadratic forms, as well as the fact that kw(X)k22 ≤ ρ4 kβk22 almost surely, we have for all t > 0,  " #      √  ρ ⊤ ⊤ 2 P Πp (z) w(X)w(X) Πp (z) ≥ 4 kβk2 1 + 2 t + 2t) w(X) ≤ exp(−t). (44)     2 Now, we combine the bounds by integrating out w(X) in line (44) and choosing t′ = 1/16 in line (45). Taking a union bound, we conclude that for any t > 0, and any fixed i = 1, . . . , n, " # √
ρ 2 2 1 P bρ (X; Xi ) ≤ n · kβk2 · 2 1 + 2 t + 2t ≥ 1 − e−t − e−n/16 . (46) Finally, another union bound shows that the maximum of the bρ (X, Xi⊤ ) satisfies P " max 1≤i≤n b2ρ (X; Xi ) ≤ ρ n · kβk22 · 1 2 (1 # √ + 2 t + 2t) ≥ 1 − e−t+log(n) − ne−n/16 , (47) which implies the √stated result after choosing t = (τ + 1) log(n + 2), and noting that since t ≥ 1, we have 21 (1 + 2 t + 2t) ≤ 5t = 5(τ + 1) log(n + 2), as well as e−t+log(n) ≤ e−τ log(n+2) ≤ n−τ for every n ≥ 1. D.2 The variance inequality (16) The following “representation lemma” will serve as the basis for controlling the variance vρ (X; Xi ) = σ 2 kXi⊤ (X ⊤ X + ρIp×p )−1 X ⊤ k22 . Lemma 6. Assume the conditions of Lemma 2. For each i ∈ {1, . . . , n}, there is a random vector vi (X) ∈ Rp and a random matrix M (X) ∈ Rp×p that are independent and satisfy the algebraic relation kXi⊤ (X ⊤ X + ρIp×p )−1 X ⊤ k22 = vi (X)⊤ M (X)vi (X). Here, the vector vi (X) can be represented in law as L vi (X) = 1 kzk2 Πp (z), (48) where z ∈ Rn is a standard Gaussian vector and Πp (z) = (z1 , . . . , zp ). Also, the matrix M (X) satisfies the algebraic relation tr(M (X)) = kX(X ⊤X + ρIp×p )−1 X ⊤ k2F . An explicit formula for M (X) is given below. 12 (49) Proof. Define the matrix A := (X ⊤ X + ρIp×p )−1 X ⊤ X(X ⊤X + ρIp×p )−1 . Then, ⊤ kXi⊤ (X ⊤ X + ρIp×p )−1 X ⊤ k22 = e⊤ i XAX ei . (50) Using the notation in the proof of the previous lemma, let X = ZΣ1/2 where Z ∈ Rn×p is a standard Gaussian random matrix. Furthermore, let Z = HLG⊤ be a signed s.v.d. for Z, as defined in Appendix E.2. Then, we define vi (X) and M (X) according to ⊤ ⊤ e⊤ LG⊤ Σ1/2{z AΣ1/2 GL⊤} H ⊤ ei . i XAX ei = ei H | | {z } (51) =:M(X) =:vi (X)⊤ Some algebra shows that M (X) satisfies the relation (49). As in the proof of Lemma 3, the argument is completed using two properties of the signed s.v.d. of a standard Gaussian matrix: The rows of H can be represented as Πp (z)/kzk2 where z ∈ Rn is a standard Gaussian vector, and the matrices H, L, and G⊤ are independent. (See Lemma 17 in Appendix E.2.) To show that vi (X) and M (X) are independent, first note that vi (X) only depends on H. Also, it is simple to check that M (X) only depends on G and L, because A is a function only of X ⊤ X = Σ1/2 GL2 G⊤ Σ1/2 . D.2.1 Concentration of the variance and bounds on its expected value Due to Lemma 6, for each i = 1, . . . , n, we have the representation L vρ (X; Xi ) = 1 Π (z)⊤ M (X)Πp (z), kzk22 p (52) where z ∈ Rn is a standard Gaussian vector, independent of M (X). Conditionally on M (X), the quadratic form Πp (z)⊤ M (X)Πp (z) concentrates around tr(M (X)) by Lemma 4. The same 2 lemma p also implies that kzk2 concentrates around n. In the next three subsections, we will show that tr(M (X)) concentrates around its expected value, and obtain upper and lower bounds on the expected value. We will need two-sided bounds in preparation for Theorem 4. D.2.2 Concentration of p tr(M (X)) Lemma 7. Assume the conditions of Lemma 2. Then for every t > 0, and every n ≥ 1, # " p p 64 n1−γ t2 tr(M (X)) − E tr(M (X)) ≥ t ≤ 2 exp(− 54 P kΣkop ). (53) p Proof. We will show that tr(M (X)) is a Lipschitz function of a standard Gaussian matrix. Define s2 , which satisfies the Lipschitz condition the function gρ : R+ → [0, 1] by gρ (s) = s2 +(ρ/n) |gρ (s) − gρ (s′ )| ≤ Ln |s − s′ |, for all s, s′ ≥ 0, where Ln := √ 3 3 8 √1 . ρ/n If σ(A) = (σ1 (A), . . . , σk (A)) denotes the vector of singular values of a rank k matrix A, then we define gρ to act on σ(A) component-wise, i.e. gρ (σ(A)) = (gρ (σ1 (A)), . . . , gρ (σk (A))). Recall from Lemma 6 that p (54) tr(M (X)) = kX(X ⊤X + ρIp×p )−1 X ⊤ kF and note that the ith singular value of the matrix X(X ⊤ X +ρIp×p )−1 X ⊤ is given by gρ (σi ( √1n X)). Viewing the Frobenius norm of a matrix as the ℓ2 norm of its singular values, we have p tr(M (X)) = kgρ (σ( √1n X))k2 . (55) Write X ⊤ = Σ1/2 Z ⊤ for a standard Gaussian matrix Z ∈ Rn×p , and let f : Rn×p → R be defined according to p f (Z) := tr(M (X)). 13 We claim that f is Lipschitz with respect to the Frobenius norm. Let W ⊤ ∈ Rp×n be a generic matrix, and put A = √1n Σ1/2 Z ⊤ and B = √1n Σ1/2 W ⊤ . Then, |f (Z) − f (W )| = kgρ (σ(A))k2 − kgρ (σ(B))k2 ≤ gρ (σ(A)) − gρ (σ(B)) (56) (57) 2 ≤ Ln kσ(A) − σ(B)k2 (58) ≤ Ln kA − BkF (59) √1 Σ1/2 n = Ln ≤ Ln √ n (Weilandt-Hoffman)
Z⊤ − W ⊤ F p kΣkop · Z ⊤ − W ⊤ F , (60) (61) where we have used a version of the Weilandt-Hoffman inequality for singular values [19, p.186], as well as the inequality kM1 M2 kF ≤ kM1 kop kM2 kF , which holds for any square matrix M1 that is compatible with M2 . (See Lemma 12 in Appendix E.) The statement of the lemma now follows from the Gaussian concentration inequality. (See Lemma 15 in Appendix E). p D.2.3 Upper bound on E tr(M (X)) Lemma 8. Assume the conditions of Lemma 2. Then, the matrix M (X) satisfies ( 1 p n(γ−η)+ 2 if η ∈ (0, 21 ) γ E tr(M (X)) . if η > 21 . n 2η (62) p Proof. By Jensen’s inequality, it is enough to bound E[tr(M (X))] from above. Define the unis variate function ψ : R+ → R+ by ψ(s) := (√s+ρ/n) 2 , and observe that  2  tr(M (X)) = tr (X ⊤ X + ρIp×p )−1 X ⊤ X = p X b λ2i (Σ) 2 b (λi (Σ)+ρ/n) i=1 b 2 )) ψ(λi (Σ i=1 = p X (63)
b 2) . = tr ψ(Σ Here where we use the “operator calculus” notation ψ(A) = U ψ(D)U ⊤ where A is a symmetric matrix with spectral decomposition A = U DU ⊤ , and ψ(D) is the diagonal matrix whose ith diagb 2 )) is a concave matrix onal entry is ψ(Dii ). It is simple to check that ψ is a concave, and so tr(ψ(Σ 2 b functional of Σ by Lemma 13 in Section E.1 of Appendix A. Therefore, Jensen’s inequality implies b 2 ])) E[tr(M (X))] ≤ tr(ψ(E[Σ (64) P = pi=1 ψ(λi (S)), b 2 ]. Since X is Gaussian, Σ b is a Wishart matrix up to scaling, where we define the matrix S := E[Σ and so Lemma 16 in Appendix E shows that this expectation may be evaluated exactly as S = (1 + n1 )Σ2 + tr(Σ) n Σ. (65) We will now use this relation to apply an integral approximation to the right side of line (64). Clearly, the eigenvalues of S are given by λi (S) = (1 + n1 )λ2i (Σ) + ≍ i−2η + tr(Σ) −η . n i 14 tr(Σ) n λi (Σ) (66) Let r ∈ (0, 1) be a constant to be specified later. On the set of indices 1 ≤ i ≤ ⌈nr ⌉ we use the bound 1 ψ(λi (S)) ≤ 1, and on the set of indices i > ⌈nr ⌉ we use the bound ψ(λi (S)) ≤ (ρ/n) 2 λi (S). −γ 2 Recalling the assumption ρ/n = n , we may decompose the inequality (64) as E[tr(M (X))] ≤ ⌈nr ⌉ X ψ(λi (S)) + i=1 . nr + n2γ Z p X ψ(λi (S)) (67) i=⌈nr ⌉+1 p x−2η + nr =: nr + n2γ hn (η, r). tr(Σ) −η
dx n x (68) (69) where the function hn is defined in the last line. The bound is optimized when the two terms on the right are of the same order; i.e. when r solves the rate equation nr ≍ n2γ hn (η, r). Noting that tr(Σ) n ≍  −η n n−1 (70) if η ∈ (0, 1) if η > 1, the quantity hn (η, r) may be computed directly as  1−2η n hn (η, r) ≍ nr(1−2η)  r(1−2η) n + nr(1−η)−1 if η ∈ (0, 12 ), if η ∈ ( 12 , 1), if η > 1. (71) (72) If we let r = r∗ (η, γ) denote the solution of the rate equation (70), then some calculation shows that under the assumption γ ∈ (0, 1), ( 2(γ − η) + 1 if η ∈ (0, 12 ), (73) r∗ (η, γ) = γ if η > 21 . η When η ∈ (0, 1) this is straightforward. To show the details for η > 1, note that the rate equation (70) may be written as nr ≍ n2γ+r(1−2η) + n2γ+r(1−η)−1 , (74) which is the same as 1 ≍ n2(γ−ηr) + n2γ−ηr−1 . In order for both terms on the right to be O(1), the number r must satisfy the constraints r ≥ γη , r≥ γ η + (75) (76) γ−1 η . (77) Since Lemma 2 assumes γ ∈ (0, 1), only the first constraint matters. Furthermore, when r ≥ γη , the second term in line (75) is o(1), and we are reduced to choosing r so that 1 ≍ n2(γ−ηr) , which gives r = r∗ (η, γ) = γη . Substituting this value into line (69) completes the proof. (Note from the discussion preceding line (69) that r must lie in the interval (0, 1), and this requires γ/η < 1, which explains the assumption γ < min{η, 1} in Lemma 2.) p D.2.4 Lower bound on E tr(M (X)) Lemma 9. Assume the conditions of Lemma 2. Then, the matrix M (X) satisfies p γ E tr(M (X)) & n 2η . (78) 2 Note that if p/n ≍ 1, it is possible that nr > p for small values of n. Since we want nr ≤ p for the integral in line (68), Lemma 2 is stated for “all large n”. 15 p Proof. The variable tr(M (X)) may be written as kX ⊤X(X ⊤ X +ρIp×p )−1 kF . Since the Frobenius norm is a convex matrix functional, Jensen’s inequality implies i h p E tr(M (X)) ≥ E X ⊤ X(X ⊤ X + ρIp×p )−1 F (79) i h
b −1 −1 , = E Ip×p + nρ Σ F b := where the last step follows algebraically with Σ If we define the univariate function f : R+ → R+ by f (s) = (1 + nρ s)−1 , then last inequality is the same as p
  b −1 . (80) E tr(M (X)) ≥ E f Σ F 1 ⊤ n X X. It is a basic fact that f is operator convex on the domain of positive semidefinite matrices [20, p.117]. This yields an operator version of Jensen’s inequality with respect to the Loewner ordering (Lemma 14 in Appendix E):
  −1 
 b b −1  f E Σ . (81) E f Σ Furthermore, if two matrices satisfy A  B  0, then kAkF ≥ kBkF [21, Corollary 7.7.4]. Using this fact, as well as the formula for the expected inverse of a Wishart matrix [22, p. 97], we obtain p  −1 
b E tr(M (X)) ≥ f E Σ F = f = n −1 n−p−1 Σ p X i=1 = p X i=1 1+ ρ n ·
F 1
n −1 ) 2 n−p−1 λi (Σ λ2i (Σ) λi (Σ) + ρ n ·
2 n n−p−1 !1/2 !1/2 (82) .  n . For any i ∈ J, the ith summand in Define the index set J = i ∈ {1, . . . , p} : λi (Σ) ≥ nρ n−p−1 the previous line is at least 1/4. Also, assumption A6 that p/n is bounded strictly between 0 and 1, as well as the decay condition on the λi (Σ), imply that |J| ≍ nγ/η , which completes the proof. D.2.5 Putting the variance pieces together Combining Lemmas 7, 8, and 9 with the Gaussian concentration inequality (Lemma 15 in Section E.2 of Appendix A) p immediately gives the following result. (We choose t to be proportional to the relevant bound on E[ tr(M (X))] in the Gaussian concentration inequality.) Lemma 10. Assume the conditions of Lemma 2 and let tr(M (X)) be as in line (49). Then, there are absolute constants κ1 , κ2 , . . . , κ6 > 0 such that the following upper-tail bounds hold for all large n, i h (83) P tr(M (X)) ≥ κ1 n2(γ−η)+1 ≤ exp(−κ2 n2(1−η)+γ ), if η ∈ (0, 21 ), and i h γ(1−η) P tr(M (X)) ≥ κ3 nγ/η ≤ exp(−κ4 n1+ η ), and the following lower-tail bound holds for all large n, i h γ(1−η) P tr(M (X)) ≤ κ5 nγ/η ≤ exp(−κ6 n1+ η ), if η > 12 , (84) if η > 0. (85) Remarks. Note that in order for the last two probabilities to be small for large values of η > 0, it is necessary that γ < 1, as assumed in Lemma 2. The next result completes the assembly of the results in this Subsection D.2. Although the first two bounds in Lemma 11 are not necessary for the statement of Theorem 4, they show that the variance vρ (X; Xi ) tends 0 as n → ∞ when γ < η, as assumed in Theorem 4. In other words, we imposed the assumption γ < η so that confidence intervals based on Φρ (Fb̺ ; Xi ) have width that tends to 0 asymptotically. 16 Lemma 11. Assume the conditions of Theorem 4 and let tr(M (X)) be as in line (49). Assume γ < min{η, 1}. Then, there are absolute constants k1 , k2 , . . . , k6 > 0 such that the following upper-tail bounds hold for all large n, i h γ P max vρ (X; Xi ) ≤ k1 n2(γ−η) ≥ 1 − 4n exp(−k2 n η ), if η ∈ (0, 21 ) (86) 1≤i≤n and i h γ γ P max vρ (X; Xi ) ≤ k3 n η −1 ≥ 1 − 4n exp(−k4 n η ), 1≤i≤n and h P max 1 1≤i≤n vρ (X;Xi ) i γ γ ≤ k5 n1− η ≥ 1 − 4n exp(−k6 n η ), if η > 12 , (87) if η > 0. (88) Proof. We only prove the last inequality (88), since the other two inequalities are proven in a similar way. By Lemma 6, we have L vρ (X; Xi ) = 1 Π (z)⊤ M (X)Πp (z) kzk22 p (89) where z ∼ N (0, Ip×p ) and z ⊥ ⊥ M (X). To apply the lower-tail bound for Gaussian quadratic p forms, note that Hölder’s inequality implies kM (X)kF ≤ tr(M (X)) since kM (X)kop ≤ 1 almost surely. Therefore, letting t = t′ tr(M (X)) with t′ ∈ (0, 1) in inequality (41) gives  i h √ 
 ′ (90) P Πp (z)⊤ M (X)Πp (z) ≥ (1 − 2 t′ ) tr(M (X))  M (X) ≥ 1 − exp − t tr(M (X)) Next, observe that inequality (41) with t = t′ · n for t′ ∈ (0, 1) gives, h i √ P kzk22 ≤ (1 + 4 t′ )n ≥ 1 − exp(−t′ n). If we define the event (91) ) (92) then the previous two inequalities imply     P E1  M (X) ≥ 1 − exp(−t′ tr(M (X))) − exp(−t′ · n) (93) E1 := ( 1 kzk22 1 ≤ Πp (z)⊤ M (X)Πp (z) √ n 1+4√t′ (1−2 t′ ) tr(M(X)) ≥ 1 − 2 exp(−t′ tr(M (X))), since tr(M (X)) ≤ n almost surely. Next, let κ5 , κ6 > 0 be as in the previous lemma, and define the event o n γ n (94) E2 := tr(M(X)) ≤ κ15 n1− η , which has probability P(E2 ) ≥ 1 − exp(−κ6 n1+ γ(1−η) η ). We now put these items together. Starting with line (89), if we work on the intersection of E1 and E2 , then for any fixed i = 1, . . . , n we have P h 1 vρ (X;Xi ) ≤ √ 1− γ 1+4√t′ 1 n η (1−2 t′ ) κ5 i h i ≥ E 1E1 · 1E2 h i ≥ 1 − E 1E1c + 1E2c      c = 1 − E E 1E1c  M (X) − P(E2 ) h i ≥ 1 − 2E exp(−t′ tr(M (X)) − P(E2c) h ′ i = 1 − 2E exp(−t tr(M (X))) · (1E2 + 1 ) − E2c γ ≥ 1 − exp(−t′ κ5 n η ) − 3P(E2c ) γ 1+ ≥ 1 − exp(−t′ κ5 n η ) − 3 exp(−κ6 n γ ≥ 1 − 4 exp(− min{t′ κ5 , κ6 } · n η ) 17 γ(1−η) η ) (95) P(E2c ) where we have used the previous lemma to bound P(E2c ), and also the assumption γ ∈ (0, 1) to conclude that γη ≤ 1 + γ(1−η) . Taking a union bound over i = 1, . . . , n, proves the claim. η The last component of Lemma 2 is to prove the MSPE inequalities. D.3 Proof of the MSPE inequalities The proof of Theorem (3) shows that for any realization of X we have h i   mspe(βb̺ |X) := n1 E kX(βb̺ − β)k22  X . n−θ kβk22 + 1 n p∧n X i=1
2 li li +̺ , (96) where li = λi ( n1 X ⊤ X). Now observe that the second term on the right side matches the expression for tr(M (X)) given in line (63) by replacing ρ with ̺ and multiplying by a factor of n1 . Therefore, using Lemma 10 and recalling ̺/n = n−θ shows that there are absolute constants κ1 , κ2 , κ3 , κ4 > 0 such that for all large n, and h
i P mspe(βb̺ |X) ≥ κ1 n−θ + n2(θ−η) ≤ exp(−κ2 n2(1−η)+θ ), if η ∈ (0, 12 ). h θ(1−η)
i θ P mspe(βb̺ |X) ≥ κ3 n−θ + n η −1 ≤ exp(−κ4 n1+ η ), if η > 21 . (97) (98) In line (97), the bound mspe(βb̺ |X) is optimized when n−θ ≍ n2(θ−η) , which explains the choice θ in line (98), the bound is optimized when n−θ ≍ n η −1 , which explains the θ = 2η 3 . Similarly, η choice θ = η+1 . Substituting in these values θ yields the stated result. E Background results E.1 Results on matrices and convexity Lemma 12. Let M1 ∈ Rk1 ×k1 and M2 ∈ Rk1 ×k2 . Then, kM1 M2 kF ≤ kM1 kop kM2 kF . (99) Proof. Observe that kM1 M2 k2F = tr(M2⊤ M1⊤ M1 M2 ) = tr((M1⊤ M1 )(M2 M2⊤ )) ≤ k1 X i=1 λi (M1⊤ M1 ) · λi (M2 M2⊤ ) ≤ kM1 k2op k1 X (100) λi (M2 M2⊤ ) i=1 2 kM1 kop kM2 k2F , = where we have used von Neumann’s trace inequality (also known as Fan’s inequality) [23, p.10] in the third line. A result on convex trace functionals. In the following lemma, an interval of the real line refers to any set of the form (a, b),(a, b],[a, b), or [a, b], where −∞ ≤ a ≤ b ≤ ∞. We also define spec(M ) p×p to be the set of eigenvalues of a square matrix M . The collection of symmetric is Pmatrices in R p×p denoted by S . For a univariate function ϕ, the symbol tr(ϕ(M )) denotes i ϕ(λi (M )). Lemma 13. Let I ⊂ R be an interval, and let M ⊂ Sp×p be a convex set such that spec(M ) ⊂ I for all M ∈ M. Let ϕ : I → R be a convex function. Then, the functional M 7→ tr(ϕ(M )) (101) is convex on M. A proof may be found in the paper [24, Proposition 2]. 18 Operator Jensen inequality. A function f : Sp×p → Sp×p is said to be operator convex if for all λ ∈ [0, 1], and all A, B ∈ Sp×p , f (λA + (1 − λ)B)  λf (A) + (1 − λ)f (B), (102) where A  B means that B − A is positive semidefinite. Lemma 14 (Operator Jensen inequality). Suppose f : Sp×p → Sp×p is operator convex, and let A be a random Sp×p -valued matrix that is integrable. Then, f (E[A])  E[f (A)]. (103) Proof. It is enough to show that for all x ∈ Rp , x⊤ f (E[A])x ≤ x⊤ E[f (A)]x. (104) For any fixed x, consider the function g : Sp×p → R defined by g(A) = x⊤ f (A)x. It is clear that g is a convex function in the usual sense, and so the ordinary version of Jensen’s inequality implies g(E[A]) ≤ E[g(A)], which is the same as (104). E.2 Results on Gaussian vectors and matrices The following lemma is standard and is often referred to as the Gaussian concentration inequality [25]. Lemma 15. Let Z ∈ Rp be a standard Gaussian vector and let f : Rp → R be an L-Lipschitz function with respect to the ℓ2 norm. Then for all t > 0,  
−t2 (105) P |f (Z) − E[f (Z)]| ≥ t ≤ 2 exp 2L 2 . Next, we give a formula for the expected square of a Wishart matrix. b= Lemma 16. Let X ∈ Rn×p have rows drawn i.i.d. from N (0, Σ), and let Σ b = Proof. Write Σ then 1 n Pn i=1 b 2 ] = (1 + 1 )Σ2 + E[Σ n 1 ⊤ n X X. Then, tr(Σ) n Σ. Xi Xi⊤ where Xi⊤ ∈ Rp is the ith row of X. If we put Mi = Xi Xi⊤ , b 2 = 12 Pn M 2 + 12 P Mi Mj . Σ i i6=j i=1 n n Clearly, the Mi are independent with E[Mi ] = Σ for all i, and so
b 2 ] = 1 E[M 2 ] + 22 n Σ2 . E[Σ 1 n n 2 It remains to compute E[M12 ]. Write Xi = Σ1/2 Zi where Zi is a standard Gaussian vector in Rp , and let Σ1/2 = U Λ1/2 U ⊤ be a spectral decomposition of Σ1/2 where U ∈ Rp×p is orthogonal, and Λ is diagonal with Λii = λi (Σ). By the orthogonal invariance of the normal distribution, L Xi = U Λ1/2 Zi , and so h i E[M12 ] = E[X1 X1⊤ X1 X1⊤ ] = U Λ1/2 E Z1 Z1⊤ ΛZ1 Z1⊤ Λ1/2 U ⊤ . (106) Define the matrix M̃1 := Z1 Z1⊤ ΛZ1 Z1⊤ . It is straightforward to verify that E[M̃1 ] is diagonal, and its jth diagonal entry is E[M̃1 ]jj = tr(Σ) + 2λj . Therefore, E[M̃1 ] = tr(Σ)Ip×p + 2Λ, and combining this with line (106) gives E[M12 ] = tr(Σ)Σ + 2Σ2 . 19 Signed s.v.d. The following lemma describes the factors of an s.v.d. of a standard Gaussian matrix Z ∈ Rn×p with n ≥ p. To make the statement of the lemma more concise, we define the term signed s.v.d. below. This is merely a particular form of the s.v.d. that ensures uniqueness. Specifically, if Z ∈ Rn×p is a full rank matrix with n ≥ p, then the signed s.v.d. of Z is given by Z = HLG⊤ , (107) where H ∈ Rn×p has orthonormal columns, the matrix L ∈ Rp×p is diagonal with L11 ≥ L22 ≥ · · · ≥ Lpp > 0, and G ∈ Rp×p is orthogonal with its first row non-negative, i.e. G1i ≥ 0 for all i = 1, . . . , p. It is a basic fact from linear algebra that the signed s.v.d. of any full rank matrix in Rn×p exists and is unique [21, Lemma 7.3.1]. This fact applies to Gaussian matrices, since they are full rank with probability 1. Lemma 17. Suppose n ≥ p, and let Z ∈ Rn×p be a random matrix with entries drawn i.i.d. from N (0, 1). Let Z = HLG⊤ , (108) be the unique signed s.v.d. of Z as defined above. Then, the matrices H, L and G are independent. Furthermore, if Hi⊤ ∈ Rp denotes the ith row of H, then for each i = 1, . . . , n, the marginal law of Hi⊤ is given by L Hi⊤ = 1 kzk2 · Πp (z), (109) where z ∈ Rn is a standard Gaussian vector, and Πp (z) is the projection operator onto the first p coordinates, i.e. Πp (z) = (z1 , . . . , zp ). Proof. We first argue that H, L, and G are independent, and then derive the representation for Hi⊤ in the latter portion of the proof. Due to the fact that the transformation Z 7→ (H, L, G) is invertible, it is possible to obtain the joint density of (H, L, G) from the density of Z by computing the matrix Jacobian of the factorization Z = HLG. (See the references [22], [26], and [27] for more background on Jacobians of matrix factorizations.) To speak in more detail about the joint density, let Vn×p denote the Stiefel manifold of n×p matrices with orthonormal columns. Also, let Dp×p denote the set of p×p diagonal matrices, and Op×p the set of orthogonal p × p matrices. The subset of Op×p with non-negative entries in the n×p . first row will be denoted by O+ Let fH,L,G : Vn×p × Dp×p × Op×p → [0, ∞) denote the joint density of (H, L, G), where the base + measure is the product of Haar measure on Vn×p , Lebesgue measure on Dp×p , and Haar measure on Op×p restricted to Op×p + . (See the book [28] for background on these measures). From lines 8.8-8.10 in the paper [29], it is known that such a density exists, and that fH,L,G factors according to fH,L,G(h, l, g) = fH (h) · fL (l) · fG (g), (110) where the function fH denotes the density of H with respect to Haar measure on Vn×p , the function fL denotes the density of L with respect to Lebesgue measure on Dp×p , and the function fG denotes the density of G with respect to the restriction of Haar measure to Op×p + . This proves that H, L, and G are independent. We now prove the representation (109). From line 8.10 in the paper [29], it is known that fH is constant on Vn×p . In other words, the matrix H follows the Haar distribution on Vn×p . Consequently, Theorem 2.2.1(ii) in the book [28] implies that the rows Hi⊤ can be represented as L Hi⊤ = Πp (Ji⊤ ) (111) where Ji⊤ ∈ Rn is the ith row of a Haar-distributed random matrix in On×n . Furthermore, the rows Ji⊤ are uniformly distributed on the unit sphere on Rn , and hence can be represented as z/kzk2, where z ∈ Rn is a standard Gaussian vector. 20 References [1] C.-H. Zhang and S. S. Zhang. Confidence intervals for low dimensional parameters in high dimensional linear models. Journal of the Royal Statistical Society: Series B, 76(1):217–242, 2014. [2] A. Javanmard and A. Montanari. Hypothesis testing in high-dimensional regression under the Gaussian random design model: Asymptotic theory. arXiv preprint arXiv:1301.4240, 2013. [3] A. Javanmard and A. Montanari. Confidence intervals and hypothesis testing for highdimensional regression. arXiv preprint arXiv:1306.3171, 2013. [4] P. Bühlmann. Statistical significance in high-dimensional linear models. 19(4):1212–1242, 2013. Bernoulli, [5] S. van de Geer, P. Bühlmann, and Y. Ritov. On asymptotically optimal confidence regions and tests for high-dimensional models. arXiv preprint arXiv:1303.0518, 2013. [6] J. D. Lee, D. L. Sun, Y. Sun, and J. E. Taylor. Exact inference after model selection via the lasso. arXiv preprint arXiv:1311.6238, 2013. [7] A. Chatterjee and S. N. Lahiri. Rates of convergence of the adaptive lasso estimators to the oracle distribution and higher order refinements by the bootstrap. The Annals of Statistics, 41(3):1232–1259, 2013. [8] H. Liu and B. Yu. Asymptotic properties of lasso+mls and lasso+ridge in sparse highdimensional linear regression. Electronic Journal of Statistics, 7:3124–3169, 2013. [9] V. Chernozhukov, D. Chetverikov, and K. Kato. Gaussian approximations and multiplier bootstrap for maxima of sums of high-dimensional random vectors. The Annals of Statistics, 41(6):2786–2819, 2013. [10] E. L. Lehmann and J. P. Romano. Testing Statistical Hypotheses. Springer, 2005. [11] D. A. Freedman. Bootstrapping regression models. The Annals of Statistics, 9(6):1218–1228, 1981. [12] P. J. Bickel and D. A. Freedman. Bootstrapping regression models with many parameters. In Festschrift for Erich L. Lehmann, pages 28–48. Wadsworth, 1983. [13] N. R. Draper and H. Smith. Applied Regression Analysis. Wiley-Interscience, 1998. [14] P. J. Bickel and D. A. Freedman. Some asymptotic theory for the bootstrap. The Annals of Statistics, pages 1196–1217, 1981. [15] S. Bobkov and M. Ledoux. One-dimensional empirical measures, order statistics, and Kantorovich transport distances. preprint, 2014. [16] A. B. Tsybakov. Introduction to Nonparametric Estimation. Springer, 2009. [17] L. Wasserman. All of Nonparametric Statistics. Springer, 2006. [18] B. Laurent and P. Massart. Adaptive estimation of a quadratic functional by model selection. Annals of Statistics, 28(5):1302–1338, 2000. [19] R. A. Horn and C. R. Johnson. Topics in matrix analysis. Cambridge University Press, 1991. [20] R. Bhatia. Matrix Analysis. Springer, 1997. [21] R. A. Horn and C. A. Johnson. Matrix Analysis. Cambridge University Press, 22nd printing edition, 2009. [22] R. J. Muirhead. Aspects of Multivariate Statistical Theory. John Wiley & Sons, 1982. [23] J. M. Borwein and A. S. Lewis. Convex Analysis and Nonlinear Optimization Theory and Examples. CMS Bookks in Mathematics. Canadian Mathematical Society, 2000. 21 [24] D. Petz. A survey of certain trace inequalities. Functional analysis and operator theory, 30:287–298, 1994. [25] S. Boucheron, G. Lugosi, and P. Massart. Concentration Inequalities: A Nonasymptotic Theory of Independence. Oxford University Press, 2013. [26] A. M. Mathai. Jacobians of Matrix Transformations and Functions of Matrix Argument. World Scientific, 1997. [27] A. Edelman and N. R. Rao. Random matrix theory. Acta Numerica, 14:233–297, 2005. [28] Y. Chikuse. Statistics on Special Manifolds, volume 174. Springer Science & Business Media, 2003. [29] A. T. James. Normal multivariate analysis and the orthogonal group. The Annals of Mathematical Statistics, pages 40–75, 1954. 22 

10math.ST

MARKOFF SURFACES AND STRONG APPROXIMATION: 1 arXiv:1607.01530v1 [math.NT] 6 Jul 2016 JEAN BOURGAIN, ALEXANDER GAMBURD, AND PETER SARNAK 1. Introduction This is the first of three papers giving detailed proofs of the results announced in [BGS16]. The main result here is strong approximation for Markoff triples for prime moduli. In paper two this is extended to more general moduli and applied to seiving in Markoff numbers. In the third paper, we formulate a strong approximation conjecture for more general affine Markoff surfaces and develop the techniques to obtain similar results in these cases. We review briefly the notation and setup from [BGS16] where background and references can be found. The Markoff surface X is the affine surface in A3 given by (1) X : x21 + x22 + x23 − 3x1 x2 x3 = 0. The Markoff triples M are the positive integer solutions to (1). Let Γ be a group of affine integral morphisms of A3 generated by the permutations of the coordiantes and the Vieta involutions R1 , R2 , , R3 , where R3 (x1 , x2 , x3 ) = (x1 , x2 , 3x1 x2 − x3 ) and R2 , R3 are defined similarly. The orbit of (1, 1, 1) under Γ yields all of M [Mar79], [Mar80]. If ∆ is the group of integral morphisms generated by Γ and the involutions which replace two of the coordinates of x by their negatives, then X(Z) consists of two ∆ orbits, namely of (0, 0, 0) and (1, 1, 1). For p a prime number the action of Γ and ∆ on X descends to a permutation action on the finite set X(Z/pZ) of solutions to (1) in Z/pZ. Our interest is in the orbits of this action which we often refer to as the components. Conjecture 1 (Strong Approximation Conjecture). For any prime p, X(Z/pZ) consists of two Γ orbits, namely {(0, 0, 0)} and X ∗ (Z/pZ) = X(Z/pZ)\{(0, 0, 0)} Clearly the above conjecture implies strong approximation for M and X(Z) in the form that their reductions mod p M → X ∗ (p) 1 2 JEAN BOURGAIN, ALEXANDER GAMBURD, AND PETER SARNAK and X(Z) → X(Z/pZ) are onto. In particular, one of the consequences is that the only prime congruence obstruction for a Markoff number m (that is a number which is a coordinate of an x ∈ M) is the one noted in [Fro13]: m 6= 23 , 0 mod p if p = 3 mod 4, p 6= 3. Our first result asserts that there is a very large orbit. Theorem 1. Fix ε > 0. Then for p large there is a Γ orbit C(p) in X ∗ (p) for which |X ∗ (p)\C(p)| ≤ pε (note that |X ∗ (p)| ∼ p2 ), and any Γ orbit D(p) satisfies 1 |D(p)| ≫ (log p) 3 . The proof of Theorem (1) establishes the strong approximation conjecture unless p − 1 is a very smooth number. In particular, the set of primes for which the strong approximation conjecture fails is very small. 2 Theorem 2. Let E be the set of primes for which the strong approximation conjecture fails. For ε > 0, the number of primes p ≤ T with p ∈ E is at most T ε , for T large. We end the introduction by outlining the rest of the paper. In Section 2 we define the fundamental rotations in Γ that are associated to an x ∈ X ∗ (p) and one of its coordinates. These act on the conic sections gotten by intersecting X ∗ (p) with the plane corresponding to the particular coordinate. Some basic properties of the incidence graph of the intersections of the conic sections are established. In Section 3, which we call the endgame, we define the cage C(p) which is shown to be a large component of X ∗ (p). Specifically any x ∈ X ∗ (p) for which the rotation associated to one of its coordinates (see Section 2 for definitions) has order at least 1 p 2 +δ0 (δ0 > 0) is shown to be in C(p). In Section 4 , the middle game, the last statement is extended to x’s for which the corresponding rotation has order pε0 (ε0 > 0 any fixed small number). The methods used in Sections 3 and 4 rely on nontrivial upper bounds for the number of points lying on curves over finite fields. In Section 3 Weil’s Riemann Hypothesis [Wei41] is a key tool, but this is not strong enough when the order is less 1 than p 2 +δ0 . In its stead we use Stepanov’s auxiliary polynomial method [Ste69] in the Appendix, or the recent gcd(u − 1, v − 1) bounds of Corvaja-Zannier [CS13], or the MARKOFF SURFACES AND STRONG APPROXIMATION: 1 3 combinatorial method based on the projective Szemeredi-Trotter Theorem [Bou12] developed in section 4. Section 5, the opening, deals with x’s for which the orders of the associated rotations are very small (for example, being uniformly bounded). This is done by lifting the equations to characteristic zero which leads to an equation in (Q̄)3 in roots of unity. In general (that is, in the setting of Paper III) we invoke Lang’s Gm conjecture at this point (proven in [Lau83] for example), however for the special case at hand one can show directly that X∗ (Q̄) has no finite Γ orbits (the last is a necessary condition for strong approximation, in the form of Conjecture 1, to hold). This Q̄ analysis leads to to part 2 of Theorem 1. In Section 6 we assemble the various stages of our argument, explicated in Sections 3, 4, 5 and give a proof of Theorem 1 in a form from which strong approximation follows if p2 − 1 is not very smooth: see (70). To prove Theorem 2 we combine the above with a variant of the results in [CKSZ14] concerning the multiplicative orders of coordinates of points of curves on A2 over Fp . 2. Preliminaries 2.1. Analysis of the conic sections. Theorem 1 in the weaker form that |C(p)| ∼ |X ∗ (p)| as p → ∞, can be viewed as the finite field analogue of [Gol03], where it is shown that the action of Γ on the compact real components of the character variety of the mapping class group of the once punctured torus is ergodic. As in [Gol03], our proof makes use of the rotations τij ◦ Ri , i 6= j where τij permutes xi and xj . For example, τ2,3 ◦ R2 (x1 , x2 , x3 ) = (x1 , x3 , 3x1 x3 − x2 ), so the action on (x2 , x3 ) for fixed x1 is given by the rotation rot(3x1 ) ! ! ! ! x2 x3 0 1 x2 (2) rot(3x1 ) = = . x3 3x1 x3 − x2 −1 3x1 x3 This rotation preserves the conic section obtained by intersecting X ∗ (p) with the plane defined by the first coordinate being equal to the value of x1 ; in general, we define the conic section Cj (a) as follows: (3) Cj (a) = {xj = a} ∩ X ∗ (p). We give an explicit description of this action, depending  2on whether x = 3x1 is parabolic (x2 − 4 = 0, that is x = ±2), hyperbolic ( x p−4 = 1) or elliptic  2 
( x p−4 = −1) with ·· being the Legendre symbol. 4 JEAN BOURGAIN, ALEXANDER GAMBURD, AND PETER SARNAK   = −1, that is if p ≡ −1 mod 4, then C1 (x) is Lemma 3. Let x = ±2. If −1 p   empty. If −1 = 1, that is if p ≡ 1 mod 4, then p (4) C1 (2) = (2, t, t ± 2i), where i2 ≡ −1 mod p; (5) C1 (−2) = (−2, t, −t ± 2i), which are pairs of disjoint lines. The action of ρ1 = rot(x) is (6) ρ1 (2, t, t ± 2i) = (2, t ± 2i, t ± 4i), (7) ρ1 (−2, t, −t ± 2i) = (−2, −t ± 2i, −t ∓ 4i), so rot(2) preserves each line and rot(−2)interchanges them. Now when x 6= ±2 we write x = χ + χ−1 ,  2   2  where χ ∈ Fp if x p−4 = 1 and χ ∈ Fp2 if x p−4 = −1. Note that ! ! ! ! ! !−1 1 −1 1 −1 1 1 χ 0 χ 0 1 1 1 1 χ −χ = rot(x) == χ χ χ1 0 χ1 0 χ1 χ χ1 χ χ1 −χ 1 and consequently ! 1 1 χℓ rot(x)ℓ = χ χ1 0 and hrot(x)i = 0 1 χℓ 1 χ ! 1 1 χ χ1 −1 −χ !−1 ( = χ1 χ − χ1 − Consequently C1 (x) contains all elements  1 1 χ χ1 χ χ1 1 χ1 ! χℓ 0 1 − χ1 χ1 1 − χχ1 χ1 0 1 χℓ ! ! ! −1 −1  1 −χ χ −χ 1 1 χ ) ; χ1 ∈ hχi .      1 −1  x2  1 1 −1  x3  1  χ− x3 − χ1 + (χx2 − x3 ) , χ− (χx3 − x2 )χ1 + x2 − χ χ χ1 χ χ χ1 with χ1 ∈ hχi. Note that where o n b Projx2 (Cx ) ⊃ aχ1 + ; χ1 ∈ hχi χ1   x2  1 −1 1 −1  x3 − and b = χ − a= χ− (χx2 − x3 ) χ χ χ MARKOFF SURFACES AND STRONG APPROXIMATION: 1 satisfy σ = ab =  5 1 x 2  χ + χ 2 = 6 1. = χ − χ1 χ − χ1 Denoting by ρ a primitive root of Fp , a hyperbolic element x can be written in the form x = ρj + ρ−j . (8) For a hyperbolic element we let ord(x) = p−1 . j An elliptic element x can be written in the form x = ξ j + ξ −j , (9) where ξ is an element in Fp2 , ξ = (ρ̃)p+1 , where ρ̃ is a generator or the multiplicative group of Fp2 . For an elliptic element we let ord(x) = p+1 . j Lemma 4. Let x be hyperbolic; write x = w + w −1, (10) where w = ρj ∈ Fp . Then C1 (x) is a hyperbola with p − 1 points. Set (11) κ(x) = Let (12)  x2 . x2 − 4 κ(x) H(x) = { t, t  | t ∈ Fp ∗ }. Then we have the following map from H(x) to C1 (x):     κ(x) κ(x) κ(x) → x, t + . , tw + (13) t, t t tw In these coordinates (14)     κ(x) κ(x) = tw, . ρ1 t, t tw Lemma 5. Let x be elliptic; write (15) x = v + vp, where v ∈ Fp2 − Fp , v p+1 = 1. Then C1 (x) is an ellipse with p + 1 points. Set (16) Let (17) κ(x) = x2 . x2 − 4 E(x) = {(t, tp ) | t ∈ Fp2 , tp+1 = κ(x)}. 6 JEAN BOURGAIN, ALEXANDER GAMBURD, AND PETER SARNAK Then we have the following map from E(x) to C1 (x):   κ(x) κ(x) p . , tv + (18) (t, t ) → x, t + t tv In these coordinates (19) ρ1  κ(x) t, t    κ(x) = tv, . tv 2.2. Incidence graph for the conic sections. We treat the case of p ≡ 3(mod4) (the case of p ≡ 1(mod4) is simpler because of the special point in Lemma 3). Let X ∗ (p) be the Markoff triples mod p; ξ any coordinate of a triple, ξ 6= 0, ± 23 . For j 6= k and ξ, η (20) |Cj (ξ) ∩ Ck (η)| = 0, 1, 2. To determine which it is, the intersection consists of all z’s such that ξ 2 + η 2 + z 2 = 3ξηz, (21) which has a solution if is a square in Fp . In terms of Legendre’s symbol (22) 9ξ 2 η 2 − 4(ξ 2 + η 2 ) |Cj (ξ) ∩ Ck (η)| = 1 +  9ξ 2 η 2 − 4(ξ 2 + η 2 ) p  . So each Cj (ξ) meets p−1 Ck (η)’s. Define the incidence graph I(p) of X ∗ (p) to have 2 vertices Cj (ξ)’s with the number of edges between Cj (ξ) and Ck (η) being |Cj (ξ) ∩ Ck (η)|. Proposition 6. For p large (p > 10) the incidence graph I(p) is connected and in fact diam(I(p)) = 2. Proof. Fix ξ1 , ξ2 and i, j, say i, j ∈ {1, 2}. We seek y ∈ Fp such that C3 (y)∩Ci (ξ1 ) 6= φ and C3 (y) ∩ Cj (ξ2 ) 6= φ. This leads to solving the pair of equations: (23) ( (9ξ12 − 4)y 2 − λ2 = 4ξ12 (9ξ22 − 4)y 2 − µ2 = 4ξ22 for y, λ, µ ∈ Fp . In ξ12 = ξ22 then (23) reduces to the first equation (take λ = µ) and since 9ξ12 − 4 6= 0 and ξ12 6= 0, it defines a conic section. Thus for p large it has a solution and provides us with our y. If ξ12 6= ξ22 then (23) defines an irreducible curve in A3 . Thus again for p large it has solutions over Fp providing us with our desired MARKOFF SURFACES AND STRONG APPROXIMATION: 1 7 y. It follows that the distance in I(p) between any two points is at most 2. On the other hand, Ci (ξ) and Ci (η) are not joined if ξ 6= η nor is Ci (ξ) joined to half of the Cj (η)’s for j 6= i. Hence diam(I(p)) = 2.  3. Endgame 3.1. Use of Weil’s bound. We begin with the following Proposition 7. If x = (x1 , x2 , x3 ) is in X ∗ (p) and for some j ∈ {1, 2, 3} the order of 1 the induced rotation rot(xj ) is at least p 2 +δ (δ > 0 fixed) then x is joined to a point y in X ∗ (p) one of whose induced rotations is of maximal order. Proof. Consider first the case that x1 (say j = 1) is hyperbolic. In light of the discussion in section 2, x = (x1 , x2 , x3 ) is connected to the points in X ∗ (p) of the form (24) (x1 , α1 t + α2 t−1 , α3 t + α4 t−1 ) with t ∈ H, a cyclic subgroup of Fp ∗ . Here |H| | (p − 1); we set eH = p−1 . Our aim |H| ∗ is to produce t’s in H for which there is a primitive root y ∈ Fp satisfying (25) α1 t + α2 t−1 = y + y −1. Let P (H)(= Pα1 ,α2 (H)) denote the number of such solutions. A subgroup K of Fp ∗ is determined by its order |K| which divides p−1; let dK = Let fH (K) = fH (dK ) be the number of solutions to (26) p−1 . |K| α1 t + α2 t−1 = y + y −1 , t ∈ H , y ∈ K (note that the traces of the matrices that we produce, namely the common values of the left- and right-hand side of (26), are hit with multiplicity 2 in both t and y in our counting). Clearly (27) fH (K) ≤ 2 min(|K|, |H|). 1 We can estimate fH (K), at least if |H| ≥ p 2 +δ , using Weil’s Riemann Hypothesis for curves over finite fields. The map ξ → ξ dK , η → η eH sends solutions of (28) Cα1 ,α2 : α1 η eH + α2 η −eH = ξ dK + ξ −dK 8 JEAN BOURGAIN, ALEXANDER GAMBURD, AND PETER SARNAK to solutions of (26) and it is eH dK to 1. Hence if N(Cα1 ,α2 ) is the number of solutions to (28) then (29) fH (K) = N(Cα1 ,α2 ) . eH dK As we prove below (see Lemma 8) the curve Cα1 ,α2 i is absolutely irreducible. Since its genus is O(eH dK ), applying Weil bound yields √ (30) N(Cα1 ,α2 ) = p + O( peH dK ). Hence (31) fH (K) = p √ + O( p). eH dK By inclusion/exclusion (32) X P (H) = µ(d)fH (Kd ), d|(p−1) where µ is the Mobius function. Hence (33)   X µ(d) X 1 1 √ ϕ(p − 1) |H| + O( p) = |H| +Oε (p 2 +ε ) = |H| +Oε (p 2 +ε ). P (H) = µ(d) d d p−1 d|(p−1) d|(p−1) Here ϕ is the Euler function and it satisfies ϕ(n) ≫ε n1−ε and hence from (33) 1 we deduce that P (H) > 1 under the assumption that |H| ≥ p 2 +δ . This proves Proposition 7 in the hyperbolic case. Now consider the case of x elliptic. Let D be a non-square element in Fp . Then √ Fp2 = Fp [ D] and we can parametrize the subgroup H1 as follows √ (34) {(ξ + Dη)d1 ; ξ, η ∈ Fp ; ξ 2 − Dη 2 = 1}, where d1 = p+1 . |H1 | The conic section C1 (x) is an ellipse which can be parametrized as α2 − Dβ 2 = κ(x), (35) where κ(x) = root in Fp ∗ . Now (36) x2 . x2 −4 We seek α which can be written as α = u + u−1 with u a primitive (ξ + √ √ D)n = gn (ξ) + hn (ξ) D, where gn , hn are integral polynomials (37) gn (ξ) = ⌊n/2⌋  X i=0  n D i ξ n−2i , 2i MARKOFF SURFACES AND STRONG APPROXIMATION: 1 (38) hn (ξ) = ⌊n/2⌋  X i=0 9  n D i ξ n−2i+1 . 2i + 1 Let (39) gn (ξ, η) = ⌊n/2⌋  X i=0 (40) hn (ξ, η) =  n D i ξ n−2i η 2i , 2i  n D i ξ n−2i+1 η 2i−1 . 2i + 1 ⌊n/2⌋  X i=0 Then we have (41) (ξ + √ √ Dη)n = gn (ξ, η) + hn (ξ, η) D. Now we seek to bound f (H1 , K) with K subgroup of Fp ∗ with d2 = p−1 . |K| As in the hyperbolic case this is given by M d(d11d,d2 2 ) where M(d1 , d2) now counts the number of points on the following curve in Fp 3 : ( (42) ξ 2 − Dη 2 = 1 gd1 (ξ, η) = µd2 + µ−d2 This is a curve of genus O(d1 d2 ) and we can apply Weil bound and inclusionexclusion as in the hyperbolic case to produce the primitive u. This completes the proof of Proposition 7.  Lemma 8. Suppose α1 α2 6= 1 mod p . Then the curve α1 y e + α2 y −e = xd + x−d is absolutely irreducible. Proof. Consider P (x, y) = α1 xd y 2e + α2 xd − x2d − y e ∈ Fp [X, Y ]. For d = 1 P is clearly irreducible. Let d > 1 and assume that P is not irreducible P and f (x, y) = ajk xj y k ∈ Fp [X, Y ] an irreducible factor. Assume d ≥ e and u a d-th root of unity. Since for 0 ≤ s ≤ d, P (x, y) = P (us x, y), also X fs (x, y) = f (us x, y) = ajk usj xj y k 10 JEAN BOURGAIN, ALEXANDER GAMBURD, AND PETER SARNAK is an irreducible component of P . Thus either fs and fs′ are coprime or fs ∼ fs′ . Since a0k 6= 0 for some k (otherwise x would be a factor of f (x, y)), it follows that fs = fs′ if fs ∼ fs′ . Case 1 The fs are not pairwise coprime. The f (x, y) = fs (x, y) for some 0 < s < d, implying that usj = 1, i.e. sj ≡ 0( mod d) if ajk 6= 0. It follows that d has a divisor d1 > 1 such that d1 |j if ajk 6= 0 and hence f (x, y) has form f (x, y) = g(xd1 , y). The polynomial g(x, y) is therefore a factor of Q(x, y) = α1 xd2 y 2e + α2 xd2 − x2d2 y e − y e with d2 = dd1 and we lowered d. Case 2 The fs (0 ≤ s ≤ d) are mutualy coprime. Define P1 (x, y) = d−1 Y fs (x, y), s=0 which divides P . Degree considerations show that 2d ≥ d degx f, 2e ≥ d degy f, 2d + e ≥ d deg f. Case 2.1 degx f > 1, degy f > 1. It follows that degx f = 2, degy f = 2, e = d, deg f = 3, and P (x, y) = P1 (x, y). With u as above, ϕ(x, y) = f (x, uy) is an irreducible factor of P (x, y). Therefore for some 0 ≤ s ≤ d X X ajk uk xj y k ∼ fs (x, y) = ajk usj xj y k . ϕ(x, y) = j,k≤2 Consequently, there is some 0 ≤ l ≤ d such that k − sj ≡ l( mod d) if ajk 6= 0. Since Y fs (x, y), α1 xd y 2d + α2 xd − x2d y d − y d = 0≤s≤d clearly a0,1 = 6 0, a1,0 6= 0, and therefore 1 ≡ l ≡ −s( mod d), i.e. k + j = 1( mod d) if ajk 6= 0. Since deg f = 3, 2 ≡ 0( mod d), hence d = 2 and a1,1 = a2,0 = a0,2 = 0. Thus α1 x2 y 4 + ga1 x2 − x4 y 2 − y 2 ∼ (a21 x2 y + a12 xy 2 + a10 x + a01 y)(a21 x2 y − a12 xy 2 − a10 x + a01 y) ∼ y 2(a21 x2 + a01 )2 − x2 (a12 y 2 + a10 )2 . Setting a0,1 = 1 gives −y 2 (a21 x2 +1)2 +x2 (a12 y 2 +a10 )2 and a221 = 1, a21 −a12 a10 = 0, a212 = α1 , a210 = α2 . But this contradicts the assumption α1 α2 6= 1. Case 2.2 degx f = 1 or degy f = 1. Assume degy f = 1, say. Then there are coprime a(x), b(x) ∈ Fp [X] such that P (x, a(x) ) = 0, that is b(x) α1 xd a(x)2e + α2 xd b(x)2e − x2d a(x)e b(x)e − a(x)e b(x)e = 0. MARKOFF SURFACES AND STRONG APPROXIMATION: 1 11 Since a(x), b(x) are coprime, it follows that a(x)e |xd , b(x)e |xd , hence a(x) or b(x) is constant. If, say, b(x) is constant, previous equation implies xd |a(x)e , hence a(x)e = γxd and α1 γ 2 x2d + α2 b2e − γb2 x2d − γb2 = 0. It follows that α1 γ = b2 , α2 b2 = γ, hence α1 α2 = 1, contradicting the assumptions that α1 α2 6= 1.  3.2. The Cage. A point x = (x1 , x2 , x3 ) ∈ X ∗ (p) is called maximal if ord(rot(xj )) is maximal for some j. Note that the condition that the order of rot(3xj ) be maximal depends only on xj and not on the other coordinates of x (since it depends on the order of λ, where λ + λ−1 = 3x1 in F∗p of F∗p2 ). We call ξ ∈ Fp maximal if it is of maximal order. By the cage we mean a set of maximal elements in X ∗ (p). We claim that the cage is connected, that is to say if x̂ and ŷ are in the cage then x̂ is connected to ŷ. Let ξ be the coordinate of maximal order of x̂ and η be the coordinate of maximal order of ŷ, so that x̂ is connected to all points in Cj (ξ) and similarly ŷ is connected to all the points in Ck (η). Now according to Proposition 6, which when extended with and inclusion/exclusion argument as in Proposition 7 gives a y of maximal order such that P ∈ Cj (ξ) ∩ Cl (y); Q ∈ Ck (η) ∩ Cl (y). Since y is maximal, P and Q are connected by Γ and P is Γ- connected to x̂ and Q to ŷ. Thus x̂ is Γ connected to ŷ. Denote by C(p) the connected component of X ∗ (p) (under the Γ action) that contains the cage, then C(p) is our large component. 4. Middle Game In the endgame (section 3) we connected any x ∈ X ∗ (p) of order (that is max(ord(rot(xj )))) 1 l ≥ p 2 +δ0 (δ0 > 0) to the cage in one step. In this section we allow any number of moves to do the connecting. In particular any x of order at least pε is shown to be in the giant component. As in section 3 the y’s which are joined to a given x whose order is l1 via the corresponding rotation and which have orders l2 (here l1 and l2 divide p − 1 or p + 1) correspond to solutions of an equation (with σ ∈ Fp , σ 6= 1):  1 σ  = h2 + , σ 6= 1 h1 + h1 h2 (43)  with h1 ∈ H1 , h2 ∈ H2 with H1 , H2 subgroups of Fp ∗ or Fp2 ∗ . 12 JEAN BOURGAIN, ALEXANDER GAMBURD, AND PETER SARNAK In (43) we have |H1 | = l1 , |H2 | = l2 . If we have an upper bound on the number of solutions to (43) so that on summing over all l2 ≤ l1 with l2 dividing p − 1 or p + 1, yields a quantitiy which is less than l1 , then there is at least one h1 for which the corresponding y will have order bigger than l1 . We then repeat this procedure replacing x by this y and so on, until the order of 1 the element is at least p 2 +δ0 . At that point we are in the endgame and can finish. The key therefore is a suitable upper bound to the number of solutions to (43). Our original treatment used Stepanov’s technique of auxiliary polynomials in his elementary proof of this Riemann Hypothesis for curves. This yields explicit and reasonably sharp estimates which are ample for our application. We carry this out in the Appendix partly to illustrate the flexibility of this method. Subsequently the recent powerful technique for estimating from above the g.c.d. of (u − 1, v − 1), of Corvaja and Zannier [CS13] yields sharper bounds. This is relevant for the purpose of giving efffective bounds on p after which Theorem 1 takes effect. The precise upper bound to (43) established by [CS13] is n |H1 |.|H2 | o 20 max (|H1|.|H2 |)1/3 , . p The third treatment, and the one which we develop in this section, while special to (43), is robust in that the upper bound requires little further structure and it is suitable for generalisation for more general moduli as demonstrated in Paper II. It is based on the following projective Szemeredi-Trotter theorem (Proposition 2 in [Bou12]). Theorem 9. Let Φ : Fp → Mat2 (Fp ) such that det Φ does not vanish identically and ImΦ ∩ PGL2 (Fp ) is not contained in a set of the form F∗p · gH for some g ∈ SL2 (Fp ) and H a proper subgroup of SL2 (Fp ). Then the following holds. Given ε > 0, r > 1, there is δ > 0 such that if A ⊂ P 1 (Fp ) and L ⊂ Fp satisfy (44) 1 ≪ |A| < p1−ε (45) log |A| < r log |L|. Then (46) where for g = |{(x, y, t) ∈ A × A × L; y = τΦ(t) (x)}| < |A|1−δ |L|, ! a b ax+b . , τg (x) = cx+d c d Using Theorem 9 we prove the following: MARKOFF SURFACES AND STRONG APPROXIMATION: 1 13 Proposition 10. Given δ > 0 there is τ < 1 and Cτ depending on δ such that if pδ < |H1 | < p1−δ then the number of solutions to (43) is at most Cτ |H1 |τ . Proof. For h ∈ H, a subgroup of F∗p or F∗p2 , denote by h̃ = h + h−1 . Similarly we e = {h̃ | h ∈ H}. denote by H Suppose that (43) has T solutions. Then  σ =u   h1 σ  = v, h1 t + h1 t h1 + (47) f2 , has at least T 2 solutions. where h1 , t ∈ H1 and u, v ∈ H Elimination of h1 in (47) yields (48)   1 2 1 uv + σ t − =0 u2 + v 2 − t + t t f2 × H f2 . which, by assumption, has at least T 2 solutions in (t, u, v) ∈ H1 × H Next, let u = f˜1 , v = f˜2 with f1 and f2 ∈ H2 and define the following elements f2 : x, y ∈ H 1 ^ , x = (f 1 f2 ) = f1 f2 + f1 f2 f1 f2 −1 y = (f^ + . 1 f2 ) = f2 f1 Thus uv = x + y, u2 + v 2 = xy + 4 and equation (48) gets transformed into (49) xy − t̃(x + y) + σ(t̃)2 + 4(1 − σ) = 0. Denoting (50) α = t̃ and β = σ(t̃)2 + 4(1 − σ) we obtain (51) y= with g= αx − β = τg (x) x−α α β 1 −α ! = g(t̃) and τg the Mobius transformation. e2 × H e2 × H e1. Equation (51) has at least T 2 solutions in (x, y, t̃) ∈ H 14 JEAN BOURGAIN, ALEXANDER GAMBURD, AND PETER SARNAK We apply Theorem 9 taking Φ(t) (52) Φ(t) = t −σt2 − 4(1 − σ) 1 −t ! e2, L = H e1. and A = H We verify the assumption on Φ. Since σ 6= 1, det Φ(t) = (1 − σ)(4 − t2 ) does not vanish identically. It remains to show that n det Φ(s) Φ(s)−1 Φ(t) ; s, t ∈ Fp } det Φ(t) is not contained in a proper subgroup H of SL2 (p). By (52) ! ! 2 2 −s σs + 4(1 − σ) t −σt − 4(1 − σ) Φ(s)−1 Φ(t) = −1 s 1 −t = −st + σs2 + 4(1 − σ) s−t (s − t)(4(1 − σ) − σst) −st + σt2 + 4(1 − σ)) ! Taking s = σt + 4(1 − σ) t gives  σ(1 + σ)t + (53) (1 − σ) − t t 1 4 4(1−σ) t −σ 2 t2 + 4(1 − σ)2 0 ! = (1 − σ) 4 t  − t gt . As the proper subgroups of SL(2, Fp ) have trivial second commutator [Suz82], it suffices to show that (54) )(gt2 gt1 gt−1 gt−1 )(gt4 gt3 gt−1 gt−1 ) gt−1 )(gt3 gt4 gt−1 gt−1 (gt1 gt2 gt−1 4 3 1 2 3 4 2 1 is not identically one for t1 , t2 , t3 , t4 ∈ F∗p . If this were the case, the same would be true for t1 , t2 , t3 , t4 taken in an extension field of Fp so as to make t2 = (55) 4(1 − σ)2 − ε (ε = ±1) σ2 solvable. Taking t = ±κ satisfying (55), we get 2 −ε (56) g±κ = ± σκ [(1 + σ) 4(1−σ) σ2 1 ! + 4(1 − σ)] ε . 0 MARKOFF SURFACES AND STRONG APPROXIMATION: 1 15 We choose ε = ±1 as to ensure that (1 + σ)(4(1 − σ)2 − ε) + 4σ 2 (1 − σ) 6= 0 and obtain matrices (57) g± = ±η ε 1 0 ! that clearly generate SL2 (p). Consequently, Theorem 9 is applicable, yielding the bound T 2 ≪ |H2|1−τ |H1 |.  5. Opening The analysis of the previous sections shows that we can connect x ∈ X ∗ (p) whose order is at least pε (or smaller if the divisors of p2 − 1 are not too numerous) to the cage. To deal with all x’s and in particular ones whose orders are uniformly bounded (indepndent of p) we lift to characteristic zero. In this connection we observe first that if the action of Γ on X ∗ (Q̄) has a finite orbit F then the strong approximation conjecture cannot hold. To see this consider more generally any finite orbit F of the Γ action on A3 (C). The coordinate of any ξ in such an F must lie in a cyclotomic field Ln = Q(ζn ), where ζn is a primitive n-th root of 1. For if ξ = (ξ1 , ξ2, ξ3 ) then ord(rot(ξj )) must be finite and hence 3ξj = tj + t−1 j (58) with tj a root of unity. If n is the least common multiple of all the orders of all the tj ’s corresponding to the ξ’s in F then ξj ∈ Ln and hence F ⊂ A3 (Ln ). In fact the ξj ’s (3) are all integral except possibly for denominators powers of 3, so that F ⊂ A3 (OLn ) (3) where OLn is the ring of S-integers in Ln , with S consisting of the primes dividing 3. If p is a rational prime (p 6= 3) which splits completely in Ln and P is a prime of (3) (3) Ln with P |(p) then OLn /P ∼ = Z/pZ). The Γ action of A3 (OLn ) factors through = Fp (∼ the reduction π mod P (3) A3 (OLn )  π y (3) Γ −−−→ Γ̄ (3) A3 (OLn )  π y (3) A3 (OL /P ) −−−→ A3 (OL /P ) and hence (59) F̄ = π(F ) ⊂ A3 (Fp ), is Γ̄ − invariant. 16 JEAN BOURGAIN, ALEXANDER GAMBURD, AND PETER SARNAK Since Γ preserves the level sets Xk : x21 + x22 + x23 − 3x1 x2 x3 = k, (60) (3) any such F is contained in Xk (OL ) for a suitable k. Thus for any such F , there is a positive density of p’s which split completely in Ln , and hence for which F̄ ⊂ Xk (Fp ) is a fixed size Γ̄ -orbit (|F̄ | ≤ |F |). That is, the finite Γ-orbits in A3 (Q̄) must be part of any description of the Γ̄-orbits on A3 (Fp ), for p large. In our setting of this paper, k = 0 and we have (we thank E. Bombieri for this simple proof) Proposition 11. X ∗ (Q̄) has no finite Γ-orbit. Proof. As in the discussion above, if F is such an orbit and ξ ∈ F then the ξj satisfy (58) with tj an lj -th root of one. The Markoff equation for t1 , t2 , t3 becomes (61) −1 −1 −1 −1 2 2 −1 2 (t1 + t−1 1 ) + (t2 + t2 ) + (t3 + t3 ) − (t1 + t1 )(t2 + t2 )(t3 + t3 ) = 0. Now (61) has no solutions with |tj | = 1 (let alone being roots of unity) except for tj = ±i, j = 1, 2, 3. To see this note that if −1 −1 ¯ a = t1 + t−1 1 (= t1 + t1 ), b = t2 + t2 , c = t3 + t3 then a, b, c, lie in [−2, 2] and by the inequality of the geometric and arithmetic means 2 |a|3 + |b|3 + |c|3 ≤ (a2 + b2 + c2 ). (62) 0 ≤ a + b + c = |abc| ≤ 3 3 Hence the only solutions to (61) correspond to a = b = c = 0 or tj = ±i. In terms of ξj this gives ξ = (0, 0, 0), which is the only invariant set for the action of Γ on X(Q̄).  We remark that in the context of the general surfaces that are studied in Paper III, for example the surfaces Xk in (60) with k 6= 0, there can be a continuum of solutions to the analogue of equation (61) with |tj | = 1. However the solutions with tj a root of unity (with unspecified order) are still restricted to a finite number of nondegenerate solutions. This follows from Lang’s Gm conjecture, see [Lau83] and [SA94] for proofs which give the solutions effectively. In various special cases these finite Q̄ orbits for the Γ-action correspond to the determination of the algebraic Painleve VI solutions ([DM00], [LT14]); we leave the details to paper III . Returning to the Markoff surface X, let ξ = (ξ1 , ξ2 , ξ3) ∈ X ∗ (p) with ord(rot(ξj )) = lj for j = 1, 2, 3. Let n = lcm(l1 , l2 , l3 ) and Ln = Q(ζn ) and let ζl1 , ζl2 , ζl3 be primitive roots of one respectively. Let 2 2 2 (63) η = (ζl1 +ζl−1 )2 +(ζl2 +ζl−1 )2 +(ζl3 +ζl−1 )2 −(ζl1 +ζl−1 )(ζl2 +ζl−1 )(ζl3 +ζl−1 ) ∈ OLn . 2 2 3 2 2 3 MARKOFF SURFACES AND STRONG APPROXIMATION: 1 17 According to Proposition 11 unless l1 = l2 = l3 = 2 (i.e. ζlj = ±i), η 6= 0. Now |η| ≤ 20 and hence (64) Norm(η) ≤ 20φ(n) ≤ 20n . If P is a prime in OLn and η ∈ P , then P |(η) and hence (65) N(P ) ≤ Norm(η) ≤ 20n . Put differently, if log20 N(P ) > n then (66) η 6= 0( mod P ). For our point ξ in X ∗ (p), 3ξj = λj + λ−1 j with λj in Fp or Fp2 and λj an lj -th root of 1, and (l1 , l2 , l3 ) 6= (2, 2, 2) since ξ 6= (0, 0, 0). If all the λj ’s are in Fp then Q(ζn ) splits completely at p, that is there is a prime P dividing (p) such that OLn /P ∼ = Fp , N(P ) = p and π(ζlj ) = λj in OLn /P and η ≡ 0( mod P ). Hence from (66) we conclude that (67) log20 p ≤ n. If the field generated by λj ’s (over Fp ) is Fp2 then there is a prime P of OLn dividing (p) such that OL /P ∼ = Fp2 , N(P ) = p2 n and π(ζlj ) = λj in OLn /P and η ≡ 0( mod P ). Hence again from (66) we conclude that (68) 2 log20 p ≤ n. Hence in either case n ≥ log20 p where n = lcm(l1 , l2 , l3 ), and hence (69) 1 max(l1 , l2 , l3 ) ≥ (log20 p) 3 . We have proven Proposition 12. Let ξ ∈ X ∗ (p) have maximal order l, i.e. max(l1 , l2 , l3 ) = l with 1 lj = ord(rot(ξj )), then l ≥ (log20 p) 3 . In particular, any component F of X ∗ (p) satisfies 1 |F | ≥ (log20 p) 3 . 18 JEAN BOURGAIN, ALEXANDER GAMBURD, AND PETER SARNAK 6. Proofs of Theorems 1 and 2 Proposition 12 establishes the second part of Theorem 1 and combined with the analysis in sections 3 and 4 yields a proof of the strong approximation conjecture if p2 − 1 is not very smooth. For example, the strong approximation conjecture is true for X ∗ (p) if the prime p satisfies (70) X 2 d 3 < y; for any y. 1 (log p) 3 ≤d≤y d|(p2 −1) We do this by using the arguments and results in [Cha13] and [CKSZ14] concerning points (x, y) on irreducible curves over Fp for which ord(x) + ord(y) is small (here ord(x) is the order of x in F∗p ). Theorem 13. Fix d ∈ Z+ and δ > 0. There is an ε > 0, ε = ε(d, δ), such that for all primes p ≤ z (z sufficiently large) with the exception of at most z δ of them, the following property holds. Let f (x, y) ∈ Fp [x, y] be of degree at most d and not divisible by any polynomial of the form ρxα y β − 1 or ρy β − xα for any ρ ∈ F¯p and integers α and β. Then all solutions (x, y) ∈ (F¯p × F¯p )∗ of f (x, y) = 0 satisfy (71) ord(x) + ord(y) ≥ pε except for at most 11d3 + d of them. Proof. Theorem 1.2 in [CKSZ14] establishes what we want except that pε in (71) is replaced by the stronger bound pα(d) , with α(d) = 2 , 89d2 + 3d + 14 while the exceptional set of primes is of zero density. For our purpose the exponent in (71) is allowed to be small and in exchange we want the exceptional set to be much smaller. To this end we follow verbatum the discussion in Section 4 of [CKSZ14]. For d fixed and T a large parameter they show that there is a U = U(d, T ) which has at 1 most O(T α(d) ) prime factors (their log T in the denominator is irrelevant for us) with the property: If p does not divide U and f as in Theorem 13 and f (x, y) = 0 in Fp , then (72) ord(x) + ord(y) ≥ T MARKOFF SURFACES AND STRONG APPROXIMATION: 1 19 except for at most 11d3 + d such (x, y) in Fp ∗ × Fp ∗ . For our given δ > 0 and large parameter z choose T to be T = z δα(d) . (73) Then the number of primes p with p|U is O(z δ ) , and if p does not divide U then Theorem 13 holds with (72) and (73), that is with ε = δα(d).  To prove Theorem 2 we apply Theorem 13 to the curves fσ (x, y) given by equation x+ 1 σ =y+ x y 1 with σ 6= 1. If (log20 p) 3 > 1000, then according to Proposition 12, for any ξ ∈ X ∗ (p) we have ord(rot(ξj0 )) is at least 1000 for j0 one of the j’s in {1, 2, 3}. Hence if p is not in the exceptional set in Theorem 13 with d = 4, then in the typical equation x + σx = y + y1 corresponding to the orders of the rotations in the rot(ξj0 ) orbit, there is (x, y) which is not one of the exceptional 11d3 + d < 1000 possible points. For such (x, y) the induced rotation has order at least pε and hence ξ is joined to the cage by the middle game. Our methods fall short of dealing with all p, specifically for those rare p’s for which p −1 is very smooth. The following hypothesis which is a strong variant of conjectures of M.C. Chang and B. Poonen [Cha13, Vol10] would suffice to deal with all large p’s. 2 Hypothesis: Given d ∈ N, there is δ > 0 and K = K(d) such that for p large and f (x, y) absolutely irreducible over Fp and of degree d and f (x, y) = 0 is not a translate of a subtorus of (F̄∗p )2 , the set of (x, y) ∈ (F∗p )2 for which f (x, y) = 0 and max(ordx, ordy) ≤ pδ , is at most K. Appendix A. Stepanov’s auxiliary polynomial method [Ste69] for bounding the number of solutons to equations like (43) is quite flexible. We demonstrate this for some special cases (the general case can be handled similarly). The proposition below is an extension of the approach and bounds in [HK00] (where S(x) = x, T (x) = 1 − x and t1 = t2 ). In what follows S(x) and T (x) are rational functions in Fp (x) of total degree d1 and d2 respectively and with disjoint divisors; e = d1 + d2 is fixed. Proposition 14. For p a large prime, t1 , t2 dividing p − 1, t1 ≥ t2 , let Y = {y ∈ Fp : S(y)t1 = T (y)t2 = 1}. 20 JEAN BOURGAIN, ALEXANDER GAMBURD, AND PETER SARNAK 1 Then if t1 ≪e p1− 2e , − 1 |Y | ≪e min{t2 , t1 t2 4e }. Remarks: (1) The trivial bound is O(t2) so the Proposition gives an improvement (power 4e saving) if t2 ≥ t14e−1 . (2) If h(ξ, η) = 0 is a plane curve of genus 0 over Fp , then the Proposition gives an upper bound on the number of solutions with ξ t1 = η t2 = 1 (cf. [CS13]). Applying Proposition 14 with t1 = t2 , S(y) = y, T (y) = ay+b cy+d yields 3 Corollary 15. For p large prime, t|(p − 1), t ≤ p 4 and Ut = {y ∈ Fp : y t = 1} the t-th roots of 1, 3 |σ(Ut ) ∩ Ut | ≪ t 4 for σ ∈ PGL2 (Fp ), σ 6= 1. 3 Corollary 16. For t|(p − 1), t ≤ p 4 , b ∈ Fp , b 6= 1, 3 |{w, ρ ∈ Fp : w + w −1 = ρ + bρ−1 , w t = ρt = 1}| ≪ t 4 . Proof. Put ρw = ξ, wρ = η, then ξ t = η t = 1 and each such solution with ξ = bη−1 η−1 corresponds to at most two solutions (w, ρ) above. Applying Corollary 15 yields Corollary 16.  Proof of Proposition 14: First we need a generalization of Proposition 3.2 in [VS12] where their common t is replaced by t0 , t1 , . . . , tn . The result is the following Lemma, whose proof is the same Lemma 17. Let t0 , t1 , . . . , tn , as well as B and J be integers, p a large prime, and α1 , . . . , αn distinct elements in Fp ∗ . Assume that 1 min(t0 , . . . , tn ) ≥ (n − 1)B 2n + JB 2 and that p ≥ (2nB + 2) max(t0 , t1 , . . . , tn ). Then xai xt0 b0,i (x − α1 )t1 b1,i . . . (x − αn )tn bn,i with aj ≤ J and b0,i , . . . bn,i ≤ B are linearly independent in Fp [x]. MARKOFF SURFACES AND STRONG APPROXIMATION: 1 21 Let α1 , . . . , αk ∈ Fp be distinct and ν1 , ν2 , . . . νk ∈ Z; set Rν (x) = (x − α1 )ν1 . . . (x − αk )νk . For m ≥ 1, X m dj1 dj k dm ν1 [R (x)] = [(x − α ) ] . . . [(x − α1 )νk ] = ν 1 j1 jk dxm dx dx j j1 +...jk =m X Bm,j (x − α1 )ν1 −j1 . . . (x − αk )νk −jk . j1 +j2 +...jk =m Hence dm (74) [(x − α1 )(x − α2 ) . . . (x − αk ] Rν (x) = Rν (x)Pm,ν (x), dxm where Pm,ν is a polynomial of degree at most km. Stepanov’s polynomial method is based on constructing a polynomial which vanishes to high order on Y . Let λa,b1 ,b2 be in Fp with 0 ≤ a ≤ J and 0 ≤ bj ≤ B and form X (75) φ(x) = λa,b1 ,b2 xa (S(x))t1 b1 (T (x))t2 b2 . m Write S(x), T (x) in the form (we assume that both factor into linear factors Fp [x]): A(x − α1 ) . . . (x − αt ) , (x − β1 ) . . . (x − βτ ) B(x − γ1 ) . . . (x − γµ ) T (x) = . (x − δ1 ) . . . (x − δν ) S(x) = (76) For simplicity we assume that S(x) and T (x) are square-free and we are assuming that the α, β, γ, δ’s are all distinct. The constants A and B can be absorbed into the λ’s, so without loss of generality we can take A = B = 1. For m ≥ 0 (77) [(x − α1 ) . . . (x − αt )(x − β1 ) . . . (x − βτ ) . . . (x − δν )]m  dm  a t1 b1 t2 b2 x (S(x)) (T (x)) = dxm xa S(x)t1 b1 T (x)t2 b2 Pm,a,b1 ,b2 (x) with Pm of degree at most em. Hence for x = y ∈ Y and m ≤ M (78) X X  dm dm  a t1 b1 t2 b2 λa,b1 ,b2 y a Pm,a,b1 ,b2 (y), φ(x)| = λ x S(x) T (x) = x=y a,b1 ,b2 x=y dxm dxm a,b ,b 1 2 by the definition of Y . We can make (78) equal to 0 for all y in Y by noting that y a Pm (y) is a polynomial of degree J + em. So (78) can be made 0 with not all of the 22 JEAN BOURGAIN, ALEXANDER GAMBURD, AND PETER SARNAK λa,b1 ,b2 ’s equal to 0 and for all m ≤ M as long as (J + eM)M < B 2 J. (79) Assuming that this is satisfied, we have φ(x) which is not identically zero and has degree (as a rational function) at most J + eBt1 . Hence if φ(x) is not identically zero, then its order of vanishing on Y is at least M and hence M|Y | ≤ J + eBt1 or (80) |Y | ≤ J + eBt1 . M We now check that under suitable constraints on the sizes of parameters, φ(x) does not vanish identically. We have φ(x) = consequently X λa,b1 ,b2 xa (x − α1 )t1 b1 . . . (x − αt )t1 b1 (x − γ1 )t2 b2 . . . (x − γµ )t2 b2 , (x − β1 )t1 b1 . . . (x − βτ )t1 b1 (x − δ1 )t2 b2 . . . (x − δν )t2 b2 (x − β1 )B . . . (x − βτ )B (x − δ1 )B . . . (x − δν )B φ(x) = X λa,b1 ,b2 xa (x − α1 )t1 b1 . . . (x − αt )t1 b1 (x − β1 )(B−b1 )t1 . . . (x − βτ )(B−b1 )t1 · · (x − γ1 )t2 b2 . . . (x − γµ )t2 b2 (x − δ1 )(B−b2 )t2 . . . (x − δν )(B−b2 )t2 . Now the monomials appearing in the last expression are linearly independent over Fp [x] according to Lemma 17 as long as (81) p ≥ (2eB + 2)t1 1 t2 ≥ eB 2e + JB e . 2 Thus, if (81) and (79) hold, so does (80). Choose J ≤ M and M 2 = Ce B 2 J. Then M= p √ Ce B J 1 1 and M ≥ J iff J ≪e B 2 . Now choose B + t12e and if t1 ≪ p1− 2e then (81) and (79) hold and − 1 |Y | ≪e t1 t2 4e . This completes the proof of Proposition 14. MARKOFF SURFACES AND STRONG APPROXIMATION: 1 23 Acknowledgements: While working on this paper, the authors were supported, in part, by the following NSF DMS awards: 1301619 (Bourgain), 1603715 (Gamburd), 1302952 (Sarnak). References [BS02] F. Beukers and C. J. Smyth, Cyclotomic points on curves, Number theory for the [Bom07] millenium (Urbana, Illinois, 2000), I, A.K. Peters, (2002), 67-85. E. Bombieri, Continued fractions and the Markoff tree, Expo. Math. 25 (2007), no [Bou12] 3, 187–213. J. Bourgain, A modular Szemeredi-Trotter theorem for hyperbolas, C.R. Acad. Sci. Paris Ser 1, 350 (2012), 793–796. [BG08] J. Bourgain and A. Gamburd, Uniform expansion bounds for Cayley graphs of SL2 (Fp ), Ann. Math. 167 (2008), 625–642. [BGS10] J. Bourgain, A. Gamburd and P. Sarnak, Affine linear sieve, expanders and sum product, Invent. Math. 179 (2010), 559–644. [BGS16] J. Bourgain, A. Gamburd, P. Sarnak, Markoff triples and strong approximation, C. R. Math. Acad. Sci. Paris 354 (2016), no. 2, 131-135. [Cas78] [Cha13] J. W. S. Cassels, Rational Quadratic Forms, Academic Press, 1978. Chang, Mei-Chu Elements of large order in prime finite fields. Bull. Aust. Math. [CKSZ14] Soc. 88 (2013) 169-176. M-C. Chang, B. Kerr, I. Shparlinski and U. Zannier, Elements of large orders on [CS13] varieties over prime finite fields, J. Theor. Nombres Bordeaux 26 (2014) 579-594. P. Corvaja, U. Zannier, Greatest common divisors of u − 1, v − 1 in positive charac- teristic and rational points on curves over finite fields, J. Eur. Math. Soc. (JEMS) 15 (2013) 1927-1942. [DKS13] C. D’Andrea, T. Krick and M. Sombra, Heights of varieties in multiprojective spaces [DM00] and arithmetic Nullstellensatz, Annales Sci. de l’ENS, 46, (2013),549-627. B. Dubrovin, M. Mazzocco, Monodromy of certain Painleve-VI transcendents and [Fro13] reflection groups, Invent. Math. 141 (2000) 55-147. G. Frobenius, Über die Markoffschen Zahlen, Akad. Wiss. Berlin, (1913), 458–487. [Gol03] W. Goldman, The modular group action on real SL(2)-characters of a one-holed torus, Geom. and Top. Vol. 7 (2003), 443–486. [HK00] R. Heath-Brown and S. Konyagin, New bounds for Gauss sums derived from k-th powers and for Heilbronn’s exponential sum, QUAR. J MATH (2000), 221–235. [Lau83] [Mar79] M. Laurent, Exponential diophantine equations, CR Acad Sc. 296 (1983), 945–947. A. Markoff, Sur les formes quadratiques binaires indéfinies, Math. Ann. 15 (1879) [Mar80] 381–409. A. Markoff, Sur les formes quadratiques binaires indéfinies, Math. Ann. 17 (1880) [LT14] 379–399. O. Lisovyy and Y. Tykhyy, Algebraic solutions of the sixth Painleve equation, Journal of Geometry and Physics, 85 (2014) 124-163. 24 JEAN BOURGAIN, ALEXANDER GAMBURD, AND PETER SARNAK [MW13] D. McCullough and M. Wanderley, Nielsen equivalence of generating pairs in [SA94] SL(2, q), Glasgow Math. J. 55 (2013), 481–509. P. Sarnak and S. Adams, Betti numbers of congruence groups, with an appendix [Sch04] by Z. Rudnick, Israel J. Math, 88, 1994, 31-72. Schmidt, W. M Equations over finite fields: an Elementary approach, Kendrick [Ste69] Press 2004. S.A. Stepanov, The number of points of a hyperelliptic curve over a prime field, [Suz82] MATH USSR-IZV 3:5 (1969), 1103–1114. M. Suzuki, Group Theory I, Springer-Verlag, Berlin-Heidelberg-New York, 1982. [Vol10] Voloch, Elements of high order on finite fields from elliptic curves, Bull. Aust. Math. Soc. 81 (2010), 425-429. [VS12] [Wei41] Vyugin, I. V., Shkredov, I. D., Sb. Math. 203(2012), 844-863. A. Weil, On the Riemann Hypothesis in function fields, Proc. Nat. Acad. Sci. USA 27 (1941), 345–347. IAS E-mail address: bourgain@math.ias.edu The Graduate Center, CUNY E-mail address: agamburd@gc.cuny.edu IAS and Princeton University E-mail address: sarnak@math.princeton.edu 

4math.GR

arXiv:1706.02858v2 [math.ST] 14 Mar 2018 On rumour propagation among sceptics Farkhondeh A. Sajadi and Rahul Roy1 University of Isfahan, Isfahan and Indian Statistical Institute, Delhi Keywords: complete coverage, eventual coverage, rumour process. AMS Classification: 60K35 Abstract: Junior, Machado and Zuluaga (2011) introduced a model to understand the spread of a rumour. This model consisted of individuals situated at i.i.d. points of the line N. An individual at the origin 0 starts a rumour and passes it to all individuals in the interval [0, ρ0 ], where ρ0 is a positive random variable. An individual located in this interval receives the rumour and transmits it further. The rumour process is said to survive if all the individuals in the model receive the rumour. We study this model, when the individuals are more sceptical and they transmit only if they receive the rumour from at least two different sources. In stochastic geometry the equivalent of this rumour process is the study of coverage of the space Nd by random sets. Our study here extends the study of coverage of space and considers the case when each vertex of Nd is covered by at least two distinct random sets. 1 Introduction Let {Xi : 1 ≥ 1} be a collection of i.i.d. Bernoulli (p) random variables, i.e., ( 1 with probability p Xi = 0 with probability 1 − p. (1) Also let {ρi : 1 ≥ 1} be a collection of i.i.d. N valued random variables, independent of the collection {Xi : 1 ≥ 1}. Let ρ denote a generic random variable with the same distribution as ρi . In addition, let ρ0 an independent N valued random variables, independent of the collections {Xi : 1 ≥ 1} and {ρi : 1 ≥ 1}, with ρ0 having the same distribution as ρ. Taking X0 ≡ 1, consider the region [ ([i, i + ρi ]). C := {i≥0:Xi =1} 1 f.sajadi@sci.ui.ac.ir, rahul@isid.ac.in 1 Junior, Machado and Zuluaga (2011) takes {i : Xi = 1} as the location of individuals on N and a rumour is started at the origin which is received by individuals located in [0, ρ0 ]. An individual at i on receiving the rumour spreads it among all individuals in the region [i, i + ρi ], and thus the rumour propagates. This model they call the firework process. They also introduce another model of rumour propagation, called the reverse firework process, where they study the region [ ([i − ρi , i]). C̃ := {i≥0:Xi =1} Their object of interest is under what conditions on the processes {Xi : 1 ≥ 1} and {ρi : 1 ≥ 1} are the regions C and C̃ unbounded connected regions with positive probability, i.e. the rumour process ‘percolates’. Remark 1.1. A simple argument using Kolmogorov’s 0-1 law yields that C (or C̃) being an unbounded connected region with positive probability is equivalent to C (or C̃) containing a region [t, ∞), for some t ≥ 1 with probability 1. We study the spread of rumour among sceptics. A sceptic individual at i spreads it among all individuals in the region [i, i + ρi ] only if s/he received the rumour from at least two distinct sources, i.e., (i) for the firework process, there are two individuals at j and k (say) with j 6= k and −1 ≤ j, k < i such that i ∈ [j, j + ρj ] ∩ [k, k + ρk ]; here we assume that there is an individual at location −1, who spreads the rumour in the region [−1, −1 + ρ−1 ], where ρ−1 is a random variable independent of all other random variables and having the same distribution as ρ, (ii) for the reverse firework process, there are two individuals at j and k (say) with j 6= k and −1 ≤ j, k < i such that j, k ∈ [i − ρi , i]. Towards this we define the regions D := {x ∈ R : there exist j, k ≥ −1 with j 6= k and Xj = Xk = 1 such that x ∈ ([j, j + ρj ] ∩ [k, k + ρk ])}, D̃ := {x ∈ R : x ∈ [i − ρi , i] for some i with Xi = 1 and there exist − i ≤ j, k < i with j 6= k and Xj = Xk = 1 such that j, k ∈ [i − ρi , i]}. We look for conditions on on the processes {Xi : 1 ≥ 1} and {ρi : 1 ≥ 1} such that the regions D and D̃ are unbounded connected regions with positive probability. Remark 1.2. As in the case of rumour propagation in C and C̃, a simple argument using Kolmogorov’s 0-1 law yields that D (or D̃) being an unbounded connected region with positive probability is equivalent to D (or D̃) containing a region [t, ∞), for some t ≥ 1 with probability 1. Thus we define 2 Definition 1.3. The firework process (respectively, reverse firework process,) percolates among sceptics if, with probability 1, D (respectively, D̃,) contains a region [t, ∞), for some t ≥ 1. Remark 1.4. Using this equivalent definition obtained by the tail event properties allows us to study the model without considering the influence of the two initiators X−1 and X0 . Proposition 1.5. Let l := lim inf jP(ρ ≥ j) > 1 and L := lim sup jP(ρ ≥ j) < ∞. j→∞ j→∞ We have that both the firework process and the reverse firework process percolate if p > 1/l and neither of them percolates if p < 1/L. We note here that this is the same condition that Junior, Machado and Zuluaga (2011) obtain for the percolation among ‘non-sceptical’ individuals. Indeed, the above proposition goes through among more radical sceptics too, i.e. if individuals need to receive the rumour from k ≥ 1 distinct sources before they transmit the rumour. As noted in the review article Junior, Machado and Ravishankar (2016), the above model is related to the study of coverage processes in stochastic geometry. Athreya, Roy and Sarkar (2004) introduce a notion of ‘eventual coverage’ which, for 1-dimension, is identical to the equivalent formulation of the percolation of rumour process as given in Remark 1.1. We state the model in brief here and present the results obtained. Let {Xi : i ∈ Nd } be a collection of i.i.d. Bernoulli (p) random variables and {ρi : i ∈ Nd } a collection of i.i.d. N valued random variables, independent of the collection {Xi : i ∈ Nd }. Let ρ denote a generic random variable with the same distribution as ρi . Let C := ∪i:Xi =1 (i + [0, ρi ]d ) denote the covered region of Nd . Athreya, Roy and Sarkar (2004) define a notion of eventual coverage as follows: Nd is eventually covered if there exists a t ∈ Nd such that t + Nd ⊆ C. For our purposes we say Nd is eventually doubly covered if there exists a t ∈ Nd such that t + Nd ⊆ D, where D := {x ∈ Rd : there exist i, j ∈ Nd with i 6= j and Xi = Xj = 1 such that x ∈ (i + [0, ρi ]d ) ∩ (j + [0, ρj ]d )}. We have Proposition 1.6. (i) For d = 1, let l := lim inf jP(ρ ≥ j) > 1 and L := lim sup jP(ρ ≥ j) < ∞. j→∞ j→∞ We have Pp (N is eventually doubly covered) = 3 ( 1 0 if p > 1/l if p < 1/L. (ii) For d ≥ 2, we have Pp (Nd is eventually doubly covered) = ( 1 0 if lim inf j→∞ jP(ρ ≥ j) > 0 if limj→∞ jP(ρ ≥ j) = 0. In stochastic geometry the notion of coverage of space has received widespread attention. In particular Hall (1988) and Chiu, et al (2013) provide a review of the topics studied. Our endeavour in this paper may be viewed as an effort to introduce a notion of ‘reinforced coverage’. 2 Proof of Proposition 1.5 Let S be the event that "the firework process percolates" and S̃ be the event that "the reverse firework process percolates". Observe that, S= ∞ [ ∞ \ ∞ [ ([ρn−i ≥ i + 1] ∩ [ρn−j ≥ j + 1]) , n=1 i=1 j=1,j6=i and S̃ = ∞ [ ∞ \ ∞ [ ([ρn+i ≥ i] ∩ [ρn+j ≥ j]) . n=1 i=1 j=1,j6=i We will show that if p > 1/l, then we have P(S) = 1, P(S̃) = 1 and if p < 1/L, then P(S) = 0, P(S̃) = 0. We prove the proposition for the firework process, a similar proof works for the reverse firework process. Using De Morgan’s law, we have Sc = ∞ \ ∞ [ ∞ \ ([ρn−i ≤ i] ∪ [ρn−j ≤ j]) . n=1 i=1 j=1,j6=i 4 Since p > 1/l, there exists η > 1 such that l > η/p. Also let i0 be such that iP(ρ ≥ i) > ηp , for all i ≥ i0 . Therefore, for all i, j ≥ i0 and each fixed n,   ∞ ∞ \ \ P ([ρn−i ≤ i] ∪ [ρn−j ≤ j]) i=i0 j=i0 ,j6=i     ∞ ∞ ∞ ∞   \ \ \ \ [ρn−i ≤ i]  [ρn−j ≤ j]) ∪ = P     j=i0 i=j+1 i=i0 j=i+1     ∞ ∞ ∞ ∞ \ \ \ \ [ρn−i ≤ i] [ρn−j ≤ j] + P  = P j=i0 i=j+1 i=i0 j=i+1  ≤ P = ≤ ≤ ∞ \ ∞ Y [1 − j=i0 +1 ∞ Y [1 − [ρn−i ≤ i] i=i0 +1 [1 − P(ρ ≥ j + 1)] + j=i0 +1 ∞ Y 1 j+1 ] ∞ \ [ρn−j ≤ j] + P j=i0 +1 j=i0 +1 Q But, ∞ j=i0 +1 [1 − Finally  ∞ Y ! [1 − P(ρ ≥ i + 1)] i=i0 +1 ∞ Y η η [1 − ]+ ] p(j + 1) p(i + 1) i=i +1 0 ∞ Y 1 ]+ j+1 i=i = 0 as well as [1 − 0 +1 Q∞ i=i0 +1 [1 1 ]. i+1 − 1 i+1 ] = 0. Therefore, P(S c ) = 0. P(percolation among sceptic individuals) ≤ P(percolation among ‘non sceptical’ individuals) =0 for p < 1/L , where the equality above follows from Theorem 2.1 of Junior, Machado and Ravishankar (2016). 3 Proof of Proposition 1.6 As mentioned earlier (i) of Proposition 1.6 is just Proposition 1.5 rephrased. Thus we need to prove (ii). First note that from Proposition 3.2(a) of Athreya, Roy and Sarkar (2004), we know that if limj−→∞ jP(ρ ≥ j) = 0 then Pp (Nd is eventually covered) = 0, and so Pp (Nd is eventually doubly covered) = 0. 5 We prove Proposition 1.6 (ii) for the case d = 2; the proof carries through in a straightforward fashion for higher dimensions. Fix 0 < p < 1 and we assume that lim inf j−→∞ jP(ρ ≥ j) > 0. For (i, j) ∈ N2 , define Bi,j := {(i, j) 6∈ D}. If we show that, for some N ≥ 1, X Pp (Bi,j ) < ∞, i,j≥N then Borel-Cantelli lemma guarantees that, with probability 1, there exists N0 ≥ 1 such that (i, j) ∈ D for all i, j ≥ N0 , i.e. we have eventual coverage. For u := (i − l1 , j − l2 ) ∈ N2 with 1 ≤ l1 ≤ i, 1 ≤ l2 ≤ j, we observe that (i, j) ∈ u + [0, ρu ]2 if and only if ρu ≥ max{l1 , l2 }. Also note that for i ≥ j and 0 ≤ t ≤ j − 1 there are exactly 2t + 1 vertices (i − l1 , j − l2 ) with max{l1 , l2 } = t (viz. {(i − l1 , j − l2 ) : l1 = t and 0 ≤ l2 ≤ t or l2 = t and 0 ≤ l1 ≤ t}), while for j ≤ t ≤ i there are exactly j vertices (i − l1 , j − l2 ) with max{l1 , l2 } = t (viz. {(i − l1 , j − l2 ) : l1 = t and 0 ≤ l2 ≤ j − 1}). Thus, for Ai,j := {(i, j) 6∈ C}, and, for i ≥ j, Pp (Bi,j ) = Pp (Ai,j )+ i−j Y [1 − pG(k + j − 1)] k=1 j−1 Y [1 − pG(t)] 2t+1  j−1 X j  pG(t) [1 − pG(t)] t=0 j−1 Y 2t [1 − pG(l)] 2l+1   +  l6=t,l=0 t=0 ×  i−j X  pG(k + j − 1) [1 − pG(k + j − 1)] j−1 i−j Y [1 − pG(l + j − 1)] Noting that for i ≥ j Pp (Ai,j ) = j−1 Y [1 − pG(l)] 2l+1 l=0 i−j Y [1 − pG(k + j − 1)] . (2)  l6=k,l=1 k=1 j   j k=1 (the last product is taken to be 1 if i = j), (2) simplifies to Pp (Bi,j ) = Pp (Ai,j ) + pPp (Ai,j ) i−1 X t=0 G(t) . 1 − pG(t) (3) Before we proceed we fix some quantities. Since lim inf jP(ρ ≥ j) > 0 we may choose j−→∞ η > 0 such that lim inf jP(ρ ≥ j) > η. Also, for this η and our fixed p ∈ (0, 1) let a be j−→∞ 6 such 0 < e−pη < a < 1. Now we choose N ≥ 1 such that for all j ≥ N the following hold: (i) jP(ρ ≥ j) > η, (ii) (1 − pηj −1 )j < a, (4) (iii) pjη > 1. Note that (ii) above guarantees that for all j ≥ N , we have (1 − pG(j))j < a. From the proof of Proposition 3.2(b) of Athreya, Roy and Sarkar (2004) we know that, for j ≥ N , ∞ X Pp (Ai,j ) < ∞. i=1 Thus if we show that, for j ≥ N , ∞ X Pp (Ai,j ) i−1 X t=0 i=1 G(t) <∞ 1 − pG(t) (5) then we will have ∞ X Pp (Bi,j ) < ∞ whenever j ≥ N. (6) i=1 Note that, i−1 X t=0 ∞ X Pp (Ai,j ) i=j = j−1 k=1 i−j X k=1 ∞ j−1 Y X G(k + j − 1) 1 − pG(k + j − 1) [1 − pG(t)]2t+1 i=j t=0 = j−1 Y j−1 Y i−j Y [1 − pG(l + j − 1)]j l=1 [1 − pG(t)]2t+1 t=0 = i−j X G(t) X G(k + j − 1) G(t) = + . Now, for fixed j ≥ N , 1 − pG(t) 1 − pG(t) 1 − pG(k + j − 1) t=0 ∞ i−1 Y X k=1 [1 − pG(l)]j−1 2t+1 t=0 ∞ X ∞ i−1 X Y  i−1 X  i=j l=j [1 − pG(t)] i−j X G(k + j − 1) 1 − pG(k + j − 1) G(k) j−1 [1 − pG(h)] h6=k,h=j k=j [1 − pG(l)] i−1 Y G(k) i−1 Y [1 − pG(h)] for j ≥ N. Taking ∞ i−1 X Y [1 − pG(l)]  h6=k,h=j k=j i=k+1 l=j ek :=   j−1 G(k) i−1 Y h6=k,h=j i=k+1 l=j 7 [1 − pG(h)] (7) as the inner sum in (7), we have ek =G(k) k Y [1 − pG(l)]j−1 l=j k−1 Y [1 − pG(h)] h=j × ∞ X 1+ i Y [1 − pG(l)] i=k+1 l=k+1 =G(k) k Y [1 − pG(l)]j−1 l=j k−1 Y i Y j−1 [1 − pG(h)] h=k+1 ! [1 − pG(h)] 1 + [1 − pG(k)]j h=j h × 1+ ∞ X i Y [1 − pG(l)] j−1 i=k+2 l=k+2 i Y [1 − pG(h)] h=k+2 i ! . (8) ! (9) Similarly, for ek+1 , as in the term in the first equality above, we have ek+1 =G(k + 1) k+1 Y [1 − pG(l)]j−1 l=j × k Y [1 − pG(h)] h=j 1+ ∞ X i Y [1 − pG(l)] j−1 i=k+3 l=k+3 i Y [1 − pG(h)] . h=k+3 Taking C(k, j) := 1 + ∞ X i Y [1 − pG(l)] j−1 i=k+2 l=k+2 i Y [1 − pG(h)] h=k+2 we see that j−1 ek+1 [1 − pG(k)] G(k + 1) [1 − pG(k + 1)] = C(k, j) j . ek G(k) 1 + C(k, j) [1 − pG(k + 1)] Now, G(k+1) G(k) (10) ≤ 1 for all k ≥ 0, and, for fixed j, j−1 (i) for k large enough, 0.9 < [1 − pG(k + 1)] [1 − pG(k + 1)] ≤ 1; (ii) for j ≥ N , from equation (3.5) of Athreya, Roy and Sarkar (2004), we have P∞ Qi j−1 < ∞ and hence i=k+2 l=k+2 [1 − pG(l)] ∞ X i Y [1 − pG(l)] i=k+2 l=k+2 j−1 i Y [1 − pG(h)] < ∞; h=k+2 which ensures that, for k large enough, 1 ≤ C(k, j) < 1.1. 1.1 Thus, for j ≥ N and k large enough, we have ek+1 ek < 1.9 , and so, by ratio test, ∞. This shows, from (7), that (5) and thereby (6) hold. 8 P∞ k=1 ek < P Now we show that, for N as above, i,j≥N Pp (Bi,j ) < ∞. Towards this, we first observe that, for i, j ≥ 1, by symmetry we have Pp (Bi,j ) = Pp (Bj,i ), thus we need to show X Pp (Bi,j ) = 2 i,j≥N ∞ X i−1 X Pp (Bi,j ) + i=N j=N ∞ X (11) Pp (Bi,i ) < ∞. i=N We will show separately that ∞ X i−1 X Pp (Bi,j ) < ∞ and i=N j=N ∞ X (12) Pp (Bi,i ) < ∞. i=N First, for any i, j ≥ 1, " Pp (Bi,j ) = 1 + p i−1 X t=0 # G(t) Pp (Ai,j ). 1 − pG(t) From the proof of Proposition 3.2(b) of Athreya, Roy and Sarkar (2004) we know that, ∞ X i−1 X Pp (Ai,j ) < ∞ and i=N j=N ∞ X Pp (Ai,i ) < ∞, i=N thus we need to show that ∞ X i−1 X i−1 X i=N j=N t=0 ∞ X i−1 X G(t) G(t) Pp (Ai,j ) < ∞ and Pp (Ai,i ) < ∞. 1 − pG(t) 1 − pG(t) t=0 (13) i=N For the first sum, interchanging the order of summation we have ∞ X i−1 X i−1 X i=N j=N t=0 ∞ ∞ X X G(t) G(t) Pp (Ai,j ) = 1 − pG(t) 1 − pG(t) t=0 ∞ X Pp (Ai,j ). j=N +1 i=max(t+1,j+1) Also breaking up the inner two sums according to the values taken by i, writing out the expression for Pp (Ai,j ) and collecting terms together, we have ∞ X ∞ X Pp (Ai,j ) j=N +1 i=max(t+1,j+1) = ∞ X ∞ X Pp (AN +r,N +m ) m=1 r=max(t−N +1,m+1) = t−N X ∞ X m=1 r=t+1−N + ∞ X N +m−1 Y [1 − pG(l)]2l+1 l=0 ∞ X m=t−N +1 r=m+1 r−m Y [1 − pG(k + N + m − 1)]N +m k=1 N +m−1 Y [1 − pG(l)] 2l+1 l=0 r−m Y k=1 9 N +m [1 − pG(k + N + m − 1)] ). G(t) 1−pG(t) , To simplify the expressions we take σt := N +m g(k, m) := [1 − pG(k + N + m − 1)] equations we have ∞ X i−1 X i−1 X i=N j=N t=0 [1 − pG(l)] 2l+1 and l=0 . Using this notation, from the previous two ∞ t−N X X G(t) Pp (Ai,j ) = σt 1 − pG(t) t=1 m=1 + N +m−1 Y sm := ∞ X ∞ X sm r=t+1−N ∞ X σt r−m Y g(k, m) k=1 ∞ X sm m=t−N +1 r=m+1 t=1 r−m Y g(k, m). (14) k=1 We start with the first term on the right in the above equation. Reordering the sums, we have ∞ X t=1 σt t−N X ∞ X sm m=1 r=t+1−N r−m Y g(k, m) = ∞ X sm m=1 k=1 Let ∞ X αm := sm ∞ X ∞ X σt t=m+N σt t=m+N ∞ X r−m Y r−m Y g(k, m). (15) r=t+1−N k=1 g(k, m) r=t+1−N k=1 denote the summand. Observe that ∞ X sm+1 t=m+1+N αm+1 = ∞ αm sm X r=t+1−N ∞ X r−m−1 Y σt t=m+N = [1 − pG(N + ∞ X σt k=1 r−m Y ∞ X σt t=m+1+N ∞ r−m X Y r=t+1−N ∞ X r−m−1 Y k=1 r−m Y σt = [1 − pG(N + m)]N +m ∞ X ∞ X σt t=m+N σm+N [1−pG(N +m)]N +m+1 g(k, m) r=t+1−N k=1 t=m+1+N m)]2(N +m)+1 ∞ X × g(k, m + 1) ∞ X g(k, m) + g(k, m) r=t+1−N k=1 r=t+1−N r=m+1 k=1 g(k, m + 1) r−m−1 Y k=1 ∞ X g(k, m + 1) t=m+1+N σt ∞ X r=t+1−N r−m−1 Y . g(k, m + 1) k=1 By our choice of N and 0 < a < 1 as in (4), we have [1 − pG(N + m)]N +m < a. All P∞ P∞ Qr−m−1 terms above being non-negative, if t=m+1+N σt r=t+1−N k=1 g(k, m + 1) < ∞, 10 then ααm+1 < a; and so an application of the ratio test will yield that the sum in (15) is m finite. P∞ P∞ Qr−m−1 To show t=m+1+N σt r=t+1−N k=1 g(k, m + 1) < ∞, we again apply a ratio P∞ Qr−m−1 test. Let τt := σt r=t+1−N k=1 g(k, m + 1) < ∞. Then, G(t + 1) τt+1 = [1 − pG(t)] τt G(t) [1 − pG(t + 1)] N +m ∞ X 1+ r Y ! g(k, m + 1) r=t+2−N −m k=t+2−N −m ∞ r X Y × N +m+1 1+ 1 + [1 − pG(t + 1)] g(k, m + 1) r=t+2−N −m k=t+2−N −m Note that ∞ X r Y !. g(k, m + 1) < ∞ (from Athreya, Roy and Sarkar (2004)) r=t+2−N −m k=t+2−N −m and so we may obtain a t0 such that, for all t ≥ t0 , (i) 1 ≤ 1 + ∞ X r Y g(k, m + 1) < 1.1 and r=t+2−N −m k=t+2−N −m N +m+1 (ii) [1 − pG(t + 1)] 1+ ∞ X r Y ! g(k, m + 1) r=t+2−N −m k=t+2−N −m This choice of t0 ensures that for all t ≥ t0 we have ∞ X t=m+1+N σt ∞ X r=t+1−N r−m−1 Y τt+1 τt > 0.9. < 1.1/1.9, and hence g(k, m + 1) < ∞. k=1 Thus the first term on the right of (14) is finite. A similar calculation and a use of ratio test shows that the second term on the right of (14) is finite, thereby showing that the sum in (14) is finite. Reordering the second sum in (11) and using the notation we introduced earlier, we have ∞ X i−1 X i=N t=0 N −1 X ∞ ∞ X ∞ X X G(t) σt Pp (Ai,i ). Pp (Ai,i ) = σt Pp (Ai,i ) + 1 − pG(t) t=0 i=t+1 i=N (16) t=N From the end of Section 3.1 of Athreya, Roy and Sarkar (2004) we know that with N as PN −1 P∞ P∞ above, i=N σt Pp (Ai,i ) < ∞ and hence t=0 i=N σt Pp (Ai,i ) < ∞. Expanding the term Pp (Ai,i ), we have ∞ X ∞ X t=N i=t+1 σt Pp (Ai,i ) = ∞ X ∞ X t=N i=t+1 11 σt i−1 Y l=0 [1 − pG(l)] 2l+1 (17) Taking at := σt see that P∞ i=t+1 at := σt Qi−1 t Y l=0 [1 − pG(l)] [1 − pG(l)] 2l+1 2l+1 l=1 from which we have and lt := P∞ r=t+2 Qr l=t+2 [1 − pG(l)] 2l+1 , we n o 2t+3 1 + [1 − pG(t + 1)] (1 + lt ) ; 2t+3 at+1 [1 − pG(t)] [1 − pG(t + 1)] (1 + lt ) G(t + 1)   = 2t+3 at G(t) [1 − pG(t + 1)] 1 + [1 − pG(t + 1)] (1 + lt ) 2t+2 = (1 + lt ) G(t + 1) [1 − pG(t)] [1 − pG(t + 1)] 2t+3 G(t) 1 + [1 − pG(t + 1)] (1 + lt ) ≤ (1 + lt ) G(t + 1) [1 − pG(t + 1)] 2t+3 G(t) 1 + [1 − pG(t + 1)] (1 + lt ) 2t+2 Also, with a as in (4), Qr+1 l=t+2 Qr l=t+2 [1 − pG(l)] 2l+1 [1 − pG(l)] 2l+1 2r+3 = [1 − pG(r + 1)]  pη < 1−p r+1 2r+3 < a for r ≥ N, so, by the ratio test lt < ∞ for t ≥ N . Now, since lt → 0 as t → ∞, from the remark 2t+2 following (4) we may obtain a t0 such that for t ≥ t0 , we have [1 − pG(t + 1)] (1 + lt ) < G(t+1) a < 1. Also G(t) < 1 for all t, thus at+1 < a for all t ≥ t0 , at and by ratio test we have that the sum in (16) is finite. This completes the proof of Proposition 1.6 (ii). References 1. Athreya, S., Roy, R. and Sarkar, A. (2004). On the coverage of space by random sets. Adv. Appl. Probab. (SGSA), 36, 1–18. 2. Chiu, S.N, Stoyan, D., N., Kendall, W.S., V. and Mecke, J. (2013). Stochastic geometry and its applications 3rd Ed., John Wiley, Chichester. 12 3. Gallo, S., Garcia, N., Junior, V. and Rodríguez, P. (2014). Rumor Processes on N and discrete renewal processes. J. Stat. Physics, 155, 591–602. 4. Hall, P. (1988). Introduction to the theory of coverage processes John Wiley, New York. 5. Junior, V., Machado, F. and Zuluaga, M. (2011), Rumor processes on N. J.. Appl. Probab., 48, 624–636. 6. Junior, V., Machado, F. and Ravishankar, K. (2016) The rumor percolation model and its variations. arXiv:1612.03853. 13 

10math.ST

A Note on Fault Tolerant Reachability for Directed Graphs Loukas Georgiadis1 Robert E. Tarjan2 arXiv:1511.07741v1 [cs.DS] 24 Nov 2015 February 26, 2018 In this note we describe an application of low-high orders [2] in fault-tolerant network design. Baswana et al. [1] study the following reachability problem. We are given a flow graph G = (V, A) with start vertex s, and a spanning tree T = (V, AT ) rooted at s. We call a set of arcs A′ valid if the subgraph G′ = (V, AT ∪ A′ ) of G has the same dominators as G. The goal is to find a valid set of minimum size. Baswana et al. [1] show that there is a valid set containing at most n − 1 arcs, and they give an algorithm to compute a minimum-size valid set in O(m log n) time, where n = |V | and m = |A|. Their algorithm is based on generalizing the notion of semi-dominators [3] from a depth-first spanning tree to an arbitrary spanning tree. Here we give a simple algorithm, Algorithm 9 below, to compute a minimum-size valid set in O(m) time, given the dominator tree D and a low-high order of it. Since D and a low-high order can be computed in O(m) time [2], so can a minimum-size valid set. Algorithm 9: Construction of a minimum-size valid set A′ Initialize A′ to be empty. For each vertex v 6= s, apply the appropriate one of the following cases, where t(v) and d(v) are the parent of v in T and D respectively (d(v) is the immediate dominator of v): Case 1: t(v) = d(v). Do nothing. Case 2: t(v) 6= d(v) and (d(v), v) ∈ A. Add (d(v), v) to A′ . Case 3: (d(v), v) 6∈ A. Subcase 3a: t(v) > v. Add to A′ an arc (x, v) with x < v. Subcase 3b: t(v) < v. Add to A′ an arc (x, v) with x > v and x not a descendant of v in D. Theorem 1. The set A′ computed by Algorithm 9 is a minimum-size valid set. Proof. First, we show that in each occurrence of Cases 2 and 3, an arc will be added to A′ . This is obvious for Case 2. To show that this is true for Case 3, we use the definition of a low-high order [2, Section 2]: for any vertex v 6= s such that (d(v), v) 6∈ A, there are two arcs (u, v) and (w, v) such that u < v < w and w is not a descendant of v in D. Only one of these arcs can be in T . It 1 Department of Computer Science & Engineering, University of Ioannina, Greece. E-mail: loukas@cs.uoi.gr. Department of Computer Science, Princeton University, 35 Olden Street, Princeton, NJ, 08540, and Intertrust Technologies. E-mail: ret@cs.princeton.edu. 2 1 follows that in both Case 3a and Case 3b, there is an arc that satisfies the stated constraint, and hence such an arc will be added to A′ . Second, we show that A′ is valid. To do this, we use the notion of a pair of divergent spanning trees from [2]. Two spanning trees B and R, both rooted at s, are divergent if, for every vertex v, the paths from s to v in B and R have only the dominators of v in common. Algorithm 1′ below, which simplifies Algorithm 1 in [2] as discussed just before Theorem 2.10 in [2], constructs two divergent spanning trees, provided that each step is successful: the proof of Theorem 2.8 in [2] applies to Algorithm 1′ equally as well as to the original Algorithm 1. Indeed, as does Algorithm 1, Algorithm 1′ constructs a pair of strongly divergent spanning trees, but here we need only the weaker property of divergence. Algorithm 1′ : Construction of Two Divergent Spanning Trees B and R Let D be the dominator tree of flow graph G = (V, A), with a preorder that is a low-high order of G. For each vertex v 6= s, apply the appropriate one of the following cases to choose arcs (b(v), v) in B and (r(v), v) in R: Case 1: (d(v), v) ∈ A. Set b(v) = r(v) = t(v). Case 2: (d(v), v) 6∈ A. Choose two arcs (u, v) and (w, v) such that u < v < w in low-high order and w is not a descendant of v in D. Set b(v) = u and r(v) = w. Suppose we apply Algorithm 1′ to G, but only allow it to choose arcs that are in subgraph G′ . Suppose Case 1 of Algorithm 1′ applies to v. Then (d(v), v) is in G. But the construction of A′ guarantees that (d(v), v) is in G′ as well. Hence this case will add an arc in G′ , namely (d(v), v), to both B and R. Suppose Case 2 of Algorithm 1′ applies to v. Then (d(v), v) is not in G. Algorithm 9 will apply Case 3 to v, which guarantees that G′ contains arcs (u, v) and (w, v) such that u < v < w (in low-high order) and w is not a descendant of v in D: if t(v) > v, t(v) cannot be a descendant of v in D since there is a v-avoiding path in T from s to t(v). Hence this case can successfully choose arcs in G′ to add to B and R. We conclude that G′ contains two spanning trees that are divergent in G. It follows that if v dominates w in G′ , v dominates w in G. The converse is immediate, since G′ is a subgraph of G. Thus A′ is valid. Finally, we show that A′ is minimum-size. Set A′ contains an arc (u, v) only if t(v) 6= d(v). Any valid set must contain an arc entering v, since otherwise t(v) dominates v in G′ , contradicting validity. References [1] S. Baswana, K. Choudhary, and L. Roditty. Fault tolerant reachability for directed graphs. In Yoram Moses, editor, Distributed Computing, volume 9363 of Lecture Notes in Computer Science, pages 528–543. Springer Berlin Heidelberg, 2015. [2] L. Georgiadis and R. E. Tarjan. Dominator tree certification and divergent spanning trees. ACM Transactions on Algorithms, 12(1):11:1–11:42, November 2015. [3] T. Lengauer and R. E. Tarjan. A fast algorithm for finding dominators in a flowgraph. ACM Transactions on Programming Languages and Systems, 1(1):121–41, 1979. 2 

8cs.DS

Finite Element Model Updating Using Fish School Search Optimization Method Boulkabeit I *, Mthembu L * Marwala T * and De Lima Neto, F † *The Centre For Intelligent System Modelling (CISM), Electrical and Electronics Engineering Department, University of Johannesburg, PO Box 524, Auckland Park 2006, South Africa, e-mail: iboulkaibet@student.uj.ac.za † Polytechnic School of Pernambuco, University of Pernambuco, Recife, Brazil, e-mail: fbln@ecomp.poli.br Abstract—A recent nature inspired optimization algorithm, Fish School Search (FSS) is applied to the finite element model (FEM) updating problem. This method is tested on a GARTEUR SMAG19 aeroplane structure. The results of this algorithm are compared with two other metaheuristic algorithms; Genetic Algorithm (GA) and Particle Swarm Optimization (PSO). It is observed that on average, the FSS and PSO algorithms give more accurate results than the GA. A minor modification to the FSS is proposed. This modification improves the performance of FSS on the FEM updating problem which has a constrained search space. Keywords: Finite Element Model (FEM); Fish School Search (FSS);Genetic Algorithm (GA); Particle Swarm Optimization (PSO) 1. INTRODUCTION A finite element model (FEM) is a numerical method used to provide approximate solutions to the analysis of complex engineering systems [1, 2, 20]. In mechanical engineering, the FEM method is often used for computing displacements, stresses and strains in structures under different input conditions. In electrical engineering it might be used to model systems that experience electromagnetic fields. However, FEMs only produce accurate solutions to simple engineering problems. In complex systems, FEM results are different from those obtained from physical experiments [3, 4, 20]. These differences can be the result of; the approximations made in the geometric construction the system model and/or the description of some system parameters e.g. joints. Therefore the initial FE model needs to be automatically and intelligently updated so as to match the measured experimental data. There are currently two main approaches to FEM updating; direct and indirect methods [1, 6]. In the direct method the measured data are directly equated to the FEM output. This naturally constrains the updating to FE system matrices (mass, stiffness etc) only. In the indirect (or iterative) approach the FEM outputs are not constrained to equal the measured data thus allowing both the system matrices and the model output to be variable. The updating problem is then how to minimise the difference between the measured data and the FEM output, given these “free” variables i.e. this is a classic optimization problem. Another way of looking at this model updating problem is that we know what the model results should be, the problem is to identify the system that generate the results. So FEM updating can also be seen as a system identification problem which is made more difficult by the often large number of uncertain parameters. Furthermore in FEM unlike in artificial models, the problem under consideration is a real physical system. Thus the updated parameters have physical meaning. This places significant limits on possible uncertain-parameter update values. This is one of the main criticisms of direct updating methods discussed above [1]. In the recent years, nature inspired optimization techniques, commonly referred to as metaheuristic, have shown promising results in systems identification research [5]. Many optimization methods such as Simulated annealing (SA), Genetic algorithm (GA) and Particle Swarm optimization (PSO) have been used for FEM updating [6, 7, 8]. In this paper we compare the performance of the recently introduced Fish School Search (FSS) algorithm [10] to the performance of the former algorithms on the FEM updating problem. 2. EVOLUTION SEARCH 2.1 Introduction All nature-inspired search algorithms are population based and have the same structure [9]. They all have a population of individuals which move around the problem space searching for the solution to the problem. The dimension of the search space is normally defined by the number of uncertain function/problem parameters. All the individuals are potential solutions to the problem at hand. The following generally occurs for the individuals during the search process in all the above algorithms:    Individuals are randomly initiated, in the beginning there is no bias to any particular point in the search space; Individuals interact to varying degrees at a local (exploitation) and global (exploration) level; Individuals are all evaluated during the search process;  There is a level of randomness at all stages of the search process. The novelty between algorithms is mainly how the potential solutions search/move (or update) in the problem space and how they interact (and or communicate) to each other. The next section briefly describes the different algorithms compared in this paper. In section three the modeled structure is described and in section four the simulation results are presented. Section five concludes the paper. 2.2 Particle Swarm Optimization (PSO) PSO was introduced by Kennedy and Eberhart [7]. It is based on the foraging behavior of a flock of birds (particles). As a standard in all PSO algorithms [6, 7, 8] each individual (i) is composed of three vectors: its position in the search space (size d) is given by xi  ( xi1 , xi2 , xi3 ,.....xid ) , the best position it has found pi  ( pi , pi , pi ,..... pi ) 1 2 3 d and its velocity v i  ( v i , v i , v i ,..... v i ) . The position and the velocity are initialized randomly in the search space. 1 2 3 d The particles move around the search space by updating their velocity and position vectors. For the current context each position on the search space is a candidate solution to the FEM updating problem. The velocity can be updated using various equations; in this paper we use the PSO version with the inertia weight not the constriction factor [7]. The velocity update equation for particle i is given by: v i  wv d d i  c 1 r1 ( p i  x i )  c 2 r2 ( p g  x i ) d d d d (1) where w is the inertia weight introduced to adjust the influence of the previous velocities on the optimization process. A large inertia weight value helps in global exploration, while a small one helps in searching within nearby areas. The inertia weight used in this paper is given by: w  max ter  iter (2) max iter The position vector is then updated using xi  xi  vi d d d (3) (detailed in §3) the problem space d =8. In GAs an individual’s make-up ( x i ) is described by the chromosome vector itself. Unlike in PSO and FSS where there is a defining equation for how an individual moves through the search space, in GAs that notion is achieved by selection, reproduction and mutation between and of individuals. This allows the algorithm to gradually converge to better solutions over generations. 2.4 Fish School Search (FSS) This method was inspired from the behavior of a school of fish and was first thought by Bastos Filho and Lima-Neto in 2007, subsequently proposed in 2008 [10]. Similar to PSO and GA, the search process in FSS is carried out by a population of individuals, i.e. fish, which have a certain level of local and global behavior. The problem space is said to have food (possible solutions) scattered in different concentrations of the aquarium (i.e. search space) and the fish are considered to ‘feed’ at these different positions as they search. In FSS, searching is referred to as swimming. This action is achieved via three effectors; individual movement, collectiveinfluence movement and school performance (called collectivevolitive) movement. The initial individual position is random. Thereafter the movement is given by: x i ( t )  x i ( t  1 )  rand (  1,1 ) step (4) The larger this value is the more successful the fish has been in searching for food. The weight value is constrained to vary between 1 and w scale . All fish are initialized with a weight of half w scale . The weight of each individual in the population is updated using Equation 5: w i ( t )  w i ( t  1)  f [ x i ( t )]  f [ x i ( t  1 )] where f (5) f [ x i ( t )]  f [ x i ( t  1 )] |} max{| c 2 are the local and global influence factors respectively. Perhaps, the Genetic Algorithm (GA) is one of the first metaheuristic (of the nature inspired algorithms) used for optimization [6, 9, 22]. The idea stems from the way biological genes (more specifically chromosomes) of parents are combined to produce better and better off-springs over generations. ( t  1) where rand is a random number uniformly generated in the interval [-1,1] and step ind is an individual step size. The step size decreases linearly during the search process so as to exploit later positions in the search process. Before movement to any particular point is taken the fish evaluates whether the food in that direction is better than at its current position. The fish then feeds. Individuals in FSS ‘store’ their previous performance information via their weight w i (t ) . The r1 and r 2 are independent random numbers and c 1 and 2.3 Genetic Algorithm ind is the finite element model and f (xi ) computes the fitness of each fish in the school. After the individual movements a weighted average of all the fish is calculated. This biases the future movement of fish towards those fish that had a successful previous movement. The resultant fish movement is calculated using: N In the current context a potential solution to the FEM updating problem is modeled as a vector of chromosomes. Each chromosome is an uncertain parameter of the function to be optimized. Again for the current FEM updating structure  x i ( t )  x i ( t  1)   x ind { f [ x i ( t )]  f [ x i ( t  1 )] } i 1 N  { f [x i 1 i ( t )]  f [ x i ( t  1 )]} (6) where  x ind is the individual displacement of the fish and N is the number of fish in the school. This is the collective – influence effector. The fish complete the swim movement with the collective - volitive adjustment step. This step takes into account how the whole school is performing so far. This step requires the calculation of the average weighted school position called the barycenter at time t: N  B (t )  x i (t ) w i (t ) i 1 (7) N  x i (t ) i 1 The final fish position will depend on whether the weight of the school has increased or decreased, this is given by Equations 8 and 9 respectively. The Equation 8 will be evaluated in the case that the weight has been increased. However, the Equation 9 will be used when the weight decreases: x i ( t )  x i ( t  1 )  step vol rand x i ( t )  x i ( t  1 )  step vol rand { x i ( t  1 )  B ( t  1 )} (8) dist ({ x i ( t  1 ), B ( t  1 )} { x i ( t  1 )  B ( t  1 )} 3. MODELLED STRUCTURE AND FE MODEL All the finite element modeling was simulated using version 6.2 of the Structural Dynamics Toolbox (SDT®) under MATLAB® environment. In this paper, a GARTEUR SMAG19 aeroplane structure is used to investigate the optimization capability of the four algorithms. The GARTEUR SM-AG19 structure was used as a benchmark study by 12 members of the GARTEUR Structures and Materials Action Group 19 [13, 14, 19]. One of the aims of the study was to compare the different computational model updating procedures on a single common test structure [12, 13, 15, 16, 19, 18]. The benchmark study also allowed participants to test a single representative structure using their own test equipment. The experimental test data used in our analysis is data obtained by the University of Manchester (U MAN) . The above aeroplane has a length of 1.5 m and a width of 3m. The depth of the fuselage is 15cm with a thickness of 5cm. Figure 1 shows the FE model of the aeroplane. In our models all element materials are considered standard isotropic. The model elements are Euler–Bernoulli. The measured natural frequency (Hz) data is: 6.51, 16.37, 33.44, 33.97, 36.17, 49.41, 50.2, 55.61, 64.04, 69.39. (9) dist ({ x i ( t  1 ), B ( t  1 )} The parameter rand is a random number uniformly generated in the interval [0,1].The step vol is used to control the displacement from or to the barycenter and it is decreased linearly as the search proceeds. The dist ( x , y ) is the Euclidian distance between x and y. This final positioning based on the average school weight increasing or decreasing effectively implements the space exploitation or exploration concept respectively. The fish exploit a particular area of space if the school collectively gains weight (i.e. school radius reduce) or otherwise explores other areas by expanding away from the barycenter if the school looses weight. 2.5 Fish School Search biased (FSSb) The original FSS algorithm is slightly modified by giving more influence to results that seems to be more promising. This was done by simply altering the feeding of fish. A variable , which allows for selective allocation of more weight to better search results is introduced to Equation 5. Obviously, this parameter is problem specific. The new equation is then: w i ( t )  w i ( t  1)   f [ x i ( t )]  f [ x i ( t  1 )] max{| Figure 1 FEM Garteur Structure The ( parameters to be updated are the right wing ), the left wing ( ), Vertical tail ) and the overall structure’s density ). The search space is thus x d  8 . The initial position vector, x i (t ) , for all algorithms is given by E =[ , Table 1 The initial vector position for all algorithms Parameter (10) f [ x i ( t )]  f [ x i ( t  1 )] |} ( 2700 8.3 8.3 8.3 4.0 Parameter where ( 4.0 { ] (11) 8.3 I. 4. SIMULATION RESULTS The weight is bounded by: In all simulations the number of iterations is set to 500. 4.1 PSO Settings In the PSO algorithm, the population number is set to 20, and constants are set to be 2. The particle position and velocity elements are bounded as shown in Table 2 Table 2 Particle position and velocity bounds Max_position Min_position Max_velocity 10 Min_velocity -10 w scale . Each algorithm is executed 30 times where the average solutions of these runs are presented in Table 5 and 6. Table 5 presents the initial value (the mean material or geometric values) of the update vector , as well as the updated values obtained by each method. The updated parameters obtained by GA, PSO, FSS and FSSb techniques are close to the mean values i.e. they are physically realistic for the given structure. In general, the updated vector obtained from GA, PSO and FSS algorithms are different from each other and this because of the way that each algorithm functions to generate the solution. The FSS and FSSb have very close update vector where involving the global and the local best position slightly affected the updated vector . PSO and FSS techniques give updated vectors that are more similar to each to those obtained by the GA algorithm. Table 5 Initial and updated parameter values Initial E vector, GA Method 2573.7 vector, PSO Method 2682.7 vector, FSS Method vector, FSSb Method 4.2 GA settings Table 3 shows the parameter setting used in the GA algorithm. Table 3 GA parameter settings Population (N) Mutation rate (mutr) Selection Rate (selcr) 20 0.2 0.5 As in the PSO algorithm the chromosomes in the genetic algorithm are bounded by the values shown in Table 2. 4.3 FSS settings In FSS algorithm the fish population is 20. The population positions are bounded according to table 2. The and are a percentage of the search space amplitude and are bounded by two vector parameters according to Table 4: Table 4 FSS parameter settings 30 -30 . Table 6 shows the output errors of the different optimization algorithms. In general, the results show that the updated FEM natural frequencies are better than the initial FEM for all methods (see the total (T) error in Table 6). The error between the second measured natural frequency and that of the initial model was 6.303%. With the GA method this error was reduced to 0.915% and by implementing the PSO it was reduced to 0.763%, and when the FSS and FSSb are implemented, the error reduced to 0.727% and 0.651% respectively. The same comment can be made for the third, fourth, seventh, eighth, ninth and the tenth natural frequencies. The PSO, FSS and FSSb give results that are better than the GA algorithm. On the other hand, GA converges faster to a local minimum, see figure 2. FSS and FSSb converge within the first 15 iterations. PSO, FSS and FSSb algorithms have almost the same asymptotic point (around 70 to 75 iterations). There is no significant difference between the total errors obtained for the PSO, FSS and FSSb [21]. Table 6 Natural Frequencies and Errors when GA, PSO, FSS and FSSb methods are used to update the FEM Mode Measured Frequency (Hz) Initial FEM Frequency (Hz) Error (%) Frequencies GA Method (Hz) Error (%) Frequencies PSO Method (Hz) Error (%) Frequencies FSS Method (Hz) Error (%) Frequencies FSSb Method (Hz) Error (%) 1 6.51 5.726 12.047 6.247 4.037 6.237 4.195 6.232 4.266 6.224 4.396 2 16.37 15.338 6.303 16.22 0.915 16.495 0.763 16.489 0.727 16. 77 0.651 3 33.44 32.457 2.939 33.086 1.057 33.306 0.399 33.278 0.486 33.350 0.269 4 33.97 35.323 3.984 34.949 2.882 34.154 0.542 34.127 0.461 34.222 0.742 5 36.17 36.020 0.414 36.284 0.316 35.908 0.724 35.896 0.759 35.925 0.678 6 49.41 44.992 8.941 49.05 0.728 48.839 1.156 48.799 1.237 48.730 1.377 7 50.20 54.685 8.934 53.964 7.499 52.938 5.455 52.903 5.384 52.834 5.248 8 55.61 55.753 0.257 54.603 1.811 55.512 0.176 55.574 0.065 55.603 0.012 9 64.04 60.021 6.276 63.695 0.538 64.16 0.187 64.130 0.140 64.068 0.043 10 69.39 68.745 0.929 70.326 1.349 68.922 0.675 68.868 0.752 68.768 0.859 T error % ______ ______ 51.024 ______ 21.132 ______ 14.272 ______ 14.277 ______ 14.275 The modified FSS algorithm, FSSb, outrank PSO on six out of the 10 modes. For example the error of the second mode was reduced to 0.763 % by the PSO algorithm while the FSSb method reduced it to 0.651 %. The same comment can be made for the third, fifth, seventh, eighth and ninth natural frequencies. Figure 3 shows a magnified plot of the cost function at higher iteration counts. Figure 2 FEM cost function vs. iterations It is difficult to make a decisive conclusion from Table 6 and Figure 2 when comparing the FSS, FSSb and PSO algorithms. The difference between the algorithm results is on the different natural modes. The FSS algorithm has smaller errors than PSO in five of the modes and PSO has smaller errors than FSS for a different set of five modes. On average the total error between these algorithms is similar as shown in figure 2. Figure 3 Magnified FEM cost function vs. iterations Figure 3 also shows that the total error of the PSO algorithm is slightly better than both FSS and FSSb algorithms, even though as commented it does not converge as fast. The monification proposed in FSSb seemed to slightly improve results over “vanilla” FSS. The algorithm errors do not improve after the magnified region. 5. CONCLUSION [7] [8] In this paper the finite element model updating problem is addressed using a recently proposed nature inspired optimization algorithm. Three of these metaheuristic methods; the genetic algorithm (GA), particle swarm optimization (PSO) and fish school search (FSS) algorithms are tested on the updating of the GARTEUR SM-AG19 FEM structure. A slight modification to the recent FSS algorithm is proposed and tested; we called it FSSb (‘b’ for bias). [9] The simulation results show that the FSS, FSSb and the PSO algorithms give more accurate results than those obtained by the GA algorithm. On the other hand, the GA algorithm has a faster convergence rate than the former although it prematurely converges. The modification on FSS algorithm improves the results of the FEM over the PSO algorithm. However, the PSO algorithm gives slightly better total error than those obtained by both FSS and FSSb algorithms. [12] [10] [11] [13] [14] [15] Further work will consider the differences between the above methods and the density based FSS algorithm (dFSS) [21]. Our preliminary analysis reveals that the FSSb algorithm may be more suited for applications that have a discretized search space, such as in FEM. But this has to be further verified in other domains. [16] [17] [18] REFERENCES [1] Friswell M.I, Mottershead J.E, Finite element model updating in structural dynamics, Kluwer Academic Publishers: Boston, 1995. [2] Leonard J.W, Khouri B.R, System identification using a standard finite element program, Engineering Structures 7 (3) (1985), 190-197. [3] Levin M.I, Lieven N.A.J, Dynamic finite element model updating using simulated annealing and Genetic Algorithms, Mechanical Systems and Signal Processing 12 (1) (1998) 91–120. Marwala T, Finite element model updating using wavelet data and Genetic Algorithm, Journal of Aircraft, 39 (4) (2002) 709-711. [19] [20] [21] [4] [5] [6] Panigrahi S.K, Chakraverty S, Mishra B.K. Vibration based damage detection in a uniform strength beam using genetic algorithm, Meccanica, Springer, 44, (2009), 697 – 710. Marwala T, Finite element model updating using computational intelligence techniques: Applications to structural dynamics, 2010 Heidelberg: Springer ISBN 978-1-84996-322-0. [22] Kennedy J, Eberhart R, Particle swarm optimization, In Proceedings of the International Conference on Neural Networks IV (1995) 19421948 IEEE. Bratton D, Kennedy J, Defining a standard for Particle Swarm Optimimzation. Proceedings of the 2007 IEEE Swarm Intelligence Symposium (SIS 2007). Goldberg, D.E, Genetic Algorithms in search, optimization and machine learning, Reading, MA: Addison Wesley, 1989 ISBN: 0201157675. C. J. A. B Filho., F. B. de Lima Neto, A. J. C. C.. Lins, A. I. S. Nascimento., and M. P. Lima, , "A novel search algorithm based on fish school behavior," Systems, Man and Cybernetics, SMC 2008. IEEE International Conference on, 2008, pp. 2646-2651. Link M, Friswell MI, Generation of validated structural dynamic models - Results of a benchmark study utilizing the GARTEUR SMAG19 Testbed, Mechanical Systems and Signal Processing, COST Action Special Issue, 17(1) January 2003, 9-20. Mottershead J.E, Friswell M.I, Model updating in structural dynamics - A Survey, Journal of Sound and Vibration, 167 (2) (1993) 347-375. Carvallo J, Datta B.N, Gupta A, Lagadapati M, A direct method for model updating with incomplete measured data and without spurious modes, Mechanical Systems and Signal Processing 21 (7) 2715-2731. Datta B.N, Finite element model updating, eigenstructure assignment and eigenvalue embedding techniques for vibrating systems, Mechanical Systems and Signal Processing 16 (2002), 83-96. Adhikari S, Friswell M.I, Distributed parameter model updating using the Karhunen-Loeve expansion, Mechanical Systems and Signal Processing 24(2) (2010) 325-339. Horibe T, Takahashi K, Crack Identification in beam using Genetic Algorithm and three dimensional P-FEM, Journal of Solid Mechanics and Materials Engineering 1 (7) (2007) 886-894. Hirotugu A, A new look at the statistical model identification, IEEE Transactions on Automatic Control 19 (6) (1974), 716–723. Leonard J.W, Khouri B.R, System identification using a standard finite element program, Engineering Structures 7 (3) (1985), 190197. Guyon I, Elisseeff A, An introduction to variable and feature selection, Journal of Machine Learning Research 3 (2003), 1157-1182. Boulkaibet I., Marwala T., Mthembu L., Friswell M. I. and Adhikari S.: Sampling Techniques in Bayesian Finite Element Model Updating, Proceedings of the Society for Experimental Mechanics, Vol. 29 pp. 7583, 2012. Madeiro S. S., Lima Neto F. B., Filho C.J.A.B, and Figueiredo E. M. N , ”Density as the segregation mechanism in fish school search for multimodal optimization problems”, Advances in Swarm Intelligence. ICSI 2011. Int’l Conference on, vol. 2, 2011, pp. 563-572. Holland, J. H.: Adaptation in Natural and Artificial Systems. Ann Arbor: University of Michigan Press, 1975. 

9cs.NE

Asymptotic equivalence of paired Hotelling test and conditional logistic regression arXiv:1610.06774v1 [math.ST] 21 Oct 2016 Félix Balazard1,2 1 2 Sorbonne Universités, UPMC Univ Paris 06, CNRS INSERM U1169, Hôpital Bicêtre, Université Paris-Sud, felix.balazard@inserm.fr Abstract Matching, the stratification of observations, is of primary importance for the analysis of observational studies. We show that the score test of conditional logistic regression and the paired Student/Hotelling test, two tests for paired data, are asymptotically equivalent. Keywords: matching, conditional logistic regression, asymptotic equivalence, paired Student test, observational studies. 1 Introduction Observational studies are the main source of information in many areas of science such as epidemiology or the social and political sciences (Rosenbaum, 2002). Such studies aim at estimating the effect of a treatment on an outcome and they differ from randomized experiments by the absence of random allocation of treatment. This lack of randomization is due to ethical or economical considerations and as a consequence, observational studies are subject to confounding. In the typical context of an observational study, we have n observations of a random outcome Y and a random predictor or treatment X. The analysis is complicated by the possibility of confounding i.e., the influence of a third variable Z on X or Y . To take an example, suppose Y is lung cancer, X is yellow coloration of the teeth and Z is smoker status. If we do not take into account Z, we will find an association between Y and X but if we control for Z, the association will disappear since both lung cancer and yellow coloration of the teeth are caused by smoking. In a randomized experiment, randomization will ensure that the treated and non-treated group have similar values for all confounders. Matching is an important method to control for bias in observational studies. Matching is the grouping of observations in strata. Such a grouping is necessary when the observations in a strata cannot be considered independent. For example, consider a continuous variable, such as blood pressure, which is measured before and after the beginning of a medical treatment given to a group of patients. In order to show that the treatment is effective, an incorrect 1 approach would be to compare the mean of the group before treatment to the mean after treatment using a two-samples Student test. This naive approach relies on the unwarranted assumption of independence of the measurement of the same patient before and after the treatment. The correct approach is to compute the difference between blood pressure before and after treatment for each patient and then to test if the differences have zero mean using a Student test. This is called the paired Student test. In this simple example, the stratas are the pairs of observations of the same patient. In presence of confounding, matching on the values of the confounder remove bias (Rubin, 1973) or when there are several confounders, other techniques based on matching can be used (Rosenbaum and Rubin, 1983). Matching can also guide the data collection of an observational study. This is best illustrated by the case-control study in epidemiology. Such a study begins by gathering disease cases for example at hospitals. The recruitment of controls is less straightforward than the recruitment of cases and must try and limit confounding. One way of achieving this is to choose controls that are naturally matched to cases: for each case, controls are chosen among friends or neighbors whose –potentially unobserved– confounders are similar to that of the case. As the case and his controls have been chosen for their similarity, they cannot be considered independent and must be grouped in a strata. Once a matching is obtained, adapted statistical procedures are needed and especially statistically sound testing procedures. A good reference on statistical methods for matched data is the monograph by Breslow and Day (1981). For strata limited to pairs of one case and one control (or before and after treatment) and when the predictor X is a binary variable, the so-called McNemar test can be used (McNemar, 1947). In this simple case, there are only 4 possibilities for each pair. The data can be summarized in the following two-by-two contingency table. Table 1: McNemar test Y =1 X=0 X=1 X=0 a b Y =0 X=1 c d In the table above, a + d is the number of pairs of a case and a control concordant for the predictor X and b + c is the number of discordant pairs. The statistic ξmc of the McNemar test depends only on the numbers of discordant pairs b and c and has the following form: ξmc = (b − c)2 . b+c Under the null hypothesis of independence between Y and X, ξmc follows the χ2 distribution asymptotically (McNemar, 1947). This test statistic has the advantage of being computationally simple. However, in many settings, we have access to more than one control for each case. For example, in the Isis-Diab study (Balazard et al., 2016), each patient was asked to recruit two controls . A test for strata of arbitrary size and a binary predictor was proposed in Mantel and Haenszel (1959) and is now referred to as the Cochran-Mantel-Haenszel (CMH) test. It 2 also provides an estimate of the odd-ratio associated with the predictor. The original article contains an interesting discussion of practical aspects of the prospective study of disease. If the predictor is not binary but continuous, appropriate tests are available if the matching is a pairing i.e., each case has exactly one control. If the vector of differences between cases and controls can be assumed to be normally distributed, a Student test can be applied on this vector to decide if its mean is zero as in our first example. The resulting test is called the paired Student test (McDonald, 2009, p180-185). The paired Hotelling test is the generalization of the paired Student test when there are several predictors. When the normality assumption is not respected, a non-parametric alternative is available via the Wilcoxon signed rank test. None of the above procedures allow for analysis of continuous variables with arbitrary strata size. This was made possible by conditional logistic regression (CLR) first introduced in Breslow et al. (1978). CLR is however not limited to continuous variables and can be applied in all settings of the previous tests. As CLR is based on logistic regression, it is more flexible than the previous tests and allows for multivariate analyses. When different statistical tests are available for the same data, it is desirable to prove that their results are similar. For procedures based on maximum likelihood such as CLR, the three tests of significance for predictors (the Wald test, the likelihood ratio test and the score test) have been shown to be asymptotically equivalent (Engle, 1984). This result of asymptotic equivalence is with respect to the null hypothesis as well as to a sequence of local alternatives. For the analysis of matched data with a binary predictor, the available tests have been shown to be identical. When the strata are pairs, both the CMH test and the McNemar test can be applied and their statistic is identical (Agresti and Kateri, 2011, p413). When the strata size is arbitrary, CMH and CLR can be applied and the CMH statistic is identical to the score test statistic of CLR (Day and Byar, 1979) and therefore if the strata are pairs, the McNemar statistic will be identical to the score test statistic of CLR. When the strata are pairs and the predictor (resp. predictors) is continuous, the paired Student test (resp. paired Hotelling test) and CLR can be applied. However, to the best of the author’s knowledge, there are no results comparing the two procedures in that case. The objective of this note is to prove that the paired Hotelling test is asymptotically equivalent to the score test of CLR and therefore also to the Wald and likelihood ratio tests. We start by deriving the exact form of the two test statistics in section 2. Finally, we present our result and prove it in section 3. 2 Test statistics Let Yi` be the label of the `-th individual of the i-th stratum and Xi` ∈ Rp its continuous predictor value. The paired Hotelling test can be applied when stratas are pairs of discordant observations and therefore ` ∈ {1, 2}. Without loss of generality, we can assume that Yi1 = 1 and Yi2 = 0. We are interested in testing the null hypothesis: H0 : There is no association between Y and X 3 (1) against the alternative hypothesis: H1 : There is an association between Y and X. To test this null hypothesis in this setting, we have two options: the paired Hotelling test whose test statistic we denote ξhot or CLR whose score test statistic we denote ξsc . We now derive the expressions of those two test statistics. As we will show below, the paired Hotelling test and CLR depend only on the vector of differences between pairs. We therefore denote Zi = Xi1 − Xi2 the difference between pairs. Paired Hotelling test The hypotheses tested by the paired Hotelling test are expressed by the parameter µ = (E[Zj ])j∈{1,...,p} , the mean of the difference between pairs. Using this parameter, the hypothesis (2) becomes H0 : µ = 0. Let the unbiased sample covariance be defined as: n C= where Z̄ = 1 n Pn i=1 1 X (Zi − Z̄)(Zi − Z̄)> , n − 1 i=1 Zi and ∗> is the transposition operator. The statistic of the test is: ξhot = nZ̄ > C −1 Z̄. (2) When Z follows a centered Gaussian distribution, the distribution of ξhot is called the Hotelling distribution. Conditional Logistic Regression score test In logistic regression, the likelihood of an observation, given the parameters α ∈ R and β ∈ Rp , equals: > ey(α+β X) P(Y = y|X) = , > 1 + eα+β X with y ∈ {0, 1}. Logistic regression can take into account stratification by having a different constant term αi for each stratum. However, when the strata are small e.g., pairs, this leads to a large number of parameters and biases parameter estimation (Breslow and Day, 1981, p249-251). Conditional logistic regression adresses the problem by conditioning the likelihood on the number of cases in each stratum. This eliminates the need to estimate the αi . In the case of pairings, when the first observation is a case and the second is a control, the conditional likelihood of the i-th strata is: 4 P(Yi1 = 1, Yi2 = 0|Xi1 , Xi2 , Yi1 + Yi2 = 1) P(Yi1 = 1|Xi1 )P(Yi2 = 0|Xi2 ) = P(Yi1 = 1|Xi1 )P(Yi2 = 0|Xi2 ) + P(Yi1 = 0|Xi1 )P(Yi2 = 1|Xi2 ) = = = exp(αi +β > Xi1 ) 1+exp(αi +β > Xi1 ) × 1 1+exp(αi +β > Xi2 ) exp(αi +β > Xi1 ) exp(αi +β > Xi2 ) × 1+exp(α 1+β> X ) + 1+exp(α 1+β> X ) × 1+exp(α > 1+exp(αi +β > Xi1 ) i i2 i i1 i +β Xi2 ) exp(β > Xi1 ) (The αi have been eliminated.) exp(β > Xi1 ) + exp(β > Xi2 ) 1 . 1 + exp(β > (Xi2 − Xi1 )) In the general case, when there is k cases in a strata of size m, with the cases being the first k observations, the conditional likelihood of a strata can be written: P m X exp( kj=1 β > Xij ) P(Yij = 1 for j ≤ k, Yij = 0 for j > k|Xi1 , . . . , Xim , Yij = k) = P , P > exp(β X ) m ij J∈C j∈J j=1 k where Ckm is the set of all subsets of size k of the set {1, . . . , m}. To obtain the full conditional likelihood, we multiply over each stratum and take the log to obtain the log-likelihood L(β, Z): L(β, Z) = − n X log(1 + e−β > Zi ). i=1 The score s(β, Z) is then defined by: n X Zi ∂L (β, Z) = s(β, Z) = . > ∂β exp(β Zi ) + 1 i=1 If β = β 0 , the covariance matrix of s(β 0 , Z) is Fisher’s information matrix: !   n X Zij Zik exp(β > Z ) ∂ 2L i 0 I(β 0 ) = E[− (β , z)|β 0 ] = > ∂β j ∂β k 0 (exp(β Z ) + 1)2 0 i j,k∈{1,...,p} i=1 . j,k∈{1,...,p} To test the hypothesis H0 : β = β 0 , the score test statistic is: ξsc = s(β 0 , Z)> I(β 0 , z)−1 s(β 0 , Z) Expressed using β, hypothesis (2) becomes H0 : β = 0. As shown above, the null hypothesis of the paired Hotelling test is expressed using µ, the mean of the difference between pairs, rather than β, the regression coefficient. There is no general correspondence between the two parameters. However, the two null hypotheses that are of interest in practice are the same: µ = 0 ⇐⇒ β = 0. 5 For β = 0, the score, Fisher’s information matrix and the test statistic become: n s(0, Z) = Z̄, 2 n 1X Zi Zi> I(0, Z) = 4 i=1 and n ξsc = nZ̄ > 1X Zi Zi> n i=1 !−1 Z̄. (3) Now that we have expressed analytically ξhot and ξsc , we state our result in the following section. 3 Asymptotic equivalence In this section, we introduce the sequence of local alternatives and state the result of asymptotic equivalence between the two tests. We adopt the framework of Engle (1984) for asymptotic equivalence. The motivation for considering a sequence of local alternatives is that any reasonable test will have the right size and will reject a fixed alternative when the number of observations becomes large. The sequence of local alternatives considers deviations from the null that approach the null as sample size increases. This allows to compare tests more precisely. (n) We model the sequence of local alternatives by a triangular array of observations (Zi )n∈N,i∈{1,...,n} that respects the following assumption. Let δ ∈ Rp and Σ be a positive definite matrix of dimension p. (n) Assumption 1. For all n ∈ N, i ∈ {1, . . . , n}, we have E[Zi ] = (n) In addition, Wi (n) = Zi − √δ n √δ n (n) and Cov[Zi ] = Σ. are independent and identically distributed. The triangular array of Assumption 1 is the sequence of local alternatives. The parameter δ controls the deviation from the null. The √ null hypothesis is the special case of the sequence of local alternatives when δ = 0. The n in the denominator of the deviation shrinks the deviation towards the null as the sample size increases. This particular power of n ensures that the test statistics are well-behaved as they converge in distribution. We denote ξsc,n (resp. ξhot,n ) the statistic of the score test (resp. of the paired Hotelling test) associated with Z (n) , the n-th line of the triangular array of Assumption 1. In the same fashion as for the Student distribution and the χ2 distribution, the Hotelling distribution with degrees of freedom p and n − 1 and the χ2p distribution are different for P finite samples but they are the same asymptotically. Therefore, ξsc,n −ξhot,n − → 0 is enough to guarantee asymptotic equivalence. Our theorem is more precise as it includes the convergence rate and limiting distribution of this difference. We note the convergence in distribution d. Theorem 1. Under the null or a sequence of local alternatives, the paired Hotelling test and the score test of CLR are asymptotically equivalent. More precisely, under Assumption 1, we have: d n(ξsc,n − ξhot,n ) → − K 6 
where K = (δ + V )> Σ−1 Σ − (δ + V )(δ + V )> Σ−1 (δ + V )> with V ∼ N (0, Σ) and N refers to the normal distribution. Examples of distribution of K for p = 1, Σ = σ 2 = 1 and different values of δ are shown in figure 1. We see that for small values of δ, K is concentrated between 0 and 1/4. As δ grows, the distribution shifts to large, negative values. For the practitioner, the result of Theorem 1 will not affect the choice of test but ensures that similar conclusions are reached regardless of that choice. The result also implies that CLR can be considered a generalization of the paired Hotelling or Student test when the conditions for the latter do not apply. Proof of Theorem 1 To facilitate the notation, we omit the dependence in n in most quantities, writing Z instead of Z (n) . The quantity we want to estimate is: n(ξsc,n − ξhot,n ) = n2 Z̄ > (Ĩ −1 − C −1 )Z̄ = n2 Z̄ > Ĩ −1 (C − Ĩ)C −1 Z̄, (4) P where Ĩ = n1 ni=1 Zi Zi> . The Lindeberg-Feller Central Limit Theorem for triangular arrays (Hall and Heyde, 1980, (n) Theorem 3.2 p58) applied on the Wi gives: √ d nZ̄ → − (δ + V ), (5) with V ∼ N (0, Σ). The weak law of large numbers for triangular arrays implies: P C− → Σ. (6) 1X 1X P Ĩ = Zi Zi> = (Zi − Z̄)(Zi − Z̄)> + Z̄ Z̄ > − →Σ n i=1 n i=1 (7) Combining (5) and (6), we see: n n and, using the second identity in (7), we obtain: d n(C − Ĩ) = C − nZ̄ Z̄ > → − Σ − (δ + V )(δ + V )> . Applying Slutsky’s lemma to all of the above, we conclude: d n(ξsc,n − ξhot,n ) → − K
where K = (δ + V )> Σ−1 Σ − (δ + V )(δ + V )> Σ−1 (δ + V )> with V ∼ N (0, Σ). This concludes the proof of Theorem 1. Acknowledgements The author thanks Gérard Biau for his supervision and his many comments on the manuscript. Funding: The author acknowledges a PhD grant from École Normale Supérieure. 7 Delta=1 Density 3 6 0 0 1 2 2 4 Density 4 8 5 6 10 Delta=0 −0.8 −0.6 −0.4 −0.2 0.0 0.2 −1.0 −0.8 −0.6 −0.4 −0.2 K K Delta=2 Delta=3 0.0 0.2 0.0 0.2 0.0 0.00 0.2 0.05 0.10 Density 0.6 0.4 Density 0.8 1.0 0.15 1.2 −1.0 −1.0 −0.8 −0.6 −0.4 −0.2 0.0 0.2 −1.0 K −0.8 −0.6 −0.4 −0.2 K Figure 1: Distribution of K. Distribution of the limit variable for p = 1, Σ = σ 2 = 1 and δ ∈ {0, 1, 2, 3}. 8 References References P. Rosenbaum. Observational Studies. Springer, New York, 2002. D. Rubin. Matching to remove bias in observational studies. Biometrics, 29:159–183, 1973. P. Rosenbaum and D. Rubin. The central role of the propensity score in observational studies for causal effects. Biometrika, 70:41–55, 1983. N.E. Breslow and N.E. Day. Statistical methods in cancer research. The analysis of casecontrol studies. International Agency for Research on Cancer, Lyon, 1981. Q. McNemar. Note on the sampling error of the difference between correlated proportions or percentages. Psychometrika, 12:153–157, 1947. F. Balazard, S. Le Fur, S. Valtat, Isis Diab collaborative group, A.-J. Valleron, and P. Bougnères. Association of environmental markers with childhood type 1 diabetes mellitus revealed by a long questionnaire on early life exposures and lifestyle in a case-control study. BMC Public Health, 2016. N. Mantel and W. Haenszel. Statistical aspects of the analysis of data from retrospective studies. Journal of the national cancer institute, 22:719–748, 1959. J.H. McDonald. Handbook of Biological Statistics. Sparky House Publishing, Baltimore, 2009. N.E. Breslow, N.E. Day, K.T. Halvorsen, R.L. Prentice, and C. Sabai. Estimation of multiple relative risk functions in matched case-control studies. American Journal of Epidemiology, 108:299–307, 1978. R.F. Engle. Wald, likelihood ratio, and lagrange multiplier tests in econometrics. Handbook of Econometrics, 2:775–826, 1984. A. Agresti and M. Kateri. Categorical Data Analysis. Springer, Berlin, Heidelberg, 2011. N.E. Day and D.P. Byar. Testing hypotheses in case-control studies-equivalence of mantelhaenszel statistics and logit score tests. Biometrics, 35:623–630, 1979. P. Hall and C.C. Heyde. Martingale limit theory and its application. Academic Press., New York, 1980. 9 

10math.ST

Flexible and Creative Chinese Poetry Generation Using Neural Memory Jiyuan Zhang1,2 , Yang Feng3,1,4 , Dong Wang1∗, Yang Wang1 , Andrew Abel6 , Shiyue Zhang1,5 , Andi Zhang1,5 1 Center for Speech and Language Technologies(CSLT), RIIT, Tsinghua University, China 2 Shool of Software & Microelectronics, Peking University, China 3 Key Laboratory of Intelligent Information Processing,Institute of Computing Technology,CAS 4 Huilan Limited, Beijing, China 5 Beijing University of Posts and Telecommunications, China 6 Xi’an Jiaotong-Liverpool University, China zhangjy ml@pku.edu.cn, wangdong99@mails.tsinghua.edu.cn arXiv:1705.03773v1 [cs.AI] 10 May 2017 Abstract It has been shown that Chinese poems can be successfully generated by sequence-to-sequence neural models, particularly with the attention mechanism. A potential problem of this approach, however, is that neural models can only learn abstract rules, while poem generation is a highly creative process that involves not only rules but also innovations for which pure statistical models are not appropriate in principle. This work proposes a memory-augmented neural model for Chinese poem generation, where the neural model and the augmented memory work together to balance the requirements of linguistic accordance and aesthetic innovation, leading to innovative generations that are still rule-compliant. In addition, it is found that the memory mechanism provides interesting flexibility that can be used to generate poems with different styles. 1 Introduction Classical Chinese poetry is a special cultural heritage with over 2,000 years of history and is still fascinating us today. Among the various genres, perhaps the most popular one is the quatrain, a special style with a strict structure (four lines with five or seven characters per line), a regulated rhythmical form (the last characters in the second and fourth lines must follow the same rhythm), and a required tonal pattern (tones of characters in some positions should satisfy a predefined regulation) (Wang, 2002). This genre flourished mostly in the Tang Dynasty, and so are often called ‘Tang 1 Corresponding author: Dong Wang; RM 1-303, FIT BLDG, Tsinghua University, Beijing (100084), P.R. China. poems’. An example of a quatrain written by Wei Wang, a famous poet in the Tang Dynasty, is shown in Table 1. Due to the stringent restrictions in both rhythm and tone, it is not trivial to create a fully rule-compliant quatrain. More importantly, besides such strict regulations, a good quatrain should also read fluently, hold a consistent theme, and express a unique affection. Therefore, poem generation is widely recognized as a very intelligent activity and can be performed only by knowledgeable people with a lot of training. 乐游原 Climbing the Paradise Mound 向晚意不适， (* Z Z P Z) As I was not in a good mood this evening round, 驱车登古原。 (P P P Z P) I went by cart to climb the Ancient Paradise Mound. 夕阳无限好， (* P P Z Z) It is now nearing dusk, 只是近黄昏。 (* Z Z P P) When the setting sun is infinitely fine, which is a must. Table 1: An example of a 5-char quatrain. The tonal pattern is shown at the end of each line, where ’P’ indicates a level tone, ’Z’ indicates a downward tone, and ’*’ indicates the tone can be either. The translation is from (Tang, 2005). In this paper we are interested in machine poetry generation. Several approaches have been studied by researchers. For example, rule-based methods (Zhou et al., 2010), statistical machine translation (SMT) models (Jiang and Zhou, 2008; He et al., 2012) and neural models (Zhang and Lapata, 2014; Wang et al., 2016a,c). Compared to previous approaches (e.g., rule-based or SMT), the neural model approach tends to generate more fluent poems and some generations are so natural that even professional poets can not tell they are the work of machines (Wang et al., 2016a). In spite of these promising results, neural models suffer from a particular problem in poem generation, a lack of innovation. Due to the statistical nature of neural models, they pay much more attention to high-frequency patterns, whereas they ignore low-frequency ones. In other words, the more regular and common the patterns, the better the neural model is good at learning them and tends to use them more frequently at run-time. This property certainly helps to generate fluent sentences, but it is not always useful: the major value of poetry is not fluency, but the aesthetic innovation that can stimulate some unique feelings. This is particularly true for Chinese quatrains that are highly compact and expressive: it is nearly impossible to find two similar works in the thousands of years of history in this genre, demonstrating the importance of uniqueness or innovation. Ironically, the most important thing, innovation, is largely treated as trivial, if not noise, by present neural models. Actually this problem is shared by all generation models based on statistics (although it is more serious for neural models) and has aroused a long-standing criticism for machine poem generation: it can generate, and sometimes generate well, but the generation tends to be unsurprising and not particularly interesting. More seriously, this problem exists not only in poem generation, but also in all generation tasks that require innovation. This paper tries to solve this extremely challenging problem. We argue that the essential problem is that statistical models are good at learning general rules (usage of regular words and their combinations) but are less capable of remembering special instances that are difficult to cover with general rules. In other words, there is only rule-based reasoning, no instance-based memory. We therefore present a memory-augmented neural model which involves a neural memory so that special instances can be saved and referred to at run-time. This is like a human poet who creates poems by not only referring to common rules and patterns, but also recalls poems that he has read before. It is hard to say whether this combination of rules and instances produces true innovation (which often requires real-life motivation rather than simple word reordering), but it indeed offers interesting flexibility to generate new outputs that look creative and are still rule-compliant. Moreover, this flexibility can be used in other ways, e.g., generating poems with different styles. In this paper, we use the memory-augmented neural model to generate flexible and creative Chinese poems. We investigate three scenarios where adding a memory may contribute: the first scenario involves a well trained neural model where we aim to promote innovation by adding a memory, the second scenario involves an over-fitted neural model where we hope the memory can regularize the innovation, and in the third scenario, the memory is used to encourage generation of poems of different styles. 2 Related Work A multitude of methods have been proposed for automatic poem generation. The first approach is based on rules and/or templates. For example, phrase search (Tosa et al., 2009; Wu et al., 2009), word association norm (Netzer et al., 2009), template search (Oliveira, 2012), genetic search (Zhou et al., 2010), text summarization (Yan et al., 2013). Another approach involves various SMT methods, e.g., (Jiang and Zhou, 2008; He et al., 2012). A disadvantage shared by the above methods is that they are based on the surface forms of words or characters, having no deep understanding of the meaning of a poem. More recently, neural models have been the subject of much attention. A clear advantage of the neural-based methods is that they can ‘discover’ the meaning of words or characters, and can therefore more deeply understand the meaning of a poem. Here we only review studies on Chinese poetry generation that are mostly related to our research. The first study we have found in this direction is the work by Zhang and Lapata (2014), which proposed an RNN-based approach that produces each new line character-by-character using a recurrent neural network (RNN), with all the lines generated already (in the form of a vector) as a contextual input. This model can generate quatrains of reasonable quality. Wang et al. (2016b) proposed a much simpler neural model that treats a poem as an entire character sequence, and poem generation is conducted character-by-character. This approach can be easily extended to various genres such as Song Iambics. To avoid theme drift caused by this long-sequence generation, Wang et al. (2016b) utilized the neural attention mechanism (Bahdanau et al., 2014) by which human intention is encoded by an RNN to guide the generation. The same model was used by Wang et al. (2016a) for Chinese quatrain generation. Yan (2016) proposed a hierarchical RNN model that conducts iterative generation. Recently, Wang et al. (2016c) proposed a similar sequence generation model, but with the difference that attention is placed not only on the human input, but also on all the characters that have been generated so far. They also proposed a topic planning scheme to encourage a smooth and consistent theme. All the neural models mentioned above try to generate fluent and meaningful poems, but none of them consider innovation. The memory-augmented neural model proposed in this study intends to address this issue. Our system was built following the model structure and training strategy proposed by Wang et al. (2016a) due to its simplicity and demonstrated quality, but the memory mechanism is general and can be applied to any of the models presented above. The idea of memory argumentation was inspired by the recent advance in neural Turing machine (Graves et al., 2014, 2016) and memory network (Weston et al., 2014). These new models equip neural networks with an external memory that can be accessed and manipulated via some trainable operations. In comparison, the memory in our work plays a simple role of knowledge storage, and the only operation is simple pre-defined READ. In this sense, our model can be regarded as a simplified neural Turing machine that omits training. Figure 1: The memory-augmented neural model used for Chinese poetry generation. In this section, we first present the idea of memory augmentation, and then describe the model structure and training method. able. The memory component involves a set of ‘direct’ mappings from input to output, and therefore can be used to memorize some special cases of the generation that can not be represented by the neural model. For poem generation, the memory stores the information regarding which character should be generated in a particular context. The output from the two components are then integrated, leading to a consolidated output. There are several ways to understand the memory-augmented neural model. Firstly, it can be regarded as a way of combining reasoning (neural model) and knowledge (memory). Secondly, it can be regarded as a way of combining rule-based inference (neural model) and instance-based retrieval (memory). Thirdly, it can be regarded as a way of combining predictions from complementary systems, where the neural model is continuous and parameter-shared, while the memory is discrete and contains no parameter sharing. Finally, the memory can be regarded as an effective regularization that constrains and modifies the behavior of the neural model, resulting in generations with desired properties. Note that this memory-augmented neural model is inspired by and related to the memory network proposed by Weston et al.(2014) and Graves et al.(2016), but we more focus on an accompanying memory that plays the role of assistance and regularization. 3.1 3.2 3 Memory-augmented neural model Memory augmentation The idea of memory augmentation is illustrated in Fig. 1. It contains two components, the neural model component on the left, and the memory component on the right. In this work, the attention-based RNN generation model presented by (Wang et al., 2016a) is used as the neural model component, although any neural model is suit- Model structure Using the Chinese poetry generation model shown in Fig. 1 as an example, this section discusses the creation of a memory-augmented neural model. Firstly, the neural model part is an attentionbased sequence-to-sequence model (Bahdanau et al., 2014). The encoder is a bi-directional RNN (with GRU units) that converts the in- put topic words, denoted by the embeddings of the compositional characters (x1 , x2 , ..., xN ), into a sequence of hidden states (h1 , h2 , ..., hN ). The decoder then generates the whole quatrain character-by-character, denoted by the corresponding embeddings (y1 , y2 , ...). At each step t, the prediction for the state st is based on the last generation yt−1 , the previous status st−1 of the decoder, as well as all the hidden states (h1 , h2 , ...) of the encoder. Each hidden state hi contributes to the generation according to a relevance factor αt that measures the similarity between st−1 and hi . This is written as: st = fd (yt−1 , st−1 , N X αt,i hi ) i=1 where αt,i represents the contribution of hi to the present generation, and can be implemented as any function. The output of the model is a posterior probability over the whole set of characters, written by zt = σ(st W ) where W is the projection parameter. The memory consists of a set of elements {mi }K i=1 , where K is the size of the memory. Each element mi involves two parts, the source part mi (s), that encodes the context, i.e. when this element should be selected, and the target part mi (g), that encodes what should be output if this element is selected. In our study, the neural model is firstly trained, and then the memory is created by running fd (the decoder of the neural model). Specifically, for the k-th poem selected to be in the memory, the character sequence is input to the decoder one by one, with the contribution from the encoder set to zero. Denoting the starting position of this poem in the memory is pk , the status of the decoder at the j-th step is used as the source part of the (pk +j)-th element of the memory, and the embedding of the corresponding character, xj , is set to be the target part. this is formally written as: mi (s) = fd (xj−1 , sj−1 , 0) (1) st , and then the outputs of the selected elements are averaged as the output of the memory component. We choose cosine distance to measure the fitting degree, and have1 : vt = K X cos(st , mi (s))mi (g). (2) i=1 The output of the neural model and the memory can be combined in various ways. Here, a simple linear combination before the softmax is used, i.e., zt = σ(st W + βvt E) (3) where β is a pre-defined weighting factor, and E contains word embeddings of all the characters. Although it is possible to train β from the data, we found that the learned β is not better than the manually-selected one. This is probably because β is a factor to trade-off the contribution from the model and the memory, and how to make the trade-off should be a ‘prior knowledge’ rather than a tunable parameter. In fact, if it is trained, than it will be immediately adapted to match the training data, which will nullify our effort to encourage innovative generation. 3.3 Model Training In our implementation, only the neural model component is required to be trained. The training algorithm follows the scheme defined in (Wang et al., 2016a), where the cross entropy between the distributions over Chinese characters given by the decoder and the ground truth is used as the objective function. The optimization uses the SGD algorithm together with AdaDelta to adjust the learning rate (Zeiler, 2012). 4 Memory augmentation for Chinese poetry generation This section describes how the memory mechanism can be used to trade-off between the requirements for rule-compliant generation and aesthetic innovation, and how it can also be used to do more interesting things, for example style transfer. and mi (g) = xj where i = pk + j. 4.1 In this section, we describe how the memory mechanism promotes innovation. Monitoring the 1 At run-time, the memory elements are selected according to their fit to the present decoder status Memory for innovative generation In fact, we run a parallel decoder to provide st in Eq.(2). This decoder does not accept input from the encoder and so is consistent with the memory construction process as Eq.(1). training process for the attention-based model, we found that the cost on the training set will keep decreasing until approaching zero, but on the validation set, the degradation stops after only one iteration. This can be explained by the fact that Chinese quatrains are highly unique, so the common patterns can be fully learned in one iteration, resulting in overfitting with additional iterations. Due to the overfitting, we observe that with the one-iteration model, reasonable poems can be generated, and with the over-fitted model, the generated poems are meaningless, in that they do not resemble feasible character sequences. The energy model perspective helps to explain this difference. For the one-iteration model, the energy surface is smooth and the energy of the training data is not very low, as illustrated in plot (a) in Fig. 2, where the x-axis represents the input and y-axis represents the output, and the z-axis represents the energy. With this model, inputs with small variance will be attracted to the same low-energy area, leading to similar generations. These generations are trivial, but at least reasonable. If the model is overfitted, however, the energy at the locations of the training data becomes much lower than their surrounding areas, leading to a bumpy energy surface as shown in plot (b) in Fig. 2. With this model, inputs with a small variation may be attracted to very different low-energy areas, leading to significantly different generations. Since many of the low-energy areas are nothing to do with good generations but are simply caused by the complex energy function, the generations can be highly surprising for human readers, and the quality is not guaranteed. In some sense, these generations can be regarded as ‘innovative’ , but based on observations made in our experiments, most of them are meaningless. The augmented memory introduces a new energy function, which is combined with the energy function of the neural model to change the energy surface of the generation system. This can be seen in Eq. (3), where st W and βvt E can be regarded as the energy function of the neural model component and the memory component, respectively, and the energy function of the memory-augmented system is the sum of the energy functions of these two components. For this reason, the effect of the memory mechanism can be regarded as a regularization of the neural model that will adjust its generation behavior. This regularization effect is illustrated in Fig. 2, where the energy function of the memory shown in plot (c) is added to the energy function of the one-iteration model and the overfitted model, as shown in plot (e) and plot (f) respectively. It can be seen that with the memory involved, the energy surface becomes more bumpy with the one-iteration model, and more smooth with the overfitted model. In the former case, the effect of the memory is to encourage innovation, while still focusing on rule-compliance, and in the latter case, the effect is to encourage rule compliance, while keeping the capability for innovation. It is important to notice that the energy function of the memory component is a linear combination of the energy functions of the compositional elements (see Eq.(2)), each of which is convex and is minimized at the location represented by the element. This means that the energy surface of the memory is rather ‘healthy’, in the sense that low-energy locations mostly correspond to good generations. For this reason, the regularization provided by the memory is safe and helpful. 4.2 Memory for style transfer The effect of the memory is easy to control. For example, the complexity of the behavior can be controlled by the memory size, the featured bias can be controlled by memory selection, and the strength of the impact can be controlled by the weighting parameter β. This means that the memory mechanism is very flexible and can be used to produce poems with desired properties. In this work, we use these capabilities to generate poems with different styles. This has been illustrated in Fig. 2, where the energy function of the style memory shown in plot (d) is biased towards a particular style, and once it is added to energy function of the one-iteration model, the resulting energy function shown in plot (g) obtains lower values at locations corresponding to the locations of the memory, which encourages generation of poems with similar styles as those poems in the memory. 5 Experiments This section describes the experiments and results carried out in this paper. Here, The baseline system was a reproduction of the Attention-based system presented in (Wang et al., 2016a). the model in This system has been shown to be rather flexi- Figure 2: The energy surface for (a) one-iteration model (b) overfitted model (c) memory (d) style memory (e) one-iteration model augmented with memory (f) overfitted model augmented with memory (g) one-iteration model augmented with style memory. ble and powerful: it can generate different genres of Chinese poems, and when generating quatrains it has been shown to be able to fool human experts in many cases (Wang et al., 2016a) and the authors had did a thorough comparison with competitive methods mentioned in the related work of this paper. We obtained the database and the source code (in theano), and reproduced their system using Tensorflow from Google2 . We didn’t make comparisons with some previous methods such as NNLM, SMT, RNNPG as they had been fully compared in (Wang et al., 2016a) and all of them were much worse than the attention-based system. Another reason was that the experts were not happy to evaluate poems with clearly bad quality. We also reproduced the model in (Wang et al., 2016c) with the help of the first author. However, since their implementation did not involve any restrictions on rhythm and tone, the experts were reluctant to recognize them as good poems. With a larger dataset (e.g., 1 Million poems), it is assumed that the rhythm and tone can be learned and their system would be good in both fluency and rule compliance. It should be also emphasized that the memory approach proposed in this paper is a general technique and is complementary to other efforts such as the planning approach (Wang et al., 2016c) and the recursive approach (Yan, 2016). Based on the baseline system, we built the memory-augmented model, and conducted two 2 https://www.tensorflow.org/ experiments to demonstrate its power. The first is an innovation experiment which employs memory to promote or regularize the generation of innovative poems, and the second is a style-transfer experiment which employs memory to generate flexible poems in different styles. We invited 34 experts to participate in the experiments, and all of them have rich experience not only evaluating poems, but also in writing them. Most of the experts are from prestigious institutes, including Peking university and the Chinese Academy of Social Science (CASS). Following the suggestions of the experts, we use five metrics to evaluate the generation, as listed below: • Compliance: if regulations on tones and rhymes are satisfied; • Fluency: if the sentences read fluently and convey reasonable meaning; • Theme consistency: if the entire poem adheres to a single theme; • Aesthetic innovation: if the quatrain stimulates any aesthetic feeling with elaborate innovation; • Scenario consistency: if the scenario remains consistent. 5.1 Datasets The baseline system was built with two customized datasets. The first dataset is a Chinese poem corpus (CPC), which we used in this work to train the embeddings of Chinese characters. Our CPC dataset contains 284,899 traditional Chinese poems in various genres, including Tang quatrains, Song Iambics, Yuan Songs, and Ming and Qing poems. This large quantity of data ensures reliable learning for the semantic content of most Chinese characters. Our second dataset is a Chinese quatrain corpus (CQC) that we have collected from the internet, which consists of 13, 299 5-char quatrains and 65, 560 7-char quatrains. This corpus was used to train the attention-based RNN baseline. We filtered out the poems whose characters are all low-frequency (less than 100 counts in the database). After the filtering, the remaining corpus contains 9,195 5-char quatrains and 49,162 7-char quatrains. We used 9,000 5-char and 49,000 7-char quatrains to train the attention model, and the rest for validation. Another two datasets were created for use in the memory-augmented system. Our first dataset, MEM-I, contains 500 quatrains randomly selected from our CQC corpus. This dataset was used to produce the memory in the innovation experiment; the second dataset, MEM-S, contains 300 quatrains with clear styles, including 100 pastoral, 100 battlefield and 100 romantic quatrains. It was used to generate memory with different styles in the style-transfer experiment. All the datasets will be released online3 . 5.2 Evaluation Process We invited 34 experts to evaluate the quality of the poem generation. In the innovation experiment, the evaluation consisted of a comparison between different systems and configurations in terms of the five metrics. The innovation questions presented the expert with two poems, and asked them to judge which of the poems was better in terms of the five metrics; in the style-transfer experiment, the evaluation was performed by identifying the style of a generated poem. The evaluation was conducted online, with each questionnaire containing 11 questions focusing on innovation and 4 questions concerned with style-transfer. Each of the style-transfer questions presented the expert with a single poem and asked them to score it between 1 to 5, with a larger score being better, in terms of compliance, aesthetic innovation, 3 http://vivi.cslt.org scenario consistency, and fluency. They were also asked to specify the style of the poem. Using the poems generated by our systems, we generated many different questions of both types, and then created a number of online questionnaires that randomly selected from these questions. This meant that as discussed above, each questionnaire had 11 randomly selected innovation questions, and 4 randomly selected style transfer questions. Each question was only used once, meaning that it was not duplicated on multiple questionnaires, and so each questionnaire was different. Experts could choose to answer multiple questionnaires if they wished, as each one was different. From the 34 experts, we collected 69 completed questionnaires, which equals to 759 innovation questions and 276 style-transfer questions. 5.3 Innovation experiment This experiment focuses on the contribution of memory for innovative poem generation. We experimented with two configurations: one is with a one-iteration model (C1 ) and the other is with an overfitted model (C∞ ). The memory was generated from the 500 quatrains in MEM-I, and the weighting factor was defined empirically as 16 for C1 and 49 for C∞ . The topics of the generation were 160 keywords randomly selected from Shixuhanyinge (Liu, 1735). Given a pair of poems generated by two different configurations using the same topic, the experts were asked to choose which one they preferred. The evaluation is therefore pair-wised, and each pair of configurations contains at least 180 evaluations. The results are shown in Table 2, where the preference ratio for each pair of configurations was tested in terms of the 5 metrics. From the first row of Table 2, we observe that the experts have a clear preference for the poems generated by the C1 model, the one that can produce fluent yet uninteresting poems. In particular, the ‘aesthetic innovation’ score for C∞ is not better than C1 , which was different from what we expected. Informal offline discussions with the poetry experts found that the experts identified some innovative expression in the C∞ condition, but most of the them was regarded as being nonsense in the opinion of many of the experts. In comparison to sparking innovation, fluency and being meaningful is more important not only for non-expert readers, but also for professional po- C1 vs C∞ C1 vs C1 +Mem C∞ vs C∞ +Mem C1 vs C∞ +Mem Compliance Fluency 0.59:0.41 0.41:0.59 0.40:0.60 0.43:0.57 0.68:0.32 0.36:0.64 0.26:0.74 0.58:0.42 Preference Ratio Theme Consistency 0.70:0.30 0.37:0.63 0.32:0.68 0.59:0.41 Aesthetic Innovation 0.68:0.32 0.33:0.67 0.30:0.70 0.50:0.50 Scenario Consistency 0.69:0.31 0.43:0.57 0.36:0.64 0.59:0.41 Table 2: Preference ratios for systems with or without overfitting and with or without memory augmentation. ets. In other words, only meaningful innovation is regarded as innovation, and irrational innovation is simply treated as junk. From the second and third rows of Table 2, it can be seen that involving memory significantly improves both C1 and C∞ , particularly for C∞ . For C1 , the most substantial improvement is observed in terms of ‘Aesthetic innovation’, which is consistent with our argument that memory can help encourage innovation for this model. For C∞ , ‘Fluency’ seems to be the most improved metric. This is also consistent with our argument that involving memory constrains over-innovation for over-fitted models. The last row of Table 2 is an extra experiment that investigates if C∞ is regularized well enough after introducing the memory. It seems that with the regularization, the overfitting problem is largely solved, and the generation is nearly as fluent and consistent as the C1 condition. Interestingly, the score for aesthetic innovation is also significantly improved. Since the regularization is not supposed to boost innovation, this seems confusing at first glance (in comparison to the result on the same metric in the first row), but this is probably because the increased fluency and consistency makes the innovation more appreciated, therefore doubly confirming our argument that true innovation should be reasonable and meaningful. 5.4 Style-transfer experiment In the second experiment, the memory mechanism is used to generate poems in different styles. We chose three styles: pastoral, battlefield, and romantic. A style-specific memory, which we call style memory, was constructed for each style by the corresponding quatrains in the MEM-S dataset. The system with one-iteration model C1 was used as the baseline. Two sets of topics were used in the experiment, one is general and the other is style-biased. The experiments then investigate if the memory mechanism can produce a clear style if the topic is general, and can transfer to a different style if the topic is style-biased already. The experts were asked to specify the style from four options including the three defined above and a ‘unclear style’ option. In addition, the experts were asked to score the poems in terms of compliance, fluency, aesthetic innovation, and scenario consistency, which we can use to check if the style transfer impacts the quality of the poem generation. Note that we did not ask for the theme consistency to be scored in this experiment because the topic words were not presented to the experts, in order to prevent the topic affecting their judgment regarding the style. The score ranges from 1 to 5, with a larger score being better. Table 3 presents the results with the general topics. The numbers show the probabilities that the poems generated by a particular system were labeled as having various styles. Since the topics are unbiased in types, the generation of the baseline system is assumed to be with unclear styles. For other systems, the style of the generation is assumed to be the same as the style of their memories. The results in Table 3 clearly demonstrates these assumptions. The tendency that romantic poems are recognized as pastoral poems is a little surprising. Further analysis shows that experts tend to recognize romantic poems as pastoral poems only if there are any related symbols such as trees, mountain, river. These words are very general in Chinese quatrains. The indicator words of romantic poems such as skirt, rouge, and singing are not as popular and their indication power is not as strong, leading to less labeling of romantic poems, as shown in the results. Model C1 (Baseline) C1 + Pastoral Mem C1 + Battlefield Mem C1 + Romantic Mem Pastoral 0.09 0.94 0.05 0.17 Probability Battlefield Romantic 0.04 0.18 0.00 0.06 0.93 0.00 0.00 0.61 Unclear 0.69 0.00 0.02 0.22 Table 3: Probability that poems generated by each configuration with general topics are labeled as various styles. We also tested transferring from one style to another. This was achieved by generating poems with some style-biased topics, and then using a style memory to force the generation to change the style. Our experiments show that in 73% cases the style can be successfully transferred. Finally, the scores of the poems generated with and without the style memories are shown in Table 4, where the poems generated with both general and style-biased topics are accounted for. It can be seen that overall, the style transfer may degrade fluency a little. This is understandable, as enforcing a particular style has to break the optimal generation with the baseline, which is assumed to be good at generating fluent poems. Nevertheless the sacrifice is not significant. Method C1 C1 C1 C1 C1 Compliance Fluency 4.10 4.07 3.82 4.00 3.95 3.01 3.00 2.63 2.78 2.80 (baseline) + Pastoral Mem + Battlefield Mem + Romantic Mem + All Mem Aesthetic Innovation 2.53 3.07 2.60 2.59 2.74 Scenario Consistence 2.94 3.17 2.95 3.00 3.05 Table 4: Averaged scores for systems with or without style memory. 5.5 Examples Table 5 to Table 7 shows example poems generated by the system C1 , C1 +Mem and C1 +Style Mem where the style in this case is set to be romantic. The three poems were generated with the same, very general, topic (‘(oneself)’). More examples are given in the supporting material. 自从此意无心物， Nothing in my heart, 一日东风不可怜。 Spring wind is not a pity. 莫道人间何所在， Don’t ask where it is, 我今已有亦相传。 I’ve noticed that and tell others. 一山自有无人语， Nobody speaking in the mountain, 不是青云入水边。 Also no green cloud stepping into the river. 莫把春风吹落叶， Spring wind does not stir leaves, 花开绿树满江船。 But flowers blooming in trees and flying to boats. Table 6: Example poems generated by the C1 +Mem system. 花香粉脸胭脂染， Beautiful face addressed by rouge, 帘影鸳鸯绿嫩妆。 Mandarin duck outside the curtain. 翠袖红蕖春色冷， Green sleeves and red flowers in cold spring, 柳梢褪叶暗烟芳。 Willow leaves gone in fragrant mist. Table 7: Example poems generated by the C1 +Style Mem system where the style is romantic. ory can boost innovation from two opposite directions: either by encouraging creative generation for regularly-trained models, or by encouraging rule-compliance for overfitted models. Both strategies work well, although the former generated poetry that was preferred by experts in our experiments. Furthermore, we found that the memory can be used to modify the style of the generated poems in a flexible way. The experts we collaborated with feel that the present generation is comparable to today’s experienced amateur poets. Future work involves investigating a better memory selection scheme. Other regularization methods (e.g., norm or drop out) are also interesting and may alleviate the over-fitting problem. Acknowledgments Table 5: system. 6 Example poems generated by the C1 Conclusions In this paper, we proposed a memory mechanism to support innovative Chinese poem generation by neural models augmented with a memory. Experimental results demonstrated that mem- This paper was supported by the National Natural Science Foundation of China (NSFC) under the project NO.61371136, NO.61633013, NO.61472428. References Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473 . Alex Graves, Greg Wayne, and Ivo Danihelka. 2014. Neural turing machines. arXiv preprint arXiv:1410.5401 . Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka GrabskaBarwińska, Sergio Gómez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, et al. 2016. Hybrid computing using a neural network with dynamic external memory. Nature 538(7626):471–476. Jing He, Ming Zhou, and Long Jiang. 2012. Generating Chinese classical poems with statistical machine translation models. In Twenty-Sixth AAAI Conference on Artificial Intelligence. Long Jiang and Ming Zhou. 2008. Generating Chinese couplets using a statistical mt approach. In Proceedings of the 22nd International Conference on Computational Linguistics. Association for Computational Linguistics, volume 1, pages 377–384. Wenwei Liu. 1735. ShiXueHanYing. Yael Netzer, David Gabay, Yoav Goldberg, and Michael Elhadad. 2009. Gaiku: Generating haiku with word associations norms. In Proceedings of the Workshop on Computational Approaches to Linguistic Creativity. Association for Computational Linguistics, pages 32–39. H Oliveira. 2012. Poetryme: a versatile platform for poetry generation. In Proceedings of the ECAI 2012 Workshop on Computational Creativity, Concept Invention, and General Intelligence. Yihe Tang. 2005. English Translation for Tang Poems (Ying Yi Tang Shi San Bai Shou). Tianjin People Publisher. Naoko Tosa, Hideto Obara, and Michihiko Minoh. 2009. Hitch haiku: An interactive supporting system for composing haiku poem. Entertainment Computing-ICEC 2008 pages 209–216. Springer. Li Wang. 2002. A Summary of Rhyming Constraints of Chinese Poems (Shi Ci Ge Lv Gai Yao), volume 1. Beijin Press. Qixin Wang, Tianyi Luo, and Dong Wang. 2016a. Can machine generate traditional Chinese poetry? a feigenbaum test. In BICS 2016. Qixin Wang, Tianyi Luo, Dong Wang, and Chao Xing. 2016b. Chinese song iambics generation with neural attention-based model. In IJCAI 16. Zhe Wang, Wei He, Hua Wu, Haiyang Wu, Wei Li, Haifeng Wang, and Enhong Chen. 2016c. Chinese poetry generation with planning based neural network. In COLING 2016. Jason Weston, Sumit Chopra, and Antoine Bordes. 2014. Memory networks. arXiv preprint arXiv:1410.3916 . Xiaofeng Wu, Naoko Tosa, and Ryohei Nakatsu. 2009. New hitch haiku: An interactive renku poem composition supporting tool applied for sightseeing navigation system. Entertainment Computing-ICEC 2009 pages 191–196. Springer. Rui Yan. 2016. i, Poet: Automatic poetry composition through recurrent neural networks with iterative polishing schema. In IJCAI2016. Rui Yan, Han Jiang, Mirella Lapata, Shou-De Lin, Xueqiang Lv, and Xiaoming Li. 2013. i, Poet: automatic Chinese poetry composition through a generative summarization framework under constrained optimization. In Proceedings of the Twenty-Third international joint conference on Artificial Intelligence. AAAI Press, pages 2197–2203. Matthew D Zeiler. 2012. Adadelta: an adaptive learning rate method. arXiv preprint arXiv:1212.5701 . Xingxing Zhang and Mirella Lapata. 2014. Chinese poetry generation with recurrent neural networks. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). pages 670–680. Cheng-Le Zhou, Wei You, and Xiaojun Ding. 2010. Genetic algorithm and its implementation of automatic generation of Chinese Songci. Journal of Software 21(3):427–437. 

2cs.AI

1 Software Engineers vs. Machine Learning Algorithms: An Empirical Study Assessing Performance and Reuse Tasks arXiv:1802.01096v2 [cs.SE] 7 Feb 2018 Nathalia Nascimento, Carlos Lucena, Paulo Alencar and Donald Cowan Abstract—Several papers have recently contained reports on applying machine learning (ML) to the automation of software engineering (SE) tasks, such as project management, modeling and development. However, there appear to be no approaches comparing how software engineers fare against machine-learning algorithms as applied to specific software development tasks. Such a comparison is essential to gain insight into which tasks are better performed by humans and which by machine learning and how cooperative work or human-in-the-loop processes can be implemented more effectively. In this paper, we present an empirical study that compares how software engineers and machine-learning algorithms perform and reuse tasks. The empirical study involves the synthesis of the control structure of an autonomous streetlight application. Our approach consists of four steps. First, we solved the problem using machine learning to determine specific performance and reuse tasks. Second, we asked software engineers with different domain knowledge levels to provide a solution to the same tasks. Third, we compared how software engineers fare against machine-learning algorithms when accomplishing the performance and reuse tasks based on criteria such as energy consumption and safety. Finally, we analyzed the results to understand which tasks are better performed by either humans or algorithms so that they can work together more effectively. Such an understanding and the resulting human-in-the-loop approaches, which take into account the strengths and weaknesses of humans and machine-learning algorithms, are fundamental not only to provide a basis for cooperative work in support of software engineering, but also, in other areas. Index Terms—Machine learning, human-in-the-loop, software engineer, automatic software engineering, internet of things, empirical study F 1 I NTRODUCTION S OFTWARE engineering processes can be very complex, costly and time-consuming [1]. They typically consist of a collection of related tasks [2] such as designing, implementing, maintaining, testing and reusing software applications [3]. In addition, as software has become embedded in systems of all kinds, millions of computer programs have to be corrected, adapted, and • • • Laboratory of Software Engineering (LES) at the Department of Informatics, Pontifical Catholic University of Rio de Janeiro, Brazil. CAPES scholarship/Program 194/Process: 88881.134630/2016-01 E-mail: nnascimento,lucena@inf.puc-rio.br See http://www.inf.puc-rio.br/∼nnascimento/ David R. Cheriton School of Computer Science. University of Waterloo, Canada. E-mail: palencar,dcowan@csg.uwaterloo.ca enhanced [2]. As a result, the field of software engineering requires millions of skilled Information Technology (IT) professionals to create millions of lines of code, which must be installed, configured, tuned, and maintained. According to Kephart (2005) [4], in the near future, it will be extremely challenging to manage IT environments, even for the most skilled IT professionals. Several researchers have proposed the use of artificial intelligence, especially machine-learning (ML) techniques, to automate different software engineering (SE) tasks [3], [5]–[20]. For example, Zhang has extensively studied this theme recently and in [3] he stated that: ”The field of software engineering turns out to be a fertile ground where many software development and maintenance 2 tasks could be formulated as learning problems and approached in terms of learning algorithms.” However, there is a lack of approaches to compare how software engineers fare against machine-learning algorithms for specific software development tasks. This comparison is critical in order to evaluate which S.E. tasks are better performed by automation and which require human involvement or human-in-the-loop approaches [21], [22]. In practice, because there are no explicit comparisons between the tasks performed by engineers and automated procedures, including machine learning, it is often not clear when to use automation in a specific setting. For example, a Brazilian company acquired a software system to select petroleum exploration models automatically, but the engineers decided they could provide a better solution manually. However, when there was a comparison of the manual solution with the one provided automatically by the system, it became clear that the automated solution was better. This illustrates that a lack of comparisons makes choosing a manual or an automated solution or a combined human-in-the-loop approach difficult. This paper, contains an empirical study [23] to compare how software engineers and machinelearning algorithms achieve performance and reuse tasks. The empirical study uses a case study involving the creation of a control structure for an autonomous streetlight application. The approach consists of four steps. First, the problem was solved using machine learning to achieve specific performance and reuse of tasks. Second, we asked software engineers with different domain-knowledge levels to provide a solution to achieve the same tasks. Third, we compared how software engineers compare with machinelearning algorithms when accomplishing the performance and reuse tasks based on criteria such as energy consumption and safety. Finally, the results were analyzed to understand which tasks are better performed by either humans or algorithms so that they can work together more effectively. Such an understanding is essential in realizing novel human-in-the-loop approaches in which machine-learning procedures assist software developers in achieving tasks. Such humanin-the-loop approaches, which take into account the strengths and weaknesses of humans and machine-learning algorithms, are fundamental not only to provide a basis for cooperative work in software engineering, but also in other application areas. This paper is organized as follows: Section 2 presents the empirical study describing research questions, hypotheses and the objective of the study. Section 3 presents the method selected to collect our empirical data. Sections 4 and 5 present the experimental results. Section 6 presents the threats to the validity of our experiment. Section 7 presents the related work. The paper ends with concluding remarks and suggestions for future work. 1.1 Motivation The theme of this paper, namely whether artificial intelligence such as machine learning, can benefit software engineering, has been investigated since 1986, when Hebert A Simon published a paper entitled “Whether software engineering needs to be artificially intelligent” [24]. In this paper, Simon discussed “the roles that humans now play versus the roles that could be taken over by artificial intelligence in developing computer systems.” Notwithstanding, in 1993, Ian Sommerville raised the following question [25]: “What of the future - can Artificial Intelligence make a contribution to system engineering?” In this paper [25], Sommervile performed a literature review in applications of artificial intelligence to software engineering, and concluded that: “the contribution of AI will be in supporting...activities that are characterized by solutions to problems which are neither right nor wrong but which are more or less appropriate for a particular situation...For example, requirements specification and analysis which involves extensive consultation with domain experts and in project management.” Several papers have since investigated the use of Machine Learning (ML) [26] in solving different software engineering (SE) tasks [5]–[20], [27]– [103]. These investigations include approaches to: i) project management [27]–[49], dealing with problems related to cost, time, quality prediction, and resource management; ii) defect prediction [50]–[84]; iii) requirements management, focusing on problems of classifying or representing requirements [85]–[88], or generating requirements [89]; iv) software development, such as code generation [20], [68], [90]–[96], synthesis [97]–[101], and code evaluation [102], [103]. 3 Most of these papers present successful applications of machine learning in software engineering, showing that ML techniques can provide correct automatic solutions to some SE problems. However, very few papers discuss whether or not a domain expert could propose a manual solution more appropriate for the particular situation. “More appropriate”, means a solution that provides better performance or increases another quality that is important to a particular application scenario, such as user preference [104]. For example, in the medical and aviation engineering fields, trust [105] in a solution provided to the end-user is an important factor to consider for a solution to be more appropriate. However, although many authors [106]–[109] have been promoting the use of neural networks [110] in medicine, Abbas et al. [105] and Castelvecchi [111] are among the few authors who questioned: “what is the trustworthiness of a prediction made by an artificial neural network?” In other application scenarios, such as many of those related to the Internet of Things (IoT) [112], [113], numerous authors [93], [95], [101], [114] consider the reuse of a solution as an important quality. They agree that to achieve the goal of billions of things connected to the Internet over the next few years [112], it is necessary to find ways to reduce time to market. For example, it is desirable that the solution or parts of the solution to design autonomous streetlights [96] for a specific scenario could be reused to design streetlights for another scenario. In particular, the Internet of Things has considerably increased the number of approaches that propose the use of machine learning to automate software development [93]–[96], [101], [115]. None of this research contains a comparison of their results to experiments designed by IoT experts. For example, do Nascimento and Lucena [101], [116] developed a hybrid framework that uses learning-based and manual program synthesis for the Internet of Things (FIoT). They generated four instances of the framework [101], [108], [117], [118] and used learning techniques to synthesize the control structure automatically. These authors stated that the use of machine learning made feasible the development of these applications. However, they did not present any experiment without using learning techniques. In contrast, most of the solutions released for the Internet of Things, such as Apple’s HomeKit’s approach [119] and Samsung Smart Things, [120] consider a software developer synthesizing the control structure for each thing manually. 1.2 Objective In this context, we decided to ask the following question: “How do software engineers compare with machine-learning algorithms?” To explore this question, we selected the Internet of Things as our application domain and then, compared a solution provided by a skilled IoT professional with a solution provided by a learning algorithm with respect to performance and reuse tasks. In short, Figure 1 depicts the theory [121] that we investigate in this paper. According to the theory, the variables that we intend to isolate and measure are the performance and reusability achieved from three kinds of solutions: i) solutions provided by learning techniques; ii) solutions provided by software engineers with IoT skills; and iii) solutions provided by software engineers without IoT skills. To evaluate the relationship among these variables, we performed an empirical study, using FIoT [101]. As shown in Figure 1, we raised four research questions (RQx) to investigate our theory’s propositions (e.g hypotheses (H-RQx)). We present these questions and hypotheses in Section 2. To collect and analyze our empirical data, we performed a controlled experiment. To perform this experiment, we reproduced the problem of synthesizing the control structure of autonomous streetlights using neuroevolution (i.e. “a learning algorithm which uses genetic algorithms to train neural networks” [122]) presented in [118]. Then, we invited 14 software engineers to provide a solution for the same problem using the same architecture and environment. Lastly, we compared the solution provided by the learning algorithm against the solutions provided by the software engineers. In this application of autonomous streetlights, we are considering a “more appropriate” solution as one that presents a better performance in the main scenario [118] or can be satisfactorily reused in a new scenario, based on criteria such as minimal energy consumption and safety (that is, maximum visual comfort in illuminated areas). 4 Actor SE Tasks Technology Synthesize and reuse the control structure of autonomous things Machine Learning Techniques Software Engineers with IoT skills without IoT skills produce solutions better than (H-RQ1, H-RQ2) Software System Internet of Things produce solutions better than (H-RQ3, H-RQ4) Improve (H-RQ1, H-RQ3) Increase (H-RQ2, H-RQ4) Performance Reusability Fig. 1. Theory [121]: Machine Learning can create solutions more appropriate than software engineers in the context of the Internet of Things. 2 H OW DO SOFTWARE ENGINEERS COM WITH MACHINE - LEARNING ALGO RITHMS ? A N E MPIRICAL S TUDY A D DRESSING P ERFORMANCE AND R EUSE IN THE I OT D OMAIN RQ4. ...software engineers without IoT skills with respect to their re-usability? PARE The experimental goal, based on the GoalQuestion-Metric (GQM) method [123] is to use the Framework for the Internet of Things (FIoT) for the purpose of comparing the use of an automated approach against a manual approach when synthesizing the control of autonomous things with respect to their performance and reuse. For this purpose, we asked four research questions (RQs) and performed a controlled experiment [23] (section 3) to investigate them. 2.2 Hypotheses Each RQ is based on one or more hypotheses, which are described next. H - RQ1. • • H - RQ2. • 2.1 Questions In terms of synthesizing the control structure of autonomous things, how does the result from a machine learning-based solution differ from solutions provided by... RQ1. ...software engineers with IoT skills with respect to their performance? RQ2. ...software engineers with IoT skills with respect to their re-usability? RQ3. ...software engineers without IoT skills with respect to their performance? H0. An ML-based approach does not improve the performance of autonomous things compared to solutions provided by IoT expert software engineers. HA. An ML-based approach improves the performance of autonomous things compared to solutions provided by IoT expert software engineers. • H0. An ML-based approach does not increase the reuse of autonomous things compared to solutions provided by IoT expert software engineers. HA. An ML-based approach increases the reuse of autonomous things compared to solutions provided by IoT expert software engineers. H - RQ3. • H0. An ML-based approach does not improve the performance of autonomous things compared to solutions provided by software engineers without experience in IoT development. 5 • HA. An ML-based approach improves the performance of autonomous things compared to solutions provided by software engineers without experience in IoT development. TABLE 1 Implementing FIoT flexible points to synthesize streetlight controllers using a ML-based approach. FIoT Framework Controller H - RQ4. • • H0. An ML-based approach does not increase the reuse of autonomous things compared to solutions provided by software engineers without experience in IoT development. HA. An ML-based approach increases the reuse of autonomous things compared to solutions provided by software engineers without experience in IoT development. 2.3 The object of the study: The Framework for the Internet of Things (FIoT) The Framework for the Internet of Things (FIoT) [101] is a hybrid software framework that allows the developer to generate controller structures for autonomous things through learning or procedural algorithms. If a researcher develops an application using FIoT, the application will contain a Java software component already equipped with modules for detecting autonomous things in an environment, assigning a controller to a specific thing, creating software agents, collecting data from devices and supporting the communication structure among agents and devices. Some features are variable and may be selected/developed according to the application type, as follows: (i) a control module such as “if-else”, neural network or finite state machine; (2) an adaptive technique to synthesize the controller at design-time, such as reinforcement learning [124] or genetic algorithm; and (iii) an evaluation process to evaluate the behavior of autonomous things that are making decisions based on the controller. For example, Table 1 summarizes how the “Streetlight Control” application will adhere to the proposed framework using a machine learning-based approach, while extending the FIoT flexible points. Table 2 summarizes how the “Streetlight Control” application will adhere to the proposed framework using a solution provided by a software engineer, while extending the FIoT flexible points. Our goal is to provide both solutions to the same application and compare the results based on the same evaluation process. Making Evaluation Controller Adaptation at design time Light Control Application Three Layer Neural Network Collective Fitness Evaluation: the solution is evaluated based on the energy consumption, the number of people that finished their routes after the simulation ends, and the total time spent by people to move during their trip Evolutionary Algorithm: Generate a pool of candidates to represent the neural network parameters TABLE 2 Implementing FIoT flexible points to synthesize streetlight controllers using a solution provided by a Software Engineer FIoT Framework Controller Making Evaluation Controller Adaptation at design time 3 Light Control Application if-else module provided by a software engineer Collective Fitness Evaluation: the solution is evaluated based on the energy consumption, the number of people that finished their routes after the simulation ends, and the total time spent by people to move during their trip None C ONTROLLED E XPERIMENT The first step of the experiment was to reproduce the experiment presented in [118] by using a not supervised learning approach. Then, we invited 14 software engineers to provide a solution for the same problem. Finally, we compared the solution provided through the learning algorithm against solutions provided by the participants. 3.1 Participant Analysis As we have described previously, the knowledge in the application domain is an important variable in our empirical study. Therefore, before performing the controlled experiment, we asked participants to describe their experience with the development of applications based on the Internet of Things, that is, developing distributed systems with embedded characteristics, such as providing each element of the system with sensors and actuators. As shown in Figure 2, 43% of 6 Medium High 14% Low 22% 21% 43% None Fig. 2. Experience of participants in developing applications based on the Internet of Things. participants have never developed an application based on the Internet of Things and 57% have developed at least one application. 3.2 Experiment: Streetlight Application In short, our experiment involves developing autonomous streetlights. The overall goal of this application is to reduce the energy consumption while maintaining appropriate visibility in illuminated areas [118]. For this purpose, we provided each streetlight with ambient brightness and motion sensors, and an actuator to control light intensity. In addition, we also provided streetlights with wireless communicators as shown in Figure 3. Therefore, the streetlights are able to cooperate with each other to establish the most likely routes for passers-by and thus achieve the goal of minimizing energy consumption. Lighting (Dark/DIM/Light) Presence (Yes/No) Data collected from the closest street light (0.0/0.5/1.0) Previous Listening Decision (Yes/No) Each streetlight has to execute three tasks every second: data collection, decision-making and action enforcement. The first task consists of receiving data related to people flow, ambient brightness, data from the neighboring streetlights and current light status (activation level of sensors and the previous output value of listeningDecision). The second task consists of analyzing collected data and making decisions about actions to be enforced. The last task is the action enforcement, which consists of setting the value of three output variables: (i) listeningDecision, that enables the streetlight to receive signals from neighboring streetlights in the next cycle; (ii) wirelessTransmitter, a signal value to be transmitted to neighboring streetlights; and (iii) lightDecision, that activates the light’s OFF/DIM/ON functions. The interested reader may consult a more extensive paper about the application scenario [118] 1 . 3.2.1 The Challenge As we explained to the participants, the tasks of collecting data and enforcing actions have already been implemented. The challenge was to provide a solution for the task of making decisions, as depicted in Figure 4. Light Decision (OFF/DIM/ON) Wireless Transmitter (0.0/0.5/1.0) Listening Decision (Yes/No) Fig. 3. Variables collected and set by streetlights. Each streetlight in the simulation has a microcontroller that is used to detect the proximity of a person, and control the closest streetlight. A streetlight can change the status of its light to ON, OFF or DIM. Fig. 4. The challenge: how does a streetlight make decisions based on collected data? We provided pseudocode that considered all possible combinations of input variables. Then, participants decided how to set output variables according to the collected data. Part 2 of this pseudocode is depicted in Figure 5. Each participant provided a different solution. Therefore, we conducted the experiment by 1. All documents that we prepared to explain this application scenario to participants are available at http://www.inf.puc-rio.br/.̃nnascimento/projects.html 2. The pseudocode that we provided to participants is available at: http://www.inf.puc-rio.br/.̃nnascimento/projects.html 7 Fig. 5. Small portion of the pseudocode of the decision module that was filled by participants. if (lighting_sensor = Medium AND detected_person = NO AND data_collected = 0.0 AND previous_listening_decision = YES) then { light_decision = __X_OFF____DIM____ON wireless_transmitter = __X_0.0 ___0.5____1.0 listening_decision = _X__YES ____NO } if (lighting_sensor = Medium AND detected_person = YES AND data_collected = 0.0 AND previous_listening_decision = YES) then { light_decision = ___OFF__X__DIM____ON wireless_transmitter = _X__0.0 ___0.5____1.0 listening_decision = __X_YES ____NO } if (lighting_sensor = Dark AND detected_person = NO AND data_collected = 0.0 AND previous_listening_decision = NO) then { light_decision = __X_OFF____DIM____ON wireless_transmitter = ___0.0 __X_0.5____1.0 listening_decision = ___YES __X__NO } Fig. 6. Small portion of the rule decisions that was synthesized according to the learning-based approach. using each one. In addition, we also considered a “zeroed” solution, which always sets all values to zero. This zeroed solution is supposed to be the worst solution, since streetlights will always switch their lights to OFF. 3.2.2 The solution generated by a machinelearning algorithm We compared the results from all of these approaches to the result produced using the machine learning approach. As do Nascimento and Lucena explain in [118], the learning approach uses a three-layer feedforward neural network combined with an evolutionary algorithm to generate decision rules automatically. Figure 6 depicts some of the rules that were generated by the evolved neural network. The interested reader can consult more extensive papers [101], [118] or read Nascimento’s dissertation [116] (chap. ii, sec. iii). Based on the generated rules and the system execution, we observe that using the solution provided by the neural network, only the street- lights with broken lamps emit “0.5” from their wireless transmitters. In addition, we also observed that a streetlight that is not broken switches its lamp ON if it detects a persons proximity or receives “0.5” from a wireless transmitter. 3.2.3 Scenario constraints Before starting a solution, each participant should consider the following constraints: • • • Do not take light numbering into account, since your solution may be used in different scenarios (see an example of a scenario in Figure 7). Three streetlights will go dark during the simulation. People walk along different paths starting at random departure points. Their role is to complete their routes by reaching a destination point. The number of people that finished their routes after the simulation ends, and the total time spent by people 8 • • • • moving during their trip are the most important factors for a good solution. A person can only move if his current and next positions are not completely dark. In addition, we also consider that people walk slowly if the place is partially devoid of light. The energy consumption also influences the solution evaluation. The energy consumption is proportional to the light status (OFF/DIM/ON). We also consider the use of the wireless transmitter to calculate energy consumption (if the streetlight emits something different from “0.0”, it consumes 0.1 of energy). Therefore, each solution is evaluated after the simulation ends based on the energy consumption, the number of people that finished their routes after the simulation ends, and the total time spent by people moving during their trip. Figure 7 depicts the elements that are part of the application namely, streetlights, people, nodes and edges. Execution: 12 seconds - A person moves from one point to another in one second or a second and a half. - Street lights execute cycles of 1 second ON DIM OFF Broken Lamps departure points target points Fig. 7. Simulated Neighborhood. Nascimento and Lucena [118] modeled the scenario as a graph, in which a node represents a streetlight position and an edge represents the smallest distance between two streetlights. (completedP eople × 100) The graph representing the streetlight network pP eople = (1) consists of 18 nodes and 34 edges. Each node totalP eople represents a streetlight. In the graph, the yellow, (totalEnergy × 100) pEnergy = 11×(timeSimulation×totalSmartLights) gray, black and red triangles represent the street) light status (ON/DIM/OFF/Broken Lamp). Each ( 10 (2) edge is two-way and links two nodes. In addi(totalT imeT rip × 100) tion, each edge has a light intensity parameter pT rip = 3×timeSimulation (( ) × totalP eople) that is the sum of the environmental light and the (2) (3) brightness from the streetlights in its nodes. Their f itness = (1.0 × pP eople) − (0.6 × pT rip)− goal is to simulate different lighting in different (0.4 × pEnergy) neighborhood areas. (4) Equations (1) through (4) show the values to be calculated for the evaluation in which pP eople is the percentage of the number of people that completed their routes before the end of the simulation out of the total number of people in the simulation; pEnergy is the percentage of energy that was consumed by streetlights out of the maximum energy value that could be consumed during the simulation. We also considered the use of the wireless transmitter to calculate energy consumption; pT rip is the percentage of the total duration time of peoples trips out of the maximum time value that their trip could spend; and f itness is the fitness of each candidate that encodes the proposed solution. 3.2.4 Example - Simulating the environment We showed participants the same simulated neighborhood scenario that was used by the genetic algorithm to evolve the neural network. 1 second/1 second and a half Fig. 8. Person moving in the simulated Neighborhood. As depicted in Figure 8, only one person was started in the scenario that we showed to participants. For instance, the person starting at point 0 has point 11 as a target. We ask participants to provide a solution to streetlights to assure that 9 this person will conclude his route before the simulation ends after 12 seconds. 3.2.5 New Scenario: Unknown environment The second step of the experiment consists of executing solutions from participants and the learning approach in a new scenario, but with the same constraints. This scenario, that is depicted in Figure 9 was not used by the learning algorithm and was not presented to participants. The goal of this new part of the experiment is to verify if the decision module that was designed to control streetlights in the first scenario can be reused in another scenario. the solution provided by each one of the 14 participants 3 . To provide a controlled experiment and be able to compare the different solutions, we started with only one person in the scenario and manually we set the parameters that were supposed to be randomly selected, such as departure and target points and broken lamps. Each experiment execution consists of executing the simulated scenario three times: (i) night (environmental light is equal to 0.0); (ii) late afternoon (environmental light is equal to 0.5); and (iii) morning (environmental light is equal to 1.0). The main idea is to determine how the solution behaves during different parts of the day. Figure 10 depicts the percentage of energy that was spent according to the environmental light for each one of the 16 different solutions. As we described previously, we also considered the use of the wireless transmitter to calculate energy consumption. As expected, as streetlights using the zeroed decision never switch their lights ON and never emit any signal, the energy consumed using this solution is always zero. It is possible to observe that only the solutions provided by the learning algorithm and by the 5th and 11th participants do not expend energy when the environmental light is maximum. In fact, according to the proposed scenario, there is no reason to turn ON streetlights during the period of the day with maximum illumination. 100 90 . Fig. 9. Simulating a new neighborhood. 80 0 70 60 ( 50 .)1 In this new scenario, we also only started one person, who has the point 18 (yellow point) as departure and the point 8 as target. As the scenario is larger, we established a simulation time of 30 seconds. 40 30 20 10 0 . 4 E XPERIMENT - PART 1 - R ESULTS We executed the experiment 16 times, only changing the decision solution of the autonomous streetlights. In the first instance, we set all outputs to zero (the zeroed solution) during the whole simulation, which is supposed to be the worst solution. For example, streetlights never switch their lights ON. In the second instance, we executed the experiment using the best solution that was found by the learning algorithm, according to the experiment presented in [118]. Then, we executed the simulation for .)1 ).5 .)1 ).5 % .)1 ).5 Fig. 10. Scenario1: Percentage of energy spent in different parts of the day according to the participant solutions. Figure 11 depicts the percentage of time that was spent by the unique person in each one of the simulations. As shown, the higher difference between solutions occurs at night. If the time is 3. All files that were generated during the development of this work, such as executable files and participants’ solutions results, are available at http://www.inf.puc-rio.br/.̃nnascimento/projects.html 10 100%, it means that the person did not complete the route, thus the solution did not work 4.1 Discussion: Participants Knowledge in IoT Versus Results 100 90 80 ) ( ). 70 60 40 simulation time did not allow the person to finish the route. After executing the solution proposed by each participant, we connect that solution’s results with the participant’s knowledge in the IoT domain, as shown in Table 3. 50 30 20 10 0 TABLE 3 Correlation between participants expertises in the Internet of Things with their solution results. 1) 0 ). ( ). ( % ). ( Fig. 11. Scenario1: Percentage of time spent by person to conclude his route based on different parts of the day according to the participant solutions. Besides presenting the results of the different solutions in different parts of the day, the best solution must be the one that presents the best result for the whole day. Thus, we calculated the average of each one of the parameters (energy, people, trip and fitness) that was achieved by solutions in different parts of the day. Figure 12 depicts a common average. We also calculated a weighted average, taking into account the duration of the parts of the day (we considered 12 hours for the night period, 3h for dim and 9h for the morning), but the results were very similar. 100 Software Engineer 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Learning zeroed Experience with IoT Development (None/Low/ Medium/ High) High None High Low None Low Medium None High None None Medium None Low Solution Performance (Fitness Average) Does the solution work? 55.48 26.99 62.88 62.49 30.50 51.09 54.37 16.59 28.62 61.60 29.67 47.81 30.32 56.91 59.53 28.33 Y N Y Y N Y Y N N Y N Y N Y Y N 90 80 70 % 60 62.88 55.48 61.60 62.49 51.09 56.91 59.53 54.37 47.81 50 40 30 20 26.99 30.50 28.62 29.67 30.32 28.33 16.59 10 0 Fitness We observe a significant difference between results from software engineers with any experience in IoT development and results from software engineers without experience in IoT development. Participant 10 is the only individual without knowledge of IoT that provided a solution that works and participant 9 is the only individual with any knowledge of IoT that did not provide a working solution. Fig. 12. Scenario1: Average of energy, trip and fitness calculated for the different parts of the day according to the participant solutions. 4.2 As shown in Figure 12, based on the fitness average, three participants namely 3, 4 and 10 provided a solution slightly better than the solution provided by the learning algorithm. Five other participants provided a solution that works and the remaining six provided a solution that does not work. As explained earlier, we have been considering an incorrect solution as one in which the person did not finish the route before the simulation ends. Even increasing the In this section, we investigate the hypotheses related to the solutions’ performance evaluation (i.e H-RQ1 and H-RQ3), as presented in subsection 2.2. Thus, we performed statistical analyses, as described by Peck and Devore [125], of the measures presented in Table 3. As shown in Table 4, we separated the results of the experiments into two groups: i) software engineers with IoT knowledge and ii) software engineers without IoT knowledge. Then, we calculated the mean and the standard deviation of Hypothesis Testing 11 the results achieved by each group performing the experiment and compare each result against the value achieved using the ML-based solution. 4.2.1 How does the evaluation result from a machine learning-based solution differ from solutions provided by IoT expert software engineers with respect to their performance? H - RQ1. • • H0. An ML-based approach does not improve the performance of autonomous things compared to solutions provided by IoT expert software engineers. HA. An ML-based approach improves the performance of autonomous things compared to solutions provided by IoT expert software engineers. The first null hypothesis H0 claims that there is no difference between the mean performance for IoT expert software engineers’ solutions and the ML-based approach solution. The alternative hypothesis claims that the ML-based approach solution improves the performance of the application in comparison to IoT expert software engineers’ solutions. Thus, the claim is that the true IoT expert software engineers’ solutions mean is below the performance achieved by the MLbased approach, that is = 59.53. Therefore, we used the ML performance as our hypothesizedvalue to test the following onesided hypothesis: H0: µs e = 59.53 H1: µs e < 59.53 where µs e denotes the true mean performance for all IoT expert software engineers’ solutions. For instance, we restricted the population sample to the number of software engineers that confirmed having experience with developing applications for the Internet of Things. As shown in Table 4, the performance mean (x) of the IoT expert software engineers’ solutions is 52.46 and the standard deviation (σ ) is 10.91. To verify if the data that we have is sufficient to accept the alternative hypothesis, we need to verify the probability of rejecting the null hypothesis [125]. Assuming that the H0 is true, and using a statistical significance level [125] of 0.01 (the chance of one in 100 of making an error), we computed the test statistic (t − statistic), as follows [125]: t − statistic : t = According to t-statistic theory, we can safely reject our null hypothesis if the t − statistic value is below the negative t − criticalvalue (threshold) [125]. This negative t − criticalvalue bounds the area of rejection of a T-distribution, as shown in Figure 13. In our experiment, as we specified a statistical significance level of 0.01, the probability of getting a T-value less or equal than the negative t − criticalvalue is 1%. We calculated the tcriticalvalue of this T-distribution according to the T-table presented in Peck and Devore (2011, pg 791) [125]. Accordingly, for a distribution with 7 degrees of freedom (see Table 4) and a confidence level of 99%, the negative tcriticalvalue is -3.00. As we depicted in Figure 13, the test statistic of our sample is higher than the tcriticalvalue. norm.php 850×350 pixels 12/11/17, 6:12 PM Fig. 13. Hypothesis H - RQ1 Test Graph. As the test statistic does not fall in the critical region, we cannot safely reject this null hypothesis. Based on a t-value of -1.83 and a degree of freedom of 7, we could reject our null hypothesis only if we had reduced the precision of our experiment to 85%. Thus, we would fail to reject the null hypothesis and would not accept the alternative hypothesis. Therefore, we cannot state that an ML-based approach improves the performance of autonomous things compared to solutions provided by IoT expert software engineers. http://www.imathas.com/stattools/norm.php Page 1 of 1 4.2.2 How does the evaluation result from a machine learning-based solution differ from solutions provided by software engineers without IoT skills with respect to their performance? H - RQ3. • (x − hypothesizedvalue) t − statistic : t = (5) ( √σn ) (52.46 − 59.53) = −1.83 (6) √ ) ( 10.91 8 H0. An ML-based approach does not improve the performance of autonomous things compared to solutions provided by 12 TABLE 4 Data to perform test statistic. Variable Software Engineers Software Engineers with IoT knowledge Software Engineers without IoT knowledge Machinelearning based approach • Standard deviation σ Degrees of freedom (n-1) t critical value (.99%) 49.45 16.00 13 2.65 52.46 54.92 10.91 7 3.00 32.61 30.00 15.15 5 3.37 n samples Highest value Mean x Median 14 62.88 43.95 8 62.88 6 61.60 1 59.53 software engineers without experience in IoT development. HA. An ML-based approach improves the performance of autonomous things compared to solutions provided by software engineers without experience in IoT development. For instance, we restricted the population sample to the number of software engineers that confirmed not having experience with developing applications for the Internet of Things. As shown in Table 4, the performance mean (x) of the solutions from software engineers without experience in IoT development is 32.61 and the standard deviation (σ ) is 15.15. Thus, the t − statistic : t = (32.61 − 59.53) = −4.35 (7) √ ) ( 15.15 6 As shown in Table 4, this T-distribution has 5 degrees of freedom. Thus, for a confidence level of 99%, the negative tcriticalvalue is -3.37. As we depicted in Figure 14, the test statistic of our sample is below the tcriticalvalue . norm.php 850×350 pixels Fig. 14. Hypothesis H - RQ3 Test Graph. 12/11/17, 6:14 PM As the t − statistic value is below the negative tcriticalvalue (-4.35 < −3.37), we can safely reject the null hypothesis, assuring that the error chancing of making an error is lower than 1%. Therefore, we accepted the alternative hypothesis: An ML-based approach improves the performance of autonomous things compared to solutions provided by software engineers with no experience in IoT development. 5 E XPERIMENT - PART 2 - R ESULTS As explained previously, the second part of the experiment consists of translating the solution provided by machine learning and participants to an unknown environment. In this second part of the experiment, we also executed the simulation 16 times: for each one of the participants’ solutions, for the machine-learning solution and for the zeroed solution. Table 5 shows the results that were achieved by the different solutions at night in a simulation of 30 seconds. As shown, most of the solutions did not work. The person in these simulations did not finish the route even when we increased the simulation time. Only the solution provided by the machine-learning algorithm and by participant 12 worked. Remember, this scenario was not used by the machine-learning algorithm during the training process. This solution was provided through machine learning for the first scenario and it was just reused in this new scenario. In other words, we did not restart the machinelearning process. We selected only those solutions that worked and verified their results for the other periods of 13 the day (morning and late afternoon). As shown in Table 6, when considering the whole day, the machine-learning approach presented the best result. Because the average time for the trip was a little higher using the machine-learning approach, the difference in energy consumption between the two solutions is considerably higher. of 14 engineers, only one participant, who has experience with IoT development, provided a solution that worked. Bad 93% TABLE 5 Using the same solution in a different environment - only at night. Software Engineer 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Learning zeroed Energy% People% Trip% Fitness 6.50 2.77 6.62 4.30 2.58 6.88 8.33 2.33 3.77 3.78 11.36 50.56 2.77 4.50 24.44 0 0 0 0 0 0 0 0 0 0 0 0 100 0 0 100 0 100 100 100 100 100 100 100 100 100 100 100 42.22 100 100 61.11 100 -42.60 -41.11 -42.65 -41.72 -41.03 -42.75 -43.33 -60.26 -41.51 -60.18 -44.54 54.43 -41.11 -41.80 53.55 -40 TABLE 6 Using the same solution in a different environment - day average. Average Participant 12 Average Learning 5.1 Energy% People% Trip% Fitness 50.52 100 38.14 56.90 8.46 100 46.29 68.83 0% Work better than ML 7% Work Fig. 15. Participants’ solution results in the second scenario. Therefore, we can safely reject the null hypothesis and accept both alternative hypotheses: 1) 2) 6 H-RQ2: H1: An ML-based approach increases the reuse of autonomous things compared to solutions provided by IoT expert software engineers. H-RQ4: H1: 2) An ML-based approach increases the reuse of autonomous things compared to solutions provided by software engineers with no experience with IoT development. D ISCUSSION In this section, we analyze the empirical experimental results to understand which tasks are better performed by humans and which by algorithms. This is important for selecting whether software engineers or machine learning can accomplish a specific task better. In our empirical study, in which we have assessed performance and reuse tasks, we accepted three alternative hypotheses and rejected one: Hypothesis Testing In this section, we investigate the hypotheses related to the solutions’ reuse evaluation, that is HRQ2 and H-RQ4, as presented in subsection 2.2. Their alternative hypotheses state that an MLbased approach improves the performance of autonomous things compared to solutions provided by software engineers, software engineers with experience in IoT development, and software engineers without experience in IoT development, respectively. We planned to perform a statistical development to evaluate these hypotheses. However, as depicted in Figure 15, in the new scenario, 0% of participants provided a result better than the result provided by the machinelearning solution. In addition, from the group Accepted: 1) 2) 3) An ML-based approach improves the performance of autonomous things compared to solutions provided by software engineers without experience with IoT development. An ML-based approach increases the reuse of autonomous things compared to solutions provided by IoT expert software engineers. An ML-based approach increases the reuse of autonomous things compared to solutions provided by software engineers without experience with IoT development. 14 Based on these results, we have found evidence that the use of machine-learning techniques can perform some SE tasks better than software engineers, considering solutions that improve performance and increase reuse. As illustrated in the experimental results, only one of the 14 software engineers provided a solution that could be reused in a new scenario. Further none of those software engineers provided a solution that works better than the ML’s solution in this new scenario. If the flexibility of the application is the most important factor, based on our results, we can safely recommend the use of machine learning. However, if we had considered performance as the only important factor to evaluate the quality of these solutions, we have found evidence that software engineers can perform SE tasks better than machine learning, considering “better” as a solution that improves performance. As described in our experiments, we cannot state that ML improves the performance of an application in comparison to solutions provided by domain expert software engineers. This is also an interesting result as many researchers, especially in the IoT domain, have strictly focused on automating software development. In brief, our experiment indicates that in some cases, software engineers outperform machinelearning algorithms, whereas in other cases, they do not. The evidence shows that it is important to know which one performs better in different situations in order to determine ways for software engineers to work cooperatively and effectively with automated machine-learning procedures. (2016) [127] report the diversity of participants as another possible threat. Therefore, in our study, we needed to be aware of at least two threats to validity namely: we have selected a sample of only 14 participants, which may not be enough to achieve conclusive results; and our sample consisted of only graduate students from two Brazilian universities. Such a group may not be representative of all software engineers, who may have substantially more professional experience and background. To mitigate the problems of the number of participants and their diversity, we selected our participants carefully. All of them have at least two years of experience with software development. In addition, we allowed participants to solve the problem by manipulating a pseudocode version, thereby avoiding gaps in the participants’ knowledge, such as experience with a particular programming language or architecture. Note that a survey was used to select participants and they all indicated a level of experience with pseudocode. The pseudocode provided by each participant was carefully translated into Java as this is the language supported by the Framework for the Internet of Things. Oizumi et al. (2017) reported a third threat to validity in [126], namely, possible misunderstandings during the study. To mitigate this problem of misunderstandings, we asked all participants to write about their understanding of the problem both before and after applying the solution. All participants indicated that they understood the task completely. We also asked them about their confidence in their proposed solution. Most of them evaluated their own solution with the highest grade, allowing us to increase our confidence in the experimental results. In addition, we assisted the participants during the entire study, making sure they understood the experimental task 7 8 Rejected: 1) An ML-based approach improves the performance of autonomous things compared to solutions provided by IoT expert software engineers. T HREATS TO VALIDITY Although we have designed and conducted the experiments carefully, there are always factors that can challenge the experiments validity. Some threats to validity as described in [123] could indeed limit the legitimacy of our results. In this section, we present the actions taken to mitigate their impact of these factors on the research results. As Oizumi et al. (2017) report in [126], the number of participants in the study can be a threat to validity. In addition, Fernandes et al R ELATED W ORK Comparing intelligent machines to the ability of a person to solve a particular problem is not a new approach. This kind of discussion has been promoted since the beginning of Artificial Intelligence. For example, in 1997, an important moment in the history of technology happened with Garry Kasparov’s 1997 chess match against the IBM supercomputer Deep Blue [128]. Recently, Silver et al. (2016, 2017) [129], [130] published a paper in the Nature Science Journal, comparing the performance of a ML technique 15 against the results achieved by the world champion in the game of Go. In [130], Silver et al. (2017) state that their program “achieved superhuman performance.” Whiteson et al. [122] indirectly performed this comparison, by evaluating the use of three different approaches of the neuroevolution learning algorithm to solve the same tasks: (i) coevolution, that is mostly unassisted by human knowledge; (ii) layered learning, that is highly assisted; and (iii) concurrent layered learning, that is a mixed approach. The authors state that their results “demonstrate that the appropriate level of human assistance depends critically on the difficulty of the problem.” Furthermore, there is also a new approach in machine learning, called Automatic Machine Learning (Auto-ML) [100], which uses learning to set the parameters of a learning algorithm automatically. In a traditional approach, a software engineer with machine learning skills is responsible for finding a good configuration for the algorithm parameters. Zoth and Lee [100] present an Auto-ML-based approach to design a neural network to classify images of a specific dataset. In addition, they compared their results with the previous state-of-the-art model, which was designed by an ML expert engineer. According to Zoth and Lee [100] , their AutoML-based approach “can design a novel network architecture that rivals the best human-invented architecture in terms of test set accuracy.” Zoth and Lee also showed that a machine-learning technique is capable of beating a software engineer with ML skills in a specific software engineering task, but the authors do not discuss this subject in the paper. Our paper appears to be the first to provide an empirical study to investigate the use of a machine-learning techniques to solve a problem in the field of Software Engineering, by comparing the solution provided by a MLbased approach against solutions provided by software engineers. 9 C ONCLUSION AND F UTURE W ORK Several researchers have proposed the use of machine-learning techniques to automate software engineering tasks. However, most of these approaches do not direct efforts toward asking whether ML-based procedures have higher success rates than current standard and manual practices. A relevant question in this potential line of investigation is: “Could a software engineer solve a specific development task better than an ML algorithm?”. Indeed, it is fundamental to evaluate which tasks are better performed by engineers or ML procedures so that they can work together more effectively and also provide more insight into novel human-in-the-loop machinelearning approaches to support SE tasks. This paper appears to be the first to provide an empirical study comparing how software engineers and machine-learning algorithms achieve performance and reuse tasks. In brief, as a result of our experiment, we have found evidence that in some cases, software engineers outperform machine-learning algorithms, and in other cases, they do not. Further, as is typical in experimental studies, although we have designed and conducted the experiment carefully, there are always factors that can threaten the experiment’s validity. For example, some threats include the number and diversity of the software engineers involved in our experiment. Understanding how software engineers fare against ML algorithms is essential to support new methodologies for developing human-inthe-loop approaches in which machine learning automated procedures assist software developers in achieving their tasks. For example, methodologies to define which agent (engineers or automated ML procedure) should execute a specific task in a software development set. Based on this understanding, these methodologies can provide a basis for software engineers and machine learning algorithms to cooperate in Software Engineering development more effectively. Future work to extend the proposed experiment includes: (i) conducting further empirical studies to assess other SE tasks, such as design, maintenance and testing; (ii) experimenting with other machine-learning algorithms such as reinforcement learning and backpropagation; and (iii) using different criteria to evaluate task execution. Possible tasks that could be investigated (refer to (i)) include programming tasks, in which case tasks performed by software development teams and ML algorithms are compared. For example, we could invite software developers from the team with the highest score in the last ACM International Collegiate Programming Contest [131], which is one of the most important programming championships in the world, to be involved in this comparison. This competition evaluates the capability of software engineers to solve complex 16 software problems. Software engineers are classified according to the number of right solutions, performance of the solutions and development time. Another line of investigation could address the use of different qualitative or quantitative methodologies. For example, the task execution comparison could rely on reference performances, such as the performance of highly successful performers [100], [129], [130] . This research work can also be extended by proposing, based on the comparison between the performance of engineers and ML algorithms, a methodology for more effective task allocation. This methodology could, in principle, lead to more effective ways to allocate tasks such as software development in cooperative work involving humans and automated procedures. Such human-in-the-loop approaches, which take into account the strengths and weaknesses of humans and machine learning algorithms, are fundamental to provide a basis for cooperative work in software engineering and possibly in other areas. ACKNOWLEDGMENTS This work has been supported by the Laboratory of Software Engineering (LES) at PUC-Rio. Our thanks to CAPES, CNPq, FAPERJ and PUCRio for their support through scholarships and fellowships. We would also like to thank the software engineers who participated in our experiment. R EFERENCES [1] [2] [3] [4] [5] [6] F. Brooks and H. Kugler, No silver bullet. April, 1987. R. S. Pressman, Software engineering: a practitioner’s approach. Palgrave Macmillan, 2005. Q. Zhang, “Software developments,” Engineering Automation for Reliable Software, p. 292, 2000. J. O. Kephart, “Research challenges of autonomic computing,” in Software Engineering, 2005. ICSE 2005. Proceedings. 27th International Conference on. IEEE, 2005, pp. 15–22. J. Mostow, “Foreword what is ai? and what does it have to do with software engineering?” IEEE Transactions on Software Engineering, vol. 11, no. 11, p. 1253, 1985. D. Barstow, “Artificial intelligence and software engineering,” in Proceedings of the 9th [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] international conference on Software Engineering. IEEE Computer Society Press, 1987, pp. 200–211. D. Partridge, “Artificial intelligence and software engineering: a survey of possibilities,” Information and Software Technology, vol. 30, no. 3, pp. 146–152, 1988. L. C. Cheung, S. Ip, and T. Holden, “Survey of artificial intelligence impacts on information systems engineering,” Information and Software Technology, vol. 33, no. 7, pp. 499–508, 1991. D. Partridge, Artificial Intelligence in Software Engineering. Wiley Online Library, 1998. A. Van Lamsweerde and L. Willemet, “Inferring declarative requirements specifications from operational scenarios,” IEEE Transactions on Software Engineering, vol. 24, no. 12, pp. 1089–1114, 1998. G. D. Boetticher, “Using machine learning to predict project effort: Empirical case studies in data-starved domains,” in Model Based Requirements Workshop. Citeseer, 2001, pp. 17–24. F. Padberg, T. Ragg, and R. Schoknecht, “Using machine learning for estimating the defect content after an inspection,” IEEE Transactions on Software Engineering, vol. 30, no. 1, pp. 17–28, 2004. D. Zhang, “Applying machine learning algorithms in software development,” in The Proceedings of 2000 Monterey Workshop on Modeling Software System Structures, 2000, pp. 275–285. ——, “Machine learning in value-based software test data generation,” in Tools with Artificial Intelligence, 2006. ICTAI’06. 18th IEEE International Conference on. IEEE, 2006, pp. 732–736. D. Zhang and J. J. Tsai, Machine learning applications in software engineering. World Scientific, 2005, vol. 16. D. Zhang, “Machine learning and valuebased software engineering: a research agenda.” in SEKE, 2008, pp. 285–290. T. M. Khoshgoftaar, “Introduction to the special issue on quality engineering with computational intelligence,” 2003. D. Zhang, “Machine learning and valuebased software engineering,” in Software Applications: Concepts, Methodologies, Tools, and Applications. IGI Global, 2009, pp. 3325–3339. 17 [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] D. Zhang and J. J. Tsai, “Machine learning and software engineering,” in Tools with Artificial Intelligence, 2002.(ICTAI 2002). Proceedings. 14th IEEE International Conference on. IEEE, 2002, pp. 22–29. M. D. Kramer and D. Zhang, “Gaps: a genetic programming system,” in Computer Software and Applications Conference, 2000. COMPSAC 2000. The 24th Annual International. IEEE, 2000, pp. 614–619. A. Holzinger, M. Plass, K. Holzinger, G. C. Crişan, C.-M. Pintea, and V. Palade, “Towards interactive machine learning (iml): applying ant colony algorithms to solve the traveling salesman problem with the human-in-the-loop approach,” in International Conference on Availability, Reliability, and Security. Springer, 2016, pp. 81–95. A. Holzinger, “Interactive machine learning for health informatics: when do we need the human-in-the-loop?” Brain Informatics, vol. 3, no. 2, pp. 119–131, 2016. S. Easterbrook, J. Singer, M.-A. Storey, and D. Damian, “Selecting empirical methods for software engineering research,” Guide to advanced empirical software engineering, pp. 285–311, 2008. H. A. Simon, “Whether software engineering needs to be artificially intelligent,” IEEE Transactions on Software Engineering, no. 7, pp. 726–732, 1986. I. Sommerville, “Artificial intelligence and systems engineering,” Prospects for Artificial Intelligence: Proceedings of AISB’93, 29 March-2 April 1993, Birmingham, UK, vol. 17, p. 48, 1993. R. S. Michalski, J. G. Carbonell, and T. M. Mitchell, Machine learning: An artificial intelligence approach. Springer Science & Business Media, 2013. A. Marchetto and A. Trentini, “Evaluating web applications testability by combining metrics and analogies,” in Information and Communications Technology, 2005. Enabling Technologies for the New Knowledge Society: ITI 3rd International Conference on. IEEE, 2005, pp. 751–779. S. Bouktif, F. Ahmed, I. Khalil, and G. Antoniol, “A novel composite model approach to improve ality prediction,” Information and Software Technology, vol. 52, no. 12, pp. 1298–1311, 2010. L. Radlinski, “A survey of bayesian net models for software development effort [30] [31] [32] [33] [34] [35] [36] [37] [38] [39] prediction,” International Journal of Software Engineering and Computing, vol. 2, no. 2, pp. 95–109, 2010. W. Zhang, Y. Yang, and Q. Wang, “Handling missing data in software effort prediction with naive bayes and em algorithm,” in Proceedings of the 7th International Conference on Predictive Models in Software Engineering. ACM, 2011, p. 4. Ł. Radliński, “A framework for integrated software quality prediction using bayesian nets,” Computational Science and Its Applications-ICCSA 2011, pp. 310–325, 2011. P. O. O. Sack, M. Bouneffa, Y. Maweed, and H. Basson, “On building an integrated and generic platform for software quality evaluation,” in Information and Communication Technologies, 2006. ICTTA’06. 2nd, vol. 2. IEEE, 2006, pp. 2872–2877. M. Reformat and D. Zhang, “Introduction to the special issue on:“software quality improvements and estimations with intelligence-based methods”,” Software Quality Journal, vol. 15, no. 3, pp. 237– 240, 2007. B. Twala, M. Cartwright, and M. Shepperd, “Applying rule induction in software prediction,” in Advances in Machine Learning Applications in Software Engineering. IGI Global, 2007, pp. 265–286. V. U. Challagulla, F. B. Bastani, and I.-L. Yen, “High-confidence compositional reliability assessment of soa-based systems using machine learning techniques,” in Machine Learning in Cyber Trust. Springer, 2009, pp. 279–322. R. C. Veras, S. R. Meira, A. L. Oliveira, and B. J. Melo, “Comparative study of clustering techniques for the organization of software repositories,” in Hybrid Intelligent Systems, 2007. HIS 2007. 7th International Conference on. IEEE, 2007, pp. 372–377. I. Birzniece and M. Kirikova, “Interactive inductive learning service for indirect analysis of study subject compatibility,” in Proceedings of the BeneLearn, 2010, pp. 1–6. D. B. Hanchate, “Analysis, mathematical modeling and algorithm for software project scheduling using bcga,” in Intelligent Computing and Intelligent Systems (ICIS), 2010 IEEE International Conference on, vol. 3. IEEE, 2010, pp. 1–7. Z. Xu and B. Song, “A machine learning ap- 18 [40] [41] [42] [43] [44] [45] [46] [47] [48] [49] plication for human resource data mining problem,” Advances in Knowledge Discovery and Data Mining, pp. 847–856, 2006. J. Wen, S. Li, Z. Lin, Y. Hu, and C. Huang, “Systematic literature review of machine learning based software development effort estimation models,” Information and Software Technology, vol. 54, no. 1, pp. 41– 59, 2012. E. Rashid, S. Patnayak, and V. Bhattacherjee, “A survey in the area of machine learning and its application for software quality prediction,” ACM SIGSOFT Software Engineering Notes, vol. 37, no. 5, pp. 1–7, 2012. H. A. Al-Jamimi and M. Ahmed, “Machine learning-based software quality prediction models: state of the art,” in Information Science and Applications (ICISA), 2013 International Conference on. IEEE, 2013, pp. 1–4. Ł. Radliński, “Enhancing bayesian network model for integrated software quality prediction,” in Proc. Fourth International Conference on Information, Process, and Knowledge Management, Valencia. Citeseer, 2012, pp. 144–149. F. Pinel, P. Bouvry, B. Dorronsoro, and S. U. Khan, “Savant: Automatic parallelization of a scheduling heuristic with machine learning,” in Nature and Biologically Inspired Computing (NaBIC), 2013 World Congress on. IEEE, 2013, pp. 52–57. D. Novitasari, I. Cholissodin, and W. F. Mahmudy, “Optimizing svr using local best pso for software effort estimation,” Journal of Information Technology and Computer Science, vol. 1, no. 1, 2016. Ł. Radliński, “Towards expert-based modelling of integrated software quality,” Journal of Theoretical and Applied Computer Science, vol. 6, no. 2, pp. 13–26, 2012. T. Rongfa, “Defect classification method for software management quality control based on decision tree learning,” in Advanced Technology in Teaching-Proceedings of the 2009 3rd International Conference on Teaching and Computational Science (WTCS 2009). Springer, 2012, pp. 721–728. R. Rana and M. Staron, “Machine learning approach for quality assessment and prediction in large software organizations,” in Software Engineering and Service Science (ICSESS), 2015 6th IEEE International Conference on. IEEE, 2015, pp. 1098–1101. H. Wang, M. Kessentini, W. Grosky, and [50] [51] [52] [53] [54] [55] [56] [57] [58] [59] H. Meddeb, “On the use of time series and search based software engineering for refactoring recommendation,” in Proceedings of the 7th International Conference on Management of computational and collective intElligence in Digital EcoSystems. ACM, 2015, pp. 35–42. V. U. Challagulla, F. B. Bastani, I.-L. Yen, and R. A. Paul, “Empirical assessment of machine learning based software defect prediction techniques,” in Object-Oriented Real-Time Dependable Systems, 2005. WORDS 2005. 10th IEEE International Workshop on. IEEE, 2005, pp. 263– 270. K. Kaminsky and G. Boetticher, “Building a genetically engineerable evolvable program (geep) using breadth-based explicit knowledge for predicting software defects,” in Fuzzy Information, 2004. Processing NAFIPS’04. IEEE Annual Meeting of the, vol. 1. IEEE, 2004, pp. 10–15. ——, “How to predict more with less, defect prediction using machine learners in an implicitly data starved domain,” in The 8th world multiconference on systemics, cybernetics and informatics, Orlando, FL. Citeseer, 2004. K. Kaminsky and G. D. Boetticher, “Better software defect prediction using equalized learning with machine learners,” Knowledge Sharing and Collaborative Engineering, 2004. O. Kutlubay and A. Bener, “A machine learning based model for software defect prediction,” working paer, Boaziçi University, Computer Engineering Department, 2005. X. Ren, “Learn to predict “affecting changes” in software engineering,” 2003. E. Ceylan, F. O. Kutlubay, and A. B. Bener, “Software defect identification using machine learning techniques,” in Software Engineering and Advanced Applications, 2006. SEAA’06. 32nd EUROMICRO Conference on. IEEE, 2006, pp. 240–247. Y. Kastro and A. B. Bener, “A defect prediction method for software versioning,” Software Quality Journal, vol. 16, no. 4, pp. 543–562, 2008. O. Kutlubay, B. Turhan, and A. B. Bener, “A two-step model for defect density estimation,” in Software Engineering and Advanced Applications, 2007. 33rd EUROMICRO Conference on. IEEE, 2007, pp. 322–332. A. S. Namin and M. Sridharan, “Bayesian 19 [60] [61] [62] [63] [64] [65] [66] [67] [68] [69] reasoning for software testing,” in Proceedings of the FSE/SDP workshop on Future of software engineering research. ACM, 2010, pp. 349–354. C. Murphy and G. Kaiser, “Metamorphic runtime checking of non-testable programs,” Columbia University Dept of Computer Science Tech Report cucs-042-09, p. 9293, 2009. W. Afzal, R. Torkar, R. Feldt, and T. Gorschek, “Genetic programming for cross-release fault count predictions in large and complex software projects,” Evolutionary Computation and Optimization Algorithms in Software Engineering, pp. 94–126, 2010. C. Murphy et al., “Using metamorphic testing at runtime to detect defects in applications without test oracles,” 2008. D. Qiu, S. Fang, and Y. Li, “A framework to discover potential deviation between program and requirement through mining object graph,” in Computer Application and System Modeling (ICCASM), 2010 International Conference on, vol. 4. IEEE, 2010, pp. V4– 110. C. Murphy, G. E. Kaiser et al., “Automatic detection of defects in applications without test oracles,” Dept. Comput. Sci., Columbia Univ., New York, NY, USA, Tech. Rep. CUCS027-10, 2010. W. Afzal, “Search-based approaches to software fault prediction and software testing,” Ph.D. dissertation, Blekinge Institute of Technology, 2009. M. K. Taghi, B. Cukic, and N. Seliya, “An empirical assessment on program moduleorder models,” Quality Technology & Quantitative Management, vol. 4, no. 2, pp. 171– 190, 2007. J. H. Wang, N. Bouguila, and T. Bdiri, “Empirical evaluation of selected algorithms for complexity-based classification of software modules and a new model,” in Intelligent Systems: From Theory to Practice. Springer, 2010, pp. 99–131. H. Jin, Y. Wang, N.-W. Chen, Z.-J. Gou, and S. Wang, “Artificial neural network for automatic test oracles generation,” in Computer Science and Software Engineering, 2008 International Conference on, vol. 2. IEEE, 2008, pp. 727–730. J. Ferzund, S. N. Ahsan, and F. Wotawa, “Automated classification of faults in pro- [70] [71] [72] [73] [74] [75] [76] [77] [78] [79] [80] [81] gramms using machine learning techniques,” in Artificial Intelligence Techniques in Software Engineering Workshop, 2008. O. Maqbool and H. Babri, “Bayesian learning for software architecture recovery,” in Electrical Engineering, 2007. ICEE’07. International Conference on. IEEE, 2007, pp. 1–6. A. Okutan, “Software defect prediction using bayesian networks and kernel methods,” Ph.D. dissertation, ISIK UNIVERSITY, 2012. D. Cotroneo, R. Pietrantuono, and S. Russo, “A learning-based method for combining testing techniques,” in Proceedings of the 2013 International Conference on Software Engineering. IEEE Press, 2013, pp. 142–151. D. Zhang, “A value-based framework for software evolutionary testing,” in Advances in Abstract Intelligence and Soft Computing. IGI Global, 2013, pp. 355–373. A. Okutan and O. T. Yıldız, “Software defect prediction using bayesian networks,” Empirical Software Engineering, vol. 19, no. 1, pp. 154–181, 2014. S. Agarwal and D. Tomar, “A feature selection based model for software defect prediction,” assessment, vol. 65, 2014. G. Abaei and A. Selamat, “Important issues in software fault prediction: A road map,” in Handbook of Research on Emerging Advancements and Technologies in Software Engineering. IGI Global, 2014, pp. 510–539. A. Okutan and O. T. Yildiz, “A novel kernel to predict software defectiveness,” Journal of Systems and Software, vol. 119, pp. 109– 121, 2016. X.-d. Mu, R.-h. Chang, and L. Zhang, “Software defect prediction based on competitive organization coevolutionary algorithm,” Journal of Convergence Information Technology (JCIT) Volume7, Number5, 2012. J. Cahill, J. M. Hogan, and R. Thomas, “Predicting fault-prone software modules with rank sum classification,” in Software Engineering Conference (ASWEC), 2013 22nd Australian. IEEE, 2013, pp. 211–219. R. Rana, M. Staron, C. Berger, J. Hansson, M. Nilsson, and W. Meding, “The adoption of machine learning techniques for software defect prediction: An initial industrial validation,” in Joint Conference on Knowledge-Based Software Engineering. Springer, 2014, pp. 270–285. T. Schulz, Ł. Radliński, T. Gorges, and 20 [82] [83] [84] [85] [86] [87] [88] [89] [90] [91] W. Rosenstiel, “Predicting the flow of defect correction effort using a bayesian network model,” Empirical Software Engineering, vol. 18, no. 3, pp. 435–477, 2013. E. Rashid, “R4 model for case-based reasoning and its application for software fault prediction,” International Journal of Software Science and Computational Intelligence (IJSSCI), vol. 8, no. 3, pp. 19–38, 2016. ——, “Improvisation of case-based reasoning and its application for software fault prediction,” International Journal of Services Technology and Management, vol. 21, no. 4-6, pp. 214–227, 2015. J. K. Chhabra and A. Parashar, “Prediction of changeability for object oriented classes and packages by mining change history,” in Electrical and Computer Engineering (CCECE), 2014 IEEE 27th Canadian Conference on. IEEE, 2014, pp. 1–6. G. Spanoudakis, A. S. d. Garcez, and A. Zisman, “Revising rules to capture requirements traceability relations: A machine learning approach.” in SEKE, 2003, pp. 570–577. M. Shin and A. Goel, “Modeling software component criticality using a machine learning approach,” Artificial Intelligence and Simulation, pp. 440–448, 2005. J. S. Shirabad, “Predictive techniques in software engineering,” in Encyclopedia of Machine Learning. Springer, 2011, pp. 782– 789. A. A. Araújo, M. Paixao, I. Yeltsin, A. Dantas, and J. Souza, “An architecture based on interactive optimization and machine learning applied to the next release problem,” Automated Software Engineering, pp. 1–49, 2016. T. Tourwé, J. Brichau, A. Kellens, and K. Gybels, “Induced intentional software views,” Computer Languages, Systems & Structures, vol. 30, no. 1, pp. 35–47, 2004. J. S. Di Stefano and T. Menzies, “Machine learning for software engineering: Case studies in software reuse,” in Tools with Artificial Intelligence, 2002.(ICTAI 2002). Proceedings. 14th IEEE International Conference on. IEEE, 2002, pp. 246–251. J. Fu, F. B. Bastani, and I.-L. Yen, “Automated ai planning and code pattern based code synthesis,” in Tools with Artificial Intelligence, 2006. ICTAI’06. 18th IEEE International Conference on. IEEE, 2006, pp. 540– 546. J. Fu, F. B. Bastani, I.-L. YEN et al., “Semantic-driven component-based automated code synthesis,” Semantic Computing, pp. 249–283, 2010. [93] A. Katasonov, O. Kaykova, O. Khriyenko, S. Nikitin, and V. Y. Terziyan, “Smart semantic middleware for the internet of things.” ICINCO-ICSO, vol. 8, pp. 169–178, 2008. [94] L. Baresi, S. Guinea, and A. Shahzada, “Short paper: Harmonizing heterogeneous components in sesame,” in Internet of Things (WF-IoT), 2014 IEEE World Forum on. IEEE, 2014, pp. 197–198. [95] L. Zhu, H. Cai, and L. Jiang, “Minson: A business process self-adaptive framework for smart office based on multi-agent,” in e-Business Engineering (ICEBE), 2014 IEEE 11th International Conference on. IEEE, 2014, pp. 31–37. [96] J. F. De Paz, J. Bajo, S. Rodrı́guez, G. Villarrubia, and J. M. Corchado, “Intelligent system for lighting control in smart cities,” Information Sciences, vol. 372, pp. 241–255, 2016. [97] I. Birzniece, “The use of inductive learning in information systems,” in Proceedings of the 16th International Conference on Information and Software Technologies, 2010, pp. 95– 101. [98] D. Alrajeh, A. Russo, and S. Uchitel, “Inferring operational requirements from scenarios and goal models using inductive learning,” in Proceedings of the 2006 international workshop on Scenarios and state machines: models, algorithms, and tools. ACM, 2006, pp. 29–36. [99] A. M. Sharifloo, A. Metzger, C. Quinton, L. Baresi, and K. Pohl, “Learning and evolution in dynamic software product lines,” in Proceedings of the 11th International Symposium on Software Engineering for Adaptive and Self-Managing Systems. ACM, 2016, pp. 158–164. [100] B. Zoph and Q. V. Le, “Neural architecture search with reinforcement learning,” arXiv preprint arXiv:1611.01578, 2016. [101] N. M. do Nascimento and C. J. P. de Lucena, “Fiot: An agent-based framework for self-adaptive and self-organizing applications based on the internet of things,” Information Sciences, vol. 378, pp. 161–176, 2017. [102] F. Jacob and R. Tairas, “Code template in[92] 21 [103] [104] [105] [106] [107] [108] [109] [110] [111] [112] [113] ference using language models,” in Proceedings of the 48th Annual Southeast Regional Conference. ACM, 2010, p. 104. B. Amal, M. Kessentini, S. Bechikh, J. Dea, and L. B. Said, “On the use of machine learning and search-based software engineering for ill-defined fitness function: a case study on software refactoring,” in International Symposium on Search Based Software Engineering. Springer, 2014, pp. 31– 45. Y. Peng, G. Wang, and H. Wang, “User preferences based software defect detection algorithms selection using mcdm,” Information Sciences, vol. 191, pp. 3–13, 2012. H. A. Abbass, E. Petraki, K. Merrick, J. Harvey, and M. Barlow, “Trusted autonomy and cognitive cyber symbiosis: Open challenges,” Cognitive computation, vol. 8, no. 3, pp. 385–408, 2016. W. G. Baxt, “Use of an artificial neural network for the diagnosis of myocardial infarction,” Annals of internal medicine, vol. 115, no. 11, pp. 843–848, 1991. M. A. Mazurowski, P. A. Habas, J. M. Zurada, J. Y. Lo, J. A. Baker, and G. D. Tourassi, “Training neural network classifiers for medical decision making: The effects of imbalanced datasets on classification performance,” Neural networks, vol. 21, no. 2, pp. 427–436, 2008. N. M. do Nascimento, M. L. Viana, and C. J. P. de Lucena, “An iot-based tool for human gas monitoring,” in IXV Congresso Brasileiro de Informatica em Saude (CBIS), vol. 1. SBIS, 2016, pp. 96–98. R. Morejón, M. Viana, and C. Lucena, “Generating software agents for data mining: An example for the health data area,” in International Conference on Software Engineering & Knowledge Engineering-SEKE, 2017. S. Haykin, Neural Networks: A Comprehensive Foundation. Macmillan, 1994. [Online]. Available: http://books.google. com.br/books?id=PSAPAQAAMAAJ D. Castelvecchi, “Can we open the black box of ai?” Nature News, vol. 538, no. 7623, p. 20, 2016. L. Atzori, A. Iera, and G. Morabito, “The internet of things: A survey,” Computer networks, vol. 54, no. 15, pp. 2787–2805, 2010. M. A. Salahuddin, A. Al-Fuqaha, M. Guizani, K. Shuaib, and F. Sallabi, [114] [115] [116] [117] [118] [119] [120] [121] [122] [123] [124] [125] “Softwarization of internet of things infrastructure for secure and smart healthcare,” Computer, vol. 50, no. 7, pp. 74–79, 2017. I. Ayala, M. Amor, L. Fuentes, and J. M. Troya, “A software product line process to develop agents for the iot,” Sensors, vol. 15, no. 7, pp. 15 640–15 660, 2015. J.-P. Briot, N. M. de Nascimento, and C. J. P. de Lucena, “A multi-agent architecture for quantified fruits: Design and experience,” in 28th International Conference on Software Engineering & Knowledge Engineering (SEKE’2016). SEKE/Knowledge Systems Institute, PA, USA, 2016, pp. 369–374. N. M. Nascimento, “FIoT: An agentbased framework for self-adaptive and self-organizing internet of things applications,” Master’s thesis, PUC-Rio, Rio de Janeiro, Brazil, August 2015. N. M. d. Nascimento, C. J. P. d. Lucena, and H. Fuks, “Modeling quantified things using a multi-agent system,” in IEEE / WIC / ACM International Conference on Web Intelligence and Intelligent Agent Technology (WIIAT), vol. 1. IEEE, 2015, pp. 26–32. N. M. NASCIMENTO and C. J. P. LUCENA, “Engineering cooperative smart things based on embodied cognition,” in NASA/ESA Conference on Adaptive Hardware and Systems (AHS 2017). IEEE, 2017. Apple, “Homekit,” https://developer.apple.com/homekit/, March 2017. Samsung, “Samsung smart things,” https://www.smartthings.com, March 2017. D. I. Sjøberg, T. Dybå, B. C. Anda, and J. E. Hannay, “Building theories in software engineering,” Guide to advanced empirical software engineering, pp. 312–336, 2008. S. Whiteson, N. Kohl, R. Miikkulainen, and P. Stone, “Evolving soccer keepaway players through task decomposition,” Machine Learning, vol. 59, no. 1-2, pp. 5–30, 2005. C. Wohlin, P. Runeson, M. Höst, M. C. Ohlsson, B. Regnell, and A. Wesslén, Experimentation in software engineering. Springer Science & Business Media, 2012. R. S. Sutton and A. G. Barto, Reinforcement learning: An introduction. MIT press Cambridge, 1998, vol. 1, no. 1. R. Peck and J. Devore, Statistics: The Exploration & Analysis of Data. Nelson Educa- 22 tion, 2011. [126] W. Oizumi, L. Sousa, A. Garcia, R. Oliveira, A. Oliveira, O. Agbachi, and C. Lucena, “Revealing design problems in stinky code: a mixed-method study,” in Proceedings of the 11th Brazilian Symposium on Software Components, Architectures, and Reuse. ACM, 2017, p. 5. [127] E. Fernandes, F. Ferreira, J. A. Netto, and E. Figueiredo, “Information systems development with pair programming: An academic quasi-experiment,” in Proceedings of the XII Brazilian Symposium on Information Systems on Brazilian Symposium on Information Systems: Information Systems in the Cloud Computing Era-Volume 1. Brazilian Computer Society, 2016, p. 64. [128] G. Kasparov, Deep Thinking: Where Machine Intelligence Ends and Human Creativity Begins. Hachette UK, 2017. [129] D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. Van Den Driessche, J. Schrittwieser, I. Antonoglou, V. Panneershelvam, M. Lanctot et al., “Mastering the game of go with deep neural networks and tree search,” Nature, vol. 529, no. 7587, pp. 484–489, 2016. [130] D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, A. Guez, T. Hubert, L. Baker, M. Lai, A. Bolton et al., “Mastering the game of go without human knowledge,” Nature, vol. 550, no. 7676, pp. 354–359, 2017. [131] A. Trotman and C. Handley, “Programming contest strategy,” Computers & Education, vol. 50, no. 3, pp. 821–837, 2008. 

2cs.AI

Joint optimization of transmission and propulsion in aerial communication networks arXiv:1710.01529v1 [cs.SY] 4 Oct 2017 Omar J. Faqir, Eric C. Kerrigan, and Deniz Gündüz Abstract— Communication energy in a wireless network of mobile autonomous agents should be considered as the sum of transmission energy and propulsion energy used to facilitate the transfer of information. Accordingly, communication-theoretic and Newtonian dynamic models are developed to model the communication and locomotion expenditures of each node. These are subsequently used to formulate a novel nonlinear optimal control problem (OCP) over a network of autonomous nodes. It is then shown that, under certain conditions, the OCP can be transformed into an equivalent convex form. Numerical results for a single link between a node and access point allow for comparison with known solutions before the framework is applied to a multiple-node UAV network, for which previous results are not readily extended. Simulations show that transmission energy can be of the same order of magnitude as propulsion energy allowing for possible savings, whilst also exemplifying how speed adaptations together with power control may increase the network throughput. U2 U1 a1 p a22 + δ22 AP (0, 0) Fig. 1: Geometric configuration for simulation setups featuring N = 1 (black) and N = 2 (green) nodes. Speeds along these paths may be variable or fixed. The altitudes and lateral displacements of U1 , U2 are a1 = a2 = 1000 m and δ1 = 0, δ2 = 1000 m, respectively. I. I NTRODUCTION We aim to derive a control strategy to minimize communication energy in robotic networks. In particular, uninhabited aerial vehicle (UAV) networks are considered, with results being generalizable to broader classes of autonomous networks. A dynamic transmission model, based on physical layer communication-theoretic bounds, and a mobility model for each node is considered alongside a possible network topology. As a cost function, we employ the underused interpretation of communication energy as the sum of transmission energy and propulsion energy used for transmission, i.e. when a node changes position to achieve a better channel. For simulation purposes we consider the two wireless network setups shown in Figure 1. We first present the most basic scenario consisting of a single agent U1 moving along a predefined linear path while offloading its data to a stationary access point (AP). We compare results for variable and fixed speeds, before studying a two-agent single-hop network. For UAV networks, research efforts largely break down into two streams: the use of UAVs in objective based missions (e.g. search and pursuit [1], information gathering/mobile sensor networks [2], [3]), and use as supplementary network links [4]. Optimal completion of these macro goals has been addressed in the literature, but there The support of the EPSRC Centre for Doctoral Training in High Performance Embedded and Distributed Systems (HiPEDS, Grant Reference EP/L016796/1) is gratefully acknowledged. O. J. Faqir and Deniz Gündüz are with the Department of Electrical & Electronics Engineering, Imperial College London, SW7 2AZ, U.K. ojf12@ic.ac.uk, d.gunduz@ic.ac.uk Eric C. Kerrigan is with the Department of Electrical & Electronic Engineering and Department of Aeronautics, Imperial College London, London SW7 2AZ, U.K. e.kerrigan@imperial.ac.uk is no necessary equivalence between optimal task-based and energy-efficient operations. Efforts concerning mobility focus on mobile (in which node mobility models are random) or vehicular (where mobility is determined by higher level objectives and infrastructure) ad-hoc networks [5]. Since neither are fully autonomous networks, mobility is not available as a decision variable. The work in [6] introduced the concept of proactive networks, where certain nodes are available as mobile relays. However, the focus is on relay trajectory design and a simplistic transmission model is assumed, inherently prohibiting energy efficiency. The related problem of router formation is investigated in [7] using realistic models of communication environments. We assume hard path constraints, possibly due to the existence of higher level macro objectives, but allow changes in trajectory along the path by optimizing their speed (as in [8], we define a trajectory as being a time-parameterized path). Use of fixed paths does not restrict our results as most UAV path planning algorithms operate over longer time horizons and are generally restricted to linear or circular loiter trajectories [8]. A linear program (LP) is used in [9] to determine how close a rolling-robot should move before transmission in order to minimize total energy. However, the linear motion dynamics used restricts applicability of the model. Similarly to our current work, [10] uses a single mobile UAV relay to maximize data throughput between a stationary source-destination pair. An optimal trajectory for an a priori transmission scheme is iteratively found. Similarly, for a given trajectory, the optimal relaying scheme may be obtained through water-filling over the source-torelay and relay-to-receiver channels. Our contribution differs from the above works in terms of the formulation of a more general nonlinear convex OCP for finding joint transmission and mobility strategies to minimize communication energy. We solve this problem, exemplifying possible savings for even just a single node. As a final point, we show analytically and numerically that, even at fixed speeds, the optimal transmission scheme for a twouser multiple-access channel(MAC) is counter-intuitive and not captured by naı̈ve transmission policies. II. P ROBLEM D ESCRIPTION Consider N homogeneous mobile nodes Un , n ∈ N , {1, . . . , N }, traveling along linear non-intersecting trajectories at constant altitudes an and lateral displacements δn over a time interval T , [0, T ]. The trajectory of node Un is denoted by t 7→ (qn (t), δn , an ), relative to a single stationary AP U0 at position (0, 0, 0) in a three dimensional space. At t = 0, Un is initialized with a data load of Dn bits, which must all be offloaded to U0 by time t = T . We consider a cooperative network model, in which all nodes cooperate to offload all the data in the network to the AP by relaying each other’s data. Each node has a data buffer of capacity M bits, which limits the amount of data it can store and relay. CN (q, p) , {r ≥ 0 | fm (q, p, r, S) ≤ 0, ∀S ⊆ N } , We employ scalar additive white Gaussian noise (AWGN) channels. For UAV applications, we assume all links are dominated by line-of-sight (LoS) components, resulting in flat fading channels, meaning all signal components undergo similar amplitude gains [11]. All nodes have perfect information regarding link status, which in practice may be achieved through feedback of channel measurements, while the overhead due to channel state feedback is ignored. Similar to [12], for a given link from source node Un to receiver node Um , the channel gain ηnm (·) is expressed as (1) where qnm , qn − qm , constant G represents transmit and receive antenna gains and α ≥ 1 the path loss exponent. We define anm and δnm in a similar fashion. The channel gain is inversely related to the Euclidean distance between nodes. Each node has a single omnidirectional antenna of maximum transmit power of Pmax Watts. We consider half duplex radios; each node transmits and receives over orthogonal frequency bands. Accordingly, a different frequency band is assigned for each node’s reception, and all messages destined for this node are transmitted over this band, forming a MAC. We do not allow any coding (e.g. network coding) or combining of different data packets at the nodes, and instead consider a decode-and-forward-based routing protocol at the relay nodes [13]. The resulting network is a composition of Gaussian MACs, for each of which the set of achievable rate tuples defines a polymatroid capacity region [14]. If N (2) where q is the tuple of the differences qnm in positions between the N users and the receiver, p ∈ P N is the tuple of transmission powers allocated by the N users on this channel, and P , [0, Pmax ] is the range of possible transmission powers for each user. fm (·) is a nonlinear function bounding CN (q, p), given by X fm (q, p, r, S) , rn − n∈S Bm log2 A. Communication Model G ηnm (qnm ) , p 2α , 2 + q2 a2nm + δnm nm nodes simultaneously transmit independent information to the same receiver, the received signal is a superposition of the transmitted signals scaled by their respective channel gains, plus an AWGN term. We model the achievable data rates using Shannon capacity, which is a commonly used upper bound on the practically achievable data rates subject to average power constraints. Due to the convexity of the capacity region, throughput maximization does not require time-sharing between nodes [14], but may be achieved through successive interference cancellation (SIC). Consider a single MAC consisting of N users Un , n ∈ N , transmitting to a receiver Um , m 6∈ N . The capacity region CN (·, ·), which denotes the set of all achievable rate tuples r, is defined as X ηnm (qnm )pn 1+ σ2 n∈S ! , (3) where rn is the nth component of r, Bm is the bandwidth allocated to Um , and σ 2 is the receiver noise power. Consider the example (Section IV-B) where we do not allow relaying. This gives rise to a MAC with N = 2 transmitters U1 , U2 and the AP U0 . The capacity region C2 (q, p) is the set of non-negative tuples (r1 , r2 ) that satisfy   η10 (q10 )p1 (4a) r1 ≤ B0 log2 1 + σ2   η20 (q20 )p2 r2 ≤ B0 log2 1 + (4b) σ2   η10 (q10 )p1 + η20 (q20 )p2 (4c) r1 + r2 ≤ B0 log2 1 + σ2 for all (p1 , p2 ) ∈ P 2 . The first two bounds restrict individual user rates to the single-user Shannon capacity. Dependence between U1 and U2 leads to the final constraint, that the sum rate may not exceed the point-to-point capacity with full cooperation. For transmit powers (p1 , p2 ) these constraints trace out the pentagon shown in Figure 2. The sum rate is maximized at any point on the segment L3 . Referring to SIC, the rate pair at boundary point R(1) is achieved if the signal from source U2 is decoded entirely before source U1 , resulting in the signal from U2 being decoded at a higher interference rate than the signal from U1 . At R(2) the opposite occurs. B. Propulsion Energy Model The electrical energy used for propulsion in rolling robots has been modeled as a linear or polynomial function of speed achievable data rates is bounded above by a set of 2|N | − 1 nonlinear submodular functions fm (·, ·, ·, ·), where |·| applied to a set denotes the cardinality operator. Exponential growth in the number of nodes is a computational intractability. Hence, results are limited to small or structured networks where only a subset of nodes use each MAC. The trajectory of node Un is denoted by the tuple r1 L1 R(1) L3 R(2) L2 Yn , (pn , rn , sn , qn , vn , v̇n , Fn ), r2 Fig. 2: Capacity region for a given power policy across two parallel channels, with corner rate pairs labeled as R(1) = (1) (1) (2) (2) (r1 , r2 ) and R(2) = (r1 , r2 ) and line segments labeled as L1 , L2 , L3 . (7) where qn (t) is the node’s position at time t and sn (t) the state of its storage buffer subject to maximum memory of M bits. The optimal control problem that we want to solve is N Z T X min pn (t) + vn (t)Fn (t)dt (8a) p,r,s,q,v,F n=1 0 s.t. ∀n ∈ N , m ∈ {N , N + 1}, t ∈ T , S ⊆ N in [9], [15] respectively. We take a more general approach, restricting the fixed wing UAV to moving at strictly positive speeds and using Newtonian laws as a basis, as in [16]. The function Ω(·) models the resistive forces acting on node Un in accordance with the following assumption. Assumption 1: The resistive forces acting on each node Un may be modeled by the function x 7→ Ω(x) such that x 7→ xΩ(x) is convex on x ∈ [0, ∞) and ∞ on x ∈ (−∞, 0). Comparatively, in the fixed wing model proposed in [17], the drag force of a UAV traveling at constant altitude at subsonic speed v is Ω(v) = ρCD0 Sv 2 2L2 + 2 (πe0 AR )ρSv 2 (5) where the first term represents parasitic drag and the second term lift-induced drag. Parasitic drag is proportional to v 2 , where ρ is air density, CD0 is the base drag coefficient, and S is the wing area. Lift induced drag is proportional to v −2 , where e0 is the Oswald efficiency, AR the wing aspect ratio and L the induced lift [17]. For fixed-altitude flight, L must be equal to the weight of the craft W = mg. The power required to combat drag is the product of speed and force. The propulsion force Fn (·) must satisfy the force balance equation Fn (t) − Ω(vn (t)) = mn v̇n (t), (6) where mn is the node mass, vn (t) is the speed and v̇n (t) is the acceleration. The instantaneous power used for propulsion is the product vn (t)Fn (t), with the total propulsion energy taken as the integral of this power over T . We assume vn (t) ≥ 0, ∀t ∈ T , which is valid for fixed wing aircrafts. Thrust is restricted to the range [Fmin , Fmax ]. C. General Continuous-Time Problem Formulation We formulate the problem in continuous-time. At time t, node Un , n ∈ N can transmit to any node Um , m ∈ {N , N + 1}\{n} at a non-negative data rate rnm (t) using transmission power pnm (t). The sum power used in all outgoing transmissions from Un is denoted by pn (t). From this, the set of fm (q(t), p(t), r(t), S \ {m}) ≤ 0 ṡn (t) = N X rmn (t) − N +1 X (8b) rnm (t) (8c) sn (0) = Dn , sn (T ) = 0 qn (0) = Qn,init , qn (T ) = Qn,final (8d) (8e) vn (0) = vn,init Fn (t) = mn v̇n (t) + Ω(vn (t)) (8f) (8g) q̇n (t) = ζn vn (t) Yn,min ≤ Yn (t) ≤ Yn,max (8h) (8i) m6=n m6=n The cost function (8a) is the sum of nodal transmission and propulsion energies. Constraint (8b) bounds the achievable data rate to within the receiving nodes’ capacity region, and (8c) updates the storage buffers with sent/received data. Constraints (8d) act as initial and final constraints on the buffers, while (8e)–(8h) ensure all nodes travel from their initial to final destinations without violating a Newtonian force-acceleration constraint; ζn ∈ {−1, 1} depending on whether the position qn (t) decreases or increases, respectively, if the speed vn (t) ≥ 0. The final constraint (8i) places simple bounds on the decision variables, given by Yn,min , (0, 0, 0, −∞, Vmin , −∞, Fmin ), (9a) Yn,max , (Pmax , ∞, M, ∞, Vmax , ∞, Fmax ), (9b) where 0 ≤ Vmin ≤ Vmax and Fmin ≤ Fmax . The above optimal control problem may then be fully discretized using optimal control solvers, such as ICLOCS [18]. Before simulation results are presented we prove that this problem admits an equivalent convex form under certain conditions. III. C ONVEXITY A NALYSIS Efficient convex programming methods exist, which may be used in real-time applications. We first show that the nonlinear data rate constraints (8b) are convex in both positions and transmission power. We then show that the non-linear equality constraint (8g) may be substituted into the cost function, convexifying the cost function. This, however, turns the previously simple thrust bound Fmin ≤ Fn (t) into a concave constraint, resulting in a convex OCP if thrust bounds are relaxed. The absence of thrust bounds arises when considering a fixed trajectory, or is a reasonable assumption if the speed range is sufficiently small. Lemma 1: The rate constraints (8b) are convex in powers and positions for all path loss exponents α ≥ 1. Proof: By writing the channel gains as an explicit function of node positions, for receiver Um each of the capacity region constraints is of the form X rn (t)− n∈S Bm log2 G X pn (t) 1+ 2 2 +q 2 α σ (a2nm + δnm nm (t) ) n∈S ! ≤ 0. (10) Since the non-negative weighted sum of functions preserves convexity properties, without loss of generality we take S to be a singleton, and drop subscripts. We also drop time dependencies. The above function is the composition of two functions φ1 ◦ φ2 (·), respectively defined as φ1 (r, φ2 (·)) , r − B log2 (1 + φ2 (·)), G p φ2 (p, q) , 2 2 . σ (a + δ 2 + q 2 )α (11) (12) The function (p, q) 7→ φ2 (p, q) is concave on the domain R+ ×R. We show this by dropping constants and considering the simpler function h(x, y) , xy −2α with Hessian # " −2α 0 y −2α−1 2 (13) ∇ h(x, y) = 2α(2α−1)x , −2α y −2α−1 y −2α−2 which is negative semi-definite, because it is symmetric with non-positive sub-determinants. Therefore, φ2 is jointly concave in both power and the difference in positions over the specified domain. φ1 is convex and non-increasing as a function of φ2 . Since the composition of a convex, non-increasing function with a concave function is convex [19], all data rate constraint functions are convex functions of (r, p, q). The posynomial objective function is not convex over the whole of its domain and the logarithmic data rate term prevents the use of geometric programming (GP) methods. Lemma 2: The following problem Z T min Fn (t)vn (t)dt (14a) vn ,Fn 0 s.t. ∀t ∈ T Fn (t) − Ω(vn (t)) = mn v̇n (t) Fmin ≤ fm (t) ≤ Fmax (14b) (14c) vn (t) ≥ 0 (14d) vn (0) = vn,init (14e) of minimizing propulsion energy of a single node Un , subject to initial and final conditions, admits an equivalent convex form for mappings vn (t) 7→ Ω(vn (t)) satisfying Assumption 1 and force bounds (Fmin , Fmax ) = (−∞, ∞). Proof: By noting that Fn (t) = Ω(vn (t)) + mn v̇n (t), we move the equality into the cost function, rewriting the problem as (15) min φ(vn ) s.t. (14c)–(14e), vn where φ(vn ) , Z |0 T Z T vn (t)v̇n (t)dt . vn (t)Ω(vn (t))dt + {z } |0 {z } φ1 (vn ) (16) φ2 (vn ) We now show that both φ1 (·) and φ2 (·) are convex. Starting with the latter, by performing a change of variable, the analytic cost is derived by first noting that φ2 (vn ) is the change in kinetic energy Z vn (T )
mn 2 φ2 (vn ) = mn vdv = vn (T ) − vn2 (0) , (17) 2 vn (0) which is a convex function of vn (T ) subject to fixed initial conditions (14d); in fact, it is possible to drop the vn2 (0) term completely without affecting the argmin. By Assumption 1, vn (t) 7→ vn (t)Ω(vn (t)) is convex and continuous on the admissible domain of speeds. Since integrals preserve convexity, the total cost function φ(·) is also convex. Removal of thrust F as a decision variable results in the set VF , {vn | Fmin ≤ Ω(vn (t)) + mn v̇n (t) ≤ Fmax }. (18) Even if Ω(·) is convex on the admissible range of speeds, the lower bound represents a concave constraint not admissible within a convex optimization framework. Therefore, dropping constraints on thrust results in a final convex formulation of Z T
mn 2 min vn (t)Ω(vn (t))dt + vn (T ) − vn2 (0) (19a) vn 2 0 s.t. ∀t ∈ T Vmin ≤ vn ≤ Vmax (19b) vn (0) = vn,init . (19c) Addition of bounds vn ∈ VF naturally results in a difference of convex (DC) problem [20] that may be solved through exhaustive or heuristic procedures. Theorem 1: In the absence of constraints on thrust, the general problem (8) admits an equivalent convex form. Proof: Non-convexities in this formulation arise from the posynomial function of speed v(t) and thrust Fm (t) in the cost function (8a), the nonlinear force balance equality (8g), and the capacity region data rate constraints (8b). The cost function is a superposition of the energies used by each node for propulsion and transmission. By noting that there is no coupling between nodes or between propulsion and transmission powers in this cost, the transformation used in Lemma 2 may be used to eliminate the nonlinear equality. We eliminate Fn (t) and v̇n (t) and move the nonlinear equality into the objective function, simultaneously convexifying the objective to get "Z # N T X mn 2 v (T ) pn (t) + vn (t)Ω(vn (t))dt + min p,r,s,q,v 2 n 0 n=1 s.t. ∀n ∈ N , m ∈ {N , N + 1}, t ∈ T , v ∈ V N , S ⊆ N 35 100 30 80 25 60 20 15 40 10 20 5 0 (8b)–(8f), (8h), Ỹn,min ≤ Ỹn (t) ≤ Ỹn,max 0 0 where Ỹn (t) , (pn (t), rn (t), sn (t), qn (t), vn (t)), and the bounds Ỹn,min , and Ỹn,max are similarly changed. It follows from Lemma 1 that all data rate constraints in (8b) are also convex, therefore the whole problem is convex. IV. S IMULATION R ESULTS A. Single Node A single mobile node U1 of mass 3 kg traveling at fixed altitude a = 1000 m and lateral displacement δ = 0 m, depicted in Figure 1, is considered first. In this section, simulation results are presented for the problem of minimizing the total communication energy to offload all data to U0 . This is compared to a water-filling solution [21] for minimizing the transmission energy. Subscripts denoting different nodes have been dropped in the remainder of this section. Specifically, we use Ω(·) of the form  ∞, ∀x ∈ (−∞, 0) Ω(x) , (21) CD1 x2 + CD2 x−2 , ∀x ∈ [0, ∞), where CD1 = 9.26 × 10−4 is the parasitic drag coefficient and CD2 = 2250 is the lift induced drag coefficient [17]. Simulation results are shown in Figure 3 for a storage buffer initialized to D = 75 MB and speeds restricted in the range [Vmin , Vmax ] = [30, 100] km/h. This results in a total energy expenditure of 309.50 kJ, where 105.05 kJ is due to transmission and 204.51 kJ is due to propulsion. Of this, only 48.01 kJ of extra propulsion energy is used to vary speed on top of the base energy required to traverse the distance at a constant speed. Furthermore, the problem would have been infeasible if the node was restricted to a constant speed of 65 km/h. We note that, with the given parameterization, it is possible to transmit up to 78 MB of data in the defined time interval. B [Hz] 105 M [GB] 1 Pmax [W] 100 α 1.5 400 600 800 1000 1200 (a) Optimal transmission power and propulsion force used by U1 . 10 5 10 35 8 30 25 The open source primal dual Interior Point solver Ipopt v.3.12.4 has been used through the MATLAB interface. Table I contains parameters common to the following experiments. Force constraints are relaxed in all experiments. From [5], the speed of a typical UAV is in the range 30 to 460 km/h. All nodes are initialized to their average speeds vn,init = (Vmax + Vmin )/2, We assume all nodes move in symmetric trajectories around the AP such that Qn,final = −Qn,init = (T /2)vn,init . σ2 [W] 10−10 200 T [min] 20 TABLE I: Dynamic model parameters that have been used across all simulation results. 6 20 15 4 10 2 5 0 0 0 200 400 600 800 1000 1200 (b) Associated achieved data rate and velocity profile of U1 . Fig. 3: Simulation results for the single-node problem, with trajectories shown as solid, and bounds shown as dashed lines. In comparison, if the speed of U1 is fixed, then the maximum transmittable data is approximately 56 MB, using 120.00 kJ of transmission energy. Although considerably more energy is used, the optimal power policy for a fixed trajectory is characterized by a water-filling solution, an equivalent proof of which may be found in [21]. This problem results in a one dimensional search space, easily solved through such algorithms as binary search. B. Multiple Nodes We now investigate the transmission energy problem for two nodes, traveling in parallel trajectories at fixed speeds such that Vmax = Vmin = 65 km/h, as depicted by the green lines in Figure 1. Relaying is not allowed, as may be the case if no bandwidth is allocated to U1 and U2 to receive each other’s transmissions, equivalently turning them into pure source nodes. Simulation results are presented in Figure 4. U1 is closer to the AP at all times, and therefore is advantaged in that it experiences more favorable channel conditions. The disadvantaged node U2 transmits for a longer duration due to the smaller relative change in its channel gain. The interior point algorithm converged after 42 iterations to a minimum energy of 52.707kJ and 26.77kJ for U1 and U2 , respectively, for a starting data load of D1 = D2 = 25 MB. It is notable that the advantaged node uses considerably more transmission energy than the disadvantaged node. Referring to [22], which derives two-user optimal power allocations that achieve arbitrary rate tuples on the boundary of C we explain this as follows. From Figure 2, the optimal rate pairs for given transmit powers p1 and p2 lie on the R EFERENCES 100 80 60 40 20 0 200 400 600 800 1000 1200 (a) Transmit powers of nodes U1 and U2 . 7 6 5 4 3 2 1 0 200 400 600 800 1000 1200 (b) Associated transmission rates achieved by nodes U1 and U2 . Fig. 4: Simulation results for the two-node transmission power problem. segment L3 . Equivalently, ∃̺ ∈ [0, 1] such that the rate (∗) (∗) pair for an arbitrary point R(∗) = (r1 , r2 ) on L3 is given by the interpolation R(∗) = ̺ · R(1) + (1 − ̺) · R(2) . We may interpret ̺ as being the priority assigned to each transmitting node by the U0 when SIC is being carried out. ̺ = 1 means that data from U1 is being decoded second, subject to a lower noise rate, while ̺ = 0 means the opposite decoding order. We may think of the mapping t 7→ ̺(t) as a time-varying priority. However, by calculating ̺(t) from the optimum powers and rates seen in Figure 4, we find that ̺(t) = 0, ∀t ∈ T such that p1 (t) > 0, p2 (t) > 0. In other words, the disadvantaged node is always given priority, which is why it uses less energy at the optimum, even though it always experiences a worse channel gain. V. C ONCLUSIONS We have presented a general optimization framework for joint control of propulsion and transmission energy for single/multi-hop communication links in robotic networks. The relaxation of transmission constraints to theoretic capacity bounds, with relatively mild assumptions on the mobility model, results in a nonlinear but convex OCP. We showed that optimizing over a fixed path, as opposed to a fixed trajectory, increases the feasible starting data by at least 30% for just a single node. For the fixed-trajectory two-node MAC simulation, the optimal solution has been presented and analyzed. Immediate extensions of this work include higher fidelity models, and analysis of the relay network encompassed in problem (8). Considering the overarching goal of real-time control, further developments will be closedloop analysis of the control strategy, and consideration of the computational burden and energy expenditure [3], [23] in the network. [1] J. Ko, A. Mahajan, and R. Sengupta, “A network-centric UAV organization for search and pursuit operations,” in Aerospace Conference Proceedings, 2002. IEEE, vol. 6, pp. 6–6, IEEE, 2002. [2] S. Wang, A. Gasparri, and B. Krishnamachari, “Robotic message ferrying for wireless networks using coarse-grained backpressure control,” in Globecom Workshops (GC Wkshps), 2013 IEEE, pp. 1386–1390, IEEE, 2013. [3] M. Thammawichai, S. P. Baliyarasimhuni, E. C. Kerrigan, and J. B. Sousa, “Optimizing communication and computation for multi-UAV information gathering applications,” arXiv preprint arXiv:1610.04091, 2016. [4] P. Zhan, K. Yu, and A. L. Swindlehurst, “Wireless relay communications with unmanned aerial vehicles: Performance and optimization,” IEEE Transactions on Aerospace and Electronic Systems, vol. 47, no. 3, pp. 2068–2085, 2011. [5] I. Bekmezci, O. K. Sahingoz, and Ş. Temel, “Flying ad-hoc networks (FANETs): A survey,” Ad Hoc Networks, vol. 11, no. 3, pp. 1254– 1270, 2013. [6] W. Zhao and M. H. Ammar, “Message ferrying: Proactive routing in highly-partitioned wireless ad hoc networks,” in Distributed Computing Systems, 2003. FTDCS 2003. Proceedings. The Ninth IEEE Workshop on Future Trends of, pp. 308–314, IEEE, 2003. [7] Y. Yan and Y. Mostofi, “Robotic router formation in realistic communication environments,” IEEE Transactions on Robotics, vol. 28, no. 4, pp. 810–827, 2012. [8] P. Sujit, S. Saripalli, and J. B. Sousa, “Unmanned aerial vehicle path following: A survey and analysis of algorithms for fixed-wing unmanned aerial vehicles,” IEEE Control Systems, vol. 34, no. 1, pp. 42–59, 2014. [9] Y. Yan and Y. Mostofi, “To go or not to go: On energy-aware and communication-aware robotic operation,” IEEE Transactions on Control of Network Systems, vol. 1, no. 3, pp. 218–231, 2014. [10] Y. Zeng, R. Zhang, and T. J. Lim, “Throughput maximization for mobile relaying systems,” arXiv preprint arXiv:1604.02517, 2016. [11] D. Tse and P. Viswanath, Fundamentals of wireless communication. Cambridge university press, 2005. [12] L. Ren, Z. Yan, M. Song, and J. Song, “An improved water-filling algorithm for mobile mimo communication systems over time-varying fading channels,” in Electrical and Computer Engineering, 2004. Canadian Conference on, vol. 2, pp. 629–632, IEEE, 2004. [13] D. Gunduz and E. Erkip, “Opportunistic cooperation by dynamic resource allocation,” IEEE Transactions on Wireless Communications, vol. 6, no. 4, 2007. [14] D. N. C. Tse and S. V. Hanly, “Multiaccess fading channels. I. polymatroid structure, optimal resource allocation and throughput capacities,” IEEE Transactions on Information Theory, vol. 44, no. 7, pp. 2796–2815, 1998. [15] Y. Mei, Y.-H. Lu, Y. C. Hu, and C. G. Lee, “Energy-efficient motion planning for mobile robots,” in Robotics and Automation, 2004. Proceedings. ICRA’04. 2004 IEEE International Conference on, vol. 5, pp. 4344–4349, IEEE, 2004. [16] U. Ali, H. Cai, Y. Mostofi, and Y. Wardi, “Motion and communication co-optimization with path planning and online channel estimation,” arXiv preprint arXiv:1603.01672, 2016. [17] Y. Zeng and R. Zhang, “Energy-efficient UAV communication with trajectory optimization,” IEEE Transactions on Wireless Communications, vol. 16, no. 6, pp. 3747–3760, 2017. [18] P. Falugi, E. Kerrigan, and E. Van Wyk, “Imperial College London Optimal Control Software User Guide (ICLOCS).” http://www.ee.ic.ac.uk/ICLOCS/, 2010. [19] S. Boyd and L. Vandenberghe, Convex optimization. Cambridge university press, 2004. [20] A. L. Yuille and A. Rangarajan, “The concave-convex procedure,” Neural computation, vol. 15, no. 4, pp. 915–936, 2003. [21] S. Wolf, “An introduction to duality in convex optimization,” Network, vol. 153, 2011. [22] R. S. Cheng and S. Verdú, “Gaussian multiaccess channels with ISI: Capacity region and multiuser water-filling,” IEEE Transactions on Information Theory, vol. 39, no. 3, pp. 773–785, 1993. [23] S. Nazemi, K. K. Leung, and A. Swami, “QoI-aware tradeoff between communication and computation in wireless ad-hoc networks,” in Personal, Indoor, and Mobile Radio Communications (PIMRC), 2016 IEEE 27th Annual International Symposium on, pp. 1–6, IEEE, 2016. 

3cs.SY

AMAS: optimizing the partition and filtration of adaptive seeds to speed up read mapping Ngoc Hieu Tran1,* Email: nhtran@ntu.edu.sg 1 Xin Chen Email: chenxin@ntu.edu.sg 1 School of Physical and Mathematical Sciences, Nanyang Technological University, Singapore * Corresponding author 1 ABSTRACT Background: Identifying all possible mapping locations of next-generation sequencing (NGS) reads is highly essential in several applications such as prediction of genomic variants or protein binding motifs located in repeat regions, isoform expression quantification, metagenomics analysis, etc. However, this task is very time-consuming and majority of mapping tools only focus on one or a few best mapping locations. Results: We propose AMAS, an alignment tool specialized in identifying all possible mapping locations of NGS reads in a reference sequence. AMAS features an effective use of adaptive seeds to speed up read mapping while preserving sensitivity. Specifically, an index is designed to pre-store the locations of adaptive seeds in the reference sequence, efficiently reducing the time for seed matching and partitioning. An accurate filtration of adaptive seeds is further applied to substantially tighten the candidate alignment space. As a result, AMAS runs several times faster than other state-of-the-art read mappers while achieving similar accuracy. Conclusions: AMAS provides a valuable resource to speed up the important yet time-consuming task of identifying all mapping locations of NGS reads. AMAS is implemented in C++ based on the SeqAn library and is freely available at https://sourceforge.net/projects/ngsamas/. Keywords: next-generation sequencing, read mapping, sequence alignment, adaptive seeds, seed partition, filtration 2 Background Recent advances in next-generation sequencing (NGS) technologies have produced massive amounts of short reads data, bringing up promising opportunities in several biomedical research areas such as RNA-seq, ChIP-seq, de novo genome sequencing, resequencing, metagenome sequencing, etc [1]. In many applications, the first key step is to map NGS reads to a reference sequence of interest. Thus, dozens of rapid alignment algorithms have been developed to improve the read mapping [2, 3], especially to address the problem of errors tolerance (mismatches, indels) and the increasing read length. To achieve an optimal combination of speed, accuracy, and memoryefficiency, many popular mapping tools such as Bowtie 1, 2 [4, 5], BWA [6-8], etc, apply heuristic approaches to prune the search space, prioritize candidate locations, and finally return one or a few best mapping locations for each read. They can be referred to as best-mappers. Another class of mapping tools, which can be referred to as all-mappers, are specialized in identifying as many as possible, if not all, matches within a reasonable time. This all-mapping task is highly essential in several applications such as prediction of genomic variants or protein binding motifs located in repeat regions, isoform expression quantification, metagenomics analysis [9-12]. Some prominent allmappers include SOAP 2 [13], SHRiMP 2 [14], mrsFAST [10, 15], mrFAST [9, 16], RazerS 3 [17], GEM [18], Masai [19], Hobbes 1, 2 [20, 21]. Exhaustive search for all mapping locations of NGS reads in a reference sequence is a computationally intensive task. The problem is even more complicated for large genomes with highly repetitive regions, e.g. the human genome, and when sequencing errors as well as genomic variants are taken into account. In order to narrow down the search space, most all-mappers are designed with the seed-and-extend strategy. In particular, a read is first partitioned into a few non-overlapping seeds. If the read can be mapped to the reference sequence with some allowed errors, its seeds must also have exact/approximate matches in the reference sequence, and moreover, the read’s mapping locations can be found in the neighbourhood regions of its seeds’ locations. Hence, the seeds are quickly mapped to the reference sequence and their identified locations are then used as candidates for further extending the alignment to the rest of the read by using standard dynamic programming algorithms such as Smith-Waterman [22] or Needleman-Wunsch [23]. The mapping sensitivity is guaranteed by the pigeonhole principle. For example, to map a read to the reference sequence with up to e errors, one can partition it into e+1 non-overlapping seeds. By the pigeonhole principle, at least one of those e+1 seeds must have an exact match in the reference sequence. Thus, all possible mapping locations with up to e errors of the read can be found by extending all exact matches of those seeds. Most all-mappers partition a read into seeds of equal lengths, where the seed length is specified by users or may also be determined from the number of errors allowed and the pigeonhole principle. In any cases, the seed length is several times shorter than the read length and the seeds are usually mapped exactly (i.e. without errors) to the reference sequence. To further speed up the seed mapping process, most all-mappers (except for RazerS 3) index the reference sequence using some special data structures. In particular, SHRiMP 2, mrsFAST, mrFAST, and Hobbes use hash tables to index 3 the reference sequence. For a fixed length k (which is less than or equal to the seed length), the hash tables contain all possible k-mers as the keys and each key is associated with a list of locations where the corresponding k-mer is observed in the reference sequence. Masai, on the other hand, constructs a suffix tree of the reference sequence to perform the seed matching. Both suffix array and the much more memory-efficient Ferragina-Manzini (FM) index [24] with Burrows-Wheeler Transform (BWT) can be used in Masai to mimic the suffix tree top-down traverse. Hash tables can also be combined to accelerate the searching in the BWT-FM index, as demonstrated in SOAP 2. Among those commonly used data structures, the FM index is well-known for its small memory footprint, which is only ~3-4 GB for the human genome. Interestingly, in the latest version of mrsFAST [15], the authors proposed a new compact structure for hash tables significantly reducing the index size to ~2GB for the human genome, which is even smaller than the FM index. In addition to indexing the reference sequence, some all-mappers also index the input reads so that multiple reads can be handled simultaneously rather than in a one-by-one fashion. For instance, in the read index of mrsFAST, each k-mer is associated with a list of read IDs, locations of the k-mer in the reads, etc. In the extension step, for each k-mer, the two lists (from the reference index and from the read index) are compared in a divide-and-conquer fashion to achieve cache efficiency. Similarly, in Masai, all seeds of the input reads are organized in a radix tree. Then, by walking through the reference suffix tree along with the read radix tree, all seeds can be mapped to the reference sequence simultaneously. After the seeds of a read have been mapped to the reference sequence, their identified locations are then used as candidates for further extending the alignment to the rest of the read. This step is usually called the extension or the verification step, and is performed by using standard dynamic programming algorithms such as Smith-Waterman [22] or Needleman-Wunsch [23]. In fact, this step is the bottleneck and consumes the most computations of the whole mapping task. As the seed length is several orders of magnitudes shorter than the reference length and the reference sequence often contains highly repetitive regions, the candidate space generated by the seeds is huge. Moreover, majority of the locations actually are false positives which will not lead to acceptable alignment results yet waste computing resources. Hence, filtration strategies are needed to reduce the number of locations The first filtration strategy widely used by the all-mappers is based on the pigeonhole principle. In particular, if a read is aligned with up to e errors and exact seed matching is employed, only e+1 nonoverlapping seeds are required to guarantee the full mapping sensitivity. More advanced strategies are further implemented in different all-mappers. For instance, the FastHASH algorithm in the latest version of mrFAST [16] applies two procedures called Cheap K-mer Selection and Adjacency Filtering to reduce the number of locations while preserving sensitivity. In the first procedure, the seeds of a read are sorted according to their frequencies in the reference sequence and e+1 seeds with the lowest frequencies are selected. The second procedure filters out the obviously false locations based on the idea that the potentially true locations (i.e. those that may return correct alignments) of adjacent seeds must be close to each other in the reference sequence. 4 In Hobbes [20] the authors proposed a more complicated procedure of seed selection. Instead of partitioning the read into contiguous seeds, Hobbes considers all possible partitions and applies a dynamic programming to select the one that minimizes the sum of frequencies of the seeds. In the latest version, Hobbes 2 [21], the authors further suggested to use extra seeds to improve filtering specificity: instead of e+1, e+2 (or more) non-overlapping seeds can be employed. Then, by the pigeonhole principle, at least two of those seeds must have exact matches in the reference sequence and hence only those locations reported by at least two seeds will be selected as candidates for the extension step. This technique has also been applied previously in GSNAP [25]. The filtration in Masai [19] is based on the use of long approximate seeds rather than short exact seeds to achieve better specificity. For example, if a read of length 100bp is aligned to the reference sequence with up to e=5 errors, by the pigeonhole principle, it can be partitioned into 6 nonoverlapping seeds of length 16bp and the seeds are mapped exactly to the reference genome. To increase the seed length, Masai partitions the read into 3 seeds of 33bp and each seed is mapped to the reference genome with up to 1 error. The approximate matching of all seeds is performed simultaneously by applying a multiple backtracking algorithm on the reference index and the read index. It is worth to note that most all-mappers partition a read into equal-length seeds and prefer to use short k-mers as seeds. However, the occurrences of equal-length seeds in a reference sequence may not be uniformly distributed and the seed frequencies may differ by orders of magnitudes (see Supplementary Figure S1 for the frequency distribution of 10-mers in the human genome). Highly repetitive seeds may substantially enlarge the candidate space with false positives and hence waste computing resources. This problem was addressed by introducing seed selection mechanisms in mrFAST, Hobbes, or by using long approximate seeds in Masai. However, the seed frequencies are still not fully controllable. A more efficient strategy of seed partition was proposed in GEM [18] using adaptive seeds. Specifically, GEM allows seeds to have variable lengths while restricting the number of candidate locations generated by each seed to be less than a predefined threshold. Figure 1A shows an example of such efficient partition where the number of candidate locations of the read was reduced by 44.66 times by using adaptive seeds instead of equal-length seeds. In general, adaptive seeds provide better control on the candidate space while consuming similar computing resources as equal-length seeds. Similar ideas have also been used in some best-mappers [7, 8, 26, 27] to avoid unnecessary extension for highly repetitive seeds. As we have briefly reviewed above, intensive efforts have been put into optimizing the use of equal-length seeds in the all-mapping task, including sequence indexes for fast exact/approximate seed matching and filtration strategies to reduce candidate locations. However, little attention has been paid into adaptive seeds and their advantages have not been fully explored to speed up the read mapping. Hence, in this work we carefully study the strengths and weaknesses of adaptive seeds and attempt to find an optimal strategy for their mapping, partition and filtration. The main contributions of our all-mapping tool with adaptive seeds, AMAS, are summarized below. 5 Firstly, to speed up the mapping of adaptive seeds, we pre-compute all possible adaptive seeds and their locations in the reference sequence for a given frequency threshold. Then, our index of the reference sequence consists of all possible adaptive seeds as the keys and each key is associated with a list of locations where the corresponding seed is observed. The index is stored in a local file and can be reused for every mapping task. The idea is similar to hash tables, but instead of fix-length k-mers, the keys in our index are variable-length substrings with frequencies less than a predefined threshold. Secondly, in the adaptive seed partition, the number of seeds is not controllable and varies across different reads. For some reads, there may not be sufficient seeds for the pigeonhole principle to guarantee the mapping sensitivity. Hence, we refine the partition of adaptive seeds by adding one more constraint on the seed lengths in order to guarantee the minimum number of seeds in the partition. On the other hand, there may also be more than enough seeds available from the partition. In such cases, we make effective use of the extra seeds to filter out false positives from the candidate space. Thirdly, due to the shortage of nucleotides assigned to the last seed in the adaptive partition of a read, that seed may be too short and the number of candidate locations it reported may far exceed the desired threshold. Majority of such locations are false positives which increase the candidate space dramatically. However, blindly ignoring all of the last seeds will lead to a considerable loss of mapping sensitivity. Hence, we also pay special attention to an accurate filtration of the last seeds. Overall, by optimizing the partition and filtration of adaptive seeds, our tool AMAS runs several times faster than other state-of-the-art all-mappers while achieving comparable sensitivity and accuracy. Detailed methods and performance results are presented in the next sections. AMAS was implemented in C++ based on the source code of Masai and the SeqAn library. Our main contributions include the index, the partition, and the filtration of adaptive seeds. For the extension step with edit distance, we borrowed Masai’s implementation [17, 19] of the Myers’ bitvector dynamic programming algorithm [28]. We also borrowed I/O components of Masai for handling the reference, the read sequences, and the alignment results in SAM format. AMAS is freely available at https://sourceforge.net/projects/ngsamas/. A user guide and sample data sets are also provided. Methods To find an effective strategy for the partition and filtration of adaptive seeds, we first performed experiments on a data set of 100k reads of length 100bp, which were simulated from the human genome (UCSC hg19) using Mason [29] with the Illumina model. The reads were mapped to the human genome with up to e=5 errors. In the following, we shall describe our strategy and use the simulated data set for illustration. Partition of adaptive seeds An example of adaptive seeds versus equal-length seeds is illustrated in Figure 1A. To partition a read into adaptive seeds, we scan the read against the reference sequence in the forward (left-to-right) 6 direction, base by base. When the number of matches of the current seed drops below a predefined frequency threshold F, its locations are added to the candidate space and a new seed is started. That process is continued until the desired number of seeds is obtained or the read’s end is reached. In the latter case, the number of matches of the last seed might be still higher than the threshold F. In the example in Figure 1A, the read was mapped to the human genome and F was set at 300. The number of candidate locations of the read was reduced by 44.66 times by using 6 adaptive seeds instead of 6 seeds of equal lengths 16bp. Similarly, for the whole simulated data set of 100k reads, the total number of candidate locations was reduced by 6.61 times (Supplementary Table S1). Index of adaptive seeds To search for the locations of adaptive seeds in the reference sequence, GEM uses the FM index to achieve both speed and low memory footprint [18]. As all-mapping is a very time-consuming task, it may be desirable to further boost up the mapping speed at the expense of a reasonable amount of memory, like using hash tables in Hobbes, mrsFAST, mrFAST, or suffix arrays, enhanced suffix arrays in Masai. Here we pre-compute all possible adaptive seeds and their locations in the reference sequence for a given frequency threshold F. Then, our index of the reference sequence consists of all possible adaptive seeds as the keys and each key is associated with a list of locations where the corresponding seed is observed. The idea is similar to hash tables, but instead of fix-length k-mers, the keys in our index are variable-length substrings with frequencies less than a predefined threshold F. To represent all possible adaptive seeds in the reference sequence, we implement a lookup table and extend it with a radix tree (Figure 1B). In particular, we first pick a length k, organize all k-mers in the lookup table and search for their locations in the reference sequence. Then, for each k-mer, we further extend the search to its 4 child nodes which correspond to adding the 4 bases “A”, “C”, “G”, and “T”. The search and extension are continued until the numbers of matches of the nodes drop below the threshold F (shaded leaf nodes in Figure 1B). Thus, all leaf nodes in the radix tree correspond to all possible adaptive seeds and each of them is linked to an array of its locations in the reference sequence. A similar idea has been used in SOAP 2 [13] where a hash table was combined to accelerate the searching in the FM index. Note that the leaf nodes in our index represent adaptive seeds with desired frequencies, not suffixes like in common suffix-array based data structures. By default, we consider adaptive seeds with a minimum length of k=10 bases. In general, using lower values of F helps to tighten the candidate space reported by the seeds but also enlarges the index. We choose F to be about 10 times less than the average frequency of k-mers in the reference sequence. For instance, we set F=300 for the human genome (Figure 1B), F=10 for the worm genome, and F=17 for the fruit-fly genome (Supplementary Table S2). The index only needs to be built once for each genome and it can be stored on a local disk for future mapping tasks. 7 Constraint on the maximum seed length As we use exact seed matching, by the pigeonhole principle, at least e+1 seeds are needed to achieve full sensitivity for mapping a read with up to e errors. The number of adaptive seeds, however, is not controllable and varies across different reads (see Supplementary Figure S2 for the seed count’s distribution of the simulated data set). Some reads may not have sufficient seeds to guarantee the mapping sensitivity. Hence, in addition to the frequency constraint, we also impose a constraint on the maximum length of the seeds. The maximum length threshold l allows to control the minimum number of seeds in the partition and hence to improve the mapping sensitivity. For example, for a read of length 100bp, setting l=33 guarantees that there will be at least 3 seeds in the partition and all matching locations with up to 2 errors will be found (Supplementary Figure S2). This feature greatly improves the sensitivity of our tool over GEM, as we will show later in the section Results. Filtration using the last seeds in the adaptive partition We also pay special attention to the last seeds in the adaptive partition. Due to the lack of nucleotides remaining, the last seed in the adaptive partition of a read may be too short and its number of matches may far exceed those of the other seeds as well as the threshold F. For instance, in the simulated data set, the number of candidate locations reported by the last seeds contributed up to 64% of the total candidate space. Apparently, majority of those locations were false positives, but removing all of them may lead to a considerable loss of mapping sensitivity. Our tool chooses to filter out only those last seeds whose numbers of matches are higher than the expected frequency of kmers and contribute to more than 95% of the candidate locations of their respective reads (i.e. the last seed reports too many locations compared to other seeds in the same read). Nevertheless, we avoid removing the last seeds from those reads that have too few seeds required by the pigeonhole principle. For the simulated data set, this filtration step only affected a small amount of reads (1.6%, Supplementary Figure S3), but significantly reduced the total candidate space by more than 50% (Supplementary Table S1). Filtration using the extra seeds in the adaptive partition Last but not least, when there are more than e+1 seeds available from the adaptive partition, we make effective use of the extra seeds to filter out false positives. For instance, the simulated reads were mapped with up to e=5 errors and hence e+1=6 seeds were required. However, the reads could be partitioned up to 9 seeds, where majority of them were partitioned into 8 seeds (Supplementary Figure S2). If a read was partitioned into 7, 8, or 9 seeds, only those locations reported by at least 2, 3, or 4 seeds, respectively, were selected as candidates. This filtration step further reduced the total candidate space by more than 70% (Supplementary Table S1). Our tool stores candidate locations in a binary search tree that allows quickly sorting and identifying those reported by multiple seeds. Using more extra seeds increases the filtering specificity but also requires longer filtration time. Our experiment with the simulated data set showed that using only one extra seed, i.e. total e+2=7 seeds, 8 achieved the best overall mapping time (Supplementary Table S1). This is similar to the observation in Hobbes 2 [21], GSNAP [25]. Overall, the combination of two filtration steps using the last seeds and the extra seeds substantially reduced the candidate space of the simulated data set by more than 85% (Supplementary Table S1). Obviously, such effective filtration will save a lot of computing resources in the extension step. Seed extension To perform seed extension with edit distance (i.e. including both mismatches and indels), we reused Masai’s implementation [17, 19] of the Myers’ bit-vector dynamic programming algorithm [28]. Results and discussion In this section, we compare the performance of AMAS and the latest state-of-the-art all-mappers GEM [18], Masai [19], Hobbes 2 [21], mrFAST (with FastHASH) [16], and mrsFAST-ultra [15]. We also included two popular best-mappers Bowtie 2 and BWA in the comparison. The mappers were configured to search for all possible mapping locations with up to e=5 errors and then output the alignment results in the SAM format. Details of parameters configuration can be found in the Supplementary Table S3. The experiments were performed on a Linux server with 12 Intel Xeon processors (E5-2640, 2.50 GHz), 64 GB of RAM, CentOS 6.3, GCC 4.4.7. All mappers were tested using one single thread and eight threads, except for Masai and mrFAST which do not support multithreading. Rabema benchmark Following previous studies of Masai and Hobbes 2 [19, 21], we used the Rabema benchmark [30] to measure the sensitivity of the mappers. First, RazerS 3 [17] was run in its full-sensitive mode to build the gold-standard set of all possible mapping locations with up to 5 errors. Any other mappers that can guarantee full mapping sensitivity can also be used to build the gold-standard set [30]. The goldstandard set was then used by Rabema to assess the sensitivity of each mapper in three categories “All”, “All-best”, and “Any-best”. For the “All” category, the mappers need to find all mapping locations in the gold-standard set. Similarly, for the “All-best” and “Any-best” categories, the mappers need to find all of the best mapping locations and any of the best mapping locations of the reads, respectively. Note that the term “best” here is in terms of the edit distance between the read and the reference sequence. Some mappers such as Bowtie 2 or BWA have their own scoring schemes to evaluate the mapping quality of the alignments. As original locations were available for the simulated reads, we also assessed the recall and the precision of the mappers on the simulated data set. Recall is defined in Rabema as the fraction of the input reads that were correctly mapped and precision is defined as the fraction of the uniquely mapped reads that were correctly mapped. A read is said to be mapped correctly if its original location was reported by the mapper. A read is said to be mapped uniquely if only one location was reported by the mapper. 9 Performance on the simulated data set Table 1 shows the Rabema benchmark for mapping the simulated data set of 100k reads to the human genome (UCSC hg19). The column “All” shows that AMAS was able to identify 98.11% of all mapping locations and achieved full sensitivity (100%) for those locations with up to 2 errors. Hobbes 2 achieved the highest sensitivity 99.85%, i.e. 1.74% more sensitive than AMAS, and was followed by Masai (99.83%), mrFAST (99.33%), AMAS (98.11%), GEM (97.65%), and mrsFAST (78.52%). AMAS and GEM had lower sensitivity than Masai, Hobbes 2, and mrFAST, especially when searching for mapping locations with 4 or 5 errors. This is because the adaptive seed partition does not guarantee the number of seeds and hence there may not be sufficient seeds required by the pigeonhole principle. The additional constraint on the maximum seed length introduced in AMAS guaranteed the full sensitivity (100%) for all mapping locations with up to 2 errors and also improved the overall sensitivity for locations with 3, 4, 5 errors. Hence, AMAS was 0.46% more sensitive than GEM. mrsFAST had much lower mapping rate (79.35%) and sensitivity (78.52%) than the others because it only allows mismatches and cannot detect indels. In the “All-best” and “Any-best” categories, AMAS achieved 99.99% and 100% sensitivity, respectively. That was similar to Hobbes 2 and better than the rest. AMAS also achieved the best recall rate 99.09%, while its precision was 99.99%, the second best among the tools. The best-mappers Bowtie 2, BWA, and GEM-best only searched for one or a few best mapping locations. Hence, they ran faster than the all-mappers while having lower sensitivity. We noticed that GEM-best outperformed Bowtie 2 and BWA in terms of both running time and sensitivity, demonstrating the advantages of adaptive seeds. We tried to configure Bowtie 2 and BWA as allmappers (Supplementary Table S3) but their running times were too long and the results were not reported here. We also configured Bowtie 2 in its -k mode to report up to -k=100 alignments for each read so that it could finish within a reasonable time. Its sensitivity was 96.04%, i.e. 2.07% lower than AMAS. Performance on real data sets Next, we run the mappers on a real NGS data set SRR063408 from the 1000 Genomes Project (individual HG01495). The data set includes 25.6 million reads of length 100bp which were generated from the Illumina Genome Analyzer II. Table 2 shows the running time, memory footprint, and Rabema benchmark of mapping the first one million reads to the human genome. AMAS took about 30 minutes to finish the job using one single thread and 5 minutes using eight threads. AMAS was able to map 93.75% of the reads and identified 98.14% of all mapping locations, including nearly all of the best mapping locations (99.98%). AMAS only lose behind Masai and Hobbes 2, which achieved highest sensitivity, 99.93% and 99.90% of all mapping locations, 100% of all best mapping locations. In both single-thread and eight-thread tests, AMAS was more than 3 times faster than Hobbes 2. AMAS was also more than 3 times faster than Masai in the single-thread test (Masai does not support multi-threading). While using the same idea of adaptive seeds, AMAS outperformed GEM in terms of mapping rate (0.06% higher), sensitivity (0.25% higher in the “All” category), and running time (2-7 10 times faster). This clearly shows the benefits of the optimized partition and filtration implemented in AMAS. AMAS was more than 6 times faster than mrFAST in the single-thread test (mrFAST does not support multi-threading). In this experiment, mrFAST produced invalid SAM output with many inconsistent CIGAR strings, which could not be converted to BAM for further analysis with Rabema. mrsFAST had similar running time as AMAS in the single-thread test but was more than 2 times slower than AMAS in the eight-thread test. AMAS also had better mapping rate (1.37% higher) and sensitivity (1.02% higher in the “All” category) than mrsFAST. To map the full data set SRR063408 of 25.6 million reads to the human genome, AMAS took about 12 hours using one single thread and 2 hours using eight threads (Table 3). Masai required 31 hours and Hobbes 2 required nearly 50 hours, that is, 2.6 times and 4.2 times slower than AMAS when using one single thread. When eight threads were deployed, Hobbes 2 took about 8 hours, 4 times slower than AMAS. mrsFAST was the fastest in the single-thread test (9 hours), but was 2 times slower than AMAS in the eight-thread test (4 hours). mrFAST could not finish the mapping within 3 days, thus we did not record its results. For this data set, GEM was not able to output its alignment results in the SAM format. Thus we were only able to record its mapping time, which was already 1.55 times slower than the total running time of AMAS. We also performed similar experiments on Caenorhabditis elegans (data set SRR065388 mapped to the worm genome UCSC ce10) and Drosophila melanogaster (data set SRR497711 mapped to the fruit-fly genome UCSC dm3). The results are presented in Supplementary Tables S4-S7. AMAS took 9 minutes to map 36 million reads to the worm genome. AMAS outperformed GEM and mrsFAST in terms of running time, mapping rate, and sensitivity. It was nearly 2 times faster than Hobbes 2, 2.5 times faster than Masai and 4.6 times faster than mrFAST. AMAS fell behind Masai and Hobbes 2 by no more than 0.04% of reads and 1.35% of all mapping locations. Similar results were also obtained for the fruit-fly genome. Most all-mappers use the seed-and-extend strategy in which locations reported by the seeds of a read are used as candidates for extending the alignment to the rest of the read. Many locations reported by different seeds of a read may actually correspond to one single location of the read in the reference sequence. Thus it is important to mark duplicate candidates to reduce the time for seed extension, and more importantly, to avoid duplicate alignments in the output. We measured the percentage of the true alignments (according to Rabema benchmark) found by the mappers among the alignments they reported. Table 4 shows that AMAS and mrsFAST were the most accurate while Masai reported quite a large number of duplicate alignments. Finally, we tested the mappers on a longer read length and more errors. A data set of one million reads of length 200bp were simulated from the human genome and then were mapped with up to e=10 errors (i.e. 5% error rate). The results are shown in Table 5. AMAS was able to map 99.90% of the reads and identified 99.02% of all mapping locations, including nearly all (99.99%) of the best mapping locations. Masai and Hobbes 2 achieved full sensitivity, but were 4-5 times slower than AMAS. GEM did not scale well with this experiment. In the single-thread test its memory footprint increased dramatically and in the eight-thread test it crashed due to memory errors. mrFAST was 3 11 times slower than AMAS and produced invalid SAM output. mrsFAST had similar running time as AMAS, but much lower mapping rate (62.56%) and sensitivity (61.42% of all mapping locations) because mrsFAST cannot handle indels. Index and memory footprint As AMAS processes reads in blocks of one million, the memory foot-print is kept stable regardless of the size of the input NGS data (Tables 2 and 3). To map the full data set SRR063408 of 25.6M 100bp reads to the human genome, AMAS requires 19.5 GB. The memory consumption includes 15.9 GB for the index and 3 GB for the reference sequence. The memory footprint is not well optimized yet and shall be the focus of our future development. Hobbes 2 uses hash tables and requires 16.7 GB. Masai uses suffix arrays and requires 25 GB. GEM uses FM index with much less memory footprint 5.7 GB for mapping. However, the memory footprint of GEM increases significantly or even crashes when outputting the alignments to SAM format (using “gem-to-sam” tool) or when handling longer reads with more errors (Tables 3, 5). mrsFAST uses hash tables but only needs 6.6 GB. Note that for smaller data sets (with 1M reads), mrsFAST and mrFAST have very low memory footprint, 2 GB, which is even less than the FM index (Table 2). Our index of the human genome takes about 3 hours to build, and it can be stored on a local disk with 13.8 GB for future mapping tasks (Supplementary Table S2). As our index needs to identify all possible adaptive seeds with desired frequencies and their locations in the reference sequence, it may take more time to build than other indexes. However, this indexing task only needs to be done once for each genome and does not affect the mapping time. Conclusions In this study, we proposed an effective strategy for the partition and filtration of adaptive seeds to speed up the exhaustive search for all possible mapping locations of NGS reads in a reference sequence. Our tool, AMAS, features a new index designed specifically to accelerate the alignment of adaptive seeds to the reference sequence. Frequency and length constraints on adaptive seeds allow better control on the candidate alignment space while preserving enough seeds to guarantee the mapping sensitivity. Finally, an accurate filtration strategy based on the last seeds and the extra seeds is further applied to substantially tighten the candidate alignment space, efficiently reducing the time for seed extension. As a result, AMAS runs several times faster than other competitors while achieving comparable sensitivity and accuracy. Multi-mapping reads are highly essential in several applications such as prediction of genomic variants or protein binding motifs located in repeat regions, isoform expression quantification, metagenomics analysis, etc. However, searching for them is a very time-consuming task. Current stateof-the-art all-mappers such as GEM, Masai, Hobbes 2, mrsFAST, mrFAST, and best-mappers such as Bowtie 2, BWA, etc, may take up to 1-2 days to find all possible mapping locations of reads from a single NGS run for the human genome. Our tool offers a more efficient way to significantly speed up that process. We have shown that AMAS run 2-7 times faster and was more sensitive than GEM. 12 AMAS was also up to 4 times faster than Masai, Hobbes 2, and mrFAST. Although AMAS did not achieve nearly full sensitivity like Masai, Hobbes 2, and mrFAST, it only missed out less than 1.76% of all mapping locations and 0.02% of all best mapping locations, while guaranteeing full sensitivity for matches with small numbers of errors. Most of missing locations by AMAS were those with high numbers of errors and may not be of interest in practice. AMAS scales well in multi-threading environment and keeps the memory footprint stable regardless of the size of input data. AMAS also supports pair-end read mapping, best-mapping mode, and –k mapping mode (reporting up to –k alignments for each read). Hence, we believe that AMAS will become a useful resource for read mapping in NGS data analysis. Competing interests The authors declare that they have no competing interests. Authors’ contributions XC and NHT designed the project. NHT developed the software and performed the analysis. NHT and XC wrote the manuscript. Both authors read and approved the manuscript. Acknowledgements We thank the SeqAn team, especially the authors of Masai package for making their source codes available. Our tool uses the SeqAn library and some well-developed components of Masai. This work was supported by the Singapore National Medical Research Council (CBRG11nov091) and the Singapore Ministry of Education Academic Research Fund (MOE2012-T2-1-055) References 1. 2. 3. 4. 5. 6. 7. 8. 9. Metzker ML: Sequencing technologies - the next generation. Nat Rev Genet 2010, 11(1):3146. Fonseca NA, Rung J, Brazma A, Marioni JC: Tools for mapping high-throughput sequencing data. Bioinformatics 2012, 28(24):3169-3177. Li H, Homer N: A survey of sequence alignment algorithms for next-generation sequencing. Brief Bioinform 2010, 11(5):473-483. Langmead B, Salzberg SL: Fast gapped-read alignment with Bowtie 2. Nat Methods 2012, 9(4):357-359. Langmead B, Trapnell C, Pop M, Salzberg SL: Ultrafast and memory-efficient alignment of short DNA sequences to the human genome. Genome Biol 2009, 10(3):R25. Li H, Durbin R: Fast and accurate short read alignment with Burrows-Wheeler transform. Bioinformatics 2009, 25(14):1754-1760. Li H, Durbin R: Fast and accurate long-read alignment with Burrows-Wheeler transform. Bioinformatics 2010, 26(5):589-595. Li H: Aligning sequence reads, clone sequences and assembly contigs with BWA-MEM. arXiv: 13033997 [q-bioGN] 2013. Alkan C, Kidd JM, Marques-Bonet T, Aksay G, Antonacci F, Hormozdiari F, Kitzman JO, Baker C, Malig M, Mutlu O et al: Personalized copy number and segmental duplication maps using next-generation sequencing. Nat Genet 2009, 41(10):1061-1067. 13 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21. 22. 23. 24. 25. 26. 27. 28. 29. 30. Hach F, Hormozdiari F, Alkan C, Birol I, Eichler EE, Sahinalp SC: mrsFAST: a cache-oblivious algorithm for short-read mapping. Nat Methods 2010, 7(8):576-577. Newkirk D, Biesinger J, Chon A, Yokomori K, Xie X: AREM: aligning short reads from ChIPsequencing by expectation maximization. J Comput Biol 2011, 18(11):1495-1505. Roberts A, Pachter L: Streaming fragment assignment for real-time analysis of sequencing experiments. Nat Methods 2013, 10(1):71-73. Li R, Yu C, Li Y, Lam TW, Yiu SM, Kristiansen K, Wang J: SOAP2: an improved ultrafast tool for short read alignment. Bioinformatics 2009, 25(15):1966-1967. David M, Dzamba M, Lister D, Ilie L, Brudno M: SHRiMP2: sensitive yet practical SHort Read Mapping. Bioinformatics 2011, 27(7):1011-1012. Hach F, Sarrafi I, Hormozdiari F, Alkan C, Eichler EE, Sahinalp SC: mrsFAST-Ultra: a compact, SNP-aware mapper for high performance sequencing applications. Nucleic Acids Res 2014, 42(Web Server issue):W494-500. Xin H, Lee D, Hormozdiari F, Yedkar S, Mutlu O, Alkan C: Accelerating read mapping with FastHASH. BMC Genomics 2013, 14 Suppl 1:S13. Weese D, Holtgrewe M, Reinert K: RazerS 3: faster, fully sensitive read mapping. Bioinformatics 2012, 28(20):2592-2599. Marco-Sola S, Sammeth M, Guigo R, Ribeca P: The GEM mapper: fast, accurate and versatile alignment by filtration. Nat Methods 2012, 9(12):1185-1188. Siragusa E, Weese D, Reinert K: Fast and accurate read mapping with approximate seeds and multiple backtracking. Nucleic Acids Res 2013, 41(7):e78. Ahmadi A, Behm A, Honnalli N, Li C, Weng L, Xie X: Hobbes: optimized gram-based methods for efficient read alignment. Nucleic Acids Res 2012, 40(6):e41. Kim J, Li C, Xie X: Improving read mapping using additional prefix grams. BMC Bioinformatics 2014, 15:42. Smith TF, Waterman MS: Identification of common molecular subsequences. J Mol Biol 1981, 147(1):195-197. Needleman SB, Wunsch CD: A general method applicable to the search for similarities in the amino acid sequence of two proteins. J Mol Biol 1970, 48(3):443-453. Ferragina P, Manzini G: Opportunistic data structures with applications. In: Proceedings of the 41st Annual Symposium on Foundations of Computer Science. IEEE Computer Society; 2000: 390. Wu TD, Nacu S: Fast and SNP-tolerant detection of complex variants and splicing in short reads. Bioinformatics 2010, 26(7):873-881. Kielbasa SM, Wan R, Sato K, Horton P, Frith MC: Adaptive seeds tame genomic sequence comparison. Genome Res 2011, 21(3):487-493. Liu Y, Popp B, Schmidt B: CUSHAW3: sensitive and accurate base-space and color-space short-read alignment with hybrid seeding. PLoS One 2014, 9(1):e86869. Myers G: A fast bit-vector algorithm for approximate string matching based on dynamic programming. J ACM 1999, 46:395-415. Holtgrewe M: Mason - a read simulator for second generation sequencing data. Technical report TR-B-10-06, Institut fur Mathematik und Informatik, Freie Universitat Berlin 2010. Holtgrewe M, Emde AK, Weese D, Reinert K: A novel and well-defined benchmarking method for second generation read mapping. BMC Bioinformatics 2011, 12:210. 14 Figure Legends Figure 1. Partition of adaptive seeds. (A) Partition of the read “AGGTCAAGAAATTGAGACCATCCCGGCCAACTTGGTGAAACCCCGTCTCGTCTCTACTAAAAGT AAAAAAAATTAGCTGGGCGTGGTGGCGGGTGCATAT” against the human genome using 6 equallength16bp seeds (upper) and 6 adaptive seeds with frequencies less than F=300 (lower). Using adaptive seeds significantly reduced the total number of candidate locations by ~44.66 times. (B) A partial sub-tree of our index with k=10, F=300 for the human genome. Nodes were extended until their numbers of matches dropped below 300 (shaded leaf nodes). The path corresponding to the first adaptive seed of the read in (A) with length 13bp and frequency 157 is illustrated with bold lines. 15 Tables Table 1. Rabema benchmark for mapping the simulated data set of 100k reads of length 100bp to the human genome with up to e=5 errors. The first seven tools are all-mappers and the last three are best-mappers. The column “All” shows the percentage of all mapping locations that the mappers were able to find. Numbers in bold show the overall percentage, numbers in the smaller font size show detailed percentages for mapping locations with 0, 1, 2 (upper) and 3, 4, 5 (lower) errors. Similarly, the two columns “All-best” and “Any-best” respectively show the sensitivity of the mappers for identifying all of the best mapping locations and any of the best mapping locations of the reads. The last two columns, “Rec” and “Prec”, show the recall and the precision of the mappers, respectively (see the main text for the definitions of recall and precision). Mappers Reads mapped (%) AMAS 99.11 All (%) 100 Bowtie2 (k=100) 99.06 GEM (all) 99.06 Masai 99.06 Hobbes2 mrFAST mrsFAST 98.55 79.35 Bowtie2 (best) 98.58 GEM (best) 99.06 BWA 98.04 100 98.11 100 100 Any-best (%) 100 99.99 98.71 90.76 72.53 99.97 99.82 98.73 96.04 93.02 76.36 52.47 100 99.99 99.82 97.65 99.94 99.77 99.94 99.71 99.69 99.34 67.40 100 100 100 99.74 99.38 97.46 100 100 100 99.83 99.20 98.92 97.26 100 99.87 99.78 99.64 99.24 96.52 100 100 100 99.70 99.05 98.35 99.94 97.28 100 100 100 99.33 100 100 100 99.99 87.95 100 74.50 53.65 78.52 100 99.97 99.92 100 100 100 37.93 48.97 97.91 96.59 92.00 90.12 100 100 55.16 100 74.09 52.27 51.96 20.70 100 99.29 97.25 94.19 31.88 11.55 02.46 96.58 96.27 94.61 67.59 32.80 97.91 96.68 92.71 75.48 42.92 14.71 89.43 99.92 99.70 99.65 99.27 97.50 100 99.98 99.87 99.82 99.44 96.86 100 100 100 99.70 99.05 98.47 100 100 100 100 100 99.92 100 100 100 100 100 55.31 100 74.16 52.33 31.93 11.60 02.50 100 99.47 97.53 97.25 96.13 95.49 100 99.96 99.87 99.77 99.50 96.94 100.00 99.68 99.20 93.00 80.16 68.20 99.27 94.15 93.14 92.20 100 99.86 99.76 99.85 90.26 100 79.76 96.01 79.43 100 99.44 79.72 39.40 99.89 100 99.44 100 99.97 99.95 100 100 100 99.92 99.95 99.85 100 99.88 99.85 88.26 100 100 99.59 97.27 99.11 100 All-best (%) 99.91 99.54 99.21 96.50 96.58 96.54 96.21 90.13 77.24 65.54 95.53 98.76 16 Rec (%) Prec (%) 99.09 99.99 98.40 99.76 98.72 99.73 99.05 100 98.99 99.89 98.00 99.43 78.14 98.75 94.88 96.25 98.61 99.70 93.92 95.79 Table 2. Rabema benchmark, running time, and memory footprint for mapping the first one million reads of the real Illumina data set SRR063408 to the human genome with up to e=5 errors. The read mapping was performed using 1 thread and 8 threads. Masai and mrFAST do not support multi-threading. The column “All” shows the percentage of all mapping locations that the mappers were able to find. Similarly, the two columns “All-best” and “Any-best” respectively show the sensitivity of the mappers for identifying all of the best mapping locations and any of the best mapping locations of the reads. In this experiment, mrFAST produced invalid SAM output with many inconsistent CIGAR strings, which could not be converted to BAM for further analysis. NA: not available. Mappers 1-thread Time (mm:ss) 8-thread Time (mm:ss) Memory (GB) Reads mapped (%) All (%) All-best (%) Any-best (%) AMAS 29:43 04:59 19.4 93.75 98.14 99.98 99.99 Bowtie2 (k=100) 55:30 07:31 3.6 93.70 95.89 99.60 99.91 GEM (all) 71:41 36:45 5.3 93.69 97.89 99.91 99.92 Masai 91:12 NA 16.7 93.75 99.93 100 100 Hobbes2 110:24 15:16 14.9 93.76 99.90 100 100 mrFAST 189:47 NA 2.0 NA NA NA NA mrsFAST 29:41 11:28 2.1 92.38 97.12 98.48 98.50 Table 3. Details of mapping the full data set SRR063408 of 25.6 million reads to the human genome with up to e=5 errors. The read mapping was performed using 1 thread and 8 threads. Masai and mrFAST do not support multi-threading. For this full data set, GEM was not able to output the alignment results in the SAM format. Thus we only recorded its running time and memory consumption for the mapping step, but not the writing step. mrFAST could not finish the mapping within 3 days, thus we did not record its results. NA: not available. Mappers 1-thread Time (hh:mm) 8-thread Time (hh:mm) Memory (GB) AMAS Bowtie2 (k=100) GEM* (all) Masai Hobbes2 mrFAST mrsFAST 12:09 24:23 18:07 31:02 49:56 NA 08:48 01:59 03:19 10:41 NA 08:13 NA 04:06 19.5 3.8 5.7 25.0 16.7 NA 6.6 17 Table 4. Percentage of the true alignments found by the mappers among the alignments they reported. For the real data set of one millions reads, mrFAST produced invalid SAM output with many inconsistent CIGAR strings, which could not be converted to BAM for further analysis. NA: not available. AMAS GEM (all) Masai Hobbes 2 mrFAST mrsFAST Number of true alignments found 8,886,145 8,624,695 13,813,089 13,735,743 13,806,005 10,785,780 Number of alignments reported 8,886,270 8,628,751 21,954,696 13,762,313 13,918,070 10,785,975 100 99.95 62.92 99.81 99.20 100 Number of true alignments found 99,339,852 105,459,737 148,827,680 148,526,618 NA 134,412,384 Number of alignments reported 99,447,106 105,602,112 243,162,835 148,743,923 149,368,650 134,506,255 99.89 99.87 61.21 99.85 NA 99.93 Mappers Simulated data set of 100k reads Percentage Real data set of 1M reads Percentage Table 5. Rabema benchmark, running time, and memory footprint for mapping the simulated data set of one million reads of length 200bp to the human genome with up to e=10 errors. The read mapping was performed using 1 thread and 8 threads. Masai and mrFAST do not support multithreading. The multi-threading mode of GEM also crashed for this data set due to memory problem. The column “All” shows the percentage of all mapping locations that the mappers were able to find. Similarly, the two columns “All-best” and “Any-best” respectively show the sensitivity of the mappers for identifying all of the best mapping locations and any of the best mapping locations of the reads. In this experiment, mrFAST produced invalid SAM output with many inconsistent CIGAR strings, which could not be converted to BAM for further analysis. NA: not available. Mappers 1-thread Time (mm:ss) 8-thread Time (mm:ss) Memory (GB) Reads mapped (%) All (%) All-best (%) Any-best (%) Recall (%) Precision (%) AMAS 38:58 05:55 19.6 99.90 99.02 99.99 100 99.89 100 Bowtie2 (k=100) 185:30 24:41 3.6 99.90 98.20 99.76 99.94 99.64 99.90 GEM (all) 332:47 NA 19.7 99.90 98.71 99.95 99.99 99.42 99.55 Masai 217:06 NA 17.0 99.90 100 100 100 99.90 100 Hobbes2 180:35 24:21 14.2 99.90 99.91 100 100 99.81 99.91 mrFAST 122:26 NA 2.3 NA NA NA NA NA NA mrsFAST 33:53 07:08 2.9 62.56 61.42 62.28 62.31 60.92 97.71 18 Additional files Additional file 1 – Supplementary figures Additional file 2 – Supplementary tables 19 (A) 16 16 16 16 16 16 55 3 35,778 1,585 6,534 3,246 13 15 17 13 15 24 157 134 184 293 282 7 ∑=47,201 ∑=1,057 (B) AAAAAAAAAA: 3,000,786 A: 612 A: 2,272 A: 99 A: 208 C: 386 C: 700 C: 35 C: 323 AGGTCAAGAA: 5,466 G: 438 G: 1,136 G: 165 G: 148 T: 836 T: 1,358 TTTTTTTTTT: 3,015,555 T: 24 T: 157 

5cs.CE

A LINEAR TIME ALGORITHM FOR THE 3-NEIGHBOUR TRAVELLING SALESMAN PROBLEM ON A HALIN GRAPH AND EXTENSIONS arXiv:1504.02151v5 [cs.DM] 4 Sep 2017 BRAD WOODS, ABRAHAM PUNNEN, AND TAMON STEPHEN Abstract. The Quadratic Travelling Salesman Problem (QTSP) is to find a least cost Hamiltonian cycle in an edge-weighted graph, where costs are defined for all pairs of edges contained in the Hamiltonian cycle. The problem is shown to be strongly NP-hard on a Halin graph. We also consider a variation of the QTSP, called the k-neighbour TSP (TSP(k)). Two edges e and f , e 6= f , are k-neighbours on a tour τ if and only if a shortest path (with respect to the number of edges) between e and f along τ and containing both e and f , has exactly k edges, for k ≥ 2. In (TSP(k)), a fixed nonzero cost is considered for a pair of distinct edges in the cost of a tour τ only when the edges are p-neighbours on τ for 2 ≤ p ≤ k. We give a linear time algorithm to solve TSP(k) on a Halin graph for k = 3, extending existing algorithms for the cases k = 1, 2. Our algorithm can be extended further to solve TSP(k) in polynomial time on a Halin graph with n nodes when k = O(log n). The possibility of extending our results to some fully reducible class of graphs is also discussed. TSP(k) can be used to model the Permuted Variable Length Markov Model in bioinformatics as well as an optimal routing problem for unmanned aerial vehicles (UAVs). 1. Introduction The Travelling Salesman Problem (TSP) is to find a least cost Hamiltonian cycle in an edge weighted graph. It is one of the most widely studied combinatorial optimization problems and is well-known to be NP-hard. The TSP model has been used in a wide variety of applications. For details we refer the reader to the well-known books [2, 8, 18, 23, 27] as well as the papers [4, 13, 21, 22]. For some applications, more than linear combinations of distances between consecutive nodes are desirable in formulating an objective function. Consider the problem of determining an optimal routing of an unmanned aerial vehicle (UAV) which has a list of targets at specific locations. This can be modelled as a TSP which requires a tour that minimizes the distance travelled. However, such a model neglects to take into account the physical limitations of the vehicle, such as turn radius or momentum. To illustrate this idea, in Figure 1 we give a Hamiltonian path, in Figure 2 we give the corresponding flight path, and Figure 3 shows a route which is longer but can be travelled at a greater speed and hence reducing the overall travel time. To model the traversal time, we can introduce penalties for pairs of (not necessarily adjacent) edges to force a smooth curve for its traversal. In this paper we consider a generalization of the TSP which can be used to model similar situations and contains many variations of the TSP, such as the angular-metric TSP [1] and the TSP [5] as special cases. Let G = (V, E) be an undirected graph on the node set V = {0, 1, . . . , n − 1} with the convention that all indices used hereafter are taken modulo n. For each edge (i, j) ∈ E a nonnegative cost cij is given. Let F be the set of all tours (Hamiltonian cycles) in G and let τ = (v0 , v1 , . . . , vn−1 , v0 ) ∈ F. Note that a tour τ can be represented either by its sequence (permutation) of the vertices, or by the sequence, or simply the set, of edges it traverses. In this paper, it is usually more convenient to work with the edge representation. The edges e = (vi , vi+1 ) and f = (vj , vj+1 ), e 6= f , are k-neighbours on τ , if and only if a shortest path between e and f on τ containing these edges has exactly k edges, for k ≥ 2. Here the shortest path refers to the path with the least number of edges, rather than the minimum cost path. Thus e and f are 2-neighbours in τ if and only if they share a common node in τ . Let q(e, f ) be the cost of the pair (e, f ) of edges and δ(k, τ ) = {(e, f ) : e, f ∈ τ and e and f are pneighbours on τ for some 2 ≤ p ≤ k}. Assume that q(e, f ) = q(f, e) for every pair of edges e, f ∈ E. Then the k-neighbour TSP (TSP(k)) is defined as in [31] 1 Figure 2. Smoothing of the optimal TSP tour. Figure 1. Optimal TSP tour. Figure Figure Figure 3. A Figure 3. A route which cantour be 1. Optimal Figure which can 1. Optimal TSP tour. TSP tour 2. Smoothbe travFigure 2. Smoothing of the optimal TSP tour. with respect ing of the elled more to length. tour. quickly. T SP (k) : X Minimize q(e, f ) + (e,f )∈δ(k,τ ) Subject to X travelled more quickly c(e) e∈τ τ ∈ F. A closely related problem, the Quadratic TSP (QTSP), is defined as follows: QT SP : X Minimize (e,f )∈τ ⊗τ Subject to q(e, f ) + X c(e) e∈τ τ ∈ F. where τ ⊗ τ = τ × τ \ of {(e,the e) : e ∈ τ }. Note: TSP tour. Figure 2. Smoothing optimal ( Figure 3. A route which can be iftravelled more 1 quickly. δ(n/2, τ ) n is even gure 3. τ ⊗τ = δ((n + 1)/2, τ ) if n is odd. Thus when k ≥ n/2 (for n even) or k ≥ (n + 1)/2 (for n odd), the k-neighbour TSP reduces to the Quadratic TSP [33]. Define TSP(1) to be the original TSP. Elsewhere in the literature (e.g. [17], [14]), the term Quadratic TSP is sometimes used for what we refer to as TSP(2). That is, quadratic terms are allowed, but only for pairs of edges that share a node. The bottleneck version of TSP(k) was introduced by Arkin et al. in [3], denoted as the k-neighbour maximum scatter TSP. Jäger and Molitor [19] encountered TSP(2) while studying the Permuted Variable Length Markov Model. Several heuristics are proposed and compared in [16, 19] as well as a branch and bound algorithm for TSP(2) in [16]. A column generation approach to solve TSP(2) is given in [28], lower bounding procedures discussed in [29], and polyhedral results were reported by Fischer and Helmberg [17], Fischer [14], and Fischer and Fischer [15]. The k-neighbour TSP is also related to the k-peripatetic salesman problem [12, 20] and the watchman problem [7]. Algorithms for maximization and minimization versions of TSP(2) were studied by Staněk [30] and Oswin et al. [26]. To the best of our knowledge, no other works in the literature address TSP(k). Referringwhich to the UAV discussed earlier, it isquickly. clear that the flight subpaths depend on both the A route canexample be travelled more angle and distances between successive nodes. By precalculating these and assigning costs to q(e, f ), for e, f ∈ E, we see that QTSP is a natural model for this problem. In fact, the flight paths may be affected by edges further downstream. Thus we can get successively better models by considering TSP(1), TSP(2),. . ., 12 TSP(k) in turn. In practice we expect diminishing returns to take hold quickly and hence TSP(k) with small values of k are of particular interest. In this paper we show that QTSP is NP-hard even if the costs are restricted to 0-1 values and the underlying graph is Halin. In contrast, TSP and TSP(2) on a Halin graph can be solved in O(n) time [9, 33]. Interestingly, we show that TSP(3) can also be solved on a Halin graph in O(n) time, although as we move from TSP(2) to TSP(3), the problem gets much more complicated. In fact, our approach can be extended to obtain polynomial time algorithms for TSP(k) whenever k = O(log n). We note that while Halin graphs have treewidth 3, the results on graphs with bounded treewidth (e.g. [6, 11]) usually cannot easily be extended to optimization problems with quadratic objective functions. The paper is organized as follows. In Section 2 we introduce some preliminary results and notations for the problem. The complexity result for QTSP on Halin graphs is given in Section 3. An O(n) algorithm to solve TSP(3) on Halin graphs is given in Section 4.1, which can be extended to obtain an O(n2(k−1)/2 ) algorithm for TSP(k). Further extensions of this result to fully reducible classes of graphs are briefly discussed in Section 5. An earlier version of the NP-completeness results presented here were included as part of the M.Sc. thesis of the first author [31]. 2. Notations and definitions A Halin graph H = T ∪ C is obtained by embedding a tree with no nodes of degree two in the plane and connecting the leaf nodes of T in a cycle C so that the resulting graph remains planar. Unless otherwise stated, we always assume that a Halin graph or its subgraphs are given in the planar embedded form. The non-leaf nodes belonging to T are referred to as tree or internal nodes and the nodes in C are referred to as cycle or outer nodes of H. A Halin graph with exactly one internal node is called a wheel. If H has at least two internal nodes and w is an internal node of T which is adjacent to exactly one other internal node, then w is adjacent to a set of consecutive nodes of C, which we denote by C(w). Note that |C(w)| ≥ 2. The subgraph of H induced by {w} ∪ C(w) is referred to as a fan, and we call w the centre of the fan. See Figure 4. F w edge in C edge in T fan in H Figure 1. A Halin graph with 3 fans. w is the centre of fan F . Figure 4. A Halin graph H with 3 fans. w is the centre of fan F . Lemma 2.1 (Cornuejols et al. [9]). Every Halin graph which is not a wheel has at least two fans. Let G = (V, E) be a graph and let S ⊆ V be a connected subgraph of G. Let ϕ(S) be the cutset of S, that is, the smallest set of edges whose removal disconnects S from the vertices in V \ S. Let G/S be the graph obtained by contracting S into a single node, called a ‘pseudonode’ denoted by vS [9]. The edges in G/S are obtained as follows: 3 (1) An edge with both ends in S is deleted; AinHalin with 3 fans. w is the centre (2) An edge Figure with both 1. ends G − Sgraph remains unchanged; (3) An edge (v1 , v2 ) with v1 ∈ G − S, v2 ∈ S is replaced by the edge (v1 , vs ). of fan F . Lemma 2.2 (Cornuejols et al. [9]). If F is a fan in a Halin graph H, then H/F is a Halin graph. Note that each time a fan F is contracted using the graph operation H/F , the number of non-leaf nodes of the underlying tree is reduced by one. That is, after at most d(n − 1)/2e fan contractions, a Halin graph will be reduced to a wheel. Let w be the centre of a fan F , and label the outer nodes in F in the order they appear in C as, u1 , u2 , . . . , ur (r ≥ 2). Let (j, k, l) be the 3-edge cutset ϕ(F ) which disconnect F from G such that j is adjacent to u1 , k is adjacent to w but not adjacent to ui for any i, 1 ≤ i ≤ r, and l is adjacent to ur (See Figure 5, r = 4). Note that every Hamiltonian cycle τ in H contains exactly two edges of {j, k, l}. The pair of edges chosen gives us a small number of possibilities for traversing F in a tour τ . For example, if τ uses k and l, it contains the subsequence w, u1 , u2 , . . . , ur (call this a left-traversal of F ), if τ uses j and k it contains the subsequence u1 , u2 , . . . ur , w (call this a right-traversal of F ) and if τ uses j and l, it contains a subsequence of the form u1 , u2 , . . . , ui , w, ui+1 , . . . , ur , for some i ∈ {1, 2, . . . , r − 1} as it must detour through the centre of F (call this a centre-traversal of F ). j F u1 u2 u3 v k edge in C edge in T fan in H u4 l Figure A Halin H containing fan F a. 3-edge {j, k,cutset l} is which a 3-edge cutset Figure 5. 2. A Halin graphgraph H containing fan F . {j, k, l} is disconnects disconnects F from H. Fwhich from H. 1 3. Complexity of QTSP on Halin graphs Many optimization problems that are NP-hard on a general graph are solvable in polynomial time on a Halin graph [9, 24, 25]. In particular, TSP on a Halin graph is solvable in linear time. Unlike this special case, we show that QTSP is strongly NP-hard on Halin graphs. The decision version of QTSP on a Halin graph, denoted by RQTSP, can be stated as follows: P e∈τ c(e) + P “Given a Halin graph H and a constant θ, does there exist a tour τ in H such that q(e, f ) ≤ θ?” e,f ∈τ Theorem 3.1 (Woods [31]). RQTSP is NP-complete even if the values c(e) ∈ {0, 1} and q(e, f ) ∈ {0, 1} for e, f ∈ H. Proof. RQTSP is clearly in NP. We now show that the 3-SAT problem can be reduced to RQTSP. The 3-SAT problem can be stated as follows: “Given a Boolean formula R in Conjunctive Normal Form (CNF) containing a finite number of clauses C1 , C2 , . . . , Ch on variables x1 , x2 , . . . , xt such that each clause contains 4 exactly three literals (L1 , . . . , L3h where for each i, Li = xj or Li = ¬xj for some 1 ≤ j ≤ t), does there exists a truth assignment such that R yields a value ‘true’ ?” From a given instance of 3-SAT, we will construct an instance of RQTSP. The basic building block of our construction is a 4-fan gadget obtained as follows. Embed a star on 5 nodes with center v and two specified nodes ` and r on the plane and add a path P from ` to r covering each of the pendant nodes so that the resulting graph is planar (see Figure 6). Call this special graph a 4-fan gadget. µ1 µ2 µ3 r ` v Figure 4-fan gadget constructed by embedding starinon nodes the Figure 6.3.4-fan gadget constructed by embedding a star on 5anodes the4plane andinadding a path.and adding a path connecting the pendant nodes so that the resulting plane graph is planar. The nodes on path P of this gadget are called outer nodes and edges on P are called outer edges. Let µ1 , µ2 , µ3 be edges with distinct end points in P . Note C1 that any `-r Hamiltonian path of the gadget must contain all the outer edges except one which is skipped to detour through v. We will refer to an `-r µ 2 µ1 µ3 Hamiltonian path in a 4-fan gadget as a center-traversal as r1 before. l1 We will construct a Halin graph H using one copy of the gadget for each clause and let µ1 , µ2 , µ3 correspond to literals contained in that clause. We will assign costs to pairs of edges such that every Hamiltonian cycle with cost 0 must contain a centre-traversal for each clause. To relate a Hamiltonian cycle to a truth vy which does not contain edge µi corresponds assignment, a centre-traversal to an assignment of a true value l2 µ4 to literal Li . v1 y clause C1 , . . . , Ch , create a copy of µthe 4-fan gadget. The r, `, and v Now construct H as follows. For each 5 nodes of the 4-fan gadget corresponding to the clause Ci are denoted by ri , `i andCv2i respectively. Connect the 6 node ri to the node `i+1 , i = 1, 2, . . . , h. Introduce nodes vxv2and vy and theµedges (`1 , vx ), (vx , vy ), (vy , rh ). vx w and connect it to vw Also introduce a new node x , vy and vi for i = 1, 2, . r.2. , h − 1. The resulting graph is the required Halin graph H. See Figure x 7. Assign the cost c(e) = 0 for every edge in H. Let x = (vx , w) and y = (vy , w). Note that every tour which contains edges x and y traverses every gadget using a centre-traversal. For each gadget: assign costs of v4 v3 q(e, f ) = 1 for pairs of edges which are neither outer edges nor both adjacent to the same literal edge µ1 , µ2 or µ3 , and for all other pairs of edges within the gadget assign cost 0. For each variable xj , j = 1, . . . , n, and all literals Lm , Lq (m 6=r4q) if xj = Lm = ¬Lq , assign cost q(µ0ml,3µ0q ) = 1 where µ0m and µ0q are edges connecting µm (and µq ) to the paired costs are assumed to be 0. µ7 µ12respective 4-fan gadget centre v. All other Suppose B is a valid truth assignment. Then in each clause there exists at least one true literal. Consider µ 8 µ11 C3gadget such that τ detours around C4 the a tour τ in H which contains edges r3 µ9 every µ10 l4x and y and traverses exactly one literal edge which corresponds to a literal which is true in B. Since the truth assignment is valid, such a τ exists. Clearly τ has cost 0, since no costs are incurred by pairs of edges contained in a single Figure 4. Example of the Halin graph constructed from F = C1 ∧ C2 ∧ C3 ∧ C4 . gadget, nor are costs incurred of the form q(µ0a , µ0b ) where La = ¬Lb . The latter must be true because in any truth assignment, the variable corresponding to La , say xa , must be either assigned a value of true or false. Suppose a cost of 1 is incurred by q(µ0a , µ0b ) and hence La = ¬Lb . If xa is true, and xa = La , then Lb clearly must be false, so τ cannot detour to miss both µ0a and µ0b . The same contradiction arises, if xa is false. Hence a yes instance of 3-SAT can be used to construct a yes instance for RQTSP with θ = 0. Now suppose there is a tour which solves RQTSP with θ = 0. Suppose τ 0 is such a tour. Clearly it must use edges x and y, and hence must traverse every gadget via a centre-traversal. Such a detour must 5 v Figure 3. 4-fan gadget constructed by embedding a star on 4 nodes in the plane and adding a path connecting the pendant nodes so that the resulting graph is planar. C1 µ1 µ2 µ3 r1 l1 vy l2 v1 y µ5 µ6 v2 vx C2 r2 w x v4 v3 r4 l3 µ12 µ7 C4 µ4 µ8 µ11 r3 µ 9 µ10 l4 C3 Figure 4. Example of the Halin graph constructed from F = C1 ∧ C2 ∧ C3 ∧ C4 . Figure 7. Example of the Halin graph constructed from F = C1 ∧ C2 ∧ C3 ∧ C4 . skip a literal edge in every gadget, otherwise a cost of 1 is incurred. Suppose D = {L1 , . . . , Ls } is the set of literals which are skipped. Li 6= ¬Lj for any i, j, otherwise a cost of 1 is incurred. This implies that a truth assignment which results in every literal in D being true is a valid truth assignment to the variables x1 , . . . , xt . That is, for each literal edge which is skipped in τ 0 , assign true or false to the corresponding variable such that the literal evaluates to true (if Li = xj , set xj = true and if Li = ¬xj , set xj = f alse). The truth values for any remaining variables can be assigned arbitrarily. This truth assignment returns true for each clause since exactly one literal in each clause is detoured, and evaluates to true. Hence this truth assignment is a valid assignment for 3-SAT.   4. Complexity of k-neighbour TSP on Halin graphs Let G be a planar embedding of a planar graph and e and f are two distinct edges of G. Then e and f 2 contains both e and f . This may include the outer are said to be cofacial if there exists a face of G which face. Theorem 4.1. Let τ be a tour in the planar embedding H = T ∪ C of a Halin graph. Then, any two edges adjacent in τ must be cofacial. Proof. Suppose the result is not true. Then, there exists a tour τ in H containing two adjacent edges e = (u, x), f = (x, v) such that e and f are not cofacial. From our previous discussion on fan traversals, we can assume that x 6∈ C. Since e and f are not cofacial at x, there exists edges g = (y, x) and h = (x, z) in H such that the clockwise ordering of edges incident on x is of the form f, . . . , g, . . . , e, . . . , h (See Figure 8). Without loss of generality, assume T is rooted at x. Then T has at least four subtrees Tu , Tv , Ty , Tz rooted respectively at u, v, y and z. Since τ is a tour containing the edge e, it must contain a path, say P1 , through the subtree Tu from u to uc ∈ C. Note that uc could be the same as u and in this case the subtree Tu is the isolated node u. Similarly, τ must contain a path P2 in Tv from v to vc ∈ C (See Figure 8). Note that P = P1 ∪ P2 ∪ {e, f } is a path in τ . Deleting the vertex set V (P ) of P and its incident edges from H yields a disconnected graph. Thus, τ − V (P ) must be disconnected, a contradiction.   A preliminary version of this result is given in [31]. 6 yc y g uc P1 u e x v f vc P2 h z zc Figure8.5.A A Halin graph H non-consecutive with non-consecutive edges andx.f Hat− node x. Figure Halin graph H with edges e and f at enode P where P P2 ∪ {e,Pf }=has hence no components, τ contains bothhence edges enoand f. H =−PP1 ∪where P1 two ∪ Pcomponents, two τ contains 2 ∪ {e, f } has both edges e and f . x A path in the planar embedding H = T ∪ C of a Halin graph is called a candidate paths if its consecutive α α 5 edges are cofacial. A candidate path with k-edges is called 6a candidate k-path. Note that only candidate paths can be subpaths of a tour but it is possible that kthere are candidate paths that are not part of a tour. Corollary 4.2. Let H = T ∪ C be a planar embedding of a Halin graph a specified edge of H. α4 and eαbe α2 3 w Then, H has at most k α · 21k−1 candidate k-paths containing e. Proof. Consider extending e, if necessary in either directions in H, to a candidate k-path P . For a given u u0 end point of a sub path of P , say vertex u, tby are onlyr+1 two possible edges incident on 1 Theorem 4.1,tthere r u, which may belong to P . Thus, there are at most 2k−1 candidate k-paths when the position of e is fixed. Since e can take any of the k positions in a candidate k-path, there arel at most k · 2k−1 candidate k-paths j u1 ur containing e.   yr−1 y1 As an immediate consequence of Corollary 4.2, we have an upper bound of 2n−2 on the number of Hamiltonian cycles in a Halin graph. To see this, let r be the number of Hamiltonian cycles in H. From Figure 6. A fan F with centre w. Every τ containing j and k contains each such cycle, we can generate n distinct Hamiltonian paths (candidate (n − 1)-paths) by ejecting an edge. edges . . ,Hamiltonian yr−1 , tr . Every l contains y1and , . . . all , yr−1 , t1paths . 1 , .all Repeating this yfor cyclesτ incontaining H, we get k rnand candidate (n − edges 1)-paths these are Every τ containing j and l must contain one of the subpaths from the set distinct. Hence rn is a lower bound on the number of candidate (n − 1)-paths in H. By Corollary 4.2, (n − 1)2n−2 upper Thus, {α1is−anj − l − bound α3 , α1on − the j −number l − α4 ,ofαcandidate − −α1)-paths l− α4 }.rn ≤ (n − 1)2n−2 and 2 − j − l (n 3 , α2 − jin−H. n−1 n−2 n−2 hence r ≤ n 2 ≤2 . Noting that there are at most 2(n − 1) edges in H and that the number of quadratic costs which are relevant is the number of candidate k-paths, it follows from Corollary 4.2, that the number of quadratic costs which are relevant is bounded above by k · 2k · (n − 1) = O(n) for any fixed k and O(nt+2 ) if k ≤ t log n. Note that any face of H must contain an outer edge. Moreover, the following Corollary will prove useful. Corollary 4.3. If H is embedded in the plane such that it is planar and C defines the outer face, for any outer edge e which is contained in the outer face and face Fe , every tour which does not contain e must contain all other edges of Fe . 7 3 4.1. TSP(3) on Halin graphs. As indicated earlier, TSP(1) is the same as TSP, which is solvable in linear time on Halin graphs [9]. TSP(2) can also be solved in linear time by appropriate modifications of the algorithm of [25] as indicated in [33]. However, for k ≥ 3, such modifications do not seem a viable option. We now develop a linear time algorithm to solve TSP(3). Let us start with an alternative formulation of TSP(3). For any subgraph G of H, let P3 (G) be the collection of all distinct candidate 3-paths in G. For each candidate 3-path (e, f, g) ∈ P3 (H), define q(e, f, g) = q(e, g) + q(e, f ) + q(f, g) c(e) + c(f ) + c(g) + . 2 3 (4.1) Now consider the simplified problem: ST SP (3) : Minimize X q(e, f, g) (e,f,g)∈P3 (τ ) Subject to τ ∈ F. Theorem 4.4. Any optimal solution to the STSP(3) is also optimal solution to TSP(3). Proof. For any τ ∈ F, X q(e, f, g) (e,f,g)∈P3 (τ )  q(e, f ) + q(f, g) c(e) + c(f ) + c(g) + = q(e, g) + 2 3 (e,f,g)∈P3 (τ ) X X = q(e, f ) + c(e). X (e,f )∈δ(3,τ )  e∈δ(τ ) Thus, the objective function values of STSP(3) and TSP(3) are identical for identical solutions. Since the family of feasible solutions of both these problems are the same, the result follows.   In view of Theorem 4.4, we restrict our attention to STSP(3). For TSP(1), Cornuejols et al. [9] identified costs of new edges generated by a fan contraction operation by solving a linear system of equations. This approach cannot be extended for any k-neighbour TSP for k ≥ 3 as it leads to an over-determined system of equations which may be infeasible. Instead, we extend the penalty approach used in Phillips et al. [25]. The idea here is to introduce a node-weighted version of the problem STSP(3) where we use a penalty function for the nodes of C, the value of which depends on the edges chosen to enter and exit the node, along with some other ‘candidate’ edges. We iteratively contract the fans in H, storing the appropriate values of suitable subpaths as we traverse the fans in a recursive way. Once we reach a wheel, we can compute an optimal tour for the resulting problem. Backtracking by recovering appropriate subpaths from contracted fans in sequence, an optimal solution can be identified. To formalize the general idea discussed above, let us first discuss the case where H is a Halin graph which is not a wheel. In this case, H will have at least two fans. Let F be an arbitrary fan in H with w as the centre. Label the outer nodes of F in the order they appear in C, say, u1 , u2 , . . . , ur (r ≥ 2). Let {j, k, l} be the 3-edge cutset ϕ(F ) which disconnects F from H such that j is adjacent to u1 , k is adjacent to w and l is adjacent to ur . Let j = (u1 , u0 ), k = (w, x) and l = (ur , ur+1 ). There are exactly two edges not connected to F which are cofacial with k and incident on x. The first edge which follows k in the clockwise orientation of edges incident on x is denoted α5 , and the other edge incident on x and cofacial with k is denoted by α6 . (See Figure 9.) There are exactly two edges not connected to F and incident on u0 . These edges are labelled α1 , α2 . Likewise, there are exactly two edges not connected to F and incident on ur+1 . These edges are labelled α3 , α4 . (See Figure 9.) Without loss of generality α1 , α3 are in C and α2 , α4 are in T . It is possible that α2 could be the same as α5 and also possible that α4 could be the same as α6 . To complete a fan contraction operation, we consider the 3 types of traversals of F . We define a penalty function stored at nodes (pseudonodes) of C which contains attributes of a minimum traversal of F of each type. For any left- or right-traversal of F , there is a single path through F using all cycle edges. Any tour which includes j and k must pass through one edge of the pair incident on u0 lying outside F together with edges y1 , . . . , yr−1 , tr , k, α6 . Similarly, any tour which includes k and l must pass through one edge of the pair incident on ur+1 lying outside F together with edges yr−1 , . . . , y1 , t1 , k, α5 . Any tour which includes j and l must also pass through one edge in each of the pairs of edges incident on u0 (or ur+1 ) lying outside 8 z zc F . That is, every tour τ containing j and l must contain a path containing one collection of edges from the set {(α1 ,Figure j, l, α3 ), (α5. l, α , j, l, α3 ), α4 )}. We refer to aedges centre-traversal of Fnode whichx.bypasses 1 , j,A 4 ), (α2graph 2 , j, l, non-consecutive Halin H(αwith e and f at y1 ∈ F ∩ C as a left path, one which bypasses ys ∈ F ∩ C for some s ∈ [2, r − 2] as a middle path, and one H − P where P = P ∪ P ∪ {e, f } has two components, hence no τ contains which bypasses yr−1 ∈ F ∩ C as 1a right2 path. both edges e and f . α5 x α6 k α2 α1 α4 w u0 t1 ur+1 tr j u1 α3 ur l yr−1 y1 Figure9. 6. A Ffan F centre withw.centre Every jτand containing j and Figure A fan with Every τw. containing k contains edges y1 ,k. . .contains , yr−1 , tr . Every τ containing k and l contains edges y , . . . , y , t . Every τ containing j and l must edges y1 , . . . , yr−1 , tr . Every τ containing k and , t1 . 1 r−1 l 1contains edges y1 , . . . , yr−1 contain one of the subpaths from the set {α − j − l − α , α − j − l − α , α − j − l − α 1 4 2 from the 3 , αset 2− Every τ containing j and l must contain one3 of1 the subpaths j − l − α4 }. {α1 − j − l − α3 , α1 − j − l − α4 , α2 − j − l − α3 , α2 − j − l − α4 }. Let S be the set of nodes (pseudonodes) in C at some iteration in the contraction process. In TSP(3), quadratic costs are ‘absorbed’ during each fan contraction operation, depending both on the edges in F and within a distance 2 from F , so care must be taken to retain the proper information. That is, in order to develop an extension of the penalty approach used in Phillips et al. [25] for TSP(3), we extend the penalty function stored at the nodes (pseudonodes) in C that depends on an additional parameter ρ, which specifies the structure of edges around each pseudonode. Note that due to the recursive property of pseudonodes where a fan contraction operation may ‘absorb’ pseudonodes, the parameter ρ specifies a shape of the structure rather than explicitly stating the edges surrounding a pseudonode. Further, we define a function β which stores the penalty values associated with ρ. 3 0), L = (1, 0), R = (0, 1), B = (1, 1)} and A2 be the Let A1 be the collection of ordered pairs {M = (0, collection of ordered pairs {(1, 3), (1, 4), (1, 6), (2, 3), (2, 4), (2, 6), (5, 3), (5, 4)}. At each node i ∈ S we define 12 1 2 21 22 2 ρi = (ρ1i , ρ2i ) where ρ1i = (ρ11 i , ρi ) ∈ A and ρi = (ρi , ρi ) ∈ A . ρi indicates which penalty value (to be defined shortly) stored at pseudonode i is to contribute to the objective function value. The first component of ρi is a binary vector of length 2 which specifies the inner structure of i (edges y1 , t1 , yr−1 and tr prior to any fan contractions such that the first component is 0 if y1 is selected, and 1 if t1 is instead, and the second component is 0 if yr−1 is selected, and 1 if tr is instead), and the second component of ρi , the outer structure (α1 to α6 prior to any contraction of adjacent pseudonodes). Let ρ be the vector containing ρi for every i ∈ S. Let ρH/F be the restriction of ρ to the vertices in H/F and augmented by ρvF ∈ {(a, b)}. Let βi (ρi ) be the penalty that is incurred if ρi occurs at i. For i ∈ C define: ( βi (ρi ) if ρi is defined, Pi (τ, ρi ) = 0 otherwise. Since the penalties stored at i ∈ S depend on edges which are not incident with i, and the dependent edges may be ‘absorbed’ into adjacent pseudonodes so that not every ρ is feasible for a given τ . That is, the inner and outer structures of adjacent pseudonodes in C must agree. Formally, we say that ρ is feasible for τ if the following conditions are satisfied for every pseudonode i (1) j, k ∈ τ ⇐⇒ ρ22 i = 6 and (2) k, l ∈ τ ⇐⇒ ρ21 i = 5, 9 Figure 7. A tour τ in a wheel, which skips edge ci . ci−1 ci+1 ti+1 ti w ui+1 i and for every pair of consecutive upseudonodes i, i + 1 ∈ C (1) (2) (3) (4) ρ12 i ρ12 i ρ22 i ρ22 i =1 =0 =3 =4 ⇐⇒ ⇐⇒ ⇐⇒ ⇐⇒ ρ21 i+1 ρ21 i+1 ρ11 i+1 ρ11 i+1 = 2, ci = 1, = 0 and = 1. ti+1 ti ci ui+1 ui c ci+1 For an example i−1 Let F 0 be the set of all feasible (τ, ρ) pairs. of a feasible (τ, ρ) pair, see Figure 10. Figure 8. Wheel H with center w considered as a fan. F1 F2 F3 Figure of Hofwhich becomes a fan after contracting F1 , F2 and F3F. ,The Figure10.9. AAsubgraph subgraph H which becomes a fan after contracting 1 F2 bold edges depict a centre-traversal of F contained in Hamiltonian cycle τ . Since τ contains and F3 . The bold edges depict a centre-traversal of F contained in Hamilton the edges t1 and yr−1 (relative to F1 ), the inner structure of vF1 must be (1, 0), and since cycle τ . A (τ, ρ F11 ),=the ((1, 0), (2, 3)), ρvofF2 v= ((0, 1), (1, 3)) τ contains edges ρ) α2 pair and αcontaining outer structure 3 (relative to vF F1 must be (2, 3). and ρis,vF3a = ((0, 0), 6)) is feasible. which correspond That feasible (τ, (2, ρ) pair must contain The ρvF1 edges = ((1, 0), (2, 3)). Consideringtoallthe of Finner 1 , F2 and F , every feasible (τ, ρ) pair contains ρ = ((1, 0), (2, 3)), ρ = ((0, 1), (1, 3)) and 3 vF2 correspond to the structure of ρvF2 are coloured in blue vand the edges which F1 ρ = ((0, 0), (2, 6)). The edges which correspond to the inner structure of ρ are coloured vF3 vF2 outer structure of ρvF2 are coloured in red. in blue and the edges which correspond to the outer structure of ρvF2 are coloured in red. The problem now contains a cost for every triplet of consecutive edges in tour τ and additionally, a penalty 4 at each outer node i. Consider the modified 3-Neighbour TSP on a Halin graph defined as follows: M T SP (3) : Minimize X z(τ, ρ) = (e,f,g)∈P3 (τ ) Subject to q(e, f, g) + X i∈C Pi (τ, ρi ) (τ, ρ) ∈ F 0 . The necessary costs to construct MTSP(3) can be obtained as required by applying (4.1) or by first embedding H in the plane, and evaluating the candidate k-paths from Corollary 4.2. Note that Pv (τ, ρv ) can be computed in O(1) time by storing penalty 24-tuples containing the βv -values described in Table 1, at each cycle node v. Also note that there may be O(2n ) feasible ρ vectors for a given τ , however, we show that the optimal (τ, ρ) pair can be found in O(n)-time. It is also important to note that the set of pseudonodes is retained for reasons which will become apparent. In the initial graph, and for all ρi , i ∈ C, set βi (ρi ) = 0. For fan F in H, the penalties must be updated to store the costs of traversing F when F is contracted to pseudonode vF . Let K represent P the traversal of F which contains only edges in C. That is, K = j − y1 − · · · − yr−1 − l. Then q(K) = e−f −g∈K q(e, f, g) represents the cost incurred by selecting the edges in K. Let τ (F ) and ρ(F ) be the restrictions of τ and ρ to F , respectively. Assign the minimum cost of the right-traversal (which contains αs , s ∈ {1, 2} and α6 ), with inner structure of the first pseudonode a ∈ {L = (1, 0), M = (0, 0)} to βvF (a, (s, 6)). That is, assign to βvF (a, (s, 6)) the sum of the costs along the traversal, q(αs − j − y1 − · · · − yr−1 − tr − k − α6 ), together with the minimum feasible set of penalties on the outer nodes contained in F , u1 , u2 . . . , ur . Note that for the case that u1 6∈ S, 10 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Penalty Description βvF ((0, ∗), (1, 6)) traversal of F with inner structure (0, ∗) and outer structure βvF ((1, ∗), (1, 6)) traversal of F with inner structure (1, ∗) and outer structure βvF ((0, ∗), (2, 6)) traversal of F with inner structure (0, ∗) and outer structure βvF ((1, ∗), (2, 6)) traversal of F with inner structure (1, ∗) and outer structure βvF ((∗, 0), (5, 3)) traversal of F with inner structure (0, ∗) and outer structure βvF ((∗, 1), (5, 3)) traversal of F with inner structure (1, ∗) and outer structure βvF ((∗, 0), (5, 4)) traversal of F with inner structure (0, ∗) and outer structure βvF ((∗, 1), (5, 4)) traversal of F with inner structure (1, ∗) and outer structure βvF ((0, 0), (1, 3)) traversal of F with inner structure (0, 0) and outer structure βvF ((0, 1), (1, 3)) traversal of F with inner structure (0, 1) and outer structure βvF ((1, 0), (1, 3)) traversal of F with inner structure (1, 0) and outer structure βvF ((1, 1), (1, 3)) traversal of F with inner structure (1, 1) and outer structure βvF ((0, 0), (1, 4)) traversal of F with inner structure (0, 0) and outer structure βvF ((0, 1), (1, 4)) traversal of F with inner structure (0, 1) and outer structure βvF ((1, 0), (1, 4)) traversal of F with inner structure (1, 0) and outer structure βvF ((1, 1), (1, 4)) traversal of F with inner structure (1, 1) and outer structure βvF ((0, 0), (2, 3)) traversal of F with inner structure (0, 0) and outer structure βvF ((0, 1), (2, 3)) traversal of F with inner structure (0, 1) and outer structure βvF ((1, 0), (2, 3)) traversal of F with inner structure (1, 0) and outer structure βvF ((1, 1), (2, 3)) traversal of F with inner structure (1, 1) and outer structure βvF ((0, 0), (2, 4)) traversal of F with inner structure (0, 0) and outer structure βvF ((0, 1), (2, 4)) traversal of F with inner structure (0, 1) and outer structure βvF ((1, 0), (2, 4)) traversal of F with inner structure (1, 0) and outer structure βvF ((1, 1), (2, 4)) traversal of F with inner structure (1, 1) and outer structure Table 1. Description of penalty 24-tuple stored at pseudonodes in C. (1, 6) (1, 6) (2, 6) (2, 6) (5, 3) (5, 3) (5, 4) (5, 4) (1, 3) (1, 3) (1, 3) (1, 3) (1, 4) (1, 4) (1, 4) (1, 4) (2, 3) (2, 3) (2, 3) (2, 3) (2, 4) (2, 4) (2, 4) (2, 4) it is not possible to have an inner structure L or B, and βvF (L, (s, 6)) = βvF (B, (s, 6)) = ∞. Otherwise ( r ) X βvF (a, (s, 6)) = q(αs − j − y1 − · · · − yr−1 − tr − k − α6 ) + min 0 βui (ρui ) (τ (F ),ρ(F ))∈F (F ): ρ1u =a or B,ρ22 ur =6 i=1 1 = q(αs , j, y1 ) + q(K) − q(yr−2 , yr−1 , l) + q(yr−2 , yr−1 , tr ) + q(yr−1 , tr , k) + q(tr , k, α6 ) (4.2) ( r ) X + min 0 βui (ρui ) . (τ (F ),ρ(F ))∈F (F ): ρ1u =a or B,ρ22 ur =6 i=1 1 We will explain how the minimum in (4.2) can be calculated efficiently later in this paper. Similarly, assign the minimum cost of the left-traversal (which contains αt , t ∈ {3, 4}, and α5 ) with inner structure a ∈ {M = (0, 0), R = (0, 1)} to βvF (a, (5, t)). In the case that ur 6∈ S, it is not possible to have an inner structure R or B, and βvF (R, (5, t)) = βvF (B, (5, t)) = ∞. Otherwise βvF (a, (5, t)) = q(α5 , k, t1 ) + q(k, t1 , y1 ) + q(t1 , y1 , y2 ) + q(K) − q(j, y1 , y2 ) + q(yr−1 , l, αt ) ( r ) X + min 0 βui (ρui ) . (τ (F ),ρ(F ))∈F (F ): ρ1ur =a or B,ρ21 u =5 (4.3) i=1 1 Let K(yi ), i ∈ {1, . . . , r − 1}, be the centre-traversal of F which does not contain yi . Then q(K(yi )) represents the cost incurred by the edges in K(yi ). That is, q(K(y1 )) = q(K) + q(j, t1 , t2 ) + q(t1 , t2 , y2 ) + q(t2 , y2 , y3 ) − q(j, y1 , y2 ) − q(y1 , y2 , y3 ), 11 q(K(yp )) = q(K) + q(yp−2 , yp−1 , tp ) + q(yp−1 , tp , tp+1 ) + q(tp , tp+1 , yp+1 ) + q(tp+1 , yp+1 , yp+2 ) − q(yp−2 , yp−1 , yp ) − q(yp−1 , yp , yp+1 ) − q(yp , yp+1 , yp+2 ), for p ∈ {2, . . . , r − 2}, and q(K(yr−1 )) = q(K) + q(yr−3 , yr−2 , tr−1 ) + q(yr−2 , tr−1 , tr ) + q(tr−1 , tr , l) − q(yr−3 , yr−2 , yr−1 ) − q(yr−2 , yr−1 , l). Assign the minimum cost of the centre-traversal which contains αs , s ∈ {1, 2}, and αt , t ∈ {3, 4}, which has inner structure a ∈ {L = (1, 0), M = (0, 0), R = (0, 1), B = (1, 1)} to βvF (a, (s, t)). In the case that u1 6∈ S and a = L, there is a single path traversing F with inner structure L, namely, j − t1 − t2 − y2 − · · · − yr−1 − l, so ( r ) X βvF (L, (s, t)) = q(αs , j, t1 ) + q(K(y1 )) + q(yr , l, αt ) + min 0 βui (ρui ) , (4.4) (τ (F ),ρ(F ))∈F (F ): 22 ρ12 ur =0,ρu2 =6 i=2 and when u1 ∈ S, we assign the cost of the minimum centre-traversal with inner structure L. Note that this path detours some yg , g ∈ {2, . . . , r − 1}.          ( r )     X q(αs , j, t1 ) + q(K(yg )) + q(yr , l, αt ) + min 0 . βui (ρui ) βvF (L, (s, t)) = min  g∈{2,...,r−1}  (τ (F ),ρ(F ))∈F (F ):   i=1   12   ρ11 u1 =1,ρr =0,     21 22 ρug =5,ρu g+1 =6 (4.5) Similarly, when a = R and ur 6∈ S, there is a single path traversing F with inner structure R, namely, j − y1 − · · · − yr−2 − rr−1 − tr − l so (r−1 ) X βvF (R, (s, t)) = q(αs , j, y1 ) + q(K(y1 )) + q(tr , l, αt ) + min 0 βui (ρui ) , (4.6) (τ (F ),ρ(F ))∈F (F ): 22 ρ11 =6 u =0,ρu 1 i=1 r−1 and when ur 6∈ S, we assign the cost of the minimum centre-traversal with structure R. Note that this path detours some yg , g ∈ {1, . . . , r − 2}.          ( r )     X βvF (R, (s, t)) = q(αs , j, y1 ) + q(K(yg )) + q(yr , l, αt ) + min 0 βui (ρui ) . min  g∈{1,...,r−2}  (τ (F ),ρ(F ))∈F (F ):   i=1   11 12   ρu1 =0,ρr =1,     21 22 ρug =5,ρu g+1 =6 (4.7) Now consider a = M . If u1 ∈ S, ur ∈ S then βvF (M, (s, t)) is assigned the cost of the minimum path detouring yg for g ∈ {1, . . . , r − 1}. If u1 (ur ) is not a pseudonode then the centre-traversal must contain y1 (yr−1 ) and 1 (r − 1) is removed from g from the following equation.          (r−1 )     X βvF (M, (s, t)) = min q(αs , j, y1 ) + q(K(yg )) + q(yr , l, αt ) + min 0 βui (ρui ) .  g∈{1,...,r−1}  (τ (F ),ρ(F ))∈F (F ):   i=1   12   ρ11 u1 =0,ρur =0,     21 22 ρug =5,ρu g+1 =6 (4.8) 12 Now consider a = B. This is the same as the case where a = M except for the inner structure.          (r−1 )     X βvF (B, (s, t)) = min q(αs , j, y1 ) + q(K(yg )) + q(yr , l, αt ) + min 0 βui (ρui ) .  g∈{1,...,r−1}  (τ (F ),ρ(F ))∈F (F ):   i=1   11 12   ρu1 =1,ρur =1,     21 22 ρug =5,ρu g+1 =6 (4.9) All βvF which have not been assigned are not associated with a feasible ρvF and are assigned a value ∞. Theorem 4.5. Suppose H = T ∪ C is a Halin graph which is not a wheel and F is a fan in H. If (τ ∗ , ρ∗ ) is an optimal tour pair in H then there exists a feasible ρ∗H/F in H/F such that (τ ∗ /F, ρ∗H/F ) is optimal in H/F and z(τ ∗ , ρ∗ ) = z(τ ∗ /F, ρ∗H/F ). Proof. Let S be the set of pseudonodes in H. Let vF be the pseudonode which results from the contraction of F , and label the vertices and edges of F as in Figure 9. Given (τ ∗ , ρ∗ ), we construct (τ ∗ /F, ρ∗H/F ) as follows. Let ρ∗H/F,i = ρ∗i ∀i 6∈ F and ρ∗H/F,vF corresponding to the structure of (τ ∗ , ρ∗ ) around F in H. That is, τ contains an αi , αj path through F in H, so let ρ∗H/F,vF = ((a, b), (c, d)) where ( 0 a= 1 if (u1 6∈ S and y1 ∈ τ ∗ ) or (u1 ∈ S and ρ∗11 u1 = 0) ∗ if (u1 ∈ 6 S and y1 6∈ τ ) or (u1 ∈ S and ρ∗11 u1 = 1), ( if (ur 6∈ S and yr−1 ∈ τ ∗ ) or (ur ∈ S and ρ∗12 u1 = 0) ∗ ∗12 if (ur ∈ 6 S and yr−1 6∈ τ ) or (ur ∈ S and ρu1 = 1), b= 0 1   1 c= 2   5 if (u1 6∈ S and α1 ∈ τ ∗ ) or (u1 ∈ S and ρ∗21 u0 = 1) ∗ if (u1 ∈ 6 S and α2 ∈ τ ) or (u1 ∈ S and ρ∗21 u0 = 2) if (u1 ∈ 6 S and α5 ∈ τ ∗ ) or (u1 ∈ S and ρ∗21 u0 = 5),   3 d= 4   6 if (ur 6∈ S and tr 6∈ τ ∗ ) or (ur ∈ S and ρ∗22 ur = 3) ∗ if (ur ∈ 6 S and tr ∈ τ ) or (ur ∈ S and ρ∗22 ur = 4) ∗ ∗22 if (ur ∈ 6 S and tr ∈ τ ) or (ur ∈ S and ρur = 6). and Then (τ ∗ /F, ρ∗H/F,vF ) is feasible in H/F . Using equations (4.2)-(4.9) and noting that all new triples which contain vF have q(e, f, g) = 0, we get X X z(τ ∗ , ρ∗ ) = q(e, f, g) + Pi (τ ∗ , ρ∗i ) (e,f,g)∈P3 (τ ∗ ) = X = X X q(e, f, g) + (e,f,g)∈ P3 (τ ∗ −F ) = i∈C (e,f,g)∈ P3 (τ ∗ )\P3 (τ ∗ −F ) q(e, f, g) + (e,f,g)∈ P3 (τ ∗ /F −vF ) z(τ ∗ /F, ρ∗H/F ). X X q(e, f, g) + i∈C\F βi (ρ∗i ) + X βi (ρ∗i ) i∈F βi (ρ∗i ) + βvF (ρvF ) i∈C\F It remains to show that (τ ∗ /F, ρ∗H/F ) is optimal in H/F . Towards a contradiction, suppose there exists a tour pair (τ 0 /F, ρ0H/F ) 6= (τ ∗ /F, ρ∗H/F ) such that z(τ 0 /F, ρ0H/F ) < z(τ ∗ /F, ρ∗H/F ). The calculations of the minimums in equations (4.2)-(4.9) imply that τ 0 /F 6= τ ∗ /F . Using equations (4.2)-(4.9), we can expand F , extending (τ 0 /F, ρ0H/F ) to (τ 00 , ρ00 ) in H with z(τ 0 /F, ρ0H/F ) = z(τ 00 , ρ00 ). Then z(τ 00 , ρ00 ) = z(τ 0 /F, ρ0H/F ) < 13 z(τ ∗ /F, ρ∗H/F ) = z(τ ∗ , ρ∗ ), contradicting the optimality of (τ ∗ , ρ∗ ). Hence (τ ∗ /F, ρ∗H/F ) is optimal in H/F .   We now show that βvF can be updated in O(|F |)-time, by introducing a structure which allows to optimally chain together the β-values for consecutive nodes in F ∩ C. Refer to the subgraph induced by the nodes {w} ∪ {ua , ua+1 , . . . , ub } as pseudo-fan P Fa,b for (1 ≤ a ≤ b ≤ r). Define the minimum penalty associated with pseudo-fan P Fa,b to be P Fa,b (c, d) for a, b ∈ {1, . . . , r} with inner structure c ∈ {0, 1}2 and outer structure d = (d1 , d2 ) ∈ {(1, 3), (1, 4), (2, 3), (2, 4)}. Note that the inner structure of pseudo-fan P Fa,b refers to the edges y1 , t1 and yr−1 , tr within fans Fa and Fb (which are contracted to pseudonodes ua and ub ), respectively. The outer structure refers to the edges α1 , α2 and α3 , α4 with respect to fans Fa and Fb , if ua and ub are considered as pseudonodes. To compute P Fa,b (c, d) we use the following recursion which chains together the minimum β-values through consecutive nodes in C while maintaining feasibility of ρ. For c1 , c2 ∈ {0, 1}, d1 ∈ {1, 2} and d2 ∈ {3, 4}, and 1 ≤ i ≤ r − 1 P F1,1 ((c1 , c2 ), (d1 , d2 )) = βu1 ((c1 , c2 ), (d1 , d2 )) (4.10) P Fr,r ((c1 , c2 ), (d1 , d2 )) = βur ((c1 , c2 ), (d1 , d2 )) (4.11) P F1,i+1 ((c1 , c2 ), (d1 , d2 )) = min {P F1,i ((c1 , s), (d1 , t)) + βui+1 ((t − 3, c2 ), (s + 1, d2 ))} (4.12) min {βui−1 ((c1 , t − 1), (d1 , s + 3)) + P Fi,r ((s, c2 ), (t, d2 ))} (4.13) s∈{0,1},t∈{3,4} and P Fi−1,r ((c1 , c2 ), (d1 , d2 )) = s∈{0,1},t∈{1,2} To prove that the recursions defined by (4.10)-(4.13) are correct, first consider the optimal P F1,i+1 ((c1 , c2 ), (d1 , d2 )). The minimum feasible assignment of penalties for the nodes within pseudo-fan P F1,i+1 is simply the minimum among optimal assignments of penalties for the nodes within pseudo-fan P F1,i and the penalty at pseudonode ui+1 , such that these penalties are feasible. That is, the outer structure of P F1,i must match the inner structure of ui+1 and vice-versa. By definition, this is precisely min P F1,i ((c1 , s), (d1 , t)) + βui+1 ((t − 3, c2 ), (s + 1, d2 )). An analogous argument holds for s∈{0,1},t∈{3,4} P Fi−1,r ((c1 , c2 ), (d1 , d2 )). It is now possible to compute the minimum traversals of F used in equations (4.2)-(4.9). For example, middle paths through F have a cost which minimizes the sum of the penalties on P F1,i−1 + βi and βi+1 + P Fi+2,r , such that both pairs are feasible. Note that the minimum ρ(F ) has been found while performing the recursion defined by equations (4.10)-(4.13). In the worst case, when all outer nodes in F belong to S, we must compute P F1,1 , . . . , P F1,r−1 and P Fr,r , . . . , P F2,r , which can be done in O(|F |)-time. By precomputing these, one can evaluate the minimum traversals of F used in equations (4.2)-(4.9), in O(|F |)-time and hence one can update βvF in O(|F |)-time. We iteratively perform the fan contraction operation, updating costs and penalties until we are left with a wheel. The optimal tour in H skirts the cycle C and detours exactly once through centre w, skipping exactly one edge of C. Orient the cycle C in the clockwise direction. τ contains all edges in C except for the skipped edge, say ci = (ui , ui+1 ), together with the two edges which detour around ci . Define function φ(ci ) for each edge ci = (ui , ui+1 ) ∈ E(C). Let ti and ti+1 be the tree edges adjacent to ui and ui+1 , respectively. We will define φ(ci ) = q(ci−2 , ci−1 , ti ) + q(ci−1 , ti , ti+1 ) + q(ti , ti+1 , ci+1 ) +q(ti+1 , ci+1 , ci+2 ) − q(ci−2 , ci−1 , ci ) − q(ci−1 , ci , ci+1 ) −q(ci , ci+1 , ci+2 ). Then, the optimal tour pair has     X z(τ ∗ , ρ∗ ) = q(C) + min φ(ci ) + min βuj (ρuj ) . i:ci ∈C   feasible ρ uj ∈C Suppose that we fix an edge ci in τ . Then H can be considered to be a fan as shown in Figure 12. 14 ci ui ci−1 ti ci−2 ui ci−1 ci+1 ti+1 ci w ti ci−2 ui+1 ui+1 ci+2 ci+1 ti+1 ci+2 w Figure wheel,which which skips Figure7. 11.AAtour tour ττ in in aa wheel, skips edgeedge ci . ci . ci−1 ci+1 ti+1 ti w Figure ui 7. ci−1 ui+1 edge c . A tour τ in a wheel, which skips i ti+1 ti ci ci ci+1 ti+1 ti u u i+1 ui w i ci−1 ci+1 ui+1 t Figure 8. Wheel i+1 H with center twi considered as a fan. ci ci ui+1 ui ci−1 ci+1 Figure 8. Wheel H with center w considered as a fan. Figure 12. Wheel H with centre w considered as a fan. Note that edge ci is fixed and the fan ‘wraps’ around the wheel, reusing edges ci−1 , ci , ci+1 , ti−1 , ti and ti+1 . F1 F2 F3 Figure 9. A subgraph of H which becomes a fan after contracting F1 , F2 and F3 . The bold edges depict a centre-traversal of F contained in Hamilton cycle τ . A (τ, ρ) pair containing ρvF1 = ((1, 0), (2, 3)), ρvF2 = ((0, 1), (1, 3)) and ρvF3 = ((0, 0), (2, 6)) is feasible. The edges which correspond to the inner F1 F2 F3 structure of ρvF2 are coloured in blue and the edges which correspond to the outer structure of ρvF2 are red. Figure 9. A subgraph of coloured H whichinbecomes a fan after contracting F , F Fix edge cr in τ and consider H to be a fan Fcr . Then the minimum tour in H can be determined by 1 2 calculating theFminimum of theedges minimum centre-traversal of Fcr and which bypasses cr (using, say, and depict a centre-traversal ofthe F tour contained in Hamilton 3 . The bold fan Fc1 ). cycle This can in O(r)-time using technique above. τ . be A computed (τ, ρ) pair containing ρvFthe =pseudofan ((1, 0), (2, 3)), ρvdescribed = ((0, 1), (1, 3)) F2 1 The preceeding discussion yields the following algorithm. and ρvF3 = ((0, 0), (2, 6)) is feasible. The edges which correspond to the inner 15 4 structure of ρvF2 are coloured in blue and the edges which correspond to the outer structure of ρvF2 are coloured in red. Algorithm 1 HalinTSP(3)(H, q, β) Input: Halin graph H, quadratic cost function q, and penalty function β if H is a wheel then Use the wheel procedure to find an optimal tour τ in H else Let F be a fan in H Contract F to a single node vF , using the Case 1 procedure. That is, assign penalties β to vF , assign costs of 0 to all triples in H/F and assign costs of q to all remaining triples which are 3-neighbours in H. HalinTSP(3)(H/F , q, β) end if Expand all pseudonodes in reverse order and update τ return τ Each time that a fan is contracted, the number of tree nodes is reduced by 1, and hence the fan contraction operation is performed one less than the number of non-leaf nodes in H. The fan contraction operation can be performed in O(|F |)-time, and each time it is performed, the number of nodes in H is reduced by |F | − 1. Since the wheel procedure takes O(n)-time, the total time for the algorithm is O(n). 4.2. TSP(k). We now show that the previous ideas can be extended to solve TSP(k). For any subgraph G of H, let Pk (G) be the collection of all distinct candidate k-paths in G. For each candidate k-path (e1 , e2 , . . . ek ) ∈ Pk (H), define q(e1 , ek−1 ) + q(e2 , ek ) q(e1 , ek−2 ) + q(e2 , ek−1 ) + q(e3 , ek ) + 2 3 c(e1 ) + c(e2 ) + . . . + c(ek ) + ... + . k Now consider the simplified problem: X ST SP (k) : Minimize q(e1 , e2 , . . . , ek ) q(e1 , e2 , . . . , ek ) =q(e1 , ek ) + (4.14) (e1 ,e2 ,...,ek )∈Pk (τ ) Subject to τ ∈ F. Theorem 4.6. Any optimal solution to the STSP(k) is also optimal for TSP(k). Proof. Using equation (4.14), the proof of this follows along the same way as that of Theorem 4.4 and hence is omitted.   As a result of Corollary 4.2, the preceding algorithm can be extended to solve TSP(k) by extending the penalty functions at outer nodes to accommodate subpaths of length 2d(k+1)/2e . A complete description of these varies as the information that needs to be stored is more involved and is hence omitted. Some details however will be available in [32]. The complexity increases by a factor of (2d(k+1)/2e ), which is constant for fixed k and polynomially bounded when k = t log n. 5. TSP(k) on fully reducible graph classes We say that a class C of 3-connected graphs is fully reducible if it satisfies the following: (1) If G ∈ C has a 3-edge cutset which partitions G into components S and S̄, then both G/S and G/S̄ are in C and we call G a reducible graph in C; and (2) TSP can be solved in polynomial time for the graphs in C that do not have non-trivial 3-edge cutsets, i.e. 3-edge cutsets that leave both components Hamiltonian. We call such graphs irreducible. For instance, Halin graphs can be understood as graphs built up from irreducible fans connected to the remainder of the graph via 3-edge cutsets. Cornuejols et al. [10] show that the ability to solve TSP in polynomial time on irreducible graphs in C allows to solve TSP in polynomial time on all of C using facts about the TSP polyhedron. 16 We remark that the algorithm of Section 4 can be used to show a somewhat similar statement for TSP(k). Here we consider a graph class C that is fully k-reducible in the sense that either it can be subdivided into irreducible graphs via 3-edge cutsets, or it is irreducible and it is possible to solve the k-neighbour Hamiltonian path problem in polynomial time. This requires solving the following problem: X X M T SP (k) : Minimize q(e1 , . . . , ek ) + Pi (τ ) (e1 ,...,ek )∈Pk (τ ) Subject to i∈V τ ∈F where Pi (τ ) is a penalty function for the pseudonode which depends on how tour τ traverses i, analogous to the construction for the 3-neighbour TSP of section 4. We recursively perform the contraction operation on the irreducible subgraphs of G, storing the necessary tour information in the penalty at the resulting pseudonode. A similar result to Corollary 4.2 may be derived to show that for any fixed k, this requires a polynomial number of penalties. The least cost traversals of S can be computed in polynomial time using a generalization of the pseudo-fan strategy above. Suppose the contraction operation is performed on a subgraph of size r in time O(P (r)), where P (r) is a polynomial in r. Each time this operation is performed, the number of nodes in the graph is reduced by r. This operation is performed at most n times and it follows that the entire algorithm can be performed in polynomial time. 6. Conclusions In this paper, we have shown that QTSP is NP-hard even when the costs are restricted to taking 0-1 values on Halin graphs. We have presented a polynomial time algorithm to solve a restriction of QTSP, denoted TSP(k) on any fully k-reducible graph for any fixed k. To illustrate this, we have given an algorithm which solves TSP(3) on a Halin graph in O(n) time. The k-neighbour bottleneck TSP on a Halin graph can be solved by solving O(log(n)) problems of the type TSP(k). However, it is possible to solve the problem faster. Details will be reported elsewhere. We would also like to thank Ante Ćustić for his useful comments. References [1] A. Aggarwal, D. Coppersmith, S. Khanna, R. Motwani, and B. Schieber. The angular-metric traveling salesman problem. SIAM Journal on Computing, 29:697–711, 2000. [2] D. Applegate, R. Bixby, V. Chvatal, and W. Cook. The traveling salesman problem: a computational study. Princeton University Press, 2011. [3] E. Arkin, Y. Chiang, J. Mitchell, S. Skiena, and T. Yang. On the maximum scatter traveling salesperson problem. SIAM Journal on Computing, 29:515–544, 1999. [4] E. Balas, R. Carr, M. Fischetti, and N. Simonetti. New facets of the STS polytope generated from known facets of the ATS polytope. Discrete Optimization, 3:3–19, 2006. [5] E. Balas, M. Fischetti, and W. Pulleyblank. The precedence-constrained asymmetric traveling salesman polytope. Mathematical Programming, 68:241–265, 1995. [6] H. Bodlaender. Dynamic programming on graphs with bounded treewidth. In Automata, languages and programming (Tampere, 1988),Lecture Notes in Comput. Sci., 317:105–118. Springer, Berlin, 1988. [7] W. Chin and S. Ntafos. Optimum watchman routes. In Proceedings of the Second Annual Symposium on Computational Geometry, 24–33. ACM, 1986. [8] W. Cook. In pursuit of the traveling salesman: mathematics at the limits of computation. Princeton University Press, 2012. [9] G. Cornuéjols, D. Naddef, and W. Pulleyblank. Halin graphs and the travelling salesman problem. Mathematical Programming, 26:287–294, 1983. [10] G. Cornuéjols, D. Naddef, and W. Pulleyblank. The traveling salesman problem in graphs with 3-edge cutsets. Journal of ACM, 32:383–410, April 1985. [11] B. Courcelle. The monadic second-order logic of graphs. I. Recognizable sets of finite graphs. Inform. and Comput., 85:12– 75, 1990. [12] F. Della Croce, V. Paschos, and R. Calvo. Approximating the 2-peripatetic salesman problem. In Proc. Workshop on Modelling and Algorithms for Planning and Scheduling Problems, MAPSP, 114–116, 2005. [13] O. Ergun and J. Orlin. A dynamic programming methodology in very large scale neighborhood search applied to the traveling salesman problem. Discrete Optimization, 3:78–85, 2006. The Traveling Salesman Problem. 17 [14] A. Fischer. An analysis of the asymmetric quadratic traveling salesman polytope. SIAM Journal on Discrete Mathematics, 28:240–276, 2014. [15] A. Fischer and F. Fischer. An extended approach for lifting clique tree inequalities. Journal of Combinatorial Optimization, 30:489–519, 2015. [16] A. Fischer, F. Fischer, G. Jäger, J. Keilwagen, P. Molitor, and I. Grosse. Exact algorithms and heuristics for the quadratic traveling salesman problem with an application in bioinformatics. Discrete Applied Mathematics, 166:97–114, 2014. [17] A. Fischer and C. Helmberg. The symmetric quadratic traveling salesman problem. Mathematical Programming, 142:205– 254, 2013. [18] G. Gutin and A. Punnen, editors. The traveling salesman problem and its variations, volume 12 of Combinatorial Optimization. Springer New York, 2002. [19] G. Jäger and P. Molitor. Algorithms and experimental study for the traveling salesman problem of second order. In B. Yang, D. Du, and C. Wang, editors, Combinatorial Optimization and Applications, Lecture Notes in Computer Science, 5165:211–224. Springer Berlin, 2008. [20] J. Krarup and I. Spadille. The peripatetic salesman and some related unsolved problems. Combinatorial programming: methods and applications: proceedings of the NATO Advanced Study Institute held at the Palais des Congrès, Versailles, France, 2-13 September 1974, 173, 1975. [21] J. LaRusic and A. Punnen. The asymmetric bottleneck traveling salesman problem: algorithms, complexity and empirical analysis. Computers & Operations Research, 43:20–35, 2014. [22] J. LaRusic, A. Punnen, and E. Aubanel. Experimental analysis of heuristics for the bottleneck traveling salesman problem. Journal of heuristics, 18:473–503, 2012. [23] E. Lawler, J. Lenstra, A. Kan, and D. Shmoys. The Traveling Salesman Problem: a guided tour of combinatorial optimization, volume 3. Wiley New York, 1985. [24] D. Lou. An algorithm to find the optimal matching in Halin graphs. IAENG International Journal of Computer Science, 34:220–226, 2007. [25] J. Phillips, A. Punnen, and S. Kabadi. A linear algorithm for the bottleneck traveling salesman problem on a Halin graph. Information Processing Letters, 67:105–110, 1998. [26] A. Oswin, A. Fischer, F. Fischer, F. Meier, U. Pferschy, A. Pilz, and R. Staněk Minimization and maximization versions of the quadratic travelling salesman problem. Optimization, 66:521–546, 2017. [27] G. Reinelt. The traveling salesman: computational solutions for TSP applications. Springer-Verlag, 1994. [28] B. Rostami, F. Malucelli, P. Belotti and S. Gualandi. Quadratic TSP: A lower bounding procedure and a column generation approach Computer Science and Information Systems (FedCSIS), 2013 Federated Conference on, 377-384, September 2013. [29] B. Rostami, F. Malucelli, P. Belotti, and S. Gualandi. Lower bounding procedure for the asymmetric quadratic traveling salesman problem. European Journal of Operational Research, 253:584–592, 2016. [30] R. Staněk, Problems on tours and trees in combinatorial optimization, PhD Thesis, Graz University of Technology, 2016. [31] B. Woods. Generalized travelling salesman problems on Halin graphs. Master’s thesis, Simon Fraser University, Canada, 2010. [32] B. Woods. The quadratic travelling salesman problem: complexity, approximations, and exponential neighbourhoods. PhD. thesis In preparation, Simon Fraser University, Canada. [33] B. Woods and A. Punnen. The quadratic travelling salesman problem on Halin graphs. Working paper, Simon Fraser University, Canada, 2016. (SFU) Department of Mathematics, Simon Fraser University, 250-13450 102 Avenue, Surrey, British Columbia, V3T 0A3, Canada 18 

8cs.DS

arXiv:1708.03981v1 [cs.SY] 14 Aug 2017 PSSE Redux: Convex Relaxation, Decentralized, Robust, and Dynamic Approaches Vassilis Kekatos, Gang Wang, Hao Zhu, and Georgios B. Giannakis ∗ August 15, 2017 1 Introduction With the advent of digital computers, power system engineers in the 1960s tried computing the voltages at critical buses based on readings from current and potential transformers. Local personnel manually collected these readings and forwarded them by phone to a control center. Nevertheless, due to timing, modeling, and instrumentation inaccuracies, the power flow equations were always infeasible. In a seminal contribution [61], the statistical foundations were laid for a multitude of grid monitoring tasks, including topology detection, static state estimation, exact and linearized models, bad data analysis, centralized and decentralized implementations, as well as dynamic state tracking. Since then, different chapters, books, and review articles have nicely outlined the progress in the area; see for example [74, 55, 1]. The revolutionary monitoring capabilities enabled by synchrophasor units have been put forth in [58]. This chapter aspires to glean some of the recent advances in power system state estimation (PSSE), though our collection is not exhaustive by any means. The Cramér-Rao bound, a lower bound on the (co)variance ∗ V. Kekatos is with the Bradley Department of Electrical and Computer Engineering, Virginia Tech, Blacksburg, VA 24061, USA. G. Wang is with the Department of Electrical and Computer Engineering, University of Minnesota, Minneapolis, MN 55455, USA, and also with the State Key Lab of Intelligent Control and Decision of Complex Systems, Beijing Institute of Technology, Beijing 100081, P. R. China. H. Zhu is with the Department of Electrical and Computer Engineering, University of Illinois, Urbana, IL 61801. G. B. Giannakis is with the Department of Electrical and Computer Engineering, University of Minnesota, Minneapolis, MN 55455, USA. E-mails: kekatos@vt.edu; {gangwang, georgios}@umn.edu; haozhu@illinois.edu. 1 of any unbiased estimator, is first derived for the PSSE setup. After reviewing the classical Gauss-Newton iterations, contemporary PSSE solvers leveraging relaxations to convex programs and successive convex approximations are explored. A disciplined paradigm for distributed and decentralized schemes is subsequently exemplified under linear(ized) and exact grid models. Novel bad data processing models and fresh perspectives linking critical measurements to cyber-attacks on the state estimator are presented. Finally, spurred by advances in online convex optimization, model-free and model-based state trackers are reviewed. Notation: Lower- (upper-) case boldface letters denote column vectors (matrices), and calligraphic letters stand for sets. Vectors 0, 1, and en denote respectively the all-zero, all-one, and the n-th canonical vectors of suitable dimensions. The conjugate of a complex-valued object (scalar, vector or matrix) x is √ denoted by x∗ ; <{x} and ={x} are its real and imaginary parts, and j := −1. Superscripts T and H stand for transpose and conjugate-transpose, respectively, while Tr(X) is the trace of matrix X. A diagonal matrix having vector x on its main diagonal is denoted by dg(x); whereas, the vector of diagonal entries of X is dg(X). The range space of X is denoted by range(X); and its null space (kernel) by null(X). The notation N (µ, Σ) represents the Gaussian distribution with mean µ and covariance matrix Σ. 2 Power Grid Modeling This section introduces notation and briefly reviews the power flow equations; for detailed exposition see e.g., [1], [19] and references therein. A power system can be represented by the graph G = (B, L), where the node set B comprises its Nb buses, and the edge set L its Nl transmission lines. Given the focus on alternating current (AC) power systems, steady-state voltages and currents are represented by their single-phase equivalent phasors per unit. A transmission line (n, k) ∈ L running across buses n, k ∈ B is modeled by its total series admittance ynk = gnk + jbnk , and total shunt susceptance jbsnk . If Vn is the complex voltage at bus n, the current Ink flowing from bus m to bus n over line (m, n) is Ink = (ynk + jbsnk /2)Vn − ynk Vn . (1) The current Inm coming from the other end of the line can be expressed symmetrically. That is not the case if the two buses are connected via a 2 transformer with complex ratio ρnk followed by a line, where ynk + jbsnk /2 ynk Vn − ∗ Vn 2 |ρnk | ρnk ynk = (ynk + jbsnk /2)Vn − Vn . ρnk Ink = (2a) Inm (2b) Kirchhoff’s current law dictates that the current injected into bus n is P In = n∈Bn Ink , where Bn denotes the set of buses directly connected to bus n. If vector i ∈ CNb collects all nodal currents, and v ∈ CNb all nodal voltages, the two vectors are linearly related through the bus admittance matrix Y = G + jB as i = Yv. (3) Similar to (3), line currents can be stacked in the 2Nl -dimensional vector if , and expressed as a linear function of nodal voltages if = Yf v (4) for some properly defined 2Nl × Nb complex matrix Yf [cf. (1)–(2)]. The complex power injected into bus n will be denoted by Sn := Pn + jQn . Since by definition Sn = Vn In∗ , the vector of complex power injections s = p + jq can be expressed as s = dg(v)i∗ = dg(v)Y∗ v∗ . (5) ∗ . The power flowing from bus n to bus n over line (m, n) is Snk = Vn Ink jθ If voltages are expressed in polar form Vn = Vn e n , the power flow equations in (5) per real and imaginary entry can be written as Nb X Pn = n=1 Nb X Qn = Vn Vn [Gnk cos(θn − θn ) + Bnk sin(θn − θn )] (6a) Vn Vn [Gnk sin(θn − θn ) − Bnk cos(θn − θn )] . (6b) n=1 Since power injections are invariant if voltages are shifted by a common angle, the voltage phase is arbitrarily set to zero at a particular bus called the reference bus. Alternatively to (6), if voltages are expressed in rectangular coordinates Vn = Vr,m + jVi,m , power injections are quadratically related to voltages Pn = Vr,m Nb X (Vr,n Gnk − Vi,n Bnk ) + Vi,m n=1 Nb X (Vi,n Gnk + Vr,n Bnk ) n=1 3 (7a) Qn = Vi,m Nb X (Vr,n Gnk − Vi,n Bnk ) − Vr,m Nb X (Vi,n Gnk + Vr,n Bnk ). (7b) n=1 n=1 To compactly express (7), observe that Sn∗ = Vn∗ In = (vH en )(eTn i) = vH en eTn Yv from which it readily follows that Pn = vH HPn v (8a) Qn = vH HQn v (8b) where the involved matrices are defined as
1 en eTn Y + YH en eTn 2
1 en eTn Y − YH en eTn . := 2j HPn := (9a) HQn (9b) Similar expressions hold for the squared voltage magnitude at bus n: Vn2 = vH HVn v, where HVn := en eTn . (10) Realizing that a line current can also be provided as Ink = eTnk if , the ∗ = V ∗I power flow on line (n, k) as seen from bus n is expressed as Snk n nk = H T H T (v en )(enk if ) = v en enk Yf v, from which it follows that Pnk = vH HPnk v H Qnk = v HQnk v (11a) (11b) where HPnk and HQnk are defined by substituting eTn Y and YH en by eTnk Yf and YfH enk in (9), accordingly. Equations (8), (10) and (11) explain how power injections, flows, and squared voltage magnitudes are quadratic functions of voltage phasors as described by vH Hm v for certain complex Nb × Nb matrices Hm . Regardless if Y and/or Yf are symmetric or Hermitian, Hm are Hermitian by definition. T = <{H }, and This means that Hm = HH m m , or equivalently, <{Hm } ={Hm }T = −={Hm }. It can be easily verified that the quadratic functions can be expressed in terms of real-valued quantities as vH Hm v = v̄T H̄m v̄ (12) for the expanded real-valued voltage vector v̄ := [<{v}T ={v}T ]T , and the real-valued counterpart of Hm , namely   <{Hm } −={Hm } H̄m := . (13) ={Hm } <{Hm } 4 3 Problem Statement It was seen in Section 2 that given grid parameters collected in Y and Yf , all power system quantities can be expressed in terms of the voltage vector v, which justifies its term as the system state. Meters installed across the grid measure electric quantities, and forward their readings via remote terminal units to a control center for grid monitoring. Due to lack of synchronization, conventional meters cannot utilize the angle information of phasorial quantities. For this reason, legacy measurements involve phaseless power injections and flows along with voltage and current magnitudes at specific buses. The advent of the global positioning system (GPS) facilitated a precise timing signal across large geographical areas, thus enabling the revolutionary technology of synchrophasors or phasor measurement units (PMUs) [58]. Recovering bus voltages given network parameters and the available measurements constitutes the critical task of power system state estimation. This section formally states the problem, provides the Cramér-Rao bound on the variance of any unbiased estimator, and reviews the Gauss-Newton iterations. Solvers based on semidefinite relaxation and successive convex approximations are subsequently explicated, and the section is wrapped up with issues germane to PMUs. 3.1 Weighted Least-Squares Formulation Consider M real-valued measurements {zm }M m=1 related to the complex power system state v through the model zm = hm (v) + m (14) where hm (v) : CNb → R is a (non)-linear function of v, and m captures the measurement noise and modeling inaccuracies. Collecting measurements and noise terms in vectors z and  accordingly, the vector form of (14) reads z = h(v) +  (15) for the mapping h : CNb → RM . Model (15) is instantiated for different types of measurements next. Traditionally, the system state v is expressed in polar coordinates, namely nodal voltage magnitudes and angles. Then h(v) maps the 2Nb -dimensional state vector to SCADA measurements through the nonlinear equations (6). Expressing the states in polar form has been employed primarily due to two reasons. First, the Jacobian matrix of h(v) is amenable to approximations. 5 Secondly, voltage magnitude measurements are directly related to states. Nevertheless, due to recent computational reformulations, most of our exposition models voltages in the rectangular form. Then, as detailed in (12), the m-th SCADA measurement zm involves the quadratic function of the state hm (v) = vH Hm v for a Hermitian matrix Hm . Expressing voltages in rectangular coordinates is computationally advantageous when it comes to synchrophasors too. As evidenced by (3)–(4), PMU measurements feature linear mappings hm (v). If PMU measurements are expressed in rectangular coordinates, the model in (14) simplifies to z = Hv +  (16) for an M × Nb complex matrix H, and complex-valued noise . Following the notation of (12)–(13), the linear measurement model of (16) can be expressed in terms of real-valued quantities as z̄ = (H∗ )v̄ + ¯. (17) The random noise vector  in (15) is usually assumed independent of h(v), zero-mean and circularly symmetric, that is E[H ] = Σ and E[T ] = 0. The last assumption holds if for example the real and imaginary components of  are independent and have identical covariance matrices. This is true for a PMU measurement, where the actual state lies at the center of a spherically-shaped noise cloud on the complex plane. Moreover, the entries of  are oftentimes assumed uncorrelated yielding 2 }) with σ 2 being the variance of the a diagonal covariance Σ = dg({σm m m-th entry m . However, that may not always be the case. For example, active and reactive powers at the same grid location are derived as products between the readings of a current transformer and a potential transformer. Further, noise terms may be correlated between the real and imaginary parts of the same phasor in a PMU. Adopting the weighted least-squares (WLS) criterion, power system state estimation can be formulated as minimize kΣ−1/2 (z − h(v))k22  v∈CNb (18) −1/2 where Σ is the matrix square root of the inverse noise covariance matrix. If the noise is independent across measurements, then (18) simplifies to M X (zm − hm (v))2 minimize . 2 v σm m=1 6 (19) Either way, the PSSE task boils down to a (non)-linear least-squares (LS) fit. When the mapping h(v) is linear or when the entries of  are uncorrelated, the measurement model in (15) can be prewhitened. For example, the linear measurement model z = Hv +  can be equivalently transformed to Σ−1/2 z = (Σ−1/2 H)v + Σ−1/2   (20) −1/2 so that the associated noise Σ  is now uncorrelated. To ease the presentation, the noise covariance will be henceforth assumed Σ = IM , yielding v̂ := arg min v M X (zm − hm (v))2 . (21) m=1 For Gaussian measurement noise  ∼ N (0, IM ), the minimizer of (21) coincides with the maximum likelihood estimate (MLE) of v [29]. 3.2 Cramér-Rao Lower Bound Analysis According to standard results in estimation theory [29], the variance of any unbiased estimator is lower bounded by the Cramér-Rao lower bound (CRLB). Appreciating its importance as a performance benchmark across different estimators, the ensuing result shown in the Appendix derives the CRLB for any unbiased power system state estimator based on the so-termed Wirtinger’s calculus for complex analysis [41]. Proposition 1. Consider estimating the unknown state vector v ∈ CNb from the noisy SCADA data {zm }M m=1 of (15), where the Gaussian measurement error m is independent across meters with mean zero and variance 2 . The covariance matrix of any unbiased estimator v̂ satisfies σm Cov(v̂)  [F† (v, v∗ )]1:Nb ,1:Nb (22) where the Fisher information matrix (FIM) is given as " PM # PM 1 1 H ∗ ∗ H 2 (Hm v)(Hm v) 2 (Hm v)(Hm v ) m=1 σm m=1 σm ∗ PM F(v, v )= PM . 1 1 ∗ ∗ H ∗ ∗ ∗ ∗ H 2 (Hm v )(Hm v) 2 (Hm v )(Hm v ) m=1 σm m=1 σm (23) ∗ In addition, matrix F(v, v ) has at least rank-one deficiency even when all possible SCADA measurements are available. Although rank-deficient, the pseudo-inverse of F(v, v∗ ) qualifies as a valid lower bound on the mean-square error (MSE) of any unbiased estimator [63]. Rank deficiency of the FIM originates from the inherent voltage 7 angle ambiguity: SCADA measurements remain invariant if nodal voltages are shifted globally by a unimodular phase constant. Fixing the angle of the reference bus waives this issue. It is also worth stressing that the CRLB in Prop. 1 is oftentimes attainable and benchmarks the optimal estimator performance [63]. Having derived the CRLB for the PSSE task, our next subsection deals with PSSE solvers. 3.3 Gauss-Newton Iterations Consider for specificity model (16), though the real-valued model in (17) or the model involving polar coordinates could be employed as well. When the noise covariance matrix Σ = IM , the PSSE task in (18) reduces to the nonlinear LS problem minimize kz − h(v)k22 (24) v∈CNb for which the Gauss-Newton iterations are known to offer the “workhorse” solution [3, Ch. 1], [1, Ch. 2]. According to the Gauss-Newton method, the function h(v) is linearized at a given point vi ∈ CNb using Taylor’s expansion as h̃(v, vi ) := h(vi ) + Ji (v − vi ) where Ji := ∇h(vi ) is the M × Nb Jacobian matrix of h evaluated at vi , whose (m, n)-th entry is given by the Wirtinger derivative ∂hm /∂Vn ; see e.g., [41] for Wirtinger’s calculus. The Gauss-Newton method subsequently approximates the nonlinear LS fit in (24) with a linear one of h̃, and relies on its minimizer to obtain the next iterate as vi+1 ∈ arg min v = arg min v z − h̃(v, vi ) z − h(vi ) 2 2 − 2(v − vi )H (Ji )H (z − h(vi )) + (v − vi )H (Ji )H Ji (v − vi ). (25) When matrix (Ji )H Ji is invertible, vi+1 can be found in closed form as  −1 vi+1 = vi + (Ji )H Ji (Ji )H (z − h(vi )). (26) The state estimate is iteratively updated using (26) until some stopping criterion is satisfied. If, on the other hand, the WLS cost (18) is minimized, the Gauss-Newton −1/2 −1/2 iterations can be similarly obtained by treating Σ z as z and Σ h(v) as h(v) in (24), yielding   i −1 i H −1 (J ) Σ (z − h(vi )). (27) vi+1 = vi + (Ji )H Σ−1  J 8 It is well known that the pure Gauss-Newton iterations in (26) or (27) may not guarantee convergence, which in fact largely depends on the starting point v0 [3, Ch. 1.5]. A common way to improve convergence and ensure descent of the cost in (18) consists of including a backtracking line search in (27) to end up with   i −1 i H −1 vi+1 = vi + µi (Ji )H Σ−1 (J ) Σ (z − h(vi ))  J (28) where the step size µi > 0 is found through the backtracking line search rule [3, Ch. 1.2]. Due to its intimate relationship with ordinary gradient descent alternatives for nonconvex optimization however, this Gauss-Newton iterative procedure can be trapped by local solutions [3, Ch. 1.5]. In a nutshell, the grand challenge remains to develop PSSE solvers capable of attaining or approximating the global optimum at manageable computational complexity. A few recent proposals in this direction are presented next. 3.4 Semidefinite Relaxation A method to tackle the nonlinear measurement model that can convert the PSSE problem of (21) to a convex semidefinite program (SDP) has been introduced in [83], [84]. Consider first expressing each measurement in z linearly in terms of the outer-product matrix V := vvH . In this way, the quadratic models in (8), (10), and (11) can be transformed to linear ones in terms of the matrix variable V. Thus, each noisy measurement in (14) can be written as zm = vH Hm v + m = Tr(Hm V) + m . Rewriting the PSSE task in (21) accordingly in terms of V reduces to V̂1 := arg min V∈CNb ×Nb M X  2 zm − Tr(Hm V) (29a) m=1 s. to V  0, and rank(V) = 1 (29b) where the positive semi-definite (PSD) and the rank-1 constraints jointly ensure that for any V obeying (29b), there always exists a vector v such that V = vvH . Although zm and V are linearly related as in (29), nonconvexity is still present in two aspects: (i) the cost function in (29a) has degree 4 in the entries of V; and (ii) the rank constraint in (29b) is nonconvex. Aiming for an SDP reformulation of (29), Schur’s complement lemma, see e.g., [5, Appx. 5.5], can be leveraged to tightly bound each summand in (29a) using 9 an auxiliary variable χm > 0. Collecting all χm ’s in χ ∈ Rm , the problem in (29) can be expressed as {V̂2 , χ̂2 } := arg min 1T χ (30a) V, χ s. to V  0, and rank(V) = 1 (30b)   χm zm − Tr(Hm V)  0, ∀m. (30c) zm − Tr(Hm V) 1 The equivalence among all three SE problems (21), (29), and (30) has been shown in [83], where their optimal solutions satisfy:  2 V̂1 = V̂2 = v̂v̂H , and χ̂2,m = zm − Tr(Hm V̂2 ) , ∀m. (31) The only source of nonconvexity in the equivalent SE problem of (30) comes from the rank-1 constraint. Motivated by the technique of semidefinite relaxation (SDR); see e.g., the seminal work of [21], one can obtain the following convex SDP upon dropping the rank constraint {V̂, χ̂} := arg min 1T χ V, χ s. to V  0, and (30c). (32a) (32b) For the SDR-PSSE formulation in (32), a few assumptions have been made in [84] to establish its global optimality in a specific setup. (as1) The graph G = (B, L) has a tree topology; (as2) Every bus is equipped with a voltage magnitude meter; and (as3) All measurements in z are noise-free, that is  = 0. Proposition 2. Under (as1)-(as3), solving the relaxed problem (32) attains the global optimum of the original PSSE problem (30) or (21); that is, rank(V̂) = 1. Assumptions (as1)-(as3) may offer a close approximation of the realistic PSSE scenario, thanks to characteristics of transmission systems such as sparse connectivity, almost flat voltage profile, and high metering accuracy. Although they do not hold precisely in realistic transmission systems, nearoptimality of the relaxed problem (32) has been numerically supported by extensive tests [84]. A more crucial issue is to recover a feasible SE solution from the relaxed problem (32), as V̂ is very likely to have rank greater than 1. This is possible either by finding the best rank-1 approximation to V̂ via eigenvalue decomposition, or via randomization [46]. 10 SDR endows SE with a convex SDP formulation for which efficient schemes are available to obtain the global optimum using for example the interior-point solver. The computational complexity for eigen-decomposition is in the order of matrix multiplication, and thus negligible compared to solving the SDP; see [46] and references therein. However, the polynomial complexity of solving the SDP could be a burden for real-time power system monitoring, which motivates well the distributed implementation of Section 4.2. 3.5 Penalized Semidefinite Relaxation Building on an alternative formulation of the power flow problem, a penalized version of the aforementioned SDP-based state estimator has been devised in [48, 47, 78]. Commencing with the power flow task, it can be interpreted as a particular instance of PSSE, where: • Measurements (henceforth termed specifications) are noiseless; • Excluding the reference bus, buses are partitioned into the subset BPV for which active injections and voltage magnitudes are specified, and the subset BPQ , for which active and reactive injections are specified. The power flow task can be posed as the feasibility problem; that is, find v ∈ CNb s. to Pn = vH HPn v, H Qn = v HQn v, Vn2 = vH HVn v, (33) ∀n ∈ BPV ∪ BPQ ∀n ∈ BPQ 2 = V0 . ∀n ∈ BPV , and Vref Using the SDP reformulation presented earlier, the power flow task can be equivalently expressed as find V ∈ CNb ×Nb (34) s. to zm = Tr(Hm V), m = 1, . . . , 2Nb − 1 V  0, and rank(V) = 1 where the specifications (constraints) of (33) have been generically captured b −1 by the pairs {(zm ; Hm )}2N m=1 . Although the optimization in (34) is non-convex, a convex relaxation can be obtained by dropping the rank constraint. To promote rank-one solutions, the feasibility problem is further reduced to [48] V̂ := arg min Tr(H0 V) (35) V0 11 s. to zm = Tr(Hm V), m = 1, . . . , 2Nb − 1. (36) If the Hermitian matrix H0 is selected such that H0  0, rank(H0 ) = Nb −1, and H0 1 = 0; then any state v close to the flat voltage profile 1 + j0 can be recovered from the rank-one minimizer V̂ of (35); see [48, Thm. 2]. The stated conditions exclude the case H0 = INb that would have led to the nuclear-norm heuristic commonly used in low-rank matrix completion. Spurred by this observation, the PSSE task can be posed as [47] V̂µ := arg min Tr(H0 V) + µ V0 M X fm zm − Tr(Hm V)
(37) m=1 for some regularization parameter µ ≥ 0, where M now can be larger than 2Nb − 1. The second term in the cost of (37) is a data-fitting term ensuring that the recovered state is consistent with the collected measurements based on selected criteria. Two cases for fm of special interest are the LS fit LS () := 2 , and the least-absolute value (LAV) one f LAV () := ||, ∀m = fm m 1, . . . , M . On the other hand, the first term in (37) can be understood as a regularizer to promote rank-one solutions V̂µ ; see more details in [47]. In the noiseless setup, where all measurements comply with the model zm = v0H Hm v0 , the minimization in (37) has been shown to possess a rankLS and f LAV ; see one minimizer V̂µ = v̂µ v̂µH for all µ ≥ 0 under both fm m details in [47]. Interestingly, the solution v̂µ obtained under the LS fit does not coincide with v0 for any µ ≥ 0, whereas the LAV solution v̂µ provides the actual state v0 for a sufficiently large µ. Error bounds between V̂µ and v0 v0H under the regularized LAV solution for noisy measurements are established in [78]. 3.6 Feasible Point Pursuit The feasible point pursuit (FPP) method studied in [51] offers another computationally manageable solver for approximating the globally optimal PSSE. As a special case of the convex-concave procedure [44], FPP is an iterative algorithm for handling general nonconvex quadratically constrained quadratic programs (QCQPs) [77]. It approximates the feasible solutions of a nonconvex QCQP by means of a sequence of convexified QCQPs obtained with successive convex inner-restrictions of the original nonconvex feasibility set [77]. The first step in applying FPP to PSSE is a reformulation of (18) into 12 a standard QCQP [71, 70] minimize v∈CNb , χ∈RM χT Σ−1 χ (38a) s. to vH Hm v ≤ zm + χm , H v (−Hm ) v ≤ −zm + χm , 1≤m≤M (38b) 1≤m≤M (38c) where vector χ ∈ RM collects the auxiliary variables {χm ≥ 0}M m=1 . For power flow and power injection measurements, the corresponding Hermitian measurement matrices {Hm } are indefinite in general; thus, both constraints (38b) and (38c) are nonconvex. On the contrary, squared voltage magnitude measurements relate to positive semidefinite matrices {Hm }, so that only the related constraint (38c) is nonconvex. Either way, problem (38) is NPhard, and hence computationally intractable [57]. Using eigen-decomposition, every measurement matrix Hm can be expressed as the sum of a positive and a negative semidefinite matrix as − Hm = H+ m + Hm , so that the constraints in (38) are rewritten as H − vH H+ m v + v Hm v ≤ zm + χm v H H+ mv +v H H− mv ≥ z m − χm (39a) (39b) for m = 1, . . . , M . Observe now that since vH H− m v is a concave function of v, it is upper bounded by its first-order (linear) approximation at any point vi ; that is, i H − i H − i vH H− m v ≤ 2<{(v ) Hm v} − (v ) Hm v . The concave function vH (−H+ m )v can be upper bounded similarly. Based on this observation, the FPP technique replaces the concave functions in the constraints of (39) with their linear upper bounds. The point of linearization at every iteration is chosen to be the previous state estimate. Specifically, initializing with v0 , the FPP produces the iterates {vi+1 , χi+1 } := arg min χT Σ−1  χ v, χ≥0 (40) i H − i H − i s. to vH H+ m v + 2<{(v ) Hm v} ≤ zm + (v ) Hm v + χm , ∀m i H + i H + i vH H− m v + 2<{(v ) Hm v} ≥ zm + (v ) Hm v − χm , ∀m. At every iteration, the FPP technique solves the now convex QCQP in (40). The procedure has been shown to globally converge to a stationary point of the WLS formulation (18) of the PSSE task [71]. Extensions of the developed FPP solver to cope with linear(ized) measurements and bad data are straightforward; see [71] for details. 13 100 Mean-square error (p.u.) Angle error (degrees) 102 102 10-2 Magnitude error (p.u.) 10-4 100 10-5 GN-based SE SDR-based SE FPP-based SE 5 10 15 20 25 10 100 10-1 10-2 10 30 GN-based SE SDR-based SE FPP-based SE Cramer-Rao bound 1 -3 1 2 3 4 5 Types of measurements bus index Figure 1: Left: Voltage magnitude and angle estimation errors per bus for the IEEE 30-bus system. Right: MSEs and CRLB versus types of measurements used for the IEEE 14-bus system using: (i) Gauss-Newton iterations; ii) the SDR-based PSSE; and iii) the FPP-based PSSE. Figure 1 compares Gauss-Newton iterations, the SDR-based solver, and the FPP-based solver on the IEEE 14- and 30-bus systems [64]. The actual nodal voltage magnitudes and angles were generated uniformly at random over [0.9, 1.1] and [−0.4π, 0.4π], respectively. Independent zero-mean Gaussian noise with standard deviation 0.05 for power and 0.02 for voltage measurements was assumed, and all reported results were averaged over 100 independent Monte Carlo realizations. The measurements for the IEEE 30bus system include all nodal voltage magnitudes and the active power flows at both sending and receiving ends. The left panel of Fig. 1 depicts that the magnitude and angle estimation errors attained by the FPP solver are consistently below its competing alternatives [71]. The second experiment examines the MSE performance of the three approaches relative to the CRLB of (22) for the IEEE 14-bus test system. Initially, all voltage magnitudes as well as all sending- and receiving-end active power flows were measured, which corresponds to the base case 1 in the x-axis of the right panel of Fig. 1. To show the MSE performance relative for an increasing number of measurements, additional types of measurements were included in a deterministic manner. All types of SCADA measurements were ordered as {Vn2 , Pkn , Pnk , Qkn , Qnk , Pn , Qn }. Each x-axis value in the right panel of Fig. 1 implies that the number of ordered types of measurements was used in the experiment to obtain the corresponding MSEs. 14 3.7 Synchrophasors To incorporate synchrophasors into the PSSE formulation, let ζn = Φn v+εn collect the noisy PMU data at bus n [cf. (16)]. The related measurement matrix is Φn , and the measurement noise εn is assumed to be complex zeromean Gaussian, independent from the noise  in legacy meters and across buses. Following the normalization convention in (21), the noise vector εn is assumed prewhitened, such that all PMU measurements exhibit the same accuracy. The PSSE task now amounts to estimating v given both z and {ζn }n∈P , where P ⊆ B denotes the subset of the PMU-instrumented buses. Hence, the MLE cost in (21) needs to be augmented by the log-likelihood induced by PMU data as v̂ := arg min v M X (zm − hm (v))2 + m=1 X kζn − Φn vk22 . (41) n∈P The SDR methodology is again well motivated to convexify the augmented PSSE problem (41) into X  H minimize 1T χ + Tr(ΦH (42a) n Φn V) − 2<{ζn Φn v} V, v, χ n∈P  s. to V v vH 1   0, and (30c). (42b) By Schur’s complement, the left SDP constraint in (42b) can be expressed equivalently as V  vvH . If the latter constraint is enforced with equality, the matrix V becomes rank-one. Imposing a rank-one constraint in (42) renders it equivalent to the augmented PSSE task of (41). The SDP here also offers the advantages of (32), in terms of the near-optimality and the distributed implementation deferred to Section 4.2. To recover a feasible solution, one can again use the best rank-1 approximation or adopt the randomization technique as elucidated in [84]. Alternatively, the two types of measurements can be jointly utilized upon interpreting the SCADA-based estimate as a prior for PMU-based estimation [58, 32]. Specifically, if v̂s is the SCADA-based estimate, the prior probability density function of the actual state can be postulated to be a circularly symmetric complex Gaussian with mean v̂s and covariance Σ̂s . Given PMU data and the SCADA-based prior, the state can be estimated following a maximum a-posteriori probability (MAP) approach as X v̂ := arg min (v − v̂s )H Σ̂−1 kζn − Φn vk22 (43) s (v − v̂s ) + v n∈P 15 where the first summand is the negative logarithm of the prior distribution, and the second one is the negative log-likelihood from the PMU data. In essence, the approach in (43) treats the SCADA-based estimate as pseudomeasurements relying on the model v̂s = v +  with circularly symmetric zero-mean noise having E[H ] = Σ̂s . 4 Distributed Solvers Upcoming power system requirements call for decentralized solvers. Measurements are now collected at much finer spatio-temporal scales and the number of states increases exponentially as monitoring schemes extend to low-voltage distribution grids [13]. Tightly interconnected power systems call for the close coordination of regional control centers [22], while operators and utilities perform their computational operations on the cloud. This section reviews advances in distributed PSSE solvers. As the name suggests, distributed PSSE solutions spread the computational load across different processors or control centers to speed up time, implement memoryintensive tasks, and/or guarantee privacy. A network of processors may be coordinated by one or more supervising control centers in a hierarchical fashion, or completely autonomously, by exchanging information between processors. To clarify terminology, the latter architecture will be henceforth identified as decentralized. Distributed solvers with a hierarchical structure have been proposed since the statistical formulation of PSSE [61, Part III]. Different versions of this original scheme were later developed in [10], [27], [80], [22], [38]. Decentralized schemes include block Jacobi iterations [43], [9]; an approximate algorithm building on the related optimality conditions [18]; or matrixsplitting techniques for facilitating matrix inversion across areas running Gauss-Newton iterations [53]. Most of the aforementioned approaches presume local identifiability (i.e., each area is identifiable even when shared measurements are excluded) or their convergence is not guaranteed. Assuming a ring topology, every second agent updates its state iteratively through the auxiliary problem principle in [16]. Local observability is waived in the consensus-type solver of [75], where each control center maintains a copy of the entire high-dimensional state vector resulting in slow convergence. For a relatively recent review on distributed PSSE solves, see also [23]. 16 Figure 2: Left: The IEEE 14-bus system partitioned into four areas [64, 38]. Dotted lassos show the buses belonging to area state vectors vk ’s. PMU bus voltage (line current) measurements depicted by green circles (blue squares). Right: The matrix structure for the left system with the distributed SDR solver: green square denotes the overall V, while dark and light blue ones correspond to the four area submatrices {Vk } and their overlaps. 4.1 Distributed Linear Estimators Consider an interconnected system partitioned in K areas supervised by separate control centers. Without loss of generality, an area may be thought of as an independent system operator region, a balancing authority, a power distribution center, or a substation [58]. Area k collects Mk measurements obeying the linear model zk = Hk vk + k (44) where vector vk ∈ CNk collects the system states related to zk through the complex matrix Hk . The random noise vector k is zero-mean with identity covariance upon prewhitening, if measurements are uncorrelated across areas. The model in (44) is exact for PMU measurements, but it may also correspond to a single Gauss-Newton iteration as explained in Section 3.3. Performing PSSE locally at area k amounts to solving minimize fk (vk ) vk ∈Xk (45) where the convex set Xk captures possible prior information, such as zero17 injection buses or short circuits [55], [1]. If fk (vk ) = kzk − Hk vk k22 /2, the minimizer of (45) is the least-squares estimate (LSE) of vk , which is also the MLE of vk for Gaussian k . As illustrated in Fig. 2, the per-area state vectors {vk }K k=1 overlap partially. Although area 4 supervises buses {9, 10, 14}, it also collects the current reading on lines (10, 11). Thus, its state vector v4 extends to bus {11} that is nominally supervised by area 3. To setup notation, define Skl as the shared states for a pair of neighboring areas (k, l). Let also vk [l] (vl [k]) denote the sub-vector of vk (vl ) consisting of their overlapping variables ordered as they appear in v. For example, v3 [4] = v4 [3] contain the bus voltages of {11}. Solving the K problems in (45) separately is apparently suboptimal since the estimates of shared states will disagree, tie-line measurements have to be ignored, and boundary states may thus become unobservable. Coupling the per-area PSSE tasks can be posed as minimize {vk ∈Xk } K X fk (vk ) (46) k=1 s. to vk [l] = vl [k], ∀l ∈ Bk , ∀k where Nk is the set of areas sharing states with area k. The equality constraints of (46) guarantee consensus over the shared variables. Although (46) is amenable to decentralized implementations (cf. [16]), areas need a coordination protocol for their updates. To enable a truly decentralized solution, we follow the seminal approach of [60, 82]. An auxiliary variable vkl is introduced per pair of connected areas (k, l); the symbols vkl and vlk are used interchangeably. The optimization in (46) can then be written as minimize {vk ∈Xk }, {vkl } K X fk (vk ) (47) k=1 s. to vk [l] = vkl , ∀l ∈ Bk , k = 1, . . . , K. Problem (47) can be solved using the alternating direction method of multipliers (ADMM) [4], [20]. In its general form, ADMM tackles convex optimization problems of the form minimize f (x) + g(z) x∈X , z∈Z s. to Ax + Bz = c 18 (48a) (48b) for given matrices and vectors (A, B, c) of proper dimensions. Upon assigning a Lagrange multiplier λ for the coupling constraint in (48a), the (x, z) minimizing (48a) are found through the next iterations for some µ > 0 µ kAx + Bzi − c + λi k22 2 µ := arg min g(z) + kAxi+1 + Bz − c + λi k22 z∈Z 2 := λi + Axi+1 + Bzi+1 − c. xi+1 := arg min f (x) + (49a) zi+1 (49b) x∈X λi+1 (49c) Towards applying the ADMM iterations to (46), identify variables {vk } as x in (48a) and {vkl } as z with g(z) = 0. Moreover, introduce Lagrange multipliers λk,l for each constraint in (47). Observe that λk,l and λl,k correspond to the distinct constraints vk [l] = vkl and vl [k] = vkl , respectively. According to (49a), the per-area state vectors {vk } can be updated separately as µX i vki+1 := arg min fk (vk ) + kvk [l] − vkl + λik,l k22 . (50) vk ∈Xk 2 l∈Bk From (49b) and assuming every state is shared by at most two areas, the auxiliary variables vkl can be readily found in closed form given by i+1 vkl =
1 i+1 vk [l] + vli+1 [k] + λik,l + λil,k 2 (51) while the two related multipliers are updated as i+1 i+1 i λi+1 k,l := λk,l + (vk [l] − vkl ) λi+1 l,k := λil,k + (vli+1 [k] − i+1 ). vkl (52a) (52b) i+1 Adding (52) by parts and combining it with (51) yields λi+1 k,l = −λl,k at all iterations i if the multipliers are initialized at zero. Hence, the auxiliary variable vkl ends up being the average of the shared states; that is, i+1 vkl =
1 i+1 vk [l] + vli+1 [k] . 2 (53) To summarize, at every iteration i: (i) Each control area solves (50). If fk (vk ) is the LS fit and for unconstrained problems, the per-area states are updated as the LSEs using legacy software. The second summand in (50) can be interpreted as pseudo-measurements on the shared states forcing them to consent across areas. 19 −2 10 Error curves averaged over areas LSE D−RPSSE −3 10 −4 10 −5 10 −6 10 0 10 20 30 40 50 60 70 80 Iteration P300 t P t Figure 3: Average error curves 300 k=1 ek,o /300 k=1 ek,c /300 (bottom) and (top) for the LSE, and its robust counterpart D-RPSSE (see Section 5.1) on a 4,200-bus grid. (ii) Neighboring areas exchange their updated shared states. This step involves minimal communication, and no grid models need to be shared. Every area updates its copies of the auxiliary variables vkl using (53). (iii) Every area updates the Lagrange multipliers λkl based on the deviation of the local from the auxiliary variable as in (52). For convex pairs {fk (vk ), Xk }K k=1 , the aforementioned iterates reach the optimal cost in (46), under mild conditions. If the overall power system is observable, the ADMM iterates converge to the unique LSE. The approach has been extended in [33] for joint PSSE and breaker status verification. The decentralized algorithms were tested on a 4,200-bus power grid synthetically built from the IEEE 14- and 300-bus systems. Each of the 300 buses of the latter was assumed to be a different area, and was replaced by a copy of the IEEE 14-bus grid. Additionally, every branch of the IEEE 300-bus grid was an inter-area line whose terminal buses are randomly selected from the two incident to this line areas. Two performance metrics were adopted: the per area error to the centralized solution of (46), denoted (c) by etk,c :=kvk − vkt k2 /Nk , and the per-area error to the true underlying state defined as etk,o :=kvk − vkt k2 /Nk . Figure 3 shows the corresponding error curves averaged over 300 areas. The decentralized LSE approached the underlying state at an accuracy of 10−3 in approximately 10 iterations 20 or 6.2 msec on an Intel Duo Core @ 2.2 GHz (4GB RAM) computer using MATLAB; while the centralized LSE finished in 93.4 msec. 4.2 Distributed SDR-based Estimators Although the SDR-PSSE approach incurs polynomial complexity √ when im4 plemented as a convex SDP, its worst-case complexity is still O(M Nb log(1/)) for a given solution accuracy  > 0 [46]. For typical power networks, the number of measurements M is on the order of the number of buses Nb , and thus the worst-case complexity becomes O(Nb4.5 log(1/)). This complexity could be prohibitive for large-scale power systems, which motivates accelerating the SDR-PSSE method using distributed parallel implementations. Following the area partition in Fig. 2, the m-th measurement per area k can be written as zk,m = hk,m (vk ) + k,m = Tr(Hk,m Vk ) + k,m , ∀k, m where Vk denotes a submatrix of V formed by extracting the rows and columns corresponding to buses in area k; and likewise for each Hk,m . Due to the overlap among the subsets of buses, the outer-product Vk of area k overlaps also with Vl for each neighboring area l ∈ Nk , as shown in Fig. 2. By reducing the measurements atP area k to submatrix Vk , one can de2 k fine the PSSE error cost fk (Vk ) := M m=1 [zk,m − Tr(Hk,m Vk )] per area k, which only involves the local matrix Vk . Hence, the centralized PSSE problem in (32) becomes equivalent to V̂ = arg min K X V0 fk (Vk ). (54) k=1 This equivalent formulation effectively expresses the overall PSSE cost as the superposition of each local cost fk . Nonetheless, even with such a decomposable cost, the main challenge to implement (54) in a distributed manner lies in the PSD constraint that couples the overlapping local matrices {Vk } (cf. Fig. 2). If all submatrices {Vk } were non-overlapping, the cost would be decomposable as in (54), and the PSD of V would boil down to a PSD constraint per area k, as in V̂ = arg min {Vk 0} K X fk (Vk ). (55) k=1 Similar to PSSE for linearized measurements in (46), the formulation in (55) can be decomposed into sub-problems, thanks to the separable PSD 21 constraints. It is not always equivalent to the centralized (54) though, because the PSD property of all submatrices does not necessarily lead to a PSD overall matrix. Nonetheless, the decomposable problem (55) is still a valid SDR-PSSE reformulation, since with the additional per-area constraints rank(Vk ) = 1, it is actually equivalent to (30). While it is totally legitimate to use (55) as the relaxed SDP formulation for (30), the two relaxed problems are actually equivalent under mild conditions. The fresh idea here is to explore valid network topologies to facilitate such PSD constraint decomposition. To this end, it will be instrumental to leverage results on completing partial Hermitian matrices to obtain PSD ones [24]. Upon obtaining the underlying graph formed by the specified entries in the partial Hermitian matrices, these results rely on the so-termed graph chordal property to establish the equivalence between the positive semidefiniteness of the overall matrix and that of all submatrices corresponding to the graph’s maximal cliques. Interestingly, this technique was recently used for developing distributed SDP-based optimal power flow (OPF) solvers in [28, 12, 42]. Construct first a new graph B 0 over B, with all its edges corresponding to the entries in {Vk }. The graph G 0 amounts to having all buses within each subset Nk to form a clique. Furthermore, the following are assumed: (as4) The graph with all the control areas as nodes, and their edges defined by the neighborhood subset {Nk }K k=1 forms a tree. (as5) Each control area has at least one bus that does not overlap with any neighboring area. Proposition 3. Under (as4)-(as5), the two relaxed problems (54) and (55) are equivalent. Proposition 3 can be proved by following the arguments in [84] to show that the entire PSD matrix V can be “completed” using only the PSD submatrices Vk . The key point is that in most power networks even those not obeying (as4) and (as5), (55) can achieve the same accuracy as the centralized one. At the same time, decomposing the PSD constraint in (55) is of paramount importance for developing distributed solvers. One can adopt the consensus reformulation to design the distributed solver for (55) as in (46) of Section 4.1. Accordingly, the ADMM iterations can be employed to solve (55) through iterative information exchanges among neighboring areas, and this is the basis of the distributed SDR-PSSE method. This distributed SDR-PSSE method was tested on the IEEE 118-bus system using the three-area partition in [31]. All three areas measure their 22 1 10 Area 1 Area 2 Area 3 Local matrix error 10 0 Area 1 Area 2 Area 3 10 Local estimation error 0 −1 10 −2 −1 10 10 −3 10 0 20 40 60 0 80 20 40 60 80 Iteration index Iteration index Figure 4: (Left) Per area state matrix error and (Right) state vector estimation error, versus the number of ADMM iterations for the distributed SDR-PSSE solver using the IEEE 118-bus system. local bus voltage magnitudes, as well as real and reactive power flow levels at all lines. The overlaps among the three areas form a tree communication graph used to construct the area-coupling constraints. To demonstrate convergence of the ADMM iterations to the centralized SE solution V̂ of (55), i − V̂ k is plotted versus the the local matrix Frobenius error norm kV̂(k) (k) F iteration index i in the left panel of Fig. 4 for every control area k. Clearly, all local iterates converge to (approximately with a linear rate) their counterparts in the centralized solution. As the task of interest is to estimate the i − v k is also voltages, the local estimation error for the state vector kv̂(k) (k) 2 depicted in the right panel of Fig. 4, where v̂(k) is the estimate of bus volti using the eigen-decomposition ages at area k obtained from the iterate V̂(k) method. Interestingly, the estimation error costs converge within the estimation accuracy of around 10−2 after about 20 iterations (less than 10 iterations for area 1), even though the local matrix has not yet converged. In addition, these error costs decrease much faster in the first couple of iterations. This demonstrates that even with only a limited number of iterations, the PSSE accuracy can be greatly boosted in practice, which in turn makes inter-area communication overhead more affordable. 5 Robust Estimators and Cyber Attacks Bad data, also known as outliers in the statistics parlance, can challenge PSSE due to communication delays, instrument mis-calibration, and/or line 23 parameter uncertainty. In today’s cyber-enabled power systems, smart meter and synchrophasor data could be also purposefully manipulated to mislead system operators. This section reviews conventional and contemporary approaches to coping with outliers. 5.1 Bad Data Detection and Identification Bad data processing in PSSE relies mainly on the linear measurement model z = Hv + , where H ∈ RM ×N . Recall that this model is exact for PMU measurements [cf. (16)–(17)], but approximate per Gauss-Newton iteration or under the linearized grid model. In addition, the aforementioned model assumes real-valued states and measurements, slightly abusing the symbols introduced in (17). This is to keep the notation uncluttered and cover both cases of exact and inexact grid models. Albeit the nominal measurement noise vector is henceforth assumed zero-mean with identity covariance, results extend to colored noise as per (20). To capture bad data, the measurement model is now augmented as z = Hv + o +  (56) where o ∈ RM is an unknown vector whose m-th entry om is deterministically non-zero only if zm is a bad datum [30, 39, 15]. Therefore, vector o is sparse, i.e., many of its entries are zero. Under this outlier-cognizant model in (56), the unconstrained LSE given as v̂LSE = (HT H)−1 HT z, yields the residual error r := z − Hv̂LSE = Pz = P(o + ) (57) with P := IM − H(HT H)−1 HT being the so-called projection matrix onto the orthogonal subspace of range(H). The last equality in (57) stems from the fact that PH = 0. As a projection matrix, P is idempotent, that is, P = P2 ; Hermitian PSD with (M − N ) eigenvalues equal to one and N zero eigenvalues; while its diagonal entries satisfy Pm,m ∈ [0, 1] for m = 1, . . . , M ; see e.g., [6]. For  ∼ N (0, IM ), it apparently holds that P ∼ N (0, P). The meansquared residual error is (see also [39] for its Bayesian counterpart) E[krk22 ] = E[kPk22 ] + kPok22 = (M − N ) + kPok22 . (58) In the absence of bad data, or if o ∈ range(H), the squared residual error follows a χ2 distribution with mean (M − N ). The χ2 -test compares krk22 against a threshold to detect the presence of bad data [1, 55]. 24 Finding both v and o from measurements in (56) may seem impossible, given that the number of unknowns exceeds the number of equations. Leveraging the sparsity of o though, interesting results can be obtained [30]. If τ0 bad data are expected, one would ideally wish to solve o n1 {v̂, ô} ∈ arg min (59) kz − Hv − ok22 : kok0 ≤ τ0 . v, o 2 But the `0 -(pseudo) norm kok0 counting the number of non-zero entries of o, renders (59) NP-hard in general; see also Definition 2 later in Section 5.2. For the special case of τ0 = 1, problem (59) can be efficiently handled. Consider the scenario where the only non-zero entry of ô is the m-th one, and denote the related v̂ minimizer by v̂(m) . Apparently, the m-th entry of the ô minimizer is ôm := zm − hTm v̂(m) . This choice nulls the m-th residual (zm − hTm v̂(m) − ôm = 0). With the m-th residual zeroed, the cost in (59) becomes kr(m) k22 := kz(m) − H(m) v̂(m) k22 , where z(m) is obtained from z upon dropping its m-th entry and H(m) by removing the m-th row of H. The problem in (59) is then equivalent to minimize m 1 kr k2 . 2 (m) 2 (60) Problem (60) can be solved by exhaustively finding all M LSEs excluding one measurement at a time. Fortunately, a classical result from the adaptive filtering literature relates the error kr(m) k22 to the error attained using all outlier-free measurements krk22 := kPzk22 ; see e.g., [25, Ch. 9] krk22 = kr(m) k22 + rm ôm . (61) The same result links the a-posteriori error rm to the a-priori error ôm as rm = Pm,m ôm . Through these links, solving (60) is equivalent to: |rm | rmax := maximize p . m Pm,m (62) In words, a single bad datum can be identified by properly normalizing the entries of the original residual vector r = Pz. Interestingly, the task in (62) coincides with the largest normalized residual (LNR) test that compares rmax to a prescribed threshold to identify a single bad datum [1, Sec. 5.7]. The p threshold is derived after recognizing that in the absence of bad data, rm / Pm,m is standard normal for all m. The LNR test does not generalize for multiple bad data and problem (59) becomes computationally intractable for larger τ0 ’s. Heuristically, if 25 a measurement is deemed as outlying, PSSE is repeated after discarding this bad datum, the LNR test is re-applied, and the process iterates till no corrupted data are identified. Alternatively, the least-median squares and the least-trimmed squares estimators have provable breakdown points and superior efficiency under Gaussian data; see e.g., [52] and references therein. Nevertheless, their complexity scales unfavorably with the network size. Leveraging compressed sensing [7], a practical robust estimator can be found if the `0 -pseudonorm is surrogated by the convex `1 -norm as [30, 31] n1 o minimize kz − Hv − ok22 : kok1 ≤ τ1 (63) v, o 2 for a preselected constant τ1 > 0, or in its Lagrangian form {v̂, ô} ∈ arg min v, o 1 kz − Hv − ok22 + λkok1 2 (64) for some tradeoff parameter λ > 0. The estimates of (64) offer joint state estimation and bad data identification. Even when some measurements are deemed as corrupted, their effect has been already suppressed. The optimization task in (64) can be handled by off-the-shelf software or solvers customized to the compressed sensing setup. When λ → ∞, the minimizer ô becomes zero, and thus v̂ reduces to the LSE. On the contrary, by letting λ → 0+ , the solution v̂ coincides with the least-absolute value (LAV) estimator [40, 6, 17, 68]; presented earlier in (37), namely v̂LAV := arg min kz − Hvk1 . v (65) For finite λ > 0, the v̂ minimizer of (64) is equivalent to Huber’s Mestimator; see [30] and references therein. Based on this connection and for Gaussian , parameter λ can be set to 1.34, which makes the estimator 95% asymptotically efficient for outlier-free measurements [50, p. 26]. Huber’s estimate can be alternatively expressed as the v-minimizer of [49], minimizev, ω 21 kωk22 + λkv − Hv − ωk1 . The bad data identification performance of this minimization has been analyzed in [76]. Table 1 compares several bad data analysis methods on the IEEE 14bus grid of Fig. 2 under the next four scenarios: (S0) no bad data; (S1) bad data on line (4, 7); (S2) bad data on line current (4, 7) and bus voltage 5; and (S3) bad data on bus voltage 5 and line currents (4, 7) and (10, 11). In all scenarios, bad data are simulated by multiplying the real and imaginary parts of the actual measurement by 1.2. The performance metric here is the `2 -norm between the true state and the PSSE, which is 26 Table 1: Mean-Square Estimation Error in the Presence of Bad Data Method GA-LSE LSE LNRT Huber’s 0.0278 0.0313 0.0336 0.0367 0.0278 0.0318 0.1431 0.1434 0.0286 0.0331 0.0404 0.0407 0.0281 0.0322 0.0390 0.0390 (S0) (S1) (S2) (S3) averaged over 1,000 Monte Carlo runs. Four algorithms were tested: (a) an ideal but practically infeasible genie-aided LSE (GA-LSE), which ignores the corrupted measurements; (b) the regular LSE; (c) the LNR test-based (LNRT) estimator with the test threshold set to 3.0 [1]; and (d) Huber’s estimator of (64) with λ = 1.34. For (S0)-(S1), the estimators perform comparably. The few corrupted measurements in (S2)-(S3) can deteriorate LSE’s performance, while Huber’s estimator performs slightly better than LNRT. Computationally, Huber’s estimator was run within 1.3 msec, while the LNRT required 1.5 msec. The computing times were also measured for the IEEE 118-bus grid without corrupted data. Interestingly, the average time on the IEEE 118-bus grid without corrupted data are 3.2 msec and 81 msec, respectively. Towards a robust decentralized state estimator, the ADMM-based framework of Section 4.1 can be engaged here too. If the measurement model for the k-th area is zk = Hk vk + ok + k , the centralized problem boils down to minimize {vk ∈Xk , ok } K X 1 k=1 2 kzk − Hk vk − ok k22 + λkok k1 . (66) To allow for decentralized implementation, the optimization in (66) can be reformulated as minimize K X 1 k=1 2 kzk − Hk vk − ok k22 + λkωk k1 (67a) over {vk ∈ Xk , ok , ωk }, {vkl } (67b) s. to vk [l] = vkl , for all l ∈ Bk , k = 1, . . . , K. (67c) ok = ωk , for all k = 1, . . . , K. (67d) As in Section 4.1, the constraints in (67c) and the auxiliary variables {vkl } enforce consensus of shared states. On the other hand, the variables {ok } 27 are duplicated as {ωk } in (67d). Then, variables {vk , ok } are put together in the x-update of ADMM in (49a), whereas {vkl , ωk } fall into the z-update in (49b). In this fashion, costs are separable over variable groups, and the minimization involving the `1 -norm enjoys a closed-form solution expressed in terms of the soft thresholding operator [31]. 5.2 Observability and Cyber Attacks In the cyber-physical smart grid context, bad data are not simply unintentional errors, but can also take the form of malicious data injections [54]. Amid these challenges, the intertwined issues of critical measurements and stealth cyber-attacks on PSSE are discussed next. It has been tacitly assumed so far that the power system is observable. A power system is observable if distinct states v 6= v0 are mapped to distinct measurements h(v) 6= h(v0 ) under a noiseless setup. Equivalently, if the so-called measurement distance function is defined as [81] D(h) := minimize kh(v) − h(v0 )k0 0 v6=v (68) the power system is observable if and only if D(h) ≥ 1. Given the network topology and the mapping h(v), the well-studied topic of observability analysis aims at determining whether the system state is uniquely identifiable, at least locally in a neighborhood of the current estimate [1, Ch. 4]. If not, mapping observable islands, meaning maximally connected sub-grids with observable internal flows, is important as well. Observability analysis relies on the decoupled linearized grid model, and is accomplished through topological or numerical tests [8, 56]. Apparently, under the linear or linearized model h(v) = Hv, the state v is uniquely identifiable if and only if H is full column-rank. Phase shift ambiguities can be waived by fixing the angle at a reference bus. In the presence of bad data and/or cyber attacks, observability analysis may not suffice. Consider the noiseless measurement model z = h(v) + o, where the non-zero entries of vector o correspond to bad data or compromised meters; and let us proceed with the following definitions. Definition 1 (Observable attack [81]). The attack vector o is deemed as observable if for every state v there is no v0 6= v, such that h(v)+o = h(v0 ). Definition 2 (Identifiable attack [81]). The attack vector o is identifiable if for every v there is no (v0 , o0 ) with v0 6= v and ko0 k0 ≤ kok0 , such that h(v) + o = h(v0 ) + o0 . 28 If the outlier vector o is observable, the operator can tell that the collected measurements do not correspond to a system state, and can hence decide that an attack has been launched. Nevertheless, the attacked meters can be pinpointed only under the stronger conditions of Definition 2. The resilience of the measurement mapping h(v) against attacks can be characterized through D(h) in (68): The maximum number of counterfeited meters for an attack to be observable is Ko = D(h)−1 and to be identifiable, it is Ki = b D(h)−1 c; see [81, 76]. Here, the floor function bxc returns the 2 greatest integer less than or equal to x. Consider the linear mapping h(v) = Hv. Measurement m is termed critical if once removed from the measurement set, it renders the power system non-identifiable. In other words, although H is full column-rank, its submatrix H(m) is not. It trivially follows that D(h) = 1, and the system operator can be arbitrarily misled even if only measurement m is attacked. Due to the typically sparse structure of H, critical measurements or multiple simultaneously corrupted data do exist [1]. It was pointed out in [45] that if an attack o can be constructed to lie in the range(H), it comprises a ‘stealth attack.’ Although finding D(h) is not trivial in general, a polynomial-time algorithm leveraging a graph-theoretic approach is devised in [39]. 6 Power System State Tracking The PSSE methods reviewed so far ignore system dynamics and do not exploit historical information. Dynamic PSSE is well motivated thanks to its improved robustness, observability, and predictive ability when additional temporal information is available [26]. Recently proposed model-free and model-based state tracking schemes are outlined next. 6.1 Model-free State Tracking via Online Learning In complex future power systems, one may not choose to explicitly commit to a model for the underlying system dynamics. The framework of online convex optimization (OCO), particularly popular in machine learning, can account for unmodeled dynamics and is thus briefly presented next [62]. The OCO model considers a multi-stage game between a player and an adversary. In the PSSE context, the utility or the system operator assumes the role of the player, while the loads and renewable generations can be viewed as the adversary. At time t, the player first selects an action Vt from a given action set V, and the adversary subsequently reveals a convex loss 29 function ft : V → R. In this round, the player suffers a loss ft (Vt ). The ultimate goal for the player is to minimize the regret Rf (T ) over T rounds: Rf (T ) := T X ft (Vt ) − minimize V∈V t=1 T X ft (V). (69) t=1 The regret is basically the accumulated cost incurred the player relative Pby T 0 to that by a single fixed action V := arg minV∈V t=1 ft (V). This fixed action is selected with the advantage of knowing the loss functions {ft }Tt=1 in hindsight. Under appropriate conditions, judiciously designed online optimization algorithms can achieve sublinear regret; that is, Rf (T )/T → 0 as T → +∞. Building on the SDR-PSSE formulation of Section 3.4, the ensuing method considers streaming data for real-time PSSE. The data referring to and colt lected over the control period t are {(zmt ; Hmt )}M mt =1 with t = 1, . . . , T . The number and type of measurements can change over time, while the matrix corresponding to measurement m may change over time as indicated by t {Hmt }M mt =1 due to topology reconfigurations. The online PSSE task can be now formulated as T X minimize ft (V) (70) V0 t=1 PMt where ft (V) := mt =1 [zmt − Tr(Hmt V)]2 . Online PSSE aims at improving the static estimates by capitalizing on previous measurements as well as tracking slow time-varying variations in generation and demand. Minimizing the cost in (70) may be computationally cumbersome for real-time implementation. An efficient alternative based on online gradient descent amounts to iteratively minimizing a regularized first-order approximation of the instantaneous cost instead [37] Vt+1 := arg min Tr(VH ∇ft (Vt )) + V0 1 kV − Vt k2F 2µt (71) for t = 1, . . ., and suitably selected step sizes µt > 0. Interestingly, the optimization in (71) admits a closed-form solution given by Vt+1 = ProjS+ [Vt − µt ∇ft (Vt )] (72) with ProjS+ denoting the projection onto the positive semidefinite cone, which can be performed using eigen-decomposition followed by setting negative eigenvalues to zero. It is worth mentioning that the online PSSE in 30 (72) enjoys sublinear regret [37]. Upon finding Vt , a state estimate vt can be obtained by eigen-decomposition or randomization as in Section 3.4. With an additional nuclear-norm regularization term promoting low-rank solutions in (71), online ADMM alternatives were devised in [36]. Interestingly, online learning tools has recently been advocated for numerous real-time energy management tasks in [35], [34], [69]. 6.2 Model-based State Tracking Although the previous model-free solver can recover slow time-varying states, model-based approaches facilitate tracking of fast time-varying system states. A typical state-space model for power system dynamics is [65] vt+1 = Ft vt + gt + ωt (73a) zt = h(vt ) + t (73b) where Ft denotes the state-transition matrix, gt captures the process mismatch, and wt is the additive noise. The nonlinear mapping h(·) comes from conventional SCADA measurements. Values {(Ft , gt )} can be obtained in real-time using for example Holt’s system identification method [11]. Two common dynamic tracking approaches to cope with the nonlinearity in the measurement model of (73b) include the (extended or unscented) Kalman filters and moving horizon estimators [65, 14, 26, 73, 72], and they are outlined in order next. The extended Kalman filter (EKF) handles the nonlinearity by linearizing h(v) around the state predictor. To start, let v̂t+1|t stand for the predicted estimate at time t + 1 given measurements {zτ }tτ =1 up to time t. Let also v̂t+1|t+1 be the filtered estimate given measurements {zτ }t+1 τ =1 . If the noise terms ωt and t in (73) are assumed zero-mean Gaussian with known covariance matrices Qt  0 and Rt  0, respectively, the EKF can be implemented with the following recursions   v̂t+1|t+1 = v̂t+1|t + Kt+1 zt+1 − h(v̂t+1|t ) (74) where the state predictor v̂t+1|t and the Kalman gain Kt+1 are given by v̂t+1|t = Ft v̂t|t + gt (75a) H Kt+1 = Pt+1|t JH t+1 Jt+1 Pt+1|t Jt+1 + Rt+1 Pt+1|t+1 = Pt+1|t − Kt+1 Jt+1 Pt+1|t Pt+1|t = Ft Pt|t FH t + Qt 31
−1 (75b) (75c) (75d) with Jt+1 being the measurement Jacobian matrix of h evaluated at v̂t+1|t , and Pt+1|t+1  0 (Pt+1|t  0) denoting the corrected (predicted) state estimation error covariance matrix at time t + 1. To improve on the approximation accuracy of the EKF, extended Kalman filters (UKF) have been reported in [65]; see also [79] for their robust versions. Particle filtering may also be useful if its computational complexity can be supported during real-time power systems operations. Because the EKF and UKF are known to diverge for highly nonlinear dynamics, moving horizon estimation (MHE) has been suggested as an accurate yet tractable alternative with proven robustness to bounded model errors [59]. Different from Kalman filtering, the initial state v0 , and noises ωt and t in MHE are viewed as deterministic unknowns taking values from given bounded sets S, W, and E, respectively. The sets W and E model disturbances with truncated densities [59]. The idea behind MHE is to perform PSSE by exploiting useful information present in a sliding window of the most recent observations. Consider here a sliding window of length L + 1. Let v̂t−L|t denote the smoothed estimate at time t − L given L past measurements, as well as the current one, namely {zτ }tτ =t−L . MHE aims at obtaining the most recent L t state estimates {v̂t−L+s|t }L s=0 based on {zτ }t−L and the available estimate v̌t−L := v̂t−L|t−1 from time t − 1 and for t ≥ L. A key simplification is that once v̂t−L|t becomes available, the other L recent estimates at time t can be recursively obtained through ‘noise-free’ propagation based on the dynamic model (73a); that is, v̂t−L+s|t = Ft−L+s−1 v̂t−L+s−1|t (76) for s = 1, . . . , L. By relating all recent estimates to v̂t−L|t via successive multiplications of transition matrices, the update in (76) simplifies to v̂t−L+s|t = Tt−L+s v̂t−L|t (77) where Tt−L+s := Ft−L+s−1 Tt−L+s−1 for s = 1, . . . , L, with Tt−L = I. The MHE-based state estimate v̂t−L|t is then given by v̂t−L|t := arg min v L X zt−L+s − h(Tt−L+s v) 2 2 + λkv − v̌t−L k22 (78) s=0 where λ > 0 can be tuned relying on our confidence in the state predictor v̌t−L , and the measurements {zτ }tt−L . Given the quadratic dependence of 32 the SCADA measurements {h(vt )} and the state v, the optimization problem in (78) is non-convex. Finding the MHE-based state estimates in real time entails online solutions of dynamic optimization problems. The MHE formulation can be convexified by exploiting the semidefinite relaxation: vector v is lifted to the matrix V := vvH  0, and the m-th entry of h(vt−L+s ) for s = 0, . . . , L, is expressed as H hm (Tt−L+s v) = vH TH t−L+s Hm Tt−L+s v = Tr(Tt−L+s Hm Tt−L+s V). Upon dropping the nonconvex rank constraint rank(V) = 1, the SDP-based MHE yields V̂t−L|t := arg min V0 L X zt−L+s −Tr TH t−L+s Hm Tt−L+s V
2 +λkv− v̌t−L k22 2 s=0 which can be solved in polynomial time using off-the-shelf toolboxes. Rankone state estimates can be obtained again through eigen-decomposition or randomization. The complexity of solving the last problem is rather high in its present form on the order of Nb4.5 [46]. Therefore, developing faster solvers for the SDP-based MHE by exploiting the rich sparsity structure in {Hm } matrices is worth investigating. Decentralized and localized MHE implementations are also timely and pertinent. Devising FPP-based solvers for the MHE in (78) constitutes another research direction. 7 Discussion This chapter has reviewed some of the recent advances in PSSE. After developing the CRLB, an SDP-based solver, and its regularized counterpart were discussed. To overcome the high complexity involved, a scheme named feasible point pursuit relying on successive convex approximations was also advocated. A decentralized PSSE paradigm put forth provides the means for coping with the computationally-intensive SDP formulations, it is tailored for the interconnected nature of modern grids, while it can also afford processing PMU data in a timely fashion. A better understanding of cyber attacks and disciplined ways for decentralized bad data processing were also provided. Finally, this chapter gave a fresh perspective to state tracking under model-free and model-based scenarios. Nonetheless, there are still many technically challenging and practically pertinent grid monitoring issues to be addressed. Solving power grid data 33 processing tasks on the cloud has been a major trend to alleviate data storage, communication, and interoperability costs for system operators and utilities. Moreover, with the current focus on low- and medium-voltage distribution grids, solvers for unbalanced and multi-phase operating conditions [2] are desirable. Smart meters and synchrophasor data from distribution grids (also known as micro-PMUs [67]) call for new data processing solutions. Advances in machine learning and statistical signal processing, such as sparse and low-rank models, missing and incomplete data, tensor decompositions, deep learning, nonconvex and stochastic optimization tools, and (multi)kernel-based learning to name a few, are currently providing novel paths to grid monitoring tasks while realizing the vision of smarter energy systems. Acknowledgements G. Wang and G. B. Giannakis were supported in part by NSF grants 1423316, 1442686, 1508993, and 1509040; and by the Laboratory Directed Research and Development Program at the NREL. H. Zhu was supported in part by NSF grants 1610732 and 1653706. Appendix 2 })). Proof of Prop. 1. Consider the AGWN model (15) with  ∼ N (0, dg({σm The data likelihood function is   M Y 1 (zm − vH Hm v)2 p p(z; v) = (79) exp − 2 2 2σm 2πσm m=1 and the negative log-likelihood function denoted by f (v) = − ln p(z; x) is f (v) =  M  X
2 1
1 H 2 z − v H v + ln 2πσ . m m m 2 2σm 2 (80) m=1 The Fisher information matrix (FIM) is defined as the Hessian of the real-valued function f (v) with respect to the complex vector v ∈ CNb . Deriving the CRLB amounts to finding the Hessian of a real-valued function with respect to a complex-valued vector. Wirtinger’s calculus confirms that f (v) can be equivalently rewritten as f (v, v∗ ); see e.g., [41]. Upon introducing the conjugate coordinates [vT (v∗ )T ]T ∈ C2Nb , the Wirtinger derivatives, namely the first-order partial differential operators of functions over 34 complex domains, are given by [41] ∂f ∂f (v, v∗ ) := ∂v ∂vT ∂f ∂f (v, v∗ ) := ∂v∗ ∂(v∗ )T  ∂f ∂f ··· = ∂V1 ∂VN constant v∗ v∗   ∂f ∂f . = ··· ∗ ∂V1∗ ∂VN v constant v  constant constant These definitions follow the convention in multivariate calculus that derivatives are denoted by row vectors, and gradients by column vectors. Define for notational brevity φm (v, v∗ ) := zm − (v∗ )T Hm v for m = 1, . . . , M . Accordingly, the Wirtinger derivatives of f (v, v∗ ) in (80) are obtained as M L X X ∂f 1 ∂φm ∂f 1 ∂φm = φ and = φ 2 m ∂vT 2 m ∂(v∗ )T ∂v σm ∂v∗ σm m=1 (82) m=1 and the Wirtinger derivatives of φm (v, v∗ ) can be found likewise ∂φm ∂φm = −(Hm v)H and = −(H∗m v∗ )H . ∂vT ∂(v∗ )T (83) In the conjugate coordinate system, the complex Hessian of f (v, v∗ ) with respect to the conjugate coordinates [vT (v∗ )T ]T is defined as   Hvv Hv∗ v ∗ 2 ∗ H(v, v ) := ∇ f (v, v ) = (84) Hvv∗ Hv∗ v∗ whose blocks are given as  Hvv ∂ := ∂vT  Hvv∗ ∂ := ∂vT ∂f ∂v H ∂f ∂v∗ ,  Hv∗ v ∂f ∂v  Hv ∗ v ∗ ∂ := ∂(v∗ )T ∂f ∂v∗ H , H ∂ := ∂(v∗ )T H . Substituting (82) and (83) into the last equations and after algebraic manipulations yields Hvv = Hv ∗ v = M X m=1 M X   −2 σm Hm v(Hm v)H − φm Hm (85a) −2 σm Hm v(H∗m v∗ )H (85b) m=1 35 Hvv∗ = Hv ∗ v ∗ = M X m=1 M X −2 ∗ ∗ σm Hm v (Hm v)H (85c)   −2 σm H∗m v∗ (H∗m v∗ )H − φm H∗m . (85d) m=1 Evaluating the Hessian blocks of (85) at the true value of v, and taking the expectation with respect to , yields E[φm ] = 0. Hence, the φm -related terms in (85) disappear, and the FIM F(v, v∗ ) := E[H(v, v∗ )] simplifies to the expression in (23); see also [66]. To show that the FIM is rank-deficient, define gm := [(Hm v)H (H∗m v∗ )H ]H , PM −2 g gH . Observe now that the nonso that the FIM becomes F = m=1 σm m m T ∗ T T zero vector d(v) := [v − (v ) ] is orthogonal to gm for m = 1, . . . , M ; that is, H gm d = vH Hm v − (vH Hm v)∗ = 0. Based on the latter, it is not hard to verify that Fd = 0, which proves that the null space of F is non-empty. References [1] A. Abur and A. Gómez-Expósito, Power System State Estimation: Theory and Implementation. New York, NY: Marcel Dekker, 2004. [2] M. Bazrafshan and N. Gatsis, “Comprehensive modeling of three-phase distribution systems via the bus admittance matrix,” arXiv:1705.06782, 2017. [3] D. P. Bertsekas, Nonlinear Programming, 2nd ed. Athena Scientific, 1999. Belmont, MA: [4] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed optimization and statistical learning via the alternating direction method of multipliers,” Found. Trends Mach. Learn., vol. 3, pp. 1–122, 2010. [5] S. Boyd and L. Vandenberghe, Convex Optimization. New York, NY: Cambridge University Press, 2004. [6] M. K. Celik and A. Abur, “A robust WLAV state estimator using transformations,” IEEE Trans. Power Syst., vol. 7, no. 1, pp. 106–113, Feb. 1992. 36 [7] S. S. Chen, D. L. Donoho, Michael, and A. Saunders, “Atomic decomposition by basis pursuit,” SIAM J. Sci. Comput., vol. 20, pp. 33–61, Jul. 1998. [8] K. A. Clements, G. R. Krumpholz, and P. W. Davis, “Power system state estimation with measurement deficiency: An observability/measurement placement algorithm,” IEEE Trans. Power App. Syst., vol. 102, no. 7, pp. 2012–2020, Jul. 1983. [9] A. J. Conejo, S. de la Torre, and M. Canas, “An optimization approach to multiarea state estimation,” IEEE Trans. Power Syst., vol. 22, no. 1, pp. 213–221, Feb. 2007. [10] T. V. Cutsem and M. Ribbens-Pavella, “Critical survey of hierarchical methods for state estimation of electric power systems,” IEEE Trans. Power App. Syst., vol. 102, no. 10, pp. 3415–3424, Oct. 1983. [11] A. L. Da Silva, M. Do Coutto Filho, and J. De Queiroz, “State forecasting in electric power systems,” IEE Proc. Generat. Transm. Distrib., vol. 130, no. 5, pp. 237–244, Sep. 1983. [12] E. Dall’Anese, H. Zhu, and G. B. Giannakis, “Distributed optimal power flow for smart microgrids,” IEEE Trans. Smart Grid, vol. 4, no. 3, pp. 1464–1475, Sep. 2013. [13] J. De La Ree, V. A. Centeno, J. Thorp, and A. Phadke, “Synchronized phasor measurement applications in power systems,” IEEE Trans. Smart Grid, vol. 1, no. 1, pp. 20–27, Jun. 2010. [14] A. S. Debs and R. E. Larson, “A dynamic estimator for tracking the state of a power system,” IEEE Trans. Power App. Syst., no. 7, pp. 1670–1678, Sep. 1970. [15] D. Duan, L. Yang, and L. L. Scharf, “Phasor state estimation from PMU measurements with bad data,” in Proc. IEEE Workshop on Comp. Adv. in Multi-Sensor Adaptive Proc., San Juan, Puerto Rico, Dec. 2011. [16] R. Ebrahimian and R. Baldick, “State estimation distributed processing,” IEEE Trans. Power Syst., vol. 15, no. 4, pp. 1240–1246, Nov. 2000. 37 [17] A. A. El-Keib, J. Nieplocha, H. Singh, and D. Maratukulam, “A decomposed state estimation technique suitable for parallel processor implementation,” IEEE Trans. Power Syst., vol. 7, no. 3, pp. 1088–1097, Aug. 1992. [18] D. M. Falcao, F. F. Wu, and L. Murphy, “Parallel and distributed state estimation,” IEEE Trans. Power Syst., vol. 10, no. 2, pp. 724–730, May 1995. [19] G. B. Giannakis, V. Kekatos, N. Gatsis, S.-J. Kim, H. Zhu, and B. Wollenberg, “Monitoring and optimization for power grids: A signal processing perspective,” IEEE Signal Process. Mag., vol. 30, no. 5, pp. 107–128, Sept. 2013. [20] G. B. Giannakis, Q. Ling, G. Mateos, I. D. Schizas, and H. Zhu, “Decentralized learning for wireless communications and networking,” Splitting Methods in Communication, Imaging, Science, and Engineering, pp. 461–497, Springer, 2016. [21] M. X. Goemans and D. P. Williamson, “Improved approximation algorithms for maximum cut and satisfiability problems using semidefinite programming,” J. ACM, vol. 42, no. 6, pp. 1115–1145, Nov. 1995. [22] A. Gómez-Expósito, A. Abur, A. de la Villa Jaén, and C. Gómez-Quiles, “A multilevel state estimation paradigm for smart grids,” Proc. IEEE, vol. 99, no. 6, pp. 952–976, Jun. 2011. [23] A. Gómez-Expósito, A. de la Villa Jaén, C. Gómez-Quiles, P. Rousseaux, and T. V. Cutsem, “A taxonomy of multi-area state estimation methods,” Electr. Pow. Syst. Res., vol. 81, pp. 1060–1069, Apr. 2011. [24] R. Grone, C. R. Johnson, E. M. Sá, and H. Wolkowicz, “Positive definite completions of partial hermitian matrices,” Linear Algebra & its Appl., vol. 58, pp. 109–124, Apr. 1984. [25] S. Haykin, Adaptive Filter Theory. New York, NY: Prentice Hall, 2002. [26] S.-J. Huang and K.-R. Shih, “Dynamic-state-estimation scheme including nonlinear measurement function considerations,” IEE Proc. Generat. Transm. Distrib., vol. 149, no. 6, pp. 673–678, Nov. 2002. 38 [27] S. Iwamoto, M. Kusano, and V. H. Quintana, “Hierarchical state estimation using a fast rectangular-coordinate method,” IEEE Trans. Power Syst., vol. 4, no. 3, pp. 870–880, Aug. 1989. [28] R. A. Jabr, “Exploiting sparsity in SDP relaxations of the OPF problem,” IEEE Trans. Power Syst., vol. 27, no. 2, pp. 1138–1139, May 2012. [29] S. M. Kay, Fundamentals of Statistical Signal Processing, Vol. I: Estimation Theory. Prentice Hall, 1993. [30] V. Kekatos and G. B. Giannakis, “From sparse signals to sparse residuals for robust sensing,” IEEE Trans. Signal Process., vol. 59, no. 7, pp. 3355–3368, Jul. 2011. [31] ——, “Distributed robust power system state estimation,” IEEE Trans. Power Syst., vol. 28, no. 2, pp. 1617–1626, May 2013. [32] V. Kekatos, G. B. Giannakis, and B. F. Wollenberg, “Optimal placement of phasor measurement units via convex relaxation,” IEEE Trans. Power Syst., vol. 27, no. 3, pp. 1521–1530, Aug. 2012. [33] V. Kekatos, E. Vlachos, D. Ampeliotis, G. B. Giannakis, and K. Berberidis, “A decentralized approach to generalized power system state estimation,” in Proc. IEEE Wkshp. on Comp. Adv. in MultiSensor Adaptive Proc., Saint Martin, Dec. 2013. [34] V. Kekatos, G. Wang, A. J. Conejo, and G. B. Giannakis, “Stochastic reactive power management in microgrids with renewables,” IEEE Trans. Power Syst., vol. 30, no. 6, pp. 3386–3395, Nov. 2015. [35] S. J. Kim, G. B. Giannakis, and K. Y. Lee, “Online optimal power flow with renewables,” in Asilomar Conf. on Signals, Syst. and Comput., Pacific Grove, CA, Nov 2014, pp. 355–360. [36] S.-J. Kim, “Online power system state estimation using alternating direction method of multipliers,” in Proc. IEEE PES Summer Meeting, Denver, CO, 2015. [37] S.-J. Kim, G. Wang, and G. B. Giannakis, “Online semidefinite programming for power system state estimation,” in Proc. IEEE Intl. Conf. on Acoustics, Speech, and Signal Process., Florence, Italy, May 2014, pp. 6024–6027. 39 [38] G. N. Korres, “A distributed multiarea state estimation,” IEEE Trans. Power Syst., vol. 26, no. 1, pp. 73–84, Feb. 2011. [39] O. Kosut, L. Jia, J. Thomas, and L. Tong, “Malicious data attacks on the smart grid,” IEEE Trans. Smart Grid, vol. 2, no. 4, pp. 645–658, Dec. 2011. [40] W. W. Kotiuga and M. Vidyasagar, “Bad data rejection properties of weighted least-absolute-value techniques applied to static state estimation,” IEEE Trans. Power App. Syst., no. 4, pp. 844–853, Apr. 1982. [41] K. Kreutz-Delgado, “The complex gradient operator and the CRcalculus,” arXiv:0906.4835, 2009. [42] A. Y. Lam, B. Zhang, and T. David, “Distributed algorithms for optimal power flow problem,” in Proc. IEEE Conf. on Decision and Control, Maui, HI, Dec. 2012, pp. 430–437. [43] S.-Y. Lin and C.-H. Lin, “An implementable distributed state estimator and distributed bad data processing schemes for electric power systems,” IEEE Trans. Power Syst., vol. 9, no. 3, pp. 1277–1284, Aug. 1994. [44] T. Lipp and S. Boyd, “Variations and extension of the convex–concave procedure,” Optim. Eng., vol. 17, no. 2, pp. 263–287, Jun. 2016. [45] Y. Liu, P. Ning, and M. K. Reiter, “False data injection attacks against state estimation in electric power grids,” ACM Trans. Info. and System Security, vol. 14, no. 1, pp. 13:1–13:33, May 2011. [46] Z.-Q. Luo, W.-K. Ma, A. M.-C. So, Y. Ye, and S. Zhang, “Semidefinite relaxation of quadratic optimization problems,” IEEE Signal Process. Mag., vol. 27, no. 3, pp. 20–34, May 2010. [47] R. Madani, A. Ashraphijuo, J. Lavaei, and R. Baldick, “Power system state estimation with a limited number of measurements,” in Proc. IEEE Conf. on Decision and Control, Las Vegas, NV, Dec. 2016. [48] R. Madani, J. Lavaei, and R. Baldick, “Convexification of power flow problem over arbitrary networks,” in Proc. IEEE Conf. on Decision and Control, Osaka, Japan, Dec. 2015. [49] O. L. Mangasarian and D. R. Musicant, “Robust linear and support vector regression,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 22, no. 9, pp. 950–955, Sep. 2000. 40 [50] R. A. Maronna, R. D. Martin, and V. J. Yohai, Robust Statistics: Theory and Methods. Wiley, 2006. [51] O. Mehanna, K. Huang, B. Gopalakrishnan, A. Konar, and N. Sidiropoulos, “Feasible point pursuit and successive approximation of non-convex QCQPs,” vol. 22, no. 7, pp. 804–808, Nov. 2015. [52] L. Mili, M. G. Cheniae, and P. J. Rousseeuw, “Robust state estimation of electric power systems,” IEEE Trans. Circuits Syst. I, vol. 41, no. 5, pp. 349–358, May 1994. [53] A. Minot, Y. M. Lu, and N. Li, “A distributed Gauss-Newton method for power system state estimation,” IEEE Trans. Power Syst., vol. 31, no. 5, pp. 3804–3815, Sep. 2016. [54] B. Y. Mo, T. H.-J. Kim, K. Brancik, D. Dickinson, H. Lee, A. Perrig, and B. Sinopoli, “Cyber-physical security of a smart grid infrastructure,” Proc. IEEE, vol. 100, no. 1, pp. 195–209, Jan. 2012. [55] A. Monticelli, “Electric power system state estimation,” Proc. IEEE, vol. 88, no. 2, pp. 262–282, Feb. 2000. [56] A. Monticelli and F. F. Wu, “Network observability: Theory,” IEEE Trans. Power App. Syst., vol. 104, no. 5, pp. 1042–1048, May 1985. [57] P. M. Pardalos and S. A. Vavasis, “Quadratic programming with one negative eigenvalue is NP-hard,” J. Global Optim., vol. 1, no. 1, pp. 15–22, Mar. 1991. [58] A. G. Phadke and J. S. Thorp, Synchronized Phasor Measurements and Their Applications. New York, NY: Springer, 2008. [59] C. V. Rao, J. B. Rawlings, and D. Q. Mayne, “Constrained state estimation for nonlinear discrete-time systems: Stability and moving horizon approximations,” IEEE Trans. Autom. Contr., vol. 48, no. 2, pp. 246–258, Feb. 2003. [60] I. D. Schizas, A. Ribeiro, and G. B. Giannakis, “Consensus in ad hoc WSNs with noisy links: Part I: Distributed estimation of deterministic signals,” IEEE Trans. Signal Process., vol. 56, no. 1, pp. 350–364, Jan. 2008. [61] F. C. Schweppe, J. Wildes, and D. Rom, “Power system static state estimation: Parts I, II, and III,” IEEE Trans. Power App. Syst., vol. 89, pp. 120–135, Jan. 1970. 41 [62] S. Shalev-Shwartz, “Online learning and online convex optimization,” Found. and Trends Mach. Learning, vol. 4, no. 2, pp. 107–194, Mar. 2011. [63] P. Stoica and T. L. Marzetta, “Parameter estimation problems with singular information matrices,” IEEE Trans. Signal Process., vol. 49, no. 1, pp. 87–90, Jan. 2001. [64] Power systems test case archive. Univ. of Washington. [Online]. Available: http://www.ee.washington.edu/research/pstca/ [65] G. Valverde and V. Terzija, “Unscented Kalman filter for power system dynamic state estimation,” IEE Proc. Generat. Transm. Distrib., vol. 5, no. 1, pp. 29–37, Jan. 2011. [66] A. Van den Bos, “A Cramér-Rao lower bound for complex parameters,” IEEE Trans. Signal Process., vol. 40, no. 10, Oct. 1994. [67] A. Von Meier, D. Culler, A. McEachern, and R. Arghandeh, “Microsynchrophasors for distribution systems,” in Proc. IEEE Conf. on Innovative Smart Grid Technologies, 2014, pp. 1–5. [68] G. Wang, G. B. Giannakis, and J. Chen, “Robust and scalable power system state estimation via composite optimization,” preprint, 2017. [69] G. Wang, V. Kekatos, A. J. Conejo, and G. B. Giannakis, “Ergodic energy management leveraging resource variability in distribution grids,” IEEE Trans. Power Syst., vol. 31, no. 6, pp. 4765–4775, Nov. 2016. [70] G. Wang, A. S. Zamzam, G. B. Giannakis, and N. D. Sidiropoulos, “Power system state estimation via feasible point pursuit,” in Proc. IEEE Global Conf. on Signal and Inf. Process., Washington, D.C., Nov. 2016. [71] ——, “Power system state estimation via feasible point pursuit: Algorithms and Crmér-Rao bound,” arXiv:1203.0853, 2017. [72] G. Wang, S.-J. Kim, and G. B. Giannakis, “Moving-horizon dynamic power system state estimation using semidefinite relaxation,” in Proc. IEEE PES General Meeting, Washington, DC, Jul. 2014, pp. 1–5. [73] S. Wang, W. Gao, and A. S. Meliopoulos, “An alternative method for power system dynamic state estimation based on unscented transform,” IEEE Trans. Power Syst., vol. 27, no. 2, pp. 942–950, May 2012. 42 [74] A. J. Wood and B. F. Wollenberg, Power Generation, Operation, and Control, 2nd ed. New York, NY: Wiley & Sons, 1996. [75] L. Xie, D.-H. Choi, and S. Kar, “Cooperative distributed state estimation: Local observability relaxed,” in Proc. IEEE Power & Energy Society General Meeting, Detroit, MI, Jul. 2011. [76] W. Xu, M. Wang, J. F. Cai, and A. Tang, “Sparse error correction from nonlinear measurements with applications in bad data detection for power networks,” IEEE Trans. Signal Process., vol. 61, no. 24, pp. 6175–6187, Dec. 2013. [77] A. S. Zamzam, N. D. Sidiropoulos, and E. Dall’Anese, “Beyond relaxation and Newton-Raphson: Solving AC OPF for multi-phase systems with renewables,” IEEE Trans. Smart Grid, 2017. [78] Y. Zhang, R. Madani, and J. Lavaei, “Conic relaxations for power system state estimation with line measurements,” IEEE Trans. Control Netw. Syst., 2017 (to appear). [79] J. Zhao, M. Netto, and L. Mili, “A robust iterated extended Kalman filter for power system dynamic state estimation,” IEEE Trans. Power Syst., 2016 (to appear). [80] L. Zhao and A. Abur, “Multiarea state estimation using synchronized phasor measurements,” IEEE Trans. Power Syst., vol. 20, no. 2, pp. 611–617, May 2005. [81] H. Zhu and G. B. Giannakis, “Robust power system state estimation for the nonlinear AC flow model,” in Proc. North American Power Symposium, Urbana-Champaign, IL, Sep. 2012. [82] H. Zhu, G. B. Giannakis, and A. Cano, “Distributed in-network decoding,” IEEE Trans. Signal Process., vol. 57, no. 10, pp. 3970–3983, Oct. 2009. [83] H. Zhu and G. B. Giannakis, “Estimating the state of AC power systems using semidefinite programming,” in Proc. North American Power Symposium, Boston, MA, Aug. 2011. [84] ——, “Power system nonlinear state estimation using distributed semidefinite programming,” IEEE J. Sel. Topics Signal Process., vol. 8, no. 6, pp. 1039–1050, Dec. 2014. 43 

3cs.SY

arXiv:1604.06525v3 [cs.GR] 9 Sep 2017 Opt: A Domain Specific Language for Non-linear Least Squares Optimization in Graphics and Imaging ZACHARY DEVITO, Facebook Research MICHAEL MARA, Stanford University MICHAEL ZOLLHÖFER, Max-Planck-Institute for Informatics GILBERT BERNSTEIN, Stanford University JONATHAN RAGAN-KELLEY, UC Berkeley CHRISTIAN THEOBALT, Max-Planck-Institute for Informatics PAT HANRAHAN, Stanford University MATTHEW FISHER, Adobe Research MATTHIAS NIESSNER, Technical University of Munich Many graphics and vision problems can be expressed as non-linear least squares optimizations of objective functions over visual data, such as images and meshes. The mathematical descriptions of these functions are extremely concise, but their implementation in real code is tedious, especially when optimized for real-time performance on modern GPUs in interactive applications. In this work, we propose a new language, Opt1 , for writing these objective functions over image- or graph-structured unknowns concisely and at a high level. Our compiler automatically transforms these specifications into state-of-the-art GPU solvers based on Gauss-Newton or LevenbergMarquardt methods. Opt can generate different variations of the solver, so users can easily explore tradeoffs in numerical precision, matrix-free methods, and solver approaches. In our results, we implement a variety of real-world graphics and vision applications. Their energy functions are expressible in tens of lines of code, and produce highly-optimized GPU solver implementations. These solvers are competitive in performance with the best published hand-tuned, application-specific GPU solvers, and orders of magnitude beyond a generalpurpose auto-generated solver. CCS Concepts: • Software and its engineering → Domain specific languages; • Computing methodologies → Image processing; Graphics systems and interfaces; Graphics processors; Procedural animation; General Terms: Domain-specific Languages, Non-linear least squares, Levenberg-Marquardt, Gauss-Newton 1 INTRODUCTION Many problems in graphics and vision can be concisely formulated as least squares optimizations on images, meshes, or graphs. For example, Poisson image editing, shape-from-shading, and as-rigidas-possible warping have all been formulated as non-linear least 1Opt is open sourced and publicly available under http://optlang.org. This is the author’s version of the work. It is posted here for your personal use. Not for redistribution. squares optimizations, allowing them to be described tersely as energy functions over pixels or vertices [47, 50, 68]. In many of these applications, high performance is critical for interactive feedback, requiring efficient parallel or GPU-based solvers [11, 30, 54, 68, 73]. These solvers require optimizations that are not expressible in generic sparse linear algebra software. For example, these solvers are matrix-free; that is, they compute matrix values on-the-fly rather than loading data from matrices materialized (i.e., stored) in memory. They also implicitly represent the sparse connectivity of the matrices based on the structure of images or graphs, rather than store it explicitly in memory. This approach can also be applied to solvers that use the Levenberg-Marquardt algorithm [36, 39]. However, the efficiency of these solvers comes at enormous implementation cost: the simple energy function must be manually transformed into a complex product of partial derivative matrices (e.g., JT F and JT Jp). Furthermore, the code tightly intertwines the calculation of partial derivatives with operations performed by the solver. Finally, this code has to be written by hand in GPU kernels; the result is hundreds of lines of highly-tuned CUDA code which is hard to maintain and modify. This paper presents a new language, Opt, which makes this type of high performance-optimization accessible to a wider community of graphics and vision practitioners. Programmers write high-level sum-of-squares energy functions over pixels or graphs, such as the example shown in Fig. 1 (left) for as-rigid-as-possible image warping. Our compiler can transform these energies into efficient GPU routines which compute products of the derivatives (e.g., JT J or JT Jp). We provide a suite of solvers that use these routines to apply either Gauss-Newton (GN) or Levenberg-Marquardt (LM) methods. The solvers use a parallel preconditioned conjugate gradient inner loop based on either matrix-free or materialized approaches. The resulting code can be used within both matrix-free and materialized solvers. Our system is able to achieve this due to four key ideas. First, we provide an optimization framework that separates the details of a particular energy from the details of the GN or LM solver approach. The framework is general enough to allow both matrix-free and materialized implementations. Second, our language provides key 2 • DeVito, Z. et al Image Warping Energy in Opt !""#$%&'()*+$','-+./$01&2&34&'-+./$01&3&34 "56'.&7'.)'-%$)/.+'8'89&2:&'8;9&2:&'82&9:&'82&;9:':'<5 ''''6','0!""#$%02&24';'!""#$%0.&744';' '''''''''=5%>%$0()*+$024&'!6.*?5#02&24';'!6.*?5#0.&744 ''''@>+.<','()<0A)B5C)<#0.&74&D>#E02&24&D>#E0.&744''''' ''''F)$6*G0-$+$/%0@>+.<&HI6J6&244 $)< /','HI"J0!""#$%#02&24';'K5)#%6>.)%#02&244 F)$6*G0-$+$/%0L>+.<0K5)#%6>.)%#02&244&/&244 Opt Compiler Mesh Deformation Shape From Shading Intrinsic Images Poisson Image Editing Optical Flow Volumetric Deformation Fast GPU Solver Fig. 1. From a high-level description of an energy, Opt produces high-performance GPU-based optimizers for many graphics problems. abstractions for representing energies at a high-level. Unknowns and other data are arranged on 2D or 3D grids, meshes, or general graphs. Energies are defined over these domains and access data through stencil patterns (fixed-size and shift-invariant local windows). Third, our compiler exploits the regularity of stencils and graphs to automatically generate efficient GPU routines that can compute products of derivatives such as JT J. Derivative terms required by these routines are created using hybrid symbolic-automatic differentiation based on a simplified version of the D algorithm [26]. Finally, we use a specialized code generator to emit efficient GPU code for the derivative terms and use metaprogramming to weave the solver code with the generated routines to avoid runtime overhead. Our method provides significantly better performance than traditional general-purpose solver libraries and matches state-of-the-art custom applications. It is easy to change details in the solver (GN vs. LM), the numeric precision (float vs. double), and matrix storage (matrix-free vs. materialized) without rewriting the energies or solvers. Programmers can quickly figure out the best settings for a particular problem depending on the need for numerical stability and available computational resources. In particular, we present the following contributions: • We propose a high-level programming model for defining energies over image and graph domains. • We introduce a generic framework for Gauss-Newton and Levenberg-Marquardt optimization on GPUs that is capable of abstracting the efficient matrix-free methods used in state-of-the-art application-specific solvers. • We provide algorithms based on symbolic differentiation that exploit the regularity of energies defined on images and graphs to produce efficient GPU routines that plug into our optimization framework. Our optimizations produce code competitive with hand-written routines. • We implement a variety of state-of-the-art graphics problems, including mesh/image deformations, smoothing, and shape-from-shading refinement using Opt. We provide an evaluation that shows that our implementations outperform state-of-the-art application-specific solvers and are up to two orders of magnitude faster than the CPU-based Ceres solver [1]. • We show how Opt’s abstraction allows the flexible generation of many solver variants for these applications that explore tradeoffs in GN vs. LM, single vs. double precision, matrix-free vs. materialized, and even hybrid solvers. 2 BACKGROUND Non-linear Least Squares Optimization. Optimization methods are used in the graphics and vision community to solve a wide range of problems. We specifically focus on unconstrained non-linear least squares optimizations [6], where a solver minimizes an energy function that is expressed as a sum of squared residual terms: E(x) = 2 R  r =1 fr (x) . The residuals fr (x) are generic functions, making the problems potentially non-linear and non-convex [9]. There has been an extensive effort in the literature to solve these problems with a large variety of numerical optimization approaches [13, 14, 21, 33, 43, 44, 66]. Gauss-Newton and Levenberg-Marquardt [36, 39] are two common methods for solving problems in computer graphics and vision. GN and LM are specifically tailored towards these kind of problems. Their second-order optimization approach has been shown to be well-suited for the solution of a large variety of problems [38, 67], and has also been successfully applied in the context of real-time optimization [68, 73]. If the non-linear energy is convex, then GN and LM will converge to the global minimum; otherwise they will converge to some local minimum. Furthermore, GN and LM internally solve a linear system. While these systems can generally be solved with direct methods, our solvers need to scale to large problem sizes and run on massively parallel GPUs; hence, we implement GN/LM with a preconditioned conjugate gradient (PCG) [45] in the inner loop. In our current implementation, we focus on GN and LM rather than other variants such as L-BFGS [45], since they reflect the approaches used in state-of-the-art hand-written GPU implementations, allowing us to compare our performance to existing solvers directly. However, we believe our approach can be generalized in future work to support such backends. Application-specific GPU Solvers. Application-specific Gauss-Newton solvers written for GPUs have been frequently used in the last two years. Wu et al. [68] use a blocked version of GN to refine depth from RGB-D data using shape-from-shading. Zollhöfer et al. [73] minimize an as-rigid-as-possible energy [50] on a mesh as part of a framework for real-time non-rigid reconstruction. Zollhöfer et Opt: A Domain Specific Language for Non-linear Least Squares Optimization in Graphics and Imaging al. [72] use a similar solver to enforce shading constraints on a volumetric signed-distance field in order to refine over-smoothed geometry with RGB data. Thies et al. [54–56] transfer local facial expressions between people in a video by optimizing photo-consistency between the video and synthesized output. Dai et al. [11] solve a global bundle adjustment problem to achieve real-time rates for globally-consistent 3D reconstruction, and Innmann et al. [30] optimize the a warp field for of non-rigid surface reconstruction. These solvers achieve high-performance by working matrix-free on the problem domain. That is, during the PCG step, they never form, or materialize the entire Jacobian J of the energy. Instead, they compute it on demand, for instance by reading neighboring pixels to compute the derivative of a regularization energy. Performance improves in two ways: first, they do not explicitly store and load sparse matrix connectivity; rather, this is implied by pixel relationships or meshes. Second, reconstructing terms is often faster than storing them, since the size of the problem data is smaller than the full matrix implied by the energy. Unfortunately, these application-specific solvers are tedious to write because they mix code that calculates complicated matrix products with partial derivatives based on the energy. High-level Solvers. Higher-level solvers such as CVX [19, 20], OpenOF [64], or ProxImaL [28] work directly from an energy specified in a domain-specific language. CVX uses disciplined programming to ensure that modeled energy functions are convex, then constructs a specialized solver for the given type of convex problem. Ceres [1] uses template meta-programming and operator overloading to solve non-linear least squares problems on the CPU using backwards auto-differentiation. Unlike Opt, these solvers do not generate efficient GPU implementations and only work with materialized matrices. OpenOF does run on GPUs, but uses materialized sparse matrices [64]. In contrast, Opt’s abstraction allows solvers to use either matrix-free or materialized approaches; we can even provide hybrids where only part of the energy is materialized. Matrix-free approaches can be significantly faster than explicit matrices due to less memory transfer (Sec. 8.3). CPU libraries such as Alglib [7], GTSAM [12], and g2o [34] abstract the solver, requiring users to provide numeric routines for energy evaluation and, optionally, gradient calculation. All of these solvers create materialized Jacobians, and then use standard numerical linear algebra methods on these matrices to compute the Newton step. They cannot optimize the compilation of energy terms and solver code, unlike application-specific solvers, and require hand-written gradients to run fast. Similar to high-level solvers, Opt only requires a description of the energy, but it uses code transformations to generate application-specific matrix-free (or hybrid) GPU solvers automatically. Simulation DSLs. Ebb [5] and Simit [32] are domain-specific languages that allow the user to express and abstract linear algebra compute operations over graphs (and in Ebb’s case, arbitrary relations such as regular grids) on heterogeneous architectures. Ebb & Simit both focus on simulation, but could be used to write non-linear least squares solvers like those produced by Opt. §3 Cost Expression §4 Generic GN/LM Optimization Framework evalF() evalJTF() applyJTJ() §5 • 3 Opt Frontend Energy (F) IR Solver Routine IR Gen F, JTJ, JTF, J IR §6 IR Optimizer & Scheduler evalJ() Fig. 2. An overview of the architecture of Opt, labeled with the sections where each part is described. The user interface to Opt (specifying concise energy functions) is fundamentally a higher-level abstraction than the direct specification of arithmetic to execute in Simit and Ebb. To write an equivalent optimization in Simit or Ebb, a user must (1) write the solver algorithm from scratch; (2) manually derive efficient arithmetic code from the energy function (particularly tricky for fused Jacobian kernels); and (3) decide up front which parts of matrix multiplication are precomputed and cached vs. recomputed on the fly. Because Opt automates (1 & 2) and allows (3) to be specified post-hoc, users can iterate far more rapidly on their energy functions. We prototyped a Gauss-Newton solver using Ebb for the Image Warping example, but found the solver generated by Opt was over 5x faster then the counterpart in Ebb. The implementation effort of the problem-specific solver in Ebb is similar to a hand-written CUDA implementation, thus significantly higher than specifying the energy in Opt. A future version of Opt could emit Simit or Ebb code, and that may be a practical solution to avoid maintaining multiple back-ends, but does not change the basic system design or address issues (1,2,3) laid out above. Differentiation Methods. Matrix-free approaches require efficient derivative computation since the derivatives are evaluated in the inner iteration of the PCG loop. Numeric differentiation, which uses finite differences to estimate derivatives, is numerically unreliable and inefficient [26]. Instead, packages like Mathematica [65] allow users to compute symbolic derivatives using rewrite rules. Because they frequently represent math as trees, they do not handle common sub-expressions well, making them impractical for large expressions [26]. Automatic-differentiation is transformation on programs rather than symbols [22, 24]. They replace numbers in a program with “dual”-numbers that track a specific partial derivative using the chain rule. However, because the transform does not work on symbols, simplifications that result from the chain rule are not always applied. We use a hybrid symbolic-automatic approach similar to D?, which represents math symbolically but stores it as a directed acyclic graph (DAG) of operators to ensure scalability to large problems [26]. A symbolic representation of derivatives is important for Opt since solver routines use many derivative terms that share common expressions. This can not be addressed by auto-differentiation methods. 4 • DeVito, Z. et al W,H = Dim("W",0), Dim("H",1) X = Unknown("X",float,{W,H},0) A = Array("A",float,{W,H},1) w_fit,w_reg = .1,.9 Energy(w_fit*(X(0,0) - A(0,0)), --fitting w_reg*(X(0,0) - X(1,0)), --regularization w_reg*(X(0,0) - X(0,1))) Fig. 3. Laplacian smoothing energy for a one-component image, implemented in Opt. Note that weights are the square root values, since the Energy function squares its inputs. void SolveLaplacian(int width, int height, float* unknown, float* target) { OptState* state = Opt_NewState(); // load the Opt DSL file containing the cost description OptProblem* problem = Opt_ProblemDefine(state,"laplacian.opt"); // describe the dimensions of the instance of the problem uint32_t dims[] = { width, height }; OptPlan* plan = Opt_ProblemPlan(state, problem, dims); // run the solver void* problem_data[] = { unknown_pixel_data, target_pixel_data }; Opt_ProblemSolve(state, plan, problem_data, NULL); } Fig. 4. Opt API calls that use the Laplacian smoothing program. 3 PROGRAMMING MODEL An overview of Opt’s architecture is given in Fig. 2. In this section, we describe our programming model to construct the problem specific energy functions. Sec. 4 describes our generic solver framework on GPUs, and describes the Gauss-Newton and LevenbergMarquardt solvers we implemented using this framework. To operate matrix-free, this framework requires application-specific solver routines (evalF(),evalJ(),evalJTF(), applyJTJ()). Sec. 5 describes how our compiler generates these routines from the energy. We introduce our programming model using the example of Laplacian smoothing of an image. A fitting term encourages a pixel X to be close to its original value A: E fit (i, j) = [X (i, j) − A(i, j)]2 A regularization term encourages neighboring pixels to be similar: Õ E reg (i, j) = [X (i, j) − X (l, m)]2 (l,m)∈N(i, j) where N (i, j) = {(i + 1, j), (i, j + 1)} The energy is a weighted sum of both terms: Õ E∆ = w fit E fit (i, j) + w reg E reg (i, j) (i, j)∈I While this example is linear, Opt supports arbitrary non-linear energy expressions. Language. Similar to shading languages such as OpenGL, Opt programs are composed of a “shader” file that describes the energy and a set of C APIs for running the problem. Fig. 3 expresses the Laplacian energy in Opt. Opt is embedded in the Lua programming language and operator overloading is used to create a symbolic representation of the energy. The first line specifies problem dimensions. Line 2 uses the function Unknown to declare the pixel array that represents the unknown X . Array is used to declare constant values such as the image A that will be fixed during optimization. The last argument of these declarations is a numeric index that associates the array with actual data provided by the C API. Energy adds residual expressions to the problem’s energy. A key part of Opt’s abstraction is that residuals are described at elements of images or graphs and are implicitly mapped over the entire domain. The term w_fit*(X(0,0) - A(0,0)) defines an energy at each pixel that is the difference between the images. We support arrays and energies that include both vector and scalar terms. The Energy function implicitly squares the terms and sums them over the domain to enforce the linear least-squares model. Terms can also include a N X A G = = = = Dim("N",0) Unknown("X", float3,{N},0) Array("A", float3,{N},1) Graph("Edges", 2, "vertex0", {N}, 3, "vertex1", {N}, 4) w_fit,w_reg = .1,.9 Energy(w_fit*(X(0) - A(0)), w_reg*(X(G.vertex0) - X(G.vertex1))) Fig. 5. The Laplacian cost defined on the edges of a mesh instead of an image. The graph represents explicit connectivity. statically-defined stencil of neighboring pixels. The regularization term w_reg*(X(0,0) - X(1,0)) defines an energy that is the difference between a pixel and the pixel to its right. Our solver framework exploits this regularity to produce efficient code. API. Applications interact with Opt programs using a C API. Fig. 4 shows an example using this API. To amortize the cost of preparing a problem used multiple times, we separate the compilation (Opt_ProblemDefine), memory allocation (Opt_ProblemPlan), and execution (Opt_ProblemSolver) of a problem into different API calls. Mesh-based problems. Opt also includes primitives for defining energies on graphs to support meshes or other irregular structures. Fig. 5 shows an example that smooths a mesh rather than an image. The Graph function defines a set of hyper-edges that connect entries in the unknown together. In this example, each edge connects two entries vertex0 and vertex1, but in general our edges allow an arbitrary number of entries to represent elements, such as three-element hyper edges to define triangles. Energies can be defined on these elements, as seen in the regularization term (line 10), which defines an energy on the edge between two vertices. Boundaries. Defining energies on arrays of pixels requires handling boundaries. By default an entire energy term is considered to have zero energy if any of its accesses would be out of bounds, but we also provide the ability to have custom behavior by querying whether a pixel is valid (InBounds) and selecting a different expression if it is not (Select): term = w_reg*(X(0,0) - X(1,0)) Energy(Select(InBounds(1,0),term,0)) Boundary handling is optimized later in the compilation process to ensure that it does not cause excessive overhead. Opt: A Domain Specific Language for Non-linear Least Squares Optimization in Graphics and Imaging Pre-computing shared expressions. Energy functions for neighboring pixels can share expensive-to-compute expressions. For instance, our shape-from-shading example (Sec. 8) uses an expensive lighting calculation that is shared by neighboring pixels. We allow the user to turn these calculations into computed arrays, which behave like arrays when used in energy functions, but are defined as an expression of other arrays: computed_lighting = ComputedArray(W,H,lighting_calculation(0,0)) Computed arrays can include computations using unknowns, and are recalculated as necessary during the optimization. Similar to scheduling annotations in Halide [48], they allow the user to balance recompute with locality at a high-level. 4 NON-LINEAR LEAST SQUARES OPTIMIZATION FRAMEWORK Our optimization framework is a generalization of the design of application-specific GPU solvers based on the Gauss-Newton approach [30, 54, 55, 68, 72, 73]. However, our solver API abstracts away the specific algorithm details, allowing us to provide options for both Gauss-Newton and Levenberg-Marquardt approaches [6, 36, 39]. We first describe the approach our specific solvers use, and then show how we separate out the details of the application-specific energy from the solver being used. In the context of non-linear least square problems, we consider the optimization objective E : RN → R, which is a sum of squares in the following canonical form: E(x) = R Õ  fr (x) • 5 over-constrained and quadratic, the least squares minimizer is the solution of a linear system of equations. This system is obtained by setting the partial derivatives to zero, which results in the well known normal equations: 2 · J(xk )T J(xk )δk∗ = −2 · J(xk )T F(xk ) This process is iterated for K steps to obtain an approximation to the optimal solution x∗ ≈ xK . The GN approach can be interpreted as a variant of Newton’s method that only requires first-order derivatives and requires less computation. To this end, it uses a first-order Taylor approximation 2(JT J) instead of the real second-order Hessian H. LM additionally introduces a steering parameter λ to switch between GN and Steepest Descent (SD). To this end, the normal equations are augmented with an additional diagonal term. This is similar to Tikhonov regularization and leads to: 2(J(xk )T J(xk ) + λ diag (J(xk )T J(xk )))δk∗ = −2J(xk )T F(xk ) The inverse of λ defines the radius of the trust region. LM guarantees convergence by shrinking the trust region radius and resolving the linear system when a proposed step fails to decrease cost. In the current LM solvers generated by Opt, we allow users to specify an initial trust region radius, minimum and maximum radii, and minimum and maximum values to clamp entries of the diagonal damping matrix. Our specific LM strategy is adapted from the Ceres solver [1]. In our comparisons, we use the same parameter configuration for both solvers. 4.1 Parallelizing the Optimization with PCG 2 r =1 The R scalar residuals fr can be general linear or non-linear functions of the N unknowns x. The objective takes the traditional form used in the Gauss-Newton method: E(x) = F(x) 2 , F(x) = [f 1 (x), . . . , f R (x)]T 2 The R-dimensional vector field F : RN → RR stacks all scalar residuals fr . The minimizer x∗ of E is given as the solution of the following optimization problem: 2 x∗ = argmin E(x) = argmin F(x) 2 x x It is solved based on a fixed-point iteration that incrementally computes a sequence of better solutions {xk }kK=1 given an initial estimate x0 . Here, K is the number of iterations; i.e., x∗ ≈ x K . In every iteration step, a linear least squares problem is solved to find the best linear parameter update. The vector field F is first linearized using a first-order Taylor expansion around the last solution xk : F(xk + δk ) ≈ F(xk ) + J(xk )δk Here, J is the Jacobian matrix and contains the first-order partial derivatives of F. By applying this approximation, the original nonlinear least squares problem is reduced to a quadratic problem: 2 δk∗ = argmin F(xk ) + J(xk )δk 2 δk After the optimal update δk∗ has been computed, a new solution xk +1 = xk + δk can be easily obtained. Since this problem is highly The core of the GN/LM methods is the iterative solution of linear least squares problems for the computation of the optimal linear updates δk∗ . This boils down to the solution of a system of linear equations in each step, i.e., the normal equations. While it is possible to use direct solution strategies for linear systems, they are inherently sequential, while our goal is a fast parallel solution on a many-core GPU architecture with conceptually several thousand independent threads of execution. Consequently, we use a parallel preconditioned conjugate gradient (PCG) solver [63, 73], which is fully parallelizable on modern graphics cards. The PCG algorithm and our strategy to distribute the computations across GPU kernels is visualized in Fig. 6. We run a PCGInit kernel (one time initialization) and three PCGStep kernels (inner PCG loop). Before the PCG solve commences, we initialize the unknowns δ 0 to zero. For preconditioning, we employ the Jacobi preconditioner, which scales the residuals with the inverse diagonal of JT J. Jacobi preconditioning is especially efficient if the system matrix is diagonally dominant, which is true for many problems; for instance, the Laplacian operator and most of its variations are diagonally dominant. When the matrix is not diagonally dominant, we fall back to a standard conjugate gradient descent by user selection. More general preconditioners could be provided as a parameter at code generation time but are not a focus of this paper. A detailed overview of different preconditioning approaches in parallel solvers is given in [23], and matrix-free preconditioners are proposed by [4, 69]. We also default to single-precision floating point numbers throughout, which matches the approach of the recent application-specific 6 • DeVito, Z. et al PCGInit Kernel 2JT F r0 = M 1 F J evalJTF() = 1/Diag(2JT J) 1 p0 = M r0 ↵n0 = reduce(rT0 p0 ) 0 rk M Residual in the k-th iteration Pre-conditioner (remains constant) pk Decent step in the k-the iteration ↵k = ↵nk /↵dk =0 k Load(X,0,0) Load(A,0,0) Load(X,1,0) Load(X,0,1) w_fit Apply(-) Apply(-) Apply(-) Apply(*) Apply(*) Apply(*) Vector of original energy terms Jacobian matrix of F Step size in k-th iteration Vector of PCG unknowns in iteration k for i = 0 to num_linear_iterations: residuals = { fit , h_reg , v_reg } PCGStep1 Kernel gk = 2JT Jpk applyJTJ() ↵dk = reduce(pTk (gk )) PCGStep2 Kernel ↵k = ↵nk /↵dk k+1 = ↵k (gk ) zk+1 = M nk 1 rk+1 = reduce(zTk+1 rk+1 ) PCGStep3 Kernel k = nk /↵nk pk+1 = zk+1 + ↵nk+1 = Fig. 7. The Laplacian example represented in our IR. Application-specific solver routines + ↵k p k k rk+1 = rk w_reg k pk k Fig. 6. Generic GPU architecture for Gauss-Newton and LevenbergMarquardt solvers whose linearized iteration steps are solved in parallel using the preconditioned conjugate gradient (PCG) method. solvers. We believe this strategy is a good compromise between computational effort and efficiency. Stencil-based Array Access. Our techniques for parallelizing work are different for array and graph residuals. For arrays, we group the computation required for each element in the unknown domain onto one GPU thread. For a matrix product such as −2JT F, each row of the output is generated by the thread associated with the unknown. If the unknown is a vector (e.g., RGB pixel), all channels are handled by one thread since these values will frequently share sub-expressions. The computations in a GPU thread can work matrix-free. For instance, if they conceptually require a particular partial derivative from matrix J, they can compute it from the original problem state, which includes the unknowns and any supplementary arrays. Matrices such as J, which are conceptually larger than the problem state, do not need to be written to memory, which minimizes memory accesses. Section 5 describes how we automatically generate these computations from our stencil- and graph-based energy specification. Graph-based Array Access. For graph-based domains, such as 3D meshes, the connectivity is explicitly encoded in a user-provided data structure. Users specify the mapping from graph edges to vertices. Residuals are defined on graph (hyper-) edges and access unknowns on vertices. To make it easy for the user to change the graph over time, we do not require a reverse mapping from unknowns to residuals for graphs. Kernels that use the residuals (PCGInit and PCGStep1) assign one edge in the graph to one GPU thread. Since the output vectors have the same dimension as the unknowns, we have to scatter the terms in the residual evaluations into these values. All threads involving partial sums for a given variable then scatter into the corresponding parts of variables using a floating-point atomic addition. 4.2 Modularizing the Solver A key contribution of our approach is the modularization of the application-specific components of GPU Gauss-Newton or Levenberg-Marquardt solvers into compartmentalized solver routines. The first routine, evalF(), simply generates the application specific energy for each residual. It only runs outside of the main loop to report progress. evalJTF. The second routine appears in the PCGInit kernel and is shown in red in Fig. 6. Here, the initial descent direction p0 is computed using the application-specific evalJTF() routine, which is generated by our compiler. It computes a matrix-free version of −2JT F. evalJTF() is also responsible for computing the preconditioner M, which is simply the dot product of a row of JT with itself. For arrays, a thread computes the rows of an output associated with one element of the unknown. For graphs, each thread only computes the parts of the dot product between JT and F which belong to the handled residual. applyJTJ. The third routine, applyJTJ(), is part of the inner PCG iteration. It computes the multiplication of 2JT J with the current descent direction pk , and incorporates the steering factor λ when using Levenberg-Marquardt. Handling arrays and graphs is similar to evalJTF(). It tends to use more values since it needs to compute entries from both J and JT . For many problems this routine is the most expensive step, so it has to be optimized well. evalJ. While evalJTF and applyJTJ are used in matrix-free code, in some cases materialized matrices are faster. In these cases, our solvers can use the evalJ routine which calculates individual entries in J that the solver can materialize in memory. 5 GENERATING SOLVER ROUTINES A key idea of Opt is that we can exploit the regularity of stenciland graph-based energies to automatically generate applicationspecific solver routines. We represent the mathematical form of the energy as a DAG of operators which we refer to as our intermediate representation (IR). We transform the IR to create new IR expressions needed for evalJTF(), applyJTJ(), and evalJ(). This process requires partial derivatives of the energy. We then optimize this IR and generate code that calculates it. 5.1 Intermediate Representation Since the Opt language is embedded in Lua, we generate the IR by running the Lua program which uses overloaded operators to build the graph. Fig. 7 shows the IR that results from the Laplacian Opt: A Domain Specific Language for Non-linear Least Squares Optimization in Graphics and Imaging -- generates the derivative of expression with respect to variable function derivative(expression, variable) if a cached ver sion of this par t ial der ivat ive ex ist s then return the cached ver sion elseif expression == variable then return 1 end result = 0 for i = 0, the number of ar дument s used by expr ession do result += derivative(argument[i],variable)*partial[i] end cache and r etur n result end Fig. 8. Pseudocode of the OnePass algorithm for generating derivatives. partial[i] is the partial derivative of the particular operator (e.g., *) with respect to the argument i, which is defined for each operator. example. Roots of the IR are residuals that we want to compute. Leaves are constants (e.g., w_fit), input data (e.g the known image A(0,0)), and the the unknown image (e.g., X(0,0)). We de-duplicate the graph as it is built, ensuring common-subexpressions are eliminated. We scalarize vectors from our frontend in the IR to improve the simplification of expressions that become zeros during differentiation. 5.2 Differentiating IR Since we do not always store the Jacobian J in memory, we need to generate residuals on-the-fly. The approach we use for differentiation is similar to Guenter’s D? [26]. It symbolically generates new IR that represents a partial derivative of an existing IR node. Unlike traditional symbolic differentiation (e.g., Mathematica), differentiation is done on a graph where terms can share common sub-expressions. In our implementation we use OnePass, a simplification of D? that can achieve good results by doing the symbolic equivalent of forward auto-differentiation [27]. Pseudocode for the algorithm is given in Fig. 8. It works by memoizing a result for each partial derivative and generates a new derivative of an expression by propagating derivatives from its arguments via the chain rule. 5.3 Generating IR for Matrix Products The IR for evalF() is simply the input energy IR. We generate IR for evalJTF(), applyJTJ(), and evalJ() as transformations of this input IR. The first two terms are conceptually derived from matrix-matrix or matrix-vector multiplications of the Jacobian. Since we compute these values matrix-free, we must generate the IR that will calculate the output given our specific problem. Each term has two versions: one for handling stencil-based and one for graph-based residuals. 5.3.1 Stencil Residuals. Our solver calls applyJTJ() to calculate a single entry of g, where g = 2JT Jp per thread. We need to determine which values from J are required and create IR that calculates them. The non-zero entries in J are determined by the stencil of a particular problem. Fig. 9 illustrates the process of discovering the non-zeros. In the Laplacian case, the partials used in these expressions are actually constants because it is a linear system. However, Opt supports the generic non-linear case, where the partials will be functions of the unknown. • 7 The pseudocode to generate JT J for stencils is shown in Fig. 10. It first finds the residuals that use unknown x0,0 because they correspond to the non-zeros of JT . Some of these residuals are not actually defined at pixel (0, 0), but use x0,0 from neighboring pixels. To find them, we exploit the fact that stencils are invertible. For each residual template in the energy, we examine each place it uses an unknown xi, j . We then shift that residual on the pixel grid, taking each place it loads a stencil value and changing its offset by (−i, −j), which generates a residual in the grid that uses x0,0 . We find all the residuals using x0,0 by repeating the process for each use of an unknown in the template. While we only allow constant stencil offsets, in principle this approach will work for any neighborhood function which is invertible. For each discovered residual, we need the other unknowns it uses which are found by examining the IR symbolically. We then generate the expressions for the part of the matrix-vector products that calculate g0,0 . In this code, we symbolically compute the partial derivatives that are the entries of J. 5.3.2 Graph Residuals. For graphs, residuals are defined on hyperedges rather than on the domain of the unknown and our solver routines are mapped over residuals directly so we do not need an inverse mapping from unknown to residual. Instead one thread computes the part of an output that relates to the residual. Pseudocode to generate applyJTJ() for graphs is given in Fig. 11. At each residual, it generates one row of Jp, and then performs the part of the multiplication for the rows of g that include partials for that residual. The output of this routine is a list of IR nodes that are atomically added into entries of g. 5.3.3 Variants. The approach to generate evalJTF() and evalJ() is similar to that of applyJTJ(). The routine create_jtf() is used to generate the expression r = −2JT F. Each row of JT can be obtained using the same approach previously described. The partials in this row are then multiplied directly with their corresponding residual term in F. Similarly, the routine create_j() simply produces all the non-zero derivatives for a particular residual that can be stored by the solver for use in materialized approaches. In LM we additionally need the term λ diag(JT J), which is inserted into the applyJTJ() routine when needed. 6 OPTIMIZING GENERATED SOLVER ROUTINES We need to translate the IR for evalF(), evalJTF(), applyJTJ(), evalJ() into efficient GPU functions. We simplify the IR based on polynomial simplification rules, optimize the handling of boundary condition statements, and schedule the IR which generates GPU code. Polynomial Simplification. Taking the derivative of the IR tends to introduce more complicated IR. In particular, the application of the multi-variable chain rules introduces statements of the form d 1 ∗ p1 + d 2 ∗ p2 + ... for each argument of an operator. Often some partials are zero, and terms in the sum can be grouped together. We take the approach of other libraries like SymPy [52] and represent primitive math operations as polynomials. In particular, additions and multiplications are represented as n-ary operators rather than binary, and we include a pow operator that raises an expression to a constant ac . Where possible, primitives are represented in terms 8 • DeVito, Z. et al (b) Actual residuals mapped over the entire image. (a) Example residual terms fit: w_fit*(X(0,0) - A(0,0)) h_reg: w_reg*(X(0,0) - X(1,0)) v_reg: w_reg*(X(0,0) - X(0,1)) (c) Residuals using a specific unknown x0,0 x0,1 v reg0,0 x0,0 x-1,0 v reg0,0 Residual template fit0,0 h reg-1,0 h reg0,0 v reg0,-1 h reg0,0 fit0,0 h reg0,0 h reg-1,0 dfit0,0 dx0,0 dh reg0,0 dh reg-1,0 dx0,0 dx0,0 fit0,0 dv reg0,0 dv reg0,-1 dx0,0 dx0,0 h reg0,0 J dh reg0,0 dx0,0 dh reg-1,0 dx0,0 v reg0,0 dv reg0,0 dx0,0 v reg0,-1 dv reg0,-1 dx0,0 residuals → Rows are required for each non-zero column required in T residuals → x1,0 x-1,0 p x0,1 x0,-1 dfit0,0 dx0,0 h reg-1,0 2 unknowns → unknowns → x0,0 v reg0,0 v reg0,-1 Row corresponding to g0,0 has non-zeros for each residual containing x0,0 = fit0,0 dh reg0,0 dx1,0 p0,0 x0,0 p1,0 x1,0 p-1,0 x-1,0 dh reg-1,0 dx-1,0 dv reg0,0 dx0,1 p0,1 dv reg0,-1 dx0,-1 x0,1 p0,-1 x0,-1 unknowns → ←required row→ x1,0 x0,-1 (d) Representation of non-zero entries in the expression g = 2JT Jp that are required to calculate g0,0 g J JT g0,0 relative index to center pixel non-zero columns where each individual residual has support unknowns → (e) Matrix free expression for g0,0 g0,0 = 2 dh reg0,0 dh reg0,0 dh reg0,0 dfit0,0 dfit0,0 p0,0 + 2 ( p0,0 + p1,0 ) + ... = 2w fit2 p0,0 + 2w reg(w regp0,0 + dx0,0 dx0,0 dx0,0 dx0,0 dx1,0 from J T w regp1,0 ) + ... from J Fig. 9. The process our compiler uses for generating applyJTJ() at a high-level. (a) The input to this transformation is a list of individual residuals defined in the IR (fit, h_reg, v_reg) that form a template. (b) The residual template is repeated over the image to generate the actual energy function. (c) The compiler considers a specific unknown x0, 0 , here shown with the residuals that refer to it, which are computed by the compiler. Unknowns and residuals are named relative to this pixel (e.g., h_reg-1, 0 is the horizontal residual from the pixel to the left). (d) The compiler then symbolically forms the result of g = 2JT Jp, here shown with the components needed to generate g0, 0 . The row of JT corresponding to unknown x0, 0 is needed. It has one non-zero for each residual in (c). This row will be multiplied against Jp. The only rows of Jp needed correspond to the residuals appearing in JT since other rows will be multiplied by 0. A row of Jp is calculated by multiplying non-zero entries in a row of J, which occur each time a residual uses an unknown, against the corresponding row of p. (e) Finally, the compiler forms a matrix-free version of the expression for g0, 0 implied by the matrix multiplications, calculating each partial using one-pass differentiation. of these operators. For instance a/b is represented as ab −1 and a − b as a + −1 ∗ b. Polynomial representation makes it easier to find opportunities for optimization such as constant propagation when the optimization first requires re-associating, commuting, or factoring expressions. Importantly, the polynomial representation also gives our scheduler freedom to reorder long sums and products to achieve other goals, such as grouping terms with the same boundary statement into a single if-statement or minimizing register pressure. During construction we optimize non-polynomial terms using constant propagation and applying algebraic identities. Before lowering into code, we also apply a factoring pass that applies a greedy multi-variate version of Horner’s scheme [10] to pull common factors out of large sums. . Bounds Optimization. Boundary conditions introduce another source of inefficiency. Opt uses InBounds and Select to create boundary conditions and masks. Translating these expressions to code can introduce inefficiency in two ways. First, it is possible for the same bound to be checked multiple times. This frequently occurs in applyJTJ() when two partials are multiplied together since both partials often contain the same bound. Redundant checks also occur when reading from arrays since Opt must always check array bounds to avoid crashes. This check is often redundant with a Select already in the energy. Secondly, without optimization, Select statements need to execute both the true and false expressions. For many cases, this means that large parts of the IR, including expensive reads from global memory, do not actually need to be calculated but are performed anyway. Opt: A Domain Specific Language for Non-linear Least Squares Optimization in Graphics and Imaging function create_jtj(residual_templates,X,P) P_hat = 0 residuals = residuals_including_x00(residual_templates) foreach residual do dr_dx00 = differentiate(residual,X(0,0)) foreach unknown u used by residual do dr_du = differentiate(residual,u) P_hat += dr_dx00*dr_dx*P(u.offset_i,u.offset_j) end end return 2*P_hat end function residuals_including_x00(residual_templates) residuals = {} foreach residual_template do foreach unknown x appear inд in residual_template do -- shift the template such that x is centered (i.e. it is x00) R = shift_exp(residual_template,-x.offset_i,-x.offset_j) table.insert(residuals,R) end end return residuals end function shift_exp(exp, shift_i, shift_j) r epl ace each access of any imaдe at (x,y) in exp wit h an access at (x + shift_i,j + shift_j) end Fig. 10. Pseudocode of the compiler transformation that Opt uses to generate JT J from residual templates. function create_jtj_graph(graph_residuals) foreach graph_residual do Jp = 0 -- handle Jp multiply against this residual foreach unknown u appear inд in graph_residual do dr_du = differentiate(graph_residual,u) Jp += dr_du*P(u.index) end -- handle partial sums for Jt*Jp foreach unknown u appear inд in graph_residual do dr_du = differentiate(graph_residual,u) inser t at omic scat t er : P_hat(u.index) += 2*dr_du*Jp end end return set of atomic scatters end Fig. 11. Pseudocode of the compiler transformation that Opt uses to generate JT J for graph residual terms. The common approach of generating two versions of code, one for the boundary region and a bounds-free one for the interior, is less effective on GPUs because they group threads into wide vector lanes of 32 elements, which increases the size of the boundary by the vector width. For smaller sized problems, large portions of the image fall in the boundary region. Instead, we address these two sources of inefficiency directly. We address the redundant bounds checks by augmenting our polynomial simplification routines to handle bounds as well. We represent bounds internally as polynomials containing boolean values b that are either 0 or 1. A Select(b,e_0,e_1) is then represented as b*e_0 + ~b*e_1. We simplify booleans raised to a power b e to b. This representation allows polynomial simplification rules to remove redundant bounds through factoring. We favor booleans over other values during factoring to ensure this simplification occurs. • 9 We address excessive computation and memory use due to bounds by determining when values in the IR need to be calculated. We associate a boolean condition with each IR node that conservatively bounds when it is used. These conditions are generated at Select statements and propagated to their arguments. To improve the effectiveness of this approach, we split large sums into individual reductions that update a summation variable. Each reduction can then be assigned a different condition. When we actually schedule code, we will only execute the code if its condition is true. Scheduling and Code Generation. We translate optimized IR into actual GPU code by scheduling the order in which the code executes the IR. Our scheduler uses a greedy approach that is aware of our boundary optimizations. It starts with the instructions that generate the output values and schedules backwards, maintaining a list of nodes that are ready to be scheduled according to their dependencies. It iteratively chooses an instruction from the ready list that has the lowest cost, schedules it, and updates the list. Our cost function first prioritizes scheduling an instruction with the same condition as the previous instruction, grouping expressions that have the same bounds together into a single if-statement. It then prioritizes choices that greedily minimize the set of live variables at that point in the program, which can provide a small benefit for large expressions. We also prioritize the instruction that has been ready the longest, which also helps reduce the required registers [57]. We translate the scheduled instructions into GPU code using Terra [16]. Terra is a multi-stage programming language with metaprogramming features that allow it to generate high-performance code dynamically. We use its GPU backend to produce CUDA code for the solver routines. To improve the performance, we automatically generate code to bind and load input data from GPU textures. In addition to having better caching behavior, textures also can perform the bounds check for loads automatically. Finally, the solver routines are inlined into the generic solver framework presented earlier. Because this code is compiled together, there is no overhead when invoking solver routines. 7 METAPROGRAMMING FLEXIBILITY Our architecture separates the specification of the energy using the Opt language from the specification of the LM/GN solvers, which interact with the energy only through the abstract solver routines generated by the compiler. This design facilitates various forms of experimentation to produce fast and effective solvers. These choices can be made quickly by changing flags in Opt. LM vs. GN. For matrix-free code, it can take significant effort to add LM extensions to a custom GN solver just to check if they are needed. The application of JT J specifically requires additional energy-specific code that can involve potentially complicated derivatives. In Opt, a single flag enables LM while leveraging the GN solver routine generators that are combined with the LM-specific extensions. This allows for the speed of GN when possible, but the convergence guarantees of LM when necessary. Matrix-free vs. Materialized. A key insight of previous hand-written GPU methods [73] adapted by our framework is that it is sometimes more efficient to compute J in-place rather than store J or JT J as a 10 • DeVito, Z. et al sparse matrix. This approach can be faster for two reasons. First, the locations of the non-zero entries in the matrix are implicitly represented by the problem domain (either an image or a graph), and are not loaded explicitly. Second, entries in the matrix can often be recomputed using less total memory bandwidth than loading the full J matrix. In the extreme case, such as Poisson Image Editing described in the next section, the matrix is constant and the non-zeros can be folded directly into the code. However, this does not apply for compute-intense problems such as fully-differentiated Cotangentweighted Laplacian Smoothing, described in the appendix, where the compute cost dominates bandwidth. Opt addresses this trade off by allowing the (gradual) choice between matrix-free and fully-materialized operations. Opt can use either a matrix-free or materialized approach. The matrix-free approach uses applyJTJ(), while the materialized approach uses evalJ() to materialize J to GPU memory. In our current implementation, we then use cuSPARSE, a highperformance GPU sparse matrix library, inside our materialized PCG solver [46]. This pathway can be extended to work with any GPU sparse solver. We can also represent hybrid approaches where expensive intermediates are materialized using ComputedArray annotations in the energy. The remaining parts of the computation are still computed on demand. This middle ground is sometimes more efficient and is easy to investigate using annotations; see our Shape from Shading example in Fig. 17. Numerical Precision. Opt also allows users to switch between float and double precision depending on the needs of the application and the capability of the GPU compute platform. Although most our graphics example problems are well-conditioned enough for floating point precision, one might want to trade speed for more stability for ill-conditioned problems; for a detailed numeric evaluation on a standard optimization benchmark, we refer to the appendix. Variants of Standard LM/GN. Many other kinds of solvers are also just variants of LM/GN that can fit into the Opt model. For instance, `p problems of the form E(x) = R Õ  fr (x) p r =1 which solve for norms other than L 2 can be computed using Iteratively re-weighted least squares (IRLS), which iteratively solves the least squares problem E(x) = R Õ  2 w i fr (x) r =1   p−2 where fr (x) is a normal least squares function and w i = fr (x) is fixed for the current iteration. Opt realizes these iterative solves by specifying the w i computation as fixed rather than part of the unknown; thus, no derivatives are computed. The example of Intrinsic Images, described in detail in the appendix, uses this type of solver to enforce sparsity on the solution. Robust kernels are another common approach for non-linear least squares optimization problems in computer vision. Here, auxiliary variables are introduced in order to determine the relevance of a  2 data term in part of the optimization formulation. The Robust Mesh Deformation example in the appendix shows how this approach can be naturally expressed in Opt. For some problems, such as Dense Optical Flow, described in the appendix, unknowns are used to sample values from constant images. We support this pattern using a sampled image operator, which can be accessed with arbitrary (u, v) coordinates. When these coordinates are dependent on the unknown image, the user provides the directional derivatives of the sampled image as other input images, which will be used to lookup the partials for the operator in the symbolic differentiation. Domains. Opt is able to exploit the implicit structure and connectivity of general n-dimensional arrays. In addition to images, optimizations are often performed on volumetric grids (e.g., [30, 72] or time-space (e.g., [62]) domains, all of which are subsets of n-D arrays and fall within the scope of Opt. Volumetric Mesh Deformation, as described in the appendix, is an example of solving for unknowns on a 3-D array. Opt efficiently handles large numbers of unknowns at each location in a regular array. For example, the Embedded Deformation example, described in the appendix, uses 12 unknowns per vertex. Opt also efficiently handles explicit structure, provided in the form of general graphs. These domains include manifold meshes and general non-manifolds. For instance, non-rigid mesh deformation approaches (e.g., [50, 51]) fall into this category, as well as widely-used global bundle adjustment methods [2, 49, 58]. Cotangent Laplacian Smoothing, described in the appendix, provides a graph connecting the wedge of triangles at each edge together using graph hyper-edges. Our abstraction also allows the energies on mixed domains. For example, an objective may contain dense regularization terms affecting every pixel of an image and a sparse set of correspondences from a fitting term. Here, the regularization energy is implicitly encoded in a 2D image domain, and the data term may be provided by a sparse graph structure. On all of these domains, Opt provides automatic derivation of objective terms, and generates GPU solvers specifically optimized for a given energy function at compile time. Multi-pass Optimization. In many scenarios, solving a single optimization is not enough, but instead requires multiple passes of different non-linear solves. Often, hierarchal, coarse-to-fine solves are used to achieve better convergence, or sometimes problem-specific flip-flop iteration can be applied (e.g., ARAP flip-flop by Sorkine and Alexa [50]). Another common case are dynamic changes in the structure of the optimization problem. For instance, fitting a mesh to point-cloud data in a non-rigid fashion is typically achieved by searching for correspondences between optimization passes (e.g., non-rigid iterative closest point) [38, 73]. Changes to the correspondences also change the structure of the sparse fitting terms. In all of these examples, custom code is required at specific stages during optimization. To support this code in Opt, we take an approach similar to multi-pass rendering in OpenGL. Between iterations of the Opt solver or between entire solves, users can perform arbitrary modifications to the underlying problem state in C/C++. Optimization weights can be changed (e.g., for parameter Opt: A Domain Specific Language for Non-linear Least Squares Optimization in Graphics and Imaging Energy Length in Opt ARAP Image Warping Interactively edit 2D images by warping them using an as-rigid-as-possible warping energy. ARAP Mesh Deformation Deform a mesh using an asrigid-as-possible warping energy. Shape From Shading Refine depth data from RGB-D scanners using a detailed color image and an estimate of lighting based on spherical harmonics. Poisson Image Editing Splice a source image into a target image without introducing seams. 3D LARAP Mesh Deformation Warp a mesh using an underlying 3D volumetric grid. Embedded Deformation Perform mesh deformation by solving a full affine transformation per vertex. Cotangent Mesh Smoothing Smooth a mesh while preserving the areas of triangles adjacent to each edge. Optical Flow Compute the apparent motion of objects between frames of video at the pixel level. Robust Mesh Deformation A version of ARAP Mesh Deformation that adds a robust kernel. Intrinsic Image Decomposition Separate an image into its reflectance and shading components. 21 lines vs. 280 custom 18 lines vs. 200 custom 96 lines vs. 445 custom 13 lines vs. 67 custom 21 lines 34 lines 32 lines 20 lines 27 lines 32 lines Fig. 12. Example applications written in Opt used in our evaluation. As a proxy for simplicity of implementation, lines of code for the energy in Opt are listed on the right, along with lines of code for the energy-specific code required by handwritten custom solvers when available. Both numbers do not include CPU code for data marshaling and setup. relaxation), underlying data structures may be dynamically updated (e.g., correspondence search or feature match pruning in bundle adjustment problems), or hierarchical and flip-flop strategies can be applied using multiple-passes. This approach allows Opt to support a wide range of solver approaches, while still providing an efficient optimization backend for their inner kernels. 8 EVALUATION To evaluate Opt, we implemented several optimization problems in the language which are summarized in Fig. 12, and described in more detail in the appendix. These include variants of image, volume, and mesh-based problems from the graphics and vision literature. We evaluate overall performance by comparing Opt to four state-of-the-art application-specific matrix-free solvers optimized for GPUs and to five solvers using the high-level Ceres library [1]. We further evaluate the benefits and tradeoffs of Opt’s ability to generate matrix-free, fully-materialized, or intermediate solver variants. We also show the efficiency of our automaticallygenerated solver routines (e.g., applyJTJ) by comparing them to hand-optimized equivalents. Finally, we implement five other problems which demonstrate the generality and expressiveness of Opt, referenced in Sec. 7, and described in detail in the appendix. The Opt code used for the energies of each example is also provided in the appendix. Results are reported as throughput on entire solve steps using a GeForce 1080 GTX, and for CPU results, an Intel Core i7-6700K CPU 4.00 GHz. 8.1 • 11 Comparison with Custom Solvers We compare solvers generated by Opt to existing state-of-the-art CUDA-based application-specific matrix-free solvers optimized for GPUs for ARAP Image Warping, Shape From Shading, ARAP Mesh Deformation, and Poisson Image Editing. Each of the original solvers took months to write, debug, and optimize in CUDA. As a concrete example, debugging the hand-written matrix-free application routine in the custom ARAP image warping originally took weeks due to the complicated cross terms that create dependencies between offsets of one pixel and the angles at a neighbor. In these comparisons, we select the Gauss-Newton backend of Opt to match the algorithmic design in the hand-written reference implementations. The Opt solvers are both significantly easier to write and faster than the handwritten application-specific solvers. In Opt, the energies for each problem could be written in tens of lines of code (Fig 12). Furthermore, Opt outperforms the handwritten solvers for all these example problems by 10-75%, see Fig. 13. Opt can outperform custom solvers because all Opt solvers benefit from optimizations made to the system. The Opt solver for ARAP Mesh Deformation runs 55% faster than custom code due to our reduction-based approach for calculating residuals. In the original solver, the authors only tried the simpler approach of using one pass to compute t = (Jp) and a second for JT t. Opt’s high-level model allowed us to experiment with different approaches more easily during development. In Shape From Shading, the Opt solver runs more than 30% faster than the handwritten CUDA solver. Some of this improvement is due to using texture objects to represent the images, which is an optimization that the original authors did not have time to implement. The ARAP Image Warping solver generated by Opt runs about 75% faster (likely due to better bounds handling) than the handwritten CUDA solver we compare against. Since Poisson Image Editing is a linear problem, we also compare against a custom Cholesky solver with pre-ordering using Eigen [25], a high-performance linear-algebra library for CPUs. The Gauss-Newton method handles linear least-squares problems in a unified way that does not require algorithmic changes. When all residuals are linear functions of the unknowns, J just becomes a constant matrix independent of x. All second order derivatives are zero, which implies that the Gauss-Newton approximation is exact and the optimum can be reached after a single non-linear iteration. The entire Opt solve was 50 times faster than Eigen’s matrix solve (not including its matrix setup time), due to Opt’s ability to implicitly represent the connectivity of the matrix. 8.2 Comparison with General Purpose Solver We also compare Opt against the high-level Ceres library, which is also able to generate a solver using only an energy specification but does not support GPU or matrix-free execution. The solvers generated by Opt are 1–3 orders of magnitude faster than Ceres on our example problems (Fig 13). For accurate comparison, we setup both Opt and Ceres to use the same LM configuration, and plot their convergence over time in Fig 14. To get the fastest results for the internal linear system, we configure Ceres to use its parallel PCG solver for Image Warping and Shape From Shading, and Cholesky factorization for Mesh Deformation. 12 • DeVito, Z. et al Image Warping Ceres (CPU) .00236MP/s Hand-written CUDA Opt 0! 0.05! 0.1! Ceres (CPU) 0.15! 0.2! 0.25! 0.3! Throughput (mega-pixels/s) ARAP Mesh Deformation .528 KV/s of each non-linear iteration, using the ComputedArray construct described in Sec. 7. We show the performance of the linear iterations for different materialization strategies on Shape From Shading in Fig. 17. The linear iterations are most efficient when we materialize the (computeintensive) lighting term and its gradient, but recompute the rest of the Jacobian every linear iteration in a matrix-free approach. Hand-written CUDA 8.4 Opt 0! 5! Shape From Shading Ceres (CPU) 10! Throughput (kilo-vertices/s) 15! 20! .00356 MP/s Hand-written CUDA Opt 0! 0.5! 1! 1.5! 2! 2.5! 3! 25! 30! Throughput (mega-pixels/s) Poisson Image Editing .501 MP/s Eigen (CPU) Hand-written CUDA Opt 0! 5! 10! 15! 20! Throughput (mega-pixels/s) Fig. 13. The solvers generated by Opt perform better than applicationspecific GPU solvers, despite requiring significantly less effort to implement. Additionally, they outperform Ceres implementations by up to three orders of magnitude, despite requiring similar implementation effort. One reason Opt is faster than Ceres is because Opt can represent the connectivity of problems on image and n-D array domains implicitly through stencil relations, while Ceres requires the user to specify energies using a graph formulation. The performance difference for Mesh Deformation is less dramatic than image-based examples because in this case Opt needs to load the connectivity of the problem from the graph data structure. However, Opt still benefits from repetitive stencil terms that are embedded in the generated code, as well as the massive parallelization of the GPU-based solver and on-the-fly computations. The Opt generated solver runs over 720 times faster on Shape From Shading, a relatively complex problem, due in part to a smart materialization strategy enabled by Opt, see Sec. 8.3. The performance benefits of Opt become more pronounced as problem size increases (Fig 15). 8.3 Matrix-free vs. Materialized Solvers As mentioned in Sec. 7, a powerful property of Opt is the ability to use matrix-free representations or hybrid representations while still supporting fully materialized solvers. We show the difference between matrix-free and materialized approaches in Fig 16. Here, Opt uses cuSPARSE for the inner multiply of the PCG solver [46]. Note that cuSPARSE only provides the functionality for linear algebra and by itself it cannot tackle non-linear least squares problems due to a lack of auto- or symbolic differentiation. Except for the highly non-linear Cotangent-weighted Laplacian Smoothing problem, all examples perform between 1.16 and 3 times faster using matrix-free approaches. Opt also allows for intermediate materialization strategies, which allows users to choose which terms to precompute at the beginning Implicit vs. Explicit Connectivity The examples throughout the paper demonstrate that Opt can handle both implicit connectivity on regular grids and explicit connectivity as specified by hypergraphs. It is difficult to quantify the performance improvement due to using an implicit representation of connectivity in general, but we provide a comparison between the two approaches on the Image Warping example in Fig. 18. We compare the performance of Opt using a Gauss-Newton solver over the standard regular grid representation of an image, and Opt using a Gauss-Newton solver over an explicit graph representation of the image. For very small image sizes the performance difference is minimal, but as image size increases the explicit approach takes about twice the amount of time to complete. The implicit approach saves both memory and bandwidth. 8.5 Evaluation of Generated Solver Routines Our approach relies on the symbolic translations of energy functions into efficient solver routines using the optimizations described in Sec. 6. Compared to hand-written code, this code is much easier to write and maintain, but inefficient translations could make it too slow. To show the effectiveness of our symbolic translations and optimization, we compare our generated solver routines to handwritten versions that were taken from the pre-existing CUDA code and slotted into our solver. Fig. 19 shows the results of our optimizations compared to the hand-written versions of solver routines ported from the CUDA examples and modified to use texture loads. The baseline (labeled "none" in the figure) roughly simulates how an auto-differentiation approach based on dual numbers would perform. Our optimizations increase performance up to 8x in the case of Shape From Shading, and are necessary for Opt to perform at or above the speed of hand-written code. Performing polynomial simplifications improves the results of all examples. The improvement is more pronounced for the image-based examples, probably because graph-based examples are bottle-necked by fetching sparse data from memory rather than by the expressions themselves. Our optimizations remove redundant bounds checks and unnecessary reads that can occur when compiling expressions that test boundary conditions. They include representing bounds as booleans, factoring the bounds out of polynomial terms, and scheduling expressions to run conditionally. They provide a significant improvement for both Shape From Shading and Image Warping. Mesh Deformation does not improve because it does not use Select. Shape From Shading shows a significant benefit from texture use, and our register minimization heuristic provides a small benefit to Shape From Shading’s JTJ function. Opt: A Domain Specific Language for Non-linear Least Squares Optimization in Graphics and Imaging • 13 } 756x } 33.5x } 101x } 722x Fig. 14. Convergence of both Opt and Ceres over time, including both double/single precision and GN/LM solvers for Opt. Per non-linear iteration (left), Opt LM and Ceres converge at the similar rates, but Opt converges faster over time (right) by completing each iteration up to several orders of magnitude quicker. Cost and time are both presented using log scale, while iteration count is linear. Vertical lines are drawn for each solver type at the iteration and time when their cost dips below the final Ceres iteration. The performance gap for the fastest Opt solver variant (single-precision GN) versus Ceres is highlighted on each of the graphs on the right. Even the full double-precision Opt LM implementation (often the slowest variant) outperforms Ceres by over an order of magnitude on all problems. 14 • DeVito, Z. et al } } 34x 215x Fig. 15. Performance of Opt compared to Ceres (using a direct Cholesky solver or an iterative PCG for the linear solve) as problem size increases on two example problems. Both unknown count and time to convergence are presented in log scale. Image Warping (left), which uses implicit connectivity in Opt, has more drastic performance differences than ARAP Mesh Deformation (right). For small problems (<5k unknowns in these examples), GPU solvers are inefficient, but the Opt generated solvers rapidly become significantly faster than their Ceres equivalent as problem size increases. Performance of Materialization Strategies Shape From Shading Materialization Fig. 17. Opt lets programmers specify hybrid materialization approaches, which are sometimes more efficient than either full matrix-free, or fully materialized approaches. Here, we show different strategies for the Shape From Shading example. Materializing just the lighting term (Lighting) outperforms fully matrix free, and fully materialized by 2-7x. Fig. 16. Comparison of the performance between the fully materialized and the best matrix-free variant available in Opt, measuring the speed of a linear iteration in PCG. The matrix-free approach is more efficient in all cases besides Cotangent-weighted Laplacian Smoothing. The materialized version creates the JT J matrix outside of the PCG loop. This incurs a onceper-PCG-solve cost not captured by these graphs, so materialized versions will perform even worse than reported here when there are a small number of inner PCG iterations. our specific optimizations are tailored to GPUs, the overall approach of symbolically calculating and simplifying functions needed by the solver is applicable to other platforms such as multi-core CPUs, or even networked clusters of machines for large problems. Finally, there are a lot of optimization problems in graphics that are not suited to the Gauss-Newton or Levenberg-Marquardt approach. Many optimization problems in the graphics literature are more efficiently solved using other techniques such as shape deformation with an interior-point optimizer [37] or mesh parametrization using quadratic programming [31]. Although these problems are not the focus of this paper, their solvers would also benefit from the architecture proposed in Opt, where a general solver library is augmented with automatically derived application-specific routines. 8.6 Limitations and Future Work 9 CONCLUSION Currently the Opt language limits what energies can be expressed efficiently. On images, our implementation limits energies to a constant-sized neighboring stencil. However, we can extend Opt to support other neighborhood functions such as affine transformations of indices as long as the neighborhood function is invertible. We also plan to extend our graph language to support the ability to reference a variable number of neighbors (such as the edges around a vertex) to make certain energies easier to express. While some of We have introduced Opt, a domain specific language that generates high-performance, application-specific GPU solvers from a high-level energy description based on stencils and graphs. Solvers generated with Opt are not only orders of magnitude faster than Ceres, but also outperform state-of-the-art hand-coded applicationspecific solvers which have been tuned with many month of tedious implementation effort. Opt is also highly flexible: it can generate solvers with either floating or double point precision, solvers that Opt: A Domain Specific Language for Non-linear Least Squares Optimization in Graphics and Imaging • } 2.05x [6] [7] [8] [9] [10] [11] 1.26x } [12] [13] [14] Fig. 18. Performance comparison between using the implicit connectivity of a regular grid for the Image Warping problem versus using an explicit graph representation. Both unknown count and time to convergence are presented in log scale. For medium to large size images, the explicit approach takes twice as long. Here, we configured Opt to produce Gauss-Newton solvers that run for 8 nonlinear iterations of 100 linear iterations each. [15] [16] [17] [18] are matrix-free, materialized, and even intermediate hybrids, and variations of GN and LM such as IRLS or robust solvers. Further, Opt provides its own parallel PCG routines to solve for the linear intermediate systems; however, it can also hand off the linear solve to other GPU solvers such as cuSPARSE. Overall, we believe that Opt’s approach of using abstracted solvers with automatically-generated application-specific routines can be extended to work with more expressive energy functions, more platforms beyond GPUs, and more kinds of solvers. Eventually, we hope that computer graphics and vision practitioners can put most energy functions from the literature into a system like Opt and automatically get a high-performance solver. We believe that Opt is a significant first step in this direction. ACKNOWLEDGMENTS This work was supported by the DOE Office of Science ASCR in the ExMatEx and ExaCT Exascale Co-Design Centers, program manager Karen Pao; DARPA Contract No. HR0011-11-C-0007; fellowships and grants from NVIDIA, Intel, and Google; the Max Planck Center for Visual Computing and Communications, and the ERC Starting Grant 335545 CapReal; and the Stanford Pervasive Parallelism Lab (supported by Oracle, AMD, Intel, and NVIDIA). We also gratefully acknowledge hardware donations from NVIDIA Corporation. [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] REFERENCES [1] Sameer Agarwal, Keir Mierle, and Others. 2010. Ceres Solver. http://ceres-solver. org. (2010). [2] Sameer Agarwal, Noah Snavely, Ian Simon, Steven M Seitz, and Richard Szeliski. 2009. Building rome in a day. In Computer Vision, 2009 IEEE 12th International Conference on. IEEE, 72–79. [3] Marc Alexa, Daniel Cohen-Or, and David Levin. As-rigid-as-possible Shape Interpolation (SIGGRAPH ’00). New York. [4] Stefania Bellavia, Jacek Gondzio, and Benedetta Morini. 2013. A matrix-free preconditioner for sparse symmetric positive definite systems and least-squares problems. SIAM Journal on Scientific Computing 35, 1 (2013), A192–A211. [5] Gilbert Louis Bernstein, Chinmayee Shah, Crystal Lemire, Zachary Devito, Matthew Fisher, Philip Levis, and Pat Hanrahan. 2016. Ebb: A DSL for Physical [31] [32] [33] [34] 15 Simulation on CPUs and GPUs. ACM Transactions on Graphics (TOG) 35, 2 (2016), 21. Ake Björck. 1996. Numerical methods for least squares problems. Siam. Sergey Bochkanov. 1999. ALGLIB. http://www.alglib.net. (1999). Nicolas Bonneel, Kalyan Sunkavalli, James Tompkin, Deqing Sun, Sylvain Paris, and Hanspeter Pfister. 2014. Interactive Intrinsic Video Editing. 33, 6 (Nov. 2014), 197:1–10. DOI:https://doi.org/10.1145/2661229.2661253 Stephen Boyd and Lieven Vandenberghe. 2004. Convex Optimization. Cambridge University Press, New York, NY, USA. Martine Ceberio and Vladik Kreinovich. 2004. Greedy Algorithms for Optimizing Multivariate Horner Schemes. SIGSAM Bull. 38, 1 (March 2004), 8–15. DOI: https://doi.org/10.1145/980175.980179 Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. 2017. BundleFusion: Real-time Globally Consistent 3D Reconstruction using On-the-fly Surface Reintegration. ACM Transactions on Graphics (TOG) 36, 3 (2017), 24. Frank Dellaert. 2012. GTSAM. https://collab.cc.gatech.edu/borg/gtsam. (2012). Ron S Dembo, Stanley C Eisenstat, and Trond Steihaug. 1982. Inexact newton methods. SIAM Journal on Numerical analysis 19, 2 (1982), 400–408. Ron S Dembo and Trond Steihaug. 1983. Truncated-newtono algorithms for large-scale unconstrained optimization. Mathematical Programming 26, 2 (1983), 190–212. Mathieu Desbrun, Mark Meyer, Peter Schröder, and Alan H. Barr. 1999. Implicit Fairing of Irregular Meshes Using Diffusion and Curvature Flow (SIGGRAPH ’99). New York. Zachary DeVito, James Hegarty, Alex Aiken, Pat Hanrahan, and Jan Vitek. 2013. Terra: A Multi-stage Language for High-performance Computing (PLDI ’13). New York, 105–116. Marek Dvoroznak. 2014. Interactive As-Rigid-As-Possible Image Deformation and Registration. In The 18th Central European Seminar on Computer Graphics. R. Fourer, D. M. Gay, and B. Kernighan. 1989. Algorithms and Model Formulations in Mathematical Programming. Springer-Verlag New York, Inc., New York, Chapter AMPL: A Mathematical Programming Language, 150–151. http://dl.acm.org/citation.cfm?id=107479.107491 Michael Grant and Stephen Boyd. 2008. Graph implementations for nonsmooth convex programs. In Recent Advances in Learning and Control. Springer-Verlag Limited, 95–110. Michael Grant and Stephen Boyd. 2014. CVX: Matlab Software for Disciplined Convex Programming, version 2.1. http://cvxr.com/cvx. (March 2014). Serge Gratton, Amos S Lawless, and Nancy K Nichols. 2007. Approximate GaussNewton methods for nonlinear least squares problems. SIAM Journal on Optimization 18, 1 (2007), 106–132. Andreas Griewank and Andrea Walther. 2008. Evaluating Derivatives: Principles and Techniques of Algorithmic Differentiation (second ed.). SIAM, Philadelphia, PA, USA. Marcus J Grote and Thomas Huckle. 1997. Parallel preconditioning with sparse approximate inverses. SIAM Journal on Scientific Computing 18, 3 (1997), 838–853. F. Grund. 1982. Automatic Differentiation: Techniques and Applications. Lecture Notes in Computer Science 120. ZAMM 62, 7 (1982), 355–355. DOI:https://doi. org/10.1002/zamm.19820620735 Gaël Guennebaud, Benoît Jacob, and others. 2010. Eigen v3. http://eigen.tuxfamily.org. (2010). Brian Guenter. 2007. Efficient Symbolic Differentiation for Graphics Applications (SIGGRAPH ’07). Article 108. Brian Guenter, John Rapp, and Mark Finch. 2011. Symbolic Differentiation in GPU Shaders. Technical Report MSR-TR-2011-31. http://research.microsoft.com/apps/ pubs/default.aspx?id=146019 Felix Heide, Steven Diamond, Matthias Nießner, Jonathan Ragan-Kelley, Wolfgang Heidrich, and Gordon Wetzstein. 2016. ProxImaL: Efficient image optimization using proximal algorithms. ACM Transactions on Graphics (TOG) 35, 4 (2016), 84. Berthold K. P. Horn and Brian G. Schunck. 1981. Determining Optical Flow. ARTIFICAL INTELLIGENCE 17 (1981), 185–203. Matthias Innmann, Michael Zollhöfer, Matthias Nießner, Christian Theobalt, and Marc Stamminger. 2016. VolumeDeform: Real-time volumetric non-rigid reconstruction. In European Conference on Computer Vision. Springer, 362–379. Liliya Kharevych, Boris Springborn, and Peter Schröder. 2006. Discrete conformal mappings via circle patterns. ACM Transactions on Graphics (TOG) 25, 2 (2006), 412–438. Fredrik Kjolstad, Shoaib Kamil, Jonathan Ragan-Kelley, David IW Levin, Shinjiro Sueda, Desai Chen, Etienne Vouga, Danny M Kaufman, Gurtej Kanwar, Wojciech Matusik, and others. 2016. Simit: A language for physical simulation. ACM Transactions on Graphics (TOG) 35, 2 (2016), 20. Dana A Knoll and David E Keyes. 2004. Jacobian-free Newton–Krylov methods: a survey of approaches and applications. J. Comput. Phys. 193, 2 (2004), 357–397. Rainer Kummerle, Giorgio Grisetti, Hauke Strasdat, Kurt Konolige, and Wolfram Burgard. 2011. g2o: A general framework for graph optimization. In Robotics and 16 • DeVito, Z. et al Shape From Shading speedup over none 8! 8! JTF 2.5! JTF 2! 2! 1.5! Mesh Deformation JTJ handwritten 2.5! 2.5! 2.5 JTF + register + texture + bounds 0! handwritten none 0! JTJ + poly 0! + register 0! + texture 0.5! 0.5 + bounds 0.5! none 0.5! + poly 0.5! + register 1! + texture 1! + bounds 1! none 1! + poly 1.5! 1.5 + register 1.5! + texture 1.5! + bounds 2! none 2! + poly + register + texture + bounds 0! handwritten none 0! JTJ + poly 2! 2 + register 2! + texture 4! 4 + bounds 4! none 6! + poly 6! Image Warping 2.5! Fig. 19. The effect of our optimizations on generated code for solver routines including polynomial simplification (poly), bounds optimization (bounds), texture loads (texture), and register minimization (register), compared to a baseline with no optimization (none). The optimizations are necessary to match the performance of handwritten equivalents, and in many cases the functions run significantly faster. Automation (ICRA), IEEE Int. Conf. on. IEEE, 3607–3613. [35] Edwin H. Land and John J. McCann. 1971. Lightness and Retinex Theory. Journal of the Optical Society of America 61, 1 (Jan. 1971), 1–11. DOI:https://doi.org/10. 1364/JOSA.61.000001 [36] Kenneth Levenberg. 1944. A method for the solution of certain non-linear problems in least squares. Quarterly of applied mathematics 2, 2 (1944), 164–168. [37] Zohar Levi and Denis Zorin. 2014. Strict minimizers for geometric optimization. ACM Transactions on Graphics (TOG) 33, 6 (2014), 185. [38] Hao Li, Bart Adams, Leonidas J Guibas, and Mark Pauly. 2009. Robust single-view geometry and motion reconstruction. In ACM Transactions on Graphics (TOG), Vol. 28. ACM, 175. [39] Donald W Marquardt. 1963. An algorithm for least-squares estimation of nonlinear parameters. Journal of the society for Industrial and Applied Mathematics 11, 2 (1963), 431–441. [40] Hermann Matthies and Gilbert Strang. 1979. The solution of nonlinear finite element equations. International journal for numerical methods in engineering 14, 11 (1979), 1613–1626. [41] Abhimitra Meka, Michael Zollhöfer, Christian Richardt, and Christian Theobalt. 2016. Live Intrinsic Video. ACM Transactions on Graphics (Proceedings SIGGRAPH) 35, 4 (2016). DOI:https://doi.org/10.1145/2897824.2925907 [42] Mark Meyer, Mathieu Desbrun, Peter SchrÃűder, and AlanH. Barr. 2003. Discrete Differential-Geometry Operators for Triangulated 2-Manifolds. In Visualization and Mathematics III, Hans-Christian Hege and Konrad Polthier (Eds.). Springer Berlin Heidelberg, 35–57. DOI:https://doi.org/10.1007/978-3-662-05105-4_2 [43] Stephen G Nash. 2000. A survey of truncated-Newton methods. J. Comput. Appl. Math. 124, 1 (2000), 45–59. [44] Jorge Nocedal and Stephen Wright. 2006. Numerical optimization. Springer Science & Business Media. [45] J. Nocedal and S. J. Wright. 2006. Numerical Optimization (2nd ed.). Springer, New York. [46] NVIDIA 2012. CUDA CUSPARSE Library. NVIDIA. [47] Patrick Pérez, Michel Gangnet, and Andrew Blake. 2003. Poisson Image Editing. In ACM SIGGRAPH 2003 Papers (SIGGRAPH ’03). ACM, New York, NY, USA, 313–318. DOI:https://doi.org/10.1145/1201775.882269 [48] Jonathan Ragan-Kelley, Andrew Adams, Sylvain Paris, Marc Levoy, Saman Amarasinghe, and Frédo Durand. 2012. Decoupling Algorithms from Schedules for Easy Optimization of Image Processing Pipelines. ACM Trans. Graph. 31, 4, Article 32 (July 2012), 12 pages. DOI:https://doi.org/10.1145/2185520.2185528 [49] Noah Snavely, Steven M Seitz, and Richard Szeliski. 2006. Photo tourism: exploring photo collections in 3D. In ACM transactions on graphics (TOG), Vol. 25. ACM, 835–846. [50] Olga Sorkine and Marc Alexa. 2007. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, Vol. 4. [51] Robert W. Sumner, Johannes Schmid, and Mark Pauly. 2007. Embedded Deformation for Shape Manipulation (SIGGRAPH ’07). New York, Article 80. [52] SymPy Development Team. 2014. SymPy: Python library for symbolic mathematics. http://www.sympy.org [53] The MathWorks Inc. 2015. MATLAB. (2015). [54] J. Thies, M. Zollhöfer, M. Nießner, L. Valgaerts, M. Stamminger, and C. Theobalt. 2015. Real-time Expression Transfer for Facial Reenactment. ACM Transactions on Graphics (TOG) 34, 6 (2015). [55] J. Thies, M. Zollhöfer, M. Stamminger, C. Theobalt, and M. Nießner. 2016. Face2Face: Real-time Face Capture and Reenactment of RGB Videos. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE. [56] Justus Thies, Michael Zollhöfer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. 2016. FaceVR: Real-Time Facial Reenactment and Eye Gaze Control in Virtual Reality. arXiv preprint arXiv:1610.03151 (2016). [57] Michael D. Tiemann. 1989. The GNU instruction scheduler. Technical Report. Cambridge, MA. [58] Bill Triggs, Philip F McLauchlan, Richard I Hartley, and Andrew W Fitzgibbon. 1999. Bundle adjustmentâĂŤa modern synthesis. In Vision algorithms: theory and practice. Springer, 298–372. [59] Daniel Vlasic, Ilya Baran, Wojciech Matusik, and Jovan Popović. 2008. Articulated mesh animation from multi-view silhouettes. In ACM Transactions on Graphics (TOG), Vol. 27. ACM, 97. [60] Daniel Vlasic, Pieter Peers, Ilya Baran, Paul Debevec, Jovan Popović, Szymon Rusinkiewicz, and Wojciech Matusik. 2009. Dynamic shape capture using multiview photometric stereo. ACM Transactions on Graphics (TOG) 28, 5 (2009), 174. [61] A. Walther and A. Griewank. 2012. Getting started with ADOL-C. In Combinatorial Scientific Computing, U. Naumann and O. Schenk (Eds.). Chapman-Hall CRC Computational Science, Chapter 7, 181–202. [62] Michael Wand, Philipp Jenke, Qixing Huang, Martin Bokeloh, Leonidas Guibas, and Andreas Schilling. 2007. Reconstruction of deforming geometry from timevarying point clouds. In Symposium on Geometry processing. Citeseer, 49–58. [63] Daniel Weber, Jan Bender, Markus Schnoes, André Stork, and Dieter Fellner. 2013. Efficient GPU data structures and methods to solve sparse linear systems in dynamics applications. In Computer Graphics Forum, Vol. 32. Wiley Online Library, 16–26. [64] Cornelius Wefelscheid and Olaf Hellwich. 2013. OpenOF: Framework for sparse non-linear least squares optimization on a GPU. (In VISAPP). [65] Wolfram Research. 2000. Mathematica. (2000). [66] SJ Wright and John Norman Holt. 1985. An inexact levenberg-marquardt method for large sparse nonlinear least squres. The Journal of the Australian Mathematical Society. Series B. Applied Mathematics 26, 04 (1985), 387–403. [67] Changchang Wu, Sameer Agarwal, Brian Curless, and Steven M Seitz. 2011. Multicore bundle adjustment. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on. IEEE, 3057–3064. [68] Chenglei Wu, Michael Zollhöfer, Matthias Nießner, Marc Stamminger, Shahram Izadi, and Christian Theobalt. 2014. Real-time Shading-based Refinement for Consumer Depth Cameras. ACM Transactions on Graphics (TOG) 33, 6 (2014). [69] Wei Xu, Ning Zheng, and Ken Hayami. 2016. Jacobian-Free Implicit Inner-Iteration Preconditioner for Nonlinear Least Squares Problems. Journal of Scientific Computing (2016), 1–27. [70] Yuting Yang, Sam Prestwood, and Connelly Barnes. 2016. VizGen: accelerating visual computing prototypes in dynamic languages. ACM Transactions on Graphics (TOG) 35, 6 (2016), 206. [71] Christopher Zach. 2014. Robust bundle adjustment revisited. In Computer Vision– ECCV 2014. Springer, 772–787. [72] Michael Zollhöfer, Angela Dai, Matthias Innmann, Chenglei Wu, Marc Stamminger, Christian Theobalt, and Matthias Nießner. 2015. Shading-based Refinement on Volumetric Signed Distance Functions. ACM Trans. Graph. 34, 4, Article 96 (July 2015), 14 pages. DOI:https://doi.org/10.1145/2766887 [73] Michael Zollhöfer, Matthias Nießner, Shahram Izadi, Christoph Rhemann, Christopher Zach, Matthew Fisher, Chenglei Wu, Andrew Fitzgibbon, Charles Loop, Christian Theobalt, and Marc Stamminger. 2014. Real-time Non-rigid Reconstruction using an RGB-D Camera. ACM Transactions on Graphics (TOG) 33, 4 (2014). Opt: A Domain Specific Language for Non-linear Least Squares Optimization in Graphics and Imaging APPENDIX In this Appendix, we provide additional details and evaluations for Opt. In Sec. A, we detail the descriptions of our test example problems that we have implemented and evaluated in Opt. The problems are split into two categories; first, those where we compare the performance against other solvers, and second, those that demonstrate the flexibility of Opt. In Sec. B, we evaluate Opt’s numerical behavior on a standard optimization benchmark. We provide numbers for different solver variations generated with Opt: float vs. double and Gauss-Newton vs Levenberg-Marquardt. In Sec. C, we show how our example problems are written in Opt. We provide the Opt energies, which are similar to graphics shaders. A DETAILED DESCRIPTION OF EXAMPLE PROBLEMS Along with the main contribution of Opt, we provide 10 different example optimization problems in Opt. We chose the first four example applications (see Sec. A.1) since they are commonly used in graphics research and optimized GPU code previously existed or could be easily adapted for the problem. For these problems, we implemented their energies in Opt and compare against previously written custom CUDA implementations. The hand-written CUDA baselines are (improved) versions of the authors’ state-of-the-art implementations. We also implemented these applications in the Ceres solver for direct comparison to another high-level solver [1]. The remaining applications detailed in Sec. A.2 show how a variety of different applications can be solved using the Opt programming model. A.1 Performance Examples • 17 Opt performs around 25 times faster than a Ceres example which is implemented in around 100 lines of code. A.1.3 Shape From Shading. In the Shape From Shading example, we use an optimizer to refine depth data captured by RGB-D scanners [68]. It uses a detailed color image and an estimate of the lighting based on spherical harmonics to refine the lower resolution depth information. Shape from Shading, which is adapted from Wu et al.’s work [68], is our most complex problem. The original implementation was a patch solver variation of a Gauss Newton solver that used shared memory at the expense of per-iteration convergence. For a more direct comparison, we ported the original code into a non-patch solver, which actually improved the convergence time over the author’s implementation. The CUDA code includes 445 lines to express the energy, applyJTJ, and evalJTF calculations. It took several months for a group of researchers to implement and optimize. In comparison, the Opt solver code is around 100 lines and runs more than 30% faster. Some of this improvement is due to using texture objects to represent the images, which is an optimization that the original authors did not have time to do. Shape from Shading also benefits from using pre-computed arrays. We instruct Opt to pre-compute a lighting term and a boundary term that are expensive to calculate and used by the energy of multiple pixels. Without this annotation, Opt runs over 7 times slower. We expect that other complicated problems will have similar behavior and pre-computed arrays will give the user an easy way to experiment with how the computation is scheduled. A.1.1 Image Warping. As-rigid-as-possible Image Warping is used to interactively edit 2D shapes in a way that minimizes a warping energy. It penalizes deviations from local rigidity, while warping to a set of user-specified constraints [17]. It co-optimizes the new pixel coordinates along with the per-pixel rotation. The CUDA implementation was adapted from the hand-written solver created by Zollhöfer et al. [73] for real-time non-rigid reconstruction. It requires around 480 lines of code to implement. Of that, 200 were devoted to the Gauss-Newton solver, and 280 to expressions for the solver routines. In this comparison, we jointly solve for rotations and translations, following the hand-written reference implementation. Note that alternating between rotation and translation in a global-local flip-flop solve is also feasible in Opt; however, overall convergence is typically worse than the joint solve [73]. In comparison, the solver generated by Opt runs about 75% faster (likely due to better bounds handling), and only requires about 20 lines of code to describe. Ceres code is more comparable in size to Opt, at around 100 lines, but it runs 100 times slower. A.1.4 Poisson Image Editing. Poisson Image Editing is used to splice a source image into a target image without introducing seams [47]. Its energy function preserves the gradients of the source image while matching the boundary to gradients in the target image. The energy formulation in this function makes J constant, and our matrix-free solver is able to inline those constants into the solver routines rather than load them from memory. To compare against a CUDA version, we adapted the Image Warping CUDA example to use the Poisson Image Editing objective function, which uses about 67 lines for the energy, evalJTF, and applyJTJ. Opt performs about 10% faster and uses only about 15 lines of code. Since this problem boils down to solving a linear system of equations, we also compare against Eigen [25], a highperformance linear-algebra library for CPUs using Cholesky with pre-ordering since it was fastest. The entire Opt solve was 50 times faster than Eigen’s matrix solve (not including its matrix setup time), due to Opt’s ability to implicitly represent the connectivity of the matrix. A.1.2 Mesh Deformation. As-rigid-as possible mesh deformation [50] is a variant of the previous example that shows Opt’s ability to run on mesh-based problems using its graph abstraction. It defines a warping energy on the edges of the mesh rather than neighboring pixels and uses 3D coordinate frames. The CUDA solver was also adapted from Zollhöfer et al. [73]. It is similar in size to the previous example, with around 200 lines devoted to expressing the energy, applyJTJ, and evalJTF calculations. A.2 Expressiveness Examples Our performance evaluations (see Sec. A.1) focus on examples from the literature where state-of-the-art hand-written code previously existed and can be compared. Opt’s programming model is also able to handle a wider variety of general non-linear least squares problems, which is at the core of many computer graphics and vision problems. 18 • DeVito, Z. et al A.2.1 Embedded Deformation. Embedded Deformation is a popular alternative method to as-rigid-as-possible deformation [51]. Rather than solving for a per-vertex rotation parameterized by Euler angles, Embedded Deformation solves for a full transformation matrix and enforces the rotations via additional soft-constraints. Compared to writing a solver by hand, writing Embedded Deformation in Opt was an easy process, since it only required increasing the number of unknowns and changing the energy terms of our as-rigid-as-possible energy, which amounted to tens of lines of code and under an hour of work. Gauss-Newton solvers are fragile on this energy function, so we configured Opt to generate an LM solver instead. The solver it produces can deform a 12k vertex mesh at an interactive rate of 5 frames per second and only requires an energy function of around 40 lines. A.2.2 Cotangent-weighted Laplacian Smoothing. Cotangentweighted Laplacian Smoothing is a method for smoothing meshes that tries to preserves the area of triangles adjacent to each edge [15]. It adapts well to meshes with non-uniform tessellations. This example highlights Opt’s ability to define residuals on larger components of a mesh by defining a hyper-edge in our graph representation that contains all the vertices in a wedge at each edge. We show the power of Opt by implementing a small variant that allows the cotangent weights to be recomputed during deformation instead of using the values from the original mesh. This variant is normally hard to write since it introduces complicated derivative terms. Opt generates them automatically, making it easier to experiment with small variants on existing energy functions. Opt generates a solver from around 45 lines of code that can smooth a 44k vertex mesh at 3.3 fps. A.2.3 Dense Optical Flow. Dense Optical Flow computes the apparent motion of objects between frames in a video at the pixel level. We implement a hierarchical version of Horn and Schunck’s algorithm [29] using an iterative relaxation scheme. Because optical flow is searching for correspondences in images, its unknowns are used to sample values from the input frames. We support this pattern using a sampled image operator, which can be accessed with arbitrary (u, v) coordinates. When these coordinates are dependent on the unknown image, the user provides the directional derivatives of the sampled image as other input images, which will be used to lookup the partials for the operator in the symbolic differentiation. The solver Opt generates from around 20 lines of code solves optical flow at 5.38MP/s. A.2.4 Intrinsic Image Decomposition. Intrinsic Image Decomposition separates an input image into its reflectance and shading components. This problem dates back to the seminal Retinex [35] work, which assumes that large gradients are more likely to correspond to reflectance than shading variation. Current approaches [8, 41] implement this assumption based on energy minimization by enforcing reflectance sparseness and shading smoothness. We reimplemented a basic version of Meka et al. [41] in Opt to show its ability to solve optimization problems that involve sparsity inducing norms; see Fig. 20. Specifying the three objectives (reproduction of input image, p reflectance sparsity and 2 shading smoothness) only requires 37 lines of code in the Opt language. Opt solves the intrinsic image decomposition problem for 461k unknowns (pixel resolution of 640 × 360, 2 unknowns per pixel), in less than 25.1 ms, fast enough to operate in real time on 30Hz video. Internally, Opt minimizes the energy based on Iteratively Reweighted Least Squares (IRLS). Fig. 20. Opt easily allows exploration of different p -norms by generating IRLS solvers. In Meka et al. [41], smaller values of p better separate reflectance and shading. A.2.5 Volumetric Mesh Deformation. Volumetric Mesh Deformation is an alternative approach to mesh deformation that warps an underlying 3-dimensional grid that the mesh is embedded in; e.g., see Innmann et al. [30]. The structure of the problem is implicit in the grid representation, so we stand to gain more than standard graph problems by using Opt over a solver that materializes the intermediate matrices. To demonstrate this performance benefit, we implemented this problem in both Ceres and Opt. The solver Opt generates from 27 lines of code solves for the deformation of a 4961 voxel grid at 21.4ms, sufficient for real-time applications. This is more than 760x faster than the Ceres solver. A.2.6 Robust Non-Rigid Alignment. Robust optimization is an alternative to robust norms that is often used for computer vision problems. The idea is to introduce auxiliary variables that weight data points as part of the optimization process. For instance, this strategy is often used in bundle adjustment or non-rigid deformation frameworks to determine the reliability of correspondences [38, 58, 71, 73]. In Opt, it is easy to add these terms using additional unknowns for energy functions on single and mixed domains. We implemented a robust term along with a fitting term for our Mesh Deformation example in only 3 extra lines of solver code; the solver it produces can deform a 10k vertex mesh at an interactive rate of 3.7 frames per second. In this example, we include a data fitting term that constrains the deformation based on a target point cloud using a point-to-plane term; i.e., a non-rigid ICP. As ground truth 3D captures, we use the data from Vlasic et al. [59, 60]. In order to make the problem harder, we introduce artificial noise by adding spurious correspondences that are far off from the surface. During the optimization, the robust optimization minimizes the weight of these outliers, and is able to achieve a robust non-rigid alignment of the mesh with respect to the target point cloud. A visual comparison with and without robust optimization is shown in Fig. 21, bottom right. B NUMERIC EVALUATION In this section, we provide a numeric evaluation of Opt. Our aim is not to evaluate performance, but rather to exercise numerics. To this end, we implemented all univariate energies from the National Institute of Standards and Technology (NIST) benchmark for non-linear Opt: A Domain Specific Language for Non-linear Least Squares Optimization in Graphics and Imaging least squares problems2 , a standardized benchmark for optimization. These objectives are non-linear regression problems that contain few unknowns (2-9) each with up to a few hundred data points; they are classified by three levels of difficulty (lower, average, higher). On all of these problems, we run four Opt solvers, float vs. double and GN vs. LM, as well as the Ceres solver which uses LM with double precision. The results are visualized in Fig. 22 and Fig. 23. As we can see there is virtually no difference between the solver types on easier problems, on some of the harder problems combinations of the numerically more stable double precision and LM are necessary – our reference is to match Ceres. C EXAMPLE ENERGY SPECIFICATIONS Fig. 24 through Fig. 33 list the Opt code for each of the ten energy functions described in Section A. Opt code is based on the Lua programming language, but uses specialized operations to specify the dimensions and types of unknowns and other variables. All energy functions are centered at a local origin, with all access being relative to this origin, which is replicated over the domain of each energy function. 2 http://www.itl.nist.gov/div898/strd/nls/nls_main.shtml • 19 20 • DeVito, Z. et al Image Warping Shape From Shading 539k unknowns, 3 unknowns per pixel, 5-point stencil 0.250 MP/s, 1.8x faster than CUDA, 101x Ceres 192k unknowns, 1 per pixel, 9-point stencil 2.99 MP/s, 1.3x faster than CUDA, 722x Ceres Cotangent Mesh Smoothing Optical Flow Robust Non-Rigid Alignment 132k unknowns, 3 per vertex, triangle and vertex energies, 147 kverts/s 614k unknowns, 2 per pixel, 5-point stencil 5.38 MP/s 70k unknowns, 7 per vertex,vertex and edge energies 37.1 kverts/s ARAP Mesh Deformation 360k unknowns, 6 per vertex,vertex and edge energies 18.1 kverts/s, 1.6x faster than CUDA, 33.6x Ceres Poisson Image Editing 1.3M unknowns, 4 unknowns per pixel, 5-point stencil 26.3 MP/s 1.1x faster than CUDA, 50.0x Eigen Embedded Mesh Deformation 154k unknowns, 12 per vertex, vertex and edge energies 66.4 kverts/s Intrinsic Image Decomposition Volumetric Mesh Deformation Fig. 21. Visualization of our example problems implemented in Opt. 461k unknowns, 2 per pixel, 5-point stencil 9.19 MP/s 30k unknowns, 7-point stencil (3D) 232 kvoxels/s, 769x faster than Ceres Opt: A Domain Specific Language for Non-linear Least Squares Optimization in Graphics and Imaging • 21 Fig. 22. Cost vs. iterations on the first 15 univariate problems from the NIST non-linear least squares optimization benchmark: http://www.itl.nist.gov/div898/ strd/nls/nls_main.shtml. 22 • DeVito, Z. et al Fig. 23. Cost vs. iterations on the last 11 univariate problems from the NIST non-linear least squares optimization benchmark: http://www.itl.nist.gov/div898/ strd/nls/nls_main.shtml. Opt: A Domain Specific Language for Non-linear Least Squares Optimization in Graphics and Imaging local W,H = Dim("W",0), Dim("H",1) local Offset = Unknown("Offset",float2,{W,H},0) local Angle = Unknown("Angle",float,{W,H},1) local UrShape = Array("UrShape", float2,{W,H},2) --original mesh position local Constraints = Array("Constraints", float2,{W,H},3) -- user constraints local Mask = Array("Mask", float, {W,H},4) -- validity mask for mesh local w_fitSqrt = Param("w_fitSqrt", float, 5) local w_regSqrt = Param("w_regSqrt", float, 6) Exclude(Not(eq(Mask(0,0),0))) --regularization for x,y in Stencil { {1,0}, {-1,0}, {0,1}, {0, -1} } do local e_reg = w_regSqrt*((Offset(0,0) - Offset(x,y)) - Rotate2D(Angle(0,0),(UrShape(0,0) - UrShape(x,y)))) local valid = InBounds(x,y) * eq(Mask(x,y),0) * eq(Mask(0,0),0) Energy(Select(valid,e_reg,0)) end --fitting local e_fit = (Offset(0,0)- Constraints(0,0)) local valid = All(greatereq(Constraints(0,0),0)) Energy(w_fitSqrt*Select(valid, e_fit , 0.0)) Fig. 24. Image Warping energy function in Opt. local N = opt.Dim("N",0) local w_fitSqrt = Param("w_fitSqrt", float, 0) local w_regSqrt = Param("w_regSqrt", float, 1) local Offset = Unknown("Offset", float3,{N},2) local Angle = Unknown("Angle",float3,{N},3) local UrShape = Array("UrShape",float3,{N},4) local Constraints = Array("Constraints",float3,{N},5) local G = Graph("G", 6, "v0", {N}, 7, "v1", {N}, 9) UsePreconditioner(true) --vertex.xyz, rotation.xyz <- unknown --original position: vertex.xyz --user constraints --fitting local e_fit = Offset(0) - Constraints(0) local valid = greatereq(Constraints(0,0), -999999.9) Energy(Select(valid,w_fitSqrt*e_fit,0)) --regularization local ARAPCost = (Offset(G.v0) - Offset(G.v1)) - Rotate3D(Angle(G.v0),UrShape(G.v0) - UrShape(G.v1)) Energy(w_regSqrt*ARAPCost) Fig. 25. ARAP Mesh Deformation energy function in Opt. • 23 24 • DeVito, Z. et al local DEPTH_DISCONTINUITY_THRE = 0.01 local W,H = Dim("W",0), Dim("H",1) local w_p = sqrt(Param("w_p",float,0))-- Fitting weight local w_s = sqrt(Param("w_s",float,1))-- Regularization weight local w_g = sqrt(Param("w_g",float,2))-- Shading weight local f_x = Param("f_x",float,3) local f_y = Param("f_y",float,4) local u_x = Param("u_x",float,5) local u_y = Param("u_y",float,6) local L = {} for i=1,9 do L[i] = Param("L_" .. i .. "",float,6+i) end -- lighting model parameters local X = Unknown("X",opt_float, {W,H},16) -- Refined Depth local D_i = Array("D_i",opt_float, {W,H},17) -- Depth input local Im = Array("Im",opt_float, {W,H},18) -- Target Intensity local edgeMaskR = Array("edgeMaskR",uint8, {W,H},19) -- Edge mask. local edgeMaskC = Array("edgeMaskC",uint8, {W,H},20) -- Edge mask. local posX,posY = Index(0),Index(1) -- equation 8 function p(offX,offY) local d = X(offX,offY) local i = offX + posX local j = offY + posY return Vector(((i-u_x)/f_x)*d, ((j-u_y)/f_y)*d, d) end -- equation 10 function normalAt(offX, offY) local i = offX + posX local j = offY + posY local n_x = X(offX, offY - 1) * (X(offX, offY) - X(offX - 1, offY)) / f_y local n_y = X(offX - 1, offY) * (X(offX, offY) - X(offX, offY - 1)) / f_x local n_z = (n_x * (u_x - i) / f_x) + (n_y * (u_y - j) / f_y) - (X(offX-1, offY)*X(offX, offY-1) / (f_x*f_y)) local sqLength = n_x*n_x + n_y*n_y + n_z*n_z local inverseMagnitude = Select(greater(sqLength, 0.0), 1.0/sqrt(sqLength), 1.0) return inverseMagnitude * Vector(n_x, n_y, n_z) end function B(offX, offY) local normal = normalAt(offX, offY) local n_x = normal[0] local n_y = normal[1] local n_z = normal[2] return L[1] + L[2]*n_y + L[3]*n_z + L[4]*n_x + L[5]*n_x*n_y + L[6]*n_y*n_z + L[7]*(-n_x*n_x - n_y*n_y + 2*n_z*n_z) + L[8]*n_z*n_x + L[9]*(n_x*n_x-n_y*n_y) end function I(offX, offY) return Im(offX,offY)*0.5 + 0.25*(Im(offX-1,offY)+Im(offX,offY-1)) end local function DepthValid(x,y) return greater(D_i(x,y),0) end local function B_I(x,y) local bi = B(x,y) - I(x,y) local valid = DepthValid(x-1,y)*DepthValid(x,y)*DepthValid(x,y-1) return Select(InBoundsExpanded(0,0,1)*valid,bi,0) end B_I = ComputedArray("B_I", {W,H}, B_I(0,0)) -- do not include unknowns for where the depth is invalid Exclude(Not(DepthValid(0,0))) -- fitting term local E_p = X(0,0) - D_i(0,0) Energy(Select(DepthValid(0,0),w_p*E_p,0)) -- shading term local E_g_h = (B_I(0,0) - B_I(1,0))*edgeMaskR(0,0) local E_g_v = (B_I(0,0) - B_I(0,1))*edgeMaskC(0,0) Energy(Select(InBoundsExpanded(0,0,1),w_g*E_g_h,0)) Energy(Select(InBoundsExpanded(0,0,1),w_g*E_g_v,0)) -- regularization term local function Continuous(x,y) return less(abs(X(0,0) - X(x,y)),DEPTH_DISCONTINUITY_THRE) end local valid = DepthValid(0,0)*DepthValid(0,-1)*DepthValid(0,1)*DepthValid(-1,0)*DepthValid(1,0)* Continuous(0,-1)*Continuous(0,1)*Continuous(-1,0)*Continuous(1,0)*InBoundsExpanded(0,0,1) local validArray = ComputedArray("valid", {W,H},valid) valid = eq(validArray(0,0),1) local E_s = 4.0*p(0,0) - (p(-1,0) + p(0,-1) + p(1,0) + p(0,1)) Energy(Select(valid,w_s*E_s,0)) Fig. 26. Shape from Shading energy function in Opt. Opt: A Domain Specific Language for Non-linear Least Squares Optimization in Graphics and Imaging local W,H = Dim("W",0), Dim("H",1) local X = Unknown("X", float4,{W,H},0) -- unknown, initialized to base image local T = Array("T", float4,{W,H},1) -- inserted image local M = Array("M", float, {W,H},2) -- mask, excludes parts of base image UsePreconditioner(false) -- do not include unmasked pixels in the solve Exclude(Not(eq(M(0,0),0))) for x,y in Stencil { {1,0},{-1,0},{0,1},{0,-1} } do local e = (X(0,0) - X(x,y)) - (T(0,0) - T(x,y)) Energy(Select(InBounds(x,y),e,0)) end Fig. 27. Poisson Image Editing energy function in Opt. local N = Dim("N",0) local w_fitSqrt = Param("w_fitSqrt", float, 0) local w_regSqrt = Param("w_regSqrt", float, 1) local w_rotSqrt = Param("w_rotSqrt", float, 2) local X = Unknown("X", float12,{N},3) --vertex.xyz, rotation_matrix <- unknown local UrShape = Image("UrShape", float3,{N},4) --urshape: vertex.xyz local Constraints = Image("Constraints", float3,{N},5) --constraints local G = Graph("G", 6, "v0", {N}, 7, "v1", {N}, 9) UsePreconditioner(true) --really needed here local Offset = Slice(X,0,3) -- select part of unknown for position --fitting local e_fit = Offset(0) - Constraints(0) local valid = greatereq(Constraints(0)(0), -999999.9) Energy(Select(valid, w_fitSqrt*e_fit, 0)) --rot local RotMatrix = Slice(X,3,12) -- extract rotation matrix local R = RotMatrix(0) local c0 = Vector(R(0), R(3), R(6)) local c1 = Vector(R(1), R(4), R(7)) local c2 = Vector(R(2), R(5), R(8)) Energy(w_rotSqrt*Dot(c0,c1)) Energy(w_rotSqrt*Dot(c0,c2)) Energy(w_rotSqrt*Dot(c1,c2)) Energy(w_rotSqrt*(Dot(c0,c0)-1)) Energy(w_rotSqrt*(Dot(c1,c1)-1)) Energy(w_rotSqrt*(Dot(c2,c2)-1)) local regCost = (Offset(G.v1) - Offset(G.v0)) Matrix3x3Mul(RotMatrix(G.v0), (UrShape(G.v1) - UrShape(G.v0))) Energy(w_regSqrt*regCost) Fig. 28. Embedded Mesh Deformation energy function in Opt. • 25 26 • DeVito, Z. et al local N = Dim("N",0) local local local local local w_fitSqrt = Param("w_fit", float, 0) w_regSqrt = Param("w_reg", float, 1) X = Unknown("X", float3,{N},2) A = Array("A", float3,{N},3) G = Graph("G", 4, "v0", {N}, 5, --current vertex "v1", {N}, 7, --neighboring vertex "v2", {N}, 9, --prev neighboring vertex "v3", {N}, 11) --next neighboring vertex UsePreconditioner(true) function cot(v0, v1) local adotb = Dot3(v0, v1) local disc = Dot3(v0, v0)*Dot3(v1, v1) - adotb*adotb disc = Select(greater(disc, 0.0), disc, 0.0001) return Dot3(v0, v1) / Sqrt(disc) end -- fit energy Energy(w_fitSqrt*(X(0) - A(0))) local local local local a b c d = = = = normalize(X(G.v0) normalize(X(G.v1) normalize(X(G.v0) normalize(X(G.v1) - X(G.v2)) --float3 X(G.v2)) --float3 X(G.v3)) --float3 X(G.v3)) --float3 --cotangent laplacian; Meyer et al. 03 local w = 0.5*(cot(a,b) + cot(c,d)) w = Sqrt(Select(greater(w, 0.0), w, 0.0001)) Energy(w_regSqrt*w*(X(G.v1) - X(G.v0))) Fig. 29. Cotangent Mesh Smoothing energy function in Opt. local local local local local local local local local W,H = Dim("W",0), Dim("H",1) w_fitSqrt = Param("w_fit", float, 0) w_regSqrt = Param("w_reg", float, 1) X = Unknown("X", float2,{W,H},2) I = Array("I",float,{W,H},3) I_hat_im = Array("I_hat",float,{W,H},4) I_hat_dx = Array("I_hat_dx",float,{W,H},5) I_hat_dy = Array("I_hat_dy",float,{W,H},6) I_hat = SampledImage(I_hat_im,I_hat_dx,I_hat_dy) local i,j = Index(0), Index(1) UsePreconditioner(false) -- fitting local e_fit = w_fitSqrt*(I(0,0) - I_hat(i + X(0,0,0),j + X(0,0,1))) Energy(e_fit) -- regularization for nx,ny in Stencil { {1,0}, {-1,0}, {0,1}, {0,-1} } do local e_reg = w_regSqrt*(X(0,0) - X(nx,ny)) Energy(Select(InBounds(nx,ny),e_reg,0)) end Fig. 30. Optical Flow energy function in Opt. Opt: A Domain Specific Language for Non-linear Least Squares Optimization in Graphics and Imaging local W,H,D = Dim("W",0), Dim("H",1), Dim("D",2) local Offset = Unknown("Offset", float3,{W,H,D},0) local Angle = Unknown("Angle",float3,{W,H,D},1) local UrShape = Array("UrShape",float3,{W,H,D},2) local Constraints = Array("Constraints",float3,{W,H,D},3) local w_fitSqrt = Param("w_fitSqrt", float, 4) local w_regSqrt = Param("w_regSqrt", float, 5) UsePreconditioner(true) --vertex.xyz, rotation.xyz <- unknown --original position: vertex.xyz --user constraints --fitting local e_fit = Offset(0,0,0) - Constraints(0,0,0) local valid = greatereq(Constraints(0,0,0)(0), -999999.9) Energy(Select(valid,w_fitSqrt*e_fit,0)) for i,j,k in Stencil { {1,0,0}, {-1,0,0}, {0,1,0}, {0,-1,0}, {0,0,1}, {0,0,-1}} do local ARAPCost = (Offset(0,0,0) - Offset(i,j,k)) - Rotate3D(Angle(0,0,0),UrShape(0,0,0) - UrShape(i,j,k)) local ARAPCostF = Select(InBounds(0,0,0), Select(InBounds(i,j,k), ARAPCost, 0.0), 0.0) Energy(w_regSqrt*ARAPCostF) end Fig. 31. Volumetric Mesh Deformation energy function in Opt. local N = Dim("N",0) local w_fitSqrt local w_regSqrt local w_confSqrt local Offset local Angle local RobustWeights local UrShape local Constraints local ConstraintNormals local G UsePreconditioner(true) = = = = = = = = = = Param("w_fitSqrt", float, 0) Param("w_regSqrt", float, 1) 0.1 Unknown("Offset", float3,{N},2) --vertex.xyz, rotation.xyz <- unknown Unknown("Angle", float3,{N},3) --vertex.xyz, rotation.xyz <- unknown Unknown("RobustWeights", float,{N},4) Array("UrShape", float3, {N},5) --urshape: vertex.xyz Array("Constraints", float3,{N},6) --constraints Array("ConstraintNormals", float3,{N},7) Graph("G", 8, "v0", {N}, 9, "v1", {N}, 10) local robustWeight = RobustWeights(0) --fitting local e_fit = robustWeight*ConstraintNormals(0):dot(Offset(0) - Constraints(0)) local validConstraint = greatereq(Constraints(0), -999999.9) Energy(w_fitSqrt*Select(validConstraint, e_fit, 0.0)) --RobustWeight Penalty local e_conf = 1-(robustWeight*robustWeight) e_conf = Select(validConstraint, e_conf, 0.0) Energy(w_confSqrt*e_conf) --regularization local ARAPCost = (Offset(G.v0) - Offset(G.v1)) - Rotate3D(Angle(G.v0),UrShape(G.v0) - UrShape(G.v1)) Energy(w_regSqrt*ARAPCost) Fig. 32. Robust Mesh Deformation energy function in Opt. • 27 28 • W,H = local local local local local local local local DeVito, Z. et al opt.Dim("W",0), opt.Dim("H",1) w_fitSqrt = Param("w_fitSqrt", float, 0) w_regSqrtAlbedo = Param("w_regSqrtAlbedo", float, 1) w_regSqrtShading = Param("w_regSqrtShading", float, 2) pNorm = Param("pNorm", float, 4) r = Unknown("r", float3,{W,H},5) r_const = Array("r_const", float3,{W,H},5) -- A constant view of the unknown i = Array("i", float3,{W,H},6) s = Unknown("s", float,{W,H},8) -- reg Albedo for x,y in Stencil { {1,0}, {-1,0}, {0,1}, {0,-1} } do local diff = (r(0,0) - r(x,y)) local diff_const = (r_const(0,0) - r_const(x,y)) -- The helper L_p function takes diff_const, raises it's length to the (p-2) power, -- and stores it in a computed array, so its value remains constant during the nonlinear iteration, -- then multiplies it with diff and returns local laplacianCost = L_p(diff, diff_const, pNorm, {W,H}) local laplacianCostF = Select(InBounds(0,0),Select(InBounds(x,y), laplacianCost,0),0) Energy(w_regSqrtAlbedo*laplacianCostF) end -- reg Shading for x,y in Stencil { {1,0}, {-1,0}, {0,1}, {0,-1} } do local diff = (s(0,0) - s(x,y)) local laplacianCostF = Select(InBounds(0,0),Select(InBounds(x,y), diff,0),0) Energy(w_regSqrtShading*laplacianCostF) end -- fit local fittingCost = r(0,0)+s(0,0)-i(0,0) Energy(w_fitSqrt*fittingCost) Fig. 33. Intrinsic Image Decomposition energy function in Opt. 

6cs.PL

DEPTH OF F -SINGULARITIES AND BASE CHANGE OF RELATIVE CANONICAL SHEAVES arXiv:1207.1910v3 [math.AG] 4 Feb 2013 ZSOLT PATAKFALVI AND KARL SCHWEDE Abstract. For a characteristic p > 0 variety X with controlled F -singularities, we state conditions which imply that a divisorial sheaf is Cohen-Macaulay or at least has depth ≥ 3 at certain points. This mirrors results of Kollár for varieties in characteristic zero. As an application, we show that relative canonical sheaves are compatible with arbitrary base change for certain families with sharply F -pure fibers. 1. Introduction In the paper [Kol11b], Kollár proved that sheaves OX (−D) satisfy strong depth conditions if D is locally Q-linearly equivalent to a divisor ∆ such that (X, ∆) is SLC or KLT. These results generalized [Ale08, Lemma 3.2], [Fuj09] and [KM98, Corollary 5.25]. Because depth conditions can be interpreted as vanishing of local cohomology, these results were described as a local version of the Kawamata-Viehweg vanishing theorem. In this paper, we obtain characteristic p > 0 analogs of the main results of [Kol11b]. This is particularly interesting because the (global) Kawamata-Viehweg vanishing theorem is false in positive characteristic [Ray78]. We replace the KLT and SLC conditions by strongly F regular and sharply F -pure singularities respectively (such characteristic p > 0 singularity classes are known as F -singularities). For the convenience of the reader, we recall that by [HW02, MS12] ◦ KLT pairs correspond philosophically to strongly F -regular pairs, and ◦ SLC pairs correspond philosophically to sharply F -pure pairs. Similar to [Kol11b], we can apply our results on depth to prove base change for relative canonical sheaves. (A special case of ) Corollary 4.13. (cf. [Kol11b, 4.3]) Let f : X −→ Y be a flat morphism of finite type with S2 , G1 equidimensional fibers to a smooth variety and let ∆ ≥ 0 be a Qdivisor on X avoiding all the codimension zero and the singular codimension one points of the fibers. Further suppose that KX + ∆ is Q-Cartier, p 6 | ind(KX + ∆) and (Xy , ∆y ) is sharply F -pure for every y ∈ Y . Then ωX/Y is flat over Y and compatible with arbitrary base change. Corollary 4.13 is hoped to be useful in constructing a moduli space for varieties of general type in positive characteristics. See [Pat10] for further explanation, and also for examples for which the above compatibility does not hold. We also remark here that ωX/Y behaves surprisingly well with respect to base-change. It obeys base-change for example when the fibers are Cohen-Macaualay [Con00, Theorem 3.6.1]. In particular, this pertains to families of normal [m] surfaces. In contrast, the higher reflexive powers, ωX/Y for m > 1, are not compatible with base change in the surface case [HK10, Section 14.A]. Similar differences between canonical and pluricanonical sheaves have been observed earlier [Kol87, page 2], [Kol90, Remark 4.4]. 2010 Mathematics Subject Classification. 13A35, 14J10, 14J17, 14F18, 13C14, 13C15. Key words and phrases. depth, Cohen-Macaulay, F -singularities, base change, relative canonical sheaf. The second author was partially supported by the NSF grant DMS #1064485. 1 2 ZSOLT PATAKFALVI AND KARL SCHWEDE The technical result on depth used to prove Corollary 4.13 is as follows. It is a characteristic p > 0 version of [Kol11b, Theorem 3(1)], also compare with [Ale08, Lemma 3.2], [Fuj09, Theorem 4.21], [AH11, Theorem 1.5] and [Kov11, Theorem 1.2, 1.5]. Theorem 3.8. (cf. [Kol11b, Theorem 3(1)]) Suppose that R is local, S2 and G1 and that 0 ≤ ∆ is an R-divisor on X = Spec R with no common components with the singular locus of X and such that (X, ∆) is sharply F -pure. Set x ∈ X to be the closed point and assume that x is not an F -pure center of (X, ∆). Suppose that 0 ≤ ∆′ ≤ ∆ is another R-divisor and that r∆′ is integral for some r > 0 relatively prime to p. Further assume that M is any rank-1 reflexive subsheaf of K(X) such that M (−r) ∼ = OX (r∆′ ) (here (·) denotes reflexive power). Then depthx M ≥ min{3, codimX x}. Another interesting depth statement, again completely analogous to a theorem of Kollár is below. In the introduction we phrase it in the language of Frobenius splittings [BK05], but in the text it is phrased slightly more generally. Theorem 3.6, Remark 3.7. (cf. [Kol11b, Theorem 3(2)]) Suppose that (R, m) is an S2 local ring with Frobenius splitting ϕ : F∗e R −→ R which is not compatibly split with m. Additionally suppose that Z is any union of compatibly ϕ-split subvarieties of X = Spec R such that no irreducible component of Z coincides with an irreducible component of X. Suppose that IZ ⊆ R is the ideal defining Z, then depthm IZ ≥ min{3, 1 + codimZ x}. The other main statement on depth we obtain, Theorem 3.1, asserts that if (X, ∆) is strongly F -regular and (pe − 1)D is an integral divisor linearly equivalent to (pe − 1)∆, then OX (−D) is Cohen-Macaulay. Compare with [Kol11b, Theorem 2]. Remark 1.1. One should also compare the above results on depth, as well as the related characteristic zero results, to [AE05, Theorem 4.8(vi)] where Aberbach and Enescu showed that the depth of an F -pure ring R is always ≥ than the dimension of the minimal F -pure center (ie, of the dimension of R modulo the splitting prime, which we know is equal to the s-dimension of [AE05] by [BST11]). Acknowledgements: The authors began working on this project at the workshop ACC for minimal log discrepancies and termination of flips held at the American Institute of Mathematics and organized by Tommaso de Fernex and Christopher Hacon. The authors would also like to thank Florian Enescu for valuable discussions as well as thank the referee, János Kollár and Sándor Kovács for many useful comments on a previous draft. 2. Preliminaries on F -singularities Notation 2.1. Throughout this paper, all schemes are Noetherian, separated, of equal characteristic p > 0 and F -finite.1 Note that any such scheme X is automatically locally excellent by [Kun76] and also has a dualizing complex by [Gab04]. In particular, we are implicitly assuming all schemes are locally excellent and possess dualizing complexes. Little will be lost to the reader if he or she considers only schemes that are essentially of finite type over a perfect field. 1Meaning the Frobenius morphism is a finite morphism. DEPTH OF F -SINGULARITIES AND BASE CHANGE OF RELATIVE CANONICAL SHEAVES 3 We remind the reader of some special divisors on non-normal schemes. Definition 2.2 (Divisors on non-normal schemes). We follow the notation of [K+92, Section 16]. For an S2 reduced local ring R, set X = Spec R. We define a W-divisor (or Weil divisor ) to be a formal sum of codimension one subsets of X whose generic points are not singular points of X. This has the same data as divisors on the regular locus of X or as rank-1 S2 submodules M of K(R) (the total ring of fractions of R) such that Mη = Rη as a subset of K(R), for every codimension 1 singular point η of X. Later in the paper, we will need to instead work with the more general notion of Weil divisorial sheaves WSh(X), rank one reflexive subsheaves of K(X) that are invertible in codimension 1. In the non-local setting, such divisors are simply formal sums of irreducible subschemes that satisfy this definition locally. We now set WDivQ (X) := WDiv(X)⊗Z Q and WDivR (X) := WDiv(X) ⊗Z R. Note we have containments: WDiv(X) ⊆ WDivQ (X) ⊆ WDivR (X). One can also form WShQ (X) := WSh(X) ⊗Z Q and WShZ(p) (X) := WShZ(p) (X) ⊗Z Z(p) similarly, but the natural maps WSh(X) − → WShZ(p) (X) − → WShQ (X) are not necessarily P injective [K+92, Section 16]. Given i ai Di = ∆ ∈ WDivR (X), we use ⌈∆⌉ to denote P ⌈a ⌉D (such roundings are not necessarily well defined for WShZ(p) (X) or WShQ (x)). i i i Finally, given D ∈ WDiv(X), we use OX (D) (or R(D)) to denote the corresponding subsheaf of K (X) (or of K(R)) in the usual way. Note that D is effective if and only if OX (D) ⊇ OX . Now we move away from divisors. Suppose that R is a ring of characteristic p > 0. Following [Sch11, Bli09], we say a Cartier subalgebra C is a graded subring of the graded ring M HomR (F∗e R, R) =: C R e≥0 where multiplication is done by Frobenius twisted composition2 such that the zeroth graded piece [C ]0 = HomR (R, R) ∼ = R. We note that even though we call C a Cartier subalgebra, it is not an R-algebra because R ∼ = [C ]0 is not necessarily central. Example 2.3 (Cartier subalgebra associated to a divisor). Given an S2 and G1 ring R, set X = Spec R and assume that 0 ≤ ∆ ∈ WDivR (X) on X (for example, if R is normal, ∆ is simply an R-divisor). We can form the Cartier subalgebra C ∆ where [C ∆ ]e := HomR (F∗e R(⌈(pe − 1)∆⌉), R) ⊆ HomR (F∗e R, R). Example 2.4 (Cartier subalgebra generated by a map). Suppose that C R is as above and ϕ ∈ [C R ]e for some e > 0. Then we can form the Cartier subalgebra Rhϕi generated by R = [C ]0 and ϕ. Explicitly, this is the direct sum R ⊕ (ϕ · (F∗e R)) ⊕ (ϕ2 · (F∗2e R)) ⊕ · · · . Now we define sharply F -pure pairs and F -pure centers. Definition 2.5 (Sharply F -pure pairs). If C is a Cartier subalgebra on R, then we say that the pair (R, C ) is sharply F -pure if there exists some ϕ ∈ [C ]e for some e ≥ 1 such that ϕ(F∗e R) = R. In particular, if (R, ∆) is a pair as in Example 2.3, then we say that (R, ∆) is sharply F -pure if the associated (R, C ∆ ) is sharply F -pure. If (R, C R ) is sharply F -pure, then we simply say that R is F -pure. Definition 2.6 (Compatible ideals and F -pure centers). If (R, C ) is a pair as above, then an ideal I ⊆ R is called C -compatible if ϕ(F∗e I) ⊆ I for all ϕ ∈ [C ]e and all e ≥ 0. In the case that C = Rhϕi, we will sometimes simply say that I is ϕ-compatible. 2If ϕ ∈ [C R ] and ψ ∈ [C R ] , then ϕ · ψ = ϕ ◦ (F e ψ). See the aforementioned sources for more details. e d ∗ 4 ZSOLT PATAKFALVI AND KARL SCHWEDE An irreducible closed set W = V (Q) ⊆ Spec R = X, for some Q ∈ Spec R, is called an F -pure center if the following two conditions hold: (a) The localization (RQ , CQ ) is sharply F -pure, and (b) For every for e ≥ 0 and ϕ ∈ [C ]e , we have ϕ(F∗e Q) ⊆ Q (in other words, if Q is C -compatible). Likewise we say that W is an F -pure center of (R, ∆) if it is an F -pure center of (R, C ∆ ) where C ∆ is associated to ∆ as in Example 2.3. We also define strongly F -regular pairs. Definition 2.7 (Strongly F -regular pairs). If R is a local ring, a pair (R, C ) is called strongly F -regular if the only proper C -compatible ideals of R are 0 and R itself. If R is not local, then we say (R, C ) is strongly F -regular if every localization is. A pair (R, ∆) is strongly F -regular if (R, C ∆ ) is strongly F -regular. Remark 2.8. Given a pair (X, ∆), all of the above definitions generalize to the non-affine setting by requiring them to hold at each stalk. The notion of Cartier subalgebras is somewhat more subtle in the non-affine setting however (but we will not need such generalities). We recall some facts about compatible ideals and F -pure centers. Lemma 2.9. Suppose that (R, C ) is a pair and I ⊆ J ⊆ R are ideals. (i) The set of C -compatible ideals are closed under sum and intersection. (ii) A prime ideal Q is C -compatible if and only if QRQ is CQ -compatible. (iii) If C ′ ⊆ C are Cartier subalgebras and I is C -compatible, then I is C ′ -compatible. (iv) Given ϕ ∈ HomR (F∗e R, R), we have ϕ(F∗e J) ⊆ J if and only if J is Rhϕi-compatible. (v) Suppose that ϕ : F∗e R −→ R is surjective. Some Q ∈ Spec R is ϕ-compatible if and only if it is ϕn -compatible where (n−1)e ϕn := ϕ ◦ (F∗e ϕ) ◦ · · · ◦ (F∗ {z | n-times ϕ) } (vi) If I is ϕ-compatible, then there exists a map ϕ/I : F∗e (R/I) −→ (R/I) such that the following diagram commutes: F∗e R ϕ  F∗e (R/I) ϕ/I // R  // (R/I). Furthermore, J ⊇ I is ϕ-compatible if and only if J/I is ϕ/I-compatible. (This statement can also be done with Cartier subalgebras, but we will not need it). (vii) (R, C ) is strongly F -regular if and only if for every c ∈ R \ {minimal primes}, there exists a ϕ ∈ [C ]e for e > 0, in fact one may take e to be any larger multiple, such that ϕ(F∗e c) = 1. Proof. (v) follows from the argument of [Sch10, Proposition 4.1]. (vii) can be found in this  generality in [Sch11, Proposition 3.23]. The rest are obvious. Our next goal is to give an example of a ϕ-compatible ideal that will be crucial in later sections. The main idea is that Frobenius maps and Frobenius splittings induce maps on local cohomology. Those induced maps can then be thought of as acting directly and explicitly on Čech classes. For the convenience of the reader not already familiar with this construction, recall that if X = Spec R and U = Spec R \ {m}, then for any coherent OX -module M , we DEPTH OF F -SINGULARITIES AND BASE CHANGE OF RELATIVE CANONICAL SHEAVES 5 have Hmi (M ) = H i−1 (U, M ) for i > 1 and also that Hm1 (M ) = H 0 (U, M )/ image(H 0 (X, M )). We can then use the Čech cohomology description of sheaf cohomology to define Hmi (M ). For a more thorough description of local cohomology by the Čech complex, see for example [BH93, Section 3.5]. We now consider Frobenius action on local cohomology. The Frobenius map R − → F∗ R → Hmi (F∗ R) ∼ yields Ψ : Hmi (R) − = Hmi (R). Given a Čech class [z] = [. . . , zj , . . . ] ∈ Hmi (R), we e e have Ψ([z]) ∈ Hmi (F∗ R). But certainly Ψ([z]) = F∗e [z]p = F∗e [. . . , zjp , . . . ] is identified with raising the entries of [z] to the pe th power. Now we do the same computation with a Frobenius splitting. Suppose that ϕ : F∗e R − →R ϕ i e i → Hm (R) induced by ϕ. is an R-linear Frobenius splitting, and so we have a map Hm (F∗ R) − Certainly ϕ(F∗e [. . . , yj , . . . ]) = [. . . , ϕ(F∗e yj ), . . . ]. But now observe that for any [z] ∈ Hmi (R) and r ∈ R we have that e e ϕ(F∗e (r · [z]p )) = ϕ(F∗e [. . . , rzjp , . . . ]) = [. . . , ϕ(F∗e r)zj , . . . ] = ϕ(F∗e r)[z]. Lemma 2.10. (cf. [EH08, Theorem 4.1]) Suppose that (R, m) is a local ring. Then AnnR Hmi (R) is compatible with every splitting3 ϕ : F∗e R −→ R of Frobenius R −→ F∗e R. Proof. We have the following composition 1 ϕ // F e R ∗ R ✤ // F e 1 ∗ // R // 1. ✤ Now suppose that r ∈ AnnR Hmi (R). Then choose [z] ∈ Hmi (R). We want to show that ϕ(F∗e r).[z] = 0. Now, it follows from the Čech cohomology description of local cohomology, and ϕ’s action on it, that e 0 = ϕ(F∗e 0) = ϕ(F∗e (r.[z]p )) = ϕ(F∗e r).[z] which completes the proof.  We also recall the following fact. We include the proof because the method will be generalized later. Lemma 2.11. If (R, C ) is strongly F -regular, then R is normal and Cohen-Macaulay. Proof. We first note that by Lemma 2.9(vii), the strong F -regularity hypothesis implies that ψ → R is an isomorphism for some there exists a Frobenius splitting ψ such that R − → F∗e R − e > 0. It then easily follows that R must be reduced since if not, the map R − → F∗e R is not injective. Normality follows since the conductor ideal is compatible with every ϕ ∈ C R by the argument of [BK05, Proposition 1.2.5]. For the Cohen-Macaulay condition, by working locally we assume that (R, m) is a local domain. By local duality [Har66, Chapter V, Theorem 6.2], q q each Hmi (R) is Matlis dual to hi−dim R ωR for some normalized dualizing complex ωR . Since an element c ∈ R annihilates a finitely generated R-module if and only if c annihilates the Matlis dual of a module, it follows that there exists 0 6= c ∈ R such that c · Hmi (R) = 0 for all i < dim R. Lemma 2.9(vii) then implies that there exists ϕ ∈ [C ]e such that the composition F e (·c) R− → F∗e R −−∗−−→ F∗e R − →R is an isomorphism. Taking local cohomology for i < dim R gives us an isomorphism: F e (·c) Hmi (R) − → Hmi (F∗e R) −−∗−−→ Hmi (F∗e R) − → Hmi (R) where the middle map is the zero map. Thus Hmi (R) = 0 which completes the proof. 3A splitting is simply a map ϕ : F e R − → R that sends F∗e 1 to 1. Splittings are necessarily surjective. ∗  6 ZSOLT PATAKFALVI AND KARL SCHWEDE We also state a generalization of [SS10, Theorem 4.3], similar computations were done in [MS12]. Lemma 2.12. Suppose that X = Spec R is S2 and G1 and sharply F -pure. Then there exists an element 0 ≤ ∆ ∈ WDivQ (X) such that (pe − 1)(KX + ∆) is Q-Cartier and (X, ∆) is sharply F -pure. Proof. A surjective map ϕ ∈ HomR (F∗e R, R) ∼ = H 0 (X, F∗e OX ((1 − pe )KX )) induces an effec4 tive Weil divisorial sheaf Γϕ by [Har94, Proposition 2.9] such that (pe − 1)KX + Γϕ ∼ 0. We would like to show that Γϕ can be identified with an element of WDiv(X). At the singular height one points η of X, OX,η is already Gorenstein. Thus we can consider the map Φη which generates HomOX,η (F∗e OX,η , OX,η ) as an F∗e OX,η -module. Set m to be the maximal ideal of OX,η and notice that m is the conductor ideal since F -pure rings are seminormal [HR76] and in particular the conductor is radical. But then Φ(F∗e m) ⊆ m by the proof of → OX,η is equal to Φ(F∗e (r · )) for [BK05, Proposition 1.2.5]. Now, we know ϕη : F∗e OX,η − some r ∈ OX,η . We want to show that r is a unit, which would prove that Γϕ is trivial at η. Since ϕη is surjective, we see that r ∈ / m and thus r is a unit. This implies that the Weil divisorial sheaf (Γϕ )η coincides with OX,η and thus Γϕ ∈ WDiv(X) as desired. Finally, set ∆ = pe1−1 Γϕ .  We conclude by recalling a well known lemma on the height of annihilators of local cohomology modules. However, because we lack a reference, we provide a proof. Lemma 2.13. Suppose that (R, m) is a local ring and suppose that M is a
finitely generated R-module which is Sn in the sense5 of [BH93]. Set Yi = V AnnR (Hmi (M )) ⊆ Spec R = X. Suppose that d is the minimum over the dimensions of the components of Supp M . Then dim Yi ≤ i − n for i < d. Remark 2.14. Note Yi may not be the same as Supp Hmi (M ) since Hmi (M ) is not finitely generated. Proof. Set ωX to be a normalized dualizing complex on X (recall that all our rings are excellent and possess dualizing complexes). By local duality in the form of [Har66, Chapter q V, Theorem 6.2], it is equivalent to prove that dim Supp h−i R HomR (M, ωR ) ≤ i−n. Suppose q this is false, and thus that W ⊆ Supp h−i R HomR (M, ωR ) is an irreducible component of dimension t > i − n for some i < d. Set γ to be the generic point of W (which we also view as a prime ideal). By localizing at γ, we see that q (h−i R HomR (M, ωR ))γ = h−i R HomRγ (Mγ , ωRγ [t])) = h−i+t R HomRγ (Mγ , ωRγ )) q q q is supported at a point. The shift by [t] is necessary to keep the dualizing complexes normalized. Thus Hγi−t (Mγ ) 6= 0 by local duality again. Now, i − t < n. Also observe that dim Mγ ≥ d − t (this is why the d is necessary since we do not know what component of Supp M we will be restricting to). Since Mγ is still Sn , we see that Hγj (Mγ ) = 0 for j < min(n, dim Mγ ). But then min(n, dim Mγ ) ≥ min(n, d − t) > i − t since n > i − t and d > i. Setting j = i − t we obtain a contradiction.  4in the terminology of [K+92, Section 16] 5In other words, depth M ≥ min(n, dim M ) for all z ∈ Spec R. Note that here we use dim M not dim R . z z z z DEPTH OF F -SINGULARITIES AND BASE CHANGE OF RELATIVE CANONICAL SHEAVES 7 3. Depth and F -singularities Our goal in this section is to prove several results on the depths of sheaves on schemes with controlled F -singularities. First we prove our result for pairs (R, ∆) which are strongly F -regular, this is the simplest case. Theorem 3.1. (cf. [Kol11b, Theorem 2]) Suppose that (R, m) is local and that (X = Spec R, ∆) is strongly F -regular. Further suppose that 0 ≤ ∆′ ≤ ∆ is such that rD ∼ r∆′ for some integral divisor D and some integer r > 0 realtively prime to p. Then OX (−D) is CohenMacaulay. Proof. By possibly multiplying r with an integer, we may assume that r = pe − 1. Choose, using Lemma 2.13, 0 6= c ∈ R \ {minimal primes} such that c · Hmi (OX (−D)) = 0 for all ′ i < dim R. Note that since ∆′ ≤ ∆, C ∆ ⊆ C ∆ and then (X, ∆′ ) is strongly F -regular as well by Lemma 2.9(iii). Therefore, by Lemma 2.9(vii) there exists an e > 0 and a splitting ϕ such that the composition F e (·c) ϕ OX − → F∗e OX − → F∗e OX ((pe − 1)∆′ ) −−∗−−→ F∗e OX ((pe − 1)∆′ ) − → OX is an isomorphism. By replacing e by a multiple if necessary, we may assume that this e > 0 also satisfies the condition from the hypothesis. Twisting by OX (−D), reflexifying, and applying Hmi ( ) we obtain the following composition which is also an isomorphism. Hmi (OX (−D)) − → Hmi (F∗e OX ((pe − 1)(∆′ − D) − D)) F e (·c) −−∗−−→ Hmi (F∗e OX ((pe − 1)(∆′ − D) − D)) ϕ − → Hmi (OX (−D)) However, the map labeled F∗e (·c) is the zero map for i < dim X since Hmi (F∗e OX ((pe − 1)(∆′ − D) − D)) = Hmi (F∗e OX (−D)). Thus Hmi (OX (−D)) = 0 as desired.  Remark 3.2. If one assumes that (X, ∆) is purely F -regular (an analog of purely log terminal [Tak08]), the same result holds by the same proof. The point is that we may take c annihilating Hmi (R) and which simultaneously doesn’t vanish along the support of any component of ∆′ . Corollary 3.3. (cf. [KM98, Corollary 5.25]) If (R, m) is local and (X = Spec R, ∆) is strongly F -regular, then for every Q-Cartier integral divisor D, OX (−D) is Cohen-Macaulay. Proof. If the index of D is not divisible by p then the statement is a special case of Theorem 3.1 by setting ∆′ := 0. Hence assume that the index m of D is divisible by p. Choose then an effective divisor E linearly equivalent to D and set r := ms + 1, ∆′ := 1r E for some integer s ≫ 0. In this situation r is relatively prime to p and   1 E = r∆′ . rD = (ms + 1)D ∼ D ∼ E = r r
Furthermore, for s ≫ 0, X, ∆ + 1r E is strongly F -regular. Hence, we may apply Theorem 3.1 for ∆ replaced by ∆+ 1r E and the above choices of r, D and ∆′ . This concludes our proof.  Before moving on to the sharply F -pure pairs, we need a Lemma on the existence of certain Frobenius splittings. Lemma 3.4. Suppose that (R, C ) is any pair where C is a Cartier subalgebra on a local ring (R, m). Suppose m is not C -compatible. Then there exists some Frobenius splitting ϕ : F∗e R −→ R such that ϕ(F∗e m) = R ) m. 8 ZSOLT PATAKFALVI AND KARL SCHWEDE Proof. There obviously exists a map in [C ]e , ψ : F∗e R − → R, such that ψ(F∗e m) 6⊆ m. It follows e that ψ(F∗ m) = R. We have two cases: Case 1: Suppose that there is a unit d ∈ R \ m such that ψ(F∗e d) = u ∈ / m. Thus e e ψ(F∗e (u−p d)) = 1. Consider the map ϕ(F∗e ) = ψ(F∗e ((u−p d)· )) and notice that ϕ(F∗e 1) = 1 which shows that ϕ is a splitting. Also notice that m is not ϕ-compatible since ϕ is a unit multiple of ψ. Thus we have found our ϕ. Case 2: Since we have already handled Case 1, we may assume that ψ(F∗e d) ∈ m for all units d ∈ R. Choose c ∈ m such that ψ(F∗e c) = 1. Now then ψ(F∗e 1) ∈ m since 1 is a unit. Thus ψ(F∗e (c + 1)) = ψ(F∗e c) + ψ(F∗e 1) ∈ 1 + m 6⊆ m is a unit. But this is a contradiction since c is assumed not be a unit and so c+ 1 is a unit.  Remark 3.5. If, in addition to the hypotheses of Lemma 3.4, R is S2 and G1 , then by applying the argument of Lemma 2.12 to the splitting ϕ constructed in the proof of Lemma 3.4, we obtain a Q-divisor ∆ on X = Spec R such that ◦ (pe − 1)(KX + ∆) is Cartier. ◦ (X, ∆) is sharply F -pure. ◦ x = V (m) is not an F -pure center of (X, ∆). The second two statements follow since ϕ ∈ [C ∆ ]e . Theorem 3.6. (cf. [Kol11b, Theorem 3(2)]) Suppose that (R, m) is a local S2 -ring, and C is a Cartier-subalgebra on R such that (R, C ) is sharply F -pure and V (m) is not an F -pure center. Additionally suppose that Z ( X = Spec R is any union of F -pure centers of (R, C ). We also assume that Z and X have no common irreducible components. If IZ is the (radical) ideal defining Z, then depthm IZ ≥ min{3, 1 + codimZ x}. → R Proof. Since shrinking C is harmless, we set C = Rhϕi for some splitting ϕ : F∗e R − which by Lemma 3.4 is not compatible with the origin V (m). Indeed, by Lemma 2.9(iii) we can only increase the number of centers when restricting a Cartier subalgebra. We have the long exact sequence: · · · Hm1 (IZ ) − → Hm1 (R) − → Hm1 (R/IZ ) − → Hm2 (IZ ) − → Hm2 (R) − → Hm2 (R/IZ ) − → Hm3 (IZ ) − → ··· and recall we are trying to show that Hmi (IZ ) = 0 for i < min{3, 1 + codimZ x}. Since R is S2 , Hm1 (R) = Hm0 (R) = 0. Since Z 6= V (m) and Z is reduced, Hm0 (OZ ) = 0 and so Hm1 (IZ ) = 0. Thus depthm IZ ≥ 2. It is now sufficient to prove the case when codimZ x ≥ 2. This implies that dim X ≥ 3 (since Z and X have no common components). Furthermore, we can assume that every component of Z has dimension at least 2. Indeed, suppose that Z1 is an irreducible component of Z such that dim Z1 = 1. If Z2 is the union of the other components of Z and Z2 6= ∅, then Z1 ∩ Z2 = x (for dimension reasons and since we working in a local ring). But this implies that x is an F -pure center since intersections of F -pure centers are unions of F -pure centers by Lemma 2.9(i). Thus we can assume that Z1 = Z is 1-dimensional. But then codimZ x = 1, which contradictions our assumption. By Lemma 2.10, we know that AnnR (Hm2 (R)) is compatible with p (R, C ) = (R, Rhϕi). 2 However, if Hm (R) 6= 0, then since R is S2 and of dimension ≥ 3, AnnR (Hm2 (R)) = m by Lemma 2.13. But AnnR (Hm2 (R)) is radical (since C is sharply F -pure) so AnnR (Hm2 (R)) = m. But V (m) is not an F -pure center, this is a contradiction. We conclude that Hm2 (R) = 0. Now we come to Hm1 (R/IZ ). Again, since R/IZ is reduced, R/IZ is S1. Furthermore, since Z has no 1-dimensional components we can apply Lemma 2.13 to conclude that AnnR Hm1 (R/IZ ) can either be m-primary or R. Suppose it is m-primary. Since ϕ|Z is still a splitting, it follows that AnnR/IZ Hm1 (R/IZ ) is ϕ|Z -compatible and also radical and so equal to m/IZ . But then DEPTH OF F -SINGULARITIES AND BASE CHANGE OF RELATIVE CANONICAL SHEAVES 9 m is ϕ-compatible by basic facts about Frobenius splitting or by Lemma 2.9(vi). We conclude that Hm1 (R/IZ ) = 0. This forces Hm2 (IZ ) to be zero and completes the proof.  Remark 3.7. Another way to state a special case of Theorem 3.6 using the language of Frobenius splittings is as follows: Suppose that (R, m) is an S2 local ring with Frobenius splitting ϕ : F∗e R −→ R which is not compatibly split with m. Additionally suppose that Z is any union of compatibly split subvarieties of X = Spec R such that no irreducible component of Z coincides with an irreducible component of X. Suppose that IZ ⊆ R is the ideal defining Z, then depthm IZ ≥ min{3, 1 + codimZ x}. Now we come to our main technical result on depth. It is a characteristic p > 0 version of [Kol11b, Theorem 3(1)] but also compare with [Ale08, Lemma 3.2], [Fuj09, Theorem 4.21], [AH11, Theorem 1.5] and [Kov11, Theorem 1.2, 1.5]. Theorem 3.8. (cf. [Kol11b, Theorem 3(1)]) Suppose that R is local, S2 and G1 , X = Spec R and that 0 ≤ ∆ ∈ WDivR (X) is such that (R, ∆) is sharply F -pure. Set x ∈ X to be the closed point and assume that x is not an F -pure center of (R, ∆). Suppose that 0 ≤ ∆′ ≤ ∆ is another element of WDivR (X) and that r∆′ is integral for some r > 0 relatively prime to p. Further assume that M is any rank-1 (along each component of X) reflexive coherent subsheaf of K(X) such that M (−r) ∼ = OX (r∆′ ) (here (·) denotes reflexive power).6 Then depthx M ≥ min{3, codimX x} = min{3, dim R}. Proof. First observe that it is harmless to assume that dim R ≥ 3 since otherwise the statement is trivial since M is reflexive and thus S2 by [Har94, Theorem 1.9]. We may also assume that M ⊆ OX is an ideal sheaf since we are working locally. We thus identify M with an ideal of R also denoted by M . Finally, by replacing r by a power if necessary, we may assume that r = pe − 1 for some e > 0. Using Lemma 3.4, we can find ϕ a splitting, not compatible with m, making the following composition an isomorphism: ϕ OX − → F∗e OX ֒→ F∗e OX ((pe − 1)∆′ ) ֒→ F∗e OX (⌈(pe − 1)∆⌉) − → OX . Twisting by M and reflexifying (which we denote by ∗∗ ), we obtain M e − → F∗e (M (p −1) ⊗ M )∗∗ e ֒→ F∗e (OX ((pe − 1)∆′ ) ⊗ M (p −1) ⊗ M )∗∗ e ֒→ F∗e (OX (⌈(pe − 1)∆⌉) ⊗ M (p −1) ⊗ M )∗∗ ϕ −−M → M. Using the fact that (OX ((pe − 1)∆′ ) ⊗ M (p e −1) )∗∗ ∼ = OX , we have a composition M− → F∗e M − →M that is an isomorphism (note the first map is not the usual inclusion of ideal sheaves via Frobenius). Certainly Hm1 (M ) = 0 since M is reflexive and thus S2 by [Har94, Theorem 1.9]. We now study Hm2 (M ). Since M is S2 , it follows that either AnnR Hm2 (M ) is equal to R or it is mprimary by Lemma 2.13. Since we have an injection Hm2 (M ) − → Hm2 (F∗e M ), it follows that e 2 2 p AnnR Hm (M ) is at the very least radical (since if r kills Hm (M ), so does r). In particular, if AnnR Hm2 (M ) 6= R, then it must be m. 6Note that it is also common to use the notation confused with Frobenius power. [·] . We do not use that notation since it might be 10 ZSOLT PATAKFALVI AND KARL SCHWEDE Fix [z] ∈ Hm2 (M ), and recall that we are considering M as an ideal. If zj ∈ Γ(U, M )7, then since ∆′ ≥ 0 we have e e zjp ∈ Γ(U, M (p ) ) ⊆ Γ U, (OX ((pe − 1)∆′ ) ⊗ M (p e −1)
⊗ M )∗∗ . Thus we have a class e [F∗e z p ] ∈ Hm2 F∗e (OX ((pe − 1)∆′ ) ⊗ M (p e −1)
⊗ M )∗∗ . e Now, when we apply ϕM to this class, it is just applied component-wise. Thus ϕM ([F∗e z p ]) = e [z]. For any r ∈ m, it follows that ϕM ((F∗e r).[F∗e z p ]) = ϕ(F∗e r).[z]. In particular, if an e arbitrary F∗e r ∈ F∗e m annihilates all classes [y] ∈ Hm2 (F∗e (OX ((pe − 1)∆′ ) ⊗ M (p −1) ⊗ M )∗∗ ) ∼ = Hm2 (F∗e M ), then ϕ(F∗e r) also annihilates all such [z] ∈ Hm2 (M ). This proves that m is ϕ-compatible, a contradiction.  4. Applications Here we list the most important corollaries of the results of Section 3. The characteristic zero analog of many of them are already mentioned in [Kol11b]. We still state them here for the sake of completeness and we give a full proof of our main motivation, the compatibility of the relative canonical sheaf with base change. In Section 4.A some lemmas are gathered while in Section 4.B the promised corollaries are presented. 4.A. Auxilliary results. In this section, we prove a series of lemmas culminating with a base change statement for relative canonical sheaves for families with sharply F -pure fibers Lemma 4.7. 4.A.1. Basic lemmas on depth and relative canonical sheaves. We begin with a short section where we make note of some simple results on depth and relative canonical sheaves that we will use. Fact 4.1. [BH93, Theorem 1.2.5] Let F be a coherent sheaf on a Noetherian scheme X, H a Cartier divisor on X containing a point P , such that the local equation of H at P is not a zero divisor of FP (in other words, it is a regular element for Fp ). Then (a) depth FP ≥ d ⇔ depth(F |H )P ≥ d − 1, (b) depth FP ≥ min{d, dim FP } ⇔ depth(F |H )P ≥ min{d − 1, dim(F |H )P }. Lemma 4.2. If f : X −→ Y is a morphism of Noetherian schemes, F 6= 0 is a coherent sheaf on X flat over Y , such that F |Xy is Sd for every y ∈ Y (i.e., F is relatively Sd over Y ) and G 6= 0 is an Sd coherent sheaf on Y , then F ⊗ f ∗ G is Sd as well. 7Here U = X \ {x}, and local cohomology classes are treated as Čech classes on U as explained earlier. DEPTH OF F -SINGULARITIES AND BASE CHANGE OF RELATIVE CANONICAL SHEAVES 11 Proof. Fix an arbitrary P ∈ X and set Q := f (P ) and F := XQ . Then depthOX,P (F ⊗ f ∗ G )P = depthOF,P (F |F )P + depthOY,Q GQ | {z } F is flat over Y & [Gro65, Proposition 6.3.1]  ≥ min d, dimOF,P (F |F )P {z | F |F is Sd  + depthOY,Q GQ } = min d + depthOY,Q GQ , dimOF,P (F |F )P + depthOY,Q GQ  ≥ min d + depthOY,Q GQ , dimOF,P (F |F )P + min{d, dimOY,Q GQ } | {z } G is Sd  = min d + depthOY,Q GQ , min{dimOF,P (F |F )P + d, dimOF,P (F |F )P + dimOY,Q GQ }  = min d + depthOY,Q GQ , min{dimOF,P (F |F )P + d, dimOX,P (F ⊗ f ∗ G )P } | {z } F is flat over Y and [Gro65, Corollaire 6.1.2]  ≥ min d + depthOY,Q GQ , min{d, dimOX,P (F ⊗ f ∗ G )P }  ≥ min d, dimOX,P (F ⊗ f ∗ G )P .  Lemma 4.3. If f : X −→ Y is a flat morphism of finite type to a Gorenstein scheme, then q q ωX/Y and ωX [− dim Y ] are locally isomorphic. In particular if X is relatively Gorenstein over Y , then it is Gorenstein, and ωX/Y and ωX are locally isomorphic over Y . Proof. Locally on Y the following isomorphisms hold. q q q ωX/Y = f ! OY ∼ = ωX [− dim Y ] = f ! ωY ∼ = f ! ωY [− dim Y ] ∼  Before continuing, let us remind ourselves of how F -adjunction works and how it can allow us to restrict divisors. 4.A.2. Restricting divisors by F -adjunction: the F -different. Suppose that ∆ ≥ 0 is a Qdivisor on an S2 and G1 variety X and that (pe − 1)(KX + ∆) is Cartier. In fact, everything we say even holds more generally if ∆ ≥ 0 is a Z(p) -Weil divisorial sheaf, which is intuitively something like a Weil divisor having components also in the singular locus8. Further suppose that D is a reduced Cartier divisor on X that is itself S2 and G1 and which has no common components with ∆. We now explain how we can construct a canonical Z(p) -Weil divisorial sheaf (not necessarily a Z(p) -Weil divisor) which we call Diff F,D ∆ on D. Here the subindex F means that this is the F -singularity counterpart of the usual different known from minimal model program theory. However, contrary to the usual different, the construction of Dif fF,D ∆ goes through without any further assumption requiring that ∆ is Q-Cartier at certain points. Without loss of generality we can assume that X = Spec R and that R is a local ring and that D = V (f ). The fact that the divisor (pe − 1)(KX + ∆) is Cartier implies that HomR (F∗e R((pe − 1)∆), R) is a free F∗e R-module. Choose a generator of this module ϕ, which 8See [K+92] and [MS12] for definitions, in the latter source these are called Z -AC-divisors. (p) 12 ZSOLT PATAKFALVI AND KARL SCHWEDE we can also view as an element of HomR (F∗e R, R) since ∆ is effective. Now define a new map e e ψ : F∗e R − → R by the rule ψ(F∗e ) = ϕ(F∗e (f p −1 · )). Certainly ψ(F∗e hf i) = ϕ(F∗e hf p i) ⊆ hf i and thus ψ induces a map ψ ∈ HomR/hf i (F∗e R/hf i, R/hf i). Note that for every height one prime η containing f ∈ R, i.e., a minimal associated prime of D, we have that ∆η = 0 (since ∆ and D have no common components). Furthermore, Rη is regular (since Rη /hf i is reduced and zero dimensional and hence regular). It follows from inspection that ψ is non-zero at every such η. By [MS12, Theorem 2.4], it follows that ψ induces an effective Z(p) -Weil divisorial sheaf on D. It is straightforward to verify that ∆|D is independent of the choice of e and ϕ and so: Definition 4.4. We use Diff F,D ∆ to denote the effective Z(p) -Weil divisorial sheaf described above which coincides with ψ. We also observe: Observation 4.5 (F -adjunction [Sch09]). Notice now additionally that ψ (corresponding to Diff F,D ∆) is surjective if and only if ψ (corresponding to ∆+D) is surjective. In other words, (X, ∆ + D) is sharply F -pure near D if and only if (D, Diff F,D ∆) is sharply F -pure. It will be useful for us to note that if (D, Diff F,D ∆) is sharply F -pure, then Diff F,D ∆ is in fact an honest Z(p) -Weil divisor and not just a Z(p) -divisorial sheaf. Suppose not, then the Weil divisorial sheaf (pe − 1)(Diff F,D ∆) must properly contain OD , even at some generic point of the non-normal locus. A contradiction can then be obtained from the fact that the conductor is already compatible with every ϕ ∈ C R (this last fact follows from the argument of [BK05, Proposition 1.2.5]). The above introduced F -different Diff F,D ∆ is equal to ∆|D in most cases when the latter is defined. Lemma 4.6. With the notation above, suppose additionally that ∆ is Z(p) -Cartier at all of the height-two primes primes of R containing f (the codimension 2 points of X that are contained inside D). Then Diff F,D ∆ coincides with the restriction ∆|D of ∆ to D. Proof. At each of those codimension 2 points q ∈ Spec R, Rq is already Gorenstein (since R/hf i is G1 and so Rq /hf i is Gorenstein). It is enough to prove the result at each such q, so fix one such q. Further choose e > 0 as above and also sufficiently divisible such that (pe − 1)D is Cartier at q. We can thus write (pe − 1)∆ = divSpec Rq (g) for some g ∈ Rq . Since Rq is Gorenstein, we can choose Φ ∈ HomRq (F∗e Rq , Rq ) generating the set as an F∗e Rq e → Rq defined by the rule Ψ(F∗e ) = Φ(F∗e (f p −1 · )). module. Consider the map Ψ : F∗e Rq − Certainly Ψ restricts to a map Ψ ∈ HomRq /hf i (F∗e (Rq /hf i), Rq /hf i) as above. Furthermore, Ψ generates the F∗e (Rq /hf i)-module HomRq /hf i (F∗e (Rq /hf i), Rq /hf i) by the diagrams in [Sch09, Proof of Proposition 7.2]. It follows that ψ = (F∗e g) · Ψ restricts to ϕ = (F∗e g) · Ψ and hence corresponds to the naive restriction ∆|W . This proves the lemma.  4.A.3. The relative canonical sheaf. We apply the above ideas on F -different to the following. It is the inductional step in the proof of Corollary 4.13. Lemma 4.7. Let f : X −→ Y be a flat morphism of finite type with S2 , G1 equidimensional fibers to a smooth variety9 Y and ∆ ∈ WDivQ (X), such that KX + ∆ is Q-Cartier and p 6 | ind(KX + ∆). Assume also that Z ⊆ Y is a smooth Cartier divisor such that for W := X ×Y Z, ∆ does not contain any component of W and (W, Diff F,W ∆) is sharply F -pure.10 Then ωX/Y |W ∼ = ωW/Z . 9Variety here means a separated, integral scheme of finite type over an algebraically closed base-field. 10The fact that (W, Diff F,W ∆) is sharply F -pure implies that Diff F,W ∆ is a Z(p) -divisor and not simply a Z(p) -Weil divisorial sheaf. DEPTH OF F -SINGULARITIES AND BASE CHANGE OF RELATIVE CANONICAL SHEAVES 13 Proof. By Lemma 4.2, both X and W are S2 . Similarly, both are G1 by Lemma 4.3. By F adjunction Observation 4.5, cf. [Sch09, Main Theorem, Proposition 7.2, Remark 7.3], (X, ∆+ W ) is sharply F -pure in a neighborhood of W . Hence, so is (X, ∆). We now claim: Claim 4.8. No F -pure center of (X, ∆) is contained in W . Proof of claim. Suppose that Z ⊆ X was an F -pure center of (X, ∆) contained in W . Let η denote the generic point of Z and now we work in R = OX,η with maximal ideal m → OX,η of [C ∆ ]e , we corresponding to η. For any element ϕ : F∗e OX,η ⊆ F∗e OX,η (⌈(pe −1)∆⌉) − e notice that ϕ(F∗ m) ⊆ m since Z is an F -pure center. Choose f ∈ m to be the defining equation e of the Cartier divisor W in OX,η . It follows from construction that (F∗e f p −1 ) · [C ∆ ]e = e [C ∆+W ]e . In other words, for any ψ ∈ [C ∆+W ]e , we can write ψ(F∗e ) = ϕ(F∗e (f p −1 · )) e for some ϕ ∈ [C ∆ ]e . With this notation, for any r ∈ R we have ψ(F∗e r) = ϕ(F∗e (f p −1 r)) ∈ ϕ(F∗e m) ⊆ m. This proves that (X, ∆ + W ) is not sharply F -pure at η, the generic point of Z. But we assumed that (X, ∆ + W ) was sharply F -pure, a contradiction which proves the claim.  We return to the proof of Lemma 4.7. By Theorem 3.8 then, for every x ∈ W , depthx ωX ≥ min{3, codimX x} = min{3, dim ωX,x }. However, by Lemma 4.3, ωX and ωX/Y are isomorphic locally, and then in the above inequality ωX can be replaced by ωX/Y . Then by Fact 4.1, ωX/Y |W is S2 . To be precise, to apply Fact 4.1, one needs to prove a priori that the local equation of W is not a zero-divisor of ωX/Y . For this it is enough to show that ωX/Y is S1 , which follows using again that locally ωX/Y and ωX are isomorphic and that ωX is S2 by [KM98, Corollary 5.69]. Therefore ωX/Y |W is S2 indeed. However, so is ωW/Z by using [KM98, Corollary 5.69] again. Furthermore, ωX/Y |W and ωW/Z are isomorphic on the relative Gorenstein locus, since the relative canonical sheaf is compatible with base-change for Gorenstein morphisms [Con00, Theorem 3.6.1]. Therefore  [Har94, Theorems 1.9 and 1.12] yields the statement of the lemma. The next lemma is used in Corollary 4.15. It allows us to cite [Kol11a]. Lemma 4.9. Suppose that Y is a scheme of finite type over an algebraically closed field. If f : X −→ Y is a projective, flat, relatively S2 and G1 , equidimensional morphism, then ωX/Y is reflexive. Proof. According to [HK04, Corollary 3.7], it is enough to exhibit an open set U contained in the relative Gorenstein locus, such that (a) for Z := X \ U , codimXy Zy ≥ 2 for every y ∈ Y and (b) for the inclusion of open set j : U ֒→ X, the natural homomorphism ωX/Y − → j∗ (ωX/Y |U ) is an isomorphism. Let W be the non-relatively Gorenstein locus. Fix a finite surjective morphism π : X − → PnY over Y , after possibly shrinking Y (cf. [Kol11b, proof of Corollary 24]). Set then Z := π −1 (π(W )), V := PnY \ π(W ). Let q : V − → PnY be the natural inclusion. With the above choices, codimXy Zy ≥ 2 is satisfied for all y ∈ Y . For the other condition, notice that it is enough to prove that the natural homomorphism π∗ ωX/Y − → π∗ j∗ (ωX/Y |U ) is an isomorphism. However π∗ j∗ (ωX/Y |U ) ∼ = q∗ ((π∗ ωX/Y )|V ) ∼ = q∗ Hom V ((π∗ OX )|V , ωV /Y ) {z } | Grothendieck duality ∼ = Hom X (π∗ OX , q∗ ωV /Y ) ∼ = | {z } adjoint functors Hom X (π∗ OX , ωPn /Y ) | {z Y } [HK04, Proposition 3.5] using that ωPn /Y is flat and relatively S2 Y ∼ = π∗ ωX/Y , 14 ZSOLT PATAKFALVI AND KARL SCHWEDE and the composition of the above isomorphisms is the natural homomorphism π∗ ωX/Y − → π∗ j∗ (ωX/Y |U ).  4.B. Consequences. We begin with a simple consequence on the depth of OX and ωX . Corollary 4.10. (cf. [Ale08, Lemma 3.2], [Fuj09, Theorem 4.21], [Kol11b, 4.1, 4.2, 4.3], [AH11, Theorem 1.5], [Kov11, Theorem 1.2, 1.5]) Suppose that X = Spec R is S2 and G1 . If X is F -pure and x ∈ X is not an F -pure center of X, then depthx OX ≥ {3, codimX x} and depthx ωX ≥ min{3, codimX x}. Proof. We may assume that X = Spec R for a local ring (R, m) with x = V (m). Since X is S2 and G1 , by using Remark 3.5, we can assume that there exists some ∆ ≥ 0 such that (pe − 1)(KX + ∆) is Cartier, such that (X, ∆) is sharply F -pure and such that x is not an F -pure center of (X, ∆). Now the second statement follows from Theorem 3.8 by setting M = OX (KX ) and setting ∆′ = ∆. The first statement also follows from Theorem 3.8 by setting M = OX and ∆′ = 0.  Question 4.11. Suppose that (R, m) is F -injective. If m is not an annihilator of any F -stable submodule of Hmi (R), does that imply any depth conditions on R or ωR ? To prove our main corollary, we need to introduce a generalization of Diff F,D ∆ to the case when D has higher codimension. We focus only on our case of interest, that is, when D is the fiber over a smooth base. Definition 4.12. Let f : X − → Y be a flat morphism of finite type with S2 , G1 equidimensional fibers. Further suppose that Y is a smooth variety, and ∆ ∈ WDivQ (X) is such that ∆ does not contain any component of any fiber. Fix a point y ∈ Y . Working locally, we may assume that Y = Spec A and X = Spec R for local rings (A, n) and (R, m). Further assume that n = hf1 , . . . , fn i is regular system of generators with Yi = V (f1 , . . . , fi ) regular. Then, let ∆i := Diff F,Yi ∆i−1 and define then Diff F,Xy ∆ := ∆n . The only question is whether this construction of Diff F,Xy ∆ is independent of the choice of fi . Following the method of F -adjunction, multiplying by each fi successively, we take a map corresponding to ∆ and ϕ : F∗e R − → R and obtain another map ψf : F∗e R − → R defined by e e e the rule ψf (F∗ ) = ϕ(F∗ ((f1 · · · fn )p −1 · )). We then restrict this map to Xy by modding out by n and so obtain ψ f . Choosing different Yi ’s is simply choosing a different set of generators {g1 , . . . , gn } for n which yields ψ g . To complete the proof of the claim, it is sufficient to show that these maps e differ only by multiplication by a unit. We use n[p ] to denote the ideal generated by the pe th e e e e e powers of the generators of n. Since n[p ] : n = h(f1 · · · fn )p −1 i+n[p ] = h(g1 · · · gn )p −1 i+n[p ] , e e see for example [Fed83, Proposition 2.1], it follows that (f1 · · · fn )p −1 = u(g1 · · · gn )p −1 + P e vi hpi for some unit u ∈ A, elements vi ∈ A and hi ∈ n. But now we see that ψ f = (F∗e u)·ψ g e since any multiple of hpi will be sent into nR. This completes the proof. Corollary 4.13. (cf. [Kol11b, 4.3]) Let f : X −→ Y be a flat morphism of finite type with S2 , G1 , equidimensional fibers to a smooth variety and let ∆ ∈ WDivQ (X) be such that it does not contain any component of any fiber. Additionally assume that KX + ∆ is Q-Cartier, p 6 | ind(KX + ∆) and (Xy , Diff F,Xy ∆) is sharply F -pure for every y ∈ Y . Then ωX/Y is flat over Y and compatible with arbitrary base change. Proof. We claim that ωX/Y is flat over Y and relatively S2 . By [BHPS12, Lemma 2.13], flatness follows as soon as we prove that the restriction of ωX/Y to every fiber is S1 . On DEPTH OF F -SINGULARITIES AND BASE CHANGE OF RELATIVE CANONICAL SHEAVES 15 the other hand relatively S2 means the stronger condition that the above restrictions are S2 . Therefore to show the claim, it is enough to prove that ωX/Y |Xy is S2 for every y ∈ Y . By [KM98, Corollary 5.69], ωXy is S2 and hence, it is enough to show that ωX/Y |Xy ∼ = ωXy locally around every point x ∈ Xy . We thus replace both X and Y by Spec OX,x and Spec OY,y , respectively. Therefore, we may assume that there is a sequence of smooth subvarieties: Y = Y0 ⊇ Y1 ⊇ · · · ⊇ Ym−1 ⊇ Ym = {y}, such that Yi−1 is a Cartier divisor in Yi . Set Xi := XYi and ∆i := Diff F,Xi ∆i−1 with ∆0 = ∆. Note that then Deltam = Diff F,Xy ∆. By applying Observation 4.5 (backwards) inductively and possibly further restricting X around x, one obtains that (Xi , ∆i + Xi−1 ) and hence (Xi , ∆i ) is sharply F -pure for all i (in fact, this also implies all the Z(p) -Weil divisorial sheaves ∆i are honest divisors). Finally, applying Lemma 4.7 inductively again yields that ωXi /Yi |Xi−1 ∼ = ωXi−1 /Yi−1 for all i, and consequently ωX/Y |Xy ∼ = ωXy . This finishes the proof of our claim. By our claim and [HK04, Corollary 3.8] ωX/Y and all its pullbacks are reflexive. Hence, by restricting to the relatively Cohen-Macaulay locus (whose complement has codimension ≥ 2) and using [HK04, Proposition 3.6], for any morphism Z − → Y , ωXZ /Z ∼  = (ωX/Y )Z . Remark 4.14. By Lemma 4.6, the appearance of Diff F,Xy ∆ can be replaced by an actual “geometric” restriction, if we assume the following: (1) for each y ∈ Y , there is some r > 0 relatively prime to p such that r∆ is Cartier at the codimension 1 points of the fiber Xy ⊆ X. In particular, this is satisfied if Supp ∆ does not contain the singular codimension one points of the fibers. Indeed, let ξ be a codimesnion 1 point of a fiber Xy . If ξ is in the singular locus of Xy , then ξ 6∈ Supp ∆ and hence ∆ is Cartier at ξ. Otherwise, X is smooth around ξ, and hence KX is Cartier at ξ. In particular then by p 6 | ind(KX + ∆), we obtain that ∆ is Z(p) Cartier at ξ. In either cases ∆ satisfies (1), and therefore in the special case of Corollary 4.13 stated in Section 1, the use of ordinary restriction of ∆ was legitimate. When f is projective, the compatibility of Corollary 4.13 follows for arbitrary reduced base by an important result of Kollár [Kol11a]. Corollary 4.15. Let f : X −→ Y be a flat projective morphism with S2 , G1 equidimensional fibers. Further suppose that Y is a reduced, separated scheme of finite type over an algebraically closed field, and ∆ a Q-Weil-divisor that avoids all the codimension zero and the singular codimension one points of the fibers. Additionally assume that there is a p 6 | N > 0, such [N ] that N ∆ is Cartier in relative codimension one and ωX/Y (N ∆) 11 is a line bundle and that (Xy , ∆y ) is sharply F -pure for every y ∈ Y . Then ωX/Y is flat and compatible with arbitrary base change. Proof. First, we need some preparation about pulling back ∆. Suppose τ : Y ′ − → Y is a ′ ′ ′ ′ ′ ′ morphism and set X := X ×Y Y , π : X − → X, and f : X − → Y the induced morphisms. Then a natural pullback ∆′ of ∆ can be defined as follows. Let U ⊆ X be the open set where f is Gorenstein and ∆ is Q-Cartier. Then, pull ∆|U back to π −1 U , and finally extend it uniquely over X ′ . This extension is unique, since codimX ′ X ′ \ π −1 U ≥ 2. We claim that (2) [N ] [N ] π ∗ ωX/Y (N ∆) ∼ = ωX ′ /Y ′ (N ∆′ ). 11 Let U be the intersection of the relative Gorenstein locus and the locus where N ∆ is Cartier. Set [N] N ι : U −→ X for the natural inclusion. The sheaf ωX/Y (N ∆) is the reflexive hull of ωU/Y (N ∆|U ), i.e. the N unique reflexive sheaf that restricts on U to the above sheaf. It can be obtained as ι∗ (ωU/Y (N ∆|U )). Indeed, N ι∗ (ωU/Y (N ∆|U )) is reflexive by [HK04, Corollary 3.7] and it is unique by [HK04, Proposition 3.6]. 16 ZSOLT PATAKFALVI AND KARL SCHWEDE [N ] [N ] Indeed, notice that by construction π ∗ ωX/Y (N ∆) and ωX ′ /Y ′ (N ∆′ ) agree over π −1 U , that [N ] is, in relative codimension one. Notice also that since ωX/Y (N ∆) is assumed to be a line [N ] [N ] bundle, so is π ∗ ωX/Y (N ∆), and therefore π ∗ ωX/Y (N ∆) is reflexive. On the other hand, since [N ] ωX ′ /Y ′ (N ∆′ ) is defined as a pushforward of a line bundle from relative codimension one, it is reflexive by [HK04, Corollary 3.7]. Therefore by [HK04, Proposition 3.6], (2) holds. In [N ] particular, ωX ′ /Y ′ (N ∆′ ) is a line bundle. The main consequence of the previous paragraph, is that the conditions of the corollary are invariant under pullback to another reduced, separated scheme Y ′ of finite type over k. That is, X ′ , f ′ and ∆′ defined above satisfy all the assumptions of the corollary. Let us turn now to the actual proof of the corollary. First, we may assume that Y is connected. Second, according to [Kol11a, Corollary 24] and Lemma 4.9, there is a locally closed decomposition ∐Yi − → Y , such that if T − → Y , then ωXT /T is flat and commutes with base-change if and only if T − → Y factors through some Yi − → Y . Now, for every irreducible component Y ′ of Y , →S there is a regular alteration S − → Y ′ [dJ96, Thorem 4.1]. By the above discussion XS − satisfies the assumptions of the corollary, and hence also of Corollary 4.13. Therefore, ωXS /S is flat and compatible with arbitrary base-change. Hence S − → Y factors through one of the Yi . In particular, since the image of S − → Y is the component Y ′ , Y ′ ⊆ Yi . That is, every irreducible component of Y is contained in one Yi . However, Y is connected, therefore all irreducible components of Y are contained in the same Yi , and hence by the reducedness of Y , Yi = Y .  Remark 4.16. In the case of dim Y = 1, if instead of assuming that Xy is sharply F -pure, one assumes that (X, Xy ) is F -pure for all y ∈ Y , the p 6 | ind(KX + ∆) assumption can be dropped from the above corollaries using the trick of Lemma 2.12. Question 4.17. Does the compatibility of the relative canonical sheaf with base change stated in Corollary 4.13 hold for singular Y (with the adequate modification in the setup as in Corollary 4.15)? From the modular point of view, especially interesting would be the case of non-reduced Y . This case is open even in the projective case. Remark 4.18. It should be noted that the characteristic zero analogue of Corollary 4.13 is known if f is projective and Y is arbitrary [KK10, Theorem 7.9]. That is, the answer to the characteristic zero analogue of Question 4.17 is positive when f is projective. Question 4.19. Can one replace sharply F -pure by log-canonical (still assuming positive characteristic) in the statement of Corollary 4.13? This would also be important from the modular point of view, since sharply F -pure varieties can be deformed to log-canonical but not sharply F -pure varieties. Question 4.20. Can one remove the divisibility by p condition from the statement of Corollary 4.13? Remark 4.21. [Kol11b, 4.10] The sheaf OX (−D) in Theorem 3.1 and Theorem 3.8 cannot be replaced by OX (D) as shown in [Kol11b]. We refer to [Kol11b] for the actual example. References [AE05] I. M. Aberbach and F. Enescu: The structure of F -pure rings, Math. Z. 250 (2005), no. 4, 791–806. MR2180375 [Ale08] V. Alexeev: Limits of stable pairs, Pure Appl. Math. Q. 4 (2008), no. 3, Special Issue: In honor of Fedor Bogomolov. Part 2, 767–783. 2435844 (2009j:14020) [AH11] V. Alexeev and C. Hacon: Non-rational centers of log canonical singularities, arXiv:1109.4164. [BHPS12] B. Bhatt, W. Ho, Zs. Patakfalvi, and C. Schnell: Moduli of products of stable varieties, arXiv:1206.0438. DEPTH OF F -SINGULARITIES AND BASE CHANGE OF RELATIVE CANONICAL SHEAVES [Bli09] [BST11] [BK05] 17 M. Blickle: Test ideals via algebras of p−e -liner maps, arXiv:0912.2255, to appear in J. Algebraic Geom. M. Blickle, K. Schwede, and K. Tucker: F -signature of pairs and the asymptotic behavior of Frobenius splittings, arXiv:1107.1082. M. Brion and S. Kumar: Frobenius splitting methods in geometry and representation theory, Progress in Mathematics, vol. 231, Birkhäuser Boston Inc., Boston, MA, 2005. MR2107324 (2005k:14104) [BH93] [Con00] [dJ96] [EH08] [Fed83] W. Bruns and J. Herzog: Cohen-Macaulay rings, Cambridge Studies in Advanced Mathematics, vol. 39, Cambridge University Press, Cambridge, 1993. MR1251956 (95h:13020) B. Conrad: Grothendieck duality and base change, Lecture Notes in Mathematics, vol. 1750, Springer-Verlag, Berlin, 2000. MR1804902 (2002d:14025) A. J. de Jong: Smoothness, semi-stability and alterations, Inst. Hautes Études Sci. Publ. Math. (1996), no. 83, 51–93. 1423020 (98e:14011) F. Enescu and M. Hochster: The Frobenius structure of local cohomology, Algebra Number Theory 2 (2008), no. 7, 721–754. MR2460693 (2009i:13009) R. Fedder: F -purity and rational singularity, Trans. Amer. Math. Soc. 278 (1983), no. 2, 461–480. MR701505 (84h:13031) [Fuj09] [Gab04] [Gro65] O. Fujino: Introduction to the log minimal model program for log canonical pairs, arXiv:0907.1506. O. Gabber: Notes on some t-structures, Geometric aspects of Dwork theory. Vol. I, II, Walter de Gruyter GmbH & Co. KG, Berlin, 2004, pp. 711–734. A. Grothendieck: Éléments de géométrie algébrique. IV. Étude locale des schémas et des morphismes de schémas. II, Inst. Hautes Études Sci. Publ. Math. (1965), no. 24, 231. 0199181 (33 #7330) [HK10] [HW02] [Har66] [Har94] [HK04] [HR76] [Kol87] C. D. Hacon and S. J. Kovács: Classification of higher dimensional algebraic varieties, Oberwolfach Seminars, vol. 41, Birkhäuser Verlag, Basel, 2010. 2675555 (2011f:14025) N. Hara and K.-I. Watanabe: F-regular and F-pure rings vs. log terminal and log canonical singularities, J. Algebraic Geom. 11 (2002), no. 2, 363–392. MR1874118 (2002k:13009) R. Hartshorne: Residues and duality, Lecture notes of a seminar on the work of A. Grothendieck, given at Harvard 1963/64. With an appendix by P. Deligne. Lecture Notes in Mathematics, No. 20, Springer-Verlag, Berlin, 1966. MR0222093 (36 #5145) R. Hartshorne: Generalized divisors on Gorenstein schemes, Proceedings of Conference on Algebraic Geometry and Ring Theory in honor of Michael Artin, Part III (Antwerp, 1992), vol. 8, 1994, pp. 287–339. MR1291023 (95k:14008) B. Hassett and S. J. Kovács: Reflexive pull-backs and base extension, J. Algebraic Geom. 13 (2004), no. 2, 233–247. MR2047697 (2005b:14028) M. Hochster and J. L. Roberts: The purity of the Frobenius and local cohomology, Advances in Math. 21 (1976), no. 2, 117–172. MR0417172 (54 #5230) J. Kollár: Subadditivity of the Kodaira dimension: fibers of general type, Algebraic geometry, Sendai, 1985, Adv. Stud. Pure Math., vol. 10, North-Holland, Amsterdam, 1987, pp. 361–398. MR946244 (89i:14029) [Kol90] J. Kollár: Projectivity of complete moduli, J. Differential Geom. 32 (1990), no. 1, 235–268. 1064874 (92e:14008) [Kol11a] [Kol11b] [K+92] J. Kollár: Hulls and husks, arXiv:0805.0576. J. Kollár: A local version of the Kawamata-Viehweg vanishing theorem, Pure Appl. Math. Q. 7 (2011), no. 4, Special Issue: In memory of Eckart Viehweg, 1477–1494. 2918170 J. Kollár and 14 coauthors: Flips and abundance for algebraic threefolds, Société Mathématique de France, Paris, 1992, Papers from the Second Summer Seminar on Algebraic Geometry held at the University of Utah, Salt Lake City, Utah, August 1991, Astérisque No. 211 (1992). MR1225842 (94f:14013) [KK10] [KM98] [Kov11] [Kun76] J. Kollár and S. J. Kovács: Log canonical singularities are Du Bois, J. Amer. Math. Soc. 23 (2010), no. 3, 791–813. 2629988 J. Kollár and S. Mori: Birational geometry of algebraic varieties, Cambridge Tracts in Mathematics, vol. 134, Cambridge University Press, Cambridge, 1998, With the collaboration of C. H. Clemens and A. Corti, Translated from the 1998 Japanese original. MR1658959 (2000b:14018) S. Kovács: Irrational centers, Pure Appl. Math. Q. 7 (2011), no. 4, Special Issue: In memory of Eckart Viehweg, 1495–1515. 2918171 E. Kunz: On Noetherian rings of characteristic p, Amer. J. Math. 98 (1976), no. 4, 999–1013. MR0432625 (55 #5612) 18 ZSOLT PATAKFALVI AND KARL SCHWEDE [MS12] [Pat10] [Ray78] [Sch09] [Sch10] [Sch11] [SS10] [Tak08] L. E. Miller and K. Schwede: Semi-log canonical vs F -pure singularities, J. Algebra 349 (2012), 150–164. 2853631 Zs. Patakfalvi: Base change behavior of the relative canonical sheaf related to higher dimensional moduli, arXiv:1005.5207, to appear in Algebra Number Theory. M. Raynaud: Contre-exemple au “vanishing theorem” en caractéristique p > 0, C. P. Ramanujam—a tribute, Tata Inst. Fund. Res. Studies in Math., vol. 8, Springer, Berlin, 1978, pp. 273–278. 541027 (81b:14011) K. Schwede: F -adjunction, Algebra Number Theory 3 (2009), no. 8, 907–950. 2587408 (2011b:14006) K. Schwede: Centers of F -purity, Math. Z. 265 (2010), no. 3, 687–714. 2644316 (2011e:13011) K. Schwede: Test ideals in non-Q-Gorenstein rings, Trans. Amer. Math. Soc. 363 (2011), no. 11, 5925–5941. K. Schwede and K. E. Smith: Globally F -regular and log Fano varieties, Adv. Math. 224 (2010), no. 3, 863–894. 2628797 (2011e:14076) S. Takagi: A characteristic p analogue of plt singularities and adjoint ideals, Math. Z. 259 (2008), no. 2, 321–341. MR2390084 (2009b:13004) Department of Mathematics, Princeton University, Princeton, NJ, 08542, USA E-mail address: pzs@princeton.edu Department of Mathematics, The Pennsylvania State University, University Park, PA, 16802, USA E-mail address: schwede@math.psu.edu 

0math.AC

A Unified View of Entropy-Regularized Markov Decision Processes Gergely Neu gergely.neu@gmail.com arXiv:1705.07798v1 [cs.LG] 22 May 2017 Universitat Pompeu Fabra, Barcelona, Spain Vicenç Gómez vicen.gomez@upf.edu Universitat Pompeu Fabra, Barcelona, Spain Anders Jonsson anders.jonsson@upf.edu Universitat Pompeu Fabra, Barcelona, Spain Abstract We propose a general framework for entropy-regularized average-reward reinforcement learning in Markov decision processes (MDPs). Our approach is based on extending the linear-programming formulation of policy optimization in MDPs to accommodate convex regularization functions. Our key result is showing that using the conditional entropy of the joint state-action distributions as regularization yields a dual optimization problem closely resembling the Bellman optimality equations. This result enables us to formalize a number of state-of-the-art entropy-regularized reinforcement learning algorithms as approximate variants of Mirror Descent or Dual Averaging, and thus to argue about the convergence properties of these methods. In particular, we show that the exact version of the TRPO algorithm of Schulman et al. (2015) actually converges to the optimal policy, while the entropy-regularized policy gradient methods of Mnih et al. (2016) may fail to converge to a fixed point. Finally, we illustrate empirically the effects of using various regularization techniques on learning performance in a simple reinforcement learning setup. 1. Introduction Reinforcement learning is the discipline of model-based optimal sequential decisionmaking in unknown stochastic environments. In average-reward reinforcement learning, the goal is to find a behavior policy that maximizes the long-term average reward, taking into account the effect of each decision on the future evolution of the decisionmaking process. In known environments, this optimization problem has been studied (at least) since the influential work of Bellman (1957) and Howard (1960): the optimal behavior policy can be formulated as the solution of the Bellman optimality equations. In unknown environments with partially known or misspecified models, greedily solving these equations often results in policies that are far from optimal in the true environment. Rooted in statistical learning theory (Vapnik, 2013), the notion of Neu Gómez Jonsson regularization offers a principled way of dealing with this issue, among many others. In particular, entropy regularization has proven to be one of the most successful tools of machine learning and related fields (Littlestone and Warmuth, 1994; Vovk, 1990; Freund and Schapire, 1997; Kivinen and Warmuth, 2001; Arora et al., 2012). The idea of entropy regularization has also been used extensively in the reinforcement learning literature (Sutton and Barto, 1998; Szepesvári, 2010). Entropyregularized variants of the classic Bellman equations and the entailing reinforcementlearning algorithms have been proposed to induce safe exploration (Fox et al., 2016) and risk-sensitive policies (Howard and Matheson, 1972; Marcus et al., 1997; Ruszczyński, 2010), or to model observed behavior of imperfect decision-makers (Ziebart et al., 2010; Ziebart, 2010; Braun et al., 2011), among others. Complementary to these approaches rooted in dynamic programming, another line of work proposes direct policy search methods attempting to optimize various entropy-regularized objectives (Williams and Peng, 1991; Peters et al., 2010; Schulman et al., 2015; Mnih et al., 2016; O’Donoghue et al., 2017), with the main goal of driving a safe online exploration procedure in an unknown Markov decision process. Notably, the state-of-the-art methods of Mnih et al. (2016) and Schulman et al. (2015) are both based on entropy-regularized policy search. In this work, we connect these two seemingly disparate lines of work by showing a strong Lagrangian duality between the entropy-regularized Bellman equations and a certain regularized average-reward objective. Specifically, we extend the linearprogramming formulation of the problem of optimization in MDPs to accommodate convex regularization functions, resulting in a convex program. We show that using the conditional entropy of the joint state-action distribution gives rise to a set of nonlinear equations resembling the Bellman optimality equations. Observing this duality enables us to establish a connection between regularized versions of value and policy iteration methods (Puterman and Shin, 1978) and incremental convex optimization methods like Mirror Descent (Nemirovski and Yudin, 1983; Beck and Teboulle, 2003) or Dual Averaging (Xiao, 2010; McMahan, 2014; Hazan et al., 2016; Shalev-Shwartz, 2012). For instance, the convex-optimization view we propose reveals that the TRPO algorithm of Schulman et al. (2015) and the regularized policy-gradient method of Mnih et al. (2016) are approximate versions of Mirror Descent and Dual Averaging, respectively, and that both can be interpreted as regularized policy iteration methods. Our work provides a theoretical justification for various algorithms that were first derived heuristically. In particular, our framework reveals that the exact version of TRPO is identical to the MDP-E algorithm of Even-Dar et al. (2009). This establishes the fact that the policy updates of TRPO converge to the optimal policy, improving on the theoretical results claimed by Schulman et al. (2015). We also argue that our formulation is useful for pointing out possible inconsistencies of heuristic learning algorithms. In particular, we show that the approximation steps employed by Mnih et al. (2016) may break the convexity of the objective, thus possibly leading to convergence to bad local optima or even divergence. This observation is in accordance 2 A unified view of entropy-regularized MDPs with the very recent findings of Asadi and Littman (2016), who show that value iteration with poorly chosen approximate updates may lead to divergence. To complement these results, we suggest an alternative objective that can be optimized consistently, avoiding the possibility of diverging. A similar Lagrangian duality between the Bellman equations and entropy maximization has been previously noted by Ziebart (2010, Sec. 5.2) and Rawlik et al. (2012) for a special class of episodic Markov decision processes where the time index within the episode is part of the state representation. In this particular setting, the convexity of the conditional entropy is more obvious. One of our key observations is pointing out the convexity of the conditional entropy of distribution functions defined over general state spaces, which enables us to develop a much broader theory of regularized Markov decision processes. We note that our theory also readily extends to discounted MDPs by replacing the stationary state-action distributions we consider by discounted state-action occupancy measures. For consistency, we will discuss each particular algorithm in their most natural average-reward version, noting that all conclusions remain valid in the simpler discounted and episodic settings. The rest of the paper is organized as follows. In Section 2, we provide background on average-reward Markov decision processes, briefly discussing both linear-programming and dynamic-programming derivations of the optimal control. In Section 3, we provide a convex-programming formulation of regularized average-reward Markov decision processes, and show the connection to the regularized Bellman equations. Section 4 provides a brief summary of the complementary dynamic-programming formulation and discusses regularized equivalents of related concepts, such as expressions of the regularized policy gradient. In Section 5, we describe several existing learning algorithms in our framework. We provide an empirical evaluation of various regularization schemes in a simple reinforcement learning problem in Section 6. P P set S, we will often use s as shorthand for s∈S , and we Notation. Given a finite P use ∆(S) = {µ ∈ RS : s µ(s) = 1, µ(s) ≥ 0 (∀s)} to denote the set of all probability distributions on S. 2. Preliminaries on Markov decision processes We consider a finite Markov decision process (MDP) M = (X , A, P, r), where X is the finite state space, A is the finite action space, P : X × A × X → [0, 1] is the transition function, with P (y|x, a) denoting the probability of moving to state y from state x when taking action a, and r : X × A → R is the reward function mapping state-action pairs to rewards. In each round t, the learner observes state Xt ∈ X , selects action At ∈ A, moves to the next state Xt+1 ∼ P (·|Xt , At ), and obtains reward r(Xt , At ). The goal is to select actions as to maximize some notion of cumulative reward.iIn this paper we consider the h P T 1 average-reward criterion lim inf T →∞ E T t=1 rt (Xt , At ) . A stationary state-feedback 3 Neu Gómez Jonsson policy (or policy for short) defines a probability distribution π(·|x) over the learner’s actions in state x. MDP theory (see, e.g., Puterman (1994)) stipulates that under mild conditions, the average-reward criterion can be maximized by stationary policies. Throughout the paper, we make the following mild assumption about the MDP: Assumption 1 The MDP M is unichain: All stationary policies P π induce a unique stationary distribution νπ over the state space satisfying νπ (y) = x,a P (y|x, a)π(a|x)νπ (x) for all y ∈ X . In particular, this assumption is satisfied if all policies induce an irreducible and aperiodic Markov chain (Puterman, 1994). For ease of exposition in this section, we also make the following simplifying assumption: Assumption 2 The MDP M admits a single recurrent class: All stationary policies π induce stationary distributions strictly supported on the same set X 0 ⊆ X . In general, this assumption is very restrictive in that it does not allow policies to cover different parts of the state space. We stress that our results in the later sections do not require this assumption to hold. With the above assumptions in mind, we can define the average reward of any policy π as " # T X 1 ρ(π) = lim E rt (Xt , At ) , T →∞ T t=1 where At ∼ π(·|Xt ) in each round t and the existence of the limit is ensured by Assumption 1. Furthermore, the average reward of any policy π can be simply P written as ρ(π) = x,a νπ (x)π(a|x)r(x, a), which is a linear function of the stationary state-action distribution µπ = νπ π. This suggests that finding the optimal policy can be equivalently written as a linear program (LP) where the decision variable is the stationary state-action distribution. Defining the set of all feasible stationary distributions as ( ) X X ∆ = µ ∈ ∆(X × A) : µ(y, b) = P (y|x, a)µ(x, a) (∀y) , (1) x,a b the problem of maximizing the average reward can be written as µ∗ = arg max ρ(µ). (2) µ∈∆ Just as a policy π induces stationary distributions νπPand µπ , a stationary distribution µ induces a state distribution νµ defined as νµ (x) = a µ(x, a) and a policy πµ defined as πµ (a|x) = µ(x, a)/νµ (x), where the denominator is strictly positive for recurrent states by Assumption 2. Since ∆ is a compact polytope (non-empty by Assumption 1) 4 A unified view of entropy-regularized MDPs the maximum in (2) is well-defined and induces an optimal policy πµ∗ in recurrent states. Due to Assumption 2, πµ∗ can be arbitrarily defined in transient states. The linear program specified in Equation (2) is well studied in the MDP literature (see, e.g., Puterman, 1994, Section 8.8), although most commonly as the dual of the linear program minρ∈R ρ subject to ρ + V (x) − (3) X P (y|x, a)V (y) ≥ r(x, a), ∀(x, a). (4) y Here, the dual variables V are commonly referred to as the value functions. By strong LP duality and our Assumption 1, the solution to this LP equals the optimal average reward ρ∗ and the dual variables V ∗ at the optimum are the solution to the average-reward Bellman optimality equations ! X V ∗ (x) = max r(x, a) − ρ∗ + P (y|x, a)V ∗ (x) , (∀x) . (5) a y Note that V ∗ is not unique as for any solution V , a constant shift V − c for any c ∈ R is also a solution. However, we can obtain a unique solution V ∗ by imposing the P additional constraint x,a µ∗ (x, a)V ∗ (x) = 0, which states that the expected value should equal 0. 3. Regularized MDPs: A convex-optimization view Inspired by the LP formulation of the average-reward optimization problem (2), we now define a regularized optimization objective—a framework that will lead us to our main results. Our results in this section only require the mild Assumption 1. Our regularized optimization problem takes the form ( ) X 1 max ρeη (µ) = max µ(x, a)r(x, a) − R(µ) , (6) µ∈∆ µ∈∆ η x,a where R(µ) : RX ×A → R is a convex regularization function and η > 0 is a learning rate that trades off the original objective and regularization. Note that η = ∞ recovers the unregularized objective. Unlike previous work on LP formulations for MDPs, we find it useful to regard (6) as the primal. We focus on two families of regularization functions: the negative Shannon entropy of (X, A) ∼ µ, X RS (µ) = µ(x, a) log µ(x, a), (7) x,a 5 Neu Gómez Jonsson and the negative conditional entropy of (X, A) ∼ µ, X X µ(x, a) = νµ (x)πµ (a|x) log πµ (a|x). RC (µ) = µ(x, a) log P b µ(x, b) x,a x,a (8) In what follows, we refer to these functions as the relative entropy and the conditional entropy. We also make use of the Bregman divergences induced by RS and RC which take the respective forms X X µ(x, a) πµ (a|x) DS (µkµ0 ) = µ(x, a) log 0 and DC (µkµ0 ) = µ(x, a) log . 0 (a|x) µ (x, a) π µ x,a x,a While the form of DS is standard (it is the relative entropy between two state-action distributions), the fact that DC is the Bregman divergence of RC (or even that RC is convex) is not immediately obvious1 . The following proposition asserts this statement, which we prove in Appendix A.1. The only work we are aware of that establishes a comparable result is the recent paper of Neu and Gómez (2017). Proposition 1 The Bregman divergence corresponding to the conditional entropy RC is DC . Furthermore, DC is nonnegative on ∆, implying that RC is convex and DC is convex in its first argument. We proceed to derive the dual functions and optimal solutions to (6) for our two choices of regularization functions. Without loss of generality, we assume that the reference policy πµ0 has full support, which implies that the corresponding stationary distribution µ0 is strictly positive on the recurrent set X 0 . We only provide the derivations for the Bregman divergences; the calculations are analogous for RS and RC . Both of these solutions will be expressed with the help of dual variables V : RX → R which are useful to think about as value functions, as in the case of the LP formulation (2). We also define the corresponding advantage functions A(x, a) = P r(x, a) + y P (y|x, a)V (y) − V (x). 3.1. Relative entropy The choice R = DS (·kµ0 ) has been studied before by Peters et al. (2010) and Zimin and Neu (2013); we defer the proofs to Appendix A.3. The optimal state-action distribution for a given value of η is ∗ µ∗η (x, a) ∝ µ0 (x, a)eηAη (x,a) , (9) where A∗η is the advantage function for the optimal dual variables Vη∗ . The dual function is X 1 g(V ) = log µ0 (x, a)eηA(x,a) , (10) η x,a 1. In the special case of loop-free episodic environments, showing the convexity of RC is straightforward (Lafferty et al., 2001; Ziebart, 2010; Rawlik et al., 2012). 6 A unified view of entropy-regularized MDPs that now needs to be minimized on RX with no constraints in order to obtain Vη∗ . By strong duality, g is convex in V and takes the value ρe∗η = maxµ∈∆ ρeη (µ) at its optimum. 3.2. Conditional entropy The choice R = DC (·kµ0 ) leads to our main contributions. Similar to above, the optimal policy is ∗ πη∗ (a|x) ∝ πµ0 (a|x)eηAη (x,a) . (11) In this case, the dual problem closely resembles the average-reward Bellman optimality equations (5): Proposition 2 The dual of the optimization problem (6) when R = DC (·kµ0 ) is given by minλ∈R λ !! X X 1 P (y|x, a)V (y) , (∀x) . πµ0 (a|x) exp η r(x, a) − λ + subject to V (x) = log η y a We defer the proofs to Appendix A.4. Using strong duality, the optimum of the above problem is ρe∗η , which implies that the optimal dual variables Vη∗ are given as a solution to the system of equations !! X X 1 P (y|x, a)Vη∗ (y) , (∀x) . πµ0 (a|x) exp η r(x, a) − ρe∗η + Vη∗ (x) = log η y a (12) By analogy with the Bellman optimality equations (5), we call this the regularized average-reward Bellman optimality equations. Since ρe∗η is guaranteed to be finite (because it is the maximum of a bounded function on a compact domain), the solution to the above optimization problem is well-defined, bounded, and unique up to a constant shift (as in the case of the LP dual variables). Again, we can make the solution unique by imposing the constraint that the expected value should equal 0. 4. Dynamic programming in regularized MDPs We now present a dynamic-programming view of the regularized optimization problem (Bertsekas, 2007) for the choice R = DC (·kµ0 ). Similar derivations have been done several times for discounted and episodic MDPs (Littman and Szepesvári, 1996; Ruszczyński, 2010; Azar et al., 2011; Rawlik et al., 2012; Asadi and Littman, 2016; Fox et al., 2016), but we are not aware of any work that considers the average-reward case. That said, the generalization is straightforward, and the existence and unicity of 7 Neu Gómez Jonsson the optimal solution to the Bellman optimality equation (12) follows from our results in the previous section. We first define the regularized Bellman equations for an arbitrary policy π and a reference policy π 0 : ! X X π(a|x) 1 − ρeη (π) + P (y|x, a)Vηπ (y) (∀x) , Vηπ (x) = π(a|x) r(x, a) − log 0 η π (a|x) y a (13) where ρeη (π) is the regularized average reward of policy π defined as in Equation (6). By our Assumption 1 and Proposition 4.2.4 of Bertsekas (2007), it is easy to show that this system of equations has a unique solution satisfying the additional constraint P ∗|π 0 π and x,a µπ (x, a)Vη (x) = 0. We also define the Bellman optimality operator Tη π|π 0 the Bellman operator Tη that correspond to the Bellman equations (12) and (13), 0 respectively, as well as the greedy policy operator Gπη that corresponds to Equation (11) (for completeness, the formal definitions appear in Appendix B). We include two results that are useful for deriving approximate dynamic programming algorithms. We first provide a counterpart to the performance-difference lemma (Burnetas and Katehakis, 1997, Prop. 1, Kakade and Langford, 2002, Lemma 6.1, Cao, 2007). This statement will rely on the regularized advantage function Aπη defined for each policy π as Aπη (x, a) = r(x, a) − X 1 log π(a|x) P (y|x, a)Vηπ (y) − Vηπ (x), − ρ e (π) + η η log π 0 (a|x) y where Vηπ is the regularized value function corresponding to π with baseline π 0 . Lemma 3 For any pair of policies π, π 0 , we have X µπ0 (x, a)Aπη (x, a). ρe(π 0 ) − ρe(π) = x,a For completeness, we provide the simple proof in Appendix B.1. Second, we provide an expression for the gradient of ρeη , thus providing a regularized counterpart of the policy gradient theorem of Sutton et al. (1999). To formalize this statement, let us consider a policy πθ parametrized by a vector θ ∈ Rd and assume that the gradient ∇πθ (a|x) exists for all x, a and all θ. The form of the policy gradient is given by the following lemma, which we prove in Appendix B.2: Lemma 4 Assume that ∂πθ∂θ(a|x) /πθ (a|x) > 0 for all θi , x, a. The gradient of ρeη exists i and satisfies X ∇e ρη (θ) = µπθ (x, a)∇ log πθ (a|x)Aπη θ (x, a). x,a 8 A unified view of entropy-regularized MDPs 5. Algorithms In this section we derive several reinforcement learning algorithms based on our results. For clarity of presentation, we assume that the MDP M is fully known, an assumption that we later relax in the experimental evaluation. We will study a generic sequential optimization framework where a sequence of policies πk are computed iteratively. Inspired by the online convex optimization literature (see, e.g., ShalevShwartz, 2012; Hazan et al., 2016) and by our convex-optimization formulation, we study two families of algorithms: Mirror Descent and Dual Averaging (also known as Follow-the-Regularized-Leader). 5.1. Iterative policy optimization by Mirror Descent A direct application of the Mirror Descent algorithm (Nemirovski and Yudin, 1983; Beck and Teboulle, 2003; Martinet, 1978; Rockafellar, 1976) to our case is defined as   1 µk+1 = arg max ρ(µ) − DR (µkµk ) , (14) η µ∈∆ where DR is the Bregman divergence associated with the convex regularization function R. We now proceed to show how various learning algorithms can be recovered from this formulation. 5.1.1. Mirror Descent with the relative entropy We first remark that the Relative Entropy Policy Search (REPS) algorithm of Peters et al. (2010) can be formulated as an instance of Mirror Descent with the Bregman divergence DS . This is easily seen by comparing the form of the update rule (14) with the problem formulation of Peters et al. (2010, pp. 2), with the slight difference that our regularization is additive and theirs is enforced as a constraint. It is easy to see that this only amounts to a change in learning rate. This connection is not new: it has been first shown by Zimin and Neu (2013)2 , and has been recently rediscovered by Montgomery and Levine (2016). Independently of each other, Zimin and Neu (2013) and Dick et al. (2014) both show that Mirror Descent achieves near-optimal regret guarantees in an online learning setup where the transition function is known, but the reward function is allowed to change arbitrarily between decision rounds. This implies that REPS duly converges to the optimal policy in our setup. 5.1.2. Mirror Descent with the conditional entropy We next show that the Dynamic Policy Programming (DPP) algorithm of Azar et al. (2012) and the Trust-Region Policy Optimization (TRPO) algorithm of Schulman et al. (2015) are both approximate variants of Mirror Descent with the Bregman 2. Although they primarily referred to Mirror Descent as the “Proximal Point Algorithm” following (Rockafellar, 1976; Martinet, 1978). 9 Neu Gómez Jonsson divergence DC . To see this, note that a full Mirror Descent update requires computing the optimal value function Vη∗ for the baseline µk , e.g. by regularized value iteration or regularized policy iteration (see Appendix B). Since a full update for Vη∗ is expensive, DPP and TRPO provide two ways to approximate it. We remark that the algorithm of Rawlik et al. (2012) can also be viewed as an instance of Mirror Descent for the finite-horizon episodic setting, in which the exact update can be computed efficiently by dynamic programming. Dynamic Policy Programming. We first claim that each iteration of DPP is a single regularized value iteration step: Starting from the previous value function Vk , it extracts the greedy policy πk+1 = Gπη k [Vk ] and applies the Bellman optimality ∗|π ∗|π operator Tη k to obtain Vk+1 = Tη k [Vk ]. This follows from comparing the form of DPP presented in Appendix A of Azar et al. (2012): their update rules (19) and (20) precisely match the discounted analogue of our expressions (26) in Appendix B with π 0 = πk . The convergence guarantees proved by Azar et al. (2012) demonstrate the soundness of this approximate update. Trust-Region Policy Optimization. Second, we claim that each iteration of TRPO is a single policy iteration step: TRPO first fully evaluates the policy πk to compute its unregularized value function Vk = V∞πk and then extracts the regularized greedy policy πk+1 = Gπη k [Vk ] with πk as a baseline. This can be seen by inspecting the TRPO update3 that takes the form (  ) X X π(a|x) 1 πk πk+1 = arg max νπk (x) π(a|x) A∞ (x, a) − log . η πk (a|x) π x a This objective approximates Mirror Descent by ignoring the effect of changing the policy on the state distribution. Surprisingly, using our formalism, this update can be expressed in closed form as πk πk+1 (a|x) ∝ πk (a|x)eηA∞ (x,a) . We present the detailed derivations in Appendix B.3. A particularly interesting consequence of this result is that TRPO is completely equivalent to the MDP-E algorithm of Even-Dar et al. (2009) (see also (Neu et al., 2010, 2014)), which is known to minimize regret in an online setting, thus implying that TRPO also converges to the optimal policy in the stationary setting. This guarantee is much stronger than the ones provided by Schulman et al. (2015), who only claim that TRPO produces a monotonically improving sequence of policies (which may still converge to a suboptimal policy). 3. As in the case of REPS, we discuss here the additive-regularization version of the algorithm. The entropy-constrained update actually implemented by Schulman et al. (2015) only differs in the learning rate. 10 A unified view of entropy-regularized MDPs 5.2. Iterative policy optimization by Dual Averaging We next study algorithms arising from the Dual Averaging scheme (Xiao, 2010; McMahan, 2014), commonly known as Follow-the-Regularized-Leader in online learning (Shalev-Shwartz, 2012; Hazan et al., 2016). This algorithm is defined by the iteration   1 (15) µk+1 = arg max ρ(µ) − R(µ) , ηk µ∈∆ where ηk is usually an increasing sequence to ensure convergence in the limit. We are unaware of any pure instance of dual averaging using relative entropy, and only discuss conditional entropy below. 5.2.1. Dual Averaging with the conditional entropy Just as for Mirror Descent, a full update (15) requires computing the optimal value function Vη∗ . Various approximations of this update have been long studied in the RL literature—see, e.g., (Littman and Szepesvári, 1996) (with additional discussion by (Asadi and Littman, 2016)), (Perkins and Precup, 2002; Ruszczyński, 2010; Petrik and Subramanian, 2012; Fox et al., 2016). In this section, we focus on the stateof-the-art algorithms of Mnih et al. (2016) and O’Donoghue et al. (2017) that were originally derived from an optimization formulation resembling our Equation (6). Our main insight is that this algorithm can be adjusted to have a dynamic-programming interpretation and a convergence guarantee. Entropy-regularized policy gradients. The A3C algorithm of Mnih et al. (2016) aims to maximize X 1 X ρ(π) − νπk (x) π(a|x) log π(a|x) ηk x a by taking policy gradient steps. Interestingly, our formalism implies a connection between TRPO and A3C. Due to P Lemma 4,Pthe gradient of ρ(πθ ) with respect to θ coincides with the gradient of x νπk (x) a πθ (a|x)Aπ∞k (x, a), so A3C actually attempts to optimize the objective   X X 1 πk log πθ (a|x) . (16) νπk (x) πθ (a|x) A∞ (x, a) − η k x a This objective can be seen as the dual-averaging counterpart of the TRPO objective. As in the case of TRPO, the maximizer of this objective can be computed in closed form as πk πk+1 (a|x) ∝ eηk A∞ . Unlike for TRPO, we are not aware of any convergence results for A3C, and we believe the algorithm does not converge. Indeed, the objective (16) is non-convex in either 11 Neu Gómez Jonsson of the natural parameters µ or π, which can cause premature convergence to a bad local optimum. An even more serious concern is that the objective function changes between iterations, so gradient descent may fail to converge to any stationary point. This problem is avoided by TRPO since the sum of the TRPO objectives is a sensible optimization objective (Even-Dar et al., 2009, Theorem 4.1). However, there is no such clear interpretation for the objective (16). O’Donoghue et al. (2017, Section 3.1) study the stationary points of the objective (16) and, similarly to us, show a connection between a certain type of value function and a policy achieving the stationary point. However, they do not show that this stationary point is unique nor that gradient descent converges to a stationary point4 . As we argue above, this may very well not be the case. These observations are consistent with those of Asadi and Littman (2016), who show that softmax policy updates may lead to inconsistent behavior when used in tandem with unregularized advantage functions. To overcome these issues, we advocate for directly optimizing the objective (15) instead of (16) via gradient descent. Due to the fact that (15) is convex in µ and to standard results regarding dual averaging (McMahan, 2014), this scheme is guaranteed to converge to the optimal policy. Estimating the gradients can be done analogously as for the unregularized objective, by our Lemma 4. 6. Experiments In this section we analyze empirically several of the algorithms described in the previous section, with the objective of illustrating the interplay of regularization and model-estimation error in a simple reinforcement learning setting. We consider an iterative setup where in each episode k = 1, 2, . . . , N , we execute a policy πk , observe the sample transitions and update the estimated model via maximum likelihood. We focus on the regularization aspect, with no other approximation error than that introduced by model estimation. It is important to emphasize that the comparison may not extend to other variants of the algorithms or in the presence of other sources of approximation. We consider a simple MDP, defined on a grid (Fig. 1, left), where an agent has four possible actions (up, down, left and right) that succeed with probability 0.9 but fail with probability 0.1. In case of failure, the agent does not move or goes to any random adjacent location. Negative (or positive) rewards are given after hitting a wall (or reaching one of the white diamond locations, respectively). In both cases, the agent is sent back to one of the starting locations (marked with ’X’ in the figure). The reward of the diamonds is proportional to the distance from the starting locations. Therefore, the challenge of this experiment is to discover the path towards the top-right reward while learning the dynamics incrementally, and then exploit it. 4. Strictly speaking, O’Donoghue et al. (2017) do not even show that a stationary point (16) exists. 12 A unified view of entropy-regularized MDPs 8 0.05 8 0.04 0.03 0.02 6 6 DA-RV DA TRPO DPP RegVI Opt 4 0.01 2 0 -0.01 0 10 -4 10 -3 10 -2 10 -1 10 0 10 1 2 avg reward 10 avg reward 0.06 4 DPP ModRegPI-2 ModRegPI-20 TRPO 2 0 10 0 10 1 10 2 iteration k Figure 1: Left: the MDP used for evaluation. Reward is −0.1 at the walls and 5, 50, 200 at the diamonds. The optimal policy is indicated by red arrows. The cell colors correspond to the stationary state distribution for open locations. Middle: Average reward as a function of the learning rate η for all algorithms (see text for details). Number of iterations N and samples per iteration S are N = S = 500. Results are taken over 20 random runs per value of η. Right: Performance of DPP, TRPO and two version of modified regularized Policy Iteration for a fixed η ≈ 0.1. Note that the optimal agent ignores the intermediate reward at the center and even prefers to hit a wall in locations too far away from the largest reward (bottom-left). We fix the number of iterations N and samples per iteration S and analyze the average reward of the final policy as a function of η. We compare the following algorithms: regularized Value Iteration with a fixed reference uniform policy and fixed η (RegVI); several variants of approximate Mirror Descent, including DPP and TRPO (Section 5.1); and two Dual Averaging methods (DA and DA-RV). DA corresponds to optimizing the objective (16), which is not guaranteed to lead to an optimal policy, and DA-RV corresponds to the iteration (15), which has convergence guarantees (Section 5.2). For both variants, we use a linear annealing schedule ηk = η · k. Fig. 1 (middle) shows results as a function of η. The maximum reward is depicted in blue at the top. For very small η (strong regularization), all algorithms perform poorly and do not even reach the intermediate reward. In contrast, for very large η, they converge prematurely to the greedy policy that exploits the intermediate reward. Typically, for an intermediate value of η, the algorithms occasionally discover the optimal path and exploit it. Note that this is not the case for RegVI, which never obtains the optimal policy. This shows that using both a fixed value of η and a fixed reference policy is a bad choice. In this MDP, we observe that the performance of both Dual Averaging methods (DA and DA-RV) is very similar, and in general slightly better than the approximate Mirror Descent variants. 13 Neu Gómez Jonsson We also show an interesting relationship between the Mirror Descent approximations. Our analysis in Section 5.1.2 suggests an entire array of algorithms lying between DPP and TRPO, just as Modified Policy Iteration lies between Value Iteration and Policy Iteration (Puterman and Shin, 1978; Scherrer et al., 2012). Fig. 1 (right) illustrates this idea, showing the convergence of DPP and TRPO for a fixed value of η. TRPO tends to converge faster than DPP to a locally optimal policy, since DPP uses a single value update per iteration. Using more value updates leads to a modified regularized Policy Iteration algorithm (we call it ModRegPI-2 and ModRegPI-20, for 2 and 20 updates, respectively) that interpolates between DPP and TRPO. 7. Conclusion We have presented a unifying view of entropy-regularized MDPs from a convexoptimization perspective. We believe that such unifying theories can be very useful in moving a field forward: We recall that in the field of online learning theory, the convex-optimization view has enabled a unified treatment of many existing algorithms and acts today as the primary framework for deriving new algorithms (see the progress from Cesa-Bianchi and Lugosi (2006) through Shalev-Shwartz (2012) to Hazan et al. (2016)). In this paper, we argued that the convex-optimization view may also be very useful in analyzing algorithms for reinforcement learning: In particular, we demonstrated how this framework can be used to provide theoretical justification for state-of-the-art reinforcement learning algorithms, and how it can highlight potential problems with them. We expect that this newly-found connection will also open the door for constructing more advanced reinforcement learning algorithms by borrowing further ideas from the convex optimization literature, such as Composite Objective Mirror Descent Duchi et al. (2010) and Regularized Dual Averaging (Xiao, 2010). Finally, we point out that our work does not provide a statistical justification for using entropy regularization in reinforcement learning. In the case of online learning in known Markov decision processes with changing reward functions, entropyregularization has been known to yield near-optimal learning algorithms (Even-Dar et al., 2009; Neu et al., 2010, 2014; Zimin and Neu, 2013; Dick et al., 2014). It remains to be seen if this technique also provably helps in driving the exploration process in unknown Markov decision processes. References S. Arora, E. Hazan, and S. Kale. The multiplicative weights update method: A meta-algorithm and applications. Theory of Computing, 8:121–164, 2012. Kavosh Asadi and Michael L. Littman. A new softmax operator for reinforcement learning. CoRR, abs/1612.05628, 2016. 14 A unified view of entropy-regularized MDPs Mohammad Gheshlaghi Azar, Vicenç Gómez, and Hilbert J Kappen. Dynamic policy programming with function approximation. In International Conference on Artificial Intelligence and Statistics, pages 119–127, 2011. Mohammad Gheshlaghi Azar, Vicenç Gómez, and Hilbert J Kappen. Dynamic policy programming. Journal of Machine Learning Research, 13(Nov):3207–3245, 2012. A. Beck and M. Teboulle. Mirror descent and nonlinear projected subgradient methods for convex optimization. Operations Research Letters, 31(3):167–175, 2003. R. Bellman. Dynamic Programming. Princeton University Press, Princeton, New Jersey, 1957. D. P. Bertsekas. Dynamic Programming and Optimal Control, volume 2. Athena Scientific, Belmont, MA, 3 edition, 2007. Daniel A Braun, Pedro A Ortega, Evangelos Theodorou, and Stefan Schaal. Path integral control and bounded rationality. In Adaptive Dynamic Programming And Reinforcement Learning (ADPRL), 2011 IEEE Symposium on, pages 202–209. IEEE, 2011. A. N. Burnetas and M. N. Katehakis. Optimal adaptive policies for Markov Decision Processes. 22(1):222–255, 1997. Xi-Ren Cao. Stochastic Learning and Optimization: A Sensitivity-Based Approach. Springer, New York, 2007. N. Cesa-Bianchi and G. Lugosi. Prediction, Learning, and Games. Cambridge University Press, New York, NY, USA, 2006. T. Dick, A. György, and Cs. Szepesvári. Online learning in markov decision processes with changing cost sequences. In ICML 2014, 2014. John C Duchi, Shai Shalev-Shwartz, Yoram Singer, and Ambuj Tewari. Composite objective mirror descent. In COLT, pages 14–26, 2010. E. Even-Dar, S. M. Kakade, and Y. Mansour. Online Markov decision processes. Mathematics of Operations Research, 34(3):726–736, 2009. Roy Fox, Ari Pakman, and Naftali Tishby. Taming the noise in reinforcement learning via soft updates. In Proceedings of the Thirty-Second Conference on Uncertainty in Artificial Intelligence, 2016. Y. Freund and R. E. Schapire. A decision-theoretic generalization of on-line learning and an application to boosting. Journal of Computer and System Sciences, 55: 119–139, 1997. 15 Neu Gómez Jonsson Elad Hazan et al. Introduction to online convex optimization. Foundations and Trends R in Optimization, 2(3-4):157–325, 2016. R. A. Howard. Dynamic Programming and Markov Processes. The MIT Press, Cambridge, MA, 1960. Ronald A Howard and James E Matheson. Risk-sensitive markov decision processes. Management science, 18(7):356–369, 1972. Sham M. Kakade and John Langford. Approximately optimal approximate reinforcement learning. In Proceedings of the Nineteenth International Conference on Machine Learning (ICML 2002), pages 267–274, 2002. J. Kivinen and M. Warmuth. Relative loss bounds for multidimensional regression problems. Machine Learning, 45:301–329, 2001. John Lafferty, Andrew McCallum, and Fernando Pereira. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In ICML 2001, pages 282–289, 2001. URL citeseer.ist.psu.edu/lafferty01conditional.html. N. Littlestone and M. Warmuth. The weighted majority algorithm. Information and Computation, 108:212–261, 1994. M.L. Littman and Cs. Szepesvári. A Generalized Reinforcement Learning Model: Convergence and applications. In Int. Conf. on Machine Learning, pages 310–318, 1996. Steven I Marcus, Emmanual Fernández-Gaucherand, Daniel Hernández-Hernandez, Stefano Coraluppi, and Pedram Fard. Risk sensitive markov decision processes. In Systems and control in the twenty-first century, pages 263–279. Springer, 1997. B. Martinet. Perturbation des méthodes d’optimisation. applications. ESAIM: Mathematical Modelling and Numerical Analysis - Modélisation Mathématique et Analyse Numérique, 12(2):153–171, 1978. URL http://eudml.org/doc/193317. H Brendan McMahan. A survey of algorithms and analysis for adaptive online learning. arXiv preprint arXiv:1403.3465, 2014. Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International Conference on Machine Learning, pages 1928–1937, 2016. William H Montgomery and Sergey Levine. Guided policy search via approximate mirror descent. In NIPS-29, pages 4008–4016, 2016. 16 A unified view of entropy-regularized MDPs A. Nemirovski and D. Yudin. Problem Complexity and Method Efficiency in Optimization. Wiley Interscience, 1983. G. Neu, A. György, and Cs. Szepesvári. The online loop-free stochastic shortest-path problem. In Proceedings of the 23rd Annual Conference on Learning Theory (COLT), pages 231–243, 2010. Gergely Neu and Vicenç Gómez. Fast rates for online learning in Linearly Solvable Markov Decision Processes. 2017. Gergely Neu, András György, Csaba Szepesvári, and András Antos. Online Markov decision processes under bandit feedback. IEEE Transactions on Automatic Control, 59:676–691, 2014. Brendan O’Donoghue, Remi Munos, Koray Kavukcuoglu, and Volodymyr Mnih. PGQ: Combining policy gradient and Q-learning. In 5th International Conference on Learning Representations, 2017. Theodore J. Perkins and Doina Precup. A convergent form of approximate policy iteration. In Advances in Neural Information Processing Systems, pages 1595–1602, 2002. Jan Peters, Katharina Mülling, and Yasemin Altun. Relative entropy policy search. In Proceedings of the Twenty-Fourth AAAI Conference on Artificial Intelligence, pages 1607–1612, 2010. Marek Petrik and Dharmashankar Subramanian. An approximate solution method for large risk-averse markov decision processes. In Proceedings of the Twenty-Eighth Conference on Uncertainty in Artificial Intelligence, pages 805–814, 2012. Martin L. Puterman. Markov Decision Processes. Wiley, 1994. ISBN 9780471727828. URL http://books.google.com/books/about/Markov_decision_ processes.html?id=Y-gmAQAAIAAJ. Martin L. Puterman and Moon Chirl Shin. Modified policy iteration algorithms for discounted Markov decision processes. Management Science, 24:1127–1137, 1978. Konrad Rawlik, Marc Toussaint, and Sethu Vijayakumar. On stochastic optimal control and reinforcement learning by approximate inference. In Proceedings of Robotics: Science and Systems VIII, 2012. R. Tyrrell Rockafellar. Monotone Operators and the Proximal Point Algorithm. SIAM Journal on Control and Optimization, 14(5):877–898, 1976. Andrzej Ruszczyński. Risk-averse dynamic programming for Markov decision processes. Mathematical programming, 125(2):235–261, 2010. 17 Neu Gómez Jonsson B. Scherrer, V. Gabillon, M. Ghavamzadeh, and M. Geist. Approximate modified policy iteration. In ICML 2012, pages 1207–1214, 2012. John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region policy optimization. In Proceedings of the 32nd International Conference on Machine Learning (ICML-15), pages 1889–1897, 2015. Shai Shalev-Shwartz. Online learning and online convex optimization. Foundations and Trends in Machine Learning, 4(2):107–194, 2012. R. S. Sutton, D. A. McAllester, S. P. Singh, and Y. Mansour. Policy gradient methods for reinforcement learning with function approximation. In NIPS-12, pages 1057–1063, 1999. R.S. Sutton and A.G. Barto. Reinforcement Learning: An Introduction. MIT Press, 1998. Cs. Szepesvári. Algorithms for Reinforcement Learning. Synthesis Lectures on Artificial Intelligence and Machine Learning. Morgan & Claypool Publishers, 2010. Vladimir Vapnik. The nature of statistical learning theory. Springer science & business media, 2013. V. Vovk. Aggregating strategies. In Proceedings of the third annual workshop on Computational learning theory (COLT), pages 371–386, 1990. Ronald J Williams and Jing Peng. Function optimization using connectionist reinforcement learning algorithms. Connection Science, 3(3):241–268, 1991. Lin Xiao. Dual averaging methods for regularized stochastic learning and online optimization. Journal of Machine Learning Research, 11(Oct):2543–2596, 2010. Brian D Ziebart. Modeling Purposeful Adaptive Behavior with the Principle of Maximum Causal Entropy. PhD thesis, Carnegie Mellon University, 2010. Brian D Ziebart, J Andrew Bagnell, and Anind K Dey. Modeling interaction via the principle of maximum causal entropy. In International Conference on Machine Learning (ICML), pages 1247–1254, 2010. A. Zimin and G. Neu. Online learning in episodic Markovian decision processes by relative entropy policy search. In NIPS-26, pages 1583–1591, 2013. 18 A unified view of entropy-regularized MDPs Appendix A. Complementary Technical Results A.1. Convexity of the negative conditional entropy Let us consider the joint state-action distribution µ on the finite set X × A. We denote P νµ (x) = a µ(x, a) and πµ (a|x) = µ(x, a)/νµ (x) for all x, a. We study the negative conditional entropy of (X, A) ∼ µ as a function of µ: RC (µ) = X x,a X µ(x, a) µ(x, a) = µ(x, a) log . µ(x, a) log P ν (x) µ b µ(x, b) x,a We will study the Bregman divergence DRC corresponding to RC : DRC (µkµ0 ) = RC (µ) − RC (µ0 ) − ∇RC (µ0 )> (µ − µ0 ), P where the inner product between two vectors v, w ∈ RX ×A is w> v = x,a v(x, a)w(x, a). Our aim is to show that DRC is nonnegative, which will imply the convexity of RC . We begin by computing the partial derivative of RC (µ) with respect to µ(x, a): X µ(x, b) µ(x, a) µ(x, a) ∂RC (µ) = log +1− = log , ∂µ(x, a) νµ (x) ν ν µ (x) µ (x) b where we used the fact that ∂νµ (x)/∂µ(x, a) = 1 for all a. With this expression, we have RC (µ0 ) + ∇RC (µ0 )> (µ − µ0 ) = X = X µ0 (x, a) log µ0 (x, a) µ0 (x, a) X + (µ(x, a) − µ0 (x, a)) log νµ0 (x) νµ0 (x) x,a µ(x, a) log µ0 (x, a) . νµ0 (x) x,a x,a Thus, the Bregman divergence takes the form   X µ0 (x, a) µ(x, a) 0 DRC (µkµ ) = µ(x, a) log − log νµ (x) νµ0 (x) x,a = X x,a µ(x, a) log X X πµ (a|x) πµ (a|x) = ν(x) πµ (a|x) log . πµ0 (a|x) πµ0 (a|x) x a This proves that the Bregman divergence corresponding to RC coincides with DC , as claimed. To conclude the proof, note that DC is the average relative entropy between the distributions πµ and πµ0 —that is, a sum a positive terms. Indeed, this shows that DC is nonnegative on the set of state-action distributions ∆(X × A), proving that RC (µ) is convex. 19 Neu Gómez Jonsson A.2. Derivation of optimal policies Here we prove the results stated in Equations (9)-(11) and Proposition 2, which give the expressions for the dual optimization problems and the optimal solutions corresponding to the primal optimization problem (6), for the two choices of regularization function DS (·kµ0 ) and DC (·kµ0 ). We start with generic derivations that will be helpful for analyzing both cases and then turn to studying the individual regularizers. Recall that the primal optimization objective in (6) is given by ( ) X 1 max ρeη (µ) = max µ(x, a)r(x, a) − R(µ) , µ∈∆ µ∈∆ η x,a where ∆, the feasible set of stationary distributions, is defined by the following constraints: X X µ(y, b) = µ(x, a)P (y|x, a), ∀y ∈ X , (17) x,a b X µ(x, a) = 1, (18) µ(x, a) ≥ 0, ∀(x, a) ∈ X × A. (19) x,a We begin by noting that for all state-action pairs where µ0 (x, a) = 0, the optimal solution µ∗η (x, a) will also be zero, thanks to the form of our regularized objective. Thus, without loss of generality, we will assume that all states are recurrent under µ0 : µ0 (x, a) > 0 holds for all state-action pairs. For any choice of regularizer R, the Lagrangian of the primal (6) is given by X X X 1 µ(y, b) V (y) µ(x, a)P (y|x, a) − L(µ; V,λ, ϕ) = µ(x, a)r(x, a) − R(µ) + η x,a y x,a b ! X X +λ 1− µ(x, a) + ϕ(x, a)µ(x, a) X x,a x,a ! = X = X x,a x,a µ(x, a) r(x, a) + X P (y|x, a)V (y) − V (x) − λ + ϕ(x, a) y 1 − R(µ) + λ η 1 µ(x, a) (A(x, a) − λ + ϕ(x, a)) − R(µ) + λ, η where V , λ and ϕ are the Lagrange multipliers5 , and A is the advantage function for V . Setting the gradient of the Lagrangian with respect to µ to 0 yields the system of 5. Technically, these are KKT multipliers as we also have inequality constraints. However, these will be eliminated by means of complementary slackness in the next sections. 20 ! A unified view of entropy-regularized MDPs equations ∂L 1 ∂R(µ) = (A(x, a) − λ + ϕ(x, a)) − , ∂µ(x, a) η ∂µ(x, a) ∂R(µ) = η (A(x, a) − λ + ϕ(x, a)) , (20) ⇔ ∂µ(x, a) for all x, a. By the first-order stationary condition, the unique optimal solution µ∗η satisfies this system of equations. To obtain the final solution we have to compute the optimal values Vη∗ , λ∗η and ϕ∗η of the Lagrange multipliers by optimizing the dual optimization objective g(V, λ, ϕ) = L(µ∗η ; V, λ, ϕ), and insert into the expression for µ∗η . V and λ are unconstrained in the dual, while ϕ satisfies ϕ(x, a) ≥ 0 for each (x, a) ∈ X × A. We give the derivations for each regularizer below. 0= A.3. The relative entropy P Here we prove the results for R(µ) = DS (µkµ0 ) = x,a µ(x, a) log µµ(x,a) 0 (x,a) . The gradient of R is ∂R(µ) µ(x, a) = log 0 + 1. ∂µ(x, a) µ (x, a) The optimal state-action distribution µ∗η is now directly given by Equation (20): µ∗η (x, a) = µ0 (x, a) exp (η (A(x, a) − λ + ϕ(x, a)) − 1) . (21) For µ∗η to belong to ∆, it has to satisfy Constraints (18) and (19). Because of the exponent in (21), µ∗η (x, a) ≥ 0 trivially holds for any choice of ϕ(x, a), and complementary slackness implies ϕ∗η (x, a) = 0 for each (x, a). Eliminating ϕ and inserting µ∗η into Constraint (18) gives us X 1= µ0 (x, a) exp (ηA(x, a)) e−ηλ−1 , x,a 1 ⇔ λ= η ! log X µ0 (x, a) exp (ηA(x, a)) − 1 . (22) x,a Since the value of λ is uniquely determined by (22), we can optimize the dual over V only. The dual function is given by   X µ∗η (x, a) 1 1 ∗ ∗ g(V ) = L(µη ; V, λ) = µη (x, a) A(x, a) − λ − log 0 +λ= +λ η µ (x, a) η x,a X 1 = log µ0 (x, a) exp (ηA(x, a)) . η x,a This is precisely the dual given in Equation (10). Note that this dual function has no associated constraints. The expression for the optimal state-action distribution in Equation (9) is obtained by inserting the advantage function A∗η corresponding to the optimal value function Vη∗ into (21). 21 Neu Gómez Jonsson A.4. The conditional entropy We next prove the results for R(µ) = DC (µkµ0 ) = gradient of R is P x,a µ(x, a) log ππµ0(a|x) . The (a|x) µ ∂R(µ) πµ (a|x) X µ(x, b) ∂πµ (b|x) = log + · . ∂µ(x, a) πµ0 (a|x) πµ (b|x) ∂µ(x, a) b Since the policy is defined as πµ (a|x) = µ(x,a) , νµ (x) its gradient with respect to µ is
I{a=b} µ(x, b) ∂πµ (b|x) 1 = − I − π (b|x) , ∀x ∈ X , a, b ∈ A. = µ {a=b} ∂µ(x, a) νµ (x) νµ (x)2 νµ (x) Inserting into the expression for the gradient of R yields
πµ (a|x) X πµ (b|x) ∂R(µ) = log + I{a=b} − πµ (b|x) ∂µ(x, a) πµ0 (a|x) πµ (b|x) b X πµ (a|x) πµ (a|x) πµ (b|x) = log = log +1− . πµ0 (a|x) πµ0 (a|x) b The optimal policy πµ∗ is now directly given by Equation (20): πη∗ (a|x) = πµ0 (a|x) exp (η (A(x, a) − λ + ϕ(x, a))) . (23) For µ∗η to belong to ∆, it has to satisfy Constraint (19). Because of the exponent in (23) and the fact that µ∗η (x, a) ∝ πη∗ (a|x), µ∗η (x, a) ≥ 0 trivially holds for any choice of ϕ(x, a), implying that ϕ∗η (x, a) = 0 for each (x, a) by complementary slackness. Since πη∗ (a|x) = µ∗η (x,a) , νη∗ (x) X we also obtain the following set of constraints: πη∗ (a|x) = a X µ∗η (x, a) a νη∗ (x) = νη∗ (x) = 1, ∀x ∈ X . νη∗ (x) Inserting the expression for πη∗ yields 1= X πµ0 (a|x) exp (ηA(x, a)) e−ηλ , ∀x ∈ X . a If we expand the expression for A(x, a) and rearrange the terms we obtain !! X X 1 P (y|x, a)V (y) πµ0 (a|x) exp η r(x, a) − λ + , ∀x ∈ X . V (x) = log η a y (24) 22 A unified view of entropy-regularized MDPs The dual function is obtained by inserting the expression for µ∗ into the Lagrangian:   X πη∗ (a|x) 1 ∗ ∗ g(V, λ) = L(µη ; V, λ) = µη (a|x) A(x, a) − λ − log + λ = λ. (25) 0 (a|x) η π µ x,a Together, Equations (24) and (25) define the dual optimization problem in Proposition 2. The expression for the optimal policy in Equation (11) is obtained by inserting the optimal advantage function A∗η into (23). We remark that to recover the optimal stationary state-action distribution µ∗η , we would have to insert the expression for the optimal policy πη∗ into Constraints (17) and (18), and solve for the stationary state distribution νη∗ . However, this is not necessary since µ∗η and νη∗ are not required to solve the dual function or to compute the optimal policy. Appendix B. The regularized Bellman operators In this section, we define the regularized Bellman operator Tπ|π0 corresponding to the policy π and regularized with respect to baseline π 0 as ! X π(a|x) X 0 0 π|π 0 P (x |x, a)V (x ) (∀x) . + Tη [V ](x) = π(a|x) r(x, a) − log 0 π (a|x) a x0 Similarly, we define the regularized Bellman optimality operator T∗|π0 with respect to baseline π 0 as !! X X 1 0 Tη∗|π [V ](x) = log π 0 (a|x) exp η r(x, a) + P (y|x, a)V (y) (∀x) , η a y and the regularized greedy policy with respect to the baseline π 0 as !! 0 Gπη [V ](a|x) ∝ π 0 (a|x) exp η r(x, a) + X P (y|x, a)V (y) − V (x) . y With these notations, we can define the regularized relative value iteration algorithm with respect to π 0 by the iteration 0 πk+1 = Gπη [Vk ] 0 Vk+1 (x) = Tη∗|π [Vk ](x) − δk+1 (26) ∗|π 0 for some δk+1 lying between the minimal and maximal values of Tη [Vk ]. A common ∗|π 0 technique is to fix a reference state x0 and choose δk+1 = Tη [Vk ](x0 ). Similarly, we can define the regularized policy iteration algorithm by the iteration  ∞ π0 πk+1 |π 0 πk+1 = Gη [Vk ] Vk+1 (x) = Tη [Vk ](x) − δk+1 , (27) 23 Neu Gómez Jonsson with δk+1 defined analogously. For establishing the convergence of the above procedures, it is crucial to ensure ∗|π 0 that the operator Tη is a non-expansion: For any value functions V1 and V2 , we need to ensure 0 0 Tη∗|π [V1 ] − Tη∗|π [V2 ] ≤ kV1 − V2 k for some norm. We state the following result claiming that the above requirement indeed holds and present the simple proof below. We note that analogous results have been proven several times in the literature, see, e.g., (Fox et al., 2016; Asadi and Littman, 2016). ∗|π 0 is a non-expansion for the supremum norm kf k∞ = maxx |f (x)|. P Proof For simplicity, let us introduce the notation Q1 (x, a) = r(x, a)+ y P (y|x, a)V1 (y), with Q2 defined analogously, and ∆ = Q1 − Q2 . We have Proposition 5 Tη ∗|π 0 Tη ∗|π 0 [V1 ](x) − Tη ! 1 [V2 ](x) = η log X π 0 (a|x) exp (ηQ1 (x, y)) − log X π 0 (a|x) exp (ηQ2 (x, y)) a a P   0 π (a|x) exp (ηQ (x, y)) 1 1 = log Pa 0 π (a|x) exp (ηQ η 2 (x, y)) Pa 0   exp (η (Q2 (x, y) + ∆(x, y))) 1 a π (a|x) P 0 = log η a π (a|x) exp (ηQ2 (x, y)) X 1 p(x, a) exp (η∆(x, a)) (with an appropriately defined p) = log η a 1 log max exp (η∆(x, a)) a η X = max ∆(x, a) = max P (y|x, a) (V1 (y) − V2 (y)) ≤ a a y ≤ max |V1 (y) − V2 (y)| . y With an analogous technique, we can also show the complementary inequality 0 0 Tη∗|π [V2 ](x) − Tη∗|π [V1 ](x) ≤ max |V2 (y) − V1 (y)| , y which concludes the proof. ∗|π 0 Together with the easily-seen fact that Tη is continuous, this result immediately ∗|π 0 implies that Tη has a fixed point by Brouwer’s fixed-point theorem. Furthermore, this insight allows us to treat the value iteration method (26) as an instance of generalized value iteration, as defined by Littman and Szepesvári (1996). ∗|π 0 We now argue that regularized value iteration converges to the fixed point of Tη . If the initial value function V0 is bounded, then so is Vk for each k since the operator 24 A unified view of entropy-regularized MDPs ∗|π 0 Tη is a non-expansion. Similar to Section 3, we assume without loss of generality that the initial reference policy π0 has full support, i.e. π0 (a|x) > 0 for each recurrent 0 state x and each action a. Inspecting the greedy policy operator Gπη , it is easy to show by induction that πk has full support for each k. In particular, πk+1 (a|x) only equals 0 if either πk (a|x) equals 0 or if the exponent Ak (x, a) equals −∞, which is only possible if Vk is unbounded. Now, since πk has full support for each k, any trajectory always has a small probability of reaching a given recurrent state. We can now use a similar argument as Bertsekas (2007, Prop. 4.3.2) to show that regularized value iteration converges to the ∗|π 0 fixed point for Tη . B.1. The proof of Lemma 3 Let µ and µ0 be the respective stationary distributions of π and π 0 . The statement follows easily from using the definition of Aπη : X x,a µ 0 (x, a)Aπη (x, a) X 1 log π(a|x) = µ (x, a) r(x, a) − − ρ e (µ) + P (y|x, a)Vηπ (y) − Vηπ (x) 0 η log π (a|x) x,a y ! X X µ0 (x, a) P (y|x, a)Vηπ (y) − Vηπ (x) = ρe(µ0 ) − ρe(µ) + X 0 x,a y 0 = ρe(µ ) − ρe(µ), where the last step follows from the stationarity of µ0 . B.2. Regularized policy gradient Here we prove Lemma 4 which gives the gradient of the regularized average reward ρeη (θ) when the policy πθ is parameterized on θ. Following Sutton et al. (1999), we first compute the gradient of Vηπθ : ! X 1 πθ (a|x) r(x, a) − log 0 − ρeη (θ) + P (y|x, a)Vηπθ (y) η π (a|x) y ! X ∂Vηπθ (y) 1 ∂πθ (a|x) ∂ ρeη X + πθ (a|x) − − + P (y|x, a) . ηπθ (a|x) ∂θi ∂θi ∂θi a y ∂Vηπθ (x) X ∂πθ (a|x) = ∂θi ∂θi a 25 ! Neu Gómez Jonsson Rearranging the terms gives us X 1 1 πθ (a|x) r(x, a) − log 0 − ρeη (θ) + P (y|x, a)Vηπθ (y) − η π (a|x) η y ∂ ρeη X ∂πθ (a|x) = ∂θi ∂θi a + X πθ (a|x) X a = P (y|x, a) y X ∂πθ (a|x) ∂θi a Aπη θ (x, a) + ! ∂Vηπθ (y) ∂Vηπθ (x) − ∂θi ∂θi X πθ (a|x) a X y ∂Vηπθ (y) ∂Vηπθ (x) P (y|x, a) − . ∂θi ∂θi P The last equality follows from the fact that since a ∂πθ (a|x)/∂θi = 0 for each x, we can add any state-dependent constant to the multiplier of ∂πθ (a|x)/∂θi ; adding the term η1 − Vηπθ (x) results in the given expression. Summing both sides over the stationary state distribution νπθ yields ∂ ρeη X ∂πθ (a|x) πθ ∂ ρeη X = νπθ (x) = νπθ (x) Aη (x, a) ∂θi ∂θ ∂θ i i x x,a + XX y = X = X x,a x,a νπθ (x)πθ (a|x)P (y|x, a) x,a ∂Vηπθ (y) X ∂Vηπθ (x) − νπθ (x) ∂θi ∂θi x X ∂Vηπθ (y) X ∂Vηπθ (x) ∂πθ (a|x) πθ Aη (x, a) + νπθ (y) − νπθ (x) νπθ (x) ∂θi ∂θi ∂θi y x νπθ (x) ∂πθ (a|x) πθ Aη (x, a). ∂θi To conclude the proof it is sufficient to note that µπθ (x, a) ∂ log πθ (a|x) µπ (x, a) ∂πθ (a|x) ∂πθ (a|x) = θ = νπθ (x) . ∂θi πθ (a|x) ∂θi ∂θi B.3. The closed form of the TRPO update Here we derive the closed-form solution of the TRPO update. To do so, we first briefly summarize the mechanism of the algorithm. The main idea of Schulman et al. (2015) is replacing ρ(µ0 ) by the surrogate X X Lπ (π 0 ) = ρ(π) + νπ (x) π 0 (a|x)Aπ∞ (x, a), x a 26 A unified view of entropy-regularized MDPs where Aπ∞ is the unregularized advantage function corresponding to policy π. Furthermore, TRPO uses the regularization term DTRPO (µkµ0 ) = X νµ0 (x) X x πµ (a|x) log a 6 πµ (a|x) . πµ0 (a|x) The difference between L + DTRPO and ρ + DC is that the approximate version ignores the impact of changing the policy π on the stationary distribution. Given this surrogate objective, TRPO approximately computes the distribution   1 (28) µk+1 = arg max Lµk (µ) − DTRPO (µkµk ) . η µ∈∆ Observing that the TRPO policy update can be expressed equivalently as (  ) X X 1 π(a|x) π(a|x) Aπ∞k (x, a) − log πk+1 = arg max νµk (x) , η πk (a|x) π a x we can see that the policy update can be expressed in closed form as πk πk+1 (a|x) ∝ πk (a|x)eηA∞ (x,a) . This update then can be seen as a regularized greedy step with respect to the value function of the previous policy πk . 6. This form is inspired by the well-known identity we state as Lemma 3. 27 

2cs.AI

Learning Policies for Adaptive Tracking with Deep Feature Cascades Chen Huang Simon Lucey Deva Ramanan Robotics Institute, Carnegie Mellon University arXiv:1708.02973v2 [cs.CV] 13 Sep 2017 chenh2@andrew.cmu.edu {slucey,deva}@cs.cmu.edu Pixel layer + correlation filters Abstract Visual object tracking is a fundamental and time-critical vision task. Recent years have seen many shallow tracking methods based on real-time pixel-based correlation filters, as well as deep methods that have top performance but need a high-end GPU. In this paper, we learn to improve the speed of deep trackers without losing accuracy. Our fundamental insight is to take an adaptive approach, where easy frames are processed with cheap features (such as pixel values), while challenging frames are processed with invariant but expensive deep features. We formulate the adaptive tracking problem as a decision-making process, and learn an agent to decide whether to locate objects with high confidence on an early layer, or continue processing subsequent layers of a network. This significantly reduces the feedforward cost for easy frames with distinct or slow-moving objects. We train the agent offline in a reinforcement learning fashion, and further demonstrate that learning all deep layers (so as to provide good features for adaptive tracking) can lead to near real-time average tracking speed of 23 fps on a single CPU while achieving state-of-the-art performance. Perhaps most tellingly, our approach provides a 100X speedup for almost 50% of the time, indicating the power of an adaptive approach. Deep convolutional layers Template frame Current frame Easy frame Agent Output at this layer or forward to the next layer Hard frame Figure 1. Learning policies for adaptive tracking with deep feature cascades. The agent decides whether we can accurately locate objects on an early layer of the cascade. This avoids the need to always wait for the last layer to make decisions, saving a substantial amount of feed-forwarding time. designing a tracker that has both high accuracy and high speed. Real-time tracking methods like TLD [21] and correlation filters [19] usually rely on low-level features that are not descriptive enough to disambiguate target and background. Several recent works [15, 27, 11, 31, 10] overcome this limitation by learning correlation filters on hierarchical deep features, but the real-time capacity largely fades away. Other deep trackers [41, 30, 40] take full advantage of the end-to-end learning, and fine-tune the deep network online to achieve top performance. However, even on a high-end GPU, the speed of such trackers is usually around 1 fps which is too slow for practical use. 1. Introduction Visual Object Tracking (VOT) is a fundamental problem in vision. We consider the single object tracking task, where an object is identified in the first video frame and should be tracked in subsequent frames, despite large appearance changes due to object scaling, occlusion, and so on. VOT is the basic building block of many time-critical systems such as video surveillance and autonomous driving. Thus it is important for a visual tracker to meet the strict constraints of time and computational budget, especially on mobile or embedded computing architectures where real-time analysis perception is often required. Although much progress has been made in the tracking literature, there still exist tremendous challenges in Two recent deep trackers [3, 18] significantly boost their speed by deploying fixed convolutional networks without any online learning. During tracking, the pre-trained network is simply traversed in a feed-forward pass for similarity evaluation or location regression, allowing to track objects at real-time speeds on GPU. Nevertheless, on a modern CPU, smartphone or tablet with much less computing power, such deep trackers can only process a couple of frames per second, far below the normal video frame-rate 30 fps. Obviously, the major computational burden comes from the forward pass through the entire network, and can 1 be larger with deeper architectures. We aim to improve both the accuracy and speed of deep tracking, instead of trading off the quality for speed by e.g. using compressed models [1]. We propose to learn to speed up deep trackers in an adaptive manner. Our adaptive approach builds on the observation that the tracking complexity varies across frames. For example, it is usually effective to use features from the last layer of a deep network to track objects undergoing large appearance change (e.g. abrupt motion) - as these higher-level features are more tolerant to dramatic appearance variation [27]. However, when the object is visually distinct or barely moves, early layers are in most scenarios sufficient for precise localization - offering the potential for substantial computational savings. In the extreme case, the “zeroth” pixel-level layer may suffice for such easy frames, as evidenced by the success of pixelbased correlation filters. Such an adaptive strategy crucially depends on making the right decision - should the tracker stop at the current feature layer or keep computing features on the next layer? To this end, we learn an agent to automatically achieve this goal as illustrated in Fig. 1. The agent learns to find the target at each layer, and decides if it is confident enough to output and stop there. If not, it moves forward to the next layer to continue. This is equivalent to learning a “timely” decision policy to choose the optimal layer for tracking. We will show such policy learning is much more robust than heuristically thresholding the response map of the current active layer. It is also in sharp contrast to the layer selection mechanism in [40], which only selects from two fixed levels of convolutional layers after the entire forward pass is finished. Instead, we formulate this problem as a decision-making process, making sequential decisions with early stopping ability. Specifically, we learn the policy in a reinforcement learning [29] fashion during the training phase, and simply apply the policy for adaptive tracking at test time. In doing so, we are able to provide a speedup of around 10× (on average) over the baseline deep tracker [3] and achieve even higher accuracies on existing OTB [43] and VOT [22] tracking benchmarks. Perhaps most tellingly, our approach provides a 100× speedup almost 50% of the time, indicating the power of an adaptive approach - it turns out that most frames are quite easy to track! Accuracy is improved because each layer of the network is directly trained to be informative for adaptive tracking, similar to the past approaches for “deep supervision” [25]. Concretely, our adaptive tracker works by defining object templates across multiple layers of a network, including the “zeroth” pixel layer. Templates are evaluated across Region of Interest (ROI) with a convolutional filter. We use fast correlation filters [19] to compute response maps for lower layers of our network, where Fourier processing significantly speeds up the convolutional procedure (Fig. 1). We refer to our approach as EArly-Stopping Tracker (EAST). On a single CPU, it has a near real-time average speed of 23.2 fps, and is about 190 fps almost 50% of the time. This makes it the first CPU-friendly deep tracker among the top benchmark performers. It is worth noting that our policy learning method is quite generic. Further, it is readily applied to train end-to-end with an existing deep network for other time-critical vision tasks besides visual tracking. 2. Related Work Real-time tracking and correlation filters: Visual tracking methods can rely on either generative (e.g. [32]) or discriminative (e.g. [16]) models. Discriminative models are often found to outperform in accuracy by discriminating the target from background. Such trackers can usually run fast using hand-crafted features (e.g. HOG [9]) and various learning methods of P-N learning [21], structured SVM [16], multi-expert entropy minimization [45], and correlation filter [4]. Among them, Discriminative Correlation Filter (DCF)based methods [4, 19] are the family of tracking methods with high efficiency and high accuracy as well. The fast speed of DCF is due to its efficient use of all spatial shifts of a training sample by exploiting the discrete Fourier transform. The pioneering MOSSE [4] and improved Kernelized Correlation Filter (KCF) [19] trackers can operate at 669 fps and 292 fps respectively on a single CPU, which far exceeds the real-time requirement. Recent advances of DCF have achieved great success by the use of multi-feature channels [14, 15, 27, 11, 31, 10], scale estimation [13, 26], longterm memory [28], and boundary effect alleviation [11, 12]. However, with increasing accuracy comes a dramatic decrease in speed (0.3–11 fps on a high-end GPU). Tracking by deep learning: Directly applying correlation filters on the multi-dimensional feature maps of deep Convolutional Neural Networks (CNNs) is one straight-forward way of integrating deep learning for tracking. Usually the deep CNN is fixed, and the DCF trackers trained on every convolutional layer are combined by a hierarchical ensemble method [27] or an adaptive Hedge algorithm [31]. Danelljan et al. [15] recently introduced a continuous spatial domain formulation C-COT, to enable efficient integration of multi-resolution deep features. C-COT and its improved ECO [10] can achieve top performance in the VOT challenge [22], but the tracking speed is still slow due to the high dimension of the entire deep feature space. Another category of deep trackers [41, 30, 40] update a pre-trained CNN online to account for the target-specific appearance at test time. Such trackers usually take a classification approach to classify many patches and choose the highest scoring one as the target object. Unfortunately, online training and exhaustive search severely hamper their speed. The top performing tracker MDNet [30] has the GPU speed of only 1 fps or so. Recent advances include using Recurrent Neural Networks (RNNs) [20, 7] to model temporal information using an attention mechanism, but the speed remains slow. One common reason for the slow speed of the abovementioned deep trackers is they always conduct a complete feed-forward pass to the last CNN layer. This ignores the fact that the tracking complexity differs across varying conditions. One of our conclusions is that most frames in current video benchmarks are rather easy. For those frames, forwarding to only early layers may suffice. In principle, such an insight can be used to speed up many recent realtime deep trackers, such as GOTURN [18] (165 fps on GPU) and SiamFC [3] (86 fps on GPU), to make them more CPU-friendly with near frame-rate speed. Feature selection in tracking: Good features are important to tracking. The initial DCF trackers were limited to a single feature channel, e.g. a grayscale image in MOSSE [4]. The DCF framework was later extended to multi-channel features such as HOG [13, 19], Haar-like features [16], binary patterns [21] and Color Attributes [14]. Generally, the hand-crafted features are cheap to compute, but they are not discriminative enough to handle complex tracking scenarios. Many recent deep trackers (e.g. [41, 30]) exploit the semantically robust features from the last CNN layer (fullyconnected). However, spatial details of the tracked object are lost in the last layer which is not optimal for visual tracking. Danelljan et al. [11] found the first convolutional layer is very suitable for tracking. Other works [15, 27, 31, 10] choose to utilize all the hierarchical convolutional features, where early layers can preserve high spatial resolution and deep layers are more discriminative. In this paper, we make the best of hand-crafted and deep convolutional features in a cascaded structure, and learn an agent to select a minimum sequence of feature layers for fast tracking purposes. Unlike FCNT [40] that selects features from two pre-defined layers only after a complete forward pass, our selection is sequential and can stop early at any layer with enough confidence. Feature cascades: CNNs are a natural cascaded architecture with increasingly abstract feature representations. Contemporary works either improve the cascade’s optimality by deep layer supervision [25], or stack multiple CNNs into a deeper cascade for coarse-to-fine [35] or multi-task [8] predictions. Our work differs in learning a decision policy of using only early feature layers in a cascade, and in combining feature cascades and reinforcement learning [36] to achieve this goal. Our approach bears some similarity to the “attentional cascade” structure [39] which uses a cascade of gradually more complex classifiers. The difference is that attentional cascade aims to use early classifiers to eliminate easy negatives and reduce the burden of complex classifier evaluation, whereas we aim to use these early layers to make a strong decision as early as possible. Reinforcement learning for tracking: Reinforcement learning (RL) [29, 36] is capable of learning good policies to take a sequence of actions based on trail and error. It has been successfully applied to vision tasks (e.g. object detection [5]) by treating them as a decision-making process. For visual tracking, there are two recent works that use RL to temporally attend target regions [44] and select suitable template [6]. Our work is the first one to use RL to learn an early decision policy for speeding up deep tracking. 3. Method We use the deep Convolutional Neural Network (CNN) as a rich feature cascade for robust visual tracking. Given an input video frame at time t, the process of tracking an object with bounding box can be seen as the problem of maximizing a confidence function fl : Ht → R over the set of hypothesized object regions Rt : rt∗ = arg max fl (r), r∈Ht (1) where Ht usually consists of regions around the object location in the previous frame, and l denotes the feature layer at which the object confidence is obtained. Many deep trackers [41, 30] exploit l = L as the last fully-connected layer or last convolutional layer of AlexNet [24] or deeper VGGNet [34]. Other trackers [15, 27, 31, 10] exploit a full layer set {l}L l=1 of all the convolutional layers to take the best advantage of feature hierarchies. While these methods have been successful and effective, they are still slow and may not be needed when tracking during easy frames. In this paper, we propose a principled sequential method that accumulates confidence fl=1,...,Lt (r) from an adapt tively small set of feature layers {l}L l=1 in order to track efficiently without losing accuracy. For example, Lt = 2 convolutional layers C1–C2 can suffice to track a distinct face in Fig. 3; but for a blurry face in cluttered backgrounds, we may want to gather more evidence from a deeper layer C5. During such sequential search, our method needs to adjust the bounding box to progressively localize the object using more and more robust features. Ideally, we would like to minimize the number of forwarded layers necessary to locate objects. A naive approach might use a heuristic for determining when to advance to the next layer: for example, one might advance if the maximum value of the current response map is below a threshold. However, defining such heuristics can be difficult when the response map is ambiguous or has multiple peaks (e.g. Fig. 3). Instead, we propose to train a functioning agent end to end by deep reinforcement learning [29]. The agent learns the action and search policy (including the early stopping criteria) so that it can make decisions under uncertainty to reach the target. Fig. 2 shows the framework of our policy learning. Pixel layer Template frame Cropped image 127*127*3 HOG layer Deep convolutional layers C2 C3 C4 C1 6*6*128 Q-Net C5 Global scaling Action history 32-d 17*17 Averaged 128-d 128-d score map 8 actions Q-Net Q-Net Q-Net Q-Net Q-Net Q-Net Local scaling Reshaped Fully-connected layers score 289-d Current frame Search image 255*255*3 No scaling 22*22*128 Stop Figure 2. System framework of our EArly-Stopping Tracker (EAST) by policy learning. C1 C2 C1 C2 C3 C4 C5 1 1,2 0 0 4,5 2,3 1 0 3 2 0,1 More examples of used layers and box actions Figure 3. Example sequences of actions taken to adjust the bounding box scale over deep convolutional layers. Box translation is determined by the relative position of maximum score on the score map. Note each score map is averaged with all the maps from preceding layers. Our agent learns to wisely act upon the score maps. It terminates the search early if the confidence score is highly peaked. When the score map is ambiguous (e.g. has two peaks for a blurry face in cluttered backgrounds), the agent postpones decision and enlarges the box according to the more unambiguous score map at the next layer. Further actions of box scaling are performed with stronger confidence. 3.1. Fully-Convolutional Siamese Network In this section, we review the fully-convolutional Siamese tracker [3], which we make use of for its good trade-off between speed (86 fps on GPU) and accuracy in exploiting deep convolutional layers. Other trackers are either too slow [15, 27, 31, 10] as a baseline (<11 fps on GPU), or do not provide explicit response maps (direct regression instead in [18]) for our policy learning purposes. The Siamese network [3] is trained offline to locate a 127 × 127 template image z within a larger 255 × 255 search image x. A similarity function is learned to compare the template image z to a candidate region of the same size in the search image x, so as to return a high score for the truth region and low score otherwise. Such similarity evaluation is fully-convolutional with respect to x in the network, much more efficient than exhaustive search. Specifically, a cross-correlation layer is proposed to compute similarity for all translated sub-regions in x in one pass: Fl (z, x) = ϕl (z) ∗ ϕl (x) + vI, (2) where ϕl is the convolutional feature embedding at layer l = 5 (i.e. C5 layer), and v ∈ R is an offset value. Here Fl (·, ·) is a confidence score map of size 17×17, as opposed to the single confidence score fl (·) in Eq. 1. During tracking, this Siamese network simply evaluates the similarity online between the template image in previous frame and the search region in current frame, leading to fast speed. The relative position of the maximum score multiplied by the stride of the network, gives the object translation from frame to frame. Our goal is to learn an early decision policy from these confidence score maps Fl , to adaptively 1) predict the object bounding box across layers, and 2) stop early at a layer l < 5 when sufficiently confident about the prediction. Note the score map dimension depends on the size of input feature maps. The score map Fl on an early layer l will have a larger resolution than 17 × 17, and so we downsample to this size to facilitate learning. Also, SiamFC [3] searches over multiple scales of the search image to handle scale variations. We only work on the original scale for high efficiency, and learn to gradually infer the box scale from heatmaps computed during the single forward-pass. 3.2. Learning Policies with Reinforcement Learning We treat the tracking problem as a Markov Decision Process (MDP) where an agent can make a sequence of actions across feature layers, see Fig. 2. This agent learns when to stop advancing to the next layer, as well as how to to gradually deform a bounding box once per layer, which is initialized to the estimated box from the previous frame. The ultimate goal is to output a tight box around the object with as few as layers as possible. The challenge is to be able to operate with rewards that rule out supervision at each step, and at the same time, to minimize the number of steps to locate the target given its changing complexity. We train the agent in a reinforcement learning (RL) setting to learn decision policies. In the RL setting there are a set of states S and actions A, and a reward function R. At each step on layer l, the agent examines the current state Sl , and decides on the action Al to either transform the box or stop with a box output. The action Al is expected to reduce the uncertainty in localizing the object, and receives positive or negative reward Rl reflecting how well the current box covers the object and how few steps are used before action stop. By maximizing the expected rewards, the agent learns the best policy to take actions and can explicitly balance accuracy (search for more layers) and efficiency (stop early if highly confident). Actions: Our action set A includes seven anisotropic scaling transformations to a box and one stop action to terminate search. We do not use the agent to predict the centroid of the box, and instead compute it directly from the relative position of the maximum score on score map as in Eq. 1. We also experimented with requiring the agent to report box translations, but found directly inferring them from the score maps simplified training and increased convergence (due to the smaller space of actions A). For scaling actions, there are two global and four local (modify aspect ratio) transformations as shown in Fig. 2. Similar to [5], any of these actions makes a combined horizontal and vertical change to the box by a factor of 0.2 relative to its current size. We also introduce a no scaling action that does not scale the box at all. This action allows the agent to postpone decision when the current score map is ambiguous or a decision simply cannot be made. Fig. 3 exemplifies this case where two peaks exist on the first-layer score map for a cluttered scene. The agent decides not to act on this map but waits for a more unambiguous map at the next layer to act (enlarge the box). States: The state Sl is represented as a tuple (Fl0 , hl ), where Fl0 is the score map and hl is a vector of the history of taken Pl actions. We define Fl0 = k=1 Fk /l as the average of the score map at current layer l and all its preceding maps from earlier layers. Thus Fl0 not only encodes the currently observed confidence but also the confidence history that has been collected. This is found to work better empirically than using Fl only, and is similar to the hypercolumn representation [17] with the benefits of simultaneously leveraging information from early layers that capture fine-grained spatial details and deeper layers that capture semantics. Also, the resulting robustness comes at a negligible cost in averaging the score maps already obtained. The history vector hl keeps track of the past 4 actions. Each action in the vector is represented by an 8-dimensional one-hot vector or zero vector (when processing the first layer). We find that including hl ∈ R32 helps stabilize action trajectories. Rewards: The reward function R(Sl−1 , Sl ) reflects the localization accuracy improvement from state Sl−1 to Sl after taking a particular action Al . The accuracy is measured by the Intersection-over-Union (IoU) between the predicted box b and ground-truth box g. We can formally define IoU as IoU (b, g) = area(b ∩ g)/area(b ∪ g). Since each state Sl is associated with a box bl , the reward function is then defined following [5]:   sign(IoU (bl , g) − IoU (bl−1 , g)), Al 6= stop R(Sl−1 , Sl ) = +3, IoU (bl , g) ≥ 0.6, Al = stop   −3, IoU (bl , g) < 0.6, Al = stop, (3) where the accuracy improvement is quantized to ±1 if the current action is not stop. This reward scheme encourages positive transformations even with small accuracy improvement. If there is no transformation for further improvement, or if the agent already arrives at the last layer l = L, the action should be stop. In this case, the reward function will penalize the predicted box bl with IoU less than 0.6. Note such a reward scheme implicitly penalizes a large number of layers l since Q-learning (detailed next) models the expected future rewards when deciding on an action (positive or negative). Deep Q-learning: The optimal policy of selecting actions should maximize the sum of expected rewards on a given frame. Since we do not have a priori knowledge about the correct layer or action to choose, we address the learning problem through deep Q-learning [29]. This approach learns an action-value function Q(Sl , Al ) to select the action Al+1 that gives the highest reward at each layer. The learning process iteratively updates the action-selection policy by: Q(Sl , Al ) = R + γ max Q(S 0 , A0 ), 0 A (4) where Q(S 0 , A0 ) is the future reward and γ the discount factor. The function Q(S, A) is learned by a deep Q-Network as illustrated in Fig. 2. It takes as input the state representation S, i.e. the reshaped score vector and action history vector. The network consists of two 128-dimensional fullyconnected layers, finally mapping to 8 actions. Each fully- connected layer is randomly initialized, and is followed by ReLU and dropout regularization [24]. Note during training, we not only update the weights of the Q-Network, but also the pre-trained convolutional layer when the agent receives rewards on that layer. Similar to the deeply-supervised net [25], our approach provides a direct target signal for learning the feature representation at each layer, so as to improve performance of our adaptive tracker. Testing with learned policies: During testing, the agent does not receive rewards or update the Q-function. It just follows the decision policy to deform the box and output it when a stop action is performed. Our agent takes only 2.1 steps on average to locate the target between frames on the OTB-50 dataset [42]. This means we can correctly track most objects by using 2 deep layers. Only for those hard frames, the search degenerates to a full forward pass. The overall tracking algorithm follows SiamFC [3] to search over candidate regions around the estimated location from previous frame. As a result, one order of magnitude speedups are achieved over traditional non-adaptive deep trackers. Set aside efficiency, the policies are still appealing in that they mimic the dynamic attention mechanism by progressively attending to the target region in feature cascades. Implementation details: We use the AlexNet [24]-like convolutional architecture as in SiamFC [3]. The whole network including Q-Net is trained on the ImageNet Video [33] trainval set (4417 videos) for 50 epochs, each completed after the agent has interacted with all training images. We make use of an -greedy [36] optimization during Qlearning, taking a random action with probability  to encourage exploration of diverse action policies. We anneal  linearly from 1 to 0.1 over the first 30 epochs, and fix  to 0.1 in the remaining 20 epochs. We use a learning rate 1e-3, discount factor γ = 0.9 and batch size 64. The network parameters are updated with direct stochastic gradient descent using MatConvNet [38] on a single NVIDIA GeForce Titan X GPU and an Intel Core i7 CPU at 4.0GHz. 3.3. Learning with cheap features Our policy learning can be applied to a feature cascade with any type of feature layers. We explore the use of additional cheap feature layers after the pixel-layer and before the more expensive deep layers. Inspired by the success of correlation filters defined on multi-channel HOG layers [9], we explore an optional HOG layer. In theory, other cheap features such as Color Attributes [14] might apply. When processing our pixel and HOG layers, we make use of fast correlation filters. We specifically make use of the Dual Correlation Filter (DCF) [19], which exhibits a good tradeoff in CPU speed (270+ fps on CPU) and accuracy compared to alternatives such as the Kernelized Correlation Filter (2× slower), STC [46] (350 fps but lower accuracy) and SRDCF [12] (5 fps). 4. Results Before comparing our EArly-Stopping Tracker (EAST) with prior works, we first conduct an ablation study of some of its important variants. We compare EAST to variants using different feature cascades—EAST-Pixel-HOG, EASTHOG, EAST-Pixel, as well as the baseline SiamFC [3]. Fig. 4 shows the probability of stopping at different feature layers and the associated speed on OTB-50 [42]. EAST indeed learns to use only early layers over 70% of the time, as they suffice when tracking easy frames. Hard frames are processed with additional layers, degenerating to a full forward-pass (as in SiamFC) only when needed. Fig. 5 illustrates some examples of easy and hard frames (and their stopped layers) on particular video sequences. Obviously, the earlier layer we stop at, the greater the speedup. EAST-Pixel-HOG (which lacks a pixel or HOG layer) is about 4.5× faster than SiamFC on average, running at 10.7 fps on a CPU and 467.3 fps on a GPU. Recall that Pixel and HOG layers can be processed by fast correlation filters [19] that run at 278 fps and 292 fps, respectively. By adding such layers, EAST achieves a CPU speedup of 10× on average and 100× for those easy frames. Even though HOG may incur an additional feature computation cost compared to the Pixel layer, it produces a greater speedup because it enables more accurate pruning, and so is selected more often by EAST. Note that the average GPU speed is sacrificed to some extent due to larger reliance on CPU computations. Nonetheless, EAST still produces a near-real-time CPU rate of 23.2 fps, making it quite practical for CPU-bound tracking (required on many embedded devices). Table 1 (top) summarizes the speed and accuracy of these EAST variants. We use the Overlap Success (OS) rate as a strict evaluation metric for accuracy. One-Pass Evaluation (OPE) is employed to compare accuracies in terms of Area Under the Curve (AUC) of OS rates for different threshold values. Table 1 shows that the use of more and stronger feature layers systematically improves the accuracy of the tracker, reaching to the AUC score of 0.638 for our full EAST model. Speed is also improved due to the larger computational savings on cheap layers. Table 1 (middle) further compares other EAST variants: • EASTlast : tracking by always forwarding to the last feature layer. • EASTth : feed-forwarding if the maximum value of the current response map is lower than 0.9. Scale is determined by the size of thresholded region. • EAST−ch : policy learning without confidence hisPl tory, i.e. we use Fl instead of Fl0 = k=1 Fk /l. • EAST−ah : policy learning without action history hl . EASTlast is similar to the baseline SiamFC [3] in that all layers are always processed, but differs in that it is trained Avg. speed SiamFC [3] with 3 scale-search CPU: 1.9 fps GPU: 86 fps Stopping probability 1 EAST-Pixel-HOG CPU: 10.7 fps GPU: 467.3 fps EAST-HOG CPU: 14.5 fps GPU: 197.6 fps EAST-Pixel CPU: 19.9 fps GPU: 223.6 fps EAST CPU: 23.2 fps GPU: 158.9 fps 0.6 0.6 0.6 0.6 0.8 0.5 0.5 0.5 0.5 0.6 0.4 0.4 0.4 0.4 0.3 0.3 0.3 0.3 0.2 0.2 0.2 0.2 0.1 0.1 0.1 0.1 0 0 0 0.4 0.2 0 C1 C2 C3 C4 C5 C1 C2 C3 C4 C5 CPU 10.7x 5.2x 3.3x 2.4x 2.0x speedup Pixel C1 C2 C3 C4 C5 CPU 98.7x 7.9x 4.5x 3.2x 2.2x 1.9x speedup HOG C1 C2 C3 C4 C5 CPU 101.3x 8.0x 4.5x 3.2x 2.2x 2.0x speedup 0 CPU speedup Pixel HOG C1 C2 C3 C4 C5 98.7x 62.4x 6.2x 3.9x 2.9x 2.0x 1.8x Figure 4. Stopping probability at different feature layers and the resulting speed on OTB-50 dataset [42]. For each of the 5 models with different feature cascades, we show the average speed on both CPU and GPU (top), and the CPU speedup ratio over the baseline SiamFC [3] (1.9 fps) at each layer (bottom). SiamFC searches over multi-scaled images to handle scale variations, while we predict the scale in a single forward-pass, leading to a constant tracking speedup. Our early stopping policy further accelerates tracking (4.5×) by adaptively using early layers as compared to SiamFC that always uses the last layer C5. When the early layer is the cheap HOG or pixel layer with fast CPU speed (270+ fps), we are able to increase the average CPU speed by one order of magnitude and operate at around 100× faster speed for almost 50% of the time. Our full model EAST operates at near real-time speed 23.2 fps on CPU. On the other hand, it is found the more reliance on CPU computations will generally increase the CPU speed, but also lose the benefits of GPU speed to some acceptable extent. C5 C4 C3 C2 C1 HOG Pixel C5 C4 C3 C2 C1 HOG Pixel Table 1. One-Pass Evaluation (OPE) results of Area Under the Curve (AUC) score and speed (CPU/GPU fps) on OTB-50 dataset. frame 50 frame 100 50 150 100 200 150 250 200 300 350 250 400 300 450 Method EAST-Pixel-HOG EAST-HOG EAST-Pixel EASTlast EASTth EAST−ch EAST−ah SiamFC [3] EAST AUC 0.619 0.628 0.631 0.645 0.597 0.621 0.610 0.612 0.638 Speed 10.7 / 467.3 14.5 / 197.6 19.9 / 223.6 3.4 / 76.3 45.7 / 172.0 21.1 / 137.2 17.4 / 127.6 1.9 / 86 23.2 / 158.9 350 Figure 5. Stopped feature layers for the frames in videos David (frames 300-770 of the original video) and Tiger2. Easy and hard frames are marked in green and red. with deep supervision and does not require an image pyramid to model scale. As a result, it both works better and faster (in CPU terms) than SiamFC. Interestingly, EASTlast outperforms EAST but is much slower. This suggests that our adaptive strategy is slightly hurting accuracy. EASTth makes use of manually-designed heursitics for stopping, avoiding the need to predict actions with a Q-Net, making it 2× faster on CPU. However, heuristic policies fail to work as well as the one learned by Q-learning. Finally, eliminating scoremap and action histories also hurts performance, likely because such histories stabilize the search process. OTB-50 results: The OTB-50 [42] benchmark contains 50 video sequences for evaluation. Table 2 compares the AUC scores of our EAST and the state-of-the-art trackers: TLD [21], GOTURN [18], Struck [16], KCF [19], DSST [13], MEEM [45], RTT [7], FCNT [40], Staple [2], HDT [31], HCF [27], LCT [28], SiamFC [3] and SINT [37]. CPU/GPU speeds (fps) are also reported. EAST achieves the highest AUC of any method. It does so while being significantly faster. For example, the runner-up SINT is 4 fps on GPU, while other GPU-based real-time deep trackers (GOTURN and SiamFC) are significantly slower on a CPU (2-3 fps). When compared to fast correlation trackers e.g. KCF defined on cheap features, EAST is significantly more accurate while still maintaining near-real-time speeds. This is in contrast to correlation filters defined on deep features (HCF and HDF), who are not real-time even with a GPU. The Staple tracker combines HOG and color features by a ridge regression, while FCNT tracks by selecting features from deep layers. EAST outperforms both in terms of accuracy. The Recurrently Targetattending Tracker (RTT) trains Recurrent Neural Networks (RNNs) to capture attentions as a regularization on correlation filter maps. However, it is noticeably slower and less accurate than EAST (3 fps on CPU, and an AUC of 0.588). OTB-100 results: The OTB-100 [43] dataset is the extension of OTB-50 and is more challenging. We test on the full 100 videos to compare with recent related trackers: RDT [6], SRDCF [12], MDNet [30], C-COT [15], ECO [10]. Table 3 summarizes their AUC scores and Table 2. The Area Under the Curve (AUC) score for One-Pass Evaluation (OPE), and speed (fps, * indicates GPU speed, otherwise CPU speed) on the OTB-50 dataset. The best results are shown in bold. Method AUC Speed TLD [21] 0.438 22 GOTURN [18] 0.450 165* Struct [16] 0.474 10 KCF [19] 0.516 172 DSST [13] 0.554 24 MEEM [45] 0.572 10 RTT [7] 0.588 3 Table 3. The Area Under the Curve (AUC) score for One-Pass Evaluation (OPE), and speed (fps, * indicates GPU speed, otherwise CPU speed) on the OTB-100 dataset. The best results are shown in bold. Method AUC Speed RDT [6] 0.603 43* SRDCF [12] 0.605 5 MDNet [30] 0.685 1* C-COT [15] 0.686 0.3 ECO [10] 0.694 6 EAST 0.629 23/159* FCNT [40] 0.599 3* Staple [2] 0.600 80 HDT [31] 0.603 10* HCF [27] 0.605 11* LCT [28] 0.612 27 SiamFC [3] 0.612 86* SINT [37] 0.625 4* EAST 0.638 23/159* Table 4. Raw scores and speed for our method and the top 4 trackers of the VOT-15 Challenge. The * indicates the speed in EFO units. The CPU/GPU speeds for our EAST are given. Tracker MDNet [30] EAST DeepSRDCF [11] EBT [47] SRDCF [12] Accuracy 0.60 0.57 0.56 0.47 0.56 Robustness 0.69 1.03 1.05 1.02 1.24 Overlap 0.38 0.34 0.32 0.31 0.29 Speed (fps) 1 21/148 <1* 5 5 35 30 Overall Rank DCF 25 20 15 10 EAST 5 0 0 EASTlast 0.2 0.4 0.6 0.8 1 Runtime (EFO/frame) Figure 6. Overall rank and runtime of our tracker (red) in comparison to the 38 trackers from VOT-14 Challenge. We show the results of EAST and its two variants that always forward to the first (i.e. DCF on pixel layer) and last (i.e. EASTlast ) feature layer. CPU/GPU speeds. EAST is close to state-of-the-art in terms of accuracy and is the fastest among the top performers on OTB-100. The MDNet and correlation-filter-based advances SRDCF, C-COT and ECO all suffer from low speed, while EAST does not sacrifice the run-time performance for accuracy. RDT is a related Reinforcement Learning (RL)based method that selects the best template to track a given frame. EAST (which also makes use of RL) proves more accurate while being 4× faster on a GPU. VOT-14 results: We test EAST on the 25 videos from the VOT-14 [23] Tracking Challenge. Fig. 6 compares the speed and overall rank of EAST to that of the 38 tracker submissions. For the rank computation, the challenge first evaluates the average accuracy and robustness ranking (refer to the supplementary material for details) for each tracker, and then averages the two rank values to get the overall rank. As can be seen from Fig. 6, EAST achieves the best accuracyspeed trade-off, outperforming all competitors in the overall rank. We also show two EAST variants at the opposite extreme: always forwarding to the first (i.e. DCF [19] on pixel layer) and last (i.e. EASTlast ) feature layer. Our EAST is able to adaptively take advantage of the speed and accuracy benefits of two variants by policy learning. VOT-15 results: The VOT-15 [22] Tracking Challenge has 60 testing videos chosen from a pool of 356. Trackers are automatically re-initialized five frames after failure (zero overlap). Table 4 compares our EAST with the top 4 trackers in terms of accuracy and speed (using the vot2015challenge toolkit). Our testing speed on this benchmark are 21 fps on CPU and 148 fps on GPU, making EAST the fastest and most CPU-friendly tracker among the top performers. We achieve comparable accuracy scores to MDNet, while providing a 148× speedup on GPU, indicating the power of our adaptive policy learning approach. 5. Conclusion This paper proposes an adaptive approach to tracking with deep feature cascades. Our fundamental insight is that most frames in typical tracking scenarios turn out to be easy, in that simple features (such as pixels or HOG) suffice. That said, some challenging frames do require “heavy-duty” invariant feature processing. The challenge is in determining which is which! By formulating the tracking problem as a decision making process, we learn a reinforcementlearning agent that can make such distinctions. Importantly, the agent learns to do so in an iterative manner, making efficient use of a feature cascade that proceeds to deeper layers only when the current one does not suffice. This dramatically reduces the feed-forwarding cost for those easy frames (by 100X), leading to an overall significant speedup for tracking. Such a policy learning method is appealing in that it is trained end-to-end and can be applied to any deep network designed for time-critical tasks. Acknowledgment: This research was supported in part by NSF grants CNS-1518865, IIS-1618903 and IIS-1526033 and the DARPA grant HR001117C0051. Additional support was provided by the Intel Science and Technology Center for Visual Cloud Systems (ISTC-VCS), Google, and Autel. Any opinions, findings, conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the view(s) of their employers or the above-mentioned funding sources. References [1] A. Alfvén. Object tracking with compressed networks. Master’s Theses in Mathematical Sciences, Mathematics (Faculty of Technology) and Numerical Analysis, Lund University, 2016. [2] L. Bertinetto, J. Valmadre, S. Golodetz, O. Miksik, and P. H. S. Torr. Staple: Complementary learners for real-time tracking. In CVPR, 2016. [3] L. Bertinetto, J. Valmadre, J. F. Henriques, A. Vedaldi, and P. H. S. Torr. Fully-convolutional siamese networks for object tracking. In ECCV, 2016. [4] D. S. Bolme, J. R. Beveridge, B. A. Draper, and Y. M. Lui. Visual object tracking using adaptive correlation filters. In CVPR, 2010. [5] J. C. Caicedo and S. Lazebnik. Active object localization with deep reinforcement learning. In ICCV, 2015. [6] J. Choi, J. Kwon, and K. M. Lee. Visual tracking by reinforced decision making. arXiv preprint, 1702.06291, 2017. [7] Z. Cui, S. Xiao, J. Feng, and S. Yan. Recurrently targetattending tracking. In CVPR, 2016. [8] J. Dai, K. He, and J. Sun. Instance-aware semantic segmentation via multi-task network cascades. In CVPR, 2016. [9] N. Dalal and B. Triggs. Histograms of oriented gradients for human detection. In CVPR, 2005. [10] M. Danelljan, G. Bhat, F. S. Khan, and M. Felsberg. ECO: efficient convolution operators for tracking. In CVPR, 2017. [11] M. Danelljan, G. Häger, F. S. Khan, and M. Felsberg. Convolutional features for correlation filter based visual tracking. In ICCV Workshop, 2015. [12] M. Danelljan, G. Häger, F. S. Khan, and M. Felsberg. Learning spatially regularized correlation filters for visual tracking. In ICCV, 2015. [13] M. Danelljan, G. Hger, F. Shahbaz Khan, and M. Felsberg. Accurate scale estimation for robust visual tracking. In BMVC, 2014. [14] M. Danelljan, F. S. Khan, M. Felsberg, and J. v. d. Weijer. Adaptive color attributes for real-time visual tracking. In CVPR, 2014. [15] M. Danelljan, A. Robinson, F. Shahbaz Khan, and M. Felsberg. Beyond correlation filters: Learning continuous convolution operators for visual tracking. In ECCV, 2016. [16] S. Hare, A. Saffari, and P. H. S. Torr. Struck: Structured output tracking with kernels. In ICCV, 2011. [17] B. Hariharan, P. Arbelez, R. Girshick, and J. Malik. Hypercolumns for object segmentation and fine-grained localization. In CVPR, 2015. [18] D. Held, S. Thrun, and S. Savarese. Learning to track at 100 fps with deep regression networks. In ECCV, 2016. [19] J. F. Henriques, R. Caseiro, P. Martins, and J. Batista. Highspeed tracking with kernelized correlation filters. TPAMI, 37(3):583–596, 2015. [20] S. E. Kahou, V. Michalski, and R. Memisevic. RATM: recurrent attentive tracking model. In CVPRW, 2017. [21] Z. Kalal, K. Mikolajczyk, and J. Matas. Tracking-learningdetection. TPAMI, 34(7):1409–1422, 2012. [22] M. Kristan, J. Matas, A. Leonardis, M. Felsberg, L. Čehovin, G. Fernandez, T. Vojir, G. Häger, G. Nebehay, and R. Pflugfelder. The visual object tracking VOT2015 challenge results. In ICCV 2015 Workshop, 2015. [23] M. Kristan, R. Pflugfelder, A. Leonardis, J. Matas, L. Cehovin, G. Nebehay, T. Vojir, G. Fernandez, A. Lukezic, and A. Dimitriev. The visual object tracking VOT2014 challenge results. In ECCV 2014 Workshop, 2014. [24] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In NIPS, 2012. [25] C.-Y. Lee, S. Xie, P. W. Gallagher, Z. Zhang, and Z. Tu. Deeply-supervised nets. In AISTATS, 2015. [26] Y. Li and J. Zhu. A scale adaptive kernel correlation filter tracker with feature integration. In ECCV Workshop, 2014. [27] C. Ma, J.-B. Huang, X. Yang, and M.-H. Yang. Hierarchical convolutional features for visual tracking. In ICCV, 2015. [28] C. Ma, X. Yang, C. Zhang, and M.-H. Yang. Long-term correlation tracking. In CVPR, 2015. [29] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, and A. K. Fidjeland et al. Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 2015. [30] H. Nam and B. Han. Learning multi-domain convolutional neural networks for visual tracking. In CVPR, 2016. [31] Y. Qi, S. Zhang, L. Qin, H. Yao, Q. Huang, J. Lim, and M.-H. Yang. Hedged deep tracking. In CVPR, 2016. [32] D. A. Ross, J. Lim, R.-S. Lin, and M.-H. Yang. Incremental learning for robust visual tracking. IJCV, 77(1-3):125–141, 2008. [33] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 115(3):211–252, 2015. [34] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. [35] Y. Sun, X. Wang, and X. Tang. Deep convolutional network cascade for facial point detection. In CVPR, 2013. [36] R. Sutton and A. Barto. Reinforcement Learning: An introduction. MIT Press, 1998. [37] R. Tao, E. Gavves, and A. W. M. Smeulders. Siamese instance search for tracking. In CVPR, 2016. [38] A. Vedaldi and K. Lenc. Matconvnet: Convolutional neural networks for matlab. In ACM MM, 2015. [39] P. Viola and M. Jones. Rapid object detection using a boosted cascade of simple features. In CVPR, 2001. [40] L. Wang, W. Ouyang, X. Wang, and H. Lu. Visual tracking with fully convolutional networks. In ICCV, 2015. [41] N. Wang, S. Li, A. Gupta, and D. Yeung. Transferring rich feature hierarchies for robust visual tracking. arXiv preprint, 1501.04587, 2015. [42] Y. Wu, J. Lim, and M.-H. Yang. Online object tracking: A benchmark. In CVPR, 2013. [43] Y. Wu, J. Lim, and M. H. Yang. Object tracking benchmark. TPAMI, 37(9):1834–1848, 2015. [44] D. Zhang, H. Maei, X. Wang, and Y.-F. Wang. Deep reinforcement learning for visual object tracking in videos. arXiv preprint, 1701.08936, 2017. [45] J. Zhang, S. Ma, and S. Sclaroff. MEEM: robust tracking via multiple experts using entropy minimization. In ECCV, 2014. [46] K. Zhang, L. Zhang, Q. Liu, D. Zhang, and M.-H. Yang. Fast visual tracking via dense spatio-temporal context learning. In ECCV, 2014. [47] G. Zhu, F. Porikli, and H. Li. Tracking randomly moving objects on edge box proposals. arXiv preprint, 1507.08085, 2015. Table 5. Network architecture and convolutional layer specifics. Supplementary Material Layer A. Algorithmic Details Network architecture: We use the exact convolutional architecture of SiamFC [3]. The convolutional layers C1−C5 and their parameter details are given in Table 5. Note maxpooling is employed for the convolutional layers C1 and C2. We use the nonlinear ReLU function [24] after every convolutional layer except for C5. Batch normalization is inserted after every linear layer. Deep Q-learning: During deep Q-learning [29], the optimal action-value function Q(Sl , Al ) obeys the Bellman equation: it is optimal to select the action A0 that maximizes the expected reward Q(Sl , Al ) = R + γ max Q(S 0 , A0 ), 0 A 0 C1 C2 C3 C4 C5 Input conv1 pool1 conv2 pool2 conv3 conv4 conv5 Support Stride 11×11 3×3 5×5 3×3 3×3 3×3 3×3 2 2 1 2 1 1 1 Method LCT [28] 0.612 0.594 0.518 27 AUC-OPE AUC-TRE AUC-SRE Speed 0 A (6) The gradient of this loss function with respect to the network weights θl is as follows:  h 0 0 Q(S , A ; θ ) − Q(S , A ; θ ) ∇θl V (θl ) = E R + γ max l−1 l l l A0 i ·∇θl Q(Sl , Al ; θl ) . (7) B. Discussions and Results The main idea of our EArly-Stopping Tracker (EAST) is to track easy frames using only early layers of a deep feature cascade, e.g. pixel values, while hard frames are processed with invariant but expensive deep layers when needed. An attached video demo.mp4 exemplifies such tracking policies in video sequences. To further validate the advantages of EAST in both accuracy and speed, we compare with the top 3 trackers on OTB-50 [42] in terms of speed, and AUC score for One-Pass Evaluation (OPE), Temporal Robustness Evaluation (TRE) and Spatial Robustness Evaluation (SRE). Table 6 shows that EAST achieves the highest scores under all evaluation metrics, while maintaining fast tracking speed. Table 7 shows the detailed ranks of accuracy RA and robustness RR under baseline and region noise experiments in VOT-14 challenge. The two experiments evaluate trackers with the initial target location from ground truth and that perturbed with random noises. The table also lists the overall rank Ro and running speed to compare EAST with the best 3 trackers out of 38 submitted ones. It is evident that Search activation 255×255 123×123 61×61 57×57 28×28 26×26 24×24 22×22 Chans. 3 96 96 256 256 192 192 128 Table 6. The Area Under the Curve (AUC) score for One-Pass Evaluation (OPE), Temporal Robustness Evaluation (TRE) and Spatial Robustness Evaluation (SRE), and speed (fps, * indicates GPU speed, otherwise CPU speed) on the OTB-50 dataset. The best results are shown in bold. (5) where Q(S , A ) is the future reward and γ the discount factor. Since the action-value function is approximated by a QNet with weights θ, the Q-Net can be trained by minimizing the loss function V (θl ) at each iteration l,  2  0 0 V (θl ) = E R + γ max Q(S , A ; θ ) − Q(S , A ; θ ) . l−1 l l l 0 Template activation 127×127 59×59 29×29 25×25 12×12 10×10 8×8 6×6 SiamFC [3] 0.612 0.621 0.554 86* SINT [37] 0.625 0.643 0.579 4* EAST 0.638 0.662 0.591 23/159* Table 7. The accuracy RA , robustness RR and average R ranks under baseline and region noise experiments in VOT-14. Ro is the overall (averaged) ranking for both experiments, which is used to rank the 38 trackers in the main paper. Our CPU/GPU speeds are reported in fps, while the speeds for the top 3 trackers are in EFO units, which roughly correspond to fps (e.g. the speed of the NCC baseline is 140 fps and 160 EFO). Tracker EAST DSST SAMF KCF RA 4.95 5.41 5.30 5.05 baseline RR 5.42 11.93 13.55 14.60 R 5.19 8.67 9.43 9.82 region noise RA RR R 5.11 4.73 4.92 5.40 12.33 8.86 5.24 12.30 8.77 5.17 12.49 8.83 Ro 5.06 8.77 9.10 9.33 Speed 22/155 7.66 1.69 24.23 EAST is one of the fastest trackers, while outperforming other top performers in the overall rank. Template Update It is worth noting that, in our feature cascade we explore the pixel and HOG layers before expensive deep layers. When processing the cheap pixel and HOG layers, we make use of fast correlation filters [19]. A correlation filter w with the same size of image patch x is learned by solving the Ridge Regression loss function X
2 min wT xi − yi + λkwk2 , (8) w i where yi is the target response value, and λ is the regularization parameter. Solving this loss function is fast due to the efficient use of all shifted patches xi by exploiting the discrete Fourier transform. Besides fast speed, the correlation filter has another benefit of updating the template w over time. However, this adaptive merit is not preserved for deep layers. Recall that for the deep convolutional layers C1 − C5, we AUC 0.641 EAST 0.638 0.635 C1 C2 C3 C4 C5 Layer Figure 7. The Area Under the Curve (AUC) score for One-Pass Evaluation (OPE) of template update for deep layers C1 − C5. follow SiamFC [3] to compute the similarity of a template image z to all translated regions in search image x by Fl (z, x) = ϕl (z) ∗ ϕl (x) + vI, (9) where ϕl is the convolutional feature embedding at layer l, and v ∈ R is an offset value. Here ϕl (z) can be treated as the convolutional template to compute the target responses, but is fixed to ϕl (zt=1 ) from the first frame and is never updated during tracking. Then a question naturally arises: can we improve the performance by updating the template for deep layers? To this end, we conduct the following experiment on OTB-50: we simply update ϕl (zt ) as ϕl (zt−1 ) from the previous frame, and record the accuracy if we separately update the convolutional layer l from C1−C5. Fig.7 shows the AUC score when we update the template for each deep layer. Marginal gains are obtained on lower layers C1−C2, suggesting that they are less invariant and so would need to be updated more often. On the other hand, updating the top layer C5 leads to no difference, which is actually in line with the observations by SiamFC [3] that always uses this invariant top layer for tracking. In the future, we can consider how to learn to update template online rather than just use the previous frame. Another promising direction is to further speedup the deep convolutional process by adopting the Fourier transform techniques. 

1cs.CV

arXiv:1608.04809v1 [math.GR] 16 Aug 2016 GENERALIZING THE FUTURAMA THEOREM J. ELDER AND O. VEGA Abstract. The 2010 episode of Futurama titled The Prisoner of Benda centers around a machine that swaps the brains of any two people who use it. The problem is, once two people use the machine to swap brains with each other, they cannot swap back. In this article, we present a new proof of this theorem and also a generalization of it to what would happen if, instead, the machine swapped cyclically the brains of p characters, where p is prime. 1. Introduction In the 2010 Writers Guild Award-winning episode of Futurama, The Prisoner of Benda [5], Professor Farnsworth and Amy build a machine that can swap the brains of any two people. The two use the machine to swap brains with each other, but then discover that once two people have swapped with each other, the machine does not swap them back. More characters get involved until the group is thoroughly mixed up, and they start looking for ways to return to their own heads. Clearly, the problem of undoing what the machine has done may be studied using permutations; each brain swap can be described by a transposition in Sn , where n is the number of characters involved in the brain-swapping. Hence, the problem consists in writing the inverse of a permutation as a product of transpositions that (1) are all distinct, and (2) were not already used in constructing the original permutation. The solution given in the episode relies on involving two additional people; these two characters were not involved in any of the original swaps, and so have no restrictions on who they can swap with now. For example, using the additional people x and y, we have (1 2)−1 = (x y)(2 x)(1 y)(2 y)(1 x) Note that the transpositions used in the product above are all distinct, and that none of them is an element of S2 , which is where the original element (1 2) was taken from. This can be done in general for any permutation, and leads to the following theorem. Theorem 1 (Keeler, 2010). Let n ∈ N, n ≥ 2. The inverse of any permutation in Sn can be written as a product of distinct transpositions in Sn+2 \ Sn . 2010 Mathematics Subject Classification. Primary 20B30; Secondary 20F05. Key words and phrases. Permutations, products of cycles, Futurama. The first author thanks the support of the Division of Graduate Studies at California State University, Fresno in the form of the Robert & Norma Craig Graduate Scholarship and also for travel support. She would also like to thank the College of Science and Mathematics at California State University, Fresno for travel support. 1 Although Ken Keeler (writer and executive producer for Futurama at the time) never published the proof of his theorem, the proof given in the episode as well as a write up of the idea may be found in [3]. This was, most probably, the first time in which a theorem was proved for the purpose to advance the narrative of a TV show, and thus it was all over the internet [1, 4, 7] and was also mentioned in books, such as Simon Singh’s The Simpsons and Their Mathematical Secrets [6]. Of course, this aroused the interest of mathematicians (the authors of this article included), and so in 2014, Evans, Huang, and Nguyen [2] proved that Keeler’s solution to the problem is optimal in the sense that it uses the minimal number of cycles and the minimal number of additional elements. In that article, they also gave necessary and sufficient conditions on m and n for the identity permutation to be expressible as a product of m distinct transpositions in Sn . Although in the title of this article we used The Futurama Theorem, for the remainder of this work we will refer to Theorem 1 as Keller’s Theorem. 2. Reproving Keeler’s Theorem We want to reprove Keeler’s Theorem using a different, and possibly less optimal, method. The goal is to present a method that can be generalized to larger cycles in Sn . As in Keeler’s proof, we will need additional elements that could not have been used in the permutation we wish to work with. As expected, we will first study how to write the inverse of a cycle using ‘new’ transpositions to later putting all this together using that every permutation can be written as the product of disjoint cycles. Definition 2. Fix n ∈ N and let x, y ∈ N be such that x, y ≫ n. We consider the set that we will use to define the elements of Sn+2 to be {1, 2, . . . , n, x, y}. Throughout this paper, we will consider, without loss of generality, the k-cycle (1 2 · · · k) instead of a generic k-cycle. We start by setting some notation. Definition 3. Let m ∈ N such that m is odd and m ≤ n − 2. And let δm ∈ Sn+2 be defined as m Y (i + 1 y)(i + 2 y)(i x)(i + 1 x). δm = i=1 i odd Notice that δm , for any 1 ≤ m, will not contain any repeated transpositions. In fact, every number 1 ≤ j ≤ m will appear at most twice in the product: once in (j x) and once in (j y). Also, we can see that (x y) is not in the product. Lemma 4. Let k ≥ 7 be odd and such that k ≡ 1 (mod 3). If σ is a k-cycle then σ −1 = δk−2 . Proof. We write k = 1 + 6t, for t ∈ N. We will prove the result by induction on t. For t = 1, we get: (1 2 3 4 5 6 7)−1 = (6 9)(7 9)(5 8)(6 8)(4 9)(5 9)(3 8)(4 8)(2 9)(3 9)(1 8)(2 8) = δ5 We now assume the result for all k-cycles, where k = 1 + 6s and s < t. Now let σ be a j-cycle, where j = 1 + 6t. We let k = j − 6 and write (2.1) (1 2 · · · j)−1 = (j − 6 j − 5 j − 4 j − 3 j − 2 j − 1 j)−1 (1 2 · · · k)−1 From the induction hypothesis, we get (1 2 · · · k)−1 = δj−8 , 2 and from the base case we get (j − 6 j − 5 j − 4 j − 3 j − 2 j − 1 j)−1 = j−2 Y (i + 1 y)(i + 2 y)(i x)(i + 1 x). i=j−6 i odd Putting all this together in Equation (2.1) yields:   j−2 Y (1 2 · · · j)−1 =  (i + 1 y)(i + 2 y)(i x)(i + 1 x) δj−8 = δj−2 , i=j−6 i odd which is what we wanted to obtain.  We will use this lemma to prove all remaining cases. First for k odd. Lemma 5. Let k be an odd number and σ be a k-cycle. 1. If k ≥ 5 and k ≡ 2 (mod 3), then σ −1 = (x y)(k − 2 x)(k − 1 x)(k x)δk−4 . 2. If k ≥ 3 and k ≡ 0 (mod 3), then σ −1 = (x y)(k x)δk−2 . Proof. Let k be an odd number and σ be a k-cycle. 1. For k = 5, we see that (1 2 3 4 5)−1 = (6 7)(3 6)(4 6)(5 6)(2 7)(3 7)(1 6)(2 6) For k > 5 and k ≡ 2 (mod 3), we write (1 2 . . . k) = (1 2 . . . k − 4)(k − 4 k − 3 k − 2 k − 1 k) Since k − 4 is odd and k − 4 ≡ 1 (mod 3) we use Lemma 4 to get the desired result. 2. For k = 3, we have (1 2 3)−1 = (4 5)(3 4)(2 5)(3 5)(1 4)(2 4), and when k > 3 and k ≡ 0 (mod 3), we know that k − 2 ≡ 1 (mod 3), with k − 2 odd. Hence (1 2 . . . k) = (1 2 . . . k − 2)(k − 2 k − 1 k). At this point we use Lemma 4 to get the desired result.  Now we will proceed in a similar manner for the cases where k is even. Lemma 1. If 2. If 3. If 6. Let k be a positive even number, and σ be a k-cycle. k ≥ 6 and k ≡ 0 (mod 3), then σ −1 = (k y)(k − 1 x)(k x)δk−3 . k ≥ 4 and k ≡ 1 (mod 3), then σ −1 = (x y)(k − 1 x)(k x)δk−3 . k ≥ 8 and k ≡ 2 (mod 3), then σ −1 = (k − 2 y)(k − 1 y)(k y)(k − 3 x)(k − 2 x)δk−5 . Proof. Let k be even, and σ be a k-cycle. 1. For k = 6, we see that (1 2 3 4 5 6)−1 = (6 8)(5 7)(6 7)(4 8)(5 8)(3 7)(4 7)(2 8)(3 8)(1 7)(2 7) For k > 6, and k ≡ 0 (mod 3), we write (1 2 . . . k) = (1 2 . . . k − 5)(k − 5 k − 4 k − 3 k − 2 k − 1 k) Since k − 5 is odd and k − 5 ≡ 1 (mod 3) we can use Lemma 4 to get the desired result 3 2. For k = 4, we have (1 2 3 4)−1 = (5 6)(3 5)(4 5)(2 6)(3 6)(1 5)(2 5) and when k > 4 and k ≡ 1 (mod 3), we know that k − 3 ≡ 1 (mod 3), k − 4 is odd. Hence (1 2 . . . k) = (1 2 . . . k − 3)(k − 3 k − 2 k − 1 k) At this point we use Lemma 4 to get the desired result. 3. For k = 8, σ −1 = (6 y)(7 y)(k y)(5 x)(6 x)δ3 When k > 8 and k ≡ 2 (mod 3), we write (1 2 . . . k) = (1 2 . . . k − 7)(k − 7 . . . k − 1 k) Since k − 7 ≡ 1 (mod 3), we use Lemma 4 to get the desired result.  Remark 7. Lemmas 4, 5, and 6 allow us to write any given k-cycle as a product of transpositions. Moreover, such products consist of transpositions that are all distinct from each other, and each one of these 2-cycles is an element in Sn+2 \ Sn . Finally, given a permutation σ ∈ Sn , we write its inverse as a product of disjoint cycles, σ −1 = σ1 σ2 · · · σr , and then we write each σi as a product of transpositions in Sn+2 \ Sn using Lemmas 4, 5, and 6. It is easy to see that the only transposition that could appear more than once in this product is (x y). In this case, each (x y) would appear at the beginning of a product representing a σi (see Lemmas 5 and 6), but this situation is easy to fix as if τ is a permutation in Sn+2 \ Sn then the conjugation (x y)τ (x y) yields a permutation that is equal to τ where all the xs and the ys have been exchanged. Keeler’s Theorem follows. 3. Products of larger cycles Now we are interested in learning what would happen if Professor Farnsworth and Amy had created a machine that swapped several brains in a cyclical way. For example, the machine would do Professor −→ Amy −→ Bender −→ Leela −→ Fry −→ Professor The problem of undoing what the machine does in the example above would be easy if the machine did not have any ‘issues’ and worked even when a set of five people sat in the machine for a second time; we could use that (1 2 3 4 5)−1 = (5 4 3 2 1) or that (1 2 3 4 5)−1 = (1 2 3 4 5)4 to solve the problem. That would be too easy, there would be no TV network that would broadcast an episode featuring such a simple problem/solution (the existence of the show Scooby-Doo may go against this premise). Hence, to make this interesting, we will assume that the machine does not always work when a set of people who have already sat in the machine sits in it again. Our restriction will be technical, and so we will phrase it explicitly in the statement of our results. So, for the rest of this article, we will assume that we are working with a machine that swaps cyclically the brains of p people at a time, where p is an odd prime. Note that now, in terms of permutations, our problem will reside in An , as the brain-swapping machine yields an odd cycle every time it functions. We will start our study with what will end up being a special case: p = 3. 4 Products of 3-cycles. It turns out that, in this case, we only need one extra character to undo the mess created by the machine. Lemma 8. Let n ∈ N, n > 2, and let σ ∈ An be an odd cycle. Then, σ −1 can be written as a product of 3-cycles σ −1 = τ1 τ2 · · · τt where τi ∈ An+1 \ An , for all i = 1, 2, . . . , t, and τi ∈< / τj >, for all i 6= j. Proof. For k = 3 we get (1 2 3) = (x 3 1)(x 1 2), and for every 3 > k ≤ n, k odd, and x ≫ n, consider the product (1 2 · · · k − 1 k) = (x k 1)(x k − 2 k − 1) · · · (x 3 4)(x 1 2). The result is now immediate, as the inverse of any odd cycle is also an odd cycle, and the 3-cycles used in the products considered all fix different sets of elements.  Lemma 9. Let α, β ∈ Sn , n > 2, be two disjoint even cycles. Then, (αβ)−1 can be written as a product of 3-cycles (αβ)−1 = τ1 τ2 · · · τt , where τi ∈ An+1 \ An , for all i = 1, 2, . . . , t, and τi ∈< / τj >, for all i 6= j. Proof. Let β −1 = (a1 a2 · · · ar ) and α−1 = (b1 b2 · · · bs ) be two disjoint cycles, where r and s are even. Note that (αβ)−1 = β −1 α−1 ∈ An can be written as β −1 α−1 = (a1 a2 )(b1 b2 ) (a2 a3 · · · ar )(b2 b3 · · · bs ) = (b2 b1 x)(b1 a2 x)(a2 a1 x)(a1 b1 x) (a2 a3 · · · ar )(b2 b3 · · · bs ). Note that the last two cycles in the product are odd, and thus we can write them as a product of 3-cycles, as indicated in Lemma 8. It is easy to see now that (αβ)−1 can be written as claimed.  The previous two lemmas indicate that we should only need one extra person to undo whatever chaos the 3-brain-swapping machine might have created. We make this result explicit in the following theorem. Theorem 10. Let σ ∈ An , n > 2. Then, σ −1 can be written as a product of 3-cycles σ −1 = τ1 τ2 · · · τt , where τi ∈ An+1 \ An , for all i = 1, 2, . . . , t, and τi ∈< / τj >, for all i 6= j. Proof. Given σ ∈ An , we know we can write σ −1 as a product of disjoint cycles. Moreover, after rearrangement (if needed) we can put all the appearing even cycles (if any) in pairs. Hence, Lemmas 8 and 9 guarantee the desired factorization of σ −1 .  Products of p-cycles, p > 5. In order to give an answer to what would happen if the machine now swapped p brains cyclically, where p is a prime larger than three, we need to set some notation first. Fix n, p ∈ N, where p > 3 is prime and n > 2. Then, we define the following ordered lists of numbers [x]n,p = x1 x2 . . . xp−3 [x]−1 n,p = xp−3 . . . x2 x1 5 where xi ≫ n, for all i = 1, 2, . . . , p − 3. We will use these lists to write cycles. For example (1 2 [x]3,7 ) = (1 2 x1 x2 x3 x4 ), where x1 , x2 , x3 , x4 ≫ 3. Theorem 11. Let n ∈ N and p be a prime such that p > 3 and n > 2. If σ ∈ An , then, σ −1 can be written as a product of p-cycles σ −1 = τ1 τ2 · · · τt , where τi ∈ An+(p−3) \ An , for all i = 1, 2, . . . , t, and τi ∈< / τj >, for all i 6= j. Proof. We will first prove that every element σ ∈ An can be written as a product of p-cycles, then we will prove that the factors used satisfy the conditions claimed in the theorem. We know that every element in An can be written as a product of odd cycles and an even number of even cycles. We will first deal with odd cycles. For k = 3 we can write (1 2 3) = (1 3 2 [x]−1 n,p )(2 1 3 [x]n,p ), and so we need p − 3 new elements to write a 3-cycle as a product of p-cycles. For k > 3 we know that every k-cycle in An can be written as a product of k−1 2 distinct 3-cycles as follows: (3.1) (a1 a2 · · · ak ) = (a1 a2 a3 )(a3 a4 a5 ) · · · (ak−2 ak−1 ak ). So, if σ ∈ An is a k-cycle with k > 3, then we write it as in Equation (3.1). After that we write each of the 3-cycles in that product as a product of two p-cycles, as discussed in the case k = 3 above. Hence, we can write σ as a product of p-cycles by incorporating p − 3 new elements. For products of two even cycles we let σ = (a1 a2 · · · ar ) and τ = (b1 b2 · · · bs ), where r and s are even. Note that στ ∈ An can be written as στ = (a1 a2 )(b1 b2 ) (a2 a3 · · · ar )(b2 b3 · · · bs ) −1 = (b2 b1 a2 [x]n,p )(a2 a1 b1 [x]n,p ) (a2 a3 · · · ar )(b2 b3 · · · bs ) where the first two cycles need p − 3 new elements to have length p, and all the other cycles involved are odd, and thus can be written as products of p-cycles. In order to see that all the p-cycles obtained in the final product describing σ −1 generate distinct subgroups of An+(p−3) we notice that, in the majority of the cases, a pair of these p-cycles will not ‘move’ the same p elements, and thus the groups they generate would intersect trivially. Thus, the only case that should be analyzed is when we get two cycles of the form τ1 = (1 3 2 [x]−1 n,p ) and τ2 = (2 1 3 [x]n,p ). In this case, if we assumed that these elements generate the same subgroups then τ2i = τ1 , for some i. However, if we look at τ2i (2) we can see that the only possibility for τ2i = τ1 to happen would be when i = p − 1, but it is easy to see that τ2p−1 = τ2−1 6= τ1 . This finishes the proof.  We conclude that as long as the brain-swapping machine Professor Farnsworth and Amy build swaps a prime number of brains cyclically, they can always fix the chaos created by incorporating enough extra characters to the mix. References [1] Casey Chan, Futurama writer invented a new math theorem just to use in the show (2010) http://gizmodo.com/5618502/futurama-writer-invented-a-new-math-theorem-just-to-use-in-the-show. [2] Ron Evans, Lihua Huang, Tuan Nguyen, Keeler’s theorem and products of distinct transpositions, Amer. Math. Monthly 121 (2014) 136–144. 6 [3] Hristo Georgiev, The Futurama theorem explained. The Commutator 2 (2010) 18–20. [4] Tony Phillips, Math in the Media, Amer. Math. Soc., Original math on Futurama (2010), http://www.ams.org/news/math-in-the-media/10-2010-media. [5] Previous Nominees & Winners of the Writers Guild Awards (last accessed on 08/15/16). http://awards.wga.org/wga-awards/previous-nominees-winners. [6] Simon Singh, The Simpsons and Their Mathematical Secrets. Bloomsbury, NY, 2013. [7] The prisoner of Benda. The Infosphere, the Futurama Wiki (last accessed on 08/15/16). http://theinfosphere.org/The_Prisoner_of_Benda. Department of Mathematics. Arizona State University. Tempe, AZ 85287. E-mail address: Jennifer.E.Elder@asu.edu Department of Mathematics. California State University, Fresno. 93740. E-mail address: ovega@csufresno.edu 7 Fresno, CA 

4math.GR

Strong converse for the feedback-assisted classical capacity of entanglement-breaking channels arXiv:1506.02228v4 [quant-ph] 3 Nov 2017 Dawei Ding∗ Mark M. Wilde† November 7, 2017 Abstract Quantum entanglement can be used in a communication scheme to establish a correlation between successive channel inputs that is impossible by classical means. It is known that the classical capacity of quantum channels can be enhanced by such entangled encoding schemes, but this is not always the case. In this paper, we prove that a strong converse theorem holds for the classical capacity of an entanglement-breaking channel even when it is assisted by a classical feedback link from the receiver to the transmitter. In doing so, we identify a bound on the strong converse exponent, which determines the exponentially decaying rate at which the success probability tends to zero, for a sequence of codes with communication rate exceeding capacity. Proving a strong converse, along with an achievability theorem, shows that the classical capacity is a sharp boundary between reliable and unreliable communication regimes. One of the main tools in our proof is the sandwiched Rényi relative entropy. The same method of proof is used to derive an exponential bound on the success probability when communicating over an arbitrary quantum channel assisted by classical feedback, provided that the transmitter does not use entangled encoding schemes. 1 Introduction The classical theory of communication is one of the modern successes of applied mathematics [CT91, GK12]. It is arguably one of the foundations of our current information age and provides new ways of thinking about problems in many other fields of study, such as physics and in particular quantum mechanics. The interaction between these two fields is mutual; while some problems in quantum mechanics can be turned into communication problems, the existence of quantum phenomena strongly suggests that we should rethink many aspects of communication theory. Not only does the notion of a quantum state challenge what we mean by “information,” but the possibilities due to quantum mechanics give rise to new classes of communication protocols. With respect to this latter consideration, some fundamental motivating questions for quantum information theory have traditionally been and still are the following: 1. Is the theory of classical communication affected at a fundamental level by the consideration of quantum mechanical phenomena? ∗ Department of Applied Physics, Stanford University, Stanford, California 94305-4090, USA Hearne Institute for Theoretical Physics, Department of Physics and Astronomy, Center for Computation and Technology, Louisiana State University, Baton Rouge, Louisiana 70803, USA † 1 2. Does using quantum states and measurement for classical communication have practical advantages over using classical techniques? In an attempt to answer these questions, one of the primary goals is to study the ability of a quantum channel to communicate classical information, that is, bits. Like many communication problems, this ability is quantified by the notion of channel capacity. The classical capacity C of a quantum channel N is defined to be the maximum rate of communication such that the decoding error probability can tend to zero in the limit of many channel uses. With this definition, one natural question is to determine how to compute the classical capacity. We know that the Holevo-Schumacher-Westmoreland (HSW) theorem [Hol98, SW97] provides a lower bound: CpN q ě χpN q ” sup IpX; Bqρ , (1.1) tpX pxq,ρx u where tpX pxq, ρx u is an ensemble of quantum states such that each ρx can be input to the channel, and IpX; Bqρ ” HpXqρ ` HpBqρ ´ HpXBqρ is the quantum mutual information of the following classical-quantum state: ÿ ρXB ” pX pxq|xyxx|X b N pρx q, (1.2) x where t|xyu is an orthonormal basis for the classical reference system and HpGqσ is the von Neumann entropy of a quantum state σG on system G. The quantity χpN q is called the Holevo information of the channel. The classical capacity CpN q can actually be formally rewritten in terms of the Holevo information as well, via a procedure known as regularization. Before doing so, note that we obtain the HSW theorem by considering only encoding procedures that do not use entangled inputs1 . That is, each quantum system sent through the channel is not entangled with any other system that is sent. To get the classical capacity, it is generally necessary to incorporate entangled inputs into the calculation. The approach given by [Hol98, SW97] is to multiplex the channel such that one use of this multiplexed channel corresponds to multiple uses of the original channel. A multipartite entangled state describing several inputs across different uses of the original channel can now be simulated by the corresponding single input state to the multiplexed channel. We can therefore express the maximum rate for blocks of size n in terms of a Holevo information: 1 χpN bn q. n (1.3) To obtain the classical capacity, we simply allow for inputs entangled across arbitrarily many channel uses: 1 CpN q “ lim χpN bn q. (1.4) nÑ8 n This idea put together with a converse theorem establishes the regularized expression in (1.4) as being equal to the classical capacity. Unfortunately, computing the classical capacity this way is clearly intractable. This prompts us to look for special cases. We observe that the limit in (1.4) is equal to χpN q iff the Holevo information satisfies a tensor-power additivity property (see Appendix A for a brief derivation): @n χpN bn q “ nχpN q. 1 (1.5) The highest possible rate with this restriction, proven by the HSW theorem to be χpN q, is then a lower bound on the classical capacity. 2 Another way to restate the above is that entangled inputs do not increase capacity iff (1.5) is satisfied. However, this is not true for some channels, as was shown in [Has09]. The next question then appears: What characterizes channels that satisfy (1.5)? A sufficient condition for a quantum channel to satisfy tensor-power additivity is for it to be entanglement-breaking (EB) [HSR03]. Let NAÑB denote a quantum channel, where the arrow notation indicates that the channel maps a state of the input system A to a state in an output system B. An EB channel is defined such that for any bipartite state ρAA1 , the output state pNAÑB b idA1 q pρAA1 q (1.6) is separable. That is, the output of an EB channel can be written as a convex sum of product states. Effectively, the channel “breaks” the entanglement between A and A1 . Previous results have established that the Holevo information of EB channels is additive2 [Sho02]. This is intuitive since any entanglement of the inputs is broken by the channel. Following this line of thinking, one could consider making a stronger statement by allowing additional resources to assist the communication but not to the point that entanglement can be established. Indeed, [BN05] proves a generalization of tensor-power additivity for EB channels with noiseless classical feedback. Furthermore, their results show that classical feedback does not increase the capacity of EB channels. There are also possible stronger statements in another direction. The direct part of the original HSW theorem states that if the rate R of communication is less than the Holevo information χ, then there exists a sequence of protocols P n such that the probability of error for such a sequence satisfies lim pe pnq “ 0, (1.7) nÑ8 where n is the number of channel uses. A result from [Hay07] sharpens this claim by showing that there exists a sequence P n such that pe pnq ď 2´kn , (1.8) for some k ą 0 determined by the channel. The converse part of the original HSW theorem can be strengthened in a similar manner. It states that if R ą CpN q, for any sequence of protocols P n , then lim pe pnq ą 0. (1.9) nÑ8 This is known as the weak converse. In contrast, a strong converse is symmetric to the achievability result above and states that regardless of the protocols used, the success probability decreases to zero in the asymptotic limit whenever R ą CpN q. The strong converse can be sharpened as well whenever there is a constant separation between R and CpN q, such that the convergence of the success probability to zero is exponential in n. There are many reasons why we would want to prove a strong converse. First, a strong converse enriches our understanding of the capacity. A strong converse along with an achievability theorem shows that the capacity is a sharp boundary between reliable and unreliable communication regimes. This, amongst other results, indicates that the classical capacity of a quantum channel is a fundamental quantity of interest. Second, a strong converse is more relevant in practice than is the weak converse. A realistic quantum communication scheme has a finite blocklength; that is, 2 Shor proved that the Holevo information of a tensor product of an EB channel with any other channel is equal to the sum of their respective Holevo informations. This form of additivity is stronger than (1.5) and is the one usually found in the literature. 3 the encoding is across a finite number of channel uses. Although the weak converse does provide a lower bound on the probability of error in the non-asymptotic regime, the strong converse improves the bound considerably. It expresses a trade-off between rate, error probability, and blocklength restrictive enough to be easily checked numerically or experimentally. While a classical version of the strong converse is known for arbitrary discrete memoryless classical channels [Wol78, Ari73], it is still open whether or not strong converses hold for memoryless quantum channels. After some early work [ON99, Win99], the strong converse has been proved for special cases, in particular for channels with certain symmetry [KW09], for EB channels [WWY14], and for a wide class of quantum Gaussian channels [BGPWW15]. Given the strong converse results for EB channels [WWY14] and the weak converse for EB channels with feedback [BN05], it is natural to ask if these two statements are true at the same time. We can also ask directly for the strong converse for unentangled inputs when a feedback link is available. These are the main questions that we address in this paper. 2 Summary of Results In this paper, we derive an explicit exponential bound on the success probability of a classical communication scheme that uses an entanglement-breaking channel along with classical feedback. The same method of proof can be used to establish a bound for arbitrary quantum channels with classical feedback, provided that the inputs are not entangled across multiple uses of the channel. When the communication rate exceeds the classical capacity, these exponential bounds immediately imply strong converse theorems for these settings. We now provide an outline of the proof: 1. First, taking as a starting point the general approach of [Nag01], relating hypothesis testing to unassisted communication, we bound the success probability of an arbitrary feedbackassisted classical communication protocol by a sandwiched Rényi relative entropy [MLDS` 13, WWY14]. 2. Next, one of the main observations from [BN05] is that the sender and receiver’s systems are separable at all times throughout such a protocol, whenever the communication channel is entanglement-breaking. We use this fact and an entropy inequality from [Kin03] to split the relative entropy into two terms. The first term is bounded by an α-information radius, which is a measure of the range of states a channel can output. We then equate this to the sandwiched α-Holevo information [WWY14], which is a Rényi generalization of the Holevo information. The second term is bounded via monotonicity by another sandwiched Rényi relative entropy, to which we recursively apply the same argument. 3. This gives the following bound on the probability of success for any finite blocklength n: psucc ď 2´n supαą1 α´1 pR´χ rα pN qq α , (2.1) where R is the rate of communication and N is the EB channel. It follows from previous arguments [WWY14] that when R ą χpN q, the right hand side of (2.1) is a decaying exponential, thereby establishing the strong converse. We provide an alternate (arguably simpler) proof of this fact (similar to those in [MH11, CMW16]) by establishing that the α-Holevo information converges continuously to the Holevo information as α approaches 1. Appendix B includes a brief review of the argument for the weak converse from [BN05]. 4 3 Preliminaries In this section we provide some necessary definitions, concepts, and previous results used in the derivation of (2.1). 3.1 Quantum states, measurements, operator norms, and quantum channels We start with definitions of relevant mathematical notions from quantum mechanics. Given a finitedimensional Hilbert space H, let BpHq denote the algebra consisting of linear operators acting on H. A relevant measure of a operator X is its Schatten α-norm, which is defined as }X}α ” tTr r|X|α su1{α , (3.1) ? where α ě 1 and |X| ” X : X. The set of quantum states is a convex subset of BpHq given by SpHq “ tρ P BpHq : ρ ě 0, Tr ρ “ 1u , (3.2) where the notation ρ ě 0 means that ρ is positive semidefinite. For composite states, we consider the tensor product of two Hilbert spaces HA and HB , denoted by HA b HB . We can obtain from the overall density operator ρAB P SpHA b HB q the reduced density operator ρA corresponding to the quantum state on only the A system by performing a partial trace: ρA ” TrB pρAB q. (3.3) TrB p|x1 yxx2 |A b |y1 yxy2 |B q ” |x1 yxx2 |A xy2 |y1 y (3.4) The partial trace can be defined as for vectors |x1 y, |x2 y, |y1 y, and |y2 y, and then extended by linearity. A state ρAB is separable if it can be written as ÿ ρAB “ ppxqρxA b ρxB , (3.5) x where ppxq is a probability distribution and tρxA u and tρxB u are sets of states. A positive operator-valued measure (POVM) consists of a set tΛm u of positive semidefinite operators indexed by m and corresponding to different measurement results. The set satisfies ř m Λm “ I, which allows us to interpret the quantity pm ” TrpρΛm q (3.6) as the probability of measuring m given a quantum state ρ. We next consider maps on operators and in particular quantum states. A linear map Ψ : BpHA q Ñ BpHB q is called positive if @τ P BpHA q, τ ě 0 implies Ψpτ q ě 0. It is called completely positive if idR b Ψ is positive for an arbitrary auxiliary system R, where idR is the identity map on R. A map Ψ is a quantum channel if it is linear, completely positive, and also trace-preserving. 5 M A’0 X0 B’0 A’2 A’1 E1 A1 N B1 E2 X1 D1 A2 B2 N D2 B’1 E3 X2 A3 B3 N D3 B’2 M̂ Figure 1: A general protocol for feedback-assisted classical communication. The sender manipulates the systems labeled by A and the receiver those labeled by B. Every X system is classical and represents noiseless classical feedback from the receiver to the sender. 3.2 Classical feedback-assisted classical communication protocols Figure 1 depicts the most general three-round protocol for classical feedback-assisted classical communication. The generalization to n rounds is clear. The protocol begins with Alice preparing a classical register M with the message to be sent. This is correlated with some system A10 . Bob uses the classical feedback channel to send a classical system X0 correlated with some quantum system B01 to Alice. The global state is then ÿ ÿ ρM A10 X0 B01 ” pM pmq |my xm|M b ρm b pX0 pxq |xy xx|X0 b ρxB 1 . (3.7) 1 A 0 m 0 x We will track the state conditioned on a particular value m of the message register M , given by ÿ m ρm pX0 pxq |xy xx|X0 b ρxB 1 . (3.8) A1 X0 B 1 ” ρA1 b 0 0 0 0 x Alice performs an encoding EA1 1 X0 ÑA1 A1 . The state at this point is 0 ρm A11 A1 B01 ” 1 EA1 1 X0 ÑA1 A1 0 1 ÿ “ ´ ρm A10 X0 B01 ¯ (3.9) ´ ¯ x pX0 pxqEA1 1 X0 ÑA1 A1 ρm A1 b |xy xx|X0 b ρB 1 , 0 x 1 0 0 (3.10) 1 1 We note that the state ρm A11 A1 B01 is separable with respect to the cut A1 A1 : B0 . Next, Alice uses the EB channel NA1 ÑB1 for the first time by sending system A1 , leading to the state ´ ¯ m . (3.11) ρm ” N ρ 1 1 1 1 A1 ÑB1 A B1 B A A1 B 1 0 1 0 Since the channel is entanglement-breaking and the state before the channel was already separable with respect to the A11 A1 : B01 cut, the state after the channel is fully separable, that is, it is separable across all possible partitions. Now, consider instead that Alice uses an arbitrary channel 6 NA1 ÑB1 but with an input separable across the A11 : A1 cut (such that ρm A11 A1 B01 is fully separable). m Then, the state ρA1 B1 B 1 is again fully separable. We write it in the form 1 0 ρm A1 B1 B 1 “ 1 ÿ 0 y,m pY pyqρy,m b ρy,m B1 b ρB 1 . A1 1 y (3.12) 0 1 Bob now applies the decoding map DB 1 1 , where X1 is the classical system that is sent 1 B0 ÑX1 B1 back to Alice. The state at this point is ´ ¯ 1 m ρm (3.13) A11 X1 B11 ” DB1 B01 ÑX1 B11 ρA11 B1 B01 ÿ ÿ “ pX1 px1 q |x1 y xx1 |X1 b pZ1 |X1 pz1 |x1 qρxA11 ,z1 b ρxB11,z1 . (3.14) x1 z1 1 1 which is fully separable. Hence, after Alice applies the second encoder E 2 and then sends it through the channel, the state will still be fully separable. The key observation here is that if we use an EB channel or an arbitrary channel with separable inputs, the state is always separable across a cut that divides Alice and Bob’s systems. The only difference in subsequent rounds is the final measurement. Say there are n rounds in 1 using a POVM given by tDm u the protocol. At the last round, Bob measures the state ρBn Bn´1 with elements corresponding to different possible messages that Alice sent. 3.3 Rényi relative entropies and bounds on success probability An important classical information theoretic quantity is the Rényi relative entropy, which can be generalized to the quantum case in a number of ways. In this paper, we use the sandwiched quantum Rényi relative entropy [MLDS` 13, WWY14] which is given by ” ´` # ˘ ¯ı 1 p1´αq{p2αq ρσ p1´αq{p2αq α log Tr σ : ρ M σ ^ psupppρq Ď supppσq _ α P p0, 1qq α´1 r α pρ}σq ” D `8 : otherwise (3.15) where α P p0, 1q Y p1, 8q and ρ M σ means ρ, σ are non-orthogonal quantum states. Note that all logarithms in this paper are taken base two. We now recall some properties of the sandwiched Rényi relative entropy. For fixed ρ and σ, the r α pρ}σq is monotone non-decreasing [MLDS` 13]. It also converges to the quantum function α ÞÑ D relative entropy Dpρ}σq [Ume62] in the limit as α Ñ 1 [MLDS` 13, WWY14]: r α pρ}σq “ Dpρ}σq, lim D αÑ1 where " Dpρ}σq ” Tr rρ plog ρ ´ log σqs : supppρq Ď supppσq . `8 : otherwise (3.16) (3.17) Furthermore, it satisfies the data-processing inequality for α P r1{2, 1q Y p1, 8q [FL13, Bei13]; that is, for all quantum channels N , r α pN pρq}N pσqq ď D r α pρ}σq. D 7 (3.18) In particular, consider the following replacement channel which simply replaces the input with some state ω: Rω pρq ” Trpρqω. (3.19) r α pω}ωq “ 0, implying that It is easy to see that for all ω, D r α pρ}σq ě D r α pRω pρq}Rω pσqq “ 0, D (3.20) which shows that the sandwiched Rényi relative entropy is non-negative whenever its arguments are quantum states ρ and σ. Using the sandwiched Rényi relative entropy, we can define the sandwiched α-Holevo information [WWY14] of an ensemble, that is, a classical probability distribution of quantum states, tpX pxq, ρx u as r α pρXR }ρX b σR q rα ptpX pxq, ρx uq ” χ inf D (3.21) σR PSpHR q where ρXR ” ÿ pX pxq|xyxx|X b pρx qR (3.22) x and ρx are states of a system R. With this, we define the α-Holevo information of a quantum channel N as χ rα pN q ” sup χ rα ptpX pxq, N pρx quq . (3.23) tpX pxq,ρx u We also define the α-information radius of a channel N : SpHA q ÞÑ SpHB q as r α pN q ” K inf r α pN pρq}σq. sup D (3.24) σPSpHB q ρPSpHA q r α is monotonically non-decreasing in α, so are χ rα. We note that since D rα and K 4 Strong converse This section is dedicated to proving the main theorem of this paper. We will need the following lemmas. The first is proven using an inequality from [Nag01] in Lemma 5 of [CMW16] via monor α . The second will be used to take advantage of the separability of the quantum state tonicity of D observed in Section 3.2. The third is an equality between α-Holevo information and α-information radius. The fourth states that α-Holevo information and α-information radius respectively tend to the conventional Holevo information and information radius in the limit α Ñ 1, which we prove in Appendix C. Note that this establishes Lemma 3 as a generalization of the equality χpN q “ KpN q [OPW97, SW01, SW02]. Lemma 1. Let α ą 1, ρ, σ P SpHq, and Λ be such that 0 ď Λ ď I. Let p ” Tr rΛρs , q ” Tr rΛσs . Then r α pρ}σq ě D “ ‰ 1 log pα q 1´α . α´1 8 (4.1) (4.2) Lemma 2 ([Kin03, Hol06]). Let PAB be a positive semidefinite separable operator. Such an operator can be written in the following form: ÿ j j CA b DB , (4.3) PAB “ j j j where CA , DB ě 0 for all j. Let PB “ TrA tPAB u and let MA be a completely positive linear map acting on the A system. Then, for all α ě 1, }pMA b idB q pPAB q}α ď να pMA q ¨ }PB }α , (4.4) where να pMA q is the 1 Ñ α norm of MA , defined as να pMA q ” }MA pXq}α . }X}1 X‰0,XPBpHA q sup (4.5) Lemma 3 ([WWY14]). For α ą 1, the α-Holevo information and the α-information radius are the same: r α pN q. rα pN q “ K χ (4.6) Proof. This statement was essentially proved as Lemma 14 of [WWY14], but only for the interval α P p1, 2s. To get the statement for all α ą 1, we actually need to extend Lemma 14 in [WWY14] slightly. This follows from the proof given there and the observations that for all α ą 1, xp1´αq{α r α is jointly quasi-convex. is operator convex, Trtxα u is convex, and D For completeness and convenience, we also give a full proof here, following the steps in the r α pN q ď χ rα pN q for α ą 1. proof of Lemma 14 in [WWY14] closely. We first prove the inequality K p1´αq{2α p1´αq{2α α r α pρ}σq ” Trppσ Defining Q ρσ q q, consider that r α pN q “ inf sup D r α pN pρq}σq K σ 1 r α pN pρq}σq log Q α´1 (4.8) 1 r α pN pρq}σq log inf sup Q σ α´1 ρ (4.9) “ inf sup σ “ (4.7) ρ ρ r α quantity and find that So now we focus on the Q ż r r α pN pρq }σq inf sup Qα pN pρq }σq ď inf sup dµpρq Q σ σ µ ρ ż r α pN pρq}σq “ sup inf dµpρq Q µ σ ÿ r α pN pρx q }σq “ sup inf pX pxqQ tpX pxq,ρx u σ “ sup (4.11) (4.12) x r α pρXB }ρX b σB q inf Q tpX pxq,ρx u σB (4.10) (4.13) The first inequality follows by taking a supremum over all probability measures µ on the set of all states ρ. The first equality is a result of applying the Sion minimax theorem [Sio58]—we can do 9 ş r α pN pρq }σq is linear in the probability measure µ and convex in so because the function dµpρq Q r states σ. Convexity of Qα pN pρq}σq in σ follows because !´ ¯α ) r α pN pρq}σq “ Tr rN pρqs1{2 σ p1´αq{α rN pρqs1{2 Q , (4.14) xp1´αq{α is operator convex for α ą 1 and Trpxα q is convex for α ą 1. The second equality follows by an application of the Fenchel-Eggleston-Caratheodory theorem (see [EGK12], for example): the r α pN pρq}σq is continuous in ρ, which is a density operator acting on a d-dimensional function Q Hilbert space, so that for each µ, there exists a probability distribution pX pxq on no more than d2 letters such that ż ÿ r α pN pρq}σq “ r α pN pρx q }σq . (4.15) dµpρq Q pX pxqQ x r α and by defining The last equality in (4.13) follows from the properties of Q ÿ ρXB ” pX pxq |xy xx|X b rN pρx qsB . (4.16) x So we can then conclude that r α pN q ď K “ 1 r α pρXB }ρX b σB q log sup inf Q α´1 tpX pxq,ρx u σB (4.17) 1 r α pρXB }ρX b σB q log Q α´1 (4.18) sup inf tpX pxq,ρx u σB “χ rα pN q . (4.19) r α pN q ě χ rα pN q is simpler. Consider that The proof of the other inequality K χ rα pN q “ sup r α pρXB }ρX b σB q inf D (4.20) sup r α pρXB }ρX b σq D (4.21) r α p|xy xx| b N pρx q } |xy xx| b σq sup D (4.22) r α pN pρx q }σq sup D (4.23) tpX pxq,ρx u σB ď tpX pxq,ρx u sup ď tpX pxq,ρx u x sup “ tpX pxq,ρx u x r α pN pρq }σq . ď sup D (4.24) ρ r α for α ą 1 [FL13, Bei13], where a The second inequality follows from joint quasi-convexity of D function f is jointly quasi-convex if f pλx1 ` p1 ´ λqx2 , λy1 ` p1 ´ λqy2 q ď max tf px1 , y1 q, f px2 , y2 qu . (4.25) r α pN q ě χ rα pN q. Since the above inequality holds for all states σ, we can conclude that K Lemma 4. For a quantum channel N , the following limits hold: rα pN q “ χpN q lim χ αÑ1 10 (4.26) and r α pN q “ KpN q, lim K (4.27) αÑ1 where KpN q ” inf sup DpN pρq}σq (4.28) σPSpHq ρPSpHq is the information radius. We now state the theorem. Theorem 5. Given any n-round protocol for classical feedback-assisted classical communication over an entanglement-breaking channel N with rate R, the average probability of success is bounded from above by an exponential in n: psucc ď 2´n supαą1 α´1 pR´χ rα pN qq α , (4.29) rα is the α-Holevo information. The same bound holds for an arbitrary channel N given where χ that the encoder does not entangle inputs across different uses of the channel. Proof. We take as a starting point the approach of Nagaoka [Nag01], connecting hypothesis testing with data processing of a Rényi information quantity. Let Pn be such a protocol. We are bounding the average probability of success, so we assume Alice chooses her messages uniformly at random. Using the notation of Section 3.2, the state we have at the final round of the protocol Pn is 1 “ ρM Bn Bn´1 L 1 ÿ |myxm|M b ρm , 1 Bn Bn´1 L m“1 (4.30) where L is the number of possible messages. Following the argument in [CMW16], we can write the success probability as psucc ı ı ” 1ÿ ” m m 1 1 “ Tr DBn B 1 ρBn B 1 “ Tr TM Bn Bn´1 ρM Bn Bn´1 , n´1 n´1 L m (4.31) where 1 TM Bn Bn´1 ” ÿ m |myxm|M b DB 1 nB n´1 m . (4.32) Note that 0 ď T ď I, so that tT, I ´ T u is a POVM. 1 We now consider the state τM Bn Bn´1 defined to be the final state if we had implemented Pn using a replacement channel Rσ instead of the original channel (see Figure 2). The overall state in this alternate scenario has a simple expression since Bob’s states are now independent of m: 1 1 τM Bn Bn´1 “ τM b σBn b τBn´1 , (4.33) where τM “ IM {L. In the above, we remind that Bn is a label for the output system of the nth replacement channel Rσ , and the state of this system is equal to σ. We can then compute the following: ” ı ı 1ÿ ” m 1 1 1 1 Tr TM Bn Bn´1 τM Bn Bn´1 “ Tr DBn B 1 σBn b τBn´1 “ , (4.34) n´1 L m L 11 M A’0 X0 B’0 A’2 A’1 E1 A1 σ B1 X1 D1 E2 A2 σ X2 B2 D2 B’1 E3 A3 σ B3 B’2 D3 M̂ Figure 2: A protocol for feedback-assisted classical communication when using the replacement channel. Notice that the communication line between A1 and M̂ is broken. ř 1 where the last equality follows because m Dm “ I and TrrσBn b τBn´1 s “ 1. This equality is intuitive, since for a replacement channel, the receiver Bob cannot do any better than to guess the input message m at random. Letting T be the operator Λ in Lemma 1, we conclude that for α ą 1, „  1 1 α r 1 1 qě }τM Bn Bn´1 Dα pρM Bn Bn´1 log psucc 1´α , (4.35) α´1 L which can be re-written as follows: α r α pρM B B 1 }τM B B 1 q. logppsucc q ` log L ď D n n´1 n n´1 α´1 (4.36) We note that if the support of the state σ does not contain the support of the image of the channel r α upper bound is trivially equal to `8. So, in what follows, we choose σ such NAÑB , then the D that its support contains the support of the image of the channel in order to guarantee that all quantities involved in the forthcoming proof are finite. r α pρM B B 1 }τM B B 1 q. To do so, consider that We would therefore like to bound D n n´1 n n´1 r α pNAn ÑBn pρM A B 1 q}τM B 1 b σBn q r α pρM B B 1 }τM B B 1 q “ D D n n´1 n n´1 n n´1 n´1 ›˜ › ¸ › ´ ¯› α › p1´αq{p2αq p1´αq{p2αq › 1 “ τM B 1 log › Θ 1´α ˝ NAn ÑBn ρM An Bn´1 τM B 1 › , (4.37) α n´1 n´1 › › α´1 σBn α where ˝ denotes function composition and in the last equality, we define Θ by Θσ pρq ” σ 1{2 ρσ 1{2 , (4.38) and the identity operation on the other systems is implied. We now use the key observation from Section 3.2 (and used in [BN05]) that if N is EB or if Alice uses separable inputs, throughout Pn , 1 Alice and Bob’s systems are always separable. Furthermore, the M system is classical, so ρM An Bn´1 1 is separable with respect to the An : M Bn´1 cut. This implies that p1´αq{p2αq τM B 1 n´1 p1´αq{p2αq 1 ρM An Bn´1 τM B 1 n´1 12 (4.39) is a positive semidefinite separable operator: p1´αq{p2αq p1´αq{p2αq 1 τM B 1 ρM An Bn´1 τM B 1 n´1 n´1 “ p1´αq{p2αq τM B 1 n´1 ÿ “ ˜ ÿ ¸ ppjqρjAn b ρjM B 1 n´1 j ´ p1´αq{p2αq j ppjqρjAn b τM B 1 ρM B 1 n´1 j p1´αq{p2αq n´1 τM B 1 (4.40) n´1 p1´αq{p2αq ¯ τM B 1 . (4.41) n´1 Since conjugation by a positive semidefinite operator is clearly a completely positive map, we can apply Lemma 2 to conclude that › ›˜ ¸ › ´ ¯› › p1´αq{p2αq p1´αq{p2αq › 1 τM B 1 τM B 1 ˝ NAn ÑBn ρM An Bn´1 › › Θ 1´α α n´1 n´1 › › σBn ˜ ¸ α › › › p1´αq{p2αq p1´αq{p2αq › 1 ď να Θ 1´α ˝ N ¨ τ ρ τ › › . (4.42) 1 1 A ÑB M B n n MB n´1 M B α n´1 σBn n´1 α We then have the following chain of inequalities: ´ ¯ r α pρM B 1 }τM B 1 q r α pρM B B 1 }τM B B 1 q ď α log να Θ 1´α ˝ N ` D D n n´1 n n´1 n´1 n´1 σ α α´1 ¯ ´ α r α pρM B B 1 }τM B B 1 q ď log να Θ 1´α ˝N `D n´1 n´2 n´1 n´2 σ α α´1 ´ ¯ α r α pρM B 1 }τM B 1 q ďn log να Θ 1´α ˝N `D 0 0 σ α α´1 ¯ ´ α “n log να Θ 1´α ˝N σ α α´1 ›ˆ › ˙ › › α › “n log sup › Θ 1´α ˝ NAÑB pρA q›› . α α´1 σB ρA PSpHA q α (4.43) (4.44) (4.45) (4.46) (4.47) The first inequality follows by combining (4.37) and (4.42). The second inequality follows from monotonicity of the sandwiched Rényi relative entropy under the partial trace channel. The third r α pρM B B 1 }τM B B 1 q is the relative entropy at round inequality follows by recognizing that D n´1 n´2 n´1 n´2 n ´ 1 of Pn , which allows us to apply the above argument inductively. This is a crucial step of the argument and proves a form of additivity similar to that of (1.5). The first equality is a consequence of the fact that ρM B01 “ τM B01 , since no channels have been applied at that point in the protocol. The last equality follows from the main result of [Aud09] which allows us to take the supremum over quantum states instead of all operators (furthermore from the fact that quantum states have trace equal to one). Hence, we have that ›ˆ › ˙ › › α α › › . pρ q (4.48) log psucc pPn q ` log L ď n log sup › Θ 1´α ˝ N AÑB A › α´1 α´1 σB α ρA PSpHA q α Hence, we have proven that the upper bound holds for all states σB , so we can conclude that ›ˆ › ˙ › › α α › log psucc pPn q ` log L ď n inf sup log › Θ 1´α ˝ NAÑB pρA q›› (4.49) α α´1 σB PSpHB q ρA PSpHA q α ´ 1 σB α r α pN q “ nK (4.50) “ nr χα pN q. (4.51) 13 The first equality comes from the definition of the α-information radius, and the second equality is due to Lemma 3. Now, the protocol uses the channel n times and the rate R of Pn is defined to be the number of bits per channel use, so that R ď logn L . This allows us to introduce R into the inequality: 1 α´1 rα pN qq. log psucc pPn q ď ´ pR ´ χ n α (4.52) Since this is true for all α ą 1, we can take a supremum over α ą 1 and arrive at the bound stated in (4.29). The strong converse itself now follows from Theorem 5 and Lemma 4. Corollary 6 (Strong Converse). The probability of success of any sequence of protocols which use an entanglement-breaking channel with classical feedback at a rate greater than the classical capacity is bounded from above by a decaying exponential. The same is true for arbitrary channels with separable inputs. r α is monotonically increasing in α, so is χ Proof. Recall that since D rα . This along with Lemma 4 implies rα pN q “ lim χ rα pN q “ χpN q. inf χ αą1 αŒ1 (4.53) rα pN q as α Œ 1, there exists a value of α ą 1 such Hence, if R ą χpN q, by the continuity of χ pR ´ χ r that α´1 pN qq ą 0. Then the bound in (4.29) implies an exponential decay of the success α α probability. 5 Conclusion By studying the classical capacity of a quantum channel, quantum information theorists found that an implication of quantum mechanics for classical communication is the possibility of superior encoding schemes that use entanglement, a uniquely quantum phenomenon [Has09]. However, entanglement does not give an advantage for every channel. The results in this and previous papers show that for entanglement-breaking channels, no communication protocols, even with classical feedback, can take advantage of this extra resource. They cannot use it to increase their capacity, nor, as proved in this paper, to even lift the exponentially decaying ceiling on their success probabilities. We perhaps expect this since by definition EB channels destroy entanglement, and classical feedback channels cannot create entanglement. This is the guiding principle behind our proof, in particular the key fact that the transmitter and receiver states are separable throughout the protocol. More generally, it is known that classical feedback can give a large boost to the classical capacity of channels which are not entanglement-breaking in at least two different ways: first, there exist channels for which the Holevo information is small but the classical capacity with feedback included can be quite large [BDSS06]. Similarly, there exist channels for which the classical capacity is small but becomes large when classical feedback is available [SS09]. Thus, showing that the Holevo information is a strong-converse bound suggests that for feedback not to help, we require special channels, such as the entanglement-breaking ones. 14 Going in the opposite direction, it is known that EB channels form a proper superset of classical channels [KHH12]. We thus obtain the result of [PV10] as a direct corollary. Furthermore, since the original posting of our paper as arXiv:1506.02228, an open question that we posed has now been answered — the strong converse exponent from Theorem 5 is tight, due to the results in [MO17]. Hence, we obtain as a special case from this and our result the classical results of [Aug78, DK79, CK82] as well. A possible direction for future research is to ask the same question for Hadamard channels, which are defined to be complements of EB channels. To define the complementary channel, we first explain the interpretation of channel as a model for open quantum dynamics. That is, a channel from system A to system B can be interpreted as the restriction of a unitary interaction with larger system BR, where R is referred to as the “environment.” The complementary channel is the map from A to BR tracing out the B system instead and is itself a channel from A to R. Hence, Hadamard channels break any entanglement with the environment system that is traced out and is thus related to our guiding principle. The strong converse has already been proved for Hadamard channels [WWY14], but with the addition of classical feedback, even a weak converse has yet to be proved. Examples of Hadamard channels include generalized dephasing channels, cloning channels, and the Unruh channel [BHTW10]. Finally, we remark here that it should be possible to use the methods given here and in [CMW16] in order to characterize a particular adaptive hypothesis testing scenario. Suppose that the goal is to distinguish an entanglement-breaking channel from a replacement channel by means of adaptive, separability-preserving channels. Then the optimal strong converse exponent should be given in terms of a quantity similar to that in (4.29). However, we leave the details for future work. Acknowledgements. We are grateful to Patrick Hayden, Milan Mosonyi, and Graeme Smith for insightful discussions about the topic of this paper. We also thank an anonymous referee for many helpful suggestions for improving the paper. MMW acknowledges support from startup funds from the Department of Physics and Astronomy at LSU, the NSF under Award No. CCF-1350397, and the DARPA Quiness Program through US Army Research Office award W31P4Q-12-1-0019. DD acknowledges support from a Stanford Graduate Fellowship. A Additivity Here we justify the statement in (1.5). The backward implication is trivial. As for the forward implication, we first note that given n, for all k P Z` , 1 1 χpN bn q ď χpN bkn q, n kn (A.1) because we can split the kn channels into k blocks of n channels and encode each block independently. In particular, n1 χpN bn q ě χpN q. To show the opposite inequality, we assume the contrapositive: n1 χpN bn q ą χpN q for some n. However, this means 1 1 χpN bkn q ´ χpN q ě χpN bn q ´ χpN q ą 0, kn n for all k P Z` and thus the limit in (1.4) does not converge to χpN q. 15 (A.2) B Weak converse and finite bounds We recall for motivation and the reader’s convenience the argument for the weak converse from [BN05]. Theorem 7. Let Pn be a protocol for classical feedback-assisted classical communication over an entanglement-breaking channel N such that it uses the channel n times, has communication rate R, and has average probability of decoding error ε. Then, it satisfies the following inequality: R ď χpN q ` gpn, εq, (B.1) where χpN q is the Holevo information of the channel and gpn, εq is a real valued function such that lim lim gpn, εq “ 0. (B.2) εŒ0 nÑ8 The same is true for a protocol for communication over an arbitrary channel given that the encoder does not entangle inputs across different uses of the channel. Proof. We use the notation in Section 3.2 and take a general approach that is, for instance, presented in Section 21.5.1 of [Wil16]. Let Φ denote the following “shared randomness” state: 1 ÿ |my xm|M b |my xm|M̂ , ΦM M̂ “ (B.3) |M| m where |M| is the size of the message set M. Now, suppose the information processing task is common randomness generation instead of classical communication. If we require the state ρ at the end of the protocol to be ε1 -close to Φ kΦ ´ ρk1 ď ε1 , (B.4) the Fannes-Audenaert inequality for mutual information [Fan73, Aud07] gives (B.5) nR “ IpM ; M̂ qΦ 1 ď IpM ; M̂ qρ ` f pn, ε q, (B.6) where f pn, ε1 q is some continuous function of n and ε1 with the property: limε1 Œ0 limnÑ8 n1 f pn, ε1 q “ 0. Continuing, we have 1 IpM ; M̂ qρ ď IpM ; Bn Bn´1 qρ “ 1 IpM ; Bn´1 qρ ` (B.7) 1 IpM ; Bn |Bn´1 qρ , (B.8) where we applied the data processing inequality and the chain rule for conditional quantum mutual information. We use the chain rule again to get 1 1 IpM ; Bn |Bn´1 qρ ď IpM Bn´1 ; Bn qρ . (B.9) From Section 3.2, if N is entanglement-breaking or if Alice uses separable inputs, Alice and Bob’s systems are always separable. Hence, the global state before the nth channel use can be written as ÿ 1 ρM A1n An Bn´1 “ pM pmq|myxm|M b ρm (B.10) A1n An B 1 n´1 m ÿ “ pM pmq |my xm|M b m ÿ y 16 m,y pY |M py|mqρm,y . A1 An b ρB 1 n n´1 (B.11) Hence, the state after the channel is given by ¯ ´ ÿ ÿ m,y 1 . “ pM pmq |my xm|M b pY |M py|mqNAn ÑBn ρm,y ρM A1n Bn Bn´1 1 A An b ρB 1 n y m (B.12) n´1 We introduce an auxiliary system Y that labels the separable sum over the index y. That is, Y is chosen such that the global state is ´ ¯ ÿ m,y 1 ” pM pmq pY |M py|mq |my xm|M b |yy xy|Y b NAn ÑBn ρm,y . (B.13) ρM Y A1n Bn Bn´1 1 A An b ρB 1 n m,y n´1 We trace over A1n to get 1 ρM Y Bn Bn´1 “ ÿ m,y pM pmq pY |M py|mq |my xm|M b |yy xy|Y b ρm,y , Bn b ρB 1 where (B.14) n´1 m,y ´ ´ ¯¯ ` ˘ m,y 1 ρm,y ” Tr N ρ “ NAn ÑBn ρm,y An ÑBn An An . Bn A1 An n (B.15) This allows us to argue 1 1 IpM Bn´1 ; Bn qρ ď IpM Y Bn´1 ; Bn qρ “ IpM Y ; Bn qρ ` 1 IpBn´1 ; Bn |M Y qρ (B.16) (B.17) “ IpM Y ; Bn qρ (B.18) ď χpN q, (B.19) where the first inequality is from data processing, the first equality from the chain rule, and the 1 second equality because the Bn´1 Bn system is in a product state when conditioning on M and Y . Finally, given equation (B.15), the state ρM Y Bn is a classical-quantum state of the form: ´ ¯ 1 1 ρM Y Bn “ TrBn´1 ρM Y Bn Bn´1 (B.20) ÿ ` ˘ “ pM pmq pY |M py|mq |my xm|M b |yy xy|Y b NAn ÑBn ρm,y (B.21) An . m,y Hence, the definition of the Holevo information of the channel N gives us final inequality. Putting things together, we find that 1 IpM ; M̂ qρ ď χpN q ` IpM ; Bn´1 qρ ď χpN q ` 1 IpM ; Bn´1 Bn´2 qρ , (B.22) (B.23) where the last inequality follows from the data processing inequality. But now, we recognize the 1 1 quantity IpM ; Bn´1 Bn´2 qρ is of the same form as IpM ; Bn Bn´1 qρ in (B.7), so that we can iterate through the same sequence of arguments to get 1 IpM ; M̂ qρ ď 2χpN q ` IpM ; Bn´2 Bn´3 qρ . (B.24) Continuing all the way back to the first channel use, we find IpM ; M̂ qρ ď nχpN q 17 (B.25) since IpM ; B01 q “ 0 (see (3.7)). Thus, we conclude R ď χpN q ` 1 f pn, ε1 q. n (B.26) Now, the rate for classical communication with n channel uses and error ε is at most that of randomness generation with n uses and error ε1 , where ε1 is some function of ε such that limεŒ0 ε1 “ 0. Hence, gpn, εq ” n1 f pn, ε1 q is what we need. Corollary 8 (Weak Converse). The classical capacity of entanglement-breaking channels with classical feedback is given by the Holevo information. The same is true for arbitrary channels with separable inputs. Proof. This is immediate. C Limit value of the α-Holevo information The arguments here are essentially the same as those in [MH11, CMW16], along with an additional insight from [TWW17, Appendix A]. We first recall a minimax result due to [MH11, Corollary A2]. s where X is a compact topological space, Y is a Lemma 9. Suppose a function f : X ˆ Y Ñ R, s subset of R, and R is the extended real numbers, satisfies 1. @y P Y , f p¨, yq is lower semicontinuous. 2. @x P X, f px, ¨q is monotonic. Then, inf sup f px, yq “ sup inf f px, yq. xPX yPY yPY xPX (C.1) If the first condition was instead 1. @y P Y , f p¨, yq is upper semicontinuous, then inf sup f px, yq “ sup inf f px, yq. yPY xPX xPX yPY (C.2) We now prove Lemma 4. r α pρXR1 , ρX bσR1 q is lower semicontinuous (see, e.g., [CMW16, Proof. For fixed ρXR1 and α, σR1 ÞÑ D r α pρXR1 , ρX b σR1 q is monotone nonAppendix A]). Furthermore, for fixed ρXR1 and σR1 , α ÞÑ D decreasing. We can therefore invoke Lemma 9 with X “ SpHR1 q and Y “ p0, 1q. We use this, the definition of the Holevo information, and properties of the quantum relative entropy to find the 18 following one-sided limit of the α-Holevo information of a quantum channel N : BpHR q Ñ BpHR1 q: rα pN q “ sup χ rα pN q lim χ αÕ1 (C.3) αPp0,1q “ sup sup αPp0,1q tpX pxq,ρx u “ “ “ “ sup rα ptpX pxq, N pρx quq χ sup inf r α pρXR1 }ρX b σR1 q D (C.5) r α pρXR1 }ρX b σR1 q sup D (C.6) tpX pxq,ρx u αPp0,1q σR1 PSpHR1 q sup inf (C.4) tpX pxq,ρx u σR1 PSpHR1 q αPp0,1q sup inf tpX pxq,ρx u σR1 PSpHR1 q DpρXR1 }ρX b σR1 q IpX : R1 qρ sup (C.7) (C.8) tpX pxq,ρx u “ χpN q, (C.9) where ρXR1 ” ÿ pX pxq|xyxx|X b rN pρx qsR1 . (C.10) x The fourth equality follows from applying Lemma 9. The fifth follows from (3.16) and the fact that the sandwiched Rényi relative entropy is monotone non-decreasing in α. r α pN pρR q}σR1 q is monotone To find the other limit, we note that for fixed ρR and σR1 , α ÞÑ D non-decreasing. We then use an idea from [TWW17, Appendix A]. Note that for fixed α and r α pN pρR q}σR1 q is continuous (and thus upper semicontinuous). Let positive definite σR1 , ρR ÞÑ D ε P p0, 1q and define σ pεq ” p1 ´ εq σ ` επ, where π is the maximally mixed state. Thus σpεq is positive definite for σ positive semidefinite. Consider that σpεq ě p1 ´ εq σ, (C.11) Dpρ}σpεqq ď Dpρ} p1 ´ εq σq “ Dpρ}σq ´ log p1 ´ εq . (C.12) which implies that (This is because of the well known fact that σ ď σ 1 implies that Dpρ}σq ě Dpρ}σ 1 q.) Hence, using Lemma 9, Lemma 3, and the identification of the Holevo information as an information radius 19 [OPW97, SW01], we can conclude the following chain of equalities: rα pN q rα pN q “ inf χ lim χ (C.13) r α pN q “ inf K (C.14) αą1 αŒ1 αą1 “ inf inf sup r α pN pρR q}σR1 q D (C.15) ď inf inf sup r α pN pρR q}σpεqR1 q D (C.16) sup r α pN pρR q}σpεqR1 q D (C.17) αą1 σR1 PSpHR1 q ρ PSpH q R R αą1 σR1 PSpHR1 q ρ PSpH q R R “ “ “ ď inf inf σR1 PSpHR1 q αą1 ρR PSpHR q inf sup r α pN pρR q}σpεqR1 q inf D (C.18) inf sup DpN pρR q}σpεqR1 q (C.19) DpN pρR q}σR1 q ´ logp1 ´ εq (C.20) σR1 PSpHR1 q ρR PSpHR q αą1 σR1 PSpHR1 q ρR PSpHR q inf sup σR1 PSpHR1 q ρR PSpHR q “ χpN q ´ log p1 ´ εq . (C.21) The first inequality follows because there is an infimum with respect to σR1 , such that replacing σR1 with σpεqR1 can never decrease the quantity. The fifth equality follows from applying Lemma 9. The sixth equality follows from (3.16) and the fact that the sandwiched Rényi relative entropy is monotone non-decreasing in α. The last inequality follows from (C.12). The last equality uses χpN q “ KpN q. rα pN q ď χpN q. Combined with Given that ε ą 0 was arbitrary, we can conclude that limαŒ1 χ rα pN q ě χpN q for all α ě 1, we conclude that the fact that χ rα pN q “ χpN q. lim χ (C.22) r α pN q “ KpN q. lim K (C.23) αŒ1 This also implies that αŒ1 r α pN q “ KpN q, which follows from a similar line of reasoning: We now show that limαÕ1 K r α pN q “ sup K r α pN q lim K αÕ1 (C.24) αPp0,1q “ sup inf sup r α pN pρR q}σR1 q D (C.25) r α pN pρR q}σR1 q D (C.26) r α pN pρR q}σR1 q sup D (C.27) αPp0,1q σR1 PSpHR1 q ρR PSpHR q “ “ “ inf sup sup σR1 PSpHR1 q αPp0,1q ρR PSpHR q inf sup σR1 PSpHR1 q ρR PSpHR q αPp0,1q inf sup σR1 PSpHR1 q ρR PSpHR q “ KpN q, DpN pρR q}σR1 q (C.28) (C.29) 20 where the third equality follows from Lemma 9 and the fact that for fixed α and ρR , σR1 ÞÑ r α pN pρR q}σR1 q is lower semicontinuous and that the pointwise supremum of lower semicontinuous D functions is lower semicontinuous. References [Ari73] Suguru Arimoto. On the converse to the coding theorem for discrete memoryless channels. IEEE Transactions on Information Theory, 19(3):357–359, May 1973. [Aud07] Koenraad M. R. Audenaert. A sharp continuity estimate for the von Neumann entropy. Journal of Physics A: Mathematical and Theoretical, 40(28):8127, July 2007. arXiv:quant-ph/0610146. [Aud09] Koenraad M. R. Audenaert. A note on the p Ñ q norms of 2-positive maps. Linear Algebra and its Applications, 430(1):1436–1440, February 2009. arXiv:mathph/0505085. [Aug78] Udo Augustin. Noisy channels. Habilitation thesis, Universitat Erlangen-Nurnberg, West Germany, September 1978. [BDSS06] Charles H. Bennett, Igor Devetak, Peter W. Shor, and John A. Smolin. Inequalities and separations among assisted capacities of quantum channels. Physical Review Letters, 96(15):150502, April 2006. arXiv:quant-ph/0406086. [Bei13] Salman Beigi. Sandwiched Rényi divergence satisfies data processing inequality. Journal of Mathematical Physics, 54(12):122202, December 2013. arXiv:1306.5920. [BGPWW15] Bhaskar Roy Bardhan, Raul Garcia-Patron, Mark M. Wilde, and Andreas Winter. Strong converse for the classical capacity of all phase-insensitive bosonic Gaussian channels. IEEE Transactions on Information Theory, 61(4):1842–1850, April 2015. arXiv:1401.4161. [BHTW10] Kamil Brádler, Patrick Hayden, Dave Touchette, and Mark M. Wilde. Trade-off capacities of the quantum Hadamard channels. Physical Review A, 81(6):062312, June 2010. arXiv:1001.1732. [BN05] Garry Bowen and Rajagopal Nagarajan. On feedback and the classical capacity of a noisy quantum channel. IEEE Transactions on Information Theory, 51(1):320–324, 2005. arXiv:quant-ph/0305176. [CK82] Imre Csiszar and Janos Korner. Feedback does not affect the reliability function of a DMC at rates above capacity. IEEE Transactions on Information Theory, 28(1):92– 93, January 1982. [CMW16] Tom Cooney, Milan Mosonyi, and Mark M. Wilde. Strong converse exponents for a quantum channel discrimination problem and quantum-feedback-assisted communication. Communications in Mathematical Physics, 344(3):797–829, June 2016. arXiv:1408.3373. 21 [CT91] Thomas M. Cover and Joy A. Thomas. Elements of Information Theory. WileyInterscience, August 1991. [DK79] Gunther Dueck and Janos Korner. Reliability function of a discrete memoryless channel at rates above capacity. IEEE Transactions on Information Theory, 25(1):82–85, January 1979. [EGK12] Abbas El Gamal and Young-Han Kim. Network information theory. Cambridge University Press, January 2012. [Fan73] Mark Fannes. A continuity property of the entropy density for spin lattice systems. Communications in Mathematical Physics, 31(4):291–294, December 1973. [FL13] Rupert L. Frank and Elliott H. Lieb. Monotonicity of a relative Rényi entropy. Journal of Mathematical Physics, 54(12):122201, December 2013. arXiv:1306.5358. [GK12] Abbas El Gamal and Young-Han Kim. Network information theory. Cambridge University Press, January 2012. arXiv:1001.3404. [Has09] Matthew B. Hastings. Superadditivity of communication capacity using entangled inputs. Nature Physics, 5:255–257, April 2009. arXiv:0809.3972. [Hay07] Masahito Hayashi. Error exponent in asymmetric quantum hypothesis testing and its application to classical-quantum channel coding. Physical Review A, 76(6):062301, December 2007. arXiv:quant-ph/0611013. [Hol98] Alexander S. Holevo. The capacity of the quantum channel with general signal states. IEEE Transactions on Information Theory, 44(1):269–273, January 1998. arXiv:quant-ph/9611023. [Hol06] Alexander S. Holevo. Multiplicativity of p-norms of completely positive maps and the additivity problem in quantum information theory. Russian Mathematical Surveys, 61(2):301–339, 2006. [HSR03] Michal Horodecki, Peter W. Shor, and Mary Beth Ruskai. Entanglement breaking channels. Reviews in Mathematical Physics, 15(6):629–641, August 2003. arXiv:quant-ph/0302031. [KHH12] J. K. Korbicz, P. Horodecki, and R. Horodecki. Quantum-correlation breaking channels, broadcasting scenarios, and finite Markov chains. Physical Review A, 86(4):042319, October 2012. arXiv:1208.2162. [Kin03] Christopher King. Maximal p-norms of entanglement breaking channels. Quantum Information and Computation, 3(2):186–190, 2003. arXiv:quant-ph/0212057. [KW09] Robert Koenig and Stephanie Wehner. A strong converse for classical channel coding using entangled inputs. Physical Review Letters, 103(7):070504, August 2009. arXiv:0903.2838. 22 [MH11] Milán Mosonyi and Fumio Hiai. On the quantum Rényi relative entropies and related capacity formulas. IEEE Transactions on Information Theory, 57(4):2474–2487, April 2011. arXiv:0912.1286. [MLDS` 13] Martin Müller-Lennert, Frédéric Dupuis, Oleg Szehr, Serge Fehr, and Marco Tomamichel. On quantum Rényi entropies: A new generalization and some properties. Journal of Mathematical Physics, 54(12):122203, December 2013. arXiv:1306.3142. [MO17] Milán Mosonyi and Tomohiro Ogawa. Strong converse exponent for classicalquantum channel coding. Communications in Mathematical Physics, 355(1):373–426, October 2017. arXiv:1409.3562v7. [Nag01] Hiroshi Nagaoka. Strong converse theorems in quantum information theory. Proceedings of ERATO Workshop on Quantum Information Science, page 33, 2001. Also appeared in Asymptotic Theory of Quantum Statistical Inference, ed. M. Hayashi, World Scientific, 2005. [ON99] Tomohiro Ogawa and Hiroshi Nagaoka. Strong converse to the quantum channel coding theorem. IEEE Transactions on Information Theory, 45(7):2486–2489, November 1999. arXiv:quant-ph/9808063. [OPW97] Masanori Ohya, Denes Petz, and Noburu Watanabe. On capacities of quantum channels. Probability and Mathematical Statistics-Wroclaw University, 17:179–196, 1997. [PV10] Yury Polyanskiy and Sergio Verdú. Arimoto channel coding converse and Rényi divergence. Proceedings of the 48th Annual Allerton Conference on Communication, Control, and Computation, pages 1327–1333, September 2010. [Sho02] Peter W. Shor. Additivity of the classical capacity of entanglement-breaking quantum channels. Journal of Mathematical Physics, 43(9):4334–4340, October 2002. arXiv:quant-ph/0201149. [Sio58] Maurice Sion. On general minimax theorems. Pacific Journal of Mathematics, 8(1):171–176, 1958. [SS09] Graeme Smith and John A. Smolin. Extensive nonadditivity of privacy. Physical Review Letters, 103(12):120503, September 2009. arXiv:0904.4050. [SW97] Benjamin Schumacher and Michael Westmoreland. Sending classical information via noisy quantum channels. Physical Review A, 56(1):131–138, July 1997. [SW01] Benjamin Schumacher and Michael D Westmoreland. Optimal signal ensembles. Physical Review A, 63(2):022308, January 2001. arXiv:quant-ph/9912122. [SW02] Benjamin Schumacher and Michael D Westmoreland. Relative entropy in quantum information theory. Contemporary Mathematics, 305:265–290, 2002. arXiv:quantph/0004045. 23 [TWW17] Marco Tomamichel, Mark M. Wilde, and Andreas Winter. Strong converse rates for quantum communication. IEEE Transactions on Information Theory, 63(1):715–727, January 2017. arXiv:1406.2946. [Ume62] Hisaharu Umegaki. Conditional expectation in an operator algebra. Kodai Mathematical Seminar Reports, 14(2):59–85, 1962. [Wil16] Mark M. Wilde. From classical to quantum Shannon theory. arXiv:1106.1445v7. [Win99] Andreas Winter. Coding theorem and strong converse for quantum channels. IEEE Transactions on Information Theory, 45(7):2481–2485, 1999. arXiv:1409.2536. [Wol78] Jacob Wolfowitz. Coding Theorems of Information Theory. Springer-Verlag New York, Inc., Secaucus, NJ, USA, 3rd edition, 1978. [WWY14] Mark M. Wilde, Andreas Winter, and Dong Yang. Strong converse for the classical capacity of entanglement-breaking and Hadamard channels via a sandwiched Rényi relative entropy. Communications in Mathematical Physics, 331(2):593–622, October 2014. arXiv:1306.1586. 24 March 2016. 

7cs.IT

NORTH-SOUTH DYNAMICS OF HYPERBOLIC FREE GROUP AUTOMORPHISMS ON THE SPACE OF CURRENTS arXiv:1509.05443v3 [math.GR] 29 Jun 2017 MARTIN LUSTIG AND CAGLAR UYANIK Abstract. Let ϕ be a hyperbolic outer automorphism of a non-abelian free group FN such that ϕ and ϕ−1 admit absolute train track representatives. We prove that ϕ acts on the space of projectivized geodesic currents on FN with generalized uniform North-South dynamics. 1. Introduction Thurston’s groundbreaking work on the classification of surface homeomorphisms (NielsenThurston classification) has triggered much of the progress on the mapping class group and on the Teichmüller space in the past 30 years. This theory has inspired important developments for several related groups; most notably for the group Out(FN ) of outer automorphisms of a non-abelian free group FN of finite rank N ≥ 2. Two rather different spaces on which Out(FN ) acts serve as analogues to Teichmüller space: One is Culler-Vogtmann’s Outer space CVN [CV86], the other is the space of projectivized geodesic currents PCurr(FN ) [Bon91]. An intersection form, generalizing Thurston’s celebrated intersection form for measured laminations [Thu88], intimately intertwines the two spaces [KL09]. The space CVN is finite dimensional, and the action of Out(FN ) on CVN is properly discontinuous. Moreover, CVN has a natural “Thurston boundary”, which gives rise to a compactification CVN . The Out(FN )-action extends to this compactification, which mimics closely Thurston’s compactification of the Teichmüller space via projective measured laminations [BF93], [CL95], [CV86], [Thu88]. On the other hand, the space of projectivized geodesic currents PCurr(FN ) is infinite dimensional, but it is already compact, and has a projective structure. There is a natural “interior” of PCurr(FN ) on which Out(FN ) acts properly discontinuously [KL10a]. The space CVN has received more attention in recent years than PCurr(FN ), due to the connections with various curve complex analogues (see e.g. [BF14], [HM13]). Nevertheless, for several purposes, including some rather fundamental ones, the space PCurr(FN ) seems to have an advantage over CVN . In order get a deeper understanding one often needs to utilize the action of Out(FN ) on both spaces simultaneously, see for instance [BR15], [Ham12]. February 12, 2018. The authors gratefully acknowledge support from U.S. National Science Foundation grants DMS 1107452, 1107263, 1107367 “RNMS: GEometric structures And Representation varieties” (the GEAR Network).” The first author was partially supported by the French research grant ANR-2010-BLAN-116-01 GGAA. The second author was partially supported by the NSF grants of Ilya Kapovich (DMS 1405146 and DMS 1710868) and Christopher J. Leininger (DMS 1510034). 1 2 MARTIN LUSTIG AND CAGLAR UYANIK This paper concerns one of the fundamental features prominently present in NielsenThurston theory: the dynamics of the action of individual elements ϕ ∈ Out(FN ) on PCurr(FN ), and the study of the relationship between aspects of this dynamics and the algebraic structure of ϕ. More precisely, the paper aims to generalize Thurston’s result that any pseudo-Anosov mapping class acts on compactified Teichmüller space with North-South dynamics [Thu88, Iva92] An important result in this direction is already known for both of the above mentioned Out(FN )-analogues of Teichmüller space: If the automorphism ϕ ∈ Out(FN ) is fully irreducible, also known as iwip (“irreducible with irreducible powers”), then the ϕ-action on both, CVN and PCurr(FN ), has uniform North-South dynamics, see [LL03], [Mar95], [Uya14] and [Uya15]. The pairs of poles of these two actions are strongly related to each other through the above mentioned intersection form, see [CHL08], [KL10b], [Uya15]. In this paper we concentrate on the following class of automorphisms: An element ϕ ∈ Out(FN ) is called hyperbolic if the length of any non-trivial conjugacy class grows exponentially under iteration of ϕ. This condition turns out to be equivalent to requiring that ϕ is atoroidal, i.e. ϕ has no non-trivial periodic conjugacy class in FN . Recall that both classes, fully irreducible and hyperbolic automorphisms, can be viewed as natural analogues of pseudo-Anosov mapping classes. In a probabilistic sense, both of these notions are “generic” in Out(FN )[Riv08], as is the case for pseudo-Anosov homeomorphisms in mapping class groups [Riv08, Mah11]. Hyperbolic automorphisms have the advantage that their mapping torus groups FN oϕ Z are Gromov hyperbolic, see [BF92], [Bri00]. However, hyperbolic automorphisms can admit invariant free factors and hence these automorphisms are harder to study than fully irreducible automorphisms. As a consequence of the more complicated algebraic structure, one can not expect a classical North-South dynamics for hyperbolic automorphisms. Counterexamples for both, CVN and PCurr(FN ) are easy to construct. Nevertheless, it turns out that there are still strong North-South dynamical features in the action of any hyperbolic ϕ ∈ Out(FN ) on the space of currents. This paper aims to make this statement precise: We define “generalized North and South poles” in PCurr(FN ), which are finite dimensional simplexes ∆+ (ϕ) and ∆− (ϕ), and show: Theorem 1.1. Let ϕ ∈ Out(FN ) be a hyperbolic outer automorphism with the property that both ϕ and ϕ−1 admit (absolute) train track representatives. Then ϕ acts on PCurr(FN ) with “generalized uniform North-South dynamics from ∆− (ϕ) to ∆+ (ϕ)” in the following sense: Given a neighborhood U of ∆+ (ϕ) and a compact set K ⊂ PCurr(FN ) r ∆− (ϕ), there exists an integer M ≥ 1 such that ϕn (K) ⊂ U for all n ≥ M . Similarly, given a neighborhood V of ∆− (ϕ) and a compact set K 0 ⊂ PCurr(FN ) r ∆+ (ϕ), there exists an integer M 0 ≥ 1 such that ϕ−n (K 0 ) ⊂ V for all n ≥ M 0 . Although there exist hyperbolic automorphisms that do not admit (absolute) train track representatives (see Sections 2.3 and 2.4), the train track assumption in the above theorem turns out to be not really restrictive (see Remark 2.6). Our train track technology (see Section 4) has been purposefully set up in a way that it also applies to more general kind of train track maps. Since the paper is already technically rather loaded, we have refrained from adding such generalizations here. DYNAMICS OF HYPERBOLIC AUTOMORPHISMS OF FN 3 The second dynamics result obtained here concerns single orbits of rational currents, that is, currents given by non-trivial conjugacy classes in FN , which form a dense subset of PCurr(FN ), see Section 2.5: Theorem 1.2. Let ϕ be as in Theorem 1.1, and let µ ∈ Curr(FN ) be a non-zero rational current. Then, after possibly replacing ϕ by a positive power, there are [µ∞ ] ∈ ∆+ (ϕ) and [µ−∞ ] ∈ ∆− (ϕ) such that lim ϕn [µ] = [µ∞ ] n→∞ and lim ϕ−n [µ] = [µ−∞ ]. n→∞ It is important to point out that this pointwise convergence is in general not uniform: On the contrary, it is possible to construct examples of hyperbolic ϕ for which there exist rational currents that linger for an arbitrary long iteration time in an arbitrary close neighborhood of a (fixed) current that is different from the eventual limit current. We now give a brief outline of the paper. We’d like to point out immediately that the sections 3, 4 and 5 are written independently of each other and can be read in arbitrary order. • In Section 2 we recall some background information and set up notation that is needed in the sections to follow. • In Section 3 we define several versions of North-South dynamics (in a general context without reference to Out(FN ) or currents) and derive the main criterion (Propositions 3.3 and 3.4) used in the proof of Theorem 1.1. • Section 4 is entirely devoted to train track technology, addressing issues such as indivisible Nielsen paths (INPs), exponential growth of the number of illegal turns under backwards iteration, etc. Moreover, we recall and investigate the notion of “goodness” which plays a key role in our analysis. We conclude Section 4 with Proposition 4.20 (see also Remark 4.21). This proposition is a crucial tool in our approach, since it permits the restriction of our attention in the convergence approximation to conjugacy classes with large goodness. • In Section 5 we present some standard notions from symbolic dynamics and use the main result from our previous paper [LU17] on limit frequencies of substitutions in order to define the limit simplexes ∆+ (ϕ) and ∆− (ϕ) and show their ϕ-invariance (Corollary 5.12). • In Section 6 the crucial convergence approximation is calculated for conjugacy classes with large goodness (Lemma 6.1), done carefully over several pages. The main results are derived from “putting together the pieces” derived previously. We finish this section with a particular “free products” example of hyperbolic outer automorphisms which are not fully irreducible. Acknowledgements: We would heartfully like to thank Ilya Kapovich for his constant encouragement and support, as well as for many helpful mathematical and organizational suggestions. We would also like to thank Arnaud Hilion and Matt Clay for useful exchanges and helpful comments. We thank the anonymous referee who provided detailed feedback, 4 MARTIN LUSTIG AND CAGLAR UYANIK and prompted us to clarify several points and improve the exposition. The second author is grateful to Chris Leininger for useful discussions, support, and helpful comments. 2. Preliminaries We first recall facts about graphs, paths, and maps between graphs. For an even more detailed discussion of these notions we refer reader to Section 2 of [DKL15]. For a streamlined version which uses the standard conventions the reader is referred to [LU16]. 2.1. Graphs. A topological graph Γ is a 1-dimensional finite cell complex. The 0-cells of Γ are called vertices and the set of vertices is denoted by V Γ. The 1-cells of Γ are called topological edges and the set of topological edges is denoted by Etop Γ. Each topological edge admits exactly two orientations, and we call a topological edge with a choice of orientation an oriented edge. The set of all oriented edges is denoted by EΓ. Given e ∈ EΓ, we denote the edge with the opposite orientation by e. Given a finite set of points in a topological graph Γ, one obtains a topological graph Γ0 as the subdivision, in which the vertex set contains the original vertex set, together with the finite set of points. Let Γ be a topological graph. We choose an orientation for each edge, and denote the set of positively (resp. negatively) oriented edges by E + Γ ( resp. E − Γ). Given an edge e ∈ E + Γ, the initial vertex of e is denoted by o(e) and the terminal vertex of e is denoted by t(e). The graph Γ is equipped with a natural metric called the simplicial metric which is obtained by identifying each edge e of Γ with the interval [0, 1] via a characteristic map αe : [0, 1] → e, meaning that α is continuous on [0, 1] and the restriction to (0, 1) is an orientation preserving homeomorphism onto the interior of e. 2.2. Paths. A combinatorial edge path γ of simplicial length n ≥ 1 is a concatenation e1 e2 . . . en of oriented edges of Γ where t(ei ) = o(ei+1 ). The simplicial length of an edge path γ in Γ is denoted by |γ|Γ , and if it is clear from the context, we suppress Γ and write |γ|. A topological edge path of simplicial length n ≥ 1 is a continuous map p : [a, b] → Γ such that there exists a combinatorial edge path γp = e1 e2 . . . en and a subdivision a = c0 < c1 < c2 < . . . < cn = b such that for each 1 ≤ i ≤ n we have p(ci−1 ) = o(ei ), and p(ci ) = t(ei ) and the restriction of p to (ci−1 , ci ) is an orientation preserving homeomorphism onto the interior of ei . We say that γp is the combinatorial edge path associated to p. Let Γ be equipped with the simplicial metric as above. A topological edge path p : [a, b] → Γ is a PL edge path if for every edge ei in the combinatorial edge path γp the composition αe−1 ◦ p : (ci−1 , ci ) → (0, 1) is the unique orientation preserving affine homeomorphism from i (ci−1 , ci ) to (0, 1). By not requiring that p(a) and p(b) are vertices in Γ, the concept of PL edge paths easily extend to PL paths. A combinatorial edge path of simplicial length 0 consists of a vertex in Γ. A (degenerate) edge path γ that consists of a vertex is called trivial. A combinatorial edge path γ is called reduced if ei 6= ei+1 for all i = 1, . . . , n − 1. A reduced edge path is called cyclically reduced if t(en ) = o(e1 ) and en 6= e1 . A topological or PL edge path is called reduced (resp. trivial) if the corresponding combinatorial edge path is reduced (resp. trivial). More generally, a PL path is called reduced if p : [a, b] → Γ is locally injective. DYNAMICS OF HYPERBOLIC AUTOMORPHISMS OF FN 5 2.3. Graph maps. A topological graph map f : Γ → Γ is a continuous map that sends vertices to vertices, and edges to topological edge paths. The latter means that for any edge e and a characteristic map αe : [0, 1] → Γ, the composition f ◦ αe : [0, 1] → Γ is a topological edge path as above. A linear graph map is a topological graph map f : Γ → Γ as above which, in addition, satisfies the following condition: For every e ∈ Γ, the restriction of f to e is a PL edge path from f (o(e)) to f (t(e)). Note that given a topological graph map f : Γ → Γ there is a linear graph map from Γ to itself that is homotopic to f relative to vertices. Definition 2.1. We say that a topological graph map f : Γ → Γ is regular if for all e ∈ EΓ the (combinatorial) edge path (associated to) f (e) is non-trivial and reduced. The map f is called expanding if for every edge e ∈ EΓ there is an exponent t ≥ 1 such that the edge path f t (e) has simplicial length |f t (e)| ≥ 2. A turn in Γ is a pair (e1 , e2 ) where o(e1 ) = o(e2 ). A turn is called non-degenerate if e1 6= e2 , otherwise it is called degenerate. A regular graph map f : Γ → Γ induces a derivative map Df : EΓ → EΓ where Df (e) is the first edge of the combinatorial edge path associated to f (e). The derivative map induces a map T f on the set of turns defined as T f ((e1 , e2 )) := (Df (e1 ), Df (e2 )). A turn (e1 , e2 ) is called legal if T f n ((e1 , e2 )) is nondegenerate for all n ≥ 0. An edge path γ = e1 e2 . . . en is called legal if every turn (ei , ei+1 ) in γ is legal. Remark-Convention 2.2. (1) For the rest of this paper, a path means a PL path. Also note that, given a combinatorial edge path γ in Γ there is always a PL edge path p : [a, b] → Γ such that γp = γ. Hence, by an edge path we will either mean a PL edge path or a combinatorial edge path and won’t make a distinction unless it is not clear from the context. (2) Similarly, from now on, unless otherwise stated, a graph map means a linear graph map as defined above. Definition 2.3. A graph map f : Γ → Γ is called a train track map if for every edge e and any t ≥ 1 the edge path f t (e) is legal. 2.4. Markings and topological representatives. Let FN denote a free group of finite rank N ≥ 2 which we fix once and for all. A graph Γ is said to be marked, if it is equipped ∼ = with a marking isomorphism θ : FN −→ π1 Γ. Markings are well defined only up to inner automorphisms, which is why we suppress the choice of a base point in Γ. A homotopy equivalence f : Γ → Γ is a topological representative of an automorphism ∼ = ϕ ∈ Out(FN ) with respect to the marking θ : FN −→ π1 Γ if one has ϕ = θ−1 f∗ θ. A graph map f : Γ → Γ is called a train track representative for ϕ if f is a topological representative for ϕ and f is a train track map. Remark 2.4. If a topological graph map f : Γ → Γ represents a hyperbolic outer automorphism ϕ of FN , then the hypothesis that f be expanding is easy to satisfy: It suffices to contract all edges which are not expanded by any iterate f t to an edge path of length ≥ 2: The contracted subgraph must be a forest, as otherwise some f t would fix a non-contractible loop and hence ϕt would fix a non-trivial conjugacy class of π1 (Γ) ∼ = FN , contradicting the 6 MARTIN LUSTIG AND CAGLAR UYANIK assumption that ϕ is hyperbolic. It is easy to see that, if f is a train track map, then this property is inherited by the above “quotient map”. Given an (not necessarily reduced) edge path γ ∈ Γ, let [γ] denote the reduced edge path which is homotopic to γ relative to endpoints. The following is a classical fact for free group automorphisms: Lemma 2.5 (Bounded Cancellation Lemma). [Coo87] Let f : Γ → Γ be topological graph map that is a homotopy equivalence. There exist a constant Cf , depending only on f , such that for any reduced edge path ρ = ρ1 ρ2 in Γ one has |[f (ρ)]| ≥ |[f (ρ1 )]| + |[f (ρ2 )]| − 2Cf . That is, at most Cf terminal edges of [f (ρ1 )] are cancelled against Cf initial edges of [f (ρ2 )] when we concatenate them to obtain [f (ρ)]. Remark 2.6. In Section 4 of this paper we will consider train track maps f : Γ → Γ that are not assumed to be homotopy equivalences. We do, however, assume that f possesses a cancellation bound: there is an integer C ≥ 0 such that for any two legal edge paths γ and γ 0 with common initial vertex but distinct initial edges the backtracking path γ 00 at the image of the concatenation point of the non-reduced path f (γ ◦ γ 0 ) has combinatorial length |γ 00 | ≤ C. The smallest such integer C is denoted by C(f ). The reason for not imposing the homotopy equivalence condition is to make the results and techniques of this paper available for the study of free group endomorphisms, as well as for the use of more general type of train track maps in the context of free group automorphisms. Definition 2.7. A path η in Γ which crosses over precisely one illegal turn is called a periodic indivisible Nielsen path (or INP, for short), if for some exponent t ≥ 1 one has [f t (η)] = η. The smallest such t is called the period of η. A path γ is called pre-INP if its image under f t0 is an INP for some t0 ≥ 1. The illegal turn on η = γ 0 ◦ γ is called the tip of η, while the two maximal initial legal subpaths γ 0 and γ, of η and η respectively, are called the branches of η. A multi-INP or a Nielsen path is a legal concatenation of finitely many INP’s. If f is an expanding train track map, then any non-trivial reduced path η, such that f t (η) is homotopic to η relative to endpoints for some t ≥ 1, can be written as a legal concatenation of finitely many INP’s. This can be seen by a direct elementary argument; alternatively it follows from Proposition 4.12. The reader should be aware of the fact that a priori the endpoints of an INP may not be vertices of Γ (compare with Convention 4.7 below). 2.5. Geodesic Currents on Free Groups. Let FN be the free group of finite rank N ≥ 2, and denote by ∂FN the Gromov boundary of FN . The double boundary of FN is defined by ∂ 2 FN := (∂FN × ∂FN ) r ∆, where ∆ denotes the diagonal. The flip map on ∂ 2 FN is defined by (ξ, ζ) 7→ (ζ, ξ). A geodesic current µ on FN is a locally finite, Borel measure on ∂ 2 FN which is flip and FN -invariant. The set of all geodesic currents on FN is denoted by Curr(FN ). The space of DYNAMICS OF HYPERBOLIC AUTOMORPHISMS OF FN 7 currents is naturally equipped with the weak*-topology. In particular, a sequence of currents µn ∈ Curr(FN ) satisfies lim µn = µ iff n→∞ lim µn (S1 × S2 ) = µ(S1 × S2 ) n→∞ for any two disjoint Borel subsets S1 , S2 ⊂ ∂FN . An immediate example of a geodesic current is given by any non-trivial conjugacy class in FN : For an element g ∈ FN which is not a proper power (i.e. g 6= hk for any h ∈ FN and any k ≥ 2) define the counting current ηg as follows: for any Borel set S ⊂ ∂ 2 FN the value ηg (S) is the number of FN -translates of (g −∞ , g ∞ ) or of (g ∞ , g −∞ ) that are contained in S. For an arbitrary element h ∈ FN r {1} write h = g k , where g is not a proper power, and define ηh := kηg . It is easy to see that this definition is independent of the particular representative of the conjugacy class of h, so that ηh depends only on [h] ⊂ FN . The set of all non-negative real multiples of a counting current is called rational currents, and it is an important fact that rational currents are dense in Curr(FN ), see [Kap05], [Kap06]. Let Φ ∈ Aut(FN ) be an automorphism of FN . Since Φ induces homeomorphisms on ∂FN and on ∂ 2 FN , there is a natural action of Aut(FN ) on the space of currents: For any Φ ∈ Aut(FN ), µ ∈ Curr(FN ) and any Borel subset S ⊂ ∂ 2 FN one has Φµ(S) := µ(Φ−1 (S)). Note that any inner automorphism of FN acts trivially, so that the above action factors through the quotient of Aut(FN ) by all inner automorphisms to give a well defined action of Out(FN ) on Curr(FN ). The space of projectivized geodesic currents is defined as the quotient  PCurr(FN ) =: Curr(FN ) ∼ where µ1 ∼ µ2 in Curr(FN ) if and only µ1 = λµ2 for some λ > 0. We denote the projective class of the current µ by [µ]. We derive from the above stated facts: Proposition 2.8. (1) The space PCurr(FN ) is a compact metric space, and the above actions of Aut(FN ) and Out(FN ) on Curr(FN ) descend to well defined actions on PCurr(FN ) by setting ϕ[µ] := [ϕµ]. (2) The subset of PCurr(FN ) defined by all rational currents is dense in PCurr(FN ). For any marked graph Γ the marking isomorphism θ : FN → π1 Γ gives rise to canonical FN -equivariant homeomorphisms between the Gromov boundaries and double boundaries of e which for simplicity we also denote by θ. the group FN and of the universal cover Γ, e We define the cylinder set associated to γ Let γ e be a reduced edge path in Γ. e as follows: Cyl(e γ ) := {(ξ, ζ) ∈ ∂FN | γ e is a subpath of [∂θ(ξ), ∂θ(ζ)]} where [∂θ(ξ), ∂θ(ζ)] denotes the biinfinite reduced edge path that joins the end ∂θ(ξ) to the e end ∂θ(ζ) of Γ. 8 MARTIN LUSTIG AND CAGLAR UYANIK e We Let γ be a reduced edge path in Γ and let µ ∈ Curr(FN ). Let γ e be a lift of γ to Γ. define hγ, µi := µ(Cyl(e γ )) and note that the FN -equivariance for geodesic currents implies that the value µ(Cylα (e γ )) does not depend on the particular lift γ e we chose, so we have a well defined quantity hγ, µi. Note also that in case of a rational current µ = ηg the quantity hγ, µi is given by the number of occurrences of γ or of γ in the reduced loop [γ] which represents the conjugacy class [g]. Moreover, the action of Out(FN ) on rational currents is given explicitly by the formula (2.1) ϕηg = ηϕ(g) . An important fact about cylinder sets is that they form a basis for the topology on ∂ 2 FN , so that a geodesic current µ is uniquely determined by the set of values hγ, µi for all non-trivial reduced edge paths γ in Γ, see [Kap06]. In particular, we have lim µn = µ if and only if lim hγ, µn i = hγ, µi for all reduced edge n→∞ n→∞ paths γ in Γ. Moreover, if we denote by P(Γ) the set of reduced edge paths in Γ, then the function µΓ : P(Γ) → R≥0 , γ 7→ hγ, µi satisfies for all γ ∈ P(Γ): P (1) µΓ (γ) = µΓ (γ 0 ) = γ 0 ∈E+ (γ) P µΓ (γ 00 ), and γ 0 ∈E− (γ) (2) µΓ (γ) = µΓ (γ), where E+ (γ) ⊂ P(Γ) and E− (γ) ⊂ P(Γ) denote the set of reduced edge paths obtained from γ by adding a further edge at the beginning or at the end respectively. Any function µ0Γ : P(Γ) → R≥0 which satisfies (1) and (2) above is called a Kolmogorov function, and it is well known (see [Kap05, Kap06]) that any such µ0Γ defines uniquely a geodesic current µ on FN which satisfies µΓ = µ0Γ . Let Γ be a marked graph as above. Given a geodesic current µ ∈ Curr(FN ), define the weight of µ with respect to Γ as X (2.2) kµkΓ := he, µi. e∈E + Γ The following criterion given in [Kap06] plays a key role in our convergence estimates in Section 6. Lemma 2.9. Let ([µn ])n∈N be a sequence of currents in PCurr(FN ). Then one has lim [µn ] = n→∞ [µ] in PCurr(FN ) if and only if hγ, µn i hγ, µi = n→∞ kµn kΓ kµkΓ lim holds for all reduced edge paths γ in Γ. DYNAMICS OF HYPERBOLIC AUTOMORPHISMS OF FN 9 3. North-South dynamics In this section we describe some general considerations for maps with North-South dynamics. We will keep the notation simple and general; at no point we will refer to the specifics of currents on free groups. However, in this section we will prove the main criteria used in the remainder of the paper to establish the North-South dynamics result in our main theorem. Definition 3.1. Let f : X → X be homeomorphism of a topological space X. (a) The map f is said to have (pointwise) North-South dynamics if f has two distinct fixed points P+ and P− , called attractor and repeller, such that for every x ∈ X r {P+ , P− } one has: lim f t (x) = P+ and lim f t (x) = P− . t→∞ t→−∞ (b) The map f is said to have uniform North-South dynamics if the following holds: There exist two distinct fixed points P− and P+ of f , such that for every compact set K ⊂ X r{P− } and every neighborhood U+ of P+ there exists an integer t+ ≥ 0 such that for every t ≥ t+ one has: f t (K) ⊂ U+ . Similarly, for every compact set K ⊂ X r {P+ } and every neighborhood U− of P− there exists an integer t− ≤ 0 such that for every t ≤ t− one has: f t (K) ⊂ U− . It is easy to see that uniform North-South dynamics implies pointwise North-South dynamics. Definition 3.2. A homeomorphism f : X → X of a topological space X has generalized uniform North-South dynamics if there exist two disjoint compact f -invariant sets ∆+ and ∆− in X, such that the following hold: (i) For every compact set K ⊂ X r ∆− and every neighborhood U+ of ∆+ there exists an integer t+ ≥ 0 such that for every t ≥ t+ one has: f t (K) ⊂ U+ (ii) For every compact set K ⊂ X r ∆+ and every neighborhood U− of ∆− there exists an integer t− ≤ 0 such that for every t ≤ t− one has: f t (K) ⊂ U− More precisely, we say that the map f has generalized uniform North-South dynamics from ∆− to ∆+ . Note that we interpret the phrase “f -invariant” in its strong meaning, i.e. f (∆+ ) = ∆+ and f (∆− ) = ∆− . Proposition 3.3. Let f : X → X be a homeomorphism of a compact metrizable space X. Let Y ⊂ X be dense subset of X, and let ∆+ and ∆− be two f -invariant sets in X that are disjoint. Assume that the following criterion holds: For every neighborhood U of ∆+ and every neighborhood V of ∆− there exists an integer m0 ≥ 1 such that for any m ≥ m0 and any y ∈ Y one has either f m (y) ∈ U or f −m (y) ∈ V . Then f 2 has generalized uniform North-South dynamics from ∆− to ∆+ . 10 MARTIN LUSTIG AND CAGLAR UYANIK Proof. Let K ⊂ X r ∆− be compact, and let U and V be neighborhoods of ∆+ and ∆− respectively. See Figure 1. X V U U0 ∆− ∆+ W K Figure 1. Let W be an open neighborhood of K such that W ∩ V = ∅. Since X is compact, the closure W is compact. Then V1 := V r W is again an open neighborhood of ∆− , moreover it is disjoint from W . Let U1 be a neighborhood of ∆+ which has the property that its closure is contained in the interior of U . Such a neighborhood exists because X is a metrizable space. Let m0 be as postulated in the criterion, applied to the neighborhoods U1 and V1 , and pick any m ≥ m0 . Consider any y ∈ Y ∩ f m (W ). Notice that f −m (y) is contained in W , which is disjoint from V1 . Thus, by the assumed criterion, f m (y) must be contained in U1 . Since W is open and f is a homeomorphism, any dense subset of X must intersect f m (W ) in a subset that is dense in f m (W ). This implies that f m (f m (W )) ⊆ U1 ⊂ U . Since K ⊂ W , this shows that f 2m (K) ⊂ U . Using the analogous argument for the inverse iteration we see that f 2 has generalized uniform North-South dynamics from ∆− to ∆+ . u t Proposition 3.4. Let f : X → X be a homeomorphism of a compact space X, and let ∆+ and ∆− be disjoint f -invariant sets. Assume that some power f s with s ≥ 1 has generalized uniform North-South dynamics from ∆− to ∆+ . Then the map f , too, has generalized uniform North-South dynamics from ∆− to ∆+ . Proof. Let K ⊂ X r ∆− be compact, and let U be an open neighborhood of ∆+ . Set K 0 := K ∪ f (K) ∪ . . . ∪ f s−1 (K), which is again compact. Note that the fact that K ⊂ X r ∆− and f −1 (∆− ) = ∆− implies that K 0 ⊂ X r ∆− . Indeed, x ∈ K 0 implies that x = f t (y) for some y ∈ K and for some 0 ≤ t ≤ s − 1. Thus x ∈ ∆− would imply that y = f −t (x) ∈ f −1 (∆− ) = ∆− , contradicting the assumption K ∩ ∆− = ∅. DYNAMICS OF HYPERBOLIC AUTOMORPHISMS OF FN 11 From the hypothesis that f s has generalized uniform North-South dynamics from ∆− to 0 ∆+ it follows that there is a bound t0 such that for all t0 ≥ t0 one has f t s (K 0 ) ⊂ U . Hence, for any point x ∈ K and any integer t ≥ st0 , we can write t = k + st0 with t0 ≥ t0 and 0 ≤ k ≤ s − 1 to obtain the desired fact 0 0 f t (x) = f k+st (x) = f st f k (x) ∈ f st (K 0 ) ⊂ U. The analogous argument for f −1 finishes the proof of the Lemma. u t 4. Train-Tracks In this section we will consider train track maps which satisfy the following properties: Convention 4.1. Let Γ be a finite connected graph, and let f : Γ → Γ be an expanding train track map which possesses a cancellation bound Cf ≥ 0 as defined in Remark 2.6. We also assume that f has been replaced by a positive power so that for some integers λ00 ≥ λ0 > 1 we have, for any edge e of Γ: λ00 ≥ |f (e)| ≥ λ0 (4.1) and λ0 , λ00 is attained for some edges. 4.1. Goodness. The following terminology was introduced by R. Martin in his thesis [Mar95]. Definition 4.2. Let f : Γ → Γ, Cf and λ0 be as in Convention 4.1. Define the critical Cf constant C for f as C := 0 . Let γ be a reduced edge path in Γ. Any edge e in γ that is λ −1 at least C edges away from an illegal turn on γ is called good, where the distance (= number of edges traversed) is measured on γ. An edge is called bad if it is not good. Edge paths or loops, in particular subpaths of a given edge path, which consist entirely of good (or entirely of bad) edges are themselves called good (or bad). For any edge path or a loop γ in Γ we define the goodness of γ as the following quotient: #{good edges of γ} g(γ) := ∈ [0, 1] |γ| We will now discuss some basic properties of the goodness of paths and loops. We first consider any legal edge path γ in Γ of length |γ| = C and compute: (4.2) |f (γ)| ≥ λ0 |γ| = λ0 C = Cf + C Lemma 4.3. Let f : Γ → Γ and λ0 be as in Convention 4.1. Let γ be a reduced edge path or a cyclically reduced loop in Γ. Then any good subpath γ 0 of γ has the property that no edge of f (γ 0 ) = [f (γ 0 )] is cancelled when f (γ) is reduced, and that it consists entirely of edges that are good in [f (γ)]. Hence, for any reduced loop γ in Γ we have: #{good edges in [f (γ)]} ≥ λ0 · #{good edges in γ} Proof. If γ is legal, then every edge is good, and the lemma follows directly from the definition of λ0 in Convention 4.1. Now assume that the path γ has at least one illegal turn. Let γ = γ1 B1 γ2 B2 . . . γn Bn 12 MARTIN LUSTIG AND CAGLAR UYANIK b2 a1 γ1 a2 Figure 2. ai ’s are legal ends of maximal bad segments, γ is a good (legal) segment be a decomposition of γ into maximal good edge paths γi and maximal bad edge paths Bi . Note that each maximal bad segment Bi can be written as an illegal concatenation Bi = ai bi ai+1 where ai is a legal segment of length C and bi is an edge path that (possibly) contains some illegal turns. Note that since |f (ai )| ≥ λ0 |ai | ≥ λ0 C = Cf +C, Lemma 2.5 implies that [f (Bi )] is an edge path of the from [f (Bi )] = a0i b0i c0i where a0i , c0i are legal edge paths such that |a0i |, |c0i | ≥ C. Moreover, the turn at f (γi )a0i is legal. Since by Convention 4.1 every edge grows at least by a factor of λ0 , this implies the required result. u t On the other hand, for any cyclically reduced loop γ the number of bad edges is related to the number of illegal turns on γ, which we denote by ILT (γ), via: (4.3) ILT (γ) ≤ #{bad edges in γ} ≤ 2C · ILT (γ) Since the number of illegal turns on γ can only stay constant or decrease under iteration of the train track map, we obtain directly (4.4) #{bad edges in [f t (γ)]} ≤ 2C · ILT (γ) ≤ 2C · #{bad edges in γ} for all positive iterates f t of f . Notice however that the number of bad edges may actually grow (slightly) faster than the number of good edges under iteration of f , so that the goodness of γ does not necessarily grow monotonically under iteration of f . Nevertheless up to passing to powers one can overcome this issue: Proposition 4.4. Let f : Γ → Γ be as in Convention 4.1. (a) There exists an integer s ≥ 1 such that for every reduced loop γ in Γ one has: g([f s (γ)]) ≥ g(γ) In particular, for any integer t ≥ 0 one has g([(f s )t (γ)]) ≥ g(γ) (b) If 0 < g(γ) < 1, then for any integer s0 > s we have 0 g([f s (γ)]) > g(γ) DYNAMICS OF HYPERBOLIC AUTOMORPHISMS OF FN 13 and thus, for any t ≥ 0: 0 g([(f s )t (γ)]) > g(γ) Proof. (a) We set s ≥ 1 so that λ0 s ≥ 2C and obtain from Lemma 4.3 for the number g 0 of good edges in [f s (γ)] and the number g of good edges in γ that: s g 0 ≥ λ0 g For the number b0 of bad edges in [f s (γ)] and the number b of bad edges in γ we have from equation (4.4) that: b0 ≤ 2Cb Thus we get g0 g g0 λ0s g g ≥ and hence ≥ ≥ 0 0 0 b 2C b b g +b g+b s which proves g([f (γ)]) ≥ g(γ). The second inequality in the statement of the lemma follows directly from an iterative application of the first. (b) The proof of part (b) follows from the above given proof of part (a), since λ0 s > 2C unless there is no good edge at all in γ, which is excluded by our hypothesis 0 < g(γ). Since the hypothesis g(γ) < 1 implies that there is at least one illegal turn in γ, in the above proof we get b ≥ 1, which suffices to show g0 g > , 0 b b s and thus g([f (γ)]) > g(γ). u t From the inequalities at the end of part (a) of the above proof one derives directly, for g > 0, the inequality 1 g([f s (γ)]) ≥ . 2C 1 1 + λ0s ( g(γ) − 1) Hence we obtain: Corollary 4.5. Let f : Γ → Γ be as in Convention 4.1, and let δ > 0 and ε > 0 be given. Then there exist an integer M 0 = M 0 (δ, ε) ≥ 0 such that for any loop γ in Γ with g(γ) ≥ δ we have g([f m (γ)]) ≥ 1 − ε for all m ≥ M 0 . u t 4.2. Illegal turns and iteration of the train track map. The following lemma (and also other statements of this subsection) are already known in differing train track dialects (compare for example Lemma 4.2.5 in [BFH00]); for convenience of the reader we include here a short proof. Recall the definition of an INP and a pre-INP from Definition 2.7. Lemma 4.6. Let f : Γ → Γ be as in Convention 4.1. Then there are only finitely many INP’s and pre-INP’s in Γ for f . Furthermore, there is an efficient method to determine them. 14 MARTIN LUSTIG AND CAGLAR UYANIK Proof. We consider the set V of all pairs (γ1 , γ2 ) of legal edge paths γ1 , γ2 with common initial vertex but distinct first edges, which have combinatorial length |γ1 | = |γ2 | = C. Note that V is finite. From the definition of the cancellation bound and the inequalities (4.1) it follows directly that every INP or every pre-INP η must be a subpath of some path γ 1 ◦ γ2 with (γ1 , γ2 ) ∈ V. Furthermore, we define for i = 1 and i = 2 the initial subpath γi∗ of γi to consist of all points x of γi that are mapped by some positive iterate f t into the backtracking subpath at the tip of the unreduced path f t (γ 1 ◦ γ2 ). We observe that the interior of any INP-subpath or pre-INP-subpath η of γ 1 ◦ γ2 must agree with the subpath γ ∗1 ◦ γ2∗ . Thus any pair (γ1 , γ2 ) ∈ V can define at most one INP-subpath of γ 1 ◦ γ2 . Since V is finite and easily computable, we obtain directly the claim of the lemma. u t In light of Lemma 4.6, we obtain a subdivision Γ0 of Γ by adding the endpoints of all the finitely many INP’s or pre-INP’s, while keeping the property that f maps vertices to vertices. This gives a train track map f 0 : Γ0 → Γ0 which represents the same outer automorphism as f : Γ → Γ, thus justifying that from now on we concentrate on train track maps which satisfy the following: Convention 4.7. Let f : Γ → Γ be a train track map as in Convention 4.1. We assume in addition that every endpoint of an INP or a pre-INP is a vertex. Lemma 4.8. Let f : Γ → Γ be as in Convention 4.7. Then, there exists an exponent M1 ≥ 1 with the following property: Let γ be a path in Γ, and assume that it contains precisely two illegal turns, which are the tips of INP-subpaths or pre-INP-subpaths η1 and η2 of γ. If η1 and η2 overlap in a non-trivial subpath γ 0 , then f M1 (γ) reduces to a path [f M1 (γ)] which has at most one illegal turn. Proof. For any point x1 in the interior of one of the two legal branches of an INP or a pre-INP η there exists a point x2 on the other legal branch such that for a suitable positive power of f one has f t (x1 ) = f t (x2 ). Hence, if we pick a point x in the interior of γ 0 , there are points x0 on the other legal branch of η1 and x00 on the other legal branch of η2 such that for some positive iterate of f one has f t (x0 ) = f t (x) = f t (x00 ). It follows that the decomposition γ = γ1 ◦ γ2 ◦ γ3 , which uses x0 and x00 as concatenation points, defines legal subpaths γ1 and γ3 which yield [f t (γ)] = [f t (γ1 ) ◦ f t (γ3 )], which has at most one illegal turn. Since by Convention 4.7 the overlap γ 0 is an edge path, it follows from the finiteness result proved in Lemma 4.6 that there are only finitely many constellations for η1 and η2 . This shows that there must be a bound M1 as claimed. u t Lemma 4.9. For every train track map f : Γ → Γ as in Convention 4.7 there exists a constant M2 = M2 (Γ) ≥ 0 such that every path γ with precisely 1 illegal turn satisfies the following: Either γ contains an INP or pre-INP as a subpath, or else [f M2 (γ)] is legal. Proof. Similar to the set V in the proof of Lemma 4.6 we define the set V+ be the set of all pairs (γ1 , γ2 ) of legal edge paths γ1 , γ2 in Γ which have combinatorial length 0 ≤ |γ1 | ≤ C and 0 ≤ |γ2 | ≤ C, and which satisfy: The paths γ1 and γ2 have common initial point, and, unless one of them (or both) are trivial, they have distinct first edges. Note that V+ is finite and contains V as subset. DYNAMICS OF HYPERBOLIC AUTOMORPHISMS OF FN 15 Following ideas from [CL15] we define a map f# : V+ → V+ induced by f through declaring the f# -image of a pair (γ1 , γ2 ) ∈ V+ to be the pair f# (γ1 , γ2 ) := (γ100 , γ200 ) ∈ V+ which is defined by setting for i ∈ {1, 2} f (γi ) =: γi0 ◦ γi00 ◦ γi000 , where γi0 is the maximal common initial subpath of f (γ1 ) and f (γ2 ), where |γi00 | ≤ C, and where γi000 is non-trivial only if |γi00 | = C. Then for any (γ1 , γ2 ) ∈ V+ we see as in the proof of Lemma 4.6 that there exists an exponent t ≥ 0 such that the iterate f#t (γ1 , γ2 ) =: (γ3 , γ4 ) ∈ V+ satisfies one of the following: (1) One of γ3 or γ4 (or both) are trivial. (2) The turn defined by the two initial edges of γ3 and γ4 is legal. (3) The path γ 3 ◦ γ4 contains an INP as subpath. From the finiteness of V+ it follows directly that there is an upper bound M2 ≥ 0 such that for t ≥ M2 one of the above three alternatives must be true for f#t (γ1 , γ2 ) =: (γ3 , γ4 ). Consider now the given path γ, and write it as illegal concatenation of two legal paths γ = γ10 ◦ γ20 . Then the maximal initial subpaths γ1 of γ 01 and γ2 of γ20 of γ2 of length |γi | ≤ C form a pair (γ1 , γ2 ) in V+ . In the above cases (1) or (2) it follows directly that f t (γ) is a (possibly trivial) legal path. In alternative (3) the path f t (γ) contains an INP. u t Proposition 4.10. For any train track map as in Convention 4.7 there exists an exponent r = r(f ) ≥ 0 such that every finite path γ in Γ with ILT(γ) ≥ 1 satisfies ILT([f r (γ)]) < ILT(γ) , unless every illegal turn on γ is the tip of an INP or pre-INP subpath ηi of γ, where any two ηi are either disjoint subpaths on γ, or they overlap precisely in a common endpoint. Proof. Through considering maximal subpaths with precisely one illegal turn we obtain the claim as direct consequence of Lemma 4.9 and Lemma 4.8. u t Definition 4.11. A path γ in Γ is called pseudo-legal if it is a legal concatenation of legal paths and INP’s. From the same arguments as in the last proof we also deduce the following: Proposition 4.12. Let f : Γ → Γ be as in Convention 4.7. Then for any finite edge path γ in Γ there is a positive iterate f t (γ) which reduces to a path γ 0 := [f t (γ)] that is pseudo-legal. The analogous statement also holds for loops instead of paths. The exponent t needed in either case depends only on the number of illegal turns in γ. Proof. Since the number of illegal turns in [f t (γ)] non-strictly decreases for increasing t, we can assume that for sufficiently large t it stays constant. It follows from Lemma 4.9 (after possibly passing to a further power of f ) that every illegal turn of [f t (γ)] is the tip of some INP-subpath ηi of [f t (γ)]. From Lemma 4.8 we obtain furthermore that any two such ηi and ηj that are adjacent on [f t (γ)] can not overlap non-trivially. u t 16 MARTIN LUSTIG AND CAGLAR UYANIK In the next subsection we also need the following lemma, where illegal (cyclic) concatenation means that the path (loop) is a concatenation of subpaths where all concatenation points are illegal turns. Lemma 4.13. Let f : Γ → Γ be an expanding train track map. Let γ be a reduced loop in Γ and let γ 0 = [f (γ)] be its reduced image loop. Assume that for some t ≥ 1 a decomposition γ 0 = γ10 ◦γ20 ◦. . .◦γt0 as illegal cyclic concatenation is given. Then there exists a decomposition as illegal cyclic concatenation γ = γ1 ◦ γ2 ◦ . . . ◦ γt with the property that the reduced image paths [f (γi )] contain the paths γi0 as subpaths, for i = 1, 2, . . . , t. Proof. We cut the loop γ at some illegal turn to get a (closed) path γ = e1 e2 . . . eq . We consider the initial subpaths γ(k) = e1 e2 . . . ek of γ for k = 1, 2, . . .. Let k 0 be the smallest positive integer such that the reduced image path [f (γ(k 0 ))] contains the path γ10 as a subpath. Since k 0 is the smallest such integer, it follows that the path f (ek0 ) passes through the last edge of γ10 . Note that f (ek0 ) is a legal path and that γi0 terminates at an illegal turn, so that the endpoint of f (ek0 ) must lie somewhere in the backtracking subpath of the possibly unreduced path [f (e1 e2 . . . ek0 ) ] ◦ [f (ek0 +1 ek0 +2 . . . eq )]. It follows that the reduced image path [f (ek0 +1 ek0 +2 . . . eq )] contains the path γ20 ◦ . . . ◦ γt0 as subpath. Thus we define γ1 := γ(k 0 ), and proceed iteratively in precisely the same fashion, thus finding iteratively γ2 , γ3 , . . . γt−1 . As above, it follows that the “left-over” terminal subpath of γ has the property that its reduced image contains γt0 as subpath, so that we can define this left-over subpath to be the final factor path γt . u t 4.3. Goodness versus illegal turns. We recall from Definition 2.7 that a multi-INP is a legal concatenation of finitely many INPs along their endpoints. For the remainder of this section we will concentrate on graph maps f : Γ → Γ which satisfy the following: Convention 4.14. Let f : Γ → Γ be an expanding train track map as in Convention 4.7. We assume furthermore that there is an upper bound A(f ) to the number of INP factor paths in any multi-INP path γ ∈ Γ. Note that this condition implies in particular that in Γ there can not be a non-trivial loop which is a cyclic legal concatenation of INPs. The next lemma shows in particular that an expanding train track map which represents a hyperbolic outer automorphism naturally satisfies the conditions given in Convention 4.14. Lemma 4.15. For any expanding train track map f : Γ → Γ, which has the property that no positive power of f fixes the homotopy class of any non-trivial loop in Γ, there is an upper bound to the number of INP factors on any multi-INP path. Proof. There is an upper bound A ≥ 0 for number of INPs in Γ, see Lemma 4.6. Hence, any multi-INP with more than A(f ) := 2A factors would have to run over the same INP twice in the same direction. Thus we obtain as subpath a non-trivial loop which is fixed by some positive power of f , violating the given assumption. u t Lemma 4.16. For any train track map f : Γ → Γ as in Convention 4.14 there exists an exponent s ≥ 1 such that for any reduced path γ ∈ Γ with reduced image path γ 0 = [f s (γ)] the following holds: If g(γ 0 ) = 0 and satisfies ILT (γ 0 ) ≥ A(f ) + 1, then ILT (γ) > ILT (γ 0 ). DYNAMICS OF HYPERBOLIC AUTOMORPHISMS OF FN 17 Proof. Since the map f is expanding, for the critical constant C ≥ 0 given in Definition 4.2 there exists an integer s ≥ 1 such that |f s (e)| ≥ 2C + 1 for every edge e ∈ Γ. We can furthermore assume s ≥ r, where r = r(f ) is given by Lemma 4.10. Thus, by Lemma 4.10 we have ILT (γ) > ILT (γ 0 ) , unless every illegal turn on γ is the tip of an INP or pre-INP ηi , and any two such subpaths ηi are disjoint or they overlap precisely at a common endpoint. The case where two such paths are disjoint can be excluded as follows: If there is an edge e in γ which doesn’t belong to any of the ηi , then f s (e) is a legal subpath of [f s (γ)] = γ 0 of length greater than 2C + 1, contradicting the assumption that g(γ 0 ) = 0. Thus we can assume that any two of the subpaths ηi overlap precisely in a common endpoint, and that there is no non-trivial initial of final subpath of γ outside of the con0 catenation of all the ηi . Therefore, for some s0 ≥ 0, the iterate f s (γ 0 ) is a multi-INP with ILT (γ 0 ) ≥ A(f ) + 1 factors, which contradicts Convention 4.14. Hence the conclusion of the Lemma follows. u t Lemma 4.17. Let f : Γ → Γ and A(f ) be as in Convention 4.14. Then there exists a constant δ with 0 < δ ≤ 1 so that the following holds: Every reduced loop γ in Γ with ILT (γ) ≥ A(f ) + 1 can be written as cyclic illegal concatenation γ = γ1 ◦ γ2 . . . ◦ γ2m such that for every odd index j, the subpath γj is either trivial or satisfies g(γj ) ≥ δ . For every even index k the subpath γk is non-trivial and satisfies g(γk ) = 0; moreover, we have: A(f ) + 1 ≤ ILT (γk ) ≤ 2A(f ) + 1 Proof. Let Lγ be the collection of maximal legal subpaths γi0 of γ of length |γi0 | ≥ 2C + 1, for C ≥ 0 as given in Definition 4.2. Note that any two distinct elements γi0 , γj0 in this collection are either disjoint or overlap at at single point. In the latter case the turn at the overlap point is illegal, as otherwise they would merge into a longer legal subpath, which violates 0 the maximality of the γi0 . For any two (on γ) consecutive paths γi0 and γi+1 in Lγ , if the path 0 0 βi between them is trivial or satisfies ILT (βi ) ≤ A(f ), then we erase γi and γi+1 from the 0 0 collection Lγ and add in the new path γi βi γi+1 . We continue this process iteratively until all the complementary subpaths βi between any two consecutive elements in our collection satisfies ILT (βi ) ≥ A(f ) + 1. We call the obtained collection of subpaths Cγ . We now pick a path γj in the collection Cγ and consider its “history” as being obtained iteratively through joining what amounts to ` paths γi0 from Lγ . Thus γj can be written as illegal concatenation 0 γj = γ10 ◦ β1 ◦ γ20 ◦ . . . ◦ γ`−1 ◦ β`−1 ◦ γ`0 , where each γi0 is legal and of length |γi0 | ≥ 2C + 1. Each βi has at most A(f ) illegal turns, and the legal subpaths of βi between these illegal turns have length ≤ 2C, so that we get |βi | ≤ (A(f ) + 1)2C. Thus the set of good edges on γj is given precisely as disjoint union 18 MARTIN LUSTIG AND CAGLAR UYANIK over all the γi0 of the sets of the |γi0 | − 2C edges of γi0 that are not on the two boundary subpaths of length C of γi0 . Hence we compute for the goodness: P` P` 0 (|γ 0 | − 2C) i=1 (|γi | − 2C) g(γj ) = = P` i=1 0 i P`−1 |γj | i=1 |γi | + i=1 |βi | ` 1 = (2C + 1) · ` + (A(f ) + 1)2C · ` 2C(A(f ) + 2) + 1 We finally add to the collection Cγ a suitable set of trivial subpaths at illegal turns to get a collection Cγ0 , where these trivial paths are chosen as to get as complementary subpaths of Cγ0 only subpaths γk which satisfy: ≥ A(f ) + 1 ≤ ILT (γk ) ≤ 2A(f ) + 1 Thus setting δ= 1 2C(A(f ) + 2) + 1 u t finishes the proof. Proposition 4.18. Let f : Γ → Γ be a train track map as in Convention 4.14, and let s ≥ 1 be the integer given by Lemma 4.16 and δ > 0 be the constant given by Lemma 4.17. Then there exists a constant R > 1 such that for any reduced loop γ in Γ one has either (1) g([f s (γ)]) ≥ δ 2 or (2) ILT (γ) ≥ R · ILT ([f s (γ)]). Proof. By 4.12 there is an exponent t ≥ 0 such that for any loop γ with less than A(f ) + 1 illegal turns the loop γ 0 = [f t (γ)] is pseudo-legal. From the assumption that f satisfies Convention 4.14 it follows that γ 0 is not a legal concatenation of INPs, so that it must have at least one good edge. Since iteration of f only decreases the number of illegal turns, we obtain from equality (4.3) that g(γ) ≥ 1 . 2C(A(f ) + 1) + 1 Thus the first inequality from our assertion follows for a suitable choice of s from Proposition 4.4. 0 Thus we can now assume that ILT (γ) ≥ A(f ) + 1, and thus that [f s (γ)] = γ10 ◦ . . . ◦ γ2m is a decomposition as given by Lemma 4.17. There are two cases to consider: Assume that X X |γj0 | ≥ |γk0 |. j odd k even For any odd index j the non-trivial path γj0 has g(γj0 ) ≥ δ, which together with the last inequality implies: δ g([f s (γ)]) ≥ 2 DYNAMICS OF HYPERBOLIC AUTOMORPHISMS OF FN 19 Now, assume that: X X |γj0 | ≤ j odd |γk0 | k even Let γ = γ1 ◦ . . . ◦ γ2m be an illegal cyclic concatenation given by Lemma 4.13 so that, for each i = 1, . . . , 2m, the reduced image [f s (γi )] contains γi0 as a subpath. Hence, for every even index, we apply Lemma 4.16 to see that ILT (γk ) > ILT ([f s (γk )]) ≥ ILT (γk0 ). Since the number if illegal turns never increases when applying a train track map, for each odd index j, we have ILT (γj ) ≥ ILT ([f s (γj )]) ≥ ILT (γj0 ). Combining last two inequalities we obtain: ILT (γ) > ILT ([f s (γ)]) + m. (4.5) We also observe that the number of illegal turns in [f s (γ)] is equal to the sum of the number of illegal turns in the odd indexed subpaths, the number of illegal turns in the even indexed subpaths, and the number of illegal turns at concatenation points: X X ILT (γk0 ) + 2m ILT (γj0 ) + (4.6) ILT ([f s (γ)]) ≤ k even j odd Now, note that (4.7) X ILT (γj0 ) ≤ X j odd j odd |γj0 | ≤ X |γk0 |, k even by assumption. For each γk0 (with even index k), since Lemma 4.17 assures g(γk0 ) = 0 and ILT (γk0 ) ≤ 2(A(f ) + 1), we have |γk0 | ≤ 2C(2(A(f ) + 1) + 1), which together with (4.7) implies that: X (4.8) ILT (γj0 ) ≤ 2mC(2(A(f ) + 1) + 1) j odd Furthermore, from ILT (γk0 ) ≤ 2(A(f ) + 1) for even index k we deduce X (4.9) ILT (γk0 ) ≤ 2m(A(f ) + 1) k even Using (4.6), (4.8) and (4.9) we obtain: (4.10) ILT ([f s (γ)]) ≤ 2mC(2(A(f ) + 1) + 1) + 2m(A(f ) + 1) + 2m 20 MARTIN LUSTIG AND CAGLAR UYANIK Using (4.5) and (4.10), we have: ILT (γ) m ≥ 1 + ILT ([f s (γ)]) ILT ([f s (γ)]) m 2mC(2(A(f ) + 1) + 1) + 2m(A(f ) + 1) + 2m 1 ≥1+ 2C(2(A(f ) + 1) + 1) + 2(A(f ) + 1) + 2 ≥1+ Therefore, the conclusion of the lemma holds for: R=1+ 1 2C(2(A(f ) + 1) + 1) + 2(A(f ) + 1) + 2 u t 4.4. Uniform goodness in the future or the past. For the convenience of the reader we first prove a mild generalization of Proposition 4.18, which will be a crucial ingredient in the proof of the next proposition. Proposition 4.19. Let f : Γ → Γ be as in Convention 4.14. Given any constants 0 < δ1 < 1 and R1 > 1, there exist an integer s1 > 0 so that for any reduced loops γ and γ 0 in Γ with [f s1 (γ 0 )] = γ one has either (i) g([f s1 (γ)]) ≥ δ1 or (ii) ILT (γ 0 ) ≥ R1 · |γ|Γ . Proof. We first replace f by a positive power (say f r , cited at the end of the proof) as given by Proposition 4.4 so that for the rest of the proof we can assume that goodness is monotone. Let s, δ and R be as in Lemma 4.18 and let γ be any reduced loop in Γ. Assume that for γ that alternative (1) of Lemma 4.18 holds, i.e. g([f s (γ)]) ≥ 2δ . Then by Corollary 4.5 there is an exponent M ≥ 1 such that g([(f s )m (γ)]) ≥ δ1 for all m ≥ M . On the other hand, if g([f s (γ)]) < 2δ , then Lemma 4.18 assures that ILT (γ) ≥ R · ILT ([f s (γ)]). We now claim that ILT (γ1 ) ≥ R · ILT (γ) s for any reduced loop γ1 with [f (γ1 )] = γ. To see this, apply Lemma 4.18 to the loop γ1 . If one had δ g([f s (γ1 )]) = g(γ) ≥ , 2 then by monotonicity of goodness this would also imply g([f s (γ)]) ≥ 2δ , which contradicts with our assumption. Hence for γ1 alternative (2) of Lemma 4.18 holds, giving indeed ILT (γ1 ) ≥ R · ILT (γ). DYNAMICS OF HYPERBOLIC AUTOMORPHISMS OF FN 21 Repeating the same argument iteratively shows that for any sequence of reduced loops γM 0 , defined iteratively through [f s (γM 0 )] = γM 0 −1 for any M 0 ≥ 2, the inequality 0 γM 0 ≥ RM · ILT (γ). holds. Also, notice that since g(γ) < δ 2 we have #{bad edges in γ} δ ≥1− , |γ|Γ 2 and by the inequalities (4.3) we have furthermore #{bad edges in γ} ≤ 2C · ILT (γ). Hence we obtain 0 ILT (γM 0 ) ≥ RM · ILT (γ) 0 #{bad edges in γ} 2C δ 1 · (1 − ) |γ|Γ 2 2C ≥ RM · ≥ RM 0 0 1 ≥ R1 . Then s1 = max{M, M 0 } satisfies the Let M 0 ≥ 1 be such that RM · (1 − 2δ ) 2C requirements of the statement of Proposition 4.19 up to replacing f by the power f r as done at the beginning of the proof. u t In addition to the train track map f we now consider a second train track map f 0 : Γ0 → Γ0 , which also satisfies the requirements of Convention 4.14 and which is related to f via maps h : Γ → Γ0 and h0 : Γ0 → Γ such that f and h0 f 0 h are homotopy inverses. We also assume that the lifts of h and h0 to the universal coverings are quasi-isometries (with respect to the simplicial metrics), so that there exists a bi-Lipschitz constant B > 0 which satisfies for any two “corresponding” reduced loops γ in Γ and γ 0 := [h(γ)] the inequalities (4.11) 1 0 |γ |Γ0 ≤ |γ|Γ ≤ B |γ 0 |Γ0 B We denote the goodness for the map f 0 by g0 , and the critical constant for f 0 from Definition 4.2 by C 0 . Proposition 4.20. Let f : Γ → Γ and f 0 : Γ0 → Γ0 be as specified in the last paragraph. Given a real number δ so that 0 < δ < 1, there exist a bound M > 0 such that (up to replacing f and f 0 by a common power) for any pair of corresponding reduced loops γ in Γ and γ 0 in Γ0 , either (a) g([f n (γ)]) ≥ δ or (b) g0 ([(f 0 )n (γ 0 )]) ≥ δ holds for all n ≥ M . 22 MARTIN LUSTIG AND CAGLAR UYANIK Proof. Let B be a bi-Lipschitz constant for the transition from Γ to Γ0 as in (4.11) above. Set R1 = 4C 0 B 2 , and apply Proposition 4.19 to the loop γ. Assume first that alternative (i) of this proposition holds. Then Corollary 4.5, applied to f s1 , gives a bound L ≥ 0 so that inequality (a) holds for all n ≥ L (after having replaced f by f s1 ). Now assume that for γ alternative (ii) of Proposition 4.19 holds. Then we have the following inequalities, where γ 00 denotes the reduced loop in Γ corresponding to f 0s1 (γ 0 ) (which implies [f s1 (γ 00 )] = γ): 1 00 1 |γ |Γ ≥ ILT (γ 00 ) B B 1 ≥ R1 · |γ|Γ = 4C 0 B · |γ|Γ B = 4C 0 · |γ 0 |Γ0 |f 0s1 (γ 0 )|Γ0 ≥ This however implies that g0 (f 0s1 (γ 0 )) ≥ 1/2, since for any t ≥ 0 the number of bad edges in f 0t (γ 0 ) is bounded above by 2C 0 · ILT (f 0t (γ 0 )) ≤ 2C 0 · ILT (γ 0 ) ≤ 2C 0 · |γ 0 |Γ0 . Thus, by invoking Corollary 4.5 again, there exist L0 such that 0 g0 (f 0s1 L (γ 0 )) ≥ δ. Hence for M 0 = max{L, L0 } the conclusion of the Lemma follows. u t Remark 4.21. From the next section on we will restrict our attention to expanding train track maps which are homotopy equivalences. For any such train track representative f of some ϕ ∈ Out(FN ) and f 0 of ϕ−1 , which both satisfy Convention 4.14, the conditions stated before Proposition 4.20 are satisfied, so that this proposition is valid for them. 5. Perron Frobenius Theorem for reducible substitutions and automorphisms The purpose of this section is to define the simplex of attraction ∆+ and the simplex of repulsion ∆− explicitly by exhibiting the coordinate values for the defining currents. For this purpose we crucially need the main result from our earlier paper [LU17]; details of how to use this quote are given below. 5.1. Symbolic dynamics, substitutions and quotes from [LU17]. In this subsection we recall some terminology, well known definitions and basic facts from symbolic dynamics. Given any finite set of symbols A = {a1 , . . . , am }, called an alphabet, one considers the free monoid A∗ over A, which is given by all finite words in A (without inverses), including the empty word 1. The length |w| of a word w ∈ A∗ denotes the number of letters in w. Furthermore one considers the full shift AZ , the set of all biinfinite words in A, which is canonically equipped with the product topology and with the shift-operator. The latter maps x = . . . xi−1 xi xi+1 . . . ∈ AZ to y = . . . yi−1 yi yi+1 . . . ∈ AZ , with yi = xi+1 ∈ A for any i ∈ Z. A subset Σ ⊂ AZ is called a subshift if it is non-empty, closed and shift-invariant. The language of Σ is the collection of all finite words occurring in elements of Σ. For any DYNAMICS OF HYPERBOLIC AUTOMORPHISMS OF FN 23 w = x1 . . . xn ∈ A∗ we denote by Cylw the set of all y = . . . yi−1 yi yi+1 . . . ∈ AZ which satisfy y1 = x1 , y2 = x2 , . . . , yn = xn . An invariant measure µ on a subshift Σ is a finite Borel measure on Σ which is invariant under the shift map. It extends canonically to an invariant measure on all of AZ by setting µ(B) := µ(B ∩ Σ) for any measurable set B ⊂ AZ . For any invariant measure µ on AZ obtained this way we say that µ has support in Σ. Remark 5.1. (1) Any invariant measure µ defines a function µA : A∗ → R≥0 , w 7→ µ(Cylw ) which satisfies for any w ∈ A∗ the following Kirchhoff conditions: X X µA (w) = µA (xw) = µA (wx) x∈A x∈A ∗ ∗ Conversely, for every such weight function µ : A → R≥0 (i.e. µ∗ satisfies the Kirchhoff conditions) there exists a well defined invariant measure µ on AZ with µA = µ∗ , see [FM10]. (2) The measure µ has support in a subshift Σ ⊂ AZ if and only if µA (w) = 0 for any w ∈ A∗ with Cylw ∩ Σ = ∅. A substitution ζ on the finite alphabet A = {a1 , . . . am } is a monoid morphism ζ : A∗ → A∗ . It is well defined by knowing all ζ(ai ), and conversely, any map A → A∗ defines a substitution. A substitution ζ : A → A∗ is called expanding if for each ai ∈ A, |ζ k (ai )| ≥ 2 for some k ≥ 1. For any two words w1 , w2 ∈ A∗ , let us denote the number of occurrences of the word w1 in w2 by |w2 |w1 . By proving a generalized version of the classical Perron–Frobenius Theorem for reducible matrices, in [LU17] we obtain the following result. Proposition 5.2 (Corollary 1.3 and Remark 3.14 of [LU17]). For any expanding substitution ζ : A∗ → A∗ and any letter ai ∈ A there is a well defined invariant measure µai on AZ . For any non-empty w ∈ A∗ and the associated cylinder Cylw the value of µai is given, after possibly raising ζ to a suitable power (which does not depend on w), by the limit frequency |ζ t (ai )|w . t→∞ |ζ t (ai )| µai (Cylw ) = lim Remark 5.3. (1) We note that in [LU17] the measure µai is defined on the substitution subshift Σζ , while in the above quote we denote by µai the canonical extension to all of AZ . Accordingly, in [LU17] the set Cylw denotes set Cylw ∩ Σζ . However, as explained in Remark 3.13 of [LU17], these differences are only superficial: the statement quoted above and the one in [LU17] are indeed equivalent. (2) It can be derived from [LU17] that the positive power to which ζ has to be raised in order to get the frequency equality in Proposition 5.2 is not just independent of w but actually independent of ζ. It only depends on the size of the alphabet A. In the next subsection we also need the following: Lemma 5.4 (Remark 3.3 of [LU17]). Let ζ : A∗ → A∗ be an expanding substitution. Then (up to replacing ζ by a positive power) there exists a constant λai > 1 for each ai ∈ A such 24 MARTIN LUSTIG AND CAGLAR UYANIK that: |ζ t+1 (ai )| = λai t→∞ |ζ t (ai )| lim u t 5.2. Train track maps reinterpreted as substitutions. We now apply the material from the previous subsection to a particular substitution ζf which we construct from an expanding train track map f : Γ → Γ as in the section 4. We first set up the following general definitions: For any graph Γ we recall (see section 2.1) that the set EΓ of oriented edges of Γ is equipped with a fixed point free involution ι : EΓ → EΓ, e 7→ e where e denotes as before the edge e with reversed orientation. We now use EΓ as alphabet, and consider the free monoid EΓ∗ and the both-sided full shift EΓZ . The latter contains the subshift (of “finite type”) ΣΓ ⊂ EΓZ , which consists of all biinfinite reduced edge paths in Γ. We denote the language of ΣΓ by P(Γ), it is the set of all finite reduced edge paths in Γ. We observe that the involution ι induces an involution on EΓ∗ which for convenience will be denoted also by ι. On EΓZ this involution induces the flip involution which is given by ι(. . . xi xi+1 xi+2 . . .) = . . . yi yi+1 yi+2 . . . with yn = x−n+1 for all n ∈ Z. The subshift ΣΓ is clearly flip-invariant. A shift-invariant measure µ∗ on ΣΓ is called flip-invariant if for any measurable subset Σ0 ⊂ ΣΓ one has µ∗ (ι(Σ0 )) = µ∗ (Σ0 ). From Remark 5.1 we see that this is equivalent to requiring that the weight function µ∗EΓ associated to µ∗ satisfies µ∗EΓ (γ) = µ∗EΓ (γ) for any γ ∈ P(Γ). We now observe: Lemma 5.5. Let Γ be a graph provided with a marking θ : FN → π1 Γ. Then any shiftinvariant and flip-invariant measure µ∗ on ΣΓ defines a current µ on FN and conversely. This canonical 1-1 correspondence is given by the fact that the Kolmogorov function µEΓ associated to µ (see section 2.5) and the weight function µ∗EΓ associated to µ∗ satisfy µEΓ (γ) = µ∗EΓ (γ) for any finite reduced path γ in Γ. Proof. It is well known that a current and the associated Kolmogorov function determine each other vice versa (see section 2.5), and similarly for a shift-invariant measure and its weight function (see Remark 5.1). Hence the statement of the lemma is a direct consequence of the observation that any flip-invariant weight function on EΓ∗ , which is zero on EΓ∗ r P(Γ), restricts on P(Γ) to a Kolmogorov function, and conversely, any Kolmogorov function on P(Γ) gives canonically rise to a flip-invariant weight function on EΓ by setting it equal to 0 for any element of on EΓ∗ r P(Γ). u t DYNAMICS OF HYPERBOLIC AUTOMORPHISMS OF FN 25 We now consider the occurrences of a path γ as subpath in a path γ 0 . As before, the number of such occurrences is denoted by |γ 0 |γ . We denote the number of occurrences of γ or of γ as subpath in a path γ 0 by hγ, γ 0 i and obtain: hγ, γ 0 i = |γ 0 |γ + |γ 0 |γ (5.1) Let now f : Γ → Γ be any graph self-map. Then f induces a substitution ζf : EΓ∗ → EΓ∗ , but in general ζf -iterates of reduced paths in Γ will be mapped to non-reduced paths. An exception is given if f is a train track map and if the path γ 0 is legal: In this case all f t (γ 0 ) will be reduced as well: [f t (γ 0 )] = f t (γ 0 ) where [ρ] denotes as in section 2.1 the path obtained from an edge path ρ via reduction relative to its endpoints. Hence the occurrences of any path γ or of γ in [f t (γ 0 )] are given by (5.2) hγ, [f t (γ 0 )]i = |f t (γ 0 )|γ + |f t (γ 0 )|γ for any integer t ≥ 0. We are now ready to prove: Proposition 5.6. Let f : Γ → Γ be an expanding train-track map, and let e ∈ EΓ. Then, after possibly replacing f by a positive power, for any reduced edge path γ in Γ the limit hγ, f t (e)i = aγ n→∞ |f t (e)| lim exists, and the set of these limit values defines a unique geodesic current µ+ (e) on FN through setting hγ, µ+ (e)i = aγ for any γ ∈ P(Γ). Proof. Since every edge e of Γ is legal, we obtain for any γ ∈ P(Γ) from equation (5.2) above that hγ, f t (e)i = hγ, [f t (e)]i = |f t (e)|γ + |f t (e)|γ for all exponents t ≥ 0. On the other hand, for any element ω ∈ EΓ∗ r P(Γ) one has |f t (e)|ω + |f t (e)|ω = 0. Furthermore the assumption that f is expanding implies directly that the substitution ζf associated to f is also expanding. Hence we can apply Proposition 5.2 to the substitution ζf to obtain that, after possibly t (e)i replacing f and ζf by a suitable positive power, the limit of the hγ,f exists indeed, and |f t (e)| ∗ ∗ ∗ that it is equal to µe (Cylγ ) + µe (Cylγ ), where µe is a shift-invariant measure on EΓ∗ with support in the subshift ΣΓ . Hence γ 7→ µ∗e (Cylγ ) defines a weight function on EΓ∗ which is zero outside P(Γ), and thus similarly γ 7→ µ∗e (Cylγ ). Their sum is thus also a weight function with support in P(Γ), and furthermore it is flip invariant, so that we obtain from Lemma 5.5 a Kolmogorov function. This in turn defines the current µ+ (e), as stated in our proposition. u t We now want to show that the currents µ+ (e) are projectively ϕ-invariant. For this purpose we start by stating two lemmas; the first one is an elementary exercise: Lemma 5.7. For any graph Γ without valence 1 vertices there exists a constant K ≥ 0 such that for any finite reduced edge path γ in Γ there exists an edge path γ 0 of length |γ 0 | ≤ K such that the concatenation γ ◦ γ 0 exists and is a reduced loop. u t 26 MARTIN LUSTIG AND CAGLAR UYANIK Lemma 5.8. Let f : Γ → Γ as in Proposition 5.6, and let K1 ≥ 0 be any constant. For all integers t ≥ 0 let γt0 ∈ EΓ∗ be any element with |γt0 | ≤ K1 . Set γt := f t (e)∗ γt0 , where f t (e)∗ is obtained from f t (e) by erasing an initial and a terminal subpath of length at most K1 . Then for any reduced path γ in Γ one has hγ, γt i lim = hγ, µ+ (e)i t→∞ |γt | Proof. From the hypotheses |γt0 | ≤ K1 and |f t (e)∗ | ≥ |f t (e)| − 2K1 , and from the fact that hγ,γt i f is expanding and hence |f t (e)| → ∞ for t → ∞, we obtain directly lim hγ,f t (e)i = 1 and t→∞ t| lim |f|γt (e)| = 1. Hence the claim follows directly from Proposition 5.6. t→∞ u t Proposition 5.9. Let ϕ ∈ Out(FN ) be an automorphism which is represented by an expanding train-track map f : Γ → Γ. We assume that ϕ and f have been replaced by positive powers according to Proposition 5.6. Then there exist a constant λe > 1 such that ϕ(µ+ (e)) = λe µ+ (e). Proof. For the given graph Γ let K ≥ 0 be the constant given by Lemma 5.7, and for any integer t ≥ 0 let γt0 ∈ P(Γ) with |γt0 | ≤ K be the path given by Lemma 5.7 so that γt =: f t (e)γt0 ∈ P(Γ) is a reduced loop. Let [wt ] ⊂ FN ∼ = π1 Γ be the conjugacy class represented by γt , and note that the rational current η[wt ] satisfies kη[wt ] k = |γt |, by equality (2.2) from section 2.5. Similarly, consider f (γn ) = f t+1 (e)f (γt0 ), and notice that |f (γt0 )| is bounded above by the constant K0 = K max{|f (e)| | e ∈ EΓ}. Since f is a train track map, the path f t+1 (e) is reduced. Hence the reduced loop γt00 := [f (γn )] = [f t+1 (e)f (γt0 )] can be written as product f t+1 (e)∗ γt00 with |γt00 | ≤ K1 and |f t+1 (e)∗ | ≥ |f t+1 (e)| − 2K1 , where f t+1 (e)∗ is a subpath of f t+1 (e) and K1 is the maximum of K0 and the cancellation bound Cf of f (see Lemma 2.5). Thus we can apply Lemma 5.8 twice to obtain for any reduced path γ in Γ that hγ, γt i lim = hγ, µ+ (e)i t→∞ |γt | and hγ, γt00 i lim = hγ, µ+ (e)i . t→∞ |γt00 | The first equality implies that the rational currents η[wt ] satisfy η[wt ] lim = µ+ (e). t→∞ kη[wt ] k From the continuity of the Out(FN )-action on current space and from ϕη[wt ] = ηϕ[wt ] (see equality (2.1) from section 2.5) we thus deduce ηϕ[wt ] = ϕµ+ (e). lim t→∞ kη[wt ] k However, since the reduced loops γt00 represent the conjugacy classes ϕ[wt ], the second of the above equalities implies that ηϕ[wt ] lim = µ+ (e). t→∞ kηϕ[wt ] k DYNAMICS OF HYPERBOLIC AUTOMORPHISMS OF FN 27 |γ 00 | t| Since lim |f|γt (e)| = 1 and lim |f t+1t (e)| = 1, with |γt | = kη[wt ] k and |γt00 | = kηϕ[wt ] k, the conclusion t→∞ t→∞ follows from Lemma 5.10 below. u t Lemma 5.10. For every edge e of Γ there exists a real number λe > 1 which satisfies: |f t+1 (e)| = λe t→∞ |f t (e)| lim Proof. This is a direct consequence of Lemma 5.4 and of the definition of ζf . u t We now define ∆− (ϕ) and ∆+ (ϕ) that are used in the next section: Definition 5.11. Let ϕ ∈ Out(FN ) be an outer automorphism. Assume that ϕ is replaced by a positive power such that both, ϕ and ϕ−1 are represented by expanding train-track maps as in Proposition 5.6. Let f : Γ → Γ be the representative of ϕ. Then the simplex of attraction is defined as follows: X X ∆+ (ϕ) = {[ ai µ+ (ei )] | ai ≥ 0, ai > 0}. ei ∈E + Γ Analogously, we define the simplex of repulsion as ∆− (ϕ) = ∆+ (ϕ−1 ). It should be pointed out here that distinct edges ei of Γ may well give the same limit current µ+ (ei ). Furthermore, even if a subset of currents µ+ (ei ) which define pairwise different vertices of the “simplex” ∆+ (ϕ), this subset may a priori not be linearly independent. However, it follows from general results in dynamics (see for instance [FM10]) that the extremal points of ∆+ (ϕ) correspond to ergodic currents, and that pairwise projectively distinct ergodic currents are linearly independent, so that ∆+ (ϕ) is indeed a finite dimensional simplex. From Proposition 5.9 we obtain directly: Corollary 5.12. Let ϕ ∈ Out(FN ) be as in Definition 5.11. Then the simplexes ∆+ (ϕ) and ∆− (ϕ) are ϕ-invariant: ϕ(∆+ (ϕ)) = ∆+ (ϕ) and ϕ(∆− (ϕ)) = ∆− (ϕ) u t Remark 5.13. There seems to be an interesting question as to what the maximal possible dimension of the limit simplexes ∆+ (ϕ) (or ∆− (ϕ) = δ+ (ϕ−1 )) can be. It has been shown by G. Levitt [Lev09] that for any ϕ ∈ Out(FN ) the number of exponentially growing strata in any train track representative of ϕ is bounded above by 3N4−2 (and that this bound is attained, by certain geometric automorphisms). This gives: 3N − 6 4 We do not know whether there are hyperbolic automorphisms which realize this bound. Preliminary considerations have lead us to construct for any integer k ≥ 1 a family of k+1 hyperbolic automorphisms ϕk with rank(ϕk ) = 2 · 3k + 3 and dim(∆+ (ϕk )) = 3 2 −1 . dim(∆+ (ϕ)) ≤ 28 MARTIN LUSTIG AND CAGLAR UYANIK In the next section we will specify the automorphism in question to be hyperbolic. Recall that ϕ ∈ Out(FN ) is hyperbolic if there is no non-trivial ϕ-periodic conjugacy class in FN . Let f : Γ → Γ be an absolute train track representative of ϕ. By Remark 2.4 we can assume that f is expanding. We then replace ϕ and f by a positive power so that Convention 4.1 is satisfied, and furthermore subdivide edges in accordance to Convention 4.7. Furthermore we have: Lemma 5.14. Any expanding train track representative f : Γ → Γ of a hyperbolic automorphism ϕ ∈ Out(FN ) which satisfies Convention 4.7 also satisfies the hypotheses of Convention 4.14. Proof. This is a direct consequence of Lemma 4.15. u t Convention 5.15. A hyperbolic automorphism ϕ ∈ Out(FN ) is said to have a properly expanding train track representative f : Γ → Γ if f satisfies Conventions 4.1, 4.7 and 4.14, and if both have been raised to a suitable positive power according to Propositions 5.6 and 5.9. Remark 5.16. From Lemma 5.14 one obtains directly that every hyperbolic automorphism ϕ which possesses an absolute train track representative has a positive power ϕt which possesses a properly expanding train track representative. 6. Hyperbolic automorphisms Let ϕ ∈ Out(FN ) be a hyperbolic automorphism which possesses a properly expanding train track representative f : Γ → Γ as in Convention 5.15. Let [w] be a conjugacy class in FN . Represent [w] by a reduced loop γ in Γ. Then the goodness of w, denoted by g([w]), is defined by: g([w]) := g(γ) Lemma 6.1. Given a neighborhood U of the simplex of attraction ∆+ (ϕ) ∈ PCurr(FN ), there exist a bound δ > 0 and an integer M = M (U ) ≥ 1 such that, for any [w] ∈ FN with g([w]) ≥ δ, we have (ϕM )n [ηw ] ∈ U for all n ≥ 1. Proof. We first replace ϕ by a positive power as in Proposition 4.4 so that the goodness function for the train-track map f : Γ → Γ becomes monotone. Recall from Section 2.5 that [ν] ∈ PCurr(FN ) is close to [ν 0 ] ∈ PCurr(FN ) if there exists ε > 0 and R >> 0 such that for all reduced edge paths γ with |γ| ≤ R we have hγ, νi hγ, ν 0 i − < ε. kνkΓ kν 0 kΓ Thus, since ∆+ (ϕ) is compact, there exist ε > 0 and R ∈ R such that the above inequalities imply for ν 0 ∈ ∆+ (ϕ) that ν ∈ U . DYNAMICS OF HYPERBOLIC AUTOMORPHISMS OF FN 29 We proved the pointwise convergence for edges in Proposition 5.6. Since there are only finitely many edges and finitely many edge paths γ in Γ with |γ| ≤ R, we can pick an integer M0 ≥ 0 such that (6.1) hγ, f n (e)i − hγ, µ+ (e)i < ε/4 |f n (e)| for all n ≥ M0 , for all γ with |γ| ≤ R and for all edges e of Γ. Let λ0 , λ00 > 1 be the minimal and the maximal expansion factors respectively as given in 1 Convention 4.1. For any reduced loop c in Γ with g(c) ≥ 1+ε/4 iterative application of the fact, that f maps any good edge in c to to a path in [f (c)] which consists entirely of good edges and has length at least λ0 , implies: |f n (c)| ≥ 1 |c|(λ0 )n . 1 + ε/4 Thus for any integer M1 > logλ0 R(1 + 4ε ) and all n ≥ M1 we get the following inequalities: (6.2) R|c| R|c|Γ (1 + ε/4) R(1 + ε/4) ε R(1 + ε/4) ≤ ≤ = ≤ |f n (c)| |c|(λ0 )n (λ0 )M1 R(1 + 4/ε) 4 We note that for any integer m ≥ 1 and for each edge the minimum expansion factor for f m is at least (λ0 )m and the maximum expansion factor for f m is at most (λ00 )m . For the rest of the proof set M = max{M0 , M1 } and δ := max{ 1 1 }, , 0 λ 1 + ε/4 1 + ( λ00 )M ε/4 and let c be a reduced loop in Γ which represents a conjugacy class w with g(w) ≥ δ. The assertion of Lemma 6.1 now follows if we show that for all integers n ≥ 1 the current (ϕM )n ([ηw ]) is (ε, R)-close (in the above sense) to some point in ∆+ (ϕ). Indeed, since by the first paragraph of the proof the goodness function is monotone, it suffices to assume n = 1 and apply the resulting statement iteratively. For simplicity we denote from now on f M by f . Another auxiliary computation gives: (6.3)     (λ00 )M · #{bad edges in c} (λ00 )M (1 − g(w))|c| λ00 M 1 λ00 M 1 ε = =( 0) −1 ≤( 0) −1 ≤ . 0 M 0 M (λ ) · #{good edges in c} (λ ) g(w)|c| λ g(w) λ δ 4 We now write c = c1 c2 . . . b1 . . . cp cp+1 . . . cr b2 cr+1 . . . bk . . . cs−1 cs , where the ci denote good edges and the bj denote maximal bad subpaths of c. Note for the second of the inequalities below that the definition of “good” implies that there can be no cancellation in f (c) between adjacent f (ci ) and f (ci+1 ) nor between adjacent f (ci ) and f (bj ), see Lemma 4.3. 30 MARTIN LUSTIG AND CAGLAR UYANIK Then we calculate: hγ, f (c)i hγ, |f (c1 )|µ+ (c1 ) + . . . + |f (cs )|µ+ (cs )i Ps − |f (c)| i=1 |f (ci )| s hγ, f (c)i X hγ, f (ci )i ≤ − |f (c)| |f (c)| i=1 P s s X hγ, f (ci )i hγ, f (ci )i P − i=1 + s |f (c)| i=1 |f (ci )| i=1 P s hγ, |f (c1 )|µ+ (c1 ) + . . . + |f (cs )|µ+ (cs )i i=1 hγ, f (ci )i Ps + P − s i=1 |f (ci )| i=1 |f (ci )| k 2R|c| X hγ, [f (bj )]i + |f (c)| j=1 |f (c)| Ps Ps hγ, f (c )i i i=1 hγ, f (ci )i i=1 − P + Ps Pk s i=1 |f (ci )| j=1 |[f (bj )]| i=1 |f (ci )| + Ps hγ, |f (c1 )|µ+ (c1 ) + . . . + |f (cs )|µ+ (cs )i i=1 hγ, f (ci )i Ps + P − s i=1 |f (ci )| i=1 |f (ci )| < ε/4 + ε/4 + ε/4 + ε/4 = ε. ≤ Here the first inequality is just a triangle inequality. In the second inequality the last two terms are unchanged, and first two terms come from the first term in the previous quantity and follows from counting frequencies as follows: An occurrence of an edge path γ or its inverse can occur either inside the image of a good edge f (ci ) or inside of the image of a bad segment [f (bj )], or it might cross over the concatenation points. This observation gives the claimed inequality. In the final inequality, the first ε/4 follows from the equation (6.2). The second one follows from the equation (6.3) as follows: k X hγ, [f (bj )]i j=1 |f (c)| ≤ k X |[f (bj )]| j=1 |f (c)| ≤ k X j=1 (λ0 )M |bj |(λ00 )M ε ≤ . 4 · #{good edges in c} DYNAMICS OF HYPERBOLIC AUTOMORPHISMS OF FN 31 The third ε/4 comes from the following observation: Ps Ps i=1 hγ, f (ci )i i=1 hγ, f (ci )i − P Ps Pk s i=1 |f (ci )| i=1 |f (ci )| + j=1 |[f (bj )]|
Pk
Ps i=1 hγ, f (ci )i j=1 |[f (bj )]| = Ps
Ps
Pk i=1 |f (ci )| i=1 |f (ci )| + j=1 |[f (bj )]| Pk Ps hγ, f (c )i i j=1 |[f (bj )]| i=1 · Ps ≤ P s i=1 |f (ci )| i=1 |f (ci )| P (λ00 )M kj=1 |bj | P ≤ ≤ ε/4 (λ0 )M si=1 |ci | by (6.3). Finally, the last ε/4 can be verified using (6.1) as follows: Ps hγ, |f (c1 )|µ+ (c1 ) + . . . + |f (cs )|µ+ (cs )i i=1 hγ, f (ci )i P Ps − s i=1 |f (ci )| i=1 |f (ci )| Ps hγ, f (ci )i − hγ, µ+ (ci )i) i=1 |f (ci )|( |f (ci )| Ps = i=1 |f (ci )| Ps i=1 |f (ci )|ε/4 ≤ P = ε/4 s i=1 |f (ci )| Since after applying ϕM to a conjugacy class with g(w) ≥ δ we still have g(ϕM (w)) ≥ δ we get (ϕM )n ([ηw ]) ∈ U for all n ≥ 1. u t Lemma 6.2. For any δ > 0 and any neighborhood U of ∆+ (ϕ) there exists an integer M (δ, U ) > 0 such that (up to replacing ϕ by a positive power) for any conjugacy class w with goodness g(w) > δ we have ϕn [ηw ] ∈ U for all n ≥ M . Proof. This is a direct consequence of Corollary 4.5 and Lemma 6.1. u t Proposition 6.3. Let ϕ ∈ Out(FN ) be a hyperbolic outer automorphism such that ϕ and ϕ−1 both admit properly expanding train-track representatives. Given neighborhoods U of the simplex of attraction ∆+ (ϕ) and V of the simplex of repulsion ∆− (ϕ) in PCurr(FN ), then (up to replacing ϕ by a positive power) there exist an integer M ≥ 0 such that for any conjugacy class w ∈ FN we have ϕn [ηw ] ∈ U for all n ≥ M . or ϕ−n [ηw ] ∈ V 32 MARTIN LUSTIG AND CAGLAR UYANIK Proof. This is a direct consequence of Proposition 4.20 (see Remark 4.21) and Lemma 6.2. u t The following result also proves Theorem 1.1 from the Introduction: Theorem 6.4. Let ϕ ∈ Out(FN ) be a hyperbolic outer automorphism such that ϕ and ϕ−1 admit absolute train-track representatives. Then, ϕ acts on PCurr(FN ) with uniform NorthSouth dynamics from ∆− (ϕ) to ∆+ (ϕ): Given an open neighborhood U of the simplex of attraction ∆+ (ϕ) and a compact set K ⊂ PCurr(FN ) r ∆− (ϕ) there exists an integer M > 0 such that ϕn (K) ⊂ U for all n ≥ M. Proof. According to Remark 5.16 we can pass to common positive powers of ϕ and ϕ−1 that have a properly expanding train track representatives as in Convention 5.15. We can then combine Propositions 2.8, 3.3, 3.4 and 6.3 to obtain directly the required result. For the application of Proposition 3.3 we need that ∆+ (ϕ) and ∆− (ϕ) are disjoint, which is shown in Remark 6.5 below. We also need for Proposition 3.3 that both, ∆+ (ϕ) and ∆− (ϕ), are ϕ-invariant, which is shown in Corollary 5.12. u t Remark 6.5 (Dynamics within ∆+ (ϕ)). (1) It is proved in Proposition 5.6 that every vertex of the simplex ∆+ = ∆+ (ϕ) is an expanding ϕ-invariant current, i.e. a projectivized current [µ] for which there exist λ > 1 and t ≥ 1 such that ϕt (µ) = λµ. For the rest of this remark we replace ϕ by a suitable positive power so that every vertex current of ∆+ is projectively fixed by ϕ. Of course, this implies that ϕ fixes also every face ∆0 of ∆+ (but not necessarily pointwise). (2) A uniform face ∆0 is a face of ∆+ which is spanned by vertices [µ1+ ], . . . , [µk+ ] that all have the same stretch factor λ > 1, i.e. ϕ(µj+ ) = λµj+ for all 1 ≤ j ≤ k. (3) Since the action of ϕ on Curr(FN ) is linear, all the uniform faces of ∆+ are pointwise fixed. Any non-vertex current [µ] is always contained in the interior of some face ∆0 ⊂ ∆+ . Then the sequence of ϕn ([µ]) converges towards a point in the uniform face ∆0+ of ∆0 which is spanned by all vertices that have maximal stretch factor among all vertices of ∆0 . (4) Similarly, under backwards iteration n → −∞ the sequence of ϕn ([µ]) converges towards a point in the uniform face ∆0− of ∆0 ⊂ ∆+ which is spanned by all vertices that have minimal stretch factor among all vertices of ∆0 . (5) As a consequence, it follows that a current [µ] can not belong to both, ∆+ (ϕ) and ∆− (ϕ): By symmetry in the previous paragraph, any [µ] ∈ ∆− (ϕ) converges under backwards iteration of ϕ−1 to an expanding ϕ−1 -invariant current. But backwards iteration of ϕ−1 is the same as forward iteration of ϕ, and the limit current can not be simultaneously expanding ϕ-invariant and expanding ϕ−1 -invariant. Remark 6.6. As recalled in Remark 6.5, every vertex of the convex cell ∆+ = ∆+ (ϕ) is an expanding ϕ-invariant current. Furthermore, any point in a uniform face of ∆+ is also an expanding ϕ-invariant current, and these are the only such in ∆+ . It follows from Theorem 6.4 and from Remark 6.5 (applied to ϕ−1 ) that there are no other expanding ϕ-invariant currents in PCurr(FN ) r ∆+ . DYNAMICS OF HYPERBOLIC AUTOMORPHISMS OF FN 33 It has been shown in [BHL15] in a slightly more general context that expanding fixed currents (i.e. currents µ with ϕµ = λϕ for some λ > 1) are in 1-1 relation with the nonnegative eigenvectors with eigenvalue > 1 for the transition matrix M (f ) of any expanding train track representative f : Γ → Γ of ϕ. It seems likely that a similar result can be obtained by extending the methods of Section 5. Moreover, pointwise dynamics for a dense subset follows from our machinery: Theorem 6.7. Let ηg be a rational current in Curr(FN ). Then, there exist [µ∞ ] ∈ ∆+ such that lim ϕt [ηg ] = [µ∞ ]. t→∞ Proof. Let γ be a reduced loop representing the conjugacy class g. Apply a sufficiently large power ϕk to g, hence [f k (γ)], the reduced edge path representing ϕk (g) is a pseudo-legal loop, see Proposition 4.12. Then the arguments in in the proof of Lemma 6.1 give the pointwise convergence by looking at the set of non-INP edges in [f k (γ)]. u t Remark 6.8. From the proof of Theorem 6.7 and from the results of our previous paper [LU17] one can also derive a precise description of the limit current µ∞ for any given rational current ηg : Let [f k (γ)] be the pseudo-legal loop representing the conjugacy class ϕk (g), and assume without loss of generality that the initial point doesn’t lie on any INP-subpath. Let [f k (γ)] = c1 c1 . . . cp ρ1 cp+1 cp+1 . . . ρ2 cr cr+1 . . . cs−1 cs be a legal concatenation of edges ci and of INP’s ρj . Then the limit current µ∞ is given by X µ+ (ci )] , [µ∞ ] = [ ci ∈EΓ0 where Γ0 is the (possibly non-connected) subgraph of Γ consisting precisely of those strata Hi of Γ which satisfy one of the following two conditions, where  denotes the natural partial order on strata defined by the non-zero off-diagonal coefficients of the transition matrix M (f ) (see [LU17]): (1) There exists a chain Hi = Hd  Hd−1  . . .  H1 of pairwise distinct strata, all with maximal Perron-Frobenius eigenvalue among the diagonal blocks of M (f ). Furthermore d is the maximal length of any such chain. (2) Hi  Hj for some Hj as in (1). The analogue of Theorem 6.7 for non-rational currents is delicate, but the authors expect that it is actually true. However, as already cautioned in the Introduction, in general the pointwise convergence on PCurr(FN ) will not be uniform. To finish the paper, we would like to illustrate our main result by considering the particular case of a hyperbolic ϕ ∈ Out(FN ) which is given as a “free product” of fully irreducible hyperbolic outer automorphisms: Theorem 6.4 then shows that such a free product automorphism acts on the space of projectivized geodesic currents with generalized uniform North-South dynamics, and the simplex of attraction is spanned by the forward limit currents for each of the fully irreducible factors. More precisely: 34 MARTIN LUSTIG AND CAGLAR UYANIK Example 6.9. Let ϕi ∈ Out(FNi ) (with Ni ≥ 3) be a finite family of hyperbolic fully irreducible outer automorphisms, where each ϕi : FNi → FNi is represented by an absolute train-track map fi : Γi → Γi . For simplicity we now assume (for example by passing to a positive power) that every fi has a fixed vertex vi in Γi , and we consider the lift Φi ∈ Aut(FN ) represented by fi on π1 (Γi , vi ). We then get a global absolute train-track map f by wedging together the Γi at the fixed points vi , and f represents the outer automorphism ϕ given by Φ = Φ1 ∗ . . . ∗ Φk ∈ Aut(FN ), for FN := FN1 ∗ . . . ∗ FNk . Then the Bestvina-Feighn Combination Theorem implies that the “free product” ϕ is hyperbolic. We consider for any i the embedding κi : Curr(FNi ) ,→ Curr(FN ) as described in [KL10c]. In [Uya14] it was proven that for each factor graph Γi the current µ+ (e) defined in Proposition 5.6 is the same for every edge e of Γi . Hence the simplex of attraction for ϕ is spanned by k currents µ+ (ei ), each of which is given by an edge ei from a different Γi . In fact, each µ+ (ei ) is the κi -image of the well defined forward limit current for ϕi in Curr(FN ). Analogously, the simplex of repulsion is equal to the linear span of the images in Curr(FN ) of the backwards limit currents for all ϕ. References [BF92] [BF93] [BF14] [BFH00] [BHL15] [Bon91] [BR15] [Bri00] [CHL08] [CL95] [CL15] [Coo87] [CV86] [DKL15] M. Bestvina and M. Feighn. A combination theorem for negatively curved groups. J. Differential Geom., 35(1):85–101, 1992. Mladen Bestvina and Mark Feighn. Outer limits. 1993. http://andromeda.rutgers.edu/ feighn/papers/outer.pdf. Mladen Bestvina and Mark Feighn. Hyperbolicity of the complex of free factors. Adv. Math., 256:104–155, 2014. Mladen Bestvina, Mark Feighn, and Michael Handel. The Tits alternative for Out(Fn ). I. Dynamics of exponentially-growing automorphisms. Ann. of Math. (2), 151(2):517–623, 2000. Nicolas Bédaride, Arnaud Hilion, and Martin Lustig. Invariant measures for train track towers. arXiv preprint arXiv:1503.08000, 2015. Francis Bonahon. Geodesic currents on negatively curved groups. In Arboreal group theory, pages 143–168. Springer, 1991. Mladen Bestvina and Patrick Reynolds. The boundary of the complex of free factors. Duke Math. J., 164(11):2213–2251, 2015. P. Brinkmann. Hyperbolic automorphisms of free groups. Geom. Funct. Anal., 10(5):1071–1089, 2000. Thierry Coulbois, Arnaud Hilion, and Martin Lustig. R-trees and laminations for free groups. III. Currents and dual R-tree metrics. J. Lond. Math. Soc. (2), 78(3):755–766, 2008. Marshall M. Cohen and Martin Lustig. Very small group actions on R-trees and Dehn twist automorphisms. Topology, 34(3):575–617, 1995. Thierry Coulbois and Martin Lustig. Long turns, INP’s and indices for free group automorphisms. Illinois J. Math., 59(4):1087–1109, 2015. Daryl Cooper. Automorphisms of free groups have finitely generated fixed point sets. J. Algebra, 111(2):453–456, 1987. Marc Culler and Karen Vogtmann. Moduli of graphs and automorphisms of free groups. Invent. Math., 84(1):91–119, 1986. Spencer Dowdall, Ilya Kapovich, and Christopher J Leininger. Dynamics on free-by-cyclic groups. Geom. Topol, 19(5):2801–2899, 2015. DYNAMICS OF HYPERBOLIC AUTOMORPHISMS OF FN [FM10] [Ham12] [HM13] [Iva92] [Kap05] [Kap06] [KL09] [KL10a] [KL10b] [KL10c] [Lev09] [LL03] [LU16] [LU17] [Mah11] [Mar95] [Riv08] [Thu88] [Uya14] [Uya15] 35 Sébastien Ferenczi and Thierry Monteil. Infinite words with uniform frequencies, and invariant measures. In Combinatorics, automata and number theory, volume 135 of Encyclopedia Math. Appl., pages 373–409. Cambridge Univ. Press, Cambridge, 2010. Ursula Hamenstaedt. The boundary of the free splitting graph and the free factor graph. arXiv preprint arXiv:1211.1630, 2012. Michael Handel and Lee Mosher. The free splitting complex of a free group, I: hyperbolicity. Geom. Topol., 17(3):1581–1672, 2013. Nikolai V Ivanov. Subgroups of Teichmuller modular groups, volume 115. American Mathematical Soc., 1992. Ilya Kapovich. The frequency space of a free group. Internat. J. Algebra Comput., 15(5-6):939–969, 2005. Ilya Kapovich. Currents on free groups. In Topological and asymptotic aspects of group theory, volume 394 of Contemp. Math., pages 149–176. Amer. Math. Soc., Providence, RI, 2006. Ilya Kapovich and Martin Lustig. Geometric intersection number and analogues of the curve complex for free groups. Geom. Topol., 13(3):1805–1833, 2009. Ilya Kapovich and Martin Lustig. Domains of proper discontinuity on the boundary of outer space. Illinois J. Math., 54(1):89–108, 2010. Ilya Kapovich and Martin Lustig. Intersection form, laminations and currents on free groups. Geom. Funct. Anal., 19(5):1426–1467, 2010. Ilya Kapovich and Martin Lustig. Ping-pong and outer space. J. Topol. Anal., 2(2):173–201, 2010. Gilbert Levitt. Counting growth types of automorphisms of free groups. Geometric and Functional Analysis, 19(4):1119–1146, 2009. Gilbert Levitt and Martin Lustig. Irreducible automorphisms of Fn have north-south dynamics on compactified outer space. J. Inst. Math. Jussieu, 2(1):59–72, 2003. Martin Lustig and Caglar Uyanik. North-south dynamics of hyperbolic free group automorphisms on the space of currents. arXiv preprint arXiv:1509.05443v2, 2016. Martin Lustig and Caglar Uyanik. Perron-Frobenius theory and frequency convergence for reducible substitutions. Discrete and Continuous Dynamical Systems, 37(1):355–385, 2017. Joseph Maher. Random walks on the mapping class group. Duke Mathematical Journal, 156(3):429– 468, 2011. Reiner Martin. Non-uniquely ergodic foliations of thin-type, measured currents and automorphisms of free groups. ProQuest LLC, Ann Arbor, MI, 1995. Thesis (Ph.D.)–University of California, Los Angeles. Igor Rivin. Walks on groups, counting reducible matrices, polynomials, and surface and free group automorphisms. Duke Mathematical Journal, 142(2):353–379, 2008. William P. Thurston. On the geometry and dynamics of diffeomorphisms of surfaces. Bull. Amer. Math. Soc. (N.S.), 19(2):417–431, 1988. Caglar Uyanik. Dynamics of hyperbolic iwips. Conformal Geometry and Dynamics of the American Mathematical Society, 18(11):192–216, 2014. Caglar Uyanik. Generalized north-south dynamics on the space of geodesic currents. Geometriae Dedicata, 177(1):129–148, 2015. Aix Marseille Université, CNRS, Centrale Marseille, I2M UMR 7373 13453, Marseille, France E-mail address: Martin.Lustig@univ-amu.fr Department of Mathematics, University of Illinois at Urbana-Champaign, 1409 West Green Street, Urbana, IL 61801, USA, http://www.math.uiuc.edu/~cuyanik2 E-mail address: cuyanik2@illinois.edu 

4math.GR

Further insights into the damping-induced self-recovery phenomenon Tejas Kotwal ∗, Roshail Gerard †and Ravi Banavar ‡ arXiv:1709.05596v4 [cs.SY] 11 Jan 2018 January 12, 2018 1 Abstract In a series of papers [1, 2, 3, 4], D. E. Chang, et al., proved and experimentally demonstrated a phenomenon they termed “damping-induced self-recovery”. However, these papers left a few questions concerning the observed phenomenon unanswered- in particular, the effect of the intervening lubricantfluid and its viscosity on the recovery, the abrupt change in behaviour with the introduction of damping, a description of the energy dynamics, and the curious occurrence of overshoots and oscillations and its dependence on the control law. In this paper we attempt to answer these questions through theory. In particular, we derive an expression for the infinite-dimensional fluid-stool-wheel system, that approximates its dynamics to that of the better understood finite-dimensional case. 2 Introduction The damping-induced self-recovery phenomenon refers to the fundamental property of underactuated mechanical systems: if an unactuated cyclic variable is subject to a viscous damping-like force and the system starts from rest, then the cyclic variable will always recover to its initial state as the actuated variables are brought to rest. A popular illustration exhibiting self-recovery is when a person sits on a rotating stool with damping, holding a wheel whose axis is parallel to the stool’s axis. The wheel can be spun and stopped at will by the person (Refer to [5]). Initially, the system begins from rest; when the person starts spinning the wheel (say anticlockwise), the stool begins moving in the expected direction (i.e., clockwise). Then the wheel is brought to a halt; the stool then begins a recovery by going back as many revolutions in the reverse direction (i.e., anticlockwise) as traversed before. This phenomenon defies conventional intuition based on well-known conservation laws. Andy Ruina was the first to report this phenomenon in a talk, where he demonstrated a couple of experiments on video [6]. Independently, Chang et al., [1, 2] showed that in a mechanical system with an unactuated cyclic variable and an associated viscous damping force, a new momentum-like quantity is conserved. When ∗ Department of Mathematics of Mechanical Engineering ‡ Systems and Control Engineering, Indian Institute of Technology Bombay, Mumbai, India † Department 1 the other actuated variables are brought to rest, the conservation of this momentum leads to asymptotic recovery of the cyclic variable to its initial position. Boundedness is another associated phenomenon, in which the unactuated cyclic variable reaches a saturation eventually, when the velocity corresponding to the actuated variable is kept constant; this occurs due to the presence of damping. In the experiment explained above, this manifests as the angle of rotation of the stool reaching an upper limit, when the wheel is spinning at a constant speed. Chang et al. generalize this theory to an infinite-dimensional system in which the interaction of an intermediate fluid is considered [3, 4]; they show that the fluid layers also display self-recovery, which is confirmed via experiments as well [7]. Such a generalization of the model is considered because of the interaction of the fluid in the bearing with the recovery phenomenon of the stool, in the experiment explained above. In this work, we make the following points: • We show that the dynamics of the stool and the wheel in the infinite-dimensional fluid system can be approximated to that of the finite-dimensional case by finding an effective damping constant that takes into account the effect of the viscous fluid on the system. • We analyse the finite-dimensional system from a dynamical systems point of view, and show that a bifurcation occurs when the damping constant switches from zero to a positive value. We also derive an expression for the angle at which boundedness occurs. • In addition to the recovery phenomenon described previously, further complex behaviour is observed in the experiments reported in [5, 7]. In particular, the unactuated variable not only approaches its initial state during recovery, but also overshoots and then oscillates about this initial position, eventually reaching it asymptotically. This oscillation phenomenon has not been looked into in previous works, and is one of the points that we address as well. • In Chang et al. [5], in the experiment described, the oscillations are of significant amplitude, and this would prompt one to assume that some sort of mechanical ‘spring-like’ energy is being stored, as the stool appears to start moving after the entire system has come to a halt. The question of energy has been touched upon in this work, and we present energy balance equations for the given mechanical system. The paper unfolds as follows. Initially we present mathematical models for the stool-wheel experiment - the first is a finite-dimensional one, and the second one incorporates the intervening fluid (either in the bearing or in a tank) using the Navier-Stokes equation for a Newtonian incompressible fluid. This is followed by a section that presents an intuitive interpretation of recovery highlighting three distinct types of behaviour. Then follows a theoretical section that presents a technique to reduce the infinitedimensional fluid effect to an effective damping constant and hence model the overall system in finite dimensions. This part is followed up by a result on boundedness and the occurrence of a bifurcation in the system dynamics. In the appendix, the derived results are used in conjunction with numerical experiments to validate the expression for the effective damping constant. We then examine the case of 2 oscillations and overshoots, and present plausible explanations to why these occur, and possible sources of future investigation. 3 Mathematical models Finite-dimensional model: We first analyze a simplified, idealized version of the person with a wheel in hand, sitting on a rotatable stool whose motion is opposed by damping, which for the purpose of analysis is assumed to be linear viscous damping. This is a specific example of the model that Chang, et al. studied [1, 2]. We assume two flat disks, one for the wheel and one for the stool-person mass as shown in Fig. 1. The stool consists of an internal motor, that actuates the wheel, while the motor-rod-stool setup rotates as one piece (henceforth just called the stool). There is linear viscous damping present in the rotational motion of the stool, with damping coefficient k. The torque imparted on the wheel by the motor is denoted by u(t). Figure 1: A schematic diagram of the wheel-stool model. The inertia matrix of the system is given by  (mij ) =  m11 m21  I = w m22 Iw m12  Iw Iw + Is  , (1) where Iw and Is are the moments of inertia of the wheel and stool respectively. The kinetic energy of the described system is K.E.(t) = 1 1 Iw (θ̇w + φ̇s )2 + Is φ̇2s , 2 2 (2) where θw denotes the angle rotated by the wheel relative to the stool, and φs denotes the angle rotated by the stool relative to the ground frame. Since there is no external potential in our system, the 3 Lagrangian only comprises of the total kinetic energy. The Euler-Lagrange equations for the system are given by Iw φ̈s + Iw θ̈w = u(t) (3a) (Iw + Is )φ̈s + Iw θ̈w = −k φ̇s (3b) Although this model captures the damping-induced boundedness and recovery phenomena, further behaviour that cannot be explained with this model are the overshoot and oscillations, which were observed in experiments [5, 7]. It was suggested that the lubricating fluid inside the bearing of the rotatable stool could be one of the reasons behind this additional behaviour. As a result, a generalized model that involves the dynamics of the fluid interacting with the stool-wheel setup is analyzed [3, 4]. Infinite-dimensional fluid model: The model consists of two concentric infinitely long cylinders with inner radius Ri and outer radius Ro . The outer radius is fixed, while the inner cylinder is free to rotate (this is the stool in our case). A motor is installed in the inner cylinder (as described in the previous case, i.e., φs ), to drive the wheel (i.e., θw ) attached to it. The annulus region between the two cylinders is filled with an incompressible viscous fluid. Due to symmetry in the z direction, we may regard this as a 2D system in a horizontal plane. The moments of inertia should be considered as per unit depth, because of the infinite length in the z direction. Due to rotational symmetry, the fluid only flows coaxially, i.e, there is no fluid flow in the radial direction. Let v(r, t) denote the tangential velocity of the fluid at radius r and time t, where Ri ≤ r ≤ Ro . The subscripts t and r denote partial derivatives with respect to time and radius respectively. As described before, u(t) is the driving torque on the wheel imparted by the motor. The corresponding equations are Iw φ̈s + Iw θ̈w = u(t) (4a)
(Iw + Is )φ̈s + Iw θ̈w = 2πρνRi Ri vr (Ri , t) − v(Ri , t) (4b) vt = ν(vrr + vr v − 2) r r (4c) where ν is the kinematic viscosity and ρ is the density of the fluid. The right hand side of Eq. (4b) is the torque on the inner cylinder due to stress exerted by the surrounding fluid [8]. Equation (4c) is the Navier-Stokes equation for an incompressible viscid fluid in radial coordiantes. The system (4) is an infinite-dimensional fluid system, compared to the previous finite-dimensional system (3). The initial and boundary conditions are given by φs (0) = φ̇s (0) = θw (0) = θ̇w (0) = 0 (5a) v(r, 0) = 0 (5b) v(Ri , t) = Ri φ̇s (t) (5c) v(Ro , t) = 0 (5d) 4 where Eqs. (5c) and (5d) are the no-slip boundary conditions for the fluid. The given initial conditions imply that the entire system begins from rest. In order to control the wheel independently of the effects of inertia of the stool and damping of the fluid, we employ a method known as partial feedback linearization. This is done by partially cancelling the nonlinearities in the dynamics, i.e. the wheel equations are linearized by introducing a new input τ (t) and redefining the input u(t) as  u(t) = Iw Iw + Is  Is τ (t) + 2πρνRi Ri vr (Ri , t) − v(Ri , t)

(6) where τ (t) is precisely equal to the acceleration of the wheel, i.e. θ̈w = τ (t). We assume a PD control law for the new control variable given as

d d d τ (t) = θ̈w (t) + c1 θ̇w (t) − θ̇w + c0 θw (t) − θw (7) d where θw (t) is the desired trajectory of the wheel, and c0 and c1 are the proportional and differential gains respectively. For the standard demonstration of recovery and/or boundedness, we require the wheel to be driven from rest to constant velocity, and then abruptly braked to a stop. Such a desired trajectory

d (t) = (1/2) θ̇steady |t| − |t − tstop | + tstop where θ̇steady is of the wheel is given by a ramp function, θw the desired constant velocity of the wheel, and tstop is the time at which the wheel is instantaneously brought to rest (Refer to Fig. 2). By appropriately tuning the values of c0 and c1 , we can mimic the desired trajectory. This tuning is important in accounting for the presence of oscillations, or lack thereof, Desired trajectory of the wheel 3wd (rad) as we shall see later. slope = 3_steady 0 0 tstop Time t (s) d Figure 2: An illustration of the desired trajectory θw (t) in the control law of the wheel (7) . 4 An intuitive interpretation of recovery We present a qualitative interpretation of the boundedness and recovery phenomenon using the momentum equation of the finite dimensional stool-wheel case. A brief explanation of the recovery phenomenon was also presented by Andy Ruina [6] which we summarize: Change in angular momentum is equal to 5 the net external torque, but since it is only due to linear viscous damping, we have L̇ = −cϕ̇ (where L is the angular momentum, c is the damping constant and ϕ is the angle). If the system’s initial and final state is the rest state, then ∆L = 0 =⇒ ∆ϕ = 0. Thus the net change in angle has to be zero, implying that recovery must occur. We delve a little deeper into the phenomenon and attempt to provide an intuitive interpretation for the reversal of direction of the stool. Consider the case where the wheel follows the desired trajectory perfectly (i.e., the ramp function). This means that at t = 0, an impulsive acceleration gets the wheel spinning instantaneously (to a constant speed), and at t = tstop , an impulsive deceleration brings the wheel to a halt instantaneously (refer to Fig. 3). Consider the damping-induced momentum equation, as derived by Chang, et al. [1, 2] Acceleration of wheel 3Bw (rad=s2 ) Iw (θ̇w + φ̇s ) + Is (φ̇s ) + kφs = 0 (8) 2000 1000 0 -1000 -2000 0 2 4 6 8 10 Time t (s) Figure 3: Acceleration vs time for the wheel. The first peak corresponds to the initial impulse to get the wheel spinning to a constant velocity; the second negative impulse is to brake the wheel to a halt. At t = 0, there is a jump in θ̇w , while φs remains constant (i.e., φs = 0) (refer to Figs. 8 and 9). The momentum equation (8) then simplifies to Iw (θ̇w (0) + φ̇s (0)) + Is (φ̇s (0)) = 0 (9) which is the usual momentum conservation. Thus, at t = 0, the velocity of the stool is governed by the conservation of standard angular momentum, and the velocity of the wheel. Let this velocity be φ̇s (0) = φ̇fs . At t = tstop , there is an instantaneous jump down to zero for θ̇w , while φs remains constant. This time, the momentum equation (8) simplifies to Iw (θ̇w (tstop ) + φ̇s (tstop )) + Is (φ̇s (tstop )) = C (10) where C is a constant (C = −kφs (tstop )). Once again, this equation may be interpreted as the standard angular momentum conservation. We examine three distinct cases: • Case 1: (No damping) The speed of the stool at t = tstop − ε is the same as what it was at t = 0, 6 (i.e. φ̇fs ), implying that at t = tstop , the stool comes to rest instantaneously (i.e., φ̇s (tstop ) = 0). Recall, that this is the standard case, where the usual momentum conservation law holds. Refer to Figs. 7 and 4. Position of stool ?s (rad) 0 -0.2 -0.4 -0.6 -0.8 -1 0 2 4 6 8 10 Time t (s) Figure 4: Case 1. k = 0, tstop = 5, vw = 2. This case corresponds to the usual momentum conservation. • Case 2: (Damping present but damping induced boundedness not yet reached) During 0 < t < tstop , the damping force keeps decreasing the speed of the stool. However, the momentum equation (10) is still the same at t = tstop . Thus, the change in momentum of the stool would be the same (as in Case 1). But since |φ̇s (tstop − ε)| < |φ̇fs |, the final momentum of the wheel ends up overshooting zero at t = tstop , resulting in the change in direction of the stool. Refer to Figs. 7 and 5. • Case 3: (Damping present and damping-induced boundedness has been attained) This case can be viewed as a special instance of Case 2, where for 0 < t < tstop , the speed of the stool decreases all the way to zero. The rest of the explanation follows as in Case 2. Refer to Figs. 7 and 6. 7 Position of stool ?s (rad) 0 -0.02 -0.04 -0.06 -0.08 -0.1 0 2 4 6 8 10 Time t (s) Figure 5: Case 2. k = 1, tstop = 1, vw = 2. In this case, damping-induced boundedness is not yet reached, but the recovery phenomenon can be seen. Position of stool ?s (rad) 0 -0.05 -0.1 -0.15 0 2 4 6 8 10 Time t (s) Figure 6: Case 3. k = 1, tstop = 5, vw = 2. In this case, both—damping-induced boundedness and recovery phenomenon can be seen. 8 Change in velocity of the stool before and after braking 0 Case 1 Case 2 Case 3 Figure 7: Comparison of the three cases before and after braking. Blue dots: velocity of stool just before braking; red dots: velocity of stool immediately after braking. For cases 2, 3, the (magnitude of) velocity of the stool before braking decreases due to damping; hence momentum transferred by the wheel overshoots the stool velocity to something positive. This explains why the wheel changes direction. 9 5 Effective damping constant In this section, we demonstrate that the dynamics of the infinite-dimensional fluid system can be approximated by that of a finite-dimensional one. d Claim 5.1. Let the desired trajectory of the wheel be the ramp function, given by θw (t) = (1/2) θ̇steady |t|−

|t − tstop | + tstop . Then the solution of Eq. (4c) satisfies lim v(r, t) = lim Ri φ̇s (t) t→∞ t→∞ Ro − r . Ro − Ri Proof: We consider an analytical solution to the PDE (4c), obtained by the method of separation of variables. Since one of the boundary conditions is non-homogeneous (and in fact time dependent), we perform a change of variables in order to make it homogeneous [9]. The change of variables is given as v(r, t) = w(r, t) + Ri φ̇s (t) Ro − r Ro − Ri (11) where w(r, t) is transformed the variable. The correction term is taken as a linear interpolation between the two boundary conditions for simplicity. The transformed PDE is given as wt − ν(wrr + w wr − 2 ) = F (r, t) r r (12) where F (r, t) is the driving force of the PDE resulting from the change of variables given by F (r, t) = −Ri φ̈s (t) Ro − r Ro − νRi φ̇s (t) 2 Ro − Ri r (Ro − Ri ) (13) The transformed initial and boundary conditions are given by w(r, 0) = 0 (14a) w(Ri , t) = 0 (14b) w(Ro , t) = 0 (14c) We assume the solution w(r, t) = W (r)T (t), and obtain the homogeneous solution of the spatial ODE by analyzing the corresponding eigenvalue problem. The eigen solution obtained is given by r W (r) = aJ1  r  λ λ r + bY1 r ν ν (15) where J1 and Y1 are the Bessel’s functions of the first and second kind respectively, λ denotes the eigen value, and a and b are arbitrary constants determined by the boundary conditions. Upon plugging in the boundary conditions, and solving for λ, we find that the eigen values are positive due to the fact that roots of Bessel’s functions are real [10]. Thus we have the following solution for v(r, t) 10 v(r, t) = ∞ X r J1 n=1 |  r ! λn λn Ro − r r + kn Y1 r Tn (t) +Ri φ̇s (t) ν ν Ro − Ri {z } (16) transient term where kn and λn are constants obtained by solving the boundary conditions in the eigen value problem, and Tn (t) is the time component of the solution, which is obtained by solving the corresponding initial value problem. The series given in the right hand side of Eq. (16) is considered as a transient term for the PDE (4c), and we show that it is insignificant as t → ∞. Consider the time component of the solution (16) Z t e Tn (t) = 0 R Ro −λn (t−τ ) Ri F (r, τ )Wn (r)dr R Ro Ri Wn (r)2 dr dτ = e −λn t Z t eλn τ (mφ̈s (τ ) + lφ̇s (τ ))dτ 0 where Wn (r) is the nth term of the Fourier series, and m and l are constants obtained after evaluating the integrals in the spatial variable. For our case, plug in θ̇w (t) equal to the desired ramp trajectory

d (θw (t) = (1/2) θ̇steady |t| − |t − tstop | + tstop ) in the momentum equation (8). Upon solving for φ̇s (t), we see that it is exponentially decaying, and therefore Tn (t) → 0 as t → ∞ for all n.1 Thus, we have the result. . In the limit where the transient term is insignificant, the solution (16) is plugged in Eq. (4b) and simplified as follows (Iw + Is )φ̈s + Iw θ̈w = − 2πρνRo Ri2 φ̇s Ro − Ri (17) We would like to emphasize here that we have found an effective damping constant kef f for the infinite-dimensional fluid system (4) that approximates its dynamics to that of a finite dimension case (3). The effective damping constant is given by kef f = 2πρνRo Ri2 Ro − Ri (18) We revisit this effective damping constant and its validity in the appendix, after we introduce a couple of expressions from the boundedness and bifurcation section. 6 Boundedness and bifurcation We now revisit the finite dimensional case (3), and rewrite the corresponding momentum equation as a first order system. The damping-induced momentum equation (8) is given by (Iw + Is )φ̇s + Iw θ̇w (t) + kφs = 0 (19) 1 In fact, the result holds for any wheel trajectory that grows polynomially for finite time before eventually coming to rest. 11 where θ̇w (t) is written as a function of time, since it is an external driving force for the stool dynamics. (Using partial feedback linearization, the wheel can be controlled independently of the stool). Claim 6.1. Suppose k > 0. Then for a fixed speed of the wheel, say θ̇w (t) = vw = constant, we have the following equality lim φs (t) = φ∗s = t→∞ −Iw vw k Proof: Rewritting Eq. (8) as a first order system in φs φ̇s = −k −Iw φs + θ̇w (t) = −αφs − β θ̇w (t) = f (φs , t) Iw + Is Iw + Is (20) where α, β are constants. Since Eq. (20) is a nonautonomous system, we rewrite it in a higher dimensional autonomous form as φ̇s = −αφs − β θ̇w (t) (21a) ṫ = 1 (21b) In the differential equation for the stool, for a fixed speed of the wheel, say θ̇w (t) = θ̇steady = vw = constant, the solution is Z φs (t) = −β t [e−α(t−s) ]vw ds = 0 −βvw [1 − e−αt ] α (22) and as t → ∞, we have φ∗s = lim φs (t) = t→∞ −βvw −Iw vw = α k . The above equality corresponds to the damping-induced boundedness phenomenon, and is especially relevant in Figs. 8 and 9. Now we get to the dynamical systems aspect of our discussion. For the purpose of this discussion we present a few definitions from standard texts on dynamical systems. A fixed point for a system of differential equations Ẋ = F (X) is defined as the set of points that are solutions to the equation F (X) = 0. A fixed point for system (21) does not exist since ṫ 6= 0 for any state (φs , t). We will consider (φ∗s , t∗ ) to be a fixed point if f (φ∗s , t∗ ) = 0. A bifurcation is defined as a change in the qualitative structure of the flow as the parameters are varied. In particular, fixed points can be created or destroyed, or their stability can change [11]. In this particular case, the drastic change in behaviour occurs at k = 0. d Claim 6.2. Let the desired trajectory of the wheel be the ramp function, i.e., θw (t) = (1/2) θ̇steady |t| −

|t − tstop | + tstop . Then • If k = 0, the stool dynamics is neutrally stable • If k > 0, the stool dynamics is asymptotically stable. Thus, a bifurcation occurs when k is varied from 0 to a positive number. 12 Velocity of stool ?_ s (rad=s) 0.1 0 -0.1 -0.2 -0.8 -0.6 -0.4 -0.2 0 Position of stool ?s (rad) Figure 8: Phase portrait of stool dynamics with varying values of damping coefficient k, and steady state wheel velocity θ̇steady = 2. Green: k = 1; red: k = 0.1; blue: k = 0. Green: damping-induced boundedness attained; red: damping induced boundedness is not yet reached; blue: neutral stability, no recovery phenomenon. The final equilibrium point for the k = 0 case and k > 0 are drastically different. Proof: k = 0: From eqn. (21), k = 0 ⇒ φ̇s = −β θ̇w (t) This implies that the stool dynamics has a fixed point only when the wheel is brought to a rest; the stool is neutrally stable at any constant value φ∗s . k > 0: The fixed point φ∗s for the stool dynamics is governed by φ∗s = −Iw θ̇w (t∗ ) k (23) For a fixed point to occur, there are two possibilities: The wheel is at a constant speed vw ⇒ φ∗s = −Iw vw /k or The wheel is at rest ⇒ φ∗s = 0 Clearly, if the wheel is braked (zero speed), the stool must go back to its initial state (φ∗s = 0) governed by φ̇s = −αφs (24) This is the damping-induced self recovery equilibrium state. As seen from the above analysis, the system undergoes a change in its dynamical system behaviour, when k is switched from 0 to a positive number. Eq. (23), rewritten as φ∗s k = −Iw θ̇w (t∗ ) (25) is the equation of a hyperbola in the variables φ∗s and k, and as k → 0, φ∗s → ∞. For k > 0 (no matter how small it is), the recovery phenomenon occurs (implying that the stool must return to its initial position regardless how many rotations it has completed). But for k = 0, the stool is neutrally stable, implying a change in the stability behaviour. (Note that as φs is 2π−periodic, one should consider 13 Velocity of stool ?_ s (rad=s) 0.3 0.2 0.1 0 -0.1 -0.2 -0.15 -0.1 -0.05 0 Position of stool ?s (rad) Figure 9: Phase portrait of stool dynamics with varying values of steady state wheel velocity θ̇steady . Innermost to outermost loop: θ̇steady = 1, 2, 3 (k = 1 for all cases). φs mod 2π to obtain the actual position of the stool. However, not doing so, gives us the number of rotations the stool completes. For e.g., if φ∗s = 8π, then the damping-induced bound is attained after four rotations) . 7 Oscillations and energy description Oscillations of the unactuated variable have been reported in a few experiments on the damping induced self-recovery phenomenon [4, 5, 6, 7]. In our opinion, the cause of these oscillations has not been adequately identified in previous works. We believe that a seemingly trivial, but important source of these oscillations, is the nature of the control law used on the wheel. It is observed that if oscillations are induced in the wheel via the control law, then the stool also mimics these oscillations, with a slight time lag due to the damping. This type of oscillation can be produced in both, the finite-dimensional as well as the infinite-dimensional system. Refer to Figs. 10 and 11 for an example via simulations. We would like to emphasize that oscillations of this kind are a direct consequence of the actuation of the wheel. It is likely that the oscillations in [4, 7] are of this kind. However, the oscillations in Ref. [5] occur even though the wheel is stationary with respect to the stool; in fact, at the extreme positions of the oscillation, it appears as though the entire system (stool and wheel) begins moving again from a complete state of rest. That is, it appears that there is some kind of mechanical ‘spring-like’ energy being stored at these extreme positions. These oscillations are clearly not of the type mentioned above, as they are not a direct result of the actuation of the wheel. Additionally, the observed oscillations in the stool are of high amplitude whereas those in the wheel, if any, appear to be of small amplitude. This is contradictory to what is observed in the previous type of oscillations (seen in Figs. 10 and 11) as the inertia of the stool-wheel system is greater than the inerita of the wheel only. A simple order of magnitude analysis shows that a more complex model is probably 14 Position of wheel 3w (rad) 500 400 300 200 100 0 0 5 10 15 20 Time t (s) Figure 10: Trajectory of the wheel (solid blue line) with control parameters c0 = 1 and c1 = 3. The red dashed line denotes the desired trajectory, and the overshoot of the wheel may be seen at t = 2s. Position of stool ?s (rad) 0.02 0 -0.02 -0.04 -0.06 -0.08 0 5 10 15 20 Time t (s) Figure 11: Trajectory of the stool (solid blue line) under the influence of the wheel trajectory shown in Fig. 10. The black dashed line denotes the zero position of the stool, and the overshoot of the stool may be seen after it recovers (Note the slight delay when compared to the overshoot of the wheel; this is due to the time constant of the system). required to account for this phenomenon. In order to demonstrate this, we first consider the energy dynamics of the system. Energy description for the finite-dimensional system: In order to derive the energy equations of the finite dimensional system, (3b) is multiplied by φ̇s , and (3a) with θ̇w , to yield (Iw + Is )φ̈s φ̇s + Iw θ̈w φ̇s = −k φ̇s φ̇s (26a) Iw φ̈s θ̇w + Iw θ̈w θ̇w = uθ̇w (26b) Equations (26a) and (26b) are integrated, and then summed up, resulting in the following energy balance (in the ground frame) Z t Z uθ̇w dt = K.E.(t) + 0 0 15 t k φ̇2s dt (27) where K.E.(t) is the total kinetic energy of the system. The second term on the right hand side of represents the cumulative energy lost due to damping losses up to time t. The energy lost in damping is precisely equal to the work done by the damping force, that is Z φs (t) L.E.(t) = Z k φ̇s dφs = 0 t k φ̇2s dt. (28) 0 The term on the left hand side of Eq. (27) represents the total energy pumped into the system by the motor, up to time t. The power input into the system by the motor is the sum of power imparted to the wheel by the actuation force u(t), and the power imparted to the stool by the reaction force. The cumulative input energy (denoted by I.E.(t)) is given by Z t Z t Z (29) 0 0 0 t uθ̇w dt (−u)φ̇s dt = u(θ̇w + φ̇s ) dt + I.E.(t) = We can simplify the energy balance (27) as I.E.(t) = K.E.(t) + L.E.(t) (30) Energy description for the infinite-dimensional fluid system: The above energy equation can be extended for the infinite dimensional case as follows. 1 I.E.(t) = K.E.(t) + ρ 2 Z 1 v dV + ρν 2 2 Z tZ  0 ∂vi ∂vk + ∂xk ∂xi 2 dV dt (31) where the second term on the right hand side is the kinetic energy of the fluid, while the third term is the energy lost due to viscous damping [8]. The velocity of the fluid in cartesian coordinates is given by v = −v sinθ î + v cosθ ĵ, where v is the tangential velocity of the fluid as discussed previously (4). We consider the system at a state when the wheel is no longer being actuated and is in a state of rest with respect to the stool. The system can be found in such a state at the extreme position of an oscillation that occurs after the wheel has been brought to rest. This implies that I.E(t) in (31) is zero and hence, the kinetic energy of the stool can only be exchanged with that of the fluid, or be lost due to damping. We analyze this state with the parameters from table 1 for the fluid-stool-wheel system. At the extremum of the oscillation, the stool-wheel system is momentarily at rest, and as is clear from (31), must be imparted energy from the fluid to start rotating again. The maximum energy that the stool-wheel system could possibly regain is therefore equal to the kinetic energy of the fluid at this extreme position. This leads to an expression (32) for the minimum fluid velocity (averaged over the thickness of the annulus) in terms of system constants and the velocity of the stool.  i 2 1h  1 (Iw + Is )φ̇2s = ρπ (Ri + t)2 − (Ri )2 h vavg 2 2 (32) where t and h are the thickness and height of the bearing respectively, while vavg is the velocity of the fluid averaged over the thickness of the annulus. Even for a stool velocity as low as 1 rpm, the required average fluid velocity is around 11 rpm, which is quite high and unrealistic. Hence, it is likely that a 16 different model is required to address such behaviour. Name Moment of Inertia of Wheel-Stool System Inner diameter of bearing Thickness of bearing Height of bearing Density of Fluid Value 1 0.1 0.01 0.01 1000 Units kg/m2 m m m kg/m3 Table 1: Parameter values of fluid-wheel-stool system that are representative of experiment in [5] 8 Conclusion In this paper, we present certain aspects of the damping-induced self recovery phenomenon that have not been investigated so far in the existing literature. We present a technique to reduce the infinitedimensional fluid model to the better understood finite dimensional case, by deriving a formula for an effective damping constant. We show that a bifurcation takes place at k = 0, and upon varying the value of k from zero to a positive number, the stability of the stool changes from neutral to asymptotically stable. We also derive an expression for the angle at which boundedness occurs to validate the approximation of the fluid system (by comparing with numerical experiments). Finally, we present an energy description of the system to give an intuitive understanding of the energy dynamics, and to point out that further experimental and theoretical investigation is necessary to explain the peculiar oscillations found in experiments [5, 6]. References [1] Dong Eui Chang and Soo Jeon. Damping-induced self recovery phenomenon in mechanical systems with an unactuated cyclic variable. Journal of Dynamic Systems, Measurement, and Control, 135(2):021011, 2013. [2] Dong Eui Chang and Soo Jeon. On the damping-induced self-recovery phenomenon in mechanical systems with several unactuated cyclic variables. Journal of Nonlinear Science, 23(6):1023–1038, 2013. [3] Dong Eui Chang and Soo Jeon. On the self-recovery phenomenon in the process of diffusion. arXiv preprint arXiv:1305.6658, 2013. [4] Dong Eui Chang and Soo Jeon. On the self-recovery phenomenon for a cylindrical rigid body rotating in an incompressible viscous fluid. Journal of Dynamic Systems, Measurement, and Control, 137(2):021005, 2015. [5] Dong Eui Chang and Soo Jeon. Video of an experiment that demonstrates the self-recovery phenomenon in the bicycle wheel and rotating stool system. https://youtu.be/og5h4QoqIFs. 17 [6] Andy bikes, Ruina. arrows, Dynamic and walking muscle-smarts: 2010: Cats, astronauts, Stability, translation, and trucks, rotation. http://techtv.mit.edu/collections/locomotion:1216/videos/8007-dynamic-walking-2010-andyruina-cats-astronauts-trucks-bikes-arrows-and-muscle-smarts-stability-trans, 2010. [7] Dong Eui Chang and Soo Jeon. Video of an experiment that demonstrates the self-recovery phenomenon in the vessel and fluid system. https://youtu.be/26qGQccK4Rc. [8] L. D. Landau and E. M. Lifshitz. Fluid Mechanics, volume 6. 1959. [9] Stanley J Farlow. Partial differential equations for scientists and engineers. Courier Corporation, 1993. [10] Milton Abramowitz and Irene A Stegun. Handbook of mathematical functions: with formulas, graphs, and mathematical tables, volume 55. Courier Corporation, 1964. [11] S.H. Strogatz. Nonlinear Dynamics and Chaos: With Applications to Physics, Biology, Chemistry, and Engineering. Advanced book program. Avalon Publishing, 1994. Appendix A Validation of effective damping constant With the computed, effective damping constant kef f (18) of the infinite-dimensional fluid system (4), we find the angle at which boundedness is attained, given by φ∗s = −Iw vw (Ro − Ri ) 2πρνRo Ri2 (33) We now present some numerical experiments by solving the PDE-system (4) using the method of lines. Equation (18) is then verified by comparing the numerical values obtained using formula (33) and the numerical solutions of the original PDE. It is observed that the error tends to zero as Ro /Ri approaches 1, independent of the value of Ri . The material constants used for the simulation, and for the theoretical calculation are tabulated below 18 Ri (cm) Ro (cm) 13.5 13.5 13.5 13.5 13.5 13.5 13.5 13.5 13.5 13.51 13.68 13.75 14 14.5 15 15.5 20 27 27 27 27 27 27 27 27 27 27 27.02 27.36 27.5 28 29 30 31 40 54 0.07 1.33 1.85 3.70 7.41 11.11 14.81 48.15 100 With the PDE Angle (rad) 6.16 108.7 149.8 291.7 553.8 790.3 1004 2265 3123 With kef f Angle (rad) 6.16 109.5 151.3 297.1 573.7 831.9 1073 2704 4160 0.07 1.33 1.85 3.70 7.41 11.11 14.81 48.15 100 1.54 27.19 37.47 72.96 138.5 197.6 251.1 566.3 780.8 1.54 27.37 37.81 74.28 143.4 208 268.4 675.9 1039.9 (Ro −Ri ) Ri × 100 % Error 0.00 0.74 1.00 1.85 3.59 5.26 6.87 19.38 33.21 0.00 0.66 0.91 1.81 3.54 5.26 6.89 19.35 33.18 Table 2: A comparison of the angles at which boundedness occurs, i.e., between the numerical simulations and the theoretically derived formula (33), for different values of (Ro − Ri )/Ri . Name Moment of Inertia of Wheel Moment of Inertia of Stool Kinematic Viscosity of Fluid Density of Fluid Steady State Velocity of Wheel Proportional Control Parameter Derivative Control Parameter Value 6 × 10−3 1.96 1.17 × 10−6 1.0147 × 103 60π 1 100 Units kg/m2 kg/m2 m2 /s kg/m3 rad/s - Table 3: Parameter constants used for the simulations. 19 

3cs.SY

Rate-Splitting for Multi-Antenna Non-Orthogonal Unicast and Multicast Transmission Yijie Mao∗ , Bruno Clerckx† and Victor O.K. Li∗ ∗ arXiv:1802.05567v2 [cs.IT] 16 Feb 2018 The University of Hong Kong, Hong Kong, China † Imperial College London, United Kingdom ∗ Email: {maoyijie, vli}@eee.hku.hk, † b.clerckx@imperial.ac.uk Abstract—In a superimposed unicast and multicast transmission system, one layer of Successive Interference Cancellation (SIC) is required at each receiver to remove the multicast stream before decoding the unicast stream. In this paper, we show that a linearly-precoded Rate-Splitting (RS) strategy at the transmitter can efficiently exploit this existing SIC receiver architecture. By splitting the unicast message into common and private parts and encoding the common parts along with the multicast message into a super-common stream decoded by all users, the SIC is used for the dual purpose of separating the unicast and multicast streams as well as better managing the multi-user interference between the unicast streams. The precoders are designed with the objective of maximizing the Weighted Sum Rate (WSR) of the unicast messages subject to a Quality of Service (QoS) requirement of the multicast message and a sum power constraint. Numerical results show that RS outperforms existing Multi-User Linear-Precoding (MU–LP) and power-domain Non-Orthogonal Multiple Access (NOMA) in a wide range of user deployments (with a diversity of channel directions and channel strengths). Moreover, since one layer of SIC is required to separate the unicast and multicast streams, the performance gain of RS comes without any increase in the receiver complexity compared with MU–LP. Hence, in such non-orthogonal unicast and multicast transmissions, RS provides rate and QoS enhancements at no extra cost for the receivers. Index Terms—non-orthogonal multicast and unicast transmission, rate-splitting, rate region, WMMSE algorithm I. I NTRODUCTION The scarcity of radio resources and the heterogeneity of wireless applications in 5G and beyond motivate recent research on the non-orthogonal unicast and multicast transmission [1]– [4], which is based on Layered Division Multiplexing (LDM) in the literature of digital television systems [5]. The unicast and multicast messages are precoded and superimposed at the transmitter and then broadcast to the receivers in the same timefrequency resources. Each receiver decodes and removes the multicast message using Successive Interference Cancellation (SIC) before decoding its intended unicast message. Joint unicast and multicast beamforming of multi-cell cooperative transmission has been studied, e.g., for transmit power minimization [1], [2], Weighted Sum Rate (WSR) maximization [3] and energy efficiency maximization [4]. However, all of the above works consider the use of Multi-User Linear Precoding (MU–LP) beamforming for the unicast messages. Once the common message is successfully decoded and subtracted from the received signal, each receiver fully decodes its intended unicast message by treating the interference as noise. However, MU–LP is sensitive to the user channel orthogonality and strengths. Another method is to apply power-domain Nonorthogonal Multiple Access (NOMA) to decode the unicast This work is partially supported by the U.K. Engineering and Physical Sciences Research Council (EPSRC) under grant EP/N015312/1. messages at the cost of more layers of SICs. Power-domain NOMA relies on Superposition Coding (SC) at the transmitter and SIC at the receivers [6]. It is denoted in short as SC–SIC in this work. By using SC–SIC, some users are forced to fully decode and cancel interference created by other users. Such method is only suitable when the user channels are (semi-) aligned and exhibit a large disparity of strengths. In contrast to MU–LP and SC–SIC, linearly-precoded RateSplitting (RS) is an emerging multi-user multi-antenna transmission strategy where each unicast message is split into a common part and a private part at the transmitter [7]. The common part is required to be decoded by all the receivers and removed from the received signal using SIC before each receiver decodes its intended private part by treating the interference from other users as noise. RS can be viewed mathematically as a non-orthogonal unicast and multicast transmission strategy given the superimposed transmission of common and private messages. Hence, RS was termed joint multicasting and broadcasting in [8]. However, common message in RS has an objective different from that of a conventional multicast message. The multicast message is intended and decoded by all the users while the common message of RS is decoded by all users but is intended to a subset of users. Its presence enables the decoding of part of the multi-user interference and treating the remaining part of the interference as noise. In this work, motivated by the benefits of RS in multiantenna Broadcast Channels (BC) [7], [9], [10], we propose the use of RS in non-orthogonal unicast and multicast transmissions. To the best of our knowledge, this is the first work that applies RS to non-orthogonal unicast and multicast transmissions. In such a setup, we split the unicast messages into common and private parts and encode the common parts along with the multicast message into a super-common stream decoded by all users. A single layer of SIC in RS is then used for the dual purpose of separating the unicast and multicast streams as well as better managing the multi-user interference of the unicast streams. We design the precoders by formulating the WSR maximization problem of the unicast messages with a Quality of Service (QoS) requirement of the multicast message and a sum power constraint. The problem is transformed into an equivalent Weighted Minimum Mean Square Error (WMMSE) problem and solved using an Alternating Optimization (AO) algorithm. We demonstrate in the numerical results that the rate region of RS is always equal to or larger than that of MU– LP and SC–SIC. Importantly, this performance gain comes at no additional cost for the receivers since one layer of SIC is required to separate unicast and multicast streams in the conventional MU–LP strategy. In other words, RS makes a better use of the existing SIC architecture. The rest of the paper is organized as follows. In Section II, the existing MU–LP beamforming is overviewed. The proposed RS beamforming and the optimization framework are respectively specified in Section III and Section IV. Section V illustrates numerical results and Section VI concludes the paper. Notations: C and E{·} respectively refer to the complex space and the statistical expectation. The boldface uppercase and lowercase letters represent matrices and vectors, respectively. k·k is the Euclidean norm. The superscripts (·)T and (·)H correspond to transpose and conjugate-transpose operators. tr(·) and diag(·) are the trace and diagonal entries. II. E XISTING MU–LP B EAMFORMING In this work, we consider a BS equipped with Nt antennas serving K single-antenna users. The users are indexed by the set K = {1, . . . , K}. In each time frame, the BS wants to transmit a multicast message W0 intended for all users and K unicast messages W1 , . . . , WK intended for different users. The messages W0 , W1 , . . . , WK are independently encoded into data streams s0 , s1 , . . . , sK . The stream vector s = [s0 , s1 , . . . , sK ]T is precoded using the precoder P = [p0 , p1 , . . . , pK ], where p0 ∈ CNt ×1 and pk ∈ CNt ×1 are the respective precoders of the multicast stream s0 and the unicast stream sk , ∀k ∈ K. Assuming that E{ssH } = I, the transmit power is constrained by tr(PPH ) ≤ Pt . The resulting transmit signal x ∈ CNt ×1 is given by X x = Ps = p0 s0 + pk s k . (1) k∈K The signal received at user-k is yk = hH k x + nk , where hk ∈ CNt ×1 is the channel between the BS and user-k. nk ∼ 2 CN (0, σn,k ) is the Additive White Gaussian Noise (AWGN) at user-k. Without loss of generality, we assume the noise variances are equal to one. The transmit SNR is equal to the total power consumption Pt . We assume perfect Channel State Information at the Transmitter (CSIT) and perfect Channel State Information at the Receivers (CSIR). At user sides, each user first decodes the multicast stream by treating the signal of all the unicast streams as interference. The Signal-to-Interference-plus-Noise Ratio (SINR) of the multicast 2 2 P H + 1). stream at user-k is γk,0 = hH k p0 /( j∈K hk pj Once s0 is successfully decoded, its contribution to the original received signal yk is subtracted. After that, user-k decodes its unicast stream sk by treating the unicast streams of other users as noise. The SINR of decoding the unicast stream sk 2 2 P H at user-k is γk = hH + 1). The k pk /( j∈K,j6=k hk pj corresponding achievable rates of s0 and sk at user-k are Rk,0 = log2 (1 + γk,0 ) and Rk = log2 (1 + γk ) . (2) To ensure that s0 is successfully decoded by all users, the achievable rate of s0 shall not exceed R0 = min {R1,0 , . . . , RK,0 } . (3) In this work, we maximize the WSR of the unicast messages while the rate constraint of the multicast message and the power constraint of the BS should be met. For a given weight vector u = [u1 , . . . , uK ], the WSR achieved by the unicast messages Fig. 1: K-user RS assisted multi-antenna non-orthogonal unicast and multicast transmission model in the K-user MU–LP assisted multi-antenna non-orthogonal unicast and multicast is X uk Rk (4a) RMU−LP (u) = max P k∈K s.t. Rk,0 ≥ R0th , ∀k ∈ K (4b) H (4c) tr(PP ) ≤ Pt where R0th is the lower bound on the multicast rate. III. P ROPOSED R ATE - SPLITTING B EAMFORMING In this section, we introduce the use of RS to the system. We highlight the difference between the proposed RS beamforming and the existing MU–LP beamforming. The contents that are not specified remain consistent with Section II. The main difference between MU–LP and RS beamforming lies in the generation of the data streams. Different from MU– LP beamforming where the messages are directly encoded into independent streams, the unicast message Wk intended for user-k is split into a common part Wk,c and a private part Wk,p , ∀k ∈ K. The common parts of the unicast messages W1,c , . . . , WK,c are encoded along with the multicast message W0 into a super-common stream s0 . It is required to be decoded by all users. Note that s0 includes not only the whole multicast message, but parts of the unicast messages intended for different users. The private parts of the unicast streams W1,p , . . . , WK,p are independently encoded into the private streams s1 , . . . , sK . The stream vector s = [s0 , s1 , . . . , sK ]T are linear precoded via the precoder matrix P = [p0 , p1 , . . . , pK ] and broadcast to users. The system model is illustrated in Fig. 1. At user sides, the super-common stream and private streams are decoded using SIC as in MU–LP. The achievable rate of the super-common stream Rk,0 and the private stream Rk , ∀k ∈ K are calculated based on equation (2). To ensure that the supercommon stream s0 is successfully decoded by all users, the achievable super-common rate R0 is calculated by equation (3). R0 is shared by the rate of transmitting the multicast message W0 and the rates of transmitting the common parts of the unicast messages of all users, W1,c , . . . , WK,c . Denote C0 as the portion of R0 transmitting W0 and Ck,0 as the user-k’s portion of R0 transmitting Wk,c , the achievable super-common rate is equal to X C0 + Ck,0 = R0 (5) k∈K Following the RS structure described above, the total achievable rate of the unicast message of user-k is Rk,tot = Ck,0 + Rk . For a given weight vector u, the WSR achieved by the private messages in the K-user RS assisted multi-antenna nonorthogonal unicast and multicast is RRS (u) = max P,c s.t. C0 + X uk Rk,tot (6a) k∈K X Ck,0 ≤ Rk,0 , ∀k ∈ K (6b) k∈K C0 ≥ R0th (6c) Ck,0 ≥ 0, ∀k ∈ K (6d) tr(PPH ) ≤ Pt (6e) where c = [C0 , C1,0 , . . . , CK,0 ] is the common rate vector. We note that problem (6) boils down to problem (4) when no power is allocated to the common messages W1,c , . . . , WK,c . Hence, RS always achieves the same or superior performance to MU–LP. In the unicast-only transmission, RS requires one layer of SIC to decode the common parts of the unicast messages for the purpose of enabling the capability of decoding part of the multi-user interference and treating part of the multiuser interference as noise [7], [9], [10]. In comparison, MU– LP does not require any SIC at each receiver in the unicastonly transmission. However, one-layer of SIC is necessary at each user to decode the multicast stream before decoding the intended unicast stream in the MU–LP assisted joint unicast and multicast transmission and RS still requires one layer of SIC at no extra cost for the receivers. The SIC of RS in the joint unicast and multicast transmission is used for separating the unicast and multicast streams as well as better managing the multi-user interference between the unicast streams. to γk,0 = 1/εMMSE − 1 and γk = 1/εMMSE − 1 based on k,0 k (9). The rates become Rk,0 = − log2 (εMMSE k,0 ) and Rk = − log2 (εMMSE ). The WMSEs are given by k ξk,0 = uk,0 εk,0 − log2 (uk,0 ), ξk = uk εk − log2 (uk ), where uk,0 and uk are weights associated with the MSEs of ∗ MMSE and gk∗ = gkMMSE user-k. The optimum equalizers gk,0 = gk,0 ∂ξk,0 ∂ξk are then derived by solving ∂gk,0 = 0 and ∂g = 0. k MMSE Substituting gk,0 and gkMMSE into (10), we have
MMSE = uk,0 εMMSE − log2 (uk,0 ), ξk,0 gk,0 k,0
(11) MMSE MMSE ξk gk = u k εk − log2 (uk ). The optimum MMSE weights are obtained by solving MMSE MMSE ∂ξk (gk ∂ξk,0 (gk,0 ) ) = 0 and = 0, which are given by ∂uk,0 ∂uk −1 u∗k,0 = uMMSE , (εMMSE , u∗k = uMMSE , (εMMSE )−1 (12) k,0 k,0 ) k k The Rate-WMMSE relationships are finally established by substituting (12) into (11). They are given by MMSE , ξk,0 (7) εk = |gk |2 Tk − 2ℜ{gk hH k pk } + 1, P H 2 2 where Tk,0 , |hH k p0 | + j∈K |hk pj | + 1 and Tk , Tk,0 − 2 H |hk p0 | . The optimum MMSE equalizers are then calculated ∂ε ∂εk by solving ∂gk,0 = 0 and ∂g = 0, which are given by k,0 k −1 −1 MMSE MMSE gk,0 = pH = pH 0 hk Tk,0 , gk k hk T k . (8) Substituting (8) into (7), the MMSEs are −1 εMMSE , min εk,0 = Tk,0 Ik,0 , εMMSE , min εk = Tk−1 Ik , k,0 k gk,0 gk (9) 2 where Ik,0 = Tk and Ik = Tk − |hH p | . The SINRs of k k decoding s0 and sk at user-k can be respectively transformed (13) uk ,gk Based on the Rate-WMMSE relationships in (13), the optimization problem (6) is transformed equivalently into the WMMSE problem given by X min uk ξk,tot (14a) P,x,u,g εk,0 = |gk,0 |2 Tk,0 − 2ℜ{gk,0 hH k p0 } + 1, min ξk,0 = 1 − Rk,0 , uk,0 ,gk,0 ξkMMSE , min ξk = 1 − Rk . IV. O PTIMIZATION F RAMEWORK The WMMSE algorithm to solve the sum rate maximization problem in RS without a multicast message is proposed in [9]. It is extended to solve the problem (4) and (6). We firstly explain the procedure to solve the RS problem (6) and then specify how (4) can be solved correspondingly. User-k decodes s0 and sk sequentially using SICs. The common stream s0 is decoded first. By using the equalizer gk,0 , s0 is estimated as ŝ0 = gk,0 yk . sk is estimated using the equalizer gk as ŝk = gk (yk − hH k p0 s0 ) after s0 is successfully decoded and removed from yk . The Mean Square Errors (MSEs) of the common and private streams are defined as εk,0 , E{|ŝk,0 − sk,0 |2 } and εk , E{|ŝk − sk |2 }, respectively. They are given by (10) s.t. X0 + k∈K X Xk,0 + 1 ≥ ξk,0 , ∀k ∈ K (14b) k∈K X0 ≤ −R0th (14c) Xk,0 ≤ 0, ∀k ∈ K (14d) H tr(PP ) ≤ Pt (14e) where x = [X0 , X1,0 , . . . , XK,0 ] is the transformation of the common rate c. u = [u1,0 , . . . , uK,0 , u1 , . . . , uK ] and g = [g1,0 , . . . , gK,0 , g1 , . . . , gK ] are the weights and equalizers, respectively. ξk,tot = Xk,0 + ξk , ∀k ∈ K. Denote uMMSE and gMMSE as two vectors formed by the corresponding MMSE equalizers and weights. According to the KKT conditions of problem (14), it is easy to show that (uMMSE , gMMSE ) are optimal and unique. We can obtain (uMMSE , gMMSE ) by minimizing (14a) with respect to u and g, respectively. Problem (14) can be transformed to problem (6) based on the Rate-WMMSE relationship (13) and the common rate transformation c = −x. The solution given by (c∗ = −x∗ , P∗ ) meets the KKT optimality conditions of (6) for any point (x∗ , P∗ , u∗ , g∗ ) satisfying the KKT optimality conditions of (14). Therefore, problem (6) and problem (14) are equivalent. The joint optimization of (x, P, u, g) in problem (14) is nonconvex. With fixed (x, P, u), the MMSE equalizer gMMSE is optimal. With fixed (x, P, g), the MMSE weight uMMSE is the optimal weight. With fixed (u, g), the optimization problem (14) is a convex Quadratically Constrained Quadratic Program (QCQP) which can be solved using interior-point methods. 1 2 3 4 5 6 7 Initialize: n ← 0, P[n] , WSR[n] ; repeat n ← n + 1; P[n−1] ← P; u ← uMMSE (Pn−1 ); g ← gMMSE (Pn−1 ); update (x, P) by solving (14) using the updated u, g; until |WSR[n] − WSR[n−1] | ≤ ǫ; V. N UMERICAL R ESULTS In this section, the performance of the proposed RS beamforming is illustrated by comparing with the existing MU–LP and SC–SIC beamforming. We focus on the two-user case since different two-user rate regions are easily compared in a twodimensional figure. The SC–SIC assisted joint unicast and multicast transmission is briefed before we illustrate the results. Different from RS where the multicast message is encoded with the common parts of the unicast messages, the multicast message in SC–SIC is encoded along with the unicast message to be decoded first into a super-common stream s0 . Hence, the super-common rate R0 of SC–SIC is shared by the rate of the multicast message as well as the rate of the unicast message to be decoded first. Following the system model of RS in Section III and the difference between SC–SIC and RS, we can formulate the problem of SC–SIC and solve it by modifying the optimization framework. Note that the receiver complexity of SC–SIC increases with the number of users. More layers of SICs are required at each user to decode the interference from more users. Moreover, SC– SIC is a particular instance of the proposed RS strategy when K = 2 [10]. RS should always achieve the same or superior performance to MU–LP and SC–SIC. We investigate the influence of the multicast rate constraint, channel strength disparity and channel angle between the users on the performance. The BS is equipped with four antennas (Nt = 4) serving two single-antenna users (K = 2). The simulation setting follows the underloaded two-user deployment in [10]. The channels of users are realized as h1 = (b) θ=2π/9 10 10 8 8 R2,tot (bit/s/Hz) R2,tot (bit/s/Hz) (a) θ=π/9 6 4 2 6 4 2 0 0 0 5 10 0 R1,tot (bit/s/Hz) 10 (d) θ=4π/9 10 8 8 R2,tot (bit/s/Hz) 10 6 4 RS SC-SIC MU-LP 2 5 R1,tot (bit/s/Hz) (c) θ=π/3 R2,tot (bit/s/Hz) Hence, the AO algorithm is motivated to solve the problem. In the nth iteration, the equalizers and weights are
calculated by (u, g) = uMMSE (P[n−1] ), gMMSE (P[n−1] ) based on the precoder P[n−1] in the (n − 1)th iteration. (x, P) are then calculated by solving problem (14). (u, g) and (x, P) are iteratively updated until the convergence of the WSR. Algorithm 1 shows the steps of AO, where ǫ is the error tolerance for convergence and WSR[n] is the WSR calculated based on the updated (x, P) in nth iteration. Since WSR[n] is increasing with n and it is bounded above for a given power constraint, the AO algorithm is guaranteed to converge. Note that the global optimality of the solution cannot be guaranteed in general as the problem is non-convex. The initialization of the precoder P will influence the final results. The optimization framework described above are adopted to solve the MU–LP problem (4) by reformulating it into its equivalent WMMSE problem and using the AO algorithm to solve it. Algorithm 1: Alternating Optimization Algorithm 6 4 2 0 0 0 5 R1,tot (bit/s/Hz) 10 0 5 10 R1,tot (bit/s/Hz) Fig. 2: Achievable rate region comparison of different strategies in perfect CSIT, γ = 1, R0th = 0.5 bit/s/Hz  H H [1, 1, 1, 1] , h2 = γ × 1, ejθ , ej2θ , ej3θ . In the following results, γ = 1 and γ = 0.3, which respectively represent equal channel strength and 5 dB channel strength  difference. For each π 4π γ, we consider four different θ, θ ∈ π9 , 2π , , 9 3 9 . When θ is less than π9 , the user channels are sufficiently aligned. When θ is larger than 4π 9 , the user channels are sufficiently orthogonal. The rate region is the set of all achievable points. Its boundary is calculated by varying the weights assigned to users. We follow the weights in [11], where the weight of user-1 is fixed to u1 = 1 for each weight of user-2 in u2 ∈ 10[−3,−1,−0.95,··· ,0.95,1,3] . The precoders of RS, MU– LP and SC–SIC are initialized using the same methods as discussed in [10]. SNR is fixed to 20 dB. Fig. 2 shows the results when user-1 and user-2 have equal channel strengths (γ = 1) and the multicast rate constraint is R0th = 0.5 bit/s/Hz. In each subfigure, the rate region achieved by RS is confirmed to be equal to or larger than that of SC–SIC and MU–LP. RS performs well for any angle between the user channels. As the SC–SIC strategy is motivated by leveraging the channel strength difference of users, it is sensitive to the channel strength disparity. When users have equal channel strengths, SC–SIC has poor performance. RS exhibits a clear rate region improvement over SC–SIC and MU–LP when θ = π9 . The performance of MU–LP is poor when the user channels are closely aligned to each other. MU– LP is sensitive to the channel angle. As θ increases, the gap between the rate regions of RS and MU–LP decreases. When θ = 4π 9 , RS reduces to MU–LP. Fig. 3 shows the results when γ = 1, R0th = 1.5 bit/s/Hz. Comparing the corresponding subfigures of Fig. 2 and Fig. 3, the rate regions of all the strategies increase as the rate threshold of the multicast message decreases. Moreover, the rate region gain of RS over MU–LP and SC–SIC increases as the rate threshold of the multicast message decreases. This is due to the fact the super-common stream can absorb a larger portion of the unicast messages as the rate of the multicast message decreases. RS is able to overcome the limitations of MU–LP and SC–SIC by dynamically determining the level of the multi-user interference to decode and treat as noise. When (b) θ=2π/9 10 8 8 R2,tot (bit/s/Hz) R2,tot (bit/s/Hz) (a) θ=π/9 10 6 4 2 6 4 2 0 0 0 5 10 0 R1,tot (bit/s/Hz) (c) θ=π/3 10 (d) θ=4π/9 10 10 RS SC-SIC MU-LP 8 R2,tot (bit/s/Hz) R2,tot (bit/s/Hz) 5 R1,tot (bit/s/Hz) 6 4 2 8 6 4 2 0 0 0 5 10 0 R1,tot (bit/s/Hz) 5 10 R1,tot (bit/s/Hz) Fig. 3: Achievable rate region comparison of different strategies in perfect CSIT, γ = 1, R0th = 1.5 bit/s/Hz (b) θ=2π/9 6 5 5 R2,tot (bit/s/Hz) R2,tot (bit/s/Hz) (a) θ=π/9 6 4 3 2 1 4 3 2 1 0 0 0 5 10 0 R1,tot (bit/s/Hz) 10 R1,tot (bit/s/Hz) (c) θ=π/3 R EFERENCES (d) θ=4π/9 6 6 5 R2,tot (bit/s/Hz) 5 R2,tot (bit/s/Hz) 5 4 3 2 RS SC-SIC MU-LP 1 4 3 2 1 0 0 0 5 R1,tot (bit/s/Hz) VI. C ONCLUSIONS To conclude, we exploit the benefit of the linearly-precoded RS in the joint unicast and multicast transmission systems. Comparing with the conventional MU–LP assisted unicast and multicast transmission system where one layer of SIC is required at each receiver to remove the multicast stream before decoding the unicast stream, the proposed RS-assisted unicast and multicast transmission system further exploits the merits of the existing one layer of SIC. By utilizing a super-common stream to encapsulate the multicast message and parts of the unicast messages, RS uses one layer of SIC to not only separate the unicast and multicast streams but also dynamically manage the multi-user interference. We show in the numerical results that the performance of MU–LP and SC–SIC is more sensitive to the channel strength disparity and channel angles among users. Thanks to its ability of partially decoding the interference and partially treating the interference as noise, RS softly bridges and outperforms MU–LP and SC–SIC in any user deployments. Moreover, the performance gain of RS increases as the rate threshold of the multicast message decreases. The benefit of RS is obtained without any increase in the receiver complexity compared with MU–LP. Therefore, RS is a more powerful transmission scheme for downlink multi-antenna nonorthogonal unicast and multicast transmission systems. 10 0 5 10 R1,tot (bit/s/Hz) Fig. 4: Achievable rate region comparison of different strategies in perfect CSIT, γ = 0.3, R0th = 0.5 bit/s/Hz the rate threshold of the multicast message decreases, more power is allocated to the unicast stream. RS exhibits further benefits of dynamic interference management. Fig. 4 shows the results when γ = 0.3, R0th = 0.5 bit/s/Hz. Comparing with the corresponding subfigures of Fig. 2, the rate region of SC–SIC is closer to that of RS when there is a 5 dB channel strength difference. However, the rate region gap between RS and SC–SIC increases with θ despite the 5 dB channel strength difference. Comparing with RS, SC–SIC is more sensitive to the angle between the user channels. In Fig. 4(b), SC–SIC and MU–LP outperform each other at one part of the rate region and the rate region of RS is larger than the convex hull of the rate regions of SC–SIC and MU–LP. We can easily draw the conclusion from Fig. 4(b) that RS softly bridges and outperforms MU–LP and SC–SIC. Comparing with MU–LP and SC–SIC, RS is more robust to a wide range of channel gain difference and channel angles among users. RS always outperforms MU–LP and SC–SIC. This performance gain comes at no additional cost for the receivers since one layer of SIC is required to separate unicast and multicast streams in the conventional MU–LP strategy. [1] J. Zhao, O. Simeone, D. Gunduz, and D. Gomez-Barquero, “Nonorthogonal unicast and broadcast transmission via joint beamforming and LDM in cellular networks,” in 2016 IEEE Global Communications Conference (GLOBECOM), Dec 2016, pp. 1–6. [2] Y. F. Liu, C. Lu, M. Tao, and J. Wu, “Joint multicast and unicast beamforming for the MISO downlink interference channel,” in 2017 IEEE 18th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC), July 2017, pp. 1–5. [3] E. Chen, M. Tao, and Y.-F. Liu, “Joint base station clustering and beamforming for non-orthogonal multicast and unicast transmission with backhaul constraints,” arXiv preprint arXiv:1712.01508, 2017. [4] O. Tervo, L.-N. Tran, S. Chatzinotas, M. Juntti, and B. Ottersten, “Energy-efficient joint unicast and multicast beamforming with multiantenna user terminals,” arXiv preprint arXiv:1705.03723, 2017. [5] L. Zhang, W. Li, Y. Wu, X. Wang, S. I. Park, H. M. Kim, J. Y. Lee, P. Angueira, and J. Montalban, “Layered-division-multiplexing: Theory and practice,” IEEE Transactions on Broadcasting, vol. 62, no. 1, pp. 216–232, March 2016. [6] Y. Saito, Y. Kishiyama, A. Benjebbour, T. Nakamura, A. Li, and K. Higuchi, “Non-orthogonal multiple access (NOMA) for cellular future radio access,” in 2013 IEEE 77th Vehicular Technology Conference (VTC Spring), June 2013, pp. 1–5. [7] B. Clerckx, H. Joudeh, C. Hao, M. Dai, and B. Rassouli, “Rate splitting for MIMO wireless networks: A promising PHY-layer strategy for LTE evolution,” IEEE Communications Magazine, vol. 54, no. 5, pp. 98–105, May 2016. [8] H. Joudeh and B. Clerckx, “Sum rate maximization for MU-MISO with partial CSIT using joint multicasting and broadcasting,” in 2015 IEEE International Conference on Communications (ICC), June 2015, pp. 4733–4738. [9] ——, “Sum-rate maximization for linearly precoded downlink multiuser MISO systems with partial CSIT: A rate-splitting approach,” IEEE Transactions on Communications, vol. 64, no. 11, pp. 4847–4861, Nov 2016. [10] Y. Mao, B. Clerckx, and V. O. K. Li, “Rate-splitting multiple access for downlink communication systems: Bridging, generalizing and outperforming SDMA and NOMA,” arXiv preprint arXiv:1710.11018v3, 2017. [11] S. S. Christensen, R. Agarwal, E. D. Carvalho, and J. M. Cioffi, “Weighted sum-rate maximization using weighted MMSE for MIMO-BC beamforming design,” IEEE Transactions on Wireless Communications, vol. 7, no. 12, pp. 4792–4799, Dec 2008. 

7cs.IT

On a measurable analogue of small topological full groups François Le Maître arXiv:1608.07399v1 [math.DS] 26 Aug 2016 August 29, 2016 Abstract We initiate the study of a measurable analogue of small topological full groups that we call L1 full groups. These groups are endowed with a Polish group topology which admits a natural complete right invariant metric. We mostly focus on L1 full groups of measure-preserving Z-actions which are actually a complete invariant of flip conjugacy. We prove that for ergodic actions the closure of the derived group is topologically simple although it can fail to be simple. We also show that the closure of the derived group is connected, and that for measure-preserving free actions of non-amenable groups the closure of the derived group and the L1 full group itself are never amenable. In the case of a measure-preserving ergodic Z-action, the closure of the derived group is shown to be the kernel of the index map. If such an action is moreover by homeomorphism on the Cantor space, we show that the topological full group is dense in the L1 full group. Using Juschenko-Monod and Matui’s results on topological full groups, we conclude that L1 full groups of ergodic Z-actions are amenable as topological groups, and that they are topologically finitely generated if and only if the Z-action has finite entropy. Contents 1 Introduction 1.1 L1 full groups of measure-preserving transformations . . . . . . . . . 1.2 Some topological properties of L1 full groups . . . . . . . . . . . . . . 1.3 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 2 4 7 2 Preliminaries 8 2.1 Basic ergodic theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 2.2 Graphings, equivalence relations and full groups . . . . . . . . . . . . 10 2.3 Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 3 L1 full groups of graphings 14 3.1 Definition and first properties . . . . . . . . . . . . . . . . . . . . . . 14 3.2 The closure of the derived group is topologically generated by involutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 1 3.3 3.4 3.5 Dye’s reconstruction theorem . . . . . . . . . . . . . . . . . . . . . . 22 Closed normal subgroups of the closure of the derived group . . . . . 23 Link with topological full groups . . . . . . . . . . . . . . . . . . . . . 27 4 L1 full groups of Z-actions 4.1 The index map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 A characterization of supports of n-cycles . . . . . . . . . . . . . . . 4.3 The L1 full group is generated by induced transformations along with periodic transformations . . . . . . . . . . . . . . . . . . . . . . . . 4.4 Escape time and non-simplicity results . . . . . . . . . . . . . . . . 4.5 Density of topological full groups, amenability and finiteness of the topological rank . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 Further remarks and questions 5.1 Non-amenability . . . . . . . . . . . 5.2 Absence of a natural closure property 5.3 Lp full groups . . . . . . . . . . . . . 5.4 Questions . . . . . . . . . . . . . . . 1 1.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 . 30 . 32 . 33 . 36 . 39 . . . . 41 41 41 43 44 Introduction L1 full groups of measure-preserving transformations The study of (invertible) measure preserving transformations on standard probability spaces has a long and rich history stemming from its connections with various fields such as statistical mechanics, information theory, Riemannian geometry or number theory. Over the years, various invariants have been introduced in order to classify such transformations, although a result of Foreman, Rudolph and Weiss shows that such a classification cannot “reasonably” be obtained for all measurepreserving transformations [FRW11]. The most well-behaved measure-preserving transformations are probably the compact ones, i.e. those which arise as translation by a topological generator in a compact abelian group (for instance, an irrational rotation is a compact transformation). Indeed, ergodic compact transformations are completely classified up to conjugacy by the countable subgroup of the circle S1 consisting of all the eigenvalues of the associated Koopman unitary (for an irrational rotation this is just the group generated by its angle). To be more precise, a theorem of Halmos and von Neumann states that two such transformations T and T 0 of a standard probability space (X, µ) are conjugate iff spec(T ) = spec(T 0 ) where spec(T ) is the group consisting of all the λ ∈ S1 such that there is a non-null f ∈ L2 (X, µ) satisfying f ◦ T = λf . So an ergodic compact transformation T is entirely described by the countable group spec(T ). The purpose of this paper is to initiate the study of a Polish group [T ]1 called the L1 full group which one can associate to a measure-preserving transformation T . Its main feature resembles that of the spectrum of compact transformations: two 2 ergodic measure-preserving transformations T and T 0 are flip-conjugate1 and if and only if [T ]1 is abstractly (or topologically) isomorphic to [T 0 ]1 . The idea for this group comes from full groups, which were defined by Dye as subgroups of the group Aut(X, µ) of measure-preserving transformations of a standard probability space (X, µ) satisfying a strong closure property. The most basic example of a full group is the full group of a measure-preserving invertible aperiodic transformation T , which consists in all the measure-preserving transformations U such that for almost all x ∈ X, there exists n ∈ Z such that U (x) = T n (x). This full group is denoted by [T ] and every U ∈ [T ] is then completely described by the cocycle map cU : X → Z defined by the equation U (x) = T cU (x) (x). Note that the full group [T ] only depends on the partition of the space into orbits induced by T : it is an invariant of orbit equivalence. A theorem of Dye states that all measure-preserving ergodic transformations are orbit equivalent [Dye59], thus making the associated full group a trivial conjugacy invariant for such transformations. To obtain an interesting invariant, we will place an integrability restriction on the cocycle maps. Definition 1.1. The L1 full group of an aperiodic measure-preserving invertible transformation T on a standard probability space (X, µ) is the group of all U ∈ [T ] such that Z |cU (x)| dµ(x) < +∞. X The L1 full group is endowed with a natural metric given by the L1 metric when viewing elements of [T ]1 as cocycle maps X → Z. Such a metric is actually complete, separable and right-invariant (see Proposition 3.4) so that L1 full groups are cli2 Polish groups. Theorem 1.2 (see Thm. 4.2). Let T and T 0 be two measure-preserving ergodic transformation of a standard probability space (X, µ). Then the following are equivalent: (1) T and T 0 are flip-conjugate; (2) the groups [T ]1 and [T 0 ]1 are abstractly isomorphic; (3) the groups [T ]1 and [T 0 ]1 are topologically isomorphic. There are two results behind this theorem: the first is a reconstruction result à la Dye which says that any isomorphisms between L1 full groups must be a conjugacy by a measure-preserving transformation. The second is Belinskaya’s theorem [Bel68, Cor. 3.7], which states that if two ergodic transformations share the same L1 full group, then they must be flip-conjugate. Two measure-preserving transformation T and T 0 are flip-conjugate if there is a measurepreserving transformation S such that T 0 = ST S −1 or T 0 = ST −1 S −1 . 2 A Polish group is cli if it admits a compatible left- (or equivalently right-) invariant metric 1 3 There are natural and well-studied analogues of full groups and L1 full groups in the context of Cantor dynamics. Given a homeomorphism T of the Cantor space C, its (topological) full group [T ]t is the group of all homeomorphisms U of the Cantor space such that for all x ∈ C there is n ∈ N such that U (x) = T n (x). This full group turns out to completely determine the partition of the Cantor space into orbits up to a homeomorphisms. Such partitions are entirely described up to homeomorphism in terms of the set of invariant measure: by a theorem of Giordano, Putnam and Skau [GPS95], two minimal homeomorphisms T and T 0 of the Cantor space are orbit equivalent if and only if there is a homeomorphism S such that S(MT ) = MT 0 where MT denotes the set of T -invariant Borel probability measures. As in the measurable case, an element of the topological full group of an aperiodic homeomorphism T is completely described by the associated cocycle map cU : C → Z defined by U (x) = T cU (x) (x). The analogue of L1 full groups is sometimes called the small topological full group and is also defined by putting a restriction on cocycles. Namely, given a homeomorphism T of the Cantor space C, its small (topological) full group [T ]c is the group of homeomorphisms U of the Cantor space whose cocycle map cU is actually continuous. A result of Giordano-Putnam-Skau in complete analogy with Theorem 1.2 states that when T is minimal, the flip conjugacy class of T is completely determined by its small topological full group [T ]c (see [GPS99]). In the end, L1 full groups can be seen as the missing piece in the following picture. Topological setup Measurable setup Orbit equivalence [T ]t = {U ∈ Homeo(C) : ∀x, U (x) = T cU (x) } [T ] = {U ∈ Aut(X, µ) : ∀x, U (x) = T cU (x) } Flip conjugacy [T ]c = {U ∈ [T ]t : cU is continuous} [T ]1 = {U ∈ [T ] : cU is integrable} 1.2 Some topological properties of L1 full groups Theorem 1.2 implies that every ergodic theoretic property of a measure-preserving transformation T should be reflected algebraically in the group [T ]1 . However since the latter is an uncountable group, it is quite hard to find a purely group-theoretic property allowing us to distinguish two L1 full groups. But as we already pointed out, we also have at our disposal a Polish group topology, which provides richer tools to study [T ]1 . Before we present our results, let us mention a fundamental tool for the study of L1 full groups and topological full groups. Given an aperiodic measure-preserving transformation, one defines the index 4 map I : [T ]1 → R by putting Z cU (x)dµ(x). I(U ) = X If T is a minimal homeomorphism, we define the index map the same way picking an invariant measure µ so as to view T as a measure-preserving transformation. The resulting index map can be shown not to depend on the chosen invariant measure [GPS99]. We can now give various descriptions of the closure of the derived group of the 1 L full group of an ergodic measure-preserving transformation. Theorem 1.3 (see Cor. 4.20). Let T ∈ Aut(X, µ) be ergodic. Then D([T ]1 ) is equal to the following four groups: • the kernel of the index map, • the connected component of the identity, • the group generated by periodic elements and • the group topologically generated by involutions. Since the index map actually takes values into Z we conclude from the first item of the above theorem that the (topological) abelianization of the L1 full group of an ergodic transformation is always equal to Z. In the context of topological full groups of minimal homeomorphisms, it is also true that the index map takes values into Z, but its kernel is often much larger than the derived group. Nevertheless, Matui succesfully computed the abelianization of [T ]c in K-theoretic terms and gave examples where this abelianisation is not even finitely generated (see [Mat06]). So the situation for L1 full groups is much simpler with respect to the (topological) abelianization. An important feature of the derived group of the topological full group of a minimal homeomorphism is that it is a simple group. The following result is a natural analogue. Theorem 1.4. Let T be a measure-preserving transformation. Then T is ergodic if and only if D([T ]1 ) is topologically simple. This result is actually true in a greater generality for L1 full groups of graphings as defined by Levitt [Lev95], see section 3.4. We also prove that D([T ]1 ) is never simple using the fact that there are many Borel sets whose escape time is not integrable (see section 4.4). Let us now seek finer invariants. The topological rank of a separable topological group is the minimal number of elements needed to generate a dense subgroup. The main result of this paper is that the topological rank of the L1 full group of an ergodic measure-preserving transformation T is related to the entropy of T . Theorem 1.5. Let T ∈ Aut(X, µ) be ergodic. Then the following are equivalent 5 (1) the L1 full group [T ]1 is topologically finitely generated; (2) the closed subgroup D([T ])1 is topologically finitely generated; (3) the transformation T has finite entropy. A crucial tool in the previous theorem is Matui’s characterization of finitegeneratedness of topological full groups [Mat06, Thm. 5.4]. He proved that the derived group of the topological full group of a minimal homeomorphism T is finitely generated if and only if T is a subshift. He also characterised finite-generatedness of the whole topological full group of T in terms of a strictly more restrictive condition. In our context the situation is again simpler: Theorem 1.5 shows that for an ergodic measure-preserving transformation T , the L1 full group of T is topologically finitely generated if and only if the closure of its derived group is. In order to use Matui’s aforementioned results so as to prove Theorem 1.5, we establish the following important connection between L1 full groups and topological full groups. Theorem 1.6 (see Thm. 4.28). Let T be a minimal homeomorphism of the Cantor space and let µ be an invariant measure. Then D([T ]c ) is dense in D([T ]1 ) and [T ]c is dense in [T ]1 . Our proof of Theorem 1.5 also relies on a deep theorem of Krieger which characterizes finite entropy transformations as those which can be realised as uniquely ergodic minimal subshifts [Kri72]. It would be nice to have a purely ergodic-theoretic proof of Theorem 1.5. Theorem 1.6 has another interesting consequence. Recall that a topological group is amenable if whenever it acts continuously on a compact space, the action admits an invariant Borel probability measure. Juschenko and Monod have proven that the topological full group of any minimal homeomorphism of the Cantor space is amenable [JM13]. Using the Jewett-Krieger theorem we obtained the following result. Theorem 1.7. Let T be an ergodic measure-preserving transformation. Then the L1 full group of T is amenable. Note that since [T ]1 surjects onto Z via the index map, it is not extremely amenable. However it would be intersting to understand wether the kernel of the index map is extremely amenable. Let us now end extend these kind of questions to the setup of measure-preserving actions of finitely generated groups. Definition 1.8. Let Γ be a finitely generated group acting in a measure-preserving manner on a standard probability space (X, µ). The full group of this action is denoted by [Γ] and defined by [Γ] = {T ∈ Aut(X, µ) : ∀x ∈ X, T (x) ∈ Γ · x}. Let S be a finite generating set for Γ. Then every Γ-orbit is naturally equipped with a Schreier graph metric dS , and we let the L1 full group of the Γ-action be the group [Γ]1 of all T ∈ [Γ] such that the map x 7→ dS (x, T (x)) is integrable. 6 The definition of this group does not depend on the choice of a finite generating set, and the L1 full group [Γ]1 has a natural Polish group topology. The following result is a counterpart to Theorem 1.7. Theorem 1.9 (see Thm. 5.1). Let Γ be a finitely generated non amenable group acting freely in a measure-preserving manner on (X, µ). Then neither [Γ]1 nor D([Γ]1 ) are amenable. It would be interesting to understand which amenable finitely generated groups admit measure-preserving free actions with an amenable L1 full group. It is also natural to seek for generalizations of our results relating topological rank and entropy in this context. Finally, we should stress out that L1 full groups are not invariants of L1 orbit equivalence but only of L∞ orbit equivalence (see Corollary 5.3). I do not know wether they are complete invariants of L∞ orbit equivalence: if [Γ]1 = [Λ]1 , does it follow that Γ and Λ are L∞ orbit equivalent? 1.3 Outline This paper is organised as follows. Section 2 begins with some general facts from ergodic theory. We then recall the definition of graphings, which provide a natural setup for L1 full groups. The section ends with a few basic facts from entropy theory. In section 3 we give the definition of the L1 full group of a graphing and establish some basic properties such as stability under taking induced transformations. We show that the closure of their derived group is topologically generated by involutions and hence connected. We then establish that L1 full groups of ergodic graphings satisfy a reconstruction theorem: any isomorphism between them arises as the conjugacy by a measure preserving transformation. We also use involutions to prove that the only closed normal subgroups of the closure of their derived group are the obvious ones when the graphing is aperiodic (Theorem 3.23). From this we conclude that the graphing is ergodic if and only if the closure of the derived group of its L1 full group is topologically simple. Finally, we relate topological full groups to L1 full groups in terms of density. Section 4 contains the main results on L1 full groups of Z-actions. We first prove Theorem 1.2 and then continue by defining the index map and studying its behaviour with respect to induced transformations. We then study involutions more closely by characterizing the Borel sets which arise as supports of involutions (and more generally those which arise as supports of n-cycles). After that, we show that the L1 full group of T ∈ Aut(X, µ) is generated by transformation induced by T along with periodic transformations, using the work of Belinskaya. We then show that D([T ]1 ) is never simple. Finally, we prove Theorem 1.7 and Theorem 1.5, using the fact that the topological full group is dense in the L1 full group. In section 5 we first prove Theorem 1.9. We then show that L1 full groups are not stable under “L1 cutting and pasting”. After that we make a few observations about Lp full groups which are defined analogously and we conclude with some more questions. 7 2 2.1 Preliminaries Basic ergodic theory Let (X, µ) be a standard probability space, i.e. a standard Borel space equipped with a Borel non-atomic probability measure µ. All such spaces are isomorphic to the interval [0, 1] equipped with the Lebesgue measure, in particular they admit a Borel linear order. We will always work up to measure zero, e.g. for two Borel sets A and B if we write A ⊆ B we mean µ(B \ A) = 0. For us a partition of a Borel set A ⊆ X will be a countable family of disjoint Borel subsets of A whose reunion is equal to A up to measure zero. Note that some of these sets may have measure zero or even be empty. A Borel bijection T : X → X is called a measure-preserving transformation of (X, µ) if for all Borel A ⊆ X, one has µ(A) = µ(T −1 (A)). We denote by Aut(X, µ) the group of all measure-preserving transformations, two such transformations being identified if they coincide up to measure zero. An important fact to keep in mind is that every Borel bijection between full measure Borel subsets of X which is measurepreserving can be seen as an element of Aut(X, µ): there is a measure-preserving transformation T 0 of (X, µ) such that T (x) = T 0 (x) for almost all x ∈ X. Definition 2.1. A measure-preserving transformation T is ergodic if every Borel set which is a reunion of T -orbits has measure 0 or 1. It is aperiodic if all its orbits are infinite. A fundamental construction for us will be that of the induced transformation: given T ∈ Aut(X, µ) and A ⊆ X of positive measure, the Poincaré recurrence theorem ensures us that for almost all x ∈ A there are infinitely many n ∈ N∗ such that T n (x). We let nA,T (x) denote the smallest such n, also called the return time to A. When the transformation T we consider is clear from the context, we will denote the return time simply by nA . The transformation TA ∈ Aut(X, µ) induced by T on A by for all x ∈ X,  n T A,T (x) if x ∈ A TA (x) = x else. It can easily be checked that TA is indeed a measure-preserving transformation, and that moreover for all x ∈ A the TA -orbit of x is the intersection of the T -orbit of x with A. Also note that (TA )−1 = (T −1 )A The following lemma is a direct consequence of the fact that the orbits of the restriction of TA to A are precisely the intersections of the T -orbits with A. Lemma 2.2. Let S ∈ Aut(X, µ) be a measure-preserving transformation and A ⊆ X intersects every S-orbit, then S is aperiodic if and only if SA is aperiodic when restricted to A, and S is ergodic if and only if SA is ergodic when restricted to A. Definition 2.3. An element T ∈ Aut(X, µ) is periodic if all its orbits are finite. 8 If T is periodic, the Borel set of minimums of T -orbits (for a fixed Borel total order on X) intersects every T -orbit at exactly one point: it is a Borel fundamental domain for T . Such fundamental domains actually exist if and only if T is periodic. A nice way to build a new transformation from a periodic one is to compose it with a transformation supported on a fundamental domain. Lemma 2.4. Let T be a periodic measure-preserving transformation, and let A be a fundamental domain of T . Let U be a measure-preserving transformation supported on A. Then (U T )A = U = (T U )A . Proof. Let x ∈ A, let n ∈ N∗ be the least positive integer such that T n (x) = x. Since A is a fundamental domain for T , for all i ∈ {0, ..., n − 1} we have T i (x) 6∈ A. Since U (y) = y for all y 6∈ A, we deduce by induction that (U T )i (x) = T i (x) 6∈ A for all i ∈ {1, ..., n − 1}. In particular (U T )n−1 (x) = T n−1 (x) so (U T )n (x) = U T n (x) = U (x) ∈ A, which proves that (U T )A = U as desired. For the equality (T U )A = U , one can run a similar argument or apply the previously obtained equality to T −1 and U −1 to get (U −1 T −1 )A = U −1 , which by taking inverses yields (T U )A = U . Proposition 2.5. Let T be a periodic measure-preserving transformation, and let A be a fundamental domain of T . Let U be a measure-preserving transformation supported on A. Then the following equivalences hold: (1) The restriction of U to A is aperiodic ⇔ U T is aperiodic ⇔ T U is aperiodic. (2) T restriction of U to A is ergodic ⇔ U T is ergodic ⇔ T U is ergodic. Proof. It follows from the proof of Lemma 2.4 that A intersects every T U orbit and every U T orbit. The conclusion then follows from Lemmas 2.4 and 2.2. Let us observe that one can go the other way around in the above constructions: starting from an aperiodic transformation T and a Borel set A which intersects almost every T -orbit, the map TA−1 T is periodic with A as a Borel fundamental domain and same return time to A as T . This fact is the key idea for the following result of Belinskaya which we won’t use but which is an important step towards to the proof of her Theorem 4.1. Proposition 2.6 (Belinskaya, [Bel68, Thm. 3.6]). Let S and T be two measurepreserving transformations, let A be a Borel subset which intersects every S and every T -orbit. Suppose that SA = TA and that the return times of S and T to A are the same, i.e. nA,S = nA,T . Then S and T are conjugate. We will use the following well-known proposition several times. A proof is provided for completeness. Proposition 2.7. Let T ∈ Aut(X, µ). There is a partition (A1 , A2 , B1 , B2 , B3 ) of supp T such that T (A1 ) = A2 , T (B1 ) = B2 and T (B2 ) = B3 . 9 Proof. Since X is a standard Borel space, there is a countable partition (Xi )i∈N of supp T into Borel sets such that for all i ∈ N, T (Xi ) is disjoint from Xi (see for instance [EG14, Lem. 5.1]). We then define by recurrence an increasing family of sets (Yi )i∈N by letting Y0 = X0 and then for all i ∈ N Yi+1 = Yi t {x ∈ Xi+1 : T (x) 6∈ Yi and T −1 (x) 6∈ Yi }. S Let Y = n∈N Yn , then by construction Y and T (Y ) are disjoint. Moreover, we have supp T = T −1 (Y ) ∪ Y ∪ T (Y ). Indeed, if x ∈ supp T \ Y there is i ∈ N such that x ∈ Xi+1 , and since x 6∈ Yi+1 either T (x) ∈ Yi or T −1 (x) ∈ Yi so that in any case x ∈ T −1 (Y ) ∪ T (Y ). Since supp T is T -invariant, we conclude that supp T = Y ∪ T (Y ) ∪ T 2 (Y ). Now let A1 = {x ∈ Y : T 2 (x) ∈ Y } and B1 = {Y \ A1 }. Then since Y and T (Y ) are disjoint, the sets A1 , A2 := T (A1 ), B1 , B2 := T (B1 ) are disjoint. Let B3 := T 2 (B1 ). By the definition of B1 the set B3 is disjoint from Y . Moreover since T (B1 ) is disjoint from Y we also have T 2 (B1 ) disjoint from T (Y ) so that B3 is actually disjoint from Y t T (Y ). We conclude that A1 , A2 , B1 , B2 , B3 are as desired. Remark. Note that the proof does not use that T is measure-preserving and actually works in the purely Borel context. 2.2 Graphings, equivalence relations and full groups If (X, µ) is a standard probability space, and A, B are Borel subsets of X, a partial isomorphism of (X, µ) of domain A and range B is a Borel bijection f : A → B which is measure-preserving for the measures induced by µ on A and B respectively. We denote by dom f = A its domain, and by rng f = B its range. Given two partial isomorphisms ϕ1 : A → B and ϕ2 : C → D, we define their composition ϕ2 ◦ ϕ1 as the map ϕ−1 1 (B ∩ C) → ϕ2 (B ∩ C) given by ϕ2 ◦ ϕ1 (x) = ϕ1 ϕ2 (x). We also define the inverse ϕ−1 of a partial isomorphism ϕ : A → B by ϕ−1 : B → A and the equation ϕ−1 ◦ ϕ = idA . By definition a graphing is a countable set of partial isomorphisms. Every graphing Φ generates a measure-preserving equivalence relation RΦ , defined to be the smallest equivalence relation containing (x, ϕ(x)) for every ϕ ∈ Φ and x ∈ dom ϕ. Given a graphing Φ, a Φ-word is a composition of finitely many elements of Φ or their inverses. Obviously (x, y) ∈ RΦ iff and only if there exists a Φ-word w such that y = w(x). Given a Borel set A, its Φ-closure is the reunion of all w(A) where w is a Φ-word. Say that Φ is ergodic when the Φ-closure of any non-null Borel set A has full measure. The full group of a measure-preserving equivalence relation R is the group [R] of automorphisms of (X, µ) which induce permutations in the R-classes, that is [R] = {ϕ ∈ Aut(X, µ) : ∀x ∈ X, ϕ(x) R x}. It is a separable group when equipped with the complete metric du defined by du (T, U ) = µ({x ∈ X : T (x) 6= U (x)}. 10 The metric du is called the uniform metric and the topology it induces is called the uniform topology. One also defines the pseudo full group of R, denoted by [[R]], which consists of all partial isomorphisms ϕ such that ϕ(x) R x for all x ∈ dom ϕ. Given a measure-preserving equivalence relation R, we denote by MR the σalgebra of all Borel sets A such that T (A) = A for all T ∈ [R]. Consider the Hilbert space of R-invariant functions, i.e. the set of all f ∈ L2 (X, µ) such that for any T ∈ [R], one has f ◦ T = f a.e. We denote this closed Hilbert space by L2R (X, µ) and remark that it consists of all the MR -measurable elements of L2 (X, µ). By definition R is ergodic if L2R (X, µ) only consists of constant functions. A graphing Φ is ergodic if and only if RΦ is. The orthogonal projection ER from the Hilbert space L2 (X) onto the closed subspace L2R (X) satisfies the following equality, which defines it uniquely: for any f ∈ L2 (X, µ) and g ∈ L2R (X, µ), Z Z fg = ER (f )g. X X ER is called a conditional expectation. When A is a subset of X, its characteristic function is an element of L2 (X, µ), and we call ER (χA ) the R-conditional measure of A, denoted by µR (A) or µΦ (A) when R = RΦ . Often the equivalence relation R will be clear from the context and we will simply call µR (A) the conditional measure of A. Because ER is a contraction for the L∞ norm, µR (A) takes values in [0, 1]. For those who are familiar with the ergodic decomposition, µR (A) is just the function x 7→ µx (A) where (µx )x∈X is the ergodic decomposition of R. Given ϕ ∈ [[R]], we have µR (dom ϕ) = µR (rng ϕ). The following well-known result is a kind of converse (see [LM15, Sec. 2.1]). Proposition 2.8 (Dye). Let R be a measure-preserving equivalence relation. Then if two sets A and B have the same R-conditional measure, there exists ϕ ∈ [[R]] whose domain is A and whose range is B. Note that in the ergodic case, the previous proposition implies that any two sets of the same measure can be mapped to each other by an element of the pseudo full group of R. A measure-preserving equivalence relation is aperiodic if almost all its equivalence classes are infinite. The following proposition is well-known, see [LM15, Sec. 2.1] for a proof. Proposition 2.9 ([Dye59], Maharam’s lemma). A measure-preserving equivalence relation R is aperiodic iff for any A ⊆ X, and for any MR -measurable function f such that 0 6 f 6 µR (A), there exists B ⊆ A such that the R-conditional measure of B equals f . Corollary 2.10. Let R be an aperiodic measure-preserving equivalence relation, and let f : X → [0, 1] be an MR -measurable function. Let A be a subset of X such that for almost all x ∈ A, one has f (x) > 0. Then there is a countable partition (An )n∈N of A such that for all x ∈ X and all n ∈ N, µR (An )(x) 6 f (x). 11 Proof. The proof relies the following basic fact: there is a countable family of MR measurable positive functions fP n : X → [0, 1] such that for all x ∈ X, we have fn (x) 6 f (x) for all n ∈ N and n∈N fn (x) = µR (x). Indeed, if for all x ∈ X we let fn (x) = 0 if µR (A)(x) = 0 and else    f (x) j k if n < bµR (A)(x)c f (x) f (x) − g(x) if n = bµR (A)(x)c fn (x) =   0 else, it is easily checked that such functions satisfy the desired properties. We can now use the previous proposition to build by recurrence a sequence (An ) of disjoint subsets P of A such that for every n ∈P N, µR (An ) = fn F 6 f . Then µR (A) = n∈N µR (An ) hence by integrating µ(A) = n∈N µ(An ) = µ( n∈N An ) so that (An ) is a partition of A meeting all our requirements. Let us also record the following easy consequences of the previous results. Lemma 2.11. Let R be a measure-preserving equivalence relation. Then if two disjoint Borel sets A and B have the same R-conditional measure, there exists an involution U ∈ [R] whose support is equal to A t B such that U (A) = B. Proof. By Proposition 2.8 there exists ϕ ∈ [[R]] such that ϕ(A) = B. Then the involution U ∈ [RΦ ] defined by   ϕ(x) if x ∈ A ϕ−1 (x) if x ∈ B U (x) =  x else clearly works. Lemma 2.12. Let R be a measure-preserving aperiodic equivalence relation and let C be a Borel subset of X. Then there exists an involution U ∈ [R] such that supp U = C. Proof. By Maharam’s Lemma (Lemma 2.9), one can write C = AtB where µR (A) = µR (B) = µR (C)/2. The previous lemma can then readily be applied. Remark. In Section 4.2, we will characterise which sets arise as supports of involutions belonging to the L1 full group of T , where T is an ergodic transformation. 2.3 Entropy The entropy of a countable partition (Ai )I∈I of (X, µ) is the quantity X H((Ai )i∈I ) := − µ(Ai ) ln µ(Ai ) i∈I where we use the convention 0 ln 0 = 0. It represents the average amount of information we get when we choose an element x ∈ X at random and only learn in which piece Ai of the partition x is. The following statement was extracted from the more general Lemma 2.1 in [Aus16]. It shows that cocycle partitions coming from elements of L1 full groups of Z-actions have finite entropy. 12 Lemma 2.13. Let (An )n∈Z be a partition of X such that Then H((An )n∈Z ) < +∞. P n∈Z |n| µ(An ) < +∞. Proof. Some elementary calculus yields the inequality −t ln t 6 mt + e−m−1 for all t, m > 0. In particular for all n ∈ Z∗ we have −µ(An ) ln µ(An ) 6 |n| µ(An ) + e−|n|−1 . P Summing this P over n ∈ Z∗ and applying our hypothesis n∈Z∗ |n| µ(An ) < +∞, we obtain that − n∈Z∗ µ(An ) ln µ(An ) < +∞. So H((An )n∈N ) < +∞ as wanted. Remark. Austin’s result (Lemma 2.1 in [Aus16]) actually yields that whenever Γ is a finitely generated group, the cocycle partitions of elements of the L1 full group of any free Γ-action have finite entropy. Given two partitions (Ai ) and (Bj ) of X, we let (Ai ) ∨ (Bj ) be their join, i.e. the partition (Ai ∩ Bj ). We have the following natural inequality (for a proof see e.g. [Dow11, 1.6.10]). Lemma 2.14. Let (Ai ) and (Bj ) be two partitions of X. Then H((Ai ) ∨ (Bj )) 6 H((Ai )) + H((Bj )). Note that in particular, the join of finitely many partitions with finite entropy has finite entropy. Lemma 2.14 enables one to give the classical definition of the entropy of T ∈ Aut(X, µ). First, given a partition Q, we let ! n−1 _ 1 h(T, P) := lim H T i (P) , n→+∞ n i=0 where the limit exists by subadditivity of the sequence H( 2.14). The entropy of T is h(T ) = sup Wn−1 i=0 T i (P)) (cf. Lemma h(T, P). P partition A partition P is called generating if up to measure zero, the σ-algebra generated by the T -translates of P is equal to the Borel σ-algebra of X. A more convenient way to see this for us is in terms of the measure algebra of (X, µ). Recall that given a standard probability space (X, µ), its measure algebra is denoted by MAlg(X, µ) and consists of all Borel subsets of X, two such sets being identified if their symmetric difference has measure zero. It is naturally endowed with a metric dµ defined by dµ (A, B) = µ(A 4 B) which is complete and separable (see [Hal50, 40.A]). Then a partition P is generating if and only if the algebra generated by the T -translates of P is dense in MAlg(X, µ). 13 Kolmogorov-Sinai’s theorem states that whenever P is a generating partition of finite entropy, one has h(T ) = h(T, P) < +∞ Conversely, Krieger’s finite generator theorem states that every transformation of finite entropy admits a finite generating partition [Kri70] (moreover, there is such a partition whose cardinality k satisfies the inequality eh(T ) 6 k 6 eh(T ) + 1). Krieger then proved that every such transformation is actually conjugate to a minimal uniquely ergodic subshift [Kri72], which will be relevant to us when we connect topological full groups to L1 full groups (see Section 4.5). 3 3.1 L1 full groups of graphings Definition and first properties Let Φ be a graphing, and let RΦ be the measure-preserving equivalence relation it generates. Such a graphing induces a metric dΦ on the orbits of RΦ by putting for all (x, y) ∈ R dΦ (x, y) = min{n ∈ N : ∃ϕ1 , ..., ϕn ∈ Φ±1 such that y = ϕn · · · ϕ1 (x)}. The L1 full group of Φ, denoted by [Φ]1 , is then defined by [Φ]1 := {T ∈ [RΦ ] : x 7→ dΦ (x, T (x)) is integrable}. The triangle inequality implies that [Φ]1 is indeed a group, which we equip with a right-invariant metric d˜Φ defined by Z dΦ (S(x), T (x)). d˜Φ (S, T ) := X Our first example is fundamental since the study of single measure-preserving transformations is a central subject in ergodic theory. Example 3.1. Suppose we are given a measure-preserving aperiodic Z-action, that is, an element T ∈ Aut(X, µ), all whose orbits are infinite. Then Φ = {T } is a graphing, and we denote by [T ]1 the associated L1 full group. Given S ∈ [RT ], for all x ∈ X there is a unique cS (x) ∈ Z such that S(x) = T cS (x) (x). The map x 7→ cS (x) is called the cocycle associated to S and it completely determines S. The partition ({x ∈ X : cS (x) = n})n∈Z is called the cocycle partition associated to S. For all x ∈ X and all S, S 0 ∈ [RT ], we have the cocycle identity cS 0 S (x) = cS 0 (S(x)) + cS (x). The L1 full group [T ]1 := [RΦ ]1 consists of all the S ∈ [RT ] such that the corresponding cocycle cS is integrable. The metric d˜Φ is then given by Z 0 ˜ dΦ (S, S ) = |cS (x) − cS 0 (x)| dµ(x) X This fundamental example will be explored in Section 4. 14 Example 3.2. Suppose Φ is a graphing, all whose connected components are complete graphs. Then dΦ is the discrete metric: for all (x, y) ∈ RΦ ,  0 if x = y, dΦ (x, y) = 1 else. Hence we have [Φ]1 = [RΦ ] and d˜Φ is equal to the uniform metric du . Example 3.3. Let Γ be a finitely generated group and consider a measure-preserving Γ-action on (X, µ). If S is a finite generating set of Γ, it defines a graphing which generates RΓ , and we can consider the associated L1 full group [S]1 . Observe that if S 0 is another finite generating set, then the metrics dS and dS 0 are bilipschitzequivalent on every orbit with a uniform constant. So the metrics d˜S and d˜S 0 are bilipschitz-equivalent and we have [S]1 = [S 0 ]1 . We conclude that we can actually define the L1 full group of the Γ-action by [Γ]1 = [S]1 since it does not depend on the choice of a finite generating set. Note that the L1 full group of the action of a finitely generated group comes naturally equipped with a complete left invariant metric well-defined up to bilipschitz equivalence. Of course Example 3.1 falls into this category with Γ = Z. Moreover, if Γ acts freely, every T ∈ [Γ] is uniquely defined by the associated cocycle cT : X → Γ defined by T (x) = cT (x) · x. Such a cocycle satisfies the cocycle identity cT 0 T (x) = cT 0 (T (x))cT (x). Proposition 3.4. Let Φ be a graphing. The metric d˜Φ is complete and separable. Proof. Let (Tn ) be a Cauchy sequence for d˜Φ . Note that since a Cauchy sequence converges iff it admits a converging subsequence we can freely replace (Tn ) by a subsequence. Clearly for all T, T 0 ∈ [Φ]1 we have du (T, T 0 ) 6 d˜Φ (T, T 0 ), so the sequence (Tn ) is du -Cauchy and admits a du -limit T ∈ [RΦ ]. Up to taking a subsequence, we may assume that Tn → T pointwise and that for all n ∈ N we have d˜Φ (Tn , Tn+1 ) < 21n . We deduce that Z X dΦ (Tn (x), Tn+1 (x))dµ(x) < 2. X n∈N Now for all n ∈ N and all x ∈ X we have dΦ (Tn (x), T (x)) = lim dΦ (Tn (x), Tm (x)) m→+∞ 6 lim m→+∞ 6 X m−1 X dΦ (Ti (x), Ti+1 (x)) i=n dΦ (Ti (x), Ti+1 (x)) i∈N By integrating this inequality for n = 0 we deduce that d˜Φ (T0 , T ) < +∞ so that T ∈ [Φ]1 . Moreover this inequality allows us apply the Lebesgue dominated convergence theorem to the sequence of functions x 7→ dΦ (Tn (x), T (x)) to conclude that dΦ̃ (Tn , T ) → 0. The metric d˜Φ is therefore complete. 15 The separability follows from the line of ideas present in [CLM16, Thm. 3.17]. Namely, let S be a family of involutions such that [ {(x, s(x)) : x ∈ X} = GΦ ∪ {(x, x) : x ∈ X} s∈S and define Γ = hSi equipped with the word metric dS . We consider the space R L1 (X, µ, Γ) of measurable functions taking values in Γ such that X dS (f (x), 1Γ ) < +∞. Such a space is separable for the natural L1 metric, and [Φ]1 is a continuous quotient of a subspace of the metric space L1 (X, µ, Γ) (namely, the subspace of maps f ∈ L1 (X, µ, Γ) such that Tf : x 7→ f (x) · x belongs to Aut(X, µ)), hence separable. Remark. Because d˜Φ is right-invariant, the previous Proposition shows that [Φ]1 is a cli Polish group, which means it admits a complete left (or equivalently right)invariant metric compatible with its topology. Examples of cli Polish groups include all the locally compact second-countable groups (see [Bec98, Prop. 3.C.2]), and nonexamples include the following Polish groups: Aut(X, µ), the group of unitaries of an infinite dimensional Hilbert space and the group of permutation of the integers. Using Rohlin’s lemma, it can be shown that whenever T ∈ Aut(X, µ) is aperiodic, the L1 full group [T ]1 is never SIN (meaning that its topology does not admit a basis at the identity made of conjugacy-invariant neighborhoods). To our knowledge, such L1 full groups are the first examples of cli Polish groups which are non-SIN but at the same time far from being locally compact groups. A fundamental feature of L1 full groups is that they are stable under taking induced maps. Proposition 3.5. Let Φ be a graphing, let [Φ]1 be the associated L1 -full group. Then for all T ∈ [Φ]1 , we have TA ∈ [Φ]1 , and d˜Φ (TA , idX ) 6 d˜Φ (T, idX ). Proof. For all n ∈ N, let An := {x ∈ X : TA (x) = T n (x)}. Then the An ’s form a partition of A. For all n ∈ N and all 0 6 m < n, we let Bn,m := T m (An ). The family (Bn,m )06m<n is a collection of disjoint sets. We now fix n ∈ N. By the triangle inequality, we have that for all x ∈ An , n dΦ (x, TA (x)) = dΦ (x, T (x)) 6 n X dΦ (T m (x), T m+1 (x)). m=0 By integrating and summing this over all n ∈ N, we deduce that Z dΦ (x, TA (x)) = A n Z XX n∈N m=0 Z dΦ (x, T (x)) 6 Bn,m dΦ (x, T (x)). X To conclude, we note that the left term above is equal to d˜Φ (TA , idX ), while the right term is equal to d˜Φ (T, idX ). 16 Remark. The above proof is basically the same as that of Kac’s formula [Kac47, Thm.2’]. We will use this proof again to get later a similar statement in the case of L1 full groups of Z-actions (see Prop. 4.4). The following lemma will provide us many involutions in L1 full groups. Lemma 3.6. Let Φ be a graphing. Then for every involution U ∈ [RΦ ], there S exists an increasing sequence of U -invariant sets An ⊆ supp U such that supp U = n∈N An and for all n ∈ N, UAn ∈ [Φ]1 . Proof. Let U ∈ [RΦ ] be an involution and for all n ∈ N, let An = {x ∈ supp U : dΦ (x, U (x)) < S n}. Since U is an involution and dΦ is symmetric each An is U invariant, and n∈N An = supp U , so µ(supp U \ An ) → 0. By the definition of An , each UAn belongs to [Φ]1 . Theorem 3.7. Let Φ be a graphing. Then [Φ]1 is dense in [RΦ ] for the uniform topology. Proof. Because full groups are generated by involutions (see [Kec10, Lem. 4.5]), we only need to show that every involution in [RΦ ] can be approximated by elements of [Φ]1 . Let U ∈ [RΦ ] be such an involution. By the previous lemmaSwe have an increasing sequence (An ) of U -invariant subsets of supp U such that n∈N An = supp U and for all n ∈ N the involution UAn belongs to [Φ]1 . So µ(supp U \ An ) → 0 and we conclude that du (UAn , U ) → 0 as desired. 3.2 The closure of the derived group is topologically generated by involutions As we will see later, given a graphing Φ and T ∈ [Φ]1 , the map A ∈ MAlg(X, µ) 7→ TA ∈ [Φ]1 is not continuous in general. We however have the following very useful statement. Proposition 3.8. Let A ⊆ X, let Φ be a graphingSand take T ∈ [Φ]1 . If (An ) is an increasing sequence of subsets of A such that A = n∈N An then TA = lim TAn . n→+∞ Proof. Let k : X → N be defined by TA (x) = T k(x) (x) and similarly for all n ∈ N define kn : X → N by TAn (x) = T kn (x) (x). Since µ(An 4 A) → 0, we have that kn → k pointwise. So if we let Bn = {x ∈ X : k(x) 6= kn (x)}, we have µ(Bn ) → 0. By the definition of Bn we get Z d˜Φ (TAn , TA ) = dΦ (TAn (x), TA (x)) Bn The triangle inequality now yields Z Z ˜ dΦ (TAn , TA ) 6 dΦ (TAn (x), x) + Bn Bn 17 dΦ (TA (x), x). (1) Since An ⊆ A we have (TA )An = TAn , so Proposition 3.5 yields d˜Φ (TAn , idX ) 6 d˜Φ (TA , idX ). By developing the left-hand term we get Z Z Z dΦ (TAn (x), x) + dΦ (TAn (x), x) 6 d˜Φ (TA (x), x)). Bn X\Bn X Since for all x ∈ X \ Bn we have TA (x) = TAn (x), we may rewrite this as Z Z Z dΦ (TAn (x), x) 6 dΦ (TA (x), x) − dΦ (TA (x), x) Bn X X\Bn Z Z dΦ (TAn (x), x) 6 dΦ (TA (x), x). Bn Bn We now reinject this inequality into inequality (1) and obtain Z dΦ (TA (x), x). d˜Φ (TAn , TA ) 6 2 Bn By the dominated convergence theorem we now have conclude that dΦ (TAn , TA ) → 0 as desired. R Bn dΦ (TA (x), x) → 0 so we The above proposition is a direct consequence of the dominated convergence theorem when the sets An are already TA -invariant. We will often use it in this easier case, but the general statement will appear at a crucial point towards proving that the closure of the derived group of [Φ]1 is topologically generated by involutions. Let us for the moment make sure that every involution belongs to the closure of the derived group. Lemma 3.9. Suppose Φ is aperiodic. Then every involution in [Φ]1 belongs to the closure of derived group of [Φ]1 . Proof. Let T ∈ [Φ]1 be an involution. We can find B ⊆ X such that supp T = B t T (B). By aperiodicity, there is an involution Ũ ∈ [RΦ ] such that supp Ũ = B (see Lemma 2.12). We then find C ⊆ B such that B = C t Ũ (C). Lemma 3.6 then S yields an increasing sequence of Ũ -invariant sets Bn such that B = n∈N Bn , and for all n ∈ N, ŨBn ∈ [Φ]1 . We let Cn = C ∩ Bn , and Tn := TCn tT (Cn ) . Now we define an involution Un in the L1 full group of Φ by  if x ∈ Bn t U (Bn ),  Ũ (x) Un (x) := T Ũ T (x) if x ∈ T (Bn t U (Bn )),  x else. By construction, the commutator [Tn , Un ] is theS tranformation induced by T on the T -invariant set An := Bn t T (Bn ), and we have n∈N An = supp T . We conclude by the previous proposition that [Tn , Un ] → T as desired. 18 Remark. I do not know wether involutions actually belong to the derived group itself. As a more general question, when does it happen that D([Φ]1 ) is closed? Note that full groups of measure-preserving equivalence relations (which are a special case of L1 full groups by 3.2) are perfect if and only if the equivalence relation is aperiodic (see [LM14b, Prop. 3.6]). We will now gradually establish that conversely, the closure of D([Φ]1 ) is topologically generated by involutions. Lemma 3.10. Let Φ be an aperiodic graphing. Then every periodic element of [Φ]1 belongs to the closed subgroup of [Φ]1 generated by involutions and hence to the closure of the derived group of [Φ]1 . Proof. Let U ∈ [Φ]1 be periodic, and for all n ∈ N∗ let An denote the U -invariant set of all x ∈ X whose U -orbit has cardinality n. By Proposition 3.8, we have U = lim U n→+∞ Sn i=1 Ai = lim n→+∞ n Y UAi . i=2 so it suffices to prove the statement for cycles, i.e. elements U of [Φ]1 for which there is n ∈ N such that every non-trivial U -orbit has cardinality n. If U ∈ [Φ]1 is such an element, we may find A ⊆ X such that supp U = n G U i (A). i=1 For i ∈ {1, ..., n − 1} we let Ui ∈ [Φ]1 be the involution defined by   U (x) if x ∈ U i (A) U −1 (x) if x ∈ U i+1 (A) Ui (x) =  x else. It now follows from the well-known identity in symmetric groups (1 2 · · · n) = (1 2)(2 3) · · · (n − 1 n) that U = U1 U2 · · · Un−1 , which proves the first part of the lemma. The “as well as” part follows from Lemma 3.9. Lemma 3.11. Let Φ be an aperiodic graphing. The L1 full group [Φ]1 is generated by elements whose support has Φ-conditional measure everywhere less than 1/4. Remark. As the proof shows, 1/4 can be replaced by any  > 0, but we won’t need that. Proof. First note that by breaking a fundamental domain into pieces of small conditional measure (using Maharam’s lemma 2.9), one can show every periodic element is the product of elements whose support has Φ-conditional measure everywhere less than 1/4. 19 Now if T ∈ [Φ]1 , one may write T = Tp T 0 where Tp is periodic and all the nontrivial T 0 -orbits are infinite. By Maharam’s lemma we then find A ⊆ supp T 0 such that for all x ∈ supp T 0 , we have 0 < µT 0 (A)(x) < 1/4. Then A intersects almost every non-trivial T 0 -orbit so that T 0 T 0 −1 A is periodic. Since 0 A has T -conditional measure everywhere less than 1/4, the same is true of its Φconditional measure. So the support of T 0 A has Φ-conditional measure everywhere 0 less than 1/4 as wanted, and T = Tp (T 0 T 0 −1 A )T A can be written as a product of elements whose support has Φ-conditional measure everywhere less than 1/4. Theorem 3.12. Let Φ be an aperiodic graphing. Then the closure of the derived group of [Φ]1 is topologically generated by involutions. Proof. Denote by S the set of T ∈ [Φ]1 whose support has Φ-conditional measure everywhere less than 1/4. By the previous lemma S generates [Φ]1 . Since S is moreover invariant under conjugacy, it suffices to show that the commutator of any two elements of S belongs to the closed subgroup generated by involutions. So let T, U ∈ S and let B = supp g ∪supp h. There exists an involution V ∈ [RΦ ] such that B and V (B) are disjoint and supp V = B t V (B) =: A. By Lemma 3.6 there exists S an increasing sequence (An ) of V -invariant subsets of A such that VAn ∈ [Φ]1 and An = A. By Proposition 3.8 we have TAn → T and UAn → U , so it suffices to show that for every n ∈ N, the commutator [TAn , UAn ] belongs to the closed group generated by involutions. To this end, we fix n ∈ N and observe that the set supp TAn ∪ supp UAn is disjoint from VAn (supp TAn ∪ supp UAn ). It follows that if we let Tn0 = VAn TA−1 V , then Tn0 n An commutes with both TAn and UAn . So we have [TAn , UAn ] = [TAn Tn0 , UAn ] = [(TAn VAn TA−1 )VAn , UAn ] n and since taking a commutator means multiplying by a conjugate of the inverse, this equality implies [TAn , UAn ] is a product of involutions in [Φ]1 . As explained before [TAn , UAn ] → [T, U ] so this ends the proof. As a consequence, we can easily derive that the closure of the derived group is connected. Corollary 3.13. Let Φ be an aperiodic graphing. Then D([Φ]1 ) is connected. Proof. By the previous theorem it suffices to show that the group generated by involutions in [Φ]1 is connected, and we will actually prove that it is path connected. Let U ∈ [Φ]1 be an involution and let A be a fundamental domain for U . There is an increasing family of Borel subsets (At )t∈[0,1] such that A0 = ∅ and A1 = A and for all t ∈ [0, 1], µ(At ) = tµ(A). Then by the dominated convergence theorem the family of involutions Ut := UAt ∪U (At ) is a continuous path from U0 = idX to U1 = U . So every involution in [Φ]1 is connected by a path to the identity, and we conclude that the group generated by involutions is path connected as desired. 20 For ergodic Z-actions, D([T ]1 ) is actually the connected component of the identity (this follows from Corollary 4.20 along with the above result). We end this section by refining the fact that the closure of the derived group is topologically generated by involutions so as to obtain the same statements that Kittrell and Tsankov had obtained for full groups in [KT10, Sec. 4.2]. The proof is actually the same as Kittrell and Tsankov’s modulo the use of the dominated convergence theorem and we reproduce it here for the reader’s convenience. Let Φ be a graphing and ϕ ∈ Φ. If A is a Borel subset of dom ϕ such that A ∩ ϕ(A) = ∅, we denote by Iϕ,A the involution in Aut(X, µ) defined by: for all x ∈ X,   ϕ(x) if x ∈ A ϕ−1 (x) if x ∈ ϕ(A) Iϕ,A (x) =  x else. Note that Iϕ,A ∈ [Φ]1 since dΦ (x, Iϕ,A (x)) 6 1 for all x ∈ X. Theorem 3.14. Let Φ be a graphing. Then the group D([Φ]1 ) is topologically generated by the set {Iϕ,A : ϕ ∈ Φ, A ⊆ dom ϕ and A ∩ ϕ(A) = ∅} . Proof. By Theorem 3.12, we only need to show that involutions belong to the group topologically generated by the set {Iϕ,A : ϕ ∈ Φ, A ⊆ dom ϕ and A ∩ ϕ(A) = ∅} . So let U ∈ [Φ]1 be an involution, and let A be a fundamental domain for the restriction of U to its support. Let (wn ) enumerate the Φ-words, and for each n ∈ N let An ={x ∈ A : U (x) = wn (x) and for all m ∈ N if U (x) = wm (x) then either |wm | > |wn | or m > n} Then (An ) forms a partition of A and for all n ∈ N, if wn = ϕ1 · · · ϕk , then for all x ∈ An we have that for all i 6= j ∈ {1, ..., k} ϕi · · · ϕk (x) 6= ϕj · · · ϕk (x) by minimality of the length of wn . By the dominated convergence theorem, U = lim N Y N →+∞ Iwn ,An , n=0 so we may actually assume that U = Iϕ1 ···ϕk ,B for some B ⊆ X and ϕ1 , ..., ϕk ∈ Φ±1 such that for all i 6= j ∈ {1, ..., k} and all x ∈ B, ϕi · · · ϕk (x) 6= ϕj · · · ϕk (x). Since X is a standard Borel space, a well-known argument yields that we may actually find a partition (Bm )m∈N of B such that for all m ∈ N and all i 6= j ∈ {1, ..., k}, ϕi · · · ϕk (Bm ) ∩ ϕj · · · ϕk (Bm ) = ∅ 21 (see for instance [EG14, Lem. 5.2]). Again by the dominated convergence theorem U = lim N →+∞ N Y Iϕ1 ···ϕk ,Bm , n=0 so we may actually assume that U = Iϕ1 ···ϕk ,C for some C ⊆ X such that for all i 6= j ∈ {1, ..., k}, ϕi · · · ϕk (C) ∩ ϕj · · · ϕk (C) = ∅ Then note that for all m ∈ N and i ∈ {1, ..., k − 1}, the identity (i − 1 k) = (i − 1 i)(i k)(i − 1 i) in the symmetric group over {0, ..., k} yields Iϕi ···ϕk ,C = Iϕi ,ϕi+1 ···ϕk (C) Iϕi+1 ···ϕk ,C Iϕi ,ϕi+1 ···ϕk (C) so that by a downward recurrence on i ∈ {1, ..., k} we have that Iϕi ···ϕk ,C belongs to the group generated by {Iϕ,A : ϕ ∈ Φ, A ⊆ dom ϕ and A ∩ ϕ(A) = ∅} . In particular, U = Iϕ1 ···ϕk ,C belongs to such a group, as wanted. 3.3 Dye’s reconstruction theorem We now show that L1 full groups as abstract groups can only be isomorphic through a measure-preserving transformation, which is a key step in order to see than L1 full groups of ergodic Z-actions are a complete invariant of flip conjugacy. Fortunately, the main result that we need for this has already been proven by Fremlin in a much more general context. Definition 3.15 (Fremlin). A subgroup G 6 Aut(X, µ) has many involutions if for every Borel A ⊆ X non-null, there exists a non-trivial involution U ∈ G such that supp U ⊆ A. Proposition 3.16. Let Φ be a graphing. The following are equivalent (1) Φ is aperiodic; (2) [Φ]1 has many involutions; (3) D([Φ]1 ) has many involutions; (4) D([Φ]1 ) has many involutions. Proof. Note that (2) and (3) are equivalent since every involution in [Φ]1 belongs to D([Φ]1 ) by Lemma 3.10. Let us show by contrapositive that (2)⇒(1): if Φ is not aperiodic, then we may find a non-null set A which intersects every Φ-orbit in at most one point. Clearly 22 every element of [Φ]1 (actually of [RΦ ]) supported in A is trivial, in particular [Φ]1 does not have many involutions. For the converse (1)⇒(2), suppose that Φ is aperiodic and let A ⊆ X be nonnull. Then by Lemma 2.12, there exists an involution U ∈ [RΦ ] whose support is equal to A. Now Lemma 3.6 ensures us that there is a non-null U -invariant set A0 such that UA0 ∈ [Φ]1 , witnessing that [Φ]1 has many involutions. So the implication (1)⇒(2) also holds and we conclude that (1), (2) and (3) are equivalent. Clearly (4) ⇒ (2), so we now only need to show that (2)⇒(4). Suppose that [Φ]1 has many involutions and let A be a non-null Borel subset of X. Let U ∈ [Φ]1 be an involution supported in A, let B be a fundamental domain of the restriction of U to its support. Finally, let V ∈ [Φ]1 be an involution supported in B. Since B and U (B) are disjoint, U V U V is an involution supported in A and belonging to D([Φ]1 ) as wanted: D([Φ]1 ) has many involutions. We denote by Aut∗ (X, µ) the group of non-singular transformations of (X, µ), i.e. Borel bijections T : X → X such that for all Borel A ⊆ X, one has µ(A) = 0 if and only if µ(T (A)) = 0. We can now state and use Fremlin’s theorem, which is a generalization of Dye’s reconstruction theorem for full groups. Theorem 3.17 (Fremlin [Fre04, 384D]). Let G, H 6 Aut(X, µ) be two groups with many involutions. Then any isomorphism between G and H is the conjugacy by some non-singular transformation: for all ψ : G → H group isomorphism, there is S ∈ Aut∗ (X, µ) such that for all T ∈ G, ψ(T ) = ST S −1 . Corollary 3.18. Let Φ and Ψ be two ergodic graphings. Then every isomorphism between [Φ]1 and [Ψ]1 is the conjugacy by some measure-preserving transformation. Proof. By Proposition 3.16 and the previous theorem, if ψ : [Φ]1 → [Ψ]1 is a group isomorphism then there is S ∈ Aut∗ (X, µ) such that for all T ∈ [Φ]1 one has ψ(T ) = ST S −1 . Let f be the Radon-Nikodym derivative of the probability measure S∗ µ with respect to µ. Since µ is preserved by every T ∈ [Φ]1 , the measure S∗ µ is preserved by every T 0 ∈ [Ψ]1 . So f ◦ T 0 = f a.e. for every T 0 ∈ [Ψ]1 . But Ψ is ergodic so this implies f is constant: the measure µ is preserved by S as wanted. 3.4 Closed normal subgroups of the closure of the derived group Given a group G ⊆ Aut(X, µ), we denote by GA the subgroup of all T ∈ G such that supp T ⊆ A. Note that if Φ is a graphing and A is Φ-invariant then D([Φ]1 )A is a closed normal subgroup of D([Φ]1 ). We will see that conversely if Φ is aperiodic then all the closed normal subgroups of D([Φ]1 ) arise in this manner. Our approach is based on the study of closed normal subgroups generated by involutions. 23 Lemma 3.19. Let Φ be a graphing. If A is a Φ-invariant set then D([Φ]1 A ) = D([Φ]1 )A = D([Φ]1 )A . Proof. To simplify notation we let G = [Φ1 ]A , so that we aim to prove that D(GA ) = D(G)A = D(G)A . To this end we show that the chain of inclusions D(GA ) ⊆ D(G)A ⊆ D(G)A ⊆ D(GA ) holds. • We clearly have the inclusion D(GA ) ⊆ D(G)A so that D(GA ) ⊆ D(G)A . • Next, note that D(G)A is a closed subgroup of G which contains every element of D(G)A . It follows that D(G)A ⊆ D(G)A . • For the last remaining inclusion D(G)A ⊆ D(GA ), let T ∈ D(G)A . Then T is supported on A and is a limit of elements Tn belonging to D(G). Now note that by definition of the topology T is also the limit of the sequence of transformations (TnA )n∈N . Moreover each TnA can be rewritten as a product of commutators of elements supported in A by taking the induced transformations instead, which is harmless since A is invariant by every element appearing in such a product. We conclude that T ∈ D(GA ) as desired. We now establish a crucial lemma allowing a better understanding closures of conjugacy classes of involutions. Lemma 3.20. Let Φ be a graphing. Let U and V be two involutions such that their supports are disjoint and have the same Φ-conditional measure. Then V belongs to the closure of the conjugacy class of U . Proof. Let A and B be fundamental domains of the respective restrictions of U and V to their support. Then µΦ (A) = µΦ (B) so there exists an involution T ∈ [RΦ ] such that T (A) = B and we can moreover assume that T (x) = x for all x ∈ A ∩ B. For all n ∈ N we let An = {x ∈ A : dΦ (x, T (x)) < n} and Bn = {x ∈ B : dΦ (x, T (x)) < n}. Since T is an involution and T (A) = B, we also have B Sn = T (An ). Note that (A ) and (B ) are increasing sequences of sets and that n n n An = A, S 0 n Bn = B. We now define a new involution Tn  T (x) if x ∈ An t Bn    V T U (x) if x ∈ U (An ) Tn0 (x) = U T V (x) if x ∈ V (Bn )    x else. The definition of An and the fact that V, U ∈ [Φ]1 yield that Tn0 ∈ [Φ]1 as well. For all n ∈ N and all x ∈ X, an easy calculation yields that: • if x ∈ (A ∪ U (A)) \ (An ∪ U (An )), then Tn0 U Tn0 (x) = U (x), • else if x ∈ (Bn ∪ V (Bn ), then Tn0 U Tn0 (x) = V (x) and 24 • else Tn0 U Tn0 (x) = x. So Tn0 U Tn0 → V pointwise. Moreover the trichotomy above also shows that for all x ∈ X we have dΦ (x, Tn0 U Tn0 (x)) 6 dΦ (x, U (x)) + dΦ (x, V (x)), so by the dominated convergence theorem d˜Φ (Tn0 U Tn0 , V ) → 0. We have therefore established that V belongs to the closure of the conjugacy class of U . Lemma 3.21. Let Φ be an aperiodic graphing, let U ∈ [Φ]1 be an involution such that µΦ (supp U ) 6 1/2. Let A be the Φ-closure of the support of U . Then the closed subgroup generated by conjugates of U by elements of D([Φ]1 ) contains D([Φ]1 )A . Proof. By Lemma 3.19 we may as well assume that A = X so that our aim becomes to show that whenever U has support whose Φ-closure is equal to X and which satisfies µΦ (supp U ) 6 1/2, the closed subgroup generated by conjugates of U by involutions contains D([Φ]1 ). So let U be as such, let B = supp U and denote by G the closed subgroup generated by conjugates of U by elements of D([Φ]1 ). Since µΦ (supp U ) 6 1/2 by Maharam’s lemma we may find a set C disjoint from B such that µΦ (C) = µΦ (B). There is an involution V ∈ [RΦ ] supported on B t C such that V (B) = C. For all n ∈ N we let Bn = {x ∈ B : dΦ (x, V (x)) < n and dΦ (U (x), V U (x)) < n} Then (Bn ) is an increasing sequence of U -invariant Borel subsets of B such that S S n∈N Bn = B. Let An be the Φ-closure of Bn . Observe that n∈N An = X, so the proof will be finished once the following claim is proven. Claim. For all n ∈ N, the group G contains D([Φ])An . We now prove the claim. Let n ∈ N; the sets and functions that are to be defined depend on n but we will omit the subscript n for readability. Let F be a fundamental domain of the restriction of U to its support, let Fn be its intersection with Bn , which is a fundamental domain for the restriction of U to Bn . Let f = µΦ (Fn ) = µΦ (Bn )/2. We know that D([Φ])An is topologically generated by involutions, so it suffices to show that any involution in [Φ]1 supported in An belongs to G. So let W be an involution supported in An . Let D be a Borel fundamental domain for the restriction of W to its support: then supp W = D t W (D). By Corollary 2.10 we have a partition (Dm )m∈N of D such that for all m ∈ N, µΦ (Dm ) 6 f6 . For all m ∈ N we let Wm = WDm tW (Dm ) . By the dominated convergence theorem, W = lim k→+∞ k Y Wm m=0 so we only need to show that every Wm belongs to G. 25 B C Bn U V Fn Let us fix m ∈ N, and for notational simplicity let Γ = hU, V i ' (Z/2Z × Z/2Z) o (Fn ) Z/2Z. Since µΦ (Dm ) 6 µΦ10 we can find a subset E ⊆ Fn such that µΦ (E) = µΦ (Dm )/2 and E is disjoint from Γ(Dm t W (Dm )). Then Γ(E) is disjoint from Dm t W (Dm ) = supp Wm . Consider the Γ-invariant set Ẽ = (E t U (E)) t V (E t U (E)) and the involution Ṽ = VẼ . An easy calculation yields U Ṽ U Ṽ = UẼ so that UẼ ∈ G. Now the support of UẼ has Φ-conditional measure µΦ (Ẽ) = 4µΦ (E) = 2µΦ (Dm ) = µΦ (supp Wm ). Since moreover Ẽ and supp Wm are disjoint, we infer by Lemma 3.20 that Wm belongs to the closure of the conjugacy class of of UẼ . Since we have already established that UẼ ∈ G, we can conclude that Wm ∈ G. As already observed, this ends the proof of the claim and therefore of the theorem. Proposition 3.22. Let Φ be an aperiodic graphing, let T ∈ [Φ]1 and let A be the Φ-closure of the support of T . Then the closed subgroup generated by conjugates of T by elements of D([Φ]1 ) contains D([Φ]1 )A . Proof. Again by Lemma 3.19 we only need to show that whenever T has a support whose Φ-closure is equal to X, the closed subgroup generated by conjugates of T by involutions contains D([Φ]1 ). Let G be the closed subgroup generated by conjugates of T by elements of D([Φ]1 ). We start by applying Proposition 2.7 and find a partition (A1 , A2 , B1 , B2 , B3 ) of supp T such that A2 = T (A1 ), B2 = T (B1 ) and B3 = T 2 (B1 ). Let C = A1 t b1 , then by construction C ∩ T (C) = ∅ and C intersects every non-trivial T -orbit. Since the Φ-closure of supp T is equal to X this implies C intersects every Φ-orbit. Let G be the closed subgroup generated by conjugates of T by elements of D([Φ]1 ). Let U be an involution in [Φ]1 whose support is equal to C. For all S n ∈ N, let Cn = {x ∈ X, dΦ (x, U (x)) < n}. Then n∈N Cn = C and for all n ∈ N we have UCn ∈ [Φ]1 . Now since U (C) is disjoint from C we also have that U (Cn ) 26 and Cn are disjoint so that [T, UCn ] is an involution belonging to G whose support is Cn t T (Cn ). Applying Lemma 3.21 S we deduce that G contains D([Φ]1 )An where An is the Φ-closure of Cn . Since n∈N Cn = C and C intersects every Φ-orbit, we have that S A = X. Moreover the sequence (An ) is increasing, so a direct application of n∈N n S Proposition refprop: weak order continuity yields that n∈N D([Φ]1 )An is dense in D([Φ]1 ) so that G contains D([Φ]1 ) as wanted. We can now easily prove our main results for this section. Theorem 3.23. Let Φ be an aperiodic graphing. Then for every closed normal subgroup N of D([Φ]1 ) there is a (unique up to measure zero) Φ-invariant Borel set A such that N = D([Φ]1 )A . Proof. Let t = sup{µ(A) : A is Φ-invariant and D([Φ]1 )A ⊆ N }, and let An be a sequence of subsets of X such that µ(An ) → t and D([Φ]1 )An ⊆ N . Note that whenever A0 , A00 are two Φ-invariant subsets of X then the group generated by D([Φ]1 )A0 ∪ D([Φ]1 )A00 is equal to D([Φ]1 )A0 ∪A00 . So we may assume the sequence (An ) is increasing. S Let A = n∈N An , then µ(A) = t. By proposition 3.8 since N is closed we actually have that N contains D([Φ]1 )A . Conversely, let T ∈ N and let B be the Φ-closure of the support of T . By Proposition 3.22 the group N contains D([Φ]1 )B . So N contains D([Φ]1 )A∪B hence µ(A ∪ B) 6 t. But µ(A) = t so B ⊆ A up to measure zero, hence T ∈ D([Φ]1 A . So D([Φ]1 )A contains N and we conclude that N = D([Φ]1 )A . Finally, the uniqueness of A up to measure zero follows from the fact that given any B ⊆ X of positive measure, there is a non trivial T ∈ D([Φ]1 ) supported in B (see Proposition 3.16). Corollary 3.24. Let Φ be an aperiodic graphing. Then Φ is ergodic if and only if D([Φ]1 ) is topologically simple. We will see that D([Φ]1 ) and D([Φ]1 ) are never simple when Φ = {T } and T ∈ Aut(X, µ) is aperiodic (see Theorem 4.26). 3.5 Link with topological full groups In this section we connect L1 full groups to the well-studied topological full groups. Since we are dealing with L1 full groups associated to graphings, we need to introduce their natural topological counterparts. Let 2N be the Cantor space and let ν be a Borel probability measure on it. A continuous graphing on the measured Cantor space (2N , ν) is a graphing Φ such that for all ϕ ∈ Φ, the sets dom ϕ, rng ϕ are clopen and ϕ : dom ϕ → rng ϕ is a homeomorphism. 27 Definition 3.25. Let Φ be a continuous graphing on (2N , ν). The topological full group of Φ, denoted by [Φ]c , is the group of all homeomorphisms T of 2N such that for all x ∈ 2N there is a neighborhood U of x and a Φ-word w such that for all y ∈ U , T (x) = w(x). It follows from compactness that one can think equivalently of the topological full group of a continuous graphing Φ as the group of all maps obtained by cutting and pasting finitely many Φ-words along a partition into clopen sets. In particular, [Φ]c is always a countable group. The most interesting case is when Φ is actually a countable group acting by homeomorphisms on a compact set (or a generating set for such a countable group): we then recover the usual definition of the topological full group. We denote by [Φ]c−inv denote the group generated by involutions in a topological full group [Φ]c . The following result is an L1 full group version of a result of Miller [Mil05] for full groups, also proved by Kittrell-Tsankov [KT10, Prop. 4.1] and Medynets [Med07] in the case of a single homeomorphism. Theorem 3.26. Let ν be a Borel probability measure on 2N , and let Φ be a continuous graphing on (2N , µ). Then the group [Φ]c−inv is dense in D([Φ]1 ). Proof. Let us recall that if ϕ ∈ Φ and A is a Borel subset of dom ϕ such that A ∩ ϕ(A) = ∅, we denote by Iϕ,A the involution in [Φ]1 defined by: for all x ∈ X,   ϕ(x) if x ∈ A ϕ−1 (x) if x ∈ ϕ(A) Iϕ,A (x) =  x else. By Theorem 3.14, the group D([Φ]1 ) is topologically generated by such involutions, so we only need to check that, given ϕ ∈ Φ and a Borel subset A of dom ϕ such that A ∩ ϕ(A) = ∅, the involution Iϕ,A can be approximated by elements of [Φ]c−inv . So let ϕ ∈ Φ and let A be a Borel subset of dom ϕ such that A∩ϕ(A) = ∅. Since ν is regular, there is a sequence (An ) of clopen subsets of 2N such that ν(An 4 A) → 0. Consider the set A0n := An \ϕ(An ). Note that since ϕ is a partial homeomorphism, A0n is still clopen, and by definition A0n is disjoint from ϕ(A0n ) so the involution Iϕ,A0n belongs to [Φ]c−inv . We now have Z Z d˜Φ (Iϕ,A , Iϕ,A0n ) ) = dΦ (ϕ(x), x) + dΦ (x, ϕ(x)) A\A0n A0n \A Z Z −1 + dΦ (ϕ (x), x) + dΦ (x, ϕ−1 (x)) ϕ(A\A0n ) =2µ(A0n 4 A). ϕ(A0n \A) Moreover since ν(An 4 A) → 0, since ϕ preserves the measure and since A is disjoint from ϕ(A), a straightforward computation yields µ(A0n 4 A) → 0. We conclude that d˜Φ (Iϕ,A , Iϕ,A0n ) ) → 0 as desired. Let us now show that the similar statement that [Φ]c is dense in [Φ]1 is not true in general. 28 Example 3.27. Consider the odometer T0 on 2N . For every word s with letters in {0, 1}, we let Ns ⊆ 2N be the set of infinite words which start by s. For every n ∈ N, let ϕn be the restriction of T0 to N1n 0 , and then let Φ = {ϕn : n ∈ N}. It is easily checked that the topological full group [Φ]c is the group S2∞ of dyadic permutations. In particular [Φ]c is locally finite, hence contained in D([Φ]1 ). We conclude from Theorem 3.26 that [Φ]c = D([Φ]1 ). However the L1 full group of Φ is equal to the L1 full group of T0 since up to measure zero they induce the same graph. But T0 has index 1 (see Definition 4.3) hence cannot belong to D([T0 ]1 ) = D([Φ]1 ), and we conclude that [Φ]c is not dense in [Φ]1 . It is well-known that every countable group Γ of measure-preserving transformation of (X, µ) is up to measure zero conjugate by a measure-preserving isomorphism to a group of homeomorphism of the Cantor space preserving a Borel probability measure ν (see for instance [Gla03, Thm. 2.15]). Here we check that the same can be done for graphings so as to be able to apply our density result above in the right level of generality. Recall that the group of homeomorphisms of the Cantor space is canonically isomorphic to the group of automorphisms of the Boolean algebra of clopen subsets of the Cantor space. Furthermore, the Boolean algebra of clopen subsets of the Cantor space is the unique countable atomless Boolean algebra (see for instance [GH09, Chap. 16]). We can now state and prove that any graphing is conjugate to a continuous one up to measure zero. Proposition 3.28. Let (X, µ) be a standard Borel space, and let Φ be a graphing on X. Then there exists a Borel probability measure ν on 2N , a continuous graphing Ψ on (2N , ν), a Φ-invariant full measure Borel subset X0 of X and Borel measurepreserving injection T : (X0 , µ) → (2N , ν) and a continuous graphing Ψ on 2N such that ΨT (X0 ) = T ΦX0 T −1 . Proof. Let C be a countable separating family of Borel sets containing X. Let D be set of finite Boolean combinations of sets of the form w(C) or w−1 (C) for w a Φ-word and C ∈ C. Then D is countable separating, contains the domain and range of every ϕ ∈ Φ, and is Φ-invariant in the sense that for every ϕ ∈ Φ and D ∈ D, ϕ±1 (D) ∈ D. Each atom of D is a singleton since D contains a separating family. Let A be the reunion of the atoms of D, then A is countable and hence has measure zero. Let X0 := X \ A, and consider the Boolean algebra DX0 := {D ∩ X0 : D ∈ D}. induced by D on X0 . By Stone duality, DX0 is isomorphic to the algebra of clopen subsets of its Stone space Y . Since DX0 is a countable atomless algebra its Stone space Y is homeomorphic to 2N so we may as well assume that Y = 2N . By construction every ϕ ∈ Φ induces a Boolean algebra isomorphism between the Boolean algebra induced by DX0 on dom ϕ and the Boolean algebra induced by DX0 on rng ϕ, so every ϕ ∈ Φ induces a homeomorphism ϕ̃ between the corresponding clopen subsets of Y . Let Ψ = {ϕ̃ : ϕ ∈ Φ}, we will check that such a continuous graphing works. 29 We define a map T : X0 → Y by associating to every x ∈ X0 the unique element T (x) of the Stone space of DX0 defined by x. Since DX0 contains a separating family of X0 , the map T is injective. Moreover it conjugates the restriction ΦX0 and ΨT (X0 ) by the definition of Ψ. Finally the map T is also Borel since the preimage of a clopen set belongs to DX0 , so by putting ν = T∗ µ we get the desired statement. 4 L1 full groups of Z-actions We now focus on L1 full groups of Z-actions, for which we can say a lot more. The main feature is that for Z-actions, L1 full groups become invariants of flip-conjugacy. To see this, we need the following fundamental result of Belinskaya. Theorem 4.1 (Belinskaya, [Bel68]). Let T and T 0 be two ergodic transformations with the same orbits, and suppose furthermore that T 0 ∈ [T ]1 . Then there exists S ∈ [T ] such that either T = ST 0 S −1 or T −1 = ST 0 S −1 . With this in hand, we can now prove that L1 full groups of ergodic Z-actions are complete invariants of flip conjugacy. Theorem 4.2. Let T and T 0 be two measure-preserving ergodic transformation of a standard probability space (X, µ). Then the following are equivalent: (1) T and T 0 are flip-conjugate; (2) the groups [T ]1 and [T 0 ]1 are abstractly isomorphic; (3) the groups [T ]1 and [T 0 ]1 are topologically isomorphic. Proof. Clearly (1) implies (3) and (3) implies (2). Suppose that (2) holds and let ψ : [T ]1 → [T 0 ]1 be an abstract group isomorphism. By Corollary 3.18, there is S ∈ Aut(X, µ) such that for all U ∈ [T ]1 , we have ψ(U ) = SU S −1 . In particular, we have that ST S −1 ∈ [T 0 ]1 . Since ψ is surjective, ST S −1 must have the same orbits as T 0 . So by Belinkskaya’s Theorem 4.1, ST S −1 and T 0 are flip-conjugate, so T and T 0 are flip-conjugate. So (2) implies (1), which ends the proof. 4.1 The index map One of the specific features of L1 full groups of aperiodic Z-actions is the existence of a natural homomorphism into R, given by integrating the corresponding cocycles. We will see later that in the ergodic case, it actually takes values into Z. 30 Definition 4.3. Let T ∈ Aut(X, µ) be ergodic. The index of S ∈ [T ]1 is defined by Z cS (x)dµ(x), IT (S) := X where cS is the cocycle associated to S ∈ [T ]. When the transformation T we are considering is clear from the context, we will also simply write the index map as I. From the cocycle identity cSS 0 (x) = c0S (S(x)) + cS (x) and the fact that elements of [T ]1 are measure preserving we deduce that the index map is a homomorphism. We will denote its kernel by [T ]0 . Our first observation is that taking induced transformations does not change the index. The proof is very similar to that of Kac’s formula, which should be no surprise since Kac’s formula corresponds to the case S = T . Proposition 4.4. Let S ∈ [T ]1 , and let A ⊆ X intersect every non trivial S-orbit. Then we have the equality I(SA ) = I(S). Proof. For all n ∈ N, let An := {x ∈ X : TA (x) = T n (x)}. Then the An ’s form a partition of A. For all n ∈ N and all 0 6 m < n, we let Bn,m := T m (An ). The family (Bn,m )06m<n is a partition of the support of S since A intersects every non trivial S-orbit. We now compute Z I(SA ) = cSA (x)dµ(x) A XZ cSA (x)dµ(x) = An n∈N = n−1 X XZ cS (S m (x))dµ(x), An m=0 n∈N the last equality being a consequence of the cocycle identity. We deduce that I(SA ) = n−1 Z XX An n∈N m=0 = XZ n∈N Z cS (S m (x))dµ(x) cS (x)dµ(x) Bn,m = cS (x)dµ(x) supp S I(SA ) = I(S) The following corollary can also be obtained directly by noting that if S ∈ [T ]1 , then on every finite S-orbit the average of nS must be equal to zero. Corollary 4.5. Let T ∈ Aut(X, µ) be aperiodic. Every periodic element of [T ]1 has index zero. 31 Proof. Let S ∈ [T ]1 be a periodic element, let A ⊆ X be a fundamental domain for S (see the paragraph following Definition 2.3). Then SA = idX by definition so I(SA ) = 0. By the previous proposition we conclude that I(S) = I(SA ) = 0. One can completely describe the ergodic elements of the L1 full group of T ∈ Aut(X, µ) which are conjugate to T in terms of their index. This will be very useful to us in the next section. Theorem 4.6 (Belinskaya, [Bel68, Thm. 3.8]). Let T ∈ Aut(X, µ) be an ergodic transformation, and let T 0 ∈ [T ]1 . Then T and T 0 are flip-conjugate iff T 0 is ergodic and I(T 0 ) = ±1. 4.2 A characterization of supports of n-cycles In full groups of aperiodic equivalence relations, every Borel set is the support of an involution. Here we will see that this is not true in general for L1 full groups and give a complete characterization of sets which arise as supports of involutions, or more generally of n-cycles. Definition 4.7. A measure-preserving transformation T is an n-cycle if every T orbit has cardinality either 1 or n. Note that if p is prime, a p-cycle is precisely an element of Aut(X, µ) of order p. Our result is that n-cycles in L1 -full groups arise only in the most obvious manner: for instance there is an involution whose support is equal to X if and only if there exists a Borel set B such that X = B t T (B). Theorem 4.8. Let T be a measure-preserving ergodic transformation, let A be a Borel subset of X and let n ∈ N. Then A is equal to the support of an n-cycle if and only if there is B ⊆ A such that A = B t TA (B) · · · t TAn−1 (B) Remark. We should point out however that every Borel set arises as the support of a periodic transformation all whose orbits have cardinality at most 3 (see Lemma 4.24). Proof of Theorem 4.8. The sufficiency of the condition is easy: since TA ∈ [T ]1 by Kac’s formula (see Prop. 3.5), if A = B t TA (B) · · · t TAn−1 (B) we can define an n-cycle T0 ∈ [T ]1 whose support equates A by putting  F i if x ∈ n−2  TA (x) i=0 TA (B) −n−1 n−1 T0 (x) = T (x) if x ∈ TA (B) .  A x else. For the converse, let us recall that if G 6 Aut(X, µ) and A is a Borel subset of X, we let GA denote the group of elements of G supported in A. Observe that if T ∈ Aut(X, µ) then for all x, y ∈ A we have dT (x, y) > dTA (x, y) so that [T ]1 A 6 [TA ]1 (we will see in Theorem 5.2 that the inclusion is actually strict). This observation will be used implicitly in what follows. 32 Suppose that A is the support of an n-cycle T0 , let C be a fundamental domain for T0 , and consider the transformation S = TC T0 whose support also equates A and satisfies A = C t S(C) t · · · t S n−1 (C). By Proposition 2.5 the restriction of S to A is ergodic. Moreover note that TC = (TA )C so the the TA -index of TC is ITA (TC ) = 1 by Proposition 4.4. Since T0 is periodic, its TA -index is equal to 0 by Corollary 4.5 and we conclude that ITA (S) = 1. By applying Belinskaya’s theorem 4.6 to the restriction of S to A which is ergodic, we conclude that TA and S are conjugate by a measure-preserving transformation U supported in A. But recall that A = C t S(C) t · · · t S n−1 (C) so we have A = U (C) t TA (U (C)) · · · t TAn−1 (U (C)) hence TA satisfies the desired conclusion with B := U (C). Example 4.9. Let T ∈ Aut(X, µ) be a weakly mixing transformation. Then for all n ∈ N, there is no n-cycle in [T ]1 whose support equates X. Remark. The existence of B ⊆ A such that A = B t TA (B) · · · t TAn−1 (B) is a spectral property of TA : it is equivalent to the associated unitary operator on L2 (A) having e2iπ/n as an eigenvalue. 4.3 The L1 full group is generated by induced transformations along with periodic transformations We will now use the index map to find a nice generating set for [T ]1 . Our ideas are in the direct continuation of Belinskaya’s paper [Bel68]. Let us start by analyzing the action of elements of [T ]1 on the T -orbits. A sequence (nk )k∈N of integers is called almost positive if for all but finitely k ∈ N, we have nk > 0. It is called almost negative if (−nk )k∈N is almost positive, and a sequence which is either almost negative or almost positive is called almost sign invariant. It is an easy exercise to check that an injective sequence is almost positive iff it tends to +∞, while it is almost negative iff it tends to −∞. The following lemma thus implies that under the identification of a T -orbit with Z via k 7→ T k (x), every forward S-orbit either tends to +∞ or −∞. Lemma 4.10 ([Bel68, Thm. 3.2]). Let T ∈ Aut(X, µ) be aperiodic and S ∈ [T ]1 . For almost every x ∈ X such that the S-orbit of x is infinite, the sequence sign(cS k (x)) is almost sign invariant. Definition 4.11. An element S ∈ [T ]1 such that for all x ∈ supp S, the sequence (cS k (x))k∈N is almost positive is called almost positive. If for every x ∈ supp S the sequence (cS k (x))k∈N is almost negative, we say that S is almost negative. Note that every almost positive or almost negative element has to be aperiodic when restricted to its support. Also the inverse of an almost positive element is almost negative and vice-versa. We have a natural order on every T -orbit which allows for a more natural formulation of the previous definition. 33 Definition 4.12. Let T ∈ Aut(X, µ) be aperiodic. We define a partial order 6T on X by x 6T y if y = T n (x) for some n > 0. Note that 6T is total when restricted to a T -orbit, and that such a restriction corresponds to the natural ordering 6 on Z via n 7→ T n (x). We may now reformulate the previous definition as: S ∈ [T ]1 is almost positive if for almost every x ∈ supp S, S k (x) >T x for all but finitely many k ∈ N. Proposition 4.13. Let T ∈ Aut(X, µ) be aperiodic. Every element s ∈ [T ]1 can be written as a product S = Sp S+ S− , where Sp is periodic, S+ is almost positive, S− is almost negative and these three elements have disjoint supports. Proof. Let S ∈ [T ]1 . The set of x ∈ X such that the sequence (cS k (x))k∈N is almost positive (respectively almost negative) is S-invariant. Call this set A+ (respectively A− ). Finally let Ap denote the set of x ∈ X whose S-orbit is finite. Then (A− , A+ , Ap ) is a partition of X by the previous lemma, and since they are S-invariant we deduce that S = SA− SA+ SAp . Clearly Sp := SAp , S+ := SA+ and S− := SA− are as wanted. Definition 4.14. Let T ∈ Aut(X, µ) be aperiodic. If S ∈ [T ]1 , we say that S is positive if S(x) >T x for all x ∈ X. The proof of the following proposition uses the same ideas as in [Bel68, Lem. 3.5]. Proposition 4.15. Let T ∈ Aut(X, µ) be aperiodic, and let S ∈ [T ]1 be almost positive. Then there exists a positive element S 0 ∈ [T ]1 whose support is a subset of supp S such that SS 0 −1 and S 0 −1 S are periodic. Proof. Consider the set A := {x ∈ X : S k (x) >T x for all k > 0}. Note that TA (x) may be defined as the first element greater than x which belongs to A, so by the definition of A we have TA (x) 6T SA (x) for all x ∈ A. Let us show that A intersects every non trivial S-orbit. Let x ∈ supp S. Since S is almost positive, the sequence (nS k (x))k∈N is almost positive. Let k be the last natural integer such that nS k (x) 6 0, then by definition S k (x) ∈ A as wanted. Now for all x ∈ A we have SA (x) >T TA (x) >T x, in particular SA is positive. Since A intersects every S-orbit, we moreover have that SSA−1 and SA−1 S are periodic, so S 0 := SA is as desired. Lemma 4.16. Let T ∈ Aut(X, µ) be aperiodic, let S ∈ [T ]1 be positive and let A = supp S. Then STA−1 is positive. Proof. Since S is positive, we have A = {x ∈ X : S(x) >T x}. But then by the definition of TA ,for all x ∈ A we have S(x) >T TA (x). Since A is TA -invariant, we deduce that for all x ∈ A, STA−1 (x) >T x. But the support of STA−1 is contained in A so the previous inequality actually holds for all x ∈ X, and we conclude that STA−1 is positive. 34 Proposition 4.17. Let T ∈ Aut(X, µ) be ergodic. Then every positive element is the product of finitely many transformations which are induced by T : for all S ∈ [T ]+ 1 there exists k ∈ N and A1 , ..., Ak ⊆ X such that S = TA1 · · · TAk . In particular, the index of a positive element is a (positive) integer. Proof. Note that the index of a positive element has to belong to [0, +∞[, and that the index of a positive element S is null if and only if S = idX . We prove the result by induction on bI(S)c. First, if bI(S)c = 0, suppose towards a contradiction that S is non-trivial. let A = supp S, by the previous lemma STA−1 is positive. Since T is ergodic, A intersects almost every T -orbit. Proposition 4.4 thus yields that I(TA ) = I(T ) = 1. So I(STA−1 ) = I(S) − 1 < 0, but STA−1 is positive, a contradiction. The inductive step is basically the same: if bI(S)c = n + 1 for some n ∈ N, the support A of S is nontrivial so that S = TA S 0 where bI(S 0 )c = n. Theorem 4.18. Let T ∈ Aut(X, µ) be ergodic, then [T ]1 is generated by the set P er([T ]1 ) ∪ Ind([T ]1 ), where P er([T ]1 ) is the set of periodic elements in [T ]1 and Ind([T ]1 ) is its set of transformations induced by T . Proof. By Proposition 4.13 the group [T ]1 is generated by almost positive elements along with periodic elements. Now every almost positive element is the product of a positive element by a periodic element (Proposition 4.15) and every positive element is the product of finitely many induced transformations (Proposition 4.17). Corollary 4.19. Let T ∈ Aut(X, µ) be ergodic, then [T ]1 is generated by the set P er([T ]1 ) ∪ {T }, where P er([T ]1 ) is the set of periodic elements in [T ]1 . Proof. If A is a non-null subset of X, then by ergodicity A intersects almost every T orbit so that TA−1 T is periodic. In other words every induced transformation belongs to the group generated by periodic elements along with T , and since by the previous theorem periodic elements along with induced transformations generate [T ]1 we are done. Corollary 4.20. Let T ∈ Aut(X, µ) be ergodic. Then D([T ]1 ) is equal to the following three groups: • the group generated by periodic elements, • the group topologically generated by involutions and • the kernel of the index map. Moreover we have the following short exact sequence I 1 → D([T ]1 ) → [T ]1 → − Z → 1. 35 Proof. Denote by Gper the group generated by periodic elements in [Φ]1 , and by Ginv the group topologically generated by involutions. Note that both are normal subgroups of [T ]1 . By Lemma 3.10, we have Gper 6 Ginv 6 D([T ]1 ). Since the index map is continuous and takes values in an abelian group, we also have D([T ]1 ) 6 Ker (I). Let S ∈ [T ]1 . We may write S as a product of periodic elements and powers of T by the previous corollary. So modulo Gper , we have S = T n for some n ∈ Z. So I(S) = n and we conclude that the index map takes values into Z. Moreover if S ∈ Ker (I) we infer that n = 0 so that S ∈ Gper . So Gper = Ker (I) and since we have already established that Gper 6 Ginv 6 D([T ]1 ) 6 Ker (I), we conclude that all these four normal subgroups are equal. Remark. Note that for topological full groups, the kernel of the index map can be larger than the derived group due to the fact that some involutions may not belong to the derived group. Matui has obtained a complete description of the quotient of the kernel of the index map by the derived group, see [Mat06, Thm. 4.8]. 4.4 Escape time and non-simplicity results In this section, we show that the closure of the derived group of [T ]1 is never simple. To this end, we will use the non-integrability of certain escape times. Let us first recall that given a standard probability space (X, µ), its measure algebra is denoted by MAlg(X, µ) and consists of all Borel subsets of X, two such sets being identified if their symmetric difference has measure zero. It is naturally endowed with a metric dµ defined by dµ (A, B) = µ(A 4 B) which is complete and separable (see [Hal50, 40.A]). Definition 4.21. Let T ∈ Aut(X, µ) and A ∈ MAlg(X, µ). The T -escape time of A is the map τA : A → N defined by • τA (x) = +∞ if there is no n ∈ N such that T n (x) 6∈ A, • else τA (x) is the the first integer n ∈ N such that T n (x) 6∈ A or T −n (x) 6∈ A. By Kac’s theorem, for any Borel set A ⊆ X the T -return time to A is integrable and its integral is equal to 1. The situation for escape times is however very different. Indeed, any non-null T -invariant set fails to have an integrable escape time so a non ergodic transformation admits non-trivial sets with non-integrable escape time. This is actually true of any measure-preserving transformation in a strong sense as the following result shows. Theorem 4.22. Let T be a measure-preserving transformation. Then there is a dense Gδ of A ∈ MAlg(X, µ) whose T -escape time is not integrable. Proof. For n ∈ N, let Fn be the set of A ∈ MAlg(X, µ) whose escape time is either non integrable or has L1 norm stricly bigger than n. We have to show that each Fn is open dense. To see that Fn is open, first note that for a fixed k ∈ N, if we let Ak := {x ∈ A : nA (x) = k} then ! k [
Ak+1 = A \ Al ∩ T −(k+1) (A) ∪ T k+1 (A) l=1 36 so that Ak depends continuously P on A. Now take A ∈ Fn , by definition we may find N ∈ N big enough so that N k=1 kµ(Ak ) > n. Such an inequality will remain true 0 for A sufficiently close to A by continuity, so that Fn is indeed open. Next, to check the density, we will be done if we can build a sequence (Am ) such that kτAm k1 → +∞ but µ(Am ) → 0. Indeed, then A ∪ Am converges to A and for any n ∈ N, it belongs to Fn for m large enough. If the ergodic decomposition of T has a diffuse part, there is a sequence of T invariant sets (Am ) such that µ(Am ) → 0 and kτAm k1 = +∞ by T -invariance. If the ergodic decomposition of T does not have a diffuse part, then T has to be aperiodic. So we apply Rohlin’s lemma to T for N = 6m ,  = 21 to get Bm ⊆ X all whose first 6m translates are disjoint such that 2·61m 6 µ(Bn ) 6 61n . Let An := 4m −1 G T i (Bm ) i=0 be the reunion of the first 4m translates of the base Bm of the Rohlin tower. We now compute the integral of the escape time: 4m /2−1 Z τAm = 2µ(Bm ) Am = 2µ(Bm ) · 1 16m > m· 6 8 X (i + 1) i=0 m 4m 42 2 +1 2 which tends to +∞ while µ(Am ) 6 ( 46 )m tends to zero as m tends to +∞. Remark. For any T ∈ Aut(X, µ) of rank one, it is easily checked that the set of A ∈ MAlg(X, µ) whose T -escape time is bounded is also dense, but I do not know wether this is true for any T ∈ Aut(X, µ). The following corollary of Theorem 4.22 is in sharp contrast with the situation for full groups of ergodic measure-preserving equivalence relations, which act transitively on elements of MAlg(X, µ) of the same measure. Corollary 4.23. Let T ∈ Aut(X, µ) be ergodic. There is a dense Gδ of A ∈ MAlg(X, µ) such that for every B ∈ MAlg(X, µ) disjoint of A of the same measure, no element of [T ]1 maps A to B. Proof. If T maps A to a set disjoint from A, then by definition the L1 -norm of T is an upper bound of the integral of the escape time of A. By the previous theorem, this cannot happen for a dense Gδ of A ∈ MAlg(X, µ). We will now refine the preceding result to establish that for all T ∈ [R] aperiodic, neither the derived group of [T ]1 nor its closure are simple. We will need the following lemma, which is essentially contained in [Fat78]. Lemma 4.24. Let T ∈ Aut(X, µ) be aperiodic. For every A ∈ MAlg(X, µ), there exists a periodic U ∈ [T ]1 whose support is equal to A and all whose orbits have cardinality at most 3. 37 Proof. Consider the induced transformation TA , which belongs to [T ]1 by Proposition 3.5. Since T is aperiodic, supp TA = A. By a maximality argument, there is a partition (A1 , A2 , B1 , B2 , B3 ) of A such that TA (A1 ) = A2 , TA (B1 ) = B2 and TA (B2 = B3 ) (see [Fri70, Lem. 7.2]). We then have a naturally defined periodic element U ∈ [T ]1 given by  TA (x) if x ∈ A1 t B1 t B2    −1 TA (x) if x ∈ A2 U (x) = T −2 (x) if x ∈ B3    A x else. It is easily checked that U is as desired. Lemma 4.25. Let T ∈ Aut(X, µ) be aperiodic. For every Borel A ⊆ X, there exists U ∈ D([T ]1 ) whose support is equal to A and all whose orbits have cardinality at most 3. Proof. Consider the transformation U ∈ [T ]1 provided by the previous lemma. Observe that U = U2 U3 where U2 is the transformation induced by U on the set of points whose U -orbit has cardinality 2, and U3 is the transformation induced by U on the set of points whose U -orbit has cardinality 3. U3 is a commutator by the equality (1 2 3) = (1 2)(1 3)(1 2)(1 3) in the symmetric group over {1, 2, 3}. Let C be a fundamental domain of the restriction of U2 to its support. Applying the above lemma again, we find V ∈ [T ]1 all whose orbits have cardinality at most 3 such that supp V = C. We then let Ũ2 = U2 V U2 V −1 ∈ D([T ]1 ) which has the same support as U2 . Now U 0 = Ũ2 U3 is as desired. If every A ∈ MAlg(X, µ) were the support of an involution belonging to [T ]1 , one could take U from the previous lemma to be an involution. However X itself may fail to be the support of an involution (see Theorem 4.8 and the example thereafter). Theorem 4.26. Let T ∈ Aut(X, µ) be ergodic. Then neither D([T ]1 ) nor D([T ]1 ) are simple. The previous theorem follows trivially from the following statement which we will prove using Theorem 4.22. Lemma 4.27. Let T ∈ Aut(X, µ) be ergodic. Then there exists S1 , S2 ∈ D([T ]1 ) such that S1 cannot be written as a product of conjugates of S2±1 by elements of [T ]1 . Proof. Take A ∈ MAlg(X, µ) whose escape time is not integrable, which exists by Theorem 4.22. By Lemma 4.25, we may fix S1 ∈ D([T1 ]) whose support is equal to A and S2 ∈ D([T ]1 ) whose support B is disjoint from A. Now assume S1 is a product of S conjugates of S2±1 by T1 , ..., Tn ∈ [T ]1 . We must have A ⊆ ni=1 Ti (B), but then the integral of the escape time of A must be less than n X dT (Ti−1 , idX ) < +∞, i=1 a contradiction. 38 Remark. The two above results result are also true in the more general aperiodic case, but the non-trivial case is the ergodic one. Indeed, in the non-ergodic case we can directly use Lemma 4.25 to build two non-trivial elements in D([Φ]1 ) whose supports are T -invariant and disjoint, so the normal subgroups they generate actually intersect trivially. 4.5 Density of topological full groups, amenability and finiteness of the topological rank Theorem 4.28. Let T be a homeomorphism of the Cantor space, and let ν be a non-atomic ergodic probability measure preserved by T . Then the topological full group [T ]c is dense in the L1 full group [T ]1 . Proof. By Theorem 3.26, the group [T ]c contains D([Φ]1 ), in particular it contains every periodic element of [Φ]1 by Lemma 3.10. By Corollary 4.19 the group [T ]1 is generated by periodic elements along with T , and since T ∈ [T ]c we conclude that [T ]c = [T ]1 . So given a measure-preserving ergodic homeomorphism T of (2N , λ), we have the following chain of Polish groups where each is dense in its successor: [T ]c 6 [T ]1 6 [T ] 6 Aut(2N , ν). I don’t know whether the above theorem can be generalised to other finitely generated groups acting by measure-preserving homeomorphisms. Recall however that this result is wrong for graphings in general (see Example 3.27). Recall that a topological group is amenable if all its continuous actions on compact spaces admit invariant probability measures. Note that if a topological group contains a dense subgroup which is amenable as a discrete group, it has to be amenable itself. We view countable groups as discrete groups. Juschenko and Monod have shown that topological full groups of minimal Z-actions on the Cantor space are amenable [JM13]. Since the Jewett-Krieger theorem ensures us that every ergodic Z-actions can be modelled by a minimal uniquely ergodic homeomorphism (see e.g. [Pet89]), the previous theorem has the following interesting application. Theorem 4.29. Let T ∈ Aut(X, µ) be ergodic. Then its L1 full group [T ]1 is an amenable topological group, and so is the closed subgroup D([T ]1 ). It would be interesting to obtain a purely ergodic-theoretic proof of the previous result. The same is true of our next result where we rely on Matui’s work on finite generation of topological full groups (the commutator group of the topological full group of a minimal homeomorphism is finitely generated if and only if the homeomorphism is a subshift of finite type, see [Mat06, Thm. 5.4]). We should also point out that our proof of (1)⇒(2) is a natural adaptation of his. Theorem 4.30. Let T ∈ Aut(X, µ) be ergodic. Then the following are equivalent 1. the L1 full group [T ]1 is topologically finitely generated; 39 2. the closed subgroup D([T ])1 is topologically finitely generated; 3. the transformation T has finite entropy. Proof. (1)⇒(3) Suppose that [T ]1 is topologically finitely generated, and let S1 , ..., Sk ∈ [T ]1 generate a dense subgroup of [T ]1 . For each i ∈ {1, ..., k} and each n ∈ Z let Ai,n := {x ∈ X : cSi (x) = n}. Since S1 , ..., Sk ∈ [T ]1 , for every i ∈ {1, ..., k} we have X |n| µ(Ai,n ) < +∞ n∈Z Lemma 2.13 thus yields that every partition (Ai,n )n∈N has finite entropy. The partiW tion P := ki=1 (Ai,n )n∈Z thus has finite entropy. We will show that it is a generating partition. Let A be the algebra generated by the T -translates of P. Note that every element of the group generated by S1 , ..., Sk must have a cocycle partition whose elements belong to A. Since P has finite entropy, in order to conclude that T has finite entropy we only need to check that A is dense in MAlg(X, µ). For every Borel set A the transformation induced by T on A has support equal to A and can be approximated by elements of the group generated by S1 , ..., Sk in the metric d˜T which refines the uniform metric du . So the supports of these transformations form a sequence (An ) whose elements belong to Q such that µ(An 4 A) → 0. We can conclude that up to measure zero, Q generates the σ-algebra of X. As explained before, this yields that T has finite entropy. (3)⇒(2) : If T has finite entropy, Krieger’s finite generator theorem yields that we can take T to be a minimal subshift of finite type [Kri72]. It then follows from Matui’s result [Mat06, Thm. 5.4] that D([T ]c ) is finitely generated. So by Theorem 4.28 the Polish group D([T ]1 ) is topologically finitely generated. (2)⇒(1): Suppose D([T ]1 ) is topologically finitely generated. It follows from Corollary 4.20 that [T ]1 = D([T ]1 ) o hT i , so [T ]1 is also topologically finitely generated. We now give a concrete example where the topological rank is equal to 2. Proposition 4.31. The topological rank of the L1 full group of an irrational rotation is equal to 2. Proof. The topological generators that Marks built in [Mar16] to show that the full group of the ergodic hyperfinite equivalence relation is topologically 2-generated can readily be used. To be more precise, consider an irrational rotation Tα on [0, 1[ of irrational angle α ∈ [0, 1[. Let β ∈ (0, α) be irrational and consider the involution ITα ,[0,β[ as defined in the paragraph preceding Theorem 3.14. The exact same proof as Marks’ shows that the group topologically generated by Tα and ITα ,[0,β[ contains all the involutions ITα ,A where A is any Borel subset of [0, 1[ 40 such that A ∩ Tα (A) = ∅ (indeed the uniform and the L1 topology coincide on the set of involutions of the form ITα ,A so Marks’ proof readily applies). So the group topologically generated by Tα and ITα ,[0,β[ contains D([Tα ]1 ) by Theorem 3.14. But by Corollary 4.20 we have [Tα ]1 = D([T ]1 ) o hT i , so the group topologically generated by Tα and ITα ,A is equal to [Tα ]1 which has thus topological rank at most 2. Since [Tα ]1 is not abelian, we conclude that its topological rank is equal to 2. Remark. Matui’s work [Mat13, Prop. 3.2] can also be used to obtain the above result for irrational rotations of angle α ∈ (0, 1/6). 5 5.1 Further remarks and questions Non-amenability Theorem 5.1. Let Γ be a finitely generated non amenable group acting freely in a measure-preserving manner on (X, µ). Then neither [Γ]1 nor D([Γ]1 ) are amenable. Proof. Let Γ be a finitely generated non amenable group acting freely in a measurepreserving manner on (X, µ). It follows from Theorem 3.7 that [Γ]1 is dense in [RΓ ] for the uniform topology. Since Γ is infinite and acting freely, the equivalence relation RΓ is aperiodic so the group [RΓ ] is perfect by [LM14b, Prop. 3.6]. So D([Γ]1 ) is also dense in [RΓ ]. Furthermore, the group [RΓ ] is non-amenable for the uniform topology by [GP07, Prop. 5.6]. So the dense subgroups [Γ]1 and D([Γ]1 ) are not amenable for the uniform topology, in particular they are not amenable for the thinner L1 topology. It would be very interesting to understand which amenable groups admit free actions whose L1 full groups are amenable. Note that by Theorem 4.29, all the free ergodic Z-actions induce an amenable L1 full group. 5.2 Absence of a natural closure property Full groups as defined by Dye are stable under cutting and pasting a countable family of elements along a countable partition of X. Here L1 full groups are only stable under cutting and pasting a finite family of elements, and it seems natural to wonder if they can be more abstractly characterised. Here we show that L1 full groups are not “stable under L1 cutting and pasting”, showing that the most natural closure property one could hope for does not hold. Theorem 5.2. There exists an ergodic measure-preserving transformation T and a non-null Borel set A ⊆ X such that [TA ]1 66 [T ]1 . Remark. A reformulation of the theorem is that the inclusion [T ]1 A 6 [TA ]1 can be strict. In particular, given S ∈ [T ]1 , it can happen that [S]1 66 [T ]1 . 41 Proof. We fix a Borel partition (Xn )n>1 of X such that for every n ∈ N, µ(Xn ) = 2−n . 4n −1 We then cut each Xn into 4n pieces (Bn,m )m=0 of equal measure (so µ(Bn,m ) = 8−n ). Let T0 be a periodic transformation such that • for all n ∈ N and m ∈ {0, ..., 4n − 2}, one has T0 (Bn,m ) = Bn,m+1 and n • for all n ∈ N and x ∈ Bn,0 one has T04 +1 (x) = x. F Then the set B0 := n∈N Bn,0 is a fundamental domain for T0 . Let U be a measure-preserving transformation such that supp U = B0 and UB0 is ergodic, and consider T := U T0 . By Proposition 2.5, T is ergodic. Moreover for all n ∈ N and m ∈ {0, ..., 4n − 1} we have T (Bn,m ) = Bn,m+1 . Now for all n ∈ N we let An := n −1 2G Bn,2n m m=0 and then A := G An . n∈N Consider the transformation TA , and note that by construction for all n ∈ N, n all m ∈ {0, 2n − 2} and all x ∈ Bn,2n m we have TA (x) = T 2 (x). Now consider the involution Un defined by  4n /2 F n /2−1 Bn,2n m (x) if x ∈ 2m=0  T F2n −1 n /2 −4 Un (x) = T (x) if x ∈ m=2n /2 Bn,2n m  x else. TA T B2,0 B2,3 B2,7 B2,11 U3 Figure 1: The construction of U3 . The set A3 is the gray part of the figure. Note that for all x ∈ An we have dT (x, Un (x)) = 4n /2. We can then compute dT (Un , idX ) = n −1 Z 2X m=0 n Bn,2n m n 4 2 × n 8 2 1 = 2 = 42 dT (x, Un (x)) Also for all x ∈ An we have dTA (x, Un (x)) = 2n /2 so the same computation as before yields 1 2n 2n = n+1 dTA (Un , idX ) = n × 8 2 2 Let U be the involution obtained by gluing together all the Un ’s, namely for all x∈X  Un (x) if x ∈ An for some n ∈ N, U (x) = x else. We clearly have X X1 = +∞ 2 n∈N n∈N X X 1 1 and dTA (U, idX ) = dTA (Un , idX ) = = < +∞. n+1 2 2 n∈N n∈N dT (U, idX ) = dT (Un , idX ) = We conclude that U witnesses the fact that [TA ]1 66 [T ]1 . Recall that two actions of finitely generated groups Γ and Λ are L1 orbit equivalent if up to a measure-preserving transformation, they share the same orbits and for every λ ∈ Λ and γ ∈ Γ, the maps x 7→ dΓ (x, λ(x)) and x 7→ dΛ (x, γ(x)) are integrable, where dΓ and dΛ are respective Schreier graph distances (see e.g. [Aus16] for more on this notion). Here we note that although sharing the same L1 full group implies L1 orbit equivalence, the converse does not hold a priori. Corollary 5.3. There exists two measure-preserving actions of the free group on two generators F2 such that the identity map is an L1 -orbit equivalence between them but they do not share the same L1 full group. Proof. Write F2 = ha, bi, fix T ∈ Aut(X, µ) and A ⊆ X as in the proof of the previous theorem. Define a measure-preserving F2 -action α1 by α1 (a) = T and α2 (b) = T . Then define a second action α2 by α2 (a) = TA and α2 (b) = T TA−1 . Since TA ∈ [T ]1 these two actions are L1 orbit equivalent via the identity map. However a map U as in the proof of the previous theorem belongs to the L1 full group of α2 , but not to the L1 full group of α2 . Remark. Obviously L∞ orbit equivalence implies conjugacy of the L1 full groups. It would be interesting to understand wether the converse holds. 5.3 Lp full groups Let p ∈ [1, +∞]. One can then define Lp full group [Φ]p of a graphing Φ on a standard probability space (X, µ) by letting [Φ]p = {T ∈ [RΦ ] : the map x 7→ dΦ (T (x), x) belongs to Lp (X, µ)} . 43 The triangle inequality and Minkowski’s inequality yield that [Φ]p is indeed a group, and the natural distance d˜pΦ given by d˜pΦ (S, T ) = kx 7→ dΦ (S(x), T (x))kp is right-invariant, complete and refines the uniform topology. Moreover Hölder’s inequality implies that [Φ]q 6 [Φ]p whenever p < q. Note that the L∞ full group is discrete since dΦ takes values in Z. It is also uncountable hence we view it as a degenerate case. For p < +∞ however, one can show exactly like for p = 1 that [Φ]p is separable, so that it is also a Polish group. The proof that [Φ]1 is dense in [RΦ ] actually implies that [Φ]∞ is dense in [RΦ ] so that all the [Φ]p are dense in [RΦ ]. Also D([Φ]∞ ) is dense in D([Φ]p ) for all p ∈ [1, +∞[, but it is not clear wether [Φ]p is dense in [Φ]q for q < p. The main difference with L1 full groups is that for p > 1 the Lp full groups are not stable under taking induced transformations. Indeed, given a set A such that 0 < µ(A) < 1 and a map f : A → N of integral 1, one can use Lemma 2.4 to build a measure-preserving ergodic transformation T whose return time to A equates f . So if f was chosen not to be in Lp the transformation induced by T on A won’t belong to the Lp full group of T . If T is an ergodic measure-preserving transformation, the Lp full group of T is a complete invariant of flip conjugacy since it also has many involutions and is a subgroup of the L1 full group. The index map on [T ]1 restricts to its Lp full group and yields a surjective homomorphism [T ]p → Z. The kernel of this homomorphism clearly contains D([T ]p ) but it is unclear wether the converse holds, in which case we would have that [T ]p is dense in [T ]q for q < p by arguments similar of those used for Theorem 4.28. One way to establish the converse would be to obtain an analogue of Theorem 4.18. To begin with, we can still write every element of [T ]p as a a product of three elements of disjoint supports, the first being almost positive, the second almost negative and the third periodic (the proof of 4.13 readily adapts). But the problem is to understand wether almost positive elements admit a decomposition as a product of periodic elements and induced transformations belonging to [T ]p (cf. Prop. 4.14 and Prop. 4.16 for the case p = 1). Since D([T ]1 ) = Ker I, a positive answer to the following general question would also imply that D([T ]p ) = Ker I. Question 5.4. Let Φ be an ergodic graphing. Is it true that for all p ∈]1, +∞[ we have the equality D([Φ]1 ) ∩ [Φ]p = D([Φ]p ), where the closure in the left term is in the L1 topology while the closure in the right term is in the Lp topology? 5.4 Questions We close this paper by recalling some questions scattered through the text and asking a few more. The first was contained in the previous section. Question 5.5. Let 1 6 p < q 6 +∞. Is it true that [Φ]q is dense in [Φ]p ? 44 Note that a positive answer for q = +∞ yields a positive answer for every 1 6 p < q 6 +∞. When Φ is a continuous graphing, we obviously have [Φ]c 6 [Φ]∞ . We thus also ask: Question 5.6. Let Φ be a finite continuous graphing and 1 6 p < +∞. Is it true that [Φ]c is dense in [Φ]p ? Note that the answer is negative if we remove the condition that Φ is finite (cf. Example 3.27). We have seen that the L1 full group of an ergodic measurepreserving transformation T is topologically finitely generated if and only if T has finite entropy, and gave examples where the topological rank of [T ]1 is equal to 2. One could hope that there is a formula linking the topological rank to entropy as is true for cost and topological rank of full groups [LM14a], but for now we don’t even know the answer to the following very basic and pessimistic question. Question 5.7. Let T be an ergodic measure-preserving transformation of finite entropy. Is the topological rank of [T ]1 necessarily equal to 2? Using Lemma 2.1 from [Aus16], our proof that if [T ]1 is topologically finitely generated then T has finite entropy can easily generalised to show that given a measure-preserving ergodic free action of a finitely generated group Γ, if the L1 full group of the action or the closure of its derived group are topologically finitely generated then the Γ-action has a generating partition of finite entropy (i.e. the action has finite Rohlin entropy; see [Sew14] for more on Rohlin entropy). The following question is very general, but answering it for some non virtually cyclic infinite amenable group would already be very nice (in the amenable case the Rohlin entropy is equal to the usual entropy by a result of Seward and Tucker-Drob [STD14]). Question 5.8. Let Γ be a finitely generated group acting freely and ergodically on (X, µ). Suppose further that the action has finite Rohlin entropy. Is it true that [Γ]1 has finite topological rank? That D([Γ]1 ) has finite topological rank? It would also be interesting to try to generalise our structural results on L1 full groups of ergodic Z-actions to Zn -actions. For instance, given a measure-preserving ergodic Zn -action, one can still define an index map In : [Zn ]1 → Rn and ask for analogues of Corollary 4.20. Question 5.9. Does the index map In take values into Zn when the canonical generators of Zn act ergodically? Is the kernel of In equal to the closure of the derived group of [Zn ]? Our positive answer to this question for n = 1 relied crucially on the decomposition of elements of [T ]1 as products of almost positive, almost negative and periodic elements of disjoint supports, but it is not clear what an analogue of this in higher dimension would be so that another approach is probably needed. We should mention that in the case of a single ergodic measure-preserving transformation T , one can use the ideas from section 4.3 to show that the open unit ball in [T ]1 only consists of periodic elements. Since those belong to D([T ]1 ) (cf. Lemma 3.10), this yields another way to prove that D([T ]1 ) is open. We thus ask: 45 Question 5.10. Let a finitely generated group Γ act freely on (X, µ) with all its generators acting ergodically. When does the open unit ball in [Γ]1 only contain periodic elements? The fact that for an ergodic measure-preserving transformation T the open unit ball of the L1 full group only contains periodic elements can be used to show that [T ]1 has Rosendal’s local property (OB) and is (OB) generated, so that by Rosendal’s work it has a well defined quasi-isometry type as a topological group (see [Ros14]). This will be the subject of a later work. Finally, full groups of ergodic measure-preserving equivalence relations can be seen as special cases of L1 full groups (see Example 3.2). They enjoy the automatic continuity property as was shown by Kittrell and Tsankov [KT10, Thm. 1.5] and hence admit a unique Polish group topology. Question 5.11. Let Φ be an ergodic graphing. Does [Φ]1 satisfy the automatic continuity property? Is its Polish group topology unique? References [Aus16] Tim Austin. Behaviour of entropy under bounded and integrable orbit equivalence. ArXiv e-prints, 2016. [Bec98] Howard Becker. Polish group actions: dichotomies and generalized elementary embeddings. J. Amer. Math. Soc., 11(2):397–449, 1998. [Bel68] R. M. Belinskaya. Partitions of lebesgue space in trajectories defined by ergodic automorphisms. Functional Analysis and Its Applications, 2(3):190– 199, 1968. [CLM16] Alessandro Carderi and François Le Maître. Orbit full groups for locally compact groups. accepted for publication in Transactions of the AMS, 2016. [Dow11] Tomasz Downarowicz. Entropy in dynamical systems, volume 18 of New Mathematical Monographs. Cambridge University Press, Cambridge, 2011. [Dye59] Henry A. Dye. On groups of measure preserving transformation. I. Amer. J. Math., 81:119–159, 1959. [EG14] Amichai Eisenmann and Yair Glasner. Generic IRS in free groups, after Bowen. to appear in Proc. Amer. Math. Soc., 2014. [Fat78] Albert Fathi. Le groupe des transformations de [0, 1] qui préservent la mesure de Lebesgue est un groupe simple. Israel J. Math., 29(2-3):302– 308, 1978. [Fre04] D. H. Fremlin. Measure theory. Vol. 3. Torres Fremlin, Colchester, 2004. Measure algebras, Corrected second printing of the 2002 original. 46 [Fri70] Nathaniel A. Friedman. Introduction to ergodic theory. Van Nostrand Reinhold Co., New York-Toronto, Ont.-London, 1970. Van Nostrand Reinhold Mathematical Studies, No. 29. [FRW11] Matthew Foreman, Daniel J. Rudolph, and Benjamin Weiss. The conjugacy problem in ergodic theory. Ann. of Math. (2), 173(3):1529–1586, 2011. [GH09] Steven Givant and Paul Halmos. Introduction to Boolean algebras. Undergraduate Texts in Mathematics. Springer, New York, 2009. [Gla03] Eli Glasner. Ergodic theory via joinings, volume 101 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 2003. [GP07] Thierry Giordano and Vladimir Pestov. Some extremely amenable groups related to operator algebras and ergodic theory. J. Inst. Math. Jussieu, 6(2):279–315, 2007. [GPS95] Thierry Giordano, Ian F. Putnam, and Christian F. Skau. Topological orbit equivalence and C ∗ -crossed products. J. Reine Angew. Math., 469:51– 111, 1995. [GPS99] Thierry Giordano, Ian F. Putnam, and Christian F. Skau. Full groups of Cantor minimal systems. Israel J. Math., 111:285–320, 1999. [Hal50] Paul R. Halmos. Measure Theory. D. Van Nostrand Company, Inc., New York, N. Y., 1950. [JM13] Kate Juschenko and Nicolas Monod. Cantor systems, piecewise translations and simple amenable groups. Ann. Math. (2), 178(2):775–787, 2013. [Kac47] Mark Kac. On the notion of recurrence in discrete stochastic processes. Bull. Amer. Math. Soc., 53:1002–1010, 1947. [Kec10] Alexander S. Kechris. Global aspects of ergodic group actions, volume 160 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 2010. [Kri70] Wolfgang Krieger. On entropy and generators of measure-preserving transformations. Trans. Amer. Math. Soc., 149:453–464, 1970. [Kri72] Wolfgang Krieger. On unique ergodicity. In Proceedings of the Sixth Berkeley Symposium on Mathematical Statistics and Probability (Univ. California, Berkeley, Calif., 1970/1971), Vol. II: Probability theory, pages 327– 346. Univ. California Press, Berkeley, Calif., 1972. [KT10] John Kittrell and Todor Tsankov. Topological properties of full groups. Ergodic Theory Dynam. Systems, 30(2):525–545, 2010. 47 [Lev95] Gilbert Levitt. On the cost of generating an equivalence relation. Ergodic Theory Dynam. Systems, 15(6):1173–1181, 1995. [LM14a] François Le Maître. The number of topological generators for full groups of ergodic equivalence relations. Invent. Math., 198:261–268, 2014. [LM14b] François Le Maître. Sur les groupes pleins préservant une mesure de probabilité. PhD thesis, ENS Lyon, 2014. 2014ENSL0892. [LM15] François Le Maître. On full groups of non-ergodic probability measure preserving equivalence relations. to appear in Ergodic Theory Dyn. Syst., 2015. [Mar16] Andrew Marks. Topological generators for full groups of hyperfinite pmp equivalence relations. preprint, 2016. [Mat06] Hiroki Matui. Some remarks on topological full groups of Cantor minimal systems. Internat. J. Math., 17(2):231–251, 2006. [Mat13] Hiroki Matui. Some remarks on topological full groups of Cantor minimal systems II. Ergodic Theory Dynam. Systems, 33(5):1542–1549, 2013. [Med07] Konstantin Medynets. On approximation of homeomorphisms of a Cantor set. Fund. Math., 194(1):1–13, 2007. [Mil05] Benjamin David Miller. Density of topological full groups. Unpublished note available at https: // dl. dropboxusercontent. com/ u/ 47430894/ Web/ notes/ density. pdf , 2005. [Pet89] Karl Petersen. Ergodic theory, volume 2 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 1989. Corrected reprint of the 1983 original. [Ros14] Christian Rosendal. Coarse geometry of topological groups. preprint, 2014. [Sew14] Brandon Seward. Krieger’s finite generator theorem for ergodic actions of countable groups i. ArXiv e-prints, 2014. [STD14] Brandon Seward and Robin D. Tucker-Drob. Borel structurability on the 2-shift of a countable group. ArXiv e-prints, 2014. 48 

4math.GR