Sampling from the Gibbs Distribution in Congestion Games

Published Online:https://doi.org/10.1287/moor.2022.1322

Abstract

Logit dynamics is a form of randomized game dynamics in which players have a bias toward strategic deviations that give a higher improvement in cost. It is used extensively in practice. In congestion (or potential) games, the dynamics converge to the so-called Gibbs distribution over the set of all strategy profiles when interpreted as a Markov chain. In general, logit dynamics can converge slowly to the Gibbs distribution, but beyond that, not much is known about its algorithmic aspects, nor that of the Gibbs distribution. In this work, we are interested in the following two questions for congestion games: (i) Is there an efficient algorithm for sampling from the Gibbs distribution? (ii) If yes, does there also exist natural randomized dynamics that converge quickly to the Gibbs distribution? We first study these questions in extension parallel congestion games, a well-studied special case of symmetric network congestion games. As our main result, we show that there is a simple variation on the logit dynamics that converges quickly to the Gibbs distribution in such games. We also address the first question for the class of so-called capacitated uniform congestion games and the second question for max cut games played on a complete graph. To prove our results, we rely on the recent breakthrough work of Anari et al. (2019) regarding the approximate sampling of a base of a matroid according to a strongly log-concave probability distribution.

1. Introduction

Congestion games constitute a rich class of games that have been studied extensively since their introduction by Rosenthal [55]. An (unweighted) congestion game Γ=(N,E,(Si)iN,(ce)eE) consists of a set of players N={1,,n} and a set of resources E={1,,m}. Every player i has a strategy set Si2E, in which each strategy is a subset of resources. Furthermore, every resource eE is equipped with a cost function ce:R0R that we assume to be nonnegative and nondecreasing. The goal of a player is to choose a strategy that minimizes the player’s total cost Ci(s)=esice(e(s)), where e(s) is the number of players using resource e in profile s×iSi=S. A well-known example is the class of symmetric network congestion games, in which we are given a directed graph G=(V,E) with origin oV and destination dV. The common strategy set of all players is given by the set of all o, d-paths in G.

Rosenthal [55] proves that congestion games are (exact) potential games. He shows that the function Φ:×iSiR given by

Φ(s)=eEk=1e(s)ce(k),
satisfies for every s×iSi and siSi the equality
Ci(s)Ci(si,si)=Φ(s)Φ(si,si).(1)

Here, Ci(si,si) is used to denote the cost of player i in the strategy profile in which i chooses si, and all other players choose their strategy in s. The function Φ is often referred to as Rosenthal’s potential. The main implication of (1) is the existence of a pure Nash equilibrium (PNE): a strategy profile in which no player can deviate to another strategy and obtain an improved cost (Rosenthal [55]). This follows directly from the observation that better (or best) response dynamics converge to a PNE in a finite number of steps. Better response dynamics is defined as the procedure by which, in every step, precisely one player deviates to another strategy that yields an improved cost (until a pure Nash equilibrium is reached). For best response dynamics, the deviating player always deviates to a strategy that yields the greatest possible improvement in cost.

In the last two decades, the algorithmic aspects of (pure) Nash equilibria are studied extensively in both general and special classes of congestion games. Two of the most prominent questions concerning pure Nash equilibria are the following.

  1. Does (natural) player dynamics, such as better or best response dynamics, converge to a PNE in polynomial time?

  2. If not, can one compute a PNE in polynomial time by other means?

Ieong et al. [38] consider the convergence time of better and best response dynamics for singleton congestion games. Here, every strategy of a player consists of a single resource. It is shown that better response dynamics converge in at most O(n2m) steps, and there exist instances in which convergence of better response dynamics might take Ω(n32m) steps (Ieong et al. [38]). The result of Ieong et al. [38] is generalized by Ackermann et al. [2] to matroid congestion games, in which the strategy set of every player is the collection of bases of a matroid on the ground set E (see Section 2.3 for a formal definition). They show that best response dynamics converge in at most O(n2mr) steps, where r is the maximum rank of the matroids that form the strategy sets of the players. They also show that matroids are, in a sense, the maximal structure that allow for polynomial time convergence of best response dynamics. Furthermore, Ackermann et al. [2] show best response dynamics are not guaranteed to converge in symmetric network congestion games, in which the common strategy set of every player is the collection of paths in a given directed graph. Best response dynamics are also studied in more general congestion games, such as with player-specific cost functions (Ackermann and Röglin [1]) or with weighted players (Even-Dar et al. [24], Fanelli and Moscardelli [26]). Approximate versions of response dynamics are also studied; see, for example, Chien and Sinclair [19] and Skopalik and Vöcking [59].

The complexity of computing a pure Nash equilibrium is also well-understood. Fabrikant et al. [25] show that computing a pure Nash equilibrium is complete for the complexity class polynomial local search. Hardness of computing a PNE is also shown for various special cases of congestion games; see, for example, Ackermann et al. [2] and Del Pia et al. [21], as well as for the computation of approximate equilibria (Skopalik and Vöcking [59]). On the positive side, Fabrikant et al. [25] show that a PNE can be computed efficiently in symmetric network congestion games despite the fact that best response dynamics are not guaranteed to converge in these games (Ackermann et al. [2]). Del Pia et al. [21] generalize this result to totally unimodular congestion games; see also Kleer and Schäfer [43] for a more general framework.

In general, player dynamics roughly comes in two flavors: one deviates to another strategy profile according to either a deterministic rule or a probabilistic one. A well-known example of the latter case is noisy (randomized) best response dynamics, which has received a lot of attention in practice; see, for example, Camerer [18] and Mäs and Nax [47], but seems hard to analyze from a theoretical perspective. Here, instead of making a deviation to another strategy according to a deterministic rule, a player chooses a strategy from the player’s set according to a probability distribution that usually puts relatively more weight on strategies that result in a lower cost. Randomized dynamics can be studied from two perspectives: as either a randomized alternative for deterministic dynamics converging to a pure Nash equilibrium or as a dynamical system on its own.

One well-known example of player dynamics that can be studied as a dynamical system and which is the topic of this paper is logit dynamics. It has received a lot of attention in various communities, such as evolutionary game theory (e.g., Sandholm [56]) and experimental economics (e.g., Camerer [18]). The procedure was introduced by Blume [14] as a form of randomized game dynamics in which players update their strategy according to a logit update rule (McFadden [48]). The logit dynamics for congestion games can be formulated as follows. For a given strategy profile s and fixed rationality level (or inverse temperature in the physics literature) parameter T0, first choose a player iN uniformly at random and then have player i choose a strategy siSi with probability

eTΦ(si,si)tSieTΦ(t,si).(2)

Note that the denominator in (2) is a normalizing constant (often called the partition function).

Remark 1.

Equivalently, one can replace the Φ(si,si) by Ci(si,si) and, similarly, in the normalizing constant. We use Φ as this is more convenient for our purposes. The equivalence follows from (1).

The rationality level T0 is used to model the amount of noise players believe there to be in the system (Auletta et al. [12]). When T, players effectively only assign positive probability to best responses, whereas when T0, the distribution in (2) approaches the uniform distribution over Si. Also note that the dynamics indeed puts relatively more probability mass on strategies that give a greater improvement in cost.

Logit dynamics gives rise to an ergodic, time-reversible Markov chain on the set S of all strategy profiles that has the Gibbs distribution π given by

π(s)=eTΦ(s)t×iSieTΦ(t)
for sS as its unique stationary distribution. This simply means that, if one runs the logit dynamics for a sufficiently long time, the distribution over S converges to the Gibbs distribution. The time it takes to converge to the stationary distribution is called the mixing time of the Markov chain.

Auletta et al. [12] interpret the Gibbs distribution as a dynamic equilibrium concept, which they dubbed the logit equilibrium; see also Ferraioli [28]. This concept is well-defined for general finite games; see, for example, Auletta et al. [12]. The goal of this work is to study algorithmic aspects of the logit equilibrium/Gibbs distribution in congestion games. Auletta et al. [12] give a tight bound for the mixing time of the logit dynamics in general potential games, which can be exponential in the quantity TΦmax (as in Example 1). They also study graphical coordination games, for which they identify various types of behavior of the logit dynamics; we refer to Auletta et al. [12] for more details. Because of the general slow mixing, Auletta et al. [10] also study the concept of metastability, which means a Markov chain stays close to some distribution on a timescale shorter than the mixing time. More work on the logit dynamics in potential games by these authors can also be found in Auletta et al. [9, 11]. We also refer to the survey article of Ferraioli [28] and references therein. There are also various results addressing the inefficiency of “long-term” equilibria in the context of logit dynamics; see, for example, the works of Asadpour and Saberi [8], Mamageishvili and Penna [45], and Penna [54].

A special case of potential games for which logit-like dynamics is studied extensively is the Glauber dynamics for the Ising model, which, in game-theoretical terms, can be seen as logit dynamics for max-2-cut games that are described in Section 1.1. Whether the logit dynamics are rapidly mixing in this case depends on the parameter T0 and graph topology G; see, for example, the work of Levin et al. [44] and references therein. Jerrum and Sinclair [39] show that, nevertheless, there exists a polynomial time algorithm to sample from the Gibbs distribution for any rationality level T0 and any graph topology. This work is partially motivated by studying a similar question for special cases of congestion games as explained later.

As said before, in general, the logit dynamics converge slowly to the Gibbs distribution (Auletta et al. [12]); in particular, the number of steps needed might be Ω(eTΦmax), where Φmax is the maximum value attained by Rosenthal’s potential. We next give a simple example illustrating this fact.

Example 1.

Consider the congestion game with players N={1,2} and two resources E={a,b}. Both resources e{a,b} can be used by both players, that is, S1=S2={{a},{b}}, and have a cost function satisfying ce(1)=0 and ce(2)=ϕ for some ϕ0. Note that Φmax=ϕ.

We assume that T is fixed in this example, in particular, independent of ϕ. If ϕ is large, the Gibbs distribution assigns weight close to 1/2 to the strategy profiles (s1,s2){(a,b),(b,a)} and weight close to (1/2)eTϕ0 to (s1,s2){(a,a),(b,b)}. We summarize these facts in Figure 1. Now, informally speaking, if we consider the logit dynamics with starting profile (a, b), then in order to reach the profile (b, a), we have to go through either the profile (a, a) or (b, b). The probability of transitioning to one of those profiles in one step of the logit dynamics is O(eϕT). As both profiles (a, b) and (b, a) appear with probability close to 1/2 in the Gibbs distribution, the probability of being in either of these profiles after a (large) number of steps of the logit dynamics should be approximately equal. However, escaping from the profile (a, b) happens with probability O(eϕT). The inverse Ω(eTϕ) of this probability then forms a lower bound on the number of steps of the logit dynamics needed to get close to the stationary Gibbs distribution.

What is causing the slow convergence in Example 1? The problem is that we need to use either the profile (a, a) or (b, b), both having very small probability in the Gibbs distribution, to move from (a, b) to (b, a). However, as it turns out, if one in addition with some probability is allowed to interchange the strategies of players 1 and 2, then the resulting dynamics converges quickly to the Gibbs distribution (being a special case of Theorem 1). Note that this enables the possibility to directly transition between (a, b) and (b, a). This motivates the following question: Is there a (simple) Markov chain on S that converges rapidly to the Gibbs distribution over S at any rationality level T?

This question may be interpreted as the natural analogue of looking for other local search procedures converging quickly to a PNE when best/better response dynamics does not have this property.

When the answer to the preceding question is not directly obvious, one can take another step back and first ask whether it is at all possible to efficiently sample from the Gibbs distribution. Informally speaking, can we take “snapshots” (according to the Gibbs distribution) from the system in equilibrium in polynomial time? More formally speaking, does there exist an efficient algorithm to sample (approximately) a strategy profile sS according to the Gibbs distribution over S at any rationality level T? This question can be interpreted as a dynamic analogue of the second question posed earlier for the computation of pure Nash equilibria. That is, although (deterministic) better/best response dynamics might take a long time to converge to a pure Nash equilibrium, one still wants to know whether a PNE can be computed efficiently by other means. Similarly, although logit (or other natural) dynamics can take a long time to converge to the Gibbs distribution, we may still ask whether, by means of sampling, we can get an impression of what the Gibbs distribution over S looks like. Note that a positive answer to the first question posed earlier gives a positive answer to the second question as one can simply run the Markov chain sufficiently long in order to generate a sample from the Gibbs distribution.

These questions are made precise in Section 2. We remark that one of the aspects that makes them nontrivial is the fact that we require the questions to hold for any rationality level T, that is, T is considered part of the input. For example, in Example 1, we could set T=Θ(1/ϕ) to circumvent the problem arising there.

Figure 1. Overview of the four possible strategy profiles in Example 1.

1.1. Our Contributions

In this work, we address the questions posed for various special cases of congestion games.

1.1.1. Extension Parallel Congestion Games.

The class of extension parallel (EP) congestion games is a well-studied special case of symmetric network congestion games; see, for example, Fotakis [29], Fujishige et al. [30], and Holzman and Law-Yone [37]. Here, the common strategy set of all players is given by the set of o, d-paths P of an extension parallel graph; see Section 4 for a definition and example. Our main result is that there is a simple Markov chain converging quickly to the Gibbs distribution over S, also implying that we can sample approximately from the Gibbs distribution. We show that, if one, in addition to the logit dynamics transitions, is allowed to randomly interchange the strategies of two players (akin to the explanation given after Example 1), the resulting Markov chain converges quickly to the Gibbs distribution. We call this the relaxed logit dynamics; see Section 4.2 for a formal definition. Note that Theorem 1 gives a doubly exponential improvement with respect to the dependence on TΦmax compared with the lower bound as given in Example 1.

Theorem 1

(Informal). The relaxed logit dynamics for EP congestion games, at rationality level T, converges to a distribution ϵ-close to the Gibbs distribution in at most

n3(logn+loglog|P|+log(2TΦmaxϵ2))
steps, where n is the number of players, P the number of paths in the EP graph, and Φmax the maximum value attained by Rosenthal’s potential.1

The notion of “ϵ-close” refers to the fact that the distribution seen after the indicated number of steps differs from the Gibbs distribution at most ϵ in total variation distance (see Section 2.5), a well-known distance measure for comparing probability distributions in Markov chain theory.

In a nutshell, Theorem 1 follows from the fact that, in EP congestion games, Rosenthal’s potential is M-convex as is shown by Fujishige et al. [30]. M-convexity is a property defined in the area of discrete convex analysis (Murota [51]) (see Section 2.3).2 The link between M-convexity and sampling is, roughly speaking, established in a series of (breakthrough) papers by Anari et al. [3, 5, 6] through the theory of strongly log-concave (SLC) polynomials. The theory of strongly log-concave (or Lorentzian) polynomials dates back to the work of Gurvits [34] and is further developed by Anariet al. [3] and Brändén and Huh [17]. In particular, Anari et al. [5] give the first polynomial-time algorithm for approximately sampling and counting the number of bases of a given matroid, resolving also an old conjecture by Mihail and Vazirani [50]. In this work, we rely on the sampling result from Anari et al. [5], albeit for relatively simple matroid structures.

Before proving the preceding result, we also give another way of sampling from the Gibbs distribution in Section 4.1 that essentially is a more direct approach than the sampler induced by the Markov chain result given (Theorem 4). The high-level approach used for this more direct sampler is given in Section 3 (and also used in Section 6). Finally, we give an application of our results to the problem of (approximate) uniformly sampling pure Nash equilibria in EP congestion games in Section 4.3, which is, to the best of our knowledge, the first of its kind.

1.1.2. Max-k-Cut Game on a (Unweighted) Complete Graph.

A max-k-cut game (see, e.g., Gourvès and Monnot [33]) is given by an undirected graph G=(V,E) whose nodes V are the players. Every vV has strategy set Sv={1,,k} whose elements are referred to as colors. The utility of a player is the number of neighbors that choose a different color. They can be modeled as a special case of extension parallel congestion games when G is the complete graph. This is explained in Section 5.

As an application of Theorem 1, we show that the relaxed logit dynamics is rapidly mixing at any rationality level T for max-k-cut games when G is the complete graph (Corollary 2). This stands in stark contrast to known results regarding the logit dynamics on complete graphs for the case k = 2, whose mixing time depends heavily on the value of the rationality level T (Levin et al. [44]). We elaborate on this in Section 5.

1.1.3. Capacitated Uniform Congestion Games.

Finally, we study the class of so-called u-capacitated k-uniform congestion games for given k=(k1,,kn) and u=(u1,,um). In such a game, the strategy set of player iN is given by all subsets of E of size ki. Furthermore, for every eE, we are given a capacity ue so that ce(x)= whenever x>ue.

The motivation for studying these games comes from the class of base-matroid congestion games, in which the strategy set of every player is the set of bases Bi of a given matroid Mi over the ground set of resources E. It is well-known that best response dynamics converge to a PNE in a polynomial number of steps in this class of games (Anari et al. [2]), and so, in particular, a PNE can be computed in polynomial time. Given the base-matroid sampling result of Anari et al. [5], a natural question that comes to mind is if a similar result exists for sampling from the Gibbs distribution in base-matroid congestion games. Here, we give a first result addressing this question. (Note that, in our setting, the strategies of player iN are the bases of the ki-uniform matroid.)

We next explain why there is a need for capacity constraints in our results. A strategy profile s can be seen as a bipartite graph in which the nodes on one side correspond to the players, having degrees ki, and the nodes on the other side correspond to the resources, having degrees e(s) (the resource load on e in profile s). For a given profile s with resource load profile (s), the bipartite graph is obtained in the natural way: there is an edge between player i and resource e if player i uses resource e in profile s. Very roughly speaking, in order to apply the sampling result in Anari et al. [5], we establish strong log-concavity of a certain polynomial associated to the vectors k and u. For this, we rely on an asymptotic enumeration formula for the number of bipartite graphs with a given degree sequence, which, in terms of congestion games, gives the number of strategy profiles that have a given resource load profile . (Asymptotic enumeration of graphs with given degrees is studied extensively in the area of combinatorics.) The formula that we use is only valid for the range of k=(k1,,kn) and resource load profiles (s)=(e(s)) satisfying the imposed capacity constraints as given in Theorem 2. Our main result is as follows.

Theorem 2

(Informal). There is an (almost) polynomial time algorithm for approximately sampling from the Gibbs distribution in u-capacitated k-uniform congestion games assuming that 1kmaxumax=o(U1/4) when n, where kmax=maxiki,umax=maxjuj and U=juj.

See Remark 6 for an explanation of what we mean by “almost” polynomial time. The proof of Theorem 2 reveals an interesting connection between M-convexity and asymptotic enumeration formulas that might be of independent interest.

1.2. Technical Approach

In this section, we give an intuitive outline of the approach taken in order to prove the results as summarized in Section 1.1. For the sake of simplicity, we work with symmetric singleton congestion games in which the common strategy set contains every individual resource as a strategy, that is, Si={{e1},{e2},,{em}} for all iN. These games form a special case of EP and capacitated uniform congestion games. All technical notions not formally defined here can be found in Section 2.

Let us recall Example 1, which is a symmetric singleton congestion game with two resources. The cause of the long convergence time in this example is the fact that transitioning between strategy profiles in which both resources have the same load profile, that is, the number of players using the resource, is difficult. To be precise, in Example 1, both profiles (a, b) and (b, a) have the same load profile, but they correspond to different strategy profiles.

In order to still be able to sample from the Gibbs distribution, we take two approaches:

  1. We first sample a load profile according to the correct probability and then sample a strategy profile corresponding to that load profile (Algorithm 1 in Section 3).

  2. We augment the logit dynamics with additional transition possibilities, making it easier to transition between strategy profiles with the same load profile.

In the first approach, the “correct probability” with which a given load profile should be sampled is the sum of all the probabilities assigned to strategy profiles with the given profile. The resulting distribution is called the Gibbs distribution induced on the load profiles. For the applications in Sections 46, sampling a strategy profile with a given load profile can be done by either generating a random permutation of the players or relying on known sampling results in the literature. In particular, in the case of capacitated uniform congestion games, sampling a strategy profile for a given load profile corresponds to sampling a bipartite graph with a given degree sequence.

The main technical challenge lies in sampling a load profile with the correct probability. In the case of symmetric singleton congestion games, the collection of all load profiles can be modeled by the set R={x0m:x1+x2++xm=n}, where xe denotes the load on resource eE. The set R is a special case of a discrete polymatroid. In order to sample a (resource) load vector x from R with the correct probability, we consider a Markov chain that can transition between load vectors x and y for which i|xiyi|=2; that is, x and y differ in one unit of load on precisely two resources (but have the same total load). The transition probabilities are chosen in such a way that the stationary distribution of the Markov chain is precisely the Gibbs distribution induced on the load profiles.

The Markov chain that we consider over the set R is a “polymatroid version” of the base-exchange Markov chain studied by Anari et al. [5] for sampling a base of a given matroid. The state space of this base-exchange Markov chain is the set of all bases of a given matroid. It randomly exchanges some element of the current base with another element not in the base, resulting in another base of the matroid. This idea relies on the well-known base-exchange property of matroids (Schrijver [57]). In their breakthrough work Anari et al. [5] show this Markov chain to be rapidly mixing when the stationary distribution is strongly log-concave. Strong log-concavity is a property of a polynomial associated with the stationary distribution of a Markov chain (defined for both discrete polymatroids and matroids). We apply the result from Anari et al. [5] by giving a reduction from sampling a base of a polymatroid, that is, a load vector from R, to the problem of sampling a base of a matroid using a well-known construction of Helgason [35]. For this reduction, we rely on some properties that preserve strong log-concavity (Brändén and Huh [17]). Whereas this reduction works for arbitrary discrete polymatroids, in our applications, we use relatively simple polymatroid structures. What is left then is to show that the Gibbs distribution induced on the load profiles indeed satisfies the notion of strong log-concavity. Based on results in Brändén and Huh [17], it turns out that, for symmetric singleton congestion games, this is the case because of the fact that Rosenthal’s potential is a separable convex function.

In order to obtain our sampler for EP congestion games (Theorem 4), we use a similar approach based on load profiles, but instead, we rely on the M-convexity of Rosenthal’s potential for such games (Fujishige et al. [30]). For Theorem 2, we also use a similar approach.

In order to show that the logit dynamics converge quickly to the Gibbs distribution when we additionally allow the interchanging of the strategy of two players, we use a Markov chain decomposition argument. This approach results in the proofs of Theorem 5 and Corollary 2. The idea of Markov chain decomposition is to divide the state space into smaller sets. If one can prove that the Markov chain converges quickly to its (induced) stationary distribution on the smaller sets and that it is easy to move between these smaller sets, then the original chain converges quickly to its stationary distribution as well. We elaborate more on this approach before giving the proof of Theorem 5.

1.3. Discussion and Further Related Work

To the best of our knowledge, ours are the first (polynomial-time) sampling results for the Gibbs distribution in congestion games beyond the well-studied case of max-cut games on general graphs, also known as the Glauber dynamics for the Ising model. Given the extensive attention that logit dynamics has received in various communities as well as the topic of approximate sampling in theoretical computer science, we believe this to be an interesting line of work at the intersection of algorithmic game theory, combinatorics, and approximate sampling to pursue further.

In particular, for special cases of congestion games with a positive answer to questions 1 and 2 posed in the introduction, do there also exist positive answers for their dynamic analogues? As a concrete open question, we ask whether it is always possible to efficiently sample from the Gibbs distribution in general base matroid congestion games (Ackermann et al. [2]). If true, this provides an interesting (qualitative) game-theoretical generalization of the sampling result of Anari et al. [5].

We conclude this section with some more background and related work concerning the result of Anari et al. [5]. A long-standing open question (Mihail and Vazirani [50]) in the area of approximate sampling and counting asks if there exists a rapidly mixing Markov chain to sample approximate uniformly at random a base of a given matroid (which, in turn, also yields a polynomial-time approximation scheme for counting the number of bases of a given matroid).3 Anari et al. [5] show that the base-exchange Markov chain is indeed rapidly mixing. In fact, they prove the more general result that this chain is rapidly mixing, using appropriate transition probabilities, for any strongly log-concave stationary distribution; see Section 2.4. An (up to constant factors) tight mixing time bound for the base-exchange Markov chain is later given by Anari et al. [6]. Before Anari et al. [5], rapid mixing of the base-exchange Markov chain (with uniform stationary distribution) was only known for special cases. In particular, Feder and Mihail [27] prove rapid mixing for balanced matroids, using Sinclair’s [58] multicommodity flow method. Examples of balanced matroids are the graphic and regular matroids.

2. Preliminaries

In this section, we give all the necessary preliminaries regarding congestion games, strongly log-concave polynomials, and the relevant Markov chain notions and results. We start with some general notation.

All logarithms in this work have Euler’s number e as their base unless specified otherwise. For k>0, we write [k]={1,,k}. For two vectors x,yn, we write xy if xiyi for i=1,,n, and x < y if strict inequality holds for at least one i. Furthermore, with |x|=i=1n|xi|, we denote the modulus of x. We use (ei)i=1,,n to denote the standard basis of Rn, that is, ei(j)=1 if j = I and ei(j)=0 otherwise. For sets A, B, and C, we use the notation AB+C to denote the set (A\B)C.

2.1. Congestion Games

A capacitated congestion game Γ is given by a tuple (N,E,(Si)iN,(ce)eE,(ue)eE), where N=[n] is a finite set of players, E=[m] a finite set of resources (or facilities), Si2E the set of strategies of player iN, and ce:0 the cost function of resource eE that satisfies ce(x)=W whenever x>ue for eE with W a sufficiently large number. Unless stated otherwise, the cost functions are assumed to be nonnegative and nondecreasing. Finally, ue is a nonnegative integer modeling the capacity on resource eE. If ue = n for every resource eE, we simply call Γ a congestion game. (Extension parallel and capacitated uniform congestion games are defined in Sections 4 and 6, respectively.) For a strategy profile s=(s1,,sn)×iSi=S, we define e(s) to be the number of players using resource e, that is, e(s)=|{iN:esi}|. A game is called symmetric if Si=Sj for all i,j[n]. We then write P to denote the common strategy set of all players.

We call (s)=(e(s))eE the resource load profile corresponding to strategy profile s. We say that a strategy sS is feasible if e(s)ue for every eE and write Sf to denote the set of all feasible strategy profiles. We only consider games in which the set of feasible strategy profiles is nonempty. More generally, we say that yNm is a (feasible) resource load profile for (N,E,(Si)iN,(ue)eE) if there is some (feasible) strategy profile s such that y=(s). We write S(y) for the set of strategy profiles s×iSi whose resource load profile is y.

Similarly, for symmetric congestion games, we define the notion of a strategy load profile that models how many players are using a strategy pP in a given strategy profile sPn. More precisely, given a strategy profile sPn, we define zp(s)=|{iN:si=p}| as the number of players choosing strategy pP in strategy profile s. The vector z(s)=(zp(s))pP is called the strategy load profile of s. Similarly as for resource load profiles, we define (with a slight abuse of notation) S(x) to be the set of strategy profiles s for which x=z(s).

The cost of player iN under a strategy profile s=(s1,,sn)×iSi is given by Ci(s)=esice(e(s)). A strategy profile sS is called a (pure) Nash equilibrium if, for every iN and every siSi, it holds that Ci(s)Ci(si,si), where (si,si) denotes the strategy profile in which player i plays si and every other player ji plays sj. We write NE(Γ) to denote the set of all pure Nash equilibria of Γ.

We say that Φ:×iSiR is an exact potential function for a congestion game Γ if, for every strategy profile s×iSi, for every player iN, and every unilateral deviation siSi of i it holds that Φ(s)Φ(si,si)=Ci(s)Ci(si,si). Rosenthal [55] shows that

Φ(s)=eEk=1e(s)ce(k)(3)
is an exact potential function for any congestion game. Subsequently, we refer to this potential function as Rosenthal’s potential.

A function ϕ:{0,,n}R is called convex if ϕ(i)ϕ(i1)ϕ(i+1)ϕ(i) for all i=1,,n1. A function ψ:{0,,n}mR is called separable convex if it is of the form (x1,,xm)j=1mψj(xj), where the ψj, given by xjψj(xj), are convex. We say that ϕ is concave if ϕ is convex, and similarly, ψ is separable concave if ψ is separable convex. A simple but important observation that we use in this work is the fact that Rosenthal’s potential is a separable convex function when seen as a function from resource load profiles to the reals. That is, the function Φ˜:0mR given by

Φ˜(β)=eEk=1βece(k)(4)
for β0m is separable convex.

Proposition 1.

If the cost functions (ce)eE are nondecreasing, then Rosenthal’s potential Φ˜ is a separable convex function.

2.2. Gibbs Distribution and Logit Dynamics

The Gibbs distribution π:SR0 over the strategy profiles of a congestion game Γ is given by

π(s)=eTΦ(s)Z,
where Φ is Rosenthal’s potential, T0 rationality level, and Z is the normalizing constant (or partition function)
Z=tSfeTΦ(t).

The logit dynamics Markov chain (formal Markov chain definitions are given in Section 2.5) with current state sS proceeds as

  • Select a player iN uniformly at random.

  • For siSi, transition to (si,si) with probability eTΦ(si,si)/Zi with normalizing constant

    Zi=rSieTΦ(r,si).

It is a standard fact, which can be shown by using (1), that this Markov chain is reversible with respect to the Gibbs distribution.

2.3. Matroids and M-Concavity

Let E=[n] be a finite set called the ground set and I2E={X:XE} a collection of subsets of E (called independent sets). The pair N=(E,I) is a matroid if (i) I; (ii) AI and BA, then BI; (iii) A,BI and |A|>|B|, then there exists an aA\B such that B+aI. An independent set BI of maximum size is called a basis. We use B to denote the set of all bases of N. The set of bases B satisfies the so-called base-exchange property: if B,BB and eB\B, then there exists an eB\B such that B+eeB. It satisfies the strong base-exchange property if both B+ee,Be+eB. The rank of a matroid is the common cardinality r of all bases in B. The -truncation M=(E,I) of a matroid M is the matroid with AI if and only if AI and |A|. The partition matroid is given by a disjoint partition E=E1Eq of the ground set E and upper bounds ui for i=1,,q. A set AE is independent if and only if |AEi|ui for all i=1,,q. The k-uniform matroid is the matroid in which AE is independent if and only if |A|k.

A discrete polymatroid is a finite set of vectors R0n with the properties that (i) 0R; (ii) if yR and xy, then xR; and (iii) if x,yR with |y|>|x|, then there is a vector wR such that x<w<max{x,y} (in which the maximum is taken coordinate-wise). The set of bases BR is given by all maximal vectors in R that have a common modulus r. A polymatroid satisfies the base-exchange property; if x,yBR and xi > yi, then there exists an index j with yj > xj and xei+ejBR.

Finally, as a generalization of discrete polymatroids, we describe the notion of M-convexity for functions Murota [51, 53]. As we mostly work with its negated counterpart of M-concavity, we describe this first. Let ν:0nR{} be a function. The effective domain of ν is given by

dom(ν)={α0n:v(α)>}.

The function ν is called M-concave if it satisfies the (symmetric) exchange property: for any α,βdom(ν) and any i[n] satisfying αi>βi, there exists a j[n] such that αj<βj and

ν(α)+ν(β)ν(αei+ej)+ν(β+eiej).(5)

It is well-known that a separable concave function is M-concave (Murota [52]). The function ν is called M-concave if it is M-concave and, in addition, there is an r0 such that dom(ν){α:iαi=r}. A function ν:0nR{} is called M-convex if ν is M-concave.

2.4. Strongly Log-Concave Polynomials

We consider polynomials pR[x1,,xn] with nonnegative real coefficients. For a vector β=(β1,,βn)0n, we write

β=i=1nxiβi
to denote the partial differential operator that differentiates a function βi times with respect to xi for i=1,,n. For α0n, we write xα to denote i=1nxiαi. Furthermore, we write α!=iαi!, and for α,κ0n with αiκi for all i, we write
(κα)=i=1n(κiαi).

For a constant t0 with tmaxiαi, we write (tα)=i=1n(tαi). Let κ0n and K=×i{0,,κi}. Let w:KR0 be a weight function. The generating polynomial of w is given by

gκ(x)=αKw(α)·xα.

The support of gκ is the set supp(gκ)={αK:w(α)>0}. The generating polynomial g is called d-homogeneous if |α|=iαi=d for all α{0,,k}m with w(α)>0. It is called multiaffine if every variable xi appears with at most multiplicity one in every monomial of p. For example, q(x1,x2)=x1x2 is multiaffine, but r(x1,x2)=x12+x1x2 is not as the multiplicity of x1 in the first monomial is two. Finally, the elementary symmetric polynomial of degree d, for κ=(1,1,,1), is given by

hκ(x)=α{0,1}n:|α|=dxα.

Definition 1

(Strong Log-Concavity; Gurvits [34]). A polynomial pR[x1,,xn] with nonnegative coefficients is called log-concave on a subset SR0n if its Hessian 2log(p) is negative semidefinite on S. A polynomial p is called SLC on S if, for any β0n, we have that βp is log-concave.

For convenience, the zero polynomial is defined to be strongly log-concave always. It is interesting to note that if a d-homogeneous multiaffine polynomial p is SLC, then the support of p must form the collection of bases of a matroid, and more generally, if a (not multiaffine) homogeneous polynomial is SLC, its support forms an M-convex set (Brändén and Huh [17]). Finally, if the generating polynomial gκ is strongly log-concave, then the probability distribution π(α)w(α) is called strongly log-concave.

Remark 2.

The definition of strong log-concavity is in this work is not really needed but included for completeness. In our proofs, we essentially only rely on properties of SLC polynomials from the literature that are reviewed as follows. For homogeneous generating polynomials the notion of strong log-concavity is equivalent to that of a polynomial being Lorentzian (Brändén and Huh [17]) or completely log-concave (Anari et al. [4]). These equivalences are shown in Brändén and Huh [16].

We next state all properties of SLC polynomials that are used in this work. First, it is easy to check that the SLC property is preserved under multiplication with a nonnegative scalar, which we state here for sake of reference.

Proposition 2

(Brändén and Huh [17]). If pR[x1,,xn] is SLC and γR0, then γ·p is SLC.

We continue with the polarization operator defined by Brändén and Huh [17]. The polarization operator introduces auxiliary variables in order to turn p into a multiaffine polynomial over a larger set of variables. (We elaborate on polarization in Section 3.1.) Formally, following Brändén and Huh [17], for κ0n, let

Rκ[xi]={polynomials in R[xi]1in of degree at most κi in xi for every i},
and Rκa[xij]={multi-affine polynomials in R[xij]1in,1jκi}. The polarization operator Πκ:Rκ[xi]Rκa[xij] replaces every factor xα by
1(κα)i=1n(elementary symmetric polynomial of degree αi in the variables {xij}1jκi).

Proposition 3

(Brändén and Huh [17]). If p is d-homogeneous and SLC over Rκ[xi], then Πκ(p) is d-homogeneous and SLC over Rκa[xij].

We conclude this section by stating a large class of homogeneous polynomials that are known to be strongly log-concave.

Proposition 4

(Brändén and Huh [17]). For κ0n and w:KR0 a nonnegative weight function, consider

gκ(x)=αKw(α)α!xα,(6)
and assume that gκ is d-homogeneous. Let ν:0nR{} be defined by ν(α)=log(w(α)) for αK and ν(α)= otherwise. If ν is M-concave, then gκ is SLC.

2.5. Markov Chains

Let M=(Ω,P) be an ergodic and time-reversible Markov chain with state space Ω, transition matrix P, and stationary distribution π (which is unique because of the ergodicity assumption). Reversibility means that π(x)P(x,y)=π(y)P(y,x) for all x,yΩ. We write Pt(x,·) for the distribution over Ω at time step t with initial state xΩ. The total variation distance dTV(π,σ) of two distributions π and σ over Ω is defined as

dTV(π,σ)=maxSΩ|π(S)σ(S)|=12xΩ|π(x)σ(x)|,
where, for a distribution σ over Ω, we write σ(S)=xSσ(x). We say that two distributions π and σ are ϵ-close if dTV(π,σ)ϵ. The total variation distance of the distribution Pt(x,·) from π at time t with initial state x is denoted by Δx(t). The mixing time of M with initial state xΩ is τx(ϵ)=min{t:Δx(t)ϵ for all tt}. Informally, τx(ϵ) is the number of steps until the Markov chain is ϵ-close to its stationary distribution given that it is starting in x. A treaty of some more advanced Markov chain notions is given in Appendix A, including the definition of the modified log-Sobolev constant ρ=ρ(P), which can be used to bound the mixing time of a Markov chain. (The definition of ρ is deferred to Appendix A as we actually do not need it in the main body; we only rely on lower bounds on this constant.) It holds that (see, e.g., Bobkov and Tetali [15])
τx(ϵ)1ρ(P)(loglogπ(x)1+log(12ϵ2)).(7)

2.5.1. Markov Chain Decomposition.

Let Ω=Ω1Ωm be a disjoint partition of the state space Ω. Following Martin and Randall [46], consider π¯(i)=π(Ωi)=xΩiπ(x) and let P¯:[m]×[m][0,1] be defined by

P¯(i,j)=π¯(i)1xΩi,yΩjπ(x)P(x,y).

The Markov chain on [m] with transition matrix P¯ is called the projection chain on the partition {Ωi}i=1,,m. It is time-reversible with respect to the distribution π¯ over [m]. For i[m] the restriction chain on Ωi has transition matrix Pi:Ωi×Ωi[0,1] given by

Pi(x,y)={P(x,y)if xy,1zΩi\{x}P(x,z)if x=y.

It is time-reversible with respect to the distribution πi(x)=π(x)/π¯(i) for xΩi. In Appendix A.1, we give a Markov chain decomposition result based on the modified log-Sobolev constant ρ.

2.5.2. Base-Exchange Markov Chain.

Let N be a matroid, and let π be an SLC probability distribution over the set of bases B given by π(α)w(α) for some nonnegative weight function w:KR. Here, π(α)w(α) means that π(α)=w(α)/(αKw(α)). The base-exchange Markov chain on B is defined by the following transitions, in which BB is the current state of the Markov chain:

  • Select an element eB uniformly at random and remove it.

  • Pick a base BB with BBe with probability w(B) among all such bases B.

It is not hard to see, using the base-exchange property, that this procedure defines an ergodic, time-reversible Markov chain with stationary distribution π. Anari et al. [4] show that this chain is rapidly mixing for any matroid N. In particular, in a recent follow-up work, they give a tight mixing time bound (Anari et al. [6]).

Theorem 3

(Anari et al. [6]). Let N be matroid of rank r, and let π be an SLC probability distribution over the set of bases B given by π(α)w(α) for some weight function w:BR0. Then, the mixing time of the base-exchange random walk satisfies τ(ϵ)O(rlog(r/ϵ)).

We note that the mixing time is independent of the size n of the ground set E of the matroid N as well as the stationary distribution π. In this work, Theorem 3 is essentially only applied to partition and uniform matroids. Furthermore, Cryan et al. [20] show that the modified log-Sobolev of the base-exchange random walk satisfies ρ1/r, where r is the rank of the matroid.

2.6. Sampling Algorithms

Consider a class of (capacitated) congestion games Γ=(N,E,(Si)iN,(ce)eE,(ue)eE) with n players and m resources and cost functions ce:0 for eE. Let w:S0 be a weight function. In this work, an algorithm for sampling sS according to a distribution ϵ-close to π with π(s)w(s) is said to run in (randomized) polynomial time if the number of arithmetic operations can be upper bounded by a polynomial in n,m,log(1/ϵ),maxe,jlog(ce(j)) and maxslog(w(s)). (We assume the strategy sets can be represented in a compact form.) The generation of a uniform random 0/1 bit is considered to be one arithmetic operation.

Remark 3

(Real Numbers). In this work, we use Markov chains whose transition probabilities are, in general, not rational numbers (in particular for the Gibbs distribution). Whenever we use real numbers, it is implicitly assumed that we use sufficiently accurate approximations to these numbers. All our results remain valid when (real-valued) transition probabilities are replaced by sufficiently accurate rational approximations. We note that, roughly speaking, whenever we want to generate Markov chain transitions with probabilities proportional to eTΦ(s) for sS, our algorithms run in pseudo-polynomial time in terms of the values of the cost functions of the congestion game under consideration.

All our algorithmic results are based on running Markov chains for a sufficiently long time. We usually write our running time bounds as the product of two factors: the number of steps that we need to run the Markov chain (before it’s close to stationarity) and the complexity of implementing one such step. In particular, in all cases, the transitions probabilities of one step are determined by a sequence of rational numbers a=(a1,,az) and q=(q1,,qz), and we want to sample an index i[z] with probability

qieaiiqieai.(8)

Here, z as well as the encoding size of the qi are poly(n,m). We refer to C=C(n,m,a) as the computational complexity of sampling an index i according to (approximations of) the preceding probabilities in order not to overload our theorem statements. We say that probabilities of the form (8) are suitable.

2.7. Bipartite Graphs

An (undirected) bipartite graph G=(AB,F) is given by two disjoint sets of nodes A={a1,,an} and B={b1,,bm} with F{{a,b}:aA,bB}. We say that G has degree sequence (x,y) if d(ai)=xi for i=1,,n and d(bi)=yj for j=1,,m, where d(v) denotes the degree of node v in G. We write G(x,y) for the set of all bipartite graphs on AB with degree sequence (x,y).

3. General Approach

In Section 2.1, we give two possible definitions for the load profile of a strategy profile s=(s1,,sn)×iSi. For general congestion games, we define the resource load profile (s)=(e(s)) that keeps track of how many players use a particular resource e in s. For symmetric congestion games, we may in addition consider the strategy load profile z(s)=(zt(s))tS0 that keeps track of how many players use a particular strategy t from the common strategy set P.

A general approach for sampling a strategy profile according to the Gibbs distribution in Sections 4.1 and 6.1 is to first sample a (resource or strategy) load profile α according to approximately the right probability and then sample a strategy profile sS(α) uniformly at random. Remember that the set S(α) is used to denote all strategy profiles with the given (resource or strategy) load profile. The approach is summarized in Algorithm 1, in which we give the formulation for resource load profiles because we use Rosenthal’s potential Φ˜:0mR as given in (4). The formulation for strategy load profiles is exactly the same with Φ˜ replaced by Φ as given in (3). We note that, in order to sample a strategy load profile sS(α) uniformly at random in symmetric games, it suffices to generate a random permutation of the players in N={1,,n}.

Sampling a load profile with the correct probability in our applications corresponds to sampling a base of a discrete polymatroid according to a strongly log-concave distribution. In order to do this, we present a reduction of this problem to that of sampling a base of a matroid according to a strongly log-concave distribution in Section 3.1 (after which we can rely on Theorem 3).

Algorithm 1

(Gibbs Sampler for Congestion Game Γ)

Input: Congestion game Γ, rationality level T0, and ϵ0.

Output: Strategy profile sS according to distribution π¯ that is ϵ-close to Gibbs distribution π at rationality level T.

  • Step I: Sample load profile α according to a distribution σ that is ϵ-close to π given by

    π(α)=|S(α)|eTΦ˜(α).

  • Step II: Sample strategy profile sS(α) (approximate) uniformly at random.

3.1. Sampling Bases of Discrete Polymatroids

In this section, we describe how to generate a discrete polymatroid base, according to a strongly log-concave distribution over the set of all polymatroid bases by reducing it to the problem of generating a base of a matroid. This follows more or less directly from Theorem 3 by using the notion of polarization. Polarization can be seen as a functional version of the classic reduction from discrete polymatroids to matroids as given by Helgason [35]. (See also Schrijver [57, chapter 44.6b] for this reduction.)

For a polymatroid R0n, consider a d-homogeneous strongly log-concave polynomial

gR(x1,,xn)=αBRw(α)xα.

This is with positive coefficients and support the set of bases BR. Consider the matroid NR=(E,I) on ground set E={(i,j):1in,1jd}, where II if and only if the vector α(I)0n given by αe=|{f:(e,f)I}| satisfies α(I)R. The fact that NR is indeed a matroid follows directly from the fact that R is a polymatroid. Note that, for a given αR, we have

|{I:α(I)=α}|=(dα).(9)

We slightly abuse notation here and write d=(d,,d)n for the all d-vector.

Then, with Π(BR) the set of bases of NR, the polarization Π(g) of g can be written as

Π(g)(y11,,y1d,,yn1,,ynd)=BΠ(BR)(dα(B))1w(α(B))·yB=BΠ(BR)wΠ(α(B))·yB,
where, for BΠ(BR), we define
wΠ(α(B))=(dα(B))1w(α(B)).(10)

Polarization should be interpreted as spreading out the weight w(α) for αR equally over all bases B{A:α(A)=α}Π(BR). Proposition 3 implies that Π(g) is also strongly log-concave.

Example 2.

Let p(x1,x2)=x12x2+x1x2 so that supp(p)={(2,0),(1,1)} and take d=(2,2). Then,

Πd(p)=12x11x12(x21+x22)+14(x11+x12)(x21+x22)=12x11x12x21+12x11x12x22+14x11x21+14x11x22+14x12x21+14x12x22.

Note that, looking at the support of p, we have (dα)=2 monomials corresponding to α=(2,0) and (dα)=4 monomials corresponding to α=(1,1).

Corollary 1 now follows directly from Theorem 3. It says the following. Suppose the current state of the base-exchange Markov chain after t steps, starting from any state B0Π(BR), is the base BΠ(BR), and suppose we output the polymatroid base α(B). If t is large enough such that we are in state B with probability close to wΠ(α(B)) for every BΠ(BR), then α(B) is outputted with probability close to w(α(B)) with wΠ(α(B)) and w(α(B)) as in (10).

Corollary 1.

Let π be the distribution over BR with π(α)w(α), and let Ππ be the distribution over Π(BR) with Ππ(B)wΠ(α(B)). Let BΠ(BR), and let Πσt=Pt(B,·) be the distribution over Π(BR) after t steps of the base-exchange Markov chain M=(Π(BR),P). Let σt be the induced distribution over BR given by σt(α)=B:α(B)=αΠσt(B). If dTV(Πσt,Ππ)ϵ, then dTV(σt,π)ϵ.

Proof.

We have

2dTV(σ,π)=αBR|B:α(B)=ασ(B)w(α)|=αBR|B:α(B)=α[σ(B)(dα(B))1w(α)]|(using (9))BB|σ(B)w(α(B))|(triangle inequality)=2dTV(σ,π)2ϵ.

This gives the desired result. □

Remark 4.

It is possible to define a more direct Markov chain on the set of all bases of a given discrete polymatroid and prove that this chain is rapidly mixing (also based on Theorem 3), but this is not needed for our results.

4. Extension Parallel Congestion Games

An extension parallel congestion game is a symmetric congestion game in which the common strategy set P of the players consists of the o, d-paths in a (directed) extension parallel network G=(V,A) with source o and target d. For two given networks Gi=(Vi,Ai) with source oi and target di for i = 1, 2, let G=(V1V2,A1A2) be the union of G1 and G2. The parallel composition of G1 and G2 is the network obtained by identifying o1 with o2 and d1 with d2. These nodes are the source and target of G, respectively. The series composition of G1 and G2 is obtained by identifying d1 with o2. The node o1 becomes the source of G and d2 its target. An extension parallel network consists of (i) a single arc (o, d), (ii) two extension parallel networks in parallel, or (iii) a single arc in series with an extension parallel network. An example is given in Figure 2. For a given extension parallel graph G, we use P={p1,,pq} to denote all o, d-paths in G. Note that, for an extension parallel network, we have q|A|=m.

Figure 2. Example of an extension parallel network.

In this section we are always working with strategy load profiles (and so these are sometimes simply referred to as load profiles). The set of all possible strategy load profiles is denoted by L={α{1,,q}n:|α|=n}. We consider the potential Φ:L defined by Φ(α)=Φ(s) for some sS(α). This is well-defined as the potential value is the same for any choice of sS(α). The main result that we need in this section is the M-convexity of Rosenthal’s potential for EP congestion games.

Proposition 5

(Fujishige et al. [30]). Let Γ be an extension parallel congestion game. Then, the potential Φ:L defined by Φ(α)=Φ(s) for sS(α) is M-convex.

The M-convexity of Φ follows, in a nutshell, from the fact that the collection of sets {Qa:aA}, where Qa={pL:ap} is the set of all paths containing arc a, form a laminar family in the case of extension parallel networks. This implies that the potential Φ is laminar convex, which yields the M-convexity property of the potential. We refer the reader to Fujishige et al. [30] for more details.

Before giving our main result as sketched in Section 1.1, we first give a more direct approach for sampling from the Gibbs distribution in EP congestion games.

4.1. Sampling from the Gibbs Distribution

As mentioned in Section 3 and sketched in Algorithm 1, the high-level algorithmic idea for sampling a strategy profile according to the Gibbs distribution consists of first sampling a load profile α with the correct probability and then a strategy profile from S(α) uniformly at random. The main result of this section based on this approach is stated in Theorem 4.

Theorem 4.

Let ϵ>0 and T0, and let Γ be an extension parallel congestion game with n players. There is a randomized algorithm A with output distribution π¯ over Pn that is ϵ-close to the Gibbs distribution π at rationality level T and runs in (expected) time O(C·nlog(n/ϵ)) with C the complexity of implementing one step of a base-exchange Markov chain with suitable probabilities (see Section 2.6).

Proof.

Note that, for an extension parallel congestion game, the number of strategy profiles corresponding to a given load profile α is |S(α)|=n!/α!. (This is the number of ways in which we can assign n labeled balls to bins b1,,bq, where bi contains αi balls.)

Lemma 1.

The n-homogeneous generating polynomial

g(x1,,xn)=α[q]n:|α|=n|S(α)|eTΦ(α)xα=α[q]n:|α|=nn!α!eTΦ(α)xα(11)
is strongly log-concave. Hence, the distribution π over L given by π(α)n!α!eTΦ(α) for αL is strongly log-concave.

Proof.

Strong log-concavity is preserved under scalar multiplication by Proposition 2, so it suffices to show that

1n!g(x)=α[q]n:|α|=neTΦ(α)α!xα
is strongly log-concave. In turn, by Proposition 4, it is sufficient to show that log(eTΦ(α))=TΦ(α) is an M-concave function on its effective domain. As T0, this is equivalent to showing that Φ(α) is M-convex on its effective domain L={α[q]n:|α|=n}. This follows from Proposition 5. □

Because of Lemma 1, the polarization Π(g) of g in (11) is also strongly log-concave. The support of Π(g) can be seen as the bases of the n-uniform matroid N on ground set {(i,j):1i,jn}. Our algorithm now consists of first running the base-exchange Markov chain for O(nlog(n/ϵ)) steps, starting from any initial base. We output α(B), where B is the state we are in after the O(nlog(n/ϵ)) steps that were carried out. The resulting distribution σ(α) over L satisfies dTV(σ,π)ϵ by Corollary 1. We then uniformly at random choose a strategy profile from S(α). Let π¯ be the resulting output distribution over S. It remains to show that dTV(π,π¯)ϵ, which the following calculation shows. Note that

π¯(s)=α!n!σ(α),
where α=(s) is the load profile corresponding to strategy s. Then,
sS|π¯(s)π(s)|=αsS:(s)=α|π¯(s)π(s)|=αsS:(s)=α|α!n!σ(α)eTΦ(α)|=αn!α!|α!n!σ(α)eTΦ(α)|=α|σ(α)n!α!eTΦ(α)|2ϵ.

We conclude with analyzing the running time of the algorithm. One step of the base-exchange Markov chain can be implemented in time O(C) by definition. Generating an sS(α) uniformly at random can be done by generating a uniform random permutation μ of {1,,n}. We set si=p1 for players i=μ(1),,μ(α1),si=p2 for players μ(α1+1),,μ(α2+1), and so on. Generating a uniform random permutation can be done in time O(nlog(n)). □

4.2. Relaxed Logit Dynamics

In this section, we give the proof of Theorem 1, reformulated in Theorem 5. Formally, the relaxed logit dynamics Markov chain with current state sPn proceeds by

  • With probability 1/2: Select two players i,jN uniformly at random and transition to s given by

    sk={si if k=jsj if k=isk otherwise.

  • With probability 1/2: Perform a transition according to the logit dynamics (as in Section 2.2).

We note that, for any symmetric congestion game, this is a well-defined ergodic, time-reversible Markov chain with the Gibbs distribution as stationary distribution.

Theorem 5.

For an extension parallel congestion game Γ with common strategy set P and initial state sPn, the mixing time of the relaxed logit dynamics Markov chain at rationality level T0 satisfies 

τs(ϵ)n3(logn+loglog|P|+log(2TΦmaxϵ2)),
where Φmax=maxrSΦ(r) is the maximum value attained by Rosenthal’s potential over Pn.

Compared with the mixing time of the (nonrelaxed) logit dynamics for general games (Auletta et al. [12]), we get a doubly exponential improvement in terms of the dependence on TΦmax (at the cost of a small polynomial increase in the dependence on n).

The proof of Theorem 5 relies on a Markov chain decomposition argument. The idea underlying Markov chain decomposition is to divide the state space into a disjoint union of smaller sets. If one can show that the Markov chain is rapidly mixing when restricted to every smaller set and that it is easy to transition between the different smaller sets, then the original chain is rapidly mixing as well. In our setting, every load profile is identified with a smaller set, which is formed by all strategy profiles inducing the given load profile. To prove rapid mixing of the smaller sets, we compare the restricted chains to the so-called random transposition Markov chain, which is known to be rapidly mixing; see, for example, Goel [32]. To prove rapid mixing between the smaller sets, that is, between load profiles, we use again the connection to the base-exchange Markov chain but in a different way than in the proof of Theorem 4. In order to prove rapid mixing of all the Markov chains involved, we show that their modified log-Sobolev constant can be bounded appropriately. Recall that this is a quantity that can be used to upper bound the mixing time of a Markov chain (see Appendix A).

Proof of Theorem 5.

We use a Markov chain decomposition argument based on the two operations that define the relaxed logit dynamics Markov chain. We first partition the state space S=Pn naturally based on load profiles by setting Ωα=S(α) for αL, whereas before we have L={α:α[q]n and |α|=n}. Our proof approach is to apply the Markov chain decomposition theorem of Hermon and Salez [36] as given in Theorem A.1. In particular, for this, we need to bound the modified log-Sobolev constants of the projection and restriction chains. We start with the modified log-Sobolev constant ρ¯ of the projection chain.

The projection chain P¯ has state space L and stationary distribution π¯(α)=|S(α)|eTΦ(α) for αL. Let α,βL such that e|αeβe|=2; that is, there exist paths p and p such that

αe={βe+1if e=pβe1if e=pαeif eE\{p,p}.

In this case, we say that α and β are adjacent load profiles differing on paths p and p. Note that, if sS(α) and sS(β) are such that they differ by a deviation of some player i from path p to path p, then P(s,s)=12nexp(TΦ(p,si))Z, and this expression is the same for every such player i with si = p. Moreover, note that π(x)=π(y) for any two strategy profiles x,yS(α).

For some fixed choice of strategy profile xΩα and player i using path p, that is, si = p, the transition probabilities for adjacent load profiles can then be seen to equal (with Z and Z are the normalizing constants as in Section 2.2)

2P¯(α,β)=1π¯(α)xΩα,yΩβπ(x)P(x,y)=π(x)π¯(α)αpn|S(α)|exp(TΦ(β))Z=exp(TΦ(α))/Z|S(α)|exp(TΦ(α))/Zαpn|S(α)|exp(TΦ(β))Z=αpnexp(TΦ(β))Z=2αpP(s,s),(12)
where the last equality is true for any choice of sS(α) and sS(β). Note that this implies that, for any α,βL,sS(α) and sS(β), we have
P(s,s)P¯(α,β)=1αp1n.(13)

The lower bound of 1/n serves as our lower bound on χ as defined in Appendix A.1. In order to bound the modified log-Sobolev constant of the projection chain, one can use a comparison argument (as defined in Appendix A.2) with the base-exchange Markov chain on the support of the polarization Π(gΓ) of gΓ as in (11). In this section, the support corresponds to the set of bases of an n-uniform matroid. In particular, it holds that

ρ¯1n·ρ(Π(L))1n2,(14)
where ρ(Π(L)) is the modified log-Sobolev constant of the base-exchange Markov chain on the support of Π(g) with g as in (11). The second inequality comes from the fact that ρ(Π(L))1/n as is shown by Cryan et al. [20]. The first inequality is somewhat tedious to prove as it requires a Markov chain comparison argument between two Markov chains on different state spaces and is, therefore, deferred to Appendix B.

We continue with bounding the modified log-Sobolev constant of the restriction chains. In order to do this, we use a comparison argument with the random transposition Markov chain on the set Sk of all permutations of {1,,k}. Given a permutation σ, this chain proceeds by selecting two positions a and b uniformly at random and interchanging the positions of the elements σ(a) and σ(b). With ρrt denoting the modified log-Sobolev constant of this chain, it follows that, for every αL, we have

ραρrt1n1(15)
using the fact that ρrt1/(n1) as shown by Goel [32]. This comparison argument is also deferred to Appendix B. Now, applying Theorem A.2, it follows that ρ¯1/n3. Plugging this into (7), it then follows that
τs(ϵ)n3(logn+loglog|P|+log(2TΦmaxϵ2))
using that π(s)1|P|neTΦmax for every sS because of the nonnegativity of the cost functions. □

4.3. Uniform Sampling of Pure Nash Equilibria

In Theorem 6, we show that the result in Theorem 4 also implies that, for an extension parallel congestion game Γ, we can (approximate) uniformly at random sample a pure Nash equilibrium from the set NE(Γ) of all pure Nash equilibria of Γ in pseudo-polynomial time. That is, we sample every sNE(Γ) with probability 1/|NE(Γ)|. The (approximate) uniform sampling of combinatorial objects has received a lot of attention in the last 30 years, in particular within the area of theoretical computer science. However, to the best of our knowledge, no nontrivial results for (pure) Nash equilibria are known despite the fact that the problem of computing Nash equilibria has received much attention.

For the proof of Theorem 6, we use the fact that Nash equilibria are precisely the strategy profiles minimizing Rosenthal’s potential in EP congestion games. Furthermore, we exploit the fact that Theorem 4 holds for any rationality level T0. In particular, if we set T large enough, then most weight in the stationary distribution is assigned to profiles minimizing Rosenthal’s potential (under the assumption that the cost functions are integer-valued). This means that, with high probability, Algorithm 1 outputs a strategy profile minimizing Rosenthal’s potential with domain S. (Whenever we refer to Algorithm 1, we mean the implementation of the high-level approach as given in the previous section.) We use the Gibbs distribution with base two instead of e to avoid having to work with real numbers.

Theorem 6.

Let ϵ>0, and let Γ be an extension parallel congestion game with integer-valued cost functions and n players. There is a randomized algorithm A with output distribution π¯ over NE(Γ) that is ϵ-close to the uniform distribution over NE(Γ) and runs in (expected) time polynomial in n,m,Φmax and log(1/ϵ), where Φmax is the maximum value attained by Rosenthal’s potential.

For the proof of Theorem 6, we use the following correspondence between Nash equilibria and strategy profiles minimizing Rosenthal’s potential.

Proposition 6

(Fotakis [29], Holzman and Law-Yone [37]). The set of strategy profiles NE(Γ) of an extension parallel congestion game Γ coincides with the set of strategy profiles that minimize Rosenthal’s potential as in (3).

Proof of Theorem 6.

We first show that, for T sufficiently large in the algorithm used to prove Theorem 4, most weight is assigned to strategy profiles minimizing Rosenthal’s potential. We apply the idea in Algorithm 1 used to prove Theorem 4 with base two instead of base e. Remember that q is the number of (o, d)-paths in the extension parallel network of the game Γ. Let ϕ=Φ(s) be the common potential value of all strategy profiles sNE(Γ). For any other strategy profile sS\NE(Γ), we have

2TΦ(s)2T(ϕ+1)=2TeTϕ
by assumption that all cost functions are integer-valued. As there are q strategies to choose from for every player, we have |S|=qn=2nlog2(q). This implies that the Gibbs distribution π over S with rationality level T=nlog2(q)+log2(2/ϵ) satisfies
π(S\NE(Γ))=sNE(Γ)2TΦ(s)2nlog(q)2T2Tϕϵ2·π(NE(Γ)).(16)

The algorithm for sampling an (almost) uniform sample from NE(Γ) now works as follows. First, compute a strategy profile minimizing Rosenthal’s potential in order to determine ϕ. This can be done efficiently; see, for example, Fotakis [29]. Then, run Algorithm 1 with T=nlog2(q)+log2(2/ϵ) and ϵ=ϵ/2. If the resulting strategy profile has potential value ϕ, output this strategy profile, and otherwise, rerun Algorithm 1 until it does. Note that, with probability at least (1ϵ/2), Algorithm 1 outputs a strategy profile with potential value ϕ in one run. A simple argument then shows that the output distribution is ϵ-close to the uniform distribution over NE(Γ) as desired. □

Remark 5.

The pseudo-polynomial dependence coming from the polynomial dependence on Φmax rather than log2(Φmax) arises from the fact that we have to compute transition probabilities of the form 2ai/i2ai, where the ai are integers, which requires Ω(iai) random 0/1 bits (following the notion of suitable probabilities in Section 2.6). However, there is no a priori reason that the problem of (approximately) sampling pure Nash equilibria according to the uniform distribution requires pseudo-polynomial (in the input size of the original congestion game) time as opposed to sampling from the Gibbs distribution. We leave open the question of finding a (truly) polynomial time algorithm.

5. Max-k-Cut Games

In this section, we give an application of Theorem 5 to max-k-cut games. We first repeat their definition and show how they can be modeled as a special case of extension-parallel congestion games in case the graph G of the game is complete.

A max-k-cut game is given by an undirected graph G=(V,E) whose nodes V are the players. Every vV has strategy set [k]={1,,k} whose elements are referred to as colors. With N(i)={jV:{i,j}E} the set of neighbors of node iV, the utility of player i is the number of neighbors Ui(s)={jN(i):sisj} that choose a different color (hence, the name max-k-cut game, referring to the well-known max-cut problem (Garey and Johnson [31])). Equivalently, the cost of a player is given by the number of neighbors that choose the same color; that is, for s[k]n, we have Ci(s)=|{jN(i):si=sj}|, and in terms of (relaxed) logit dynamics, considering the cost of a player is equivalent to considering its utility because Ci(s)+Ui(s)=|N(i)| is independent of the chosen color in s.

When G=(V,E) is the complete graph, a max-k-cut game can naturally be modeled as a so-called symmetric singleton congestion game in which E=[k] is the set of resources and every resource is equipped with the cost function ce(x)=x1 for e[k]. Indeed, if players are using resource/color e, then the cost for every player choosing color is 1, which is precisely the number of neighbors choosing that color for every such node (because of the fact that G is the complete graph). Such games are easily seen to be special cases of extension parallel congestion games.

The relaxed logit dynamics may then be interpreted as the process by which either a player changes its color (according to the transitions as specified by the logit dynamics) or the colors of two players are interchanged. As Φmaxk(k1)/2, which is the total number of edges of the graph G, we obtain the following corollary of Theorem 5.

Corollary 2.

For a max-k-cut game played on the complete graph G=(V,E) with initial state s[k]n, where n=|V|, the mixing time of the relaxed logit dynamics Markov chain at rationality level T0 satisfies

τs(ϵ)n3(logn+loglogk+log(Tk(k1)ϵ2)).

Levin et al. [44] show that the mixing time of the (nonrelaxed) logit dynamics on the complete graph depends on the parameter T0 in the case k = 2. They show there is a critical value Tc so that, when TTc, the logit dynamics converge quickly to the Gibbs distribution, whereas when T>Tc, the dynamics converge slowly. What Corollary 2 shows is that, if one in addition is allowed to randomly interchange the colors of two nodes, the slow mixing for T>Tc can be resolved. We believe this to be of independent interest.

6. Capacitated Uniform Congestion Games

In this section, we consider u-capacitated k-uniform congestion games for vectors u=(u1,,um) and k=(k1,,kn). We write K=|k|=iki and U=|u|=eue. The vector u models the capacities of the resources eE, that is, the variables (ue)eE as defined in Section 2.1. The strategy set of player iN is given by all subsets SE of cardinality |S|=ki, that is, the bases of the ki-uniform matroid on E. We write Γ(u,k) for the collection of all u-capacitated k-uniform congestion games. We remark that, in this section, load profiles refer to resource load profiles as defined in Section 2.1 and no longer to path load profiles as considered in Section 4.

Note that we can naturally model a feasible strategy profile in s=(s1,,sn)S of a capacitated uniform congestion game as a (simple) bipartite graph G=(NE,F)G(u,k) on NE: there is an edge {i,e}F if and only if player iN uses resource eE in its strategy si. The main result needed in this section is stated in Proposition 7. We use the notation [x]b=x(x1)(xb+1) for x,b1. For a bipartite degree sequence (k,α), we then write Kb=i=1n[ki]b and Ab=j=1m[αj]b. Note that K=A=K1=A1.

Proposition 7

(McKay [49]). Let D be the collection of all bipartite degree sequences (k,α) for which 1kmaxαmax=o(K1/4). Then,

|G(k,α)|=K!iki!jαj!exp(K2K2·A2+O(max{kmax,αmax}4/K))
as K.

6.1. Sampling from the Gibbs Distribution

In this section, we give an (almost) polynomial time sampling algorithm that samples from a distribution that is close to the Gibbs distribution provided the game is sufficiently large. That is, we show that, for a large class of pairs (u, k), we can sample from a distribution close to the Gibbs distribution.

We follow again the high-level approach in Algorithm 1. The set of all feasible load profiles is now given by L(k,u)={α:0αu and |α|=iki} and S(α)=G(k,α) for any feasible load profile αL(k,u). Recall that we want to sample an αL(k,u) with probability proportional to (approximately) |S(α)|eTΦ(α), and then sample a strategy profile sS(α) with probability 1/|S(α)|.

A couple of problems arise here compared with the case of extension parallel congestion games. First of all, there is no polynomial time algorithm known to compute the numbers wα=|S(α)|. (In fact, it is still an open question whether this problem is # P-complete; see, e.g., Jerrum et al. [40].) Instead, we rely on a fully polynomial randomized approximation scheme for computing approximations w^α to the numbers wα up to arbitrary precision (Bezáková et al. [13], Jerrum et al. [40]). Second, in this case, the polynomial

g(x1,,xn)=αL(k,u)|S(α)|eTΦ(α)xα(17)
is in general not strongly log-concave. We overcome this problem by showing that we can restore strong log-concavity “approximately” when the game becomes large and when there are, in addition, suitable capacity constraints. We do this by using asymptotic enumeration formulas for the number of bipartite graphs with a given degree sequence, an area that has received considerable attention in combinatorics. It turns out that replacing |S(α)| by an asymptotic approximation ϕ(α) in (17) gives rise to a strongly log-concave polynomial.

Finally, the problem of sampling a strategy profile sS(α)=G(k,α) now corresponds to that of sampling a bipartite graph with degree sequence (k,α) for which many algorithms are known. The main result of this section is given in Theorem 7.

Theorem 7.

Let ϵ0, let π be the Gibbs distribution at rationality level T0, and let D be the class of all congestion games Γ(k,u) satisfying

1kmaxumax=o(K1/4).(18)

There is a randomized algorithm A for the class D and a constant K00 such that the output distribution σ¯ over S has the property that dTV(σ¯,π)ϵ whenever KK0. The algorithm runs in (expected) time

C·n(logn+loglog|P|+log(2TΦmaxϵ2))
with C(n,m,ϵ,Φmax)= poly(1/ϵ,n,m,Φmax).

Proof.

Setting

ϕ(α)=K!k!α!exp(K2K2·A2),
it follows, assuming (18) holds, that, for any 0αu, we have ϕ(α)=(1+o(1))|S(α)|, where o(1) is with respect to K. In particular, if KK0 for K0 large enough, it follows that
12|S(α)|ϕ(α)32|S(α)|.(19)

The next step is now to show that replacing |S(α)| by ϕ(α) in (20) gives rise to a strongly log-concave polynomial. The crucial observation here is to see that A2=jαj(αj1) is a separable convex function.

Lemma 2.

The K-homogeneous generating polynomial

g(x1,,xn)=αL(k,u)K!k!α!·exp(K2K2·j=1mαj(αj1))exp(TΦ(α))·xα(20)
is strongly log-concave.

Proof.

Following the proof of Lemma 1, first observe that (and note that k and all quantities involving k are considered fixed)

k!K!g(x1,,xn)=αL(k,u)1α!exp(K2K2·j=1mαj(αj1))exp(TΦ(α))xα
is strongly log-concave as well because of Proposition 2. Then, in order to apply Proposition 4, it suffices to show that
(K2K2·j=1mαj(αj1)+TΦ(α))
is M-concave over its domain L(k, u). (Formally speaking, we define it to be outside of L(k, u).) This follows directly from the fact that both j=1mαj(αj1) and Φ(α) are separable convex functions in (α1,,αm) over the (effective) domain L(k, u), and the fact that K2,K,T0. Separable convex functions (over effective domain L(k, u)) are known to be M-convex (Murota [52]). □

One can now carry out similar steps as in the proof of Theorem 4, albeit with some modifications. Again, the polarization Π(g) of g as in (20) is strongly log-concave as well. The support of Π(g) is now the set of bases of the n-truncation of a partition matroid N on ground set E=jEj, where Ej={(j,i):1in} with AE is independent if and only if |AEj|uj for j=1,,m. It follows that the modified log-Sobolev constant of this chain satisfies ρN1/n. Using a Markov chain comparison argument as described in Remark A.1 in Appendix A.2 with δ=1/2 because of (19) then yields that the modified log-Solev constant ρ of the Markov chain in which we use the original quantities |S(α)|, instead of the approximation ϕ(α), satisfies ρ1/(3n).

The algorithmic problem is now that we cannot compute the quantities wα=|S(α)| exactly in polynomial time, so one step of this base-exchange Markov chain cannot be implemented efficiently. Nevertheless we can use the approximation scheme of Bezáková et al. [13] to compute approximations w^α up to arbitrary precision in time polynomial in n, m and 1/ϵ (this gives the dependence of 1/ϵ in C). A Markov chain comparison argument (again as in Remark A.1) then implies that the base-exchange Markov chain on N using these approximations is also rapidly mixing, and in particular, it is sufficient to run the chain

3n(logn+loglog|P|+log(2TΦmaxϵ2))
steps and then output α(B), where B is the current base after having run the chain for the aforementioned number of steps.

We conclude with the sampling of a strategy profile from the set S(α) uniformly at random, which now requires sampling a bipartite graph from the set G(k,α) uniformly at random. One algorithm to do this for degree sequences satisfying the condition in Proposition 7 is that of Arman et al. [7] that runs in expected polynomial time. For an overview of algorithms that can be used to (approximately) sample a bipartite graph with a given degree sequence, see, for example, Dyer et al. [23]. □

Remark 6.

We remark here that the algorithm of Theorem 7 does not run in polynomial time as described in Section 2.6 because of the dependence of C on 1/ϵ (as opposed to the required log(1/ϵ)). This dependence arises because of the algorithmic approximations of the numbers |S(α)| used in the proof of Theorem 7. Alternatively, we could just use the approximations ϕ(α) straight away. However, these predictions only become accurate when K, so this gives a weaker result in terms of closeness to the Gibbs distribution.

Acknowledgments

The author is grateful to Prasad Tetali for pointing him to Hermon and Salez [36] and to the anonymous reviewers of the 22nd ACM Conference on Economics and Computation and Mathematics of Operations Research for their useful comments. A large part of this work has been carried out while the author was a postdoctoral fellow at the Max Planck Institute for Informatics in Saarbrücken, Germany. An abstract of this work appears in the Proceedings of the 22nd ACM Conference on Economics and Computation (Kleer [42]).

Appendix A. Markov Chains and Functional Inequalities

Three well-known quantities that can be used to upper bound the mixing time of a Markov chain are the Poincaré constant, the log-Sobolev constant, and the modified log-Sobolev constant. In this section, we define the modified log-Sobolev constant.

Let M=(Ω,P) be a time-reversible Markov chain with stationary distribution π and f,g:ΩR0. Let Eπ(f)=xΩπ(x)f(x). Furthermore, define the entropy-like quantity

Entπ(f)=Eπ[flog(f)flog(Eπ(f))]
and the Dirichlet form
EP(f,g)=12xΩyΩπ(x)P(x,y)[f(x)f(y)][g(x)g(y)].

The modified log-Sobolev constant of the Markov chain M is defined by

ρ(P)=inf{EP(f,log(f)) Entπ(f)|f:ΩR0, Entπ(f)0}.

As also mentioned in Section 2.5, it holds that

τx(ϵ)1ρ(P)(loglogπ(x)1+log(12ϵ2)).

A.1. Markov Chain Decomposition

Let M=(Ω,P) be a time-reversible Markov chain with stationary distribution π, and let Ω=Ω1Ωm be a disjoint partition of the state space. We write ρ for the modified log-Sobolev constant of this chain, ρ¯ for the modified log-Sobolev constant of the projection chain, and ρi for that of the restriction chain Pi (and define ρmin=miniρi).

Hermon and Salez [36] recently give a Markov chain decomposition theorem that applies to the Poincaré constant, the log-Sobolev constant, and the modified log-Sobolev constant. For other Markov chain decomposition theorems, see, for example, the work of Jerrum et al. [41], who, in particular, give stronger theorems for the Poincaré and log-Sobolev constants. We next describe the necessary objects to formulate their result.

Assume that, for each i,j[m] with ij and P¯(i,j)>0, we are given a coupling κij:Ωi×Ωj[0,1] of the probability distributions πi and πj. That is, κij is such that

xΩi,yΩjκij(x,y)=πi(x),yΩj,xΩiκij(x,y)=πj(y).

Based on the couplings κij, we define

χ=minxΩi,yΩj,i,j[m]{π(x)P(x,y)π¯(i)P¯(x,y)κij(x,y)},
with the range taken over all combinations for which the denumerator in the fraction is strictly positive. We state (a small variation of) the theorem of Hermon and Salez [36] for the modified log-Sobolev constant (for the other constants the statements are similar).

Theorem A.1

(Hermon and Salez [36]). With the preceding notation, it holds that ρmin{χρ¯,ρmin}. Furthermore, the parameter χ satisfies

χmaxxΩi,yΩj,i,j[m]:P¯(i,j)>0{P(x,y)P¯(i,j),P(y,x)P¯(j,i)}.

A.2. Markov Chain Comparison

Another useful property of proving mixing time bounds through Poincaré and (modified) log-Sobolev constants is that it is easy to see that small perturbations in the transition probabilities and the stationary distribution only result in mild variations in these constants by means of a Markov chain comparison argument. Goel [32] shows the following for the modified log-Sobolev constant based on similar results for the other constants by Diaconis and Saloff-Coste [22].

Theorem A.2

(based on Goel [32, Lemma 4.1]). Let M=(Ω,P) and M=(Ω,P) be two finite, reversible Markov chains with stationary distributions π and π, respectively, and modified log-Sobolev constant ρ and ρ, respectively. Assume there is a mapping ϕ mapping any (arbitrary) f:ΩR0 to f:ΩR0 and constants C,c>0 and B0 such that, for all f, we have

EP(f,logf)C·EP(f,logf) and c·Entπ(f) Entπ(f)+B·EP(f,logf).

Then,

cρC+Bρρ.

Remark A.1.

In particular, if Ω=Ω and there exists a δ>0 such that (1δ)P(x,y)P(x,y)(1+δ)P(x,y) for all x,yΩ and (1δ)π(x)π(x)(1+δ)π(x) for xΩ, it directly follows that

1ρ1+δ1δ·1ρ.

Appendix B. Comparison Arguments Omitted in Proof of Theorem 5

B.1. First Inequality in (14)

We start with showing the first inequality in (14), which is

ρ¯1n·ρ(Π(L)),
by using Theorem A.2. We heavily abuse notation and write Π(·) for many different objects in order not to overload the notation. Remember that ρ¯ is the modified log-Sobolev constant of the projection chain M=(L,P¯), where L={α[q]n:|α|=n} with stationary distribution given by
π¯(α)=|S(α)|eTΦ(α).

Furthermore, ρ(Π(L)) is the modified log-Sobolev constant of the base-exchange Markov chain MΠ=(Π(L),PΠ) on Π(L) with stationary distribution πΠ. Similar to what is explained in Section 3.1, we may write Π(L)=Fi, where Fi={(i,j):1jn} for i=1,,n, and Π(L) is then the set of bases of the n-uniform matroid on Fi. Roughly speaking, for every path pP, we introduce n auxiliary elements (corresponding to the n auxiliary variables introduced when polarizing).

For every αL, there are (nα) bases BΠ(L) corresponding to it (as in Section 3.1). We denote this set of bases by

Π(α)={AΠ(L):α(A)=α},
where α(A) is the vector given by αi=|AFi| for i=1,,n.

We start with defining the required mapping ϕ needed in Theorem A.2. For f:LR0, we define f:Π(L)R0 simply by setting f(A)=f(α(A)) for AΠ(L). It can then easily be checked that Entπ(f)= EntπΠ(f) as π(α)=AΠ(α)πΠ(A). This means that we can take c = 1 and B = 0 in Theorem A.2. In order to show the desired Dirichlet form inequality in the statement of Theorem A.2, it suffices to prove that, for any adjacent α,βL, it holds that

AΠ(α)πΠ(A)BΠ(β)PΠ(A,B)C·π¯(α)P¯(α,β).(B.1)
with C = n. The fact that this is sufficient follows from the observation that summing up (B.1) for all ordered pairs (α,β) for α,βL gives the desired result (in combination with the definition of f).

Now, fix α,βL and assume that they are adjacent (the case α=β can be dealt with similarly). Remember that γL is adjacent to α if e|αeγe|=2, that is, there exist paths p and p such that

αe={γe+1if e=pγe1if e=pαeif eE\{p,p}.

Let r be the path for which αr=βr+1 and write Nr(α) for all load profiles γ adjacent to α for which αr=γr+1 (including β). Following the definition of the base-exchange Markov chain, it then holds that, for any AΠ(α) and BΠ(β), we have

2·BΠ(β)PΠ(A,B)=αpn(nβpβ+1)(nβ)1eTΦ(β)γNr(α){α}(nγpγ+1)(nγ)1eTΦ(γ)Q,(B.2)
where pγ is used to indicate the path p=pγ for which αp=γp1 for γNp(α) and pα=r. With some care, it can be shown that, for γN(α){a}, it holds that
(nγpγ+1)(nγ)1(nβpβ+1)(nβ)1=γpγβpβ=βpγ+1βpβ1n.

Continuing the estimate in (B.2), we then get

Qn·αpneTΦ(β)γNr(α){α}eTΦ(γ)=2nP¯(α,β),
using (12) for the final equality. This gives the desired result in (B.1). Applying Theorem A.2 with c=1,C=n and B = 0 then gives the desired first inequality in (14).

B.2. First Inequality in (15)

In order to show the inequality in (15), we again use a Markov chain comparison between two chains on different state spaces. We want to show that

ραρrt,
where ρα is the modified log-Sobolev constant of the restriction chain on S(α) for aL, in which we randomly interchange the strategies of two players, and ρrt the modified log-Sobolev constant of the so-called random transposition walk. From now on, we fix some αL.

It is convenient to study these chains in terms of bipartite graphs with given degrees on node partition AB. For αL, we consider the degree sequence x=(x1,,xn) with xi = 1 for every iB, and the sequence y=(y1,,yq) with yp=αp for pA, where one should remember that q is the number of strategies, that is, paths, available in the common strategy set (denoted by A here) of all players. It follows directly that there is a one-to-one correspondence between S(α) and G(x,y), where, for a given strategy profile sS(α), there is an edge {i, p} if and only if si = p. (This is similar to the setting we consider in Section 6.) Given sS(α), our restriction chain can be interpreted as randomly selecting two edges from the bipartite graph Gs corresponding to the profile s and switching them if possible. That is, if we select {i, p} and {i,p} with pp, we delete the edges {i, p} and {i,p} and add the edges {i,p} and {i,p} (note that ii always holds as the nodes in B have degree one).

In order to introduce the random transposition Markov chain, we split up every node pjA into nodes pj1,,pjαj and consider bipartite graphs on two sets of n nodes B={1,,n} and A*=jAj* with Aj*={pj1,,pjαj}, where every node has degree one. That is, every such graph is a perfect matching between A* and B. Note that there are precisely

j=1qαj!=α!
perfect matchings corresponding to the graph Gs for sS(α) under the natural transformation in which, for a given perfect matching, we consider the graph that we get by merging all the nodes pj1,,pjαj back into one node pj for every j=1,,q. We denote this set of perfect matchings by H(s) for sS(α). The random tranposition Markov chain M=(H,P) with H=sH(s) denoting the set of all perfect matchings on the bipartition A*B proceeds by selecting two edges (of the current perfect matching) uniformly at random and switching them. Note that this is always possible here as opposed to in the case of our restriction chains on S(α).

We can now use a similar type of comparison argument as for the first inequality in (14) given earlier. We define the mapping ϕ for a given function f:S(α)R0 by setting f(M)=f(s) whenever MH(s) for sS(α). Let σ be the uniform distribution over S(α) and let σH the uniform distribution over H. Note that σ(s)=MH(s)σH(M). It then follows that, for every s,sS(α), we have

MH(s)σH(M)MH(s)P(M,M)=σ(s)Pα(s,s)(B.3)
because Pα(s,s)=1n(n1)=MH(s)P(M,M) for all MH(s) whenever ss and Pα(s,s)>0. Note that there is only one matching MH(s) such that P(M,M)>0. When s=s, we also have
MH(s)P(M,M)=Pα(s,s).

This implies that we can take C = 1. As before, we can take a = 1 and B = 0.

Endnotes

1 If one drops the assumption that the cost functions are nonnegative (see Section 2), the parameter Φmax can be replaced by ΔΦΦmaxΦmin, where Φmin is the minimum value attained by Rosenthal’s potential. This also holds for all subsequent results.

2 In fact, the result generalizes directly to “symmetric congestion games for which Rosenthal’s potential is M-convex,” but we are not aware of any other interesting class of congestion games for which this is true (and, therefore, choose to formulate our results in terms of EP congestion games). M-convexity of Rosenthal’s potential already fails to hold for the smallest nonextension parallel network congestion game, which is the graph that has two graphs, both consisting of two parallel edges, in series.

3 In fact, no efficient algorithm was known for the approximate sampling and counting of the bases of a matroid.

References

  • [1] Ackermann H, Röglin H (2008) On the convergence time of the best response dynamics in player-specific congestion games. Preprint, submitted May 8, https://arxiv.org/abs/0805.1130.Google Scholar
  • [2] Ackermann H, Röglin H, Vöcking B (2008) On the impact of combinatorial structure on congestion games. J. ACM 55(6):1–22.Google Scholar
  • [3] Anari N, Gharan SO, Vinzant C (2018) Log-concave polynomials, entropy, and a deterministic approximation algorithm for counting bases of matroids. Proc. 59th Annual Sympos. Foundations Comput. Sci., 35–46.Google Scholar
  • [4] Anari N, Liu K, Gharan SO, Vinzant C (2018) Log-concave polynomials III: Mason’s ultra-log-concavity conjecture for independent sets of matroids. Preprint, submitted July 2, https://arxiv.org/abs/1807.00929.Google Scholar
  • [5] Anari N, Liu K, Gharan SO, Vinzant C (2019) Log-concave polynomials II: High-dimensional walks and an FPRAS for counting bases of a matroid. Proc. 51st Annual ACM SIGACT Sympos. Theory Comput., 1–12.Google Scholar
  • [6] Anari N, Liu K, Gharan SO, Vinzant C, Vuong TD (2021) Log-concave polynomials IV: Approximate exchange, tight mixing times, and near-optimal sampling of forests. Proc. 53rd Annual ACM SIGACT Sympos. Theory Comput., 408–420.Google Scholar
  • [7] Arman A, Gao P, Wormald N (2019) Fast uniform generation of random graphs with given degree sequences. Proc. 60th Annual Sympos. Foundations Comput. Sci., 1371–1379.Google Scholar
  • [8] Asadpour A, Saberi A (2009) On the inefficiency ratio of stable equilibria in congestion games. Internet and Network Economics, 545–552.CrossrefGoogle Scholar
  • [9] Auletta V, Ferraioli D, Pasquale F, Persiano G (2013) Mixing time and stationary expected social welfare of logit dynamics. Theory Comput. Systems 53(1):3–40.CrossrefGoogle Scholar
  • [10] Auletta V, Ferraioli D, Pasquale F, Persiano G (2018) Metastability of logit dynamics for coordination games. Algorithmica 80(11):3078–3131.CrossrefGoogle Scholar
  • [11] Auletta V, Ferraioli D, Pasquale F, Penna P, Persiano G (2015) Logit dynamics with concurrent updates for local interaction potential games. Algorithmica 73(3):511–546.CrossrefGoogle Scholar
  • [12] Auletta V, Ferraioli D, Pasquale F, Penna P, Persiano G (2016) Convergence to equilibrium of logit dynamics for strategic games. Algorithmica 76(1):110–142.CrossrefGoogle Scholar
  • [13] Bezáková I, Bhatnagar N, Vigoda E (2007) Sampling binary contingency tables with a greedy start. Random Structures Algorithms 30(1–2):168–205.CrossrefGoogle Scholar
  • [14] Blume LE (1993) The statistical mechanics of strategic interaction. Games Econom. Behav. 5(3):387–424.CrossrefGoogle Scholar
  • [15] Bobkov SG, Tetali P (2006) Modified logarithmic Sobolev inequalities in discrete settings. J. Theoretical Probab. 19(2):289–336.CrossrefGoogle Scholar
  • [16] Brändén P, Huh J (2018) Hodge-Riemann relations for Potts model partition functions. Preprint, submitted November 5, https://arxiv.org/abs/1811.01696.Google Scholar
  • [17] Brändén P, Huh J (2020) Lorentzian polynomials. Ann. Math. 192(3):821–891.CrossrefGoogle Scholar
  • [18] Camerer CF (2010) Behavioural game theory. Behavioural and Experimental Economics, 42–50.Google Scholar
  • [19] Chien S, Sinclair A (2011) Convergence to approximate Nash equilibria in congestion games. Games Econom. Behav. 71(2):315–327.CrossrefGoogle Scholar
  • [20] Cryan M, Guo H, Mousa G (2019) Modified log-Sobolev inequalities for strongly log-concave distributions. Proc. 60th Annual Sympos. Foundations Comput. Sci., 1358–1370.Google Scholar
  • [21] Del Pia A, Ferris M, Michini C (2017) Totally unimodular congestion games. Proc. 28th Annual ACM-SIAM Sympos. Discrete Algorithms, 577–588.Google Scholar
  • [22] Diaconis P, Saloff-Coste L (1996) Logarithmic Sobolev inequalities for finite Markov chains. Ann. Appl. Probab. 6(3):695–750.CrossrefGoogle Scholar
  • [23] Dyer M, Greenhill C, Kleer P, Ross J, Stougie L (2021) Sampling hypergraphs with given degrees. Discrete Math. 344(11):112566.CrossrefGoogle Scholar
  • [24] Even-Dar E, Kesselman A, Mansour Y (2007) Convergence time to Nash equilibrium in load balancing. ACM Trans. Algorithms 3(3):32.CrossrefGoogle Scholar
  • [25] Fabrikant A, Papadimitriou C, Talwar K (2004) The complexity of pure Nash equilibria. Proc. 36th Annual ACM Sympos. Theory Comput., 604–612.Google Scholar
  • [26] Fanelli A, Moscardelli L (2011) On best response dynamics in weighted congestion games with polynomial delays. Distributed Comput. 24(5):245–254.CrossrefGoogle Scholar
  • [27] Feder T, Mihail M (1992) Balanced matroids. Proc. 24th Annual ACM Sympos. Theory Comput., 26–38.Google Scholar
  • [28] Ferraioli D (2013) Logit dynamics: A model for bounded rationality. ACM SIGecom Exchanges 12(1):34–37.CrossrefGoogle Scholar
  • [29] Fotakis D (2010) Congestion games with linearly independent paths: Convergence time and price of anarchy. Theory Comput. Systems 47(1):113–136.CrossrefGoogle Scholar
  • [30] Fujishige S, Goemans M, Harks T, Peis B, Zenklusen R (2015) Congestion games viewed from m-convexity. Oper. Res. Lett. 43(3):329–333.CrossrefGoogle Scholar
  • [31] Garey MR, Johnson DS (1990) Computers and Intractability: A Guide to the Theory of NP-Completeness (W.H. Freeman & Co., New York).Google Scholar
  • [32] Goel S (2004) Modified logarithmic Sobolev inequalities for some models of random walk. Stochastic Processes Appl. 114(1):51–79.CrossrefGoogle Scholar
  • [33] Gourvès L, Monnot J (2009) On strong equilibria in the max cut game. Proc. Fifth Internat. Workshop Internet Network Econom., 608–615.Google Scholar
  • [34] Gurvits L (2010) On multivariate Newton-like inequalities. Advances in Combinatorial Mathematics, 61–78.Google Scholar
  • [35] Helgason T (1974) Aspects of the theory of hypermatroids. Hypergraph Seminar (Springer, Berlin/Heidelberg), 191–213.Google Scholar
  • [36] Hermon J, Salez J (2019) Modified log-Sobolev inequalities for strong-Rayleigh measures. Preprint, submitted February 7, https://arxiv.org/abs/1902.02775.Google Scholar
  • [37] Holzman R, Law-Yone N (1997) Strong equilibrium in congestion games. Games Econom. Behav. 21(1–2):85–101.CrossrefGoogle Scholar
  • [38] Ieong S, McGrew R, Nudelman E, Shoham Y, Sun Q (2005) Fast and compact: A simple class of congestion games. Proc. 20th Natl. Conf. Artificial Intelligence, 489–494.Google Scholar
  • [39] Jerrum M, Sinclair A (1993) Polynomial-time approximation algorithms for the Ising model. SIAM J. Comput. 22(5):1087–1116.CrossrefGoogle Scholar
  • [40] Jerrum M, Sinclair A, Vigoda E (2004) A polynomial-time approximation algorithm for the permanent of a matrix with nonnegative entries. J. ACM 51(4):671–697.CrossrefGoogle Scholar
  • [41] Jerrum M, Son JB, Tetali P, Vigoda E (2004) Elementary bounds on Poincaré and log-Sobolev constants for decomposable Markov chains. Ann. Appl. Probab. 14(4):1741–1765.CrossrefGoogle Scholar
  • [42] Kleer P (2021) Sampling from the Gibbs distribution in congestion games. Proc. 22nd ACM Conf. Econom. Comput., 679–680.Google Scholar
  • [43] Kleer P, Schäfer G (2017) Potential function minimizers of combinatorial congestion games: Efficiency and computation. Proc. 18th ACM Conf. Econom. Comput., 223–240.Google Scholar
  • [44] Levin DA, Luczak MJ, Peres Y (2010) Glauber dynamics for the mean-field Ising model: Cut-off, critical power law, and metastability. Probab. Theory Related Fields 146(1–2):223–265.CrossrefGoogle Scholar
  • [45] Mamageishvili A, Penna P (2016) Tighter bounds on the inefficiency ratio of stable equilibria in load balancing games. Oper. Res. Lett. 44(5):645–648.CrossrefGoogle Scholar
  • [46] Martin RA, Randall D (2000) Sampling adsorbing staircase walks using a new Markov chain decomposition method. Proc. 41st Annual Sympos. Foundations Comput. Sci., 492–502.Google Scholar
  • [47] Mäs M, Nax HH (2016) A behavioral study of “noise” in coordination games. J. Econom. Theory 162:195–208.CrossrefGoogle Scholar
  • [48] McFadden D (1973) Conditional logit analysis of qualitative choice behavior. Frontiers in Econometrics, 105–142.Google Scholar
  • [49] McKay BD (1984) Asymptotics for 0-1 matrices with prescribed line sums. Enumeration and Design (Academic Press, Cambridge, MA), 225–238.Google Scholar
  • [50] Mihail M, Vazirani U (1989) On the expansion of 0-1 polytopes. J. Combin. Theory Ser. BGoogle Scholar
  • [51] Murota K (1998) Discrete convex analysis. Math. Programming 83(1–3):313–371.CrossrefGoogle Scholar
  • [52] Murota K (2003) Discrete convex analysis. SIAM Monograph on Discrete Mathematics and Applications (Society for Industrial and Applied Mathematics).CrossrefGoogle Scholar
  • [53] Murota K (2009) Recent developments in discrete convex analysis. Research Trends in Combinatorial Optimization, 219–260.Google Scholar
  • [54] Penna P (2018) The price of anarchy and stability in general noisy best-response dynamics. Internat. J. Game Theory 47(3):839–855.CrossrefGoogle Scholar
  • [55] Rosenthal RW (1973) A class of games possessing pure-strategy Nash equilibria. Internat. J. Game Theory 2:65–67.CrossrefGoogle Scholar
  • [56] Sandholm WH (2010) Population Games and Evolutionary Dynamics (MIT Press, Cambridge, MA).Google Scholar
  • [57] Schrijver A (2003) Combinatorial optimization: Polyhedra and efficiency. Matroids, Trees, Stable Sets, vol. B, Algorithms and Combinatorics (Springer-Verlag, Berlin).Google Scholar
  • [58] Sinclair A (1992) Improved bounds for mixing rates of Markov chains and multicommodity flow. Combin. Probab. Comput. 1(4):351–370.CrossrefGoogle Scholar
  • [59] Skopalik A, Vöcking B (2008) Inapproximability of pure Nash equilibria. Proc. 40th Annual ACM Sympos. Theory Comput., 355–364.Google Scholar