Thermodynamic formalism for amenable groups and countable state spaces

There are two general ways to describe a system composed of many particles: microscopically and macroscopically. The first one makes use of the exact positions of the particles, as well as their local interactions. The second one, in turn, is usually outlined by thermodynamic quantities such as energy and entropy. One could say that statistical mechanics — originated from the works of Boltzmann [10] and Gibbs [32] — is the bridge between the microscopic and the macroscopic descriptions of this kind of systems. In this connection, Gibbs measures are a central object.

It is fair to say that Gibbs measures are at the core of the “conceptual basis of equilibrium statistical mechanics” [52]. Relevant examples are the Ising model, which tries to capture the magnetic properties of certain materials; the hard-core model, that describes the distribution of gas particles in a given environment; among many others [29, 30, 31]. In these cases it is customary to consider that the many particles interacting are infinite, take a value from a state space $A$ (also called alphabet when $A$ is countable), and they are disposed in a crystalline structure. This structure and its symmetries are usually represented by a countable group $G$, possibly with some Cayley graph associated with it. A particular case is the hypercubic $d$-dimensional lattice, which can be understood as the Cayley graph of the finitely generated abelian group $G=\mathbb{Z}^{d}$ according to its canonical generators. Then, it is natural to represent an arrangement of particles as an element of the space of configurations $X=A^{G}$, the $G$-full shift. Considering this, one is interested in certain measures $\mu$ in the space $\mathcal{M}(X)$ of Borel probability measures supported on $X$. More specifically, the measures of interest are the ones that describe these kind of systems when they are in thermal equilibrium, where the energy of configurations is given by some potential $\phi:X\to\mathbb{R}$. However, there are many mathematically consistent ways to represent that situation by choosing an appropriate measure $\mu\in\mathcal{M}(X)$ and, as the theory evolved, it drew the attention from different areas of expertise such as probability [51, 27] and ergodic theory [56, 14]. Consequently, the very concept of Gibbs measure started to develop in more abstract and not always equivalent directions.

We focus mainly on four conceptualizations of the idea of thermal equilibrium, namely, DLR, conformal, Bowen-Gibbs, and equilibrium measures. We now proceed to briefly describe each of them.

Dating back to the 60’s, Dobrushin [22, 23] and, independently, Lanford and Ruelle [41] proposed a concept of Gibbs measure that extended the usual Boltzmann-Gibbs formalism to the infinite particles setting. Roughly, the idea involved looking for probability distributions compatible with a family of maps — sometimes called specification — that prescribe conditional distributions inside finite subsets of $G$ given some fixed configuration outside. More specifically, given a collection $\gamma=(\gamma_{K})_{K\in\mathcal{F}(G)}$ of probability kernels $\gamma_{K}\colon\mathcal{B}\times X\to[0,1]$, with $\mathcal{F}(G)$ the set of finite subsets of $G$ and $\mathcal{B}$ the Borel $\sigma$-algebra, one is interested in finding measures $\mu\in\mathcal{M}(X)$ such that $\mu\gamma_{K}=\mu$ for every $K\in\mathcal{F}(G)$, where $\mu\gamma_{K}$ is a new measure (a priori, different from $\mu$) obtained from $\mu$ via $\gamma_{K}$. Those distributions are called DLR measures after the above cited authors and they have received considerable attention from both mathematical physicists and probabilists (see, for example, [30, 31, 38, 52]).

Another rather classical way to define a Gibbs measure, which does not involve conditional distributions, was introduced by Capocaccia in [17]. Given a class $\mathcal{E}$ of local homeomorphisms $\tau:X\to X$ and a potential $\phi:X\to\mathbb{R}$, one is interested in measures $\mu$ such that $\frac{d(\mu\circ\tau^{-1})}{d\mu}=\mathop{\textrm{\rm exp}}\nolimits(\phi_{*}^{\tau})$ for every $\tau\in\mathcal{E}$, where $\phi_{*}^{\tau}:X\to\mathbb{R}$ is a function representing the energy difference between a configuration $x$ and $\tau(x)$ (e.g., see [38, Definition 5.2.1]). This kind of measures fits in the more general context of $(\Psi,\mathcal{R})$-conformal measures explored in [1], where $\mathcal{R}$ is a Borel equivalence relation and $\Psi\colon\mathcal{R}\to\mathbb{R}_{+}$ is a measurable function. Then, Capocaccia’s measures, that we simply call conformal measures, can be recovered by taking a function $\Psi$ related to the given potential and $\mathcal{R}$ the tail relation in the space of configurations. By considering other particular Borel relations $\mathcal{R}$ and measurable functions $\Psi$, one can recover other relevant notions of conformal measures, such as the ones presented in [20, 49, 53], that are mainly adapted to the one-dimensional setting, i.e., when $G=\mathbb{Z}$ or, considering also semigroups, when $G=\mathbb{N}$.

A third possibility, introduced by Rufus Bowen in a one-dimensional and ergodic theoretical context [14], is to define Gibbs measures by specifying bounds for the probability of cylindrical events. More concretely, one is interested in the measures $\mu\in\mathcal{M}(X)$ for which there exists constants $C>0$ and $p\in\mathbb{R}$ such that

As in [7], we call those measures Bowen-Gibbs measures to avoid confusion. This definition has been considered in the literature [18, 36, 38, 52] and also relaxed versions of it, such as the so-called weak Gibbs measures [58, 60], where the constant $C$ is replaced by a function that grows sublinearly in $n$. This and further relaxations have also played a relevant role in the multi-dimensional case, this is to say, when $G=\mathbb{Z}^{d}$ and $d>1$, for finite state spaces (e.g., see [38, Theorem 5.2.4]).

The last important definition considered in this work is the one of equilibrium measure. When $X$ is a finite configuration space, equilibrium measures are simply probability vectors that maximize the sum (or difference) of an entropy- and an energy-like quantity, that is, a quantity like

where $k=|X|$, $x_{i}\in X$, $\phi:X\to\mathbb{R}$ is a potential, $p=(p_{1},\dots,p_{k})$ is a probability vector with $p_{i}$ the probability associated with $x_{i}$, and $H(p)$ is the Shannon entropy of $p$. These measures were considered, for example, in [31, 38, 52]. On the other hand, when $X$ is an infinite configuration space and there is a robust notion of specific entropy, let’s say $h(\mu)$, we are interested in studying measures $\mu\in\mathcal{M}(X)$ that maximize the quantity $h(\mu)+\int\phi\,d\mu$ for a continuous potential $\phi:X\to\mathbb{R}$. This notion tries to capture the macroscopic behaviour of the system without making any assumption of the microscopic structure.

The problem of proving equivalences among these and other related notions has already been studied in different settings. We mention some relevant results that can be found in the literature.

In the one-dimensional case, for finite state spaces, Meyerovitch [44] proved the equivalence between conformal measures and DLR measures for some families of proper subshifts. Also, Sarig [54, Theorem 3.6] proved that any DLR measure on a mixing subshift of finite type is a conformal measure, for a different but related notion of conformal, restricted to the one-dimensional setting. In the same work, for one-sided and countably infinite state spaces, Sarig [54, Proposition 2.2] proved that conformal measures — according to his definition — are DLR measures for topological Markov shifts. In this same setting, Mauldin and Urbański [43] proved the existence of equilibrium measures and that any equilibrium measure satisfies a Bowen-Gibbs equation. Moreover, if the topological Markov shift satisfies the BIP property and the potential has summable variation, Beltrán, Bissacot, and Endo [6] proved that DLR measures and conformal measures — in the same sense as Sarig — are equivalent. Finally, for potentials with summable variation on sofic subshifts, Borsato and MacDonald [12] proved the equivalence between DLR and equilibrium measures. There are also other classes of measures in the one-dimensional case which we do not treat here, such as $g$-measures [37, 59] and eigenmeasures associated with the Ruelle operator [14, 52]. When the state space is finite, it is known that the set of DLR measures and $g$-measures do not contain each other [28, 9], but there is a characterization for when a $g$-measure is a DLR measure [7]. In addition, eigenmeasures coincide with DLR measures for continuous potentials in the one-sided setting, as proven by Cioletti, Lopes, and Stadlbauer in [19]. Pioneering works in the one-dimensional countably infinite state space setting can be found in [33, 34].

In the multi-dimensional case, some results regarding the equivalences among the four notions of Gibbs measures have been proved for finite state spaces. A first important reference is Keller [38, Theorem 5.2.4 and Theorem 5.3.1], where it is proven that when $\phi:X\to\mathbb{R}$ is regular (which includes the case of local and Hölder potentials, and well-behaved interactions), then the four definitions are equivalent. Here, by regular, we mean that

where $\delta_{n}(\phi)$ is the oscillation of $\phi$ when considering configurations that coincide in a specific finite box, namely, $[-n,n]^{d}\cap\mathbb{Z}^{d}$. Other classical references in this setting are due to Dobrushin [21] and Lanford and Ruelle [41], which, combined, establish the equivalence between DLR measures and equilibrium measures for a general class of subshifts of finite type. Kimura [40] generalized the equivalence between DLR and conformal measures for subshifts of finite type, and some of the implications are true for more general proper subshifts. In the countably infinite state space setting, Muir [45, 46] obtained all equivalences for the $G$-full shift when $G=\mathbb{Z}^{d}$. In order to do this, it was required that the potential $\phi:X\to\mathbb{R}$ is regular and satisfies a normalization criterion, namely, exp-summability:

This last condition is automatically satisfied when $A$ is finite.

Results proving equivalences between different kinds of Gibbs measures go beyond the amenable [55, 5, 2, 15] and even the symbolic setting to general topological dynamical systems [3, 36].

One of our main contributions is to exhibit conditions to guarantee that the four notions of Gibbs measures presented above are equivalent, when considering the state space $A=\mathbb{N}$ and an arbitrary countable amenable group $G$, thus extending Muir’s methods to the more general amenable case. Countable amenable groups play a fundamental role in ergodic theory [48] and include many relevant classes of groups, such as abelian (so, in particular, $G=\mathbb{Z}^{d}$), nilpotent, and solvable groups and are closed under many natural operations, namely, products, extensions, etc. (e.g., see [42]). In the more general group and finite state space setting, the equivalence between DLR and conformal measures was extended to general subshifts over a countable discrete group $G$ with a special growth property by Borsato-MacDonald [13, Theorems 5 and Theorem 6]. Recently, a different proof for the equivalence between DLR and conformal measures for any proper subshift was given by Pfister in [50]. Also, in [4], a Dobrushin-Lanford-Ruelle type theorem is proven in the case that the group is amenable and a topological Markov property holds, which is satisfied, in particular, by subshifts of finite type. Here, as Muir, we focus on the $G$-full shift case. We consider the configuration space $X=\mathbb{N}^{G}$, for $G$ an arbitrary countable amenable group, and an exp-summable potential $\phi:X\to\mathbb{R}$ with summable variation (according to some exhausting sequence). The concept of summable variation extends the one of regular potential presented before. More precisely, a potential $\phi$ has summable variation if

where $\{E_{m}\}_{m}$ is an exhausting sequence for $G$ and $\delta_{E_{m}}(\phi)$ is a standard generalization of $\delta_{m}(\phi)$.

The paper is organized as follows. First, in Section 2, we present some preliminary notions about amenable groups $G$, the corresponding symbolic space $\mathbb{N}^{G}$, and potentials. Later, in Section 3, we introduce the concept of pressure in our framework and we prove its existence. Also, we prove that it satisfies an infimum rule and that it can be obtained as the supremum of the pressures associated with finite alphabet subsystems. In order to achieve this, we use relatively new techniques for tilings of amenable groups [26] and, inspired by ideas for entropy from [25], we develop a generalization of Shearer’s inequality for pressure. In Section 4, we introduce spaces of permutations and Gibbsian specifications in order to pave the way for the definitions of conformal and DLR measures, respectively. Next, in Section 5, we prove the equivalence between the four notions of Gibbs measures mentioned above given some conditions on the potential, such as exp-summability and summable variation. We also prove related results involving equilibrium measures. In order to prove the equivalence between DLR and conformal measures we rely on the strategies presented on [45] for the $G=\mathbb{Z}^{d}$ case, which already considers an infinite state space. Moreover, using Prokhorov’s Theorem and relying on the existence of conformal measures in the compact setting [20], we prove the existence of a conformal (and DLR) measure in our context. We also prove that DLR measures are Bowen-Gibbs. If it is also the case that the measure is invariant under shift actions of the group, we prove that any Bowen-Gibbs measure is an equilibrium measure and that any equilibrium measure is a DLR measure. At last, in Section 6, we show how to recover previous results from ours and, inspired by the Potts model and considering a version of it with countably many states, we exhibit a family of examples for which all our results apply non-trivially and, in addition, a version of Dobrushin’s Uniqueness Theorem adapted to our setting holds, thus providing a regime where the uniqueness of a Gibbs measure is satisfied.

We dedicate this section to introduce the pressure of a potential. We define and work on the setting of exp-summable potentials with summable variation on a countable alphabet. The pressure — basically equivalent to the specific Gibbs free energy — is a very relevant thermodynamic quantity that helps to capture the concept of Gibbs measure in a quantitative way.

First, we prove that the pressure, which we define through a limit over sets that are becoming more and more invariant, exists in the finite alphabet case. The definition of the pressure is often done in terms of a particular Følner sequence, which is an, a priori, less robust and less overarching approach. Existence of the limit for a particular Følner sequence $\{F_{n}\}_{n}$ and the fact that it is independent on the choice of such sequence is well-known (see, for example, [57, 35, 16], in the context of absolutely summable interactions). Here, we prove something stronger: that our definition of pressure obeys the infimum rule — which is a refinement of the Ornstein-Weiss Lemma (see, for example, [39, §4.5]) —, this is to say, it can be expressed as an infimum over all finite sets of $G$. In order to conclude this, we extend the results about Shearer’s inequality in [25] for topological entropy to pressure.

Now, in the countable alphabet context, we take a similar approach. First, we consider again a definition of pressure in terms of sets that are becoming more and more invariant. Next, we prove that the infimum rule still holds and, finally, we prove that the pressure can be obtained as the supremum of the pressures associated with finite alphabet subsystems. A related result was obtained by Muir in [45] for the $\mathbb{Z}^{d}$ group case, where the pressure was defined as a limit over a particular type of Følner sequence, namely, open boxes centered at the origin of radius $n$. The existence of this limit was proven through a sub-additivity argument that exploits the property that large boxes can be partitioned into many equally sized ones, which might not be valid in more general groups. In order to generalize this idea of partitioning sets, we make use of tiling techniques introduced in [26], which, together to what is done in the finite alphabet case, allow us to prove the infimum rule for infinite alphabets over a countable amenable group. This type of result was not considered in [45].

We begin by introducing some definitions. Given a potential $\phi\colon X\to\mathbb{R}$ and $F\in\mathcal{F}(G)$, define the partition function for $\phi$ on $F$ as

where $\sup\phi_{F}([w])=\sup\{\phi_{F}(x):x\in[w]\}$. We define the pressure of $\phi$, which we denote by $p(\phi)$, as

whenever such limit as $F$ becomes more and more invariant exists. In addition, given a finite subset $A\in\mathcal{F}(\mathbb{N})$, we define $Z_{F}(A,\phi)$ as the partition function associated with the restriction of $\phi$ to $A^{G}$. More precisely,

Similarly, we define $p(A,\phi)$ as

whenever such limit exists.