We prove the continuity of the extremal decomposition measure of the free state of low temperature Potts models, and more generally of ferromagnetic finite-spin models, on a regular tree, including general clock models. The decomposition is supported on uncountably many inhomogeneous extremal states, that we call glassy states. The method of proof provides explicit concentration bounds on branch overlaps, which play the role of an order parameter for typical extremals. The result extends to the counterpart of the free state (called central state) in a wide range of models which have no symmetry, allowing also the presence of sufficiently small field terms. Our work shows in particular that the decomposition of central states into uncountably many glassy states in finite-spin models on trees at low temperature is a generic phenomenon, and does not rely on symmetries of the Hamiltonian.
Any \mu \in {\rm ex}\mathcal{G}(\gamma) is characterized by the equivalent items:
Tail-triviality: A \in \mathcal{F}_\infty \Longrightarrow \mu(A) \in \{0,1\}.
Short-range correlations
The limit (?) exists and is an extremal Gibbs measure for \mu-almost every \omega.
Assume that \mathcal{G}(\gamma)\neq \emptyset. Then \mathcal{G}(\gamma) is a convex subset of \mathcal{M}_1^+(\Omega) whose extreme boundary is denoted {\rm ex} \mathcal{G}(\gamma), and satisfies the following properties:
where \alpha_\mu \in \mathcal{M}_1^+\Big({\rm ex}\mathcal{G}(\gamma),e({\rm ex} \mathcal{G}(\gamma))\Big) is defined for all M \in e({\rm ex} \mathcal{G}(\gamma)) by
If \mu is an extremal Gibbs measure then \mu is a tree-indexed Markov chain.
Let \mu be a homogeneous tree-indexed Markov chain. Denote by P its transition matrix (which is then homogeneous, too) and define the corresponding maximal jump probability to be
Equivalently, for all v\sim 0 we have
Consider any central state \mu in the class of models with Hamiltonian of the form (?), fulfilling the bounds (?) and (?). Then, there exists \beta_0>0 large enough such that for any \beta>\beta_0, for any a\in{\mathbb Z}_q,
where \epsilon_1(\beta)=\epsilon_1(\beta,u,U,d)\downarrow0 as \beta\uparrow\infty and \epsilon_2(p_1)=\epsilon_2 (p_1,u,U,d)\downarrow0 as p_1\downarrow0.
In particular, if \mu is the free state of a clock model, the reconstruction bound
holds at large enough \beta.
Consider any central state \mu in the class of models with Hamiltonian of the form (?) fulfilling the bounds (?) and (?). Then, there exists \beta_0>0 large enough such that for any \beta>\beta_0,
where \epsilon_1(\beta)\downarrow0 as \beta\uparrow\infty and \epsilon_2 (p_1)\downarrow0 as p_1\downarrow0 are the same quantities as in Theorem 3.1. In particular, for large enough \beta and small enough p_1=p_1(\mu),
and thus \mu is not extremal.
The proof is short and elementary, given the reconstruction bound of Theorem 3.1. Observe that
which, by the quantitative reconstruction Theorem 3.1 gives the lower bound
which proves the theorem.
Consider any central state \mu in the class of models with Hamiltonian of the form (?) fulfilling the bounds (?) and (?). Then, there exist \beta_0>0 large enough and p_1^0=p_1^0(\beta,u,U,d)>0 small enough such that for any \beta>\beta_0, p_1(\mu)<p_1^0 and for \mu\otimes\mu-a.e. pair (\omega,\omega') the extremal Gibbs measures \pi(\cdot|\omega) and \pi(\cdot|\omega') are singular with respect to each other, i.e.
The decomposition measure \alpha_\mu has no atoms, i.e. \alpha_\mu(\{\nu\})=0 for all \nu\in{\rm ex}\mathcal G(\gamma). In particular there are uncountably many extremal states which enter in the extremal decomposition of \mu.
Suppose the opposite, namely that \mu(\pi(\cdot|\omega)=\mu_0)>0, for some atom \mu_0. Then
which is a contradiction.
Define the tail-measurable function, called thinned branch-overlap, measuring how much the configuration \sigma matches with \omega on the increasing sparse volumes \Lambda_n, as
For any fixed configuration \omega, this is a tail-measurable observable w.r.t. the dependence on the spins \sigma, which takes values on the interval [0,1]. Analogous tail-measurability also holds w.r.t. to the parameter \omega.
Consider any central state \mu in the class of models with Hamiltonian of the form (?) fulfilling the bounds (?) and (?). Then there are two functions \epsilon_1(\beta)\downarrow 0 as \beta\uparrow \infty and \epsilon_2(p_1)\downarrow0 as p_1\downarrow0 such that for sparse enough sets \Lambda_n=\Lambda_n^{\rm r} the following holds:
For \mu-a.e. \omega, the thinned branch-overlap \underline\phi^\omega is \pi(\cdot |\omega)-a.s. lower bounded by
Let \beta be large enough and p_1 be small enough such that
Then there is a sequence of integers {\rm r}=(r_i)_{i\in {\mathbb N}} such that for the correspondingly defined tail-measurable observable \underline\phi^\omega we have the strict inequality
for \mu\otimes\mu-a.e. (\omega,\omega').
A boundary law \lambda for a transfer matrix Q is a family of row vectors \lambda_{vw} \in \; ]0,\infty[^{\mathbb{Z}_q} which satisfy, for all oriented pairs of nearest neighbors v,w \in V, the following consistency equation: there exists some c_{vw} >0 such that for all i \in \mathbb{Z}_q,
There is a one-to-one relation between Gibbs measures \mu which are also tree-indexed Markov chains and boundary laws \lambda.
A Gibbs measure \mu is then described via its finite-volume marginals in any \Lambda\Subset V in the form
where w_{{\Lambda}} is the unique neighbor of w\in \partial \Lambda which lies in \Lambda. The Markov chain transition operator is given for all v,w \in V by the stochastic matrix
Suppose that all eigenvalues of Q_0 are different from 1/d. Then there is a neighborhood of Q_0 such that for all Q in this neighborhood there exists a continuously differentiable solution Q\mapsto\bar x(Q) of the boundary law equation (?) for Q, which has the property that \bar x(Q_0)=1.
It is a straightforward application of the implicit function theorem. We have F(1,Q_0)=0, so we can find a neighborhood around Q_0 in matrix norm such that a perturbed solution exists, if D_x F(1,Q_0)=id_{q} -d Q_0 is invertible.
Consider the continuously differentiable solution \bar{x} of Lemma 4.1. We call central state the Markov chain Gibbs state associated to the pair (\bar x(Q),Q) solving (?).
Consider a clock model with Hamiltonian in the class (?) fulfilling the bound (?). We have the lower bounds on the eigenvalues \lambda_j of the transition operator Q_0=\frac{e^{-\beta \Phi}}{\Vert e^{-\beta \Phi} \Vert_1} of the form
Clearly \Vert e^{-\beta \Phi} \Vert_1\leq 1+(q-1)e^{-\beta u}. To estimate the spectral radius of the symmetric matrix e^{-\beta \Phi}-id_q, observing that for v with \Vert v \Vert_2=1 we have
This implies that the eigenvalues of e^{-\beta \Phi} are bounded below by 1-e^{-\beta u} (q-1), which proves the lemma.
Let \omega^0\in \Omega_0^V be a fixed reference configuration. A contour for the spin configuration \omega\in \Omega_0^V relative to \omega^0 is a pair
where the support \gamma \ {\subset} \ \{v\in V: \omega_v\neq \omega_v^0\} is a connected component of the set of incorrect points for \omega (with respect to \omega^0), and \omega_\gamma=(\omega_v)_{v\in\gamma}.
Consider \omega^0 \in \Omega and a volume \Lambda \Subset V. Then
where the sum is over contours \bar\gamma in the finite volume \Lambda.
We define the set of broken bonds of the configuration \omega\in\Omega by
and denote the set of edges attached to \gamma\subset V by
We have the following upper bound on the number of connected subsets of vertices of \mathcal T^d=(V,E):
Let \mu be a central state of any model with Hamiltonian in the class (?) fulfilling the bounds (?) and (?). Let \bar\gamma=(\omega'_{\gamma},\gamma) be a contour relative to the fixed configuration \omega. Denote \omega'=(\omega'_{\gamma}\omega_{\gamma^c}) the corresponding excited spin configuration. Then the excess energy satisfies
In particular, if there exist \delta\geq0 and \epsilon>0 such that | D(\omega)\cap E(\gamma) | \leq\delta|\gamma| and \Vert\Psi\Vert_\infty\leq\epsilon/2 then
Consider any model in the class (?). Let
where u,U\in{\mathbb R}^+ are the bounds on energy costs defined in (?) and d is the branching number of the tree.
Denote by B_v the bad event that there exists a contour around v with respect to \omega which does not have good enough excess energy in the sense that
Suppose that \mu is a homogeneous tree-indexed Markov chain for a Hamiltonian in the class (?) fulfilling the bounds (?) and (?). Let p_1=p_1(\mu) be its lazyness parameter as in Defintion 2.2. Let \delta_0 as in (4.5). Define \lambda(p_1)=\lambda(p_1,u,U,d) as
Note that \lambda(p_1)\uparrow\infty or equivalently e^{-\lambda(p_1)}\downarrow 0 \text{ as } p_1\downarrow 0. In particular, if \mu_\beta is the free state of the q-clock model at inverse temperature \beta (see (?)),
Let \mu be the central state of any model with Hamiltonian in the class (?) fulfilling the bounds (?) and (?). Then for \beta large enough we have, for \mu-almost every \omega
where the event B_v is defined in (4.5), and \epsilon_1(\beta)\downarrow0 as \beta\uparrow\infty.
Moreover,
where \lambda(p_1) is defined in (4.6) and for p_1 small enough, for any v\in V
where \epsilon_2(p_1)\downarrow0 as p_1\downarrow0.
We decompose
and we use the infinite-volume version of the Peierls bound of Lemma 4.3 (removing the restriction to finite volume always gives an upper bound) applied to the complement of a bad event at v to bound the probability of a mismatch at v by the sum over contours attached at v. This gives
where \rho[\omega](\bar \gamma) is the contour activity relative to \omega. Note that \omega is not assumed to be a local ground state (in the sense of [CoKuLe22]), but we may still use the bound on the excess energy of a local perturbation given in Lemma 4.5.
This delivers, using the assumptions on the potential in the class (?) and the Definition 4.5 of B_v,
Using the entropy bound of Lemma 4.4 we deduce that there exists C',c'>0 such that
for \beta>4\log((q-1)(d+1)^2)/(u(d-1)), which proves (?).
Next we use the exponential Markov inequality to bound \mu(B_v(\gamma)) and sum over the contours afterwards to bound \mu(B_v), as will be described now.
It suffices to fix any x\in \gamma, look at the conditional measure \mu_a(\cdot ):=\mu(\cdot |\sigma_x=a ) for any fixed a\in {\mathbb Z}_q, and find an upper bound which does not depend on a.
Using the Markov chain property of the measure \mu_a, successive applications of the homogeneous transition matrix P yield
which follows from
and the bound |E( \gamma) |\leq (d+1)|\gamma|. Optimizing over t leads to the bound
which proves (?), since the definition (4.6) of \exp(-\lambda(p_1)) is made to be the value of the above infimum. Finally using the entropy bound of Lemma 4.4 on the number of contours of fixed length which are attached to a given point, the estimate for \mu(B_v) follows. Indeed, there exists C>0 such that for any c\in(0,1) we have
for p_1 sufficiently small, which proves (?).
Consider the central state \mu of any model with Hamiltonian in the class (?) fulfilling the bounds (?) and (?). Suppose that the transition matrix P of the state \mu is irreducible and aperiodic. Recall the Definition 4.5 of a bad event B_v at v\in V and the notation for its expectation \mu(B_v)=:m(p_1) (which does not depend on v). Let
Then for small enough p_1, there exist C>0 such that for any c_1\in(0,\frac13) and c_2\in(0,1), for any u,v\in V,
where |v-u| is the graph distance in the tree and
where \lambda(p_1) is defined in (4.6) and \lambda_2(P) is the second largest eigenvalue (in modulus) of the transition matrix P of \mu.
We first define several objects and events which are depicted on Figure ?.
Denote by the half-trees H_u(r) and H_v(r) the sets
We will choose r=\lfloor |u-v|/3\rfloor. Clearly H_u and H_v are disjoint, with their distance growing proportionally to |u-v|. Define
The sets B^1_u and B^1_v describe bad events at u (resp. v) which are produced by contours included in H_u (resp. H_v). The sets B^2_u and B^2_v describe bad events which are not produced by the previous contours and may connect u and v.
Bad event caused by long contours are exponentially improbable in the minimal length of such contours. Namely, by the Key Lemma 4.6 and the entropy bound given in Lemma 4.4, there exist C_1>0 such that for any c_1\in(0,\frac13) we have
Spelling out the correlation function between the bad events in terms of the decomposition (?), we therefore have
It remains to control the decay of |\mu(B^1_u;B^1_v)|. We would have the exact decorrelation \mu(B^1_u;B^1_v)=0 in the free state of a clock model, but for an non-symmetric model there is no reason to assume that such decorrelation holds. Nevertheless, we may use the Markov chain property of \mu along the path from u to v and the exponential convergence of the corresponding 1d-Markov chain, to arrive at the decorrelation which is exponential in the distance |u-v| (with our irreducibility and aperiodicity assumptions).
Indeed, choose x=x(u,v) to be the (or one of the at most two) middle sites between u and v (which means that the distances |u-x| and |v-x| differ at most by one). From the tree-indexed Markov chain property of \mu, using conditional independence, we have
Let \Delta_{u,x}(a):=\mu(B^1_u|\sigma_x=a)-\mu(B^1_u) and similarly for v. We have
By symmetry it suffices to bound |\Delta_{u,x}(a)|. Choose y on the path [x,u], between x and u, such that it is the closest to u with the property that all the contours appearing in the definition of B^1_u do not touch (x,y].
To control the difference between \mu(B^1_u|\sigma_x=a) and \mu(B^1_u) we use the exponential relaxation of the homogenous 1d Markov chain which arises as restriction of \mu, on the path from x to y. There exists C_2>0 such that for any c_2\in(0,1),
where the last line follows e.g. from [Bremaud, Example 4.3.9: Rates of Convergence via the Perron-Frobenius Theorem]. As by construction |y-x|\sim \frac{1}{6}|u-v|, as |u-v|\uparrow\infty, this implies the desired exponential decorrelation of |\mu(B^1_u;B^1_v)|, and concludes the proof.
Consider the central state \mu of any model with Hamiltonian of the form (?) fulfilling the bounds (?) and (?). For any sequence of subsets \Lambda_n=\Lambda_n^r of a branch of the tree such that \sum_{n=1}^\infty {|\Lambda_n|}^{-1}<\infty and for \mu-almost every \omega, the following limit holds
where m(p_1)=\mu(B_0).
The almost sure limit holds, if we can show that for any fixed \delta>0 the number of indices n for which
is finite, for \mu-almost every \omega. By Borel-Cantelli it suffices to show that
For this we use the quadratic Chebychev inequality to bound the probability in the above sum by
Recalling the definition (4.7) of site covariances \mbox{Cov}(u,v) and using Lemma 4.7, we deduce that there exists C>0 such that
since for any fixed vertex v\in V, the sum \sum_{u>v}\mbox{Cov}(u,v)\leq C'<\infty uniformly in v by Lemma 4.7. This proves the existence of the almost sure limit of empirical sums of indicators of bad points, along the quadratic volume sequences defined in (?).
Consider the central state \mu of any model with Hamiltonian in the class (?) fulfilling the bounds (?) and (?).
There exist a sequence of spacings r=(r_i)_{i\in {\mathbb N}}, in general depending on model parameters, such that for any sequence of subsets \Lambda_n=\Lambda^r_n of a branch of the tree such that \sum_{n=1}^\infty
{|\Lambda_n|}^{-1}<\infty, the following holds.
For \mu-almost every \omega, for \pi(\cdot |\omega)-almost every realizations of \sigma we have
It suffices to prove that, for any \delta>0 we have, for \mu-a.e. \omega,
Using the Borel-Cantelli lemma it suffices to show that, for \mu-a.e. \omega,
Now by the quadratic Chebychev inequality, it is enough to show that
holds for \mu-a.e. \omega. We would like to use the abstract fact that \pi(\cdot|\omega) is an extremal Gibbs measure, and hence decorrelates in the sense of (?). Indeed, by extremality of \pi(\cdot |\omega) at fixed \omega, for any fixed u and any sequence v\uparrow\infty (meaning that v leaves any finite ball) the following non-quantative result holds: For \mu-a.e. \omega,
The problem is that the rate of convergence may depend on \omega, while we want a speed which is uniform in all the extremals (in order for the limit (4.9) to hold along the same sparse sequence of volumes for \mu-a.e. \omega). An explicit analysis is difficult, due to the lack of spatial homogeneity. To bypass this difficulty, note that to ensure (Proof) it is sufficient to ensure that
Using monotone convergence for non-negative functions to interchange the \mu-integral and the infinite sum, we can deduce this from
By dominated convergence we have, for any u\in V,
We look at the off-diagonal terms in the double-sum (Proof) as in (Proof). Now we can achieve the desired summability of
by iteratively choosing v_i given v_1, \dots, v_{i-1} on the same branch so large that
The above choice provides the following upper bound of (Proof) as
which is finite.
On the one hand we have the following lower bound from Theorem 3.4 and monotone convergence. For sparse enough sets \Lambda_n=\Lambda_n^{\rm r}, for \mu almost every \omega,
On the other hand, we use
to write the upper bound
The first term in the r.h.s. is equal to {1}/{q} for \mu\otimes\mu-a.e. pair (\omega,\omega'), in the case of clock models. Indeed, the Borel-Cantelli argument, together with the Chebychev inequality as in the proof of Lemma 4.8 shows that the desired strong law of large numbers follows (since exponential decorrelation of the local events (1_{\omega_v=a})_{v\in\Lambda_n^r} follows from exponential decorrelation of the 1d Markov chains as in (?)). In the general case of central states this term equals \sum_{a\in {\mathbb Z}_q}\mu(\sigma_0=a)^2; it is close to 1/q for perturbed free states of clock models, see Section ?.
The second term in the r.h.s. is \pi(\cdot | \omega')-a.s. controlled via Theorem 3.4 by
So the statement of the theorem follows once
which concludes the proof.
The text on this page is from arXiv:2310.11101v1 by Loren Coquille, Christof Kuelske and Arnaud Le Ny, under the licence CC BY 4.0. The corpus made these changes: it read the items out of the LaTeX source, it expanded the macros of the paper, it put each proof after its statement, and it converted the text to HTML. The formulas are the TeX of the paper, and KaTeX renders them. A formula or a reference can look different from the PDF. The page is published under CC BY 4.0.
The record of each item is in analysis/objects/2310/2310.11101.jsonl.