Glauber dynamics for Ising models on random regular graphs: cut-off and metastability

TL;DR

This paper investigates the mixing times of Glauber dynamics for the Ising model on random regular graphs, revealing phase transitions, metastability, and cutoff phenomena depending on temperature and external magnetic field.

Contribution

It provides a detailed analysis of the cutoff and metastability phenomena for Glauber dynamics on random regular graphs, connecting these to phase transitions and critical external fields.

Findings

Metastability occurs for certain parameters with exponential mixing times.

Cutoff phenomenon at order n log n in specific regimes.

Critical external field matches the threshold for unique Gibbs measure on the d-ary tree.

Abstract

Consider random $d$-regular graphs, i.e., random graphs such that there are exactly $d$ edges from each vertex for some $d\ge 3$. We study both the configuration model version of this graph, which has occasional multi-edges and self-loops, as well as the simple version of it, which is a $d$-regular graph chosen uniformly at random from the collection of all $d$-regular graphs. In this paper, we discuss mixing times of Glauber dynamics for the Ising model with an external magnetic field on a random $d$-regular graph, both in the quenched as well as the annealed settings. Let $\beta$ be the inverse temperature, $\beta_c$ be the critical temperature and $B$ be the external magnetic field. Concerning the annealed measure, we show that for $\beta > \beta_c$ there exists $\hat{B}_c(\beta)\in (0,\infty)$ such that the model is metastable (i.e., the mixing time is exponential in the graph size $n$) when $\beta> \beta_c$ and $0 \leq B < \hat{B}_c(\beta)$, whereas it exhibits the cut-off phenomenon at $c_\star n \log n$ with a window of order $n$ when $\beta < \beta_c$ or $\beta > \beta_c$ and $B>\hat{B}_c(\beta)$. Interestingly, $\hat{B}_c(\beta)$ coincides with the critical external field of the Ising model on the $d$-ary tree (namely, above which the model has a unique Gibbs measure). Concerning the quenched measure, we show that there exists $B_c(\beta)$ with $B_c(\beta) \leq \hat{B}_c(\beta)$ such that for $\beta> \beta_c$, the mixing time is at least exponential along some subsequence $(n_k)_{k\geq 1}$ when $0 \leq B < B_c(\beta)$, whereas it is less than or equal to $Cn\log n$ when $B>\hat{B}_c(\beta)$. The quenched results also hold for the model conditioned on simplicity, for the annealed results this is unclear.