A new way to sample data at low temperatures

Researchers developed a framework to speed up sampling in systems where competing phases create bottlenecks.

A new framework has been developed for sampling at low temperatures using Glauber dynamics. This method helps when competing modes, or phases, create bottlenecks that make mixing exponentially slow.

The framework relates convergence within a phase to the geometry of disagreements between two independent configurations. These configurations are drawn from that phase.

The approach was applied to three models on random regular graphs. For the hard-core model on random bipartite dd-regular graphs, Glauber dynamics mixes within n3/2+o(1)n^{3/2+o(1)} steps. This occurs in the non-uniqueness regime λ>λc(d)\lambda > \lambda_c(d) when starting from one-sided independent sets.

In the random-cluster model on random dd-regular graphs, Glauber dynamics mixes within the ordered phase in n3/2+o(1)n^{3/2+o(1)} steps for β≥βc\beta \geq \beta_c. It mixes within the disordered phase in n3/2+o(1)n^{3/2+o(1)} steps for β≤βc\beta \leq \beta_c when starting from the all-in or all-out edge sets.

For the ferromagnetic Ising model on random dd-regular graphs, Glauber dynamics mixes within a phase in n3/2+o(1)n^{3/2+o(1)} steps. This works throughout the low-temperature regime β>βc(d)\beta > \beta_c(d) using all-plus or all-minus initialisations.

An n2+o(1)n^{2+o(1)}-time algorithm can combine phase samples into samples from the full distribution. This avoids time-costly approximations of partition functions.

Key facts

  • Glauber dynamics mixes within n3/2+o(1)n^{3/2+o(1)} steps for several models on random dd-regular graphs.
  • The framework uses the geometry of disagreements between two independent configurations.
  • An n2+o(1)n^{2+o(1)}-time algorithm combines phase samples to avoid approximating partition functions.

All sections

Edition