Simplifying Generalized Belief Propagation on Redundant Region Graphs

TL;DR

This paper introduces a simplified derivation of generalized belief propagation (GBP) equations, removing redundancies to improve convergence and applying it successfully to Ising and spin glass models.

Contribution

It presents a systematic method to eliminate redundancies in GBP equations, enhancing their applicability to complex many-body systems.

Findings

SGBP achieves satisfactory performance on Ising and spin glass models.

Redundant GBP equations can be neglected without affecting results.

The new derivation simplifies implementation and improves convergence.

Abstract

The cluster variation method has been developed into a general theoretical framework for treating short-range correlations in many-body systems after it was first proposed by Kikuchi in 1951. On the numerical side, a message-passing approach called generalized belief propagation (GBP) was proposed by Yedidia, Freeman and Weiss about a decade ago as a way of computing the minimal value of the cluster variational free energy and the marginal distributions of clusters of variables. However the GBP equations are often redundant, and it is quite a non-trivial task to make the GBP iteration converges to a fixed point. These drawbacks hinder the application of the GBP approach to finite-dimensional frustrated and disordered systems. In this work we report an alternative and simple derivation of the GBP equations starting from the partition function expression. Based on this derivation we propose a natural and systematic way of removing the redundance of the GBP equations. We apply the simplified generalized belief propagation (SGBP) equations to the two-dimensional and the three-dimensional ferromagnetic Ising model and Edwards-Anderson spin glass model. The numerical results confirm that the SGBP message-passing approach is able to achieve satisfactory performance on these model systems. We also suggest that a subset of the SGBP equations can be neglected in the numerical iteration process without affecting the final results.