Region graph partition function expansion and approximate free energy landscapes: Theory and some numerical results

TL;DR

This paper develops a theoretical framework using region graph partition function expansion to construct approximate free energy landscapes for complex graphical models, validated through numerical simulations on spin systems.

Contribution

It introduces a novel approach to approximate free energy landscapes using region graph methods, improving upon traditional Bethe-Peierls approximations.

Findings

Enhanced accuracy over Bethe-Peierls approximation

Identification of collective domains in disordered systems

Insights into low-temperature glass-like behaviors

Abstract

Graphical models for finite-dimensional spin glasses and real-world combinatorial optimization and satisfaction problems usually have an abundant number of short loops. The cluster variation method and its extension, the region graph method, are theoretical approaches for treating the complicated short-loop-induced local correlations. For graphical models represented by non-redundant or redundant region graphs, approximate free energy landscapes are constructed in this paper through the mathematical framework of region graph partition function expansion. Several free energy functionals are obtained, each of which use a set of probability distribution functions or functionals as order parameters. These probability distribution function/functionals are required to satisfy the region graph belief-propagation equation or the region graph survey-propagation equation to ensure vanishing correction contributions of region subgraphs with dangling edges. As a simple application of the general theory, we perform region graph belief-propagation simulations on the square-lattice ferromagnetic Ising model and the Edwards-Anderson model. Considerable improvements over the conventional Bethe-Peierls approximation are achieved. Collective domains of different sizes in the disordered and frustrated square lattice are identified by the message-passing procedure. Such collective domains and the frustrations among them are responsible for the low-temperature glass-like dynamical behaviors of the system.