Tensor Networks contraction and the Belief Propagation algorithm

TL;DR

This paper connects belief propagation, a message-passing algorithm for graphical models, with tensor network contraction, revealing that simple tensor network algorithms are akin to mean field approximations in graphical models.

Contribution

It demonstrates how belief propagation can be adapted for tensor network contraction and shows the equivalence of the Simple-Update algorithm to the Bethe-Peierls approximation.

Findings

Simple-Update is equivalent to Bethe-Peierls approximation.

Belief Propagation can be adapted for PEPS tensor networks.

This connection enables improvements in tensor network algorithms.

Abstract

Belief Propagation is a well-studied message-passing algorithm that runs over graphical models and can be used for approximate inference and approximation of local marginals. The resulting approximations are equivalent to the Bethe-Peierls approximation of statistical mechanics. Here we show how this algorithm can be adapted to the world of PEPS tensor networks and used as an approximate contraction scheme. We further show that the resultant approximation is equivalent to the ``mean field'' approximation that is used in the Simple-Update algorithm, thereby showing that the latter is a essentially the Bethe-Peierls approximation. This shows that one of the simplest approximate contraction algorithms for tensor networks is equivalent to one of the simplest schemes for approximating marginals in graphical models in general, and paves the way for using improvements of BP as tensor networks algorithms.