Gaussian approximation of dynamic cavity equations for linearly-coupled stochastic dynamics

TL;DR

This paper introduces a Gaussian approximation framework for analyzing the complex dynamics of sparse disordered systems modeled by linearly-coupled stochastic differential equations, enabling efficient analysis and message-passing algorithms.

Contribution

It develops a novel second-order truncation approach to derive Gaussian cavity equations for sparse systems, bridging dynamical mean-field theory and message-passing methods.

Findings

Exact dynamical equations for linear systems with additive noise.

Retrieves classical spectral density results for sparse random matrices.

Applicable to sparse graphs and can incorporate non-linearities and constraints.

Abstract

Stochastic dynamics on sparse graphs and disordered systems often lead to complex behaviors characterized by heterogeneity in time and spatial scales, slow relaxation, localization, and aging phenomena. The mathematical tools and approximation techniques required to analyze these complex systems are still under development, posing significant technical challenges and resulting in a reliance on numerical simulations. We introduce a novel computational framework for investigating the dynamics of sparse disordered systems with continuous degrees of freedom. Starting with a graphical model representation of the dynamic partition function for a system of linearly-coupled stochastic differential equations, we use dynamic cavity equations on locally tree-like factor graphs to approximate the stochastic measure. Here, cavity marginals are identified with local functionals of single-site trajectories. Our primary approximation involves a second-order truncation of a small-coupling expansion, leading to a Gaussian form for the cavity marginals. For linear dynamics with additive noise, this method yields a closed set of causal integro-differential equations for cavity versions of one-time and two-time averages. These equations provide an exact dynamical description within the local tree-like approximation, retrieving classical results for the spectral density of sparse random matrices. Global constraints, non-linear forces, and state-dependent noise terms can be addressed using a self-consistent perturbative closure technique. The resulting equations resemble those of dynamical mean-field theory in the mode-coupling approximation used for fully-connected models. However, due to their cavity formulation, the present method can also be applied to ensembles of sparse random graphs and employed as a message-passing algorithm on specific graph instances.