Gaussian approximation of dynamic cavity equations for linearly-coupled stochastic dynamics
Authors:
Mattia Tarabolo,
Luca Dall'Asta
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 num…
▽ More
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.
△ Less
Submitted 4 June, 2025; v1 submitted 20 June, 2024;
originally announced June 2024.
Small-Coupling Dynamic Cavity: a Bayesian mean-field framework for epidemic inference
Authors:
Alfredo Braunstein,
Giovanni Catania,
Luca Dall'Asta,
Matteo Mariani,
Fabio Mazza,
Mattia Tarabolo
Abstract:
We present the Small-Coupling Dynamic Cavity (SCDC) method, a novel generalized mean-field approximation for epidemic inference and risk assessment within a fully Bayesian framework. SCDC accounts for non-causal effects of observations and uses a graphical model representation of epidemic processes to derive self-consistent equations for edge probability marginals. A small-coupling expansion yield…
▽ More
We present the Small-Coupling Dynamic Cavity (SCDC) method, a novel generalized mean-field approximation for epidemic inference and risk assessment within a fully Bayesian framework. SCDC accounts for non-causal effects of observations and uses a graphical model representation of epidemic processes to derive self-consistent equations for edge probability marginals. A small-coupling expansion yields time-dependent cavity messages capturing individual infection probabilities and observational conditioning. With linear computational cost per iteration in the epidemic duration, SCDC is particularly efficient and valid even for recurrent epidemic processes, where standard methods are exponentially complex. Tested on synthetic networks, it matches Belief Propagation in accuracy and outperforms individual-based mean-field methods. Notably, despite being derived as a small-infectiousness expansion, SCDC maintains good accuracy even for relatively large infection probabilities. While convergence issues may arise on graphs with long-range correlations, SCDC reliably estimates risk. Future extensions include non-Markovian models and higher-order terms in the dynamic cavity framework.
△ Less
Submitted 10 April, 2025; v1 submitted 6 June, 2023;
originally announced June 2023.
