When dealing with stochastic processes on graphs, two elements can render the estimation of observables computationally hard: first, the presence of reweighting terms that bias the probability measure toward rare events and, second, the possibility for variables to return to a previously visited state. While in the past, several approximation methods have successfully dealt with either of these two factors, our work aims at addressing both simultaneously.

This talk will introduce the Matrix Product Belief Propagation approximation, which can be applied to Markov processes biased by arbitrary reweighting factors that concentrate most of the probability mass on a small subspace of trajectories. Our approach builds on a version of the dynamic cavity method where distributions of pairs of trajectories are parametrized by matrix product states, a well-known tool in the field of many-body quantum systems. Two applications will be mentioned: Bayesian inference on epidemic models and the analysis of Glauber dynamics of Ising models. 

报告人简介

Stefano Crotti is a PhD candidate in physics at the Polytechnic University of Turin, Italy, under the supervision of prof. Alfredo Braunstein. His research focuses on employing techniques from the statistical physics of disordered systems to problems in optimization and inference. In the past two years he has been working on how to reconstruct partially observed realizations of stochastic dynamics on graphs.

邀请人

张潘 研究员