Derivations of Survey Propagation and the road to generalitzations.

by Alejandro Lage Lage Castellanos (Università del Havana (Cuba))

Wednesday 10 Feb 2016, 14:30 → 16:30 Europe/Rome

Aula Careri (Dip. di Fisica - Edificio G. Marconi)

The mean field character of SP equations make it hard to apply (or even converge) them in finite dimensional lattices. We can get more precise approximations to finite dimensional systems by including larger regions in the approximation. This path resulted in Generalized Belief Propagation equations for the Cluster Variational Method. However, any practical generalization of SP equations to such context remains a challenge. We show three different formal approaches to SP equations, including one of our own, that we call variational. Among them, two, the one based on the Replica Method of Rizzo et al. [2010] and our variational one, leads easily to generalizations of SP equations to CVM. We will also discuss how to create gauge-free CVM approximations from scratch, and finally discuss the challenges raised while trying to do GSP.