The collapse of particle filters

Peter J. Bickel
Statistics Department
University of California, Berkeley

Abstract

        Particle filters are Monte Carlo approximations to the conditional distribution of the hidden process in HMM and state space models.  They have become very popular in situations where the state space is large.  Unfortunately it's been observed in climate forecasting and other situations, where the dimension of the state space is large that they can fail after 1 step of the approximation.  Using simple models I will try to explain why and when this phenomenon can be expected. 

This is joint work with Thomas Bengtsson and Bo Li.