Curse-of-dimensionality revisited: Collapse of the particle filter in very large scale systems

TL;DR

This paper provides theoretical insights into why particle filters fail in very large systems, showing that weights tend to collapse to one as system size and sample size grow, especially under certain growth conditions.

Contribution

It offers a theoretical analysis of weight collapse in high-dimensional particle filters, highlighting conditions under which the weights converge to one, indicating failure.

Findings

Weights converge to one as system and sample size increase

Collapse occurs unless ensemble size grows super-exponentially

Results apply to models with Gaussian and Cauchy likelihoods

Abstract

It has been widely realized that Monte Carlo methods (approximation via a sample ensemble) may fail in large scale systems. This work offers some theoretical insight into this phenomenon in the context of the particle filter. We demonstrate that the maximum of the weights associated with the sample ensemble converges to one as both the sample size and the system dimension tends to infinity. Specifically, under fairly weak assumptions, if the ensemble size grows sub-exponentially in the cube root of the system dimension, the convergence holds for a single update step in state-space models with independent and identically distributed kernels. Further, in an important special case, more refined arguments show (and our simulations suggest) that the convergence to unity occurs unless the ensemble grows super-exponentially in the system dimension. The weight singularity is also established in models with more general multivariate likelihoods, e.g. Gaussian and Cauchy. Although presented in the context of atmospheric data assimilation for numerical weather prediction, our results are generally valid for high-dimensional particle filters.