Abstract

Particle flow has been described in the literature as a very efficient particle filter with no need for resampling. It has been documented as achieving impressive results against state-of-the-art baselines when applied to high dimensional dynamic inference problems. However, the existing literature makes it difficult to understand how, why and when particle flow works. This talk will put particle flow on a solid theoretical footing. Links to Langevin diffusions, marginal particle filters, SMC samplers and data augmentation will be highlighted. The conditions that result in no resampling being necessary will be made clear. To highlight the benefits, results will demonstrate the efficiency of the technique in both dynamic contexts and static contexts (where particle flow offers untapped potential to outperform, for example, MCMC and SMC samplers).