Path storage in the particle filter
In the train to Annecy, I read the recently arXived paper by my former PhD student Pierre Jacob (now at NUS), along with Lawrence Murray (Perth), and Sylvain Rubenthaler (Nice), where they obtain precise degeneracy rates of the regular particle filter applied to hidden Markov models with a compact observation space, precise enough to consider storing the entire paths at a linear occupancy rate. Interestingly, the distance to the most common ancestor is of order N log N, if N is the number of particles. And the number of nodes is O(N log N) as well. This means indeed that the whole paths can be stored, which offers a lot of potential in terms of Rao-Blackwellisation and parallelisation. I was first bemused by a silly misunderstanding about the purpose of the algorithm: it is directed at inference at the current time index, not over the whole past and not over the parameters of the model for, else how could we consider the algorithm has converged when it degenerates to a single path at some finite horizon the past? Pierre has also written a further comment of the paper on Statistfaction.