Title:Optimizing weighted ensemble sampling of steady states

Abstract:We show how to optimize resampling methods for Markov chains based on stratification, our aim being to reduce variance in steady-state computations. We consider collections of replicas of a Markov chain, with each replica carrying a weight, and where the replicas occupy a number of strata, or bins. We refer to this as as weighted ensemble (WE), after a resampling method from the computational chemistry literature. Traditionally, WE refers to resampling based on an ad hoc rule: maintaining a roughly spatially uniform distribution of replicas in the bins. We derive, from first principles, a strategy for optimizing the replica allocation for a pre-selected set of bins. In numerical tests on simple problems, our optimized allocation strategy outperforms uniform allocation and naive sampling. We further propose a synthetic variance calculation that may be used to help choose WE parameters, including the collection of bins.