Markov Kernels Local Aggregation for Noise Vanishing Distribution Sampling

Abstract

A novel strategy that combines a given collection of $\pi$-reversible Markov kernels is proposed. At each Markov transition, one of the available kernels is selected via a state-dependent probability distribution. In contrast to random-scan type approaches that assume a constant (i.e. state-independent) selection probability distribution, the state-dependent distribution is specified so as to privilege moving according to a kernel which is relevant for the local topology of the target distribution. This approach leverages paths or other low dimensional manifolds that are typically present in noise vanishing distributions. Some examples for which we show (theoretically or empirically) that a locally-weighted aggregation converges substantially faster and yields smaller asymptotic variances than an equivalent random-scan algorithm are provided.