Combined-chain nested sampling for efficient Bayesian model comparison

Model comparison problems arise in many fields of science and engineering, including signal processing. In these problems, we wish to quantify how well each of a set of possible models describes a set of observations. Many numerical techniques exist to perform model comparison, but this paper focuses on nested sampling, which is a numerical integration algorithm for evaluating probabilities of models. The original formulation of nested sampling is a strictly sequential algorithm. Most modern advances in computing are via parallel processing, however, and we therefore present a novel method for parallelizing nested sampling. This paper sets out the mathematical foundation for this parallelization, as well as ideas for implementing it. Three examples demonstrate the effectiveness of the present parallel technique in realistic scientific and engineering data analysis problems.