Massively parallel transdimensional sampling with data assimilation: an application to LISA galactic binaries

Abstract

Transdimensional Bayesian inference is becoming a key ingredient in gravitational-wave astronomy. In many relevant applications, the number of components needed to describe the data is not known a priori and must be inferred jointly with their parameters. We present a transdimensional Sequential Monte Carlo (SMC) framework designed to exploit massive parallelism. The method evolves a population of particles through a sequence of tempered distributions and uses the No-U-Turn Sampler for the exploration of fixed-dimensional parameter space while reversible-jump birth and death moves allow the number of components to vary. The same SMC construction also enables posterior samples obtained from a shorter data segment to be updated as additional data become available, enhancing the overall efficiency of the analysis. We validate the approach on two proof-of-concept problems: the recovery of a sequence of Gaussian pulses and a simplified LISA Galactic-binary inference problem. In both cases, the method recovers the injected number of components and produces consistent posterior estimates. These results indicate that the proposed method is a promising avenue toward scalable and parallel Bayesian transdimensional inference for current and future gravitational-wave astronomy.