Massively parallel Bayesian inference for transient gravitational-wave astronomy

TL;DR

This paper introduces a scalable, parallelized nested sampling method for Bayesian inference in gravitational-wave astronomy, significantly reducing computation time and enabling flexible, accurate analysis of complex astrophysical models.

Contribution

It presents a novel parallelized nested sampling approach that scales efficiently with high-performance computing resources, improving inference speed without sacrificing model flexibility.

Findings

Wall-time scales almost linearly with the number of CPUs.

Enables use of complex models without additional optimization.

Implementation available in open-source library pBilby.

Abstract

Understanding the properties of transient gravitational waves and their sources is of broad interest in physics and astronomy. Bayesian inference is the standard framework for astro-physical measurement in transient gravitational-wave astronomy. Usually, stochastic sampling algorithms are used to estimate posterior probability distributions over the parameter spaces of models describing experimental data. The most physically accurate models typically come with a large computational overhead which can render data analysis extremely time consuming, or possibly even prohibitive. In some cases highly specialized optimizations can mitigate these issues, though they can be difficult to implement, as well as to generalize to arbitrary models of the data. Here, we propose an accurate, flexible and scalable method for astro-physical inference: parallelized nested sampling. The reduction in the wall-time of inference scales almost linearly with the number of parallel processes running on a high-performance computing cluster. By utilizing a pool of several hundreds or thousands of CPUs in a high-performance cluster, the large wall times of many astrophysical inferences can be alleviated while simultaneously ensuring that any gravitational-wave signal model can be used "out of the box", i.e., without additional optimization or approximation. Our method will be useful to both the LIGO-Virgo-KAGRA collaborations and the wider scientific community performing astrophysical analyses on gravitational waves. An implementation is available in the open source gravitational-wave inference library $\texttt{pBilby}$ (parallel $\texttt{bilby}$).