JAXNS: a high-performance nested sampling package based on JAX

Since its debut by John Skilling in 2004, nested sampling has proven a valuable tool to the scientist, providing hypothesis evidence calculations and parameter inference for complicated posterior distributions, particularly in the field of astronomy. Due to its computational complexity and long-running nature, in the past, nested sampling has been reserved for offline-type Bayesian inference, leaving tools such as variational inference and MCMC for online-type, time-constrained, Bayesian computations. These tools do not easily handle complicated multi-modal posteriors, discrete random variables, and posteriors lacking gradients, nor do they enable practical calculations of the Bayesian evidence. An opening thus remains for a high-performance out-of-the-box nested sampling package that can close the gap in computational time, and let nested sampling become common place in the data science toolbox. We present JAX-based nested sampling (JAXNS), a high-performance nested sampling package written in XLA-primitives using JAX, and show that it is several orders of magnitude faster than the currently available nested sampling implementations of PolyChord, MultiNEST, and dynesty, while maintaining the same accuracy of evidence calculation. The JAXNS package is publically available at \url{https://github.com/joshuaalbert/jaxns}.

PDF Abstract