Gravitational wave populations and cosmology with neural posterior estimation
We apply neural posterior estimation for fast-and-accurate hierarchical Bayesian inference of gravitational wave populations. We use a normalizing flow to estimate directly the population hyper-parameters from a collection of individual source observations. This approach provides complete freedom in event representation, automatic inclusion of selection effects, and (in contrast to likelihood estimation) without the need for stochastic samplers to obtain posterior samples. Since the number of events may be unknown when the network is trained, we split into sub-population analyses that we later recombine; this allows for fast sequential analyses as additional events are observed. We demonstrate our method on a toy problem of dark siren cosmology, and show that inference takes just a few minutes and scales to $\sim 600$ events before performance degrades. We argue that neural posterior estimation therefore represents a promising avenue for population inference with large numbers of events.