Accelerating multimodel Bayesian inference, model selection, and systematic studies for gravitational wave astronomy

Abstract

Gravitational wave models are used to infer the properties of black holes in merging binaries from the observed gravitational wave signals through Bayesian inference. Although we have access to a large collection of signal models that are sufficiently accurate to infer the properties of black holes, for some signals, small discrepancies in the models lead to systematic differences in the inferred properties. In order to provide a single estimate for the properties of the black holes, it is preferable to marginalize over the model uncertainty. Bayesian model averaging is a commonly used technique to marginalize over multiple models, however, it is computationally expensive. An elegant solution is to simultaneously infer the model and model properties in a joint Bayesian analysis. In this work we demonstrate that a joint Bayesian analysis can not only accelerate but also account for model-dependent systematic differences in the inferred black hole properties. We verify this technique by analysing 100 randomly chosen simulated signals and also the real gravitational wave signal GW200129_065458. We find that not only do we infer statistically identical properties as those obtained using Bayesian model averaging, but we can sample over a set of three models on average $2.5\times$ faster. In other words, a joint Bayesian analysis that marginalizes over three models takes on average only $20\%$ more time than a single model analysis. We then demonstrate that this technique can be used to accurately and efficiently quantify the support for one model over another, thereby assisting in Bayesian model selection.