Factorized Parameter Estimation for Real-Time Gravitational Wave Inference

Abstract

We present a parameter estimation framework for gravitational wave (GW) signals that brings together several ideas to accelerate the inference process. First, we use the relative binning algorithm to evaluate the signal-to-noise-ratio timeseries in each detector for a given choice of intrinsic parameters. Second, we decouple the estimation of the intrinsic parameters (such as masses and spins of the components) from that of the extrinsic parameters (such as distance, orientation, and sky location) that describe a binary compact object coalescence. We achieve this by semi-analytically marginalizing the posterior distribution over extrinsic parameters without repeatedly evaluating the waveform for a fixed set of intrinsic parameters. Finally, we augment samples of intrinsic parameters with extrinsic parameters drawn from their appropriate conditional distributions. We implement the method for binaries with aligned spins, restricted to the quadrupole mode of the signal. Using simulated GW signals, we demonstrate that the method produces full eleven-dimensional posteriors that match those from standard Bayesian inference. Our framework takes only ~200 seconds to analyze a typical binary-black-hole signal and ~250 seconds to analyze a typical binary-neutron-star signal using one computing core. Such real-time and accurate estimation of the binary source properties will greatly aid the interpretation of triggers from gravitational wave searches, as well as the search for possible electromagnetic counterparts. We make the framework publicly available via the GW inference package cogwheel.