Incorporating neutron star physics into gravitational wave inference with neural priors

TL;DR

The paper tackles the problem that gravitational-wave parameter estimation often relies on agnostic priors that underutilize neutron-star (NS) physics. It introduces neural priors built from NS population models and equation-of-state (EOS) constraints using normalizing flows to model joint distributions over NS masses and tidal deformabilities, enabling EOS-informed GW inference and Bayesian model selection. Applied to GW170817, GW190425, and GW230529, the approach yields clear source classifications (BNS vs NSBH) and tighter posteriors, with GW170817 favoring a soft EOS and GW230529 favoring NSBH, while distance estimates shift in ways consistent with the EOS-informed priors. This data-driven framework integrates NS physics into GW analyses, enabling more informed inferences as NS theory and observations advance, and is readily extensible to future detectors and multimessenger data.

Abstract

Bayesian inference, widely used in gravitational-wave parameter estimation, depends on the choice of priors, i.e., on our previously existing knowledge. However, to investigate neutron star mergers, priors are often chosen in an agnostic way, leaving valuable information from nuclear physics and independent observations of neutron stars unused. In this work, we propose to encode information on neutron star physics into data-driven prior distributions constructed with normalizing flows, referred to as neural priors. These priors take input from constraints on the nuclear equation of state and neutron star population models. Applied to GW170817, GW190425, and GW230529, we highlight two contributions of the framework. First, we demonstrate its ability to provide source classification and to enable model selection of equation of state constraints for loud signals such as GW170817, directly from the gravitational-wave data. Second, we obtain narrower constraints on the source properties through these informed priors. As a result, the neural priors consistently recover higher luminosity distances compared to agnostic priors. Our method paves the way for classifying future ambiguous low-mass mergers observed through gravitational waves and for continuously incorporating advances in our understanding of neutron star properties into gravitational-wave data analysis.

Figures (7)

• Figure 1: Schematic overview of the neural priors. For both BNS and NSBH binaries (left column), a training dataset of masses and tidal deformabilities is generated, where the former are obtained by assuming a specific NS population (middle panel) and the latter from a set of EOS constraints (right panel). A normalizing flow (NF) is trained on these samples to render their distribution tractable. The learned densities are used as priors, which we refer to as neural priors. Further details are given in Sec. \ref{['sec: methods: neural priors construction details']}.
• Figure 2: Neural priors on source-frame masses and tidal deformabilities considered in this work. The top row (bottom row) shows the priors for BNS (NSBH) systems, while the columns show different population models and the colors denote the EOS constraints. The contour lines show the $68\%$ and $95\%$ probability areas of the distributions.
• Figure 3: Histograms of the log-likelihood values of the posterior samples obtained with the uninformed prior (gray) and the neural priors (colors). For each GW event, we only show the results for the preferred NS population model.
• Figure 4: Inference results of some selected parameters for GW170817 using the BNS hypothesis with the Gaussian population model and varying EOS constraints (colors) compared to the inference using uninformed priors (in gray). The star indicates the reference model with the highest evidence across the different neural priors. The dark (light) contours denote the $68\%$ ($95\%$) credible area.
• Figure 5: Inference results of some selected parameters for GW190425 using the BNS hypothesis with the uniform population model and varying EOS constraints (colors) compared to the inference using uninformed priors (in gray). The star indicates the reference model with the highest evidence across the different neural priors. The dark (light) contours denote the $68\%$ ($95\%$) credible area.