Robust parameter estimation within minutes on gravitational wave signals from binary neutron star inspirals

TL;DR

This work addresses the computational bottleneck of extracting nuclear matter information from binary neutron star gravitational waves by adapting the GPU-accelerated Jim parameter estimation pipeline to include tidal effects. By combining relative binning, gradient-based MCMC, and normalizing flows within flowMC, the authors demonstrate rapid Bayesian parameter estimation for GW170817 and GW190425 across two waveform families without pretraining. Validation against LALSuite and Bilby shows consistent posteriors with favorable effective sample sizes, and real-event runtimes are reduced to tens of minutes on a single GPU, with a smaller environmental footprint than CPU-based approaches. The approach promises rapid, scalable EOS inference and efficient multi-messenger analyses for current and next-generation gravitational wave detectors.

Abstract

The gravitational waves emitted by binary neutron star inspirals contain information on nuclear matter above saturation density. However, extracting this information and conducting parameter estimation remains a computationally challenging and expensive task. Wong et al. introduced Jim arXiv:2302.05333, a parameter estimation pipeline that combines relative binning and jax features such as hardware acceleration and automatic differentiation into a normalizing flow-enhanced sampler for gravitational waves from binary black hole (BBH) mergers. In this work, we extend the Jim framework to analyze gravitational wave signals from binary neutron stars (BNS) mergers with tidal effects included. We demonstrate that Jim can be used for full Bayesian parameter estimation of gravitational waves from BNS mergers within a few tens of minutes, which includes the training of the normalizing flow and computing the reference parameters for relative binning. For instance, Jim can analyze GW170817 in 26 minutes (33 minutes) of total wall time using the TaylorF2 (IMRPhenomD_NRTidalv2) waveform, and GW190425 in around 21 minutes for both waveforms. We highlight the importance of such an efficient parameter estimation pipeline for several science cases as well as its ecologically friendly implementation of gravitational wave parameter estimation.

Figures (7)

• Figure 1: Schematic overview of the training loop of the flowMC sampler. Each loop starts with running the local sampler. For the local sampling, we use the MALA algorithm, which exploits the gradient of the target distribution (green and gray lines) to evolve the Markov chains (red). With the samples obtained from the local sampler, a normalizing flow (NF) is trained to approximate the distribution of the Markov chains. During the global sampling phase, we use the density learned by the NF (purple) as a proposal. We accept or reject the proposed samples (red) with a Metropolis-Hastings step, which relies on both the proposal density as well as the target density. This local-global procedure is repeated until the NF has converged. Afterward, we perform a fixed number of production loops, where the weights of the NF are frozen and the local and global sampler output the final production samples.
• Figure 2: Distribution of mismatch values between ripple and LALsuite implementations of the IMRPhenomD (IMRD), IMRPhenomD_NRTidalv2 (IMRD_NRTv2) and TaylorF2 (TF2) waveforms. For both newly added waveform models, the probability mass peaks at a mismatch below $10^{-8}$, indicating consistency with the LALsuite implementation.
• Figure 3: p-p plot for the injections done with TaylorF2 (left) and IMRPhenomD_NRTidalv2 (right), each created from 100 injections. For both of the waveform models, the p-p plots agree well with the diagonal, demonstrating the robustness of the Jim pipeline on analyzing BNS signals. The shaded regions indicate the $1\sigma$, $2\sigma$, and $3\sigma$ credible levels due to finite sample size.
• Figure 4: Comparison of the posterior distributions of GW170817, obtained with the TaylorF2 waveform, using Jim and pBilby.
• Figure 5: Comparison of the posterior distributions of GW170817, obtained with the IMRPhenomD_NRTidalv2 waveform, using Jim and pBilby.