Costs of Bayesian Parameter Estimation in Third-Generation Gravitational Wave Detectors: an Assessment of Current Acceleration Methods

TL;DR

The paper addresses the looming computational bottleneck of Bayesian parameter estimation for compact binary mergers in the new generation of gravitational wave detectors. It systematically compares standard stochastic-sampling PE with acceleration methods Relative Binning, Multibanding, and Reduced Order Quadrature using hundreds of trials and the ET MDC-1 catalog to model how cost scales with SNR and duration. The main finding is that standard PE would require billions of CPU hours for a month of 3G data, while acceleration can reduce this to millions, with ROQ delivering the fastest speed at the cost of upfront training and potential high-SNR accuracy trade-offs. The work highlights the need for efficient, scalable inference strategies and suggests that machine learning and other innovations will be essential to enable catalog-scale analyses in the 3G era while stressing the importance of validating accuracy in high-SNR regimes.

Abstract

Bayesian inference with stochastic sampling has been widely used to obtain the properties of gravitational wave (GW) sources. Although computationally intensive, its cost remains manageable for current second-generation GW detectors because of the relatively low event rate and signal-to-noise ratio (SNR). The third-generation (3G) GW detectors are expected to detect hundreds of thousands of compact binary coalescence (CBC) events every year with substantially higher SNR and longer signal duration, presenting significant computational challenges. In this study, we systematically evaluate the computational costs of CBC source parameter estimation (PE) in the 3G era by modeling the PE time cost as a function of SNR and signal duration. We examine the standard PE method alongside acceleration methods including relative binning, multibanding, and reduced order quadrature. We predict that PE for a one-month-observation catalog with 3G detectors could require at least billions of CPU core hours with the standard PE method, whereas acceleration techniques can reduce this demand to less than millions of core hours, which is as high as the cost of analyzing GW events in the past 10 years. These findings highlight the necessity for more efficient PE methods to enable cost-effective and environmentally sustainable data analysis for 3G detectors. In addition, we assess the accuracy of accelerated PE methods, emphasizing the need for careful treatment in high-SNR scenarios.

Figures (4)

• Figure 1: The CPU core time of stochastic sampling in our experiments, along with the bilinear fitting. The error bars represent the 16% to 84% percentiles of the sampling time across 20 events for each configuration, with the mean indicated in the center. The dashed lines correspond to the bilinear fits of the sampling time cost.
• Figure 2: The Jensen-Shannon Divergence (JSD) between posteriors obtained using acceleration methods and the standard method. Each subfigure corresponds to a different source type, and each patch represents a different parameter. The color of each grid represent the value of JSD in bits. A JSD of 0 indicates that the two posterior distributions are identical, while a JSD of 1 indicates they are completely different.
• Figure 3: Top: Cumulative distributions of network SNR of different sources (in different columns) and different detector networks (in different linestyles). The black dashed lines mark the SNR=8 threshold. Bottom: Cumulative distributions of signal durations, measured from 5 Hz, for different source types.
• Figure 4: Top: Cumulative distributions of estimated PE cost for detected events (SNR$>$8, in one-month observation) in CPU hours with different detector networks. The solid line represents ET-only network, dashed line represents the ET-CE network, and solid-dashed line represents the ET-2CE network. Different source types are in separate columns, with different colors representing various PE methods. Bottom: Total CPU hours required to perform Bayesian PE for the one-month observation with different detector networks. The numbers indicate the corresponding CPU core hours for each bar. A dot and a cross above each bar represent the cost for $f_\mathrm{low}=$3Hz and 2Hz, respectively.