Accelerated Sequential Posterior Inference via Reuse for Gravitational-Wave Analyses

TL;DR

The paper tackles the computational bottleneck of reanalyzing gravitational-wave events under different models by reusing existing results with ASPIRE, which couples a normalizing-flow density estimator with a generalized sequential Monte Carlo that evolves from the flow to the true posterior under a new model. It preserves unbiasedness and reproduces full Bayesian results when changing waveform models or adding physics such as spin precession and orbital eccentricity, with $4$–$10\times$ reductions in likelihood evaluations per sample. The authors validate the approach with simulated injections and real data (e.g., GW150914), including P–P tests showing unbiased posteriors and robust evidence estimates. The framework enables scalable waveform-systematics studies and catalog updates, and is released as open-source software to support rapid, unbiased reanalyses across gravitational-wave astronomy and beyond.

Abstract

We introduce Accelerated Sequential Posterior Inference via Reuse (ASPIRE), a broadly applicable framework that transforms existing posterior samples and Bayesian evidence estimates into unbiased results under alternative models without rerunning the original analysis. ASPIRE combines normalizing flows with a generalized Sequential Monte Carlo scheme, enabling efficient updates of existing results and reducing the computational cost of reanalyses by 4-10 times. This addresses a growing problem in gravitational-wave astronomy, where events must be repeatedly reanalyzed under different models or physical hypotheses. We show that ASPIRE reproduces full Bayesian results when switching waveform models or adding physical effects such as spin precession and orbital eccentricity. With this statistical robustness, ASPIRE turns repeated reanalyses into fast, reliable updatespaving the way for systematic studies of waveform systematics, scalable reanalyses across large event catalogs, and broadly applicable Bayesian reanalysis across other scientific domains.

Figures (5)

• Figure 1: Comparison of traditional Bayesian inference (top), which must perform full inference for a new model $M_2$ despite existing posterior samples with $M_1$, with (bottom), which reuses the existing posterior samples under model $M_1$ via flow approximation and an update to obtain unbiased results under a new model $M_2$.
• Figure 2: Posterior samples obtained from analyzing two simulated signals with different waveform models. Top: GW150914-like system with comparable masses and low spins. Bottom: highly spinning asymmetric system with $m_1/m_2=4$. Results are shown for IMRPhenomXPHM using standard sampling with dynesty (black dotted line), IMRPhenomXO4a using , initialized from the IMRPhenomXPHM samples (orange dashed line) and IMRPhenomXO4a using standard sampling with dynesty (blue solid line). Parameters shown are chirp mass, mass ratio, effective aligned spin $\chi_\mathrm{eff}$ and effective precessing spin $\chi_p$. The in millinats (mnats) between the two IMRPhenomXO4a posteriors is quoted above each subplot and show that the results are consistent.
• Figure 3: Posterior samples obtained when adding new physics to an initial analysis. Top: spin-precessing analysis of a GW150914 injection, results are shown for the Gray: IMRPhenomD (aligned spins) baseline with dynesty (black dotted line), IMRPhenomPv2 (precessing spins) using initialized from the aligned-spin samples (orange dashed-line) and an IMRPhenomPv2 using standard sampling with dynesty (blue solid line). Bottom: eccentric analysis of an aligned-spin GW150914 injection, results are shown for the TaylorF2 baseline with dynesty (black dotted line), TaylorF2Ecc using initialized from the TaylorF2 result (orange dashed line) and TaylorF2Ecc using standard sampling with dynesty (blue solid line). Parameters shown are chirp mass, mass ratio, effective aligned spin $\chi_\mathrm{eff}$, and effective precessing spin $\chi_p$ (top row) or eccentricity at 20 Hz $e_{20\textrm{Hz}}$ (bottom row). The in millinats (mnats) between the posteriors with same waveform is quoted above each subplot and demonstrate consistency.
• Figure 4: Posterior distributions for GW150914 with chirp mass and mass ratio and aligned spin $\chi_\mathrm{eff}$ and effective precessing spin $\chi_\mathrm{p}$. Results are shown for an initial analysis with IMRPhenomXPHM and dynesty (black dotted line), an analysis with IMRPhenomXO4a and , updated from the initial analysis (orange dashed line) and a baseline analysis IMRPhenomXO4a and dynesty (blue solid line). The in millinats (mnats) between the posteriors with same waveform is quoted above each subplot and confirm that the results are consistent.
• Figure 5: Probability–probability (P–P) plot for 100 simulated binary black hole signals analyzed with . Signals were generated with IMRPhenomXPHM in a three-detector network and analyzed using bilby with dynesty. Initial samples were drawn from the prior to initialize . Shaded regions indicate 1-, 2-, and 3-$\sigma$ confidence intervals. The combined $p$-value of 0.574 demonstrates that the posteriors are statistically unbiased.