Assessing the robustness of amortized simulation-based inference to transient noise in gravitational-wave ringdowns

Abstract

Gravitational waves (GW) emitted by binary systems allow us to perform precision tests of general relativity in the strong field regime. Ringdown signals allow for probing black hole mass and spin with high precision in GW astronomy. With improvements in current and next-generation GW detectors, developing likelihood-free parameter inference methods is crucial. This is especially important when facing challenges such as non-standard noise, partial data, or incomplete signal models that prevent the use of analytical likelihood functions. In this work, we propose an amortized simulation-based inference strategy to estimate ringdown parameters directly. Specifically, our method is based on amortized neural posterior estimation, which trains a neural density estimator of the posterior for all data segments within the prior range. The results show that our trained amortized network achieves statistically consistent parameter estimates with valid confidence coverage compared to established Markov-chain methods, while offering inference speeds that are orders of magnitude faster. Furthermore, we evaluate the robustness of the method against transient noise contamination. Our analysis reveals that the timing of glitch injection has a decisive impact on estimation bias, particularly during the tail of a signal with sparse information. Glitch strength is positively correlated with estimation error, but has limited effect at low signal-to-noise ratios. Mass and spin parameters are most sensitive to noise. This study not only provides an efficient and accurate inference framework for ringdown analysis but also lays a foundation for developing robust data-processing pipelines for future GW astronomy in realistic noise environments.

Figures (6)

• Figure 1: Comparison of parameter estimates and confidence intervals derived from MCMC and amortized NPE. The figure shows the mean values for parameters $\mathcal{M}_{f}$, $\chi_f$, $\mathcal{A}_{220}$, and $\phi_{220}$ under the two methods, alongside their $68.3\%$, $95.5\%$, and $99.7\%$ credible intervals from the full posterior distribution. The black line marks the values of the injection parameters.
• Figure 2: The figure illustrates the relationship between credible level and expected coverage for models trained under different training epochs. The blue curve represents the 100 epochs training model while the yellow curve corresponds to the 200 epochs model, trained for a higher number of epochs.
• Figure 3: The JSD values between posterior distributions obtained from data contaminated with glitches and those from clean data at the time point of noise insertion. The data points in the graph are the mean JSD calculated from five independent runs. Error bars represent the standard deviation (SD) of these five runs. The horizontal dashed line indicates a threshold of JSD = 0.5, above which the posterior deviation is considered significant.
• Figure 4: The JSD value of marginal distribution different time point between posterior distributions obtained from data contaminated with glitches and those from clean data at the time point of noise insertion. The same JSD threshold as in Fig. \ref{['Time_ALL']} is shown for reference.
• Figure 5: The JSD value between posterior distributions obtained from data contaminated with glitches and those from clean data for different glitches SNR. The data points in the graph are the mean JSD calculated from five independent runs. Error bars represent the SD of these five runs. The same JSD threshold as in Fig. \ref{['Time_ALL']} is shown for reference.