Joint inference for gravitational-wave signal and noise glitch: Method and application

Technical Summary: Joint Inference for Gravitational-Wave Signal and Noise Glitch

Problem Statement
Gravitational-wave (GW) observatories are frequently contaminated by non-Gaussian, non-stationary noise transients known as "glitches." These artifacts, occurring approximately once per minute, degrade the accuracy of astrophysical inference by confounding search pipelines and biasing parameter estimation when they overlap with true GW events. While existing mitigation strategies include data excision (removing affected segments) or glitch subtraction (modeling and removing the glitch prior to inference), these methods have limitations. Excision sacrifices astrophysical information, and subtraction methods can leave residual glitch power that still biases results. Furthermore, the interpretation of specific events, such as GW191109 and GW200129, has been shown to depend heavily on the chosen glitch mitigation strategy and waveform approximant, sometimes leading to conflicting conclusions regarding black hole spin properties.

Methodology
The authors present bilby_glitch, a modular analysis pipeline designed for the simultaneous Bayesian inference of GW signals and noise glitches. Built upon the widely used bilby framework, the pipeline modifies the standard Whittle likelihood to treat the data $d$ as a linear sum of stationary Gaussian noise $n$, a glitch model $g(\gamma)$, and a signal waveform $h(\theta)$:
$$d = n + g(\gamma) + h(\theta)$$
To address the computational complexity of jointly sampling high-dimensional signal and glitch parameters, the pipeline analytically marginalizes over the luminosity distance.

The framework supports transdimensional inference, where the number of glitch components $N$ is treated as a discrete parameter sampled directly from the data using the tBilby package (which employs ghost parameters). The paper implements three distinct glitch models to accommodate different morphologies:

1. Slow Scattering Model: A physically motivated model for scattered light glitches, characterized by sinusoidal motion in the microseism band generating multiple scattering arches.
2. Sine-Gaussian Wavelets: A flexible superposition of sine-Gaussian wavelets (similar to bayeswave) capable of fitting a broad range of unmodeled glitch morphologies.
3. Chirplets: A generalization of wavelets including a chirp rate parameter, allowing frequency evolution over time. While less effective for general glitch subtraction due to increased dimensionality, they are noted for their ability to mimic GW signals, serving as a rigorous test case for bias mitigation.

Key Results
The pipeline was validated and applied through the following analyses:

• Simulation Validation:

  • Slow Scattering: Using 145 simulated binary black hole (BBH) signals with overlapping slow scattering glitches, the pipeline successfully recovered signal parameters within expected credible intervals. The joint inference increased computational runtime by approximately 50% compared to signal-only analysis.
  • Chirplet Bias Mitigation: In a test case where a chirplet glitch was engineered to bias astrophysical inference, the joint inference framework successfully reconstructed both the signal and the glitch. The results demonstrated that modeling the glitch concurrently with the signal is necessary to produce unbiased estimates of binary parameters (e.g., chirp mass, mass ratio, and spin parameters), whereas neglecting the glitch led to significant biases.

• Application to Real Events:

  • GW191109: Re-analyzed using the transdimensional slow scattering model. The results were consistent with previous analyses (Udall et al.), recovering similar posteriors for intrinsic parameters. Notably, the transdimensional approach favored $N=2$ scattering arches rather than the $N=5$ assumed in prior work, though this did not significantly alter the astrophysical conclusions.
  • GW200129: Re-analyzed using joint inference with sine-Gaussian wavelets and two waveform approximants: NRSur7dq4 and IMRPhenomXPHM. The study found that the evidence for spin-precession is highly dependent on the interplay between the waveform model and the glitch model.
    • When using NRSur7dq4 combined with wavelet-based glitch modeling, the evidence for spin-precession was significantly weaker ($\chi_p \approx 0.52$), consistent with findings by Payne et al.
    • Conversely, using IMRPhenomXPHM with the same wavelet-based glitch modeling yielded strong evidence for spin-precession ($\chi_p \approx 0.67$).
    • This divergence highlights that waveform systematics play a critical role in the interpretation of this event, and that joint inference can alter the strength of evidence for specific physical phenomena depending on the waveform choice.

Significance and Claims
The authors position bilby_glitch as a significant step toward integrating robust glitch mitigation directly into standard GW parameter estimation workflows. By leveraging the existing bilby infrastructure, the tool allows users to perform joint inference without adopting a separate analysis framework. The paper claims that while joint inference is computationally expensive, it is essential for producing unbiased astrophysical inferences, particularly for events where glitches overlap with the signal in time and frequency.

The authors explicitly state that the paper serves as a "milestone" in facilitating this type of joint inference. They emphasize the modularity of the code, noting that while the current implementation includes three specific models, the framework is designed to allow researchers to easily incorporate additional glitch models (such as those by Bondarescu et al. or Malz and Veitch) in the future. The paper concludes modestly, acknowledging that while it improves robustness, the computational cost remains a challenge, and future work should focus on developing samplers better suited for high-dimensional, multi-modal, and transdimensional parameter spaces.