Inferring the spins of merging black holes in the presence of data-quality issues

This paper demonstrates that both low-SNR glitches and the residuals from glitch subtraction can significantly bias black hole spin measurements, necessitating joint inference of glitch and gravitational wave parameters to mitigate these uncertainties.

The Gist

The Big Picture: Listening to the Universe with a Broken Ear

Imagine the LIGO and Virgo detectors as incredibly sensitive ears trying to hear a whisper from a black hole collision happening billions of light-years away. These collisions create gravitational waves (ripples in space-time). By listening to these ripples, scientists can figure out how fast the black holes were spinning before they crashed.

However, the Earth is a noisy place. Just like trying to hear a whisper in a crowded room, the detectors get interrupted by "glitches"—sudden bursts of noise caused by things like a truck driving by, a distant earthquake, or even a fly hitting the equipment.

This paper asks a scary question: What if a glitch happens at the exact same time as a black hole collision, but it's so quiet that we don't even notice it? Does that tiny, unnoticed noise mess up our calculation of how fast the black holes were spinning?

The Core Problem: The "Ghost" Left Behind

The authors found two major problems with how we currently handle these glitches:

1. The "Eraser" Leaves a Smudge
When scientists spot a loud glitch, they try to "subtract" it from the data, like using an eraser to remove a pencil mark from a drawing.

• The Analogy: Imagine you are trying to copy a beautiful painting, but someone scribbled a messy line over it. You try to erase the scribble. Even if you do a perfect job, the paper is still slightly rough where you erased. That roughness is the "residual."
• The Finding: The paper proves that no matter how loud or quiet the original glitch was, the act of subtracting it always leaves behind a tiny bit of "roughness" (residual noise). This leftover noise is strong enough to trick the computer into thinking the black holes were spinning differently than they actually were.

2. The "Invisible" Glitch
Usually, scientists only try to fix glitches that are loud enough to be flagged by their software (like a "Check Engine" light).

• The Analogy: Imagine you are trying to hear a singer, and there is a tiny, quiet hum in the background. Your "Check Engine" light doesn't go off because the hum is too quiet. So, you ignore it.
• The Finding: The authors showed that even these super quiet glitches (ones too faint to be detected) can still significantly distort the answer. It's like a tiny, invisible smudge on a lens that makes the whole picture look blurry. They found that these quiet glitches can make us think a black hole is spinning backward when it's actually spinning forward, or vice versa.

The "Phase" Dance

The paper also looked at when and how the glitch hits the signal.

• The Analogy: Think of the black hole signal as a song and the glitch as a drumbeat.
  • If the drumbeat hits exactly when the singer takes a breath, it might hide the breath.
  • If it hits when the singer hits a high note, it might make the note sound flat.
  • If it hits in between, it might not matter at all.
• The Finding: The damage depends entirely on the timing and the "rhythm" (phase) of the glitch relative to the black hole signal. A wide variety of different glitches, hitting at different times, can all cause the same kind of confusion.

The Solution: Don't Just Erase, Listen Together

The paper suggests that the old method (find glitch -> erase glitch -> analyze signal) is flawed because the "erasing" process itself creates errors.

• The New Idea: Instead of trying to erase the glitch first, scientists should try to listen to the glitch and the black hole signal at the same time.
• The Analogy: Instead of trying to clean the painting before copying it, you should look at the messy painting and the clean version simultaneously, asking the computer: "Is this line part of the art, or is it a scribble?" By modeling both at once, the computer can figure out the true spin of the black holes without being tricked by the leftover "smudges."

Why Should You Care?

Black hole spins tell us how these cosmic monsters are born and how they live. If our measurements are wrong because of a tiny, unnoticed glitch, our entire understanding of the universe's history could be slightly off.

The Takeaway:
The universe is noisy. We can't just ignore the quiet noises or think that "erasing" the loud ones fixes everything. To get the truth, we need to be smarter about how we separate the signal from the noise, acknowledging that even the quietest static can change the story we tell about the cosmos.

Technical Summary

1. Problem Statement

Gravitational wave (GW) observations of binary black hole (BBH) mergers provide critical insights into astrophysical formation channels through the measurement of component spins (specifically the effective aligned spin $\chi_{\text{eff}}$ and effective precessing spin $\chi_p$). However, ground-based detectors (LIGO, Virgo, KAGRA) are susceptible to non-Gaussian, transient noise artifacts known as "glitches."

Current inference pipelines typically assume data consists of a GW signal plus stationary Gaussian noise. When glitches occur, standard mitigation involves inferring the glitch's morphology and subtracting it from the data. The paper addresses two critical gaps in this process:

1. Sub-threshold Glitches: Glitches with low Signal-to-Noise Ratios (SNR $\rho_g \lesssim 5$) often go undetected by automated flagging tools (e.g., Omicron) but may still bias parameter inference.
2. Residual Bias: Even when glitches are detected and subtracted, statistical uncertainties in the inference process prevent perfect subtraction, leaving behind "residual power" that can systematically bias the inferred spin parameters.

2. Methodology

The authors conducted a series of controlled simulations using simulated data to isolate statistical uncertainties from systematic modeling errors.

• Simulations:

  • CBC Signals: Four distinct BBH configurations were simulated based on real events GW191109 and GW200129, ranging from high-mass/moderate-spin to moderate-mass/extreme-spin.
  • Glitch Models: Two models were used:
    1. Slow Scattering Model: A physically motivated model based on light scattering off a harmonic oscillator (3 arches).
    2. Sum-of-Wavelets: A flexible model using Morlet-Gabor wavelets (implemented via bayeswave and bilby).
  • Noise: Analyses were performed in both zero-noise (to isolate glitch effects) and Gaussian noise environments.

• Inference Pipeline:

  • Standard CBC inference was performed using the bilby pipeline with the Dynesty sampler and the NRSur7dq4 waveform approximant.
  • Mitigation Strategies Tested:
    1. No Mitigation: Analyzing data with the glitch present.
    2. Point-Estimate Subtraction: Inferring glitch parameters, generating a point estimate (Maximum Likelihood, Median, or Random Draw), subtracting it, and then inferring CBC parameters.
    3. Joint Inference: Simultaneously inferring CBC and glitch parameters (using bilby with fixed wavelets or bayeswave with variable dimensions).

• Metrics:

  • Residual SNR: The SNR of the glitch remaining after subtraction ($\rho_{\text{res}}$).
  • Bias Quantification: Measured via the credible level of the true parameter value in the posterior and the Jensen-Shannon (JS) divergence between the glitch-affected posterior and the reference (glitch-free) posterior.

3. Key Contributions and Results

A. Inevitable Residual Power (Section III)

The study confirms a theoretical expectation derived from the Fisher information matrix: glitch subtraction inevitably leaves residual power due to statistical uncertainty, regardless of the original glitch's SNR.

• Result: For a model with $N_p$ parameters, the expected residual SNR is $\rho_{\text{res}} \approx \sqrt{N_p}$.
• Finding: Even for high-SNR glitches, the residual SNR remains non-trivial (typically 3–7). This residual is an inescapable byproduct of the subtraction process.
• Point Estimates: The median of the posterior distribution for the glitch provides the most reliable point estimate for subtraction, minimizing residual SNR compared to Maximum Likelihood or random draws.

B. Bias from Sub-Threshold Glitches (Section IV)

The authors demonstrated that glitches with low SNR ($\rho_g \lesssim 5$), which are likely undetected in real data, can cause significant biases in spin inference.

• Result: Glitches with $\rho_g = 5$ (below the typical detection threshold) caused the true spin values to drop to low credibility (e.g., <20%) and produced JS divergences exceeding standard thresholds for disagreement.
• Implication: Spin measurements for massive BBHs (like GW191109 and GW200129) may be subtly biased by "silent" glitches that analysts are unaware of.

C. Bias from Residuals (Section V)

The study investigated whether the residuals left behind after standard subtraction (which have SNRs comparable to the low-SNR glitches studied above) also bias results.

• Result: Yes. Residuals from point-estimate subtraction can induce biases comparable to those of unsubtracted low-SNR glitches.
• Mechanism: The bias depends heavily on the specific realization of the glitch and the choice of point estimate. Random draws from the posterior often resulted in larger biases than the median.
• Solution: Joint inference (simultaneously modeling the glitch and the CBC) effectively mitigates these biases by marginalizing over the uncertainty of the glitch realization, returning results consistent with the true values.

D. Dependence on Time and Phase (Section VI)

The authors explored how the alignment of the glitch relative to the CBC signal affects bias.

• Result: Biases are not limited to glitches perfectly aligned with the merger. A wide range of time and phase overlaps can produce significant biases.
• Pattern: The magnitude and direction of the bias in $\chi_{\text{eff}}$ vary cyclically with the glitch phase. For $\chi_p$, biases are less cyclic but can still be severe.
• Insight: The bias arises because the glitch + signal combination mimics a different astrophysical configuration (specifically, a different spin configuration) rather than just adding noise.

4. Significance and Conclusions

• Reliability of Spin Measurements: The paper challenges the reliability of spin measurements for BBH events that have undergone standard glitch subtraction. Even "cleaned" data may contain residual biases that affect the interpretation of astrophysical populations.
• The "Silent" Glitch Danger: A major finding is that the most dangerous glitches may be those too quiet to be flagged by current data-quality tools. These can systematically skew the inferred spin distribution of the universe.
• Methodological Recommendation:
  • Joint Inference: The most robust solution for known glitches is joint CBC-glitch inference, which accounts for statistical uncertainties in the glitch morphology.
  • Glitch Marginalization: For low-SNR glitches, future work should focus on developing population priors for glitches to allow for marginalization over potential glitch realizations, rather than relying on detection and subtraction.
• Future Outlook: The authors emphasize that while they used "worst-case" glitch configurations, the lack of precise fine-tuning required to produce these biases suggests that such errors are a genuine risk in current GW catalogs.

In summary, this work provides a rigorous statistical framework showing that glitch mitigation via subtraction is insufficient to guarantee unbiased spin inference. It calls for a paradigm shift toward joint inference and better modeling of low-SNR noise populations to ensure the accuracy of gravitational wave astrophysics.