Coalescing Compact Binary Parameter Estimation with Gravitational Waves in the Presence of non-Gaussian Transient Noise

This study quantifies how non-Gaussian transient noise glitches in LIGO detectors bias the parameter estimation of compact binary coalescence signals, revealing significant distortions in mass, spin, and sky position—particularly when glitches occur shortly before the merger—and establishing time-separation thresholds to identify when glitch subtraction is necessary for unbiased results.

The Gist

The Cosmic "Pop" in the Data: How Glitches Trick Our Ear to the Universe

Imagine you are trying to listen to a very faint, beautiful symphony played by two black holes colliding billions of light-years away. This is what gravitational wave detectors like LIGO do. They are essentially giant, ultra-sensitive ears trying to hear the "chirp" of these cosmic collisions.

But here's the problem: The universe is noisy. Not just the hum of the cosmos, but sudden, loud, weird noises on Earth. A truck driving over a pothole, a lightning storm, or even a tiny vibration in the detector's mirror can create a sudden "pop" or "crack" in the data. Scientists call these glitches.

This paper is about a critical question: If a cosmic "chirp" and a terrestrial "pop" happen at the same time, does our computer get confused? And if so, how badly does it mess up our understanding of the black holes?

Here is the breakdown of their findings using simple analogies.

---

1. The Setup: A Stolen Spotlight

The researchers took real data from LIGO (specifically from the Livingston, Louisiana detector, which is the "sensitive ear" of the trio). They picked three types of annoying glitches that happen often:

• The "Blip": A tiny, sharp pop that lasts less than a second.
• The "Thunder": A rumble caused by actual thunderstorms shaking the detector, lasting a few seconds.
• The "Fast-Scattering": A rapid series of pops caused by light bouncing off dust, lasting minutes.

They then took computer simulations of real black hole collisions (based on famous events like GW150914 and GW190521) and "injected" them into the data right on top of these glitches. It's like playing a recording of a violin concerto while someone is banging on a drum next to the microphone.

2. The Result: The Computer Gets Drunk on Noise

When the computer tried to figure out the properties of the black holes (how heavy they are, how fast they were spinning, and where they were in the sky), the glitches made it hallucinate.

• The Mass Mix-up: The computer thought the black holes were much heavier or lighter than they actually were. In some cases, it guessed the mass was off by 100 times the mass of our Sun. That's like trying to weigh a feather while standing on a scale that someone is kicking.
• The Spin Spin-out: The black holes in the simulation weren't spinning at all. But because of the glitch, the computer insisted they were spinning at maximum speed. It's like seeing a still photo of a person and the computer insisting they are doing a breakdance because of a blur in the background.
• The Sky Scramble: This is the most dangerous part. If you want to point a telescope at the collision to see what's left behind, you need to know where it is. The glitches made the computer point the telescope in the completely wrong direction. In many cases, the computer was 99% sure the black holes were in a part of the sky where they actually weren't.

3. The "Safe Zone" Rule

The researchers wanted to know: How close does the glitch have to be to the signal to cause this mess?

They found a "Safe Zone."

• The Time Prior: Imagine the black hole collision is a song. The computer listens to a tiny window of time before the song starts to get ready.
• The Finding: If the glitch happens inside that tiny window before the song starts, the computer gets completely confused. It thinks the glitch is part of the song.
• The Surprise: Even if the glitch happens before the song starts, if it's close enough to be in that "listening window," the computer still gets it wrong. It's like if someone sneezes right before you start speaking; the computer thinks the sneeze was the first word of your sentence.

The Verdict: If a glitch happens within the "time prior" of the signal, the data is usually corrupted. You can't just ignore it; you have to fix the data first.

4. The "Glitch-Tracker" Phenomenon

The most fascinating discovery was a behavior they called "Glitch-Tracking."

Sometimes, the computer is so confused that it completely ignores the black hole signal. Instead, it tries to explain the glitch as if it were the black hole.

• The Analogy: Imagine you are trying to identify a specific bird by its song. But a car backfires nearby. Instead of saying, "That's a car," the computer says, "That bird must be a car!" and starts calculating the "mass" and "spin" of the car as if it were a bird.
• In their study, for certain types of glitches (Blips) and certain heavy black holes, the computer literally tracked the glitch's timing, thinking the glitch was the event.

5. Why This Matters

This paper is a warning label for the future of astronomy.

• The Good News: We now know exactly when glitches cause problems. If a glitch is far away in time, we are safe.
• The Bad News: As detectors get more sensitive (hearing fainter signals), we will hear more glitches. The chance of a glitch overlapping with a real signal is going up.
• The Solution: We can't just trust the computer's first guess anymore. If a signal overlaps with a glitch, we have to use special, expensive math to "subtract" the glitch before we can trust the results. Otherwise, we might publish papers about black holes that don't exist or have the wrong properties.

The Bottom Line

Listening to the universe is like trying to hear a whisper in a hurricane. This paper tells us that if a gust of wind (a glitch) hits at the exact moment the whisper starts, we might think the wind was the whisper. We now have a map of exactly how close that wind needs to be to trick us, so we can stop making mistakes and start finding the real secrets of the cosmos.

Technical Summary

1. Problem Statement

Gravitational-wave (GW) detectors (LIGO, Virgo, KAGRA) frequently encounter non-Gaussian transient noise artifacts known as glitches. These glitches can overlap in time with genuine GW signals from compact binary coalescences (CBCs), such as binary black hole (BBH) mergers.

• The Core Issue: When a GW signal coincides with a glitch, standard parameter estimation (PE) algorithms—which assume Gaussian background noise—can produce biased posterior distributions. This leads to incorrect inferences about the source's mass, spin, sky location, and distance.
• Context: During the fourth observing run (O4), 29% of GW candidates had overlapping or nearby glitches. As detector sensitivity improves, the likelihood of signal-glitch overlap increases, posing a risk to the accuracy of astrophysical conclusions and follow-up observations (e.g., electromagnetic telescopes).
• Gap: While glitch subtraction methods exist (e.g., BayesWave), they are computationally expensive and do not perfectly remove biases. There is a need to quantify exactly when and how glitches bias parameters to determine if subtraction is necessary and to define "safe" time separations between signals and glitches.

2. Methodology

The study employs a "worst-case scenario" approach using real detector data and simulated signals to quantify biases across ten observable parameters.

• Data Source: Strain data from the third observing run (O3), specifically from the LIGO Livingston Observatory (LLO), which was the most sensitive detector during that period.
• Glitch Selection: Three specific glitch types were selected based on their high occurrence rates and significant time-frequency overlap with CBC signals:
  1. Blip: Sub-second duration, wide frequency bandwidth.
  2. Thunder: Several seconds duration, low frequency (<200 Hz), caused by acoustic coupling from thunderstorms.
  3. Fast-scattering: Sub-second transients occurring in rapid succession over minutes, correlated with ground motion.
• Signal Injection:
  • Three simulated BBH signals were injected, modeled after real detections: GW190521 (high mass, short duration), GW150914 (standard mass), and GW231123 (extremely high mass).
  • Signals were injected into the LLO data at various time offsets relative to the glitch start time ("glitch-front").
  • Control Group: Injections were also performed into "clean" (glitch-free) Gaussian data segments preceding the glitches to establish baseline posterior distributions.
• Parameter Estimation: Bayesian inference was performed using the Bilby library with the IMRPhenomXPHM waveform model. The analysis assumed only a CBC signal and Gaussian noise (standard PE), intentionally ignoring the glitch to simulate real-world conditions where glitches might go unmitigated.
• Statistical Metric: A cost function was used to quantify the deviation of the inferred posterior median ($\bar{x}$) from the true injected value ($\mu$), normalized by the posterior standard deviation ($\sigma$):
  $$c = \frac{|\bar{x} - \mu|}{\sigma}$$
  Biases were deemed "significant" if the cost function or standard deviation exceeded thresholds derived from the control group (specifically, the global extrema of the inner 75th percentile of control data).
• Glitch-Front Definition: The study defined a reproducible "glitch-front" (start time) for each glitch type using time-frequency analysis (Omegascan) to measure the earliest point where the glitch energy significantly impacted the data.

3. Key Contributions

• Quantification of Bias: The study provides the first comprehensive quantification of parameter biases for short-duration CBC signals interacting with three distinct, high-rate glitch types across a wide range of masses (up to ~249 $M_\odot$).
• Definition of "Safe" Time Separation: It establishes the temporal window within which a glitch must occur to cause significant bias, offering guidelines for when glitch subtraction is mandatory.
• Identification of "Glitch-Tracking": The authors discovered a specific failure mode where the PE model fits the glitch itself rather than the GW signal, causing the inferred geocenter time to "track" the glitch's position in time.
• Validation of Time-Prior Hypothesis: The study confirms that biases are most severe and frequent when glitch energy falls within the time prior of the CBC signal (the window used by the PE algorithm to define the signal's start).

4. Key Results

A. Parameter Bias Magnitude

• Mass and Spin: All mass parameters ($m_1, m_2, \mathcal{M}, q$) and spin parameters ($\chi_{eff}, \chi_p, a_1, a_2$) showed statistically significant biases for all glitch types.
  • Deviations in chirp mass reached up to 100 $M_\odot$.
  • Spin parameters, which were injected as zero, were frequently biased toward maximal spin values (near 1.0), indicating the model interprets glitch energy as high-spin precession.
• Luminosity Distance ($d_L$): Significant biases were observed, with some cases deviating by over 5,000 Mpc. Biases were often negative (underestimating distance), likely because the high-SNR glitch mimics a closer, louder source.
• Sky Position: Glitch-affected injections were four times more likely to have sky position posteriors with negligible support (<20%) for the true sky location compared to control injections. This poses a severe risk for electromagnetic follow-up.

B. Temporal Dependence and "Safe" Separation

• The Time Prior is Critical: The study found that 98.1% of significant biases occurred when glitch energy was present within the time prior of the CBC signal (the 0.1s window before the merger time used in the analysis).
• Severity: Biases occurring within the time prior were, on average, 20.5 times larger (in cost function value) than those occurring outside the time prior.
• Glitch Types:
  • Blip and Thunder: Significant biases occurred even before the glitch-front (negative time offsets), suggesting the glitch's influence extends into the signal's time prior.
  • Fast-scattering: Biases were predominantly observed after the glitch-front, as the early transient arches did not significantly overlap with the signal's time prior.

C. Glitch-Tracking Behavior

• For Blip glitches overlapping with GW150914-like and GW190521-like signals, the PE model exhibited "glitch-tracking."
• Mechanism: When the glitch entered the signal's time prior, the inferred geocenter time posterior distribution shifted to align with the glitch's time, effectively treating the glitch as the signal.
• Robustness: This behavior was less pronounced for the GW231123-like signal, likely due to its lower frequency overlap with the blip glitches.

5. Significance and Implications

• Operational Guidelines: The study provides concrete criteria for data analysts: if a GW candidate overlaps with a known glitch type (Blip, Thunder, Fast-scattering) within the signal's time prior, the inferred mass, spin, and sky location are unreliable without glitch subtraction.
• Follow-up Astronomy: The severe bias in sky localization highlights the risk of misdirecting telescopes. Low-latency localization methods (like PyCBC Live) may be more robust than full Bayesian PE, but full PE results for glitch-contaminated events must be treated with extreme caution.
• Future Research: The findings suggest that simply increasing detector sensitivity without addressing glitch mitigation will lead to a higher rate of biased astrophysical conclusions. The paper advocates for the development of faster, more accurate glitch-subtraction pipelines or joint inference methods (modeling signal and glitch simultaneously) to handle the increasing rate of signal-glitch overlaps in future observing runs (O4 and beyond).

In conclusion, this paper demonstrates that short-duration CBC signals are extremely susceptible to non-Gaussian noise. The presence of glitch energy within the analysis time prior is the primary driver of catastrophic parameter estimation errors, necessitating robust mitigation strategies to ensure the scientific integrity of future GW detections.