A Gaussian process framework for testing general relativity with gravitational waves

This paper introduces a Gaussian process framework with a time-localized kernel to test general relativity using gravitational wave data from binary black hole mergers, finding no evidence of deviations in GWTC-3 events and constraining fractional strain deviations to as low as 7%.

The Gist

The Big Picture: Listening for a "Ghost" in the Machine

Imagine you are a detective trying to solve a mystery. You have a very specific, perfect script for how a crime should happen (this is General Relativity, Einstein's theory of gravity). You also have a recording of the crime scene (this is the Gravitational Wave data from colliding black holes).

Usually, when you play the recording, it matches the script perfectly. But sometimes, there might be a tiny, unexpected sound—a creak, a whisper, or a glitch—that doesn't fit the script. That extra sound could be a clue that the script is wrong and that there is "new physics" happening.

The problem is, we don't know what that extra sound should sound like. It could be a whisper, a shout, a high-pitched squeal, or a low rumble. If you only listen for a specific type of noise, you might miss the real clue.

This paper introduces a new detective tool: A "Gaussian Process" framework. Instead of guessing what the weird noise sounds like, this tool acts like a highly flexible, shape-shifting net. It casts a wide net to catch any kind of unexpected sound, as long as it follows a few basic rules about how it behaves.

How the Tool Works: The "Smart Net"

The scientists built a mathematical "net" (called a Kernel) with three specific rules based on what they think a "new physics" signal would look like:

1. It happens at the crash: The weird noise is expected to happen right when the two black holes smash together (the merger), not long before or long after.
2. It has a rhythm: The noise likely oscillates (wiggles back and forth) at a specific speed, similar to the frequency of the crash itself.
3. It's a bit messy: It's not a perfect, clean sine wave; it has some randomness to it, like static on a radio.

By programming these rules into their computer model, they created a system that can say, "I see a pattern here that fits our idea of 'new physics,' even though we didn't know exactly what it would look like beforehand."

The Experiment: Testing the Net

The team tested their new net in three ways:

1. The "Fake Signal" Test: They took real, quiet noise from the LIGO detectors and secretly injected a fake "new physics" signal into it.

   • Result: The net caught it immediately. It correctly identified, "Hey, there's something here that doesn't fit the standard script!" and even reconstructed what the fake signal looked like.

2. The "Silence" Test: They looked at 174 segments of pure noise where no signal was injected.

   • Result: The net stayed quiet. It didn't scream "GHOST!" when there was nothing there. This proved the tool isn't just hallucinating signals out of random static.

3. The "Different Script" Test: They tried to catch a signal that was different from the rules their net was built on (a signal that changed its rhythm over time).

   • Result: Even though the signal was slightly different from their expectations, the net was flexible enough to still catch it and say, "Something is wrong here."

The Real Investigation: Checking 60 Black Hole Collisions

Finally, they applied their tool to 60 real gravitational wave events from the third catalog of black hole collisions (GWTC-3). They took the data, subtracted the perfect Einstein script, and looked at what was left over (the "residuals").

• The Verdict: They found no evidence of new physics.
• The Conclusion: For all 60 events, the leftover noise looked exactly like what you'd expect from random static or minor imperfections in the recording equipment. It matched Einstein's script perfectly.

How Precise Are They?

Even though they didn't find a ghost, they set a very strict limit on how "loud" a ghost could be hiding in the data.

They calculated that if there were a deviation from Einstein's theory, it would have to be incredibly small. Specifically, for one event (GW190701 203306), they can say with 90% confidence that any deviation is less than 7% of the total signal strength.

Think of it this way: If the gravitational wave signal was a giant ocean wave, they are saying, "If there is a tiny ripple caused by new physics, it is smaller than 7% of the height of that giant wave."

The Bottom Line

This paper doesn't discover new physics. Instead, it builds a better, more flexible "net" to catch it. They tested this net on simulated data and found it works great. When they used it on real data from 60 black hole collisions, the net came up empty.

The takeaway: Einstein's theory of gravity still holds up perfectly under the most extreme conditions we can observe. If new physics is hiding in the gravitational waves, it is hiding very well, and we need even more sensitive tools to find it.

Technical Summary

Technical Summary: A Gaussian Process Framework for Testing General Relativity with Gravitational Waves

Problem Statement
Gravitational-wave (GW) astronomy offers a unique opportunity to test General Relativity (GR) in the strong-field regime. However, in the absence of a compelling, specific alternative theory to GR, conducting analyses that allow for a wide range of potential deviations is challenging. Existing methods, such as parameterized post-Newtonian tests or residual strain tests using spline models and wavelets (e.g., BayesWave), often rely on specific assumptions about the form of the deviation or require flexible models that may not capture localized, coherent excess power effectively. The authors propose a method to search for deviations from GR in binary black hole merger signals with minimal assumptions regarding the specific nature of the deviation.

Methodology
The authors introduce a Gaussian process (GP) framework to model potential deviations from GR, denoted as $\delta s$, within the GW strain data $h$. The data is modeled as the sum of a GR-predicted signal ($s_{GR}$), instrumental noise ($n$), and the deviation signal ($\delta s$).

• Kernel Design: The core of the framework is a time-domain kernel $K(t_i, t_j)$ designed to encode prior beliefs about the nature of a deviation. The authors assume that if a deviation exists, it is likely:

  1. Localized in time near the merger.
  2. Characterized by a specific frequency $f_0$ (similar to the merger/ringdown frequency).
  3. Not necessarily symmetric in time.

  The kernel combines a Gaussian term (for temporal localization), a cosine term (for oscillation at frequency $f_0$), and a radial basis function (for stochasticity). The kernel is parameterized by a scale factor $k_0$ (amplitude), width $w$ (duration), characteristic frequency $f_0$, and coherence length $l$.

• Likelihood Construction: The framework operates in the frequency domain. The time-domain kernel is transformed into a frequency-domain kernel $K(f)$ using a discrete Fourier transform and projected onto the observing band (50–512 Hz). The total covariance matrix $C(\Lambda)$ is the sum of the noise covariance $N$ and the signal covariance $S(\Lambda)$, where $S$ is constructed from the kernel and detector antenna response functions. The likelihood is computed for the residual data $\delta h = h - s_{GR}$, where $s_{GR}$ is the best-fit GR waveform (IMRPhenomPv2).

• Inference: The authors use nested sampling (via dynesty and Bilby) to infer the posterior distributions of the GP hyper-parameters ($k_0, w, f_0, l$). They define a signal-to-noise Bayes factor ($B$) to compare the hypothesis of a deviation ($k_0 > 0$) against the noise-only hypothesis ($k_0 = 0$).

Key Contributions and Demonstrations

1. Simulated Signal Recovery: The framework was tested on simulated signals injected into O3 noise data. The GP successfully recovered the injected deviations, with posterior distributions for hyper-parameters consistent with true values. The Bayes factor strongly supported the signal hypothesis ($\ln B = 21.9$) for an injected signal with a network signal-to-noise ratio (SNR) of 9.4.
2. Robustness to Noise: The authors analyzed 174 segments of instrumental noise (pre-event) to establish a background distribution for the Bayes factor. They found that realistic noise segments generally yielded low Bayes factors, though they noted the presence of unflagged glitches could produce outliers.
3. Model Flexibility: The framework was tested against a GR-violating signal not drawn from the kernel (a modified post-Newtonian phase coefficient $\beta_2$). The GP successfully detected the deviation ($\ln B = 13.02$), demonstrating that the model can capture deviations even if they do not perfectly match the kernel's specific functional form, particularly for the post-merger portion of the signal.

Results from GWTC-3
The authors applied the framework to 60 binary black hole events from the Gravitational-Wave Transient Catalog 3 (GWTC-3) detected by both LIGO Hanford (H1) and Livingston (L1).

• No Evidence for Deviation: None of the 60 events showed evidence for a deviation from GR. The distribution of Bayes factors for the GW events was consistent with the background distribution derived from noise segments.
• Most Significant Event: The event with the highest Bayes factor was GW190916 200658 ($\ln B = 4.71$, $p=0.03$). The authors note this is statistically expected given the number of events tested and that the significance decreased slightly ($\ln B = 3.85$) when marginalizing over uncertainties in the GR waveform parameters.
• Upper Limits: In the absence of detections, the authors calculated upper limits on the fractional deviation in GW strain. For the event GW190701 203306, they constrained the maximum fractional deviation amplitude to as low as 7% (90% credibility) of the GR strain amplitude.

Significance and Claims
The paper claims to provide a flexible, minimal-assumption tool for detecting deviations from GR in GW residual data. The authors emphasize that their framework is a detection tool rather than a definitive proof of new physics. They explicitly state that if a deviation were detected, it could arise from either a true violation of GR or, more likely, misspecification in the analysis models (e.g., noise modeling or waveform approximants).

The authors conclude that while their method is computationally expensive (scaling as $O(n^3)$ due to covariance matrix inversion), it successfully constrains deviations in current data. They suggest that repeated, unexplained deviations from the best template banks would be required to seriously consider the alternative hypothesis of a GR violation. Future work is proposed to improve computational scaling and to potentially adapt the framework for modeling noise glitches.