⚛️ general relativity

Testing General Relativity with GWTC-4.0 through mixture models

This paper introduces a flexible mixture-model framework to test General Relativity using GWTC-4.0 data, revealing that while current observations remain consistent with GR, the significantly lower Bayes factors compared to standard analyses highlight the necessity of adopting more adaptable models to avoid biasing evidence against potential deviations.

Original authors: Koustav Chandra, Juan Calderón Bustillo

Published 2026-07-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Koustav Chandra, Juan Calderón Bustillo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery: Is the universe playing by the rules of Einstein's General Relativity (GR), or is there a secret rulebook being broken?

For the last decade, scientists have been listening to the "chirps" of black holes and neutron stars colliding. These collisions create gravitational waves, and scientists check if these waves match Einstein's predictions. So far, they mostly match. But to be sure, scientists want to combine the evidence from many different collisions to see if a pattern of "rule-breaking" emerges.

Here is the problem with how they've been doing it so far, and how this new paper offers a smarter way to look at the data.

The Old Way: The "All-or-Nothing" Team

Previously, when scientists combined data from 50 or 100 collisions, they used two main methods:

The "One Truth" Method: They assumed that if Einstein's rules were broken, every single collision broke them in the exact same way. It's like assuming that if one student in a school cheated on a test, every student in the school cheated on the exact same question.
The "Average Cheater" Method: They assumed that deviations from the rules were random but followed a neat, bell-curve pattern (like heights in a population).

The Flaw: These methods are very rigid. If even one event has a weird glitch or a tiny, unique quirk that looks like a rule-break, these rigid methods get confused. They might either say, "Aha! Einstein is wrong!" (because one weird event skewed the average) or "Einstein is definitely right!" (because the weird event got washed out by the assumption that everyone must be the same).

The New Way: The "Mixture Model" (The Smart Detective)

The authors of this paper, Koustav Chandra and Juan Calderón Bustillo, propose a new, more flexible approach called a Mixture Model.

Think of it like a fruit salad.

The Old Way assumed the salad was either 100% apples (Einstein is right) or 100% oranges (Einstein is wrong).
The New Way asks: "What if the salad is a mix?" Maybe 95% of the fruit is apples (Einstein is right), but 5% is oranges (Einstein is wrong), or maybe there are some weird, exotic fruits (like boson stars) mixed in.

In this new model, there is a variable called $\zeta$ (zeta).

If $\zeta = 1$ , the whole salad is apples (Everything follows Einstein).
If $\zeta = 0.5$ , half the salad is apples and half is something else.
The model doesn't force the "something else" to be the same for every event. It just asks, "Is there any part of this group that doesn't fit?"

What They Found

The team applied this new "fruit salad" logic to real data from the LIGO-Virgo-KAGRA collaboration (covering events from 2015 to 2023).

The Verdict: The data still looks like a salad made almost entirely of apples. Einstein's General Relativity holds up. There is no strong evidence that the universe is breaking the rules.
The Surprise: However, the "confidence" in this result is different.
- The old, rigid methods said, "We are 99.9% sure Einstein is right!" (Very high confidence).
- The new, flexible model said, "We are pretty sure Einstein is right, but let's not be too confident. The evidence is about 10 to 20 times stronger for Einstein than for breaking the rules."

Why the difference?
The old methods were "overconfident" because they forced the data into a rigid box. The new method admits, "Nature might be messy. Maybe some events have weird glitches or unique properties." By allowing for this messiness, the new model finds that the evidence for Einstein is still strong, but it's not as overwhelmingly "decisive" as the old methods claimed.

A Real-World Analogy: The Glitchy Microphone

Imagine you are listening to a choir.

Old Method: You assume that if one singer hits a wrong note, the whole choir is singing off-key. Or, if the microphone has a static crackle, you assume the entire choir is singing a different song.
New Method: You realize, "Wait, maybe 99 singers are perfect, and only one singer has a bad microphone, or maybe one singer is just having an off day."

The paper shows that when you account for the possibility of "bad microphones" (systematic errors) or "off-day singers" (unique deviations), you don't need to throw out the whole theory. You just need a more flexible way to count the votes.

The Bottom Line

The universe still seems to be playing by Einstein's rules. But this paper argues that scientists should stop using "rigid" math that assumes everyone is identical. Instead, they should use "flexible" math that allows for a mix of perfect matches and weird outliers.

This doesn't mean Einstein is wrong; it means the scientists are being more humble and realistic about how they count the evidence. It's a better way to listen to the universe without getting tricked by a single loud noise.

Technical Summary: Testing General Relativity with GWTC-4.0 through Mixture Models

Problem Statement
Gravitational-wave (GW) observations of compact binary mergers have enabled precision tests of General Relativity (GR) in the strong-field dynamical regime. Current methodologies for combining results from multiple events to search for deviations from GR rely on two primary frameworks: the "joint posterior" approach and the "hierarchical" approach.

The joint posterior approach assumes a common deviation parameter value across all events (effectively assuming $\sigma=0$ ).
The hierarchical approach treats deviation parameters as random variables drawn from a population distribution, typically modeled as a Gaussian with mean $\mu$ and covariance $\Sigma$ .

Both frameworks implicitly impose a restrictive "all-or-nothing" assumption: they assume that either all events satisfy GR (with deviations being zero or drawn from a specific population) or all events deviate from GR in a consistent manner. The authors argue that these assumptions are overly restrictive for theory-agnostic fractional deviation parameters, as beyond-GR (bGR) theories often predict deviations dependent on intrinsic source properties in non-linear ways. Furthermore, these frameworks cannot accommodate scenarios where a sub-population of events (e.g., exotic compact objects like boson stars) deviates from GR while the majority remains consistent with it. This rigidity risks biasing inferred evidence, potentially underestimating heterogeneous departures or overestimating evidence against GR due to unmodeled systematics.

Methodology
The authors introduce a flexible Bayesian framework based on mixture models to test GR across a catalogue of events. The core of this framework is a single parameter, $\zeta \in [0, 1]$ , representing the fraction of events consistent with GR.

Model Structure: The likelihood for a single event $k$ is modeled as a mixture:
$L(d_k | \zeta) = \zeta Z_{GR}(d_k) + (1-\zeta) Z_{bGR}(d_k)$
where $Z_{GR}$ and $Z_{bGR}$ are the Bayesian evidences for the GR and alternative models, respectively.
Hypothesis Testing: The framework tests the hypothesis that the entire catalogue is consistent with GR ( $\zeta = 1$ ) against a general model where $\zeta$ is a free parameter. The catalogue-level Bayes factor is computed as $B_{GR}^{bGR} = B_{\zeta=1}^{\zeta \neq 1}$ .
Agnosticism: Unlike the hierarchical approach, the default form of this mixture model is agnostic regarding the population distribution of deviations for the non-GR fraction. It does not assume a specific functional form (e.g., Gaussian) for the deviations of the $1-\zeta$ fraction, nor does it assume deviations are independent of source properties.
Inputs: The method requires only the set of Bayes factors against GR obtained from individual event analyses, making it applicable to existing public results from the LIGO-Virgo-KAGRA (LVK) collaboration.

Key Contributions and Validation

Framework Unification: The authors demonstrate that the existing joint and hierarchical frameworks used by LVK are restricted versions of this general mixture model.
Synthetic Validation: Using toy models and mock catalogues, the authors show that:
- The mixture model can detect GR violations even when individual events provide inconclusive evidence (accumulating small deviations across the population).
- The framework is robust against systematic errors. In scenarios where a few loud events exhibit apparent GR violations due to systematics, the mixture model naturally dilutes this false evidence as more consistent events are added, whereas standard methods might be misled by the outliers.
- When deviations are truly consistent across all events (satisfying the assumptions of current frameworks), the mixture model recovers the deviation effectively, though the hierarchical model may yield higher (potentially infinite) Bayes factors due to the tightening of the shared mean constraint.

Results: Application to GWTC-4.0
The authors applied this framework to publicly available results from the LVK collaboration covering O1–O4a observing runs, utilizing three specific test types:

Flexible Theory Independent (FTI): Probing deviations in the inspiral phase.
pSEOBNR: Testing consistency of post-inspiral signals with GR predictions.
KerrPostMerger (TEOBPM): Searching for deviations in the post-merger ringdown spectrum.

Findings:

Consistency with GR: For all three tests, the posterior distributions for $\zeta$ peaked near unity, indicating that the data are consistent with all events satisfying GR.
Bayes Factors: The catalogue-level Bayes factors favoring GR ( $B_{\zeta=1}^{\zeta \neq 1}$ $B_{ζ = 1}^{ζ \neq = 1}$ ) were found to be:
- FTI: $\simeq 20$
- pSEOBNR: $\simeq 10$
- TEOBPM: $\simeq 15$
Comparison with LVK Analyses: These values are significantly smaller than those inferred from existing LVK analyses using joint and hierarchical frameworks (which often reported Bayes factors in the range of $10^2$ to $10^4$ , indicating "decisive" evidence).
Interpretation: The authors argue that the lower Bayes factors do not imply a failure to find GR, but rather that the data require greater flexibility in modeling potential deviations. The standard frameworks, by enforcing stricter assumptions about the population distribution of deviations, artificially inflate the evidence for GR when the data actually contain a mix of consistent events and potential outliers or systematics.

Significance and Recommendations
The paper concludes that the current "all-or-nothing" approaches may systematically underestimate the evidence for heterogeneous departures from GR or overstate the evidence for GR when systematics are present. The mixture model offers a necessary flexibility to accommodate a population where a finite fraction of events may deviate from GR without imposing restrictive assumptions on the nature of those deviations.

The authors recommend adopting flexible mixture models as a standard practice for testing GR with growing gravitational-wave catalogues, unless there are specific physical motivations (such as in graviton-mass estimates) to impose more restrictive models. They emphasize that this flexibility comes at no cost: if deviations were truly uniform across the population, the mixture model would still detect them, but it avoids the risk of bias when the underlying physics is more complex or heterogeneous.