⚛️ general relativity

Deglobalising the fit: how to search for physics beyond vacuum General Relativity in LISA catalogues

This paper proposes a two-step post-processing strategy to efficiently search for deviations from vacuum General Relativity in LISA data by quantifying how overlapping signals mask new physics through a "masking ratio" and identifying a minimal set of correlated nuisance sources that must be re-analyzed to recover the full statistical significance of potential deviations.

Original authors: Francisco Duque, Jonathan Gair

Published 2026-10-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Francisco Duque, Jonathan Gair

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

The universe speaks to us through ripples in the fabric of space and time, known as gravitational waves. For decades, scientists have listened to these whispers from colliding black holes and neutron stars, using them to test the laws of gravity. The most successful theory we have is Einstein's General Relativity, which describes gravity as the curvature of space caused by mass. However, physicists suspect that this theory might be incomplete, perhaps needing a small correction to explain the mysterious dark energy or dark matter that fills the cosmos. To find these corrections, researchers must listen to the gravitational waves with extreme precision, looking for tiny deviations from the patterns Einstein predicted.

The challenge is that the universe is not quiet. A future space-based observatory called LISA, scheduled to launch in the 2030s, will not just hear a few clear signals; it will be flooded with millions of overlapping ripples from binary star systems across the galaxy. Imagine trying to hear a single violin solo while sitting in the middle of a stadium where every seat is occupied by another musician playing a slightly different tune. To make sense of this cacophony, scientists must perform a "global fit," a massive computational effort that tries to model and subtract every single source simultaneously to isolate the one they are interested in. This process is so complex that running it from scratch for every new theory of gravity the scientific community wants to test is impossible. Instead, the plan is to do the heavy lifting once with the standard theory, publish the results, and then let other scientists re-analyze specific targets later. But this two-step approach carries a hidden danger: the very act of subtracting the other stars in the first step might accidentally remove the tiny signal of new physics the researchers are looking for.

Francisco Duque and Jonathan Gair have developed a method to detect and fix this problem before it ruins the search for new physics. They realized that when the initial analysis subtracts the millions of background stars, it might inadvertently "eat" part of the strange signal hidden within a target star. Because the background stars are so numerous and their signals overlap with the target, they can absorb the unusual features of the new physics, making it look like the standard theory is correct even when it is not. The researchers showed that this masking effect can be measured by a single number, which they call the masking ratio. This number tells scientists exactly how much of the evidence for a new theory is lost when they skip the full, expensive analysis and just look at the leftover data. If the number is close to one, the simple method works fine. If it is low, the new physics is being hidden, and the scientists know they must go back and re-analyze the target star together with a few specific background stars that are doing the hiding.

To prove their idea works, the team built a simplified simulation involving twenty-one overlapping signals, where one was the target and the other twenty were the background noise. They injected a small, fake deviation from standard gravity into the target signal and then ran their test. They found that if they simply subtracted the background stars and looked at the target alone, the strange signal was significantly weakened, and the scientists would have missed it or reported it with much less confidence. However, by using their new algorithm, they could identify which of the twenty background stars were responsible for absorbing the signal. In their simulation, just two of the background stars were doing most of the hiding. When the researchers re-analyzed the target star together with just those two culprits, they recovered the full strength of the signal, matching the result they would have gotten if they had run the impossible, full-scale analysis of all twenty-one stars at once.

The beauty of this method is that it does not require scientists to guess which stars are the problem or to run expensive new simulations. They can calculate the masking ratio and identify the troublemakers using the data already produced by the initial analysis. This means that when LISA finally launches and begins its work, the scientific community can efficiently test new theories of gravity without getting bogged down by the sheer volume of data. They can quickly decide whether a simple re-check of a single star is enough, or if they need to bring a small group of neighbors into the conversation to ensure they are not missing a fundamental truth about the universe. By knowing exactly how much information is lost in the subtraction process, researchers can ensure that their search for the next great discovery in physics remains sharp, accurate, and free from the blind spots created by the crowded universe around them.

Technical Summary: Deglobalising the fit: how to search for physics beyond vacuum General Relativity in LISA catalogues

Problem Statement
The Laser Interferometer Space Antenna (LISA) is expected to observe approximately $10^7$ overlapping gravitational-wave signals, necessitating a "global fit" to extract all sources simultaneously. The current plan involves an initial global fit using vacuum General Relativity (GR) models for all binary systems. Subsequent scientific investigations, such as testing for deviations from vacuum GR, are expected to rely on post-processing these public catalogues rather than rerunning the computationally prohibitive global fit for every new physical model.

However, the authors identify a critical flaw in a two-step post-processing strategy (fitting a target source alone after subtracting nuisance sources): correlations between overlapping sources can cause nuisance sources to "absorb" part of a beyond-vacuum effect during the initial vacuum fit. When these biased nuisance sources are subtracted, the absorbed effect is removed from the data, never reaching the reanalysis. Consequently, post-processing may fail to detect existing deviations (masking) or report upper limits that are spuriously tight (overconfident) because it treats nuisance sources as perfectly known rather than as fitted parameters that could have absorbed the signal.

Methodology
The paper develops a framework based on the linear-signal approximation to quantify and mitigate this masking effect.

Geometric and Statistical Formulation:
The authors model the data space using inner products weighted by the noise covariance. They define the "vacuum manifold" (parameters $\theta_v$ and nuisance sources $\theta_c$ ) and the "beyond-vacuum" direction ( $\theta_b$ ). A beyond-vacuum effect displaces the model in data space. The component of this displacement lying within the tangent space of the vacuum manifold can be absorbed by adjusting vacuum and nuisance parameters. Only the orthogonal component is observable.
- In a Global Fit, the orthogonal component is projected onto the space spanned by the target's vacuum parameters and the nuisance sources ( $V \oplus C$ ).
- In Post-Processing, the nuisance sources are subtracted (effectively fixed), leaving only the target's vacuum parameters ( $V$ ) to absorb the effect.
Detection Statistics and Ratios:
The authors derive detection statistics for both strategies. They introduce two key dimensionless quantities to characterize the loss of information in post-processing:
- The Masking Ratio ( $R$ ): For a strong deviation, $R$ is the factor by which the detection significance is reduced. If a global fit detects a deviation at $n\sigma$ , post-processing reports it at $R n\sigma$ .
- The Null Ratio ( $R_0$ ): In the absence of a deviation, $R_0$ quantifies the overconfidence of the upper limits. It represents the factor by which the posterior width shrinks compared to the global fit, leading to bounds that are too tight by a factor of $\sqrt{R_0}$ .
- The Survival Ratio: A weighted average of $R^2$ and $R_0$ that describes the fraction of the expected statistic surviving post-processing for effects of any size.
Computational Estimation and Source Selection:
Crucially, the authors show that $R$ and $R_0$ can be estimated directly from the posterior samples of the initial vacuum global fit without rerunning the fit. They propose a greedy algorithm to identify which specific nuisance sources contribute most to the masking:
- The algorithm ranks nuisance sources based on how much of the "beyond-vacuum" direction they can reproduce (absorb) after the target's own vacuum parameters are projected out.
- It iteratively adds the most correlated nuisance sources to a set $S$ until the "residual masking ratio" (the masking remaining after refitting $S$ jointly with the target) is sufficiently close to 1.
- This allows for a "down-scaled" reanalysis where only the target and a small subset of critical nuisance sources are refitted jointly, recovering the global fit's sensitivity at a fraction of the computational cost.
Averaged vs. Simple Post-Processing:
The paper compares "simple" post-processing (subtracting the best-fit nuisance parameters) with "averaged" post-processing (marginalizing over the posterior uncertainty of the nuisance parameters).
- Simple post-processing is computationally cheaper and better for detecting deviations (higher significance) but yields tighter, potentially overconfident upper limits in null tests.
- Averaged post-processing broadens the posterior, recovering upper limits that are statistically consistent with the global fit, but reduces detection significance by at most a factor of $\sqrt{2}$ .

Key Results
The methodology is validated using a toy model of 21 overlapping chirping sinusoids (1 target, 20 nuisance sources) with full Markov-chain Monte Carlo (MCMC) inference.

Validation of Ratios: The linear-signal predictions for $R$ and $R_0$ matched the MCMC results within statistical errors. For example, in a vacuum injection, the post-processing uncertainty was narrower than the global fit by a factor matching the predicted $R$ .
Bias Recovery: When a beyond-vacuum effect was injected, simple post-processing recovered a biased value (shifted toward vacuum) with reduced significance. Refitting the target jointly with the specific nuisance sources identified by the greedy algorithm (Group B in the toy model) successfully recovered the true value and significance, closely matching the full global fit.
Non-linearity and Expansion Points: The study noted that the ranking of nuisance sources can depend on the expansion point (the value of the beyond-vacuum parameter). While the null-point ranking identified one group as dominant, the ranking inverted when evaluated at the true injected value. The authors propose sweeping the expansion point or iteratively updating it to ensure robust source selection.
Averaging Strategy: The averaged post-processing strategy successfully calibrated upper limits in the vacuum case but reduced detection significance in the beyond-vacuum case, confirming the theoretical trade-off.

Significance and Claims
The paper claims to provide a computationally efficient, practical "recipe" for testing General Relativity and other beyond-vacuum models in LISA data without the need to rerun the full global fit for every hypothesis.

Feasibility: It demonstrates that the loss of sensitivity in post-processing can be quantified and minimized by selectively refitting only a small subset of correlated nuisance sources.
Applicability to LISA: The method is designed to work with the posterior samples released from the initial vacuum global fit. It allows the scientific community to assess which sources require a down-scaled reanalysis and which can be safely analyzed in isolation.
Robustness: The authors argue that their estimator, derived from posterior samples, is robust to deviations from the linear-signal approximation and can be applied to realistic mock datasets (such as the Mojito dataset) to validate strategies before LISA's launch.
Systematics: The framework also offers a way to study how waveform systematics or instrumental artifacts in one source can propagate biases to overlapping sources, potentially mimicking deviations from GR.

The authors maintain a modest tone, acknowledging that the linear-signal approximation has limitations (e.g., curvature effects, non-Gaussian posteriors) and that the final selection of sources may require iterative refinement or a conservative threshold. They emphasize that while their toy model was simple, the formalism is directly applicable to the complex, high-dimensional data expected from LISA.

Technical Summary: Deglobalising the fit: how to search for physics beyond vacuum General Relativity in LISA catalogues

More like this