⚛️ general relativity

Quasinormal modes from numerical relativity with Bayesian inference

This paper introduces a flexible Gaussian-process model to quantify numerical uncertainties in state-of-the-art numerical relativity waveforms, enabling efficient Bayesian inference of quasinormal modes without relying on expensive Markov chain Monte Carlo methods.

Original authors: Richard Dyer, Christopher J. Moore

Published 2026-06-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Richard Dyer, Christopher J. Moore

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 Big Picture: Listening to the "Ring" of a Black Hole

Imagine two black holes crashing into each other. When they merge, the resulting black hole doesn't just sit there; it "rings" like a bell that has just been struck. This ringing is called the ringdown.

According to Einstein's theory of gravity, this ring isn't a single tone. It's a complex chord made of specific notes, called Quasinormal Modes (QNMs). Just like a bell's tone tells you its size and shape, these "notes" tell us about the mass and spin of the new black hole. If we can hear these notes clearly, we can test if Einstein's theory holds up in the most extreme conditions in the universe.

The Problem: The "Static" on the Recording

Scientists use supercomputers to simulate these black hole crashes. These simulations produce a "recording" of the gravitational waves. However, because these are computer simulations, they aren't perfect. They have numerical noise—tiny errors that look like static or fuzz on a radio recording.

Traditionally, when scientists tried to find the "notes" (QNMs) in this recording, they used a method called least-squares fitting.

The Analogy: Imagine trying to find a specific melody in a song that has static on it. The old method was like saying, "I'll just draw a line through the noise to find the melody." It gives you one answer (the best guess), but it doesn't tell you how much you should trust that answer. It also assumes the static is random and simple, which isn't true for these complex simulations.

The Solution: A Smart "Noise Map" and a New Way to Listen

This paper introduces two major upgrades to how we listen to these black hole rings:

1. The "Smart Noise Map" (Gaussian Processes)

Instead of assuming the static is random, the authors built a Gaussian Process (GP) model.

The Analogy: Think of the computer simulation errors not as random static, but as a specific, predictable pattern of "fuzz" that changes over time. The authors took a huge library of 13 different black hole simulations and studied the "fuzz" in all of them. They trained a smart AI (the GP) to learn exactly what this fuzz looks like.
The Result: Now, when they analyze a new simulation, they don't just guess the noise; they have a custom-made map of where the errors are likely to be. This allows them to separate the real "ringing" from the computer "fuzz" much more accurately.

2. The "Instant Probability" Method (Bayesian Inference)

Once they have a map of the noise, they need to find the notes. The old way (least-squares) gives you a single point: "The note is here."

The New Way: The authors use Bayesian inference. Instead of giving you one point, it gives you a cloud of possibilities.
The Analogy: Imagine you are trying to guess the weight of a mystery object.
- Old Method: "It weighs exactly 5kg." (No idea if you are right or wrong).
- New Method: "It weighs between 4.8kg and 5.2kg, and I'm 95% sure it's in that range."
Why it's fast: Usually, calculating these "clouds" takes a supercomputer days to run. The authors found a mathematical trick (linearizing the model) that lets them calculate this cloud instantly—in milliseconds. It's as fast as the old method but gives you a full picture of the uncertainty.

What They Found

The team tested this new method on a specific black hole simulation (SXS:CCE:0001).

Better Precision: Because they used their "Smart Noise Map," the "cloud" of possible answers for the black hole's mass and spin was much tighter and more precise than with the old methods.
Detecting Faint Notes: They could tell with high confidence which "notes" (QNMs) were actually there and which were just noise. They introduced a new "Significance Score" (0 to 1) to say, "We are 99% sure this note exists," or "This note is probably just noise."
Checking the Fit: They created a new test (Posterior Predictive Check) to see if their model was trying too hard to fit the noise (overfitting) or missing the real signal (underfitting). They found that if they started listening too early in the ringdown, the model got confused by the crash noise, but if they waited a bit, the fit was perfect.

The Bottom Line

This paper doesn't discover a new black hole or prove a new theory of physics. Instead, it provides a better toolkit for the tools scientists already use.

Old Toolkit: A ruler and a guess.
New Toolkit: A laser scanner and a probability calculator that runs instantly.

By using this new toolkit, scientists can now analyze the "ringing" of black holes with much higher confidence, allowing them to test Einstein's theories with greater precision than ever before. The code and methods are now open for everyone to use.

Technical Summary: Quasinormal Modes from Numerical Relativity with Bayesian Inference

Problem Statement
Numerical relativity (NR) provides high-fidelity predictions for gravitational-wave (GW) signals from merging black holes, particularly during the ringdown phase where the remnant settles into a Kerr state. This phase is characterized by quasinormal modes (QNMs)—damped sinusoidal oscillations containing information about the remnant's mass and spin. While traditional analysis relies on linear least-squares fitting to extract QNM parameters, these methods face significant limitations:

Uncertainty Quantification: Least-squares fitting implicitly assumes stationary white noise and yields only point estimates without robust uncertainty quantification. This is insufficient for probing subdominant or nonlinear effects (e.g., overtones, quadratic modes) where numerical uncertainties in NR waveforms are non-negligible.
Computational Scalability: Bayesian approaches, which offer full posterior distributions, typically rely on Markov chain Monte Carlo (MCMC) methods. These are computationally expensive and do not scale well to models involving many QNMs.
Model Selection: Point estimates cannot determine whether the data supports the inclusion of specific QNMs, making it difficult to distinguish between physical signals and overfitting.

Methodology
The authors introduce a flexible Bayesian framework for fitting QNM models to NR waveforms, specifically utilizing Cauchy-Characteristic Evolved (CCE) waveforms from the Spectral Einstein Code (SpEC) catalog. The methodology consists of three core components:

Gaussian Process (GP) Model for Numerical Uncertainty:
Instead of assuming white noise, the authors construct a physically motivated GP model to characterize the numerical uncertainty in NR waveforms.
- Training: The GP is trained on the residuals between the highest (L5) and second-highest (L4) resolution simulations from a catalog of 13 CCE waveforms.
- Kernel Design: The kernel combines a stationary component (capturing periodicity related to the fundamental QNM frequency) with time-dependent amplitude modulation (modeling exponential decay during ringdown and constant noise floors at late times).
- Parameters: The model uses pooled hyperparameters ( $\lambda, \mu$ ) learned across the entire catalog, ensuring the uncertainty estimate is conservative and physically grounded.
Efficient Bayesian Inference via Linearization:
The QNM model is expressed as a sum of damped sinusoids. To avoid expensive MCMC sampling:
- Linearization: The model is linearized around a reference point (obtained via a preliminary least-squares fit with fixed asymptotic Bondi data mass and spin). The remnant mass ( $M_f$ ) and spin ( $\chi_f$ ) are treated as free parameters via Taylor expansion.
- Analytic Sampling: With a flat prior on the real and imaginary parts of the amplitudes (Prior 1), the posterior distribution becomes exactly multivariate Gaussian. This allows for the analytic calculation of the posterior mean and covariance (equivalent to the Fisher information matrix), enabling extremely fast sampling ( $O(1)$ ms for single-mode fits).
- Prior Reweighting: To accommodate priors flat on amplitude magnitudes (Prior 2), the authors employ importance sampling to reweight the analytic samples from Prior 1.
Significance and Model Validation:
- Significance Metric ( $S_\alpha$ ): A Bayesian measure of significance is defined to quantify how confidently the posterior excludes the hypothesis that a specific QNM amplitude is zero. This replaces ambiguous point-estimate comparisons.
- Posterior Predictive Checks (PPC): A PPC is implemented using the $L_2$ norm of the residuals as a test statistic. By comparing observed residuals against the sampling distribution of a generalized chi-squared variable (derived from the GP kernel eigenvalues), the method assesses whether a model underfits or overfits the data.

Key Results
The framework was applied to the SXS:CCE:0001 simulation, fitting overtone models to the $(2,2)$ spherical harmonic mode (and including the $(3,2)$ mode in a secondary model).

Uncertainty Modeling: The GP kernel significantly outperformed a simple white-noise (WN) kernel in describing the numerical residuals, as evidenced by higher log-likelihoods and lower Kullback-Leibler divergences.
Posterior Constraints: The GP-based posteriors for QNM amplitudes and remnant parameters ( $M_f, \chi_f$ ) were significantly narrower (more constraining) than those obtained with the WN kernel, demonstrating the value of accurate uncertainty modeling.
Stability and Significance:
- The fundamental mode showed high significance ( $S_\alpha \approx 1$ ) across a wide range of start times.
- Higher overtones showed decreasing significance at later start times, consistent with their faster decay rates.
- The inclusion of the $(3,2,0,+)$ mode (Model 2) stabilized the significance of lower overtones, resolving fluctuations observed in simpler models.
Model Fit Quality: PPCs revealed that while a model with up to $n=6$ overtones fits well at early start times ( $t_0 \lesssim 5M$ ), it begins to overfit the data at later times ( $t_0 \gtrsim 8M$ ). This diagnostic capability highlights the necessity of selecting appropriate start times and model complexity.

Significance and Claims
The paper positions this work as a methodological advancement rather than a discovery of new physical phenomena. The primary contributions are:

Scalability: The method achieves computational costs comparable to least-squares fitting while providing full Bayesian posterior distributions, making it viable for large-scale catalog analyses.
Robust Uncertainty Quantification: By training a GP on NR residuals, the approach provides a conservative, physically motivated estimate of numerical noise, moving beyond the unrealistic white-noise assumption.
Rigorous Model Selection: The introduction of the significance metric ( $S_\alpha$ ) and PPCs offers a statistically rigorous way to determine the presence of subdominant modes and prevent overfitting, addressing a known challenge in QNM extraction.
Reproducibility: The authors provide public access to the code, plotting scripts, and processed data, facilitating the application of these techniques to future GW catalogs and more complex NR simulations.

The authors conclude that this framework supports the analysis of subdominant and higher-order effects in the ringdown, such as nonlinear quadratic modes and power-law tails, by providing a reliable statistical foundation for interpreting NR data.