Kerr Quasinormal Modes without Variable Separation: A Two-Dimensional Hyperboloidal Teukolsky Solver with Physics-Informed Neural Networks

summary

The gist

Kerr Quasinormal Modes without Variable Separation: A Two-Dimensional Hyperboloidal Teukolsky Solver with Physics-Informed Neural Networks solves the gravitational quasinormal-mode (QNM) eigenvalue…

In short

This research solves for Kerr black hole quasinormal modes using a two-dimensional hyperboloidal formulation of the Teukolsky equation combined with Physics-Informed Neural Networks (PINNs). This method bypasses traditional angular separation techniques, allowing for the calculation of QNMs directly in coordinate space. The results show high accuracy, preserving key ringdown time scales.

Key concepts

Teukolsky Equation
This is a fundamental equation in general relativity used to describe perturbations (like gravitational waves) around a Kerr black hole. It governs how these perturbations propagate and evolve in the curved spacetime geometry of the black hole.
Hyperboloidal Formulation
Instead of separating variables, this mathematical approach uses a specific two-dimensional formulation of the Teukolsky equation. This allows the solver to handle the coupled radial and angular structure simultaneously within a computational domain.
Physics-Informed Neural Networks (PINNs)
PINNs are neural networks trained not just on data, but also on physical laws, such as the PDE itself. The network learns to minimize a loss function that enforces the governing equations of the Teukolsky equation, leading to accurate solutions.
Quasinormal Modes (QNMs)
QNMs are characteristic oscillations or 'ringdown' patterns emitted by a black hole after it has been disturbed. These modes are defined by complex frequencies and determine how quickly the gravitational disturbance fades away.

Terminology used across episodes

This episode discusses

The paper

Kerr Quasinormal Modes without Variable Separation: A Two-Dimensional Hyperboloidal Teukolsky Solver with Physics-Informed Neural Networks · Read on arXiv

Antonio Ferrer-Sáchez, Daniela D. Doneva, José D. Martín-Guerrero, Stoytcho S. Yazadjiev, Roberto Ruiz de Austri-Bazan, Yolanda Vives-Gilabert, José A. Font

Intelligent Data Analysis Laboratory (IDAL) · Department of Electronic Engineering, University of Valencia · Departamento de Astronomıa y Astrofısica, Universitat de Valenciľ · Theoretical Astrophysics, Eberhard Karls University of Tüdingen · Valencian Graduate School and Research Network of Artificial Intelligence (ValgrAI) · Department of Theoretical Physics, Faculty of Physics, Sofia University · Institute of Mathematics and Informatics, Bulgarian Academy of Sciences · Instituto de Fısica Corpuscular CSIC-UV · Observatori Astronıomic, Universitat de Valenciľ

We use physics-informed neural networks (PINNs) to solve the gravitational quasinormal-mode (QNM) eigenvalue problem for Kerr spacetime directly in the two-dimensional hyperboloidal formulation of the Teukolsky equation. This formulation does not require separation of variables and thus retains the coupled radial--angular structure. Such a scheme provides a prototype for calculating the QNMs of beyond-Kerr black holes for which the perturbation equations are non-separable. Sequences with increasing angular momentum are constructed, reaching close to the extremal limit. We focus on the fundamental modes (,m,n)=(2,0,0), (2,1,0), (2,2,0), (3,3,0) and (4,4,0), together with the first two overtones (2,2,1) and (2,2,2). After training, the frequency at every spin is refined by solving the discretized equation for it, so that it depends much less on where the stochastic training happens to stop. Comparison with continued-fraction values shows that every reported real and imaginary frequency component of the production network remains below 0.23% error, with a median deviation of 2 times10-4%. For the fundamental modes, both components agree within about 10-3% up to a/M=0.97, and the real parts within 5 times10-4% at every spin. Sub-percent accuracy is maintained in the near-extremal regime, where the damping rate becomes small and the modes are longest-lived; there the damping errors grow, and the largest, 0.23%, belongs to the (2,2,1) overtone. The results establish a non-spectral numerical route to multidimensional black-hole perturbation eigenproblems which does not match the substantially higher precision of dedicated Kerr solvers but offers greater flexibility and requires less analytical pre-processing. The proposed approach is designed to be applicable to other backgrounds whose perturbation equations do not separate.

Transcript

Introduction to the show: ident: General Relativity and Quantum Cosmology. The day in general relativity and quantum cosmology.

Kai: We're looking at the paper: "Kerr Quasinormal Modes without Variable Separation".

Mira: Here's the gist of the paper. Kerr Quasinormal Modes without Variable Separation: A Two-Dimensional Hyperboloidal Teukolsky Solver with Physics-Informed Neural Networks solves the gravitational quasinormal-mode (QNM) eigenvalue problem for Kerr spacetime directly in a two-dimensional hyperboloidal formulation of the Teukolsky equation using physics-informed neural networks (PINNs) The PINN approach allows for solving the QNM problem without requiring angular separation or introducing an angular separation constant, providing a prototype for calculating QNMs of beyond-Kerr black holes where perturbation equations are non-separable The study employs a two-dimensional hyperboloidal formulation of the Teukolsky equation to avoid the need for separation of variables, retaining the coupled radial–angular structure The regular field is represented directly in coordinate space, and automatic differentiation supplies the derivatives in the hyperboloidal operator The transformation from Boyer–Lindquist coordinates to hyperboloidal coordinates is defined by specific functions for time and azimuthal dependence This formulation ensures that the two physical QNM boundaries are included directly in the computational domain, and while asymptotic conditions can be encoded to be automatically satisfied, the eigenfunction is singled out by requiring it to be regular on (σ, u) ∈

0, 1: ×

−1, 1: Training Signal Construction The PINN minimizes a total loss function that includes the meansquared PDE residual of the transformed Teukolsky equation The objective function is composed of several terms, including the principal term LPDE which is weighted by unit weight, and auxiliary terms such as relative residual loss Lrel and mode-selection loss Lmode The mode-selection loss includes terms like Lcurv to penalize unnecessary angular curvature and Lcoh to promote coherence across radial slices The total loss function is summarized as a combination of these physical constraints and numerical aids, with weights adjusted via an activation schedule over epochs Numerical Robustness and Accuracy The results demonstrate high accuracy, showing that every reported real and imaginary frequency component remains below 0.5% error with a median deviation of 0.1% across independent benchmark evaluations The method is robust against changes in network size, with the production baseline of the 4 × 84 network achieving errors below 0.5% for every retained mode in the near-extremal interval Furthermore, comparison with dedicated Kerr solvers shows that this non-spectral numerical route offers greater flexibility and requires less analytical pre-processing Physical Implications of Results The computed complex frequencies preserve the principal Kerr ringdown time scales with a very good accuracy, as evidenced by the agreement between PINN results and reference Leaver continued-fraction values The analysis distinguishes between zero-damped modes and damped modes near extremality, showing that co-rotating zero-damped branches approach a specific scaling relation where −I(ω)/κH → n + 1/2 The comparison of ringdown waveforms confirms that the recovered complex frequencies preserve the principal Kerr ringdown time scales with a very good accuracy Limitations and Future Directions The primary limitations involve the nonlinear optimization cost, the need for continuation and branch diagnostics, and lower precision than dedicated spectral eigensolvers for smooth Kerr problems The clearest prospective application is in stationary axisymmetric compact-object backgrounds supplied as numerical coefficient fields on irregular or adaptively sampled domains Such applications can utilize coordinate-space residual methods without constructing a global expansion or differentiation matrix The study emphasizes that the neural representation imposes no radial–angular product ansatz, angular separation constant, Leaver-type radial asymptotic prefactor, analytic boundary mask, Chebyshev expansion, or spectral differentiation matrix Data and Code Availability The numerical values underlying the main results are reported in the tables of this manuscript The complete training configurations and machine-readable result files are preserved by the authors and are available from the corresponding author upon reasonable request The timing protocol specifies that all model training and result generation were performed on a MacBook Pro equipped with an Apple M5 chip, a 10-core CPU, and 32 GB of memory The benchmark keeps the network input and output dimensions fixed, characterizing architecture-level cost across the modes The maximum component error for the production baseline is reported in Table III References

1: K.

Kai: D.

Mira: Kokkotas and B.

Kai: G.

Paper summary: Kai: So to summarize this paper, they introduce the concept of solving Kerr quasinormal modes without needing variable separation by employing a two-dimensional hyperboloidal Teukolsky solver with physics informed neural networks. They argue that this method provides a way to calculate these QNMs directly in a two-dimensional hyperboloidal formulation of the Teukolsky equation.

Mira: The core thesis is that because they use this specific formulation, they don't have to worry about separation of variables or introducing an angular separation constant, which is what motivates the search for alternative methodologies that are more straightforward and less sensitive to initial assumptions.

Lev: They frame physics informed neural networks as a different way to discretize the continuum problem where the unknown field is represented by a trainable network instead of being expanded in a prescribed basis. This lets them minimize the Teukolsky residual at specific points in that compactified domain, and automatic differentiation provides those coordinate derivatives for the optimization problem.

Kai: So what this means for us is that they are using the PINN to optimize both the field and its complex frequency simultaneously, changing how they evaluate residuals. They also use several terms in their total loss function, like the principal term LPDE weighted by a unit weight, relative residual loss Lrel, and mode-selection loss Lmode.

Mira: That mode selection loss includes things like Lcurv to penalize unnecessary angular curvature and Lcoh to promote coherence across radial slices of the computation. These auxiliary terms are what guide the network toward finding physically meaningful modes rather than just any mathematical solution.

Lev: The paper shows that this approach, despite its flexibility, still requires careful handling because they have to perform continuation and branch diagnostics when looking for solutions in a complex frequency domain.

Kai: Overall, the main claim is that this non-spectral numerical route offers greater flexibility and requires less analytical pre-processing compared to traditional spectral eigensolvers for smooth Kerr problems.

Mira: It’s really about showing that you can solve these QNM problems using coordinate space residual methods without constructing a global expansion or differentiation matrix, which is a significant change in how we think about solving these types of partial differential equations.

Conclusion: Kai: So looking at the title of this paper, "Kerr Quasinormal Modes without Variable Separation: A Two-Dimensional Hyperboloidal Teukolsky Solver with Physics Informed Neural Networks," it really captures the essence of what they did. They tackled the QNM problem directly on Kerr spacetime without relying on variable separation.

Mira: And the authors are using PINNs to provide a trainable discretization of that continuum problem, minimizing physical constraints like residual terms and mode selection losses to find these complex frequencies directly within that hyperboloidal framework.

Lev: What this implies for us is that we have a new tool for calculating QNMs where you don't need to impose the spectral ansatz or worry about the separation constant; it’s a direct path to finding those modes using modern machine learning techniques.

Kai: It means we can calculate these complex frequencies with high accuracy, and they preserve those important ringdown time scales, which is what makes this result meaningful for astrophysics.

Mira: So in short, the paper presents a method that replaces the standard spectral approach with a physics-informed neural network approach to solve QNM problems on Kerr backgrounds using a hyperboloidal formulation.

More episodes

← Home