Factorized neural posterior estimation for rapid and reliable inference of parameterized post-Einsteinian deviation parameters in gravitational waves

The direct detection of gravitational waves (GWs) by LIGO has strikingly confirmed general relativity (GR), but testing GR via GWs requires estimating parameterized post-Einsteinian (ppE) deviation parameters in waveform models. Traditional Bayesian inference methods like Markov chain Monte Carlo (MCMC) provide reliable estimates but suffer from prohibitive computational costs, failing to meet the real-time demands and surging data volume of future GW detectors. Here, we propose a factorized neural posterior estimation framework: we construct independent normalizing flow models for each of the nine ppE deviation parameters and effectively integrate prior information from other source parameters via a conditional embedding network. Leveraging a hybrid neural network with a convolutional neural network and a Residual Neural Network for feature extraction, our method performs rapid and statistically reliable posterior inference directly from binary black hole signals. Compared to conventional MCMC, our approach achieves millisecond-scale inference time with a speedup factor of $9 \times 10^4$. Comprehensive validations show that the posterior estimates pass the Kolmogorov-Smirnov test and achieve empirical coverage probabilities close to theoretical targets. This work demonstrates the great potential of deep learning for GW parameter estimation and provides a viable technical solution for real-time GR tests with next-generation detectors.

💡 Research Summary

The paper addresses the pressing need for rapid and reliable estimation of parameterized post‑Einsteinian (ppE) deviation parameters in gravitational‑wave (GW) data, a task that is essential for testing General Relativity (GR) in the strong‑field regime. Traditional Bayesian inference methods such as Markov chain Monte Carlo (MCMC) deliver accurate posterior distributions but are computationally prohibitive, taking minutes to hours per event—far too slow for the anticipated data deluge from next‑generation detectors like the Einstein Telescope and Cosmic Explorer.

To overcome this bottleneck, the authors develop a factorized Neural Posterior Estimation (NPE) framework that decomposes the full 16‑dimensional posterior (seven intrinsic/extrinsic source parameters plus nine ppE deviation parameters) into a product of a basic‑parameter posterior and nine conditional posteriors for each ppE term. Mathematically, they approximate

 P(θ|x) ≈ P(θ_basic|x) ∏_{i=1}^{9} P(δχ_i | x, θ*_i),

where θ*_i denotes all parameters except the i‑th deviation. This conditional independence assumption reduces the high‑dimensional joint estimation problem to nine parallel, low‑dimensional tasks, dramatically simplifying training and inference.

Dataset and preprocessing
The authors generate a large synthetic dataset using PyCBC with the IMRPhenomPv2 waveform model, which supports nine ppE phase correction terms. They uniformly sample 2 × 10⁶ parameter sets within physically motivated bounds (masses 20–100 M⊙, spins −1 to 1, luminosity distance 1000–3000 Mpc, SNR 14–20). Each sample consists of a 1‑second strain time series sampled at 500 Hz. Colored Gaussian noise is added using the aLIGOZeroDetHighPower power‑spectral density, and a frequency‑domain whitening operation is applied to suppress low‑frequency noise. The strain is then normalized to zero mean and unit variance before being fed to the network.

Network architecture
Feature extraction is performed by a three‑layer one‑dimensional convolutional neural network (CNN) that downsamples the raw waveform and captures local temporal patterns. The resulting feature tensor is flattened and passed to a deep Residual Network (ResNet) comprising 128 residual blocks. Each block contains two fully‑connected layers with 512 neurons, batch normalization, dropout, and skip connections that preserve gradient flow. The ResNet outputs a compact embedding vector that serves as the conditioning input for a normalizing‑flow model (e.g., RealNVP or Neural Spline Flow) dedicated to a single ppE parameter. Thus, nine independent flow models are trained in parallel, each learning the conditional density Q_ϕ(δχ_i | x, θ*_i).

Training and validation
Training maximizes the conditional log‑likelihood log Q_ϕ, effectively minimizing the Kullback‑Leibler divergence between the learned and true posteriors. Validation employs two statistical diagnostics: (1) the Kolmogorov‑Smirnov (KS) test to assess the overall shape agreement of the estimated marginal distributions with the ground‑truth posteriors, and (2) empirical coverage probability (ECP) tests for the 90 % credible intervals. All nine ppE parameters achieve KS p‑values above 0.05 and ECP values between 0.88 and 0.94, indicating well‑calibrated posteriors.

Performance
On an NVIDIA RTX 3090 GPU, inference for a single GW event takes roughly 2 ms, a speed‑up factor of ≈ 9 × 10⁴ compared with a typical MCMC run (~180 s). The mean absolute error in the estimated δχ_i values is below 0.02, and the 95 % credible intervals contain the true values in 93 % of the test cases. This demonstrates that the factorized NPE approach delivers both speed and statistical fidelity.

Limitations and future work
The conditional independence assumption, while computationally advantageous, neglects potential correlations between ppE deviations and the basic source parameters. In real signals, such correlations could introduce bias if not modeled. Extending the framework to a joint flow that captures these dependencies, or employing hierarchical conditioning, is a natural next step. Moreover, the current validation uses idealized Gaussian noise and whitening; robustness to non‑stationary, non‑Gaussian detector artifacts remains to be demonstrated. Transfer learning or domain‑adaptation techniques could bridge the gap to actual LIGO‑Virgo‑KAGRA data. Finally, the study focuses on binary‑black‑hole mergers; applying the method to binary neutron star or neutron‑star–black‑hole systems will require additional waveform modeling and possibly larger training sets.

Conclusion
By factorizing the posterior, leveraging a hybrid CNN‑ResNet encoder, and training independent normalizing‑flow models for each ppE deviation, the authors present a scalable, real‑time solution for testing GR with gravitational waves. The method achieves millisecond‑scale inference, maintains rigorous statistical calibration, and offers a promising pathway toward automated, on‑the‑fly GR tests in the era of high‑rate GW astronomy.