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

Technical Summary: Factorized Neural Posterior Estimation for ppE Parameters in Gravitational Waves

Problem Statement
The direct detection of gravitational waves (GWs) has confirmed General Relativity (GR), yet testing GR under strong-field conditions requires estimating parameterized post-Einsteinian (ppE) deviation parameters within waveform models. Traditional Bayesian inference methods, such as Markov Chain Monte Carlo (MCMC) and nested sampling (e.g., LALInference, Bilby), provide reliable posterior estimates but suffer from prohibitive computational costs. These methods struggle to meet the real-time demands and handle the surging data volumes expected from next-generation GW detectors. Furthermore, high-dimensional joint estimation of ppE parameters alongside standard binary black hole (BBH) parameters introduces complex couplings, leading to optimization difficulties, mode collapse, and slow convergence, particularly when dealing with narrow posteriors or multimodal likelihood landscapes.

Methodology
The authors propose a Factorized Neural Posterior Estimation (NPE) framework designed to perform rapid and statistically reliable inference directly from binary black hole signals. The core innovation lies in decomposing the high-dimensional joint posterior estimation problem into a series of independent, conditional estimation tasks.

• Factorization Strategy: Instead of training a single high-dimensional flow model for all parameters, the framework constructs independent normalizing flow models for each of the nine ppE deviation parameters ($\delta\chi_0$ to $\delta\chi_7$). The joint posterior is approximated as:
  $$P(\theta|x) \approx P(\theta_{\text{basic}}|x) \prod_{i=7}^{15} P(\delta\chi_i | x, \theta^*_i)$$
  where $\theta^*_i$ represents all other parameters excluding the target $\delta\chi_i$. This transforms the problem into nine parallel conditional estimation tasks.

• Network Architecture:

  • Hybrid Feature Extraction: A hybrid embedding network combines a Convolutional Neural Network (CNN) and a Residual Neural Network (ResNet). Preprocessed strain data (whitened and normalized) passes through a 1D CNN for downsampling and feature extraction, followed by a deep ResNet (128 residual blocks) to capture local temporal features and global dependencies.
  • Conditional Embedding: To address physical couplings (e.g., between mass, spin, and ppE parameters), a separate Multi-Layer Perceptron (MLP) encodes the "conditioning" parameters ($\theta^*_i$) into embedding vectors. These vectors are concatenated with the data features before being fed into the NPE model. This ensures the inference for a specific ppE parameter is guided by the known physical state of the source.
  • Normalizing Flow: The NPE model consists of 10 stacked affine coupling layers, each with 5 hidden layers of 512 neurons. It learns the direct mapping from observed data and conditional parameters to the posterior distribution of the target deviation parameter.

• Training and Data:

  • Dataset: A large-scale simulated dataset of $2 \times 10^7$ independent BBH samples was generated using PyCBC with the IMRPhenomXPHM waveform model. The simulation includes 15 fundamental physical parameters and 9 ppE parameters, with noise modeled using the aLIGOZeroDetHighPower power spectral density.
  • Loss Function: The models are trained using the negative log-likelihood loss, maximizing the conditional log-likelihood of the true parameters given the data.
  • Optimization: Training utilizes the AdamW optimizer with early stopping based on validation loss to prevent overfitting.

Key Contributions

1. Factorized NPE Framework: The introduction of a parameter-independent approach that decouples the estimation of ppE parameters, significantly reducing model complexity and training difficulty compared to joint high-dimensional estimation.
2. Conditional Embedding Mechanism: The explicit integration of physical source parameters as conditional inputs to guide the inference, mitigating biases caused by parameter degeneracies and preventing artificial decoupling.
3. Hybrid Deep Learning Architecture: The use of a CNN-ResNet hybrid for feature extraction from noisy GW signals, enabling the capture of subtle correlations between waveforms and deviation parameters.
4. Fully Amortized Inference: The method learns a direct mapping from data to the posterior distribution, eliminating the need for reduced-order models, iterative MCMC sampling, or assumptions about parametric distribution forms.

Results

• Statistical Reliability: Comprehensive validation using 1,000 independent simulated events shows that the posterior estimates pass the Kolmogorov-Smirnov (KS) test. The empirical coverage probabilities are close to theoretical targets, with specific parameters ($\delta\chi_1, \delta\chi_3, \delta\chi_{5l}$) showing outstanding calibration where coverage curves nearly coincide with the ideal diagonal.
• Posterior Quality: Comparisons with MCMC (using the Dynesty nested sampler) reveal that the NPE framework produces posterior distributions with shapes consistent with MCMC results. For some parameters ($\delta\chi_0$ to $\delta\chi_2$), the NPE distributions are slightly narrower, suggesting lower uncertainty, while for others ($\delta\chi_3$ to $\delta\chi_7$), the uncertainty estimates are comparable. Crucially, both methods yield the same core physical conclusion: no significant deviation from GR was detected in the test scenarios.
• Computational Performance: The NPE method achieves a speedup factor of approximately $9 \times 10^4$ compared to traditional MCMC. While MCMC requires hours (e.g., ~27,000 seconds for $\delta\chi_0$), the NPE framework performs complete posterior sampling for a single event in less than 0.4 seconds (milliseconds scale).
• Real-Data Application: The framework was applied to the GW150914 event. The resulting posterior distributions for both ppE parameters and fundamental BBH parameters (masses, spin, distance) are in good agreement with existing official results, despite the difference between simulated Gaussian noise and the non-stationary noise of the actual event.

Significance
The paper claims that this work demonstrates the great potential of deep learning for GW parameter estimation, specifically for testing General Relativity. By providing a scalable, verifiable, and computationally efficient technical solution, the proposed framework addresses the bottleneck of real-time analysis for future GW detectors (such as the Einstein Telescope, Taiji, and TianQin). It enables rapid constraints on ppE parameters immediately following a detection, facilitating timely observational decisions and advancing the search for physics beyond GR. The method's ability to maintain statistical rigor while achieving millisecond-scale inference represents a viable path toward handling the high-volume data streams of the next generation of gravitational wave astronomy.