Testing the robustness of simulation-based gravitational-wave population inference

Abstract

Gravitational-wave population studies have become more important in gravitational-wave astronomy because of the rapid growth of the observed catalog. In recent studies, emulators based on different machine learning techniques are used to emulate the outcomes of the population synthesis simulation with fast speed. In this study, we benchmark the performance of two emulators that learn the truncated power-law phenomenological model by using Gaussian process regression and normalizing flows techniques to see which one is a more capable likelihood emulator in the population inference. We benchmark the characteristic of the emulators by comparing their performance in the population inference to the phenomenological model using mock and real observation data. Our results suggest that the normalizing flows emulator can recover the posterior distribution by using the phenomenological model in the population inference with up to 300 mock injections. The normalizing flows emulator also underestimates the uncertainty for some posterior distributions in the population inference on real observation data. On the other hand, the Gaussian process regression emulator has poor performance on the same task and can only be used effectively in low-dimension cases.

Deep learning and Bayesian inference of gravitational-wave populations: hierarchical black-hole mergers

“…Compared to previous work [40,41], the approach presented here replaces the approximation of simulated binary BH distributions via histograms with Gaussian KDEs, and the emulation of these distributions across both the source-and population-level parameter spaces via GPR with DNNs. While GPR has been shown to be an ineffective approach in higher dimensions [42], alternative deep learning techniques such as normalizing flows [44] have proven successful [43,45,46]. Rather than training on probability density evaluations, the required number of which scales with an exponent equal to the dimensionality, these models are trained directly on samples from the true distribution, in which case the size of the training data scales linearly with dimensionality.…”

“…Compressed principal components of binned simulation data were emulated over low-(typically one-or two-) dimensional source-and population-level parameter spaces using Gaussian process regression (GPR). However, this emulation approach was shown to be unsuitable for extension to more complex higher-dimensional modeling scenarios due to poor predictive accuracy and infeasible computational requirements [41,42]. These issues were tackled in Ref.…”