Scalable and Precise Population Inference of Gravitational Waves with Density Estimation
ORAL
Abstract
Population analyses of gravitational wave (GW) infers astrophysical model parameters which advance our understanding of astrophysics, including black hole formation channels, the neutron star equation of state, and cosmology.
These parameters are estimated through hierarchical Bayesian statistics that account for selection effects.
However, to maintain the accuracy of Monte Carlo integration, the number of samples required, which is constrained by the complexity of the waveform model, scales as the GW catalog size increases.
Furthermore, when inference involves complex models with exact physical relations, the posterior may contain narrow, high-probability regions that are poorly sampled in standard analyses, further constraining inference accuracy.
We evaluate the samples drawn from model on the density estimates of the posteriors to compute the likelihood.
This enables flexible sampling, bypassing the sampling limitation in Monte Carlo integration and allowing denser coverage of regions with narrow features.
We quantify the uncertainty of this approach and compare the performance of different machine learning density estimators, including Gaussian mixture and neural-based models, against the standard population-inference framework.
This method offers improved scalability for future gravitational-wave catalogs and provides a foundation for incorporating exact physical constraints in population studies, such as the neutron-star equation of state and cosmological inference.
These parameters are estimated through hierarchical Bayesian statistics that account for selection effects.
However, to maintain the accuracy of Monte Carlo integration, the number of samples required, which is constrained by the complexity of the waveform model, scales as the GW catalog size increases.
Furthermore, when inference involves complex models with exact physical relations, the posterior may contain narrow, high-probability regions that are poorly sampled in standard analyses, further constraining inference accuracy.
We evaluate the samples drawn from model on the density estimates of the posteriors to compute the likelihood.
This enables flexible sampling, bypassing the sampling limitation in Monte Carlo integration and allowing denser coverage of regions with narrow features.
We quantify the uncertainty of this approach and compare the performance of different machine learning density estimators, including Gaussian mixture and neural-based models, against the standard population-inference framework.
This method offers improved scalability for future gravitational-wave catalogs and provides a foundation for incorporating exact physical constraints in population studies, such as the neutron-star equation of state and cosmological inference.
–
Presenters
-
Man Chun Yeung
- University of Wisconsin - Milwaukee