Hierarchical Subtraction with Neural Density Estimators as a General Solution to Overlapping Gravitational Wave Signals

TL;DR

This paper presents a fast, accurate, and adaptable hierarchical subtraction method using neural density estimators and likelihood-based resampling to effectively analyze overlapping gravitational wave signals in future detectors.

Contribution

It introduces a novel combination of hierarchical subtraction, neural density estimators, and likelihood resampling for efficient overlapping GW signal inference.

Findings

Hierarchical subtraction converges with enough iterations.

Neural density estimators enable rapid posterior sampling.

Method is adaptable to various source types and separations.

Abstract

Overlapping gravitational wave (GW) signals are expected in the third-generation (3G) GW detectors, leading to one of the major challenges in GW data analysis. Inference of overlapping GW sources is complicated - it has been reported that hierarchical inference with signal subtraction may amplify errors, while joint estimation, though more accurate, is computationally expensive. However, in this work, we show that hierarchical subtraction can achieve accurate results with a sufficient number of iterations, and on the other hand, neural density estimators, being able to generate posterior samples rapidly, make it possible to perform signal subtraction and inference repeatedly. We further develop likelihood-based resampling to accelerate the convergence of the iterative subtraction. Our method provides fast and accurate inference for overlapping GW signals and is highly adaptable to various source types and time separations, offering a potential general solution for overlapping GW signal analysis.