Gravitational-Wave Detection using Multivariate Analysis

TL;DR

This paper demonstrates that multivariate analysis, specifically boosted decision trees, significantly enhances the sensitivity of gravitational-wave burst detection, especially for signals with unknown waveforms, by improving classification accuracy and detection volume.

Contribution

The study applies boosted decision tree multivariate analysis to gravitational-wave burst detection, showing substantial sensitivity improvements over traditional methods.

Findings

BDTs increase detection sensitivity by up to 50%.

The method is robust against sky position and waveform variations.

Detection of diverse waveforms beyond training set is possible.

Abstract

Searches for gravitational-wave bursts (transient signals, typically of unknown waveform) require identification of weak signals in background detector noise. The sensitivity of such searches is often critically limited by non-Gaussian noise fluctuations which are difficult to distinguish from real signals, posing a key problem for transient gravitational-wave astronomy. Current noise rejection tests are based on the analysis of a relatively small number of measured properties of the candidate signal, typically correlations between detectors. Multivariate analysis (MVA) techniques probe the full space of measured properties of events in an attempt to maximise the power to accurately classify events as signal or background. This is done by taking samples of known background events and (simulated) signal events to train the MVA classifier, which can then be applied to classify events of unknown type. We apply the boosted decision tree (BDT) MVA technique to the problem of detecting gravitational-wave bursts associated with gamma-ray bursts. We find that BDTs are able to increase the sensitive distance reach of the search by as much as 50%, corresponding to a factor of ~3 increase in sensitive volume. This improvement is robust against trigger sky position, large sky localisation error, poor data quality, and the simulated signal waveforms that are used. Critically, we find that the BDT analysis is able to detect signals that have different morphologies to those used in the classifier training and that this improvement extends to false alarm probabilities beyond the 3{\sigma} significance level. These findings indicate that MVA techniques may be used for the robust detection of gravitational-wave bursts with a priori unknown waveform.