Gravitational wave detection without boot straps: A Bayesian approach

Abstract

In order to separate astrophysical gravitational-wave signals from instrumental noise, which often contains transient non-Gaussian artifacts, astronomers have traditionally relied on bootstrap methods such as time slides. Bootstrap methods sample with replacement, comparing single-observatory data to construct a background distribution, which is used to assign a false-alarm probability to candidate signals. While bootstrap methods have played an important role establishing the first gravitational-wave detections, there are limitations. First, as the number of detections increases, it makes increasingly less sense to treat single-observatory data as bootstrap-estimated noise, when we know that the data are filled with astrophysical signals, some resolved, some unresolved. Second, it has been known for a decade that background estimation from time-slides eventually breaks down due to saturation effects, yielding incorrect estimates of significance. Third, the false alarm probability cannot be used to weight candidate significance, for example when performing population inference on a set of candidates. Given recent debate about marginally resolved gravitational-wave detection claims, the question of significance has practical consequences. We propose a Bayesian framework for calculating the odds that a signal is of astrophysical origin versus instrumental noise without bootstrap noise estimation. We show how the astrophysical odds can safely accommodate glitches. We argue that it is statistically optimal. We demonstrate the method with simulated noise and provide examples to build intuition about this new approach to significance.