Taming outliers in pulsar-timing datasets with hierarchical likelihoods and Hamiltonian sampling

TL;DR

This paper presents a robust Bayesian method using hierarchical likelihoods and Hamiltonian sampling to identify and mitigate outliers in pulsar-timing datasets, improving the reliability of gravitational wave searches.

Contribution

The authors introduce a fully consistent hierarchical Bayesian approach with Hamiltonian sampling to detect and handle outliers in pulsar-timing data, enhancing data analysis robustness.

Findings

Method effectively identifies outliers probabilistically.

Sampling approach is computationally efficient.

Improves confidence in gravitational wave detection results.

Abstract

Pulsar-timing datasets have been analyzed with great success using probabilistic treatments based on Gaussian distributions, with applications ranging from studies of neutron-star structure to tests of general relativity and searches for nanosecond gravitational waves. As for other applications of Gaussian distributions, outliers in timing measurements pose a significant challenge to statistical inference, since they can bias the estimation of timing and noise parameters, and affect reported parameter uncertainties. We describe and demonstrate a practical end-to-end approach to perform Bayesian inference of timing and noise parameters robustly in the presence of outliers, and to identify these probabilistically. The method is fully consistent (i.e., outlier-ness probabilities vary in tune with the posterior distributions of the timing and noise parameters), and it relies on the efficient sampling of the hierarchical form of the pulsar-timing likelihood. Such sampling has recently become possible with a "no-U-turn" Hamiltonian sampler coupled to a highly customized reparametrization of the likelihood; this code is described elsewhere, but it is already available online. We recommend our method as a standard step in the preparation of pulsar-timing-array datasets: even if statistical inference is not affected, follow-up studies of outlier candidates can reveal unseen problems in radio observations and timing measurements; furthermore, confidence in the results of gravitational-wave searches will only benefit from stringent statistical evidence that datasets are clean and outlier-free.