Gravitational wave burst is a catch-all category for signals whose durations are shorter than the observation period. We apply a method new to gravitational wave data analysis --- Bayesian non-parameterics --- to the problem of gravitational wave det
ection, with an emphasis on pulsar timing array observations. In Bayesian non-parametrics, constraints are set on the function space that may be reasonably thought to characterize the range of gravitational-wave signals. This differs from the approaches currently employed or proposed, which focus on introducing parametric signal models or looking for excess power as evidence of the presence of a gravitational wave signal. Our Bayesian nonparametrics analysis method addresses two issues: (1) investigate if a gravitational wave burst is present in the data; (2) infer the sky location of the source and the duration of the burst. Compared with the popular method proposed by Finn & Lommen, our method improves in two aspects: (1) we can estimate the burst duration by adding the prior that the gravitational wave signals are smooth, while Finn & Lommen ignored this important point; (2) we perform a full Bayesian analysis by marginalizing over all possible parameters and provide robust inference on the presence of gravitational waves, while Finn & Lommen chose to optimize over parameters, which would increase false alarm risk and also underestimate the parameter uncertainties.
