Abstract
Pulsar timing arrays search for stochastic processes such as gravitational waves by comparing pulse time of arrival data for millisecond pulsars to expectations from a background with a given power spectral density (PSD). To make the analysis computationally tractable, the Bayesian likelihood is usually computed using an approximation in which the signal is taken to be a sum of Fourier modes appropriate to the total time of observation, even though the true signal is not periodic. We study the difference between likelihoods computed with this Fourier approximation method for power law spectra and those computed exactly (or using more-closely spaced frequencies as a proxy for the exact result) in the NANOGrav 15-year dataset. We find that the true marginal likelihoods for power-law PSDs are on average about half as large as the likelihoods computed using the Fourier approximation. This could lead to an error of a factor of two in model comparison. However, in the important comparison of uncorrelated vs. Hellings-Downs correlated models, a very similar correction appears in both, so the model comparison is essentially unaffected. We also compare parameter estimation results for power law PSDs, finding little difference between the methods. We briefly discuss spectra with sharper features, for which the approximation could be much worse.