Researchers introduce a fast method to compare models of the nanohertz gravitational-wave background (GWB) seen by pulsar timing arrays (PTAs). The new approach combines normalizing flows, nested sampling, and a likelihood that is marginalized over all non-GWB parameters. The authors say this lets them compute Bayes factors — the usual Bayesian metric for comparing models — in minutes on consumer hardware instead of the much longer times required by existing PTA methods.

What the team did is practical and technical. They first reduce the problem size by marginalizing the PTA likelihood over all pulsar-specific and other non-GWB parameters. They then train a normalizing flow — a machine-learning model that learns how to transform simple random numbers into samples from a complicated distribution — on the resulting free-spectrum posterior. The flows are trained with a new Python package called coppuccino, which specifically learns the copula, meaning the dependence structure between frequencies after transforming each to uniform values. The trained flow acts as a cheap surrogate for the expensive likelihood inside a nested sampling routine. Nested sampling is an algorithm designed to compute the Bayesian evidence directly; evidence ratios give the Bayes factors used to compare models.

They validated the method on both analytic examples and simulated PTA data. According to the paper, the method recovered analytic evidences within the reported error bars. When the Bayes factors were treated as binary classifiers for which model produced the simulated data, they obtained an area-under-the-curve (AUC) score of 0.996, indicating very good discrimination in their tests. The authors emphasize that their workflow can compare arbitrary spectral shapes for the GWB and can be run quickly on modest hardware.