Making Bayesian gravitational-wave inference faster

parameter estimation
Bayesian inference
computational methods
Why much of the computational cost of compact-binary inference can be removed without changing the statistical problem.
Published

September 5, 2026

Bayesian parameter estimation for gravitational-wave signals repeatedly evaluates a likelihood over a high-dimensional parameter space. For compact binaries, the expensive part is usually not the statistical formalism itself but the repeated evaluation of waveform-dependent frequency-domain inner products.

A large fraction of this work is redundant.

Nearby waveforms are strongly correlated, and the quantities entering the likelihood vary much more smoothly across frequency than the raw waveform phase might suggest. Fast inference methods exploit this structure rather than evaluating every likelihood on the full native frequency grid.

Several complementary strategies are possible. Relative binning represents the ratio between nearby waveforms using a much coarser frequency grid. Meshfree methods interpolate quantities entering the likelihood across intrinsic parameter space. Adaptive grids place frequency samples according to the local variation of the waveform rather than at every Fourier frequency.

These methods are particularly important for next-generation detectors. Lowering the starting frequency from tens of hertz to a few hertz greatly increases the signal duration and the number of Fourier samples. A method that is inexpensive for a current-detector binary-neutron-star signal can otherwise become prohibitively costly.

The aim is therefore not to approximate the posterior after the fact, but to reorganize the likelihood calculation so that the expensive information is computed once and reused efficiently throughout the inference.

Papers from our work on fast inference