[ColmTalbot]

This replaces the interpolated lookup table used for distance marginalization in GravitationalWaveTransient with distmarg. distmarg evaluates the distance-marginalized likelihood directly from the matched-filter and optimal SNRs at the reference distance. The ROQ, relative-binning and multiband likelihoods use the same distance_marginalized_likelihood, so they pick up the change without modification.

What changes

• Likelihood: distance_marginalized_likelihood calls distmarg instead of the 400×800 lookup table and its spline.
  • The result follows the array backend of the inputs, so the JAX path still works with SCIPY_ARRAY_API=1.
  • Time-bin and calibration axes are handled.
• Prior: the distance prior is tabulated from the prior's own cosmology for UniformSourceFrame and UniformComovingVolume, rather than from the interpolated prob. Nothing else is precomputed; previously, building the table took ~40 s per phase setting.
• Distance reconstruction samples the exact conditional posterior. If the template is identically zero, it samples the distance prior.
• Deprecation: distance_marginalization_lookup_table is deprecated, with a DeprecationWarning, for removal in v3.1.0. A small record of the tabulated prior is still written to cached_lookup_table_filename, because bilby_pipe reads that file during data generation.
• Removals: BoundedRectBivariateSpline and its test are removed. The lookup table was its only user.
• Dependency: distmarg[rust] is added to gw_requirements.txt. The rust extra installs the compiled backend: a single-point evaluation then takes ~2 µs, against ~7 µs for the table.
• This removes the need to carry around a lookup table, so will reduce startup times for many examples and production uses by a few minutes
• bilby_pipe can stop carrying around precomputed tables, reducing the wheel size dramatically.
• It is also much more accurate (especially at large SNR) than the grid method (see below).

Accuracy

distmarg is accurate to ≲1e-4 in ln L across the full stress grid:

• d* from d_min/5 to 5 d_max;
• rho_mf from 1e-3 to 1e4, including negative values;
• with and without phase marginalization.

This is checked against a dense reference quadrature and an independent mpmath oracle.

In a version of the fast-tutorial example:

• distmarg matched a brute-force distance integral to ≤1.2e-6;
• the lookup table was off by up to 1e-2;
• sampling reproduced the evidence and distance posterior of the lookup-table likelihood within sampling noise.

Testing

• The test suite runs at least a few minutes faster because we don't have to construct any of the lookup tables.
• Nothing has been tested on a GPU yet.

🤖 Generated with assistance from Claude Code

[GregoryAshton]

I'm a fan of this approach. Is it possible to add a second maintainer (no need for code credit) to the repo in case you choose to leave us and do something else, and we need to bug-fix distmarg?