bilby-laplace

A Bilby sampler plugin that estimates posteriors via the Laplace approximation — a Gaussian fitted at the maximum a posteriori (MAP) point using the Hessian of the log-posterior — followed by optional resampling to correct for non-Gaussianity.

The method is fast, scales well to moderate dimensions, and produces asymptotically exact posterior samples when the true posterior is close to Gaussian. It is useful as a cheap cross-check against nested sampling results.

Documentation: full guides, the configuration/API reference, and background on the method are at https://gregoryashton.github.io/bilby-laplace/.

NOTE: This is currently in development and derived from bilby PR #933 (Gregory Ashton).

Installation

pip install bilby-laplace

Or from source:

git clone https://github.com/GregoryAshton/bilby-laplace
cd bilby-laplace
pip install -e .

Bilby discovers the sampler automatically via its plugin entry-point system — no further configuration is needed.

Quick start

import bilby

result = bilby.run_sampler(
    likelihood=likelihood,
    priors=priors,
    sampler="laplace",
    outdir="outdir",
    label="my_run",
)

result.plot_corner()
print(result.posterior)

Features

• Several resampling strategies to correct for non-Gaussianity: rejection, importance, inprior, smc (via aspire), or none.
• Two covariance routes: the numerical Hessian (fisher_method="hessian", any likelihood) or the genuine waveform Fisher matrix (fisher_method="waveform", for gravitational-wave likelihoods).
• Supply a precomputed covariance via sampling_cov (e.g. from gwfast / GWFish).
• Multimodal posteriors via SMC with n_modes > 1.
• Laplace log-evidence always available; rejection and SMC add independent estimates.

See the documentation for the full list of options and guidance on choosing between them.

Examples

Runnable examples are in examples/ (gaussian, rosenbrock, BBH, BNS). Each has a Makefile:

cd examples/gaussian
make laplace
make compare

Documentation development

pip install -e ".[docs]"
mkdocs serve