Nested Sampling¶
Nested-sampler is the posterior and evidence calculation path. It is the
default method for production model comparison, uncertainty estimates and
prior-volume dependent statements.
Use Inference Methods for the shared likelihood, prior and constraint definitions used by all inference methods.
When To Use It¶
Use nested sampling when the analysis requires:
Posterior samples: credible intervals, correlations and multimodal structure.
Evidence: model comparison through the nested-sampler log-evidence.
Prior-volume effects: checks where posterior support depends on prior widths or non-rectangular constraints.
Nested sampling is slower than the point-estimate methods. The quick examples therefore use small settings that are suitable for smoke tests, not final science runs.
Minimal Configuration¶
Nested sampling is selected in [Inference]:
[Inference]
method = Nested-sampler
likelihood = gaussian
sampler = cpnest
nlive = 256
maxmcmc = 256
seed = 1234
The likelihood is evaluated on the same selected NR time samples described in Inference Methods.
Sampler Backends¶
sampler can be:
Sampler |
Behaviour |
|---|---|
|
Default sampler. Output is written under |
|
Alternative nested sampler. Uses |
Settings¶
The most important nested-sampler controls are:
nlive: number of live points. Increase this for production runs, broad priors or visibly structured posteriors.
maxmcmc: maximum internal MCMC length for
cpnestproposals.seed: random seed for reproducible diagnostic runs.
nnest: number of nested-sampler processes used by
raynest.nensemble: number of ensemble processes used by
raynest.
The right nlive and maxmcmc values depend on model dimension, prior
volume, start time, error prescription and waveform degeneracies. A short test
can confirm the pipeline, but it does not establish sampler convergence.
Outputs¶
Nested-sampler outputs include:
Posterior samples: written under the sampler output directory and used by post-processing.
Evidence: saved by the sampler backend.
Prior-railing diagnostics: generated when posterior support accumulates close to prior boundaries.
Waveform reconstructions: produced during post-processing when waveform plotting is enabled.
Mismatch/SNR diagnostics: produced when the corresponding options are enabled.
If posterior railing is detected, inspect the printed warning and
Algorithm/Parameters_prior_railing*.txt before interpreting the result.
Validation Checklist¶
Before using a nested-sampler result scientifically:
Prior boundaries: check prior-railing diagnostics and corner plots.
Sampler settings: rerun with larger
nliveormaxmcmcwhen the posterior is broad, multimodal or near a boundary.Fit window: vary
t-startort-endto test ringdown stability.Error model: compare at least one alternative NR uncertainty prescription when available.
Residuals: inspect waveform residuals, not only posterior summaries.
Evidence: compare evidences only between runs with compatible data, likelihoods, priors and sampler settings.