jesterTOV.inference.samplers.eos_reweighting.EOSReweightingSampler#
- class EOSReweightingSampler(
- likelihood,
- prior=None,
- sample_transforms=None,
- likelihood_transforms=None,
- config=None,
- seed=0,
Bases:
JesterSamplerSampler that reweights a discrete EOS set by jester likelihoods.
Unlike parametric samplers implemented in
jester, this class does not sample a prior. Instead, it receives a fixed set of tabulated EOS curves and evaluates the combined likelihood on each one. The result is a discrete posterior over the EOS set together with the Bayesian evidence \(\log Z\).- Parameters:
likelihood (LikelihoodBase) – Combined likelihood (GW, NICER, radio, …) created by
create_combined_likelihood().prior (object, optional) – Accepted for API compatibility with
JesterSampler; however, ignored for the reweighting.sample_transforms (list, optional) – Accepted for API compatibility; however, ignored for the reweighting.
likelihood_transforms (list, optional) – Accepted for API compatibility; however, ignored for the reweighting.
config (EOSReweightingConfig) – Sampler configuration including EOS file paths and grid settings.
seed (int, optional) – Accepted for API compatibility; however, ignored for the reweighting.
- __init__(
- likelihood,
- prior=None,
- sample_transforms=None,
- likelihood_transforms=None,
- config=None,
- seed=0,
Methods
__init__(likelihood[, prior, ...])add_name(x)Turn an array into a dictionary.
compute_evidence(log_likelihoods)Compute Bayesian evidence and effective sample size from log-likelihoods.
evaluate_batch(f, all_masses, all_lambdas, ...)Evaluate f on all N EOS curves using
jax.lax.map().load_and_grid(paths, n_grid, m_min, m_max)Load EOS files and resample all curves onto two common mass grids.
Build a single-EOS log-likelihood callable for use with
jax.lax.map().posterior(params, data)Evaluate posterior log probability from flat array.
posterior_from_dict(named_params, data)Evaluate posterior log probability from parameter dict.
print_summary([transform])Print summary of sampling run.
sample(key)Evaluate all EOS curves and return evidence + posterior samples.
Attributes
Absolute cap (\(M_\odot\)) applied to the common mass grid upper bound in
load_and_grid()whenm_maxis not given explicitly.likelihoodpriorsample_transformslikelihood_transformsparameter_namessampler- DEFAULT_M_MAX_CAP: float = 4.0#
Absolute cap (\(M_\odot\)) applied to the common mass grid upper bound in
load_and_grid()whenm_maxis not given explicitly.
- compute_evidence(log_likelihoods)[source]#
Compute Bayesian evidence and effective sample size from log-likelihoods.
This is inspired by the lwp code: https://git.ligo.org/reed.essick/lwp/
Same computation as
lwp.utils.utils.estimate_evidence(with its default uniformprior), just carried out in log space for numerical stability. Writing \(w^{(a)} = \exp(\text{log\_likelihoods}_a)\) for the raw per-EOS likelihood and \(p^{(a)} = w^{(a)} / \sum_b w^{(b)}\) for the normalised posterior weight,lwpcomputes\[Z = \frac{1}{N}\sum_{a=1}^{N} w^{(a)} \,, \qquad Z_2 = \frac{1}{N}\sum_{a=1}^{N} \left(w^{(a)}\right)^2 \,, \qquad \delta Z = \sqrt{\frac{Z_2 - Z^2}{N}}\,,\]where \(\delta Z\) is the standard error of the mean over the
Nweights, and reports the relative uncertainty \(\delta Z / Z\) after propagating it to log space via the delta method (\(\sigma_{\log Z} \approx \delta Z / Z\)). Using \(N_\mathrm{eff} = Z^2/Z_2\cdot N = 1/\sum_a (p^{(a)})^2\) (the Kish effective sample size,lwp.stats.stats.nkde) this simplifies to\[\log Z = \mathrm{logsumexp}(\text{log\_likelihoods}) - \log N \,, \qquad \sigma_{\log Z} = \sqrt{\frac{1}{N_\mathrm{eff}} - \frac{1}{N}}\,.\]
- evaluate_batch(f, all_masses, all_lambdas, all_radii)[source]#
Evaluate f on all N EOS curves using
jax.lax.map().The curves are processed in Python-level batches of
config.batch_sizeso that progress and ETA can be logged. Each batch is evaluated with a jittedjax.lax.map(). This double batching is somewhat awkward, but it is needed for the logging and the runtime is about the same as a singlelax.mapcall.- Parameters:
f (
Callable[[tuple[Array,Array,Array]],Array]) – Single-EOS evaluator returned bymake_eos_fn().all_masses (
Float[Array, 'N L']) – Stacked JAX arrays of shape[N, L].all_lambdas (
Float[Array, 'N L']) – Stacked JAX arrays of shape[N, L].all_radii (
Float[Array, 'N L']) – Stacked JAX arrays of shape[N, L].
- Return type:
Float[Array, 'N']- Returns:
Float[Array, “ N”] – Log-likelihoods per EOS.
- load_and_grid(paths, n_grid, m_min, m_max)[source]#
Load EOS files and resample all curves onto two common mass grids.
Two grids are built from the same underlying curves: a likelihood grid (bounded by
m_min/m_max, as configured) used to evaluate the likelihoods, and a full grid that always spans the EOS set’s own natural range — from \(\min(M)\) across all curves up to \(\max(M_\mathrm{TOV})\) — regardless ofm_min/m_max. Those two bounds only restrict what jester’s likelihoods see; the full-range curves are what gets resampled into the posterior (sample()) for downstream plotting/output, so posterior M-R/M-Lambda curves and the saved result are never truncated by the likelihood grid’s bounds.- Parameters:
paths (
list[str]) – Paths to NPZ files. Each file must contain keysmasses,lambdas, andradii(1-D or 2-D arrays). For a file with N curves the arrays must be shaped[N, n_points], with matching shapes across the three keys — ragged/heterogeneous curves within a single file are not supported and raiseValueError. Curves may differ in length between files, since each curve is resampled onto the common mass grid independently.n_grid (
int) – Number of mass grid points (used for both grids).m_min (
float) – Lower bound of the likelihood grid in \(M_\odot\).m_max (
float|None) – Upper bound of the likelihood grid in \(M_\odot\). IfNone, then use \(\max(M_\mathrm{TOV})\) across all curves inpaths, capped atDEFAULT_M_MAX_CAP\(M_\odot\) (a warning is logged if the cap is applied, since likelihoods will not be evaluated above it).
- Return type:
tuple[tuple[Float[Array, 'N L'],Float[Array, 'N L'],Float[Array, 'N L']],tuple[Float[Array, 'N L'],Float[Array, 'N L'],Float[Array, 'N L']]]- Returns:
likelihood_grid (tuple of Float[Array, “N L”]) –
(masses, lambdas, radii)on the[m_min, m_max]-bounded grid used for likelihood evaluation.full_grid (tuple of Float[Array, “N L”]) –
(masses, lambdas, radii)on the EOS set’s natural[min(M), max(M_TOV)]grid, used for posterior resampling — unaffected bym_min/m_max.
- make_eos_fn()[source]#
Build a single-EOS log-likelihood callable for use with
jax.lax.map().
- sample(key)[source]#
Evaluate all EOS curves and return evidence + posterior samples.
After computing the per-EOS evidence and posterior weights, this also draws
n_resample(default: \(N_\mathrm{eff}\)) equal-weight posterior samples viaresample_eos_posterior(), so that downstream postprocessing (e.g.generate_eos_reweighting_plots()) can plot mass-radius/mass-Lambda curves directly without having to deal with the unequal per-EOS weights itself. The resampled curves are drawn from the full, uncapped mass grid (seeload_and_grid()), so plots always show each EOS out to its true \(M_\mathrm{TOV}\) even if the likelihood grid was capped.- Parameters:
key (
Union[Key[Array, ''],UInt32[Array, '2'],UInt32[Array, '4']]) – JAX random key. Only its bits are used (to seed the NumPy RNG for resampling); the likelihood evaluation itself is deterministic.- Return type:
- Returns:
SamplerOutput –
samples["eos_index"]: integer index per input EOS (length N)samples["log_likelihood"]: log-likelihood per input EOS (length N)samples["posterior_weight"]: normalised posterior weight per input EOS (length N)samples["masses_EOS"],samples["Lambdas_EOS"],samples["radii_EOS"]: resampled equal-weight posterior curves (length N_resampled)samples["resampled_eos_index"]: index into the input EOS set for each resampled drawsamples["resampled_log_likelihood"]: log-likelihood of each resampled drawlog_prob: same assamples["log_likelihood"](length N)metadata["evidence"]: evidence dict (log_Z, log_Z_std, N_eff, …)metadata["N_eos"]: total number of input EOS curvesmetadata["N_resampled"]: number of resampled posterior draws