rxmc.config.CalibrationConfig

class rxmc.config.CalibrationConfig(evidence: Evidence, model_config: ParameterConfig, likelihood_configs: list | None = None, likelihood_scaling: float | None = None)

Bases: object

End-to-end configuration for Bayesian calibration.

Combines an Evidence object with ParameterConfig instances for the physical-model parameters and any parametric likelihood parameters. The result is a flat parameterization suitable for black-box samplers (emcee, Dynesty, black-box-bayes, etc.).

Parameters:

• evidence (Evidence) – Aggregated experimental constraints.

• model_config (ParameterConfig) – Prior and proposal for the physical-model parameters.

• likelihood_configs (list of ParameterConfig, optional) – One ParameterConfig per parametric constraint in evidence.parametric_constraints. Omit or pass None when there are no parametric likelihood models.

• likelihood_scaling (float, optional) – Multiplicative scale applied to the total log-likelihood before adding the log-prior. Useful for tempering or importance re-weighting. Defaults to 1.0.

Raises:

• ValueError – If evidence contains no constraints.

• ValueError – If the model parameters in model_config do not match those in the evidence constraints.

• ValueError – If the number or parameter lists of likelihood_configs do not match evidence.parametric_constraints.

ndim

Total number of free parameters (model + all likelihood sectors).

Type:

int

dimensions

Array of sector sizes [model_ndim, lc0_ndim, lc1_ndim, ...].

Type:

ndarray

indices

Cumulative split indices derived from dimensions; used by split_parameters().

Type:

ndarray

__init__(evidence: Evidence, model_config: ParameterConfig, likelihood_configs: list | None = None, likelihood_scaling: float | None = None)

Methods

__init__(evidence, model_config[, ...])
conditional_posterior(x_lm, lm_index, ym)	Log posterior for one likelihood sector, conditioned on observed data.
log_likelihood(x)	Evaluate the scaled log likelihood at a flat parameter vector.
log_posterior(x)	Evaluate the log posterior at a flat parameter vector.
log_posterior_batch(thetas)	Evaluate the log posterior for a batch of parameter vectors.
log_prior(x)	Evaluate the joint log prior at a flat parameter vector.
predict(xmodel)	Generate model predictions for every constraint.
prior_transform(u)	Map unit-cube coordinates to physical parameters (Dynesty interface).
split_parameters(x)	Split a flat parameter vector into model and likelihood sub-vectors.
starting_location(nwalkers)	Generate initial walker positions across all parameter sectors.

Attributes

parameter_configs	model first, then likelihoods.
parameter_names	Flat parameter names in sampler order.
parameters	All Parameter objects in flat sampler order.
prior	Prior distribution objects in parameter-sector order.

property parameter_configs: list

model first, then likelihoods.

Type:

All parameter sectors in flat sampler order

property parameters: List[Parameter]

All Parameter objects in flat sampler order.

The order matches the flat parameter vector consumed by log_posterior(), log_likelihood(), and starting_location().

property parameter_names: List[str]

Flat parameter names in sampler order.

Convenience accessor equivalent to [p.name for p in self.parameters]. Provided for compatibility with external drivers such as black-box-bayes that export inference results keyed by name.

property prior: list

Prior distribution objects in parameter-sector order.

Returns one entry per sector: the model prior first, followed by one entry per likelihood sector. Each entry is whatever was passed as prior to the corresponding ParameterConfig — a list of univariate distributions, a multivariate distribution, or a custom prior object.

split_parameters(x) → tuple

Split a flat parameter vector into model and likelihood sub-vectors.

Parameters:

x (ndarray, shape (ndim,)) – Flat parameter vector in sampler order.

Returns:

• xmodel (ndarray) – Model parameter sub-vector of shape (model_config.ndim,).

• xlikelihoods (list of ndarray) – One sub-vector per likelihood sector.

log_prior(x) → float

Evaluate the joint log prior at a flat parameter vector.

Parameters:

x (ndarray, shape (ndim,)) – Flat parameter vector in sampler order.

Returns:

float – Sum of log prior densities across all sectors.

log_likelihood(x) → float

Evaluate the scaled log likelihood at a flat parameter vector.

Parameters:

x (ndarray, shape (ndim,)) – Flat parameter vector in sampler order.

Returns:

float – likelihood_scaling * evidence.log_likelihood(xmodel, xlikelihoods).

log_posterior(x) → float

Evaluate the log posterior at a flat parameter vector.

Returns -inf immediately if either the prior or likelihood is non-finite, avoiding unnecessary model evaluations.

Parameters:

x (ndarray, shape (ndim,)) – Flat parameter vector in sampler order.

Returns:

float – log_prior(x) + log_likelihood(x), or -inf if either term is non-finite.

log_posterior_batch(thetas: ndarray) → ndarray

Evaluate the log posterior for a batch of parameter vectors.

Convenience wrapper for external orchestration layers (e.g. black-box-bayes). The current implementation is serial but preserves the advertised interface for future parallelisation.

Parameters:

thetas (ndarray, shape (…, ndim)) – Batch of parameter vectors.

Returns:

ndarray, shape (n,) – Log posterior value for each row of thetas.

Raises:

ValueError – If the trailing dimension of thetas is not ndim.

starting_location(nwalkers: int) → ndarray

Generate initial walker positions across all parameter sectors.

Parameters:

nwalkers (int) – Number of walkers.

Returns:

ndarray, shape (nwalkers, ndim) – Concatenated initial positions from each sector’s proposal distribution.

prior_transform(u) → ndarray

Map unit-cube coordinates to physical parameters (Dynesty interface).

Delegates to ParameterConfig.prior_transform() for each sector, which in turn either calls ppf on each element of a list prior or calls prior_transform directly on a joint prior object.

Parameters:

u (array-like, shape (ndim,)) – Unit-cube coordinates, each in [0, 1).

Returns:

theta (ndarray, shape (ndim,)) – Physical parameter vector.

Raises:

• NotImplementedError – If any sector’s prior does not support a transform (see ParameterConfig.prior_transform()).

• ValueError – If u does not have length ndim.

predict(xmodel) → list

Generate model predictions for every constraint.

Parameters:

xmodel (ndarray, shape (model_config.ndim,)) – Physical model parameter vector.

Returns:

list of ndarray – Predicted observable for each constraint, in the order evidence.constraints + evidence.parametric_constraints.

conditional_posterior(x_lm, lm_index: int, ym) → float

Log posterior for one likelihood sector, conditioned on observed data.

Evaluates marginal_log_likelihood(ym, *x_lm) + prior_logpdf(x_lm) for the likelihood sector at lm_index. Useful for Gibbs-style updates where the likelihood parameters are sampled separately from the physical model parameters.

Parameters:

• x_lm (ndarray, shape (likelihood_configs[lm_index].ndim,)) – Parameter vector for the target likelihood sector.

• lm_index (int) – Index into likelihood_configs (and evidence.parametric_constraints).

• ym (ndarray) – Predicted observable used as the conditioning data.

Returns:

float – Log posterior for this likelihood sector.