rxmc.param_sampling.BatchedAdaptiveMetropolisSampler#

class rxmc.param_sampling.BatchedAdaptiveMetropolisSampler(params: list[Parameter], prior, starting_location: ndarray, initial_proposal_cov: ndarray, epsilon_fraction: float = 1e-06)[source]#

Bases: Sampler

Metropolis sampler that updates the proposal covariance after each batch.

After every completed batch — burn-in batches included — the proposal covariance is replaced by the empirical covariance of that batch, scaled by 2.38² / ndim. Burn-in only affects whether the samples are recorded, so the proposal adapts during burn-in as is standard.

Parameters:

  • params (list of Parameter) – Parameters to sample.

  • prior (object) – Prior distribution with a callable logpdf(x) method.

  • starting_location (np.ndarray, shape (ndim,)) – Initial parameter vector.

  • initial_proposal_cov (np.ndarray, shape (ndim, ndim)) – Initial proposal covariance matrix.

  • epsilon_fraction (float, optional) – Small regularisation fraction added to the empirical covariance diagonal. Defaults to 1e-6.

__init__(params: list[Parameter], prior, starting_location: ndarray, initial_proposal_cov: ndarray, epsilon_fraction: float = 1e-06)[source]#

Methods

__init__(params, prior, starting_location, ...)

batch_acceptance_fractions()

Acceptance fraction for each completed batch.

most_recent_batch_acceptance_fraction()

Acceptance fraction of the most recent batch.

overall_acceptance_fraction()

Overall acceptance fraction across all completed batches.

record_batch(n_steps, n_accepted, chain, ...)

Append a completed batch to the running chain.

sample(n_steps, starting_location, rng, ...)

Run the sampler for one batch, then adapt the proposal.
sample(n_steps: int, starting_location: ndarray, rng: Generator, log_posterior: Callable[[ndarray], float], burn: bool = False)[source]#

Run the sampler for one batch, then adapt the proposal.

Overrides Sampler.sample() to replace the proposal covariance with the current batch’s empirical covariance after every batch, burn-in or not. proposal and proposal_cov always reflect the proposal the next batch will use.

Parameters:

  • n_steps (int) – Number of steps to run.

  • starting_location (np.ndarray, shape (ndim,)) – Starting parameter vector.

  • rng (np.random.Generator) – Random number generator.

  • log_posterior (callable) – Function f(x) -> float returning the log posterior.

  • burn (bool, optional) – If True, discard the samples (they are not recorded); the covariance update still happens. Defaults to False.