Skip to content

diagnostics

Post-run diagnostics and resampling for the BlackJAX SMC sampler modes.

Pure functions over a finished SMC state -- no coupling to the sampling loop. Kept separate so each mode class's get_samples/_get_diagnostics stay thin composition and the actual math has a name and a docstring of its own.

Functions:

Name Description
kish_ess_history

Per-iteration Kish effective sample size for tempered modes (at/ft).

kish_log_z_error

Delta-method log-evidence standard error for tempered modes (at/ft).

persistent_ess_history

Per-iteration persistent-sampling ESS (modes ap/fp).

persistent_log_z_error

Delta-method log-evidence standard error for persistent sampling (modes ap/fp).

resample_persistent_particles

Draw the posterior sample from a finished persistent-sampling (ap/fp) state.

kish_ess_history(is_weights_history: np.ndarray) -> np.ndarray ¤

Per-iteration Kish effective sample size for tempered modes (at/ft).

Importance weights are already normalized per iteration, so Kish ESS is 1 / sum(w**2).

Parameters:

Name Type Description Default
is_weights_history ndarray

Per-iteration normalized IS weights, shape (n_iterations, n_particles).

required

Returns:

Type Description
ndarray

ESS at each iteration, shape (n_iterations,).

kish_log_z_error(is_weights_history: np.ndarray) -> float ¤

Delta-method log-evidence standard error for tempered modes (at/ft).

Var(log Z_k) = sum(w**2) - 1/N per step for normalized weights, summed across steps.

Parameters:

Name Type Description Default
is_weights_history ndarray

Per-iteration normalized IS weights, shape (n_iterations, n_particles).

required

Returns:

Type Description
float

Standard deviation of the final cumulative log_Z.

persistent_ess_history(ps, n_iterations: int) -> np.ndarray ¤

Per-iteration persistent-sampling ESS (modes ap/fp).

Parameters:

Name Type Description Default
ps

The algorithm's persistent-sampling state, as in :func:resample_persistent_particles.

required
n_iterations int

Number of completed iterations.

required

Returns:

Type Description
ndarray

ESS at each iteration 1..n_iterations, shape (n_iterations,).

persistent_log_z_error(ps, n_iterations: int) -> float ¤

Delta-method log-evidence standard error for persistent sampling (modes ap/fp).

At step k, importance weights are exp(delta_beta * logL) over all k batches of accumulated particles; Var(log Z_k) = Var(w) / (N_eff * E[w]^2), summed across steps.

Parameters:

Name Type Description Default
ps

The algorithm's persistent-sampling state, as in :func:resample_persistent_particles.

required
n_iterations int

Number of completed iterations.

required

Returns:

Type Description
float

Standard deviation of the final cumulative log_Z.

resample_persistent_particles(ps, n_dims: int) -> dict[str, np.ndarray] ¤

Draw the posterior sample from a finished persistent-sampling (ap/fp) state.

Resamples with replacement from all-temperature particles, weighted by the persistent-sampling weight formula, to a target count approximately equal to the effective sample size 1 / max(weights).

Parameters:

Name Type Description Default
ps

The algorithm's persistent-sampling state (PersistentSamplingState), i.e. state.sampler_state (mode ap) or state (mode fp).

required
n_dims int

Dimension of the sampling space.

required

Returns:

Type Description
dict[str, ndarray]

Dict with keys "samples" (shape (n, n_dims)) and

dict[str, ndarray]

"log_likelihood" (shape (n,)).