Working with results

The SMC runner returns an SMCResult. It contains the state after each completed iteration in result.states, the starting population in result.initial_state, and step outputs in result.outputs. result.final_state is the final population; result[i] accesses an earlier population. See Particle state and results for the particle-state interface.

Posterior samples and summaries

For standard and persistent Bayesian SMC runs that have reached the posterior target, use the final population to compute weighted summaries:

mean = result.weighted_mean()
covariance = result.weighted_covariance()
samples, weights = result.posterior_samples()

The default method="weights" returns the final particles and their normalized weights. To draw an equally weighted sample with replacement:

samples, weights = result.posterior_samples(
    n_samples=1_000,
    method="multinomial",
    rng=42,
)

Respect the weights when analysing the particles directly. Intermediate tempered populations do not yet represent the full posterior, and resampling does not make particles independent.

Run diagnostics

Inspect the run summary and histories to understand how the population evolved:

summary = result.summary()
ess = result.effective_sample_sizes()
beta = result.metadata_history("beta")
log_evidence = result.log_evidence()
log_evidence_history = result.log_evidence_history()

The beta history is relevant to tempered runs. ESS describes weight concentration; it does not measure the number of independent posterior draws. Step-specific diagnostics are available in result.outputs. See Evidence and uncertainty for interpreting evidence uncertainty.

Nested-sampling results

For nested-sampling SMC, use the shell-weighted result interface:

nested = result.as_nested_sampling_result()
complete = nested.complete
log_evidence = nested.log_evidence
# Once complete, with store_shell_particles=True:
samples, weights = nested.posterior_samples()

The final state represents the constrained prior. Bayesian evidence and posterior weights come from the accumulated shells, so use this interface for posterior analysis. A run suspended by num_steps must be resumed to completion before extracting posterior samples.