Prediction

Merged-grid prediction for state-space GPs.

See plans/2026-04-21-state-space-gps-design.md §Stage 4.

predict_smoothed

predict_smoothed(
    posterior,
    train_data,
    test_inputs,
    *,
    observation_mask=None,
)

Smoothed-latent prediction for state-space GPs.

Returns a GaussianDistribution over the M test points whose mean is the RTS-smoothed posterior mean and whose scale is a lx.DiagonalLinearOperator carrying the smoothed marginal variance plus prior.jitter. Test inputs are returned in caller order regardless of sorting.

See plans/2026-04-21-state-space-gps-design.md §Stage 4.

predict_filtered

predict_filtered(
    posterior,
    train_data,
    test_inputs,
    *,
    observation_mask=None,
)

Filtered (causal) prediction for state-space GPs.

Each test point conditions only on training observations at timestamps less than or equal to the test timestamp (with train < test tie-break). Marginals are extracted from the forward filter trajectory rather than the smoother.

See plans/2026-04-21-state-space-gps-design.md §Stage 4.