State inference

State inference#

Inference turns an observation into a posterior over state. The discrete scheme is mean-field fixed point iteration. The continuous scheme conditions a Gaussian or a Gaussian mixture on a linear observation exactly (on readings of independent noise, a DiagonalNoise, in state space at a cost linear in the number of readings), and rolls mixtures forward through a transition with a reduction schedule so that the number of components stays fixed. A System calls these for you through its infer_states method; the functions are exposed for direct use.

Discrete#

fpi

Mean-field posteriors over state factors given one observation per modality.

InferenceDetail

Per-modality internals of one sweep; None where inapplicable.

Linear-Gaussian and mixture#

condition

Bayesian update of g on y = H x + c + eps, eps ~ N(0, R).

DiagonalNoise

An observation noise of independent readings, N(0, diag(scale²)), in the form the state-space update reads.

condition_mixture

Exact Bayesian update of m on y = H x + c + eps, eps ~ N(0, RR^T).

predict

Push g through x' = F x + b + eps, eps ~ N(0, Q_chol Q_chol^T).

propagate

Exact grow: K components through M transitions -> K*M components.

rollout_reduced

Scan horizon steps of propagate -> reduce(k) with static shapes.

rollout_gpb1

The GPB1 rollout: at each of horizon steps, propagate() the belief through the M mode transitions, then moment-match the result to a single Gaussian.