LinearGaussianTransition#

class hybrid_aif.LinearGaussianTransition[source]#

Bases: Module

x' = F_j x + G_j u + b_j + eps_j: linear map(s) with conditioned innovations.

The mode index j runs over the state factor’s discrete parents (see innovations); F and G are shared across modes or given per mode (the switching linear dynamical system pattern).

F: Array#

The state map: (d, d), shared across modes, or (*parent_shape, d, d) matching innovations.parent_shape, one map per mode.

innovations: ConditionalGaussian#

A ConditionalGaussian whose parent axes are the state factor’s discrete parents in declared (B_deps) order: its discrete state parents and any discrete control node, each a mode index j. A state factor with no discrete parent declares the trivial parent shape (1,) (always index 0). Its location is the offset b_j.

G: Array | None = None#

The control input map of the state factor’s continuous control nodes, u their values concatenated in declared order: (d, du) shared, or (*parent_shape, d, du) per mode on the rule F follows; None when the state factor reads no continuous node.

noise_prior: InverseGamma | InverseWishart | Parameter | None = None#

A conjugate prior over the innovations’ covariance Q (InverseGamma, per-dimension variances, or InverseWishart, a full covariance), shared across the modes or with leading axes innovations.parent_shape (one prior per mode); or Parameter(name), reading a noise held in SystemParams.parameters. When set, every consumer reads the innovations’ Cholesky factor lowered from the prior’s expected-precision point (transition_point()) and the declared innovations.chol is ignored (only its shape is a declaration); learn updates the prior from the pairwise residual scatter, under the rate key transition:<factor>:noise. None: Q is declared, not learned. Design: process-noise-learning.md D1, D4.

property learnable_noise: bool#

Capability, per-field: the process noise is a posterior (noise_prior holds one).

__init__(F, innovations, G=None, noise_prior=None)#