gaussian_emission

gaussian_emission#

hybrid_aif.gaussian_emission(loc, R_chol, strength=1.0, anchor=False)[source]#

A Gaussian emission over categorical parents alone, y = c_j + noise_j with the parents’ configuration j selecting the offset and the noise: the matrix-normal inverse-Wishart family with an empty regressor block, its head the offset column only.

loc is (*parent_shape, d_obs), one offset per parent configuration (a Gaussian mixture’s component means, the parent its component label); R_chol is per configuration, (*parent_shape, d_obs, d_obs), or shared, (d_obs, d_obs). strength and anchor are LinearGaussianObservationMNIW.from_head_noise()’s. A modality declared this way names only categorical state factors in A_deps; its message to them is the exact categorical message of its predictive density, and it learns by the responsibility-weighted conjugate update at zero regressor width.

Design: docs/design/structure-learning.md S2