Expand description
Exact shared-smoothing Gaussian REML for tangent-vector responses.
The model has a scalar predictor design X (N x K) and a tangent
response Y (N x D). A coefficient matrix B is fitted under
sum_i w_i (y_i - B' x_i)' M_i (y_i - B' x_i) + sum_b lambda_b tr(B' S_b B).
The implementation deliberately never constructs the stacked
(N D) x (K D) design or a S_b (x) I_D penalty. Isotropic metrics use
only K x K normal equations. Varying Fisher metrics stream exact joint
sufficient statistics into a (K D) x (K D) Gram matrix whose storage is
independent of N.
Structs§
- Response
Geometry Curvature - Typed curvature-as-estimand record carried by a response-geometry archive.
- Response
Geometry Metadata - Persistence and presentation metadata for
ResponseGeometryModel. - Response
Geometry Model - A complete typed response-geometry model archive.
- Response
Geometry Summary - Core summary of a fitted response-geometry model.
- Shared
Tangent Penalty - One compact predictor-space smoothing penalty.
- Shared
Tangent Reml Fit - A converged, serializable shared-tangent model.
- Shared
Tangent Reml Request - Owned request for a shared-tangent REML fit.
Enums§
Constants§
Functions§
- fit_
shared_ tangent_ reml - Fit a shared-smoothing multi-output Gaussian model by exact profiled REML.