Module vector_response 

Vector-valued response support.

Many smooths sharing one latent: the shape function in the latent-variable engine maps to a reduced activation vector (tens-to-hundreds of dimensions, after a random-matrix noise cut). This module defines the response-side types, the Gaussian vector likelihood, and the connector trait the inner solver consumes.

Conventions:

• Y is shape (N, M): N rows, M output dimensions.
• eta is shape (N, M): the linear predictor with one column per output.
• For Gaussian identity-link, mean(η) = η, so the likelihood depends only on eta and Y.

The Hessian is block-structured: per-row (N independent blocks for the Gaussian case), each of size (M, M). For a Gaussian likelihood with Diagonal/Isotropic noise this per-row block is itself diagonal — exactly what the arrow Schur elimination in solver/arrow_schur.rs consumes.

Structs

GaussianVectorLikelihood

Gaussian vector likelihood with identity link.

MultinomialLogitLikelihood

Multinomial-logit (softmax) likelihood with explicit reference class.

VectorResponseTarget

Vector-valued response target.

Enums

VectorNoise

Per-output noise model for a vector response.

Traits

VectorLikelihood

Connector trait the inner solver (Piece 1) plugs into.