gam 0.3.116

Generalized penalized likelihood engine
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34

//! SAE-manifold term configuration.
//!
//! This is the formal Methodspace row for the SAE-manifold term:
//!
//! ```text
//! Z_i ~= sum_k a_ik g_k(t_ik),     g_k(t) = Phi_k(t) B_k
//! ```
//!
//! Tier assignment:
//!
//! * beta: [`SaeManifoldAtom::decoder_coefficients`] (`B_k`, one block per atom).
//! * ext-coords: [`SaeAssignment`] (`logits -> a_ik` and per-atom
//!   `LatentCoordValues`). Softmax uses the identifiable reference-logit chart
//!   with `K - 1` free assignment coordinates (`0` for `K = 1`). Per-row latent coordinates are written `t`; existing
//!   kernel-shape state remains with carriers such as `SpatialLogKappaCoords`.
//! * rho: [`SaeManifoldRho`] (`lambda_sparse`, `lambda_smooth`, `alpha_kj`) plus
//!   the discrete `K` selected by the Python `compare_models` wrapper.
//!
//! The per-row local block is exactly the audit-revised shape:
//!
//! ```text
//! ext_i = (assignment chart_i, t_i0[0..d_0], ..., t_iK[0..d_K])
//! dim(ext_i) = assignment_dim + sum_k d_k
//! ```
//!
//! [`SaeManifoldTerm::assemble_arrow_schur`] materializes the Gauss-Newton
//! bordered Hessian in that layout and hands it to
//! [`crate::solver::arrow_schur::ArrowSchurSystem`].

// Split from the original oversized module; keep included in order.
include!("sae_manifold/part_000.rs");
include!("sae_manifold/part_001.rs");
include!("sae_manifold/part_002.rs");
include!("sae_manifold/part_003.rs");