Struct AtomInnerFit 

pub struct AtomInnerFit {
    pub design: Array2<f64>,
    pub derivative_design: Array2<f64>,
    pub beta: Array1<f64>,
    pub penalty: Array2<f64>,
    pub penalized_hessian: Array2<f64>,
    pub row_scores: Array2<f64>,
    pub weights: Array1<f64>,
    pub dispersion: f64,
    pub peak_design_row: Array1<f64>,
    pub mode_design_row: Array1<f64>,
}

The fitted per-atom inner-decoder smooth, captured once at fit time so the post-PIRLS atom-inference reports reuse the same design, penalized Hessian, and per-row scores the identifiability certificate’s curvature sees.

The SAE decoder reconstructs Z_i ≈ Σ_k a_ik Φ_k(t_ik) B_k. Holding all other atoms and the assignment fixed at the fitted optimum, atom k’s own contribution along a single output channel j is the Gaussian-identity penalized smooth a_ik · Φ_k(t_ik)ᵀ β_{k,j} with roughness penalty S_k, Gauss–Newton observation weight w_i = a_ik² (the assignment mass enters the channel linearly, so the normal-equation weight is its square), and dispersion the fitted reconstruction dispersion. That is an ordinary penalized WLS smooth — exactly what crate::inference::riesz, gam_terms::inference::lawley, and the κ-profile machinery consume. The channel j is the atom’s dominant decoder output direction (largest column norm of B_k), i.e. the channel that carries the atom’s signal.

Fields

design: Array2<f64>

Φ_k evaluated on the atom’s active rows, (n_active, M_k). The inner GAM smooth design. Column 0 is the constant/intercept basis column.

derivative_design: Array2<f64>

∂Φ_k/∂t along the atom’s leading latent axis on the active rows, (n_active, M_k): the derivative design the average-derivative functional integrates.

beta: Array1<f64>

The fitted decoder coefficients for the captured output channel, β_{k,j} ∈ ℝ^{M_k}.

penalty: Array2<f64>

The atom roughness Gram S_k, (M_k, M_k).

penalized_hessian: Array2<f64>

The penalized Hessian H = ΦᵀWΦ + S_k at the fitted state, (M_k, M_k).

row_scores: Array2<f64>

Per-row Gaussian-identity scores s_i = ∂nll_i/∂β = −w_i r_i Φ_i / φ, (n_active, M_k), on the captured channel.

weights: Array1<f64>

Per-row Gauss–Newton weights w_i = a_ik² on the captured channel.

dispersion: f64

Fitted reconstruction dispersion φ (Gaussian σ²).

peak_design_row: Array1<f64>

Design row at the latent peak t_peak (largest fitted |g_k|).

mode_design_row: Array1<f64>

Design row at the latent mode t_mode (largest assignment mass).

Trait Implementations

impl Clone for AtomInnerFit

fn clone(&self) -> AtomInnerFit

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for AtomInnerFit

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.