Struct GridSpline2dFit 

pub struct GridSpline2dFit {
    pub coeffs: Vec<Vec<f64>>,
    pub log_lambda: f64,
    pub sigma2: Vec<f64>,
    pub restricted_loglik: f64,
    /* private fields */
}

Fitted penalized tensor-product smoother with its factored covariance.

Fields

coeffs: Vec<Vec<f64>>

Per response dimension: coefficients in row-major flat order g = j1·(K+3) + j2.

log_lambda: f64

Selected (or supplied) log smoothing parameter, shared by all response dimensions.

sigma2: Vec<f64>

Per response dimension: profiled (or supplied) observation variance σ².

restricted_loglik: f64

Pooled restricted log-likelihood at the optimum, up to λ- and data-independent additive constants (exact REML differences across λ).

Implementations

impl GridSpline2dFit

pub fn to_state(&self) -> GridSpline2dState

Snapshot the fit for persistence (#1031). Verbatim — every field predict reads is copied; the training design is not retained on the fit and is not needed for replay.

pub fn from_state(state: &GridSpline2dState) -> Result<Self, String>

Rebuild a predict-capable fit from a snapshot (#1031). Validates shape, finiteness, positive cell widths/counts, positive σ², and that the basis arithmetic is self-consistent (m_axis = K + 3, chol is p × p, coeffs/sigma2 agree on D), so a corrupt payload fails here rather than inside a later predict. The restored fit replays the posterior mean+variance bit-for-bit: predict reads only the snapshotted fields.

pub fn predict( &self, dim: usize, x1: f64, x2: f64, ) -> Result<(f64, f64), String>

Posterior (mean, variance) of response dimension dim at an arbitrary point: the 16-entry basis row dotted with the coefficients, and σ̂²_dim·x'(X'WX+λS)⁻¹x through the retained Cholesky factor. Outside the bounding box the boundary cell’s cubic polynomial extends.