Struct GridSpline2dDesign 

pub struct GridSpline2dDesign { /* private fields */ }

Banded sufficient statistics of one streaming pass plus the exact penalty: everything needed to evaluate the REML criterion and solve at any λ.

Implementations

impl GridSpline2dDesign

pub fn build( x1: &[f64], x2: &[f64], y: &[f64], w: &[f64], k: usize, metric: [f64; 2], ) -> Result<Self, String>

Single-response entry: see Self::build_multi.

pub fn build_multi( x1: &[f64], x2: &[f64], responses: &[&[f64]], w: &[f64], k: usize, metric: [f64; 2], ) -> Result<Self, String>

One streaming pass over the rows plus the exact per-cell quadrature assembly of the penalty. k is the number of cells per axis; metric = [a1, a2] is the diagonal anisotropy of the biharmonic form. responses holds one length-n response per dimension; the design, penalty, and the REML-shared λ are common to all dimensions (one surface smoothness), only the right-hand sides differ.

pub fn lower_corner(&self) -> [f64; 2]

Lower corner of the data bounding box per axis.

pub fn cell_widths(&self) -> [f64; 2]

Knot-cell width per axis.

pub fn penalty_value(&self, coeff: &[f64]) -> Result<f64, String>

Exact penalty quadratic form J(f) = c'Sc of a coefficient vector — the assembled anisotropic biharmonic energy of the spline it encodes.

pub fn posterior( &self, fit: &GridSpline2dFit, ) -> Result<GridSpline2dPosterior, String>

Posterior summary of a fit FROM THIS DESIGN, in the exact algebra of the solved system (no approximation):

• unit_covariance = (X'WX + λS)⁻¹ (scale-free Bayesian posterior covariance of the row-major coefficient vec, shared by dimensions);
• edf = tr[(X'WX + λS)⁻¹ X'WX] (the smoother’s effective degrees of freedom at the fitted λ);
• residual_cross_cov[d,e] = r_d'W r_e / (n − edf) assembled from the streamed sufficient statistics (y_d'Wy_e − c_d'X'Wy_e − c_e'X'Wy_d + c_d'X'WX c_e).