Struct FixedDesignRowCache 

pub struct FixedDesignRowCache { /* private fields */ }

Cached fixed design rows for GLM / changing-W PIRLS trials.

This cache owns the theta-invariant X rows once. Each trial recomputes X'WX and X'Wz because the scalar working weights and working response genuinely move during PIRLS. The Gaussian constant-Gram trick does not apply when W changes; the saved work is the expensive measure-jet basis/design construction, not the unavoidable weighted contraction over fixed rows.

Implementations

impl FixedDesignRowCache

pub fn build(x: ArrayView2<'_, f64>) -> Result<Self, String>

Cache a finite, non-empty fixed design.

pub fn n(&self) -> usize

pub fn p(&self) -> usize

pub fn design(&self) -> ArrayView2<'_, f64>

pub fn xtwx(&self, weights: ArrayView1<'_, f64>) -> Result<Array2<f64>, String>

Recompute X' diag(weights) X over cached rows.

This remains O(n p^2), the irreducible weighted contraction when W changes. It avoids rebuilding the n-row measure-jet design.

pub fn xtwz( &self, weights: ArrayView1<'_, f64>, z: ArrayView1<'_, f64>, ) -> Result<Array1<f64>, String>

Recompute X' diag(weights) z over cached rows for a PIRLS response.