Struct CompiledMap 

pub struct CompiledMap {
    pub raw_from_compiled: Array2<f64>,
    pub compiled_block_ranges: Vec<Range<usize>>,
    pub raw_block_ranges: Vec<Range<usize>>,
}

Closed-form Gram-based compile output: a single p_raw × p_compiled reparam matrix T mapping compiled coordinates back to raw width. T · θ lifts a fitted compiled-width β back to raw width; predict-time row contribution is X_raw · T · θ where X_raw is the full raw design.

compiled_block_ranges[b] gives the column range inside T (and inside the compiled-width coefficient vector) attributable to raw block b. raw_block_ranges[b] gives the corresponding raw-width column range.

Fields

raw_from_compiled: Array2<f64>

(p_raw × p_compiled) raw-from-compiled reparam matrix.

compiled_block_ranges: Vec<Range<usize>>

Per-block compiled-width column ranges, parallel to raw_block_ranges. Same length as the input ordering.

raw_block_ranges: Vec<Range<usize>>

Per-block raw-width column ranges (copied through from input).

Implementations

impl CompiledMap

pub fn p_raw(&self) -> usize

Raw coefficient width (p_raw).

pub fn p_compiled(&self) -> usize

Compiled (reduced) coefficient width (p_compiled).

pub fn reduce_design( &self, raw_design: &Array2<f64>, ) -> Result<Array2<f64>, String>

Reparameterise a raw design into compiled coordinates: X_compiled = X_raw · T (n × p_compiled). Because the lift is β_raw = T β_compiled, the compiled design predicts identically to the raw design on every compiled coefficient: X_compiled · θ = X_raw · (T θ). Families that build directly in reduced coordinates feed this compiled design (and the reduce_penalties_with_map penalties) to the solver; the rank-deficient raw basis never reaches Newton.

pub fn lift_coefficients( &self, beta_compiled: &Array1<f64>, ) -> Result<Array1<f64>, String>

Lift a fitted compiled-width coefficient vector back to raw width: β_raw = T · β_compiled. This is the exact inverse direction of the quotient reduction — the reduced coordinates are what Newton/REML operate in, and this map carries the final estimate (and any linear functional of it) back to the original parameterisation so reported coefficients and predictions match the raw design.

Trait Implementations

impl CompiledBlockMap for CompiledMap

Neutral view of this compiled reparametrisation for the gauge layer (#1521): Gauge::from_compiled_map lives DOWN in gam-problem and names only the CompiledBlockMap trait, never the concrete CompiledMap (which lives ABOVE gam-problem). This impl supplies the inverted dependency edge.

fn raw_from_compiled(&self) -> &Array2<f64>

The (p_raw × p_compiled) raw-from-compiled reparam matrix T.

fn raw_block_ranges(&self) -> &[Range<usize>]

Per-block raw-width column ranges.

fn compiled_block_ranges(&self) -> &[Range<usize>]

Per-block compiled-width column ranges, parallel to Self::raw_block_ranges.

impl Debug for CompiledMap

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

Formats the value using the given formatter.