Struct GaussianJackknifePlusStats 

pub struct GaussianJackknifePlusStats { /* private fields */ }

Test-point-independent sufficient statistics for the exact penalized Gaussian-identity jackknife+, factored ONCE from (X, y, Sλ) so any number of test points reuse the single Cholesky of M = XᵀX + Sλ.

For each training row i the leave-one-out fit is the rank-one Sherman–Morrison downdate of M, giving (in closed form, no refits):

  vᵢ = M⁻¹ xᵢ                         (p-vector, one column of M⁻¹Xᵀ)
  hᵢ = xᵢᵀ vᵢ ,  cᵢ = rᵢ / (1 − hᵢ)   (signed LOO residual)
  Rᵢ = |cᵢ|                            (LOO absolute residual)

At a test point x_* the LOO prediction is then a single inner product per row, μ̂₋ᵢ(x_*) = x_*ᵀβ̂ − (x_*ᵀvᵢ)·cᵢ, so interval is O(n·p) after the O(n·p²) factorization here. This is the substrate the predict(interval=level) magic default replays: the stats are exactly the {vᵢ, cᵢ, Rᵢ} of gaussian_jackknife_plus, which is recovered exactly when fed a single x_* (the in-module test asserts that equivalence).

Unit prior weights are required, as everywhere in this module: a reweighted training row is not exchangeable with the test row, so the finite-sample coverage proof does not apply. The constructor rejects non-unit weights and rows with 1 − hᵢ ≤ 1e-10 (no leave-one-out information).

Implementations

impl GaussianJackknifePlusStats

pub fn new( x: &Array2<f64>, y: &Array1<f64>, prior_weights: &Array1<f64>, s_lambda: &Array2<f64>, ) -> Result<Self, String>

Factor the test-point-independent jackknife+ statistics from the training design, response, prior weights, and penalty matrix.

pub fn from_design_unit_weight_normal_matrix( x: &Array2<f64>, y: &Array1<f64>, prior_weights: &Array1<f64>, m: &Array2<f64>, ) -> Result<Self, String>

Same exact jackknife+ statistics as new, but the penalized normal matrix M = XᵀX + Sλ is supplied directly rather than reassembled from Sλ. For a Gaussian-identity unit-weight fit the converged penalized Hessian stored in [FitGeometry] is this M (the working weights are unity and the matrix is dispersion-unscaled), so persisting the design + M at fit time and replaying through this constructor reproduces the certified interval with no penalty re-derivation — the seam the saved-model predict(interval=…) magic uses.

prior_weights is validated to be unity for the exchangeability guarantee, identically to new.

pub fn n(&self) -> usize

Number of training rows backing the leave-one-out construction.

pub fn p(&self) -> usize

Coefficient dimension p.

pub fn beta(&self) -> &Array1<f64>

Full-model coefficient vector β̂ = M⁻¹Xᵀy. The plug-in mean at a test point x_* is x_*ᵀ β̂. Exposed so the pyffi layer can emit the mean / linear_predictor columns alongside the jackknife+ bounds without re-running the predictor stack.

pub fn interval( &self, x_star: &Array1<f64>, alpha: f64, ) -> Result<JackknifePlusInterval, String>

Jackknife+ interval at one test row x_* and miscoverage alpha, returning the Barber et al. (2021) set with guarantee P(Y_* ∈ Ĉ) ≥ 1 − 2·alpha.

Trait Implementations

impl Clone for GaussianJackknifePlusStats

fn clone(&self) -> GaussianJackknifePlusStats

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for GaussianJackknifePlusStats

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

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for GaussianJackknifePlusStats

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for GaussianJackknifePlusStats

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.