Trait BlockExcessTarget 

pub trait BlockExcessTarget {
    // Required methods
    fn block_dim(&self) -> usize;
    fn rho_dim(&self) -> usize;
    fn block_curvatures(&self) -> &Array1<f64>;
    fn excess(&self, t: &Array1<f64>) -> f64;
    fn excess_rho_gradient(&self, t: &Array1<f64>) -> Array1<f64>;
    fn displaced_neg_score(&self, t: &Array1<f64>) -> Array1<f64>;
    fn base_neg_score(&self) -> Array1<f64>;

    // Provided methods
    fn excess_with_displaced_neg_score(
        &self,
        t: &Array1<f64>,
    ) -> (f64, Option<Array1<f64>>) { ... }
    fn excess_with_displaced_neg_score_batch(
        &self,
        draws: &Array2<f64>,
    ) -> Vec<(f64, Option<Array1<f64>>)> { ... }
}

Caller-supplied evaluator for the non-Gaussian remainder ΔF(t) of the local log-posterior, restricted to the curvature-heavy block subspace (issue #784).

Implemented by gam-solve’s Gam784BlockTarget; consumed by LaplaceMarginalSampler::block_sampled_marginal_correction. Lives in this neutral crate so both the implementor (gam-solve) and the sampler impl (the gam-inference monolith) name the same trait without an SCC edge.

Required Methods

fn block_dim(&self) -> usize

Dimension m of the block subspace (number of untrustworthy directions being sampled).

fn rho_dim(&self) -> usize

Number of outer ρ coordinates the gradient is reported against.

fn block_curvatures(&self) -> &Array1<f64>

Block curvatures λ_r (the H-eigenvalues of the sampled directions), length block_dim().

fn excess(&self, t: &Array1<f64>) -> f64

Non-Gaussian remainder ΔF(t) at whitened block displacement t (length block_dim()).

fn excess_rho_gradient(&self, t: &Array1<f64>) -> Array1<f64>

ρ-gradient ∂ΔF/∂ρ_k at the same t, length rho_dim() — the explicit penalty-score channel (a).

fn displaced_neg_score(&self, t: &Array1<f64>) -> Array1<f64>

Per-row displaced score ∂(D(η̂+s(t))/2φ)/∂η evaluated at η̂ + s(t) (length = number of observation rows): the only per-draw ingredient of the exact-gradient channels (b)–(d) the assembly side cannot reconstruct.

fn base_neg_score(&self) -> Array1<f64>

The same per-row score channel at the undisplaced mode η̂.

Provided Methods

fn excess_with_displaced_neg_score( &self, t: &Array1<f64>, ) -> (f64, Option<Array1<f64>>)

Fused (excess(t), displaced_neg_score(t)). The returned score is None exactly when the excess is non-finite (an infeasible draw the sampler discards before reading the score). The default preserves the two-call behavior; implementors override to share the displacement + jet.

fn excess_with_displaced_neg_score_batch( &self, draws: &Array2<f64>, ) -> Vec<(f64, Option<Array1<f64>>)>

Batched Self::excess_with_displaced_neg_score over many whitened draws (one draw per COLUMN, shape block_dim() × n_draws). Batching may only change HOW the shared linear algebra is computed (one BLAS-3 product over all columns), never WHAT is computed. The default preserves the per-column behavior exactly; the GLM implementor overrides it.

Dyn Compatibility

This trait is dyn compatible.

In older versions of Rust, dyn compatibility was called "object safety".

Implementors