Struct DeflationSpec 

pub struct DeflationSpec {
    pub basis: Vec<Array1<f64>>,
}

Top-subspace (Hutch++) deflation configuration for the surrogate. When a plan carries one, RationalLogdetPlan::evaluate peels an r-dimensional orthonormal subspace Q of the heavy (top) directions from the operator and splits the log-determinant by the EXACT identity (no invariance assumed)

tr log(S/c) = tr(Qᵀ log(S/c) Q) + tr(P log(S/c) P), P = I − QQᵀ,

evaluating the first block deterministically over the r basis columns and the second by Hutchinson over the PROJECTED probes u_j = P v_j (each with its own reference norm ‖u_j‖², so the k − r bookkeeping is automatic). The Hutchinson variance then rides only on the off-diagonal mass of P log(S/c) P — small once Q captures the heavy directions — collapsing the error bar that raw probes carry on a wide spectrum. The decomposition is EXACT for ANY orthonormal Q; the subspace iteration only steers Q toward the top space to reduce variance, it can never bias the estimate.

The basis is FROZEN here (built once by RationalLogdetPlan::with_deflation from the operator at the plan’s ρ), NOT rebuilt per evaluation. This is what keeps value and gradient the SAME functional: with the estimated term2, the sum term1 + term2 is Q-dependent, so a Q that moved with ρ would put an un-modelled ∂Q/∂ρ term in the true gradient. A frozen Q makes the fixed-Q directional derivative EXACT for the surrogate, at the cost of Q going slightly stale as the line search moves ρ (which only relaxes the variance reduction — never biases the value, since the decomposition is exact for any fixed orthonormal Q).

Fields

basis: Vec<Array1<f64>>

Frozen orthonormal top-subspace basis Q (columns q_i), built once from the operator. Reused verbatim across every ρ evaluation (CRN).

Trait Implementations

impl Clone for DeflationSpec

fn clone(&self) -> DeflationSpec

Returns a duplicate of the value.

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

Performs copy-assignment from source.