Struct LowRankHessian 

pub struct LowRankHessian {
    pub sigma: Number,
    pub vecs: Vec<Vec<Number>>,
    pub signs: Vec<Number>,
}

B = σI + Σᵢ signᵢ vᵢ vᵢᵀ — the L-BFGS matrix in product form, with no n × n storage (gh#1054).

Each stored pair contributes one −a aᵀ and one +b bᵀ term, a = B s / √(sᵀB s) and b = ȳ / √(sᵀȳ) with ȳ the Powell-damped y — the same rank-2 update LBfgs::as_triplet replays densely, so the two describe one matrix. A pair the dense replay skips (sᵀBs or sᵀȳ below 1e-14) is skipped here too.

Fields

sigma: Number

vecs: Vec<Vec<Number>>

The vᵢ, each of length n.

signs: Vec<Number>

−1 or +1 per vector.

Implementations

impl LowRankHessian

pub fn apply(&self, v: &[Number]) -> Vec<Number>

B v, in O(k n).

pub fn max_abs_entry(&self) -> Number

max |B_ij|. B is positive definite (BFGS with Powell damping), and for a PSD matrix the largest entry in magnitude sits on the diagonal (|B_ij| ≤ √(B_ii B_jj)), so this is the dense matrix’s max-norm exactly, at O(k n).

Trait Implementations

impl Clone for LowRankHessian

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for LowRankHessian

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

Formats the value using the given formatter.