Struct IrlbaConfig 

pub struct IrlbaConfig {
    pub rank: usize,
    pub work: Option<usize>,
    pub tol: f64,
    pub max_restarts: usize,
    pub seed: Option<u64>,
    pub mean_center: bool,
    pub require_convergence: bool,
}

Configuration for svd_with.

Fields

rank: usize

Number of singular triplets wanted.

work: Option<usize>

Basis size. Must exceed rank; defaults to rank + 7, clamped to min(rows, cols).

This is the knob that matters on tall matrices

Each restart re-orthogonalises against the whole basis, so a step costs O(work · (rows + cols)) — on a matrix with a million rows that dominates the sparse products by a wide margin. A larger work makes each step dearer but converges in far fewer restarts, and the second effect wins comfortably.

Measured on 400k × 30k restricted to 2238 columns, 50 components (cargo run --release --example tune):

work	time	matvecs	accuracy
rank + 7 (default)	30.2 s	1515	6.7e-17
rank + 30	13.7 s	701	1.9e-14
rank + 50	18.6 s	601	2.0e-14
rank + 100	21.4 s	501	2.2e-14

So rank + 30 was 2.2× faster at the same accuracy. The cost is basis memory, which is linear in work: work · rows scalars for the left basis, or 600 MiB rather than 456 MiB at a million rows. The default stays conservative because memory is the reason to choose this crate; raise it when you have the headroom.

Run the tune example on your own shape rather than trusting these numbers — the optimum moves with the aspect ratio and the spectrum.

tol: f64

Relative residual tolerance: triplet i is accepted once its residual falls below tol · σ_max.

max_restarts: usize

Cap on restart cycles before giving up.

seed: Option<u64>

Fixed seed for the starting vector; None draws from the OS.

mean_center: bool

Subtract column means without materialising the centered matrix — i.e. PCA rather than plain SVD.

require_convergence: bool

Refuse to return triplets that did not reach tol. Defaults to true.

Exhausting the restart budget means the answer is a best effort of unknown quality. Returning it as Ok puts the burden on every caller to remember to inspect the diagnostics, and a pipeline that forgets gets a silently degraded decomposition feeding whatever comes next. Failing loudly is the safer default; set this to false if a best effort is genuinely what you want, then check SvdRec::converged.

Implementations

impl IrlbaConfig

pub fn new(rank: usize) -> Self

Configuration for rank triplets, everything else defaulted.

pub fn work(self, work: usize) -> Self

Set the basis size.

pub fn tol(self, tol: f64) -> Self

Set the relative residual tolerance.

pub fn max_restarts(self, n: usize) -> Self

Set the restart cap.

pub fn seed(self, seed: u64) -> Self

Fix the seed, making the result reproducible.

pub fn mean_center(self, yes: bool) -> Self

Enable implicit mean centering.

pub fn allow_unconverged(self) -> Self

Accept a best-effort result instead of failing when the restart budget runs out.

Trait Implementations

impl Clone for IrlbaConfig

fn clone(&self) -> IrlbaConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for IrlbaConfig

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

Formats the value using the given formatter.