Struct LowRankApprox 

pub struct LowRankApprox {
    pub rank: usize,
    pub u: Array2<f32>,
    pub vt: Array2<f32>,
    pub singular_values: Array1<f32>,
    pub reconstruction_error: f32,
}

Low-rank approximation W ≈ U @ diag(S) @ V^T computed via pure-Rust power iteration (no LAPACK / C dependencies).

Fields

rank: usize

Target rank

u: Array2<f32>

Left singular vectors — shape (rows, rank)

vt: Array2<f32>

Right singular vectors (transposed) — shape (rank, cols)

singular_values: Array1<f32>

Singular values — shape (rank,)

reconstruction_error: f32

Relative Frobenius reconstruction error ||W - approx||_F / ||W||_F

Implementations

impl LowRankApprox

pub fn compute( w: &Array2<f32>, rank: usize, num_iter: usize, ) -> ModelResult<Self>

Compute a rank-rank approximation of w using power iteration.

num_iter controls the number of power-iteration steps per component. Higher values give more accurate singular vectors.

pub fn reconstruct(&self) -> ModelResult<Array2<f32>>

Reconstruct the full matrix: U @ diag(S) @ V^T.

pub fn compression_ratio(&self) -> f32

Compression ratio: (rows * cols) / (rows * rank + rank * cols).

pub fn forward(&self, x: &Array1<f32>) -> ModelResult<Array1<f32>>

Fast forward pass using factored form: (U @ diag(S)) @ (V^T @ x).

x must have length equal to the number of columns (original input dim).

Trait Implementations

impl Clone for LowRankApprox

fn clone(&self) -> LowRankApprox

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for LowRankApprox

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

Formats the value using the given formatter.