Struct SubspaceEmbedding 

pub struct SubspaceEmbedding { /* private fields */ }

A random subspace embedding S: R^n → R^k (k × n projection matrix).

Implementations

impl SubspaceEmbedding

pub fn new(config: SubspaceEmbeddingConfig) -> OptimizeResult<Self>

Construct a new embedding according to config.

Errors

Returns OptimizeError::InvalidInput if either dimension is zero.

pub fn project(&self, x: &Array1<f64>) -> OptimizeResult<Array1<f64>>

Project a vector x ∈ R^n → y = Sx ∈ R^k.

Errors

Returns OptimizeError::ValueError if x.len() != original_dim.

pub fn reconstruct(&self, y: &Array1<f64>) -> OptimizeResult<Array1<f64>>

Approximate reconstruction x ≈ S^T y (pseudo-inverse via transpose).

Errors

Returns OptimizeError::ValueError if y.len() != embedding_dim.

pub fn project_matrix(&self, a: &Array2<f64>) -> OptimizeResult<Array2<f64>>

Project a matrix A ∈ R^{n×p} → SA ∈ R^{k×p}.

The embedding matrix S is k × n, so A must have n rows (i.e., a.nrows() == original_dim).

Errors

Returns OptimizeError::ValueError if a.nrows() != original_dim.

pub fn embedding_dim(&self) -> usize

Embedding (sketch) dimension k.

pub fn original_dim(&self) -> usize

Original dimension n.

pub fn jl_epsilon(&self, n_points: usize, delta: f64) -> f64

Johnson-Lindenstrauss ε-distortion bound.

For n_points points and failure probability delta, the embedding preserves all pairwise distances within factor (1 ± ε) with probability at least 1 − delta.

Formula: ε = sqrt(2 * ln(n / δ) / k).