Module decomp 

Matrix decompositions, one per file — general-purpose kernels for future use across the library (regression, calibration, PCA, factor models), independent of any finance semantics.

• cholesky: A = L L^T for symmetric PSD matrices, plus the SPD linear solve through the factor — the fast path for normal equations and covariance sampling;
• [qr]: Householder QR (A = Q R, thin form), plus numerically stable linear least squares — the right tool for regression (e.g. Longstaff-Schwartz bases) without forming A^T A;
• [svd]: one-sided Jacobi singular value decomposition (A = U S V^T), plus the minimum-norm pseudo-inverse solve for rank-deficient problems;
• eigen: cyclic Jacobi eigendecomposition of symmetric matrices — the engine behind the PSD projection in nearest_correlation.

All matrices are Vec<Vec<f64>> row-major, matching the rest of the crate; the implementations favor clarity and robustness on the small-to-moderate sizes quant workflows use.

Re-exports

pub use cholesky::cholesky_factor;

pub use cholesky::cholesky_solve;

pub use eigen::symmetric_eigen;

pub use qr::least_squares;

pub use qr::qr;

pub use svd::pseudo_solve;

pub use svd::svd;

Modules

cholesky

Cholesky factorization A = L L^T of a symmetric positive (semi-)definite matrix, and the SPD linear solve through the factor.

eigen

Eigendecomposition of symmetric matrices by the cyclic Jacobi method — robust and accurate for the small matrices of correlation and covariance work (the PSD projection inside Higham’s nearest-correlation algorithm runs on this).

qr

Householder QR decomposition and numerically stable linear least squares (no A^T A squaring of the condition number).

svd

Singular value decomposition by one-sided Jacobi rotations — simple, accurate for the small-to-moderate matrices of quant workflows, and rank-revealing. A = U diag(S) V^T with orthonormal U (m x n, columns for nonzero singular values), non-negative S sorted descending, and orthogonal V (n x n).