Expand description
fpcoa: Principal Coordinate Analysis (PCoA) with fixed-rank randomized SVD. Assumptions:
dist
is dense, strictly symmetric, with zeros on the diagonal. Pipeline (Halko–Martinsson–Tropp range finder + symmetric core):
- Build B = -0.5 * J * (D ∘ D) * J (double-centering)
- Q = subspace_iteration_full(B, rank = k + oversample, nbiter = q) (annembed/lax)
- B_r = Qᵀ * B * Q (r × r, symmetric)
- Symmetric eig (nalgebra): B_r = V Λ Vᵀ (Λ sorted ↓)
- Zero-out negative eigenvalues (and their eigenvectors)
- U ≈ Q * V_k
- Coordinates = U * sqrt(Λ_k)
Structs§
- Fpcoa
Options - Options for randomized PCoA.
- PCoA
Result - Result of PCoA.
Functions§
- pcoa_
randomized - Perform PCoA via randomized range finder + symmetric eig on the small core.