Multilinear PCA (MPCA)

Multilinear PCA — one orthonormal factor per feature mode, H (3 × r_h), W (3 × r_w), B (4 × r_b), and a small score tensor per sample. Spanish: PCA multilineal.

The sample axis survives: these are features for each sample, not a compression of sample identities. The flattened basis is H ⊗ W ⊗ B — separable, unlike ordinary PCA.

With ranks (2, 2, 3) there are 12 scores and 24 stored factor entries, against 432 for a dense 36 × 12 basis — 18 times fewer. (Both also store the 36-entry mean.)