Higher-order singular value decomposition: Difference between revisions

Revision as of 20:18, 28 June 2025

Revision as of 20:21, 28 June 2025

Line 101:

* Compute a rank-<math>\bar R_m </math> truncated SVD <math>\mathcal{A}_{[m]} \approx U_m \Sigma_m V^T_m </math>, and store the top <math>\bar R_m </math> left singular vectors <math>U_m \in F^{I_m \times \bar R_m}</math>;

while a '''sequentially truncated M-mode SVD (HOSVD)''' (or '''successively truncated M-mode SVD(HOSVD)''') is obtained by replacing step 2 in the interlaced computation by

* Compute a rank-<math>\bar R_m </math> truncated SVD <math>\mathcal{A}_{[m]}^{m-1} \approx U_m \Sigma_m V^T_m </math>, and store the top <math>\bar R_m </math> left singular vectors <math>U_m \in F^{I_m \times \bar R_m}</math>. Unfortunately, truncation does not result in an optimal solution for the best low multilinear rank optimization problem,.<ref name=":2" /><ref name=":Vasilescu2002"/><ref name=":4" /><ref name=":fist_hosvd" /> However, both the classically and interleaved truncated M-mode SVD/HOSVD result in a '''quasi-optimal''' solution:<ref name=":4" /><ref name=":fist_hosvd" Vasilescu2003/><ref name=":5" /><ref>{{Cite journal|last=Grasedyck|first=L.|date=2010-01-01|title=Hierarchical Singular Value Decomposition of Tensors|journal=SIAM Journal on Matrix Analysis and Applications|volume=31|issue=4|pages=2029–2054|doi=10.1137/090764189|issn=0895-4798|citeseerx=10.1.1.660.8333}}</ref> if <math>\mathcal{\bar A}_t </math> denotes the classically or sequentially truncated M-mode SVD(HOSVD) and <math>\mathcal{\bar A}^* </math> denotes the optimal solution to the best low multilinear rank approximation problem, then<math display="block">\| \mathcal{A} - \mathcal{\bar A}_t \|_F \le \sqrt{M} \| \mathcal{A} - \mathcal{\bar A}^* \|_F; </math>in practice this means that if there exists an optimal solution with a small error, then a truncated M-mode SVD/HOSVD will for many intended purposes also yield a sufficiently good solution.

== Applications ==