Unifying singular value decompositions of tensors via aligned orthogonality

Abstract

We study basis-aligned two-orthogonal (bato) tensors, which can be written as a sum of critical rank-one approximations whose factors are singular vectors of their flattenings. As such, bato tensors admit a Tucker decomposition and a canonical polyadic decomposition that are related to each other. We prove that generic bato decompositions are identifiable and that their truncations are critical low-bato-rank approximations. We also compute the dimension of the set of bato tensors, and identify the irreducible components of its Zariski closure with isotopy classes of maximal partial Latin hyperrectangles. Many important tensors are bato, such as determinants, matrix multiplication tensors, and other structure tensors of algebras.