1
$\begingroup$

It seems that higher-order tensors (of order 3 or higher) generally have unique decompositions under relatively mild conditions. For example, Kruskal proved that if an order-3 Tensor $T$ can be decomposed as the outer product of matrices $A, B, C$, with Kruskal ranks $k_A, k_B, k_C$ respectively, then as long as the rank $R$ of $T$ satisfies

$$R \le \frac{k_A + k_B + k_C - 2}{2},$$

then the decomposition is unique.

My question is: aren't matrix decompositions a special case of tensor decompositions? In other words, I can think of a matrix $M$ as a tensor whose third dimension is of size 1. If I was to apply the same result to matrix decompositions, I would get that that the matrix decomposition is unique as long as its rank is:

$$R \le \frac{k_A + k_B + 1 - 2}{2} = \frac{k_A + k_B - 1}{2},$$

(where we have assumed that $C$ is just a matrix of ones), but we know that this is generally not true. You can take a matrix, and rotate its component matrices, and get a different decomposition (the so-called "rotation problem", so matrix decompositions are not unique.

How do we reconcile these facts?

$\endgroup$
1
$\begingroup$

Kruskal rank can not be higher than the canonical rank, so you can never have:

                         2*R <= 2*R - 1

which means that the decomposition is never unique in the general case.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.