2
$\begingroup$

Let's consider the tensor $t \in \mathbb{R}^{d_1 \times d_2 \times d_3}$. It is well-know statement, that multilinear rank of tensor less or equal than border rank, which means that if we consider multilinear rank $rk_{mult} = (rk_1, rk_2, rk_3)$ than $rk_i(t) \le brk(t)$.

I have the proof of this simple fact, but I can not understand some notions in it. The proof is:

Let $t_n \to t$, $rk(t_n) = brk(t)$. Let's write the canonical decomposition for every $t_n: t_n = \sum a_i^n \otimes b_i^n \otimes c_i^n$. Then $t_1^n = \sum a_i^n \otimes (b_i^n \otimes c_i^n)$ converges to $t_1$. Then because $t_1^n$ is a matrix we have that $rk(t_1) \le rk(t_1^n) \le rk(t_n) = brk(t)$.

I can not understand one thing. Here is written that $t_1^n = \sum a_i^n \otimes (b_i^n \otimes c_i^n)$. But $t_1^n$ is the first flattering, it is a matrix it equals $t_1(x) = \sum (a_i x^T)b_i \otimes c_i$, how it can be the Kronecker product of 3 vectors? Is there any typo or I misunderstanding something? If here is a typo, how I can fix it?

Thanks for the help!

$\endgroup$
2
$\begingroup$

The confusion cames from the fact that the symbol "$\otimes $" can be interpreted as the tensor product, the outer product, and the Kronecker product. So, the expression $t=\sum a_i\otimes b_i\otimes c_i$ can be interpreted as a tensor (abstract notion), three-way array and a vector, respectively.

In you case, $t^n$ is interpreted as a matrix (flattening or unfolding): $t^n=\sum a_i^n\otimes (b_i^n\otimes c_i^n)$ should be read as $t^n=\sum a_i^n (b_i^n\otimes c_i^n)^T$, i.e. $t_n$ is a sum of rank-$1$ $d_1\times d_2d_3$ matrices.

$\endgroup$

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.