Is separability a property of the homomorphism or just the matrix that represents it?

118 Views Asked by At

I've been trying to learn about separable matrices in the context of the Singular Value Decomposition. I suspect one of the following is true but don't know which:

  • If a homomorphism is separable, it remains so no matter what bases are chosen for the domain and codomain; or
  • Any homomorphism can be written as a separable matrix with the right choice of bases.

Is either of these true? Suggested references for this sort of thing would also be greatly appreciated.

1

There are 1 best solutions below

1
On BEST ANSWER

This is a little bit difficult to google for so for the sake of any readers who, like me, had not seen this term before: a separable matrix $M$ is a matrix that can be written as an outer product $v \otimes u^T$ of two vectors. Here is the relevant result:

Claim: Let $U, V$ be vector spaces with bases $B, C$, let $T : U \to V$ be a linear transformation, and let $M$ be the matrix of $T$ with respect to the given bases. Then $M$ is separable iff $T$ has rank at most $1$, and in particular this condition does not depend on the choice of bases.

$U$ and $V$ don't need to be finite-dimensional here but we can assume that condition for simplicity.

Proof. If $M = v \otimes u^T$ is separable then every column of $M$ is a scalar multiple of $v$, so $M$ (and hence $T$) has rank at most $1$.

Conversely, if $T$ has rank at most $1$, then either $T = 0$ (in which case the desired property is trivially satisfied) or any matrix $M$ of $T$ with respect to any bases has the property that, if we pick some nonzero column $v$, then every column is a scalar multiple of this column. These scalars can then be assembled into a vector $u$ such that $M = v \otimes u^T$ as desired. $\Box$

To say all this in a coordinate-independent way, a linear map $T : U \to V$ has rank at most $1$ iff it can be written $T = v \otimes f$ where $f \in U^{\ast}$ is an element of the dual space. This generalizes to any finite rank as follows: a linear map has rank at most $k$ iff it can be written $T = \sum_{i=1}^k v_i \otimes f_i$ where $v_i \in V, f_i \in U^{\ast}$. Again $U$ and $V$ don't need to be finite-dimensional.