6
$\begingroup$

Given positive semidefinite matrices $A,B \succeq 0$, $A, B \in \mathbb{R}^{n \times n}$, if we have $$\langle A, B \rangle = 0,$$ where $\langle \cdot, \cdot \rangle$ denotes the Frobenius inner product, then

  1. What are tight necessary conditions of ranks of $A, B$? For example, $\mbox{rank} (A) + \mbox{rank}(B) \leq n$? How to prove that?

  2. What can be say about the orthogonality of the eigenvectors of $A$ and $B$? For example, are they pairwise orthogonal?

$\endgroup$
  • $\begingroup$ I would start by WLOG assuming $B$ is diagonal with non-negative entries (by choice of suitable o.n. basis) and then taking $A=C^\top C$ for a symmetric matrix $C$... does this help? (Sorry, have to run and catch a bus) $\endgroup$ – Yemon Choi Jan 13 '17 at 21:51
8
$\begingroup$

If $A$ and $B$ are positive semidefinite and $A^{1/2}$ and $B^{1/2}$ their positive semidefinite square roots,

$$\text{Tr}(AB) = \text{Tr}(A^{1/2} B A^{1/2}) = \text{Tr}((B^{1/2} A^{1/2})^T B^{1/2} A^{1/2})$$ and this is $0$ if and only if $B^{1/2} A^{1/2} = 0$.
That in turn is equivalent to $\text{Ran}(A) \subseteq \text{Ker}(B)$. So indeed $\text{rank}(A) + \text{rank}(B) \le n$, and this bound is tight. Moreover, the eigenvectors of $A$ for nonzero eigenvalues are in $\text{Ker}(B)$ and therefore are orthogonal to the eigenvectors of $B$ for nonzero eigenvalues.

$\endgroup$

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.