0
$\begingroup$

Let $d',d \in \mathbb{N}$, with $d' \le d$ and let $\Sigma$ be a symmetric matrix of dimension $d$. I'm trying to solve the following problem:

$$\max_{\substack{L \in \mathcal{M}_{d'\times d}(\mathbb{R}) \\LL^T = I}} \quad trace\left(L \Sigma L^T\right)$$

I can prove that, when $\Sigma$ is positive semidefinite, a solution can be found by adding to each row of $L$ the eigenvectors of $\Sigma$ associated to its $d'$ largest eigenvalues. I have also read some papers where they use this fact assuming only that $\Sigma$ is symmetric.

However, in the proof I used when $\Sigma$ is positive semidefinite, I use an inequality that depends critically on the non negativity of the eigenvalues of $\Sigma$, so that if there were negative eigenvalues, that inequality would be false. That's why I can't generalize this proof to any symmetric matrix.

How can I make a proof for this general case? Thanks in advance.

$\endgroup$
  • $\begingroup$ It would be nice if you write your proof for the case you wrote about. $\endgroup$ – Royi Jul 4 '18 at 20:55
2
$\begingroup$

I suspect that your inequality would work just fine with negative values.

That being said, we can extend your result as follows. If $\Sigma$ is symmetric, then there exists an $\alpha > 0$ such that $\Sigma + \alpha I$ is positive semidefinite. We note that for any $L$ with $LL^T = I$, $$ \operatorname{trace}(L(\Sigma + \alpha I)L^T) = \operatorname{trace}(L\Sigma L^T) + \alpha \operatorname{trace}(LL^T) = \operatorname{trace}(L\Sigma L^T) + \alpha d' $$ Conclude that if $L$ maximizes $\operatorname{trace}(L(\Sigma + \alpha I)L^T)$, then it also maximizes $\operatorname{trace}(L\Sigma L^T)$.

$\endgroup$

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.