2
$\begingroup$

As in the object, I'm looking at the case where $x \in \mathbb R^d$ is a generic vector, and $AA^\top \in \mathbb R^{d \times d}$ is a p.s.d. matrix.

I'm investigating the following inequality

$x^\top e^{-xx^\top - AA^\top} x \leq x^\top e^{-xx^\top} x,$

where I'm considering matrix exponentials. In general, it is not true that if $BB^\top \geq CC^\top$ in p.s.d. sense, then we have $e^{BB^\top} \geq e^{CC^\top}$. However, here, I'm looking for something weaker, but I'm struggling to prove it / found a counter example!

$\endgroup$

1 Answer 1

4
$\begingroup$

Alas, this is not correct. You may try $d=2$, $x=(a, a)^\top$ for certain $a>0$, and choose $A$ of rank 1 such that $AA^\top+xx^\top=\operatorname{diag}(p, q)$. For positive $p,q$ such matrix $A$ exists iff $$0=\det(-xx^\top+\operatorname{diag}(p, q))=(p-a^2)(q-a^2)=a^4,$$ or $1/p+1/q=1/a^2$. Your inequality reads as $(e^{-p}+e^{-q})a^2\leqslant 2a^2e^{-2a^2}$. If we denote $1/p=x,1/q=y$, then $x+y=1/a^2$ is fixed and we need $f(x)+f(y)$, where $f(x)=e^{-1/x}$, be maximal when $x=y$. But $f$ is not concave on $(0,\infty)$, so this does not hold in general.

$\endgroup$
2
  • $\begingroup$ Thank you! Very nice approach! I have a follow up question though. Do you think a similar inequality could be proved, with a correction coefficient that depends on $\| A^\top x \|_2$? A case that breaks the previous inequality, suggested by your proof, is given by using column spaces of $AA^\top$ and $xx^\top$ to be "roughly orthogonal"... $\endgroup$
    – Simone256
    Commented 2 days ago
  • $\begingroup$ I do not understand what exactly do you mean $\endgroup$ Commented 2 days ago

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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