1
$\begingroup$

I am reading a paper which repetitively uses the following statement, but I don't know why this is true:

Statement Let $A$ be a symmetric $n$-by-$n$ matrix, and $B$ be a $n$-by-$n$ matrix, $A\geq 0$ is equivalent to $\text{Tr}(A^TB)\geq 0$, $\forall B$.

Update:

I am not sure in the above statement on whether $B$ is an arbitrary matrix. Actually, in the paper, the matrix $B$ takes values:

(1) $B=QQ^T$, where $Q\in\mathbb{R}_{n\times 2}$ and $\|Q_{i:}\|_2^2=1$, for $1\leq i\leq n$.

(2) $B=(u\circ q)(u \circ q)^T$, where $u$ is an arbitrary vector, and $q\in\{-1,+1\}^n$

$\endgroup$
12
  • 1
    $\begingroup$ Sorry, your statement is incomple! $\endgroup$ Jan 24, 2023 at 9:17
  • 2
    $\begingroup$ I assume you mean for the trace to be nonnegative and for B to also be psd. If so, this is just saying that the psd cone is self dual. math.stackexchange.com/questions/3358916/… $\endgroup$ Jan 24, 2023 at 9:28
  • 1
    $\begingroup$ You mean trace of $A^TB$ bigger than $0$ for every $B$? $\endgroup$ Jan 24, 2023 at 9:28
  • 3
    $\begingroup$ Note that the matrices of the form (2) are precisely the matrices of the form $xx^T$, in which case $\operatorname{Tr}(A^TB)=x^TAx$. Matrices of the form (1) contain all $xx^T$ with $x$ of norm $\sqrt{2}$ by taking the columns of $Q$ to be equal. (I'm assuming you mean $1\leq i \leq 2$.) The equivalence follows. $\endgroup$ Jan 24, 2023 at 10:03
  • 4
    $\begingroup$ Well, compiling the comments you have an answer. Taking $B = xx^T$ you obtain $x^TAx\geq 0$ and apply this to every eigenvector of $A$. Conversely, for any positive definite matrix $B$, write its Choleski decomposition $B=LL^T$ and use the fact that $A$ is psd for every column of $L$. $\endgroup$ Jan 24, 2023 at 12:42

0

Your Answer

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

Browse other questions tagged or ask your own question.