15
$\begingroup$

I wonder whether the following property holds true: For every real symmetric matrix $S$, which is positive in both senses: $$\forall x\in{\mathbb R}^n,\,x^TSx\ge0,\qquad\forall 1\le i,j\le n,\,s_{ij}\ge0,$$ then $\sqrt S$ (the unique square root among positive semi-definite symmetric matrices) is positive in both senses too. In other words, it is entrywise non-negative.

At least, this is true if $n=2$. By continuity of $S\mapsto\sqrt S$, we may assume that $S$ is positive definite. Denoting $$\sqrt S=\begin{pmatrix} a & b \\ b & c \end{pmatrix},$$ we do have $a,c>0$. Because $s_{12}=b(a+c)$ is $\ge0$, we infer $b\ge0$.

$\endgroup$

2 Answers 2

26
$\begingroup$

No. If $$A = \begin{pmatrix}10&-1&5\\-1&10&5\\5&5&10\end{pmatrix},$$ then $A$ is positive definite but does not have all entries positive, while $$ A^2 = \begin{pmatrix}126&5&95\\5&126&95\\95&95&150\end{pmatrix} $$ is positive in both senses.

$\endgroup$
31
$\begingroup$

Robert Bryant already showed via an example that the answer is "no". To come up with lots of counterexamples, recall that (under some mild assumptions) if $A$ has maximal eigenvalue $\lambda_{\text{max}}$ and corresponding unit eigenvector $\mathbf{v}$ then $(A/\lambda_{\text{max}})^k \rightarrow \mathbf{v}\mathbf{v}^*$ as $k \rightarrow \infty$.

So if you pick any matrix that is (a) positive semidefinite with a negative entry, and (b) has a unique maximal eigenvalue with corresponding entrywise positive eigenvector, then repeatedly squaring $A$ will eventually give a counterexample to the original question.

$\endgroup$
1
  • 2
    $\begingroup$ I like this conceptual analysis; it's what I had in mind when I constructed the example that I did. In order to really complete this as an argument, though, one needs to explain that, while there is no symmetric $n$-by-$n$ matrix $A$ that satisfies both conditions (a) and (b) when $n=2$, there is such a matrix when $n=3$ (and hence, when $n\ge 3$). Of course, once one has produced a counterexample when $n=3$, there's no need for the general argument. $\endgroup$ Oct 19, 2020 at 11:48

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.