6
$\begingroup$

Playing around with Matlab I noticed something very peculiar:

Take the symmetric matrix $A \in \mathbb R^{n \times n}$ defined by

$$A_{ij}= i \delta_{ij} - \frac{\varepsilon}{\sqrt{i}\sqrt{j}}\,.$$

Here $\delta_{ij}$ is the Kronecker delta.

We first note that this matrix is not diagonally dominant if $n$ is large enough.

This is because $\lim_{n \rightarrow \infty} \sum_{i=1}^{n} \vert A_{i,1}\vert=\infty >\vert A_{1,1} \vert.$

It is obvious that we require $\varepsilon<1$ in order for $A$ to be positive definite, since otherwise $A_{1,1}\le 0.$

However, I noticed that for let's say $\varepsilon=0.1$ one can make the dimension as large as one wants and the matrix remains positive definite.

Question: How can one show that $A$ is positive definite independent of the dimension if $\varepsilon$ is sufficiently small but fixed ?

$\endgroup$
4
  • $\begingroup$ It looks like a consequence of Gershgorin disks theorem. But precise computations have to be done. $\endgroup$ Jan 13, 2020 at 13:37
  • $\begingroup$ Hi Sascha, does your transition occur at $\epsilon = 6/\pi^2$? $\endgroup$
    – RaphaelB4
    Jan 13, 2020 at 14:08
  • $\begingroup$ @RaphaelB4 it is hard to say numerically to be honest. $\endgroup$
    – Sascha
    Jan 13, 2020 at 15:46
  • 2
    $\begingroup$ @JeanMarieBecker mhmm, not sure. I think Gershgorin's disk theorem is somewhat equivalent to being diagonally dominant for positive definite matrices. $\endgroup$
    – Sascha
    Jan 13, 2020 at 15:46

1 Answer 1

7
$\begingroup$

The claim is true with $\epsilon=\frac6{\pi^2}\,$.

To see this, remark that by changing variable $x_i=y_i\sqrt i\,$, this is equivalent to proving that $$\epsilon\left(\left(\frac1{ij}\right)\right)_{1\le i,j}\le I_\infty.$$ The first (infinite) matrix is $V\otimes V$ with $V=(1,\frac12\,,\ldots,\frac1n\,,\ldots)$. It is symmetric, rank-one, with eigenvalues $0$ (infinitely many times) and ${\rm Tr}(V\otimes V)=\frac{\pi^2}6$ (simple).

$\endgroup$
3
  • $\begingroup$ what a nice argument this is. Just out of curiosity. Is it also obvious to see that the random signs do not matter? Cause this would somehow break the multiplicative structure. $\endgroup$
    – Sascha
    Jan 13, 2020 at 16:08
  • $\begingroup$ @Sascha. This is not the same situation, because then the corresponding matrix is not any more rank-one. $\endgroup$ Jan 13, 2020 at 16:11
  • $\begingroup$ I agree, let me just accept this wonderful answer and ask a follow-up question. $\endgroup$
    – Sascha
    Jan 13, 2020 at 16:54

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.