6
$\begingroup$

First, in this discussion, I am only considering real matrices. Second, I have a few questions I am ruminating on related to symmetric matrices. Some of these questions I need someone to say my logic is correct while others I need you to help provide the logic.

  1. I am curious as to what the closure of the set of positive definite matrices is. I know that this set is an open cone. Just thinking about it seems to point to the closure being the positive semi-definite matrices. However, I cannot think of a way to prove this. Also, is the closure in the set of symmetric matrices, the same as the closure in the set of all matrices. I believe this to be the same as any sequence of symmetric matrices should be a symmetric matrix.

  2. Also, is every symmetric matrix, either positive definite, negative definite or in the boundary of both? Again this seems somewhat logical but I don't know why.

  3. Finally, if the closure of the positive definite matrices is the positive semi-definite ones, do all the matrices in the boundary have determinant zero? This seems true as I believe these matrices have at least one eigenvalue as $0$ implying $0=det(A-\lambda I)= det (A-0I) =det (A)$. If I was wrong about the closure is there any way to characterize the matrices in the boundary?

Thanks!

$\endgroup$
2
$\begingroup$
  1. Yes. Suppose $\{A_n\}$ is a family of PD matrices and $A_n\to A$. Then $0=A_n-A_n^\ast\to A-A^\ast$ (hence $A$ is Hermitian) and for any vector $x$, we have $x^\ast A_nx>0$ and $x^\ast A_nx\to x^\ast Ax$ (hence $x^\ast Ax\ge0$). It follows that $A$ is PSD. Conversely, if $A$ is PSD, then $A_n=A+\frac1n I$ is PD and $A_n\to A$.
  2. No. $\operatorname{diag}(-1,1)$ is the simplest counterexample.
  3. Yes. The set of PD matrices is open (this can be proved by considering $x^\ast(A+\delta A)x$). So, every matrix on the boundary of PD matrices is a PSD matrix with at least one zero eigenvalue and therfore it has zero determinant.
$\endgroup$
3
$\begingroup$

1 and 3 are correct. Remember that the symmetric matrices make a vector subspace.

What you are missing is the indefinite quadratic forms, such as $x^2 - y^2.$ This corresponds to a diagonal matrix with diagonal entries $(1,-1).$ It is not in the boundary of anything you have mentioned.

As the dimension $n$ goes up in symmetric $n$ by $n$ matrices, you get more and more separate connected components, according to signature: http://en.wikipedia.org/wiki/Sylvester%27s_law_of_inertia

$\endgroup$
  • $\begingroup$ Why is the first part of #1 true though? I can't think of a proof for it. $\endgroup$ – Leo Spencer Oct 30 '13 at 23:11
  • 1
    $\begingroup$ @Leo, here is a book to start. books.google.com/… $\endgroup$ – Will Jagy Oct 30 '13 at 23:17

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.