8
$\begingroup$

I have $A^{1/2} B A^{1/2} \preceq I$ for two PSD matrices $A$ and $B$, and I'd like to know if that implies $\|AB\|_2 \leq 1.$

The argument I was using to show this is that for any two square matrices $A$ and $B,$ it is always the case that $\|AB\|_2 = \|BA\|_2.$ I thought I read that this equality does hold in a reputable source, but I don't have access to it right now and I was unsuccessful in reproducing a proof.

I know the eigenvalues of $AB$ and $BA$ are the same modulo possibly having different numbers of zeros, so I'm worried that I might have ``remembered'' something that isn't true!

Any references/counterexamples for either question?

$\endgroup$
2
  • $\begingroup$ This is false in general. You can find $A, B$ such that $AB = 0$ but $BA \neq 0$ (although the result you cite shows that $BA$ is nilpotent). The spectral norm is defined in terms of singular values, not eigenvalues. $\endgroup$ Feb 22, 2013 at 7:27
  • $\begingroup$ Did I miss something or your assumptions imply that $BA$ is the transpose of $AB$ ? $\endgroup$ Jan 11, 2021 at 16:19

2 Answers 2

14
$\begingroup$

I suppose that your norm is either the Schur-Frobenius $\|F\|=({\rm Tr}FF^T)^{1/2}$ or that subordinated to the Euclidan norm $\|F\|=\rho(FF^T)$. In either case, the equality $\|AB\|=\|BA\|$ follows from the equality of the spectra of $ABBA$ and $BAAB$ (mind that $A$ and $B$ are symmetric). This is true because both matrices are similar to $A^2B^2$ (mind that $A$ and $B$ are non-singular).

If you drop your assumption that $A,B$ are PSD, then the equality does not hold in general. A simple counter-example is given by a pair $(A,B)$ such that $AB=0_n$ but $BA\neq 0_n$.

$\endgroup$
3
$\begingroup$

Going back to the original motivation, the following observations might still be useful.

  1. Denis already mentioned this: $A^{1/2}BA^{1/2} \sim A^{1/2}A^{1/2}BA^{1/2}A^{-1/2}$ for strictly positive $A$. This implies that $\sigma(A^{1/2}BA^{1/2})=\sigma(AB)$.

  2. If $A$ and $B$ are arbitrary (square) matrices such that their product $AB$ is normal, then one can show that $\sigma(AB) \prec_w \sigma(BA)$ (where $\prec_w$ denotes weak majorization); equivalently, for any unitarily invariant norm, we have $\|AB\| \le \|BA\|$.

  3. If $\|AB\| \le 1$ for positive definite $A$ and $B$, then $\|A^{1/2}B^{1/2}\| \le 1$.

$\endgroup$
3
  • $\begingroup$ In (1), the assumption that $A$ is strictly positive is unnecessary since strictly positive matrices are dense in all positive matrices. $\endgroup$ Feb 22, 2013 at 19:28
  • $\begingroup$ @Eric: yes indeed; I just wrote it so that it was ok to write $A^{-1/2}$ instead of $(A+\epsilon I)^{-1/2}$, and then let $\epsilon \to 0$ $\endgroup$
    – Suvrit
    Feb 22, 2013 at 19:32
  • $\begingroup$ @Suvrit: I was wondering how much can be said if $A$ is positive definite but $B$ is only symmetric. If $B$ is any symmetric matrix, using 1, I have that the eigenvalues of $A^{1/2}BA^{1/2}$ and $AB$ are equal. Now since $A^{1/2}BA^{1/2}$ is symmetric, its max absolute eigenvalue is its spectral norm. Can I then say that $\Vert AB\Vert \le \Vert A^{1/2}BA^{1/2}\Vert$? $\endgroup$ Feb 25, 2020 at 4:08

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.