The original question goes equivalently like this

For real matrices, if $A$ and $B$ are both positive-definite
Prove: all the eigenvalues of $AB$ are positive.

Facts that I know may have a role to play in the proof:

1) For real symmetrical matrices, $A$ is positive-definite $\Leftrightarrow$ all eigenvalues of $A$ are positive.

2) Any real positive-definite matrix $A$ can be diagonalized as $$A=Q^T\Lambda Q$$ where $\Lambda=\text{diag}\{\lambda_1,\lambda_2, \cdots,\lambda_n\}$ is the diagonal matrix comprising all the eigenvalues of $A$.

3).Any real positive-definite matrix $A$ can be decomposed as $$A=P^TP$$ where $P$ is also a real positive-definite matrix that shares the same size with $A$.

However, I failed to make any progress after trying them all. Could you please help me with the proof or drop me a hint? Best regards!

EDIT I'd be specially grateful if you could explain it in a way that a student who has only taken an elementary linear algebra can understand. :) By saying "elementary", I mean I have only learnt real matrices and know very little about complex ones, to the point that I don't even really know what Hermitte is (⊙o⊙)


Hint. Let $B^{1/2}$ be the positive definite square root of $B$. We have $AB=B^{-1/2}(B^{1/2}AB^{1/2})B^{1/2}$. Use matrix congruence(with Sylvester's law of inertia) and matrix similarity to finish the argument.

Note that the title and the body of your question do not match. While $AB$ always possesses a full positive spectrum when $A$ and $B$ are positive definite, $AB$ is not necessarily positive definite. For a counterexample, consider $A=\pmatrix{1\\ &3}$ and $B=\pmatrix{1&1\\ 1&1+\epsilon}$ for any small $\epsilon>0$.

  • 1
    $\begingroup$ @user1551 Now the matrices are not transpose congruent. From your expression it is clear that $AB$ is similar to the positive definite matrix $B^{1/2}AB^{1/2}$ via an invertible matrix. So they must have same eigenvalues. We do not need Sylvester's law of inertia here. $\endgroup$ – mrka Mar 25 '15 at 11:24
  • 1
    $\begingroup$ @mrka I don't understand your comment that "now the matrices (which matrices?) are not transpose congruent". I did use matrix congruence here: $AB$ is similar to $B^{1/2}AB^{1/2}$, and $B^{1/2}AB^{1/2}$ is congruent to $A$. Sylvester's law of inertia may be an overkill, but the fact that $B^{1/2}AB^{1/2}$ is positive definite IS owing to the fact that matrix congruence preserves positive definiteness. $\endgroup$ – user1551 Mar 25 '15 at 11:31
  • 1
    $\begingroup$ @Babai That's because $x^TB^{1/2}AB^{1/2}x>0$ for every nonzero $x$. Put $y=B^{1/2}x$ if you still don't see it. $\endgroup$ – user1551 Aug 8 '16 at 19:24
  • 1
    $\begingroup$ @Babai That's basic linear algebra. In general, if $X,Y$ are two rectangular matrices such that the products $XY$ and $YX$ make sense, then $XY$ has the same set of nonzero eigenvalues as $YX$. In particular, if both $X$ and $Y$ are square, then $XY$ and $YX$ have identical spectra. Therefore $AB$ and $B^{1/2}AB^{1/2}$ have identical eigenvalues. These eigenvalues must be real because $B^{1/2}AB^{1/2}$ is positive semidefinite. $\endgroup$ – user1551 Aug 8 '16 at 19:46
  • 1
    $\begingroup$ @Babai Surely you are correct. I didn't notice that you have changed the conditions for $A$. $\endgroup$ – user1551 Aug 8 '16 at 19:59

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.