Recently I've asked question about the set given as $\{x: x^T(A-B)x = 1\}$. In that question I wanted to prove that such set isn't bounded, and according to the answer I can do it if I can prove the fact from the title of this question i.e.:

If $A$ and $B$ are different, symmetric, positive definite $n \times n$ $(n > 1)$ matrices such that $\det A = \det B$ then the matrix $A - B$ is neither positive semidefinite nor negative definite

Could you please help me to start the proof? I have no idea how to do it (namely how to use the equality of determinants)


Audience request: we can begin with orthogonal $Q$ such that $Q^TB Q = D$ is diagonal with positive diagonal elements. We can then continue with a diagonal matrix $W$ with diagonal elements certain reciprocals of square roots, so that $W Q^T BQW = W^T Q^T B Q W = I,$ and name P = QW.$

there is a matrix $P$ with $\det P \neq 0$ so that $P^T B P = I.$ Then $$ P^T ( A - B) P = C - I, $$ where $$ C = P^T A P $$ is symmetric, positive definite and, since $\det B \cdot \det^2 P = 1,$ so $\det A \cdot \det^2 P = 1,$ we get $$ \det C = 1.$$

Let's see, if all eigenvalues of $C$ are equal to $1$ then it is the identity matrix; that needs a little proof using symmetry and orthogonal matrices. In this case, $A = B$ and $x^T (A-B)x = 1$ is impossible.

Otherwise, $C$ has an eigenvalue larger than $1$ and another eigenvalue smaller than $1,$ which tells us that $C-I$ is indefinite.

You should look up Sylvester's Law of Inertia, which has to do with the "congruence" I am using.

  • $\begingroup$ Why $\exists \ P: P^TBP = I$? I thought there should be a diagonal matrix, but not necessary an identity... $\endgroup$ – D F Feb 12 '18 at 19:36
  • $\begingroup$ Wow! Extremely beautiful solution. Thanks a lot! $\endgroup$ – D F Feb 12 '18 at 19:45

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.