3
$\begingroup$

I am given real, symmetric matrices $X \succ B \succ A \succ 0$ (where '$\succ$' signifies positive definiteness such that if $B \succ A$ then $B-A \succ 0$ is positive definite). Further let the '$\preceq$' operator signify negative semi-definiteness (and $\succeq$ signify positive semi-definiteness). Does the following hold? \begin{align} (X + B)^{-1} - (X + B + A)^{-1} + (2B + 2A)^{-1} - (2B)^{-1} \preceq 0 \end{align} A few observations:

  1. As the inverse is operator decreasing I know that: \begin{align} (2B + 2A)^{-1} - (2B)^{-1} \preceq 0 \;\;\;\;\;\;\mbox{and}\;\;\;\;\;\; (X + B)^{-1} - (X + B + A)^{-1} \succeq 0. \end{align}

  2. If $A = I$, $B = 2I$, $X = 3I$, where $I$ is the identity matrix, the above inequality holds.

  3. I believe the derivative of the matrix inverse should play a role here. Essentially, how does the addition of $A$ (or $2A$) affect the negative semi-definiteness of the above expression when (i) adding $A$ to $X+B$ (in the inverse) makes the expression more positive semi-definite and (ii) adding $2A$ to $2B$ (in the inverse) makes the expression more negative semi-definite. If the matrix inverse is somehow 'negative', then adding $2A$ should make the expression more negative semi-definite than adding $A$.

Thank you very much for any insights.

$\endgroup$
1
$\begingroup$

It's very easy to generate a random counterexample by computer, e.g. $$ X=I,\ B=\pmatrix{0.9&0\\ 0&0.1},\ A=\pmatrix{0.7&0.07\\ 0.07&0.01}. $$ Numerical computation shows that the eigenvalues of $C=(X + B)^{-1} - (X + B + A)^{-1} + (2B + 2A)^{-1} - (2B)^{-1}$ are approximately $0.007$ and $-0.418$. Therefore $C$ is indefinite.

$\endgroup$
  • $\begingroup$ But A is not positive definite, the eigenvalues are roughly -0.065 and 0.6165. $\endgroup$ – OttoVonBismarck Feb 11 '15 at 16:51
  • $\begingroup$ @OttoVonBismarck I didn't notice the requirement that $A\succ0$ before. It's fixed now. $\endgroup$ – user1551 Feb 11 '15 at 18:01
  • $\begingroup$ Alas, this does break the inequality. I still wish I had more insight! Do you know if there's an example where the inequality also doesn't hold when $X \succ B \succ A \succ I$, where $I$ is the identity matrix? $\endgroup$ – OttoVonBismarck Feb 11 '15 at 21:10
  • $\begingroup$ @OttoVonBismarck If you scale $X,B$ and $A$ by the same factor $\alpha>0$, then when $\alpha$ is sufficiently large, $A$ will eventually be $\succ I$. Since the new $C$ is just $\frac1\alpha$ times the old $C$, it is still indefinite and your inequality still doesn't hold. $\endgroup$ – user1551 Feb 12 '15 at 1:18
  • $\begingroup$ Ah yes yes, good. Thank you very much! $\endgroup$ – OttoVonBismarck Feb 12 '15 at 22:00

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.