10
$\begingroup$

For positive semidefinite matrices $A,B,C \in \mathbb{R}^{n\times n}$, the following inequalities are well known:

$$(\det(A+B))^{1/n} \geq (\det A)^{1/n} + (\det B)^{1/n} $$

and

$$\det(A+B+C) + \det(C) \geq \det(A+C) + \det(B+C).$$

The first is, of course, Minkowski's determinant inequality. I'm not sure whether the second has a name, but it can be found, e.g., here and here.

I would like to know whether anyone knows of any reverse forms of these inequalities. To be more specific, I note that the Minkowski determinant inequality can be viewed as a consequence of the Entropy Power Inequality applied to Gaussian random vectors with covariance matrices $A$ and $B$. Applying similar reasoning to the entropy power inequality appearing in Theorem 4 of this paper, we may conclude that the following reverse counterparts to the above inequalities hold:

$$(\det(A+B))^{1/n} \leq (\det A)^{1/n}\left(\frac{\frac{1}{n}\operatorname{Tr}(B^{-1})}{(\det B^{-1})^{1/n}} \right) + (\det B)^{1/n}\left(\frac{\frac{1}{n}\operatorname{Tr}(A^{-1})}{(\det A^{-1})^{1/n}} \right),~~~~~(\star) $$

and

$$(\det(A+B+C) \det(C) )^{1/n} + (\det(A) \det(B) )^{1/n} \leq (\det(A+C) \det(B+C) )^{1/n} .~~~(\star\star)$$

The first may be obtained from the second by taking $C = \varepsilon I$ and letting $\varepsilon \downarrow 0$. I view $(\star)$ as a reverse counterpart to Minkowski's determinant inequality, since the ratio

$$\frac{\frac{1}{n}\operatorname{Tr}(B^{-1})}{(\det B^{-1})^{1/n}}\geq 1$$

by the AM-GM inequality applied to the eigenvalues of $B^{-1}$ (and similarly for the ratio involving $A^{-1}$), with near equality if the matrix $B$ has eigenvalues all roughly the same.

Has anyone seen these inequalities before, or does anyone have a direct proof of $(\star\star)$?

$\endgroup$
7
$\begingroup$

Inequality ($\star\star$) essentially follows from the original Minkowski plus an implication of Lidkskii's inequality (Fiedler's inequality, noted below). $\newcommand{\da}{\downarrow} \newcommand{\ua}{\uparrow}$

Assume $C$ is strictly positive definite (otherwise, $\det C=0$ rendering ($\star\star$) trivial), and let $C^{1/2}$ denote the positive square root of $C$. Renaming the matrices $A \leftarrow C^{-1/2}A C^{-1/2}$ and $B \leftarrow C^{-1/2}B C^{-1/2}$, the original inequality reduces to showing that \begin{equation*} \det(A+B+I)^{1/n} + [\det A\det B]^{1/n} \le [\det(I+A)\det(I+B)]^{1/n}. \end{equation*}

Let $a^\da$ denote the eigenvalues of $A$ in decreasing order; similarly, define $b^\ua$ for $B$. Then, we have \begin{equation*} \det(I+A)\det(I+B) = \prod_i(1+a_i^\da)\prod_i(1+b_i^\ua) = \prod_i(1+a_i^\da + b_i^\ua + a_i^\da b_i^\ua), \end{equation*} whence by the Minkowski inequality we get \begin{equation*} [\det(I+A)\det(I+B)]^{1/n} \ge \prod_i(1+a_i^\da+b_i^\ua)^{1/n} + \prod_i(a_i^\da b_i^\ua)^{1/n}. \end{equation*} Now from Fiedler's (1971) determinantal inequality we know that \begin{equation*} \prod_i(1+a_i^\da+b_i^\ua) \ge \det(I+A+B), \end{equation*} while clearly $\prod_i(a_i^\da b_i^\ua) = \det(A)\det(B)$. Thus, the inequality ($\star\star$) follows.

| cite | improve this answer | |
$\endgroup$
  • 1
    $\begingroup$ If $Q^{-1}CQ=I$ then $C=I$. What do you mean?? You have shown (**) for $C=I$, but I don't see how you can get the general case from that, given that the product of two positive definite matrices is not necessarily positive definite. $\endgroup$ – Wolfgang Oct 9 '16 at 19:11
  • $\begingroup$ I think it works by just considering positive square roots: $C = C^{1/2}C^{1/2}$. Then, the argument goes through as Suvrit suggested by considering $\tilde{A} = C^{-1/2}AC^{-1/2}$, $\tilde{B} = C^{-1/2}BC^{-1/2}$, and $\tilde{C}=I$. $\endgroup$ – Tom Oct 9 '16 at 19:59
  • $\begingroup$ @Suvrit, am curious as to why you say ($*$) is not a reverse Minkowski inequality (does such a thing have a precise definition?). That ($*$) follows from Minkowski (plus Fiedler's, plus linearizing ($**$)) should not preclude such a thing. E.g., 'reverse Holder' follows from Holder's inequality. In fact, ($*$) actually implies Minkowski's inequality, so could be regarded as an improvement. Indeed, Minkowski is equivalent to concavity of $A\mapsto (\det A)^{1/n}$, or $f(t)\leq f(0) + t f'(0)$ for $t$ small, and $f(t) := (\det(A+t I))^{1/n}$. But, this is precisely ($*$) with $B=tI$. $\endgroup$ – Tom Oct 9 '16 at 23:10
  • $\begingroup$ @Tom sorry, I meant to say $(**)$ is not a reverse Minkowski inequality (because to prove it I had to use Minkowski), but you never claimed that anyways. I'll edit out that part. $\endgroup$ – Suvrit Oct 10 '16 at 0:06
  • 1
    $\begingroup$ @Tom (1st comment) yes that completes Suvrit's argument, noting that $\tilde{A} $ and $\tilde{B} $ are psd. $\endgroup$ – Wolfgang Oct 10 '16 at 11:47

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.