Let $A$ and $B$ be $n\times n$ Hermitian positive definite matrices and $k>0$ real. Then $A^k$ is well-defined and experimentally, we have $$\det(A^k+BABA^{-1})\geqslant \det(A^k+BA^{-1}BA),$$or equivalently $$\det(A^k+A^{-1}BAB)\geqslant \det(A^k+ABA^{-1}B).$$ Note that this would imply the inequality $\color{green}{\det (A^2+BAAB)\geqslant \det(A^2+ABBA)}$ found (but presumably unproved) by M. Lin, which is mentioned in a comment here.

One striking fact is that $B$ (or $A$, same outcome) can be multiplied by a positive factor, which means that the inequality remains valid for any "proportionality" between the two terms of each sum. Indeed, for any $\lambda>0$, we have equivalently $$\det(A^k+\color{red}\lambda BABA^{-1})\geqslant \det(A^k+\color{red}\lambda BA^{-1}BA).$$

Any ideas how to prove this conjecture?

  • $\begingroup$ I have some other generalizations based on your very interesting determinantal conjectures --- I wish I had time to think about these. Do you have numerical evidence for these? The second one suggests that perhaps there is even a Löwner dominance somewhere that yields the ineq for all $\lambda\ge 0$. $\endgroup$ – Suvrit Jun 23 '17 at 11:06
  • $\begingroup$ @Suvrit Yes I wouldn't have posted that without checking first for several thousands of random matrices of different sizes and several big and small values of $k$. I almost thought my few lines of PARI code had a bug... :-) $\endgroup$ – Wolfgang Jun 23 '17 at 11:14
  • $\begingroup$ have you tried Tao's strategy: proving that the spectrum of $C=A^{-(k+1)/2}BABA^{-(k+1)/2}$ is majorated by the spectrum of $D=A^{-(k-1)/2}BA^{-1}BA^{-(k-1)/2}$? $\endgroup$ – Fedor Petrov Jun 23 '17 at 11:45
  • $\begingroup$ @FedorPetrov That looks like an appealing idea, but I already have a hard time following how he re-arranges the traces. And no idea how to use that method with inverses, let alone arbitrary powers, even the k=2 case. $\endgroup$ – Wolfgang Jun 23 '17 at 12:12
  • 1
    $\begingroup$ Sorry to have confused you, I had known this $\det (A^2+BAAB)\geqslant \det(A^2+ABBA)$. It follows from COROLLARY 2.3 of the paper M. Lin, H. Wolkowicz, An eigenvalue majorization inequality for positive semidefinite block matrices, Linear Multilinear Algebra 60 (2012), no. 11-12, 1365–1368. tandfonline.com/doi/abs/10.1080/03081087.2011.651723 $\endgroup$ – M. Lin Jun 25 '17 at 0:00

Let $C = A^{-(k+1)/2} B A B A^{-(k+1)/2}$ and $D = A^{-(k-1)/2} B A^{-1} B A^{-(k-1)/2}$ .

We have to show that $det(I+C) \ge det(I+D)$ .

Now my goal is to apply equation (5.21) in Ando, Majorizations and Inequalities in Matrix Theory, http://ac.els-cdn.com/0024379594903417/1-s2.0-0024379594903417-main.pdf?_tid=1af72ca2-58c4-11e7-a518-00000aab0f01&acdnat=1498298674_bf9275307496eb46505ba4b5b84a7dcc :

Choose $E = A^{-(k+1)/2} B A^{-(k+1)/2}$ and $F = A^{k/2+1}$.

Then $C^{1/2} = \vert F E \vert$ and $D^{1/2} = \vert F^{a_1} E^{b_1} F^{a_2} E^{b_2} \vert$ , where $a_1 = (k/2)/(k/2+1)$, $b_1 = 1$, $a_2 = 1/(k/2+1)$, $b_2 = 0$.

Since $0 \le a_1 \le b_1$ and $0 \le a_2$ and $0 \le b_2$ and $a_1 + a_2 = b_1 + b_2 = 1$ it follows that $D^{1/2} \prec_{\log} C^{1/2}$ and therefore $D \prec_{\log} C$.

Since $log(1+e^x)$ is convex in $x$ it follows that $I+C$ weakly log majorizes $I+D$ and therefore $det(I+C) \ge det(I+D)$ .

| cite | improve this answer | |
  • $\begingroup$ Shouldn't it be (k+1)/2 in the definition of E? $\endgroup$ – Wolfgang Jun 24 '17 at 14:40
  • $\begingroup$ @Wolfgang : Yes, corrected. $\endgroup$ – jjcale Jun 24 '17 at 17:00
  • $\begingroup$ @jjcale, would you try Conjecture 4.2 in the paper worldscientific.com/doi/abs/10.1142/S0219199716500449 $\endgroup$ – M. Lin Jun 25 '17 at 0:09
  • $\begingroup$ @M.Lin, can you state the conjecture here, please? $\endgroup$ – Wolfgang Jun 25 '17 at 17:26
  • $\begingroup$ Here is the link to the arXiv preprint arxiv.org/pdf/1604.04141v1.pdf $\endgroup$ – M. Lin Jun 26 '17 at 5:02

EDIT (added some clarifications). The argument below provides a self-contained proof.

Introduce the shorthand $C^{-2}=A^{k+1}$. We need to show that \begin{equation*} \det(I+ CBABC) \ge \det(I + CABA^{-1}BAC). \end{equation*}

We will prove this inequality by establishing the log-majorization \begin{equation*} \prod\nolimits_{j=1}^k\lambda_j(CABA^{-1}BAC) \le \prod\nolimits_{j=1}^k\lambda_j(CBABC),\quad 1\le j \le n. \end{equation*} This log-majorization is equivalent to $\lambda_1(\wedge^k (CABA^{-1}BAC)) \le \lambda_1(\wedge^k(CBABC))$ (for $1\le k\le n$). It suffices to show this for $k=1$; the general case follows similarly upon exploiting the multiplicativity of the wedge product.

Thus, to prove $\lambda_1(CABA^{-1}BAC) \le \lambda_1(CBABC)$, we use its scale independence and observe that for positive matrices $X$ and $Y$, we have \begin{equation*} [Y \le I \implies X \le I] \implies \lambda_1(X)\le \lambda_1(Y). \end{equation*}

Thus, to prove the $\lambda_1$ inequality, it suffices to prove that \begin{equation*} CBABC \le I \Leftrightarrow\ \begin{bmatrix} A^{k+1} & B\\ B & A^{-1} \end{bmatrix} \ge 0\quad\implies CABA^{-1}BAC \le I. \end{equation*} But $\begin{bmatrix} A^{k+1} & B\\ B & A^{-1} \end{bmatrix} \ge 0$ only if $B\le \sqrt{A^{k+1}A^{-1}}=A^{k/2}$; and if this is so, then it must also be the case that $\begin{bmatrix} A^{k-1} & B\\ B & A \end{bmatrix} \ge 0$. Notice now that $A^{k-1} = C^{-2}A^{-2}$, and apply Schur complements to the latter matrix inequality to obtain \begin{equation*} BA^{-1}B \le C^{-2}A^{-2}\Leftrightarrow CABA^{-1}BAC \le I. \end{equation*} Thus, we have shown that $\lambda_1(CABA^{-1}BAC) \le \lambda_1(CBABC)$. In a similar manner we can prove the general case for $k>1$, which ends up establishing the desired log-majorization.

This finishes the proof of the claim because $\lambda(X) \prec_{\log} \lambda(Y) \implies \det(I+X) \le \det(I+Y)$.

| cite | improve this answer | |
  • 1
    $\begingroup$ How can you see that $B \le A^{k/2}$ ? $B$ is positive definite but otherwise arbitrary . $\endgroup$ – jjcale Jun 24 '17 at 6:33
  • $\begingroup$ Yes, how can you apply your first lemma if $Y\le I$ doesn't hold? $\endgroup$ – Wolfgang Jun 24 '17 at 7:09
  • $\begingroup$ @jjcale Because it's a conditional statement; if we have $CBABC\le I$, then we have $B\le A^{k/2}$. What we really wanted to prove was the eigenvalue domination; we reduced that to showing that $Y\le I \implies X \le I$ must hold (note that $\lambda(\cdot)$ is positively homogeneous). $\endgroup$ – Suvrit Jun 24 '17 at 12:10
  • $\begingroup$ Actually, I should be a bit more precise, and just write that what I am using is the observation: $(0 \le Y \le I \implies 0 \le X \le I) \implies \lambda_1(X)\le \lambda_1(Y)$ $\endgroup$ – Suvrit Jun 24 '17 at 12:14
  • $\begingroup$ So you mean your proof only works if CBABC≤I? Well, granted, but then that's a very special case only. $\endgroup$ – Wolfgang Jun 24 '17 at 12:55

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.