5
$\begingroup$

Let $\mathbf A=\left[\begin{matrix}\mathbf A_{11}&\mathbf A_{12}\\ \mathbf A_{21}&\mathbf A_{22}\end{matrix}\right]$ be a positive semi-definite matrix, $\mathbf A_{ij}\in\mathbb C^{n\times n}$ and $rank(\mathbf A)=n$.

Prove or disprove that $$ \lambda_{max}(\mathbf A) + \lambda_{max}(\mathbf A_{11}+\mathbf A_{22}) \leq \lambda_{max}\left(\left[\begin{matrix}tr(\mathbf A_{11})&tr(\mathbf A_{12})\\ tr(\mathbf A_{21})&tr(\mathbf A_{22})\end{matrix}\right]\right) + tr(\mathbf A_{11}+\mathbf A_{22}), $$ where $\lambda_{max}$ denotes the largest eigenvalue and $tr$ denotes the trace.

I tried small values of $n$. For $n=1$ we have "$=$" (the same terms in the LHS and RHS). I have a rather long proof for $n=2$ (in my proof I also replaced $\mathbb C$ by $\mathbb R$).

It seems that some results on (completely) positive maps are somehow related to the inequality ( indeed, all the maps $$ \mathbf A\rightarrow \mathbf A_{11}+\mathbf A_{22},\qquad \mathbf A\rightarrow \left[\begin{matrix}tr(\mathbf A_{11})&tr(\mathbf A_{12})\\ tr(\mathbf A_{21})&tr(\mathbf A_{22})\end{matrix}\right],\qquad \mathbf A\rightarrow tr(\mathbf A_{11}+\mathbf A_{22}) $$ are positive). By Theorem 2.3.7 (The Russo-Dye Theorem) from the book R. Bhatia, Positive definite matrices, $$ \lambda_{max}\left(\left[\begin{matrix}tr(\mathbf A_{11})&tr(\mathbf A_{12})\\ tr(\mathbf A_{21})&tr(\mathbf A_{22})\end{matrix}\right]\right)\leq n\lambda_{max}(\mathbf A), $$ which is useless since we need inequality of the type $\lambda_{max}(\mathbf A)\leq\dots$

$\endgroup$
3
  • $\begingroup$ If the off diagonal blocks are also postive definite then the claims holds "easily". Have to still think about the general case. $\endgroup$
    – Suvrit
    Sep 5, 2016 at 21:40
  • $\begingroup$ @Suvrit In my application the off diagonal blocks are not positive definite (even not symmetric/hermitian). It would be nice to see your "easy" solution for this particular case because I guess that a more general inequality holds (I will update the question soon) probably your "easy" can be valid for it. $\endgroup$
    – Ignat
    Sep 6, 2016 at 9:48
  • $\begingroup$ The "easy" solution was based on proving a stronger inequality that holds when everything is PSD, and by noting that $\|A\| \le \left\Vert \begin{bmatrix}\|A_{11}\| & \|A_{12}\|\\ \|A_{21}\| & \|A_{22}\|\end{bmatrix}\right\Vert$. Since in the meanwhile you already have an answer to the original question, this special case has been obviated. $\endgroup$
    – Suvrit
    Sep 6, 2016 at 18:28

1 Answer 1

3
+50
$\begingroup$

Write $A=\sum w_i\otimes w_i$ where $w_i$ are orthogonal and $\|w_1\|=\max_i\|w_i\|$. Also write $w_i=(u_i,v_i)$. Then the inequality in question is just $$ \|u_1\|^2+\|v_1\|^2+\max_{e:\|e\|=1}\sum_i[|(u_i,e)|^2+|(v_i,e)|^2] \\ \le \left\|\begin{pmatrix}\sum_i\|u_i\|^2&\sum_i(u_i,v_i)\\\sum_i(v_i,u_i)&\sum_i\|v_i\|^2\end{pmatrix}\right\|+\sum_i(\|u_i\|^2+\|v_i\|^2) $$ Now, cancel the first term with the corresponding term on the RHS and estimate $ \sum_{i\ge 2}[|(u_i,e)|^2+|(v_i,e)|^2] $ by $ \sum_{i\ge 2}(\|u_i\|^2+\|v_i\|^2) $. Also, note that the norm of the sum of several positive definite matrices is not less than the norm of each of them, so we can remove all $i\ge 2$ in the matrix and reduce the problem to showing that $$ |(u_1,e)|^2+|(v_1,e)|^2\le \left\|\begin{pmatrix}\|u_1\|^2&(u_1,v_1)\\(v_1,u_1)&\|v_1\|^2\end{pmatrix}\right\| $$ However, we can write the matrix on the right as the sum of positive definite matrices of the kind $\begin{pmatrix}|(u_1,e_j)|^2& (u_1,e_j)(e_j,v_1)\\(v_1,e_j)(e_j,u_1)&|(v_1,e_j)|^2\end{pmatrix}$ where $e_j$ is an orthonormal basis with $e_1=e$. Again, we can remove all matrices with $j\ge 2$, after which we get an identity.

$\endgroup$
4
  • $\begingroup$ Thank you for your nice solution. If am not mistaken, you derivation also implies that if "=" holds, then $A_{11}$ is a rank-$1$ matrix. I think the last inequality is just $|(u_1,e)|^2 + |(v_1,e)|^2\leq \|u_1\|^2+\|v_1\|^2=\|\text{matrix}\|$ (since the matrix is rank-1, its norm equals to trace). I will also try to extend your proof for $k\times k$ block matrices. $\endgroup$
    – Ignat
    Sep 6, 2016 at 9:57
  • $\begingroup$ @Ignat Not quite. It becomes rank 1 only when you replace norms by scalar products with $e$. Before the very last step it is not necessarily rank $1$ because Cauchy inequality is strict more often than not. $\endgroup$
    – fedja
    Sep 6, 2016 at 12:25
  • $\begingroup$ the inequality has an interesting application for tensor decompositions. I am preparing a short paper and I would like to mention you name in the acknowledgment part. Could you please send me your name/affilation at ignat dot domanov at kuleuven dot be? Or I can try to mention a link to this page if it will be allowed by journal's policy. $\endgroup$
    – Ignat
    Sep 15, 2016 at 18:54
  • $\begingroup$ @Ignat Just cite MO :-) $\endgroup$
    – fedja
    Sep 15, 2016 at 22:16

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.