5
$\begingroup$

Let $ \begin{bmatrix} A& B \\\\ B^* &C \end{bmatrix}$ be positive semidefinite, $A,C$ are of size $n\times n$.

Are the following plausible inequalities true? I have seen a lot of similar results, but for the following inequalities, I cannot locate them in the literature or find that they have been pointed out to be false.

1 $$\quad \sum\limits_{i=1}^k\lambda_i\begin{bmatrix} A& B \\\\ B^* &C \end{bmatrix}\le \sum\limits_{i=1}^k\left(\lambda_i(A)+\lambda_i(C)\right)\quad, $$ where $1\le k\le n$.

2 Modified $$\quad \prod\limits_{i=1}^{2k}\lambda_i\begin{bmatrix} A& B \\\\ B^* &C \end{bmatrix}\le \prod\limits_{i=1}^k \lambda_i(A)\lambda_i(C) \quad, $$ where $1\le k\le n$.

3 $$ 2\lambda_i^{1/2}(B^*B)\le \lambda_i(A+C),$$ where $1\le k\le n$.

Here, $\lambda_i(\cdot)$ means the $i$th largest eigenvalue of $\cdot\quad$. Any references or counterexamples are appreciated.

$\endgroup$
2
  • $\begingroup$ it seems to me that your first majorization is reversed?? $\endgroup$
    – Suvrit
    Jul 19, 2011 at 0:26
  • $\begingroup$ Probably you are saying $\sum\limits_{i=1}^k\lambda_i\begin{bmatrix} A&0 \\0 &C \end{bmatrix}\le \sum\limits_{i=1}^k\lambda_i\begin{bmatrix} A& B \\ B^* &C \end{bmatrix}$, which is true. $\endgroup$
    – Sunni
    Jul 19, 2011 at 0:40

1 Answer 1

18
$\begingroup$

Item 1 is true. This is part of Problem 22 (b) in Section 3.5 of Horn and Johnson [HJ94], which states that for Ky Fan norm ||⋅|| (and in fact for any unitarily invariant norm) and a positive semidefinite block matrix $\begin{pmatrix}A & B \\ B^* & C\end{pmatrix}$, it holds that $\left\|\begin{pmatrix}A & B \\ B^* & C\end{pmatrix}\right\| \le \left\|\begin{pmatrix}A & 0 \\ 0 & 0\end{pmatrix}\right\| + \left\|\begin{pmatrix}0 & 0 \\ 0 & C\end{pmatrix}\right\|$.

([Aud06] contains a proof of a slight generalization of this inequality among other results.)

Item 2 in the original question is false by considering the case where A=C=I/2, B=0, and k=1, where I is the identity matrix. (Did you mean to square the left-hand side?)

Modified item 2 is false; see Willie Wong’s comment on this answer.

Item 3 is false. A simple counterexample is n=2, i=1, $A=\begin{pmatrix}1 & 0 \\ 0 & 0\end{pmatrix}$, $B=\begin{pmatrix}0 & 1 \\ 0 & 0\end{pmatrix}$, $C=\begin{pmatrix}0 & 0 \\ 0 & 1\end{pmatrix}$. Then $2\sqrt{\lambda_1(B^*B)}=2$ but λ1(A+C)=1.

References

[Aud06] Koenraad M. R. Audenaert. A norm compression inequality for block partitioned positive semidefinite matrices. Linear Algebra and its Applications, 413(1):155–176, Feb. 2006. http://dx.doi.org/10.1016/j.laa.2005.08.017

[HJ94] Roger A. Horn, Charles R. Johnson. Topics in Matrix Analysis. Cambridge University Press, 1994.

$\endgroup$
3
  • $\begingroup$ +1 for your partial solution. I think I made a typo in Item 2, $k$ in the left hand side should be $2k$. When $k=n$, obviously item 2 is correct. $\endgroup$
    – Sunni
    Jul 19, 2011 at 15:57
  • 4
    $\begingroup$ Modify the example by Tsuyoshi Ito: take $A = I, C = I/2$, $B = 0$, $n=2$, and $k = 1$. The LHS is $1\times 1 = 1$. The right hand side is $1 \times 1/2 = 1/2$. $\endgroup$ Jul 19, 2011 at 16:05
  • $\begingroup$ Yes, so I see that item 2 is not correct. $\endgroup$
    – Sunni
    Jul 19, 2011 at 16:26

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.