2
$\begingroup$

Considering in the complex fields. Let $P$ be a nonsingular matrix, $P^* $ be its conjugate transpose, is there a relation between $P^*|D|P$ and $|P^*DP|$, where $D$ is a diagonal matrix? In particular, is it true

$P^* |D|P \ge |P^*DP|$ in the sense of Lowner order, or is there an order for eigenvalues?

Here $|A|=(A^*A)^{1/2}$, the absolute value of a complex matrix.

Edit As I know from Suvrit's answer, there is no relation like $P^* |D|P \ge |P^* DP|$ in the sense of Lowner order. So my question becomes, is the $i$th largest eigenvalue of $P^* |D|P$ larger than that of $|P^*DP|$?

$\endgroup$
1
  • $\begingroup$ Even the modified question has a negative answer; please see my update. $\endgroup$
    – Suvrit
    Oct 26, 2011 at 10:42

2 Answers 2

7
$\begingroup$

Update: In the edited question, the OP asks whether $\lambda_i^\downarrow(P^*|D|P) \ge \lambda_i^\downarrow(|P^*DP|)$ (or even the reverse direction). Such a relations do not hold either. Take for e.g., $P=\begin{bmatrix} 2 & 1 \\\\ 2 & 2\end{bmatrix}$, and use the same $D$ as below.

Then, we have $\lambda^\downarrow(P^*|D|P) = (41.22, 0.776)$, while $\lambda^\downarrow(|P^*DP|) = (23.369, 1.369)$.

However, if one assume that $P$ is a contraction, then several interesting results can be shown.


I don't think there is any useful relation.

Here is a counterexample:

\begin{equation*} P = \begin{bmatrix} 2 & 2\\\\ 2 & 4 \end{bmatrix}, \end{equation*} and \begin{equation*} D = \begin{bmatrix} -2 & 0\\\\ 0 & 4 \end{bmatrix} \end{equation*} Then, \begin{equation*} P^T|D|P = \begin{bmatrix} 12 & 16\\\\ 16 & 24 \end{bmatrix} \end{equation*} and

\begin{equation*} |P^TDP|^2 = \begin{bmatrix} 640 & 1536\\\\ 1536 & 3712 \end{bmatrix} \end{equation*} Then, \begin{equation*} \lambda(P^T|D|P - |P^TDP|) = (-3.3678, 31.4856), \end{equation*} which is indefinite.

$\endgroup$
7
  • $\begingroup$ Isn't your $P^TDP=[8, 24; 24, 56]$, then $|P^TDP|=[11.3137, 22.6274; 22.6274, 56.56858]$? $\endgroup$
    – Sunni
    Jul 12, 2011 at 23:54
  • $\begingroup$ sorry, while typesetting I reran with 'sqrt' instead of 'sqrtm' by mistake; it's still a counterexample with the matrix squareroot--- thanks for catching. $\endgroup$
    – Suvrit
    Jul 13, 2011 at 1:54
  • $\begingroup$ I think there is a relation between eigenvalues, i.e. the $k$th largest eigenvlaue of $P^* |D|P$ is no less than the $k$th largest eigenvalue of $|P^*DP| $\endgroup$
    – Russel
    Jul 13, 2011 at 17:12
  • $\begingroup$ Sure, one can have several inequalities; All I am saying is that the semidefinite ordering that you asked in your second question above, does not hold in either direction because of the above indefinite matrix. $\endgroup$
    – Suvrit
    Jul 13, 2011 at 17:40
  • 1
    $\begingroup$ Is there any reason why you do not divide everything by 2? $\endgroup$ Jul 22, 2011 at 13:40
1
$\begingroup$

Using the same example above, I found

p=[1,1;1,2]

p =

 1     1
 1     2

d=[-1,0;0,2]

d =

-1     0
 0     2

c=[1,0;0,2]

c =

 1     0
 0     2

eig(((p*d*p)^2)^(1/2))

ans =

0.2426
8.2426

eig(p*c*p)

ans =

0.1690

11.8310

But this example shows it is possible the majorization relation holds.

$\endgroup$

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.