7
$\begingroup$

Lets have two Hermitian $n\times n$ matrices $X$ and $Y$.

Are there any known properties of the singular values of $$Z = X + i Y.$$

I am the most interested in bounding from above a few first singular values of $Z$ by the eigenvalues of $X$ and $Y$. And sth that is stronger than:

$$\sum_{i=1}^k \sigma_i^2(Z)\leq \sum_{i=1}^k \left( \lambda_i^2(X)+\lambda_i^2(Y)+\lambda_i(i[X,Y]) \right)$$ for $1\leq k \leq n$ and (singular/eigen)values sorted in the decreasing order.

$\endgroup$
3

1 Answer 1

4
$\begingroup$

Some results in this direction that you might find useful are listed below.

Theorem (Bhatia and Kittaneh). Let $X$, $Y$, and $Z$ be as in the question above. Then, \begin{equation*} \| (X^2+Y^2)^{1/2} \|_p \le \|Z\|_p \le 2^{1/2-1/p}\| (X^2+Y^2)^{1/2} \|_p, \end{equation*} where $2 \le p \le \infty$, and $\|\cdot\|_p$ denotes the Schatten-$p$ norm. The inequality above gets reversed for $1\le p \le 2$. Also, these inequalities are sharp.

Even more directly relevant is the following theorem that discusses majorization of singular values of $X+Y$ by those of $Z$.

Theorem (Bhatia and Kittaneh). Let $X$, $Y$, and $Z$ be as in the question. Then \begin{equation*} \sigma(X+Y)\quad \prec_w\quad \sqrt{2}\sigma(Z) \end{equation*} If $X$ is psd, then the above weak majorization can be replaced by weak log-majorization, that is, \begin{equation*} \prod_{j=1}^k \sigma_j(X+Y) \le \prod_{j=1}^k\sqrt{2}\sigma_j(Z). \end{equation*} Finally, if both $X$ and $Y$ are psd, then we have even stronger inequalities: \begin{equation*} \sigma_j(X+Y) \le \sqrt{2}\sigma_j(Z)\quad 1 \le j \le n. \end{equation*}

Bhatia and Kittaneh also discuss some applications of the above theorem to commutator inequalities.

References

  1. R. Bhatia and F. Kittaneh. "The singular values of $A+B$ and $A+iB$." Linear Algebra and its Applications, 431(2009), pp. 1502-1508.

  2. R. Bhatia and F. Kittaneh. "Cartesian decompositions and Schatten norms." Linear Algebra and its Applications, 318(2000), pp. 109--116.

$\endgroup$
0

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.