In general, a Hermitian matrix can have complex off-diagonal terms. Given any Hermitian matrix $[A]_{n,m}$, I can construct another matrix $[\vert A\vert ]_{n,m} =\vert A_{n,m} \vert$. I would like to know if there are any known relationships between two such matrices. Some more specific questions that I was thinking about but don't have an answer to are the following:

  1. Is it possible to transform using a unitary $UAU^{\dagger}$ or using a convex combination of unitary matrices $\sum_i p_i U_iAU_i^{\dagger}$ where $\sum_i p_i =1$? It seems to be possible for $2 \times 2$ positive semi-definite matrices.

  2. Are there any inequalities relative their trace norms $\operatorname{tr}(\sqrt{A^\dagger A})$ and $\operatorname{tr}(\sqrt{\vert A \vert^\dagger \vert A \vert})$? Such as if one is always bigger than the other.

Any information will be helpful for general Hermitian matrices will be helpful, but if you would like to know, I am looking at two special cases matrices right now:

a) $A$ is a positive definite matrix.

b) $A$ is a hollow (zero main diagonal), banded, Toeplitz and Hermitian matrix that looks like the following:

\begin{bmatrix} 0 & a_1 & a_2 & \ldots & a_n \\ a_1^* & 0 & a_1 &\ldots & \vdots \\ a_2^* &a_1^* & 0 & \ldots &\vdots \\ \vdots & & & \ddots & \vdots\\ a_n^* & a_{n-1}^* & a_{n-2}^* &\ldots &0 \end{bmatrix}

Thanks for all your help.

  1. The condition $$\tag{1} B=\sum_jp_jU_jAU_j^*,\ \ \ \sum p_j=1,\ \ p_j\geq0$$ implies that $B$ has the same trace as $A$. If $$ A=\begin{bmatrix}1&1\\1&-1\end{bmatrix}, $$ neither $A$ is a convex combination of unitary conjugates of $|A|$, nor $|A|$ is a convex combination of unitary conjugates of $A$. $$\ $$ It does work, as you say, for $2\times2$ positive-semidefinite matrices. In this case the main diagonal of $A$ consists of nonnegative entries, and so if $A_{12}=e^{it}\,|A_{12}|$, $$ UAU^*=\begin{bmatrix}1&0\\0&e^{it}\end{bmatrix}\, \begin{bmatrix}A_{11}&e^{it}|A_{12}|\\ e^{-it}|A_{21}|&A_{22}\end{bmatrix}\, \begin{bmatrix}1&0\\0&e^{-it}\end{bmatrix}\, = \begin{bmatrix}A_{11}&|A_{12}|\\ |A_{21}|&A_{22}\end{bmatrix}\,=|A|. $$ $$\ $$ Note that $(1)$ is equivalent to the majorization $\lambda(B)\prec\lambda(A)$ (i.e. majorization between the vectors of eigenvalues). For the matrix $$ A=\begin{bmatrix}0&-2&-2\\ -2&0&-3\\-2&-3&0\end{bmatrix}, $$ we have $\lambda(|A|)=-\lambda(A)$. So to be able to obtain $|A|$ in the form $(1)$, we would need $-\lambda(A)\prec\lambda(A)$, which one can check explicitly it fails.$$\ $$ Similarly, it also fails when $A$ is positive-semidefinite: for $$ A=\begin{bmatrix}4&-2&-2\\ -2&5&-3\\-2&-3&6\end{bmatrix}, $$ again $\lambda(A)$ and $\lambda(|A|)$ do not majorize each other, so we cannot obtain one from the other with an expression of the form $(1)$.

  2. Since $\text{tr}(A^*A)=\text{tr}(|A|^*|A|)$ always, , $$ \text{tr}(\sqrt{A^*A})\leq n^{1/2}\,\text{tr}(A^*A)^{1/2}= n^{1/2}\,\text{tr}(|A|^*|A|)^{1/2}\leq n^{1/2}\,\text{tr}(\sqrt{|A|^*|A|}) $$


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.