4
$\begingroup$

A special case of the well known Lieb concavity theorem states that the following function is concave on positive operators A and B:

$$ (A,B) \to \text{Tr} \{A^s X B^{1-s} X^\dagger \} $$

for $s \in [0,1]$ and an arbitrary operator $X$. I am wondering if it is known whether the following function is concave as well:

$$ (A,B) \to \text{Tr} \{A^s B^{(1-s)/2} X X^\dagger B^{(1-s)/2} \} $$

If not, do you have a counterexample? Apologies if this question has already been posted somewhere else in math overflow...

I have a similar question as the above one for the Ando convexity theorem.

$\endgroup$

1 Answer 1

5
$\begingroup$

The conjectured inequality is false.

Using cyclicity of the trace, let's first write it in a slightly nicer form \begin{equation*} f(A,B) := \operatorname{tr}(X^*B^{(1-s)/2} A^sB^{(1-s)/2}X). \end{equation*}

For joint-concavity to hold, we'd like to show \begin{equation*} f\left(\tfrac{A+B}{2}, \tfrac{U+V}{2} \right) \ge \tfrac12f(A,U)+\tfrac12f(B,V).\qquad (*) \end{equation*}

But picking $A,B,U,V,X$ as follows, we have an immediate counterexample (I used $s=0.5$): \begin{equation*} A = \begin{pmatrix} 41 & 25\\ 25 & 26\end{pmatrix},\ B = \begin{pmatrix} 2 & 7\\ 7 & 25\end{pmatrix},\ U = \begin{pmatrix} 25 & 22\\ 22 & 20\end{pmatrix},\ V = \begin{pmatrix} 5 & 5\\ 5 & 10\end{pmatrix},\\ X = \begin{pmatrix} 1 & 0\\ 7 & -1\end{pmatrix}. \end{equation*}

With this choice of variables, we obtain 5.964618e+03 < 6.529930e+03 in $(*)$ above.


EDIT

In case you are interested or for others who may interested in concavity / convexity results similar in flavor to the one asked above, I'd like to mention the following nice paper: Concavity of certain matrix trace and norm functions, F. Hiai (2013).

In that paper, Hiai studies (among others), joint concavity and convexity of maps of the form \begin{equation*} (A,B) \mapsto \mbox{tr}(\Phi(A^p)^{1/2}\Psi(B^q)\Phi(A^p)^{1/2})^s, \end{equation*} for suitable choices of reals $p,q,s$, and positive linear maps $\Phi$ and $\Psi$ (e.g, with $\Phi(A) = X^*AX$, $\Psi=\mbox{Id}$, and $s=1$, we recover Lieb's concavity).

$\endgroup$
1
  • $\begingroup$ Thanks for this answer. Since we know that it is true for $X = I$, is it possible to determine conditions on X such that the inequality is true? $\endgroup$ Apr 23, 2014 at 4:14

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.