The max-relative entropy between two states is defined as

$$D_{\max }(\rho \| \sigma):=\log \min \{\lambda: \rho \leq \lambda \sigma\},$$

where $\rho\leq \sigma$ should be read as $\sigma - \rho$ is positive semidefinite. Consider the following quantity for a bipartite state $\rho_{AB}$ with reduced states $\rho_{A}$ and $\rho_{B}$.

$$I_{\max}(\rho_{AB}) = D_{\max}(\rho_{AB}||\rho_{A}\otimes\rho_{B})$$

I would like to know if this satisfies a continuity bound. That is, given $\rho_{AB}\approx_{\epsilon}\sigma_{AB}$ in some distance measure, can we bound $|I_{\max}(\rho_{AB}) - I_{\max}(\sigma_{AB})|$?

Motivation for question: Recall the quantum relative entropy $D(\rho||\sigma) = \text{Tr}(\rho\log\rho - \rho\log\sigma)$ with the convention that $0\log 0 = 0$. Let us define the mutual information as follows

$$I(\rho_{AB}) = D(\rho_{AB}||\rho_{A}\otimes\rho_{B}) = -S(\rho_{AB}) + S(\rho_A) + S(\rho_B),$$

where $S(\rho) = -\text{Tr}(\rho\log\rho)$ is the von Neumann entropy. In this case, we may use Fannes inequality to find a bound on $|I(\rho_{AB}) - I(\sigma_{AB})|$ in terms of $\|\rho_{AB} - \sigma_{AB}\|_1$. I'm wondering if the move from $D(.||.)$ to $D_{\max}(.||.)$ can be made while still having some Fannes type bound.


Unfortunately $D_{\max}$ is not a continuous function and so functions built from it tend not to be continuous. For example consider consider the two states

$$ \rho_{AB} = |00 \rangle \langle 00|, $$ and $$ \tau_{AB}(\epsilon) = (1-\epsilon) |00 \rangle \langle 00 | + \epsilon | 11\rangle \langle 11 |. $$ A quick calculation gives $I_{\max}(\rho_{AB}) = 1$ and $I_{\max}(\tau_{AB}(\epsilon)) = - \log(\epsilon)$. So $I_{\max}(\rho_{AB}) \neq \lim_{\epsilon \rightarrow 0} I_{\max}(\tau_{AB}(\epsilon)) = \infty$.

However, some authors have studied a smoothed version of $I_{\max}$ (see https://arxiv.org/pdf/1308.5884.pdf).


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.