Let $\lVert\cdot\rVert$ be the matrix/sup norm on $M_n(\mathbb{C})$: $$\lVert A \rVert = \sup\{\lVert Av\rVert~;~\lVert v\rVert = 1\}$$ which for positive operators is equal to their highest eigenvalue.

Let $\rho$ and $\pi$ be two states in $M_n(\mathbb{C})$. That is: $\rho\geq 0$ and Tr$(\rho) = 1$ and so for $\pi$. And define $\rho^{-1}$ as the generalised inverse of $\rho$: the operator with the same eigenspaces and kernel, but every nonzero eigenvalue $\lambda$ changed to $\lambda^{-1}$, and define $\rho^{-\frac{1}{2}} = \sqrt{\rho^{-1}}$, where $\sqrt{}$ is the usual positive matrix square root.

Assuming that ker$(\rho)\subseteq $ker$(\pi)$, is it true that $$\lVert \rho^{-\frac{1}{2}}\pi\rho^{-\frac{1}{2}}\rVert\geq 1$$ for any such states $\rho$ and $\phi$?

I know this is true when $\rho$ and $\pi$ can be simultaneously diagonalised, because then this question reduces to a trivial statement about probability distributions. I'm looking for a proof (or counterexample) in the general case.

edit: Proof of this last statement: Suppose $\rho$ and $\pi$ are both diagonalised by some unitary $U$: $\rho=UPU^\dagger$, $\pi = UQU^\dagger$, for some diagonal $P$ and $Q$. Since the norm is unitary invariant, the expression we want to prove is:

$$\sup_i \frac{P_i}{Q_i} \geq 1$$

Suppose this is not the case, then $P_i\leq Q_i$ for all $i$, but we have $\sum_i P_i = \sum_i Q_i = 1$ and $P_i,Q_i\geq 0$, so this is only possible when $P_i=Q_i$ for all $i$.

  • $\begingroup$ Positive means $z^* \rho z \ge 0$ for any complex vector $z$? $\endgroup$ – user251257 Oct 28 '16 at 13:06
  • $\begingroup$ Perhaps I have missed something but if $\rho=e_1 e_1^*$ (pure state) your definitions seem to set $\rho^{-1/2} = e_1 e_1^*$ as well? In which case $\rho^{-1/2}\pi\rho^{-1/2}$ could be zero? $\endgroup$ – H. H. Rugh Oct 28 '16 at 13:06
  • $\begingroup$ @H.H.Rugh hence $\ker \rho \subseteq \ker \pi$ $\endgroup$ – Omnomnomnom Oct 28 '16 at 13:08
  • $\begingroup$ You are right I did miss something. $\endgroup$ – H. H. Rugh Oct 28 '16 at 13:10
  • $\begingroup$ Yeah, that's why I require that extra property. @user251257: That is indeed what I mean by positive. Alternatively, that every eigenvalue is positive. $\endgroup$ – John Oct 28 '16 at 13:15

You can prove the general case analogously. Assume wlog. $\rho$ be invertible after dividing out $\ker \rho$.

Notice that $\rho^{-1/2} \pi \rho^{-1/2}$ is positive again. Now, assume the converse $$ \| \rho^{-1/2} \pi \rho^{-1/2} \| = \lambda_\max(\rho^{-1/2} \pi \rho^{-1/2}) < 1. $$ Then, we have $$ tr(\pi) = \sum_{i=1}^n e_i^* \rho^{1/2} (\rho^{-1/2} \pi \rho^{-1/2}) \rho^{1/2} e_i < \sum_{i=1}^n e_i^* \rho^{1/2} \rho^{1/2} e_i = tr(\rho) = 1. $$


Let $Z=\ker \rho$ and let $e_1,...,e_k$ be an orthonormal base for $Z^\perp$ which we supplement with $e_{k+1},...,e_n$ to get an orthonormal base for the full space. Then $$\rho=\sum_{i=1}^k \lambda_{i} e_i e_i^*, \ \ \ \ \rho^{-1/2} = \sum_{i=1}^k\lambda_i^{-1/2} e_i e_i^*$$ We then have for the trace norm $$ |\rho^{-1/2} \pi \rho^{-1/2}|_1 = \sum_{i=1}^k \frac{1}{\lambda_i} (e_i^* \pi e_i) \geq \sum_{i=1}^k (e_i^* \pi e_i) =\sum_{i=1}^n (e_i^* \pi e_i)= {\rm Tr} \; \pi =1$$ where we used $\ker \rho\subset \ker \pi$ in the third equality. And similarly (with $q=\rho^{-1/2} \pi \rho^{-1/2}$): $$ 1=|\pi|_1 = \sum_{i=1}^n (e_i^* \pi e_i) = \sum_{i=1}^k \lambda_i (e_i^* q e_i)\leq \|q\| \sum_{i=1}^k \lambda_i = \|q\| \; |\rho|_1 = \|q\|$$

  • $\begingroup$ @H.H.Rugh it looks like you're using some kind of trace norm instead of the operator norm $\endgroup$ – Omnomnomnom Oct 28 '16 at 13:29

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.