This was an interesting result I found while reading Holevo's Quantum Information theory book. They call it the non-commutative operator Cauchy Schwarz.

Let $S$ be a state i.e. a non negative operator with unit trace on a Hilbert space $\mathcal{H}$ in $\mathbb{C}$. Then for arbitrary operators $X$ and $Y$ on $\mathcal{H}$, \begin{equation} |Tr(SX^*Y)|^2 \leq Tr(SX^*X)Tr(SY^*Y)\label{Thi} \end{equation}

I was able to show this by mirroring the usual Cauchy Schwarz inequality (I assumed RHS is finite). Then a friend suggested that it might follow easily if we show that $Tr(SX^*Y)$ is an inner product on the operator space. I was able to show that it is a semi inner product. To prove it is an inner product, I needed to show $$Tr(SX^*X) = 0 \Rightarrow X=0$$

I wasn't successful with this one. I was able to show it if $S$ is a strictly positive operator (at least I think I did) but not for non negative. My guess is that in general, $Tr(SX^*Y)$ is a seminorm but not necessarily a norm.

But I was unable to furnish an $X$ such that $X\ne 0$ but $Tr(SX^*X) = 0$. I seek such an example. Kindly give me some ideas. We could think in terms of matrices here.

Update: Upon further inspection, I realized that my "proof that mirrors cauchy schwarz" was in fact flawed. I assumed implicitly that both terms in RHS were $>0$. In one of the steps, I divide by the square root of RHS, which would be illegal if it were $0$. This wasn't a problem in usual Cauchy schwarz because we were given that LHS is square of mod of an inner product. This changes my perception and I now believe that the above is an inner product but I need to show the final step.

Update: I'll show my work so far. Let $\langle X,Y\rangle_T \triangleq Tr(SX^*Y)$. Then

  1. Linearity in second argument (In Quantum theory, it is the second argument) and Conjugation: Easy to show.
  2. Non-Negative for equal arguments: We have $$Tr(SX^*X) = Tr(XSX^*)$$ Since $S \geq 0$, $XSX^* \geq 0$ and hence $Tr(XSX^*) \geq 0$. Also if $X=0$, then $\langle X,X\rangle_T = 0$.
  3. Need to show if $Tr(SX^*X) = 0$ then $X=0$. If $S$ is strictly positive definite, then for any orthonormal basis $e_i$

$$ 0 = \langle X,X\rangle_T = Tr(\sqrt{S}X^*X\sqrt{S}) = \sum_{i=1}^d \langle e_i,\sqrt{S}X^*X\sqrt{S}e_i\rangle \\ = \sum_{i=1}^d \langle X\sqrt{S}e_i,X\sqrt{S}e_i\rangle = \sum_{i=1}^d \|X\sqrt{S}e_i\|^2$$

Hence $\|X\sqrt{S}e_i\| =0 $ for every $i$ and this further implies $X\sqrt{S} = 0$. Since $\sqrt{S}$ is invertible owing to strict positivity, we get $X=0$.This completes the proof when $S$ is strictly positive definite. For non-negative definite, I don't know how to tackle it as $S$ may have a non-trivial kernel/nullspace.

  • $\begingroup$ Could you just work with blocks? $\endgroup$ Jun 15, 2018 at 3:46
  • $\begingroup$ Perhaps. But I just now realized that my proof had a small but important flaw. Please see my update above. In this case, I have to resume my usual inner product proof. $\endgroup$ Jun 15, 2018 at 3:50
  • $\begingroup$ Cauchy - Schwartz inequality holds for any semi-inner product $\endgroup$ Jun 15, 2018 at 5:58
  • $\begingroup$ @Kavi: i didn't know that. Let me try that psrt again. $\endgroup$ Jun 15, 2018 at 8:27

1 Answer 1


Looks like I figured it out. I think one doesn't require 3) in my second update to prove the result. If $S$ has a nontrivial nullspace, it isn't true. So we tackle that case by showing that if either $Tr(SX^*X)=0$ or $Tr(SY^*Y)=0$ then $Tr(SX^*Y)=0$.

Suppose $Tr(SX^*X)=0$, then given a basis $e_i$, we have $$0=Tr(SX^*X) = Tr(XSX^*)=\sum_{i=1}^d \langle e_i, XSX^*e_i \rangle =\sum_{i=1}^d \langle \sqrt{S}X^*e_i, \sqrt{S}X^*e_i \rangle \\= \sum_{i=1}^d\| \sqrt{S}X^*e_i\|^2$$ The last step implies $\|\sqrt{S}X^*e_i \|=0$ and hence $\sqrt{S}X^*e_i=0$ for all $i$. Hence $\sqrt{S}X^*=0$. Now $$Tr(SX^*Y) = Tr(Y\sqrt{S}\sqrt{S}X^*) = 0$$ The other case is similar. For the case where $Tr(SX^*X)\ne 0,Tr(SY^*Y)\neq 0$ but finite, we can mirror the Cauchy Schwarz proof.


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.