(Notation) Let $\Phi$ be a generic quantum map sending states in $\mathbb{C}^n$ into states in $\mathbb{C}^m$. We say that $\Phi$ is positive when $\Phi(X)\ge0$ for any positive linear operator $X\in\mathrm{Lin}(\mathbb{C}^n)$. We say that $\Phi$ is completely positive (CP) when $\Phi\otimes \operatorname{Id}_k$ is a positive map for all $k\ge0$.

(A standard approach to proving that CP $\iff$ $n$-positive) It is well-known that $\Phi$ is CP iff its Choi representation, $$J(\Phi)\equiv (\Phi\otimes \operatorname{Id}_n)(|m\rangle\!\langle m|)\in\operatorname{Lin}(\mathbb{C}^m\otimes \mathbb{C}^n),$$ is positive semidefinite. Here $|m\rangle\equiv \sum_{i=1}^n |i,i\rangle$ is the (unnormalised) maximally entangled state. A standard way to show this is to observe that

  1. If $J(\Phi)$ is positive semidefinite then it admits an eigendecomposition $J(\Phi)=\sum_a v_a v_a^\dagger$ for some collection of vectors $v_a\in\mathbb{C}^m\otimes \mathbb{C}^n$;
  2. The eigendecomposition for $J(\Phi)$ corresponds to a Kraus-like decomposition for $\Phi$ itself: $\Phi(X)=\sum_a A_a X A_a^\dagger$ with $A_a$ being the linear operators with the same components as the vectors $v_a$.
  3. Such a Kraus-like decomposition can always be rewritten as a Stinespring-like representation $\Phi(X)=\operatorname{Tr}_1(VXV^\dagger)$ with $V\equiv \sum_a |a\rangle\otimes A_a$, and any map with such a representation is CP, because $$(\Phi\otimes \operatorname{Id}_k)(\mathbb{P}(|\Psi\rangle)) = \operatorname{Tr}_1\!\!\big[\mathbb{P}((V\otimes I_k)|\Psi\rangle)\big], \qquad \mathbb{P}(|\psi\rangle)\equiv |\psi\rangle\!\langle\psi|, \qquad \forall|\Psi\rangle\in \mathbb{C}^{n+k},$$ meaning the action of any finite extension of $\Phi$ on unit-rank projections returns the partial trace of a unit-rank projection, which is always a positive semidefinite operator.

(The question) Now, suppose I'm interested in proving the fact that $\Phi\otimes \operatorname{Id}_n$ sending maximally entangled states to valid states is sufficient to know that any extension of $\Phi$ sends physical states into physical states (i.e. that $\Phi$ is CP). The above approach does of course work, but it involves quite a bit of machinery to show something that on the face of it seems a rather simple statement.

Is there a simpler or more direct way to show that $(\Phi\otimes \operatorname{Id}_n)(|m\rangle\!\langle m|)$ being positive is sufficient to know that $\Phi\otimes \operatorname{Id}_k$ is a positive map for all $k$?

  • $\begingroup$ This is not particularly complicated (and indeed you seem to be done after 2 as you get the Kraus operators, what's the point of 3?) This seems primarily opinion-based. $\endgroup$ Commented May 8, 2022 at 23:20
  • $\begingroup$ @NorbertSchuch well, sure, the third point is not crucial, I probably use it because I prefer to think of maps via Stinespring. But the main rational of the question was to avoid having to pass by the eigendecomposition of the Choi altogether. Or at least, a more direct way to relate positivity on general states to the positivity on the maximally entangled state, that doesn't involve having to discuss different representations of the map etc $\endgroup$
    – glS
    Commented May 8, 2022 at 23:31
  • $\begingroup$ I agree that the full Choi proof is quite complicated; indeed, 1 and 2 together don't prove at all that the channel you get is the channel you started with (nor does 3 prove it?), so the full proof would likely be longer. (At least, when I prove the full Choi isomorphism this is taking a while.) -- I posted an argument which might well be considered easier - all you need to know is Schmidt decompositions, I believe. $\endgroup$ Commented May 8, 2022 at 23:37
  • $\begingroup$ ... I still don't get at all what 3 is about. I mean, it is clear that a map in Kraus form is CP, is it? (What does Kraus-"like" even mean? This is entirely a Kraus form, isn't it?) $\endgroup$ Commented May 8, 2022 at 23:39
  • $\begingroup$ @NorbertSchuch yes I agree that 2 and 3 here are essentially equivalent. Both ultimately rely on the fact that $APA$ is positive if $P$ is. The reason I said "Kraus-like" is because I was not technically working with channels, i.e. $\Phi$ could be not trace-preserving, and then the decomposition is not strictly speaking a "Kraus decomposition" I think? i.e. the "Kraus operators" do not need to satisfy the normalisation $\sum_a A_a^\dagger A_a=I$. I'm not sure whether people would talk about a proper Kraus decomposition in this case $\endgroup$
    – glS
    Commented May 9, 2022 at 21:51

2 Answers 2


One alternative argument would be as follows:

  1. If there exists a $\rho\ge0$ such that $\Phi\otimes \mathrm{Id}_k(\rho)\not\ge0$, then there will also exist a pure $\vert\chi\rangle$ such that $\Phi\otimes \mathrm{Id}_k(\vert\chi\rangle\langle\chi\vert)\not\ge0$. (This follows immediately from convexity, e.g. by taking an ensemble decomposition of said $\rho$ -- it will contain one such $\vert\chi\rangle$.)

  2. Take the Schmidt decomposition of $\vert\chi\rangle$, and denote its Schmidt rank by $\ell$. Clearly, $\ell\le n$. Then, when considering $\Phi\otimes \mathrm{Id}_k(\vert\chi\rangle\langle\chi\vert)$, the extending space (the one with the $\mathrm{Id}$) can be compressed to a space with dimension $\ell$ (spanned by the Schmidt vectors). Call the compressed state $\vert\chi'\rangle$.

  3. Clearly, in the compressed space $\Phi\otimes Id_\ell(\vert\chi'\rangle\langle\chi'\vert)\not\ge0$.

This shows that $n$-positive implies $k$-positive for $k>n$. If, in addition, you want to make sure that $n$-positivity when applied only to the maximally entangled state $\vert\Omega\rangle$ is sufficient, then you can do the following:

  1. Assume wlog $\ell = n$. (Otherwise, embed into an $n$-dimensional space.) Write $\vert\chi'\rangle = (\mathrm{I}\otimes M)\vert\Omega\rangle$. Then, $$ (\mathrm{I}\otimes M)\,\big[(\Phi\otimes \mathrm{Id}_n)(\vert\Omega\rangle\langle\Omega\vert)\big]\,(\mathrm{I}\otimes M^\dagger) = (\Phi\otimes \mathrm{Id}_n)(\vert\chi'\rangle\langle\chi'\vert) \not\ge 0\ , $$ which proves that $(\Phi\otimes \mathrm{Id}_n)(\vert\Omega\rangle\langle\Omega\vert)\ge0$ implies $n$-positivity.

(Fun fact on the side: This argument can also be used to show that in order to check CP, it is sufficient to evaluate the action of the channel on any $\vert\chi'\rangle$ with maximal Schmidt rank, since in that case, $M$ is invertible, and the argument works both ways.)

  • $\begingroup$ ah, nice, this is the kind of thing I was hoping for. One could also make a trivial modification to the argument to make it "positive" I think: $$(\Phi\otimes\operatorname{Id}_k)(|\chi\rangle\!\langle\chi|)=(I\otimes D)[(\Phi\otimes\operatorname{Id}_n)(|\Omega\rangle\!\langle\Omega|)](I\otimes D)$$ with $D$ diagonal matrix with the Schmidt coefficients of $\chi$, so that we get a direct relation between positivity on $|\Omega\rangle$ and positivity on arbitrary (pure) states $\endgroup$
    – glS
    Commented May 9, 2022 at 21:47
  • $\begingroup$ If D is singular this is a one-way relation. But isn't this precisely my point 4? $\endgroup$ Commented May 9, 2022 at 22:54
  • $\begingroup$ sure. I just like this equation because it gives a direct explicit relation answering the question: "how does the action of $\Phi\otimes\operatorname{Id}_k$ on a generic state $|\chi\rangle$ relate to the action on maximally entangled states?". It's essentially your point 4, yes, observing that the matrix $D$ is just the diagonal matrix whose elements are the Schmidt coefficients $\endgroup$
    – glS
    Commented May 10, 2022 at 7:07
  • 1
    $\begingroup$ @gls ... or my 4 using the SVD of M :) In fact, this is one of the advanced questions I like asking in oral exams, if everything goes smoothly ;) (Will have to remember to delete this comment! :-o) $\endgroup$ Commented May 10, 2022 at 7:17
  • $\begingroup$ @NorbertSchuch Actually something that always bugs me is whether this theorem implies 2-positivity is enough to prove CP, since your proof essentially says that if it's 2-positive then it's true for $k > 2$. I have never seen this connection stated clearly. $\endgroup$ Commented Jun 1, 2022 at 1:37

You can skip point 3 since $\Phi(X) = AXA^\dagger$ is a CP map and a sum of CP maps is again CP. The map $\Phi$ is CP because $(\Phi \otimes {\rm Id})(Z)= (A\otimes I)Z(A\otimes I)^\dagger$ – clearly a positive map.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.