On the other hand, if the map is linear, completely positive, trace preserving, we prove that the map has Kraus decomposition.
Let dA≡dim(HA),dB≡dim(HB) and R is an auxiliary system with the same orthonormal basis with A. Let ∣Γ⟩RA denote the following unnormalized maximally entangled vector:
∣Γ⟩RA≡i=0∑dA−1∣i⟩R⊗∣i⟩A
then we have
(I⊗NA→B)(∣Γ⟩⟨Γ∣RA)=i,j=0∑dA−1∣i⟩⟨j∣R⊗NA→B(∣i⟩⟨j∣A)
Since ∣Γ⟩⟨Γ∣RA is positive semi-definite, since NA→B is completely positive, the RHS is positive semi-definite, too. So it can be diagonalized:
i,j=0∑dA−1∣i⟩⟨j∣R⊗NA→B(∣i⟩⟨j∣A)=l=0∑d−1∣ϕl⟩⟨ϕl∣RB
where d≤dAdB, dAdB is the dimension of system RB, and ∣ϕl⟩ does not necessarily have to be orthonormal.
we expand the ∣ϕl⟩ in terms of an orthonormal basis of RB:
∣ϕ⟩RB=i=0∑dA−1j=0∑dB−1αij∣i⟩R⊗∣j⟩B
let VA→B denote the following linear operator:
VA→B≡i=0∑dA−1j=0∑dB−1αi,j∣j⟩B⟨i∣A
Then we see:
(I⊗VA→B)∣Γ⟩RA=(I⊗i=0∑dA−1j=0∑dB−1αi,j∣j⟩B⟨i∣A)(k=0∑dA−1∣k⟩R⊗∣k⟩A)=i=0∑dA−1j=0∑dB−1k=0∑dA−1αi,j∣k⟩R⊗∣j⟩B⟨i∣k⟩A=i=0∑dA−1j=0∑dB−1αi,j∣i⟩R⊗∣j⟩B=∣ϕ⟩RB
So we see that for all vector ∣ϕ⟩RB, we can find a linear operator VA→B such that (I⊗VA→B)∣Γ⟩RA=∣ϕ⟩RB.
And we have:
(⟨i∣R⊗I)∣ϕ⟩RB=(⟨i∣R⊗I)⋅(I⊗VA→B)∣Γ⟩RA=(⟨i∣R⊗I)(I⊗VA→B)j=0∑dA−1∣j⟩R⊗∣j⟩A=j=0∑dA−1⟨i∣j⟩R⊗VA→B∣j⟩A=VA→B∣i⟩A
Finally we have:
NA→B(∣i⟩⟨j∣A)=(⟨i∣R⊗I)⋅p,q=0∑dA−1∣p⟩⟨q∣R⊗NA→B(∣p⟩⟨q∣A)⋅(∣j⟩R⊗I)=(⟨i∣R⊗I)⋅l=0∑d−1∣ϕl⟩⟨ϕl∣RB⋅(∣j⟩R⊗I)=l=0∑d−1(⟨i∣R⊗I)∣ϕl⟩⋅⟨ϕl∣(∣j⟩R⊗I)=l=0∑d−1(Vl)A→B∣i⟩⟨j∣A(Vl)A→B†
We still need to prove that ∑lVl†Vl=I. Since NA→B is trace preserving, then:
tr(NA→B(∣i⟩⟨j∣A))=tr(∣i⟩⟨j∣A)=δij
Consider:
δij=tr(NA→B(∣i⟩⟨j∣A))=tr(l∑Vl∣i⟩⟨j∣AVl†)=tr(l∑Vl†Vl∣i⟩⟨j∣A)=⟨j∣Al∑Vl†Vl∣i⟩A
That is, ∑lVl†Vl=I.