7
$\begingroup$

All quantum operations $\mathcal{E}$ on a system of Hilbert space dimension $\mathcal{d}$ can be generated by an operator-sum representation containing at most $\mathcal{d^2}$ elements. Extending further, an operation from space with dimension $\mathcal{m}$ to space with dimension $\mathcal{n}$ has an operator sum representation in terms of Kraus operators. Refer: http://en.wikipedia.org/wiki/Quantum_operation#Kraus_operators

The follow up proof is in terms of an exercise in Nielsen and Chuang's book, exercise 8.10 enter image description here

Proving the first part is easy enough... $ W_{jk} = W_{kj}^{*} $ if you expand the definition of $ W_{jk}$ and using properties of transpose conjugate. So W is in fact Hermitian. That it is of rank at most $ d^2 $ is what I'm not able to prove.

I read on wikipedia about the properties of rank of a matrix, and it says that a matrix of rank M can be expressed as the sum of M rank-1 matrices. In that case I would need to prove that the individual terms in the sum:

$ \sum_{j,k} tr(E_j^{+}E_k)\left|j\right \rangle \left\langle k \right| $

are of rank 1. Since $tr(E_j^{+}E_k) $ is in fact a scalar, and $ \left|j\right \rangle$ & $\left\langle k \right| $ are the eigenbasis for the input and output spaces, there can be d such terms of $ \left|j\right \rangle$ and $ \left|k\right \rangle$ respectively, the product of which forms a rank-1 matrix.

Hence there would be at max d*d such terms, if all the terms $tr(E_j^{+}E_k) $ are non-zero.

Is that the correct proof? Am I making a mistake somewhere?

The number of such terms is also called the Kraus rank as given in: https://en.wikipedia.org/wiki/Quantum_channel#Pure_channel

$\endgroup$
  • $\begingroup$ yes, you are making a mistake: A matrix of rank k kann be expressed as a sum of k rank-1 matrices (that's via singular value decomposition, just as Nielsen&Chuang say), but that doesn't mean that ANY decomposition into rank-1 matrices has at most k matrices. Rather the opposite: it has at least k matrices. Thus proving (which is trivial) that the summands in your some are of rank-1 doesn't get you anywhere. $\endgroup$ – Martin Oct 8 '14 at 17:04
4
$\begingroup$

Since there still isn't an answer but the question has attracted a few upvotes, let me elaborate on my comment. This is more maths, than physics, but anyway.

Writing $\sum_{jk} \mathrm{tr}(E_j^{\dagger}E_i)|j\rangle\langle k|$ doesn't give you anything. This is indeed a rank-one decomposition, but the theorem does not tell you that ANY rank-1 decomposition has at most d^2 terms. This would be true, if $|j\rangle$ was the eigenbasis of $\mathbb{C}^d$ - but it is not. $W$ is an $M\times M$ matrix, hence $|j\rangle$ is the eigenbasis of $\mathbb{C}^M$ and $M$ could be much larger than this.

However, what is true is that $E_j\in\mathbb{C}^{d\times d}$. The key observation is that at most $d^2$ of these $E_j$ can therefore be linearly independent and this entails that $W$ can only be of rank at most $d^2$. Here is a proof (albeit not that pretty):

There is a basis of $\mathbb{C}^{d\times d}$ with $d^2$ elements, call it $F_j$, which is orthonormal with respect to the trace inner-product. Now, since the $F_j$ form a basis, for every $E_j$ we have:

$$ E_j=\sum_i a_i^{(j)} F_i \qquad a_i^{(j)}\in\mathbb{C}\quad \forall\,j\in \{1,\ldots,N\}$$

the $E_j$ are linear combinations of the $F_j$. But then, we can reexpress the columns of W by $F_i$ and obtain:

$$ \operatorname{tr}(E_i^{\dagger}E_j)=\sum_{k=1}^{d^2} {a_k^{(i)}}^* a_k^{(j)} $$

By definition of a basis, only $d^2$ of the $E_i$ can be linearly independent. Without loss of generality, we suppose the first $d^2$ $E_i$ were linearly independent. Let's then have a look at the $(d^2+1)$th column. Since $E_{d^2+1}$ is linearly dependent, its coefficients $a_k^{(d^2+1)}$ are linear combinations of the other $a^{(j)}$, say

$$ a^{(d^2+1)}_k=\sum_{j=1}^{d^2} b_ja_k^{(j)}$$

But then, we can see that

$$ \operatorname{tr}(E_i^{\dagger}E_{d^2+1})=\sum_j b_j \operatorname{tr}(E_i^{\dagger}E_j) $$

hence the whole column is a linear combination of the former columns. Putting everything together, $W$ can have at most $d^2$ linearly independent columns. We can then diagonalize $W$, since it is Hermitian and proceed as indicated.

$\endgroup$
  • 1
    $\begingroup$ I can see the reasoning for determining the maximal possible rank of W. Still, I fail to see the usefulness in using W for proving the particular statement. That is, why is W constructed the way it is? What makes the unitary matrix U significant by its ability to make W diagonal? I'm a bit confused by the premises suggested for this problem. $\endgroup$ – Invoker Aug 20 '16 at 12:41

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.