On Quantum Channels.

Frank Verstraeteab and Henri Verscheldea

a
Department of Mathematical Physics and Astronomy, Ghent University, Belgium
b
Department of Electrical Engineering (SISTA), KULeuven, Belgium
(Dated: February 21, 2002)
One of the most challenging open problems in quantum information theory is to clarify and
quantify how entanglement behaves when part of an entangled state is sent through a quantum
channel. Of central importance in the description of a quantum channel or completely positive map
(CP-map) is the dual state associated to it. The present paper is a collection of well-known, less
known and new results on quantum channels, presented in a unified way. We will show how this
dual state induces nice characterizations of the extremal maps of the convex set of CP-maps, and
how normal forms for states defined on a Hilbert space with a tensor product structure lead to
arXiv:quant-ph/0202124v2 22 Jan 2003

interesting parameterizations of quantum channels.

The existence of non-local correlations or entanglement in multipartite quantum systems [1, 2] is one of the corner-
stones on which the newly established field of quantum information theory is build. The main gain of quantum over
classical information processing stems from the fact that we are allowed to perform operations on entangled states:
through the quantum correlations, an operation on a part of the system affects the whole system. One of the most
challenging open problems is to clarify and quantify how entanglement behaves when part of an entangled state is
sent through a quantum channel.
Of central importance in the description of a quantum channel or completely positive map (CP-map) is the dual
state associated to it. This state is defined over the tensor product of the Hilbert space itself (the input of the channel)
with another one of the same dimension (the output of the channel). It is clear that there appears a natural tensor
product structure, and indeed the notion of entanglement will be crucial in the description of quantum channels.
In a typical quantum information setting, Alice wants to send one qubit (eventually entangled with other qubits)
to Bob through a quantum channel. The channel acts linearly on the input state, and the consistency of quantum
mechanics dictates that this map be completely positive (CP) [3]. This implies that the map is of the form [4]
X
Φ(ρ) = Ai ρA†i .
i

Moreover the map is trace-preserving if no loss of the particle can occur. A natural way of describing the class of CP-
maps is by using the duality between maps and states, first observed by Jamiolkowski [5] and since then rediscovered
by many. We review some nice properties of CP-maps based on this dual description, and show how to obtain the
extreme points of the convex set of trace-preserving CP-maps.
The dual state is defined on a Hilbert space that is the tensor product of two times the original Hilbert space on
which the map acts, and is therefore naturally endowed with a notion of entanglement. Unitary evolution for example
corresponds to maximal correlations between the in- and output state, and this kind of evolution leads to a dual
state that is maximally entangled. We will show how normal forms derived for entangled states lead to interesting
parameterizations of CP-maps, and will discuss some issues concerning the use of quantum channels to distribute
entanglement.
It thus turns out that the techniques developed for describing entanglement can directly be applied for describing
the evolution of a quantum system. Concepts as quantum steering and teleportation have a direct counterpart. A
quantum channel for example will be useful for distributing entanglement if and only if the dual state associated to it
is entangled, and optimal decompositions of states as derived in the case of entanglement of formation will yield very
appealing parameterizations of quantum channels.

I. CHARACTERIZATION OF CP-MAPS

The most general evolution of a quantum system is described by a linear CP-map [3]. In this section we will give
a self-contained description of CP-maps or quantum channels. Most of the mathematics presented originate from the
seminal papers of de Pillis [6] and Choi [4]. The fact that the evolution of quantum systems is described by linear
completely positive maps is a consequence of the assumption of the linearity of the evolution (the complete positivity
follows from consistency arguments once the linearity is accepted).
Let us now recall some notations and useful tricks. Consider a pure state |χi in a Hilbert space that is a tensor
 2

product of two Hilbert spaces of dimension n

n
X
|Ai = aij |ii|ji.
ij

Define
n
X
|Ii = |ii|ii
i

an unnormalized maximally entangled state and A the operator with elements hi|A|ji = aij , then
|Ai = A ⊗ In |Ii.
Moreover it holds that
X ⊗ Y |Ai = XA ⊗ Y |Ii = XAY T ⊗ In |Ii = In ⊗ Y AT X T |Ii.
P
The symbol |Ii will solely be used to denote the unnormalized maximally entangled state |Ii = i |iii. We are
now ready for the following fundamental Theorem of de Pillis[6]:
Theorem 1 A linear map Φ acting on a matrix X is Hermitian-preserving if and only if there exist operators {Ai }
and real numbers λi such that
X
Φ(X) = λi Ai XA†i
i

Proof: Suppose the map Φ acts on a n × n matrix. Then due to linearity, Φ is completely characterized if we know
how it acts on a complete basis of n × n matrices, for example on all matrices |ei ihej |, 1 ≤ i, j ≤ n with |ei i a complete
orthonormal base in Hilbert space. Let us define the n2 × n2 positive matrix
 
|e1 ihe1 | · · · |e1 ihen |
|IihI| =  · · · ··· ··· , (1)
|en ihe1 | · · · |en ihen |
being the matrix notation of a maximally entangled state in a n ⊗ n Hilbert space. It follows that all the information
of a map Φ is encoded in the state
ρΦ = In ⊗ Φ(|IihI|), (2)
2
as the n n × n blocks represent exactly the action of the map on the complete basis |ei ihej |. If Φ is Hermitian-
preserving, then Φ(|ei ihej |) has to be equal to the Hermitian conjugate of Φ(|e
P j ihei |), and this implies that ρΦ is
Hermitian. Let us therefore consider the eigenvalue decomposition of ρΦ = i λi |χi ihχi |. Using the trick |Ai =
P
(A ⊗ I)|Ii, we easily arrive at the conclusion that Φ(X) = i λi Ai XA†i , where {λi } are the eigenvalues and where
T
the operators {Ai } are the reshaped versions of the eigenvectors of ρΦ .
A central ingredient in the proof was the introduction of the matrix
ρΦ = In ⊗ Φ(|IihI|)
P
with |Ii = i |ii|ii a maximally entangled state. We define this Hermitian matrix ρΦ as being the dual state
corresponding to the map Φ. It was already explained that it encodes all the information about the map, and its
eigenvectors give rise to the operators Ai . The above lemma characterizes all possible Hermitian preserving maps,
and therefore surely all positive and completely positive maps. For example, let us consider the positive map that
corresponds to taking the transpose of the density operators of a qubit:
 
1 0
λ1 = 1 A1 = (3)
0 0
 
0 0
λ2 = 1 A2 = (4)
0 1
 
0 1 √
λ3 = 1 A3 = / 2 (5)
1 0
 
0 1 √
λ4 = −1 A4 = / 2 (6)
−1 0
 3

Not all Hermitian-preserving maps are physical in quantum mechanics however: if a map acts on a subsystem,
then it should conserve positivity of the complete density operator. This extra assumption leads to the condition of
complete positivity, meaning that Im ⊗ Φ is positive for all m. Of course, this implies that the dual state ρΦ is not
only Hermitian but also positive (i.e. all its eigenvalues are positive), as it is defined as the action of the map In ⊗ Φ
on a maximally entangled state. The positive eigenvalues can then be absorbed into the (Kraus) operators {Ai }, and
we have therefore proven the Kraus representation Theorem (Choi[4]):
Theorem 2 A linear map Φ acting on a density operator ρ is completely positive if and only if there exist operators
{Ai } such that
X
Φ(ρ) = Ai ρA†i .
i

Remarks:
P
• A CP-map is trace-preserving iff i A†i Ai = In ; this property is easily verified using the cyclicity of the trace.
In terms of the (unique) dual state ρΦ associated to the map Φ, this trace-preserving condition amounts to:

T r2 (ρΦ ) = In .

Here the notation T r2 means the partial trace over the second subsystem. A CP-map is furthermore called
bistochastic if also the condition

T r1 (ρΦ ) = In

holds; this property is equivalent to the fact that the map is identity-preserving, i.e. Φ(In ) = In .
• The dual state ρΦ corresponding to a CP-map Φ is uniquely defined. The Kraus operators are obtained by
considering the columns of a square root of ρΦ (Ai is obtained by making a matrix out of the i’th column of a
square root of X, with ρΦ = XX † ). As the square root of a matrix is not uniquely defined, the Kraus operators
are not unique. Each different “square root” X of ρΦ (ρΦ = XX † ) gives rise to a different set of equivalent
Kraus operators. This implies that all equivalent sets of Kraus operators are related by an isometry, and that
the minimal number of Kraus operators is given by the rank of the density operator ρΦ . Therefore we define the
rank of a map to be the rank of the dual operator ρΦ . This rank is bounded above by n2 with n the dimension of
the Hilbert space. A unique Kraus representation can be obtained by for example enforcing the Kraus operators
to be orthogonal, as these would correspond to the unique eigenvectors of ρΦ . Note that a similar reasoning
applies to all Hermitian preserving and all positive maps, although there an additional sign should be taken into
account.
• By construction, we have proven that a map Φ acting on a n-dimensional Hilbert space is completely positive
iff In ⊗ Φ is positive: there is no need to consider auxiliary Hilbert spaces with dimension larger than the
original one. The reasoning is as follows: if In ⊗ Φ is positive, then ρΦ is positive, and therefore Φ has a Kraus
representation, which implies complete positivity.
• Suppose Φ is positive but not completely positive. Then there exists a completely positive map Φ̃ and a positive
scalar ǫ such that

Φ(ρ) = (1 + nǫ)Φ̃(ρ) − ǫTr(ρ)In .

The proof of this fact is elementary: take ǫ to be the opposite of the smallest eigenvalue of ρΦ (this eigenvalue is
negative as otherwise Φ would be completely positive), and define the CP-map Φ̃(ρ) = (Φ(ρ)+ nǫTr(ρ)I/n)/(1 +
nǫ) (this map is completely positive because the dual state AΦ̃ associated to it is positive and has therefore a
Kraus representation). Note that the whole reasoning is also valid for general Hermitian-preserving maps. As
an example, consider again the transpose map on a qubit. Then it can be checked that the minimal value of ǫ
is 1 (this is true for the PT operation in arbitrary dimensions) and that the Kraus operators corresponding to
Φ̃ become
r   r   r  
2 1 0 2 0 0 1 0 1 √
{Ai } = { , , / 2}.
3 0 0 3 0 1 3 1 0
 4

• To make the duality between maps and states more explicit, it is useful to consider the following identity:
 
Φ(ρ) = T r2 ρTΦ1 (ρ ⊗ In ) , (7)

where T1 means partial transposition with relation to the first subsystem. This can be proven by explicitly
writing the map Φ into Kraus operator form, and exploiting the cyclicity of the trace. Due to the partial
transpose condition of Peres [7], it is clear that ρTΦ1 will typically not longer be positive. This identity is very
useful, and was used in the section on optimal teleportation with mixed states.

II. EXTREME POINTS OF CP-MAPS

The set of completely positive maps is a convex set: indeed, if Φ1 and Φ2 are CP-maps, then so is xΦ1 + (1 − x)Φ2 .
Due to the one to one correspondence between maps Φ and states ρΦ , it is trivial to obtain the extreme points of the
set of completely positive maps: these are the maps with one Kraus operator, corresponding to ρΦ having rank 1.
If however we consider the convex set of trace-preserving maps, the characterization of extreme points becomes
more complicated. The knowledge of the set of extreme points of the trace-preserving CP-maps is very interesting
from a physical perspective in the following way: suppose one has a multipartite state of qudits and one wants to
maximize some convex functional of the state (e.g. the fidelity, ...) by performing local operations. Due to convexity,
the optimal operation will correspond to an extreme point of the set of trace-preserving maps.
Let us now characterize all extremal trace-preserving maps:
Theorem 3 Consider a TPCP-map Φ acting on a Hilbert space of dimension n and of rank m. Consider the dual
state ρΦ = XX † with X a n2 × m matrix, and the n2 matrices Xi = X † (σi ⊗ In )X (the matrices {σi } form a complete
basis for the Hermitian n × n matrices). Then Φ is extremal if and only if m ≤ n and if the set of linear equations
∀i : Tr(QXi ) = 0 has only the trivial solution Q = 0.
This condition is equivalent to the following one given by Choi[4]: given m2 Kraus operators {Ai } of a map Φ, then
the map is extremal iff the m2 matrices {A†i Aj }, 1 ≤ i, j ≤ m are linearly independent.
Proof: The map Φ is extremal if and only if there does not exist a R with the property that RR† 6= I and such that
Tr2 (XRR† X † ) = I. This condition is equivalent to the fact that the set of equations
 

Tr X (RR† − I) X † σi ⊗ I  = 0
 
| {z }
Q

does only have the trivial solution Q = 0. As there are n2 independent generators σi and due to the fact that Q has
m2 degrees of freedom, it is immediately clear that there will always be a non-trivial solution if m > n, ending the
proof.
It remains to be proven that he condition obtained is equivalent to the one derived[43] by Choi [4]. This can be seen
P P
as follows: the condition Tr2 (XRR† X † ) = I is equivalent to the condition jk A†k Aj ( i Rji Rki ∗
− δjk ) = 0 (this is
readily obtained using the trick |Ai = A ⊗ I|Ii). Therefore a nontrivial solution of Q is possible iff the set of matrices
{A†i Aj }, 1 ≤ i, j ≤ m are linearly dependent.
Note that the given proof is constructive and can therefore be used for decomposing a given TPCP-map into a
convex combination of extremal maps: √ once a non-trivial Q and therefore R is obtained, one can scale it such that
RR† ≤ I, and define another S = I − RR† . This S is guaranteed to be another trace-preserving map up to a
constant factor, and the original map is the sum of the maps parameterized by XRR† X † and XSS † X † .
All TPCP maps Φ of rank 1 are of course extreme and correspond to unitary dynamics. One easily verifies that this
implies that the dual ρΦ is a maximally entangled state. The intuition behind this is as follows: by equation (7), ρΦ
characterizes the correlation between the output and the input of the channel. Maximal correlation happens iff the
evolution occurs reversibly and thus unitarily, and therefore corresponds to maximal “entanglement” between in- and
output. We will explore this connection between maps and entanglement more thoroughly in the following section.
One could go one step further, and try to characterize all extreme points of the convex set defined by all trace-
preserving channels for which the extra condition holds that Φ(ρ1 ) = ρ2 with ρ1 and ρ2 given density operators. (Note
that ρ1 and ρ2 can be chosen completely arbitrary, as there will always exist at least one TPCP-map that transforms
a given state into another given one: consider for example the map with its associated dual state ρΦ = I ⊗ ρ2 .)
Bistochastic channels are a special subset of this convex set of maps (in that case ρ1 = ρ2 ≃ I). An adaption of
Theorem 3 leads to the following:
 5

Theorem 4 Consider the convex set of trace-preserving CP-maps Φ for which Φ(ρ1 ) = ρ2 with ρ1 , ρ2 given. Suppose
Φ is of rank m, its dual state is ρΦ = XX † with X a n × n matrix, and that there are m Kraus operators {Ai }. Then
this map is extremal if and only if the set of 2m2 linear equations

Tr(QX † (σi ⊗ I)X) = 0 Tr(QX † (ρT1 ⊗ σi )X) = 0 (8)

has only the trivial solution Q = 0, or equivalently if and only if the m2 operators {A†i Aj ⊕ Aj ρ1 A†i } (1 ≤ i, j ≤ m)
are linearly independent.

Proof: The proof is completely analogous to the proof of Theorem 3, but here we have the extra condition


Tr X(RR† − I)X(ρT1 ⊗ σi ) = 0.

In terms of Kraus operators, this additional condition becomes

X X
Aj ρ1 A†k ( ∗
Rji Rki − δjk ) = 0
kj i

which ends the proof.

A similar Theorem was stated by Landau and Streater [8] in the special case of bistochastic maps. In analogy with
the conclusions of Theorem 3, we conclude that the√number of Kraus operators in an extremal TPCP-map of the kind
considered in the above Theorem is bounded by ⌊ 2n2 ⌋.
Let us for example consider the case of qubits. Then the rank of an extremal Φ is bounded by 2, and extremal rank
2 TPCP-maps obeying the condition Φ(ρ1 ) = ρ2 typically exist. There is however a notable exception if ρ1 = ρ2 = I/2
(i.e. when Φ is bistochastic): a bistochastic qubit map has a corresponding dual ρΦ that is Bell-diagonal. A Bell-
diagonal state is a convex sum of maximally entangled states, and therefore a rank 2 bistochastic map cannot be
extremal. Note however that this is an accident, and for Hilbert space dimensions larger than 2 there exist extremal
bistochastic channels that are not unitary [8]. Sometimes the name “unital” is also used instead of “bistochastic”.
The foregoing argument however shows that this terminology is not completely justified.
One could now add more constraints Φ(ρ2i ) = ρ2i+1 , and this would lead to similar conditions for extremality in
terms of the Kraus operators. Note however that the ρi appearing in the constraints cannot be chosen completely
arbitrary, as in general non-compatible constraints can arise due to the complete positivity condition on the physical
maps (Deciding whether a set of conditions Φ(ρ2i ) = ρ2i+1 is physical can be solved using the techniques of semidefinite
programming [9]).
Let us now formulate another interesting Theorem:
Theorem 5 Given a Hilbert space of dimension n and a trace-preserving map Φ of rank m ≤ n, then there exist pure
states |ψi such that Φ(|ψihψ|) are states of rank m − 1.
Proof: Let us first consider the case m = n, and define m Kraus operators {Ai } corresponding to Φ. Given a pure
state |ψi, then Φ maps this state to one that is not full rank iff there exists a pure state |χi such that
X
hχ|Φ(|ψihψ|)|χi = 0 = |hχ|Ai |ψi|2 .
i
P P
Writing |χi = i yi |ii, |ψi = i xi |ii and hj|Ai |ki = Ajik , then the previous equation amounts to solving the following
set of bilinear equations:
n m=n
X X
∀i = 1 : n, ( xj Ajik )yk = 0.
k=1 j=1

This set of equations always has a non-trivial solution. Indeed, the parameters xj can always be chosen such that
P
the matrix à = j xj Ajik is singular (if all Ai are full rank then this can be done by fixing all but one of them, and
then choosing the remaining parameter such that the determinant vanishes; if one of the Ai is rank deficient then
the solution is of course direct). Then the parameters yk can be chosen such that the vector y is in the right kernel
of à (the right kernel is not zero-dimensional as the dimension of the matrix à is n × n), and therefore Φ(|ψihψ|) is
not full rank. If m < n, then the right kernel of à is at least n − m + 1 dimensional, such that n − m + 1 linearly
independent |χi can be found such that hχ|Φ(|ψihψ|)|χi = 0, which ends the proof.
In general , it is thus proven that one can always find states |ψi such that the rank of Φ(|ψihψ|) is smaller than the
rank of the map, which is surprising. Note that the bound in the Theorem is generically tight, i.e. the minimal rank
 6

of the output state will typically be m − 1; this follows from the fact that decreasing the rank of the matrix à with
two units would need n(n − 1)/2 independent degrees of freedom, while there are only n − 1 available.
Note that extremal TPCP-maps always fulfil the conditions of the Theorem. In particular, extremal qubit channels
are generically of rank 2, and the previous Theorem implies that there always exist pure states that remain pure after
the action of a rank 2 extremal map (This was also observed by Ruskai et al.[10]).
The above Theorem has also some consequences for the study of entanglement. Applying the foregoing proof to the
dual state ρΦ , we can easily prove the following: if the rank of a mixed state ρ defined in a n × n dimensional Hilbert
space is given by m ≤ n, then there always exist at least (n − m + 1) linearly independent product states orthogonal
to it.
Let us now consider an example of the use of extremal maps. Suppose we want to characterize the optimal local
trace-preserving operations that one has to apply locally to each of the qubits of a 2-qubit entangled mixed state,
such as to maximize the fidelity (i.e. the overlap with a maximally entangled state). This problem is of interest in the
context of teleportation [11, 12] as the fidelity of the state used to teleport is the standard measure of the quality of
teleportation. Badziag and the Horodecki’s [13] discovered the intriguing property that the fidelity of a mixed state
can be enhanced by applying an amplitude damping channel to one of the qubits. This is due to the fact that the
fidelity is both dependent on the quantum correlations and on the classical correlations, and enhancing the classical
correlations by mixing (and hence losing quantum correlations) can sometimes lead to a higher fidelity.
With the help of the previous analysis of extremal maps, we are in the right position to find the optimal trace-
preserving map that maximizes the fidelity. Indeed, the optimization problem is to find the trace-preserving CP-maps
ΦA , ΦB such as to maximize the fidelity F defined as
n  o
F (ρ, ΦA , ΦB ) = hψ|ΦA ⊗ ΦB (ρ)|ψi = Tr ρ Φ†A ⊗ Φ†B (|ψihψ|) (9)

with |ψi the maximally entangled state. This problem is readily seen to be jointly convex in ΦA and ΦB , and therefore
the optimal strategy will certainly consist of applying extremal (rank 2) maps ΦA , ΦB . As we just have derived an
easy parametrization of these maps, it is easy to devise a numerical algorithm that will yield the optimal solution.
Note that the problem, although convex in ΦA and ΦB , is bilinear and therefore can have multiple (local) maxima.
This problem disappears when only one party (Alice or Bob) applies a map (i.e. ΦB = I). This problem was
studied in more detail by Rehacek et al.[14], where a heuristic algorithm was proposed to find the optimal local
trace-preserving map to be applied by Bob. As the optimization problem is however convex, the powerful techniques
of semidefinite programming [9] should be applied, for which an efficient algorithm exists that is assured to converge
to the global optimum. Indeed, due to linearity the problem now consists of finding the 2-qubit state ρΦ† ≥ 0 with
constraint TrB (ρΦ† ) = I such that the fidelity is maximized. As we already know, the algorithm will converge to
a ρΦ of maximal rank 2 in the case of qubits. Exactly the same reasoning holds for systems in higher dimensional
Hilbert spaces: if only one party is to apply a trace-preserving operation to enhance the fidelity, the above semidefinite
program will produce the optimal local map that maximally enhances the fidelity.
Other situations in which extremal maps will be encountered are for example the problem of optimal cloning[15,
16, 17]: given an unknown input state ρ, one wants to construct the optimal trace-preserving CP-map such as to yield
an output for which the fidelity with ρ ⊗ ρ is maximal. This can again be rephrased as a semidefinite program whose
unique solution will be given by an extremal trace-preserving CP-map.

III. QUANTUM CHANNELS AND ENTANGLEMENT

The physical interpretation of the dual state corresponding to a CP-map or quantum channel is straightforward.
It is the density operator that corresponds to the state that can be made as follows: Alice prepares a maximally
entangled state |Ii, and sends one half of it to Bob through the channel Φ. This results into ρΦ .
A perfect quantum channel is unitary and the corresponding state ρΦ is a maximally entangled state. This corre-
sponds to the case of perfect transmission of qudits, and indeed a maximally entangled state is the state with perfect
quantum correlations. Consider now a completely depolarizing channel. In that case it is possible to transmit a
classical bit perfectly, and indeed ρΦ corresponds to a separable state with maximal classical correlations. As a third
example, consider the complete amplitude damping channel. Then ρΦ is a separable pure state with no correlations
whatever between Alice and Bob. It is therefore clear that the study of the character of correlation present in the
quantum state ρΦ tells us a lot about the character of the quantum channel.
This way of looking at quantum channels gives a nice way of unifying statics and dynamics in one framework: the
future is entangled (or at least correlated) with the past. Just as a measurement in the future gives us information
about the prepared system (through the use of the quantum Bayes rule), a measurement on Bob’s side enables Alice
to refine her knowledge of her local system (through the use of the quantum steering Theorem)[44]. It is therefore
 7

clear that the description of entanglement will shed new light on the question of describing correlations between the
states of the same system at two different instants of time, and vice-versa. Therefore we expect that many useful
results concerning entanglement can directly be applied to quantum channels. On the other hand, a lot of work has
been done concerning the quantification of the classical capacity of a quantum channel. These results offer a nice
starting point for the study of classical correlations present in a quantum state.

A. Quantum capacity

The quantum capacity of a quantum channel is related to the asymptotic number of uses of the channel needed for
obtaining states whose fidelity tends to one. To transmit quantum information with high fidelity, one indeed needs
almost perfect singlets. It is immediately clear that ideas of entanglement distillation will be crucial: sending one
part of an EPR through the channel will result in a mixed state, and these mixed states will have to be purified.
Let us first establish a result that was already intrinsically used by many [12, 18, 19, 20]:
Theorem 6 A quantum channel Φ can be used to distribute entanglement if and only if ρΦ is entangled. If ρΦ is
separable, then the Kraus operators of the map Φ can be chosen to be projectors, and the map Φ is entanglement
breaking.
Proof: The if part is obvious, as ρΦ is the state obtained by sending one part of a maximally entangled state through
the channel. To prove the only if part, assume that ρΦ is separable. Then all Kraus-operators can be chosen to be
projectors (corresponding to the decomposition with separable pure states), destroying all entanglement.
It is also possible to make a quantitative statement:
Theorem 7 Suppose we want to use the channel Φ to distribute entanglement by sending one part of an entangled
state through the channel. The maximal attainable fidelity (i.e. overlap with a maximally entangled state) corresponds
to the largest eigenvalue of ρΦ . This maximal fidelity is obtained if Alice sends one half of the state described by the
eigenvector of ρΦ corresponding to its largest eigenvalue.
Proof: Suppose Alice prepares the entangled state |χi and sends the second part to Bob through the channel Φ with
Kraus-operators {Ai }. We want to find the state |χi such that
X
hI| I ⊗ Ai |χihχ|I ⊗ A†i |Ii = hχ|ρΦ |χi (10)
i

is maximized, which immediately gives the stated result.

The above result is amazing: it tells us that it is not always the best strategy to send one part of a maximally
entangled state through the channel. It would be tempting to conjecture that the entanglement of distillation of the
obtained state represents the quantum capacity of the given channel.
Note that the eigenvalues and eigenvectors of ρΦ got an appealing interpretation: these represent the fidelities that
are obtained by sending one half of the eigenvectors through the channel. Note also that the reduction criterion
[21, 22],
I
I ⊗ Tr2 (ρΦ ) − ρΦ = − ρΦ
n
implies that ρΦ is entangled if its largest eigenvalue exceeds 1/n. This is of course in complete accordance with the
previous Theorem, as the maximal fidelity for a separable state is also given by 1/n.
A more sophisticated treatment of the quantum capacity of a quantum channel would involve ideas of coding and
of quantum error correction, although only partial results have been obtained yet; the following is an incomplete list
of papers where interesting results have been obtained [18, 23, 24, 25, 26, 27, 28].

B. Classical Capacity

Let us now move towards the well-studied problem of classical capacity of a quantum channel. The central result
is the Holevo- Schumacher- Westmoreland Theorem [29, 30], which tells us that the classical product state capacity
of a quantum channel Φ is given by
 
 X X 
χ(Φ) = max S(Φ( pj ρj )) − pj S(Φ(ρj )) . (11)
pj ,ρj  
j j
 8

Let us now ask the following question: what would be the analogy and the interpretation of this formula in the dual
picture of states ρΦ ? Using formula (7), it holds that

Φ(ρj ) = Tr1 (ρΦ (ρTj ⊗ I)).

Suppose Alice and Bob share the state ρΦ . Then the above formula describes how Bob has to update his local
density operator when Alice did a measurement with corresponding POVM-element ρTj . Reasoning along the lines of
the HSW-Theorem, the natural interpretation would now be that formula (11) will give us a measure of how much
(secret) classical randomness Alice and Bob can create using the state ρΦ : if Alice implements a POVM measurement
with elements {pj , ρTj }, this drives the system at Bob’s side into a particular direction, and a measurement of Bob
will reveal some information about the (random) outcome of Alice. Note that we interpret the presence of a bipartite
state as being a particular kind of quantum channel. Note that the question of creating shared randomness has also
been discussed in [31, 32].
The foregoing discussion suggests the following definition for the classical random correlations C cl present in a
quantum state ρ:
X
CBcl
(ρAB ) = max S(ρB ) − pj S(ρjB ) (12)
{Ej }
j
pj = Trρ(Ej ⊗ I) (13)
1
ρjB = Tr1 (ρ(Ej ⊗ I)) . (14)
pj

Here {Ej } presents the elements of the POVM implemented by Alice. Observe that there is an asymmetry in the
cl cl
definition, in that CA is not necessarily equal to CB . This definition coincides with the one given by Henderson
and Vedral [33], where they introduced this measure because it fulfilled the condition of monotonicity under local
operations.
In general, the classical mutual information obtained by the actions of Alice and Bob to obtain classical randomness
will be smaller than the derived quantity (12), as coding is needed to achieve the Shannon capacity. This coding
could be implemented by doing joint measurements, but we do not expect that the upper bound is tight; a better
rate could be obtained if also public classical communication is allowed (A. Winter, unpublished).

IV. ONE-QUBIT CHANNELS

In the case of qubit channels, much more explicit results can be obtained, due to the fact that we have a fairly
good insight into the properties of mixed states of two qubits. In this section we highlight some questions about qubit
channels that can be solved analytically.
Recall formula (7)
 
Φ(ρ) = T r1 ρTΦ1 (ρ ⊗ In ) (15)

which is almost exactly the same expression as if Alice were measuring the POVM-element ρ on the joint state ρΦ ;
the difference it that the partial transpose of this state has to be taken. It is now natural to look at the R-picture of
the dual state ρΦ associated to the map [34], where ρ is parameterized by a real 4 × 4 matrix

Rij = Tr (ρσi ⊗ σj ) ,

0 ≤ σi ≤ 3. In the R-representation, a partial transpose corresponds to a multiplication of the third column or row
with a minus sign. Let us therefore define RΦ to be the parameterization of ρTΦ1 in the R-picture, i.e. the R-picture of
ρΦ in which the third row is multiplied by −1. Note that the first row of RΦ is given by [1; 0; 0; 0], as this corresponds
to the trace-preserving condition.
If x is the Bloch vector corresponding, then the action of the map with corresponding ρTΦ1 or RΦ is the following:
   
1 1
= R Φ (16)
x′ x

. One can easily prove that the image of the Bloch sphere yields an ellipsoid, where the local density operator of
Alice is represented by the center of the ellipsoid. This implies that the knowledge of the ellipsoid corresponds to the
 9

complete knowledge of the quantum channel up to local unitaries at the input. (Note that not all ellipsoids correspond
to physical maps, but that there is some restriction on the ratio of the axis).
Let us now consider the analogue of local unitary (LU) and local filtering (SLOCC) equivalence classes as known
for mixed states of two qubits [34]. What we are looking for are normal forms Ω (where Ω is a map) such that
Φ(ρ) = BΩ(AρA† )B † with A, B ∈ SU (2) or ∈ SL(2, C).
The LU case is very easy: each RΦ can be brought into the unique form
 
1 0 0 0
x λ1 0 0 
RΦ = 
y 0 λ2 0 
z 0 0 ±λ3

by local unitary transformations, where λ1 ≥ λ2 ≥ |λ3 | and x, y ≥ 0; one just has to take the singular value
decomposition of the lower 3 × 3 block of R, taking into account that the orthogonal matrices have determinant +1
(see also Fujiwara and Algoet [35] and King and Ruskai [36] for a different approach but with the same result).
Let us next move to SLOCC equivalence classes; it is clear that the Lorentz singular value decomposition [34] is all
we need:

Theorem 8 Given a 1-qubit trace-preserving CP-map Φ and its dual RΦ . Then the SLOCC normal form Ω of RΦ
is proportional to one of the following unique normal forms:
   1 0√ 0 0
  
1 0 0 0 1 0 0 0
 0 s1 0 0   0 x/ 3 0√ 0  0 0 0 0
0 0 s 0     0 0 . (17)
2 0 0 x/ 3 0  0 0
0 0 0 s3 2/3 0 0 1/3 1 0 0 0

Here 1 ≥ s1 ≥ s2 ≥ |s3 |, 1 − s1 − s2 − s3 ≥ 0 and 0 ≤ x ≤ 1. For maps with a normal from of the first kind, one can
choose the Kraus operators equal to

{Ai } = {p0 Aσ0 B, p1 Aσ1 B, p2 Aσ2 B, p3 Aσ3 B} (18)

with A, B complex 2 × 2 matrices and pi ≥ 0, related to the {si } by the formula relating the eigenvalues of a Bell
diagonal state to its Lorentz singular values. The Kraus operators of maps with a normal form of the second kind can
be chosen to be of the form
r   r   r  
1+x 1 0 1−x 1 0 2 0 1
{Ai } = { A 0 √1
B, A 0 − √13 B, A B}, (19)
2 3 2 3 0 0

again with A, B complex 2 × 2 matrices. In the third case, the map is trivial as it maps everything to the same point.
{si }, x, A, B, {pi } can be calculated explicitly by calculating the Lorentz singular value decomposition of the state ρΦ .

Proof: The proof is immediate given the Lorentz singular value decomposition. The first case corresponds to a
diagonalizable R, and a diagonal R corresponds to a bistochastic channel. The second and third case correspond to
non-diagonalizable cases (note that there are 2 normal forms in the case of states that do not apply here as they
cannot lead to trace-preserving channels).
This gives a nice classification of all the classes of TPCP-maps on qubits: the generic class is the one that can be
brought into unital form by adding appropriate filtering transformations A, B, i.e. the ellipsoid can be continuously
deformed to an ellipsoid whose center is the maximally mixed state. The non-generic class however cannot be deformed
in this way: it is easy to show that the ellipsoid corresponding to the normal form touches the Bloch sphere at one
and only at one point; there is no filtering operation that can change this property. We conclude that the ellipsoids
in the non-generic case are not (and cannot be made by filtering operations) symmetric around the origin and that
they touch the Bloch sphere at exactly one point.
We depict both types of normal ellipsoids in figure 1. Note that this geometrical picture will be very useful in
guessing input states that maximize the classical capacity of the state (see e.g. [36]).

A. Extremal maps for qubits

In the case of a qubit channel Φ, the dual state ρΦ is a mixed state of two qubits. It is possible to obtain an explicit
parameterization of all extremal qubit maps (see also Ruskai et al. [10] for a different approach):
 10

FIG. 1: The image of a channel in generic normal form (left) or in non-generic normal form (right).

Theorem 9 The set of dual states ρΦ corresponding to extreme points of the set of completely positive trace preserving
maps Φ on 1 qubit is given by the union of all maximally entangled pure states, and all rank 2 states ρ for which
Tr2 (ρΦ ) is equal and Tr1 (ρΦ ) is not equal to the identity. The Kraus operators corresponding to the rank 1 extreme
points are unitary, while the ones corresponding to the rank 2 extreme points have a representation of the form:
   p 
s0 0 0 1 − s21
A1 = U V† A2 = U p V† (20)
0 s1 1 − s20 0
with U, V unitary.
Proof: We have already proven that extremal TPCP-maps have maximal rank 2. Due to the duality between maps
and states, it is sufficient to consider rank 2 density operators of two qubits ρΦ for which T r2 (ρΦ ) = I2 . A real
parameterization of all 2-qubit density operators ρ is given by the real 4 × 4 matrix R with coefficients
Rij = T r (ρσi ⊗ σj ) (21)
where 0 ≤ i, j ≤ 3. An appropriate choice of local unitary bases can always make the R1:3,1:3 block diagonal, and the
trace-preserving condition translates into R0,1:3 = 0. Therefore R is given by:
 
1 0 0 0
t λ 0 0 
R= 1 1 .
t2 0 λ2 0 
t3 0 0 λ3
The corresponding ρ is given by
 
1 + t3 + λ3 0 t1 − it2 λ1 − λ2
1 0 1 + t3 − λ3 λ1 + λ2 t1 − it2 
ρ=  ,
4 t 1 + it2 λ1 + λ 2 1 − t 3 − λ3 0
λ1 − λ2 t1 + it2 0 1 − t3 + λ3
and the positivity of ρ constrains the allowed range of the 6 parameters. Let us now impose that the rank of the
corresponding ρ is 2. This implies that linear combinations of 3 × 3 minors of ρ be zero, and after some algebra one
obtains the following conditions:
t3 (λ3 + λ1 λ2 ) = 0
t2 (λ2 + λ1 λ3 ) = 0
t1 (λ1 + λ2 λ3 ) = 0
These equations, supplemented with the fact that diagonal elements of a positive semidefinite matrix are always bigger
than the elements in the same column, lead to the conclusion that all ti but one have to be equal to zero if ρ is rank
2. Without loss of generality, we can choose t1 = t2 = 0 and parameterize λ1 = cos(α), λ2 = cos(β). We thus arrive
at the canonical form
 
1 0 0 0
 0 cos(α) 0 0 
R= . (22)
0 0 cos(β) 0
sin(α) sin(β) 0 0 − cos(α) cos(β)
 11

Suppose that sin(α) sin(β) = 0 (this condition is equivalent to T r1 (ρΦ ) = I/2. Then the state corresponding to this
R is Bell-diagonal and thus a convex sum of two maximally entangled states, and therefore the map corresponding
to this state cannot be extremal. In the other case, an extremal
p rank 2 TPCP-map is obtained,
p which can easily be
shown to yield the given Kraus representation, where s0 = 1 − cos(α + β)/2 and s1 = 1 − cos(α − β)/2.
Note that the corresponding Theorem for bistochastic qubit channels is not very useful, as extremal TPCP qubit
channels are always unitary. Theorem 5 however is very interesting, and indicates that there always exist pure states
that remain pure after the action of the extremal qubit channel: indeed, ifp the basis vectors
p {|ii} are chosen according
to the unitary V in (20), then it is easily checked that the states |ψi ≃ s2 1 − s22 |0i ± s1 1 − s21 remain pure by the
action of the extremal map. Note that these two states are the only ones with this property, and note also that they
are not orthogonal to each other.

B. Quantum capacity

Let us now move on to the relation between 1-qubit quantum channels and entanglement. We can now make use
of the plethora of results derived for mixed states of two qubits. Let us first consider Theorem 6 about entanglement
breaking channels. In the case of mixed states of two qubits, a state is entangled iff it violates the reduction criterion
I ⊗ ρB − ρ ≥ 0. But in the case of the dual state ρΦ , it holds that ρB = I/2, and therefore it holds that a quantum
channel Φ can be used to distribute entanglement iff the maximal eigenvalue of ρΦ exceeds 1/2 [45]. In the light of
Theorem 7, it follows that such a non-entanglement breaking channel can always be used to distribute an entangled
state with fidelity larger than 1/2, which implies on its turn that it can be used to distill entanglement[18].
Consider now an entanglement breaking channel, i.e. a channel for which ρΦ is separable. In this case all the Kraus
operators can be chosen to be projectors. An explicit way of calculating this Kraus representation exists. Indeed, in
the section about entanglement of formation of two qubits, a constructive way of decomposing a separable mixed state
of two qubits as a convex combination of separable pure states was given. It was furthermore proven that a separable
state of rank 2 or 4 can always be written as a convex combination of 2 respectively 4 separable pure states, thus giving
rise to 2 respectively 4 rank one Kraus operators. Surprisingly, most separable rank 3 mixed states of two qubits
can only be written as a convex combination of 4 separable pure states. This implies that a generic entanglement
breaking channel of rank 3 needs 4 Kraus operators if these are to be chosen rank 1. Let us also mention that the
set of separable states is not of measure zero, implying that the set of entanglement breaking channels is also not of
measure zero.
The results of Wootters [37] can of course also be applied to non-entanglement-breaking channels. A direct appli-
cation of the formalism of Wootters yields the following Theorem:

Theorem 10 Given a 1-qubit channel Φ and the state ρΦ associated to it. If C is the concurrence of ρΦ , then the
channel has a Kraus representation of the form:
X
Φ(ρ) = pi (Ui C̃Vi )ρ(Ui C̃Vi )† (23)
i
√ √ 
1 1+C + 1−C √ 0√
C̃ = (24)
2 0 1+C − 1−C

where Ui , Vi are unitary matrices.

Proof: The Theorem is a direct consequence of the fact that a mixed state with concurrence C can be written as a
convex sum of pure states all with concurrence equal to C.
The geometrical meaning in the context of channels is the following: each trace-preserving CP-map is a convex
combination of contractive maps in unique different directions, where each contraction has the same magnitude.
Let us next address the question of calculating the quantum capacity of the one-qubit channel. Clearly, Theorem
7 tells us what states to send through the channel such as to maximize the fidelity of the shared entangled states. In
general, the quantum capacity cannot be calculated as we even don’t have a way of calculating the entanglement of
distillation of mixed states of two qubits (which is a simpler problem).
In the case of unital channels of rank 2 however, the eigenvectors of ρΦ are maximally entangled and the quantum
capacity can be calculated explicitly:

Theorem 11 Consider a bistochastic qubit channel Φ of rank 2. Then its quantum capacity is given by CQ = 1−H(p),
where p is the maximal eigenvalue of ρΦ and H(p) = −p log2 (p) − (1 − p) log2 (1 − p).
 12

Proof: A unital qubit channel exhibits the nice property that no loss whatever occurs by sending a maximally entangled
state through the channel: it can easily be shown (see Bennett et al.[18]) that sending a quantum system through
the channel is equivalent to using the standard teleportation channel induced by the (non-maximally entangled state)
ρΦ . Because we can use the state ρΦ , obtained by sending a Bell state through the channel, to perfectly simulate the
channel, this is clearly the optimal thing to do, and the quantum capacity of the channel is therefore equal to the
distillable entanglement of ρΦ . Now Rains [38] has proven that the distillable entanglement of a Bell diagonal state
of rank 2 is given by Edist (ρ) = 1 − S(ρ), which ends the proof of the Theorem.
More general, the quantum capacity of bistochastic qubit channel is always equal to the entanglement of distillation
of the corresponding dual states (due to the arguments in the previous proof).
As a last remark, we observe that the channels of the non-generic kind that touch the Bloch sphere at exactly one
point are never entanglement-breaking: this follows from the fact that the concurrence of ρΦ always exceeds 0 in that
case.

C. Classical capacity

Far more progress has been made concerning the classical capacity of quantum channels: it is known that the
classical capacity using product inputs is given by the Holevo-χ quantity. Here the geometrical picture derived in
section 6.4 can sharpen our intuition. Consider for example the case of a unital channel. It is immediately clear that
Holevo-χ will be maximized by choosing a mixture of two states that lie on the opposite side of the major axis of
the ellipsoid. This implies that the optimal input states are orthogonal. King and Ruskai [36, 39] even proved that
entangled inputs cannot help in the case of unital channels, and we conclude that the classical capacity of the unital
channels is completely understood.
Consider however a non-unital channel of the generic kind. As proven before, this channel can be interpreted as the
succession of a filter, a unital channel, and another filter. The critical source of noise or decoherence and irreversibility
in a channel is the mixing, and the previous analysis tells us that this mixing can always be interpreted to happen in
a unital way, whereas the in- and output of the unital channel is reversibly but non-orthogonally filtered. It follows
that orthogonal inputs will not appear orthogonally in the unital channel, and typically orthogonal inputs will not
achieve capacity. This strange fact was indeed discovered by Fuchs [40], and it appears to be generic for non-unital
channels.
Let us now have a look at the non-generic family of channels, whose ellipsoids touch the Bloch sphere at exactly
one point. It happens that the so-called stretched channel belongs to this family, and this channel has the property
that its (product) capacity is only achieved for an input ensemble with three states[41]. This is surprising but not too
surprising given the geometrical picture, as one of the input states corresponds to the pure output state, while the
other two ones are chosen to lie symmetric around the axis connecting the maximally entangled state with the pure
output state. Note however that most of the non-generic states achieve capacity with 2 input states.
Let us now move to calculate the classical capacity of the extremal qubit channels. In the case of extremal qubit
channels, it is possible to reduce the problem of calculating the classical (Holevo) capacity to an optimization problem
over the ensemble average. The problem to be solved is as follows: find the optimal ensemble {ρi , pi } such that
X X
S( pi Φ(ρi )) − pi S(Φ(ρi ))
i i

is maximized. We assume that Φ is rank 2 and therefore has a Kraus representation of the form (20). It is clear that
only pure states {ρi } have to be considered. It is easily seen that in the case of qubits, the entropy of ap
state is a convex
monotonously increasing function of the determinant of the density operator: S(ρ) = H(1/2(1 − 1 − 4 det(ρ)2 ))
with H(p) = p log(p) + (1 − p) log(1 − p) the Shannon entropy function. Inspired by the analysis of 2-qubit channels
by Uhlmann in terms of anti-linear operators [42], we make the following observation:
 
det A1 |ψihψ|A†1 + A2 |ψihψ|A†2 = |ψ T (AT1 σy A2 − AT2 σy A1 )ψ|. (25)

Here ψ is the vector notation (in the computational basis) of |ψi, and σy is a Pauli matrix. Suppose now that we
add an additional constraint to the problem, namely that the ensemble average ρ is given. Taking a square root
X of ρ = XX † , all possible pure state decompositions can be written as X ′ = XU with U an arbitrary isometry
(note that the columns of XU represent all unnormalized pure states in the decomposition). With this additional
constraint, the problem can be solved exactly as we solved the entanglement of formation problem. A constructive
way of obtaining the optimal decomposition of ρ is as follows: take a square root X of ρ, and calculate the singular
value decomposition of the symmetric matrix X T (AT1 σy A2 − AT2 σy A1 )X = V ΣV T . Call C = σ1 − σ2 the concurrence
 13

with {σi } the singular values of the above symmetric matrix. Then the optimal decomposition is obtained by
choosing U = V ∗ O with O the real orthogonal matrix that is chosen such that the diagonal entries of the matrix
R = OT (Diag[σ1 , −σ2 ] − Cρ)O) vanish. For given ensemble average ρ, the classical capacity is therefore given by the
following formula: S(Φ(ρ)) − f (C) (see also Uhlmann [42]).

To derive an explicit formula for the classical capacity of the extremal channels, we still have to do an optimization
over all possible ensemble averages ρ. Note that the previous analysis already learned us that the capacity will always
be reached with an ensemble of two input states. Both the terms Φ(ρ) and C can easily be extremized separately,
but unfortunately even if the eigenvalues of ρ are fixed, the optimal eigenvectors for maximizing S(ρ) and minimizing
C are not compatible. However, the capacity can easily be calculated numerically, as it just an optimization problem
over three real parameters.
On the other hand, we have seen that the definition of the classical capacity had a direct counterpart in giving an
appealing definition for the number of classical correlations present in a (mixed) bipartite state Ccl (see 12). The
techniques used in the foregoing paragraph are perfectly adequate to give an exact expression of this quantity if the
shared quantum state is a rank 2 bipartite state ρ of qubits. Indeed, a mixed bipartite state of two qubits can just
be seen as a more general kind of quantum channel.

V. CONCLUSION

We have shown that the natural description of quantum channels or positive linear maps is given by a dual quantum
state associated to the map. This dual state is defined over a Hilbert space that is naturally endowed with a tensor
product structure of the in- and output of the channel. We showed that the techniques developed in the context of
entanglement are of direct use in describing positive maps. We derived a characterization of the extreme points of the
convex set of trace-preserving completely positive maps, and gave some generalizations. We discussed some new results
about the classical and quantum capacity of a quantum channel, and in the case of one-qubit channels we showed
how to exploit the duality between qubit channels and mixed states of two qubits to obtain useful parameterizations.

[1] A. Einstein, B. Podolsky, and N. Rosen. Can quantum-mechanical description of physical reality be considered complete?
Phys. Rev., 47:777–780, 1935.
[2] E. Schrödinger. Discussion of probability distributions between separated systems. Proc. Camb. Phil. Soc., 31:555, 1935.
[3] K. Kraus. States, Effects and Operations: Fundamental Notions of Quantum Theory. Springer-Verlag, 1983.
[4] M.-D. Choi. Completely positive linear maps on complex matrices. Linear Algebra and Its Applications, 10:285–290, 1975.
[5] A. Jamiolkowski. Linear transformations which preserve trace and positive semidefiniteness of operators. Rev. of Mod.
Phys., 3:275–278, 1972.
[6] J. dePillis. Linear transformations which preserve hermitian and positive semidefinite operators. Pacific J. Math., 23:129,
1967.
[7] A. Peres. Separability criterion for density matrices. Phys. Rev. Lett., 77:1413–1415, 1996.
[8] L.J. Landau and R.F. Streater. On birkhoff’s theorem for doubly stochastic completely positive maps of matrix algebras.
Lin. Alg. Appl., 193:107, 1993.
[9] L. Vandenberghe and S. Boyd. Semidefinite programming. SIAM Review, 38:49, 1996.
[10] M.B. Ruskai, S. Szarek, and E. Werner. An analysis of completely-positive trace-preserving maps on 2x2 matrices. Lin.
Alg. Appl., 347:159–187, 2002.
[11] C.H. Bennett, G. Brassard, C. Crépeau, R. Jozsa, A. Peres, and W.K. Wootters. Teleporting an unknown quantum state
via dual classical and Einstein-Podolsky-Rosen channels. Phys. Rev. Lett., 70:1895–1899, 1993.
[12] M. Horodecki, P. Horodecki, and R. Horodecki. General teleportation channel, singlet fraction, and quasidistillation. Phys.
Rev. A, 60:1888–1898, 1999.
[13] P. Badzia̧g, M. Horodecki, P. Horodecki, and R. Horodecki. Local environment can enhance fidelity of quantum telepor-
tation. Phys. Rev. A, 62:012311, 2000.
[14] J. Rehácek, Z. Hradil, J. Fiurásek, and C. Brukner. Designing optimum cp maps for quantum teleportation. Phys. Rev.
A, 64:060301, 2001.
[15] V. Bužek and M. Hillery. Quantum copying: Beyond the no-cloning theorem. Phys. Rev. A, 54:1844, 1996.
[16] N. Cerf. Pauli cloning of a quantum bit. Phys. Rev. Lett., 84:4497, 2000.
[17] K. Audenaert and B. De Moor. Optimizing completely positive maps using semidefinite programming. Phys. Rev. A,
65:030302, 2002.
[18] C.H. Bennett, D.P. DiVincenzo, J.A. Smolin, and W.K. Wootters. Mixed state entanglement and quantum error correction.
Phys. Rev. A, 54:3824–3851, 1996.
[19] E.M. Rains. A semidefinite program for distillable entanglement. IEEE Trans. on Inf. Theory, 47:2921, 2001.
 14

[20] I. Cirac, W. Dür, B. Kraus, and M. Lewenstein. Entangling operations and their implementation using a small amount of
entanglement. Phys. Rev. Lett., 86:544, 2001.
[21] M. Horodecki and P. Horodecki. Reduction criterion of separability and limits for a class of distillation protocols. Phys.
Rev. A, 59:4206, 1999.
[22] N. J. Cerf, C. Adami, and R. M. Gingrich. Quantum conditional operator and a criterion for separability. Phys. Rev. A,
60:893–898, 1999.
[23] B. W. Schumacher. Sending entanglement through noisy quantum channels. Phys. Rev. A, 54:2614, 1996.
[24] D.P. DiVincenzo, P.W. Shor, and J.A. Smolin. Quantum-channel capacity of very noisy channels. Phys. Rev. A, 57:830,
1998.
[25] A. Winter. Coding theorems of quantum information theory. PhD thesis, Bielefeld University, 1999. quant-ph/9907077.
[26] H. Barnum, E. Knill, and M. A. Nielsen. On quantum fidelities and channel capacities. IEEE Trans. Inf. Theory, 46:1317,
2000.
[27] M. Hamada. Lower bounds on the quantum capacity and error exponent of general memoryless channels. quant-ph/0112103.
[28] M. Hamada. A lower bound on the quantum capacity of channels with correlated errors. quant-ph/0201056.
[29] A.S. Holevo. The capacity of quantum channel with general signal states. IEEE Trans. on Inf. Theory, 44:269, 1998.
[30] B. Schumacher and M. Westmoreland. Sending classical information via noisy quantum channels. Phys. Rev. A, 56:131–138,
1997.
[31] B.M. Terhal, M. Horodecki, D.W. Leung, and D.P. DiVincenzo. The entanglement of purification. quant-ph/0202044.
[32] N.J. Cerf, S.Massar, and S. Schneider. Multipartite classical and quantum secrecy monotones. quant-ph/0202103.
[33] L. Henderson and V. Vedral. Classical, quantum and total correlation. Jour. of Phys. A: Math. and Gen., 34(35):6899–6905,
2001.
[34] F. Verstraete, J. Dehaene, and B. De Moor. Local filtering operations on two qubits. Phys. Rev. A, 64:010101(R), 2001.
[35] A. Fujiwara and P. Algoet. Affine parameterization of completely positive maps on a matrix algebra. Phys. Rev. A,
59:3290, 1999.
[36] C. King and M.-B. Ruskai. Minimal entropy of states emerging from noisy quantum channels. IEEE Trans. on Inf. Theory,
47:192–209, 2001.
[37] W.K. Wootters. Entanglement of formation of an arbitrary state of two qubits. Phys. Rev. Lett., 80:2245, 1998.
[38] E. M. Rains. Rigorous treatment of distillable entanglement. Phys. Rev. A, 60:173, 179, 1999.
[39] C. King. Additivity for a class of unital qubit channels. quant-ph/0103156.
[40] C.A. Fuchs. Nonorthogonal quantum states maximize classical information capacity. Phys. Rev. Lett., 79:1162, 1997.
[41] C. King, M. Nathanson, and M.B. Ruskai. Qubit channels can require more than two inputs to achieve capacity. Phys.
Rev. Lett., 88:057901, 2002.
[42] A. Uhlmann. On 1-qubit channels. J. Phys. A, 34:7047, 2001.
[43] Actually, Choi derived the different problem of characterizing the extremal points of the (not necessarily trace-preserving)
CP-maps that leave the identity unaffected, but his arguments are readily translated to the present situation. Note also
that his proof was much more involved.
[44] In some sense one could argue that this was expected due to the fact that space and time play analogous roles in the
theory of relativity. It is very nice however that in the non-relativistic case considered here, the duality is already present.
This gives hope that it should be possible to generalize the current findings to the relativistic case.
[45] This was first observed by Michael Horodecki