Quantum information theory (MAT4430) Spring 2021

Lecture 17: The quantum capacity and channels for which it vanishes
Lecturer: Alexander Müller-Hermes

In this lecture we will start our discussion of the quantum capacity. Similar to the classical
capacity, the quantum capacity quantifies what communication rates are achievable for the
transmission of quantum information, i.e., of quantum states preserving their entanglement
with any reference system. To measure the communication error it will be convenient to use
the diamond-norm, which we introduced in the exercises.

1 Channel fidelity and diamond norm

Recall the so-called diamond norm, which we introduced in the exercises:
Definition 1.1 (Diamond norm). The diamond norm of a linear map L : B (HA ) → B (HB )
is defined as
kLk = sup kidn ⊗ Lk1→1 ,
n∈ N
where idn : B( C
n) → B(C n) denotes the identity map.
The following properties where proved in the exercises:
• For any linear map L : B (HA ) → B (HB ) we have

kLk = kidB(HA ) ⊗ Lk1→1 ≤ dim(HA )kLk1→1 .

• For linear maps L1 : B(HA1 ) → B(HB1 ) and L2 : B(HA2 ) → B(HB2 ) we have

kL1 ⊗ L2 k = kL1 k kL2 k .

• The transpose map ϑd : B( Cd) → B(Cd) satisfies

kϑd k = d.

We will often need to relate different measures of the distance between a quantum channel
and the identity channel. To prove the necessary estimates we will start with a lemma:
Lemma 1.2. For any linear map L : B(H) → B(H) we have

kLk1→1 ≤ 2 max{kL(|vihv|)k1 : |vi ∈ H, hv|vi = 1}.

Recall from Lecture 7 that there is a stronger statement when L is assumed to be positive.
Unfortunately, we need to deal with non-positive maps in the following.

Proof. Consider an operator X ∈ B(H) satisfying kXk1 = 1. Then, we have the decompo-
sition
X = H1 + iH2 ,
with selfadjoint operators H1 , H2 ∈ B(H)sa given by
1 
H1 = X + X† ,
2
1  
H2 = X − X† .
2i

1
It is easy to see that kH1 k1 ≤ 1 and kH2 k1 ≤ 1, and that

kL(X)k1 ≤ kL(H1 )k1 + kL(H2 )k1 . (1)

Next, we consider the spectral decomposition

dim(H)
X
H1 = λi |vi ihvi |,
i=1

with eigenvalues λi ∈ R and normalized vectors |vii ∈ H. Note that

dim(H)
X
|λi | = kH1 k1 ≤ 1,
i=1

and by the triangle inequality we conclude that

dim(H)
X
kL(H1 )k1 ≤ |λi |kL(|vi ihvi |)k1 ≤ max{kL(|vihv|)k1 : |vi ∈ H, hv|vi = 1}.
i=1

Repeating the same argument for H2 and combining these estimates with (1) finishes the
proof.

Using the previous estimate, we can show the following lemma:

Lemma 1.3. Let H denote a complex Euclidean space and T : B(H) → B(H) a quantum
channel. If we have
F (|vihv|, T (|vihv|)) ≥ 1 − ,
for every normalized |vi ∈ H, then we have

kidB(H) − T k ≤ 8(2)1/4 .

Proof. To prove this statement, we will convert several times between different fidelities and
trace norms. By the Fuchs-van-de-Graaf inequalities and Bernoulli’s inequality we have
1
1 − k|vihv| − T (|vihv|)k21 ≥ (1 − )2 ≥ 1 − 2,
4
and hence √
k|vihv| − T (|vihv|)k1 ≤ 2 2,
for every normalized |vi ∈ H. By Lemma 1.2 we find that
√
kidB(H) − T k1→1 ≤ 4 2.

We could now use a result from the exercises to derive an upper bound on the diamond
norm. Unfortunately, this bound contains a dimension factor, and to obtain a better bound,
we will convert back to a fidelity. Consider any |ψi ∈ H ⊗ H with hψ|ψi = 1. By the Schmidt
decomposition, we can write
Xd
|ψi = λi |ai i ⊗ |bi i,
i=1

2
where d = dim(H) and with Schmidt coefficients λi ∈ + R
0 satisfying
Pd 2
i=1 λi = 1 and
orthonormal bases {|ai i}di=1 and {|bi i}di=1 of H. We can check that

1 − F (|ψihψ|, (idB(H) ⊗ T )(|ψihψ|))2 = hψ|idB(H) ⊗ (idB(H) − T )(|ψihψ|)|ψi

d
X
λ2i λ2j |hbi | idB(H) − T (|bi ihbj |) |bj i|


=
i,j=1
√
≤ kidB(H) − T k1→1 ≤ 4 2,

where we used that the squares of Schmidt coefficients sum up to 1 and that



|hbi | idB(H) − T (|bi ihbj |) |bj i| ≤ k idB(H) − T (|bi ihbj |) k∞


≤ k idB(H) − T (|bi ihbj |) k1 ≤ kidB(H) − T k1→1 ,

for any i, j ∈ {1, . . . , d}. Next, we use the Fuchs-van-de-Graaf inequalities again to see that

k idB(H) ⊗ idB(H) − T (|ψihψ|) k1 ≤ 4(2)1/4 ,

for every normalized vector |ψi ∈ H ⊗ H. Finally, we apply Lemma 1.2 and a result from
the exercises to conclude
kidB(H) − T k ≤ 8(2)1/4 .

It should be noted that the constants in the previous lemma are not optimal. Using some
tricks (see Watrous) the final bound can be slightly improved to

kidB(H) − T k ≤ 2(2)1/4 .

However, these constants will not play any role in the following, since we will always consider
situations where  → 0 exponentially fast and where we are interested in the rates of this
exponential convergence.

2 Definition of the quantum capacity

As always, we begin with the definition of coding schemes:

Definition 2.1. Let T : B(HA ) → B(HB ) denote a quantum channel. An (n, m, δ)-coding
scheme for quantum communication over T is a pair of quantum channels

C
E : B ( 2 )⊗m → B(HA

⊗n
) and D : B(HB ⊗n
) → B ( 2 )⊗m ,


C
such that
kid⊗m
2 − D ◦ T ⊗n ◦ Ek ≤ δ.

In Figure 1 you can see a schematic visualizing a coding scheme for the transmission of
quantum information.

3
Figure 1: Coding scheme for the transmission of quantum information with reference system

Then, we may define the quantum capacity as follows:

Definition 2.2 (Quantum capacity). We call a rate R ≥ 0 achievable for quantum com-
N
munication over the quantum channel T : B(HA ) → B(HB ) if for every n ∈ , there exists
an (n, mn , δn )-coding scheme such that
mn
R = lim and lim δn = 0.
n→∞ n n→∞

The quantum capacity of a quantum channel T : B(HA ) → B(HB ) is given by

Q(T ) = sup{R ≥ 0 : R achievable rate for quantum communication}.

We will omit the proof of the following lemma since it is very similar to a lemma involving
the classical capacity of a quantum channel, which was proved in the exercises.

Lemma 2.3. For any quantum channel T : B(HA ) → B(HB ) and any k ∈ N we have
Q(T ⊗k ) = kQ(T ).

3 Channels with vanishing quantum capacity

We have seen that the constant quantum channels are the only quantum channels with van-
ishing classical capacity. In the case of the quantum capacity, the situation is very different:
There are at least two different reasons for a quantum channel to have zero quantum capacity
leading to two distinct classes of quantum channels with vanishing quantum capacity. At
the point of writing, it is not known whether there are quantum channels outside of these
classes for which the quantum capacity vanishes. Deciding whether there are such examples
is a major open problem in quantum Shannon theory!

3.1 The transposition bound

The following theorem gives an upper bound on the quantum capacity of a quantum channel:

Theorem 3.1 (The transposition bound). For any quantum channel T : B (HA ) → B (HB )
we have
Q(T ) ≤ log (kϑB ◦ T k ) .

4
Proof. Note that the diamond norm of the transpose map ϑB : B(HB ) → B(HB ) is given by
kϑB k = dim(HB ). Consider now an (n, m, δ)-coding scheme for quantum communication
over the quantum channel
T : B (HA ) → B
(HB ) given by
a pair of quantum channels
C
E : B ( 2 )⊗m → B HA

⊗m
and D : B HA⊗m
C
→ B ( 2 )⊗m satisfying

δ = kid⊗m
2 − D ◦ T ⊗n ◦ Ek .

Using multiplicativity of the diamond norm and the triangle inequality, we may compute
⊗m
2m = kϑ2 km
 = kϑ2 k
= kϑ⊗m ◦ id⊗m − D ◦ T ⊗n ◦ E + D ◦ T ⊗n ◦ E k


2 2
≤ kϑ⊗m ◦ id⊗m − D ◦ T ⊗n ◦ E k + kϑ⊗m ◦ D ◦ T ⊗n ◦ Ek


2 2 2
≤ δkϑ⊗m ⊗m
2 k + k(ϑ2 ◦ D ◦ ϑ⊗n ⊗n
B ) ◦ ϑB ◦ T
⊗n
◦ Ek
≤ δ2m + kϑB ◦ T kn .

Rearranging this inequality and taking the logarithm shows that

m log(1 − δ)
+ ≤ log (kϑB ◦ T k ) .
n n
Applying this inequality for a sequence of (n, mn , δn )-coding schemes with
mn
lim =R and lim δn = 0,
n→∞ n n→∞

gives the inequality

 
mn log(1 − δ)
R = lim + ≤ log (kϑB ◦ T k ) ,
n→∞ n n

which finishes the proof.

Besides giving an upper bound on a complicated quantity, the previous theorem has an
important consequence:

Corollary 3.2. If the composition ϑB ◦ T is completely positive for a quantum channel

T : B(HA ) → B(HB ), then we have Q(T ) = 0.

Proof. If ϑB ◦ T is completely positive, then we can compute

kϑB ◦ T k = kidHA ⊗ ϑB ◦ T k1→1 = k1HA ⊗ (T ∗ ◦ ϑ∗B )(1HB )k∞ = k1HA ⊗ 1HA k∞ = 1,

where we used the Russo-Dye theorem and the fact that T is trace-preserving. By Theorem
3.1 we have
Q(T ) ≤ log (kϑB ◦ T k ) = 0.

In Lecture 6 we have seen an example of an entangled quantum state with positive partial
transpose. Whenever the Choi operator of a quantum channel yields such a state, then the
channel has vanishing quantum capacity. There are many examples of such quantum channels
showing that the structure of quantum channels with vanishing quantum capacity is much
richer than for the classical capacity. Surprisingly, the channels staying completely positive
under composition with the transpose are not even the only quantum channels with zero
quantum capacity as we will see now.

5
3.2 Antidegradable channels have zero quantum capacity
Consider a quantum channel T : B(HA ) → B(HB ) given in its Stinespring form
 
T (X) = TrE V XV † ,

for any X ∈ B(HA ) and with an isometry V : HA → HB ⊗ HE . Recall the complementary

channel T c : B(HA ) → B(HE ) given by
 
T c (X) = TrB V XV † ,

for any X ∈ B(HA ). The channel T is called

• degradable if there is a quantum channel S : B(HB ) → B(HE ) such that S ◦ T = T c .

• antidegrable if there is a quantum channel R : B(HE ) → B(HB ) such that R ◦ T c = T .

Degradable and antidegradable quantum channels will play an important role in the study
of the quantum capacity. For now, we just note the following theorem:

Theorem 3.3. If the quantum channel T : B(HA ) → B(HB ) is antidegradable, then we

have Q(T ) = 0.

Proof. Assume for contradiction that there are quantum channels En : B( C2) → B(HA⊗n)
⊗n
and Dn : B(HB C N
) → B( 2 ) for each n ∈ such that

id2 = lim Dn ◦ T ⊗n ◦ En .
n→∞

Such channels can easily be obtained from coding schemes achieving some non-zero rate.
C C C
Consider now the quantum channels Sn : B( 2 ) → B( 2 ⊗ 2 ) given by

Sn (X) = (Dn ⊗ Dn ◦ R) ◦ Ad⊗n

V ◦En ,

where V : HA → HB ⊗ HE denotes the Stinespring isometry of T and R : B(HE ) → B(HB )

such that R ◦ T c = T . By compactness of the set of quantum channel, there exists a
convergent subsequence Snk and a quantum channel S = limk→∞ Snk . It is now easy to see
C C C
that S : B( 2 ) → B( 2 ⊗ 2 ) is a quantum channel satisfying

(id2 ⊗ Tr) ◦ S = (Tr ⊗id2 ) ◦ S = id2 ,

contradicting the no-cloning theorem.

4 Quantum communication and entanglement generation

To prove the coding theorem for the quantum capacity it will be useful to consider the task
of entanglement generation. Instead of transmitting part of an arbitrary quantum state,
entanglement generation concerns the creating of a maximally entangled quantum state
between two distant parties. At first sight, this seems to be a strictly weaker task than
communication, but it will turn out that the corresponding capacities, i.e., the suprema of
achievable rates for these two tasks are the same.

Definition 4.1 (Coding schemes for generating entanglement). Let T : B(HA ) → B(HB )
denote a quantum channel. An (n, m, δ)-coding scheme for entanglement generation using
T is a quantum channel
D : B(HB ⊗n
C
) → B ( 2 )⊗m ,

6
together with a quantum state
ρ∈D ( C2)⊗m ⊗ (HA)⊗n
,
such that
F ω2⊗m , id⊗m ⊗ D ◦ T ⊗n (ρ) ≥ 1 − δ.



2

When considering coding schemes for generating entanglement we will often restrict to
the case of pure quantum states ρ = |ψihψ|. We do not loose generality by doing so. Indeed,
the function

2
ρ 7→ F ω2⊗m , id⊗m
2 ⊗ D ◦ T ⊗n (ρ) ,
for fixed quantum channels D and T is convex and continuous, and hence its maximum over
the set of quantum states is attained in a pure state. Therefore, we do not make a given
coding scheme for entanglement generation worse by replacing its quantum state ρ by a pure
state.
Definition 4.2 (Entanglement-generation capacity). We call a rate R ≥ 0 achievable for
generating entanglement using the quantum channel T : B(HA ) → B(HB ) if for every
N
n ∈ , there exists an (n, mn , δn )-coding scheme for entanglement generation using T such
that
mn
R = lim and lim δn = 0.
n→∞ n n→∞
The entanglement-generation capacity of a quantum channel T : B(HA ) → B(HB ) is
QEG (T ) = sup{R ≥ 0 : R achievable rate for entanglement generation using T }.
We will now show that this quantity coincides with the quantum capacity. For this, we
will need a lemma:
Lemma 4.3. Let H, H0 denote complex Euclidean spaces with dim(H), dim(H0 ) ≥ 2. Con-
sider a quantum channel T : B (H0 ) → B(H) satisfying


F ωH , (idB(H) ⊗ T ) (|φihφ|) ≥ 1 − δ,
for some pure state |φi ∈ H ⊗ H0 and where ωH ∈ H ⊗ H denotes the normalized maximally
entangled state. Then, for any complex Euclidean space K with dim(K) ≤ dim(H)/2, there
exist quantum channels
E : B(K) → B(H0 ) and D : B(H) → B(K)
such that
kidB(K) − D ◦ T ◦ Ek ≤ 8 (32δ)1/4 .
The proof of this lemma will use the following lemma. We will postpone its proof to the
exercises:
Lemma 4.4. Let H, H0 denote complex Euclidean spaces and consider a quantum channel
T : B (H0 ) → B(H) satisfying


F ωH , (idB(H) ⊗ T ) (|φihφ|) ≥ 1 − δ,
for some pure state |φi ∈ H ⊗ H0 . Then, there exists a number r ≤ dim(H) and a pure state
|Ωr i ∈ H ⊗ H of the form
r
1 X
|Ωr i = √ |ai i ⊗ |ai i
r
i=1
for some orthonormal set ⊂ H, and a quantum channel E : B(H) → B(H0 ) with
{|ai i}ri=1


F ωH , (idB(H) ⊗ T ◦ E) (|Ωr ihΩr |) ≥ 1 − 4δ.

7
Proof. Exercises.

Now we can proceed with the proof of the main lemma:

Proof of Lemma 4.3. Using Lemma 4.4 and considering the quantum channel T ◦ E (see the
statement of Lemma 4.4) instead of T we can restrict to the case where H0 = H and
r
1 X
|φi = √ |ai i ⊗ |ai i ∈ H ⊗ H
r
i=1

for some orthonormal set {|a1 i, . . . , |ar i} ⊂ H where we write d = dim(H). Remember, that
by doing so we have

F ωH , (idB(H) ⊗ T ) (|φihφ|) ≥ 1 − 4δ.
We start by defining an orthonormal set of vectors {|w1 i, . . . , |wr i} ⊂ H generating nested
subspaces V1 ⊂ V2 ⊂ · · · ⊂ Vr = span{|a1 i, . . . , |ar i} as
span{|w1 i, . . . , |wk i} = Vk , for each k ∈ {1, . . . , r}. (2)
Initially, we set Vr = span{|a1 i, . . . , |ar i} and we recursively define

R
 +
αk = min{h|vihv|, T (|vihv|)iHS : |vi ∈ Vk , hv|vi = 1} ∈ 0

|wk i ∈ Vk some normalized vector satisfying h|wk ihwk |, T (|wk ihwk |)iHS = αk

Vk−1 = {|vi ∈ Vk : hwk |vi = 0} ⊂ Vk ,


for all k from r to 1. Clearly, the set of vectors {|w1 i, . . . , |wr i} ⊂ H constructed in this way
satisfies (2). Moreover, since this set forms an orthonormal basis of span{|a1 i, . . . , |ar i} we
have
r
1 X
|φihφ| = |wk ihwl | ⊗ |wk ihwl |,
r
k,l=1
√
since |φi = vec(Π)/ r, where
r
X r
X
Π= |ai ihai | = |wi ihwi |,
i=1 i=1

is the projection onto span{|a1 i, . . . , |ar i}. After extending {|w1 i, . . . , |wr i} to an orthonor-
mal basis {|w1 i, . . . , |wd i} of H, we also have
d
1 X
ωH = |wk ihwl | ⊗ |wk ihwl |.
d
k,l=1

Using these expressions, we find that

r

2 1 X
F ωH , (idB(H) ⊗ T ) (|φihφ|) = h|wk ihwl |, T (|wk ihwl |)iHS .
dr
k,l=1
PN
Using the Kraus decomposition T = n=1 AdKn we find that
N
X
|h|wk ihwl |, T (|wk ihwl |)iHS | = | hwk |Kn |wk ihwl |Kn |wl i|
n=1
v v
uN uN
uX uX
≤t |hwk |Kn |wk i|2 t |hwl |Kn |wl i|2
n=1 n=1
p p
= h|wk ihwk |, T (|wk ihwk |)iHS h|wl ihwl |, T (|wl ihwl |)iHS
√
= αk αl ,

8
and hence that
r

1 X√
F ωH , (idB(H) ⊗ T ) (|φihφ|) ≤ √ αk .
dr k=1
Using the Cauchy-Schwarz inequality, we conclude that
v
r u r

1 X √ u1 X
1 − 4δ ≤ F ωH , (idB(H) ⊗ T ) (|φihφ|) ≤ √ αk ≤ t αk ,
dr k=1 d
k=1

by which we find that

r
1X
αk ≥ (1 − 4δ)2 ≥ 1 − 8δ.
d
k=1

Define the number

m = max{k ∈ {1, . . . , r} : αk ≥ 1 − 16δ},
and, since 1 ≥ α1 ≥ α2 ≥ · · · ≥ αr , we have
m r
1X 1 X m r−m m d−m
1 − 8δ ≤ αk + αk ≤ + (1 − 16δ) ≤ + (1 − 16δ) ,
d d d d d d
k=1 k=m+1

from which we conclude that m ≥ d/2. Consider now some complex Euclidean space K of
dimension dim(K) ≤ d/2. There exists an isometry V : K → Vm ⊆ H and we may define
a quantum channel E : B(K) → B(H) by E = AdV . Furthermore, we may define another
quantum channel D : B(H) → B(K) by
h i
D(X) = V † XV + Tr (1H − V V † )X σ,

for some quantum state σ ∈ D(K). For any |xi ∈ K we have

F (E(|xihx|), (T ◦ E)(|xihx|)) = F (|vihv|, T (|vihv|)) ≥ αm ≥ 1 − 16δ,

where |vi = V |xi ∈ Vm . Using the data-processing inequality of the fidelity and the fact
that (D ◦ E)(|xihx|) = |xihx| for any |xi ∈ K, we conclude that

F (|xihx|, (D ◦ T ◦ E)(|xihx|)) = F ((D ◦ E)(|xihx|), (D ◦ T ◦ E)(|xihx|))

≥ F (E(|xihx|), (T ◦ E)(|xihx|))
≥ 1 − 16δ,

for any |xi ∈ K. Using Lemma 1.3 we have

kidB(K) − D ◦ T ◦ Ek ≤ 8 (32δ)1/4 .

Finally, we can prove that the quantum capacity of a quantum channel coincides with
the entanglement-sharing capacity:

Theorem 4.5. For any quantum channel T : B(HA ) → B(HB ) we have

Q(T ) = QEG (T ).

9
Proof. Fix some quantum channel T : B(HA ) → B(HB ) and let R ≥ 0 be an achievable
rate for transmitting quantum information over T such that for every n ∈ , there exists an N
(n, mn , δn )-coding scheme with
mn
R = lim and lim δn = 0.
n→∞ n n→∞

Let us denote by
En : B (C2)⊗m n

⊗n
→ B(HA ) and ⊗n
Dn : B(HB )→B ( C2)⊗m n


,
the coding channels making up the coding schemes such that
kid⊗m
2
n
− Dn ◦ T ⊗n ◦ En k = δn → 0,
as n → ∞. Now, note that
kω2⊗mn − id⊗m ⊗ Dn ◦ T ⊗n ◦ En ω2⊗mn k1 ≤ kid⊗m − Dn ◦ T ⊗n ◦ En k → 0
n


n
2 2

as n → ∞. Using the Fuchs-van-de-Graaf inequalities, we conclude that

F ω2⊗mn , id⊗m ⊗ Dn ◦ T ⊗n ◦ En ω2⊗mn → 1,
n




2

as n → ∞. This shows that R is an achievable rate for generating entanglement by using

the coding schemes defined by the quantum channels Dn and the quantum states
ρn = (id⊗m ⊗ En ) ω2⊗mn .
n


2

We have shown that

Q(T ) ≤ QEG (T ).
To show the other direction, consider (n, mn , δn )-coding schemes for every n ∈ N for entan-
glement sharing over T such that δn → 0 as n → ∞ and achieving a rate
mn
R = lim .
n→∞ n

Assume that the quantum channels

⊗n
Dn : B(HB )→B ( C2)⊗m n


,
together with pure quantum states
|φn i ∈ ( C2)⊗m n
⊗ (HA )⊗n
make up these coding schemes such that
F ω2⊗mn , id⊗m ⊗ Dn ◦ T ⊗n (|φn ihφn |) ≥ 1 − δn → 1,
n



2

N
as n → ∞. For every n ∈ we now use Lemma 4.3 with the complex Euclidean space K =
C C
( 2 )⊗(mn −1) and H = ( 2 )⊗mn satisfying dim(K) = dim(H)/2. This shows the existence of
quantum channels
Ẽn : B((C2)⊗(m −1)) → B(HA⊗n)
n
and D̃n : B((C2)⊗m n
C2)⊗(m −1))
) → B(( n

such that
⊗(mn −1)
kid2 − D̃n ◦ Dn ◦ T ⊗n ◦ Ẽn k ≤ 8 (32δn )1/4 → 0,
as n → ∞. We conclude that the rate
mn − 1
R = lim ,
n→∞ n
is achievable for quantum communication via the quantum channel T .

10