Quantum information theory (MAT4430) Spring 2021

Lecture 18: LSD theorem (part 2), superactivation and classical assistance
Lecturer: Alexander Müller-Hermes

In the final lecture, we will prove the converse part of the LSD theorem. We will then
study a striking consequence of this theorem: Superactivation. Although two channels each
have zero quantum capacity their tensor product can have non-zero quantum capacity. One
way of putting this is that allowing for a side-channel with zero capacity to be used by the
sender and the receiver can increase the quantum capacity. Along similar lines, we will then
look at classical side-channels either from the sender to the receiver or from the receiver to
the sender. We will see that in the first case the classical side-channel does not increase the
quantum capacity, but in the second case it does.

1 The proof of the LSD theorem

We will need the following theorem, which we have shown in the exercises:
Theorem 1.1 (Data-processing inequality for the coherent information). For quantum chan-
nels T : B(H1 ) → B(H2 ) and S : B(H2 ) → B(H3 ) we have

Ic (σ, S ◦ T ) ≤ Ic (σ, T ),

for any σ ∈ D(H1 ).

Proof. Exercise.

Now, we can show the main result of this lecture:

Theorem 1.2 (Lloyd, Shor, Devetak). For any quantum channel T : B(HA ) → B(HB ) we
have
1  
Q(T ) = lim Ic T ⊗k .
k→∞ k

Proof. In the previous lecture, we have seen that

1  
lim sup Ic T ⊗k ≤ QEG (T ) = Q(T ),
k→∞ k

for any quantum channel T : B(HA ) → B(HB ). To prove the remaining inequality assume
that R > 0 is an achievable rate for entanglement generation via T . Therefore, we have for
N
every n ∈ there is an (n, mn , δn )-coding scheme for entanglement generation consisting of
a quantum channel
Dn : B(HB ⊗n
C
) → B ( 2 )⊗mn ,

together with a pure quantum state

|φn i ∈ (C2)⊗m n
⊗ (HA )⊗n ,

such that R = limn→∞ mn /n and

F ω2⊗mn , id⊗m ⊗ Dn ◦ T ⊗n (|φn ihφn |) ≥ 1 − δn → 1,

n



2

as n → ∞. By the Fuchs-van-de-Graaf inequalities we find that

kω2⊗mn − id⊗m ⊗ Dn ◦ T ⊗n (|φn ihφn |) k1 = n → 0,

n


2

1
as n → ∞, and (by taking the partial trace) we also have

1⊗m
2
n
⊗n


k m
− D n ◦ T (ρn ) k1 ≤ n ,
2 n
where we set ρn ∈ (HA )⊗n to denote the partial traces of |φn ihφn | over the reference system.
Using Fannes’ inequality (which we proved in exercise 4 on sheet 8), we find that

H id⊗m ⊗ Dn ◦ T ⊗n (|φn ihφn |) ≤ 2mn n + 1

n



2

and that
Dn ◦ T ⊗n (ρn ) ≤ mn n + 1,



mn − H
whenever n ∈

√
N
is chosen large enough (to guarantee that η(n ) ≤ 1). Since both |φn i
and vec ρn are purifications of the same quantum state ρn , we can use the Schmidt-
decomposition to show that

id⊗m ⊗ Dn ◦ T ⊗n (|φn ihφn |) ,

n


2

and
 √ √ 
id⊗n
A ⊗ Dn ◦ T
⊗n
vec ( ρn ) vec ( ρn )†
have the same non-zero spectrum. Therefore, we have

H id⊗m ⊗ Dn ◦ T ⊗n (|φn ihφn |)

n



2

 √ √ † 
= H id⊗n A ⊗ D n ◦ T ⊗n
vec ( ρn ) vec ( ρn ) .

Using the data processing inequality of the coherent information (see exercises), we find that

Ic (T ⊗n ) ≥ Ic (Dn ◦ T ⊗n )
≥ Ic (ρn , D ◦ T ⊗n )

 √ √ † 
= H Dn ◦ T ⊗n (ρn ) − H id⊗n ⊗n



A ⊗ Dn ◦ T vec ( ρn ) vec ( ρn )
= H Dn ◦ T ⊗n (ρn ) − H id⊗m ⊗ Dn ◦ T ⊗n (|φn ihφn |)


n



2
≥ mn (1 − 3δn ) − 2.

We conclude that
 
1 mn 2
lim inf Ic (T ⊗n ) ≥ lim inf (1 − 3δn ) − = R.
n→∞ n n→∞ n n
We conclude that
1   1
lim sup Ic T ⊗k ≤ Q(T ) ≤ lim inf Ic (T ⊗k ),
k→∞ k k→∞ k

and hence we have shown that

1  ⊗k 
Q(T ) = lim Ic T .
k→∞ k

The LSD-theorem derives a formula for the quantum capacity of a quantum channel,
but again this formula is not satisfactory as it involves a regularization. It is currently not
known whether there exists a different type of formula not involving any regularization, and
it is not even known whether the quantum capacity is a computable quantity in the sense of
Turing.
In the following, we collect some examples of quantum channels T : B(HA ) → B(HB )
for which the capacity is known:

2
 • If ϑB ◦ T is completely positive, then Q(T ) = 0.

• If T is antidegradable, then Q(T ) = 0.

• If T is degradable, then Q(T ) = Ic (T ).

• In particular, the erasure channel Eλ : B( Cd) → B(Cd+1) for λ ∈ [0, 1] given by

Eλ (X) = (1 − λ)X ⊕ 0 + λ Tr [X] |d + 1ihd + 1|.

satisfies
Q(Eλ ) = (1 − 2λ) log(d).

• If HA = HB = C2 and the quantum channel T is Pauli-diagonal, i.e., we have

T = p0 id2 + p1 AdσX +p2 AdσY +p3 AdσZ ,

for p ∈ P({0, 1, 2, 3}), then

Q(T ) ≥ 1 − H(p).
This inequality is called the Hashing bound.

2 Superactivation
The following theorem is due to Graeme Smith and John Yard.

Theorem 2.1. Consider an isometry

V : CdA
→ Cd B
Cd ,
⊗ E

and the complementary quantum channels T : B(Cd ) →A B( Cd B ) and T c : B( Cd

A ) →
C
B( dE ) given by
h i h i
T (X) = TrE V XV † and T c (X) = TrB V XV † .

Moreover, let {pi , ρi }Ni=1 denote an ensemble of quantum states with a probability distribution
C
p ∈ P ({1, . . . , N }) and ρi ∈ D( dA ) for every i ∈ {1, . . . , N } and let D = N dA . Then, we
have  1
 1
Ic T ⊗ E 1 ≥ χ ({pi , T (ρi )}) − χ ({pi , T c (ρi )}) ,
2 2 2
where E 1 : B(
2
CD ) → B(CD+1) is the erasure channel defined by
1 1
E 1 (X) = X ⊕ 0 + |D + 1ihD + 1|.
2 2 2
Proof. Consider the classical-quantum state
N
CN ⊗ Cd
X
ρCA = pi |iihi| ⊗ ρA
i ∈ D(
A
)
i=1

and its purification

N
√ 0
CN ⊗ CN ⊗ Cd Cd
X
|ψC 0 CA0 A i = pi |ii ⊗ |ii ⊗ |φA
i
A
i∈ A
⊗ A
,
i=1

3
where each |φA
i
0
Ai ∈ Cd A ⊗ CdA denotes a purification of ρA
i ∈ D( Cd A ). Next, we define the
pure quantum state

|τC 0 CA0 EB i = (1C 0 ⊗ 1C ⊗ 1A0 ⊗ V ) (|ψC 0 CA0 A i) .

It is easy to verify that

N
X
pi |iihi| ⊗ T c ρA


τCE = TrC 0 A0 B [|τC 0 CA0 EB ihτC 0 CA0 EB |] = i
i=1
XN
pi |iihi| ⊗ T ρA


τCB = TrC 0 A0 E [|τC 0 CA0 EB ihτC 0 CA0 EB |] = i
i=1
XN
pi T c ρA


τE = TrC 0 CA0 B [|τC 0 CA0 EB ihτC 0 CA0 EB |] = i
i=1
N
X
pi T ρA


τB = TrC 0 CA0 E [|τC 0 CA0 EB ihτC 0 CA0 EB |] = i ,
i=1

and by the definition of the Holevo quantity (and properties of the von Neumann entropy)
we have
χ ({pi , T (ρi )}) = H(τB ) − H(τCB ) + H(p),
and
χ ({pi , T c (ρi )}) = H(τE ) − H(τCE ) + H(p).
Finally, we consider the quantum state

ρC 0 A0 A = TrC [|ψC 0 CA0 A ihψC 0 CA0 A |] ,

and, by identifying CD = CN ⊗ Cd A , we can compute

Ic (E 1 ⊗ T ) ≥ Ic (ρC 0 A0 A , E 1 ⊗ T )
2 2
     
= H E 1 ⊗ T (ρC 0 A0 A ) − H E 1 ⊗ T c (ρC 0 A0 A )
2 2
1 1
= H ((idC 0 A0 ⊗ T ) (ρC 0 A0 A )) + H (T (ρA ))
2 2
1 1
− H ((idC 0 A0 ⊗ T c ) (ρC 0 A0 A )) − H (T c (ρA ))
2 2
1 1 1 1
= H(τC 0 A0 B ) + H(τB ) − H(τC 0 A0 E ) − H(τE )
2 2 2 2
1 1 1 1
= H(τCE ) + H(τB ) − H(τCB ) − H(τE )
2 2 2 2
1 1
= χ ({pi , T (ρi )}) − χ ({pi , T c (ρi )})
2 2

To find an explicit example for superactivation, we could use a quantum channel T :

B(HA ) → B(HB ) such that ϑB ◦ T is completely positive and such that
1 1
χ ({pi , T (ρi )}) − χ ({pi , T c (ρi )}) > 0,
2 2
for some ensemble {pi , ρi }N
i=1 with ρi ∈ D(HA ). Indeed such quantum channels exist, and
you may look into the book by Watrous to see an explicit example of a quantum channel

4
 C C
T : B( 4 ) → B( 4 ) satisfying these properties. From the transposition bound we know
that Q(T ) = 0, but the previous theorem implies that
 
Q(T ⊗ E 1 ) ≥ Ic T ⊗ E 1 > 0,
2 2

C C
for some erasure channel E 1 : B( D ) → B( D+1 ) with erasure probability 1/2. We have
2
seen earlier that E 1 is antidegradable and hence we have Q(E 1 ) = 0. From this we can
2 2
conclude that surprising fact that the tensor product T1 ⊗ T2 of two quantum channels can
have strictly positive quantum capacity although both T1 and T2 have zero quantum capacity.
This phenomenon is called superactivation!
By a similar argument (see the book by Watrous for the details) to what we did for the
classical capacity, the following corollary can be obtained from an example of superactivation:
Corollary 2.2. There is a quantum channel T : B(HA ) → B(HB ) satisfying

Ic (T ⊗ T ) > 2Ic (T ).

The coherent information is not additive!

3 Does classical communication help?

In most quantum communication scenarios happening on earth, it is a reasonable assumption
that the sender and receiver can communicate classically. How does this affect the achivable
rates for quantum communication? We will consider to basic scenarios of such an assistance:
Forward classical communication from the sender to the receiver and backward classical
communication from the receiver to the sender. We will see that while the former does not
increase quantum communication rates, the latter does.

3.1 Classical forward communication does not help!

How should we model forward classical communication from the sender to the receiver in
the quantum communcation scenario? Well, instead of using an encoding quantum channel
the sender could use a general encoding instrument and the decoding channel that the
receiver applies could depend on the classical information in the output of the instrument.
Indeed, this captures the most general scenario of classical forward communication. The
next definition makes this precise:
Definition 3.1 (Coding schemes assisted by classical forward communication). Let T :
B(HA ) → B(HB ) denote a quantum channel. An (n, m, δ)-coding scheme for quantum
communication over T assisted by forward classical communication is given by an instrument
{Ei }K
i=1 of completely positive maps

C
Ei : B ( 2 )⊗m → B(HA

⊗n
)
PK
such that i=1 Ei is a quantum channel, and quantum channels
⊗n
Di : B(HB ) → B ( 2 )⊗m ,


C
for each i ∈ {1, . . . , K} such that
K
X
kid⊗m
2 − Di ◦ T ⊗n ◦ Ei k ≤ δ.
i=1

As always, we define a capacity as follows:

5
Definition 3.2 (Quantum capacity assisted by classical forward communication). We call
a rate R ≥ 0 achievable for quantum communication assisted by classical forward commu-
N
nication 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 assisted by classical forward communication of a quantum channel

T : B(HA ) → B(HB ) is given by

Q→ (T ) = sup{R ≥ 0 : R achievable rate for quantum comm. assisted by class. forward comm.}.

We will need the following lemma about the fidelity:

Lemma 3.3. Let T : B(HA ) → B(HR ) denote a quantum channel and E : B(HR ) → B(HA )
a completely positive map satisfying Tr [E(1R )] = dR and

F (ωR , (idR ⊗ T ◦ E) (ωR ))2 ≥ 1 − ,

for some  ∈ (0, 1]. Then, there exists a pure quantum state |ψRA i ∈ HR ⊗ HA such that

F (ωR , (idR ⊗ T ) (|ψRA ihψRA |))2 ≥ 1 − 2.

Proof. Consider Kraus decompositions

N
X N
X
T = AdKi and E= AdLj ,
i=1 j=1

with Kraus operators (which might be zero)

Ki : HA → HR and Li : HR → HA ,

for every i, j ∈ {1, . . . , N } such that

Tr [Ki Lj ] = 0,

whenever i 6= j. The existence of such Kraus decompositions follows by considering general

sets of Kraus operators {K̃i }N M
i=1 and {L̃j }j=1 and the matrix X ∈ B( C C
N , M ) with entries

Xij = Tr [Ki Lj ] .

By the singular value decomposition, there exist unitaries U ∈ U CN
and U 0 ∈ U CM

such that X h i
Uik Ulj0 Tr K̃k L̃l = 0,
kl

whenever i 6= j. Now, the operators Ki = N

P PM 0
k=1 Uik K̃k and Lj = l=1 Ulj L̃l define new sets
of Kraus operators with the desired property.
By a lemma from lecture 9 we have
N
1 X
2
F (ωR , (idR ⊗ T ◦ E) (ωR )) = 2 | Tr [Ki Li ] |2 ,
dR i=1

and in the following we may assume that Ki and Li are non-zero (otherwise restrict the sum
to only include the non-zero terms). Then, we have
N
1 X | Tr [Ki Li ] |2
F (ωR , (idR ⊗ T ◦ E) (ωR ))2 = pi ≥ 1 − ,
d2R i=1 pi

6
where we introduced
1 h i
pi = Tr Li L†i > 0.
dR
P
Since i pi = 1, there exists some i ∈ {1, . . . , N } such that

| Tr [Ki Li ] |2 | Tr [Ki Li ] |2
= i ≥ 1 − .
d2R pi
h
dR Tr Li L†i

Now, we define the operator √

dR Li
L= r h i,
†
Tr Li Li

and the pure quantum state

|ψi = (1R ⊗ L)|ΩR i,
where |ΩR i denotes the normalized maximally entangled state. It is then easy to verify that

1 | Tr [Ki Li ] |2
F (ωR , (idR ⊗ T ) (|ψRA ihψRA |))2 ≥ 2 | Tr [Ki L] |2 = h i ≥ 1 − .
dR dR Tr Li L†i

Now, we show that forward classical communication does not help:

Theorem 3.4. For any quantum channel T : B(HA ) → B(HB ) we have

Q→ (T ) = Q(T ).

Proof. Clearly, we have Q(T ) ≤ Q→ (T ) since the parties can just choose not to communicate.
To see the other inequality assume that R ≥ 0 is an achievable rate for quantum communica-
tion assisted by forward classical communication. Consider a sequence of (n, mn , δn )-coding
schemes for quantum communication assisted by forward classical communication for each
N
n ∈ such that
mn
R = lim ,
n→∞ n
(n)
given by a sequence of encoding instruments {Ei }K n
i=1 with

Ei
(n)
:B ( C2)⊗m n

⊗n
→ B(HA )

such that
PKn (n)
i=1 Ei is a quantum channel for each n ∈ N, and decoding quantum channels
Di
(n) ⊗n
: B(HB )→B ( C2)⊗m n


,

for each i ∈ {1, . . . , Kn } such that

Kn
(n) (n)
X
kid⊗m
2
n
− Di ◦ T ⊗n ◦ Ei k = δn → 0,
i=1

as n → ∞. We can use these coding schemes to send maximally entangled quantum states
and by the Fuchs-van-de-Graaf inequalities we find that
Kn
!! !
(n) (n)
X
F ω2⊗mn , id⊗m Di ◦ T ⊗n ◦ Ei ω2⊗mn


2
n
⊗ → 1,
i=1

7
 (n)
as n → ∞. Without loss of generality we may assume that Ei 6= 0 for each i and each n.
Using the formula for the fidelity with a pure state, we find that
Kn (n)
! !2
X (n) (n) Ei
ω2⊗mn , id⊗m ⊗n
ω2⊗mn


pi F 2
n
⊗ Di ◦T ◦ (n)
=: 1 − n → 1,
i=1 pi

as n → ∞, and where we introduced

h i
(n)
Tr Ei (1⊗m
2
n
)
(n)
pi = > 0.
2mn

Since
PKn (n)
i=1 pi = 1 for each n ∈ N, there exists in ∈ {1, . . . , Kn} for each n ∈ N such that
(n)
! !2
(n) Ein
ω2⊗mn , id⊗m ⊗n
ω2⊗mn


F 2
n
⊗ D in ◦T ◦ (n)
≥ 1 − n .
pin

Now, applying Lemma 3.3 we find a sequence of pure quantum states |ψn i ∈ ( C2)⊗mn ⊗n
⊗ HA
such that    2
(n)
F ω2⊗mn , id⊗m2
n
⊗ Din ◦ T ⊗n
(|ψn ihψ n |) → 1,
(n)
as n → ∞. We conclude that the Din and |ψn i form a sequence of coding schemes for
entanglement generation over T achieving the rate R. We conclude that

Q→ (T ) ≤ QEG (T ) = Q(T ),

and the proof is finished.

3.2 Classical backward communication helps!

we will now see that coding schemes allowing for classical information to be send backwards
from the receiver to the sender can achieve higher communication rates than the unassisted
quantum capacity. For simplicity, we will restrict to the task of entanglement generation
assisted by backward communication.

Definition 3.5 (Entanglement generation schemes assisted by backward communication).

Let T : B(HA ) → B(HB ) denote a quantum channel. An (n, m, δ)-coding scheme for en-
tanglement generation over T assisted by backward classical communication is given by a
⊗n
quantum state ρRA ∈ D(HR ⊗ HA ) for some complex Euclidean space HR , an instrument
N
{Di }i=1 consisting of completely positive maps
⊗n
Di : B(HB ) → B(( C2)⊗m),
PN
such that i=1 Di is a quantum channel, and quantum channels

Ei : B(HR ) → B(( C2)⊗m)

such that
N
!
X
ω2⊗n , ⊗n



F Ei ⊗ Di ◦ T (ρRA ) ≥ 1 − δ.
i=1

As always, we define a capacity as follows:

8
Definition 3.6 (Quantum capacity assisted by classical backward communication). We
call a rate R ≥ 0 achievable for entanglement generation assisted by classical backward
communication over the quantum channel T : B(HA ) → B(HB ) if for every n ∈ , there N
exists an (n, mn , δn )-coding scheme such that
mn
R = lim and lim δn = 0.
n→∞ n n→∞

The quantum capacity assisted by classical backward communication of a quantum channel

T : B(HA ) → B(HB ) is given by

Q←,EG (T ) = sup{R ≥ 0 : R achievable rate for quantum comm. assisted by class. backward comm.}.

Surprisingly, we have the following theorem:

C C
Theorem 3.7. The erasure channel Eλ : B( 2 ) → B( 3 ) with erasure probability λ ∈ [0, 1]
satisfies
Q←,EG (Eλ ) ≥ 1 − λ > 1 − 2λ = QEG (Eλ ).

Proof. The last equality was shown in the exercises. We will now show that any rate 0 ≤
R < (1 − λ) is achievable for entanglement generation assisted by backward communication.
The strategy to achieve such a rate is very simple: The sender just sends n halfs of maximally
entangled states through Eλ⊗n . Then, the receiver identifies which of the maximally entangled
states were transmitted correctly. If more than bRnc maximally entangled states have been
transmitted correctly the receiver discards enough of them to be left with exactly bRnc
of them. If less than bRnc maximally entangled states have been transmitted correctly the
receiver outputs some fail state σF,n . Finally, the receiver communicates which tensor factors
contain the bRnc maximally entangled states (if there were enough of them) and the sender
restricts to those tensor factors as well. We will see that the probability of failure for this
scheme approaches 0 as n → ∞.
To do the above strategy formally, we will need to define an instrument and some addi-
tional quantum channels. Consider first the operator W : 3 → 2 such that C C
(
|ii, if i ∈ {1, 2}
W |ii =
0, otherwise,

and the quantum channel Γ : B( C3) → B(C2 ⊗ C2) given by

12
Γ(X) = AdW (X) ⊗ |1ih1| + h3|X|3i ⊗ |0ih0|.
2
Furthermore, for every n ∈ N and i1 , . . . , in ∈ {0, 1} we define an operator Ui1 ,...,in :
C C
( 2 )⊗n → ( 2 )⊗n such that
O O
Ui1 ,...,in (|a1 i ⊗ · · · ⊗ |an i) = |ak i ⊗ |al i.
k:ik =1 l:il =0

C C
Then, we define completely positive maps Di1 ,...,in : B(( 2 ⊗ 2 )⊗n ) → B(( 2 )⊗bRnc ) C
( ⊗bRnc ⊗(n−bRnc)

◦ AdUi1 ···in ◦ id⊗n


id2 ⊗ Tr2 2 ⊗ Adhi1 ,...,in | , if i1 + · · · + in ≥ bRnc,
Di1 ,...,in =
σF,n Tr ◦ id⊗n


2 ⊗ Adhi1 ,...,in | , otherwise,

C
where σF,n ∈ D ( 2 )⊗bRnc denotes some quantum state. It is easy to check that {Di1 ,...,in }i1 ,...,in

C
is an instrument. Finally, we define quantum channels Ei1 ,...,in : B(( 2 )⊗n ) → B(( 2 )⊗bRnc ) C
as follows: If i1 + · · · + in ≥ bRnc, then Ei1 ,...,in by traces over all tensor factors k for which

9
ik = 0 and maybe more until only bRnc tensor factors survive. If i1 + · · · + in < bRnc, then
Ei1 ,...,in traces out all tensor factors and outputs σF,n from above. Now, note that

12 12
(id2 ⊗ Γ ◦ Eλ )(ω2 ) = (1 − λ)ω2 ⊗ |1ih1| + λ ⊗ ⊗ |0ih0|,
2 2
and therefore we have
X
Ei1 ,...,in ⊗ (Di1 ,...,in ◦ (Γ ◦ Eλ )⊗n ) ω2⊗n

i1 ,...,in
⊗bRnc
= pn ω2 + (1 − pn )σF,n ⊗ σF,n ,

where
pn = Prob (X1 + · · · + Xn ≥ bRnc) ,
for some sequence of random variables (Xk )k∈N independently and identically distributed
E
with Prob(X1 = 1) = [X1 ] = 1 − λ. Since R < 1 − λ we may use the weak law of large
numbers to conclude that pn → 1 as n → ∞. Therefore, we have
 
bRnc
X √
Ei1 ,...,in ⊗ (Di1 ,...,in ◦ (Γ ◦ Eλ )⊗n ) ω2⊗n  ≥ pn → 1,



F ω2 ,
i1 ,...,in

as n → ∞ showing that R is achievable for entanglement generation assisted by classical

backward communication.

10