Quantum information theory (MAT4430) Spring 2021

Lecture 12: The Holevo-Schumacher-Westmoreland theorem

Lecturer: Alexander Müller-Hermes

In the previous lecture, we have introduced the classical capacity of quantum channels,
which quantifies the ability of such a channel to transmit classical information. We have
seen that the bound from Holevo’s theorem can be attained by taking product states as
input to many copies of the channel but allowing for global POVMs to be applied to many
communication lines. In particular, we have seen that such schemes can achieve the Holevo
information
χ (T ) = sup χ ({p(x), T (ρx )}x∈Σ ) .
{p(x),ρx }x∈Σ

of a quantum channel T : B(HA ) → B(HB ). Building on this result, we will now prove
the main theorem about the classical capacity of quantum channels, which is known as
the Holevo-Schumacher-Westmoreland theorem due to its discoverers. After proving this
theorem, we will discuss the problem of computing the classical capacity, and we will identify
classes of quantum channels for which it can be evaluated.

1 The Holevo-Schumacher-Westmoreland theorem

We will need the following lemma, which we will prove in the exercises:

Lemma 1.1. For k ∈ {1, . . . , N } and δ 6 k1 let Pδ,k (N ) denote the set of all probability
distributions p ∈ P ({1, . . . , N }) such that p(1), · · · , p(k) > δ. Then, we have
 
1 − δk
max = −kδ log(δ) − (1 − δk) log ,
p∈Pδ (N ) N −k

and the optimum is achieved for the probability distribution with

1 − δk
p(1) = · · · = p(k) = δ and p(k + 1) = · · · = p(N ) = .
N −k
Proof. Exercises.

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

Theorem 1.2 (HSW). For any quantum channel T : B(HA ) → B(HB ) we have
1  ⊗k 
C(T ) = lim χ T .
k→∞ k

Proof. Using the coding schemes constructed for product codes in the previous lecture, we
find that
χ (T ) 6 C(T ),
for any quantum channel T : B(HA ) → B(HB ). Now, we apply this equation for the
quantum channel T ⊗k and using an exercise we obtain
1  ⊗k  1
χ T 6 C(T ⊗k ) = C(T ).
k k

1
Finally, taking limits of subsequences shows that
1  
lim sup χ T ⊗k 6 C(T ).
k→∞ k

We will now derive an upper bound on the classical capacity C(T ) from which the theorem
N
will follow. For this, we assume that for every n ∈ there exists an mn ∈ such that N
mn
R = lim ,
n→∞ n
and such that there is a map
⊗n
En : {0, 1}mn → D(HA ),

and a measurement
µn : {0, 1}mn → B(HB )⊗n ,
with
hµ(b1 , . . . , bmn ), T ⊗n ◦ E(b1 , . . . , bmn )iHS = 1 − δn ,
for all b1 , . . . , bmn ∈ {0, 1} and such that limn→∞ δn = 0. To apply Holevo’s theorem, we
consider the uniform probability distribution pAn ∈ P ({0, 1}mn ) given by
1
pAn (b1 , . . . , bn ) = ,
2mn
and consider the ensemble

{pAn (a1 , . . . , amn ), T ⊗n ◦ E(a1 , . . . , amn )}.

We may now define a joined probability distribution by

pAn B n (a1 , . . . , amn , b1 , . . . , bmn ) = pAn (a1 , . . . , amn ) hµ(b1 , . . . , bmn ), T ⊗n ◦E(a1 , . . . , amn )iHS ,

and by Holevo’s theorem, we have

I (An : B n )pA 6 Iacc {pA (a1 , . . . , amn ), T ⊗n ◦ E(a1 , . . . , amn )} 6 χ T ⊗n .

n Bn

Let us estimate the mutual information on the left-hand side of the previous equation. Since
pAn is uniform on {0, 1}mn , we have

H(pAn ) = mn .

Note that
X 1
pB n (b1 , . . . , bmn ) = hµ(b1 , . . . , bmn ), T ⊗n ◦ E(a1 , . . . , amn )iHS
a1 ,...,amn
2mn
1 − δn
> ,
2m n
for each (b1 , . . . , bmn ) ∈ {0, 1}mn and hence

pB n = (1 − δn )u + δn q,

with the uniform distribution u ∈ P ({0, 1}mn ) given by u(b1 , . . . , bmn ) = 1/2mn and some
other probability distribution q ∈ P ({0, 1}mn ). By concavity of Shannon’s entropy we have

H(pB n ) > (1 − δn )mn .

2
Finally, note that
1 − δn
,
pAn B n (b1 , . . . , bmn , b1 , . . . , bmn ) =
2mn
for every (b1 , . . . , bmn ) ∈ {0, 1}mn and by Lemma 1.1 we conclude that
   
1 − δn δn
H(pAn B n ) 6 −(1 − δn ) log − δn log
2mn 22mn − 2mn
6 (1 + δn )mn + H((1 − δn , δn )) 6 (1 + δn )mn + 1.

With this we compute that

I (An : B n )pA = H(pAn ) + H(pB n ) − H(pAn B n ) > (1 − 2δn )mn − 1.

n Bn

We conclude that
mn 1 1 1
− 6 I (An : B n )pA B 6 χ T ⊗n ,


(1 − 2δn )
n n n n n n
for any n ∈ N. Taking limits of subsequences we conclude that
1
χ T ⊗n .


R 6 lim inf
n→∞ n
By combining this with the lower bound, we find that
1   1  
lim sup χ T ⊗k 6 C(T ) 6 lim inf χ T ⊗k ,
k→∞ k k→∞ k

and we conclude that

1  ⊗k 
C(T ) = lim χ T .
k→∞ k

2 The capacity of entanglement breaking channels

We need the following definition, which already appeared in Lecture 6:

Definition 2.1. A linear map T : B(HA ) → B(HB ) is called entanglement breaking (EB)
if CT ∈ Sep (HA , HB ). We denote the set of entanglement breaking maps by EB(HA → HB ).

The following theorem contains a few equivalent characterizations of these linear maps:

Theorem 2.2 (Characterizing entanglement breaking channels). For a linear map T :

B(HA ) → B(HB ) the following are equivalent:

1. We have T ∈ EB(HA → HB ).

2. We have
(idR ⊗ T ) (XRA ) ∈ Sep (HR , HB )
for every Euclidean space HR and every XRA ∈ B(HR ⊗ HA )+ .

3. We have the Kraus decomposition

N
X
T = AdKn ,
n=1

with rk (Kn ) = 1 for each n ∈ {1, . . . , N }.

3
 4. We have a decomposition
L
X
T = hAl , ·iHS Bl ,
l=1

with Al ∈ B(HA )+ and Bl ∈ B(HB )+ .

The decomposition in the fourth point is sometimes called the Holevo representation of the
entanglement breaking map.

Entanglement breaking quantum channels, as the name suggests, break the entanglement
between any system they are applied to and any other system, and they can be thought of as
a POVM measurement followed by a preparation conditioned on the measurement outcome
(see point 4. in the previous theorem). The entanglement breaking channels were among the
first classes of quantum channels for which the regularized Holevo information, and hence
the classical capacity, could be computed. We will now discuss how this was achieved:

Theorem 2.3. Let T : B(HA ) → B(HB ) denote an entanglement breaking quantum channel
and S : B(HC ) → B(HD ) any quantum channel. We have

χ (T ⊗ S) = χ (T ) + χ (S) .

Proof. For any ensembles {p(x), ρx }x∈Σ1 with ρx ∈ D(HA ) and {q(y), σy }y∈Σ2 with σy ∈
D(HC ), we have

χ ({p(x), T (ρx )}) + χ ({q(y), S(σy )}) = χ ({p(x)q(y), T (ρx ) ⊗ S(σy )}) 6 χ (T ⊗ S) .

Hence, we have
χ (T ⊗ S) > χ (T ) + χ (S) .
For the remaining inequality, we consider an ensemble {p(x), ρx }x∈Σ with ρx ∈ D(HA ⊗ HC )
such that
χ (T ⊗ S) = χ ({p(x), (T ⊗ S)(ρx )}) ,
which exists by an exercise. Since T is an entanglement breaking quantum channel, we have
X X
(T ⊗ idC ) (ρx ) = q 0 (x, y 0 )τxy0 ⊗ ηxy0 = q(x, y)|bxy ihbxy | ⊗ |cxy ihcxy |,
y0 y

for probability distributions q 0 (x, ·) and q(x, ·), quantum states τxy0 ∈ D(HB ) and ηxy0 ∈
D(HC ) and some vectors |bxy i ∈ HB and |cxy i ∈ HC obtained by the spectral decomposition,
for any x ∈ Σ. Next, we consider the quantum state
(x)
X
σABD = q(x, y)|yihy| ⊗ |bxy ihbxy | ⊗ S (|cxy ihcxy |)
y

By strong subadditivity of the von-Neumann entropy, we have

(x) (x) (x) (x)
H(σBD ) > H(σABD ) − H(σAB ) + H(σB ). (1)

Let us compute the entropies appearing in the previous inequality: It is easy to see that
(x)
X
σBD = q(x, y)|bxy ihbxy | ⊗ S (|cxy ihcxy |) = (T ⊗ S) (ρx ) ,
y

and hence we have

(x)
H(σBD ) = H ((T ⊗ S) (ρx )) . (2)

4
Next, note that
!
(x)
X
H(σABD ) =H q(x, y)|yihy| ⊗ |bxy ihbxy | ⊗ S (|cxy ihcxy |)
y
X
= H(q) + q(x, y)H (|bxy ihbxy | ⊗ S (|cxy ihcxy |))
y
X
= H(q) + q(x, y)H (S (|cxy ihcxy |)) ,
y

and similarly we also have

   
(x) (x)
H σAB = H σA = H(q).

Combining the previous two equations shows that

(x) (x)
X
H(σABD ) − H(σAB ) = q(x, y)H (S (|cxy ihcxy |)) . (3)
y

Finally, we have
!
(x)
X
H(σB ) =H q(x, y)|bxy ihbxy | = H (T (TrC [ρx ])) . (4)
y

Combining (1),(2),(3) and (4) we find that

X
H ((T ⊗ S) (ρx )) > q(x, y)H (S (|cxy ihcxy |)) + H (T (TrC [ρx ])) .
y

Summing over x ∈ Σ shows that

X XX X
p(x)H ((T ⊗ S) (ρx )) > p(x)q(x, y)H (S (|cxy ihcxy |)) + p(x)H (T (TrC [ρx ])) .
x x y x

Now, we compute
!
X X
χ ({p(x), (T ⊗ S)(ρx )}) = H p(x)(T ⊗ S)(ρx ) − p(x)H ((T ⊗ S)(ρx ))
x x
" !#! " !#!
X X
6H T TrC p(x)(ρx ) +H S TrA p(x)(ρx )
x x
XX X
− p(x)q(x, y)H (S (|cxy ihcxy |)) − p(x)H (T (TrC [ρx ]))
x y x

= χ ({p(x), T (TrC [ρx ])}) + χ ({p(x)q(x, y), S (|cxy ihcxy |)}) ,

since " #
XX X
p(x)q(x, y)|cxy ihcxy | = TrA p(x)ρx .
x y x
We conclude that
χ (T ⊗ S) = χ ({p(x), (T ⊗ S)(ρx )}) 6 χ(T ) + χ(S).
This finishes the proof.

As a corollary of the previous theorem and the HSW theorem we have:

Corollary 2.4. For any entanglement breaking channel T : B(HA ) → B(HB ) we have
C(T ) = χ(T ).

5
3 Other examples where the Holevo information is additive
Without proof, we mention the following classes of quantum channels T : B(H) → B(H) for
which we have
χ (T ⊗ S) = χ (T ) + χ (S) ,
for any quantum channel S : B(H0 ) → B(H00 ). For each of these classes we we have C(T ) =
χ(T ).

Theorem 3.1 (Additivity examples). The Holevo quantity is additive if T is

• a unital qubit channel.

• a depolarizing channel, i.e., such that T : B(H) → B(H) is given by

1H
T (X) = (1 − p) Tr [X] + pX,
dim(H)

for any X ∈ B(H).

• a Schur multiplier channel, i.e., such that T : B(H) → B(H) is given by

T (X) = A X,

for any X ∈ B(H), where A ∈ B(H)+ and where denotes the entrywise product in
the computational basis.