a

r
X
i
v
:
q
u
a
n
t
-
p
h
/
0
4
0
3
0
7
2

v
1



9

M
a
r

2
0
0
4
A quantum channel with additive minimum output entropy
Nilanjana Datta,
1,
Alexander S. Holevo,
2,
and Yuri Suhov
1,
1
Statistical Laboratory, Centre for Mathematical Science,
University of Cambridge, Wilberforce Road, Cambridge CB3 0WB, UK
2
Steklov Mathematical Institute, Gubkina 8, 119991 Moscow, Russia
We give a direct proof of the additivity of the minimum output entropy of a particular quantum
channel which breaks the multiplicativity conjecture. This yields additivity of the classical capacity
of this channel, a result obtained by a dierent method in [10]. Our proof relies heavily upon certain
concavity properties of the output entropy which are of independent interest.
PACS numbers: 03.67.Hk, 03.67.-a
INTRODUCTION
A number of important issues of quantum information theory would be greatly claried if several resources
and parameters were proved to be additive. However, the proof of additivity of such resources as the minimum
output entropy of a quantum memoryless channel and its classical capacity remains in general an open problem,
see e.g. [8]. Recently Shor [4] provided a new insight into how several additivitytype properties are related to
each other. He proved that: (i) additivity of the minimum output entropy of a quantum channel, (ii) additivity
of the classical capacity of a quantum channel, (iii) additivity of the entanglement of formation, and (iv) strong
superadditivity of the entanglement of formation are equivalent in the sense that if one of them holds for all
channels then the others also hold for all channels.
In this paper we study the additivity of the minimum output entropy for a channel which is particularly
interesting because it breaks a closely related multiplicativity property [2]. For this channel the additivity of the
classical capacity and of the minimum output entropy are equivalent, which allows us to derive an alternative
proof of the result in [10], where additivity of its capacity was established. The problem of additivity of
the minimum output entropy is interesting and important in its own right (it is straightforward, addresses a
fundamental geometric feature of a channel and may provide insight into more complicated channel properties).
In this paper, the key observation that ensures the additivity is that the output entropy of the product channel
exhibits specic concavity properties as a function of the Schmidt coecients of the input pure state. It is our
hope that a similar mechanism might be responsible for the additivity of the minimum output entropy in other
interesting cases.
THE ADDITIVITY CONJECTURE
A channel in the nite dimensional Hilbert space H C
d
is a linear trace-preserving completely positive
map of the algebra of complex d dmatrices. A state is a density matrix , that is Hermitian matrix such
that 0, Tr = 1. The minimum output entropy of the channel is dened as
h() := min

S(()), (1)
where the minimization is over all possible input states of the channel. Here S() = log is the von
Neumann entropy of the channel output matrix = (). The additivity problem for the minimum output
entropy is to prove that
h(
1

2
) = h(
1
) +h(
2
), (2)
where
1
,
2
are two channels in H
1
, H
2
respectively, denotes tensor product.
A channel is covariant, if there are unitary representations U
g
, V
g
of a group G such that
(U
g
U

g
) = V
g
()V

g
; g G. (3)
If both representations are irreducible, then we call the channel irreducibly covariant. In this case there is a
simple formula

C() = log d h(), (4)

relating the Holevo capacity

C() of the channel with h() [9]. Since the tensor product of irreducibly covariant
channels (with respect to possibly dierent groups G
1
, G
2
) is again irreducibly covariant (with respect to the
group G
1
G
2
)), it follows that if (2) holds for two such channels, then

C(
1

2
) =

C(
1
) +

C(
2
). (5)
Notice that this does not follow from the result of [4] which asserts that if (2) holds for all channels, then (5)
also holds for all channels. In the latter case also

C(
1

n
) =

C(
1
) + +

C(
n
), which implies that

C() is equal to the classical capacity of the channel (see [8] for more detail).
The concavity of the von Neumann entropy implies that the minimization in (1) can be restricted to pure
input states, since the latter correspond to the extreme points of the convex set of input states. Hence, we can
equivalently write the minimum output entropies in the form
h() = min
|H
||||=1
S(([)[)); (6)
h(
1

2
) = min
|
12
H
1
H
2
||
12
||=1
S((
1

2
)([
12
)
12
[)). (7)
Here [
12
)
12
[ is a pure state of a bipartite system with the Hilbert space H
1
H
2
, where H
i
C
di
for
i = 1, 2. In order to prove (2), it is sucient to show that the minimum in (7) is attained on unentangled
vectors [
12
). Consider the Schmidt decomposition
[
12
) =
d

=1
_

[; 1)[; 2), (8)

where d = mind
1
, d
2
, [; j) is an orthonormal basis in H
j
; j = 1, 2, and = (
1
, . . . ,
d
) is the vector of
the Schmidt coecients. The state [
12
)
12
[ can then be expressed as
[
12
)
12
[ =
d

,=1
_

[; 1); 1[ [; 2); 2[. (9)

The Schmidt coecients form a probability distribution:

0 ;
d

=1

= 1, (10)
thus the vector varies in the (d 1)dimensional simplex
d
, dened by these constraints. Extreme points
(vertices) of
d
correspond precisely to unentangled vectors [
12
) = [
1
) [
2
) H
1
H
2
. The proof of (2)
becomes straightforward if we can prove that for every choice of the bases, the function

d
S(M( )) (11)
2
attains its minimum at the vertices of
d
. Here S(M( )) is the von Neumann entropy of the channel matrix
M() := (
1

2
) ([
12
)
12
[) =
d

,=1
_

1
([; 1); 1[)
2
([; 2); 2[). (12)
Two special properties of a function can guarantee this: one is concavity, and another is Shur concavity (see
the Appendix). Both of them appear useful in consideration of the particular channel we pass to.
THE CHANNEL
The channel considered in this paper was introduced in [2]. It is dened by its action on d d matrices as
follows:
() =
1
d 1
_
I tr()
T
_
(13)
where
T
denotes the transpose of the matrix , and I is the unit matrix in H C
d
. It is easy to see that the
map is linear and trace-preserving. For the proof of complete positivity see [2]. Moreover, is irreducibly
covariant since for any arbitrary unitary transformation U
(UU

) =

U()

, (14)
hence the relation (4) holds for this channel.
Our aim will be to prove the additivity relation
h( ) = 2h(), (15)
for the channel (13). For d = 2, (13) is a unital qubit channel, for which property (15) follows from [5]. For
d 3, (15) can be deduced from additivity of the Holevo capacity (5), established in [10], by a dierent method.
Here we provide a direct proof based on the idea described at the end of the previous section.
For a pure state = [)[ the channel output is given by
([)[) =
1
d 1
_
I [) [
_
, (16)
where the entries of vector [) are complex conjugates of the corresponding entries of vector [). The matrix
([)[) has a non-degenerate eigenvalue equal to 0 and an eigenvalue 1/(d1) which is (d1)fold degenerate.
The von Neumann entropy S(([)[)) is obviously the same for all pure states, and so
h() = log(d 1). (17)
As argued in the previous section, in order to prove (15), it is sucient to show that the minimum in
h( ) = min
|
12
C
d
C
d
||
12
||=1
S(( )([
12
)
12
[))
is attained on unentangled vectors [
12
). Consider the Schmidt decomposition (8) of [
12
). Owing to the
property (14), we can choose for [; j), ; j = 1, 2, the canonical basis in C
d
. As it was shown in the previous
section, it suces to check that S(M()) attains its minimum at the vertices of
d
. Here M() is the matrix
dened in (12) for the channel under consideration:
M() =
d

,=1
_

([)[) ([)[),
3
where by (13)
([)[) =
1
d 1
(

I [)[) ,
owing to the fact that [) and [) are real.
Using (8) and the completeness relations:
I =
d

=1
[)[, I I =
d

,=1
[)[,
we obtain
M() =
1
(d 1)
2
_
_
d

,=1
[)[(1

) +
d

,=1
_

[)[
_
_
. (18)
In order to nd the eigenvalues of M(), it is instructive to rst study the secular equation of a more general
n n matrix:
A =
n

j=1

j
[j)j[ +
n

j,k=1

k
[j)k[.
Matrix A gives ( )([
12
)
12
[) for a particular choice of the parameters
j
and
j
[see eq.(21) below]. It
has the form:
_
_
_
_
_
_
_

1
+
1

1

2
+
2

n
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

1

n
+
n
_
_
_
_
_
_
_
. (19)
The secular equation det(A I) = 0 can be written as
F() = 0, (20)
where
F() =

j
(
j
)
_
1 +

1

+. . . +

n

_
.
Solving eq.(20) would be in general non-trivial. However, representing the matrix
_
(d 1)
2
M()

in the form
(19) results in a convenient expression for F(). This allows us to identify many of the eigenvalues of (d 1)
2
M(). More precisely, we identify j with a pair (, ) and obtain

= 1

;
j

, , = 1, . . . , d. (21)
Therefore,
F() =
d

,=1
(1

)
_
_
1 +
d

=1

(1

)
_
_
=
d

,=1=
(1

)
_
_
d

=1
(1 2

)
_
_
_
1 +
d

=1

(1 2

)
_
_
_
_
_
.
(22)
4
Eq.(20) yields the following equations:
(1

) = 0, ,= , , = 1, 2, . . . , d, (23)
where

denote the Schmidt coecients (8). Equation (23) implies that there are d(d 1) eigenvalues of
the form
= 1

, ,= , = 1, . . . , d. (24)
The roots of the equation
d

=1
(1 2

)
_
1 +
d

=1

(1 2

)
_
= 0 (25)
give the remaining d eigenvalues of the matrix
_
(d 1)
2
M()

.
For the case d = 3 the roots of (25) can be explicitly evaluated. This is done in the next section. The case
of arbitrary d > 3 is discussed in sections that follow. Note that the sum of all eigenvalues of
_
(d 1)
2
M()

equals
tr
_
(d 1)
2
M()

= (d 1)
2
Tr M() = (d 1)
2
,
since M() is a density matrix acting in C
d
2
.
EIGENVALUES FOR d = 3
For d = 3, there are d(d 1) = 6 eigenvalues of the matrix
_
(d 1)
2
M()

= 4M(), which are given by

(24). The sum of these eigenvalues is:
3

,=1
=
(1

) = 2
_
3 2(
1
+
2
+
3
)

= 2
since

1
+
2
+
3
= 1. (26)
The remaining three eigenvalues of 4M() are given by the roots of the equation
3

=1
(1 2

)
_
1 +
3

=1

(1 2

)
_
= 0. (27)
Since the sum of all the eigenvalues is equal to (d 1)
2
4, these remaining three eigenvalues sum up to
4 2 = 2. Using (26), we can cast (27) as:

3
+a
2

2
+a
1
+a
0
= 0 (28)
where
a
0
= 4
1

3
; a
1
= 1 ; a
2
= 2. (29)
5
The three roots of (28) are given by

1
:=
a
2
3
+ (T
1
+T
2
),

2
:=
a
2
3

1
2
(T
1
+T
2
) +
1
2
i

3 (T
1
T
2
),

3
:=
a
2
3

1
2
(T
1
+T
2
)
1
2
i

3 (T
1
T
2
). (30)
Here
T
1
:=
_
R +

D
_
1/3
and T
2
:=
_
R

D
_
1/3
,
and
R =
1
54
(9a
1
a
2
27a
0
2a
3
2
), D = Q
3
+R
2
, Q :=
1
9
(3a
1
a
2
2
). (31)
Thus, the matrix M() has six eigenvalues of the form (1/4)(1

), where , = 1, 2, 3 and ,= ,
and three eigenvalues
1
,
2
and
3
, with
i
:=
i
/4. The output entropy S
_
M()
_
can be expressed as the
sum:
S(M()) = S
1
() +S
2
().
Here
S
1
() =
3

,=1
=
1
4
(1

) log
_
1
4
(1

)
_
=
1
2
3

=1

log

4
=
1
2
H() + 1 (32)
where H() =

d
=1

log

denotes the Shannon entropy of , and

S
2
() =
3

i=1

i
log
i
.
Since H() is a concave function of = (
1
,
2
,
3
), so is S
1
(). Hence S
1
() attains its minimum at the
vertices of
3
.
Let us now evaluate the summand S
2
(). Substituting the values of a
0
, a
1
and a
2
from (29) into (31), we get
R =
1
27
+ 2t, Q =
1
9
, D = 4t(
1
27
t) 0.
Here t =
1

3
, 0 t 1/27. Hence, we can write
R +

D = R +i
_
[D[ = re
i
,
where r =
_
R
2
+[D[ = 1/27 and = arctan(
_
[D[/R), 0 , so that
tan =
_
t(1/27 t)
t 1/54
.
6
Considering the sign of this expression we nd that t = 0 corresponds to = , while t = 1/27 to = 0. In
terms of the eigenvalues
k
, k = 1, 2, 3 can now be expressed as:

k
=
1
6
_
1 + cos
_

3

2(k 1)
3
__
=
1
3
cos
2
_

6

2(k 1)
6
_
.
Hence,
S
2
() =
3

k=1
1
3
cos
2
_

6

2(k 1)
6
_
log
_
1
3
cos
2
_

6

2(k 1)
6
__
. (33)
An argument similar to Lemma 3 of [6] shows that the RHS of (33) has a global minimum, equal to 1, at
= corresponding to t =
1

3
= 0. Hence, S
2
() attains its minimal value 1 at every point of the boundary

3
, in particular at its vertices:
i
= 1,
j
= 0 for j ,= i, i = 1, 2, 3. The summand S
1
(), given by (32 ), also
attains its minimum, equal to 1, at the vertices. Therefore the sum S(M()) attains its minimum, equal to 2,
at the vertices of
3
. Hence, h( ) = 2, and the additivity (15) holds, as h() = 1 by (17).
We conjecture that the entropy S(M()) as a function of is concave. This is supported by a 3D-plot
of S(M()) as a function of two independent Schmidt coecients
1
and
2
; here
i
0 for i = 1, 2 and

1
+
2
1. See Figure 1 below.
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
2
3
4
0
0.2
0.4
0.6
0.8
FIG. 1: The entropy S(M()) as a function of two independent Schmidt
coecients 1 and 2.
However S
2
() is not concave as can be seen e.g. by taking
2
=
1
,
3
= 1 2
1
. See Figure 2.
MINIMUM OUTPUT ENTROPY IN d > 3 DIMENSIONS
In a previous section we found that the matrix
_
(d 1)
2
M()

, where M() is the output density matrix of

the channel and is given by (18), has d(d 1) eigenvalues of the form
(1

) , with ,= , , = 1, 2, . . . , d, (34)
7
0.1 0.2 0.3 0.4 0.5
1.02
1.04
1.06
1.08
1.1
1.12
FIG. 2: S2() as a function of = 1 = 2.
and the remaining d eigenvalues are given by the roots
1
, . . . ,
d
of (25). Hence, the matrix M() has d(d 1)
eigenvalues of the form
e

:=
1
(d 1)
2
_
1

_
, ,= , , = 1, 2, . . . , d,
and d eigenvalues of the form
g
i
:=

i
(d 1)
2
, i = 1, 2, . . . , d.
Note that the
i
s and g
i
s are functions of
d
. Accordingly, we write the von Neumann entropy of the
output density matrix as a sum
S(M()) = S
1
() +S
2
() (35)
where
S
1
() :=
d

=1
d

=1
=
e

log e

, S
2
() :=
d

i=1
g
i
log g
i
. (36)
Note that
d

=1
d

=1
=
e

=
d

=1
d

=1
=
1

(d 1)
2
=
d 2
d 1
. (37)
Dene the following variables:
e

(= e

()) :=
d 1
d 2
e

=
1
(d 1) (d 2)

1d
=,

, ,= , , = 1, . . . , d. (38)
g
i
(= g
i
()) := (d 1) g
i
=
1
(d 1)

i
, i = 1, . . . , d. (39)
8
For d 3 we have e

0, and from (37) it follows that

d

=1
d

=1
=
e

=
d 1
d 2
d

=1
d

=1
=
e

= 1,
d

i=1
g
i
= (d 1)
_
1
d 2
d 1
_
= 1.
Hence, e := e

[ ,= , , = 1, 2, . . . , d and g = g
i
[i = 1, . . . , d are probability distributions. In terms
of these variables
S
1
() =
d 2
d 1
H( e)
d 2
d 1
log
_
d 2
d 1
_
, (40)
where H( e) denotes the Shannon entropy of e. In view of (36)
S
2
() =
1
(d 1)
H( g) +
1
d 1
log (d 1) . (41)
From (40) it follows that S
1
() in (35) is a concave function of the variables e

. These variables are ane

functions of the Schmidt coecients
1
, . . . ,
d
. Hence, S
1
is a concave function of and attains its minimum
at the vertices of
d
, dened by the constraints (10).
Let us now analyze S
2
(). We wish to prove the following:
Theorem . The function S
2
is Schur-concave in
d
i.e.,

= S
2
() S
2
_

_
, where denotes
the majorization order (see the Appendix).
Since every
d
is majorized by the vertices of
d
, this will imply that S
2
() also attains its minimum at
the vertices. Thus S() = S
1
() +S
2
() is minimized at the vertices, which correspond to unentangled states.
As was observed, this implies the additivity.
PROOF OF THE THEOREM
We will use the quite interesting observation made in [7], that the Shannon entropy H(x) is a monotonically
increasing function of the elementary symmetric polynomials s
k
(x
1
, x
2
, . . . , x
d
), k = 0, . . . , d, in the variables
x = (x
1
, x
2
, . . . , x
d
). The latter are dened by equations (65) of the Appendix. Hence the Shannon entropy
H( g) in (41) is a monotonically increasing function of the symmetric polynomials
s
k
() := s
k
( g
1
, g
2
, . . . , g
d
)
1
(d 1)
k
s
k
(
1
,
2
, . . . ,
d
), k = 0, . . . , d, (42)
Therefore, to prove the Theorem it is sucient to prove that the functions s
k
() are Schur concave in
d
.
Here the variables g
i
are given by (39), and the variables
i
are the roots of eq. (25). Dene the variables

:= 1 2

, = 1, 2, . . . , d.
Note that 1

1, owing to the inequality 0

1. Moreover,
d

=1

= d 2,
since

d
=1

= 1. In terms of the variables

, (25) can be expressed as

d

=1
(

)
_
1 +
1
2
d

=1
1

)
_
= 0. (43)
9
Since the roots
1
, . . . ,
d
of (25) are identied, trivially, as the zeroes of the product (
1
)(
2
) . . . (
d
),
equation (43) can be expressed in terms of these roots as follows:
d

k=0

k
(1)
k
s
dk
(
1
,
2
, . . . ,
d
) = 0. (44)
In terms of the elementary symmetric polynomials s
l
, of the variables
1
,
2
, . . . ,
d
, (43) can be rewritten as
d

k=0

k
(1)
k
s
dk
(
1
,
2
, . . . ,
d
) +
d1

k=0

k
(1)
k
d

l=1
s
d1k
(
1
, . . . , ,
l
. . . ,
d
)
(1
l
)
2
= 0, (45)
where the symbol ,
l
means that the variable
l
has been omitted from the arguments of the corresponding
polynomial. Equating the LHS of (44) with the LHS of (45) yields, for each 0 k d 1 :
s
dk
(
1
,
2
, . . . ,
d
) = s
dk
(
1
,
2
, . . . ,
d
) +
d

l=1
s
d1k
(
1
, . . . , ,
l
. . . ,
d
)
(1
l
)
2
. (46)
Note that in (46), values s
dk
(
1
,
2
, . . . ,
d
) are expressed in terms of values of elementary symmetric polyno-
mials in the variables
1
,
2
, . . . ,
d
(which are themselves linear functions of the Schmidt coecients
1
, . . . ,
d
).
Our aim is to prove that s
k
() is Schur concave in the Schmidt coecients
1
, . . . ,
d
. Eq.(42) implies that
this amounts to proving Schur concavity of s
dk
(
1
,
2
, . . . ,
d
) as a function of
1
, . . . ,
d
, for all 0 k d.
The functions

k
(
1
, . . . ,
d
) := s
dk
(
1
, . . . ,
d
) +
d

l=1
s
d1k
(
1
, . . . , ,
l
. . . ,
d
)
(1
l
)
2
RHS of (46) (47)
are symmetric in the variables
1
,
2
, . . . ,
d
, and hence in the variables
1
, . . . ,
d
. By eq.(64) (see the Appendix)
it remains to prove
(
i

j
)
_

j
_
(
i

j
)
_

j
_
0, 1 i, j d. (48)
By (66) we have

k
(
1
, . . . ,
d
) =

i
s
dk
(
1
, . . . ,
d
) +

i
d

l=1
s
d1k
(
1
, . . . , ,
l
. . . ,
d
)
(1
l
)
2
= s
d1k
(
1
, .., ,
i
, ..,
d
) +
d

l=1
l=i
s
d1k
(
1
, . . . , ,
i
, .., ,
l
. . . ,
d
)
(1
l
)
2

1
2
s
d1k
(
1
, . . . , ,
i
. . . ,
d
). (49)
Therefore,
_

j
_
(
1
, . . . ,
d
) = s
d1k
(
1
, .., ,
i
, ..,
d
) s
d1k
(
1
, .., ,
j
, ..,
d
)
+
d

l=1
l=i
s
d1k
(
1
, .., ,
i
, .., ,
l
..,
d
)
(1
l
)
2

d

l=1
l=j
s
d1k
(
1
, .., ,
j
, .., ,
l
. . . ,
d
)
(1
l
)
2

1
2
_
s
d1k
(
1
, . . . , ,
i
. . . ,
d
) s
d1k
(
1
, . . . , ,
j
. . . ,
d
)

. (50)
10
Using (67) we get
_

j
_
(
1
, . . . ,
d
) =
1
2
_

j

i
)s
dk2
(
1
, .., ,
i
, .., ,
j
. . . ,
d
)
+
(
i

j
)
2
s
dk2
(
1
, .., ,
i
, .., ,
j
. . . ,
d
)
+
d

l=1
l=i,j
(1
l
)
2
_
s
dk2
(
1
, .., ,
i
, .., ,
l
. . . ,
d
) s
dk2
(
1
, .., ,
j
, .., ,
l
. . . ,
d
)

=
d

l=1
l=i,j
(1
l
)
2
_
s
dk2
(
1
, .., ,
i
, .., ,
l
. . . ,
d
) s
dk2
(
1
, .., ,
j
, .., ,
l
. . . ,
d
)

=
d

l=1
l=i,j
(1
l
)
2
(
j

i
)s
dk3
(
1
, .., ,
i
, .., ,
j
, ..,
l
. . . ,
d
). (51)
Substituting (51) in (48) we obtain that Schur concavity holds if and only if
d

l=1
l=i,j
(1
l
)s
dk3
(
1
, .., ,
i
, .., ,
j
, ..,
l
. . . ,
d
) 0, 1 i, j d. (52)
The variables
i
and
j
do not appear in (52). Owing to symmetry, without loss of generality, we can choose
i = d 1 and j = d. Then omitting
d1
and
d
results in replacing (52) by
d2

l=1
(1
l
)s
dk3
(
1
, .., ,
l
. . . ,
d2
) 0.
By setting n = d 2, we can express the condition for Schur concavity by the following lemma.
Lemma . The functions
k
, dened in (47), are Schur concave in the Schmidt coecients
1
, . . . ,
d
if
n

l=1
(1
l
)s
nk1
(
1
, .., ,
l
. . . ,
n
) 0, 0 k d 3, (53)
where the variables
i
:= 1 2
i
, 1 i n, satisfy
1
i
1,
n

l=1

l
n 2. (54)
Note: The constraints (54) follow from the relations:
l
0 l, and
n

l=1

l
=
d2

l=1

l
1.
Proof of the Lemma
The constraints (54) imply that at most one of the variables
1
, . . . ,
n
can be negative. Note that (1
l
) is
always nonnegative since
l
1. Thus if all
1
, . . . ,
n
0, (53) obviously holds. Hence, we need to prove (53)
only in the case in which one, and only one, of the variables
1
, . . . ,
n
is negative.
To establish the latter fact, we rst prove the inequality
n

l=1
1
l

l
0, or
n

l=1

l
1 2
l
0. (55)
11
Without loss of generality we can choose
1
< 0 and
l
> 0 for all l = 2, 3, . . . , n. Hence,
1
> 1/2 and
l
< 1/2
for all l = 2, 3, . . . , n. Write:
LHS of (55) =

1
1 2
1
+
n

l=2

l
1 2
l
:= T
1
+T
2
.
Note that T
1
0 since
1
> 1/2. The function
f(
i
) :=

i
1 2
i
, 0
i
< 1/2,
is convex. Hence, T
2
(
2
, ,
n
), as a sum of convex functions, is convex on the simplex dened by

2
+ +
n
1
1
, 0
i
< 1/2 , i = 2, . . . , n,
with xed
i
> 1/2.
Hence, T
2
achieves its maximum on the vertices of the simplex. One vertex is (0, , 0) , for which T
2
= 0,
and hence T
1
+T
2
< 0. Other vertices are obtained by permutations from (1
1
, 0, , 0) and give
T
2
=
1
1
1 2(1
1
)
=
1
1
1 2
1
.
Thus the maximal value of T
1
+T
2
is

1
1 2
1

1
1
1 2
1
=
1 2
1
1 2
1
= 1
which proves (55).
To prove (53), using the denition (65) of elementary symmetric polynomials, we write:
s
nk1
(
1
, .., ,
l
. . . ,
n
) =
c
n
(k 1)!
n

j
1
=1
j
1
=l
n

j
2
=1
j
2
=l,j
1

n

j
k
=1
j
k
=l
j
k
=j
i
1ik
1

j1

j
k

l
.
Here c
n
:=
1
, ..,
l
. . . ,
n
< 0. Hence, the required inequality (53) becomes
n

l=1
n

j
1
=1
j
1
=l
n

j
2
=1
j
2
=l,j
1

n

j
k
=1
j
k
=l
j
k
=j
i
1ik1
1
l

j1

j
k

l
0. (56)
Once again, without loss of generality we can choose
1
< 0 and
l
> 0 for all l = 2, 3, . . . , n. Then
LHS of (56) =
n

j1=2
n

j
2
=2
j
2
=j
1

n

j
k
=2
j
k
=j
i
1ik1
1

j1

j
k
_

_
1
1

1
+
k

r=1
1
jr

1
+
n

l=2
l=j
i
1ik
1
l

l
_

_ (57)
Equation (57) can be derived as follows: Let
T(l, j
1
, j
2
, . . . , j
k
) :=
1
l

j1

j
k

l
, (58)
with l, j
1
, j
2
, . . . , j
k
1, 2, . . . , n and l, j
1
, j
2
, . . . , j
k
all dierent.
12
Without loss of generality we can choose
1
< 0 and
l
> 0 for all l = 2, 3, . . . , n. Then
LHS of (56) =
n

l=1
n

j
1
=1
j
1
=l
n

j
2
=1
j
2
=l,j
1

n

j
k
=1
j
k
=l
j
k
=j
i
1ik1
T(l, j
1
, j
2
, . . . , j
k
)
=
n

j1=2
n

j
2
=2
j
2
=j
1

n

j
k
=2
j
k
=j
i
1ik1
T(1, j
1
, j
2
, . . . , j
k
) +
n

l=2
n

j
2
=2
j
2
=l

n

j
k
=2
j
k
=l
j
k
=j
i
1ik1
T(l, 1, j
2
, . . . , j
k
)
+
n

l=2
n

j
1
=2
j
1
=l
n

j
3
=2
j
3
=j
1
,l

n

j
k
=2
j
k
=l
j
k
=j
i
1ik1
T(l, j
1
, 1, j
3
, . . . , j
k
) +
+
n

l=2
n

j
1
=2
j
1
=l

n

j
k1
=2
j
k1
=l
j
k1
=j
i
1ik2
T(l, j
1
, . . . , j
k1
, 1) +
n

l=2
n

j
1
=2
j
1
=l

n

j
k
=2
j
k
=l
j
k
=j
i
1ik1
T(l, j
1
, . . . , j
k
).
(59)
Now,
n

l=2
n

j
1
=2
j
1
=l

n

j
i1
=2
j
i1
=l
j
i1
=jr1ri2
n

j
i+1
=2
j
i+1
=l
j
i+1
=jr1ri1

n

j
k
=2
j
k
=l
j
k
=jr1rk1
T(l, j
1
, . . . , j
i1
, 1, j
i+1
, . . . , j
k
)
=
n

l=2
n

j
1
=2
j
1
=l

n

j
i1
=2
j
i1
=l
j
i1
=jr1ri2
n

j
i+1
=2
j
i+1
=l
j
i+1
=jr1ri1

n

j
k
=2
j
k
=l
j
k
=jr1rk1
1
l

j1
. . .
ji1

ji+1
. . .
j
k

l
=
n

j1=2

n

j
i1
=2
j
i1
=jr1ri2
n

j
i
=2
j
i
=jr1ri1
n

j
i+1
=2
j
i+1
=jr1ri

n

j
k
=2
j
k
=jr1rk1
1
ji

j1
. . .
ji1

ji

ji+1
. . .
j
k
=
n

j1=2
n

j
2
=2
j
2
=j
1

n

j
i
=2
j
i
=jr1ri1
n

j
k
=2
j
k
=jr1rk1
1

j1

j
k
_
1
ji

1
_
. (60)
In the second last line on the RHS of (60) , we have changed the dummy variable from l to j
i
. Hence,
RHS of (59) =
n

j1=2
n

j
2
=2
j
2
=j
1

n

j
k
=2
j
k
=j
i
1ik1
1

j1

j
k
_

_
1
1

1
+
k

r=1
1
jr

1
+
n

l=2
l=j
i
1ik
1
l

l
_

_
= RHS of (57) (61)
From (55) it follows that for given j
1
, j
2
, . . . , j
k
, with 2 j
r
n for r = 1, 2, . . . , k, and j
m
,= j
k
for all m ,= k:
n

l=2
l=j
i
1ik
1
l

l
+
k

r=1
1
jr

jr
+
1
1

1
0.
13
Hence,
n

l=2
l=j
i
1ik
1
l

l

_
k

r=1
1
jr

jr
+
1
1

1
_
(62)
Substituting (62) on the RHS of (57) yields
RHS of (57)
n

j1=2
n

j
2
=2
j
2
=j
1

n

j
k
=2
j
k
=j
i
1ik
1

j1

j
k
_
k

r=1
(1
jr
)
_
1

jr
_
_
=
n

j1=2
n

j
2
=2
j
2
=j
1

n

j
k
=2
j
k
=j
i
1ik
1

j1

j
k
_
k

r=1
(1
jr
)
_

jr

1

jr
_
_
0. (63)
This proves (56) and hence (53) for n 3 and all k = 1, 2, . . . , n 2.
APPENDIX
A realvalued function on R
n
is said to be Schur concave (see [11]) if:
x y = (x) (y).
Here the symbol x y means that x = (x
1
, x
2
, . . . , x
n
) is majorized by y = (y
1
, y
2
, . . . , y
n
) in the following
sense: Let x

be the vector obtained by rearranging the coordinates of x in decreasing order:

x

= (x

1
, x

2
, . . . , x

n
) means x

1
x

2
. . . x

n
.
For x, y R
n
, we say that x is majorized by y and write x y if
k

j=1
x

j

k

j=1
y

j
, 1 k n,
and
n

j=1
x

j
=
n

j=1
y

j
.
In the simplex
d
, dened by the constraints (10), the minimal point is (1/d, . . . , 1/d) (the baricenter of
d
),
and the maximal points are the permutations of (1, 0, . . . , 0) (the vertices).
A dierentiable function (x
1
, x
2
, . . . , x
n
) is Schur concave if and only if :
1. is symmetric
2.
(x
i
x
j
)
_

x
i

x
j
_
0, 1 i, j n. (64)
14
The l
th
elementary symmetric polynomial s
l
in the variables x
1
, x
2
, . . . , x
n
is dened as
s
0
(x
1
, x
2
, . . . , x
n
) = 1,
s
l
(x
1
, x
2
, . . . , x
n
) =

1i1<i2<i
l
d
x
i1
x
i2
. . . x
i
l
for l = 1, 2, . . . , n. (65)
We shall use the following identities:

x
j
s
k
(x
1
, x
2
, . . . , x
n
) = s
k1
(x
1
, . . . , ,x
j
, . . . , x
n
) (66)
and
s
k
(x
1
, . . . , ,x
i
, . . . , x
n
) s
k
(x
1
, . . . , ,x
j
, . . . , x
n
) = (x
j
x
i
)s
k1
(x
1
, . . . , ,x
i
, .., ,x
j
, . . . , x
n
) (67)
Acknowledgments This work was initiated when A.S.H. was an overseas visiting scholar at St Johns Col-
lege, Cambridge. He gratefully acknowledges the hospitality of the Statistical Laboratory of the Centre for
Mathematical Sciences. N.D. and Y.S. worked in association with the CMI.

n.datta@statslab.cam.ac.uk

holevo@mi.ras.ru

yms@statslab.cam.ac.uk
[1] K.Matsumoto, T.Shimono and A.Winter, Remarks on additivity of the Holevo channel capacity and the entangle-
ment of formation, quant-ph/0206148
[2] R.F.Werner and A.S.Holevo, Counterexample to an additivity conjecture for output purity of quantum channels,
Jour. Math. Phys., 43, 2002
[3] G.G.Amosov and A.S.Holevo, On the multiplicativity conjecture for quantum channels, quant-ph/0103015
[4] P. Shor, Equivalence of Additivity Questions in Quantum Information Theory, quant-ph/0305035
[5] C.King, Additivity for a class of unital qubit channels, quant-ph/0103156
[6] M.Sasaki, S.Barnett, R.Jozsa, M.Osaki and O.Hirota, Accessible information and optimal strategies for real sym-
metrical quantum sources,quant-ph/9812062
[7] G.Mitchison and R. Jozsa, Towards a geometrical interpretation of quantum information compression, quant-
ph/0309177
[8] A. S. Holevo, Additivity of classical capacity and related problems, http://www.imaph.nat.tu-bs.de/qi/problems/problems.html
[9] A. S. Holevo, Remarks on the classical capacity of quantum channel covariant channels quant-ph/0212025
[10] K.Matsumoto, F. Yura, Entanglement cost of antisymmetric states and additivity of capacity of some channels,
quant-ph/0306009
[11] R. Bhatia, Matrix analysis, SpringerVerlag, New York, 1997.
15