a

r
X
i
v
:
q
u
a
n
t
-
p
h
/
0
4
0
1
0
2
6

v
1



7

J
a
n

2
0
0
4
Comments on multiplicativity of
maximal p-norms when p = 2
Christopher King

Department of Mathematics
Northeastern University, Boston MA 02115
[email protected]
Mary Beth Ruskai

Department of Mathematics , Tufts University

Medford, Massachusetts 02155
[email protected]
January 8, 2004
Dedicated to A.S. Holevo on the occasion of his 60th birthday
Abstract
We consider the maximal p-norm associated with a completely positive
map and the question of its multiplicativity under tensor products. We
give a condition under which this multiplicativity holds when p = 2, and
we describe some maps which satisfy our condition. This class includes
maps for which multiplicativity is known to fail for large p.
Our work raises some questions of independent interest in matrix the-
ory; these are discussed in two appendices.

Partially supported by the National Science Foundation under Grant DMS-0101205.

Partially supported by the National Science Foundation under Grant DMS-0314228.

1
1 Introduction
In quantum information theory, noise is modeled by a completely positive and
trace-preserving (CPT) map acting on the states of the quantum system. In
general takes pure states into mixed states, and in order to assess the noisiness
of the map one is interested in knowing how close the image states may come to
pure states. Amosov, Holevo and Werner (AHW) [2] observed that this could be
measured by the quantity

p
() = sup |()|
p
: > 0, Tr = 1 (1)
where ||
p
=
_
Tr ()
p

1/p
and 1 p . For any density matrix , ||
p
1
with equality if and only if is pure. Hence
p
() 1 with equality if and only
if there is a pure state for which () is also pure (since by convexity the sup
in (1) is achieved on a pure state).
The p-norm dened above can be extended to arbitrary matrices as |A|
p
=
_
Tr [A[
p

1/p
with [A[ =

A. The following useful relationships, which were

established in [1], can be readily veried.

p
() = sup
>0,Tr =1
|()|
p
= sup
A>0
|(A)|
p
Tr A
(2)
= sup
A=A

|(A)|
p
Tr [A[
= sup
A=A

|(A)|
p
|A|
1
(3)
sup
A
|(A)|
p
|A|
1
. (4)
(The equivalence of (2) and (3) follows from the convexity of the p-norm and
the fact that [(A
+
A

)[ = (A
+
) + (A

) when A = A
+
A

is the
decomposition of a self-adjoint operator into its positive and negative parts.) It
follows immediately from (3) that for any self-adjoint A
|(A)|
p

p
() Tr [A[. (5)
The representation (4) suggests viewing as a map between spaces of com-
plex matrices with dierent p-norms. As suggested in [1] one can generalize this
by dening
||
qp
= sup
A
|(A)|
p
|A|
q
(6)
and let ||
R
qp
denote the same quantity when the supremum is restricted to
the real vector space of self-adjoint operators. Then
p
() is precisely ||
R
1p
.
2
In general, ||
R
qp
||
qp
and one would expect that the inequality could be
strict for some . However, the second part of Theorem 1 states that equality
holds in all dimensions when p = q = 2; and in Appendix B.3, we show that
equality holds for CPT maps on qubits when q = 1, p 2. This raises the
question of whether equality always holds and, if not, for what types of maps
strict inequality is possible.
Amosov, Holevo and Werner conjectured [2] that
p
is multiplicative on tensor
products, i.e., that

p
( ) =
p
()
p
() (7)
This has been veried in a number of special cases, although it is now known
to be false in general. Amosov and Holevo [1] proved (7) when both and
are products of depolarizing CPT maps and p is integer. King proved (7) for all
p 1 and arbitrary under the additional assumption that is a unital qubit
CPT map [10], is a depolarizing channel in any dimension [11], or is an
entanglement breaking map [12]. However, Holevo and Werner [19] also showed
that (7) need not hold in general by giving a set of explicit counterexamples for
p > 4.79 and d 3.
Amosov and Holevo conjectured [1] that the quantity in (6) should also be
multiplicative for 1 q p, i.e., that
| |
qp
= ||
qp
||
qp
(8)
Beckner [5] established an analogous multiplicativity for commutative systems
when 1 q p. Curiously, Junge [8] proved (8) for completely positive (CP)
maps with p and q in the opposite order, that is for the case 1 p q. However,
our main interest is the case q = 1 < p, and Junges result does not seem to shed
any direct light on this question.
The conjecture (7) is of greatest interest for p near 1 since taking the limit
as p 1 yields the von Neumann entropy of , another natural measure of
purity, and the validity of (7) for p in an interval of the form [1, 1 +) with > 0
would imply additivity of the minimal entropy. Moreover, it has been shown
[18] that additivity of minimal entropy is equivalent to several other important
conjectures in quantum information theory, including additivity of Holevo ca-
pacity and additivity of entanglement of formation. Audenaert and Braunstein
[3] have also observed a connection between multiplicativity for CP maps, and
super-additivity of entanglement of formation.
In view of the Holevo-Werner example, it is natural to conjecture that (7)
holds in the range 1 p 2. This is precisely the range of values of p for which
the function f(x) = x
p
is operator convex, and it is also the range for which a
3
number of convexity inequalities hold. Verifying (7) for the special case of p = 2
would suggest its validity for 1 p 2. Unfortunately, even this seemingly
simple case is not as straightforward as one might hope. In this note, we prove
(7) when p = 2 for a special class of CP maps. Although this is a rather limited
result, it gives some insight into the diculties one encounters in the general
case.
Note that the multiplicativity for CPT maps follows if it holds for all CP
maps. We consider this more general case, as it does not seem more dicult. In
fact, it is not hard to show that multiplicativity (7) holds for all p 1 whenever
is an extreme CP map. This is because these are precisely the CP maps which
can be written in the form () = A

A, i.e, with one Kraus operator, and one

can then assume without loss of generality that A is diagonal. Thus, the extreme
CP maps fall into the diagonal maps considered in Example 1 below, for which
multiplicativity has been proved.
Although the rst part of the following theorem is included in Junges result,
we include an elementary argument here.
Theorem 1 ||
22
is multiplicative, i.e., ||
22
= ||
22
||
22
. More-
over, ||
22
= ||
R
22
Proof: First, recall that the complex n n matrices form a Hilbert space with
respect to the inner product A, B) = Tr A

B, and let

denote the adjoint of
the linear operator with respect to this inner product. Since
_
||
22
_
2
= sup
A
Tr [(A)]

(A)
Tr A

A
= sup
A
A, (

)(A))
A, A)
, (9)
it follows that ||
22
is the usual operator sup norm on this Hilbert space or,
equivalently, the largest singular value of . Thus, ||
22
is the square root of
the largest eigenvalue of (

). This is the same as the largest eigenvalue of

(

) I; therefore, |I|
22
= ||
22
. The main result then follows from
the submultiplicativity of the Hilbert space operator norm under composition
since
| |
22
= |( I) (I )|
22
|( I)| |(I )|
22
= ||
22
||
22
.
Note that since (

) has real eigenvalues (in fact, they are non-negative), the

solutions of the eigenvector equation (

)(B) = B are self-adjoint (or can be

so chosen if is degenerate). This implies that the supremum in (9) is achieved
with a self-adjoint A, which implies the second statement in the Theorem.
4
2 Main Theorem
We now nd it convenient to introduce some notation. When e
j
is an or-
thonormal basis for C
d
, we will let E
jk
= [e
j
)e
k
[ denote the matrix with a 1
in the j-th row and k-th column and 0s elsewhere. Then the set of operators
E
jk
also form an orthonormal basis for the d d matrices with respect to the
Hilbert-Schmidt inner product. Moreover, if is a matrix on C
d
C
d

, we can
write =

jk
E
jk
M
jk
where M
jk
= Tr
1
(E

jk
I). This is equivalent to
saying that is a block matrix with blocks M
jk
.
If is a CP map, then (I )() =

jk
E
jk
(M
jk
) > 0 which implies that
any 2 2 submatrix
_
(M
jj
) (M
jk
)
(M
kj
) (M
kk
)
_
is positive semi-denite. This implies
in turn that
(M
jk
) = (M
jj
)
1/2
R
jk
(M
kk
)
1/2
(10)
where R
jk
is a contraction. Hence
Tr (M

jk
)(M
jk
) |(M
jj
)|
2
|(M
kk
)|
2
. (11)
Theorem 2 Let and be CP maps one of which (say ) satises the condition
Tr (E
ik
)

(E
j
) 0 i, j, k, . (12)
Then
2
( ) =
2
()
2
().
Proof: Writing an arbitrary density matrix as above, one nds
( )() =

jk
(E
jk
) (M
jk
) (13)
Thus
Tr [( )()]

( )()
=

ik

j
Tr (E
ik
)

(E
j
) Tr (M
ik
)

(M
j
) (14)

ik

Tr (E
ik
)

(E
j
)

|(M
ik
)|
2
|(M
j
)|
2

ik

Tr (E
ik
)

(E
j
)

_
|(M
ii
)|
2
|(M
jj
)|
2
|(M
kk
)|
2
|(M

)|
2
[
2
()]
2

ik

Tr (E
ik
)

(E
j
)

_
Tr M
ii
Tr M
jj
Tr M
kk
Tr M

(15)
5
where we have used (11) and (5). Now, note that the matrix
N =

ik
_
Tr M
ii
Tr M
kk
E
ik
(16)
=
_
_
_
_
_

Tr M
11

Tr M
22
.
.
.

Tr M
dd
_
_
_
_
_
_
Tr M
11

Tr M
22
. . .

Tr M
dd
_
(17)
is positive semi-denite and

ik

j
Tr (E
ik
)

(E
j
)
_
Tr M
ii
Tr M
jj
Tr M
kk
Tr M

= Tr (N)

(N) (18)
[
2
()]
2
(Tr N)
2
(19)
= [
2
()]
2
_

i
Tr M
ii
_
2
= [
2
()]
2
(Tr )
2
. (20)
When (12) holds the absolute value bars are redundant in (15). One can then
substitute (20) in (15) to yield
Tr [( )()]

( )() [
2
()]
2
[
2
()]
2
(Tr )
2
. (21)
Taking the square root and dividing both sides by Tr , one nds
|( )()|
2
Tr

2
()
2
() 0. (22)
Taking the supremum over gives the desired result.
Although condition (12) is simple, it is basis dependent. In order that be
multiplicative, it suces that (12) holds for the matrices E
jk
= [e
j
)e
k
[ associ-
ated with some basis for C
d
. The question of when such a basis can be found
gives rise to some interesting questions in matrix theory which we remark on
in Appendix A. We remark here only that, although we do not expect (12) to
hold for all CP maps, we also do not have a counter-example. Thus, one cannot
exclude the possibility that the hypothesis of Theorem 2 is actually satised by
all CP, or by all CPT, maps.
6
3 Special cases
It order to show that our results are not vacuous, we now give some examples of
maps which satisfy (12).
1. Maps with only diagonal Kraus operators. In this case (E
jk
) = a
jk
E
jk
for some positive matrix A = (a
jk
), and the condition (12) follows from
the orthonormality of the E
jk
. This class of CP maps was studied by
Landau and Streater who named them the diagonal maps. In fact a more
complicated analysis using the Lieb-Thirring inequality can be used to show
multiplicativity for all p 1 for these maps [14].
2. Maps for which (I )(M) =

jk
E
jk
(E
jk
) has non-negative elements,
where M =

jk
E
jk
E
jk
is the maximally entangled state. (This is the
block matrix with blocks (E
jk
); it is sometimes called the Choi matrix or
Jamiolkowski state representative of .) Although this condition is clearly
sucient to satisfy (12), it is not necessary. For example, let A be a positive
semi-denite matrix with some a
jk
< 0. Then for that particular j, k the
corresponding map in Example 1 has (E
jk
) = a
jk
E
jk
with one strictly
negative element.
3. Multiplicativity at p = 2 has been proven for all qubit CPT maps [9].
However, we can verify the condition (12) only for a subset of qubit CPT
maps. This subset is described using the parametrization of qubit maps
that was derived in [15] and summarized in Appendix B.1.
In terms of that notation, the condition (12) is satised when t
1
= t
2
= 0
(since in this case (E
jk
) is diagonal when j = k and skew diagonal when
j ,= k). It is interesting to note that multiplicativity is known to hold for
all p 1 under the stronger condition t
1
= t
2
= t
3
= 0 [10], so this result
suggests that it may hold for p 1 for a larger class of CPT maps.
Another class of qubit maps which satisfy (12) are those with
1

2
,
t
2
= 0, and t
1
0 (again using the notation in Appendix B.1). These
maps belong to Example 2 since (E
jk
) has non-negative elements for all
j, k. Furthermore, in Theorem 3 of [9], King proved that multiplicativity
holds for these channels for all integer p 1, and later [13] extended this
to all p 2.
4. The special class of maps satisfying (23) and discussed below. This class
includes maps for which multiplicativity does not hold for some p > 2.
7
Let M denote a d d Hermitian matrix with elements m
jk
= x
jk
+iy
jk
with
x
jk
, y
jk
real. Let : M (M) denote a linear map with the following very
special properties:
[(M)]
jk
=
_

d
j
m

when j = k,
a
jk
(x
jk
+ i
jk
y
jk
) when j ,= k
(23)
where d
j
0, a
jk
are the o-diagonal elements of a xed Hermitian matrix and

jk
=
kj
= 1. The map is trace-preserving if and only if the matrix D with
elements d
j
is column stochastic. Not every map of the form (23) will necessarily
be CP. However, certain special subclasses can be identied.
a) a
jk
= 0 j ,= k. In this case, is a QC map consisting of the projection
onto the diagonal part of M followed by the action of a column stochastic
matrix on the classical probability vector corresponding to the diagonal.
b) A > 0 is a xed positive semi-denite matrix, d
j
=
j
a
jj
and
jk
= +1.
This is exactly the diagonal class described above.
c) d
j
= 1
j
, a
jk
= 1 j ,= k and
jk
= 1. In this case, (M) =
(Tr M)I M
T
and
1
d1
(M) is the CPT map for which Holevo and Werner
[19] showed that (7) does not hold for large p.
Since King [9, 14] showed that multiplicativity holds for all p 1 for maps of
type (a) and (b), multiplicativity at p = 2 may not seem very signicant. How-
ever, maps of type (c) are precisely those used to establish that multiplicativity
does not hold for suciently large p. Moreover, the full class includes convex
combinations of maps of type (a) with one of type (b) or type (c), and Kings
results do not apply to this class. Thus this class of maps is neither trivial nor
uninteresting.
Although one can verify that CP maps satisfying (23) always satisfy the
hypothesis of Theorem 2, we state and prove their multiplicativity as a separate
result. Inequality (11) again plays a key role in the proof.
Theorem 3 Let be a CP map satisfying (23) and let be an arbitrary CP
map. Then
2
( ) =
2
()
2
().
Proof: As before, let =

jk
E
jk
M
jk
be the matrix with blocks M
jk
. Then
( )() > 0 has blocks
[( )()]
jk
=
_

d
j
(M

) when j = k,
a
jk
((M
jk
) + i
jk
(M
jk
) when j ,= k
(24)
8
where (M
jk
) =
1
2
(M
jk
+M

jk
) and (M
jk
) =
i
2
(M
jk
M

jk
). A straightforward
calculation gives
Tr [( )()]

( )() (25)
=

n
d
j
d
jn
Tr (M

)(M
nn
) +

j=k
[a
jk
[
2
Tr (M

jk
)(M
jk
)

n
d
j
d
jn
|(M

)|
2
|(M
nn
)|
2
+

j=k
[a
jk
[
2
|(M
jj
)|
2
|(M
kk
)|
2
[
2
()]
2
_

n
d
j
d
jn
Tr M

Tr M
nn
+

j=k
[a
jk
[
2
Tr M
jj
Tr M
kk
_
(26)
where we have used the Schwarz inequality for the Hilbert-Schmidt inner product
and (11). Now, the term in parentheses in (26) is precisely Tr (N)

(N) where
N is the matrix with elements N
jk
= (Tr M
jj
Tr M
kk
)
1/2
. The desired result then
follows as in the proof of Theorem 2 since
Tr [( )()]

( )() [
2
()]
2
Tr (N)

(N)
[
2
()]
2
[
2
()]
2
(Tr N)
2
= [
2
()]
2
[
2
()]
2
(Tr )
2
.
4 Concluding remarks
If one could replace the operator basis E
jk
by a more general orthonormal
operator basis G
m
for C
dd
, one could always satisfy the analogue of (12).
One need only choose G
m
to be the basis which diagonalizes the positive semi-
denite operator

, i.e., for which
(

)(G
m
) =
2
m
G
m
. (27)
where
m
are the singular values of . Then Tr (G
m
)

(G
n
) =
2
n

mn
0.
Moreover, as noted at the end of the proof of Theorem 1, one can always choose
the basis so that each G
m
= G

m
is self-adjoint.
Using the orthogonality condition Tr G

m
G
n
=
mn
, one can show that a
density matrix on a tensor product space can be written in the form
=

m
G
m
W
m
with W
m
= Tr
1
(G

m
I). (28)
Note G
m
self-adjoint implies that W
m
is also self-adjoint.
9
We now try to imitate the proof of Theorem 2. Since ( )() =

m
(G
m
) (W
m
),
Tr [( )()]

( )() =

mn
Tr (G
m
)

(G
n
) Tr (W
m
)

(W
n
)
=

2
n
Tr (W
n
)

(W
n
) (29)
[
2
()]
2

2
n
(Tr [W
n
[)
2
(30)
= [
2
()]
2
Tr (N

)(N)
[
2
()]
2
[
2
()]
2
(Tr [N[)
2
(31)
where N =

m
G
m
Tr [W
n
[ (and the rst inequality implicitly used the assump-
tion that G
m
is self-adjoint so that W
m
is).
Unfortunately, we can not conclude that Tr [N[ Tr as needed to complete
the proof. (If we had instead N =

m
G
m
Tr W
n
, then we would have N =
Tr
2
> 0 and Tr N = Tr [N[ = Tr .) This is a real problem. Using the Pauli
basis for qubits, for which G
k
= 2
1/2

k
, consider the maximally entangled Bell
state = G
0
G
0
+ G
1
G
1
G
2
G
2
+ G
3
G
3
. Then Tr
2
=
1
2
I, but
N =
1

2
_
2 2 +i
2 i 0
_
is not positive semi-denite and Tr [N[ > 1.
Although our results do not prove it, we conjecture that multiplicativity does
hold for all CP maps at p = 2. If this conjecture turns out to be false, then it
seems unlikely that any other value of p between 1 and 2 would play a special
role, and there would probably be counterexamples to multiplicativity all the
way down to p = 1. In this case additivity of minimal entropy would be an
isolated result and the attempt to prove it using p-norms would probably be
futile.
A Comments on positivity condition (12):
To nd conditions under which (12) holds, note that it is equivalent to the
requirement that
x
ik,j
= Tr E

ik
(

)(E
j
) 0 i, j, k, (32)
which is precisely the condition that the matrix X representing the positive semi-
denite linear operator

in the orthonormal operator basis E
jk
also has
non-negative elements x
ik,j
.
10
If [f
j
) = U[e
j
) denotes another O.N. basis for C
d
, then F
jk
= [f
j
)f
k
[ =
UE
jk
U

is an orthonormal operator basis for C

dd
, and the corresponding matrix
representative of

is (U
T
U

)X(U U). In the proof of Theorem 2 above

we require only that E
jk
have the form [e
j
)e
k
[ for some orthonormal basis e
j

for C
d
so that we can use (11).
Therefore, multiplicativity of
2
() will hold if there is a unitary operator
U on C
d
such that the matrix (U
T
U

)X(U U) has non-negative elements

[where X is the matrix with elements dened by (32)]. Unfortunately, this is not
a very tractable condition.
One way to nd an example of a CP map satisfying (12) on C
d
is to nd a
d
2
d
2
positive semi-denite matrix X with non-negative elements. Regarding
the d d blocks of X as (E
jk
) denes a CP map of the type considered in
Example 2. However, as noted before, CP maps satisfying (12) need not have
this form.
Ad
2
d
2
matrix X with elements x
ik,j
0 also denes a positive semi-denite
linear map which one can write as = (

). The map then satises (12),

but it need not be CP. We only know that =
U

for some linear operator

U
on C
dd
which is unitary in the sense Tr
U
(A)

U
(B) = Tr A

B for all dd
matrices A, B. This implies that must have the form (A) = U

(A)U for
some unitary matrix U.
1
Hence, is CP if and only if

is; however, this does

not seem easy to check.
B Qubit maps
B.1 Notation:
It will be useful to summarize some basic facts about the representation of matri-
ces and CP maps on qubits using the identity and Pauli matrices as bases. One
can write an arbitrary matrix as A = z
0
I +z with z
0
C, z = w+iu and w, u
vectors in R
3
. When z
0
,= 0, any norm satises |z
0
I + z| = [z
0
[|I +
1
z
0
z|.
Therefore, we will present most results only for z
0
= 1; results for z
0
= 0 are
generally straightforward. The most general CP map has the form
(I +z) = (1 +s w + i s u)I + (t + Tw + iTu) (33a)
(z) = s zI + (Tz) (33b)
1
The fact that
U
must have the form
U
(A) = U

AU is probably well-known, but was

rst brought to the attention of MBR by Nicolas Boulant, who includes a proof in his paper
[6]. If one writes
U
in Kraus form, one can then use the fact that

U
= I to show that
the Kraus operators can be chosen to be a single unitary.
11
where s, t are vectors in R
3
and T is a real 3 3 matrix. is TP if and only if
s = 0; and is unital if and only if t = 0.
As observed in [15], one can use the singular value decomposition to assume
without loss of generality that T is diagonal with real (but not necessarily posi-
tive) elements
k
. This leads to the canonical form
(I +w) = I +

k
(t
k
+
k
w
k
)
k
(34)
for CPT maps introduced in [15]. Conditions on the parameters t
k
,
k
which
guarantee that is CPT are given in [17]; some special cases were considered
earlier in [4].
B.2 Useful formulas
We now restrict attention to CPT maps acting on A = I +z for which (A) =
I + (t + Tw + iTu). Then
A

A = (1 +[z[
2
)I + 2(w +u w) (35)
with
[w +u w[ = [w[
2
+[w[
2
[u[
2
(u w)
2
. (36)
Therefore, the eigenvalues of A

A are 1 +[z[
2
2[w +u w[ or, equivalently,
1 +[z[
2
2
_
[w[
2
+[w[
2
[u[
2
(u w)
2
. (37)
and those of (A)

(A) are

= 1 +[t + Tw[
2
+[Tu[
2
2
_
[t + Tw[
2
(1 +[Tu[
2
) [(t + Tw) (Tu)[
2
. (38)
When (t + Tw) (Tu) = 0, (38) becomes
_
[t + Tw[ +
_
1 +[Tu[
2
_
2
.
We now wish to evaluate and bound |A|
2
1
= (Tr [A[)
2
Note that (37) implies
that the eigenvalues of [A[ =

A are
_
1 +[z[
2
2
_
[w[
2
+[w[
2
[u[
2
(u w)
2
, (39)
and observe that their product can be written as
(1 +[z[
2
)
2
4([w[
2
+[w[
2
[u[
2
(u w)
2
) = (1 [w[
2
+[u[
2
)
2
+ 4(u w)
2
.
12
Therefore,
(Tr [A[)
2
= 2
_
1 +[z[
2
+
_
(1 [w[
2
+[u[
2
)
2
+ 4(u w)
2
_
(40)
2
_
1 +[w[
2
+[u[
2
+

1 [w[
2
+[u[
2

_
(41)

_
4(1 +[u[
2
) if [w[
2
1 +[u[
2
4[w[
2
if [w[
2
> 1 +[u[
2
.
(42)
B.3 Equality for CPT maps when p 2:
We now show that ||
R
1p
= ||
1p
for CPT maps on qubits when p 2. For
A = I +z we prove the somewhat stronger result that
|[I + (w + iu)]|
2
p
|I + (w + iu)]|
2
1

|[I + w )|
2
p
|I + w |
2
1
(43)
where w is the unit vector dened by w = [w[ w. Our argument will use the
following easily veried results. When a 0 and m 1,
f(x) = [x + a[
m
+[x a[
m
is increasing for x > 0 (44)
g(x) =
[f(x)]
2/m
x
2
is decreasing for x > 0. (45)
Note that f(x) is symmetric in x and a from which it follows that the expression
on the right in (44) is also increasing in a.
It follows from (42) and (38) that
|(A)|
2
p
|A|
2
1

_
(
+
)
p/2
+[

[
p/2
_
2/p
4 max([w[
2
, 1 +[u[
2
)
(46)
Since the numerator has the form of f in (44) with m = p/2, the monotonicity
in a implies that dropping the dot product terms in (38) increases the right side
of (46). Hence
|(A)|
2
p
|A|
2
1

_
_
[t + Tw[ +
_
1 +[Tu[
2
_
p
+

[t + Tw[
_
1 +[Tu[
2

p
_
2/p
4 max([w[
2
, 1 +[u[
2
)
(47)
Since [Tu[ [u[, (44) again implies that the right side of (47) increases when
[Tu[ is replaced by [u[ in the numerator. Also we can only increase the ratio
13
on the right side of (47) by choosing t Tw to be positive. Therefore, we can
conclude from (44) that this ratio is increasing in [w[ for [w[
2
1 + [u[
2
, and
from (45) that it is decreasing in [w[ for [w[
2
1 + [u[
2
. Hence this ratio is
maximized when [w[
2
= 1 +[u[
2
. Therefore the ratio in (47) is less than
_
_
[t + Tw[ +[w[
_
p
+

[t + Tw[ [w[

p
_
2/p
4[w[
2
, (48)
which we want to show is smaller than the RHS of (43). Since [w[
2
= 1+[u[
2
1,
[t + Tw[

[w[t + Tw

= [w[ [t + T w[. (49)

Using (44) again to replace the [t + Tw[ term in (48), we nd
|(A)|
2
p
|A|
2
1

_
_
[t + T w[ + 1
_
p
+

[t + T w[ 1

p
_
2/p
4
(50)
=
|(I + w )|
2
p
|I + w |
2
1
(51)

_
||
R
1p
_
2
=
p
()
2
. (52)
We next consider the case z
0
= 0, for which A = z, [A[ = [z[I and (A) =
(Tz). Then |(A)|
p
= [Tz[ for all p so that
|(A)|
p
|A|
1
=
|(z)|
p
|z|
1
=
[Tz[
[z[
max
k

k

p
() (53)
where the last inequality follows from the fact that [
2
()]
2

1
2
(1+[t+Tw[)
2

[Tw[
2
which can be made equal to max
k

2
k
for some w with [w[ = 1.
To complete the proof, recall that for z
0
,= 0, |z
0
I +z|
p
= [z
0
[ |I +
1
z
0
z|
p
and note that the factor [z
0
[ will cancel in any ratio of norms. Therefore, if we
take the supremum over all complex matrices A, we can use (52) and (53) to
conclude that ||
1p

p
() = ||
R
1p
when p 2. The reverse inequality
||
R
1p
||
1p
always holds; therefore, we must have equality for p 2.
B.4 Remarks
Suppose that both
p
() = ||
R
1p
and ||
1p
are multiplicative for some p, .
Suppose also that ||
R
1p
= ||
1p
in dimension d (e.g., d = 2.) Then in
dimension d
2
(e.g., d = 4),
| |
R
1p
=
_
||
R
1p
_
2
=
_
||
1p
_
2
= | |
1p
. (54)
14
Thus, equality also holds in dimension d
2
for maps of the form .
The argument in Section B.3 breaks down for 1 p < 2. Although one does
not expect (43) to hold, the weaker inequality |(A)|
p
|A|
1

p
() might still
hold, and this is all that is needed to show ||
1p
||
R
1p
. However, even
for p = 1, we have been unable to verify (or nd a counter-example to) this.
For the general CP form (33a) with s ,= 0,
1
() = 1 + [s[ > 1 is achieved
with w =
s
|s|
, and the eigenvalues of (A)

(A) are
(S
2
+[t + Tw[
2
+[Tu[
2
2
_
[t + Tw[
2
[(S
2
+[Tu[
2
] [(t + Tw) (Tu)[
2
with S
2
= (1 +s w)
2
+(s u)
2
. If one tries to use the argument in the previous
section, the RHS of (47) becomes
_
_
[t + Tw[ +
_
S
2
+ [Tu[
2
_
p
+

[t + Tw[
_
S
2
+[Tu[
2

p
_
2/p
4 max([w[
2
, 1 +[u[
2
)
. (55)
This does not have the form [x + a[
m
+ [x a[
m
because a =
_
S
2
+[Tu[
2
and
S
2
depends on w.
References
[1] G. G. Amosov and A. S. Holevo, On the multiplicativity conjecture for
quantum channels, Theor. Probab. Appl. 47, 143146 (2002).
[2] G. G. Amosov, A. S. Holevo, and R. F. Werner, On Some Additivity Prob-
lems in Quantum Information Theory, Problems in Information Transmis-
sion, 36, 305313 (2000).
[3] K. M. R. Audenaert and S. L. Braunstein, On strong superadditivity of
the entanglement of formation, preprint quant-ph/0303045; to appear in
Commun. Math. Phys. 244, xxxx (2004).
[4] A. Fujiwara and P. Algoet, One -to-one parametrization of quantum chan-
nels Physical Review A, 59, 32903294 (1999).
[5] W. Beckner, Inequalities in Fourier Analysis Ann. Math. 102, 159182
(1975).
[6] N. Boulant, private communication.
15
[7] M. Horodecki, P. W. Shor and M. B. Ruskai, General Entanglement Break-
ing Channels, Rev. Math. Phys. 15, 629641 (2003).
[8] M. Junge, private communication.
[9] C. King, Maximization of capacity and p-norms for some product channels,
J. Math. Phys., 43, 12471260 (2002).
[10] C. King, Additivity for unital qubit channels, J. Math. Phys., 43, 4641
4653 (2002).
[11] C. King, The capacity of the quantum depolarizing channel, IEEE Trans.
Info. Theory, 49, 221229, (2003).
[12] C. King, Maximal p-norms of entanglement breaking channels, Quantum
Information and Computation, 3, 186190 (2003).
[13] C. King, Inequalities for Trace Norms of 2 2 Block Matrices Commun.
Math. Phys. 242, 531545 (2003).
[14] C. King, An application of the Lieb-Thirring inequality in quantum infor-
mation theory, to appear in Proceedings of ICMP 2003.
[15] C. King and M. B. Ruskai, Minimal Entropy of States Emerging from Noisy
Quantum Channels, IEEE Trans. Info. Theory 47, 192209 (2001).
[16] L. J. Landau and R. F. Streater, On Birkhos theorem for doubly stochas-
tic completely positive maps of matrix algebras, Lin. Alg. Appl. 193, 107
127 (1993).
[17] M. B. Ruskai, S. Szarek and E. Werner, An Analysis of Completely-Positive
Trace-Preserving Maps on 2x2 Matrices, Lin. Alg. Appl. 347, 159187
(2002).
[18] P. W. Shor, Equivalence of Additivity Questions in Quantum Information
Theory, preprint quant-ph/0305035; to appear in Commun. Math. Phys.
244, xxxx (2004).
[19] R. F. Werner and A. S. Holevo, Counterexample to an additivity conjec-
ture for output purity of quantum channels, J. Math. Phys. 43, 43534357
(2002).
16