A Class of Linear Positive Maps in Matrix Algebras

arXiv:quant-ph/0307132v1 18 Jul 2003

A. Kossakowski
Institute of Physics
Nicolaus Copernicus University
Grudziadzka 5, 87–100 Toruń, Poland
e-mail: [email protected]

Abstract
A class of linear positive, trace preserving maps in Mn is given in terms of
n2 −1
affine maps in R which map the closed unit ball into itself.

1 Introduction
A linear mapping from a C ∗ -algebra A into a C ∗ -algebra B is called positive if it
carries positive elements of A into positive elements of B. Such a map is said to be
normalised if the image of the unity of A coincides with the unity of B.
Positive maps have become a common interest to mathematicians and physicist,
c.f. [1 – 31] and [32, 33]. There are several reasons for this:

(i) The notion of positive map generalizes that of state, ∗-homomorphism, Jordan
homomorphism, and conditional expectation.

(ii) By duality, a normalised positive map defines an affine mapping between sets
of states of C ∗ -algebras. If these algebras coincide with the C ∗ -algebras of
observables assigned to a physical system, then this affine mapping corresponds
to some operation which can be performed on the system.

(iii) It has been recently shown that there exists a strong connection between the
classification of the entanglement of quantum states and the structure of posi-
tive maps.

Let Mk (A) be the algebra of k × k matrices with elements from A and Mk+ (A) be
the positive cone in Mk (A). For k ∈ N and ϕ : A → B define maps ϕk , ϕk : Mk (A) →

1
Mk (B), where ϕk ([aij ]) = [ϕ(aij )] and ϕk ([aij ]) = [ϕ(aji )]. The map ϕ is said to be
k-positive (k-copositive) if the map ϕk (ϕk ) is positive, respectively.
A map ϕ is said to be k-decomposable if ϕk is positive whenever [aij ] and [aji ]
belong to Mk+ (A). A map ϕ : A → B is called completely positive, completely copos-
itive and decomposable if it is k-positive, k-copositive and and k-decomposable for
all k ∈ N, respectively. According to the Størmer theorem [10], every decomposable
map ϕ : A → B(H) (B(H) is the set of bounded linear operators on a complex
Hilbert space H) has the form ϕ(a) = v ∗ j(a)v, where j : A → B(K) is a Jordan
homomorphism and v is a bounded operator from H into K. Since every Jordan ho-
momorphism is the sum of a ∗-homomorphism and a ∗-antihomomorphism [3] every
decomposable map ϕ : A → B(H) has the form ϕ = ϕ1 + ϕ2 , where ϕ1 is completely
positive and ϕ2 is completely copositive. A positive map which is not decomposable
is called an indecomposable one.
In the case A = B(H) = Mn (C) every decomposable map ϕ : Mn (C) → Mn (C)
can be written in the form
b∗k aT bk ,
X X
ϕ(a) = a∗k aak + (1.1)
k k

where a1 , a2 , . . . , b1 , b2 , . . . ∈ Mn (C) and T is the transpose map in Mn . It is known

[2, 12, 16] that every positive map ϕ : M2 (C) → M2 (C) is decomposable. The first
example of an indecomposable map in M3 (C) was given by Choi [7, 8] and its gener-
alizations can be found in [17 – 31].
There are two known types of indecomposable maps in Mn for n ≥ 3:
(i) A series of maps ϕk : Mn (C) → Mn (C), k = 1, . . . , n − 2; c.f. [21, 31]. Let ǫ
be the projection of norm one in Mn (C) to the diagonal part with respect to
a given basis in Cn and s be the unitary shift in Mn (C) such that s = [δi,j+1 ],
where the indices are understood to be mod n. The maps ϕk are defined as
follows
k
ǫ(si as∗i ) − a .
X
ϕk (a) = (n − k)ǫ(a) + (1.2)
i=1
Since ϕk (1ln ) = (n − 1)1ln and Trϕk (a) = (n − 1)Tra, the maps
1
τk (a) = ϕk (a) , k = 1, . . . , n − 2 , (1.3)
n−1
are bistochastic ones.
(ii) A class of maps ϕ(p0 , p1 , . . . , pn ) : Mn → Mn , c.f. [28, 31] of the form
ϕ(p0 , p1 , . . . , pn )(a) = a′ with
a′11 = p0 a11 + pn ann

2
 a′22 = p0 a22 + p1 a11
··· ···
a′nn = p0 ann + pn−1 an−1,n−1
a′ij = −aij , (1.4)

where

p0 , p1 , . . . , pn > 0 , n − 1 > p0 ≥ n − 2 , p1 · . . . · pn ≥ (n − 1 − p0 )n ,
n ≥ 3.
(1.5)
The above maps are atomic, i.e., they cannot be decomposed into the sum of a
2-positive and a 2-copositive maps [23, 31].

2 A Class of Positive Maps in Mn

Let ϕ be a positive, trace preserving map in Mn (C) i.e., a dynamical map [34]. Since
ϕ is self-adjoint, i.e., ϕ(a∗ ) = ϕ(a)∗ it is enough to consider it in the space

Hn = {a ∈ Mn : a = a∗ } . (2.1)

The linear space Hn is the real Hilbert space with the scalar product (a, b) = Tr(ab)
and the norm kak2 = (a, a).
If a ∈ Hn and Tra = p > 0, the necessary condition for positivity of a is Tra2 ≤
(Tr a)2 = p2 , which can also be written in the form
1ln 2 n−1
a− Tra ≤ (Tra)2 . (2.2)
n n
Let us define the following convex sets in Hn :
n 1ln 2 n − 1 2 o
Bn (p) = a ∈ Hn :
p ≤ a−p , Tra = p , (2.3)
n no n
Sn (p) = a ∈ Bn (p) : a ≥ 0 , (2.4)
n 1ln 2 1 o
Bn0 (p) = a ∈ Hn : a− p ≤ p2 , Tra = p . (2.5)
n n(n − 1)

The set Sn (p = 1) is just the set of all density operators on Cn .

Lemma 1 If a ∈ Bn0 (p) then a ≥ 0, i.e., the inclusions

Bn0 (p) ⊆ Sn (p) ⊆ Bn (p) (2.6)

hold.

3
Proof. Let us observe that Bn0 is invariant with respect to the mapping ρ → U ∗ ρU ,
U ∈ SU (n), and the set
n  n
n X 1 2 p2 X o
b0n = p1 , . . . , pn : pi − p ≤ , pi = p
i=1
n n(n − 1) i=1

is the ball inscribed in the set

n n
X o
sn (p) = p1 , . . . , pn : pi ≥ 0 , i = 1, . . . , n , pi = p > 0 .
i=1

Since the eigenvalues of a ∈ Bn0 lie in b0n , they are non-negative.

It should be pointed out that in the case n = 2 one has

B20 (p) = S2 (p) = B2 (p) . (2.7)

Let εp : Bn (p) → Bn0 (p) be defined as follows:

1ln 1  1ln  1  1  1ln

εp (a) = p+ a− p = a+ 1− p. (2.8)
n n−1 n n−1 n−1 n
The map ǫp is positive and trace preserving. It is convenient to introduce an orthonor-
√
mal base in Hn : f1 , . . . , fn2 such that (fα , fβ ) = δαβ and fn2 = 1ln / n, i.e. Trfα = 0
for α = 1, . . . , n2 − 1. Every element of a ∈ Hn can be written in the form
n2
X
a = xα f α , xα = Tr(afα ) (2.9)
α=1

or
1ln
a = Tra + hf, xi , (2.10)
n
2 −1
where x = (x1 , . . . , xn2 −1 ) ∈ Rn , f = (f1 , . . . , fn2 −1 ), and
2 −1
nX
hf, xi = f α xα . (2.11)
α=1

Let B(Rn , r) be the closed ball in Rn with the radius r, i.e.,

B(Rn , r) = {x ∈ Rn : hx, xi ≤ r} . (2.12)

Let Dn denote the set of affine maps (T, b) : Rn → Rn , where

(T, b)x = T x + b , T ∈ Mn (R) , b ∈ Rn , (2.13)

4
which maps the closed unit ball B(Rn , 1) into itself. Since Dn is convex, compact and
finite-dimensional, each element of Dn can be written as a finite convex combination
of its extreme elements.The extreme elements of Dn have been classified in [16] in
thefollowing manner:
n
Extr Dn = (T, b) : T = R1 ΛR2 , b = R1 c , R1 , R2 ∈ O(n) ,
Λ = kλi δij k , λ1 = . . . = λn−1 = [1 − δ2 (1 − κ2 )]1/2 , λn = κ[1 − δ2 (1 − κ2 )]1/2 ,
o
c1 = . . . = cn−1 = 0 , cn = δ(1 − κ2 ) , 0 ≤ κ ≤ 1 , 0 < δ ≤ 1 . (2.14)

It is clear that the map (T, rb) maps the ball B(Rn , r) into itself provided (T, y) ∈ Dn .
Let us observe that the subset ExtrDn0 of ExtrDn corresponding to κ = 1 has the
form
Extr Dn0 = {(T, y) : T ∈ O (n) , y = 0} . (2.15)
Taking into account (2.3), (2.9) and (2.10) one finds that an element a ∈ Bn (p) can
be represented in the form
1ln  2 −1
 n − 1 1/2 
a = p + hf, xi , p = Tra , where x ∈ B Rn ,p , (2.16)
n n
while the map εp , c.f. (2.8) takes the form

1ln 1
εp (a) = p+ hf, xi . (2.17)
n n−1
Composing the map εp with the map ψ : Bn (p) → Bn (p) given by
h  n − 1 1/2 i 1ln D  n − 1 1/2 E
ψ T, yp (a) = p + f, yp (2.18)
n n n
one finds that the composed map ϕ[(T, y)] : Bn (p) → Bn0 (p) ⊆ Bn (p) has the form

1ln 1 D  n − 1 1/2 E
ϕ[T, y](a) = p+ f, T x + py , (2.19)
n n−1 n
where (T, y) ∈ Dn2 −1 , is a positive one by the construction. Since the map (2.19) is
homogeneous it can be extended to Hn , and its extension to Mn is given by

ϕ[T, y](c + id) = ϕ[T, y](c) + iϕ[T, y](d) for c, d ∈ Hn . (2.20)

The main result can be summarized in the following

5
 2 −1
Theorem 1 Every affine map (T, y) which maps the closed unit ball in Rn into
itself induces a positive, trace preserving map ϕ(T, y) : Mn → Mn

1ln 1 D  n − 1 1/2 E
ϕ[T, y](a) = Tra + f, T x + y Tra , (2.21)
n n−1 n
2
where x = (x1 , . . . , xn2 −1 ) ∈ Rn −1 , xα = Tr(afα ), and f = (f1 , f2 , . . . , fn2 −1 ),
fα = fα∗ , Tr(fα fβ ) = δαβ , Trfα = 0, α, β = 1, . . . , n2 − 1.

It should be pointed out that in the case n = 2 every positive, trace preserving map
has the form (2.21).
A systematic construction of the operators f1 , . . . , fn2 −1 ∈ Hn (generators of
SU(n)) is well known, c.f. [35]. They are given by

(f1 , . . . , fn2 −1 ) = (dℓ , ukℓ , vkℓ ) , ℓ = 1, . . . , n − 1 , 1 ≤ k < ℓ ≤ n, (2.22)

ℓ
1 X
dℓ = p (ekk − kek+1,k+1 ) , ℓ = 1, . . . , n − 1 , (2.23)
ℓ(ℓ + 1) k=1
1
ukℓ = √ (ekℓ − eℓk ) , (2.24)
2
−i
vkℓ = √ (ekℓ − eℓk ) , (2.25)
2
ekℓ = ek (eℓ , · ) , (2.26)

where (e1 , . . . , en ) is an orthonormal base in Cn .

As an illustration of the formula (2.21) let us consider the case n = 3. Using the
notation
x1 = Tr(ad1 ) , x2 = Tr(ad2 ) (2.27)
and
xkℓ = Tr(aukℓ ) , ykℓ = Tr(avkℓ ) (2.28)
let us consider the rotation R(α) ∈ SO(8) given by

x′1 = x1 cos α − x2 sin α ,

x′2 = x1 sin α + x2 cos α ,
x′kℓ = −xkℓ , (2.29)
′
ykℓ = −ykℓ , 1 ≤ k < ℓ ≤ n.

Taking into account the explicit form of d1 , d2 , ukℓ , vkℓ , 1 ≤ k < ℓ ≤ 3, one finds
that the corresponding map ϕ[α] ≡ ϕ[R(α), 0] : M3 (C) → M3 (C) has the form

6
a′ = ϕ[α](a), where
1
a′11 = [λ(α)a11 + µ(α)a22 + ν(α)a33 ] ,
2
1
a′22 = [ν(α)a11 + λ(α)a22 + µ(α)a33 ] ,
2
1
a′33 = [µ(α)a11 + ν(α)a22 + λ(α)a33 ] ,
2
1
a′ij = − aij , i 6= j , (2.30)
2
and
2
λ(α) = (1 + cos α) ,
3 √
2 1 3 
µ(α) = 1 − cos α − sin α ,
3 2 √2
2 1 3 
ν(α) = 1 − cos α + sin α ,
3 2 2
λ(α) + µ(α) + ν(α) = 2 . (2.31)

The special cases are:

(i) a′ = ϕ[α = π/3](a)
1
a′11 = (a11 + a33 ) ,
2
1
a′22 = (a22 + a11 ) ,
2
1
a′33 = (a33 + a22 ) ,
2
1
a′ij = − aij .
2

(ii) a′ = ϕ[α = −π/3](a)

1
a′11 = (a11 + a22 ) ,
2
1
a′22 = (a22 + a33 ) ,
2
1
a′33 = (a33 + a11 ) ,
2
1
a′ij = − aij .
2

7
(iii) a′ = ϕ[α = π](a) = 1/2(1l3 Tra − a)

(iv) a′ = ϕ[α = 0](a) = 1/3ϕ[α = π](a) + (1 − 1/3)ψ(a)

a′′ = ψ(a) ,
a′′11 = a11 , a′′22 = a22 , a′′33 = a33 ,
1
a′′ij = − aij .
2

The maps (i) and (ii) are indecomposable Choi’s maps, the map (iii) is completely
copositive, while (iv) is decomposable.
Let [xij ] ∈ M3 (M3 ) be the matrix
 
1 0 0 0 1 0 0 0 1
 

 0 p 0 0 0 0 0 0 0 

 

 0 0 p−1 0 0 0 0 0 0 

0 0 0 p−1 0 0 0 0 0
 
 
 
[xij ] = 

1 0 0 0 1 0 0 0 1 
 p > 0. (2.32)
 

 0 0 0 0 0 p 0 0 0 

 

 0 0 0 0 0 0 p 0 0 

 

 0 0 0 0 0 0 0 p−1 0 

1 0 0 0 1 0 0 0 1

Then both [xij ] and [xji ] belong to M3 (M3 )+ . It is easily seen that the matrix
[ϕ[α](xij )] is not positive for 0 < α < π, i.e., the map ϕ[α] is indecomposable for
0 < α < π.
It should be pointed out that the indecomposable maps (1.4) does not belong to
the class considered above.

References
[1] W. F. Stinespring, Proc. Amer. Math. Soc. 6, 211 (1955).

[2] E. Størmer, Acta Math. 110, 233 (1963).

[3] E. Størmer, Trans. Amer. Math. Soc. 120, 438 (1965).

[4] E. Størmer, in: Lecture Notes in Physics 29, Springer Verlag, Berlin, 1974,
pp. 85–106.

8
 [5] W. Arverson, Acta Math. 123, 141 (1969).

[6] M. D. Choi, Lin. Alg. Appl. 10, 285 (1975).

[7] M. D. Choi, Lin. Alg. Appl. 12, 95 (1975).

[8] M. D. Choi, J. Operator Theory 4, 271 (1980).

[9] E. Størmer, Math. Ann. 247, 21 (1980).

[10] E. Størmer, Proc. Amer. Math. Soc. 86, 402 (1982).

[11] E. Størmer, J. Funct. Anal. 66, 235 (1985).

[12] S. L. Woronowicz, Rep. Math. Phys. 10, 165 (1976).

[13] S. L. Woronowicz, Comm. Math. Phys. 51, 243 (1976).

[14] A. Jamiokowski, Rep. Math. Phys. 3, 275 (1972).

[15] L. E. Labuschagne, W. A. Majewski, M. Marciniak, to appear.

[16] V. Gorini, E. C. G. Sudarshan, Comm. Math. Phys. 46, 43 (1976).

[17] B. M. Terhal, Lin. Alg. Appl. 323, 61 (2001).

[18] A. G. Robertson, Math. Proc. Camb. Phil. Soc. 94, 291 (1983).

[19] A. G. Robertson, Proc. Roy. Soc. Edinburgh 94 A, 71 (1983).

[20] A. G. Robertson, J. Lond. Math. Soc. (2) 32, 133 (1985).

[21] T. Ando, seminar note 1985.

[22] J. Tomiyama, Contemporary Math. 62, 357 (1987).

[23] K. Tanahashi, J. Tomiyama, Canadian Math. Bull. 31, 308 (1988).

[24] T. Takasaki, J. Tomiyama, Math. Japonica 1, 129 (1982).

[25] S. H. Kye, in: Elementary operators and applications, M. Mathion, ed., World
Scientific, 1992.

[26] S. J. Cho, Lin. Alg. Appl. 171, 213 (1992).

[27] H. Osaka, Lin. Alg. Appl. 153, 73 (1991).

[28] H. Osaka, Publ. RIMS Kyoto Univ. 28, 747 (1992).

9
[29] H. Osaka, Lin. Alg. Appl. 186, 45 (1992).

[30] K. C. Ha, Publ. RIMS Univ. 34, 591 (1998).

[31] K. C. Ha, Lin. Alg. Appl. 359, 277 (2003).

[32] M. Horodecki, P. Horodecki, R. Horodecki, Phys. Rev. Lett A 223, 1 (1996).

[33] G. Albert, T. Beth, M. Horodecki, P. Horodecki, R. Horodecki, M. Rötter, H.

Weinfurter, R. Werner, A. Zeinlinger, Quantum Information, Springer Tracts in
Modern Physics 173, Springer Verlag, Berlin, 2001.

[34] E. C. G. Sudarshan, P. M. Mathews, J. Rau, Phys. Rev. 121, 920 (1961).

[35] F. T. Hioe, J. H. Eberly, Phys. Rev. Lett. 47, 838 (1981).

10