Qit 230730

Positive Maps in Quantum Information Theory
Seung-Hyeok Kye
Department of Mathematics
Seoul National University
Seoul 151-742, Korea
Lecture notes in the fall semester 2022

at Seoul National University
July 30, 2023

2
Preface
Positive linear functionals and positive linear maps had been playing important
roles in the theory of operator algebras, which reflect noncommutative order struc-
tures. Such order structures provide basic mathematical frameworks for current
quantum information theory. The main purpose of this lecture note is to introduce
basic notions like separability/entanglement and Schmidt numbers from quantum
information theory in terms of positive maps between matrix algebras.
Basic tools are Choi matrices and duality arising from bilinear pairing between
matrices. We begin with concrete examples of positive maps Ads which sends x
to s˚ xs, and define separability/entanglement and Schmidt numbers in terms of
Choi matrices. We also use duality to introduce various kinds of positivity, like k-
positivity and complete positivity. Positive maps which are not completely positive
are indispensable tools to detect entanglement through the duality.
In Chapter 1, we introduce the above notions and exhibit nontrivial examples of
positive maps. We also provide a unified argument to recover various known criteria
through ampliation. We will focus in Chapter 2 on the issue how positive maps
detect entanglement. Through the discussion, exposed faces of the convex cones of
all positive maps play important roles. We exhibit three classes of positive maps,
by Choi, Woronowicz and Robertson in 1970’s and 1980’s which generate exposed
extreme rays of the convex cone of all positive linear maps.
This is the collection of lecture notes during the fall semester of 2022, at Seoul
National University, Seoul, Korea. The author tried to minimize preliminaries, re-
quiring only undergraduate linear algebra. The author is grateful to all the audiences
for their feedbacks on the notes. Special thanks are due to Kyung Hoon Han for
his careful reading of the drafts. Nevertheless, any faults in this lecture notes are
responsibility of the author.
February 2023
3
4
Contents
1 Positive Maps and Bi-partite States 7

1.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.1.1 Vectors and matrices . . . . . . . . . . . . . . . . . . . . . . . 8
1.1.2 Positive matrices . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.1.3 Tensor products . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.1.4 Singular value decomposition . . . . . . . . . . . . . . . . . . 12
1.1.5 Ranks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.1.6 Boundary of a convex set . . . . . . . . . . . . . . . . . . . . . 14
1.1.7 convex cones . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.2 Positive maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.2.1 Positive maps between matrix algebras . . . . . . . . . . . . . 17
1.2.2 Extremal positive maps . . . . . . . . . . . . . . . . . . . . . 18
1.2.3 k-superpositive maps . . . . . . . . . . . . . . . . . . . . . . . 21
1.3 Positive maps between 2 ˆ 2 matrices . . . . . . . . . . . . . . . . . . 23
1.3.1 Congruence maps and orthogonal transformations . . . . . . . 23
1.3.2 Decomposability of positive maps . . . . . . . . . . . . . . . . 26
1.4 Choi matrices and separable states . . . . . . . . . . . . . . . . . . . 28
1.4.1 Choi matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.4.2 Schmidt numbers and entanglement . . . . . . . . . . . . . . . 30
1.4.3 Partial transposes . . . . . . . . . . . . . . . . . . . . . . . . . 33
1.4.4 Entanglement with positive partial transpose . . . . . . . . . . 35
1.5 Duality and completely positive maps . . . . . . . . . . . . . . . . . . 38
1.5.1 Dual cones of convex cones . . . . . . . . . . . . . . . . . . . . 38
1.5.2 k-positive maps and completely positive maps . . . . . . . . . 40
1.5.3 Faces for completely positive maps . . . . . . . . . . . . . . . 43
1.5.4 Decomposable maps and PPT states . . . . . . . . . . . . . . 44
1.6 Nontrivial examples of positive maps . . . . . . . . . . . . . . . . . . 46
1.6.1 Tomiyama’s example of k-positive maps . . . . . . . . . . . . 46
1.6.2 The Choi map between 3 ˆ 3 matrices . . . . . . . . . . . . . 49
1.6.3 The Woronowicz map from 2 ˆ 2 matrices to 4 ˆ 4 matrices . 51
5
1.7 Isotropic states and Werner states . . . . . . . . . . . . . . . . . . . . 52
1.7.1 Isotropic states . . . . . . . . . . . . . . . . . . . . . . . . . . 52
1.7.2 Werner states . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
1.8 Mapping cones and tensor products . . . . . . . . . . . . . . . . . . . 59
1.8.1 Mapping cones of positive maps . . . . . . . . . . . . . . . . . 59
1.8.2 Tensor products and compositions of linear maps . . . . . . . 61
1.8.3 Roles of ampliation . . . . . . . . . . . . . . . . . . . . . . . . 62
1.8.4 Tensor products of positive maps . . . . . . . . . . . . . . . . 64
1.8.5 Entanglement breaking maps . . . . . . . . . . . . . . . . . . 65
1.9 Historical remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
2 Detecting Entanglement by Positive Maps 71

2.1 Exposed faces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
2.1.1 Dual faces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
2.1.2 Exposed positive maps . . . . . . . . . . . . . . . . . . . . . . 75
2.2 The Choi map revisited . . . . . . . . . . . . . . . . . . . . . . . . . . 79
2.2.1 The Choi map and completely positive maps . . . . . . . . . . 79
2.2.2 The Choi map and decomposable maps . . . . . . . . . . . . . 82
2.3 Maximal faces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
2.3.1 Boundary of convex cones . . . . . . . . . . . . . . . . . . . . 83
2.3.2 Maximal faces and minimal exposed faces . . . . . . . . . . . 86
2.3.3 Boundaries of k-positive maps . . . . . . . . . . . . . . . . . . 87
2.4 Entanglement detected by positive maps . . . . . . . . . . . . . . . . 89
2.4.1 Optimal entanglement witnesses . . . . . . . . . . . . . . . . . 89
2.4.2 Spanning properties . . . . . . . . . . . . . . . . . . . . . . . . 94
2.5 Positive maps of Choi type . . . . . . . . . . . . . . . . . . . . . . . . 97
2.5.1 Choi type positive maps between 3 ˆ 3 matrices . . . . . . . . 97
2.5.2 Spanning properties of Choi type positive maps . . . . . . . . 101
2.5.3 Lengths of separable states . . . . . . . . . . . . . . . . . . . . 105
2.6 Exposed positive maps by Woronowicz . . . . . . . . . . . . . . . . . 107
2.6.1 A dimension condition for exposed positive maps . . . . . . . 107
2.6.2 Exposed positive maps between 2 ˆ 2 and 4 ˆ 4 matrices . . . 110
2.7 Exposed positive maps by Robertson . . . . . . . . . . . . . . . . . . 113
2.7.1 Robertson’s positive maps between 4 ˆ 4 matrices . . . . . . . 113
2.7.2 Exposedness of the Robertson map . . . . . . . . . . . . . . . 117
Index 131
6
Chapter 1
Positive Maps and Bi-partite

States
In this chapter, we introduce various kinds of positive maps between matrix algebras
together with the corresponding notions for the tensor products of matrix algebras.
We begin with the most elementary positive maps Ads for given matrices s, which
send x to s˚ xs. These maps give rise to important classes of positive maps according
to the ranks of s, by taking convex combinations. Positive maps arising in this
way with matrices whose ranks are at most k are called k-superpositive maps, and
all of them make the class of completely positive maps which play crucial roles.
Another important positive map is given by the transpose, which makes the class of
decomposable positive maps together with completely positive maps.
For a given linear map between matrix algebras, we assign a matrix in the
tensor product of matrices, or equivalently a block matrix, which is called the Choi
matrix. The block matrices corresponding to completely positive maps are precisely
positive (semi-definite) matrices, which represent bi-partite states after normalizing
by the trace. The class of 1-superpositive maps corresponds to separable states.
Nonseparable states are called entanglement, which is now recognized as one of the
most important resources in the current quantum information theory.
Another main tool is the bilinear pairing between matrices which gives rise to
the bilinear pairing between linear maps through the Choi matrices. The dual
notion of k-superpositivity by this bilinear pairing is k-positivity, which has been
studied among operator algebraists since Stinespring’s representation theorem in the
1950’s. The usual positivity of linear maps coincides with 1-positivity arising in this
way. Therefore, the positivity of linear maps between matrix algebras is just the
dual notion of separability of bi-partite states, and this is why positive maps play
important roles to detect entanglement.
After we fix notations together with several preliminaries in Section 1.1, we show
in Section 1.2 that the map Ads generates an extreme ray of the convex cone of all
7
positive linear maps, and introduce the notion of k-superpositive maps and decom-
posable maps. In Section 1.3, we show that all positive maps are decomposable in
low dimensional cases like maps between 2 ˆ 2 matrices. In Section 1.4, we define
Choi matrices through which we introduce various kinds of bi-partite states, under
the name of Schmidt numbers. We define the bilinear pairing between linear maps in
Section 1.5 and get the notion of k-positivity which is dual to k-superpositivity. Dur-
ing the discussion, we see that complete positivity of maps corresponds to positivity
of block matrices through the Choi matrices. Sections 1.6 and 1.7 will be devoted
to introduce various examples which distinguish k-superpositivity and k-positivity
of maps together with the corresponding states which include Werner states and
isotropic states. We also see concrete examples of indecomposable positive maps.
Finally, we introduce the notion of mapping cones in Section 1.8 with which we
obtain various characterizations of aforementioned notions in terms of ampliation
which is the tensor product with the identity map. Our exposition of topics reverses
historical development in parts, and this is why we add historical remarks in Section
1.9.
1.1 Preliminaries
1.1.1 Vectors and matrices
A vector in the vector space Cn over the complex field is denoted by a ket |vy, which
may be understood as a column vector or an n ˆ 1 matrix;
¨ ˛
v1
˚ v2 ‹
|vy “ ˚ . ‹ P Cn .
˚ ‹
˝ .. ‚
vn
The adjoint of ket |vy is denoted by a braxv|, that is,
xv| “ pv̄1 , v̄2 , ¨ ¨ ¨ , v̄n q P Cn .
For given two vectors |vy and |wy in Cn , we have
xv|wy “ v̄1 w1 ` v̄2 w2 ` ¨ ¨ ¨ ` v̄n wn ,
and so xv|wy is the standard inner product of Cn which is linear in the second
variable and conjugate-linear in the first variable. So, xv| may be considered as the
linear functional on Cn which sends |wy to xv|wy. On the other hand, we see that
¨ ˛
v1 w̄1 v1 w̄2 ¨ ¨ ¨ v1 w̄n
˚ v2 w̄1 v2 w̄2 ¨ ¨ ¨ v2 w̄n ‹
˚ ‹
|vyxw| “ ˚ . .. .. .. ‹
˝ .. . . . ‚
vm w̄1 vm w̄2 ¨ ¨ ¨ vm w̄n
8
is an m ˆ n matrix of rank one, for |vy P Cm and |wy P Cn . This is nothing but
|vyxw| : |uy ÞÑ |vyxw|uy “ xw|uy|vy P Cm , |uy P Cn ,
as a linear map from Cn into Cm with the one dimensional range space spanned by
|vy. The set Mmˆn of all m ˆ n matrices is a vector space over the complex field.
When m “ n, we use the notation Mn , which is usually called the matrix algebra
since two square matrices with the same size can be multiplied.
The vectors appearing in the standard orthonormal basis tei : i “ 1, 2, . . . , nu of
n
C are denoted by ¨ ˛
0
˚ .. ‹
˚.‹
˚ ‹
˚0‹
˚ ‹
˚1‹ Ð i-th,
|iy “ ˚ ‹
˚0‹
˚ ‹
˚ .. ‹
˝.‚
0
řn
so we may write |vy “ i“1 vi |iy. If |iy P Cm and |jy P Cn then
|iyxj|
is the mˆn matrix whose entries are zeros except for pi, jq-entry 1. So, the collection
t|iyxj| : i “ 1, 2, . . . , m, j “ 1, 2, . . . , nu
is the standard matrix units. The m ˆ n matrix A “ raij s may be written by

¨ ˛
a11 a12 ¨ ¨ ¨ a1n
ÿ ˚ a21 a22 ¨ ¨ ¨ a2n ‹
A“ aij |iyxj| “ ˚ . .
˚ ‹
. . .
i,j
˝ .. .. .. .. ‹
‚
am1 am2 ¨ ¨ ¨ amn
When we consider concrete examples, we sometimes use notations |0y, |1y, . . . , be-
ginning with |0y for orthonormal basis. In this case, we have
ˆ ˙ ˆ ˙
1 2 0
|0y “ PC , |1y “ P C2
0 1
for the two dimensional case.
1.1.2 Positive matrices

A self-adjoint n ˆ n matrix A is called positive semi-definite or just positive if
xx|A|xy ě 0 (1.1)
9
for every |xy P Cn . Every |xy P Cn gives rise to the rank one matrix |xyxx|, which is
positive since
xy|xyxx|yy “ |xx|yy|2 ě 0
for every |yy P Cn . If |xy is a unit vector then |xyxx| is the projection onto the one
dimensional space spanned by |xy. By the spectral decomposition, we know that
every positive matrix can be written by
ÿ
|xι yxxι |
ι
with a finite family t|xι yu of vectors. It should be noted that this expression is far
from being unique. By another application of the spectral decomposition, we know
that a positive matrix is the square of a matrix. We write A ď B for Hermitian
matrices A and B if B ´ A is positive.
For two matrices A “ raij s and B “ rbij s with the same sizes, we define
ÿ
xA, By “ Tr pAB T q “ aij bij ,
i,j
where B T and Tr pAB T q denote the transpose of B and the trace of AB T , respec-
tively. This is a bilinear pairing which is non-degenerate, that is, satisfies
xA, By “ 0 for every B ðñ A “ 0.
It should be noted that xA, Ay may be negative. For an example, we take a self-
adjoint ˆ ˙
0 i
A“ ,
í 0
to get xA, Ay “ ´2 ă 0. We have the following relations
xA, BCy “ xAC T , By “ xB T A, Cy,
whenever the sizes of matrices are given so that the above relation is meaningful.
We also note that the identity
|xyxy|T “ xy|T |xyT “ |ȳyxx̄|

ř
holds with the notation |x̄y “ i x̄i |iy. Therefore, we also have
xA, |x̄yxȳ|y “ Tr pA|x̄yxȳ|T q

“ Tr pA|yyxx|q (1.2)
“ Tr pxx|A|yyq “ xx|A|yy.
Especially, we have
xx|A|xy “ xA, |x̄yxx̄|y
10
for every A P Mn and |xy P Cn . Therefore, we see that A P Mn is positive if and
only if xA, By ě 0 for every positive B.
We note that every linear functional % on the space Mn is of the form
% : A ÞÑ xA, B% y, A P Mn , (1.3)
for an n ˆ n matrix B% . Then % is positive, that is, sends a positive matrix to a

nonnegative real numbers if and only if B% is positive. Furthermore, it is easily seen
that % is unital, that is, sends the identity matrix to 1, if and only if B% is of trace
one. A positive matrix % of trace one is called a density matrix, or a state, which is
of the form ÿ
%“ pi |ξi yxξi |, (1.4)
i
ř
for a probability distribution tpi u with i pi “ 1 and a family t|ξi yu of unit vectors.
We note that the set, denoted by Dn , of all states in Mn is a convex set whose
elements are convex combinations of one dimensional projections t|ξyxξ|u with unit
vectors t|ξyu, which are called pure states. Sometimes, a positive matrix itself is
called an (unnormalized) state.
1.1.3 Tensor products

ř
An m ˆ n matrix A “ i,j aij |iyxj| corresponds to the vector
ÿ
xÃ| “ aij xi|xj| P Cm b Cn , (1.5)
i,j
where we usually delete the tensor notation in xi| b xj| to write as xi|xj|, or even as
xij|. If we endow txi|xj|u with the lexicographic order then xÃ| is the concatenation
xÃ| “ pa11 , . . . , a1n , a21 , . . . , a2n , . . . , . . . , am1 , . . . , amn q P Cmn
of row vectors of the matrix A.

The tensor product A b B of two matrices A and B is defined by
˜ ¸ ˜ ¸
ÿ ÿ ÿ
AbB “ aij |iyxj| b bk` |kyx`| “ aij bk` |iy|kyxj|x`|.
ij k i,j,k,`
If we endow t|iy|kyu and t|jy|`yu with the lexicographic orders again, then A b B
has the matrix form. For example, when both A and B are 2 ˆ 2 matrices, we have
¨ ˛
a11 b11 a11 b12 a12 b11 a12 b12
˚a11 b21 a11 b22 a12 b21 a12 b22 ‹
AbB “˚ ˝a21 b11 a21 b12 a22 b11 a22 b12 ‚.
‹
a21 b21 a21 b22 a22 b21 a22 b22
11
In this way, a matrix in Mm b Mn may be identified as an m ˆ m block matrix
whose entries are n ˆ n matrices. So, we may identify Mm b Mn with Mm pMn q.
Especially, Im b B is the m ˆ m diagonal block matrix whose diagonal block is given
by B P Mn , where Im denotes the m ˆ m identity matrix. Sometimes, we denote
by I for the identity matrix. On the other hand, A b In is the m ˆ m block matrix
whose pi, jq block is given by aij In P Mn . Note that
ÿ
pA b Bq|jy|`y “ aij bk` |iy|ky
i,k
˜ ¸ ˜ ¸
ÿ ÿ
“ aij |iy b bk` |ky “ A|jy b B|`y,
i k
and so A b B sends |jy|`y to A|jy b B|`y as a linear map from Cm b Cn into itself.
The following identities are easily checked;
pC b DqT “ C T b DT ,
pA b BqpC b Dq “ AC b BD,
Tr pA b Bq “ Tr pAqTr pBq,
whenever they are defined. Suppose that A and C are m ˆ n matrices, and B, D
are k ˆ ` matrices. Using the above relations, we also have the identity
xA b B, C b Dy “ xA, CyxB, Dy. (1.6)
For linear maps φ1 : Mm1 Ñ Mn1 and φ2 : Mm2 Ñ Mn2 , we define the linear map
φ1 b φ2 : Mm1 b Mm2 Ñ Mn1 b Mn2

řm1
as follows: Every element of Mm1 b Mm2 is given by A “ i,j“1 |iyxj| b Aij with
Aij P Mm2 in a unique way. We define
m1
ÿ
pφ1 b φ2 qpAq “ φ1 p|iyxj|q b φ2 pAij q P Mn1 b Mn2 .
i,j“1
Then we see that φ b ψ sends A b B P Mm1 b Mm2 to φpAq b ψpBq P Mn1 b Mn2 .
1.1.4 Singular value decomposition

Let A be an arbitrary m ˆ n matrix. Then A˚ A is an n ˆ n positive matrix with
eigenvalues
λ1 ě λ2 ě ¨ ¨ ¨ ě λr ą 0 “ λr`1 “ ¨ ¨ ¨ “ λn ,
with r ď m and r ď n. Denoting by D the r ˆ r diagonal matrix with the diagonal
entries λ1 , . . . , λr , we have
ˆ ˙
˚ ˚ D 0
V pA AqV “
0 0
12
` ˘
for an n ˆ n unitary matrix V . We write V “ V1 V2 with the n ˆ r matrix V1
and the n ˆ pn ´ rq matrix V2 . Then we have
V1˚ pA˚ AqV1 “ D P Mr , V2˚ pA˚ AqV2 “ 0 P Mn´r .
Define the m ˆ r matrix U1 by U1 “ AV1 D´1{2 with the obvious meaning for D´1{2 .
Then we have
U1˚ U1 “ D´1{2 V1˚ A˚ AV1 D´1{2 “ Ir ,
` ˘
and so we can take an m ˆ pm ´ rq matrix U2 so that U “ U1 U2 is an m ˆ m
unitary. We have
U1˚ AV1 “ U1˚ AV1 D´1{2 D1{2 “ U1˚ U1 D1{2 “ D1{2 ,

U2˚ AV1 “ U2˚ AV1 D´1{2 D1{2 “ U2˚ U1 D1{2 “ 0.
From V2˚ pA˚ AqV2 “ 0, we also have AV2 “ 0, and so we have

ˆ ˙ ˆ ˙
˚ U1˚ AV1 U1˚ AV2 D1{2 0
U AV “ “ ,
U2˚ AV1 U2˚ AV2 0 0
as m ˆ n matrices.
Therefore, the m ˆ n matrix A is written by
A “ U ΣV ˚ , (1.7)
with the matrix Σ given by

r
ÿ 1{2
Σ“ λk |kyxk| P Mmˆn ,
k“1
and unitary matrices U P Mm and V P Mn . This is known as the singular value

1{2 1{2
decomposition of A, and λ1 , . . . , λr are called the singular values of A. By the
identity ˆ 1{2 ˙ ˆ ˙ ˆ 1{2 ˙
D 0 Ir 0 D 0
Σ“ “ ,
0 0 0 0 0 In´r
we also see that A can be written as
ˆ ˙
Ir 0
A“U W, (1.8)
0 0
with invertible matrices U P Mm and W P Mn . This can be also obtained from row
and column elementary operations.
References: [11]
13
1.1.5 Ranks
For a matrix A, we denote by Im A the range space which is the span of column
vectors. The span of the conjugates of row vectors is the orthogonal complement
of the kernel space ker A. By the dimension theorem, the span of columns and the
span of rows share the same dimension, which is called the rank of A and is denoted
by rank A. We see that ker A˚ A “ ker A which implies rank A “ rank A˚ A, and so
the number r in the singular value decomposition (1.7) is the rank of A.
ř
When A “ i,j aij |iyxj|, the j-th column vector is given by
ÿ
A|jy “ aij |iy.
i
ř
Take any basis t|ξk y : k “ 1, 2, . . . , rank Au of Im A and write A|jy “ k sjk |ξk y,
then we have ˜ ¸
ÿn ÿ ÿ
A “ AIn “ A|jyxj| “ |ξk y sjk xj| .
j“1 k j
Therefore, we have
rank
ÿA
A“ |ξk yxηk | (1.9)
k“1
ř
with xηk | “ j sjk xj|, and we see that it is possible to express A with the sum
of a family of rank one matrices with the cardinality rank A. In the singular value
decomposition (1.7), we take column vectors t|ξi yu and t|ηj yu of the unitary matrices
U and V , respectively. Then we have
´ř ¯ ´ř ¯˚
ř 1{2
A “ p i |ξi yxi|q λ
k k |kyxk| j |ηj yxj|
´ř ¯ ´ř ¯ rank
ÿA 1{2
ř 1{2
“ p i |ξi yxi|q λ
k k |kyxk| j |jyxηj | “ λk |ξk yxηk |.
k“1
Therefore, vectors t|ξk yu and t|ηk yu in (1.9) can be chosen to be orthogonal.

A vector of the form |ξy b |ηy in Cm b Cn is called a product vector which
corresponds to the m ˆ n matrix |ξyxη̄| of rank one under the correspondence (1.5).
The Schmidt rank of a vector |ζy P Cm b Cn , denoted by SR |ζy, is the smallest
number of product vectors whose sum is |ζy. By the correspondence (1.5) and
the expression (1.9), we see that the Schmidt rank coincides with the rank of the
corresponding matrix.
1.1.6 Boundary of a convex set

Suppose that C is a nonempty convex set in a real vector space Rν . A point x of C
is called a relative interior point or just an interior point of C if for each y P C there
14
rz

y
r hhrr
h
rx0
hhhh
hhhh
hhhhr
Figure 1.1: If the line segment from an interior point x0 to y can be extended, then
the line segment from any point z to y also can be extended
is t ą 1 such that p1 ´ tqy ` tx P C. This means that the line segment from every
y P C to x may be extended within C. The set of all interior points of C will be
denoted by int C, which coincides with the relative interior of C with respect to the
affine manifold generated by C. From this, one may see that int C is nonempty for
any nonempty convex set C. A point y P C which is not an interior point is called
a boundary point, and the set of all boundary points of C is denoted by BC.
We fix an interior point x0 of a convex set C. Suppose that y P C satisfies the
condition suptt : p1 ´ tqx0 ` ty P Cu “ 1. That is, we suppose that the line segment
from x0 to y cannot be extended within C. Then it is clear that y is a boundary
point of C. Suppose that y does not satisfy the condition. Then there is t ą 1
such that p1 ´ tqx0 ` ty P C. For an arbitrary z P C we can take s ă 0 such that
p1 ´ sqx0 ` sz P C since x0 is an interior point. Put r “ spt´1q
t´s
. Then we see that
r ă 0 and
1´s t´1
p1 ´ rqy ` rz “ rp1 ´ tqx0 ` tys ` rp1 ´ sqx0 ` szs P C.
t´s t´s
This shows that y is an interior point of C. See Figure 1.1. Therefore, we conclude
that y is an interior point of C if and only if the line segment from a single interior
point x0 to y can be extended within C.
A nonempty convex subset F of a convex set C is called a face of C if the
following property
x0 , x1 P C, 0 ă t ă 1, p1 ´ tqx0 ` tx1 P F ùñ x0 , x1 P F
holds. This means that if an interior point of a line segment in C belongs to F

then the line segment itself is contained in F . An extreme point is nothing but a
face consisting of a single point. A face F of a convex set C is called proper if it
is a proper subset of C. If a point of a face F is an interior point of C then we
see that F “ C. Therefore, a proper face must be contained in the boundary. The
intersection of an arbitrary family of faces is again a face whenever it is nonempty,
and so every point x of a convex set has the smallest face F containing the point.
15
We see that x is an interior point of F , and a convex set is completely partitioned
into interiors of faces.
Suppose that C is a compact convex set. If x P C is not an extreme point then it
is an interior point of a line segment. We extend this line segment until it meets the
boundary, and so we see that x is a convex combination of two boundary points. If
one of them is not an extreme point and contained in a proper face then we express
it as a convex combination of two boundary points of the face. We continue this
process to see that every point of a compact convex set in a finite dimensional space
is a convex combination of extreme points. Carathéodory theorem tells us that every
point of a d-dimensional convex set is the convex combination of extreme points with
the cardinality at most d ` 1. An important consequence is that the convex hull of
a compact set in a finite dimensional space is again compact. It should be noted
that this is not the case for infinite dimensional spaces.
References: [99], [127], [74]
1.1.7 convex cones

Recall that a subset of a real vector space is called a convex cone if it is closed under
summations and scalar multiplications by nonnegative numbers. A point x0 P K is
said to generate an extreme ray of a convex cone K or to be extremal in K if the
ray ttx0 : t ě 0u itself is a face of K. This is the case if and only if x0 “ x1 ` x2
with x1 , x2 P K implies that x1 is a nonnegative scalar multiplication of x0 . In this
case, the ray ttx0 : t ě 0u is called an extreme ray. Sometimes, the point x0 itself is
called extreme.
The set of all n ˆ n positive matrices, denoted by Mn` , is a closed convex cone
in the real vector space Mnh of all Hermitian matrices. We note that a positive
matrix generates an extreme ray of Mn` if and only if it is of rank one. We see that
A P Mn` is an interior point of Mn` if and only if rank A “ n, or equivalently A is
non-singular. Especially, the identity matrix In is an interior point of Mn` . The ray
generated by In may be considered to be located at the center of the convex cone
Mn` . In fact, if we take a projection P onto a proper subspace then we have
" *
1
sup t P R : p1 ´ tqP ` t ¨ In ě 0 “ 2.
2
If we extend the line segment from P to 12 In until the line meets the boundary of
Mn` , then the line segment meets at the orthogonal complement In ´ P , and 12 In is
located at the center of the line segment.
For a positive matrix A, we consider the range vectors of A and take the convex
cone F generated by t|ξyxξ| : |ξy P Im Au, then A is an interior point of F . Especially,
16
for any |ξy P Im A there exists t ą 1 such that p1 ´ tq|ξyxξ| ` tA is positive. The
convex cone F is the smallest face containing A, and every face of Mn` arises in this
way.
We note that a convex cone K itself has no extreme point except zero. The set Dn
of all density matrices is the intersection of the convex cone Mn` and the hyperplane
determined by the condition that the trace is one. If we take a hyperplane H which
does not contain zero but meets an extreme ray of K then the intersection is just a
single point which is an extreme point of the convex set K X H.
1.2 Positive maps

We first note that every positive map between matrix algebras is sitting in the real
vector space of all Hermiticity preserving maps. We will show that the map Ads
defined by Ads pxq “ s˚ xs generates an extreme ray of the convex cone P1 of all
positive linear maps. Using those maps together with the transpose map, we define
the notions of k-superpositivity and decomposability for positive maps.
1.2.1 Positive maps between matrix algebras

The vector space of all linear maps from Mm into Mn will be denoted by LpMm , Mn q
which is of m2 n2 dimension over the complex field. A linear map φ : Mm Ñ Mn
is called Hermiticity preserving if φpaq P Mn is Hermitian whenever a P Mm is
Hermitian. When φ is Hermiticity preserving, we write a “ b ` ic with Hermitian
matrices b and c, then we have
φpa˚ q “ φpb ´ icq “ φpbq ´ iφpcq “ pφpbq ` iφpcqq˚ “ φpaq˚ .
In short, φ satisfies the following relation
φpa˚ q “ φpaq˚ , a P Mm . (1.10)
If φ satisfies the property (1.10) and a is Hermitian then we have φpaq˚ “ φpa˚ q “
φpaq, and so φpaq is Hermitian. Therefore, we see that φ is Hermiticity preserving
if and only if it satisfies (1.10). The restriction of a Hermiticity preserving map
h h
φ : Mm Ñ Mn on Mm is a real linear map from Mm into Mnh . Conversely, every real
h
linear map φ : Mm Ñ Mnh can be extended to the complex linear map φ̃ : Mm Ñ Mn
by
φ̃pa ` ibq “ φpaq ` iφpbq. (1.11)
The set HpMm , Mn q of all Hermiticity preserving maps is a vector space over the
real field with the real dimension m2 n2 .
17
A linear map φ : Mn Ñ Mm is called positive if it sends positive matrices to
positive matrices. Because every Hermitian matrix is the difference of two positive
matrices, it is clear that every positive linear map is Hermiticity preserving. The
set of all positive maps from Mm into Mn will be denoted by P1 rMm , Mn s, which is
a closed convex cone in the real vector space HpMm , Mn q.
For a given linear map φ : Mm Ñ Mn , the adjoint map φ˚ : Mn Ñ Mm is defined
by
xφ˚ pxq, yy “ xx, φpyqy, x P Mn , y P Mm .
It is easy to see that the following relation
pφ ˝ ψq˚ “ ψ ˚ ˝ φ˚
holds whenever the composition φ ˝ ψ is possible. A linear map φ : Mm Ñ Mn is

positive if and only if xφpxq, yy ě 0 for every positive x P Mm and positive y P Mn
if and only if xx, φ˚ pyqy ě 0 for every positive x and y if and only if φ˚ : Mn Ñ Mm
is a positive linear map. In short, we have seen that φ is positive if and only if its
adjoint map φ˚ is positive.
1.2.2 Extremal positive maps

For a given m ˆ n matrix s, we define the linear map Ads : Mm Ñ Mn by
Ads pxq “ s˚ xs, x P Mm .
It is clear that Ads is a positive map for any matrix s. The map Adu is called a
congruence map when u is a unitary. It is easy to see that the following identities
Adst “ Adt ˝ Ads , pAds q˚ “ AdsT
hold for matrices s and t, whenever st is defined. It is also clear that if s P Mn is

nonsingular then Ads : Mn Ñ Mn is an order isomorphism. Recall that a bijective
map φ : Mn Ñ Mn is called an order isomorphism when both φ and φ´1 are positive,
that is, φ satisfies
s ď t ðñ φpsq ď φptq. (1.12)
A linear map which generates an extreme ray of the convex cone P1 is called
an extremal positive map. It is clear that φ is extremal positive if and only if so is
φ˚ . If φ is extremal positive and σ is a linear order isomorphism then both σ ˝ φ
and φ ˝ σ are also extremal whenever they are defined, because both φ ÞÑ σ ˝ φ and
φ ÞÑ φ ˝ σ define affine isomorphisms between P1 rMm , Mn s. We will show that the
map Ads is an extremal positive map. To do this, we need the following:
18
Proposition 1.2.1 Suppose that a P Mn commutes every b P Mn . Then a is a
scalar multiple of the identity matrix.
ř
Proof. Write a “ i,j aij |iyxj|. Then |kyx`|a “ a|kyx`| implies
x`|a|ky “ xk|kyx`|a|ky “ xk|a|kyx`|ky “ 0,
when k ‰ `, and so a must be a diagonal matrix. We also have
x`|a|`y “ xk|kyx`|a|`y “ xk|a|kyx`|`y “ xk|a|ky,
for each k, `. ˝
The identity map id “ AdI with the identity matrix I is the key case for the
extremeness of the maps Ads . Sometimes, we denote by idn for the identity map on
Mn .
Lemma 1.2.2 The identity map idn “ AdIn is extremal in P1 rMn , Mn s.
Proof. Suppose that both φ and idn ´ φ are positive. Then for every unit vector
|ξy P Cn , we have 0 ď φp|ξyxξ|q ď |ξyxξ|, and so there exists λξ with 0 ď λξ ď 1 such
that
φp|ξyxξ|q “ λξ |ξyxξ|,
because |ξyxξ| generates an extreme ray of the convex cone Mn` . We have to show
that λξ “ λη for any unit vectors |ξy and |ηy. We note that φppq ď p for every
projection in Mnh , which implies pIn ´ pqφppq “ 0. We replace p by In ´ p, to get
pφpIn ´ pq “ 0. Therefore, we have
0 “ pIn ´ pqφppq ´ pφpIn ´ pq “ φppq ´ pφpIn q.
Taking the adjoint, we also have
0 “ pφppq ´ pφpIn qq˚ “ φppq ´ φpIn qp,
and conclude that φpIn q commutes with every projection. Since every matrix is the
linear combination of projections, we see that φpIn q “ λIn by Proposition 1.2.1.
Taking an orthonormal basis t|ξi yu, we have
ÿ ÿ ÿ
λξi |ξi yxξi | “ φp|ξi yxξi |q “ φpIn q “ λIn “ λ|ξi yxξi |.
i i i
Putting |ξi y at the right-side of the above identity, we have λξi “ λ, and we conclude
that φ “ λ ¨ idn . This shows that idn is an extremal positive map. ˝
19
Lemma 1.2.3 Suppose that m ď n and s is an m ˆ n matrix of the form
` ˘
s “ Im 0 P Mmˆn
with the m ˆ m identity matrix Im in the left corner. If φ : M` Ñ Mm is extremal

positive then Ads ˝φ : M` Ñ Mn is also extremal.
Proof. Note that σ “ Ads is of the form:

ˆ ˙
a 0
σ : a ÞÑ P Mn .
0 0
We define τ : Mn Ñ Mm by
ˆ ˙
a b
τ: ÞÑ a P Mm .
c d
Then we have τ ˝ σ “ idm . Suppose that σ ˝ φ “ ψ1 ` ψ2 with ψ1 , ψ2 P P1 rM` , Mn s.

Then we have
φ “ τ ˝ σ ˝ φ “ τ ˝ ψ1 ` τ ˝ ψ2 .
Since φ is extremal, there exists λ ě 0 such that τ pψ1 paqq “ λφpaq for each a P M` .
Because ˆ ˙
φpaq 0
ψ1 paq ď σpφpaqq “ , a P M`` ,
0 0
we have pσ ˝ τ qpψ1 paqq “ ψ1 paq for a P M` . Therefore, it follows that
ψ1 paq “ σpτ pψ1 paqqq “ σpλφpaqq “ λσpφpaqq,
for each a P Mm , and so σ ˝ φ “ Ads ˝φ is extremal. ˝
Theorem 1.2.4 For any m ˆ n matrix s, the linear map Ads : Mm Ñ Mn is

extremal positive.
Proof. By singular
ˆ ˙ value decomposition (1.8), we have s “ upqw with the m ˆ k
Ik ` ˘
matrix p “ and k ˆ n matrix q “ Ik 0 together with invertible matrices
0
u, w. Then we have
Ads “ Adw ˝ Adq ˝ Adp ˝ Adu .
By lemma 1.2.3 and Lemma 1.2.2, we see that AdpT “ AdpT ˝id is extremal, and so
Adp “ pAdpT q˚ is also extremal. Therefore, Adq ˝ Adp is extremal by Lemma 1.2.3
again. On the other hand, Adw and Adu are order isomorphisms since w and v are
invertible. Therefore, we conclude that Ads is extremal. ˝
20
It is very difficult in general to find out all the extreme rays of the convex
cone P1 rMm , Mn s, and it is an open problem even for low dimensional cases like
m “ n “ 3. But, it is not so difficult to find boundary points of P1 rMm , Mn s. We
`
see that if φpaq is singular for a nonzero a P Mm then φ must be a boundary point
of P1 . Indeed, the following is immediate from the definition of interior points.
Proposition 1.2.5 If φ P int P1 rMm , Mn s then φ sends every nonzero element in

`
Mm to an interior point of Mn` .
`
Proof. Take a nonzero a P Mm and b P Mn` , and consider the map ψ : Mm Ñ Mn
given by ψpxq “ Tr pxq
Tr paq
b. Then there exists t ą 1 such that p1 ´ tqψ ` tφ P P1 , and
so we have p1 ´ tqb ` tφpaq ě 0. Since b P Mn` was arbitrary, we see that φpaq is an
interior point of Mn` . ˝
We will see later that the converse of Proposition 1.2.5 also holds. See Proposi-
tion 2.3.3. Therefore, φ is on the boundary of P1 rMm , Mn s if and only if there is a
`
nonzero a P Mm such that φpaq P Mn is singular. It is easily seen that the positive
map Ads : Mm Ñ Mn satisfies this property for every s P Mmˆn . Especially, if
s P Mnˆn itself is singular then Ads paq is singular for every a P Mn .
References: [111], [118]
1.2.3 k-superpositive maps

We denote by SPk the convex hull of tAds : rank s ď ku;
SPk :“ conv tAds : rank s ď ku Ă P1 ,
where conv S denotes the convex hull of S. We also use the notation SPk rMm , Mn s
when we need to specify the domain and range. If k ă ` ď m ^ n and rank s “ `
then Ads R SPk since it is extremal, where m ^ n denotes the the minimum of m
and n. Therefore, we have the following chain of strict inclusions
SP1 Ř SP2 Ř ¨ ¨ ¨ Ř SPk Ř ¨ ¨ ¨ Ř SPm^n
of convex cones of positive maps. A positive map in SPk is called k-superpositive.

A 1-superpositive map may be called just superpositive.
Proposition 1.2.6 For each k “ 1, 2, . . . , m ^ n, the convex cone SPk is closed in

HpMm , Mn q.
Proof. We consider the linear functional τ on the space HpMm , Mn q defined by
τ pφq “ Tr φpIm q, φ P HpMm , Mn q.
21
ř ř
Then we have τ pAds q “ ij |sij |2 for s “ ij sij |iyxj|. Next, we consider the set
ř
Ck :“ ts P Mmˆn : rank s ď k, ij |sij |2 “ 1u,
which is compact. Then the image of Ck under the continuous map s ÞÑ Ads is a
compact subset of HpMm , Mn q whose convex hull is tφ P SPk : τ pφq “ 1u, which is
also compact. Therefore, we conclude that the convex cone SPk is closed. ˝
It is clear that the transpose map T : x ÞÑ xT is a positive map between Mn .
The transpose map is extremal in P1 rMn , Mn s, since T “ T ˝ id and T is an order
isomorphism. When n ě 2, it is easy to see that T is not a member of the convex
cone SPn rMn , Mn s, whose extreme rays consist of Ads . If we assume that T P SPn
ř
then we have T “ Ads for a matrix s “ i,j sij |iyxj|. We have
ÿ ÿ
Ads p|iyxj|q “ s̄qp sk` |pyxq|iyxj|kyx`| “ s̄ip sj` |pyx`|.
k,`,p,q p,`
Since Tp|iyxj|q “ |jyxi|, we have s̄ij sji “ 1 for every i, j, especially s has no zero
entries. On the other hand, we have s̄ip sj` “ 0 whenever pp, `q ‰ pj, iq, which is
absurd.
Since the composition of positive maps is positive, we see that
SPk :“ tT ˝ φ : φ P SPk u
is a convex cone in P1 , and we have another chain of strict inclusions
SP1 Ř SP2 Ř ¨ ¨ ¨ Ř SPk Ř ¨ ¨ ¨ Ř SPm^n (1.13)
of convex cones in P1 . By the identity
T ˝ Ads paq “ ps˚ asqT “ sT aT s̄ “ Ads̄ paT q “ Ads̄ ˝Tpaq,
we see that φ P SPk if and only if T ˝ φ P SPk if and only if φ ˝ T P SPk .

In the identity Ad|ξyxη| paq “ |ηyxξ|a|ξyxη|, we see that xξ|a|ξy is a scalar and
Tp|ηyxη|q “ |η̄yxη̄|. Therefore, we have T ˝ Ad|ξyxη| paq “ Ad|ξyxη̄| paq, and so get the
identity
T ˝ Ad|ξyxη| “ Ad|ξyxη̄| . (1.14)
Hence, it follows that two convex cones SP1 and SP1 coincide, and we have the
following inclusion:
SP1 Ă SPm^n X SPm^n . (1.15)
Proposition 1.2.7 Suppose that s P Mmˆn is nonzero. Then Ads P SPm^n if and
only if s is of rank one.
22
Proof. It remains to prove the ‘only if’ part. Suppose that Ads belongs to the
convex cone SPm^n . Since Ads generates an extreme ray of P1 , it also generates an
extreme ray of the smaller cone SPm^n , and so Ads “ Adt ˝T for an m ˆ n matrix
t, which implies that the identity
¯
s˚ |ξyxη|s “ t˚ |η̄yxξ|t
holds for every |ξy and |ηy in Cm . We fix a vector |ηy so that t˚ |η̄y ‰ 0 and xη|s ‰ 0.
For arbitrary nonzero |ξy, the left side is a rank one matrix with the range vector
s˚ |ξy. On the other hand, the right side is a rank one matrix with the range vector
t˚ |η̄y, and so s˚ |ξy is a scalar multiple of t˚ |η̄y for arbitrary |ξy P Cm . Therefore s˚
is of rank one. ˝
Proposition 1.2.7 tells us that an extreme ray of SPm^n belongs to SPm^n then it
actually belongs to SP1 “ SP1 . Note that this does not imply that SPm^n X SPm^n
coincides with SP1 . The convex hull of two convex cones K1 and K2 is nothing but
K1 ` K2 “ tx1 ` x2 : xi P Ki u.
Now, we have concrete examples of positive maps in SPm^n ` SPm^n . Those maps
in this class, which are called decomposable, can be expressed by the combinations
of Ads and T. One question arises naturally: Does every positive map arise in this
way?
References: [3], [106], [108]
1.3 Positive maps between 2 ˆ 2 matrices

In this section, we show that all positive maps between 2 ˆ 2 matrices can be ex-
pressed with Ads and the transpose map. The main step is to realize the congruence
map AdU by a 2 ˆ 2 special unitary matrix U as a rotation in the three dimensional
real vector space.
1.3.1 Congruence maps and orthogonal transformations

In order to show that every positive map from M2 into itself is decomposable, we
first look at the 4-dimensional real vector space M2h whose elements may be written
by ˆ ˙
t ` z x ´ iy
%“ “ tI2 ` xσx ` yσy ` zσz ,
x ` iy t ´ z
with real numbers t, x, y, z and Pauli matrices
ˆ ˙ ˆ ˙ ˆ ˙
0 1 0 í 1 0
σx “ , σy “ , σz “ . (1.16)
1 0 i 0 0 ´1
23
The following four matrices
! )
?1 I2 , ?1 σx , ?1 σy , ?1 σz
2 2 2 2
make an orthonormal basis of M2h with respect to the inner product given by
ÿ
pa, bq “ Tr pab˚ q “ aij b̄ij , (1.17)
for a, b P M2h . We see that % P M2` if and only if t ě |z| and t2 ´ z 2 ě x2 ` y 2 if and
only if t ě 0 and x2 ` y 2 ` z 2 ď t2 holds. Therefore, we see that % P D2 if and only if
1
x2 ` y 2 ` z 2 ď t 2 , t“ (1.18)
2
holds, and so D2 is the three dimensional ball, called the Bloch ball.
We recall that the matrix 21 I2 is located at the center of the convex cone M2` ,
and consider the distance from % P D2 satisfying (1.18) to
1
%˚ :“ I2 P D2 .
2
The distance }x ´ y}HS arising from the inner product (1.17) is called the Hilbert–
Schmidt distance. We have
›ˆ ˙›2
2
› z x ´ iy ›
› “ 2px2 ` y 2 ` z 2 q.
}% ´ %˚ }HS “ ››
x ` iy ´z ›HS
Therefore, we see that % P D2 if and only if t, x, y and z satisfy the relation (1.18)
if and only if the inequality
1
}% ´ %˚ }HS ď ?
2
holds, as it was expected. We also see that % is on the boundary, that is, satisfies
the equality }%˚ ´ %}HS “ ?12 if and only if % “ |ξyxξ| is of rank one if and only if %
generates an extreme ray of the convex cone M2` . The 2-dimensional sphere which
is the boundary of D2 is called the Bloch sphere. We also note that each boundary
point of D2 corresponds to the range vector |ξy P C2 up to scalar multiplications,
which corresponds to a point of the complex projective space CP1 .
We say that φ : Mm Ñ Mn is trace preserving if Tr pφpxqq “ Tr pxq for every
x P Mm . By the identity Tr a “ xI, ay, we see that φ is trace preserving if and only
if φ˚ is unital. A trace preserving unital positive map between M2 has an interesting
geometric interpretation. We begin with a Hermiticity preserving map φ on M2h ,
which is the orthogonal sum of RI2 and the subspace
Z :“ span tσx , σy , σz u
24
of M2h . Suppose that φ is unital and trace preserving. Then both RI2 and Z are
invariant under φ, and φ|Z is a linear transform in the three dimensional space Z
over the real field.
We examine what happens for AdU with a 2 ˆ 2 special unitary
ˆ ˙
α ´β̄
U“ , |α|2 ` |β|2 “ 1.
β ᾱ
It is clear that the congruence map AdU is unital and trace preserving. We have
ˆ ˙ˆ ˙ˆ ˙ ˆ ˙
ᾱ β̄ 0 1 α ´β̄ 2Re pαβ̄q ᾱ2 ´ β̄ 2
AdU pσx q “ “
´β α 1 0 β ᾱ α2 ´ β 2 ´2Re pαβ̄q
We choose real variables a, b, c, d with α “ a ` ib, β “ c ` id and a2 ` b2 ` c2 ` d2 “ 1,

then we have
α2 ´ β 2 “ p1 ´ 2b2 ´ 2c2 q ` ip2ab ´ 2cdq,
2αβ̄ “ p2ac ` 2bdq ` ip´2ad ` 2bcq,
and so, it follows that
AdU pσx q “ p1 ´ 2b2 ´ 2c2 qσx ` p2ab ´ 2cdqσy ` p2ac ` 2bdqσz .
Similarly, we also have
AdU pσy q “ p´2ab ´ 2cdqσx ` p1 ´ 2b2 ´ 2d2 qσy ` p2ad ´ 2bcqσz ,

AdU pσz q “ p´2ac ` 2bdqσx ` p´2ad ´ 2bcqσy ` p1 ´ 2c2 ´ 2d2 qσz .
Therefore, the map φ determines a linear transform between three dimensional space
Z spanned by Pauli matrices, which corresponds to the following 3 ˆ 3 matrix
¨ ˛
1 ´ 2b2 ´ 2c2 ´2ab ´ 2cd ´2ac ` 2bd
˝ 2ab ´ 2cd 1 ´ 2b2 ´ 2d2 ´2ad ´ 2bc ‚.
2ac ` 2bd 2ad ´ 2bc 1 ´ 2c2 ´ 2d2
This is an orthogonal matrix with the determinant 1, and sends pd, ć, bqT to itself.
Therefore, it represents the rotation in R3 around the direction pd, ć, bqT P R3 . It
rotates by the angles 2θ satisfying cos θ “ a and sin2 θ “ b2 ` c2 ` d2 , and every
rotation in R3 arises in this way.
Note that the transpose map T : M2h Ñ M2h is also unital and trace preserving.
Since Tpσx q “ σx , Tpσy q “ ´σy and Tpσz q “ σz , we see that the trace map induces
the reflection ¨ ˛
1 0 0
˝0 ´1 0‚,
0 0 1
which is an orthogonal matrix with the determinant ´1. We recall that every
orthogonal matrix which is not a rotation can be expressed by the composition of a
25
rotation and the above reflection. We also recall that a linear map φ between Rn is
an isometry, that is, }φpxq} “ }x} for every x, if and only if φ preserves the inner
product if and only if it is represented by an orthogonal matrix. Therefore, we have
the following:
Proposition 1.3.1 Let φ : M2 Ñ M2 is unital and trace preserving. If φ|Z is an

isometry then φ “ AdU or φ “ AdU ˝T for a unitary matrix U .
References: [88], [8], [6], [7]
1.3.2 Decomposability of positive maps

Now, suppose that φ is positive as well as unital and trace preserving. Then φ sends
D2 into itself, that is, we have
% P D2 ùñ φp%q P D2 .
For a given % P Z, we have % ` %˚ P D2 if and only if }%}HS ď ?1 . By the relation

2
φp% ` %˚ q “ φp%q ` %˚ , we have
1 1
% P Z, }%}HS ď ? ùñ }φp%q}HS ď ? .
2 2
When we identity pZ, } }HS q with the normed space pR3 , } }q, we see that φ|Z sends
the unit ball in R3 into itself. Such a linear map is called a contraction. Then, the
map φ|Z belongs to the convex set C3 of all real linear contractions between R3 , and
so, φ|Z is a convex combination of extreme points of C3 . It is easy to see that φ P C3
is an extreme point of C3 if and only if φ is an isometry.
Proposition 1.3.2 Let Cn be the convex set of all contractions between Rn . A linear
map φ P Cn is an extreme point of Cn if and only if φ is an isometry.
Proof. By the singular value decomposition, φ is the composition of isometries and

a diagonal map δ P Cn . If φ is extreme in Cn then δ is also extreme in Cn . It is
evident that a diagonal map which is extreme in Cn is an isometry. For the converse,
we suppose that φ is an isometry of Rn and φ “ p1 ´ tqφ0 ` tφ1 with φ0 , φ1 P Cn .
When }x} “ 1, we have p1 ´ tqφ0 pxq ` tφ1 pxq “ φpxq is an extreme point of the unit
ball. Therefore, it follows that φ0 pxq “ φ1 pxq “ φpxq for every x with }x} “ 1, and
we conclude that φ “ φ0 “ φ1 . ˝
By Proposition 1.3.1 and Proposition 1.3.2, we conclude that every unital trace
preserving positive map between 2 ˆ 2 matrices is decomposable. For general cases,
we begin with a map in the interior of P1 . We recall that if φ P int P1 rMm , Mn s then
26
`
φ sends every nonzero element in Mm to an interior point of Mn` by Proposition
1.2.5. In short, φpaq is nonsingular whenever a is nonzero. It is easy to see that φ
is an interior point of P1 if and only if φ˚ is an interior point of P1 .
Proposition 1.3.3 Let φ be an interior point of the convex cone P1 rMn , Mn s. Then
there exist nonsingular a, b P Mn` such that the map φ̃ :“ Ada ˝φ ˝ Adb is positive,
unital and trace preserving.
Proof. We note that φ̃ is unital if and only if aφpb2 qa “ I if and only if the identity
φpb2 q´1 “ a2 (1.19)
holds, and φ̃ is trace preserving if and only if AdbT ˝φ˚ ˝ AdaT is unital if and only if
φ˚ pa2T q´1 “ b2T , (1.20)
where a2T “ pa2 qT “ paT q2 has the obvious meaning. We define the map f from Dn
into itself by
1
f psq “ T
φrφ˚ psT q´T s´1 , s P Dn ,
Tr φrφ ps q s
˚ ´T ´1
which is a continuous self map on the compact convex set Dn . By Brower’s fixed
point theorem, there exists s0 P Dn such that f ps0 q “ s0 . We take a, b P Mn` so that
a2 “ s 0 , b2 “ φ˚ pa2T q´T .
The second identity tells us that a and b satisfy (1.20), and so φ̃ is trace preserving.
Furthermore, we have
1 1
a2 “ s0 “ φrφ˚ pa2T q´T s´1 “ φpb2 q´1
t t
with t “ Tr φrφ˚ pa2T q´T s´1 , which implies In “ taφpb2 qa “ tφ̃pIn q. Because φ̃ is
trace preserving, we have t “ 1. Therefore, we conclude that φ̃ is unital by (1.19).
˝
Therefore, we conclude that every interior point φ of the convex cone P1 rM2 , M2 s
is decomposable. By the exactly same argument as in Proposition 1.2.6, it is easily
seen that the convex cone SPm^n ` SPm^n is also closed. Therefore, we have the
following:
Theorem 1.3.4 Every positive map in P1 rM2 , M2 s is decomposable.
From this, we see that every extremal positive map between 2 ˆ 2 matrices is of
the form Ads or Ads ˝T for a 2 ˆ 2 matrix s. It is known [129] that every positive
map in P1 rM2 , M3 s and P1 rM3 , M2 s is also decomposable. We will see in Section
1.5 that this is not the case for P1 rM3 , M3 s. Concrete examples of indecomposable
positive linear maps will be given in Section 1.6.
References: [111], [129], [6], [7]
27
1.4 Choi matrices and separable states
For a given linear map φ from Mm into Mn , we associate the Choi matrix Cφ in the
tensor product Mm b Mn which plays a central role throughout the topics. With
this isomorphism φ ÞÑ Cφ , 1-superpositive maps correspond to separable states, and
composition with the transpose map corresponds to taking the partial transpose.
1.4.1 Choi matrices

We note that a linear map φ : Mm Ñ Mn is completely determined by φp|iyxj|q for
i, j “ 1, 2, . . . , m, because t|iyxj| : i, j “ 1, . . . , mu is a basis of Mm . We define the
Choi matrix Cφ of a linear map φ P LpMm , Mn q by
m
ÿ
Cφ “ |iyxj| b φp|iyxj|q P Mm b Mn .
i,j“1
Then it is clear that
φ ÞÑ Cφ : LpMm , Mn q Ñ Mm b Mn
is a linear isomorphism, which is called the Jamiolkowski–Choi isomorphism. By

ř
the relation i,j xa, |iyxj|yφp|iyxj|q “ φpaq, we have
ÿ
xa b b, Cφ y “ xa, |iyxj|yxb, φp|iyxj|qy “ xb, φpaqy (1.21)
i,j
for a P Mm and b P Mn . The identity (1.21) may be used as the definition of the
Choi matrix, because it determines the entries of the Choi matrix when a “ |iyxj|
and b “ |kyx`|.
If φ maps from Mm into Mn then φ˚ interchanges the domain and the range,
and so Cφ˚ belongs to Mn b Mm . For given a P Mm and b P Mn , we have
xb b a, Cφ˚ y “ xa, φ˚ pbqy “ xφpaq, by “ xa b b, Cφ y. (1.22)
Therefore, Cφ˚ P Mn b Mm is the flip of Cφ P Mm b Mn , where the flip operation is

defined by x b y ÞÑ y b x.
Now, we look for properties of the Choi matrices which correspond to Hermiticity
preserving maps and positive maps.
Proposition 1.4.1 A linear map φ : Mm Ñ Mn is Hermiticity preserving if and

only if Cφ is Hermitian.
28
Proof. By the definition of Choi matrix, we see that Cφ is Hermitian if and only if
the following
φp|iyxj|q˚ “ φp|jyxi|q, i, j “ 1, 2, . . . , m (1.23)
holds. Recall that φ is Hermiticity preserving if and only if φpaq˚ “ φpa˚ q for every
a P Mm , which implies (1.23). Conversely, the relation (1.23) implies
φpaq˚ “ φ p aij |iyxj|q˚ “ āij φp|iyxj|q˚ “ āij φp|jyxi|q “ φpa˚ q,

ř ř ř
ř
for a “ aij |iyxj| P Mm . ˝
A matrix % P Mm b Mn is called block-positive when
xζ|%|ζy ě 0
holds for every product vector |ζy P Cm b Cn . Compare with the definition (1.1)
of positive matrices. The convex cone of all block-positive matrices is denoted by
BP 1 rMm b Mn s or just by BP 1 .
Proposition 1.4.2 A linear map φ : Mm Ñ Mn is positive if and only if Cφ is

block-positive in Mm b Mn .
Proof. We note that φ is positive if and only if φp|ξyxξ|q is positive in Mn for every
|ξy P Cm if and only if
¯
x|ηyxη|, φp|ξyxξ|qy “ x|ξyxξ| b |ηyxη|, Cφ y “ x|ξy|ηyxξ|xη|, Cφ y “ xξ|xη̄|C ¯
φ |ξy|η̄y
is nonnegative for every |ξy P Cm and |ηy P Cn , where we used (1.2). ˝

ř
For an m ˆ n matrix s “ i,j sij |iyxj|, we have
ř ´ř ¯ ´ř ¯
CAds “ p,q |pyxq| b k,` s̄`k |kyx`| |pyxq| i,j sij |iyxj|
´ř ř ¯ ´ ř ř ¯
“ s̄
k,` `k p x`|py|py|ky s
i,j ij q xq|iyxq|xj|
´ř ¯ ´ř ¯
“ k,` s̄`k |`y|ky i,j sij xi|xj| .
Therefore, we see that

CAds “ |s̃yxs̃|, (1.24)
under the correspondence between s P Mmˆn and the vector |s̃y P Cm b Cn given
by (1.5).
ˆ It is worthwhile to write down the Choi matrix of the map Ads when
˙
a b
s“ . We have xs̃| “ pa, b, c, dq, and
c d
¨ 2 ˛
|a| āb āc ād
˚ b̄a |b|2 b̄c b̄d ‹
CAds “ |s̃yxs̃| “ ˚ ˝ c̄a c̄b |c|2 c̄d ‚.
‹
¯
da ¯
db ¯ |d|2
dc
29
T | “ pa, c, b, dq, and
We also have xsĂ
¨ ˛
|a|2 āc āb ād
˚ c̄a |c|2 c̄b c̄d ‹
CpAds q˚ “ CAdsT ˝ b̄a b̄c |b|2 b̄d ‚,
“˚ ‹
¯
da ¯
dc ¯ |d|2
db
¯ and so we
which is the flip of CAds . On the other hand, we have xsr˚ | “ pā, c̄, b̄, dq,
see that CAds˚ is the conjugation of the flip of CAds .
References: : [32], [67], [21],
1.4.2 Schmidt numbers and entanglement

The identity (1.24) tells us that the Choi matrix of the map Ads is of rank one
mn ˆ mn positive matrix onto the vector in Cm b Cn whose Schmidt rank coincides
with the rank of the matrix s. We define
Sk rMm b Mn s : “ tCφ P Mm b Mn : φ P SPk rMm , Mn su

“ conv t|ζyxζ| P Mm b Mn : |ζy P Cm b Cn , SR |ζy ď ku,
for k “ 1, 2, . . . , m^n. Then we see that % P Sk if and only if % is an (unnormalized)

state of the form
ÿ
%“ |ζi yxζi |
i
with SR |ζi y ď k. Furthermore, a linear map φ belongs to SPm^n if and only if

Cφ is positive. Because a positive matrix generates an extreme ray of the cone
pMm b Mn q` if and only if it is of rank one, we see that Ads generates an extreme
ray of the cone SPm^n . The point of Theorem 1.2.4 is that Ads generates an extreme
ray of the much larger convex cone P1 . We may summarize in the following diagram:
LpMm , Mn q : SP1 Ř SPk Ř SPm^n Ř P1
Ó JC Ó Ó Ó Ó (1.25)
Mm b Mn : S1 Ř Sk Ř pMm b Mn q` Ř BP 1
A state % P Sk zSk´1 is called to have Schmidt number k. A state of Schmidt

number one is also called separable, and a state which is not separable is called
entangled. Recall that a vector in Cm b Cn with Schmidt rank one is called a
product vector. Then, a state in pMm b Mn q` is separable if and only if it is the
convex combination of rank one projections onto product vectors. In other words,
30
a state % is separable if and only if there exists a finite family of product vectors
t|ξi y|ηi yu satisfying ÿ
%“ |ξi y|ηi yxξi |xηi |. (1.26)
i
If we restrict ourselves to the case of normalized states with trace one, then we have
ÿ
%“ pi |ξi y|ηi yxξi |xηi |
i
with unit vectors |ξi y, |ηi y and probability distribution tpi u. A state in the tensor
product Mm b Mn is called a bi-partite state, or more precisely an m b n state to
emphasize the size of matrices.
`
We denote by Mm b Mn` the convex cone generated by a b b with a P Mm `
and
`
b P Mn ;
`
Mm b Mn` :“ conv ta b b P Mm b Mn : a P Mm `
, b P Mn` u.
`
Since |ξy|ηyxξ|xη| “ |ξyxξ| b |ηyxη| belongs to Mm `
b Mn` , we have S1 Ă Mm b Mn` .
ř `
ř
Conversely, if a “ p |ξp yxξp | P Mm and b “ q |ηq yxηq | P Mn` then we have
ÿ
abb“ |ξp y|ηq yxξp |xηq |
p,q
is separable. Therefore, we have

`
S1 “ Mm b Mn` .
In short, a bi-partite state is separable if and only if it is the sum of product states
`
of the forms a b b with positive a P Mm and b P Mn` . So, entanglement consists of
the difference
pMm b Mn q` zMm `
b Mn` .
In the case of the space CpXq of all continuous functions on a compact Haus-
dorff space X, the algebraic tensor product CpXq b CpY q consists of finite sums of
functions of separable variables, that is, functions of the form f pxqgpyq. By Stone–
Weierstrass theorem, we see that CpXq b CpY q is dense in CpX ˆ Y q. We write
P :“ CpXq` b CpY q` . Then P ´ P :“ tf ´ g : f, g P P u is a unital subalgebra of
CpX ˆ Y, Rq which separates points in X ˆ Y . Applying the real version of Stone–
Weierstrass theorem, we see that P ´ P is dense in CpX ˆ Y, Rq. Therefore, we
conclude that P “ CpXq` b CpY q` is dense in CpX ˆ Y q` , and there exists no
entanglement in function spaces.
If a state % P Mm b Mn is of rank one, then it is easy to find its Schmidt number
by the Schmidt rank of its range vector. For example,
¨ ˛
1 0 1 0
˚0 0 0 0‹
˚
˝1 0 1 0‚
‹ (1.27)
0 0 0 0
31
is of Schmidt number one, or equivalently separable, because the range vector
p|0y ` |1yq|0y “ p1, 0, 1, 0qT P C2 b C2
corresponds to the 2 ˆ 2 matrix ˆ ˙

1 0
1 0
which is of rank one. On the other hand,
¨ ˛
1 0 0 1
˚0 0 0 0‹
˚
˝0 0
‹ (1.28)
0 0‚
1 0 0 1
T
is of Schmidt number
ˆ ˙two, and so entangled, because the range vector p1, 0, 0, 1q
1 0
corresponds to which is of rank two.
0 1
Note that the matrix in (1.28) is the Choi matrix of the identity map on M2 .
The Choi matrix
n
ÿ
Cid “ |iyxj| b |iyxj| “ |ωyxω|
i,j“1
of the identity map on Mn with

n
ÿ
|ωy “ |iy|iy P Cn b Cn
i“1
plays an important role. It is a bi-partite state with the maximal Schmidt number.
ř
In general, a pure state % “ |ζyζ| is called maximally entangled if |ζy “ |ξi y b |ηi y
with orthonomal bases t|ξi yu and t|ηi yu. On the other hand, the identity matrix in
Mm b Mn is the Choi matrix of the trace map
Tr : X ÞÑ Tr pXqIn , X P Mm .
The trace map may be considered to be located at the center of the convex cone P1
of all positive maps, by the location of the identity matrix in Mm b Mn . In fact,
we will see in Section 2.3 that the converse of Proposition 1.2.5 holds, and so the
trace map is an interior point of the convex cone P1 rMm , Mn s. On the other hand,
the identity map between Mn is located at a corner of the convex cone P1 rMn , Mn s,
since it generates an extreme ray of P1 .
For given complex numbers α, β with modulus one, we take a product vector
|ζy “ p1, αβ̄qT b p1, βqT “ p1, β, αβ̄, αqT P C2 b C2 ,
32
to get the pure separable state
¨ ˛
1 β̄ ᾱβ ᾱ
˚β 1 ᾱβ 2 ᾱβ ‹
‹.
|ζyxζ| “ ˚
˝αβ̄ αβ̄ 2 1 β̄ ‚
α αβ̄ β 1
Taking β “ 1, i, ´1, í, and averaging the corresponding states, we get the following
separable state ¨ ˛
1 ¨ ¨ ᾱ
˚¨ 1 ¨ ¨‹
˝ ¨ ¨ 1 ¨ ‚ P M2 b M2 . (1.29)
˚ ‹
α ¨ ¨ 1
For double indices pi, kq and pj, `q from t1, . . . , muˆt1, . . . , nu with pi, kq ‰ pj, `q,
we consider the bi-partite state %pi,kq,pj,`q pαq in Mm b Mn defined by
%pi,kq,pj,`q pαq “ Imn ` ᾱ|ikyxj`| ` α|j`yxik|,
for a complex number α with modulus one. It is clear that %pi,kq,pj,`q pαq is positive.
If i “ j then the entries α and ᾱ appear in the i-the diagonal block, and so it is
separable. This is also the case when k “ `. In order to deal with the case i ‰ j
and k ‰ `, we consider the product vector
|ζy “ p|iy ` αβ̄|jyq b p|ky ` β|`yq “ |iky ` β|i`y ` αβ̄|jky ` α|j`y. (1.30)
By the exactly same way to get the separable state (1.29), we see that %pi,kq,pj,`q pαq is
separable whenever |α| “ 1, by adding diagonal entries. Now, it is easily seen that
%pi,kq,pj,`q p`1q ´ Imn , %pi,kq,pj,`q pìq ´ Imn , pImn ` |ikyxik|q ´ Imn
make a linearly independent family with the cardinality pmnq2 . Therefore, we con-
clude that S1 rMm b Mn s has nonzero volume in pMm b Mn q` .
References: [128], [34], [124], [40], [106]
1.4.3 Partial transposes

The Choi matrix of the transpose map is given by
n
ÿ n
ÿ
CT “ |iyxj| b T p|iyxj|q “ |iyxj| b |jyxi|.
i,j“1 i,j“1
When n “ 2, its matrix form is given by

¨ ˛
1 0 0 0
˚0 0 1 0‹
CT “ ˚˝0
‹,
1 0 0‚
0 0 0 1
33
which is not positive.
We see that CT P Mn pMn q is the block-wise transpose of Cid as an n ˆ n matrix
with entries from Mn . Every % P Mm b Mn is uniquely expressed by
m
ÿ
%“ |iyxj| b %ij
i,j“1
with %ij P Mn . We define the partial transpose %Γ by

m
ÿ m
ÿ
%Γ “ |jyxi| b %ij “ |iyxj| b %ji .
i,j“1 i,j“1
We note that CT is the partial transpose of Cid . In general, we have

ÿ
Cφ˝T “ |iyxj| b φp|jyxi|q “ pCφ qΓ . (1.31)
i,j
For a given linear map φ : Mm Ñ Mn , we define the linear map φ̄ : Mm Ñ Mn

by
φ̄pxq “ φpx̄q, x P Mm ,
where x̄ is the matrix whose entries are the conjugates of the corresponding entries
of x. Then, we have
Cφ̄ “ Cφ .
If φ is Hermiticity preserving then we have
φpxT qT “ φpxT q˚ “ φpxT˚ q “ φpx̄q “ φ̄pxq.
In short, we have
T ˝ φ ˝ T “ φ̄,
for a Hermiticity preserving map φ. Therefore, we also have
CT˝φ “ Cφ̄˝T “ CΓφ̄ “ CΓφ , (1.32)
for a Hermiticity preserving map φ. We also note the following identity
pa b bqΓ “ aT b b
holds. If a and b are Hermitian, then we have
pa b bqΓ “ aT b b “ a˚ b pb˚ qT “ a b bT .
`
Because S1 “ Mm b Mn` and the transpose of a positive matrix is again positive,
we have the following:
34
Theorem 1.4.3 The partial transpose of a separable state is positive.
A positive matrix is called of positive partial transpose or just PPT if its partial
transpose is positive. Theorem 1.4.3, which is called the PPT criterion for sepa-
rability says that every separable state is of PPT. It is also can be seen from the
inclusion relation (1.15). On the other hand, Proposition 1.2.7 tells us that if a rank
one state in Mm b Mn is of PPT then it should be separable. More generally, it is
known [63] that if a PPT state in Mm b Mn has rank at most maxtm, nu then it is
separable.
Note that the matrix in (1.27) is of PPT, but the matrix Cid in (1.28) is not of
PPT, from which we may infer that Cid is not separable. In fact, we know that the
positive map id “ AdI is not 1-superpositive, because rank I ą 1. The convex cone
of all m b n PPT state will be denoted by PPT rMm b Mn s or just by PPT . Then
we have S1 Ă PPT Ă Sm^n “ pMm b Mn q` in the diagram (1.25). The converse of
Theorem 1.4.3 does not hold in general.
References: [24], [94], [63]
1.4.4 Entanglement with positive partial transpose

For given positive matrices a and b, it is easy to see that kerpa ` bq Ă ker a by the
identity xξ|pa ` bq|ξy “ xξ|a|ξy ` xξ|b|ξy. Therefore, we have Im a Ă Im pa ` bq for
a, b P Mn` , and so we also have
Im a ` Im b “ Im pa ` bq, a, b P Mn` .
This simple fact is useful to show that a given state is not separable, because the
ř
expression % “ |ζi yxζi | gives rise to the restriction |ζi y P Im %. In fact, we note
that the following
Im % “ span t|ζi yu
ř
holds for a positive matrix % “ |ζi yxζi |.
It is easily seen that the following matrix
¨ ˛
1 ¨ ¨ ¨ 1 ¨ ¨ ¨ 1
˚ ¨ 1 ¨ 1 ¨ ¨ ¨ ¨ ¨ ‹
˚ 2 ‹
˚ ¨ ¨ 2 ¨ ¨ ¨ 1 ¨ ¨ ‹
˚ ‹
˚ ‹
˚ ‹
˚ ¨ 1 ¨ 2 ¨ ¨ ¨ ¨ ¨ ‹
˚ ‹
%“˚ ˚ 1 ¨ ¨ ¨ 1 ¨ ¨ ¨ 1 ‹
‹ (1.33)
1
˚ ¨ ¨ ¨
˚ ¨ ¨ 2
¨ 1 ¨ ‹
‹
˚ ‹
˚ 1
‹
˚ ¨ ¨ 1 ¨ ¨ ¨ ¨ ¨ ‹
˚ 2 ‹
˝ ¨ ¨ ¨ ¨ ¨ 1 ¨ 2 ¨ ‚
1 ¨ ¨ ¨ 1 ¨ ¨ ¨ 1
35
is a PPT state. In fact, we have %Γ “ %. We see that the range is the 4-dimensional
space which is spanned by
|0y|0y ` |1y|1y ` |2y|2y,

1 ? 1 ? 1 ? (1.34)
? |0y|1y ` 2|1y|0y, ? |1y|2y ` 2|2y|1y, ? |2y|0y ` 2|0y|2y,
2 2 2
with the corresponding 3 ˆ 3 matrices
? ˛
0 ?12
¨ ˛ ¨ ˛ ¨ ˛ ¨
1 0 0 0 0 0 0 0 0 2
?
˝0 1 0‚, ˝ 2 0 0 ,
‚ ˝ 0 0 ?1 ‚, ˝0 0 0 ‚. (1.35)
? 2
0 0 1 0 0 0 0 2 0 ?1 0 0
2
If % is separable, then the four dimensional space spanned by these four matrices
must contain rank one matrices. In other words, there exist a, b, c, d P C such that
¨ ? ˛
a ?b d 2
˚b?2 a2 c
? ‚
‹
˝ ? 2
?d c 2 a
2
is of rank one. Considering various 2 ˆ 2 submatrices, one see easily that this is not
possible. Therefore, we conclude that % is an entangled state.
A nonzero subspace of Cm b Cn is called entangled when it has no nonzero
product vector. A state is entangled whenever its range space is entangled. We
have seen that the four dimensional subspace of C3 b C3 spanned by four vectors in
(1.34) is entangled.
Even though the range space of a state is not entangled, a PPT state % may be
shown to be entangled by considering the range space of the partial transpose %Γ
as well as the range space of % itself. Suppose that % is a separable state with the
expression in (1.26) whose range space is spanned by the product vectors |ξi y|ηi y.
By the relation
|ξy|ηyxξ|xη|Γ “ p|ξyxξ| b |ηyxη|qΓ

“ |ξyxξ|T b |ηyxη|
¯ ξ|
“ |ξyx ¯ b |ηyxη| “ |ξy|ηyx
¯ ¯
ξ|xη|,
we have the following identity

ÿ
%Γ “ |ξ¯i y|ηi yxξ¯i |xηi |,
i
ř
for % “ i |ξi y|ηi yxξi |xηi |. If the partial transpose %Γ is positive then we see that
the range space %Γ is spanned by product vectors |ξ¯i y|ηi y. Therefore, we have the
following necessary condition for separability.
36
Theorem 1.4.4 Suppose that % is a PPT state. If % is separable then there exists
a finite family t|ξi y|ηi yu of product vectors satisfying
Im % “ span t|ξi y|ηi yu, Im %Γ “ span t|ξ¯i y|ηi yu.
¯
The vector |ξy|ηy is called the partial conjugate of the product vector |ξy|ηy.
Theorem 1.4.4 is called the range criterion for separability which is especially useful
for PPT states whose range is not full. For an example, we consider the following
PPT state %p defined by
¨ ˛
p ¨ ¨ ¨ p ¨ ¨ ¨ p
˚ ¨ p2 ¨ ¨ ¨ ¨ ¨ ¨ ¨ ‹
˚ ‹
˚ ¨ ¨ 1 ¨ ¨ ¨ ¨ ¨ ¨ ‹
˚ ‹
˚ ¨ ¨ ¨ 1 ¨ ¨ ¨ ¨ ¨ ‹
˚ ‹
%p “ ˚˚ p ¨ ¨ ¨ p ¨ ¨ ¨ p ‹,
‹ (1.36)
˚ ¨ ¨ ¨ ¨ ¨ p2 ¨ ¨ ¨ ‹
˚ ‹
˚ ¨ ¨ ¨ ¨ ¨ ¨ p2 ¨ ¨ ‹
˚ ‹
˝ ¨ ¨ ¨ ¨ ¨ ¨ ¨ 1 ¨ ‚
p ¨ ¨ ¨ p ¨ ¨ ¨ p
for a positive number p ą 0. The partial transpose is given by

¨ ˛
p ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨
˚ ¨ p2 ¨ p ¨ ¨ ¨ ¨ ¨ ‹
˚ ‹
˚ ¨ ¨ 1 ¨ ¨ ¨ p ¨ ¨ ‹
˚ ‹
˚ ¨ p ¨ 1 ¨ ¨ ¨ ¨ ¨ ‹
Γ
˚ ‹
˚ ¨ ¨ ¨ ¨ p ¨2 ¨
%p “ ˚ ¨ ¨ ‹‹.
˚ ¨ ¨ ¨ ¨ ¨ p ¨ p ¨ ‹
˚ ‹
˚ ¨ ¨ p ¨ ¨ ¨ p2 ¨ ¨ ‹
˚ ‹
˝ ¨ ¨ ¨ ¨ ¨ p ¨ 1 ¨ ‚
¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ p
We note that ranks of %p and %Γp are 7 and 6, respectively. We also note that the
kernels of %p and %Γp are spanned by
¨ ˛ ¨ ˛ ¨ ˛ ¨ ˛ ¨ ˛
1 ¨ ¨ ¨ ¨
˚ ¨ ‹ ˚ ¨ ‹ ˚´1‹ ˚ ¨ ‹ ˚ ¨ ‹
˚ ‹ ˚ ‹ ˚ ‹ ˚ ‹ ˚ ‹
˚ ¨ ‹ ˚ ¨ ‹ ˚ ¨ ‹ ˚ ¨ ‹ ˚p‹
˚ ‹ ˚ ‹ ˚ ‹ ˚ ‹ ˚ ‹
˚ ¨ ‹ ˚ ¨ ‹ ˚p‹ ˚ ¨ ‹ ˚ ¨ ‹
˚ ‹ ˚ ‹ ˚ ‹ ˚ ‹ ˚ ‹
˚´1‹ , ˚ 1 ‹ and ˚ ¨ ‹, ˚ ¨ ‹, ˚ ¨ ‹,
˚ ‹ ˚ ‹ ˚ ‹ ˚ ‹ ˚ ‹
˚ ¨ ‹ ˚ ¨ ‹ ˚ ¨ ‹ ˚´1‹ ˚ ¨ ‹
˚ ‹ ˚ ‹ ˚ ‹ ˚ ‹ ˚ ‹
˚ ¨ ‹ ˚ ¨ ‹ ˚ ¨ ‹ ˚ ¨ ‹ ˚´1‹
˚ ‹ ˚ ‹ ˚ ‹ ˚ ‹ ˚ ‹
˝ ¨ ‚ ˝ ¨ ‚ ˝ ¨ ‚ ˝p‚ ˝ ¨ ‚
¨ ´1 ¨ ¨ ¨
respectively.
37
If %p is separable then there exists a product vector |ξy|ηy P Im % such that
¯
|ξy|ηy P Im %Γ . We write |ξy “ px1 , x2 , x3 qT P C3 and |ηy “ py1 , y2 , y3 qT P C3 then we
have
|ξy|ηy “ px1 y1 , x1 y2 , x1 y3 , x2 y1 , x2 y2 , x2 y3 , x3 y1 , x3 y2 , x3 y3 qT P C3 b C3 .
Therefore, we have the following system of equations
x1 y 1 “ x2 y 2 ,
x2 y 2 “ x3 y 3 ,
x̄1 y2 “ px̄2 y1 , (1.37)
x̄2 y3 “ px̄3 y2 ,
x̄3 y1 “ px̄1 y3
We note that this system of equations has four unknowns up to scalar multiplications
and five equations, and so we may expect that there is no nontrivial solution like
|ξy|ηy ‰ 0. In fact, one can easily see that (1.37) has a nontrivial solution only when
p “ 1. Therefore, we conclude that %p is a PPT entangled states for p ‰ 1.
References: [129], [24], [113], [62], [83], [68], [69]
1.5 Duality and completely positive maps

We define in this section a bilinear pairing between Hermiticity preserving linear
maps through Choi matrices. The dual object of k-superpositivity is nothing but k-
positivity which had been studied since 1950’s. Through the discussion, we naturally
get the correspondence between complete positivity of linear maps and positivity of
their Choi matrices.
1.5.1 Dual cones of convex cones

For finite dimensional vector spaces X and Y , we recall that every linear map
φ P LpX, Y ˚ q corresponds to the linear functional Lφ P pX b Y q˚ given by
Lφ px b yq “ φpxqpyq, x P X, y P Y, (1.38)
where X ˚ denotes the dual space of X consisting of all linear functionals on X. In

case of matrix algebras, we employ the identification (1.3) between Mn and its dual
Mn˚ . If φ : Mm Ñ Mn˚ then the Choi matrix Cφ plays the role of a linear functional
on Mm b Mn by the relation (1.21):
xa b b, Cφ yMm bMn “ xb, φpaqyMn , a P Mm , b P Mn ,
38
which is nothing but the reformulation of (1.38) in terms of bilinear pairing between
matrices. Therefore, it is natural to define the bilinear pairing between Mm b Mn
and LpMm , Mn q by
xa b b, φy1 :“ xb, φpaqy, a P Mm , b P Mn , φ P LpMm , Mn q. (1.39)
Then we have x%, φy1 “ x%, Cφ y by (1.21) for % P Mm b Mn and φ P LpMm , Mn q.

Because every % P Mm b Mn can be written by % “ Cψ for ψ P LpMm , Mn q, it is also
natural to define the bilinear pairing
xψ, φy2 “ xCψ , Cφ y, φ, ψ P LpMm , Mn q (1.40)
between mapping spaces. If % “ Cψ , then we have
x%, φy1 “ x%, Cφ y “ xCψ , Cφ y “ xψ, φy2 ,
and so we do not distinguish the bilinear pairings x ¨ , ¨ y, x ¨ , ¨ y1 and x ¨ , ¨ y2 , and use

the same notation x ¨ , ¨ y for them.
The following identity
xφ, ψy “ xφ˚ , ψ ˚ y (1.41)
is immediate, because Cφ˚ P Mn b Mm is the flip of Cφ P Mm b Mn by (1.22). For
given linear maps φ P LpMA , MB q, ψ P LpMB , MC q and σ P LpMA , MC q, we also
have
ÿ ÿ
xψ ˝ φ, σy “ xψpφpeij qq, σpeij qy “ xφpeij q, ψ ˚ pσpeij qqy “ xφ, ψ ˚ ˝ σy,
i,j i,j
which also implies xφ, ψ ˚ ˝ σy “ xσ ˚ ˝ ψ, φ˚ y “ xψ, σ ˝ φ˚ y by (1.41). Therefore, we

have
xψ ˝ φ, σy “ xφ, ψ ˚ ˝ σy “ xψ, σ ˝ φ˚ y. (1.42)
Suppose that X and Y are finite-dimensional real spaces which are dual to
each other through a non-degenerate bilinear pairing x , y, that is, x P X satisfies
xx, yy “ 0 for every y P Y implies x “ 0, and same for y P Y . To be precise, we define
the linear map σX : X Ñ Y ˚ ; for a given x P X, we define σX pxq P Y ˚ as the linear
functional which sends y to xx, yy. Then non-degeneracy condition tells us that σX
is injective, and so we have dim X ď dim Y . By the similar map σY : Y Ñ X ˚ , we
have dim X “ dim Y , and both σX and σY are linear isomorphisms. Throughout
this note, we always assume that a bilinear pairing is non-degenerate.
For a given subset C of X, we define the dual cone C ˝ in Y by
C ˝ “ ty P Y : xx, yy ě 0 for each x P Cu,
39
and the dual cone D˝ Ă X similarly for a subset D of Y . If C is a closed convex
cone in a finite dimensional real vector space X and x0 R C, then it is known that
there exists a linear functional f on X such that f pxq ě 0 for every x P C and
f px0 q ă 0. From this, we have the following:
Proposition 1.5.1 For any subset C of X, the bidual cone C ˝˝ is the smallest
closed convex cone containing C.
Proof. We denote by X temporarily the smallest closed convex cone containing C,

then it is clear that X Ă C ˝˝ . Suppose that x0 R X . Then there exists y P Y such
that xx, yy ě 0 for every x P X but xx0 , yy ă 0. Since C Ă X , we have y P C ˝ . But
the relation xx0 , yy ă 0 tells us that x0 R C ˝˝ . Therefore, we conclude that C ˝˝ “ X .
˝
References: [99], [127], [34], [108], [107], [36]
1.5.2 k-positive maps and completely positive maps

By Proposition 1.5.1, we see that every closed convex cone C of X is the dual cone
of C ˝ Ă Y , that is C ˝˝ “ C. This means that C is determined by the intersection
of ‘half-spaces’ tx P X : xx, yy ě 0u given by y P C ˝ . We are going to determine the
dual cone of the convex cone SPk with respect to the pairing (1.40). This is same
as the dual cone of Sk with respect to the pairing (1.39).
For any |ζy “ ki“1 |xi y|yi y P Cm b Cn with SR |ζy ď k, we have
ř
k
ÿ
|ζyxζ| “ |xi yxxj | b |yi yxyj |,
i,j“1
and so, we also have

k
ÿ
x|ζyxζ|, Cφ y “ x|yi yxyj |, φp|xi yxxj |qy
i,j“1
C G
k
ÿ k
ÿ
“ |iyxj| b |yi yxyj |, |iyxj| b φp|xi yxxj |q .
i,j“1 i,j“1
Putting
k
ÿ k
ÿ
k m
|ξy “ |iy|xi y P C b C , |ηy “ |iy|yi y P Ck b Cn ,
i“1 i“1
we see that
x|ζyxζ|, Cφ y “ x|ηyxη|, pidk b φqp|ξyxξ|qy. (1.43)
Therefore, we have the following:
40
Theorem 1.5.2 A map φ : Mm Ñ Mn belongs to SP˝k if and only if the map idk b φ
from Mk b Mm into Mk b Mn is positive.
We say that φ is k-positive if the map idk b φ is positive, and completely positive
if it is k-positive for every k “ 1, 2, . . . . The convex cone of all k-positive maps
(respectively completely positive maps) from Mm into Mn is denoted by Pk rMm , Mn s
(respectively CPrMm , Mn s).
Theorem 1.5.3 For a linear map φ : Mm Ñ Mn , the following are equivalent:
(i) φ is completely positive,
(ii) idm^n b φ is positive,
(iii) φ belongs to SPm^n ,
(iv) Cφ is positive.
Proof. Since the convex cone pMm b Mn q` is self-dual, the equivalences between
(ii), (iii) and (iv) follows from Theorem 1.5.2. By the relation
pidk b Ads qpa b bq “ a b ps˚ bsq “ pIk b sq˚ pa b bqpIk b sq,
we see that every map in SPm^n is completely positive. ˝

ř ř
For a “ i,j aij |iyxj| and b “ i,j bij |iyxj| in Mn , the Schur product a ˝ b is given
ř
by a ˝ b “ i,j aij bij |iyxj|. For a positive a P Mn` , the map
Σa : x ÞÑ a ˝ x, x P Mn (1.44)
is a completely positive map whose Choi matrix is given by

ÿ ÿ
CΣa “ aij |iyxj| b |iyxj| “ aij |iy|iyxj|xj|,
i,j i,j
which is positive. Considering the partial transpose CΓΣa , we see that CΣa is separable
if and only if it is diagonal by Theorem 1.4.3.
By the statement (iii) of Theorem 1.5.3, every completely positive map φ is of
the form ÿ
φ“ Adsi
iPI
for a finite family tsi : i P Iu of mˆn matrices. This is called a Kraus decomposition
ř
of φ. This decomposition is obtained from a decomposition of Cφ “ |ξi yxξi | into
the sum of rank one positive matrices |ξi yxξi |, and so it is far from being unique.
41
If we take t|ξi yu to be linearly independent then we can also take a decomposition
ř
φ “ Adsi so that tsi u is linearly independent m ˆ n matrices.
A state % P Mm b Mn is called k-block-positive if xζ|%|ζy ě 0 for every vector
|ζy P Cm b Cn with SR |ζy ď k. Since x|ζyxζ|, Cφ y “ xζ̄|Cφ |ζ̄y, we also see that φ
is k-positive if and only if Cφ is k-block-positive. The convex cone of all k-block-
positive matrices in Mm b Mn is denoted by BP k rMm b Mn s. We complement the
diagram (1.25) as follows:
LpMm , Mn q : SP1 Ř SPk Ř CP Ř Pk Ř P1
Ó JC Ó Ó Ó Ó Ó (1.45)
Mm b Mn : S1 Ř Sk Ř pMm b Mn q` Ř BP k Ř BP 1
In this diagram, SPk and Pk are dual to each others with respect to the pairing
(1.40). On the other hand, the duality relation between Sk and Pk with respect to
the pair (1.39) may be stated as follows:
Theorem 1.5.4 For k “ 1, 2, . . . , m ^ n, we have the following:
(i) A linear map φ : Mm Ñ Mn is k-positive if and only if x%, φy ě 0 for every

state % with Schmidt number ď k,
(ii) A state % P Mm b Mn is of Schmidt number ď k if and only if x%, φy ě 0 for

every k-positive map φ : Mm Ñ Mn .
Compositions of k-positive maps by the transpose map give rise another impor-
tant class of positive maps. We denote by Tk the transpose map on Mk .
Proposition 1.5.5 For a linear map φ : Mm Ñ Mn , the following are equivalent:
(i) φ ˝ Tm is k-positive,
(ii) Tn ˝ φ is k-positive,
(iii) Tk b φ is positive,
(iv) CΓφ is k-block-positive.
Proof. We note that % is k-block-positive if and only if %̄ is k-block-positive, and

so the equivalence between (i), (ii) and (iv) follows from the identities (1.31) and
(1.32). On the other hand, the identity
idk b pTn ˝ φq “ pTk b Tn q ˝ pTk b φq
42
tells us that (ii) and (iii) are equivalent, because both Tk b Tm and its inverse are
positive maps. ˝
We say that φ is k-copositive when φ satisfies the conditions in Proposition 1.5.5,
and completely copositive when it is k-copositive for every k “ 1, 2, . . . . The convex
cone of all completely copositive maps will be denoted by CCP, or CCPrMm , Mn s
to specify the domain and the range. Then we have
CCPrMm , Mn s “ SPm^n ,
because CP “ SPm^n . We also see that φ is completely copositive if and only if

φ ˝ Tm is completely positive if and only if CΓφ is positive. We say that % P Mm b Mn
is copositive when %Γ is positive.
References: [72], [73], [21], [66], [34], [124], [108],
1.5.3 Faces for completely positive maps

Recall that every face of the convex cone pMm b Mn q` is determined by a subspace
of Cm b Cn ; for a subspace V , the set of all positive matrices whose range spaces are
contained in V is a face of pMm b Mn q` , and every face is in this form. Therefore,
every face of the convex cone CPrMm , Mn s is also determined by a subspace of the
space Mmˆn of all m ˆ n matrices.
Proposition 1.5.6 For a given subspace V of Mmˆn , the set
FV :“ conv tAdsi : si P V u (1.46)
is a face of CPrMm , Mn s, and every face arises in this way.
For a given p P Mn` , we denote by CPrMm , Mn ; ps the convex set of all completely
ř
positive maps φ with φpIq “ p. We note that φ “ Adsi belongs to CPrMm , Mn ; ps
ř ř
if and only if s˚i si “ p. For a given completely positive map φ “ iPI Adsi in
CPrMm , Mn ; ps, we take the subspace V “ span tsi : i P Iu. Then every map in FV
ř ř
is of the form ψ “ k Adtk with tk P V . Write tk “ i αik si . Then we have
˜ ¸ ˜ ¸
ÿ ÿ ÿ ÿ
ψpxq “ t˚k xtk “ ᾱik s˚i x αjk sj
k k i j
˜ ¸
ÿ ÿ
“ ᾱik αjk s˚i xsj .
i,j k
ř
We write βij “ k ᾱik αjk . Then we see that ψ P CPrMm , Mn ; ps if and only if
ř ˚
ř ˚
i,j βij si sj “ p, and φpIq “ ψpIq if and only if i,j pβij ´ δij qsi sj “ 0, where δij “ 1
43
for i “ j, and δij “ 0 for i ‰ j. If ts˚i sj : i, j P Iu is linearly independent then we
see that ψ “ φ, that is, the the intersection of FV and CPrMm , Mn ; ps consists of a
single point tφu. Therefore, we conclude that φ is an extreme point of the convex
set CPrMm , Mn ; ps whenever ts˚i sj u is linearly independent. The converse is also
true. We will use the following lemma whose simple proof is omitted.
Lemma 1.5.7 Suppose that t|vi y : i “ 1, . . . , nu is linearly independent in Cn . Then

t|vi yxvj | : i, j “ 1, . . . , nu is linearly independent in Mn .
ř
Theorem 1.5.8 Suppose that p P Mn` and φ “ iPI Adsi with linearly independent
ř
family tsi u of m ˆ n matrices with iPI s˚i si “ p. Then φ is an extreme point of the
convex set CPrMm , Mn ; ps if and only if ts˚i sj : i, j P Iu is linearly independent.
ř
Proof. Suppose that φ “ Adsi is an extreme point of the convex set CPrMm , Mn ; ps
ř
and i,j αij s˚i sj “ 0. We have to show that αij “ 0. We first consider the case when
α “ rαij s is a Hermitian matrix. We may assume that Í ď rαij s ď I by taking a
scalar multiplication. Define
ÿ ÿ
φ˘ pxq “ Adsi pxq ˘ αij s˚i xsj , x P Mm .
i i,j
ř
If we write δij ` αij “ βij , then we have φ` pxq “ i,j βij s˚i xsj . Writing rβij s “ γ ˚ γ,
we have ÿÿ ÿ
φ` pxq “ γ̄ki γkj s˚i xsj “ t˚k xtk ,
i,j k k
ř
with tk “ j γkj sj . Therefore, we see that φ` is completely positive. By the same
way, we also see that φ´ is also completely positive. The relation φ “ 21 pφ` `φ´ q tells
ř
us that φ “ φ` “ φ´ . Considering the Choi matrices of the map x ÞÑ i,j αij s˚i xsj ,
we conclude that α “ rαij s “ 0 by Lemma 1.5.7. For the general cases, we take
ř ř
conjugate to get i,j ᾱji s˚i sj “ 0. Then we have i,j pαij ` ᾱji qs˚i sj “ 0. Since
α ` α˚ is self-adjoint, we see that α ` α˚ “ 0. By the same reasoning, we also have
iα ´ iα˚ “ 0. Therefore, we conclude that α “ 21 pα ` α˚ q ` 2i1 piα ´ iα˚ q must be
zero. ˝
1.5.4 Decomposable maps and PPT states

Recall that a map in SPm^n `SPm^n is called decomposable. Then φ is decomposable
if and only if it is the sum of a completely positive map and a completely copositive
map. The convex cone of all decomposable maps will be denoted by DECrMm , Mn s
or just by DEC, and the corresponding convex cone of matrices in Mm b Mn is
44
denoted by DECpMm b Mn q or DEC whose member is the sum of a positive matrix
and a copositive matrix. On the other hand, the linear map corresponding to a
PPT matrix will be also called a PPT map, and the convex cone of all PPT maps
is denoted by PPTrMm , Mn s or PPT. A map is a PPT map if and only if it is both
completely positive and completely copositive. We may summarize as follows:
LpMm , Mn q : SP1 Ă PPT Ă CP Ă DEC Ă P1
Ó JC Ó Ó Ó Ó Ó (1.47)
Mm b Mn : S1 Ă PPT Ă pMm b Mn q` Ă DEC Ă BP 1
Suppose that C1 and C2 are closed convex cones in a finite dimensional real
vector space X. Then the identity pC1 ` C2 q˝ “ C1˝ X C2˝ is easily seen. We also
have a clear inclusion pC1 X C2 q˝ Ą C1˝ ` C2˝ , from which we have
C1 X C2 Ă pC1˝ ` C2˝ q˝ “ C1˝˝ X C2˝˝ “ C1 X C2 .
Hence, we have the relations
pC1 ` C2 q˝ “ C1˝ X C2˝ , pC1 X C2 q˝ “ C1˝ ` C2˝ , (1.48)
for closed convex cones C1 and C2 . Therefore, we see that two convex cones PPT
and DEC are dual to each other with respect to the pairing (1.40). On the other
hand, PPT and DEC are dual with respect to the pairing (1.39).
Theorem 1.5.9 A linear map φ : Mm Ñ Mn is decomposable if and only if the

inequality x%, φy ě 0 holds for every PPT state % P Mm b Mn .
In view of two diagrams (1.45) and (1.47), it is natural to look for any relations
between PPT and Sk , or equivalently those between DEC and Pk . We first note that
S1 rMm b Mn s “ PPT rMm b Mn s holds if and only if DECrMm , Mn s “ P1 rMm , Mn s
holds. Theorem 1.3.4 tells us the relation DECrM2 , M2 s “ P1 rM2 , M2 s holds, by
which we also have
S1 rM2 b M2 s “ PPT rM2 b M2 s.
It was also shown in [129] that the following identities
DECrM2 , M3 s “ P1 rM2 , M3 s, DECrM3 , M2 s “ P1 rM3 , M2 s
hold which are equivalent to S1 rM2 b M3 s “ PPT rM2 b M3 s.
45
By the example (1.33), we know that the strict inclusion
S1 rM3 b M3 s Ř PPT rM3 b M3 s
holds in M3 b M3 , and so we have
DECrM3 , M3 s Ř P1 rM3 , M3 s.
We will see a concrete example for this strict inclusion in the next section. It was
shown in [18] that the following inclusion
P2 rM3 , M3 s Ă DECrM3 , M3 s (1.49)
holds for a special class of linear maps, and the equivalent dual inclusion relation
S2 rM3 b M3 s Ă PPT rM3 b M3 s
had been conjectured in [100]. The relation (1.49) was proved in [133], and the
relation Pn´1 rMn , Mn s Ă DECrMn , Mn s is conjectured in [26].
It is also an interesting question to ask how large Schmidt numbers are attained
by PPT states. It is known [17, 65, 14] that PPT states may have arbitrary large
Schmidt numbers when we increase the size of matrices. It is equivalent to look for
indecomposable k-positive maps for large k. See [10]. In the case of m ě 3 and
n “ 4, it is unknown if there exists a PPT state with Schmidt number greater than
2, or equivalently if there exists an indecomposable 2-positive map.
References: [129], [18], [100], [133], [17], [65], [26], [14], [10]
1.6 Nontrivial examples of positive maps

In this section, we exhibit Tomiyama’s examples of linear maps which distinguish
k-positivity for different k’s, by considering the line segment between the trace map
and the identity map, which are located at the center and the boundary of the
convex cone P1 , respectively. We also give two examples of indecomposable positive
maps by Choi and Woronowicz.
1.6.1 Tomiyama’s example of k-positive maps

In this section, we will explore examples of positive maps which are not completely
positive. By the strict inclusion (1.13), we also have the following strict inclusions
of convex cones in the space HpMn , Mn q;
CP “ Pn Ř Pn´1 Ř ¨ ¨ ¨ Pk Ř ¨ ¨ ¨ Ř P1 .
46
We first look for explicit examples to distinguish the above convex cones.
We recall that the trace map n1 Tr n is located in the center of the convex cone
CP, and the identity map idn generates an extreme ray of P1 . We consider the line
segment between these two maps to define
1
φλ “ p1 ´ λq Tr n ` λ idn , ´8 ă λ ă `8. (1.50)
n
It is clear that φλ P CP for λ P r0, 1s, and φλ P P1 implies λ ď 1. We have
n n
ÿ 1´λ ÿ
Cφλ “ |iyxj| b δij In ` λ |iyxj| b |iyxj|
i,j“1
n i,j“1
1´λ
“ In b In ` λ|ωyxω|,
n
with |ωy “ ni“1 |iy|iy P Cn b Cn . We take |ζy “ ki“1 |iy|iy with SR |ζy “ k. If
ř ř
φλ P Pk then we have
1´λ k
0 ď xCφλ , |ζyxζ|y “ k ` λk 2 “ r1 ` λpnk ´ 1qs,
n n
from which we get a necessary condition
´1
ďλď1 (1.51)
nk ´ 1
´1
for φλ P Pk . When λ “ nk´1
, the map φλ is the scalar multiplication of the map
τn,k “ kTr n ´ idn , k “ 1, 2, . . . , n, (1.52)
whose Choi matrix is given by
Cτn,k “ kIn b In ´ |ωyxω|.
In order to show that (1.51) is also sufficient for φλ P Pk , it suffices to show that τn,k
is k-positive. For this purpose, we need the following:
Proposition 1.6.1 For a linear map φ : Mm Ñ Mn and k “ 1, 2, . . . , n, the follow-

ing are equivalent:
(i) φ is k-positive,
(ii) Adp ˝φ˚ is completely positive for every rank k projection p in Mm ,
(iii) pIn b pqCφ˚ pIn b pq P Mn b Mm is positive for every rank k projection p in

Mm .
(iv) pp b In qCφ pp b In q P Mm b Mn is positive for every rank k projection p in Mm .
47
|ξ¯i yxξ¯j |bφp|ξi yxξj |q P Mm bMn is positive for every orthornormal family
řk
(v) i,j“1
tξi : i “ 1, 2, . . . , ku in Cm .
řk
(vi) i,j“1 |iyxj| b φp|ξi yxξj |q P Mk b Mn is positive for every orthornormal family
tξi : i “ 1, 2, . . . , ku in Cm .
Proof. In order to prove (i) ùñ (ii), we suppose that φ is k-positive and p P Mm

is a rank k projection. Take a unitary u P Mm which sends Ck ˆ t0u Ă Cm onto the
range of p. Then Adu ˝ Adp ˝φ˚ is a k-positive map from Mn into Mk , and so it is
completely positive by Theorem 1.5.3. Therefore, we see that Adp ˝φ˚ is completely
positive. For the reverse direction, we suppose that (ii) holds, and s is arbitrary
rank k matrix in Mmˆn . Take a projection p P Mm such that rank p ď k and ps “ s,
to see that
xφ, Ads y “ xφ, Adps y “ xφ, Ads ˝ Adp y “ xAdsT ˝φ, Adp y “ xAdsT , Adp ˝φ˚ y.
By the assumption, Adp ˝φ˚ is completely positive, and so we have xφ, Ads y ě 0 and
conclude that φ is k-positive. The equivalence (ii) ðñ (iii) also comes out from
Theorem 1.5.3, since the Choi matrix of Adp ˝φ˚ is just pIn b pqCφ˚ pIn b pq. Since
Cφ˚ is the flip of Cφ , we see that (iii) and (iv) are equivalent.
We take arbitrary projection p of rank k, and write p “ ki“1 |ξ¯i yxξ¯i | for or-
ř
thonormal family t|ξi y : i “ 1, 2, . . . , ku of Cm . Then we have

m
ÿ
pp b In qCφ pp b In q “ p|syxt|p b φp|syxt|q
s,t“1
m ÿ
ÿ k
“ |ξ¯i yxξ¯i |syxt|ξ¯j yxξ¯j | b φp|syxt|q
s,t“1 i,j“1
˜ ¸
k
ÿ m
ÿ
“ |ξ¯i yxξ¯j | b φ xξ¯i |sy|syxt|xt|ξ¯j y
i,j“1 s,t“1
k
ÿ
“ |ξ¯i yxξ¯j | b φp|ξi yxξj |q,
i,j“1
which shows (iv) ðñ (v). The remaining equivalence between (v) and (vi) is easily
seen by considering the isometry from Ck into Cm which sends |iy to |ξ¯i y. ˝
Proposition 1.6.2 For k “ 1, 2, . . . , n, the map τn,k : Mn Ñ Mn is k-positive.
Proof. In order to apply Proposition 1.6.1 (vi), we take an orthonormal family

t|ξi y : i “ 1, 2, . . . , ku of vectors in Cn . Then we have τn,k p|ξi yxξj |q “ kδij In ´ |ξi yxξj |,
48
and so it follows that
ÿk k
ÿ k
ÿ
|iyxj| b φp|ξi yxξj |q “ k |iyxi| b In ´ |iyxj| b |ξi yxξj |
i,j“1 i“1 i,j“1
“ kIk b In ´ |ξyxξ|,
řk
with |ξy “ i“1 |iy|ξi y P Ck b Cn . Now, we conclude that kIk b In ´ |ξyxξ| ě 0 holds
since xξ|ξy “ k. ˝
Theorem 1.6.3 For k “ 1, 2, . . . , n, the map φλ : Mn Ñ Mn is k-positive if and

only if λ satisfies the inequality
´1
ď λ ď 1.
nk ´ 1
We note that φλ is completely copositive if and only if pCφλ qΓ is positive if and
only if 1´λ
n
ě |λ| if and only if
´1 1
ďλď .
n´1 n`1
Summarizing, we have the following:
Proposition 1.6.4 The linear φλ in (1.50) satisfies the following:
´1
(i) φλ is positive if and only if n´1
ď λ ď 1,
´1
(ii) φλ is completely positive if and only if n2 ´1
ď λ ď 1,
´1 1
(iii) φλ is completely copositive if and only if n´1
ďλď n`1
.
Therefore, φλ is either completely positive or completely copositive whenever it
is positive. Especially, there is no indecomposable positive maps among φλ ’s.
References: [20], [125], [120], [126]
1.6.2 The Choi map between 3 ˆ 3 matrices

In the case of n “ 3, the map τ3,1 “ 2φ´1{2 is located at the end point of the interval
on which φλ is positive, and its Choi matrix is given by
¨ ˛
¨ ¨ ¨ ¨ ´1 ¨ ¨ ¨ ´1
˚ ¨ 1 ¨ ¨ ¨ ¨ ¨ ¨ ¨ ‹
˚ ‹
˚ ¨ ¨ 1 ¨ ¨ ¨ ¨ ¨ ¨ ‹
˚ ‹
˚ ‹
˚ ‹
˚ ¨ ¨ ¨
˚ 1 ¨ ¨ ¨ ¨ ¨ ‹
‹
Cτ3,1 “ ˚
˚ ´1 ¨ ¨ ¨ ¨ ¨ ¨ ¨ ´1 ‹.
‹
˚ ¨ ¨ ¨
˚ ¨ ¨ 1 ¨ ¨ ¨ ‹ ‹
˚ ‹
˚ ‹
˚ ¨ ¨ ¨
˚ ¨ ¨ ¨ 1 ¨ ¨ ‹ ‹
˝ ¨ ¨ ¨ ¨ ¨ ¨ ¨ 1 ¨ ‚
´1 ¨ ¨ ¨ ´1 ¨ ¨ ¨ ¨
49
In order to get an example of indecomposable positive map, we adjust the diagonal
part of Cτ3,1 as follow:
¨ ˛
1 ¨ ¨ ¨ ´1 ¨ ¨ ¨ ´1
˚ ¨ 1 ¨ ¨ ¨ ¨ ¨ ¨ ¨ ‹
˚ ‹
˚ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ‹
˚ ‹
˚ ‹
˚ ‹
˚ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ‹
˚ ‹
˚ ´1 ¨ ¨
˚ ¨ 1 ¨ ¨ ¨ ´1 ‹.
‹ (1.53)
˚ ¨ ¨ ¨
˚ ¨ ¨ 1 ¨ ¨ ¨ ‹
‹
˚ ‹
˚ ‹
˚ ¨ ¨ ¨
˚ ¨ ¨ ¨ 1 ¨ ¨ ‹
‹
˝ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ‚
´1 ¨ ¨ ¨ ´1 ¨ ¨ ¨ 1
Note that neither the above matrix nor its partial transpose is positive. This is the
Choi matrix of the map defined by
¨ ˛
x11 ` x33 ´x12 ´x13
φch pxq “ ˝ ´x21 x22 ` x11 ´x23 ‚ “ ψpxq ´ x, (1.54)
´x31 ´x32 x33 ` x22
for x “ rxij s P M3 , with

¨ ˛
2x11 ` x33 ¨ ¨
ψpxq “ ˝ ¨ 2x22 ` x11 ¨ ‚.
¨ ¨ 2x33 ` x22
In order to show that φch is positive, we begin with the following:
Proposition 1.6.5 Let A be an n ˆ n nonsingular positive matrix, and |ξ0 y a unit

vector. Then A ě |ξ0 yxξ0 | if and only if xξ0 |A´1 |ξ0 y ď 1.
Proof. We have A ě |ξ0 yxξ0 | if and only if I ě A´1{2 |ξ0 yxξ0 |A´1{2 if and only if
}A´1{2 |ξ0 y} ď 1 if and only if xξ0 |A´1 |ξ0 y ď 1. ˝
Now, we see that φch is positive if and only if ψp|ξ0 yxξ0 |q ě |ξ0 yxξ0 | for every unit
vector |ξ0 y if and only if
xξ0 |ψp|ξ0 yxξ0 |q´1 |ξ0 y ď 1
for every unit vector |ξ0 y if and only if the inequality
α β γ
` ` ď1 (1.55)
2α ` γ 2β ` α 2γ ` β
holds for all positive real numbers α, β and γ.
Lemma 1.6.6 The inequality (1.55) holds for all positive real numbers α, β and γ.
50
Proof. Taking x “ αγ , y “ α
β
and z “ βγ , it suffices to show the inequality
1 1 1
` ` ď 1,
2`x 2`y 2`z
or equivalently xy ` yz ` zx ě 3 under the constraint xyz “ 1. This comes out by
comparing the arithmetic and geometric means. ˝
Therefore, we conclude that the map φch is a positive map. One may check that
the bilinear pairing of the matrix (1.53) with the PPT matrix in (1.33) is strictly
negative, and so we also conclude that φch is not decomposable. This is another way
to see that the PPT state in (1.33) is entangled. The map φch in (1.54) is usually
called the Choi map.
Theorem 1.6.7 The Choi map φch defined in (1.54) is a positive map which is not
decomposable.
The Choi map φch had been extended in higher dimensions in various directions
[121, 89, 18, 132, 90, 41]. See also Section 2.5.
References: [22], [25], [23], [121], [89], [18], [132], [90], [41]
1.6.3 The Woronowicz map from 2 ˆ 2 matrices to 4 ˆ 4 ma-

trices
We also give an example of an indecomposable positive map from M2 into M4 . We
define the linear map φwo from M2 into M4 by
¨ ˛
ˆ ˙ 4x ´ 2y ´ 2z ` 3w ´2x ` 2z ¨ ¨
x y ˚ ´2x ` 2y 2x z ¨ ‹
φwo : ÞÑ ˚ ‹ . (1.56)
z w ˝ ¨ y 2w ´2z ´ w‚
¨ ¨ ´2y ´ w 4x ` 2w
Then the map φwo sends the rank one matrix
ˆ ˙
1 ᾱ
pα “
α |α|2
to ¨ ˛
|α ´ 2|2 ` 2|α|2 ´2 ` 2α ¨ ¨
˚ ´2 ` 2ᾱ 2 α ¨ ‹
2 ‚.
˚ 2
‹
˝ ¨ ᾱ 2|α| ´2α ´ |α|
¨ ¨ ´2ᾱ ´ |α|2 4 ` 2|α|2
The k ˆ k principal submatrix of the left–upper corner has the determinant ∆k as
follows:
∆1 “ 2|α|2 ` |α ´ 2|2 ,
∆2 “ 2|α|2 ` 4,
∆3 “ |α|2 |α ` 2|2 ,
∆4 “ 0.
51
Therefore, we see that φwo ppα q is positive when α ‰ 0, ´2. For α “ 0, α “ ´2
and α “ 8, the map φwo sends
ˆ ˙ ˆ ˙ ˆ ˙
1 ¨ 1 ´2 ¨ ¨
p0 “ , p´2 “ and p8 “
¨ ¨ ´2 4 ¨ 1
to the matrices
¨ ˛ ¨ ˛ ¨ ˛
4 ´2 ¨ ¨ 24 ´6 ¨ ¨ 3 ¨ ¨ ¨
˚´2 2 ¨ ¨‹ ˚´6 2 ´2 ¨ ‹ ˚¨ ¨ ¨ ¨ ‹
˚ ‹, ˚ ‹ and ˚ ‹,
˝ ¨ ¨ ¨ ¨‚ ˝ ¨ ´2 8 ¨‚ ˝¨ ¨ 2 ´1‚
¨ ¨ ¨ 4 ¨ ¨ ¨ 20 ¨ ¨ ´1 2
respectively, which are positive. Therefore, we see that the map φwo is a positive
map. We note that φppα q is of rank three for every α P C Y t8u. The map φwo is
called the Woronowicz map. The Choi matrix of φwo is given by
¨ ˛
4 ´2 ¨ ¨ ´2 ¨ ¨ ¨
˚´2 2 ¨ ¨ 2 ¨ ¨ ¨ ‹
˚ ‹
˚ ¨
˚ ¨ ¨ ¨ ¨ 1 ¨ ¨ ‹
‹
˚ ¨
˚ ¨ ¨ 4 ¨ ¨ ´2 ¨ ‹
‹
Cφwo “ ˚
˚
‹.
‹
˚´2 2 ¨ ¨ 3 ¨ ¨ ¨ ‹
˚ ‹
˚ ¨
˚ ¨ 1 ¨ ¨ ¨ ¨ ¨ ‹
‹
˝ ¨ ¨ ¨ ´2 ¨ ¨ 2 ´1‚
¨ ¨ ¨ ¨ ¨ ¨ ´1 2
We will see in Section 2.6 that the map φwo in (1.56) is an extremal positive
map. If φwo is decomposable, then it must be of the form Ads or T ˝ Ads , and so
the Choi matrix is of rank one or its partial transpose is of rank one. We see that
neither is the case, and so φwo is not decomposable.
1.7 Isotropic states and Werner states

We continue to look for conditions for k-superpositivity of the maps by Tomiyama in
the last section, or equivalently, determine the Schmidt numbers of the corresponding
Choi matrices and their partial transposes. These are isotropic states and Werner
states, respectively.
1.7.1 Isotropic states

We return to the map φλ defined in (1.50) which is completely positive if and only
if n´1
2 ´1 ď λ ď 1 by Theorem 1.6.3. In this case, we have the state
1´λ 1´λ
%λ :“ Cφλ “ In b In ` λ|ωyxω| “ CTr ` λCid (1.57)
n n
52
which is called the isotropic state. We recall that the vector |ωy was defined by
|ωy “ ni“1 |iiy. We have already seen in Proposition 1.6.4 that %Γλ is positive if and
ř
´1 1
only if φλ is completely copositive if and only if n´1 ď λ ď n`1 , and so %λ is a PPT
state if and only if
´1 1
2
ďλď
n ´1 n`1
holds.
When λ “ n´1 2 ´1 , one can show that
n 1
%λ “ In b In ´ 2 |ωyxω| (1.58)
n2 ´1 n ´1
is separable. In fact, we modify the vector |ζy given in (1.30) to define
|ζik y “ p|iy ` β|kyq b p|iy ´ β̄|kyq “ |iiy ´ β̄|iky ` β|kiy ´ |kky
with a complex number β of modulus one, and put

1 ÿ
%ik “ |ζik yxζik |
4 β“˘1,˘i
“ |iiyxii| ` |ikyxik| ` |kiyxki| ` |kkyxkk| ´ |iiyxkk| ´ |kkyxii|
for each ordered pair pi, kq with i ‰ k and i, k “ 1, 2, . . . , n. Summing up all of
them, we get the separable state
˜ ¸
ÿn ÿ n
ÿ
2pn ´ 1q |iiyxii| ` 2 |ikyxik| ´ 2 |ωyxω| ´ |iiyxii|
i“1 i‰k i“1
n
ÿ ÿ
“2n |iiyxii| ` 2 |ikyxik| ´ 2|ωyxω|.
i“1 i‰k
Adding the separable state 2pn ´ 1q ni‰j |ijyxij|, we finally get the state %λ in (1.58)
ř
up to scalar multiplication. Therefore, we see that %λ P Sn if and only if %λ P S1 , in

case of λ ă 0.
We proceed to determine the Schmidt number of the state %λ , or equivalently
look for the condition on λ for which φλ is k-superpositive. In order to determine
the Schmidt number of the state %λ , we consider the bilinear pairing with φµ for
´1
µ “ nk´1 . We compute
ˆ ˙ˆ ˙
1´λ 1´µ 1´λ 1´µ
x%λ , %µ y “ n `λ ` µ ` pn2 ´ nq ` pn2 ´ nq λµ
n n n n
“ λµpn2 ´ 1q ` 1.
´1
Therefore, %λ P Sk implies that x%λ , %µ y ě 0 with µ “ nk´1 , which holds only when
nk´1
λ ď n2 ´1 . Therefore, we have the following necessary condition
´1 nk ´ 1
ď λ ď . (1.59)
n2 ´ 1 n2 ´ 1
53
for φλ P Sk . Since ˆ ˙
nk ´ 1 ´1 k ´1
“ ` 1´ 2 ,
n2 ´ 1 n2 ´ 1 n n ´1
“ ‰
we see that λk :“ nk´1
n2 ´1
divides the interval n´1
2 ´1 , 1 into n subintervals with the
same lengths.
In order to get sufficient conditions for φλ P SPk , we will use special kinds of
symmetries of the identity map idn and its Choi matrix Cid . We call U b V acting
on Cm b Cn a local unitary when both U and V are unitaries acting on Cm and Cn ,
respectively. It is clear that both pU b V q|ζy and |ζy P Cm b Cn share the same
Schmidt ranks, and so it follows that both states
AdU bV p%q “ pU ˚ b V ˚ q%pU b V q
and % have the same Schmidt numbers. For a linear map φ : Cm Ñ Cn , we have
xa b b, AdU bV pCφ qy “ xAdU T bV T pa b bq, Cφ y
“ xAdU T paq b AdV T pbq, Cφ y
“ xAdV T pbq, φ ˝ AdU T paqy
“ xb, AdV ˝φ ˝ AdU T paqy
“ xa b b, CAdV ˝φ˝AdU T y,
and so it follows that
AdU bV pCφ q “ CAdV ˝φ˝AdU T . (1.60)
Two nˆn unitaries U and V satisfy AdU bV pCid q “ Cid if and only if AdV ˝ AdU T “ id
if and only if U T V “ λIn with |λ| “ 1. Especially, we have
AdU bŪ pCid q “ Cid ,
for every unitary U .

Now, we look for linear maps φ P LpMn , Mn q satisfying AdU bŪ pCφ q “ Cφ for
every unitary U . This holds if and only if AdŪ ˝φ ˝ AdU T “ φ if and only if
x|kyx`|, φp|iyxj|qy “ xU |kyx`|U ˚ , φpŪ |iyxj|U T qy. (1.61)
Now, we suppose that φ satisfies (1.61) for every unitary U , and proceed to determine
the pk, `q entry
x|kyx`|, φp|iyxj|qy
of φp|iyxj|q for those φ. We first consider the case when one of i, j, k and `, say i, is
different from others. In this case, we take the unitary U satisfying U |iy “ ´|iy and
U |ιy “ |ιy for ι “ j, k and `, to see that x|kyx`|, φp|iyxj|qy “ 0. Therefore, it remains
to determine
x|iyxj|, φp|iyxj|qy, x|jyxi|, φp|iyxj|qy, x|kyxk|, φp|iyxi|qy, x|iyxi|, φp|iyxi|qy,
54
with i ‰ j and i ‰ k. Taking the unitary U with U |iy “ |iy and U |jy “ i|jy, we see
that x|jyxi|, φp|iyxj|qy “ 0. Taking the unitaries permuting |iy’s, we also see that the
values of
α :“ x|kyxk|, φp|iyxi|qy, β :“ x|iyxj|, φp|iyxj|qy
are independent on the choices of i, j and i, k, respectively. Now, we take the unitary
U with U |iy “ ?12 p|iy ` |jyq and U |jy “ ?12 p|iy ´ |jyq to see that
α ` β “ x|iyxj|, φp|iyxj|qy ` x|jyxj|, φp|iyxi|qy

1
“ px|iyxi|, φp|iyxi|qy ` x|jyxj|, φp|jyxj|qyq
2
for i ‰ j. Taking finally the unitary U with U |iy “ |jy and U |jy “ |iy for i ‰ j, we
have
x|iyxi|, φp|iyxi|qy “ x|jyxj|, φp|jyxj|qy,
for i ‰ j, and so we also have
x|iyxi|, φp|iyxi|qy “ α ` β
for every i. Therefore, we have the following:
Proposition 1.7.1 For a map φ : Mn Ñ Mn , the following are equivalent:
(i) AdU bŪ pCφ q “ Cφ for every unitary U ,
(ii) AdŪ ˝φ ˝ AdU T “ φ for every unitary U ,
(iii) φ “ α Tr n ` β idn for complex numbers α and β.
For a given φ P P1 rMn , Mn s, we define the linear map

ż
T rφs :“ AdŪ ˝φ ˝ AdU T dU,
U pnq
where the integration is taken over the unitary group U pnq of all nˆn unitaries with
respect to the Haar measure. More precisely, this map is defined by the relation
ż
xT rφs, ψy “ xAdŪ ˝φ ˝ AdU T , ψy dU, ψ P LrMn , Mn s.
U pnq
Because φ is k-superpositive if and only if AdŪ ˝φ˝AdU T is k-superpositive for every

unitary U , we see that if φ is k-superpositive then T rφs is also k-superpositive. It
is also clear that T rφs satisfies the condition (ii) of Proposition 1.7.1, and so we
55
have T rφs “ αTr n ` βidn for complex numbers α and β. In order to determine the
coefficients α and β, we first note the relations
xAdŪ ˝φ ˝ AdU T , Tr y “ xφ, AdU ˚ ˝Tr ˝ AdU y “ xφ, Tr y

xAdŪ ˝φ ˝ AdU T , idy “ xφ, AdU ˚ ˝id ˝ AdU y “ xφ, idy,
from which we have

xφ, idy “ xT rφs, idy “ xαTr ` βid, idy “ nα ` n2 β
xφ, Tr y “ xT rφs, Tr y “ xαTr ` βid, Tr y “ n2 α ` nβ.
Solving this equations, we have

1 1
α“ xφ, Tr n ´ n1 idn y, β“ xφ, idn ´ n1 Tr n y. (1.62)
n2 ´1 n2 ´1
We summarize as follows:
Proposition 1.7.2 For a linear map φ : Mn Ñ Mn , we define complex numbers α

and β by (1.62). Then we have T rφs “ αTr n ` βidn .
We take the positive map φ whose Choi matrix is given by Cφ “ nk |ζyxζ| with
|ζy “ ki“1 |iiy P Cn b Cn . Then Cφ has Schmidt number k, and φ is k-superpositive.
ř
We also have
n´k nk ´ 1
α“ 2 , β“ 2
n ´1 n ´1
and T rφs “ φλk with λk “ nk´1 n2 ´1
. Therefore, we conclude that φλk is k-superpositive.
We have already seen that %λ with λ “ n´1 2 ´1 is separable, and so it belongs to Sk for
every k “ 1, 2, . . . , n. Therefore, we conclude that (1.59) gives rise to the necessary

and sufficient condition for %λ P Sk , and summarize as follows:
Theorem 1.7.3 The linear map φλ defined by (1.50) is k-superpositive if and only
if the isotropic state %λ in (1.57) belongs to Sk if and only if λ satisfies
´1 nk ´ 1
ď λ ď .
n2 ´ 1 n2 ´ 1
References: [128], [59], [124], [9], [55]
1.7.2 Werner states

It remains to look for conditions on λ for which %λ P S k , or equivalently, %Γλ P Sk .
1
We recall that %Γλ is positive if and only if n´1
´1
ď λ ď n`1 . For those λ’s, the state
Γ
%λ is called the Werner state, which is given by
%Γλ “ pCφλ qΓ “ Cφλ ˝T “ Cψλ ,
56
where the map ψλ : Mn Ñ Mn is defined by
Tr
ψλ :“ φλ ˝ T “ p1 ´ λq ` λT.
n
´1
Note that ψλ is positive if and only if φλ is positive if and only if n´1 ď λ ď 1.
Γ
We also recall that %λ is separable if and only if %λ is separable if and only if
´1 1
n2 ´1
ď λ ď n`1 by Theorem 1.7.3.
In order to determine the Schmidt number of %Γλ , we take |ζij y “ |ijy ´ |jiy P
Cn b Cn with Schmidt rank two. Summing up all of them, we have
ÿ ÿ
|ζij yxζij | “ p|ijyxij| ´ |ijyxji|q “ CTr ´ CT ,
iăj i‰j
which is the Choi matrix of the map

ˆ ˙
´1 1 ´1
ψ´1{pn´1q “ φ´1{pn´1q ˝ T “ 1´ Tr n ` Tn ,
n´1 n n´1
up to a scalar multiple.
Theorem 1.7.4 For k “ 2, 3, . . . , n, the Werner state %Γλ belongs to Sk if and only
if ψλ is k-superpositive if and only if if and only if the following
´1 1
ďλď
n´1 n`1
holds.
Proof. Suppose that %Γλ P Sk . Then the positivity of %Γλ implies the condition.
1
When λ “ n´1 ´1
, we have seen that %Γλ P S2 Ă Sk . For λ “ n`1 , we know that
Γ
%λ P S1 Ă Sk . ˝
Therefore, we see that the state %Γλ has Schmidt number two whenever it is not
separable. We exhibit the properties of Werner states corresponding to isotropic
states. By the relation (1.60), we have
AdU bU pCT q “ CT .
This property actually characterizes the Werner states as follows:
Proposition 1.7.5 For a map ψ : Mn Ñ Mn , the following are equivalent:
(i) AdU bU pCψ q “ Cψ for every unitary U ,
(ii) AdU ˝ψ ˝ AdU T “ ψ for every unitary U ,
(iii) ψ “ αTr n ` βTn for complex numbers α and β.
57
Proof. We replace φ in Proposition 1.7.1 by ψ “ T ˝ φ to get the condition (iii).
From Proposition 1.7.1 (ii), we also have
ψ “ T ˝ φ “ T ˝ AdŪ ˝φ ˝ AdU T “ AdU ˝T ˝ φ ˝ AdU T “ AdU ˝ψ ˝ AdU T ,
which is (ii). This is equivalent to (i) by (1.60). ˝

For a given ψ P P1 rMn , Mn s, we define the linear map
ż
W rψs :“ AdU ˝ψ ˝ AdU T dU,
U pnq
as the definition of T rφs. Then we have
W rψs “ αTr n ` βTn ,
with
1 1
α“ xψ, Tr n ´ n1 Tn y, β“ 2 xψ, Tn ´ n1 Tr n y
n2
´1 n ´1
by the exactly same method as that of Proposition 1.7.2. We take |ζy “ |12y´|21y P
Cn b Cn with SR |ζy “ 2. We also take ψ P LpMn , Mn q with Cψ “ |ζyxζ|. Then we
have xψ, Tn y “ ´2 and xψ, Tr n y “ 2, and so, we see that the following linear map
npn ` 2q 1
W rψs “ pTr ´ Tq “ ψ´1{pn´1q
2pn ´ 1q n´1
is 2-superpositive, which recovers the Schmidt number of the corresponding Werner
states.
Finally, we determine λ for which φλ is k-copositive, or equivalently, ψλ is k-
positive, for k “ 2, 3, . . . , n. Suppose that ψλ P Pk . Then we have
0 ď xψλ , ψ´1{pn´1q y “ ´pn ` 1qλ ` 1,

1
and so λ ď n`1 . This implies that ψλ is completely positive by Proposition 1.6.4,
and so k-positive. Because φ´1{pn´1q is both 1-copositive and completely copositive,
we have the following:
Theorem 1.7.6 For k “ 2, 3, . . . , n, the map φλ is k-copositive if and only if ψλ is

k-positive if and only if the following inequality
´1 1
ďλď
n´1 n`1
holds.
Together with Theorem 1.7.5, we see that 2-superpositivity and 2-positivity co-
incide for the linear maps ψλ ’s. We summarize in Figure 1.2.
58
P1
Pk
CP
SPk
SP1
q q q 0r q q r λ
´1 ´1 ´1 1 nk´1
n´1 nk´1 2
n ´1 n`1 n2 ´1
1
SP2 “ CCP “ P2
Figure 1.2: Various kinds of positivity of the map φλ “ p1 ´ λq n1 Tr n ` λidn
1.8 Mapping cones and tensor products

Various convex cones of positive maps considered so far are invariant under compo-
sitions by completely positive maps. This property implies further characterizations
of various kinds of positive maps in terms of composition, ampliation and tensor
products. In order to explain them in a unified framework, we exhibit an iden-
tity which relates compositions and tensor products of linear maps between matrix
algebras.
In this section, we use the notations MA and MB for matrix algebras which
are acting on the finite dimensional spaces CA and CB , respectively, in order to
emphasize the roles of parties in the tensor product. The convex cone CPrMA , MB s
will be abbreviated by CPAB , sometimes just by CP; the identity map on MA will
be denoted by idA . For sets K1 and K2 of linear maps, K1 ˝ K2 will denote the set
of all φ1 ˝ φ2 with φi P Ki for i “ 1, 2. We also denote by CK the set of of all Cφ
with φ P K.
1.8.1 Mapping cones of positive maps

A closed convex cone K of positive linear maps in HpMA , MB q is called a right
mapping cone if K ˝ CPAA Ă K, and a left mapping cone if CPBB ˝ K Ă K. A
closed convex cone K is also called a mapping cone when CPBB ˝ K ˝ CPAA Ă K
holds. Since idA P CPAA and idB P CPBB , it is clear that K is a mapping cone if
and only if it is both left and right mapping cones. The following proposition is an
immediate consequence of the identity (1.42);
xψ ˝ φ, σy “ xφ, ψ ˚ ˝ σy “ xψ, σ ˝ φ˚ y.
Proposition 1.8.1 Suppose that K is a closed convex cone satisfying the relation
SP1 Ă K Ă P1 . Then K is a right mapping cone if and only if K ˝ is a right mapping
cone.
59
Proof. Note that K is a right mapping cone if and only if xφ ˝ σ, ψy ě 0 for every
φ P K, σ P CPAA and ψ P K ˝ . By the identity (1.42), this is the case if and only if
xφ, ψ ˝ σ ˚ y ě 0 for every φ P K, σ P CPAA and ψ P K ˝ if and only if K ˝ is a right
mapping cone, since pCPAA q˚ “ CPAA . ˝
It is clear that SPk is a mapping cone, and so all the convex cones in the mapping
space HpMA , MB q appearing in the diagrams (1.45) are mapping cones. If K1 and
K2 are mapping cones then their intersection K1 X K2 and their convex hull K1 ` K2
are also mapping cones. Therefore, all the convex cones of HpMA , MB q appearing
in (1.47) are also mapping cones. We note that K Ă K ˝ CP holds automatically,
and so we have pK ˝ CPq˝ Ă K ˝ in general.
Proposition 1.8.2 A closed convex cone K of positive maps in HpMA , MB q is a

left mapping cone if and only if K ˝ “ pCPBB ˝ Kq˝ , and it is a right mapping
cone if and only if K ˝ “ pK ˝ CPAA q˝ . It is a mapping cone if and only if K ˝ “
pCPBB ˝ K ˝ CPAA q˝ .
As another application of the identity (1.42), we see that φ P pK ˝ CPq˝ if and

only if xφ, ψ ˝ σy ě 0 for every ψ P K and σ P CP if and only if xψ ˚ ˝ φ, σy ě 0 for
every ψ P K and σ P CP if and only if ψ ˚ ˝ φ P CP for every ψ P K. In short, we
have
pK ˝ CPq˝ “ tφ P HpMm , Mn q : ψ ˚ ˝ φ P CP for every ψ P Ku,
which coincide with K ˝ if and only if K is a right mapping cone by Proposition
1.8.2. Therefore, we see that the dual cone of a mapping cone K can be described
in terms of composition as well as bilinear pairing, and this is possible only when K
is a one-sided mapping cone. We summarize as follows:
Theorem 1.8.3 For a closed convex cone K of positive maps in the space HpMA , MB q,
the following are equivalent:
(i) K is a right mapping cone,
(ii) φ P HpMA , MB q belongs to K ˝ if and only if ψ ˚ ˝ φ is completely positive for

every ψ P K.
The following are also equivalent:
(iii) K is a left mapping cone,
(iv) φ P HpMA , MB q belongs to K ˝ if and only if φ ˝ ψ ˚ is completely positive for

every ψ P K.
References: [114], [116], [106], [107], [117], [36]
60
φ˚1
MA1 - MB1
φ1
σ φ1 b φ2 -
- φ2 ˝ σ ˝ φ˚1
? φ2 ?
MA2 - MB2
Figure 1.3: The map φ1 b φ2 sends Cσ to Cφ2 ˝σ˝φ˚1 .
1.8.2 Tensor products and compositions of linear maps

So far, we have described pK ˝ CPq˝ and pCP ˝ Kq˝ in terms of compositions. We are
going to describe them in terms of tensor products of linear maps. To do this, we
need an identity which connects compositions and tensor products of linear maps.
For given linear maps φi : MAi Ñ MBi with i “ 1, 2, we note that every element of
the domain MA1 b MA2 of the map φ1 b φ2 can be written as the Choi matrix Cσ of
a linear map σ : MA1 Ñ MA2 , and Cφ2 ˝σ˝φ˚1 belongs to the range MB1 b MB2 . See
Figure 1.3.
For every b1 P MB1 and b2 P MB2 , we have
xb1 b b2 , Cφ2 ˝σ˝φ˚1 yB1 B2 “ xb2 , φ2 pσpφ˚1 pb1 qqqyB2

“ xσ ˚ pφ˚2 pb2 qq, φ˚1 pb1 qyA1
“ i,j xσ ˚ pφ˚2 pb2 qq, eA A1
ř ˚
i,j yA1 xφ1 pb1 q, ei,j yA1
1
“ i,j xb2 , φ2 pσpeA A1

ř
i,j qqyB2 xb1 , φ1 pei,j qyB1
1
“ i,j xb1 b b2 , φ1 peA A1

ř
i,j q b φ2 pσpei,j qqyB1 B2 ,
1
where teA
i,j u denotes the matrix units of MA . Therefore, we have
φ1 peA A1
b φ2 qpeA A1
ř ř
Cφ2 ˝σ˝φ˚1 “ i,j i,j q b φ2 pσpei,j qq “
1
i,j pφ1 i,j b σpei,j qq,
1
which gives rise to the following:
Theorem 1.8.4 For linear maps φi : MAi Ñ MBi for i “ 1, 2 and σ : MA1 Ñ MA2 ,
we have the identity
Cφ2 ˝σ˝φ˚1 “ pφ1 b φ2 qpCσ q. (1.63)
Since AdU bV “ AdU b AdV , we see that the identity (1.60) is, in fact, a special
case of (1.63).
References: [80]
61
1.8.3 Roles of ampliation
We are going to describe the dual cones in terms of ampliation idA b φ or φ b idB .
To do this, we consider the following diagrams
id φ
MA ÝÝÝAÑ MA MA ÝÝÝÑ MB
§ §
§ ˚ §ψ
đσ đ
φ B id
MA ÝÝÝÑ MB MB ÝÝÝÑ MB
which will give us formulae involving idA b φ and φ b idB . Indeed, we have
xφ, ψ ˝ σy “ xφ ˝ σ ˚ , ψy “ xCφ˝σ˚ , Cψ y “ xpidA b φqpCσ˚ q, Cψ y,

xφ, σ ˝ ψy “ xψ ˝ φ˚ , σ ˚ y “ xCψ˝φ˚ , Cσ˚ y “ xpφ b idB qpCψ q, Cσ˚ y.
From the first line, we see that φ P pK ˝ CPAA q˝ if and only if the ampliation idA b φ
sends pMA b MA q` to CK ˝ . We also see that φ P pCPBB ˝ Kq˝ if and only if φ b idB
sends CK to pMB b MB q` from the second line. Therefore, we have the following:
Theorem 1.8.5 For a closed convex cone K Ă HpMA , MB q of positive maps, the
following are equivalent:
(ii) φ P K ˝ if and only if idA b φ sends positive matrices to CK ˝ .
The following are also equivalent:
(iv) φ P K ˝ if and only if φ b idB sends CK to positive matrices.
When K ˝ is given by the mapping cone SP1 , the statement (ii) of Theorem 1.8.5
tells us that φ P SP1 rMA , MB s if and only if idA b φ sends every state in MA b MB to
a separable state. In this context, a 1-superpositive map is also called entanglement
breaking. On the other hands, the statement (iv) with K ˝ “ DEC and K ˝ “ Pk
gives rise to following:
Corollary 1.8.6 For φ : MA Ñ MB , we have the following:
(i) φ is decomposable if and only if φ b idB sends PPT states to positive matrices,
(ii) φ is k-positive if and only if φ b idB sends states with Schmidt numbers ď k
to positive matrices.
62
Every vector |ζy “ ì“1 |ξi y|ηi y P CA b CA has Schmidt number ď k whenever
ř
` ď k. Therefore, we see that φ is k-positive if and only if

`
ÿ
pφ b idA qp|ζyxζ|q “ φp|ξi yxξj |q b |ηi yxηj |
i,j“1
is positive whenever ` ď k if and only if ì,j“1 |ηi yxηj | b φp|ξi yxξj |q is positive when-
ř
ever ` ď k. Compare with the conditions (v) and (vi) of Proposition 1.6.1.
The dual cone of a mapping cone can be also explained in terms of Choi matrices
of maps in the dual cone. To do this, we consider the following diagrams
id ψ
MA ÝÝÝAÑ MA MA ÝÝÝÑ MB
§ §
§φ §φ
đ đ
ψ˚ B id
MB ÝÝÝÑ MA MB ÝÝÝÑ MB
to get the identities
xφ, ψ ˝ σy “ xψ ˚ ˝ φ, σy “ xpidA b ψ ˚ qpCφ q, Cσ y,

xφ, σ ˝ ψy “ xφ ˝ ψ ˚ , σy “ xpψ b idB qpCφ q, Cσ y.
Therefore, we see that φ P pK ˝ CPAA q˝ if and only if idA b ψ ˚ sends Cφ to a positive

matrix for every ψ P K, and φ P pCPBB ˝ Kq˝ if and only if ψ b idB sends Cφ to a
positive matrix for every ψ P K.
Theorem 1.8.7 For a closed convex cone K Ă HpMA , MB q of positive maps, the
following are equivalent:
(ii) φ P K ˝ if and only if pidA b ψ ˚ qpCφ q ě 0 for every ψ P K.
Furthermore, the following are also equivalent:
(iv) φ P K ˝ if and only if pψ b idB qpCφ q ě 0 for every ψ P K.
When we take K “ SP˝k , this gives us the following characterization of Schmidt

numbers of states:
Corollary 1.8.8 A state % P MA b MB belongs to Sk if and only if pidA b ψqp%q ě 0

for every k-positive map ψ : MB Ñ MA if and only if pψ b idB qp%q ě 0 for every
k-positive map ψ : MA Ñ MB .
63
The most important case occurs when k “ 1; a state % is separable if and only
if pψ b idB qp%q ě 0 for every positive map ψ : MA Ñ MB . If we take ψ “ T, the
transpose map, then we have the PPT criterion; if % is separable then pTbidqp%q ě 0.
We see that x%, Ty ě 0 if and only if x%Γ , Cid y ě 0. This condition tells us that the
sum of some coefficients of %Γ is nonnegative. Therefore, we see that the condition
pψ b idB qp%q ě 0 is much stronger than x%, ψy ě 0 to detect entanglement.
References: [70], [113], [114], [60], [34], [124], [27], [36], [80],
1.8.4 Tensor products of positive maps

The dual cone may be also described in terms of state Cid with the maximal Schmidt
number. For this purpose, we consider the following diagrams
φ φ˚
MA ÝÝÝÑ MB MB ÝÝÝÑ MA
§ §
§id §id
đ A đ B
ψ ψ˚
MA ÝÝÝÑ MB MB ÝÝÝÑ MA
together with the identity
xφ, σ ˝ ψy “ xφ˚ , ψ ˚ ˝ σ ˚ y “ xψ ˝ φ˚ , σ ˚ y “ xpφ b ψqpCidA q, Cσ˚ y,
xφ, ψ ˝ σy “ xψ ˚ ˝ φ, σy “ xpφ˚ b ψ ˚ qpCidB q, Cσ y,
to get the following:
Theorem 1.8.9 Suppose that K is a closed convex cone of positive maps in HpMA , MB q.
Then the following are equivalent:
(i) K is a left mapping cone,
(ii) φ P K ˝ if and only if pφ b ψqpCidA q ě 0 for every ψ P K.

Furthermore, the following are also equivalent:
(iii) K is a right mapping cone,
(iv) φ P K ˝ if and only if pφ˚ b ψ ˚ qpCidB q ě 0 for every ψ P K.
We note that the dual cones are described in terms of images of Cid under the
tensor product φ b ψ in Theorem 1.8.9. Finally, we explore properties of φ b ψ
themselves in terms of dual cones. To do this, we put arbitrary τ in the places of
idA to get
φ
MA ÝÝÝÑ MB
§
§τ
đ
ψ
MA ÝÝÝÑ MB
64
together with the identity
xσ ˚ ˝ ψ ˝ τ, φy “ xψ ˝ τ, σ ˝ φy “ xψ ˝ τ ˝ φ˚ , σy “ xpφ b ψqpCτ q, Cσ y.
Then we see that φ P pCPBB ˝K ˝CPAA q˝ if and only if φbψ sends positive matrices
to positive matrices for every ψ P K, that is, φbψ is a positive map for every ψ P K.
Therefore, we have the following:
Theorem 1.8.10 For a given closed convex cone K Ă HpMA , MB q of positive

maps, the following are equivalent:
(i) K is a mapping cone,
(ii) φ P K ˝ if and only if φ b ψ is a positive map for every ψ P K.
When K “ SPk , Theorem 1.8.10 tells us that ψ b φ is positive for every k-

superpositive map ψ if and only if idk b φ is positive. Noting that idk is a typical
example of SPk , one may ask if idk in the definition of k-positivity may be replaced
by another k-superpositive map. When we fix a matrix s with rank k, it is easily
seen that the map idk b φ is positive if and only if Ads bφ is positive, using singular
value decomposition of s.
We also put two identity maps in the diagram as follows:
id
MA ÝÝÝAÑ MA
§
§id
đ A
φ
MA ÝÝÝÑ MB
Then we get the identity pidA b φqpCidA q “ Cφ , and have the following:
Proposition 1.8.11 A linear map φ : MA Ñ MB belongs to a convex cone K if

and only if pidA b φqpCidA q belongs to CK .
1.8.5 Entanglement breaking maps

It is worthwhile to collect applications of results in this section for K ˝ “ SPk to
get equivalent conditions for k-superpositivity of a map φ, or equivalently Schmidt
numbers of the state Cφ .
Corollary 1.8.12 For a linear map φ : MA Ñ MB , the following are equivalent:

ř
(i) φ is k-superpositive, that is, φ “ i Adsi with rank si ď k,
65
(ii) Cφ belongs to Sk , that is, has the Schmidt numbers ď k,
(iii) ψ ˚ ˝ φ is completely positive for every k-positive map ψ : MA Ñ MB ,
(iv) φ ˝ ψ ˚ is completely positive for every k-positive map ψ : MA Ñ MB ,
(v) idA b φ sends every state into Sk ,
(vi) φ b idB sends every matrix in BP k to a positive matrix,
(vii) pidA b ψqpCφ q ě 0 for every k-positive map ψ : MB Ñ MA ,
(viii) pψ b idB qpCφ q ě 0 for every k-positive map ψ : MA Ñ MB ,
(ix) pφ b ψqpCidA q ě 0 for every k-positive map ψ : MA Ñ MB ,
(x) pφ˚ b ψ ˚ qpCidB q ě 0 for every k-positive map ψ : MA Ñ MB ,
(xi) φ b ψ is positive for every k-positive map ψ : MA Ñ MB .
(xii) pidA b φqpCidA q belongs to Sk .
If s “ |ξyxη| is of rank one, then we have
¯ ξ|y
Ads paq “ |ηyxξ|a|ξyxη| “ xξ|a|ξy |ηyxη| “ xa, |ξyx ¯ |ηyxη|.
Therefore, we see that φ is 1-superpositive if and only if φ is of the following form

ÿ
φpaq “ xa, vk yuk (1.64)
k
with positive matrices uk and vk . This is called the Holevo form.

References: [114], [56], [61], [3], [27], [36], [80]
1.9 Historical remarks

Positive linear maps have played crucial roles in the theory of operator algebras
since the Gelfand–Naimark–Segal construction in the 1940’s, by which positive linear
functionals give rise to ˚-representations of abstract C ˚ -algebras. This construction
was extended for completely positive linear maps into operators in the 1955 paper
by Stinespring [110], where a linear map φ was defined to be completely positive
when idk b φ is positive for every k “ 1, 2, . . . . Note that a linear functional on a
matrix algebra into scalars is completely positive if and only if it is just positive by
Theorem 1.5.3. See also [125] for more general situations. Stinespring representation
66
theorem tells us that every completely positive map is of the form V ˚ πp ¨ qV for a
˚-homomorphism π and a bounded linear map V . See [92].
Størmer defined the notion of decomposability in his paper [111] of 1963 by maps
of the form V ˚ πp ¨ qV with Jordan homomorphisms π which preserve Hermiticity and
square of Hermitian elements. It was shown later in [112] that a map is decomposable
in this sense if and only if it is the sum of a completely positive map and a completely
copositive map. In the paper [111], extremeness of the map Ads in Theorem 1.2.4
had been considered in more general situations, and all the extreme points of the
convex set P1 rM2 , M2 , Is of all unital positive maps between 2ˆ2 matrices have been
found explicitly, from which Theorem 1.3.4 follows. We followed [118] and [6, 7] for
the proofs of Theorem 1.2.4 and Theorem 1.3.4, respectively. The maps Ads also
play important roles in matrix theory. Suppose that a map φ : Mn Ñ Mn satisfies
φp0q “ 0 and (1.12). If we suppose further that φ is bijective or continuous then it
is known [109, 87, 104] that φ “ Ads or φ “ Ads ˝T for a nonsingular s P Mn .
Choi matrices have been introduced by de Pillis [32] in 1967, where Proposition
1.4.1 was shown. Propositions 1.4.2 and Theorem 1.5.3 were given by Jamiolkowski
[67] in 1972 and Choi [21] in 1975, respectively, after whom the one-to-one cor-
respondence φ ÞÑ Cφ is now called the Jamiolkowski–Choi isomorphism. Kraus
decomposition of completely positive maps were found independently [72, 73]. We
ř
note that the matrix i,j |jyxi| b φp|iyxj|q has been considered in [32] and [67],
ř
instead of Cφ “ i,j |iyxj| b φp|iyxj|q. Recall that the notions of Hermitcity and
1-block-positivity are invariant under taking partial transposes. Further variants
has been considered in [93, 80, 79, 54]. The Choi matrix can be defined for infi-
nite dimensional cases [57, 58, 84, 119, 35, 50, 85, 37] in various situations, and
multi-linear maps between matrix algebras [78, 52, 53]. Since Arveson’s extension
theorem for completely positive maps [4] and Choi’s Theorem 1.5.3 on the corre-
spondence between completely positivity of maps and positivity of Choi matrices,
it has been widely accepted by operator algebraists that completely positive maps
serve as morphisms for operator algebras reflecting noncommutative order struc-
tures, and important notions like nuclearity and injectivity have been described in
terms of completely positive maps. See the survey article [33]. Theorem 1.5.8 on
the extreme points of unital completely positive maps is also taken from [21].
The notion of duality between tensor products and linear maps in (1.39) goes
back to the work of Woronowicz [129] in 1976, where he showed that every positive
maps between M2 and M3 are decomposable. For this purpose, he utilized the
duality to get the equivalent claim that for any PPT state % P M2 b M3 there
¯
exists a product vector |ξy|ηy P Im % such that |ξy|ηy P Im %Γ . This is equivalent in
principle to the claim that every 2 b 3 PPT state in M2 b M3 is separable, as it was
noticed later by Horodecki’s in the paper [60] of 1996, where the duality between
67
S1 and P1 was also given. Theorem 1.5.4 (i) on the duality between k-positivity
and states of Schmidt number ď k was obtained in [34]. The duality between linear
maps through (1.40) was defined in [107]. Duality between separability of states and
positivity of maps can be extended for infinite dimensional cases [115].
The linear map τn,n´1 in (1.52) with k “ n ´ 1, which was given by the 1972
paper of Choi [20], is the first concrete example of a linear map which distinguishes
complete positivity and k-positivity. The equivalent condition (iv) of Proposition
1.6.1 to the k-positivity is implicit in [20] and stated in [120] clearly. The condition
(vi) of Proposition 1.6.1 together with Theorem 1.6.3 is due to Tomiyama [126].
After the first example of indecomposable positive map in [22], Choi and Lam [25]
constructed the Choi map φch in (1.54), which is known to be extreme [42]. The
linear map from M2 into M4 given by (1.56) was constructed by Woronowicz [130] in
the contexts of non-extendibility which has a close relationship with extremeness of
positive maps. There are other examples [122] of indecomposable maps between M2
and M4 . See also [112] for infinite family of indecomposable positive maps in other
dimensions. Further interesting properties of two maps by Choi and Woronowicz
will be discussed in the next chapter.
The notion of entanglement had been originated from the work by Einstein,
Podolsky and Rosen in the 1930’s. It was defined for general mixed states by Werner
[128] in 1989 under the name of Einstein-Podolsky-Rosen correlation. Separable
states, which were defined as convex combinations of product states, were called
to be classically correlated in the paper. The notion of Schmidt numbers of bi-
partite states has been considered in [34] under the name k-simple vectors following
Woronowicz’ terminology [129], together with Theorem 1.5.2. The term Schmidt
numbers were introduced in [124], where Corollary 1.8.8 was shown. It was shown
in [40] that every state in the unit ball around the identity Imn P Mm b Mn with
respect to the norm } ¨ }HS is separable. It is very hard in general to prove that
a given state is entangled, since we have to show that some kinds of expressions
are not possible. In fact, it is now known [39] that it is an N P hard problem in
general to decide if a given state is separable or entangled. Positive maps play
crucial roles in this problem by duality: A state % is entangled if and only if there
exists a positive map φ such that x%, φy ă 0. This will be the main topic of the
next chapter. Theorem 1.4.3, which is now called the PPT criterion, was presented
by Choi [24] in 1980 together with the example (1.33) of a PPT entangled state,
and rediscovered by Peres [94] in 1996. We note that the first example of a PPT
entangled state was given by Woronowicz [129] in 1976 among 2 b 4 states in the
context of decomposability of positive maps. For more criteria for separability, we
refer to survey articles [38, 64]
The Werner states had been introduced by Werner [128] together with Proposi-
68
tion 1.7.5 and the technique using the integration over the unitary group to show
that a given state is separable. It is another problem to look for decomposition
into finitely many product states. See [91]. On the other hands, Horodecki’s [59]
introduced the isotropic states and gave the condition for separability using the cor-
responding invariant properties in Proposition 1.7.1. Theorem 1.7.3 on the Schmidt
numbers of isotropic states was obtained in [124]. The k-copositivity of the map φλ ,
or equivalently k-positivity of ψλ in Theorem 1.7.6 was obtained in [126] by another
application of Proposition 1.6.1.
The notion of superpositive maps has arisen more recently, even though positive
maps of the form (1.64) had been considered by Størmer [114] in 1980’s. In the paper
[61], a map φ was called entanglement breaking when φ satisfies the the condition
(v) of Corollary 1.8.12 with k “ 1, and it was shown that conditions (i), (iii), (iv)
and (xii) are equivalent to (v) in Corollary 1.8.12 together with the Holevo form
[56] in (1.64). On the other hand, a map was called superpositive in [3] when its
Choi matrix is separable. See [5] for the characteristic property of unital completely
positive maps which is not entanglement breaking. Equivalent conditions (ii), (iii)
and (v) of Corollary 1.8.12 for k-superpositive maps were given in [27]. We follow
[106, 108] for the definition of k-superpositivity. We note that Theorem 1.5.2 on the
dual of k-superpositivity recovers the definition of k-positivity.
Mapping cones were introduced by Størmer [114] in the 1980’s in more general
contexts to study the extension problem of positive maps. It was known [107, 117]
that the dual cones of mapping cones can be described in terms of compositions and
tensor products; φ P K ˝ if and only if φ ˝ ψ is completely positive for every ψ P K if
and only if ψ b φ is positive for every ψ P K if and only if pψ b φqpCid q ě 0 for every
for every ψ P K. One-sided mapping cone was introduced quite recently in [36] to see
that some of the above equivalences hold only when K is a one-sided mapping cone.
Results in Section 1.8, which were taken from [36] and [80], recover many known
results in a systematic way. For examples, statements (i) and (ii) of Corollary 1.8.6
were given in [70, 113] and [34], respectively. On the other hand, Corollary 1.8.8 with
k “ 1 gives rise to the Horodecki’s separability criterion [60]; a state % is separable
if and only if pψ b idB qp%q ě 0 for every positive map ψ : MA Ñ MB . The analogous
characterization of states in Sk in terms of k-positive maps is due to [124]. See also
[115] for an infinite dimensional analogue for Horodecki’s separability criterion.
69
70
Chapter 2
Detecting Entanglement by
Positive Maps
We will focus in this chapter on the duality between positive maps and separable
states: A bi-partite state % P Mm b Mn is separable if and only if x%, φy ě 0 holds
for every positive map φ from Mm into Mn . It is clear that this is the case if and
only if the inequality holds for every positive map which generates an extreme ray
of the convex cone P1 of all positive maps. But, it is not so easy to find extremal
positive maps, because we do not know facial structures of the convex cone P1 . In
fact, it is enough to check the inequality for positive maps which generate exposed
rays of P1 , by Straszewicz’s Theorem which tells us that the set of exposed points
of a compact convex set is dense among all extreme points.
We will also consider the question how much entanglement a given positive map
may detect. A positive map will be called optimal when it detects a maximal set of
entanglement. It turns out that the notion of optimality involves facial structures of
P1 again; a positive map is optimal if and only if the smallest face containing it has
no completely positive map. We give a sufficient condition on positive maps which
detect maximal set of entanglement, in terms of exposed faces. This condition,
called the spanning property, is much easier to check, and this explains why we
are interested in exposed faces. We also note that the boundary of a convex cone
consists of maximal faces which are always exposed.
A face is exposed with respect to the bilinear pairing if and only if it is a dual face,
which is determined by a hyperplane given by an element of the dual cone. In order
to detect entangled states with positive partial transposes, we need indecomposable
positive maps, and so it is important to find indecomposable exposed positive maps,
in this contexts. We will exhibit such positive maps in low dimensional cases.
We will see in Section 2.1 that a face F is exposed in the convex if and only if it
coincides with the bidual face F 2 . We also see that every face for completely positive
maps turns out to be exposed, and this is also the case for PPT states. Especially,
71
we show that the map Ads is exposed in the convex cone P1 consisting of all positive
maps. In Section 2.2, we show that the Choi map φch has very special properties; the
smallest face containing φch is not exposed; the smallest exposed face φ2ch containing
φch has a nontrivial intersection with completely positive maps, but has the trivial
intersection with completely copositive maps. We will show in Section 2.3 that a
face is maximal if and only if its dual face is minimal among exposed faces. With
this, we can understand relative locations of convex cones Pk of all k-positive maps.
Section 2.4 will be devoted to explain optimality and spanning properties of positive
maps, and explain them in terms of faces and exposed faces, respectively.
In the remaining part, we will exhibit several examples of indecomposable pos-
itive maps in low dimensions. In Section 2.5, we exhibit parameterized examples
of indecomposable positive maps between 3 ˆ 3 matrices, which are variants of the
Choi map. We will find positive maps among them which distinguish various kinds
of optimality. We also provide in Section 2.6 a sufficient condition for exposedness
of a positive map which is not so difficult to check, and we show that the Woronow-
icz map from M2 into M4 is exposed. We will work with parameterized examples.
Finally, we exhibit in Section 2.7 an example of indecomposable exposed positive
map between 4 ˆ 4 matrices, which was constructed by Robertson.
2.1 Exposed faces

A face is exposed if and only if it is a dual face. We will see that every face for
completely positive maps is exposed. This is also the case for PPT states. We also
show that the map Ads generates an exposed ray of the convex cone of all positive
maps.
2.1.1 Dual faces

Suppose that X and Y are finite dimensional vector spaces with a non-degenerate
bilinear pairing between them. For a subset F of a closed convex cone C of X, we
define the subset F 1 of C ˝ by
F 1 “ ty P C ˝ : xx, yy “ 0 for each x P F u Ă C ˝ Ă Y.
It is then clear that F 1 is a face of C ˝ , which is said to be the dual face of F . For
an arbitrary subset F of X, we have F Ă F 2 , from which we also have F 3 Ă F 1 .
Therefore, we have F 1 “ pF 2 q1 in general, and we see that every dual face is the dual
face of a face. If F is a face with an interior point x0 then it is also easy to see that
F 1 “ ty P C ˝ : xx0 , yy “ 0u.
72
When tx0 u is a singleton, tx0 u1 is denoted just by x10 .
We say that a face F of a closed convex cone C in X is an exposed face if it is a
dual face of a subset in Y . This means that the face is determined by the hyperplane
tx P X : xx, y0 y “ 0u in X given by a point y0 P C ˝ . In other words, we have
F “ C X tx P X : xx, y0 y “ 0u.
Proposition 2.1.1 For a subset F of a convex cone C in X, we have the following:
(i) F 2 is the smallest exposed face containing F ,
(ii) F is an exposed face if and only if F “ F 2 .
Proof. Suppose that an exposed face S 1 satisfies the relation F Ă S 1 Ă F 2 . Then

we have F 3 Ă S 2 Ă F 1 , and so F 1 “ S 2 and S 1 “ F 2 . This proves (i). The second
statement (ii) follows form (i). ˝
We recall that every face of the closed convex cone CPrMm , Mn s is of the form
FV “ conv tAds : s P V u
for a subspace V of Mmˆn . By the relation

¯
xAds , Adt y “ x|s̃yxs̃|, |t̃yxt̃|y “ xt̃|s̃yxs̃|t̃¯y “ |xs̃|t̃¯y|2 “ |xs, ty|2 , (2.1)
we see that pFV q2 “ FV , and we have the following:
Proposition 2.1.2 Every face of the convex cone CP is exposed.
We proceed to show that every face of the convex cone PPT of all PPT states
is also exposed. We begin with more general situations. Recall that we have the
relations (1.48);
pC1 ` C2 q˝ “ C1˝ X C2˝ , pC1 X C2 q˝ “ C1˝ ` C2˝ ,
for closed convex cones C1 and C2 . Suppose that Fi is a subset of Ci for i “ 1, 2.

Then it is easy to see that the following identity
pF1 ` F2 q1 “ F11 X F21 (2.2)
holds, where pF1 ` F2 q1 is a face of the convex cone pC1 ` C2 q˝ “ C1˝ X C2˝ and Fi1 is
a face of Ci˝ for i “ 1, 2.
Now, we suppose that Fi is a face of a convex set Ci for i “ 1, 2. Then it is clear
that F1 X F2 is a face of the convex set C1 X C2 whenever it is not empty. Conversely,
suppose that F is a face of C1 X C2 . Take an interior point x of F . We also take the
73
face Fi of Ci in which x is an interior point for i “ 1, 2. Then x is also an interior
point of F1 X F2 , which is a face of C1 X C2 . Because x is an interior point of both
faces F and F1 X F2 of C1 X C2 , we conclude that F “ F1 X F2 . In short, we have
F “ F1 X F2 , int F Ă int F1 , int F Ă int F2 . (2.3)
It is clear that faces Fi of Ci satisfying (2.3) are uniquely determined.
Proposition 2.1.3 Suppose that Ci is a closed convex cone in a finite dimensional

space, for i “ 1, 2. For every face F of C1 X C2 , there exist unique faces F1 and F2
of C1 and C2 , respectively, satisfying (2.3). In this case, we have
F 1 “ F11 ` F21 , F 2 “ F12 X F22 .
Proof. The inclusion F11 ` F21 Ă F 1 follows from Fi1 Ă F 1 . For the reverse inclusion,
take y P F 1 . Since y P pC1 X C2 q˝ “ C1˝ ` C2˝ , we may write y “ y1 ` y2 with yi P Ci˝
for i “ 1, 2. If we take an interior point x of F1 X F2 , then x P int Fi Ă Ci by (2.3),
and xx, yi y ě 0 for i “ 1, 2. By the relation
0 “ xx, yy “ xx, y1 y ` xx, y2 y,
we see that xx, yi y “ 0 and yi P Fi1 , since x is an interior point of Fi . Therefore, we

have y P F11 ` F21 . This proves the first identity. The second identity follows from
(2.2). ˝
Because every face of CP is of the form FV “ tAds : s P V u2 for a subspace V
of Mmˆn , we see that every face of CCP is of the form
F V :“ tAds ˝T : s P V u2 .
Therefore, every face of PPT “ CP X CCP must be of the form
FV X F W , (2.4)
for a pair pV, W q of subspaces of Mmˆn . Note that the intersection of two exposed
faces is exposed again by (2.2). By Proposition 2.1.2 and Proposition 2.1.3, we have
the following:
Proposition 2.1.4 Every face of the convex cone PPT is exposed.
Every face of pMm b Mn q` is also of the form
FV “ t% P pMm b Mn q` : Im % Ă V u (2.5)
74
for a subspace of Cm b Cn , with the same notation as in (1.46), and so every face
of the convex cone PPT of all PPT states is also of the form
FV X F W ,
for a pair pV, W q of subspaces Cm b Cn , with
F W “ t%Γ : % P FW u. (2.6)
It is not so easy in general to determine if FV X F W is nontrivial or not. If we denote

by V the subspace of C3 b C3 spanned by vectors in (1.34), then we see that the
PPT state % defined by (1.33) belongs to the face FV X F V .
As for the convex cone P1 of all positive maps, not every face is exposed even
in the low dimensional cases like P1 rM2 , M2 s “ DECrM2 , M2 s. See [13]. It is not
known whether every face of the convex cone S1 is exposed or not.
References: [76], [34], [13], [43]
2.1.2 Exposed positive maps

A ray ttx0 : t ě 0u of a closed convex cone C generated by x0 P C is called an
exposed ray if it is an exposed face. In this case, we say that x0 is exposed in C.
We proceed to show that an extreme ray Ads is exposed in P1 as well as in CP for
every s P Mmˆn . Recall that the dual cone of P1 is SP1 , and so φ P P1 belongs to
pAdσ q2 if and only if the following
s P Mmˆn , rank s “ 1, xAdσ , Ads y “ 0 ùñ xφ, Ads y “ 0 (2.7)
holds. In order to show that Adσ generates an exposed ray of the convex cone P1 ,
we have to show that the condition (2.7) implies that φ is a scalar multiple of Adσ
by Proposition 2.1.1 (ii).
In the following discussion, we identify Mm b Mn with Mm pMn q “ Mmn , and
the entries of Cφ P Mmn will be denoted by cpi,kq,pj,`q with i, j “ 1, . . . , m and
k, ` “ 1, . . . , n, where pi, kq’s and pj, `q’s are endowed with the lexicographic orders.
Therefore,
cpi,kq,pj,`q “ x|ikyxj`|, Cφ y “ x|iyxj| b |kyx`|, Cφ y “ x|kyx`|, φp|iyxj|qy
is the pk, `q entry of the pi, jq block of Cφ P Mm pMn q.

By singular value decomposition, every m ˆ n matrix is of the form uσv ˚ with
invertible matrices u P Mm , v P Mn and
r
ÿ
σ“ |iyxi| P Mmˆn . (2.8)
i“1
75
Note that Aduσv˚ “ Adv˚ ˝ Adσ ˝ Adu , and
φ ÞÑ Adv˚ ˝φ ˝ Adu
is an affine isomorphism between HpMm , Mn q. Therefore, we see that Aduσv˚ is

exposed in P1 rMm , Mn s if and only if Adσ is exposed in P1 rMm , Mn s. We note that
the entries of CAdσ are 0 or 1, and the pi, kq, pj, `q entry is 1 if and only if i “ k and
j “ ` with i, j “ 1, . . . , r.
From now on, we will prove that Adσ is exposed in P1 rMm , Mn s when σ is given
by (2.8). For this purpose, we suppose that φ P pAdσ q2 . We first consider the m ˆ n
matrix s “ |iyxk| in (2.7), to see that the diagonal entries of Cφ are given by
cpi,kq,pi,kq “ 0, pi, kq P t1, . . . , mu ˆ t1, . . . nuztp1, 1q, . . . , pr, rqu. (2.9)
When J is a subset of t1, 2, . . . , mu ˆ t1, 2, . . . , nu, we denote by AJ the principal

submatrix of A by taking pi, kq rows and columns for pi, kq P J. In order to determine
off-diagonal entries cpi,kq,pj,`q of Cφ with pi, kq ‰ pj, `q, we consider
J “ tpi, kq, pi, `q, pj, kq, pj, `qu, (2.10)
and the principal submatrix of rCφ sJ of Cφ . If i “ j or k “ ` then rCφ sJ is a

positive 2 ˆ 2 matrix and so cpi,kq,pj,`q “ 0 by considering the diagonal entries in
(2.9). Otherwise, rCφ sJ is a block matrix in M2 b M2 “ M2 pM2 q. We proceed to
show that this is a block-positive matrix.
Suppose that m “ 2, 3, . . . and i, j “ 1, 2, . . . , m with i ‰ j. We consider the
linear map λmi,j : M2 Ñ Mm defined by
ˆ ˙
a11 a12
λm
i,j : ÞÑ a11 |iyxi| ` a12 |iyxj| ` a21 |jyxi| ` a22 |jyxj| P Mm .
a21 a22
Then it is easily seen that the adjoint map pλm ˚

k,` q : Mm Ñ M2 is given by
m ˆ ˙
ÿ akk ak`
pλm
k,` q
˚
: aij |iyxj| ÞÑ P M2 .
a`k a``
i,j“1
If φ : Mm Ñ Mn then pλnk,` q˚ ˝ φ ˝ λm
i,j is a map from M2 into M2 , whose Choi matrix
is given by just rCφ sJ with J in (2.10). Therefore, we see that if Cφ is block-positive
then rCφ sJ is also block-positive in M2 b M2 .
Lemma 2.1.5 Suppose that % P BP 1 rM2 b M2 s has at most one nonzero diagonal
entry. Then all the other entries of % are zero.
76
Proof. We first consider the case when all the diagonal entries are zero except
for the left upper corner. Then by the definition of block-positivity, the 2 ˆ 2
principal submatrices %tp1,1q,p1,2qu , %tp2,1q,p2,2qu , %tp1,1q,p2,1qu and %tp1,2q,p2,2qu are positive.
Therefore, we see that % is of the form
¨ ˛
a ¨ ¨ α
˚¨ ¨ β ¨‹
Cφ “ ˚ ˝ ¨ β̄ ¨ ¨ ‚ P M2 pM2 q,
‹
ᾱ ¨ ¨ ¨
with a ě 0 and α, β P C. Take xξ| “ p1, x, y, xyq P C2 b C2 whose Schmidt rank is
one. Then we have
0 ď x%, |ξyxξ|y “ a ` 2Re rypxα ` x̄βqs
for every y P C. This implies that xα ` x̄β “ 0 for every x P C, and so it follows
that α “ β “ 0. The remaining cases can be done in the same way. ˝
Therefore, we see that cpi,kq,pj,`q ‰ 0 only when J contains at least two nonzero
diagonal entries of rCφ sJ . By diagonal entries given in (2.9), this happens only for
cpi,iq,pj,jq and cpi,jq,pj,iq with i, j “ 1, 2, . . . , r and i ‰ j. In this case, J is given by
J “ tpi, iq, pi, jq, pj, iq, pj, jqu
and the corresponding principal submatrices of CAdσ and Cφ are of the forms
¨ ˛ ¨ ˛
1 ¨ ¨ 1 a ¨ ¨ α
˚¨ ¨ ¨ ¨‹ ˚¨ ¨ β ¨‹
Cid2 “ ˚
˝¨
‹ and % a,b,α,β “ ˝ ¨ β̄ ¨ ¨ ‚,
˚ ‹
¨ ¨ ¨‚
1 ¨ ¨ 1 ᾱ ¨ ¨ b
respectively, with a, b ě 0 and α, β P C. They are Choi matrices of the maps
pλni,j q˚ ˝ Adσ ˝λm n ˚ m
i,j and pλi,j q ˝ φ ˝ λi,j , respectively.
We proceed to show that pλnk,` q˚ ˝ φ ˝ λm n ˚ m 2
i,j belongs to ppλk,` q ˝ Adσ ˝λi,j q in
P1 rM2 , M2 s. To do this, we suppose that s P M2ˆ2 is of rank one matrix satisfying
the relation xpλnk,` q˚ ˝Adσ ˝λm n m ˚
i,j , Ads y “ 0. Then we have xAdσ , λk,` ˝Ads ˝pλi,j q y “ 0.
Since λnk,` is completely positive, we see that λnk,` ˝ Ads ˝pλm ˚
i,j q belongs to SP1 . In
fact, it is easily seen that λnk,` ˝ Ads ˝pλm ˚
i,j q “ Adŝ , with
ŝ “ s11 |iyxk| ` s12 |iyx`| ` s21 |jyxk| ` s22 |jyx`| P Mmˆn ,
which is of rank one. Therefore, we have
0 “ xλnk,` ˝ Ads ˝pλm ˚ n ˚ m

i,j q , φy “ xAds , pλk,` q ˝ φ ˝ λi,j y,
to see that pλnk,` q˚ ˝ φ ˝ λm n ˚ m 2

i,j belongs to ppλk,` q ˝ Adσ ˝λi,j q in P1 rM2 , M2 s. The
following two lemmas will show that the Choi matrix of pλni,j q˚ ˝ φ ˝ λm i,j must be a
scalar multiple of Cid2 .
77
Lemma 2.1.6 Suppose that a, b ě 0 and α, β P C. Then %a,b,α,β P M2 pM2 q is
?
block-positive if and only if |α| ` |β| ď ab.
Proof. Suppose that %a,b,α,β is block-positive, and take xξ| “ pp, peiτ , qeipθ´τ q , qeiθ q
in C2 b C2 with real numbers p, q. Note that SR xξ| “ 1. Then we have
“ ‰
0 ď x%a,b,α,β , |ξyxξ|y “ p2 a ` q 2 b ` 2pqRe eiθ pα ` βe´2iτ q
for every p, q P R. Therefore, we have
ˇRe e pα ` βe´2iτ q ˇ2 ď ab
ˇ “ iθ ‰ˇ
for every θ and τ , which implies the condition. For the converse, we put
|α|a |β|a |α|b |β|b

a1 “ , a2 “ , b1 “ , b2 “ .
|α| ` |β| |α| ` |β| |α| ` |β| |α| ` |β|
Then we see that %a,b,α,β “ %a1 ,b1 ,α,0 ` %a2 ,b2 ,0,β is the Choi matrix of a decomposable
map. ˝
Lemma 2.1.7 The identity map id2 on M2 is exposed in P1 rM2 , M2 s.
Proof. Suppose that φ P id22 . Taking s “ |iyxj| P M2ˆ2 ˆ with i ‰iθj˙in (2.7), we see
1 e
that Cφ must be of the form %a,b,α,β . We also take s “ in (2.7), to see
éíθ ´1
that
0 “ x|s̃yxs̃|, %a,b,α,β y “ a ` b ´ 2Re pα ` βe´2iθ q,
for every θ. Therefore, we have β “ 0, and we also have
?
0 “ a ` b ´ 2Re α ě a ` b ´ 2|α| ě 2 ab ´ 2|α| ě 0,
by Lemma 2.1.6. This happens only when a “ b “ α. ˝

Now, we return to the discussion before Lemma 2.1.6, to see that the Choi matrix
of pλni,j q˚ ˝ φ ˝ λm
i,j must be a scalar multiple of Cid2 , and it is clear that all the choices
i, j share a common scalar multiple. Therefore, we see that Cφ is a scalar multiple
of CAdσ , and conclude that Adσ is exposed. Using singular value decomposition
s “ uσv ˚ , we have the following theorem which provides another proof of Theorem
1.2.4, because an exposed positive map must be extremal.
Theorem 2.1.8 For every s P Mmˆn , the map Ads is exposed in P1 rMm , Mn s.
References: [134], [86], [53], [81]
78
2.2 The Choi map revisited
In this section, we consider the bidual of the Choi map φch , which is the smallest
exposed face of P1 containing φch . Especially, we determine completely positive
maps and decomposable maps which belong to this bidual.
2.2.1 The Choi map and completely positive maps

In order to calculate the dual face of a positive map with respect to the duality
between SP1 and P1 , we take an extreme ray Ads in SP1 with s “ |ξyxη| P Mmˆn . If
we write |ξy “ m
ř řn ř
i“1 ξ i |iy and |ηy “ j“1 ηj |jy then we have s “ i,j ξi η̄j |iyxj|, and
ř
so we have xs̃| “ i,j ξi η̄j xi|xj| and
ÿ
CAd|ξyxη| “ |s̃yxs̃| “ ξ¯i ηj ξk η̄` |iy|jyxk|x`|
i,j,k,`
˜ ¸ ˜ ¸
ÿ ÿ
“ ξ¯i ξk |iyxk| b ηj η̄` |jyx`| ¯ ξ|
“ |ξyx ¯ b |ηyxη|.
i,k j,`
Therefore, we have
¯ ξ|
xAds , φy “ x|ξyx ¯ b |ηyxη|, φy “ x|ηyxη|, φp|ξyx
¯ ξ|qy
¯ “ xη̄|φp|ξyx
¯ ξ|q|η̄y.
¯ (2.11)
¯ such that φp|ξyx
We first find all |ξy’s ¯ ξ|q
¯ is singular, and find the kernel |η̄y of
¯ ξ|q,
φp|ξyx ¯ to get extreme rays Ads in the dual face φ1 in SP1 .
We apply the above method to the Choi map φch : M3 Ñ M3 defined in (1.54).
¯ “ px, y, zqT P C3 , we calculate to get
For given |ξy
¯ ξ|q
det φch p|ξyx ¯ “ |x|4 |y|2 ` |y|4 |z|2 ` |z|4 |x|2 ´ 3|xyz|2 .
This is nonnegative by arithmetic-geometric inequality, and this may give rise to

another proof for positivity of the Choi map. It vanishes when and only when two
¯ ξ|q
of variables are zero or |x| “ |y| “ |z|. Therefore, φch p|ξyx ¯ is singular if and only
¯ is one of the following:
if |ξy
|ξ¯1 y “ p1, 0, 0qT , |ξ¯2 y “ p0, 1, 0qT , |ξ¯3 y “ p0, 0, 1qT , |ξ¯4 y “ peia , eib , eic qT ,
¯ ξ|q
and the corresponding kernel vectors of φch p|ξyx ¯ are given by
|η̄1 y “ p0, 0, 1qT , |η̄2 y “ p1, 0, 0qT , |η̄3 y “ p0, 1, 0qT , |η̄4 y “ peia , eib , eic qT ,
respectively, up to scalar multiplications. Then the dual face φ1ch is the convex
cone generated by superpositive maps Ads ’s with following 3 ˆ 3 rank one matrices
s “ |ξyxη|:
¨ ˛ ¨ ˛ ¨ ˛ ¨ ˛
0 0 1 0 0 0 0 0 0 1 α γ
˝0 0 0‚, ˝1 0 0‚, ˝0 0 0‚, sα,β,γ “ ˝α 1 β ‚,
0 0 0 0 0 0 0 1 0 γ β 1
79
where αβγ “ 1 with |α| “ |β| “ |γ| “ 1. We denote
S “ te21 , e32 , e13 u Y tsα,β,γ : αβγ “ 1, |α| “ |β| “ |γ| “ 1u, (2.12)
with matrix units. Then the convex cone φ1ch is generated by tAds : s P Su.
We note that matrices in S belong to the 7-dimensional subspace
D “ traij s P M3 : a11 “ a22 “ a33 u. (2.13)
We also see that the following four matrices

¨ ˛ ¨ ˛ ¨ ˛ ¨ ˛
1 1 1 1 ´1 1 1 1 ´1 1 ´1 ´1
˝1 1 1‚, ˝´1 1 ´1‚, ˝ 1 1 ´1‚, ˝´1 1 1‚
1 1 1 1 ´1 1 ´1 ´1 1 ´1 1 1
together with e21 , e32 , e13 are linearly independent rank one matrices belonging to D.
Therefore, we conclude that S spans the 7-dimensional space D defined by (2.13).
We note that all maps in FD , which is a face of CP, do not belong to SP1 . We also
note φ1ch Ř FD X SP1 , because there are rank one matrices in D which do not belong
to S.
We proceed to show that the smallest face Fφch of P1 containing φch is strictly
smaller than the bidual face φ2ch of the Choi map φch P P1 rM3 , M3 s. From this, we
will conclude that Fφch is not exposed face. We see that Ads P φ2ch if and only if
xs, ty “ 0 for every t P S if and only if xs, ty “ 0 for every t P D if and only if s
belongs to the 2-dimensional subspace
E “ ts P M3ˆ3 : s is diagonal, Tr s “ 0u.
Therefore, we conclude that φ2ch X CP “ FE . We consider three matrices

¨ ˛ ¨ ˛ ¨ ˛
1 ¨ ¨ ¨ ¨ ¨ ´1 ¨ ¨
s1 “ ˝ ¨ ´1 ¨‚, s2 “ ˝¨ 1 ¨ ‚ s3 “ ˝ ¨ ¨ ¨ ‚
¨ ¨ ¨ ¨ ¨ ´1 ¨ ¨ 1
in E, and take
φ0 “ Ads1 ` Ads2 ` Ads3 P φ2ch . (2.14)
We put φt “ p1 ´ tqφ0 ` tφch then we have
¨ ˛
p3 ´ tqx11 ` tx33 ¨ ¨
φt pxq “ ˝ ¨ p3 ´ tqx22 ` tx11 ¨ ‚´ rxij s,
¨ ¨ p3 ´ tqx33 ` tx22
for x “ rxij s P M3 . By Proposition 1.6.5, we see that φt is a positive map if and
only if 0 ď t ď 3 and
α β γ
` ` ď1
p3 ´ tqα ` tγ p3 ´ tqβ ` tα p3 ´ tqγ ` tβ
80
1
for every positive α, β and γ. We take β “ α
and γ Ñ 0, to get a necessary condition
p2 ´ tqtα2 ` p3 ´ tqp1 ´ tq ě 0, αą0
for positivity of φt . Therefore, we see that φt is positive only when 0 ď t ď 1. This

tells us that φch is on the boundary of the convex cone φ2ch , and the smallest face
Fφch determined by φch is strictly smaller than φ2ch . Especially, we conclude that the
face Fφch of P1 is not exposed. One can also show that Fφch contains no completely
positive map. In fact, it is known [42] that the Choi map φch generates an extreme
ray of P1 . We summarize as follows:
Theorem 2.2.1 We define the subset S of M3ˆ3 by (2.12), and put
D “ span S, E “ tt P M3ˆ3 : xs, ty “ 0 for every s P Su.
The smallest exposed face φ2ch and the smallest face Fφch of P1 containing the Choi
map φch have the following properties;
(i) rank s “ 1 and Ads P φ1ch if and only if s P S,
(ii) φ1ch is generated by tAds : s P Su, and φ1ch Ř FD X SP1 ,
(iii) Fφch Ř φ2ch and φ2ch X CP “ FE .
(iv) Fφch has no nontrivial completely positive map.
Proof. It remains to prove (iv). If a completely positive map belongs to Fφch , then
there is s P E such that Ads P Fφch . Since φch is an interior point of Fφch , there is
λ ą 1 such that p1 ´ λq Ads `λφch is a positive map. Multiplying a scalar, we may
assume that φch ´ Ads is a positive map. Comparing the 2 ˆ 2 principal submatrices
of the inequality φch p|xyxx|q ě Ads p|xyxx|q, we have
ˆ 2 ˙ ˆ ˙
|x1 | ` |x3 |2 ´x1 x̄2 |s1 |2 |x1 |2 ´s̄1 s2 x1 x̄2
ě
´x2 x̄1 |x2 |2 ` |x1 |2 ´s̄2 s1 x2 x̄1 |s2 |2 |x2 |2
for every x1 , x2 , x3 P C, where s1 , s2 , s3 are diagonal entries of s P M3 . Taking

x3 “ 0, we have
p1 ´ |s1 |2 qpp1 ´ |s2 |2 q|x2 |2 ` |x1 |2 q ě |x2 |2 |1 ´ s̄1 s2 |2
for every x1 , x2 P C. Taking x1 “ 0, we have p1 ´ |s1 |2 qp1 ´ |s2 |2 q ě |1 ´ s̄1 s2 |2 , and

|s1 ´ s2 | “ 0. By the same way, we have s2 “ s3 , and s “ 0 since s P E. ˝
References: [25], [19], [29], [42]
81
2.2.2 The Choi map and decomposable maps
In the above discussion, we have seen that the smallest exposed face φ2ch meets
nontrivial completely positive maps by FE . We look for decomposable maps in φ2ch .
We note that if an extreme ray T ˝ Ads belongs to φ2ch then xT ˝ Ads , Adt y “ 0 for
every t P S, or equivalently xAds , T ˝ Adt y “ 0 for every t P S. We also note that
T ˝ Ad|ξyxη| “ Ad|ξyxη̄| . We will see that t|ξyxη̄| : |ξyxη| P Su spans the whole space,
which implies that there exists no nonzero T ˝ Ads in φ2ch .
For this purpose, we take
¨ ˛ ¨ ˛ ¨ ˛ ¨ ˛ ¨ ˛ ¨ ˛
1 1 1 1 1 1
|ξ1 y “ ˝ 1‚ , |ξ2 y “ ˝ ´1 ‚, |ξ3 y “ ˝ 1 ‚ , |ξ4 y “ ˝ í‚ , |ξ5 y “ ˝ 1 ‚, |ξ6 y “ ˝í ‚
1 1 ´1 1 í í
and |ηi y “ |ξi y for i “ 1, 2, 3, 4, 5, 6. We also define

¨ ˛ ¨ ˛ ¨ ˛ ¨ ˛ ¨ ˛ ¨ ˛
1 0 0 0 1 0
|ξ7 y “ 0 , |ξ8 y “ 1 , |ξ9 y “ 0 , |η7 y “ 0 , |η8 y “ 0 , |η9 y “ 1‚.
˝ ‚ ˝ ‚ ˝ ‚ ˝ ‚ ˝ ‚ ˝
0 0 1 1 0 0
Then we have |ξi yxηi | P S for i “ 1, 2, . . . , 9. We show that the set t|ξi yxη̄i | : i “
1, 2, . . . 9u is linearly independent. Suppose that B “ 9i“1 ai |ξi yxη̄i | “ 0, and look
ř
at the entries of the matrix B “ rBij s. Then we have
B11 “ a1 ` a2 ` a3 ` a4 ` a5 ` a6 “ 0,
B12 “ a1 ´ a2 ` a3 ´ ia4 ` a5 ´ ia6 “ 0,
B22 “ a1 ` a2 ` a3 ´ a4 ` a5 ´ a6 “ 0,
B23 “ a1 ´ a2 ´ a3 ´ ia4 ´ ia5 ´ a6 “ 0,
B31 “ a1 ` a2 ´ a3 ` a4 ´ ia5 ´ ia6 “ 0,
B33 “ a1 ` a2 ` a3 ` a4 ´ a5 ´ a6 “ 0.
With this relation, we have ai “ 0 for i “ 1, 2, . . . , 6, from which we also have

a7 “ a8 “ a9 “ 0. Therefore, we have the following:
Theorem 2.2.2 For the Choi map φch : M3 Ñ M3 , we have
DEC X φ2ch “ conv tAds : s P Eu Ă CP,
where E is the 2-dimensional subspace of M3ˆ3 consisting of all diagonal matrices

with trace zero.
We also note that t|ξi y|ηi y : i “ 1, 2, . . . , 9u spans the whole space C3 b C3 , but
their partial conjugates span the 7-dimensional subspace.
82
We close this section with a geometric interpretation of Theorem 2.2.1. To begin
with general situations, we suppose that F is a subset of a closed convex cone C.
For a given x P C, we see that x P F 2 if and only if xx, yy “ 0 for every y P F 1 if and
only if xx, yy “ 0 for every y P EpF 1 q, where EpCq denotes the set of all extreme rays
of C. Since F 1 is a face of C ˝ , we have EpF 1 q “ F 1 X EpC ˝ q. Furthermore, we see
that y P F 1 if and only if F Ă y 1 . This is also equivalent to F 2 Ă y 1 by Proposition
2.1.1. Therefore, we have
č č
F 2 “ ty 1 : y 1 Ą F 2 , y P EpC ˝ qu “ ty 1 : y 1 Ą F, y P EpC ˝ qu
for an arbitrary subset F of a closed convex cone C.

We restrict our attention to the case when C “ P1 . We know that every extreme
ray of SP1 “ P˝1 is of the form Ads , which is also exposed in SP1 . Therefore, we have
č
F 2 “ tAd1s : Ad1s Ą F u. (2.15)
In the next section, we will see that a face is maximal if and only if it is a dual
face of an exposed face which is minimal among all exposed faces. This means that
every maximal face of P1 is of the form Ad1s for a rank one matrix s. Therefore, we
see that the smallest exposed face F 2 containing a subset F of P1 is the intersection
of all maximal faces of P1 containing F .
When F is a singleton tφch u, we see that maximal faces containing φch consist
of Ad1s with s P S by Theorem 2.2.1 (i). They are also maximal faces containing
φ2ch . On the other hand, a maximal face Ad1s with a rank one matrix s contains the
subset FE of φ2ch if and only if s belongs to D, with notations in Theorem 2.2.1. We
note that D Ś S. Further, we note that there are rank one matrices in D which
does not belong to S. For example, the maximal face Ad1e12 contains FE , but does
not contain φ2ch , since e12 P DzS.
References: [19]
2.3 Maximal faces

In this section, we show that a face is maximal if and only if its dual face is minimal
among all exposed faces. Especially, every maximal face is a dual face. Since we
know all the exposed ray of P˝k “ SPk , we can characterize all maximal faces of the
convex cone Pk .
2.3.1 Boundary of convex cones

Suppose that X and Y are finite-dimensional real vector spaces, with a non-degenerate
bilinear pairing x , y on X ˆ Y . We also suppose that C is a closed convex cone in
83
X, and the bilinear pairing is non-degenerate on C, that is, the following
x P C, xx, yy “ 0 for each y P C ˝ ùñ x “ 0 (2.16)
holds. In the simplest case of X “ Y “ R2 , the bilinear pairing xx, yy “ x1 y1 ` x2 y2

is not non-degenerate on the half-plane
C :“ tx P X : xx, y0 y ě 0u, (2.17)
with y0 P Y . In this case, the dual cone is given by the single ray tλy0 P Y : λ ě 0u
generated by y0 P Y . In case of X “ Y “ Rn , if a closed convex cone C contains
two rays R` x0 and R` p´x0 q with the opposite directions then the standard bilinear
pairing xx, yy “ ni“1 xi yi is not non-degenerate on C.
ř
If C is a closed convex cone on which the bilinear pairing is non-degenerate then

for every x P C with x ‰ 0 there exists y P C ˝ such that xx, yy ą 0. Every finite
dimensional space may be endowed with a norm, and so we may use compactness
argument to conclude that there exists η P C ˝ with the property
x P C, x ‰ 0 ùñ xx, ηy ą 0, (2.18)
which is seemingly stronger than (2.16). The bilinear pairings defined in (1.39) and
(1.40) are non-degenerate on any convex cones appearing in (1.45) or (1.47). In fact,
either the identity matrix in Mm b Mn or the trace map in LpMm , Mn q play the roles
of η in (2.18).
Proposition 2.3.1 Let X and Y be finite-dimensional spaces, and C a closed con-

vex cone in X on which x , y is a non-degenerate bilinear pairing. For a given
y P C ˝ , the following are equivalent:
(i) y is an interior point of C ˝ ,
(ii) xx, yy ą 0 for each nonzero x P C.
(iii) xx, yy ą 0 for each x P C which generates an extreme ray.
Proof. If y is an interior point of C ˝ then we may take t P r0, 1q and z P C ˝ such

that y “ p1 ´ tqη ` tz, where η P C ˝ is a point with the property (2.18). Then we
see that
xx, yy “ p1 ´ tqxx, ηy ` txx, zy ą 0
for each nonzero x P C. This proves (i) ùñ (ii). It is clear that (ii) and (iii) are
equivalent. Now, we suppose that y P C ˝ satisfies (ii), and take an arbitrary point
z P C ˝ . Put C “ tx P C : }x} “ u. Then since C1 is compact, α “ suptxx, zy : x P
84
C1 u is finite, and we see that xx, zy ď 1 for each x P C1{α . By (ii), we also take δ
with 0 ă δ ă 1 such that xx, yy ě δ for each x P C1{α . Put
ˆ ˙
1 1
w “ 1´ z` y.
1´δ 1´δ
We check xx, wy ě 0 for each x P C1{α , and so have w P C ˝ . Because z was

an arbitrary point of C ˝ , we conclude that y is an interior point of C ˝ , as it was
required. ˝
The condition of non-degeneracy of the bilinear pairing on C is important in
Proposition 2.3.1. In case of the convex cone C given in (2.17), we note that any
point in C ˝ does not satisfy the condition (ii) of Proposition 2.3.1. But, all the
points of C ˝ “ tλy0 : λ ě 0u is an interior point of C ˝ except the origin.
Corollary 2.3.2 Let X and Y be finite-dimensional spaces, and C a closed convex

cone in X on which x , y is a non-degenerate bilinear pairing. Then we have the
following:
(i) If F is a face of C satisfying F 1 “ C ˝ then we have F “ t0u,
(ii) If F is a face of C ˝ satisfying F 1 “ t0u then we have F “ C ˝ .
Proof. The statement (i) is a trivial consequence of (2.16). To prove (ii), we suppose
that F Ř C ˝ . Then we have F Ă BC ˝ , and can take a nonzero y P int F Ă BC ˝ . By
the implication (ii) ùñ (i) of Proposition 2.3.1, there exists a nonzero x P C such
that xx, yy “ 0. Since y P int F , we have x P F 1 . This shows that F 1 is nonzero. ˝
We apply Proposition 2.3.1 to the duality S1˝ “ P1 , to see that φ is an interior
point of P1 rMm , Mn s if and only if x%, φy ą 0 for every nonzero % “ |ζyxζ| P S1 .
Taking |ζy “ |ξy|ηy, we see that φ is an interior point of P1 if and only if
x|ηyxη|, φp|ξyxξ|qy ą 0
for every nonzero |ξy P Cm and |ηy P Cn if and only if φp|ξyxξ|q is nonsingular for
every nonzero |ξy if and only if φpxq is nonsingular for every nonzero x P Mn` if and
only if φpxq is an interior point of Mn` for every nonzero x P Mm `
. This shows that
the converse of Proposition 1.2.5 holds.
Proposition 2.3.3 A positive linear map φ : Mm Ñ Mn is an interior point of P1

if and only if φpxq is an interior point of Mn` for every nonzero x P Mm
`
.
References: [34]
85
2.3.2 Maximal faces and minimal exposed faces
We say that L is a minimal exposed face if it is an exposed face and minimal among
all exposed proper faces. It is easy to see that if L is a minimal exposed face of the
cone C then L1 is a maximal face of C ˝ . To see this, we suppose that F is a face of
C ˝ such that F Ą L1 . Then we have
L “ L2 Ą F 1 .
Since F 1 is an exposed face and L is minimal among exposed faces, we have either
F 1 “ t0u or F 1 “ L. If F 1 “ t0u then we have F “ C ˝ by Corollary 2.3.2 (ii). In
case of F 1 “ L, we have F Ă F 2 “ L1 , which implies F “ L1 . Therefore, we conclude
that L1 is a maximal face. The following theorem tells us that the converse is also
true. Especially, every maximal face is a dual face.
Theorem 2.3.4 Let X and Y be finite-dimensional normed spaces with a non-

degenerate bilinear pairing x , y on a closed convex cone C in X. Then we have the
following:
(i) If L is a minimal exposed face of C then L1 is a maximal face of C ˝ ,
(ii) every maximal face of C ˝ is the dual face of a unique minimal exposed face of
C.
Proof. It remains to prove (ii). To do this, suppose that F is a maximal face of C ˝ .

Note that we have F Ă F 2 Ă C ˝ . If F 2 “ C ˝ then we have F 1 “ F 3 “ pC ˝ q1 “ t0u,
and so we have F “ C ˝ by Corollary 2.3.2 (ii), which is not possible. By the
maximality of F , we have F “ pF 1 q1 is the dual face of F 1 . In order to show that
F 1 is minimal among exposed faces, we suppose that L is an exposed face satisfying
L Ă F 1 . Then we have F “ F 2 Ă L1 . By the maximality of F , we have either
L1 “ F or L1 “ C ˝ . If L1 “ F then L “ L2 “ F 1 . If L1 “ C ˝ then L “ t0u by
Corollary 2.3.2 (i). Therefore, F 1 is minimal among exposed faces. In order to show
the uniqueness, we suppose that L11 “ L12 “ F for exposed faces L1 and L2 . Then
we have
L1 “ L21 “ F 1 “ L22 “ L2 ,
as it was required. ˝
We recall that Ads generates an exposed ray of the convex cone SP1 ˝ “ P1 .
Therefore, every m ˆ n matrix σ gives rise to a maximal face
conv tAds P SP1 : rank s “ 1, xs, σy “ 0u
86
s
........................
s
...............................
..........
...........
........ .......
......... ....
....
.
.... ...
...
... ...
... ...
.... ..
..
.. ..
.. ..
.. ..
.. ..
... .
.. ..
.. ..
...
... ....
..
...
... ...
... ...
... ..
.... ......
...... ....
....... ......
......... ........
...............
s
................ s
........................
...........
Figure 2.1: Line segments are minimal exposed faces. Four points are extreme points
which are not exposed, but they are exposed in the line segments.
of the convex cone SP1 rMm , Mn s by the identity (2.1). The exposed map Adσ ˝T of
P1 also gives rise to the corresponding maximal face. But, not every maximal face
arises in this way, because there are indecomposable positive map which generates
exposed ray, as we will see later.
We note that an exposed ray of a convex cone is automatically a minimal exposed
face. But the converse is not true. Consider the convex cone whose section looks like
an athletics track which consists of two semicircles and two parallel line segments.
We note that a line segment generates a minimal exposed face. This exposed face
contains an exposed ray in itself, but it is not an exposed ray in the whole convex
cone. See Figure 2.1.
References: [34]
2.3.3 Boundaries of k-positive maps

We apply Theorem 2.3.4 to SP˝k “ Pk to see that
pAds q1 “ tφ P Pk : xφ, Ads y “ 0u
is a maximal face of Pk whenever rank s ď k, since Ads generates an exposed ray

of SPk . Furthermore, every maximal face of the convex cone Pk rMm , Mn s is of the
form pAds q1 for an m ˆ n matrix s with rank at most k. Equivalently, we may apply
Theorem 2.3.4 to Sk˝ “ Pk to get the following:
Corollary 2.3.5 For a vector |ζy P Cm b Cn with SR |ζy ď k,
Fk rζs :“ tφ P Pk : x|ζyxζ|, φy “ 0u
is a maximal face of Pk rMm , Mn s. Conversely, every maximal face of Pk rMm , Mn s

is of the form Fk rζs for a vector |ζy with SR |ζy ď k.
By the uniqueness part of Theorem 2.3.4 (ii), the vector |ζy in the second part
of Corollary 2.3.5 is determined uniquely. This can be also seen as follows: We note
87
@ P` F` rζ1 s Pk Fk rζ1 s
@
@
@
@
@ Fk rζ2 s
@
@
@
@
@
@
@
Figure 2.2: When SR |ζ2 y ą `, the maximal face Fk rζ2 s of Pk is located inside of P` .
that the Choi matrices of the maps in Fk rζs X CP are positive matrices % satisfying
x|ζyxζ|, %y “ 0. Therefore, we see that Fk rζs “ Fk rωs implies that |ζy and |ωy are
parallel. In the case of k “ 1, we see that every maximal face of P1 is determined
by a product vector |ζy “ |ξy b |ηy P Cm b Cn , and we have
F1 rζs “ tφ P P1 : xη̄|φp|ξyxξ|q|η̄y “ 0u.
We consider a maximal face Fk rζs of the convex cone Pk with SR |ζy “ s ď k,

and investigate how Fk rζs is located in the bigger convex cone P` when ` ă k. In
the case of s ď `, we have Fk rζs “ F` rζs X Pk . Because Fk rζs Ă BPk , we have
Fk rζs “ F` rζs X BPk . Especially, we have Fk rζs Ă BP` . In general situations, it is
easy to see that if C1 is a convex subset of a convex set C, then either C1 Ă BC or
int C1 Ă int C holds. We show int Fk rζs Ă int P` in case of s ą `. See Figure 2.2.
Proposition 2.3.6 Suppose that SR |ζy “ s with s ď k and ` ă k. Then the

maximal face Fk rζs of Pk satisfies the following:
(i) if s ď ` then we have Fk rζs “ F` rζs X BPk Ă BP` ,
(ii) if ` ă s then we have int Fk rζs Ă int P` .
Proof. It remains to prove (ii). Assume that there exists φ P int Fk rζs such that
φ R int P` . Then φ P BP` , and so there exists |ωy with SR |ωy ď ` such that φ P F` rωs.
Since φ P int Fk rζs Ă BPk , we have φ P F` rωsXBPk “ Fk rωs by (i). Now, φ P int Fk rζs
implies that two faces Fk rζs and Fk rωs satisfy the relation Fk rζs Ă Fk rωs, and we
have Fk rζs “ Fk rωs by maximality. Therefore, we conclude that |ζy and |ωy are
parallel to each other, which implies that s ď `. ˝
The rays generated by Ads and Ads ˝T are exposed rays of DEC as well as P1 ,
since DEC Ă P1 . Because DEC is the convex hull of them, they exhaust all the
88
exposed rays of DEC. Since x%, Ads̄ y “ xs̃|%|s̃y, we see that every maximal face of
PPT is of the form
t% P PPT : xζ|%|ζy “ 0u or t%Γ P PPT : xζ|%|ζy “ 0u, (2.19)
for |ζy P Cm b Cn . This faces of PPT correspond to faces FV X F W , with nota-

tions in (2.5) and (2.6), which are determined by pairs pζ K , Mmˆn q and pMmˆn , ζ K q,
respectively.
References: [74], [75], [76]
2.4 Entanglement detected by positive maps

A self-adjoint matrix W in Mm b Mn is called an entanglement witness if x%, W y ě 0
for every separable state %, but x%0 , W y ă 0 for a state %0 . In this case, %0 must be
an entangled state which is detected by W . An entanglement witness is nothing but
the Choi matrix of a positive map which is not completely positive, or equivalently, a
block-positive matrix which is not positive. In this section, we look for conditions on
positive maps which detect maximal set of entanglement. We will begin with a pair
pC, Dq of convex cones satisfying C Ă D, motivated by inclusions S1 Ă pMm b Mn q`
and S1 Ă PPT .
2.4.1 Optimal entanglement witnesses

We recall that a state % is entangled if and only if there exists a positive map φ
such that x%, φy ă 0 with respect to the bilinear pairing (1.39). For a given positive
linear map φ : Mm Ñ Mn , we denote by Wφ the set of all (unnormalized) states
% P pMm b Mn q` such that x%, φy ă 0. If Wφ0 Ă Wφ1 then φ0 is of little use as a
detector of entanglement. In this context, it is natural to seek conditions for a pair
pφ0 , φ1 q so that Wφ0 and Wφ1 are comparable.
Let X and Y be real vector spaces with a bilinear pairing. Suppose that C Ă D
are closed convex cones in X, with the dual cones D˝ Ă C ˝ in Y . Throughout this
section, we also suppose that the bilinear pairing is non-degenerate on both D and
C ˝ . This implies that there exist ξ0 P C and η0 P D˝ satisfying
x P D, x ‰ 0 ùñ xx, η0 y ą 0, y P C ˝ , y ‰ 0 ùñ xξ0 , yy ą 0,
by (2.18). Then we see that η0 is an interior point of both D˝ and C ˝ by Proposition

2.3.1, and so we have int D˝ Ă int C ˝ . In the same way, we also see that int C Ă
int D.
89
..........................................................................................
..................... ..............
..............
C ..........
.......... ........
............. C .......
...... .............................................................................. ....
....
................ ..........
.
........
.
...
...........
C
xr1 ....
......
....
...
....
...
...
r!
C ...
... xλ! Cr ! ...
... ..
! C xt
... .. ..
..
... .. .
.. ... ..
.. ..
... !! C .... ...
....
....... r
! !
C Wy ...
....
..
...
.
.
........
.............
.... ............. C x0
.... ................................................................................
C ......
........ .
.
....
...
...
....... ...
......... .......
........... .........
...........
D
...............
C
........................... ...............
.............................................................................
C
C
Figure 2.3: If xxλ , yy “ 0, then xt is a common interior point of Wy and D.
We know that x P X does not belong to C if and only if there exists y P C ˝ such
that xx, yy ă 0. For a given y P C ˝ , we define
Wy rD; Cs :“ tx P D : xx, yy ă 0u
which is a convex subset of DzC. We use the notation Wy if there is no confusion.

When x P Wy , we say that y is a witness for x P DzC, or x is detected by y. We
note that Wy is nonempty if and only if y R D˝ . In this case, we take x1 P int Wy ,
x0 P int C Ă int D, and consider the line segment xt “ p1 ´ tqx0 ` tx1 . Since
xx0 , yy ą 0 and xx1 , yy ă 0, there exists λ P p0, 1q such that xxλ , yy “ 0. Then we
see that xt is a common interior point of Wy and D for every t P pλ, 1q, and so we
see that int Wy Ă int D. See Figure 2.3.
Furthermore, the set Wy rD; Cs has nonzero volume in D whenever it is nonempty.
To see this, it suffices to show that D is contained in the affine manifold generated
by Wy rD; Cs. If z1 P D with xz1 , yy ě 0 then we take z0 P Wy rD; Cs and put
zt “ p1 ´ tqz0 ` tz1 . Then xzt , yy “ p1 ´ tqxz0 , yy ` txz1 , yy ă 0 for sufficiently small
t ą 0, and so we see that z1 “ 1t zt ´ 1´t t
z0 belongs to the affine manifold generated
by Wy rD; Cs.
We take y0 , y1 P C ˝ zD˝ and consider the line through yt “ p1 ´ tqy0 ` ty1 . If
this line touches the convex set D˝ then the intersection with D˝ must be on either
py´8 , y0 q or py0 , y1 q or py1 , y`8 q, with the obvious meanings of notations. Suppose
that the intersection is on py´8 , y0 q, that is, there exists z P D˝ such that y0 is
between z and y1 . If we take x P D, then we have xx, zy ě 0. Therefore, xx, y0 y ă 0
implies xx, y1 y ă 0, that is Wy0 Ă Wy1 . In short, we can say that the further y is
from D˝ , the bigger Wy is. See Figure 2.4. We show that the converse also holds up
to scalar multiplications of y0 . We begin with the following simple lemma.
Lemma 2.4.1 Suppose that Wy0 Ă Wy1 . If x P D and xx, y0 y “ 0, then xx, y1 y ď 0.
Proof. Assume that there exists x P D such that xx, y0 y “ 0 and xx, y1 y ą 0. If
we take x1 P Wy0 then xx1 ` λx, y0 y ă 0 for every real number λ. But, we have
90
..........................................................................................
..................... ..............
..............
y´8 .............
..........
... ..............................................................................
..........
........
.......
....
... .......... ....
PP ................... .... ....
......
PP z ........... .... ...
r
. . ...
PP ........ ...
... ...
... ... ..
...PP .. ..
..
... PP y0 ..
..
.
..
r
.. .
.. ..
.... ...
PP ry1
... PP
.... .. .
.
....... ... ...
........ .... ...
˝ ............. ...... ....
D
PP ........
PP y`8
.... .............
.... ................................................................................
....... ....
....
......... .......
.........
...........
...............
........................... C˝
.............................................................................
...............
...........
Figure 2.4: The further y is from D˝ , the bigger Wy is.
xx1 ` λx, y1 y ě 0 for sufficiently large λ. This shows that Wy0 is not contained in
Wy1 . ˝
The author is grateful to Yoonje Jeong for simplification of the proof for the
statement (i) in the following proposition.
Proposition 2.4.2 Let X and Y be real vector spaces with a bilinear pairing.
Suppose that C Ă D are closed convex cones X, and the bilinear pairing is non-
degenerate on C ˝ and D. For y0 , y1 P C ˝ zD˝ , we have the following:
(i) Wy0 Ă Wy1 holds if and only if there exists λ ą 0 and z P D˝ such that
λy0 “ y1 ` z,
(ii) Wy0 “ Wy1 holds if and only if there exists λ ą 0 such that λy0 “ y1 ,
(iii) Wy0 Ř Wy1 holds if and only if there exists λ ą 0 and nonzero z P D˝ such
that λy0 “ y1 ` z.
Proof. For the statement (i), it remains to show the ‘only if’ part. Suppose that
Wy0 Ă Wy1 . For x P D and x1 P Wy0 , we have xxx, y0 yx1 ´ xx1 , y0 yx, y0 y “ 0, and so
we also have
xx, y0 yxx1 , y1 y ´ xx1 , y0 yxx, y1 y “ xxx, y0 yx1 ´ xx1 , y0 yx, y1 y ď 0,
by Lemma 2.4.1. Therefore, we have

ˇ 1 ˇ
1
ˇ xx , y1 y ˇ xx, y1 y
xx, y0 y ą 0, x P Wy0 ùñ ˇ 1
ˇ ˇě . (2.20)
xx , y0 y ˇ xx, y0 y
Define "ˇ 1 ˇ *
ˇ xx , y1 y ˇ 1
λ :“ inf ˇ 1
ˇ ˇ : x P Wy0 .
xx , y0 y ˇ
xξ,y1 y
Taking ξ P C such that xξ, yy ą 0 for every nonzero y P C ˝ , we see that λ ě xξ,y0 y
,
and so we have λ ą 0. By (2.20), we have
λxx, y0 y ´ xx, y1 y ě 0, (2.21)
91
whenever x P D satisfies xx, y0 y ą 0. By definition of λ, we also have (2.21) for
x P D with xx, y0 y ă 0. If xx, y0 y “ 0 then we also have (2.21) by Lemma 2.4.1.
Therefore, we have λy0 ´ y1 P D˝ , which completes the proof of (i).
To prove (ii), suppose that Wy0 “ Wy1 . Then by (i), there exists λ, µ ą 0 and
z, w P D˝ such that λy0 “ y1 ` z and µy1 “ y0 ` w, and so λpµy1 ´ wq “ y1 ` z.
Therefore, we have pλµ ´ 1qy1 “ λw ` z. Because λw ` z P D˝ and y1 P C ˝ zD˝ is
nonzero, we have λµ ´ 1 “ 0 and w “ z “ 0. The converse is clear.
Finally, the statement (iii) is an immediate consequence of (i) and (ii). ˝
For y0 and y1 in Proposition 2.4.2, we note that yi1 “ tx P C : xx, yi y “ 0u for
i “ 0, 1. The condition of Proposition 2.4.2 (i) shows that λxx, y0 y “ xx, y1 y ` xx, zy
for every x P C, and so we see that Wy0 Ă Wy1 implies y01 Ă y11 . If Wy0 “ Wy1 then
y01 “ y11 by Proposition 2.4.2 (ii). It should be noted that y01 “ y11 does not imply
Wy0 “ Wy1 . Indeed, we consider the pair S1 Ă pMm b Mn q` with the dual pair
CP Ă P1 , and the map φ1{2 “ 21 φ0 ` 21 φch with φ0 in (2.14). We see that φ11{2 “ φ1ch ,
but Wφ1{2 Ř Wφch by Proposition 2.4.2 (iii), because φ0 P CP.
Theorem 2.4.3 Suppose that C Ă D are closed convex cones in a real vector space
X, and Y is a real vector space with a bilinear pairing on X ˆ Y which is non-
degenerate on C ˝ and D. For y0 P C ˝ zD˝ , the following are equivalent:
(i) Wy0 is maximal among tWy : y P C ˝ zD˝ u,
(ii) the smallest face of C ˝ containing y0 has no nontrivial intersection with D˝ .
Proof. We denote by F the smallest face of C ˝ containing y0 , then y0 is an interior

point of F . Suppose that there exists a nonzero z P F X D˝ . Then there exist t ă 0
such that yt :“ p1 ´ tqy0 ` tz P F Ă C ˝ . This implies p1 ´ tqy0 “ yt ` p´tqz with
p´tqz P D˝ , and we have Wy0 Ř Wyt by Proposition 2.4.2 (iii). This tells us that
Wy0 is not maximal. For the converse, suppose that Wy0 is not maximal, and there
exists y1 P C ˝ such that Wy0 Ř Wy1 . Then there exist λ ą 0 and nonzero z P D˝
such that λy0 “ y1 ` z, which implies z P F . Therefore, we conclude that F X D˝ is
nontrivial. ˝
For a given y0 P C ˝ , we consider the hyperplane Hy0 “ tx P X : xx, y0 y “ 0u.
The relations between the set Wy0 rD; Cs, the hyperplane Hy0 and the convex cone
D depend on the location of y0 . See Figure 2.5. We recall that D˝ Ă C ˝ .
• If y0 is an interior point of D˝ then Hy0 has no nontrivial intersection with D

and Wy0 rD; Cs is empty.
• If y0 is a boundary point of D˝ then Hy0 X D is nontrivial but Wy0 rD; Cs is

still empty.
92
PP
PP
P
..........................................PP ..........................................
..................... ........... .................... ...........
.
......
..
..
.. ........ P .......... ........
.
..
.. ..... ........ .....
.
... .... P ..... ....
.... ..
...
..
...
....
...
........................................................
Hy0 .
.
...
.
.
...
................. .................... ..
... ....
....
..
......................
... .................. .... .......
.... ..
..
..
........... ...
... ... ....
... .
......... ...
...
..... .....
..
...
..
.
....
.
...
.
.. ...
..
..
...
... .
....
...
....
ry0 ...
..
..
..
.. D .. C ..
.. .. C ˝ D˝
...
..
..
..
PP
PP ...................................... ......................................
....................... ........... ....................... ...........
........... ...........
PP ........ ........
........ ...... ........ ......
.......
.
PP .... .......
. ....
...
. PH .... ...
. ....
r
. .. . ..
.... ..... ......
..... .....
. ........................... y0 .... ..... ......
..... .....
. ...........................
.... ........................ ... .... ........................ ...
...
... ....
.
........... .
........
..
...
... ...
... ....
.
........... .
........
..
y0 ...
...
... .... ... ... .... ...
... .... .. ... ˝ .... ˝ ..
.. D .. C ..
.. .. C .. D ..
..
PP ....................................... .......................................
...................... ......................
ry0
........... ...........
........... ...........
PP ........ ........
........ ..... ........ .....
.......
.
PP .... .......
. ....
..
... PP ....
..
... ....
... .. ... ..
. PH ..................................... ..
... .........................
. . ..................................... ..
... .........................
.
.. .
....... ... .. .
....... ...
.
.... ...
.... .
........
... y0 .
.... ...
... .
.........
...
... . .
. ..... ... ... . ...... ...
. .
... .... ..
.. ... .... ..
..
.... ....
...
.. D .. C ..
.. .. C ˝ D˝
...
..
..
..
......................................
.......................
r
......................................
.......................
.....
.
...........
........
PP
PP
...........
........
......
.... .
...........
........
.....
y0 ...........
........
......
....
.
... .... .
... ....
..
. PP .. ..
. ..
...
. ... ......
..... .....
. ........................... ...
. ... ......
..... .....
. ...........................
..
.
PP ..
........
..................... ... ..
. ..
............................
. ...
.... PH ...
.......... ... .... ...
.......... ...
..... ... .... ...
....
.. ....
.. .
.
... y0 ....
.. ....
.. ..
...
.... .. .... ..
...
.. D .. C ..
.. .. C ˝ D˝
...
..
..
..
Figure 2.5: The location of the hyperplane Hy0 depends on the location of y0 P C ˝
• When we take y0 P int C ˝ zD˝ , the set Wy0 rD; Cs is nonempty and int pHy0 X
Dq Ă int D. But Hy0 X C is still trivial.
• If we take y0 P BC ˝ zD˝ then y01 “ Hy0 X C is nontrivial. In this case, we take

the smallest face F Ă BC ˝ of C ˝ containing y0 . Theorem 2.4.3 tells us that
the set Wy0 is maximal if and only if F X D˝ is trivial.
In order to detect entangled states, we first apply Theorem 2.4.3 to the pair of
convex cones S1 Ř pMm b Mn q` with the dual pair CP Ř P1 of linear maps, or
equivalently pMm b Mn q` Ř BP 1 . We say that a positive map φ or its Choi matrix
Cφ is optimal if the set Wφ rpMm b Mn q` ; S1 s is maximal. Theorem 2.4.3 tells us
that φ is optimal if and only if the smallest face of P1 containing φ has no nonzero
completely positive map. We see that the Choi map φch is optimal by Theorem 2.2.1
(iv). Every extremal positive map is optimal if it is not completely positive, and
so the completely copositive maps Ads ˝T are optimal whenever rank s ě 2. Such
optimal positive maps are actually of little use to detect entanglement, because they
can detect only non-PPT states. If we want to detect PPT entanglement then we
have to use the pair S1 Ă PPT with the dual pair DEC Ă P1 . In this case, we see
that the set Wφ rPPT ; S1 s is maximal if and only if the smallest face of P1 containing
φ has no nontrivial decomposable maps.
93
References: [123], [82], [102]
2.4.2 Spanning properties

Because we do not know the facial structures of the convex cone P1 , it is not so easy
to determine if a given positive map is optimal or not. We consider a slightly stronger
condition than Theorem 2.4.3 (ii); the smallest exposed face φ2 of P1 containing φ
has no nonzero completely positive maps, that is, φ2 X CP “ t0u.
We begin with general situations. We compare the definition of Wy rD; Cs with
1
y “ tx P C : xx, yy “ 0u, which must be considered as a face of C, not a face of D
since y P C ˝ zD˝ .
Proposition 2.4.4 Suppose that C Ă D are closed convex cones in a real vector
space X, and Y is a real vector space with a bilinear pairing on X ˆ Y . For y0 P
C ˝ zD˝ , the following are equivalent:
(i) the smallest exposed face of C ˝ containing y0 has no nontrivial intersection

with D˝ , that is, y02 X D˝ “ t0u,
(ii) int y01 Ă int D.
Proof. For the convex subset y01 of D, there are two possibilities; either int y01 Ă int D
or y01 Ă BD. We will show that y01 Ă BD if and only if y02 X D˝ ‰ t0u. Suppose that
y01 Ă BD. This happens if and only if the face y01 is contained in a maximal face y11
of D for a point y1 P D˝ . We see that y01 Ă y11 if and only if y12 Ă y02 if and only if
y1 P y02 . This implies y1 P y02 X D˝ . Conversely, if y1 P y02 X D˝ is nonzero then we
have y01 Ă y11 Ă BD. ˝
The location of the nontrivial face y01 “ Hy0 X C with y0 P BC ˝ zD˝ depends on
the exposed face y02 . Proposition 2.4.4 tells us that y01 is located inside of D if and
only if y02 does not touch D˝ except zero. Whenever this is the case, the set Wy0 is
maximal. See Figure 2.6.
Now, we look for positive maps φ satisfying φ2 X CP “ t0u. Since a completely
positive map is the sum of Ads ’s, we see that φ2 X CP ‰ t0u holds if and only if
there exists s P Mmˆn such that Ads P φ2 , which is equivalent to the following:
SR |ζy “ 1, x|ζyxζ|, φy “ 0 ùñ x|ζyxζ|, Ads y “ 0. (2.22)
Motivated by this, we define the set
P rφs :“ t|ζy “ |ξy|ηy P Cm b Cn : x|ζyxζ|, φy “ 0u,
for a positive map φ. Then, we see that φ2 has no nonzero completely positive map
¯ 2
if and only if P rφs spans the whole space, by the relation x|ζyxζ|, Ads y “ |xζ|s̃y|
94
.................................................................... r
............... ...............
............ ....... .......
.........
.......
....
PP
PP .........
.......
....
.... ...
.... y1 ...
....
..
. .... ... ...
.... ... ...
PP
.
....
.
.. PPr .
....... .
...............................................................................
... ...
..
...
..
... 1 PH ......... ... .. ..
. ... ...
.
..
... .
....
...
. ...
y1 ..
..
..
..
y1 .
..
..
.
..
..
.... .. ..
...
.. D .. C ..
..
. ..
.. C˝ ..
.. D˝
.........
...................................................................
............ .........
.......
r
.....
....
....
.............
.....
....
....
.............
.......
.....
....
.... ..y2
... ...
..
r
.. .... . .
.... ... ...
.
....
.
.. .
............................................................................
.... Hy0 ...
..
...
..
...
.
.. ....
....
...........
........
y21 ...
...
..
..
..
...
.
..
..
...
.
..
... .... .. .. ..
... .... .. .. ..
.. D .. C ..
.. ..
.. C˝ ..
.. D˝
Figure 2.6: y12 touches D˝ , but y22 does not.
as in (2.1). A positive map φ satisfying this property is said to have the spanning
property. We note that
|ξy|ηy P P rφs ðñ xφ, Ad|ξ̄yxη| y “ 0 ðñ xη̄|φp|ξyxξ|q|η̄y “ 0 (2.23)
by (2.11). It is clear that a positive map with the spanning property is optimal.
We have seen in Theorem 2.2.1 that the Choi map φch does not have the spanning
property. By Proposition 2.4.4, we see that a positive linear map φ : Mm Ñ Mn has
the spanning property if and only if int φ1 Ă int pMn b Mn q` .
For a product vector |ζy “ |ξy|ηy P Cm b Cn , we recall the identity
¯
|ζyxζ|Γ “ |ξy|ηyx ¯
ξ|xη|.
If we replace Ads by Ads ˝T in (2.22), then we have
x|ζyxζ|, Ads ˝Ty “ x|ζyxζ|Γ , Ads y.
Therefore, we see that the following are equivalent;
• φ2 has no nonzero completely copositive maps,
• the partial conjugates of P rφs spans the whole space,
A positive map with such properties is called to have the co-spanning property. It
is clear that φ has the spanning property if and only if φ ˝ T has the co-spanning
property. Theorem 2.2.2 tells us that the Choi map φch has the co-spanning property.
We say that a positive map φ is co-optimal when the smallest face of P1 contain-
ing φ has no nonzero completely copositive map, or equivalently φ ˝ T is optimal.
We also say that a positive map is bi-optimal if it is both optimal and co-optimal.
We note that φ is bi-optimal if and only if the smallest face of P1 containing φ has
no nonzero decomposable map, and so we have the following by Theorem 2.4.3.
95
Proposition 2.4.5 For a positive map φ : Mm Ñ Mn , the following are equivalent:
(i) φ is bi-optimal,
(ii) the smallest face of P1 containing φ has no decomposable map,
(iii) φ detects a maximal set of PPT entanglement.
Therefore, an bi-optimal positive map should be indecomposable automatically.

On the other hand, a positive map φ has the bi-spanning property when it has both
the spanning property and the co-spanning property. We apply Proposition 2.4.4 to
the pair S1 Ă PPT to get the following:
Theorem 2.4.6 For a positive linear map φ, the following are equivalent:
(i) φ has the bi-spanning property,
(ii) φ2 has no nonzero decomposable maps,
(iii) int φ1 Ă int PPT .
The property (iii) of Theorem 2.4.6 is useful to find PPT entanglement. If we

take a finite family I of product vectors in P rφs so that both they and their partial
conjugates span the whole space, then the average %1 of t|ζi yxζi | : i P Iu lies on the
boundary of S1 , but in the interior of PPT . If we take any %0 in the interior of S1
then the line segment %λ “ p1 ´ λq%0 ` λ%1 extends to PPT entanglement.
A positive map with the the bi-spanning property must be indecomposable by the
condition (ii) of Theorem 2.4.6. An indecomposable positive map which generates
an exposed ray has automatically the bi-spanning property. The following shows
implication relations among various notions we have discussed so far.
exposed indecomposable ùñ bi-spanning ùñ spanning
ó ó ó
extreme indecomposable ùñ bi-optimal ùñ optimal
indecomposable
References: [123], [82], [45], [77]
96
2.5 Positive maps of Choi type
We exhibit in this section variants of the Choi map between 3 ˆ 3 matrices. Among
them, we find positive maps with the bi-spanning property, and indecomposable
positive maps with the spanning property without bi-optimality. Motivated by these
examples, we briefly discuss the length of a separable state %, which is the smallest
number of pure products states whose sum gives rise to %.
2.5.1 Choi type positive maps between 3 ˆ 3 matrices

For a given triplet pa, b, cq of nonnegative real numbers and a 3 ˆ 3 matrix X “ rxij s,
we define
¨ ˛
ax11 ` bx22 ` cx33 ´x12 ´x13
φra,b,cs pXq “ ˝ ´x21 cx11 ` ax22 ` bx33 ´x23 ‚.
´x31 ´x32 bx11 ` cx22 ` ax33
Note that φr1,0,1s is nothing but the Choi map φch defined in (1.54), and the map τ3,k
in (1.52) is given by φrk´1,k,ks . Especially, we have seen in Section 1.6 that φr1,2,2s
is a 2-positive map which is not completely positive, and φr0,1,1s is a completely
copositive map which is not 2-positive. The Choi matrix of the map φra,b,cs is given
by
¨ ˛
a ¨ ¨ ¨ ´1 ¨ ¨ ¨ ´1
˚ ¨ c ¨ ¨ ¨ ¨ ¨ ¨ ¨ ‹
˚ ‹
˚ ¨ ¨ b ¨ ¨ ¨ ¨ ¨ ¨ ‹
˚ ‹
˚ ¨ ¨ ¨ b ¨ ¨ ¨ ¨ ¨ ‹
˚ ‹
˚ ´1 ¨ ¨ ¨ a ¨ ¨ ¨ ´1 ‹ .
%ra,b,cs :“ ˚ (2.24)
‹
˚ ¨ ¨ ¨ ¨ ¨ c ¨ ¨ ¨ ‹
˚ ‹
˚ ¨ ¨ ¨ ¨ ¨ ¨ c ¨ ¨ ‹
˚ ‹
˝ ¨ ¨ ¨ ¨ ¨ ¨ ¨ b ¨ ‚
´1 ¨ ¨ ¨ ´1 ¨ ¨ ¨ a
Note that φra,b,cs is completely positive if and only if %ra,b,cs is positive if and only
if the following
aě2
holds. For example, the map
φr2,0,0s “ Ade11 é22 ` Ade22 é33 ` Ade33 é11
has been discussed in (2.14), where we use teij u for the standard matrix units. On
97
the other hand, φra,b,cs is completely copositive if and only if
¨ ˛
a ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨
˚ ¨ c ¨ ´1 ¨ ¨ ¨ ¨ ¨ ‹
˚ ‹
˚ ¨ ¨
˚ b ¨ ¨ ¨ ´1 ¨ ¨ ‹
‹
˚ ¨ ´1 ¨
˚ b ¨ ¨ ¨ ¨ ¨ ‹
‹
Γ
%ra,b,cs “ ˚˚ ¨ ¨ ¨ ¨ a ¨ ¨ ¨ ¨ ‹
‹
˚ ¨ ¨
˚ ¨ ¨ ¨ c ¨ ´1 ¨ ‹
‹
˚ ¨ ¨ ´1 ¨ ¨ ¨
˚ c ¨ ¨ ‹
‹
˝ ¨ ¨ ¨ ¨ ¨ ´1 ¨ b ¨ ‚
¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ a
is positive if and only if

bc ě 1.
The completely copositive map φr0,1,1s may be written by
φr0,1,1s “ T ˝ pAde12 é21 ` Ade23 é32 ` Ade31 é13 q .
We note that the map φra,b,cs may be expressed by
φra,b,cs “ ψra,b,cs ´ id3
with the map ψra,b,cs , which sends x “ rxij s to the diagonal matrix with the diagonal
entries
pa ` 1qx11 ` bx22 ` cx33 ,
cx11 ` pa ` 1qx22 ` bx33 ,
bx11 ` cx22 ` pa ` 1qx33 .
We apply Proposition 1.6.5, to see that φra,b,cs is positive if and only if
ψra,b,cs p|ξ0 yxξ0 |q ě |ξ0 yxξ0 |
for every unit vector |ξ0 y if and only if
xξ0 |ψra,b,cs p|ξ0 yxξ0 |q´1 |ξ0 y ď 1
for every unit vector |ξ0 y if and only if the inequality
α β γ
` ` ď1 (2.25)
pa ` 1qα ` bβ ` cγ pa ` 1qβ ` bγ ` cα pa ` 1qγ ` bα ` cβ
holds for all positive real numbers α, β and γ.

We take α “ β “ γ “ 1 to get a ` b ` c ě 2. We also take β “ α´1 and γ Ñ 0,
then we have
caα4 ` pa2 ´ 1 ` bcqα2 ` ab ě 0
98
for each positive α. If a ě 1 then this is true. If 0 ď a ď 1 then this implies
a2 ´ 1 ` bc ě 0 or pa2 ´ 1 ` bcq2 ´ 4bca2 ď 0,
which implies bc ě p1 ´ aq2 . Therefore, we get necessary conditions
a ` b ` c ě 2, 0 ď a ď 1 Ñ bc ě p1 ´ aq2 , (2.26)
where p Ñ q means that p implies q. Proving the following lemma, we may conclude
that the map φra,b,cs is positive if and only if the conditions in (2.26) are satisfied.
Lemma 2.5.1 If nonnegative numbers a, b and c satisfy the conditions in (2.26)

then inequality (2.25) holds for all positive α, β, γ.
Proof. Suppose that a, b and c satisfy (2.26), and α, β and γ are positive numbers.
Put
bβ ` cγ bγ ` cα bα ` cβ
x“ , y“ , z“ ,
α β γ
and consider the system of linear equations
xα ´ bβ ´ cγ “ 0,
ćα ` yβ ´ bγ “ 0,
´bα ´ cβ ` zγ “ 0,
with unknowns α, β and γ. This system of equations has already a nontrivial

solution, and so we have
xyz ´ bcpx ` y ` zq ´ pb3 ` c3 q “ 0. (2.27)
On the other hand, the left-side of (2.25) becomes

1 1 1
` ` ,
a`1`x a`1`y a`1`z
and so, it suffices to show the inequality
F px, y, zq “ xyz ` apxy ` yz ` zxq ` pa2 ´ 1qpx ` y ` zq ` pa ` 1q2 pa ´ 2q ě 0
holds under the condition (2.27). To see this, we slice the surface (2.27) by the plane
x ` y ` z “ d. (2.28)
We first note that the surface (2.27) and the plane (2.28) has nonempty inter-
section if and only if d ě 3pb ` cq. When d “ 3pb ` cq, the intersection is the one
point pb ` c, b ` c, b ` cq. When d ą 3pb ` cq, the intersection is a compact curve.
99
One can show that if px, y, zq is a critical point of F under the constraints (2.27)
and (2.28) then
x “ y, y “ z or z “ x. (2.29)
Therefore, it suffices to find minimum of F on the curve given by (2.27) and (2.29).
In case of x “ y, the curve is given by px2 ´ bcqz “ 2bcx ` b3 ` c3 from which we have
?
x ą bc, and F becomes a function of one variable x. By a direct computation, we
have
1 “ 2 ‰
F 1 pxq “ 2 2
x ` pb ` cqx ` pb2 ´ bc ` c2 q
px ´ bcq
ˆ rx ´ pb ` cqsrax2 ` pa2 ´ 1 ` bcqx ` abcs.
The first two factors are nonnegative, and the last factor is also nonnegative by the
second conditions in (2.26). Now, we conclude that F has the minimum at x “ b`c,
which implies y “ z “ b ` c by (2.27) and x “ y. Therefore, we have
F ě 3apb ` cq3 ` 3pa2 ´ 1 ` bcqpb ` cq ` ppa ` 1q2 pa ´ 1q ` b3 ` c3 q

“ pa ` b ` c ´ 2qpa ` b ` c ` 1q2 ě 0,
by the first condition of (2.26). The cases of y “ z and z “ x can be done in the
same way. ˝
In order to deal with decomposability of the map φra,b,cs , we consider the PPT
state %p defined in (1.36) We take bilinear pairing with %ra,b,cs to get
0 ď x%p , φra,b,cs y “ 3rcp2 ` pa ´ 2qp ` bs
for every p ą 0. Therefore, we get the following necessary condition

ˆ ˙2
2á
0 ď a ď 2 Ñ bc ě . (2.30)
2
Compare with the condition (2.26). Now, we show that the condition (2.30) implies
that the map φra,b,cs is decomposable. If a ě 2 or bc ě 1 then there is nothing to
prove since φra,b,cs is completely positive or completely copositive in each case. In
case of 0 ď a ă 2 and bc ă 1, we have
? ?
φra,b,cs “ p1 ´ bcq φrα,0,0s ` bc φr0,?b{c,?c{bs ,
with α “ 1á?bc . We note that α ě 2 by (2.30), and so φrα,0,0s is completely positive.

Since φr0,?b{c,?c{bs is completely copositive, we have seen that φra,b,cs is decompos-
able. We summarize as follows:
Theorem 2.5.2 Let a, b and c be nonnegative real numbers. Then the linear map
φra,b,cs is
100
(i) positive if and only if a ` b ` c ě 2 and 0 ď a ď 1 Ñ bc ě p1 ´ aq2 ,
(ii) completely positive if and only if a ě 2,
(iii) completely copositive if and only if bc ě 1,

` 2á ˘2
(iv) decomposable if and only if 0 ď a ď 2 Ñ bc ě 2
.
For further examples of Choi type positive linear maps in various directions, see
Section 7 of the survey paper [30] and references there. See also [71, 105] for recent
development.
References: [18], [30], [71], [105]
2.5.2 Spanning properties of Choi type positive maps

In order to figure out the region given by (2.26), we look at the curve given by
0 ď a ă 1, a ` b ` c “ 2, bc “ p1 ´ aq2 .
b
We parameterize by t “ 1á
. Then 0 ă t ă 8, and we have
at ` b “ t, p1 ´ tqa ` c “ 2 ´ t, a ` ct “ 1, (2.31)
from which we have

ˆ ˙
p1 ´ tq2 t2 1
γptq “ paptq, bptq, cptqq “ , , .
1 ´ t ` t2 1 ´ t ` t2 1 ´ t ` t2
2 2 2
One may
b easily check that γ makes a part of the circle centered at p 3 , 3 , 3 q with the
radius 23 . See Figure 2.7. When t “ 0, we have γp0q “ p1, 0, 1q gives rise to the
Choi map φch , and we also get the completely copositive map φr0,1,1s at t “ 1. We
write
φt :“ φraptq,bptq,cptqs , 0 ă t ă 8.
Now, we use (2.23) to look for product vectors in P rφt s. We write |ξy “
px, y, zqT P C3 . When |x| “ |y| “ |z| “ 1, we see that
¨ ˛
2 ´xȳ ´xz̄
φt p|ξyxξ|q “ ˝ýx̄ 2 ýz̄ ‚
´z x̄ ´z ȳ 2
has the kernel vector |η̄y “ px, y, zqT P C3 . Therefore, we have
|ξy b |ηy “ pα, β, γqT b pᾱ, β̄, γ̄qT P P rφt s, (2.32)
101
φ0 “ φr1,0,1s “ φch..........r........................................................................................
.....
."
.....
.
.......
.......
...... rφt
......
....
:“ φrγptqs
".............
" ....
....
" . ....
...
" ....... ...
...
. ...
" ..
.. ..
" .. ..
.. ..
φr2,0,0s r" rφ1

... ..
"
...
..
..
..
..
. “ φr0,1,1s
b .. ....
.. ...
b .. .
... .
b ... ...
. ...
b ...........
b ..
.
....
b ....... ..
....
b ........... ....
......
b................ ......
b....r....................... ......
.
........
..........
φ8 “ φr1,1,0s ...............................................
Figure 2.7: Two surfaces a ` b ` c “ 2 and bc “ p1 ´ aq2 make a circle.
whenever |α| “ |β| “ |γ| “ 1. When one of x, y, z, say z “ 0, the determinant

det φt p|ξyxξ|q is given by
pb|x|2 ` c|y|2 qpac|x|4 ` |x|2 |y|2 a2 ` |x|2 |y|2 bc ` ab|y|4 ´ |y|2 |x|2 q
pt ´ 1q2
“ 2 3
pt2 |x|2 ` |y|2 qp´|y|2 t ` |x|2 q2 ,
pt ´ t ` 1q
?
which becomes zero when |x| “ t and |y| “ 1. In this case,
¨ ˛
at ` b ´xȳ 0
φt p|ξyxξ|q “ ˝ ýx̄ ct ` a 0 ‚
0 0 bt ` c
has the kernel vector |η̄y “ px, ty, 0qT P C3 . Therefore, we have
? ?
|ξy b |ηy “ p tα, β, 0qT b p tᾱ, tβ̄, 0qT P P rφt s. (2.33)
By the same way, we also have

? ?
|ξy b |ηy “ p0, tα, βqT b p0, tᾱ, tβ̄qT P P rφt s,
? ? (2.34)
|ξy b |ηy “ pβ, 0, tα, qT b ptβ̄, 0, tᾱqT P P rφt s,
with |α| “ |β| “ 1.

Among them, we take ten product vectors with real coefficients as follows: We
first take
|ξ1 y b |η1 y “ p1, 1, 1qT b p1, 1, 1qT ,
|ξ2 y b |η2 y “ p1, 1, ´1qT b p1, 1, ´1qT ,
(2.35)
|ξ3 y b |η3 y “ p1, ´1, 1qT b p1, ´1, 1qT ,
|ξ4 y b |η4 y “ p´1, 1, 1qT b p´1, 1, 1qT ,
102
among product vectors in (2.32). They span 4-dimensional space whose orthogonal
complement is spanned by
|12y ´ |21y, |23y ´ |32y, |31y ´ |13y, |11y ´ |22y, |22y ´ |33y.
Next, we also take

? ?
|ξ5 y b |η5 y “ p t, 1, 0qT b p t, t, 0qT ,
? ? (2.36)
|ξ6 y b |η6 y “ p t, ´1, 0qT b p t, ´t, 0qT ,
among (2.33), and
? ?
|ξ7 y b |η7 y “ p0, t, 1qT b p0, t, tqT ,
? ?
|ξ8 y b |η8 y “ p0, t, ´1qT b p0, t, ´tqT ,
? ? (2.37)
|ξ9 y b |η9 y “ p1, 0, tqT b pt, 0, tqT ,
? ?
|ξ10 y b |η10 y “ p´1, 0, tqT b p´t, 0, tqT ,
among (2.34). These six vectors span a 6-dimensional space whose orthogonal com-
plement is spanned by
|12y ´ t|21y, |23y ´ t|32y, |31y ´ t|13y. (2.38)
It is easily seen that the above ten product vectors span the whole space C3 b C3
unless t ‰ 1. Therefore, we have the following:
Theorem 2.5.3 The positive map φra,b,cs has the bi-spanning property, whenever
pa, b, cq satisfies
0 ă a ă 1, a ` b ` c “ 2, bc “ p1 ´ aq2 .
In fact, it is known [47] that the maps in Theorem 2.5.3 generates an exposed
ray of the convex cone P1 .
Now, we turn our attention to the surface bc “ p1 ´ aq2 with 0 ď a ď 1 in the
condition (2.26) of positivity of the map φra,b,cs . To do this, we fix t ą 0 with t ‰ 1,
and consider the line segment
" *
´ s¯ t
Lt :“ Lt psq “ 1 ´ s, st, : ďsď1 ,
t 1 ´ t ` t2
which is a part of the line segment from p1, 0, 0q to δptq :“ p0, t, 1t q. Precisely, it is
the line segment between two points between γptq and δptq. It is easily seen that this
line segment is contained in the surface bc “ p1 ´ aq2 in (2.26). See Figure 2.8. We
note that Lt p1q “ δptq “ p0, t, 1t q gives rise the the completely copositive map φr0,t, 1 s ,
t
and so it is clear that maps on the line segment Lt do not satisfy the co-optimal
property. Nevertheless, we will show that they satisfy the spanning property.
103
p2, 0, 0q
p1, 0, 1q
γptq
t
p1, 1, 0q δptq
Figure 2.8: The surface bc “ p1 ´ aq2 consists of line segments.
Theorem 2.5.4 The positive map φ :“ φra,b,cs has the spanning property, whenever
pa, b, cq satisfies
0 ď a ă 1, a ` b ` c ą 2, bc “ p1 ´ aq2 .
Proof. We define vectors in C3 as follows;

1
|ξθ,σ y “p0, eiθ b1{4 , eiσ c1{4 qT , 1
|ηθ,σ y “p0, eíθ pbcq1{4 , eíσ b1{2 qT ,
2
|ξθ,σ y “peiσ c1{4 , 0, eiθ b1{4 qT , 2
|ηθ,σ y “peíσ b1{2 , 0, eíθ pbcq1{4 qT ,
3
|ξθ,σ y “peiθ b1{4 , eiσ c1{4 , 0qT , 3
|ηθ,σ y “peíθ pbcq1{4 , eíσ b1{2 , 0qT .
It is easily checked that
k k k k
xξθ,σ |xηθ,σ |Cφ |ξθ,σ y|ηθ,σ y “ ´2p1 ´ aqbc1{2 ` 2b3{2 c “ 0
k k
for k “ 1, 2, 3. Therefore, the vectors |ξθ,σ y|ηθ,σ y belong to P rφra,b,cs s for k “ 1, 2, 3.
We take σ1 “ 0, σ2 “ π{2 and σ3 “ π, and consider the 9 ˆ 9 matrix whose columns
k k
consist of nine vectors |ξ0,σ `
y|η0,σ `
y for k, ` “ 1, 2, 3. Then the absolute value of the
9 9
determinant is given by 128 b 2 c 4 which is nonzero. ˝
Therefore, every triplet pa, b, cq in Theorem 2.5.4 gives rise to an indecomposable

positive map φ :“ φra,b,cs with the following properties;
104
................................................
.................... .........
...........
......... .......
....... ......
.
...
.
.......
..
........
pM3 b M3 q` ....
....
....
....
.. ....
XXX ...
. ....
.... .. ....
.... . . .
...
.................................... ....
... XXX ................ ........
..
.
...
... ..
..
....... ......
..
...
r
.... ...
... .......... XXX .... ...
.. ....
. .
.. .....
..... ..
............ ...
.... ....
. X ..
.............. .... ...
1 XX ....... ....
... ...
... ...
...
......
.
...
φ ....
....
XXX ...
...
...
...... X
......
...... ..
.
.
.
.
..
...
....
....
...
...
...
Hφ
...... ... ... ...
...... .... ... ...
... .... ... ...
PPT
...
...
...
..
S1
..
..
...
...
.
....
...
...
..
..
..
... ... ... ..
. ..
. .
Figure 2.9: Entanglement Wφ rpMm b Mn q` , S1 s detected by φ is maximal, but PPT

entanglement Wφ rPPT , S1 s detected by φ is not maximal.
• φ has the spanning properties, so it is optimal and detects a maximal set

Wφ ppM3 b M3 q` , S1 pM3 b M3 qq of entanglement. The interior of the face φ1 of
S1 is sitting in the interior of pM3 b M3 q` by Proposition 2.4.4.
• φ does not have co-optimal properties, especially the smallest face of P1 con-
taining φ has a nonzero decomposable map, so it does not detect a maximal set
Wφ pPPT rM3 b M3 s, S1 pM3 b M3 qq of PPT entanglement by Theorem 2.4.3.
The face φ1 of S1 is sitting on the boundary of PPT by Proposition 2.4.4. See
Figure 2.9.
References: [31], [44], [45], [46], [47], [48]
2.5.3 Lengths of separable states

We return to the ten product vectors |ζi y :“ |ξi y|ηi y listed in (2.35), (2.36) and
(2.37). We note that
¨ ˛
α ¨ ¨ ¨ 1 ¨ ¨ ¨ 1
˚ ¨
˚ γ ¨ 1 ¨ ¨ ¨ ¨ ¨ ‹
‹
˚ ¨
˚ ¨ β ¨ ¨ ¨ 1 ¨ ¨ ‹
‹
ÿ10 ˚ ¨
˚ 1 ¨ β ¨ ¨ ¨ ¨ ¨ ‹
‹
%t :“ ˚ 1 ¨ ¨ ¨ α ¨ ¨ ¨ 1
|ζi yxζi | “ ˚ ‹,
‹
i“1 ˚ ¨
˚ ¨ ¨ ¨ ¨ γ ¨ 1 ¨ ‹
‹
˚ ¨
˚ ¨ 1 ¨ ¨ ¨ γ ¨ ¨ ‹
‹
˝ ¨ ¨ ¨ ¨ ¨ 1 ¨ β ¨ ‚
1 ¨ ¨ ¨ 1 ¨ ¨ ¨ α
with
2t2 ` 2 t`2 t3 ` 2
α“ , β “ , γ “ ,
t2 ` 2 t2 ` 2 t2 ` 2
up to scalar multiplications. So far, we have seen that %t is a separable state on the
boundary of S1 . Furthermore, %t is in the interior of PPT unless t “ 1. In fact, one
may see both %t and %Γt have full ranks when t ‰ 1.
105
We have found product vectors |ξy|ηy P P rφt s in (2.32), (2.33) and (2.34). It is
known [46] that they are all product vectors in P rφt s, and there is no more product
vectors in P rφt s. Now, we consider the map
ψt :“ φt ` φt ˝ T, t ą 0, t ‰ 1.
Then |ξy|ηy P P rψt s if and only if |ξy|ηy P P rφt s and |ξy|ηy ¯ P P rφt s if and only if
|ξy|ηy is one of ten product vectors |ζ1 y, . . . , |ζ10 y listed in (2.35), (2.36) and (2.37),
with |ζi y “ |ξi y b |ηi y. Therefore, the convex cone generated by ten product states
|ζi yxζi | with i “ 1, . . . , 10 is a face of S1 , which is the dual face
ψt1 “ φ1t X pφt ˝ Tq1 (2.39)
of the positive map ψt . Because these ten product vectors span the whole space
C3 b C3 , we see that the map ψt has the bi-spanning property for t ‰ 1. Note that
ψt does not generate an extreme ray of P1 .
Even though ten vectors |ζ1 y, . . . , |ζ10 y are linearly dependent in C3 b C3 , the
corresponding ten states are linearly independent in M3 b M3 . To see this, we
suppose that 10
ř 3 3
i“1 αi |ζi yxζi | “ 0. Take |ωy P C b C among vectors in (2.38), then
we have
ÿ10 4
ÿ
0“ αi |ζi yxζi |ωy “ αi xζi |ωy|ζi y.
i“1 i“1
Because t|ζi y : i “ 1, 2, 3, 4u is linearly independent, we have αi xζi |ωy “ 0, and

αi “ 0 for i “ 1, 2, 3, 4. Since t|ζi y : i “ 5, 6, . . . , 10u is also linearly independent, we
have αi “ 0 for i “ 5, 6, . . . , 10.
Therefore, the separable state %t with t ‰ 1 is decomposed into the sum of pure
product states in a unique way. Especially, we need ten pure product states in order
to express %t as the sum of them. The length of a separable state % is defined as
the smallest number of pure product states whose convex sum is %. We see that the
length of the separable state %t is 10, which is strictly greater than 9 “ 3 ¨ 3. It is
known [101, 15] that the lengths of 2 b 2 and 2 b 3 separable states are less than or
equal to 4 and 6, respectively.
Linear independence of t|ζi yxζi | : i “ 1, 2, . . . , 10u has a geometric interpretation.
We restrict ourselves on the normalized states. Then the face ψt1 in (2.39) is the 9-
dimensional simplex with the ten extreme points which is normalized states |ζi yxζi |
with i “ 1, 2, . . . , 10. All the states in this face has also a unique decomposition into
the convex sum of pure product states.
References: [101], [1], [106], [2], [15], [46], [16], [47], [48]
106
2.6 Exposed positive maps by Woronowicz
In this section, we exhibit positive linear maps from M2 to M4 which generate
exposed rays of the convex cone P1 of all positive maps. For this purpose, we also
provide a sufficient condition for exposedness which is relatively easy to check.
2.6.1 A dimension condition for exposed positive maps

For a linear map φ : Mm Ñ Mn , we define the linear map φ̂ : Mm b Cn Ñ Cn by
φ̂pa b ξq “ φpaq|ξy, a P Mm , ξ P Cn ,
and the subspace Nφ in Mm b Cn by

`
Nφ “ span ta b |ηy P Mm b Cn : φpaq|ηy “ 0u.
Then we have Nφ Ă ker φ̂, in general. By the relation (2.11), we have
|ξyxξ| b |ηy P Nφ ðñ |ξy b |η̄y P P rφs,
for |ξy P Cm and |ηy P Cn .

`
Because S1 is generated by Mm b Mn` , we see that ψ P φ2 is equivalent to
`
a P Mm , |ηy P Cn , xa b |ηyxη|, φy “ 0 ùñ xa b |ηyxη|, ψy “ 0, (2.40)
with the duality between S1 and P1 . By the identity
xη|φpaq|ηy “ x|η̄yxη̄|, φpaqy “ xa b |η̄yxη̄|, φy,
we also see that a b |η̄yxη̄| P φ1 if and only if a b |ηy P Nφ for a P Mm

`
and |ηy P Cn .
Therefore, we see that
ψ P φ2 ðñ Nφ Ă Nψ .
If φ satisfies the condition
ker φ̂ “ Nφ , (2.41)
then we have ker φ̂ “ Nφ Ă Nψ Ă ker ψ̂. Therefore, we conclude that there exists
a linear map X : Cn Ñ Cn such that ψ̂ “ X ˝ φ̂, or equivalently, there exists
X P Mn such that ψpaq|ηy “ Xφpaq|ηy for each a P Mn and |ηy P Cm . This implies
ψpaq “ Xφpaq. Therefore, if a positive linear map φ : Mm Ñ Mn satisfies the
condition (2.41) together with the following condition
(C) If X P Mn and a ÞÑ Xφpaq is positive then X is a scalar matrix,
then we conclude that φ is exposed. Because pXφpaqq˚ “ Xφpa˚ q if and only if

Xφpaq “ φpaqX ˚ , we see that the following two conditions
107
(Ch ) If X P Mn and Xφpaq “ φpaqX ˚ for every a P Mm then X is a scalar matrix,
(C1h ) If X P Mn and a ÞÑ Xφpaq is Hermiticity preserving then X is a scalar matrix
are equivalent. It is clear that (Ch ) or equivalent condition (C1h ) implies (C).
For a subset S of Mn , we define the commutant S 1 by
S 1 :“ ta P Mn : ax “ xa for every x P Su.
Proposition 1.2.1 tells us that the commutant of Mn itself is trivial, that is, consists
of scalar matrices. It is clear that S 1 is a subalgebra of Mn , and it is a ˚-subalgebra
if S is ˚-invariant, that is, x P S implies x˚ P S. It is also clear that S Ă S 2 and
S 1 “ S 3 . A positive linear map φ : Mm Ñ Mn is called irreducible if the range
φpMm q of φ has the trivial commutant. Note that the range of a positive map
φ : Mm Ñ Mn is a ˚-invariant subspace, and so its commutant is a ˚-subalgebra of
Mn . Suppose that the range of an irreducible positive map φ contains the identity.
If we take a P Mm with φpaq “ Im then Xφpaq “ φpaqX ˚ implies that X “ X ˚ , and
so the irreducibility implies the condition (Ch ). Therefore, we have the following:
Theorem 2.6.1 (Woronowicz [131]) Suppose that φ : Mm Ñ Mn is an irreducible

positive linear map satisfying the condition (2.41). If the range of φ contains the
identity then φ generates an exposed ray of the convex cone P1 of all positive maps.
If the range of φ contains the idnetity then we see that the map φ̂ is surjective.
Therefore, we have dim ker φ̂ “ npm2 ´ 1q, and so the condition ker φ̂ “ Nφ in (2.41)
holds in Theorem 2.6.1 if and only if the following dimension condition
dim Nφ “ npm2 ´ 1q (2.42)
is satisfied. We call this the Woronowicz dimension condition.

It is easy to see that the condition (Ch ) of a positive map φ implies irreducibility.
To see this, we suppose that φ satisfies (Ch ) and Xφpaq “ φpaqX for every a P Mm .
Because the range of the positive map φ is ˚-invariant, we also have X ˚ φpaq “
φpaqX ˚ for every a P Mm . If we write Y “ 12 pX ` X ˚ q and Z “ 2i1 pX ´ X ˚ ),
then we have Y φpaq “ φpaqY “ φpaqY ˚ , which implies that Y is a scalar matrix by
condition (Ch ). By the same argument, we also see that Z is a scalar matrix, and
so we conclude that φ is irreducible. It should be noted that irreducibility does not
imply the condition (C) in general. To see this, we define φ : M2 Ñ M2 by
ˆ ˙ ˆ ˙
a b a a
φ: ÞÑ ,
c d a a`d
108
which is easily seen to be irreducible. But, we have
ˆ ˙ˆ ˙ ˆ ˙ ˆ ˙ˆ ˙
1 0 a a a a a a 1 1
“ “ ,
1 0 a a`d a a a a`d 0 0
and so we see that φ violates the condition (Ch ) as well as (C). It is straightforward
to check that the Woronowicz map φwo given by (1.56) satisfies the condition (Ch ),
and so it is irreducible.
In the remainder of this section, we show that the identity map idn : Mn Ñ Mn
satisfies the dimension condition (2.42). This will give rise to an another proof of
Theorem 2.1.8. We fix i “ 2, 3, . . . , n for a moment, and put
|ξy “ p1, α2 , . . . , αi , . . . , αn qT P Cn ,
|ηi y “ pᾱi , 0, . . . , 0, ´1, 0, . . . , 0qT P Cn ,
with ´1 at the i-th entry. Then we have |ξyxξ| b |ηi y P Nidn . We identity |ξyxξ| b |ηy
in Mn b Cn and |ξy b |ξy ¯ b |ηy in Cn b Cn b Cn to count dimensions. For each
i “ 2, . . . , n, we define the space Vi by
¯ b |ηi y : α2 , α3 , . . . , αn P Cu Ă Cn b Cn b Cn .
Vi “ span t|ξy b |ξy
Lemma 2.6.2 The set tαk ᾱ` : k, ` “ 0, 1, 2, . . . u of monomials is linearly indepen-

dent.
Proof. Suppose that P pαq is a linear combination of finitely many monomials

k `
among α ᾱ . Then P ptαq is a polynomial with respect to the variable t whose
coefficients are homogeneous polynomials in the variables α and ᾱ. Therefore, it is
enough to show that homogeneous polynomials with a fixed degree n are linearly
independent. We multiply αn to all of homogeneous polynomials of degree n, and
use the identity theorem on the unit circle, to get the conclusion. ˝
Because monomials with variables 1, αi and ᾱi , i “ 2, 3, . . . , n, are linearly
independent, we see that the dimension of Vi is the number of monomials which
¯ b |ηi y. Noting that |ξy
appear in the entries of |ξy b |ξy ¯ and |ηi y have n and 2
¯ b |ηi y has 2n ´ 1
variables and the variable ᾱi appears two times, we see that |ξy
monomials in the entries, and
dim Vi “ np2n ´ 1q, i “ 2, . . . , n.
The orthogonal complement ViK is spanned by orthogonal vectors
|j11y ` |jiiy, j “ 1, 2, . . . , n,
(2.43)
|xyzy, x, y “ 1, 2, . . . , n, z “ 2, . . . , i ´ 1, i ` 1, . . . , n,
109
whose cardinality is n ` n2 pn ´ 2q “ n3 ´ np2n ´ 1q.
In order to count the dimension of NidKn “ V2K X V3K X ¨ ¨ ¨ X VnK , we write down
the generators of ViK as follows
V2K : |111y ` |122y, |211y ` |222y, . . . , |n11y ` |n22y, |xy3y, |xy4y, ¨ ¨ ¨ , |xyny,
V3K : |111y ` |133y, |211y ` |233y, . . . , |n11y ` |n33y, |xy2y, |xy4y, ¨ ¨ ¨ , |xyny,
¨¨¨
VnK : |111y ` |1nny, |211y ` |2nny, . . . , |n11y ` |nnny, |xy2y, |xy3y, ¨ ¨ ¨ , |xypn ´ 1qy.
Then we see that V2K X V2K X ¨ ¨ ¨ X VnK is spanned by following n orthogonal vectors
|ζ1 y :“|111y ` |122y ` |133y ` ¨ ¨ ¨ ` |1nny,

|ζ2 y :“|211y ` |222y ` |233y ` ¨ ¨ ¨ ` |2nny,
¨¨¨
|ζn y :“|n11y ` |n22y ` |n33y ` ¨ ¨ ¨ ` |nnny.
Therefore, we have dim Nidn “ n3 ´ n which coincides with the dimension of ker id
xn .
By Theorem 2.6.1, we conclude that the identity map idn is exposed.
References: [130], [131], [49], [81]
2.6.2 Exposed positive maps between 2ˆ2 and 4ˆ4 matrices

The Woronowicz map φwo in (1.56) can be parameterized as follows: We first take
positive numbers a, b, c and d with ab ą 1, and we define positive numbers e, f, g, h
and k by the relations
ˆ ˙ ˆ ˙
e c
pab ´ 1q “ apc ` dq , g 2 “ acd, h “ be ´ c2 , k “ bf ´ d2 . (2.44)
f d
We define φ : M2 Ñ M4 by
¨ ˛
ˆ ˙ hx ´ cdpy ` zq ` kw ´gx ` gz 0 0
x y ˚ ´gx ` gy ax z 0 ‹
φ “˚ ‹ , (2.45)
z w ˝ 0 y bw ćz ´ dw‚
0 0 ćy ´ dw ex ` f w
If pa, b, c, dq “ p2, 2, 2, 1q, then we get the Woronowicz map φwo .

For each complex number α, we consider the following positive rank one matrix
ˆ ˙
1 ᾱ
|ξα yxξα | “
α |α|2
onto the vector |ξα y “ p1, αqT P C2 , and consider the determinant ∆i pαq of right-
below i ˆ i submatrix of φp|ξα yxξα |q for each i “ 1, 2, 3, 4. By a direct calculation,
110
we have
∆1 pαq “ e ` f |α|2 ,
“ ‰
∆2 pαq “ |α|2 h ´ cdpα ` ᾱq ` k|α|2 ,
∆3 pαq “ acd|α|2 |1 ´ α|2 ,
∆4 pαq “ 0.
By the relations between e, f, g, h and k in (2.44), we see that
∆1 pαq ą 0, ∆2 pαq ą 0, ∆3 pαq ą 0, ∆4 pαq “ 0
for any complex numbers α ‰ 0, 1, and so, we see that φp|ξα yxξα |q is positive for
α ‰ 0, 1. We also have
¨ ˛
h ´g 0 0
˚´g a 0 0‹
φp|ξ0 yxξ0 |q “ ˚
˝0
‹
0 0 0‚
0 0 0 e
and ¨ ˛
h ´ 2cd ` k 0 0 0
˚ 0 a 1 0 ‹
φp|ξ1 yxξ1 |q “ ˚ ‹.
˝ 0 1 b ć ´ d‚
0 0 ć ´ d e ` f
We write |ξ8 y :“ p0, 1qt P C2 , then we have
¨ ˛
k 0 0 0
˚0 0 0 0‹
φp|ξ8 yxξ8 |q “ ˚
˝0
‹.
0 b ´d‚
0 0 ´d f
Since φp|ξα yxξα |q is positive for every α P C Y t8u, we conclude that φ is a positive
map from M2 into M4 .
It is straightforward to see that φ satisfies the condition (Ch ). Take a matrix
σ “ rσij s P M4 , and suppose that
ˆ ˙ ˆ ˙
x y x y
rσij s φ “φ rσ̄ji s
z w z w
for every x, y, z, w P C. Compare the entries of both sides, one may check that σ is
a scalar matrix.
We also see that φp|ξα yxξα |q P M4 has the one dimensional kernel space which is
generated by ¨ ˛
gαp1 ´ αq
˚α rh ´ cdpα ` ᾱq ` k|α|2 s‹
|ηα y :“ ˚ ‹ P C4 (2.46)
˝ é ´ f |α|2 ‚
´ᾱpc ` dαq
111
for α P C, and the kernel of φp|ξ8 yxξ8 |q is spanned by
|η8 y “ p0, 1, 0, 0qT .
Now, we proceed to determine the dimension of the space Nφ , or equivalently,
the dimension of the space
span t|ζα y :“ |ξα y b |ξ¯α y b |ηα y : α P C Y t8uu Ă C2 b C2 b C4 .
We first note that each entries of |ζα y are linear combinations of the following 12
monomials
1, α, ᾱ, α2 , αᾱ, ᾱ2 , α3 , α2 ᾱ, αᾱ2 , α3 ᾱ, α2 ᾱ2 , α3 ᾱ2 . (2.47)
In order to show that φ satisfies the condition (2.42), we consider the vector
pf1 pαq, f2 pαq, . . . , fn pαqq P Cn , (2.48)
for complex functions f1 , . . . , fn . In order to find the dimension of the span of the
vectors in (2.48) through α P C, it suffices to consider the rank of the ‘coefficient
matrix’. The coefficient matrix of |ξα y|ξ¯α y is given by
¨ ˛
1 ¨ ¨ ¨ ¨ ¨ ¨¨¨
˚ ¨ ¨ 1 ¨ ¨ ¨ ¨ ¨ ¨‹
˚ ‹
˝ ¨ 1 ¨ ¨ ¨ ¨ ¨ ¨ ¨‚
¨ ¨ ¨ ¨ 1 ¨ ¨¨¨
with respect to the monomials 1, α, ᾱ, α2 , αᾱ, ᾱ2 , . . . , which is rank four. On the
other hand, the coefficient matrix for |ηα y is given by
¨ ˛
¨ g ¨ ´g ¨ ¨ ¨ ¨ ¨ ¨¨¨
˚ ¨ h ¨ ćd ćd ¨ ¨ k ¨ ¨ ¨ ¨‹
˚ ‹
˝é ¨ ¨ ¨ ´f ¨ ¨ ¨ ¨ ¨ ¨ ¨‚
¨ ¨ ć ¨ ´d ¨ ¨ ¨ ¨ ¨¨¨
with respect to monomials 1, α, ᾱ, α2 , αᾱ, ᾱ2 , α3 , α2 ᾱ, . . . . This is of rank 4, which
means that the set t|ηα y : α P Cu spans the whole space C4 . Now, one may produce
the coefficient matrix of |ζα y :“ |ξα y b |ξ¯α y b |ηα y with respect to the monomials in
(2.47) which is 16 ˆ 12 matrix, and check that this matrix has rank 12. This shows
that the map φ satisfies the Woronowicz dimension condition (2.42), and we have
the following:
Theorem 2.6.3 The map φ : M2 Ñ M4 defined by (2.44) and (2.45) is an inde-
composable exposed positive map for every positive a, b, c and d with ab ą 1.
Proof. It remains to show that the map φ is not decomposable. If it were decom-
posable, then φ “ Ads or φ “ Ads ˝T. Then either Cφ or CΓφ is of rank one, which
is not the case. ˝
References: [49]
112
2.7 Exposed positive maps by Robertson
In this section, we exhibit the Robertson map between 4 ˆ 4 matrices which is an
indecomposable positive map generating an exposed ray. The construction use the
notion of Jordan homomorphisms.
2.7.1 Robertson’s positive maps between 4 ˆ 4 matrices

In this section, we construct an example of indecomposable exposed positive map
between 4 ˆ 4 matrices. We begin with the map τ2,1
ˆ ˙ ˆ ˙
α β δ ´β
τ2,1 “ Tr 2 ´ id2 : ÞÑ
γ δ ´γ α
between 2 ˆ 2 matrices, as it was defined in (1.52). This is an anti-automorphism of

order two, that is, it satisfies τ2,1 pxyq “ τ2,1 pyqτ2,1 pxq for x, y P M2 and τ2,1 ˝τ2,1 “ id,
as well as it is a completely copositive map.
We recall that quaternion numbers may be expressed as
"ˆ ˙ *
α β
H“ P M2 pCq : α, β P C . (2.49)
´β̄ ᾱ
In other words, the quaternion a ` bi ` cj ` dk P H corresponds to

ˆ ˙
a ` ib c ` id
a ` bi ` cj ` dk ÐÑ P M2 pCq.
ć ` id a ´ ib
It is worthwhile to note the following correspondences

ˆ ˙ ˆ ˙ ˆ ˙
i 0 0 1 0 i
iØ “ iσz , jØ “ iσy , kØ “ iσx ,
0 í ´1 0 i 0
in terms of Pauli matrices in (1.16).

We define the linear map π : M4 pCq Ñ M4 pCq by
1
π “ pid4 ` T2 b τ2,1 q,
2
that is, we define
ˆ ˙ ˆ ˙ ˆ ˙
x y 1 x ` τ2,1 pxq y ` τ2,1 pzq x y
π “ , P M4 pCq.
z w 2 z ` τ2,1 pyq w ` τ2,1 pwq z w
Then π is a positive map since τ2,1 is completely copositive. With the expression
(2.49), we have
H “ tx P M2 pCq : τ2,1 pxq “ x˚ u,
113
and one may check that the range of M4h pCq under π is
"ˆ ˙ *
h x y h
M2 pHq “ : x, y P H X M2 pCq, w P H
y˚ w
$¨ ˛ ,
’
’ a 0 α β /
/ (2.50)
0 a ´β̄ ᾱ‹
&˚ ‹ .
“ ˚ : a, b P R, α, β P C Ă M4h pCq,
’
’
˝ ᾱ ´β b 0 ‚ /
/
β̄ α 0 b
% -
and π|M2h pHq is the identity. In short, π is a positive projection from M4h pCq onto
M2h pHq, which is a six dimensional subspace of the real space M4h pCq.
We note that M2h pHq is a JC-subalgebra of M4 pCq, that is, it is a real subspace
of Hermitian matrices in M4h pCq which is closed under the binary operation
1
a ˝ b “ pab ` baq.
2
By the relation
1
a ˝ b “ rpa ` bq2 ´ a2 ´ b2 s,
2
we see that a real subspace of Hermitian matrices is a JC-algebra if and only if it
is closed under the operation of square. Furthermore, a linear map φ between JC-
algebras is a Jordan homomorphism, that is, φpa ˝ bq “ φpaq ˝ φpbq if and only if it
preserves the operation of square. Because a Hermitian matrix is positive if and only
if it is the square of a Hermitian matrix, we see that every Jordan homomorphism
defined on the full matrix algebra is positive.
We define the Jordan automorphism θ : M2h pHq Ñ M2h pHq by
ˆ ˙ ˆ ˙
x y w y
θ ˚ “
y w y˚ x
of order two. This is also positive, since the 4 ˆ 4 matrix in (2.50) is positive if and
only if ab ě |α|2 ` |β|2 . Finally, we define the map
φrb :“ θ ˝ π
on M4h pCq, which is an extension of θ defined on M2h pHq to the whole Hermitian
parts M4h pCq. See Figure 2.10. We extend the map φrb to the whole matrix algebra
M4 pCq by (1.11). Then we have
ˆ ˙ ˆ ˙
x y 1 w ` τ2,1 pwq y ` τ2,1 pzq
φrb “
z w 2 z ` τ2,1 pyq x ` τ2,1 pxq
ˆ ˙ (2.51)
1 Tr 2 pwq y ` τ2,1 pzq
“ ,
2 z ` τ2,1 pyq Tr 2 pxq
114
M4h pCq P
PP
P PφP
rb
π P PP
? P P
q
P
M2h pHq θ - M h pHq ãÑ M h pCq
2 4
Figure 2.10: The Robertson map φrb is an extension of Jordan automorphism θ.
for x, y, z and w in M2 pCq. This map φrb is called the Robertson map which is a
positive map between M4 pCq. It should be noted that the map
ˆ ˙ ˆ ˙
x y w y
ÞÑ
z w z x
is not a positive map on the whole matrix algebra M4 pCq, although it is positive on
the JC-subalgebra M2h pHq of M4 pCq.
The Choi matrix of 2φrb is a 16 ˆ 16 matrix, whose first four rows coming from
φrb p|1yxj|q with j “ 1, 2, 3, 4 are given by
» fi
¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ 1 ¨ ¨ ¨ ¨ 1
—¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ffi
— ffi .
–¨ ¨ 1 ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ´1 ¨ ¨ fl
¨ ¨ ¨ 1 ¨ ¨ ¨ ¨ ¨ 1 ¨ ¨ ¨ ¨ ¨ ¨
The next four rows are given by

» fi
¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨
—¨ ¨ ¨
— ¨ ¨ ¨ ¨ ¨ ¨ ¨ 1 ¨ ¨ ¨ ¨ 1ffi
ffi .
–¨ ¨ ¨ ¨ ¨ ¨ 1 ¨ ¨ ¨ ¨ ¨ 1 ¨ ¨ ¨ fl
¨ ¨ ¨ ¨ ¨ ¨ ¨ 1 ´1 ¨ ¨ ¨ ¨ ¨ ¨ ¨
The other rows are given by

» fi
¨ ¨ ¨ ¨ ¨ ¨ ¨ ´1 1 ¨ ¨ ¨ ¨ ¨ ¨ ¨
—¨ ¨ ¨ 1 ¨ ¨ ¨ ¨ ¨ 1 ¨ ¨ ¨ ¨ ¨ ¨ffi
— ffi
–1 ¨ ¨ ¨ ¨ 1 ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨fl
¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨
and » fi
¨ ¨ ¨ ¨ ¨ ¨ 1 ¨ ¨ ¨ ¨ ¨ 1 ¨ ¨ ¨
—¨
— ¨ ´1 ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ 1 ¨ ¨ffi
ffi .
–¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨fl
1 ¨ ¨ ¨ ¨ 1 ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨
In order to show that φrb is not decomposable, we define
ˆ ˙ ˆ ˙ ˆ ˙ ˆ ˙
0 i 0 j 0 k 0 1
σ1 “ ˚ , σ2 “ ˚ , σ3 “ , σ4 “ ,
i 0 j 0 k˚ 0 1 0
115
where 1 is the identity for quaternion. We also put
ˆ ˙
1 0
σ5 “ σ1 σ2 σ3 σ4 “ .
0 ´1
Then we have
φrb pσi q “ σi , i “ 1, 2, 3, 4, φrb pσ5 q “ ´σ5 .
Now, we define Σ P M4 pM4 q by

¨ ˛
0 0 0 σ1
˚σ4 0 0 0 ‹
Σ“˚
˝ 0 σ3 0 0 ‚.
‹
0 0 σ2 0
Then we have ¨ ˛
σ5 0 0 0
˚ 0 ´σ5 0 0 ‹
Σ4 “ ˚
˝0
‹, Σ8 “ I16 ,
0 σ5 0 ‚
0 0 0 ´σ5
and eigenvalues of Σ consist of eighth roots of unity. Define
?
%˘ “ Σ ` Σ˚ ˘ Σ4 ` p1 ` 2qI16
¨ ˛
˘σ5 σ4˚ 0 σ1
˚ σ4 ¯σ5 σ3˚ 0 ‹ ?
“˚˝ 0 ˚ ‚ ` p1 `
‹ 2qI16 ,
σ3 ˘σ5 σ2
σ1˚ 0 σ2 ¯σ5
then we see that

pid4 b φrb qp%` q “ %´
The eigenvalues of %˘ are given by

?
tω ` ω̄ ˘ ω 4 ` 1 ` 2 : ω 8 “ 1, ω P Cu.
Therefore, we see that %` is of PPT, because the partial transpose of %` is nothing

but its conjugate. We also note that pid4 b φrb qp%` q “ %´ is not positive. By
Corollary 1.8.6 (i), we conclude that φrb is not decomposable.
Theorem 2.7.1 The map φrb defined in (2.51) is an indecomposable positive map.
References: [95], [96], [97], [98], [118]
116
2.7.2 Exposedness of the Robertson map
In this section, we show that the Robertson map φrb generates an exposed ray of
he convex cone P1 rM4 , M4 s. First of all, it is straightforward to check that the map
φrb is irreducible. ˆ ˙
I2 0
We write V “ P M4 . Then we have
0 Í2
ˆ ˙ ˆ ˙
x y 1 Tr 2 pwq ý ´ τ2,1 pzq
AdV ˝φrb “ .
z w 2 ´z ´ τ2,1 pyq Tr 2 pxq
We note that
ˆ ˙
Tr 2 pwq ý ´ τ2,1 pzq
´z ´ τ2,1 pyq Tr 2 pxq
ˆ ˙ ˆ ˙ ˆ ˙
Tr 2 pxq ` Tr 2 pwq 0 x y τ2,1 pxq τ2,1 pzq
“ ´ ´
0 Tr 2 pxq ` Tr 2 pwq z w τ2,1 pyq τ2,1 pwq
ˆ ˙
˚ T 0 1
We also note that τ2,1 pxq “ u x u for x P M2 , with u “ P M2 , and so, it
´1 0
follows that
1` ˘
AdV ˝φrb pXq “ Tr 4 pXq ´ X ´ U ˚ X T U
2
with ¨ ˛
ˆ ˙ ¨ 1 ¨ ¨
u 0 ˚´1 ¨ ¨ ¨ ‹
U“ “˚ ˝ ¨ ¨ ¨ 1‚ P M4 .
‹ (2.52)
0 u
¨ ¨ ´1 ¨
We note that both φrb and AdV ˝φrb are unital, and the map AdV ˝φrb is also
irreducible. We proceed to show that AdV ˝φrb is also exposed, which implies that
φrb is exposed.
Theorem 2.7.2 For the matrix U in (2.52), the map
φU :“ Tr 4 ´ id4 ´ AdU ˝T (2.53)
is exposed.
Proof. We show that φU satisfies the dimension condition (2.42). We note that
¯ ξ|U,
φp|ξyxξ|q “ I4 ´ |ξyxξ| ´ U ˚ |ξyx ¯
for a unit vector |ξy P C4 . Because |ξy and U ˚ |ξy¯ are mutually orthogonal, we see
¯ Therefore,
that φp|ξyxξq|ηy “ 0 if and only if |ηy belongs to the span of |ξy and U ˚ |ξy.
we have to show that
¯ b |ξy, |ξy b |ξy
Nφ “ span t|ξy b |ξy ¯ b U ˚ |ξy
¯ P C4 b C4 b C4 : |ξy P C4 u
117
is of 60 dimensional subspace of C4 b C4 b C4 . To see this, we write |ξy “
p1, α, β, γqT P C4 . Then the following 16 monomials
1, α, β, γ,
ᾱ, αᾱ, β ᾱ, γ ᾱ,
(2.54)
β̄, αβ̄, β β̄, γ β̄,
γ̄, αγ̄, βγ̄, γγ̄
¯ To list up monomials appearing in the
appear to write down the entries of |ξy b |ξy.
¯ b |ξy, we multiply the monomials in the first row of (2.54) with
entries of |ξy b |ξy
T
p1, α, β, γq to get ten monomials
1, α, β, γ, α2 , β 2, γ 2, αβ, βγ, γα.
We also get ten monomials from each row in (2.54), and so we have 40 monomials
¯ b |ξy. Therefore, we see that the span V of |ξy b |ξy
in the entries of |ξy b |ξy ¯ b |ξy
is of 40 dimension. We have
¯
|ξy|ξy|ξy “ p1, α, β, γqT b p1, ᾱ, β̄, γ̄qT b p1, α, β, γqT ,
and so we also see that vectors in the family
t|ijky ´ |kjiy : i, j, k “ 1, 2, 3, 4, i ‰ ku
belong to V K . These 24 vectors, up to scalar multiplications, are easily seen to

be mutually orthogonal, and so we see that it is a basis of V K . The orthogonal
complement W K of the space W spanned by
¯ b U ˚ |ξy
|ξy b |ξy ¯ “ p1, α, β, γqT b p1, ᾱ, β̄, γ̄qT b p´ᾱ, 1, ´γ̄, β̄qT
has the following 24 vectors
|i11y ` |i22y, |i33y ` |i44y,

|i13y ` |i42y, |i14y ´ |i32y,
|i23y ´ |i41y, |i24y ` |i31y,
with i “ 1, 2, 3, 4. They make a basis of W K .

One may check directly that V K X W K is a four dimensional subspace of C4 b
C4 b C4 which is spanned by the following four vectors
|114y ´ |411y ` |224y ´ |422y ` |231y ´ |132y,

|423y ´ |324y ` |133y ´ |331y ` |144y ´ |441y,
|233y ´ |332y ` |244y ´ |442y ` |314y ´ |413y,
|142y ´ |241y ` |113y ´ |311y ` |223y ´ |322y.
118
Hence, we conclude that dimpV ` W q “ 60, as it was required. ˝
Therefore, we see that the Robertson map φrb generates an exposed ray of the
convex cone P1 rM4 , M4 s of all positive maps between 4 ˆ 4 matrices. It is known
[28] that the map φU defined in (2.53) is exposed whenever U is a anti-symmetry.
We note that
φU “ τ4,1 ´ AdU ˝T
with the map τ4,1 in (1.51). The map φU is usually called the Breuer–Hall map. For
further related examples of positive maps, see [103] and Section 8 of [30].
References: [12], [51], [28], [103], [30]
119
120
Bibliography
[1] E. Alfsen and F. Shultz, Unique decompositions, faces, and automorphisms of

separable states, J. Math. Phys. 51 (2010), 052201.
[2] E. Alfsen and F. Shultz, Finding decompositions of a class of separable states,

Linear Alg. Appl. 437 (2012), 2613–2629.
[3] T. Ando, Cones and norms in the tensor product of matrix spaces, Linear Alg.
Appl. 379 (2004), 3–41.
[4] W. Arveson, Subalgebras of C ˚ -algebras, Acta Math. 123 (1969), 141–224.
[5] W. Arveson, Quantum channels that preserve entanglemen, Math. Ann. 343
(2009), 757–771.
[6] G. Aubrun and S. J. Szarek, Two proofs of Størmer’s theorem, arXiv

1512.03293.
[7] G. Aubrun and S. J. Szarek, “Alice and Bob Meet Banach: The Interface
of Asymptotic Geometric Analysis and Quantum Information Theory”, Math.
Surveys Monog. Vol 223, Amer. Math. Soc., 2017.
[8] I. Bengtsson and K. Życzkowski, “Geometry of Quantum States”, Cambridge

Univ. Press, 2006.
[9] B. V. R. Bhat, Linear maps respecting unitary conjugation, Banach J. Math.

Anal. 5 (2011), 1-5.
[10] B. V. R. Bhat and H. Osaka, A factorization property of positive maps on

C ˚ -algebras, Intern. J. Quantum Inform. 18 (2020), 2050019.
[11] R. Bhatia, “Matrix Analysis”, Grad. Texts Math. Vol. 169, Springer-Verlag,
1997.
[12] H.-P. Breuer, Optimal Entanglement Criterion for Mixed Quantum States,
Phys. Rev. Lett. 97 (2006), 080501.
121
[13] E.-S. Byeon and S.-H. Kye, Facial structures for positive linear maps in the two
dimensional matrix algebra, Positivity, 6 (2002), 369–380.
[14] D. Cariello, Inequalities for the Schmidt number of bipartite states, Lett. Math.
Phys. 110 (2020), 827–833.
[15] L. Chen and D. Ž. Djoković, Qubit-qudit states with positive partial transpose,
Phys. Rev. A 86 (2012), 062332.
[16] L. Chen and D. Ž. Djoković, Dimensions, lengths and separability in finite-
dimensional quantum systems, J. Math. Phys. 54 (2013), 022201.
[17] L. Chen, Y. Yang and W.-S. Tang, Schmidt number of bipartite and multipartite
states under local projections, Quantum Inform. Process 16 (2017), 75.
[18] S.-J. Cho, S.-H. Kye and S. G. Lee, Generalized Choi maps in 3-dimensional
matrix algebras, Linear Alg. Appl. 171 (1992), 213–224.
[19] H.-S. Choi and S.-H. Kye, Facial structures for separable states, J. Korean
Math. Soc. 49 (2012), 623–639.
[20] M.-D. Choi, Positive linear maps on C ˚ -algebras, Canad. Math. J. 24 (1972),
520–529.
[21] M.-D. Choi, Completely positive linear maps on complex matrices, Linear Alg.
Appl. 10 (1975), 285–290.
[22] M.-D. Choi, Positive semidefinite biquadratic forms, Linear Alg. Appl. 12
(1975), 95–100.
[23] M.-D. Choi, Some assorted inequalities for positive maps on C ˚ -algebras, J.
Operator Theory 4 (1980), 271–285.
[24] M.-D. Choi, Positive linear maps, Operator Algebras and Applications
(Kingston, 1980), pp. 583–590, Proc. Sympos. Pure Math. Vol 38. Part 2, Amer.
Math. Soc., 1982.
[25] M.-D. Choi and T.-T. Lam, Extremal positive semidefinite forms, Math. Ann.
231 (1977), 1–18.
[26] M. Christandl, A. Müller-Hermes and M. M. Wolf, When do composed maps

become entanglement breaking? Ann. Henri Poincaré 20 (2019), 2295–2322.
[27] D. Chruściński and A. Kossakowski, On Partially Entanglement Breaking Chan-

nels, Open Sys. Inform. Dynam. 13 (2006), 17–26.
122
[28] D. Chruściński and G. Sarbicki, Exposed positive maps in M4 pCq, Open Sys.
Inform. Dynam. 19 (2012), 1250017.
[29] D. Chruściński and G. Sarbicki, Optimal entanglement witnesses for two qutrits,
Open Sys. Inform. Dynam. 20 (2013), 1350006.
[30] D. Chruściński and G. Sarbicki, Entanglement witnesses: construction, analysis

and classification, J. Phys. A: Math. Theor. 47 (2014), 483001.
[31] D. Chruściński and F. A. Wudarski, Geometry of entanglement witnesses for

two qutrits, Open Sys. Inform. Dynam. 18 (2011), 375–387.
[32] J. de Pillis, Linear transformations which preserve Hermitian and positive

semidefinite operators, Pacific J. Math. 23 (1967), 129–137.
[33] E. G. Effros, Aspects of non-commutative order, C ˚ -Algebras and Applications

to Physics (Los Angeles, 1977), pp. 1–40, Lecture Note in Math. Vol. 650,
Spronger, 1978.
[34] M.-H. Eom and S.-H. Kye, Duality for positive linear maps in matrix algebras,
Math. Scand. 86 (2000), 130–142.
[35] S. Friedland, Infinite dimensional generalizations of Choi’s Theorem, Spec. Ma-

trices 7 (2019); 67–77.
[36] M. Girard, S.-H. Kye and E. Størmer, Convex cones in mapping spaces between
matrix algebras, Linear Alg. Appl. 608 (2021), 248–269.
[37] S. Gudder, Operator isomorphisms on Hilbert space tensor products, arXiv

2010.15901.
[38] O. Gühne and G. Toth, Entanglement detection, Phys. Rep. 474 (2009), 1–75.
[39] L. Gurvits, Classical complexity and quantum entanglement, J. Comput. System

Sci. 69 (2004), 448–484.
[40] L. Gurvits and H. Barnum, Largest separable balls around the maximally mixed
bipartite quantum state, Phys. Rev. A 66 (2002), 062311.
[41] K.-C. Ha, Atomic positive linear maps in matrix algebras, Publ. Res. Inst. Math.
Sci. 34 (1998), 591–599.
[42] K.-C. Ha, Notes on extremality of the Choi map, Linear Alg. Appl. 439 (2013),
3156–3165.
123
[43] K.-C. Ha and S.-H. Kye, Construction of entangled states with positive partial
transposes based on indecomposable positive linear maps, Phys. Lett. A 325
(2004), 315–323.
[44] K.-C. Ha and S.-H. Kye, One parameter family of indecomposable optimal en-
tanglement witnesses arising from generalized Choi maps, Phys. Rev. A, 84
(2011), 024302.
[45] K.-C. Ha and S.-H. Kye, Optimality for indecomposable entanglement witnesses,
Phys. Rev. A 86 (2012), 034301.
[46] K.-C. Ha and S.-H. Kye, Entanglement witnesses arising from Choi type positive
linear maps, J. Phys. A: Math. Theor. 45 (2012), 415305.
[47] K.-C. Ha and S.-H. Kye, Exposedness of Choi type entanglement witnesses
and applications to lengths of separable states, Open Syst. Inform. Dynam. 20
(2013), 1350012.
[48] K.-C. Ha and S.-H. Kye, Separable states with unique decompositions, Commun.
Math. Phys. 328 (2014), 131–153.
[49] K.-C. Ha and S.-H. Kye, Construction of exposed indecomposable positive linear
maps between matrix algebras, Linear Multilinear Alg. 64 (2016), 2188–2198.
[50] E. Haapasalo, The Choi–Jamiolkowski isomorphism and covariant quantum

channels, Quantum Stud.: Math. Found. 8 (2021), 351–373.
[51] W. Hall, A new criterion for indecomposability of positive maps, J. Phys. A:

Math. Theor. 39 (2006) 14119–14131.
[52] K. H. Han and S.-H, Kye, Various notions of positivity for bi-linear maps and
applications to tri-partite entanglement, J. Math. Phys. 57 (2016), 015205.
[53] K. H. Han and S.-H. Kye, Construction of multi-qubit optimal genuine entan-
glement witnesses, J. Phys. A: Math. Theor. 49 (2016), 175303.
[54] K. H. Han and S.-H. Kye, Choi matrices revisited, II, arXiv 2307.09247.
[55] A.-L. K. Hashagen, Symmetry Methods in Quantum Information Theory, Dis-

sertation, Tech. Univ. München, 2018.
[56] A. S. Holevo, Quantum coding theorems, Russian Math. Surveys 53 (1998),

1295–1331.
124
[57] A. S. Holevo, Entropy gain and the Choi–Jamiolkowski correspondence for infi-
nite dimensional quantum evolutions, Theor. Math. Phys. 166 (2011), 123–138.
[58] A. S. Holevo, The Choi–Jamiolkowski forms of quantum Gaussian channels, J.

Math. Phys. 52 (2011), 042202.
[59] M. Horodecki, P. Horodecki, Reduction criterion of separability and limits for

a class of distillation protocols, Phys. Rev. A 59 (1999), 4206–4216.
[60] M. Horodecki, P. Horodecki and R. Horodecki, Separability of mixed states:

necessary and sufficient conditions, Phys. Lett. A 223 (1996), 1–8.
[61] M. Horodecki, P. W. Shor and M. B. Ruskai, Entanglement braking channels,

Rev. Math. Phys. 15 (2003), 629–641.
[62] P. Horodecki, Separability criterion and inseparable mixed states with positive
partial transposition, Phys. Lett. A 232 (1997), 333–339.
[63] P. Horodecki, M. Lewenstein, G. Vidal and I. Cirac, Operational criterion and

constructive checks for the separability of low rank density matrices, Phys. Rev.
A 62 (2000), 032310.
[64] R. Horodecki, P. Horodecki, M. Horodecki and K. Horodecki, Quantum entan-

glement, Rev. Mod. Phys. 81 (2009), 865–942.
[65] M. Huber, L. Lami, C. Lancien and A. Múller-Hermes, High-Dimensional En-

tanglement in States with Positive Partial Transposition, Phys. Rev. Lett. 121
(2018), 200503.
[66] T. Itoh, Positive maps and cones in C ˚ -algebras, Math. Japonica 31 (1986),
607–616.
[67] A. Jamiolkowski, Linear transformations which preserve trace and positive

semidefinite operators, Rep. Math. Phys. 3 (1972), 275–278.
[68] Y.-H. Kiem, S.-H. Kye and J. Lee, Existence of product vectors and their partial
conjugates in a pair of spaces, J. Math. Phys. 52 (2011), 122201.
[69] Y.-H. Kiem, S.-H. Kye and J. Na, Product vectors in the ranges of multi-partite
states with positive partial transposes and permanents of matrices, Commun.
Math. Phys. 338 (2015), 621–639.
[70] E. Kirchberg, The “Strong Kadison inequality” of Woronowicz, Rep. Math.

Phys. 18 (1980), 111–116.
125
[71] P. Kopszak, M. Mozrzymas and M. Studziński, Positive maps from irreducibly
covariant operators, J. Phys. A: Math. Theor. 53 (2020), 395306.
[72] K. Kraus, General State Changes in Quantum Theory, Ann. of Phys. 64 (1971),
311–335.
[73] K. Kraus, Operations and effects in the Hilbert space formulation of quantum
theory, Foundations of quantum mechanics and ordered linear spaces (Marburg,
1973), pp. 206–229, Lecture Notes in Phys., Vol. 29, Springer, 1974.
[74] S.-H. Kye, Facial structures for positive linear maps between matrix algebras,
Canad. Math. Bull. 39 (1996), 74–82.
[75] S.-H. Kye, Boundaries of the cone of positive linear maps and subcones in matrix
algebras, J. Korean Math. Soc. 33 (1996), 669–677.
[76] S.-H. Kye, On the convex set of all completely positive linear maps in matrix
algebras, Math. Proc. Cambridge Philos. Soc. 122 (1997), 45–54.
[77] S.-H. Kye, Facial structures for various notions of positivity and applications
to the theory of entanglement, Rev. Math. Phys. 25 (2013), 1330002.
[78] S.-H. Kye, Three-qubit entanglement witnesses with the full spanning properties,
J. Phys. A: Math. Theor. 48 (2015), 235303.
[79] S.-H. Kye, Choi matrices revisited, J. Math. Phys. 63 (2022), 092202.
[80] S.-H. Kye, Compositions and tensor products of linear maps between matrix
algebras, Linear Alg. Appl. 658 (2023), 283–309.
[81] S.-H. Kye, Exposedness of elementary positive maps between matrix algebras,
arXiv 2301.05788.
[82] M. Lewenstein, B. Kraus, J. Cirac and P. Horodecki, Optimization of entangle-

ment witness, Phys. Rev. A 62 (2000), 052310.
[83] M. Lewenstein, B. Kraus, P. Horodecki and J. Cirac, Characterization of sepa-

rable states and entanglement witnesses, Phys. Rev. A 63 (2000), 044304.
[84] Y. Li and H.-K. Du, Interpolations of entanglement breaking channels and equiv-
alent conditions for completely positive maps, J. Funct. Anal. 268 (2015), 3566–
3599.
[85] B. Magajna, Cones of completely bounded maps, Positivity 25 (2021), 1–29.
126
[86] M. Marciniak, Rank properties of exposed positive maps, Linear Multilinear Alg.
61 (2013), 970–975.
[87] L. Molnár, Order–automorphisms of the set of bounded observables, J. Math.

Phys. 42 (2001), 5904.
[88] M. A. Nielsen and I. L. Chuang, “Quantum Computation and Quantum Infor-

mation”, Cambridge Univ. Press, 2000.
[89] H. Osaka, Indecomposable Positive Maps in Low Dimensional Matrix Algebras,

Linear Alg. Appl. 153 (1991), 73-83.
[90] H. Osaka, A series of absolutely indecomposable positive maps in matrix alge-

bras, Linear Alg. Appl. 186 (1993), 45–53.
[91] S. K. Pandey , V. I. Paulsen , J. Prakash and M. Rahaman, Entanglement

breaking rank and the existence of SIC POVMs, J. Math. Phys. 61 (2020),
042203.
[92] V. Paulsen, “Completely Bounded Maps and Operator Algebras”, Cambridge

Univ. Press, 2002.
[93] V. I. Paulsen and F. Shultz, Complete positivity of the map from a basis to its
dual basis, J. Math. Phys. 54 (2013), 072201.
[94] A. Peres, Separability Criterion for Density Matrices, Phys. Rev. Lett. 77
(1996), 1413–1415.
[95] A. G. Robertson, Automorphisms of spin factors and the decomposition of pos-

itive maps, Quart. J. Math. Oxford 34 (1983), 87–96.
[96] A. G. Robertson, Positive extensions of automorphisms of spin factors, Proc.

Royal Soc. Edinburgh A, 94 (1983), 71–77.
[97] A. G. Robertson, Schwarz inequalities and the decomposition of positive maps

on C ˚ -algebras, Math. Proc. Cambridge Philos. Soc. 94 (1983), 291–296.
[98] A. G. Robertson, Positive projections on C ˚ -algebras and extremal positive

maps, J. London Math. Soc. 32 (1985), 133–140.
[99] R. T. Rockafellar, “Convex Analysis”, Princeton Univ. Press, 1970.
[100] A. Sanpera, D. Bruß and M. Lewenstein, Schmidt number witnesses and bound
entanglement, Phys. Rev. A 63 (2001), 050301.
127
[101] A. Sanpera, R. Tarrach and G. Vidal, Local description of quantum insepara-
bility, Phys. Rev. A 58 (1998), 826–830.
[102] G. Sarbicki, General theory of detection and optimality, arXiv:0905.0778.
[103] G. Sarbicki and D. Chruściński, A class of exposed indecomposable positive

maps, J. Phys. A: Math. Theor. 46 (2013), 015306.
[104] R. Šemrl and R. Sourour, Order preserving maps on Hermitian matrices, J.

Austral. Math. Soc. 95 (2013), 129–132.
[105] S. Singh and I. Nechita, Diagonal unitary and orthogonal symmetries in quan-
tum theory, Quantum 5 (2021), 519.
[106] L. Skowronek, Dualities and positivity in the study of quantum entanglement,

Intern. J. Quant. Inform. 8 (2010), 721–754.
[107] L. Skowronek, Cones with a mapping cone symmetry in the finite-dimensional

case, Linear Alg. Appl. 435 (2011), 361–370.
[108] L. Skowronek, E. Størmer, and K. Życzkowski, Cones of positive maps and

their duality relations, J. Math. Phys. 50, (2009), 062106.
[109] H. Schneider, Positive operators and an inertia theorem, Numer. Math. 7

(1965), 11–17.
[110] W. F. Stinespring, Positive functions on C ˚ -algebras, Proc. Amer. Math. Soc.

6, (1955), 211–216.
[111] E. Størmer, Positive linear maps of operator algebras, Acta Math. 110 (1963),
233–278.
[112] E. Størmer, Decomposition of positive projections on C ˚ -algebras, Math. Ann.

247 (1980), 21–41.
[113] E. Størmer, Decomposable positive maps on C ˚ -algebras, Proc. Amer. Math.

Soc. 86 (1982), 402–404.
[114] E. Størmer, Extension of positive maps into BpHq, J. Funct. Anal. 66 (1986),
235-254.
[115] E. Størmer, Separable states and positive maps, J. Funct. Anal. 254 (2008),
2303–2312.
[116] E. Størmer, Duality of cones of positive maps, Münster J. Math. 2 (2009),

299–309.
128
[117] E. Størmer, Tensor products of positive maps of matrix algebras, Math. Scand.
111 (2012), 5–11.
[118] E. Størmer, “Positive Linear Maps of Operator Algebras”, Springer-Verlag,

2013.
[119] E. Størmer, The analogue of Choi matrices for a class of linear maps on Von
Neumann algebras, Intern. J. Math. 26 (2016), 1550018.
[120] T. Takasaki and J. Tomiyama, On the geometry of positive maps in matrix

algebras, Math. Z. 184 (1983), 101–108.
[121] K. Tanahashi and J. Tomiyama, Indecomposable positive maps in matrix al-

gebras, Canad. Math. Bull. 31 (1988), 308–317.
[122] W.-S. Tang, On positive linear maps between matrix algebras, Linear Alg.
Appl. 79 (1986), 33–44.
[123] B. M. Terhal, Bell Inequalities and the Separability Criterion, Phys. Lett. A
271 (2000), 319–326.
[124] B. M. Terhal and P. Horodecki, A Schmidt number for density matrices, Phys.
Rev. A 61 (2000), 040301.
[125] J. Tomiyama, On the difference of n-positivity and complete positivity in C ˚ -

algebras, J. Funct. Anal. 49 (1982), 1–9.
[126] J. Tomiyama, On the geometry of positive maps in matrix algebras. II, Linear
Alg. Appl. 69 (1985), 169–177.
[127] R. Webster, “Convexity”, Oxford Univ. Press, 1994.
[128] R. F. Werner, Quantum states with Einstein-Podolsky-Rosen correlations ad-

mitting a hidden-variable model, Phys. Rev. A, 40 (1989), 4277–4281.
[129] S. L. Woronowicz, Positive maps of low dimensional matrix algebras, Rep.

Math. Phys. 10 (1976), 165–183.
[130] S. L. Woronowicz, Nonextendible positive maps, Commun. Math. Phys. 51

(1976), 243–282.
[131] S. L. Woronowicz, Exposed positive maps, private communication, July 29,

2011.
[132] S. Yamagami, Cyclic inequalities, Proc. Amer. Math. Soc, 118 (1993), 521–
527.
129
[133] Y. Yang, D. H. Leung and W.-S. Tang, All 2-positive linear maps from M3 pCq
to M3 pCq are decomposable, Linear Alg. Appl. 503 (2016), 233–247.
[134] D. A. Yopp and R. D. Hill, Extremals and exposed faces of the cone of positive
maps, Linear Multilinear Alg. 53 (2005), 167–174.
130
Index
Ads , 18 SR |ζy, 14
BP 1 rMm b Mn s, 29 SPk , 22
BP k rMm b Mn s, 42 SPk rMm , Mn s, 21
C ˝ , 39 T, 10
Cφ , 28 Tr , 10, 32
CCPrMm , Mn s, 43 Tk , 42
conv S, 21 Wφ , 89
CPrMm , Mn ; ps, 43 Wy rD; Cs, 90
CPrMm , Mn s, 41 X ˚ , 38
Dn , 11
DECrMm , Mn s, 44 %Γ , 34
DECpMm b Mn q, 45 φ̄, 34
Fk rζs, 87 φ˚ , 18
FV , 43, 74 φra,b,cs , 97
F V , 74 φch , 50
F W , 75 φrb , 114
F 1 , 72 φwo , 51
H, 113 BC, 15
HpMm , Mn q, 17 x ¨ , ¨ y, 10, 39
|iy, 9 xÃ|, 11
I, 12 a ˝ b, 114
Im , 12 adjoint, 8
id, 19 adjoint map, 18
idn , 19 affine isomorphism, 18, 76
int C, 15 affine manifold, 15
LpMm , Mn q, 17 ampliation map, 62
Lφ , 38 anti-automorphism, 113
Mm ` b M ` , 31
n arithmetic mean, 51
Mn , 9 arithmetic-geometric inequality, 79
Mn` , 16
Mnh , 16 bi-optimal property, 95, 96
Mmˆn , 9 bi-partite state, 31
m ^ n, 21 bi-spanning property, 96, 103, 106
m b n state, 31 bidual cone, 40
P rφs, 94 bidual face, 79, 80
P1 rMm , Mn s, 18 bilinear pairing, 39, 51, 53, 60
Pk rMm , Mn s, 41 Bloch ball, 24
PPT rMm b Mn s, 35 Bloch sphere, 24
PPTrMm , Mn s, 45 block matrix, 12
Sk rMm b Mn s, 30 block-positive matrix, 29, 76, 78, 89
131
block-wise transpose, 34 entanglement breaking map, 62, 69
boundary, 24, 88, 96, 105 entanglement witness, 89
boundary point, 15, 21, 92 exposed face, 73, 79, 81, 83, 94
bra, 8 exposed point, 75
Breuer–Hall map, 119 exposed positive map, 78, 107, 108, 117
Brower’s fixed point theorem, 27 exposed ray, 75, 86–88, 96, 103
extremal positive map, 16, 18, 20, 27, 52, 93
Carathéodory theorem, 16 extreme point, 15, 16, 26, 44, 67
Choi map, 51, 79, 81, 82, 95, 97 extreme ray, 16–18, 30, 32, 79, 81, 83, 84,
Choi matrix, 28, 32, 33, 47, 54, 61, 63, 67, 106
69, 77, 88, 89, 93, 97, 115
classically correlation, 68 face, 15, 43, 72, 74, 81, 85, 92–95
co-optimal property, 95, 103, 105 flip operation, 28, 39
co-spanning property, 95 function space, 31
coefficient matrix, 112
column elementary operation, 13 Gelfand–Naimark–Segal construction, 66
column vector, 8, 14 geometric mean, 51
commutant, 108 half-space, 40
completely copositive map, 43, 44, 49, 67, Hermitian matrix, 10, 17, 28, 34, 114
74, 93, 95, 98, 101, 113 Hermiticity preserving map, 17, 24, 28, 34,
completely positive map, 41, 43, 44, 47, 49, 108
52, 58, 66, 67, 69, 74, 79, 82, 89, 93, Hilbert–Schmidt distance, 24
94, 97, 101, 113 Holevo form, 66
completely positive projection, 114 homogeneous polynomial, 109
complex projective space, 24 Horodecki’s separability criterion, 69
composition, 61, 69 hyperplane, 17, 73, 92
congruence map, 18 ˚-homomorphism, 67
conjugate, 34
contraction, 26 identity map, 19, 32, 47, 54, 78, 109, 110
convex cone, 16, 18, 23, 27, 40, 62, 63, 72– identity matrix, 11, 12, 16, 19, 32
74, 92 indecomposable exposed positive map, 112,
convex hull, 21, 23 113
convex set, 14 indecomposable positive map, 27, 46, 50, 51,
copositive matrix, 43, 45 68, 87, 96, 97, 104, 116
inner product, 8, 24, 26
decomposable map, 23, 26, 27, 44–46, 62, interior, 88, 94, 96, 105
67, 78, 82, 93, 95, 96, 101 interior point, 14–16, 21, 27, 32, 72, 74, 84,
density matrix, 11, 17 85, 90, 92
detect, 90 irreducible positive linear map, 108
dimension theorem, 14 isometry, 26
dual cone, 39, 40, 60, 64, 69 isotropic state, 53, 56, 57, 69
dual face, 72, 73, 79
dual space, 38 Jamiolkowski–Choi isomorphism, 28, 67
duality, 67, 68, 79, 107 JC-subalgebra, 114, 115
Jordan homomorphism, 67, 114
eigenvalue, 12
Einstein-Podolsky-Rosen correlation, 68 kernel, 79
entangled state, 30, 32, 68, 93 kernel space, 14, 37
entangled subspace, 36 ket, 8
132
Kraus decomposition, 41, 67 PPT map, 45
k-block-positive matrix, 42 PPT state, 36, 37, 45, 46, 62, 67, 73, 75, 100
k-copositive map, 43, 58 principal submatrix, 51, 76, 77
k-positive map, 41, 42, 46, 47, 49, 58, 62, probability distribution, 11, 31
63, 66, 68, 69 product state, 31, 68, 69
k-simple vectors, 68 product vector, 14, 29–33, 36–38, 67, 95, 96,
k-superpositive map, 21, 41, 53, 55–57, 63, 101, 102, 106
65, 69 projection, 10, 19
proper face, 15
left mapping cone, 59, 60, 62–64 pure product state, 106
length, 106 pure separable state, 33
lexicographic order, 75 pure state, 11
linear functional, 8, 11, 21, 38–40, 66
linear map, 9, 17, 68 quaternion, 113, 116
local unitary, 54
range, 36, 108
mapping cone, 59, 60, 62, 65, 69 range criterion, 37
matrix, 9, 11 range space, 14, 36, 43
matrix algebra, 9, 67 range vector, 24, 31, 32
matrix units, 9, 61, 80 rank of a matrix, 14, 30, 35
maximal face, 83, 86–88 rank one matrix, 9, 14, 16, 22, 30, 36, 77,
minimal exposed face, 86, 87 79, 80, 83, 110
monomial, 109, 112, 118 rank one projection, 30
multi-linear map, 67 ray, 16
reflection, 25
non-degenerate bilinear pairing, 39, 72, 83,
relative interior point, 14
84, 92
right mapping cone, 59, 60, 62–64
non-extendibile positive map, 68
Robertson map, 113, 115, 119
operation of square, 114 rotation, 25
optimal property, 93, 94, 105 row elementary operation, 13
order isomorphism, 18 row vector, 11, 14
orthogonal matrix, 25, 26
scalar matrix, 107, 108, 111
partial conjugate, 37, 82, 95, 96 Schmidt number, 30–32, 42, 46, 53, 54, 57,
partial transpose, 34–37, 50, 67 58, 62, 65, 66, 68, 69
Pauli matrix, 23, 25, 113 Schmidt rank, 14, 30, 31, 54, 77
polynomial, 109 Schur product, 41
positive linear functional, 11, 66 self-adjoint matrix, 9
positive linear map, 18, 23, 27, 29, 32, 41, separable state, 30, 31, 33, 35–37, 45, 46, 53,
42, 45–47, 51, 60, 62, 65–69, 75, 85, 56, 64, 67–69, 89, 105
89, 93, 94, 101, 107 separable variables, 31
positive matrix, 9, 11, 12, 16, 29, 30, 35, 41, simplex, 106
42, 45, 50, 62, 65, 67, 89 singular value, 13
positive partial transpose, 35 singular value decomposition, 13, 14, 20, 26,
positive semi-definite matrix, 9 65, 75, 78
PPT, 35, 116 spanning property, 95, 104, 105
PPT criterion, 35, 64, 68 special unitary matrix, 23
PPT entangled state, 35, 36, 38, 68, 93, 96, spectral decomposition, 10
105 state, 11, 30
133
state of rank one, 31, 35
Stinespring representation theorem, 67
Stone–Weierstrass theorem, 31
superpositive map, 21, 69, 79
tensor product of function spaces, 31

tensor product of linear maps, 12, 61, 64, 69
tensor product of matrices, 11, 31
trace, 10
trace map, 32, 47
trace preserving, 24, 26, 27
transpose, 10
transpose map, 22, 25, 33, 42, 64
unital linear functional, 11

unital positive map, 24, 26, 27, 67
unitary matrix, 26
vector, 8, 11
Werner state, 56–58, 68

witness, 90
Woronowicz dimension condition, 108, 112,
117
Woronowicz map, 52, 110
134

Qit 230730

Uploaded by

Copyright:

Available Formats

Qit 230730

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Qit 230730

Uploaded by

Copyright:

Available Formats

Positive Maps in Quantum Information Theory

Lecture notes in the fall semester 2022

July 30, 2023

1 Positive Maps and Bi-partite States 7

2 Detecting Entanglement by Positive Maps 71

Positive Maps and Bi-partite

|vyxw| : |uy ÞÑ |vyxw|uy “ xw|uy|vy P Cm , |uy P Cn ,

is the standard matrix units. The m ˆ n matrix A “ raij s may be written by

1.1.2 Positive matrices

xA, By “ 0 for every B ðñ A “ 0.

xA, BCy “ xAC T , By “ xB T A, Cy,

|xyxy|T “ xy|T |xyT “ |ȳyxx̄|

xA, |x̄yxȳ|y “ Tr pA|x̄yxȳ|T q

for an n ˆ n matrix B% . Then % is positive, that is, sends a positive matrix to a

1.1.3 Tensor products

xÃ| “ pa11 , . . . , a1n , a21 , . . . , a2n , . . . , . . . , am1 , . . . , amn q P Cmn

of row vectors of the matrix A.

a21 b21 a21 b22 a22 b21 a22 b22

xA b B, C b Dy “ xA, CyxB, Dy. (1.6)

φ1 b φ2 : Mm1 b Mm2 Ñ Mn1 b Mn2

1.1.4 Singular value decomposition

V1˚ pA˚ AqV1 “ D P Mr , V2˚ pA˚ AqV2 “ 0 P Mn´r .

U1˚ AV1 “ U1˚ AV1 D´1{2 D1{2 “ U1˚ U1 D1{2 “ D1{2 ,

From V2˚ pA˚ AqV2 “ 0, we also have AV2 “ 0, and so we have

with the matrix Σ given by

and unitary matrices U P Mm and V P Mn . This is known as the singular value

Therefore, vectors t|ξk yu and t|ηk yu in (1.9) can be chosen to be orthogonal.

1.1.6 Boundary of a convex set

holds. This means that if an interior point of a line segment in C belongs to F

1.1.7 convex cones

1.2 Positive maps

1.2.1 Positive maps between matrix algebras

φpa˚ q “ φpb ´ icq “ φpbq ´ iφpcq “ pφpbq ` iφpcqq˚ “ φpaq˚ .

In short, φ satisfies the following relation

φpa˚ q “ φpaq˚ , a P Mm . (1.10)

It is easy to see that the following relation

holds whenever the composition φ ˝ ψ is possible. A linear map φ : Mm Ñ Mn is

1.2.2 Extremal positive maps

Ads pxq “ s˚ xs, x P Mm .

Adst “ Adt ˝ Ads , pAds q˚ “ AdsT

hold for matrices s and t, whenever st is defined. It is also clear that if s P Mn is

x`|a|ky “ xk|kyx`|a|ky “ xk|a|kyx`|ky “ 0,

when k ‰ `, and so a must be a diagonal matrix. We also have

x`|a|`y “ xk|kyx`|a|`y “ xk|a|kyx`|`y “ xk|a|ky,

Lemma 1.2.2 The identity map idn “ AdIn is extremal in P1 rMn , Mn s.

0 “ pIn ´ pqφppq ´ pφpIn ´ pq “ φppq ´ pφpIn q.

Taking the adjoint, we also have

0 “ pφppq ´ pφpIn qq˚ “ φppq ´ φpIn qp,

with the m ˆ m identity matrix Im in the left corner. If φ : M` Ñ Mm is extremal

Proof. Note that σ “ Ads is of the form:

Then we have τ ˝ σ “ idm . Suppose that σ ˝ φ “ ψ1 ` ψ2 with ψ1 , ψ2 P P1 rM` , Mn s.

ψ1 paq “ σpτ pψ1 paqqq “ σpλφpaqq “ λσpφpaqq,

for each a P Mm , and so σ ˝ φ “ Ads ˝φ is extremal. ˝

Theorem 1.2.4 For any m ˆ n matrix s, the linear map Ads : Mm Ñ Mn is

Proposition 1.2.5 If φ P int P1 rMm , Mn s then φ sends every nonzero element in

1.2.3 k-superpositive maps

SPk :“ conv tAds : rank s ď ku Ă P1 ,

SP1 Ř SP2 Ř ¨ ¨ ¨ Ř SPk Ř ¨ ¨ ¨ Ř SPm^n

of convex cones of positive maps. A positive map in SPk is called k-superpositive.