What Are Operator Spaces?

What are operator spaces?
Operator algebra group

G. Wittstock,
B. Betz, H.-J. Fischer, A. Lambert,
K. Louis, M. Neufang, I. Zimmermann
Universität des Saarlandes
January 7, 2001
Contents
1 Short History 3
2 Operator Spaces and Completely Bounded Maps 3

2.1 Basic facts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Ruan’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Elementary constructions . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.4 The space CB (X, Y ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.5 The dual . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.6 Direct sums . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.7 MIN and MAX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.8 Injective operator spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.8.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.8.2 Examples and elementary constructions . . . . . . . . . . . . . . 12
2.8.3 Characterizations . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3 Operator Systems and Completely Positive Maps 13

3.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2 Characterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.3 Matrix order unit norm . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.4 Injective operator systems . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4 Hilbertian Operator Spaces 15

4.1 The spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.2 The morphisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.3 The column Hilbert space CH . . . . . . . . . . . . . . . . . . . . . . . . 17
4.3.1 Characterizations . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.4 Column Hilbert space factorization . . . . . . . . . . . . . . . . . . . . . 19
1
What are operator spaces ? January 7, 2001 http://www.math.uni-sb.de/∼wittck/ 2
5 Multiplicative Structures 20
5.1 Operator modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
5.2 Completely bounded module homomorphisms . . . . . . . . . . . . . . . 22
5.3 Operator algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
6 Tensor Products 26
6.1 Operator space tensor products . . . . . . . . . . . . . . . . . . . . . . . 27
6.2 Injective operator space tensor product . . . . . . . . . . . . . . . . . . . 30
6.2.1 Exact operator spaces . . . . . . . . . . . . . . . . . . . . . . . . 31
6.3 Projective operator space tensor product . . . . . . . . . . . . . . . . . . 32
6.4 The Haagerup tensor product . . . . . . . . . . . . . . . . . . . . . . . . 33
6.5 Completely bounded bilinear mappings . . . . . . . . . . . . . . . . . . . 38
6.6 Module tensor products . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.6.1 Module Haagerup tensor product . . . . . . . . . . . . . . . . . . 42
7 Complete Local Reflexivity 43
8 Completely Bounded Multilinear Mappings 44
9 Automatic Complete Boundedness 46
10 Convexity 48
10.1 Matrix convexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
10.1.1 Separation theorems . . . . . . . . . . . . . . . . . . . . . . . . . 50
10.1.2 Bipolar theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
10.2 Matrix extreme points . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
10.3 C ∗ -convexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
10.3.1 Separation theorems . . . . . . . . . . . . . . . . . . . . . . . . . 54
10.4 C ∗ -extreme points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
11 Mapping Spaces 56
11.1 Completely nuclear mappings . . . . . . . . . . . . . . . . . . . . . . . . 57
11.2 Completely integral mappings . . . . . . . . . . . . . . . . . . . . . . . . 58
12 Appendix 59
12.1 Tensor products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
12.1.1 Tensor products of operator matrices . . . . . . . . . . . . . . . . 59
12.1.2 Joint amplification of a duality . . . . . . . . . . . . . . . . . . . 60
12.1.3 Tensor matrix multiplication . . . . . . . . . . . . . . . . . . . . 61
12.2 Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
13 Symbols 63
1 Short History
The theory of operator spaces grew out of the analysis of completely positive and
completely bounded mappings. These maps were first studied on C ∗ -algebras, and later
on suitable subspaces of C ∗ -algebras. For such maps taking values in B(H) represen-
tation and extension theorems were proved [Sti55], [Arv69], [Haa80], [Wit81], [Pau82].
Many of the properties shared by completely positive mappings can be taken over to
the framework of operator systems [CE77]. Operator systems provide an abstract de-
scription of the order structure of selfadjoint unital subspaces of C ∗ -algebras. Paulsen’s
monograph [Pau86] presents many applications of the theory of completely bounded
maps to operator theory. The extension and representation theorems for completely
bounded maps show that subspaces of C ∗ -algebras carry an intrinsic metric structure
which is preserved by complete isometries. This structure has been characterized by
Ruan in terms of the axioms of an operator space [Rua88]. Just as the theory of C ∗ -
algebras can be viewed as noncommutative topology and the theory of von Neumann
algebras as noncommutative measure theory, one can think of the theory of operator
spaces as noncommutative functional analysis.
This program has been presented to the mathematical community by E.G. Effros
[Eff87] in his address to the ICM in 1986. The following survey articles give a fairly
complete account of the development of the theory: [CS89], [MP94], [Pis97].
2 Operator Spaces and Completely Bounded Maps

2.1 Basic facts
The spaces
Let X be a complex vector space. A matrix seminorm [EW97b] is a family of mappings
k · k : Mn (X) → lR, one on each matrix level1 Mn (X) = Mn ⊗ X for n ∈ IN, such that
(R1) kαxβk ≤ kαkkxkkβk for all x ∈ Mn (X), α ∈ Mm,n , β ∈ Mn,m

(R2) kx ⊕ yk = max{kxk, kyk} for all x ∈ Mn (X), y ∈ Mm (X).2
Then every one of these mappings k · k : Mn (X) → lR is a seminorm. If one (and

then every one) of them is definite, the operator space seminorm is called a matrix
norm.
1
The term matrix level is to be found for instance in []
2
It suffices to show one of the following two weaker conditions:
(R10 ) kαxβk ≤ kαkkxkkβk for all x ∈ Mn (X), α ∈ Mn , β ∈ Mn ,

(R2) kx ⊕ yk = max{kxk, kyk} for all x ∈ Mn (X), y ∈ Mm (X),
which is often found in the literature, or
(R1) kαxβk ≤ kαkkxkkβk for all x ∈ Mn (X), α ∈ Mm,n , β ∈ Mn,m ,

(R20 ) kx ⊕ yk ≤ max{kxk, kyk} for all x ∈ Mn (X), y ∈ Mm (X),
which seems to be appropriate in convexity theory.

A matricially normed space is a complex vector space with a matrix norm. It

can be defined equivalently, and is usually defined in the literature, as a complex vector
space with a family of norms with (R1) and (R2) on its matrix levels.
If Mn (X) with this norm is complete for one n (and then for all n), then X is called
an operator space3 ([Rua88], cf. [Wit84a]).
For a matricially normed space (operator space) X the spaces Mn (X) are normed
spaces (Banach spaces).4 These are called the matrix levels of X (first matrix level,
second level . . . ).
The operator space norms on a fixed vector space X are partially ordered by the
pointwise order on each matrix level Mn (X). One says that a greater operator space
norm dominates a smaller one.
The mappings
A linear mapping Φ between vector spaces X and Y induces a linear mapping Φ(n) =
idMn ⊗ Φ ,
Φ(n) : Mn (X) → Mn (Y )
[xij ] 7→ [Φ(xij )] ,
the nth amplification of Φ.

For matricially normed X and Y , one defines
n o
kΦkcb := sup kΦ(n) k n ∈ IN .

Φ is called completely bounded if kΦkcb < ∞ and completely contractive if

kΦkcb ≤ 1.
Among the complete contractions, the complete isometries and the complete quotient
mappings play a special role. Φ is called completely isometric if all Φ(n) are isometric,5
and a complete quotient mapping if all Φ(n) are quotient mappings.6
The set of all completely bounded mappings from X to Y is denoted by CB (X, Y )
[Pau86, Chap. 7].
An operator space X is called homogeneous if each bounded operator Φ ∈
B(M1 (X)) is completely bounded with the same norm: Φ ∈ CB(X), and kΦkcb = kΦk
[Pis96].
3
In the literature, the terminology is not conseqent. We propose this distinction between matricially
normed space and operator space in analogy with normed space and Banach space.
4
In the literature, the normed space M1 (X) usually is denoted also by X. We found that a more
distinctive notation is sometimes usefull.
5
I. e.: kxk = kΦ(n) (x)k for all n ∈ IN, x ∈ Mn (X).
6 −1
I. e.: kyk = inf{kxk | x ∈ Φ(n) (y)} for alle n ∈ IN and y ∈ Mn (Y ), or equivalently
Φ (Ball Mn (X)) = Ball Mn (Y ) for all n ∈ IN, where Ball◦ Mn (X) = {x ∈ Mn (X) | kxk < 1}.
(n) ◦ ◦
Notations
Using the disjoint union
˙
[
M (X) := Mn (X),
n∈IN
the notation becomes simpler.7
Examples
B(H) is an operator space by the identification Mn (B(H)) = B(Hn ). Generally, each
C ∗ -algebra A is an operator space if Mn (A) is equipped with its unique C ∗ -norm. Closed
subspaces of C ∗ -algebras are called concrete operator spaces. Each concrete operator
space is an operator space. Conversely, by the theorem of Ruan, each operator space is
completely isometrically isomorphic to a concrete operator space.
Commutative C ∗ -algebras are homogeneous operator spaces.
The transposition Φ on l2 (I) has norm kΦk = 1, but kΦkcb = dim l2 (I). If I is
infinite, then Φ is bounded, but not completely bounded.
If dim H ≥ 2, then B(H) is not homogeneous [Pau86, p. 6].
Smith’s lemma
For a matricially normed space X and a linear operator Φ : X → Mn , we have kΦkcb =
kΦ(n) k. In particular, Φ is completely bounded if and only if Φ(n) is bounded [Smi83,
Thm. 2.10].
Rectangular matrices
For a matricially normed space X, the spaces Mn,m (X) = Mn,m ⊗ X of n × m-matrices
over X are normed by adding zeros so that one obtains a square matrix, no matter of
which size.
Then
Mn,m (B(H)) = B(Hm , Hn )
holds isometrically.
7
The norms on the matrix levels Mn (X) are then one mapping M (X) → lR. The amplifications of
Φ : X → Y can be described as one mapping Φ : M (X) → M (Y ). We have
kΦkcb = sup{kΦ(x)k | x ∈ M (X), kxk ≤ 1}.
Φ is completely isometric if kxk = kΦ(x)k for all x ∈ M (X), and Φ is a complete quotient mapping
if kyk = inf{kxk | x ∈ Φ−1 (y)} for all y ∈ M (Y ) or Φ(Ball◦ X) = Ball◦ Y , where Ball◦ X = {x ∈
M (X) | kxk < 1}.
2.2 Ruan’s theorem

Each concrete operator space is an operator space. The converse is given by
Ruan’s theorem: Each (abstract) operator space is completely isometrically isomorphic
to a concrete operator space [Rua88].
More concretely, for a matricially normed space X let Sn be the set of all complete
contractions from X to Mn . Then the mapping
M M
Φ:X → Mn
n∈IN Φ∈Sn
x 7→ (Φ(x))Φ
is a completely isometric embedding of X into a C ∗ -algebra [ER93].
A proof relies on the separation theorem for absolutely matrix convex sets.
This theorem can be used to show that many constructions with concrete operator
spaces yield again concrete operator spaces (up to complete isometry).
2.3 Elementary constructions

Subspaces and quotients
Let X be a matricially normed space and X0 ⊂ X a linear subspace. Then Mn (X0 ) ⊂
Mn (X), and X0 together with the restriction of the operator space norm again is a
matricially normed space. The embedding X0 ,→ X is completely isometric. If X is an
operator space and X0 ⊂ X is a closed subspace, then Mn (X0 ) ⊂ Mn (X) is closed and
X0 is an operator space.
Algebraically we have Mn (X/X0 ) = Mn (X)/Mn (X0 ). If X0 is closed, then X/X0
together with the quotient norm on each matrix level is matricially normed (an operator
space if X is one). The quotient mapping X → X/X0 is a complete quotient mapping.
More generally, a subspace of a matricially normed space (operator space) X is
a matricially normed space (operator space) Y together with a completely isometric
operator Y → X. A quotient of X is a matricially normed space (operator space) Y
together with a complete quotient mapping X → Y .
Matrices over an operator space

The vector space Mp (X) of matrices over a matricially normed space X itself is matri-
cially normed in a natural manner: The norm on the nth level Mn (Mp (X)) is given by
the identification
Mn (Mp (X)) = Mnp (X)
[BP91, p. 265]. We8 write
lMp (X)
8
In the literature, the symbol Mp (X) stands for both the operator space with first matrix level
Mp (X) and for the pth level of the operator space X. We found that the distinction between lMp (X)
and Mp (X) clarifies for instance the definition of the operator space structure of CB(X, Y ).
for Mp (X) with this operator space structure. In particular,
M1 (lMp (X)) = Mp (X)
holds isometrically. Analogously Mp,q (X) becomes a matricially normed space lMp,q (X)
by the identification
Mn (Mp,q (X)) = Mnp,nq (X).
By adding zeros it is a subspace of lMr (X) for r ≥ p, q.

Examples: For a C ∗ -algebra A, lMp (A) is the C ∗ -Algebra of p × p-matrices over A
with its natural operator space structure.
The Banach space Mp (A) is the first matrix level of the operator space lMp (A).
The complex numbers have a unique operator space structure which on the first
matrix level is isometric to C
l , and for this Mp (Cl ) = Mp holds isometrically. We write
lMp := lMp (Cl ).
Then lMp always stands for the C ∗ -algebra of p × p-matrices with its operator space
structure. The Banach space Mp is the first matrix level of the operator space lMp .
Columns and rows of an operator space

The space X p of p-tupels over an operator space X can be made into an operator space
for instance by reading the p-tupels as p × 1- or as 1 × p-matrices. This leads to the
frequently used columns and rows of an operator space X:
Cp (X) := lMp,1 (X) and Rp (X) := lM1,p (X).
The first matrix level of these spaces are
M1 (Cp (X)) = Mp,1 (X) and M1 (Rp (X)) = M1,p (X), respectively.
If X 6= {0}, the spaces Cp (X) and Rp (X) are not completely isometric. In general even
the first matrix levels Mp,1 (X) and M1,p (X) are not isometric.
Cp := Cp (Cl ) is called the p-dimensional column space and Rp := Rp (Cl ) the
p-dimensional row space.
The first matrix levels of Cp and Rp are isometric to l2p , but Cp and Rp are not
completely isometric.
2.4 The space CB (X, Y )

Let X and Y be matricially normed spaces. A matrix [Tij ] ∈ Mn (CB (X, Y )) determines
a completely bounded operator
T : X → lMn (Y )
x 7→ [Tij (x)] .
Defining k[Tij ]k = kT kcb , CB (X, Y ) becomes a matricially normed space. It is an

operator space, if Y is one. The equation
cb
lMp (CB (X, Y )) = CB (X, lMp (Y ))
holds completely isometrically.
2.5 The dual

The dual9 of a matricially normed space X is defined as X ∗ = CB(X, Cl ) [Ble92a].10
Its first matrix level is the dual of the first matrix level of X: M1 (X ∗ ) = (M1 (X))∗ .
The canonical embedding X ,→ X ∗∗ is completely isometric [BP91, Thm. 2.11].
Some formulae
For matricially normed spaces X and Y , m ∈ IN, y ∈ Mm (Y ) and T ∈ CB (X, Y ) we
have
kyk = sup {kΦ(y)k | n ∈ IN, Φ ∈ CB (Y, lMn ), kΦkcb ≤ 1}
and
kT kcb = sup{kΦ(n) ◦ T kcb | n ∈ IN, Φ ∈ CB (Y, lMn ), kΦkcb ≤ 1}.
A matrix [Tij ] ∈ Mn (X ∗ ) defines an operator
T : Mn (X) → Cl
X
[xij ] 7→ Tij xij .
i,j
Thus we have an algebraic identification of Mn (X ∗ ) and Mn (X)∗ and further of Mn (X ∗∗ )

and Mn (X)∗∗ . The latter even is a complete isometry ([Ble92b, Cor. 2.14]):
cb
lMn (X ∗∗ ) = lMn (X)∗∗ .
The isometry on the first matrix level is shown in [Ble92a, Thm. 2.5]. This already
implies11 the complete isometry. More generally we have12
cb
CB (X ∗ , lMn (Y )) = CB (lMn (X)∗ , Y ).
9
In the literature, this dual was originally called standard dual [Ble92a].
10
The norm of a matricially normed space X is given by the unit ball BallX ⊂ M (X). Here,
BallX ∗ = {Φ : X → Mn | n ∈ IN, Φ completely contractive}.
11
Mk (lMn (X)∗∗ ) = Mk (Mn (X))∗∗ = Mkn (X)∗∗ = Mkn (X ∗∗ ) = Mk (lMn (X ∗∗ )).
cb
12
This follows from lMn (X ∗∗ ) = lMn (X)∗∗ and the above mentioned formula
kT kcb = sup{kΦ(n) ◦ T kcb | n ∈ IN, Φ ∈ CB (Y, lMn ), kΦkcb ≤ 1}.
cb
X is called reflexive, if X = X ∗∗ . An operator space X is reflexive if and only if
its first matrix level M1 (X) is a reflexive Banach space.
The adjoint operator

For T ∈ CB (X, Y ), the adjoint operator T ∗ is defined as usual. We have: T ∗ ∈
CB (Y ∗ , X ∗ ), and kT kcb = kT ∗ kcb . The mapping
∗
: CB (X, Y ) → CB (Y ∗ , X ∗ )
T 7→ T ∗
even is completely isometric [Ble92b, Lemma 1.1].13

T ∗ is a complete quotient mapping if and only if T is completely isometric; T ∗ is
completely isometric if T is a complete quotient mapping. Especially for a subspace
X0 ⊂ X we have [Ble92a]:
cb
X0∗ = X ∗ /X0⊥
and, if X0 is closed,
cb
(X/X0 )∗ = X0⊥ .
2.6 Direct sums

∞-direct sums
Let I be an index set and Xi for each i ∈ I an operator space. Then there are an
operator space X and complete contractions πi : X → Xi with the following universal
mapping property: For each family of complete contractions ϕi : Z → Xi there is exactly
one complete contraction ϕ : Z → X such L that ϕi = πi ◦ ϕ for all i. X is called ∞-
direct sum of the Xi and is denoted by ∞ (Xi | i ∈ I). The πi are complete quotient
mappings.
Q construct a ∞-direct sum for instance as the linear subspace X =
One can
{(xi ) ∈ i∈I Xi | sup{kxi k | i ∈ I} < ∞} of the cartesian product of the
X
Qi , the π i being the projections on the components. We have Mn (X) = {(xi ) ∈
i∈I Mn (Xi ) | sup{kxi k | i ∈ I} < ∞}, and the operator space norm is given by
k(xi )k = sup{kxi k | i ∈ I}.
13
The isometry on the matrix levels follows from the isometry on the first matrix level using the
cb
above mentioned formula CB (lMn (X)∗ , Y ) = CB (X ∗ , lMn (Y )):
Mn (CB (X, Y )) = M1 (CB (X, lMn (Y )))

,→ M1 (CB (lMn (Y )∗ , X ∗ ))
= M1 (CB (Y ∗ , lMn (X ∗ )))
= Mn (CB (Y ∗ , X ∗ )).
1-direct sums
Let I be an index set and Xi for each i ∈ I an operator space. Then there are an
operator space X and complete contractions ιi : Xi → X with the following universal
mapping property: For each family of complete contractions ϕi : Xi → Z there is exactly
one complete contraction ϕ : X →LZ such that ϕi = ϕ ◦ ιi for all i. X is called 1-direct
sum of the Xi and is denoted by 1 (Xi | i ∈ I).The ιi are completely isometric.
One can construct a 1-direct sum∗ for instance as the closure of the sums of the
π
Xi∗∗ →i ( ∞ (Xi∗ | i ∈ I))∗ , where πi is the projection
L
images
Lof the∗ mappings Xi ,→
from ∞ (Xi | i ∈ I) onto Xi∗ .
The equation
M M
( (Xi | i ∈ I))∗ = (Xi∗ | i ∈ I)
1 ∞
p-direct sums
p-direct sums of operator spaces for 1 < p < ∞ can be obtained by interpolation between
the ∞- and the 1-direct sum.
2.7 MIN and MAX

Let E be a normed space. Among all operator space norms on E which coincide on the
first matrix level with the given norm, there is a greatest and a smallest. The matricially
normed spaces given by these are called MAX (E) and MIN (E). They are characterized
by the following universal mapping property:14 For a matricially normed space X,
M1 (CB (MAX (E), X)) = B(E, M1 (X))
and
M1 (CB (X, MIN (E))) = B(M1 (X), E)
We have [Ble92a]
cb
MIN (E)∗ = MAX (E ∗ ),
cb
MAX (E)∗ = MIN (E ∗ ).
For dim(E) = ∞,
idE : MIN (E) → MAX (E)
14
MAX is the left adjoint and MIN the right adjoint of the forgetfull functor which maps an operator
space X to the Banach space M1 (X).
is not completely bounded[Pau92, Cor. 2.13].15

Subspaces of MIN -spaces are MIN -spaces: For each isometric mapping ϕ : E0 → E,
the mapping ϕ : MIN (E0 ) → MIN (E) is completely isometric.
Quotients of MAX -spaces are MAX -spaces: For each quotient mapping ϕ : E → E0 ,
the mapping ϕ : MAX (E) → MAX (E0 ) is a complete quotient mapping.
Construction of MIN :
For a commutative C ∗ -algebra A = C(K), each bounded linear mapping Φ : M1 (X) → A
is automatically completely bounded with kΦkcb = kΦk [Loe75].16
Each normed space E is isometric to a subspace of the commutative C ∗ -algebra
l∞ (Ball(E ∗ )). Thus the operator space MIN (E) is given as a subspace of l∞ (Ball(E ∗ )).
For x ∈ Mn (MIN (E)) we have
n o
kxk = sup kf (n) (x)k f ∈ Ball(E ∗ ) .

The unit ball of MIN (E) is given as the absolute matrix polar of Ball(E ∗ ).
Construction of MAX :
For a index set I, l1 (I) = c0 (I)∗ . l1 (I) is an operator space as dual of the commutative
C ∗ -algebra c0 (I), and each bounded linear mapping Φ : l1 (I) → M1 (X) is automatically
completely bounded with kΦkcb = kΦk.17
Each Banach space18 E is isometric to a quotient of l1 (Ball(E)). Thus the operator
space MAX (E) is given as a quotient of l1 (Ball(E)).
For x ∈ Mn (MAX (E)) we have
kxk = sup{kϕ(n) (x)k | n ∈ IN, ϕ : E → Mn , kϕk ≤ 1}.
The unit ball of MAX (E) is given as the absolute matrix bipolar of Ball(E ) .
2.8 Injective operator spaces

2.8.1 Definition
A matricially normed space X is called injective if completely bounded mappings into
X can be extended with the same norm. More exactly:
For all matricially normed spaces Y0 and Y , each complete contraction ϕ : Y0 → X
and each complete isometry ι : Y0 → Y there is a complete contraction ϕ̃ : Y → X such
that ϕ̃ι = ϕ.
15
Paulsen uses in his proof a false estimation for the projection constant of the finite dimensional
Hilbert spaces; the converse estimation is correct [Woj91, p. 120], but here useless. The gap can be
filled [Lam97, Thm. 2.2.15] using the famous theorem of Kadets-Snobar: The projection constant of an
√
n-dimensional Banach space is less or equal than n [KS71].
16
I. e. A is a MIN -space.
17
I. e. l1 (I) is a MAX -space.
18
A similar construction is possible for normed spaces.
It suffices to consider only operator spaces Y0 and Y . Injective matricially normed

spaces are automatically comlete, so they are also called injective operator spaces.
2.8.2 Examples and elementary constructions

B(H) is injective [].
Completely contractively projectable subspaces of injective operator spaces are in-
jective.
L∞ -direct sums of injective operator spaces are injective.
Injective operator systems, injective C ∗ -algebras and injective von Neumann-
algebras are injective operator spaces.
2.8.3 Characterizations
For a matricially normed space X the following conditions are equivalent:
a) X is injective.
b) For each complete isometry ι : X → Z there is a complete contraction π : Z → X
such that πι = idX . I. e. X is completely contractively projectable in each space
containing it as a subspace.
c) For each complete isometry ι : X → Z and each complete contraction ϕ : X →
Y there is a complete contraction ϕ̃ : Z → Y such that ϕ̃ι = ϕ. I. e. Complete
contractions from X can be extended completely contractively to any space conaining
X as a subspace.19
d) X is completely isometric to a completely contractively projectable subspace of
B(H) for some Hilbert space H.
e) X is completely isometric to pAq, where A is an injective C ∗ -algebra and p and
q are projections in A.[]
Robertson[] characterized the infinite dimensional injective subspaces of L B(l2 ) up
L∞
to isometry (not complete isometry!). They are B(l2 ), l∞ , l2 , l∞ ⊕ l2 and n∈IN l2 .
∞
(Countable L -direct sums of such are again comletely isometric to one of these.) If an
injective subspace of B(l2 ) is isometric to l2 , it is completely isometric to Rl2 or Cl2 .[]
Injective envelopes
Let X be a matricially normed space.
An operator space Z together with a completely isometric mapping ι : X → Z is
called an injective envelope of X if Z is injective, and if idZ is the unique extension
of ι onto Z.
This is the case if and only if Z is the only injective subspace of Z which contains
the image of X.[]
Each matricially normed space has an injective envelope. It is unique up to a canon-
ical isomorphism.
19
Equivalently: Completely bounded mappings from X can be extended with the same norm.
A matricially normed space Z together with a completely isometric mapping ι :

X → Z is called an essential extension of X if a complete contraction ϕ : Z → Y is
completely isometric if only ϕ ◦ ι is completely isometric.
ι : X → Z is an injective envelope if and only if Z is injective and ι : X → Z is an
essential extension.
Every injective envelope ι : X → Z is a maximal essential extension, i. e. for each
essential extension ι̃ : X → Z̃, there is a completely isometric mapping ϕ : Z̃ → Z such
that ϕ ◦ ι̃ = ι.
3 Operator Systems and Completely Positive Maps

3.1 Definitions
Let V be a complex vector space. An involution on V is a conjugate linear map
∗ : V → V , v 7→ v ∗ , such that v ∗∗ = v. A complex vector space is an involutive vector
space if there is an involution on V . Let V be an involutive vector space. Then Vsa is
the real vector space of selfadjoint elements of V , i.e. those elements of V , such that
v ∗ = v. An involutive vector space is an ordered vector space if there is a proper cone20
V + ⊂ Vsa . The elements of V + are called positive and there is an order on Vsa defined
by v ≤ w if w − v ∈ V + for v, w ∈ Vsa .
An element 1l ∈ V + is an order unit if for any v ∈ Vsa there is a real number t > 0,
such that −t1l ≤ v ≤ t1l. If V+ has an order unit then Vsa = V+ − V+ .
The cone V + is called Archimedian if w ∈ −V + whenever there exists v ∈ Vsa such
that tw ≤ v for all t > 0. If there is an order unit 1l then V+ is Archimedian if w ∈ −V +
whenever tw ≤ 1l for all t > 0.
Let V be an ordered vector space. If V+ is Archimedian and contains a distinguished
order unit 1l then (V, 1l) is called an ordered unit space.
Let V, W be involutive vector spaces. We define an involution ? on the space L(V, W )
of all linear mappings from V → W by ϕ? (v) = ϕ(v ∗ )∗ , ϕ ∈ L(V, W ). If moreover V, W
are ordered vector spaces, then ϕ is positive if ϕ? = ϕ and ϕ(V + ) ⊂ W + . If (V, 1l) and
(W, 1l0 ) are ordered unit spaces a positve map ϕ : V → W is called unital if ϕ(1l) = 1l0 .
Let V be an involutive vector space. Then Mn (V ) is also an involutive vector space by
[vij ]∗ = [vji
∗ ]. V is a matrix ordered vector space if there are proper cones M (V )+ ⊂
n
Mn (V )sa for all n ∈ IN, such that α∗ Mp (V )+ α ⊂ Mq (V )+ for all α ∈ Mpq and p, q ∈ IN
holds21 . This means that (Mn (V )+ )n∈IN is a matrix cone.
Let V, W be matrix ordered vector spaces. A linear mapping φ : V → W is
completely positive if φ(n) : Mn (V ) → Mn (W ) is positive for all n ∈ IN. A complete
order isomorphism22 from V to W is a completely positive map from V → W that
is bijective, such that the inverse map is completely positive.
20
A cone K is a subset of a vector space, such that K + K ⊂ K and lR+ K ⊂ K. If moreover
(−K) ∩ K = {0} holds, K is a proper cone.
21
If V + = M1 (V )+ is a proper cone, by the matrix condition all the cones Mn (V )+ will be proper.
22
Note that some authors don’t include surjectivity in the definition.
The well-known Stinespring theorem for completely positive maps reads [Pau86,
Theorem 4.1]:
Let A be a unital C ∗ -algebra and let H be a Hilbert space. If ψ : A → B(H)
is completely positive then there are a Hilbert space Hπ , a unital ∗-homomorphism
π : A → B(Hπ ) and a linear mapping V : H → Hπ , such that ψ(a) = V ∗ π(a)V for all
a ∈ A.
Let V be an involutive vector space. Then V is called an operator system if it is a

matrix ordered ordered unit space, such that Mn (V )+ is Archimedian for all n ∈ IN. In
this case Mn (V ) is an ordered unit space with order unit 1ln = 1l ⊗ idMn for all n ∈ IN,
where 1l ∈ V+ is the distinguished order unit of V and idMn is the unit of lMn .
Example
Let H be a Hilbert space. Then, obviously, B(H) is an ordered unit space with
order unit the identity operator. Using the identification Mn (B(H)) = B(H n ) we let
Mn (B(H))+ = B(H n )+ . So we see that B(H) is an operator system.
Let L be an operator system. Then any subspace S ⊂ L that is selfadjoint, i.e.
∗
S ⊂ S, and contains the order unit of L is again an operator system with the induced
matrix order. So unital C ∗ -algebras and selfadjoint subspaces of unital C ∗ -algebras
containig the identity are operator systems.
Note that a unital complete order isomorphism between unital C ∗ -algebras must be
a ∗-isomorphism [Cho74, Corollary 3.2]. So unital C ∗ -algebras are completely charac-
terized by their matrix order. They are not characterized by their order. For instance
take the opposite algebra Aop of a unital C ∗ -algebra A. Then A+ = Aop
+ but A and A
op
op
are not ∗-isomorphic. Obviously M2 (A)+ 6= M2 (A )+ .
3.2 Characterization
Choi and Effros [CE77, Theorem 4.4] showed the following characterization theorem:
Let V be an operator system. Then there are a Hilbert space H and a unital complete
order isomorphism from V to a selfadjoint subspace of B(H).
A unital complete order isomorphism is obtained by
M M
Φ:V → lMn
n∈IN ϕ∈Sn
x 7→ (ϕ(x))ϕ ,
where Sn is the set of all unital completely positive maps ϕ : V → lMn .
3.3 Matrix order unit norm

Let L be an operator system. We define norms by

r1ln x
kxkn := inf r ∈ lR| ∈ M2n (L)+ (1)
x∗ r1ln
for all n ∈ IN and x ∈ Mn (L). With these norms L becomes an operator space.
If Φ is any unital completely positve embedding from L into some B(H) (cf. section
3.2) then kΦ(n) (x)k = kxkn for all n ∈ IN and x ∈ Mn (L). This holds because kyk ≤ 1
if and only if

1l y
0≤
y ∗ 1l
for all y ∈ B(H) and all Hilbert spaces H.
Let L and S be operator systems and let ψ : L → S be completely positive. We
supply L and S with the norms from equation (1). Then ψ is completely bounded and
kψ(1l)k = kψk = kψkcb (cf. [Pau86, Proposition 3.5]).
3.4 Injective operator systems

An operator system R is called injectiv if given operator systems N ⊂ M each com-
pletely positive map ϕ : N → R has a completely positive extension ψ : M → R.
If an operator system R is injective then there is a unital complete order isomorphism
from R onto a unital C ∗ -algebra. The latter is conditionally complete23 . (cf. [CE77,
Theorem 3.1])
4 Hilbertian Operator Spaces

4.1 The spaces
An operator space X is called hilbertian, if M1 (X) is a Hilbert space H. An operator
space X is called homogeneous, if each bounded operator T : M1 (X) → M1 (X) is
completely bounded and kT kcb = kT k [Pis96].
Examples The minimal hilbertian operator space MIN H := MIN (H) and the
maximal hilbertian operator space MAX H := MAX (H), the column Hilbert space
CH := B(Cl , H) and the row Hilbert space RH := B(H, Cl ) are homogeneous hilber-
tian operator spaces on the Hilbert space H.
Furthermore, for two Hilbert spaces H and K, we have completely isometric isomor-
phisms [ER91, Thm. 4.1] [Ble92b, Prop. 2.2]
cb cb
CB (CH , CK ) = B(H, K) and CB (RH , RK ) = B(K, H).
These spaces satisfy the following dualities [Ble92b, Prop. 2.2] [Ble92a, Cor. 2.8]:
∗ cb
CH = RH
cb
R∗H = CH
cb
MIN ∗H = MAX H
cb
MAX ∗H = MIN H .
23
An ordered vector space V is conditionally complete if any upward directed subset of Vsa that is
bounded above has a supremum in Vsa .
For each Hilbert space H there is a unique completely self dual homogeneous operator
space, the operator Hilbert space OH H [Pis96, §1]:
cb
OH ∗H = OH H .
The intersection and the sum of two homogeneous hilbertian operator spaces are
again homogeneous hilbertian operator spaces [Pis96].
4.2 The morphisms

The space CB (X, Y ) of completely bounded mappings between two homogeneous hilber-
tian operator spaces X, Y enjoys the following properties (cf. [MP95, Prop. 1.2]):
1. (CB (X, Y ), k · kcb ) is a Banach space.
2. kAT Bkcb ≤ kAkkT kcb kBk for all A, B ∈ B(H), T ∈ CB (X, Y ).
3. kT kcb = kT k for all T with rank(T ) = 1.
Consequently, CB (X, Y ) is a symmetrically normed ideal (s.n. ideal) in the sense of

Calkin, Schatten [Sch70] and Gohberg [GK69].
The classical examples for s.n. ideals are the famous Schatten ideals:
Sp := {T ∈ B(H) | the sequence of singular values of T is in `p } (1 ≤ p < ∞).
Many, but not all s.n. ideals can be represented as spaces of completely bounded maps
between suitable homogeneous hilbertian operator spaces.
The first result in this direction was
CB (RH , CH ) = S2 (H) = HS (H)
isometrically [ER91, Cor. 4.5].

We have the following characterizations isometrically or only isomorphically (')
[Mat94], [MP95], [Lam97]:
CB (↓, →) MIN H CH OH H RH MAX H

MIN H B(H) HS (H) ' HS (H) HS (H) ' N (H)
CH B(H) B(H) S4 (H) HS (H) HS (H)
OHH B(H) S4 (H) B(H) S4 (H) ' HS (H)
RH B(H) HS (H) S4 (H) B(H) HS (H)
MAX H B(H) B(H) B(H) B(H) B(H)
cb
As a unique completely isometric isomorphism, we get CB (CH ) = B(H) (cf. [Ble95,
Thm. 3.4]). The result
CB (MIN H , MAX H ) ' N (H)
is of special interest. Here, we have a new quite natural norm on the nuclear operators,
which is not equal to the canonical one.
Even in the finite dimensional case, we only know

n n
≤ kid : MIN (`n2 ) → MAX (`n2 )kcb ≤ √
2 2
[Pau92, Thm. 2.16]. To compute the exact constant is still an open problem. Paulsen
conjectured that the upper bound is sharp [Pau92, p. 121].
Let E be a Banach space and H a Hilbert space. An operator T ∈ B(E, H) is
completely bounded from MIN (E) to CH , if and only if T is 2-summing [Pie67] from E
to H (cf. [ER91, Thm. 5.7]):
M1 (CB (MIN (E), CH )) = Π2 (E, H)
with kT kcb =π2 (T ).
4.3 The column Hilbert space CH

For the Hilbert space H = `2 , we can realize the column Hilbert space CH as a column,
by the embedding
H ,→ B(H),
   . 
.. .. 0 ··· 0 ···
 .   .. .. 
 ξi 
 ξi . .
 
 ..  7→  . .
  
 .   .. 0 ··· 0 
..
   
. .. .. ..
. . .
Via this identification, C`n2 is the n-dimensional column space Cn .
CH is a homogeneous hilbertian operator space: All bounded maps on H are com-
cb
pletely bounded with the same norm on CH . Actually we have CB (CH ) = B(H) com-
pletely isometrically [ER91, Thm. 4.1].
CH is a injective operator space (cf. [Rob91]).
Tensor products
Let X be an operator space. We have complete isometries [ER91, Thm. 4.3 (a)(c)]
[Ble92b, Prop. 2.3 (i)(ii)]:
cb ∨
CH ⊗h X = CH ⊗ X
and ∧
cb
X ⊗h CH = X ⊗ CH .
∨ ∧
Herein, ⊗h is the Haagerup tensor product, ⊗ the injective tensor product and ⊗ the
projective tensor product.
For Hilbert spaces H and K, we have complete isometries [ER91, Cor. 4.4.(a)]
[Ble92b, Prop. 2.3(iv)]
cb ∨ cb ∧ cb
CH ⊗h CK = CH ⊗ CK = CH ⊗ CK = CH⊗2 K .
4.3.1 Characterizations
In connection with the column Hilbert space, it is enough to calculate the row norm
kT krow := sup sup k[T x1 . . . T xn ]kM1,n (Y )

n∈IN k[x1 ...xn ]kM ≤1
1,n (X)
or the column norm

 
T x1

 .. 
kT kcol := sup sup  . 
n∈IN 
x

1
T xn
 .. 

 .  ≤1 M n,1 (X)

xn
Mn,1 (Y )
of an operator T , instead of the cb-norm, to ascertain the complete boundedness.

Let X be an operator space and S : CH → X bzw. T : X → CH . Then we have
kSkcb = kSkrow resp. kT kcb = kT kcol ([Mat94, Prop. 4] resp. [Mat94, Prop. 2]).
The column Hilbert space is characterized as follows,
(A) as a hilbertian operator space [Mat94, Thm. 8]:

For an operator space X on an Hilbert space H, we have the following equivalences:
1. X is completely isometric to CH .
2. For all operator spaces Y and all T : X → Y we have kT kcb = kT krow , and
for all S : Y → X we have kSk = kSkrow . For all operator spaces Y and
all T : Y → X we have kT kcb = kT kcol , and for all S : X → Y we have
kSk = kSkcol .
3. X coincides with the maximal hilbertian operator space on columns and with
the minimal hilbertian operator space on rows. That means isometrically
Mn,1 (X) = Mn,1 (MAX (H))

M1,n (X) = M1,n (MIN (H))
(B) as an operator space: For an operator space X TFAE:

cb
1. There is a Hilbert space H, such that X = CH completely isometrically.
2. We have
Mn,1 (X) = ⊕2 M1 (X)
and
M1,n (X) = M1,n (MIN (M1 (X)))
isometrically. [Mat94, Thm. 10].
3. CB (X) with the composition as multiplication is an operator algebra [Ble95,
Thm. 3.4].
4.4 Column Hilbert space factorization

Let X, Y be operator spaces. We say that a linear map T : M1 (X) → M1 (Y ) factors
through a column Hilbert space, if there is a Hilbert space H and completely bounded
maps T2 : X → CH , T1 : CH → Y with T = T1 ◦ T2 . We define
γ2 (T ) := inf kT1 kcb kT2 kcb ,
where the infimum runs over all possible factorizations. If no such factorisation exists
we say γ2 (T ) := ∞. Γ2 (X, Y ) is the Banach space of all linear maps T : X → Y with
γ2 (T ) < ∞ [ER91, Chap. 5],[Ble92b, p. 83].
Let X1 and Y1 be operator spaces and T ∈ Γ2 (X, Y ), S ∈ CB (X1 , X), R ∈
CB (Y, Y1 ). Then we have the CB ideal property
γ2 (RT S) ≤ kRkcb γ2 (T )kSkcb .
We interpret a matrix T = [Tij ] ∈ Mn (Γ2 (X, Y )) as a mapping from X to Mn (Y ):
[Tij ](x) := [Tij (x)]. T has a factorization in completely bounded mappings
T T
X →2 M1,n (CH ) →1 Mn (Y ).
Again, we define
γ2 (T ) := inf kT1 kcb kT2 kcb ,
where the infimum is taken over all factorizations. So we get an operator space structure
on Γ2 (X, Y ) [ER91, Cor. 5.4].
Let X, Y be operator spaces and Y0 an operator subspace of Y . Then the inclusion
Γ2 (X, Y0 ) ,→ Γ2 (X, Y ) is completely isometric [ER91, Prop. 5.2].
Let X, Y be operator spaces. It is well known that every linear map
T :X →Y∗
defines a linear functional
fT : Y ⊗ X → Cl
via
hfT , y ⊗ xi := hT (x), yi.
This identification determines the complete isometry [ER91, Thm. 5.3] [Ble92b, Thm.
2.11]
cb
Γ2 (X, Y ∗ ) = (Y ⊗h X)∗ .
Let X, Y , Z be operator spaces. We get a complete isometry
cb
Γ2 (Y ⊗h X, Z) = Γ2 (X, Γ2 (Y, Z))
via the mapping
T 7→ Te
Te(x)(y) := T (y ⊗ x)
[ER91, Cor. 5.5].
5 Multiplicative Structures
For an abstract C ∗ -algebra the GNS construction provides a concrete representation
of its elements as bounded operators on a Hilbert space. For non-selfadjoint algebras
there is, hitherto, no analogue in the framework of classical functional analysis. But
endowed with an operator space structure (which is compatible with the multiplica-
tive structure), these non-selfadjoint algebras do have a representation in some B(H)
(theorem of Ruan type for operator algebras).
The so-called operator modules (over algebras) are also characterized by
Axioms of Ruan type; here, matrices whose entries are algebra elements
take the place of the scalar ones. The corresponding morphisms are the
completely bounded module homomorphisms, the most important properties of which
make their appearance in Representation, Decomposition and Extension Theorems.
5.1 Operator modules

Let A1 , A2 ⊂ B(H) be C ∗ -algebras with 1lH ∈ A1 , A2 . A closed subspace X of B(H) is
called a concrete (A1 , A2 )-operator module , if A1 X ⊂ X and XA2 ⊂ X. Whenever
A1 = A2 = A, we call X a concrete A-operator bimodule (cf. [ER88, p. 137]).
In analogy to the operator space or the operator algebra situation, there is an
abstract characterization of operator modules (cf. [Pop00, Déf. 4.1]):
Consider, as above, two unital C ∗ -algebras A1 , A2 ⊂ B(H) with 1lH ∈ A1 , A2 , and an
(algebraic) (A1 , A2 )-module X. We call X an abstract (A1 , A2 )-operator module, if it
carries an operator space structure satisfying the following axioms (of Ruan type):
(R1) kaxbkm ≤ kakkxkm kbk

x 0
(R2)
0 y = max{kxkm , kykn },
m+n
where m, n ∈ IN, a ∈ Mm (A1 ), x ∈ Mm (X), y ∈ Mn (Y ), b ∈ Mm (A2 ).

For abstract operator modules holds a representation theorem of Ruan type
(cf. [Pop00, Thm. 4.7]):
Let V be an abstract (A1 , A2 )-operator module. Then there exist a Hilbert space K, a
complete isometry Θ : X ,→ B(K) and ∗ -representations π1 , π2 of A1 resp. A2 in B(K)
such that:
Θ(axb) = π1 (a)Θ(x)π2 (b),
where x ∈ X, a ∈ A1 , b ∈ A2 . In case A1 = A2 , one can even choose π1 = π2 .
Basic examples of operator modules

Let A be a unital C ∗ -algebra, X a normed space, and Y an operator space. Then
B(X, A) resp. CB (Y, A) are operator spaces via the identifications Mn (B(X, A)) =
B(X, Mn (A)) resp. Mn (CB (Y, A)) = CB (Y, Mn (A)). These become A-operator bimod-
ules, when endowed with the natural module operations as follows ([ER88, p. 140]):
(a · ϕ · b)(x) = aϕ(x)b
for all a, b ∈ A, ϕ ∈ B(X, A) resp. CB (Y, A), x ∈ X resp. x ∈ Y .

In the category of operator modules, the morphisms are the
completely bounded module homomorphisms. For these we have a representation and
an extension theorem.
Representation theorem (cf. [Hof95, Kor. 1.4]):
Let H be a Hilbert space, M a C ∗ -algebra in B(H), and A, B C ∗ -subalgebras of M .
Then the following hold true:
(a) (cf. [Pau86, Thm. 7.4]) For each completely bounded (A, B)-module homo-
morphism Φ : M → B(H), there exist a Hilbert space K, a ∗ -representation
π : M → B(K) and linear operators v, w ∈ B(H, K) sharing the following proper-
ties:
(a1) Φ(x) = v ∗ π(x)w for all x ∈ M , i.e. (K; π; v ∗ ; w) is a representation of Φ

(a2) kΦkcb = kvkkwk
(a3) lin(π(M )vH) = lin(π(M )wH) = K
(a4) v ∗ π(a) = av ∗ for all a ∈ A and π(b)w = wb for all b ∈ B.
(b) (cf. also [Smi91, Thm.3.1]) If, in addition, M ⊂ B(H) is a von Neumann algebra
and Φ : M → B(H) is a normal completely bounded (A, B)-module homomor-
phism, one can require the ∗ -representation π of part (a) to be normal. There
exist families (ai )i∈I and (bi )i∈I in the commutant of A and B, respectively, with
the following properties (the sums are to be taken in the WOT topology):
P
(b1) Φ(x) = i∈I ai xbi for all x ∈ M
∗ ∗
P P
(b2) i∈I ai ai ∈ B(H), i∈I bi bi ∈ B(H) and kΦkcb =
1 P 1
∗ ∗
P
k i∈I ai ai k k i∈I bi bi k .
2 2
Extension theorem ([Wit84a, Thm. 3.1], cf. also [MN94, Thm. 3.4] and [Pau86, Thm.
7.2]):
Let F be an injective C ∗ -algebra, and let A, B ⊂ F be two unital C ∗ -subalgebras.
Consider furthermore two (A, B)-operator modules E0 and E with E0 ⊂ E. Then for
each φ0 ∈ CB (A,B) (E0 , F ), there exists an extension φ ∈ CB (A,B) (E, F ) with φ|E0 = φ0
and kφkcb = kφ0 kcb .
The decomposition theorem for completely bounded module homomorphisms can be
found in the corresponding chapter.
5.2 Completely bounded module homomorphisms

Let A, B ⊂ B(H) be C ∗ -algebras with 1lH ∈ A, B, and let E and F be two (A, B)-
operator modules, i.e. (algebraic) A-left-B-right-modules. A mapping φ ∈ L(E, F )
is called (A, B)-module homomorphism (in case A = B A-bimodule homomor-
phism) if
φ(axb) = aφ(x)b
for all a ∈ A, b ∈ B, x ∈ E.
Furthermore we will write CB (A,B) (E, F ) for the set of all completely bounded (A, B)-
module homomorphisms between E and F . The space CB (A,B) (E) with the composition
of operators as multiplication is a Banach algebra.
Let A1 , A2 ⊂ B(H) be C ∗ -algebras such that 1lH ∈ A1 , A2 . Let further A ⊂ A1 ∩ A2
be a unital ∗ -subalgebra of A1 and A2 with 1lH ∈ A. An A-bimodule homomorphism
Φ : A1 → A2
is called self-adjoint if
Φ(x)∗ = Φ(x∗ )
for all x ∈ A1 .
Dealing with completely bounded module homomorphisms, we have at our disposal
a representation theorem, an extension theorem and the following decomposition
theorem of Wittstock ([Wit81, Satz 4.5] and cf. [Pau86, Thm. 7.5]):
Let A, E and F be unital C ∗ -algebras. Let moreover F be injective, and A be a
subalgebra of E and F with 1lE = 1lF = 1lA . Then for each self-adjoint completely
bounded A-bimodule homomorphism φ : E → F , there exist two completely positive
A-bimodule homomorphisms φ1 and φ2 sharing the properties φ = φ1 −φ2 and kφkcb =
kφ1 + φ2 kcb .
Consider two von Neumann algebras M and N , and two C ∗ -algebras A1 , A2 ⊂ B(H),
where 1lH ∈ A1 , A2 and A1 ∪ A2 ⊂ M ∩ N . We then have the decomposition theorem
of Tomiyama-Takesaki (cf. [Tak79, Def. 2.15]): Each operator φ ∈ CB (A1 ,A2 ) (M, N )
has a unique decomposition φ = φσ + φs , φσ , φs ∈ CB (A1 ,A2 ) (M, N ) normal resp.
singular, where kφσ kcb , kφs kcb ≤ kφkcb . We thus obtain the algebraically direct sum
decomposition:
CB (A1 ,A2 ) (M, N ) = CB σ(A1 ,A2 ) (M, N ) ⊕ CB s(A1 ,A2 ) (M, N ). (2)
Here, the notions ”normal” and ”singular”, repectively, are built in analogy to the
framework of linear functionals on a von Neumann algebra M .24
We list some basic facts about the spaces and mappings mentioned in (2):
24
Let M∗ denote the (unique) predual of M . Then we have the `1 -direct sum decomposition
M ∗ = M∗ ⊕`1 (M ∗ )s
of M ∗ into normal (i.e. w∗ -continuous) and singular functionals. [In the literature, one usually writes
M∗⊥ instead of M ∗s , corresponding to M∗ (= M ∗σ ).] Analogously, an operator φ ∈ B(M, N ), M , N von
Neumann algebras, is called normal (i.e. w∗ -w∗ -continuous), if φ∗ (N∗ ) ⊂ M∗ , and it is called singular,
if φ∗ (N∗ ) ⊂ M ∗s .
(a) In case M = N , all the spaces in (2) are Banach algebras.
(b) The following properties of φ are hereditary for the normal part φσ and the singular
part φs : completely positive, homomorphism, ∗ -homomorphism.
(c) If α ∈ Aut(M ) and β ∈ Aut(N ) are ∗ -automorphisms, we have (βφα)σ = βφσ α

and (βφα)s = βφs α .
(d) For φ ∈ CB (B(H)), H a Hilbert space, we have: φ ∈ CB s (B(H)) ⇔ φ|K(H) ≡ 0 .
Let H be a Hilbert space, and let A1 , A2 ⊂ B(H) be two C ∗ -algebras with 1lH ∈ A1 , A2 .
Then we obtain [Pet97, Prop. 4.2.5]:
cb
CB σ(A1 ,A2 ) (B(H)) = CB (A1 ,A2 ) (K(H), B(H)) (3)
cb
CB s(A1 ,A2 ) (B(H)) = CB (A1 ,A2 ) (Q(H), B(H)) (4)
completely isometrically, where Q(H) = B(H)/K(H) denotes the Calkin algebra.

Let X be an arbitrary operator space. Then the space of all completely bounded
(A1 , A2 )-module homomorphisms between X and B(H) can be identified with the dual
of a module Haagerup tensor product in the following way ([Pet97, p. 67], cf. also
[ER91, Cor. 4.6], [Ble92b, Prop. 2.3]):
cb
CB (A1 ,A2 ) (X, B(H)) = (RH ⊗hA1 X ⊗hA2 CH )∗
completely isometrically. Hence we see that CB (B(H)) itself and (looking at (3), (4)),
just so, CB σ(A1 ,A2 ) (B(H)) and CB s(A1 ,A2 ) (B(H)) are dual operator spaces [Pet97, p. 70].
5.3 Operator algebras

In analogy to concrete operator spaces we define (cf. [BRS90, Def. 1.1]): An operator
algebra is a closed, not necessarily self-adjoint subalgebra X of B(H) (H a Hilbert
space).
Example: For selfadjoint X we have the theory of C ∗ -algebras.
As in the operator space situation, one can also adopt an abstract point of view: here,
this leads to considering Banach algebras which are operator spaces and are equipped
with a multiplication compatible with the operator space structure. These provide an
abstract characterization of the (concrete) operator algebras (cf. below: analogue of
Ruan’s theorem for operator algebras).
If X, Y , Z are operator spaces, and if Φ : X × Y → Z is bilinear, we can define
another bilinear map in the following way (cf.: Amplification of bilinear mappings):
Φ(n,l) : Mn,l (X) × Ml,n (Y ) → Mn (Z)

 
Xl
([xij ] , [yjk ]) 7→  Φ(xij , yjk ) (l, n ∈ IN).
j=1
This is called the bilinear amplification25 of Φ.

Φ is called completely bounded if kΦkcb := supn kΦ(n,n) k < ∞, and completely
contractive if kΦkcb ≤ 1.26 Compare this definition with the approach presented in
Completely bounded bilinear Mappings.
[In the sequel, for Banach algebras with unit e, we will require kek = 1.] An operator
space (X, k·kn ) with a bilinear, associative and completely contractive map m : X×X →
X, the multiplication, is called an abstract operator algebra (cf. [BRS90, Def. 1.4]).
Here, the multiplication on Mn (X) is just the matrix multiplication mn .
In the unital case m is automatically associative [BRS90, Cor. 2.4].
We have an analogue of Ruan’s theorem ([Ble95, Thm. 2.1], cf. also [BRS90, Thm.
3.1]): Let A be a unital Banach algebra and an operator space. Then A is completely
isometrically isomorphic to an operator algebra if and only if the multiplication on A is
completely contractive.
This yields the following stability result:
1.) The quotient of an operator algebra with a closed ideal is again an operator algebra
[BRS90, Cor. 3.2].
With this at hand, one deduces another important theorem on hereditary properties of
operator algebras:
2.) The class of operator algebras is stable under complex interpolation [BLM95, (1.12),
p. 320].
Adopting a more general point of view than in the Ruan type Representation The-
orem above, one obtains the following [Ble95, Thm. 2.2]: Let A be a Banach algebra
and an operator space. Then A is completely isomorphic to an operator algebra if and
only if the multiplication on A is completely bounded. (cf. the chapter Examples!)
Basic examples of operator algebras are provided by the completely bounded maps
on some suitable operator spaces. More precisely, for an operator space X, one obtains
the following [Ble95, Thm. 3.4]: CB (X) with the composition as multiplication, is
completely isomorphic to an operator algebra if and only if X is completely isomorphic
to a column Hilbert space. – An analogue statement holds for the isometric case.
In the following result, for operator algebras A and B, the assumption that A and
B be (norm-) closed, is essential (in contrast to the whole rest) [ER90b, Prop. 3.1]:
A unital complete isometry ϕ between A ⊂ B(H) and B ⊂ B(K) (H, K Hilbert spaces),
where 1lB(H) ∈ A, 1lB(K) ∈ B, is already an algebra homomorphism.
Examples
In the sequel we will equip the spaces `p (1 ≤ p ≤ ∞) with the pointwise product and
consider them as Banach algebras. We will further consider the Schatten classes Sp
(1 ≤ p ≤ ∞) endowed with either the usual multiplication or the Schur product.
25
In the literature, e.g. in [BRS90, p. 190], the bilinear amplification Φ(n,n) is often referred to as the
amplification and is noted by Φ(n) .
26
In order to define the notion of complete boundedness of bilinear mappings, it suffices to consider
only the Φ(n,n) instead of all the Φ(n,l) ; this definition is usually chosen in the literature about completely
bounded bi- and, analogously, multilinear maps [BRS90, p. 190], [CES87, p. 281].
1. The space `2 [BLM95, Thm. 2.1]

With the following operator space structures, `2 is completely isometrically iso-
morphic to an operator algebra: R`2 , C`2 , OH `2 , R`2 ∩ C`2 , MAX `2 .
More generally: The space `2 is completely isometrically isomorphic to an operator
algebra, if endowed with an operator space structure which dominate dominates
both R`2 and C`2 .
With the following operator space structures, `2 is not completely isomorphic to
an operator algebra: R`2 + C`2 , MIN `2 .
More generally: The space `2 is not completely isomorphic to an operator algebra,
if endowed with an operator space structure which is dominate dominated by
both R`2 and C`2 .
2. The spaces 27 MIN (`p ), MAX (`p ) and O`p = (MIN (`∞ ), MAX (`1 )) 1
p
In the extreme cases p = 1 resp. p = ∞, we have two opposite results [BLM95,

Thm. 3.1]:
(a) Equipped with any operator space structure, `1 is completely isometrically

isomorphic to an operator algebra.
(b) MIN (`∞ ) is, up to complete isomorphy, the only operator algebra structure
on `∞ .
For 1 ≤ p ≤ ∞ the following holds true (cf. [BLM95, Thm. 3.4]):
(a) MIN (`p ) is completely isomorphic to an operator algebra if and only if p = 1

or p = ∞.
(b) MAX (`p ), in case 1 ≤ p ≤ 2, is completely isometrically isomorphic to an
operator algebra. In all the other cases MAX (`p ) is not completely isomorphic
to an operator algebra.
On the contrary, the operator space structure on the `p spaces obtained via
complex interpolation always defines an operator algebra structure. More
precisely [BLM95, Cor. 3.3]:
For each 1 ≤ p ≤ ∞, O`p is completely isometrically isomorphic to an operator
algebra.
3. The Schatten classes Sp

We write OSp for the operator space structure defined on Sp by G. Pisier. This
operator space structure is obtained by complex interpolation between S∞ = K(`2 )
and S1 = K(`2 )∗ .
27
For the construction of the operator spaces O`p compare the chapter on complex interpolation.
(a) Let us first consider the usual product on the Schatten classes Sp . Here we
have the following negative result [BLM95, Thm. 6.3]: For each 1 ≤ p < ∞,
the operator space OSp with the usual product is not completely isomorphic
to an operator algebra.
(b) Consider now the Schur product on the Schatten classes Sp . Here we obtain
positive results, even for different operator space structures:
(b1) MAX (Sp ) with the Schur product is, in case 1 ≤ p ≤ 2, completely
isometrically isomorphic to an operator algebra [BLM95, Thm. 6.1].
(b2) OSp with the Schur product is, in case 2 ≤ p ≤ ∞, completely isometri-
cally isomorphic to an operator algebra [BLM95, Cor. 6.4].
Caution is advised: The space OS1 (and likewise OS1op ), whether endowed with the
usual or the Schur product, is not completely isomorphic to an operator algebra
[BLM95, Thm. 6.3].
6 Tensor Products
An operator space tensor product is an operator space whose structure is deduced from
the operator space structure of the factors. Operator space tensor products are defined
for all operator spaces and have functorial properties. On the tensor product of two fixed
operator spaces one usually considers operator space norms which are cross norms.
A lot of spaces, especially spaces of mappings, may be considered as operator space
tensor products of simpler ones. The theory of operator space tensor products follows
the lines of the theory of tensor products of Banach spaces. But at some points tensor
products of operator spaces have new properties not found for tensor products of Banach
spaces or even better properties as their counterparts. So in some cases the theory of
operator space tensor products gives solutions to problems not solvable within the theory
of Banach spaces (cp. [ER90a, Thm. 3.2]).
The Haagerup tensor product ⊗h has a variety of applications in the theory of op-
erator spaces and completely bounded operators.
∨
The injective operator space tensor product ⊗ is the least28 and the projective oper-
∧
ator space tensor product ⊗ is the greatest 29 among all operator space tensor products.
[BP91, Prop. 5.10].
On the algebraic tensor product X ⊗ Y of operator spaces X, Y one can compare an
operator space tensor norm k · kα with the injective tensor norm k · kλ and the projective
tensor norm k · kγ of normed spaces:
k · kλ ≤ k · k∨,1 ≤ k · kα,1 ≤ k · k∧,1 ≤ k · kγ .

28
i.e. the injective operator space tensor norm is minimal among all operator space tensor norms.
29
i.e. the projective operator space tensor norm is maximal among all operator space tensor norms.
6.1 Operator space tensor products

An operator space tensor product is the completion of the algebraic tensor product with
respect to an operator space tensor norm.
An operator space tensor norm k · kα is defined for each pair (X, Y ) of opera-
tor spaces and endows their algebraic tensor product X ⊗ Y with the structure of an
matrix normed space (X ⊗ Y, k · kα ) thuch that the following two properties 1 and 2
[BP91, Def. 5.9]. hold.
The completion is called the α-operator space tensor product of X and Y and
is denoted by X ⊗α Y .
1. For the complex numbers holds
Cl ⊗α Cl = C
l.
2. For all S ∈ CB (X1 , X2 ) and T ∈ CB (Y1 , Y2 ) the operator S⊗T : X1 ⊗Y1 → X2 ⊗Y2
has a continuous extension
S ⊗α T ∈ CB (X1 ⊗α Y1 , X2 ⊗α Y2 ).
The bilinear mapping
⊗α : CB (X1 , X2 ) × CB (Y1 , Y2 ) → CB (X1 ⊗α Y1 , X2 ⊗α Y2 )

(S, T ) 7→ S ⊗α T
is jointly completely contractive.30
Property 2 may be replaced by the assumptions 3 and 4.31

30
i.e., let [Sij ] ∈ Mp (CB (X1 , X2 )), [Tkl ] ∈ Mq (CB (Y1 , Y2 )), p, q ∈ IN, then the norm of the linear
operator
[Sij ⊗α Tkl ] ∈ Mpq (CB (X1 ⊗α Y1 , X2 ⊗α Y2 ))
is estimated by
k[Sij ⊗ Tkl ]kcb ≤ k[Sij ]kcb k[Tkl ]kcb .
Remark: This is indeed an equality.

31
(2) ⇒(3),(4): Condition 3 is a special case of 2. Let I ∈ Mp (CB (lMp (X), X)), J ∈
Mq (CB (lMq (Y ), Y )) be matrices, which are algebraically the the identical mappings of the vector spaces
Mp (X) respectivly Mq (Y ). By assumption 2 we have
I ⊗α J ∈ Mpq (CB (lMp (X) ⊗α lMq (Y ), X ⊗α Y ))

= M1 (CB (lMp (X) ⊗α lMq (Y ), lMpq (X ⊗α Y ))),
kI ⊗α Jkcb ≤ kIkcb kJkcb = 1.
Now I ⊗α J : lMp (X) ⊗α lMq (Y ) → lMpq (X ⊗α Y ) is the shuffle-map in 4.

(3),(4)⇒(2): Let [Sij ] ∈ Mp (CB (X1 , X2 )), [Tkl ] ∈ Mq (CB (Y1 , Y2 )), p, q ∈ IN, and S ∈
3. An operator space tensor product ⊗α is functorial: For all S ∈ CB (X1 , X2 ) and

T ∈ CB (Y1 , Y2 ) the operator S ⊗T : X1 ⊗Y1 → X2 ⊗Y2 has a continuous extension
S ⊗α T ∈ CB (X1 ⊗α Y1 , X2 ⊗α Y2 ),
and
kS ⊗α T kcb ≤ kSkcb kT kcb .
Remark: This is indeed an equality kS ⊗α T kcb = kSkcb kT kcb .
4. The algebraic shuffle isomorphism Mp (X)⊗Mq (Y ) ∼

= Mpq (X ⊗Y ) has a continuous
extension to a complete contraction:
lMp (X) ⊗α lMq (Y ) → lMpq (X ⊗α Y ).
This complete contraction is called the shuffle map of the α-operator space tensor
product.
Condition 4 is equivalent to the following two conditions: The shuffle mappings
lMp (X) ⊗α Y → lMp (X ⊗α Y ),

X ⊗α lMq (Y ) → lMq (X ⊗α Y )
are completely contractive.

Operator space tensor producte may have further special properties:
An operator space tensor product ⊗α is called

cb
symmetric, if X ⊗α Y = Y ⊗α X is a complete isometry;
cb
associative, if (X ⊗α Y ) ⊗α Z = X ⊗α (Y ⊗α Z) is acomplete isometry;
injective, if for all subspaces X1 ⊂ X, Y1 ⊂ Y the map X1 ⊗α Y1 ,→ X ⊗α Y is

acomplete isometry;
projective, if for all subspaces X1 ⊂ X, Y1 ⊂ Y the map X ⊗α Y → X/X1 ⊗α Y /Y1

is a complete quotient map;
CB (X1 , lMp (X2 )), T ∈ CB (Y1 , lMq (Y2 )) the corresponding operators. By 3 holds
S ⊗α T ∈ CB (X1 ⊗α Y1 , lMp (X2 ) ⊗α lMq (Y2 )),

kS ⊗α T kcb ≤ kSkcb kT kcb = k[Sij ]kcb k[Tkl ]kcb .
We apply the shuffle map A : lMp (X2 ) ⊗α lMq (Y2 ) → lMpq (X2 ⊗α Y2 ) and obtain from 4
[Sij ⊗α Tkl ] = A(S ⊗α T ) ∈ M1 (CB (X1 ⊗α Y1 , lMpq (X2 ⊗α Y2 ))) = Mpq (CB (X1 ⊗α Y1 , X2 ⊗α Y2 ),
k[Sij ⊗α Tkl ]kcb ≤ kS ⊗α T kcb ≤ k[Sij ]kcb k[Tkl ]kcb .
Hence ⊗α is jointly completely bounded.

self dual, if the algebraic embedding X ∗ ⊗ Y ∗ ⊂ (X ⊗α Y )∗ has a completely isometric

extension X ∗ ⊗α Y ∗ ⊂ (X ⊗α Y )∗ .
In many applications one finds the Haagerup-tensor product. It is not symmetric,

but associative, injective, projective and self dual.
cross norms
Sometimes one considers an operator space norm on the algebraic tensor product of two
fixed operator spaces. Then one usually demands that this norm and its dual norm are
at least cross norms. Operator space tensor norms always have these properties.
An operator space norm k · kα on the algebraic tensor product X ⊗ Y of two operator
spaces X and Y is said to be a cross norm, if
kx ⊗ ykα,pq = kxkp kykq
for all p, q ∈ IN, x ∈ Mp (X), y ∈ Mq (Y ) holds.

cb
For cross norms Cl ⊗α X = X is completely isometric.
For an operator space norm k·kα on the algebraic tensor product of two fixed operator
spaces X and Y one usually asks for the following three properties (i)–(iii).32
(i) k · kα is a cross norm.
(ii) Let ϕ ∈ X ∗ , ψ ∈ Y ∗ be linear functionals and
ϕ⊗ψ :X ⊗Y → Cl
hx ⊗ y, ϕ ⊗ ψi := hx, ϕihy, ψi
where x ∈ X, y ∈ Y . their tensorproduct. The tensor product ϕ ⊗ ψ has a

continuous linear extension to X ⊗α Y .
Then the dual operator space norm k · kα∗ is defined on the algebraic tensor product
X∗ ⊗ Y ∗ by the algebraic embedding
X ⊗ Y ⊂ (X ∗ ⊗α Y ∗ )∗ .
(iii) The dual operator space norm k · kα∗ is a cross norm.
There is a smallest operator space norms among the operator space norms on X ⊗ Y ,
for which k · kα and the dual norm k · kα∗ are cross norms. This is the injective operator
space tensor norm k · k∨ . [BP91, Prop. 5.10].
32
The conditions (i)-(iii) are equivalent to the following: The bilinear maps
X × Y → X ⊗α Y, (x, y) 7→ x ⊗ y
∗ ∗ ∗
X × Y → (X ⊗α Y ) , (ϕ, ψ) 7→ ϕ ⊗ ψ
are jointly completely bounded.

There is a greatest operator space norm among the operator space norms on X ⊗ Y ,
for which k · kα and the dual norm k · kα∗ are cross norms. This is the projective operator
space tensor norm k · k∧ [BP91, Prop. 5.10].
On the algebraic tensor product X ⊗ Y one can compare the operator space norms
k · kα for which k · kα and the dual norm k · kα∗ are cross norms with the injective tensor
norm k · kλ and the projective tensor norm k · kγ of normed spaces:
k · kλ ≤ k · k∨,1 ≤ k · kα,1 ≤ k · k∧,1 ≤ k · kγ .
6.2 Injective operator space tensor product

The representations of two operator spaces X in B(H) and Y in B(K) yield a repre-
sentation of the algebraic tensor product of X and Y in B(H ⊗2 K). The operator
space structure obtained in this way turns out to be independent of the representations
chosen. It is called the injectiveoperator space tensor product of X and Y and
∨
is denoted by X ⊗ Y [BP91, p. 285]. Hence, in the case of C ∗ –algebras, the injective
operator space tensor product and the minimal C ∗ –tensor product coincide.33
By means of the duality of tensor products we obtain a formula [BP91, Thm. 5.1]
for the injective operator space tensor norm of an element u ∈ Mn (X ⊗ Y ) which is
representation free:
kuk∨ = sup khu, ϕ ⊗ ψikMnkl ,
where k, l ∈ IN, ϕ ∈ Ball(Mk (X ∗ )) and ψ ∈ Ball(Ml (Y ∗ )).

Interpreting, as is usual, the elements of the algebraic tensor product as finite rank
operators we have the completely isometric embeddings [BP91, Cor. 5.2]
∨ ∨
X ⊗ Y ,→ CB (X ∗ , Y ) resp. X ⊗ Y ,→ CB (Y ∗ , X).
The injective operator space tensor norm is the least cross norm whose dual norm
again is a cross norm.
The injective operator space tensor product is symmetric, associative and injective.
But it is not projective [BP91, Cor. 5.2].
The injective norm is the dual norm of the projective operator space tensor norm
[BP91, Thm. 5.6]; but the projective operator space tensor norm is not in general the
dual of the injective operator space tensor norm even if one of the two spaces involved
is finite dimensional [ER90a, p. 168], [ER91, p. 264].
33
For this reason the injective operator space tensor product is also called spatial tensor product of
operator spaces and denoted by X ⊗min Y .
Some formulae for the injective operator space tensor product

If E and F are normed spaces, then [BP91, Prop. 4.1]
∨ cb
M IN (E) ⊗ M IN (F ) = M IN (E ⊗λ F )
6.2.1 Exact operator spaces

We consider now the exact sequence
0 ,→ K(`2 ) ,→ B(`2 ) → Q(`2 ) → 0.
An operator space X is said to be exact [Pis95, §1], if the short sequence of injective
tensor products
∨ ∨ ∨
0 ,→ X ⊗ K(`2 ) ,→ X ⊗ B(`2 ) → X ⊗ Q(`2 ) → 0
is again exact. Then tensorizing with such an operator space preserves the exactness
of arbitrary exact sequences of C ∗ -algebras (for C ∗ -algebras cf [Kir83]). Obviously, all
finite dimensional operator spaces are exact.
Exactness is inherited by arbitrary subspaces. The injective tensor product of two
exact operator spaces is again exact. For an exact space X we are given a degree of
exactness by the quantity
∨ ∨ ∨
ex(X) = kX ⊗ Q(H) → (X ⊗ B(H))/(X ⊗ K(H))k.
We have 1 ≤ ex(X) < ∞ [Pis95, §1], because the mapping
∨ ∨ ∨
(X ⊗ B(H))/(X ⊗ K(H)) → X ⊗ Q(H)
is a complete contraction. For non-exact operator spaces X we put ex(X) = ∞.
For an exact C ∗ -algebra 34 A we have ex(A) = 1.
For an operator space X we have:
ex(X) = sup{ex(L) : L ⊂ X, dim L < ∞}.
so we can confine our examinations to finite dimensional spaces. From this it is also
immediate that: ex(X0 ) ≤ ex(X) if X0 ⊂ X. One has for finite dimensional operator
spaces X1 and X2 the complete variant of the Banach-Mazur distance
dCB (X1 , X2 ) = inf{kϕkcb kϕ−1 kcb }
( the infimum is taken over all isomorphisms ϕ from X1 to X2 ). Via this Banach-Mazur
distance we can define the quantity
dSK (X) := inf{dCB (X, L), dim(L) = dim(X), L ⊂ Mn , n ∈ IN}
p
According to [Pis95, Thm. 1] ex(X) = dSK (X) holds, and ex(X) ≤ dim(X).
34
A characterization of exact C ∗ -algebras is given in [Kir94] and [Kir95].
6.3 Projective operator space tensor product

∧
The projective operator space tensor product X ⊗ Y of two operator spaces X
and Y is characterized by the following complete isometry:
∧ cb cb
(X ⊗ Y )∗ = CB (X, Y ∗ ) = CB (Y, X ∗ ).
where linear mappings are identified in the usual way with bilinear forms.
One can also characterize the projective operator space tensor product by the fol-
lowing universal property [BP91, Def. 5.3]
∧ cb
CB (X ⊗ Y, Z) = JCB (X × Y ; Z),
where Z is an operator space.

Here, JCB (X × Y ; Z) denotes the operator space of jointly completely bounded bi-
linear mappings.
One also has an explicit expression for the projective operator space tensor norm of
an element u ∈ Mn (X ⊗ Y ): (cf. [ER91, Formel (2.10)])
kuk∧ = inf {kαkkxkp kykq kβk : u = α(x ⊗ y)β} ,
where p, q ∈ IN, x ∈ Mp (X), y ∈ Mq (Y ) and α ∈ Mn,pq , β ∈ Mpq,n .

The projective operator space tensor norm is symmetric, associative and projective
[ER91, p. 262]. But it is not injective.
The projective operator space tensor norm is the greatest operator space tensor norm
which is a cross norm [BP91, Thm. 5.5].
Its dual norm is the injective operator space tensor norm [BP91, Thm. 5.6]; but the
projective operator space tensor norm is not in general the dual of the injective operator
space tensor norm even if one of the two spaces involved is finite dimensional [ER90a,
p. 168], [ER91, p. 264].
Some formulae for the projective operator space tensor product

1. If E and F are normed spaces, we have [BP91, Prop. 4.1]
∧ cb
M AX(E) ⊗ M AX(F ) = M AX(E ⊗γ F )
completely isometrically.
2. Taking the projective operator space tensor product of various combinations of

the column Hilbert spaces C, the row Hilbert spaces R and an arbitrary operator
space X one obtains the following completely isometric identifications: [ER91],
[Ble92b, Prop. 2.3]
∧ cb
(i) X ⊗ CH = X ⊗h CH
∧ cb
(ii) RH ⊗ X = RH ⊗h X
∧ cb ∨ cb cb
(iii) CH ⊗ CK = CH ⊗ CK = CH ⊗h CK = CH⊗2 K
∧ cb ∨ cb cb
(iv) RH ⊗ RK = RH ⊗ RK = RH ⊗h RK = RH⊗2 K
∧ cb cb
(v) RH ⊗ CK = RH ⊗h CK = T (H, K)
∧ cb ∧ ∧ cb
(vi) X ⊗ T (H, K) = RH ⊗ X ⊗ CK = RH ⊗h X ⊗h CK
cb ∧ ∧ cb
(vii) CB (X, B(K, H)) = (RH ⊗ X ⊗ CK )∗ = (RH ⊗h X ⊗h CK )∗ ,
where the space of trace class operators T (H, K) is endowed with its natural op-
cb
erator space structure: T (H, K) := K(K, H)∗ .
3. Let M and N be von Neumann algebras and denote by M ⊗N the von Neumann
tensor product35 . For the preduals one has
∧ cb
M∗ ⊗ N∗ = (M ⊗N )∗
completely isometrically [ER90a].

Let G and H be locally compact topological groups and denote by VN (G),
VN (H) the corresponding group von Neumann algebras.36 It is well-known that
VN (G) ⊗ VN (H) = V N (G × H). Since the Fourier algebra37 A(G) can be iden-
tified with the predual of the group von Neumann algebra VN (G) [Eym64], this
implies that
∧ cb
A(G) ⊗ A(H) = A(G × H)
holds completely isometrically38 [ER90a].
6.4 The Haagerup tensor product

The Haagerup tensor product was first introduced by Effros and Kishimoto [EK87] for
C∗ -algebras generalizing the original work of U. Haagerup [Haa80].
The Haagerup tensor product X⊗h Y of two operator spaces X and Y is characterized
by the complete isometry
cb
(X ⊗h Y )∗ = CB (X × Y ; Cl ),
35
If M ⊂ B(H), N ⊂ B(K) are von Neumann algebras, M ⊗N is defined to be the closure in the weak
operator topology of the algebraic tensor product M ⊗ N ⊂ B(H ⊗2 K).
36
The group von Neumann algebra VN (G) of a locally compact group G is defined to be the von
Neumann algebra generated by the left regular representation of G in B(L2 (G)).
37
For a locally compact group G, the set {f ∗ ǧ|f, g ∈ L2 (G)} ⊂ C0 (G) turns out to be a linear
space and even an algebra (with pointwise multiplication). Its completion with respect to the norm
kuk = inf{kf k2 kgk2 | u = f ∗ ǧ} is a Banach algebra and is called the Fourier algebra of G.
38
If G is a locally compact abelian group, then A(G) is identified – via the Fourier transform –
with L1 (G),
b where G b denotes the dual group of G. Thus, for locally compact groups G and H, the
∧ cb
identification A(G) ⊗ A(H) = A(G × H) can be thought of as a non commutative analogue of the
well-known classical identification L1 (G) ⊗γ L1 (H) = L1 (G × H).
where bilinear forms are identified with linear maps in the usual fashion.
We can also characterize the Haagerup tensor product by the following universal
property: For an operator space Z we have
cb
CB (X ⊗h Y ; Z) = CB (X × Y ; Z)
Here, CB (X × Y ; Z) denotes the operator space of completely bounded bilinear
mappings.
For operator spaces X and Y the Haagerup operator space tensor norm of
u ∈ Mn (X ⊗ Y ) is explicitly given by (cf. [ER91, Formel (2.11)], [BP91, Lemma 3.2])
kukh = inf kxkn,l kykl,n ,
where l ∈ IN, x ∈ Mn,l (X), y ∈ Ml,n (Y ) and u is the tensor matrix product u = x y.
The Haagerup tensor product X ⊗h Y then of course is the completion of the algebraic
tensor product X ⊗Y with respect to this operator space tensor norm. There are several
other useful formulae 2 3, 4 for the Haagerup norm.
The Haagerup tensor product is not symmetric as shown by concrete examples. But
it is associative, injective [PS87, p. 272; Thm. 4.4], [BP91, Thm. 3.6], projektiv [ER91,
Thm. 3.1] and selfdual [ER91, Thm. 3.2]. Thus the embedding
X ∗ ⊗h Y ∗ ,→ (X ⊗h Y )∗
is a complete isometry.
The extension of the identity mapping on the algebraic tensor product of two oper-
ator spaces X, Y from the Haagerup tensor product into the injective tensor product is
injective. One therefore obtains a canonical embedding
∨
X ⊗h Y ⊂ X ⊗ Y.
The complex interpolation of operator spaces and the Haagerup tensor product com-
mute [Pis96, Thm. 2.3]. Let (X0 , X1 ) and (Y0 , Y1 ) be compatible pairs of operator
spaces. Then (X0 ⊗h Y0 , X1 ⊗h Y1 ) is a compatible pair of operator spaces and we have
completely isometrically
cb
(X0 ⊗h Y0 , X1 ⊗h Y1 )ϑ = (X0 , X1 )ϑ ⊗h (Y0 , Y1 )ϑ
for 0 ≤ ϑ ≤ 1.
On normed spaces there is no tensor norm which at the same time is associative,
injective, projective and selfdual. The Haagerup tensor product can be interpreted as
a generalization of the H-tensor product introduced by Grothendieck39 for normed
39
This tensor norm also is known as γ2 [Pis86].
spaces E and F [BP91, pp. 277-279, Prop. 4.1]. In fact, on the first matrix level we
have:
MIN (E) ⊗h MIN (F ) = E ⊗H F,
MAX (E) ⊗h MAX (F ) = E ⊗H ∗ F
isometrically. The non-associativity of the H-tensor product is reflected by the fact that
in general MIN (E) ⊗h MIN (F ) and MIN (E ⊗H F ) are not completely isometric.
Some formulae for the Haagerup tensor product

1. Let A and B be C∗ -algebras in B(H). Then on the algebraic tensor product A ⊗ B
the Haagerup tensor norm is explicitly given by
1 1
n P o
kukh := inf k nν=1 aν a∗ν k 2 k nν=1 b∗ν bν k 2 : u = nν=1 aν ⊗ bν ,
P P
where n ∈ IN, aν ∈ A, bν ∈ B.
The Haagerup norm of Pnν=1 aν ⊗bν ∈ A⊗B equals the cb–norm of the elementary
P
operator B(H) 3 x 7→ nν=1 aν xbν . The Haagerup tensor product A ⊗h B is the
completion of the algebraic tensor product A ⊗ B with respect to the above norm.
The following more general definition in particular yields a completely isometric
embedding A ⊗h B ,→ CB (B(H)).
2. We have
kukh = inf{ kκ=1 kxκ kkyκ k : u = kκ=1 xκ yκ }
P P
where k, l ∈ IN, xκ ∈ Mn,l (X), yκ ∈ Ml,n (Y ). In fact, one summand suffices [BP91,
Lemma 3.2].
3. For elements u ∈ Mn (X ⊗ Y ) in the algebraic tensor product there is an l ∈ IN
and elements x ∈ Mn,l (X), y ∈ Ml,n (Y ) such that
u = xy
kukh = kxkkyk.
The infimum occuring in the formula describing the norm in this case is actually
a minimum [ER91, Prop. 3.5].
4. The Haagerup norm of an element u ∈ Mn (X ⊗h Y ) can also be expressed using
a supremum:40
kukh = sup khu, ϕ ψikMn2 ,
40
For x ∈ X, y ∈ Y , we have:
" l #
X
(ϕ ψ)(x ⊗ y) = hx ⊗ y, ϕ ψi = hx ⊗ y, ϕij ⊗ ψjk i = ϕ(x)ψ(y) ∈ Mn .
j=1
Here, we used the definitions of two fundamental notions in operator space theory: the
tensor matrix multiplication ϕ ψ of mappings ϕ, ψ and the joint amplification of the duality of tensor
products.
where l ∈ IN, ϕ ∈ Mn,l (X ∗ ), ψ ∈ Ml,n (Y ∗ ), kϕk = kψk = 1 [ER91, Prop. 3.4].

5. From the definition of the Haagerup norm one easily deduces that the shuffle-map
lMp (X) ⊗h lMq (Y ) → lMpq (X ⊗h Y )

is a complete contraction. Hence the Haagerup tensor product enjoys property 2
of an operator space tensor product.
But the shuffle-map is not an isometry in general as shown by the following ex-
ample:41
cb cb
lMn (Cl ) ⊗h lMn (Rl ) = lMl (lMn ((Tn )) 6= lMl (lMn (lMn )) = lMn (lMn (Cl ⊗h Rl )).
Since the bilinear mapping
⊗h : CB (X1 , X2 ) × CB (Y1 , Y2 ) → CB (X1 ⊗h X2 , Y1 ⊗h Y2 )
(S, T ) 7→ S ⊗h T
is contractive, it is jointly completely contractive.42
In fact, using 8, we see that it is even completely contractive.
6. For the row and column structure the shuffle-map even is a complete isometry.
We have the Lemma of Blecher and Paulsen [BP91, Prop. 3.5]:
cb
Cn (X) ⊗h Rn (Y ) = lMn (X ⊗h Y ).
In many cases it suffices to prove a statement about the Haagerup tensor product
on the first matrix level and then to deduce it for all matrix levels using the above
formula.43
41
Algebraically, we have on both sides the same spaces of matrices. But on the left side one obtains
the finer operator space structure Tn := lM∗n of the trace class:
cb cb
lMn (Cl ) ⊗h lMn (Rl ) = Cl (lMn ) ⊗h Rl (lMn ) = lMl (lMn ⊗h lMn )
cb
= lMl (Cn ⊗h Rn ⊗h Cn ⊗h Rn ) = lMl (Cn ⊗h Tn ⊗h Rn ) = lMl (Mn (Tn )).
On the right side, we get the coarser operator space structure of the matrices Mn :
cb
lMn (lMn (Cl ⊗h Rl ) = lMn (lMn (lMl )) = lMl (lMn (lMn )).
42
This follows from the equivalence of the property 2 and the properties 3 and 4 of operator space
tensor products.
43
This method can be applied to obtain this complete isometry itself. It is easy to see that on the
first matrix level we have
M1 (Cn (X) ⊗h Rn (Y )) = Mn (X ⊗h Y )
isometrically. From this the complete isometry follows – for all p ∈ IN we have:
Mp (Cn (X) ⊗h Rn (Y )) = M1 (Cp (Cn (X)) ⊗h Rp (Rn (Y ))) = M1 (Cpn (X) ⊗h Rpn (Y ))
= Mpn (X ⊗h Y ) = Mp (lMn (X ⊗h Y ))
isometrically .
Here we list some special cases of the Lemma of Blecher and Paulsen:
cb
Cn ⊗h Rn = lMn ,
cb
Cn ⊗h X = Cn (X),
cb
X ⊗h Rn = Rn (X),
cb
Cn ⊗h X ⊗h Rn = lMn (X).
7. In contrast to 6, for Rn ⊗h Cn one obtains the finer operator space structure of

the trace class ∧
cb cb
Tn := lM∗n = Rn ⊗ Cn = Rn ⊗h Cn .
For an operator space X we have [Ble92b, Prop. 2.3]:
cb ∧
Rn ⊗h X = Rn ⊗ X,
cb ∧
X ⊗h Cn = Cn ⊗ X,
cb ∧
Rn ⊗h X ⊗h Cn = Tn ⊗ X,
cb
Rn ⊗h X ∗ ⊗h Cn = lMn (X)∗ .
8. By the very construction the bilinear mapping X × Y → X ⊗h Y , (x, y) 7→ x ⊗ y

is a complete contraction.
Hence its amplification, the tensor matrix product
h : lMn,l (X) × lMl,n (Y ) → lMn (X ⊗h Y )

(x, y) 7→ x y,
also is a complete contraction. The linearization of the tensor matrix product

gives the complete contraction
lMn,l (X) ⊗h lMl,n (Y ) → lMn (X ⊗h Y ).
9. The bilinear mapping44
⊗h : CB (X1 , X2 ) × CB (Y1 , Y2 ) → CB (X1 ⊗h Y1 , X2 ⊗h Y2 )

(S, T ) 7→ S ⊗h T
44
The amplification of ⊗h is nothing but the tensor matrix multiplication h of operator matrices.
is completely contractive45 and gives rise to a complete contraction
CB (X1 , X2 ) ⊗h CB (Y1 , Y2 ) → CB (X1 ⊗h Y1 , X2 ⊗h Y2 ).
10. Let H and K be Hilbert spaces. Taking the Haagerup tensor product of the
column Hilbert space C and the row Hilbert space R one obtains completely iso-
metrically the space of compact operators K resp. of trace class46 operators T
[ER91, Cor. 4.4]:
cb
RH ⊗h CK = T (H, K),
cb
CK ⊗h RH = K(H, K).
This example also shows that the Haagerup tensor product is not symmetric.
6.5 Completely bounded bilinear mappings

In the case of bilinear mappings between operator spaces one has to distinguish between
two different notions of complete boundedness: on one hand we have the jointly com-
pletely bounded [BP91, Def. 5.3 (jointly completely bounded)] and, on the other
hand, the completely bounded bilinear mappings [CS87, Def. 1.1].
The class of completely bounded bilinear maps is is contained in the first one. These
notions are in perfect analogy to those of bounded bilinear forms on normed spaces. For
completely bounded bilinear mappings, we have at our disposal similar representation and
extension theorems47 as in the case of completely bounded linear maps. There are two
tensor products corresponding to the above two classes of bilinear mappings, namely
the projective and the Haagerup tensor product. Depending on the class of bilinear
maps, one uses different methods to define the amplification of a bilinear mapping
Φ : X × Y → Z.
45
Let S ∈ Mn (CB (X1 , X2 )), T ∈ Mn (CB (Y1 , Y2 )). For x ∈ Mp,q (X1 ), y ∈ Mq,p (Y1 ) we have
(S T )(p) (x y) = (S (p,q) (x)) (T (q,p) (y)),

k(S T )(p) (x y)kMpn (X2 ⊗h Y2 ) ≤ kS (p,q) (x)kkT (q,p) (y)k ≤ kSkcb kT kcb kxkkyk.
By the definition of the Haagerup norm (in X1 ⊗h Y1 ) we obtain
k(S T )(p) (x y)kMpn (X2 ⊗Y2 ) ≤ kSkcb kT kcb kx ykMp (X1 ⊗h Y1 ) ,

kS h T kcb ≤ kSkcb kT kcb .
This means that ⊗h is completely contractive.

cb
46
T is endowed with its natural operator space structure T (H, K) := K(K, H)∗ .
47
Extension theorems for completely bounded bilinear (and, more generally, multilinear) maps can be
derived from the injectivity of the Haagerup tensor product.
Amplification
In the literature, there are two different notions of an amplification of a bilinear mapping.
We shall call the first kind 1 of amplification the joint amplification. This joint
amplification is needed to obtain a matrix duality – which is fundamental in the duality
theory of operator spaces –, starting from an ordinary duality hX, X ∗ i.
The notion of joint amplification leads to the jointly completely bounded bilinear
maps as well as the projective operator space tensor product.
We will speak of the second kind 2 of an amplification as the amplification of a
bilinear mapping. This notion leads to the completely bounded bilinear maps and the
Haagerup tensor product.
In the sequel, we will use the notation Φ : X × Y → Z for a bilinear mapping and
Φ̃ : X ⊗ Y → Z for its linearization.
Both notions of an amplification of a bilinear map Φ are formulated in terms of the
amplification of its linearization:
Φ̃(n) : Mn (X ⊗ Y ) → Mn (Z).
1. The joint amplification of Φ produces the bilinear mapping
Φ(p×q) : (x, y) 7→ Φ̃(pq) (x ⊗ y) = [Φ(xij , ykl )] ∈ Mp (Mq (Z)) = Mpq (Z).
of the operator matrices x = [xij ] ∈ Mp (X) and y = [ykl ] ∈ Mq (Y ). Here, the
tensor product of operator matrices is defined via
x ⊗ y = [xij ] ⊗ [ykl ] := [xij ⊗ ykl ] ∈ Mpq (X ⊗ Y ) = Mp (Mq (X ⊗ Y )).
2. In the case of completely bounded bilinear maps one deals with the tensor matrix
multiplication [Eff87]
hP i
l
x y = [xij ] [yjk ] := j=1 xij ⊗ y jk ∈ Mn (X ⊗ Y )
of operator matrices x = [xij ] ∈ Mn,l (X) and y = [yjk ] ∈ Ml,n (Y ).

For more formulae, see: tensor matrix multiplication.
The (n,l)-th amplification of a bilinear map Φ : X × Y → Z is defined by
Φ(n,l) : Mn,l (X) × Ml,n (Y ) → Mn (Z)
hP i
l
(x, y) 7→ Φ̃(n) (x y) = j=1 Φ(xij , yjk ) ∈ Mn (Z)
for l, n ∈ IN, x = [xij ] ∈ Mn,l (X), y = [yjk ] ∈ Ml,n (Y ). In case n = l, we shortly

write48
Φ(n) := Φ(n,n) : Mn (x) ⊗ Mn (Y ) → Mn (Z)
(x, y) 7→ Φ(n) (x, y) := Φ̃(n) (x y).
48
In the literature, the amplification of a bilinear map often is defined only for quadratic operator
matrices and is called the amplification Φ(n) .
Nevertheless, we will also be dealing with the amplification for rectangular matrices since this permits
the formulation of statements where for fixed n ∈ IN all amplifications Φ(n,l) , l ∈ IN, are considered.
Jointly complete boundedness

Let X, Y, Z be operator spaces. A bilinear mapping Φ : X × Y → Z is called jointly
completely bounded [BP91, Def. 5.3 (jointly completely bounded)] if the norms
of the joint amplifications of Φ are uniformly bounded:
kΦkjcb := sup kΦ(p×q) (x ⊗ y)k < ∞,
where p, q ∈ IN, x ∈ Ball(Mp (X)), y ∈ Ball(Mq (Y )) [BP91, Def. 5.3]. The norm kΦkjcb
equals the norm kΦ̃kcb of the linearization
∧
Φ̃ : X ⊗ Y → Z
on the projective operator space tensor product.
JCB (X × Y ; Z) denotes the operator space consisting of the jointly completely
bounded bilinear maps. One obtains a norm on each matrix level by the identifica-
tion
Mn (JCB (X × Y ; Z)) = JCB (X × Y ; lMn (Z)).
We have
cb cb
JCB (X × Y ; Z) = CB (X, CB (Y, Z)) = CB (Y, CB (X, Z)).
completely isometrically. By taking the transposition Φt (y, x) := Φ(x, y) we obtain a
complete isometry
cb
JCB (X × Y ; Z) = JCB (Y × X; Z).
Complete boundedness
For the definition of the completely bounded bilinear maps we need the amplification
Φ(n) , the linearization Φ̃ : X ⊗ Y → Z. and the tensor matrix multiplication x y of
operator matrices x, y.
A bilinear mapping Φ : X × Y → Z, n ∈ IN is called completely bounded if
kΦkcb := sup kΦ(n) (x, y)k < ∞
where n ∈ IN, x ∈ Ball(Mn (X)), y ∈ Ball(Mn (Y )).
The norm kΦkcb equals the norm kΦ̃kcb of the linearization
Φ̃ : X ⊗h Y → Z
on the Haagerup tensor product.
Furthermore, the norms kΦkn are obtained using the tensor matrix products of all
49 rectangular matrices of n rows resp. n columns:
kΦkn := sup kΦ(n,l) (x, y)k

49
Note that the norm of the bilinear map
Φ(n) : Mn (x) ⊗ Mn (Y ) → Mn (Z)
in general is smaller than the norm kΦkn .
where l ∈ IN, x ∈ Ball(Mn,l (X)), y ∈ Ball(Ml,n (Y )).

We have
kΦkcb = sup {kΦkn : n ∈ IN} .
The norm kΦkn equals the norm kΦ̃(n) k of the amplification of the linearization
Φ̃(n) : Mn (X ⊗h Y ) → Mn (Z)
on the Haagerup tensor product.

A bilinear form Φ is already seen to be completely bounded if kΦk1 < ∞. Then we
have kΦkcb = kΦk1 .50
CB (X × Y ; Z) denotes the operator space consisting of completely bounded bilinear
maps. One obtains a norm on each matrix level using the identification
Mn (CB (X × Y ; Z)) = CB (X × Y ; lMn (Z)).
Corresponding to the completely bounded bilinear maps we have the linear maps
which are completely bounded on the Haagerup tensor product. The identification
cb
CB (X × Y ; Z) = CB (X ⊗h Y ; Z)

Completely bounded bilinear mappings are in particular jointly completely bounded.
The embedding CB (X × Y ; Z) ⊂ JCB (X × Y ; Z) is a complete contraction.
The transpose Φt (y, x) := Φ(x, y) of a completely bounded bilinear mapping Φ in
general is not completely bounded.51
For completely bounded bilinear (and, more generally, multilinear) maps Φ ∈
CB (A × B; B(H)) we have some generalizations of Stinespring’s representation theo-
rem.
Representation
Completely bounded bilinear forms were first studied on C∗ -algebras A, B [EK87]. For
a bilinear form Φ : A × B → Cl the following properties are equivalent:
(1) Φ is completely bounded.
(2) There is a constant c such that

∗ 12 1
Pl Pl Pl ∗
| j=1 Φ(aj , bj )| ≤ ck j=1 aj aj k k j=1 bj bj k
2
for all l ∈ IN, aj ∈ A, bj ∈ B.

50
More generally, the equation kΦkcb = kΦk1
holds for bilinear maps with values in a commutative C∗ -algebra A since every bounded linear map
taking values in A is automatically completely bounded and kΦkcb = kΦk [Loe75, Lemma 1]. For
bilinear maps Φ : X × Y → Mn (A) we have kΦkcb = kΦkn .
51
To this corresponds the fact that the Haagerup tensor product is not symmetric.
(3) There is a constant c and states ω ∈ S(A), ρ ∈ S(B) such that

1 1
|Φ(a, b)| ≤ c ω(aa∗ ) 2 ρ(b∗ b) 2
for all a ∈ A, b ∈ B.
(4) There exist ∗ -representations πω : A → B(Hω ) and πρ : A → B(Hρ ) with cyclic

vectors ξω ∈ Hω , ξρ ∈ Hρ and an operator T ∈ B(Hω , Hρ ) such that Φ(a, b) =
hT πω (a)ξω , πρ (b)ξρ i for all a ∈ A, b ∈ B.
One can choose c = kT k = kΦkcb and kξω k = kξρ k = 1.52
6.6 Module tensor products

So far the only module tensor product of operator modules that has been studied is the
module Haagerup tensor product [Rua89] [BMP].
6.6.1 Module Haagerup tensor product

Let X be a right operator module over a C ∗ -algebra A, Y a left A-operator module, and
W an operator space. A bilinear mapping Ψ : X × Y → W is called balanced, if the
equation
Ψ(x · a, y) = Ψ(x, a · y)
obtains for all x ∈ X, y ∈ Y, a ∈ A.
The module Haagerup tensor product is defined to be the operator space
X ⊗hA Y (which is unique up to complete isometry) together with a bilinear, completely
contractive, balanced mapping
⊗hA : X × Y → X ⊗hA Y,
such that the following holds true: For each bilinear, completely bounded balanced map
Ψ:X ×Y →W
there is a unique linear completely bounded map

e : X ⊗hA Y → W
Ψ
e ◦ ⊗hA = Ψ and kΨk
satisfying Ψ e cb = kΨkcb .
The module Haagerup tensor product can be realized in different ways:
1. Let
N := lin{(x · a) ⊗ y − x ⊗ (a · y) | a ∈ A, x ∈ X, y ∈ Y }.

The quotient space (X ⊗h Y ) N with its canonical matrix norms is an operator
space which satisfies the defintion of X ⊗hA Y [BMP].
52
For further references see [CS89, Sec. 4].
2. Let us denote byX ⊗A Y the algebraic module tensor product, i.e. the quotient
space (X ⊗alg Y ) N . For n ∈ IN and u ∈ Mn (X ⊗A Y ), by
( " l # )
X
pn (u) := inf kSkkT k u = Sik ⊗A Tkj , l ∈ IN, S ∈ Mnl (X), T ∈ Mln (Y )

k=1
we define a semi-norm on Mn (X ⊗A Y ). We obtain
Kern(pn ) = Mn (Kern(p1 )),
and the semi-norms pn give an operator space norm on (X⊗A Y )/Kern(p1 ) [Rua89].
The completion of this space satisfies the definition of X ⊗hA Y [BMP].
Examples
Let X be an operator space. Then the space of completely bounded (A1 , A2 )-
module homomorphisms between X and B(H) can be identified with the dual of a
module Haagerup tensor product in the following way ([Pet97, p. 67], cf. also [ER91,
Cor. 4.6], [Ble92b, Prop. 2.3]):
cb
CB (A1 ,A2 ) (X, B(H)) = (RH ⊗hA1 X ⊗hA2 CH )∗
7 Complete Local Reflexivity

An operator space X is called completely locally reflexive [EJR98, §1], if to each
finite dimensional subspace there is L ⊂ X ∗∗ a net of completely contractive mappings
ϕα : L → X that converges to the embedding L → X ∗∗ in the point weak∗ topology.53
This property is inherited by arbitrary subspaces. In general it is not preserved by
quotients. In the case that for example the kernel is an M-ideal (e.g. the twosided ideal
of a C ∗ -algebra) and that the original space is completely reflexive we have that the
quotient space is completely locally reflexive [ER94, Thm. 4.6].
Banach spaces are always locally reflexive (Principle of local reflexivity [Sch70]).
On the contrary, not all operator spaces are completely locally reflexive. For exam-
ple the full C ∗ -algebra of the free group on two generators C ∗ (F2 ) and B(`2 ) are not
completely locally reflexive [EH85, p. 124-125].
An operator space X is completely locally reflexive, if and only if one (and then
every) of the following conditions is satisfied for all finite dimensional operator spaces
L [EJR98, §1, 4.4, 5.8]:
∨ cb ∨ ∨
1. L ⊗ X ∗∗ = (L ⊗ X)∗∗ , where ⊗ denotes the injective operator space tensor
product,
53
The difference to the definition of local reflexivity is the fact that the ϕα are not only supposed to
be contractive, but even completely contractive.
cb
2. CB (L∗ , X ∗∗ ) = CB (L∗ , X)∗∗ ,
∧ cb ∨ ∧
3. L∗ ⊗ X ∗ = (L ⊗ X)∗ , where ⊗ denotes the projective operator space tensor
∨
product and ⊗ denotes the injective operator space tensor product,
cb
4. CN (X, L∗ ) = CI (X, L∗ ), where CN (·, ·) denotes the completely nuclear and
CI (·, ·) the completely integral mappings,
5. ι(ϕ) = ι(ϕ∗ ) for all ϕ ∈ CI(X, L∗ ).
In the conditions 1), 2) and 3) it suffices to check the usual isometry to prove the
complete isometry.
Examples
The following classes of operator spaces are completely locally reflexive [EJR98, 6.1,
6.2], [EH85, Prop. 5.4]:
1. reflexive operator spaces (e.g. finite dimensional operator spaces L),
2. nuclear C ∗ - algebras (e.g. K(H) or commutative C ∗ - algebras),54
3. preduals of von Neumann algebras, especially duals of C ∗ -algebras (e.g. T (H) =
K(H)∗ = B(H)∗ ).
8 Completely Bounded Multilinear Mappings

Going beyond the linear case, one can introduce the concept of complete bound-
edness for multilinear maps. The motivation mainly lies in the study of higher
dimensional Hochschild cohomology over C ∗ - and von Neumann algebras55 [Chris-
tensen/Effros/Sinclair ’87].
As in the linear case, the most important properties of the completely bounded
multilinear maps make their appearance56 in representation, extension and decom-
position theorems.
Let X1 , . . . , Xk , Y be operator spaces and Φ : X1 × · · · × Xk → Y a multilinear
mapping. We define [Christensen/Effros/Sinclair ’87, p. 281] a multilinear mapping
Φ(n) : Mn (X1 ) × · · · × Mn (Xk ) → Mn (Y )
 
n
j ,m
Φ(xl,j j1 ,j2
X
(x1 , . . . , xk ) 7→  1
1 , x2 , . . . , xkk−1 ) ,
j1 ,j2 ,...,jk−1 =1
54 ∗
a survey of the theory of nuclear C -algebras is e.g. to be found in [Mur90] and [Pat88].
There is a close connection to the long-standing still open derivation problem for C ∗ -algebras (or,
55
equivalently [Kirchberg ’96, Cor. 1], the similarity problem). One should note that in the framework
of operator spaces and completely bounded maps, considerable progress has been made in attacking
these problems. For instance, Christensen [Christensen ’82, Thm. 3.1] was able to show that the inner
derivations from a C ∗ -algebra into B(H) are precisely the completely bounded ones.
56
possibly footnote!
where n ∈ IN, the nth amplification of Φ.

Φ is called completely bounded if kΦkcb := supn kΦ(n) k < ∞. It is called com-
pletely contractive if kΦkcb ≤ 1.
Also compare the chapter Completely bounded bilinear maps.
Example: For bilinear forms on commutative C ∗ -algebras we have the following
result on automatic complete boundedness [Christensen/Sinclair ’87, Cor. 5.6]:
Let A be a commutative C ∗ -algebra. Then each continuous bilinear form Φ : A×A →
Cl is automatically completely bounded and
kΦk ≤ kΦkcb ≤ KG kΦk,
where KG denotes the Grothendieck constant. Furthermore, KG is the least such con-
stant.
One often studies completely bounded multilinear maps by considering the lineariza-
tion on the Haagerup tensor product, where the following relation holds [Paulsen/Smith
’87, Prop. 1.3; cf. also Sinclair/Smith ’95, Prop. 1.5.1]: If X1 , . . . , Xn are operator spaces
and H is a Hilbert space, then a multilinear mapping Φ : X1 × · · · × Xn → B(H) is
completely bounded if and only if its linearization ϕ is a completely bounded mapping
on X1 ⊗h · · · ⊗h Xn . In this case, kΦkcb = kϕkcb .
Also compare the chapter: Completely bounded bilinear mappings.
Representation theorem [Paulsen/Smith ’87, Thm. 3.2, cf. also Thm. 2.9; Sin-
clair/Smith, Thm. 1.5.4]:
Let A1 , . . . , Ak be C ∗ -algebras, X1 ⊂ A1 , . . . , Xk ⊂ Ak operator spaces and H a Hilbert
space. Let further be Φ : X1 × · · · × Xk → B(H) a completely contractive multi-
linear mapping. Then there exist Hilbert spaces Ki (i = 1, . . . , k), ∗ -representations
πi : Ai → B(Ki ) (i = 1, . . . , k), contractions Ti : Ki+1 → Ki (i = 1, . . . , k − 1) and two
isometries Vi : H → Ki (i = 1, k) such that
Φ(x1 , . . . , xk ) = V1∗ π1 (x1 )T1 π2 (x2 )T2 · · · Tk−1 πk (xk )Vk .
Following [Ylinen ’90, p. 296; cf. also Christensen/Effros/Sinclair ’87] it is possible to

eliminate the “bridging maps” Ti . One obtains the following simpler form for the rep-
resentation theorem:
Let A1 , . . . , Ak be C ∗ -algebras, X1 ⊂ A1 , . . . , Xk ⊂ Ak operator spaces and H a
Hilbert space. Let further Φ : X1 × · · · × Xk → B(H) be a completely contractive
multilinear mapping. Then there exist a Hilbert space K, ∗ -representations πi : Ai →
B(K) and two operators V1 , Vk ∈ B(H, K) such that
Φ(x1 , . . . , xk ) = V1∗ π1 (x1 )π2 (x2 ) · · · πk (xk )Vk .
From this result one can deduce the following:

Extension theorem [cf. Paulsen/Smith ’87, Cor. 3.3 and Sinclair/Smith ’95, Thm.
1.5.5]:
Let Xi ⊂ Yi (i = 1, . . . , k) be operator spaces and H a Hilbert space. Let further
Φ : X1 × . . . Xk → B(H) be a completely contractive multilinear mapping. Then there
exists a multilinear mapping Φe : Y1 × · · · × Yk → B(H) which extends Φ preserving the

cb-norm: kΦkcb = kΦk e cb .
Let A and B be C ∗ -algebras. For a k-linear mapping Φ : Ak → B we define
[Christensen/Sinclair ’87, pp. 154-155] another k-linear mapping Φ∗ : Ak → B by
Φ∗ (a1 , . . . , ak ) := Φ(a∗k , . . . , a∗2 , a∗1 )∗ ,
where a1 , . . . , ak ∈ A. A k-linear map Φ : Ak → B is called symmetric if Φ = Φ∗ . In

∗
this case, Φ(n) = Φ(n) (n ∈ IN).
A k-linear map Φ : Ak → B is called completely positive if
Φ(n) (A1 , . . . Ak ) ≥ 0
for all n ∈ IN and (A1 , . . . , Ak ) = (A∗k , . . . , A∗1 ) ∈ Mn (A)k , where A k+1 ≥ 0 for odd k.
2
Caution is advised: In the multilinear case complete positivity does not necessar-
ily imply complete boundedness! For an example (or more precisely a general method
of constructing such), cf. Christensen/Sinclair ’87, p. 155.
There is a multilinear version of the decomposition theorem for completely bounded
symmetric multilinear mappings:
Decomposition theorem [Christensen/Sinclair ’87, Cor. 4.3]:
Let A and B be C ∗ -algebras, where B is injective, and let further Φ : Ak → B be a
completely bounded symmetric k-linear mapping. Then there exist completely bounded,
completely positive k-linear mappings Φ+ , Φ− : Ak → B such that Φ = Φ+ − Φ− and
kΦkcb = kΦ+ + Φ− kcb .
9 Automatic Complete Boundedness

Completely bounded linear and multilinear mappings share strong structural proper-
ties. Thanks to the complete boundedness they have a very specific form (cf. the corre-
sponding representation theorems) whence they are much more accessible than arbitrary
bounded linear (or multilinear) mappings.57 Moreover, even in the multilinear case we
have at our disposal an extension theorem for completely bounded mappings – which
again is in striking contrast to the situation of arbitrary bounded (multi)linear mappings.
Because of the very nice structure theory of completely bounded linear and multi-
linear mappings, respectively, it is highly interesting to decide whether or not a given
bounded (multi)linear mapping actually is completely bounded. The most elegant way
to proceed is of course to check some simple conditions concerning the spaces involved
and/or some (purely) algebraic properties of the mapping which automatically imply
the complete boundedness of the latter.
In the following we shall collect various such criteria relying
(1) on the initial and/or target space

57
For example, the representation theorems provide a very useful tool in calculating cohomology
groups (cf. for example the monograph [SS95]).
(2) mainly on algebraic properties of the mapping (being, e.g., a ∗ -homomorphism of

C ∗ -algebras or a module homomorphism).
(1) “Criterion: spaces”
(1.1) The linear case

∗ Smith’s lemma: If X is a matricially normed space and Φ : X → Mn ,
where n ∈ IN, is a bounded linear operator, then we have kΦkcb = kΦ(n) k.
In particular, Φ is completely bounded if and only if Φ(n) is bounded
[Smi83, Thm. 2.10].
∗ Let X be an operator space, A a commutative C ∗ -algebra and Φ : X → A
a bounded linear operator. Then Φ is completely bounded and kΦkcb =
kΦk ([Loe75, Lemma 1], cf. also [Arv69, Prop. 1.2.2]).
In particular, every bounded linear functional is automatically completely
bounded with the same cb-norm.
∗ The following theorem shows that, roughly speaking, the above situa-
tions are the only ones where every bounded linear operator between
(arbitrary) C ∗ -algebras is automatically completely bounded. More pre-
cisely:
Let A and B be C ∗ -algebras. In order to have the complete boundedness
of very bounded linear operator from A to B it is necessary and sufficient
that either A is finite dimensional or B is a subalgebra of Mn ⊗ C(Ω) for
a compact Hausdorff space Ω [HT83, Cor. 4], [Smi83, Thm. 2.8], cf. also
[Smi83, p. 163].
(1.2) The bilinear case
∗ Let A be a commutative C ∗ -algebra. Then every bounded bilinear form
Φ : A × A → Cl is automatically completely bounded and
kΦk ≤ kΦkcb ≤ KG kΦk,
where KG denotes the Grothendieck constant; furthermore, KG is the

least possible constant [CS87, Cor. 5.6]
(2) “Criterion: mappings”
(2.1) The linear case

∗ Every ∗ -homomorphism between C ∗ -algebras is completely contractive.58
∗ Let A and B be C ∗ -subalgebras of B(H), where H is a Hilbert space. Let
further E ⊂ B(H) be an (A, B)-operator module and Φ : E → B(H) be a
bounded (A, B)-module homomorphism. Let A and B be quasi-cyclic59
, i.e., for every finite set of vectors ξi , ηj ∈ H, 1 ≤ i, j ≤ n, there exist
This is evident because π (n) : Mn (A) → Mn (B) is a ∗ -homomorphism for every n ∈ IN.
58
For example, cyclic C ∗ -algebras are quasi-cyclic. A von Neumann algebra M is quasi-cyclic if every
59
normal state on the commutant M 0 is a vector state [Smi91, Lemma 2.3].

ξ, η ∈ H such that ξi ∈ Aξ, ηj ∈ Bη, 1 ≤ i, j ≤ n. Then Φ is completely

bounded and kΦkcb = kΦk ([SS95, Thm. 1.6.1]; cf. also [Smi91, Thm.
2.1] and [Smi91, Remark 2.2], [Sat82, Satz 4.16]). – For special cases of
this result obtained earlier see [Haa80], [EK87, Thm. 2.5], [PPS89, Cor.
3.3] and [DP91, Thm. 2.4].
(2.2) The bilinear case
∗ Let A ⊂ B(H) be a C ∗ -algebra and E ⊂ B(H) an A-operator bimodule.
A bilinear mapping Φ : E × E → B(H) is called A-multimodular if
Φ(aeb, f c) = aΦ(e, bf )c
holds for all a, b, c ∈ A and e, f ∈ E.

We have a bilinear analogue of the above theorem:
Let A ⊂ B(H) be a quasi-cyclic C ∗ -algebra and E ⊂ B(H) be an A-
operator bimodule. Let further Φ : E × E → B(H) be a bilinear, A-
e : E⊗
multimodular mapping such that the corresponding linearization Φ
E → B(H) is bounded on E ⊗h E.
Then Φ is completely bounded and kΦkcb = kΦk e [SS95, Thm. 1.6.2].
10 Convexity
Since convex sets are important in the study of ordered or normed vector spaces, it
is only natural to ask for a non-commutative version of convexity that is more suited
for vector spaces of operators. Thus matrix convex sets were introduced, which play
the same role in operator space theory as ordinary convex sets in classical functional
analysis.
Some time earlier C ∗ -convex sets were defined for C ∗ -algebras. Both notions are
different but similar, so a section about C ∗ -convex sets is also included in this survey.
Some publications about non-commutative convexity are [EW97b], [WW99], [FZ98],
[Mor94], [Fuj94].
10.1 Matrix convexity

Let V be a complex vector space. A set K of matrices over V consists of subsets
Kn ⊂ Mn (V ) for all n ∈ IN. Subsets K1 ⊂ V can be considered as sets of matrices over
V by putting Kn = ∅ for n ≥ 2. More generally, if for some n ∈ N the corresponding
set Kn is not given a priori, put this Kn as the empty set. For sets of matrices, we have
the following notions of convexity:60
60 S
With the notation M (V ) = {Mn (V ) | n ∈ IN}, a set K of matrices over V is simply a subset of
M (V ). The sets Kn can be regained as Kn = K ∩ Mn (V ).
Then the definitions have the following form:
Let K ⊂ M (V ). K is called matrix convex if
K ⊕ K ⊂ K and α∗ Kα ⊂ K for all matrices α with α∗ α = 1l.

A set K of matrices over V is called matrix convex or a matrix convex set

[Wit84b] if for all x ∈ Kn and y ∈ Km
x ⊕ y ∈ Kn+m ,
and for all x ∈ Kn and α ∈ Mn,m with α∗ α = 1lm
α∗ xα ∈ Km .
A set K of matrices over V is called absolutely matrix convex [EW97a] if for all
x ∈ Kn and y ∈ Km
x ⊕ y ∈ Kn+m ,
and for all x ∈ Kn , α ∈ Mn,m and β ∈ Mm,n with kαk, kβk ≤ 1
αxβ ∈ Km .
A set K of matrices over V is called a matrix cone [Pow74] if for all x ∈ Kn and
y ∈ Km
x ⊕ y ∈ Kn+m ,
and for all x ∈ Kn and α ∈ Mn,m
α∗ xα ∈ Km .
A set K of matrices over V is matrix convex if and only if all matrix convex combi-
nations of elements of K are again in K. K is absolutely matrix convex if and only if all
absolutely matrix convex combinations of elements of K are again in K. Here, Pn a∗ma-
trix convex combination of x1 , . .P . , xn (xi ∈ Kki ) is a sum of the form i=1 αi xi αi
with matrices αi ∈ Mki ,j such that ni=1 αi∗ αi = P 1lj . An absolutely matrix convex
n
combination of x1 , . . . , P
xn is a sum of the form i=1 αi xi βi with matrices αi ∈ Mj,ki
and βi ∈ Mki ,j such that ni=1 αi αi∗ ≤ 1lj and ni=1 βi∗ βi ≤ 1lj .
P
If V is a topological vector space, topological terminology is to be considered at all
matrix levels: For instance, a set K of matrices over V is called closed if all Kn are
closed.
K is called absolutely matrix convex if
K ⊕ K ⊂ K and αKβ ⊂ K for all matrices α, β with kαk, kβk ≤ 1.
K is called a matrix cone if
K ⊕ K ⊂ K and α∗ Kα ⊂ K for all matrices α.

Examples
matrix convexity
1. The unit ball of an operator space X, given by the family Ball(X)n = {x ∈

Mn (X) | kxk ≤ 1} for all n ∈ IN,61 is absolutely matrix convex and closed.
2. The set of matrix states of a unital C ∗ -algebra A, given by the family CS(A)n =
{ϕ : A → Mn | ϕ completely positive and unital} for all n ∈ IN, is matrix convex
and weak-∗-compact.
3. If A is a C ∗ -algebra, the positive cones Mn (A)+ for all n ∈ IN form a closed matrix
cone.
10.1.1 Separation theorems

An important tool in the theory are the following separation theorems.
Let hV, W i be a (non degenerate) duality of complex vector spaces. Thus V and
W have weak topologies, and the matrix levels have the corresponding product topol-
ogy. For v = [vi,j ] ∈ Mn (V ) and w = [wk,l ] ∈ Mm (W ), hv, wi is defined by the joint
amplifications of the duality:
hv, wi = [hvi,j , wk,l i](i,k),(j,l) .
Note that the matrices, ordered by the cone of the positive semidefinite matrices,
are not totally ordered; 6≤ does not imply ≥.
Theorem:62 Let hV, W i be a duality of complex vector spaces, K a closed set of
matrices over V and v0 ∈ Mn (V ) \ Kn for some n.
a) [WW99, Thm. 1.6] If K is matrix convex, then there are w ∈ Mn (W ) and α ∈

(Mn )sa such that for all m ∈ IN and v ∈ Km
Rehv, wi ≤ 1lm ⊗ α, but Rehv0 , wi 6≤ 1ln ⊗ α.

61
Also Ball = {x ∈ M (X) | kxk ≤ 1}.
62
From this theorem one can easily get the following sharper version of the parts a), b) and d):
a) If K is matrix convex, then there are w ∈ Mn (W ), α ∈ (Mn )sa and ε > 0 such that for all m ∈ IN
and v ∈ Km
Rehv, wi ≤ 1lm ⊗ (α − ε1ln ), but Rehv0 , wi 6≤ 1ln ⊗ α.
b) If K is matrix convex and 0 ∈ K1 , then there are w ∈ Mn (W ) and ε > 0 such that for all m ∈ IN
and v ∈ Km
Rehv, wi ≤ (1 − ε)1lnm , but Rehv0 , wi 6≤ 1ln2 .
d) If K is absolutely matrix convex, then there are w ∈ Mn (W ) and ε > 0 such that for all m ∈ IN
and v ∈ Km
khv, wik ≤ 1 − ε, but khv0 , wik > 1.

b) [EW97b, Thm. 5.4] If K is matrix convex and 0 ∈ K1 , then there is w ∈ Mn (W )

such that for all m ∈ IN and v ∈ Km
Rehv, wi ≤ 1lnm , but Rehv0 , wi 6≤ 1ln2 .
c) [Bet97, p. 57] If K is a matrix cone, then there is w ∈ Mn (W ) sucht that for all
m ∈ IN and v ∈ Km
Rehv, wi ≤ 0, but Rehv0 , wi 6≤ 0.
d) [EW97a, Thm. 4.1] If K is absolutely matrix convex, then tere is w ∈ Mn (W )

such that for all m ∈ IN and v ∈ Km
khv, wik ≤ 1, but khv0 , wik > 1.
One can prove Ruan’s theorem using the separation theorem for absolutely matrix con-
vex sets, applied to the unit ball of a matricially normed space [].
If V is a complex involutive vector space, one can find selfadjoint separating func-
tionals:
Theorem: 63 Let hV, W i be a duality of complex involutive vector spaces, K a
closed set of selfadjoint matrices over V and v0 ∈ Mn (V ) \ Kn for some n.
b) If K is matrix convex and 0 ∈ K1 , then there is a w ∈ Mn (W )sa such that for all
m ∈ IN and v ∈ Km
hv, wi ≤ 1lnm , but Rehv0 , wi 6≤ 1ln2 .
10.1.2 Bipolar theorems

Let hV, W i be a duality of complex vector spaces and K a set of matrices over V .
The matrix polar of K is a set D of matrices over W , given by64
Dn = {w ∈ Mn (W ) | Rehv, wi ≤ 1lnm for all m ∈ IN, v ∈ Km }.
The absolute matrix polar of K is a set D of matrices over W , given by65
Dn = {w ∈ Mn (W ) | khv, wik ≤ 1 for all m ∈ IN, v ∈ Km }.
Polars of sets of matrices over W are defined analogously.

We have the bipolar theorems: Let hV, W i be a duality of complex vector spaces
and K a set of matrices over V .
63
This theorem can be obtained from the above separation theorem, part a) and b). Note that for
selfadjoint v, the mapping w 7→ hv, wi is selfadjoint.
64
D = {w ∈ M (W ) | Rehv, wi ≤ 1l for all v ∈ K}.
65
D = {w ∈ M (W ) | khv, wik ≤ 1 for all v ∈ K}.
a) [EW97b] K equals its matrix bipolar if and only if K is closed and matrix convex
and 0 ∈ K1 .
b) [EW97a] K equals its absolute matrix bipolar if and only if K is closed and abso-
lutely matrix convex.
The matrix bipolar of a set K of matrices over V is therefore the smallest closed and
matrix convex set which contains K and 0.
The absolute matrix bipolar of a set K of matrices over V is therefore the smallest
closed and absolutely matrix convex set which contains K.
So we get a characterization of the unit balls of MIN (E) and MAX (E) for a normed
space E.
10.2 Matrix extreme points

A part of convexity theory studies the possibility of representing all points of a convex
set as convex combinations of special points, the so called extreme points. The
well-known results about this are the Krein-Milman theorem and its sharpenings, the
Choquet representation theorems. The question arises whether there are analogous
results for non-commutative convexity. The sections Matrix extreme points and
C ∗ -extreme points give partial answers to this question.
Let V be a vector space. A matrix convex set of matrices over V is called matrix
convex set in V for short. Let A be a set of matrices over V . The matrix convex hull
of A is the smallest matrix convex set in V containing A. Its closure is the smallest
closed matrix convex set containing A because the closure of matrix convex sets is
matrix convex. Two elements x, y ∈ Mn (V ) are unitarily equivalent, if there is a
unitary u ∈ Mn such that x = u∗ yu. Let U (S) be the set of all elements, that are
unitarily equivalent to elements of S ⊂ Mn (V). x ∈ Mn (V ) is called reducible, if it is
unitarily equivalent to some block matrix y0 z0 ∈ Mn (V ). A matrix convex combination
Pk ∗
i=1 αi xi αi is called proper, if all αi are square matrices different from 0.
Let K be a matrixPconvex set in V . Then x ∈ Kn is a structural element66 of
Kn , if whenever x = ki=1 αi∗ xi αi is a proper matrix convex combination of xi ∈ Kn ,
then every xi is unitarily equivalent to x. The set of all structural elements of Kn is
denoted by str(Kn ). The set of structural elements of K is the set of matrices over V
consisting of str(Kn ) for all n ∈ IN.
Example: Let L be an operator system. The generalized state space of L is the

matrix convex set CS(L) in the dual L∗ , which consists of the matrix states
CS(L)n = {ψ | ψ : L → Mn completely positive and unital}.
The generalized state space is weak*-compact. It follows from [CE77, Lemma 2.2] that
the structural elements of CS(L)n are exactly those completely positive and unital
66
The term structural element is defined by Morenz [Mor94]. Another equivalent definition is given
in [WW99]. There the structural elements are called matrix extreme points.
mappings which are pure. To every compact and matrix convex set K there is an
operator system which has K as its generalized state space [WW99, Prop. 3.5].
Let V be a locally convex space and induce the product topology on Mn (V ). The
matrix convex Krein-Milman Theorem is: Let K be a compact matrix convex set
in V . Then K is equal to the closed matrix convex hull of the structural elements of K.
If V has finite dimension, then K is the matrix convex hull of its structural elements.
The converse result is: Let K be a compact matrix convex set in V . Let S be a
closed set of matrices, such that Sn ⊂ Kn and v ∗ Sl v ⊂ Sm for all partial isometries
v ∈ Mlm and for all n, m, l ∈ IN, l ≥ m. If the closed matrix convex hull of S equals K,
then all structural elements of K are in S.([WW99], [Fis96]).
It is possible to sharpen these results for more special matrix convex sets. A matrix
convex set K is called simple, if there are n ∈ IN and A ⊂ Mn (V ), such that K is equal
to the matrix convex hull of A. K is a simple matrix convex set, if and only if there is
n ∈ IN such that str(Km ) = ∅ for all m > n.
Suppose that K is a matrix convex set in V . Then x ∈ Km is a matrix extreme
point, if x ∈ str(Km ) and
/ ∪m<l 1l∗lm str(Kl )1llm .

x∈
Let mext(K) be the set of matrices consisting of all matrix extreme points of K.
Suppose that K is a simple compact matrix convex set in V . Then K is equal to
the closed matrix convex hull of mext(K). If V has finite dimension, then the closure
is not needed, that means K is the matrix convex hull of mext(K). In this case the
following result also holds: Let S be a set of matrices over V not containing reducible
elements such that the matrix convex hull of S equals K, then mext(K)m ⊂ U (Sm ) for
all m ∈ IN ([Mor94], [Fis96]).
If K is compact and not simple, mext(K) may be empty. As an example take the
generalized state space CS(A) of a C ∗ -algebra A. Its matrix extreme points are exactly
the irreducible finite dimensional representations of A. These need not exist in general.
10.3 C ∗ -convexity
Let A be a unital C ∗ -algebra. A subset K ⊂ A is a C ∗ -convex set, if
n
X
a∗i xi ai ∈ K
i=1
Pn ∗
Pn ∗
for all xi ∈ K and ai ∈ A such that i=1 ai ai = 1l. The sum i=1 ai xi ai is called a
C ∗ -convex combination.
A subset K ⊂ A is a C ∗ -absolutely convex set, if

n
X
a∗i xi bi ∈ K
i=1
Pn ∗
Pn ∗
for all xi ∈ K and ai , bi ∈ A such that i=1 ai ai , i=1 bi bi ≤ 1l.
In particular C ∗ -convex sets are convex and C ∗ -absolutely convex sets are absolutely
convex.
Example: Let H be a Hilbert space and T ∈ B(H). The n-th matrix range of T is
the set
W (T )n := {ϕ(T )|ϕ : B(H) → Mn completely positive and unital}.
W (T )n is a compact and C ∗ -convex subset of Mn and to every compact and C ∗ -convex

subset K ⊂ Mn there exist a separable Hilbert space H and S ∈ B(H) such that
K = W (S)n ([LP81, Prop. 31]).
The sets
Ball(B(H)) = {x ∈ B(H)|kxk ≤ 1} and P = {x ∈ B(H)|0 ≤ x ≤ 1l}
are C ∗ -convex und wot-compact subsets of B(H). Ball(B(H)) is also C ∗ -absolutely

convex.
Loebl and Paulsen introduced C ∗ -convex sets in [LP81]. At the beginning of the
nineties Farenick and Morenz studied C ∗ -convex subsets of Mn ([Far92],[FM93]). Even-
tually Morenz succeeded in proving an analogue of the Krein-Milman theorem for a
compact C ∗ -convex subset of Mn ([Mor94]). At the end of the nineties Magajna gen-
eralized the notion of C ∗ -convex sets to the setting of operator modules and proved
some separation theorems [Mag00, Th. 1.1] and also an analogue of the Krein-Milman
theorem [Mag98, Th. 1.1].
10.3.1 Separation theorems

Suppose that A, B ⊂ B(H) are unital C ∗ -algebras and Y ⊂ B(H) is a (A, B)-bimodul.
Then K ⊂ Y is (A, B)-absolutely convex, if
n
X
a∗i xi bi ∈ K
i=1
Pn ∗
Pn ∗
for all xi ∈ K and ai ∈ A, bi ∈ B such that i=1 ai ai , i=1 bi bi ≤ 1l. Let Y be a
A-bimodul. Then K is A-convex, if
n
X
a∗i xi ai ∈ K
i=1
for all xi ∈ K and ai ∈ A such that ni=1 a∗i ai = 1l. In the case Y = A this definition is
P
equivalent to the definition of C ∗ -convex sets.
There are following separation theorems: Let A, B ⊂ B(H) be unital C ∗ -algebras

and Y ⊂ B(H) a A, B-bimodul. Let K ⊂ Y be norm closed and y0 ∈ Y \ K.
1) If A = B, 0 ∈ K and K is A-convex, then there is a Hilbert space Hπ , a cyclic

representation π : A → B(Hπ ) and a completely bounded A-bimodul-homomorphism,
such that for all y ∈ K
Re φ(y) ≤ 1l, but Re φ(y0 ) 6≤ 1l.
2)If K is (A, B)-absolutely convex, then there is a Hilbert space Hπ , representations
π : A → B(Hπ ) and σ : B → B(Hπ ) and a completely bounded (A, B)-bimodul-
homomorphism φ : Y → B(Hπ ), such that for all y ∈ K
kφ(y)k ≤ 1, but kφ(y0 )k > 1.
10.4 C ∗ -extreme points

Let A be an unital C ∗ -algebra and Y ⊂ A. The C ∗ -convex hull of Y is the smallest
C ∗ -convex set that contains Y .
Suppose
Pn K ⊂ A is C ∗ -convex. Then x ∈ K is a C ∗ -extreme point, if whenever
x = i=1 ai xi ai is a C ∗ -convex combination of xi ∈ K with invertible ai ∈ A, then
∗
there are unitaries ui ∈ A such that x = u∗i xi ui for i = 1, . . . , n.
Suppose now A = Mn and let K ⊂ Mn be compact and C ∗ -convex. Let K̃ be the

matrix convex hull of K. Then K̃ is a simple compact and matrix convex set in C l , such
that K̃n = K ([Fis96]). Thus it is possible to conceive a C ∗ -convex subset of Mn as a
matrix convex set in Cl . Now the matrix convex Krein-Milman theorem can be used.
Moreover, it follows from the work of Farenick and Morenz that the structural elements
of K̃n are exactly the not reducible C ∗ -extreme points of K. So following theorem
holds: Let K ⊂ Mn be compact and C ∗ -convex, then K is equal to the C ∗ -convex hull
of its C ∗ -extreme points.
In order to get a somewhat more general result, the definition of the extreme points
can be changed. Suppose that R is a hyperfinite factor andPthat K ⊂ R is C ∗ -convex.
n ∗
Then x ∈ K is a R-extreme point, if whenever x = i=1 ai xi ai is a C -convex
combination of xi ∈ K such that all ai ∈ A are positive and invertible, then it follows
that x = xi and ai x = xai forr i = 1, . . . , n.67
With this definition following theorem hold: Let K ⊂ R be C ∗ -convex and weak*
compact. Then K is equal to the weak* closure of the C ∗ -convex hull of its R-extreme
points.
67
If R = Mn the R-extreme points are exactly the C ∗ -extreme points. In general every R-extreme
point is also C ∗ -extreme, but not vice versa.
11 Mapping Spaces
Let E, F be Banach spaces. We consider a linear subspace A(E, F ) of the space B(E, F )
of the continuous operators between E and F which contains all finite rank maps and is
a Banach space with respect to a given norm. Furthermore, it is usually required that
A(E, F ) be defined for all pairs of Banach spaces E and F . Such a space is called –
according to A. Grothendieck – a mapping space .
Analogously, we call an operator space A(X, Y ) which is a linear subspace of
CB (X, Y ) a CB -mapping space. Note that generally the algebraic identification
of Mn (A(X, Y )) with A(X, Mn (Y )) fails to be isometric and that the norms on
A(X, Mn (Y )) do not generate an operator space structure for A(X, Y ).
There is a close relationship between mapping spaces and tensor products: The space
F (X, Y ) of all finite rank maps between X and Y and the algebraic tensor product of
X ∗ with Y are isomorphic:
X ∗ ⊗alg Y ∼
= F (X, Y ).
This identification enables us to transfer norms from one space to the other one. To this
end, we consider the extension of the mapping X ∗ ⊗ Y → F (X, Y ) to the completion
with respect to an operator space tensor norm X ∗ ⊗Ye :
Φ : X ∗ ⊗Y
e → CB (X, Y ).
Φ is in general neither injective nor surjective. As a CB-mapping space one obtains
Im(Φ) ⊂ CB (X, Y )
with the operator space norm of
(X ∗ ⊗Y
˜ )/Ker(Φ).
We consider now assignments that assign a mapping space A(·, ·) with operator space
norm α(·) to every pair of operator spaces. In the Banach space theory A. Pietsch inten-
sified the notion of mapping spaces to that of the operator ideals [Pie78]. Analogously,
we consider operator ideals which are mapping spaces with the CB -ideal property
[ER94], i.e , the composition
CB (X1 , X2 ) × A(X2 , Y2 ) × CB (Y2 , Y1 ) → A(X1 , Y1 )

(Ψ1 , Φ, Ψ2 ) 7→ Ψ2 ◦ Φ ◦ Ψ1
is for all operator spaces X1 , X2 , Y1 , Y2 defined and jointly completely contractive.

A CB -ideal is called local [EJR98], if its norm satisfies:
α(ϕ) = sup{α(ϕ|L ) : L ⊂ X, dim L < ∞}.

11.1 Completely nuclear mappings

Let X and Y be operator spaces. The completely nuclear mappings from X to Y
[ER94, §2], [EJR98, §3] are defined by the projective operator space tensor norm. We
consider the extension of the canonical identity X ∗ ⊗ Y = F (X, Y ):
∧ ∨
Φ : X ∗ ⊗ Y → X ∗ ⊗ Y ⊂ CB(X, Y ).
∨ ∧
⊗ is the injective tensor product and ⊗ the projective tensor product.
A mapping in the range of Φ is called completely nuclear. One denotes by
∧
CN (X, Y ) := (X ∗ ⊗ Y )/Ker(Φ)
the space of the completely nuclear mappings and endows it with the quotient operator
space structure. The operator space norm is denoted by ν(·). Mn (CN (X, Y )) and
CN (X, Mn (Y )) are in general not isometric.
Nuclear 68 mappings are completely nuclear. [ER94, 3.10]
In general, the projective tensor norm does not respect complete isometries.
Hence,even for subspaces Y0 ⊂ Y the canonical embedding CN (X, Y0 ) → CN (X, Y )
is generally only completely contractive and not isometric. Since the projective tensor
norm respects quotient mappings, every nuclear map ϕ : X0 → Y on a subspace X0 ⊂ X
with ν(ϕ) < 1 has an extension ϕ̃ to the whole of X satisfying ν(ϕ̃) < 1.
The completely nuclear mappings enjoy the CB -ideal property. Furthermore, the
adjoint ϕ∗ is completely nuclear if ϕ is, and the inequality: ν(ϕ∗ ) ≤ ν(ϕ) [EJR98,
Lemma 3.2] holds.
A mapping ϕ is completely nuclear, if and only if there is a factorization of the form
M (a,b)
B(`2 ) → T (`2 )
↑r ↓s
ϕ
X → Y
Here a, b are Hilbert-Schmidt operators defining the mapping M (a, b) : x 7→ axb. For
the completely nuclear norm we have: ν(ϕ) = 1 precisely if for all > 0 there exists a
factorization with krkcb kak2 kbk2 kskcb ≤ 1 + 69 [ER94, Thm. 2.1].
68
The completely nuclear mappings owe their definition to the one of the nuclear mappings of the
Banach space theory. There, one considers a corresponding mapping ΦB : E ∗ ⊗γ F → B(E, F ) for two
Banach spaces E and F .
69
In the Banach space theory one has an analogous statement: A mapping ϕ is nuclear, if and only if
there’s a diagram
d
`∞ → `1
r↑ ↓s
ϕ
E → F,
where d is a diagonal operator, i.e. , there is a (di ) ∈ `1 , such that d((ai )) = (di ·ai ) for all (ai ) ∈ `∞ . The
nuclear norm is then computed as: νB (ϕ) = inf krkkdk`1 ksk, where the infimum runs over all possible
factorizations.
The completely nuclear mappings are not local.
11.2 Completely integral mappings

Completely integral mappings are defined by the aid of completely nuclear mappings.
A mapping ϕ : X → Y is said to be completely integral, if there are a constant c > 0
and a net of finite rank maps ϕα ∈ CN (X, Y ) with ν(ϕα ) ≤ c converging to ϕ in the
point norm topology 70 .
The set of all these mappings forms the space CI (X, Y ) of the completely integral
mappings.
The infimum of all those constants c satisfying the above condition is actually at-
tained and denoted by ι(ϕ). ι(·) is a norm turning CI (X, Y ) into a Banach space. The
unit sphere of CI (X, Y ) is merely the point norm closure of the unit sphere of CN (X, Y ).
One obtains the canonical operator space structure by defining the unit sphere of
Mn (CI (X, Y )) as the point norm closure of unit sphere of Mn (CN (X, Y )).
By definition we have ι(ϕ) ≤ ν(ϕ); for finite dimensional X we have moreover
[EJR98, Lemma 4.1]
cb
CI (X, Y ) = CN (X, Y ).
Integral 71 mappings are completely integral [ER94, 3.10].
The canonical embedding
∨
CI (X, Y ) ,→ (X ⊗ Y ∗ )∗
is a complete isometry [EJR98, Cor. 4.3]. One has moreover that [EJR98, Cor. 4.6] ϕ is
completely integral, if and only if there is a factorization of the form:
M (ω)
B(H) → B(K)∗
↑r ↓s
ϕ
X → Y ,→ Y ∗∗
with weak∗ -continuous s . The mapping M (ω) : B(H) → B(K)∗ is for two elements
a ∈ B(H), b ∈ B(K) given by (M (ω)(a))(b) = ω(a ⊗ b). We have for the norm ι(ϕ) = 1,
if there is a factorization with krkcb kωkkskcb = 1 (note that generally kM (ω)kcb 6= kωk).
The completely integral mappings enjoy also the CB -ideal property. Contrasting
the situation of completely nuclear mappings they are local. In general one only has
ι(ϕ) ≤ ι(ϕ∗ ) [EJR98].
70
ϕα → ϕ in the point norm topology, if kϕα (x) − ϕ(x)k → 0 for all x ∈ X.
71
The unit ball of the integral mappings of the Banach space theory is just the point norm closure
of the unit ball of the nuclear mappings. One should note that the formulas ιB (ϕ) = ιB (ϕ∗ ) and
IB (E, F ∗ ) = (E ⊗λ F )∗ have no counterparts for completely integral mappings.
12 Appendix
12.1 Tensor products
12.1.1 Tensor products of operator matrices
As usual we define the algebraic tensor product of operator matrices x = [xij ] ∈ Mp (X),
y = [ykl ] ∈ Mq (Y ) by setting
x ⊗ y := [xij ⊗ [ykl ]k,l ]i,j ∈ Mp (X ⊗ Mq (Y )).
Here we have used the definition Mp (X) = Mp ⊗ X and the associative law
Mp (X) ⊗ Mq (Y ) = Mp ⊗ (X ⊗ Mq (Y )) = Mp (X ⊗ Mq (Y )). (5)
In view of the next identification one should note that the shuffle-map is an algebraic
isomorphism:
X ⊗ (Mq ⊗ Y ) → Mq ⊗ (X ⊗ Y ), (6)
x ⊗ (β ⊗ y) 7→ β ⊗ (x ⊗ y),
for β ∈ Mq , x ∈ X, y ∈ Y . The shuffle-isomorphism at hand we obtain the identification:
x ⊗ y = [xij ⊗ [ykl ]k,l ]i,j = [[xij ⊗ ykl ]k,l ]i,j ∈ Mp (Mq (X ⊗ Y )).
Finally we use the usual72 identification Mp (Mq ) = Mpq
[[xij ⊗ ykl ]k,l ]i,j = [xij ⊗ ykl ](i,k),(j,l)
to obtain
Mp (X) ⊗ Mq (Y ) = Mpq (X ⊗ Y ). (7)
We call this algebraic isomorphism the shuffle-isomorphism.

One should note that for operator space tensor products the algebraic identifications
(5) and (6) are only complete contractions:
lMp (X) ⊗α Y → lMp (X ⊗α Y ), (8)

X ⊗α lMq (Y ) → lMq (X ⊗α Y ). (9)
In general these are not isometries even for p = 1 resp. q = 1.

For an operator space tensor product the shuffle-map
lMp (X) ⊗α lMq (Y ) → lMpq (X ⊗α Y ) (10)

72
In the matrix so obtained (i, k) are the row indices and (j, l) are the column indices, where i, j =
1, . . . , p and k, l = 1, . . . , q. The indices (i, k) resp. (j, l) are ordered lexicographically.
in general is only completely contractive.

In the case of the injective operator space tensor product this is of course a complete
isometry.
More generally, one considers the shuffle–map for rectangular matrices73 :
lMm,n (X) ⊗α lMp,q (Y ) → lMmp,nq (X ⊗α Y ).
Another example is provided by the Blecher-Paulsen equation.
12.1.2 Joint amplification of a duality

The matrix dualitywhich is fundamental in the duality theory of operator spaces, is a
special case of the joint amplification of a bilinear mapping.
The joint amplification of a duality hX, X ∗ i of vector spaces is defined by
hx, ϕip×q = h[xij ], [ϕκλ ]ip×q := [hxij , ϕκλ i] ∈ Mp (Mq ) = Mpq
for x = [xij ] ∈ Mp (X), ϕ = [ϕκλ ] ∈ Mq (X ∗ ).

Interpreting ϕ as a mapping ϕ : X → Mq we have
ϕ(p) (x) = hx, ϕip×q .
Associated to the duality of tensor products74 hX ⊗ Y, X ∗ ⊗ Y ∗ i is the joint

amplification
hMp (X ⊗ Y ), Mq (X ∗ ⊗ Y ∗ )i.
Especially, the equation
hx ⊗ y, ϕ ⊗ ψi = h[xij ] ⊗ [ykl ], [ϕκλ ] ⊗ [ψµν ]i

:= [hxij , ϕκλ ihykl , ψµν i](ikκµ),(jlλν) ∈ Mp1 (Mp2 (Mq1 (Mq2 ))) = Mp1 p2 q1 q2
obtains, where x = [xij ] ∈ Mp1 (X), y = [ykl ] ∈ Mp2 (Y ), ϕ = [ϕκλ ] ∈ Mq1 (X ∗ ),

ψ = [ψµν ] ∈ Mq2 (Y ∗ ).
73
The shuffle-map
(U ⊗ X) ⊗ (V ⊗ Y ) → (U ⊗ V ) ⊗ (X ⊗ Y ),
(u ⊗ x) ⊗ (v ⊗ y) 7→ (u ⊗ v) ⊗ (x ⊗ y),
U ,V ,X,Y operator spaces, has been studied for various combinations of operator space tensor products
[EKR93, Chap. 4].
74
The duality hX ⊗ Y, X ∗ ⊗ Y ∗ i is defined by hx ⊗ y, ϕ ⊗ ψi := hx, ϕihy, ψi for x ∈ X, y ∈ Y , ϕ ∈ X ∗ ,
ψ ∈ Y ∗.
12.1.3 Tensor matrix multiplication

The definition of the completely bounded bilinear maps as well as the
Haagerup tensor product relies on the tensor matrix multiplication [Eff87]
hP i
l
x y = [xij ] [yjk ] := x
j=1 ij ⊗ y jk ∈ Mn (X ⊗ Y )
of operator matrices x = [xij ] ∈ Mn,l (X), y = [yjk ] ∈ Ml,n (Y ).

The amplification of the bilinear mapping ⊗ : X × Y → X ⊗ Y is given by
⊗(n,l) = : Mn,l (X) × Ml,n (Y ) → Mn (X ⊗ Y ).
For scalar matrices α, γ ∈ Mn , β ∈ Ml we have
(αxβ) (yγ) = α(x (βy))γ.
We use the short hand notation αxβ yγ.

For linear maps
Φ = [Φij ] : x → Mn,l (V ),
Ψ = [Φjk ] : x → Ml,n (W )
we denote by Φ Ψ the mapping

 
Xl
ΦΨ= Φij ⊗ Ψjk  : X ⊗ Y → Mn (V ⊗ W ),
j=1
 
Xl
Φ Ψ : x ⊗ y 7→  Φij (x) ⊗ Ψjk (y) .
j=1
We then have
(Φ Ψ)(p) (x y) = (Φ(p,q) (x)) (Ψ(q,p) (y))
for x ∈ Mp,q (X), y ∈ Mq,p (Y ).

Let ⊗α be an operator space tensor product. We define the tensor matrix mul-
tiplication α of operator matrices S = [Si,j ] ∈ Mn,l (CB (X1 , X2 )), T = [Tk,l ] ∈
Ml,n (CB (Y1 , Y2 )) of completely bounded maps by setting
 
X l
S α T =  Sij ⊗α Tjk  ∈ Mn (CB (X1 ⊗α Y1 , X2 ⊗α Y2 )).
j=1
12.2 Interpolation
Intersection and sum
Let X and Y be operator spaces such that M1 (X) and M1 (Y ) are embedded in a
Hausdorff topological vector space. Mn (X ∩ Y ) is given a norm via Mn (X ∩ Y ) :=
Mn (X) ∩ Mn (Y ). So we have:
n o
k[xij ]kMn (X∩Y ) = max k[xij ]kMn (X) , k[xij ]kMn (Y ) .
The operator space X ∩ Y is called the intersection of X and Y .

For operator spaces75 X and Y , by embedding X ⊕ Y in (X ∗ ⊕∞ Y ∗ )∗ we obtain an
operator space structure X ⊕1 Y . We write ♦ := {(x, −x)} ⊂ X ⊕1 Y . The quotient
operator space (X ⊕1 Y ) /♦ is called the sum of X and Y and is denoted by X + Y .
We have
k[xij ]kMn (X+Y ) = inf k[(xijX , xijY )]kMn (X⊕1 Y ) .
[xij ]=[xij ]X +[xij ]Y
Interpolation
Let E0 , E1 be Banach spaces continuously embedded in a Hausdorff topological vector
space. The pair (E0 , E1 ) is called a compatiple couple in the sense of interpolation theory
[BL76]. Then we can define the interpolation space Eθ := (E0 , E1 )θ for 0 < θ < 1.
Pisier introduced the analogous construction for operator spaces [Pis96, §2]: Let Xi
(i = 0, 1) be operator spaces continuously embedded in a Hausdorff topological vector
space V . Then we have specific norms on Mn (Xi ) and continuous linear inclusions
Mn (Xi ) ,→ Mn (V ) for all n ∈ IN.76 The interpolated operator space Xθ is defined
via Mn (Xθ ) := (Mn (X0 ), Mn (X1 ))θ .
Let X be an operator space, H a Hilbert space and V : H → X a bounded linear and
injective mapping with dense range such that the mapping77 V V ∗ : X ∗ → X also is
bounded, linear and injective with dense range. Then we have completely isometrically
[Pis96, Cor. 2.4]:
cb
(X ∗ , X) 1 = OH H .
2
Examples
cb cb
1. (RH , CH ) 1 = OH H = (MIN H , MAX H ) 1
2 2
75
Let E, F be Banach spaces. Then we have their 1-direct sum E ⊕1 F with the norm
k(xE , xF )k = kxE k + kxF k
and their sum E + F with the quotient norm

kxkE+F = inf kxE kE + kxF kF .
x=xE +xF
2
76
We identify Mn (V ) with V n .
77
As usual we identify H with its dual.
cb cb
2. (CH ⊗h RH , RH ⊗h CH ) 1 = OH H ⊗h OH H = OH H⊗H
2
In this manner one also obtains operator space structures on the Schatten ideals Sp =
(S∞ , S1 ) 1 for 1 ≤ p ≤ ∞.
p
13 Symbols
Sets
IN the set of natural numbers
ZZ the set of integers
Ql the set of rational numbers
lR the set of real numbers
Cl the set of complex numbers
Banach spaces
E, F Banach spaces
H,
LK Hilbert spaces
2 Hilbert
Ln space direct sum
Hn i=1 H
`n2 the n-dimensional Hilbert space
Algebras
B(H) the algebra of bounded linear operators on H
A, B C ∗ -algebras
Aop the opposite algebra of an algebra A
1lA the identity in A
M, N von Neumann algebras
Ideals
Sp the Schatten-p-classes
K(·, ·) the ideal of compact operators
K(·, ·) the closed linear span of the elementary operators; ideal in B ad (·, ·)
N (·, ·) the ideal of nuclear operators
HS (·, ·) the ideal of Hilbert-Schmidt operators (S2 )
Π2 (·, ·) the ideal of absolutely 2-summing operators
π2 (·) the absolutely 2-summing norm
Mapping spaces
Γ2 (·, ·) operators factorizing through the column Hilbert space
γ2 (·) the norm on Γ2
CN (·, ·) completely nuclear operators
ν(·) the norm on CN
CI (·, ·) completely integral operators
ι(·) the norm on CI
Operator spaces
X, Y operator spaces
B(H) aqlgebra of bounded linear operators on H
Mn (X) Mn ⊗ X matrices with entries from X (algebraically)
M1 (X) first level of the operator space X
CB (X, Y ) the operator space of completely bounded mappings
CB (X, Y )A the operator space of completely bounded right A-module homomor-
phisms
CB (X × Y ; Z) the operator space of completely bounded bilinear mappings
JCB (X × Y ; Z) the operator space of jointly completely bounded bilinear mappings
k · kjcb norm of a jointly completely bounded bilinear mapping
X0 , Y0 operator subspace of the corresponding operator spaces
X∗ dual of the operator space X
Special operator spaces

MIN (E) the minimal operator space on E
MAX (E) the maximal operator space on E
MIN H the minimal operator space on H
MAX H the maximal operator space on H
RH the row Hilbert space
Rn R`n2
CH the column Hilbert space
Cn C`n2
OH H the operator Hilbert space
Norms
k · kcb completely bounded norm
k · krow row norm
k · kcol column norm
k · kn norm on Mn (X)
k · km,n norm on Mm,n (X)
Matrices
Mn,m (X) n × m-matrices over X
Mn (X) Mn,n (X)
Mn,m Mn,m (Cl )
Mn Mn,n
lMn,m (X) the operator space of n × m-matrices over X
lMn (X) lMn,n (X)
lMn,m lMn,m (Cl )
lMn lMn,n
Cn (X) lMn,1 (X), the columns of an operator space
Rn (X) lM1,n (X), the rows of an operator space
Tensor products
tensor matrix product
⊗ algebraic tensor product
⊗A algebraic module tensor product
⊗
e completion of the algebraic tensor product
⊗α operator space tensor product
⊗α∗ dual operator space tensor product
⊗h the Haagerup tensor product
⊗hA the module Haagerup tensor product
∨
⊗ the injective tensor product
∧
⊗ the projective tensor product
⊗λ the injective Banach space tensor product
⊗γ the projective Banach space tensor product
⊗ext the external tensor product
⊗Θ the internal tensor product
S ⊗α T α-tensor product of the completely bounded operators S, T
Tensor norms
k · kα α-operator space tensor norm
k · kα,n α-operator space tensor norm on the nth level
k · kα∗ dual operator space tensor norm of k · kα
k · k∨ injective operator space tensor norm
k · k∧ projective operator space tensor norm
k · kh Haagerup operator space tensor norm
k · kλ injective Banach space tensor norm
k · kγ projective Banach space tensor norm
Operator modules
A1 , A2 C ∗ -algebras
M, N von Neumann algebras
Aut(M ) the set of ∗ -automorphisms of M
CB (A1 ,A2 ) (X, Y ) the operator space of completely bounded left A1 - right A2 -module
homomorphisms between X and Y
CB σ(A1 ,A2 ) (X, Y ) the operator space of normal completely bounded left A1 - right A2 -
module homomorphisms between X and Y
CB s(A1 ,A2 ) (X, Y ) the operator space of singular completely bounded left A1 - right A2 -
module homomorphisms between X and Y
Hilbert-C ∗ -modules
X conjugate Hilbert-C ∗ -module
B(·, ·)A the space of bounded right A-module homomorphisms
h·, ·iA inner product of a right Hilbert-A-module
A h·, ·i inner product of a left Hilbert-A-module
h·, ·i inner product of a Hilbert space
B ad (X, Y ) the operator space of adjointable A-module homomorphisms between
Hilbert-C ∗ -modules X and Y
Mappings
1lB(H) the identity of B(H)
1ln the identity of Mn
π representations
Φ(n) nth amplification of a linear mapping Φ
Φ(n,l) (n, l)th amplification of a bilinear mapping Φ
Φ(n) Φ(n,n) , nth amplification of a bilinear mapping Φ
Φ(n) nth amplification of a multilinear mapping Φ
Φ(p×q) joint amplification of a bilinear mapping Φ
Φ̃ linearization of a bilinear mapping Φ
Φσ the normal (=w∗ -w∗ -continuous) part of a mapping Φ between dual
spaces
Φs the singular part of a mapping Φ between dual spaces
Θ faithful non-degenerate ∗-representation of a C ∗ -algebra
Isomorphisms
cb
= completely isometrically isomorphic
cb
' completely isomorphic
' isomorphic
Miscellanea
Ball
Ln unit ball
Li=1 Xi nth direct sum
n∈IN Xn countable direct sum
T∗ the adjoint of the operator T
References
[Arv69] William B. Arveson. Subalgebras of C ∗ -algebras. Acta Math., 123:141–224,
1969. 3, 47
[Bet97] Benedikt Betz. Matrixkonvexe Analysis. Diplomarbeit, Universität der Saar-

landes, 1997. 51
[BL76] Jöran Bergh and Jörgen Löfström. Interpolation spaces. An introduc-

tion. Springer-Verlag, Berlin, 1976. Grundlehren der Mathematischen Wis-
senschaften, No. 223. 62
[Ble92a] David P. Blecher. The standard dual of an operator space. Pacific J. Math.,
153(1):15–30, 1992. 8, 8, 8, 9, 10, 15
[Ble92b] David P. Blecher. Tensor products of operator spaces. II. Canad. J. Math.,
44(1):75–90, 1992. 8, 9, 15, 15, 17, 17, 19, 19, 23, 32, 37, 43
[Ble95] David P. Blecher. A completely bounded characterization of operator algebras.

Math. Ann., 303(2):227–239, 1995. 16, 18, 24, 24, 24
[BLM95] David P. Blecher and Christian Le Merdy. On quotients of function algebras

and operator algebra structures on lp . J. Operator Theory, 34(2):315–346,
1995. 24, 25, 25, 25, 25, 26, 26, 26, 26
[BMP] David P. Blecher, Paul S. Muhly, and Vern I. Paulsen. Categories of operator
modules (Morita equivalence and projective modules). Mem. Amer. Math.
Soc. 42, 42, 43
[BP91] David P. Blecher and Vern I. Paulsen. Tensor products of operator spaces. J.
Funct. Anal., 99(2):262–292, 1991. 6, 8, 26, 27, 29, 30, 30, 30, 30, 30, 30, 31,
32, 32, 32, 32, 34, 34, 35, 35, 36, 38, 40, 40
[BRS90] David P. Blecher, Zhong-jin Ruan, and Allan M. Sinclair. A characterization

of operator algebras. J. Funct. Anal., 89(1):188–201, 1990. 23, 24, 24, 24, 24,
24, 24
[CE77] Man Duen Choi and Edward G. Effros. Injectivity and operator spaces. J.
Functional Analysis, 24(2):156–209, 1977. 3, 14, 15, 52
[CES87] Erik Christensen, Edward G. Effros, and Allan Sinclair. Completely bounded
multilinear maps and C ∗ -algebraic cohomology. Invent. Math., 90(2):279–296,
1987. 24
[Cho74] Man Duen Choi. A schwarz inequality for positive linear maps on C ∗ -algebras.
Illinois J. Math., 18:565–574, 1974. 14
[CS87] Erik Christensen and Allan M. Sinclair. Representations of completely

bounded multilinear operators. J. Funct. Anal., 72(1):151–181, 1987. 38,
47
[CS89] Erik Christensen and Allan M. Sinclair. A survey of completely bounded

operators. Bull. London Math. Soc., 21(5):417–448, 1989. 3, 42
[DP91] K. R. Davidson and S. C. Power. Isometric automorphisms and homol-

ogy for nonselfadjoint operator algebras. Quart. J. Math. Oxford Ser. (2),
42(167):271–292, 1991. 48
[Eff87] Edward G. Effros. Advances in quantized functional analysis. In Proceedings

of the International Congress of Mathematicians, Vol. 1, 2 (Berkeley, Calif.,
1986), pages 906–916, Providence, RI, 1987. Amer. Math. Soc. 3, 39, 61
[EH85] Edward G. Effros and Uffe Haagerup. Lifting problems and local reflexivity
for C ∗ -algebras. Duke Math. J., 52(1):103–128, 1985. 43, 44
[EJR98] Edward G. Effros, Marius Junge, and Zhong-Jin Ruan. Integral mappings and
the principle of local reflexivity for non-commutative l1 -spaces. Math.Ann.,
??(?):??–??, 1998. 43, 43, 44, 56, 57, 57, 58, 58, 58, 58
[EK87] Edward G. Effros and Akitaka Kishimoto. Module maps and Hochschild–
Johnson cohomology. Indiana Univ. Math. J., 36(2):257–276, 1987. 33, 41,
48
[EKR93] Edward G. Effros, Jon Kraus, and Zhong-Jin Ruan. On two quantized tensor
products. In Operator algebras, mathematical physics, and low-dimensional
topology (Istanbul, 1991), pages 125–145. A K Peters, Wellesley, MA, 1993.
60
[ER88] Edward G. Effros and Zhong-jin Ruan. Representations of operator bimodules

and their applications. J. Operator Theory, 19(1):137–158, 1988. 20, 21
[ER90a] Edward G. Effros and Zhong-jin Ruan. On approximation properties for op-
erator spaces. Internat. J. Math., 1(2):163–187, 1990. 26, 30, 32, 33, 33
[ER90b] Edward G. Effros and Zhong-jin Ruan. On nonselfadjoint operator algebras.

Proc. Amer. Math. Soc., 110(4):915–922, 1990. 24
[ER91] Edward G. Effros and Zhong-jin Ruan. Self-duality for the Haagerup tensor
product and Hilbert space factorizations. J. Funct. Anal., 100(2):257–284,
1991. 15, 16, 17, 17, 17, 17, 19, 19, 19, 19, 19, 23, 30, 32, 32, 32, 32, 34, 34,
34, 35, 36, 38, 43
[ER93] Edward G. Effros and Zhong-jin Ruan. On the abstract characterization of

operator spaces. Proc. Amer. Math. Soc., 119(2):579–584, 1993. 6
[ER94] Edward G. Effros and Zhong-Jin Ruan. Mapping spaces and liftings for op-
erator spaces. Proc. London Math. Soc. (3), 69(1):171–197, 1994. 43, 56, 57,
57, 57, 58
[EW97a] Edward G. Effros and Corran Webster. Operator analogues of locally convex
spaces. In Operator algebras and applications (Samos, 1996), pages 163–207.
Kluwer Acad. Publ., Dordrecht, 1997. 49, 51, 52
[EW97b] Edward G. Effros and Soren Winkler. Matrix convexity: operator analogues
of the bipolar and Hahn-Banach theorems. J. Funct. Anal., 144(1):117–152,
1997. 3, 48, 51, 52
[Eym64] Pierre Eymard. L’algèbre de Fourier d’un groupe localement compact. Bull.
Soc. Math. France, 92:181–236, 1964. 33
[Far92] D. R. Farenick. C ∗ -convexity and matricial ranges. Canad. J. Math.,

44(2):280–297, 1992. 54
[Fis96] Hans-Jörg Fischer. Struktur Matrix konvexer Mengen (sic!). Diplomarbeit,

Universität der Saarlandes, 1996. 53, 53, 55
[FM93] D. R. Farenick and Phillip B. Morenz. C ∗ -extreme points of some compact

C ∗ -convex sets. Proc. Amer. Math. Soc., 118(3):765–775, 1993. 54
[Fuj94] Ichiro Fujimoto. Decomposition of completely positive maps. J. Operator

Theory, 32(2):273–297, 1994. 48
[FZ98] Douglas R. Farenick and Hongding Zhou. The structure of C ∗ -extreme points
in spaces of completely positive linear maps on C ∗ -algebras. Proc. Amer.
Math. Soc., 126(5):1467–1477, 1998. 48
[GK69] I. C. Gohberg and M. G. Kreı̆n. Introduction to the theory of linear non-

selfadjoint operators. American Mathematical Society, Providence, R.I., 1969.
Translated from the Russian by A. Feinstein. Translations of Mathematical
Monographs, Vol. 18. 16
[Haa80] Uffe Haagerup. Decomposition of completely bounded maps on operator al-

gebras. September 1980. 3, 33, 48
[Hof95] Helmut Hofmeier. Vollständig beschränkte Modul-Homomorphismen über von

Neumann-Algebren. Dissertation, Universität der Saarlandes, 1995. 21
[HT83] Tadasi Huruya and Jun Tomiyama. Completely bounded maps of C ∗ -algebras.
J. Operator Theory, 10(1):141–152, 1983. 47
[Kir83] Eberhard Kirchberg. The Fubini theorem for exact C ∗ -algebras. J. Operator
Theory, 10(1):3–8, 1983. 31
[Kir94] Eberhard Kirchberg. Commutants of unitaries in UHF algebras and functorial

properties of exactness. J. Reine Angew. Math., 452:39–77, 1994. 31
[Kir95] Eberhard Kirchberg. On subalgebras of the CAR-algebra. J. Funct. Anal.,

129(1):35–63, 1995. 31
[KS71] M. Ĭ. Kadets and M. G. Snobar. Certain functionals on the Minkowski com-
pactum. Mat. Zametki, 10:453–457, 1971. 11
[Lam97] Anselm Lambert. Homogene hilbertsche Operatorräume und ihre vollständig

beschränkten Abbildungen. Diplomarbeit, Universität der Saarlandes, 1997.
11, 16
[Loe75] Richard I. Loebl. Contractive linear maps on C ∗ -algebras. Michigan Math.

J., 22(4):361–366 (1976), 1975. 11, 41, 47
[LP81] Richard I. Loebl and Vern I. Paulsen. Some remarks on C ∗ -convexity. Linear
Algebra Appl., 35:63–78, 1981. 54, 54
[Mag98] Bojan Magajna. On c∗ -extreme points. Preprint, 1998. 54
[Mag00] B. Magajna. C ∗ -convex sets and completely bounded bimodule homomor-

phisms. Proc. Roy. Soc. Edinburgh Sect. A, 130(2):375–387, 2000. 54
[Mat94] Ben Mathes. Characterizations of row and column Hilbert space. J. London
Math. Soc. (2), 50(1):199–208, 1994. 16, 18, 18, 18, 18
[MN94] Paul S. Muhly and Qi Yuan Na. Extension of completely bounded A-B bi-
module maps. Glasgow Math. J., 36(2):145–155, 1994. 21
[Mor94] Phillip B. Morenz. The structure of C ∗ -convex sets. Canad. J. Math.,

46(5):1007–1026, 1994. 48, 52, 53, 54
[MP94] B. Maurey and G. Pisier. Espaces de Banach classiques et quantiques. Société

Mathématique de France, Paris, 1994. 3
[MP95] D. Benjamin Mathes and Vern I. Paulsen. Operator ideals and operator spaces.
Proc. Amer. Math. Soc., 123(6):1763–1772, 1995. 16, 16
[Mur90] Gerard J. Murphy. C ∗ -algebras and operator theory. Academic Press Inc.,
Boston, MA, 1990. 44
[Pat88] Alan L. T. Paterson. Amenability. American Mathematical Society, Provi-

dence, RI, 1988. 44
[Pau82] Vern I. Paulsen. Completely bounded maps on C ∗ -algebras and invariant

operator ranges. Proc. Amer. Math. Soc., 86(1):91–96, 1982. 3
[Pau86] Vern I. Paulsen. Completely bounded maps and dilations. Longman Scientific
& Technical, Harlow, 1986. 3, 4, 5, 14, 15, 21, 21, 22
[Pau92] Vern I. Paulsen. Representations of function algebras, abstract operator

spaces, and Banach space geometry. J. Funct. Anal., 109(1):113–129, 1992.
11, 17, 17
[Pet97] Uwe Peters. Duale Operatorräume und der Standard-Prädual von

s
CB (B(H)). Diplomarbeit, Universität der Saarlandes, 1997. 23, 23, 23,
43
[Pie67] A. Pietsch. Absolut p-summierende Abbildungen in normierten Räumen. Stu-

dia Math., 28:333–353, 1966/1967. 17
[Pie78] Albrecht Pietsch. Operator ideals. VEB Deutscher Verlag der Wissenschaften,
Berlin, 1978. 56
[Pis86] Gilles Pisier. Factorization of linear operators and geometry of Banach spaces.
Published for the Conference Board of the Mathematical Sciences, Washing-
ton, D.C., 1986. 34
[Pis95] Gilles Pisier. Exact operator spaces. Astérisque, (232):159–186, 1995. Recent
advances in operator algebras (Orléans, 1992). 31, 31, 31
[Pis96] Gilles Pisier. The operator Hilbert space OH, complex interpolation and tensor
norms. Mem. Amer. Math. Soc., 122(585):viii+103, 1996. 4, 15, 16, 16, 34,
62, 62
[Pis97] Gilles Pisier. Espaces d’opérateurs: une nouvelle dualité. Astérisque,

(241):Exp. No. 814, 4, 243–272, 1997. Séminaire Bourbaki, Vol. 1995/96.
3
[Pop00] Ciprian Pop. Bimodules normés représentables sur des espaces hilbertiens. In
Operator theoretical methods (Timişoara, 1998), pages 331–370. Theta Found.,
Bucharest, 2000. 20, 20
[Pow74] Robert T. Powers. Selfadjoint algebras of unbounded operators. II. Trans.

Amer. Math. Soc., 187:261–293, 1974. 49
[PPS89] Vern I. Paulsen, Stephen C. Power, and Roger R. Smith. Schur products and
matrix completions. J. Funct. Anal., 85(1):151–178, 1989. 48
[PS87] V. I. Paulsen and R. R. Smith. Multilinear maps and tensor norms on operator
systems. J. Funct. Anal., 73(2):258–276, 1987. 34
[Rob91] A. Guyan Robertson. Injective matricial Hilbert spaces. Math. Proc. Cam-
bridge Philos. Soc., 110(1):183–190, 1991. 17
[Rua88] Zhong-jin Ruan. Subspaces of C ∗ -algebras. J. Funct. Anal., 76(1):217–230,

1988. 3, 4, 6
[Rua89] Zhong-jin Ruan. Injectivity of operator spaces. Trans. Amer. Math. Soc.,
315(1):89–104, 1989. 42, 43
[Sat82] Ulrich Saterdag. Vollständig positive und vollständig beschränkte modulhomo-

morphismen auf operatoralgebren. Diplomarbeit, Universität der Saarlandes,
1982. 48
[Sch70] Robert Schatten. Norm ideals of completely continuous operators. Springer-

Verlag, Berlin, 1970. Second printing. Ergebnisse der Mathematik und ihrer
Grenzgebiete, Band 27. 16, 43
[Smi83] R. R. Smith. Completely bounded maps between C ∗ -algebras. J. London

Math. Soc. (2), 27(1):157–166, 1983. 5, 47, 47, 47
[Smi91] R. R. Smith. Completely bounded module maps and the Haagerup tensor
product. J. Funct. Anal., 102(1):156–175, 1991. 21, 47, 48, 48
[SS95] Allan M. Sinclair and Roger R. Smith. Hochschild cohomology of von Neumann
algebras. Cambridge University Press, Cambridge, 1995. 46, 48, 48
[Sti55] W. Forrest Stinespring. Positive functions on C ∗ -algebras. Proc. Amer. Math.

Soc., 6:211–216, 1955. 3
[Tak79] Masamichi Takesaki. Theory of operator algebras. I. Springer-Verlag, New

York, 1979. 22
[Wit81] Gerd Wittstock. Ein operatorwertiger Hahn–Banach Satz. J. Funct. Anal.,

40(2):127–150, 1981. 3, 22
[Wit84a] G. Wittstock. Extension of completely bounded C ∗ -module homomorphisms.

In Operator algebras and group representations, Vol. II (Neptun, 1980), pages
238–250. Pitman, Boston, Mass., 1984. 4, 21
[Wit84b] Gerd Wittstock. On matrix order and convexity. In Functional analysis: sur-
veys and recent results, III (Paderborn, 1983), pages 175–188. North-Holland,
Amsterdam, 1984. 49
[Woj91] P. Wojtaszczyk. Banach spaces for analysts. Cambridge University Press,

Cambridge, 1991. 11
[WW99] Corran Webster and Soren Winkler. The Krein-Milman theorem in operator
convexity. Trans. Amer. Math. Soc., 351(1):307–322, 1999. 48, 50, 52, 53, 53
Index
Subindex completely bounded, 23, 37, see
see amplification Haagerup tensor product, 39
see axioms norm k·kcb , 39
see bilinear mapping representation, 40
see C ∗ -algebra space CB (X × Y ; Z), 40
see completelely bounded mapping transposition, 40
see examples joint amplification, 38
see Haagerup tensor product Φ(p×q) , 38
see Hilbertian operator space matrix duality, 38
see injective op. space tensor product jointly completely bounded, 37, 37,
see lemma of see
see mapping projective operator space tensor product,
see mapping space 39
see module homomorphism norm k·kjcb , 39
see multilinear mapping space JCB (X × Y ; Z), 39
see operator algebra transposition, 39
see operator module linearization Φ̃, 38
see operator space transposed mapping Φt , 39, 40
see operator space tensor product bipolar theorem, 50
see projective op. space tensor product
see special operator spaces Calkin algebra, 22
see theorem of CB -ideal property, 55
absolutely matrix convex, 48 CB -mapping space, 54
absolutely matrix convex combination, characterization
48 completely locally reflexive operator
abstract operator space, see spaces, 42
operator space Christensen
algebra homomorphism, 24 theorem of, 43
amplification, 4, bilinear mapping, see column Hilbert space, see
bilinear mapping hilbertian operator space
joint, see column norm, 17
bilinear mapping column space Cp , 7
automatic complete boundedness, 45 complete quotient mapping, see
axioms, of Ruan completely bounded mapping
operator module, 20 completely bounded mapping, 4, com-
operator space, 3 plete quotient mapping, 4
completely contractive, 4
bilinear mapping, amplification, 23, 38 completely isometric, 4
to quadratic matrices Φ(n) , 38 completely contractive, see
to rectangular matrices Φ(n,l) , 38 completely bounded mapping
completely isometric, see
73
completely bounded mapping, 9 operator space, 5, 7

completely locally reflexive, 42 operator spaces
concrete operator space, see completely locally reflexive, 43
operator space projective op. space tensor product,
convex combination 32
absolutely matrix convex combina- extension theorem
tion, 48 for completely bounded module ho-
matrix convex combination, 48 momorphisms, 21
cross norm, 28 for completely bounded multilinear
C ∗ -algebra, 5, 19, 23, commutative, 5, mappings, 44
11, 43
injective, 22 factorization
nuclear, 43 through a column Hilbert space, see
quasi-cyclic, 46 hilbertian operator space
finite dimensional mappings, 54
decomposition theorem
for completely bounded symmetric Grothendieck
multilinear mappings, 45 H-tensor product, 34
for self-adjoint completely bounded H-tensor product, 34
bimodule homomorphisms, 22 Haagerup tensor, norm, 33
of Tomiyama-Takesaki, 22 infimum formula, 33
derivations minimum formula, 34
completely bounded, 43 sum formula, 34
inner, 43 supremum formula, 35
dominate, 4, 24 product, 33
dual H-tensor product, 34
of an operator space, see interpolation, 34
operator space involving column Hilbert space,
operator space tensor norm, see 36, 37
operator space tensor norm involving row Hilbert space, 36, 37
operator space tensor product, see lemma of Blecher and Paulsen, 36
operator space tensor product shuffle map, 35
tensor matrix product h , 36
embedding
hilbertian operator space, 15
canonical, 8, 57
column Hilbert space C, 15, 17
examples, 48, Haagerup tensor product,
as an operator algebra, 24
34
characterizations, 18
hilbertian operator spaces, 15
factorization through a, 18
injective op. space tensor product,
homogeneous, 15
30
completely self dual, 15
module Haagerup tensor product,
maximal, 15
42
minimal, 15
of operator modules, 20
morphisms, 16
operator algebra, 23, 24
operator Hilbert space OH , 15, 61
as an operator algebra, 24 mapping space, 54, see

row Hilbert space R, 15 Abbildungsraum, completely in-
as an operator algebra, 24 tegral mappings, 56
Hochschild cohomology, 43 completely nuclear mappings, 55
matricially normed space, 4, see
ICM 1986, 3 space
ideal matrix
of an operator algebra, see set of matrices, 47
operator algebra matrix cone, 48
Schatten class, 16 matrix convex, 48
symmetrically normed, 16 absolutely, 48
ideal property, 55 matrix convex combination, 48
injective operator space, tensor norm, 29 matrix duality, see
tensor product, 29 bilinear mapping
and λ-tensor product, 30 matrix multiplication, 23
interpolation of operator spaces, 61 matrix polar, 50
interpolation space, 61 absolute, 50
intersection of operator spaces, 60 module homomorphism, 22
completely bounded
joint
decomposition theorem, 22
amplification, see
extension theorem, 21
bilinear mapping
module Haagerup tensor product,
lemma of, see 23, 42
theorem of representation (K; π; v ∗ ; w), 21
Blecher and Paulsen, 36 representation theorem for, 20
Smith, 5, 46 normal, 22
locally reflexive, 42 self-adjoint, 21
completely, 42 singular, 22
multilinear mapping, amplification, 44
M-ideal, 42 completely bounded, 44
mapping, complete quotient mapping, 9 extension theorem, 44
completely bounded, 3, see representation theorem, 44
completely bounded mapping completely bounded symmetric
completely integral, 56 decomposition theorem, 45
characterization, 57 linearization, 44
completely nuclear, 55 Haagerup tensor product, 44
characterization, 56 multimodular, 47
completely positive, 3
Hilbert-Schmidt, 16 normal
integral, 57 module homomorphism, see
nuclear, 16, 56 module homomorphism
shuffle, see
operator
operator space tensor product,
2-summing, 16
27
adjoint, 9 projective, see

operator algebra, 23 projective
abstract, 23 spatial, see
concrete, 23 injective
ideal in an, 24 product, 25
Ruan’s theorem, 23 associativity of, 28
operator ideal, 55 axioms, 26
operator module, see cross norm, 28
module homomorphisms, Haagerup, see
(A1 , A2 )-, 20 Haagerup tensor product
abstract, 20 injective, see
concrete, 20 injective
examples, 20 injectivity of, 28
operator bimodule maximal, see
A-, 20 projective
representation theorem, 20 minimal, see
operator space, 3, 4, abstract, 5 injective
completely locally reflexive, 42 projective, see
concrete, 5 projective
dual, 8 projectivity of, 28
finite dimensional, 42 self-duality of, 28
hilbertian, 15, see shuffle mapping, 27
hilbertian operator space spatial, see
homogeneous, 4 injective
completely self dual, see symmetry of, 28
hilbertian operator space operator space tensor norm
injective, 17 projective, 55
interpolated, 61 operator systems, 3
matrix level, 4, 6
maximal, 10, 50 Paulsen
minimal, 10, 50 conjecture of, 16
operator space norm, 3 projective operator space, tensor norm,
quotient of, 6 31
reflexive, 8 tensor product, 31
subspace of, 6 and γ-tensor product, 32
operator space tensor, norm, 26 column Hilbert space, 32
Haagerup, see Fourier algebra, 32
Haagerup tensor norm predual of a von Neumann alge-
injective, see bra, 32
injective row Hilbert space, 32
maximal, see
quasi-cyclic, 46
projective
quotient
minimal, see
injective
of an operator space, see state space

operator space generalized, 51
quotient mapping structural element, 51
complete, see subspace
mapping of an operator space, see
operator space
representation theorem sum of operator spaces, 60, 61
for completely bounded module ho-
momorphisms, 20 tensor norm
for completely bounded multilinear of operator spaces, see
mappings, 44 operator space tensor norm
operator modules, 20 tensor product
row Hilbert space, see of operator spaces, see
hilbertian operator space operator space tensor product
row norm, 17 cross norm, 28
row space Rp , 7 Haagerup, see
Ruan Haagerup
axioms theorem
operator modules, 20 bipolar theorem, 50
operator space, 3 separation theorem, 49
theorem of, 5 theorem of, see
lemma of
Schatten ideals Sp , 16, 24 Christensen, 43
operator space structure, 61 Ruan, 5
Schur product, 24 for operator algebras, 23
self-adjoint topology
algebra in B(H), 23 non commutative, 3
separation theorem, 49
set of matrices, 47
shuffle mapping, see
operator space tensor product,
27
similarity problem, 43
singular
module homomorphism, see
module homomorphism
singular values, 16
special operator spaces, CB (X, Y ), 7
intersection of operator spaces X ∩
Y , 60
operator Hilbert space OH , 61
sum of op. spaces X + Y , 61
standard dual, 8
∗ -representation, 20

What Are Operator Spaces?

Uploaded by

Copyright:

Available Formats

What Are Operator Spaces?

Uploaded by

Document Information

Original Description:

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

What Are Operator Spaces?

Uploaded by

Copyright:

Available Formats

What are operator spaces?

Operator algebra group

2 Operator Spaces and Completely Bounded Maps 3

3 Operator Systems and Completely Positive Maps 13

4 Hilbertian Operator Spaces 15

7 Complete Local Reflexivity 43

8 Completely Bounded Multilinear Mappings 44

9 Automatic Complete Boundedness 46

2 Operator Spaces and Completely Bounded Maps

(R1) kαxβk ≤ kαkkxkkβk for all x ∈ Mn (X), α ∈ Mm,n , β ∈ Mn,m

Then every one of these mappings k · k : Mn (X) → lR is a seminorm. If one (and

(R10 ) kαxβk ≤ kαkkxkkβk for all x ∈ Mn (X), α ∈ Mn , β ∈ Mn ,

which is often found in the literature, or

(R1) kαxβk ≤ kαkkxkkβk for all x ∈ Mn (X), α ∈ Mm,n , β ∈ Mn,m ,

which seems to be appropriate in convexity theory.

A matricially normed space is a complex vector space with a matrix norm. It

the nth amplification of Φ.

Φ is called completely bounded if kΦkcb < ∞ and completely contractive if

the notation becomes simpler.7

Mn,m (B(H)) = B(Hm , Hn )

kΦkcb = sup{kΦ(x)k | x ∈ M (X), kxk ≤ 1}.

2.2 Ruan’s theorem

2.3 Elementary constructions

Matrices over an operator space

for Mp (X) with this operator space structure. In particular,

M1 (lMp (X)) = Mp (X)

Mn (Mp,q (X)) = Mnp,nq (X).

By adding zeros it is a subspace of lMr (X) for r ≥ p, q.

lMp := lMp (Cl ).

Columns and rows of an operator space

Cp (X) := lMp,1 (X) and Rp (X) := lM1,p (X).

The first matrix level of these spaces are

2.4 The space CB (X, Y )

Defining k[Tij ]k = kT kcb , CB (X, Y ) becomes a matricially normed space. It is an

2.5 The dual

Thus we have an algebraic identification of Mn (X ∗ ) and Mn (X)∗ and further of Mn (X ∗∗ )

Mk (lMn (X)∗∗ ) = Mk (Mn (X))∗∗ = Mkn (X)∗∗ = Mkn (X ∗∗ ) = Mk (lMn (X ∗∗ )).

The adjoint operator

even is completely isometric [Ble92b, Lemma 1.1].13

2.6 Direct sums

Mn (CB (X, Y )) = M1 (CB (X, lMn (Y )))

2.7 MIN and MAX

M1 (CB (MAX (E), X)) = B(E, M1 (X))

M1 (CB (X, MIN (E))) = B(M1 (X), E)

is not completely bounded[Pau92, Cor. 2.13].15

kxk = sup{kϕ(n) (x)k | n ∈ IN, ϕ : E → Mn , kϕk ≤ 1}.

2.8 Injective operator spaces

It suffices to consider only operator spaces Y0 and Y . Injective matricially normed

2.8.2 Examples and elementary constructions

A matricially normed space Z together with a completely isometric mapping ι :

3 Operator Systems and Completely Positive Maps

Let V be an involutive vector space. Then V is called an operator system if it is a

where Sn is the set of all unital completely positive maps ϕ : V → lMn .

3.3 Matrix order unit norm

3.4 Injective operator systems

4 Hilbertian Operator Spaces

4.2 The morphisms

1. (CB (X, Y ), k · kcb ) is a Banach space.