Notes on duality methods and operator spaces

David P. Blecher
Department of Mathematics, University of Houston
May 7, 2007
4
Preface
These are fairly complete notes from part of a course I taught in Fall 2006 at the
University of Houston. The students had taken a course on C

-algebras with me
the previous year, and these notes assume familiarity with parts of that course.
Chapter 1 includes material usually met with in a year long functional analysis
course, and is stolen from the usual sources (particularly [24]). Chapter 2 is a
considerable expansion of most of Chapter 1 of my book [4] with Christian Le
Merdy. The presentation here is thus greatly shaped by that book (indeed some
is copied verbatim), and of course thanks go to Christian for permitting me to do
this. Since these notes were aimed at the students in the class, I have not yet taken
the trouble to compile an adequate bibliography, or to make sure that results are
always attributed, etc.
Revision of January, 2007
5
6
Contents
1 Topological vector spaces and duality 9
1.1 Topological vector spaces . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2 A linear algebraic interlude . . . . . . . . . . . . . . . . . . . . . . . 14
1.3 Weak topologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.4 The weak* topology . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.5 Extreme points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.6 The classical Shilov boundary . . . . . . . . . . . . . . . . . . . . . . 28
2 Operator spaces 35
2.1 Basic facts, examples, and constructions . . . . . . . . . . . . . . . . 35
2.2 Duality of operator spaces . . . . . . . . . . . . . . . . . . . . . . . . 53
2.3 Operator systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.4 Operator space tensor products . . . . . . . . . . . . . . . . . . . . . 67
2.5 Properties of completely bounded maps . . . . . . . . . . . . . . . . 84
2.6 Duality and tensor products of dual spaces . . . . . . . . . . . . . . 87
7
8 CONTENTS
Chapter 1
Topological vector spaces
and duality
1.1 Topological vector spaces
We recall from topology the notion of an initial topology. If X is a set, and T is a
collection of functions from X into various Hausdor topological spaces, the initial
topology induced by T on X is the weakest (coarsest) topology on X making all the
functions in T continuous. One then proves that a net x

x in X in this topology
i f(x

) f(x) for all f T. From this it is clear that a function g : Z X

from a topological space Z is continuous i f g is continuous for all f T. Also,
this topology on X is Hausdor if whenever x ,= y in X, then there exists an f T
such that f(x) ,= f(y).
By F we denote either the real or the complex eld. Vector spaces will be over
A topological vector space (TVS) is a vector space X over F with a topology such
that the vector space operations are continuous. We will always assume that the
topology is Hausdor. In terms of nets this means that if x
t
x and y
t
y in X,
and if a
t
a in F, then x
t
+y
t
x +y and a
t
x
t
ax. A TVS is locally convex if
it arises in the following way. Suppose that X is a vector space, and o is a family of
seminorms on X with the property that if x ,= y in X, then there exists an m o
such that m(x y) ,= 0. Let T be the set of scalar valued functions on X of the
form m( y), for m o and y X. By the last fact in the last paragraph, the
initial topology induced on X by T is Hausdor.
Proposition 1.1.1. If (X, ) and o are as in the last paragraph, then (X, ) is a
TVS. Also
(1) A net (x

) in X converges to x in (X, ) if and only if m(x

x) 0 for every
m o.
(2) If x X, then sets of the form

n
k=1
y X : m
k
(y x) < , m
1
, , m
n
o, > 0 (1.1)
9
10 CHAPTER 1. TOPOLOGICAL VECTOR SPACES AND DUALITY
are convex and open, and they form a neighborhood basis at x; indeed, for every
neighborhood U of x there is a set V of the form above such that x V U.
Proof. (1) If m(x

x) 0 for every m o, then by the triangle inequality, for

any y X we have
[m(x

y) m(x y)[ m(x

x) 0.
Thus m(x

y) m(x y). That is, f(x

) f(x) for any f in the set T, above.

By a statement in the rst paragraph of this section x

x in (X, ). The other

direction is much easier.
From (1) it is easy that (X, ) is a TVS. Indeed if x
t
x and y
t
y in X, and
if a
t
a in F, then m(x
t
x) 0 and m(y
t
y) 0 by (1), and so
m(x
t
+y
t
x y) m(x
t
x) +m(y
t
y) 0.
Thus x
t
+ y
t
x + y by (1) again. Also, since [a
t
[ [a[ there is a t
0
such that
[a
t
[ [a[ + 1 for t t
0
. For such t,
m(a
t
x
t
ax) m(a
t
x
t
a
t
x) +m(a
t
x ax) = [a
t
[m(x
t
x) +[a
t
a[m(x) 0.
(2) If m o and > 0, and if y
1
, y
2
X with m(y
k
x) < , then
m(
1
2
(y
1
+y
2
) x) =
1
2
m((y
1
x) + (y
2
x)) <
1
2
( +) = .
Thus y X : m(y x) < is convex. By the denition of an initial topology, this
set is also open. Clearly an intersection of convex open sets is convex and open.
This proves the rst assertion in (2).
We next claim that sets of the form (1.1) form the basis for another topology
on X, which will be coarser than . Indeed, any x X lies in a set of the form
(1.1), clearly. By basic topology, the claim follows if given two sets B
1
, B
2
of the
form (1.1), and an element z in their intersection, there is a third set B
3
of the
form (1.1) such that z B
3
B
1
B
2
. To see this, suppose that B
1
is specied
by x X, > 0 and m
1
, , m
n
o as in (1.1), and that B
2
is specied by
x

X,

> 0, and r
1
, , r
p
o. If z B
1
B
2
, choose > 0 smaller than every
number m
k
(z x), and every number

r
j
(z x

). Let
B
3
=
n
k=1
y X : m
k
(y z) <
p
k=1
y X : r
k
(y z) < .
This is a set of the form (1.1), and if y B
3
then for each k = 1, , n we have
m
k
(yx) m
k
(yz)+m
k
(z x) < +m
k
(z x) < m
k
(z x)+m
k
(z x) = .
Thus y B
1
, and similarly y B
2
. So z B
3
B
1
B
2
.
Let x

x in X in the topology . If > 0 and m o, then there exists

0
such that x

y X : m(y x) < whenever

0
. Hence for any y X,
by the triangle inequality, [m(x

y) m(x y)[ m(x

x) < . That, is each

function in T is -continuous, and so , so that = . If z U , there is a
set B of the form (1.1) with z B U. Taking B
1
= B
2
= B in the last paragraph
shows that there is a set B
3
B which is centered at z and of the required form
in (2).
1.1. TOPOLOGICAL VECTOR SPACES 11
Remarks. 1) In any TVS X, if y X and F with ,= 0 are xed, then
the following two maps are homeomorphisms: (x) = x y, and (x) = x. This
is because they are easily checked to be continuous, and their inverses, which are of
essentially the same form, are also continuous.
2) Any normed space (X, | |), with the norm topology (that is, the metric
space topology induced by the metric d(x, y) = |xy|) is a LCTVS. Indeed setting
o = | |, the sets of the form (1.1) are exactly the open balls in this metric space.
1.1.2 For a TVS X we write X

for the space of linear continuous functionals

from X into F.
Note that if a linear : X F is continuous at 0, then it is continuous
everywhere. For if x
t
x in X, then x
t
x 0, so that (x
t
x) 0, and
(x
t
) (x).
1.1.3 We recall that a Minkowski function on a vector space X is a function
m : X R which is subadditive (i.e. m(x + y) m(x) + m(y)) and positive
homogeneous (i.e. m(tx) = tm(x) if t 0).
We also recall the Hahn-Banach theorem for such m, proved in Real Analy-
sis/Functional Analysis: if Y is a linear subspace of X, and if : Y R is a linear
functional dominated by m on Y (i.e. (y) m(y) for all y Y ), then there exists a
linear extension : X R which is also dominated by m on X. That is,
|Y
= ,
and (x) m(x) for all x X.
Lemma 1.1.4. Let C be an open convex set containing 0 in a TVS X. Dene
m(x) = infs > 0 : s
1
x C, x X.
Then m is a Minkowski function on X and C = x X : m(x) < 1.
Proof. Note that if x X, then x/n 0 as n . Thus N N with x/n C
for n N. So m(x) N < .
If t > 0 then m(tx) = infs > 0 : s
1
tx C. Letting s

= s/t we have
m(tx) = infts

: s

> 0, x/s

C = tm(x).
If t = 0 it is obvious that m(tx) = tm(x).
Next, suppose that x/s C and y/t C for s, t > 0. Then
x +y
s +t
=
s
s +t
x
s
+
t
s +t
y
t
C,
since C is convex. Thus
m(x +y) infs +t : x/s C, y/t C, s, t > 0
= infs > 0 : x/s C + inft > 0 : y/t C
= m(x) +m(y).
Thus m is a Minkowski function on X. If m(x) < 1 then s (0, 1) with x/s C,
and so x = (1 s)0 + s
x
s
C. Conversely, if x C then since C is open tx C
for some t > 1 (by the continuity of the map t tx). So m(x) < 1.
12 CHAPTER 1. TOPOLOGICAL VECTOR SPACES AND DUALITY
Theorem 1.1.5. (A geometric Hahn-Banach theorem) If X be a TVS, and let
U, B be nonempty convex disjoint subsets of X, with U open. Then there exists a
continuous linear functional : X F, and a number R such that
Re (x) < Re (y), x U, y B.
If B is also open then you can take the here to be a < too.
Proof. First suppose that F = R. Choose x
0
U, y
0
B and set z = y
0
x
0
and C = U B + z = u b + z : u U, b B. Since U, B are convex it is
easy to check that C is convex. Moreover C is open, being the union of the open
sets U (b z), for b B. Also 0 C. By Lemma 1.1.4 there is a Minkowski
function m with C = x X : m(x) < 1. Since z / C we have m(z) 1. Dene

0
: Rz R by
0
(sz) = s. If s 0 then
0
(sz) = s sm(z) = m(sz). If
s < 0 then
0
(sz) = s 0 m(sz). Thus
0
is dominated by m on Rz. By
the Hahn-Banach theorem in 1.1.3, there exists a linear extension : X R with
(x) m(x) for all x X.
Let > 0 be given. For all x C we have (x) m(x) < 1. Thus if x C
then (x) < . If x ()C, then (x) > . If V = C ()C then V is open,
contains 0, and [(x)[ < for x V . Thus is continuous at 0, and hence by 1.1.2
it is continuous on X.
If x U, y B then x y +z C, so that (x y +z) < 1. But (z) = 1 and
so (x) < (y). Let t = sup(x) : x U. Then
(x) t (y), x U, y B.
We now show that the rst can never be an equality. Suppose that x
0
U
with (x
0
) = t. Choose w X with (w) = 1. Since the map r x
0
+ rw is
continuous, there exists an > 0 such that x
0
+ rw U whenever [r[ < . Hence
(x
0
) +r = (x
0
+rw) t, if 0 < r < . This contradicts (x
0
) = t.
A similar argument shows that if B is open, the other inequality is strict too.
Finally, if F = C then regard X as a real vector space. By the above there exists
an R-linear continuous : X R with (x) < t (y), for x U, y B. Let
(x) = (x) i(ix). This is a common trick, which you will have met. Clearly
is continuous, Re = , and we are done.
Theorem 1.1.6. (A second geometric Hahn-Banach theorem) If X is a LCTVS,
and if C, K are nonempty closed convex disjoint subsets of X, with K compact,
then there exists a continuous linear functional on X, and numbers , with
> 0 such
Re (x) < + Re (y), x C, y K.
Proof. If y K, then C y is closed, so its complement is an open set containing
0. By Proposition 1.1.1, there is an open convex set V
y
centered at 0, contained
in this open set. So
V
y
(C y) = . (1.2)
Consider the open cover y +
1
3
V
y
: y K of K. It has a nite subcover y
1
+
1
3
V
y
1
, , y
m
+
1
3
V
y
m
say. Then V =
k
1
3
V
y
k
is an open convex set containing 0.
1.1. TOPOLOGICAL VECTOR SPACES 13
It is easy to see that K+V =
yK
y +V is open and convex, as is C+V . We now
show that (K + V ) (C + V ) = . Indeed suppose that y K, z C, v
1
, v
2
V
with y +v
1
= z +v
2
. For some j we have y y
j
+
1
3
V
y
j
. Also, v
1
, v
2

1
3
V
y
j
, and so
z = y +v
1
v
2
= y
j
+ (y y
j
) +v
1
v
2
y
j
+
1
3
V
y
j
+
1
3
V
y
j
+
1
3
V
y
j
,
so that z y
j
+V
y
j
, contradicting (1.2).
By Theorem 1.1.5 there exists a continuous linear functional : X F, and a
number R such that
Re (x) < < Re (y), x K +V, y C +V.
In particular this holds for x K, y C. Let = infRe (x) : x K. Since K
is compact, this inmum is a minimum. Hence < , and = > 0.
1.1.7 If X is a TVS and A X we dene the polar
A

= X

: Re((x)) 1 for all x A.

This is convex, being an intersection of the convex sets X

: Re((x)) 1,
for all x A. It also contains 0. In our applications usually A has the property
that x A whenever [[ 1 and x A, and in this case it is easy to see that
A

= X

: [(x)[ 1 for all x A.

Next, if B is any subset of X

, we dene the prepolar

B

= x X : Re((x)) 1 for all B.

If B has the property that B whenever [[ 1 and B, then as above
B

= x X : [(x)[ 1 for all B.

It is easy to see that a prepolar is always closed and convex, being the intersection
of closed and convex sets x X : Re((x)) 1, and it contains 0. The bipolar
of A is (A

. Clearly A (A

, so that the bipolar of A contains the closure of

A.
Theorem 1.1.8. (Bipolar theorem) If X is a LCTVS, and if C is a nonempty
convex subset of X whose closure contains 0, then the bipolar of C is the closure of
C.
Proof. Let D be the closure of C, which is easy to see is also convex. We already
observed that the bipolar of C contains the closure of C. Suppose that x / D.
The complement of D is an open set containing x. By Proposition 1.1.1, there is
an open convex set V centered at x, contained in this open set. By the geometric
Hahn-Banach theorem 1.1.5, there is a continuous X

and R with
Re (x) < Re (y), y D.
14 CHAPTER 1. TOPOLOGICAL VECTOR SPACES AND DUALITY
Now 0 D, and so we may assume that < 0. Dividing by [[, and letting
= /[[, we have
Re (x) < 1 Re (y), y D.
Thus C

. This, together with the left side of the centered equation shows that
x / (C

. Thus D = (C

.
1.2 A linear algebraic interlude
The next few results are pure linear algebra:
Proposition 1.2.1. Let X be a vector space, and let ,
1
,
2
, ,
n
be linear
maps from X to the scalar eld F. The following are equivalent:
(i) =

k

k

k
for some xed scalars
1
, ,
n
F.
(ii) There is a constant K 0 such that [(x)[ K max
k
[
k
(x)[ for all x X.
(iii)
k
Ker(
k
) Ker().
Proof. (i) (ii)
[(x)[ =

k
(x)

k
[
k
[[
k
(x)[
_

k
[
k
[
_
max
k
[
k
(x)[.
(ii) (iii) Clear. (iii) (i): Dene T : X F
n
by T(x) = (
1
(x),
2
(x), ,
n
(x)).
By the hypothesis (iii) we easily see that Ker(T) Ker(). Let Q : X X/Ker(T)
be the canonical quotient map. Thus by the factor theorem in Linear algebra,
induces a linear map : X/Ker(T) F with Q = . By the factor the-
orem we also have a 1-1 linear

T : X/Ker(T) F
n
with

T Q = T. Let
Y = Ran(T) = Ran(

T) F
n
. Let g =

T
1
, a linear map from Y to F. It is easy
to see that we can extend any linear map g dened on a vector subspace of F
n
, to
a linear map g

dened on all of F
n
(for example if y
1
, , y
m
is a basis for the
subspace, contained in a basis y
1
, , y
m
, y
n
of F
n
, then dene g

(y
k
) = g(y
k
)
for k = 1, , m, and g

(y
k
) = 0 if m < k n). Also, by easy linear algebra, a
linear map g

: F
n
F must be of the form g

((x
1
, x
2
, , x
n
)) =

k

k
x
k
, for
some xed scalars
1
, ,
n
F. Thus,
(x) = (Q(x)) = (

T
1

TQ(x))) = g(T(x)) = g

((
1
(x),
2
(x), ,
n
(x)),
and so (x) =

k

k

k
(x), for all x X.
The sets Ker() considered in the last Proposition have a nice characterization
too:
Proposition 1.2.2. Let X be a vector space, and let H be a proper linear subspace
of X. The following are equivalent:
(i) H = Ker() for a nonzero linear map : X F.
1.2. A LINEAR ALGEBRAIC INTERLUDE 15
(ii) H is a maximal subspace of X. That is, for any x
0
X H we have Span
H, x
0
= X.
(iii) H has codimension 1. That is, X/H is one dimensional.
The in (i) is unique up to a scalar.
Proof. (ii) (iii) Assuming (ii), pick a nonzero coset v +H V/H. So v / H, and
by hypothesis Span H, v = X. If x X then x = v +h for some h H, F,
so that x +H = (v +H).
(iii) (i) Assuming (iii), let : X/H F be an isomorphism, and let Q : X
X/H be the canonical quotient map. Then Q is the desired .
(i) (ii) Assume (i), and pick x
0
X Ker(). For any x X let h =
x
(x)
(x
0
)
x
0
. Then
(h) = (x)
(x)
(x
0
)
(x
0
) = 0.
Thus h H = Ker(), and x is h plus a scalar times x
0
. Thus x SpanH, x
0
.
Since x was arbitrary, we are done.
The uniqueness of follows from Proposition 1.2.1.
We call a subspace H of X of the type characterized in Proposition 1.2.2 a
hyperplane in X through the origin. By a hyperplane in X we mean a translation
of a hyperplane H through the origin. That is, a subset of X of the form H +w =
x +w : x H, for a hyperplane H through the origin, and a xed vector w X.
It is easy to see that the hyperplanes in R
2
(resp. R
3
) are lines (resp. planes).
Corollary 1.2.3. Let X be a vector space, and let K be a subset of X. Then K
is a hyperplane in X if and only if there is a scalar t F and a nonzero linear map
: X F, such that
K = x X : (x) = t.
Proof. Suppose that t, are as above. Since is nonzero it is clear, by multiplying
through by a scalar, that there exists a w X with (w) = t. If (x) = t then
clearly (x w) = 0, so that x w Ker(), or equivalently, x Ker() + w.
Thus
x X : (x) = t = Ker() +w.
The converse is similar.
Lemma 1.2.4. Let X be a TVS. A linear functional : X F is continuous i
Ker() is closed.
Proof. The one direction is trivial. For the other, we will use several general ob-
servations. First, if N is any closed subspace of X, then X/N is a TVS with the
quotient topology, and in this case the quotient map q : X X/N is open. Indeed,
if V is an open set in X, then the saturation of V , namely the set of x X such
that x v N for some v V , is open. Indeed this set is exactly
xN
(V + x).
From general topology it follows that q is a continuous open map. Now it is easy to
16 CHAPTER 1. TOPOLOGICAL VECTOR SPACES AND DUALITY
see that X/N is a TVS. For example if W is an open neigborhood of q(x
0
) +q(y
0
)
in X/N, then since x
0
+y
0
q
1
(W), there exist open neigborhoods U, V of x
0
, y
0
respectively, with U + V q
1
(W). Hence q(U) + q(V ) W, and q(U) and q(V )
are open neigborhoods of q(x
0
) and q(y
0
) respectively. This veries one of the two
conditions for X/N to be a TVS, and the other is similar. To see that X/N is
Hausdor, suppose that q(x) ,= q(y), or, equivalently, that x y / N. Then since
N
c
is open, there exist open V, W X with x V, y W and V W N
c
. The
latter condition means that q(U) q(V ) = , so that X/N is Hausdor.
The next general observation we make is that if U is an open neighborhood of 0
in a TVS X, then there exists an open set V U with the property that if x V
and F with [[ 1, then x V . To see this, note that by denition of X being
a TVS, there is an > 0 and an open W X with W U whenever [[ . Let
V =
||1
W, this does the trick.
Next we observe that there is only one topology on F with respect to which it is
a TVS. Indeed, if X = F with such a topology , then by denition of a TVS, the
identity map F X is continuous. Thus is contained in the usual topology. Also
if U is the union of the open disk center 0 of radius 1, with the complement of the
closed disk center 0 of radius 2, then U is open, since its complement is compact
(being the image of a compact in F). On the other hand, by the last paragraph,
there is a -open set V U which is a disk center 0. That is, V = x F : [x[ < r
for some r > 0. Since X is a TVS, all open disks are in , so that is the usual
topology.
Finally, putting the observations above together with Proposition 1.2.2, we see
that may be regarded as the composition of two continuous maps: the quotient
map q : X X/Ker(), and the induced isomorphism : X/Ker() F. The
latter must be a homeomorphism by the fact in the last paragraph.
If F = R we can rephrase Theorems 1.1.5 and 1.1.6 in terms of hyperplanes. For
example, in 1.1.5 the condition
(x) < (y), x U, y B,
is just saying that U lies strictly below and B lies above the hyperplane H =
x X : (x) = . Similarly, 1.1.6 we have two parallel hyperplanes separating
C and K.
Note that in a TVS X, if : X F is continuous then the hyperplanes
associated with it as above are closed.
1.3 Weak topologies
The most frequently encountered case of a LCTVS construction is when we are given
a vector space X, and a linear subspace J Lin(X, F), such that whenever x ,= y
in X, then there exists J such that (x) ,= (y). Let o = [()[ : J.
Applying the construction before Proposition 1.1.1 to the set of seminorms o on X,
we obtain a topology on X making X into a LCTVS. We call this the weak topology
on X induced by J. Note that by Proposition 1.1.1 (1), it is clear that a net x

x
1.3. WEAK TOPOLOGIES 17
in X in this topology i [(x

) (x)[ 0 for all J, i (x

) (x) for all

J.
Proposition 1.3.1. Suppose that X, J are as in the last paragraph, and that
X is given the weak topology induced by J. A linear functional : X F is
continuous in this topology i J.
Proof. By the observation above the Proposition, any J is continuous. If
: X F is continuous in this topology then it is continuous at 0. Hence by the
last assertion of Proposition 1.1.1 there is a set U =
n
k=1
x X : [
k
(x)[ < ,
for some > 0 and
k
J, with [(x)[ < 1 whenever x U. It follows that for
any x X, we have [(x)[
1

max
k
[
k
(x)[. (This may be seen as a consequence
of the following elementary Exercise: If m
1
, m
2
: X [0, ) on a vector space X,
and if m
k
(tx) = tm
k
(x) for all x X, t 0, and if there are positive numbers ,
such that if m
1
(x) < then m
2
(x) < , then m
1

m
2
.)
By Proposition 1.2.1 we see that
k
is a linear combination of the
k
, so that
J.
Denition 1.3.2. If X is a normed space, and J = X

, then the weak topology

induced on X by J is called the weak topology of X.
Note that we are using the usual Hahn-Banach theorem here somewhere! We
also remark that some authors write (X, X

) for the weak topology on X.

Corollary 1.3.3. If X is a normed space, then
(1) A net (x

) in X converges to x in the weak topology if and only if (x

) (x)
for all X

.
(2) A linear functional : X F is continuous in the weak topology i : X F
is continuous in the norm topology.
Proof. (1) This follows from the observation just above Proposition 1.3.1.
(2) This is exactly Proposition 1.3.1 in this setting.
We often write , x) for the canonical pairing of a space with its dual, or with
its predual. In the above, , x) = (x).
Theorem 1.3.4. (Mazur) If C is a convex set in a normed space then its norm
closure equals its closure in the weak topology.
Proof. If x is in the norm closure of C then there is a net from C converging in the
norm topology, and hence by Corollary 1.3.3 (1) in the weak topology. So x is in
the weak closure of C.
Now suppose that x is not in the norm closure of C. So there is an > 0 with
B(x, ) C = . By the geometric Hahn-Banach theorem 1.1.5, taking the open set
U there equal to B(x, ), there is a X

and R with
Re (x) < Re (y), y C.
Now is continuous on X in the weak topology, hence Re is also, and so by basic
topology we have Re (y) for all y in the weak closure of C. Thus x is not in
the weak closure of C.
18 CHAPTER 1. TOPOLOGICAL VECTOR SPACES AND DUALITY
Corollary 1.3.5. If C is a nonempty set in a normed space and if x is in the weak
closure of C, then there is a sequence x
n
x in norm, with each x
n
a nite convex
combination of members of C.
Proof. Apply the last result to the convex hull of C.
1.4 The weak* topology
In this section, E, F, X and Y are Banach spaces. We write i
X
or for the canonical
embedding X X

. However we often suppress this map and simply consider X

as a subspace of X

. We can do this since it is a well known consequence of the

usual Hahn-Banach theorem that i
X
is isometric.
Denition 1.4.1. If X is the dual of a normed space E, and if J is the set of
functionals on X of the form f f(y), for y E, then the weak topology induced
on X by J is called the weak* topology or w*-topology.
Indeed, we consistently write w* for weak

. Some authors write (X

, X) for
the w*-topology on X

.
We say that a function f : X

is w*-continuous (resp. is a w*-homeomorphism)

if it is continuous (resp. is a homeomorphism) with respect to the w*-topologies on
X

and Y

.
The space X

is a LCTVS with the w*-topology. Also, we have:

Corollary 1.4.2. If X is a normed space, then
(1) A net (

) in X

converges to in the w*-topology if and only if

(x) (x)
for all x X.
(2) A linear functional : X

F is continuous with respect to the w*-topology

on X

i = x for some x X.
Proof. Just as in the proof of Corollary 1.3.3.
We omit the short, simple, and pretty proof of:
Theorem 1.4.3. (Alaoglu) If E is a normed space, then Ball(E

) is compact in
the w*-topology.
Recall the adjoint operator T

: F

, for T B(E, F), dened by T

() =
T. This has many well known properties which you will have met in Functional
Analysis, and which are easy exercises. For example, (ST)

= T

, |T| = |T

|,
and so on. Note that T

( x) =

T(x) for any x E.
Proposition 1.4.4. The operators from E

to F

which are w*-continuous (resp.

w*-homeomorphisms) are precisely those of the form S

for an S B(F, E) (resp.

for a surjective bicontinuous S B(F, E)).
1.4. THE WEAK* TOPOLOGY 19
Proof. Suppose that S B(F, E), and
t
weak* in E

. Then
S

(
t
)(y) =
t
(S(y)) (S(y)) = S

()(y),
for any y Y . That is, S

(
t
) S

() weak* in F

. Thus S

is w*-continuous.
If S is surjective bicontinuous then (S

)
1
= (S
1
)

is w*-continuous too.
If T : E

is w*-continuous, and if y F then y T is a w*-continuous

functional on E

, so that y T = x for an x E. Dene S(y) = x. Or in other

words, S = i
1
E
T

i
F
. This is clearly linear, bounded, and
S

()(y) = (S(y)) = (x) = x() = y(T()) = T()(y),

so that S

= T.
If, further, T is a w*-homeomorphism, then the same argument gives T
1
= R

where R : E F. We have (RS)

= S

= Id, so that RS = Id. Similarly

SR = Id, so that S is surjective bicontinuous.
A dual Banach space is a Banach space X which is linearly isometric to the dual
of another Banach space Y . We call Y a predual of X, and sometimes write Y = X

(although preduals may be very nonunique). Note that any predual Y though, may
be identied isometrically with a subspace Z of X

. Namely, if T : X Y

is
an isometric isomorphism, then T

: Y

is an isometric isomorphism too,

and we can identify Y with Z = T

Y ) X

. Note too that X

= Z

via the
canonical map, namely, R(x)() = (x), x X, X

. Indeed R = (
1
)

T,
a composition of two isometric isomorphisms, where : Y Z is the isometric
isomorphism above. The weak* topology on X induced by Y and T, is dened to
be the topology in which a net x
t
x in X i T(x
t
)(y) T(x)(y) for all y Y .
Note that this is equivalent to T

( y)(x
t
) T

( y)(x) for all y Y ; or equivalent

to (x
t
) (x) for all Z. Thus the weak topology on X induced by Z is the
weak* topology on X induced by Y and T, and this is also the weak* topology on
X induced by Z and R. We can therefore usually forget about Y and T and simply
work with the predual Z above. There is of course no guarantee that a dierent
predual of X should be identied with the same subspace Z of X

. Indeed if all
preduals of X give the same subspace Z of X

then we say that X has a strongly

unique predual. For example, any von Neumann algebra M (like

) have strongly
unique predualsin this case the space Z is the set of functionals : M C
which are normal in the sense that whenever (x
t
) is an increasing net in M with
supremum x then lim
t
(x
t
) = (x).
Corollary 1.4.5. The space Z discussed above is precisely the set of X

which
are continuous in the weak* topology on X induced by Y and T.
Proof. This follows from the discussion above and Proposition 1.3.1 (see also Corol-
lary 1.4.2 (2)).
Lemma 1.4.6. (Goldstine) If X is a normed space then Ball(X) is dense in
Ball(X

) in the w

-topology.
20 CHAPTER 1. TOPOLOGICAL VECTOR SPACES AND DUALITY
Proof. We apply the bipolar theorem to Ball(

X) X

, the latter equipped with

the weak* topology. In this topology the dual of X

is X

. The polar of Ball(

X)
corresponds to the set E of X

with [, x)[ = [(x)[ 1, for all x Ball(X).

Thus E = Ball(X

). The prepolar of E in X

is the set F of X

with
[()[ 1 for all E. Clearly F = Ball(X

). So by the bipolar theorem

Ball(X) is dense in Ball(X

).
The conclusion of Goldstines lemma of course implies that E is w*-dense in
E

.
Theorem 1.4.7. (KreinSmulian)
(1) Let E be a dual Banach space, and let F be a convex subset of E. Then F is
w*-closed in E if and only if F rBall(E) is closed in the w*-topology of E for
every r > 0.
(2) A linear bounded map u: E F between dual Banach spaces is w*-continuous
if and only if whenever x
t
x is a bounded net converging in the w*-topology
in E, then u(x
t
) u(x) in the w*-topology.
(3) Let E and F be as in (2), and u: E F a w*-continuous isometry. Then u
has w*-closed range, and u is a w*-w*-homeomorphism onto Ran(u).
Proof. Item (1) may be found in the standard texts. Note that the condition here
is saying precisely that whenever (x
t
) is a bounded net in F with w*-limit x, then
x F. It is easy to see that if F satises this condition then so does
1
r
(F x) for
any x F and r > 0. This is used in the proof from [24] say.
(2) We rst prove this for linear (scalar valued) functionals on E. Note that
the statement in this case is equivalent to saying that is w

-continuous if its
restriction to the unit ball is w

-continuous. This follows from by simply dividing

everything by a scalar large enough to ensure that the given bounded net now
lies in the unit ball. However, if the restriction to the unit ball is w

-continuous,
then Ker() Ball(E) is w*-closed. Hence Ker() is w*-closed by (1), and so is
w*-continuous by Lemma 1.2.4.
The result (2) as stated here follows from the scalar valued case. For suppose
that u(x
t
) u(x) w* for every bounded w*-convergent net x
t
x. If is a w*-
continuous functional on F, then (u(x
t
)) (u(x)). By the scalar functional
version of the present result, we see that u is w*-continuous on E. If z

z w*
in E then (u(z

)) (u(z)). Since this holds for all w*-continuous functionals

on F, we must have u(z

) u(z) w*. So u is w*-continuous.

(3) As in (2), a linear map is w

-continuous if its restriction to the unit ball is

w

-continuous. Next, it is easy to check using (1) that Ran(u) is w*-closed in F.

For by (1) it suces that Ball(Ran(u)) is w*-closed. However this ball is in fact w*-
compact, being the image of the w*-compact set Ball(E) under the w*-continuous
map u. Thus the restriction of u to Ball(E) takes w

-closed (and thus w

-compact)
sets to w

-compact (and thus w

-closed) sets in Ran(u). Thus the inverse of u

restricted to the ball is w

-continuous, and so u
1
is w

-continuous by (2).
If F is a subset of a normed space X then we write F

for the annihilator of F

in X

: that is, F

= X

: (F) = 0. This is a w*-closed linear subspace of

1.4. THE WEAK* TOPOLOGY 21
X

(Exercise). Indeed note that F

= (Span(F))

= Span(F)

. If W X

then
W

is the preannihilator of W in X: that is, W

= x X : (x) = 0 W.
This is a norm closed linear subspace of X (Exercise). Indeed note that W

=
(Span(W))

= Span(W)

.
Proposition 1.4.8. Suppose that X is a normed space.
(1) If F is a nonempty subset of X then (F

= Span(F).
(2) If W is a nonempty subset of X

then (W

= Span(W)
w
.
(3) The w*-closed linear subspaces of X

are precisely those of the form F

for
some linear subspace F X.
Proof. (1) and (2) both follow immediately from the bipolar theorem, and the rela-
tionship mentioned above about the connections between the perp and the polar.
For (3), we already remarked that any F

was a w*-closed linear subspace. Con-

versely, by (2), any w*-closed linear subspace W = (W

.
We recall from Functional Analysis that a 1-quotient map is a contractive linear
T : E F between normed spaces which takes the open unit ball of E onto the
open unit ball of F. These are exactly the maps whose canonically associated map
E/Ker(T) F is an isometric isomorphism (check this). (Note that this is implied
by, but is not the same as, that T takes the closed unit ball of E onto the closed
unit ball of F.) Indeed if F is a closed linear subspace of E, then the canonical
map E E/F is a 1-quotient map. The following are homework exercises: If
T : E F is an isometry then T

is a 1-quotient map (this uses the usual Hahn-

Banach theorem). If T : E F is a 1-quotient map then T

is an isometry. Hence
if T is an isometry then so is T

.
From facts in the last paragraph it follows that if F is a linear subspace of
E, then F

= E

/F

isometrically, and (E/F)

= F

isometrically. Indeed, if
i : F E is the inclusion, then i

: E

is a 1-quotient map, and so we have

F

= E

/Ker(i

) = E

/F

isometrically. The other relation follows similarly, or

from:
Corollary 1.4.9. If F is a w*-closed linear subspace of E

, for a normed space E,

then F

= (E/F

isometrically and w*-homeomorphically. In particular, F is a

dual Banach space.
Proof. Let N = F

. The quotient map q : E E/N dualizes to give a weak*

continuous isometry from (E/N)

. The range of this isometry is easily seen

to be contained in N

= F. On the other hand, if N

then by the factor

theorem there is a unique

(E/N)

with

q = . This says exactly that
q

maps onto N

. By the Krein-Smulian theorem, the isometry is also a weak*-

homeomorphism onto its range.
Also, in the notation of the last result, (F

)

= E

/(F

)

= E

/F isometri-
cally.
Lemma 1.4.10. Let F be a closed linear subspace of a Banach space E.
22 CHAPTER 1. TOPOLOGICAL VECTOR SPACES AND DUALITY
(1) As subsets of E

we have F
w
= F

.
(2) The second dual of the inclusion map F E is a w*-continuous isometry from
F

onto F

. Thus F

= F

isometrically and w*-homeomorphically, via

this canonical isometry.
(3) E F

= F.
Proof. (1) This follows from Proposition 1.4.8 (2) applied to W = i
E
(F).
(2) By principles stated above, this map is a w*-continuous isometry from F

onto a w*-closed subspace Y of E

, and by the Krein-Smulian theorem it is a w*-

homeomorphism onto this subspace. Since this homeomorphism takes the w*-dense
set

F onto i
E
(F), by topology we have that Y is the w*-closure of i
E
(F), which by
(1) equals F

.
(3) If x E with i
E
(x) F

, then x (F

= F. So i
E
(x) i
E
(F).
Lemma 1.4.11. If D is a set whose span is dense in a normed space E, then a
bounded net
t
w
in E

if and only if
t
(x) (x) for all x D.
Proof. Suppose that |
t
| C. Given x E and > 0, there exist n N and

k
C, d
k
D such that |x

n
k=1

k
d
k
| < . There exists t
0
such that t t
0
implies that [
t
(d
k
) (d
k
)[ < /(

n
k=1
[
k
[) for all k = 1, , n. Then
[
t
(x) (x)[ [
t
(x
n

k=1

k
d
k
)[ +[
n

k=1

k
(
t
(d
k
) (d
k
))[ +[(
n

k=1

k
d
k
x)[
(C +||) +
n

k=1
[
k
[[
t
(d
k
) (d
k
)[
D
for a constant D independent of . Thus
t
w*. The converse is trivial.
Lemma 1.4.12. Let u: E F

be a bounded linear map. Then there exists a

unique w

-continuous u: E

extending u. Moreover | u| = |u|.

Proof. Set u = i

F
u

where i
F
: F F

is the canonical isometry. Since T

is w

-continuous for any T B(E, F), we see that u is w

-continuous. Also, since

T

( x) =

T(x) for any x E, we have
u(i
E
(x)) = i

F
(i
F
(u(x))) = u(x),
since
i

F
(i
F
())(y) = i
F
()(i
F
(y)) = i
F
(y)() = (y), y F.
Thus u does extend u. Consequently, | u| |u|. On the other hand,
| u| = |i

F
u

| |i

F
||u

| = |i
F
||u| = |u|.
So | u| = |u|. The uniqueness follows from topology and the fact that E is w

-dense
in E

; by general topology we know there is a unique continuous extension of a

function dened on a dense subset of a topological space.
1.5. EXTREME POINTS 23
1.5 Extreme points
1.5.1 A face in a convex set C in a vector space X, is a nonempty convex subset
F of C with the property that if tx +(1 t)y F, for x, y C and 0 < t < 1, then
x F and y F. An extreme point is a one-point face. The set of extreme points
of C will be written ext(C).
Remarks. 1) We will use the easy fact that a face of a face of a convex set C
is a face of C (Exercise).
2) There are several other useful equivalent denitions of an extreme point (see
the Exercises at the end of the chapter).
3) If X is a normed space dierent from (0), then any extreme point x
0
of
Ball(X) must have norm 1. For suppose that 0 ,= x
0
Ball(X). Then 0 =
x
0
x
0
2
,
so that 0 is not an extreme point. Also, x
0
= t
x
0
x
0

+ (1 t)0, for t = |x
0
|.
We will also use the notion of the convex hull co(o) of a set o. This is the
smallest convex set containing o, and also equals the set of nite sums

n
k=1
t
k
x
k
with x
k
o, t
k
0,

n
k=1
t
k
= 1. The closed hull co(o) is the smallest closed
convex set containing o, and also equals the closure of co(o). This uses the simple
fact that the closure in a TVS of a convex set is convex (Exercise).
Theorem 1.5.2. (Krein-Milman) If K is a nonempty compact convex set in a
LCTVS X, then K is the closure in this topology of the convex hull of the set of
extreme points of K.
Proof. (From [24]) Claim 1: Every closed face C
0
of C contains an extreme point
of C. To prove this, let be the set of closed faces of C
0
, ordered by inclusion, a
partially ordered set. If ( is a chain in , let F
1
=
FC
F. By the nite intersection
property in topology, F
1
is nonempty. It is easy to check that F
1
is a face of C
0
.
Thus F
1
is a lower bound of (, and by Zorns lemma has a minimal fact F
0
say.
Claim 1 will be proved if we can prove:
Claim 2: F
0
has only one element. To see this, let X

. Since F
0
is compact,
Re() achieves a minimum value, s say, on F
0
. Let F

= x F
0
: Re((x)) = s.
This is nonempty and compact. Also, F

is a face of F
0
, and hence of C
0
. To see
this, let x, y F
0
and 0 < t < 1 with tx + (1 t)y F

. We have Re((x)) s,
Re((y)) s, and tRe((x))+(1t)Re((y)) = s. Thus Re((x)) = Re((y)) = s,
and so x, y F

. So F

is a face of F
0
.
Since F
0
is minimal, F
0
= F

, which says that Re() is constant on F

0
for every
X

. By the geometric Hahn-Banach theorem 1.1.6, we must have that F

0
has
only one element, proving Claims 2 and 1.
Finally, let B be the closed convex hull of the extreme points of C. If B ,= C, let
x C B. By Proposition 1.1.1, there is an open convex set U with x U B
c
.
By the geometric Hahn-Banach theorem 1.1.5, there exists a X

and a t R
with
Re((x)) < t Re((y)), y B.
Since C is compact, Re() achieves a minimum value, s say, on C. Since x C,
the last centered equation implies that s < t. As in the second last paragraph, let
24 CHAPTER 1. TOPOLOGICAL VECTOR SPACES AND DUALITY
F

= x F
0
: Re((x)) = s, a face in C. The last centered equation implies that
B F

= . Thus the face F

contains no extreme points of C, since the latter are

all in B. This contradicts Claim 1. Thus B = C.
Corollary 1.5.3. If K is a nonempty convex w*-closed and bounded subset of X

,
for a normed space X, then K is the closure in the w*-topology of the convex hull
of the set of extreme points of K.
Proof. If K is nonempty and bounded, then it is a subset of rBall(X

), for some
r > 0. The latter set is w*-compact by Alaoglus theorem. Thus if K is also
w*-closed then K is w*-compact. The result then follows from Theorem 1.5.2.
Corollary 1.5.4. A normed space whose unit ball has no extreme points, cannot
be a dual Banach space.
We now turn to some examples. First, the unit ball of c
0
has no extreme points.
More generally, we have
Proposition 1.5.5. The unit ball of C
0
(), for any locally compact Hausdor
space , has no extreme points unless is compact. If is compact then the
extreme points are the unimodular functions, that is the functions f with [f[ = 1
on .
Proof. Suppose that f is an extreme point of the unit ball of C
0
(). Then as
remarked above, |f| = 1. Suppose that [f(x
0
)[ < 1 for some x
0
. Choose K
with [f(x
0
)[ < K < 1, and an open set U containing x
0
with [f[ K < 1 on U.
By a well known variant of Urysohns lemma, there is a function g C
0
() with
g(x
0
) = 1, 0 g 1, and g = 0 outside of U. Then f is a convex combination:
f =
1
2
((f + (1 K)g) + (f (1 K)g)).
Notice that outside U, we have [f + (1 K)g[ = [f[ 1. On U, we have [f + (1
K)g[ [f[ + 1 K K + 1 K = 1. So f + (1 K)g Ball(C
0
()). Similarly,
f (1 K)g Ball(C
0
()). Thus f is not an extreme point of Ball(C
0
()), which
is a contradiction. Hence [f[ = 1 on . The latter, together with the fact that
f C
0
()), forces to be compact,
Finally, if f C
0
() with [f[ = 1 on , and if f =
1
2
(g + h) for g, h
Ball(C
0
()), then for any x we have
1
2
(g(x) + h(x)) = f(x). We are now
in the unit ball of F, and we must have f(x) = g(x) = h(x). So f = g = h, and so
f is extreme.
Remark. Since any L

(, ) is a commutative unital C

-algebra, it is -
isomorphic to C(K) for compact K. Hence the extreme points of its unit ball
correspond exactly to the elements of C(K) satisfying f

f = 1. Hence they are the

functions in L

(, ) satisfying f

f = 1, that is, [f[ = 1 a.e..

On the other hand, any norm 1 element f of L
p
(, ) is an extreme points of
its unit ball. Indeed if f =
1
2
(g +h), with g, h Ball(L
p
(, )) then 2 = |g +h|
p

|g|
p
+ |h|
p
, and so by the converse of the Minkowski inequality (see e.g. Taylor)
1.5. EXTREME POINTS 25
g = th for t > 0, so f =
1+t
2
h. Since |f|
p
= 1 we must have t = 1, so that
f = g = h.
Proposition 1.5.6. The unit ball of L
1
([0, 1]), has no extreme points.
Proof. Suppose that f L
1
([0, 1]) with |f|
1
=
_
1
0
[f[ dx = 1. By the intermediate
value theorem there is a c (0, 1) with
_
c
0
[f[ dx = 1/2. Let h(t) = 2f(t) on [0, c]
and h(t) = 0 on (0, 1]; and let g(t) = 0 on [0, c] and h(t) = 2f(t) on (0, 1]. Then
1
2
(g +h) = f, and g, h Ball(L
1
([0, 1])), and g ,= f. Thus f is not an extreme point
of Ball(L
1
([0, 1])).
Proposition 1.5.7. The extreme points of the positive part of the unit ball of a
unital C

-algebra A are exactly the orthogonal projections in A.

Proof. Suppose that x is an extreme point of the positive part of the unit ball of A.
Then x =
1
2
((2x x
2
) + x
2
), and by spectral theory |2x x
2
| 1 (since this can
be calculated with x replaced by the function z(t) = t in C(Sp(a))). Hence x = x
2
and so x is a projection.
For the converse, we may suppose that A is a -subalgebra of B(H). If P
is an orthogonal projection onto a subspace K H, and if S, T B(H) with
0 S, T I, and P =
1
2
(S +T), and if K with || = 1, then
1 = P, ) =
1
2
(S, ) +T, )).
This forces S, ) = T, ) = 1, which by the converse to the Cauchy-Schwarz
inequality implies that S = T = . On the other hand, if K then
0 = P, ) =
1
2
(S, ) +T, )),
which forces S, ) = T, ) = 0. Thus S
1
2
= 0 and S = 0. So S = 0 on K

, so
that S = P. Similarly, T = P, and so P is an extreme point.
Proposition 1.5.8. Let H be a Hilbert space. Any isometry on B(H) is an extreme
point of the unit ball of B(H). If H is a nite dimensional complex Hilbert space,
then the extreme points of the unit ball of B(H) are exactly the unitaries.
Proof. If R is an isometry, and R =
1
2
(S + T) for S, T Ball(B(H)) then for any
unit vector H we have
1 = R, R) =
1
2
(S, R) +S, R)).
This forces S, R) = 1, which by the converse to the Cauchy-Schwarz inequality
implies that S = R. So S = R and similarly T = R. So R is an extreme point of
the unit ball of B(H).
Now suppose that H is nite dimensional, and suppose that R is an extreme
point of the unit ball of B(H). We can write R = U[R[ by the polar decomposition
in linear algebra, with U a unitary. Then it is easy to see that [R[ must be an
26 CHAPTER 1. TOPOLOGICAL VECTOR SPACES AND DUALITY
extreme point of the unit ball of B(H) too. It follows that [R[ is an extreme point
of the positive part of the unit ball of B(H) too, so that by the previous result we
have that [R[ is a projection P. Then [R[ =
1
2
((P +P

) +(P P

)), which forces

P

= 0 and P = I. Thus R = U.
Remark. A characterization of the extreme points of the unit ball of any C

-
algebra may be found in e.g. [23, Proposition 1.4.7].
In what follows we will need to use regular Borel probability measures on a
compact Hausdor space K. We recall from measure theory that these are the
positive regular Borel measures such that (K) = 1, or equivalently, the signed
or complex regular Borel measures with (X) = [[(X) = 1. Or, via the Riesz
representation theorem for measures, these are exactly the states on C(K). It will
be useful to recall that states can be dened to be the functionals on C(K) with
(1) = || = 1 (see 2.2.1 in the C

-course). Recall that the Dirac point mass,

the probability measure supported on a single point x K, may be identied with
the state
x
, namely, point evaluation f f(x) for f C(K). By Gelfands theory
(Chapter 1 of the C

-algebras course), the set of such Dirac point masses is the

same as C(K)
#
, the nontrivial homomorphisms in C(K)

. Also, we showed there

that the map : K C(K)
#
taking x to
x
is a homeomorphism if we give C(K)
#
the (relative) weak* topology it inherits as a subspace of C(K)

.
Proposition 1.5.9. If K is a compact Hausdor space, then the set of extreme
points of the positive part of the unit ball of C(K)

equals the set of extreme points

of the regular Borel probability measures on K, and equals the set C(K)
#
of Dirac
point masses.
Proof. By the Markov-Riesz representation theorem for positive measures, the pos-
itive part C of the unit ball of C(K)

is exactly the set of positive regular Borel

measures on K with (K) 1. Any extreme point of this set is also clearly
an extreme point of the convex set of regular Borel probability measures. But the
converse of this is true too: if is an extreme point of the regular Borel probability
measures on K, and if =
1
2
( +) for , C, then 1 = (K) =
1
2
((K)+(K)).
This forces (K) = (K) = 1, so that , are probability measures, which in turn
forces = = .
Suppose that x K, and that
x
=
1
2
( + ) for states , on C(K). Then as
functionals on C(K), we have
1
2
(f)
1
2
((f) +(f)) =
x
(f) , f C(K)
+
.
If f C(K) with f(x) = 0, then we can write f =

4
k=1
i
k
f
k
with f
k
C(K)
+
and f
k
(x) = 0. Thus (f) =

4
k=1
i
k
(f
k
) = 0. Hence Ker(
x
) Ker(), which
implies by Proposition 1.2.1 that = D
x
for a scalar D. Evaluating at 1, the
identity of C(K), we have 1 = D, and so =
x
. Similarly, =
x
, and so
x
is
an extreme point of the set of states on C(K).
Conversely, let be an extreme point of the set of regular Borel probability
measures, viewed as a functional on C(K). Claim: (fg) = (f)(g) for any
1.5. EXTREME POINTS 27
f, g C(K), 0 f 1. To this end, suppose that 0 f 1, f C(K). Then
0 (f) 1. Let t = (f).
Case 1: 0 < t < 1. In this case dene (g) = t
1
(fg) and (g) = (1
t)
1
((1 f)g), for g C(K). We have (1) = 1, and
[(g)[ =
[(fg)[
t

([fg[)
t

|g|

(f)
t
= |g|

.
Thus corresponds to a probability measure on K. A similar argument shows that
(1) = 1 and [(g)[ |g|

, so that also corresponds to a probability measure

on K. Since = t+(1t), we must have = = . That is, (fg) = (f)(g)
for all g C(K).
Case 2: t = 0. In this case
[(fg)[ ([fg[) |g|

(f) = 0,
and so (fg) = (f)(g).
Case 3: t = 1. This case is similar to Case 2, since
[((1 f)g)[ ([(1 f)g[) |g|

(1 f) = 0,
and so (f)(g) = (g) = (fg).
We have now proved the Claim. Since the span of the f with 0 f 1 is all
of C(K), we deduce from the Claim that (fg) = (f)(g) for all f, g C(K).
Thus , viewed as a functional on C(K), is a homomorphism, and hence by Gelfand
theory (more particularly, a result from the rst chapter of the C

-algebra class)
we conclude that =
x
for some x K.
Theorem 1.5.10. (Banach-Stone) If T : C(K
1
) C(K
2
) is a surjective isometry,
where K
1
, K
2
are compact Hausdor spaces, then K
1
and K
2
are homeomorphic.
Indeed there exists a homeomorphism : K
2
K
1
, and a unimodular function
u C(K
2
) such that T(f)(x) = u(x)f((x)) for every x K
2
.
Proof. Let u = T(1), where 1 is the identity of C(K
1
). Since 1 is an extreme point
of Ball(C(K
1
)), we must have u = T(1) an extreme point of Ball(C(K
2
)). So u is
unimodular. Dene S(f) = u()T(f)(), then it is easy to see that S : C(K
1
)
C(K
2
) is a surjective isometry with S(1) = 1. Then S

: C(K
2
)

C(K
1
)

is
a weak* homeomorphic isometry, which restricts to a w*-homeomorphism from
the set of states on C(K
2
) onto the set of states on C(K
1
). Moreover, because S

preserves all the linear and convexity structure, as well as norms, it follows that
restricts further to a map from the set of extreme points of the state space of
C(K
2
) onto the set of extreme points of the state space of C(K
1
) Using Proposition
1.5.9, we conclude that restricts to a map from
x
: x K
2
onto
x
: x K
1
,
and we continue to write this restriction as . Now is a w*-homeomorphism. On
the other hand, the last two sets, with their w*-topology, are homeomorphic to
K
2
and K
1
respectively, as remarked earlier. Thus we obtain a homeomorphism
: K
2
K
1
such that the following diagram commutes:
28 CHAPTER 1. TOPOLOGICAL VECTOR SPACES AND DUALITY
C(K
2
)
#

C(K
1
)
#

K
2

K
1
Let f C(K
1
), x K
2
. Then
f((x)) = f(
1
((
x
))) = (
x
)(f),
since f(
1
()) = (f) for any C(K
1
)
#
. Thus
S(f)(x) =
x
(S(f)) = S

(
x
)(f) = (
x
)(f) = f((x)).
Finally, T(f)(x) = u(x)S(f)(x) = u(x)f((x)) for f C(K
1
), x K
2
.
Proposition 1.5.11. If K is a compact Hausdor space, then the set of extreme
points of the unit ball of C(K)

equals the set

x
: x K, T.
Proof. Suppose that is a Borel measure on K corresponding via the Riesz rep-
resentation theorem for measures to an extreme point in Ball(C(K)

). By the
Radon-Nikodym theorem d = fd[[, for a unimodular function f on K. It is easy
to see that [[ is an extreme point of Ball(C(K)

), and hence also of the positive

part of Ball(C(K)

). Thus [[ =
x
for some x K, and we must have = f(x)
x
.
Conversely, if
x
=
1
2
( +) for contractive functionals , on C(K), then
1 =
x
(1) =
1
2
((1) +(1)),
which forces , to be states. Thus = =
x
by Proposition 1.5.9. We leave the
rest as an exercise.
1.6 The classical Shilov boundary
1.6.1 We use the term function space for a closed subspace of C(), for a compact
Hausdor space . A unital function space is a closed subspace F of C() which
contains constant functions. More generally, we also use this term for a Banach
space E with a distinguished element e E, such that E is isometrically isomorphic
to an F as above, with the isomorphism taking e to 1. A map with the latter
property will be called unital. For a unital function space (E, e) we write S(E)
for the set of functionals E

with (e) = || = 1. This is a convex and

weak* closed (and hence weak* compact) subset of Ball(E

). Thus S(C()) is the

set of (regular Borel) probability measures on (see the discussion in and above
Proposition 1.5.9). We write

for the Dirac point mass at , and

for its
restriction to E, if E C(). A Choquet boundary point for E in is dened to
be a point satisfying one of the equivalent conditions in the next result:
Lemma 1.6.2. Suppose that E is a nonzero closed subspace of C(), for a compact
Hausdor space . Suppose that E contains constant functions and separates points
of . For a point the following are equivalent:
1.6. THE CLASSICAL SHILOV BOUNDARY 29
(i)

ext(Ball(E

)),
(ii)

ext(S(E)),
(iii)

is the only probability measure on extending

.
Moreover, if ext(S(E)), then =

for some (necessarily unique) .

Proof. The equivalence of (i) and (ii) follows as in the rst part of the proof of
Proposition 1.5.9. To prove the equivalence of (ii) and (iii), we will use the fact
from Proposition 1.5.9 that
ext(S(C())) =

: . (1.3)
Suppose that satises (iii), and that

=
1
2
( + ), with , S(E). By the
HahnBanach theorem, we may extend , to probability measures on . Then
the average of these measures, by hypothesis, equals

. By (1.3), both probability

measures coincide with

. So = =

. This yields (ii).

Let ext(S(E)), and consider the set P

of states of C() extending .

This is a nonempty subset of S(C()), by the HahnBanach theorem. Moreover,
P

is convex, since averages of states are states, and averages of extensions of

is an extension of . Also, P

is w*-closed, since if
t
weak* in S(E), and
if x E, then
t
(x) = (x) (x), so that P

. Thus P

has an extreme
point
0
by the KreinMilman theorem. If
0
=
1
2
(
1
+
2
) for
i
S(C()),
then if we restrict the last equation to E, and remember that ext(S(E)), we
see that
1
=
2
=
0
= on E. Thus
i
P

. Since
0
is an extreme point of
P

we deduce that
1
=
2
. Thus we have proved that
0
is an extreme point of
S(C()), and is hence by (1.3) it is a point mass

for some . The point

is necessarily unique by the fact that C() separates the points of . This proves
the nal assertion.
If satises (ii), and if we set =

and apply the argument in the last

paragraph to , then the above argument shows that any extreme point of P

is of the form

for some . Clearly = since E separates points of .

Thus by the KreinMilman theorem P

is a singleton set, namely

. This yields
(iii).
1.6.3 (The Choquet boundary) For any unital function space (E, e), we consider
the collection of pairs (, j) consisting of a compact Hausdor space , and a unital
linear isometry j : E C(), such that j(E) separates points of . We shall call
such a pair a function-extension of E.
We dene the Choquet boundary Ch(E) of (E, e) to be the set of extreme points
of S(E). Note that ext(S(E)) is not empty by the KreinMilman theorem. If
(, j) is a function-extension of E, then we dene the Choquet boundary in of
j(E) to be the set of w satisfying the equivalent conditions in the last lemma
for j(E) C(). As we see next, this last lemma shows that the set of Choquet
boundary points in does not depend essentially on the function-extension of E;
this set is homeomorphic to ext(S(E)) as topological spaces.
Lemma 1.6.4. If (, j) is any function extension of E, then the canonical map
g : S(E) taking to the functional

j on E, is a topological embedding.
30 CHAPTER 1. TOPOLOGICAL VECTOR SPACES AND DUALITY
Moreover g takes the Choquet boundary in of j(E) onto the set of extreme points
of S(E).
Proof. Let (, j) be any function extension of E, and dene g as above. Clearly
g is one-to-one since if g(w
1
) = g(w
2
) then j(x)(w
1
) = j(x)(w
2
) for every x E.
Since j(E) separates points we conclude that w
1
= w
2
. Next we claim that g is
continuous if one As in the proof of the last lemma there gives S(E) the weak*
topology. Indeed if w
t
w in , and if x E is xed, then j(x)(w
t
) j(x)(w).
That is, g(w
t
)(x) g(w)(x), and so g(w
t
) g(w) weak*. Since is compact, it
follows from topology that g is a topological embedding. Next, let be j regarded
as a map from E onto j(E). So is an isometric unital isomorphism, and hence

is a weak* homeomorphic isometric isomorphism. Since and

preserve all
the structure,

must restrict to a weak* homeomorphism from S(j(E)) onto

S(E), and further to a weak* homeomorphism between the extreme points of these
spaces. Thus
w
ext(S(j(E))) i

(
w
) ext(S(E)). Hence w is in the Choquet
boundary in of j(E) i

(
w
) ext(S(E)). But

(
w
) =
w
j = g(w), and so
we have shown that w is in the Choquet boundary in of j(E) i g(w) Ch(E).
1.6.5 (The Shilov boundary) As noted in Functional Analysis, any Banach space
E is linearly isometric to a closed subspace of C(), for some compact . Indeed,
this may be done so that E separates points (take = Ball(X

) with the w*-

topology. The question arises of nding the smallest which works, that is, what
is the minimal or essential topological space on which E can be supported in this
way? For unital function spaces E such a minimal does exist, and it is called the
Shilov boundary E of E.
To be a little more careful, we declare two function-extensions (, j) and (

, j

)
to be E-equivalent, if there exists a surjective homeomorphism :

such that
j

(x) = j(x), for all x E. This is an equivalence relation on the collection of

function-extensions of E. We dene a Shilov boundary for a unital function space
(E, e) to be a pair (E, i) having the universal property of the next theorem.
Theorem 1.6.6. (The boundary theorem for unital function spaces) Let (E, e)
be a unital function space. Then there exist a compact Hausdor space, written
E, and an isometric unital map i : E C(E) such that i(E) separates points of
E, with the following universal property: Given any function-extension (, j) of
E, there exists a (necessarily unique) topological embedding : E , such that
j(x) = i(x) for all x E.
Proof. Dene E to be the weak* closure of Ch(E) in S(E). The canonical map
from i : E C(S(E)), which takes an x E to the function on S(E) taking
S(E) to (x), is a unital contraction. In fact it is an isometry, since if E was a
concrete unital function space in C(K) say, then
z
: z K is a subset of S(E)
and for any x E there is a z K with |x| = [x(z)[ = [
z
(x))[ |i(x)|. By the
KreinMilman theorem, i restricts further to an isometry i : E C(E), since if
[(x)[ t < 1 for all ext(S(E)), then it is easy to see that |x| < 1). Clearly
(E, i) is a function-extension of E. We will show that this function-extension has
the universal property of the theorem. Let (, j) be any function extension of E.
1.6. THE CLASSICAL SHILOV BOUNDARY 31
As in the proof of the last lemma there exists a topological embedding g : S(E)
taking to the functional

j on E. The restriction of g
1
to the w*-closure
of Ch(E) is the desired embedding . We need to prove that
j(x)(()) = i(x)() = (x), x E,
for all Ch(E)
w
. By topology, it is enough to prove this for Ch(E). By the
second last lemma, such =

j. Thus () = (

j) = by denition of
and g above. So j(x)(()) = j(x)() = (x), and we are done.
1.6.7 We say that Ch(E)
w
is the Shilov boundary of E. More generally, if (, j) is
a function extension of E with the universal property in the theorem, then we also
say that is the Shilov boundary of E, and write as E. If (, j) is any function
extension of E, and if

is as in the third remark after the boundary theorem, then

we say that

is the Shilov boundary of E in .

1.6.8 (Remarks on the universal property) We will make ve important obser-
vations that ow from the universal property in 1.6.6. First, note that any space
(E, i) having this universal property, has the property that there is no proper
closed subset of E such that i()
|
is still an isometry on E. This may be seen
by letting j = i()
|
. If this were an isometry on E then by the universal property
there exists a topological embedding : E with i(x)((w)) = i(x)(w) for all
w E, x E. Since i(E) separates points we have (w) = w for all w E, so
that = E.
Second, we remark that it is a simple exercise to see that a function-extension
(, j) of E is E-equivalent to (E, i) if and only if (, j) also has the universal prop-
erty of the theorem. Thus the set of function-extensions having the universal prop-
erty of the theorem, is one equivalence class of the relation we called E-equivalence
above. Third, we point out that if (, j) is any function-extension of E, and if
is the associated topological embedding : E coming from the universal
property in the theorem, then letting

= (E) and j

(x) = j(x)
|
, we have that
(

, j

) is E-equivalent to (E, i). Hence by the second remark, (

, j

) also has the

universal property of the theorem. Thus every function-extension of E contains a
function-extension with the universal property of the theorem. Putting these three
remarks together we deduce our fourth remark, namely that the Shilov boundary
of E may be taken to be any function-extension (, j) of E with the property that
there is no closed subset

of such that j()

|
is still an isometry on E. Our
fth remark is that nonetheless there is a canonical choice for the Shilov boundary
(E, i) (that is, a canonical element of the equivalence class). Namely, the closure
of Ch(E) in E

with respect to the w

-topology, as in the beginning of the proof

above.
1.6.9 (Representation on the maximal ideal space) For any unital subalgebra
A C(), consider the Gelfand representation from Chapter 1 of the C

-algebras
course, which is an unital contractive homomorphism of A into C(M
A
), where M
A
is the maximal ideal space A
#
, or space of characters. Note that is isometric
32 CHAPTER 1. TOPOLOGICAL VECTOR SPACES AND DUALITY
because
|f| = sup[
x
(f)[ : x sup[(f)[ : M
A
= |(f)| |f|, f A.
So (M
A
, ) is a function extension of A. Clearly A separates points of M
A
. By the
third (or fth) observation after the boundary theorem, the Choquet and Shilov
boundaries of A may be thought of as subspaces of M
A
. Indeed ext(S(A)) M
A
.
1.6.10 (Examples)
(1) We denote as usual by T the compact space of complex numbers with
modulus equal to one, and we write z = (z
1
, . . . , z
n
) for an element of T
n
. Let E =

1
n+1
, with its canonical basis (e
k
)
n
k=0
. Consider the linear function i : E C(T
n
)
which takes e
0
to 1, and otherwise takes e
k
to the kth coordinate function z z
k
on T
n
. This is an isometry such that i(E) separates points of T
n
. However there
is no proper closed subset of T
n
so that i()
|
is still isometric. Indeed if is a
closed subset of T
n
and if = (
k
) T
n
, dene x = (1,
1
, . . . ,
n
) E. Then
|x| = n +1. Claim: |i(x)
|
| < n +1, so that i()
|
is not an isometry. Suppose to
the contrary that |i(x)
|
| = n + 1. Then by compactness there exists z with
[i(x)(z)[ = n + 1. That is, 1 =
1
n+1
[1 +z
1

1
+. . . +z
n

n
[. Since any point on T is
an extreme point, this is only possible if z
1
=
1
, , z
n
=
n
, that is, z = . This
proves the claim. By the fourth remark after the boundary theorem, this shows
that E = T
n
.
(2) Another important example is the disc algebra A(D) of complex functions
that are analytic in the open unit disk D, and continuous on its closure the closed
unit disk. We have A(D) C(D), but there is an obvious map A(D) C(T),
namely the restriction map taking f to f
|T
. By the maximum modulus theorem
from complex variables, this map is an isometry, and is therefore a function exten-
sion of A(D). (To see this, suppose that [z
0
[ < 1 with [f(z
0
)[ > [f(z)[ for every
z T. By the maximum modulus theorem we can nd a sequence z
n
: n n
0

with [z
n
[ = 1 1/n and [f(z
n
)[ > [f(z
0
)[. If z is a w*-limit point of the sequence,
then z T and [f(z)[ [f(z
0
)[, a contradiction.) We claim that this function exten-
sion is the Shilov boundary. That is, A(D) = T. To see this, by the fourth remark
after the boundary theorem it suces to show that there is no proper compact sub-
set of T which supports A(D) isometrically. Suppose that is a proper compact
subset of T, and that z
0
T . Let f(z) = (z +z
0
)/2 Ball(A(D), and note that
[f(z)[ < 1 for every z . By compactness, sup[f(z)[ : z < 1 = |f|
A(D)
.
So f f
|
is not an isometry.
This example suggests why the Shilov boundary is a boundary.
Exercises.
(1) In this question X and Y are normed spaces.
(i) Show that the map B(X, Y ) to B(Y

, X

) which takes T to T

, is isometric
and linear. Show that it is also onto, if X and Y are reexive.
(ii) Show that if T is an isometric isomorphism of X onto Y , then T

is an
isometric isomorphism of Y

onto X

.
1.6. THE CLASSICAL SHILOV BOUNDARY 33
(iii) Show that if T is an isometry of X into Y , then T

is a 1-quotient map.
(iv) Show that if T is a 1-quotient map then T

is an isometry of Y

into X

.
(v) If T B(X, Y ), and x X, calculate T

( x). Deduce that if T

is
isometric then T is isometric.
(vi) If T B(X, Y ), and T

is a 1-quotient map use (v) and (iv) to show that

T is an isometry.
(vii) If X, Y are complete, and if T B(X, Y ), and T

is an isometry, then T is
a 1-quotient map.
2. Show that a Banach space X is separable if and only if Ball(X

) with the
weak* topology is metrizable [Hint: For the () direction, mimic the proof of
Alaoglus theorem, but with a countable product of copies of the disk].
3. Suppose that X is a Banach space. Prove that X is a dual Banach space i
there is a contractive projection P : X

with range

X and weak*
closed kernel. Prove that if there exists only one such projection P then any
two preduals of X are linearly isometric to each other. (The converse is still an
open problem.)
4. Prove that F has a unique TVS topology.
5. Show that if K is a convex set in a vector space V , then e is an extreme point
of K i whenever e =
1
2
(x + y) for x, y K then x = y = e. Show that this is
also equivalent to saying that if e =

n
k=1
t
k
x
k
for x
k
K and numbers t
k
> 0
such that

n
k=1
t
k
= 1, then all the x
k
= e. Show that this is also equivalent to
saying that K e is convex. Show that this is also equivalent to saying that if
6. Show that if X is a dual Banach space, then there exists a set I and a linear
isometry T : X

(I) whose range is weak*-closed, and such that T is also

a homeomorphism with respect to the weak* topologies.
34 CHAPTER 1. TOPOLOGICAL VECTOR SPACES AND DUALITY
Chapter 2
Operator spaces
2.1 Basic facts, examples, and constructions
2.1.1 (Matrix notation) Fix m, n N. If X is a vector space, then so is M
m,n
(X),
the set of m n matrices with entries in X. This may also be thought of as the
algebraic tensor product M
m,n
X, where M
m,n
= M
m,n
(C). We write I
n
for the
identity matrix of M
n
= M
n,n
. We write M
n
(X) = M
n,n
(X), C
n
(X) = M
n,1
(X)
and R
n
(X) = M
1,n
(X).
If x is a matrix, then x
ij
or x
i,j
denotes the i-j entry of x, and we write x as
[x
ij
] or [x
i,j
]
i,j
. We write (E
ij
)
ij
for the usual (matrix unit) basis of M
m,n
(we
allow m, n innite here too). We write A A
t
for the transpose on M
m,n
, or more
generally on M
m,n
(X). We will sometimes meet large matrices with row and col-
umn indexing that is sometimes cumbersome. For example, a matrix [a
(i,k,p),(j,l,q)
]
is indexed on rows by (i, k, p) and on columns by (j, l, q), and may also be written
as [a
(i,k,p),(j,l,q)
]
(i,k,p),(j,l,q)
if additional clarity is needed. To illustrate this nota-
tion, the reader may want to write down the matrix [
i,

kj
]
(i,k),(j,)
. Here
i,j
is
Kroneckers delta.
2.1.2 (Norms of matrices with operator entries) We recall from 3.4.3 of the C

-
course that M
n
(B(H)) is a C

-algebra for any Hilbert space H, with the norm

that it gets via the -isomorphism M
n
(B(H))

= B(H
(n)
). Reviewing, recall that
H
(n)
is the Hilbert space direct sum of m copies of H, and the norm of a vector
= (
k
) there is (

n
k=1
|
k
|
2
)
1
2
. As is explained clearly in 3.4.3 of the C

-course,
this -isomorphism is the map taking a matrix [T
ij
] M
n
(B(H)) to the operator
from H
(n)
to H
(n)
:
_

_
T
11
T
12
T
1n
T
21
T
22
T
2n


T
n1
T
n2
T
nn
_

_
_

n
_

_
=
_

k
T
1k

k
T
2k

k
T
nk

k
_

_
.
35
36 CHAPTER 2. OPERATOR SPACES
Thus M
n
(B(H)) is a C

-algebra, and [T
ij
] M
n
(B(H)) has a natural norm:
|[T
ij
]|
n
= sup|[T
ij
]

| :

H
(n)
, |

| 1.
Using the principle that || = sup[, )[ : Ball(K) in any Hilbert space K,
we deduce that
|[T
ij
]|
n
= sup[

i,j
< T
ij

j
,
i
> [ :

= (
i
),

= (
i
) Ball(H
(n)
).
We can also view M
n
(B(H)) as the spatial tensor product M
n
B(H) (see 3.4.3 of
the C

-course).
Similar identities hold for rectangular matrices. Indeed if m, n N, and K, H
are Hilbert spaces, then we always assign M
m,n
(B(K, H)) the norm (written ||
m,n
)
ensuring that
M
m,n
(B(K, H))

= B(K
(n)
, H
(m)
) isometrically (2.1)
via the natural algebraic isomorphism.
2.1.3 (Completely bounded maps) Suppose that X and Y are vector spaces and
that u: X Y is a linear map. For a positive integer n, we write u
n
for the
associated map [x
ij
] [u(x
ij
)] from M
n
(X) to M
n
(Y ). This is often called the
(nth) amplication of u, and may also be thought of as the map I
M
n
u on M
n
X.
Similarly one may dene u
m,n
: M
m,n
(X) M
m,n
(Y ). If each matrix space M
n
(X)
and M
n
(Y ) has a given norm | |
n
, and if u
n
is an isometry for all n N, then
we say that u is completely isometric, or is a complete isometry. Similarly, u is
completely contractive (resp. is a complete quotient map) if each u
n
is a contraction
(resp. takes the open ball of M
n
(X) onto the open ball of M
n
(Y )). A map u is
completely bounded if
|u|
cb
def
= sup
_
|[u(x
ij
)]|
n
: |[x
ij
]|
n
1, all n N
_
< .
As in the Banach space case, it is easy to prove that |u +v|
cb
|u|
cb
+|v|
cb
,
and |u|
cb
= [[|u|
cb
for a scalar , and so on. Compositions of completely
bounded maps are completely bounded, and one has the expected relation |u
v|
cb
|u|
cb
|v|
cb
. If u: X Y is a completely bounded linear bijection, and if its
inverse is completely bounded too, then we say that u is a complete isomorphism.
In this case, we say that X and Y are completely isomorphic and we write X Y .
If, further, u and u
1
are completely contractive, then just as in the Banach space
case they are complete isometries.
2.1.4 (Operator spaces) A concrete operator space is a (usually closed) linear sub-
space X of B(K, H), for Hilbert spaces H, K (indeed the case H = K usually
suces, via the canonical inclusion B(K, H) B(H K)). However we will want
to keep track too of the norm | |
m,n
that M
m,n
(X) inherits from M
m,n
(B(K, H)),
for all m, n N. We write | |
n
for | |
n,n
; indeed when there is no danger of
confusion, we simply write |[x
ij
]| for |[x
ij
]|
n
.
2.1. BASIC FACTS, EXAMPLES, AND CONSTRUCTIONS 37
An abstract operator space is a pair (X, | |
n

n1
), consisting of a vector space
X, and a norm on M
n
(X) for all n N, such that there exists a linear complete
isometry u: X B(K, H). In this case we call the sequence | |
n

n
an operator
space structure on the vector space X. An operator space structure on a normed
space (X, | |) will usually mean a sequence of matrix norms as above, but with
| | = | |
1
.
Clearly subspaces of operator spaces are again operator spaces. We often identify
two operator spaces X and Y if they are completely isometrically isomorphic. In
this case we often write X

= Y completely isometrically, or say X

= Y as operator
spaces. Sometimes we simply write X = Y .
2.1.5 (C

-algebras) We remarked in the C

-course that by 1.3.2 of that course

and the C

-identity there can be at most one norm on a -algebra for which that
-algebra is a C

-algebra. Thus, as we said in 3.4.3 of that course, if A is a

C

-algebra then the -algebra M

n
(A) has a unique norm with respect to which
it is a C

-algebra. With respect to these matrix norms, A is an operator space.

This may be seen by noting that if A is a closed -subalgebra of B(H) then A
is a concrete operator space, and the operator space structure (i.e. matrix norms)
imputed to M
n
(A) from B(H
(n)
), also makes M
n
(A) a C

-algebra as we said in
3.4.3 of the C

-course. We call this the canonical operator space structure on a

C

-algebra. If the C

-algebra A is commutative, with A = C

0
() for a locally
compact space , and then these matrix norms are determined via the canonical
isomorphism M
n
(C
0
()) = C
0
(; M
n
). Explicitly, if [f
ij
] M
n
(C
0
()), then:
|[f
ij
]|
n
= sup
t
_
_
[f
ij
(t)]
_
_
. (2.2)
To see this, note that by the above one only needs to verify that (2.2) does indeed
dene a C

-norm on M
n
(C
0
()). Clearly the right hand side of (2.2) is a nite
number, since |[f
ij
(t)| n
2
max
i,j
[f
ij
(t)[, and each of the functions f
ij
is bounded
on . Also, it easy to check that (2.2) does dene a norm. To see that (2.2) is a
Banach algebra, note that
|[f
i,j
][g
i,j
]| = sup|[f
i,j
(t)g
i,j
(t)]| : t
sup|[f
i,j
(t)]| : t sup|[g
i,j
(t)]| : t
= |[f
i,j
]||[g
i,j
]|.
So M
n
(C
0
()) is a Banach algebra. We check the C

-identity:
|[f
i,j
][f
i,j
]|
n
= sup|[f
i,j
(t)][f
i,j
(t)]| : t
= sup|[f
i,j
(t)]

[f
i,j
(t)]| : t
= sup|[f
i,j
(t)]|
2
: t
= sup|[f
i,j
(t)]|
M
n

2
.
Thus M
n
(C
0
()) is a C

-algebra.
38 CHAPTER 2. OPERATOR SPACES
Proposition 2.1.6. For a homomorphism : A B between C

-algebras, the
following are equivalent: (i) is contractive, (ii) is completely contractive, and
(iii) is a -homomorphism. If these hold, then (A) is closed, and is a complete
quotient map onto (A); moreover is one-to-one if and only if it is completely
isometric.
Proof. (i) (iii) We proved this in the case that A, B and are unital in 2.3.11
in the C

-course, and it follows from 2.4.3 which we didnt prove completely. If A

is unital, then p = (1) is an orthogonal projection, By replacing B by pBp we
make unital and can appeal to 2.3.11 in the C

-course to see that is -linear.

In the nonunital case we will use some basic things about second duals (see Section
4.1 in the C

-course). Consider

: A

. By routine arguments

is still
a contractive homomorphism, and since A

is unital we see that

is -linear.
Hence =

|A
is -linear.
(ii) (i) Obvious.
(iii) (ii) Note that
n
is a -homomorphism, and so contractive by 2.1.7 in
the C

-course. So is completely contractive.

Clearly if is completely isometric it is one-to-one. Conversely, if is one-to-one
then
n
is a one-to-one -homomorphism and so isometric by 2.1.7 in the C

-course.
Thus is completely isometric.
By 2.4.2 in the C

-course, is a 1-quotient map onto its (closed) range. Simi-

larly,
n
is a 1-quotient map, so that is a complete quotient map.
2.1.7 (Maps into a commutative C

-algebra) If [a
ij
] M
n
then
|[a
ij
]| = sup
_

ij
a
ij
z
j
w
i

: z = [z
j
], w = [w
i
] Ball(
2
n
)
_
.
Moreover, if a
ij
B(H), for a Hilbert space H, then
|[a
ij
]| sup
_
|

ij
a
ij
z
j
w
i
| : z = [z
j
], w = [w
i
] Ball(
2
n
)
_
.
Indeed, if one uses the fact that |T| = sup [T, )[ : , Ball(H), for any
T B(H), then one sees that the right side of the last centered formula is
sup

_
[a
i,j
]
_

_
z
1

z
2

.
.
.
z
n

_
,
_

_
w
1

w
2

.
.
.
w
n

_
_

: , Ball(H),

z ,

w Ball(l
2
n
),
which is dominated by |[a
i,j
]|.
Using these formulae, it is easy to see that any continuous linear functional
: X C on an operator space X is completely bounded, with || = ||
cb
. We
2.1. BASIC FACTS, EXAMPLES, AND CONSTRUCTIONS 39
have
|[(x
i,j
)]|
n
= sup
_
_
_

i,j
(x
i,j
)z
j
w
i

:

z ,

w Ball(l
2
n
)
_
_
_
= sup
_
_
_

_
_

i,j
x
i,j
z
j
w
i
_
_

:

z ,

w Ball(l
2
n
)
_
_
_
|| sup
_
_
_
_
_
_
_
_
_

i,j
x
i,j
z
j
w
i
_
_
_
_
_
_
:

z ,

w Ball(l
2
n
)
_
_
_
|| |[x
i,j
]|.
Hence |
n
| ||, and so ||
cb
= ||.
Next we claim that |u| = |u|
cb
for any bounded linear map u from an operator
space into a commutative C

-algebra. We can assume that the commutative C

-
algebra is C
0
(), for a locally compact . For xed w let
w
X

be dened
by
w
(x) = u(x)(w). Note that [
w
(x)[ = [u(x)(w)[ |u(x)| |u||x|, if x E.
Thus |
w
| |u|. We then have
|[u(x
i,j
)(w)]| = |[
w
(x
i,j
)]| |
w
||[x
i,j
]| |u||[x
i,j
]|
Thus by equation (2.2), it follows that |[u
n
(x
i,j
)]| |u||[x
i,j
]|, and so |u|
n
|u|.
Since this is true for all n N we have |u|
cb
= |u|.
2.1.8 (Properties of matrix norms) If K, H are Hilbert spaces, and if X is a
subspace of B(K, H), then there are certain well-known properties satised by the
matrix norms | |
m,n
described in 2.1.2. Most important for us are the following
two.
(R1) |x|
n
|||x|
n
||, for all n N and all , M
n
, and x M
n
(X)
(where multiplication of an element of M
n
(X) by an element of M
n
is dened
in the obvious way).
(R2) For all x M
m
(X) and y M
n
(X), we have

_
x 0
0 y
_

m+n
= max|x|
m
, |y|
n
.
We often write x y for the matrix in (R2).
To see (R1), dene
=
_

11
I
H

12
I
H
. . .
1n
I
H
.
.
.

n1
I
H

n2
I
H
. . .
nn
I
H
_

_ = I
H
M
n
B(H) = M
n
(B(H)).
Similarly dene

= I
H
M
n
(B(H)). Note that | | = || since the -
homomorphism M
n
M
n
(B(H)) taking is one-to-one and hence is a
40 CHAPTER 2. OPERATOR SPACES
(complete) isometry. Also x = x

for all x M
n
(B(H)). Thus
|x| = | x

| | ||x||

| = |||x|||
since M
n
(B(H)) is a Banach algebra. This proves (R1).
To prove (R2), note that if a = [I
n
: 0], b = [I
n
: 0]
t
, then x = a(xy)b (using the
notation after the statement of (R2)). If we let a = [I
n
I
H
: 0] M
n,n+m
(B(H))
(that is, a is an n (n + m) matrix consisting of all zero entries except for an
I
H
in the i-i entry for i = 1, , n), and if

b = a
t
, then as in the proof of (R1),
x = a(x y)

b. Hence
|x|
n
= | a(x y)

b|
n
| a||x y|
m+n
|

b|.
Note that | a| = | a a

|
1
2
= |I|
1
2
M
n
(B(H))
= 1, and similarly |

b| = |

b|
1
2
= 1. Thus
|x|
n
|x y|
m+n
. Similarly, |y|
n
|x y|
m+n
, so that max|x|
n
, |y|
m

|x y|
m+n
.
For the other direction, let H
n
, H
m
, then
|(x y)
_

_
|
2
= |
_
x
y
_
|
2
= |x|
2
+|y|
2
|x|
2
||
2
+|y|
2
||
2
.
Clearly this is dominated by max|x|, |y|
2
(||
2
+ ||
2
). Thus we deduce that
|x y| max|x|, |y[. This proves (R2).
It follows from (R1) that switching rows (or columns) of a matrix of operators
does not change its norm. Such switching is equivalent to multiplying by a per-
mutation matrix U, namely a matrix which is all zeroes except for one 1 in each
row and each column. Such a matrix has norm 1, being unitary, and so
|Ux| |x| |U

Ux| |Ux|.
Adding (or dropping) rows of zeros or columns of zeros does not change the norm
of a matrix of operators. To see this, note that by the last paragraph we can
suppose all the zero rows (resp. columns) are at the bottom (resp. right) of the
matrix. But then it is elementary to see that the norm is unchanged if we remove
those zero rows or columns. For example

_
x
0
_

m,n
= sup|x| : Ball(H
(n)
) = |x|.
By the principle in the last paragraph, we really only need to specify the norms
for square matrices, that is, the case m = n above, since M
m,n
(X) may be viewed
as a subspace of M
k
(X) where k = maxm, n.
If X is an operator space then the canonical algebraic isomorphisms
M
n
(M
m
(X))

= M
m
(M
n
(X))

= M
mn
(X) (2.3)
are isometric. To see this, we can assume that X is a C

-algebra, and then no-

tice that these three canonical algebraic isomorphisms are -isomorphisms, hence
completely isometric.
2.1. BASIC FACTS, EXAMPLES, AND CONSTRUCTIONS 41
If X is an operator space then so is M
n
(X), the latter with the operator space
structure for which the canonical isomorphism M
m
(M
n
(X))

= M
mn
(X) becomes
an isometry. One way to see this
1
is to note that if X is a subspace of a C

-
algebra A, then M
n
(X) M
n
(A), and the latter is a C

-algebra and hence is

an operator space whose matrix norms are the ones making M
m
(M
n
(A)) a C

-
algebra. Then M
n
(X), with the inherited matrix norms, is an operator space.
However, M
m
(M
n
(A)) is -isomorphic to M
mn
(A), and hence the norm making
M
m
(M
n
(A)) a C

-algebra is exactly the one coming from the C

-algebra M
mn
(A)
via the canonical -isomorphism M
m
(M
n
(A))

= M
mn
(A). Restricting the latter
isomorphism to M
m
(M
n
(X)) gives the desired assertion.
If T B(K, H), and if z
ij
C then |[z
ij
T]| = |T||[z
ij
]|. We leave this as an
exercise.
It is easy to see from the above that max
i,j
|x
ij
| |[x
ij
]|
n
nmax
i,j
|x
ij
|.
It follows from this that a sequence (x
k
) of matrices in M
n
(X) converges i the
i-j entry of x
k
converges as k to the same entry of x, for all i, j.
2.1.9 (An operator that is not completely bounded) The canonical example of a
map that is not completely bounded is the transpose map (x) = x
t
on K = K(
2
)
thought of as innite matrices (via the prescription x [x

e
j
,

e
i
)]). Note that
is an isometric linear -antiisomorphism. Indeed if

z = (z
i
),

w = (w
i
) Ball(
2
),
then

i,j=1
x
ji
z
j
w
i

i,j=1
x
ij
w
i
z
j

|[x
ij
]|.
This says that is a contraction, and by symmetry (since
1
= ) it is an isometry.
If E
ij
is the usual basis for M
n
and if : M
n
K is the top left corner
embedding, i.e. (x) = x 0, then is a one-to-one -homomorphism, and hence
is a complete isometry. If x
n
= [(E
ji
] M
n
(K) then
|x
n
|
n
= |[(E
ji
)]| = |[E
ji
]| = 1,
as can be seen by switching rows and columns of the matrix [E
ji
] to make it an
identity matrix. On the other hand,
n
(x
n
) = [(E
ij
)], which has the same norm
as [E
ij
]. Erasing zero rows and columns, the latter becomes an n n matrix with
all entries 1. By the C

-identity the latter has the same norm as x

x = n, where
x is a column of n entries each equal to 1. Thus we have |
n
(x
n
)|
n
= n, and so
|
n
| n. Hence is not completely bounded.
Conditions (R1) and (R2) in 2.1.8 are often called Ruans axioms. Ruans theo-
rem asserts that (R1) and (R2) characterize operator space structures on a vector
space. This result is fundamental to our subject in many ways. At the most pedes-
trian level, it is used frequently to check that certain abstract constructions with
operator spaces remain operator spaces. At a more sophisticated level, it is the
foundational and unifying principle of operator space theory. We now proceed to
Eros and Ruans proof of Ruans theorem. We omit the proof of the rst lemma,
which may be found on [p. 30,ERbook].
1
Another way to see this is as in 2.1.23 (7).
42 CHAPTER 2. OPERATOR SPACES
Lemma 2.1.10. If X is a vector space, and if | |
n
is a norm on M
n
(X), for each
n N, satisfying (R1) and (R2), and if F Ball(M
n
(X)

), then there are states

, on M
n
with
[F(x)[ (

)
1
2
|x| (

)
1
2
, , M
n
, x M
n
(X).
Lemma 2.1.11. If (X, | |
n
) are as in the last lemma, if F Ball(M
n
(X)

),
and if H is the Hilbert space
2
n
, then there exist vectors , Ball(H
(n)
), and a
completely contractive u : X B(H)

= M
n
such that F = u
n
(), ).
Proof. By the last lemma there are states , on M
n
with [F(

x)[ (

)
1
2
|x|(

)
1
2
for , M
n
. By part of the proof of 4.1.26 in the C

-course, in the case

M = M
n
= B(
2
n
) (although the proof is much simpler in this case), we have
(x) =
n

k=1
x
k
,
k
) = (x I
n
), ), x M
n
,
where = (
k
) Ball(H
(n)
). It follows that for any M
n
we have
(

) = (

I
n
), ) = ( I
n
), ( I
n
)) = |( I
n
)|
2
.
Similarly, (

)
1
2
= |( I
n
)| for some Ball(H
(n)
). The inequality in the
rst line of the proof then reads
[F(

x)[ |x| |( I
n
)| |( I
n
)|, , M
n
.
Let E = (M
n
I
n
) and K = (M
n
I
n
), subspaces of C
n
2
. Fix x M
n
(X) for
a moment and dene g : E K C by g(( I
n
), ( I
n
)) = F(

x), for
, M
n
. Thus
[g(( I
n
), ( I
n
))[ |x| |( I
n
)| |( I
n
)|, , M
n
.
It is easy to see from this that g is a well-dened and a contractive sesquilinear form
on E K. By Hilbert space theory there exists an operator in B(E, K), which we
shall write as T(x), with |T(x)| |x|, and
T(x)( I
n
), ( I
n
)) = F(

x), x M
n
(X), , M
n
.
It is easy to see that T is linear. Let P be the projection from C
n
2
onto E. Since
E is invariant under M
n
I
n
, it follows from a fact in 4.1.11 in the C

-course that
P (M
n
I
n
)

. Let R = T()P B(C

n
2
)

= M
n
2. Then R B(M
n
(X), M
n
2) is a
linear contraction, since |R(x)| |T(x)||P| |x|. We have
R(x), ) = T(x)(I
n
I
n
), (I
n
I
n
)) = F(I
n
xI
n
) = F(x), x M
n
(X).
Notice next that if , , M
n
then
T(x)( I
n
), ( I
n
)) = F(

x) = T(x)( I
n
), ( I
n
)).
2.1. BASIC FACTS, EXAMPLES, AND CONSTRUCTIONS 43
That is,
T(x)h, k) = T(x)( I
n
)h, k), h E, k K,
which means that T(x) = T(x)( I
n
). Hence
R(x) = T(x)P = T(x)( I
n
)P = R(x)( I
n
), M
n
.
A similar argument shows that R(x) = ( I
n
)R(x) for M
n
. It is a simple
linear algebra exercise that if S : M
n
(Y ) M
n
(Z) is a linear map, where Y, Z
are vector spaces, then S = u
n
for a linear u : Y Z i S(x) = S(x)
for all x M
n
(Y ) and , M
n
. Hence R = u
n
for some u : X M
n
with
|u
n
| = |R| 1. By Exercise 5 at the end of this section, this forces |u
m
| 1 for
all m n, so that u is completely contractive.
Thus u
n
(x), ) = R(x), ) = F(x) for all x M
n
(X).
Corollary 2.1.12. If (X, | |
n
) are as in the last lemma, and if x M
n
(X),
then there exists a completely contractive u : X M
n
such that |u
n
(x)| = |x|
n
.
Proof. By the HahnBanach theorem there exists F Ball(M
n
(X)

) with [F(x)[ =
|x|
n
. By the last lemma, there exist vectors , Ball(C
n
2
), and a completely
contractive u : X M
n
such that F(x) = u
n
(x), ). Thus
|x|
n
= [F(x)[ |u
n
(x)||||| |u
n
(x)|.
However clearly |u
n
(x)| |x|
n
.
Theorem 2.1.13. (Ruan) Suppose that X is a vector space, and that for each
n N we are given a norm | |
n
on M
n
(X) satisfying conditions (R1) and (R2)
above. Then X is linearly completely isometrically isomorphic to a linear subspace
of B(H), for some Hilbert space H.
Proof. Suppose that (X, | |
n
) satises (R1) and (R2). Let I be the collection
of all completely contractive : X M
n
, for all n N. We write n

= n if
: X M
n
. Let M =

I
M
n

, a von Neumann algebra (see 4.1.16 in the

C

-course), and therefore certainly an operator space. Dene u : X M by

u(x) = ((x))
I
. This is a complete contraction, as is very easy to check, and so
u
n
is a contraction for each n N. Choose x M
n
(X), and by Corollary 2.1.12
choose completely contractive : X M
n
such that |
n
(x)| = |x|
n
. If x = [x
ij
]
then since the projection P : M M
n

onto the -entry is a -homomorphism,

and hence completely contractive, we have
|u
n
(x)| = |[u(x
ij
)]| |[P(u(x
ij
))]| = |[(x
ij
)]| =
n
(x)| = |x|
n
.
Thus u
n
is an isometry, and hence u is a complete isometry.
2.1.14 We next discuss some consequences and applications of Ruans theorem. It
follows immediately from this result that the abstract operator spaces are precisely
the vector spaces X with matrix norms satisfying (R1) and (R2). More precisely, a
44 CHAPTER 2. OPERATOR SPACES
sequence of norms ||
n
, with ||
n
a norm on M
n
(X), is an operator space structure
(oss) on X in the sense of 2.1.4, i they satisfy (R1) and (R2). The one direction of
this follows immediately from Theorem 2.1.13. The other follows immediately from
the fact noted earlier that every concrete, and hence every abstract, operator space
satises (R1) and (R2).
2.1.15 (Quotient operator spaces) If Y X is a closed linear subspace of an
operator space, then Ruans theorem allows one to check that X/Y is an operator
space with matrix norm on M
n
(X/Y ) coming from the identication M
n
(X/Y )

=
M
n
(X)/M
n
(Y ), the latter space equipped with its quotient Banach space norm.
Explicitly, these matrix norms are given by the formula |[x
ij

+Y ]|
n
= inf|[x
ij
+
y
ij
]|
n
: y
ij
Y . Here x
ij
X. Note that with this denition, the canonical
quotient map q : X X/Y is a complete quotient map.
To check the (R1) condition, note that more generally if M
n,m
, M
m,n
, x =
[x
ij
] M
m
(X), then q
n
(x) = q
n
(x) (an exercise in linear algebra), and so
|q
n
(x)| = |q
n
(x)| |x|
n
|||x|
m
||.
Given > 0 we may choose x above so that |x| < |q
n
(x)| +, and then
|q
n
(x)| |||x|
m
|| < ||(|q
n
(x)|
m
+)||.
Letting 0 gives |q
n
(x)| |||q
n
(x)|
m
|| as desired.
To prove (R2) let x be as in the last paragraph, and choose y M
k
(X) with
|y|
k
< |q
k
(y)| +. Then q
n+k
(x y) = q
n
(x) q
k
(y), and so
|q
n
(x) q
k
(y)|
n+k
= |q
n+k
(x y)|
n+k
|x y|
n+k
= max|x|, |y|
which is dominated by max|q
n
(x)|, |q
k
(y)| + . Letting 0 gives |q
n
(x)
q
k
(y)|
n+k
max|q
n
(x)|, |q
k
(y)|, which is one half of the (R2) condition. The
other half follows from our slightly more general version of (R1) in the last paragraph
since, for example, if = [I
n
O] then
|q
n
(x)|
n
= |(q
n
(x) q
k
(y))

| |q
n
(x) q
k
(y)|.
2.1.16 (Factor theorem) If u: X Z is completely bounded, and if Y is a closed
subspace of X contained in Ker(u), then the canonical map u: X/Y Z induced
by u is also completely bounded, with | u|
cb
= |u|
cb
. If Y = Ker(u), then u is
a complete quotient map if and only if u is a completely isometric isomorphism.
Indeed this follows exactly as in the usual Banach space case (exercise).
2.1.17 (The -direct sum) Let X

: I be a family of operator spaces, and we

write

(or

if more clarity is needed), for their -direct sum as Banach

spaces. If I = 1, . . . , n then we usually write this sum as X
1

X
n
. Thus
a tuple (x

) is in

if and only if x

for all , and sup

|x

| < . Let
us write P

for the projection of

onto X

. We assign

an operator
space structure by dening |x|
n
= sup

|x

|
M
n
(X

)
if x M
n
(

) and x

=
(P

)
n
(x). Another way to say this, is that we are assigning M
n
(

) the norm
2.1. BASIC FACTS, EXAMPLES, AND CONSTRUCTIONS 45
making the canonical linear algebraic isomorphism M
n
(

)

=

M
n
(X

) an
isometry. It is easy to check by Ruans theorem that this is an operator space
structure on

. Or one can see this directly as follows. If X

, where
A

is a C

-algebra, then

is isometrically embedded in the C

-algebra direct
sum

discussed in the C

-course. The canonical operator space structure

on the C

-algebra

is given by the formula |x|

n
= sup

|x

|
M
n
(A

)
, where
x

= (P

)
n
(x). This may be seen, as in the end of 2.1.5, by proving that the
latter formula is a C

-algebra norm on M
n
(

), which in turn follows easily

for example from the -isomorphism M
n
(

)

=

M
n
(A

). Then

inherits this operator space structure from

.
Another way to say the above, is that if X

B(H

) then

may
be regarded as the subspace of B(
2

) consisting of the operators which take

(

)
2

to (x

). It is a simple exercise to see that the norm of the latter

operator is sup

|x

|, with a similar formula for matrix norms.

By denition, the canonical inclusion and projection maps between

and
its th summand are complete isometries and complete quotient maps respectively.
As we said above, if X

are C

-algebras then this direct sum is the usual C

-algebra
direct sum. If the X

are W

-algebras then this direct sum is a W

-algebra too, as
discussed in 4.1.16 in the C

-course.
The -direct sum has the following universal property. If Z is an operator space
and u

: Z X

are completely contractive linear maps, then there is a canonical

complete contraction Z

taking z Z to the tuple (u

(z)). We leave this

as an easy exercise.
If X

= X for all I, then we usually write

I
(X) for

.
One may dene a c
0
-direct sum of operator spaces to simply be the closure in

of the set of tuples which are zero except in nitely many entries.
2.1.18 (Mapping spaces) If X, Y are operator spaces, then the space CB(X, Y )
of completely bounded linear maps from X to Y , is also an operator space, with
matrix norms determined via the canonical isomorphism between M
n
(CB(X, Y ))
and CB(X, M
n
(Y )). That is, if [u
ij
] M
n
(CB(X, Y )), then dene
|[u
ij
]|
n
= sup
_
|[u
ij
(x
kl
)]|
nm
: [x
kl
] Ball(M
m
(X)), m N
_
. (2.4)
Here the matrix [u
ij
(x
kl
)] is indexed on rows by i and k and on columns by j and
l. Then
M
n
(CB(X, Y ))

= CB(X, M
n
(Y )) isometrically. (2.5)
One may see that (2.4) denes an operator space structure on CB(X, Y ) by appeal-
ing to Ruans theorem 2.1.13. Alternatively, one may see it as follows. Consider
the set I =
n
Ball(M
n
(X)), and for x Ball(M
m
(X)) I set n
x
= m. Con-
sider the operator space direct sum (see 2.1.17)

xI
M
n
x
(Y ), which is an opera-
tor space. Then the map from CB(X, Y ) to

xI
M
n
x
(Y ) taking u to the tuple
(u
n
x
(x))
x

x
M
n
x
(Y ) is (almost tautologically) a complete isometry. For exam-
ple, note that
|(u
n
x
(x))
xI
| = sup|u
n
x
(x)| : x I = sup|u
n
| : n N = |u|
cb
.
Thus CB(X, Y ) is an operator space.
46 CHAPTER 2. OPERATOR SPACES
2.1.19 (The dual of an operator space) The special case when Y = C in 2.1.18 is
particularly important. In this case, for any operator space X, we obtain by 2.1.18
an operator space structure on X

= CB(X, C). The latter space equals B(X, C)

isometrically by 2.1.7. We call X

, viewed as an operator space in this way, the

operator space dual of X. This duality will be studied further in later sections. By
(2.5) we have
M
n
(X

)

= CB(X, M
n
) isometrically. (2.6)
(Note that the map implementing this isomorphism is also exactly the canonical
map from M
n
X

to B(X, M
n
), where (a )(x) = (x)a, for a M
n
,
X

, x X.)
2.1.20 (Minimal operator spaces) Let E be a Banach space, and consider the
canonical isometric inclusion of E in the commutative C

-algebra C(Ball(E

)).
Here E

is equipped with the w

-topology. This inclusion induces, via 2.1.5, an

operator space structure on E, which is denoted by Min(E). We call Min(E) a
minimal operator space. By (2.2), the resulting matrix norms on E are given by
|[x
ij
]|
n
= sup
_
|[(x
ij
)]| : Ball(E

)
_
(2.7)
for [x
ij
] M
n
(E). Thus every Banach space may be canonically considered to be
an operator space. Since Min(E) C(Ball(E

)), we see from 2.1.7 that for any

bounded linear u from an operator space Y into E, we have
|u: Y Min(E)|
cb
= |u: Y E|. (2.8)
From this last fact one easily sees that Min(E) is the smallest operator space struc-
ture on E. For if [[[ [[[
n
was an operator space structure on E, with [[[ [[[
1
= | |,
write X for the abstract operator space which is E with these matrix norms. Then
I
E
: Min(E) X is a linear isometry, and so by (2.8) we have |I
1
E
|
cb
= |I
1
E
| = 1.
But this says precisely that [[[ [[[
n
dominates the norm in (2.7).
Also, if is any compact space and if i : E C() is an isometry, then the
matrix norms inherited by E from the operator space structure of C(), coincide
again with those in (2.7). That is, the norms in (2.7) equal |[i(x
ij
)]|
n
. This may
be seen by applying 2.1.7 to i and i
1
(the latter dened on the range of i). By
2.1.7 we have i completely contractive. But since Min(E) C(Ball(E

))) we have
by 2.1.7 again that i
1
is completely contractive. So i is a complete isometry, which
is the desired identity.
This means that minimal operator spaces are exactly the operator spaces com-
pletely isometrically isomorphic to a subspace of a C(K)-space. Note too that the
category of Banach spaces and bounded linear maps is the same as the category
of minimal operator spaces and completely bounded linear maps.
According to 3.2.2 (see particularly (3.2)) in the C

-course, another way of

stating (2.7) is to say that
M
n
_
Min(E)
_
= M
n

E (2.9)
2.1. BASIC FACTS, EXAMPLES, AND CONSTRUCTIONS 47
isometrically via the canonical isomorphism. Indeed, if [x
ij
] M
n
(Min(E)), then

n
i,j=1
E
ij
x
ij
is the corresponding element of M
n
E, and the injective tensor
product norm of the latter sum is
sup
__
_
_
n

i,j=1
E
ij
(x
ij
)
_
_
_ : Ball(E

)
_
= sup|[(x
ij
)]|
M
n
,
which is precisely |[x
ij
]|
M
n
(Min(E))
.
2.1.21 (Maximal operator spaces) If E is a Banach space then Max(E) is the
largest operator space structure we can put on E. We dene the matrix norms on
Max(E) by the following formula
|[x
ij
]|
n
= sup
_
|[u(x
ij
)]| : u Ball(B(E, Y )), all operator spaces Y
_
. (2.10)
This may be seen to be an operator space structure on X by using Ruans theorem.
However again a direct sum argument is more elementary: Dene a map i : x
(u(x))
u
from X into the operator space Z =

u
Y
u
, where the latter sum is indexed
by every u : E Y as in (2.10), and writing such Y as Y
u
. We may assume that
the cardinality of Y is dominated by that of X so that there are no set theoretic
issues. Since there exists at least one such u which is an isometry (see e.g. 2.1.20),
it is evident that i is an isometry. Thus | |
1
is the usual norm on E. Then the
matrix norms inherited by E from the operator space structure of Z, gives E an
operator space structure. However the latter coincides again with the one in (2.10).
That is, the norms in (2.10) equal |[i(x
ij
)]|
n
.
It is clear from this formula that Max(E) has the property that for any operator
space Y , and for any bounded linear u: E Y , we have
|u: Max(E) Y |
cb
= |u: E Y |. (2.11)
Indeed to prove this we may assume that u is a contraction, and then from (2.10) we
see that |[u(x
ij
)]| is dominated by the norm in (2.10). That is, u
n
is a contraction,
so that u is completely contractive, as a map into E with the matrix norms from
(2.10). This proves (2.11).
It is also clear that Max(E) is the largest operator space structure we can put
on E. For if [[[ [[[
n
was another operator space structure on E, with [[[ [[[
1
= | |,
write X for the abstract operator space which is E with these matrix norms. Then
I
E
: Max(E) X is a linear isometry, and so by (2.11) we have |I
E
|
cb
= |I
E
| = 1.
But this says precisely that [[[ [[[
n
is dominated by the norm in (2.10).
2.1.22 (Hilbert column and row spaces) If H is a Hilbert space then there are
two canonical operator space structures on H most commonly considered. The rst
is the Hilbert column space H
c
. Informally one should think of H
c
as a column in
B(H). Thus if H =
2
n
then H
c
= M
n,1
, thought of as the matrices in M
n
which
are zero except on the rst column. We write this operator space also as C
n
, and
the row version as R
n
. Note that for such a matrix x the norm |x| = |x

x|
1
2
is
precisely the
2
n
norm of the entries in x. So C
n

=
2
n
isometrically. However C
n
is
48 CHAPTER 2. OPERATOR SPACES
not completely isometric to R
n
, and they fail to even be completely isomorphic if
n is innite (see the discussion after Proposition 2.1.25).
For a general Hilbert space H there are several simple ways of describing H
c
more precisely. For example, one may identify H
c
with the concrete operator space
B(C, H). If H write T

: C H for the operator taking 1 to . It is easy to

see that T

is the operator on C taking 1 to , ). Thus if [

ij
] M
n
(H) then
by the C

-identity
|T

ij
]| = |[
n

k=1
T

ki
T

kj
]|
1
2
= |[
n

k=1

kj
,
ki
)]|
1
2
.
Another equivalent description of Hilbert column space is as follows: If is a
xed unit vector in H, then the set H of rank one operators is a closed
subspace of B(H) which is isometric to H via the map . (By convention,
maps H to , ).) Thus we may transfer the operator space structure
on H inherited from B(H) over to H. The resulting operator space structure
is independent of and coincides with H
c
. To see this, we will use the C

-identity
in M
n
(B(H)) applied to |[
ij
]|. Note that
[
ij
]

[
ij
] = [
ji
][
ij
] = [
n

k=1
(
ki
)(
kj
)].
However, (
ki
)(
kj
) =
kj
,
ki
) R, where R = . It was left as an exercise
in 2.1.8 that |[z
ij
R]| = |R||[z
ij
]| for scalars z
ij
, and so we conclude using the
C

-identity that |[
ij
]| = |[

n
k=1

kj
,
ki
)]|
1
2
. That is,
|[
ij
]|
M
n
(H
c
)
=
_
_
_
_
n

k=1

kj
,
ki
)
__
_
_
1
2
, [
ij
] M
n
(H). (2.12)
This shows that this is the same operator space structure on H as the previous one.
If H =
2
n
and we take = (1, 0, , 0) then is precisely the matrices in
M
n
which are zero except on the rst column.
If T B(H, K) then |T| = |T|
cb
, where the latter is the norm taken in
CB(H
c
, K
c
). Indeed let [
ij
] M
n
(H
c
), and let B(
2
n
,
2
n
(H)) correspond to
this matrix via the identity M
n
(H
c
) = M
n
(B(C, H)) = B(
2
n
,
2
n
(H)). Similarly,
let B(
2
n
,
2
n
(K)) corresponding to [T
ij
]. Then = (I

2
n
T) , and hence
|| |I

2
n
T||| |T||| (see 3.4.2 in the C

-course). This shows that

|T
n
| |T|, and so |T|
cb
|T|.
More generally, we have
B(H, K) = CB(H
c
, K
c
) completely isometrically (2.13)
We will give a quick proof of this identity at the end of this section.
A subspace K of a Hilbert column space H
c
is again a Hilbert column space,
as may be seen by considering (2.12). Similarly the quotient H
c
/K
c
is a Hilbert
column space completely isometric to (H K)
c
. This may be seen by considering
2.1. BASIC FACTS, EXAMPLES, AND CONSTRUCTIONS 49
the canonical projection P from H
c
onto (H K)
c
. Note P is applying completely
contractive by the fact at the start of the second last paragraph, and is therefore
clearly a complete quotient map. Now apply 2.1.16 to see that H
c
/Ker(P) =
H
c
/K
c
= (H K)
c
completely isometrically.
We dene Hilbert row space similarly. Recalling that H

=

H is a Hilbert space
too, we identify H
r
with the concrete operator space B(

H, C). Analogues of the
above results for H
c
hold, except that there is a slight twist in the corresponding
version of (2.13). Namely, although B(H, K) = CB(H
r
, K
r
) isometrically, this
is not true completely isometrically. Instead, as we shall see, there is a canonical
completely isometric isomorphism B(H, K)

= CB(

K
r
,

H
r
). We write C and R for

2
with its column and row operator space structures respectively.
We have
(H
c
)

=

H
r
and (H
r
)

=

H
c
(2.14)
completely isometrically using the operator space dual structure in 2.1.19. The
rst relation is obtained by setting K = C in (2.13). The second relation follows
e.g. from the rst if we replace H there by K =

H, and take the operator space
dual, using the fact that Hilbert spaces are reexive, and also the rst result in
the next Section 2.2, which states that X X

completely isometrically. Thus

K
c
= (K
c
)

= (H
r
)

completely isometrically.
Just as in one of the exercises for Chapter 1, the map T T

is a complete
isometry from CB(X, Y ) into CB(Y

, X

) and this map is onto if X is reexive.

Thus if H, K are Hilbert spaces then we have
B(H, K)

= CB(H
c
, K
c
)

= CB((K
c
)

, (H
c
)

)

= CB(

K
r
,

H
r
),
using (2.14).
2.1.23 (Matrix spaces) If X is an operator space, and I, J are sets, then we
write M
I,J
(X) for the set of I J matrices whose nite submatrices have uniformly
bounded norm. We explain: By an IJ matrix we mean a matrix x = [x
i,j
]
iI,jJ
,
where x
i,j
X. For such a matrix x, and for a subset = C D I J, we
write x

for the submatrix [x

i,j
]
iC,jD
. Sometimes we also write x

for the
same matrix viewed as an element of M
I,J
(X), and with all other entries zero. We
say the submatrix is nite if is nite. We dene |x| to be the supremum of
the norms of its nite submatrices, and M
I,J
(X) consists of those matrices x with
|x| < . Similarly there is an obvious way to dene a norm on M
n
(M
I,J
(X)) by
equating this space with M
I,J
(M
n
(X)), and one has M
n
(M
I,J
(X))

= M
n.I,n.J
(X),
for n N.
We write M
I
(X) = M
I,I
(X), C
w
I
(X) = M
I,1
(X), and R
w
I
(X) = M
1,I
(X). If
I =
0
we simply denote these spaces by M(X), C
w
(X) and R
w
(X) respectively.
Also, M
n
I,J
(X) will denote the vector subspace of M
I,J
(X) consisting of nitely
supported matrices, that is, those matrices with only a nite number of nonzero
entries. We write K
I,J
(X) for the norm closure in M
I,J
(X) of M
n
I,J
(X). We set
K
I
(X) = K
I,I
(X), C
I
(X) = K
I,1
(X), and R
I
(X) = K
1,I
(X). Again we merely
write K(X), R(X) and C(X) for these spaces if I =
0
. If X = C then we write
C
I
(C) = C
I
. Similarly, R
I
= R
I
(C). We write K
I,J
for K
I,J
(C), and M
I,J
for
M
I,J
(C).
50 CHAPTER 2. OPERATOR SPACES
We leave the following assertions about matrix spaces as exercises, for the most
part. Throughout, I, J, I
0
, J
0
are sets and X, Y are operator spaces.
(1) If X Y (completely isometrically), then M
I,J
(X) M
I,J
(X) completely
isometrically. Thus if X B(H, K) then M
I,J
(X) M
I,J
(B(H, K)). This
is important, since this reduces most facts about M
I,J
(X) to facts about
M
I,J
(B(H, K)), which we shall see in (5) is a simple space to deal with.
(2) If u: X Y is completely bounded, then so is the obvious amplication
u
I,J
: M
I,J
(X) M
I,J
(Y ), and |u
I,J
|
cb
= |u|
cb
. Clearly u
I,J
also restricts
to a completely bounded map from K
I,J
(X) to K
I,J
(Y ). If u is a complete
isometry, then so is u
I,J
(see (1)). Thus the M
I,J
() and K
I,J
() constructions
are injective in some sense.
(3) M
I,J

= B(
2
J
,
2
I
) completely isometrically. Via this identication, K
I,J
=
S

(
2
J
,
2
I
) completely isometrically. Thus for any Hilbert spaces K, H we
have that B(K, H)

= M
I
0
,J
0
completely isometrically, for some sets I
0
, J
0
.
(4) We have M
I,J
(M
I
0
,J
0
)

= M
II
0
,JJ
0

= M
I
0
,J
0
(M
I,J
) completely isometrically.
(5) Putting (3) and (4) together, it follows easily that for any sets I, J, we have
M
I,J
(B(K, H))

= B(K
(J)
, H
(I)
) completely isometrically.
(6) Fix i I, j J. The map which takes x X to the matrix in M
I,J
(X) which
is all zero except for an x in the i-j-entry, is a complete isometry. The map
M
I,J
(X) X which takes a matrix to its i-j-entry, is a complete contraction.
The map which takes x X to the matrix in M
I
(X) which is all zero except
for an x in all the entries on the main diagonal, is a complete isometry.
(7) If X is an operator space then so is M
I,J
(X). Indeed if X B(H), then
by (1) and (5) we have M
I,J
(X) M
I,J
(B(H))

= B(H
(J)
, H
(I)
) completely
isometrically. If X is complete then so is M
I,J
(X). To see this, we can
suppose that X is a closed subspace of B(H). Then M
I,J
(X) M
I,J
(B(H))

=
B(H
(J)
, H
(I)
), and the latter space is complete. Suppose that a
n
M
I,J
(X),
with a
n
a M
I,J
(B(H)). Then by (6) the i-j-entry of a
n
converges to the
i-j-entry of a, and so the latter is in X. Hence a M
I,J
(X). So M
I,J
(X) is
norm closed in M
I,J
(B(K, H)).
(8) We have M
I,J
(X) = C
w
I
(R
w
J
(X)) = R
w
J
(C
w
I
(X)) completely isometrically.
One way to see this is to rst check this identity in the case X = B(H) using
(5) repeatedly, and then use this fact to do the general case.
(9) By a similar argument, M
I,J
(M
I
0
,J
0
(X))

= M
II
0
,JJ
0
(X) for any operator
space X, generalizing (4).
(10) C
w
I
(C) = C
I
= (
2
I
)
c
(see 2.1.22 for this notation). Indeed, by (5) we have
C
w
I
(C) = B(C,
2
I
) = (
2
I
)
c
, and this must equal C
I
since nitely supported
tuples are dense in
2
I
. Similarly, R
I
= R
w
I
(C) = (
2
I
)
r
.
(11) K
I,J
(X) is the set of x M
I,J
(X) such that the net (x

) converges to x,
where the net is indexed by the nite subsets = C D of I J, ordered by
inclusion.
(12) For any operator spaces X, Y we have CB(X, M
I,J
(Y ))

= M
I,J
(CB(X, Y ))
isometrically. We leave it as an exercise to write down the obvious isomorphism
here, and to check that this is a (complete) isometry.
2.1. BASIC FACTS, EXAMPLES, AND CONSTRUCTIONS 51
2.1.24 (Innite sums) Suppose that X, Y are subspaces of a C

-algebra A
B(H). Let I be an innite set. If x R
w
I
(X) and y C
I
(Y ), then the product
xy =

i
x
i
y
i
, if x and y have ith entries x
i
and y
i
respectively, actually converges in
norm to an element of A, and we have |xy| |x||y|. This is clear if I is nite, in
this case we can view x R
n
(B(H)) = B(H
(n)
, H) and similarly y B(H, H
(n)
),
and then clearly |xy| |x||y|. To see the general case, we use the following
notation. If z is an element of R
w
I
(X) or C
I
(Y ), and if I, write z

for z but
with all entries outside switched to zero. Since y C
I
(Y ), by 2.1.23 (11) given
> 0 there is a nite set I, such that |y y

| = |y

c | < . If

is a nite
subset of I not intersecting then
_
_
_

x
i
y
i
_
_
_ = |x

| |x

||y

| |x||y

| < |x|.
Hence the sum converges in norm as claimed. For any nite I, a compu-
tation identical to the rst part of the second last centered equation shows that
|

i
x
i
y
i
| |x||y|. Taking the limit over , we have |xy| |x||y|.
Proposition 2.1.25. For any operator space X and set I, we have that CB(C
I
, X)

=
R
w
I
(X) and CB(R
I
, X)

= C
w
I
(X) completely isometrically.
Proof. We sketch the proof of just the rst relation. Dene L: R
w
I
(X) CB(C
I
, X)
by L(x)(z) =

i
x
i
z
i
, for x R
w
I
(X), z C
I
. This map is well dened, by the
argument for 2.1.24 for example. It is also easy to check, by looking at the partial
sums of this series as in 2.1.24, that L is contractive. Alternatively, this can be seen
by viewing X B(H), and L(x)(z) as the product (composition) TS of the oper-
ator S

: H H
(I)
: [z
i
], and the operator T
x
: H
(I)
H : [
i
]

i
x
i

i
.
It is easy to argue that
|[L(x)(z
ij
)]| = |[T
x
S
z
ij
]| |T
x
||[S
z
ij
]| = |x||[z
ij
]|,
so that L is a contraction.
Conversely, for u in CB(C
I
, X), let x be a 1 by I matrix whose ith entry is
u(e
i
), where (e
i
) is the canonical basis. If = i
1
, i
2
, , i
m
I then
|x

| = |[u(e
i
1
) u(e
i
2
) u(e
i
m
)]| |u|
cb
|[e
i
1
e
i
2
e
i
m
]| = |u|
cb
,
since the last matrix after erasing rows and columns of zeros is an identity ma-
trix. Thus x R
w
I
(X) and |x|
R
w
I
(X)
|u|
cb
. It is easy to see that L(x)(z) =

i
u(e
i
)z
i
= u(

i
e
i
z
i
) = u(z) if z C
I
. Thus L(x) = u, and so L is a surjective
isometry. This together with (2.5) yields
M
m
(CB(C
I
, X))

= CB(C
I
, M
m
(X))

= R
w
I
(M
m
(X))

= M
m
(R
w
I
(X))
isometrically. From this one sees that L is a complete isometry.
We will use the last lemma to verify two facts that were mentioned earlier. First,
that C
I
is not completely isomorphic to R
I
if I is innite. One way to see this is to
52 CHAPTER 2. OPERATOR SPACES
note that CB(C
I
, R
I
)

= R
w
I
(R
I
)

= R
II
by Proposition 2.1.25, and (10) and (4)
of 2.1.23. This is saying that for an operator T : C
I
R
I
, the cb-norm equals
its Hilbert-Schmidt norm (that is, its norm in S
2
(
2
I
), see the end of Section 3.3
and 3.4.1 in the C

-course). Similarly if S : R
I
C
I
. So if C
I

= R
I
completely
isomorphically, then there is an invertible operator between them which is in S
2
(
2
I
).
Since S
2
(
2
I
) is an ideal (see the end of Section 3.3 in the C

-course), this implies

that the identity map (= TT
1
) is in S
2
(
2
I
), which is absurd if I is innite.
Second, we show that B(H, K)

= CB(H
c
, K
c
) completely isometrically. Indeed,
CB(C
I
, C
J
)

= R
w
I
(C
J
) = R
w
I
(C
w
J
)

= M
J,I

= B(
2
I
,
2
J
),
using Proposition 2.1.25, and 2.1.23 (10), (8), and (3).
Historical note: The results in Section 2.1 are almost all due to Arveson,
Eros, and Ruan [1, 15]. Hamana studied matrix spaces (see 2.1.23) in some of his
papers (see e.g. [18]), and they are studied in more detail by Eros, and Ruan in
[11, 12]. Maximal operator spaces were rst considered by Blecher and Paulsen [6].
Preliminary forms of some of the results in 2.1.22 were noted in the latter paper;
and the fact that B(H, K)

= CB(H
c
, K
c
) isometrically is due to Wittstock [30]. In
the generality listed here, the main source for the results towards the end of 2.1.22
is [14], although Blecher independently discovered a couple of these [3].
Exercises.
(1) If T B(K, H), and if [z
ij
] M
n
show that |[z
ij
T]| = |T||[z
ij
]|. Deduce
that C has a unique operator space structure (up to complete isometry).
(2) If X B(H, K), and if S : K K

and T : H

H are operators between

Hilbert spaces, prove that the map x SxT is completely bounded with cb-
norm dominated by |S||T|.
(3) If K is a closed subspace of a Hilbert space H, prove that the map x P
K
x
|K
is completely contractive from B(H) to B(K).
(4) Prove that if S : M
n
(Y ) M
n
(Z) is a linear map, where Y, Z are vector spaces,
then S = u
n
for a linear u : Y Z i S(x) = S(x) for all x M
n
(Y )
and , M
n
.
(5) (R. R. Smith) If u : X M
n
satises |u
n
| 1, use linear algebra to show that
|u
m
| 1 for all m n, so that u is completely contractive. [Hint: if m n
and , C
mn
, then we can write =

n
k=1

k


e
k
, where
k
C
m
. Since
Span
k
: k = 1, , n has dimension n, there is an isometry M
m,n
and
vectors

k
C
n
with

k
=
k
. Similarly, there is an isometry M
m,n
with

k
=
k
. Use this to nd an upper bound for the number [u
m
(x), )[.]
(6) Prove that (R1) and (R2) together are equivalent to requiring that: (R1)

|x|
n
|||x|
m
||, for all n, m N and all M
n,m
, M
m,n
, and
x M
m
(X), and (R2)

|x y|
n+m
max|x|
n
, |y|
m
for x M
n
(X), y
M
m
(X). Prove also that (R1)

and (R2)

actually imply that | |

n
is a norm.
(7) Suppose that u
i
: X B(K
i
, H
i
) are completely contractive, and that K =

i
K
i
and H =
i
H
i
(Hilbert space sum). Dene u : X B(K, H) by
2.2. DUALITY OF OPERATOR SPACES 53
u(x) = (u
i
(x)), where the latter denotes the operator (
i
) (u
i
(x)
i
) on H.
Show that |u
i
|
cb
|u|
cb
1.
(8) Prove that if in Ruans theorem X is also separable, then one may take the
Hilbert space there to be
2
.
(9) If X, Y are (possibly incomplete) operator spaces, and if : X Y is a linear
isomorphism such that the map is a well dened complete isometry
from Y

onto X

, then is completely isometric.

(10) Prove the facts stated in 2.1.23.
2.2 Duality of operator spaces
An operator space Y is said to be a dual operator space if Y is completely isomet-
rically isomorphic to the operator space dual (see 2.1.19) X

of an operator space
X. We also say that X is an operator space predual of Y , and sometimes we write
X as Y

. If X, Y are dual operator spaces then we write w

CB(X, Y ) for the space

of w

-continuous completely bounded maps from X to Y .

Unless otherwise indicated, in what follows the symbol X

denotes the dual

space together with its dual operator space structure as dened in 2.1.19. Of course
X

is considered as the dual operator space of X

.
Proposition 2.2.1. If X is an operator space then X X

completely isometri-
cally via the canonical map i
X
.
Proof. We can suppose that X is a subspace of B(H), for a Hilbert space H. Fix
n N and [x
ij
] M
n
(X). We rst show that |[i
X
(x
ij
)]|
n
|[x
ij
]|
n
. By denition,
the norm |[i
X
(x
ij
)]|
n
in M
n
((X

) equals
sup
_
|[i
X
(x
ij
)(f
kl
)]|
nm
: [f
kl
] Ball(M
m
(X

)), m N
_
= sup
_
|[f
kl
(x
ij
)]|
nm
: [f
kl
] Ball(M
m
(X

)), m N
_
|[x
ij
]|
n
,
the last line by denition of [f
kl
] Ball(M
m
(X

)).
Since |[i
X
(x
ij
)]|
n
equals the supremum above, and since M
m
(X

)

= CB(X, M
m
),
to see that i
X
is completely isometric, it suces to prove the Claim: for a given
n N, > 0, and [x
kl
] M
n
(X), there exists an integer m and a completely con-
tractive u: X M
m
such that |[u(x
kl
)]| |[x
kl
]| . In fact this Claim follows
immediately (with = 0 and m = n) from Corollary 2.1.12.
Remark. Because of its independent interest, we will give another alternative
proof of the Claim in the last proof. Let [x
ij
] M
n
(X) M
n
(B(H))

= B(H
(n)
).
Thus [x
ij
] may be viewed as an operator on H
(n)
. The norm of any operator
T B(K), for any Hilbert space K, is given by the formula |T| = sup[Ty, z)[ :
y, z Ball(K). Thus in our case,
|[x
ij
]|
n
= sup[[x
ij
]y, z)[ : y, z Ball(H
(n)
).
54 CHAPTER 2. OPERATOR SPACES
So, if > 0 is given, there exists y, z Ball(H
(n)
) such that [[x
ij
]y, z)[ > |[x
ij
]|
n

. If y = (
k
) and z = (
k
), with
k
,
k
H, then [x
ij
]y, z) =

i,j
x
ij

j
,
i
), and
so

i,j
x
ij

j
,
i
)

|[x
ij
]| .
Let K =Span
1
, . . . ,
n
,
1
, . . . ,
n
in H. This is nite dimensional, and so
there is an isometric -isomorphism : B(K) M
m
, where m = dim(K). Then
is completely contractive by Proposition 2.1.6. Let P
K
be the projection from H
onto K. Let T : B(H) B(K) be the function T(x) = P
K
x
|K
. By an exercise at
the end of the section, T is completely contractive. Let u = T, which will be
completely contractive too. Now [T(x
ij
)]y, z) =

i,j
T(x
ij
)
j
,
i
), and so
|[T(x
ij
)]|
n

i,j
T(x
ij
)
j
,
i
)

i,j
P
K
x
ij

j
,
i
)

i,j
x
ij

j
,
i
)

,
the last step since
i
K. Thus,
|[u(x
ij
)]|
n
= |[(T(x
ij
))]|
n
= |[T(x
ij
)]|
n

i,j
x
ij

j
,
i
)

|[x
ij
]| ,
using the fact at the end of the last paragraph. This proves the Claim.
2.2.2 From 2.2.1 we have for any [x
ij
] M
n
(X) that
|[x
ij
]|
n
= sup|[
kl
(x
ij
)]| : m N, [
kl
] Ball(M
m
(X

)) (2.15)
There is a canonical map : M
n
(X) CB(X

, M
n
), namely ([x
ij
])() =
[(x
ij
)], and (2.15) says that is an isometry. Note that if [x
ij
] M
n
(X), and
if (
t
) is a net in X

converging weak* to X

, then
t
(x
ij
) (x
ij
), and
so [
t
(x
ij
)] [(x
ij
)] in norm in M
n
, and hence also weak*. Thus the range of
is inside w

CB(X

, M
n
). On the other hand, if u w

CB(X

, M
n
), then u
corresponds to a matrix [
ij
] M
n
(X

). If (
t
) is a net in X

converging weak*
to X

, then u(
t
) = [
ij
(
t
)] u() = [
ij
()] in M
n
, and so
ij
(
t
)

ij
() for each i, j. Thus
ij
is weak* continuous and so
ij
= i
X
(x
ij
) for some
x
ij
X. Clearly ([x
ij
]) = u. In other words, is an isometry from M
n
(X) onto
w

CB(X

, M
n
):
M
n
(X)

= w

CB(X

, M
n
) CB(X

, M
n
). (2.16)
Another consequence of 2.2.1, is that if X is an operator space which as a Banach
space is reexive, then X

= X

completely isometrically.
2.2.3 (The adjoint map) The adjoint or dual u

of a completely bounded map

u: X Y between operator spaces is completely bounded from Y

to X

, with
|u

|
cb
= |u|
cb
. Indeed if [u
ij
] M
n
(CB(X, Y )) then u

ij
: Y

and
|[u

ij
]|
n
= sup|[u

ij
(
kl
)]| : [
kl
] Ball(M
m
(Y

)), m N.
2.2. DUALITY OF OPERATOR SPACES 55
However
|[u

ij
(
kl
)]| = sup|[u

ij
(
kl
)(x
rs
)]| : [x
rs
] Ball(M
p
(X)), p N,
= sup|[
kl
(u
ij
(x
rs
))]| : [x
rs
] Ball(M
p
(X)), p N,
and it follows by combining the last two centered equations, and using (2.15), that
|[u

ij
]|
n
= sup|[u
ij
(x
rs
)]| : [x
rs
] Ball(M
p
(X)), p N
= |[u
ij
]|
n
.
Thus : CB(X, Y ) CB(Y

, X

) is a complete isometry.
Direct computations from the denitions also show that if u is a complete quo-
tient map then u

is a complete isometry (exercise). It is slightly harder to see

that if u is completely isometric then u

is a complete quotient map. This requires

Wittstocks extension theorem, which we will prove later using elementary prop-
erties of the Haagerup tensor product. The crux of Wittstocks result is that if
X Y then any complete contraction w : X M
n
has a completely contractive
extension u : Y M
n
. To see that u

is a complete quotient map if u : X Y is

completely isometric, let [
ij
] Ball(M
n
(X

)). By (2.6) we may regard [

ij
] as a
complete contraction g : X M
n
. By Wittstocks extension theorem there exists
a complete contraction w : Y M
n
with w
|u(X)
= g u
1
on u(X). By (2.6) we
may regard w as a matrix [
ij
] Ball(M
n
(Y

)). We claim that [u

(
ij
)] = [
ij
].
Indeed, if x X then
[u

(
ij
)(x)] = [
ij
(u(x))] = w(u(x)) = g(u
1
(u(x)) = g(x) = [
ij
(x)].
Thus u

is a complete quotient map. Conversely, if u

is a complete quotient map

then u

is a complete isometry, so that u is a complete isometry (using 2.2.1).

Thus u is a complete isometry if and only if u

is a complete isometry.
2.2.4 (Duality of subspaces and quotients) The operator space versions of the
usual Banach duality of subspaces and quotients follow easily from 2.2.3. If X is a
subspace of Y , then we have X

= Y

/X

and (Y/X)

= X

. Indeed the dual of

the inclusion map i : X Y will be a complete quotient map i

: Y

, which
induces a complete isometry X

= Y

/Ker(i

) = Y

/X

. Similarly, the dual

of the canonical quotient map q : Y Y/X is the canonical complete isometry
q

: (Y/X)

which we know from the Banach space case has range X

.
The predual versions go through too with the same proofs as in the Banach
space case (see the discussion around Corollary 1.4.9): if X is a w

-closed subspace
of a dual operator space Y , then (Y

/X

)

= (X

= X as dual operator spaces.

Also, (X

)

= Y/(X

= Y/X completely isometrically. These use the facts in

the last paragraph, and the Banach space fact that (X

= X.
2.2.5 (Good and bad preduals) If X is an operator space which has a predual
Banach space Z, then there is only one way to give Z an operator space structure
with any hope that Z

= X completely isometrically. Namely, view Z X

and
give Z the operator space structure inherited from X

. That is, dene

|[z
ij
]|
n
= sup|[x
pq
, z
ij
)]| : [x
pq
] Ball(M
m
(X)), m N, (2.17)
56 CHAPTER 2. OPERATOR SPACES
where , ) is the pairing between X and Z. Unfortunately, even then Z

may fail
to be completely isometric to X. We shall see an example of this later. Thus there
may be good and bad Banach space preduals of an operator space X (the bad
ones having no operator space structure whose operator space dual is X).
2.2.6 (The trace class operator space) If H is a Hilbert space then B(H) is a
dual Banach space and, fortunately, its (unique) predual Banach space S
1
(H) is
good in the sense of 2.2.5. More precisely, let us equip its predual S
1
(H) (e.g.
see Section 3.3 in the C

-course) with the operator space structure it inherits from

B(H)

via the canonical isometric inclusion S

1
(H) B(H)

. Then we claim that

B(H)

= S
1
(H)

completely isometrically. Indeed the canonical map : B(H)

S
1
(H)

is completely contractive by denition. Indeed if X, Z are as in 2.2.5, then

by denition we equip Z with the the operator space structure making the canonical
map : Z X

a complete isometry. Then the map X Z

taking x X to

( x)
is a complete contraction, and this is the canonical map from X into Z

. To see
that is completely isometric, we use the second proof of 2.2.1, given in the Remark
after that result. This shows that for any n N, > 0, and [x
kl
] M
n
(B(H)),
we can nd an integer m and a completely contractive u: B(H) M
m
such that
|[u(x
kl
)]| |[x
kl
]|. We recall that u is the composition of maps and T there.
Since M
m
and B(K) are nite dimensional, is w

-continuous. If x
t
x weak*
in B(H), and if , K then
T(x
t
), ) = x
t
, ) x, ) = T(x), ),
so that T is w

-continuous. Hence u is w

-continuous. By the argument in 2.2.2, u

corresponds to a matrix [z
ij
] M
m
(S
1
(H)), and by (2.17) the norm of this matrix
equals the cb-norm of u, which is 1. Finally, we have
|[x
kl
]| |[u(x
kl
)]| = |[x
kl
, z
ij
)]| |[(x
kl
)]|
n
.
Since > 0 was arbitrary, we have proved the desired reverse inequality. Thus is
a complete isometry.
Similarly, B(K, H) is the dual operator space of the space S
1
(H, K) of trace
class operators, the latter regarded as a subspace of B(K, H)

. Henceforth, when
we write S
1
(H, K) we will mean the operator space predual of B(K, H) described
above. Similarly, we will henceforth also view S
1
n
= M

n
as an operator space.
Lemma 2.2.7. Any w

-closed subspace X of B(H) is a dual operator space. In-

deed, if Y = S
1
(H)/X

is equipped with its quotient operator space structure

inherited from S
1
(H), then X

= Y

completely isometrically.
Proof. This follows from 2.2.4 and 2.2.6.
In particular this shows that any von Neumann algebra equipped with its nat-
ural operator space structure (see 2.1.5) is a dual operator space.
The converse of 2.2.7 is true too, as we see next, so that dual operator spaces,
and the w

-closed subspaces of some B(H), are essentially the same thing.

2.2. DUALITY OF OPERATOR SPACES 57
Lemma 2.2.8. Any dual operator space is completely isometrically isomorphic,
via a homeomorphism for the w

-topologies, to a w

-closed subspace of B(H), for

some Hilbert space H.
Proof. Suppose that W is a dual operator space, with predual X. Let Y = C, and
recall from 2.1.18 the construction of a complete isometry
W = CB(X, Y )

xI
M
n
x
(Y ) =
xI
M
n
x
,
namely the map J taking w W to the tuple ([w, x
ij
)])
x
in
x
M
n
x
. Since the
maps w w, x
ij
) are w

-continuous for any xed x I, and since

n
x
S
1
n
x
is
dense in the Banach space predual
1
x
S
1
n
x
of
x
M
n
x
, it is easy to see that J
is w

-continuous too. Indeed by Lemma 1.4.11 to show that J is w

-continuous,
it suces to show that if a I is xed, and if z S
1
n
where n = n
a
, and if

a
: S
1
n
a

1
x
S
1
n
x
is the canonical inclusion map, and if
t
weak* in X

, then
J(
t
),
a
(z)) J(),
a
(z)). However
J(
t
),
a
(z)) =
n

i,j

t
(a
ij
)z
ij

n

i,j
(a
ij
)z
ij
= J(),
a
(z)).
So J is w

-continuous. Thus by the Krein-Smulian theorem 1.4.7, W is completely

isometrically and w

-homeomorphically isomorphic to a w

-closed subspace of the

W

-algebra
x
M
n
x
. If the latter is regarded as a von Neumann subalgebra of B(H)
say, then W is completely isometrically and w

-homeomorphically isomorphic to a
w

-closed subspace of B(H).

2.2.9 (W

-continuous extensions) If X and Y are two operator spaces and if

u: X Y

is completely bounded, then its (unique) w

-continuous extension
u: X

provided by 1.4.12 is completely bounded, with | u|

cb
= |u|
cb
.
Indeed recall from 1.4.12 that u = i

Y
u

; and clearly
| u|
cb
= |i

Y
u

|
cb
|i

Y
|
cb
|u

|
cb
= |u|
cb
,
using the rst paragraph in 2.2.3, whereas | u|
cb
|u|
cb
since u extends u. Note
that since u is w

-continuous, we have u(X

) u(X)
w
. The above also shows
that
CB(X, Y

) = w

CB(X

, Y

) (2.18)
isometrically via the mapping u u. Indeed note that if g w

CB(X

, Y

) then
g = g
|X
, since both of these maps are w

-continuous and they agree on the w

-dense
subset X.
By 2.2.6, the last paragraph applies in particular to B(H) valued maps.
2.2.10 (The second dual) Let X be an operator space, and x n N. We wish to
compare the spaces M
n
(X

) (equipped with its operator space dual matrix norms

as in 2.1.19), and M
n
(X)

. First note that they can be canonically identied as

topological vector spaces, as may M
n
(X

) and M
n
(X)

. Indeed note that M

n
(X)
is a direct sum of n
2
copies of X, and so we can apply the principles in Exercise
58 CHAPTER 2. OPERATOR SPACES
(3) at the end of this section. Applying these Banach space principles, we see that
we have bicontinuous isomorphisms M
n
(X)

= X

X and M
n
(X

)

=
X

. Hence we have
M
n
(X)

= (X

1

1
X

= X

= M
n
(X

).
If M
n
(X)

, let [
ij
] be the corresponding matrix in M
n
(X

), via the iso-

morphisms in the last centered equation. We will prove in 2.2.12 below that the
map [
ij
] is an isometry. As a rst easy step, let us check that it is a con-
traction. If Ball(M
n
(X)

), then by Goldstines lemma 1.4.6, there is a net

(x
s
)
s
in Ball(M
n
(X)) such that x
s
in the w

-topology of M
n
(X)

. This
means that (x
s
) () for any M
n
(X)

. Since M
n
(X)

= M
n
(X

) and
M
n
(X)

= M
n
(X

) bicontinuously, this is equivalent to

n

i,j=1

i,j
(x
s
i,j
)
n

i,j=1

i,j
(
i,j
),
i,j
X

,
which in turn is equivalent to (x
s
i,j
)
i,j
() for all i, j = 1, , n and X

.
Let [
pq
] Ball(M
m
(X

)), for some m 1. We deduce that

|[
ij
,
pq
)]| = lim
s
|[
pq
, x
s
ij
)]| 1,
by (2.4) or (2.15). Thus |[
ij
,
pq
)]| 1. By (2.4) again, we deduce that
|[
ij
]|
M
n
(X

)
1, which proves the result.
Note too that the map [
ij
] above restricts to the identity map on M
n
(X),
by the last part of the aforementioned Exercise (3) at the end of the section.
2.2.11 (The second dual of a C

-algebra) If A is a C

-algebra, then there are

at least three canonical norms one could put on M
n
(A

). Fortunately, they are

all the same, as we now show. The rst two are the ones discussed in 2.2.10.
The third are those from 2.1.5, arising from the fact that A

is a C

-algebra
(see Theorem 4.1.18 in the C

-course), and hence has a canonical operator space

structure. To see that these three are the same, we will need to use notation
and facts from around Theorem 4.1.18 in the C

-course. To avoid confusion, we

state that whenever we write M
n
(A

) below, we are equipping this space with

its operator space dual matrix norms (see 2.1.19); thus M
n
(A

)

= CB(A

, M
n
)
isometrically. Let
u
: A B(H
u
) denote the universal representation of A, and
we write A

for the von Neumann algebra

u
(A)

(see the proof of Theorem 4.1.18

in the C

-course). The claim will follow if we can prove for any xed n 1 that
M
n
(A)

= M
n
(A

)

= M
n
(A

) isometrically (2.19)
via the canonical maps. The rst of these maps is the contraction from M
n
(A)

to M
n
(A

) discussed in 2.2.10. The second map in (2.19) is (

u
)
n
, which is a
contraction since according to 2.2.9, the mapping
u
is a complete contraction. To
establish (2.19), we need only prove that the resulting contraction : M
n
(A)

M
n
(A

) is isometric. It is clearly one-to-one. We regard M

n
(A)

as a C

-algebra
2.2. DUALITY OF OPERATOR SPACES 59
by applying Theorem 4.1.18 in the C

-course to M
n
(A). Claim: is w

-continuous.
Regarding as valued in B(H
(n)
u
), we have () , ) =

i,j

u
(
ij
)
j
,
i
), for
= [
i
], = [
i
] H
(n)
u
, and = [
ij
] as in 2.2.10. If
s
weak* in M
n
(A)

then the argument in 2.2.10 shows also that

s
i,j

i,j
weak* in A

, and so

u
(
s
ij
)
j
,
i
)
u
(
ij
)
j
,
i
). Hence (
s
) , ) () , ), which implies
that is w

-continuous. Thus is the unique w

-continuous extension of (
u
)
n
to M
n
(A)

. By ADDREF in the C

-course, is a -homomorphism. Since it is

one-to-one it is isometric.
The last result has many consequences. For example, we can use it to see
that S

(H)

= S
1
(H) completely isometrically. Indeed, since S

(H)

= B(H)
completely isometrically, S

(H)

must be the unique operator space predual S

1
(H)
of B(H) (see 2.2.6). Also we obtain:
Theorem 2.2.12. If X is an operator space then M
n
(X)

= M
n
(X

) isometri-
cally for all n N (via an isomorphism extending the identity map on M
n
(X)).
Proof. Choose a C

-algebra A with X A completely isometrically. Then X

completely isometrically by 2.2.3, hence we have both M

n
(X)

M
n
(A)

,
and M
n
(X

) M
n
(A

), isometrically. Under the identications between M

n
(A)

and M
n
(A

) and between M
n
(X)

and M
n
(X

) discussed above, these two em-

beddings are easily seen to be the same. That is, the diagram below commutes:
M
n
(A)

M
n
(A

M
n
(X)

M
n
(X

)
Hence the isometry M
n
(A

) = M
n
(A)

provided by 2.2.11, implies that we also

have M
n
(X

) = M
n
(X)

isometrically.
2.2.13 (Duality of Min and Max) For any Banach space E, we have
Min(E)

= Max(E

) and Max(E)

= Min(E

), (2.20)
completely isometrically. To see this, note that an element in M
n
(Max(E)

) may be
regarded as a map in CB(Max(E), M
n
) by (2.6). By (2.11) this is exactly the same
as a map in B(E, M
n
). On the other hand, M
n
(Min(E

))

= M
n

E

= B(E, M
n
)
by (2.9) and (3.3) in the C

-course. That is, an element in M

n
(Min(E

)) may be
regarded as a map in B(E, M
n
). These identications preserve the norm, so that
M
n
(Max(E)

) = M
n
(Min(E

)). That is, Max(E)

= Min(E

). Therefore also
Max(E

= Min(E

). However we claim that Min(E

) = Min(E)

. This claim
may be seen by rst proving it in the case that E = C(K), for compact K. The claim
follows in this case from 2.2.11, since the second dual of a commutative C

-algebra
is a commutative C

-algebra, and hence is a minimal operator space. Next we use

the fact that minimal operator spaces are completely isometric to subspaces of
unital commutative C

-algebras, and the fact that the second dual of a complete

60 CHAPTER 2. OPERATOR SPACES
isometry is a complete isometry (see 2.2.4). Thus if Min(E) C(K) completely
isometrically, then dualizing this embedding we get a commuting diagram
C(K)

Min(C(K))

Min(E

) Min(E)

where all arrows except possibly the bottom one are complete isometries. Hence
the bottom one is a complete isometry, proving the claim.
Finally, Max(E

) and Min(E)

are two operator space structures on E

with
the same operator space dual, and therefore they are completely isometric, by 2.2.1.
2.2.14 (The 1-direct sum) For a family X

: I of operator spaces, we give

its canonical predual operator space structure (see 2.2.5), as the predual
of the operator space

. It is easy to argue directly from the denitions that

the canonical inclusion and projection maps

and

between
1

and its th
summand are complete isometries and complete quotient maps respectively. Or, to
see that

: X

is a complete isometry, consider the following sequence

of canonical maps:
X

,
and let u be the composition of all these maps. On the other hand, the dual of
the canonical projection map

(which is a complete quotient map),

is a complete isometry j : X

. Moreover, the range of u falls within

j(X

), and j
1
u is the complete isometry from X

into its second dual. This

implies that the rst of these maps in the sequence,

, is a complete isometry.
Next we observe that
1

is a good predual of

, in the sense of 2.2.5.

That is,
(
1

as dual operator spaces. (2.21)

Indeed, it is easy to see (if necessary by an argument early in 2.2.6), that the canon-
ical map :

(
1

is a complete contraction. On the other hand,

if an element in Ball(M
n
((
1

)) may be regarded as a complete contraction

from
1

into M
n
. Composing this map with each

, we get a tuple in the ball

of

CB(X

, M
n
). Since CB(X

, M
n
)

= M
n
(X

), we actually obtain an element

in the ball of M
n
(

)

=

M
n
(X

). It is easy to see from all this that is

a complete isometry.
Corollary 2.2.15. Any operator space X is a complete quotient of a 1-sum of
spaces of the form S
1
n
= M

n
.
Proof. The map J in the proof of 2.2.8 is a weak* continuous complete isometry.
Thus by Exercise (5) below, J = q

for a complete quotient map q from a 1-sum of

spaces of the form S
1
n
= M

n
, onto X.
2.2.16 We end this section with an example of an operator space which is a dual
Banach space, but has no good operator space predual in the sense of 2.2.5. Let
B = B(H) with its canonical matrix norms, and let K be the compact operators
2.2. DUALITY OF OPERATOR SPACES 61
on H. Then Q = B/K is the well known Calkin algebra, which is a C

-algebra and
hence has a canonical operator space structure. The only fact we will need about
the Calkin algebra is that it is not commutative. Let X = B(H) but with matrix
norms
[[[[x
ij
][[[
n
= max|[x
ij
]|
M
n
(B)
, |[q(x
ji
)]|
M
n
(Q)
, [x
ij
] M
n
(X),
where q : B Q is the canonical quotient map. One can easily check that X
is an operator space, for example by appealing to Ruans theorem. As a Banach
space X is just B, since q is a contraction (so that [[[x[[[
1
= |x|
B
). Thus X has a
unique Banach space predual S
1
(H), the trace class. We will show that this is a
bad predual.
Notice that [[[ [[[
n
restricted to the copy of M
n
(K) is just the usual norm, since
q annihilates K. Thus if Y = S
1
(H) with its canonical predual matrix norms from
2.2.5, that is, the matrix norms coming from its duality with (X, [[[ [[[
n
), then
for [y
ij
] M
n
(Y ) we have
|[y
ij
]|
M
n
(Y )
= sup|[< y
ij
, x
kl
>]| : [x
kl
] Ball(M
m
(X)), m N
sup|[< y
ij
, x
kl
>]| : [x
kl
] Ball(M
m
(K)), m N
= |[y
ij
]|
M
n
(K

)
= |[y
i,j
]|
M
n
(S
1
(H))
,
where the last norm is the usual operator space structure of S
1
(H) (see 2.2.6).
Thus, if [x
ij
] M
n
(X), then
|[x
ij
]|
M
n
(Y

)
= sup|[< x
ij
, y
kl
>]| : [y
kl
] Ball(M
m
(Y )), m N
sup|[< x
ij
, y
kl
>]| : [y
kl
] Ball(M
m
(S
1
(H))), m N
= |[x
ij
]|
M
n
(B)
,
by the fact in 2.2.6 that S
1
(H)

= B completely isometrically. Hence if Y

= X
completely isometrically, then [[[[x
ij
][[[
n
|[x
ij
]|
M
n
(B)
. The reverse inequality fol-
lows from the denition of [[[[x
ij
][[[
n
, and so [[[[x
ij
][[[
n
= |[x
ij
]|
M
n
(B)
. There-
fore, |[q(x
ji
)]|
M
n
(Q)
|[x
ij
]|
M
n
(B)
for all [x
ij
] M
n
(X). If [z
ij
] M
n
(K), then
|[q(x
ji
)]|
M
n
(Q)
= |[q(x
ji
+ z
ji
)]|
M
n
(Q)
|[x
ij
+ z
ij
]|
M
n
(B)
. Taking the inmum
over such z
ij
K, we get |[q(x
ji
)]|
M
n
(Q)
|[q(x
ij
)]|
M
n
(Q)
. Symmetry implies that
this inequality is in fact an equality. But we claim that the only unital C*-algebras A
with |[a
ji
]|
n
= |[a
ij
]|
n
, for all n N and [a
ij
] M
n
(A), are the commutative ones,
and the Calkin algebra is not commutative! Indeed if A is any unital C*-algebra,
and if A

is A with the reversed multiplication, then A is a unital C*-algebra, and

its canonical matrix norm are given by |[a
ij
]|
M
n
(A

))
= |[a
ji
]|
M
n
(A))
(we leave
this as an exercise). Thus if the identity in the claim holds, then the identity map
A A

is a complete isometry. By Corollary 2.3.5, it is a homomorphism, so that

A is commutative.
Historical note: The results in Section 2.2 are due to Blecher (see [2], which
was written close to the date of [6, 13], although it appeared much later), with the
following main exceptions. The fact that X X

completely isometrically was

62 CHAPTER 2. OPERATOR SPACES
independently noticed in [6, 13]. Eros and Ruan had noticed 2.2.8 via a dierent
route [12]. Item 2.2.16 is a simplication by Blecher and Magajna [5] of examples
of Eros-Ozawa-Ruan, and Peters-Wittstock. Le Merdy was the rst to nd an
example of a bad predual in the sense of 2.2.5.
Exercises.
(1) As in the Exercise 1 at the end of Chapter 1, show that the map T T

is a
complete isometry from CB(X, Y ) into CB(Y

, X

), and show that this map

is onto if X and Y are reexive.
(2) Show that if u is a complete quotient map then u

is a complete isometry.
(3) Show that if F is any Banach space, and if E = F F is a nite direct sum
of n copies of F, equipped with any norm such that the n canonical inclusions of
F into E are isometries, then E

= F
1

1
F

= F

F bicontinuously.
Also, if Z = F

, equipped with any norm such that the n canonical

inclusions of F

into Z are isometries, show that Z

= E

bicontinuously, via the

map (
1
, ,
n
)(x
1
, , x
n
) =

n
k=1

k
(x
k
), for
k
F

, x
k
X. Show
that a similar statement holds for W = F

equipped with any norm

such that the n canonical inclusions of F

into W are isometries. Moreover,

show that the resulting isomorphism E

= W restricts on E to the identity
map.
(4) Show that the 1-sum has the following universal property: If Z is an operator
space and if u

: X

Z are completely contractive linear maps, then there is

a canonical complete contraction u:
1

Z such that u

= u

.
(5) Show that if u : X

is a weak* continuous isometry (resp. complete

isometry), where X and Y are Banach spaces (resp. complete operator spaces),
then u = q

for a 1-quotient map (resp. complete quotient map) q : Y X.

[Hint: By Krein-Smulian, u is a weak* homeomorphism and its range N is
weak* closed. Thus we can assume X

= N, X = Y/N

, and u is the inclusion

N Y

, in which case we can take q : Y Y/N

to be the canonical quotient

map.]
(6) If A is any unital C*-algebra, and if A

is A with the reversed multiplication,

show that A is a unital C*-algebra, and its canonical matrix norms are given
by |[a
ij
]|
M
n
(A

))
= |[a
ji
]|
M
n
(A))
.
2.3 Operator systems
2.3.1 (Operator systems) An operator system is a subspace o of a unital C

-
algebra A, which contains the identity of A, and which is selfadjoint, that is, x

o if and only if x o. A subsystem of an operator system o is a selfadjoint
linear subspace of o containing the identity 1 of o. If o is an operator system, a
subsystem of a C

-algebra A, then o has a distinguished positive cone o

+
= x
o : x 0 in A. We also write o
sa
for the real vector space of selfadjoint elements x
(i.e. those satisfying x = x

) in o. Then o has an associated ordering , namely we

say that x y if x, y are selfadjoint and y x o
+
. Note that if x o then
x+x

2
2.3. OPERATOR SYSTEMS 63
and
xx

2i
are selfadjoint, and so any x o is of the form x = h +ik for h, k o
sa
.
Also, if h o
sa
then |h|1 +h and |h|1 h are positive. Thus o
sa
= o
+
o
+
.
A linear map u : o o

between operator systems is called -linear if u(x

) =
u(x)

for all x o. Some authors say that such a map is selfadjoint. We say that
u is positive if u(o
+
) o

+
. By facts at the end of the last paragraph, any x o
may be written as x = x
1
x
2
+ i(x
3
x
4
), and from this it is easy to see that a
positive map u : o o

is -linear. Indeed,
u(x

) = u(x
1
x
2
i(x
3
x
4
)) = u(x
1
) u(x
2
) i(u(x
3
) u(x
4
)),
whereas
u(x)

= (u(x
1
) u(x
2
) +i(u(x
3
) u(x
4
)))

= u(x
1
) u(x
2
) i(u(x
3
) u(x
4
)).
The operator system M
n
(o), which is a subsystem of M
n
(A), has a positive cone
too, and thus it makes sense to talk about completely positive maps between operator
systems. These are the maps u such that u
n
= I
M
n
u: M
n
(o) M
n
(o

) is positive
for all n N. Indeed the morphisms in the category of operator systems are often
taken to be the unital completely positive maps. Any -homomorphism between
C

-algebras is clearly positive (we also noted this in the C

-course), and applying

this fact to
n
shows that is completely positive. Completely positive maps are
discussed in very many places in the literature (see e.g. [8, 21]), and we shall be
brief here.
Suppose that o is a subsystem of a unital C

-algebra. By the HahnBanach

theorem, the set of states of o (that is, the set of o

with (1) = || = 1)
is just the set of restrictions of states on the containing C

-algebra to o. Using
this fact, and Proposition 2.2.4 (resp. Proposition 2.3.8 (iv)) in the C

-course, it
follows that o
sa
(resp. o
+
) is exactly the set of elements x o such that (x) R
(resp. (x) 0) for all states of o. From this it is clear that if u: o
1
o
2
is a contractive unital linear map between operator systems, then u is a positive
map (for if x o
1+
, and if is a state on o
2
then u is a state of o
1
, so
that (u(x)) 0; and so u(x) 0). Applying this principle to u
n
, we see that a
completely contractive unital linear map between operator systems is completely
positive.
Clearly an isomorphism between operator systems which is unital and completely
positive, and has completely positive inverse, preserves all the order. Such a map
is called a complete order isomorphism. The range of a completely positive unital
map between operator systems is clearly also an operator system; we say that such
a map is a complete order injection if it is a complete order isomorphism onto its
range.
The following simple fact relates the norm to the matrix order, and is an el-
ementary exercise using the denition of a positive operator. Namely, if x is an
element of a unital C

-algebra or operator system A, or if x B(K, H), then

_
1 x
x

1
_
0 |x| 1. (2.22)
Here 0 means positive in M
2
(A) (or positive in B(H K)).
64 CHAPTER 2. OPERATOR SPACES
2.3.2 It is easy to see from (2.22) that a completely positive unital map u between
operator systems is completely contractive. (For example, to see that u is contrac-
tive, take |x| 1, and apply u
2
to the associated positive matrix in (2.22). This is
positive, so that using (2.22) again we see that |u(x)| 1.) Putting this together
with some facts from 2.3.1 we see that a unital map between operator systems is
completely positive if and only if it is completely contractive; and in this case the
map is -linear. If, further, u is one-to-one, then by applying the above to u and
u
1
one sees immediately that a unital map between operator systems is a complete
order injection if and only if it is a complete isometry.
Theorem 2.3.3. (Stinespring) Let A be a unital C

-algebra. A linear map u: A

B(H) is completely positive if and only if there is a Hilbert space K, a unital
-homomorphism : A B(K), and a bounded linear V : H K such that
u(a) = V

(a)V for all a A. This can be accomplished with |u|

cb
= |V |
2
. Also,
this equals |u|. If u is unital then we may take V to be an isometry; in this case
we may view H K, and we have u() = P
H
()
|H
.
Proof. The usual proof of this may be found in many places (e.g. [1, 8, 21]), and it
is very similar to the proof of the GNS construction from the C

-algebra course.
Thus we just give a sketch. Given a completely positive u, the idea to construct ,
as in the GNS construction proof, is to nd an inner product dened on a simple
space containing H on which A has a natural algebraic representation. In this case,
the space is A H, and we dene the representation of A by (a)(b ) = ab
for a, b A, H. We dene the inner product on AH by
a , b ) = T(b

a), ) , a, b A, , H.
The rest can be left as an exercise, following the model of the GNS construction.
Proposition 2.3.4. (A KadisonSchwarz inequality) If u: A B is a unital com-
pletely positive (or equivalently unital completely contractive) linear map between
unital C

-algebras, then u(a)

u(a) u(a

a), for all a A.

Proof. By 2.3.3 we have u = V

()V , with |V | 1 and a -homomorphism.

Thus u(a)

u(a) = V

(a)

V V

(a)V V

(a)

(a)V = u(a

a).
Corollary 2.3.5. Let u: A B be a completely isometric unital surjection be-
tween unital C

-algebras. Then u is a -isomorphism.

Proof. By 2.3.4 applied to both u and u
1
we have u(x)

u(x) u(x

x), and
u
1
(u(x)

u(x)) u
1
(u(x))

u
1
(u(x)) = x

x, for all x A. Applying u to

the last inequality gives u(x)

u(x) u(x

x). Hence u(x)

u(x) = u(x

x). Now
use the polarization identity (see ADDREF in the C

-algebra course), to conclude

that u(x)

u(y) = u(x

y) for x, y A. Setting y = 1 gives u(x)

= u(x

), and so
u(x

y) = u(x

)u(y). So u is a -isomorphism.
Proposition 2.3.6. Let u: A B be as in 2.3.4. Suppose that c A, and that c
satises u(c)

u(c) = u(c

c). Then u(ac) = u(a)u(c) for all a A.

2.3. OPERATOR SYSTEMS 65
Proof. Suppose that B B(H). We write u = V

()V as in Stinesprings theorem,

with V

V = I
H
. Let P = V V

be the projection onto V (H). By hypothesis

V

(c)

P(c)V = V

(c)

(c)V . For H, set = (c)V . Then |P|

2
=
V

(c)

P(c)V , ) = ||
2
. Thus P = , and V V

(c)V = (c)V . Therefore

u(a)u(c) = V

(a)V V

(c)V = V

(a)(c)V = u(ac).
2.3.7 (Completely positive bimodule maps) An immediate consequence of 2.3.6:
Suppose that u: A B is as in 2.3.4, and that there is a C

-subalgebra C of A
with 1
A
C, such that = u
|C
is a -homomorphism. Then
u(ac) = u(a)(c) and u(ca) = (c)u(a) (a A, c C).
We recall that a map is idempotent if = .
Theorem 2.3.8. (Choi and Eros) Suppose that A is a unital C

-algebra, and
that : A A is a unital, completely positive (or equivalently by 2.3.2, completely
contractive), idempotent map. Then we may conclude:
(1) R = Ran() is a C

-algebra with respect to the original norm, involution, and

vector space structure, but new product r
1

r
2
= (r
1
r
2
).
(2) (ar) = ((a)r) and (ra) = (r(a)), for r R and a A.
(3) If B is the C

-subalgebra of A generated by the set R, and if R is given the

product

, then
|B
is a -homomorphism from B onto R.
Proof. (2) By linearity and the fact that a positive map is -linear (see 2.3.1), we
may assume that a, r are selfadjoint. Set
d = d

=
_
0 r
r

a
_
.
Then
2
(d
2
) (
2
(d))
2
by the KadisonSchwarz inequality 2.3.4, so that
_
(r
2
) (ra)
(ar)
_

_
r
2
r(a)
(a)r
_
.
Here is used for a term we do not care about. Applying
2
gives
_
(r
2
) (ra)
(ar)
_

_
(r
2
) (r(a))
((a)r)
_
.
Thus
_
0 (ra) (r(a))
(ar) ((a)r)
_
0,
which implies that (ra) (r(a)) = 0 and (ar) ((a)r) = 0.
(1) By (2) we have for r
1
, r
2
, r
3
R that
(r
1

r
2
)

r
3
= ((r
1
r
2
)r
3
) = (r
1
r
2
r
3
).
Similarly, r
1

(r
2

r
3
) = (r
1
r
2
r
3
), which shows that the multiplication is
associative. It is easy to check that R (with original norm, involution, and vector
66 CHAPTER 2. OPERATOR SPACES
space structure, but new multiplication) satises the conditions necessary to be a
C

-algebra. For example:

(r
1

r
2
)

= (r
1
r
2
)

= (r

2
r

1
) = r

1
.
We check the C

-identity using the KadisonSchwarz inequality 2.3.4:

|r

r| = |(r

r)| |(r)

(r)| = |r

r| = |r|
2
,
and conversely,
|r|
2
= |r

r| |(r

r)| = |r

r|.
(3) This will follow if we can prove that (r
1
r
2
r
n
) = r
1

r
2

r
n
, for
r
i
R. This follows in turn by induction on n. Supposing that it is true for n = k,
we see that r
1

r
2

r
k+1
equals
((r
1

r
2

r
k
)r
k+1
) = ((r
1
r
2
r
k
)r
k+1
) = (r
1
r
2
r
k
r
k+1
),
using (2) in the last equality.
2.3.9 (The Paulsen system) If X is a subspace of B(H), we dene the Paulsen
system to be the operator system
o(X) =
_
CI
H
X
X

CI
H
_
=
__
x
y

_
: x, y X, , C
_
in M
2
(B(H)), where the entries and in the last matrix stand for I
H
and I
H
respectively. The following important lemma shows that as an operator system
(i.e. up to complete order isomorphism) o(X) only depends on the operator space
structure of X, and not on its representation on H.
Lemma 2.3.10. (Paulsen) Suppose that for i = 1, 2, we are given Hilbert spaces
H
i
, K
i
, and linear subspaces X
i
B(K
i
, H
i
). Suppose that u: X
1
X
2
is a linear
map. Let o
i
be the following operator system inside B(H
i
K
i
):
o
i
=
_
CI
H
i
X
i
X

i
CI
K
i
_
.
If u is contractive (resp. completely contractive, completely isometric), then
:
_
x
y

_

_
u(x)
u(y)

_
is positive (resp. completely positive and completely contractive, a complete order
injection) as a map from o
1
to o
2
.
Proof. Suppose that z is a positive element of o
1
. Thus
z =
_
a x
x

b
_
2.4. OPERATOR SPACE TENSOR PRODUCTS 67
where a and b are positive. Since z 0 if and only if z + 1 0 for all > 0, we
may assume that a and b are invertible. Then
_
a

1
2
0
0 b

1
2
_ _
a x
x

b
_ _
a

1
2
0
0 b

1
2
_
=
_
1 a

1
2
xb

1
2
b

1
2
x

1
2
1
_
0.
Hence by (2.22), we have that |a

1
2
xb

1
2
| 1. Applying u we obtain that
|a

1
2
u(x)b

1
2
| 1. Reversing the argument above now shows that (z) 0. So
is positive, and a similar argument shows that it is completely positive if u is
completely contractive. By 2.3.2 we have that is completely contractive in that
case. If in addition u is a complete isometry, then applying the above to u and u
1
we obtain the nal assertion.
Historical notes: This section is a slight variant of [4, Section 1.3]; historical
attributions are given there.
2.4 Operator space tensor products
2.4.1 (Minimal tensor product) Let X and Y be operator spaces, and let X Y
denote their algebraic tensor product. We recall from 3.2.2 and 3.2.3 in the C

-
course that any u =

n
k=1
x
k
y
k
XY can be associated with a map u : Y

X
dened by u() =

k
x
k
(y
k
), for Y

. If u =

n
k=1
x
k

k
X Y

then
u can be associated with a map u : Y X dened by u(y) =

k
x
k

k
(y), for
y Y . Both u and u are automatically completely bounded by 2.1.7, since they
are linear combinations of scalar functionals multiplied by xed operators. Thus
the above correspondences between tensor products and nite rank mappings yield
embeddings X Y CB(Y

, X) and X Y

CB(Y, X). The minimal tensor

product X
min
Y may then be dened to be (the completion of) X Y in the
matrix norms inherited from the operator space structure on CB(Y

, X) described
in 2.1.18. That is,
X
min
Y CB(Y

, X) completely isometrically. (2.23)

Explicitly, if u =

n
k=1
x
k
y
k
X Y , then the norm of u in X
min
Y equals
sup
_
_
_
_

k
x
k

ij
(y
k
)
__
_
_
M
m
(X)
, (2.24)
the supremum taken over all [
ij
] in the ball of M
m
(Y

), and all m N. Applying

(2.15) to (2.24), we see that |u|
min
equals the more symmetric form
sup
_
_
_
_

rs
_

k
x
k

ij
(y
k
)
___
_
_
M
ms
= sup
_
_
_
_

rs
(x
k
)
ij
(y
k
)
__
_
_
M
ms
, (2.25)
the supremum taken over [
rs
] and [
ij
] in the ball of M
s
(X

) and M
m
(Y

) re-
spectively, and all m, s N. A similar formula holds in M
n
(X
min
Y ):
|[w
rs
]|
M
n
(X
min
Y )
= sup
_
|[(
kl

ij
)(w
rs
)]|
_
(2.26)
68 CHAPTER 2. OPERATOR SPACES
for [w
rs
] M
n
(X Y ), where the supremum is taken over all [
rs
] and [
ij
] in the
ball of M
s
(X

) and M
m
(Y

) respectively, and all m, s N, and where

kl

ij
denotes the obvious functional on X Y formed from
kl
and
ij
.
We see from (2.26) that
min
is commutative, that is
X
min
Y = Y
min
X
as operator spaces. The underlying reason for this is because in the formulae
above we have
rs
(x)
ij
(y) =
ij
(y)
rs
(x).
It is also easy to see from (2.26) that
min
is functorial. That is, if X
i
and Y
i
are operator spaces for i = 1, 2, and if u
i
: X
i
Y
i
are completely bounded, then
the map x y u
1
(x) u
2
(y) on X
1
X
2
has a unique continuous extension
to a map u
1
u
2
: X
1

min
X
2
Y
1

min
Y
2
, with |u
1
u
2
|
cb
|u
1
|
cb
|u
2
|
cb
.
One way to see this is to note that if [
rs
] and [
ij
] are in the ball of M
s
(Y

1
)
and M
m
(Y

2
) respectively, for m, s N, then
1
u
1

cb
[
rs
u
1
] and
1
u
2

cb
[
ij
u
2
]
are in the ball of M
s
(X

1
) and M
m
(X

2
) respectively. Hence by (2.26) we have
1
|u
1
|
cb
|u
2
|
cb
|[(
kl

ij
)((u
1
u
2
)w
rs
)]| |[w
rs
]|
M
n
(X
1

min
X
2
)
,
for [w
rs
] M
n
(X
1
X
2
), and taking the supremum over [
rs
] and [
ij
], by (2.26)
again we have
1
|u
1
|
cb
|u
2
|
cb
|[(u
1
u
2
)w
rs
]|
M
n
(Y
1

min
Y
2
)
|[w
rs
]|
M
n
(X
1

min
X
2
)
.
Thus |u
1
u
2
|
cb
|u
1
|
cb
|u
2
|
cb
. (As an exercise, the reader could check that
|u
1
u
2
|
cb
= |u
1
|
cb
|u
2
|
cb
, but we shall not need this.)
If, further, the u
i
are completely isometric, then so is u
1
u
2
. This latter fact
is called the injectivity of the tensor product. To prove it, since u
1
u
2
=
(u
1
I) (I u
2
), we may by symmetry reduce the argument to the case that
Y
2
= X
2
, and u
2
= I
X
2
. Then it is easy to see that we can suppose that X
1
Y
1
and that u
1
is this inclusion map. In this case, consider the commutative diagram
CB(X

2
, X
1
) CB(X

2
, Y
1
)

X
1

min
X
2
u
1
I
Y
1

min
X
2
where the vertical arrows are complete isometries by denition of
min
, and the
top arrow is a complete isometry (since a map into a subspace of an operator
space clearly has the same norm as when it is viewed as a map into the bigger
space). Hence the bottom arrow is a complete isometry too, which is what we
need.
For any operator spaces X, Y , we have
X
min
Y

CB(Y, X) completely isometrically, (2.27)

2.4. OPERATOR SPACE TENSOR PRODUCTS 69
via the map : u u mentioned at the start of 2.4.1. We rst prove this in the
case that X = B(H). Consider the sequence of maps
B(H)
min
Y

CB(Y, B(H))

= w

CB(Y

, B(H)) CB(Y

, B(H)),
where the

= is from (2.18). The composition of these maps is the complete

isometry u u implementing (2.23). Since the last few maps in the sequence
are isometries so is the rst one.
For a general operator space X B(H) we have a commutative diagram
B(H)
min
Y

CB(Y, B(H))

X
min
Y

CB(Y, X)
where the left vertical arrow is a complete isometry by the injectivity of
min
,
and right one is obviously a complete isometry as we observed earlier. By the last
paragraph, the top arrow is an isometry, and so the bottom arrow is an isometry
too. We leave the proof that it is a complete isometry to the interested reader.
2.4.2 (The spatial tensor product and
min
) Suppose that H
1
, H
2
are Hilbert
spaces, and consider the canonical map : B(H
1
) B(H
2
) B(H
1

2
H
2
). This
is the map taking a rank one tensor S T in B(H
1
) B(H
2
) to the map on
H
1

2
H
2
denoted also by S T in 3.4.2 of the C

-course. We claim that actually

is a complete isometry when B(H
1
) B(H
2
) is given its norm as a subspace of
B(H
1
)
min
B(H
2
). To see this, we choose a set I such that H
1
=
2
I
, so that we
both have M
I

= B(H
1
) -isomorphically, and also H
1

2
H
2

= H
(I)
2
as Hilbert
spaces. By (2.27) and 2.2.6, B(H
1
)
min
B(H
2
) CB(S
1
(H
2
), M
I
). However, by
2.2.6 and 2.1.23 (5) and (12), we have
CB(S
1
(H
2
), M
I
)

= M
I
(S
1
(H
2
)

)

= M
I
(B(H
2
))

= B(H
(I)
2
)

= B(H
1

2
H
2
),
isometrically. A similar argument proves the complete isometry, and proves the
claim.
Thus if X and Y are subspaces of B(H
1
) and B(H
2
) respectively, then by the
injectivity of this tensor product, we have that X
min
Y is completely isometrically
isomorphic to the closure in B(H
1

2
H
2
) of the span of the operators x y on
H
1

2
H
2
, for x X, y Y . Thus the minimal tensor product of X Y may
alternatively be dened to be this subspace of B(H
1

2
H
2
).
The above implies that the minimal tensor product of C

-algebras coincides
with the tensor product of the same name used in C

-algebra theory, or with the

so-called spatial tensor product, that we studied in 3.4.3 in the C

-course. Indeed
note that if A and B are C

-subalgebras of B(H
1
) and B(H
2
) respectively, then
A
min
B may be identied with the closure of a subspace of B(H
1

2
H
2
), that
we saw in 3.4.3 in the C

-course, was actually a -subalgebra. Thus A

min
B is a
C

-algebra. If A and B are also commutative, then so is A

min
B, since it is the
closure of a commutative -subalgebra.
70 CHAPTER 2. OPERATOR SPACES
From the last paragraph it is clear that for any operator space X,
M
n

min
X

= M
n
(X) (2.28)
completely isometrically, since both can be completely isometrically identied with
the same subspace of B(
2
n
H)

= B(H
(n)
), if X B(H). Similarly, M
mn

min
X

=
M
mn
(X).
2.4.3 (Uncompleted tensor products) For what follows, it is convenient to state
separately a simple property of tensor product norms. If E and F are incomplete
spaces, and is a tensor norm on

E

F, then it is usual to write

E

F for the
completion of

E

F with respect to . We will always deal with so-called cross
norms; that is, (x y) = |x||y| for x E, y F. Let us write E

F for the
(possibly incomplete) subspace EF of

E

F, equipped with the norm . Claim:

F is the closure (and also the completion) of E

F. To see this, we need to

show that any u

E

F may be approximated in the norm topology by elements
in E F. However, such u may rst be approximated in norm by a nite sum of
elementary tensors x y, with x

E and y

F. Then we can approximate x y
in norm by x

, with x

E and y

F. Hence u is approximable by elements

in E

F.
2.4.4 (Further properties of
min
) For any set I we have
K
I

min
X

= K
I
(X). (2.29)
To see this, rst note that if X B(H), then by the injectivity of
min
we have
K
I

min
X M
I

min
B(H). By 2.4.2, the latter space can be identied with
a subspace of B(
2
I
H)

= B(H
(I)
)

= M
I
(B(H)) (see 2.1.23 (5)). On the other
hand, K
I
(X) is the closure of M
n
I
(X) in M
I
(B(H)). We can express this in the
commutative diagram
B(
2
I

2
H) M
I
(B(H))

M
I

min
B(H) M
I
(X)

M
n
I

min
X M
n
I
(X).
The complete isometry in the top row, restricts to a complete isometry in the bottom
row. Taking completions, and using 2.4.3, gives (completely isometrically)
K
I

min
X = M
n
I

min
X

= M
n
I
(X) = K
I
(X).
There is a rectangular variant of (2.29): for any sets I, J we have
K
I,J

min
X

= K
I,J
(X). (2.30)
2.4. OPERATOR SPACE TENSOR PRODUCTS 71
To see this, suppose that I has a bigger cardinality than J (the contrary case is
almost identical). Then we may regard K
I,J
K
I
and K
I,J
(X) K
I
(X). By the
injectivity of
min
we have a commutative diagram
K
I

min
X K
I
(X)

K
I,J

min
X K
I,J
(X).
The complete isometry in the top row coming from (2.29), restricts to a complete
isometry in the bottom row, proving (2.30).
Similarly, it follows from the second last paragraph, and from the fact that
B((H
1

2
H
2
)
2
H
3
)

= B(H
1

2
(H
2

2
H
3
)), that
min
is associative. That is,
(X
1

min
X
2
)
min
X
3
= X
1

min
(X
2

min
X
3
). (2.31)
To see this clearly, suppose that X
i
B(H
i
), and consider the commutative diagram
B((H
1

2
H
2
)
2
H
3
) B(H
1

2
(H
2

2
H
3
))

B(H
1

2
H
2
)
min
B(H
3
) B(H
1
)
min
B(H
2

2
H
3
)

(X
1

min
X
2
)
min
X
3
X
1

min
(X
2

min
X
3
).
The vertical arrows are complete isometries by 2.4.2 and the injectivity of
min
.
The -isomorphism in the top row, which is a complete isometry, restricts to a
complete isometry in the bottom row. Taking completions, and using 2.4.3, gives
(completely isometrically)
(X
1

min
X
2
)
min
X
3
= (X
1

min
X
2
)
min
X
3

= X
1

min
(X
2

min
X
3
),
which equals X
1

min
(X
2

min
X
3
). This proves the associativity. Accordingly, the
space in (2.31) will be merely denoted by X
1

min
X
2

min
X
3
, and the proof above
shows that it can be identied with a subspace of B(H
1

2
H
2

2
H
3
). Similarly,
one may consider the N-fold minimal tensor product X
1

min

min
X
N
of any
N-tuple of operator spaces.
Proposition 2.4.5. Let E, F be Banach spaces and let X be an operator space.
(1) Min(E)
min
X = E

X as Banach spaces.
(2) Min(E)
min
Min(F) = Min(E

F) as operator spaces.
Proof. We have isometric embeddings Min(E)
min
X CB(X

, Min(E)) and
E

X B(X

, E) by (2.23) and (3.2) in the C

-course. However CB(X

, Min(E)) =
B(X

, E) by (2.8). Thus both spaces in (1) coincide isometrically with the same
subspace of B(X

, E), which proves (1). The isometry in (2) follows from (1).
Thus the complete isometry in (2) will follow if Min(E)
min
Min(F) is a minimal
72 CHAPTER 2. OPERATOR SPACES
operator space. To see this, suppose that E C(K
1
) and F C(K
2
) isometri-
cally. Then Min(E) C(K
1
) and Min(F) C(K
2
) completely isometrically. So
by the injectivity of
min
, we have that Min(E)
min
Min(F) is completely isomet-
rically contained inside C(K
1
)
min
C(K
2
) completely isometrically. However, we
observed in 2.4.2 that the minimal tensor product of commutative C

-algebras is
a commutative C

-algebra, and hence is a C(K)-space, and is a minimal operator

space.
2.4.6 (Haagerup tensor product) Before we dene this tensor product, we intro-
duce an intimately related class of bilinear maps. Suppose that X, Y , and W are
operator spaces, and that u: X Y W is a bilinear map. For n, p N, dene a
bilinear map M
n,p
(X) M
p,n
(Y ) M
n
(W) by
(x, y)
_
p

k=1
u(x
ik
, y
kj
)
_
i,j
,
where x = [x
ij
] M
n,p
(X) and y = [y
ij
] M
p,n
(Y ). If p = n we write this
map as u
n
. If the norms of these bilinear maps (as dened in ADDREF in the C

-
algebra notes) are uniformly bounded over p, n N, then we say that u is completely
bounded, and write the supremum of these norms as |u|
cb
. Sometimes this is called
completely bounded in the sense of Christensen and Sinclair. It is easy to see (by
adding rows and columns of zeroes to make p = n) that |u|
cb
= sup
n
|u
n
|. (Indeed,
if [x
ij
] M
n,p
(X) and [y
ij
] M
p,n
(Y ), and if m = maxn, p, then
_
_
_
_
p

k=1
u(x
ik
, y
kj
)
__
_
_
n
= |u
m
([x

ij
], [y

ij
])|
m
|u
m
||[x
ij
]||[y
ij
]|,
where [x

ij
] and [y

ij
] are m m matrices obtained from [x
ij
] and [y
ij
] by adding
rows or columns of zeros.)
We say that u is completely contractive if |u|
cb
1. Completely bounded
multilinear maps of three variables have a similar denition (involving the expression
[

k,l
u(x
ik
, y
kl
, z
lj
)]), and similarly for four or more variables. We remark that if
v : X B(H) and w: Y B(H) are completely bounded linear maps, then it is
easy to see that the bilinear map (x, y) v(x)w(y) is completely bounded in the
sense above, and has completely bounded norm dominated by |v|
cb
|w|
cb
. Indeed,
note that
_
_
_
_
n

k=1
v(x
ik
)w(y
kj
)
__
_
_
n
|[v(x
ij
)]|
n
|[w(y
ij
)]|
n
|v|
cb
|w|
cb
|[x
ij
]|
n
|[y
ij
]|
n
,
if [x
ij
] M
n
(X) and [y
ij
] M
n
(Y ), since M
n
(B(H)) is a Banach algebra.
Let X, Y be operator spaces. For n N and z M
n
(X Y ) we dene
|z|
h
= inf|x||y|, (2.32)
where the inmum is taken over all p N, and all ways to write z = x y, where
x M
n,p
(X), y M
p,n
(Y ). Here x y denotes the formal matrix product of x
2.4. OPERATOR SPACE TENSOR PRODUCTS 73
and y using the sign as multiplication: namely x y = [

p
k=1
x
ik
y
kj
]. To
make sense of this, we rst note that any z M
n
(X Y ) can be written as such
a x y. To see this we observe that this is clearly true if z has only one nonzero
entry. For example, if this entry were the 1-2 entry, and if z
12
=

p
k=1
x
k
y
k
, then
z = xy where x M
np
(X) has rst row consisting of the x
k
and zeros elsewhere,
and y M
pn
(Y ) has second column consisting of the y
k
and zeros elsewhere. Next
note that
x y +x

= [x : x

]
_
y
y

_
,
for matrices x, x

, y, y

of appropriate sizes. Similarly for a sum of any (nite)

number of terms of the form x y. Thus by writing z M
n
(X Y ) as a sum of
n
2
matrices, each of which has only one nonzero entry, and using the facts above,
we do indeed have z = x y as desired.
It is clear that |z|
h
= [[|z|
h
if C. Next note that the last centered
equation actually shows that |z +z

|
h
|z|
h
+|z

|
h
for z, z

M
n
(X Y ). For
suppose that z = x y and z

= x

, with |x||y| < |z|

h
+ and |x

||y

| <
|z

|
h
+ . By the trick of writing x y = tx
1
t
y with t =
_
|y|/|x|, we can
assume that |y| = |x|. Similarly, assume that |y

| = |x

|. Then
|z +z

|
h
|[x : x

]|
_
_
_
_
y
y

_
_
_
_
_
|x|
2
+|x

|
2
_
|y|
2
+|y

|
2
,
the last following from the C

-identity used four times. For example,

|[x : x

]|
2
= |xx

+x

| |xx

| +|x

| = |x|
2
+|x

|
2
.
Thus
|z +z

|
h
|x||y| +|x

||y

| |z|
h
+|z

|
h
+ 2.
Now let 0, to see that | |
h
is a seminorm.
Suppose that u: XY W be a bilinear map which is completely contractive
in the sense above. Let u: X Y W be the canonically associated linear map.
For z M
n
(X Y ), if z = x y as above, then
| u
n
|
n
=
_
_
_
_
p

k=1
u(x
ik
, y
kj
)
_
|x||y|.
Taking the inmum over x, y with z = x y, we see by the denition of | |
h
that
| u
n
(z)| |z|
h
, (2.33)
where the latter quantity is as dened in (2.32). If and are contractive function-
als on X and Y respectively, then using 2.1.7 and the fact at the end of the second
paragraph of 2.4.6, we see that the bilinear map (x, y) (x)(y) is completely
contractive. Thus from (2.33) we see that

k=1
(x
k
)(y
k
)

|z|
h
, z =
p

k=1
x
k
y
k
.
74 CHAPTER 2. OPERATOR SPACES
By the denition of the Banach space injective tensor norm of z (see 3.2.2 in the
C

-course), we deduce that the latter norm of an element z X Y is dominated

by |z|
h
. Hence indeed | |
h
is a norm.
Proposition 2.4.7. If x and Y are operator spaces, then the completion X
h
Y
of X Y with respect to | |
h
is an operator space.
Proof. We use Ruans theorem, in the form of Exercise (6) to Section 2.1. To see
(R1), suppose that M
m,n
, M
n,m
, z M
n
(X Y ) with z = x y as above.
Then z = (x) (y), and so
|z|
h
|x||y| |||x||y|||.
Taking the inmum over x, y with z = x y, we see by denition of | |
h
that
|z|
h
|||||z|
h
.
For (R2), let z

= x

M
p
(X Y ), then z z

= (x x

) (y y

), and
|z z

|
h
|x x

||y y

| = max|x|, |x

| max|y|, |y

|.
As in the proof that | |
h
is a norm, we can assume that |x| = |y| and |x

| =
|y

|. Then |z z

|
h
max|x||y|, |x

||y

|, and taking the inmum over such

x, x

, y, y

gives (R2). Note that (R1) and (R2) pass to the completion of X Y .
So X
h
Y is an operator space.
This operator space X
h
Y is called the Haagerup tensor product. Note that
the canonical bilinear map : X Y X
h
Y is completely contractive in the
sense above.
Using (2.33) we see that if u: X Y W is a bilinear map with associated
linear map u: X Y W, then u is completely bounded if and only if u extends
to a completely bounded linear map on X
h
Y . Moreover we have
|u|
cb
=
_
_
u: X
h
Y W
_
_
cb
in that case. The above property means that the Haagerup tensor product linearizes
completely bounded bilinear maps. A moments thought shows that this is a universal
property. That is, suppose that (W, ) is a pair consisting of an operator space W,
and a completely contractive bilinear map : X Y W, such that the span of
the range of is dense in W, and which possesses the following property:
Given any operator space Z and given any completely bounded bilinear map
u: X Y Z, then there exists a linear completely bounded u: W Z such
that u((x, y)) = u(x, y) for all x X, y Y , and such that | u|
cb
|u|
cb
.
Then X
h
Y

= W via a complete isometry v satisfying v = .
We leave it to the reader to check the above assertions as an exercise.
2.4.8 (More properties of the Haagerup tensor product)
2.4. OPERATOR SPACE TENSOR PRODUCTS 75
Since X
h
Y is an (uncompleted) operator space, there is a canonical norm
on M
m,n
(X
h
Y ), via viewing this space as a subspace of M
p
(X
h
Y ), for
p = maxm, n. It is easy to see that for z M
m,n
(X
h
Y ), this canonical norm
is still given by the formula (2.32), however with x M
m,p
(X), y M
p,n
(X)
there. There is a canonical linear isomorphism between C
m
(X) R
n
(Y ) and
M
m,n
(X Y ), taking [x
i
] [y
i
] [x
i
y
j
]
(i,j)
. Using the denition (2.32)
it is a very easy exercise to show that this isomorphism is actually an isometry
C
m
(X)
h
R
n
(Y )

= M
m,n
(X
h
Y ). Passing to the completion, we have C
m
(X)
h
R
n
(Y )

= M
m,n
(X
h
Y ) isometrically.
This tensor product is functorial. That is, if u
i
: X
i
Y
i
are completely bounded
maps between operator spaces, then u
1
u
2
: X
1

h
X
2
Y
1

h
Y
2
is completely
bounded, and |u
1
u
2
|
cb
|u
1
|
cb
|u
2
|
cb
. Indeed, if z = x y M
n
(X Y ),
then (u
1
u
2
)
n
(z) = (u
1
)
n
(x) (u
2
)
n
(y), and so
|(u
1
u
2
)
n
(z)|
h
|(u
1
)
n
(x)||(u
2
)
n
(y)| |u
1
|
cb
|u
2
|
cb
|x||y|.
Taking the inmum over such x, y with z = x y gives
|(u
1
u
2
)
n
(z)|
h
|u
1
|
cb
|u
2
|
cb
|z|
h
.
Thus u
1
u
2
is continuous on X
1

h
X
2
, and extends uniquely to u
1
u
2
:
X
1

h
X
2
Y
1

h
Y
2
satisfying |u
1
u
2
|
cb
|u
1
|
cb
|u
2
|
cb
.
The Haagerup tensor product is projective, that is, if u
1
and u
2
in the last item
are complete quotient maps, then so is u
1
u
2
. To see this, note that by the
functoriality, the map u
1
u
2
is a complete contraction. Let z M
n
(Y
1
Y
2
),
with |z|
h
< 1. By denition, we may write z = y
1
y
2
, where y
1
M
n,p
(Y
1
),
y
2
M
p,n
(Y
2
) both have norm < 1. Then y
1
= (u
1
)
n,p
(x
1
) and y
2
= (u
2
)
p,n
(x
2
)
for x
1
M
n,p
(X
1
), x
2
M
p,n
(X
2
), both of norm < 1. Let w = x
1
x
2

M
n
(X
1

h
X
2
), this matrix has norm < 1, and (u
1
u
2
)
n
(w) = z. By an obvious
density argument, this shows that u
1
u
2
above is a complete quotient map.
The Haagerup tensor product is not commutative. That is, in general X
h
Y
and Y
h
X are not isometric. We shall see some examples of this later.
The Haagerup tensor product is associative. That is,
(X
1

h
X
2
)
h
X
3

= X
1

h
(X
2

h
X
3
)
completely isometrically. To see this, we rst show it for the uncompleted Haagerup
tensor product, where there is an obvious algebraic linear isomorphism : (X
1

X
2
) X
3
X
1
(X
2
X
3
). If z M
n
((X
1

h
X
2
) X
3
) with |z|
h
< 1 then
z = uw, where u M
n,p
(X
1

h
X
2
) and w M
p,n
(X
3
), both of norm < 1. By
the rst few lines in 2.4.8, we have u = xy for some x M
n,k
(X), y M
k,p
(Y ),
both of norm < 1. But then
n
(z) = x (y z), and hence it is easy to see that
|
n
(z)|
h
< 1. So is a complete contraction, and similarly
1
is a complete
contraction. So is a complete isometry. Taking the completion, just as in the
proof of the associativity of
min
, gives the associativity of
h
. Accordingly, the
three-fold tensor product in the last displayed equation will be merely denoted
by X
1

h
X
2

h
X
3
. The induced norms on M
n
(X
1
X
2
X
3
) may be described
76 CHAPTER 2. OPERATOR SPACES
by the 3-variable version of (2.32). From this one may see that X
1

h
X
2

h
X
3
has the universal property of linearizing completely bounded trilinear maps (see
discussion at the end of 2.4.6). Similar assertions clearly hold for the N-fold
Haagerup tensor product X
1

h

h
X
N
of any N-tuple of operator spaces.
There are convenient norm expressions for | |
h
. Suppose that A and B are
C

-algebras. If X and Y are subspaces of A and B respectively, and if z XY ,

then to say that z = x y, is simply to say that z =

p
k=1
a
k
b
k
, where a
k
is the kth entry in the row matrix x, and b
k
is the kth entry in the column
matrix y. By the C

-identity,
|x|
2
= |xx

| =
_
_
_
p

k=1
a
k
a

k
_
_
_.
Similarly, |y|
2
=
_
_
_

p
k=1
b

k
b
k
_
_
_. Thus by the denition in 2.4.6 we have
|z|
h
= inf
_
_
_
p

k=1
a
k
a

k
_
_
_
1
2
_
_
_
p

k=1
b

k
b
k
_
_
_
1
2
(2.34)
where the inmum is taken over all ways to write z =

p
k=1
a
k
b
k
in X Y .
The following shows that the last formula extends to the completed Haagerup
tensor product X
h
Y , replacing p by in (2.34).
Proposition 2.4.9.
(1) If z X
h
Y with |z|
h
< 1 then we may write z as a norm convergent sum

k=1
a
k
b
k
in X
h
Y , with |

k=1
a
k
a

k
| < 1 and |

k=1
b

k
b
k
| < 1, and
where the last two sums converge in norm. That is, [a
1
a
2
] R(X) and
[b
1
b
2
]
t
C(Y ).
(2) If x = [a
1
a
2
] R(X) and y = [b
1
b
2
]
t
C
w
(Y ), that is if

k=1
a
k
a

k
converges in norm and if the partial sums of

k=1
b

k
b
k
are uniformly bounded
in norm, then

k=1
a
k
b
k
converges in norm in X
h
Y . Similarly if x R
w
(X)
and y C(Y ).
Proof. (1) If z is as stated, choose w
1
XY with |z w
1
|
h
<

2
and |w
1
|
h
< 1.
By (2.34) we may write w
1
=

n
1
k=1
x
k
y
k
with

k
x
k
x

k
1 and

k
y

k
y
k
1.
Repeating this argument, we may choose w
2
X Y with |z w
1
w
2
|
h
<

2
2
,
and |w
2
|
h
<

2
. By (2.34) we write w
2
=

n
2
k=n
1
+1
x
k
y
k
with

k
x
k
x

k


2
and

k
y

k
y
k


2
. Continuing so, we obtain for every m N a nite rank tensor
w
m
=

n
m
k=n
m1
+1
x
k
y
k
with |zw
1
. . .w
m
| <

2
m
,

n
m
k=n
m1
+1
x
k
x

k


2
m1
,
and

n
m
k=n
m1
+1
y

k
y
k


2
m1
. Now it is clear that the partial sums of

k=1
x
k
x

k
and

k=1
y

k
y
k
are Cauchy. For example, for any j > i n
m1
+ 1 we have
|
j

k=i
x
k
x

k
| |

k=n
m1
+1
x
k
x

k
|

k=m1

2
k
=

2
m
0
2.4. OPERATOR SPACE TENSOR PRODUCTS 77
with m. Hence

k=1
x
k
x

k
and

k=1
y

k
y
k
converge in norm to elements with
norm 1 + . Also, the partial sums of

k=1
x
k
y
k
are Cauchy, so that that
sum converges in norm (see (2) below). Since a subsequence of these partial sums
converges to z, by the rst displayed equation in the proof, we have z =

k=1
x
k
y
k
as desired.
(2) To see that the partial sums of

k=1
a
k
b
k
are Cauchy, note that from
(2.34) we have |

m
k=n
a
k
b
k
|
h
|

m
k=n
a
k
a

k
|
1
2
|

m
k=n
b

k
b
k
|
1
2
. Now use the
fact that the partial sums of

k=1
a
k
a

k
are Cauchy.
The Haagerup tensor product is injective (Theorem 2.4.10 below). In order to
establish this, we will need a simple linear algebraic fact about tensors z E F.
Suppose that X is a closed subspace of E, with z X F E F, and suppose
also that z =

n
k=1
x
k
y
k
, with y
k
a linearly independent subset of F. Then
we claim that x
k
X for all k = 1, , n. To prove this, choose by the Hahn-
Banach theorem functionals
k
F

with
k
(y
i
) =
i,k
, Kroneckers delta. Then
(I
E

k
)(z) = x
k
. However, since z X Y we must have (I
E

k
)(z) X. So
x
k
X.
Theorem 2.4.10. If u
i
: X
i
Y
i
are completely isometric maps between operator
spaces, then u
1
u
2
: X
1

h
X
2
Y
1

h
Y
2
is a complete isometry.
Proof. We may assume that X
i
Y
i
, and u
i
is the inclusion. By a two-step
argument, as in the proof of the injectivity of
min
, we can assume that X
2
= Y
2
,
and u
2
is the identity map. Also, it is enough to prove the result for the uncompleted
tensor products. Of course u
1
u
2
is a complete contraction, by the functoriality
of
h
. To prove that u
1
u
2
is an isometry, it suces to show that if z X
1
X
2
,
and that if z viewed as an element of Y
1
Y
2
has Haagerup norm < 1, then
z Ball(X
1

h
X
2
). Thus suppose that z X
1
X
2
, with z = xy =

n
k=1
x
k
y
k
,
where x = [x
k
] R
n
(Y
1
), y = [y
k
] C
n
(X
2
), with |x||y| < 1. If b
k
y
k
is
a basis for Spany
k
, and if b = [b
k
] C
m
(Y
1
), then there is a matrix of scalars
with y = b. One can in fact choose so that its rst few rows form a copy of the
identity matrix. Let = u be a polar decomposition of , where = (

)
1
2
, and
u is an isometry. In fact, it is a simple linear algebraic exercise to see that

I,
so that is invertible. Then z = x y = xu b, and
|xu||b| |x||u

ub| = |x||u

y| < 1.
Since is invertible, the entries of b are linearly independent. By the fact above
the theorem, xu R
m
(X
1
). Thus indeed z Ball(X
1

h
X
2
).
Finally, to see that u
1
I is a complete isometry, we note that by the last
paragraph we have that C
n
(X
1
)
h
R
n
(X
2
) C
n
(Y
1
)
h
R
n
(X
2
) isometrically.
By the last part of the rst bullet in 2.4.8, we conclude that M
n
(X
1

h
X
2
)
M
n
(Y
1

h
X
2
) isometrically. That is, (u
1
I)
n
is an isometry, so that u
1
I is a
complete isometry.
2.4.11 (Operator space projective tensor product) As with the Haagerup tensor
product, it is convenient to rst dene an intimately related class of bilinear maps.
78 CHAPTER 2. OPERATOR SPACES
Suppose that X, Y , and W are operator spaces and that u: X Y W is a
bilinear map. We say that u is jointly completely bounded if there exists a constant
K 0 such that
|[u(x
ij
, y
kl
)]
(i,k),(j,l)
| K|[x
ij
]||[y
kl
]|
for all m, n and [x
ij
] M
n
(X), and [y
kl
] M
m
(Y ). Here, as usual, the matrix
is indexed on rows by i and k, and on columns by j and l. The least such K is
written as |u|
jcb
. We say that u is jointly completely contractive if |u|
jcb
1.
Jointly completely bounded multilinear maps of three or more variables are dened
similarly. Any completely contractive (in the sense of 2.4.6) bilinear map u is jointly
completely contractive. This is immediate from the simple relation [u(x
ij
, y
kl
)] =
u
nm
([x
ij
] I
m
, I
n
[y
kl
]), where u
mn
is as dened at the start of 2.4.6. Indeed, for
[x
ij
] M
n
(X), and [y
kl
] M
m
(Y ) we have
|[u(x
ij
, y
kl
)]| = |u
nm
([x
ij
]I
m
, I
n
[y
kl
])| |[x
ij
]I
m
||I
n
[y
kl
]| = |[x
ij
]||[y
kl
]|.
Conceptually, the simplest way to dene the operator space projective tensor
product X

Y of two operator spaces X and Y , is to identify it (completely

isometrically) with a subspace of CB(X, Y

, via the map that takes xy to the

functional T T(x)(y) on CB(X, Y

). This gives XY an (incomplete) operator

space structure, and the completion is an operator space completely isometric to
a subspace of CB(X, Y

. One then can immediately verify results like (2.36),

(2.37), and (2.38) below. This was the approach taken in [6], and the reader might
like to try these as an exercise. However this approach does not yield the following
explicit internal formula for the tensor norm: For z M
n
(X Y ) and n N,
dene
|z|

= inf|||x||y|||, (2.35)
the inmum taken over p, q N, and all ways to write z = (x y), where
M
n,pq
, x M
p
(X), y M
q
(Y ), and M
pq,n
. Here we wrote x y for the
tensor product of matrices, namely xy = [x
ij
y
kl
]
(i,k),(j,l)
, indexed on rows by i
and k, and on columns by j and l. Suppose that z M
n
(XY ), and that we have
a jointly completely contractive bilinear map u: X Y W. Let u : X Y W
be the associated linear map. Write z = (x y) as above. A simple calculation
shows that
| u
n
(z)| = |[u(x
ij
, y
kl
)]| |||[u(x
ij
, y
kl
)]||| |||[x
ij
]||[y
kl
)]|||.
Taking the inmum over such ways to write z, we see from the denitions that
| u
n
(z)| |z|

. (2.36)
From the observation at the end of the rst paragraph of 2.4.11, it follows that this
is also true if u is completely contractive. Taking u = : X Y X
h
Y , in
which case u is the identity map, we deduce from (2.36) that |z|
h
|z|

.
We leave it as an exercise similar to the analogous statement for the Haagerup
tensor product (or see [15]), that the quantities in (2.35) dene an operator space
structure on X Y (to see that these are norms as opposed to seminorms, use the
2.4. OPERATOR SPACE TENSOR PRODUCTS 79
fact at the end of the last paragraph). Thus the completion of XY with respect to
these matrix norms is an operator space, which we call the operator space projective
tensor product, and write as X

Y .
By (2.36) we see that if u: X Y W is a bilinear map with associated linear
map u: XY W, and if u is jointly completely contractive, then u is completely
contractive with respect to | |

, and extends further to a complete contraction

u : X

Y W. Conversely, if v : X

Y W is completely contractive,
and if u: X Y W is the associated bilinear map, then u is jointly completely
contractive. To see this, note that if [x
ij
] Ball(M
n
(X)), [y
kl
] Ball(M
m
(Y )),
then [x
ij
y
kl
] Ball(M
mn
(X

Y )) (take = = I
mn
in (2.35)). Thus
|[u(x
ij
, y
kl
)]| = |v
mn
([x
ij
y
kl
])| 1.
Writing JCB(X, Y ; W) for the space of jointly completely bounded maps, we have
shown that JCB(X, Y ; W)

= CB(X

Y, W) isometrically, via the canonical map.

In other words, the operator space projective tensor product linearizes jointly com-
pletely bounded bilinear maps. As for the Haagerup tensor product this is a universal
property, and characterizes X

Y uniquely up to complete isometry.

If u: X Y W is bilinear and jointly completely bounded, write u
#
for the
map from X to the set of functions from Y to W dened by
u
#
(x)(y) = u(x, y), x X, y Y.
Then u
#
CB(X, CB(Y, W)): indeed,
|u
#
|
cb
= sup|[u
#
(x
ij
)]| : [x
ij
] Ball(M
n
(X)), n N
= sup|[u
#
(x
ij
)(y
kl
)]| : [x
ij
] Ball(M
n
(X)), [y
kl
] Ball(M
m
(Y ))
= sup|[u(x
ij
, y
kl
)]| : [x
ij
] Ball(M
n
(X)), [y
kl
] Ball(M
m
(Y ))
= |u|
jcb
.
Conversely, if v CB(X, CB(Y, W)) and if u(x, y) = v(x)(y), then reversing
the last argument shows that u is jointly completely bounded. Indeed, we have
shown that CB(X, CB(Y, W))

= JCB(X, Y ; W)

= CB(X

Y, W) isometri-
cally via the canonical map. In fact this is a complete isometry, as may be seen
by the common trick of replacing W by M
n
(W) in the isometric identity, thus
CB(X

Y, M
n
(W))

= CB(X, CB(Y, M
n
(W)). Using (2.5) we then have a string
of isometries
M
n
(CB(X

Y, W))

= CB(X

Y, M
n
(W))

= CB(X, CB(Y, M
n
(W)))

= CB(X, M
n
(CB(Y, W)))

= M
n
(CB(X, CB(Y, W))).
The isometry which is the composition of all these isometries is easily seen to
just be the nth amplication of the map CB(X

Y, W) CB(X, CB(Y, W))

above, and hence the latter is a complete isometry. A similar argument works for
CB(Y, CB(X, Z)), and thus we have
CB(X

Y, Z)

= CB(X, CB(Y, Z))

= CB(Y, CB(X, Z)) (2.37)
80 CHAPTER 2. OPERATOR SPACES
completely isometrically. In particular,
(X

Y )

= CB(X, Y

)

= CB(Y, X

) completely isometrically. (2.38)

Corollary 2.4.12. For any operator spaces X, Y , the space CB(X, Y

) is a dual
operator space, with predual X

Y .
We now list a sequence of properties of the operator space projective tensor prod-
uct. We leave it as an exercise, copying the analoguous proofs in 2.4.8, that

is
functorial, and projective. We use these words in the sense that we have used them
for the other two tensor products. We show next that

is commutative, that is,
X

Y

= Y

X completely isometrically. To see this we will use Exercise (9) to

Section 2.1. Indeed if : X Y Y X, then it is easy to check that the map
on (Y

X)

is exactly the composition of the complete isometries in the

sequence
(Y

X)

= CB(X, Y

)

= (X

Y )

provided by (2.38).
To show that

is associative, that is, (X

Y )

Z

= X

(Y

Z) completely
isometrically, two methods come to mind. First, one could show that each of these
two spaces has the universal property of linearizing jointly completely bounded
trilinear maps, and hence they must be the same. A second method is to mimic the
proof just given for commutativity, since these spaces have the same duals:
((X

Y )

Z)

= CB(X

Y, Z

)

= (X, CB(Y, Z

)),
and
(X

(Y

Z))

= CB(X, (Y

Z)

)

= CB(X, CB(Y, Z

)),
using (2.37) several times.
Proposition 2.4.13. Let E, F be Banach spaces and let Y be an operator space.
(1) Max(E)

Y = E

Y isometrically.
(2) Max(E

F) = Max(E)

Max(F) completely isometrically.
Proof. The rst item follows as in the proof of the commutativity of

above,
by computing the duals of these two tensor products, using (2.38), (2.11), and
ADDREF from the C

-course:
(Max(E)

Y )

= CB(Max(E), Y

) = B(E, Y

)

= (E

Y )

.
Then (2) follows from (1) if we can show that Max(E)

Max(F) is a maximal oper-
ator space, or equivalently that any contractive map on it is completely contractive.
To do this, observe that
B(Max(E)

Max(F), W) = B(E

F, W) = B(E, B(F, W)),

2.4. OPERATOR SPACE TENSOR PRODUCTS 81
isometrically for any operator space W, by (1) and ADDREF from the C

-course.
The latter space equals
CB(Max(E), CB(Max(F), W)) = CB(Max(E)

Max(F), W)
by (2.11) and (2.37). Thus Max(E)

Max(F) is maximal.
Proposition 2.4.14. (Comparison of tensor norms) If X and Y are operator
spaces then the various tensor norms on X Y are ordered as follows:
| |

| |
min
| |
h
| |

| |

.
Indeed the identity is a complete contraction X

Y X
h
Y X
min
Y .
Proof. The rst inequality follows easily for example from (2.26) and the denition
of

in the C

-course. The fact that | |

| |

follows from the universal

property of

in the C

-course. Indeed, since the bilinear map : XY X

Y
is jointly completely contractive, and hence contractive, it induces a contraction
X

Y X

Y . We saw in 2.4.11 the complete contraction X

Y X
h
Y .
For the remaining relation, consider unital C

-algebras A and B containing X

and Y respectively, and the following commutative diagram of uncompleted tensor
products
A
h
B A
min
B

X
h
Y X
min
Y,
where the horizontal arrows are the identity map, and the vertical arrows are com-
plete isometries by the injectivity of these tensor norms. Thus the bottom arrow
will be a complete contraction if the top one is. Now A
min
B is a C

-algebra as
observed in 2.4.2. Moreover the maps
: A A
min
B : a a 1 , : B A
min
B : b 1 b
are (completely contractive) -homomorphisms. The bilinear map (a, b) (a)(b) =
a b from A B to A
min
B is therefore completely contractive in the sense of
2.4.6. By the universal property of
h
, this bilinear map induces a linear complete
contraction from A
h
B to A
min
B. But the latter map is clearly the identity
map, which proves the desired relation.
Proposition 2.4.15. If X, Y are operator spaces, if H, K are Hilbert spaces, and
if m, n N, then we have the following complete isometries:
(1) H
r

h
X = H
r

X, and X
h
H
c
= X

H
c
.
(2) H
c

h
X = H
c

min
X, and X
h
H
r
= X
min
H
r
.
(3) C
n
(X)

= C
n

h
X = C
n

min
X, and R
n
(X)

= X
h
R
n
= X
min
R
n
.
(4) (

H
r

X

K
c
)

= (

H
r

h
X
h
K
c
)

= CB(X, B(K, H)).
(5) S

(K, H)

= H
c

min

K
r
and S

(K, H)
min
X

= H
c

h
X
h

K
r
.
82 CHAPTER 2. OPERATOR SPACES
(6) M
m,n
(X)

= C
m

h
X
h
R
n
.
(7) M
m,n
(X
h
Y )

= C
m
(X)
h
R
n
(Y ).
(8) H
c

K
c
= H
c

h
K
c
= H
c

min
K
c
= (H
2
K)
c
, and similarly for row
Hilbert spaces.
(9) S
1
(K, H)

=

K
r

H
c
.
(10) CB(S
1
(
2
I
,
2
J
), X)

= M
I,J
(X), if I, J are sets.
Proof. We will prove (1) last, although we use it to prove some of the others. To
prove (3), note that by (2.30) we have C
n

min
X

= C
n
(X). By the last proposition
there is a complete contraction C
n

h
X C
n

min
X

= C
n
(X). The inverse of
the latter map is the map u : C
n
(X) C
n

h
X dened by u(

x) =

n
k=1

e
k
x
k
,
where

x = [x
k
] C
n
(X). By denition of the Haagerup tensor norm, namely
(2.32), we have
_
_
_
n

k=1

e
k
x
k
_
_
_
h
|I
n
||

x| = |

x|.
Thus u is a contraction, and a similar argument at the matrix level shows that it is
a complete contraction. Thus C
n

h
X

= C
n
(X). A similar argument proves the
other relation in (3).
It suces to prove (2) for the uncompleted tensor products: H
c

h
X = H
c

min
X. Let us examine the norm on both sides. If z =

m
k=1

k
x
k
H
c
X, let
K = Span(
k
) H
c
. By the injectivity of
h
(resp.
min
) the norm |z|
h
(resp.
|z|
min
) is the same whether computed in KX or in H
c
X. Thus we can assume
that H is nite dimensional. A similar argument lets us make this same assumption
if z M
n
(H
c
X). Now K is a Hilbert column space, and is isometrically, and
hence completely isometrically by (2.13), isomorphic to C
n
for some n N. By (3),
we have
K
h
X

= C
n

h
X

= C
n

min
X

= K
min
X.
It is clear from the above discussion that we now have proved (2).
The rst equality in (4) is clear from (1), and the rest is clear from the complete
isometries
(

H
r

X

K
c
)

= CB(X

K
c
, H
c
)

= CB(X, CB(K
c
, H
c
)),
which equals CB(X, B(K, H)). Here we have used (2.38), (2.37), (2.14), and (2.13).
For the rst equality in (5), note that the canonical map X Y B(Y

, X)
has range which is precisely the set of nite rank operators. Thus the canonical
complete isometry
H
c

min

K
r
CB((

H
r
)

, H
c
)

= CB(K
c
, H
c
) = B(K, H)
has range that is the closure of the set of nite rank operators (we have used (2.23),
(2.14), and (2.13) here). But this closure in B(K, H) is S

(K, H). For the second

equality, note that S

(K, H)
min
X = H
c

min
X
min

K
r
by commutativity of

min
and the rst part of (5), and so the second part of (5) follows from (2).
2.4. OPERATOR SPACE TENSOR PRODUCTS 83
Item (6) follows from (5) and (2.30), and (7) follows from (6) by (3) and the
associativity of the Haagerup tensor product:
C
m
(X)
h
R
n
(Y )

= C
m

h
(X
h
Y )
h
R
n

= M
m,n
(X
h
Y ).
The middle equality in (8) follows from (2), and the rst equality from (1). Writing
H
c
= C
J
and

K
r
= R
I
for sets I, J, we have H
2
K =
2
(I J), so that
(H
2
K)
c
= C
IJ
. Then the last equality in (8) may be seen from (2.30):
H
c

min
K
c
= C
I

min
C
J
= C
I
(C
J
) = C
IJ
.
To see (9) note that (

K
r

H
c
)

= CB(H
c
, K
c
) = B(H, K) (as in the proof of (4),
for example). Thus

K
r

H
c
is the unique predual S
1
(K, H) of B(H, K). Lastly,
for (10), CB(S
1
(
2
I
,
2
J
), X) is completely isometric to
CB(R
I

C
J
, X)

= CB(C
J
, CB(R
I
, X))

= R
w
J
(C
w
I
(X))

= M
I,J
(X),
using (9), (2.37), 2.1.25 twice, and 2.1.23 (8).
Finally, for (1), there is a direct proof in [15], for example, but we give an indirect
one. We prove that X
h
H
c
= X

H
c
completely isometrically, the other relation
being similar. Clearly it suces, by Proposition 2.4.14, to show that I : X
h
H
c

H
c
is completely contractive. This will follow if we can show that any jointly
completely contractive map u: X H
c
B(L) is completely contractive (in the
sense of 2.4.6). For if the latter statement was true, take u = : XH
c
X

H
c
,
this is jointly completely contractive, so completely contractive, and thus linearizes
to a completely contractive linear map I : X
h
H
c
X

H
c
.
We may assume that H =
2
J
for some set J. Let v : X R
w
J
(B(L)) =
B(L
(J)
, L) be the linear map dened by v(x) = (u(x, e
i
))
iJ
for any x X. As in
the proof of (10), there is a sequence of isometries
CB(X

C
J
, B(L)) = CB(X, CB(C
J
, B(L))) = CB(X, R
w
J
(B(L)))
provided by (2.37) and Proposition 2.1.25, and it is easy to check that the compo-
sition of these maps takes u (viewed as an element of CB(X

C
J
, B(L))) to v.
Thus |v|
cb
= |u|
jcb
1. Then we dene a map w: C
J
C
J
(B(L)) B(L, L
(J)
)
by w() = (
j
I
L
) for = (
j
)
2
J
. It is clear that |w|
cb
= 1. Also, we have a
factorization
u(x, ) =

j
u(x, e
i
)
i
I
L
= u(x,

j
e
i

i
) = v(x)w(), x X, = (
i
) H =
2
J
.
But any such product of two completely contractive linear maps, is clearly a com-
pletely contractive bilinear map in the sense of 2.4.6) (this is also the easy part of
the later result Theorem 2.5.6). Thus u is completely contractive.
Historical notes: This section is an expansion of [4, Section 1.5]; historical
attributions are given there. The proof given here of the injectivity of the Haagerup
tensor product is from [6]; the original source is [22].
Exercises.
84 CHAPTER 2. OPERATOR SPACES
(1) Show that the following map : R
n

h
X

h
C
n
M
n
(X)

is a surjective
complete isometry:
(

c )([x
ij
]) =

r [(x
ij
)]

c ,

r R
n
,

c C
n
, X

, [x
ij
] M
n
(X).
[Hint: Show that is a complete contraction, and that

is the composition
of the canonical complete isometries (R
n

h
X

h
C
n
)

= CB(X

, M
n
)

=
M
n
(X

)

= M
n
(X)

.]
2.5 Properties of completely bounded maps
Corollary 2.5.1. (The Wittstock Hahn-Banach extension theorem) If X is a
closed subspace of an operator space Y , if H, K are Hilbert spaces, and if u :
X B(K, H) is completely contractive, then there exists a completely contractive
u : Y B(K, H) with u
|X
= u.
Proof. We identify K

=
2
J
and H

=
2
I
, for sets I, J, so that B(K, H)

= M
I,J
.
Now CB(X, M
I,J
)

= (R
I

h
X
h
C
J
)

isometrically, via the map that takes v

CB(X, M
I,J
) to the functional taking r xc to rv(x)c, for r R
I
, c C
J
, x X
(see 2.4.15 (4)). Thus u corresponds to a contractive functional in the latter space.
By the injectivity of
h
, we have R
I

h
X
h
C
J
R
I

h
Y
h
C
J
. By the usual
Hahn-Banach theorem, extends to a contractive functional (R
I

h
Y
h
C
J
)

,
and by the above, this corresponds to a complete contraction u : Y M
I,J
. We
have
r u(x)c = (r x c) = (r x c) = ru(x)c, r R
I
, c C
J
, x X.
Thus, u(x) = u(x) for x X.
2.5.2 (Injective spaces) An operator space Z is said to be injective if for any
completely bounded linear map u: X Z and for any operator space Y containing
X as a closed subspace, there exists a completely bounded extension u: Y Z such
that u
|X
= u and | u|
cb
= |u|
cb
. A similar denition exists for Banach spaces. Thus
an operator space (resp. Banach space) is injective if and only if it is an injective
object in the category of operator (resp. Banach) spaces and completely contractive
(resp. contractive) linear maps.
We remark that one version of the HahnBanach theorem may be formulated
as the statement that C is injective (as a Banach space). It follows from Corollary
2.5.1 that:
Theorem 2.5.3. If H and K are Hilbert spaces then B(K, H) is an injective
operator space.
Corollary 2.5.4. An operator space is injective if and only if it is linearly com-
pletely isometric to the range of a completely contractive idempotent map on B(H),
for some Hilbert space H.
2.5. PROPERTIES OF COMPLETELY BOUNDED MAPS 85
Proof. () Supposing X B(H), extend I
X
to a complete contraction P : B(H)
X. Clearly P P = P.
() Follows from 2.5.3 and an obvious diagram chase (we leave the details as an
exercise).
Theorem 2.5.5. (Representation of completely bounded maps) Suppose that
X is a subspace of a C

-algebra B, that H and K are Hilbert spaces, and that

u: X B(K, H) is a completely bounded linear map. Then there exists a Hilbert
space L, a -representation : B B(L) (which may be taken to be unital if B is
unital), and bounded operators S: L H and T : K L, such that u(x) = S(x)T
for all x X. Moreover this can be done with |S||T| = |u|
cb
.
Conversely, any linear map u of the form u = S()T as above, is completely
bounded with |u|
cb
|S||T|.
Proof. We may suppose that u is completely contractive. In the notation of Lemma
2.3.10, u is a corner of a completely positive map unital dened from a subsystem
of M
2
(A) into B(H K). By the extension theorem of Wittstock 2.5.1, one can
extend to a unital completely positive map M
2
(A) B(HK). This in turn, by
Stinesprings theorem, equals V

()V , for a unital representation of M

2
(A) on
another Hilbert space. It is quite easy algebra to see that any unital representation
of M
2
(A) on a Hilbert space E gives rise to a unitary operator U from that Hilbert
space onto L L, for a subspace L of E, and a unital representation of A on L,
such that the rst representation equals [a
ij
] U

[(a
ij
)]U, for [a
ij
] M
2
(A). In
our case, we obtain
_
0 T(x)
0 0
_
=
__
0 x
0 0
__
= V

_
0 (x)
0 0
_
UV = W

(x)W,
where W = [0 I]UV , with a similar formula dening W

. Pre- and post-multiplying

by the projection from H H onto H, and the inclusion from H into H H, gives
T = R()S, for appropriate contractions S, R.
The last assertion we leave as an easy exercise using Proposition 2.1.6, and
Exercise (2) of Section 2.1.
The following important result states that any completely bounded bilinear map
may be factorized as a product of two completely bounded linear maps. It is due to
Christensen and Sinclair (the C

-algebra case), and Paulsen and Smith (the general

case. Note that their injectivity of the Haagerup tensor product (Theorem 2.4.10)
immediately reduces the general case to the C

-algebra case, as we shall see).

Theorem 2.5.6. Suppose that X and Y are operator spaces, and that u: XY
B(K, H) is a bilinear map.
(1) u is completely contractive (as a bilinear map) if and only if there is a Hilbert
space L, and there are completely contractive linear maps v : X B(L, H) and
w: Y B(K, L), with u(x, y) = v(x)w(y) for all x X and y Y .
86 CHAPTER 2. OPERATOR SPACES
(2) If X and Y are subspaces of unital C

-algebras A and B respectively, and if

the conditions in (1) hold, then there exist Hilbert spaces K
1
and K
2
, unital
-representations : A B(K
1
) and : B B(K
2
), and contractions T
B(K, K
2
), S B(K
2
, K
1
) and R B(K
1
, H), such that
u(x, y) = R(x)S(y)T, x X, y Y.
There are half a dozen or more proofs of this result in the literature, which we
describe some of. First, note that the second part of this result follows immediately
from the rst part and 2.5.5. Second, note that one may assume that A and B are
C

-algebras. To see this, suppose that X and Y are subspaces of C

-algebras A and
B respectively. If u : X
h
Y B(K, H) is the associated linear map, then since
X
h
Y A
h
B, by Wittstocks extension theorem 2.5.1 we can extend u to a
completely contractive linear map on A
h
B. This yields a completely contractive
bilinear map on A B. If the C

-algebra case holds, then this gives the desired

result for A and B, and by restriction, for X and Y .
The next observation (used in some proofs), is that one may replace the target
space B(K, H) by C, by the same trick used in the proof of Theorem 2.5.1. Indeed,
as in that proof we have
CB(X
h
Y, B(K, H))

= (R
I

h
(X
h
Y )
h
C
J
)

= (X

h
Y

,
where X

= R
I

h
X and Y

= Y
h
C
J
. If the result were true in the scalar
valued case (and by the last paragraph we may assume that X

, Y

are C

-algebras),
then we see that u corresponds to a product v(x)w(y), for complete contractions
w : Y
h
C
J
B(C, L) = L
c
, and v : R
I

h
X B(L, C) = ADD. Using ADDREF
(1), and ADDREFS, w induces a complete contraction w

: Y CB(C
J
, L
c
)

=
B(H, L), and v induces a complete contraction v

: X CB(R
I
,

L
r
)

= B(L, H).
It is easy to check that u(x, y) = v

(x)w

(y), which proves the result.

At this point, proofs of Theorem 2.5.6 that use the route of the last paragraph
now have to characterize elements of (A
h
B)

. Two dierent approaches to this

may be found in [15, Section 9.4] and [7]. The former uses a geometric Hahn-
Banach separation argument similar to the proof of Lemma 2.1.10 (appearing rst
in unpublished work of Haagerup [17]). The latter crucially uses a very useful notion
due to Roger Smith called strong independence of vectors in a Hilbert space [28].
Other proofs may be found in [21, 25, 27], and of course the original papers.
2.5.7 (Completely bounded bilinear functionals) If we have a completely contrac-
tive bilinear u: X Y C, by Theorem 2.5.6 we may write u(x, y) = v(x)w(y).
Here H is a Hilbert space, and w: Y B(C, H) = H
c
and v : X B(H, C) =

H
r
are completely contractive linear maps. Regarding these as mapping into H and

H
respectively, we have
u(x, y) = w(y), v(x))
H
, x X, y Y.
Or, if H =
2
I
, then we may regard w, v as complete contractions Y C
I
and
v : X R
I
respectively, and then u(x, y) = v(x)w(y), where the multiplication
occurring here is simply multiplying a row matrix by a column matrix.
2.6. DUALITY AND TENSOR PRODUCTS OF DUAL SPACES 87
This can be rewritten in another important way. Note that v

: B(H, C)

is a complete contraction, and hence so is s = v

, where : H
c
B(H, C)

is
the canonical complete isometry (see (2.14)). Also,
s(w(y))(x) = (w(y))(v(x)) = v(x)w(y) = u(x, y) = u(y)(x),
where u : X Y

is the canonical map associated with u (see ADDREF in the

C

-algebra course). Thus u = s w is a factorization of u through H

c
:
X H
c
Y

.
The steps here are reversible, that is, if u: XY C and if u = sw, where H is a
Hilbert space, and w: Y H
c
and s : H
c
X

are completely contractive linear

maps, then u is a completely contractive bilinear functional. That is, u corresponds
in the usual way to an element of Ball((X
h
Y )

). This is a characterization of
Ball((X
h
Y )

), or, if you like, of (X

h
Y )

. Summarizing, such functionals are

nothing but the maps X Y

which factor completely contractively through

H
c
.
2.5.8 (Further remarks on Theorem 2.5.6) An analoguous result to (1) of Theorem
2.5.6 holds for multilinear completely bounded maps. Thus if X
1
, . . . , X
m
are opera-
tor spaces and if v
j
: X
j
B(H
j
, H
j1
) are completely contractive linear maps then
the N-linear mapping taking (x
1
, . . . , x
m
) X
1
X
m
to v
1
(x
1
) v
m
(x
m
)
B(H
m
, H
0
) is easily seen to be completely contractive in the sense of 2.4.6. Con-
versely, any completely contractive m-linear map u : X
1
X
2
. . .X
m
B(K, H)
has this form. The proof of this latter assertion proceeds by induction on m. As-
sume that it is true for k = 2 and k = m 1. We have an associated completely
bounded linear map dened on X
1

h
X
2

h
. . .
h
X
m

= X
1

h
(X
2

h
. . .
h
X
m
)
(the latter by a fact from the discussion on associativity in 2.4.8). By the k = 2
case, this map may be factorized as: x
1
(x
2
x
m
) v
1
(x
1
)w(x
2
x
m
).
By the k = m 1 case, w(x
2
x
m
) = v
2
(x
2
) v
m
(x
m
). Thus u is of the
required form.
Likewise, (2) of Theorem 2.5.6 has an analogous formulation for multilinear
maps, which follows immediately from (1) and Theorem 2.5.5.
Historical notes: The proof of Wittstocks Theorem 2.5.1 given here is the
modication from [6] of a proof due to Eros. Theorem 2.5.5 was rst proved by
Haagerup in unpublished work [16] from 1980. There are some interesting historical
anecdotes in this handwritten manuscript concerning this result, and related topics.
The rst published proof was the simple one that Paulsen found [20], and this is
the one given here.
2.6 Duality and tensor products of dual spaces
2.6.1 (Mapping spaces as duals) If Y is a dual operator space then we saw in
Corollary 2.4.12 that so is CB(X, Y ), for any operator space X. Indeed by (2.38)
88 CHAPTER 2. OPERATOR SPACES
an explicit predual for CB(X, Y ) is X

. From this, together with the density

of the nite rank tensors in X

, and Lemma 1.4.11, it follows that a bounded

net (u
t
)
t
in CB(X, Y ) converges in the w

-topology to a u CB(X, Y ) if and only

if u
t
(x)(z) u(x)(z) for all x X, z Y

. That is, if and only if u

t
(x) converges
in the w

-topology to u(x) in Y for all x X. Next, suppose that Y = B(K, H) for

Hilbert spaces H, K. Since the latter net is bounded, it follows from the fact that
the weak* topology coincides with the WOT on bounded sets in B(K, H), that the
above equivalent conditions are also equivalent to
u
t
(x), ) u(x), ) for all x X, K, H. (2.39)
It is easy to see that the latter condition is equivalent to the same condition, but
with , arbitrary elements of a orthonormal basis for H and for K.
2.6.2 (Dual matrix spaces) If X is a dual operator space then so is M
n
(X). Indeed
by (2.38) and (2.6) we have
(S
1
n

= CB(X

, M
n
)

= M
n
(X).
More generally the same proof, but substituting 2.1.23 (12) for (2.6), shows that for
sets I, J, M
I,J
(X) is a dual operator space with operator space predual S
1
(
2
I
,
2
J
)

, and also M
I,J
(X)

= CB(X

, M
I,J
). Alternatively, note that by (2.38) and
2.4.15 (10), we have
(S
1
(
2
I
,
2
J
)

X

= CB(S
1
(
2
I
,
2
J
), X)

= M
I,J
(X).
If I, J are sets, and if I
0
and J
0
are nite subsets of I and J respectively, write
= I
0
J
0
. The set of such is a directed set under the usual ordering. For
such , and for x M
I,J
(X), we write x

for the matrix x, but with entries x

ij
switched to zero if (i, j) / . Then (x

is a net indexed by , which we

call the net of nite submatrices of x.
Corollary 2.6.3. (Eros and Ruan) Let X be a dual operator space, and let I, J
be sets.
(1) If (x
t
)
t
is a bounded net in M
I,J
(X), then x
t
x M
I,J
(X) in the w

-topology
in M
I,J
(X), if and only if each entry in x
t
converges in the w

-topology in X
to the corresponding entry in x.
(2) If Y is a dual operator space, and if u: X Y is a w

-continuous com-
pletely bounded map, then the amplication u
I,J
: M
I,J
(X) M
I,J
(Y ) is
w

-continuous.
(3) M
n
I,J
(X) is w

-dense in M
I,J
(X). Indeed if I, J are sets, and x M
I,J
(X),
then the net of nite submatrices of x converges to x in the w

-topology.
Proof. As we said in (2.6.2), M
I,J
(X) = CB(X

, M
I,J
) = CB(X

, B(
2
J
,
2
I
)). By
(2.39) and the remark after it, it follows that a bounded net x
s
x M
I,J
(X)

=
CB(X

, B(
2
J
,
2
I
)) if and only if
[x
s
i,j
()]e
j
, e
i
) = x
s
i,j
() [x
i,j
()]e
j
, e
i
) = x
i,j
(), X

,
2.6. DUALITY AND TENSOR PRODUCTS OF DUAL SPACES 89
that is, if and only if x
s
i,j
x
i,j
weak*, for all i I, j J. This is (1).
Items (2) and (3) follow immediately from (1). For example, if u: X Y is
w

-continuous, and if we have a bounded net x

s
x M
I,J
(X), then by (1) each
entry of x
s
converges weak* to the corresponding entry of x. Also, (u
I,J
(x
s
)) is a
bounded net in M
I,J
(X), and it converges to u
I,J
(x) by (1) again, since each entry
of u
I,J
(x
s
) converges weak* to the corresponding entry of u
I,J
(x). Thus u
I,J
is
w

-continuous by 1.4.7.
Note that in 2.6.3 (2), if u is also a complete isometry then by 1.4.7 we see that
u
I,J
is a w

-homeomorphism onto its range, which is w

-closed. As a corollary we
see that for a w

-closed subspace X B(K, H), one may dene M

I,J
(X) to be
the w

-closure of M
n
I,J
(X) in M
I,J
(B(K, H)) = B(K
(J)
, H
(I)
). Indeed, taking u
to be the embedding X B(K, H), we see that M
I,J
(X) is w

-homeomorphically
completely isometric to a w

-closed subspace of M
I,J
(B(K, H)). Applying (3) of
the last result we deduce the statement about M
n
I,J
(X).
We turn next to a characterization of dual operator spaces:
Theorem 2.6.4. Let X be an operator space with a given weak* topology (coming
from a predual Banach space). The following are equivalent:
(i) X with its given weak* topology is a dual operator space.
(ii) M
n
(X) is a dual Banach space, and the n
2
canonical inclusion maps from X
into M
n
(X) are w

-continuous, for all n 2.

(iii) Whenever (x
s
)
s
is a net in Ball(M
n
(X)), x M
n
(X), and the i-j entry of x
s
converges in the weak* topology to the i-j entry of x for all i, j = 1, . . . , n,
then x Ball(M
n
(X)).
Proof. Write
ij
: X M
n
(X) for the i-j inclusion map.
(i) (ii) If x
t
x weak* in X then by Theorem 2.6.3 (i) we have
ij
(x
t
)

ij
(x) weak*. So
ij
is w

-continuous.
(ii) (iii) If x
s
, x are as in (iii), with x
s
ij
x
ij
weak* for all i, j, then by (ii)
we have
ij
(x
s
ij
)
ij
(x
ij
) weak*, so that
x
s
=
n

i,j=1

ij
(x
s
ij
)
n

i,j=1

ij
(x
ij
) = x
weak* in M
n
(X). If x
s
Ball(M
n
(X)), then since the latter ball is weak* closed,
it follows that x Ball(M
n
(X)).
(iii) (i) We may suppose that the predual Banach space W X

. We will
always regard W as an operator space by giving it the inherited matrix norms from
X

. We will use Exercise (1) of Section 2.4, namely that the following canonical
map : R
n

h
X

h
C
n
M
n
(X)

is a surjective complete isometry:

(

c )([x
ij
]) =

r [(x
ij
)]

c ,

r R
n
,

c C
n
, X

, [x
ij
] M
n
(X).
Alternatively, this fact can be proved from the later result (2.41), since using that
result and 2.4.15 (6), we have R
n

h
X

h
C
n

= (C
n

h
X
h
R
n
)

= M
n
(X)

.
90 CHAPTER 2. OPERATOR SPACES
Note that
(

e
k

e
l
)([x
ij
]) = (x
kl
), X

, [x
ij
] M
n
(X), k, l 1, . . . , n.
Since
h
is injective we deduce that R
n

h
W
h
C
n
R
n

h
X

h
C
n

= M
n
(X)

isometrically. Dene : M
n
(X) (R
n

h
W
h
C
n
)

by (x)(u) = (u)(x), for

x M
n
(X), u R
n

h
W
h
C
n
. Note that (x)(

e
k
w

e
l
) = w(x
kl
), for w W,
by the last centered equation. From this it is clear that is one-to-one, and it is
easy to see that it is onto. Note that (Ball(M
n
(X))) is w

-closed by hypothesis,
for if (x
s
) (x) weak* in (R
n

h
W
h
C
n
)

, with |x
s
| 1, then by the last
line, w(x
s
kl
) w(x
kl
) for all w W, and k, l. That is, x
s
kl
x
kl
weak*, so that
|x| 1 by (iii).
If u R
n

h
W
h
C
n
then |u| = |(u)|, which equals
sup|(u)(x)| : x Ball(M
n
(X)) = sup|(x)(u)| : x Ball(M
n
(X)).
Thus the pre-polar (Ball(M
n
(X)))

equals the unit ball of R

n

h
W
h
C
n
.
Therefore by the bipolar theorem, (Ball(M
n
(X))) = Ball((R
n

h
W
h
C
n
)

).
That is, is an isometry. The composition of with the canonical isometries
(R
n

h
W
h
C
n
)

= CB(W, M
n
) = M
n
(W

) from 2.4.15 (4) and (2.6), is the nth

amplication of the isometry X W

. Since this holds for any n 1, the latter

map is a complete isometry. Thus W

= X completely isometrically.
2.6.5 (Normal spatial tensor product) If X and Y are dual operator spaces, with
operator space preduals X

and Y

, then CB(Y

, X) is the dual operator space of

X

by 2.6.1. As in (2.27), we regard X

min
Y CB(Y

, X), and we dene

the normal minimal tensor product XY to be the w

-closure of X Y (or of
X
min
Y ) in CB(Y

, X). Equivalently, if X and Y are w

-closed subspaces of
B(H) and B(K) respectively, then we may dene XY to be the w

-closure in
B(H
2
K) of the copy of XY . This is sometimes referred to as the normal spatial
tensor product. If M and N are W

-algebras, then M N as described above is the

usual von Neumann algebra tensor product as we dened it in the C

-algebra course
ADDREF. In particular, B(H) B(K) = B(H
2
K) as we saw in ADDREF in the
C

-algebra course. To see that these two denitions of XY are the same (up to
w

-homeomorphic complete isometry), we use the following argument. Since X and

Y are w

-closed subspaces of B(H) and B(K) respectively, we know by 2.2.7 that

X

and Y

are quotients of S
1
(H) and S
1
(K) respectively. By the projectivity
property of

, we obtain a complete quotient map Q: S
1
(H)

S
1
(K) X

.
Using the identication (2.38) we see that Q

may be viewed as a w

-continuous
completely isometric embedding
CB(Y

, X) CB(S
1
(K), B(H))

= B(H
2
K),
the last relation from the rst paragraph of 2.4.2. Via the canonical identication
of X Y with a subset of CB(Y

, X), it is easy to argue that the w

-closure of
XY in B(H
2
K) may be identied with the w

-closure of XY in CB(Y

, X).
2.6. DUALITY AND TENSOR PRODUCTS OF DUAL SPACES 91
In general, CB(Y

, X) (or equivalently, (X

) is not equal to XY ;
nonetheless they do coincide in many cases of interest. In fact, (X

=
CB(Y

, X) equals the so-called normal Fubini product of X and Y (a fact proved

independently by Blecher and Ruan, see [3, Theorem 2.5], [26], inspired by the von
Neumann algebra case [12]). The normal Fubini product is dened in terms of slice
maps, and thus the question of whether XY = CB(Y

, X) holds, is related to
slice map properties.
We discuss three examples of when XY = CB(Y

, X) holds. First, it holds

when X and Y are von Neumann algebras. This was proved by Eros and Ruan
in [12]: by the last paragraph this case is equivalent to showing that the normal
Fubini product of von Neumann algebras equals their von Neumann algebra tensor
product, which can be deduced from the slice map theorem of Tomiyama [29].
Second, XY = CB(Y

, X) holds when X = B(H, K). Indeed, for any dual

operator space Y and sets I, J,
M
I,J
Y

= M
I,J
(Y ) (2.40)
as dual operator spaces. This follows by the remark after 2.6.3, and an argument
similar to the one used for (2.29). Also, M
I,J
(Y )

= CB(Y

, M
I,J
) by 2.1.23 (12)
(setting one of the spaces there equal to C). Thus M
I,J
Y

= CB(Y

, M
I,J
). Note
that taking J singleton, and Y =

K
r
, gives H
c

K
r
= B(K, H).
We leave it as an exercise that the normal spatial tensor product is associative,
and functorial for w

-continuous completely bounded maps.

2.6.6 (W

-continuous extensions of bilinear maps) Let X, Y be operator spaces,

let W be a dual operator space, and let u: XY W be a completely contractive
bilinear map. We claim that there is a unique separately w

-continuous extension
u: X

W of u, and this extension is completely contractive too. To prove

this, we may assume by Lemma 2.2.8 that W is a w

-closed subspace of some B(H).

By the Theorem 2.5.6, there exists a Hilbert space L and two completely contractive
maps v : X B(L, H) and w: Y B(H, L) such that u(x, y) = v(x)w(y) for all
x X, y Y . By 2.2.9, v and w have w

-continuous completely contractive

extensions v : X

B(L, H) and w: Y

B(H, L). Dene u: X

B(H) by setting u(, ) = v() w(), for X

, Y

. Then the easy part of

Theorem 2.5.6 ensures that u is completely contractive, and it clearly is separately
w

-continuous, and extends u. Note that for any separately w

-continuous extension
u: X

B(H) of u, we must have

u(, ) = lim
s
lim
t
u(x
t
, y
s
), if x
t
, y
s
,
where all the limits here are in the weak* topology. From this we see that u(, )
W, and also the uniqueness of the extension.
2.6.7 (Self-duality of
h
) Let X and Y be operator spaces. Then
X

h
Y

(X
h
Y )

completely isometrically (2.41)

92 CHAPTER 2. OPERATOR SPACES
via the canonical map J (that is, J( )(x y) = (x)(y)). To prove this, we
rst assume that X and Y are nite-dimensional. In this case, J is a surjection, by
linear algebra. An element U in the ball of M
n
((X
h
Y )

) corresponds by (2.6)
to a complete contraction u : X
h
Y M
n
. By 2.5.6 (1), there exist a Hilbert
space L and two complete contractions v : X B(L, C
n
) and w: Y B(C
n
, L)
such that u(x y) = v(x)w(y) for any x X and y Y . We may assume that L
is nite dimensional, by replacing L by its nite dimensional subspace [w(Y )C
n
],
and restricting v(x) to that subspace. Thus we may assume that v : X M
n,p
and w: Y M
p,n
, for an integer p 1. We let = [
ij
] M
n,p
(X

) and
= [
ij
] M
p,n
(Y

) be the two matrices corresponding to v and w respectively

(by a variant of (2.6), these have norm 1. Then
u(x y) = v(x)w(y) = [
ij
(x)][
ij
(y)] =
_

ik
(x)
kj
(y)
_
, x X, y Y.
If z = then it is easy to see that J
n
(z) = U, and |z|
h
|||| =
|v|
cb
|w|
cb
1. The converse inequality |U|
cb
|z|
h
may be obtained by revers-
ing the arguments.
In the general case, x [u
ij
] M
n
(X

). Write each u
ij
X

in the
form

m
k=1

k
, for functionals
k
X

,
k
Y

. Let W (resp. Z) be the span

of all these (nite number of) functionals in X

(resp. Y

), over all i and j too. Then

W

= (X/W

and Z

= (Y/Z

isometrically. The canonical maps X X/W

and Y Y/Z

induce a complete quotient map X

h
Y (X/W

)
h
(Y/Z

),
by the projectivity of
h
(see the third bullet in 2.4.8). By 2.2.3, the last map
dualizes to give a weak* continuous complete isometry ((X/W

)
h
(Y/Z

))

(X
h
Y )

. On the other hand, by the last paragraph, ((X/W

)
h
(Y/Z

))

=
W
h
Z completely isometrically. Thus we have the following diagram of completely
isometries (the vertical arrow coming from the injectivity of
h
, see 2.4.8):
M
n
(X

M
n
(W
h
Z) M
n
(((X/W

)
h
(Y/Z

))

) M
n
((X
h
Y )

).
We may view [u
ij
] in M
n
(W
h
Z). The composition of the maps in the last sequence
is easily seen to take [u
ij
] to [J(u
ij
)] M
n
((X
h
Y )

), and so we are done.

2.6.8 (The dual of the Haagerup tensor product) If X and Y are operator spaces
then w (X
h
Y )

if and only if there exist [

i
] R
w
I
(X

) and [
i
] C
w
I
(Y

)
such that w may be written as
w(x y) =

iI

i
(x)
i
(y), x X, y Y. (2.42)
The last sum converges absolutely in C, as may be seen by the CauchySchwarz
inequality:

iI
[
i
(x)
i
(y)[
_

iI
[
i
(x)[
2
_1
2
_

iI
[
i
(y)[
2
_1
2
= |(x)||(y)|,
2.6. DUALITY AND TENSOR PRODUCTS OF DUAL SPACES 93
where : X R
I
and : Y C
I
are the canonical maps which are associated
with [
i
] R
w
I
(X

) and [
i
] C
w
I
(Y

). Indeed, by 2.1.23 (12), for example, and

are completely contractive i [
i
] and [
i
] have norm 1. Note that (2.42) may
be rewritten as w(x y) = (x)(y), and now this relation may be seen to be a
restatement of the discussion in 2.5.7. Indeed, this argument shows that |w| 1
i , above may be chosen to be complete contractions, and i [
i
] and [
i
] in
(2.42) have norm 1.
Thus every w (X
h
Y )

is a sum of rank one tensors

w =

iI

i

i

i
X

,
i
Y

.
Viewing (X
h
Y )

as a tensor product of X

and Y

in this way, leads to the

weak* Haagerup tensor product, often called the extended Haagerup tensor product,
which we shall not discuss further here. It has properties analogous to the Haagerup
tensor product.
Historical note: Nearly all the facts about innite matrices in 2.6.2 and 2.6.3
are explicitly in [18, 11, 12]. The result 2.6.4 is due to Le Merdy [19]. See e.g.
[15] for more on the normal spatial tensor product and the Fubini tensor product.
The selfduality relation (2.41) was conjectured by Blecher and proved in full in [14].
Dierent proofs appear in [3, 7]. The dual of the Haagerup tensor product was rst
explored by Eros and Kishimoto [9] following unpublished work of Haagerup [17],
viewing this dual as a tensor product originates in [7].
For more detailed references to topics in this section see [4], and other basic
operator space texts.
94 CHAPTER 2. OPERATOR SPACES
Bibliography
[1] W. B. Arveson, Subalgebras of C

-algebras, Acta Math. 123 (1969), 141224.

[2] D. P. Blecher, The standard dual of an operator space, Pacic J. Math. 153 (1992), 1530.
[3] D. P. Blecher, Tensor products of operator spaces II, Canad. J. Math. 44 (1992), 7590.
[4] D. P. Blecher and C. Le Merdy, Operator algebras and their modulesan operator space
approach, Oxford Univ. Press, Oxford (2004).
[5] D. P. Blecher and B. Magajna, Duality and operator algebras: automatic weak

continuity
and applications, J. Funct. Anal. 224 (2005), 386407.
[6] D. P. Blecher and V. I. Paulsen, Tensor products of operator spaces, J. Funct. Anal. 99
(1991), 262292.
[7] D. P. Blecher and R. R. Smith, The dual of the Haagerup tensor product, J. London Math.
Soc. 45 (1992), 126144.
[8] J. B. Conway, A Course in Operator Theory, Graduate Studies in Mathematics, 21, Amer.
Math. Soc. Providence, RI, 2000.
[9] E. G. Eros and A. Kishimoto, Module maps and Hochschild-Johnson cohomology, Indiana
Univ. Math. J. 36 (1987), 257276.
[10] E. G. Eros and Z-J. Ruan, On matricially normed spaces, Pacic J. Math. 132 (1988),
24364.
[11] E. G. Eros and Z-J. Ruan, Representations of operator bimodules and their applications,
J. Operator Theory 19 (1988), 137157.
[12] E. G. Eros and Z-J. Ruan, On approximation properties for operator spaces, Internat. J.
Math. 1 (1990), 163187.
[13] E. G. Eros and Z-J. Ruan, A new approach to operator spaces, Canad. Math. Bull. 34
(1991), 329337.
[14] E. G. Eros and Z-J. Ruan, Self-duality for the Haagerup tensor product and Hilbert space
factorization, J. Funct. Anal. 100 (1991), 257284.
[15] E. G. Eros and Z-J. Ruan, Operator Spaces, London Mathematical Society Monographs,
New Series, 23, The Clarendon Press, Oxford University Press, New York, 2000.
[16] U. Haagerup, Decompositions of completely bounded maps on operator algebras, Unpub-
lished manuscript (1980).
[17] U. Haagerup, The tensor product for C

-algebras, Unpublished manuscript (1980).

95
96 BIBLIOGRAPHY
[18] M. Hamana, Tensor products for monotone complete C

-algebras, I, Japan J. Math. 8

(1982), 259283.
[19] C. Le Merdy, On the duality of operator spaces, Canad. Math. Bull. 38 (1995), 334346.
[20] V. I. Paulsen, Every completely polynomially bounded operator is similar to a contraction,
J. Funct. Anal. 55 (1984), 117.
[21] V. I. Paulsen, Completely bounded maps and operator algebras, Cambridge Studies in Ad-
vanced Math., 78, Cambridge University Press, Cambridge, 2002.
[22] V. I. Paulsen and R. R. Smith, Multilinear maps and tensor norms on operator systems, J.
Funct. Anal. 73 (1987), 258276.
[23] G. K. Pedersen, C

-algebras and their automorphism groups, Academic Press, London,

1979.
[24] G. K. Pedersen, Analysis now, Graduate Texts in Mathematics, 118, Springer-Verlag, New
York, 1989.
[25] G. Pisier, Introduction to operator space theory, London Math. Soc. Lecture Note Series,
294, Cambridge University Press, Cambridge, 2003.
[26] Z-J. Ruan, On the predual of dual algebras, J. Operator Theory, 27 (1993), 179192.
[27] A. M. Sinclair and R. R. Smith, Hochschild cohomology of von Neumann algebras, London
Math. Soc. Lecture Note Series, 203, Cambridge University Press, Cambridge, 1995.
[28] R. R. Smith, Completely bounded module maps and the Haagerup tensor product, J. Funct.
Anal. 102 (1991), 156175.
[29] J. Tomiyama, Tensor products and projections of norm one in von Neumann algebras,
Lecture notes, University of Copenhagen, 1970.
[30] G. Wittstock, Extensions of completely bounded C

-module homomorphisms, Operator al-

gebras and group representations, Vol. II (Neptun, 1980), pp. 238250, Monogr. Stud.
Math., 18, Pitman, Boston, MA, 1984.