Lecture Notes on

-Algebras and K-Theory

N.P. Landsman
Kortewegde Vries Institute for Mathematics
University of Amsterdam
Plantage Muidergracht 24
NL-1018 TV AMSTERDAM
THE NETHERLANDS
email npl@science.uva.nl
Abstract: The aim of these lectures is to explain the basics of the theory of C

-algebras
and their associated K-groups in the light of noncommutative geometry. Part I is an introduc-
tion to C*-algebras, covering the philosophy of noncommutative geometry, Banach algebras and
C*-algebras, commutative C*-algebras, the original Gelfand-Naimark duality theorem, categories
and functors, natural transformations and equivalence of categories, the categorical version of
the Gelfand-Naimark duality theorem, the GNS-construction, the Gelfand-Naimark representa-
tion theorem, ideals and compact operators, the structure of nite-dimensional C

-algebras, and
applications to groupoids. Part II is an introduction to the K-theory of C*-algebras, covering
projections in C*-algebras, the denition of K
0
, vector bundles and topological K-theory, K
0
for nonunital C*-algebras, some explicit computations, properties of C*-algebraic K-theory (half-
exactness, excision, etc.), higher K-groups, the index map, and Bott periodicity.
0
Draft: December 14, 2003
2
1 Introduction
Noncommutative geometry was created by Alain Connes around 1980 in an eort to relate the
theory of algebras of operators on Hilbert spaces to dierential geometry and algebraic topology.
For example, one of his goals was to generalize the famous index theorems of Atiyah and Singer
to the setting of foliated manifolds. More generally, noncommutative geometry turned out to
be a powerful tool for the study of singular spaces. This program has eventually led to a huge
edice, described in [3]. Conness book is very advanced, partly because the underlying theory is
dicult, and also because his examples typically come from the frontiers of mathematics. Thus
many parts of [3] are dicult even for experts in noncommutative geometry. As an example of a
good introductory text, we mention [8].
Noncommutative geometry relies on a certain technical apparatus, which includes homological
algebra and the theory of C

-algebras, which are certain algebras of operators on a Hilbert space.

These notes are concerned with the latter theory, which is an important eld of mathematics and
mathematical physics worth studying also independently of noncommutative geometry. To see at
what point C

-algebras enter the game, we now introduce the philosophy of noncommutative

geometry in a simplied way.
A modern view of mathematics, going back to Hilbert and Bourbaki, is that the subject is
concerned with structures dened on sets. The three basic examples of structures are topology,
algebra, and (partial) order. As you know, a topology on a set X is a choice of a collection of
so-called open subsets of X, an algebraic structure on a set A in the simplest case is a map from
AA to A, whereas a partial ordering on a set describes a certain relation between its elements.
Higher structures are dened in terms of these basic ones; for example, dierential geometry starts
from the assumption that certain open sets dening the topology of X are homeomorphic to R
n
for some n. Alternatively, the theory of groups arises by putting certain axioms on an algebraic
structure A A A, whereas an algebra over a commutative ring k in the usual sense involves
two such maps on A as well as two maps kk k on a second set k, along with a map kA A.
A key guiding thought is that whenever a set carries more than one structure, these structures
should be related by certain compatibility conditions. Almost every decent denition in mathematics
illustrates this point, the denition of a C

-algebra given below being a particularly rich example.

Algebraic topology tries to describe the topology of X in terms of certain abelian groups A(X),
which are dened by X in a functorial way.
1
That is, in the so-called covariant case a map
2
: X Y denes a homomorphism

: A(X) A(Y ), whereas in the contravariant case one

has a homomorphism

: A(Y ) A(X). These are supposed to satisfy the obvious composition

rules: given another map : Y Z with associated homomorphism

: A(Y ) A(Z) (in the

covariant case), one should have ()

: A(X) A(Z). Similarly, in the contravariant

case one requires ( )

: A(Z) A(X), where

: A(Z) A(Y ). Finally, in both

cases the identity map id
X
: X X induces the identity map from A(X) to A(X).
Consequently, if X

= Y (i.e, X is homeomorphic to Y ), then A(X)

= A(Y ) (i.e., A(X) and
A(Y ) are isomorphic as groups). This conclusion is most powerfully used in a negative way: if
A(X) and A(Y ) are not isomorphic, then X and Y cannot be homeomorphic. Thus the groups
A(X) capture some aspects of the topology of X. In any case, as the name suggests, in algebraic
topology one trades a topological structure for an algebraic one, viz. that of an abelian group.
The rst step in the program of noncommutative geometry is somewhat similar: again the
topology of X is used to dene an abelian group A(X), but this time A(X) has the richer structure
of a commutative algebra over k = C. Namely, one simply takes A(X) = C(X) := C(X, C),
the collection of all continuous complex-valued functions on X. The association X C(X) is
contravariant: a map : X Y denes a homomorphism of commutative algebras

: C(Y )
C(X) as the pullback, that is, for f C(Y ) one puts

(f) = f C(X). The question then

arises to what extent the commutative algebra C(X) captures X and its topology. For example,
the worst possible scenario arises when X has the coarse topology,
3
so that C(X)

= C, which
1
The A(X) are typically (co)homology groups or homotopy groups (though the latter are not necessarily abelian).
2
All maps between topological spaces are assumed to be continuous, unless the contrary is explicitly stated.
3
I.e., the only open sets are X and .
3
means that C(X) does not contain any information whatsoever about the space X.
On the other hand, in the most favourable situation, when X is a compact Hausdor space,
it turns out that X can be reconstructed as a topological space from C(X). This reconstruction
is done as follows. One denes a character of a commutative algebra A (over C) as a nonzero
homomorphism : A C of algebras (that is, (ab) = (a)(b)), and turns the set (A) of
all characters of A into a topological space by saying that
n
when
n
(a) (a) in C
for all a A. For A = C(X), one has an obvious map X (C(X)), written as x
x
,
dened by
x
(f) = f(x). When X is a compact Hausdor space, it turns out that this map is a
homeomorphism, so that X

= (C(X)).
This results shows that every possible further structure on X that is dened in terms of its
topology, could equally well be dened in terms of the commutative algebra C(X). The best
example is given by the notion of a vector bundle over X; roughly speaking, this is a collection
E =
xX
E
x
of vector spaces parametrized by X, each of which isomorphic to C
n
for some xed
n. More precisely, a vector bundle over X is an open surjection : E X with the property that
each ber E
x
:=
1
(x) is a vector space isomorphic to C
n
, and each x X has a neighbourhood
U
x
such that
1
(U
x
)

= U
x
C
n
, the isomorphism
1
(U
x
) U
x
C
n
being linear on each ber.
The simplest example is E = X C
n
with (x, v) = x, but there might be other possibilities.
4
Now, similar to the passage from X to C(X), one may describe E algebraically by its space of
sections
(E) = s : X E [ s = id.
The key point is that (E) is a module over C(X), the action C(X) (E) (E) being given
by f s : x f(x)s(x). Furthermore, E may be reconstructed as a vector bundle over X from
(E) as a C(X) module: E is isomorphic to
xX

E
x
, where

E
x
:= (E)/(C(X, x) (E)). Here
C(X, x) := f C(X) [ f(x) = 0.
Similar to the notion of a direct sum V W of two vector spaces, one may form the direct sum
E F of two vector bundles over a given space X. This leads to the denition of K
0
(X), which
is the abelian group with one generator for each isomorphism class [E] of vector bundles over X,
and relations [E] + [F] = [E F]. The association X K
0
(X) is contravariantly functorial (like
X C(X)), by a construction known as a pullback, too: given a vector bundle : E Y and a
map : X Y one denes the pullback bundle

E over X by

E = E
Y
X := (v, x) E X [ (v) = (x),
with projection (v, x) x. Thus a generator [E] of K
0
(Y ) is mapped into a generator [

E] of
K
0
(X), and since the pullback construction preserves direct sums, this induces a homomorphism

: K
0
(Y ) K
0
(X). For compact Hausdor spaces, the so-called topological K-theory K
0
(X)
(due to Atiyah and Hirzebruch) is an important example of the general strategy of algebraic
topology.
Here again, it should be possible to dene K
0
(X) in terms of C(X). This may indeed be done.
First, dene M

(C(X)) as the collection of all innite matrices with a nite number of entries in
C(X). This is an algebra through the usual matrix multiplication rule, where one now multiplies
functions within C(X) (i.e., pointwise). Dene an idempotent in M

(C(X)) as an element e
satisfying e
2
= e. Two idempotents can be added, like any two elements of M

(C(X)), but in
addition there is an operation of direct sum, viz.
e f =
_
e 0
0 f
_
.
We now dene an equivalence relation among all idempotents in M

(C(X)). Each element of

M

(C(X)) may also be seen as a continuous function f : X M

n
(C) for some n, where
M
n
(C) is the space of complex n n matrices. This induces a norm
5
on M

(C(X)) by |f| :=
sup
xX
|f(x)|
n
, where | |
n
is the usual norm on M
n
(C):
|a|
n
= sup|az|, z C
n
, |z| = 1, (1)
4
Unless X is contractible.
5
Note that M

(C(X)) is not complete in this norm.

4
where |z|
2
=

n
k=1
z
k
z
k
is the usual norm on C
n
. This leads to a notion of homotopy: two
idempotents e, f in M

(C(X)) are said to be homotopic or homotopy equivalent when there is

a norm-continuous path f : [0, 1] M

(C(X)) of idempotents (i.e., f(t)

2
= f(t) for all t) with
f(0) = e and f(1) = f. Finally, we dene K
0
(C(X)) as the abelian group with one generator
for each homotopy class [e] of idempotents in M

(C(X)), and relations [e] + [f] = [e f]. As

promised, it then turns out that K
0
(C(X))

= K
0
(X) in a natural way.
These results do not yet give a completely algebraic reformulation of the topological concept
of a compact Hausdor space. The second step in the program of noncommutative geometry is
to give an abstract characterization of those commutative algebras that are isomorphic to C(X),
for some compact Hausdor space X. This problem was solved, ahead of its time, by Gelfand and
Naimark in 1943 (see [6]). They noted that the space C(X) has the following additional structure
beyond just being a commutative algebra over C (see Exercises). Firstly, it has a norm, given by
|f|

:= sup[f(x)[ [ x X,
in which it is a Banach space. This Banach space structure of C(X) is compatible with its structure
as commutative algebra by the property
|fg|

|f|

|g|

. (2)
Secondly, C(X) has an involution f f

, given by f

(x) = f(x). This involution is related to

the norm as well as to the algebraic structure by the property
|f

f| = |f|
2
. (3)
We summarize these properties by saying that C(X) is a commutative C

-algebra with unit;

abstractly, a commutative C

-algebra is dened as a Banach space that at the same time is a

commutative algebra with involution, such that (2) and (3) hold.
The rst theorem of Gelfand and Naimark then reads as follows: Every commutative C

-algebra
A with unit is isomorphic to C(X) for some compact Hausdor space X. The isomorphism is
constructed as explained above: one takes X := (A), which turns out to be a compact Hausdor
space, and the map A C(X) is the so-called Gelfand transform a a, where a() := (a). It
is typical of commutative C

-algebras, as opposed to general commutative algebras over C, that

the Gelfand transform is an isomorphism.
This theorem can be expanded into a categorical statement: the category of compact Haus-
dor spaces as objects and continuous maps as arrows is dual (or anti-equivalent) to the category
with unital commutative C

-algebras as objects and unital homomorphisms as arrows. This means,

roughly speaking, the following. As we have seen, the association C : X C(X) can be contravari-
antly extended to

: C(Y ) C(X) for any : X Y . Similarly, a unital homomorphism

: A B denes a map

: (B) (A) by basically the same contravariant pullback con-

struction: a character : B C on B is mapped to the character

:= : A C on A.
Moreover, C and are inverses to each other up to isomorphism, that is, (C(X))

= X (as we
have seen), and C((A))

= A for any compact Hausdor space X and any unital commutative
C

-algebra A.
We now come to the third and decisive step of noncommutative geometry: whenever a def-
inition or construction works for commutative C

-algebras, try to extend it to noncommutative

C

-algebras. The rst thing to do here is to make sense of the notion of a noncommutative
C

-algebra itself: this simply consists of omitting the word commutative in the denition of a
commutative C

-algebra! Thus a C

-algebra is dened as a Banach space that at the same time

is an algebra with involution, such that (2) and (3) hold (see Appendix).
A key example of a noncommutative C

-algebra is the algebra A = M

n
(C) of complex n n
matrices, with norm (1) and involution a

ij
= a
ji
(see Exercises). This is a special case of A = B(H),
the algebra of all bounded operators on a Hilbert space H (cf. the Appendix below), which is a
C

-algebra in the usual operator norm (83) and the usual adjoint or Hermitian conjugate as the
involution, i.e., (z, a

w) = (az, w).
6
Furthermore, any subalgebra of B(H) that is closed in the
norm-topology and closed under Hermitian conjugation is obviously a C

-algebra as well.
6
Here ( , ) is the inner product on H (taken linear in the second entry).
5
Similar to their characterization of commutative C

-algebras, Gelfand and Naimark completely

claried the nature of general C

-algebras. Their second theorem, contained in the same paper

as their rst, reads: Every C

-algebra A is isomorphic to a norm-closed and

-closed subalgebra
of B(H) for some Hilbert space H. The proof of this theorem is based on the so-called GNS-
construction (after Gelfand, Naimark, and Segal), which is of central importance to the theory of
C

-algebras, and which basically explains why C

-algebras are naturally related to Hilbert spaces.

This construction starts with the concept of state on a C

-algebra A, which for simplicity

we assume to contain a unit. A state is a linear functional that is positive, in the sense that
(a

a) 0 for all a A, and normalized, in that (1) = 1.

7
The characters of a commutative
C

-algebra are examples of states. Let us rst suppose that (a

a) > 0 for all a, and that A has

a unit. In that case, A is a pre-Hilbert space in the inner product (a, b)

:= (a

b), which may

be completed into a Hilbert space H

. Then A acts on H

by means of

: A B(H

), given
by

(a)b := ab.
8
It is easy to see that

is injective: if

(a) = 0 then, taking b = 1, one infers

that a = 0 as an element of H

, but then (a, a)

= (a

a) = 0, contradicting the assumption that

(a

a) > 0. Moreover, one checks that

is a homomorphism of C

-algebras, for example,

(a)

(b)c =

(a)bc = abc =

(ab)c
for all c, which implies

(a)

(b) =

(ab). Thus A is isomorphic to

(A) B(H

).
9
In general
A may not possess such strictly positive functionals, but it always has suciently many states.
For an arbitrary state the Hilbert space H

is constructed by rst dividing A by the kernel of

(, )

, and proceeding in the same way. The representation

may then fail to be injective, but

by taking the direct sum of enough such representations one always arrives at an injective one.
In conclusion, the theory of (locally)
10
compact Hausdor spaces forms the commutative corner
of the world of C

-algebras, whereas in general one deals with operators on a Hilbert space. This
unies part of point-set topology with functional analysis in a nontrivial way. Conness main idea
was to rst reformulate the constructions of algebraic topology and dierential geometry in terms
of commutative C

-algebras, and then to try and generalize these constructions to arbitrary C

-
algebras. Since the latter are typically noncommutative, this explains the word noncommutative
in noncommutative geometry.
We have, in fact, already seen one such generalization, since the denition of K
0
(C(X)) may
be generalized to any (unital) C

-algebra practically without any modication.

11
This leads to
the subject of K-theory for C

-algebras, which is one of the main techniques in noncommutative

geometry.
From a historical perspective, it is interesting to remark that the ideology of noncommutative
geometry is closely related to that of quantum mechanics, and this analogy actually played an
important role in shaping the eld.
12
From 1900 onwards, physicists began to recognize that the
classical physics of Newton, Maxwell, and Lorentz could not describe all of Nature. The fascinating
era thus initiated by Planck, to be continued mainly by Einstein and Bohr, ended in 1927 with
the nal formulation of quantum mechanics by von Neumann.
13
This theory replaced classical
mechanics, and was initially discovered in two halves.
One half of quantum theory was discovered in 1925 by Heisenberg by a principle of reinterpre-
tation (Umdeutung) of the observables of classical mechanics. The latter were typically functions
7
Positive functionals on a C

-algebra are automatically continuous, with norm = (1).

8
This is initially dened on the dense subspace A H

, and subsequently extended by continuity.

9
A more technical argument shows that

is isometric, so that its image is closed.

10
If on extends the rst theorem of Gelfand and Naimark to commutative C

-algebras without a unit, one ends

up with locally compact spaces.
11
Only the denition of the norm on M

(A) requires some discussion.

12
More generally, the theory of operator algebras was decisively inuenced by quantum mechanics and quantum
eld theory.
13
John von Neumann (19031957) was a Hungarian prodigy; he wrote his rst mathematical paper at the age of
seventeen. Except for this rst paper, his early work was in set theory and the foundations of mathematics. In the
Fall of 1926, he moved to G ottingen to work with Hilbert, the most prominent mathematician of his time. Around
1920, Hilbert had initiated his Beweistheory, an approach to the axiomatization of mathematics that was doomed
to fail in view of Godels later work. However, at the time that von Neumann arrived, Hilbert was mainly interested
in quantum mechanics.
6
on some phase space, forming a commutative algebra under pointwise multiplication, similar to
C(X) above.
14
According to Heisenberg, the passage from classical mechanics to quantum me-
chanics is achieved by replacing such classical observables by quantum observables. Heisenberg
introduced a mathematical structure for the latter, including a multiplication rule which he im-
mediately recognized as being noncommutative in nature. His boss in G ottingen was Born, who
was one of the few physicists of his day to be familiar with the concept of a matrix, and saw that
Heisenbergs quantum observables were innite matrices. Thus Heisenbergs version of quantum
mechanics was initially known as matrix mechanics. Born turned to his former teacher Hilbert
for mathematical advice.
15
Aided by his assistants Nordheim and von Neumann, Hilbert thus
ran a seminar on the mathematical structure of quantum mechanics, and the three wrote a joint
paper on the subject (now obsolete). In any case, the fact that quantum mechanics involved some
noncommutative algebra was seen as a crucial feature of the theory.
The second half of quantum mechanics, discovered in 1926 by Schr odinger, was called wave
mechanics, in which the famous symbol , denoting a wave function, played an important role.
The possible relationship between these alternative formulations of quantum mechanics, which at
rst sight looked completely dierent, was much discussed at the time. It was von Neumann who,
in 1927 at the age of 23, found the mathematical structure of quantum mechanics, and claried
the relationship between the work of Heisenberg and that of Schr odinger.
In this process, he dened the abstract concept of a Hilbert space, which previously had only
appeared in some examples. These examples went back to the work of Hilbert and his pupils
in Gottingen on integral equations, spectral theory, and innite-dimensional quadratic forms.
Hilberts famous memoirs on integral equations had appeared between 1904 and 1906. In 1908,
his student E. Schmidt had dened the space
2
in the modern sense, and F. Riesz had studied
the space of all continuous linear maps on
2
in 1912. Various examples of L
2
-spaces had emerged
around the same time. However, the abstract notion of a Hilbert space was missing until von
Neumann provided it.
Von Neumann saw that Schr odingers wave functions were unit vectors in a certain Hilbert
space,
16
and that Heisenbergs quantum observables were linear operators on a dierent Hilbert
space.
17
A unitary transformation between these spaces then provided the the mathematical
equivalence between wave mechanics and matrix mechanics.
18
In a series of papers that appeared
between 19271929, von Neumann dened Hilbert space, formulated quantum mechanics in this
language, and developed the spectral theory of bounded as well as unbounded normal operators
on a Hilbert space. This work culminated in his book [13], which to this day remains the denitive
account of the mathematical structure of elementary quantum mechanics.
19
Since certain crucial
aspects of von Neumanns formulation of quantum mechanics have found their way into the theory
of C

-algebras, it is of interest to review this formulation.

The quantum observables of a given physical system are the selfadjoint linear operators a on
a Hilbert space H. The states of the system are the so-called density operators on H, that is,
the positive trace-class operators on H with unit trace. The expectation value of an observable
a in a state is given by Tr (a). For example, if is the (orthogonal) projection [v] on a unit
vector v (expressing the statement that the system is in the state v), and a is the projection [w]
on a unit vector w (seen as the elementary observable corresponding to the yes-no question: is
the system in the state w?), then the number Tr ([v][w]) = [(v, w)[
2
is what physicists call the
transition probability between v and w. This gives the probability that a system that has been
prepared in the state [v] is actually found to be in the state [w].
Let B(H) be the space of all bounded operators on H (whose selfadjoint elements are the
14
In classical mechanics one usually works with smooth functions on a manifold.
15
Hilbert had been interested in the mathematical structure of physical theories for a long time; his Sixth Problem
(1900) called for the mathematical axiomatization of physics.
16
Which in the simplest case was L
2
(R
3
).
17
Typically
2
.
18
Similar, mathematically incomplete insights had been reached by Pauli, Schrodinger, and Dirac.
19
Von Neumanns book was preceded by Diracs The Principles of Quantum Mechanics (1930), which contains
another brilliant, but mathematically questionable account of quantum mechanics in terms of linear spaces and
operators.
7
bounded observables), with unit operator 1. We have already dened the notion of a state on a
general (unital) C

-algebra A in an abstract way; this denition is due to von Neumann in the

special case A = B(H). The relation with density operators is that any such operator denes
a state

on B(H) by

(a) = Tr (a). What about the converse? Von Neumann implicitly

assumed a certain continuity condition on states, which in modern terminology singles out the
so-called normal states on B(H). When H is nite-dimensional, any state is normal. In general,
von Neumann showed that a state on B(H) is normal i it is of the form =

for some density

matrix on H.
Von Neumann was familiar with Minkowskis convexity theory, and he recognized that his and
Schr odingers notion of a quantum-mechanical state could be reformulated in those terms. The
set of all states on B(H) is obviously convex, as is its subset of all normal states on B(H). Now
Minkowski had dened the extreme boundary of a convex set K as the set of all K that are
indecomposable, in the sense that if =
1
+ (1 )
2
for some 0 < < 1 and
1
,
2
K
then
1
=
2
= . (In some examples of compact convex sets in R
n
, the extreme boundary of K
coincides with its geometric boundary; cf. the closed unit ball. However, the extreme boundary of
an equilateral triangle consists only of its corners. If K fails to be compact, its extreme boundary
may be empty, as illustrated by the open unit ball in any dimension.)
Von Neumann saw that Schr odingers wave functions precisely correspond to the points in the
extreme boundary of the normal state space of B(H). More precisely, is an extreme point in this
convex set i it is a one-dimensional projection = [v] for some unit vector (or wave function)
v H.
20
In view of this, points in the extreme boundary of some state space are nowadays called
pure states, whereas all other states are said to be mixed.
Having dened Hilbert space and linear operators, von Neumann got bored with single oper-
ators, and turned to algebras thereof. In one of his papers on Hilbert space theory (1929), von
Neumann dened a ring of operators M (nowadays called a von Neumann algebra) as a

-subalgebra
of the algebra B(H) of all bounded operators on a Hilbert space H that contains the unit 1 and is
closed (i.e., sequentially complete) in the weak operator topology. This means that M is a unital
subalgebra of B(H) under operator mulitplication, that M is closed under the involution a a

,
and that if for some sequence a
n
M one has [(v, (a
n
a)w)[ 0 for all v, w H, then a M.
For example, B(H) is itself a von Neumann algebra. When H is nite-dimensional, any direct sum
of matrix algebras containing 1 is a von Neumann algebra.
Such a ring of operators is automatically closed also in the norm-topology, so that a von
Neumann algebra is a C

-algebra when it is studied with respect to this topology. However, the

natural topology on a von Neumann algebra is a dierent one, so that the theory of such operator
algebras is quite dierent from the theory of C

-algebras. Both remain closely related to quantum

theory. In the Bibliography we list some books on C

-algebras and von Neumann algebras.

2 Commutative C

-algebras
Our rst goal is to prove a slight generalization of the rst theorem of Gelfand and Naimark quoted
above:
Theorem 2.1 Every commutative C

-algebra A is isomorphic to C
0
(X) for some locally compact
Hausdor space X. Here C
0
(X) is seen as a C

-algebra under pointwise operations, the sup-norm

|f|

:= sup[f(x)[ [ x X,
and the involution f

(x) = f(x).
To explain the formulation of this theorem, we should clarify two issues. Firstly:
20
The passage from the unit vector v to the projection [v] loses the phase information in v, but this information
drops out of the expectation values (v, av) dened by v anyway.
8
Denition 2.2 A morphism between C

-algebras A, B is a (complex-) linear map : A B

such that
(ab) = (a)(b); (4)
(a

) = (a)

(5)
for all a, b A. An isomorphism is a bijective morphism. Two C

-algebras are isomorphic

when there exists an isomorphism between them.
A morphism between C

-algebras is usually called a

-homomorphism. One immediately
checks that the inverse of a bijective morphism is a morphism. It is remarkable, however, that a
the image of a morphism between C

-algebras is automatically closed (and hence a C

-algebra),
while an injective morphism (and hence an isomorphism) is automatically isometric. We will prove
this at a later stage; see section 10. For this reason the condition that an isomorphism be isometric
is not included in the denition. As a trivial example of this denition, note that the C

-algebra
given by all 2 2 matrices of the form
_
a 0
0 a
_
, a C,
is isomorphic to C.
Secondly, recall that a space X is called locally compact when each point has a compact
neighbourhood.
Denition 2.3 Let X be a locally compact Hausdor space X. The space C
0
(X) consists of all
continuous functions on X that vanish at innity in the sense that for each > 0 there is a
compact subset K X such that [f(x)[ < for all x outside K.
So when X is compact one trivially has C
0
(X) = C(X). When X is not compact, the sup-norm
can still be dened, as C
0
(X) C
b
(X), and just as for C(X) one easily checks that C
0
(X) is a
commutative C

-algebra in this norm. For X = R, this denition just means that f C

0
(R) when
lim
x
f(x) = 0.
The proof of Theorem 2.1 relies on the theory of commutative Banach algebras (cf. Denition
11.17). This class of algebras is of some interest in itself, as the following example shows. Consider
L
1
(R), with the usual linear structure, and norm
|f|
1
:=
_
R
dx[f(x)[.
The associative product dening the Banach algebra structure is convolution, that is,
f g(x) :=
_
R
dy f(x y)g(y).
Using Fubinis theorem on product integrals, one easily estimates
|f g|
1
|f|
1
|g|
1
,
which proves that L
1
(R) is a commutative Banach algebra. Furthermore, L
1
(R) admits an invo-
lution given by f

(x) := f(x), but note that this norm does not satisfy (94), so that L
1
(R) is
not a C

-algebra with respect to these operations. There is no unit in L

1
(R); it would have been
Diracs delta-function (i.e., the measure on R which assigns 1 to x = 0 and 0 to all other x), but
this does not lie in L
1
(R).
Each p R denes a character
p
on L
1
(R) by means of the Fourier transform

p
(f) :=

f(p) =
_
R
dxf(x)e
ipx
;
9
since it is a well-known property of the Fourier transform that it diagonalizes the convolution
product, in that

fg =

f g (the product on the right-hand side being pointwise multiplication). In
particular, the Gelfand transform f

f, which we recall to be dened by

f() := (f), is nothing
but the Fourier transform on characters of the form
p
. In fact, all characters turn out to be of
this form, so that (L
1
(R))

= R.
By the RiemannLebesgue lemma, one has

f C
0
(R). Also, as is easily shown, |

f|

|f|
1
.
These are typical properties of the Gelfand transform dened on commutative Banach algebras.
For a C

-algebra, the Gelfand transform should be isometric. This is not the case for the norm
| |
1
in which, as we have already remarked L
1
(R) fails to be a C

-algebra. However, we can dene

a new norm on L
1
(R) in such a way that the Gelfand transform is isometric by construction. In
other words, we put |f| := |

f|

. It is not dicult to show that this is indeed a norm, and that

both (93) and (94) are now satised. An appeal to the StoneWeierstrass theorem shows that the
image of L
1
(R) under the Gelfand transform is dense in C
0
(R). Hence, if we dene C

(R) as the
closure of L
1
(R) in the new norm | |, it follows that C

(R)

= C
0
(R). This is consistent with
Theorem 2.1, and the given argument is basically an outline of its proof in general.
Let us now turn to the proof of Theorem 2.1. The rst few steps are valid for general commu-
tative Banach algebras A. Recall
Denition 2.4 A character of a commutative (Banach) algebra A (over C) is a nonzero homo-
morphism : A C of algebras,
21
in particular,
(ab) = (a)(b). (6)
The set of all characters of A is called the structure space (A) of A.
22
This is a topological
space in the so-called Gelfand topology, dened as the weakest topology making all functions
(a), a A, continuous. In other words, the Gelfand topology is generated by all sets of the
form a
1
(O), a A, O C open.
It is easily shown that a neighbourhood base of the Gelfand topology is given by sets of the
form
O

a
1
,...,a
n
() := (A) [ [(a
i
) (a
i
)[ < i, (7)
where (A), > 0, n < , and a
i
A.
The relevance of the following result to the proof of Theorem 2.1 should be clear.
Proposition 2.5 If a commutative Banach algebra A has a unit 1, then (A) is a compact Haus-
dor space in the Gelfand topology.
23
The Hausdor property is an easy exercise. The compactness property is much harder. It relies
on the following lemma, which will be proved later on.
Lemma 2.6 Let A be a commutative Banach algebra with unit. Then each (A) is continu-
ous, with norm
|| = (1) = 1, (8)
so that (A) A

.
Let us note that, according to this lemma, (A) may not only be equipped with the Gelfand
topology, but also with the relative norm topology on A

. The latter topology will, however, only

be used in proofs. More importantly, recall from functional analysis that the weak

-topology on
A

is the topology of pointwise convergence, in which

n
when
n
(a) (a) in C for all
a A. This is the weakest topology in which all functions (a), a A, A

, are continuous;
a neighbourhood base is given by (7), where one now puts , A

. It follows that the Gelfand

topology on (A) coincides with the relative
24
weak

-topology on A

.
21
Since (A) is a homomorphism, it is linear by denition.
22
It is not obvious that A has any character. We will show in Lemma 2.9 below that (A) is not empty when A
is a C

-algebra.
23
If A has no unit, (A) will merely be locally compact, as we shall see.
24
I.e., restricted from A

to (A).
10
Now, it is an easy exercise to show that if A has a unit, then (A) is a closed subspace of the
unit ball
25
of A

in the weak

-topology (and also in the norm topology). Combining this with the
basic Banach-Alaoglu theorem of functional analysis,
26
Proposition 2.5 follows.
Apart from such technicalities, the fundamental idea of the proof of Theorem 2.1 is very simple.
As mentioned before, it consists of the Gelfand transform.
Proposition 2.7 Let A be a commutative Banach algebra with unit, and let (A) carry the
Gelfand topology.
27
The Gelfand transform a a, dened by
a() := (a), a A, (A), (9)
is an algebra homomorphism from A to C((A)) (equipped with pointwise operations).
The homomorphism property is trivial, since
( a

b)() := a()

b() = (a)(b) = (ab),

by denition of a character. Hence the only nontrivial claim is that a is a continuous function on
(A). Recall that any Banach space A is naturally embedded in its double dual A

:= (A

, so
A A

, by the very same formula a a, with a() := (a). Hence the Gelfand transform is
simply the restriction of this embedding to (A) A

.
Once again, we evoke a basic result of functional analysis without proof (cf. [15]). Namely,
under this embedding, A is precisely the subset of elements of A

that are continuous functionals

on A

with respect to the weak

-topology.
28
Of course, elements of A A

will still be continuous

when restricted to (A) A

, so Proposition 2.7 follows.

We can sharpen Proposition 2.7 when A is a C

-algebra.
Lemma 2.8 Let A be a commutative C

-algebra with unit, and equip C((A)) with the usual

structure of C(X) as a commutative C

-algebra, X = (A). Then the Gelfand transform is a

-homomorphism (i.e., a morphism of C

-algebras).
In view of Proposition 2.7 and Denition 2.2, we only need to show that

a

= a, and for this

purpose it suces that (a) R whenever a

= a, a A. To do so, we suppose that (a) = +i,

where , R. Note that b := a 1 is selfadjoint. Since (1) = 1, one has (b) = i. Hence for
t R one computes
[(b +it1)[
2
=
2
+ 2t +t
2
.
On the other hand, using || = 1 (so that [(c)[ |c|) and (94), we estimate
[(b +it1)[
2
|b +it1|
2
= |(b +it1)

(b +it1)| = |b
2
+t
2
| |b|
2
+t
2
.
Combining the displayed equations yields
2
+ t |b|
2
for all t R. This is impossible unless
= 0, so that (a) is real when a = a

. Consequently, by (9) the function

A is real-valued.
To prove Theorem 2.1, it remains to be shown that the Gelfand transform is a bijection on a
commutative C

-algebra (given this, it is trivial that its inverse is again a

-homomorphism). We
will prove both injectivity and surjectivity from the most dicult part of the entire proof:
29
Lemma 2.9 Let A be a commutative C

-algebra with unit. Then the Gelfand transform is iso-

metric, that is,
| a|

= sup
(A)
[(a)[ = |a|.
25
The unit ball is the set of all functionals A

for which 1.
26
This states that for any Banach space A the unit ball in A

is compact in the weak

-topology (but not in the

norm topology, except when A is nite-dimensional!). See, e.g., [15].
27
Clearly, we can now say that this is the weakest topology making all a, a A, continuous.
28
Whereas A

by denition consists of all functionals on on A

that are continuous with respect to the norm-

topology on A

.
29
Dont get confused: an isomorphism between two C

-algebras is automatically isometric, but in the present

proof we prove the bijectivity property of an isomorphism by showing that the Gelfand transform is isometric.
11
Given this lemma, injectivity is obvious. Surjectivity follows from the Stone-Weierstrass theo-
rem, which we now recall.
30
Let X be a compact Hausdor space. A subalgebra

A of C(X) (regarded as a commutative
C

-algebra in the usual way) that

1. separates points on X (i.e., if x ,= y there is f

A such that f(x) ,= f(y));
2. is closed under complex conjugation (i.e., if f

A then f

A);
3. contains the unit function 1
X
(where 1
X
(x) = 1 for all x X);
is dense in C(X) in the sup-norm.
Applying this to X = C() and

A as the image of the Gelfand transform on A, it is trivial
that property 1 holds; property 2 follows from Lemma 2.8, and property 3 is a consequence of
(8). Furthermore, if the Gelfand transform is isometric,

A must be closed in C((A)), so that it
coincides with C((A)) by the Stone-Weierstrass theorem. This proves surjectivity.
To prove Lemmas 2.6 and 2.9, and also as the technical basis for much of the theory of general
C

-algebras, we need to develop some spectral theory. The basics are valid for general Banach
algebras, so we briey return to this setting, specializing to the C

-case as appropriate.
3 Spectral theory for Banach algebras
Until further notice, let A be a Banach algebra with unit.
Denition 3.1 The resolvent (a) of a A is the set of all z C for which a z1 has a
(two-sided) inverse in A.
The spectrum (a) of a A is the complement of (a) in C; in other words, (a) is the set
of all z C for which a z1 has no (two-sided) inverse in A.
When A is the algebra of n n matrices, the spectrum is just the set of eigenvalues. For
A = B(H), Denition 3.1 reproduces the usual notion of the spectrum of an operator on a Hilbert
space. The third example, which we present as a lemma, is an easy exercise.
Lemma 3.2 For f A = C(X), one has (f) = f(x), x X.
These examples have a physical interpretation (also cf. the Introduction). In classical physics
the observables are real-valued functions on some phase space X, so the spectrum of an observable
is just the set of values it can assume. In quantum physics one assumes that the observables are
selfadjoint operators on a Hilbert space H, and one assumes that the values such an observable
can assume lie in its spectrum. Hence Denition 3.1 unies these to cases to some extent.
Theorem 3.3 The spectrum (a) of any element a A is
1. contained in the set z C[ [z[ |a|;
2. compact;
3. not empty.
The proof uses two lemmas. We assume that A is unital.
Lemma 3.4 When |a| < 1 the sum

n
k=0
a
k
converges to (1 a)
1
.
Hence (a z1)
1
always exists when [z[ > |a|.
30
This theorem is contained in elementary analysis courses for X = [0, 1]; for the general case see [15].
12
We rst show that the sum is a Cauchy sequence. Indeed, for n > m one has
|
n

k=0
a
k

k=0
a
k
| = |
n

k=m+1
a
k
|
n

k=m+1
|a
k
|
n

k=m+1
|a|
k
.
For n, m this goes to 0 by the theory of the geometric series. Since A is complete, the Cauchy
sequence

n
k=0
a
k
converges for n . Now compute
n

k=0
a
k
(1 a) =
n

k=0
(a
k
a
k+1
) = 1 a
n+1
.
Hence
|1
n

k=0
a
k
(1 a)| = |a
n+1
| |a|
n+1
,
which 0 for n , as |a| < 1 by assumption. Thus
lim
n
n

k=0
a
k
(1 a) = 1.
By a similar argument,
lim
n
(1 a)
n

k=0
a
k
= 1.
so that, by continuity of multiplication in a Banach algebra, one nally has
lim
n
n

k=0
a
k
= (1 a)
1
.
The second claim of the lemma follows because (a z)
1
= z
1
(1 a/z)
1
, which exists
because |a/z| < 1 when [z[ > |a|.
To prove that (a) is compact, it remains to be shown that it is closed.
Lemma 3.5 The set
G(A) := a A[ a
1
exists
of invertible elements in A is open in A.
Given a G(A), take a b A for which |b| < |a
1
|
1
. By (88) this implies
|a
1
b| |a
1
| |b| < 1. (10)
Hence a+b = a(1+a
1
b) has an inverse, namely (1+a
1
b)
1
a
1
, which exists by (10) and Lemma
3.4. It follows that all c A for which |a c| < lie in G(A), for |a
1
|
1
.
To resume the proof of Theorem 3.3, given a A we now dene a function f : C A by
f(z) := z a. Since |f(z +) f(z)| = , we see that f is continuous (take = in the denition
of continuity). Because G(A) is open in A by Lemma 3.5, it follows from the topological denition
of a continuous function that f
1
(G(A)) is open in C. But f
1
(G(A)) is the set of all z C where
z a has an inverse, so that f
1
(G(A)) = (a). This set being open, its complement (a) is closed.
Finally, dene g : (a) A by g(z) := (z a)
1
. For xed z
0
(a), choose z C such that
[z z
0
[ < |(az
0
)
1
|
1
. From the proof of Lemma 3.5, with a replaced by az
0
and c replaced
by a z, we see that z (a), as |a z
0
(a z)| = [z z
0
[. Moreover, the power series
1
z
0
a
n

k=0
_
z
0
z
z
0
a
_
k
13
converges for n by Lemma 3.4, because
|(z
0
z)(z
0
a)
1
| = [z
0
z[ |(z
0
a)
1
| < 1.
By Lemma 3.4, the limit n of this power series is
1
z
0
a

k=0
_
z
0
z
z
0
a
_
k
=
1
z
0
a
_
1
_
z
0
z
z
0
a
__
1
=
1
z a
= g(z).
Hence
g(z) =

k=0
(z
0
z)
k
(z
0
a)
k1
(11)
is a norm-convergent power series in z. For z ,= 0 we write |g(z)| = [z[
1
|(1a/z)
1
| and observe
that lim
z
1a/z = 1, since lim
z
|a/z| = 0 by Denition 11.9.3. Hence lim
z
(1a/z)
1
=
1, and
lim
z
|g(z)| = 0. (12)
Let A

be a functional on A; since is bounded, (11) implies that the function g

: z (g(z))
is given by a convergent power series, and (12) implies that
lim
z
g

(z) = 0. (13)
Now suppose that (a) = , so that (a) = C. The function g, and hence g

, is then dened on
C, where it is analytic and vanishes at innity. In particular, g

is bounded, so that by Liouvilles

theorem it must be constant. By (13) this constant is zero, so that g = 0 by Corollary 11.16. This
is absurd, so that (a) ,= C hence (a) ,= . This nishes the proof of Theorem 3.3.
For one thing, we are now in a position to prove Lemma 2.6. Indeed, we now know that
(a z)
1
exists when [z[ > |a|. Since a character (A) is a homomorphism of algebras, also
(a z) must be invertible, so that (a z) ,= 0, i.e., (a) ,= z. It follows that [(a)[ |a|, so
that || 1. But the property (1) = 1 was an easy exercise, so that nally || = 1.
We now return to Theorem 2.1. Dene the spectral radius r(a) of a A by
r(a) := sup[z[, z (a). (14)
From Theorem 3.3.1 one immediately infers
r(a) |a|. (15)
The proof of Theorem 2.1 relies on two properties, which are valid in any commutative Banach
algebra A with unit.
Proposition 3.6 For each a A one has
r(a) = lim
n
|a
n
|
1/n
; (16)
(a) = ( a), (17)
where (cf. Lemma 3.2) ( a) = (a) [ (A).
Before proving this proposition, we show how it implies Lemma 2.9, and therefore Theorem
2.1, at least in the unital case. Let A be a commutative C

-algebra with unit, and rst take a A

sucht that a

= a. Then (94) implies |a

2
| = |a|
2
, so |a| = r(a) from (16). On the other hand,
(17) implies r(a) = | a|

, so that |a| = | a|

. It is a very easy exercise to show that this implies

the same equality for general a A. This completes the proof of Theorem 2.1 in the unital case,
up to the proof of Proposition 3.6.
14
The proof of (16) is relatively easy (at least compared to the second part!). By Lemma 3.4, for
[z[ > |a| the function g in the proof of Lemma 3.5 has the norm-convergent power series expansion
g(z) =
1
z

k=0
_
a
z
_
k
. (18)
On the other hand, we have seen that for any z (a) one may nd a z
0
(a) such that the power
series (11) converges. If [z[ > r(a) then z (a), so (11) converges for [z[ > r(a). At this point
the proof relies on the theory of analytic functions with values in a Banach space, which says that,
accordingly, (18) is norm-convergent for [z[ > r(a), uniformly in z. Comparing with (15), this
sharpens what we know from Lemma 3.4. The same theory says that (18) cannot norm-converge
uniformly in z unless |a
n
|/[z[
n
< 1 for large enough n. This is true for all z for which [z[ > r(a),
so that
lim sup
n
|a
n
|
1/n
r(a). (19)
To derive a second inequality we use the following polynomial spectral mapping property.
Lemma 3.7 For a polynomial p on C, dene p((a)) as p(z)[ z (a). Then
p((a)) = (p(a)). (20)
To prove this equality, choose z, C and compare the factorizations
p(z) = c
n

i=1
(z
i
());
p(a) 1 = c
n

i=1
(a
i
()1). (21)
Here the coecients c and
i
() are determined by p and . When (p(a)) then p(a) 1 is
invertible, which implies that all a
i
()1 must be invertible. Hence (p(a)) implies that
at least one of the a
i
()1 is not invertible, so that
i
() (a) for at least one i. Hence
p(
i
()) = 0, i.e., p((a)). This proves the inclusion (p(a)) p((a)).
Conversely, when p((a)) then = p(z) for some z (a), so that for some i one must
have
i
() = z for this particular z. Hence
i
() (a), so that a
i
() is not invertible,
implying that p(a) 1 is not invertible, so that (p(a)). This shows that p((a)) (p(a)),
and (20) follows.
To conclude the proof of (16), we note that since (a) is closed there is an (a) for
which [[ = r(a). Since
n
(a
n
) by Lemma 3.7, one has [
n
[ |a
n
| by (15). Hence
|a
n
|
1/n
[[ = r(a). Combining this with (19) yields
lim sup
n
|a
n
|
1/n
r(a) |a
n
|
1/n
.
Hence the limit must exist, and
lim
n
|a
n
|
1/n
= inf
n
|a
n
|
1/n
= r(a).
To prove (17), we need to develop the theory of ideals in commutative C

-algebras. We will do
this directly also for the general noncommutative case, specializing to the commutative situation
as appropriate.
4 Ideals in Banach algebras and C

-algebras
We proceed in two stages, rst discussing ideals in general Banach algebras, and subsequently
passing to the special case of C

-algebras.
15
Denition 4.1 Let A be a Banach algebra.
A left ideal (right ideal) in A is a closed linear subalgebra J for which a J implies ba J
(ab J) for all b A.
An ideal in A is a subspace that is both a left and a right ideal (i.e., a two-sided ideal).
A maximal ideal is a proper ideal
31
J A for which J

J A for some proper ideal

J
implies

J = J.
32
Note that an ideal is by denition two-sided. Compared to the purely algebraic denition, an
ideal in a Banach algebra has to be closed by denition. This turns out to make an enormous
dierence, restricting the possible ideals. In particular, an ideal in a Banach algebra is itself a
Banach algebra. An ideal J that contains an invertible element a must coincide with A, since
a
1
a = 1 must lie in J, so that all B = B1 must lie in J. This shows the need for considering
Banach algebras with and without unit; it is usually harmless to add a unit to a Banach algebra
A, but a given proper ideal J ,= A does not contain 1, and one cannot add 1 to J without ruining
the property that it is a proper ideal.
Denition 4.1 literally applies to C

-algebras. One would expect that an ideal in a C

-algebra
is required to be selfadjoint, but this is turns out to be a consequence of the given denition.
Proposition 4.2 Let J be an ideal in a C

-algebra A. If a J then a

J; in other words, every

ideal in a C

-algebra is selfadjoint.
The proof of this theorem is surprisingly dicult, using techniques we have not yet discussed.
Hence we postpone it until section 10.
In the commutative case, left and right ideals are the same as ideals. For example, if A = C(X)
for a compact space X, then each closed subspace Y X denes an ideal
C(X, Y ) := f C(X) [ f(x) = 0 x Y . (22)
Note that C(X, Y ) is indeed closed by denition of the sup-norm, and that C(X, Y )

= C
0
(XY ).
In fact, it is an easy exercise that all ideals in C(X) are of this form. It is not necessary to assume
that Y is closed, but this assumption entails no loss of generality, since C(X, Y ) = C(X, Y ), where
Y is the closure of Y . We will see shortly that C(X, Y ) is maximal i Y is a point, and that all
maximal ideal in C(X) are of this form.
Proposition 4.3 If J is an ideal in a Banach algebra A then the quotient A/J is a Banach algebra
in the norm
|(a)| := inf
jJ
|a +j| (23)
and the multiplication
(a)(b) := (ab). (24)
Here : A A/J is the canonical projection. If A is unital and J is proper, then A/J is unital,
with unit (1).
We omit the standard proof that A/J is a Banach space in the norm (23). As far as the Banach
algebra structure is concerned, rst note that (24) is well dened: when j
1
, j
2
J one has
(a +j
1
)(b +j
2
) = (ab +aj
2
+j
1
b +j
1
j
2
) = (ab) = (a)(b),
since aj
2
+ j
1
b + j
1
j
2
J by denition of an ideal, and (j) = 0 for all j J. To prove (88),
observe that, by denition of the inmum, for given a A, for each > 0 there exists a j J such
that
|(a)| + |a +j|. (25)
31
A proper ideal of A is an ideal that is neither A nor 0.
32
Note that a maximal ideal J A dened in the purely algebraic sense is automatically closed. To prove this,
note that the closure J of J cannot be A, since J does not contain any invertible element of A (otherwise it would
coincide with A), and the set G(A) of all invertible elements in A is open (see Lemma 3.5). Since J J A and
J is maximal, J = J.
16
For if such a j would not exist, then |(a)| |a + j| for all j J, violating (23). On the
other hand, for any j J it is clear from (23) that
|(a)| = |(a +j)| |a +j|. (26)
For a, b A choose > 0 and j
1
, j
2
J such that (25) holds for a, b, and estimate
|(a)(b)| = |(a +j
1
)(b +j
2
)| = |((a +j
1
)(b +j
2
))|
|(a +j
1
)(b +j
2
)| |a +j
1
| |b +j
2
|
(|(a)| +)(|(b)| +). (27)
Letting 0 yields
|(a)(b)| |(a)| |(b)|. (28)
When A has a unit, it is obvious from (24) that (1) is a unit in A/J. By (26) with a = 1 and
j = 0 one has |(1)| |1| = 1. On the other hand, from (28) and (24) with b = 1 and a AJ
one derives |(1)| 1. Hence |(1)| = 1.
Like Proposition 4.2, the proof of the following C

-algebraic analogue of Proposition 4.3 is

dicult, so we postpone it until section 10 as well.
Proposition 4.4 Let J be an ideal in a C

-algebra A. The quotient A/J is a C

-algebra with
respect to the norm (23), the multiplication (24), and the involution
(a)

= (a

). (29)
In the commutative case, the basic example (with X and Y compact, as above), is
C(X)/C(X, Y )

= C(Y ), (30)
as two elements f, g of C(X) are identied in C(X)/C(X, Y ) when f g C(X, Y ), i.e., when they
coincide on Y . If one looks at C(X, Y ) as the kernel of the restriction map r
Y
: C(X) C(Y ),
then elementary linear analysis shows that ran(r
Y
)

= C(X)/ ker(r
Y
), which is just (30).
We now return to the proof of Theorem 2.1, which was complete up to the proof of (17). This
hinges on the following key result.
Theorem 4.5 Let A be a unital commutative Banach algebra with structure space (A).
1. If (A) then J

:= ker() is a maximal ideal in A;

2. Every maximal ideal is of this form;
3.
1
=
2
i J

1
= J

2
.
Hence there is a bijective correspondence between (A) and the set of all maximal ideals in A.
The rst and the third claim are easy exercises. Before demonstrating the second claim, let us
show how Theorem 4.5 implies (17). Namely, precisely the dicult second claim enters the proof
of a crucial lemma.
Lemma 4.6 One has (a) ,= 0 for all (A) i a G(A) (i.e., a is invertible).
If a G(A) then (a)(a
1
) = (aa
1
) = (1) = 1, so that (a) ,= 0. Conversely, assume
there exists a / G(A) such that (a) ,= 0 for all (A). The assumption a / G(A) implies that
the ideal
J
a
= ab [ b A
17
generated by a does not contain 1, so that J
a
,= A. By abstract nonsense
33
there must be a maximal
ideal J A containing J
a
. By Theorem 4.5.2, it must be that J = J

for some (A), so that

J
a
ker(). Hence (ab) = (a)(b) = 0 for all b, and since (a) ,= 0 by assumption, it follows
that (b) = 0 for all b A, hence = 0. This contradicts the denition of (A).
Eq. (17) easily follows from this lemma: it implies that z (a), that is, a z G(A), i
(a z) ,= 0 for all (A), so that z (a) i (a) ,= z for all . Taking the complement
(a) = C(a), we obtain
(a) = (a) [ (A).
But (a) = a() by denition of the Gelfand transform, so that (17) follows from Lemma 3.2.
This concludes the proof of Theorem 2.1 in the unital case.
We now prove Theorem 4.5.2. Let J be a maximal ideal. Since J ,= A, there is a nonzero b A
which is not in J. Form
J
b
:= ba +j[, a A, j J.
Since A is commutative, J
b
is an ideal. Taking a = 0 we see J J
b
. Taking a = 1 and j = 0 we
see that b J
b
, so that J
b
,= J. Hence J
b
= A, as J is maximal. In particular, 1 J
b
, so that
1 = ba +j for some a A, j J. Applying the canonical projection : A A/J to this equation
gives
(1) = 1 = (ba) = (b)(a),
because of (24) and (J) = 0. hence (a) = (b)
1
in A/J. Since b was arbitrary (though nonzero),
this shows that every nonzero element of A/J is invertible, since every such element is of the form
(b) for some b / J.
The structure of commutative Banach algebras in which every nonzero element is invertible is
described by the GelfandMazur theorem.
Theorem 4.7 If every nonzero element of a unital Banach algebra A is invertible, then A

= C as
Banach algebras.
By Theorem 3.3.3 one has (a) ,= for each a ,= 0, hence there is a z C for which az is not
invertible. Hence a z = 0 or a = z1 by assumption, and the map a z is the desired algebra
isomorphism.
Continuing the proof of Theorem 4.5.2, we have A/J

= C, so that there is a homomorphism
: A/J C. Now dene a map : A C by (a) := ((a)). This map is clearly linear, since
and are. Also,
(a)(b) = ((a))((b)) = ((a)(b)) = ((ab)) = (ab),
because of (24) and the fact that is a homomorphism.
Therefore, is multiplicative; it is nonzero because (b) ,= 0, or because (1) = 1. Hence
(A). Finally, trivially
J = ker() ker( ) = ker().
But if b / J then (b) ,= 0, hence (b) = (b) ,= 0, since is an isomoprhism. So J
c
(ker())
c
,
or ker() J. Finally, J = ker().
33
Let C be the collection of all proper ideals containing J
a
. This is a partially ordered set by inclusion. By
Hausdors maximality theorem, C has a maximal totally ordered subcollection T. Then the union J of all members
of T is an ideal in A. It is clear that J
a
J, whereas J = A because none of the elements of C contains 1, while A
does. Since C is maximal, J is a maximal ideal. Note that this argument, and thereby the whole of Gelfands theory
of commutative Banach algebras depends on the axioms of choice (which is equivalent to Hausdors maximality
theorem).
18
5 C

-algebras without unit

We still need to prove Theorem 2.1 for the nonunital case, and will use this opportunity to introduce
a general technique for handling C

-algebras without a unit. As mentioned above, such C

-algebras
are quite important whenever ideals play a role.
The discussion of general (not necessarily commutative) Banach algebras without unit is easier
than that of C

-algebras. When a Banach algebra A does not contain a unit, we can always add
one, as follows. Form the vector space AC, and turn this into an algebra by means of
(a +1)(b +1) := ab +b +a +1, (31)
where we have written a +1 for (a, ), etc. In other words, the number 1 in C is identied with
the unit 1. Furthermore, dene a norm on AC by
|a +1| := |a| +[[. (32)
In particular, |1| = 1. Using (93) in A, as well as 11.9.3, one sees from (31) and (32) that
|(a +1)(b +1)| |a| |b| +[[ |b| +[[ |a| +[[ [[ = |a +1| |b +1|,
so that AC is a Banach algebra with unit. Since by (32) the norm of a A in A coincides with
the norm of a + 01 in AC, we have shown the following.
Proposition 5.1 For every Banach algebra without unit there exists a unital Banach algebra

A,
the unitization of A, and an isometric (hence injective) morphism A

A, such that

A/A

= C.
The unitization of A with the given properties is not unique. Indeed, when A is a C

-algebra
it is natural to equip AC with the involution
(a +1)

:= a

+1. (33)
However, if AC is normed through (32) it fails to be a C

-algebra, since (94) is not satised.

34
To guarantee that AC is a unital C

-algebra one needs a dierent norm, constructed as follows.

Recall Denition 11.12.
Proposition 5.2 Let A be a C

-algebra.
1. The map : A L(A) given by
(a)b := ab (34)
establishes an isomorphism between A and (A) L(A).
2. When A has no unit, dene a norm on AC by
|a +1| := |(a) +1|, (35)
where the norm on the right-hand side is the operator norm (82) in L(A), and 1 on the
right-hand side is the unit operator in L(A). With the operations (31) and (33), the norm
(35) turns AC into a C

-algebra with unit, called

A.
By (93) we have |(a)b| = |ab| |a| |b| for all b, so that |(a)| |a| by (82). On the
other hand, using (94) and (95) we can write
|a| = |aa

|/|a| = |(a)
a

|a|
| |(a)|;
in the last step we used (84) and |(a

/|a|)| = 1. Hence
|(a)| = |a|. (36)
34
This may already be seen in the simplest of all examples A = C.
19
Being isometric, the map must be injective; it is clearly a homomorphism, so that we have proved
the rst claim of the proposition.
It is clear from (31) and (33) that the map a + 1 (a) + 1 (where the symbol 1 on the
left-hand side is dened below (31), and the 1 on the right-hand side is the unit in L(A)) is a
morphism. Hence the norm (35) satises (93), because (88) is satised in the Banach algebra
L(A). Moreover, in order to prove that the norm (35) satises (94), by Lemma 11.20 it suces to
prove that
|(a) +1|
2
|((a) +1)

((a) +1)| (37)

for all a A and C.
To do so, we use a trick similar to the one involving (25), but with inf replaced by sup. Namely,
in view of (82), for given c L(A) and > 0 there exists a b A, with |b| = 1, such that
|c|
2
|c(b)|
2
. Applying this with c = (a) +1, we infer that for every > 0 one has
|(a) +1|
2
|((a) +1)b|
2
= |ab +b|
2
= |(ab +b)

(ab +b)|.
Here we used (94) in A. Using (34), the right-hand side may be rearranged as
|(b

)(a

+1)(a +1)b| |(b

)| |((a) +1)

((a) +1)| |b|.

Since |(b

)| = |b

| = |b| = 1 by (36) and (95), and |b| = 1 also in the last term, the inequality
(37) follows by letting 0.
Hence the C

-algebraic version of Theorem 5.1 is

Proposition 5.3 For every C

-algebra without unit there exists a unique unital C

-algebra

A and
an isometric (hence injective) morphism A

A, such that

A/A

= C.
The uniqueness of

A will follow from Corollary 8.5 below.
In the commutative case, the unitization procedure has a simple topological meaning, which
illustrates the general principle that the use of commutative C

-algebras often allows one to trade

topological properties for algebraic ones. Recall that the one-point compactication

X of a
non-compact topological space X is the set

X = X , whose open sets are the open sets in
X plus those subsets of X whose complement is compact in X. The injection i : X

X
is continuous, and any continuous function f C
0
(X) extends uniquely to a function f C(

X)
satisfying f() = 0. The space

X is the solution (unique up to homeomorphism) of a so-called
universal problem by Alexandros theorem: If : X Y is a map between locally compact
Hausdor spaces such that Y f(X) is a point and f is a homeomorphism onto its image, then
there is a unique homeomorphism :

X Y such that = i.
The proof of the following lemmas is an easy exercise.
Lemma 5.4 Let A = C
0
(X) for some noncompact locally compact Hausdor space X. Then

A

= C(

X), where 1

A is identied with 1

X
C(

X). Conversely, removing C1

X
from C(

X)
corresponds to removing C from

A = AC (as a vector space), leaving one with C
0
(X).
Hence the unitization of C
0
(X) corresponds to the one-point compactication of X.
Lemma 5.5 Let A be a commutative C

-algebra without unit.

1. Each (A) extends to a character on

A by
(a +1) := (a) +. (38)
2. The functional

on

A, dened by

(a +1) := , (39)
is a character of

A.
20
3. There are no other characters on

A.
4. (

A) is homeomorphic to the one-point compactication of (A).
In fact, this lemma is true for any commutative Banach algebra, with respect to any unitization.
We are now in a position to prove Theorem 2.1 also in the nonunital case. Applying the unital
case of Theorem 2.1 to

A and using Lemma 5.5, one nds

A

= C(

X) with X := (A). Removing
C from

A = AC precisely leaves one with C
0
(X) by Lemma 5.4, so that nally A

= C
0
(X).
Note that the Gelfand transform on a commutative C

-algebra without unit indeed takes values

in C
0
((A)), since by (39) one has a(

) =

(a + 0) = 0 for the (unique continuous) extension

of a from (A) to its one-point compactication.
6 Categorical version of the rst GelfandNaimark theorem
We have seen that any compact Hausdor space X denes a commutative C

-algebra with unit

C(X). As mentioned in the Introduction, the association X C(X) is contravariant in the
following sense: a map
35
: X Y denes a morphism

: C(Y ) C(X) of C

-algebras via
the pullback, that is, for f C(Y ) one puts

(f) = f C(X). This, of course, relies on the

fact that the composition of two continuous functions is again continuous. This association has
the property ( )

: C(Z) C(X), where

: C(Z) C(Y ) is the pullback of a

map : Y Z. Furthermore, the identity map id
X
: X X induces the identity map on C(X).
In the opposite direction, a commutative C

-algebra with unit A denes a compact Hausdor

space (A) in a similar contravariant way: since (A) A

C(A), one may equally well dene

the pullback of a unit-preserving morphism : A B as

: (B) (A); this relies on the

fact that the composition of a character with a morphism is again a character. Also here one has
()

: (C) (A), where

: (C) (B) is the pullback of a unital morphism

: B C.
What is the precise relationship between the maps C and ? They are not inverse to each
other; as we shall see, one has (C(X))

= X and C((A))

= A, where the symbol

= stands
for homeomorphism and isomorhism, respectively. Moreover, the specic homeomorphims X
(C(X)) for each X turn out to be well-behaved with respect to maps f : X Y , ans similarly
for the isomorphisms A C((A)). To describe the situation in an optimal way, we use the
formalism of category theory, which we briey recall.
Denition 6.1 A category C consists of:
A class
36
C
0
of objects;
A class C
1
of arrows;
maps s : C
1
C
0
(the source map), t : C
1
C
0
(the target map), i : C
0
C
1
(the
inclusion map), and multiplication
m : C
2
:= C
1

C
0 C
1
= (, ) C
1
C
1
[ s() = t() C
1
.
We write x

y when C
1
satises s() = x and t() = y, and or for m(, ).
These maps are subject to the following axioms:
1. s(i(x)) = t(i(x)) = x for all x C
0
;
2. If , C
2
then s() = s() and t() = t();
3. If (, ) C
2
and (, ) C
2
then () = ();
35
Recall Footnote 2; a map is continuous by denition.
36
We refrain from dening the concept of a class; this would lead us into logic and the foundations of mathematics.
We just mention that a class may be larger than a set.
21
4. i(t()) = i(s()) = for all C
1
.
We interpret as an arrow from s() to t(), so that i(x) is an arrow from x to x. The composition
of arrows is dened whenever s() = t() (so that on paper the direction of an arrow is
from right to left!). Arrow composition is associative whenever dened, and each i(x) acts as an
identity under this composition operation. The class of all arrows from x to y in a category C is
sometimes written as (x, y)
C
.
The simplest example of category is the category S of all sets, where S
0
consists of sets and S
1
consists of functions.
37
Of course, in the present context our basis examples are the category CH of
compact Hausdor spaces as objects and maps as arrows, and the category CC
1
A of commutative
C

-algebras with unit as objects and unital morphisms as arrows. The former is a subcategory (as
dened in the obvious way) of the category T of all topological spaces and maps, and the latter is
a subcategory of the category CA of all C

-algebras and morphisms.

An important feature of category theory is that the notion of isomorphism is intrinsically
dened: one calls two objects x, y C
0
isomorphic, written x

= y, when there exist x

y and
y

x such that = i(y) and = i(x). Thus two topological spaces are isomorphic in T when
they are homeomorphic,
38
and two C

-algebras are isomorphic in CA when they are isomorphic

according to Denition 2.2.
Note that a groupoid is a category
39
in which all arrows are isomorphisms, i.e., are invertible.
Groupoids are usually denoted by G, and elements of G
1
are typically called .Thus, further to
the axioms for a category, in a groupoid one has a map I : G
1
G
1
, called the inverse, also
written I(f) = f
1
, satsfying s I = t, t I = s,
1
= i(s()), and
1
= i(t()). Groupoids
provide interesting examples of categories, and may help getting used to the latter concept.
40
Moreover, at a later stage well show how, as independently discovered by Connes and Renault,
certain groupoids canonically dene C

-algebras. This construction is of paramount importance

for noncommutative geometry. Here are some examples.
1. The simplest example of a groupoid comes from a set X without additional structure. One
takes G
1
= G
0
= X, and s = t = i = id, with multiplication xx = x for all x X and
inverse x
1
= x. Thus each point of X is the unique arrow to and from itself, and there are
no other arrows. One says that as a groupoid X is a space. As we shall see, when X is a
locally compact space, the C

-algebra associated to this groupoid is C

(X) = C
0
(X).
2. A set X denes another groupoid, called the pair groupoid over X. Here one takes G
1
=
X X and G
0
= X, with s(x, y) = y, t(x, y) = x, i(x) = (x, x), (x, y) (y, z) = (x, z), and
(x, y)
1
= (y, x). When X is nite with cardinality n, the associated C

-algebra turns out

to be C

(X X) = M
n
(C).
3. An equivalence relation on a set X denes a subgroupoid (X, ) of the pair groupoid
over X; here G
0
= X, while G
1
consists of those pairs (x, y) X X for which x y.
When x y for any (x, y) one recovers the pair groupoid over X; on the other hand, when
x x and nothing else the groupoid dened by the equivalence relation is X as a space. It
is a useful exercise that the axioms for an equivalence relation imply that the denition of a
groupoid is satised.
41
37
Here one sees that a class is larger than a set, since the set of all sets does not exist, whereas the category of all
sets does.
38
In algebraic topology one works in the homotopy category hT, in which the objects are topological spaces and
the arrows are homotopy classes of continuous maps. In that case, isomorphism of objects comes down to homotopy
equivalence of spaces.
39
It is often required that that this category is small, in that its class of arrows is a set.
40
Note, though, that the categories CC
1
A and CH fail to be groupoids by a long shot!
41
The fundamental idea of Connes, which underlies about half of noncommutative geometry, is as follows. When
X is a locally compact Hausdor space, the quotient X/ may be ill-behaved in being non-Hausdor; it may even
have no nontrivial open sets. In such cases, the space C
0
(X/ ) carries practically no information about X/ , not
even as a set. In such situations, Connes proposed to describe the quotient X/ by the C

-algebra C

(X, ) of
the groupoid dened by , which may have a rich structure. Note that a complication arises: the subset of X X
dened by often needs to be retopologized in order to dene C

(X, ). Also, in foliation theory, one does not in

fact work with C

(X, ) but with a closely related C

-algebra.
22
4. A group G is a groupoid with G
1
= G and G
0
= e. When G is nite, C

(G) is the group

ring CG. More generally, when G is locally compact C

(G) is the so-called convolution

C

-algebra of G, which lies at the basis of modern harmonic analysis.

We return to general categories. According to the well-known category theorist Peter Freyd,
Topology is the study of continuous maps; category theory is the study of functors.
Denition 6.2 Let G and H be categories. A covariant functor or simply functor : G H
is a pair of maps
i
: G
i
H
i
, i = 0, 1, such that
1. i
H

0
=
1
i
G
;
2. s
H

1
=
0
s
G
and t
H

1
=
0
t
G
;
3. If (, ) G
2
then
1
() =
1
()
1
().
It follows that
0
is in fact determined by
1
, since i is injective. Nonetheless, it is useful
to keep them apart. In any case, functors are simply homomorphisms between categories in the
obvious sense. Now, the map C : CH CC
1
A dened by C
0
(X) = C(X) and C
1
() =

fails
to be a functor, because in fact one has ( )

, as we have seen. Similarly, the map

: CC
1
A CH dened by
0
(A) = (A) and
1
() =

is not a functor for the same reason.

So in our application we rather need something like an anti-homomorphism.
Denition 6.3 Let G and H be categories. A contravariant functor : G H is a pair of
maps
i
: G
i
H
i
, i = 0, 1, such that
1. i
H

0
=
1
i
G
;
2. s
H

1
=
0
t
G
and t
H

1
=
0
s
G
;
3. If (, ) G
2
then
1
() =
1
()
1
().
Note that a (possibly contravariant) functor : G H preserves isomorphism of objects: if
x

= y in G then
0
(x)

=
0
(y) in H, since if x

y is invertible then so is
0
(x)

1
()

0
(y) by
the axioms.
It should be clear that C : CH CC
1
A and : CC
1
A CH are contravariant functors. To
describe the relationship between the maps C and , we need the idea of a natural transformation
between two functors. Indeed, according to the founders of category theory Samuel Eilenberg and
Saunders Mac Lane, Categories are dened to dene functors; functors are dened to dene natural
transformations.
Denition 6.4 A natural transformation between two functors : G H and : G H
(that are either both covariant or both contravariant) is a map : G
0
H
1
such that s(
x
) =
0
(x)
and t(
x
) =
0
(x) (in other words, is a collection of maps
0
(x)

x

0
(x)
xG
0), such that the
following diagram commutes for all arrows x

y:

0
(x)

x

0
(x)

1
()

_
1
()

0
(y)

0
(y)
The given functors are called naturally isomorphic,
42
written

= , when there exists a natural
transformation between them for which all arrows
x
are invertible (i.e., are isomorphisms).
43
42
Or equivalent.
43
Hence a natural transformation of functors between groupoids automatically denes a natural isomorphism.
23
It follows that if and are naturally isomorphic, then (x)

= (x) for all x G
0
, but this
condition in itself is not sucient to render and naturally isomorphic; the isomorphisms
between (x) and (x) must be compatible with the arrows, as expressed by the diagram in the
denition.
The original motivation for this denition was to nd a way of expressing the fact that the dual
V

of any nite-dimensional vector space V (over C) is isomorphic to V , but in an unnatural

way, in that an isomorphism depends on the coice of a basis, whereas the double dual V

is
isomorphic to V in a natural way, namely through the Gelfand transform v v from V to V

,
where v() := (v) for V

. This fact may now be expressed by saying that the functor

from the category of all nite-dimensional vector spaces to itself (with linear maps as arrows) is
equivalent to the identity functor.
Now the whole point is the following denition.
Denition 6.5 Two categories G, H are called equivalent, written G H, when there exist
functors : G H and : H G such that

= id
H
and

= id
G
. Similarly, G
and H are called anti-equivalent or dual when there exist contravariant functors with the same
properties.
44
Spelling out what this means, using Denition 6.4, yields the commutative diagram
x

x

0

0
(x)

_
1
()
y

0
(y)
(40)
for all x

y H, in which all arrows
x
are invertible, and similarly for G.
Theorem 6.6 The category CH of compact Hausdor spaces and continuous maps is dual to the
category CC
1
A of unital commutative C

-algebras and unital morphisms.

We already have the contravariant functors C : CH CC
1
A and : CC
1
A CH. According
to the diagram (40), we need isomorphisms
A
for any unital commutative C

-algebra such that

A

A
C((A))

B
C((B))
(41)
commutes for any unital morphism : A B, and isomorphisms
X
for any compact Hausdor
space X such that
X

X
(C(X))

Y
(C(Y ))
(42)
commutes for any map : X Y .
For
A
we take the Gelfand transform
A
(a) = a, with a() = (a), which we already know
to be an isomorphism by Theorem 2.1. Furthermore, it is an easy exercise that Diagram (41)
commutes. For
X
we take the evaluation map, i.e.,
X
(x) =
x
, where
x
(f) = f(x). Again it
is easy to show that Diagram (42) commutes. The only remaining step is to show that
X
is an
isomorphism in the category CH, which we now do. We rst prove injectivity, then surjectivity,
and nally continuity of
X
(the latter also of its inverse).
44
Similarly, G and H are isomorphic, G

= H, when = id
H
and = id
G
. This notion is rarely useful,
however. The natural notion of isomorphism in category theory is equivalence.
24
Since a compact Hausdor space is normal,
45
Urysohns lemma says that C(X) separates points
on X (i.e., for all x ,= y there is an f C(X) for which f(x) ,= f(y)). This immediately shows
that
X
is injective.
To prove surjectivity, suppose there is (C(X)) that is not of the form =
x
for some
x X. By Theorem 4.5, there then exists a maximal ideal J in C(X) such that J ,= ker(
x
) for
all x, in other words, for all x there is a f
x
J for which f
x
(x) ,= 0. For, if for some x there
would exist no such f
x
, then f(x) = 0 for all f J, so that J ker(
x
) and hence J = ker(
x
)
by maximality of J, contradicting the assumption. For each x, the set O
x
where f
x
is nonzero
is open, because f is continuous. This gives a covering O
x

xX
of X. By compactness, there
exists a nite subcovering O
x
i

i=1,...,N
. Then form the function g :=

N
i=1
[f
x
i
[
2
. This function
is strictly positive by construction, so that it is invertible.
46
But J is an ideal, so that, with all
f
x
i
J (since all f
x
J) also g J. But an ideal containing an invertible element must coincide
with C(X), contradicting the assumption that J is maximal. Hence
X
is surjective by reductio
ad absurdum.
The space X may be equipped with the new topology induced by
X
, in which the open sets
are of the form
1
X
(O), with O open in (C(X)) in the Gelfand topology. We claim that this
new topology on X is weaker than the original one.
47
Namely, for f C(X) one has

f
X
= f.
Therefore, since the Gelfand topology on (C(X)) is the weakest topology for which all Gelfand
transforms

f are continuous,
48
the new topology on X is the weakest topology for which all f are
continuous. But f was already continuous with respect to the given topology, so the claim follows.
Without proof we now state a result from topology.
Lemma 6.7 Let a set X be Hausdor in some topology T
1
and compact in a topology T
2
. If T
1
is
weaker than T
2
then T
1
= T
2
.
Since X is in fact compact and Hausdor in both topologies, we conclude from this lemma that
the new topology on X must coincide with the original one. In other words,
X
is a homeomor-
phism. This nishes the proof of Theorem 6.6.
Corollary 6.8 The space X in Theorem 2.1 is unique up to homeomorphism.
Firstly, it follows from the comment after Denition 6.3 that if A

= C(X) in CC
1
A for some
X CH
0
, then (A)

= (C(X)) in CH. But we have just seen that (C(X))

= X in CH, so
that (A)

= X, or X

= (A).
This conclusion would be invalid when X is not required to be Hausdor, in which case C(X)

=
C(Y ) in CC
1
A no longer implies X

= Y in CH.
We now examine the case of commutative C

-algebras possibly without unit, versus locally

compact Hausdor spaces that are possibly noncompact. It is clear that a slight diculty arises
here, since a map : X Y does not, in general, pull back to a morphism

: C
0
(Y ) C
0
(X).
For example, with Y equal to a point any f C(Y )

= C pulls back to a constant function on X,
which does not vanish at innity. Hence some restriction is necessary on the class of allowed maps
between locally compact Hausdor spaces.
Denition 6.9 A map : X Y between locally compact Hausdor spaces is proper when

1
(K) is compact for any compact set K Y . The category LCH consists of locally compact
Hausdor spaces as objects and proper maps as arrows.
49
Clearly, maps like c : R pt are not proper. We list some properties of proper maps : X
Y .
50
45
Recall that in a normal space any disjoint pair of closed sets is contained in a disjoint pair of open sets.
46
Recall that f C(X) is invertible i f(x) = 0 for all x X, in which case f
1
(x) = 1/f(x).
47
A topology T
1
is called weaker than a topology T
2
on the same set if any open set of T
1
contains an open set
of T
2
. This includes the possibility T
1
= T
2
.
48
Cf. Footnote 27.
49
This denition relies on the property that the composition of two proper maps is again proper.
50
See Bourbaki, General Topology Ch. 1-4, , 10.
25
1. A proper map is automatically closed.
51
2. is proper i it is closed and
1
(pt) is compact for any point pt Y .
3. If X is compact then any : X Y is proper.
4. is proper i its canonical extension :

X

Y , given by (
X
) =
Y
, is continuous.
52
5. The pullback

maps C
0
(Y ) into C
0
(X).
The algebraic counterpart of a proper map is as follows.
Denition 6.10 A morphism : A B between C

-algebras is called nondegenerate when

(A)B (i.e., the linear span of all expressions of the form (a)b, a A, b B) is dense in B.
Clearly, a unital morphism between unital C

-algebras is nondegenerate. Conversely, a non-

degenerate morphism : A B between unital C

-algebras is automatically unital: it follows

from the denition of a morphism that p := (1) is a projection in B, so that (A)B pB. Since
B = pB (1 p)B as a vector space, (A)B can only be dense in B when p = 1. Finally, a
nondegenerate morphism : A B between nonunital C

-algebras extends uniquely to a unital

morphism

:

A

B between the unitizations in the obvious way.
Using one-point compactications it is easy to check that the pullback

: C
0
(Y ) C
0
(X) of
a proper map : X Y is nondegenerate. This eventually leads to the following generalization
of Theorem 6.6:
Theorem 6.11 The category LCH of locally compact Hausdor spaces and proper continuous
maps is dual to the category CCA of commutative C

-algebras and nondegenerate morphisms.

As an exercise, this can be derived from Theorem 6.6 and the above results. The only tricky
point is the proof that LCH contains identities, i.e., that the map id : A A is nondegenerate.
In other words, is A
2
= AA dense in A? In fact:
Lemma 6.12 In any C

-algebra A one has A

2
= A.
This will be proved in section 8.
53
7 The structure of C

-algebras
Having said everything about commutative C

-algebras there is to say, we now turn to the general

case. It is an exercise (also cf. the Appendix) to prove the following:
Proposition 7.1 Let H be a Hilbert space. The algebra B(H) of all bounded operators on H is a
C

-algebra under the operator adjoint a a

, (cf. (92)) and the operator norm (83). Moreover,

each (operator) norm-closed

-algebra in B(H) is a C

-algebra.
Our goal is to prove the converse, which is conveniently stated in terms of the following concept.
Denition 7.2 A representation of a C

-algebra A on a Hilbert space H is a morphism :

A B(H). In particular, is a linear map satisfying (ab) = (a)(b) and (a

) = (a)

.
If we just say representation of a C

-algebra without further comment, we mean a represen-

tation on a Hilbert space. It is not obvious from this denition, but true, that the image (A) is
automatically norm-closed (and hence a C

-algebra) in B(H). This is a special case of the result

that the image of a morphism between C

-algebras is always closed (and hence a C

-algebra); see
section 10 below. As an example of a representation, consider A = C
0
(R) and H = L
2
(R), with
51
That is, the image of any closed set is closed.
52
This is true even when X and/or Y are/is compact.
53
The same proof shows that A
n
= A for any n N.
26
given by (f)(x) = f(x)(x).
54
This representation is injective; an example of a representa-
tion of A = C
0
(R) that isnt is given by (f) = f(x) for some x R. As another example, we
just mention that the unitary representation theory of locally compact groups is a special case of
the representation theory of C

-algebras, since for every such group G there exists a C

-algebra
C

(G) whose representations are in bijective correspondence with the unitary representations of
G, preserving direct sums, etc.
For later use, we dene the notion of equivalence of representations: two representations
1
(A)
and
2
(A) of a C

-algebra A on Hilbert spaces H

1
and H
2
are called equivalent or unitarily
equivalent when there is a unitary map u : H
1
H
2
such that
2
(a) = u
1
(a)u

for all a A.
In the previous example, this notion reduces to the unitary equivalence of group representations
as it is usually dened.
We now state the structure of C

-algebras.
Theorem 7.3 Every C

-algebra A admits an injective

55
representation : A B(H) on some
Hilbert space H. In other words, every C

-algebra is isomorphic to a norm-closed

-algebra in
B(H), for some Hilbert space H.
Since an injective morphism (and hence an injective representation) is automatically isometric,
this theorem yields the best characterization of C

-algebras one could possibly hope for. Note

that we now have two entirely dierent characterizations of commutative C

-algebras, given by
Theorems 2.1 and 7.3. These are, of course, related by applying Theorem 7.3 to A = C
0
(X).
56
The proof of Theorem 7.3 uses a beautiful construction, which is important in its own right.
This construction, in turn, relies on a notion from quantum mechanics. Let A = B(H) and take
H with || = 1. The functional

: a (, a) has the properties

(a

a) 0 for all a A
(positivity) and

(1) = 1 (normalization). More generally, let be a unital representation of a

unital C

-algebra A on some Hilbert space H.

57
Then any unit vector H denes a functional

on A by

(a) := (, (a)), (43)

which has the same two properties. This functional is called the vector state dened by . This
suggests the following concept.
Denition 7.4 A state on a unital C

-algebra A is a functional : A C that is positive, in

that (a

a) 0 for all a A, and normalized, in that (1) = 1.

A state on a C

-algebra algebra A without unit is a functional for which the canonical exten-
sion to the unitization

A, dened by
(a +1) := (a) +, (44)
is positive.
The main properties of states, which basically yield a number of equivalent denitions, are
summarized in the following results, whose proof is a straightforward exercise.
Proposition 7.5 Any positive functional on a C

-algebra is continuous.
A continuous functional on a unital C

-algebra is positive i || = (1).

A positive functional on a C

-algebra is a state i || = 1.
A state on a C

-algebra A without unit has a unique extension to a state on the uniti-

zation

A, given by (44).
54
This may be generalized to the case A = C
0
(X) and H = L
2
(X, ), where is a Borel measure on X.
55
One could call an injective representation faithful in the usual sense of the word. However, this terminology is
sometimes used for representations for whch (a

a) = 0 implies a = 0, so we avoid it here.

56
Theorem 7.3 does not imply that for A = C
0
(X) one may always choose H = L
2
(X, ), and in fact its proof
does not lead to this choice at all. Nonetheless, except in pathological cases one may indeed nd a Borel measure
on X for which the representation of C
0
(X) on L
2
(X, ) by multiplication operators is injective.
57
When A has no unit, should be nondegenerate for

to dene a state.
27
Conversely, the restriction of a state on

A to A is a state.
We now turn to the GNS-construction, initially assuming that A is unital. First, we call
a representation cyclic if its carrier space H contains a cyclic vector for ; this means
that the closure of (A) coincides with H. Such representation are the building blocks of any
representation.
58
The following theorem and its proof form the GNS-construction.
59
Theorem 7.6 Let be a state on a C

-algebra A. There exists a cyclic representation

of A
on a Hilbert space H

with cyclic unit vector

such that
(a) = (

(a)

) a A. (45)
Using (43), we could somewhat cryptically write this as =

.
The idea of the proof is simple; the devil is in the details of step 4.
1. Given a state on A, dene a bilinear form (, )
0
on A by
(a, b)
0
:= (a

b). (46)
It is easily shown from the positivity property of that this form is positive semi-denite
60
and sesquilinear.
61
Hence (, )
0
denes a pre-inner product on A. Its null space
N

= a A[ (a

a) = 0 (47)
is a left-ideal in A.
2. The form (, )
0
projects to an inner product
62
(, )

on the quotient A/N

. If p

: A
A/N

is the canonical projection, then by denition

(p

a, p

b)

:= (a, b)
0
. (48)
The Hilbert space H

is the completion of A/N

in this inner product.

3. The representation

(A) is initially dened on A/N

, that is, on all vectors of the

form p

b, b A, by

(a)p

b := p

ab. (49)
This is well dened because N

is a left ideal in A. It is trivial to show that

is linear
and satises

(ab) =

(a)

(b) and

(a)

(a

); the latter is proved by taking inner

products with arbitrary vectors pB and pc in A/N

.
4. We will show that each

(a) is bounded on A/N

. Hence

(A) may be extended to all of

H

by continuity.
63
It is easily shown that this extension still has the dening properties of
a representation.
5. The vector dened by

:= p

1 is cyclic, since

(a)

= p

a, hence

(A)

= A/N

,
which by construction is dense in H

.
6. Finally, (45) follows by a simple computation.
58
It may be shown that any non-degenerate representation is a direct sum of cyclic representations. Here one
says that a representation (A) is non-degenerate if (a)v = 0 for all a A implies v = 0.
59
For Gelfand-Naimark-Segal. This construction is very important in quantum eld theory and quantum statistical
mechanics.
60
This means that (a, a)
0
0 for all a A.
61
That is, (a, b)
0
= (b, a)
0
, which is equivalent to the property (a

) = (a).
62
Which by denition is a positive denite hermitian (=sequilinear) form.
63
This relies on a simple but important fact in functional analysis: if B
0
is a dense subspace of a Banach space B,
and a linear map a : B
0
B
0
is bounded, then a extends to a bounded operator a : B B by continuity. Namely,
if v

v in B with all v

B
0
, then {av

} is a Cauchy sequence in B, so that one may put av := limav

.
28
When A has no unit, we simply dene

(A) as the restriction of

(

A) to A.
To prove the claim in step 4, we compute |

(a)p

b|
2
, where a, b A. By (48) and (49) one
has |

(a)p

b|
2
= (b

ab). The key inequality is now

(b

ab) |a|
2
(b

b), (50)
whose proof we postpone until the end of section 9. But (b

b) = |p

b|
2
, so that |

(a)p

b|
|a| |p

b|, so that |

(a)| |a| from (83).

We now take up the proof of Theorem 7.3. Given a state on A, the GNS-construction
provides us with a representation

: A B(H

). If this representation were injective, we would

be nished, but rstly we do not know that any state exists on A at all, and secondly we have no
guarantee that

is indeed injective. The rst problem is resolved by the following lemma, which
we will prove once we have developed the theory of states.
Lemma 7.7 For any selfadjoint element a A (i.e., a

= a), there exists a state

a
on A such
that [
a
(a)[ = |a|.
In fact, it will be shown in section 8 that this lemma follows from the more powerful
Proposition 7.8 For any a A and (a) there is a state

S(A) for which

(a) = .
To prove this, assume that A has a unit; if not, pass to the unitization and use Proposition
7.5. Given a A and (a), dene a linear map

: Ca +C1 C by

(a +1) := +.
Since (a), one has + (a + 1); this easily follows from the denition of . Hence
(15) with a replaced by a + 1 implies [

(a + 1)[ |(a + 1)|. Since

(1) = 1, it
follows from Proposition 7.5 that | | = 1. By the Hahn-Banach Theorem 11.15, there exists an
extension

of

to A of norm 1. By Proposition 7.5,

is a state on A, which clearly satises

(a) =

(a) = .
To solve the second problem of the possible lack of injectivity of

, we proceed as follows.
First, note that if
1
and
2
are representations of A on Hilbert spaces H
1
and H
2
, respectively,
then there is an obvious notion of the direct sum
1

2
, which is a representation of A on H
1
H
2
;
one simply puts
(
1

2
)(a) :=
_

1
(a) 0
0
2
(a)
_
, a A.
This immediately extends to the direct sum of a nite number of Hilbert spaces and representations.
The direct sum of a countable number of Hilbert spaces H
i
is dened by

i
H
i
:=
_
v

i
H
i
[

i
|v
i
|
2
H
i
<
_
;
Similar to the proof of completeness of
2
, it is a simple exercise to prove that this space is complete
in the inner product
(v, w) :=

i
(v
i
, w
i
)
H
i
.
However, for the proof of Theorem 7.3 we unfortunately need uncountable direct sums, since
we wish to dene the direct sum of all GNS-representations of A, which is a huge number. To
detail, let S(A) be the state space of A, that is, the set of all states on A. We will need to make
sense of constructions such as
S(A)
H

and
S(A)

. To see what this means, let us say a

few words about S(A).
This set has the structure of a topological space in two ways, since by Proposition 7.5 one has
S(A) A

, so that S(A) inherits both the norm-topology and the weak

-topology of A

. In what
follows, we always regard S(A) as a topological space with respect to the latter.
64
Furthermore,
S(A) has the structure of a convex set in the vector space A

. That is, if , S(A) and

64
Recall that for nite-dimensional A there is no dierence between these topologies.
29
[0, 1] then + (1 ), initially dened in A

, actually belongs to S(A). This is immediate

from Proposition 7.5. Repeating this process, it follows that

i
p
i

i
belongs to S(A) when all
p
i
0 and

i
p
i
= 1, and all
i
S(A). Combining the two structures, it follows from the
Banach-Alaoglu theorem that S(A) is a compact convex set when A is unital.
65
For example, for A = M
2
(C) one has S(A)

= B
3
, as compact convex sets, where the three-ball
is dened as B
3
= (x, y, z) R
3
[ x
2
+y
2
+z
2
1. Even this space is already uncountable! For
A = C(X) it follows from the Riesz theorem of measure theory that S(A) consists of all probability
measures on X.
Now, to dene
S(A)
H

and the like, let S be any set, and let f : S R

+
be a positive
function on S. Dene

S
f() := sup
FS

F
f(),
where the supremum is taken over all nite subsets F of S. It is easy to show that if the left-hand
side is nite, then it can only have a countable number of nonzero terms. This gives meaning to
H
u
:=
S(A)
H

:=
_
_
_
v

S(A)
H

S(A)
|v

|
2
H

<
_
_
_
. (51)
One can show that H
u
is complete in the inner product
(v, w) :=

S(A)
(v

, w

)
H

;
note that this sum only has a countable number of terms. We then dene the universal repre-
sentation
u
(A) by

u
(A) :=
S(A)

(A); (52)
this means, of course, that
u
(a)v for v H
u
is dened by its components (
u
(a)v)

:=

(a)v

,
which again is a countable family. It is easy to see that this indeed is a representation.
The Hilbert space H
u
will be the H in the statement of Theorem 7.3. To prove that =
u
is injective, take some xed S(A), and dene the vector as having components

= 0 for
,= , and

equal to the cyclic vector of the GNS-construction for . This vector clearly lies in
H
u
. Then (
u
(a)) =

(a)

for = , and zero otherwise. Now suppose that

u
(a) = 0 for
some a A. Then
u
(a) = 0, hence

(a)

= 0, hence
|

(a)

|
2
= (a

a) = 0
by the GNS-construction. Taking =
a

a
, Lemma 7.7 implies
|a|
2
= |a

a| =
a

a
(a

a) = 0,
hence a = 0, so that
u
is injective. This concludes the proof of Theorem 7.3, up to the proof of
(50) and Lemma 7.7.
It should be noted that this proof relies on incredible overkill, in that H
u
is far larger than nec-
essary. For example, for A = M
n
(C) any vector state in C
n
gives rise to an injective representation,
namely the dening one.
To prove the inequality (50) and Lemma 7.7 we need to develop the theory of states, which
in turn relies on the notion of positivity in C

-algebra. This notion, in turn, depends on the

the functional calculus in C

-algebras. All these topics are of great interest in their own right as
well, and no course in C

-algebras would be complete without covering these seemingly technical

developments.
65
When A has no unit, the space of all positive functionals with 1 is a compact convex set, but S(A)
isnt.
30
8 Spectrum and functional calculus
Throughout this section, A is supposed to be unital. In the nonunital case, the same considerations
apply after passing to the unitization

A.
For each element a of a unital C

-algebra A there is a smallest C

-subalgebra C

(a) of A that
contains a and 1, namely the closure of the linear span of 1 and all operators of the type a
1
. . . a
n
,
where a
i
is a or a

. Following the terminology for operators on a Hilbert space, an element a A

is called normal when aa

= a

a. The crucial property of a normal operator is that C

(a) is
commutative. In particular, when a is selfadjoint, C

(a) is simply the closure of the space of all

polynomials in a.
Theorem 8.1 Let a be a normal element of a unital C

-algebra. Then the spectrum

A
(a) of a
in A coincides with the spectrum
C

(a)
(a) of a in C

(a) (so that we may unambiguously speak of

the spectrum (a)).
Recall that G(A) consists of the invertible elements in A. Let a G(A) be normal, and consider
the C

-algebra C

(a, a
1
, 1) generated by a, a
1
, and 1. One has (a
1
)

= (a

)
1
, and a, a

,
a
1
, (a

)
1
and 1 all commute with each other. Hence C

(a, a
1
, 1) is commutative; it is the
closure of the space of all polynomials in a, a

, a
1
, (a

)
1
, and 1. By Theorem 2.1 we have
C

(a, a
1
, 1)

= C(X) for some compact Hausdor space X. Since a is invertible and the Gelfand
transform (9) is an isomorphism, a is invertible in C(X). That is, a(x) ,= 0 for all x X. However,
for any f C(X) that is nonzero throughout X we have 0 < |f|
2

ff

1 pointwise, so that
0 1
X
|f|
2

ff < 1 pointwise, hence

|1
X
ff /|f|
2

< 1.
Here f

(x) = f(x). Using Lemma 3.4, in terms of 1 = 1

X
we may therefore write
1
f
=
f

|f|
2

k=0
_
1
ff

|f|
2

_
k
. (53)
Hence a
1
is a norm-convergent limit of a sequence of polynomials in a and a

. Gelfand transform-
ing this result back to C

(a, a
1
, 1), we infer that a
1
is a norm-convergent limit of a sequence of
polynomials in a and a

. Hence a
1
lies in C

(a), and C

(a, a
1
, 1) = C

(a).
Now replace a by a z, where z C. If a is normal, then a z is normal. So if we assume
that a z G(A) the argument above applies, leading to the conclusion that the resolvent
A
(a)
in A coincides with the resolvent
C

(a)
(a) in C

(a). By Denition 3.1 we then conclude that

A
(a) =
C

(a)
(a).
Corollary 8.2 If a

= a then the spectrum (a) is a subset of R.

According to Theorem 2.1, the function a is real-valued when a = a

. Hence by (17) the

spectrum
C

(a)
(a) is real, so that by the previous result (a) is real.
The commutative C

-algebra C

(a) is clearly important, so it is gratifying that its structure

can be described explicitly.
Proposition 8.3 Let a be normal. The structure space (C

(a)) is homeomorphic with (a), so

that C

(a) is isomorphic to C((a)). Under this isomorphism the Gelfand transform a : (a) C
is the identity function id
(a)
: t t.
Given the isomorphism C

(a)

= C(X) of Theorem 2.1 (where X = (C

(a))), according to
(17) the function a is a surjective map from X to (a). We now prove injectivity. When
1
,
2
X
and
1
(a) =
2
(a), then, for all n N, we have

1
(a
n
) =
1
(a)
n
=
2
(a)
n
=
2
(a
n
)
31
by iterating (6) with b = a. Since also
1
(1) =
2
(1) = 1 by (8), we conclude by linearity that

1
=
2
on all polynomials in a. By continuity (cf. Lemma 2.6) this implies that
1
=
2
on
C

(a), since the linear span of all polynomials is dense in C

(a). Using (9), we have proved that

a(
1
) = a(
2
) implies
1
=
2
.
Since a C(X), a is continuous. To prove continuity of the inverse, one checks that for z (a)
the functional a
1
(z) (C

(a)) maps a to z (and hence a

n
to z
n
, etc.). Looking at the denition
of the Gelfand topology in Denition 2.4, one then sees that a
1
is continuous. In conclusion, a is
a homeomorphism. The nal claim of the theorem is then obvious.
An immediate consequence is the continuous functional calculus.
66
Theorem 8.4 For each selfadjoint element a A and each f C((a)) there is an operator
f(a) A, which is the obvious expression when f is a polynomial (and in general is given via the
uniform approximation of f by polynomials), such that
|f(a)| = |f|

; (54)
(f(a)) = f((a)). (55)
In particular, the norm of f(a) in C

(a) coincides with its norm in A.

Still writing X = (C

(a)), there are maps a a from C

(a) to C(X) (i.e., the Gelfand

transform GT), a

: C((a)) C(X), where a is seen as a map from X to (a), and nally

the continuous functional calculus CFC as a map from C((a)) to C

(a). These maps form a

commutative diagram:
C

((a))
CFC
-
C

(a)
@
@
@
@
@
a

R
C(X)
GT
?
By Proposition 8.3, a

is an isomorphism, and so is GT by Theorem 2.1. Hence CFC is an

isomorphism by commutativity of the diagram. Furthermore, the rst two maps are isometric; the
rst by direct computation (any pullback

: C(Y ) C(X) of a homeomorphism : X Y is

isometric), and the second by Lemma 2.9. Hence the continuous functional calculus is isometric,
too. This yields (54).
Since f((a)) is the set of values of f on (a), (55) follows from (17), with a replaced by
f(a).
This leads us to a central result in the theory of C

-algebras.
Corollary 8.5 The norm in a C

-algebra is unique (that is, given a C

-algebra A there is no
other norm in which A is a C

-algebra).
First assume a = a

, and apply (54) with f = id

(a)
. By denition, the sup-norm of id
(a)
is
r(a), so that
|a| = r(a) (a = a

). (56)
Since a

a is selfadjoint for any a, for general a A we have, using (94),

|a| =
_
r(a

a). (57)
Since the spectrum is determined by the algebraic structure alone, (57) shows that the norm is
determined by the algebraic structure as well.
66
For operators on a Hilbert space there exist holomorphic and measurable functional calculi as well.
32
Note that Corollary 8.5 does not imply that a given

-algebra can be normed only in one way
so as to be completed into a C

-algebra. In Corollary 8.5 the completeness of A is assumed from

the outset.
As an immediate application of (56), let us derive Lemma 7.7 from Proposition 7.8, as promised.
Since (a) is closed by Theorem 3.3.2, there is an (a) for which r(a) = [[. For this one
nds

from Proposition 7.8, giving [

(a)[ = [a[ = r(A) = |A| by (56).

An important consequence of Corollary 8.5 is
Proposition 8.6 An isomorphism between C

-algebras is isometric.
Let : A B an isomorphism (cf. Denition 2.2). Dene a map | |

: A R
+
by
|a|

:= |(a)|
B
. It easily follows that this is a norm on A, in which A is complete: a sequence
a

in A is a Cauchy sequence with respect to | |

when (a

) is a Cauchy sequence in B.
The latter converges to some b B, so that a

converges to
1
(b). Furthermore, (4) and (5)
guarantee that the norm | |

satises (93) and (94). Thus A is a C

-algebra in | |

, so that
| |

= | |
A
.
To close this section, we prove Lemma 6.12. It is enough to show that any selfadjoint a A is of
the form a = a
1
a
2
, where A has no unit (otherwise the claim is trivial). Embed A

A and consider
C

(a)

A. Now factorize the function t t on (a) R as t = f
1
(t)f
2
(t) for some f
i
C((a)),
so that a = a
1
a
2
by Proposition 8.3 for a
i
:= CFC(f
i
) C

(a). Now the constant function 1

(a)
maps to 1 C

(a)

A under CFC, so if f C((a)) then CFC(f f(0)1
(a)
) C

(a) A.
67
Imposing the additional condition f
i
(0) = 0 therefore guarantees that in fact the a
i
lie in A.
9 Positivity in C

-algebras
A bounded operator a on a Hilbert space H is called positive when (v, av) 0 for all v H.
68
This property is equivalent to a

= a and (a) R
+
, and clearly also applies to closed subalgebras
of B(H).
Classically, a function f on some space X is called positive when it is pointwise positive, that
is, when f(x) 0 for all x X. This applies, in particular, to elements of the commutative
C

-algebra C
0
(X) (where X is a locally compact Hausdor space).
These examples are not as dissimilar as they might seem. Let PH be the projective Hilbert
space dened by H; this space ist most easily dened as the quotient of the unit sphere H
1
:=
v H [ (v, v) = 1 by the equivalence relation v w when w = zv for some z C with [z[ = 1.
Now, an operator a B(H) denes a function a C(PH) by means of a([v]) := (v, av); the
continuity of this function is an easy exercise. Hence we see that a is real-valed i a is self-adjoint
as an operator on H, and that a is a positive operator precisely when a is positive as a function
on PH.
69
Hence we have a notion of positivity for certain concrete C

-algebras, which we would like to

generalize to arbitrary abstract C

-algebras. In particular, one is interested in nding a number

of equivalent characterizations of positivity.
Denition 9.1 An element a of a C

-algebra A is called positive when a = a

and its spectrum

is positive; i.e., (a) R
+
. We write a 0 when a is positive, and A
+
for the set of all positive
elements in A.
It is immediate from Theorems (17) and Proposition 8.3 that a is positive i its Gelfand
transform a is positive in C((a)). The basic structure of A
+
is captured by the following denition.
67
If A has no unit, then necessarily 0

A
(a).
68
In quantum mechanics this means that the expectation value of the observable a is always positive.
69
This argument exhibits a crucial dierence between classical and quantum mechanics: if one regards PH as the
phase space of a quantum system, the only observables are functions of the form a for some self-adjoint a B(H).
This class by no means exhausts C(PH); for example, for H = C
2
one has PH

= S
2
, the two-sphere, on which the
space of all real-valued continuous (or even smooth) functions is clearly innite-dimensional.
33
Denition 9.2 A convex cone in a real vector space V is a subspace V
+
such that
1. when v V
+
and t R
+
then tv V
+
;
2. when v, w V
+
then v +w V
+
;
3. V
+
V
+
= 0.
A linear partial ordering in V is a partial ordering in which v w implies v + f w + f
for all f V and tv tw for all t R
+
.
These two structures are equivalent: given V
+
V one denes by putting v w if wv V
+
,
and given one denes V
+
= v V [ 0 v.
Proposition 9.3 The set A
+
of all positive elements of a C

-algebra A is a convex cone in the

real vector space A
sa
:= a A [ a

= a.
The rst property follows from (ta) = t(a), which is a special case of (55).
Since (a) [0, r(a)], we have [c t[ c for all t (a) and all c r(a). Hence
sup
t(a)
[c1
(a)
a[ c by (17) and Proposition 8.3, so that |c1
(a)
a|

c. Gelfand
transforming back to C

(a), this implies |c1 a| c for all c |a| by Theorem 8.4. Inverting
this argument, one sees that if |c1 a| c for some c |a|, then (a) R
+
.
Use this with a a +b and c = |a| +|b|; clearly c |a +b| by 11.9.4. Then
|c1 (a +b)| |(|a| a)| +|(|b| b)| c,
where in the last step we used the previous paragraph for a and for b separately. As we have seen,
this inequality implies a +b A
+
.
Finally, when a A
+
and a A
+
it must be that (a) = 0, hence a = 0 by (56) and
(14).
For example, when a = a

one checks the validity of

|a| 1 a |a| 1 (58)
by taking the Gelfand transform of C

(a). The implication

b a b = |a| |b| (59)
then follows, because b a b and (58) for a b yield |b| 1 a |b| 1, so that (a)
[|b|, |b|], hence |a| |b| by (56) and (14). For later use we also record
Lemma 9.4 When a, b A
+
and |a +b| k then |a| k.
By (58) we have a + b k1, hence 0 a k1 b by the linearity of the partial ordering,
which also implies that k1 b k1, as 0 b. Hence, using k1 0 (since k 0) we obtain
k1 a k1, from which the lemma follows by (59).
We now come to the central result in the theory of positivity in C

-algebras, which generalizes

the cases A = B(H) and A = C
0
(X).
Theorem 9.5 One has
A
+
= a
2
[ a

= a (60)
= a

a[ a A. (61)
When (a) R
+
and a = a

then

a A is dened by the continuous functional calculus for
f =

, and satises

a
2
= a. Hence A
+
a
2
[ a

= a. The opposite inclusion follows from

(55) and Corollary 8.2. This proves (60).
The inclusion A
+
a

a[ a A is is trivial from (60).

34
Lemma 9.6 Every selfadjoint element a has a decomposition a = a
+
a

, where a
+
, a

A
+
and a
+
a

= 0. Moreover, |a

| |a|.
Apply the continuous functional calculus with f = id
(a)
= f
+
f

, where id
(a)
(t), f
+
(t) =
maxt, 0, and f

(t) = maxt, 0. Since |f

r(a) = |a| (where we used (56)), the bound

follows from (54) with a a

.
We use this lemma to prove that a

a[ a A A
+
. Apply the lemma to a = b

b (noting that
a is selfadjoint). Then
(a

)
3
= a

(a
+
a

)a

= a

aa

= a

ba

= (ba

ba

.
Since (a

) R
+
because a

is positive, we see from (55) with f(t) = t

3
that (a

)
3
0. Hence
(ba

ba

0.
Lemma 9.7 If c

c A
+
for some c A then c = 0.
We can write c = d +ie, d and e selfadjoint, so that
c

c = 2d
2
+ 2e
2
cc

. (62)
Now for any a, b A one has
(ab) 0 = (ba) 0. (63)
This is because for z ,= 0 the invertibility of ab z implies the invertibility of ba z. Namely, one
computes that (ba z)
1
= b(ab z)
1
a z
1
1. Applying this with c instead of A and c

for
b, we see that the assumption (c

c) R

implies (cc

) R

, hence (cc

) R
+
. By (62),
(60), and Proposition 9.3 (applied to Denition 9.2.2) we see that c

c 0, i.e., (c

c) R
+
, so
that the assumption c

c A
+
now yields (c

c) = 0. Hence c = 0 by Proposition 9.3 applied to

Denition 9.2.3.
The last claim before the lemma therefore implies ba

= 0. As (a

)
3
= (ba

ba

= 0 we see
that (a

)
3
= 0, and nally a

= 0 by the continuous functional calculus with f(t) = t

1/3
. Hence
b

b = a
+
, which lies in a
+
. This completes the proof of Theorem 9.5.
An important consequence of (61) is the fact that inequalities of the type a
1
a
2
for selfadjoint
a
1
, a
2
are stable under conjugation by arbitrary elements b A, so that a
1
a
2
implies b

a
1
b
b

a
2
b. This is because a
1
a
2
is the same as a
2
a
1
0; by (61) there is an a
3
A such that
a
2
a
1
= a

3
a
3
. But clearly (a
3
b)

a
3
b 0, and this is nothing but b

ab b

a
2
b. For example,
replace a in (58) by a

a, and use (94), yielding a

a |a|
2
1. Applying the above principle gives
b

ab |a|
2
b

b (64)
for all a, b A. The denition of a state easily implies that if a b, then (a) (b). Thus (50)
follows.
10 Approximate units
It remains to prove Propositions 4.2 and 4.4, as well as the comments after Denition 2.2. In fact,
these claims are closely related. The appropriate technical tool is the theory of approximate units.
For any noncompact space X, the C

-algebra C
0
(X) has no unit (the unit would be 1
X
, which
does not vanish at innity because it is constant). There is a certain substitute for the absent unit,
however. Taking X = R for simplicity, and pick a sequence of functions 1
n
, n N, that take the
value 1 on [n, n] and vanish for [x[ > n + 1. It is clear that one does not have 1
n
1
R
in the
sup-norm, but instead one has lim
n
|1
n
f f|

= 0 for all f C
0
(R). Such a construction
may be performed in any C

-algebra.
35
Denition 10.1 An approximate unit in a non-unital C

-algebra A indexed by some directed

set (i.e., a set with a partial order and a sense in which ) is a family 1

of selfadjoint
elements of A, such that
|1

| 1, (65)
and
lim

|1

a a| = lim

|a1

a| = 0 (66)
for all a A.
The following fact is very useful, though its proof is awful.
Proposition 10.2 Every non-unital C

-algebra A has an approximate unit. When A is separable,

one may choose countable.
One takes to be the set of all nite subsets of A, partially ordered by inclusion. Hence
is of the form = a
1
, . . . , a
n
, from which we build the element b

:=

i
a

i
a
i
. Clearly
b

is selfadjoint, and according to Theorem 9.5 and Proposition 9.3 one has (b

) R
+
, so that
n
1
1 +b

is invertible in the unitization

A of A. Hence we may form
1

:= b

(n
1
1 +b

)
1
. (67)
Since b

is selfadjoint and b

commutes with functions of itself (such as (n

1
1 + b

)
1
), one has
1

= 1

. Although (n
1
1 +b

)
1
is computed in

A, so that it is of the form c +1 for some c A
and C, one has I

= b

c + b

, which actually lies in A. Using the continuous functional

calculus with f(t) = t/(n+t) on b

, one sees from (55) and the positivity of b

that (1

) [0, 1].
This implies (65) because of (56).
Putting c
i
:= 1

a
i
a
i
, a simple computation shows that

i
c
i
c

i
= n
2
b

(n
1
1 +b

)
2
. (68)
We now apply (54) with a replaced by b

and f(t) = n
2
t(n
1
+ t)
2
. Since f 0 and f
assumes its maximum at t = 1/n, one has sup
tR
+ [f(t)[ = 1/4n. As (b

) R
+
, it follows that
|f|

1/4n, hence |n
2
b

(n
1
1 + b

)
2
| 1/4n by (54), so that |

i
c
i
c

i
| 1/4n by (68).
Lemma 9.4 then shows that |c
i
c

i
| 1/4n for each i = 1, . . . , n. Since any a A sits in some
directed subset of with n , it follows from (94) that
lim

|1

a a|
2
= lim

|(1

a a)

a a| = lim

|c

i
c
i
| = 0.
The other equality in (66) follows analogously.
Finally, when A is separable one may draw all A
i
occurring as elements of from a
countable dense subset, so that is countable.
We are now in a position to prove Propositions 4.2. Let J A be the given ideal, and put
J

:= a

[ a J. Note that j J implies j

j J J

: it lies in J because J is an ideal, hence

a left-ideal, and it lies in J

because J

is an ideal, hence a right-ideal. Since J is an ideal, J J

is a C

-subalgebra of A. Hence by 10.2 it has an approximate unit 1

. Take j J. Using (94)

and 1

= 1

, we estimate
|j

|
2
= |(j 1

j)(j

)| = |(jj

jj

) 1

(jj

jj

)|
|(jj

jj

)| +|1

(jj

jj

)| |(j

j j

j1

)| +|1

| |(jj

jj

)|.
As we have seen, j

j J J

, so that, also using (65), both terms vanish for . Hence

lim

|j

| = 0. But 1

lies in J J

, so certainly 1

J, and since J is an ideal it must

be that j

J for all . Hence j

is a norm-limit of elements in J; since J is closed, it follows

that j

J.
In view of Proposition 4.3, all we need to prove to establish Proposition 4.4 is the property
(94). This uses
36
Lemma 10.3 Let 1

be an approximate unit in J, and let a A. Then

|(a)| = lim

|a a1

|. (69)
It is obvious from (23) that
|a a1

| |(a)|. (70)
To derive the opposite inequality, add a unit 1 to A if necessary, pick any j J, and write
|a a1

| = |(a +j)(1 1

) +j(1

1)| |a +j| |1 1

| +|j1

j|.
Note that
|1 1

| 1 (71)
by Denition 10.1 and the proof of Proposition 9.3. The second term on the right-hand side goes
to zero for , since j J. Hence
lim

|a a1

| |a +j|. (72)
For each > 0 we can choose j J so that (25) holds. For this specic j we combine (70), (72),
and (25) to nd
lim

|a a1

| |(a)| |a a1

|.
Letting 0 proves (69).
We now prove (94) in A/J. Successively using (69), (94) in

A, (71), (69), (24), and (29), we
nd
|(a)|
2
= lim

|a a1

|
2
= lim

|(a a1

(a a1

)|
= lim

|(1 1

)a

a(1 1

)| lim

|1 1

| |a

a(1 1

)| lim

|a

a(1 1

)|
= |(a

a)| = |(a)(a

)| = |(a)(a)

|.
Lemma 11.20 then implies (94). This nishes the proof of Proposition 4.4.
We now state and prove the key result about morphisms.
Theorem 10.4 Let : A B be a nonzero morphism between C

-algebras.
1. is continuous, with norm 1.
2. ker() is an ideal in A.
3. (A) is a C

-subalgebra of B; in particular, (A) is closed in B.

4. If is injective, then it is isometric.
To prove continuity of , we may assume that A and B have units, which are preserved by .
70
If z (a), so that (a z)
1
exists in A, then (a z) is certainly invertible in B, for (4) implies
that ((a z))
1
= ((a z)
1
). Hence (a) ((a)), so that
((a)) (a). (73)
Hence r((a)) r(a), so that || 1 follows from (57). This proves continuity of , which also
immediately implies the second claim of the theorem.
71
70
If A has no unit, we rst replace by the morphism
0
: A (A), where the bar on the right-hand side
denotes the closure in the norm-topology of B. Noting that
0
is nondegenerate by Lemma 6.12, we subsequently
extend
0
to a unital morphism between the unitizations of A and (A), if necessary (cf. the comments after
Denition 6.10).
71
For a dierent proof of continuity, assume that a

= a by using (94) in the standard way, and reduce the

situation to the commutative case by working with
a
: C

(a) (C

(a)). Even though we do not know yet that

is continuous, it follows that the right-hand side is commutative as the closure of a commutative subalgebra of B.
Subsequently, use Theorem 2.1 to reduce to the case of a unital morphism : C(X) C(Y ). By Theorem 6.6,
is the pullback of some : Y X. Lemma 3.2 then immediately yields (73).
37
We now prove the last claim of the theorem. Assume there is an b A for which |(b)| , = |b|.
By (94), (4), and (5) this implies |(b

b)| , = |b

b|. Put a := b

b, noting that a

= a. By (57) and
(14) we must have (a) ,= ((a)). Then (73) implies ((a)) (a). By Urysohns lemma there
is a nonzero f C((a)) which vanishes on ((a)), so that f((a)) = 0. By Lemma 10.5 below
we have (f(a)) = 0. This is a contradiction when is injective, in which case must therefore
be isometric.
Combining the second claim with Proposition 4.4, we see that A/ ker() is a C

-algebra. This
ts into a commutative triangle of C

-algebras:
A

-
A/ ker()
@
@
@
@
@

R
B

?
Here is the canonical projection and is dened by ([a]) = (a) (where [a] is the equivalence
class in a/ ker() of a A). By the theory of vector spaces, is a vector space isomorphism
between A/ ker() and (A). Since and are morphisms between C

-algebras, so is . Since
is injective, it is isometric, as we have just shown. Hence (A/ ker()) has closed range in B.
But (A/ ker()) = (A), so that has closed range in B. Since is a morphism, its image is a

-algebra in B, which by the preceding sentence is closed in the norm of B. Hence (A), inheriting
all operations in B, is a C

-algebra.
Finally, one trivially has || = 1 and || = 1 (because is an isometry). It then follows from
= that || = 1.
We have used the following fact:
Lemma 10.5 When : A B is a morphism and a = a

then
f((a)) = (f(a)) (74)
for all f C((a)) (here f(a) is dened by the continuous continuous functional calculus, and so
is f((a)) in view of (73)).
The property is true for polynomials by (4), since for those f has its naive meaning. For general
f the result then follows by continuity.
Note that Theorem 10.4.2 has a converse: every ideal in a C

-algebra is the kernel of some

morphism. This follows from Theorem 4.2, since J is the kernel of the canonical projection :
A A/J, where A/J is a C

-algebra, and is a morphism by (24), and (29).

11 The C

-algebra of compact operators

It would appear that the appropriate generalization of the C

-algebra M
n
(C) of n n matrices
to innite-dimensional Hilbert spaces H is the C

-algebra B(H) of all bounded operators on H.

This is not the case. Firstly, unlike M
n
(C) (which can be shown to possess only one irreducible
representation up to equivalence), B(H) has a huge number of inequivalent representations; even
when H is separable, most of these are realized on non-separable Hilbert spaces. Secondly, unlike
M
n
(C), B(H) is not simple when H is innite-dimensional (i.e., it has proper ideals). Thirdly,
B(H) is non-separable in the norm-topology even when H is separable.
The appropriate generalization of M
n
(C) to an innite-dimensional Hilbert space H turns out
to be the C

-algebra B
0
(H) of compact operators on H.
72
In noncommutative geometry elements
of B
0
(H) play the role of innitesimals. In K-theory, B
0
(H) is indistinguishable from C. In the
theory of operator algebras, B
0
(H) is a basic building block for C

-algebras.
72
In the literature, one often nds the notation K(H) for B
0
(H), but this is awkward in connection to K-theory.
38
Denition 11.1 Let H be a Hilbert space. A nite-rank operator on H is an operator a B(H)
for which aH := av[ v H is nite-dimensional. Equivalently, a is a nite linear combination
nite-dimensional projections on H.
A compact operator on H is an operator that can be approximated in norm by nite-rank
operators.
73
Lemma 11.2 The collection of all nite-rank operators on H is a

-algebra in B(H). Its norm-
closure B
0
(H) is a C

-algebra, which contains precisely all compact operators on H.

In other words, B
0
(H) is the smallest C

-algebra in B(H) that contains all nite-rank operators.

The norm in B
0
(H) is, of course, the operator norm (83). It is clear that B
0
(H
1
)

= B
0
(H
2
) i
dim(H
1
) = dim(H
2
), and one often writes B
0
or K for the C

-algebra of compact operators on a

separable innite-dimensional Hilbert space H

=
2
.
To prove the lemma, note that by denition p

= p for any projection p, so that B

f
(H) is a

-algebra. Alternatively, it is obvious that if a B

f
(H), then a

B
f
(H). The second claim is
trivial as well.
Proposition 11.3 1. The unit operator 1 lies in B
0
(H) i H is nite-dimensional.
2. The C

-algebra B
0
(H) is an ideal in B(H).
74
Firstly, for any sequence (or net) a
n
B
f
(H) we may choose a unit vector v
n
(a
n
H)

. Then
(a
n
1)v = v, so that |(a
n
1)v| = 1. Hence sup
v=1
|(a
n
1)v| 1, so that |a
n
1| 0
is impossible by denition of the norm (83) in B(H).
Secondly, when a B
f
(H) and b B(H) then ab B
f
(H), since abH aH. But since
ba = (a

, and B
f
(H) is a

-algebra, one has a

B
f
(H) and hence ba B
f
(H). Hence
B
f
(H) is an ideal in B(H), save for the fact that it is not norm-closed (unless H has nite
dimension). Now if a
n
a then a
n
b ab and ba
n
ba by continuity of multiplication in B(H).
Hence B
0
(H) is an ideal by virtue of its denition.
One has a complete characterization of compact operators.
Theorem 11.4 A self-adjoint operator a B(H) is compact i a =

iI
a
i
p
i
(norm-convergent
sum), where I is countable, a
i
R, each projection p
i
is nite-dimensional (so that each nonzero
eigenvalue of a has nite multiplicity), and the set of all eigenvalues of a only has 0 as a possible
accumulation point.
For a proof see, e.g., [15].
Unlike B(H), the state space of B
0
(H) can be explicitly computed. We just state the results,
referring to [15] for details. This involves the study of a number of algebraic ideals of B(H), which
are not closed.
Denition 11.5 The Hilbert-Schmidt norm |a|
2
of a B(H) is dened by
|a|
2
2
:=

i
|ae
i
|
2
, (75)
where e
i

i
is an arbitrary basis of H; the right-hand side is independent of the choice of the basis.
The Hilbert-Schmidt class B
2
(H) consists of all a B(H) for which |a|
2
< .
The trace norm |a|
1
of a B(H) is dened by
|a|
1
:= |(a

a)
1
4
|
2
2
, (76)
where (a

a)
1
4
is dened by the continuous functional calculus. The trace class B
1
(H) consists of
all a B(H) for which |a|
1
< .
73
If a is a compact operator then aB
1
is compact in H (with the norm-topology). Here B
1
is the unit ball in H,
i.e., the set of all v H with v 1.
74
The quotient C(H) := B(H)/B
0
(H) turns out to be a very interesting C

-algebra, known as the Calkin

algebra.
39
Both B
1
(H) and B
2
(H) are complete in the pertinent norms, and are therefore Banach spaces.
Moreover, B
2
(H) is a Hilbert space in the inner product
(a, b) := Tr a

b. (77)
If a B
1
(H) then
Tr a :=

i
(e
i
, ae
i
) (78)
is nite and independent of the basis.
75
Proposition 11.6 One has the inclusions
B
f
(H) B
1
(H) B
2
(H) B
0
(H) B(H), (79)
with equalities i H is nite-dimensional.
This chain of inclusions is sometimes seen as the non-commutative analogue of

c
(X)
1
(X)
2
(X)
0
(X)

(X),
where X is an discrete set, with equalities i X is nite. Since
1
(X) =
0
(X)

and

(X) =

1
(X)

=
0
(X)

, this analogy is strengthened by the following result.

Theorem 11.7 One has B
0
(H)

= B
1
(H) and B
1
(H)

= B
0
(H)

= B(H) under the pairing

(a) = Tr a = a(). (80)
Here B
0
(H)

is identied with B
1
(H), and a B
1
(H)

is identied with a B(H).

Corollary 11.8 Any state on B
0
(H) is of the form a Tr a, where is a density matrix,
that is, an element of B
1
(H) that is positive in B
0
(H) (i.e., 0) and has unit trace (Tr = 1).
Density matrices were introduced by John von Neumann [13] as the most general possible states
in elementary quantum mechanics. Using modern terminology, he dened a state on B(H) to
be normal when (p) =

j
(p
j
) for each pairwise orthogonal family of projections p
j

j
with

j
p
j
= p.
76
He then proved that a normal state on B(H) is necessarily of the form described in
Corollary 11.8. Thus the normal states on B(H) coincide with the states on B
0
(H). In fact, this
statement turns out to be equivalent to the claim B(H) = B
1
(H)

in Theorem 11.7.
75
When a / B
1
(H), it may happen that Tr a depends on the basis; it may even be nite in one basis and innite
in another.
76
Here the sum is dened as the least upper bound of all nite sums of the p
j
, with respect to the ordering of
projections dened by p q when pH qH.
40
Appendix: Basic denitions
All vector spaces will be dened over C, and all functions will be C-valued, unless we explicitly
state otherwise. The abbreviation i means if and only if, which is the same as the symbol .
An equation of the type a := b means that a is by denition equal to b.
Denition 11.9 A norm on a vector space V is a map | | : V R such that
1. |v| 0 for all v V ;
2. |v| = 0 i v = 0;
3. |v| = [[ |v| for all C and v V ;
4. |v +w| |v| +|w| (triangle inequality).
A norm on V denes a metric d on V by d(v, w) := |v w|. A vector space with a norm that is
complete in the associated metric (in the sense that every Cauchy sequence converges) is called a
Banach space. We will denote a generic Banach space by the symbol B.
The two main examples of Banach spaces we will encounter are Hilbert spaces and certain
collections of operators on Hilbert spaces.
Denition 11.10 A pre-inner product on a vector space V is a map ( , ) : V V C such
that
1. (
1
v
1
+
2
v
2
,
1
w
1
+
2
w
2
) =
1

1
(v
1
, w
1
) +
1

2
(v
1
, w
2
) +
2

1
(v
2
, w
1
) +
2

2
(v
2
, w
2
) for
all
1
,
2
,
1
,
2
C and v
1
, v
2
, w
1
, w
2
V ;
2. (v, v) 0 for all v V .
An equivalent set of conditions is
1. (v, w) = (w, v) for all v, w V ;
2. (v,
1
w
1
+
2
w
2
) =
1
(v, w
1
) +
2
(v, w
2
) for all
1
,
2
C and v, w
1
, w
2
V ;
3. (v, v) 0 for all v V .
A pre-inner product for which (v, v) = 0 i v = 0 is called an inner product.
The equivalence between the two denitions of a pre-inner product is elementary; in fact, to
derive the rst axiom of the second characterization from the rst set of conditions, it is enough
to assume that (v, v) R for all v (use this reality with v v +iw). Either way, one derives the
Cauchy-Schwarz inequality
[(v, w)[
2
(v, v)(w, w), (81)
for all v, w V . Note that this inequality is valid even when ( , ) is not an inner product, but
merely a pre-inner product.
It follows from these properties that an inner product on V denes a norm on V by |v| :=
_
(v, v); the triangle inequality is automatic.
Denition 11.11 A Hilbert space is a vector space with inner product that is complete in the
associated norm. We will usually denote Hilbert spaces by the symbol H.
A Hilbert space is completely characterized by its dimension (i.e., by the cardinality of an
arbitrary orthogonal basis). To obtain an interesting theory, one therefore studies operators on a
Hilbert space, rather than the Hilbert space itself. Although a satisfactory mathematical theory
of unbounded operators on a Hilbert space exists, we will restrict ourselves to bounded operators.
We recall this concept in the more general context of arbitrary Banach spaces.
41
Denition 11.12 A bounded operator on a Banach space B is a linear map a : B B for
which
|a| := sup |av| [ v B, |v| = 1 < . (82)
The space of all bounded operators on a Banach space B is called L(B).
The number |a| is the operator norm, or simply the norm, of a. This terminology is justied,
as it follows almost immediately from its denition (and from the properties of the the norm on
B) that the operator norm is indeed a norm.
(It is easily shown that a linear map on a Banach space is continuous i it is a bounded operator,
but we will never use this result. Indeed, in arguments involving continuous operators on a Banach
space one almost always uses boundedness rather than continuity.)
When B is a Hilbert space H the expression (82) becomes
|a| := sup (av, av)
1
2
[ v H, (v, v) = 1. (83)
When a is bounded, it follows that
|av| |a| |v| (84)
for all v B. Conversely, when for a 0 there is a C > 0 such that |av| C|v| for all v,
then a is bounded, with operator norm |a| equal to the smallest possible C for which the above
inequality holds.
Proposition 11.13 The space L(B) of all bounded operators on a Banach space B is itself a
Banach space in the operator norm.
In view of the comments following (83), it only remains to be shown that L(B) is complete in
the operator norm. Let a
n
be a Cauchy sequence in L(B). In other words, for any > 0 there
is a natural number N() such that |a
n
a
m
| < when n, m > N(). For arbitrary v B, the
sequence a
n
v is a Cauchy sequence in B, because
|a
n
v a
m
v| |a
n
a
m
| |v| |v| (85)
for n, m > N(). Since B is complete by assumption, the sequence a
n
v converges to some w B.
Now dene a map a on B by av := w = lim
n
a
n
v. This map is obviously linear. Taking n in
(85), we obtain
|av a
m
v| |v| (86)
for all m > N() and all v B. It now follows from (82) that a a
m
is bounded. Since
a = (a a
m
) + a
m
, and L(B) is a linear space, we infer that a is bounded. Moreover, (86) and
(82) imply that |a a
m
| for all m > N(), so that a
n
converges to a. Since we have just
seen that a L(B), this proves that L(B) is complete.
We dene a functional on a Banach space B as a linear map : B C that is continuous, in
that (v)[ C|v| for some C, and all v B. The smallest such C is the norm of :
|| := sup [(v)[, v B, |v| = 1. (87)
Denition 11.14 The dual B

of B is the space of all functionals on B.

Similarly to the proof of 11.13, one shows that B

is a Banach space. For later use, we quote,

without proof, the fundamental Hahn-Banach theorem of functional anlysis and an important
consequence of it (cf. [15]).
Theorem 11.15 For a functional
0
on a linear subspace B
0
of a Banach space B there exists a
functional on B such that =
0
on B
0
and || = |
0
|. In other words, each functional dened
on a linear subspace of B has an extension to B with the same norm.
42
Corollary 11.16 When (v) = 0 for all B

then v = 0.
Recall that an algebra is a vector space with an associative bilinear operation (multiplication)
: A A A; we usually write ab for a b. We call A unital if there is an element 1 A, the
unit, such that 1a = a1 = a for all a A. It is easy to see that a unit is unique when it exists.
It is clear that L(B) is an algebra with unit under operator multiplication. One easily shows
that |ab| |a| |b|.
Denition 11.17 A Banach algebra is a Banach space A which is at the same time an algebra,
in which for all a, b A one has
|ab| |a| |b| . (88)
It follows that multiplication in a Banach algebra is separately continuous in each variable. In
a Banach algebra one needs a sharper denition of a unit than for general algebras: we require
that |1| = 1.
As we have just seen, for any Banach space B the space L(B) of all bounded operators on B is
a Banach algebra. In what follows, we will restrict ourselves to the case that B is a Hilbert space
H; this leads to the Banach algebra L(H). In the theory of C

-algebras, one usually writes B(H)

for L(H). This algebra has additional structure.
Denition 11.18 An involution on an algebra A is a real-linear map A A

such that for all

a, b A and C one has
a

= a; (89)
(ab)

= b

; (90)
(a)

= a

. (91)
A

-algebra is an algebra with an involution.
The operator adjoint a a

on a Hilbert space, dened by the property

(v, a

w) := (av, w), (92)

denes an involution on B(H). Hence B(H) is a

-algebra. The crucial property |a

a| = |a|
2
for
all a B(H). motivates the following denition.
Denition 11.19 A C

-algebra is a complex Banach space A that is at the same time a

-algebra,
such that for all a, b A one has
|ab| |a| |b|; (93)
|a

a| = |a|
2
. (94)
For C

-algebras, the property |1| = 1 for a unit follows from the axioms, and does not need
to be imposed separately. It is clear that B(H) is a C

-algebra. Moreover, each (operator) norm-

closed

-algebra in B(H) is obviously a C

-algebra.
We note a very useful lemma, whose proof is an easy exercise.
Lemma 11.20 1. If an involution on a Banach algebra A satises |a|
2
|a

a|, then A is a
C

-algebra.
2. For any element a of a C

-algebra one has

|a

| = |a|. (95)
Bibliography
[1] W.B. Arveson and R.G. Douglas (eds.), Operator Theory, Operator Algebras and Applications,
Vols. 1&2 (AMS, Providence, 1990).
[2] B. Blackadar, K-theory for Operator Algebras, 2nd ed. (Cambridge University Press, Cam-
bridge, 1999).
[3] A. Connes, Noncommutative Geometry (Academic Press, San Diego, 1994).
[4] K.R. Davidson, C

-algebras by Example (AMS, Providence, 1996).

[5] J. Dixmier, C

-algebras (North-Holland, Amsterdam, 1977).

[6] R.S. Doran (ed.), C

-algebras: 19431993, A Fifty Year Celebration (AMS, Providence, 1994).

[7] D.E. Evans and Y. Kawahigashi, Quantum Symmetries on Operator Algebras (Oxford Uni-
versity Press, New York, 1998).
[8] J.M. Gracia-Bonda, J.C. Varilly, and H. Figueroa, Elements of Noncommutative Geometry
(Birkh auser, Boston, 2001).
[9] R.V. Kadison (eds.), Operator Algebras and Applications, Vols. 1&2 (AMS, Providence, 1982).
[10] R.V. Kadison and J.R. Ringrose, Fundamentals of the Theory of Operator Algebras, Vol. I:
Elementary Theory (Academic Press, New York, 1983).
[11] R.V. Kadison and J.R. Ringrose, Fundamentals of the Theory of Operator Algebras, Vol. II:
Advanced Theory (Academic Press, New York, 1986).
[12] C. Kassel, Quantum Groups (Springer, New York, 1995).
[13] J. von Neumann, Mathematische Grundlagen der Quantenmechanik (Springer, Heidelberg,
1932).
[14] J. von Neumann, Collected Works. Vol. III: Rings of Operators, ed. A. H. Taub (Pergamon
Press, New York, 1961).
[15] G.K. Pedersen, Analysis Now (Springer, Heidelberg, 1989).
[16] B.V. Rajarama Bhat, G.A. Elliott, and P.A. Fillmore (eds.), Lectures on Operator Theory
(AMS, Providence, 1999).
[17] M. Rrdam, F. Larsen, N. Laustsen, An Introduction to K-theory for C*-algebras (Cambridge
University Press, Cambridge, 2000).
[18] S. Sakai, C

-algebras and W

-algebras (Springer, Heidelberg, 1971).

[19] M. Takesaki, Theory of Operator Algebras, Vol. I (Springer, Heidelberg, 2002).
[20] M. Takesaki, Theory of Operator Algebras, Vols. II&III (Springer, Heidelberg, 2003).
[21] N.E. Wegge-Olsen, K-theory and C

-algebras (Oxford University Press, Oxford, 1993).

43