C∗-ALGEBRAS

J. A. Erdos

Department of Mathematics
King’s College
London WC2R 2LS
ENGLAND

Abstract

These notes form an introductory account of C∗ -algebras. Sec-

tions 1, 2, and 4 provide a straightforward development of the subject
up to the Gelfand-Naimark Theorem, namely the result that every
C∗ -algebra is isometrically isomorphic to a ∗ -subalgebra of the alge-
bra of operators on some Hilbert space. Some results on more general
commutative Banach algebras, whose proofs require little extra ef-
fort, are included. In Section 3 there are accounts of two applications
of the commutative theory: the C∗ -algebra approach to the spectral
theorem for bounded normal operators on Hilbert space and a brief
introduction to the ideas of abstract harmonic analysis.
2 C*-algebras

1 Definitions, the spectrum and other basics.

1.1 Banach algebras and C∗ -algebras

In these notes we shall use the term algebra to mean a linear associative
algebra where, if no field is mentioned, the scalars will be the complex field
C. An algebra A is said to be a normed algebra if it has a norm that makes
it into a normed linear space and the norm also satisfies

(i) ||ab|| ≤ ||a||.||b||

(ii) if A has an identity e then ||e|| = 1 .

If A is a normed algebra and A is a Banach space (that is, A, with its norm,
is complete) then A is called a Banach algebra.
∗
An involution on A is a map : A → A (a 7→ a∗ ) such that

(iii) a∗∗ = a,

(iv) (λa + µb)∗ = λ̄a∗ + µ̄b∗ ,

(v) (ab)∗ = b∗ a∗ .

If A is a Banach algebra with involution and also

(vi) ||aa∗ || = ||a||2 ,

then A is called a C∗ -algebra and (vi) is called the “C∗ condition”.

The term “Banach algbra with involution” is normally reserved for the case
when the involution is continuous; if for all a ∈ A we have ||a∗ || = ||a|| we
say that the involution is isometric.
Some trivial consequences. The following results hold.
1. In a Banach algebra, multiplication A × A → A, ((a, b) 7→ ab) is contin-
uous.
This follows by applying (i) and the triangle inequality for the norm to the
identity

ab − a0 b0 = (a − a0 )(b − b0 ) + (a − a0 )b0 + a0 (b − b0 )
John Erdos 3

to obtain

||ab − a0 b0 || ≤ ||a − a0 ||.||b − b0 || + ||a − a0 ||.||b0 || + ||a0 ||.||b − b0 || .

2. In a C∗ -algebra the involution is isometric.

Using the C∗ condition and (i), we have that for any a ∈ A, ||a||2 = ||aa∗ || ≤
||a||.||a∗ || so that ||a|| ≤ ||a∗ ||. Applying this to a∗ gives the opposite in-
equality, so ||a|| = ||a∗ ||.
3. In a C∗ -algebra

||a|| = sup ||ax|| = sup ||xa|| .

||x||≤1 ||x||≤1

a∗ a∗
Clearly sup||x||≤1 ||ax|| ≤ sup||x||≤1 ||a||.||x|| ≤ ||a|| and when x = ||a∗ ||
= ||a||
,
this supremum is attained.
4. If a C∗ -algebra A has an identity e, then e = e∗ .
In any algebra, the identity is unique. Now for all a ∈ A,

ae∗ = (ea∗ )∗ = (a∗ )∗ = a

and similarly a = e∗ a so e∗ is the identity, showing that e = e∗ . Note also

that the C∗ condition makes (ii) redundant since ||e|| = ||ee∗ || = ||e||2 and
so ||e|| = 1.
5. The C∗ condition (vi) may be replaced by ||aa∗ || ≥ ||a||2 .
This is because this is enough to deduce that ||a|| = ||a∗ || in (2) above, and
so the opposite inequality is a consequence of (i) : ||aa∗ || ≤ ||a||.||a∗ || = ||a||2
In fact, (vi) may be replaced by ||aa∗ || = ||a||.||a∗ || but this is not important
(and the proof is very difficult).
Examples.
1. C(X). Let X be a compact space and let C(X) be the Banach space of all
complex-valued functions on X with the usual norm, ||f || = supx∈X |f (x)|.
Multiplication in C(X) is defined pointwise : f.g(x) = f (x).g(x) and the
involution by complex conjugation f ∗ (x) = f (x). It is easy to see that C(X)
is a commutative C∗ -algebra which has an identity, namely the function e
where e(x) = 1 for all x ∈ X.
2. C0 (X). Now X is a locally compact space and C0 (X) is the Banach space
of all complex-valued functions on X which vanish at infinity. The algebraic
4 C*-algebras

operations and norm are defined as in Example 1. Once again, C0 (X) is a

commutative C∗ -algebra but this time it has no identity.
3. B(H). Let H be a Hilbert space and let B(H) be the algebra of all
continuous linear operators on H. For A ∈ B(H) let A∗ be the usual adjoint.
Then it is clear that B(H) is a C∗ -algebra. Furthermore, any subalgebra A
of B(H) that is closed under adjoints (that is, A∗ ∈ A whenever A ∈ A) and
is closed in the norm sense (hence complete) is an example of a C∗ -algebra.
If X is a Banach space, the algebra B(X ) of all continuous operators on X is
a Banach algebra but if X is not a Hilbert space then B(X ) is not, in general,
a C∗ -algebra.
4. A(∆). Let ∆ be the open unit disc and let A(∆) be the set of functions
¯ A(∆) is a subalge-
that are analytic on ∆ and continuous on its closure, ∆.
¯ Note that A(∆) is not closed under the complex conjugation
bra of C(∆).
involution. However, there is another natural involution that can be defined
on A(∆), namely
f ∗ (z) = f (z̄) .
It is easy to verify that this involution is isometric (that is, ||f || = ||f ∗ ||).
Also, if f g = 0 then one (or both) of f and g must vanish at an infinite
number of points inside ∆ and so, by a well-known theorem of complex func-
tion theory, must be zero. Thus A(∆) is an integral domain. In particular,
f.f ∗ = 0 ⇒ f = 0. However, A(∆) is not a C∗ -algebra. To see this take, for
example, f (z) = eiz . It is trivial to verify that ||f.f ∗ || = 1 and ||f || ≥ e so
that ||f.f ∗ || 6= ||f ||2 .
5. L1 (R), L1 (T), `1 (Z) and more generally L1 (G) for a locally compact abelian
groups. We first deal with the Banach space L1 (R) of integrable functions
(with usual Lebesgue measure) on the real line and the usual norm
Z ∞
||f ||1 = f (t) dt .
−∞

Multiplication of two functions f and g is defined as their convolution :

Z ∞
(f ∗ g)(t) = f (s)g(t − s) ds .
−∞

The change of variable u = t − s in the integration shows that, also,

Z ∞
(f ∗ g)(t) = f (t − u)g(u) du .
−∞

It is a simple exercise, using the theorems of Fubini and Tonelli to show that
||f ∗ g||1 ≤ ||f ||1 .||g||1 ,
John Erdos 5

(see Lemma 3.2 for a proof of a slightly more general result). It follows that
L1 (R) is a commutative Banach algebra. It is not a C∗ -algebra (although one
can define an isometric involution by f ∗ (t) = f (−t)).
L1 (T), the integrable functions on the unit circle is defined similarly. We can
parametrise the unit circle T by {eit : 0 ≤ t < 2π} and identify functions on
T with periodic functions of period 2π on R. The norm and multiplication
are defined as above, except that the limits of integration are 0 and 2π.
P∞
`1 (Z) is the space of sequences {ξn : −∞ < n < ∞} such that −∞ |ξn |
converges. The norm of a sequence x = (ξn ) is given by
∞
X
||x|| = |ξn |
−∞

and the product of x with y = (ηn ) is z = (ζn ) where

∞
X
ζk = ξn ηk−n
n=−∞

All the above classical examples are special cases of the following situation.
A topological group is a group with a topology on it that makes the group
operations continuous. Let G be a topological group which is abelian as
a group and locally compact as a topological space. It is shown in books
on measure theory that on any such group one can find a measure µ that
is invariant under the group operation. That is to say, for any measurable
subset δ of G,
µ(δx) = µ(δ)
for all x ∈ G. Here the group operation is written as multiplication and δx
denotes the set {yx : y ∈ δ}. This measure is unique (up to a multiplicative
constant) and is called Haar measure.
The set L1 (G) of complex-valued functions that are integrable with respect
to Haar measure forms a Banach algebra under the norm
Z
||f || = |f (t)| dt .
G

and multiplication
Z
(f ∗ g)(t) = f (s)g(s−1 t) ds .
G

It is easy to verify that all the above examples fit into this framework. It is
these examples that form the subject of abstract harmonic analysis which is
treated in more detail in Section 3.
6 C*-algebras

Looking ahead. The main aim of these notes is to prove that all C∗ -
algebras are of the form of the examples above. Commutative C∗ -algebras
are like Examples 1 or 2 (depending on whether there is an identity) and all
C∗ -algebras are of the form of Example 3.
Much of the terminology of C∗ -algebras is inherited from operator theory
on Hilbert space. Thus, an element a of a C∗ -algebra A is said to be self-
adjoint if a∗ = a, and normal if aa∗ = a∗ a. If A has an identity e then
u ∈ A is called unitary if uu∗ = u∗ u = e.

1.2 The spectrum

Let A be a Banach algebra with identity e (algebras without identity will be

discussed later). The spectrum of an element a of A is defined to be the set
of all complex numbers λ such that λe − a has no inverse in A. The spectrum
of a is denoted by σ(a). Thus

σ(a) = {λ : λe − a is singular } .

It may happen that an element a can be considered as a member of more

than one algebra. If it is necessary to indicate the algebra, we write σA (a)
for the spectrum of a as a member of the algebra A.
The complement C \σ(a) of σ(a) is called the resolvent set of a and denoted
by ρ(a).
The above definitions are identical to those for operators. These in turn are
generalisations of the eigenvalues of matrix theory.
To study the spectrum, it is necessary to discuss when elements have inverses.
The simple but important Lemma 1.1 below uses a formal geometric series.
For the proof, we need to recall the concept of absolute convergence for
P
series in a Banach space. To say that xn is absolutely convergent means
P
that ||xn || is convergent as a series of real numbers. For an absolutely
P P
convergent series xn , write sn = nr=1 xr . Then
° °
° n+p ° n+p
° X ° X
°
||sn+p − sn || = ° °
xr ° ≤ ||xr || .
°r=n+1 ° r=n+1

P
Since ||xn || is convergent and hence a Cauchy series, we have that (sn ) is
a Cauchy sequence thus converges. This shows that absolutely convergent
series in a Banach space are convergent.
John Erdos 7

Lemma 1.1 If ||x|| < 1 then (e − x)−1 exists.

P∞ n
Proof. Since ||xn || ≤ ||x||n , and ||x|| < 1, it is clear that 0 x is absolutely
convergent. Denote its sum by y. Now note the identity

e−xn+1 = (e−x)(e+x+x2 +x3 +· · ·+xn ) = (e+x+x2 +x3 +· · ·+xn )(e−x) .

Let n → ∞. Then xn+1 → 0 and so

e = (e − x)y = y(e − x)

showing that y is the inverse of e − x.

The above lemma is often stated in the obviously equivalent form : if ||e −
a|| < 1 then a−1 exists.

Theorem 1.2 Let G be the set of invertible elements of a Banach algebra

A. Then

(i) G is open,

(ii) the map x 7→ x−1 (G → G) is continuous.

Proof. (i) We need to show that if a ∈ G then, for x of sufficiently small

norm, a + x ∈ G. Write a + x = a(e + a−1 x). From 1.1, if ||a−1 x|| < 1 then
(e + a−1 x) is invertible and hence so is a + x. Thus, if ||x|| < 1/||a−1 || then
a + x ∈ G and so the open ball centre a radius ||a−1 ||−1 is a subset of G.
(ii) Given a ∈ A and any ² > 0, write

||a−1 − b−1 || = ||b−1 (b − a)a−1 ||

= ||(b−1 − a−1 )(b − a)a−1 + a−1 (b − a)a−1 ||
≤ ||b−1 − a−1 ||.||b − a||.||a−1 || + ||b − a||.||a−1 ||2

So ³ ´³ ´
||a−1 − b−1 || 1 − ||b − a||.||a−1 || ≤ ||a−1 ||2 .||b − a|| .
First taking ||b − a|| sufficiently small to make 1 − ||b − a||.||a−1 || > 12 , it is
clear that ||a−1 − b−1 || may be made arbitrarily small by taking ||b − a|| small
enough.

Corollary 1.3 The closure of a proper ideal is proper. Hence maximal ideals
are closed.
8 C*-algebras

Proof. Continuity of the algebraic operations shows that the closure of an

ideal is an ideal. But a proper ideal I contains no invertible elements and
so is disjoint from some open ball centre the identity e. The closure Ī is
therefore also disjoint from this ball and so it is a proper ideal.

Theorem 1.4 The spectrum σ(a) of any element a of A is a non-empty

compact subset of C.

Proof. We first show that σ(a) is bounded. In fact, if |λ| > ||a|| then
||λ−1 a|| < 1 and so, by 1.1 λe − a = λ(e − λ−1 a) has an inverse. Hence σ(a)
is contained in the ball centre 0, radius ||a||.
To prove that σ(a) is closed, we show that ρ(a) is open. If λ ∈ ρ(a) then
λe − a ∈ G where G, the set of invertible elements, is open (1.2). Hence, for
some ² > 0, any element b with ||b − (λe − a)|| < ² is in G. Thus if |µ − λ| < ²
then µe − a ∈ G and so µ ∈ ρ(a), showing that ρ(a) is open. (More briefly
one can say that ρ(a) is open since it is the pre-image of the open set G
under the continuous map λ 7→ λe − a.)
For the proof that σ(a) 6= ∅, first note that, if λ, µ ∈ ρ(a) then

(λe−a)−1 −(µe−a)−1 = [(µe−a)(µe−a)−1 ](λe−a)−1 −[(λe−a)(λe−a)−1 ](µe−a)−1

and so, since all the elements commute,

(λe − a)−1 − (µe − a)−1 = (µ − λ)(λe − a)−1 (µe − a)−1

(the above relation is called the resolvent equation).

Now take any φ ∈ A0 (the Banach space dual of A) and let f (λ) = φ[(λe −
a)−1 ]. Then f is holomorphic on ρ(a), since

f (µ) − f (λ)
f 0 (λ) = lim
µ→λ µ−λ
( )
(µe − a)−1 − (λe − a)−1
= lim φ
µ→λ µ−λ
n o
= lim φ −(λe − a)−1 (µe − a)−1
µ→λ

= −φ(λe − a)−2 .

Note that above (and again below), we use the continuity of inversion,
(Lemma 1.2 (ii)) several times and also the continuity of φ .
John Erdos 9

Now, if σ(a) were empty, f would be entire. Also,

"µ ¶−1 #
1 a
lim |f (λ)| = lim φ e− = 0.
|λ|→∞ |λ|→∞ |λ| λ

The above shows that f is bounded and entire, so by Liouville’s Theorem

it is constant and its limit as |λ| → ∞ shows f = 0. But, by the Hahn-
Banach Theorem we can choose φ so that φ(−a−1 ) = f (0) 6= 0, giving us a
contradiction. Thus f cannot be entire and so σ(a) is not empty.

Note that it is essential for the above proof that the field of scalars of our
Banach algebra is the complex field.
10 C*-algebras

Theorem 1.5 (Gelfand-Mazur) Every complex Banach division algebra

is isometrically isomorphic to C. This isomorphism is unique.

Proof. In a division algebra A, the only singular element is 0. If a ∈ A, there

exists an element λ of σ(a) and λe − a, being singular, equals 0. Therefore
A consists of complex multiples of the identity. Clearly

||a|| = ||λe|| = |λ|

and the map a 7→ λ where a = λe is an isometric isomorphism.

Theorem 1.6 (Spectral Mapping Theorem - elementary form) Let a

be an element of a Banach algebra with identity.

(i) If p is any polynomial, then

σ (p(a)) = {p(λ) : λ ∈ σ(a)}

(ii) if 0 ∈
/ σ(a) ½ ¾
1
σ(a−1 ) = : λ ∈ σ(a) .
λ

Proof. (i) For any complex number λ,

p(λ)e − p(a) = (λe − a)P (a) = P (a)(λe − a)

where P (a) is some polynomial in a. Hence, if p(λ)e − p(a) has an inverse,

so has (λe − a). Thus, λ ∈ σ(a) ⇒ p(λ) ∈ p(σ(a)).
Conversely, if µ ∈ σ(p(a)) then µe − p(a) has no inverse. Consider the
polynomial q(t) = µ−p(t). This factorises as q(t) = k(λ1 −t)(λ2 −t) · · · (λn −t)
where k is a constant and λ1 , λ2 , · · · λn are the zeros of q. Then

µe − p(a) = k(λ1 e − a)(λ2 e − a) · · · (λn e − a)

and if all the factors (λi e − a) had inverses, so would µe − p(a) (since the
factors commute). Therefore, at least one zero λi of q is in the spectrum of
a and p(λi ) = µ; that is, µ ∈ {p(λ) : λ ∈ σ(a)}.
(ii) This is obvious from the relation
µ ¶
1
λe − a = −aλ e − a−1 .
λ
John Erdos 11

Theorem 1.7 In a C∗ -algebra,

(i) σ(a) = σ(a∗ ) ,

(ii) if u is unitary then |λ| = 1 for all λ ∈ σ(u),

(iii) if a is self-adjoint then σ(a) is real.

Proof. (i) Since e∗ = e, we have xy = yx = e if and only if y ∗ x∗ = x∗ y ∗ = e.

Thus (λe − a) is invertible if and only if (λe − a)∗ = (λ̄e − a∗ ) is invertible
showing that σ(a) = σ(a∗ ).
(ii) If u is unitary then ||u||2 = ||uu∗ || = ||e|| = 1. Now if λ ∈ σ(u) then from
1.4, |λ| ≤ ||u|| = 1. But u∗ is also unitary and Theorem 1.6 (ii) shows that
λ−1 ∈ σ(u∗ ) so that |λ−1 | ≤ 1. Thus |λ| = 1.
(iii) Suppose that a = a∗ . We need to show that if λ = α + iβ and the
imaginary part β is non-zero, then λe − a has an inverse in A. First a
reduction: note that
" #
αe − a
(α + iβ)e − a = β + ie
β

and x = αe−a β
is self-adjoint. Hence it is sufficient to show that x + ie is
invertible for every self-adjoint x.
Suppose this is not the case so that −i ∈ σ(x). Then for any ξ,

ξ + 1 ∈ {ξ + it : t ∈ σ(x)} = σ(ξ + ix)

and so, from 1.4, |ξ + 1| ≤ ||ξ + ix||. Taking ξ to be real the C∗ condition
shows that

(ξ + 1)2 ≤ ||ξ + ix||2 = ||(ξ + ix)∗ (ξ + ix)|| = ||ξ 2 + x2 || ≤ ξ 2 + ||x||2 .

This implies that 1 + 2ξ ≤ ||x||2 for all real ξ, which is impossible. Thus
−i ∈
/ σ(x) and so the spectrum of any self-adjoint element is real.
The following is a simple result on the spectrum of a product.

Lemma 1.8 Let a, b ∈ A. Then for λ 6= 0, λ ∈ σ(ab) ⇔ λ ∈ σ(ba).

Proof. If λ 6= 0 and λ ∈ ρ(ab), it can be verified by direct multiplication

that
λ−1 e + λ−1 b(λe − ab)−1 a.
12 C*-algebras

is the inverse of λe − ba. Thus λ ∈ ρ(ba).

The spectral radius ν(a) of an element a ∈ A is defined as

ν(a) = max{|λ| : λ ∈ σ(a)} .

It is clear from the proof of Theorem 1.4 that ν(a) ≤ ||a||. The next theorem
establishes a formula for ν(a). It is reminiscent of the formula
1 1
= lim |an | n
r n→∞
P∞
for the radius of convergence of the scalar power series n=0 an z n .

Theorem 1.9 (Spectral radius formula)

1
ν(a) = lim ||an || n
n→∞

Proof. Consider the function λφ[(λe − a)−1 ] where φ ∈ A0 . It follows

as in the proof of Theorem 1.4 that f is holomorphic on ρ(a) and so on
{λ : |λ| > ν(a)}. Make the substitution z = 1/λ and define
· h ¸ i
1 1
g(z) = φ ( e − a)−1 = φ (e − za)−1 .
z z
1
Then g is holomorphic in {z : |z| < ν(a)
}.
1
We find the Taylor expansion of g. For |z| < ||a||
we have from Lemma 1.1
that ∞ X
(e − za)−1 = z n an .
n=0

Since φ is continuous,
³ ´ ∞
X
g(z) = φ (e − za)−1 = z n φ(an ) .
n=0

Initially we only know that this power series expansion of g is valid for |z| <
1 1
||a||
. However, since g is holomorphic on the larger disc {z : |z| < ν(a) } and the
Taylor expansion of a holomorphic function is unique, the above expansion is
1 1
valid for |z| < ν(a) . Thus, for |z| < ν(a) then limn→∞ z n φ(an ) = 0. Actually,
we only need to know that this sequence is bounded, that is, for each φ in
the dual of A, there exists a constant Kφ such that

|φ(z n an )| < Kφ for all n .

John Erdos 13

It now follows from the Uniform Boundedness Theorem that there exists a K
1
independent of φ such that, for all n and all z with |z| < ν(a) , ||z n an || < K;
that is,
1
n n1 Kn
||a || < .
|z|
1 1
It follows that ||an || n ≤ K n ν(a) and so
1
sup||an || n ≤ ν(a) .

In the other direction, if λ ∈ σ(a) then, for all positive integers n, λn ∈

1
σ(an ) (1.6(i)) and so, (1.4) |λ|n ≤ ||an ||. That is, |λ| ≤ ||an || n . Taking the
1
maximum over λ ∈ σ(a), we have that ν(a) ≤ ||an || n for all n. Therefore,
combining all that we have shown,
1 1 1
inf ||an || n ≥ lim||an || n ≥ ν(a) ≥ sup||an || n
n

so the last three terms are equal.

Note that the existence of the limit is not part of the hypotheses of the
theorem and is established by the above proof

Corollary 1.10 If A is a C∗ -algebra and a is normal then

||a|| = ν(a) .

Proof. It x = x∗ then by the C∗ condition, ||x2 || = ||x||2 and by repeating

m m
this we have that ||x2 || = ||x||2 . Thus
m 1
||x|| = lim ||x2 || 2m = ν(x) .
m→∞

For a normal a, using the above for aa∗ ,

||a||2 = ||aa∗ || = ν(aa∗ )

1
= lim ||(aa∗ )n || n
n→∞
1 1
≤ lim ||an || n .||(a∗ )n || n
n→∞
= (ν(a))2 ≤ ||a||2

and so equality holds throughout.

14 C*-algebras

We now consider the spectrum with respect to subalgebras. Let A be a

Banach algebra. For a subalgebra B of A to be a Banach algebra in its own
right, it must be complete and, since A is complete, this is equivalent to
requiring it to be closed. The boundary of a set S will be denoted by ∂S,
that is, ∂S = S ∩ S c .

Theorem 1.11 Let B be a closed subalgebra of A containing the identity of

A. Then
∂σB (b) ⊆ σA (b) ⊆ σB (b) .

Proof. If b ∈ B and λe − b has no inverse in A then it can have no inverse

in B. This shows that σA (b) ⊆ σB (b).
Now let λ ∈ ∂σB (b) = σB (b) ∩ ρB (b) = σB (b) ∩ ρB (b). Then there exists
a sequence (λn ) of elements of ρB (b) with (λn ) → λ. If λ ∈ / σA (b) then
(λe − b)−1 exists in A and, using the continuity of inversion (1.2(ii)), the
sequence ((λn e − b)−1 ) converges to (λe − b)−1 in A. But then ((λn e − b)−1 )
is a Cauchy sequence of elements of B and so converges in B to the same
limit, (λe − b)−1 , contradicting that λ ∈ σB (b).

Corollary 1.12 For b ∈ B,

νA (b) = νB (b)

Proof. This is obvious either from the above result or from Theorem 1.9.

Corollary 1.13 If b ∈ B, and σB (b) has empty interior, then σB (b) = σA (b)

Proof. If σB (b) has empty interior, then ∂σB (b) = σB (b).

If A and B are algebras, a homomorphism π : A → B is a map which pre-
serves all the algebraic structure. If A and B are ∗-algebras and π preserves
the involution then it is called a ∗-homomorphism. If A has an identity e
then it is easy to see that π(e) acts as the identity on the range of π (in
fact, if B is a normed algebra, on the closure of the range of π). Thus there
is essentially no loss of generality to assume that, where identities exist, all
homomorphisms preserve identities - and we shall make this assumption.

Theorem 1.14 If A and B are C∗ -algebras with identities and π : A 7→ B is

a ∗-homomorphism then ||π(a)||B ≤ ||a||A . In particular, π is automatically
continuous.
John Erdos 15

Proof. If x has an inverse in A then π(x−1 ) is the inverse of π(x) in B.

Applying this to λe − a shows that, ρA (a) ⊆ ρB (π(a)) for all a ∈ A. Hence
νB (π(a)) ≤ νA (a) ≤ ||a|| . (1)

But in the C∗ -algebra B,

||π(a)||2 = ||π(a)(π(a))∗ || = ||π(a)π(a∗ )|| = ||π(aa∗ )|| .
Since π(aa∗ ) is selfadjoint in B, this, together with 1.10 shows that

||π(a)||2 = νB (π(aa∗ )) .
Replacing a by aa∗ in (1) and using the above we have
||π(a)||2 = νB (π(aa∗ )) ≤ ||aa∗ || = ||a||2 .

An examination of the above proof shows that the C∗ condition was used
only for the range algebra B and the theorem is valid with a weaker condition
on A, namely that it be a Banach algebra with isometric involution. The
following consequence is clear for algebras with identities.

Corollary 1.15 If π is a ∗-isomorphism then it is isometric. In particu-

lar, there is at most one norm that makes a Banach algebra with isometric
involution into a C∗ -algebra.

The result above is also true for algebras without identities. It can be proved
by using the procedure given in 1.3 below to embedd each algebra into an
algebra with identity and apply Corollary 1.15. The detailsd are omitted.

Theorem 1.16 If B is a closed ∗subalgebra of a C∗ -algebra A with the same

identity then for all b ∈ B, σA (b) = σB (b).

Proof. Since the spectrum of a self-adjoint element of a C∗ -algebra is real

(1.7), Corollary 1.13 shows that self-adjoint elements have the same spectra
in A and B. Now, if λ ∈ ρA (b), both λe − b and its adjoint have inverses
in A. Thus 0 is not in σA ((λe − b)(λe − b)∗ ) and since (λe − b)(λe − b)∗ is
self-adjoint, it therefore has an inverse in B. Writing x for this inverse, we
have that
e = (λe − b)(λe − b)∗ x
showing that (λe − b) has a right inverse in B. Similarly, (λe − b) has a left
inverse in B and a simple algebra argument shows that it is then invertible.
Thus σB (b) ⊆ σA (b) for all b ∈ B and since the opposite inclusion is obvious,
equality holds.
16 C*-algebras

1.3 Algebras without identity

It is a fact that, given an algebra A without identity, it is possible to construct

an algebra A1 with an identity e such that A is a subalgebra of A1 and A1 is
the span of A and e. In fact the construction makes A into an ideal of A1 . If
A is a C∗ -algebra then one can ensure that A1 is also a C∗ -algebra. We now
give a sketch of how this can be done.
From the algebraic point of view, the construction is very easy. A1 is simply
defined as A ⊕ C, the direct sum of A and C with the usual linear space
structure. Multiplication is defined in A1 by

(a, λ).(b, µ) = (ab + λb + µa, λµ)

(this definition is very natural if you think of (a, λ) as a plus λ times a formal
identity). Of course, A is identified with the set {(a, 0) : a ∈ A}. Note that
A is an ideal of A1 . The element (0, 1) = e is the identity of the algebra A1 .
There are many ways to define a norm on A1 to make it into a Banach
algebra. For example
k(a, λ)k = kak + |λ|
is complete and it is easy to prove that it satisfies the multiplicative norm
condition (i) for a Banach algebra. However, this does not make A1 into a
C∗ -algebra.
To make A1 into a C∗ -algebra, define a norm |||.||| as follows :

|||(a, λ)||| = sup kax + λxk.

kxk≤1

where x varies over A. Note that multiplication by (a, λ) acts as linear oper-
ator on the ideal A of A1 and |||(a, λ)||| is simply the norm of this operator.
From this it follows easily that |||.||| satisifies the multiplicative condition for
the norm of a Banach algebra.
Recall (Section 1.1, consequence 3) that kak = supkxk≤1 kaxk showing that
|||(a, 0)||| = kak, so that the embedding of A into A1 is isometric. To verify
the C∗ condition,

|||(a, λ)|||2 = sup kax + λxk2

kxk≤1

= sup k(x∗ a∗ + λ̄x∗ )(ax + λx)k

kxk≤1
= sup |||(x∗ , 0)(a, λ)∗ (a, λ)(x, 0)|||
kxk≤1
John Erdos 17

≤ sup |||(x∗ , 0)|||.|||(a, λ)∗ (a, λ)|||.|||(x, 0)|||

kxk≤1
= |||(a, λ)∗ (a, λ)||| .

Hence (Section 1.1, consequence 5) |||.||| satisfies the C∗ condition.

It remains to prove completeness. For this we use the fact that A has no
identity.Let ((an , λn )) be a Cauchy sequence in A1 . We first show that (λn )
is a bounded sequence. Suppose this is not the case. Then by passing to a
subsequence, we may assume that |λn | → ∞. Then
° ° µ ¶
° am an °° am an
° − ° = sup k − xk
°
λm λn kxk≤1 λm λn
°µ ¶ µ ¶°
° am an °
= sup ° ° x+x − x + x °°
kxk≤1 λm λn
≤ |λm |−1 |||(am , λm )||| + |λn |−1 |||(an , λn )||| .

Since (an , λn ) is Cauchy in A1 it is bounded and since |λn | → ∞ it follows

that ( λann ) is Cauchy in A and hence converges to some element f of A. But,
for all x ∈ A, ° °
° an ° 1
° x + x° = |||(an , λn )||| → 0
° °
λn |λn |
and so f x + x = 0. Thus A has an identity (namely −f ), contrary to hypoth-
esis. Therefore (λn ) is a bounded sequence and, passing to a subsequence,
we may assume that it is convergent to some number λ.
Now, using

kan − am k = |||(an , λ) − (am , λ)|||

= |||(an , λn ) + (0, λ − λn ) − (am , λm ) − (0, λ − λm )|||
≤ |||(an , λn ) − (am , λm )||| + |λ − λn | + |λ − λm |

it follows that (an ) is a Cauchy sequence in A and so has a subsequence

convergent to some element a ∈ A. It is now a subsequence of (an , λn )
converges to (a, λ) ∈ A1 .
Most of the proofs in the remainder of these notes will be given for algebras
with identities. The above construction will then be used to deduce results
for algebras that need not have identities.
18 C*-algebras

2 Commutative theory.
In this section we shall show that a commutative Banach algebra with iden-
tity can be mapped into the algebra of all continuous functions on some
compact space. In the case of a C∗ -algebra the embedding is an isometric
*-isomorphism.
Before embarking on the proofs, we need some remarks on quotients. Recall
that if X is a linear space and Y is a subspace of X then we define an
equivalence relation ∼ on X by

x1 ∼ x2 ⇐⇒ x1 − x2 ∈ Y .

Denote the equivalence class of x by [x] (the notation x + Y is also used).

The quotient space X /Y is defined as the set of equivalence classes, with
addition and multiplication by scalars defined in the natural way. If X is a
Banach space and Y is a closed subspace then

k[x]k = inf kx + yk
y∈Y

makes X /Y into a Banach space. All the above is standard Banach space
theory.
If A is an algebra and I ⊆ A then it can be verified that

[a].[b] = [a.b]

is a well-defined multiplication if and only if I is an ideal. In the case of

a Banach algebra A and a closed ideal I, this construction makes A/I a
Banach algebra. Furthermore, if A is a C∗ -algebra then it canbe shown that
A/I is also a C∗ -algebra (but this fact will not be needed in these notes).
From now on in this section, except where otherwise stated, A will denote
a commutative Banach algebra with identity e. Obviously any ideal I is
two-sided and any quotient algebra A/I is commutative.

Theorem 2.1 For a maximal ideal M of A, then A/M is isometrically iso-

morphic to C.

Proof. From Corollary 1.3, M is closed. Also, A/M has no proper ideals
since if I were such then Z = ∪{a : [a] ∈ I} would be a proper ideal of A
containing M .
John Erdos 19

Now if [a] 6= 0 (i.e. a ∈

/ M ) then (A/M )[a] the principal ideal generated
by [a] is non-zero and so equals A/M . Thus [e] ∈ (A/M )[a] and so there
exists [x] ∈ A/M such that [x].[a] is the identity of A/M . This shows that
A/M is a division algebra and so, by the Gelfand-Mazur Theorem (1.5) it is
isometrically isomorphic to C.

Corollary 2.2 There is a bijection between the set M of proper maximal

ideals and the set Φ of all non-zero homomorphisms of A into C.

Proof. Given M ∈ M, consider the composition of the natural map A →

A/M and the homomorphism of the Gelfand-Mazur Theorem A/M → C.
This is a non-zero homomorphism φM of A into C with kernel φ−1
M (0) = M .

In the other direction, given φ ∈ Φ we show that the corresponding maximal

ideal is its kernel φ−1 (0). Let φ, ψ ∈ Φ with φ−1 (0) ⊆ ψ −1 (0) then for all
a ∈ A,
a − φ(a)e ∈ φ−1 (0) ⊆ ψ −1 (0)
so that 0 = ψ(a − φ(a)e) = ψ(a) − φ(a), since ψ(e) = 1; that is, φ = ψ. This
shows firstly that for φ ∈ Φ, φ−1 (0) is maximal; for if φ−1 (0) were a proper
subset of an M ∈ M, the homomorphism ψ = φM would have a strictly
larger kernel. Secondly, the above shows in particular that, if φ and ψ have
the same kernel the they are equal. Hence the map φ 7→ φ−1 (0) is a bijection
Φ → M.
We call Φ (or M) the carrier space of A. Other terms often used are : the
maximal ideal space or the spectrum of A.
We now proceed to a construction similar to the embedding of a Banach
space into its bidual. For each φ ∈ Φ and a ∈ A, we have a complex number
φ(a). This is really a map from Φ × A into C. So far φ has been thought
of as a fixed function and a as the variable. However, this can be reversed.
Define, for each a ∈ A the function â : Φ → C given by

â(φ) = φ(a) .

The function â is called the Gelfand transform of a. We now establish its

properties.

Theorem 2.3 The map a 7→ â is an algebraic homomorphism of A into the

set of complex-valued functions on Φ.
20 C*-algebras

Proof. This is a routine verification using the fact that each φ is a homo-
morphism. For example, to show that ab c = â.b̂, simply note that

c
(â.b̂)(φ) = φ(a).φ(b) = φ(ab) = ab(φ) .

Theorem 2.4

(i) ê is the unit function (i.e. ê(φ) = 1 for all φ).

T T
(ii) â = 0 if and only if a ∈ {φ−1 (0) : φ ∈ Φ} = {M : M ∈ M} .

(iii) σ(a) = {â(φ) : φ ∈ Φ}

(iv) ν(a) = max {|â(φ)| : φ ∈ Φ}

(v) If φ1 6= φ2 then for some a ∈ A, â(φ1 ) 6= â(φ2 ) .

Proof. (i) For each φ ∈ Φ, φ(e) = 1, that is ê(φ) = 1.

(ii) For a ∈ A,
â = 0 ⇔ â(φ) = 0, ∀φ ∈ Φ ⇔ φ(a) = 0, ∀φ ∈ Φ ⇔ a ∈ φ−1 (0), ∀φ ∈
Φ.
(iii) For any φ ∈ Φ, φ(φ(a)e − a) = 0. Hence φ(a)e − a is in the kernel of φ
which is a proper ideal. Therefore it has no inverse, that is φ(a) ∈ σ(a).
Conversely, if λ ∈ σ(a) then λe − a has no inverse. Therefore the principal
ideal A(λe − a) is proper and so is a subset of some maximal ideal M , that
is, it is a subset of the kernel of some φ ∈ Φ. In particular φ(λe − a) = 0 and
so λ = φ(a).
(iv) This is obvious from (iii)
(v) If φ1 6= φ2 then the functions must differ at some element a ∈ A. Then
φ1 (a) 6= φ2 (a), that is, â(φ1 ) 6= â(φ2 ).
Note that the function â (where a ∈ A) in the algebra B(Φ) of bounded
functions on Φ, has norm

kâk = sup |â(φ)| .

φ∈Φ

Hence the above Theorem shows that

kâk = ν(a) ≤ kak

John Erdos 21

so the Gelfand transform is continuous A → B(Φ). The Gelfand transform

is one to one (i.e. a monomorphism) if A has the property that ν(a) = 0 ⇒
a = 0; it is isometric if kak = ν(a) for all a ∈ A, and this is so in the case
when A is a commutative C∗ -algebra (Theorem 1.10).
An element of a Banach algebra is called quasinilpotent if the only element
in its spectrum is 0. Thus (iii) above shows that the Gelfand transform of
a ∈ A is the zero function if and only if a is quasinilpotent.
We mention the Jacobson radical for general interest, although we shall
not be making any use of the concept in the rest of these notes. The Jacobson
radical R of any Banach algebra A with identity is defined by

R = ∩{all maximal left ideals of A } = ∩{all maximal right ideals of A }

(and it is a fact that the two intersections are equal). When A is commutative
then,
R = ∩{all maximal two-sided ideals of A } .
Now,

â = 0 ⇔ â(φ) = 0 ∀φ ∈ Φ ⇔ a ∈ M ∀M ∈ M ⇔ a ∈ R.

Thus the kernel of the Gelfand map is the Jacobson radical.

We now proceed to define a topology on the carrier space. If Φ is the carrier
space of the commutative Banach algebra A, the Gelfand topology on Φ is
the weakest topology (that is, the one with fewest open sets) to make each
function â : Φ → C continuous.
A base for the neighbourhoods of φ in the Gelfand topology is given by all
sets of the form

{ψ ∈ Φ : |φ(ai ) − ψ(ai )| < ², i = 1, 2, · · · , n}

that is
{ψ ∈ Φ : |âi (φ) − âi (ψ)| < ², i = 1, 2, · · · , n}
as the quantities a1 , a2 , · · · , an in A and ² > 0 range over all choices.
Convergence of sequences in this topology is given by

φn → φ ⇔ â(φn ) → â(φ) in C ∀a ∈ A
⇔ φn (a) → φ(a) ∀a ∈ A
22 C*-algebras

that is, convergence of sequences in the Gelfand topology is pointwise con-

vergence.
From here on, when we refer to the “carrier space” of a commutative banach
algebra, we shall understand the space Φ of homomorphisms with the Gelfand
topology . Note that the Gelfand topology is Hausdorff, since if φ1 6= φ2 then
there exists a ∈ A such that φ1 (a) 6= φ2 (a) so, if δ = 12 |φ1 (a) − φ2 (a)| then

N1 = {φ : |φ1 (a) − φ(a)| < δ} and N2 = {φ : |φ2 (a) − φ(a)| < δ}

are neighbourhoods of φ1 and φ2 respectively and are disjoint.

The lemma below uses Tychonoff’s theorem which states that the product
of compact spaces with the product topology is compact. Recall that if
{Xγ : γ ∈ Γ} is a family of non-empty sets then the cartesian product,
Q
γ∈Γ Xγ has, as a typical member, the function γ 7→ xγ where xγ ∈ Xγ . The
proof below exploits this by identifying each member of the carrier space
of the algebra A with an element of a suitable product space. It is the
same technique as that usually used to prove the Banach-Alaoglu Theorem
on the weak∗ compactness of the unit ball of the dual of a Banach space.
An alternative (shorter) proof could be given deducing the result from the
Banach-Alaoglu Theorem.

Lemma 2.5 The carrier space of a commutative Banach algebra with iden-
tity is compact.

Proof. Let Φ be the carrier space of A and let φ ∈ Φ. From Theorem 2.4 (iii),
φ(a) ∈ σ(a) and so |φ(a)| ≤ ν(a) ≤ kak. Thus the closed disc Za in C, centre
0 radius kak is compact and contains φ(a) for all a ∈ A. Let
Y
Z= Za .
a∈A

Then Z with the product topology is compact by Tychonoff’s Theorem.

Recall that the product topology on Z is the weakest topology to make
the projection maps pa corresponding to each element of the index set A
continuous. Note that the elements of Z are all functions f : A → C such
that |f (a)| ≤ kak. Then pa (f ) = f (a). Now the elements of Φ are particular
cases of these functions (i.e. elements of Z) and pa (φ) = φ(a) = â(φ). Thus
the topology on Φ inherited from the product topology of Z is the same as
the Gelfand topology. It is therefore sufficent to show that Φ is a closed
subset of Z.
John Erdos 23

The elements of Φ are those elements of Z that also satisfy the algebraic
properties of homomorphisms. Thus if φ ∈ Φ then for all a, b ∈ A,

φ(a + b) = φ(a) + φ(b) .

Define the function La,b : Z → C by

La,b (f ) = f (a + b) − f (a) − f (b) = pa,b (f ) − pa (f ) − pb (f ) .

Then, directly from the definition of the product topology it is clear that La,b
is continuous and hence its kernel L−1a,b (0) is closed. The algebraic condition
on φ given above is that φ is in the kernel of each La,b . Similarly we define,
for a, b ∈ A and λ ∈ C,

Ma,b (f ) = f (ab) − f (a).f (b)

Sa,λ (f ) = f (λa) − λf (a)
I(f ) = f (e) − 1

and all these functions are continuous. It is easy to see that an element of Z
is in Φ exactly when it is in the intersection of the kernels of all the functions
La,b , Ma,b , Sa,λ , I as a, b, λ take all possible values. Therefore it is a closed
subset of Z and so it is compact.
The theory developed in the results above, when put together, constitute a
proof of the following important theorem.

Theorem 2.6 (Gelfand map) Given any commutative Banach algebra A

with identity, there is a homomorphism of A into C(Φ), the algebra of all
continuous complex-valued functions on a compact Hausdorff space Φ. The
kernel of this map is the set of all quasinilpotent elements of A.

The carrier space appears to be a rather abstract topological space. However,

in some important cases it can be realised as a very concrete space. The
following lemma gives a simple and useful example of this situation.

Lemma 2.7 If A is generated by the identity e and one element a then the
carrier space Φ of A is homeomorphic to σ(a).

Proof. We show that the required homeomorphism is the function â : Φ →

σ(a). Indeed, by the definition of the Gelfand topology, â is continuous and
24 C*-algebras

from Theorem 2.4 (iii) it is onto σ(a). To show that â is injective we use our
hypothesis that the elements {p(a)}, as p ranges over all polynomials, form
a dense subset of A. Then
â(φ1 ) = â(φ2 ) ⇒ φ1 (a) = φ2 (a)
⇒ φ1 (p(a)) = φ2 (p(a)) ∀ polynomials, since φ1 , φ2 are homomorphisms
⇒ φ1 (x) = φ2 (x) ∀x ∈ A, since φ1 , φ2 are continuous
⇒ φ1 = φ2

showing that â is injective. So â is a continuous bijection. Now apply the gen-

eral theorem that a continuous bijection between compact Hausdorff spaces
is a homeomorphism.
If X and Y are topological spaces and h : X ↔ Y is a homeomorphism
between them, the mapping f ↔ f˜ given by
f˜(h(x)) = f (x) or, equivalently f˜(y) = f (h−1 (y))
is clearly an isomorphism between the function spaces C(X) and C(Y ). This
observation, together with Lemma 2.7 shows that, for the case of a singly
generated algebra, we can compose this isomorphism with the Gelfand map
to get a map of A into the space C(σ) where σ ⊆ C is the spectrum of
the generator. (Loosely speaking, we make C(σ) the range of the Gelfand
map.) Moreover (with the notation of Lemma 2.7) using the function â as
the homeomorphism, the function fx on C corresponding to an element x ∈ A
is given by
fx (s) = x̂(â−1 (s)) .
For the case of x = a this reduces to fa (s) = s. This shows that when we use
Lemma 2.7 and the Gelfand transform to map the singly generated algebra
A into C(σ), we obtain that the image of the generator a is the identity
function f (s) = s.
So far in this section our algebra A was a commutative Banach algebra. For
the special case of a C∗ -algebra A, a much stronger result holds. Here the
algebra is not merely mapped onto some subalgebra of C(Φ) but essentially
it “is” the whole of C(Φ).

Theorem 2.8 (The commutative Gelfand-Naimark Theorem) Given

any commutative C∗ -algebra A with identity, there is an isometric ∗-isomorphism
of A onto C(Φ), the algebra of all continuous complex-valued functions on a
compact Hausdorff space Φ.
John Erdos 25

Proof. Let Φ be the carrier space of A and let a 7→ â be the Gelfand map
as above. Clearly every element of a commutative C∗ -algebra is normal and
so from Corollary 1.10 the norm of each element of A is equal to its spectral
radius. Thus kak = kâk.
Now A is a Banach space and thus complete. Therefore, since the Gelfand
map is isometric, {â : a ∈ A} is a complete subset of C(Φ) and so closed.
We shall use the Stone-Weierstrass Theorem to show that it is the whole of
C(Φ).
We show that the Gelfand map preserves the involution (where the involution
in C(Φ) is defined in the natural way : f ∗ (φ) = f (φ). For a ∈ A we write
a = x + iy where x = 12 (a + a∗ ), y = 2i1 (a − a∗ ) are selfadjoint. From
Theorem 1.7 (iii), x and y have real spectra and so, since ˆ(φ) ∈ σ(x), the
functions x̂ and ŷ are real-valued. Therefore

(â)∗ = x̂ + iŷ = x̂ − iŷ = x d

− iy = ac∗ .

This shows that the image of  is a *-subalgebra and also that the Gelfand
map preserves the adjoint operation. The remaining conditions needed to
apply the Stone-Weierstrass Theorem have already been proved, namely
that  separates the points of Φ and contains the unit function (Theo-
rem 2.4 (i), (v)). Therefore  = C(Φ).

Lemma 2.9 If A is a commutative C∗ -algebra generated by the identity e and

the elements a and a∗ of A, then the carrier space Φ of A is homeomorphic
to σ(a).

Proof. We again show that the required homeomorphism is the function

â : Φ → σ(a). The only difference between this lemma and Lemma 2.7 is
that the dense subset of A is {p(a, a∗ )} as p ranges over polynomials in 2
variables. So, to prove that â is an injection we note that

â(φ1 ) = â(φ2 ) ⇒ ac∗ (φ1 ) = (â(φ1 ))∗ = â(φ1 ) = â(φ2 ) = ac∗ (φ2 )

and from this it follows that

φ1 (p(a, a∗ )) = φ2 (p(a, a∗ ))

for all polynomials in two variables. Thus, by the hypothesis, φ1 (x) = φ2 (x)
for all x ∈ A so that φ1 = φ2 . The remainder of the proof is exactly the same
as that of Lemma 2.7.
26 C*-algebras

The observations following Lemma 2.7 also apply to Lemma 2.9 above. In
view of Theorem 2.8 the conclusion here is that A is isometrically isomorphic
to C(σ). Once again, the image of the generator a is the identity function
f (s) = s

Theorem 2.10 (The Gelfand-Naimark Theorem - algebras with no identity)

Given any commutative C∗ -algebra A, there is an isometric ∗-isomorphism
of A onto C0 (Φ), the algebra of all continuous complex-valued functions van-
ishing at infinity on a locally compact Hausdorff space Φ.

Proof. As shown in Section 1, A can be embedded in an algebra A1 which

is A with an identity adjoined. The above theory can then be applied to A1 .
From the construction of adjoining an identity it is easy to see that A (strictly
speaking {(a, 0) : a ∈ A}) is a maximal ideal of A1 . Therefore, if Φ1 is the
carrier space of A1 , by Corollary 2.2 there is a unique element of Φ1 with
kernel A. Call this element φ∞ and put

Φ = Φ1 \ {φ∞ }

Since the identity of A1 is (0, 1), it follows that φ∞ ((a, λ)) = λ. Also, ev-
ery element of Φ is (by restriction) a non-zero homomorphism of A into C.
Conversely, if ψ is a non-zero homomorphism of A into C, then φ : A1 → C
defined by φ((a, λ)) = ψ(a) + λ is easily shown to be an element of Φ1 . Thus
Φ is, as before, the space of all non-zero homomorphisms of A into C and is
called the carrier space of A. Since Φ1 is compact, the topology induced on
Φ is locally compact (i.e. every point has a compact neighbourhood).
From Theorem 2.8, the Gelfand map (a, λ) 7→ (a, dλ) is an isometric iso-
morphism of A1 onto C(Φ1 ). The composition of this with the embedding
A → (A, 0) takes A onto the algebra of all functions of C(Φ1 ) that vanish at
φ∞ . Restricting the functions to Φ shows that A is mapped onto an algebra
 of continuous functions on Φ.
A function f ∈ C(Φ1 ) satisfies f (φ∞ ) = 0 if and only if, given any ² > 0,
there exists an open neighbourhood N of φ∞ such that |f (φ)| < ² for φ ∈ N .
Since the complement of N is compact, this can be rephrased in terms of
elements of Φ alone as follows : given any ² > 0, there exists a compact set
K such that |f (φ)| < ² for all φ ∈/ K. This is exactly the definition of a
“function vanishing at infinity” on a locally compact space.
John Erdos 27

3 Applications of the Commutative Theory

3.1 The Spectral Theorem for Normal Operators

Let H be a Hilbert space and let B(H) be the C∗ -algebra of all bounded
linear operators on H. It T ∈ B(H) is a normal operator then T T ∗ = T ∗ T
and so the algebra A generated by T, T ∗ and I (the identity operator) is a
commutative C∗ -subalgebra of B(H).
Let σ be the spectrum of T in B(H). Then Theorem 1.16 shows that
σA (T ) = σ .
It follows from Lemma 2.9 that the carrier space of A is homeomorphic to
σ. Using the remarks following Lemma 2.7 and Lemma 2.9 we see that
the Gelfand map composed with this homeomorphism gives an isometric *-
isomorphism of A onto C(σ). Moreover, the homeomorphism can be chosen
so that T is mapped onto the identity function; i.e. onto the function fT
where fT (t) = t. If A ∈ A, we denote the image of A under this map by fA
and if f ∈ C(σ), we denote the operator mapped onto f by Af .
The spectral theorem for T is now easily derived using two standard the-
orems. Each of these is usually referred to as the “Riesz representation
theorem”. The first is the description of the dual of C(σ) as the set of Borel
measures on σ and the second is the correspondence between sesqui-linear
functionals on H × H and bounded operators on H.
First we note that, for any x, y ∈ H, the map φ : C(σ) → C defined by
φ(f ) = hAf x, yi
is a continuous linear functional. Thus there exists a Borel measure µx,y on
σ such that, Z
hAf x, yi = f (t) dµx,y .
σ
We shall frequently use the fact that if two measures give the same integral
of continuous functions then they are the same. If f ∈ C(σ), and |f (t)| ≤ 1
then kAf k = kf k ≤ 1 and so
¯Z ¯
¯ ¯
¯ f (t) dµx,y ¯ = |hAf x, yi| ≤ kxk.kyk
¯ ¯
σ

which, by a standard argument using regularity, shows that for all Borel sets
δ,
|µx,y (δ)| ≤ kxk.kyk.
28 C*-algebras

Also, it is easy to verify that for a fixed Borel set δ, µx,y (δ) is linear in x and
conjugate linear in y. For example, if f ∈ C(σ)
Z
f dµx+x0 ,y = hAf (x + x0 ), yi = hAf x, yi + hAf x0 , yi
Z
= f d(µx,y + µx0 ,y )

showing that µx+x0 ,y (δ) = µx,y (δ) + µx0 ,y (δ) for any Borel set δ.
The next step is to extend the definition of Af to all functions f in the set
B(σ) of bounded Borel measurable functions on σ. To do this, note that for
f ∈ B(σ), Z
φ(x, y) = f dµx,y
σ

is a bounded sesquilinear form on H × H. The second “Riesz representation

theorem” states that every such form equals hAx, yi for some bounded linear
operator A. Thus we extend the definition of Af by
Z
hAf x, yi = φ(x, y) = f dµx,y .
σ

We now show that the extended map f 7→ Af preserves the algebraic struc-
ture, that is, it is a *-homomorphism.
Since for f ∈ C(σ) we have that A∗f = Af¯, it follows easily that
Z Z Z
f dµx,y = hAf x, yi = hA∗f y, xi = f¯ dµy,x = f dµy,x .
σ σ σ

This shows that integrating with respect to µx,y is the same as integrating
with respect to µy,x . The relation therefore holds for all f ∈ B(σ). Thus
µx,y = µy,x and consequently A∗f = Af¯ for all f ∈ B(σ).
The proof that f 7→ Af preserves multiplication requires two stages. If
f, g ∈ C(σ), Af g = Af .Ag and so, for all x, y
Z Z
f g dµx,y = f dµAg x,y .
σ σ

So “g dµx,y = dµAg x,y ” and the above relation holds for all f ∈ B(σ). Thus
for all f ∈ B(σ) and g ∈ C(σ),

Af g = Af .Ag .
John Erdos 29

Also
Af g = (Af¯ḡ )∗ = (Af¯Aḡ )∗ (Aḡ )∗ (Af¯)∗ = Ag Af .
But then
Z Z
gf dµx,y = hAgf x, yi = hAg Af x, yi = g dµAf x,y
σ σ

and the equality of the integrals for all g ∈ B(σ) follows as before. Thus for
all f, g ∈ B(σ) we have Af Ag = Af g = Ag Af .
The isometry of the Gelfand map is not extended to f ∈ B(σ), since f may
be large on sets of measure 0. However, it is easy to verify that kAf k ≤ kf k
for f ∈ B(σ).
We now look at the operators Af for the cases when f is the characteristic
function χδ of some Borel set δ. We write E(δ) = Aχδ . Clearly each E(δ)
commutes with T and E(σ) = I. We show that E(·) is an instance of what is
called a “spectral measure”. Since (χδ )2 = χδ and χδ is real, it is immediate
that E(δ)2 = E(δ) and E(δ)∗ = E(δ) so that each E(δ) is an orthogonal
projection on H. It is easy to see that the following properties hold :

1. E(α ∩ β) = E(α).E(β)

2. E(α ∪ β) = E(α) + E(β) − E(α).E(β)

To demonstrate the appropriate countable additivity of E(·), let {δi } be a

countable family of disjoint Borel sets whose union is δ. Then, since for all
P
x ∈ H, µx,x (.) is a measure, µx,x (δ) = ∞
1 µx,x (δi ) and so, since E(δi ).E(δj ) =
0 when i 6= j, a simple calculation shows that
n
X n
X n
X
kE(δ)x − E(δi )xk2 = kE(δ)xk2 − k E(δi )xk2 = µx,x (δ) − µx,x (δi ) .
i=1 i=1 i=1
P P
Thus, ∞ i=1 E(δi )x converges in norm to E(δ)x. Thus
∞
i=1 E(δi ) converges
to E(δ) in the strong operator topology and we express this property by
saying that

3. E(·) is strongly countably additive.

We extend the definition of E(δ) to all Borel subsets of C by E(δ) = E(δ ∩σ).
Clearly all the properties are preserved. A map from the Borel subsets of
C with values in the orthogonal projections on Hilbert space H satisfying
30 C*-algebras

conditions 1 – 3 above is called a spectral measure. In the case of the

spectral measure we have derived from the bounded normal operator T , we
have that E(σ) = I and E(·) has support σ. (The support of a Borel measure
is the complement of the union of all open sets of measure zero. Thus it is
always closed – but in this case it is also compact.)
The fact that T = Af where f is the identity function shows that, with the
new notation, Z
hT x, yi = λ hE(dλ)x, yi
σ
for all x, y ∈ H. This is a weak form of the Spectral Theorem.
R
To obtain a stronger version, we need to discuss what is meant by σ f (λ)E(dλ).
P
For a simple function f = ni=1 αi χδi , one defines

Z n
X
f (λ)E(dλ) = αi E(δi ) .
σ i=1
R
Clearly for a simple function f , k σ f (λ)E(dλ)k ≤ kf k. Now for every
f ∈ B(σ), there is a sequence (fk ) of simple functions
R
converging uniformly
to f on σ. It follows easily that the sequence ( σ fk (λ)E(dλ)) converges in
the norm of B(H) to an operator which is dependent only on f and not the
choice of the sequence (fk ). Define
Z Z
f (λ)E(dλ) = lim fk (λ)E(dλ) .
σ k→∞ σ

R
We clearly would like to have that σ f (λ)E(dλ) = Af and this is indeed the
case. To prove it we first note that the equality is true for simple functions.
Now if (fk ) is a sequence of simple functions converging uniformly to f , since
kAf − Afk k ≤ kf − fk k it is clear that Af equals the integral. Looking at the
special case when f is the identity function, we obtain the spectral theorem.
Spectral Theorem for Normal Operators. For a normal operator on
a Hilbert space H, there exists a spectral measure E(·) with support the
spectrum of T such that Z
T = λ E(dλ) .

We shall not go further with this topic in any detail in these notes. However
we should mention a few developments.
R
1. It follows that defining f (T ) = σ f (λ)E(dλ) gives a functional calculus
for bounded Borel functions of a normal operator. In fact one could
John Erdos 31

deal with functions that are essentially bounded (in the appropriate
sense) but we will not go into these technicalities.

2. It is easy to show that, when T is selfadjoint, each E(δ) commutes

with any operator that commutes with T . This fact is the foundation
stone of the theory of von Neumann algebras, since it shows that every
strongly closed *-subalgebra of B(H) is generated by the selfadjoint
projections it contains.

3. One can prove that, for any Borel set δ, the spectrum of T E(δ), re-
garded as an operator on the range of E(δ), is δ. The converse to
this is true and shows the uniqueness of the spectral measure of an
operator: namely if E(·) is a spectral measure commuting with T such
that σ(T E(δ)) ⊆ δ for all Borel subsets δ of C, the E(·) is the spectral
measure of T .

3.2 An Interlude on Harmonic Analysis

The purpose of this section is to connect the Gelfand theory and Fourier
analysis on groups. The most familiar and the most important groups for
Fourier analysis are the real line R, euclidean n-space Rn , the integers Z (all
with addition as the operation) and the circle group T of complex numbers
of modulus 1 under multiplication (equivalently (0, 2π] with addition modulo
2π). We shall refer to these groups as “our specific groups”.
The appropriate setting for elementary abstract harmonic analysis is to con-
sider functions on a locally compact abelian group G (the more advanced
portions of the subject deal with more general groups). We have seen that
such a group G has a Haar measure and that L1 (G), with convolution as
multiplication, is a Banach algebra. We shall introduce the statements of
the results in this generality. However, when the proofs are much easier (or
even trivial) in the case of our specific groups mentioned above, we shall omit
the general proofs.
For the whole of this section, G will denote a locally compact abelian group
and all integrations will be with respect to Haar measure over the whole
group. Recall (Section 1, Example 5) that L1 (G) is a Banach algebra with
multiplication Z
(f ∗ g)(t) = f (s)g(s−1 t) ds .
G

Some preliminary results

32 C*-algebras

Theorem 3.1 L1 (G) is a commutative algebra.

Proof. For the general case one needs the fact that given any subset δ of G,
then δ and δ −1 = {x−1 : x ∈ δ} have the same Haar measure. For our specific
groups this fact is clear and the rest of the proof is a trivial verification using
a change of variable in the integration.

Lemma 3.2 If f ∈ L1 (G) and g ∈ Lp (G), (1 ≤ p < ∞) then (f ∗ g)(y) exist

for almost all y and f ∗ g ∈ Lp (G) with

kf ∗ gkp ≤ kf k1 kgkp .

Proof. The proof uses the theorems of Fubini and Tonelli on double inte-
p
gration. Let q = p−1 be conjugate to p and let h be an arbitrary element
of L (G). Then f (x)g(x−1 y)h(y) is measurable on G × G. So, using the
q

invariance of Haar measure, we have that,

Z Z Z ½Z ¾ 1 ½Z ¾1
p q
−1 −1 p q
|f (x)g(x y)h(y)| dy dx ≤ |f (x)| |g(x y)| dy |h(y)| dy dx
= kf k1 kgkp khkq

Tonelli’s theorem now implies that f (x)g(x−1 y)h(y) is integrable over G × G

and Fubini’s theorem shows that its integral is
Z Z Z
h(y) f (x)g(x−1 y) dx dy = (f ∗ g)(y)h(y) dy .
R
Essentially the same calculation as above now shows that h 7→ h(y).(f ∗ g)(y) dy
is in the dual of Lq with norm at most kf k1 kgkp and so f ∗ g ∈ Lp with

kf ∗ gkp ≤ kf k1 kgkp .

The above result shows, in particular, that if Tf : L2 (G) → L2 (G) is defined

by Tf g = f ∗ g the map f 7→ Tf is continuous from L1 (G) to B(L2 (G)). Easy
calculations show that this map is an algebraic homomorphism and that, if
f ∗ is defined by f ∗ (x) = f (x−1 ), then Tf∗ = Tf ∗ .
We shall need the fact that Tf 6= 0 whenever f 6= 0. For general groups,
the standard proof of this involves the concept of an “approximate identity”
John Erdos 33

but we shall not formally introduce this. We indicate proofs for our specific
groups.
In `1 (Z), the sequence e = (²n ) where ²0 = 1 and ²n = 0 for n 6= 0 acts as an
identity (and also e ∈ `2 (Z)). Thus if x = (ξn ) 6= 0, Tx e = x 6= 0.
The proof for L1 (R) is similar to the general case and involves some tech-
nicalities. We first deal with a continuous function f of compact support.
If f 6= 0, choose ² > 0 such that ² < kf k1 . Clearly it is enough to find a
function g ∈ L2 (R) ∩ L1 (R) such that kf ∗ g − f k1 < ². Since f vanishes
outside a compact set, there exists k > 0 such that f (x) = 0 for |x| > k.
Also an elementary result shows that f is uniformly continous so there is a
²
δ > 0 with δ < 1 such that for |s| < δ and all t, |f (t − s) − f (t)| < 2(k+1) .
The condition δ < 1 is imposed so that if |s| < δ, |f (t − s) − f (t)| vanishes
for |t| > k + 1. Therefore, for |s| < δ,
Z
|f (t − s) − f (t)| dt < ² .

Define g by (
1
if |s| < δ
g(t) = 2δ
0 otherwise.
R R
Then g(s) ds = 1 so f (t) = f (t)g(s) ds and
Z ¯Z ¯
¯ ¯
kf ∗ g − f k1 = ¯ [f (t − s) − f (t)]g(s) ds¯ dt
¯ ¯
Z Z
≤ |f (t − s) − f (t)| dt g(s) ds
< ².
For an arbitrary function f ∈ L1 (R), we use the fact that the continuous
functions of compact support are dense in L1 (R). It then follows by a routine
approximation argument that, for f 6= 0, the operator Tf is non-zero.
The proof for L1 (T) (and indeed for a general locally compact abelian group
G) follows along the same lines. Instead of the interval [−δ, δ] one needs a
suitable compact neighbourhood of 1. The details are omitted.

Abstract harmonic analysis

Lemma 3.3 The Gelfand map on L1 (G) is injective.

Proof. In view of Theorem 2.4.(iv), it is sufficient to show that if f is a

non-zero element of L1 (G), then ν(f ) 6= 0
34 C*-algebras

But if f 6= 0, then the operator Tf : L2 (G) → L2 (G) given by Tf g = f ∗ g,

is non-zero and kTf k ≤ kf k. Since Tf∗ = Tf ∗ , we have that Tf is normal and
so, using Lemma 2.7,
1
n n
0 6= kTf k = lim
n
kT f k
n times
z }| { 1
≤ lim kf ∗ f ∗ · · · ∗ f k n = ν(f ).
n
This proves the lemma.
Note that another way of stating the above lemma is : L1 (G) is semi-simple.
A continuous homomorphism of G into the group T of complex numbers of
modulus 1 is called a character of G. Denote the set of characters of G by
Ĝ. If ξ, η ∈ G, it is easy to verify that the function ξη : G 7→ C defined by,
(ξη)(x) = ξ(x).η(x)
is also a character and that with this operation, Ĝ is an abelian group. We
shall establish a correspondence between Ĝ and the carrier space Φ of L1 (G).
For any function f on G, we define the function fy (where y ∈ G) by fy (x) =
f (yx). It is an easy exercise to show that if f ∈ L1 (G) the map y 7→ fy is
continuous (G → L1 (G)). (First consider the case when f is a continuous
function of compact support and then use the fact that such functions are
dense in L1 (G).)

Theorem 3.4 Given any character ξ of G, the function φξ : L1 (G) → C,

given by Z
φξ (f ) = f (x)ξ(x) dx
G
is a non-zero homomorphism. The map ξ 7→ φξ is a bijection between the
group Ĝ of characters and the carrier space Φ of L1 (G).

Proof. Since characters are continuous and bounded, there is no problem

concerning integrability in the definition of φξ . It is clear that φξ is linear
on L1 (G). If f, g ∈ L1 (G), using the invariance of Haar measure, Fubini’s
theorem and the fact that ξ(xy) = ξ(x)ξ(y),
Z Z
φξ (f ∗ g) = f (y)g(y −1 x)ξ(x) dy dx
Z Z
= f (y)g(x)ξ(yx) dy dx
Z Z
= f (y)ξ(y) dy g(x)ξ(x) dx
= φξ (f )φξ (g)
John Erdos 35

and so φξ is a homomorphism.
Since ξ is a non-zero function, clearly φξ 6= 0. Similarly, it is clear that the
map ξ 7→ φξ is an injection.
To show that ξ 7→ φξ is a surjection, we use the fact that the dual of L1 is
L∞ . Suppose that φ ∈ Φ. Then φ is a continuous linear functional on L1 (G)
and so, for some α ∈ L∞ (G),
Z
φ(f ) = f (x)α(x) dx .

Choose g ∈ L1 (G) such that φ(g) 6= 0. Then,

φ(f )φ(g) = φ(f ∗ g)

Z
= (f ∗ g)(x)α(x) dx
Z Z
= f (y)g(y −1 x)α(x) dy dx
Z
= f (y)φ(gy−1 ) dy .

Thus Z
φ(gy−1 )
φ(f ) = f (y) dy.
φ(g)

It therefore remains to prove that if ξ is defined by

φ(gy−1 )
ξ(y) = ,
φ(g)

then ξ is a character. Continuity of ξ is clear. Also, an easy calculation

shows that for x, y ∈ G,

gx−1 ∗ gy−1 = g(xy)−1 ∗ g

and so φ(gx−1 ).φ(gy−1 ) = φ(g(xy)−1 ).φ(g) . Therefore,

ξ(x).ξ(y) = ξ(xy) . (2)

R
Since φ(f ) = f (y)ξ(y) dy and |φ(f )| ≤ ν(f ) ≤ kf k, we have that |ξ(x)| ≤ 1
except on a set of measure zero. Continuity of ξ shows that |ξ(x)| ≤ 1 for
all x ∈ G. But also, from (1), |ξ(x−1 )| = |ξ(x)|−1 ≤ 1 and so |ξ(x)| = 1 all
x ∈ G. Thus ξ is a character of G.
36 C*-algebras

φ(g −1 )
Note that it is implicit in the above proof that φ(g) y
is independent of of
the choice of g (provided φ(g) 6= 0). This can also be seen from the relation
f ∗ gy = fy ∗ g. Note also that ξ(y) = ξ(y −1 ) = φ(g
φ(g)
y)
.

The group Ĝ inherits a topology from the Gelfand topology on Φ via the
bijection of the above theorem. That is, a set δ is open in Ĝ when {φξ : ξ ∈ δ}
is open in Φ. It is a fact that this makes Ĝ into a topoplogical group (i.e.
with this topology, the group operations are continuous). We shall verify this
for our specific groups but the general proof will not be given in these notes.
We call the group Ĝ with this topology the dual group of the group G.
For f ∈ L1 (G), the Gelfand transform of f is a function with domain Φ, the
carrier space of L1 (G). Since Φ and Ĝ are homeomorphic, we may therefore
regard the transform of f as a function on Ĝ. Thus

fˆ(ξ) = φξ (f )

where the map ξ 7→ φξ (Ĝ → Φ) is as given in the Theorem above. Therefore,

fˆ may be written as Z
ˆ
f (ξ) = f (x)ξ(x) dx .
G

The function fˆ (regarded as a function on Ĝ) is called the Fourier trans-

form of f . Since the Fourier transform is essentially the same as the Gelfand
transform, the following properties follow immediately from the general the-
ory.
For f, g ∈ L1 (G),

∗ g = fˆ.ĝ
1. fd

2. fˆ is continuous and vanishes at infinity,

3. kfˆk∞ ≤ kf k1 .

We now examine our specific groups and will show that the above indeed
gives the classical Fourier transform. The results quoted should be familiar
in the case of the classical theory.
The additive group R of real numbers.
We first identify the characters. If ξ is a character, |ξ(x)| = 1 for all x ∈ R,
ξ(0) = 1 and ξ is continuous. Therefore there exists a δ such that | arg ξ(x)| <
π/2 for |x| < δ.
John Erdos 37

For any x, ξ(x) = ξ( x2 + x2 ) = (ξ( x2 ))2 and so,

arg ξ( x2 ) = 12 arg ξ(x) or arg ξ( x2 ) = 12 arg ξ(x) + π .

But |x| < δ implies that | x2 | < δ and so | arg ξ( x2 )| < π

2
and the second
possibility is excluded for such x.
Choose one particular number t 6= 0 with |t| < δ and suppose that ξ(t) = eiθ .
Then from above, we see that
θ
ξ( 2tn ) = ei 2n

for every positive integer n. Also, for every integer m

³ ´m mθ
ξ( mt
2n
) = ξ( 2tn ) = ei 2n
P mk
and so for all dyadic rationals α, (i.e. finite sums of the form )
2nk
ξ(αt) = eiαθ .

Since dyadic rationals are dense in R, using the continuity of ξ we have that
the above hold for all real numbers α. Now let y = θt . The for any x ∈ R

ξ(x) = ξ( xt .t) = eixy .

Hence every character is of the form

ξ(x) = eixy

for some y ∈ R.
Conversely it is easy to see that distinct y ∈ R give distinct characters. Also
it is clear that multiplication of characters corresponds to addition of the
numbers that give rise to them. Thus, as far as algebraic properties are
concerned, the dual group of R is R.
We now verify that the topology induced on R by the Gelfand topology
coincides with the usual topology. Note that a base for the neighbourhoods
of 0, (that is, of the character ξ0 corresponding to the map x 7→ eixy when
y = 0, hence ξ0 (x) ≡ 1) in the Gelfand-induced topology is given by
½ ¯Z ¯ ¾
¯ ixy
¯
N (ξ0 , f1 , f2 , · · · , fn ; ²) = ¯ ¯
y : ¯ fr (x)(1 − e ) dx¯ < ², 1 ≤ r ≤ n
R
for all choices of f1 , f2 , · · · , fn and ². Since the map y →
7 f (x)(1 − eixy ) dx
is continuous, it is clear that every neighbourhood of this type contains an
38 C*-algebras

open interval about 0. Conversely, choosing f1 = χ[−1,1] , an explicit calcu-

lation shows that N (ξ0 , f1 ; ²) = {y : | siny y | > 1 − 2² } which, for suitable ² is
contained in any open interval containing 0. Thus the neighbourhoods of 0
are the same in both topologies. An easy translation argument shows that
all neighbourhoods are the same in the two topologies. Thus the dual of R is
R (strictly one should say that the dual of R is algebraically isomorphic and
homeomorphic to R).
When dealing with Fourier transforms, Haar measure on R is usually taken
as √12π times Lebesgue measure. Then, from above, we have that the Fourier
transform fˆ of a function f ∈ L1 (R) is

1 Z∞
fˆ(y) = √ f (x) e−ixy dy .
2π −∞

The properties 1 – 3 above are then familiar results of Fourier theory, Prop-
erty 2 being the Riemann-Lebesgue lemma.
The additive group of Rn .
Note that each subspace of Rn is a subgroup. If ξ is a character of Rn , it
induces a character on each one-dimensional subspace. Let {er : 1 ≤ r ≤ n}
be the usual basis of Rn . Then x 7→ ξ(xer ) is a character of R and so from
above we have that for some yr ,

ξ(xer ) = eixyr .

Thus, if x = (x1 , x2 , · · · , xn ),
³X ´
ξ(x) = ξ xr er = eix.y

where y ∈ Rn and x.y is the usual inner product. It follows easily that the
dual of Rn is Rn and the Gelfand transform of f ∈ L1 (Rn ) is therefore
à !n Z
1
fˆ(y) = √ f (x)e−ix.y dx ,
2π n
R

that is, the Fourier transform.

The circle group T.
It follows just as for R, that every character of ξ of T is of the form

ξ(x) = eixy .
John Erdos 39

However, since x and x + 2π represent the same member of T, we must have

eixy = ei(x+2π)y . Therefore y is an integer. Thus the dual of T is the additive
group Z of integers. The Gelfand transform of f ∈ L1 (T) is
Z 2π
fˆ(n) = f (x) e−inx dx .
0

The doubly infinite sequence {fˆ(n)} is usually called the sequence of Fourier
coefficients of f .
The group Z of integers.
Once again, the characters are of the form ξ(n) = einy but now any two
numbers differing by 2π give the same character. Thus the dual of Z is T.
(The proof that the topology of T arising from the Gelfand topology is the
usual topology is an easy verification).
Haar measure on Z gives unit mass to every point and so L1 (Z) is the space
of absolutely convergent doubly infinite sequences. The Fourier transform of
a member (an )−∞<n<∞ of L1 (Z) is the function f ∈ C(T) given by,
∞
X
f (y) = an einy
−∞

and this is just a Fourier series.

40 C*-algebras

4 Representations - the Gelfand-Naimark The-

orem

4.1 The GNS construction

Until further notice, we shall be considering a C∗ -algebra A with identity

e. We begin with a lemma which is the first of many occasions when the
commutative Gelfand-Naimark theorem is used to prove a fact about general
C∗ -algebras.

Lemma 4.1 Let x be a selfadjoint element of A such that σ(x) ⊆ R+ . Then

there exists a unique y ∈ A such that y is selfadjoint, σ(y) ⊆ R+ and y 2 = x.

Proof. Let C be the commutative C∗ -subalgebra of A generated by e and

x. Then C is isometrically isomorphic, via the Gelfand map, to C(S), the
continuous functions on a compact space S. Since the range of x̂ is σ(x) ⊆ R+ ,
the function x̂ is non-negative and so there exists a unique non-negative
function f ∈ C(S) such that f 2 = x̂. As the Gelfand map is onto, f = ŷ for
some y ∈ C. Since ŷ is a non-negative function, it is clear that y = y ∗ and
σ(y) ⊆ R+ .
To prove uniqueness, let z be any element satisfying the conclusion of the
lemma. Then z commutes with x and so the C∗ -algebra generated by x, z
and e is commutative and contains C. Under the Gelfand map applied to
this algebra, both y and z are represented by positive functions whose square
is the same (namely the function corresponding to x). Therefore y = z.
We now introduce the main concept of this section. A representation1 ρ
of A is a *-homomorphism of A into the C∗ -algebra B(H) of all bounded
linear operators on some Hilbert space H such that ρ(e) = I. That is, a
representation is a map ρ : A 7→ B(H) such that

ρ(λa + µb) = λρ(a) + µρ(b)

ρ(ab) = ρ(a).ρ(b)
ρ(a∗ ) = ρ(a)∗
ρ(e) = I.

1
In a more general context this would be called a *-representation of A on a Hilbert
space.
John Erdos 41

Note that in many other treatments ρ(e) = I is not included in the definition
of the term representation. However, the condition is convenient and, as will
be seen later, there is no essential loss in generality.
A representation is said to be faithful if it is an injection. Note that Theo-
rem 1.14 shows that (with the above definition) any representation is auto-
matically contiuous with kρ(a)k ≤ kak. The following lemma is included to
motivate the development.

Lemma 4.2 Let ρ be a representation of A on H and let ξ be a unit vector

of H. Then the function f : A 7→ C defined by

f (a) = hρ(a)ξ, ξi

satisfies

(i) f is a linear functional,

(ii) f (a∗ a) ≥ 0 for all a ∈ A,

(iii) f (e) = 1,

(iv) f is continuous

(v) kf k = 1 .

Proof. The only part that is not completely trivial is (ii) whose proof is as
follows :

f (a∗ a) = hρ(a∗ a)ξ, ξi = hρ(a∗ )ρ(a)ξ, ξi = kρ(a)ξk2 ≥ 0.

Parts (iv) and (v) follow from the fact (Theorem 1.14 ) that kρ(a)k ≤ kak.

A linear functional f : A 7→ C is said to be a positive linear functional if

f (a∗ a) ≥ 0 for all a ∈ A. If f satisfies properties (i) – (v) of Lemma 4.2 then
f is said to be a state of A. In Lemma 4.3 below we shall prove that every
positive linear functional is continuous and is a positive scalar multiple of a
state.
From the above, we have that every representation gives rise to a state. The
object of the GNS construction is to reverse the process; that is, to construct
a representation from any given state.
42 C*-algebras

Lemma 4.3 Let f be a positive linear functional on A. Then for all a, b ∈ A,

(i) f (b∗ a) = f (a∗ b) ,

(ii) |f (b∗ a)|2 ≤ f (a∗ a).f (b∗ b)

Proof. These results arise from the fact that φ(x, y) = f (y ∗ x) is a non-
negative sesquilinear form on A. The details follow.
(i) This is proved using polarization. We have

4b∗ a = (a + b)∗ (a + b) − (a − b)∗ (a − b) + i(a + ib)∗ (a + ib) − i(a − ib)∗ (a − ib)

4a∗ b = (a + b)∗ (a + b) − (a − b)∗ (a − b) − i(a + ib)∗ (a + ib) + i(a − ib)∗ (a − ib)

and, since f is positive and linear,

4f (a∗ b) = f [(a + b)∗ (a + b)] − f [(a − b)∗ (a − b)]

+if [(a + ib)∗ (a + ib)] − if [(a − ib)∗ (a − ib)] = 4f (b∗ a).

(ii) This is the Cauchy-Schwartz inequality. For all λ ∈ C,

f [(λa + b)∗ (λa + b)] = |λ|2 f (a∗ a) + λf (b∗ a) + λf (a∗ b) + f (b∗ b) ≥ 0 .

Choosing λ = ke−iθ where θ = arg f (b∗ a) with k real gives, using (i), that

k 2 f (a∗ a) + 2k|f (b∗ a)| + f (b∗ b) ≥ 0

for all k ∈ R, and the result merely states the fact that the discriminant of
the above quadratic is negative.

Corollary 4.4

(i) f (a∗ ) = f (a) ,

(ii) |f (a)|2 ≤ f (e)f (a∗ a) .

Proof. Put b = e in the lemma.

Lemma 4.5 If f is a positive linear functional on A then f is continuous

and kf k = f (e).
John Erdos 43

Proof. Suppose a = a∗ and kak ≤ 1. Then σ(a) ⊆ [−1, 1] and the spectral
mapping theorem shows that σ(e−a) ⊆ [0, 2]. Thus by Lemma 4.1 A contains
a selfadjoint element b such that b2 = e − a. Therefore f (e − a) = f (b∗ b) ≥ 0
and so
f (a) ≤ f (e) .
x∗ x
Now if x ∈ A, applying the above to and using Corollary 4.4 (ii) we
kx∗ xk
have that

|f (x)|2 ≤ f (e).f (x∗ x) ≤ [f (e)]2 kx∗ xk ≤ [f (e)]2 kxk2 .

Thus f is continuous with kf k ≤ f (e). But also since kek = 1,

f (e) = |f (e)| ≤ kf k

and so kf k = f (e).
Note that the only time that the C∗ condition, via Lemma 4.1, is used in this
section is in the above proof. By justifying the formal binomial expasion of
1
(e − x) 2 for kxk ≤ 1, the above lemma can be proved for Banach *-algebras
with isometric involution. This shows that the GNS construction is valid for
these more general algebras.
The following lemma gives an inner product space on which the GNS con-
struction will provide a representation.

Lemma 4.6 Let f be a positive linear functional on A and let N = {x :

f (x∗ x) = 0}. Then N is a left ideal and if [a] denotes the equivalence class
of a in the quotient space A/N then

h[a], [b]i = f (b∗ a)

defines an inner product on A/N .

Proof. If x ∈ N then from Lemma 4.3 (ii), for a ∈ A,

|f (ax)|2 ≤ f (x∗ x).f (a∗ a) = 0 .

Therefore we have the following characterisation of N :

N = {x : f (ax) = 0 for all a ∈ A} .

It is clear from this that N is a left ideal.

44 C*-algebras

We now show that f (b∗ a) depends only on the equivalence classes of a and
b. For all x, y ∈ N , using Lemma 4.3 (i) and the above,

f [(b + y)∗ (a + x)] = f (b∗ a) + f (b∗ x) + f [y ∗ (a + x)]

= f (b∗ a) + f (b∗ x) + f [(a + x)∗ y]
= f (b∗ a) .

Hence the relation

h[a], [b]i = f (b∗ a)
gives well-defined form on A/N and is clearly a non-negative sesquilinear
form. Also h[a], [a]i = 0 if and only if [a] = [0] (the zero of A/N ). Thereforeh·, ·i
is an inner product on A/N .

Lemma 4.7 If f is a positive linear functional on A, then for all x, y ∈ A

|f (y ∗ xy)| ≤ kxkf (y ∗ y) .

Proof. let g(a) = f (y ∗ ay). Then g(a∗ a) = f (y ∗ a∗ ay) ≥ 0 and so g is

positive. From Lemma 4.5 both f and g are continuous and kgk = g(e) =
f (y ∗ y). Hence

|f (y ∗ xy)| = |g(x)| ≤ kgk.kxk = kxkf (y ∗ y) ,

and the lemma is proved.

Recall that if K is any inner product space then its completion H is a Hilbert
space and K is (strictly one should say “can be identified with”) a dense
subspace of H.
The following theorem is called after Gelfand, Naimark and Segal.

Theorem 4.8 (GNS construction) Given any state on a C∗ -algebra with

identity, there exists a representation ρf on some Hilbert space Hf such that

f (a) = hρf (a)ξ, ξi

for some unit vector ξ of Hf .

Proof. Let N = {x : f (x∗ x) = 0}. Define Hf as the completion of the inner

product space A/N constructed as in Lemma 4.6 . Then Hf is a Hilbert
space.
John Erdos 45

For each [x] ∈ A/N and a ∈ A define

ρf (a)[x] = [ax] .
Since N is a left ideal it is clear that ρf (a) is a well- defined map form A/N
to itself. Clearly ρf (a) is linear and we now show that kρf (a)k ≤ kak. (Note
that this could not be deduced from Theorem 1.14 since A/N is not complete
and so B (A/N ) is not a C∗ -algebra.) Using Lemma 4.7 we have that,
kρf (a)xk2 = k[ax]k2 = f (x∗ a∗ ax)
≤ ka∗ akf (x∗ x)
≤ kak2 h[x], [x]i
= kak2A .k[x]k2A/N .

Therefore ρf (a) is a continuous linear operator on a dense subspace of Hf

and so it has a unique continuous extension to Hf . It should not cause any
confusion if the extension is also denoted by ρf (a).
Now a trivial verification shows that the map a 7→ ρf (a) is a homomorphism
and ρf (e) is easily seen to be the identity operator on A/N . In addition
hρf (a)[x], [y]i = = h[ax], [y]i = f (y ∗ ax) = f ((a∗ y)∗ x)
= h[x], [a∗ y]i
= h[x], ρf (a∗ )[y]i

which proves that ρf (a)∗ = ρf (a∗ ) on the dense subspace A/N of Hf and
hence, since both ρf (a)∗ and ρf (a∗ ) are bounded, they are equal.
We have thus proved that ρf is a representation of A on Hf . Also
hρf (a)[e], [e]i = h[a], [e]i = f (a)
and, since f is a state, k[e]k2Hf = f (e∗ e) = f (e) = 1 so [e] is a unit vector.
This completes the proof.

4.2 Positive elements and the Gelfand-Naimark theo-

rem

In this section, A continues to denote a C∗ -algebra with identity e. The

question of showing that A has a sufficiently rich supply of states hinges on
46 C*-algebras

the study of elements of the form a∗ a. The major tool in the investigation
of this is the commutative Gelfand-Naimark theorem.
An element a of A is said to be positive if it is selfadjoint and its spectrum
consists of non-negative real numbers. The main technical problem will be
to show that every element of the form x∗ x is positive. Note that, from
Theorem 1.16 , if B is a C∗ -subalgebra of A, with the same identity as A,
then σB (b) = σA (b) for all b ∈ B. Thus an element a is positive in A if
and only if it is positive in any one *-subalgebra containing it and so it is
sufficient to look at its spectrum in the C∗ -algebra generated by a and e.
The set of all positive elements of A will be denoted by A+ . Note that the
spectral mapping theorem (Theorem 1.6 ) shows that if a ∈ A+ then an ∈ A+
for all positive integers n.
John Erdos 47

Lemma 4.9 If a is a selfadjoint element of A then

(i) ke − ak ≤ 1 ⇒ a ∈ A+ ,
(ii) a ∈ A+ , kak ≤ 1 ⇒ ke − ak ≤ 1 ,
° °
(iii) a ∈ A+ ⇐⇒ °°kake − a°° ≤ kak .

Proof. Let C be the C∗ -algebra generated by a and e and suppose that Φ is

the carrier space of C. Since the range of the Gelfand transform â of a is the
spectrum of a, it follows that a ∈ A+ if and only if â is a positive function
on Φ.
To prove (i), if ke − ak ≤ 1 then kê − âk ≤ 1 and so |1 − â(φ)| ≤ 1 for all
φ ∈ Φ. Since a is selfadjoint, â is real-valued and so it follows that â(φ) ≥ 0
for all φ ∈ Φ. Therefore a ∈ A+ .
For (ii), if a ∈ A+ and kak ≤ 1 then kâk ≤ 1 and so 0 ≤ â(φ) ≤ 1 for all
φ ∈ Φ. Therefore ke − ak = kê − âk ≤ 1.
a
Finally (iii) results from applying (i) and (ii) to .
kak

Theorem 4.10 A+ is a closed convex cone with A+ ∩ (−A+ ) = (0).

Proof. If a ∈ A+ then clearly λa ∈ A+ for all λ ≥ 0. To prove that A+ is

convex suppose a, b ∈ A+ and λ, µ ≥ 0 with λ + µ = 1. If a = b = 0 then
a b
λa + µb = 0 ∈ A+ . Otherwise, and are both positive
kak + kbk kak + kbk
and have norm at most 1. Hence, using Lemma 4.9 ,
° ° ° Ã !° ° Ã !°
° λa + µb °° ° a ° ° b °
° ° ° ° °
°e − ° ≤ °λ e − ° + °µ e − °
° kak + kbk ° ° kak + kbk ° ° kak + kbk °
≤ λ+µ=1

λa + µb
and so by part (i) of Lemma 4.9 , ∈ A+ . Consequently, λa + µb ∈
kak + kbk
A+ and so A+ is a convex cone.
If x ∈ A+ ∩ (−A+ ) then x is selfadjoint and σ(x) = {0}. Hence ν(x) = 0 and
so from Corollary 1.10, x = 0.
To show that A+ is closed, note that the set S of all selfadjoint elements of
A is closed. Lemma implies that A+ is a closed subset of S and so A+
is closed.
48 C*-algebras

Theorem 4.11 The following conditions on an element a of A are equiva-

lent.

(i) a ∈ A+ ,

(ii) a = x2 for some selfadjoint element x of A ,

(iii) a = y ∗ y for some element y of A .

Proof. Lemma 4.1 establishes (i)⇒ (ii) and (ii) ⇒ (iii) is trivial. The
substance of the Theorem is to prove that (iii) ⇒ (i).
If a = y ∗ y then a is self-adjoint. Let Φ be the carrier space of the commutative
C∗ -algebra C generated by a and e. Then â is a real function on Φ and,
writing f (φ) = max[â(φ), 0] and g(φ) = − min[â(φ), 0], we have that f and
g are non-negative functions on Φ with

â = f − g and f g = gf = 0 .

Since f, g ∈ C(Φ), from the commutative Gelfand-Naimark theorem (Theo-

rem 2.8 ) we have that for some b, c ∈ C, b̂ = f and ĉ = g. Thus

a = y∗y = b − c

where bc = cb = 0. Since an element of C is in A+ if and only if its Gelfand

transform is non-negative, we have that b, c ∈ A+ . To prove that a ∈ A+ we
shall show that c = 0.
Let u and v be, respectively, the real and imaginary parts of cy ∗ . Since u
and v are selfadjoint the spectrum of each is real and the spectral mapping
theorem shows that u2 and v 2 are positive. Since A+ is a convex cone,

(cy ∗ )∗ cy ∗ + cy ∗ (cy ∗ )∗ = 2(u2 + v 2 ) ∈ A+ . (1)

Also
cy ∗ (cy ∗ )∗ = cy ∗ yc = cbc − c3 = −c3 ∈ −A+ . (2)
Therefore, using (1), (2) and Theorem 4.10 ,

(cy ∗ )∗ cy ∗ = c3 + 2(u2 + v 2 ) ∈ A+ . (3)

However, since (cy ∗ )∗ cy ∗ and cy ∗ (cy ∗ )∗ have the same non-zero numbers in
their spectrum (Lemma 1.8), (3) implies that

cy ∗ (cy ∗ )∗ = −c3 ∈ A+ .
John Erdos 49

Thus c3 ∈ A+ ∩ (−A+ ) and so, using Theorem 4.10 , c3 = 0. This implies

that c = 0 and the Theorem is proved.
The above theorem shows that the various notions of positivity are identical.
As a result we now have that a linear functional is positive if and only if it
is positive on positive elements. This is needed for the following result.

Lemma 4.12 Let f be a linear functional on A such that kf k = f (e) = 1.

Then f is a state of A.

Proof. We show that f is positive and therefore a state. First we show that
if a is selfadjoint then f (a) is real. We may suppose that kak ≤ 1. Then, for
any real k, using the C∗ - condition,

ka + ikek2 = k(a − ike)(a + ike)k = ka2 + k 2 ek ≤ 1 + k 2 .

Now, if f (a) = α + iβ, using kf k = f (e) = 1 it follows that

|f (a + ike)|2 = |α + i(k + β)|2 = α2 + k 2 + β 2 + 2kβ ≤ ka + ikek2 ≤ 1 + k 2 .

If β 6= 0, this produces a contradiction for positive or negative k of sufficiently

large modulus. Hence β = 0 and f (a) is real. Now suppose that a is positive
with kak ≤ 1. Then ke − ak ≤ 1 so, using kf k = 1, −1 ≤ f (e − a) ≤ 1. Since
f (e − a) = 1 − f (a) this implies that f (a) ≥ 0.

Lemma 4.13 For any non-zero element x of A there exists a state f of A

such that f (x∗ x) 6= 0.

Proof. Let C be the C∗ -algebra generated by e and x∗ x. The Gelfand

mapping theorem shows that for any λ in the spectrum of x∗ x, there is a
homomorphism φ of C into C such that φ(x∗ x) = λ. In particular, φ is a
linear functional on C such that kφk = φ(e) = 1 and clearly, by the choice of
λ we can ensure that φ(x∗ x) 6= 0. A simple application of the Hahn-Banach
theorem gives an extension f of φ to A such that kf k = kφk. It follows from
Lemma 4.12 that f is a state.

Theorem 4.14 (Gelfand-Naimark) Any C∗ -algebra is isometrically *-isomorphic

to a closed *-subalgebra of operators on some Hilbert space.
50 C*-algebras

Proof. We first prove the theorem for the case of a C∗ -algebra A with identity
e. For each non-zero x ∈ A let fx be a state (as found in the preceding lemma)
such that fx (x∗ x) 6= 0. Let ρx be the representation of A on a Hilbert space
Hx constructed from fx by the GNS construction Theorem 4.8 . Then

kρx (x)[e]k2 = fx (ex∗ xe) = fx (x∗ x) 6= 0

and so ρx (x) 6= 0.
Let H = ⊕{Hx : x ∈ A} and define ρ : A → B(H) by

ρ(a) = ⊕{ρx (a) : x ∈ A} .

Then ρ is clearly a representation and, since ρa (a) 6= 0, ρ is injective. Hence,

from Corollary 1.15 , it is isometric. Thus the theorem is proved for algebras
with identity.
If A has no identity and A1 is the C∗ -algebra formed by adjoining an identity
to A, then A is a closed *-subalgebra of A1 . Hence, applying the result
proved above to A1 we find an isometric *-representation ρ1 of A1 into B(H)
for some Hilbert space H. The result now follows just by restricting ρ1 to A.

The above theorem may be rephrased as follows : every C∗ -algebra has a

faithful representation.
John Erdos 51

Suggestions for further reading

1. R.V. Kadison and J.R. Ringrose, Fundamentals of the theory of opera-

tor algebras, Academic Press, New York, 1983.

2. L.H. Loomis, Abstract harmonic analysis, Van Nostrand, Princeton,

1953.

3. G.J. Murphy, C∗ -algebras and operator theory, Academic Press, San

Diego, 1990.

4. G.K. Pedersen C∗ -algebras and their automorphism groups, Academic

Press, London, 1979.

5. W. Rudin, Fourier analysis on groups, Interscience, New York, 1962.