Lecture Notes on

Operator Algebras

Draft: December 14, 2011

N.P. Landsman
Institute for Mathematics, Astrophysics, and Particle Physics
Radboud University Nijmegen
Heyendaalseweg 135
6525 AJ NIJMEGEN
THE NETHERLANDS
email: [email protected]
website: http://www.math.ru.nl/∼landsman/OA2011.html
tel.: 024-3652874
office: HG03.740
1 HISTORICAL NOTES AND OVERVIEW 2

1 Historical notes and overview

The theory of operator algebras tends to be rather technical. Some textbooks start
straight away with the definition of at least three (and at most eleven) different
topologies on the space of all bounded operators on a Hilbert space. Instead, we
begin with an informal survey, which is partly historical in nature. One of our aims
is to emphasize the origins of operator algebras in (quantum) physics.
In fact, the theory of operator algebras has two quite different sources in the
1930s and 1940s, respectively, with two associated great mathematicians:

• Hilbert spaces (John von Neumann),1 leading to the theory of von Neumann
algebras (originally called rings of operators by von Neumann himself).

• Commutative Banach algebras (Israel Gelfand),2 giving rise to C ∗ -algebras.

So, roughly speaking, the theory of operator algebras is the same as the theory of
von Neumann algebras and C ∗ -algebras.3 Let us elaborate on each of these in turn.

1.1 John von Neumann and quantum mechanics

John von Neumann (1903–1957) was a Hungarian prodigy; he wrote his first math-
ematical paper at the age of seventeen. Except for this first paper, his early work
was in set theory and the foundations of mathematics. In the Fall of 1926, he
moved to Göttingen 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 Gödel’s later
work. However, at the time that von Neumann arrived, Hilbert was also interested
in quantum mechanics.
From 1900 onwards, physicists had begun to recognize that the classical physics of
Newton and Maxwell could not describe all of Nature. The fascinating era that was
thus initiated by Planck, to be continued mainly by Einstein and Bohr, ended in 1925
with the discovery of quantum mechanics. This theory replaced classical mechanics,
and was initially discovered in two guises. Schrödinger was led to a formulation
called ‘wave mechanics,’ in which the famous symbol Ψ, denoting a ‘wave function,’
played an important role. Heisenberg discovered a form of quantum mechanics that
at the time was called ‘matrix mechanics.’ The relationship and possible equivalence
between these alternative formulations of quantum mechanics, which at first sight
looked completely different, was much discussed at the time.
Heisenberg’s paper initiating matrix mechanics was followed by the ‘Dreimän-
nerarbeit’ of Born, Heisenberg, and Jordan (1926); all three were in Göttingen at
the time. Born was one of the few physicists of his day to be familiar with the
1
Some of von Neumann’s papers on operator algebras were coauthored by F.J. Murray.
2
Some of Gelfand’s papers on operator algebras were coauthored by M. Naimark.
3
More recently, operator algebras that are not closed under involution have been studied, in
close connection to so-called operator spaces. See, for example, [12].
1 HISTORICAL NOTES AND OVERVIEW 3

concept of a matrix; in previous research he had even used infinite matrices.4 Born
turned to his former teacher Hilbert for mathematical advice. 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. 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 (which is now obsolete).
It was von Neumann alone who, at the age of 23, recognized the mathematical
structure of quantum mechanics. In this process, he defined 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 Göttingen on integral
equations, spectral theory, and infinite-dimensional quadratic forms. Hilbert’s fa-
mous memoirs on integral equations had appeared between 1904 and 1906. In 1908,
his student E. Schmidt had defined 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 L2 -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 Schrödinger’s wave functions were unit vectors in a
Hilbert space of L2 type, and that Heisenberg’s observables were linear operators
on a different Hilbert space, of `2 type. A unitary transformation between these
spaces provided the the mathematical equivalence between wave mechanics and
matrix mechanics. (Similar, mathematically incomplete insights had been reached
by Pauli and Dirac.) In a series of papers that appeared between 1927–1929, von
Neumann defined 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 Mathematische Grundlagen
der Quantenmechanik (1932), which to this day remains the definitive account of
the mathematical structure of elementary quantum mechanics.5
More precisely, von Neumann proposed the following mathematical formulation
of quantum mechanics (cf. Ch. 2 below for minimal background on Hilbert spaces).

1. The observables of a given physical system are the self-adjoint linear opera-
tors a on a Hilbert space H.

2. The states of the system are the so-called density operators ρ̂ on H, that
is, the positive trace-class operators on H with unit trace.

3. The expectation value < a >ρ̂ of an observable a in a state ρ̂ is given by

< a >ρ̂ = Tr (ρ̂a). (1.1)

4
Heisenberg’s fundamental equations of quantum mechanics, viz. pq − qp = −~i, which initially
were quite mysterious, could only be satisfied by infinite-dimensional matrices.
5
Von Neumann’s book was preceded by Dirac’s The Principles of Quantum Mechanics (1930),
which contains another brilliant, but this time mathematically questionable account of quantum
mechanics in terms of linear spaces and operators. See e.g. [29, 34] for modern accounts.
1 HISTORICAL NOTES AND OVERVIEW 4

As a special case, take a unit vector Ψ in H and form the associated projection
pΨ = |ΨihΨ|, (1.2)
where we use the following notation (due to Dirac): for any two vectors Ψ, Φ in H,
the operator |ΨihΦ| is defined by6
|ΨihΦ|Ω = (Φ, Ω)Ψ. (1.3)
In particular, if Φ = Ψ is a unit vector, then (1.2) is the projection onto the one-
dimensional subspace C · Ψ of H spanned by Ψ (see Exercise 1 below for a rappèl
on projections). In that case, it is easily shown (see Exercise 2) that the density
operator ρ̂ = pΨ leads to
< a >pΨ ≡< a >Ψ = Tr (pΨ a) = (Ψ, aΨ). (1.4)
Special states like pΨ (often confused with Ψ itself, which contains additional phase
information) are called pure states, whereas all other states are said to be mixed.
Let B(H) be the space of all bounded operators on H, with unit operator simply
denoted by 1. A functional ω : B(H) → C (which is linear by definition) is called:
• positive when ω(a∗ a) ≥ 0 for all a ∈ B(H);
• normalized when ω(1) = 1.
We will see that positivity implies continuity (see Exercise 3). For reasons to become
clear shortly, a normalized positive functional on B(H) is called a state on B(H);
our earlier use of this word should accordingly be revised a little. Indeed, it is trivial
to show that for any density operator ρ̂, the functional ρ : B(H) → C defined by
ρ(a) = Tr (ρ̂a) (1.5)
is a state on B(H). Conversely, are all states on B(H) of this kind?
Von Neumann implicitly assumed a certain continuity condition on states, which
in modern terminology is called σ-weak (or ultraweak) continuity, which implies
that the answer is yes; states à la (1.5) (in other words, σ-weakly continuous states)
are called normal states on B(H). However, without this continuity condition the
set of states on B(H) turns out not to be exhausted by density operators on H
(unless H is finite-dimensional), although it is hard to give explicit examples.
The set of all states on B(H) is obviously convex (within B(H)∗ ), as is its subset
of all normal states on B(H). The extreme boundary of a convex set K is the set
of all ω ∈ K that are indecomposable, in the sense that if ω = λω1 + (1 − λ)ω2 for
some 0 < λ < 1 and ω1 , ω2 ∈ K, then ω1 = ω2 = ω.7 Von Neumann saw that states
(1.4) precisely correspond to the points in the extreme boundary of the convex set
Sn (B(H)) of normal states on B(H): on other words, a density operator ρ̂ yields an
extreme point ρ in Sn (B(H)) iff it is of the form (1.4). See also Exercise 4.
6
Dirac wrote |Ωi for Ω, etc., which is superfluous, but in this special case it leads to the neater
expression |ΨihΦ|Ωi = hΦ|Ωi|Ψi, where the inner product is written as hΦ|Ωi ≡ (Φ, Ω).
7
In some examples of compact convex sets in Rn , 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 even be
empty, as illustrated by the open unit ball in any dimension.
1 HISTORICAL NOTES AND OVERVIEW 5

1.2 Von Neumann algebras

In one of his papers on Hilbert space theory (1929), von Neumann defined 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 in the weak operator topology. This means that:
• M is a subalgebra of B(H) with unit under operator multiplication;
• M is closed under taking the adjoint (or Hermitian conjugate) a 7→ a∗ ;
• If (v, (aλ − a)w) → 0 for all v, w ∈ H for some net (aλ ) in M , then a ∈ M .
For example, B(H) is itself a von Neumann algebra. When H is finite-dimensional,
any direct sum of matrix algebras containing 1 is a von Neumann algebra.
In the same paper, von Neumann proved what is still the first and most basic
theorem of the subject, called the Double Commutant Theorem:8
Let M be a unital ∗ -subalgebra of B(H). Then the following condtions
are equivalent (and hence each defines M to be a von Neumann algebra):
• M 00 = M ;
• M is closed in the weak operator topology;
• M is closed in the strong operator topology;
• M is closed in the σ-weak operator topology.
Here, for any Hilbert space H, let S ⊂ B(H) be some subset. The commutant of
S is defined by
S 0 := {b ∈ B(H) | ab = ba ∀a ∈ S}. (1.6)
Note that S 0 is a subalgebra of B(H). Similarly, one defines the bicommmutant
S 00 = (S 0 )0 of S (it makes no sense to go on, since S 000 = S 0 , see Exercise 5).
The strong operator topology on B(H) may be defined in saying that a net aλ
converges to a iff aλ v → av for all v ∈ H. It should be mentioned that, though easily
defined, neither the weak topology on a von Neumann algebra M nor the strong
one is a natural one; the natural topology on a von Neumann algebra M ⊆ B(H)
turns out to be the σ-weak or ultraweak one. This topology is provided by the
seminorms kakρ̂ = |Tr (ρ̂a)|, where ρ̂ is an element of the trace-class B1 (H), cf. §2.10.
Hence aλ → a σ-weakly when Tr (ρ̂(aλ − a)) → 0 for all ρ̂ ∈ B1 (H). Here it turns
out that one could equally well restrict ρ̂ to be a density operator on H, so that a
physicist would be justified in saying that the σ-weak topology is the topology of
pointwise convergence of quantum-mechanical expectation values.
Von Neumann’s motivation in studying rings of operators was plurifold; beyond
quantum theory, we mention probability theory, entropy, ergodic theory, discrete
groups, representation theory, projective geometry, and lattice theory.
8
This theorem is remarkable, in relating a topological condition to an algebraic one; one is
reminded of the much simpler fact that a linear subspace K of H is closed iff K ⊥⊥ = K, where
K ⊥ is the orthogonal complement of K in H.
1 HISTORICAL NOTES AND OVERVIEW 6

1.3 C*-algebras
Following the pioneering work of von Neumann, an important second step in the the-
ory of operator algebras was the initiation of the theory of C ∗ -algebras by Gelfand
and Naimark in 1943. It turns out that von Neumann’s “rings of operators” are
special cases of C ∗ -algebras, but von Neumann algebras also continue to be studied
on their own. A fruitful mathematical analogy is that C ∗ -algebras provide a non-
commutative generalization of topology, whereas von Neumann algebras comprise
noncommutative measure theory.9 To understand this, we state the most important
fact about C ∗ -algebras, namely that there are two totally different way of approach-
ing them:

• As norm-closed ∗ -subalgebras A ⊂ B(H);

• As noncommutative generalizations of the space C(X) ≡ C(X, C) of complex-

valued continuous functions on a compact space X.

We start with the second. In 1943, Gelfand and Naimark 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 (a ∈ C(X))

kak∞ := sup{|a(x)|, x ∈ X},

in which it is a Banach space. This Banach space structure of C(X) is compatible

with its structure as an algebra by the property

kabk ≤ kakkbk, (1.7)

where we have written k·k for k·k∞ , and a, b ∈ C(X). But more structure is needed!
An involution on an algebra A is a real-linear map A → A∗ such that a∗∗ = a,
(ab)∗ = b∗ a∗ , and (λa)∗ = λa∗ for all a, b ∈ A and λ ∈ C. An algebra with involution
is also called a ∗ -algebra. Secondly, then, C(X) has an involution a 7→ a∗ , given
by a∗ (x) = a(x). This involution is related to the norm as well as to the algebraic
structure by the property
ka∗ ak = kak2 . (1.8)
We summarize these properties by saying that C(X) is a commutative C ∗ -algebra
with unit, in the following sense:

A (commutative) C ∗ -algebra is a Banach space that at the same time

is a (commutative) algebra with involution, such that the compatibility
conditions (1.7) and (1.8) hold.

The first theorem of Gelfand and Naimark then reads as follows:

Every commutative C ∗ -algebra A with unit is isomorphic to C(X) for

some compact Hausdorff space X, unique up to homeomorphism.
9
These analogies form the basis of noncommutative geometry as developed by Connes [5].
1 HISTORICAL NOTES AND OVERVIEW 7

The isomorphism is constructed as follows. The space X is often denoted by

Σ(A) and is called the Gelfand spectrum of A. It may be realized as the set
of multiplicative functionals or characters on A, that is, nonzero linear maps
ω : A → C that satisfy ω(ab) = ω(a)ω(b).10 Thus one takes X := ∆(A), which turns
out to be a compact Hausdorff space in the topology of pointwise convergence, and
the map A → C(X) is the so-called Gelfand transform a 7→ â, where â(ω) := ω(a).
Similar to their characterization of commutative C ∗ -algebras above, Gelfand and
Naimark also quite brilliantly clarified the nature of general C ∗ -algebras. Let us
first note that B(H) is a C ∗ -algebra (See Exercise 6). Moreover, any norm-closed
∗
-subalgebra of B(H) is a C ∗ -algebra (which is trivial, given the previous result).
Perhaps surprisingly, this is also the most general kind of C ∗ -algebra, for the second
theorem of Gelfand and Naimark (contained in the same paper as their first) reads:
Every C ∗ -algebra A is isomorphic to a norm-closed ∗ -subalgebra of B(H),
for some Hilbert space H.
Note that, in contrast to the previous theorem, H is by no means unique in any
sense! The proof of this theorem is based on the so-called gns-construction (after
Gelfand, Naimark, and the American mathematician I.E. Segal), which basically
explains why C ∗ -algebras are naturally related to Hilbert spaces, and which pervades
the subject in every conceivable way. This construction starts with the concept of
a state on a C ∗ -algebra A, which we have already encountered for A = B(H):
A state on a C ∗ -algebra A with unit is a functional ω : A → C that
satisfies ω(a∗ a) ≥ 0 ∀a ∈ A, (positivity) and ω(1) = 1 (normalization).11
The characters of a commutative C ∗ -algebra are examples of states. Let us first
suppose that ω(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)ω := ω(a∗ b), which may be completed
into a Hilbert space Hω . Then A acts on Hω by means of πω : A → B(Hω ), given
by πω (a)b := ab.12 It is easy to see that πω is injective: if πω (a) = 0 then, taking
b = 1, one infers that a = 0 as an element of Hω , but then (a, a)ω = ω(a∗ a) = 0,
contradicting the assumption that ω(a∗ a) > 0. Moreover, one checks that πω is a
homomorphism of C ∗ -algebras, for example,
πω (a)πω (b)c = πω (a)bc = abc = πω (ab)c
for all c, which implies πω (a)πω (b) = πω (ab). Thus A is isomorphic to πω (A) ⊂
B(Hω ).13 In general, A may not possess such strictly positive functionals, but it
always has sufficiently many states. For an arbitrary state ω, the Hilbert space Hω
is constructed by first dividing A by the kernel of (·, ·)ω , and proceeding in the same
way. The representation πω may then fail to be injective, but by taking the direct
sum of enough such representations one always arrives at an injective one.
10
These functionals lie in the dual space A∗ of A. The topology in which the theorem holds is
the (relative) weak∗ topology on Σ(A) ⊂ A∗ , also called the Gelfand topology.
11
Positive functionals on a C ∗ -algebra are continuous, with norm kωk = ω(1). A state on a
∗
C -algebra A without unit is defined as a positive functional ω : A → C that satisfies kωk = 1.
12
Initially defined on the dense subspace A ⊂ Hω , and subsequently extended by continuity.
13
A more technical argument shows that πω is isometric, so that its image is closed.
1 HISTORICAL NOTES AND OVERVIEW 8

Exercises for Lecture 1

1. A (bounded) operator p : H → H is called a projection when p2 = p∗ = p.
Prove that there is a bijective correspondence between projections in B(H)
and closed subspaces of H (which by definition are linear ).
Hint: the closed subspace Kp corresponding to p is Kp = pH ≡ Range(p).
Find the projection pK corresponding to some given closed subspace K ⊂ H
and show that pKp = p and KpK = K.

2. Prove (1.4).

3. Prove that a state ω on a C ∗ -algebra A (with unit) satisfies |ω(a)| ≤ kak (and
hence is continuous with norm kωk = ω(1)).

4. The convex structure of the state space is nicely displayed by H = C2 , so that

B(H) = M2 (C), the C ∗ -algebra of 2 × 2 complex matrices. Put
 
1 1 + z x + iy
ρ(x, y, z) = 2 . (1.9)
x − iy 1 − z

(a) Show that ρ(x, y, z) is a density operator on C2 iff (x, y, z) ∈ R3 with

x2 + y 2 + z 2 ≤ 1.

(b) Show that every state ω on M2 (C) is of the form ωρ , with

ωρ (a) = Tr (ρa). (1.10)

(c) Conclude that the state space S(M2 (C)) of M2 (C) is isomorphic (as a
convex set) to the three-ball B 3 = {(x, y, z) ∈ R3 | x2 + y 2 + z 2 ≤ 1}.
(d) Under this isomorphism, show that the extreme boundary of S(M2 (C))
corresponds to the two-sphere S 2 = {(x, y, z) ∈ R3 | x2 + y 2 + z 2 = 1}.
(e) Verify that the states in the extreme boundary of S(M2 (C)) are exactly
those of the form (1.4), where Ψ ∈ C2 is a unit vector.

5. Prove that S 000 = S 0 , for any subset S ⊂ B(H).

6. Prove that B(H) is a C ∗ -algebra.

Hint: the only reasonably difficult part is the proof of ka∗ ak = kak2 .
2 REVIEW OF HILBERT SPACES 9

2 Review of Hilbert spaces

2.1 Inner product, norm, and metric
The following definitions are basic to all of functional analysis. Note that the concept
of a metric applies to any set (i.e., not necessarily to a vector space).
Definition 2.1 Let V be a vector space over C.
1. An inner product on V is a map V × V → C, written as hf, gi 7→ (f, g),
satisfying, for all f, g, h ∈ V , t ∈ C:
(a) (f, f ) ∈ R+ := [0, ∞) (positivity);
(b) (g, f ) = (f, g) (symmetry);
(c) (f, tg) = t(f, g) (linearity 1);
(d) (f, g + h) = (f, g) + (f, h) (linearity 2);
(e) (f, f ) = 0 ⇒ f = 0 (positive definiteness).
2. A norm on V is a function k · k : V → R+ such that for all f, g, h ∈ V , t ∈ C:
(a) kf + gk ≤ kf k + kgk (triangle inequality);
(b) ktf k = |t|kf k (homogeneity);
(c) kf k = 0 ⇒ f = 0 (positive definiteness).
3. A metric on V is a function d : V × V → R+ satisfying, for all f, g, h ∈ V :
(a) d(f, g) ≤ d(f, h) + d(h, g) (triangle inequality);
(b) d(f, g) = d(g, f ) for all f, g ∈ V (symmetry);
(c) d(f, g) = 0 ⇔ f = g (definiteness).
These structures are related in the following way:
Proposition 2.2 1. An inner product on V defines a norm on V by
p
kf k = (f, f ). (2.11)

2. This norm satisfies the Cauchy–Schwarz inequality

|(f, g)| ≤ kf kkgk. (2.12)

3. A norm k · k on a complex vector space comes from an inner product iff

kf + gk2 + kf − gk2 = 2(kf k2 + kgk2 ), (2.13)
in which case
(f, g) = 41 (kf + gk2 − kf − gk2 + ikf − igk2 − ikf + igk2 ). (2.14)

4. A norm on V defines a metric on V through d(f, g) := kf − gk.

2 REVIEW OF HILBERT SPACES 10

2.2 Completeness
Many concepts of importance for Hilbert spaces are associated with the metric rather
than with the underlying inner product or norm. The main example is convergence:

Definition 2.3 1. Let (xn ) := {xn }n∈N be a sequence in a metric space (V, d).
We say that xn → x for some x ∈ V when limn→∞ d(xn , x) = 0, or, more
precisely: for any ε > 0 there is N ∈ N such that d(xn , x) < ε for all n > N .
In a normed space, hence in particular in a space with inner product, this
therefore means that xn → x if limn→∞ kxn − xk = 0.

2. A sequence (xn ) in (V, d) is called a Cauchy sequence when d(xn , xm ) → 0

when n, m → ∞; more precisely: for any ε > 0 there is N ∈ N such that
d(xn , xm ) < ε for all n, m > N .
In a normed space, this means that (xn ) is Cauchy when kxn − xm k → 0 for
n, m → ∞, in other words, if limn,m→∞ kxn − xm k = 0.

Clearly, a convergent sequence is Cauchy: from the triangle inequality and symmetry
one has d(xn , xm ) ≤ d(xn , x) + d(xm , x), so for given ε > 0 there is N ∈ N such that
d(xn , x) < ε/2, et cetera. However, the converse statement does not hold in general,
as is clear from the example of the metric space (0, 1) with metric d(x, y) = |x − y|:
the sequence xn = 1/n does not converge in (0, 1). In this case one can simply
extend the given space to [0, 1], in which every Cauchy sequence does converge.

Definition 2.4 A metric space (V, d) is called complete when every Cauchy se-
quence in V converges (i.e., to an element of V ).

• A vector space with norm that is complete in the associated metric is called a
Banach space. In other words: a vector space B with norm k · k is a Banach
space when every sequence (xn ) such that limn,m→∞ kxn − xm k = 0 has a limit
x ∈ B in the sense that limn→∞ kxn − xk = 0.

• A vector space with inner product that is complete in the associated metric is
called a Hilbert space. In other words: a vector space H with innerpproduct
( , ) is a Hilbert space when it is a Banach space in the norm kxk = (x, x).

A subspace of a Hilbert space may or may not be closed. A closed subspace

K ⊂ H of a Hilbert space H is by definition complete in the given norm on H (i.e.
any Cauchy sequence in K converges to an element of K).14 This implies that a
closed subspace K of a Hilbert space H is itself a Hilbert space if one restricts the
inner product from H to K. If K is not closed already, we define its closure K as
the smallest closed subspace of H containing K; once again, this is a Hilbert space.
14
Since H is a Hilbert space we know that the sequence has a limit in H, but this may not lie in
K even when all elements of the sequence do. This is possible precisely when K fails to be closed.
2 REVIEW OF HILBERT SPACES 11

2.3 Geometry of Hilbert space

n
The vector spaces
Pn C from linear algebra are Hilbert spaces in the usual inner prod-
uct (z, w) = k=1 zk wk . Indeed, a finite-dimensional vector space is automatically
complete in any possible norm. More generally, Hilbert spaces are the vector spaces
whose geometry is closest to that of Cn , because the inner product yields a notion
of orthogonality: we say that two vectors f, g ∈ H are orthogonal, written f ⊥ g,
when (f, g) = 0.15 Similary, two subspaces16 K ⊂ H and L ⊂ H are said to be
orthogonal (K ⊥ L) when (f, g) = 0 for all f ∈ K and all g ∈ L. A vector f is
called orthogonal to a subspace K, written f ⊥ K, when (f, g) = 0 for all g ∈ K,
etc. We define the orthogonal complement K ⊥ of a subspace K ⊂ H as

K ⊥ := {f ∈ H | f ⊥ K}. (2.15)

This set is linear, so that the map K 7→ K ⊥ , called orthocomplementation, is an

operation from subspaces of H to subspaces of H. Clearly, H ⊥ = 0 and 0⊥ = H.
Closure is an analytic concept, related to convergence of sequences. Orthogo-
nality is a geometric concept. However, both are derived from the inner product.
Hence one may expect connections relating analysis and geometry on Hilbert space.

Proposition 2.5 Let K ⊂ H be a subspace of a Hilbert space.

1. The subspace K ⊥ is closed, with

⊥
K ⊥ = K = K ⊥. (2.16)

2. One has
K ⊥⊥ := (K ⊥ )⊥ = K. (2.17)

3. Hence for closed subspaces K one has K ⊥⊥ = K.

Definition 2.6 An orthonormal basis (o.n.b.) in a Hilbert space is a set (ek )

of vectors satisfying (ek , el ) = δkl and being such that
Pany v ∈ H can be written as
P N
v = k vk ek for some vk ∈ C, in that limN →∞ kv − k=1 vk ek k = 0.

If v = k vk ek , then, as in linear algebra, vk = (ek , v), and k |vk |2 = kvk2 . This is

P P
called Parseval’s equality; it is a generalization of Pythagoras’s Theorem.
Once more like in linear algebra, all o.n.b. have the same cardinality, which de-
fines the dimension of H. We call an infinite-dimensional Hilbert space separable
when it has a countable o.n.b. Dimension is a very strong invariant: running ahead
of the appropriate definition of isomorphism of Hilbert spaces in §2.4, we have

Theorem 2.7 Two Hilbert spaces are isomorphic iff they have the same dimension.
15
By definition of the norm, if f ⊥ g one has Pythagoras’ theorem kf + gk2 = kf k2 + kgk2 .
16
A subspace of a vector space is by definition a linear subspace.
2 REVIEW OF HILBERT SPACES 12

2.4 The Hilbert spaces `2

We say that H1 and H2 are isomorphic as Hilbert space when there exists an
invertible linear map u : H1 → H2 that preserves the inner product, in that
(uf, ug)H2 = (f, g)H1 for all f, g ∈ H1 ; this clearly implies that also the inverse
of u preserves the inner product. Such a map is called unitary.
To prove Theorem 2.7, we first introduce a Hilbert spaces `2 (S) for any set S (in
the proof, S will be a set labeling some o.n.b., like S = N in the countable case).
• If S is finite, then `2 (S) = {f : S → C} with inner product
X
(f, g) = f (s)g(s). (2.18)
s∈S

The functions (δs )s∈S , defined by δs (t) = δst , t ∈ S, clearly form an o.n.b. of `2 (S).
Now let H be an n-dimensional Hilbert space; a case in point is H = Cn .
By definition, H has an o.n.b. (ei )ni=1 . Take S = n = {1, 2, . . . , n}. The map
u : H → `2 (n), given by linear extension of uei = δi is unitary and provides an
isomorphism H ∼ = `2 (n). Hence all n-dimensional Hilbert space are isomorphic.
• If S is countable, then `2 (S) = {f : S → C | kf k2 < ∞}, with
!1/2
X
kf k2 := |f (s)|2 , (2.19)
s∈S

with inner product given by (2.18); this is finite for f, g ∈ `2 (S) by the Cauchy–
Schwarz inequality. Once again, the functions (δs )s∈S form an o.n.b. of `2 (S), and
the same argument shows that all separable Hilbert space are isomorphic to `2 (N)
and hence to each other. A typical example is `2 (Z).
• If S is uncountable, then `2 (S) is defined as in the countable case, where
the sum in (2.19) is now defined as the supremum of the same expression
evaluated on each finite subset of S. Similarly, the sum in (2.18) is defined by
first decomposing f = f1 − f2 + i(f3 − f4 ) with fi ≥ 0, and g likewise; this
decomposes (f, g) as a linear combination of 16 non-negative terms (fi , gj ),
each of which is defined as the supremum over finite subsets of S, as for kf k2 .
The previous construction of an o.n.b. of `2 (S) still applies verbatim, as does the
proof that any Hilbert space of given cardinality is isomorphic to `2 (S) for some S
of the same cardinality. In sum, we have proved (von Neumann’s) Theorem 2.7.
Let us note that for infinite sets S we may regard `2 (S) as the closure in the
norm (2.19) of the (incomplete) space `c (S) of functions that are nonzero at finitely
many s ∈ S; this means that for any f ∈ `2 (S) there is a sequence (fn ) in `c (S) such
that limn→∞ kfn − f k2 = 0. In what follows, we also encounter the Banach space
`∞ (S) = {f : S → C | kf k∞ < ∞}; (2.20)
kf k∞ := sup{|f (s)|}, (2.21)
s∈S

which is evidently the closure of `c (S) in the supremum-norm k · k∞ , in that for

any f ∈ `∞ (S) there is a sequence (fn ) in `c (S) such that limn→∞ kfn − f k∞ = 0.
2 REVIEW OF HILBERT SPACES 13

2.5 The Hilbert spaces L2

A more complicated example of a Hilbert space is L2 (Rn ), familiar from quantum
mechanics. which can be defined either directly through measure theory (see §2.6), or
indirectly, as a completion of Cc (Rn ), the vector space of complex-valued continuous
functions on Rn with compact support.17 Two natural norms on Cc (Rn ) are:
kf k∞ := sup{|f (x)|, x ∈ Rn }, (2.22)
Z 1/2
n 2
kf k2 := d x |f (x)| . (2.23)
Rn
The first norm is called the supremum-norm or sup-norm; see §2.7. The second
norm is called the L2 -norm. It is, of course, derived from the inner product
Z
(f, g) := dn x f (x)g(x). (2.24)
Rn
Now, Cc (Rn ) fails to be complete in either norm k · k∞ or k · k2 .
• The completion of Cc (Rn ) in the norm k · k∞ turns out to be C0 (Rn ).18
• The completion of Cc (Rn ) in the norm k · k2 is L2 (Rn ), defined in two steps.
Definition 2.8 The space L2 (Rn ) consists of all functions f : Rn → C for which
there exists a Cauchy sequence (fn ) in Cc (Rn ) with respect to k·k2 such that fn (x) →
f (x) for all x ∈ Rn \N , where N ⊂ Rn is a set of (Lebesgue) measure zero.19
We can extend the inner product on Cc (Rn ) to L2 (Rn ) by (f, g) = limn→∞ (fn , gn ),
where (fn ) and (gn ) are Cauchy sequences in L2 (Rn ) w.r.t. the L2 -norm. However,
this sesquilinear form fails to be positive definite (take a function f on Rn that is
nonzero in finitely—or even countably—many points). To resolve this, introduce
L2 (Rn ) := L2 (Rn )/N , (2.25)
where
N := {f ∈ L2 (Rn ) | kf k2 = 0}. (2.26)
Using measure theory, it can be shown that f ∈ N iff f (x) = 0 for all x ∈ Rn \N ,
where N ⊂ Rn is some set of measure zero. If f is continuous, this implies that
f (x) = 0 for all x ∈ Rn . It is clear that k · k2 descends to a norm on L2 (Rn ) by
k[f ]k2 := kf k2 , (2.27)
where [f ] is the equivalence class of f ∈ L2 (Rn ) in the quotient space. However, we
normally work with L2 (Rn ) and regard elements of L2 (Rn ) as functions instead of
equivalence classes thereof. So in what follows we should often write [f ] ∈ L2 (Rn )
instead of f ∈ L2 (Rn ), which really means f ∈ L2 (Rn ), but who cares . . .
17
The support of a function is defined as the smallest closed set outside which it vanishes.
18
This is the space of all continuous functions f : Rn → C that vanish at infinity in the sense
that for each  > 0 there is a compact subset K ⊂ Rn such that |f (x)| <  for all x outside K.
19
A subset N ⊂ Rn has measure zero if P for any  > 0 therePexists a covering of N by an at
most countable set (In ) of intervals for which n |In | <Q, where n |In | is the sum of the volumes
n
of the In . (Here an interval in Rn is a set of the form k=1 [ak , bk ]). For example, any countable
n
subset of R has measure zero, but there are many, many others.
2 REVIEW OF HILBERT SPACES 14

2.6 Measure theory and Hilbert space

The construction of L2 (Rn ) may be generalized to Hilbert spaces L2 (X, µ) defined for
arbitrary locally compact Hausdorff spaces X; the concept of a measure µ underlying
this generalization is very important also for (commutative) C ∗ -algebras.
Let P (X) be the power set of X, i.e., the set of all subsets of X, and denote
the topology of X (i.e., the set of open subsets of X) by O(X). A σ-algebra on
X is a subset Σ of P (X) such that ∪n An ∈ Σ and ∩n An ∈ Σ whenever An ∈ Σ,
n ∈ N. Note that O(X) is generally not a σ-algebra on X; it is closed under taking
arbitrary unions, but finite intersections only. Let B(X) be the smallest σ-algebra
on X containing O(X); elements of B(X) are called Borel sets in X.
Definition 2.9 A (Radon) measure on X is a map µ : B(X) → [0, ∞] satisfying:
P
1. µ(∪n An ) = n µ(An ) whenever An ∈ B(X), n ∈ N, Ai ∩ Aj = ∅ for all i 6= j;
2. µ(K) < ∞ for each compact subset K of X;
3. µ(A) = sup{µ(K), K ⊂ A, Kcompact} for each A ∈ B(X).
R R
An integral on Cc (X) is a (complex) linear map X : Cc (X) → C such that X f
is in R+ whenever f (x) ∈ R+ for all x ∈ X (in which case we say f ≥ 0).
The Riesz–Markov Theorem states that these concepts are equivalent:
Theorem 2.10 There is a bijective correspondence between integrals and measures:
R
• A measure µ on X defines an integral X dµ on Cc (X), given on f ≥ 0 by
Z Z 
dµ f := sup dµ g | 0 ≤ g ≤ f, gsimple , (2.28)
X X
where a simple function is a finite linear
P combinationRof characteristic
P func-
tions χK , K ⊂ X compact, and if g = i λi χKi , then X dµ g := i λi µ(Ki ).
R
• An integral X on Cc (X) defines a measure µ on X, given on compact K by
Z 
µ(K) = inf f | f ∈ Cc (X), χK ≤ f ≤ 1 . (2.29)
X

For any p > 0, we define Lp (X, µ) as the space of Borel functions20 on X for which
Z 1/p
p
kf kp := dµ |f | < ∞, (2.30)
X

where the integral is defined à la (2.28). The map k · kp : Lp (X, µ) → R+ has a p-

independent null space N , with associated Banach space Lp (X, µ) := Lp (X, µ)/N .
For p = 2, the Banach space L2 (X, µ) is actually a Hilbert space with inner product
Z Z
(f, g) := dµ f g ≡ dµ(x) f (x)g(x), (2.31)
X X

where similarly ambiguous notation has been used as for L2 (Rn ) (cf. the end of §2.5).
20
Here f : X → C is Borel when fi−1 ((s, t)) ∈ B(X) for each 0 ≤ s < t, i = 1, 2, 3, 4, where
f = f1 − f2 + i(f3 − f4 ) is the unique decomposition with fi ≥ 0 (e.g., f1 (x) = max{Re(f (x)), 0}).
2 REVIEW OF HILBERT SPACES 15

2.7 Operators on Hilbert space

An operator a : H1 → H2 between two Hilbert space is simply a linear map (i.e.,
a(λv + µw) = λa(v) + µa(w) for all λ, µ ∈ C and v, w ∈ H1 ). We write av for a(v).
Taking H1 = H1 = H, an operator a : H → H is just called an operator on H.
Taking H1 = H and H2 = C, we obtain a functional on H. For example, any f ∈ H
yields a functional ϕ : H → C by ϕ(g) = (f, g). By Cauchy–Schwarz, |ϕ(g)| ≤ Ckgk
with C = kf k. Conversely, the Riesz–Fischer Theorem states that if some ϕ
satisfies this bound, then it is of the above form, for a unique f ∈ H.
As in real analysis, where one deals with functions f : R → R, it turns out to
be useful to single out functions with good properties, notably continuity. So what
does one mean by a ‘continuous’ operator a : H1 → H2 ? One answer come from
topology: the inner product on a Hilbert space defines a norm, the norm defines a
metric, and finally the metric defines a topology, so one may use the usual definition
of a continuous function f : X → Y between two topological spaces. We use an
equivalent definition, in which continuity is replaced by boundedness:
Definition 2.11 a : H1 → H2 be an operator. Define kak ∈ R+ ∪ {∞} by
kak := sup {kavkH2 , v ∈ H1 , kvkH1 = 1}, (2.32)
p
where kvkH1 = (v, v)H1 , etc. We say that a is bounded when kak < ∞, in which
case the number kak is called the norm of a.
If a is bounded, then it is immediate that
kavkH2 ≤ kak kvkH1 (2.33)
for all v ∈ H1 . This inequality is very important. For example, it implies that
kabk ≤ kakkbk, (2.34)
where a : H → H and b : H → H are any two bounded operators, and ab := a ◦ b,
so that (ab)(v) := a(bv). Eq. (2.33) also implies the easy half of:
Proposition 2.12 An operator on a Hilbert space H is bounded iff it is continuous
in the sense that fn → f implies afn → af for all convergent sequences (fn ) in H.
When H is finite-dimensional, any operator on H is bounded (and may be rep-
resented by a matrix). For an infinite-dimensional example, take H = `2 (S) and
a ∈ `∞ (S), for some set S. It is an exercise to show that if f ∈ `2 (S), then
af ∈ `2 (S). Hence we may define a multiplication operator â : `2 (S) → `2 (S) by
â(f ) := af, (2.35)
that is, (âf )(x) = a(x)f (x). This operator is bounded, with
kâk = kak∞ . (2.36)
Similarly, take H = L2 (Rn ) and a ∈ C0 (Rn ). Once again, (2.35) defines a bounded
multiplication operator â : L2 (Rn ) → L2 (Rn ), satisfying (2.36).
More generally, for locally compact X, a function a ∈ C0 (X) defines a multi-
plication operator â on H = L2 (X, µ) satisfying kâk ≤ kak∞ , with equality iff the
support of the measure µ is X (i.e., every open subset of X has positive measure).
2 REVIEW OF HILBERT SPACES 16

2.8 The adjoint

Let a : H → H be a bounded operator. The inner product on H gives rise to a map
a 7→ a∗ , which is familiar from linear algebra: if a is a matrix (aij ) w.r.t. some o.n.b.,
then a∗ = (aji ). In general, the adjoint a∗ is uniquely defined by the property21

(a∗ f, g) = (f, ag) for all f, g ∈ H. (2.37)

Note that a 7→ a∗ is anti-linear: one has (λa)∗ = λa for λ ∈ C. Also, one has

ka∗ k = kak; (2.38)

ka∗ ak = kak2 . (2.39)

The adjoint allows one to define the following basic classes of bounded operators:
1. n : H → H is normal when n∗ n = nn∗ .
2. a : H → H is self-adjoint when a∗ = a (hence a is normal).
3. a : H → H is positive, written a ≥ 0, when (f, af ) ≥ 0 for all f ∈ H.
4. p : H → H is a projection when p2 = p∗ = p (hence p is positive).
5. u : H → H is unitary when u∗ u = uu∗ = 1 (hence u is normal).
6. v : H → H is an isometry when v ∗ v = 1, and a partial isometry when v ∗ v
is a projection (in which case vv ∗ is automatically a projection, too).

Proposition 2.13 1. An operator a is self-adjoint a iff (f, af ) ∈ R for all f ∈ H

(and hence positive operators are automatically self-adjoint).
2. There is a bijective correspondence p ↔ K between projections p on H and
closed subspaces K Pof H: given p, put K := pH, and given K ⊂ H, define p
on f ∈ H by pf = i (ei , f )ei , where (ei ) is an arbitrary o.n.b. of K.
3. An operator u is unitary iff it is invertible (with u−1 = u∗ ) and preserves the
inner product, i.e., (uf, ug) = (f, g) for all f, g ∈ H.
4. An operator v is a partial isometry iff v is unitary from (ker v)⊥ to ran(v).
5. An operator v is an isometry iff (vf, vg) = (f, g) for all f, g ∈ H.

Similar definitions apply to (bounded) operators between different Hilbert spaces:

e.g., the adjoint a∗ : H2 → H1 of a : H1 → H2 satisfies (a∗ f, g)H1 = (f, ag)H2 for
all f ∈ H2 , g ∈ H1 , and unitarity of u : H1 → H2 means u∗ u = 1H1 and uu∗ = 1H2 ;
equivalently, u is invertible and (uf, ug)H2 = (f, g)H1 for all f, g ∈ H1 (cf. §2.4).
21
To prove existence of a∗ , the Riesz–Fischer Theorem is needed. For fixed a : H → H and
f ∈ H, one defines a functional ϕaf : H → C by ϕaf (g) := (f, ag). By Cauchy–Schwarz and (2.33),
one has |ϕaf (g)| = |(f, ag)| ≤ kf kkagk ≤ kf kkakkgk, so kϕaf k ≤ kf kkak. Hence there exists a
unique h ∈ H such that ϕaf (g) = (h, g) for all g ∈ H. Now, for given a the association f 7→ h is
clearly linear, so that we may define a∗ : H → H by a∗ f := h; eq. (2.37) then trivially follows.
2 REVIEW OF HILBERT SPACES 17

2.9 Spectral theory

The spectrum of an operator a generalizes the range of a (complex-valued) function,
and is its only invariant under unitary transformations a 7→ u∗ au. To get started, we
first restate the spectral theorem of linear algebra. In preparation, we call a family
(pi ) of projections on a Hilbert space H mutually orthogonal if pi H P ⊥ pj H for
i 6= j; this is the case iff pi pj = δij pi . Such a family is called complete
P if i pi f = f
for all f ∈ H; of course, if dim(H) < ∞, this simply means i pi = 1.
Proposition 2.14 Let a : Cn → Cn be a self-adjoint operator on Cn (i.e., an
hermitian matrix). There
P exists a complete family (pi ) of mutually orthogonal pro-
jections so that a = i λi pi , where λi are the eigenvalues of a. Consequently, pi is
the projection onto the eigenspace of a in H with eigenvalue λi , and the dimension
of the subspace pi H is equal to the multiplicity of the eigenvalue λi .
This is no longer true for self-adjoint operators on infinite-dimensional Hilbert
spaces. For example, if a ∈ C0 (R, R), then the associated multiplication opera-
tor â on L2 (R) has no eigenvectors at all! However, is has approximate eigenvectors,
2
in the following sense: for fixed x0 ∈ R, take fn (x) := (n/π)1/4 e−n(x−x0 ) /2 , so that
fn ∈ L2 (R) with kfn k = 1. The sequence fn has no limit in L2 (R).22 Nonetheless,
an elementary computation shows that limn→∞ k(â − λ)fn k = 0 for λ = a(x0 ), so
that the fn form approximate eigenvectors of â with ‘eigenvalue’ a(x0 ).
Definition 2.15 Let a : H → H be a normal operator. The spectrum σ(a) con-
sists of all λ ∈ C for which there exists a sequence (fn ) in H with kfn k = 1 and

lim k(a − λ)fn k = 0. (2.40)

n→∞

1. If λ is an eigenvalue of a, in that af = λf for some f ∈ H with kf k = 1, then

we say that λ ∈ σ(a) lies in the discrete spectrum σd (a) of a.
2. If λ ∈ σ(a) but λ ∈
/ σd (a), it lies in the continuous spectrum σc (a) of a.
3. Thus σ(a) = σd (a) ∪ σc (a) is the union of the discrete and the continuous part.
Indeed, in the first case (2.40) clearly holds for the constant sequence fn = f (for
all n), whereas in the second case λ by definition has no associated eigenvector.
If a acts on a finite-dimensional Hilbert space, then σ(a) = σd (a) consists of
the eigenvalues of a. On the other hand, in the above example of a multiplication
operator â on L2 (R) we have σ(â) = σc (â). Our little computation shows that σc (â)
contains the range ran(a) of the function a ∈ C0 (R), and it can be shown that
σ(â) = ran(a)− (i.e., the topological closure of the range of a : R → R as a subset
of R). In general, the spectrum may have both a discrete and a continuous part.23
22
It converges to Dirac’s delta function δ(x − x0 ) in a ‘weak’ sense, viz. limn→∞ (fn , g) = g(x0 )
for each fixed g ∈ Cc∞ (R), but the δ ‘function’ is not an element of L2 (R) (it is a distribution).
23
If a is the Hamiltonian of a quantum-mechanical system, the eigenvectors corresponding to
the discrete spectrum are bound states, whereas those related to the continuous spectrum form
wavepackets defining scattering states. Just think of the hydrogen atom. It should be mentioned
that such Hamiltonians are typically unbounded operators.
2 REVIEW OF HILBERT SPACES 18

2.10 Compact operators

Even if H is infinite-dimensional, there is a class of operators whose spectrum is
discrete. First, a finite-rank operator is an operator with finite-dimensional
range. Using Dirac’s notation, for f, g ∈ H we write |f ihg| for the operator h 7→
(g, h)f . An important special case is g = f with kf k = 1, so that |f ihf | is the
one-dimensional projection onto the subspace spanned by Pf . More generally, if (ei )
is an o.n.b. of some finite-dimensional subspace
P K, then i |ei ihei | is the projection
onto K. Any finite linear combination i |fi ihgi | is finite-rank, and vice versa.
Definition 2.16 A bounded operator Hilbert space is called compact iff it is the
norm-limit of a sequence of finite-rank operators.
Note that multiplication operators of the type â on L2 (Rn ) for 0 6= a ∈ C0 (Rn )
are never compact. On the other hand, typical examples
R n of compact operators
2 n
on L (R ) are integral operators of the kind af (x) = d y K(x, y)f (y) with K ∈
L2 (R2n ).
Theorem 2.17 Let a be a self-adjoint compact operator on a Hilbert space H. Then
the spectrum σ(a) is discrete. All nonzero eigenvalues have finite multiplicity, so
that only λ = 0 may have infinite multiplicity (if it occurs), and in addition 0 is
the only possible accumulation point of σ(a) = σd (a). If pP i is the projection onto
the eigenspace corresponding to eigenvalue λi ,P then a = i λi pi , where the sum
converges strongly, i.e., in the sense that af = i λi pi f for each fixed f ∈ H.
The compact operators are closed under multiplication and taking adjoints, so that,
in particular, a∗ a is compact whenever a is. Hence Theorem 2.17 applies to a∗ a.
Note that a∗ a is self-adjoint and that its eigenvalues are automatically non-negative.
say that a compact operator a : H → H is trace-class if the
Definition 2.18 We P
√
trace-norm kak1 := k µk is finite, where the µk are the eigenvalues of a∗ a.
Theorem 2.19 Suppose a is trace-class. Then the trace of a, defined by
X
Tr (a) := (ei , aei ), (2.41)
i

is absolutely convergent and independent of the

P orthonormal basis (ei ). In particular,
∗
if a = a with eigenvalues (λi ), then Tr a = i λi . Furthermore:
1. If b is bounded and a is trace-class, then ab and ba are trace-class, with
Tr (ab) = Tr (ba). (2.42)

2. If u is unitary and a is trace-class, then uau−1 is trace-class, with

Tr (uau−1 ) = Tr (a). (2.43)

The following notion plays a fundamental role in von Neumann algebra theory:
Definition 2.20 A trace-class operator ρ : H → H is called a density
P operator if
ρ is positive and Tr (ρ) = 1 (so that kρk1 = 1).
P Equivalently, ρ = i λi pi (strongly)
with dim(pi ) < ∞ for all i, 0 < λi ≤ 1, and i λi = 1.
3 C*-ALGEBRAS 19

3 C*-algebras
3.1 Basic definitions
If a and b are bounded operators on H, then so is their sum a + b, defined by
(a + b)(v) = av + bv, and their product ab, given by (ab)(v) = a(b(v)). This
follows from the triangle inequality for the norm and from (2.34), respectively. Also,
homogeneity of the norm yields that ta is bounded for any t ∈ C. Consequently, the
set B(H) of all bounded operators on a Hilbert space H forms an algebra over the
complex numbers, having remarkable properties. To begin with (cf. (2.32)):
Proposition 3.1 The space B(H) of all bounded operators on a Hilbert space H is
a Banach space in the operator norm
kak := sup {kaf kH , f ∈ H, kf kH = 1}. (3.1)
This is a basic result from functional analysis; it even holds if H is a Banach space.
Definition 3.2 A Banach algebra is a Banach space A that is simultaneously an
algebra in which kabk ≤ kak kbk for all a, b ∈ A.
According to (2.34), we see that B(H) is not just a Banach space but even a Banach
algebra. Also this would still be the case if H were merely a Banach space, but the
fact that it is a Hilbert space gives a crucial further ingredient of the algebra B(H).
Definition 3.3 1. An involution on an algebra A is a real-linear map A → A∗
such that a∗∗ = a, (ab)∗ = b∗ a∗ , and (λa)∗ = λa∗ for all a, b ∈ A and λ ∈ C.
An algebra with involution is also called a ∗ -algebra.
2. A C ∗ -algebra is a Banach algebra A with involution in which for all a ∈ A,
ka∗ ak = kak2 . (3.2)

3. A homomorphism between C ∗ -algebras A en B is a linear map ϕ : A → B

that satisfies ϕ(ab) = ϕ(a)ϕ(b) and ϕ(a∗ ) = ϕ(a)∗ for all a ∈ A, b ∈ B.
4. An isomorphism between two C ∗ -algebras is an invertible homomorphism.24
In view of (2.39), we conclude that B(H) is a C ∗ -algebra (with the identity operator
as its unit) with respect to the involution defined by the operator adjoint (2.37).
Similarly, if A ⊂ B(H) is a norm-closed subalgebra of B(H) such that if a ∈ A,
then a∗ ∈ A (so that A is an algebra with involution), then A is obviously a C ∗ -
algebra (not necessarily with unit). A case in point is A = K(H), the C ∗ -algebra of
compact operators on H. If dim(H) = ∞, there is a strict inclusion K(H) ⊂ B(H);
for one thing, the unit operator lies in B(H) but not in K(H), which has no unit.
If dim(H) < ∞, though, one has K(H) = B(H) = Mn (C), the n × n matrices.
On the other hand, the set B1 (H) of trace-class operators satisfies (3.2) in the
operator norm (3.1) but fails to be complete in that norm, whereas in the trace-norm
k · k1 it is complete but (3.2) fails. Either way, B1 (H) fails to be a C ∗ -algebra.
24
We will shortly prove that an isomorphism is automatically isometric.
3 C*-ALGEBRAS 20

3.2 Commutative C ∗ -algebras

The C ∗ -algebras K(H) and B(H) are highly noncommutative. For the opposite
case, let X be a locally compact Hausdorff space (physicists may keep X = Rn in
mind). The space C0 (X) of all continuous functions f : X → C that vanish at
infinity25 is an algebra under pointwise operations.26 It has a natural involution
f ∗ (x) = f (x), (3.3)
and a natural supremum-norm or sup-norm given by (cf. 2.22)
kf k∞ := sup{|f (x)|, x ∈ X}. (3.4)
Then C0 (X) is a commutative C ∗ -algebra; the axioms are easily checked. Let us note
that C0 (X) has a unit (namely the function equal to 1 for any x) iff X is compact.
The converse, due to Gelfand and Naimark (1943), is a fundamental result:
Theorem 3.4 Every commutative C ∗ -algebra A is isomorphic to C0 (X) for some
locally compact Hausdorff space X, which is unique up to homeomorphism.
This space X is often denoted by Σ(A) and is called the Gelfand spectrum of
A. It may be realized as the set of all nonzero linear maps ω : A → C that satisfy
ω(ab) = ω(a)ω(b) (i.e., of nonzero homomorphisms A → C as C ∗ -algebras).27 The
Gelfand transform maps each a ∈ A to a complex-valued function â on Σ(A) by
â(ω) := ω(a) (a ∈ A, ω ∈ Σ(A)). (3.5)
The Gelfand topology is the weakest topology on Σ(A) making all functions â
continuous (i.e., the topology generated by the sets â−1 (U ), U ∈ C open, a ∈ A). In
this topology, Σ(A) is compact iff 1 ∈ A (exercise), and locally compact otherwise
(later).28 The isomorphism A → C0 (Σ(A)) is just given by the Gelfand transform.
It is immediate from the definition of Σ(A) that a 7→ â is an algebra homomor-
phism; the proof that ω(a∗ ) = ω(a), and hence that ab∗ = (â)∗ , is an exercise. It
If 1 ∈ A, injectivity of the Gelfand transform as a map from A to C(Σ(A)) results
from the difficult fact (proved in §3.4 below) that it is isometric, i.e., kâk∞ = kak.
Surjectivity then easily follows from the Stone–Weierstrass Theorem (exercise).
The hard part of the proof of Theorem 3.4, i.e., the isometry of the Gelfand
transform, may be approached in two rather different ways. One, going back to
Gelfand himself, heavily relies on the theory of (maximal) ideals in Banach algebras.
The other, pioneered by Kadison and Segal, uses the state space (and especially the
pure state space) of A in a central way. Since the same technique also applies to the
proof of the second great theorem about C*-algebras (and also because it is closer to
quantum-mechanical thinking), in these notes we favour the state space approach.
25
I.e., for each  > 0 there is a compact subset K ⊂ X such that |f (x)| <  for all x outside K.
26
Addition is given by (f + g)(x) = f (x) + g(x), multiplication is (f g)(x) = f (x)g(x), etc.
27
For example, if A is already given as A = C(X), then each x ∈ X defines a functional ωx on
A by ωx (f ) = f (x), which is multiplicative by the pointwise definition of multiplication in A.
28
In §3.3 we show that Σ(A) ⊂ A∗ , where A∗ is the dual of A, but the Gelfand topology on Σ(A)
(in which Theorem 3.4 holds) is not the norm-topology but the (relative) weak∗ topology. From
that perspective, continuity of â follows from basic functional analysis: a 7→ â maps A into the
double dual A∗∗ of A, and A ⊂ A∗∗ precisely consists of all w∗ -continuous functionals on on A∗ .
3 C*-ALGEBRAS 21

3.3 States
The concept of a state originates with quantum physics, but also purely mathemat-
ically it came to play a dominant (and beautiful) role in operator algebra theory.
Definition 3.5 A state on a unital C ∗ -algebra A is a linear map ω : A → C that
is positive, in that ω(a∗ a) ≥ 0 for all a ∈ A, and normalized, in that ω(1) = 1.
If we define the dual A∗ of A as the space of linear maps ϕ : A → C for which
kϕk = sup{|ϕ(a)|, a ∈ A, kak = 1} (3.6)
is finite (cf. (2.32)), then it can be shown that any state ω on A lies in A∗ , with
kωk = 1. This leads to an extension of Definition 3.5 to general (i.e., not necessarily
unital) C ∗ -algebras: a state on a C ∗ -algebra A is a functional ω : A → C that is
positive and normalized in the sense that kωk = 1. This implies ω(1) = 1 whenever
A does have a unit, so that the two definitions are consistent when they overlap.
The state space S(A) of A (i.e., the set of all states on A) is a convex set: if
ω1 and ω2 are states, then so is λω1 + (1 − λ)ω2 for any λ ∈ [0, 1]. It follows that
if
P(ω1 , ω2 , . . . , ωn P
) are states, and (λ1 , λ2 , . . . , λn ) are numbers in [0, 1] such that
λ
i i = 1, then i λi ωi is a state. This extends to infinite sums if we equip S(A)
with the weak topology inherited from A∗ (in which ωn → ω if ωn (a) → ω(a) for
∗

each a ∈ A). If A has a unit, then S(A) is a compact convex set in this topology.29
Definition 3.6 A state ω is pure if ω = λ ω1 + (1 − λ)ω2 for some λ ∈ (0, 1) and
certain states ω1 and ω2 implies ω1 = ω2 . The pure states on A comprise the pure
state space of A, denoted by P (A) or ∂S(A). If a state is not pure, it is mixed.30
The convex structure of the state space is nicely displayed in the noncommutative
case by A = M2 (C), the C ∗ -algebra of 2 × 2 complex matrices. Put
 
1 + z x + iy
ρ = 12 ; (3.7)
x − iy 1 − z
then ρ is a density matrix on C2 iff (x, y, z) ∈ R3 with x2 + y 2 + z 2 ≤ 1; this set is
the three-ball B 3 in R3 . It is easy to see that ρ defines a state ωρ on the M2 (C) by
ωρ (a) = Tr (ρa). (3.8)
Conversely, every state on M2 (C) is of this form, so that the state space S(M2 (C))
is isomorphic (as a convex set) to B 3 . The pure states ∂B 3 then correspond to the
two-sphere S 2 = {(x, y, z) ∈ R2 | x2 + y 2 + z 2 = 1} (see exercises).
In the commutative case, we have the key behind the proof of Theorem 3.4:
Lemma 3.7 The pure state space P (A) ⊂ S(A) ⊂ A∗ of a commutative C ∗ -algebra
A coincides with its Gelfand spectrum Σ(A) (seen as a subspace of A∗ ).
The proof is an exercise, but one can see the point from the example A = C(X).
29
This follows from the Banach–Alaoglu Theorem of functional analysis; see [8] or exercises.
30
The Krein–Milman Theorem of functional analysis [8] guarantees the abundance of pure states
in compact convex sets in that any state is a convex sumof pure states (or limit thereof).
3 C*-ALGEBRAS 22

3.4 Spectrum
To prove isometry of the Gelfand transform (and hence Theorem 3.4) from Lemma
3.7, we need a nice result with an ugly proof based on the Axiom of Choice (ac):
Lemma 3.8 Let A be a C ∗ -algebra with unit. For any self-adjoint a ∈ A, there is
a pure state ω0 ∈ P (A) such that |ω0 (a)| = kak.
This will be proved in a minute; for now, we just point out that for A = C(X),
X compact, this is immediate from Weierstrass’ Theorem stating that a continuous
function on a compact set assumes its maximum (and its minimum). Given Lemma
3.8, if a∗ = a, then kak = |ω0 (a)| = |â(ω0 )| ≤ kâk∞ ≤ kak, the last inequality
arising because kâk∞ = sup{|â(ω)|, ω ∈ Σ(A)}, (3.5), and |ω(a)| ≤ kak (since ω is
a state). Hence kâk∞ = kak for self-adjoint a, and therefore for any a (exercise).
The proof of Theorem 3.4 is now complete up to Lemma 3.8. To prove the latter,
and for many other reasons, we introduce the following extremely important notion.
Definition 3.9 Let A be a Banach algebra with unit. The spectrum σ(a) of a ∈ A
is the set of all z ∈ C for which a − z ≡ a − z · 1 has no (two-sided) inverse in A.
The spectral radius r(a) of a ∈ A is defined as r(a) := sup{|z|, z ∈ σ(a)}.
We quote two basic results from functional analysis [8]:31
Proposition 3.10 1. The spectrum σ(a) is a nonempty compact subset of C.

2. The spectral radius is given by r(a) = limn→∞ kan k1/n .

If A is a C ∗ -algebra, then ka2 k = kak2 for self-adjoint a, so the second property
implies kak = r(a) whenever a∗ = a. The first part ensures the existence of λ0 ∈ σ(a)
with |λ0 | = r(a), so jointly we have kak = |λ0 | for some λ0 ∈ σ(a). Furthermore:
Lemma 3.11 Let A be a C ∗ -algebra with unit. For any a ∈ A and λ ∈ σ(a), there
is a pure state ω ∈ P (A) such that ω(a) = λ.
Choosing λ = λ0 as above immediately yields Lemma 3.8. The existence of a
general state ω ∈ S(A) achieving ω(a) = λ is an exercise, based on the Hahn–
Banach Theorem of functional analysis (which in turn relies on ac).32 Furthermore,
a clever use of the Krein–Milman Theorem (which once again relies on ac) shows
that ω may be chosen in P (A); see [16, Corollary 1.4.4, Theorem 4.3.8].
We have now completed the proof of Theorem 3.4, up to the uniqueness of X
(up to homeomorphism).33 This is easily seen to be equivalent to the property

Σ(C(X)) ∼
= X, (3.9)

where the pertinent map X → Σ(C(X)) is given by x 7→ ωx , ωx (f ) = f (x). The

proof is a highly nontrivial but dull exercise in topology; see [8, Theorem vii.8.7].
31
For completeness’ sake, the proofs are given in the appendix below.
32
Gelfand and Naimark’s proof of Theorem 3.4 also relies on ac through a Zorn’s Lemma argu-
ment finding some maximal ideal. A constructive version of Theorem 3.4 exists; see [9].
33
The nonunital case follows from the unital one, using a technique called unitization. See §3.6.
3 C*-ALGEBRAS 23

Exercises for Lecture 2

Unless stated otherwise, A is a commutative C ∗ -algebra with unit. To do nos. 4 and 7 in the
simplest way, you may use a few facts about spectral theory and about positive operators we will
prove later on. First, σ(µ · 1 + λa) = µ + λσ(a) ≡ {µ + λz : z ∈ σ(a)} for all a ∈ A and µ, λ ∈ C.
Second, a ∈ A is positive (in that a = b∗ b for some b ∈ A) iff a is self-adjoint with σ(a) ⊂ [0, ∞). In
particular, a functional ω : A → C is positive iff ω(a) ≥ 0 for all self-adjoint a with σ(a) ⊂ [0, ∞).
We use the notation a ≥ 0 for positive a and write a ≥ b if a − b is positive. Similarly, we write
ω ≥ τ for functionals ω and τ such that ω − τ is positive.
1. Show that if ω ∈ Σ(A), then ω(a∗ ) = ω(a) for all a ∈ A.
2. Given that kâk∞ = kak for self-adjoint a ∈ A, prove the same equality for general a ∈ A.
3. For a ∈ C(X), X compact Hausdorff, show that σ(a) = {a(x), x ∈ X}.
4. Show that Σ(A) = P (A), as follows:
(a) Prove that each ω ∈ Σ(A) is a state, and use the Cauchy-Schwarz inequality for an
appropriate semi-inner product to show that ω is pure. Hint: Start by showing that
ω = λω1 + (1 − λ)ω2 with λ 6= 0, 1 implies that ω1 (a) = ω2 (a) for self-adjoint a.
(b) Prove for all self adjoint a ∈ A that there is a positive scalar λ such that a + λ and
λ − a are positive in A.
(c) Let ω be a pure state. Prove that if τ : A → C is a functional such that 0 ≤ τ ≤ ω,
then we can find a scalar β such that τ = βω.
(d) For 1 ≥ b ≥ 0 use ω0 (a) := ω(ab) to show that

ω(ab) = ω(a)ω(b). (3.10)

(e) Finally, prove that (3.10) holds for general b ∈ A.

5. Using isometry, prove surjectivity of the Gelfand transform from the Stone-Weierstrass
theorem [8]: Let X be a compact Hausdorff space. Let  be a subalgebra of C(X) (regarded
as a commutative C ∗ -algebra) that:

(a) separates points on X (i.e., if x 6= y there is f ∈ Â such that f (x) 6= f (y));

(b) is closed under complex conjugation (i.e., if f ∈ Â then f ∈ Â);
(c) contains the unit function 1X (where 1X (x) = 1 for all x ∈ X).
Then  is dense in C(X) in the sup-norm.
6. Use the Banach–Alaoglu Theorem [8] to prove that Σ(A) is compact in the Gelfand topology.
You may assume that Σ(A) ⊂ A∗ , so it only remains to be proved that Σ(A) is a closed
subset of the unit ball in A∗ .
7. For an arbitrary C ∗ -algebra with unit, and self-adjoint a ∈ A, show that for each λ ∈ σ(a)
there is a state ω ∈ S(A) for which ω(a) = λ. You may proceed as follows:
(a) Define ω on an appropriate two-dimensional subspace of A and use the Hahn-Banach
Theorem to extend it to all of A.
(b) Let b ∈ A be positive and write ω(b) = α + iβ with α, β ∈ R. Use the spectral radius
to show that α ≥ 0.
(c) Consider the sequence an = a − α + inβ to show that β = 0.
3 C*-ALGEBRAS 24

3.5 Continuous functional calculus

Let A be a C ∗ -algebra with unit, and let a ∈ A be normal (i.e., aa∗ = a∗ a). In that
case, the “C ∗ -algebra C ∗ (a, 1) generated (within A) by a and the unit” is well defined
as the smallest C ∗ -subalgebra of A containing a and 1: for this is simply given by the
norm-closure of all polynomials in a and a∗ . By normality of a, all such polynomials
commute, and by the Banach algebra axiom kabk ≤ kakkbk this commutativity is
preserved by norm-limits of sequences (or even nets) of polynomials. Hence C ∗ (a, 1)
is commutative, and according to Theorem 3.4 we have C ∗ (a, 1) ∼ = C(X) for some
compact Hausdorff space X. What is X explicitly?
Theorem 3.12 Let a ∈ A be normal. Then Σ(C ∗ (a, 1)) is homeomorphic with
σ(a), so that C ∗ (a, 1) ∼
= C(σ(a)). This isomorphism may be chosen such that (and
is uniquely defined if ) a ∈ C ∗ (a, 1) maps to the function idσ(a) : t → t in C(σ(a)).
We prove this in case that a∗ = a; the general case involves some extra notational
complications only. Define a map f 7→ f (a) from C(σ(a)) to C ∗ (a, 1) as follows:
1. Polynomials p are mapped into the corresponding polynomials p(a) ∈ C ∗ (a, 1);34
2. Arbitrary functions f ∈ C(σ(a)) are first approximated in the sup-norm by
polynomials pn , i.e., limn pn = f uniformly on σ(a), upon which f (a) is defined
as the norm-limit of pn (a) in A.35 By construction, f (idσ(a) ) = a.
To show that f 7→ f (a) is an isomorphism, we first note that f (a) = GT−1 ◦ â∗ (f )
(exercise),36 where both the Gelfand transform GT : A → C(Σ(A)) and the map
â : Σ(A) → σ(a) are given by (3.5), in this case with A = C ∗ (a, 1). Note that â is
initially defined as a function from Σ(A) to C, but since ω(a) ∈ σ(a) for all ω ∈ Σ(A)
(exercise), it actually takes values in σ(a) ⊂ C. Since GT is an isomorphism, it
remains to be shown that â∗ : C(σ(a)) → C(Σ(C ∗ (a, 1))) is an isomorphism, which
in turn will be the case iff â is a homeomorphism. This is indeed the case: the inverse
of â : Σ(A) → σ(a), ω 7→ ω(a), is λ 7→ ωλ , where λ ∈ σ(a) and ωλ ∈ Σ(C ∗ (a, 1))
is defined by ωλ (f (a)) = f (λ). Finally, continuity of â and its inverse is an easy
consequence of the definition of the Gelfand topology. 
It immediately follows from Theorem 3.12 that
kf (a)k = kf k∞ ; (3.11)
σ(f (a)) = f (σ(a)). (3.12)
For f = idσ(a) this recovers a result we already knew, viz. r(a) = kak if a∗ = a, and
p
kak = r(a∗ a) (3.13)
in general. This shows that the norm is determined by the spectrum, so that:
Corollary 3.13 The norm in a C ∗ -algebra is unique. That is, given a C ∗ -algebra
A (and especially its norm), there is no other norm in which A is a C ∗ -algebra.
34
That is, if p(t) = n cn tn , then p(a) = n cn an .
P P
35
This procedure is validated by Weierstrass’ Theorem (recall that σ(a) is compact) and the fact
that if also qn → f , then limn qn (a) = limn pn (a) (exercise, based on kf (a)k ≤ kf k∞ ).
36
Here g ∗ : C(Y ) → C(X) is the pullback of a continuous map g : X → Y , i.e., g ∗ ϕ = ϕ ◦ g.
3 C*-ALGEBRAS 25

3.6 C ∗ -algebras without unit

We still need to prove Theorem 3.4 for the nonunital case, and will use this opportu-
nity to introduce a general technique for handling C ∗ -algebras without a unit. When
a Banach algebra A does not contain a unit, we can always add one, as follows. Form
the vector space A ⊕ C, and turn this into an algebra by means of
(a + λ1)(b + µ1) := ab + λb + µa + λµ1, (3.14)
where we have prophetically written a + λ1 for (a, λ), et cetera. In other words, the
number 1 in C is identified with the unit 1. Now, the norm on A ⊕ C defined by
ka + λ1k := kak + |λ| makes A ⊕ C a Banach algebra with unit, but that isn’t the
right one for C ∗ -algebras (exercise). Indeed, when A is a C ∗ -algebra we equip A ⊕ C
with the natural and obvious involution
(a + λ1)∗ := a∗ + λ1, (3.15)
and the correct C ∗ -norm is as follows (by Corollary 3.13, this is the only possibility).
Theorem 3.14 Let A be a C ∗ -algebra without unit.37
1. The map ρ : A → L(A) given by ρ(a)b := ab yields an isomorphism (of Banach
algebras) between A and ρ(A) ⊂ L(A). (N.B. This is true also for unital A.)
2. Define a norm on A ⊕ C by ka + λ1k := kρ(a) + λ1kL(A) , where 1 on the right-
hand side is the unit operator in L(A). With the natural algebraic operations
(given above), this norm turns A ⊕ C into a C ∗ -algebra with unit, called Ȧ.
3. There is an isometric (hence injective) morphism A → Ȧ, such that Ȧ/A ∼ =C
∗ ∗
as C -algebras, and Ȧ is the unique unital C -algebra with this property.
The proof of the first claim is an exercise. It is clear from (3.14) and (3.15) that
the map a + λ1 → ρ(a) + λ1 (where the symbol 1 on the left-hand side is defined
below (3.14), and the 1 on the right-hand side is the unit in L(A)) is a morphism.
Hence the norm defined in claim 2 satisfies (2.39), because the latter is satisfied in
the Banach algebra L(A). Moreover, in order to prove that the norm on Ȧ satisfies
(3.2), by Lemma 3.19 (see exercises) it suffices to prove that for all a ∈ A and λ ∈ C,
kρ(a) + λ1k2 ≤ k(ρ(a) + λ1)∗ (ρ(a) + λ1)k. (3.16)
Indeed, for given c ∈ L(A) and  > 0 there exists a b ∈ A, with kbk = 1, such that
kck2 −  ≤ kc(b)k2 . Applying this with c = ρ(a) + λ1, we infer that for every  > 0,
kρ(a) + λ1k2 −  ≤ k(ρ(a) + λ1)bk2 = kab + λbk2 = k(ab + λb)∗ (ab + λb)k.
Here we used (3.2) in A. The right-hand side may be rearranged as
kρ(b∗ )ρ(a∗ + λ1)ρ(a + λ1)bk ≤ kρ(b∗ )k k(ρ(a) + λ1)∗ (ρ(a) + λ1)k kbk.
Since kρ(b∗ )k = kb∗ k = kbk = 1 by claim 1 and (3.22), and kbk = 1 also in the last
term, the inequality (3.16) follows by letting  → 0. 
37
For any Banach space A, the Banach algebra of all bounded linear maps A → A is called L(A).
3 C*-ALGEBRAS 26

3.7 Commutative C ∗ -algebras without unit

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 compactification Ẋ of a non-compact topological space X is the set Ẋ =
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 ,→ Ẋ is continuous, and any
continuous function f ∈ C0 (X) extends uniquely to a function f ∈ C(Ẋ) satisfying
f (∞) = 0. The space Ẋ is the solution (unique up to homeomorphism) of a so-called
universal problem by Alexandroff’s theorem: If ϕ : X → Y is a map between locally
compact Hausdorff spaces such that Y \f (X) is a point and f is a homeomorphism
onto its image, then there is a unique homeomorphism ψ : Ẋ → Y such that
ϕ = ψ ◦ i. The proof of the following lemmas is an easy exercise.

Lemma 3.15 Let A = C0 (X) for some noncompact locally compact Hausdorff space
X. Then Ȧ ∼ = C(Ẋ), where 1 ∈ Ȧ is identified with the constant function 1Ẋ
in C(Ẋ). Conversely, removing C1Ẋ from C(Ẋ) corresponds to removing C from
Ȧ = A ⊕ C (as a vector space), leaving one with C0 (X).

Hence the unitization of C0 (X) corresponds to the one-point compactification of X.

Lemma 3.16 Let A be a commutative C ∗ -algebra without unit.38

1. Each ω ∈ Σ(A) extends to a character ω̇ on Ȧ by

ω̇(a + λ1) := ω(a) + λ. (3.17)

2. The functional ω∞ on Ȧ, defined by

ω∞ (a + λ1) := λ, (3.18)

is a character of Ȧ.

3. There are no other characters on Ȧ.

4. Σ(Ȧ) is homeomorphic to the one-point compactification of Σ(A).

We may now prove Theorem 3.4 also in the nonunital case. Applying the unital
case of Theorem 3.4 to Ȧ and using Lemma 3.16, one finds Ȧ ∼ = C(Ẋ) with X :=
Σ(A). Formally, we now use a little lemma stating that if A and B are C ∗ -algebras
without unit, then Ȧ ∼= Ḃ iff A ∼
= B. Informally: removing C from Ȧ = A ⊕ C
precisely leaves one with C0 (X) by Lemma 3.15, so that finally A ∼
= C0 (X). 

Note that the Gelfand transform on a commutative C ∗ -algebra without unit

indeed takes values in C0 (Σ(A)), since by (3.18) one has â(ω∞ ) = ω∞ (a + 0) = 0 for
the (unique continuous) extension of â from Σ(A) to its one-point compactification.
38
In fact, this lemma is true for any commutative Banach algebra, with respect to any unitization.
3 C*-ALGEBRAS 27

3.8 Commutative harmonic analysis

One of the most beautiful applications of Theorem 3.4 is to commutative harmonic
analysis. Let G be a locally compact abelian group (think of G = R, G = Z, or G =
T if you like), with HaarR measure dx (think of Lebesgue measure on R,R the counting
R 2π dθ
f (eiθ )).
P
measure on Z, so that Z dx f (x) = n∈Z f (n), whilst T dx f (x) = 0 2π
−1
For f, g ∈ Cc (G), the convolution product f ∗ g is defined by (x − y ≡ xy )
Z
f ∗ g(x) := dy f (x − y)g(y). (3.19)
G

Using the invariance of Haar measure, it is trivial to verify that this product is com-
mutative if G is abelian. We also define an involution on Cc (G) by f ∗ (x) = f (−x),
where −x ≡ x−1 .39 Of course, we would now like to turn Cc (G) into a commutative
C ∗ -algebra, but the most obvious norms like the Lp -ones do not accomplish this.
Instead, for f ∈ Cc (G) we define an operator π(f ) on the Hilbert space L2 (G)
by π(f )ψ = f ∗ ψ; here we initially pick ψ ∈ Cc (G) and show that
Z
kπ(f )k ≤ kf k1 := dx |f (x)|, (3.20)
G

so that π(f ) is bounded and may be extended to all of L2 (G) by continuity. As-
sociativity of convolution then implies π(f ∗ g) = π(f )π(g), and also one has
π(f ∗ ) = π(f )∗ . The map f 7→ π(f ) from Cc (G) to B(L2 (G)) is injective (exer-
cise), so that kf k = kπ(f )k defines a norm on Cc (G). One immediately sees that
the axioms (1.7) and (1.8) are satisfied, so that the completion of Cc (G) in this
norm, called C ∗ (G), is a commutative C ∗ -algebra.40 What is its Gelfand spectrum?
Recall that, for any locally compact abelian group G, the dual group or character
group Ĝ is defined as Ĝ = Hom(G, T), i.e., the continuous group homomorphisms
from G to T, equipped with the compact-open topology.41 For example, for G = R
we have Ĝ ∼= R, where p ∈ R defines a character χp ∈ R̂ by χp (x) = exp(ipx). On
the other hand, for G = T one finds Ĝ ∼= Z, where n ∈ Z defines χn (z) = z n , z ∈ T.
Theorem 3.17 Let G be a locally compact abelian group. The Gelfand spectrum
Σ(C ∗ (G)) is homeomorphic to Ĝ, so that C ∗ (G) ∼
= C0 (Ĝ), and the Gelfand transform
ˆ
f 7→ f implementing this isomorphism coincides with the Fourier transform
Z
fˆ(χ) = dx χ(x)f (x). (3.21)
G

The homeomorphism in question maps χ ∈ Ĝ to ωχ ∈ Σ(C ∗ (G)), given by ωχ (f ) =

fˆ(χ), as in (3.21). We defer a proof of this beautiful theorem to the exercises.
39
This choice, rather than the more natural f ∗ (x) = f (x), is made in order to satisfy the axioms
for an involution with respect to the convolution product (as opposed to the pointwise one).
40
Because of (1.7), commutativity is preserved by the completion procedure.
41
The compact-open topology on Ĝ is the restriction to Hom(G, C) of the topology on C(G, C)
generated by the open sets O(K, U ) = {ϕ ∈ C(G, C) | ϕ(K) ⊂ U, K ⊂ G compact, U ⊂ C open}.
In general, G is compact iff Ĝ is discrete, as exemplified by G = T and Ĝ = Z. The space Ĝ is
itself a locally compact abelian group under pointwise multiplication, and the famous Pontryagin
ˆ
Duality Theorem states that Ĝ ∼ = G. We will not need this group structure, though.
3 C*-ALGEBRAS 28

Exercises for Lecture 3

1. Prove the following theorem.

Theorem 3.18 Let a be a normal element of a unital C ∗ -algebra A. Then

the spectrum of a in A coincides with the spectrum of a in C ∗ (a, 1), so that we
may unambiguously speak of the spectrum σ(a).

2. Let A be a Banach algebra without unit. Show that A ⊕ C is a Banach algebra

with unit in the norm ka + λ1k := kak + |λ|. Give an example where A is a
C ∗ -algebra showing that A ⊕ C is not a C ∗ -algebra in this norm.

3. Prove Part 1 of Theorem 3.14.

4. Prove:

Lemma 3.19 (a) If an involution a 7→ a∗ on a Banach algebra A satisfies

the inequality kak2 ≤ ka∗ ak, then kak2 = ka∗ ak and hence A is a C ∗ -
algebra.
(b) For any element a of a C ∗ -algebra one has

ka∗ k = kak. (3.22)

5. Prove Lemma 3.15.

6. Prove Lemma 3.16.

7. Prove Theorem 3.17 by showing that:

(a) Σ(C ∗ (G)) = Σ(L1 (G)), where the Banach algebra L1 (G) is the comple-
tion of Cc (G) in the L1 -norm k · k1 .
(b) Σ(L1 (G)) ∼ = Ĝ, first by proving that each ωχ is a character of L1 (G),
χ ∈ Ĝ, and secondly that there are no others.
3 C*-ALGEBRAS 29

3.9 The structure of C ∗ -algebras

Having understood the structure of commutative C ∗ -algebras, we now turn to the
general case. We already know that the algebra B(H) of all bounded operators
on some Hilbert space H is a C ∗ -algebra in the obvious way (i.e., the algebraic
operations are the natural ones, the involution is the operator adjoint a 7→ a∗ , and
the norm is the operator norm of Banach space theory). Moreover, each (operator)
norm-closed ∗ -algebra in B(H) is a C ∗ -algebra. Our goal is to prove the converse:
Theorem 3.20 Each C ∗ -algebra A is isomorphic to a norm-closed ∗ -algebra in
B(H), for some Hilbert space H. In other words, for any C ∗ -algebra A there exist
a Hilbert space H and an injective homomorphism π : A → B(H).
A homomorphism π : A → B(H) is often called a representation of A on H. The
equivalence between the two statements in the theorem is a consequence of:
Proposition 3.21 Let ϕ : A → B be a nonzero morphism between C ∗ -algebras.
1. ϕ is continuous, with norm ≤ 1 (i.e., kϕ(a)kB ≤ kakA for all a ∈ A);
2. If ϕ is injective, then it is isometric (i.e., kϕ(a)kB = kakA for all a ∈ A).
The proof is an exercise. Let us note that Theorems 3.4 and 3.20 are compatible,
in that any measure µ on X satisfying µ(U ) > 0 for each open U ⊂ X leads to an
injective representation of C0 (X) on L2 (X, µ) by multiplication operators, that is,
π(f )ψ(x) = f (x)ψ(x), where f ∈ C0 (X) and ψ ∈ L2 (X, µ). See [8, §VIII.5].
The proof of Theorem 3.20 uses the beautiful gns-construction,42 which is
important in its own right. We assume that A is unital (and return to the non-
unital case at the end). First, we call a representation π cyclic if its carrier space
H contains a cyclic vector Ω for π; this means that the closure of π(A)Ω coincides
with H. Such representation are the building blocks of any representation.43
Theorem 3.22 Let ω be a state on a C ∗ -algebra A. There exists a cyclic represen-
tation πω of A on a Hilbert space Hω with cyclic unit vector Ωω such that
ω(a) = (Ωω , πω (a)Ωω ) ∀a ∈ A. (3.23)
We first give the idea of the proof in the special case that ω(a∗ a) > 0 for any
a 6= 0. Define a sesquilinear form (−, −) on A by (a, b) := ω(a∗ b). This form
is positive definite by assumption, so that we may complete A to a Hilbert space
called Hω . For each a ∈ A we then define a map πω (a) : A → A by πω (a)b = ab.
Regarding A as a dense subspace of Hω , this defines an operator πω (a) on a dense
domain in Hω . This operator turns out to be bounded, so that it may be extended
from A to Hω by continuity and we obtain a map πω : A → B(Hω ). Trivial
computations show that πω is a representation. The special vector Ωω is simply
1 ∈ A, seen as an element of Hω . Indeed, kΩω k2 = (Ωω , Ωω ) = ω(1∗ 1) = ω(1) = 1
and (Ωω , πω (a)Ωω ) = ω(1∗ a1) = ω(a). Hence the only difficulty of the proof lies in
the boundedness of πω (a) and in the removal of the assumption ω(a∗ a) > 0.
42
For Gelfand-Naimark-Segal. This construction is very important also in mathematical physics.
43
Any non-degenerate representation π is a direct sum of cyclic representations. Here one says
that a representation π(A) is non-degenerate if π(a)v = 0 for all a ∈ A implies v = 0.
3 C*-ALGEBRAS 30

3.10 Proof of Theorems 3.22 and 3.20

The inequality establishing boundedness of πω (a), or specifically kπω (a)k ≤ kak, is
ω(b∗ a∗ ab) ≤ kak2 ω(b∗ b), (3.24)
which is a transcription of the inequality kπω (a)bk2Hω ≤ kak2A kbk2Hω . We defer the
proof of this to subsection 3.12. Under our standing assumption, i.e., ω(a∗ a) > 0 iff
a 6= 0, this not only proves Theorems 3.22, but also Theorem 3.20: for πω (a) = 0
implies kπω (a)Ωω k2 = 0, whose left-hand side is precisely (Ωω , πω (a∗ a)Ωω ) = ω(a∗ a).
In general, a C ∗ -algebra may lack such states, and we must adapt the proof of
both theorems. The gns-construction is easy: for an arbitrary state ω, we introduce
Nω = {a ∈ A | ω(a∗ a) = 0}, which is a left-ideal in A (i.e. a C ∗ -subalgebra of A such
that ab ∈ Nω whenever a ∈ A and b ∈ Nω ). If pω a is the image of a ∈ A in A/Nω ,
we may define an inner product on the latter by (pω a, pω b) = ω(a∗ b); this is well
defined and positive definite, and we define the Hilbert space Hω as the completion
of A/Nω in this inner product. Furthermore, we define πω (a) : A/Nω → Hω by
πω (a)pω b := pω ab; this is indeed well defined because Nω is a left ideal in A. Finally,
we define Ωω := pω 1. The proof that everything works is then an exercise.
When A has no unit, use the gns-construction for the unitization Ȧ and simply
restrict πω̇ (Ȧ) to A to define πω (A). This completes the proof of Theorem 3.22. 
We now take up the proof of Theorem 3.20. To solve the problem ofL the possible
lack of injectivity of πω , we replace Hω by the crazy Hilbert space Hc = ω∈P (A) Hω ,
where P (A) is the pure state space of A. The elements of this space are sequences
Ψ ≡ (Ψω )ω∈P (A) , such that: (i) Ψω ∈ Hω ; (ii) only countably many vectors Ψω
2
P
are nonzero; and (iiii) ω∈P (A) kΨω kHω < ∞ (note that the sum makes sense!).
Addition and scalar multiplication are defined pointwise, and the inner product is
X
(Ψ, Φ)Hc = (Ψω , Φω )Hω . (3.25)
ω∈P (A)

The crazy space Hc carries a representation π(A) defined by (π(a)Ψ)ω = πω (a)Ψω ,

and the point of all this is that π is injective: if π(a) = 0, then (π(a)Ψ)ω = 0 for
each Ψ ∈ Hc and each ω ∈ P (A), hence also for the pure state ω0 of Lemma
3.8, and for the vector Ψ given by Ψω0 = Ωω0 and Ψω = 0 for all ω 6= ω0 . But
this implies πω0 (a)Ωω0 = 0, so that |(Ωω0 , πω0 (a)Ωω0 )| = |ω0 (a)| = 0. If a∗ = a
this implies kak = 0 by Lemma 3.8, hence a = 0. For general a, we replace a
in the above calculation by a∗ a, which is self-adjoint. This yields the inference
π(a) = 0 ⇒ π(a∗ a) = 0 ⇒ ka∗ ak = 0. The C*-axiom equates ka∗ ak to kak2 , so that
once again kak = 0. It follows that π is injective, and Theorem 3.20 is proved. 
It should be noted that this proof relies on incredible overkill, in that Hc is far
larger than necessary (indeed, in all but the most trivial cases, H is non-separable).
For example, already for A = M2 (C) we have P (A) ∼ = S 2 , so that Hc = ω∈S 2 C2 ;
L
this Hilbert space is non-separable, whereas A has an injective representation on C2 .
More generally, K(H) or B(H) has an injective representation on H by definition,
whereas Hc is non-separable. In the commutative case, A = C0 (X) yields the non-
separable Hc = x∈X C, although A has an injective representation on L2 (X, µ).
L
3 C*-ALGEBRAS 31

3.11 Easy examples of the gns-construction

In the following examples we say that two representations π1 (A) and π2 (A) of a
C ∗ -algebra A on Hilbert spaces H1 and H2 are (unitarily) equivalent (π1 ∼ = π2 ),
∗
when there is a unitary map u : H1 → H2 such that π2 (a) = uπ1 (a)u for all a ∈ A.
The gns-representation is often equivalent to some ‘familiar’ representation:
• Let A = C0 (X). The Riesz–Markov Theorem 2.10 implies that the state space
of A is isomorphic
R to the convex set of (Radon) probability measures µ on X; we
write ωµ (f ) = X dµ f for this correspondence, cf. (2.28). The pure states (which
correspond to point or Dirac measures) take the form ωx (f ) = f (x) for some x ∈ X,
and clearly Nx ≡ Nωx = {f ∈ C0 (X) | f (x) = 0}. It follows that A/Nx ∼ = C
under the unitary map pωx f 7→ f (x), and the corresponding gns-representation is
equivalent to πx (f ) = f (x) on Hx = C (exercise). In the ‘opposite’ case where
µ(U ) > 0 for any open U ⊂ X, we have Nωµ = 0 and the Hilbert space Hωµ ≡ Hµ is
given (‘on the nose’, i.e. literally rather than up to equivalence)
R by the completion
2
L (X, µ) of C0 (X) with respect to the inner product (f, g) = X dµ f g. The corre-
sponding gns-representation πµ is obviously given by multiplication operators, and
the cyclic vector Ωµ is simply the R function identically equal to 1. We verify2 (3.23)
by computing (Ωµ , πµ (f )Ωµ ) = X dµ f = ωµ (f ); for f = 1X this yields kΩµ k = 1.
• For a noncommutative example, take A = Mn (C), with a state necessarily of the
∗
P for some density matrix ρ̂. So Nρ = {a ∈ A | Tr (ρ̂a a) = 0}. If
form ρ(a) = Tr (ρ̂a),
we expand ρ̂ = i λi pi (cf. §2.10), and for simplicity assume that pi = |ei ihei | with
respect to the standard basis (ei ) of Cn , then two cases of special interest arise:
1. If ρ̂ = |ej ihej | is pure, the state is just ρ(a) = (ej , aej ), having null space
Nρ = {a ∈ A | aej = 0}. Hence a ∈ Nρ iff the j’th column Cj (a) of a vanishes, so
that a − b ∈ Nρ iff Cj (a) = Cj (b). Thus the equivalence class pρ a ∈ A/Nρ may be
identified with Cj (a), so that Hρ ≡ A/Nρ ∼ = Cn (with the standard inner product)
under the unitary u : pρ a 7→ Cj (a) from A/Nρ to Cn , with inverse u−1 : z 7→ pρ a,
where aij = zi and aik = 0 for all i and k 6= j (that is, a has Cj (a) = z and zeros
elsewhere). We likewise write u−1 w = pρ b, with bij = wi and bik = 0 for all i and
k 6= j. Hence u(pρ a) = z and u(pρ b) = w, and unitarity follows by computing
X X
(pρ a, pρ b)Hρ = ρ(a∗ b) = aij bij = z i wi = (z, w)Cn = (u(pρ a), u(pρ b))Cn .
i i

(Physicists beware: no sum over j!) The gns-representation, originally given on Hρ

by πρ (a)pρ b = pρ (ab), is transformed to uπρ (a)u−1 ≡ π̂ρ on Cn , which is given by
π̂ρ (a)w = uπρ (a)pρ b = upρ (ab) = Cj (ab) = aw.
The cyclic vector uΩρ = Ω̂ρ in Cn is just the basis vector ej from which we started.
More generally, for a pure state ψ(a) = (Ψ, aΨ) the gns-representation πψ (Mn (C))
induced by ψ is equivalent to the defining representation on Cn , with Ω̂ψ = Ψ.
2. The ‘opposite’ case where ρ(a∗ a) > 0 for all a 6= 0 occurs when λi > 0 for all i.
In that case, Hρ ∼= Mn (C) with inner product (a, b) = Tr (a∗ b), and πρ ∼
= π̂ρ with
1/2
π̂ρ (a)b = ab. The cyclic vector in Mn (C) then becomes Ω̂ρ = ρ (see exercises).
3 C*-ALGEBRAS 32

3.12 Positivity in C ∗ -algebras

To prove the inequality (3.24), as well as some deeper results not contained in this
course, we need to develop the theory of positivity in C ∗ -algebras. This theory will
unify the following special cases: a bounded operator a on a Hilbert space H is
called positive when (v, av) ≥ 0 for all v ∈ H, whilst 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.
Definition 3.23 An element a of a C ∗ -algebra A is positive when a = b∗ b for
some b ∈ A. We then write a ≥ 0, and define a ≥ b for self-adjoint a, b if a − b ≥ 0.
If A is commutative, this clearly reproduces the second definition above, and if
A ⊂ B(H), the equivalence with the first notion is a consequence of the following:
Proposition 3.24 An element a ∈ A is positive iff a∗ = a and σ(a) ⊂ R+ .
The proof is an exercice. Hilbert space theory then equates the characterization of
a ≥ 0 given in this proposition to the condition that (v, av) ≥ 0 for all v ∈ H.
We are now going to prove the inequality (3.24). Here is the first step.
Lemma 3.25 1. If C is a C ∗ -subalgebra of A and a ≤ b in C, then a ≤ b in A.
2. If ϕ : A → B is an isomorhpism, then a ≥ b in A iff ϕ(a) ≥ ϕ(b) in B.
3. If a ≥ d, then b∗ ab ≥ b∗ db for any b ∈ A.
4. If a∗ = a and A is unital, then −kak · 1 ≤ a ≤ kak · 1.
The first three claims are trivial. The fourth is first proved in C(σ(a)), where it
reads −r(a) · 1σ(a) ≤ idσ(a) ≤ r(a) · 1σ(a) (here 1σ(a) is the function t → 1 whilst
idσ(a) : t → t, and we have used kidσ(a) k∞ = r(a)). This obviously holds pointwise,
and hence also in the sense of Definition 3.23. Under the inverse Gelfand transform
C(σ(a)) → C ∗ (a, 1) ⊂ A, i.e., by the continuous functional calculus (see §3.5), the
function 1σ(a) is mapped to 1 ∈ A, whereas idσ(a) is mapped to a ∈ A. In combination
with r(a) = kak and the second part of the lemma, this gives the fourth part. 
For any a, the third and the fourth part give b∗ a∗ ab ≤ kak2 · b∗ b. Since states by
definition preserve inequalities as in Definition 3.23, we finally obtain (3.24). 
Finally, an interesting perspective on positivity is given by the concept of a
convex cone in a real vector space V , which is a subspace V + such that:
(i): if v ∈ V + and t ∈ R+ then tv ∈ V + ; (ii): if v, w ∈ V + then v + w ∈ V + ;
(iii): V + ∩ −V + = 0.
Such a convex cone is equivalent to a linear partial ordering in V , which is a
partial ordering ≤ in which v ≤ w implies v + f ≤ w + f for all f ∈ V and tv ≤ tw
for all t ∈ R+ . Indeed, given V + ⊂ V one defines ≤ by putting v ≤ w if w − v ∈ V + ,
and given ≤ one defines V + = {v ∈ V | 0 ≤ v}. Without proof, we state:
Proposition 3.26 The set A+ of all positive elements of a C ∗ -algebra A is a convex
cone in the real vector space Asa := {a ∈ A | a∗ = a}.
3 C*-ALGEBRAS 33

Exercises for Lecture 4

1. Prove Proposition 3.21 (you may assume that A and B have units).
Hint for part 1: show that σ(ϕ(a)) ⊆ σ(a). Hint for part 2: Assume there is
an b ∈ A for which kϕ(b)k 6= kbk. Show that this implies kϕ(b∗ b)k 6= kb∗ bk.
Put a := b∗ b and show that σ(ϕ(a)) ⊂ σ(a). By Urysohn’s lemma there is a
nonzero f ∈ C(σ(a)) that vanishes on σ(ϕ(a)), so that f (ϕ(a)) = 0. Prove
that ϕ(f (a)) = 0, which contradicts injectivity of ϕ.

2. Fill in the details of the gns-construction for general states ω:

(a) Show that the inner product (pω a, pω b) = ω(a∗ b) on A/Nω is well defined;
(b) Show that the representation πω (a)pω b := pω ab is well defined on A/Nω ;
(c) Show that πω (a) is bounded.

3. Show in detail that the gns-representation of C0 (X) induced by a pure state

ωx (f ) = f (x) for some x ∈ X is equivalent to πx (f ) = f (x) on H = C.

4. Show in detail that the gns-representation

P of Mn (C) induced by a mixed state
ρ(a) = Tr (ρ̂a), where ρ̂ = i λi pi with λi > 0 for all i, is equivalent to the
representation π(a)b = ab on H = Mn (C) with inner product (a, b) = Tr (a∗ b).

5. Prove Proposition 3.24.

6. A representation π of a C ∗ -algebra A on a Hilbert space H is called irre-

ducible if a closed subspace K of H that is stable under π(A) (in the sense
that if ψ ∈ K, then π(a)ψ ∈ K for all a ∈ A) is either H or 0.

(a) Prove that each of the following conditions is equivalent to irreducibility:

• π(A)0 = C · 1 (where S 0 is the commutant of S ⊂ B(H));
• π(A)00 = B(H);
• Every vector in H is cyclic for π(A).
Hint: from the theory of von Neumann algebras, use the fact that if
π(A)0 6= C · 1, then π(A)0 contains a nontrivial projection (i.e. 0 6= p 6= 1).
(b) Prove that if ω is pure, then the gns-representation πω is irreducible.
(c) Prove that if πω is irreducible, then ω is pure (difficult!).
3 C*-ALGEBRAS 34

Appendix 1: proof of Proposition 3.10

The proof of part 1 uses two lemmas. We assume that A is unital.
Pn k −1
Lemma 3.27 When kak < 1 the sum k=0 a converges to (1 − a) . Hence
(a − z1)−1 always exists when |z| > kak, so that

r(a) ≤ kak. (3.26)

We first show that the sum is a Cauchy sequence. Indeed, for n > m one has
n
X m
X n
X n
X n
X
k k k k
k a − a k=k a k ≤ ka k ≤ kakk .
k=0 k=0 k=m+1 k=m+1 k=m+1

For n, m → ∞ this goes to 0 by

Pnthe ktheory of the geometric series. Since A is
complete, the Cauchy sequence k=0 a converges for n → ∞. Now compute
n
X n
X
k
a (1 − a) = (ak − ak+1 ) = 1 − an+1 .
k=0 k=0

Hence n
X
k1 − ak (1 − a)k = kan+1 k ≤ kakn+1 .
k=0
Pn
which → 0 for n → ∞, as kak < 1 by assumption. Thus limn→∞ k=0 ak (1−a) = 1.
By a similar argument,
n
X
lim (1 − a) ak = 1.
n→∞
k=0

so that, by continuity of multiplication in a Banach algebra, one finally has

n
X
lim ak = (1 − a)−1 .
n→∞
k=0

The second claim of the lemma follows because (a − z)−1 = −z −1 (1 − a/z)−1 ,

which exists because ka/zk < 1 when |z| > kak. 

To prove that σ(a) is compact, it remains to be shown that it is closed.

Lemma 3.28 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 kbk < ka−1 k−1 . This implies

ka−1 bk ≤ ka−1 k kbk < 1. (3.27)

Hence a + b = a(1 + a−1 b) has an inverse, namely (1 + a−1 b)−1 a−1 , which exists by
(3.27) and Lemma 3.27. It follows that all c ∈ A for which ka − ck <  lie in G(A),
for  ≤ ka−1 k−1 . 
3 C*-ALGEBRAS 35

To resume the proof of Proposition 3.10.1, given a ∈ A we now define a function

f : C → A by f (z) := z − a. Since kf (z + δ) − f (z)k = δ, we see that f is
continuous (take δ =  in the definition of continuity). Because G(A) is open in A
by Lemma 3.28, it follows from the topological definition 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)) = ρ(a). This set being open, its complement σ(a)
is closed. Finally, define g : ρ(a) → A by g(z) := (z − a)−1 . For fixed z0 ∈ ρ(a),
choose z ∈ C such that |z − z0 | < k(a − z0 )−1 k−1 . From the proof of Lemma
3.28, with a replaced by a − z0 and c replaced by a − z, we see that z ∈ ρ(a), as
ka − z0 − (a − z)k = |z − z0 |. Moreover, the power series
n  k
1 X z0 − z
z0 − a k=0 z0 − a

converges for n → ∞ by Lemma 3.27, because

k(z0 − z)(z0 − a)−1 k = |z0 − z| k(z0 − a)−1 k < 1.

By Lemma 3.27, the limit n → ∞ of this power series is

∞  k   −1
1 X z0 − z 1 z0 − z 1
= 1− = = g(z).
z0 − a k=0 z0 − a z0 − a z0 − a z−a

Hence ∞
X
g(z) = (z0 − z)k (z0 − a)−k−1 (3.28)
k=0

is a norm-convergent power series in z. For z 6= 0 we write kg(z)k = |z|−1 k(1 −

a/z)−1 k and observe that limz→∞ 1 − a/z = 1, since limz→∞ ka/zk = 0. Hence
limz→∞ (1 − a/z)−1 = 1, and
lim kg(z)k = 0. (3.29)
z→∞
∗
Let ρ ∈ A be a functional on A; since ρ is bounded, (3.28) implies that the function
gρ : z → ρ(g(z)) is given by a convergent power series, and (3.29) implies that

lim gρ (z) = 0. (3.30)

z→∞

Now suppose that σ(a) = ∅, so that ρ(a) = C. The function g, and hence gρ , is
then defined on C, where it is analytic and vanishes at infinity. In particular, gρ is
bounded, so that by Liouville’s theorem it must be constant. By (3.30) this constant
is zero, so that g = 0.44 This is absurd, so that ρ(a) 6= C hence σ(a) 6= ∅. This
finishes the proof of Proposition 3.10.1. 
44
This follows by a basic result in Banach spaces B: if v ∈ B is such that ρ(v) = 0 for all ρ ∈ B ∗ ,
then v = 0.
3 C*-ALGEBRAS 36

The proof of Proposition 3.10.2 is as follows. By Lemma 3.27, for |z| > kak the
function g in the proof of Lemma 3.28 has the norm-convergent power series
∞
1 X  a k
g(z) = . (3.31)
z k=0 z
On the other hand, we have seen that for any z ∈ ρ(a) one may find a z0 ∈ ρ(a)
such that the power series (3.28) converges. If |z| > r(a) then z ∈ ρ(a), so (3.28)
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, (3.31) is norm-
convergent for |z| > r(a), uniformly in z. Comparing with (3.26), this sharpens what
we know from Lemma 3.27. The same theory says that (3.31) cannot norm-converge
uniformly in z unless kan k/|z|n < 1 for large enough n. This is true for all z for
which |z| > r(a), so that
lim sup kan k1/n ≤ r(a). (3.32)
n→∞
To derive a second inequality we use the following polynomial spectral mapping
property.
Lemma 3.29 For a polynomial p on C, define p(σ(a)) as {p(z)| z ∈ σ(a)}. Then
p(σ(a)) = σ(p(a)). (3.33)
To prove this equality, choose z, α ∈ C and compare the factorizations
n
Y
p(z) − α = c (z − βi (α));
i=1
n
Y
p(a) − α1 = c (a − βi (α)1). (3.34)
i=1

Here the coefficients c and βi (α) are determined by p and α. When α ∈ ρ(p(a))
then p(a) − α1 is invertible, which implies that all a − βi (α)1 must be invertible.
Hence α ∈ σ(p(a)) implies that at least one of the a − βi (α)1 is not invertible, so
that βi (α) ∈ σ(a) for at least one i. Hence α ∈ p(βi (α)) − α = 0, i.e., α ∈ p(σ(a)).
This proves the inclusion σ(p(a)) ⊆ p(σ(a)).
Conversely, when α ∈ p(σ(a)) then α = p(z) for some z ∈ σ(a), so that for some
i one must have βi (α) = z for this particular z. Hence βi (α) ∈ σ(a), so that a−βi (α)
is not invertible, implying that p(a) − α1 is not invertible, so that α ∈ σ(p(a)). This
shows that p(σ(a)) ⊆ σ(p(a)), and (3.33) follows. 
To conclude the proof of Proposition 3.10.2, we note that since σ(a) is closed
there is an α ∈ σ(a) for which |α| = r(a). Since αn ∈ σ(an ) by Lemma 3.29, one has
|αn | ≤ kan k by (3.26). Hence kan k1/n ≥ |α| = r(a). Combining this with (3.32)
yields
lim sup kan k1/n ≤ r(a) ≤ kan k1/n .
n→∞
Hence the limit must exist, and
lim kan k1/n = inf kan k1/n = r(a). 
n→∞ n
 4 VON NEUMANN ALGEBRAS 37

4 von Neumann algebras

See §1.2 for notation and preliminary remarks on von Neumann algebras. Apart from
von Neumann’s own motivation, later developments brought contact with quantum
field theory [13], quantum statistical mechanics [2], number theory [5, 6], knot the-
ory [15], and algebraic geometry [6].45 In 1982 the Fields Medal was awarded to
Alain Connes for his contributions to the classification of von Neumann algebras
[5], whereas Vaughan Jones received the same prize in 1990 for his new invariant in
knot theory, obtained through a study of von Neumann algebras of type ii1 .

4.1 The Double Commutant Theorem

Operator algebra theory started with the following theorem of von Neumann:46

Theorem 4.1 Let M be a unital ∗ -subalgebra of B(H). Then the following cond-
tions are equivalent (and hence each defines M to be a von Neumann algebra):

(1) M 00 = M ;
(2) M is closed in the weak operator topology;
(3) M is closed in the strong operator topology;
(4) M is closed in the σ-weak operator topology.

The essence of the proof is already contained in the finite-dimensional case H = Cn ,

where the nontrivial claim in Theorem 4.1 is: if M is a unital ∗ -subalgebra of Mn (C),
then M 00 = M . In fact, all we need to prove is M 00 ⊆ M , since the converse inclusion
is trivial. The idea is to take n arbitrary vectors v1 , . . . , vn in H, and, given a ∈ M 00 ,
find some b ∈ M such that avi = bvi for all i = 1, . . . , n. Hence a = b, so a ∈ M .
For fixed v ∈ H, form the linear subspace M v = {mv | m ∈ M } of H, with
associated projection p (i.e. pw = w if w ∈ M v and pw = 0 if w ∈ (M v)⊥ ). Then
p ∈ M 0 (exercise). Hence a ∈ M 00 commutes with p. Since 1 ∈ M , we have v ∈ M v,
so v = pv, and av = apv = pav ∈ M v. Hence av = bv for some b ∈ M .
Now run the same argument with the substitutions H H n = H ⊕· · ·⊕H (with
n terms), M Mn = {diag(m, . . . , m) | m ∈ M }, and v v = ⊕i vi ≡ (v1 , . . . , vn ).
00 00
We then have (Mn ) = (M )n (exercise), so for any matrix a = diag(a, . . . , a) in
(M 00 )n the previous argument yields a matrix b = diag(b, . . . , b) ∈ Mn such that
av = bv. But this is avi = bvi for all i = 1, . . . , n, so that a = b and hence M 00 ⊆ M .
The implication (3) ⇒ (1), then, is easily shown by adapting the above proof
to infinite-dimensional H (exercise). Furthermore, (1) ⇒ (2) ⇒ (3) is quite trivial,
whereas (1) ⇔ (4) may be proved either as in exercise 6 or through the following
functional-analytic result (cf. [33, Vol. I, Thm. ii.2.6] or [28, Thm. 4.6.7]): A linear
subspace of B(H) is strongly closed iff it is weakly closed iff it is σ-weakly closed. 
45
Jakob Lurie (hailed by some as a successor to the algebraic geometer and category theorist
Alexandre Grothendieck) is now teaching a course on von Neumann algebras at Harvard [19].
46
See [23], and compare Proposition 2.5, reformulated as: A subspace K ⊂ H is closed iff
⊥⊥
K = K. In either case, a topological condition is equivalent to an algebraic one.
4 VON NEUMANN ALGEBRAS 38

4.2 From spectral theory to von Neumann algebras

If we combine the continuous functional calculus (i.e. Theorem 3.12) with the Riesz–
Markov Theorem 2.10 from measure theory (see also §3.11), we obtain:
Proposition 4.2 Let a∗ = a ∈ B(H). For each density operator ρ̂ on H (with
associated state ρ) there exists a unique probability measure µρ on σ(a) such that
Z
ρ(f (a)) ≡ Tr (ρ̂f (a)) = dµρ f for all f ∈ C(σ(a)). (4.35)
σ(a)

Since the (Lebesgue) integral appearing here is defined for the far larger class B(σ(a))
of bounded Borel functions on σ(a), we may try to extend the continuous functional
calculus f 7→ f (a) to a “Borel functional calculus.” To do so, we regard B(σ(a)) as
a commutative C ∗ -algebra under pointwise operations and the sup-norm, and intro-
duce the commutative von Neumann (and hence C ∗ ) algebra W ∗ (a) = C ∗ (a, 1)00 .
Theorem 4.3 Let a∗ = a ∈ B(H). The isomorphism C(σ(a)) → C ∗ (a, 1) of Theo-
rem 3.18 has a unique extension to a homomorphism B(σ(a)) → W ∗ (a) that satisfies
kf (a)k ≤ kf k∞ for each f ∈ B(σ(a)), and (4.35) remains valid for such f .
The proof [28, Thm. 4.5.4] is based on the following fact [28, Lemma 3.2.2].
Lemma 4.4 There is a bijective correspondence between bounded sesquilinear forms
Q : H × H → C on a Hilbert space H and bounded operators on H, as follows:
• Any b ∈ B(H) defines a bounded sesquilinear form Qb by Qb (v, w) = (v, bw),
which satisfies the bound |Qb (v, w)| ≤ Ckvkkwk (indeed, take C = kbk).
• Conversely, for any sequilinear form Q that satisfies |Q(v, w)| ≤ Ckvkkwk
there is unique b ∈ B(H) such that Q = Qb , with kbk ≤ C.
It suffices to define Q on the diagonal in H × H, that is, Q(v,
Pw) is determined by
3
Q(v, v) ≡ Q(v) through the polarization identity Q(v, w) = 4 k=0 ik Q(w + ik v).
1

This correspondence applies to (4.35) byR taking the special case ρ̂ = |ΨihΨ|, so
that ρ(b) ≡ ψ(b) = (Ψ, bΨ): if Q(Ψ) = σ(a) dµψ f , then Q = Qf (a) , at least for
f ∈ C(σ(a)). The point now is that for f ∈ B(σ(a)) the (Lebesgue) integral on the
right-hand side of (4.35) remains well defined, yielding a sesquilinear form bounded
by C = kf k∞ . This, then, defines the operator f (a) through Q = Qf (a) .
To prove that f (a) ∈ W ∗ (a), we use a much deeper lemma [28, Prop. 6.2.9]:47
Lemma 4.5 B(σ(a), R) is the bounded monotone sequential completion of C(σ(a), R).
Now define B! = {f ∈ B(σ(a), R) | f (a) ∈ W ∗ (a)}, so C(σ(a), R) ⊆ B! ⊆ B(σ(a), R).
If (fn ) is some bounded monotone sequence in B! with pointwise limit f , then—since
the map f 7→ f (a) is a homomorphism and hence preserves positivity—(fn (a)) is a
bounded monotone sequence in B(H), which strongly converges to f (a) [28, Prop.
4.5.2]. Since W ∗ (a) is closed under strong limits, it follows that B! is monotone
sequentially complete, so that by Lemma 4.5, B! = B(σ(a), R). Complexifying, we
obtain f (a) ∈ W ∗ (a) for all f ∈ B(σ(a)). The rest of the proof is an exercise. 
47
I.e., B(σ(a), R) is the smallest space of bounded real functions on σ(a) that contains C(σ(a), R)
and is closed under bounded pointwise limits of monotone (increasing or decreasing) sequences.
4 VON NEUMANN ALGEBRAS 39

4.3 Projections in von Neumann algebras

By Theorem 4.3, the spectral projection P (∆) ≡ χ∆ (a) is defined for any Borel
subset ∆ ⊂ σ(a) (or even ∆ ⊂ R, in which case P (∆)R is defined as χ∆∩σ(a) ). For
any f ∈ B(σ(a)), we are going to define an operator dP f , in such a way that
Z Z
f (a) = dP f ≡ dP (λ) f (λ). (4.36)
σ(a)
R
The special case f = 1σ(a) gives the resolution of the identity 1H = σ(a) dP (λ),
R
whilst f = idσ(a) yields the spectral decomposition of a, i.e., a = σ(a) dP (λ) λ
(one could equally well integrate over R here). To define the integral, we need a
result like Lemma 4.5, involving thePset E(σ(a), R) of elementary functions on σ(a),
i.e., finite linear combinations f = k ck χBk , with ck ∈ R and Bk ⊂ σ(a) Borel.
Lemma 4.6 B(σ(a), R) is the bounded monotone sequential completion of E(σ(a), R).
Moreover, any f ∈ B(σ(a), R+ ) is a pointwise limit of a bounded (from above) mono-
tone increasing sequence fn in E(σ(a), R+ ), denoted by fn % f .
See e.g. [11, Prop. 4.1.5]. So first assume f ≥ 0, and find (fn ) as in this lemma.
As in the proof of Lemma 4.5, the bounded monotonicity of the sequence (fn ) in
B(σ(a), R) is inherited by the sequence (fn (a)) in B(H), which strongly converges
to f (a). Since the underlying principle will recur, we isolate it [28, Prop. 4.5.2]):
Lemma 4.7 Any monotone increasing (decreasing) sequence (bn ) of self-adjoint op-
erators on H that is bounded from above (below) converges strongly to a unique limit
b = b∗ ∈ B(H). In that case we write bn % b (bn & b).
So the applications of this Lemma we have had so far were all of the following nature:
if fn % f in B(σ(a), R), then fn (a) % f (a) in W ∗ (a) ⊂ B(H). We now use Lemma
4.7 once more to turn (4.36) into a tautology (!), as follows. Initially, assume f ≥ 0.
1. Approximate f by a sequence fn % f , as in Lemma 4.6.
P R P
2. For elementary functions fn = k ck χBk , define dP fn = k ck P (Bk ).
R
3. The ensuing sequence bn = dP fn of self-adjoint operators isRbounded mono-
tone in W ∗ (a) and hence has a strong limit b. By definition, dP f = b.
Finally, write a general (complex-valued) f as a sum of positive terms and define
the integral by linearity. Thus (4.36) holds just by definition of the right-hand side.
Crucially, it follows that a von Neumann algebra is generated by its projections:48
Theorem 4.8 Let P(M ) = {p ∈ M | p2 = p∗ = p} be the projection lattice in a
von Neumann algebra M ⊂ B(H). Then M = P(M )00 .
The inclusion P (M )00 ⊆ M follows from P (M ) ⊂ M and M 00 = M . Conversely, by
Theorem 4.3, the spectral projections of a∗ = a ∈ M lie in W ∗ (a) ⊆ M , and hence
in P(M ). Let E(M ) be the finite linear span of P(M ); our proof of (4.36) yields a
as a strong limit of some sequence in E(M ), so that a ∈ E(M )00 = P (M )00 . 
48
This is not true for general C ∗ -algebras! Just think of A = C0 (R), with P(A) = {0}.
4 VON NEUMANN ALGEBRAS 40

4.4 von Neumann algebras as C ∗ -algebras

Since the strong and the weak topologies on B(H) are weaker than the norm topol-
ogy, it is clear from Theorem 4.1 that a von Neumann algebra is norm-closed and
hence is a C ∗ -algebra. In the spirit of Definition 3.3, one may look for a character-
ization of von Neumann algebras as C ∗ -algebras with some extra property that is
independent of the embedding M ⊂ B(H). This problem was solved by Sakai [30].49
Theorem 4.9 A C ∗ -algebra M is isomorphic to a von Neumann algebra iff it is
the (Banach) dual of a Banach space M∗ (called the predual of M ).
There is a canonical embedding M∗ ,→ M ∗ , ϕ̂ 7→ ϕ, with ϕ(a) = a(ϕ̂). Elementary
functional analysis [8, Thm. v.1.3] shows that (the image of) M∗ consists precisely
of the weak∗ -continuous functionals on M (where the weak∗ -topology on M is the
topology of pointwise convergence, seeing M as the dual of M∗ ).
For example, in the commutative case M = L∞ (X, µ), acting on H = L2 (X, µ)
as multiplicationR operators, standard measure theory gives M∗ = L1 (X, µ) under the
pairing ϕ(f ) = X dµ f ϕ̂ [28, Thm. 6.5.11]. In the noncommutative case, Theorem
4.42 in Appendix 2 or [28, Thm. 3.4.13] yields M∗ = B1 (H) for M = B(H) under
the familiar identification ρ(a) = Tr (ρ̂a), the trace-norm kρ̂k1 on B1 (H) coinciding
with the norm kρk in B(H)∗ . This also shows that the σ-weak topology on B(H)
coincides with the weak∗ -topology, which is true for all von Neumann algebras:
Theorem 4.10 Let M ⊂ B(H) be a von Neumann algebra. The predual M∗ of M
(seen as a subspace of M ∗ ) coincides with the space of σ-weakly continuous func-
tionals on M (so that the σ-weak topology on M coincides with the weak∗ -topology).
We just sketch the proof, leaving details to the exercises and Appendix 3. Let
M ⊥ := {ρ̂ ∈ B(H)∗ | ρ(a) = 0 ∀a ∈ M };
M ⊥⊥ := {a ∈ B(H) | ρ(a) = 0 ∀ρ̂ ∈ M ⊥ }.
Assuming the theorem for M = B(H) (i.e. Theorem 4.42), the key is to show that
M ⊥⊥ = M ; (4.37)
M∗ ∼= B(H)∗ /M ⊥ , (4.38)
where (4.38) denotes an isometric isomorphism of normed spaces. Since the right-
hand side of (4.38) is a Banach space, so is the left-hand side. This yields the first
claim. Combining (4.38) with the duality B(H) = B1 (H)∗ and (4.37), we have
M∗ ∼
= (B(H)∗ /M ⊥ )∗ = M ⊥⊥ = M.
∗

This is the second claim. The first equality sign is true, because if Y is a closed
subspace of a Banach space Y , then (X/Y )∗ = {ϕ ∈ X ∗ | ϕ  Y = 0}.
For the remainder of the theorem, note that aλ → a σ-weakly in M whenever
ρ(aλ − a) → 0 for all ϕ ∈ B(H)∗ . By (4.38), this is equivalent to aλ → a in the
weak∗ -topology, since a possible component of ϕ in M ⊥ drops out. 
Corollary 4.11 Each ϕ ∈ M∗ is of the form ϕ(a) = Tr (ρ̂a), for some ρ̂ ∈ B1 (H).
49
A proof of this theorem will be sketched Appendix 3.
4 VON NEUMANN ALGEBRAS 41

Exercises for Lecture 5

1. (a) Let S ⊂ B(H) be such that a ∈ S iff a∗ ∈ S. Show that S 0 is a von
Neumann algebra.
(b) Apply this to show that U (G)0 is a von Neumann algebra for any unitary
group representation U : G → B(H) on a Hilbert space H.
(c) Show that any von Neumann algebra arises in this way. In other words,
given a von Neumann algebra M ⊂ B(H), give a group G and a unitary
representation U : G → B(H) such that M = U (G)0 .

2. Derive the following useful reformulation of Theorem 4.1 from the latter:
Let M be a unital ∗ -subalgebra of B(H). Then the closures of M in the strong,
weak, and σ-weak topologies coincide with each other and with M 00 .

3. Let M be a unital ∗ -algebra in B(H), take a vector v ∈ H, and let p be the

(orthogonal) projection onto the closure of M v ⊂ H. Prove that p ∈ M 0 .

4. In the notation of the proof of Theorem 4.1, prove that (Mn )00 = (M 00 )n .

5. In the proof of Theorem 4.1 we showed that if M is a unital ∗ -subalgebra of

Mn (C), then M 00 = M . Prove the following claim for infinite-dimensional H:
if M is a unital ∗ -subalgebra of B(H), then M 00 = M iff M is strongly closed.
∞
6. For possibly infinite-dimensional H, define a new Hilbert space P H 2 whose
(v1 , v2 , . . .) in H with i kvi k < ∞;
elements v are infinite sequences of vectors P
the inner product is given by (v, w)H ∞ = i (vi , wi )H . (Note: H ∞ = H ⊗ `2 .)
There is an obvious (diagonal) embedding of B(H) in B(H ∞ ), whose image is
denoted by B(H)∞ . Similarly, the image of M ⊂ B(H) is denoted by M∞ .

(a) Show that the σ-weak topology on B(H) is the relative weak topology
on B(H)∞ (i.e., the weak topology on B(H ∞ ) restricted to B(H)∞ ).
(b) Define a new topology on B(H), called the σ-strong topology, by re-
stricting the strong topology on B(H ∞ ) to B(H)∞ . Use the same trick
as above (i.e., the passage of H to H d ) to prove the following version of
the Double Commutant Theorem (and hence (1) ⇔ (4) in Theorem 4.1):
Theorem 4.12 Let M be a unital ∗ -algebra in B(H). The following
conditions are equivalent:
i. M is a von Neumann algebra;
ii. M is closed in the σ-weak operator topology;
iii. M is closed in the σ-strong operator topology.

7. Complete the proof of Theorem 4.3 by showing that the map f 7→ f (a),
initially from B(σ(a)) to B(H), is a homomorphism of C ∗ -algebras.

8. Prove (4.37) and (4.38) (the answer is in Appendix 3).

4 VON NEUMANN ALGEBRAS 42

4.5 Isomorphisms between von Neumann algebras

An isomorphism between C ∗ -algebras in the sense of ∗ -algebras is automatically iso-
metric; see Proposition 3.21. An even better result holds for von Neumann algebras:

Theorem 4.13 An isomorphism ϕ : M → N between von Neumann algebras (as

∗
-algebras) is an isomorphism of Banach spaces and a homeomorphism w.r.t. the σ-
weak topologies on M and N (hence M∗ and N∗ are isomorphic as Banach spaces).

The claim about the norm-topology follows from Proposition 3.21, for von Neumann
algebras are C ∗ -algebras. Since ϕ is isometric, it induces a dual isomorphism (of
Banach spaces) ϕ∗ : N ∗ → M ∗ , with the property that M ∼ = (ϕ∗ (N∗ ))∗ under
∗
the map a 7→ (ϕ (ω) 7→ ω(ϕ(a))), a ∈ M , ω ∈ N∗ . Uniqueness of the predual
(cf. [33, Vol. I Cor. iii.3.9]) then yields ϕ∗ (N∗ ) ∼
= M∗ , which in turn implies that
ϕ preserves pointwise convergent nets: if ω (aλ ) → ω 0 (a) for all ω 0 ∈ M∗ , then
0

ω(ϕ(aλ )) → ω(ϕ(a)) for all ω ∈ N∗ . Hence ϕ is σ-weakly continuous. 

A second instructive proof is based on the projection lattice; cf. Theorem 4.8.

Proposition 4.14 The set P(M ) of projections in a von Neumann algebra M is a

complete lattice under the partial ordering p 6 q iff pq = qp = p.

Since p 6 q in M ⊂ B(H) iff pH ⊆ qH (exercise 1), the supremum p ∨ q is the

projection on pH + qH, whilst the infimum p ∧ q is the projection on pH ∩ qH (see
exercise 2 for intrinsic expressions independent of H). For arbitrary families (pλ )λ∈Λ
of projections, ∨λ pλ equals the projection on the closure of the linear span of all
subspaces Hλ ≡ pλ H, whereas ∧λ pλ ≡ p is the projection on their intersection. To
show that the latter lies in M (provided all the pλ do, of course), note that each
unitary u ∈ M 0 satisfies uHλ = Hλ for all λ, so that also u(∩λ Hλ ) = ∩λ Hλ . Hence
pu = up and so p ∈ M 00 = M (for each element of a von Neumann algebra is a linear
combination of at most four unitaries in it). Finally, by de Morgan’s Law we have
∨λ pλ = (∧λ p⊥ ⊥ ⊥
λ ) , with f = 1 − f for any f ∈ P(M ). Hence also ∨λ pλ ∈ M . 

Definition 4.15 A map ϕ : M → N of von Neumann algebras is called completely

additive if ϕ(∨λ pλ ) = ∨λ ϕ(pλ ) for any family (pλ ) in P(M ).

Theorem 4.13 then follows from two key properties of completely additive maps:

Proposition 4.16 Let ϕ : M → N be a homomorphism of von Neumann algebras.

1. ϕ is σ-weakly continuous iff it is completely additive.50

2. If ϕ is an isomorphism, then it is completely additive.

The proof of claim 2 is an exercise, as is the implication from σ-weak continuity

to completely additivity in claim 1. The converse implication, however, is quite
difficult; we refer to [10, §i.4.2], [33, Vol. I, Cor. iii.3.11], [19, Lecture 11].
50
Similarly, a functional ρ : M → C is σ-weakly continuous iff it is completely additive.
4 VON NEUMANN ALGEBRAS 43

4.6 Classification of abelian von Neumann algebras

Theorem 4.13 shows that the notion of isomorphism to be used in the classification
of von Neumann algebras M is unambiguous. There are two totally different cases:51
• Abelian von Neumann algebras, which equal their center (M ∩ M 0 = M );
• Factors, i.e., von Neumann algebras with trivial center (M ∩ M 0 = C · 1).
Using the technique of direct integrals, the classification of general von Neumann
algebras may be reduced to these cases [10, Part ii], [33, Vol. I, Ch. iv, v].
As to the abelian case, we first sharpen Theorem 4.3 (cf. [8, §ix.8.10]).
Theorem 4.17 Let a∗ = a ∈ B(H). There exists a measure µ on σ(a) such that
the isomorphism C(σ(a)) → C ∗ (a, 1) of C ∗ -algebras (cf. Theorem 3.18) has a unique
extension to an isomorphism L∞ (σ(a), µ) → W ∗ (a) of von Neumann algebras.52
Such a measure µ has the defining property that µ(∆) = 0 iff µψ (∆) = 0 for all unit
vectors Ψ ∈ H (see §4.2 for notation), and hence is defined up to equivalence by its
measure class (if H is separable, µ may even be taken to be a probability measure).
We omit a complete proof,53 but would like to clarify the key points (cf. [8, §ix.8]):
1. The Borel functional calculus Ba : f 7→ f (a), B(σ(a)) → W ∗ (a), is surjective.
2. It is not injective; for real-valued f one has f (a) = 0 iff (Ψ, f (a)Ψ) = 0 for
Rall unit vectors Ψ ∈ H, which by the last claim of Theorem 4.3 isess,µthe case iff
σ(a)
dµψ f = 0 for all Ψ, which by construction is the case iff kf k∞ = 0.

3. Hence W ∗ (a) ∼
= B(σ(a))/ ker Ba ∼
= L∞ (σ(a), µ) as Banach spaces (via Ba ).
Theorem 4.18 Let M ⊂ B(H) be an abelian von Neumann algebra (H separable).
Then M ∼
= L∞ (X, µ) for some compact space X and probability measure µ on X.
This follows from the previous theorem by a result of von Neumann himself [23]:
Proposition 4.19 Let M ⊂ B(H) be an abelian von Neumann algebra (H separa-
ble). Then M = W ∗ (a) for some a = a∗ ∈ B(H) (i.e., M is “singly generated”).
The proof is a (difficult) exercise. Finally, the most general result is as follows:54
Theorem 4.20 Let M ⊂ B(H) be an abelian von Neumann algebra. Then one has
M∼= L∞ (X, µ) for some locally compact space X and Borel measure µ on X.
Conversely, L∞ (X, µ) defines a von Neumann algebra of multiplication operators on
H = L2 (X, µ), with operator norm equal to the norm k · kess,µ
∞ [8, Thm. ii.1.5], so
that we have found a complete characterization of abelian von Neumann algebras!
51
Of course, M = C falls in both classes, but is unique as such.
52
Here L∞ (σ(a), µ) denotes the space of (equivalence classes of) Borel functions f : σ(a) → C for
which f is bounded on σ(a)\∆0 for some Borel subset ∆0 ⊂ σ(a) with µ(∆0 ) = 0. The canonical
norm on L∞ (σ(a), µ) is kf kess,µ
∞ = inf{sup{|f (x)|, x ∈ σ(a)\∆}∆ ⊂ σ(a), | µ(∆) = 0}.
53
For the existence of µ see [8, §ix.8] for separable H, and [10, §i.7.2] in general.
54
The proof is technical (cf. [10, Thm. i.7.3.1] or [33, Vol. I, §iii.1]), but the idea is to find an
abelian C ∗ -algebra A for which M = A00 , upon which X = Σ(A), and the measure µ is constructed
such that µ(∆) = 0 iff µψ (∆) = 0 for all unit vectors Ψ ∈ H, with µψ defined similarly to (4.35).
One cannot take A = M , since Σ(M ) may not support such measures, cf. [33, Vol. I, Thm. iii.1.18].
4 VON NEUMANN ALGEBRAS 44

4.7 Abelian von Neumann algebras and Boolean lattices

Theorem 4.20 is not as good as its counterpart Theorem 3.4 for C ∗ -algebras, since in
the latter X is unique up to homeomorphism, whereas in the former the pair (X, µ)
lacks intrinsic uniqueness properties.55 Thus it also makes sense to apply Theorem
3.4 to abelian von Neumann algebras, so that M ∼ = C(X). Since by Theorem 4.8, M
has plenty of projections, which as elements of C(X) are realized by characteristic
functions χ∆ (∆ ⊂ X), the space X must have lots of clopen (i.e. closed and open)
sets. It can be shown that X arises as the Gelfand spectrum of some abelian von
Neumann algebra iff it is hyperstonean, which (besides compact Hausdorff) means:
• X is Stone if the only connected subsets of it are single points.56
• X is Stonean if it is Stone and the closure of every open set is (cl)open;
• X is hyperstonean if it is Stonean, and for any nonzero f ∈ C(X, R+ ) there
exists a completely additive positive measure µ such that µ(f ) > 0.
See [33, Vol. I, §iii.1]. Now, the only other area of mathematics where such crazy
spaces appear is logic. Indeed, recall that a Boolean lattice is an orthocomplemented
distributive lattice,57 and also recall Stone’s Theorem, which states that a lattice L
is Boolean iff it is isomorphic to the lattice of all clopen subsets of a Stone space
X (where the partial ordering is given by set-theoretic inclusion, so that the lattice
operations are U ∨ W = U ∪ W and U ∧ V = U ∩ W ). The space X ≡ Σ̂(L) is called
the Stone spectrum of L, and is determined by L up to homeomorphism [14].
Theorem 4.21 The projection lattice P(M ) of a von Neumann algebra M is Boolean
iff M is abelian, in which case the Gelfand spectrum Σ(M ) of M (as a commuta-
tive C ∗ -algebra) is homeomorphic to the Stone spectrum Σ̂(P(M )) of P(M ). Hence
M∼ = C(Σ̂(P(M ))), whilst P(M ) is isomorphic to the lattice of clopens in Σ(M ).
Towards a proof, recall that the Stone spectrum Σ̂(L) is the set of homomorphisms
ϕ̂ : L → {0, 1} of Boolean lattices (where {0, 1} is a lattice under 0 6 1), topologized
by saying that the basic opens in ΣL are those of the form Ux = {ϕ̂ ∈ Σ̂(L) | ϕ̂(x) =
1}, for each x ∈ L. The homeomorphism Σ(M ) ∼ = Σ̂(P(M )) then arises as follows:
• ϕ ∈ Σ(M ), ϕ : M → C, restricts to ϕ̂ ∈ Σ̂(P(M )), ϕ̂ : P(M ) → {0, 1}.
• ϕ̂ ∈ Σ̂(P(M )) extends to a character ϕ ∈ Σ(M ) by the spectral theorem.

55
Under special assumptions, at least some good models for (X, µ) obtain. To explain the two
main examples, let us call a projection p ∈ P(M ) minimal if p 6= 0 and there exists no q ∈ P(M )
such that 0 < q < p (where q < p iff q 6 p and q 6= p). Then M is called atomless if it has no
minimal projections, whereas it is said to be atomic if for any nonzero p ∈ P(M ) there is a minimal
projection q ∈ P(M ) such that q 6 p. Under the assumption that H is separable, we then have
M∼ = L∞ (0, 1) (w.r.t. Lebesgue measure) iff M is atomless [33, Vol. I, Thm. iii.1.22], whereas if M
is atomic, it is isomorphic to either `∞ or to a finite direct sum of copies of C (exercise).
56
Equivalently, a Stone space is compact, T0 , and has a basis of clopen sets.
57
Recall that a lattice L is called distributive when x ∨ (y ∧ z) = (x ∨ y) ∧ (x ∨ z), and ortho-
complemented when there exists a map ⊥ : L → L that satisfies x⊥⊥ = x, y ⊥ 6 x⊥ when x 6 y,
x ∧ x⊥ = 0, and x ∨ x⊥ = 1. For example, P(M ) is orthocomplemented by p⊥ = 1 − p.
4 VON NEUMANN ALGEBRAS 45

Exercises for Lecture 6

1. For any two projections p, q on H, show that pq = qp = p iff pH ⊆ qH.
2. Show that for any two projections p, q ∈ M ⊂ B(H) one has
M abelian
p∧q = lim (pq)n = pq; (4.39)
n→∞
M abelian
p ∨ q = (p⊥ ∧ q ⊥ )⊥ = p + q − pq, (4.40)
where the limit is taken in either the strong or the σ-weak topology on M ,
p⊥ = 1 − p, and the expressions on the right apply when M is abelian.
3. Prove that an isomorphism of von Neumann algebras is completely additive.
4. Prove that a σ-weakly continuous homomorhism of von Neumann algebras is
completely additive. Hint 1: Use the following sharpening of Lemma 4.7:
Lemma 4.22 Any monotone increasing net (aµ ) of self-adjoint operators that
is bounded from above strongly converges to its supremum ∨µ aµ in B(H).
See [16, Lemma 5.1.4]. Hint 2: Show that strong convergence in this lemma
may be replaced by σ-weak convergence, as follows: show that strong con-
vergence of a bounded monotone increasing net in B(H) implies strong con-
vergence of the corresponding net of diagonal operators in B(H ∞ ), which by
definition establishes σ-strong and hence σ-weak convergence in B(H).
5. Use Lemma 4.22 to give a new proof of Proposition 4.14, in which the existence
of ∨λ pλ is shown first (from which the existence of ∧λ pλ is derived). Hint: With
λ ∈ Λ, use the net Pf (Λ) of finite subsets of Λ (ordered by inclusion).
6. Around Theorem 4.17, assumeP∞ that H is separable with o.n.b. (en ) and prove
−n
that the measure µ(∆) = n=1 2 (en , p(∆)en ) is a probability measure on
σ(a), which satisfies µ(∆) = 0 iff µψ (∆) = 0 for all unit vectors Ψ ∈ H.
7. Prove Proposition 4.19. Hint: give a self-contained version of von Neumann’s
own proof [23, Satz 10] (see www.math.ru.nl/∼landsman/Johnny1929.pdf).58
8. Show that a Boolean lattice may equivalently be defined as a commutative
ring in which x2 = x for all x (see [19, Lecture 14]).
9. Give an example illustrating that P(M ) is not distributive (and hence not
Boolean) whenever M is noncommutative.
10. Is the canonical example of a Stone space, viz. the Cantor set C,59 Stonean?
58
This proof has been rephrased and streamlined by Takesaki [33, Vol. I, Prop. iii.1.21]. See
also Exercise f on p. 134 of [10], which we paraphrase as follows. Since H is separable, M is
generated by a countable family (pn ) of projections. From this, construct a new family (p0 )r∈Q∩[0,1]
of projections, such that if r ≤ r0 then p0r 6 p0r0 , p00 = 0, p01 = 1, and s-limr↑r0 pr = pr0 . Then
construct a self-adjoint a ∈ M such that for each r ∈ Q∩[0, 1], p0r is the greatest spectral projection
for a such that p0r a 6 rQ
· p0r . Finally, prove that a generates M , i.e., that M = W ∗ (a).
P∞
59
This is either C = N 2 or, homeomorphically, f (C) ⊂ [0, 1] defined by f (x) = n=1 xn 2−n .
4 VON NEUMANN ALGEBRAS 46

4.8 The von Neumann algebra of a discrete group

We now turn to the classification of factors, following the original work of Murray and
von Neumann [20]. First, we show that there is actually something worth classifying
by displaying a class of nontrivial examples of factors. In contrast, ‘trivial’ examples
of factors are those of the form M = B(H), or, slightly more involved (whence the
name ‘factor’): M = B(H1 )⊗1H2 ⊂ B(H1 ⊗H2 ) or its commutant M 0 = 1H1 ⊗B(H2 ).
The first is isomorphic to B(H1 ) by the map a⊗1H2 7→ a, and similar for the second.
To get something new, define the group algebra CG of a discrete group G as
the ∗ -algebra of functions f : G → C with finite support, with operations
X
f ∗ g(x) = f (xy −1 )g(y); f ∗ (x) = f (x−1 ). (4.41)
y∈G

Note that CG has a unit, namely δe (where δP e (x) = δe,x ). Next, represent CG on the
Hilbert space H = `2 (G) by f 7→ πL (f ) = x f (x)UL (x), where the (left) regular
representation UL of G is defined by UL (x)ψ(y) = ψ(x−1 y). The von Neumann
algebra of G is defined as W ∗ (G) = πL (CG)00 , so that W ∗ (G) ⊂ B(`2 (G)).
To see what is going on here, consider the case where G is finite,60 with an asso-
ciated finite set Ĝ of (unitary) equivalence classes of irreducible representations.61
∼ dγ
L γ ∈ Ĝ, realized on Hγ = C (with dγ < ∞),
Take some representative Uγ of each
and define a Hilbert space HĜ = γ∈Ĝ Mdγ (C) consisting of matrix-valued functions
ψ̂ on Ĝ, ψ̂(γ) ∈ Mdγ (C), with inner product (ψ̂, ϕ̂) = γ Tr (ψ̂(γ)∗ ϕ̂(γ)). Then UL
P

and W ∗ (G) are block-diagonalized by a unitary62 transformation u : `2 (G) → HĜ ,

p X
uψ(γ) ≡ ψ̂(γ) = dγ ψ(x)Uγ (x); (4.42)
x∈G
1 Xp
u−1 ψ̂(x) ≡ ψ(x) = dγ Tr (Uγ (x)∗ ψ̂(γ)). (4.43)
|G|
γ∈Ĝ

Indeed, writing ÛL (x) ≡ uUL (x)u−1 , we easily find ÛL (x)ψ̂(γ) = Uγ (x)ψ̂(γ), and
similarly π̂L (f )ψ(γ) = (1/ dγ )fˆ(γ)ψ(γ), where π̂L (f ) = uπL (f )u−1 . Using Schur’s
p

lemma (i.e., πγ (G)00 = Mdγ (C)), we obtain uW ∗ (G)u−1 = HĜ , seen as a ∗ -algebra,
acting on itself (now in its original guise as a Hilbert space) by fˆψ̂(γ) = fˆ(γ)ψ̂(γ).
Thus W (G) ∼
∗
L
= γ∈Ĝ Mdγ (C) is isomorphic to a finite direct sum of matrix algebras,
whose center W ∗ (G)∩W ∗ (G)0 is isomorphic to C(Ĝ). InPfact, ˜
p f : Ĝ → C corresponds
to ψ̂(γ) = f˜(γ)·1dγ in HĜ and hence to f (x) = (1/|G|) γ dγ f˜(γ)χγ (x) in W ∗ (G).
In particular, f lies in the center of W ∗ (G) iff it is a class function. This also
follows from (4.41), since g ∗ g = g ∗ f for all g iff f (yxy −1 ) = f (x) for all x, y ∈ G.
60
See, for example, [31] for the finite group theory used here (all based on the Peter–Weyl theory).
61
Recall that the cardinality of Ĝ equals the number kG of conjugacy classes in G. This is
because the class functions on G by definition form a kG -dimensional subspace C(G)c of `2 (G),
whilst the characters x 7→ χγ (x) = Tr Uγ (x), γ ∈ Ĝ, form a basis of C(G)c .
62
Unitarity follows from the fact that the matrix elements x 7→ (ei , Uγ (x)ej ), where (ei ) is an
o.n.b. of Hγ , form a basis of `2 (G), combined with Schur’s orthogonality relations. This implies in
turn that each irreducible representation Uγ of G, is contained in UL with multiplicity dγ .
4 VON NEUMANN ALGEBRAS 47

4.9 Nontrivial examples of factors

As we have seen, for finite G the group von Neumann algebra W ∗ (G) is not a factor
(unless G = {e}). However, our computation of the center W ∗ (G)∩W ∗ (G)0 suggests:
Theorem 4.23 The group von Neumann algebra W ∗ (G) of a countable group is a
factor iff all nontrivial conjugace classes in G (i.e., all except {e}) are infinite.
Since πL (CG) consists of all finite sums x f (x)UL (x), it is likely that W ∗ (G) =
P
πL (CG)00 incorporates all strongly convergent sums of this kind. Indeed, the map
τ (f ) = f (e) (4.44)
from CG to C extends to W ∗ (G) by P(strong) continuity and has the property that
for any a ∈ W ∗ (G) one has a = x f (x)UL (x) with f (x) = τ (aUL (x −1
))UL (x),
where the sum over G converges strongly (exercise). We identify such a with the
corresponding function f , and compute τ (f ∗ ∗ f ) = kf k22 , so that W ∗ (G) ⊂ `2 (G).
The computation at the very end of the previous section still holds, from which we
infer that f ∈ W ∗ (G) ∩ W ∗ (G)0 iff f is constant on each conjugacy class of G. If so,
and if the condition in the theorem holds, then f ∈ `2 (G) iff f = λδe , or f ∈ C · 1.
Conversely, any f that is constant on some finite conjugacy class (different from
{e}) and zero elsewehere is central without being a multiple of the unit. 
Do such “icc” groups actually exist? In fact, there are infinitely many of them:
each free group on n > 1 generators is an example. Another example is the group
S∞ of finite permutations of N.63 In any case, we would now like to determine if
such a factor W ∗ (G) is “trivial” or not. The simples way to do so involves the trace.
Definition 4.24 A trace on a von Neumann algebra M is a map τ : M+ → [0, ∞]
satisfying τ (λ·a+b) = λ·τ (a)+τ (b) for all a, b ∈ M+ and λ ≥ 0, and τ (aa∗ ) = τ (a∗ a)
for all a ∈ M (equivalently τ (uau∗ ) = τ (a) for all a ∈ M+ and unitary u ∈ M ).
A trace is finite if τ (a) < ∞ for all a ∈ M+ , semifinite if for any a ∈ M+
there is a nonzero b ≤ a in M+ for which τ (b) < ∞, and infinite otherwise.
A von Neumann algebra is called (semi)finite if it admits a faithful σ-weakly
continuous (semi)finite trace,64 and purely infinite otherwise.
The usual trace Tr is a trace on M = B(H) in this sense; B(H) is finite iff dim(H) <
∞ and semifinite otherwise. So is the map τ in (4.44), making W ∗ G) finite.
To distinguish W ∗ (G) from B(H), we state without proof [33, Vol. I, Cor. v.2.32]:
Theorem 4.25 Any two nonzero σ-weakly continuous (semi)finite traces τ, τ 0 on a
(semi)finite factor are proportional, i.e., τ 0 = λτ for some λ ∈ R+ .
Since W ∗ (G) is infinite-dimensional for icc G, W ∗ (G)  Mn (C). For infinite-
dimensional H, however, we have Tr (1) = ∞ on B(H), whereas τ (1) = 1 on
W ∗ (G). Hence, as Tr and τ are σ-weakly continuous, W ∗ (G)  B(H) for any H.
63
A j-cycle is a cyclic permutation of j objects (called the carrier of the cycle in question). Any
element p of S∞ = ∪n Sn is a finite product of j-cycles with disjoint carriers, and for each j ∈ N the
number of j-cycles in such a decomposition of p is uniquely determined by p. Two permutations
in S∞ , then, are conjugate iff they have the same number of j-cycles, for all j ∈ N (cf., e.g., [31]).
64
A finite trace on a factor is automatically σ-weakly continuous [33, Vol. I, Prop. v.2.5].
4 VON NEUMANN ALGEBRAS 48

4.10 Equivalence of projections

Murray and von Neumann actually started with the projection lattice P (M ). In
Proposition 4.14 we have already introduced a partial ordering 6 on P (M ), but
another ordering exists, which turns out to be total whenever M is a factor.
Definition 4.26 Let P (M ) be the projection lattice of a von Neumann algebra M .
We say that p ∼ q in P (M ) when there exists u ∈ M such that u∗ u = p and uu∗ = q,
and that p . q if there is p0 ∈ P (M ) with p ∼ p0 and p0 6 q.
It is an exercise to show that ∼ is an equivalence relation. Through its stated
properties, the operator u in this definition is unitary from pH to qH, vanishes on
(pH)⊥ , and has range qH. Such an operator is called a partial isometry, with
initial projection p and final projection q. It follows that a necessary condition
for p ∼ q is that dim(pH) = dim(qH), but (unless M = B(H)) this is by no means
sufficient, since the unitary mapping pH to qH is required to lie in M . For example,
if H = C ⊕ C then p = diag(1, 0) is equivalent to q = diag(0, 1) with respect to
M = M2 (C), but not with respect to M = C ⊕ C (i.e., the diagonal 2 × 2 matrices).
To see how natural this definition is, consider a unitary representation U of a
group G on H. If Hi ⊂ H is stable under U (G), i = 1, 2, then the restrictions Ui of U
to Hi are unitarily equivalent precisely when [H1 ] ∼ [H2 ] with respect to M = U (G)0
(where [Hi ] is the projection onto H). Furthermore, U1 is unitarily equivalent to
a subrepresentation of U2 iff [H1 ] . [H2 ]. More generally, if N ⊂ B(H) is a von
Neumann algebra, with stable subspaces Hi , i = 1, 2, then the restrictions Ni to Hi
are unitarily eqiuvalent iff [H1 ] ∼ [H2 ] with respect to M = N 0 , et cetera.
One may compare projections in M with sets and compare 6, ∼, and . with
⊆ (inclusion), ∼= (isomorphism), and ,→ (the existence of an injective map), respec-
tively. The Schröder–Bernstein Theorem of set theory states that if X ,→ Y and
Y ,→ X, then X ∼ = Y . Similarly (with the proof as an exercise):
Proposition 4.27 If p . q and q . p, then p ∼ q.
The special role of factors with respect to the partial ordering . now emerges.
Proposition 4.28 If M is a factor, then . is a total ordering (i.e., p . q or q . p).
The property of a factor that leads to this result (with the proof as an exercise) is:
Lemma 4.29 Let M be a factor. For any nonzero projections p, q ∈ P (M ) there
are nonzero projections p0 , q 0 ∈ P (M ) such that p0 6 p, q 0 6 q, and p0 ∼ q 0 .
In other words, any pair p, q of nonzero projections in a factor has a pair of equivalent
subprojections. Accordingly, from Zorn’s Lemma (brr . . . ) there exist maximal
orthogonal families (piP) and (qj ) with pi 6 p,Pqj 6 q, and pi ∼ qj , for all i, j.
Defining p0 = ∨i pi = i pi and q0 = ∨j qj = j qj , one has p0 6 p, q0 6 q, and
p0 ∼ q0 (exercise). But then, by maximality, p − p0 and q − q0 have no equivalent
subprojections. To avoid a contradiction with Lemma 4.29, at least one of these
must vanish. If p − p0 = 0, then p ∼ q0 6 q, so p . q. If q − q0 = 0, one similarly
has q . p (both may be zero, in which case p ∼ q by Proposition 4.27). 
4 VON NEUMANN ALGEBRAS 49

4.11 The Murray–von Neumann classification of factors [20]

Definition 4.30 A projection p in M is called finite if q ∼ p and q 6 p for some
q ∈ P (M ) implies q = p, and minimal if q ≤ p, q ∈ P (M ), implies q = p or q = 0.
As an alternative to Definition 4.24, we then have [33, Vol. I, Thms. v.2.4, v.2.15]:
Proposition 4.31 A factor M is finite iff 1 is finite, semifinite iff 1 majorizes a
finite projection, and purely infinite iff all projections are infinite.
For M = B(H), we have p (in)finite iff dim(pH) = Tr (p) is (in)finite. However, if
G is icc, all projections p ∈ W ∗ (G) are finite, despite Tr (p) = ∞. All minimal pro-
jections in B(H) are one-dimensional, whereas W ∗ (G) has no minimal projections.
Definition 4.32 A factor M is said to be of type:
• i if it has at least one minimal projection, subdivided into:
– type in (n ∈ N) if M is finite and 1 is the sum of n minimal projections;
– type i∞ if M is type i and semifinite but not finite;
• ii if it has no minimal projections but some nonzero finite projection, with:
– type ii1 if M is type ii and finite;
– type ii∞ if M is type ii and semifinite but not finite;
• iii if all nonzero projections are infinite.
A nice understanding of these types arises from a construction similar to the trace.
Definition 4.33 A dimension function on a von Neumann algebra M is a func-
tion d : P (M ) → [0, ∞] such that d(p) < ∞ iff p is finite, d(p + q) = d(p) + d(q) if
pq = 0 (i.e., pH ⊥ qH), and d(p) = d(q) if p ∼ q.
Paraphrasing results in Murray and von Neumann’s papers [20, 25], we now have:
Theorem 4.34 For any von Neumann algebra M , the restriction of a trace to
P (M ) is a dimension function. If M ⊂ B(H) is a factor, with H separable:
1. Any σ-weakly continuous trace on M restricts to a completely additive dimen-
sion function with the additional property that p ∼ q if d(p) = d(q).
2. Any dimension function with this additional property arises from a σ-weakly
continuous trace, and hence is completely additive and unique up to scaling.
3. In that case, the dimension function d induces an isomorphism between P (M )/ ∼
and some subset of [0, ∞]. Suitably normalizing d, this subset must be one of:
• {0, 1, 2, . . . , n} for some n ∈ N, in which case M is type in ;
• N ∪ ∞ (type i∞ );
• [0, 1] (type ii1 );
• [0, ∞] (type ii∞ );
• {0, ∞} (type iii).
4 VON NEUMANN ALGEBRAS 50

Exercises for Lecture 7

1. This exercise sharpens the polar decomposition [8, viii.3.11]:

Lemma 4.35 For any a ∈ B(H) there exists √ a unique partial isometry w
∗
such that a = w|a| = |a |w, where |a| = a∗ a (defined by the continuous
functional calculus), the initial projection w∗ w is [(ker a)⊥ ] = [ran(a∗ )], and
the final projection ww∗ is [ran(a)].

Prove that if a ∈ M ⊂ B(H), where M is a von Neumann algebra, then

w ∈ M and |a| ∈ M (answer: [33, Vol. I, Prop. ii.3.14]).
In all exercises below, M is a factor on a separable Hilbert space.
2. Prove Lemma 4.29. Hint (assuming p, q ∈ P (M ) nonzero):
(a) First show that pM q 6= 0.
(b) Then show that the polar decomposition of some nonzero a ∈ pM q gives
an equivalence between nonzero subprojections of p and q.
3. Prove the following statements (relative to M ):
(a) If p is finite and q ∼ p, then q is finite.
(b) If p is finite and q . p, then q is finite.
(c) If P (M ) contains some infinite (i.e., not finite) p, then dim(H) = ∞.
(d) A minimal projection is finite.
4. (a) Show that for all p, q ∈ P (M ), q 6= 0, there exists an index set I and an
orthogonal family (pi )i∈I in P (M ) with pi ∼ q for
P all i, as well as some
r ∈ P (M ) with r . q, and r 6= q, such that p = i∈I pi + r.
(b) Show that if p is finite, then |I| < ∞, and the cardinality |I| of I is
independent of the choices of the pi and r.
(c) Show that if q is minimal, then r = 0.
5. Show that if p and q are infinite, then p ∼ q. Hint: use the previous exercise.
6. Let M = B(H). Show from Theorem 4.34 that M is type idim(H) .
Hint: prove that d(p) := dim(pH) defines a dimension function on M .
7. Let M be a type i factor with some minimal projection q. Define d(p) = |I|,
where I is defined as in exercise 4.
(a) Show that d is a dimension function.
(b) Show that the range d(P (M )) of this function d is of the form listed in
Theorem 4.34 (part 3) either as type in , or as type i∞ .
8. Let G be icc. Show that W ∗ (G) is type ii1 .
4 VON NEUMANN ALGEBRAS 51

4.12 Classification of hyperfinite factors

Murray and von Neumann completely classified two sorts of factors. We say that a
von Neumann algebra M is hyperfinite if M = (∪n Mn )00 , where each von Neumann
subalgebra Mn ⊂ M is finite-dimensional and Mn ⊂ Mn+1 . For example, if H is
separable, then M = B(H) is hyperfinite. Also, if some countable group G is
the union of an increasing sequence of finite subgroups Gn , i.e., G = ∪n Gn with
Gn ⊂ Gn+1 , then the associated von Neumann algebra W ∗ (G) is hyperfinite. This
applies, for example, to the (icc) group S∞ = ∪n Sn of finite permutations of N.
Theorem 4.36 Let M ⊂ B(H) be a factor on a separable Hilbert space.
• If M is type i, then M ∼
= B(H), for some Hilbert space H.
• If M is type ii1 and hyperfinite, then M ∼
= W ∗ (S∞ ), henceforth called R.
The proof of the first claim is a nontrivial exercise. The second is much more difficult
[21]; it follows that W ∗ (G) ∼= R for any finitely generated icc group G. An example
of a ii∞ factor is also quickly found, namely M = N ⊗ B(`2 ), where N is ii1 . In
fact, any ii∞ factor on a separable Hilbert space is of this sort (exercise), but if M
is hyperfinite, it is a priori unclear if N is, too (see below). Hence Murray and von
Neumann were unable to classifiy even hyperfinite ii∞ factors. About type iii they
knew almost nothing, except for a couple of examples from ergodic theory [25].
Between 1971–1975, Alain Connes made two decisive contributions [3, 4, 5]:
1. Dividing type iii factors into iiiλ , λ ∈ [0, 1], by means of a new invariant;
2. Completely classifying hyperfinite type ii∞ and type iii factors, as follows:
• There is a unique hyperfinite ii∞ factor, namely R ⊗ B(`2 );
• There is a unique hyperfinite iii1 factor;
• There is a unique hyperfinite iiiλ factor for each λ ∈ (0, 1);
• There is an infinite family of hyperfinite iii0 factors, completely classified
by the so-called flow of weights of Connes and Takesaki [33, Vol. II].
We list iii1 separately from iiiλ for λ ∈ (0, 1) for two reasons: first, “hyperfinite iii1 ”
turns out to be the factor occurring in quantum field theory and quantum statistical
mechanics of infinite systems [13] (while iiiλ for λ ∈ (0, 1) seems artificial), and
second, the proof of uniqueness of the hyperfinite iii1 factor is much more difficult.65
Apart from the Tomita–Takesaki theory (cf. §4.13), an important technical tool of
Connes was his own profound discovery that a von Neumann algebra M ⊂ B(H), H
separable, is hyperfinite iff it is injective in that there exists a σ-weakly continuous
conditional expectation E : B(H) → M , that is, a linear map E : B(H) → B(H)
such that E(a) ∈ M and E(a∗ ) = E(a)∗ for all a ∈ B(H), E 2 = E, and kEk = 1.66
65
There is an entire book about this proof [35]. In his review MR1030046 (91a:46059) of this
book for Mathematical Reviews in 1991, E. Størmer wrote: ‘At the time of writing this review, by
far the deepest and most difficult proof in von Neumann algebra theory is the one of Connes and
Haagerup on the uniqueness of the injective factor of type iii1 with separable predual.’
66
It follows that E(abc) = aE(b)c for all a, c ∈ M , b ∈ B(H). The equivalence of hyperfiniteness
and injectivity implies, for example, that if M = N ⊗ B(`2 ) is hyperfinite, then so is N .
4 VON NEUMANN ALGEBRAS 52

4.13 Tomita–Takesaki Theory

At a conference 1967, the Japanese mathematician Tomita and the German-Dutch
mathematical physics trio Haag–Hugenholtz–Winnink (hhw) independently dis-
tributed two preprints that ‘would later completely change the scope of operator
algebra’ [32, p. 235]. Tomita’s original work was not quite correct, but after elab-
oration by Takesaki it formed the basis of both Connes’s subsequent work on the
classification of type iii factors and, through hhw, of a powerful mathematically
rigorous approach to the quantum statistical mechanics of infinite systems [13, 18].
Definition 4.37 A von Neumann algebra M ⊂ B(H) is in standard form if H
contains a unit vector Ω that is cyclic and separating for M .
Here we say that Ω is separating for M if aΩ 6= 0 for all nonzero a ∈ M (being cyclic
for M , Ω is also separating for M 0 ). The point is that any von Neumann algebra can
be brought into standard form [33, Vol. II, §ix.1]. For separable H, this follows by
picking an injective density operator ρ̂ on H, whose associated state ρ : a 7→ Tr (ρ̂a)
is faithful in that ρ(a∗ a) > 0 for all nonzero a ∈ M , and passing to πρ (M ) ∼
= M.
For example, M = B(H) acting on H is not in standard form, but acting on
B2 (H) by left multiplication it is, where B2 (H) is the Hilbert space of Hilbert–
∗
Schmidt operators on H with inner product √ (a, b) = Tr (a b). If ρ̂ ∈ B1 (H) is an
injective density operator on H, then Ω = ρ̂ ∈ B2 (H) brings M in standard form.
In this case, M 0 ∼
= B(H)op (where the suffix “op” means that multiplication is done
in the opposite order, i.e. ab in B(H)op is equal to ba in B(H)), which acts on B2 (H)
by right multiplication. If H = Cn , one simply has B(H) = B2 (H) = Mn (C).
Let M ⊂ B(H) be in standard form. Tomita introduced the (unbounded) anti-
linear operator S with initial domain D(S0 ) = M Ω and action S0 (aΩ) = a∗ Ω; this
domain is dense because Ω is cyclic for M , and the action is well defined since Ω is
separating for M . This operator turns out to be closable, with closure S. Any closed
operator
√ a has a polar decomposition1/2 a = v|a|, where v is a partial isometry and
∗
|a| = a a. For S, we write S = J∆ , where J is an antilinear partial isometry
and ∆ = S ∗ S. Since S is injective with dense range, J is actually anti-unitary, satis-
fying J ∗ = J and J 2 = 1. Furthermore, ∆ ≥ 0, so that ∆it is well defined for t ∈ R:
writing ∆ = exp(h) for the self-adjoint operator h = log ∆, we have ∆it = exp(ith).
Theorem 4.38 (Tomita and Takesaki) Let M ⊂ B(H) be a von Neumann al-
gebra in standard form, let S be the closure of the antilinear operator S0 defined by
S0 (aΩ) = a∗ Ω, and let S = J∆1/2 be the polar decomposition of S. Then:
• M 0 = JM J ≡ {JaJ | a ∈ M }.
• For each t ∈ R and a ∈ M , the operator αt (a) := ∆it a∆−it lies in M .
• The map t 7→ αt is a group homomorphism from R to Aut(M ) (i.e., the group
of all automorphisms of M ), which is continuous in that for each a ∈ M the
function t 7→ αt (a) from R to M is continuous w.r.t. the σ-weak topology.
This theorem was originally stated and proved in the language of Hilbert algebras
[33, Vol. II]; for a relatively short proof that avoids this formalism, see [1, §2.5].
The exercises illustrate this deep result in the context of some simple examples.
4 VON NEUMANN ALGEBRAS 53

4.14 Connes’s “cocycle Radon–Nikodym Theorem”

The image of R in Aut(M ) by α is called the modular group of M associated
with the cyclic and separating vector Ω (or rather, with the corresponding σ-weakly
continuous faithful state ω). Simple examples (as in the exercises) show that the
modular group explicitly depends on the vector Ω. In his thesis [3], Connes analyzed
the dependence of α on Ω. To state the simplest version of his result (to which the
case involving two different Hilbert spaces can be easily reduced), assume that H
contains two different vectors Ω1 and Ω2 , each of which is cyclic and separating for
(i)
M . We write αt for the time-evolution derived from Ωi , i = 1, 2.
Theorem 4.39 There is a family Ut of unitary operators in M , t ∈ R, such that
(1) (2)
αt (a) = Ut αt (a)Ut∗ ; (4.45)
Ut+s = Us αs(2) (Ut ). (4.46)
The proof of this theorem (Connes’s favourite [7]) is based on the following idea.
Extend M to Mat2 (M ), i.e., the von Neumann algebra of 2 × 2 matrices with entries
in M , and let Mat2 (M ) act on H2 = H ⊕ H in the obvious way. Subsequently, let
Mat2 (M ) act on H4 = H ⊕ H ⊕ H ⊕ H = H2 ⊕ H2 by simply doubling the action on
H2 . The vector Ω = (Ω1 , 0, 0, Ω2 ) ∈ H4 is then cyclic and separating for Mat2 (M ),
with corresponding ∆ = diag(∆1 , ∆4 , ∆3 , ∆2 ) (exercise). Here ∆1 and ∆2 are just
the operators on H originally defined by Ω1 and Ω2 , respectively, and  ∆3 and∆4
a11 a12
are certain operators on H. Denoting elements of Mat2 (M ) by a = ,
a21 a22
!
  (1)
a 0 α̃ (a) 0
∆it ∆−it = t
(2) , (4.47)
0 a 0 α̃t (a)
∆1 a11 ∆−it ∆it1 a12 ∆−it
 it 
(1) 1 4
α̃t (a) := ; (4.48)
∆it4 a21 ∆−it
1 ∆it4 a22 ∆−it
4
∆3 a11 ∆−it ∆it3 a12 ∆−it
 it 
(2) 3 2
α̃t (a) := . (4.49)
∆it2 a21 ∆−it
3 ∆it2 a22 ∆−it
2

But by Theorem 4.38, the right-hand side of (4.47) must be of the form diag(b, b) for
(1) (2)
some b ∈ Mat2 (M ), so that α̃t (a) = α̃t (a). This allows us to replace ∆it4 a22 ∆−it
4
in (4.48) by ∆it2 a22 ∆−it
2 . We then put U t = ∆it −it
1 ∆4 , which, unlike either ∆it
1 or
−it (1)
∆4 , lies in M , because each entry in α̃t (a) must lie in M if all the aij do, and
here we have taken a12 = 1. All claims of the theorem may then be verified using
elementary computations with 2 × 2 matrices. For example, combining
     
a 0 0 1 0 0 0 0
=
0 0 0 0 0 a 1 0
(1) (1) (1)
with the property α̃t (ab) = α̃t (a)α̃t (b), we recover (4.45). Using the identity
    
0 Ut 0 1 0 0
=
0 0 0 0 0 Ut ,
and evolving each side to time s, we arrive at (4.46). A ‘Proof from the Book’ !
4 VON NEUMANN ALGEBRAS 54

4.15 Connes’s parametrization of type iii factors

Theorem 4.39 shows that the modular group of M is independent of the vector Ω
used to define it, “up to inner transformations”. To make this precise, we say that
an automorphism γ : M → M is inner if there exists a unitary element u ∈ M
such that γ(a) = uau∗ for all a ∈ M . The inner automorphisms of M form a
normal subgroup Inn(M ) of the group Aut(M ) of all automorphisms, with quotient
Out(M ) = Aut(M )/Inn(M ). Hence the image π(α(R)) of the modular group in
Out(M ) under the canonical projection π : Aut(M ) → Out(M ) is independent of
Ω, and invariants of this image will be invariants of M itself. Such invariants are
trivial if M is a factor of type i or ii, since in that case π(α(R)) = {e}; to see this
in the finite case (i.e., type in or type ii1 ), take a finite trace τ on M and check that
∆ = 1 for πτ (M ) ∼ = M . For the semifinite but not finite case (i.e., type i∞ or type
ii∞ ), a slight generalization of the gns-construction leads to the same conclusion.
To find invariants for type iii factors, we therefore need to extract information
from the modular group t 7→ αt “up to inner automorphisms”. Recall Def. 2.15.
Definition 4.40 Let α : R → Aut(M ) be an action of R on a von Neumann algebra.
• The Arveson spectrum sp(α) of α consists of all k ∈ R for which there is a
sequence (xn ) in M with kxn k = 1 and limn→∞ kαt (xn ) − eikt xn k = 0 ∀t ∈ R.
• Let M α = {x ∈ M | αt (x) = x ∀t ∈ R}. If e ∈ P (M α ) and Me := {x ∈ M |
xe = ex = x}, then αt : M → M restricts to αte : Me → Me , defining a (group)
homomorphism αe : R → Aut(Me ), t 7→ αte . The Connes spectrum of α is
Γ (α) := 06=e∈P (M α ) sp(αe ) ⊂ R, or, equivalently, Γ(α) := exp(Γ0 (α)) ⊂ R+
0
T
∗.

The Connes spectrum Γ(α) is a closed subgroup of R+ ∗ , which has the great virtue
0 0
that if π(α(R)) = π(α (R)), then Γ(α) = Γ(α ). So if α is the modular group of
M with respect to some state ω, then Γ(α) is independent of ω and may therefore
be called Γ(M ). This invariant can also be defined through the usual spectrum of
self-adjoint operators on Hilbert space. To this effect, Connes defined and proved
\ \
S(M ) := σ(∆ω ) = σ(∆ϕe ), (4.50)
ω 06=e∈P (M α )

where the first intersection is over all σ-weakly continuous faithful states ω on M ,
whereas in the second one takes a fixed σ-weakly continuous faithful state ϕ on M ,
and restricts it to ϕe := ϕ|Me . Furthermore, ∆ω denotes the operator ∆ on Hω ,
defined w.r.t. the usual cyclic unit vector Ωω of the gns-construction, etc. If M is
a type i or ii factor (on a separable Hilbert space) one has S(M ) = {1}, whereas
0 ∈ S(M ) iff M is type iii. Connes showed that Γ(M ) = S(M ) ∩ R+ ∗ , and the known
classification of closed subgroups of R+
∗ yields his parametrization of type iii factors:
Definition 4.41 Let M be a type iii factor. Then M is said to be of type:
• iii0 if Γ(M ) = {1};
• iiiλ , where λ ∈ (0, 1), if Γ(M ) = λZ ;
• iii1 if Γ(M ) = R+
∗.
4 VON NEUMANN ALGEBRAS 55

Bonus exercises for Lecture 8.67

1. Let M ⊂ B(H) be a von Neumann algebra. Show that Ω ∈ H is cyclic for M 0
iff it is separating for M (equivalently, Ω is cyclic for M iff it is separating for
M 0 ). Hence M is in standard form iff Ω is cyclic for both M and M 0 .
2. Let the von Neumann algebra M = Mn (C) act on the Hilbert space H =
Mn (C) (with inner product (a, b) = Tr (a∗ b)) by left multiplication.
√
(a) Take Ω = 1n / n as a unit vector in H. Show that Ω is cyclic and
separating for M , that Sa = a∗ , J = S, and ∆ = 1. Conclude that
αt (a) = a for all t ∈ R, a ∈ M (so that α : R → Aut(M ) is trivial).
(b) Now take an arbitrary injective density matrix ρ̂, and define Ω = ρ̂1/2 .
(ei ) of Cn and there
In other words (diagonalizingPρ̂), there is an o.n.b. P
√
are numbers pi ∈ (0, 1) with i pi = 1 so that Ω = ni=1 pi |ei ihei | as a
unit vector in H (the choice pi = 1/n for all i gives the previous case).
Show that Ω is cyclic and separating for M , that SΨ = ρ̂−1/2 Ψ∗ ρ̂1/2 ,
JΨ = Ψ∗ , and ∆Ψ = ρ̂Ψρ̂−1 , where Ψ ∈ H. For a ∈ M , check that
Ja∗ JΨ = Ψa, so that, given ρ̂∗ = ρ̂, we have ∆ = ρ̂J ρ̂−1 J.
As in physics [18], write ρ̂ = exp(h) for some self-adjoint h ∈ M , and
write ã for Ja∗ J, so that ã ∈ M 0 if a ∈ M . Hence ∆ = exp(h − h̃) =
exp(h) exp(−h̃), since h and h̃ commute. Conclude that αt (a) = eith ae−ith
for all t ∈ R, a ∈ M (so that α : R → Aut(M ) is nontrivial, but inner).
3. Suppose a von Neumann algebra M admits a faithful finite trace τ . Show that
S = J (and hence ∆ = 1) for πτ (M ) ⊂ B(Hτ ) with respect to Ω = Ωτ .
4. In the proof of Theorem 4.39, check that the operator ∆ takes the diagonal
form ∆ = diag(∆1 , ∆4 , ∆3 , ∆2 ), where ∆1 and ∆2 are the operators on H
originally defined by Ω1 and Ω2 .
In the exercises below, M is a factor on a separable Hilbert space.
5. Prove that a type i factor M ⊂ B(H) is isomorphic to B(K), for some Hilbert
space K. Hint. Show that:
(a) pM p = C · p whenever p is a minimal projection.
(b) Any two minimal projections in M are equivalent.
P
(c) 1 = i∈I pi for some orthogonal family (pi )i∈I of minimal projections;
(d) If u∗i ui = p1 and ui u∗i = pi , then pi xpj = λij ui u∗j for some λij ∈ C, x ∈ M .
Finally, let K be Cn if |I| = n or K = `2 if I ∼ = N, with o.n.b. (pi )i∈I ,
and let H = p1 H. Show that there is a unique unitary u : K ⊗ H 0 → H
0

such that u(ei ⊗ ψ) = ui ψ for all ψ ∈ H 0 and i ∈ I, and finally prove that
u∗ M u = B(K) ⊗ 1H 0 .
67
If you like, submit nos. 5 and 6 by mail to A.J. Lindenhovius, imapp, fnwi, ru, Heyen-
daalseweg 135, 6525 aj Nijmegen, in order to replace your lowest mark for the earlier weeks.
4 VON NEUMANN ALGEBRAS 56

6. Show that a ii∞ factor M is isomorphic to N ⊗ B(`2 ), where N is type ii1 :

(a) Pick a nonzero finite projection p in M and show that thereP exists an
orthogonal family (pi )i∈N of projections with pi ∼ p for all i and i pi = 1.
In what follows, ui is a partial isometry in M such that u∗i ui = p and
ui u∗i = pi .
(b) Let H 0 = pH and consider N = pM p as a von Neumann algebra on
B(H 0 ). Prove that N is a factor, type ii1 factor.
(c) Show that the operator u : H 0 ⊗ `2 → H defined by u(ψ ⊗ ei ) = ui ψ is
unitary and satisfies uN ⊗ B(`2 )u∗ = M .
4 VON NEUMANN ALGEBRAS 57

Appendix 2: Trace class and Hilbert–Schmidt operators

We now deepen our understanding of the trace on B(H), introduced in §2.10. Our
aim is to prove the duality B1 (H)∗ = B(H), which in detail reads as follows.68
Theorem 4.42 1. As usual, for ρ̂ ∈ B1 (H), define ρ ∈ B(H)∗ by
ρ(a) := Tr (ρ̂a). (4.51)
The map ρ̂ 7→ ρ from B1 (H) to B(H)∗ is an isomorphism of Banach spaces
onto its image,69 which therefore is a norm-closed subspace of B(H)∗ .
2. The same map from B1 (H) to B0 (H)∗ is an isomorphism of Banach spaces.
3. For a ∈ B(H), define ã ∈ B1 (H)∗ by
ã(ρ̂) := Tr (ρ̂a). (4.52)
One then has B(H) ∼
= B1 (H)∗ as Banach spaces.
The proof will be given at the end of this appendix, preceded by some preliminaries.
Let {ei }i be an o.n.b. in a Hilbert space H. For a ∈ B(H) and a ≥ 0, define
X
Tr (a) := (ei , aei ). (4.53)
i

Elementary computations (cf. [28]) give

Tr (a∗ a) = Tr (aa∗ ) (4.54)
for all a ∈ B(H) (whether or not these expressions are finite), and
Tr (uau∗ ) = Tr (a) (4.55)
for a ≥ 0 and u unitary. In particular, (4.53) is independent of the choice of the
basis. For any a ∈ B(H), define the trace norm of a by
√
kak1 := Tr (|a|) = Tr ( a∗ a), (4.56)
and the trace class B1 (H) ⊂ B(H) by
B1 (H) := {a ∈ B(H) | kak1 < ∞}. (4.57)
The Hilbert–Schmidt norm is defined by
kak22 := Tr (|a|2 ) = Tr (a∗ a), (4.58)
with associated Hilbert–Schmidt class B2 (H) ⊂ B(H) defined by
B2 (H) := {a ∈ B(H) | kak2 < ∞}. (4.59)
With our usual (Dirac) notation |vihw| for the operator u 7→ (w, u)v, it follows that
k|vihw| k1 = k|vihw| k2 = kvk kwk. (4.60)
68
This may be seen as a noncommutative generalization of the dualities `∗0 ∼
= `1 with (`1 )∗ ∼
= `∞ .
69
An isomorphism of Banach spaces is by definition isometric.
4 VON NEUMANN ALGEBRAS 58

Proposition 4.43 1. For all a ∈ B(H) one has

kak ≤ |ak2 ≤ kak1 . (4.61)

2. Every trace-class operator and every Hilbert–Schmidt operator is compact, so

that one has the inclusions

B1 (H) ⊂ B2 (H) ⊂ B0 (H) ⊂ B(H). (4.62)

3. The Hilbert–Schmidt operators B2 (H) form a Hilbert space in the inner product

(a, b) := Tr (a∗ b); (4.63)

in particular, B2 (H) is a Banach space in the norm (4.58).

4. B2 (H) is a two-sided ∗ -ideal in B(H).

5. For a ∈ B(H) and b ∈ B1 (H) one has

|Tr (ab)| ≤ kakkbk1 . (4.64)

6. B1 (H) is a Banach space in the norm (4.56).

7. B1 (H) is a two-sided ∗ -ideal in B(H).

8. If a ∈ B1 (H), then (4.53) converges absolutely and is independent of the basis

chosen.

9. For either a and b in B2 (H), or a ∈ B1 (H) and b ∈ B(H), or a ∈ B(H) and

b ∈ B1 (H), one has
Tr (ab) = Tr (ba). (4.65)

Proof.

1. Although kbk ≥ kbvk for all b ∈ B(H) and all unit vectors v, for every  > 0
there is a v ∈ H of norm 1 such that kbk2 ≤ kbvk2 + . Put b = (a∗ a)1/4 , and
note that k(a∗ a)1/4 k2 = kak. Completing v to a basis {ei }i , we have
X
kak = k(a∗ a)1/4 k2 ≤ k(a∗ a)1/4 vk2 +  ≤ k(a∗ a)1/4 ei k2 +  = kak1 + .
i

Since this holds for all  ≥ 0, one has

kak ≤ |ak1 . (4.66)

Similarly, kak ≤ |ak2 . The remaining inequality in (4.61) will follow from the
next item.
4 VON NEUMANN ALGEBRAS 59

P
2. Let aP ∈ B1 (H). Since i (ei , |a|ei ) < ∞, for every  > 0 we can find n such
that i>n (ei , |a|ei ) < . Let pn be the projection onto the linear span of
{ei }i=1,...,n . Using (4.66), we have

kp⊥
n |a| k = kp⊥
1/2 2 ⊥ ⊥ ⊥
n |a|pn k ≤ kpn |a|pn k1 < .

Since p⊥
n = 1 − pn , it follows that pn |a|
1/2
→ |a|1/2 in the norm topology. Using
continuity of the involution and of multiplication, it follows that pn |a|pn → |a|
in norm. Since each operator pn |a|pn obviously has finite rank, |a| is compact.
But a has polar decomposition a = u|a|, so that a is compact (since B0 (H) is
a two-sided ideal in B(H)). The proof for B2 (H) is analogous.
P p
From the spectral theorem for compact operators, one now has kakp = λi
for p = 1, 2, where the λi ≥ 0 are the eigenvalues of |a| ≥ 0. The second
inequality in (4.61) is now immediate. Hence (4.61) has been proved.
3. The polarization formula (a + b)∗ (a + b) + (a − b)∗ (a − b) = 2(a∗ a + b∗ b)
yields the inequality (a + b)∗ (a + b) ≤ 2(a∗ a + b∗ b), since c∗ c ≥ 0 for all c,
including c = a − b. Hence a, b ∈ B2 (H) implies a + b ∈ B2 (H). With the
obvious λa ∈ B2 (H) when a ∈ B2 (H) for all λ ∈ C, this proves that B2 (H)
is a vector space. The sesquilinear form (4.63) is clearly positive semidefinite,
so the Cauchy–Schwarz inequality holds, and B2 (H) is a pre-Hilbert space.
In fact, by (4.61) the form (4.63) is positive definite, since k · k is a norm.
Finally, we show that B2 (H) is complete. Pick a basis {ei }i∈I in H (not
necessarily countable), and note
P that B2 (H) is the closure of the linear span of
all operators of the form a = i,j aij |ei ihej |. This is because of the continuity
of the inclusions in (4.62) and the fact that B0 (H) is itself the closure of this
linear span. An easy calculation gives
X X
k aij |ei ihej |k22 = |aij |2 .
i,j i,j

Hence B2 (H) is isomorphic to the space of square-summable sequences indexed

by J = I × I; this is well known to be a Hilbert space for any index set J.
4. With a ∈ B2 (H) one has a∗ ∈ B2 (H) by (4.54). Furthermore, for a unitray
u it is immediate from the definition of the Hilbert–Schmidt norm that ua ∈
B2 (H). By the linearity of B2 (H), this implies ba ∈ B2 (H) for all b ∈ B(H).
Taking the star yields the same conclusion for ab.
5. Assume a ∈ B(H) and b ∈ B1 (H), and let b = u|b| be the polar decomposition
of b. We write Tr (ab) = ((au|b|1/2 )∗ , |b|1/2 ) in the inner product (4.63). This
makes sense, since by definition of B1 (H) b ∈ B1 (H) implies |b| ∈ B1 (H).
Hence |b|1/2 ∈ B2 (H) by definition of B2 (H). The latter being a ∗ -ideal in
B(H), one the also has (au|b|1/2 )∗ ∈ B2 (H). The Cauchy–Schwarz inequality
then yields
X
|Tr (ab)|2 ≤ k|b|1/2 k22 kau|b|1/2 k22 = kbk1 kau|b|1/2 ei k2 .
i
4 VON NEUMANN ALGEBRAS 60

Since kau|b|1/2 ei k ≤ kaukk|b|1/2 ei k and kauk ≤ kak, the claim follows.

6. Let a, b ∈ B1 (H), and let a + b = u|a + b| be the polar decomposition. Then

ka + bk1 = Tr (u∗ (a + b)) = Tr (u∗ a) + Tr (u∗ b).

Applying (4.64) with ku∗ k ≤ 1, one has ka + bk1 ≤ kak1 + kbk1 . Hence B1 (H)
is a vector space and k · k1 is a norm; it is positive definite by (4.61). We
now prove completeness of B1 (H). Let {an } be a Cauchy sequence in k · k1 .
By (4.61), this is also a Cauchy sequence for k · k, which converges to some
a ∈ B0 (H). Let a − an = u|a − an |, so that |a − an | = u∗ (a − an ). Writing pN
for the projection on the linear span of e1 , . . . , eN , one has, for N < ∞,
N
X
(ei , |a − an |ei ) = lim Tr (pN u∗ (am − an )) ≤ lim sup kam − an k1 ,
m→∞ m
i=1

where we used (4.64) to derive the inequality. Since the right-hand side is
independent of N , we can let N → ∞ on the left, to obtain

ka − an k1 ≤ lim sup kam − an k1 .

m

Since {an } is a Cauchy sequence, it follows that limn ka − an k1 = 0, so that the

sequence converges to a ∈ B1 (H) also in the trace norm. Note that a similar
proof can be given for B2 (H).

7. For u unitary, one has |ua| = |a|, so if a ∈ B1 (H) then ua ∈ B1 (H). By

linearity and the fact that each bounded operator is a linear combination of at
most four unitaires (exercise!), one has ba ∈ B1 (H) for all b ∈ B(H). Similarly,
|au| = u−1 |a|u0 , so by (4.55) one has au ∈ B1 (H), etc. Finally, if a = u|a| and
a∗ = ũ|a∗ | then |a∗ | = ũ∗ |a|u∗ , which is in B1 (H) by the previous argument in
this item, so that a∗ ∈ B1 (H).

8. It is clear from (4.64) that Tr (a) < ∞. To prove absolute convergence, with
a = u|a|1/2 |a|1/2 one has

|(ei , aei )| ≤ k|a|1/2 u∗ ei kk|a|1/2 ei k.

Taking the sum and using the Cauchy–Schwarz inequality, one obtains
X
|(ei , aei )| ≤ k|a|1/2 u∗ k2 k|a|1/2 k2 .
i

By the argument in the proof of (4.64) (with b replaced by a), the right-hand
side is finite. Independence of the basis is now an easy exercise.

9. It is enough to consider the case where a or b (as appropriate) is unitary, in

which case one uses (4.55) and item 4 or 7. 
4 VON NEUMANN ALGEBRAS 61

At last, we are now in a position to prove Theorem 4.42.

1. For ρ̂ ∈ B1 (H) and ρ ∈ B(H)∗ , it is clear from (4.64) and (4.65) that

kρk ≤ kρ̂k1 . (4.67)

For the converse inequality, write ρ̂ = u|a|, and estimate

kρ̂k1 = Tr (|ρ̂|) = Tr (u∗ ρ̂) = ρ(u∗ ) ≤ kρkku∗ k ≤ kρk.

Hence kρ̂k1 = kρk for ρ ∈ B(H)∗ .

2. For part 2 of the theorem we also need this equality for ρ ∈ B0 (H)∗ . To that
effect, and replace the above estimate by

Tr (pN |ρ̂|) = Tr (pN u∗ ρ̂) = ρ(pN u∗ ) ≤ kρkkpN u∗ k ≤ kρk. (4.68)

/ B0 (H) in general, one now has pN u∗ ∈ B0 (H).

The point is that, whereas 1 ∈
Letting N → ∞ then yields the desired result.
Now take ϕ ∈ B0 (H)∗ and a ∈ B2 (H) ⊆ B0 (H). Then

|ϕ(a)| ≤ kϕkkak ≤ kϕkkak2 .

Since B2 (H) is a Hilbert space, by Riesz–Fischer there is an operator ρ̂ ∈

B2 (H) such that ϕ(a) = Tr (ρ̂a) for all a ∈ B2 (H). Letting N → ∞ in (4.68),
this can be sharpened to ρ̂ ∈ B1 (H). Hence the map ρ̂ 7→ ρ is surjective onto
B0 (H)∗ and isometric, so it must be an isomorphism of Banach spaces.

3. Now let a ∈ B(H) and ã ∈ B1 (H)∗ . From (4.64) and (4.65) one has kãk ≤ kak.
For the converse, take v, w ∈ H. On the one hand, one has

ã(|wihv|) = (v, aw),

and on the other one has

|ã(|wihv|)| ≤ kãkk|wihv|k1 .

Combining these and using (4.60) yields

|(v, aw)| ≤ kãk kvkkwk. (4.69)

Taking v, w of unit length and using the identity

kak = sup{|(v, aw)| | v, w ∈ H, kvk = kwk = 1} (4.70)

for any a ∈ B(H) gives kak ≤ kãk, so that kãk = kak.

A given ϕ ∈ B1 (H)∗ defines an operator a on H by ϕ(|wihv|) = (v, aw). Since
the operators |wihv| span B1 (H), one has ϕ = ã, and by (4.69) and (4.70) it
follows that a ∈ B(H). The proof of Theorem 4.42 is now complete. 
4 VON NEUMANN ALGEBRAS 62

Appendix 3
Half of Theorem 4.9 evidently follows from Theorem 4.10. The converse (‘if’) im-
plication uses a refinement of the gns-construction, where the state ω is assumed
to be σ-weakly continuous (such states are also called normal). In that case, using
the theory of σ-weakly closed ideals of von Neumann algebras, it can be shown that
πω (M ) coincides with πω (M )00 and hence is a von Neumann algebra [1, Thm. 2.4.24].
Since normal pure state on a von Neumann algebra may not exist, the ‘crazy’
Hilbert space Hc in the proof ofL Theorem 3.20 (see §3.10) must be replaced by the
even crazier direct sum Hec = ω∈Sn (M ) Hω , where this time the sum is over all
normal states on M . Similarly, in Lemma 3.8 one should now have a normal state
instead of a pure state. Otherwise, the proof that M has a faithful representation
as a von Neumann algebra on a Hilbert space is essentially follows the proof of
Theorem 3.20 (see Sakai’s own book [30, Thm. 1.16.7] for details). 

We next prove (4.37). The inclusion M ⊂ M ⊥⊥ is trivial. For the converse, pick
a∈/ M ; since M is a von Neumann algebra, it is σ-weakly closed, so its complement
M c in B(H) is σ-weakly open. Hence there are ϕ ∈ B(H)∗ and  > 0 such that the
open neighbourhood

O(a) := {b ∈ B(H) | |ϕ(a) − ϕ(b)| < }

of a entirely lies in M c . So |ϕ(a) − ϕ(b)| ≥  for all b ∈ M . This implies ϕ(b) = 0

by linearity in b. Hence |ϕ(a)| ≥ , so a ∈/ M ⊥⊥ , hence M ⊥⊥ ⊂ M .
For (4.38), first note that M ⊥ is a norm-closed subspace of B(H)∗ = B1 (H),
which is a Banach space in the trace-norm (which coincides with the norm inher-
ited from B(H)∗ , since the injection B1 (H) ,→ B(H)∗ is an isometry). Hence the
quotient B(H)∗ /M ⊥ is a Banach space in the canonical norm

kϕ̇k := inf{kϕ + ψk | ψ ∈ M ⊥ }.

where ϕ̇ is the image of ϕ ∈ B(H)∗ under the canonical projection, and the norm
is the one in B(H)∗ . Let ϕ := ϕ  M be the restriction of ϕ ∈ B(H)∗ to M . It
is clear that the map ϕ 7→ ϕ̇ is well defined and is a linear bijection from M∗ to
B(H)∗ /M ⊥ . In fact, this map is isometric. Firstly, one trivially has

kϕ k = sup{|ϕ(a)| | a ∈ Mu } = inf sup{|ϕ(a) + ψ(a)| | a ∈ Mu },

ψ∈M ⊥

since ψ(a) = 0. But this is clearly majorized by

kϕ̇k = inf sup{|ϕ(a) + ψ(a)| | a ∈ B(H)1 },

ψ∈M ⊥

since now the supremum is taken over a larger set. Hence kϕ k ≤ kϕ̇k. Conversely,
for any ϕ ∈ B(H)∗ with kϕ̇k = 1 there exists, by a version of the Hahn–Banach
theorem, an a ∈ B(H) with â ∈ M ⊥⊥ , ϕ(a) = 1 and kak = 1. From (4.37) one then
infers that kϕ k ≥ |ϕ(a)| = 1 = kϕ̇k. This finishes the proof of Theorem 4.10. 
REFERENCES 63

References
[1] Bratteli, O. & Robinson, D.W. (1987). Operator Algebras and Quantum Statisti-
cal Mechanics. Vol. I: C ∗ - and W ∗ -Algebras, Symmetry Groups, Decomposition
of States. 2nd Ed. Berlin: Springer.

[2] Bratteli, O. & Robinson, D.W. (1981). Operator Algebras and Quantum Sta-
tistical Mechanics. Vol. II: Equilibrium States, Models in Statistical Mechanics.
Berlin: Springer.

[3] A. Connes, Une classification des facteurs de type iii, Ann. Sci. École Norm.
Sup. 6, 133–252 (1973).

[4] A. Connes, Classification of injective factors. Cases ii1 , ii∞ , iiiλ , λ 6= 1, Ann.
Math. 104, 73–115 (1976).

[5] A. Connes, Noncommutative Geometry (Academic Press, San Diego, 1994).

[6] A. Connes & M. Marcolli, Noncommutative Geometry, Quantum Fields and

Motives (AMS, 2008).

[7] G. Cornelissen, N.P. Landsman, and W. van Suijlekom, ‘The flashes

of insight never came for free’: Interview with Alain Connes,
Nieuw Archief voor Wiskunde 5/11, 250–256 (2010). Available at
www.math.ru.nl/∼landsman/ConnesNAW.pdf.

[8] J.B. Conway, A Course in Functional Analysis (Springer, New York, 1990).

[9] T. Coquand and B. Spitters, Constructive Gelfand duality for C*-algebras,

Math. Proc. Camb. Phil. Soc. 147, 339–344 (2009).

[10] J. Dixmier, Von Neumann Algebras (North-Holland, Amsterdam, 1981).

[11] R. Dudely, Real Analysis and Probability (Wadsworth & Brooks/Cole, Pacific
Grove, 1989).

[12] E.G. Effros and Z.-J. Ruan, Operator spaces (Oxford University Press, New
York, 2000).

[13] R. Haag, Local Quantum Physics (Springer, Berlin, 1992).

[14] P. T. Johnstone, Stone Spaces (Cambridge University Press, 1982).

[15] V.F.R. Jones, Subfactors and knots (AMS, 1991).

[16] R.V. Kadison and J.R. Ringrose, Fundamentals of the Theory of Operator Al-
gebras, Vol. I: Elementary Theory (Academic Press, New York, 1983).

[17] R.V. Kadison and J.R. Ringrose, Fundamentals of the Theory of Operator Al-
gebras, Vol. II: Advanced Theory (Academic Press, New York, 1986).
REFERENCES 64

[18] N.P. Landsman and Ch.G. van Weert, Real- and imaginary-time field theory
at finite temperature and density, Phys. Rep. 145, 141–249 (1987).

[19] J. Lurie, Math261y: von Neumann algebras (Harvard University), Lecture notes
(in progress) available at www.math.harvard.edu/∼lurie/261y.html.

[20] F. J. Murray and J. von Neumann, On rings of operators, Ann. Math. 37,
116–229 (1936).

[21] F. J. Murray and J. von Neumann, On rings of operators, ii, Trans. Amer.
Math. Soc. 41, 208–248 (1936).

[22] F. J. Murray and J. von Neumann, On rings of operators, iv, Ann. Math. 44,
716–808 (1943).

[23] J. von Neumann, Zur Algebra der Funktionaloperatoren und theorie der nor-
malen Operatoren, Math. Ann. 102, 370–427 (1929).

[24] J. von Neumann, Mathematische Grundlagen der Quantenmechanik (Springer,

Berlin, 1932).

[25] J. von Neumann, On rings of operators, iii, Ann. Math. 41, 94–161 (1940).

[26] J. von Neumann, On rings of operators. Reduction theory, Ann. Math. 50,
401–485 (1949).

[27] J. von Neumann, Collected Works. Vol. III: Rings of Operators, ed. A. H. Taub
(Pergamon Press, New York, 1961).

[28] G.K. Pedersen, Analysis Now, 2nd printing (Springer, New York, 1995).

[29] M. Reed and B. Simon, Methods of Modern Mathematical Physics, Vol. I: Func-
tional Analysis (Academic Press, San Diego, 1980).

[30] S. Sakai, C*-algebras and W*-algebras (Springer, Berlin, 1971).

[31] B. Simon, Representations of Finite and Compact Groups (AMS, 1996).

[32] M. Takesaki, Twenty-five years in the theory of type iii von Neumann algebras,
in: R.S. Doran (ed.), C ∗ -algebras: 1943–1993, A Fifty Year Celebration, pp.
233–239 (AMS, Providence, 1994).

[33] M. Takesaki, Theory of Operator Algebras, Vols. I–III (Springer, Heidelberg,

2002–2003).

[34] J. Weidmann, Lineare Operatoren in Hilberträumen, Vol. i (Teubner, Stuttgart,

2000). (Teubner, Stuttgart, 2003).

[35] S. Wright, Uniqueness of the injective iii1 factor (Lecture Notes in Mathematics
1413, Springer-Verlag, Berlin, 1989).