Professional Documents
Culture Documents
Math 582
John Roe
FALL 2015
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
2
Lecture 1
Introduction
This is a course about C ∗ -algebras. The study of these algebras combines techniques
from functional analysis, topology, algebra, and lattice theory.
The course text is Murphy, C ∗ -algebras and operator theory. This book gives
an economical presentation of the theory but it does not contain many examples.
Supplementary reading will therefore come from Davidson, C ∗ -algebras by example.
A background in analysis on the level of the MATH 501-503 sequence is assumed.
The specific items that we will need from the Math 503 syllabus are:
(1) Normed spaces. Banach spaces. Linear operators. Examples.
(2) Spaces of bounded linear operators. The uniform boundedness principle and
the open mapping theorem.
(3) Bounded linear functionals. Dual spaces. The Hahn-Banach extension the-
orem. Separation of convex sets.
(4) Spaces of continuous functions. Ascoli’s theorem, Stone-Weierstrass theo-
rem.
(5) Hilbert spaces. Perpendicular projections. Orthonormal bases. Self-adjoint
operators.
(6) Compact operators on a Hilbert space. Fredholms alternative. Spectrum
and eigenfunctions of a compact, self-adjoint operator.
We will need some extra background in functional analysis relating to duality in
locally convex spaces. We’ll wait until later to introduce some of these results (e.g.
the Krein-Milman theorem) but for now let’s just set up the basic terminology.
Definition 1.1. Let E be a vector space1. A seminorm on E is a function p : E →
R+ such that
(a) p(λx) = |λ|p(x) for all λ ∈ C, x ∈ E;
(b) p(x + y) 6 p(x) + p(y) for all x, y ∈ E.
A seminorm is a norm if, in addition, p(x) = 0 implies x = 0.
Many examples of topological vector spaces arise from the following construction.
Let {pα }α∈A be a family of seminorms on a vector space E. We can define a topology
on E by calling U ⊆ E open if for every x0 ∈ U there exist α1 , . . . , αn ∈ A and real
ε1 , . . . , εn > 0 such that
{x ∈ E : pαi (x − x0 ) < εi , (i = 1, . . . , n)} ⊆ U.
This topology makes E into a TVS; it is called the locally convex topology defined
by the seminorms {pα }. It is the weakest vector topology (fewest open sets) which
makes all the seminorms continuous. The family of seminorms is called separating
if for every nonzero x ∈ E there is a pα with pα (x) 6= 0: this is equivalent to the
Hausdorffness of the associated locally convex topology.
1All vector spaces in this course are over C unless otherwise specified.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
3
Example 1.2. Suppose that E and F are vector spaces equipped with a bilinear
or sesquilinear pairing
h·, ·i : E × F → C.
Then for each y ∈ F , the expression py (x) = |hx, yi| gives a seminorm on E. The
locally convex topology on E generated by these seminorms is called the σ(E, F )-
topology on E. In the same way we can define the σ(F, E)-topology on F . These
topologies are Hausdorff if the pairing is nondegenerate.
Example 1.3. Let E be a Banach space and let F = E ∗ , the dual space of E.
It follows from the Hahn-Banach theorem that the natural pairing between E and
E ∗ is nondegenerate. The σ(E, E ∗ )-topology on E is called the weak topology. The
σ(E ∗ , E)-topology on E ∗ is called the weak-star topology.
Proposition 1.4. (Banach-Alaoglu theorem) Let E be a Banach space. Then
the closed unit ball of E ∗ is weak-star compact.
Proof. Let B be the closed unit ball of E, and let B ∗ be the closed unit ball of
E ∗ equipped with its weak-star topology. Let D be the closed unit disc in the
complex plane. Let X be the space of all maps B → D, equipped with the product
topology; by Tychonoff’s theorem (1.13), X is a compact space. Observe that B ∗ is
identified with a subspace of X (the subspace consisting of linear maps) and that
this identification is topological: the product topology on X restricts to the weak-
star topology on B ∗ . It suffices therefore to show that B ∗ is closed in X. But B ∗
is just the subset of X consisting of those maps ϕ satisfying the uncountably many
relations
ϕ(λ1 x1 + λ2 x2 ) − λ1 ϕ(x1 ) − λ2 ϕ(x2 ) = 0;
for λ1 , λ2 , x1 , x2 fixed the LHS depends continuously on ϕ, so it cuts out a closed
subset, and the intersection of all these closed subsets is B ∗ .
Questions related to nonmetrizable vector topologies in operator algebras are
traditionally discussed using the language of nets (Moore-Smith convergence). We
review this language now (without detailed proofs: the reader may supply them as
exercises).
Definition 1.5. A directed set Λ is a partially ordered set in which any two elements
have an upper bound. A net in a topological space X is a function from a directed
set to X. Usually we denote a net like this: {xλ }λ∈Λ .
For example, a sequence is just a net based on the directed set N. Notice that in
general it is not required that there exist a least upper bound for any two elements.
Definition 1.6. Let U ⊆ X. A net {xλ }λ∈Λ in X is eventually in U if there is some
λ0 ∈ Λ such that xλ ∈ U for all λ > λ0 . It is frequently in U if it is not eventually
in X \ U .
Definition 1.7. If there is some point x ∈ X such that {xλ } is eventually in every
neighborhood of x, we say that {xλ } converges to x.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
4
As an exercise in manipulating this definition, the reader may prove that a space
X is Hausdorff if and only if every net converges to at most one point. Similarly
a function f : X → Y is continuous if and only if, whenever {xλ } is a net in X
converging to x, then net {f (xλ )} converges in Y to f (x). The basic idea here is to
use the system of open neighborhoods of a point, directed by (reverse) inclusion, as
the parameter space of a net.
Lemma 1.8. Let E be a vector space with a locally convex topology defined by
seminorms pα . Then a net {xλ } in E converges to x if and only if pα (xλ − x)
converges to 0 for all α.
Remark 1.9. Watch out for the fact that a convergent net in a metric space need
not be bounded! For example, consider the directed set Λ = {1 − 1/n : n ∈
N} ∪ {3 − 1/n : n ∈ N} with the usual ordering. A convergent net parameterized by
Λ can do “whatever it wants” on the first segment (from 0 to 1) of the directed set.
A subnet of a net is defined as follows. Let {xλ }λ∈Λ be a net based on the directed
set Λ. Let Ξ be another directed set. A function λ : Ξ → Λ is called a final function
if, for every λ0 ∈ Λ, there is ξ0 ∈ Ξ such that λ(ξ) > λ0 whenever ξ > ξ0 . For such a
function the net {xλ(ξ) }ξ∈Ξ is a net based on the directed set Ξ; it is called a subnet
of the originally given one. Notice that this notion of subnet is more relaxed than
the usual one of subsequence; we allow some repetition and backtracking.
Definition 1.10. A net in X is universal if, for every subset A of X, the net is
either eventually in A or eventually in the complement of A.
Lemma 1.11. Every net has a universal subnet.
Proof. This requires the axiom of choice. See the appendix to this lecture for an
account of the proof.
Proposition 1.12. Let X be a topological space. The following are equivalent:
(a) X is compact.
(b) Every universal net in X converges.
(c) Every net in X has a convergent subnet.
This makes possible a rather short proof of Tychonoff’s theorem.
Theorem 1.13. A product of compact topological spaces is compact.
Q
Proof. Let X = Xα be such a product, πα the coordinate projections. Pick a
universal net {xλ } in X; then a one-line proof shows that each of the nets {πα (xλ )}
is universal in Xα , and hence is convergent (a universal net with a convergent subnet
must itself be convergent), say to xα . But then by definition of the product topology
the net {xλ } converges to the point x whose coordinates are xα .
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
5
Proof. (of Lemma 1.11.) This depends unavoidably on the axiom of choice.
Let N : D → X be a net in X, parameterized by a directed set D. A collection
F of subsets of X will be called an N -filter if N is frequently in every member
of F , and if F is closed under finite intersections and the formation of supersets.
Such objects exist: for example, {X} is an N -filter. The collection of N -filters is
partially ordered by inclusion, and every chain in this partially ordered set has an
upper bound (the union). Thus Zorn’s Lemma provides a maximal N -filter; call it
F0 .
Suppose that S ⊆ X has the property that N is frequently in A ∩ S for every
A ∈ F0 . Then the union of F0 with the set of all sets A ∩ S, A ∈ F0 , is again an
N -filter. By maximality we deduce that S itself belongs to F0 .
We will use this property to construct the desired universal subnet. Let D0 be
the collection of pairs (A, i) with A ∈ F0 , i ∈ D, and N (i) ∈ A; it is a directed set
under the partial order
(B, j) > (A, i) ⇔ B ⊆ A, j > i.
The map (A, i) 7→ i is final so defines a subnet N 0 of the net N . We claim that this
subnet is universal.
Let S ⊆ X have the property that N 0 is frequently in S. Let A ∈ F0 and let
i be arbitrary. By definition, there exist B ∈ F0 , B ⊆ A, and j > i, such that
N (j) = N 0 (B, j) ∈ B ∩ S ⊆ A ∩ S. We conclude that N is frequently in A ∩ S for
every A ∈ F0 and hence, as observed above, that S itself belongs to F0 .
Now let S be arbitrary. It is not possible that N 0 be frequently both in S and in
X \ S, for then (by the above) both S and X \ S would belong to F0 , and then their
intersection, the empty set, would do so as well, a contradiction. Thus N 0 fails to
be frequently in one of these sets, which is to say that it is eventually in the other
one.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
6
Lecture 2
Hilbert space
Thus
kT k2 6 kT ∗ T k 6 kT ∗ kkT k.
This gives kT k 6 kT ∗ k; but since ∗ is an involution we deduce that kT k = kT ∗ k
and also that
(2.1) kT ∗ T k = kT k2 .
Equation 2.1 is called the C ∗ -identity and will become one of the axioms for abstract
C ∗ -algebras. The operator T ∗ is called the adjoint of T .
We recall some standard terminology for special kinds of operators on a Hilbert
space.
• T ∈ B(H) is selfadjoint if T = T ∗ .
• U ∈ B(H) is unitary if U U ∗ = U ∗ U = I. Geometrically, U is an isometric
isomorphism from H to itself.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
7
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
8
where the series converges in norm. The spectrum of T consists of the set {λj }∪{0};
each λj is an eigenvalue with a finite-dimensional eigenspace.
In brief, a compact selfadjoint operator can be “diagonalized” relative to a suitable
orthonormal basis.
Definition 2.7. Let T be a selfadjoint operator on a Hilbert space H. One says
that T is positive if the quadratic form defined by T is positive semidefinite, that
is, if
hT v, vi > 0 for all v ∈ H.
Clearly, the positive operators form a cone — the sum of two positive operators
is positive, as is any positive real multiple of a positive operator.
Example 2.8. If T = S ∗ S, then hT v, vi = hSv, Svi > 0 and thus T is positive. In
fact, this condition is necessary and sufficient for positivity. To see this, we will need
a version of the spectral theorem that applies to all selfadjoint operators, not merely
compact ones. This will be proved as a consequence of our work on C ∗ -algebras.
It is helpful to know something about Hilbert-Schmidt and trace-class operators.
These are important subclasses of the compacts.
Let H and H 0 be (separable, infinite dimensional) Hilbert spaces, and choose
orthonormal bases (ei ) and (e0j ) in H and H 0 . A bounded linear operator A : H → H 0
can be represented by an “infinite matrix” with coefficients2
cij (A) = he0j , Aei i.
Proposition 2.9. The quantity
X
kAk2HS = |cij (A)|2 ∈ [0, ∞]
i,j
which is certainly independent of the choice of basis in H 0 . But since cij (A) =
cji (A∗ ), kAk2HS = kA∗ k2HS which is independent of the choice of basis in H by the
same argument.
Definition 2.10. An operator A such that kAkHS < ∞ is called a Hilbert-Schmidt
operator, and kAkHS is called its Hilbert-Schmidt norm.
Proposition 2.11. Hilbert-Schmidt operators have the following properties
2Ofcourse, not every such infinite matrix represents a bounded operator; but this does not
matter here.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
9
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
10
Proof. Think about the Hilbert-Schmidt case first. Choose an orthonormal basis
(ei ) for H, and write
X
Tr(AB) = hei , ABei i
i
X
= hA∗ ei , Bei i
i
X
= cij (B)cij (A) (by Parseval’s theorem).
i,j
(It is not the case that every continuous kernel defines a trace-class operator;
but every smooth kernel on a compact manifold defines a trace-class operator with
respect to the smooth measure class).
Remark 2.19. We can use Hilbert-Schmidt operators to develop the theory of Hilbert
space tensor products. Specifically, let H and H 0 be Hilbert spaces. Let HS(H, H 0 )
be the space of Hilbert-Schmidt operators from H to H 0 ; it is a Hilbert space.
When H = C this Hilbert space is canonically identified with H 0 ; when H 0 = C it is
canonically identified with the dual space H ∗ . This leads us to make the following
definition: the Hilbert space tensor product of H and H 0 is
H ⊗ H 0 = HS(H ∗ , H 0 ).
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
11
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
12
Lecture 3
Fredholm theory
The material in this lecture goes back to the roots of functional analysis in the
theory of integral equations. It also provides a fundamental example in C ∗ -algebra
theory, specifically relating to the Calkin algebra, which is the quotient B(H)/K(H)
of the bounded operators on Hilbert space by the ideal of compact operators.
Definition 3.1. Let V and W be Hilbert spaces and let T : V → W be a bounded
linear operator. We say T is a Fredholm operator if
(a) The kernel of T has finite dimension,
(b) The range of T has finite codimension.
Proposition 3.2. The kernel and the range of a Fredholm operator on Hilbert space
are closed subspaces.
Proof. The kernel of any bounded linear operator is closed, since it is the inverse
image of a closed set, namely {0}, under a continuous map. As for the range, let
T : V → W be Fredholm. Let {w1 , . . . , wn } be a basis for a complement of Im T in
W . Define a bounded linear operator
Xn
⊥ n
L : (Ker T ) ⊕ C → W, (v, λ1 , . . . , λn ) 7→ T v + λi wi .
i=1
−1
L is bijective, and therefore has a bounded inverse M = L by the Closed Graph
Theorem. Then Im T is the inverse image M −1 ((Ker T )⊥ ) of the closed subspace
(Ker T )⊥ , so it is closed.
Definition 3.3. The index Index(T ) of a Fredholm operator T is the difference of
dimensions Nullity T − Corank T . (By definition, the corank of T is the dimension
of a complementary subspace to Im(T ), i.e. it is dim(Im(T )⊥ ) = dim(Ker(T ∗ )))
The rank-nullity theorem can be restated as follows: if V, W are finite-dimensional
and T : V → W is a linear map, then Index(T ) = dim V − dim W . In other words,
the index does not depend on T at all! In particular, an operator from a finite-
dimensional vector space to itself must have index zero. This statement is not true
for maps from an infinite-dimensional space to itself, as the following important
example shows.
Example 3.4. Let V = W = `2 , the Hilbert space of square-summable sequences.
Let T : V → W be the linear operator defined by
T (a0 , a1 , a2 , a3 , . . .) = (0, a0 , a1 , a2 , . . .)
called the unilateral shift. Clearly, Nullity T = 0 while Corank T = 1, so T is a
Fredholm operator of index −1. The adjoint operator T ∗ (the unilateral backward
shift) defined by
T ∗ (a0 , a1 , a2 , a3 , . . .) = (a1 , a2 , a3 , a4 , . . .)
has index +1.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
13
Lemma 3.5. Let H be a Hilbert space and let S ∈ B(H) with kSk < 1. Then I − S
is an invertible operator.
Proof. Define R by the series
∞
X
2
R = I + S + S + ··· = S n.
n=0
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
14
By our assumption, (T11 + L11 ) is invertible (Corollary 3.6). Now we are going per-
form “elementary row and column operations” on the matrix T + L. The operations
we want to carry out are these:
−1
• Add
−L10 (T11 + L11 ) −1times row 2 to row 1; that is, multiply on the left by
1 −L10 (T11 + L11 )
.
0 1
−1
• Add −(T 11 + L11 ) L01 times column
2 to column 1; that is, multiply on the
1 0
right by .
−(T11 + L11 )−1 L01 1
These operations are effected by invertible matrices, so they don’t change the di-
mension of the kernel or the codimension of the image (and in particular they don’t
change the Fredholm index). Their effect is to reduce T + L to the matrix
L00 − L10 (T11 + L11 )−1 L01
0
.
0 T11 + L11
The index of this diagonal matrix is clearly the sum of the indices of the diagonal
entries. But the first diagonal entry is just a linear transformation between finite-
dimensional vector spaces, so its index is dim V0 − dim W0 = Index(T ), and the
second entry is invertible so it has index zero. Thus T + L is Fredholm and has the
same index as T , completing the proof.
Proposition 3.8. (Atkinson’s theorem) Let T be a bounded operator on a Hilbert
space H. The following conditions are equivalent
(a) T is Fredholm.
(b) T is invertible modulo finite-rank operators: there is a bounded operator S such
that I − ST and I − T S are of finite rank.
(c) T is invertible modulo compact operators: there is a bounded operator S such
that I − ST and I − T S are compact operators.
Proof. Suppose that T is Fredholm (a). Then T maps the orthogonal comple-
ment (Ker(T ))⊥ bijectively onto Im(T ). Let Q be the inverse map from Im(T )
to (Ker(T ))⊥ ; by the Closed Graph Theorem, Q is a bounded operator. Let P be
the orthogonal projection from H onto the closed subspace Im(T ) and let S = QP .
Then by construction, I − T S and I − ST are the orthogonal projections onto
Im(T )⊥ and Ker(T ) respectively. Since these are finite-dimensional, the associated
projections have finite rank. Thus T is invertible modulo finite-rank operators (b).
It is obvious that (b) implies (c). Suppose (c), that T is invertible modulo com-
pacts, and let S be such that I − ST ∈ K and I − T S ∈ K. There is a finite rank
operator F such that kI − ST − F k < 21 . By Lemma 3.5, this implies that ST + F
is invertible. Consider now the identity map
I = (ST + F )−1 (ST + F )
when restricted to the kernel of T ; the restriction of ST + F to Ker(T ) has finite
rank, whence the restriction of I to Ker(T ) has finite rank, and thus Ker(T ) is
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
15
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
16
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
17
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
18
Lecture 4
Banach algebras
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
19
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
20
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
21
must converge on D, so for each λ ∈ D the set Sλ = {λn an } has the property
that ϕ(Sλ ) is bounded for all ϕ ∈ A∗ . In other words, Sλ ⊂ A is point bounded
(when considered as a set of linear functionals on A∗ ). By the Uniform Boundedness
Principle (with E = A∗ ), Sλ is uniformly bounded, i.e., it is a bounded subset of
A ⊆ A∗∗ . Consequently
lim sup kan k1/n = |λ|−1 lim sup kλn an k1/n 6 |λ|−1 .
This holds for all λ ∈ D, that is, whenever |λ| < spr(a)−1 , so we finally get
lim sup kan k1/n 6 spr(a) as required.
Remark 4.21. P The basic idea of the above proof is the following: the series λ 7→
−1
(1 − λa) = an λn has radius of convergence equal to spr(a)−1 . We apply the
Cauchy-Hadamard formula for the radius of convergence (supposing that this is
valid for Banach space valued holomorphic functions) to get
1
spr(a)−1 = .
lim sup kan k1/n
The point of the argument we gave is to use the functionals ϕ to avoid having
to develop the theory of Banach valued holomorphic functions, reducing to the
scalar case. Essentially, we went through part of the proof of the Cauchy-Hadamard
formula and then applied the uniform boundedness principle.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
22
Lecture 5
∗
C -algebra basics and the Gelfand-Naimark theorem
Gelfand used the theory of maximal ideals to analyze the structure of commutative
unital Banach algebras. If A is such an algebra then its Gelfand dual A b is the space
of maximal ideals of A; equivalently, by 4.16 and 4.13 above, it is the space of
algebra-homomorphisms A → C. Every such homomorphism is of norm 6 1, as we
saw, so Ab may be regarded as a subset of the unit ball of the Banach space dual
∗
A of A (namely, it is the subset comprising all those linear functionals α which are
also multiplicative in the sense that α(xy) = α(x)α(y)). This is a weak-star closed
subset of the unit ball of A∗ , and so by the Banach-Alaoglu Theorem 1.4 it is a
compact Hausdorff space in the weak-star topology. If a ∈ A then we may define a
continuous function â on A b by the usual double dualization:
â(α) = α(a).
In this way we obtain a contractive3 algebra-homomorphism G : a 7→ â, called the
Gelfand transform, from A to C(A),b the algebra of continuous functions on the
Gelfand dual.
Theorem 5.1. Let A be a commutative unital Banach algebra. An element a ∈ A
is invertible if and only if its Gelfand transform â = Ga is invertible. Consequently,
the Gelfand transform preserves spectrum: the spectrum of Ga is the same as the
spectrum of a.
Proof. If a is invertible then Ga is invertible, since G is a homomorphism. On the
other hand, if a is not invertible then it is contained in some maximal ideal m,
which corresponds to a point of the Gelfand dual on which Ga vanishes. So Ga isn’t
invertible either.
Example 5.2. Let A = `1 (Z), the Banach algebra of summable two-sided sequences
under convolution. Each such sequence (a n ) may be identified with the correspond-
an z n , z ∈ T, and in this way we obtain
P
ing absolutely convergent Fourier series
a homomorphism
A → C(T).
In fact, this homomorphism is the Gelfand transform: the dual space A b is identified
with the circle. (Each character on A is determined by what it does to the generator
of the group Z, and it must send this generator to a complex number of modulus
one.) Thus the Gelfand transform generalizes some of the basic ideas of Fourier
analysis.
Remark 5.3. Norbert Wiener proved that if a continuous function f is nowhere
vanishing and has an absolutely convergent Fourier series, then its pointwise inverse
has an absolutely convergent Fourier series too. This is a classic application of “soft
analysis”: it follows immediately from Theorem 5.1 above.
3This means kGak 6 kak.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
23
The Gelfand theory works for all commutative Banach algebras: an involution is
not required. However, in that generality the Gelfand transform need not be either
injective or surjective. (We have already seen that it is not surjective for `1 (Z); for
injectivity, consider the 2-dimensional commutative algebra of matrices of the form
( a0 ab ) for a, b ∈ C; the Gelfand dual consists of only one point.)
Definition 5.4. A homomorphism ϕ of Banach ∗-algebras is a ∗-homomorphism if
it preserves the involution, that is, ϕ(a∗ ) = ϕ(a)∗ for all a.
Let A be a unital Banach ∗-algebra. We say that A is symmetric if, for all a ∈ A
such that a = a∗ , we have SpectrumA (a) ⊆ R. For example, if X is a compact
Hausdorff space, then C(X) is symmetric. On the other hand, the disk algebra D
is not symmetric (every element with all Taylor coefficients real has a = a∗ , but
the points of U give elements of the Gelfand dual, and no nonconstant holomorphic
function is real-valued at all these points).
Lemma 5.5. Let A be a commutative unital Banach ∗-algebra. The Gelfand trans-
form G : A → C(A)
b is a ∗-homomorphism if and only if A is symmetric.
Proof. If G is a ∗-homomorphism, then for a = a∗ ∈ A, the function â is real-valued.
Since the spectrum of a is the same as the spectrum (that is, the closure of the
range) of â (by theorem 5.1), it is a subset of R. Conversely, if A is symmetric, then
for any a ∈ A, â + ab∗ and i(â − ab∗ ) are real-valued functions: a bit of algebra shows
that ab∗ is the complex conjugate of â, so G is a ∗-homomorphism.
Lemma 5.6. Let A ⊆ B be symmetric unital Banach ∗-algebras. Let a ∈ A. Then
SpectrumA (a) = SpectrumB (a).
Proof. It is enough to show that A is invertible in A if and only if it is invertible
in B. “Only if” is obvious. To show “if”, suppose first that a = a∗ . Suppose that
a has an inverse b ∈ B. By the symmetric property of the algebra A, (a − (i/n)1)
is invertible in A for each n, and (a − (i/n)1)−1 → b as n → ∞ (by continuity of
inversion in B). But then b is a limit of members of A, so it belongs to A after all.
We have verified the claim when a = a∗ . To get the general case from this,
suppose that a is invertible in B, with inverse b. Then, in B, bb∗ a∗ a = b · 1 · a = 1,
and similarly a∗ abb∗ = 1, so a∗ a is invertible in B, hence in A by the special case
that we have already proved. That is, bb∗ ∈ A. Now it follows that b = bb∗ a∗ ∈ A
as required.
We now focus on algebras satisfying the C ∗ -identity 2.1.
Definition 5.7. A Banach ∗-algebra A is a C ∗ -algebra if ka∗ ak = kak2 for all a ∈ A.
A normed ∗-algebra that satisfies the C ∗ -identity but need not be complete is
sometimes called a pre-C ∗ -algebra. Such an algebra can be completed to a C ∗ -
algebra in the usual way.
Proposition 5.8. Let A be a unital C ∗ -algebra. If a ∈ A is normal (that is,
commutes with a∗ ) then kak = spr(a). For every a ∈ A, kak = spr(a∗ a)1/2 .
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
24
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
25
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
26
Lecture 6
Functional calculus and positivity
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
27
case we have a functional calculus for continuous functions on the spectrum that
vanish at zero. The Gelfand-Naimark theorem says that every nonunital commuta-
tive C ∗ -algebra is of the form C0 (X) for some locally compact Hausdorff space X.
The process of unitalization corresponds to forming the one-point compactification
of X.
Remark 6.3. Let A be a unital C ∗ -algebra and let a be a normal element of A. For
every connected component K of Spectrum(a) there is defined a spectral projection
p = χK (a) ∈ A, which is given by the functional calculus using the (continuous)
function χK , the characteristic function of K; we have p = p2 = p∗ . It is sometimes
important to know that p can be given by the Cauchy integral formula
Z
1
(6.4) p= (z1 − a)−1 dz,
2πi Γ
where Γ is a cycle in the resolvent set of a which has winding number 1 about K
and 0 about all other components of Spectrum(a). (The integrand is a continuous
function on a compact set with values in a Banach space; there is no difficulty
about defining such integrals as limits of Riemann sums in the usual way.) To prove
equation 6.4 it suffices, by the Gelfand-Naimark theorem, to consider the case where
A = C(X) for some compact X; then the result follows from the integral formula
for the winding number Z
1 dz
wn(Γ; w) =
2πi Γ z − w
together with an easy “Fubini-type” theorem that equates the result of evaluating a
C(X)-valued integral at x0 ∈ X with the ordinary integral of the function obtained
by evaluating the integrand at x0 .
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
28
Lecture 7
Positivity and order
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
29
Here we are using the notation Im(T ) for the closure of the range of the operator
T . Note that V and P need not commute, so that there are really two notions of
polar decomposition (left and right handed); we have made the conventional choice.
Proposition 7.4. Every operator T ∈ B(H) has a unique polar decomposition
1
T = V P . The operator P = |T | = (T ∗ T ) 2 belongs to C ∗ (T ), the C ∗ -subalgebra of
B(H) generated by T . If T happens to be invertible, then V is unitary, and it also
belongs to C ∗ (T ) in this case.
Proof. Let T = V P be a polar decomposition. Then T ∗ T = P ∗ V ∗ V P = P 2 , so P is
the unique positive square root of T ∗ T , which we denote by |T |. Fixing this P let
us now try to construct a polar decomposition. For ξ ∈ H,
kP ξk2 = hP ∗ P ξ, ξi = hT ∗ T ξ, ξi = kT ξk2
so an isometry Im(P ) → Im(T ) is defined by P v 7→ T v. Extending by zero on the
orthogonal complement we get a partial isometry V with the property that T = V P ;
and it is uniquely determined since the definition of polar decomposition requires
that V (P v) = T v and that V = 0 on the orthogonal complement of Im(P ).
1
If T is invertible then so is T ∗ T , and we have V = T (T ∗ T )− 2 ; this shows that
V ∈ C ∗ (T ) in this case, and unitarity is a simple check.
A key challenge in setting up the theory of abstract C ∗ -algebras is to prove the
equivalence of (a)–(c) above without using quadratic forms on Hilbert space as a
crutch. This was done by Kelley and Vaught some years after C ∗ -algebras were in-
vented! (This time lapse is responsible for a now-obsolete terminological difference in
the older literature: there was a distinction between ‘B ∗ -algebras’ and ‘C ∗ -algebras’,
and Kelley and Vaught’s theorem showed that this distinction was otiose.)
Proposition 7.5. Let A be a unital C ∗ -algebra. The following conditions on a
selfadjoint a ∈ A are equivalent:
(a) The spectrum of a is contained in the positive reals.
(b) a is the square of a selfadjoint element of A.
(c) kt1 − ak 6 t for all t > kak.
(d) kt1 − ak 6 t for some t > kak.
Proof. Conditions (a) and (b) are equivalent by a functional calculus argument.
Suppose that (a) holds. Then Spectrum(a) is contained in the interval [0, kak] of
the real axis and so the supremum of the function t1 − λ, for λ ∈ Spectrum(a), is
at most t; this proves (c) which implies (d). Finally, assume (d). Then |t − λ| 6 t
for all λ ∈ Spectrum(a); so λ is positive on Spectrum(a), yielding (a).
We will call a ∈ A positive if it is selfadjoint and satisfies one of these four
conditions. Part (d) of this characterization implies that the positive elements form
a cone:
Proposition 7.6. The sum of two positive elements of A is positive.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
30
Proof. Let a0 , a00 be positive elements of A. Let t0 , t00 be as in (d). Let a = a0 + a00
and t = t0 + t00 . Then by the triangle inequality t > kak and kt1 − ak 6 kt0 1 − a0 k +
kt00 1 − a00 k 6 t0 + t00 = t. So a is positive.
We can therefore define a (partial) order on Asa by setting a 6 b if b − a is posi-
tive. This order relation is very useful, especially in connection with von Neumann
algebras, but it has to be handled with care — not every ‘obvious’ property of the
order is actually true. For instance, it is not true in general that if 0 6 a 6 b, then
a2 6 b2 . In Proposition 7.8 we will give some true properties of the order relation.
These depend on the following tricky result.
Theorem 7.7. (Kelley–Vaught) An element a of a unital C ∗ -algebra A is positive
if and only if a = b∗ b for some b ∈ A.
Proof. Any positive element can be written in this form with b = a1/2 . To prove the
converse we will argue first that b∗ b cannot be a negative operator (except zero).
Indeed, put b = x + iy with x and y selfadjoint; then
b∗ b + bb∗ = 2(x2 + y 2 ) > 0.
Thus if b∗ b is negative, bb∗ = 2(x2 + y 2 ) − b∗ b is positive. But the spectra of b∗ b and
bb∗ are the same (apart possibly from zero), by a well-known algebraic result; we
deduce that Spectrum(b∗ b) = {0}, and hence b∗ b = 0.
Now for the general case: suppose that a = b∗ b; certainly a is selfadjoint, so we
may write a = a+ − a− where the positive elements a+ and a− are defined as in
1/2
Equation 7.1. We want to show that a− = 0. Let c = ba− . Then we have
1/2 1/2
c∗ c = a− aa− = −a2−
so c∗ c is negative. Hence it is zero, by the special case above; so a− = 0 and a is
positive.
Proposition 7.8. Let x and y be positive elements of a unital C ∗ -algebra A, with
0 6 x 6 y. Then
(a) We have x 6 kxk1 (and similarly for y); moreover kxk 6 kyk.
(b) For any a ∈ A, a∗ xa 6 a∗ ya.
(c) If x and y are invertible, we have 0 6 y −1 6 x−1 .
Proof. The first part of (a) is obvious from the functional calculus, and it implies
the second part by arguing that
x 6 y 6 kyk1
so x 6 kyk1 whence kxk 6 kyk. Item (b) becomes obvious if we use Theorem 7.7;
if x 6 y then y − x = c∗ c, so a∗ ya − a∗ xa = (ca)∗ ca is positive. For (c), if x 6 y
put z = x1/2 y −1/2 ; then z ∗ z = y −1/2 xy −1/2 6 1 (by (b)). But z ∗ z 6 1 if and only if
zz ∗ 6 1 (because the spectra of these are the same apart possibly from {0}), which
is to say that x1/2 y −1 x1/2 6 1; and applying (b) again we get y −1 6 x−1 .
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
31
which is valid for 0 < α 6 1. Finally, (iii) follows from (ii) since λα → 1 uniformly
on [ε, M ] for any fixed M as α → 0.
Remark 7.10. The function λ 7→ λα is not operator monotone for any α > 1. It
suffices to consider α = 2. Let p = ( 10 00 ) and q = 21 ( 11 11 ); these are orthogonal
projections. Clearly p 6 p + q. But p2
(p + q)2 since (p + q)2 − p = q + pq + qp =
1 3 2
( ) has negative determinant and so cannot be a positive operator.
2 2 1
Let A be a C ∗ -algebra (unital or not) and let Λ(A) be the set of positive a ∈ A
such that kak < 1. This is a partially ordered set.
Lemma 7.11. For any A, the partially ordered set Λ(A) is directed (Definition 1.5).
Proof. We must show that any two elements x, y ∈ Λ(A) have an upper bound. Let
a = (1 − x)−1 x and b = (1 − y)−1 y, which are positive elements of A (even though
the functional calculus takes place in the unitalization A).
e Let c = a + b and let
−1
z = (1 + c) c, so that z ∈ Λ(A). I claim z is an upper bound for x and y. Indeed
z = g1 (c) (in the notation of Lemma 7.9) and x = g1 (a); since a 6 c, operator
monotonicity gives x 6 z, and similarly y 6 z, as required.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
32
Lecture 8
Ideals and quotients
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
33
Proof. Let x ∈ J and let {uλ } be an approximate unit for the C ∗ -algebra J ∩ J ∗ .
Using the C ∗ -identity
kx∗ (1 − uλ )k2 = k(1 − uλ )xx∗ (1 − uλ )k 6 kxx∗ (1 − uλ )k
which tends to zero since xx∗ ∈ J ∩ J ∗ . It follows that x∗ uλ → x∗ , whence
x∗ ∈ J.
It is an important fact that the quotient of a C ∗ -algebra by an ideal (which is
certainly a Banach algebra in the induced norm) is in fact a C ∗ -algebra. We need
approximate units to check this.
Theorem 8.4. Let A be a C ∗ -algebra and J an ideal in A. Then A/J is a C ∗ -algebra
with the induced norm and involution.
Proof. A/J is a Banach algebra with involution (note that we need J = J ∗ here).
Recall that the quotient norm is defined as follows:
k[a]k = inf ka + xk.
x∈J
Here [a] denotes the coset (in A/J) of the element a ∈ A. I claim that
(8.5) k[a]k = lim ka(1 − u)k
u
where u runs over an approximate unit for J. Granted this it is plane sailing to the
C ∗ -identity, since we get
k[a∗ ][a]k = lim ka∗ a(1 − u)k > lim k(1 − u)a∗ a(1 − u)k
u u
= lim ka(1 − u)k2 = k[u]k2 .
u
To establish the claim notice that certainly ka(1 − u)k > k[a]k for every a. On the
other hand, given ε > 0 there exists x ∈ J with ka − xk 6 k[a]k + ε. It is eventually
true that kx(1 − u)k 6 ε also and then
ka(1 − u)k 6 k(a − x)(1 − u)k + kx(1 − u)k 6 k[a]k + 2ε
by the triangle inequality. This establishes the claim.
Remark 8.6. It is instructive to think about how this proof works where A = C(X)
and J is the ideal of functions that vanish on a closed subset F ⊆ X.
We can use this result to set up the fundamental ‘isomorphism theorems’ in the
category of C ∗ -algebras.
Corollary 8.7. Let α : A → B be any homomorphism of C ∗ -algebras. Then α is
contractive, and it can be factored as
A / A/ ker(α) / B
where the first map is a quotient map and the second is an isometric injection whose
range is a C ∗ -subalgebra of B.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
34
Proof. By passing to the unitalization, we may assume that A and B are unital and
that α is a unital ∗-homomorphism. By Corollary 5.13, α is contractive, hence in
particular it is continuous and ker(α) is a closed ideal. The standard construction
from algebra now factors α through an injective ∗-homomorphism β : A/ ker(α) →
B. Since A/ ker(α) is a C ∗ -algebra from the previous result, Corollary 5.13 shows
that β is an isometry of A/ ker(α) onto a subalgebra of B. This subalgebra is
complete (by the isometry property), hence closed. This completes the proof.
We can also prove the ‘diamond isomorphism theorem’.
Proposition 8.8. Let A be a C ∗ -algebra. Let J be an ideal in A and let B be a
C ∗ -subalgebra of A. Then B +J is a C ∗ -subalgebra of A and there is an isomorphism
B/(B ∩ J) = (B + J)/J.
Proof. The ∗-homomorphism B → A → A/J has range a C ∗ -subalgebra of A/J,
isomorphic to B/B ∩ J by the previous theorem. Thus the inverse image in A of
this C ∗ -subalgebra (namely B + J) is closed, and so is a C ∗ -algebra also. The result
now follows by the usual algebra argument.
Remark 8.9. When B is another ideal J there is a useful addendum to this result:
for two ideals I and J in A, their intersection is equal to their product. To see this,
observe
(I ∩ J)2 ⊆ IJ ⊆ I ∩ J.
But for any C ∗ -algebra B, we have B = B 2 , by an easy functional calculus argument.
Definition 8.10. Let A be a C ∗ -algebra. A C ∗ -subalgebra B of A is called heredi-
tary if for positive elements a ∈ A, b ∈ B, the inequality 0 6 a 6 b implies a ∈ B.
Example 8.11. Let p ∈ A be a projection. The subalgebra B = pAp (the “corner”
defined by p) is hereditary. Indeed, suppose 0 6 a 6 b = pbp, and write a = x∗ x.
Then
kx(1 − p)k2 = k(1 − p)a(1 − p)k 6 k(1 − p)b(1 − p)k = 0,
so x(1 − p) = 0 and thus pap = a.
Proposition 8.12. Every closed ideal of a C ∗ -algebra is a hereditary subalgebra.
Proof. Let J C A be such an ideal, and let 0 6 a 6 b with b ∈ J. Let {uλ } be an
approximate unit for J. Then lim(1 − uλ )b = 0 and hence
lim k(1 − uλ )a(1 − uλ )k 6 lim k(1 − uλ )b(1 − uλ )k = 0.
It follows that a1/2 = lim uλ a1/2 ∈ J, so a ∈ J.
A theorem of Effros says that every hereditary subalgebra is of the form L ∩ L∗ ,
where L is a closed left ideal, and that every such left ideal gives rise to a hereditary
subalgebra. This includes the above result.
Exercise 8.13. Let A be a C ∗ -algebra and let B = xAx where x is a positive
element of A. Prove that B is a hereditary subalgebra.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
35
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
36
Lecture 9
Examples of C ∗ -algebras
Finite-dimensional algebras
The obvious examples of finite-dimensional C ∗ -algebras are the algebras of bounded
operators on a finite-dimensional Hilbert space, in other words, the matrix algebras
Mn (C) of n × n matrices. The norm of a matrix A is of course the square root of
the largest eigenvalue of A∗ A, as follows from the relationship between norm and
spectral radius (5.8).
Lemma 9.1. Every finite-dimensional C ∗ -algebra is unital.
Proof. Let A be such a C ∗ -algebra. For any selfadjoint a ∈ A let a0 be the self-
adjoint projection f (a∗ a), where f (λ) equals 0 at λ = 0 and equals 1 for all λ > 0.
Finite-dimensionality implies that a has discrete spectrum, so f is continuous on
the spectrum of a and a0 is well defined. For any x ∈ A we have
kx(1 − (x∗ x)0 )k2 = k(1 − (x∗ x)0 )x∗ x(1 − (x∗ x)0 )k = 0,
so x = x(x∗ x)0 . Now suppose that p is a projection with p > (x∗ x)0 ; then
kx(1 − p)k2 = kx(1 − p)x∗ k 6 kx(1 − (x∗ x)0 )x∗ k = 0,
so xp = x. Choose then a vector space basis of A consisting of self-adjoint elements
xj and let p = (x21 + · · · + x2N )0 . By Lemma 7.9 (iii), p > (x2j )0 for each j so then
xj p = pxj = xj . Linearity now shows that p is a unit.
Definition 9.2. A C ∗ -algebra is called simple if it has no non-trivial 2-sided (closed)
ideals.
Lemma 9.3. Every simple finite-dimensional C ∗ -algebra is a matrix algebra.
We will prove this later, after we have developed some ideas about representations
of C ∗ -algebras. It follows that
Proposition 9.4. Every finite-dimensional C ∗ -algebra is isomorphic to a direct sum
of matrix algebras.
Proof. Let A be such an algebra and let B = Z(A) be the center of A, which is a
finite-dimensional commutative C ∗ -algebra. The Gelfand-Naimark theorem shows
that B ∼ = Ce1 ⊕ · · · ⊕ Cek where {e1 , . . . , ek } are mutually orthogonal central pro-
jections. Thus the Aj = ej A = Aej are subalgebras of A and A is their direct sum.
Moreover, I claim that each Aj is simple. For suppose not and let J C Aj be a non-
trivial ideal. As a finite-dimensional C ∗ -algebra, J is unital: its unit, considered as
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
37
/
2 /5
describes the homomorphism from C⊕M2 (C) to M2 (C)⊕M5 (C) that sends a⊕( db ec )
to
a 0 0 0 0
0 b c 0 0
a 0
⊕
0 d e 0 0 .
0 0 0 0 0
b c
0 0 0 d e
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
38
Example 9.6. One can see this measure theory/topology contrast even in the com-
mutative case. Let X be a compact metric space (the circle for example) and let µ
be a Radon measure on X. Every continuous complex-valued function on X acts
on H = L2 (X, µ) as a multiplication operator, and therefore we obtain an embed-
ding6 of the C ∗ -algebra A into B(H). The commutant A0 of this representation
of A is precisely the algebra L∞ (X, µ) of essentially bounded measurable functions
on X (proving this is a good exercise). This algebra is almost independent of the
topological structure of X, it only reflects the measure theory.
In the same situation, the essential commutant of A is the algebra
D(A) = {T ∈ B(H) : T a − aT ∈ K(H) ∀a ∈ A}.
In contrast to the von Neumann commutant, we will see that this algebra en-
codes significant topological information about A, even in the commutative example
that we mentioned above. For example, let X = S 1 as above, A = C(S 1 ), and
H = L2 (S 1 ). The essential commutant of A contains in particular the algebra of
pseudodifferential operators of order zero on the circle, closely related to analysis
and PDE.
Since D(A) ⊇ K by definition, we can define the reduced commutant to be the
∗
C -algebra quotient
QD(A) = D(A)/K.
An important example is the Calkin algebra
Q(H) = B(H)/K(H),
which is the reduced commutant of the identity. This arises in index theory: by
Atkinson’s theorem 3.8, an operator T ∈ B(H) is Fredholm precisely when it maps
to an invertible in Q(H).
Group algebras
Let G be a locally compact group. To keep things simple, we will assume for now
that G is a discrete group; but most of what we say extends easily to unimodular
locally compact groups (just by replacing sums over G with integrals with respect to
Haar measure); and indeed, with a bit of extra care about left v right, to all locally
compact groups whatsoever.
The group algebra C[G] is the vector space with basis G, given a multiplication
operation which linearly extends the group multiplication on basis elements. To put
it another way, C[G] consists of finitely supported functions f : G → C equipped
with the convolution product
X X
(9.7) (f1 ∗ f2 )(g) = f1 (g1 )f2 (g2 ) = f1 (h)f2 (h−1 g).
g1 g2 =g h∈G
6Provided µ is strong-starly positive on each open set, as Lebesgue measure is for example.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
39
This algebra also has an involution, which is the antilinear extension of the map
g 7→ g −1 , or in other words
f ∗ (g) = f¯(g −1 ).
This algebra has a faithful representation on a Hilbert space. Namely, let H = `2 (G)
and define the convolution of f1 ∈ C[G] and f2 ∈ H by the same formula as in
Equation 9.7 above. Then f1 ∗ f2 belongs to H also. The operator
Tf1 : f2 7→ f1 ∗ f2
is bounded on H, and the map f 7→ Tf is an injective ∗-homomorphism from C[G]
to B(`2 (G)), which we may use to regard C[G] as a ∗-subalgebra of B(`2 (G)). The
closure of C[G] inside B(`2 (G)) is therefore a C ∗ -algebra.
Definition 9.8. The C ∗ -algebra so defined is called the reduced group C ∗ -algebra
of G, denoted Cr∗ (G).
Example 9.9. Let G be the group Z. Then the group algebra C[G] can be identified
with the algebra C[z, z −1 ] of finite Laurent series, or trigonometric polynomials
(think of z = eiθ ). Similarly, Parseval’s theorem from Fourier analysis identifies
H = `2 (Z) with the space L2 (S 1 ) of square integrable functions on the circle, on
which the trigonometric polynomials operate by pointwise multiplication. The norm
of such an operator is the supremum norm of the multiplying function. Thus Cr∗ (Z)
is the closure, in the sup norm, of the trigonometric polynomials on S 1 . By the
Stone-Weierstrass theorem, this is the C ∗ -algebra C(S 1 ) of continuous functions
on the circle. As this example suggests, it is easy to see that if the group G is
commutative, so is the C ∗ -algebra Cr∗ (G).
Remark 9.10. The reason for the terminology “reduced” is the close association of
our construction with one particular representation of G, the regular representation
on `2 (G). In the classical representation theory of finite or compact groups, the
regular representation has a universal property (every irreducible representation is
a subrepresentation of it). But this is not true for all groups, even in a weak sense.
When it is not, the full scope of the representation-theory of G is not encompassed
by Cr∗ (G); one needs to use a larger construction, the so-called maximal C ∗ -algebra
of G. We will talk about this in much more detail later.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
40
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
41
example, we will construct the Toeplitz algebra, which is the universal algebra gen-
erated by a single isometry in roughly the same sense that C(S 1 ) is the universal
algebra generated by a single unitary. More generally we shall make the following
definition:
Definition 9.12. Let n = 2, 3, 4, . . .. The Cuntz algebra On is the universal Pn C ∗-
algebra generated by n isometries s1 , . . . , sn subject to the single relation i=1 si s∗i =
1.
For more about when one can and can’t make such definitions, see Remark 20.11.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
42
Lecture 10
Hilbert modules and multipliers
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
43
If hx, xi = 0, then taking λ large and negative in this inequality we see that hx, yi = 0
also. Otherwise, put λ = −khx, xik−1 to get Cauchy-Schwarz.
Remark 10.3. Note that the proof of Cauchy-Schwarz only requires (i)–(iii). It
follows that given an A-module M satisfying (i)–(iii), the set N of vectors of norm
zero forms a submodule, and dividing by that N we obtain a module M/N satisfying
(i)–(iv). Then by completing we may obtain a Hilbert module.
Example 10.4. An important example of a Hilbert A-module is the ‘universal’
module HA which is comprised of sequences {an } of elements of A such that a∗n an
P
converges in A. Using the Cauchy-Schwarz inequality it is not hard to show that
this is a Hilbert module. P Note the convergence condition carefully however: it is
not equivalent to say that kan k2 converges (a series of positive elements of a C ∗ -
algebra
P1 can converge in norm without converging absolutely, for instance the series
n
e n , where en is the orthogonal projection onto the n’th basis vector, converges
in norm but not absolutely in K(`2 )).
Lemma 10.5. Let M be a Hilbert A-module. Then M A is dense in M . In fact
M hM, M i is dense in M , where hM, M i denotes the subspace of A spanned by inner
products.
Notice that hM, M i need not be dense in A (ideals give obvious examples). We
say that M is full if this is so.
Proof. The closure of hM, M i is a C ∗ -ideal J in A. Let uα be an approximate unit
for J. Then for x ∈ M put xα = xuα . We have
hx − xα , x − xα i = hx, xi − uα hx, xi − hx, xiuα + uα hx, xiuα → 0
and thus xα → x, giving the result.
In the theory of Hilbert space a critical role is played by the existence of orthogonal
complements to closed subspaces. If M is a Hilbert space (or Hilbert module) and
N a closed subspace, then the orthogonal subspace is
N ⊥ = {y ∈ M : hx, yi = 0 ∀x ∈ N }.
The Projection Theorem states that if M is a Hilbert space then N ⊥ is complemen-
tary to N ; each x ∈ M can be decomposed uniquely into a component in N and a
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
44
component in N ⊥ . The Projection Theorem however fails for Hilbert modules. For
instance if A = C[0, 1], M = A, and N = C0 (0, 1] considered as a Hilbert A-module,
then N ⊥ = {0} and we have no direct sum decomposition of M into N ⊕ N ⊥ . This
is the fundamental difficulty of Hilbert module theory.
Definition 10.6. Let M and N be Hilbert A-modules. An adjointable map from
M to N is a linear map T : M → N for which there exists an adjoint T ∗ : N → M ,
necessarily unique, satisfying
hT x, yi = hx, T ∗ yi, x ∈ M, y ∈ N.
The set of adjointable maps will be denoted B(M, N ), or B(M ) if M = N .
Because of the failure of the Projection Theorem, it need not be the case that a
norm bounded linear map M → N is adjointable. On the other hand, an adjointable
map must be bounded (a Uniform Boundedness Principle argument).
Proposition 10.7. Let M be a Hilbert A-module. Then B(M ) is a C ∗ -algebra.
Proof. B(M ) is clearly a normed ∗-algebra. The C ∗ -identity can be proved by
mimicking the Hilbert space proof: for x ∈ M , kxk = 1,
kT xk2 = khT x, T xik = khx, T ∗ T xik 6 kT ∗ T k.
The supremum of the left side over all possible x is kT k2 , giving the result. It
follows as usual that kT k = kT ∗ k for all T ∈ B(M ), and so we see that B(M ) is a
closed subspace of the Banach algebra of all bounded C-linear maps M → M , and
therefore is complete.
Let E and F be Hilbert A-modules. A rank one operator from E to F is a linear
map E → F of the form
θx,y (z) = xhy, zi, x ∈ F, y ∈ E.
It is adjointable, with adjoint θy,x . The closed linear span of rank one operators is
denoted K(E, F ) and called the space of compact operators from E to F . (Warning:
They need not be compact in the sense of Banach space theory!) If E = F , the
subspace K(E) of compact operators in B(E) is a C ∗ -ideal, in fact an essential ideal
(same proof as Remark 8.15).
Lemma 10.8. Let M be a Hilbert A-module, and let uα be an approximate unit for
the C ∗ -algebra K(M ). Then uα (ξ) → ξ for all ξ ∈ M .
Proof. By lemma 10.5 it suffices to show this assuming that ξ ∈ M hM, M i. But if
ξ = xhy, zi then
uα ξ = (uα θx,y )(z) → θx,y (z) = ξ
as required.
Example 10.9. If A is a unital C ∗ -algebra, and we consider A as a Hilbert module
over itself, the it is easy to see that B(A) = K(A) = A.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
45
Example 10.10. Now suppose that A is a non-unital C ∗ -algebra. It is still the case
that A = K(A); the isomorphism sends a ∈ A to the operator of left multiplication
La (a0 ) = aa0 on the Hilbert module A. It is clear that La is right A-linear; to see
that it is compact, choose an approximate unit {uλ } for A and write La = lim θa,uλ .
In the non-unital case the algebra B(A) is a unital C ∗ -algebra containing A =
K(A) as an ideal. It is called the multiplier algebra of A, and denoted M (A); by the
preceding remarks, it contains A as an essential ideal.
Proposition 10.11. The multiplier algebra M (A) of a non-unital C ∗ -algebra A has
a universal property: if A is also embedded as an ideal in a C ∗ -algebra B, then there
is a unique ∗-homomorphism B → M (A) that restricts to the identity on A; and
this extension is injective if A is essential in B.
Proof. Left multiplication Lb by b ∈ B defines an adjointable Hilbert A-module map
A → A (the adjoint is left multiplication by b∗ ), that is, an element of M (A). If A
is essential in B then Lb = 0 implies b = 0, which is to say that the induced map
B → M (A) is injective.
Example 10.12. Let A = C0 (X) where X is a locally compact Hausdorff space.
Then M (A) = Cb (X), the algebra of all bounded continuous functions on X. In-
deed, it is clear that A is an essential ideal in Cb (X); on the other hand, if T is a
Hilbert module map A → A, then using linearity one sees that for each x ∈ X, the
quantity f (x) = (T ux )(x) is independent of the choice of ux ∈ A having ux (x) = 1;
the function f is continuous and bounded, and T is given by multiplication by f .
Thus the construction of the multiplier algebra corresponds to the Stone-Cech com-
pactification in the commutative case. (Contrast the process of unitalization, which
corresponds to the one-point compactification.)
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
46
Lecture 11
Review of locally convex spaces
The key objects of study in functional analysis are Banach spaces (of which C ∗ -
algebras are a particular case, of course). However, more general topological vector
space structures arise naturally in the study of C ∗ -algebras. In this lecture we’ll do
a quick review.
Definition 11.1. A topological vector space is a vector space E over the field F = R
or C equipped with a topology in which the vector space operations (addition and
scalar multiplication) are continuous as maps E × E → E, F × E → E.
Topological vector spaces form a category in which the morphisms are the con-
tinuous linear maps. Many authors (e.g. Rudin) restrict attention to Hausdorff
topological vector spaces, but it is occasionally useful to consider non-Hausdorff
examples.
Exercise 11.2. A TVS is Hausdorff if and only if the origin {0} is a closed subset.
There is a natural notion of Cauchy net in a topological vector space (a net {xλ }
is Cauchy if xλ − xµ → 0 as λ, µ → ∞) and therefore we may speak of a complete
TVS (one in which every Cauchy net converges).
Definition 11.3. A seminorm on a vector space E is a function p : X → R+ that
satisfies p(λx) = |λ|p(x) for all λ ∈ F, and p(x + y) 6 p(x) + p(y) (the triangle
inequality). It is a norm if in addition p(x) = 0 implies x = 0.
Many examples of topological vector spaces arise from the following construction.
Let {pα }α∈A be a family of seminorms on a vector space E. We can define a topology
on E be calling U ⊆ E open if for every x0 ∈ U there exist α1 , . . . , αn ∈ A and real
ε1 , . . . , εn > 0 such that
{x ∈ E : pαi (x − x0 ) < εi , (i = 1, . . . , n)} ⊆ U.
This topology makes E into a TVS; it is called the locally convex topology defined
by the seminorms {pα }.
Exercise 11.4. Let xj be a net in a LCTVS whose topology is defined by a family
of seminorms {pα }. Then xj → x if and only if, for each α, pα (xj − x) → 0. What
condition on the family of seminorms will ensure that E is Hausdorff?
There is a geometric reason for the “locally convex” terminology (namely, a TVS
is locally convex iff there is a basis for the neighborhoods of 0 comprised of convex
sets). To see this, suppose that p is a seminorm on the vector space E. Then the
set A = {x ∈ E : p(x) 6 1} has the following properties:
• A is convex : for any finite P subset {a1 , . . . , an } of A, and any
P positive real
numbers λ1 , . . . , λn with λi = 1, the convex combination λi ai belongs
to A.
• A is balanced : if a ∈ A and λ ∈ F with |λ| 6 1, then λa ∈ A also.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
47
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
48
ψ
ϕ
R
must commute.
Proof. Let C be the collection of pairs (G, χ), where F 6 G 6 E, and χ : G → R
is a linear functional extending ϕ and having χ 6 p. Order this set by saying that
(G1 , χ1 ) 6 (G2 , χ2 ) if G1 6 G2 and χ2 extends χ1 .
In this partially ordered set, each chain has an upper bound (the union). So by
Zorn’s Lemma, there is a maximal element (G, χ) in C. We claim that this maximal
G is the whole of E.
Suppose not. Choose x1 ∈ E \ G and try to extend χ to a functional χ1 on
G1 = hG, x1 i. Such an extension is determined by α = χ1 (x1 ); in fact, it is given by
the formula
χ1 (x + λx1 ) = χ(x) + λα (x ∈ G);
we need to pick α so that the condition χ1 6 p is satisfied.
Taking λ = ±1 it is evident that we must have, for all x ∈ G,
χ(x) + α 6 p(x + x1 ), χ(x) − α 6 p(x − x1 )
and conversely a scaling argument shows that this would be sufficient. Can we
choose α to make this condition true? Equivalently, can we choose α so that
χ(y) − p(y − x1 ) 6 α 6 p(x + x1 ) − χ(x) ∀x, y ∈ G?
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
49
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
50
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
51
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
52
Proof. Notice that both G and K are compact metrizable spaces. Let P (G) be the
space of probability measures on G, equipped with its vague topology (that is, the
weak-star topology as a subset of the dual of C(G)). It is a compact convex set.
Define a map Φ : P (G) → E ∗ by
Z
Φ(µ)(x) = σ(x) dµ(σ).
G
If µ is a convex combination of Dirac measures, Φ(µ) is a convex combination of
points of G. Since every µ is a vague limit of such convex combinations, one sees
easily that the range of Φ is exactly K.
Suppose that ψ is an extreme point of K. Let
S = {µ ∈ P (G) : Φ(µ) = ψ}.
It is nonempty (since the range of Φ is K), and since ψ is an extreme point of K, it
follows that S is an extreme set in P (G) (see the proof of 11.18 for the definition of
extreme set). Every extreme set contains an extreme point, as we saw above, and
the extreme points of P (G) are Dirac measures, so there is some Dirac measure in
S, say the unit mass δϕ at ϕ ∈ G. But then
Z
ψ(x) = Φ(δϕ )(x) = σ(x) dδϕ (σ) = ϕ(x),
G
so ψ = ϕ and belongs to G as required.
Proposition 11.20. (Goldstine’s theorem) Let E be a Banach space, and regard
E as a subspace of its double dual E ∗∗ via the canonical embedding. The unit ball
of E is weak-∗ dense in the unit ball of E ∗∗ .
Proof. Let B be the unit ball of E; we want to know what is the weak-∗ closure of
B in E ∗∗ . Since B is convex, this is the same as finding its weak-∗ closed convex
hull, i.e., the intersection of all the weak-∗ closed half-spaces that contain it. But
the weak-∗ continuous linear functionals on a dual space F ∗ are just the points of
F . In other words, the weak-∗ closure of B is the intersection of all the half-spaces
{x ∈ E ∗∗ : ϕ(x) 6 1},
as ϕ runs over the unit ball of E ∗ , and this is just the unit ball of E ∗∗ .
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
53
Lecture 12
Topologies on B(H)
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
54
the ultraweak topology arises from the fact that the trace-class operators can be
made into a Banach space, and the above pairing makes B(H) the dual of this
Banach space: the ultraweak topology is just the weak-star topology associated to
this duality pairing. But I will try not to mention it any more. (Reading over these
notes makes it clear that this resolution was not successful.)
There is an intimate relation between positivity and the strong (and weak) oper-
ator topologies:
Lemma 12.4. If a norm-bounded net of positive operators converges weakly to zero,
then it converges strongly to zero.
Proof. Let {Tλ } be such a net in B(H). Let ξ ∈ H. By the Cauchy-Schwarz
inequality
1/2 3/2
kTλ ξk4 = hTλ ξ, Tλ ξi2 6 hTλ ξ, ξihTλ3 ξ, ξi 6 kTλ k3 kξk2 hTλ ξ, ξi
and the right-hand side tends to zero by weak convergence.
The “monotone convergence theorem” holds in the strong topology.
Proposition 12.5. Let {Tλ } be an increasing net of selfadjoint operators in B(H),
and suppose that {Tλ } is bounded above by a selfadjoint operator S. Then {Tλ } is
strongly convergent to an operator T with T 6 S.
Proof. We need to review some standard material on operators, sesquilinear forms,
and quadratic forms. Let T be a bounded operator on a Hilbert space. Then the map
σT : H × H → C defined by σT (ξ, η) = hT ξ, ηi is called the sesquilinear form associ-
ated to T . Clearly it is bounded in the sense that |σT (ξ, η)| 6 Ckξkkηk; conversely,
it follows from the Riesz representation theorem that any bounded sesquilinear form
arises from a unique bounded operator T .
The quadratic form associated to a sesquilinear form σ is the map ϕ defined by
ϕ(ξ) = σ(ξ, ξ). One can recover σ from ϕ via the polarization identity
3
X
4σ(ξ, η) = ik ϕ(ik ξ + η)
k=0
and, abstractly, one can define a quadratic form as a mapping ϕ such that the
polarization identity yields a sesquilinear form. A quadratic form is hermitian if
it has values in R and positive if it has values in R+ ; in the bounded case these
correspond to the hermitian and positive properties of the corresponding operator
T.
After this preparation, return to the proof. For each fixed ξ ∈ H the limit of
the increasing net hTλ ξ, ξi of real numbers exists and is bounded by hSξ, ξi. The
function ξ 7→ limλ hTλ ξ, ξi is a bounded quadratic form, and using the polarization
identity one may define a positive bounded operator T such that
hT ξ, ξi = limhTλ ξ, ξi.
λ
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
55
Proof. Let uα be an approximate unit for K(M ). By Lemma 10.8, uα ξ → ξ for all
ξ ∈ M . Thus for each T ∈ B(M ),
k(T uα − T )ξk → 0, k(uα T ∗ − T ∗ )ξk → 0,
and therefore the net T uα , which consists of compact operators of norm 6 kT k,
tends strong-starly to T . The result follows.
Since the Riesz representation theorem is not true for Hilbert modules in general,
there is no correspondence between sesquilinear forms and operators as there is in
the Hilbert space case. Thus Proposition 12.5 fails for B(M ) where M is a Hilbert
module (i.e., as we shall see, B(M ) is not a von Neumann algebra in general).
However, there is a partial result:
Lemma 12.8. Let M be a Hilbert A-module and T an adjointable operator on M .
Then T is positive (in the C ∗ -algebra B(M )) if and only if hx, T xi > 0 (in A) for
all x ∈ M .
Proof. Suppose that T is adjointable and satisfies hx, T xi > 0 for all x. Using the
obvious analog of the polarization identity it follows that T is self-adjoint. write
T = T+ − T− in the usual way. Then for all x,
0 6 hx, T−3 xi = −hT− x, T T− xi 6 0,
whence T− = 0 and T is positive.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
56
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
57
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
58
Lecture 13
Von Neumann Algebras
Proof. We have already proved (i). As for (ii), let C be strong-starly closed and
suppose Tα is a net in C converging weakly to T . Write Tα = Xα + iYα where
Xα = 21 (Tα + Tα∗ ) and Yα = 2i (Tα∗ − Tα ) are selfadjoint nets in C and converge weakly
to X and Y with T = X + iY . By part (i) we may replace Xα , Yα by convex
combinations which will converge strongly to X and Y , hence strong-starly since
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
59
T 7→ ..
.
T
(in a more sophisticated language this is T → 7 T ⊗ In ). A routine calculation
shows that ρ(A)0 consists of n × n matrices all of whose elements belong to A0 , and
therefore that ρ(A)00 = ρ(A00 ). Now apply the previous case to the C ∗ -algebra ρ(A),
the operator ρ(T ), and the vector ξ. We find that there is ρ(R) ∈ ρ(A) such that
k(ρ(R) − ρ(T ))ξk < 1, and when written out componentwise this is exactly what we
want.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
60
Remark 13.5. Our hypothesis that A must be unital is stronger than necessary.
Suppose simply that A is non-degenerate, meaning that AH is dense in H, or equiv-
alently that Rη = 0 for all R ∈ A implies η = 0. Then the conclusion of the double
commutant theorem holds. To see this, we need only check that the vector ξ still
belongs to Aξ — the rest of the proof will go through as before. Indeed, one can
write ξ = P ξ + (1 − P )ξ, where P is the projection onto Aξ as above, and since
P ∈ A0 we have, for every R ∈ A,
R(1 − P )ξ = (1 − P )Rξ = Rξ − Rξ = 0;
so by non-degeneracy (1 − P )ξ = 0 and ξ = P ξ ∈ Aξ.
Let H be a Hilbert space and T ∈ B(H). Recall (Definition 7.3) that a polar
decomposition for T is a factorization T = V P , where P is a positive operator and
V is a partial isometry with initial space Im(P ) and final space Im(T ). We proved in
Proposition 7.4 that a polar decomposition exists and is unique, and that moreover
P = |T | belongs to the C ∗ -algebra generated by T . The partial isometry V does
not always belong to this C ∗ -algebra (example?), but we do have:
Lemma 13.6. Let T ∈ B(H) have polar decomposition T = V P . Then the partial
isometry V belongs to W ∗ (T ), the von Neumann algebra generated by T .
Proof. Using the bicommutant theorem, it is enough to show that V commutes with
every operator S that commutes with T and T ∗ . We have H = Ker(P ) ⊕ Im(P )
so it suffices to check that V Sξ = SV ξ separately for ξ ∈ Ker(P ) = Ker(T ) and
for ξ = P η. In the first case, ST ξ = T Sξ = 0 so Sξ ∈ Ker(T ), and then V ξ = 0,
V Sξ = 0. In the second case
V Sξ = V SP η = V P Sη = T Sη = ST η = SV P η = SV ξ
since P belongs to the C ∗ -algebra generated by T .
Remark 13.7. It follows of course that the orthogonal projection V V ∗ onto Im(T )
also belongs to the von Neumann algebra generated by T (this is also easy to see
directly). A corollary is that every von Neumann algebra of dimension > 1 contains
nontrivial projections. For suppose T is a self-adjoint element of such a von Neu-
mann algebra B which is not a multiple of the identity. Then the spectrum of T
contains at least two points, λ0 and λ1 . Let f be a continuous function equal to zero
in a neighborhood of λ0 and equal to one in a neighborhood of λ1 . Then f (T ) ∈ B
and the range of f (T ) is neither 0 nor all of H (the latter because there is g(T )
with nonzero range such that f (T )g(T ) = 0). The projection onto Im(f (T )) then
is nontrivial and belongs to B.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
61
Lecture 14
Von Neumann Algebras, representations, and states
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
62
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
63
Proof. Let E = ρ(A)H and let y ∈ E. Let {uλ } be an approximate unit for A. We’ll
construct a ∈ A, x ∈ E such that y = ρ(a)x by an inductive process. It’s convenient
to carry out the intermediate steps of the induction inside the unitalization A e (notice
that ρ extends to a unital action of A e on E), but the final result will lie in A.
To start the induction define x0 = y ∈ E and a0 = 1 ∈ A. e Supposing that
x0 , . . . , xn−1 and a0 , . . . , an−1 have been defined, select un from the approximate
unit so that kρ(un )xn−1 − xn−1 k < 2−n (this can be done for any element of E). Let
Xn
−n −n
an = an−1 − 2 (1 − un ) = 2 1 + 2−n un
k=1
and note that the latter expression is a sum of positive terms so an is invertible in
e with ka−1 k 6 2n . Set xn = ρ(a−1 )y. Now
A n n
xn − xn−1 = ρ(a−1 −1
n (an−1 − an )an−1 )y = 2
−n
ρ((1 − un ))xn−1 ,
so kxn − xn−1 k 6 2−n . Thus {xn }, {an } are Cauchy sequences convergent
P to x−kand a
respectively, and y = ρ(a)x since y = ρ(an )xn for each n. Finally, a = ∞
k=1 2 uk ∈
A as required.
Remark 14.8. The construction used in the lemma above is called the Cohen fac-
torization theorem. Inspection of the proof shows that no Hilbert space properties
of H are used; it follows in the same way that for any Banach space X which is a
module over a C ∗ -algebra A, the subspace AX is closed in X. For instance, suppose
that A is a subalgebra of a C ∗ -algebra B and J C B. Then AJ is closed in J.
Representations are studied by way of states.
Definition 14.9. A linear functional ϕ : A → C on a C ∗ -algebra A is called positive
if ϕ(a∗ a) > 0 for all a ∈ A. A state is a positive linear functional of norm one.
Example 14.10. Let ρ : A → B(H) be a representation and let ξ ∈ H be a unit
vector. Then the functional
ϕ(a) = hξ, ρ(a)ξi
is a state, called the vector state associated to ξ.
Any positive linear functional is bounded; for, if ϕ were positive and unbounded,
P −k
we could find elements ak ∈ A+ of norm one with ϕ(ak ) > 4k , and then a = 2 ak
−k k
is an element of A+ satisfying ϕ(a) > 2 ϕ(ak ) > 2 for all k, contradiction. Thus
there is little loss of generality in restricting attention to states. A state on C(X)
is a Radon probability measure on X, by the Riesz representation theorem from
measure theory. Note that the state defined by the Radon probability measure µ is
a vector state of C(X); take H = L2 (X, µ), ρ the representation by multiplication
operators, and ξ the constant function 1.
If σ is a state (or just a positive linear functional) on A then we may define a
(semidefinite) ‘inner product’ on A by
ha, biσ = σ(a∗ b).
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
64
The Cauchy–Schwarz inequality continues to hold (with the usual proof): we have
|ha, biσ |2 6 ha, aiσ hb, biσ .
Note that ha, aiσ is different from the C ∗ -norm kak2 .
Proposition 14.11. A bounded linear functional ϕ : A → C is positive if and only
if lim ϕ(uλ ) = kϕk for some approximate unit uλ for A; and in particular, if and
only if ϕ(1) = kϕk in case A happens to be unital.
Proof. There is no loss of generality in supposing that kϕk = 1.
Suppose that ϕ is positive. Then ϕ(uλ ) is an increasing net of real numbers,
bounded above by 1, and therefore convergent to some real r 6 1. For each a in the
unit ball of A, we have by Cauchy–Schwarz
|ϕ(auλ )|2 6 ϕ(a∗ a)ϕ(u2λ ) 6 ϕ(uλ ) 6 r.
Thus, since {uλ } is an approximate unit, |ϕ(a)|2 6 r, whence r > 1 on taking the
supremum over a in the unit ball of A. Since we already know r 6 1, this gives
r = 1 as required.
Conversely suppose that lim ϕ(uλ ) = 1. First we will prove that ϕ takes real
values on selfadjoint elements. Take a selfadjoint element a ∈ A, of norm one, and
write ϕ(a) = x + iy ∈ C; we want to prove that y = 0; assume for a contradiction
that y > 0. For each n one can extract a un from the approximate unit which has
ϕ(un ) > 1 − 1/n3 and commutes well enough with a that we have the estimate
knun + iak2 6 n2 + 2
(of course if it were the case that un a = aun exactly then the bound would be n2 +1).
Therefore 2
2 3 2
x + n(1 − 1/n ) + y 6 |ϕ(nun + ia)|2 6 n2 + 2
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
65
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
66
Lecture 15
The GNS Construction
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
67
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
68
for all y ∈ A.
By following through the GNS construction exactly as above, one can show that to
any representable linear functional ϕ there corresponds a Hilbert space representation
ρ : A → B(H) and a cyclic vector u such that ϕ(a) = hρ(a)u, ui.
Representability is automatic for positive linear functionals on unital Banach
algebras:
Lemma 15.7. Any positive linear functional ϕ on a unital Banach ∗-algebra7 A is
continuous and representable.
Proof. Suppose x = x∗ ∈ A with kxk < 1. Using Newton’s binomial series we may
construct a selfadjoint
1 1 3
y = (1 − x) 2 = 1 − x + x2 − · · ·
2 8
2
with y = 1 − x. We conclude that ϕ(1 − x) > 0 by positivity, so ϕ(x) 6 ϕ(1). Now
for any a ∈ A we have
|ϕ(a)|2 6 ϕ(1)ϕ(a∗ a) 6 ϕ(1)2 ka∗ ak 6 ϕ(1)2 kak2
(the first inequality following from Cauchy-Schwarz and the second from the pre-
ceding discussion), so ϕ is continuous and kϕk 6 ϕ(1). To show representability, we
apply the preceding discussion to the positive linear functional a 7→ ϕ(y ∗ ay) to get
the desired conclusion with cx = kx∗ xk.
It was shown by Varopoulos that this theorem extends to non-unital Banach
algebras having bounded approximate units.
7By definition, we require that the involution on a Banach ∗-algebra should be isometric.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
69
Lecture 16
Abelian von Neumann algebras and the spectral theorem
Now we will discuss the GNS construction in the abelian case, and what can be
obtained from it. Let us begin by constructing a standard example of an abelian
von Neumann algebra.
Let (X, µ) be a finite measure space. For each essentially bounded function f ∈
∞
L (X, µ), the corresponding multiplication operator Mf on the Hilbert space H =
L2 (X, µ) is bounded, with operator norm equal to the L∞ -norm of f ; and in this
way one identifies L∞ (X, µ) with a ∗-subalgebra of B(H).
Proposition 16.1. The algebra L∞ (X, µ) is a von Neumann algebra of operators
on L2 (X, µ).
Proof. We will show that M = L∞ (X, µ) is equal to its own commutant. Suppose
that T ∈ B(H) commutes with each Mf , and let g = T · 1, where 1 denotes the
constant function 1, considered as an element of H. Then for all f ∈ L∞ (X, µ) ⊆
L2 (X, µ) we have
T f = T Mf 1 = Mf T 1 = Mf g = f g.
∞ 2
Since L is dense in L it follows that T = Mg , and standard estimates show that
kgk∞ 6 kT k, so that g ∈ L∞ . Thus M 0 ⊆ M ; but M ⊆ M 0 since M is abelian, and
consequently M = M 0 .
The proposition still holds for a σ-finite measure space (X, µ), with essentially
the same proof.
The content of Spectral Theorem is that this is essentially the only example of
an abelian von Neumann algebra. However, to explain what is meant here we need
to distinguish between two notions of isomorphism for von Neumann algebras (or
other algebras of operators on Hilbert space). Let M ⊆ B(H) and N ⊆ B(K) be
von Neumann algebras. We will say that they are abstractly isomorphic if there is a
∗-isomorphism (an isomorphism of C ∗ -algebras) from M to N ; and we will say that
they are spatially isomorphic if there is a unitary U : H → K such that U M U ∗ = N .
Clearly spatial isomorphism implies abstract isomorphism but the converse is not
true: the 1-dimensional von Neumann algebras generated by the identity operators
on two Hilbert spaces of different finite dimension are abstractly isomorphic, but
they are not spatially isomorphic.
Proposition 16.2. Let X be a compact metrizable space. Every cyclic representa-
tion of C(X) is unitarily equivalent to a representation of C(X) by multiplication
operators on L2 (X, µ), where µ is some regular Borel probability measure on X.
Proof. Let ρ be a cyclic representation of A = C(X) with cyclic vector ξ. Then
σ(f ) = hξ, ρ(f )ξi defines a state on A. By the uniqueness of the GNS construction
(Remark 15.2), ρ is unitarily equivalent to the GNS representation constructed from
the state σ. Now we appeal to the Riesz Representation Theorem from measure
theory (see Rudin, Real and complex analysis, Theorem 2.14, for example): each
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
70
state σ of C(X) is the functional of integration with respect to some regular Borel
probability measure µ. The GNS construction then produces the space L2 (X, µ)
and the multiplication representation.
Let M be an abelian von Neumann algebra in B(H). We say that M is maximal
abelian if it is contained in no larger abelian von Neumann algebra. It is easy
to see that this condition is equivalent to M 0 = M . In particular, the proof of
Proposition 16.1 shows that L∞ (X, µ) is a maximal abelian von Neumann algebra
of operators on L2 (X, µ).
Theorem 16.3. Let M be an abelian von Neumann algebra in B(H), where H is
separable. Then the following are equivalent:
(a) M has a cyclic vector;
(b) M is maximal abelian;
(c) M is spatially equivalent to L∞ (X, µ) acting on L2 (X, µ), for some finite mea-
sure space X.
Proof. Suppose that M has a cyclic vector. We begin by remarking that M has
a separable C ∗ -subalgebra A such that A00 = M . (Proof: The unit ball of B(H)
is separable and metrizable in the strong topology, hence second countable. Thus
the unit ball of M is second countable, hence M itself is separable, in the strong
topology. Let A be the C ∗ -algebra generated by a countable strongly dense subset
of M .) The representation of A on B(H) is cyclic (since A is strongly dense in
M ) so by Proposition 16.2 it is unitarily equivalent to the representation of C(X)
on L2 (X, µ) by multiplication operators. Now L∞ (X, µ) is a von Neumann algebra
which contains C(X) as a strongly dense subset (by the Dominated Convergence
Theorem) so the unitary equivalence must identify L∞ (X, µ) with M . This proves
that (a) implies (c). We have already remarked that (c) implies (b). Finally, suppose
that M is any abelian von Neumann algebra of operators on H. Decompose H as
a countable direct sum of cyclicP subspaces for the commutant M 0 , say with unit
cyclic vectors ξn ; and let ξ = 2−n ξn . The projection Pn : H → Hn belongs to
00 0
M = M ⊆ M , using the bicommutant theorem and the fact that M is abelian.
Hence each ξn = 2n Pn ξ belongs to M 0 ξ, so ξ is a cyclic vector for M 0 . We have shown
that the commutant of any abelian von Neumann algebra on a separable Hilbert
space has a cyclic vector; in particular if M is maximal abelian, then M = M 0 and
M has a cyclic vector.
We can use this to get the structure of a general abelian von Neumann algebra,
up to abstract (not spatial!) isomorphism:
Theorem 16.4. Each abelian von Neumann algebra on a separable Hilbert space is
abstractly isomorphic to some L∞ (X, µ).
Proof. Let M be an abelian von Neumann algebra on B(H). Recall from the last
part of the preceding proof that the commutant M 0 has a cyclic vector ξ; let P
denote the orthogonal projection onto M ξ. Then P is a projection in M 0 . The
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
71
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
72
For the proof, one need only verify that the definition E(S) = χS (T ) (using the Borel
functional calculus described above) describes a resolution of the identity. To show
that T has an integral decomposition as above one approximates the integral by
Riemann sums and uses the fact that, when restricted to the range of the projection
E((k/n, (k + 1)/n]), the operator T lies within 1/n in norm of the operator of
multiplication by k/n.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
73
Remark 16.9. The point of the spectral theorem is that it provides a set of canonical
models (up to unitary equivalence) for normal operators: every normal operator is
equivalent to one of the models. For a complete picture, this existence statement
needs to be supplemented by a uniqueness statement saying when two of the models
are or are not unitarily equivalent to one another. This is not a completely straight-
forward matter. Even in the finite dimensional case, two normal operators with the
same spectrum need not be unitarily equivalent (because the multiplicities of the
eigenvalues may differ). In the infinite-dimensional case this “multiplicity” idea has
to be elaborated with a heavy dose of measure theory. (Example: Let X denote the
multiplication operator by x on L2 [−1, 1]. Show that X and X ⊕X are not unitarily
equivalent. Note that both operators have spectrum [−1, 1] and no eigenvalues at
all.) The final result is called spectral multiplicity theory (see Halmos’ book with
that title). We won’t discuss it further here.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
74
Lecture 17
Pure States and Irreducible Representations
In this section we will begin a more detailed study of the representation theory of
a C ∗ -algebra A. By application to group C ∗ -algebras, this theory will include the
theory of unitary representations of groups. We’ll discuss this connection.
Let A be a C ∗ -algebra (unital for simplicity). Recall that a linear functional
σ : A → C is a state if it has norm one and moreover σ(1) = 1. Thus the collection
of states on A is a weak-star compact8, convex subset of A∗ . It is called the state
space of A, and denoted by S(A).
The general yoga of functional analysis encourages us to consider the extreme
points of S(A), called the pure states. By the Krein–Milman Theorem, S(A) is the
closed convex hull of the pure states.
Lemma 17.1. Let σ be a state of a C ∗ -algebra A, and let ρ : A → B(H) be the
associated GNS representation, with unit cyclic vector ξ. For any positive functional
ϕ 6 σ on A there is an operator T ∈ ρ[A]0 such that
ϕ(a) = hξ, ρ(a)T ξi
and 0 6 T 6 1.
This is a kind of ‘Radon-Nikodym Theorem’ for states.
Proof. Recall that H is obtained by completing the inner product space A/N , where
N is the set {a ∈ A : σ(a∗ a) = 0}. Define a sesquilinear form on A/N by (a, b) 7→
ϕ(a∗ b). This form is well-defined, positive semidefinite, and bounded by 1, so there
is an operator T on H such that
ϕ(a∗ b) = h[a], T [b]i = hρ(a)ξ, T ρ(b)ξi.
The computation
h[a], T ρ(b)[c]i = ϕ(a∗ bc) = h[b∗ a], T [c]i = h[a], ρ(b)T [c]i
now shows that T commutes with ρ[A].
Proposition 17.2. A state is pure if and only if the associated GNS representation
is irreducible.
Proof. Suppose first that the representation ρ : A → B(H) is reducible, say H =
H1 ⊕ H2 where H1 and H2 are closed, A-invariant subspaces. Write a cyclic vector
ξ = ξ1 + ξ2 where ξi ∈ Hi , i = 1, 2. We cannot have ξ1 = 0, otherwise ρ[A]ξ ⊆ H2 ;
and for similar reasons we cannot have ξ2 = 0. Now we have
σ(a) = hξ, ρ(a)ξi = hξ1 , ρ(a)ξ1 i + hξ2 , ρ(a)ξ2 i
8Warning: In the non-unital case the state space is not closed in the unit ball of A∗ , and hence
is not compact, because the condition lim σ(uλ ) = 1 is not preserved by weak-star convergence—
the zero functional can arise as a weak-star limit of states (consider the evaluations at the points
1, 2, 3, . . . as functionals on C0 (R)). One must replace the state space by the quasi-state space
Q(A) consisting of positive functionals of norm 6 1. See Pedersen’s book for details.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
75
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
76
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
77
ideal. Consequently, unless all prime ideals are maximal (a rare occurrence) the
topology is not even T1 , let alone Hausdorff. However, it is T0 — given any two
distinct points, at least one of them is not in the closure of the other one. (Proof is
an easy exercise, but a good check that you are following the definitions.)
Exercise 17.12. If A is unital show that Prim(A) is compact.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
78
Lecture 18
Around Fell’s theorem
(these are the “kernels” of U and Prim(A) \ U in the terminology mentioned above).
Since Prim(A) \ U is closed, it is equal to the hull of its kernel J; in particular,
since U is nonempty, J 6= 0. Note that I ∩ J is the intersection of the kernels of
all irreducible representations; so by Remark 17.3 I ∩ J = 0. Since nontrivial ideals
must intersect nontrivially, and J is nontrivial, it follows that I = 0, and therefore
that its hull is all of Prim(A). But Hull(I) = U so U is dense.
Now applying the Baire category theorem to a countable base for the topology we
find that there is a dense point in Prim(A). This point corresponds to a primitive
ideal which is contained in every primitive ideal, and since the irreducible represen-
tations separate points of A, the only possibility for such an ideal is (0) which is
therefore primitive, as required.
Remark 18.4. It is worthwhile to ask when the natural map  → Prim(A) will be a
bijection (and therefore a homeomorphism). In such a circumstance an irreducible
representation of A is determined up to spatial equivalence by its kernel. We know
this is true for commutative C ∗ -algebras but we will (perhaps?) see that it extends
to a much wider class of algebras, the so-called type I algebras. These include many
algebras occurring in the representation theory of Lie groups, for example. The
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
79
antithesis of type I algebras are antiliminary algebras. UHF algebras (to be studied
later) provide examples of these.
Exercise 18.5. Show that  → Prim(A) is bijective iff  is a T0 space.
Now we will prove Theorem 18.1.
Lemma 18.6. Let A be a C ∗ -algebra, let E be a collection of states of A, and suppose
that for every a > 0 in A we have
sup{σ(a) : σ ∈ E} = kak.
Then the weak-∗ closure of E contains every pure state.
Proof. Once again, we stick with unital A. We’ll show in fact that the weak-∗ closed
convex hull C of E is the entire state space S of A. This will suffice, since the extreme
points of the closed convex hull of E must belong to the closure of E by Milman’s
theorem 11.19.
Think of Asa as a real Banach space, and think of C and S as compact convex
subsets of the LCTVS E which is the dual of Asa equipped with the weak-∗ topology.
The dual of E is then Asa once again. Suppose for a contradiction that C 6= S,
so that there exists ϕ ∈ S \ C. By the Hahn-Banach theorem applied to the space
E, there exist a ∈ Asa and c ∈ R such that ϕ(a) > c whereas σ(a) < c for all
σ ∈ C. We may assume a is positive (add a large multiple of 1). Now we get
ϕ(a) > sup{σ(a) : σ ∈ E} = kak which is a contradiction.
Remark 18.7. A similar but easier Hahn-Banach argument proves that if S is the
state space of a unital C ∗ -algebra A, then the convex hull of S ∪ (−S) is the space of
all real linear functionals on A of norm 6 1. For suppose, if possible, that ψ is a real
linear functional of norm 1 not belonging to the convex hull K of S ∪ (−S). Then
by the Hahn–Banach theorem there exists a ∈ Asa and α ∈ R with ψ(a) > α and
ϕ(a) 6 α for all ϕ ∈ K. The latter condition implies in particular that |σ(a)| 6 α
for all σ ∈ S. But kak = sup{|σ(a)| : σ ∈ S} by Lemma 15.3 and the remark
following, so we get ψ(a) > kak, a contradiction.
Proof of Theorem 18.1. Let Π ⊆ A b be a set of irreducible representations of A, and
let ρ be another such representation. Then ρ belongs to the closure of Π for the
quotient topology from P (A) if and only if
(a) some vector state associated to ρ is a weak-∗ limit of states associated to repre-
sentations π ∈ Π.
On the other hand, ρ belongs to the closure of Π for the pull-back of the Jacobson
topology from Prim(A) if and only if
T
(b) the kernel Ker(ρ) contains {Ker(π) : π ∈ Π}.
We need therefore to prove that conditions (a) and (b) are equivalent.
Suppose
T (b). By passing to the quotient, we may assume without loss of generality
that {Ker(π) : π ∈ Π} = 0. The direct sum of all the representations π is then
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
80
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
81
Lecture 19
Representations by compact operators
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
82
Remark 19.3. We can use these ideas to pay a debt from 9.3, namely, to give the
proof that every simple finite-dimensional C ∗ -algebra is a matrix algebra. Indeed, let
A be such an algebra. Choose a non-trivial state and apply the GNS construction;
we obtain a finite-dimensional representation H which is non-trivial and therefore
faithful (its kernel must be zero). A standard argument shows that if V is a sub-
representation, so is V ⊥ ; so, by induction, H is a direct sum of irreducible faithful
representations. Picking just one of these we see that A is a matrix algebra by
Lemma 19.2.
Corollary 19.4. The algebra K(H) is simple (it has no non-trivial ideals).
Proof. Let J ⊆ K(H) be a non-zero ideal. The subspace J · H of H is invariant
under K(H), so it is all of H. Thus J is non-degenerately represented on H. By
Lemma 19.1, this representation of J is irreducible. By Lemma 19.2, J = K(H).
Corollary 19.5. If A is an irreducible C ∗ -subalgebra of B(H) which contains a
single non-zero compact operator, then A contains all of K(H).
Proof. Consider the ideal J = A ∩ K(H) in A. The space J · H is A-invariant, and
A is irreducible, so JH = H and J is non-degenerate. Now we argue as before: by
Lemma 19.1 J is irreducible, and then by Lemma 19.2, J = K(H).
Remark 19.6. Of course it follows that Prim(K(H)) consists of a single point (cor-
responding to the zero ideal). Let us also go on to prove that K(H) has only one
[ also consists just of a single point. Let
irreducible representation, so that K(H)
0
ρ : K(H) → B(H ) be an irreducible representation. Let P be a rank-one projection
in K(H). Since K(H) is simple, ρ is faithful; so ρ(P ) is a nonzero projection on H 0 .
Now let ξ 0 ∈ H 0 be a unit vector in the range of the projection ρ(P ), and let σ be
the associated state, that is
σ(T ) = hξ 0 , ρ(T )ξ 0 i.
Since ρ is irreducible, ξ 0 is a cyclic vector. By the GNS uniqueness theorem, ρ is
unitarily equivalent to the GNS representation associated to the state σ. But
σ(T ) = hξ 0 , ρ(P T P )ξ 0 i = λT = Tr(P T P )
and the GNS representation associated to this state is the standard one.
A C ∗ -algebra A is called liminal if, for every irreducible representation ρ : A →
B(H), one has ρ(A) = K(H). (Kaplansky originally called these CCR algebras,
for “completely continuous representations”, “completely continuous” being an old
term for what are nowadays called compact operators; but this abbreviation is not
used much anymore.) The argument in the above remark shows that for every
liminal C ∗ -algebra A, A
b = Prim(A); irreducible representations are determined up
to spatial equivalence by their kernels.
Definition 19.7. A positive element x of a C ∗ -algebra A is abelian if the hereditary
subalgebra xAx is commutative.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
83
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
84
Lecture 20
The Toeplitz algebra and its representations
Let H = `2 be the usual sequence space. The unilateral shift operator V is the
isometry H → H given by
V (x0 , x1 , x2 , . . .) = (0, x0 , x1 , . . .).
Notice that V ∗ V = 1, whereas V V ∗ = 1 − P , where P is the rank one projection
onto the span of the basis vector e0 = (1, 0, 0, . . .). Thus V is a Fredholm operator,
of index −1.
Definition 20.1. The Toeplitz C ∗ -algebra T is the C ∗ -subalgebra of B(H) generated
by the unilateral shift operator.
Lemma 20.2. The algebra T is irreducible and contains the compact operators.
Proof. Suppose that K ⊆ H is a closed invariant subspace for T. If K 6= 0, then K
contains a nonzero vector, and by applying a suitable power of V ∗ if necessary we
may assume that K contains a vector v = (x0 , x1 , x2 , . . .) with x0 6= 0. Since P ∈ T
we find that P v = v0 e0 ∈ K and so e0 ∈ K. Now applying powers of V we find
that each en = V n e0 ∈ K, so K = H. Thus T is irreducible. Since it contains the
compact operator P , the second statement follows from Corollary 19.5.
The unilateral shift operator is unitary modulo the compacts, and so its essen-
tial spectrum, that is the spectrum of its image in the Calkin algebra Q(H) =
B(H)/K(H), is a subset of the unit circle S 1 .
Lemma 20.3. The essential spectrum of V is the whole unit circle.
Proof. This is a consequence of Fredholm index theory. Suppose the contrary: then
there is a continuous path of complex numbers λ, running from λ = 0 to λ = 2,
that avoids the essential spectrum of V . But then λ 7→ (V − λ) is a continuous
path of Fredholm operators, and Index V = −1 as we observed above, whereas
Index(V − 2) = 0 since V − 2 is invertible. This is a contradiction.
We see therefore that the C ∗ -algebra T/K is commutative, generated by a single
unitary element whose spectrum is the whole unit circle. Consequently there is a
short exact sequence of C ∗ -algebras
0 → K → T → C(S 1 ) → 0
called the Toeplitz extension.
It is helpful to make the link with complex function theory.
Definition 20.4. Let S 1 denote the unit circle in C. The Hardy space H 2 (S 1 ) is
the closed subspace of L2 (S 1 ) spanned by the functions z n , for n > 0. A Toeplitz
operator on H 2 (S 1 ) is a bounded operator Tg of the form
Tg (f ) = P (gf ) (f ∈ H 2 (S 1 )),
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
85
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
86
Remark 20.7. There is also a matrix version of the Toeplitz index theorem. Taking
n × n matrices over the Toeplitz extension yields
0 → K → Mn (T) → Mn (C(S 1 )) → 0,
since Mn (K) ∼
= K. Thus, any continuous map g : S 1 → GL(n, C) is the symbol of a
matrix Toeplitz operator which is Fredholm, and we can ask about its index. The
answer is minus the winding number of the determinant det g, which is a loop in
C\{0}. To see this, use induction on n, together with a fact from algebraic topology:
every loop S 1 → GL(n), n > 2, is homotopic to the direct sum of a constant loop
and a loop S 1 → GL(n − 1). In turn, this follows from the homotopy equivalence
GL(n) → U (n) and the fact that
U (n − 1) → U (n) → S 2n−1
is a fibration.
Now let us classify the irreducible representations of the Toeplitz algebra T. Let
ρ : T → B(H) be such a representation. We consider two cases:
(a) ρ annihilates the ideal K ⊆ T;
(b) ρ does not annihilate the ideal K ⊆ T.
In case (b), ρ(K) · H is a T-invariant subspace, hence is equal to H; so the
restriction of ρ to K(H) is a non-degenerate representation. Hence it is irreducible
(Lemma 19.1) and thus it is unitarily equivalent to the identity representation. It
follows (Lemma 19.1 again) that the representation ρ of T is unitarily equivalent to
the identity representation of T on H 2 (S 1 ).
In case (a), ρ passes to an irreducible representation of T/K = C(S 1 ). Such a
representation is 1-dimensional, determined by a point of S 1 ; conversely all such
1-dimensional representations of T are (of course!) irreducible.
Thus Tb = Prim(T) consists of the union of S 1 and a disjoint point. What is the
topology? It is easy to see that the circle S 1 gets its usual topology. On the other
hand, the disjoint point corresponds to the zero ideal, which is of course contained
in every other ideal; so that the only neighborhood of that point is the whole space.
Thus the ‘extra point’ cannot be separated from any other representation. We
summarize:
Proposition 20.8. The space T b of irreducible representations of the Toeplitz algebra
T is of the form S t {p0 }, where S 1 has its usual topology and the only open set
1
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
87
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
88
There is a nontrivial involutive ring satisfying these relations, but no C ∗ -algebra can
satisfy them except the zero C ∗ -algebra.
Definition 20.14. A collection R of C ∗ -relations is compact if they imply norm
bounds: that is, if there is some constant K such that, whenever x1 , . . . , xn and their
adjoints satisfy the relations R in a C ∗ -algebra, then kxj k 6 K for j = 1, . . . , n.
Lemma 20.15. If R is a compact collection of C ∗ -relations, then it has a universal
representation. The universal representation is algebraically faithful if and only if
some representation of R is algebraically faithful.
Proof. We need the direct product α Aα of a family of C ∗ -algebras. This is defined
Q
to be the collection of all bounded sequences a = (aα ), aα ∈ Aα , with the supremum
norm
kak = sup{kaα k}.
Making whatever obeisance you find necessary to the gods of set theory, take the
direct product of all the representations of R. This is a C ∗ -algebra and, because of
the compactness condition, the product of all the representatives of xn is an element
of this C ∗ -algebra. Clearly, we have obtained a universal representation.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
89
Lecture 21
The Cuntz Algebras
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
90
T 7→ ... .
T
It follows that the C ∗ -subalgebra An of On defined as the closed linear span of
the words of weight zero is the direct limit of matrix algebras
C → Mn (C) → Mn2 (C) → Mn3 (C) → · · ·
which is called the uniformly hyperfinite (UHF) algebra of type n∞ . We will discuss
UHF algebras, and more general inductive limits of matrix algebras, later in the
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
91
course; we won’t need any of their theory in our discussion of the Cuntz algebra for
now.
Definition 21.4. Let A be a C ∗ -algebra and B a C ∗ -subalgebra of A. A conditional
expectation of A onto B is a positive linear map Φ : A → B which is the identity
on B. (Positive means that Φ carries positive elements to positive elements.) The
expectation Φ is faithful if, whenever a is positive and non-zero, Φ(a) is non-zero
also.
Our analysis of the Cuntz algebra will be based on a certain faithful conditional
expectation of On onto An . We will represent this expectation in a certain sense as
a limit of inner endomorphisms, and we will use this representation to show that On
is simple.
Let z ∈ T be a complex number of modulus one. The elements zs1 , . . . , zsn of On
are isometries which satisfy the defining relation for the Cuntz algebra, so by the
universal property there is an automorphism
θz : On → On
which sends si to zsi for each i. One has θz (sµ s∗ν ) = z |µ|−|ν| sµ s∗ν , so that An is
precisely the fixed-point algebra for the one-parameter group of automorphisms θz .
It is easy to see that z 7→ θz (a) is continuous for each a ∈ On .
Lemma 21.5. With notation as above, the map Φ defined by
Z 2π
1
Φ(a) = θeit (a) dt
2π 0
is a faithful conditional expectation of On onto An .
The proof is easy. We will now show how to approximate Φ by inner endomor-
phisms:
Lemma 21.6. For each positive k one can find an isometry wk ∈ On with the
property that
Φ(a) = wk∗ awk
for every a in the span of those words sµ s∗ν where |µ|, |ν| < k.
Proof. Let us begin with the following observation. Suppose that m > k, that
∗ ∗
x = sm 1 s2 , and that y = sµ sν is a word where |µ|, |ν| 6 k. Then x yx = 0 in all
cases except when y is of the form s`1 s∗` ∗
1 , when x yx = 1. Indeed, in order that x sµ
∗
not equal zero it is necessary that µ be of the form (1, 1, 1, . . .) and in order that
s∗ν x not equal zero it is necessary that ν be of the form (1, 1, 1, . . .). Thus we need
0
only consider y = s`1 s∗`1 and direct computation yields the stated result.
∗
Now put x = s2k
P
1 s 2 and let w = s γ xs γ , where the sum runs over all words sγ
of length k. Notice that w is an isometry. Indeed, we have
X X X
w∗ w = sδ x∗ s∗δ sγ xs∗γ = sγ x∗ xs∗γ = sγ s∗γ = 1.
γ,δ γ γ
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
92
Let us consider what is w∗ yw for words y = sµ s∗ν with |µ|, |ν| 6 k. If |µ| =
6 |ν| then
our previous observation shows that each of the terms in the double sum that makes
up w∗ yw is equal to zero, so the whole sum is zero. Suppose now that |µ| = |ν| = k.10
Then y is a matrix unit sµ s∗ν . We have
X
wy = sγ xs∗γ y = sµ xs∗ν ,
γ
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
93
1
and so v ∗ a∗ av is invertible. Let y = v(v ∗ a∗ av)− 2 and let x = y ∗ a∗ . Then
xay = y ∗ a∗ ay = 1
as required.
Remark 21.9. Nothing we have said so far rules out the possibility that On and Om
might be isomorphic for m 6= n. That is not the case, however. This was proved by
Cuntz using K-theory in a famous paper (Annals, 1981).
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
94
Lecture 22
∗
Basics on Group C -Algebras and Harmonic Analysis
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
95
also, however, make a similar calculation for the right action of G on itself:
x y a b ax bx + y
= .
0 1 0 1 0 1
The Jacobian of (x, y) 7→ (ax, bx+y) is a, which tells us that the right Haar measure
is da db/a.
We’ll use the notation dg instead of dµ(g) for integration with respect to the Haar
measure. Notice the identities (for fixed h)
d(hg) = dg, d(gh) = ∆(h)dg, d(g −1 ) = ∆(g −1 )dg.
To prove the last identity note that both sides define right Haar measures, so they
agree up to a scalar multiple; consider integrating the characteristic function of a
small symmetric neighborhood of the identity (where ∆ ≈ 1) to show that the scalar
is 1.
Definition 22.3. We make L1 (G, µ) into a Banach ∗-algebra by defining
Z
f1 ∗ f2 (g) = f1 (h)f2 (h−1 g)dh, f ∗ (g) = ∆(g)−1 f¯(g −1 ).
G
Exercise: Check that the involution is isometric and that (f1 ∗ f2 )∗ = f2∗ ∗ f1∗ .
Theorem 22.4. There is a one-to-one correspondence between strongly continuous
unitary representations of G and nondegenerate ∗-representations of the Banach
algebra L1 (G).
Proof. To keep matters simple let us assume first of all that G is discrete. (In
this case the Haar measure on G is just counting measure, and we write `1 (G) for
L1 (G).) Let π : G → U (H) be a unitary representation. We define a representation
ρ : L1 (G) → B(H) by
X X
ρ( ag [g]) = ag π(g);
the series on the right converges in norm. This representation has ρ([e]) = 1,
so it is non-degenerate. Conversely, given a representation ρ as above, the map
π : G → U (H) defined by
π(g) = ρ([g])
is a unitary representation of G.
In the general case we must replace summation by integration, handle the compli-
cations caused by the modular function, and use an approximate unit to deal with
the fact that the ‘Dirac masses’ at points of G are no longer elements of L1 (G). This
is routine analysis, the details are on page 183 of Davidson’s book.
The L1 -norm on L1 (G) is not a C ∗ -norm in general. To manufacture a C ∗ -algebra
related to the representation theory of G, we should form a completion of L1 (G) in an
appropriate C ∗ -norm. There are several different ways to form such a completion (as
will become apparent, this kind of issue turns up many times in C ∗ -algebra theory.)
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
96
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
97
Lemma 22.7. Let G be a discrete (or a locally compact) group. The commutant of
the left regular representation of G on L2 (G) is the von Neumann algebra generated
by the right regular representation (and vice versa of course).
Proof. Obviously the left and right regular representations commute. The commu-
tant R = λ[G]0 of the left regular representation consists of those operators on
L2 (G) that are invariant under the left-translation action of G on L2 (G), and the
commutant L = ρ[G]0 of the right regular representation consists of those operators
that are invariant under the right-translation action. It will be enough to show that
L and R commute, since then ρ[G] ⊆ R ⊆ (L)0 = ρ[G]00 .
So let S ∈ L and T ∈ R. For simplicity, assume that G P is discrete. Then (by
translation invariance) S is completelyP determined by S[e] = sx [x], and similarly
T is completely determined by T [e] = ty [y].
For group elements g and h we have
DX X E X
hT S[g], [h]i = hS[g], T ∗ [h]i = sx [gx], t̄y−1 [yh] = sx thx−1 g−1 .
By a similar calculation
X
hST [g], [h]i = ty sg−1 y−1 h .
The substitution x = g −1 y −1 h shows that these sums agree, and thus T S = ST .
Let’s take a moment to develop an important fact about these group von Neumann
algebras. For definiteness consider the algebra L(G) generated by the left regular
representation of G (i.e. the commutant of the right regular representation). There
is an important linear functional τ on L(G) defined by
τ (T ) = h[e], T [e]i,
where [e] is the basis element of `2 (G) corresponding to the identity of G. Clearly
τ is a state (in fact a vector state).
Lemma 22.8. The state τ is a trace on L(G); that is, τ (ST ) = τ (T S) for all
S, T ∈ L(G).
Proof. Let T ∈ L(G). Since T is in the commutant of the right regular representa-
tion, for any g ∈ G,
h[g], T [g]i = h[g], T [e.g]i = h[e]g, (T [e])gi = h[e], T [e]i = τ (T ).
Thus τ (g −1 T g) = τ (T ) for all T . Since τ is weakly continuous it follows that
τ (ST ) = τ (T S) for all S ∈ L(G).
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
98
Lecture 23
Some More Harmonic Analysis
Proof. All we need to do is to show that C ∗ (G) = Cr∗ (G), and for this it will be
enough to show that every character α of G extends to a continuous linear map
Cr∗ (G) → C.
Let us first establish the following: pointwise multiplication by a character of G
preserves the norm on the regular representation. In other words, if f ∈ L1 (G) and
α is a character, then
kλ(αf )k = kλ(f )k.
To see this write
λ(αf ) = Uα λ(f )Uα∗
where Uα : L2 (G) → L2 (G) is the unitary operator of pointwise multiplication by α.
It follows that if α is any character of G, and β is a character which extends
continuously to Cr∗ (G) (that is, β ∈ Gbr ), then the pointwise product αβ belongs to
G
br also. Since G b is a group under pointwise product, and G br is nonempty (after all,
Cr∗ (G) must have some characters), we deduce that G b=G br .
We can use our results to discuss the bare bones of the Pontrjagin duality theory
for abelian groups. For starters let G be a discrete abelian group. As we saw above,
Gb is a compact abelian group also, and C ∗ (G) = Cr∗ (G) = C(G). b The implied
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
99
The phrase ‘for an appropriate choice of Haar measure’ conceals all the various
powers of 2π which show up in the classical theory of Fourier analysis.
Proof. The regular representation of Cr∗ (G) is a cyclic representation with cyclic
vector [e]. The associated vector state
τ (f ) = h[e], f ∗ [e]i
corresponds to a measure on G b (by the Riesz representation theorem). Translation
by a character α on C(G) b corresponds to pointwise multiplication by α on Cr∗ (G);
and since α(e) = 1 we see that the measure corresponding to τ is translation invari-
ant, so it is (up to a multiple) the Haar measure ν on G.b By the uniqueness of the
∗
GNS construction, the regular representation of Cr (G) is spatially equivalent to the
multiplication representation of C(G) b on L2 (G,
b ν), with cyclic vector equal to the
constant function 1. In particular there is a unitary U : L2 (G) → L2 (G)
b such that
U (f ∗ h) = (Ff ) · (U h)
for f ∈ L1 (G), h ∈ L2 (G). In particular, taking h = [e] (which is the cyclic vector,
so that U h = 1), we see that U f = Ff for f ∈ L1 . Thus the unitary U is an
extension of the Fourier transform, as asserted.
The Plancherel theorem extends to non-discrete locally compact groups too. Now
the representing measure ν on G
b is not finite, and we cannot appeal directly to the
GNS theory. Instead we build the representing measure by hand, in the following
way. For each x ∈ L1 ∩ L2 (G) we define a measure νx by the GNS construction, so
that Z
Ff dνx = hx, f ∗ xiL2 (G) .
G
b
If both x and y belong to L ∩ L2 (G) then a simple argument using commutativity
1
shows that
|Fx|2 νy = |Fy|2 νx
as measures on G.
b For any compact subset K of G b one can find an x ∈ L1 ∩ L2 (G)
such that Fx is positive on K. Now we may define an unbounded measure ν on
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
100
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
101
Lecture 24
Amenability I
From now on, therefore, let us restrict attention to discrete and countable, but not
necessarily abelian, groups. Let G be such a group. A function p on such a group is
called positive definite if for every positive integer n and every n-tuple (x1 , . . . , xn ) of
group elements, the n × n matrix whose entries are p(x−1 j xi ) is positive as a matrix;
that is, for all α1 , . . . , αn ∈ C we have
n
X
(24.1) αi ᾱj p(x−1
j xi ) > 0.
i,j=1
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
102
Proof. To prove that the product of positive definite functions is again positive def-
inite, it is necessary to prove that the pointwise product (Schur product) of positive
matrices is positive. This is a standard result of matrix theory. We reproduce the
proof. Let A = (aij ) and B = (bij ) be positive matrices. Their Schur product
C =P (aij bij ). Now because B is positive, B = X ∗ X for some matrix X; that is,
bij = k xik x̄jk . It follows that for any vector ξ,
XX
ξ ∗ Cξ = aij (x̄jk ξ¯j )(xik ξi )
k i,j
and for each k the inner sum is positive. Hence C is a positive matrix, as required.
Remark 24.4. Since (normalized) positive definite functions correspond to states on
C ∗ (G), which in turn correspond via the GNS construction to (cyclic) representa-
tions of G, it is natural to ask: is there an operation on representations of G which
corresponds to the pointwise product of positive definite functions on G? The an-
swer of course is the tensor product. Suppose that ρk : G → U (Hk ) are unitary
representations with cyclic vectors ξk ∈ Hk , k = 1, 2. The corresponding positive
definite functions on G are then
ϕk (g) = hξk , ρk (g)ξk i.
Their pointwise product is the positive definite function associated to the tensor
product ρ = ρ1 ⊗ ρ2 on H = H1 ⊗ H2 , with cyclic vectot ξ = ξ1 ⊗ ξ2 .
It follows that B(G) is an algebra under pointwise multiplication. In fact it is a
Banach algebra: that is, with the norm described above (as the dual space of C ∗ (G))
we have kpqk 6 kpkkqk for p, q ∈ B(G). This is obvious if p, q are positive definite
because positive definite functions attain their norm at the identity element of G,
as we mentioned above. The general case follows from this. For example, suppose p
and q are real (i.e., correspond to self-adjoint linear functionals on C ∗ (G)) and have
norm 6 1. Then by Remark 18.7, each of them can be written
p = sp+ − (1 − s)p− , q = tq+ − (1 − t)q− ,
where s, t ∈ [0, 1] and p± , q± are positive definite functions of norm 1. It follows that
pq can also be written in this way, that is as λ(pq)+ − (1 − λ)(pq)− , where (pq)± are
positive definite of norm 1 and λ ∈ [0, 1]. Therefore pq has norm 6 1 as required.
Definition 24.5. B(G) is called the Fourier-Stieltjes algebra of G.
To explain this terminology, consider the case G = Z. The positive definite
functions on Z are the Fourier series of positive measures on the circle, and thus the
algebra B(G) is just the algebra of Fourier series of measures on the circle, classically
known as the Stieltjes algebra.
Can every positive definite function on G be approximated (in some sense) by
compactly supported positive definite functions? This turns out to be a delicate
question, and its answer leads us to consider the relationship between the maximal
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
103
C ∗ -algebra C ∗ (G) and the reduced C ∗ -algebra Cr∗ (G) which is associated to the
regular representation. To see why, suppose that p(g) = hξ, λ(g)ξi is the positive
definite function associated to a vector state of the regular representation, ξ ∈ `2 (G).
If ξ has finite support, so does the function p. Moreover, we have
kpkB(G) 6 kξk2 ;
and therefore, since vectors v of finite support are dense in `2 (G), we find that every
positive definite function associated to a vector state of the regular representation
of G belongs to the closure (in B(G)) of the space of functions of finite support.
Conversely we also have:
Lemma 24.6. Any positive definite function of finite support is associated to a
vector state of the regular representation.
Proof. Let p be such a function. The operator T η = η ∗ p of right convolution by p
is bounded on `2 (G), by finiteness of the support; and the fact that p is a positive
definite function shows11 that T is a positive operator. Moreover, T commutes with
1
the left regular representation λ[G]. Now put ξ = T 2 [e] (notice that v need not
have finite support). Then for g ∈ G,
1 1
hξ, λ(g)ξi = hT 2 [e], λ(g)T 2 [e]i = hT [e], [g]i = hp, [g]i = p(g)
as required.
Denote by A+ (G) the space of positive definite functions on G of the form
X
p(g) = hξi , λ(g)ξi i
i
2
kξi k2 < ∞ (notice that this implies that
P
where ξi is a sequence in ` (G) having
the above sequence converges uniformly to a positive definite function.) More canon-
ically we may write
p(g) = Tr(λ(g)T )
where T is a positive trace-class operator on `2 (G). We have kpkB(G) = kξi k2 =
P
kT k1 , the trace norm of T , and it follows from the fact that the space of trace-class
operators is a Banach space under the trace norm that A+ (G) is a closed cone in
B+ (G). In fact the arguments above prove
Proposition 24.7. A+ (G) is the closure of C[G] ∩ B+ (G) (the compactly supported
positive definite functions on G) in B+ (G).
Indeed, Lemma 24.6 shows that every compactly supported element of B+ (G)
belongs to A+ (G), and the remarks preceding the lemma show that every element
of A+ (G) can be approximated by compactly supported elements of B+ (G).
11Check this!
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
104
We can also introduce the notation A(G) for the subalgebra of B(G) generated
by A+ (G). It is the closure of C[G] in B(G). Elements of A(G) are functions of the
form X
p(g) = hξi , λ(g)ηi i = Tr(λ(g)T )
where kξi k2 < ∞ and kηi k2 < ∞ (the trace-class operator T need no longer be
P P
positive). In other words, the corresponding elements of C ∗ (G)∗ are precisely those
that arise, via the regular representation, from the ultraweakly continuous linear
functionals on B(`2 (G)).
Remark 24.8. The algebra A(G) is called the Fourier algebra of G; remember that
B(G) is the Fourier–Stieltjes algebra. In the example G = Z, the algebra A(G) is
generated by sequences of the form
Z 2π
n 7→ eint u(t)v̄(t) dt
0
where u, v ∈ L (T). This is the sequence of Fourier coefficients of the L1 function
2
uv̄, and the collection of products of this sort has dense span in L1 , so we see that
A(Z) is just the algebra of Fourier transforms of L1 functions on the circle.
Lemma 24.9. Let G be a discrete group. The weak-∗ topology on a bounded subset
of C ∗ (G)∗ is equivalent to the topology of pointwise convergence on the corresponding
bounded subset of B(G).
Proof. It is apparent that weak-∗ convergence in the dual of C ∗ (G) implies point-
wise convergence in B(G). Conversely suppose that pλ is a bounded net in B(G)
converging pointwise to p. Then the corresponding functionals
X X
ϕλ ag [g] = ag pλ (g)
converge to X X
ϕ ag [g] = ag p(g)
ag [g] belonging to `1 (G). Since the {ϕλ } are uniformly bounded and
P
for all a =
` (G) is norm dense in C ∗ (G), it follows that ϕλ (a) → ϕ(a) for all a ∈ C ∗ (G).
1
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
105
Lecture 25
Amenability II
Theorem 25.1. Let G be a countable discrete group. Then the following are equiv-
alent:
(a) Cr∗ (G) = C ∗ (G),
(b) The Banach algebra A(G) possesses a bounded approximate unit,
(c) Every function in B(G) can be pointwise approximated by a norm bounded se-
quence of compactly supported functions.
Proof. Suppose (c). Then (by lemma 24.9) each state σ of C ∗ (G) is a weak-∗ limit of
vector states a 7→ hξ, λ(a)ξi, where ξ is a unit vector in the regular representation.
It follows that for each such state σ,
|σ(a)| 6 kλ(a)k
for all a ∈ C (G); and consequently kakC ∗ (G) 6 kλ(a)k, whence C ∗ (G) = Cr∗ (G),
∗
which is (a).
Now suppose (a), that C ∗ (G) = Cr∗ (G). Let a ∈ C ∗ (G) have ϕ(a) = 0 for each
ultraweakly continuous linear functional ϕ on B(`2 (G)). Then λ(a) = 0 and thus
a = 0 since λ is faithful on C ∗ (G). It follows that the space of ultraweakly con-
tinuous functionals on the regular representation is weak-∗ dense in C ∗ (G)∗ . Using
lemma 24.9, this implies that A+ (G) is dense in B+ (G) in the topology of pointwise
convergence. Now the constant function 1 belongs to B+ (G) (it corresponds to the
trivial representation of G) and thus there is a net in A+ (G) converging pointwise
to 1. Since A(G) is generated by compactly supported functions, such a net is nec-
essarily an approximate unit for A(G), and it is bounded because the norm in A(G)
is given by evaluation at e. This proves (b).
Finally, suppose (b). Let fλ be an approximate unit for A(G). We may assume
without loss of generality that fλ consists of compactly supported functions. Let
f ∈ B(G); then f fλ is a net in A(G) and it converges pointwise to f .
Let G be a discrete group. An invariant mean on G is a state m on the com-
mutative C ∗ -algebra `∞ (G) which is invariant under left translation by G: that is
m(Lg f ) = m(f ), where Lg f (x) = f (g −1 x).
Definition 25.2. A group G that possesses an invariant mean is called amenable.
Example 25.3. The group Z is amenable. Indeed, consider the states of `∞ (Z)
defined by
n
1 X
σn (f ) = f (j),
2n + 1 j=−n
and let σ be a weak-∗ limit point of these states (which exists because of the com-
pactness of the state space). Since for each k ∈ Z, and each f ,
2k
|σn (Lk f − f )| 6 kf k
2n + 1
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
106
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
107
and these facts together show that we may assume without loss of generality that
each ϕλ is a positive function of norm one. For each g ∈ G we have Lg ϕλ − ϕλ → 0
weakly in `1 (G) as λ → ∞. Now we use the fact (a consequence of the Hahn-Banach
Theorem) that the weak and norm topologies on `1 have the same closed convex
sets to produce a sequence {ψ1 , ψ2 , . . .} of convex combinations of the {ϕλ } such
that for each g ∈ G, Lg ψn − ψn → 0 in the norm topology of `1 (G). Each ψn is a
positive function of `1 -norm equal to one. Take ξn to be the pointwise square root
of ψn .
Remark 25.7. Examining the proof, we see that we can assume without loss of
generality that the vectors ξn are finitely supported.
Now we can finally prove
Theorem 25.8. A discrete group G is amenable if and only if Cr∗ (G) = C ∗ (G).
Proof. Suppose that G is amenable. Then by Proposition 25.6 there is a sequence
{pn } of positive definite functions on G of norm one , associated to vector states of
the regular representation, that tend pointwise to 1. These functions pn constitute
a bounded approximate unit for A(G), so by Theorem 25.1, C ∗ (G) = Cr∗ (G).
Conversely suppose that C ∗ (G) = Cr∗ (G). Then A(G) has a bounded approximate
unit, which we may assume without loss of generality is made up of compactly
supported positive definite functions of norm one. By Lemma 24.6, such functions
are associated to vector states of the regular representation. By Proposition 25.6,
G is amenable.
To conclude let us prove Følner’s theorem.
Proposition 25.9. Any (discrete) amenable group admits a Følner sequence.
Proof. It will suffice to show that for any finite subset S ⊆ G and any ε > 0 there
exists a finite subset F ⊆ G with
#(xF 4 F )
< ε ∀x ∈ S;
#F
a diagonal argument then completes the proof. By Proposition 25.6, there exists a
positive, finitely supported function ψ with kψk`1 = 1 on G such that kψ −Lx ψk`1 <
ε/#S for all x ∈ S.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
108
PNow
k
we use what Terry Tao calls the “layer-cake decomposition” to write ψ =
i=1 ci χEi for some nested non-empty sets E1 ⊃ E2 ⊃ . . . ⊃ Ek and some positive
constants ci . Since ψ has norm one,
X
ci #Ei = 1.
i
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
109
Lecture 26
Free Group C ∗ -Algebras
Let G be a discrete group. The canonical trace on Cr∗ (G)) is the vector state τ
associated with [e] ∈ `2 (G); that is,
τ (a) = h[e], λ(a)[e]i.
The associated positive definite function is equal to 1 at the identity and 0 at every
other group element. Notice that we may also write
τ (a) = h[g], λ(a)[g]i
for any fixed g ∈ G (the right-hand side agrees with τ on group elements, and it
is continuous, so it agrees with τ everywhere). As we already observed in the von
Neumann algebra context (Lemma 22.8), this implies that the state τ is a trace,
that is, τ (aa0 ) = τ (a0 a) for every a, a0 ∈ Cr∗ (G).
Remark 26.1. The trace τ is faithful, meaning that if a ∈ Cr∗ (G) is positive and
τ (a) = 0, then a = 0. Indeed, this is a consequence of the fact that the regular
representation (a faithful representation) is the GNS representation associated to τ .
To prove it from first principles, use the Cauchy–Schwarz inequality to estimate the
matrix coefficients of λ(a):
1 1
|h[g], λ(a)[g 0 ]i| 6 hg, λ(a)gi 2 hg 0 , λ(a)g 0 i 2 = τ (a).
We are going to use the faithful trace τ as a key tool in analyzing the reduced
group C ∗ -algebra Cr∗ (F2 ). Notice that since the free group is non-amenable, this
is different from the full C ∗ -algebra C ∗ (F2 ). We are going to distinguish them by
C ∗ -algebraic properties. Specifically, we will prove that Cr∗ (F2 ) is simple. This is
very, very different from C ∗ (F2 ): let G be any finite group with two generators, then
according to the universal property of C ∗ (F2 ) there is a surjective ∗-homomorphism
C ∗ (F2 ) → C ∗ (G), and of course the kernel of this ∗-homomorphism is a non-trivial
ideal.
Here is the key lemma. Denote by u, v the unitaries in Cr∗ (F2 ) corresponding to
the generators x, y of the group.
Lemma 26.2. Let a ∈ Cr∗ (F2 ). Then the limit
m n
1 X X i j −j −i
lim u v av u
m,n→∞ mn
i=1 j=1
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
110
Proof. Start with a special case: assume that T maps H (and not just H2 ) into H1 .
Write
n n
X
2
X
2
Ui T Ui∗
=
T + U1∗ Ui T Ui∗ U1
.
i=1 i=2
The first term T in the sum above has range contained in H1 , whereas the remaining
sum has range contained in H2 (since U1∗ Ui maps H1 into H2 ). Thus the ranges of
the two terms are orthogonal and we get
n n
X
2
X
2
Ui T Ui∗
6 kT k2 +
Ui T Ui∗
.
i=1 i=2
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
111
By induction we get
n
X
2
Ui T Ui∗
6 nkT k2
i=1
giving the result (without the factor 2) in the special case. The general case follows
since any operator T that maps H2 into H1 can be written as T = T 0 + (T 00 )∗ , where
kT 0 k 6 kT k, kT 00 k 6 kT k, and both T 0 and T 00 map H into H1 (to see this, put
T 0 = T P2 and T 00 = P1 T ∗ , where P1 and P2 are the orthogonal projections of H
onto H1 and H2 respectively.
Now we will use this to prove
Pm Lemma
Pn 26.2. It will be enough to prove that if
1
a = [g], then the average mn i=1 j=1 u i j
v av −j u−i converges in norm to 1 if g = e
and to 0 otherwise. For simplicity work with one sum at a time, so consider the
expression
n
1 X j −j
Sn (a) = v av .
n j=1
1
Clearly, Sn ([g]) = [g] if g is a power of y. Otherwise, I claim, kSn ([g])k 6 2n− 2 .
To see this decompose the Hilbert space H = `2 (G) as H1 ⊕ H2 , where H1 is the
subspace of `2 spanned by those free group elements represented by reduced words
that begin with a non-zero power of x (positive or negative), and similarly H2 is
spanned by those free group elements represented by reduced words that begin with
a non-zero power of y (positive or negative), together with the identity. Suppose
that g is not a power of y. Then the reduced word for g contains at least one x and
by multiplying on the left and right by suitable powers of y (an operation which
commutes with Sn and does not change the norm) we may assume that g begins
and ends with a power of x. Now we apply Lemma 26.6 with T the operation of left
multiplication by g and Ui the operation of left multiplication by v i . Once we have
checked that these elements indeed map H2 and H1 in the way that Lemma 26.6
applies, the result follows immediately.
To complete the proof of Lemma 26.2 all we need to do is to reiterate the above
argument to show that
m n
1 X X i j −j −i 4
k u v av u k 6 √ √
mn i=1 j=1 min{ m, n}
whenever a = [g] and g is not the identity.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
112
Lecture 27
Free Group C ∗ -Algebras II
Now we are going to prove that the reduced C ∗ -algebra Cr∗ (G) for G = F2 has no
non-trivial projections. Thus Cr∗ (G) is an example of a simple, unital C ∗ -algebra
with no non-trivial projections. Such examples had long been sought in C ∗ -algebra
theory. This is not the first such example (that was constructed by Blackadar in the
seventies using the theory of AF algebras), but it is the ‘most natural’.
We begin with some abstract ideas which lie at the root of K-homology theory for
C ∗ -algebras. They are developments of a basic notion of Atiyah (1970) to provide
a ‘function-analytic’ setting for the theory of elliptic pseudodifferential operators.
Definition 27.1. Let A be a C ∗ -algebra. A Fredholm module M over A consists
of the following data: two representations ρ0 , ρ1 of A on Hilbert spaces H0 and
H1 , together with a unitary operator U : H0 → H1 that ‘almost intertwines’ the
representations, in the sense that U ρ0 (a) − ρ1 (a)U is a compact operator from H0
to H1 for all a ∈ A.
It is often convenient to summarize the information contained in a Fredholm
module in a ‘supersymmetric’ format: we form H = H0 ⊕ H1 , which we consider
as a ‘super’ Hilbert space (that is, the direct sum decomposition, which may be
represented by the matrix γ = ( 10 −1
0
), is part of the data), we let ρ = ρ0 ⊕ ρ1 , and
we put
0 U∗
F = ,
U 0
which is a self-adjoint operator, odd with respect to the grading (that is, anticom-
muting with γ), having F 2 = 1, and commuting modulo compacts with ρ(a) for each
a ∈ A. Thus we may say that a Fredholm module is given by a triple M = (ρ, H, F ).
Lemma 27.2. Suppose that a Fredholm module over the C ∗ -algebra A is given (and
is denoted as above). Then for each projection e ∈ A the operator Ue = ρ1 (e)U ρ0 (e)
is Fredholm, when considered as an operator from the Hilbert space ρ0 (e)H0 to the
Hilbert space ρ1 (e)H1 .
Proof. We have
Ue Ue∗ = ρ1 (e)U ρ0 (e)U ∗ ρ1 (e) ∼ ρ1 (e)U U ∗ ρ1 (e) = ρ1 (e)
where we have used ∼ to denote equality modulo the compact operators. Similarly
Ue∗ Ue ∼ ρ0 (e). Thus Ue is unitary modulo the compacts.
It follows that each Fredholm module M on A gives rise to an integer-valued
invariant
IndexM : Proj(A) → Z
defined on the space of projections in A.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
113
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
114
Thus
Ue∗ Ue − e0
0
τM (e) = Tr(γρ(e)[F, ρ(e)]F ρ(e)) = Tr .
0 e1 − Ue Ue∗
as asserted.
Definition 27.5. We will say that a Fredholm module over a unital C ∗ -algebra A
is summable if its domain of summability is a dense, unital subalgebra of A.
Theorem 27.6. Let A be a unital C ∗ -algebra. Suppose that there exists a summable
Fredholm module M over A, for which the associated trace τM on the dense subal-
gebra A ⊆ A is positive, faithful, and has τM (1) = 1. Then A has no non-trivial
projections.
shows that the integrand (z1 − x)−1 belongs to A and is a continuous function, in
the norm of A, of z ∈ Γ∗ . Therefore the integral converges in A and q ∈ A as
asserted.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
115
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
116
of `(v) only for those finitely many vertices v that lie along the geodesic segment
from e to ge. Thus λ(g)U λ(g −1 ) − U is of finite rank. It follows that U commutes
modulo finite rank operators with elements of C[G], and hence by an approximation
argument that U commutes modulo compact operators with elements of Cr∗ (G). We
have shown that our data define a summable Fredholm module.
To complete the proof it is necessary to identify the trace τM on the domain
of summability. We need only compute τM ([g]) for group elements g. If g = e
is the identity, we easily calculate τM ([e]) = 1. (For instance this follows from
Proposition 27.4 applied to the projection [e] = 1.) If g is not the identity, then to
compute τM ([g]) = 12 Tr(γF [F, [g]]), we must compute traces of operators such as
λ(g) − U ∗ λ(g)U
on H0 . Represent this operator by an infinite matrix relative to the standard basis
of H0 . The diagonal entries of the matrix are all zeros, so the trace is zero. It follows
that X
τm ( ag [g]) = ae
and thus τM agrees with the canonical trace on Cr∗ (F2 ), which we know to have the
properties listed in Theorem 27.6.
This completes the proof of
Theorem 27.7. The C ∗ -algebra Cr∗ (F2 ) has no non-trivial projections.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
117
Lecture 28
Completely Positive Maps and Nuclearity I
Recall
Definition 28.1. Let A and B be C ∗ -algebras. A linear map (usually bounded)
ϕ : A → B is said to be positive if ϕ(a) > 0 whenever a > 0. If A and B are unital
and ϕ(1) = 1 it is unital positive.
A unital positive map is automatically bounded, with norm 1. Since an element
a ∈ A is positive iff a = x∗ x, we see that ∗-homomorphisms are positive maps.
Let ϕ : A → B be a linear map. Then ϕ also induces linear maps (inflations)
ϕn : Mn (A) → Mn (B). If ϕ is a ∗-homomorphism (and therefore positive) then all
its inflations are too. However, in general the inflation of a positive map need not
be positive: the standard example is the transposition map on A = M2 (C). Indeed,
consider the matrices in M4 (C) = M2 (M2 (C))
1 0 0 1 1 0 0 0
0 0 0 0 0 0 1 0
0 0 0 0 and 0 1 0 0 ;
1 0 0 1 0 0 0 1
the first is positive, the second is not, but the second is obtained by transposing the
2 × 2 blocks of the first.
Definition 28.2. A positive linear map ϕ is completely positive if all its inflations
ϕn are positive also.
Remark 28.3. We use the abbreviations cp = “completely positive” and ucp =
“unital completely positive”.
Clearly a ∗-homomorphism is completely positive. To get a more general under-
standing of complete positivity we need to know something about positive elements
of matrix algebras.
Lemma 28.4. Let A be a C ∗ -algebra and let X = (xij ) ∈ Mn (E). The following
conditions are equivalent:
(i) X is positive in Mn (A);
(ii) X belongs to the span of the matrices aa∗ = (ai a∗j ), where a = (ai ) is a column
vector in An .
(iii) For every column vector b ∈ An , the element b∗ Xb = i,j b∗i xij bj is positive
P
in A.
∗
Proof. For (i)⇒(ii), suppose that X is positive. Then
∗
P X ∗= Y Y where Y ∈ Mn (A).
Let a1 , . . . , an be the columns of Y . Then Y Y = i ai ai (check this!). This proves
(ii).
For (ii)⇒(iii), note that if X = aa∗ , then b∗ Xb = (b∗ a)(b∗ a)∗ > 0.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
118
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
119
It follows that ϕ(a∗ a) = ϕ(a∗ )ϕ(a) if and only if β(a) = 0. Similarly, ϕ(aa∗ ) =
ϕ(a)ϕ(a∗ ) if and only if α(a) = 0. Thus the set M of a ∈ A with ϕ(a∗ a) = ϕ(a)∗ ϕ(a)
and ϕ(aa∗ ) = ϕ(a)ϕ(a)∗ is a subalgebra, characterized by α(a) = β(a) = 0; it is
called the multiplicative domain of ϕ. If a ∈ M and a0 ∈ A then
ϕ(aa0 ) = ϕ(a)ϕ(a0 ), ϕ(a0 a) = ϕ(a0 )ϕ(a)
(the bimodule property).
As might be expected, cp maps involving matrix algebras are particularly impor-
tant and explicit. These characterizations are due to Choi.
Example 28.8. (CP maps from matrices) Let B be a unital C ∗ -algebra and
ϕ : Mn (C) → B a linear map. Then ϕ is completely positive if and only if the
matrix Φ = [ϕ(eµν )] is positive in Mn (B), where eµν are the standard matrix units.
To prove this, go back to the working definition (Corollary 28.5). We see that ϕ is
completely positive if and only if the expression
X
S= b∗i ϕ(a∗i aj )bj
i,j
is positive. To see this, we once again use the “working definition” 28.5 of complete
positivity. First, suppose that ϕ is a cp map. Then for all lists b1 , . . . , bn in Mn (C)
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
120
Take bν = eν1 (the matrix unit). Then the left side is Φ(a)e11 , so we must have
Φ(a) > 0.
For the converse, suppose that Φ is known to be a positive map. The map ϕ is
completely positive if and only if for all lists ai in A and bi in Mn (C), and all ξ ∈ Cn ,
the expressions X
S0 = ξ ∗ b∗i ϕ(a∗i aj )bj ξ > 0
i,j
are
P positive in C. Write each b as a linear combination of matrix units, say bi =
µν 0
c
µ,ν i e µν Then we can expand S as
.
X µν
S0 = ξ¯ν c̄i ϕ(a∗i aj )µρ cρσ
j ξσ .
P ρσ
Put yρ = cj aj ξσ ∈ A; then
X
S0 = ϕ(yµ∗ yρ )µρ = Φ(Y) > 0,
since the matrix Y = y∗ y is positive.
Here is an important corollary.
Lemma 28.11. Suppose that A ⊆ B are C ∗ -algebras. Every ucp map from A to a
matrix algebra extends to a ucp map from B to the same matrix algebra.
Proof. By Example 28.10, ucp maps from A to Mn (C) are in one-to-one correspon-
dence with states on Mn (A). Thus, we need only show that every state ϕ on A
extends to a state on B. By the Hahn-Banach theorem, there is a linear functional
ψ on B of norm one and extending ϕ. We must show that ψ is positive, but this
follows from Proposition 14.11.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
121
Lecture 29
Completely Positive Maps and Nuclearity II
We will be interested in limits of completely positive maps. Such limits are con-
sidered in the point-norm sense: a sequence (or net) ϕn of completely positive maps
A → B converges (in this sense) to ϕ if and only if ϕn (a) → ϕ(a) (in norm) for each
a ∈ A. Clearly, complete positivity is preserved under this sort of limit.
Definition 29.1. A completely positive map ϕ : A → B is factorable if it is equal
to a composite of completely positive maps A → Mn (C) → B (for some n ∈ N). It
is nuclear if it is the point-norm limit of factorable maps. A C ∗ -algebra A is nuclear
if the identity map A → A is nuclear.
Example 29.2. Every finite-dimensional C ∗ -algebra is a direct sum of matrix al-
gebras, and is therefore nuclear. In fact, we obtain the same definition if we replace
Mn (C) by a general finite-dimensional algebra in the definitions above.
Example 29.3. Every commutative (unital) C ∗ -algebra A = C(X) is nuclear, by
the following argument. Let {U1 , . . . , UN } be a finite open cover of X, let xi ∈ Ui
and let {fi } be a subordinate partition of unity. Then the maps
A → CN , f 7→ (f (x1 ), . . . , f (xN ))
and X
CN → A, (λ1 , . . . , λN ) 7→ λi f i
i
are completely positive and, as the mesh of the cover decreases, their composite
point-norm approximates the identity.
Nuclearity has many applications in C ∗ -algebra theory. Here we will develop one
which is particularly relevant to K-homology.
Let 0 → J → A → B → 0 be a short exact sequence of C ∗ -algebras and ∗-
homomorphisms (an extension). A splitting is a map B → A which is a right
inverse to the quotient map A → B.
We know from algebra that the existence of a splitting which is a homomorphism
implies that the extension is a direct sum, A = J ⊕ B. But we might ask for a
splitting which has weaker properties, such as complete positivity. An extension
that admits a completely positive splitting is called semisplit.
Proposition 29.4. (Choi-Effros lifting theorem) Every extension (as above)
for which the quotient B = A/J is separable and nuclear is semisplit.
Proving this will require some new technology, foreshadowed in Remark 8.17.
Recall (Definition 8.1) that an approximate unit for a C ∗ -algebra is an increasing net
of positive contractions uλ such that the operations of left (and right) multiplication
by uλ tend to 1 in the point-norm topology. Now suppose that the algebra J for
which we are forming an approximate unit is in fact an ideal in some larger C ∗ -
algebra A. Consider the expressions uλ a and auλ , for a ∈ A. We should not expect
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
122
these to converge to a in general (after all, they belong to the ideal J), but we can
still ask what happens to their difference:
Definition 29.5. An approximate unit {uλ } for an ideal J C A is quasicentral if
uλ a − auλ → 0 for all a ∈ A.
Proposition 29.6. Quasicentral approximate units always exist.
Proof. Let D = {uλ } be the canonical approximate unit (Theorem 8.2). I claim that
for each fixed a ∈ A the commutators auλ − uλ a tend to zero in the weak topology
of the Banach space A. Indeed, we must show that ϕ(auλ − uλ a) → 0 for all ϕ ∈ A∗ ,
and it suffices to consider the case where ϕ is a state. Let π be the corresponding
GNS representation of A on a Hilbert space H. Note that π(uλ ) tends strongly
to the orthogonal projection onto the closed (14.7) subspace K = π(J)H (check
this separately for ξ ∈ K and ξ ∈ K ⊥ ). But K is an A-invariant subspace, so P
commutes with A, giving the result.
Now we apply the Hahn-Banach theorem to prove the following assertion: for
fixed a ∈ A and ε > 0 and any u ∈ D, there is v > u in D such that kav − vak < ε.
Indeed, let C = {av − va : v > u, v ∈ D}; it is a closed convex set (as D is closed
under the formation of convex combinations) and we want to prove it contains 0. If
it didn’t, by the Hahn-Banach theorem there would exist a linear functional ϕ that
has absolute value > 1 on C, and that contradicts the first part.
A simple “inflation” argument extends the previous paragraph to finite subsets:
for fixed a1 , . . . , ak ∈ A and ε > 0 and any u ∈ D, there is v > u in D such
that kaj v − vaj k < ε for j = 1, . . . , k. Now use this to define a subnet {uµ } of D
(parameterized by µ ∈ F × R+ × D, where F is the set of finite subsets of A) for
which kauµ − uµ ak → 0 for all a; that is, a quasicentral approximate unit.
Let J C A and let B be another C ∗ -algebra. A unital completely positive map
B → A/J is liftable if there exists another unital completely positive map B → A
making the obvious diagram commute. We will use quasicentral approximate units
to prove
Proposition 29.7. In the above situation, suppose B is a separable, unital C ∗ -
algebra. Then the collection of liftable ucp maps B → A/J is closed in the point-
norm topology.
Proof. Let ϕn : B → A/J be a sequence of ucp maps converging in the point-norm
topology to ϕ, and suppose that ψn : B → A are liftings of ϕn . We want to modify
the ψn inductively to new liftings ψn0 which also converge in the point-norm topology;
then their limit ϕ will be a lifting of ψ.
Let {bj } be a dense subset of the unit ball of B. Without loss of generality we
can assume a “fast convergence” condition that for each N = 2, 3, . . . we have
kϕN (bj ) − ϕN −1 (bj )k < 2−N ∀j < N.
We’ll modify the lifts ψn inductively to new ucp lifts ψn0 satisfying
0 0 −N
kψN (bj ) − ψN −1 (bj )k < 2 · 2 ∀j < N.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
123
This condition (together with the density of the bj and the fact that ucp maps
have norm 1) implies that the sequence ψn0 converges in the point-norm topology,
as required.
Suppose that the lifts ψ10 , . . . , ψN
0 0
have been constructed. Our formula for ψN +1 is
1 1 1 1
0 2 0
ψN +1 (b) = (1 − u) ψN +1 (b)(1 − u) + u ψN (b)u ,
2 2 2
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
124
Lecture 30
Nuclearity and Group Algebras
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
125
series converges and defines an element of Cr∗ (G) (indeed of C[G]). This is a com-
pletely positive map.
Calculation shows that the composite of these three maps is the multiplier Mpn .
Since pn tends to 1 pointwise, these multipliers tend to the identity operator in the
point-norm topology. Thus Cr∗ (G) is nuclear.
Proposition 30.5. Suppose Cr∗ (G) is nuclear. Then G is an amenable group.
Proof. By definition of nuclearity, there are nets of ucp maps
ϕα : Cr∗ (G) → Mnα (C), ψα : Mnα (C) → Cr∗ (G)
whose composites tend point-normwise to the identity map. Using Lemma 28.11,
extend the maps ϕλ to all of B(`2 (G)). We obtain composite maps αλ : B(`2 (Γ)) →
Cr∗ (Γ) ⊆ L(Γ) (the group von Neumann algebra). Now use the compactness of the
unit ball of B(`2 (Γ)) in the weak topology to pass to a point-weak limit
α : B(`2 (Γ)) → L(Γ)
which restricts to the identity on Cr∗ (Γ).
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
126
the first equality because of the bimodule property (Remark 28.7) and the second
because τ is a trace. Applied to `∞ (G) this shows that m is an invariant mean.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
127
Lecture 31
Approximate Equivalence and the Weyl-von Neumann Theorem
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
128
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
129
Corollary 31.9. The essential spectrum of a diagonal operator is the set of limit
points of its eigenvalue sequence.
Lemma 31.10. Let X be a metric space. Suppose that (xn ) and (yn ) are two
sequences in X that have the same limit points. Then there is a bijection α : N → N
such that d(xn , yα(n) ) → 0 as n → ∞.
Proof. An exercise, using induction.
Proposition 31.11. Two diagonal operators that have the same essential spectrum
are essentially unitarily equivalent.
Proof. Let D and D0 be diagonal operators, with corresponding eigenvalue sequences
(λn ), (λ0n ) and eigenbases (en ), En0 ). Using lemma 31.10, choose a permutation α of
N such that
|λn − λ0α(n) | → 0
as n → ∞. Then the unitary U defined by U en = e0α(n) implements an approximate
equivalence between D and D0 .
Definition 31.12. An extension τ : A → Q(H) splits if there is a commutative
diagram of ∗-homomorphisms
B(H)
=
A
τ / Q(H)
We are going to prove that split extensions are unique up to equivalence. A
consequence (in the single operator case) will be that an extension is split if and
only if it comes from an operator of the form normal plus compact.
Proposition 31.13. Let X be a compact metric space. Then any two split exten-
sions of C(X) are equivalent.
Proof. We’ll define a special class of split extensions called diagonal extensions. To
construct such an extension, let (xn ) be a sequence in X which is dense and has every
point of X as a limit point. Let H = `2 and let ϕ(f ) be the diagonal operator on
H with diagonal elements (f (xn )). Then ϕ : C(X) → B(H) is a ∗-homomorphism
and τ = π ◦ ϕ is injective, defining a split extension.
Given two sequences (xn ) and (x0n ), we can always construct by Lemma 31.10 a
bijection α : N → N such that d(xn , x0α(n) ) → 0. The unitary U on `2 defined by
α then implements an equivalence between the corresponding extensions τ and τ 0 .
Thus all diagonal extensions are equivalent and our problem is to show that every
split extension is equivalent to a diagonal one.
Suppose that Y → X is a surjective map of compact metric spaces. Then the
induced map C(X) → C(Y ) is injective so every extension of C(Y ) gives rise to an
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
130
C(X)
τ / Q(H)
As we remarked above, it suffices to show that the extension π ◦ ψ of C(Y ) is
diagonal. But C(Y ) is generated by countably many projections, whence Y is a zero-
dimensional compact metrizable space. By a standard result of general topology,
such a space is homeomorphic to a compact subset of R. Let g : Y → R be such a
homeomorphism. Then T = ψ(g) is a self-adjoint operator which generates ψ(C(Y ))
as a C ∗ -algebra. By the Weyl-von Neumann theorem, T is a compact perturbation of
a diagonal operator T 0 whose essential spectrum is g(Y ) ∼
= Y . Thus the extension
π ◦ ψ is equivalent to one generated by a diagonal operator (and moreover, by
Proposition 31.11, we may assume without loss of generality that the eigenvalue
sequence of that operator belongs to Y ). Such an extension is, by construction,
diagonal.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
131
Lecture 32
Brown-Douglas-Fillmore Theory
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
132
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
133
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
134
Lecture 33
More about BDF Theory
I want to relate the Ext theory that we have been learning about to the Fredholm
modules that showed up in our discussion of Cr∗ (F2 ) (Definition 27.1). Let A be a
separable nuclear C ∗ -algebra and τ an extension of A. According to Theorem 32.4,
there is a Hilbert space H = H 0 ⊕ H 00 and a ∗-homomorphism ϕ : A → B(H) with
the property that in the 2 × 2 matrix decomposition
ϕ11 ϕ12
ϕ= ,
ϕ21 ϕ22
we have τ = π ◦ ϕ11 . Moreover, we saw in the proof of Theorem 32.4 that the
off-diagonal terms ϕ12 and ϕ21 take values in the compact operators. Equivalently,
if we define
1 0
F = ,
0 −1
then F 2 = 1, F = F ∗ , and F commutes modulo compacts with ϕ(a) for all a ∈ A.
These are exactly the criteria that we used to define a Fredholm module in Defini-
tion 27.1. The only difference is that in that section, we demanded an additional
piece of structure, namely a grading on our Hilbert space with respect to which
F is odd and ϕ is even. Let us agree to call a Fredholm module equipped with a
grading (as in 27.1) an even Fredholm module and one without this structure an
odd Fredholm module. Then we have shown that every extension of a separable
nuclear C ∗ -algebra gives rise to an odd Fredholm module.
Conversely, suppose (H, F, ϕ) is an odd Fredholm module over an algebra A. Then
P = 12 (1 + F ) is a projection; let H 0 be the range of P . For each a ∈ A we can define
a “Toeplitz” operator Ta = P ϕ(a)P on H 0 , and the map a 7→ Ta is a homomorphism
modulo compacts and therefore gives rise to an extension of A. Thus extensions and
odd Fredholm modules are equivalent (for nuclear C ∗ -algebras).
We saw in Lecture 20 that Toeplitz operators with invertible symbol are Fredholm
and therefore have an index. The corresponding notion in this general case makes
use of K-theory.
Let A be a unital C ∗ -algebra. The group U (A) of unitary elements of A is a
topological group (in the norm topology) and therefore π0 U (A), the collection of
connected components of U (A), is a discrete group. Moreover for any n, Mn (A) is
also a C ∗ -algebra and there is a natural embedding of U (Mn (A)) in U (Mn+1 (A))
(“direct sum with 1”).
Definition 33.1. For a unital C ∗ -algebra A, the K1 group K1 (A) is the (abelian)
group defined as a direct limit
K1 (A) = lim π0 (Mn (A)).
n→∞
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
135
A1 / B
there is an exact sequence relating the K-theory groups of C, A0 and A1 , and
B.
The key to BDF’s results is that Ext has the same properties. In contrast to K-
theory, where the homotopy invariance (for example) is a more or less trivial conse-
quence of the definitions, the homotopy invariance of Ext is a major theorem.
Remark 33.3. As well as the “odd” pairing defined in Proposition 33.2 there is also
an “even” pairing between graded Fredholm modules and the even K-group K0 .
We have, in fact, already made use of this pairing in a special case, lemma 27.2.
That lemma shows how to give an integer-valued pairing between even Fredholm
modules and projections in the algebra A. But it can easily be extended to give a
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
136
pairing between such modules and projections in matrix algebras Mn (A), and such
projections generate the even K-group K0 (A).
To finish our brief discussion of Ext I will give an example of a C ∗ -algebra B
(necessarily non-nuclear) for which Ext(B) is not a group. Let Γ be the discrete
group SL(3, Z). The key fact from representation theory that we need to use is that
Γ has Kazhdan’s Property T.
Definition 33.4. A finitely generated discrete group Γ has property T if the fol-
lowing is true: for each finite generating set S (or for just one such set) there is a
constant εS > 0 such that, if π : Γ → U (H) is a unitary representation and there is
a unit vector ξ ∈ H which is “almost fixed” by Γ, in the sense that
kξ − π(g)ξk < εs ∀g ∈ S,
then there is also a unit vector ξ 0 ∈ H that is fixed (that is, π(g)ξ 0 = ξ 0 for all
g ∈ Γ).
Example 33.5. Finite groups have property T.
Example 33.6. No infinite amenable group has property T: the normalized char-
acteristic functions of a Folner sequence, thought of as vectors in the regular rep-
resentation, become more and more “almost fixed”, but the regular representation
has no fixed vectors (because the group is infinite).
It is a deep fact that any infinite property T groups exist; but they do, and
SL(3, Z) is an example.
Remark 33.7. A fixed vector in a representation of Γ is the same thing as a subrep-
resentation which is a copy of the trivial representation. Therefore, the definition
of property T can be rephrased: if a unitary representation of Γ “approximately”
contains a copy of the trivial representation, then it actually does contain a copy of
the trivial representation. When formulated this way, Property T has an extension
to any finite-dimensional irreducible representation. Let π : Γ → U (Hπ ) be such a
representation and let ρ : Γ → U (Hρ ) be any unitary representation. We say that ρ
“approximately contains” π if there is an isometry V : Hπ → Hρ such that
kπ(g) − V ∗ ρ(g)V k < ηS ∀g ∈ S,
where S is a generating set as before. Then, for a property T group, there is an
ηS such that this approximate containment implies that ρ actually does contain a
subrepresentation isomorphic to π. To prove this, apply the definition of property T
to the space of bounded operators Hπ → Hρ ; because Hπ is finite-dimensional, this
is actually a Hilbert space in the Hilbert-Schmidt norm, and its invariant elements
are the intertwining operators.
Remark 33.8. It follows from the previous remark that for each finite-dimensional
irreducible representation π of a property T group, there is a nonzero projection
pπ ∈ C ∗ (Γ) (the maximal group C ∗ -algebra) with the following property: for any
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
137
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
138
Lecture 34
UHF Algebras
In this section we shall study C ∗ -algebras (and associated von Neumann algebras)
that are constructed as unital inductive limits of matrix algebras. An important
example is the algebra An ⊆ On which was used in our earlier discussion of Cuntz
algebras.
Lemma 34.1. Let Mm (C) and Mn (C) be matrix algebras. There exists a unital
∗-homomorphism Mm (C) → Mn (C) if and only if m|n; and, when this condition is
satisfied, all such unital ∗-homomorphisms are unitarily equivalent to that given by
T
T 7→ ..
.
T
with T repeated n/m times down the diagonal and zeroes elsewhere.
Proof. This is a special case of Lemma 9.5.
Suppose now that k1 |k2 |k3 | · · · is an increasing sequence of natural numbers, each
of which divides the next. By the lemma, there is associated to this sequence
a unique (up to unitary equivalence) sequence of matrix algebras and unital ∗-
homomorphisms
α1 α2
Mk1 (C) / Mk2 (C) / Mk3 (C) / ··· .
Let A denote the (algebraic) inductive limit of this sequence. (Reminder about
what this means: Think of unions. More formally, the elements of A are equivalence
classes of sequences {Ti }, Ti ∈ Mki (C), which are required to satisfy αi (Ti ) = Ti+1
for all but finitely many i, and where two sequences are considered to be equivalent
if they differ only in finitely many places. These may be added, multiplied, normed
(remember that the αi are isometric inclusions!), and so on, by pointwise operations.)
The algebra A is a pre-C ∗ -algebra; that is, a normed algebra which satisfies all the
C ∗ -axioms except that it need not be complete. Its completion is a C ∗ -algebra which
is called the uniformly hyperfinite (UHF) algebra (or Glimm algebra) determined by
the inductive system of matrix algebras.
Proposition 34.2. All UHF algebras are simple, separable, unital, and have a
unique trace.
Proof. Obviously they are separable and unital. Let A be a UHF algebra and let
α : A → B be a unital ∗-homomorphism to another C ∗ -algebra. Since each ma-
trix subalgebra Mki (C) of A is simple, the restriction of α to it is an injective
∗-homomorphism and is therefore isometric. Thus α is isometric on the dense sub-
algebra A of A, whence by continuity α is isometric on the whole of A and in
particular has no kernel. Thus, A is simple.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
139
For each matrix algebra Mki (C) let τki : Mki (C) → C denote the normalized trace,
that is
1
τki (T ) = Tr(T ).
ki
The τki then are consistent with the embeddings αi so they pass to the inductive limit
to give a trace τ on the dense subalgebra A of A. Moreover, τ is norm continuous on
A (since the τki are all of norm one on the matrix algebras Mki (C)) and so extends
by continuity to a trace, indeed a tracial state, on A. The uniqueness of the trace
on matrix algebras12 shows that every other trace must agree, up to normalization
of course, with this one.
Lemma 34.3. Any nonzero tracial state on a simple C ∗ -algebra is faithful (Re-
mark 26.1). In particular, the unique trace on a Glimm algebra is faithful.
Proof. Let τ be a nonzero tracial state on A and let N = {a ∈ A : τ (a∗ a) = 0}.
Since τ is a state, N is a left ideal in A (see the GNS construction, Theorem 15.1).
But N = N ∗ by the trace property, so N is in fact a two-sided ideal, which must
vanish by simplicity. Hence τ is faithful.
Every UHF algebra contains many projections (because the matrix subalgebras
contain many projections). We are going to study the structure of UHF algebras by
looking at the projections that they contain. We need a couple of general lemmas
first.
Lemma 34.4. Let p and q be projections in a unital C ∗ -algebra A, and suppose that
kp − qk < 1. Then τ (p) = τ (q) for any trace τ on A. In fact, p and q are unitarily
equivalent.
Proof. Define x = qp + (1 − q)(1 − p). Then xp = qx, and a simple calculation shows
that
x − 1 = 2qp − q − p = (2q − 1)(p − q).
1
Thus kx − 1k < 1 and so x is invertible. We may define a unitary u = x(x∗ x)− 2 by
the functional calculus. Now x∗ x commutes with p, and so
1 1
up = xp(x∗ x)− 2 = qx(x∗ x)− 2 = qu
as required.
Lemma 34.5. Let A be a unital C ∗ -algebra and let x ∈ A be a self-adjoint element
which is approximately a projection in the sense that kx2 − xk < 14 . Then there is a
projection p ∈ A such that kx − pk < 12 .
Proof. From the functional calculus, the inequality kx2 − xk < 14 implies that x has
spectrum in (− 21 , 12 ) ∪ ( 12 , 32 ). Let f be the function which equals 0 on (− 21 , 12 ) and
1 on ( 12 , 32 ). Then f is continuous on the spectrum of x, f (x) is a projection, and
kf (x) − xk < 21 .
12And how do you prove that? Compute with matrix units to show that every trace-zero finite
matrix lies in the span of commutators.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
140
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
141
Proposition 34.11. If two UHF algebras are isomorphic, then any two matrix
algebra sequences generating them have the same supernatural number.
Proof. First, let us reformulate the idea that two matrix algebra sequences have the
same supernatural number. If the two sequences have matrix orders k1 |k2 |k3 | · · ·
and l1 |l2 |l3 | · · · respectively, to say that the associated supernatural numbers are
the same is just to say that for each i there is a j such that ki divides lj , and that
for each i there is a j such that li divides kj .
Let us prove proposition 34.10. If the two sequences above have the same super-
natural number, then by passing to subsequences (which doesn’t change the direct
limit) we can arrange that ki |li |ki+1 . Then consider embeddings of matrix algebras
as in the diagram below
/ / /
Mk1 (C) Mk2 (C)
9
Mk3 (C)
9 ;···
Ml1 (C) / Ml2 (C) / Ml3 (C) / ···
We obtain a ∗-isomorphism between dense subalgebras of A and B, which extends
automatically to a ∗-isomorphism from A to B.
Now let us prove proposition 34.11. Let the two algebras A and B be obtained
from generating sequences k1 |k2 |k3 | · · · and l1 |l2 |l3 | · · · respectively. Let α : A → B
be an isomorphism. Notice that α takes the unique trace τA on A to the unique
trace τB on B.
We will show that for each i there is a j such that ki divides lj . Indeed, let p
denote the matrix unit e11 in Mki (C) ⊆ A. Then α(p) is a projection in B. By
density there is x = x∗ ∈ Mlj (C) ⊆ B such that kx − α(p)k < 20 1
. By lemma 34.5
there is a projection q ∈ Mlj (C) such that kq − α(p)k < 1. By the lemma 34.4 and
the uniqueness of the trace
1
τB (q) = τB (α(p)) = τA (p) = .
ki
Since the normalized trace of every projection in Mlj (C) is a multiple of 1/lj , we
conclude that ki divides lj , as required.
Notice as a corollary that there are uncountably many isomorphism classes of
UHF algebras.
Example 34.12. The Fermion algebra (CAR algebra) arises in quantum field the-
ory. The second quantization of a fermion field leads to the following problem: given
a Hilbert space H, find a C ∗ -algebra A generated by operators cv , v ∈ H, such that
v 7→ cv is linear and isometric and
cv cw + cw cv = 0, c∗v cw + cw c∗v = hw, vi1.
These are called the canonical anticommutation relations. Compare the definition
of a Clifford algebra. We will show that the C ∗ -algebra A arising in this way is the
unique UHF algebra with supernatural number 2∞ .
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
142
Let v be a unit vector on H. Then cv has square zero, and c∗v cv + cv c∗v = 1.
Multiplying through by c∗v cv we find that c∗v cv is a projection, call it ev , and that cv
is a partial isometry with range projection ev and domain projection 1 − ev . We can
therefore identify the C ∗ -subalgebra of A generated by cv : it is a copy of M2 (C),
generated by matrix units e21 = cv , e11 = ev , e22 = 1 − ev , e12 = c∗v .
The idea now is to apply the process inductively to the members v1 , v2 , . . . of
an orthonormal basis for the Hilbert space H. Let A1 , A2 , . . . be the subalgebras
of A (all isomorphic to M2 (C)) generated by cv1 , cv2 , . . .. If the algebras Ai , Aj all
commuted with one another for i 6= j then we would be done, because the subalgebra
spanned by A1 , . . . , An would then be the tensor product M2 (C) ⊗ · · · ⊗ M2 (C) =
M2n (C) and we would have represented A as an inductive limit of matrix algebras.
Unfortunately the Ai ’s do not commute; but they do ‘supercommute’ when they are
considered as superalgebras in an appropriate way. Rather than developing a large
machinery for discussing superalgebras, we will inductively modify the algebras Aj
to algebras Bj which are each individually isomorphic to 2 × 2 matrix algebras,
which do commute, and which have the property that the subalgebra of A spanned
by A1 , . . . , An is equal to the subalgebra of A spanned by B1 , . . . , Bn . This, then,
will complete the proof.
Let γi be the grading operator for Ai , which is to say the operator corresponding
to the 2 × 2 matrix ( 10 −10
). Notice that γi is self-adjoint, has square 1 and commutes
with cvj for j 6= i, while it anticommutes with cvi ; in particular γi commutes with
γj . Let
B1 = A1 = C ∗ (cv1 ), B2 = C ∗ (cv2 γ1 ), . . . Bn = C ∗ (cvn γ1 · · · γn−1 ).
By the same argument as for the Ai above, each Bi is isomorphic to M2 (C). Moreover
the generators of the algebras Bi all commute. Finally, it is clear that the subalgebra
spanned by B1 , . . . , Bn is the same as the subalgebra spanned by A1 , . . . , An . The
proof is therefore complete.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
143
Lecture 35
UHF Algebras and von Neumann Factors
We are going to use UHF algebras to construct interesting examples of von Neu-
mann algebras. (This is something of a reversal of history: Glimm’s UHF-algebra
construction is explicitly patterned after a construction of Murray and von Neu-
mann, that of the so-called hyperfinite II1 factor.) Let H be a Hilbert space.
Recall that a von Neumann algebra of operators on H is a unital, weakly closed
∗-subalgebra M ⊆ B(H). Its commutant
M 0 = {T ∈ B(H) : T S = ST ∀ S ∈ M }
is also a von Neumann algebra.
Definition 35.1. If the center
Z(M ) = M ∩ M 0
consists only of scalar multiples of the identity, then the von Neumann alegbra M
is called a factor.
There is a sophisticated version of the spectral theorem which expresses every von
Neumann algebra as a ‘direct integral’ of factors.
We are going to manufacture examples of factors by the following procedure: take
a C ∗ -algebra A (in our examples, a UHF algebra) and a state σ of A. By way of the
GNS construction, form a representation ρσ : A → B(Hσ ) and let Mσ be the von
Neumann algebra generated by the representation, that is Mσ = ρσ [A]00 . If Mσ is a
factor, then we call σ a factorial state. As we shall see, many different examples of
factors can be generated from the factorial states on one and the same C ∗ -algebra.
By way of a warm-up let us see that UHF C ∗ -algebras have trivial center.
Lemma 35.2. Let A be a UHF algebra; then Z(A) = C1.
Proof. The proof applies to any unital C ∗ -algebra with a unique faithful trace, τ .
Let z ∈ Z(A). Then the linear functional a 7→ τ (az) is a trace on A, so by uniqueness
it is equal to a scalar multiple of τ , say λτ . But then
τ (a(z − λ1)) = 0
for all a ∈ A, in particular for a = (z − λ1)∗ , so by faithfulness z − λ1 = 0 and z is
a multiple of the identity.
Remark 35.3. Of course this does not prove that various von Neumann algebra
completions of A must have trivial center. We need more work to see that! Moreover,
the argument above does not give much insight into how the von Neumann algebra
proof is going to go. Consider the following alternative argument: let z ∈ Z(A).
Then for any ε > 0 there is a matrix subalgebra Mn (C) ⊆ A and an x ∈ Mn (C)
with kx − zk < ε. Let Z
x̂ = u∗ xu dλ(u),
Un (C)
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
144
where λ is Haar measure on the compact group Un (C). Then x̂ is a multiple of the
identity, but kx̂ − zk < ε. Since ε is arbitrary, z is a multiple of the identity. This
proof is more like the von Neumann algebra argument we will give in Proposition 36.1
below.
We will discuss a special class of states on UHF algebras, the product states.
Recall that a UHF algebra A is a completed inductive limit of matrix algebras
Mk1 (C) ⊆ Mk2 (C) ⊆ Mk3 (C) ⊆ · · · .
Putting d1 = k1 , di+1 = ki+1 /ki for i > 1, we can write the inductive sequence
equivalently as
Md1 (C) ⊆ Md1 (C) ⊗ Md2 (C) ⊆ Md1 (C) ⊗ Md2 (C) ⊗ Md3 (C) ⊆ · · · .
Notice that Mdi+1 (C) is the commutant of Mki (C) inside Mki+1 (C). For each i, let σi
be a state of Mdi (C). (We will discuss the explicit form of these states in a moment.)
Then one can easily define a state σ on the algebraic inductive limit A by
σ(a1 ⊗ a2 ⊗ · · · ) = σ1 (a1 )σ2 (a2 ) · · · .
Note that all but finitely many of the ai are equal to one. When each of the σi
are the normalized trace, this is exactly the construction we performed in the last
lecture to obtain τ . The functional σ has norm one, so it extends by continuity to
a state on the UHF algebra A. Such a state is called a product state.
Lemma 35.4. Let A be a UHF algebra and let σ be a product state on it. Let
M = Mki (C) be one of the matrix algebras in an inductive sequence definining
A. If x ∈ M and y ∈ M 0 (by which I mean the commutant of M in A), then
σ(xy) = σ(x)σ(y).
Proof. The result is apparent if y ∈ Mkj (C) for some j > i. For then the commuta-
tion condition allows us to write x = X ⊗ 1, y = 1 ⊗ Y relative to the tensor product
decomposition of the matrix algebra, and the result follows from the definition of
a product state. The proof is completed by an approximation argument, similar to
Remark 35.3 above.
Let y ∈ A commute with Mki (C) and let ε > 0. Then there is a y 0 ∈ Mkj (C) (for
some big j) with ky 0 − yk < ε. Now y 0 probably does not commute with Mki (C)
anymore, but
Z
ŷ = u∗ y 0 u dλ(u)
U (ki )
does, and it lies within ε of y. Moreover ŷ ∈ Mkj (C) and therefore σ(xŷ) = σ(x)σ(ŷ).
Thus
|σ(xy) − σ(x)σ(y)| < 2ε|σ(x)|.
Since ε is arbitrary this gives us the result.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
145
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
146
unitary u ∈ Mkn (C) such that σq (a) = σq0 (u∗ au) for all a ∈ Mkn (C). This equality
extends to all a ∈ A by construction, and then it extends to A by continuity. So
the associated irreducible representations of A are unitarily equivalent.
Conversely, suppose that σq (a) = σq0 (u∗ au) for some unitary u. Find an element
x ∈ Mkn (C) (n sufficiently large) with ku − xk < 21 . Suppose for a contradiction
that qj = 1, qj0 = 2 for some j > n, and let e ∈ Mkj (C) be of the form 1 ⊗ e11 ,
where e11 is the matrix unit in Mdj (C). Now σ 0 (e) = 0, so σ 0 (x∗ ex) = σ 0 (x∗ xe) =
σ 0 (x∗ x)σ 0 (e) = 0 using the multiplicative property of product states and the fact
that e commutes with x. Now write
1 = σ(e) = σ 0 (u∗ eu) 6 |σ 0 ((u∗ − x∗ )eu)| + |σ 0 (x∗ e(u − x))| + |σ 0 (x∗ ex)| < 1
to obtain a contradiction.
Corollary 35.8. Let A be a UHF algebra. Then the unitary dual A
b is uncountable,
but has the indiscrete topology. Moreover, A is antiliminal.
Proof. Proposition 35.7 shows that a UHF algebra A has uncountably many in-
equivalent unitary representations; but since it is simple the primitive ideal space is
trivial, and hence A
b has the indiscrete topology by Theorem 18.1.
Let us show that A is not postliminal (see Remark 19.9 for the definitions here).
Consider an irreducible representation ρ of A. If A is postliminal, then ρ(A) con-
tains an abelian element of norm one, that is, a rank one projection. But then it
contains all the compact operators (Corollary 19.5). Simplicity implies A ∼ = K, a
contradiction.
Since A is not postliminal, its maximal postliminal ideal I is not equal to A.
Remark 19.9 tells us that A/I is antiliminal. But by simplicity I = 0 so A is
antiliminal as required.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
147
Lecture 36
The Classification of Factors
Irreducible representations are (of course!) factorial. However UHF algebras also
possess factorial representations that are not irreducible, In fact we have
Proposition 36.1. Each product state of a UHF algebra is factorial.
Proof. Let σ be a product state of the UHF algebra A and let (ρσ , Hσ , uσ ) be a
cyclic representation associated with σ via the GNS construction. Suppose that
z ∈ Z(ρσ [A]00 ). By the Kaplansky Density Theorem 14.2, there is a sequence yj in
A with ρ(yj ) → z weakly and kyj k 6 kzk for all j. Fix n, and let zj be obtained
from yj by averaging over the unitary group of Mkn (C) ⊆ A, as in the proof of
Lemma 35.4. For each fixed such unitary u we have
ρ(uyj u∗ ) → ρ(u)zρ(u∗ ) = z
in the weak topology, boundedly in norm; so by applying the Dominated Conver-
gence Theorem we see that ρ(zj ) → z weakly. Since zj commutes with Mkn (C) we
have for x, y ∈ Mkn (C),
hz[x], [y]i = limh[x], ρ(zj )[y]i =
lim σ(x∗ zj y) = σ(x∗ y)σ(z) = h[x], [y]iσ(z)
using Lemma 35.4. Since this holds for all n, it follows that z = σ(z)1 is a scalar
multiple of the identity. Thus ρσ (A)00 is a factor.
Many different interesting factors can be obtained from this construction. But
how are these factors to be distinguished? In this section we will briefly outline the
classification of factors that was developed by Murray and von Neumann.
Definition 36.2. Let M be a von Neumann algebra. A weight on M is a function
ϕ : M+ → [0, ∞] (notice that the value ∞ is permitted) which is positive-linear in
the sense that
ϕ(α1 x1 + α2 x2 ) = α1 ϕ(x1 ) + α2 ϕ(x2 )
for α1 , α2 ∈ R+ and x1 , x2 ∈ M+ .
The canonical example is the usual trace on B(H). We say that ϕ is semi-finite
if the domain
M+ϕ = {x ∈ M+ : ϕ(x) < ∞}
is weakly dense in M+ , and that ϕ is normal if
lim ϕ(xλ ) = ϕ(lim xλ )
whenever xλ is a monotone increasing net in M that is bounded above (and hence
weakly convergent). If ϕ(u∗ xu) = ϕ(x) for all x ∈ M+ and unitary u we say that ϕ
is a tracial weight, or sometimes just a trace (so long as we remember that it does
not have to be bounded).
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
148
Lemma 36.3. For each weight ϕ on M , the linear span M ϕ of M+ϕ is a hereditary
∗-subalgebra to which ϕ extends as a positive linear functional. If ϕ is tracial, this
extension of ϕ has the trace property ϕ(xy) = ϕ(yx).
Definition 36.4. A von Neumann algebra M is said to be finite if it has a faithful
normal finite trace (i.e. a tracial state). It is semifinite if it has a faithful normal
semifinite trace.
Recall the definition of finiteness for projections (Remark 34.7).
Theorem 36.5. (Murray-von Neumann) A factor M is finite if and only if every
projection in M is finite. Any finite or semifinite factor contains finite projections.
Murray and von Neumann introduced the idea of studying factors by looking
at the comparison properties of projections. Let p and q be projections in a von
Neumann algebra M . One says that p is weaker that q (or that q is stronger than p)
if p is equivalent to a subprojection of q: that is, if there is a partial isometry v ∈ M
with v ∗ v = p and vv ∗ 6 q. We write p - q in this case. (Thus, an infinite projection
is one that is strictly weaker than itself.) Murray and von Neumann showed that, if
M is a factor, then the projections form a totally ordered set under the relation -.
This led to the classification of factors:
Finite faithful Semifinite No faithful
normal trace faithful nor- normal trace
mal trace
Minimal projec- Type In Type I∞ Impossible
tion
No minimal pro- Type II1 Type II∞ Type III
jection
It is easy to give examples of factors of type I. In fact, the matrix algebras Mn (C)
are factors of type In , and the bounded operators B(H) form a factor of type I∞ .
These can be shown to be the only examples.
Lemma 36.6. The completion of a UHF algebra relative to the canonical trace is a
type II1 factor.
Proof. Let’s check the definition.
The canonical trace is a product state, so we have obtained a factor R by Propo-
sition 36.1.
The canonical trace extends to a linear functional on R which is weakly continuous
(by construction!) and hence normal. Weak continuity also shows that the extended
functional is still a trace.
The extended trace is still faithful: Suppose that T ∈ R with τ (T ∗ T ) = 0. Then
(letting ξ denote the cyclic vector), for any unitary u ∈ A
0 = τ (ρ(u∗ )T ∗ T ρ(u)) = hξ, ρ(u∗ )T ∗ T ρ(u)ξi = kT ρ(u)ξk2 ,
so T ρ(u)ξ = 0. But as u runs over the unitaries of A, the ρ(u)ξ span H, so T = 0.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
149
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
150
Notice that ϕα (u∗n xun ) = ϕα (x) for all x. Notice also that if x, y ∈ M2m (C) ⊆ A
then u∗n xun commutes with y for all n > m. Now suppose that x, y, z ∈ M2k (C) ⊆ A.
Then
ϕα (z ∗ u∗n xun y) = ϕα (u∗n xun z ∗ y) = ϕα (u∗n xun )ϕα (z ∗ y) = ϕα (x)ϕα (z ∗ y)
using the multiplicative property of product states (Lemma 35.4). In terms of the
GNS representation this means
h[z], πα (u∗n xun )[y]i → ϕα (x)h[z], [y]i.
But the vectors [z], [y] span a dense subspace of the representation space, so this
shows πα (u∗n xun ) → ϕα (x)1 weakly as required.
It follows immediately that Mα has no finite normal trace (so it is not a factor of
type II1 ). For, supposing for a moment that τ was such a trace, we would have
τ (u∗n xun ) = τ (x) ∀n.
A normal functional is weakly continuous on bounded sets (dominated convergence
theorem) and so we would have
τ (x) = τ ( lim u∗n xun ) = ϕα (x)τ (1).
n→∞
But ϕα is not a trace so this is ridiculous.
To extend this argument to show there is no semifinite trace either (i.e. Mα is
not a factor of type II∞ ) requires some techniques for working with (unbounded)
normal functionals. I won’t get started on that here, but simply refer to Theorem
6.5.15 in Pedersen.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
151
Lecture 37
Weak containment: loose ends
The objective of these notes is to tidy up a few loose ends that were left dangling
regarding
• the Fell topology on the spectrum of a C ∗ -algebra (definition 17.4);
• the definitions of amenability and property T;
• the notion of “weak containment” of representations, defined below;
• and the proof of the trace property Tr(AB) = Tr(BA) under minimal as-
sumptions, see Lecture 2.
Throughout this section Γ denotes a discrete, finitely generated group. (The results
extend to the locally compact case.) We will consider unitary representations of
Γ, which are the same thing as unitary representations of the maximal group C ∗ -
algebra C ∗ (Γ). If π is such a representation, a (normalized) positive definite function
associated to that representation is a function on Γ of the form
γ 7→ hξ, π(γ)ξi
for some unit vector ξ ∈ Hπ . In other words, a positive definite function associated
to π corresponds to a vector state of C ∗ (Γ) coming from that representation. (Via
the GNS construction, one can go back from the positive definite function to an
associated representation, though what you get may not be the whole of the original
π; you will get its restriction to the cyclic subspace generated by ξ.)
Let P (π) denote the collection of normalized positive definite functions associated
to π.
Definition 37.1. Let π and ρ be representations of Γ. Then π is weakly contained
in ρ (in symbols: π ρ) if every member of P (π) is a pointwise limit of a sequence
of convex combinations of members of P (ρ).
Remark 37.2. Weak containment is clearly a transitive relation, and containment
(π is a direct summand in ρ) implies weak containment. On the other hand, weak
containment takes no account of multiplicities: if π ρ then ⊕n π ρ also. More
generally, if a family of representations are each individually contained in ρ, so is
their sum.
Remark 37.3. A simple normalization argument shows that it is equivalent to de-
mand that every positive definite function associated to π (normalized or not) is in
the pointwise closed subspace of `∞ (Γ) spanned by P (ρ). We will use this below.
Lemma 37.4. Suppose that π has a cyclic vector. Then, in order to check that π
ρ, it suffices to prove that the single positive definite function ϕ(γ) = γ 7→ hξ, π(γ)ξi
is in the closed linear span of P (ρ).
Proof. Consider a vector η of the form
X n
η= λi π(γi )ξ,
i=1
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
152
Call the right side of this expression A(ϕ). Then A is an operation which maps the
span of P (ρ) to itself. Consequently, all the positive definite functions γ 7→ hη, π(γ)ηi
belong to the pointwise-closed linear span of P (ρ). But the collection of η of this
sort is dense in H, since ξ is a cyclic vector.
Lemma 37.5. Suppose that π is an irreducible representation. Then π ρ if and
only if each member of P (π) can be pointwise approximated by members of P (ρ) (it
is not necessary to take linear or convex combinations).
Proof. We work in the state space S = S(C ∗ (Γ)). Let σ be a vector state associated
to π. Let X denote the collection of vector states associated to ρ (i.e., the states
corresponding to the members of P (ρ)) and let Y be its closed convex hull. The
weak containment relation π ρ says exactly that σ ∈ Y . But σ is a pure state,
that is an extreme point of S, and in particular therefore it is an extreme point of Y .
It follows from Milman’s theorem 11.19 that σ belongs to the closure of X, which
is what was required.
The unitary dual of Γ is the collection of equivalence classes of irreducible unitary
representations of Γ, i.e., the spectrum of C ∗ (Γ). It is given the Fell topology.
Definition 37.6. Let ρ be any representation of Γ. The support of ρ is the subset
of Γ
b consisting of those irreducible representations weakly contained in ρ. It is a
closed subset of Γ.
b
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
153
Conversely suppose that Γ does not have property T. Then (fixing the generating
set S) for each ε = 1/n, say, there is L
a representation ρn that contains an ε-fixed
vector but no fixed vector. Form ρ = n ρn . Then π0 ρ but ρ contains no fixed
vector.
Proposition 37.9. A group Γ has property T if and only if the point representing
the trivial representation is isolated (that is, the one-point set containing it is open)
in the unitary dual Γ.
b
(We moved the extra Pn ’s into the middle because they commute with A, and
round to the far end using the trace property.) Now consider Pn APn and Pn BPn as
operators on the Hilbert space Hn which is the range of Pn . Their composite equals
Pn ABPn so is of trace class. But Pn APn is invertible on Hn by construction. Thus
Pn BPn is trace class on Hn . Consequently
Tr(Pn APn · Pn BPn ) = Tr(Pn BPn · Pn APn )
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25
154
by the easy case of the result. Unwinding the limits as before completes the proof
for self-adjoint A.
Now for the general case, we use polar decomposition to write A = V P with P
positive and V a partial isometry. Then P B = V ∗ AB is of trace class. Consequently
Tr(AB) = Tr(V P B) = Tr(P BV ) = Tr(BV P ) = Tr(BA)
where the first step (commuting V) uses the ”easy case” and the second step (com-
muting P) uses the self-adjoint case proved above.
AMS Open Math Notes: Works in Progress; Reference # OMN:201701.110660; 2017-01-01 20:43:25