The Mathematics Any Physicist Should Know

The Mathematics any Physicist
Should Know
Thomas Hjortgaard Danielsen
Contents
Preface 5
I Representation Theory of Groups and Lie Algebras 7
1 Peter-Weyl Theory 9
1.1 Foundations of Representation Theory . . . . . . . . . . . . . . . 9
1.2 The Haar Integral . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.3 Matrix Coecients . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.4 Characters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.5 The Peter-Weyl Theorem . . . . . . . . . . . . . . . . . . . . . . 24
2 Structure Theory for Lie Algebras 29
2.1 Basic Notions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.2 Semisimple Lie Algebras . . . . . . . . . . . . . . . . . . . . . . . 35
2.3 The Universal Enveloping Algebra . . . . . . . . . . . . . . . . . 42
3 Basic Representation Theory of Lie Algebras 49
3.1 Lie Groups and Lie Algebras . . . . . . . . . . . . . . . . . . . . 49
3.2 Weyls Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4 Root Systems 59
4.1 Weights and Roots . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.2 Root Systems for Semisimple Lie Algebras . . . . . . . . . . . . . 62
4.3 Abstract Root Systems . . . . . . . . . . . . . . . . . . . . . . . . 68
4.4 The Weyl Group . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5 The Highest Weight Theorem 75
5.1 Highest Weights . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.2 Verma Modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.3 The Case sl(3, C) . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6 Innite-dimensional Representations 91
6.1 Grding Subspace . . . . . . . . . . . . . . . . . . . . . . . . . . 91
6.2 Induced Lie Algebra Representations . . . . . . . . . . . . . . . . 95
6.3 Self-Adjointness . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.4 Applications to Quantum Mechanics . . . . . . . . . . . . . . . . 102
II Geometric Analysis and Spin Geometry 109
7 Cliord Algebras 111
7.1 Elementary Properties . . . . . . . . . . . . . . . . . . . . . . . . 111
7.2 Classication of Cliord Algebras . . . . . . . . . . . . . . . . . . 117
3
4
7.3 Representation Theory . . . . . . . . . . . . . . . . . . . . . . . . 121
8 Spin Groups 125
8.1 The Cliord Group . . . . . . . . . . . . . . . . . . . . . . . . . . 125
8.2 Pin and Spin Groups . . . . . . . . . . . . . . . . . . . . . . . . . 128
8.3 Double Coverings . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
8.4 Spin Group Representations . . . . . . . . . . . . . . . . . . . . . 135
9 Topological K-Theory 139
9.1 The K-Functors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
9.2 The Long Exact Sequence . . . . . . . . . . . . . . . . . . . . . . 144
9.3 Exterior Products and Bott Periodicity . . . . . . . . . . . . . . 149
9.4 Equivariant K-theory . . . . . . . . . . . . . . . . . . . . . . . . . 151
9.5 The Thom Isomorphism . . . . . . . . . . . . . . . . . . . . . . . 155
10 Characteristic Classes 163
10.1 Connections on Vector Bundles . . . . . . . . . . . . . . . . . . . 163
10.2 Connections on Associated Vector Bundles* . . . . . . . . . . . . 166
10.3 Pullback Bundles and Pullback Connections . . . . . . . . . . . . 172
10.4 Curvature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
10.5 Metric Connections . . . . . . . . . . . . . . . . . . . . . . . . . . 178
10.6 Characteristic Classes . . . . . . . . . . . . . . . . . . . . . . . . 180
10.7 Orientation and the Euler Class . . . . . . . . . . . . . . . . . . . 186
10.8 Splitting Principle, Multiplicative Sequences . . . . . . . . . . . . 190
10.9 The Chern Character . . . . . . . . . . . . . . . . . . . . . . . . . 197
11 Dierential Operators 201
11.1 Dierential Operators on Manifolds . . . . . . . . . . . . . . . . . 201
11.2 The Principal Symbol . . . . . . . . . . . . . . . . . . . . . . . . 205
11.3 Dirac Bundles and the Dirac Operator . . . . . . . . . . . . . . . 210
11.4 Sobolev Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
11.5 Elliptic Complexes . . . . . . . . . . . . . . . . . . . . . . . . . . 227
12 The Atiyah-Singer Index Theorem 233
12.1 K-Theoretic Version . . . . . . . . . . . . . . . . . . . . . . . . . 233
12.2 Cohomological Version . . . . . . . . . . . . . . . . . . . . . . . . 236
A Table of Cliord Algebras 245
B Calculation of Fundamental Groups 247
Bibliography 251
Index 252
Preface
When following courses given by Ryszard Nest at the Copenhagen University,
you can be almost certain that a reference to the Atiyah-Singer Index Theorem
will appear at least once during the course. Thus it was an obvious project for me
to nd out what this, apparently great theorem, was all about. However, from
the beginning I was well aware that this was not an easy task and that it was
necessary for me to delve into a lot of other subjects involved in its formulation,
before the goal could be reached. It has never been my intension to actually
prove the theorem (well except for a few moments of utter over ambitiousness)
but merely to pave a road for my own understanding. This road leads through
as various subjects as K-theory, characteristic classes and elliptic theory. I have
tried to treat each subject as thoroughly and self-contained as I could, even
though this meant including stu which wasnt really necessary for the Index
Theorem.
The starting point is of course my own prerequisites when I began my work
half a year ago, that is a solid foundation in Riemannian geometry, algebraic
topology (notably homology and cohomology) and pseudodierential calculus on
Euclidean space. From here we develop at rst, in a systematic way, topological
K-theory. The approach is via vector bundles as it can be found in for instance
[Atiyah] or [Hatcher], no C
-algebras are involved. In the rst two sections

the basic theory will be outlined and most proofs will be given. In the third
section we present the famous Bott-periodicity Theorem, without giving a proof.
The last two sections are dedicated to the Thom Isomorphism. To this end we
introduce equivariant K-theory (that is, K-theory involving group actions), a
slight generalization of the K-theory treated in the rst sections. I follow the
outline given in the classical article by [Segal]. One could argue, that equivariant
K-theory could have been introduced from the very beginning, however I have
chosen not to, in order not to blur the introductory presentation with too many
technicalities.
The second chapter deals with the Chern-Weil approach to characteristic
classes of vector bundles. The rst four sections are devoted to the study of the
basic theory of connections on vector bundles. From the curvature forms and
invariant polynomials we construct characteristic classes, in particular Chern
and Pontrjagin classes and their relationships will be discussed. In the following
section the Euler class of oriented bundles is dened. I have relied heavily on
[Morita] and [Milnor, Stache] when working out these sections but also [Mad-
sen, Tornehave] has provided valuable inspiration. The chapter ends with a
discussion of certain characteristic classes constructed, not from invariant poly-
nomials but from invariant formal power series. Examples of such classes are
the Todd class and the total

A-class and the Chern character. No eort has
been made to include great theorems, in fact there are really no major results
in this chapter. It serves as a tool box to be applied to the construction of the
topological index.
The third chapter revolves around dierential operators on manifolds. In the
5
6
standard literature on this subject not much care is taken, when transferring the
dierential operators and principal symbols from Euclidean space to manifolds.
Ive tried to remedy this, giving a precise and detailed treatment. To this I have
added a lot of examples of classical dierential operators, such as the Lapla-
cian, Hodge-de Rham operators, Dirac operators etc. calculating their formal
adjoints and principal symbols. To shed some light on the analytic properties we
introduce Sobolev spaces. Essentially there are two dierent denitions: in the
rst one, Sobolev spaces are dened in terms of connections, and in the second
they are dened as the clutching of local Euclidean Sobolev spaces. We prove
that the two denitions agree, when the underlying manifold is compact, and we
show how to extend dierential operators to continuous operators between the
Sobolev spaces. The major results such as the Sobolev Embedding Theorem,
the Rellich lemma and Elliptic Regularity are given without proofs. We then
move on to elliptic complexes, which provides us with a link to the K-theory
developed in the rst chapter.
In the fourth and nal chapter the Index Theorem is presented. We construct
the so-called topological index map from the K-group K(TM) to the integers
and state the index theorem, which says that the index function when evaluated
on the specic K-class determined from the symbol of an elliptic dierential op-
erator, is in fact equal to the Fredholm index. I give a short sketch of the proof
based on the original 1968-article by Atiyah and Singer. Then by introducing
the cohomological Thom isomorphism, Thom Defect classes etc. and drawing
heavily on the theory developed in the previous chapters we manage to deduce
the famous cohomological index formula. To demonstrate the power of the Index
Theorem, we prove two corollaries, namely the generalized Gauss-Bonnet The-
orem and the fact that any elliptic dierential operator on a compact manifold
of odd dimension has index 0.
I would like to thank Professor Ryszard Nest for his guidance and inspiration,
as well as answers to my increasing amount of questions.
Copenhagen, March 2008. Thomas Hjortgaard Danielsen.
Part I
Representation Theory of
Groups and Lie Algebras
7
Chapter 1
Peter-Weyl Theory
1.1 Foundations of Representation Theory
We begin by introducing some basic but fundamental notions and results regard-
ing representation theory of topological groups. Soon, however, we shall restrict
our focus to compact groups and later to Lie groups and their Lie algebras.
We begin with the basic theory. To dene the notion of a representation, let V
denote a separable Banach space and equip B(V ), the space of bounded linear
maps V V , with the strong operator topology i.e. the topology on B(V )
generated by the seminorms |A|
x
= |Ax|. Let Aut(V ) B(V ) denote the
group of invertible linear maps and equip it with the subspace topology, which
turns it into a topological group.
Denition 1.1 (Representation). By a continuous representation of a topo-
logical group G on a separable Banach space V we understand a continuous
group homomorphism : G Aut(V ). We also say that V is given the struc-
ture of a G-module. If is an injective homomorphism the representation is
called faithful .
By the dimension of the representation we mean the dimension of the vector
space on which it is represented. If V is innite-dimensional the representation
is said to be innite-dimensional as well.
In what follows a group without further specication will always denote a
locally compact topological group, and by a representation we will always un-
derstand a continuous representation. The reason why we demand the groups
to be locally compact should be apparent in the next section.
We will distinguish between real and complex representations depending on
whether V is a real or complex Banach space. Without further qualication, the
representations considered will all be complex.
The requirement on to be strongly continuous can be a little hard to handle,
so here is an equivalent condition which is more applicable:
Proposition 1.2. Let : G Aut(V ) be a group homomorphism. Then the
following conditions are equivalent:
1) is continuous w.r.t. the the strong operator topology on Aut(V ), i.e.
is a continuous representation.
2) The map GV V given by (g, v) (g)v is continuous.
For a proof see [1] Proposition 18.8.
10 Chapter 1 Peter-Weyl Theory
Example 1.3. The simplest example one can think of is the trivial repre-
sentation: Let G be a group and V a Banach space, and consider the map
G g id
V
. This is obviously a continuous group homomorphism and hence
a representation.
Now, let G be a matrix Lie group (i.e. a closed subgroup of GL(n, C)).
Choosing a basis for C
n
we get an isomorphism Aut(C
n
)

GL(n, C), and
we can thus dene a representation of G on C
n
simply by the inclusion map
G GL(n, C). This is obviously a continuous representation of G, called the
dening representation.
We can form new representations out of old ones. If (
1
, V
1
) and (
2
, V
2
) are
representations of G on Banach spaces we can form their direct sum
1

2
to be the representation of G on V
1
V
2
(which has been given the norm
|(x, y)| = |x| +|y|, turning V
1
V
2
into a Banach space) given by
(
1
2
)(g)(x, y) = (
1
(g)x,
2
(g)y).
If we have a countable family (H
i
)
iI
of Hilbert spaces we can form the direct
sum Hilbert space

iI
H
i
to be the vector space of sequences (x
i
), x
i
H
i
,
satisfying

iI
|x
i
|
2
Hi
< . Equipped with the inner product (x
i
), (y
i
) =
iI
x
i
, y
i
this is again a Hilbert space. If we have a countable family (
i
, H
i
)
of representations such that sup
iI
|
i
(g)| < for each g G, then we can
form the direct sum of the representations

iI

i
on

iI
H
i
by
_
iI
i
_
(g)(x
i
) = (
i
(g)x
i
).
Finally, if (
1
, H
1
) and (
2
, H
2
) are representations on Hilbert spaces, we can
form the tensor product, namely equip the tensor product vector space H
1
H
2
with the inner product
x
1
x
2
, y
1
y
2
= x
1
, y
1
x
2
, y
2
which turns H
1
H
2
into a Hilbert space, and dene the tensor product repre-
sentation
1
2
by
(
1
2
)(g)(x y) =
1
(g)x
2
(g)y.
Denition 1.4 (Unitary Representation). By a unitary representation of a
group G we understand a representation on a Hilbert space H such that (g)
is a unitary operator for each g G.
Obviously the trivial representation is a unitary representation. As is the
dening representation of any subgroup of the unitary group U(n). In the next
section we show unitarity of some more interesting representations.
Denition 1.5 (Intertwiner). Let two representations (
1
, V
1
) and (
2
, V
2
) of
the same group G be given. By an intertwiner or an intertwining map between
1
and
2
we understand a bounded linear map T : V
1
V
2
rendering the
following diagram commutative
V
1
1(g)
V
2
2(g)
V
1
V
2
i.e. satisfying T
1
(g) =
2
(g) T for all g G. The set of all intertwining
maps is denoted Hom
G
(V
1
, V
2
).
1.1 Foundations of Representation Theory 11
A bijective intertwiner with bounded inverse between two representations is
called an equivalence of representations and the two representations are said to
be equivalent. This is denoted
1

=
2
.
Its easy to see that Hom
G
(V
1
, V
2
) is a vector space, and that Hom
G
(V, V ) is
an algebra. The dimension of Hom
G
(V
1
, V
2
) is called the intertwining number of
the two representations. If
1

=
2
via an intertwiner T, then we have
2
(g) =
T
1

1
(g) T. Since we thus can express the one in terms of the other, for
almost any purpose the two representations can be regarded as the same.
Proposition 1.6. Hom
G
respects direct sum in the sense that
Hom
G
(V
1
V
2
, W)
= Hom
G
(V
1
, W) Hom
G
(V
2
, W) and (1.1)
Hom
G
(V, W
1
W
2
)
= Hom
G
(V, W
1
) Hom
G
(V, W
2
). (1.2)
Proof. For the rst isomorphism we dene
: Hom
G
(V
1
V
2
, W) Hom
G
(V
1
, W) Hom
G
(V
2
, W)
by (T) := (T[
V1
, T[
V2
). It is easy to check that this is indeed an element of the
latter space. It has an inverse
1
given by
1
(T
1
, T
2
)(v
1
, v
2
) := T
1
(v
1
) +T
2
(v
2
),
and this proves the rst isomorphism. The latter can be proved in the same
way.
Denition 1.7. Given a representation (, V ) of a group G, we say that a
linear subspace U V is -invariant or just invariant if (g)U U for all
g G.
If U is a closed invariant subspace for a representation of G on V , we
automatically get a representation of G on U, simply by restricting all the (g)s
to U (U should be a Banach space, and therefore we need U to be closed). This
is clearly a representation, and we will denote it [
U
(although we are restricting
the (g)s to U and not ).
Here is a simple condition to check invariance of a given subspace, at least in
the case of a unitary representation
Lemma 1.8. Let (, H) be a representation of G, let H = U U
be a de-
composition of H and denote by P : H U the orthogonal projection onto U.
If U is -invariant then so is U
. Furthermore U is -invariant if and only if

P (g) = (g) P for all g G.
Proof. Assume that U is invariant. To show that U
is invariant let v U
.
We need to show that (g)v U
, i.e. u U : (g)v, u = 0. But thats

easy, exploiting unitarity of (g):
(g)v, u = (g
1
)((g)v), (g
1
)u = v, (g
1
)u
which is 0 since (g
1
)u U and v U
. Thus U
is invariant
Assume U to be invariant. Then also U
is invariant by the above. We split

x H into x = Px + (1 P)x and calculate
P (g)x = P((g)(Px + (1 P)x)) = P(g)Px +P(g)(1 P)x.
The rst term is (g)Px, since (g)Px U, and the second term is zero, since
(g)(1 P)x U
. Thus we have the desired formula.

Conversely, assume that P (g) = (g) P. Every vector u U is of the
form Px for some x H. Since
(g)u = (g)(Px) = P((g)x) U,
U is an invariant subspace.
For any representation (, V ) it is easy to see two obvious invariant subspaces,
namely V itself and 0. We shall focus a lot on representations having no
invariant subspaces except these two:
Denition 1.9. A representation is called irreducible if it has no closed invari-
ant subspaces except the trivial ones. The set of equivalence classes of nite-
dimensional irreducible representations of a group G is denoted

G.
A representation is called completely reducible if it is equivalent to a direct
sum of nite-dimensional irreducible representations.
Any 1-dimensional representation is obviously irreducible, and if the group is
abelian the converse is actually true. We prove this in Proposition 1.14
If (
1
, V
1
) and (
2
, V
2
) are irreducible representations then the direct sum
2
is not irreducible, since V
1
is an
1
2
-invariant subspace of V
1
V
2
:
(
1
2
)(g)(v, 0) = (
1
(g)v, 0).
The question is more subtle when considering tensor products of irreducible
representations. Whether or not the tensor product of two irreducible repre-
sentations is irreducible and if not, to write is as a direct sum of irreducible
representations is a branch of representation theory known as Clebsch-Gordan
theory.
Lemma 1.10. Let (
1
, V
1
) and (
2
, V
2
) be equivalent representations. Then
1
is irreducible if and only if
2
is irreducible.
Proof. Given the symmetry of the problem, it is sucient to verify that ir-
reducibility of
1
implies irreducibility of
2
. Let T : V
1
V
2
denote the
intertwiner, which by the Open Mapping Theorem is a linear homeomorphism.
Assume that U V
2
is a closed invariant subspace. Then T
1
U V
1
is closed
and
1
-invariant:
1
(g)T
1
U = T
1
2
(g)U T
1
U
But this means that T
1
U is either 0 or V
1
, i.e. U is either 0 or V
2
.
Example 1.11. Consider the group SL(2, C) viewed as a real (hence 6-dimensional)
Lie group. We consider the following 4 complex representations of the real Lie
group SL(2, C) on C
2
:
(A) := A, (A) := A,
(A) := (A
T
)
1
,

(A) := (A
)
1
,
where A simply means complex conjugation of all the entries. All four are clearly
irreducible. They are important in physics where they are called spinorial rep-
resentations. The physicists have a habit of writing everything in coordinates,
thus will usually be written
, where = 1, 2 but the exact notation will

vary according to which representation we have imposed on C
2
(i.e. according
to how transforms as the physicists say). In other words they view C
2
not as
a vector space but rather as a SL(2, C)-module. The notations are
C
2
,

C
2
,

C
2
,

C
2
.
1.1 Foundations of Representation Theory 13
The representations are not all mutually inequivalent, actually the map :
C
2
C
2
given by the matrix
_
0 1
1 0
_
intertwines with and intertwines
with
. On the other hand and are actually inequivalent as we will se in Sec-

tion 1.4. These two representations are called the fundamental representations
of SL(2, C).
In short, representation theory has two goals: 1) given a group: nd all the
irreducible representations and 2) given a representation of this group: split it (if
possible) into a direct sum of irreducibles. The rest of this chapter deals with the
second problem (at least for compact groups) and in the end we will achieve some
powerful results (Schur Orthogonality and the Peter-Weyl Theorem). Chapter
5 revolves around the rst problem of nding irreducible representations.
But already at this stage we are able to state and prove two quite interesting
results. The rst result is known as Schurs Lemma. We prove a slightly more
general version than is usually seen, allowing the representations to be innite-
dimensional.
Theorem 1.12 (Schurs Lemma). Let (
1
, H
1
) and (
2
, H
2
) be two irre-
ducible unitary representations of a group G, and suppose that F : H
1
H
2
is an intertwiner. Then either F is an equivalence of representations or F is the
zero map.
If (, H) is an irreducible unitary representation of G and F B(H) is a
linear map which commutes with all (g), then F = id
H
.
Proof. The proof utilizes a neat result from Gelfand theory: suppose that
A is a commutative unital C*-algebra which is also an integral domain (i.e.
ab = 0 implies a = 0 or b = 0), then A

= Ce. The proof is rather simple.
Gelfands Theorem states that there exists a compact Hausdor space X such
that A

= C(X). To reach a contradiction, assume that X is not a one-point
set, and pick two distinct points x and y. Then since X is a normal topological
space, we can nd disjoint open neighborhoods U and V around x and y, and
the Urysohn Lemma gives us two nonzero continuous functions f and g on X,
the rst one supported in U and the second in V , the product, thus, being
zero. This contradicts the assumption that A = C(X) was an integral domain.
Therefore X can contain only one point and thus C(X)
= C.
With this result in mente we return to Schurs lemma. F being an intertwiner
means that F
1
(g) =
2
(g) F, and using unitarity of
1
(g) and
2
(g) we get
that
F

2
(g) =
1
(g) F
where F
is the hermitian adjoint of F. This yields

(FF
)
2
(g) = F
1
(g) F
=
2
(g) (FF
).
In the last equality we also used that F intertwines the two representations.
Consider the C
-algebra A = C
(id
H2
, FF
), the C
-algebra generated by id
H2
and FF
. Its a commutative unital C
-algebra, and all the elements are of the

form

n=0
a
n
(FF
)
n
. They commute with
2
(g):
_

n=1
a
n
(FF
)
n
_
2
(g) =
n=1
(a
n
(FF
)
n
2
(g)) =
n=1
a
n
(
2
(g)(FF
)
n
)
=
2
(g)
n=1
a
n
(FF
)
n
.
We only need to show that A is an integral domain. Assume ST = 0. Since
2
(g)S = S
2
(g) its easy to see that ker S is
2
-invariant.
2
is irreducible
so ker S is either H
2
or 0. In the rst case S = 0, and we are done, in the
second case, S is injective, and so T must be the zero map. This means that
A = Cid
H2
, in particular, there exists a C so that FF
= id
H2
. Likewise,
one shows that F
F =
t
id
H1
. Thus, we see
F = F(F
F) = (FF
)F =
t
F
which implies F = 0 or =
t
. In the second case if =
t
= 0 then F
Fv = 0
for all v H
1
, and hence
0 = v, F
Fv = Fv, Fv,
i.e. F = 0. If =
t
and ,= 0 then it is not hard to see that
1
2
F is unitary,
and that F therefore is an isomorphism.
The second claims is an immediate consequence of the proof of the rst.
The content of this can be summed up to the following: If
1
and
2
are irre-
ducible unitary representations of G on H
1
and H
2
, then Hom
G
(H
1
, H
2
)
= C if
1
and
2
are equivalent and Hom
G
(H
1
, H
2
) = 0 if
1
and
2
are inequivalent.
Corollary 1.13. Let (, H
1
) and (, H
2
) be nite-dimensional unitary repre-
sentations which decompose into irreducibles
=
iI
m
i
i
and =
iI
n
i
i
.
Then dimHom
G
(H
1
, H
2
) =
iI
n
i
m
i
.
Proof. Denoting the representations spaces of the irreducible representations
by V
i
we get from (1.1) and (1.2) that
Hom
G
(H
1
, H
2
) =
iI
jI
n
i
m
j
Hom(V
i
, V
j
),
and by Schurs Lemma the dimension formula now follows.
Now for the promised result on abelian groups
Proposition 1.14. Let G be an abelian group and (, H) be a unitary repre-
sentation of G. If is irreducible then is 1-dimensional.
Proof. Since G is abelian we have (g)(h) = (h)(g) i.e. (h) is an in-
tertwiner. Since is irreducible, Schurs Lemma says that (h) = (h) id
H
.
Thus, each 1-dimensional subspace of H is invariant, and by irreducibility H is
1-dimensional.
Example 1.15. With the previous lemma we are in a position to determine the
set of irreducible complex representations of the circle group = 1/Z. Since this
is an abelian group, we have found all the irreducible representations when we
know all the 1-dimensional representations. A 1-dimensional representation is
just a homomorphism 1/Z C
, so lets nd them: It is well-known that the

only continuous homomorphisms 1 C
are those of the form x e

2iax
for some a 1. But since we also want it to be periodic with periodicity 1,
only integer values of a are allowed. Thus,

consists of the homomorphisms
n
(x) = e
2inx
for n Z.
Proposition 1.16. Every nite-dimensional unitary representation is com-
pletely reducible.
1.2 The Haar Integral 15
Proof. If the representation is irreducible then we are done, so assume we
have a unitary representation : G Aut(H) and let 0 , = U H be an
invariant subspace. The point is that U
is invariant as well cf. Lemma 1.8. If

both [
U
and [
U
are irreducible we are done. If one of them is not, we nd
an invariant subspace and perform the above argument once again. Since the
representation is nite-dimensional and since 1-dimensional representations are
irreducible, the argument must stop at some point.
1.2 The Haar Integral
In the representation theory of locally compact groups (also known as harmonic
analysis) the notions of Haar integral and Haar measure play a key role.
Some preliminary denitions: Let X be a locally compact Hausdor space
and C
c
(X) the space of complex valued functions on X with compact support.
By a positive integral on X is understood a linear functional I : C
c
(X) C
such that I(f) 0 if f 0. The Riesz Representation Theorem tells us that to
each such positive integral there exists a unique Radon measure on the Borel
algebra B(X) such that
I(f) =
_
X
fd.
We say that this measure is associated with the positive integral.
Now, let G be a group. For each g
0
G we have two maps L
g0
and R
g0
, left
and right translation, on the set of complex-valued functions on G, given by
(L
g0
f)(g) = f(g
1
0
g), (R
g0
f)(g) = f(gg
0
).
These obviously satisfy L
g1g2
= L
g1
L
g2
and R
g1g2
= R
g1
R
g2
.
Denition 1.17 (Haar Measure). Let G be a locally compact group. A
nonzero positive integral I on G is called a left Haar integral if I(L
g
f) = I(f)
for all g G and f C
c
(X). Similarly a nonzero positive integral is called a
right Haar integral if I(R
g
f) = I(f) for all g G and f C
c
(X). An integral
which is both a left and a right Haar integral is called a Haar integral .
The measures associated with left and right Haar integrals are called left and
right Haar measures. The measure associated with a Haar integral is called a
Haar measure.
Example 1.18. On (1
n
, +) the Lebesgue integral is a Haar integral: it is ob-
viously positive, and it is well-known that the Lebesgue integral is translation
invariant:
_
R
n
f(x +a)dx =
_
R
n
f(a +x)dx =
_
R
n
f(x)dx.
The associated Haar measure is of course the Lebesgue measure m
n
.
On the circle group (, ) we dene an integral I by
C() f
1
2
_
2
0
f(e
it
)dt.
As before this is obviously a positive integral and since
I(L
e
iaf) =
1
2
_
2
0
f(e
ia
e
it
)dt =
1
2
_
2
0
f(e
i(a+t)
)dt
=
1
2
_
2
0
f(e
it
)dt
again by exploiting translation invariance of the Lebesgue measure, I is a left
Haar integral on . Likewise one can show that it is a right Haar integral as
well, and hence a Haar integral. The associated Haar measure on is also called
the arc measure.
In both cases the groups were abelian and in both cases the left Haar integrals
were also right Haar integrals. This is no mere coincidence for if G is an abelian
group we have L
g0
= R
g
1
0
and thus a positive integral is a left Haar integral if
and only if it is a right Haar integral.
The following central theorem attributed to Alfred Haar and acclaimed as
one of the most important mathematical discoveries in the 20th century states
existence and uniqueness of left and right Haar integrals on locally compact
groups.
Theorem 1.19. Every locally compact group G possesses a left Haar integral
and a right Haar integral, and these are unique up to multiplication by a positive
constant.
If G is compact then the two integrals coincide, and the corresponding Haar
measure is nite.
It would be far beyond the scope of this thesis to delve into the proof of this.
The existence part of the proof is a hard job so we just send some acknowledging
thoughts to Alfred Haar and accept it as a fact of life.
Now we restrict focus to compact groups on which, as we have just seen, we
have a nite Haar measure. The importance of this niteness is manifested in
the following result:
Theorem 1.20 (Unitarization). Let G be a compact group and (, H) are
representation on a Hilbert space (H, , ). Then there exists an inner product
,
G
on H equivalent to , which makes a unitary representation.
Proof. Since the measure is nite, we can integrate all bounded measurable
functions over G. Let us assume the measure to be normalized, i.e. that (G) =
1. For x
1
, x
2
H the map g (g)x
1
, (g)x
2
is continuous (by Proposition
1.2), hence bounded and measurable, i.e. integrable. Now dene a new inner
product by
x
1
, x
2
G
:=
_
G
(g)x
1
, (g)x
2
dg. (1.3)
That this is a genuine inner product is not hard to see: it is obviously sesqui-
linear by the properties of the integral, it is conjugate-symmetric, as the original
inner product is conjugate-symmetric. Finally, if x ,= 0 then (g)x ,= 0 ((g) is
invertible) and thus |(g)x| > 0 for all g G. Since the map g |(g)x|
2
is
continuous we have x, x
G
=
_
G
|(g)x|dg > 0.
By the translation of the Haar measure we get
(h)x
1
, (h)x
2
G
=
_
G
(gh)x
1
, (gh)x
2
dg
=
_
G
(g)x
1
, (g)x
2
dg
= x
1
, x
2
G
.
Thus, is unitary w.r.t. this new inner product.
We just need to show that the two norms | | and | |
G
corresponding to
the two inner products are equivalent, i.e. that there exists a constant C so that
|| C||
G
and ||
G
C||. To this end, consider the map g |(g)x|
2
for
some x H. Its a continuous map, hence sup
gG
|(g)x|
2
< for all x, and
1.2 The Haar Integral 17
the Uniform Boundedness Principle now says that C := sup
gG
|(g)| < .
Therefore
|x|
2
=
_
G
|x|
2
dg =
_
G
|(g
1
)(g)x|
2
dg C
2
_
G
|(g)x|
2
= C
2
|x|
2
G
.
Conversely we see
|x|
2
G
=
_
G
|(g)x|
2
_
G
|(g)|
2
|x|
2
dg C
2
_
G
|x|
2
dg = C
2
|x|
2
.
This proves the claim.
If we combine this result with Proposition 1.16 we get
Corollary 1.21. Every nite-dimensional representation of a compact group is
completely reducible.
The Peter-Weyl Theorem which we prove later in this chapter provides a
strong generalization of this result in that it states that every Hilbert space
representation of a compact group is completely reducible.
We end this section by introducing the so-called modular function which is a
function that provides a link between left and right Haar integrals.
Let G be a topological group and I : f
_
G
f(g)dg a left Haar integral.
Let h G and consider the integral

I
h
: f
_
G
f(gh
1
)dg. This is positive
and satises
I
h
(L
g0
f) =
_
G
f(g
1
0
gh
1
)dg =
_
G
f(gh
1
)dg =

I
h
(f)
i.e. is a left Haar integral. By the uniqueness part of Haars Theorem there exists
a positive constant c such that

I
h
(f) = cI(f). We dene the modular function
: G 1
+
by assigning this constant to the group element h i.e.
_
G
f(gh
1
)dg = (h)
_
G
f(g)dg.
It is not hard to see that this is indeed a homomorphism: on one hand we have
_
G
f(g(hk)
1
)dg = (hk)
_
G
f(g)dg,
and on the other hand we have that this equals
_
G
f(gk
1
h
1
)dg = (h)
_
G
f(gk
1
)dg = (h)(k)
_
G
.f(g)dg
Since this holds for all integrable functions f we must have (hk) = (h)(k).
One can show that this is in fact a continuous group homomorphism and thus
in the case of G being a Lie group, a Lie group homomorphism.
If is identically 1, that is if every right Haar integral satises
_
G
f(hg)dg =
_
G
f(g)dg (1.4)
for all h, then the group G is called unimodular. Eq. (1.4) says that an equivalent
condition for a group to be unimodular is that all right Haar integrals are also
left Haar integrals. As we have seen previously in this section abelian groups
and compact groups are unimodular groups.
1.3 Matrix Coecients
Denition 1.22 (Matrix Coecient). Let (, V ) be a nite-dimensional rep-
resentation of a compact group G. By a matrix coecient for the representation
we understand a map G C of the form
m
v,
(g) = ((g)v)
for xed v V and V
.
If we pick a basis e
1
, . . . , e
n
for V and let
1
, . . . ,
n
denote the correspond-
ing dual basis, then we see that m
ei,j
=
j
((g)e
i
) precisely are the entries of
the matrix-representation of (g), therefore the name matrix coecient.
If V comes with an inner product , , then by the Riesz Theorem all matrix
coecients are of the form m
v,w
= (g)v, w for xed v, w V . By Theorem
1.20 we can always assume that this is the case.
Denote by C(G)
the space of linear combinations of matrix coecient. Since

a matrix coecient is obviously a continuous map, C(G)
C(G) L
2
(G).
Thus, we can take the inner product of two functions in C(G)
. Note, however
that the elements of C(G)
need not all be matrix coecients for .

The following technical lemma is an important ingredient in the proof of the
Schur Orthogonality Relations which is the main result of this section.
Lemma 1.23. Let (, H) be a nite-dimensional unitary representation of a
compact group G. Dene the map T
: End(H) C(G) by
T
(A)(g) = Tr((g) A). (1.5)

Then C(G)
= imT
.
Proof. Given a matrix coecient m
v,w
we should produce a linear map A :
H H, such that m
v,w
= T
(A). Consider the map L

v,w
: H H dened
by L
v,w
(u) = u, wv, the claim is that this is the desired map A. To see this
we need to calculate Tr L
v,w
and we claim that the result is v, w. Since L
v,w
is sesquilinear in its indices (L
av+bv
,w
= aL
v,w
+ bL
v
,w
), its enough to check
it on elements of an orthonormal basis e
1
, . . . , e
n
for H.
Tr L
ei,ei
=
n
k=1
L
ei,ei
e
k
, e
k
=
n
k=1
e
k
, e
i
e
i
, e
k
= 1
while for i ,= j
Tr L
ei,ej
=
n
k=1
L
ei,ej
e
k
, e
k
=
n
k=1
e
k
, e
j
e
i
, e
k
= 0.
Thus, Tr L
v,w
= v, w. Finally since
L
v,w
(g)u = (g)u, wv = u, (g
1
)wv = L
v,(g
1
)w
u
we see that
T
(L
v,w
)(g) = Tr((g) L
v,w
) = Tr(L
v,w
(g)) = v, (g
1
)w = (g)v, w
= m
v,w
(g).
Conversely, we should show that any map T
(A) is a linear combination of

matrix coecients. Some linear algebraic manipulations should be enough to
1.3 Matrix Coecients 19
convince the reader that we for any A End(H) have A =
n
i,j=1
Ae
j
, e
i
L
ei,ej
w.r.t some orthonormal basis e
1
, . . . , e
n
. But then we readily see
T
(A)(g) = T
_
n
i,j=1
Ae
j
, e
i
L
ei,ej
_
(g) =
n
i,j=1
Ae
j
, e
i
T
(L
ei,ej
)(g)
=
n
i,j=1
Ae
j
, e
i
m
ei,ej
(g).
Theorem 1.24 (Schur Orthogonality I). Let (
1
, H
1
) and (
2
, H
2
) be two
unitary, irreducible nite-dimensional representations of a compact group G. If
1
and
2
are equivalent, then we have C(G)
1
= C(G)
2
. If they are not, then
C(G)
1
C(G)
2
inside L
2
(G).
Before the proof, a few remarks on the integral of a vector valued function
would be in order. Suppose that f : G H is a continuous function into
a nite-dimensional Hilbert space. Choosing a basis e
1
, . . . , e
n
for H we can
write f in its components f =
n
i=1
f
i
e
i
, which are also continuous, and dene
_
G
f(g)dg :=
n
i=1
_
G
f
i
(g)dg e
i
.
Its a simple change-of-basis calculation to verify that this is independent of the
basis in question. Furthermore, one readily veries that it is left-invariant and
satises
_
_
G
f(g)dg, v
_
=
_
G
f(g), vdg and A
_
G
f(g)dg =
_
Af(g)dg
when A End(H).
Proof of Theorem 1.24. If
1
and
2
are equivalent, there exists an isomor-
phism T : H
1
H
2
such that T
1
(g) =
2
(g)T. For A End(H
1
) we see
that
T
2
(TAT
1
)(g) = Tr(
2
(g)TAT
1
) = Tr(T
1
2
(g)TA) = Tr(
1
(g)A)
= T
1
(A)(g).
Hence the map sending T
1
(A) to T
2
(TAT
1
) is the identity id : C(G)
1

C(G)
2
proving that the two spaces are equal.
Now we show the second claim. Dene for xed w
1
H
1
and w
2
H
2
the
map S
w1,w2
: H
1
H
2
by
S
w1,w2
(v) =
_
G
1
(g)v, w
1
2
(g
1
)w
2
dg.
S
w1,w2
is in Hom
G
(H
1
, H
2
) since by left-invariance
S
w1,w2
1
(h)(v) =
_
G
1
(gh)v, w
1
2
(g
1
)w
2
dg =
_
G
1
(g)v, w
1
2
(hg
1
)w
2
dg
=
2
(h)
_
G
1
(g)v, w
1
2
(g
1
)w
2
dg
=
2
(h)S
w1,w2
(v).
Assume that we can nd two matrix coecients m
v1,w1
and m
v2,w2
for
1
and
2
that are not orthogonal, i.e. we assume that
0 ,=
_
G
m
v1,w2
(g)m
v2,w2
(g)dg =
_
G
1
(g)v
1
, w
1
2
(g)v
2
, w
2
dg
=
_
G
1
(g)v
1
, w
1
2
(g
1
)w
2
, v
2
dg.
From this we read S
w1,w2
v
1
, v
2
, = 0, so that S
w1,w2
,= 0. Since its an inter-
twiner, Schurs Lemma tells us that S
w1,w2
is an isomorphism. By contraposition,
the second claim is proved.
In the case of two matrix coecients for the same representation, we have the
following result
Theorem 1.25 (Schur Orthogonality II). Let (, H) be a unitary, nite-
dimensional irreducible representation of a compact group G. For two matrix
coecients m
v1,w1
and m
v2,w2
we have
m
v1,w1
, m
v2,w2
=
1
dimH
v
1
, v
2
w
2
, w
1
. (1.6)
Proof. As in the proof of Theorem 1.24 dene S
w1,w2
: H H by
S
w1,w2
(v) =
_
G
1
(g)v, w
1
2
(g
1
)w
2
dg =
_
G
(g
1
)L
w2,w1
(g)v dg.
We see that
m
v1,w1
, m
v2,w2
=
_
G
(g)v
1
, w
1
(g)v
2
, w
2
dg
=
_
G
(g)v
1
, w
1
(g
1
)w
2
, v
2
dg
=
_
G
(g)v
1
, w
1
(g
1
)w
2
, v
2
= S
w1,w2
v
1
, v
2
.
Furthermore, since S
w1,w2
commutes with (g), Schurs Lemma yields a com-
plex number (w
1
, w
2
), such that S
w1,w2
= (w
1
, w
2
) id
H
. The operator S
w1,w2
is linear in w
2
and anti-linear in w
1
, hence (w
1
, w
2
) is a sesquilinear form on
H. We now take the trace on both sides of the equation S
w1,w2
= (w
1
, w
2
) id
H
:
the right hand side is easy, its just (w
1
, w
2
) dimH. For the left hand side we
calculate That is, we get (w
1
, w
2
) = (dimH)
1
w
1
, w
2
, and hence
S
w1,w2
= (dimH)
1
w
1
, w
2
id
H
.
By substituting this into the equation m
v1,w1
, m
v2,w2
= S
w1,w2
v
1
, v
2
the
desired result follows.
1.4 Characters
Denition 1.26 (Class Function). For a group G, a class function is a func-
tion on G which is constant on conjugacy classes. The set of square-integrable
resp. continuous class functions on G are denoted L
2
(G, class) and C(G, class).
It is not hard to see that the closure of C(G, class) inside L
2
(G) is L
2
(G, class).
Thus, L
2
(G, class) is a Hilbert space. Given an irreducible nite-dimensional
representation the set of continuous class functions inside C(G)
is very small:
Lemma 1.27. Let (, H) be a nite-dimensional irreducible unitary represen-
tation of a compact group G, then the only class functions inside C(G)
are
complex scalar multiples of T
(id
H
).
1.4 Characters 21
Proof. To formulate the requirement on a class function, consider the repre-
sentation of G on C(G) by ((g)f)(x) = f(g
1
xg), then in terms of this a
function f is a class function if and only if (g)f = f for all g.
For reasons which will be clear shortly, we introduce another representation
of G on End(H) by
(g)A = (g)A(g
1
).
Equipping End(H) with the inner product A, B := Tr(B
A), it is easy to see

that becomes unitary. The linear map T
: End(H) C(G)
which we
introduced in Lemma 1.23 is an intertwiner of the representations and :
T
((g)A)(x) = Tr
_
(x)(g)A(g
1
)
_
= Tr((g
1
xg)a)
= (g) Tr((x)A) = (g)T
(A)(x).
T
was surjective by Lemma 1.23. To show injectivity we dene

T
:=
dimH T
and show that this is unitary. Since the linear maps L

v,w
span End(H) it is
enough to show unitarity on these. But rst we need some facts concerning
L
v,w
:
L
v,w
x, y = x, wv, y = x, wv, y = x, v, yw
= x, y, vw = x, L
w,v
showing that L
v,w
= L
w,v
. Furthermore
L
w
,v
L
v,w
x = L
w
,v
(x, wv) = x, wv, v
t
w
t
= v, v
t
x, ww
t
= v, v
t
L
w
,w
x.
With the inner product on End(H) these results now yield
L
v,w
, L
v
,w
= Tr(L
w
,v
L
v,w
) = Tr(v, v
t
L
w
,w
) = v, v
t
w, w
t
.
Since T
(L
v,w
)(x) = m
v,w
(x), and using Schur Orthogonality II we see
(L
v,w
),

T
(L
v
,w
) = dimHm
v,w
, m
v
,w
= v, v
t
w, w
t
= L
v,w
, L
v
,w
.
Thus

T
is unitary and in particular injective.

Now we come to the actual proof: let C(G)
be a class function.

T
is bijective, so there is a unique A End(H) for which =

T
(A). That

T
intertwines and leads to

(g
1
xg) = ((g))(x) = (g)
(A)(x) =

T
((g)A)(x) =

T
((g)A(g
1
)),
and since was a class function we get that (g)A(g
1
) = A, i.e. A intertwines
. But was irreducible, which by Schurs Lemma implies A = id
H
, and hence
= T
(id
H
).
In particular there exists a unique class function
0
which is positive on e and
which has L
2
-norm 1: namely we have
|
0
|
2
2
= |
(A)|
2
2
= |A|
2
= Tr(A
A)
so if
0
should have norm 1 and be positive on e, then A is forced to be
(dimH)
1
id
H
, so that
0
is given by
0
(g) = Tr (g). This is a function of
particular interest:
Denition 1.28 (Character). Let (, V ) be a nite-dimensional representa-
tion of a group G. By the character of we mean the function
: G C
given by
(g) = Tr (g).
If is a character of an irreducible representation, is called an irreducible
character.
The character is a class function, so in the case of two representations
1
and
2
being equivalent via the intertwiner T:
2
(g) = T
1
(g)T
1
, we have
1
=
2
. Thus, equivalent representations have the same character. Actually,
the converse is also true, we show that at the end of the section.
Suppose that G is a topological group, and that H is a Hilbert space with
orthonormal basis e
1
, . . . , e
n
. Then we can calculate the trace as
Tr (g) =
n
i=1
(g)e
i
, e
i
which shows that
C(G)
. In due course we will prove some powerful or-

thogonality relations for irreducible characters. But rst we will see that the
character behaves nicely with respect to direct sum and tensor product opera-
tions on representations.
Proposition 1.29. Let (
1
, V
1
) and (
2
, V
2
) be two nite-dimensional repre-
sentations of the group G. The characters of
1
2
and
1
2
are then given
by
12
(g) =
1
(g) +
2
(g) and
12
(g) =
1
(g)
2
(g). (1.7)
Proof. Equip V
1
and V
2
with inner products and pick orthonormal bases (e
i
)
and (f
j
) for V
1
and V
2
respectively. Then the vectors (e
i
, 0), (0, f
j
) form an
orthonormal basis for V
1
V
2
w.r.t. the inner product
(v
1
, v
2
), (w
1
, w
2
) := v
1
, w
1
+v
2
, w
2
.
Thus we see
12
(g) = Tr
1
2
(g)
=
m
i=1
2
(g)(e
i
, 0), (e
i
, 0)
_
+
n
j=1
2
(g)(0, f
j
), (0, f
j
)
_
=
m
i=1
1
(g)e
i
, e
i
+
n
j=1
2
(g)f
j
, f
j
=
1
(g) +
2
(g).
Likewise, the vectors e
i
f
j
constitute an orthonormal basis for V
1
V
2
w.r.t.
the inner product
v
1
v
2
, w
1
w
2
:= v
1
, w
1
v
2
, w
2
,
and hence
12
(g) = Tr
1
2
(g) =
m,n
i,j=1
2
(g)(e
i
f
j
), (e
i
f
j
)
_
=
m,n
i,j=1
1
(g)e
i
, e
i
2
(g)f
j
, f
j
=
m
i=1
1
(g)e
i
, e
i
j=1
2
(g)f
j
, f
j
=
1
(g)
2
(g).
1.4 Characters 23
The following lemma, stating the promised orthogonality relations of char-
acters, shows that irreducible characters form an orthonormal set in C(G).
The Schur Orthogonality Relations are important ingredients in the proof, thus
henceforth we need the groups to be compact.
Lemma 1.30. Let (
1
, V
1
) and (
2
, V
2
) be two nite-dimensional irreducible
representations of a compact group G. Then the following hold:
1)
1

=
2
implies
1
,
2
= 1.
2)
1

2
implies
1
,
2
= 0.
Proof. In the rst case, we have a bijective intertwiner T : V
1
V
2
. Choose
an inner product on V
1
and an orthonormal basis (e
i
) for V
1
. Dene an inner
product on V
2
by declaring T to be unitary. Then (Te
i
) is an orthonormal basis
for V
2
. Let n = dimV
1
= dimV
2
. The expressions
1
=
n
i=1
1
(g)e
i
, e
i
and
2
=
n
j=1
2
(g)Te
j
, Te
j
along with (1.6) yield
1
,
2
=
n
i,j=1
_
G
1
(g)e
i
, e
i
2
(g)Te
j
, Te
j
dg
=
n
i,j=1
_
G
1
(g)e
i
, e
i
T
1
(g)e
j
, Te
j
dg
=
n
i,j=1
_
G
1
(g)e
i
, e
i
1
(g)e
j
, e
j
dg
=
1
n
n
i,j=1
e
i
, e
j
e
i
, e
j
=
1
n
n
i=1
1 = 1.
In the second case, if
1
and
2
are non-equivalent then by Theorem 1.24
we have C(G)
1
C(G)
2
. Since
1
C(G)
1
and
2
C(G)
2
, the result
follows.
This leads to the main result on characters:
Theorem 1.31. Let be an nite-dimensional representation of a compact
group G. Then decomposes according to
,
i
i
.
Proof. Proposition 1.16 says that

=
m
i
i
where
i
is irreducible and m
i
is the number of times that
i
occurs in . From Lemma 1.30 it follows that
i
m
i
i
and hence by orthonormality of the irreducible characters that
m
i
=
,
i
.
Example 1.32. A very simple example to illustrate this is the following. Con-
sider the 2-dimensional representation of given by
x
1
2
_
e
2inx
+e
2imx
e
2inx
+e
2imx
e
2inx
+e
2imx
e
2inx
+e
2imx
_
for n, m Z. It is easily seen to be a continuous homomorphism Aut(C
2
)
with character
(x) = e
2imx
+ e
2inx
. But the two terms are irreducible
characters for , cf. Example 1.15, and by Theorem 1.31 we have

=
n

m
.
Corollary 1.33. For nite-dimensional representations
1
,
2
and of a com-
pact group we have:
1)
1

=
2
if and only if
1
=
2
.
2) is irreducible if and only if
= 1.
Proof. For the rst statement, the only-if part is true by the remarks following
the denition of the character. To see the converse, assume that
1
=
2
.
Then for each irreducible representation we must have
1
,
=
2
,
and therefore
1
and
2
are equivalent to the same decomposition of irreducible
representations, hence they are equivalent.
If is irreducible then Lemma 1.30 states that
= 1. Conversely,
assume
= 1 and decompose into irreducibles:

=
m
i
i
. Or-
thonormality of the irreducible characters again gives
m
2
i
. From
this it is immediate that there is precisely one m
i
which is 1, while the rest are
0, i.e.

=
i
. Therefore is irreducible.
Considering the representations and from Example 1.11 we see that the
corresponding characters satisfy
and since
is actually a complex
map, they are certainly not equal. Hence the representations are inequivalent.
1.5 The Peter-Weyl Theorem
The single most important theorem in the representation theory of compact
topological groups is the Peter-Weyl Theorem. It has numerous consequences,
some of which we will mention at the end of this section.
Theorem 1.34 (Peter-Weyl I). Let G be a compact group. Then the subspace
M(G) :=
G
C(G)
of C(G) is dense in L
2
(G).
In other words the linear span of all matrix coecients of the nite-dimensional
irreducible representations of G is dense in L
2
(G).
Proof. We want to show that M(G) = L
2
(G). We prove it by contradiction
and assume that M(G)
,= 0. Now, suppose that M(G)
(which is a closed
subspace of L
2
(G) and hence a Hilbert space itself) contains a nite-dimensional
R-invariant subspace W (R is the right-regular representation) such that R[
W
is irreducible (we prove below that this is a consequence of the assumption
M(G)
,= 0). Then we can pick an nite orthonormal basis (

i
) for W, and
then for 0 ,= f W
f(x) =
N
i=1
f,
i
i
(x).
This is a standard result in Hilbert space theory. Then we see that
f(g) = (R[
W
(g)f)(e) =
N
i=1
R[
W
(g)f,
i
i
(e).
Since R[
W
is a nite-dimensional irreducible representation, the map g
R[
W
(g),
i
is a matrix coecient. But this means that f M(G), hence a
contradiction.
1.5 The Peter-Weyl Theorem 25
Now, lets prove the existence of the nite-dimensional right-invariant sub-
space. Let f
0
M(G)
be nonzero. As C(G) is dense in L

2
(G) we can nd a
C(G) such that , f
0
, = 0 where (g) = (g
1
). Dene K C(GG) by
K(x, y) = (xy
1
) and let T : L
2
(G) L
2
(G) be the integral operator with
K as its kernel:
Tf(x) =
_
G
K(x, y)f(y)dy.
According to functional analysis, this is a well-dened compact operator, and it
commutes with R(g):
T R(g)f(x) =
_
G
K(x, y)R(g)f(y)dy =
_
G
(xy
1
)f(yg)dy
=
_
G
(xgy
1
)f(y)dy =
_
G
K(xg, y)f(y)dy
= R(g)(Tf)(x).
In the third equation we exploited the invariance of the measure under the right
translation y yg
1
.
Since R(g) is unitary, also the adjoint T
of T commutes with R(g):

T
R(g) = T
R(g
1
)
= (R(g
1
) T)
= (T R(g
1
))
= R(g) T
.
Thus, the self-adjoint compact operator T
T commutes with R(g). The Spectral

Theorem for compact operators yields a direct sum decomposition of L
2
(G):
L
2
(G) = ker(T
T)
_
,=0
E
_
where all the eigenspaces E
are nite-dimensional. They are also R-invariant,

for if f E
then
T
T(R(g)f) = R(g)(T
T)f = R(g)(f) = (R(g)f) (1.8)

i.e. R(g)f E
. Actually M(G) is R-invariant: all functions are of the form
n
i=1
a
i
i
(x)
i
,
i
and since
R(g)f(x) = f(xg) =
n
i=1
a
i
i
(x)(
i
(g)
i
),
i
we see that R(g)f M(G). But then also M(G)
is invariant. If P : L
2
(G)
M(G)
denotes the orthogonal projection, then by Lemma 1.8, P commutes

with R(g), and a calculation like (1.8) reveals that PE
are all R-invariant sub-

spaces of M(G)
. These are very good candidates to the subspace we wanted:

they are nite-dimensional and R-invariant, so we can restrict R to a represen-
tation on these. We just need to verify that at least one of them is nonzero. So
assume that PE
are all 0. This means by denition of P that
M(G)
and hence that M(G)
= ker T
T ker T, where the last inclu-

sion follows since f ker T
T implies 0 = T
Tf, f = Tf, Tf, i.e. Tf = 0.

But applied to the f
0
M(G)
we picked at the beginning, we have

Tf
0
(e) =
_
G
(ey
1
)f
0
(y)dy =
_
G
(y)f
0
(y)dy = , f
0
, = 0,
and as Tf
0
is continuous, Tf
0
,= 0 as an L
2
function. Thus, we must have at
least one for which PE
,= 0. If R restricted to this space is not irreducible, it

contains a nontrivial subspace on which it is. Thus, we have proved the result.
What we actually have shown in the course of the proof is that we for each
nonzero f can nd a nite-dimensional subspace U L
2
(G) which is R-invariant
and, restricted to which, R is irreducible. We can show exactly the same thing
for the left regular representation L, all we need to alter is the denition of K,
which should be K(x, y) = (x
1
y). This observation will come in useful now,
when we prove the promised generalization of Corollary 1.21:
Theorem 1.35 (Peter-Weyl II). Let (, H) be any (possibly innite-dimen-
sional ) representation of a compact group G on a Hilbert space H. Then

=
i
where
i
is a nite-dimensional irreducible representation of G, i.e. is
completely reducible.
Proof. By virtue of Theorem 1.20 we can choose a new inner product on H
turning into a unitary representation.
Then we consider the set of collections of nite-dimensional invariant sub-
spaces of H restricted to which is irreducible, i.e. an element (U
i
)
iI
in is
a collection of subspaces of H satisfying the mentioned properties. We equip
with the ordering dened by (U
i
)
iI
(U
j
)
jJ
if
i
U
i

j
U
j
. It is easily
seen that (, ) is inductively ordered, hence Zorns Lemma yields a maximal
element (V
i
)
iI
. To show the desired conclusion, namely that
H =
iI
V
i
,
we assume that W := (
V
i
)
,= 0. We have a contradiction if we in W can nd

a nite-dimensional -invariant subspace on which is irreducible, so thats our
goal.
First we remark that W is -invariant since its the orthogonal complement
to an invariant subspace, thus we can restrict to a representation on W. Now,
we will dene an intertwiner T : W L
2
(G) between [
W
and the left regular
representation L. Fix a unit vector x
0
H and dene (Ty)(g) = y, (g)x
0
.
Ty : G C is clearly continuous, and since Tx
0
(e) = |x
0
| , = 0, Tx
0
is
nonzero in L
2
(G), hence T is nonzero as a linear map. T is continuous, as the
Cauchy-Schwartz inequality and unitarity of (g) give
[Ty(g)[ = [y, (g)x
0
[ |y||x
0
|
that is |T| |x
0
|. T is an intertwiner:
(T (h))y(g) = (h)y, (g)x
0
= y, (h
1
g)x
0
= L(h) (Ty)(g).
The adjoint T
: L
2
(G) W (which is nonzero, as T is) is an intertwiner as
well, for taking the adjoint of the above equation yields (h)
= T
L(h)
for all h. Using unitarity we get (h

1
) T
= T
L(h
1
), i.e. also T
is an
intertwiner.
As T
is nonzero, there is an f
0
L
2
(G) such that T
f
0
,= 0. But by the
remark following the proof of the rst Peter-Weyl Theorem we can nd a non-
trivial nite-dimensional L-invariant subspace U L
2
(G) containing f
0
. Then
T
U W is nite-dimensional, nontrivial (it contains T
f
0
) and -invariant,
for if T
f T
U, then (h) T
f = T
L(h)f T
U. Inside T
U we can now
nd a subspace on which is irreducible, hence the contradiction.
An immediate corollary of this is:
Corollary 1.36. An irreducible representation of a compact group is automat-
ically nite-dimensional.
1.5 The Peter-Weyl Theorem 27
In particular the second Peter-Weyl Theorem says that the left regular rep-
resentation is completely reducible. In many textbooks this is the statement of
the Peter-Weyl Theorem. The proof of this is not much dierent from the proof
we gave for the rst version of the Peter-Weyl Theorem, and from this it would
also be possible to derive our second version of the Peter-Weyl Theorem. I chose
the version with matrix coecients since it can be used immediately to provide
elegant proofs of some results in Fourier theory, which we now discuss.
Theorem 1.37. Let G be a compact group. The set of irreducible characters
constitute an orthonormal basis for the Hilbert space L
2
(G, class). In particular
every square integrable class function f on G can be written
f =
G
f,
,
the convergence being L
2
-convergence.
Proof. Let P
: L
2
(G) C(G)
denote the orthogonal projection onto

C(G)
. It is not hard to see that P
maps class functions to class functions, hence

P
(L
2
(G, class)) C(G)
C(G, class), the last space being the 1-dimensional

C
by Lemma 1.27. Hence the space

M(G, class) := M(G) C(G, class) =
G
C(G)
C(G, class)
has as orthonormal basis the set of irreducible characters of G. To see that the
characters also form an orthonormal basis for the Hilbert space L
2
(G, class)
assume that there exists an f L
2
(G, class) which is orthogonal to all the
characters. Then since P
f is just a scalar multiple of
we see
P
f = P
f,
= f,
= 0
where in the third equality we exploited self-adjointness of the projection P
.
Thus we must have f M(G)
which by Peter-Weyl I implies f = 0.

Specializing to the circle group yields the existence of Fourier series. First
of all, since is abelian, all functions dened on it are class functions, and
functions on are nothing but functions on 1 with periodicity 1. Specializing
the above theorem to this case then states that the irreducible characters e
2inx
constitute an orthonormal basis for L
2
(, class) and that we have an expansion
of any such square integrable class function
f =
nZ
c
n
(f)e
2inx
(1.9)
where c
n
is the nth Fourier coecient
c
n
(f) = f,
n
=
_
1
0
f(x)e
2inx
dx.
Its important to stress that the convergence in (1.9) is only L
2
-convergence. If
we put some restrictions to f such as dierentiability or continuous dierentia-
bility we can achieve pointwise or uniform convergence of the series. We will not
travel further into this realm of harmonic analysis.
Chapter 2
Structure Theory for Lie
Algebras
2.1 Basic Notions
Although we succeeded in Chapter 1 to prove some fairly strong results, we
must realize that it is limited how much we can say about topological groups,
compact or not. For instance the Peter-Weyl Theorem tells us that every rep-
resentation of a compact group is completely reducible, but if we dont know
the irreducible representations then whats the use? Therefore we change our
focus to Lie groups. The central dierence, when regarding Lie groups, is of
course that we have their Lie algebras at our disposal. Often these are much
easier to handle than the groups themselves, while at the same time saying
quite a lot about the group. Therefore we need to study Lie algebras and their
representation theory.
In this section we focus solely on Lie algebras, developing the tools necessary
for the representation theory of the later chapters. We will only consider Lie
algebras over the elds 1 and C (commonly denoted K) although many of the
results in this chapter carry over to arbitrary (possibly algebraically closed)
elds of characteristic 0.
Denition 2.1 (Lie Algebra). A Lie algebra g over K is a K-vector space g
equipped with a bilinear map [ , ] : g g g satisfying
1) [X, Y ] = [Y, X] (antisymmetry)
2) [[X, Y ], Z] + [[Y, Z], X] + [[Z, X], Y ] = 0 (Jacobi identity).
A Lie subalgebra h of g is a subspace of g which is closed under the bracket, i.e.
for which [h, h] h. A Lie subalgebra h for which [h, g] h is called an ideal .
In this thesis all Lie algebras will be nite-dimensional unless otherwise spec-
ied.
Example 2.2. The rst examples of Lie algebras are algebras of matrices. By
gl(n, 1) and gl(n, C) we denote the set of real resp. complex n n matrices
equipped with the commutator bracket. It is trivial to verify that these are
indeed Lie algebras. The list below contains the denition of some of the classical
Lie algebras. They are all subalgebras of the two Lie algebras just mentioned.
It is a matter of routine calculations to verify that these examples are indeed
29
30 Chapter 2 Structure Theory for Lie Algebras
closed under the the commutator bracket.
sl(n, 1) = X gl(n, 1) [ Tr X = 0
sl(n, C) = X gl(n, C) [ Tr X = 0
so(n) = X gl(n, 1) [ X +X
t
= 0
so(m, n) = X, gl(m+n, 1) [ X
t
I
m,n
+I
m,n
X = 0
so(n, C) = X gl(n, C) [ X +X
t
= 0
u(n) = X gl(n, C) [ X +X
= 0
u(m, n) = X gl(m+n, C) [ X
I
m,n
+I
m,n
X = 0
su(n) = X gl(n, C) [ X +X
= 0, Tr X = 0
su(m, n) = X gl(m+n, C) [ X
I
m,n
+I
m,n
X = 0, Tr X = 0
where I
m,n
is the block-diagonal matrix whose rst mm block is the identity
and the last n n block is minus the identity.
Another interesting example is the endomorphism algebra End
K
(V ) for some
K-vector space V , nite-dimensional or not. Equipped with the commutator
bracket [A, B] = ABBA this becomes a Lie algebra over K, as one can check.
To emphasize the Lie algebra structure of this, it is sometimes denoted gl(V ).
We stick to End(V ).
We always have the trivial ideals in g, namely 0 and g itself. If g is a Lie
algebra and h is an ideal in g, then we can form the quotient algebra g/h in the
following way: The underlying vector space is the vector space g/h and this we
equip with the bracket
[X + h, Y + h] = [X, Y ] + h.
Using the ideal-property it is easily checked that this is indeed well-dened and
satises the properties of a Lie algebra.
Denition 2.3 (Lie Algebra Homomorphism). Let g and g
t
be Lie algebras
over K. A K-linear map : g g
t
is called a Lie algebra homomorphism if it
satises [(X), (Y )] = [X, Y ] for all X, Y g. If is bijective it is called a
Lie algebra isomorphism.
An example of a Lie algebra homomorphism is the canonical map : g g/h
mapping X to X+h. It is easy to see that the image of a Lie algebra homomor-
phism is a Lie subalgebra of g
t
and that the kernel of a homomorphism is an
ideal in g. Another interesting example is the so-called adjoint representation
ad : g End(V ) given by ad(X)Y = [X, Y ]. We see that ad(X) is linear,
hence an endomorphism, and that the map X ad(X) is linear. By virtue
of the Jacobi identity it respects the bracket operation and is thus ad is a Lie
algebra homomorphism.
In analogy with vector spaces and rings we have the following
Proposition 2.4. Let : g g
t
be a Lie algebra homomorphism and h g an
ideal which contains ker , then there exists a unique Lie algebra homomorphism
: g/h g
t
such that = . In the case that h = ker and g
t
= im the
induced map is an isomorphism.
If h and k are ideals in g then there exists a natural isomorphism (h+k)/k

h/(h k).
Denition 2.5 (Centralizer). Finally, for any element X g we dene the
centralizer C(X) of X to be the set of elements in g which commute with X. Let
h be any subalgebra of g. The centralizer C(h) of h is the set of all elements of
2.1 Basic Notions 31
g that commute with all elements of h. The centralizer of g is called the center
and is denoted :(g).
For a subalgebra h of g we dene the normalizer N(h) of h to be all elements
X g for which [X, h] h.
We immediately see that the centralizer of X is just ker ad(X), hence C(X)
is an ideal. Furthermore we see that
C(h) =
Xh
C(X)
and that :(g) = ker ad. Hence also the center is an ideal. Finally, a subalgebra
of g is an ideal if and only if its normalizer is g.
Now consider the so-called derived algebra: Tg := [g, g] which clearly is an
ideal. g is called abelian if Tg = 0, i.e. if [X, Y ] = 0 for all X, Y g. Every
1-dimensional Lie algebra is abelian by antisymmetry of the bracket.
Denition 2.6 (Simple Lie Algebra). A nontrivial Lie algebra is called
indecomposable if the only ideals are the trivial ones: g and 0. A nontrivial Lie
algebra is called simple if it is indecomposable and Tg ,= 0.
Any 1-dimensional Lie algebra is indecomposable and as the next proposition
shows, the requirement Tg ,= 0 is just to get rid of these trivial examples:
Proposition 2.7. A Lie algebra is simple if and only if it is indecomposable
and dimg 2.
Proof. If g is simple then it is not abelian, hence we must have dimg 2.
Conversely, assume that g is indecomposable and dimg 2. As Tg is an ideal
we can only have Tg = 0 or Tg = g. In the rst case, g is abelian and hence all
subspaces are ideals, and since dimg 2, nontrivial ideals exist, contradicting
indecomposability. Therefore Tg = g ,= 0.
Now, lets consider the following sequence of ideals T
1
g := Tg, T
2
g :=
[Tg, Tg], . . . , T
n
g := [T
n1
g, T
n1
g], the so-called derived series. Obviously we
have T
m+n
g = T
m
(T
n
g). To see that they are really ideals we use induction:
We have already seen that T
1
g is an ideal, so assume that T
n1
g is an ideal.
Let X, X
t
T
n1
g and let Y g be arbitrary. Then by the Jacobi identity
[[X, X
t
], Y ] = [[X
t
, Y ], X] [[Y, X], X
t
].
Since T
n1
g is an ideal, [X
t
, Y ], [Y, X] T
n1
g showing that [[X, X
t
], Y ]
T
n
g.
Denition 2.8 (Solvable Lie Algebra). A Lie algebra is called solvable if
there exists an N such that T
N
g = 0.
Abelian Lie algebras are solvable, since we can take N = 1. On the other
hand, simple Lie algebras are denitely not solvable, for we showed in the proof
of Proposition 2.7 that Tg = g which implies that T
n
g = g for all n.
Proposition 2.9. Let g be a Lie algebra.
1) If g is solvable, then so are all subalgebras of g.
2) If g is solvable and : g g
t
is a Lie algebra homomorphism, then im
is solvable.
3) If h g is a solvable ideal so that g/h is solvable, then g is solvable.
4) If h and k are solvable ideals of g, then so is h + k.
Proof. 1) It should be clear that Th Tg. Hence, by induction, D
i
h T
i
g
and since T
N
g = 0 for some N, then D
N
h = 0 as well.
2) Since is a Lie algebra homomorphism, we have T((g)) = (Tg), and
again by induction T
i
((g)) = (T
i
g). Thus, T
N
g = 0 implies T
N
((g)) = 0.
3) Assume there is an N for which T
N
(g/h) = 0 and consider the canonical
map : g g/h. Like above we have T
i
(g/h) = T
i
((g)) = (T
i
g). Thus,
since T
N
(g/h) = 0, we have (T
N
g) = 0 i.e. T
N
g h. But h was also solvable,
so we can nd an M for which T
M
h = 0. Then
T
M+N
g = T
M
(T
N
g) T
M
h = 0
i.e. g is solvable.
4) By 3) of this proposition it is enough to prove that (h+k)/k is solvable. By
Proposition 2.4 there exists an isomorphism (h+k)/k

h/(hk), and the right
hand side is solvable since it is the image of the canonical map h h/(hk).
The last point of this proposition yields the existence of a maximal solvable
ideal in g, namely if h and k are solvable ideals, then h + k will be a solvable
ideal containing both. Thus the sum of all solvable ideals is a solvable ideal. This
works since the Lie algebra is nite-dimensional. By construction, it is unique.
Denition 2.10 (Radical). The maximal solvable ideal, the existence of which
we have just veried, is called the radical of g and is denoted Rad g.
A Lie algebra g is called semisimple if Rad g = 0.
Since all solvable ideals are contained in Rad g another way of formulating
semisimplicity would be to say that it has no nonzero solvable ideals. In this
sense, semisimple Lie algebras are as far as possible from being solvable.
In the next section we prove some equivalent conditions for semisimplicity.
Proposition 2.11. Semisimple Lie algebras have trivial centers.
Proof. The center is an abelian, hence solvable, ideal, and is therefore trivial
by denition.
We now consider a concept closely related to solvability. Again we consider
a sequence of ideals: g
0
:= g, g
1
:= Tg, g
2
:= [g, g
1
], . . . , g
n
:= [g, g
n1
]. It
shouldnt be too hard to see that T
i
g g
i
.
Denition 2.12 (Nilpotent Lie Algebra). A Lie algebra g is called nilpotent
if there exists an N such that g
N
= 0.
Since T
i
g g
i
nilpotency of g implies solvability of g. The converse statement
is not true in general. So schematically:
abelian nilpotent solvable
in other words, solvability and nilpotency are in some sense generalizations of
being abelian.
Here is a proposition analogous to Proposition 2.9
Proposition 2.13. Let g be a Lie algebra.
1) If g is nilpotent, then so are all its subalgebras.
2) If g is nilpotent and : g g
t
is a Lie algebra homomorphism, then
im is nilpotent.
2.1 Basic Notions 33
3) If g/:(g) is nilpotent, then g is nilpotent.
4) If g is nilpotent, then :(g) ,= 0.
Proof. 1) In analogy with the proof of Proposition 2.9 a small induction ar-
gument show that if h g is a subalgebra, then h
i
g
i
. Thus, g
N
= 0 implies
h
N
= 0.
2) We have already seen that (g)
1
= (Tg). Furthermore
(g)
2
= [(g), (g)
1
] = [(g), (Tg)] = ([g, Tg]) = (g
2
)
and by induction we get (g)
i
= (g
i
). Hence nilpotency of g implies nilpotency
of (g).
3) Letting : g g/:(g) denote the canonical homomorphism, we see that
(g/:(g))
i
= ((g))
i
= (g
i
) = g
i
/:(g).
Thus, if (g/:(g))
N
= 0 then g
N
:(g). But then g
N+1
= [g, g
N
] [g, :(g)] =
0, hence g is nilpotent.
4) As g is nilpotent there is a smallest n such that g
n
,= 0 and g
n+1
= 0. This
means that [g, g
n
] = 0 i.e. everything in g
n
commutes with all elements of g.
Thus, 0 ,= g
n
:(g).
Denition 2.14. An element X g is called ad-nilpotent if ad(X) is a nilpotent
linear map, i.e. if there exists an N such that ad(X)
N
= 0.
If the Lie algebra is a subalgebra of an algebra of endomorphisms (for in-
stance End(V )), it makes sense to ask if the elements themselves are nilpotent.
In this case nilpotency and ad-nilpotency of an element X need not be the same.
For instance in End(V ) we have the identity I, which is obviously not nilpo-
tent. However, ad(I) = 0, and thus I is ad-nilpotent. The reverse implication,
however, is true:
Lemma 2.15. Let g be a Lie algebra of endomorphisms of some vector space.
If X g is nilpotent, then it is ad-nilpotent.
Proof. We associate to A g two linear maps
A
,
A
: End(V ) End(V )
by
A
(B) = AB and
A
(B) = BA. Its easy to see that they commute, and
that ad(A) =
A
A
.
As A is nilpotent,
A
and
A
are also nilpotent, so we can nd an N for which
N
A
=
N
A
= 0. Since they commute, we can use the binomial formula and get
ad(A)
2N
= (
A
A
)
2N
=
2N
j=0
(1)
j
_
2N
j
_
2Nj
A

j
A
which is zero since all terms contain either
A
or
A
to a power greater than
N.
An equivalent formulation of nilpotency of a Lie algebra is that there exists an
N such that ad(X
1
) ad(X
N
)Y = 0 for all X
1
, . . . , X
N
, Y g. In particular, if
g is nilpotent, then there exists an N such that ad(X)
N
= 0 for all X g, i.e. X
is ad-nilpotent. Thus, for a nilpotent Lie algebra g, all elements are ad-nilpotent.
That the converse is actually true is the statement of Engels Theorem, which
will be a corollary to the following theorem.
Theorem 2.16. Let V be a nite-dimensional vector space and g End(V ) be
a subalgebra consisting of nilpotent linear endomorphisms. Then there exists a
nonzero v V which is an eigenvector for all A End(V ).
Proof. We will prove this by induction over the dimension of g. First, assume
dimg = 1. Then g = KA for some nonzero A g. As A is nilpotent there is a
smallest N such that A
N
,= 0 and A
N+1
= 0, i.e. we can nd a vector w V
with A
N
w ,= 0 and A(A
N
w) = A
N+1
w = 0. Since all elements of g are scalar
multiples of A the vector A
N
w will qualify.
Now assuming that the theorem holds for all Lie algebras of dimension strictly
less than n, we should prove that it holds for n-dimensional algebras as well.
The algebra g consists of nilpotent endomorphisms on V , hence by the previous
lemma, all elements are ad-nilpotent. Consider a subalgebra h ,= g of g which
thus also consists of ad-nilpotent elements. For A h we have that ad(A)h h
since h as a subalgebra is closed under brackets. We can form the vector space
g/h and dene a linear map ad(A) : g/h g/h by
ad(A)(B + h) = (ad(A)B) + h.
This is well dened for if B + h = B
t
+ h, then B B
t
h and therefore
ad(A)(B
t
+ h) = ad(A)B
t
+ h = ad(A)B
t
+ ad(A)(B B
t
) + h = ad(A)B + h
= ad(A)(B + h).
This map is again nilpotent since ad(A)
N
(B + h) = (ad(A)
N
B) + h = h = [0].
So, the situation now is that we have a subalgebra ad(h) of End(g/h) with
dimad(h) dimh < dimg = n. Our induction hypothesis then yields an ele-
ment 0 ,= [B
0
] = B
0
+h g/h on which ad(A) is zero for all A h. This means
that [A, B
0
] h for all h, i.e. the normalizer N(h) of h is strictly larger than h.
Now assume that h is any maximal subalgebra h ,= g. Then since N(h) is
a strictly larger subalgebra we must have N(h) = g and consequently h is an
ideal. Then g/h is a Lie algebra with canonical Lie algebra homomorphism
: g g/h and g/h must have dimension 1 for, assuming otherwise, we could
nd a 1-dimensional subalgebra k ,= g/h in g/h, and then
1
(k) ,= g would be
a subalgebra strictly larger that h. This is a contradiction, hence dimg/h = 1
and g
= h KA
0
for some nonzero A
0
g h.
So far, so good. Now we come to the real proof of the existence of the nonzero
vector v V . h was an ideal of dimension n 1 hence the induction hypothesis
assures that the subspace W := v V [ B h : Bv = 0 is nonempty. We
will show that each linear map A g (which maps V V ) can be restricted
to a map W W and that it as such a map is still nilpotent. This will, in
particular, hold for A
0
which by nilpotency will have the eigenvalue 0 and hence
a nonzero eigenvector v W associated to the eigenvalue 0. This will be the
desired vector, for all linear maps in g can according to the decomposition above
be written as B +A
0
for some B h, and Bv = 0 since v was chosen to be in
W.
Thus, to nish the proof we only need to see that W is invariant. So let A g
be any map. Since h is an ideal, [A, h] h and hence for w W
B(Aw) = A(Bw) [A, B]w = 0
for all B h. This shows that Aw W and hence that W is invariant. A
restriction of a nilpotent map is clearly nilpotent. This completes the proof.
From this we can prove
Corollary 2.17 (Engels Theorem). A Lie algebra is nilpotent if and only if
all its elements are ad-nilpotent.
2.2 Semisimple Lie Algebras 35
Proof. We have already showed the only if part. To show the if part we
again invoke induction over the dimension of g. If dimg = 1 then g is abelian,
hence nilpotent.
Now set n = dimg and assume that the result holds for all Lie algebras with
dimension strictly less than n. All the elements of g are ad-nilpotent, hence
ad(g) is a subalgebra of End(g) consisting of nilpotent elements and the previous
theorem yields an element 0 ,= X g for which ad(Y )(X) = 0 for all Y g i.e.
X is contained in the center :(g) which is therefore a nonzero ideal and g/:(g)
is a Lie algebra whose dimension strictly less than n. Furthermore all elements
of g/:(g) are ad-nilpotent, since by denition of the quotient bracket
ad([A])[B] = ad(A)B +:(g)
we have that ad(A)
N
= 0 implies ad([A])
N
= 0. Thus g/:(g) consists solely
of ad-nilpotent elements. Consequently the induction hypothesis assures that
g/:(g) is nilpotent. Then by Proposition 2.13 g is nilpotent.
2.2 Semisimple Lie Algebras
The primary goal of this section is to reach some equivalent formulations of
semisimplicity. Our approach to this will be via the so-called Cartan Criterion
for solvability which we will prove shortly.
First we need a quite powerful result from linear algebra regarding advanced
diagonalization:
Theorem 2.18 (SN-Decomposition). Let V be a nite-dimensional vector
space over K and let A End(V ). Then there exist unique commuting linear
maps S, N End(V ), S being diagonalizable and N being nilpotent, satisfying
A = S+N. This is called the SN-decomposition In fact S and N can be realized
as polynomials in A without constant terms.
Furthermore, if A = S+N is the SN-decomposition of A, then ad(S)+ad(N)
is the SN-decomposition of ad(A).
We will not prove this
1
.
Cartans Criterion gives a sucient condition for solvability based on the trace
of certain matrices. Therefore the following lemma is necessary.
Lemma 2.19. Let V be a nite-dimensional vector space, W
1
and W
2
be sub-
spaces of End(V ) and dene M := B End(V ) [ ad(A)W
1
W
2
. If A M
satises Tr(AB) = 0 for all B M then A is nilpotent.
Proof. Let A M satisfy the required condition, and consider the SN-decom-
position of A = S + N. We are done if we can show that S = 0. Well, S is
diagonalizable, so we can nd a basis e
1
, . . . , e
n
for V in which S has the form
diag(a
1
, . . . , a
n
). We will show that all these eigenvalues are 0, and we do so in
a curious way: We dene E := span
Q
a
1
, . . . , a
n
K to be the subspace of K
over the rationals spanned by the eigenvalues. If we can show that this space, or
equivalently its dual space E
, consisting of -linear maps E , is 0, then

we are done.
So, let E
be arbitrary. The basis we chose for V readily gives us a

basis for End(V ), consisting of E
ij
where E
ij
is the linear map determined by
E
ij
e
j
= e
i
and E
ij
e
k
= 0 for k ,= j. Then we see
(ad(S)E
ij
)e
j
= [S, E
ij
]e
j
= SE
ij
e
j
E
ij
Se
j
= Se
i
a
j
E
ij
e
j
= (a
i
a
j
)e
i
1
For a proof the reader is referred to for instance [5] Section 4.3.
while [S, E
ij
]e
k
= 0 for k ,= j i.e. ad(S)E
ij
= (a
i
a
j
)E
ij
.
Now, let B End(V ) denote the linear map which in the basis e
1
, . . . , e
n
is
diag((a
1
), . . . , (a
n
)). As with S we have that ad(B)E
ij
= ((a
i
) (a
j
))E
ij
.
There exists a polynomial p =
N
n=1
c
n
X
n
without constant term which maps
a
i
a
j
to (a
i
a
j
) = (a
i
) (a
j
) (its a matter of solving some equations
to nd the coecients c
n
). Then we have
p(ad S)E
ij
= c
n
(ad S)
n
E
ij
+ +c
1
(ad S)E
ij
= c
n
(a
i
a
j
)
n
E
ij
+ +c
1
(a
i
a
j
)E
ij
= p(a
i
a
j
)E
ij
= ((a
i
) (a
j
))E
ij
which says that p(ad S) = ad B. A statement in the SN-decomposition was that
ad S, being the diagonalizable part of ad A, is itself a polynomial expression
in ad A without constant term, which implies that ad B is a polynomial in
ad A without constant term. Since A M we have that ad(A)W
1
W
2
, and
since ad(B) was a polynomial expression in ad(A) then also ad(B)W
1
W
2
,
i.e. B M, and therefore by assumption Tr(AB) = 0. The trace of AB is
the sum

n
i=1
a
i
(a
i
) and applying to the equation Tr(AB) = 0 we get
n
i=1
(a
i
)
2
= 0, i.e. (a
i
) = 0 ((a
i
) are rationals hence (a
i
)
2
0). Therefore
we must have = 0 which was what we wanted.
Theorem 2.20 (Cartans Criterion). Let V be a nite-dimensional vector
space and g End(V ) a subalgebra. If Tr(AB) = 0 for all A g and all B Tg
then g is solvable.
Proof. As T
n
g = T
n1
(Tg) (Tg)
n1
we see that g will be solvable if Tg is
nilpotent. To show that Tg is nilpotent we invoke Engels Theorem and Lemma
2.15 which combined say that Tg is nilpotent if all X Tg are nilpotent.
To this end we use the preceding lemma with W
1
= g and W
2
= Tg and
M = B End(V ) [ [B, g] Tg. Notice that g M. The reverse inclusion
need not hold.
Now, let A Tg be arbitrary, we need to show that it is nilpotent, and by
virtue of the previous lemma it suces to verify that Tr(AB) = 0 for all B M.
A is of the form [X, Y ] for X, Y g and we have, in general that
Tr([X, Y ]B) = Tr(XY B) Tr(Y XB) = Tr(Y BX) Tr(BY X)
= Tr([Y, B]X) = Tr(X[Y, B]). (2.1)
Since B M and Y g we have by construction of M that [Y, B] Tg. But
then by assumption in the theorem we have that Tr(AB) = Tr([X, Y ]B) =
0.
With this powerful tool we can prove the promised equivalent conditions for
a Lie algebra to be semisimple. One of them involves the so-called Killing form:
Denition 2.21 (Killing Form). By the Killing form for a Lie algebra g over
K we understand the bilinear form B : g g K given by
B(X, Y ) = Tr(ad(X) ad(Y )).
Proposition 2.22. The Killing form is a symmetric bilinear form satisfying
B([X, Y ], Z) = B(X, [Y, Z]). (2.2)
Furthermore, if is any Lie algebra automorphism of g then B((X), (Y )) =
B(X, Y ).
Proof. B is obviously bilinear, and symmetry is a consequence of the property
of the trace: Tr(AB) = Tr(BA). Eq. (2.2) follows from (2.1). If : g
g is a Lie algebra automorphism, then another way of writing the equation
[(X), (Y )] = ([X, Y ]) is ad((X)) = ad(X). Therefore
B((X), (Y )) = Tr( ad(X) ad(Y )
1
) = Tr(ad(X) ad(Y ))
= B(X, Y ).
Calculating the Killing form directly from the denition is immensely compli-
cated. Fortunately, for some of the classical Lie algebras we have a much simpler
formula:
B(X, Y ) =
_
_
2(n + 1) Tr(XY ), for X, Y sl(n + 1, K), sp(2n, K)
(2n 1) Tr(XY ), for X, Y so(2n + 1, K)
2(n 1) Tr(XY ), for X, Y so(2n, K).
(2.3)
Lemma 2.23. If g is a Lie algebra with Killing form B and h g is an ideal,
then B[
hh
is the Killing form of h.
Proof. First a general remark: If : V V is a linear map, and W
V is a subspace for which im W, then Tr = Tr([
W
): Namely, pick a
basis e
1
, . . . , e
k
for W and extend it to a basis e
1
, . . . , e
k
, . . . , e
n
for V .
Let
1
, . . . ,
n
denote the corresponding dual basis. As (v) W we have
k+i
((v)) = 0 and hence
Tr =
n
i=1
i
((e
i
)) =
k
i=1
i
((e
i
)) = Tr([
W
).
Now, let X, Y h, then as h is an ideal: ad(X)g h and ad(Y )g h, which
means that the image of ad(X) ad(Y ) lies inside h. It should be obvious that
the adjoint representation of h is just ad(X)[
h
for X h. Therefore
B
h
(X, Y ) = Tr(ad(X)[
h
ad(Y )[
h
) = Tr((ad(X) ad(Y ))[
h
)
= B[
hh
(X, Y ).
Theorem 2.24. If g is a Lie algebra, then the following are equivalent:
1) g is semisimple i.e. Rad g = 0.
2) g has no nonzero abelian ideals.
3) The Killing form B of g is non-degenerate.
4) g is a direct sum of simple Lie algebras: g = g
1
g
n
.
Proof. We rst prove that 1 and 2 are equivalent. If g is semisimple, then g
has no nonzero solvable ideals and since abelian ideals are solvable, no nonzero
abelian ideals either. Conversely, if Rad g ,= 0, then, since Rad g is solvable,
there is a smallest N for which T
N
(Rad g) ,= 0 and T
N+1
(Rad g) = 0. Then
T
N
(Rad g) and is an abelian ideal hence a solvable ideal. So by contraposition,
if no solvable ideals exist, then g is semisimple.
Now we show that 1 implies 3. We consider the so-called radical of the Killing
form B namely the subspace
h := X g [ Y g : B(X, Y ) = 0.
h is an ideal for if X h and Y g, then for all Z g:
B([X, Y ], Z) = B(X, [Y, Z]) = 0
i.e. [X, Y ] h. Obviously B is non-degenerate if and only if h = 0.
Now we assume that Rad g = 0 and want to show that h = 0. We can do
this by showing that h is solvable for then h Rad g. First we use the Cartan
Criterion on the Lie algebra ad(h) showing that this is solvable: By denition
of h we have that 0 = B(X, Y ) = Tr(ad(X) ad(Y )) for all X h and Y g,
in particular it holds for all X Th. In other words we have Tr(AB) = 0 for
all A ad(Th) = T(ad h) and all B ad h. Hence the Cartan Criterion tells us
that ad h is solvable, i.e. 0 = T
N
(ad h) = ad(T
N
h). This says that T
N
h :(g)
implying that T
N+1
h = 0. Thus, h is solvable and consequently equals 0.
Then we prove 3 implies 2. Assume that h = 0, and assume k to be an abelian
ideal and let X k and Y g. Since the adjoint representations maps according
to (exploiting the ideal property of k)
g
ad(Y )
g
ad(X)
k
ad(Y )
k
ad(X)
Tk = 0
we have (ad(X) ad(Y ))
2
= 0 that is ad(X) ad(Y ) is nilpotent. Since nilpotent
matrices have zero trace we see that 0 = Tr(ad(X) ad(Y )) = B(X, Y ). This
implies X h, i.e. k h and thus the desired conclusion.
We then proceed to show that 1 implies 4. Suppose g is semisimple, and
let h g be any ideal. We consider its orthogonal complement w.r.t. B:
h
:= X g[ Y h : B(X, Y ) = 0. This is again an ideal in g for if X h
and Y g, then for all Z g we have [Y, Z] h and hence

B([X, Y ], Z) = B(X, [Y, Z]) = 0
saying that [X, Y ] h
. To show that we have a decomposition g = h h
we
need to show that the ideal h h
is zero. We can do this by showing that it

is solvable for then semisimplicity forces it to be zero. By some remarks earlier
in this proof, solvability of h h
would be a consequence of ad(h h
) being
solvable. To show that ad(hh
) is solvable we invoke the Cartan Criterion: For

X T(hh
) hh
and Y hh
we have Tr(ad(X) ad(Y )) = B(X, Y ) = 0

since, in particular, X h and Y h
. Thus, the Cartan Criterion renders

solvability of ad(hh
) implying hh
= 0. Hence, hh
= 0 and g = hh
.
After these preliminary remarks we proceed via induction over the dimension
of g. If dimg = 2, then g is simple, for any nontrivial ideal in g would have
to be 1-dimensional, hence abelian, and such do not exist. Assume now that
dimg = n and that the result is true for Lie algebras of dimension strictly less
than n. Suppose that g
1
is a minimal nonzero ideal in g then g
1
is simple since
dimg
1
2 and since any nontrivial ideal in g
1
would be an ideal in g properly
contained in g
1
contradicting minimality. Then we have g = g
1
g
1
with g
1
semisimple, for if k is any abelian ideal in g
1
then it is an abelian ideal in g and
these do not exist. Then by the induction hypothesis we have g
1
= g
2
g
n
, a
sum of simple Lie algebras, hence g = g
1
g
2
g
n
, a sum of simple algebras.
Finally we show that 4 implies 2. So consider g := g
1
g
n
and let h g
be an abelian ideal. It is not hard to verify that h
i
:= h g
i
is an abelian ideal
in g
i
, thus h
i
= g
i
or h
i
= 0. As h
i
is abelian and g
i
is not, we can rule out the
rst possibility, i.e. h
i
= 0 and hence h = 0.
During the proof we saw that any ideal in a semisimple Lie algebra has a
complementary ideal. This is important enough to be stated as a separate result:
Proposition 2.25. Let g be a semisimple Lie algebra and h g an ideal. Then
h
:= X g [ Y h : B(X, Y ) = 0 is an ideal in g and g = h h
.
Another very important concept in the discussion to follow is that of com-
plexication.
Denition 2.26 (Complexication). Let V be a real vector space. By the
complexication V
C
of the vector space V we understand V
C
:= V iV which
equipped with the scalar multiplication
(a +ib)(v
1
+iv
2
) = (av
1
bv
2
) +i(av
2
+bv
1
)
becomes a complex vector space.
If g is a real Lie algebra, the complexication g
C
of g is the vector space gig
equipped with the bracket
[X
1
+iX
2
, Y
1
+iY
2
] = ([X
1
, Y
1
] [X
2
, Y
2
]) +i([X
1
, Y
2
] + [X
2
, Y
1
])
(note that this in not the usual direct sum bracket!). It is easily checked that
g
C
is a complex Lie algebra.
Other presentations of this subject dene the complexication of g by g
C
=
g
R
C, where C is considered a 2-dimensional real space. By writing C = 1i1
and use distributivity of the tensor product, this denition is equivalent to ours.
Example 2.27. The classical real Lie algebras mentioned earlier have the fol-
lowing complexications
gl(n, 1)
C

= gl(n, C)
sl(n, 1)
C

= sl(n, C)
so(n)
C

= so(n, C)
so(m, n)
C

= so(m+n, C)
u(n)
C

= gl(n, C)
u(m, n)
C

= gl(m+n, C)
su(n)
C

= sl(n, C)
su(m, n)
C

= sl(m+n, C).
Lets prove a few of them. For the rst one, pick an element X of gl(n, C)
and split it in real and imaginary parts X = X
1
+iX
2
. It is an easy exercise to
verify that the map X X
1
+iX
2
is a Lie algebra isomorphism gl(n, C)

gl(n, 1)
C
.
To prove u(n)
C

= gl(n, C), let X gl(n, C) and write it as
X =
X X
2
+i
X +X
2i
.
It is not hard to see that both
1
2
(XX
) and
1
2i
(X+X
) are skew-adjoint, i.e.

elements of u(n). Again it is a trivial calculation to show that
X
X X
2
+i
X +X
2i
is a Lie algebra isomorphism gl(n, C)

u(n)
C
. The other identities are veried
in a similar fashion.
Proposition 2.28. A Lie algebra g is semisimple if and only if g
C
is semisim-
ple.
Proof. Let B denote the Killing form of g and B
C
the Killing form of g
C
.
Our rst task is to relate them. If X
1
, . . . , X
n
is a basis for g as an 1-vector
space, then X
1
, . . . , X
n
is also a basis for g
C
as a C-vector space. Therefore
for X, Y g the linear map ad(X) ad(Y ) will have the same matrix whether it
is considered a linear map on g or g
C
. In particular their traces will be equal
which amounts to say that B(X, Y ) = B
C
(X, Y ). In other words
B
C
[
gg
= B. (2.4)
Now assume g to be semisimple, or, equivalently B to be non-degenerate. Then
B(X, Y ) = 0 for all Y g implies X = 0. To show that B
C
is non-degenerate,
let X g
C
satisfy B
C
(X, Y ) = 0 for all Y g
C
. Then it particularly holds for
all Y g. Write X = A
1
+iA
2
where A
1
, A
2
g, then by (2.4)
0 = B
C
(A
1
, Y ) +B
C
(A
2
, Y ) = B(A
1
, Y ) +iB(A
2
, Y )
for all Y g. Hence by non-degeneracy of B we have A
1
= A
2
= 0, i.e. X = 0.
Thus B
C
is non-degenerate.
Now assume B
C
to be non-degenerate and suppose B(X, Y ) = 0 for all Y g.
This particularly holds for the basis elements: B(X, X
k
) = 0 for k = 1, . . . , n.
By (2.4) we also have B
C
(X, X
k
) = 0, and since X
1
, . . . , X
n
was also a basis
for g
C
we get B
C
(X, Y ) = 0 for all Y g
C
and thus by non-degeneracy of B
C
that X = 0, i.e. B is non-degenerate.
Up till now we have talked a lot about semisimple Lie algebras and their
amazing properties. But we have not yet encountered one single example of a
semisimple Lie algebra. The rest of this section tends to remedy that. The rst
thing we do is to introduce a class of Lie algebras which contains the semisimple
ones:
Denition 2.29 (Reductive Lie Algebra). A Lie algebra g is called reductive
if for each ideal a g there is an ideal b g such that g = a b.
From Proposition 2.25 it follows that semisimple Lie algebras are reductive.
So schematically we have
simple semisimple reductive.
Note how these classes of Lie algebras are somehow opposite to the classes of
abelian, solvable or nilpotent algebras.
The next proposition characterizes the semisimple Lie algebras among the
reductive ones
Proposition 2.30. If g is reductive, then g = Tg:(g) and Tg is semisimple.
Thus a reductive Lie algebra is semisimple if and only if its center is trivial.
Proof. Let be the set of direct sums a
1
a
k
where a
1
, . . . , a
k
are
indecomposable ideals (i.e. they contain only trivial ideals). The elements of
are themselves ideals. Let a be an element of maximal dimension. As g is
reductive, there exists an ideal b such that g = a b. We want to show that
b = 0 (and hence g = a) so assume for contradiction that b ,= 0 and let
b
t
b be the smallest nonzero indecomposable ideal (which always exists, for
if b contains no proper ideals, then b is indecomposable). But then a b
t

contradicting maximality of a, and therefore g = a .
Now lets write
g = a
1
a
j
. .
g1
a
j+1
a
k
. .
g2
where a
1
, . . . , a
j
are 1-dimensional and a
j+1
, . . . , a
k
are higher dimensional and
thus simple. Therefore g
1
is abelian and g
2
is semisimple (by Theorem 2.24) and
by denition of the direct sum bracket we have
Tg = T(a
1
a
k
) = Ta
1
Ta
k
= Ta
j+1
Ta
k
= g
2
.
This shows that Tg is semisimple. We now only have to justify that g
1
equals
the center. We have g
1
:(g) for in the decomposition g = g
1
g
2
[(X, 0), (Y, Z)] = [X, Y ] + [0, Z] = 0.
Conversely, let X :(g). We decompose it X = X
1
+ + X
k
according to
the decomposition of g in indecomposable ideals. Then X
i
:(a
i
) which means
that X
i
= 0 for j > i and hence X g
1
.
The next result will help us mass-produce examples of reductive Lie algebras
Proposition 2.31. Let g be a Lie subalgebra of gl(n, 1) or gl(n, C). If g has
the property that X g implies X
g (where X
is the conjugate transpose of

X), then g is reductive.
Proof. Dene a real inner product on g by X, Y = Re Tr(XY
). This is a
genuine inner product: its symmetric:
Y, X = Re Tr(Y X
) = Re Tr(Y
) = Re Tr((XY
)
= Re Tr(XY
) = Re Tr(XY
) = X, Y ,
and its positive denite, for Tr(XX
) is nothing but the sum of the square of

the norm of the columns in X which is 0 if and only if X = 0.
Assuming a to be an ideal in g, let a
be the complementary subspace w.r.t.

the inner product just dened. Then as vector spaces it holds that g = a a
.
For this to be a Lie algebra direct sum we need a
to be an ideal. Let X a
and Y g, then for all Z a

[X, Y ], Z = Re Tr(XY Z
Y XZ
) = Re Tr(XZ
Y XY Z
)
= Re Tr(X(Y
Z)
X(ZY
) = X, [Y
, Z]
which is 0 as X a
and [Y
, Z] a since Y
g. Thus a
is an ideal.
Obviously gl(n, 1) and gl(n, C) are closed under the operation conjugation
transpose and are therefore reductive. They are not semisimple as their centers
contain the scalar matrices diag(a, . . . , a) for a 1 or a C respectively,
violating Proposition 2.11.
The Lie algebras so(n) are semisimple for n 3. Recall that so(n) is the set
of real n n matrices X for which X + X
= 0. From the denition it is clear

that if X so(n) then also X
so(n). Hence so(n) is reductive for all n. so(2)

is a 1-dimensional (hence abelian) Lie algebra and thus is not semisimple. Let
us show that so(3) is semisimple. Thanks to Proposition 2.30 this boils down to
verifying that its center is trivial. So assume
X =
_
_
0 a b
a 0 c
b c 0
_
_
to be in an element in the center of so(3). In particular it has to commute with
the two matrices
A
1
=
_
_
0 1 0
1 0 0
0 0 0
_
_
and A
2
_
_
0 0 1
0 0 0
1 0 0
_
_
.
We have
A
1
X =
_
_
a 0 c
0 a b
0 0 0
_
_
and XA
1
=
_
_
a 0 0
0 a 0
c b 0
_
_
.
As these two matrices should be equal we immediately get that b = c = 0.
Furthermore
A
2
X =
_
_
0 0 0
0 0 0
0 a 0
_
_
and XA
2
=
_
_
0 0 0
0 0 a
0 0 0
_
_
and we get a = 0. Thus X = 0, and the center is trivial. Generalizing this to
higher dimensions one can show that so(n) is semisimple for n 3. Now since
so(n, C) = so(n)
C
(cf. Example 2.27) Proposition 2.28 says that also so(n, C) is
semisimple for n 3.
The Lie algebra u(n) is reductive. These are just the n n complex matrices
satisfying X+X
= 0 and again it is clear that u(n) is closed under the operation

conjugate transpose and hence reductive. It is not semisimple, since the matrices
diag(ia, . . . , ia) for a 1 are all in the center. However the subalgebra su(n) is
semisimple for n 2 (su(1) is zero-dimensional) as can be seen by an argument
analogous to the one given above. Since its complexication is sl(n, C) this is also
semisimple for n 2. But sl(n, C) is also the complexication of sl(n, 1) which
is therefore also semisimple for n 2. By the same argument also so(m, n) for
m+n 3 and su(m, n) for m+n 2 are semisimple since their complexications
are. Wrapping up, the following Lie algebras are semisimple
sl(n, 1) for n 2
sl(n, C) for n 2
so(n) for n 3
so(m, n) for m+n 3
so(n, C) for n 3
su(n) for n 2
su(m, n) for m+n 2.
2.3 The Universal Enveloping Algebra
For a nite-dimensional vector space V we have the tensor algebra T(V ) dened
by
T(V ) =
n=0
V
n
.
From this one can form various quotients. One of the more important is the
symmetric algebra S(V ) where we mod out the ideal J generated by elements
of the form X Y Y X. The resulting algebra is commutative by con-
struction. If X
1
, . . . , X
n
is a basis for V , then one can show that the set
X
i1
1
X
in
n
[ i
1
, . . . , i
n
N
0
(we dene X
0
= 1) will be a basis for S(V ) which
is thus (unlike the exterior algebra) innite-dimensional. If we set I = (i
1
, . . . , i
k
)
we will use the short-hand notation X
I
for X
i1
X
i
k
. We dene the length of
I to be [I[ = k and write j I if j i
1
, . . . , i
k
.
Denition 2.32 (Universal Enveloping Algebra). Let g be a Lie alge-
bra. By a universal enveloping algebra of g we understand a pair (U, i) of an
associative unital algebra U and a linear map i : g U with i([X, Y ]) =
i(X)i(Y ) i(Y )i(X) satisfying that for any pair (A, ) of an associative unital
algebra A and a linear map : g A with ([X, Y ]) = (X)(Y )(Y )(X)
there is a unique algebra homomorphism : U A with = i.
2.3 The Universal Enveloping Algebra 43
In other words any linear map : g A satisfying the above condition
factorizes through U rendering the following diagram commutative
g
i
?
?
?
?
?
?
?
?
U
A
As for the symmetric algebra, multiplication in a universal algebra is written by
juxtaposition.
Proposition 2.33. Let g be a Lie algebra and J the two-sided ideal in T(g)
generated by elements of the form X Y Y X [X, Y ]. If i denotes the
restriction of the canonical map : T(g) T(g)/J to g then (T(g)/J, i) is a
universal enveloping algebra for g. It is unique up to algebra isomorphism.
Proof. Uniqueness rst. Assume that (U, i) and (
U,
i) are universal enveloping

algebras for g. Since
i : g

U is a linear map satisfying the bracket condition,
the universal property of (U, i) yields an algebra homomorphism : U

U
so that
i = i. Likewise for i : g U the universal property of (
U,
i) yields
an algebra homomorphism :

U U so that i =
i. Composing these gives

that i = i, i.e. makes the following diagram commutative
g
i
?
?
?
?
?
?
?
?
U
U
But obviously id
U
also makes the diagram commute and by uniqueness =
id
U
. Likewise one shows that = id
U
, thus U and

U are isomorphic.
To show existence we just need to verify that (T(g)/J, i) is really a universal
enveloping algebra. Well, rst of all
i([X, Y ]) = ([X, Y ]) = [X, Y ] + J
= [X, Y ] +X Y Y X [X, Y ] + J = X Y Y X + J
= (X + J) (Y + J) (Y + J) (X + J)
= (X)(Y ) (Y )(X) = i(X)i(Y ) i(Y )i(X).
Now, suppose that : g A is a linear map satisfying ([X, Y ]) =
(X)(Y ) (Y )(X). Consider the following diagram.
g

N
N
N
N
N
N
N
N
N
N
N
N
N
N
T(g)
T(g)/J
m
m
m
m
m
m
m
m
m
m
m
m
m
m
A
Since is linear it factorizes uniquely through T(g) yielding an algebra homo-
morphism
t
: T(g) A. On the generators of J we see that
t
(X Y Y X [X, Y ]) =
t
(X Y )
t
(Y X)
t
([X, Y ])
= (X)(Y ) (Y )(X) ([X, Y ]) = 0.
Thus, vanishing on J,
t
factorizes uniquely through T(g)/J by an algebra ho-
momorphism : T(g)/J A, i.e. = i. This proves existence.
If g is abelian, then the ideal J is generated by elements X Y Y X, i.e.
U(g) is just the symmetric algebra S(g).
For the tensor algebra we have a ltration if we dene
T
m
(g) :=
m
k=0
g
k
.
We can carry this over to U(g) if we dene U
m
(g) := (T
m
(g)). We see that
U
m
(g)U
n
(g) = (T
m
(g))(T
n
(g)) = (T
m
(g)T
n
(g)) (T
m+n
(g)) = U
m+n
(g).
Since we have U
m1
U
m
it makes sense to dene the vector spaces G
m
(g) =
U
m
(g)/U
m1
(g) with canonical map q
m
: U
m
(g) G
m
(g) and
G(g) :=
m=0
G
m
(g).
To shorten the notation we will just write T
m
, U
m
, G
m
and G when no confusion
is possible.
The product in U(g) denes a graded algebra structure on G. Namely, dene
a product map G
m
G
n
G
m+n
by
q
m
(v)q
n
(w) = q
m+n
(vw)
where v U
m
and w U
n
are representatives for the elements in G
m
and G
n
.
For this to make sense we used that U
m
U
n
U
m+n
as showed above. A simple
argument shows that this product is well-dened (i.e. independent of choice of
representatives). Thus, G becomes a graded algebra.
Now the composition
g
m

U
m
qm
G
m
gives a linear map
m
: g
m
G
m
which is surjective as both and q
m
are
surjective. Then also the linear map =
m
: T(g) G is surjective.
Lemma 2.34. The map : T(g) G is a surjective algebra homomorphism
that vanishes on J. Thus, it induces a surjective map : S(g) G.
Proof. Let x g
m
and y g
n
. Then
(x y) = q
m+n
((x y)) = q
m+n
((x)(y)) = q
m
((x))q
n
((y)) = (x)(y).
Hence, is an algebra homomorphism. Now, consider a generator XY Y X
for the ideal J. We have that X Y Y X g
2
and therefore (X Y
Y X) U
2
. Hence
(X Y Y X) = q
2
((X Y Y X)).
But by denition of we have (X Y Y X) = ([X, Y ]) U
1
and thus
q
2
(([X, Y ])) = 0. Hence vanishes on J.
This induced homomorphism is exceedingly interesting. The rest of this sec-
tion is devoted to prove the following theorem and some of its consequences.
Theorem 2.35 (Poincar-Birkho-Witt). The induced map : S(g) G
is an isomorphism of algebras.
For brevity we will refer to it as the PBW-Theorem. We split up the proof of
this into a series of lemmas. For the rst one let X
1
, . . . , X
n
denote a basis
for g, let X
I
[ I increasing be the associated basis for the symmetric algebra
and recall the ltration S
m
(g).
Lemma 2.36. For each m N
0
there is a unique linear map f
m
: g S
m

S(g) satisfying
1) f
m
(X
j
X
I
) = X
j
X
I
for j I and X
I
S
m
(meaning [I[ m),
2) f
m
(X
j
X
I
) X
j
X
I
S
k
when X
I
S
k
and k m,
3) f
m
(X
i
f
m
(X
j
X
I
)) = f
m
(X
j
f
m
(X
i
X
I
)) +f
m
([X
i
, X
j
] X
I
) for
X
I
S
m1
,
and f
m
restricted to g S
m1
equals f
m1
.
Proof. For the time being we assume that we have shown uniqueness. Since
S
m2
S
m1
S
m
and since f
m
satises 1) and 2) for X
I
S
m
, f
m
s
restriction to g S
m1
clearly satises 1) and 2) for X
I
S
m1
. Likewise
f
m
satises 3) for X
I
S
m2
. By uniqueness this restriction must equal f
m1
.
To show existence and uniqueness we use induction over m. For m = 0 S
0
is
spanned by 1 so we only need to dene f
0
on X
i
1. If 1) should be fullled
we can only have f
0
(X
i
1) = X
i
. Obviously 2) is also satised, and 3) is an
empty statement as S
1
= 0. Thus f
0
exists and is unique.
Now for the induction step we assume that we have a unique map f
m1
satisfying 1)-3). It is enough to dene f
m
on elements of the form X
j
X
I
with
[I[ = m (and I increasing) since the remarks at the beginning and uniqueness
of f
m1
implies that f
m
restricted to g S
m1
is f
m1
.
If j I we must have by 1) that f
m
(X
j
X
I
) = X
j
X
I
. If we dont have
j I then we write I = (i, J) where i J, i < j and [J[ = m1. We therefore
have X
I
= X
i
X
J
and as i J we have X
I
= X
i
X
J
= f
m
(X
i
X
J
). Now we
can exploit 3) to dene f
m
on X
j
X
I
by
f
m
(X
j
X
I
) = f
m
(X
j
f
m
(X
i
X
J
))
= f
m
(X
i
f
m
(X
j
X
J
)) +f
m
([X
j
, X
i
] X
J
). (2.5)
One question arises: are the terms on the right-hand side already dened? Well,
in the rst term we have from 2) that
f
m
(X
j
X
J
) = f
m1
(X
j
X
J
) = X
j
X
J
+y
where y S
m1
. Thus the rst term in (2.5) becomes f
m
(X
i
X
j
X
I
)+f
m
(X
i
y) which is already dened since i (j, J). Also the second term in (2.5) is well-
dened as X
J
S
m1
. Thus, we have in a unique way dened a linear map f
m
which clearly satises 1) and 2). Its a bit more problematic to show that it
satises 3): If j < i and j I then f
m
was dened through 3) so it obviously
holds in this case. If i < j and i I we get it from the previous case by
exchanging X
i
with X
j
and use that the Lie bracket is anti-commutative. If
i = j, then 3) is true as the rst 2 terms are equal and the last is 0. Finally we
need to check 3) in the situation where neither i I nor j I. Write I = (i
0
, J)
whence i
0
< i and i
0
< j. As 3) is valid for m 1 we get (after some boring
calculations)
f
m
(X
j
X
I
) = f
m
(X
j
f
m
(X
i0
X
J
))
= f
m
(X
i0
f
m
(X
j
X
J
)) +f
m
([X
j
, X
i0
] X
J
)
since [J[ = m2. Furthermore we have by 2) that
f
m
(X
j
X
J
) = X
j
X
J
+w (2.6)
where w S
m2
. Since i
0
J and i
0
j we may use 3) on the expression
f
m
(X
i
f
m
(X
i0
X
j
X
J
)) and w can be written as a linear combination of X
K
s
with [K[ = m2 and therefore, by linearity, we may use 3) on f
m
(X
i
f
m
(X
j
w)). By (2.6) we can use 3) a couple of times on f

m
(X
i
f
m
(X
j
f
m
(X
i0
X
J
)))
and get
f
m
(X
i
f
m
(X
j
X
I
)) = f
m
(X
i
f
m
(X
j
f
m
(X
i0
X
J
)))
= f
m
(X
i0
f
m
(X
i
f
m
(X
j
X
J
))) +f
m
([X
i
, X
i0
] f
m
(X
i
X
J
))
+f
m
([X
j
, X
i0
] f
m
(X
i
X
J
)) +f
m
([X
i
, [X
j
, X
i0
]] X
J
).
Interchanging i and j in the expression above and subtracting we obtain
f
m
(X
i
f
m
(X
j
X
I
)) f
m
(X
j
f
m
(X
i
X
I
)) = f
m
([X
i
, X
j
] X
I
)
which is exactly 3).
Lemma 2.37. There is a Lie algebra homomorphism : g End(S(g))
satisfying
1) (X
i
)X
J
= X
i
X
J
when i J.
2) (X
i
)X
J
X
i
X
J
S
m
when [J[ = m.
Proof. By Lemma 2.36 we can dene a linear map f : gS(g) End(S(g))
to be the unique map whose restriction to g S
m
equals f
m
. If J has length m,
we have from Lemma 2.36 3) that
f([X
i
, X
j
] X
J
) = f
m
([X
i
, X
j
] X
J
)
= f
m
(X
i
f
m
(X
j
X
J
)) f
m
(X
j
f
m
(X
i
X
J
)).
Therefore we get a Lie algebra homomorphism by (X
i
)X
J
= f(X
i
X
J
). 1)
and 2) are fullled as an immediate consequence of 1) and 2) in Lemma 2.36.
Lemma 2.38. Let Y T
m
J and let Y
m
be the component of Y in g
m
, i.e.
its component of pure degree m. Then Y
m
J.
Proof. We need to show that Y
m
is 0 in the symmetric algebra. Y
m
is a linear
combination of elements of the form X
1
X
m
, X
1
, . . . , X
m
g, so for
simplicity we assume Y
m
is of this form. From the previous lemma we have a
Lie algebra homomorphism : g End(S(g)). Now consider the commutative
diagram
g

P
P
P
P
P
P
P
P
P
P
P
P
P
P T(g)
U(g)
m
m
m
m
m
m
m
m
m
m
m
m
m
m
End(S(g))
induces an algebra homomorphism : U(g) End(S(g)) which by compo-
sition with yields an algebra homomorphism
t
: T(g) End(S(g)) with
ker
t
J. Since Y
m
J we thus have
t
(Y
m
) = (X
1
) (X
m
) = 0. But we
also have by successive use of property 2) satised by that
(X
1
) (X
m
)1 = X
1
X
m
+y
where y is an element in S
m1
. Thus we have X
1
X
m
+ y = 0 and since
X
1
X
m
is the only term of degree m it has to be 0. Thus Y
m
is 0 in S(g).
Now we are nally ready to do what we set out for: proving the Poincar-
Birkho-Witt Theorem:
Proof of Theorem 2.35. The only thing we need to show is that is injec-
tive. This is equivalent to J = ker . From Lemma 2.34 we have J ker . Thus
we only need to show the inclusion . So assume that we have y ker , i.e.
y g
m
and
m
(y) = 0. This implies q
m
((y)) = 0 which implies (y) U
m1
.
Since U
m1
= (T
m1
) there is a y
t
T
m1
so that (y) = (y
t
). But this means
that y y
t
J = ker , hence y y
t
T
m
J. As y is the mth component of
y y
t
the previous lemma says that y J.
This apparently innocent-looking theorem has some useful consequences of
which we will mention a few.
Corollary 2.39. Let W g
m
be a subspace such that [
W
: W

S
m
(g)
is an isomorphism. Then (W) and U
m1
are algebraically complementary sub-
spaces in U
m
, i.e. U
m
= (W) U
m1
.
Proof. Lets consider the following diagram
g
m

U
m
qm
S
m
(g)
G
m
(2.7)
Its commutative since both compositions equal
m
. By assumption and since
is bijective, [
W
: W G
m
is a bijection. Thus also q
m
: W G
m
is
a bijection. Since U
m1
is the kernel of q
m
, it is clear that (W) U
m1
= 0,
namely if x = (y) (W) U
m1
then 0 = q
m
(x) = q
m
(y) and since
q
m
is a bijection, y = 0 and therefore x = 0. Finally we need to show that
U
m
= (W) + U
m1
. The inclusion is obvious, so let x U
m
. If x is in the
kernel of q
m
, it is in U
m1
and we are done, so assume that x is not in the
kernel. Then q
m
(x) ,= 0 and due to bijectivity of q
m
there exists 0 ,= y W
such that q
m
(y) = q
m
(x). But this exactly means that there is an y
t
U
m1
such that x = (y) +y
t
, hence the inclusion is established.
Corollary 2.40. The map i : g U(g) is injective, i.e. g can be considered
a subspace of U(g).
Proof. We clearly have that S
1
(g) = g, and : g S
1
(g) is just the injection
of g into the symmetric algebra, i.e. its bijective. Also is bijective by the
PBW-Theorem. Thus by diagram (2.7) also q
1
is bijective, in particular
: g U
1
U is injective. But restricted to g is nothing but i which is
therefore injective.
Corollary 2.41. Let X
1
, . . . , X
n
be a basis for g. Then the set consisting of
elements of the form X
i1
1
X
in
n
, i
k
N
0
(X
0
k
= 1) is a basis for U(g).
Proof. We will nd a basis for each subspace in the ltration U
m
for U(g),
then since U(g) =
m=0
U
m
we get a basis for U(g).
For U
0
which is just the scalar eld, we have an obvious basis, namely 1.
Now we use induction over m: assume U
m1
has a basis consisting of elements
X
i1
1
X
in
n
with i
1
+ +i
n
m1. Let W g
m
be the subspace spanned
by X
i1
1
X
in
n
with i
1
+ + i
n
= m. This is isomorphic to S
m
(g) via
(they have the same kind of basis) and hence by Corollary 2.39 we have that
U
m
= U
m1
(W). W is via mapped bijectively to the set
spanX
i1
1
X
in
n
[ i
1
+ +i
n
= m.
Thus we get the desired basis for U
m
.
A basis for U(g) as given above is often referred to as a PBW-basis. Sometimes
this last corollary is the PBW-Theorem.
Corollary 2.42. Let g be a Lie algebra and h a subalgebra of g. Then U(h) is
canonically isomorphic to the subalgebra of U(g) generated by 1 and h.
Proof. Let : h g be the injection and compose it with : g U(g) to
get Lie algebra homomorphism h U(g). By the universal property of U(h)
this map factorizes to a unital algebra homomorphism : U(h) U(g). If we
pick a basis X
1
, . . . , X
k
for h and extend it to a basis X
1
, . . . , X
k
, . . . , X
n
for g then as (X
i
) = X
i
for i k and (X
j
) = 0 for j > k and since U(h)
has a PBW-basis generated by X
1
, . . . , X
k
, it should be clear that : U(h)
(U(h)) is an isomorphism.
Corollary 2.43. Let g be the Lie algebra direct sum g
1
g
2
, then there is a
vector space isomorphism U(g)

U(g
1
)
C
U(g
2
).
Proof. From the preceding corollary we can view U(g
1
) and U(g
2
) as subspaces
of U(g). Therefore it makes sense to dene the map : U(g
1
) U(g
2
) U(g)
by (XY ) = XY . By choosing bases X
1
, . . . , X
k
for g
1
and X
k+1
, . . . , X
n
for g
2
, then X
1
, . . . , X
n
will be a basis for g and from the PBW-Theorem it
is not hard to see, that is a vector space isomorphism.
Chapter 3
Basic Representation Theory
of Lie Algebras
3.1 Lie Groups and Lie Algebras
The denition of a representation of a Lie group is not dierent from the def-
inition of a representation of a topological group: It is still a continuous map
G Aut(V ) where Aut(V ) is equipped with the strong operator topology. If
V is nite-dimensional, then Aut(V ) is a Lie group itself and so we can prove
Lemma 3.1. Let (, V ) be a nite-dimensional representation of a Lie group
G. Then the map : G Aut(V ) is a Lie group homomorphism.
Proof. For a nite-dimensional Banach space V the strong operator topology
on B(V ) is equal to the norm topology. Hence the subspace topology on Aut(V )
induced by the strong operator topology on B(V ) is equal to the topology on
Aut(V ) given by the smooth structure. But then, since is a continuous group
homomorphism, a classical result on Lie groups states that is automatically
smooth
1
.
Now, Lie group homomorphisms induce Lie algebra homomorphisms. In order
to exploit this we need to study representation theory on the Lie algebra level.
If V is some xed nite-dimensional vector space, it is well-known that End(V )
with the commutator bracket is the Lie algebra of Aut(V ).
Denition 3.2 (Group Representation). Let g be a Lie algebra and V
a vector space. A representation of g on V is a Lie algebra homomorphism
: g End(V ). We also say that we have given V the structure of a g-
module. If is an injective map, the representation is called faithful .
The dimension of the Lie algebra representation is the dimension of the vector
space, on which it is represented.
Let G be a Lie group with Lie algebra g. If : G Aut(V ) is a nite-
dimensional representation of G, then is a smooth map (by Lemma 3.1) and
it thus induces a Lie algebra homomorphism:
: g End(V ). This renders

the following diagram commutative:
g

exp
End(V )
exp
Aut(V )
(3.1)
1
See for instance [14] Theorem 3.39.
50 Chapter 3 Basic Representation Theory of Lie Algebras
Thus, a nite-dimensional Lie group representation automatically yields a repre-
sentation of the Lie algebra of G. This we call the induced Lie algebra represen-
tation (in physics texts this is sometimes called the innitesimal representation
of g). The converse statement that a nite-dimensional Lie algebra representa-
tion automatically returns a Lie group representation is true in the case where
G is simply connected.
All the concepts introduced in Section 1.1 have Lie algebra analogs:
Denition 3.3. Let : g End(V ) be a Lie algebra representation. A
subspace U V is called -invariant, if for all X g: (X)U U. The
representation is called irreducible if the only invariant subspaces are 0 and g.
Let (
t
, V
t
) be another Lie algebra representation. A linear map T : V V
t
is called an intertwiner between and
t
if T (X) =
t
(X) T for all X g.
The set of all intertwiners is denoted Hom
g
(V, V
t
). If T is an isomorphism, T is
called an equivalence of representations and the two representations are said to
be equivalent (denoted
=
t
).
The analogy is actually quite close as can be seen from the following propo-
sition which compares a Lie group representation with its Lie algebra represen-
tation
Proposition 3.4. Let (, V ) and (
t
, V
t
) be nite-dimensional representations
of a connected Lie group G, and (
, V ) and (
t
, V
t
) the Lie algebra represen-
tations. Then the following hold:
1) A subspace U V is -invariant if and only if it is
-invariant.
is irreducible.
3) T : V V
t
is an intertwiner between and
t
if and only if it is an
intertwiner between
and
t
.
4)

=
t
if and only if

=
t
.
Proof. 1) If G
e
denotes the connected component of G containing the identity
element, then its a fact from elementary Lie group theory that G
e
is the sub-
group of G generated by all elements of the form exp(X) for X g. Since G is
connected, G
e
= G, i.e. all elements of G are on the form exp(X
1
) exp(X
k
).
Now assume that U V is
-invariant. We want to show that (g)U U.

g is of the form exp(X
1
) exp(X
k
) and hence (recall the diagram above)
(g) = (exp(X
1
)) (exp(X
k
)) = exp(
(X
1
)) exp(
(X
k
)).
Since exp(
(X
i
)) =
n=0
1
n!
(
(X
i
))
n
and U is
(X
i
)-invariant, we have
exp(
(X
i
))U U and therefore (g)U U.
Conversely, assume U is -invariant and v U. Then as
(X)v =
d
dt
t=0
((exp(tX))v), (3.2)
(a formula which can be found in introductory textbooks on dierential geom-
etry) and (exp(tX))v U, we have
(X)v U.
2) This now follows from 1).
3) Assume T Hom
g
(V, V
t
), i.e. T
(X) =
t
(X) T. We consider T (g)

and once again use that g = exp(X
1
) exp(X
k
) so that
T (g) = T exp(
(X
1
)) exp(
(X
k
)).
3.1 Lie Groups and Lie Algebras 51
Again since exp(
(X
i
)) =
n=0
1
n!
(
(X
i
))
n
and T is continuous, it inter-
twines each term, and so we have T exp(
(X
i
)) = exp(
(X
i
)) T. Therefore
T (g) =
t
(g) T.
Conversely, assume T (g) =
t
(g) T. Then
(T
(X))v = T
d
dt
t=0
(exp(tX))v =
d
dt
t=0
(T (exp tX)v)
=
d
dt
t=0
(
t
(exp tX) T)v = (
t
(X) T)v.
4) This follows directly from 3).
We even have a version of Schurs Lemma
Theorem 3.5 (Schurs Lemma). Let (, V ) and (
t
, V
t
) be nite-dimensional
irreducible representations of a Lie algebra g. If : V V
t
is an intertwiner,
then is either the zero map or an equivalence of representations.
If (, V ) is a nite-dimensional irreducible complex representation and :
V V a linear map commuting with all (X) then = id
V
for a C.
Proof. ker is an invariant subspace of V , for if v ker then
((X)v) =
t
(X)(v) = 0.
Thus ker = V or ker = 0. In the rst case is the zero map and in the
second it is injective. Likewise im is an invariant subspace of V
t
, so that im =
0 or im = V
t
. In the case of being injective (and V being nontrivial), we
cannot have im = 0. Hence im = V
t
and is a bijection.
For the second assertion, observe that has an eigenvalue C, i.e. id
V
in not injective. Since id
V
intertwines we must have id
V
= 0.
In Chapter 1 we introduced direct sum and tensor product of representations.
How do these operations behave when passing to the Lie algebra level?
Proposition 3.6. Assume that we have two nite-dimensional representations
(, V )and (
t
, W) of a Lie group G. The direct sum representation
t
induces
a representation of g on V W given by
(
t
)
(X)(v, w) = (
(X)v,
t
(X)w),
i.e. (
t
)
(X) =
(X)
t
(X).
The tensor product representation
t
induces a representation of g on
V W given by
(
t
)
(X)(v w) =
(X)v w +v
t
(X)w,
i.e. (
t
)
(X) =
(X) id
W
+id
V

t
(X).
Proof. From (3.2) we have that
(
t
)
(X)(v, w) =
d
dt
t=0
(
t
)(exp(tX))(v, w)
=
d
dt
t=0
_
(exp(tX))v,
t
(exp(tX))w
_
= (
(X)v,
t
(X)w).
For the second part we need to calculate
(
t
)
(X)(v w) =
d
dt
t=0
(
t
)(e
tX
)(v w)
=
d
dt
t=0
(e
tX
)v
t
(e
tX
)w . (3.3)
In other words, we need to be able to dierentiate a tensor product of two smooth
curves. A calculation completely similar to the one carried out when proving
the formula for dierentiation of an ordinary product shows that (v w)
t
(t) =
v
t
(t) w(t) + v(t) w
t
(t) whenever v : I V and w : I W are smooth
vector valued functions. This and (3.3) yield:
(
t
)
(X)(v w) =
d
dt
t=0
_
(e
tX
)v
_
t
(e
0X
)w
+(e
0X
)v
d
dt
t=0
_
t
(e
tX
)w
_
=
(x)v w +v
(X)w
which was the desired expression.
Based on this we simply dene the direct sum of two Lie algebra represen-
tations (, V ) and (
t
, W) to be (
t
)(X) := (X)
t
(X) and the tensor
product of the two by (
t
)(X) := (X) id
W
+id
V

t
(X).
Now consider a complex representation (, V ) of g. We can extend this to a
representation
C
of g
C
on V by
C
(X+iY )v := (X)v+i(Y )v. This we call the
complexication of the representation . For the transition from a representation
to its complexication we have an analogy of Proposition 3.4:
Proposition 3.7. Let g be a nite-dimensional real Lie algebra, and (, V )
and (
t
, V
t
) complex Lie algebra representations. Let g
C
,
C
and
t
C
denote the
associated complexications. Then the following hold:
1) A subspace W V is -invariant if and only if it is
C
-invariant.
C
is irreducible.
3) A linear map T : V V
t
intertwines and
t
if and only if it intertwines
C
and
t
C
.
4) and
t
are equivalent if and only
C
and
t
C
are equivalent.
Proof. 1) If W is -invariant then
C
(X +iY )W = (X)W +i(X)W W.
Conversely, if W is
C
-invariant then for X g we have (X)W =
C
(X)W
W. This proves 1).
2) Follows immediately from 1).
3) If T intertwines and
t
then
T
C
(X +iY ) = T ((X) +i(Y )) = T (X) +T i(Y )
=
t
(X) T +i
t
(Y ) T =
t
C
(X +iY ) T.
Conversely, if T intertwines
C
and
t
C
then for X g:
T (X) = T
C
(X) =
t
C
(X) T =
t
(X) T
which proves the claim.
4) This is an immediate consequence of 3).
3.1 Lie Groups and Lie Algebras 53
In the chapters to follow we will develop a powerful theory to determine
irreducible representations of complex semisimple Lie algebras. But at this point
it is fruitful to examine the simplest semisimple Lie algebra we know: sl(2, C).
Apart from being interesting because of its connection with the group SU(2)
and quantum mechanics, it will also serve as a simple example for the more
general theory.
sl(2, C) is the complex Lie algebra consisting of complex 2 2 matrices with
zero trace. An obvious basis is the following
H :=
_
1 0
0 1
_
E :=
_
0 1
0 0
_
F :=
_
0 0
1 0
_
which obeys the commutation relations
[H, E] = 2E , [H, F] = 2F , [E, F] = H.
Theorem 3.8. For each integer m 1 there is, up to equivalence, only one
irreducible complex representation (
m
, V
m
) of sl(2, C) of dimension m and for
V
m
there exists a basis v
0
, . . . , v
m1
so that
1)
m
(H)v
j
= (m2j 1)v
j
.
2)
m
(E)v
j
= j(mj)v
j1
for j > 0 and
m
(E)v
0
= 0.
3)
m
(F)v
j
= v
j+1
for j < m1, and
m
(F)v
m1
= 0.
Proof. We rst prove that any irreducible complex representation is of the
form mentioned in the theorem. So let (, V ) be an arbitrary complex irre-
ducible representation of sl(2, C) and let m be its dimension. (H) is a complex
endomorphism of V and therefore has an eigenvalue
t
with eigenvector v. We
see that
(H)(E)v = (E)(H)v +([H, E])v =
t
(E)v + 2(E)v = (
t
+ 2)(E)v.
Thus either (E)v is zero or its an eigenvector for (H) with an eigenvalue
dierent from
t
. Since the representation is nite-dimensional there can only
be nitely many distinct eigenvalues, i.e. there is a j
0
such that (E)
j0
v ,= 0
and (E)
j0+1
v = 0. Lets dene v
0
:= (E)
j0
v and lets denote the eigenvalue
by . Thus v
0
satises
(H)v
0
= v
0
and (E)v
0
= 0.
We now dene v
j
:= (F)
j
v
0
. These satisfy
(H)v
j
= (H)(F)
j
v
0
= ( 2j)v
j
and (F)v
j
= v
j+1
.
Again, as there are only nitely many distinct eigenvalues, there exists a max-
imal n such that v
n
,= 0 and v
n+1
= 0. Let W = spanv
0
, . . . , v
n
. We want
to show that W = V , and as is irreducible, it suces to verify that W is
an invariant subspace. All the vectors in W are eigenvectors for (H), so W is
(H)-invariant. By construction W is also (F)-invariant. For (E)-invariance
we use induction over j to show that (E)v
j
= j( j + 1)v
j1
for j > 0 and
(E)v
0
= 0. This last information we already have, so we proceed immediately
to the induction step: Assume (E)v
j
= j( j + 1)v
j1
. Then
(E)v
j+1
= (E)(F)v
j
= (F)(E)v
j
+([E, F])v
j
= (F)(j( j + 1)v
j1
) +(H)v
j
= j( j + 1)v
j
+ ( 2j)v
j
= (j + 1)( j)v
j
,
and this was what we wanted. Hence W is also (E)-invariant and therefore
W = V and n + 1 = m = dimV . The only thing left is to show that = n,
i.e. (H)v
0
= nv
0
. To show this we calculate the trace of (H) in two dierent
ways. First (H) = ([E, F]) = [(E), (F)] and the trace of a commutator
is always 0. Thus Tr (H) = 0. On the other hand, in the basis v
0
, . . . , v
n
(H) is diagonal with eigenvalues , 2, . . . , 2n and therefore the trace is

+ ( 2) + + ( 2n). The only way this could ever equal 0 is if = n.
This shows that (, V ) is equivalent to (
m
, V
m
) and the uniqueness part of
the theorem is proved.
To prove existence, let V be a complex vector space of dimension m and let
v
0
, . . . , v
m1
denote a basis. We dene a linear map : sl(2, C) End(V )
by
(H)v
j
:= (m2j 1)v
j
(E)v
j
:= j(mj)v
j1
and (E)v
0
= 0
(F)v
j
:= v
j+1
and (F)v
m1
= 0
i.e. we brutally force to have the properties we want. It is not hard to verify
that this is really a representation of sl(2, C).
To see that is irreducible, let W V be a nontrivial invariant subspace.
As W is (H)-invariant and (H) has the v
j
s as eigenvectors, W must contain
at least one of these, say v
k
. Then as W is (E)-invariant, and v
0
is a multiple
of (E)
k
v
k
which is in W, we must have v
0
W. And nally as W is (F)-
invariant we have v
j
= (F)v
0
W for all j = 0, . . . , m1. Hence W = V and
is irreducible. This completes the proof.
3.2 Weyls Theorem
In this section we will show a nice theorem due to Weyl, stating that any nite-
dimensional representation of a semisimple Lie algebra is completely reducible,
i.e. can be decomposed into irreducible representations.
Lemma 3.9. Let g be a semisimple Lie algebra, and (, V ) a nite-dimensional
faithful representation of g. Then the bilinear form on g given by (X, Y ) =
Tr((X)(Y )) is symmetric, non-degenerate and ([X, Y ], Z) = (X, [Y, Z]).
Note that if is just the adjoint representation of g, then is nothing but
the Killing form of g.
Proof. That is symmetric and has the associativity property ([X, Y ], Z) =
(X, [Y, Z]) can be proved in exactly the same way as we proved Proposition
2.22.
Now dene, as for the Killing form, the radical
Rad := X g [ Y g : (X, Y ) = 0.
Obviously is non-degenerate if and only if Rad = 0. We will use Cartans
Criterion to show that (Rad ) is solvable. For X, X
t
, Y Rad we have
Tr([(X), (X
t
)](Y )) = ([X, X
t
], Y ) = 0
by denition of Rad , hence (Rad ) is solvable. Since : g (g) is a
Lie algebra isomorphism, (g) is semisimple. As (Rad ) (g) is solvable,
we must have (Rad ) = 0 and consequently Rad = 0, i.e. is non-
degenerate.
3.2 Weyls Theorem 55
Now let B = X
1
, . . . , X
n
be a basis for g. Since is non-degenerate we get a
corresponding dual basis B
t
= Y
1
, . . . , Y
n
for g determined by (X
i
, Y
j
) =
ij
.
For each X g we can expand [X, X
i
] and [X, Y
i
] in the bases B and B
t
respectively
[X, X
i
] =
n
j=1
a
ij
X
j
and [X, Y
i
] =
n
j=1
b
ij
Y
j
.
By associativity of we get the following connection between the coecients
a
ik
=
n
j=1
a
ij
(X
j
, Y
k
) = ([X, X
i
], Y
k
) = ([X
i
, X], Y
k
) = (X
i
, [X, Y
k
])
=
n
j=1
b
kj
(X
i
, Y
j
) = b
ki
.
Denition 3.10 (Casimir element). Still assuming to be faithful we dene
the Casimir element of relative to the basis B by
c
=
n
i=1
(X
i
)(Y
i
) End(V ).
One can show that this is in fact independent of the choice of basis. Since
ad((X)) is a derivation of End(V ) i.e.
[(X), (Y )(Z)] = [(X), (Y )](Z) +(Y )[(X), (Z)]
c
commutes with (X):

[(X), c
] =
n
i=1
[(X), (X
i
)(Y
i
)]
=
n
i=1
[(X), (X
i
)](Y
i
) +
n
i=1
(X
i
)[(X), (Y
i
)]
=
n
i,j=1
a
ij
(X
j
)(Y
i
) +
n
i,j=1
b
ij
(X
i
)(Y
i
) = 0.
In the case of being irreducible Schurs Lemma says that c
is nothing but
multiplication by a constant . We can even calculate this particular constant:
On one hand we must have Tr c
= dimV , but on the other hand

Tr c
=
n
i=1
Tr (X
i
)(Y
i
) = (X
i
, Y
i
) = n = dimg
therefore =
dimg
dimV
so in this case we see that the Casimir element is indeed
independent of the choice of basis.
If is not faithful, the above construction is meaningless since is no longer
non-degenerate and the dual basis is therefore ill dened. Nonetheless we can
still dene a Casimir element of . The kernel ker is an ideal in g and is
therefore a direct sum of simple ideals (g is still assumed to be semisimple).
Let g
t
denote the direct sum of the remaining simple ideals in g, then g
t
is
semisimple, g = ker g
t
and [
g
is injective. Choosing a basis for g
t
we dene
the Casimir element c
of to be the Casimir element of [

g
relative to the
chosen basis for g
t
. Obviously, c
commutes with (X) just as before. Therefore

if is irreducible also [
g
is irreducible and by the same reasoning as above we
see that c
is just multiplication by the constant = dimg

t
/ dimV which is
non-zero unless the representation is trivial.
Lemma 3.11. Let (, V ) be a representation of a semisimple Lie algebra g.
Then (g) sl(V ), i.e. Tr (X) = 0 for each X g. In particular any 1-
dimensional representation will be trivial : (X)v = 0.
Proof. As g is semisimple, we have that g = [g, g] i.e. an arbitrary X g
can be written X = [Y, Y
t
]. Therefore (X) = [(Y ), (Y
t
)] and commutators
always have zero trace.
Theorem 3.12 (Weyl). Every nite-dimensional representation of a semisim-
ple Lie algebra is completely reducible.
Proof. At rst we work out the case in which W V is a codimension 1
invariant subspace on which is irreducible. As mentioned, the Casimir element
c
commutes with (X) and from this it is not hard to extract that ker c
is an
invariant subspace of V and that (by Schurs Lemma) c
(W) W. The induced

representation of on V/W is trivial by the preceding lemma (for V/W is 1-
dimensional) and this says that (X) maps V into W. Thus since c
is just
the sum of compositions of certain (X)s we also have that c
maps V into
W. Therefore c
has a nontrivial kernel. But on c
[
W
is the Casimir element of
[
W
and by the remarks above this is just multiplication by a nonzero constant.
Therefore W ker c
= 0, and hence V = W ker c
, i.e. ker c
is the desired
invariant subspace.
Next we will investigate the case in which V contains an invariant subspace
W of codimension 1 (i.e. dimV dimW = 1) on which is not irreducible. We
want to nd a (necessarily 1-dimensional) invariant subspace Z of V such that
V = WZ. We do accomplish this by induction over dimW. If dimW = 0 then
we can take Z to be V . For the induction step let dimW > 0 and assume the
conclusion to hold for all dimensions less than Ws. Let W
t
W be a proper
non-zero invariant subspace, then W
t
is also an invariant subspace of V and
W/W
t
V/W
t
is a subspace of codimension 1. induces a representation on
V/W
t
by (X)[v] = [(X)v] (it is easily checked that it is well-dened). Then
W/W
t
is an invariant subspace of V/W
t
of codimension 1 and since dimW/W
t
<
dimW the induction hypothesis (if restricted to W/W
t
is not irreducible) or
the remarks above (if restricted to W/W
t
is irreducible) yields a 1-dimensional
invariant subspace

W/W
t
V/W
t
such that V/W
t
= W/W
t
W/W
t
. But this
means that W
t

W is a subspace of codimension 1 and again since dimW
t
<
dimW we get a 1-dimensional invariant subspace Z such that

W = W
t
Z.
To see that we also have V = W Z we observe that W Z = 0 for if
W Z ,= 0 we would have that Z W and consequently

W/W
t
W/W
t
contradicting the decomposition above. Thus W Z = 0 and therefore (for
dimension reasons) V = W Z.
Now for the general case where W V is just some invariant subspace.
gives rise to a representation of Hom(V, W) by the denition
(X)f = (X)f f(X).
Let U Hom(V, W) be the set of linear maps f which restricted to W are just
multiplication by a constant, i.e. f U if w W : f(w) = w for some C.
U is a -invariant subspace since for w W
(X)w = (X)f(w) f((X)w) = (X)w (X)w = 0
i.e. ( (X)f)[
W
= 0 for f U. Now let

U U be the set of linear maps
f : V W such that f[
W
= 0. Also

U is an invariant subspace of U, and

U
has codimension 1 since any function f U can be written as a sum

f +f
0
where
f

U and where f
0
is id
W
extended by 0 to all of V . From the analysis in the
3.2 Weyls Theorem 57
beginning of the proof we get a 1-dimensional invariant subspace Z U such
that U =

U Z. Let g
0
: V W be a function that spans Z. By a scaling we
can obtain g
0
[
W
= id
W
. Since Z is 1-dimensional the lemma says that (X)f
0
=
0, i.e. g
0
commutes with (X) and therefore ker g
0
is an invariant subspace of V .
Again, since g
0
[
W
= id
W
we must have W ker g
0
= 0 and as img
0
= W the
Dimension Theorem of linear algebras says that dimker g
0
+dimimg
0
= dimV
and therefore V = W ker g
0
.
What we have actually proved now, is that any invariant subspace has a
complementary invariant subspace. Having observed this the rest of the proof
is identical to the proof of Proposition 1.16.
Armed with this theorem we only need to know the irreducible representations
of a semisimple Lie algebra to classify the nite-dimensional representations.
Classifying the irreducible representations of complex semisimple Lie algebras
is the goal of a later chapter. For now we state and prove to corollaries regarding
sl(2, C) representations which will be useful at a later stage.
Corollary 3.13. If is a nite-dimensional complex representation of sl(2, C)
on V , then (H) has only integer eigenvalues.
Proof. sl(2, C) is semisimple and so by Weyls Theorem V is a direct sum
V = V
1
V
n
so that [
Vi
is irreducible. Assume v V to be an eigenvector
for (H) and the corresponding eigenvalue. Then v = v
1
+ + v
n
with
v
i
V
i
and we see that (H)v
i
= v
i
i.e. is an eigenvalue for, say, [
V1
. But
from Theorem 3.8 it should be apparent that such an eigenvalue has to be an
integer.
Corollary 3.14. Let : sl(2, C) End(V ) be a, possibly innite-dimensional,
representation such that for each v V there is a nite-dimensional invariant
subspace containing v, then is a, possibly innite, direct sum of irreducible
representations.
Proof. Every element v V is contained in a nite-dimensional invariant sub-
space, and restricted to this subspace can, by Weyls Theorem, be decomposed
into a direct sum of irreducible representations. Thus we have V =
iI
U
i
where U
i
is a nite-dimensional irreducible invariant subspace of V . We need
to show that this sum of subspaces is in fact a direct sum. To this end call a
subspace J I an independent set if

iJ
U
i
is a direct sum. One-point sets
are examples of independent sets. Its not hard to see that the set of indepen-
dent subsets of I, ordered by inclusion, is inductively ordered (i.e. every totally
ordered subset has an upper bound). Thus Zorns Lemma gives a maximal in-
dependent subset J
0
I. Put
V
0
:=
iJ0
U
i
=
iJ0
U
i
.
We obviously have V
0
V . To show the reverse inclusion consider U
i
for an
arbitrary i I. We want to show that U
i
V
0
and hence that V V
0
. This is
trivially true if i J
0
. If not then, by maximality, V
0
+U
i
cannot be a direct sum,
i.e. V
0
U
i
,= 0. But V
0
U
i
U
i
is an invariant subspace and by irreducibility
of U
i
we must have V
0
U
i
= U
i
i.e. U
i
V
0
.
Chapter 4
Root Systems
4.1 Weights and Roots
First some terminology:
Denition 4.1 (Maximal Torus). Let g be a nite-dimensional Lie algebra.
A torus t in g is a commutative subalgebra of g. A maximal torus in g is a torus
which is not contained in any strictly larger torus.
Lemma 4.2. Let t be any maximal torus of g. Then t equals its own centralizer.
Proof. As t is commutative then every element of t commutes with all of t,
hence t sits in its centralizer. Conversely, assume that X is in the centralizer,
then obviously t+KX is a torus. Since t was maximal this implies that X t.
Every Lie algebra possesses a nonzero maximal torus: Pick any 1-dimensional
Lie subalgebra, this is automatically commutative, and nd a maximal commu-
tative subalgebra that contains it.
Denition 4.3 (Cartan Subalgebra). Let g be a complex Lie algebra. A
Cartan subalgebra h of g is a maximal torus such that ad(H) End(g) is
diagonalizable for all H h.
It is not obvious at all that non-trivial Cartan subalgebras exist. One can
show that they do in fact exist if g is semisimple. Furthermore, if both h and
h
t
are Cartan subalgebras, one can show that there is an automorphism of g
mapping h to h
t
. In particular all Cartan subalgebras have the same complex
dimension. This common dimension is called the rank of the Lie algebra.
Proposition 4.4. If g
i
are Lie algebras and h
i
are corresponding Cartan sub-
algebras, then h
1
h
2
is a Cartan subalgebra of g
1
g
2
.
Proof. Obviously h
1
h
2
is an abelian subalgebra and its maximally compact,
for if (X, Y ) g
1
g
2
commutes with every (H
1
, H
2
) h
1
h
2
, then
(0, 0) = [(X, Y ), (H
1
, H
2
)] = ([X, H
1
], [Y, H
2
])
and since h
i
is maximally abelian, we get X h
1
and Y h
2
. Furthermore, we
see that for H = (H
1
, H
2
) h
1
h
2
that ad(H) : g
1
g
2
g
1
g
2
is the map
given by
(X, Y ) (ad(H
1
)X, ad(H
2
)Y )
and since ad(H
i
) is diagonalizable, ad(H) is diagonalizable as well.
59
60 Chapter 4 Root Systems
Denition 4.5 (Weight). Let g be a complex Lie algebra and consider and a
(possibly innite-dimensional) representation (, V ) of g and let h be any xed
torus in g. A functional h
for which
V
:= v V [ (H)v = (H)v for all H h

is non-zero is called a weight for w.r.t. h. The set of weights is denoted (, h)
(or just ()). For (, h) the space V
is called the weight space associated

with and the elements are called weight vectors.
It is not hard to see that the set of weights is preserved under equivalence
of representations: Consider two representations (, V ) and (
t
, V
t
) of g which
are equivalent through the bijective intertwiner T : V V
t
, and let ()
and (
t
) be the set of weights relative to a xed torus h in g. If (),
then V
and T(V
) are both nonzero. For 0 ,= v V
we see that
t
(H)(Tv) =
T((H)v) = (H)Tv, i.e. T(V
) V
t
which is therefore nonzero. Thus ()

(
t
). A similar argument gives the reverse inclusion and hence that equivalent
representations have the same weights.
We record some important properties of weights
Theorem 4.6. Let (, V ) be a nite-dimensional representation of g and h a
torus in g. Then (, h) is nonempty and nite.
If furthermore (H) is diagonalizable for each H h, then we have the weight
space decomposition
V =
()
V
. (4.1)
Proof. Let X
1
, . . . , X
n
be a basis for h. As is a complex representation,
(X
1
) must have an eigenvalue,
1
and associated with this the eigenspace E
1
.
Since h is abelian it is easy to see that (H) maps E
1
into E
1
for any H h,
and by the same argument (X
2
)[
E1
must have an eigenvalue
2
with eigenspace
E
2
,= 0. By induction we get a subspace 0 , = E
j
E
1
, j = 1, . . . , n on which
(X
j
) acts by multiplication by
j
. Thus, if we dene the functional h
by
(X
j
) =
j
we get (H)v = (H)v for all H h and v E
n
. Thus E
n
V
so
that V
,= 0 and (, h) ,= .
If is a weight then (X
i
) is an eigenvalue for (X
i
). Since this acts on a
nite-dimensional space, there can be only nitely many eigenvalues. Thus there
can be only nitely many weights.
Now assume that (H) is diagonalizable for each H h. In particular this
holds for X
1
. Therefore we can decompose V into eigenspaces for (X
1
). Again
as h is abelian (X
2
) maps these eigenspaces into themselves and as (X
2
) is di-
agonalizable we can decompose each of the eigenspaces of (X
1
) into eigenspaces
for (X
2
). Each of these eigenspaces can then be decomposed into eigenspaces
for (X
3
) and so on, until we have a decomposition V = V
1
V
N
(which
is nite as V is nite-dimensional) such that (X
i
) acts on V
j
by the scalar
ij
,
i = 1, . . . , n, j = 1, . . . , N.
Now dene functionals
j
h
, j = 1, . . . , N by
j
(X
i
) =
ij
. We claim that
(, h) =
1
, . . . ,
N
. To show that
j
is a weight we only need to see that
V
j
= v V [ (H)v =
j
(H)v is non-zero. But we see that 0 , = V
j
V
j
for if v V
j
then for any i = 1, . . . , n we have
(X
i
)v =
ij
v =
j
(X
i
)v. (4.2)
Thus
j
is a weight.
If is a weight how is V
related to the subspaces V

j
? The claim is that
V
j:=j
V
j
.
4.1 Weights and Roots 61
The inclusion is obvious from (4.2). To see let v V
and split it
v = v
1
+ + v
N
with v
j
V
j
. Since V
j
is -invariant we also have (H)v
j
=
(H)v
j
. Let J be an index such that
J
,= . Then there exists an I such that
(X
I
) ,=
J
(X
I
). But on the other hand we have
(X
I
)v
J
=
IJ
=
J
(X
I
)v
J
.
The only way this can be compatible with (X
I
) ,=
J
(X
I
) is if v
J
= 0. This
shows the reverse inclusion.
Now suppose is a weight. Then we just showed that V
j=
V
j
. As V
is non-trivial there is at least one j such that =

j
Therefore we have (, h) =
1
, . . . ,
N
. In particular we have the weight space decomposition.
In particular, due to the weight space decomposition, the number of weights
cannot exceed dim
C
V .
From the proof of this proposition we saw that diagonalizability of the en-
domorphisms (X) was crucial for the weight space decomposition to hold. By
denition of the Cartan subalgebra the adjoint representation exactly possesses
this property. To exploit this we make the following denition
Denition 4.7 (Root). Let g be a complex Lie algebra and h be any xed
Cartan subalgebra in g. A nonzero functional h
for which
g
:= X g [ [H, X] = (H)X for all H h

is non-zero is called a root for g w.r.t. h. The set of roots is denoted R(g, h)
(or just R for brevity). For R(g, h) the space g
is called the root space

associated with and the elements are called root vectors for .
Thus roots are nothing more that weights of the adjoint representation. Notice
that we do not recognize the zero functional as a root. This, however, doesnt
mean that g
0
is zero. In fact g
0
= h: By denition of g
0
, this is just the centralizer
of h. But by denition h is itself a maximal torus and thus equals its own
centralizer g
0
.
Now a direct translation of Theorem 4.6 gives
Corollary 4.8. Let g be a complex Lie algebra and h a Cartan subalgebra. Then
R(g, h) is nonempty and nite and we have the so-called root space decomposi-
tion of g relative to h:
g = h
_
R
g
_
. (4.3)
In particular we see from the root space decomposition that the number of
roots cannot exceed dim
C
g dim
C
h.
In the case of a semisimple Lie algebra, roots carry a lot of information about
the structure of the Lie algebra, in fact they can be used to classify semisimple
Lie algebras, and thus it pays o to investigate the set of roots somewhat closer.
Thats the purpose of the next section. For now we just prove some elementary
results without the assumption of g being semisimple.
Proposition 4.9. Let (, V ) be a (possibly innite-dimensional ) representation
of a complex Lie algebra g and let h g be a Cartan subalgebra. For (, h)
and R(g, h) we have
(g
)V
V
+
.
Thus, if + is not a weight, then (g
) annihilates V
.
Proof. Let X g
and v V
. Then for all H h we have

(H)((X)v) = (X)(H)v + [(H), (X)]v
= (H)(X)v +([H, X])v
= (H)(X) +((H)X)v = ((H) +(H))(X)v,
i.e. (X)v V
+
. If + is not a weight, then V
+
= 0 and the last assertion
follows.
In the case of the adjoint representation of g we immediately get
Corollary 4.10. For , R(g, h) we have
[g
, g
] g
+
.
4.2 Root Systems for Semisimple Lie Algebras
In this section we record some further results about roots relying heavily on
semisimplicity of the Lie algebras in question.
In this section g will be a complex semisimple Lie algebra and h a Cartan
subalgebra. B will denote the Killing form, i.e. the bilinear form g g C
given by B(X, Y ) = Tr(ad(X) ad(Y )). On a semisimple Lie algebra, this is
non-degenerate (cf. Theorem 2.24).
Lemma 4.11. Let R = R(g, h) be the set of roots, then:
1) Let , R0 so that + ,= 0, X g
and Y g
, then B(X, Y ) =
0, in other words: g
and g
are B-orthogonal.
2) For R 0 B[
gg
is non-singular, i.e. B(X, Y ) = 0 for every
Y g
implies X = 0.
3) If R then R.
4) B[
hh
is non-degenerate, thus to any R there is a unique H
h
satisfying (H) = B(H, H
) for all H h.
5) For R, X g
and Y g
, then [X, Y ] = B(X, Y )H
.
6) For , R, (H
) is a rational multiple of (H
).
7) The roots span h
.
8) If R then (H
) ,= 0.
Proof. 1) From Corollary 4.10 we have that [g
, g
] g
+
or equivalently
ad(g
)g
g
+
, and hence
ad(g
) ad(g
)g
g
++
.
As + ,= 0, g
++
,= g
, so ad(X) ad(Y ) maps g
to a totally dierent
subspace of g. Therefore by picking bases for the root spaces and for g
0
= h
and putting them together to a basis for g (the root space decomposition),
the matrix representation of ad(X) ad(Y ) must have zeros in the diagonal and
therefore the trace is 0.
2) Let X g
and assume B(X, Y ) = 0 for all Y g
. If we can show that

B(X, Y ) = 0 for all Y g, then non-degeneracy of B renders X = 0. But if
R 0 and ,= and Y g
then 1) tells us that B(X, Y ) = 0 and by

the root space decomposition we get that B(X, Y ) = 0 for all Y g.
4.2 Root Systems for Semisimple Lie Algebras 63
3) If R then there is a non-zero X g
, and assuming g
to be 0 we get
from 2) that X = 0 since B(X, Y ) = 0 for all Y g
. Thus by contradiction
g
is non-zero and is thus a root.

4) Non-degeneracy of B on h h is an immediate consequence of 2) with
= = 0. Hence we get an isomorphism

B : h h
by mapping X to the
functional B(X, ). Therefore if we put H
:=

B
1
() we get an element in h
satisfying
(H) =

B(H
)(H) = B(H
, H).
5) By Corollary 4.10 we get [g
, g
] h and so for Y g
, we have
[X, Y ] h. For H h we see that
B([X, Y ], H) = B(Y, [X, H]) = (H)B(Y, X)
= B(H, B(X, Y )H
) = B(B(X, Y )H
, H)
and as B[
hh
is non-degenerate, [X, Y ] = B(X, Y )H
.
6) Pick X
,= 0 in g
. The by 2) there exists a 0 ,= Y
so that
B(X
, Y
) ,= 0. By a proper scaling we can assume B(X
, Y
) = 1. Then 4)
says that [X
, Y
] = H
.
Now, put g
t
:=
nZ
g
+n
(observe that only nitely many of these spaces
are non-trivial). Since ad(H
) maps any root space to to itself, ad(H
) will also
map g
t
into itself. Thus by restriction of ad(H
) we get a map g
t
g
t
. We
will calculate the trace of this in two ways. On one hand the action of ad(H
)
on g
+n
is [H
, X] = ( +n)(H
)X so the trace becomes

Tr(ad(H
)) =
nZ
((H
) +n(H
)) dimg
+n
.
Again, this sum is nite as only nitely many of the dimensions are non-zero.
On the other hand, as ad(X
) maps g
+n
into g
+(n+1)
, g
t
is ad(X
)-
invariant. Similarly it is ad(Y
)-invariant, so as a map on g
t
we have
ad(H
) = ad X
ad Y
ad Y
ad X
and consequently the trace of ad H
is 0. Solving the equation

nZ
((H
) +
n(H
)) dimg
+n
= 0 yields
(H
) =
nZ
ndimg
+(n+1)
nZ
dimg
+(n+1)
(H
)
and this is a rational multiple of (H
).
7) Pick a basis H
1
, . . . , H
n
for h, then the functionals
k
, k = 1, . . . , n given
by
k
(H
j
) =
kj
will constitute a basis for h
. Assume that the roots do not

span h
, then there is a
k
which is not in the span of the roots. This can only
happen if (H
k
) = 0 for all roots R. Thus for any root and X g
we
have 0 = (H
k
)X = [H
k
, X], and since H
k
commutes with every element of h
H
k
must be in the center of g. But for a semisimple Lie algebra the center is
trivial, i.e. H
k
= 0 and hence a contradiction.
8) Assume (H
) = 0. Then by 6) (H
) = 0 for all roots . By 7) the roots

span h
, and this implies H
= 0. But then (H) = B(H, H
) = 0 for all H h,
i.e. H
= 0 and hence = 0 which is a contradiction.

Observe that for point 1) it wasnt necessary for g to be semisimple.
The particular element H
h whose existence was asserted in 4) is called a

co-root. Unlike X
or Y
which were arbitrarily chosen, H
is really unique,
due to semisimplicity of g. In fact for any functional h
we have a unique
element H
=

B
1
() h. This gives rise to a bilinear form , on h
by
, := B(H
, H
). (4.4)
Thus we see that (H
) = (H
) = , . Note, however, that this is not an

inner product on h
, since it is symmetric and not conjugate symmetric (recall

that h
is a complex vector space). At the end of this section we restrict it to a

real subalgebra on which it is an inner product.
Our next objective is to break up g into copies of sl(2, C) the representation
theory of which we have already studied. The rst result states that the only
choice in picking X
is the choice of a constant.

Proposition 4.12. Let g be semisimple. If is a root, then dimg
= 1 and
the only integer multiples of in R are .
Proof. As before we pick X
and Y
so that B(X
, Y
) = 1
and therefore [X
, Y
] = H
.
Now, dene
g
tt
= CX
CH
n<0
g
n
.
We progress as in the previous proof: We show that the space is invariant un-
der ad H
, ad X
and Y
and calculate the trace of ad H
= ad X
ad Y

ad Y
ad X
on this space, obtaining an equation which yields the desired re-

sult. It is obviously ad H
-invariant. As
[X
, X
] = 0 , [X
, H
] = (H
)X
and since ad X
maps g
n
into g
(n+1)
(which is at most h) it is also ad X
-
invariant. Finally as
[Y
, X
] = H
, [Y
, H
] = (H
)Y
and since ad Y
maps g
n
into g
(n1)
, it is also ad Y
-invariant.
Now we calculate the trace. Like before ad H
is a commutator and as such

has zero trace. On the other hand ad H
acts on g
tt
by eigenvalues (H
) ,= 0,
0 and n(H
) respectively and therefore the trace becomes

Tr ad H
= (H
) +
n=1
n(H
) dimg
n
.
Equating this with zero yields

n=1
ndimg
n
= 1. Since g
is a non-trivial
space we must have dimg
= 1 and dimg
n
= 0 for n > 1. Since n R if
an only if n R (cf. Lemma 4.11) the claim is proved.
In addition to the formula (2.3) we have the following relatively simple formula
for computing the Killing form when restricted to a Cartan subalgebra. This is
of particular interest because of Lemma 4.11 which says that B, due to non-
degeneracy, provides an isomorphism from h to its dual.
Corollary 4.13. Restricted to h h the Killing form is given by
B(H, H
t
) =
R
(H)(H
t
).
Proof. Pick a basis H
1
, . . . , H
k
for h and pick for each root a non-zero
element X
. Then the set H

1
, . . . , H
k
, X
1
, . . . , X
n
is a basis for g.
ad H ad H
t
acts by eigenvalue 0 on H
1
, . . . , H
k
and with eigenvalue (H)(H
t
)
on X
. Hence
B(H, H
t
) = Tr(ad H ad H
t
) =
R
(H)(H
t
).
With this at hand we are actually capable of computing the Killing form, cf.
Example 4.26 in the next section.
As observed in the proof of Proposition 4.12 we have for each root a triple
X
, Y
, H
satisfying
[H
, X
] = (H
)X
, [H
, Y
] = (H
)Y
, [X
, Y
] = H
.
Thus they span a 3-dimensional complex subalgebra. If we normalize, i.e. dene
H
t
=
2H
(H
)
, X
t
=
2X
(H
)
, Y
t
= Y
,
then we see that
[H
t
, X
t
] = 2X
t
[H
t
, Y
t
] = 2Y
t
[X
t
, Y
t
] = H
t
i.e. the triple X

t
, Y
t
, H
t
is isomorphic, as a Lie algebra, to sl(2, C). There-

fore we denote sl
:= spanX
t
, Y
t
, H
t
.
Let R and R 0. By an -string containing we understand the
set of roots of the form +n for n Z. This is also called a root string.
Proposition 4.14. Let R and R 0.
1) The -string containing is of the form
+n [ p n q (4.5)
for xed p, q N
0
, that is, there are no gaps in the string. Furthermore
p q = 2
,)
,)
, i.e. 2
,)
,)
is an integer.
2) If is not an integer multiple of then we dene a representation of
sl
on g
t
:=
nZ
g
+n
by restriction of ad
g
to sl
. This representation
is irreducible.
Proof. As we saw in Proposition 4.12 the only integer multiples of which
are in R are . Thus if is an integer multiple of then = and the root
string is then either 0, 2 or 2, 0. In both cases it is readily checked that
1) holds.
Under the assumption that is not an integer multiple of we check 1) and
2) simultaneously. It is evident that ad H
t
acts diagonally on g
t
with eigenvalues
( +n)(H
t
) =
2( +n)(H
)
,
=
2,
,
+ 2n.
Thus any ad H
t
-invariant subspace of g
t
is a sum of g
+n
-spaces. Since -
invariant subspaces are in particular ad H
t
-invariant, this holds for -invariant

subspaces as well. Now, let V g
t
be a -invariant subspace on which is
irreducible, and let p and q be the smallest resp. greatest integer n such that
g
+n
is a summand in V . From the representation theory of sl(2, C) we know
that this is the unique irreducible representation of sl
of dimension dimV , and

thus that the eigenvalues of ad H
t
are dimV + 1, dimV + 3, . . . , dimV

3, dimV 1. Comparing with the eigenvalues above we see that there can be no
gaps, i.e. the -string contains the roots in 4.5. Furthermore we see that
dimV 1 =
2,
,
+ 2q and 1 dimV =
2,
,
2p.
Adding these two equations yields p q = 2
,)
,)
.
By Weyls Theorem can be decomposed into irreducible representations.
Assume that V
t
is an irreducible summand in g
t
dierent from V and let again
p
t
and q
t
be the smallest resp. greatest n such that g
+n
is a summand in
V
t
. But then p
t
q
t
= pq = 2
,)
,)
. Since V V
t
= 0 we must have p
t
> q
or q
t
< p. Assuming the rst we get
q
t
p
t
> q p
i.e. p > q
t
. Adding this to q > p
t
(which we have by assumption) we get
p q > p
t
q
t
which is a contradiction. Similarly if we assume q
t
< p. Thus,
there cannot exist other irreducible summands in g
t
than V and consequently
g
t
= V , so the representation is irreducible. Hence g
t
is the direct sum of g
+n
where p n q, thus (4.5) is indeed equal to the -string.
The last thing we need to see is that p and q are non-negative. If = 0 then
is in the root string and hence p, q 1. If ,= 0 then is in the root string
and therefore p, q 0. Thus in any case p and q are non-negative.
We can use this to elaborate on the result of Corollary 4.10
Corollary 4.15. If , R 0 and + ,= 0, then [g
, g
] = g
+
.
Proof. Since we cannot have = = 0, we can assume ,= 0. If at rst,
= n, we only have 2 possibilities for n: 0 and 1 (we cannot have = ). If
= 0, then g
= h and by denition of g
: [g
, h] = g
= g
+
. If = then
g
+
= g
2
= 0 and the equality follows immediately from Corollary 4.10.
Assuming is not an integer multiple of we use the preceding proposition.
For each n between p and q we have a representative X
+n
for g
+n
. These
correspond to the vectors v
0
, . . . , v
N
we found when we studied sl(2, C), since
ad X
maps X
+n
to X
+(n+1)
. The only one of these listed elements which
is mapped to 0 by ad X
is X
+q
.
Now, if [X
, X
] ,= 0 then g
+
,= 0 i.e. + R and since the root spaces
are 1-dimensional we must have the desired equality. If [X
, X
] = 0, then we
must have q = 0 (for X
+q
was the only element mapped to 0 by ad X
). But
that means the + is not in the root string, i.e. + / R so that g
+
= 0.
Thus we must have equality.
Corollary 4.16. Let , R and assume them not to be multiples of each
other. Let p and q be the integers from Proposition 4.14 then for X
,
Y
and X
[Y
, [X
, X
]] =
q(1 +p)
2
(H
)B(X
, Y
)X
. (4.6)
Proof. Both sides equal 0 if one of X
, Y
and X
is 0, so lets assume them

to be non-zero. It is enough (by Proposition 4.12) to show that the formula
holds for X
t
and Y
t
and since B(X

t
, Y
t
) =
2
(H)
the formula we want to
verify is
[Y
, [X
, E
]] = q(1 +p)X
.
Comparing with the situation for sl(2, C) we have (up to constants) that v
0
corresponds to X
+q
, v
1
to X
+(q1)
and v
N
to X
p
(where N = dimg
1 = p + q + 1) and ad Y
maps X
+n
to X
+(n1)
(up to a scale factor).
Therefore X
= c(ad Y
)
q
X
+q
for some constant c, i.e. X
corresponds to
v
q
. Hence
[Y
, [X
, X
]] = (ad Y
)(ad X
)X
= q(N q + 1)X
= q(1 +p)X
.
This proves (4.6).
By a real form of a complex vector space W, we shall understand a real vector
space V , such that W = V
C
, i.e. W is the complexication of V .
Proposition 4.17. Let g be a complex semisimple Lie algebra and R(g, h) the
set of roots w.r.t. a Cartan subalgebra h. If h
0
denotes the 1-linear span of the
co-roots, then h
0
is a real form of the vector space h and all the roots are real
valued on h
0
.
Proof. Let be a root. By Corollary 4.13 we have
, = B(H
, H
) =
R
(H
)(H
) =
R
,
2
.
From Proposition 4.14 we get for each root integers p
and q
associated with
the -string containing , and these satisfy , =
1
2
(p
), . Substi-
tuting this into the expression above yields , =
R
1
4
(p
)
2
,
2
and thereby
, =
4
R
(p
)
2
.
Hence , is a rational number. Since , is a rational multiple of ,
(cf. Lemma 4.11) also , is rational.
By Lemma 4.11 7) the roots span h
and by non-degeneracy of the Killing

form the co-roots will span h. Let n be the complex dimension of h, we can there-
fore nd co-roots H
1
, . . . , H
n
which constitute a basis for h. Let
1
, . . . ,
n
denote the corresponding dual basis for h
. Set V := span
R
1
, . . . ,
n
. As
1
, . . . ,
n
is a complex basis for h
, V is a real form for h
. Furthermore all
the roots lie in V : indeed let be a root and write =
n
i=1
c
j
i
for c
j
C,
then
(H
j
) =
n
i=1
c
i
i
(H
j
) = c
j
and hence by the arguments above that c
j
is rational, in particular real. Thus
the roots span V .
Now put h
0
:= span
R
H
[ R. We have an isomorphism h

h given
by H
. This isomorphism obviously maps V to h

0
, thus h
0
is a real form
of h. For H =
n
i=1
c
i
H
i
h
0
we have real coecients c
i
and therefore
(H) =
n
i=1
c
i
(H
i
)
which is real since (H
i
) is rational as proved above.
Restricted to the real space h
0
the bilinear form , introduced in (4.4) (this
was just, in some sense, the dual of the Killing form) is a genuine inner product.
It is also positive denit for if h
0
is non-zero then
, = B(H
, H
) =
R
(H
f
)
2
and this is strictly positive by 8) of Lemma 4.11 since is real valued and since
the roots span h
0
.
Lemma 4.18. The space h
0
equipped with the bilinear form (4.4) is a real inner
product space.
Finally we will introduce a set of orthogonal transformations, known as root
reections: Let be a root and dene the map s
: h
0
h
0
by
s
() = 2
,
,
The action of s
on is simply that it reects in the hyperplane orthogonal

to , hence the name. Clearly, it is an orthogonal linear map, for s
() =
whereas for
we have s
() = . Thus by choosing a proper

orthonormal basis the matrix of s
is diag(1, 1, . . . , 1), i.e. s
is orthogonal.
Proposition 4.19. The root reections map R into R.
Proof. For , R we see (by Proposition 4.14) that
s
() = 2
,
,
= (p q) = + (q p)
where p and q are the integers given by Proposition 4.14. Since p q p q
we see that +(q p) is in the -string containing and hence it is in R0.
But as ,= 0 also s
() ,= 0, i.e. s
() R.
4.3 Abstract Root Systems
To be able to handle root systems in more abstract terms we introduce the notion
of an abstract root system and in the next section we dene the Weyl group.
The purpose of these sections is not to make a complete exposition of the vast
theory of abstract root systems and classication of semisimple Lie algebras,
rather it is intended to serve as a tool box providing the results necessary to
prove the Highest Weight Theorem in the next chapter.
Denition 4.20 (Abstract Root System). Let V be a real nite-dimensional
inner product space. By an abstract root system we understand a nite set R V
of non-zero elements satisfying
1) R spans V .
2) R is invariant under root reections, i.e. if , R then
s
() := 2
,
||
2
R.
3) If , R then 2
,)
||
2
is an integer.
Elements of R are called roots. The dimension of V is called the rank of the
root system.
The root system is called reduced if R implies 2 / R. An element R
is called reduced if
1
2
/ R.
Phrased in this new language a root system R(g, h) of a complex semisimple
Lie algebra w.r.t. a Cartan subalgebra is an abstract reduced root system in
h
0
(the inner product given by the dual of the Killing form): 1) follows from
4.3 Abstract Root Systems 69
Lemma 4.11, 2) follows from Proposition 4.19, 3) follows from Proposition 4.14,
and nally R(g, h) is reduced as a consequence of Proposition 4.12.
Let R V and R
t
V
t
be two abstract root systems. We say that the two
root systems are isomorphic if there is an orthogonal linear map : V V
t
which maps R to R
t
.
A root system R is called reducible if there is a decomposition R = R
t
R
tt
where any element of R
t
is orthogonal to all the elements of R
tt
. If not so, the
root system is called irreducible. It is a non-trivial fact, that the root system of
a complex semisimple Lie algebra is irreducible if and only if the Lie algebra is
simple.
In the next lemma we present some basic facts about abstract root systems
Lemma 4.21. Let R V be an abstract root system.
1) If R, then R.
2) If R, then 0,
1
2
, , 2 are the only possible elements in R 0
which are proportional to . If is reduced, only 0, and 2 are
possible. If R is reduced, then only 0 and are possible.
3) If R and R 0, then the integer 2
,)
||
2
equals either 0, 1,
2, 3 or 4. The possibility 4 only occurs when R is non-reduced and
= 2.
4) If , R are non-proportional and || ||, then 2
,)
||
2
equals 0 or
1.
5) Let , R. If , > 0, then R 0. If , < 0, then
+ R 0.
Proof. 1) This follows since
R s
() = 2
,
||
2
= .
2) By 1) we have that 0 and are elements in R 0 proportional to
and if R is non-reduced then also 2 are possible elements in R 0.
Now assume c R for some non-zero real number c, then
2
c,
||
2
and 2
, c
|c|
2
are integers. This implies that 2c and
2
c
are integers. The last condition says
that c is of the form c =
2
n
for n Z 0 whereas the rst condition further
reduces the possible values to 2, 1,
1
2
,
1
2
, 1, 2. If is reduced, c =
1
2
cannot
occur.
3) If = 0 then 2
,)
||
2
= 0, so assume ,= 0. From the Cauchy-Schwarz
inequality we have
2
,
||
2
2
,
||
2

4
||
2
||
2
|||||||| = 4.
If 2
,)
||
2
is non-zero, then also 2
,)
||
2
is non-zero and as their product is less
than 4, both integers must have absolute value at most 4. If and are non-
proportional the inequality is strict, hence the integers have absolute value at
most 3. Therefore for 4 to occur we must have = c and from 2) we see
that the only possibility is = 2. This, of course, can only happen if R is
non-reduced.
4) Again, since the inequality is strict, we get
2
,
||
2
=
2
||
2
[, [ <
2
||
2
|||| = 2
||
||
2
and so the conclusion follows.
5) If and are proportional, i.e. = c then from 2) we know that c
can only be
1
2
, 1 or 2. If , < 0 then c < 0. For c = 2 we have
+ = R. If c = 1, then + = 0 R 0, and if c =
1
2
, then
+ =
1
2
= R. For , we do exactly the same.
Now suppose and are non-proportional. W.l.o.g. we may assume ||
||, then by 4) 2
,)
||
2
is either 0 or 1. If , > 0 then we must have
2
,)
||
2
= 1 and consequently s
() = + R. Conversely, if , < 0 then

2
,)
||
2
= 1 and s
() = R.
For the rest of this section we will introduce the two closely connected notions
of a positive system and a fundamental system for a root system in V and show
how they are related. For this purpose we divide V into half spaces. By an open
half space of V we mean a subset of V of the form v V [ v, > 0 where
is some xed non-zero vector.
Denition 4.22 (Positive System). Let R be a root system in V . By a
positive system or a system of positive roots we understand a subset R
+
R
satisfying
1) There exists an open half space in V containing R
+
,
2) R = R
+
(R
+
).
The elements of R
+
are called positive roots.
If R is a root system and R
+
a system of positive roots, then a root is called
simple if it positive and if it is not possible to express it as a sum of positive
roots.
Lemma 4.23. If and are distinct simple roots, then is not a root.
Consequently , 0. Geometrically this amounts to say that the angle be-
tween and is at least /2.
Proof. Assume to be a root. If it is positive, then = ()+ is a sum
of positive roots. If it is not positive, then is positive and = ( ) +
is a sum of positive roots. In any case we get a contradiction, so is not a
root. The last claim is an immediate consequence of Lemma 4.21 5).
Now for the related notion
Denition 4.24 (Fundamental System). Let R be a root system in V . A
subset R is called a fundamental system (or a simple system or a basis) for
R if
1) is a basis for V ,
2) R N
0
(N
0
).
An equivalent formulation of the last condition is that any root is expressible
as a linear combination =
n
i
i
where
i
and the coecients n
i
are
all either non-positive or non-negative integers. If =
n
i
i
is positive then
n
i
is a positive integer called the level of w.r.t. the positive system R
+
.
4.3 Abstract Root Systems 71
Proposition 4.25. Let R be a root system in V . If R
+
is a positive system,
then the set of simple roots is a fundamental system. Conversely, if is a
fundamental system, then R
+
:= N
0
R is a positive system.
Proof. We show that is a fundamental system. Let be a positive root.
Either is a simple root or it can be decomposed = + . Now and
are either simple roots or they can be decomposed into sums of positive roots,
etc. And so we continue until is written as a N
0
-linear combination of simple
roots. Thus R
+
N
0
and therefore also R = R
+
(R
+
) N
0
(N
0
),
and as R spans V also is a spanning set. To see that it is linearly independent,
let , be distinct. Then neither nor are in R
+
, hence is
not a root. By Lemma 4.21 5) , 0. is contained in an open half space
meaning that there exists a vector such that , > 0 for all . Assume
= 0 and put
:= [
> 0 and e
:=
[,
where e
is dened to be 0, if
is empty. We see that e

+
= e
. Since
+
and
are disjoint , 0 if
+
and
. From this it follows that

e
+
, e
+
= e
+
, e
0, (as e
+
, e
can be written as a sum of elements of the

form [
[, with
+
and
) and consequently e
+
= e
= 0.
Therefore
0 = , e
[, .
Since , is strictly negative, we must have
= 0 for all . Hence is

a linearly independent set and is thus a fundamental system for R.
Now let =
1
, . . . ,
n
be a fundamental system. To see that it is contained
in an open half space put :=
1
+ +
n
. Then we have
i
, > 0 for all
i = 1, . . . , n and therefore that thus N
0
hence R
+
is contained in an open half
space. We obviously have R
+
(R
+
) R. Conversely let R. By property
2) of a fundamental system is of the form =
n
i
i
where n
i
are all either
non-positive or non-negative. But that exactly means that R
+
(R
+
).
Positive systems, and consequently also fundamental systems, do exist. One
way of dening a positive system by virtue of the so-called lexicographic order-
ing: Let R be a root system in V and v
1
, . . . , v
n
a spanning set for V . We say
that R is positive if there exists a 1 k n such that , v
i
= 0 for i < k
and , v
k
> 0. As one can check this is a positive system and consequently:
any root system has positive system and thus also a fundamental system.
Example 4.26. Lets consider the Lie algebra sl(3, C). It is conceptually easy
(though tedious) to verify that the following subspace
h :=
_
_
_
h
1
0 0
0 h
2
0
0 0 h
3
_
_
h
1
, h
2
, h
3
C, h
1
+h
2
+h
3
= 0
_
is in fact a Cartan subalgebra of sl(3, C). Obviously, it has dimension 2, and
a basis is given by H
1
:= diag(1, 1, 0) and H
2
:= diag(0, 1, 1). Dening for
i = 1, 2, 3 the functional e
i
h
by
e
i
_
_
h
1
0 0
0 h
2
0
0 0 h
3
_
_
= h
i
it can be seen that the roots of sl(3, C) w.r.t. the Cartan subalgebra h are just
e
i
e
j
for ij. This set of roots is given the ordering that e
i
e
j
is positive
when i < j, i.e. the positive roots are e
1
e
2
, e
1
e
3
and e
2
e
3
. Since
e
1
e
3
= (e
1
e
2
) + (e
2
e
3
)
we see that e
1
e
2
and e
2
e
3
are simple roots. We want now to calculate the
Killing form and the co-roots. Calculating the Killing form is easy thanks to
Corollary 4.13 and we get for the basis elements
1
:
[H
1
[
2
= B(H
1
, H
1
) = 12
B(H
1
, H
2
) = 6
[H
2
[
2
= B(H
2
, H
2
) = 12.
The co-root of e
1
e
2
is the unique vector H
e1e2
satisfying (e
1
e
2
)(X) =
B(H
e1e2
, X) for all X h. Writing H
e1e2
= aH
1
+ bH
2
, this condition with
X = H
1
gives us the equation 2 = 12a 6b, and when X = H
2
the condition
reads 1 = 6a+12b. This system of equations has the solution a =
1
6
and b = 0,
hence H
e1e2
=
1
6
H
1
. In a similar fashion we calculate that H
e2e3
=
1
6
H
2
.
Since e
1
e
3
was just (e
1
e
2
) + (e
2
e
3
) the corresponding co-root is just
1
6
(H
1
+H
2
). The remaining 3 roots are just the negatives of these three, hence
we have calculated all the co-roots.
4.4 The Weyl Group
Denition 4.27 (Weyl Group). Let R be a root system in V . The subgroup
W(R) (or just W for short) of O(V ) generated by the root reections s
for
R is called the Weyl group of the root system. If R is the root system of a
Lie algebra g w.r.t. a Cartan subalgebra h we denote it W(g, h).
The Weyl group is a nite group: Indeed any element of the Weyl group maps
R to R and if two elements agree on the roots, they agree on a spanning set,
and hence they are equal. Since R is nite, there are only nitely many dierent
orthogonal transformations of V mapping R into R, hence the group is nite.
If r is any orthogonal linear map on V we see
s
r
() = 2
, r
|r|
2
r = 2
r
1
,
||
2
r
= r
_
r
1
2
r
1
,
||
2

_
= rs
(r
1
)
In particular, if = r then
s
= rs
r
1
. (4.7)
Lemma 4.28. Let R be a root system and =
1
, . . . ,
n
the set of simple
roots. If is a positive root proportional to
i
(i.e. =
i
or = 2
i
) then
s
i
() = , otherwise s
i
() is positive.
Proof. As is positive =
i
n
i
i
where n
i
N
0
. If =
i
then of course
s
i
() = . Similarly, if = 2
i
, one calculates that s
i
() = . If is not
proportional to
i
, then there exists a j ,= i such that n
j
> 0. But then the jth
coecient in
s
i
() =
n
i=1
n
i
i
2
,
i
|
i
|
2

i
1
In light of our notion of positivity of roots we see that we can simplify our formula to
B(H, H
) = 2
R
+ (H)(H
).
4.4 The Weyl Group 73
equals n
j
> 0. Hence by the property of all coecients are automatically
positive and thus s
i
() is a positive root.
Proposition 4.29. If R is a root system and =
1
, . . . ,
n
the simple
elements, then s
[ is a set of generators for W(R). Furthermore if

is reduced, then there exists
j
and s W(R) such that = s(
i
).
Proof. We prove the theorem backwards. Let W
t
W be the subgroup gen-
erated by the root reections of the simple roots. Let =
n
i
i
be a reduced
root, and assume it to be positive. We do induction on the level of . If the level
of equals 1, then =
i
for some i, and we can pick the identity element id
of W
t
, so that = id
i
. Now let the level of be strictly greater than 1 and
assume the result to hold for any root which is of a strictly lower level. We have
0 < ||
2
=
n
i
,
i
hence there must exist an index i

0
such that ,
i0
> 0. We cannot have
=
i0
(since the level was greater than 1) and we cannot have = 2
i0
(since
was reduced). By the preceding lemma we get that := s
i
0
is positive. But
as
=
i,=i0
n
i
i
+
_
n
i0
2
,
i0
|
i0
|
2
_
i0
and since ,
i0
> 0, is of a strictly lower level than . Hence by the
induction hypothesis = s(
j
) with s W
t
and = s
1
i
0
() = s
i
0
s(
j
).
Since s
i
0
s W
t
we have the desired result.
Now let R. We can assume to be reduced for if not,
1
2
will be reduced
and a quick calculation reveals that s
= s1
2
. Thus = s(
j
) for s W
t
. By
(4.7) s
= ss
j
s
1
, hence s
W
t
. As all the generators of W lie in W
t
the
two groups are equal.
Proposition 4.30. Let R be a root system and and
t
two simple systems
for R. Then there exists s W(R) such that
t
= s.
Proof. Let R
+
and R
+
t
denote the sets of positive roots w.r.t. resp.
t
.
Then we have [R
+
[ = [R
+
t
[ =
1
2
[R[ =: q and of course we have R
+
= R
+
t
if and
only if =
t
.
Now put r := [R
+
R
+
t
[ = q n for some n N
0
. We verify the claim of
the proposition by induction over n. If n = 0 then R
+
= R
+
t
and consequently
=
t
and we can choose s = id W.
Assume then that n > 0, that is R
+
,= R
+
t
. As generates R
+
we cant
possibly have R
+
t
for then R
+
R
+
t
and hence they would be equal. Thus
we can pick an which is not in R
+
t
. But then R
+
t
. If R
+
R
+
t
,
then by Lemma 4.28 s
() R
+
(for as / R
+
t
, cannot be proportional to
), i.e. s
() R
+
s
(R
+
t
). Consequently R
+
s
R
+
t
contains at least q n
elements. But also R
+
s
R
+
t
. Thus [R
+
s
R
+
t
[ q (n 1), and so,
as s
R
+
t
has s
t
as fundamental system, the induction hypothesis yields an
element s W so that s
t
= s, i.e.
t
= s
s.
Denition 4.31 (Dominant Element). An element V is called dominant
w.r.t. a fundamental system if , 0 for all (and hence also for all
R
+
).
Proposition 4.32. For any V there is a fundamental system with respect
to which is dominant.
Proof. If = 0 then it is obviously dominant, so we assume ,= 0. Put
v
1
:= and complete this to an orthogonal basis v
1
, . . . , v
n
for V . Via the
lexicographic ordering this yields a fundamental system =
1
, . . . ,
n
. By
denition we have, in particular, that
i
, v
1
0 for all i and thus = v
1
is
dominant w.r.t. .
Proposition 4.33. Let R V be a root system and R
+
a positive system. If
V , then there exists s W(R) such that s() is dominant.
Proof. Since R has a positive system it has a fundamental system . According
to Proposition 4.32 there is a fundamental system
t
w.r.t. which is dominant.
Proposition 4.30 yields an element s W which maps to
t
, i.e. if =
1
, . . . ,
n
then
t
= s
1
, . . . , s
n
. Since
s
1
,
i
= , s
i
0
we see that s
1
is dominant.
The nal result of this chapter is
Proposition 4.34. Let R be a reduced root system and dene :=
1
2
R
+ .
If is a simple root then s
() = and consequently we have

2
,
||
2
= 1. (4.8)
Proof. From Lemma 4.28 we have that s
maps positive roots to positive roots

with the single exception of which by s
is mapped to (as R is reduced

are the only roots proportional to ). We see
s
(2) = s
(2 ) +s
().
2 is just the sum of all positive roots except for . By the remarks above,
s
just permutes the positive roots, i.e. s
() = and thus s
(2 ) = 2 .
Therefore
s
(2) = (2 ) = 2( )
and the conclusion is reached.
Chapter 5
The Highest Weight Theorem
5.1 Highest Weights
The main result of this chapter will be, as promised, the Theorem of the Highest
Weight a classication theorem for the irreducible representations of complex
semisimple Lie algebras. Apart from its importance in representation theory, it
plays a prominent role in particle physics, in that the bound states of quarks
known as hadrons are modeled by representations with certain highest weights.
At rst we set out to dene an ordering of the roots. The rst step in this
process was the construction of the real form for the Cartan subalgebra carried
out in Proposition 4.17. The next step is to introduce the notion of a Weyl
chamber: Let g be a complex semisimple Lie algebra and R(g, h) its root system
relative to the Cartan subalgebra h and let h
0
be the real form of h. As we have
R h
0
, and thus for each R the kernel of is a hyperplane in h
0
, and the
connected components of h
0
R
ker (which are open and nite in number)
are called the Weyl chambers of the root system R.
Now pick a Weyl chamber ( and let R be an arbitrary root. Then is
non-zero and real on ( and since ( is connected, is either strictly positive
or strictly negative. The set of roots which are positive on ( are called positive
roots relative to the Weyl chamber ( and the set is denoted R
+
(() or just R
+
.
This is a positive system as dened in the preceding chapter for let x ( be
nonzero and let be the functional in h
0
corresponding to x through the inner
product , . Then for R
+
we have , = (x) > 0, so the positive
roots are contained in an open half space of h
0
. Furthermore Lemma 4.11 3)
says that R = R
+
(R
+
) and hence R
+
is a positive system. We denote the
corresponding set of simple roots by .
For two roots and we write if is a positive root. This is the
promised ordering of the roots.
Lemma 5.1. Let N
0
R
+
be the set of linear combinations of positive roots with
nonnegative integer coecients, then N
0
R
+
(N
0
R
+
) = 0.
Proof. If N
0
R
+
then 0 on the Weyl chamber (. If (N
0
R
+
),
then likewise 0 on (. Since the Weyl chamber is an open subset of h
0
then
is zero on ( and hence zero as a linear functional.
Lemma 5.2. The spaces
g
+
:=
R
+
g
and g
:=
R
+
g
are Lie subalgebras of g and we have the decomposition

g = g
+
h g
. (5.1)
75
76 Chapter 5 The Highest Weight Theorem
Proof. We will show that g
+
is a subalgebra of g. Let , R
+
and assume
[g
, g
] ,= 0. This means that + is a root and as it is obviously positive on

( then + R
+
, hence [g
, g
] g
+
, so its a subalgebra. Now (4.3) readily
translates into (5.1).
Now we can dene the important notion of a highest weight and a highest
weight vector.
Denition 5.3 (Highest Weight Vector). By a highest weight vector for a
Lie algebra representation (, V ) of g we understand a nonzero weight vector
v V
(relative to some Cartan subalgebra h) such that (X)v = 0 for all

X g
+
. A weight whose weight space contains a highest weight vector is called
a highest weight.
Well, do such objects exist at all? Yes they do:
Proposition 5.4. If (, V ) is a nite-dimensional representation of g, then a
highest weight vector and thus a highest weight exist.
Proof. Let ( h
0
be the Weyl chamber determining R
+
and pick an element
X
0
(. Then (X
0
) > 0 for all R
+
. As () is nite, there is a weight
such that Re((X
0
)) is maximal. Then + cant be a weight when R
+
,
i.e. V
+
= 0. By Proposition 4.9 we get that (g
)V
= 0 for all R
+
so that (g
+
)V
= 0. Therefore any nonzero v V
will be a highest weight

vector with highest weight .
For a general representation we cannot hope for uniqueness of the highest
weight, a representation can have several highest weights. However for irre-
ducible representations we do have uniqueness. This is the rst step towards
classication of irreducible representations through their highest weights. The
rest will follow in the sequel.
To prove uniqueness we will introduce the following technical tool which is
used only in proofs.
Denition 5.5 (Cyclic Vector). A highest weight vector v for a representation
(, V ) is called cyclic if the only -invariant subspace containing v is V itself.
Lemma 5.6. Let (, V ) be a nite-dimensional representation of g and let v V
be a cyclic highest weight vector for associated with the weight .
1) The weight space is 1-dimensional : V
= Cv.
2) V = Cv span(X
1
) (X
n
)v [X
k
g
, n N.
3) Every weight () is of the form for N
0
R
+
.
Proof. We prove rst 2). We dene V
0
:= Cv and inductively V
n+1
:= V
n
+
(g
)V
n
and let W be the union of all the V
n
s. It should be clear that W
actually equals Cv span(X
1
) (X
n
)v [X
k
g
, n N, so we need to
show that W = V . This we can accomplish by showing that W, which contains
v, is invariant. Since we have the decomposition (5.1) it is enough to show that
W is g
- and h-invariant (meaning that (X)W W for X g
or X h).
Since we have, by construction that (g
)V
n
V
n+1
, it is clear that W is
g
-invariant. That was the easy part. To show the other invariances we use
induction. As v is a weight vector, then (h)V
0
V
0
and (g
+
)V
0
= 0, in
particular V
0
is h- and g
+
-invariant. Now assume that V
n
is h- and g
+
-invariant.
As V
n+1
= V
n
+ (g
)V
n
it is enough to show that (h)((g
)V
n
) V
n+1
and
(g
+
)((g
)V
n
) V
n+1
. For the rst one, let H h, Y g
and w V
n
. Then
(H)(Y )w = (Y )(H)w +([H, Y ])w.
5.1 Highest Weights 77
By induction we had that V
n
was h- and g
+
-invariant, implying that (H)w V
n
and ([H, Y ])w V
n+1
(the last following from [H, Y ] g
), hence (Y )(H)w
V
n+1
, thus (H)(Y )w V
n+1
. Therefore W is h-invariant.
Now let w V
n
, X g
+
and Y g
. Then again
(X)(Y )w = (Y )(X)w +([X, Y ])w.
As V
n
is g
+
-invariant, we have (X)w V
n
, so that (Y )(X)w V
n+1
. Since
V
n
is h- and g
+
-invariant and (g
)V
n
V
n+1
then the decomposition (5.1)
says that (g)V
n
V
n+1
, i.e. ([X, Y ])w V
n+1
. This proves the induction step
and hence 2).
Now we prove 3). Lets consider the vector w = (X
1
) (X
n
)v for X
j

g
j
with
j
R
+
. For H h we then see that
(H)w = (H)(X
1
) (X
n
)v
=
n
j=1
((X
1
) ([H, X
j
]) (X
n
)v) +(X
1
) (X
n
)(H)v
=
_
n
j=1
j
(H)
_
(X
1
) (X
n
)v +(H)(X
1
) (X
n
)v
= ( )(H)w
where =
1
+ +
n
N
0
R
+
. Thus if w ,= 0 then is a weight. Now if w is
an arbitrary weight vector with weight it must, by 2) be a linear combination
of elements of the form as above. These are all weight vectors and since dierent
weight spaces are linearly independent they must be weight vectors of the same
weight. Thus = for some as above. This proves 3)
Now for 1). Let w V
and assume for contradiction that w / Cv. Then if

w = (X
1
) (X
n
)v for n > 0 and a calculation as above shows that (H)w =
(
1

n
)(H)w. If w is a linear combination of such elements they must
all be weight vectors of the same weight where ,= 0 since w / Cv. Thus we
cannot possibly have (H)w = (H)w for all H h, hence a contradiction.
With this technical lemma at our disposal we can prove the rst part of the
Theorem of the Highest Weight
Proposition 5.7. Let (, V ) be an irreducible representation of g. Then has
a unique highest weight . If v V
is non-zero, then the following hold

1) V
= Cv.
2) V = Cv span(X
1
) (X
n
)v [X
k
g
, n N.
3) Every weight () is of the form for N
0
R
+
.
4) For R
+
and each X g
we have (X)v = 0.
Proof. By Proposition 5.4 a highest weight and highest weight vector v for
exist. Since is irreducible it is automatically a cyclic highest weight vector
and by the previous lemma, 1)-3) are valid.
Assume that is another highest weight for . By 3), since is a highest
weight, = for N
0
R
+
. Similarly, as is a highest weight =
t
for
t
N
0
R
+
. Thus we both have =
t
N
0
R
+
and = N
0
R
+
,
hence = 0 by Lemma 5.1.
The proof of point 4 is easy: (X)V
V
+
and since is positive + is
strictly higher than and therefore V
+
= 0.
Note that 3) justies the name highest weight, indeed it is higher than any
of the other weights. It also shows that the highest weight is independent of the
ordering of the roots, i.e. independent of the choice of Weyl chamber. It is also
worth mentioning that the weight spaces for non-highest weights need not be
1-dimensional.
As a partial converse to point 4) of the proposition above we have
Proposition 5.8. Let (, V ) be a representation with highest weight . If v V
satises (X)v = 0 for each X g
with R
+
, then v V
.
Proof. Assume for contradiction that v / V
satises that (X)v = 0 for each

root vector X g
with R
+
. Without loss of generality we can assume that
v has no component in V
. Now consider the weight spaces in which v has a

component and let
0
be the highest of these weights.
0
is (by uniqueness of the
highest weight for irreducible representations) strictly lower than . As in the
proof of Lemma 4.11 we see that Cv span(X
1
) (X
n
)v [X
k
g
, n N
is a non-trivial invariant subspace of V and hence by irreducibility equals V .
But all the weight vectors are associated with weights which are lower than
0
hence they are all strictly lower than and thats a contradiction.
Earlier we saw that weights are preserved under equivalence of representa-
tions. The same is true for highest weights: equivalent representations have the
same highest weights. This is not hard to see: Let (, V ) and (
t
, V
t
) be repre-
sentations which are equivalent through an intertwiner T : V

V
t
. If is a
highest weight for , then (g
+
)V
= 0. Now was also a weight for

t
with
weight space T(V
). Since
t
(g
+
)(TV
) = T((g
+
)V
) = 0,
is also a highest weight for
t
. A similar argument works for the other direction.
For irreducible representations the situation is interesting. Not only will two
equivalent irreducible representations have the same (unique) highest weight,
also the converse is true: two irreducible representations having the same highest
weight are equivalent.
Proposition 5.9. Let (, V ) and (
t
, V
t
) be two irreducible representations of
a complex semisimple Lie algebra g. If they have the same highest weight, then
they are equivalent.
Proof. Let v and v
t
be non-zero highest weight vectors for and
t
respectively.
Form the subspace
S := C(v, v
t
) span(
t
)(X
1
) (
t
)(X
n
)(v, v
t
) [ X
i
g, n N
By the same arguments as of Lemma 5.6 this space equals
C(v, v
t
) span(
t
)(X
1
) (
t
)(X
n
)(v, v
t
) [ X
i
g
, n N
Obviously this is a
t
-invariant subspace of V V
t
and we will now show
that (
t
)[
S
is irreducible. To this end let T S be a nontrivial invariant
subspace on which
t
is irreducible. This exists by Weyls Theorem. Then
(
t
)[
T
has a unique highest weight by Proposition 5.7. Let (v
0
, v
t
0
) ,= 0 be
an associated highest weight vector. Now for each X g
with R
+
we have
by Proposition 5.7 4) that
0 = (
t
)(X)(v
0
, v
t
0
) = ((X)v,
t
(X)v
t
),
i.e. (X)v =
t
(X)v
t
= 0 and by the same proposition we have v
0
= cv and
v
t
0
= c
t
v
t
with c, c
t
C. Therefore we have (v, v
t
) T and hence also in S. But
5.2 Verma Modules 79
since v
0
and v
t
0
are highest weight vectors and the X
i
s in g
push the weights

down we must have (v
0
, v
t
0
) C(v, v
t
). Proposition 5.7 says that
T = C(v
0
, v
t
0
) span(
t
)(X
1
) (
t
)(X
n
)(v
0
, v
t
0
) [ X
i
g
, n N
but since (v
0
, v
t
0
) C(v, v
t
) we see that T = S and thus that (
t
)[
S
is
irreducible.
Now it is not hard to see that the projection
1
: V V
t
V intertwines
(
t
)[
S
and and that
2
: V V
t
V
t
intertwines (
t
)[
S
and
t
. Since
these are all irreducible the are all mutually equivalent by Schurs Lemma, in
particular and
t
are equivalent.
5.2 Verma Modules
Lets introduce an important class of functionals in h
Denition 5.10 (Integral Element). An element h
is called an integral
element if the number
2
,
,
is an integer for all roots R.
By Proposition 4.14 all roots are integral elements. The importance of integral
elements is due to the following
Proposition 5.11. Let be a complex representation of a complex semisim-
ple Lie algebra g, then the weights of w.r.t. a Cartan subalgebra are integral
elements. The highest weights of are dominant integral elements.
Proof. Let R, and consider the Lie subalgebra sl
= spanX
t
, Y
t
, H
t
which is isomorphic to sl(2, C). From Corollary 3.13 we know that (H

t
) has
only integer eigenvalues. Since 0 ,= v V
implies
(H
t
)v = (H
t
)v = 2
(H
)
,
v = 2
,
,
v,
2
,)
,)
is an eigenvalue for (H
t
), and must therefore be an integer.

Let be a highest weight and v V
a nonzero weight vector. Let further-

more be a simple root and consider the subspace sl
= spanH
t
, X
t
, Y
t
which is isomorphic to sl(2, C). Let W be the span of elements of the form
(Y
t
)
n1
(H
t
)
n2
(X
t
)
n3
v. Since v is a highest weight vector, W equals the
span of the elements (Y
t
)
n
v. But on elements of this form (H
t
) acts by the
eigenvalues
( n)(H
t
) = 2
,
||
2
2n.
Obviously 2
,)
||
2
is the greatest of these eigenvalues and from Theorem 3.8 this
greatest eigenvalue has to be non-negative.
Our nal task concerning the Highest Weight Theorem is to construct for
each dominant integral element in h
an irreducible representation of g having

the integral element as highest weight.
We retain the notation g
+
=
R
+ g
and g
R
+ g
and dene
b := h g
+
and :=
1
2
R
+ . Before we proceed we need the following
lemma linking the universal enveloping algebra of g to representation theory.
Proposition 5.12. Let g be a complex Lie algebra. There is a 1-1 correspon-
dence between (possibly innite-dimensional ) complex representations of g and
unital left modules over U(g).
By a unital module we mean a U(g)-module such that 1v = v.
Proof. Let : g End
C
(V ) be a complex representation of g. Then by the
universal property it factorizes through U(g) to a unital algebra homomorphism
: U(g) End
C
(V ), so V is given the structure of a unital left U(g)-module
by uv = (u)v.
Conversely, if V is a left U(g)-module, then it is a complex vector space since C
sits in U(g), and we dene a representation by (X)v = (X)v ( : g U(g)
is the embedding of g). These two constructions are easily seen to be inverses
of each other.
Let V be a complex vector space and a U(g)-module. Referring to Proposition
5.12, to V corresponds a unique complex representation of g on V and therefore
we already have the notions of irreducibility, weights, weight vectors and weight
spaces for U(g) modules (relative to some Cartan subalgebra of g). We let (V )
denote the set of weights. A direct translation of Proposition 4.9 tells us that
g
V
+
when (V ). Hence
(V )
V
is a g-invariant subspace of V .
In the same spirit we dene a highest weight vector for V to be a nonzero weight
vector v V
for some such that g

+
v = 0. A weight whose weight space
contains a highest weight vector is called a highest weight. If v V is a highest
weight vector for some U(g)-module the highest weight module generated by v
is the U(g)-submodule U(g)v of V .
The following lemma on highest weight modules is somewhat an analog to
Proposition 5.7. The only (but important) dierence is that we now consider
innite-dimensional representations/modules as well.
Lemma 5.13. Let V be a U(g)-module, v V be a highest weight vector as-
sociated with highest weight and let W = U(g)v be the highest weight module
generated by v. Then we have
1) W = U(g
)v.
2) W =
h
W
where dimW
< and dimW
= 1.
3) Every weight of W is of the form
n
i=1
n
i
i
with
i
R
+
and n
i
N
0
.
Proof. 1) We have that g = g
h g
+
and from the PBW-Theorem we
see that any element of U(g) is a (a linear combination) of elements of the
form Y HX where X U(g
+
), H U(h) and Y U(g
). Since we have
U(g
+
)

= g
+
U(g) C (by the PBW-Theorem) and Xv = 0 for X g
+
U(g
+
)
(since v is a highest weight vector) we get U(g
+
)v = Cv. Similarly, Hv is just a
constant times v, so U(h)v = Cv. Thus, only elements in U(g
) give something
new and therefore U(g)v = U(g
)v.
2) and 3) We clearly have

W
W. By Proposition 4.9 we get for

R0 that g
W
+
and hence (by the PBW-Theorem) U(g)(
. As v
we see
W = U(g)v U(g)(
.
Thus W =
.
U(g
) is generated by X
[ R
+
and hence a basis for W = U(g
)v is
X
n1
1
X
n
k
k
v. In particular any weight vector must be of this form and for
H h we see
H(X
n1
1
X
n
k
k
v) = ( n
1
1
n
k
k
)(H)v.
This proves 3). Observe that for any given only nitely many combinations
give and therefore dimW
< . There is only one possibility to get , and

thus dimW
= 1, in fact W
= Cv. This proves 2).

Certain innite-dimensional modules, called Verma modules are necessary to
construct the irreducible representations we seek. The construction of the Verma
module goes as follows: Let V
1
and V
2
be two complex vector spaces and let A
and B be complex, associative unital algebras. Assume furthermore that V
1
is a
right B-module, V
2
is a left B-module and that V
1
is a left A-module such that
(av)b = a(vb) for all a A, b B and v V
1
.
Denote by I the two-sided ideal in V
1
C
V
2
generated by all elements of the
form v
1
b v
2
v
1
bv
2
, and dene the tensor product
V
1
B
V
2
:= (V
1
C
V
2
)/I.
What we do is that we identify v
1
b v
2
with v
1
bv
2
, one might say we have
made it associative. The equivalence class in V
1

B
V
2
containing v
1
v
2
will
still be denoted v
1
v
2
, now we just have the above identication.
V
1
B
V
2
is given the structure of an A-module by dening a(v
1
v
2
) = av
1
v
2
.
With this V
1
B
V
2
has the following universal property:
Proposition 5.14. Let W be a complex vector space and : V
1
V
2
W a
bilinear map satisfying (v
1
b, v
2
) = (v
1
, bv
2
), then there exists a unique linear
map

: V
1
B
V
2
W such that (v
1
, v
2
) =

(v
1
v
2
).
Proof. The proposition is easily proved when considering the following com-
muting diagram
V
1
V
2

L
L
L
L
L
L
L
L
L
L
V
1
V
2
V
1
B
V
1
.q
q
q
q
q
q
q
q
q
q
q
W
As is bilinear it descends uniquely to the linear map
t
, which is 0 on the
ideal I since
t
(v
1
b v
2
) = (v
1
b, v
2
) = (v
1
, bv
2
) =
t
(v
1
bv
2
) and therefore
descends uniquely to

such that (v
1
, v
2
) =

(v
1
v
2
).
Let h
be arbitrary. We dene a representation of b = h g

+
on C by
(H)z = ( )(H)z for H h
(X)z = 0 for X g
+
.
and this gives C the structure of a U(b)-module. If we want to stress the module
structure of C we write it as C
. Now multiplication turns U(g) into both a

left U(g)-module and a right U(b)-module, and therefore it makes sense to dene
the Verma module associated with :
V () := U(g)
U(b)
C
. (5.2)
This is an innite-dimensional left U(g)-module and thus corresponds to some
innite-dimensional representation of g.
In the following proposition we outline some properties of the Verma module
Proposition 5.15. Let h
be arbitrary.
1) The Verma module V () is a highest weight module generated by 1 1
(called the canonical generator of V ()) which is a highest weight vector
with weight .
2) The map U(g
) V () given by u u(1 1) is a linear bijection.

3) If M is another highest weight module over U(g) generated by a highest
weight vector v of weight , then there exists a unique U(g)-module
homomorphism : V () M with (11) = v. This map is surjective
and it is injective if and only if u ,= 0 in U(g
) implies uv ,= 0 in M.
Proof. 1) Since by denition V () = U(g)
U(b)
C
it should be clear from

the module structure that V () = U(g)(1 1). For X g
+
and 1 the unit in C
we have that X 1 = 0 (this is how we dened the module structure on C) and
therefore
X(1 1) = (X 1) 1 = (1 X) 1 = 1 X 1 = 0
(in the third equality we used that X b U(b)). Thus 1 1 is a highest
weight vector. For H h we have
H(1 1) = 1 H 1 = 1 ( )(H) = ( )(H)(1 1)
and therefore 1 1 is a weight vector of weight .
2) Since g = g
b we have from Corollary 2.43 a vector space isomorphism

U(g)

= U(g
)
C
U(b). If X = X
1
X
k
X
k+1
X
n
where X
1
, . . . , X
k
g
and X
k+1
, . . . , X
n
b the isomorphism is simply given by X X
1
X
k

X
k+1
X
n
(and extended by linearity of course). We now get a string of vector
space isomorphisms
V () = U(g)
U(b)
C

= (U(g
)
C
U(b))
U(b)
C
= U(g
)
C
(U(b)
U(b)
C
= U(g
)
C
C
= U(g
)
given by composition of the maps
u(1 1) u 1 (u 1) 1 u (1 1) u 1 u.
3) First we dene a map : U(g) C
M by (u, z) u(zv). For

X g
+
or X h we see
(uX, z) = uX(zv) = zu(Xv)
whereas
(u, Xz) = u((Xz)v) = zu(Xv)
i.e. (uX, z) = (u, Xz) for all X U(b). By the universal property of
U(b)
(Proposition 5.14) there is a unique map : U(g)
U(b)
C
M satisfying
(u z) = (u, z) = u(zv). Phrased a little dierently, is the unique map
satisfying (u(11)) = (u1) = uv. Thus is a U(b)-module homomorphism
and (1 1) = v, and existence and uniqueness is veried.
Since M is generated by v, any element in M is of the form uv for some
u U(g), and from this it follows that is surjective.
Now assume that uv = 0 for some nonzero u U(g) then (u(1 1)) =
uv = 0 and as u(1 1) is nonzero, is not injective. Conversely, assume that
u ,= 0 implies uv ,= 0. Since V () is, by Lemma 5.13 1), generated by the
elements u(1 1) for u U(g
), then u ,= 0 implies u(1 1) is nonzero and

(u(1 1)) = uv ,= 0. Hence the only element which is mapped to zero by is
0, and hence is injective.
Another way of formulating property (3) of this proposition is that any highest
weight module with a certain weight is a quotient of the Verma module with
that same weight. Thus the Verma modules are in some sense the biggest among
the highest weight modules.
We continue to let h
and put
V ()
+
=
,=
V ()
.
We claim that any proper U(g)-submodule of V () is contained in V ()
+
: By
Lemma 5.13 V ()
= C(1 1), so if a submodule contains V ()
then it
contains the canonical generator 1 1 and hence equals all of V ().
Let S denote the sum of all proper submodules of V (). Obviously this is a
submodule contained in V ()
+
, so we can form the quotient L() = V ()/S
which is equipped with the module structure X[v] = [Xv] (which is well-dened
as S is a submodule) for v V (). Let q : V () L() denote the quotient
map. L() is an irreducible module, for if W L() is a proper submodule then
q
1
(W) is a proper submodule of V () and hence contained in S, i.e. W = 0.
Finally if X g
+
and H h then we have [1 1] ,= 0 and X[1 1] =
[X(11)] = 0 and H[11] = [H(11)] = [()(H)11] = ()(H)[11]
hence we have proved
Proposition 5.16. L() is an irreducible highest weight module over U(g) and
[1 1] is a highest weight vector with highest weight .
Thus if the Verma module V () is the biggest highest weight module with
highest weight , then L() is the smallest highest weight module with
weight and this one is irreducible. Hence, for each h
we can produce
an irreducible (possibly innite-dimensional) representation of g
1
.
The nal task is to seek out those which are actually nite-dimensional. We
will show that if is an integral element which is real on h
0
then L( + ) is
actually nite-dimensional and hence the module we set out to nd. We break
up the proof in some intermediate lemmas.
Recall that in the universal enveloping algebra U(g) (or indeed in any algebra)
we can dene a bracket [ , ] by [X, Y ] = XY Y X. If X and Y happens to be
in g then the value of this bracket is the same as the Lie bracket of X and Y
(this is part of the denition of U(g)).
Lemma 5.17. In U(sl(2, C)) we have that [E, F
n
] = nF
n1
(H(n1)) where
E, F and H denote the canonical basis vectors for sl(2, C).
Proof. Let us by R
F
denote the map on U(g) which multiplies by F from the
right, and similarly by L
F
the map that multiplies with F from the left, and
dene ad(F)E := (L
F
R
F
)E = [F, E]. By the binomial formula applied to
(R
F
)
n
= (L
F
ad(F))
n
we get
(R
F
)
n
E =
n
j=0
_
n
j
_
(L
F
)
nj
(ad(F))
j
E.
This sum terminates after 3 terms since
ad(F)
3
E = [F, [F, [F, E]]] = [F, [F, H]] = [F, 2F] = 0.
Therefore we get
(R
F
)
n
E = (L
F
)
n
E n(LF)
n1
[F, E]
n(n 1)
2
(L
F
)
n2
[F, [F, E]]
= (L
F
)
n
E +nH(L
F
)
n1
n(n 1)(L
F
)
n2
F
= (L
F
)
n
E +nF
n1
(H (n 1)).
1
It can be shown that any irreducible representation of g is in fact equivalent to one of the
form L() but proving that is a little outside the scope of these notes.
Subtract (L
F
)
n
E and we have the result we want.
Lemma 5.18. Let g be a complex semisimple Lie algebra and h a Cartan subal-
gebra. Let h
and be chosen such that m := 2

,)
||
2
is a positive integer.
Let v
denote the canonical generator of the Verma module V () and M be the

U(g)-submodule of V () generated by Y
m
v
, where Y := Y
t
U(g).
Then M is isomorphic, as a U(g)-module, to V (s
).
Proof. As Y g
we have Y
m
U(b) and consequently v := Y
m
v
,= 0
according to Proposition 5.15 2). We have v
V ()
and hence that

v V ()
m
. But m = s
() and therefore v V ()
s()
.
We now show that Xv = 0 for X := X
t
where is a simple root, for then v

will be a highest weight vector with weight s
(). As M is the corresponding

highest weight module generated by v it follows from Proposition 5.15 3) that
M and V (s
()) are isomorphic.

If ,= then g
= 0 according to Lemma 4.23 and therefore [X, Y ] = 0

and thus we get
Xv = XY
m
v
= Y
m
Xv
= 0.
If = then we have the triple sl
= X, Y, H
t
and the preceding lemma

tells us that
Xv = XY
m
v
= [X, Y
m
]v
= mY
m1
(H
t
(m1))v
= m(( )(H
t
) (m1)) Y
m1
v
= m
_
2
,
||
2
(m1)
_
Y
m1
v
= 0
where the last equality is a consequence of Proposition 4.34.
Proposition 5.19. Let h
be a dominant integral element such that [

h0
is
real-valued. Then L( +) is a nite-dimensional irreducible U(g)-module with
highest weight .
Proof. From Proposition 5.16 we know already that L(+) is an irreducible
highest weight module with highest weight . Thus we can nd a non-zero
element [v
] L( + )
. Let be a simple root and put Y := Y

t
and
n := 2
+,)
||
2
. As is a dominant integral element and 2
,)
||
2
= 1 (Proposition
4.34) n is a positive integer. From the preceding lemma we have that Y
n
v

V (s
( +)) V ( +). But s
( +) = ( +) n and thus V (s
( +))
has highest weight n which is strictly lower than . Thus V (s
( + ))
cannot equal V ( + ) and therefore the former is a proper submodule of the
latter. Consequently Y
n
v
S (which was the sum of all proper submodules)

and therefore in the quotient, i.e. Y
n
[v
] = [Y
n
v
] = 0.
Next we prove that the set of weights for L(+) is invariant under the action
of the Weyl group. For now we just write v
instead of [v
]. Let be a simple root

and sl
= X, Y, H denote the corresponding sl(2, C)-triple. Put v

i
:= Y
i
v
.
As we have just seen, for i big enough, v
i
is zero. Therefore there is a maximal
n such that v
n
,= 0 and v
n+1
= 0. Consider the space W := Cv
0
+ + Cv
n
.
Obviously this space is Y - and H-invariant. Since
Xv
k
= XY
k
v
=
k
i=1
Y
i1
[X, Y ]Y
ki
+Y
k
Xv
.
The last term is 0, as v
is a highest weight, while the rst k terms give a

multiple of v
k1
(recall that [X, Y ] = H). Thus W is sl
-invariant, and thus it

is a nite-dimensional U(sl
)-submodule of L( + ). Consider the sum of all

nite-dimensional U(sl
)-submodules of L( + ). The space is g-invariant, for

if T is a nite-dimensional U(sl
)-module then gT is nite-dimensional (it has

dimension at most dimg dimT) and it is sl
-invariant for if t T, A sl
and
B g then
ABt = BAt + [A, B]t = Bt
t
+ [A, B]t gT,
i.e. gT is itself a nite-dimensional U(sl
)-submodule. Thus as L( +) is irre-

ducible this sum of nite-dimensional U(sl
)-modules (which is non-empty since

it contains v
for example) must equal L( + ). This implies that each vector

in L(+) has components in nitely many nite-dimensional U(sl
)-modules.
The sum of these modules is again a nite-dimensional U(sl
)-module, conclu-
sion: every vector in L( + ) is in a nite-dimensional sl
-invariant subspace.
Corollary 3.14 then gives a decomposition of L( + ) into nite-dimensional
irreducible U(sl
)-modules.
Let be an arbitrary weight for L( + ), let w L( + )
be a nonzero
weight vector and a xed simple root. Then w will have components in nitely
many U(sl
)-submodules and we decompose accordingly: w = w

1
+ + w
k
.
Then we have
k
i=1
Hw
i
= Hw = (H)w =
k
i=1
(H)w
i
i.e. Hw
i
= 2
,)
||
2
w
i
. Observe that m := 2
,)
||
2
is an integer since is, in
particular, an integral element. If , > 0 then w
i
is a weight vector of weight
m for the corresponding representation of sl
. The weights of this representation

is distributed symmetrically around 0 and acting on w
i
by Y
m
we push it
down to a nonzero weight vector of weight m. Thus Y
m
w
i
,= 0 and therefore
Y
m
w ,= 0. But as Y
m
w L( + )
m
we see that this space is nontrivial
i.e. m = s
() is a weight. If, on the other hand, , < 0 then likewise

X
m
w ,= 0 is an element of L( +)
s()
. If , = 0 then s
() = . In any
case we see that s
() is again a weight. As the Weyl group is generated by

root reections from simple roots (cf. Proposition 4.29) we see that the set of
weights is invariant under the action of the Weyl group.
Finally we show that the set of weights of L( + ) is nite. Since by 2) of
Lemma 5.13 the weight spaces are nite-dimensional the weight space decom-
position tells us that L( +) is nite-dimensional.
From Proposition 4.29 any functional in h
0
is of the form w() where w is
an element of the Weyl group and is dominant. In particular any weight can
be written as w() where is a dominant weight. Thus the number of weights
is at most the number of elements of W (which is nite) times the number of
dominant weights. Any weight is of the form
k
i=1
n
i
i
(Lemma 5.13) where
1
, . . . ,
k
is the set of simple roots. If, in addition,
k
i=1
n
i
i
is dominant
then it follows that

k
i=1
n
i
i
, 0
and hence
,
k
i=1
n
i
i
, =
k
i=1
n
i
i
, .
As
i
is simple, we get from Proposition 4.34 that
i
, =
1
2
|
i
|
2
> 0. There-
fore we must have that
k
i=1
n
i
is bounded by a constant which is independent
of the specic weight in question. Thus there can only be nitely many dominant
weights.
Wrapping up, the results of this chapter can be stated as
Theorem 5.20 (Highest Weight Theorem). Let g be a complex semisimple
Lie algebra and h a Cartan subalgebra. There is a 1-1 correspondence between
nite-dimensional complex irreducible representations of g and dominant inte-
gral elements in h
0
.
5.3 The Case sl(3, C)
In this last section of the chapter we apply the machinery developed thus far
in this chapter to the special case of the Lie algebra sl(3, C). This is a complex
semisimple Lie algebra as we have seen. It is the complexication of su(3), hence
by Proposition 7.22 the representation theory of these two Lie algebras is the
same. Furthermore SU(3) is simply connected (Theorem B.6) and thus there is
a 1-1 correspondence between nite-dimensional representations of SU(3) and
ditto representations of su(3). Combining these we see that a complete knowl-
edge of the representation theory of sl(3, C) yields complete knowledge of the
representation theory of SU(3). In particular, if we can determine the irreducible
representations of sl(3, C) we have determined the irreducible representations
of SU(3). The group SU(3) is important due to its relation to quantum eld
theory and particle physics.
We start by picking the basis
H
1
=
_
_
1 0 0
0 1 0
0 0 0
_
_
, H
2
=
_
_
0 0 0
0 1 0
0 0 1
_
_
,
X
1
=
_
_
0 1 0
0 0 0
0 0 0
_
_
, X
2
=
_
_
0 0 0
0 0 1
0 0 0
_
_
, X
3
=
_
_
0 0 1
0 0 0
0 0 0
_
_
,
Y
1
=
_
_
0 0 0
1 0 0
0 0 0
_
_
, Y
2
=
_
_
0 0 0
0 0 0
0 1 0
_
_
, Y
3
=
_
_
0 0 0
0 0 0
1 0 0
_
_
.
Note, that physicists would usually replace H
2
by
1
3
_
_
1 0 0
0 1 0
0 0 2
_
_
.
Below we state the commutation relations involving H
1
and/or H
2
[H
1
, H
2
] = 0
[H
1
, X
1
] = 2X
1
[H
1
, Y
1
] = 2Y
1
[H
2
, X
1
] = X
1
[H
2
, Y
1
] = Y
1
[H
1
, X
2
] = X
2
[H
1
, Y
2
] = Y
2
[H
2
, X
2
] = 2X
2
[H
2
, Y
2
] = 2Y
2
[H
1
, X
3
] = X
3
[H
1
, Y
3
] = Y
3
[H
2
, X
3
] = X
3
[H
2
, Y
3
] = Y
3
It is not hard to see that spanH
1
, X
1
, Y
1
and spanH
2
, X
2
, Y
2
are subalgebras
of sl(3, C) and that they are both isomorphic to sl(2, C).
Dene h := spanH
1
, H
2
. Since [H
1
, H
2
] = 0 this is an abelian subalgebra.
Assume that
X = a
1
X
1
+a
2
X
2
+a
3
X
3
+b
1
Y
1
+b
2
Y
2
+b
3
Y
3
5.3 The Case sl(3, C) 87
commutes with H
1
. Then
0 = [H
1
, X] = 2a
1
X
1
a
2
X
2
+a
3
X
3
2b
1
Y
1
+b
2
Y
2
b
3
Y
3
and since the X
i
s and Y
i
s are linear independent, we must have a
1
= a
2
=
a
3
= b
1
= b
2
= b
3
= 0. From this it follows that h is a maximal torus. From the
commutation relations above we see that ad(H
1
) and ad(H
2
) are diagonal in the
chosen basis and hence that h is a Cartan subalgebra of sl(3, C). We identify h
with C
2
by identifying h
with the pair ((H

1
), (H
2
)).
Lets determine the set of roots R(sl(3, C), h), i.e. pairs (a
1
, a
2
) for which there
exists an X ,= 0 such that
[H
1
, X] = a
1
X and [H
2
, X] = a
2
X.
From the commutation relations above we see that we, at least, have the follow-
ing roots
(2, 1), (1, 2), (1, 1), (2, 1), (1, 2), (1, 1)
with corresponding root vectors X
1
, X
2
, X
3
, Y
1
, Y
2
and Y
3
. By the root space
decomposition there can be no more roots than these.
As fundamental system we pick :=
1
,
2
where
1
:= (2, 1) and
2
:=
(1, 2). These are obviously linearly independent and as
(2, 1) =
1
(1, 1) =
1
2
(1, 1) =
1
+
2
(2, 1) =
1
(1, 2) =
2
(1, 2) =
2
.
Hence R N
0
(N
0
) and thus is a genuine fundamental system. The
positive roots are
1
,
2
and
1
+
2
with corresponding root vectors X
1
, X
2
, X
3
and the negative roots are
1
,
2
and
1

2
with corresponding root
vectors Y
1
, Y
2
and Y
3
. Therefore
sl(3, C)
+
= spanX
1
, X
2
, X
3
and sl(3, C)
= spanY
1
, Y
2
, Y
3
.
Now consider a nite-dimensional complex representation (, V ) of sl(3, C).
A weight for this is (under the identication above) a pair (
1
,
2
) C
2
for
which there exists 0 ,= v V with
(H
1
)v =
1
v and (H
2
)v =
2
v.
We can restrict to H
1
, X
1
, Y
1
thus obtaining a representation of sl(2, C).
From Corollary 3.13 we know that the eigenvalues of (H
1
) are integers. Simi-
larly, the eigenvalues of (H
2
) are integers. Thus a weight is necessarily a pair
of integers.
If we, moreover, require the functional (
1
,
2
) to be dominant, i.e. ,
1
+
2
0 for all positive roots , then we get that , =
1
+
2
0 and that
,
1
= 2
1
2
0 and ,
2
=
1
+ 2
2
0. Hence
1
=
1
3
((2
1
2
) + (
1
+
2
)) 0,
2
=
1
3
((
1
+ 2
2
) + (
1
+
2
)) 0.
In other words the dominant integral elements of h
are precisely the (function-

als represented by the) pairs (
1
,
2
) with
i
N
0
. Now the Highest Weight
Theorem tells us that there is a 1-1-correspondence between these pairs and
irreducible representations of sl(3, C). But how do we, given a dominant inte-
gral element, compute the corresponding irreducible representation? In principle
we could compute the Verma module and then take the quotient as described
in the previous section. However, this is too complicated, so here we outline a
somewhat simpler method which is applicable in this case.
The rst thing we do is to nd the so-called fundamental representations
of sl(3, C), namely the irreducible representations having (1, 0) and (0, 1) as
highest weights. The rst of these is simply the dening representation i.e. the
representation on C
3
given by
1
(X)v = Xv. This representation is irreducible:
assume that W is a non-zero invariant subspace and 0 ,= v = (v
1
, v
2
, v
3
) W.
We see that (X
1
)v = X
1
v = (v
2
, 0, 0) and X
3
v = (v
3
, 0, 0), so if either v
2
,= 0
or v
2
,= 0 then e
1
W. If v
2
= v
3
= 0 then v
1
,= 0 and
1
(H
1
)v = (v
1
, 0, 0).
Thus in any case e
1
W. Likewise one shows that e
2
and e
3
are in W as well
and thereby that
1
is irreducible.
We easily calculate that (1, 0), (1, 1) and (0, 1) are weights of
1
with
corresponding weight vectors e
1
, e
2
and e
3
(the standard basis vectors for C
3
).
By the weight space decomposition the number of weights cant exceed 3, so we
have found all the weights. Since (1, 1) = (1, 0)
1
and (0, 1) = (1, 0)
2
we see that (1, 0) is the highest weight. Thus
1
is the irreducible representation
having (1, 0) as highest weight.
Now consider the representation
2
on C
3
given by
2
(X)v = X
t
v.
Since
2
([X, Y ]) = [X, Y ]
t
= [X
t
, Y
t
] = [
2
(X),
2
(X)] this is really a
representation. By the same kind of reasoning as before, this representation is
irreducible and e
1
, e
2
and e
3
are weight vectors of
2
with corresponding weights
(1, 0), (1, 1) and (0, 1). This time (0, 1) is the highest weight since
(1, 0) = (0, 1)
1
2
and (1, 1) = (0, 1)
2
.
Thus
2
is the irreducible representation having (0, 1) as highest weight.
Now let a dominant integral element (m
1
, m
2
) be given. Form the tensor
product representation
m1m2
:=
m1
1

m2
2
(where by
m1
1
we mean the
m
1
-fold tensor product of
1
with itself). It is not hard to see that the vec-
tor v
m1m2
:= e
m1
1
e
m2
3
is a highest weight vector for
m1m2
with weight
(m
1
, m
2
). Let W denote the smallest invariant subspace of (C
3
)
(m1+m2)
con-
taining v
m1m2
, then v
m1m2
is a cyclic highest weight vector and
m1m2
[
W
is an
irreducible representation having (m
1
, m
2
) as highest weight.
Lets be even more concrete in the case of highest weight (1, 1). To make
things a little easier we dene a new basis by
f
1
:= e
3
, f
2
:= e
2
, f
3
:= e
1
for then
2
(Y
2
)f
1
= f
2
and
2
(Y
1
)f
2
= f
3
whereas all other possible actions of
the Y
i
s on the f
j
s are 0. The highest weight vector for
2
is f
1
.
We should form the tensor product
1
2
on C
3
C
3
and nd the smallest
invariant subspace containing e
1
f
1
. We nd this subspace simply by acting
on e
1
f
1
with (
1

2
)(Y
1
) and (
1

2
)(Y
2
) (acting by (
1

2
)(X
i
) just
gives 0 as e
1
and f
1
are highest weight vectors). The resulting vectors we act on
again by (
1
2
)(Y
1
) and (
1
2
)(Y
2
) and so forth until we hit 0. If ones does
so, one would nd that the smallest invariant subspace W is the span of the
following 8 elements (for the sake of brevity we have omitted the tensor product
symbol)
e
1
f
1
, e
2
f
1
, e
1
f
2
, e
3
f
1
+e
2
f
2
, e
2
f
2
+e
1
f
3
, e
2
f
3
, e
3
f
2
, e
3
f
3
.
Since the vectors e
i
f
j
constitute a basis for C
3
C
3
it shouldnt be too hard to
see that the vectors above are linearly independent.
1
2
restricted to this set
is thus the irreducible representation corresponding to the highest weight (1, 1).
5.3 The Case sl(3, C) 89
But this representation is nothing but the adjoint representation of sl(3, C).
Indeed consider the isomorphism : W sl(3, C) which maps
e
1
f
1
X
3
, e
1
f
2
X
1
, e
2
f
1
X
2
e
3
f
1
+e
2
f
2
H
2
,
e
2
f
2
+e
1
f
3
H
1
, e
2
f
3
Y
1
, e
3
f
2
Y
2
, e
3
f
3
Y
3
.
This is an intertwiner of (
1
2
)[
W
and ad. For instance we see that
(
1
2
)(Y
1
)e
1
f
1
=
1
(Y
1
)e
1
f
1
+e
1
2
(Y
1
)f
1
= e
2
f
1
+e
1
0 = e
2
f
1
whereas
ad(Y
1
)((e
1
f
1
)) = ad(Y
1
)(X
3
) = [Y
1
, X
3
] = X
2
= (e
2
f
1
).
Thus the adjoint representation of sl(3, C) is (equivalent to) the irreducible
representation with highest weight (1, 1).
Chapter 6
Innite-dimensional
Representations
6.1 Grding Subspace
In Chapter 3 we discussed nite-dimensional representations of Lie groups and
their induced Lie algebra representations. All that was based on the fact that
Aut(V ) is a Lie group when V is a nite-dimensional vector space. Its a natural
question to ask what will happen if we consider representations on innite-
dimensional Hilbert spaces. Do they descend to representations of the associated
Lie algebra? Under certain modications the answer is yes, but it is clear that
the situation is not as simple as in the previous chapter.
Recall the situation: Given a nite-dimensional representation (, V ) of an
arbitrary Lie group G, it induces a representation
of g on V given explicitly
by
(X)v =
d
dt
t=0
(exp(tX))v. (6.1)
This is well-dened, since was automatically smooth (Lemma 3.1). In the case
of a representation on an innite-dimensional Hilbert space H we would like to
dene a g-representation by (6.1) but a priori this expression does not make
sense for all vectors in H. Actually, we have to restrict the representation of g
to a subspace of H, namely to the space of so-called C
-vectors. To dene these

we need to discuss dierentiability of H-valued maps.
Denition 6.1. Let U 1
n
be an open set, and f : U H be a map into a
Banach space. f is said to be dierentiable at x
0
U if the limit
lim
x0
f(x
0
+x) f(x
0
)
|x|
exists. In the armative case we dene f
t
(x
0
) to be that limit. The function f
is said to be dierentiable if it is dierentiable at all the points of U. We say
that f is C
k
if f is k times dierentiable and f
(k)
is continuous. We say that f
is C
if it is C
k
for all k.
We dene the directional derivative of f at x
0
in the direction of v 1
n
to
be
lim
t0
f(x
0
+tv) f(x
0
)
t
.
Like in ordinary calculus one can show that f is C
k
if and only if all the
directional derivatives are C
k1
.
91
92 Chapter 6 Innite-dimensional Representations
Now we consider a map f : G H from a Lie group G (or, for that
matter, any smooth manifold) to H. We can dene dierentiability of this by
demanding that its composition with any inverse coordinate map is dierentiable
in the sense of Denition 6.1, i.e. if x
0
G and (U, ) is an arbitrary coordinate
map for G around x
0
then we demand the map f
1
: (U) H to be
dierentiable.
If X g is a left-invariant vector eld on G we can interpret (Xf)(g) as the
directional derivative of f in g in the direction of X
g
because we have
(Xf)(g) = lim
t0
f(g exp tX) f(g)
t
.
Why is this true? Let be the ow of X, then
1
we have (t, g) = g exp tX.
Hence
X
g
=

t
t=0
(t, g) =
d
dt
t=0
g exp tX,
and therefore
(Xf)(g) = X
g
f =
_
d
dt
t=0
g exp tX
_
f =
d
dt
t=0
f(g exp tX)
= lim
t0
f(g exp tX) f(g)
t
which was to be proved.
To any left-invariant vector eld X there is an associated right-invariant vector
eld Y . Left invariance of X simply means that X
g
= (L
g
)
X
e
, (where L
g
is
the Lie group automorphism h g
1
h) i.e. X is determined from its value at
e. Similarly any right-invariant vector eld Y satises Y
g
= (R
g
)
R
e
where R
g
is the automorphism h hg. Thus, from X we get a right-invariant vector
eld Y by Y
g
= (R
g
)
X
e
. We see that Y can also be regarded as a directional
derivative:
(Y f)(g) = Y
g
f = (R
g
)
X
e
f = X
e
(f R
g
)
=
d
dt
t=0
(f R
g
)(exp tX) =
d
dt
t=0
f(exp(tX)g)
= lim
t0
f(exp(tX)g) f(g)
t
.
We can now dene the subspace of C
-vectors.
Denition 6.2. Let (, H) be a representation of G on a Hilbert space H.
A vector v H is called a C
-vector for if the map G H given by

g (g)v is C
. The space of C
-vectors for is denoted H
.
The goal of the rest of this section is to prove that H
is dense in H. The
strategy of the proof is to introduce a subspace of H
, the so-called Grding

subspace, which is more manageable than H
itself, and show that this is dense.

A key ingredient in the proof will be the following lemma:
Lemma 6.3. Let (, H) be a representation of G and let S be a subspace of H
such that for all v S the limit
(X)v := lim
t0
(exp(tX))v v
t
(6.2)
exists and is in S for all X g. Then S H
.
1
See [9] Proposition 20.8 g).
6.1 Grding Subspace 93
Proof. Let v S. We should verify that v H
i.e. that the function f

v
:
G H given by f
v
(g) := (g)v is C
.
We will use induction. First, lets show that it is C
1
. The requirement that
the limit (6.2) exists for all X g, says that f
v
is dierentiable in e G. We
need to show that it is dierentiable in any arbitrary point g G. But as (g)
is a continuous map
lim
t0
f
v
(g exp(tX)) f
v
(g)
t
= lim
t0
(g exp(tX))v (g)v
t
= (g) lim
t0
(exp(tX))v v
t
= (g)(X)v.
Thus, all the directional derivatives exist and are continuous (by Proposition
1.2). Hence f
v
is C
1
.
Now for the induction step: Assuming that f
v
is C
k
, we will show that it
is C
k+1
. Since v is in S, also (X)v is in S, hence, by assumption f
(X)v
is
C
k
. But f
(X)v
(g) = (g)(X)v is the directional derivative of f
v
. Since all the
directional derivatives of f
v
are C
k
, f
v
itself must be C
k+1
.
Before introducing the Grding subspace we need to be able to talk about
integration in innite-dimensional Hilbert spaces. So let (X, ) be a measure
space, H a Hilbert space, and assume that f : X H is a compactly supported
continuous map, then there exists a unique vector v H satisfying
2
v, w =
_
X
f(x), wd(x).
This unique vector v is what we shall understand by the integral of the function
f, and thus we write
_
X
f(x)d(x) := v.
This is integration in the weak sense, and sometimes also called the weak integral
For this kind of integration we have the usual inequality
3
_
_
_
_
X
f(x)d(x)
_
_
_
_
X
|f(x)|d(x). (6.3)
Furthermore, if T : H H is a continuous linear map, then
4
T
_
X
f(x)d(x) =
_
X
Tf(x)d(x). (6.4)
We also have dominated convergence:
Theorem 6.4 (Dominated Convergence). Let f
n
: X H be a sequence
of integrable functions converging pointwise to f, and suppose g : X 1
+
is an integrable function such that |f
n
(x)| g(x) for all x X, then f is
integrable and
lim
n
_
X
f
n
d =
_
X
fd.
Proof. For any x X we have |f
n
(x)| g(x) for all n and thus we must also
have |f(x)| g(x), i.e. f is integrable. We see that
_
_
_
_
X
f
n
d
_
X
fd
_
_
_ =
_
_
_
_
X
(f
n
f)d
_
_
_
_
X
|f
n
(x) f(x)|d(x)
2
For this implication, see [Rudin] Denition 3.26 and Theorem 3.27.
3
See [Rudin] Theorem 3.29.
4
See [Rudin] Exercise 3.24.
By the usual dominated convergence we know that
lim
n
_
X
|f
n
(x) f(x)|d(x) =
_
X
lim
n
|f
n
(x) f(x)|d(x)
and this is 0, since the function x lim
n
|f
n
(x) f(x)| is identically
0.
With this in mind we can, given a unitary representation on H and a
f C
c
(G), dene a linear map (f) : H H by
(f)v :=
_
G
f(g)(g)vd(g)
where is the left Haar measure on G. This map is continuous for
| (f)v| =
_
_
_
_
G
f(g)(g)v dg
_
_
_
_
G
[f(g)[|v| dg = |f|
1
|v|.
I.e. | (f)| |f|
1
< since f is smooth and compactly supported.
Denition 6.5 (Grding Subspace). Let (, H) be a unitary representation
of G on the Hilbert space H. The Grding subspace H
G
of is the subspace
of H spanned by all elements of the form (f)v for f C
c
(G) and v H.
Elements of this space are called Grding vectors.
Lemma 6.6. The Grding subspace is a subset of H
.
Proof. We will show that H
G
satises the requirements of Lemma 6.3, i.e.

that for all v H
G
the limit
(X)v = lim
t0
(exp(tX))v v
t
exists and is of the form (f)v
t
. We will show that we actually have the explicit
expression
(X) (f)v = (Y f)v (6.5)
where Y is the right-invariant vector eld corresponding to X and Y f is the
C
-function on G obtained by letting Y act on f.

Now x a t ,= 0, then
(exp(tX)) id
H
t
(f)v =
1
t
(exp(tX))
_
G
f(g)(g)v dg
1
t
_
G
f(g)(g)v dg
=
1
t
_
G
f(g)(exp(tX)g)v dg
1
t
_
G
f(g)(g)v dg.
In the last equality we used (6.4) to put (exp(tX)) inside the integral.
The Haar measure on G is left-invariant, and hence the rst integral will be
unaltered if we change the variable from g to exp(tX)g. Then the expression
above becomes
_
G
f(exp(tX)g) f(g)
t
(g)v dg. (6.6)
We want to investigate the limit as t 0. The fraction inside the integral
sign looks familiar, it converges to (Y f)(g) when t 0, so if we can exchange
limit and integration we are done. Because of this convergence we can, given an
> 0, nd a such that for [t[ < we have
f(exp(tX)g) f(g)
t
[(Y f)(g)[
f(exp(tX)g) f(g)
t
(Y f)(g)
<
6.2 Induced Lie Algebra Representations 95
which implies that
f(exp(tX)g) f(g)
t
[(Y f)(g)[ +.
Y f is a smooth compactly supported function (just like f), hence the right hand
side is a t-independent integrable majorant for the expression (6.6). Thus we
have dominated convergence and for t 0 (6.6) tends to
_
G
(Y f)(g)(g)v dg = (Y f)v
which proves the statement.
Theorem 6.7 (Grding). The Grding subspace is dense in H. In particular
H
is dense in H.
Proof. Let v H be arbitrary. For any given > 0 we should nd a Grding
vector which is -close to v.
By Proposition 1.2 the map G g (g)v v H is continuous for all
v H. Hence the set
U
:= g G[ |(g)v v| < H
is open. It is then possible to nd a compactly supported smooth positive func-
tion f in G which is supported in U and satises
_
G
f(g)dg = 1
5
. We claim
that the Grding vector (f)v is -close to v:
| (f)v v| =
_
_
_
_
G
f(g)(g)vdg v
_
_
_ =
_
_
_
_
G
f(g)((g)v v)dg
_
_
_
_
G
f(g)|(g)v v|dg <
_
G
f(g)dg = .
In the last inequality we used that f was supported in U where, by denition,
|(g)v v| < .
6.2 Induced Lie Algebra Representations
For the rest of this chapter denotes a unitary representation of G on a Hilbert
space H.
We have for each X g a map (X) dened by (6.2) on a dense subset of H,
namely H
. In this section we will see that in fact, is a representation of g

on H
. We will develop this result in a few propositions. The rst thing to be

veried is that (X) is linear:
Proposition 6.8. (X) is a linear map and im(X) H
, i.e. (X)
End(H
).
Proof. To show linearity we just we simply calculate
(X)(av +bw) = lim
t0
(exp(tX))(av +bw) (av +bw)
t
= a lim
t0
(exp(tX))v v
t
+b lim
t0
(exp(tX))w w
t
= a(X)v +b(X)w.
5
Since G is locally compact, we can nd an open set V U with V compact and contained
in U. Then by standard manifold theory, we can nd a smooth positive function

f on G which
is supported in V ([9] Proposition 2.26). Since

f is supported in V and V was compact,

f is
compactly supported, hence integrable. By scaling

f with a proper normalizing constant we
obtain a smooth positive compactly supported function f with

G
fdg = 1.
For the second claim dene the map f : G H by f(g) = (g)v for
v H
. Thus constructed, f is a C
-map. That (X)v is a C
-vector means
by denition that g (g)(X)v is C
. To see that this is true, we calculate:

(g)(X)v = (g) lim
t0
(exp(tX))v v
t
= lim
t0
(g exp(tX))v (g)v
t
= (Xf)(g).
But Xf is C
, hence (X)v is a C
-vector.
In short, (X) is an endomorphism of H
, or a densely dened operator on

H.
Proposition 6.9. The map : g End(H
) mapping X (X) is a Lie

algebra representation of g.
Proof. There are two things to prove: showing that is linear, and showing
that it respects Lie brackets.
Verifying linearity of is straightforward:
(X +Y )v = lim
t0
(exp(tX +tY ))v v
t
= lim
t0
(exp(tX) exp(tY ))v (exp(tY ))v
t
+ lim
t0
(exp(tY ))v v
t
= lim
t0
(exp(tX)) id
H
t
lim
t0
(exp(tY ))v +(Y )v
= lim
t0
(exp(tX)) id
H
t
v +(Y )v = ((X) +(Y ))v,
and
(cX)v = lim
t0
(exp t(cX))v v
t
= c lim
t0
(exp(ctX))v v
ct
= c(X)v.
Hence is linear.
Now we should prove that ([X, Y ]) = (X)(Y ) (Y )(X). Note that
since (X) and (Y ) map H
into H
it makes sense to compose them. To

nd out how the compositions (X)(Y ) and (Y )(X) act, we see that
(X)(Y )v, w = lim
s0
lim
t0
_
(exp sX) id
H
s
(exp tY )v v
t
, w
_
(6.7)
(6.8)
In ordinary calculus one shows that if a function has continuous partial deriva-
tives, the function is dierentiable. By a similar argument one can show that
the equation above implies that
(X)(Y )v, w = lim
(s,t)(0,0)
_
(exp sX) id
H
s
(exp tY )v v
t
, w
_
In fact we can put s = t, and hence we get
(X)(Y )v, w = lim
t0
(exp tX)(exp tY )v (exp tX)v (exp tY )v +v, w
t
2
.
This is valid for all w H and so we conclude that
(X)(Y )v = lim
t0
1
t
2
_
(exp tX)(exp tY )v (exp tX)v (exp tY )v +v
_
.
6.3 Self-Adjointness 97
A similar calculation and argument render
(Y )(X)v = lim
t0
1
t
2
_
(exp tY )(exp tX)v (exp tX)v (exp tY )v +v
_
.
Therefore by subtracting we get
[(X),(Y )]v = lim
t0
((exp tX exp tY ))v (exp tY exp tX)v
t
2
= lim
t0
(exp tX exp tY )
(exp(tX) exp(tY ) exp tX exp tY )v v
t
2
= lim
t0
(exp tX exp tY )
(exp(t
2
[X, Y ] +O(t
3
)))v v
t
2
= lim
t0
(exp(t[X, Y ] +O(t
3/2
)))v v
t
= ([X, Y ])v
in that O(t
n
) denotes terms which are bounded by a positive constant times t
n
.
Why does the last equality hold? If is an O(t
3/2
)-function then it is easy to
see that
t
(0) = 0. Thus if F is a real function then
d
dt
t=0
F(t +(t)) = F
t
(0 +(0))
d
dt
t=0
(t +(t)) = F
t
(0).
Thus an O(t
3/2
)-term does not inuence the derivative at 0. An elaboration
of this argument gives the same result for derivations of functions in a Hilbert
space.
Thus (X) is a densely dened operator on H.
Proposition 6.10. The operator (X) is anti-symmetric, i.e. (X) (X)
.
Proof. It makes sense to talk about the adjoint of (X) as it is densely de-
ned. For w in the domain of (X)
, (X)
w is the unique vector satisfying

(X)v, w = v, (X)
w for all v H
. In the following computation we

exploit unitarity of (g). Let t ,= 0 and v, w H
1
t
((exp tX)v v), w = (exp tX)v,
1
t
w v,
1
t
w
= v,
1
t
(exp(tX))w v,
1
t
w
= v,
1
t
_
(exp(tX))w w
_
,
and so in the limit t 0 we get
(X)v, w = v, (X)w.
This precisely states that (X) is anti-symmetric.
6.3 Self-Adjointness
In this section we proceed to investigate the operator (X). Our goal is to apply
it to the quantum mechanical theory of momentum and angular momentum,
and for that we would like it to be self-adjoint. But that cannot be, thats what
Proposition 6.10 says. What we will show at the end, however, is that (X) is
essentially skew-adjoint i.e. that the closure is skew-adjoint: (X)
= (X).
This implies that the operator i(X) will be self-adjoint.
Proving this, however, is rather complicated and we have to go through a
series of technical lemmas before we can prove the result. The rst result revolves
around how (f) acts on (X)(H
G
). Recall from Chapter 1 that we have dened

the modular function : G 1
+
by
_
G
f(gh
1
)dg = (h)
_
G
f(g)dg
where dg is the left Haar measure on G. is a Lie group homomorphism, and
therefore it makes sense to dene a map : g 1
+
by
(X) =
d
dt
t=0
(exp(tX)).
Lemma 6.11. Let y = (k)v be a Grding vector and letf C
c
(G) be arbi-
trary. Then we have
(f)(X)y = (Xf)y +(X) (f)y. (6.9)
Proof. Recall (6.5) by which we have
(f)(X) (k)v = (f)( (Y k)v) =
__
G
f(g)(Y k)(h)(g)(h)v dhdg
=
__
G
f(g) lim
t0
1
t
_
k(exp(tX)h) k(h)
_
(g)(h)v dhdg.
As in the proof of Lemma 6.6 we have dominated convergence so we can inter-
change limit and integration, and hence get
lim
t0
1
t
__
G
f(g)
_
k(exp(tX)h) k(h)
_
(g)(h)v dhdg.
Now, x a nonzero t and split the integral in two:
1
t
__
G
f(g)k(exp(tX)h)(g)(h)v dhdg and
1
t
__
G
f(g)k(h)(g)(h)v dhdg.
In the rst we replace h by exp(tX)h (which does not alter the integral,
since the Haar measure is left-invariant), and in the second we replace g by
g exp(tX) (which we can do if we compensate by introducing (exp(tX)) in
the expression). Thus we get
1
t
__
G
f(g)k(h)(g)(exp(tX)h)(h)v dhdg
1
t
__
G
(exp(tX))f(g exp(tX))k(h)(g)(exp(tX))(h)v dhdg
=
__
G
1
t
_
f(g) f(g exp(tX))
_
k(h)(g)(exp(tX))(h)v dhdg
(exp(tX)) 1
t
__
G
f(g exp(tX))k(h)(g)(exp(tX))(h)v dhdg.
Inside the rst integral we have
lim
t0
f(g) f(g exp(tX))
t
= (Xf)(g),
and in the second we have
lim
t0
(exp(tX)) 1
t
= (X).
Thus, by using dominated convergence once again to bring the limit inside the
integrals, we get in the rst case
__
G
(Xf)(g)k(h)(g)(h)v dhdg = (Xf) (k)v = (Xf)y,
while the second gives
(X)
__
G
f(g)k(h)(g)(h)v dhdg = (X) (f) (k)v = (X) (f)y.
This completes the proof.
To nd the adjoint of the continuous linear map (f) put f
(g) := f(g
1
)(g):
(f)v, w =
_
_
G
f(g)(g)v dg, w
_
=
_
G
f(g)(g)v, wdg
=
_
G
v, f(g)(g
1
)wdg =
_
G
v, f(g
1
)(g)(g)wdg
=
_
G
v, f
(g)(g)wdg =
_
G
f
(g)(g)w, vdg
=
_
_
G
f
(g)(g)wdg, v
_
= v, (f
)w.
Hence the adjoint is (f
). In the next lemma we calculate the adjoint in the

case of a function of the form Y f for a right-invariant vector eld Y .
Lemma 6.12. Let f C
c
(G) and X g and let Y be the right-invariant
vector eld corresponding to X. Then the adjoint of (Y f) is given by
(Y f)
= (Xf
) (X) (f
). (6.10)
Proof. First of all f
is smooth, as is a Lie group homomorphism, and like

f it is compactly supported.
Now we want to show that
(Y f)v, w = v, (Xf
)w (X)v, (f
)w
for all v, w H, but since (Y f), (Xf
) and (f
) are all continuous linear

maps, it is sucient to verify the statement for Grding vectors v and w. For
v H
G
we have by (6.5) that (Y f)v = (X) (f)v, and so

(Y f)v, w = (X) (f)v, w = (f)v, (X)w = v, (f
)(X)w
= v, (Xf
)w +(X) (f
)w.
In the last equation we invoked Lemma 6.11.
Lemma 6.13. There exists a sequence of real, non-negative functions f
n

C
c
(G) and a corresponding descending sequence (U
n
) of neighborhoods around
e, such that |f
n
|
1
= 1, supp f
n
U
n
and

n
U
n
= e, and such that the
sequence |Y f
n
Xf
n
|
1
is a bounded sequence for each X g with corresponding
right-invariant vector eld Y .
We will skip the proof of this lemma. It is long, technical and not very il-
luminating. For a proof consult [3] Lemma 7.26. However we see that the re-
quirements on f
n
implies that f
n
converges to the delta function
e
in the sense
of distributions: namely let C
c
(G) be arbitrary. It is continuous in e and
since the supports of f
n
shrink to e we can, given an > 0 nd an n large
enough such that [(g) (e)[ < for g U
n
and therefore
_
G
f
n
(g)(g) dg (e)
_
Un
f
n
(g)((g) (e))dg
_
Un
f
n
(g)[(g) (e)[dg
<
_
Un
f
n
(g)dg = .
Lemma 6.14. Let (f
n
) be a sequence of compactly supported smooth functions
satisfying the demands of Lemma 6.13. Then the sequence of bounded operators
( (Y f
n
) (Xf
n
) (X) (f
n
)) converges strongly to the zero-operator.
Proof. Strong convergence is the same as pointwise convergence so we need to
show that
( (Y f
n
) (Xf
n
) (X) (f
n
))v =
_
G
(Y f
n
Xf
n
(X)f
n
)(g)(g)v dg
converges to 0 for all v H. Putting (g) := (g)v then : G H is a
continuous function and (e) = v. We split the above expression in two
_
G
(Y f
n
Xf
n
(X)f
n
)(g)((g)(e))dg+(e)
_
G
(Y f
n
Xf
n
(X)f
n
)(g)dg
To see that the rst term tends to 0 we remark that |f
n
| and |Y f
n
Xf
n
|
1
are bounded sequences by Lemma 6.13 and hence by the triangle inequality
|Y f
n
Xf
n
(X)f
n
|
1
is bounded by some constant M. Therefore
_
_
_
_
G
(Y f
n
Xf
n
(X)f
n
)(g)((g) (e))dg
_
_
_
_
G
[(Y f
n
Xf
n
(X)f
n
)(g)[|((g) (e))|dg
sup
gUn
|(g) (e)|
_
G
[(Y f
n
Xf
n
(X)f
n
)(g)[dg
M sup
gUn
|(g) (e)|.
We only need to take the sup over U
n
since this is where f
n
is supported.
As is continuous we can to a given > 0 nd a neighborhood U around e
so that |(g) (e)| < /M for g U. As U
n
is a descending sequence of sets
we will eventually have U
n
U, and thus sup
gUn
|(g) (e)| < /M, i.e.
_
_
_
_
G
(Y f
n
Xf
n
(X)f
n
)(g)((g) (e))dg
_
_
_ < .
Therefore the rst term tends to zero.
Now we need to show that
_
G
(Y f
n
Xf
n
(X)f
n
)(g)dg tends to zero. The
rst term converges to zero:
_
G
(Y f
n
)(g)dg =
_
G
lim
t0
1
t
(f
n
(exp(tX)g) f
n
(g))dg
= lim
t0
1
t
_
_
G
f
n
(exp(tX)g)dg
_
G
f
n
(g)dg
_
= 0,
where the second equality follows from dominated convergence and the last one
follows from left invariance of the Haar measure.
The two last terms happen to cancel each other:
_
G
(Xf
n
)(g)dg =
_
G
lim
t0
1
t
(f
n
(g exp(tX)g) f
n
(g))dg
= lim
t0
1
t
_
(exp(tX)) 1
_
_
G
f
n
(g)dg = (X)
_
G
f
n
(g)dg,
proving the lemma.
Now for the nal lemma, before the main results of this section.
Lemma 6.15. For all X g we have (X)
(X)
.
Proof. We will show that (X)
x, y+x, (X)
y = 0 for all x, y D((X)
).
The strategy is to write this as a sum of terms which are all zero (the function
f C
c
(G) is arbitrary):
(X)
x, y +x, (X)
y =(X)
x, y (X)
x, (f
)y (6.11)
+(X)
x, (f
)y x, (X) (f
)y (6.12)
+x, (X) (f
)y x, (Y f
)y (6.13)
+x, (Y f
)y x, (Xf
+(X)f
)y (6.14)
+x, (Xf
+(X)f
)y + (Y f)x, y (6.15)
(Y f)x, y +(X) (f)x, y (6.16)
(X) (f)x, y + (f)x, (X)
y (6.17)
(f)x, (X)
y +x, (X)
y. (6.18)
The terms (8.2) and (6.17) are 0 simply by denition of the adjoint. The terms
(6.13) and (6.16) are 0 by Eq. (6.5). The term (6.15) is 0 by Lemma 6.12. In
(6.14) we put f = f
n
and let n and the two terms converge to zero by
Lemma 6.14. For (8.1) we do something similar: we put f = f
n
and show that
(f
n
)y y for n . Namely for any v H we have
lim
n
(f)y, v = lim
n
(f
n
)y, v = lim
n
_
G
f
n
(g)(g)y dg, v
_
= lim
n
_
G
f
n
(g)(g)y dg, vdg
= lim
n
_
G
f
n
(g)(g)y, vdg
=
e
((g)y, v) = y, v.
Thus for n the two terms in (8.1) go to zero. An identical argument shows
that the two terms in (6.18) cancel as well.
Finally, we have collected enough results to prove the following theorem
Theorem 6.16. (X) is essentially skew-adjoint for all X g.
Proof. We will see that (X)
= (X). Since (X) is densely dened, (X)
is closed, and as (X) (X)
by Proposition 6.10 also we have

(X) (X)
(6.19)
because (X) is the smallest closed extension of (X). Also (X)
is densely de-
ned, so (X)
is closed and thus equal to (X). (6.19) now implies (X)

(X)
i.e. (X) (X)
.
Conversely, Lemma 6.15 yields (X) = (X)
(X)
which immediately
renders (X)
(X)
= (X). Thus, (X) is skew-adjoint.

Thus if we dene
(X) = (X) we get a map
from g to the set of skew-

adjoint operators on H. This latter space we denote O(H).
Theorem 6.17. For any X g we have exp(
(X)) = (exp(X)), i.e. the

following diagram is commutative (compare with (3.1))
g

exp
O(H)
exp
Aut(H)
Proof. Consider the following two families of unitary operators V
t
:= (exp tX)
and W
t
:= exp(t
(X)). By Proposition 1.2 V

t
is a strong continuous 1-parameter
group of unitaries, thus by Stones Theorem there exists a self-adjoint operator
T on H such that V
t
= exp(itT) and it can be constructed in the following way:
Let D(A) := v H [ lim
t0
1
t
(V
t
v v) exists and put for v D(A):
Av = lim
t0
1
t
(V
t
v v).
Then T := iA will do the job. Since we have for v H
:
lim
t0
1
t
(V
t
v v) = lim
t0
(exp tX)v v
t
=
(X)v
we have that H
D(A) and that Av =
(X)v. This implies
(X) iT.
Since both sides are skew-adjoint we get
(X) =
(X)
(iT)
= iT
and thus iT
(X), i.e. T = i
(X). Thus V
t
= W
t
. In particular this
holds for t = 1 and this is the desired formula.
6.4 Applications to Quantum Mechanics
In this nal chapter we bring the Peter-Weyl Theorem, the Highest Weight
Theorem and the theory from the previous section into play in a brief description
of the quantum mechanical theory of momentum and angular momentum
6
.
First lets discuss quantum mechanical momentum. Quantum mechanics is
modeled on operators on Hilbert spaces: to each quantum mechanical system
there is associated a Hilbert space (usually an L
2
-space) and to any observable
for that system such as momentum, position or energy, there is a unique self-
adjoint operator on the associated Hilbert space. The only possible values of a
given observable one can measure lie in the spectrum of the given operator, and
since the operator is self-adjoint only real values occur in the spectrum. Using the
theory from the previous section we will motivate why the momentum operators
look as they do and nd a domain of essentially self-adjointness.
Lets consider a free particle in space. Then the proper Hilbert space is L
2
(1
3
).
On this we consider the operator T (x
0
)f(x) = f(x x
0
) which translates the
coordinate system along x
0
. This corresponds to an actual translation of the
system along x
0
. Its called the translation operator. It is not hard to see (using
Proposition 1.2) that T is a continuous representation of the additive group
1 and translation invariance of the Lebesgue measure on 1
3
yields unitarity of
6
For a more physical but highly non-rigorous discussion of this, the reader is referred to
Chapter 3 of [12].
6.4 Applications to Quantum Mechanics 103
this representation
7
. Hence it induces a Lie algebra representation T
of 1
3
on
L
2
(1
3
) (or rather on the space of C
-vectors L
2
(1
3
)
T
), where the Lie algebra
of 1
3
is just 1
3
with the zero-bracket and the exponential map is the identity
1
3
1
3
. Choosing the standard basis e
1
, e
2
, e
3
for 1
3
we can calculate
T
(e
i
):
T (e
1
)f(x) =
d
dt
t=0
T (exp te
1
)f(x) =
d
dt
t=0
f(x +te
1
)
=
d
dt
t=0
f(x
1
+t, x
2
, x
3
) =
f
x
1
that is T
(e
1
) =

x1
and likewise T
(e
i
) =

xi
.
Well, what are the C
-vectors anyway? That is not an easy question and we

wont answer it completely. First of all a C
-vector for T is an L
2
-function sat-
isfying that the map x (T (x)f) is C
from 1
3
to L
2
(1
3
). More concretely,
this means, in particular, that the expression
T (te
i
)f f
t
converges, in L
2
, to some L
2
-function h when x tends to 0. Lets show that this
implies that h is the weak derivative of f, i.e. satises f, = h, for all
test functions . Obviously, we have (for a xed t) that
_
T (te
i
)f f
t
,
_
h, .
On the other hand we see that
_
T (te
i
)f f
t
,
_
=
1
t
_
T (te
i
)f, f,
_
=
1
t
_
f, T (te
i
) f,
_
=
_
f,
T (te
i
)f
t
_
(the rst equality being a consequence of translation invariance of the Lebesgue
measure) and this converges to f,
i
by dominated convergence, since the
fraction inside the inner product converges to
i
. Thus h is the weak derivative
of f and this is in L
2
. This we can repeat as often as we like, i.e. f has weak
derivatives in L
2
of any order. By the Sobolev Embedding Theorem this implies
that f is a C
-function. In other words, if f is a C
-vector for T , it is a
C
-function. Next we will show that the C
-functions with compact support

are actually Grding vectors and thus in particular C
-vectors. First let f, h

C
c
(1
3
). We want to show that

T (f)h = h f (i.e. the convolution of h with
f). By denition
T (f)h =
_
R
3
f(x)T (x)hdx
is the unique L
2
-function satisfying
T (f)h, =
_
R
3
f(x)T (x)f, dx
for all L
2
. We see (by a proper use of Fubinis Theorem) that
_
R
3
f(x)T (x)f, =
_
R
3
f(x)h(y x)(y) dydx = h f,
7
Unitarity is a requirement on symmetry operators in quantum mechanics since they pre-
serve probability.
and hence that

T (f)h = hf. Thus, the Grding vectors are convolutions. Now
a quite deep theorem due to Dixmier and Malliavin states that any function in
C
c
(1
3
) can be written as
n
i=1
u
i
v
i
where u
i
, v
i
C
c
(1
3
). By the remarks above this means that C
-functions
with compact support are Grding vectors and thus our operators T
(e
i
) are at
least dened on this space.
But why are these operators interesting at all? For the motivation we give
the following slightly non-rigorous argument based on Theorem 6.17: If is
a small number we can consider what physicists call an innitesimal transla-
tion e
i
. Then by Theorem 6.17 T (e
i
) = exp(T
(e
i
)). By series expansion
of exp and neglecting terms with to a power higher than 1 we get approxi-
mately T (e
i
) id +T
(e
i
). Thus T
(e
i
) is a measure of how much the system
changes when translated a little bit along the x
i
-axis. In the Hamiltonian formu-
lation of classical mechanics, this precisely has the interpretation of momentum
of the system. In generalizing to quantum mechanics, we therefore dene mo-
mentum in this way.
However the operator T
(e
i
) is only anti-symmetric. To make it self-adjoint
and to make it have the right units we dene p
j
:= i/T
(e
j
)
8
. These are the
momentum operators in quantum mechanics.
Now following the same pattern, we analyze rotation and how it inuences a
quantum mechanical system. We consider the rotation group SO(3) which is a
compact Lie group. Its Lie algebra is so(3) i.e. the 3 3 skew-symmetric real
matrices. An obvious basis for so(3) is the following
A
1
=
_
_
0 0 0
0 0 1
0 1 0
_
_
, A
2
=
_
_
0 0 1
0 0 0
1 0 0
_
_
, A
3
=
_
_
0 1 0
1 0 0
0 0 0
_
_
,
and it is easily checked that they possess the following commutation relations
[A
1
, A
2
] = A
3
, [A
2
, A
3
] = A
1
, [A
3
, A
1
] = A
2
.
The exponential map is no longer as trivial as in the past example. To calculate
exp A for a matrix A so(3) we have to calculate the Taylor series
n=0
1
n!
A
n
.
If we apply this formula for exp to the matrices tA
i
for some real number t, we
get
exp(tA
1
) =
_
_
1 0 0
0 cos t sin t
0 sin t cos t
_
_
, exp(tA
2
) =
_
_
cos t 0 sin t
0 1 0
sin t 0 cos t
_
_
(6.20a)
exp(tA
3
) =
_
_
cos t sin t 0
sin t cos t 0
0 0 1
_
_
. (6.20b)
That is, the element exp(tA
i
) SO(3) is the rotation of angle t around the
x
i
-axis.
Consider a particle (without spin!) which can move inside some massive sphere
1
3
, or for that matter inside = 1
3
. The proper Hilbert space is then
L
2
(). A rotation R SO(3) of the system corresponds to a rotation R
1
of
the coordinate system, hence the rotation operator D(R) (where D stands for
the German word Drehung) changing the system according to the rotation is
8
/ = 1.0546 10
34
Js is the so-called reduced Planck constant.
D(R)f(x) = f(R
1
x). The map D : SO(3) Aut(L
2
()) is easily seen to be a
unitary representation of SO(3) (unitarity is a consequence of the transformation
theorem for integrals). Lets calculate the induced Lie algebra representation D
for the basis elements A

1
, A
2
and A
3
:
D
(A
3
)f(x) =
d
dt
t=0
f(exp(tA
1
)x)
=
d
dt
t=0
f(x
1
cos t x
2
sin t, x
1
sin t +x
2
cos t, x
3
)
=
d
dt
t=0
(x
1
cos t x
2
sin t)
f
x
1
+
d
dt
t=0
(x
1
sin t +x
2
cos t)
f
x
2
= x
2
f
x
1
+x
1
f
x
2
.
By a similar calculation for A
1
and A
2
we thus get
D
(A
1
) = x
2
f
x
3
x
3
f
x
2
(6.21a)
D
(A
2
) = x
3
f
x
1
x
1
f
x
3
(6.21b)
D
(A
3
) = x
1
f
x
2
x
2
f
x
1
. (6.21c)
From the commutation relations of the basis elements we immediately get similar
commutation relations for the operators above.
What are the C
-vectors in this case? Assume f L

2
() to be a C
-vector
and lets write its arguments in polar coordinates f(r, , ). Then the eect of
a rotation of the coordinate system (i.e. of the action of D(R)) will be to add
constants to the last variables:
(D(R)f)(r, , ) = f(r,
0
,
0
)
By the same argument as for translations C
-vectors have to be C
in the last
two variables.
To motivate why these operators are interesting we once again take a small
number and look at the innitesimal rotation r = exp(A
i
). Exploiting The-
orem 6.17 we get D(r) = exp(D
(A
i
)) id +D
(A
i
), i.e. D
(A
i
) is the rate
of change when rotating the system a little bit. This we interpret as orbital
angular momentum. Again to obtain self-adjoint operators with the right units
we dene the orbital angular momentum operators to be
L
x
:= i/ D
(A
1
) , L
y
:= i/ D
(A
2
) , L
z
:= i/ D
(A
3
).
Since SO(3) is a compact group the Peter-Weyl Theorem says that D can
be decomposed into irreducible representations. So rst of all lets determine
the irreducible representations of SO(3). First we observe that the following
matrices, known in physics as the Pauli spin matrices,
1
=
1
2
_
0 1
1 0
_
,
2
=
1
2
_
0 i
i 0
_
,
3
=
1
2
_
i 0
0 i
_
constitute a basis for su(2). One can calculate that
exp(t
1
) =
_
e
1
2
it
0
0 e
1
2
it
_
, exp(t
2
) =
_
cos(
1
2
t) sin(
1
2
t)
sin(
1
2
t) cos(
1
2
t)
_
exp(t
3
) =
_
cos(
1
2
t) i sin(
1
2
t)
i sin(
1
2
t) cos(
1
2
t)
_
.
Letting H, E and F denote the usual basis for the sl(2, C) (the complexication
of su(2)) we have the relations
E =
1
+i
2
, F =
1
+i
2
, H = 2i
3
. (6.22)
Furthermore we see that the Pauli matrices satisfy the commutation relations
[
1
,
2
] =
3
, [
2
,
3
] =
1
, [
3
,
1
] =
2
,
and thus we get a Lie algebra isomorphism : su(2) so(3) by
i
A
i
.
Since SU(2) is simply connected there is a unique Lie group homomorphism
: SU(2) SO(3) which induces . Since is an isomorphism is a smooth
covering map, namely the universal double covering of SO(3). In particular we
have I = ker .
SU(2)
N
N
N
N
N
N
N
N
N
N
N
N
N
SO(3)
G
If F : SU(2) G is any Lie group homomorphism it induces a Lie group
homomorphism

F : SO(3) G if and only if I ker F ( is a surjective
submersion so F is smooth if and only if

F is smooth).
Now the irreducible representations of su(2) are in 1-1 correspondence with
the irreducible representations of sl(2, C) and these we know. For each k N
there is exactly one representation
C
k
of sl(2, C) of dimension k, and thus for
su(2) there is for each k N one irreducible representation
k
of dimension
k. Since SU(2) is simply connected these lift to irreducible representations of
SU(2). Lets call them
k
.
From above we know that
k
induces a representation of SO(3) if and only
if I ker
k
, and the claim is that this happens exactly when k is odd. To
see this observe that exp(2
3
) = I and therefore
k
(I) =
k
(exp(2
3
)) = exp(
k
(2
3
)) = exp(i
C
k
(H))
where we used that
3
=
i
2
H. But
C
k
(H) is diagonal and the eigenvalues are
k 1, k 3, . . . , k + 3, k + 1 and therefore
exp(i
C
k
(H)) = exp(diag(i(k 1), . . . , i(k + 1)))
= diag(e
2i
k1
2
, . . . , e
2i
k+1
2
) =
_
I, k odd
I, k even.
The induced representation of SO(3) is irreducible, for the corresponding rep-
resentation of so(3)

= su(2) is just
k
which is irreducible. Furthermore there
can be no other irreducible representations of SO(3), for assuming the existence
of a such would give an irreducible representation of so(3) which is not among
the
k
s. Wrapping up: for each positive odd integer k there is one irreducible
representation of SO(3) of dimension k.
We now return to the representation D. By the Peter-Weyl Theorem we can
decompose this representation into irreducible representations. We dont know
how many times a specic irreducible representation occurs in this decomposi-
tion, it depends on the particular space . Let V
k
L
2
() denote an invariant
irreducible subspace of dimension k. Then we know that k is odd. If the state
vector of a particle happens to be in V
k
the physical interpretation is that the
square of the total orbital angular momentum of the particle is
2
4
(k 1)(k + 1).
The values above with k odd are the only possible values of the total orbital
angular momentum of a particle one can measure. But what about the angular
momentum along the axes? Let D
k
and D
k
denote the restriction of D and D
to the subspace V
k
. We see that on this subspace:
L
z
= i/D
k
(A
3
) = i/D
k
(
3
) =
1
2
/(D
k
)
C
(H)
and we know the eigenvalues of (D
k
)
C
(H): they are k1, k3, . . . , k+3, k+1
and therefore the eigenvalues of L
z
on V
k
are
k 1
2
/,
k 3
2
/, . . . ,
k + 3
2
/,
k + 1
2
/.
Observe that they are all integer multiples of /. In physics a state in V
k
which
is an eigenvector for L
z
with eigenvalue m/ is denoted [km.
By symmetry similar results hold for the other axes. But observe that L
x
, L
y
and L
z
do not commute, due to the commutation relations on so(3) and there-
fore the angular momentum along the axes cannot be simultaneously measured
unless they are all 0: Indeed assume that v V
k
is a common eigenvector for
L
x
, L
y
and L
z
and that the eigenvalue of, say, L
z
is non-zero. Then
0 ,= L
z
v = [L
x
, L
y
]v = 0
and we have a contradiction.
Part II
Geometric Analysis and Spin
Geometry
109
Chapter 7
Cliord Algebras
7.1 Elementary Properties
In this chapter we will introduce the Cliord algebra and discuss some of its
elementary properties. The setting is the following:
Let V be a nite-dimensional vector space over the eld K (predominantly 1
or C) and : V V K a symmetric bilinear form on V . is said to be
positive (resp. negative) denite if for all 0 ,= v V we have (v, v) > 0 (resp.
(v, v) < 0). is called non-degenerate if (v, w) = 0 for all v V implies
w = 0. From a bilinear form we construct a quadratic form : V K given
by (v) := (v, v). We can recover the original bilinear form by the polarization
identity:
(u +v) = (u +v, u +v) = (u) + (v) + 2(u, v),
hence
(u, v) =
1
2
((u +v) (u) (v)). (7.1)
Thus we have a 1-1 correspondence between symmetric bilinear forms and
quadratic forms. Thus a quadratic form is called positive denite/negative
denite/non-degenerate if is. Very important examples of non-degenerate bi-
linear forms on 1
n
are
p,q
with p + q = n where
p,q
relative to the standard
basis for 1
n
is given by the diagonal matrix diag(1, . . . , 1, 1, . . . , 1) of p pos-
itives and q negatives.
Denition 7.1 (Cliord Algebra). Let (V, ) be a vector space with a
quadratic form . The associated Cliord algebra Cl(V, ) (abbreviated Cl())
is an associative, unital algebra over K with a linear map i
: V Cl()
obeying the relation i(v)
2
= (v) 1 (1 is the unit element of Cl()). Further-
more (Cl(), i
) should have the property that for every unital algebra A, and
every linear map f : V A satisfying f(v)
2
= (v) 1 there exists a unique
algebra homomorphism

f : Cl() A, such that f =

f i
.
Two questions immediately arise: given (V, ), do such objects exist and if
they do, are they unique? Fortunately, the answer to both questions is yes:
Proposition 7.2. For any vector space V with a quadratic form let be the
two-sided ideal in the tensor algebra T(V ) spanned by all elements of the form
a (v v (v) 1) b (for a, b T(V ) and v V ). Then T(V )/, with the
map i
: V T(V )/ being where : V T(V ) is the injection of V

into T(V ), and : T(V ) T(V )/ is the quotient map, is a Cliord algebra,
and any other Cliord algebra over (V, ) is isomorphic to this one.
111
112 Chapter 7 Cliord Algebras
Proof. Uniqueness. Assume that Cl
1
() and Cl
2
() with linear maps i
1
:
V Cl
1
() and i
2
: V Cl
2
() are Cliord algebras. Since Cl
1
() is
a Cliord algebra, and i
2
is linear and satises i
2
(v)
2
= (v) 1, it induces
an algebra homomorphism

i
2
: Cl
1
() Cl
2
(); likewise i
1
induces an al-
gebra homomorphism
i
1
: Cl
2
() Cl
1
() such that the following diagram
commutes:
V
i1
.x
x
x
x
x
x
x
x
x
i2
F
F
F
F
F
F
F
F
F
Cl
1
()
i2
Cl
2
()
i1
We see that
i
2
=
i
2
i
1
=
i
2
i
1
i
2
and since
i
2
i
1
is the unique map satisfying this, it must be id
Cl2()
. Likewise
i
1

i
2
= id
Cl1()
which means that the two Cliord algebras are isomorphic.
This proves uniqueness.
Existence. We now show that Cl() := T(V )/ is indeed a Cliord algebra. i
is easily seen to satisfy i
(v)
2
= (v) 1, where 1 Cl() is the coset containing
the unit element of T(V ). Now let f : V A be linear with f(v)
2
= (v)1. By
the universal property of the tensor algebra this map factorizes uniquely through
T(V ) to an algebra homomorphism f
t
: T(V ) A, such that f = f
t
. f
t
inherits the property (f
t
(v))
2
= (v) 1, and consequently
f
t
(v v (v) 1) = (f
t
(v))
2
(v)f
t
(1) = (f
t
(v))
2
(v) 1
so f
t
vanishes on . Therefore it factorizes uniquely through Cl() to

f :
Cl() A, such that f =

f =

f i
. Thus Cl() = T(V )/ is a

Cliord algebra.
We immediately see that if the quadratic form on V is identically 0, the
Cliord algebra Cl() is nothing but the well-known exterior algebra
(V ).
From now on we will write i instead of i
where no confusion is possible. A

simple calculation reveals that for all u, v V : i(u +v)
2
= i(u)
2
+i(v)
2
+i(u)
i(v) +i(v) i(u). A comparison of this with the polarization identity using that
(u+v) 1 = i(u+v)
2
, (u) 1 = i(u)
2
and (v) 1 = i(v)
2
yields the following
useful formula:
i(u) i(v) +i(v) i(u) = 2(u, v) 1. (7.2)
It can be used to prove the following:
Proposition 7.3. Let e
1
, . . . , e
n
be an orthogonal basis for V , then the set
consisting of 1 and all products of the form i(e
j1
) i(e
j
k
), where j
1
< < j
k
and 1 k n is a basis for Cl(). In particular i : V Cl() is injective,
and the dimension of Cl() is 2
n
.
Proof. Since e
i1
e
i
k
span the tensor algebra and since i(e
j
) are subject to
the relations (7.2) we see that elements of the form e
i1
e
i
k
where i
1
< < i
k
span Cl(). In particular dimCl() 2
n
.
For each v V dene endomorphisms (v), (v) and c(v) on
V by
(v) : v
1
v
k
v v
1
v
k
.
(v) : v
1
v
k

k
j=1
(1)
j+1
(v, v
j
)v
1
v
j
v
k
.
c(v) = (v) +(v).
7.1 Elementary Properties 113
Its a matter of calculations to show that (v)
2
= (v)
2
= 0 and that we have
c(v)
2
= (v) 1
c(u)c(v) +c(v)c(u) = 2(u, v) 1.
By the universal property of Cliord algebras there exists an algebra homo-
morphism c : Cl() End
K
(
V ) extending c. In particular we get a linear

map
: Cl()
V
by x c(x)1, called the symbol map. We observe that
(e
i1
e
i
k
) = c(e
i1
) c(e
i
k
)1 = e
i1
e
i
k
and thus that is surjective. Hence dimCl() 2
n
.
Since i is injective we can imagine V as sitting as a subspace of Cl(). Thus,
henceforth we will write v instead of i(v).
The inverse of the symbol map Q :
V Cl(), mapping e
i1
e
i
k
to
e
i1
e
i
k
is called the quantization map.
If (V, ) and (W, ) are two vector spaces with bilinear forms, a linear map
f : V W is called orthogonal w.r.t. the bilinear forms if (f(v), f(w)) =
(v, w) (or equivalently, if (f(v)) = (v)). We say that (V, ) and (W, ) are
isomorphic if there exists an orthogonal linear map f : V W which is also a
vector space isomorphism and such a map we call an orthogonal isomorphism. If
is non-degenerate, one can show that an orthogonal endomorphism f : V
V is automatically an isomorphism. Thus the set of orthogonal endomorphisms
of V is a group O(V, ) (or just O()), called the orthogonal group. This is a
closed subgroup of the Lie group GL(V ), and thus itself a Lie group. Picking
only those orthogonal endomorphisms having determinant 1 gives the special
orthogonal group SO(V, ) or just SO(). This is again a Lie group, being a
closed subgroup of O().
A curious fact is that there exists an isomorphism among the orthogonal
groups. To see this, let : (1
p+q
,
p,q
) (1
p+q
,
q,p
) be an anti-orthogonal
linear map, i.e. satisfying
q,p
((x)) =
p,q
(x)
and consider the map : O(p, q) O(q, p) given by (A) = A
1
. It is
easily checked that this is a Lie group isomorphism. Restricting to SO(p, q),
it maps into SO(q, p) since the determinant is unaltered by a conjugation. Thus
we also have an isomorphism SO(p, q)

SO(q, p).
The next proposition shows that Cl is a covariant functor from the category
of vector spaces with bilinear forms and orthogonal linear maps to the category
of associated unital algebras and algebra homomorphisms.
Proposition 7.4. An orthogonal linear map f : V W between (V, ) and
(W, ) induces a unique algebra homomorphism f : Cl() Cl() satisfying
f(i
(v)) = i
(f(v)). If f is an orthogonal isomorphism, then f is an algebra

isomorphism.
If, furthermore, we have a vector space U with quadratic form and linear
maps f : V U and g : U W satisfying (f(v)) = (v) and (g(u)) =
(u), then g f = g f.
Proof. It is a simple application of the universal property of the Cliord alge-
bra. Consider the following commutative diagram:
V
i
W
i
Cl()
f
Cl()
The map i
f : V Cl() satises the condition to factorize through Cl():

((i
f)(v))
2
= i
(f(v))
2
= (f(v)) 1 = (v) 1.
This gives the unique map f. If f is bijective, f
1
gives the unique homomor-
phism f
1
: Cl() Cl(). An argument similar to the one in the proof of
Proposition 7.2 shows that f and f
1
are inverses of each other. The last claim
of the lemma follows from the uniqueness, in that gf : Cl() Cl() clearly
satises (g f)(i
(v)) = i
((g f)(v)).
The most interesting examples of Cliord algebras show up when the bilin-
ear form is non-degenerate. For the time being we consider only real vector
spaces and Cliord algebras. A prominent example of a vector space with non-
degenerate bilinear form is (1
p+q
,
p,q
) where
p,q
(e
i
, e
j
) = 0 if i ,= j and
p,q
(e
i
, e
i
) =
_
1, i p
1, i > p
(e
1
, . . . , e
p+q
is the standard basis of 1
p+q
). The associated quadratic form is
denoted
p,q
and the corresponding Cliord algebra is denoted Cl
p,q
. In this
case O(
p,q
) and SO(
p,q
) are the well-known orthogonal groups O(p, q) and
SO(p, q).
Example 7.5. 1) Let us consider the vector space 1 with the single basis
element e
1
:= 1 and the quadratic form
0,1
(x
1
e
1
) = x
2
1
. By Proposition 7.3
1, e
1
(where 1 is now the unit element of Cl
0,1
) is a basis for Cl
0,1
. The fact
that e
2
1
=
0,1
(e
1
) 1 = 1 shows that the linear map Cl
0,1
1 1 C,
e
1
i C denes an algebra isomorphism Cl
0,1
C and that the injection

1 Cl
0,1
is given by x ix, i.e. 1 sits inside Cl
0,1

= C as the imaginary
part. Thus, Cl
0,1
is just the eld of complex numbers.
2) As another example, consider 1
2
with the standard basis e
1
, e
2
and the
quadratic form
0,2
(x
1
e
1
+x
2
e
2
) = x
2
1
x
2
2
. The Cliord algebra has the basis
1, e
1
, e
2
, e
1
e
2
. It is an easy application of formula (7.2) to show that the linear
map from Cl
0,2
to the algebra of quaternions H given by 1 1, e
1
i,
e
2
j and e
1
e
2
k is in fact an algebra isomorphism Cl
0,2
H.
These two will be our role models throughout the text. Because of this last
example, elements of Cliord algebras are sometimes called generalized quater-
nions.
Now, let (V, ) be any real vector space with non-degenerate bilinear form ,
and let e
1
, . . . , e
n
be a basis for V . We will consider the matrix (
ij
) where
ij
= (e
i
, e
j
). Since is symmetric the matrix (
ij
) is symmetric as well, i.e.
it can be diagonalized. Let
1
, . . . ,
n
be the eigenvalues and f
1
, . . . , f
n
a basis
of diagonalizing eigenvectors. This means that
(f
i
, f
j
) = 0 if i ,= j and (f
i
, f
i
) =
i
.
7.1 Elementary Properties 115
Observe, that none of the eigenvalues are 0 (a 0 eigenvalue would violate non-
degeneracy of ). Arrange the eigenvalues so that
1
, . . . ,
k
are all strictly
positive and
k+1
, . . . ,
n
are all strictly negative. Dene
f
i
=
1
_
[
i
[
f
i
then we see that
(
f
i
,

f
j
) = 0 if i ,= j and (
f
i
,
f
i
) =
_
1, i k
1, i > k
.
A basis satisfying this is called a (real) orthonormal basis w.r.t. . Thus we have
proven
Theorem 7.6 (Classication of Real Bilinear Forms). Let (V, ) be a real
vector space with non-degenerate bilinear form. Then there exists an orthonor-
mal basis for V and the map sending this basis to the standard basis for 1
n
is
an orthogonal isomorphism (V, ) (1
n
,
k,nk
).
In particular any real Cliord algebra originating from a non-degenerate
quadratic form is isomorphic to Cl
p,q
for a certain p and q (Proposition 7.4).
The complex case is a bit dierent. On C
n
we have bilinear forms
n
given by
n
(e
i
, e
j
) =
ij
. The corresponding quadratic form is denoted
n
. For an arbi-
trary complex vector space (V, ), chose a basis e
1
, . . . , e
n
and write in this
basis (
ij
). Again, since it is symmetric, it can be diagonalized. Let
1
, . . . ,
n
be the eigenvalues and f
1
, . . . , f
n
a diagonalizing basis of eigenvectors. None
of the eigenvalues are 0 and hence we can dene
f
i
=
1
i
f
i
.
This basis satises (
f
i
,

f
j
) =
ij
, and is thus called a (complex) orthonormal
basis. Thus we have shown
Theorem 7.7 (Classication of Complex Bilinear Forms). Let (V, ) be
a complex vector space with non-degenerate bilinear form. Then there exists an
orthonormal basis for V and the map sending this basis to the standard basis
for 1
n
is an orthogonal isomorphism (V, ) (C
n
,
n
).
Appealing to Proposition 7.4 we see that a complex Cliord algebra is iso-
morphic to Cl(
n
) for some n.
Again we consider the general situation of Cliord algebras over K. Now we
want to equip the Cliord algebra with two involutions t and which we will
need later in the construction of various subgroups of Cl().
Proposition 7.8. Each Cliord algebra Cl() admits a canonical anti-auto-
morphism, i.e. a linear map t : Cl() Cl() that for all x, y Cl()
satises
t(x y) = t(y) t(x) , t t = id
Cl()
, t[
V
= id
V
.
Proof. Consider the involution J of the tensor algebra given by v
1
v
k
J
v
k
v
1
(and extended by linearity). Now J : T(V ) Cl() is an
anti-homomorphism that vanishes on since
J(a v v b a ((v) 1) b)
= J(b) v v J(a) J(b) ((v) 1) J(a)
= J(b) (v v (v) 1) J(a) .
But then J induces a unique anti-homomorphism t : Cl() Cl() deter-
mined by t[x] = J(x) (where [x] Cl() is the coset containing x T(V )).
It is easy to see that J(x y) = J(y) J(x), so we also have t(x y) = t(y) t(x).
J is clearly an involution, i.e. J J = id
T(V )
. On one hand id
T(V )
induces the
map id
Cl()
, and on the other hand J J induces the map t t. By uniqueness
we conclude t t = id
Cl()
. The last property, t[
V
= id
V
follows from the fact
that J(v) = v for v V .
Now for the construction of the second involution:
Proposition 7.9. Each Cliord algebra Cl() admits a canonical automor-
phism : Cl() Cl() which satises = id
Cl()
and [
V
= id
V
.
Furthermore we have (1) = 1 and (e
i1
e
i
k
) = (1)
k
e
i1
e
i
k
.
Proof. Consider the linear bijection : V V given by v v. By the
functorial property of Cl it induces an automorphism : Cl() Cl(), and
we see that
= = id
V
= id
Cl()
.
The second property of is obtained from the identity i = i which is
seen from the commutative diagram in the proof of Proposition 7.4. This gives
(i(v)) = i(v) = i(v), and by considering v an element of Cl() (by virtue
of the injectivity of i) we have proven the claim.
The the calculation of t and on our model algebras C and H we refer to
Example 8.6 in the next chapter.
Due to involutivity of , we can split the the Cliord algebra in a direct sum
of two subspaces:
Cl() = Cl
0
() Cl
1
() (7.3)
where Cl
i
() = x Cl() [ (x) = (1)
i
x for i = 0, 1. It is easily seen that
Cl
i
() Cl
j
() Cl
i+j (mod 2)
() which says that the Cliord algebra is a Z
2
-
graded algebra or a super algebra. Cl
0
() is called the bosonic subalgebra (note
that it is actually a subalgebra), and Cl
1
() is called the fermionic subspace.
We see that a product of the form v
1
v
k
for v
i
V is bosonic if k is even
and fermionic if k is odd. Elements of Cl
0
() Cl
1
() are called homogenous
elements and the degree of a homogenous element x is denoted by [x[.
If A and B are two Z
2
-graded algebras we could form the tensor product
A B with the usual product (a b)(a
t
b
t
) = (aa
t
) (bb
t
). However, this
is usually not of great interest since the resulting algebra is not Z
2
-graded (at
least not non-trivially). Therefore we dene the so-called graded tensor product
or super tensor product A
B in the following way: As a vector space it is just

the ordinary tensor product A B but the product is given on homogenous
elements a, a
t
A and b, b
t
B by
(a b)(a
t
b
t
) = (1)
]a
]]b]
(aa
t
) (bb
t
).
This gives A
B a natural grading by dening

(A
B)
0
= (A
0
B
0
) (A
1
B
1
)
(A
B)
1
= (A
0
B
1
) (A
1
B
0
).
With this at hand we can accomplish out nal task of this section, showing how
Cl reacts to a direct sum of vector spaces. By an orthogonal decomposition of
(V, ), we understand a decomposition V = V
1
V
2
such that if v = v
1
+v
2
we
have (v) =
1
(v
1
) +
2
(v
2
) (equivalently if (V
1
, V
2
) = 0).
7.2 Classication of Cliord Algebras 117
Proposition 7.10. Assume we have an orthogonal decomposition V = V
1
V
2
of (V, ), then Cl(V, )
= Cl(V
1
,
1
)
Cl(V
2
,
2
) as algebras.
Proof. We use the universal property of Cliord algebras to cook up a map.
First, dene g : V Cl(V, )
= Cl(V
1
,
1
)
Cl(V
2
,
2
) by
g(v) = v
1
1 + 1 v
2
.
A quick calculation shows that g(v)
2
= (v)(1 1) and thus by the univer-
sal property of Cl there exists an algebra homomorphism g : Cl(V, )
Cl(V
1
,
1
)
Cl(V
2
,
2
) extending g. To see that this is indeed an isomorphism
choose a basis e
1
, . . . , e
m
for V
1
and a basis f
1
, . . . , f
n
for V
2
and put them
together to a basis e
1
, . . . , f
n
for V . A basis for Cl(V, ) consists of elements of
the form e
i1
e
i
k
f
j1
f
j
l
where i
1
< < i
k
, k m and j
1
< < j
l
, l n.
Similarly a basis for Cl(V
1
,
1
)
Cl(V
2
,
2
) is given by e
i1
e
i
k
f
j1
f
j
l
(with the same restrictions on the indices as above). One can verify that
g(e
i1
e
i
k
f
j1
f
j
l
) = e
i1
e
i
k
f
j1
f
j
l
i.e. it maps basis to basis. Thus it is an algebra isomorphism.
7.2 Classication of Cliord Algebras
In this section we set out to classify Cliord algebras originating from non-
degenerate bilinear forms. At rst we concentrate on real Cliord algebras. We
have taken the rst step in the classication in that we have shown in the
previous section, that it is enough to concentrate our treatment of Cliord
algebras to the case where the vector space is 1
p+q
equipped with the quadratic
form
p,q
(x
1
e
1
+ +x
p+q
e
p+q
) := x
2
1
+ +x
2
p
(x
2
p+1
+ +x
2
p+q
)
(where e
1
, . . . , e
p+q
denotes the usual standard basis for 1
p+q
). The associated
Cliord algebra was denoted Cl
p,q
. As we saw in Example 7.5, we have
Cl
0,1

= C and Cl
0,2

= H. (7.4)
Likewise, one can show that the following isomorphisms hold
Cl
1,0

= 1 1 and Cl
2,0

= Cl
1,1

= 1(2) (7.5)
(K(n) is the algebra of n n matrices over K, where K can be either 1, C or
H). As is apparent, all four Cliord algebras are isomorphic to either an algebra
of matrices over 1, C or H, or to a direct sum of two such algebras. The goal of
this section is to show that this is no coincidence. Actually, it is a consequence
of the Cartan-Bott Periodicity Theorem which we will prove at the end of this
section that this holds for every Cliord algebra Cl
p,q
.
Before proving it, we need two lemmas:
Lemma 7.11. We have the following algebra isomorphisms:
1(m) 1(n)
= 1(mn) C
R
C
= C C C
R
H
= C(2) H
R
H
= 1(4).
If K denotes either C or H, then
1(n)
R
K
= K(n).
This is well-known so we wont prove it here.
1
Instead we show the next
lemma which really does most of the work:
Lemma 7.12. We have the following three algebra isomorphisms:
Cl
0,n+2

= Cl
n,0
Cl
0,2
, Cl
n+2,0

= Cl
0,n
Cl
2,0
and
Cl
p+1,q+1

= Cl
p,q
Cl
1,1
.
for all p, q, n N 0.
Proof. To prove the rst isomorphism, the strategy is the following: we will
construct a linear map f : 1
n+2
Cl
n,0
Cl
0,2
, show that f factorizes
through Cl
0,n+2
, and that the induced map

f : Cl
0,n+2
Cl
n,0
Cl
0,2
is
an isomorphism. Letting e
1
, . . . , e
n+2
denote the standard basis for 1
n+2
,
e
t
1
, . . . , e
t
n
the usual basis for 1
n
, and e
tt
1
, e
tt
2
the usual basis for 1
2
, and
thereby generators for the Cliord algebras Cl
0,n+2
, Cl
n,0
and Cl
0,2
respectively,
we dene f by
f(e
i
) =
_
e
t
i
e
tt
1
e
tt
2
if 1 i n
1 e
tt
in
if n + 1 i n + 2
For 1 i, j n we compute, using the rules for multiplication in a Cliord
algebra, that
f(e
i
) f(e
j
) +f(e
j
) f(e
i
) = (e
t
i
e
tt
1
e
tt
2
) (e
t
j
e
tt
1
e
tt
2
) + (e
t
j
e
tt
1
e
tt
2
) (e
t
i
e
tt
1
e
tt
2
)
= (e
t
i
e
t
j
) (e
tt
1
e
tt
2
e
tt
1
e
tt
2
) + (e
t
j
e
t
i
) (e
tt
1
e
tt
2
e
tt
1
e
tt
2
)
= (e
t
i
e
t
j
+e
t
j
e
t
i
) (e
tt
1
e
tt
1
e
tt
2
e
tt
2
) = 2
ij
1 1
because e
t
i
e
t
j
+e
t
j
e
t
i
= 2
ij
1, as e
t
1
, . . . , e
t
n
is basis for 1
n
orthonormal w.r.t.
n,0
. For n + 1 i, j n + 2 we have
f(e
i
) f(e
j
) +f(e
j
) f(e
i
) = (1 e
tt
in
) (1 e
tt
jn
) + (1 e
tt
jn
) (1 e
tt
in
)
= 1 e
tt
in
e
tt
jn
+ 1 e
tt
jn
e
tt
in
= 1 (e
tt
in
e
tt
jn
+e
tt
jn
e
tt
in
) = 2
ij
1 1
where the last minus is due to e
tt
1
, e
tt
2
being a basis for 1
2
orthonormal w.r.t.
0,2
. A similar computation shows that f(e
i
)f(e
j
) +f(e
j
)f(e
i
) = 0 if 1 i n
and n + 1 j n + 2. But then for x = x
1
e
1
+ + x
n+2
e
n+2
we have by
linearity of f that
f(x)
2
= (x
2
1
+ +x
2
n+2
)1 1 =
0,n+2
(x) 1 1.
Therefore f factorizes uniquely through Cl
0,n+2
to an algebra homomorphism
f : Cl
0,n+2
Cl
n,0
Cl
0,2
. f maps to a set of generators for Cl
n,0
Cl
0,2
;
thus

f maps to a set of generators for Cl
n,0
Cl
0,2
. Since

f is an algebra ho-
momorphism,

f must then be surjective. Since
dimCl
0,n+2
= 2
n+2
= 2
n
2
2
= (dimCl
n,0
)(dimCl
0,2
) = dim(Cl
n,0
Cl
0,2
),
the Dimension Theorem from linear algebra tells us that

f is also injective. Thus
f is the desired isomorphism.

The second isomorphism is proved in exactly the same way, and we avoid
repeating ourselves.
1
For a proof consult for instance [Lawson and Michelson], pp 26-27, Proposition 4.2.
7.2 Classication of Cliord Algebras 119
The proof of the third isomorphism is essentially the same as the two rst. Let
e
1
, . . . , e
p+1
,
1
, . . . ,
q+1
be an orthogonal basis for 1
p+q+2
(i.e. (v, w) = 0
when v ,= w) with the quadratic form
p+1,q+1
, such that
p+1,q+1
(e
i
) = 1,
p+1,q+1
(
i
) = 1, and let e
t
1
, . . . , e
t
p
,
t
1
, . . . ,
t
q
and e
tt
1
,
tt
1
be similar bases
for 1
p+q
and 1
1+1
(and thereby generators for the Cliord algebras Cl
p+1,q+1
,
Cl
p,q
and Cl
1,1
respectively). We now dene a linear map f : 1
p+q+2
Cl
p,q
Cl
1,1
by
f(e
i
) =
_
e
t
i
e
tt
1
tt
1
if 1 i p
1 e
tt
1
if i = p + 1
, f(
j
) =
_
t
j
e
tt
1
tt
1
if 1 j q
1
tt
1
if j = q + 1
.
Just like before it can be shown that
f(x)
2
=
p+1,q+1
(x) 1 1,
and thus f induces an isomorphism

f : Cl
p+1,q+1
Cl
p,q
Cl
1,1
.
Now we are ready to state and prove the Cartan-Bott Theorem:
Theorem 7.13 (Cartan-Bott I). We have the following isomorphisms:
Cl
0,n+8

= Cl
0,n
1(16) and Cl
n+8,0

= Cl
n,0
1(16).
Proof. Using the two rst isomorphisms from Lemma 7.12 a couple of times
yields
Cl
0,n+8

= Cl
n+6,0
Cl
0,2

=
= Cl
0,n
Cl
2,0
Cl
0,2
Cl
2,0
Cl
0,2
= Cl
0,n
Cl
2,0
Cl
2,0
Cl
0,2
Cl
0,2
where the last isomorphism follows from the fact that for arbitrary real algebras
A and B we have AB

= BA. From (7.4) we have Cl
0,2

= H, and from (7.5)
that Cl
2,0

= 1(2). Thus, using H
R
H
= 1(4) from Lemma 7.11, we get

Cl
2,0
Cl
2,0
Cl
0,2
Cl
0,2

= 1(2) 1(2) (H
R
H)
= 1(4) 1(4)
= 1(16).
This completes the proof of the rst isomorphism. The proof of the second
isomorphism is identical to this.
Now its evident that once we know the Cliord algebras Cl
0,0
, Cl
0,1
, . . . , Cl
0,7
and Cl
0,0
, Cl
1,0
, . . . , Cl
7,0
we know all of them: Consider Cl
p,q
and assume that
p q. Then by the third isomorphism in Lemma 7.12 we have that Cl
p,q
is iso-
morphic to Cl
pq,0
(Cl
1,1
)
q
= Cl
pq,0
1(2
q
), and Cl
pq,0
can be expressed
as a tensor product of Cl
k,0
(with 0 k 7) and some copies of 1(16). We can
do the same if q p. Using the isomorphisms from Lemma 7.11 one obtains
the table of Cliord algebras in Appendix A. From this table we see that any
Cliord algebra is either a matrix algebra or a sum of two such algebras, as we
pointed out in the beginning of this section.
Example 7.14. As an example, let us show that Cl
2,11

= C(64):
Cl
2,11

= Cl
0,9
Cl
1,1
Cl
1,1

= Cl
0,1
1(16) 1(2) 1(2)
= C
R
1(64)
= C(64) .
As yet another example, let us show that Cl
3,2

= 1(4) 1(4):
Cl
3,2

= Cl
1,0
Cl
1,1
Cl
1,1

= (1 1) 1(4)
= (1 1(4)) (1 1(4))
= 1(4) 1(4),
where we have used that the tensor product is distributive w.r.t. and that for
any real algebra A the isomorphism 1
R
A
= A holds.
By repeating the arguments of Example 7.14 in a more general setting we
obtain:
Corollary 7.15. Cl
p,q
is a direct sum of two matrix algebras exactly when
p q 1 (mod 4).
We conclude this treatment by proving
Proposition 7.16. We have the following algebra isomorphism Cl
p,q

= Cl
0
p,q+1
.
Proof. Let e
1
, . . . , e
p+q+1
denote an orthogonal basis for 1
p+q+1
which sat-
ises (e
i
) = 1 for i = 1, . . . , p and (e
j
) = 1 for j = p + 1, . . . , p + q + 1.
Assume the basis has been chosen so that e
1
, . . . , e
p+q
is a basis for 1
p+q
.
Now dene a linear map f : 1
p+q
Cl
0
p,q+1
by
f(e
i
) = e
p+q+1
e
i
for i p+q. Like in the proof of Lemma 7.12 one checks that f satises f(x)
2
=
(x) 1 and thus factorizes to an algebra homomorphism

f : Cl
p,q
Cl
0
p,q+1
.
By inspection, this is the desired isomorphism.
For the rest of this section we will consider complex Cliord algebras. It turns
out that complex Cliord algebras behave even nicer than their real counter-
parts. As we have already seen a complex Cliord algebra associated with a
non-degenerate bilinear form is isomorphic to Cl(
n
). The fact that there is
only one index on and not two as in the real case, indicates some sort of
simplication.
But rst we introduce a way of turning real vector spaces/algebras into com-
plex ones:
Denition 7.17 (Complexication). By the complexication of a real vector
space V we mean the real tensor product V
C
:= V
R
C. If If the vector space
V carries a bilinear form , then the complexication of the bilinear form is
C
(v , v
t

t
) :=
t
(v, v
t
). The complexication of the corresponding
quadratic form is then
C
(v ) =
2
(v).
In the same way we dene the complexication of a real algebra A by A
C
:=
A
R
C carrying the product
(a )(a
t
t
) = (aa
t
) (
t
).
Assume (V, ) to be a real vector space with a non-degenerate bilinear form.
Then
C
is a non-degenerate bilinear form on V
C
for assume v
0
0
to satisfy
C
(v
0
0
, v ) =
0
(v
0
, v) ,= 0 for all v . Then
0
,= 0 and (v
0
, v) ,= 0
for all v and by non-degeneracy of this implies that v
0
,= 0. Thus v
0
0
,= 0.
In particular, if p + q = n we have that
C
p,q
is equivalent to
n
and thus
C
p,q
is equivalent to
n
.
Now, one can pose the question: is the complexication of a real Cliord
algebra a complex Cliord algebra? By the following lemma the answer is yes.
Lemma 7.18. For p +q = n we have Cl(
C
p,q
)
= Cl(
n
)
= Cl
C
0,n
.
Proof. The rst isomorphism is due to the fact that the complexication of
p,q
is equivalent to
n
and thus the corresponding Cliord algebras are iso-
morphic.
To verify the second isomorphism we construct a linear map : C
n
Cl
C
0,n
and show that it factorizes to an isomorphism : Cl(
n
)

Cl
C
0,n
. At rst,
7.3 Representation Theory 121
we remark that C
n
= 1
n
R
C. We then dene by (v z) = i(v) z where
i : 1
n
Cl
0,n
denotes the usual embedding. Since
(v z)
2
= (i(v) z)
2
= i(v)
2
z
2
=
0,n
(v)z
2
1 1 =
C
n
(v z) 1 1,
factorizes uniquely to an algebra homomorphism : Cl(
n
) Cl
C
0,n
. Both
algebras have complex dimension 2
n
so its enough to show that is surjective.
But is surjective since is an algebra homomorphism and maps onto a
set of generators of Cl
C
0,n
. Namely, the set of elements of the form i(v) generate
Cl
0,n
, and 1 C generates C.
Thus, henceforth we will stick to the notation Cl
C
n
for the complex Cliord
algebra over C
n
equipped with any non-degenerate quadratic form, since the
preceding lemma guarantees that they are all isomorphic.
This result, in combination with the classication results for real Cliord
algebras, we get the complex version of the Cartan-Bott Theorem:
Theorem 7.19 (Cartan-Bott II). We have the following2-periodicity: Cl
C
n+2
=
Cl
C
n
C
Cl
C
2
, and furthermore that Cl
C
2
= C(2).
Proof. Invoking Lemma 7.12 and Lemma 7.18 we obtain the following chain
of isomorphisms:
Cl
C
n+2
= Cl
0,n+2
R
C
= Cl
n,0
R
C
R
Cl
0,2
= Cl
n,0
R
(C
C
C)
R
Cl
0,2

= (Cl
n,0
R
C)
C
(C
R
Cl
0,2
)
= Cl
C
n
C
Cl
C
2
.
For the second isomorphism, just recall that Cl
0,2

= H and H
R
C
= C(2).
Remembering that Cl
1,0

= 1 1 so that Cl
C
1
= C C, we obtain:
Corollary 7.20. If n = 2k, then Cl
C
n
= C(2
k
). If n = 2k + 1, then Cl
C
n
=
C(2
k
) C(2
k
).
7.3 Representation Theory
In this section we will turn our attention to the representation theory of Cliord
algebras. For the following denition, let us denote by K either 1, C or H:
Denition 7.21 (Algebra Representation). Let A be an algebra over K and
V a vector space over K. A K-representation of A is an algebra homomorphism
: A End
K
(V ). A subspace U of V is called invariant under if (x)U U
for all x A. The representation is called irreducible if the only invariant
subspaces are 0 and V .
By an intertwiner of two representations and
t
of A on V and V
t
we
understand a linear map f : V V
t
satisfying
t
(x) f = f (x)
for all x A. Two representations are called equivalent if there exists an inter-
twiner between them which is also an isomorphism of vector spaces.
Just as we had complexication of a real algebra, we can complexify complex
representations: If : A End(V ) is a representation of a real algebra on a
complex vector space V , we dene the complexication
C
: A
C
End(V ) by
C
(x) = (v). The notions of invariant subspaces and irreducibility of a rep-
resentation and its complexied are closely related as the following proposition
shows
Proposition 7.22. Let A be a real unital algebra, and an algebra representa-
tions on the complex V . Let A
C
and
C
denote the associated complexications.
Then the following hold:
1) A subspace W V is -invariant if and only if it is
C
-invariant.
C
is irreducible.
Proof. 1) If W is -invariant then
C
(x )W = (x)W W.
Conversely, if W is
C
-invariant then for x A we have (x)W =
C
(x1)W
W. This proves 1).
2) Follows immediately from 1).
A representation of A on V gives V the structure of a left A-module simply
by dening a v := (a)v. This is compatible with addition in V .
The next proposition actually contains all the information we need to de-
termine all the irreducible representations (up to equivalence) of the Cliord
algebras:
Proposition 7.23. The matrix algebra K(n) has only one irreducible K-repre-
sentation, namely the dening representation i.e. the natural isomorphism
n
:
K(n)

End
K
(K
n
). The algebra K(n) K(n) has exactly 2 inequivalent irre-
ducible K-representations, namely:
0
n
(x
1
, x
2
) :=
n
(x
1
) and
1
n
(x
1
, x
2
) :=
n
(x
2
). (7.6)
We saw in the previous section that Cl
p,q
is of the form K(n) K(n) i
p q 1 (mod 4) and a matrix algebra otherwise. This observation along side
with the preceding proposition yields the number of irreducible representations
of the real Cliord algebras.
But there is a slight problem here, in that the real Cliord algebra is not
always a real matrix algebra or a sum of real matrix algebras. Thus the irre-
ducible representations of Proposition 7.23 need not be real! For instance Cl
1,4

=
H(2)H(2), and Proposition 7.23 gives us two irreducible H-representations over
H
2
! But fortunately we can always turn a complex or quaternionic representa-
tion into a real representation if we just remember to adjust the dimension.
2
Without going further into details with keeping track of the dimensions we have
shown:
Theorem 7.24. Consider the real Cliord algebra Cl
p,q
. If p q 1 (mod
4) Cl
p,q
has up to equivalence two real irreducible representations and up to
equivalence exactly one real irreducible representation otherwise.
In particular if n 1 (mod 4) Cl
0,n
has two irreducible representations,
0
n
and
1
n
, and otherwise only one
n
. If n 1 (mod 4) we dene
n
:=
0
n
such that
to each real Cliord algebra Cl
0,n
we associate a real irreducible representation
n
called the real spin representation. The elements of the corresponding vector
spaces are called spinors.
The similar situation for complex Cliord algebras is simpler due to the fact
that each complex Cliord algebra decomposes into complex matrix algebras
(cf. Corollary 7.20). Thus all irreducible representations are complex. This will
make it a lot easier to keep track of the dimensions.
2
C
n
is naturally isomorphic to 1
2n
via the isomorphism : C
n
1
2n
, (
1
, . . . , n)
(Re
1
, Im
1
, . . . , Re n, Imn). If is a complex representation of an algebra A, then we
just dene a real representation on 1
2n
by (x)(v) = ((x)
1
(v)). Its easy to see that
is irreducible i is. Likewise with a quaternionic representation we exploit the natural
isomorphism H
n
= 1
4n
.
7.3 Representation Theory 123
Theorem 7.25. Consider the complex Cliord algebra Cl
C
n
. If n = 2k we have
(up to equivalence) exactly one irreducible complex representation
n
on C
2
k
,
namely the isomorphism Cl
C
2k
End(C
2
k
).
If n = 2k + 1 we have (up to equivalence) exactly two irreducible representa-
tions
0
n
and
1
n
on C
2
k
.
For n = 2k or n = 2k + 1 like above, dene
n
:= C
2
k
. The elements of
n
are called Dirac spinors or complex n-spinors. The irreducible representations
of Cl
C
n
are representations on
n
.
In the case where n is odd we want to single out
0
n
and dene
n
:=
0
n
which is just the composition
n
=
0
n
: Cl
C
n
End
C
(
n
) End
C
(
n
)
1
End
C
(
n
)
of the isomorphism with the projection
1
onto the rst component. Hence, for
each n we have an irreducible complex representation on
n
called the complex
spin representation.
Finally, lets try to break up the action of the Cliord algebra into smaller
pieces and see how they act on the spinors. This requires introduction of the
so-called volume element. It is well-known that,
(1
n
) has a unique volume
element given unambiguously, in any orthonormal basis e
1
, . . . , e
n
, by =
e
1
e
n
. Applying the quantization map to this yields an element :=
Q() Cl
0,n
, also called the volume element, given by = e
1
e
n
. For the
complex Cliord algebra Cl
C
n
we dene the volume element by
C
:= i
n+1
2

.
In the case n = 2k we note that
2
= (1)
k
and that commutes with every
element of Cl
0
0,2k
, while anti-commutes Cl
1
0,2k
, for instance we see that
e
1
= (e
1
e
2k
)e
1
= (1)
2k1
e
2
1
e
2
e
2k
= e
1
.
If n = 2k + 1 then commutes with everything.
We are only interested in the even case, so assume n = 2k, dene
C
:= i
k
and consider the map
f := i
k
2k
() :
2k

2k
.
From this we see that f commutes with
2k
():
f(
2k
()) = i
k
2k
()(
2k
()) = i
k
2k
() = i
k
2k
()
=
2k
()i
k
2k
() =
2k
()f()
for Cl
0
0,2k
and
2k
. Furthermore f is an involution:
f f = i
2k
2k
(
2
) = (1)
k
2k
((1)
k
) = (1)
2k
id
2k
= id
2k
.
Then it is well-known that f has the eigenvalues 1 and corresponding eigenspaces
2k
(of equal dimension) such that
2k

=
+
2k
2k
. The elements of
2k
are
called positive and negative Weyl spinors or even and odd chiral spinors, respec-
tively.
We can use the map f to induce another splitting of the complexied Clif-
ford algebra (in even dimension!), for left multiplication by
C
on Cl
C
2k
is an
involution, hence the algebra splits into eigenspaces
Cl
C
2k
= (Cl
C
2k
)
+
(Cl
C
2k
)

where, in fact
(Cl
C
2k
)
=
1
2
(1
C
) Cl
C
2k
.
We can combine this splitting with the splitting given by the involution to
obtain the spaces
(Cl
C
2k
)
0
:=
1
2
(1
C
)(Cl
C
2k
)
0
and (Cl
C
2k
)
1
:=
1
2
(1
C
)(Cl
C
2k
)
1
.
For the next proposition we will identify End(
+
2k
) as the subspace of End(
2k
)
consisting of maps
2k

2k
which map
+
2k
to itself and which map
2k
to
0, and in a similar way we identify End(
2k
) and Hom(
2k
,
2k
) as subspaces
of End(
2k
).
Proposition 7.26. The spin representation
2k
: Cl
C
2k
End(
2k
) restricts
to the following isomorphisms
(Cl
C
2k
)
0
+
= End(
+
2k
), (Cl
C
2k
)
0

= End(
2k
)
(Cl
C
2k
)
1
+
= Hom(
2k
,
+
2k
), (Cl
C
2k
)
1

= Hom(
+
2k
,
2k
).
Proof. First assume
+
2k
, i.e. f() = , then for (Cl
C
2k
)
0
+
:
f(
2k
()) =
2k
(
C
) =
2k
(
C
)
=
2k
()f() =
2k
(),
i.e.
2k
()
+
2k
. For this result we only used that ( Cl
C
2k
)
0
, but to show
that
2k
() is 0 on
2k
we need that =
1
2
(1 +
C
). Let
2k
, i.e.
2k
(1 +
C
) = 0, then
2k
() =
1
2
2k
((1 +
C
)) =
1
2
2k
((1 +
C
)) = 0.
This shows that
2k
maps (Cl
C
2k
)
0
+
into End(
+
2k
). The reasoning is the same in
the other 4 cases. But then, since
2k
is an isomorphism, the restricted maps
must be isomorphisms as well, and this proves the proposition.
Chapter 8
Spin Groups
8.1 The Cliord Group
Having introduced the Cliord algebra Cl(), we proceed to dene its Cliord
group (). The point of doing this is that the Cliord group has two particu-
larly interesting subgroups, the pin and spin groups.
Let V be a nite-dimensional real vector space, and let be a quadratic form
on V.
Denition 8.1 (Cliord Group). Let Cl
() denote the multiplicative group

of invertible elements of Cl(). The Cliord group (by some also called the
Lipschitz group) of Cl() is the group
() := x Cl
() [ (x)vx
1
V for all v V .
One mechanically veries that () is truly a group. The group Cl
() is
an open subgroup of Cl(), just as Aut(V ) is an open subgroup of End(V ) (at
least when V is nite-dimensional). In the latter case the Lie algebra of Aut(V )
is just End(V ) with the commutator bracket. In the same way the Lie algebra
cl
() of the group Cl
() is just Cl() with the commutator bracket.

It is very conspicuous from the denition that we are interested in a particular
representation : () Aut(V ), namely () x
x
where
x
: V
V is given by
x
(v) = (x)vx
1
and called the twisted adjoint representation
(indeed, the form of
x
is reminiscent of the adjoint representation of a Lie
group).
One reason for considering
x
(v) = (x)vx
1
instead of Ad
x
(v) = xvx
1
is that the twisted adjoint representation keeps track of an otherwise annoying
sign. W.r.t. the bilinear form we dene, for x V with (x) ,= 0 the reection
s
x
through the hyperplane orthogonal to x by
s
x
(v) := v 2
(v, x)
(x)
x.
We then have the following geometric interpretation of
x
which in addition to
being a pretty fact, is crucial in the proof of Lemma 8.10.
Proposition 8.2. For any x V with (x) ,= 0 we have x (), and the
map
x
: V V given by
x
(v) = (x)vx
1
is the reection through the
hyperplane orthogonal to x.
Proof. First, since x
2
= (x) 1 ,= 0, x is invertible in Cl() with inverse
125
126 Chapter 8 Spin Groups
x
1
=
1
(x)
x V . Using this and Eq. (7.2) we see that
(x)(x)vx
1
= xv((x)x
1
) = xvx
= 2(v, x)x + (x)v V
i.e. we have x () and
x
v = v 2
(v, x)
(x)
x = s
x
(v) V.
The following proposition is extremely useful:
Proposition 8.3. Let be a non-degenerate bilinear form on V. Then for the
twisted adjoint representation ker = 1
1 (where 1
= 1 0).
Proof. Since is non-degenerate, we can choose an orthonormal basis e
1
, . . . , e
n
for V such that (e
i
) = 1 and (e
i
, e
j
) = 0 when i ,= j. Let x ker ; this
means (x)v = vx for all v V. Because of the Z
2
-grading of Cl() we can
write x = x
0
+ x
1
where x
0
and x
1
belong to Cl
0
() and Cl
1
(), respectively.
This gives us the equations
vx
0
= x
0
v (8.1)
vx
1
= x
1
v. (8.2)
The terms x
0
and x
1
can be written as polynomials in e
1
, . . . , e
n
. By successively
applying the identity e
i
e
j
+ e
j
e
i
= 2(e
i
, e
j
) we can express x
0
in the form
x
0
= a
0
+ e
1
a
1
where a
0
and a
1
are both polynomial expressions in e
2
, . . . , e
n
.
Applying to this equality shows that a
0
Cl
0
() and a
1
Cl
1
(). Setting
v = e
1
in Eq. (8.1) we get
e
1
a
0
+e
2
1
a
1
= a
0
e
1
+e
1
a
1
e
1
= e
1
a
0
e
2
1
a
1
where the last equality follows from a
0
Cl
0
() and a
1
Cl
1
(). We deduce
0 = e
2
1
a
1
= (e
1
)a
1
; since (e
1
) ,= 0, we have a
1
= 0. So, the polynomial
expression for x
0
does not contain e
1
. Proceeding inductively, we realize that x
0
does not contain any of the terms e
1
, . . . , e
n
and so must have the form x
0
= t 1
where t 1.
We can apply an analogous argument to x
1
to conclude that neither does
the polynomial expression for x
1
contain any of the terms e
1
, . . . , e
n
. However,
x
1
Cl
1
(), so x
1
= 0.
Thus, x = x
0
+ x
1
= t 1. Since x ,= 0, we must have t 1
. This shows
ker 1
1; the reverse inclusion is obvious.

The assumption that is non-degenerate is not redundant. Consider a real
vector space V with dimV 2. If 0, then Cl(V, ) =
V, the exterior
algebra of V. Consider the element x = 1 + e
1
e
2
. Clearly, x
1
= 1 e
1
e
2
, and
we have
(1 +e
1
e
2
) v (1 +e
1
e
2
)
1
= (1 +e
1
e
2
) v (1 e
1
e
2
) = v,
i.e. 1 + e
1
e
2
ker , yet 1 + e
1
e
2
is not a scalar multiple of 1. Thus, since
the following propositions use Proposition 8.3 in their proofs we will, from this
point on, always assume to be non-degenerate. This is not a severe restriction
since practically all interesting Cliord algebras originate from non-degenerate
bilinear forms.
We now introduce the important notions of conjugation and norm.
8.1 The Cliord Group 127
Denition 8.4. For any x Cl(), the conjugate of x is dened as x := t((x)).
Moreover, the norm of x is dened as N(x) := xx.
Note that t = t (it clearly holds on e
i1
e
ir
, and by linearity on
any element of Cl()). Also note that x = x. The term norm is justied in the
following lemma, displaying some elementary properties of the norm
Lemma 8.5. When is non-degenerate, the norm possesses the following prop-
erties:
1) If v V then N(v) = (v) 1, i.e. N is a negative extension of to the
algebra.
2) If x (), then N(x) 1
1.
3) When restricted to (), the norm N : () 1
1 is a homomorphism.
Moreover, N((x)) = N(x).
Proof. 1) This is just a simple calculation
N(v) = v(t (v)) = v
2
= (v) 1.
2) According to Proposition 8.3, its enough to show that N(x) ker . By
denition of the Cliord group, x () implies
(x)vx
1
V for all v V.
As t[
V
= id
V
, we thus have
(x)vx
1
= t((x)vx
1
) = t(x
1
)vt((x)).
Isolating the v on the right-hand-side we get, using t = t,
v = t(x)(x)v(t((x))x)
1
= (xx)v(xx)
1
,
i.e. xx ker . But then xx = xx ker .
3) Two simple calculations:
N(xy) = xyy x = xN(y)x = xxN(y) = N(x)N(y)
and, since (x) = (x), (following from t = t)
N((x)) = (x)(x) = (xx) = (N(x)) = N(x).
Example 8.6. Lets calculate the conjugate and the norm on the two model
Cliord algebras. First Cl
0,1

= C. Recall that 1 sits inside C as the imaginary
line and that is the involution satisfying (1) = 1 and () = for 1.
Thus is just conjugation: (z) = z. Since t is just the identity on Cl
0,1
it
follows that conjugation in the Cliord sense is just usual conjugation
t((z)) = z.
Therefore the norm becomes N(z) = zz = [z[
2
=
0,1
(z), i.e. the square of
the usual norm and minus the quadratic form.
For Cl
0,2

= Hthe situation is similar. For Cl
0,2
we have the basis 1, e
1
, e
2
, e
1
e
2
where e
1
, e
2
is the usual basis for 1
2
, and 1
2
sits inside Cl
0,2
as spane
1
, e
2
.
We have that (1) = 1, (e
i
) = e
i
and (e
1
e
2
) = (e
1
)(e
2
) = e
1
e
2
. Further-
more t(1) = 1, t(e
i
) = e
i
and t(e
1
e
2
) = t(e
2
)t(e
1
) = e
2
e
1
= e
1
e
2
. Thus
t((q
0
1 +q
1
e
1
+q
2
e
2
+q
3
e
1
e
2
)) = q
0
1 q
1
e
1
q
2
e
2
q
3
e
1
e
2
.
As before, conjugation in the Cliord sense is just usual conjugation q. Further-
more we see that
N(q) = qq = [q[
2
again the square of the usual norm.
For the next proposition recall the denition of the orthogonal group O()
(when is non-degenerate!) as the endomorphisms f : V V satisfying
(f(v)) = (v).
Proposition 8.7. For any x (), the map
x
is an orthogonal transforma-
tion of V. That is, (()) O().
Proof. Let x () and use the fact that N is a homomorphism:
N(
x
v) = N((x)vx
1
) = N((x))N(v)N(x
1
) = N(x)N(v)N(x)
1
= N(v),
Since N(v) = (v) 1, this shows that
x
is -preserving. Along with the
linearity, this exactly shows that
x
O().
8.2 Pin and Spin Groups
Denition 8.8 (Spin Group). The pin group is dened as
Pin() := x () [ N(x) = 1.
The spin group consists of those elements of Pin() that are linear combinations
of even-degree elements:
Spin() := Pin() Cl
0
().
Its not too dicult to verify that Pin() and Spin() really are groups. We
will write Pin(p, q) and Spin(p, q) for the pin and spin groups associated with
Cl
p,q
and likewise Pin(n) := Pin(0, n) and Spin(n) := Spin(0, n) (not to be
confused with the complex pin and spin groups Pin(
n
) and Spin(
n
) sitting
inside Cl
C
n
!). Recall that the algebra structure in Cl
0,n
is that generated by the
relations v v = |v|
2
1.
Since is assumed to be non-degenerate, any real Cliord algebra is (up to
isomorphism) of the form Cl
p,q
and thus we can in fact always assume the real
pin and spin groups to be of the form Pin(p, q) and Spin(p, q).
In addition to the complex pin and spin groups there are also complexica-
tions of the real pin/spin groups, namely let (V, ) be a real quadratic vector
space and dene Pin
c
() Cl()
C
= Cl()C to be the subgroup of invertible
elements in Cl()C generated by Pin()1 and 1U(1). Similarly, Spin
c
()
is dened as the subgroup generated by Spin() 1 and 1 U(1). These are
not to be confused with Spin(
C
)!
Proposition 8.9. There is a Lie group isomorphism
Spin
c
()

Spin()
1
U(1)
i.e. the Spin
c
() is quotient of Spin() U(1) where we identify (1, 1) and
(1, 1). In particular Spin
c
() is connected if Spin() is connected.
Proof. We consider the smooth map Spin() U(1) Spin
c
() given by
(g, z) gz. This is a surjective Lie group homomorphism. The kernel consists
of elements (g, z) such that gz = 1 i.e. g = z
1
U(1). But these elements are
rare, there are only 1. Thus the kernel equals (1, 1), and the map above
descends to an isomorphism.
The group Spin
c
() is important in the study of complex manifolds.
We have the following important lemma, to be used in the proof of Theorem
8.16.
8.2 Pin and Spin Groups 129
Lemma 8.10. The map : Pin() O() is a surjective homomorphism
with kernel 1, 1. Similarly, the map : Spin() SO() is a surjective
homomorphism with kernel 1, 1.
Proof. First, Proposition 8.7 guarantees (Pin()) O(). It is trivial to
verify that is a homomorphism. To prove that is surjective we use the
Cartan-Dieudonne Theorem
1
which states that any element T of O() can be
written as the composition T = s
1
s
p
of reections where s
j
is the reection
through the hyperplane orthogonal to some vector u
j
V But according to
Proposition 8.2 s
j
=
uj
. Since N(u
j
) = (u
j
) 1, replacing u
j
by
uj
](uj)]
(this doesnt change the reection) leaves N(u
j
) = 1 and thus N(u
1
u
p
) =
1, i.e. u
1
u
p
Pin(). Hence T =
u1

up
=
u1up
which proves that
[
Pin()
is surjective. The kernel is easily calculated:
ker [
Pin()
= ker Pin() = t 1
1 [ N(t) = 1 = 1, 1.
To prove the analogous statement for Spin() we need rst to show that maps
Spin() to SO(). Assume, for contradiction, that this is not the case, i.e. that
an element f O() SO() exists such that
x
= f for some x Spin(). By
the Cartan-Dieudonne Theorem f can be written as an odd number of reections
f = s
1
s
2k+1
, and to each such reection s
j
corresponds a vector u
j
so
that s
j
=
uj
. In other words we have
x
=
u1u
2k+1
or id =
x
1
u1u
2k+1
.
By Proposition 8.3 x
1
u
1
u
2k+1
= 1 for some 1
, i.e. x =
1
u
1
u
2k+1
and this is a contradiction, since u
1
u
2k+1
Cl
1
(). Thus maps Spin()
to SO() and since 1 Spin() the kernel is still Z
2
.
As a consequence of the proof of this lemma we note, that the set v
V [ (v) = 1 of unit vectors generate Pin() as a subgroup of (), i.e. any
element of Pin() can be written as a product of unit vectors in V . Similarly
elements of Spin() can be written as a product of an even number of unit
vectors in V .
Lets warm up by doing a few simple examples
Example 8.11. Lets calculate Pin(1) and Spin(1). They are subgroups of
Cl
0,1

= C and the vector space, from which the Cliord algebra originates, is
1 which sits inside C as the imaginary line (cf. Example 7.5). In 1 we have
just two unit vectors, namely 1. They sit in C as i and they generate Pin(1)
which are thus seen to be isomorphic to Z
4
(the fourth roots of unity). Spin(1)
is then generated by products of two such unit vectors, i.e. Spin(1) = Z
2
.
Now, lets calculate Pin(2) and Spin(2). Recall from Example 7.5 that Cl
0,2

=
H and that the vector space 1
2
sits inside H as the i, j-coordinates. Thus Pin(2)
is generated by elements in H of the form ai +bj with a
2
+b
2
= 1. We see that
(ai +bj)(a
t
i +b
t
j) = (aa
t
+bb
t
)1 + (ab
t
a
t
b)k,
and one can check that (aa
t
+bb
t
)
2
+ (ab
t
a
t
b)
2
= 1. Thus Pin(2) is the set
Pin(2) = ai +bj [ a
2
+b
2
= 1 c1 +dk [ c
2
+d
2
= 1,
and the product is the one induced from the product on the quaternions. Spin(2)
is the subgroup of this consisting of a product of an even number of unit vectors,
i.e. specically
Spin(2) = c1 +dk [ c
2
+d
2
= 1.
One can check that the product inherited from H is the same as the one on the
circle group so that Spin(2) = U(1).
1
See for instance [Gallier] ([5]), Theorem 7.2.1.
At this point, it is probably not very clear why there should be anything
particularly interesting about pin and spin groups. The explanation is that they
are double coverings of O() and SO() (Theorem 8.16). When n 3 then
Spin(n) is even the universal double covering of SO(n) (Corollary 8.20). The
following more tedious example can be thought of as a special case of this fact:
it shows that Spin(3) is isomorphic to SU(2) which is known to be the universal
double covering of SO(3).
Example 8.12. Calculation of Spin(3). Choose an orthonormal basis e
1
, e
2
, e
3
of 1
3
. By Proposition 7.3, the list of elements
1, e
1
, e
2
, e
3
, e
1
e
2
, e
1
e
3
, e
2
e
3
, e
1
e
2
e
3
forms a basis for the Cliord algebra Cl
0,3
. The group Spin(3) can be written
Spin(3) = x Cl
0
0,3
[ v V : (x)vx
1
V and N(x) = 1,
as N(x) = 1 implies x Cl
0,3
(explicitly, x
1
= x). The rst thing we want
to show is that x Spin(3) if and only if x Cl
0
0,3
and N(x) = 1. The only
non-trivial statement to be proven is that the conditions x Cl
0
0,3
and N(x) = 1
imply that (x)vx
1
V for any v V.
Since x Cl
0
0,3
and v Cl
1
0,3
, we have xvx
1
Cl
1
0,3
. Thereby,
xvx
1
= u +e
1
e
2
e
3
with u V and 1. Moreover, observe that v = v for all v V, and
x = x
1
since N(x) = 1, so
xvx
1
= x
1
v x = xvx
1
.
But e
1
e
2
e
3
= e
3
e
2
e
1
= e
1
e
2
e
3
, so
u +e
1
e
2
e
3
= xvx
1
= xvx
1
= u e
1
e
2
e
3
;
hence = 0, and so xvx
1
V. This is equivalent to (x)vx
1
V, as desired.
Knowing that x Spin(3) i x Cl
0
0,3
and N(x) = 1, we can characterize
the elements x of Spin(3) in a very handy way. Namely, since x Cl
0
3,0
, x must
have the form
x = a1 +be
1
e
2
+ce
1
e
3
+de
2
e
3
where a
2
+b
2
+c
2
+d
2
= N(x) = 1. Thus, Spin(3) consists of all elements x of
the form x = a1 +be
1
e
2
+ce
1
e
3
+de
2
e
3
with a
2
+b
2
+c
2
+d
2
= 1. This allows
us to establish an isomorphism Spin(3)
= SU(2) like follows:

e
1
e
2

_
i 0
0 i
_
, e
1
e
3

_
0 1
1 0
_
, e
2
e
3

_
0 i
i 0
_
,
i.e. x
_
a +ib c id
c id a ib
_
. Thus, Spin(3) is isomorphic to SU(2). One
can use similar arguments to prove Spin(4)
= SU(2) SU(2) and Spin(3, 1)

0

=
SL(2, C) (cf. [Lawson and Michelson], p. 56, Theorem 8.4).
Example 8.13. Lets calculate some of the spin
c
groups also. This is rather
easy, given Proposition 8.9 and the calculations above.
First we show that Spin
c
(3)
= U(2). Dene the map

Spin(3) U(1)
= SU(2) U(1) U(2)

8.3 Double Coverings 131
by (A, z) zA. It is well-known that this map is a surjective Lie group
homomorphism, and it is easily seen that the kernel is (I
2
, 1) (I
2
is just the 22
identity matrix). Thus a Lie group isomorphism is induced on the quotient, i.e.
an isomorphism Spin
c
(3)

U(2).
We can argue almost similarly to calculate Spin
c
(4). For brevity, put
S(U(2) U(2)) :=
_
_
A
1
0
0 A
2
_
[ A
1
, A
2
U(2), det A
1
det A
2
= 1
_
Dene a map
Spin(4) U(1)
= SU(2) SU(2) U(1) S(U(2) U(2))

(A
1
, A
2
, z)
_
zA
1
0
0 zA
2
_
.
This map is again a surjective Lie group homomorphism, as above, and the
kernel is (I
2
, I
2
, 1), and thus it induces an isomorphism Spin
c
(4)

S(U(2)
U(2)).
Note that Pin() and Spin() are Lie groups. This is because the multiplica-
tive group Cl
() of invertible elements is an open subset of Cl() (this is a

general result for algebras) which is a nite-dimensional linear space, hence a
manifold. Thus, Cl
() is a manifold, and since multiplication and inversion are

smooth maps, it is a Lie group. As Pin() is a closed subgroup of Cl
() (since
N is continuous) and Spin() is a closed subgroup of Pin() (since Cl
0
() is a
closed subspace of Cl()), they are Lie groups.
8.3 Double Coverings
In this section we will prove that Pin() and Spin() are double coverings of
O() and SO(), respectively. This will allow us to prove furthermore that
Spin(n) is the universal double covering of SO(n) which is our main result. We
rst recall the notion of a covering space in the general setting of two topological
spaces:
Denition 8.14 (Covering Map). Let Y and X be topological spaces. A
covering map is a continuous and surjective map p : Y X with the property
that for any x X there is an open neighborhood U of x so that p
1
(U) can
be written as the disjoint union of open subsets V
(called the sheets) with the

restriction p[
V
: V
U a homeomorphism. We say that U is evenly covered

by p and call Y a covering space of X. When all the bers p
1
(x) have the
same nite cardinality n, we call p an n-covering. If Y is simply connected, the
covering is called a universal covering.
When X is pathwise connected, all bers p
1
(x) will have the same cardinal-
ity. If the covering is the universal covering this is precisely the cardinality of
the fundamental group
1
(X).
Quite often, covering maps between groups arise from group actions GY
Y. We now introduce the notion of an even action or covering action as it
allows an elegant proof of Theorem 8.16. A group G is said to act evenly on
the topological space Y if each point y Y has a neighborhood U such that
g U h U = if g ,= h. As usual, Y/G denotes the orbit space under the action
of G, equipped with the quotient topology.
Lemma 8.15. Let G be a nite group acting evenly on a topological space Y .
Then the canonical map p : Y Y/G is a [G[-covering.
Proof. p is obviously continuous and surjective. Let [y] = p(y) Y/G for
some y Y . We shall produce a neighborhood of [y] which is evenly covered
by p. As G is acting evenly there exists a neighborhood V Y of y such that
g V h V = if g ,= h; dene U := p(V ). U is open, for we have
p
1
(U) =
_
gG
g V,
where g V are all open (as the map Y x g x is a homeomorphism
for all g G). Consequently, p
1
(U) is open and by denition of the quotient
topology, U is open. Thus, we have that U is a neighborhood of [y] and that
p
1
(U) is a disjoint union of sets homeomorphic to V , and that the number of
sheets is equal to the order of the group G. The only thing we need to show is
that p[
V
: V U is a homeomorphism. The map is obviously surjective. To
show injectivity assume that p(x) = p(y) (for x, y V ) that is there exists a
g G such that y = g x. But then y g V . From the fact that V g V =
if g ,= e, we deduce g = e and consequently x = y.
Continuity of p[
V
is obvious, since its a restriction of a continuous map. We
now only need to show that p[
V
is a open map. But p is itself an open map, for
let O Y be any open subset. By denition of the quotient topology, p(O) is
open if and only if p
1
(p(O)) is open. But
p
1
(p(O)) =
_
gG
g O
is open, being the union of open sets gO. Being a restriction of an open map p to
an open set V , p[
V
is an open map, and p[
V
is therefore a homeomorphism.
Theorem 8.16. The map : Pin() O() is a double covering. Moreover,
: Spin() SO() is a double covering.
Proof. By Lemma 8.10, the homomorphism : Pin() O() is sur-
jective and has the kernel 1, 1 = Z
2
. By standard results we have that
O() and Pin()/Z
2
are isomorphic as Lie groups. We let Z
2
act on Pin()
by multiplication which is obviously an even action. By the preceding Lemma
8.15, the quotient map Pin() Pin()/Z
2
is a double covering. Since
: Pin() O() can be identied with this map via the above isomorphism
of Lie groups, is a double covering. The proof for : Spin() SO() is
completely analogous.
Corollary 8.17. The groups Pin(n) and Spin(n) are compact groups.
Proof. Since O(n) and SO(n) are compact and Pin(n) and Spin(n) are nite
coverings, the result follows from standard covering space theory.
This does not hold in general, for instance Spin(3, 1)
0
= SL(2, C) which is
denitely not compact. If the identity component in non-compact, the entire
group must be non-compact.
In Chapter 7 we constructed explicit isomorphisms : O(p, q)

O(q, p)
and : SO(p, q)

SO(q, p). A natural question would be if these isomor-
phisms lift to isomorphisms on the level of pin and spin groups? For the pin
groups the answer is no, in general. But for the spin groups the answer is ar-
mative:
8.3 Double Coverings 133
Proposition 8.18. There exists a Lie group isomorphism : Spin(p, q)

Spin(q, p) such that the following diagram commutes

Spin(p, q)

Spin(q, p)
SO(p, q)

SO(q, p)
Proof. Let be an anti-orthogonal linear map as mentioned in Chapter 7. By
the same sort of argument as in the proof of Proposition 7.4, it is seen that
extends to a map on the corresponding Cliord algebras. This is also denoted
. Dene : Spin(p, q) Spin(q, p) by (recall that Spin(p, q) is generated by
an even number of unit vectors)
(v
1
v
2k
) = (1)
k
(v
1
) (v
2k
)
where v
1
, . . . , v
2k
1
p+q
. This is easily seen to be a Lie group isomorphism.
Thus we only need to check commutativity of the diagram. First we observe
(
vi
)v =
vi

1
(v) = (v
i
1
(v)v
1
i
)
= (v
i
)v(v
1
i
) = (v
i
)v(v
i
)
1
=
(vi)
v.
Then we immediately get
((v
1
v
2k
)) = ((1)
k
(v
1
) (v
2k
))
=
(v1)

(v
2k
)
= (
v1
) (
v
2k
)
= (
v1

v
2k
) = ((v
1
v
2k
))
Restricting attention to Spin(n) we can prove another delicious fact concern-
ing its topology:
Theorem 8.19. Spin(n) is path-connected when n 2 and simply connected
when n 3.
Proof. We rst remark that Spin(n) is pathwise connected. Consider the el-
ement v
1
v
2k
where v
j
1
n
and
0,n
(v
j
) = 1 and note that each v
j
can
be connected to v
1
by a continuous path running on the unit sphere S
n1
=
v 1
n
[
0,n
(v) = 1 in 1
n
. Thus there is a continuous path from v
1
v
2k
to
v
1
v
1
= 1 and thereby Spin(n) is path-connected.
Next we need to show, that the fundamental group at each point of Spin(n)
is the trivial group. Since Spin(n) is path-connected it suces to show this for
just a single point. So let x
0
Spin(n). By standard covering space theory
1
(Spin(n), x
0
)
1
(SO(n), (x
0
)) is injective, and the index of the subgroup
(
1
(Spin(n), x
0
)) in
1
(SO(n), (x
0
)) is equal to the number of sheets of the
covering , which we have just showed was 2. For n 3 the fundamental group
1
(SO(n), (x
0
)) is Z
2
, so as
:
1
(Spin(n), x
0
) has index 2, it must be the
trivial subgroup. Since
was injective also

1
(Spin(n), x
0
) is trivial, proving
that Spin(n) is simply connected.
Putting p = 0 in Theorem 8.16 and combining with Theorem 8.19 we get the
main result:
Corollary 8.20. For n 3, the group Spin(n) is the universal (double) cover-
ing of SO(n).
Its a classical fact from dierential geometry that for connected Lie groups G
and H, if F : G H is a covering map, then the induced map F
: g h is
an isomorphism.
2
Spin(n) is connected and simply connected (Theorem 8.19),
2
See for instance [Warner], Proposition 3.26.
and SO(n) is connected. By Theorem 8.16, the homomorphism : Spin(n)
SO(n) is a covering map, and thus we have an isomorphism
: spin(n)

so(n),
in particular, dimspin(n) = dimso(n) =
n(n1)
2
.
Lets investigate this map a little further. Recall that the Lie algebra of Cl
0,n
is just the Cliord algebra Cl
0,n
itself with the commutator bracket. Spin(n) is
a Lie subgroup of Cl
0,n
and hence the Lie algebra spin(n) is a Lie subalgebra of
Cl
0,n
.
Proposition 8.21. Let e
1
, . . . , e
n
be an orthonormal basis for 1
n
,
then spin(n) Cl
0,n
is spanned by elements of the form e
i
e
j
where 1 i <
j n. Furthermore
maps e
i
e
j
to the matrix 2B
ij
so(n) where B
ij
is the
n n-matrix which is 1 in its ijth entry and 1 in its jith entry.
Proof. Consider the curve
t (e
i
cos t +e
j
sin t)(e
i
cos t +e
j
sin t) = cos(2t) + sin(2t)e
i
e
j
.
It is a curve in Spin(n) since it is the product of two unit vectors, and its value at
t = 0 is the neutral element 1. Upon dierentiating at t = 0 we get 2e
i
e
j
, which
is then an element of T
1
Spin(n)

= spin(n). They are all linearly independent i
Cl
0,n
hence also in spin(n), and there are exactly
n(n1)
2
of them, i.e. they span
spin(n).
Now, is the restriction of the twisted adjoint representation to Spin(n), and
since Spin(n) Cl
0
0,n
we get (g)v = gvg
1
for g Spin(n) and v 1
n
. As for
the usual adjoint representation one can calculate
(
X)v = Xv vX (8.3)
in particular we get
(e
i
e
j
)e
k
= e
i
e
j
e
k
e
k
e
i
e
j
=
_
_
0, k ,= i, j
2e
j
, k = i
2e
i
, k = j
We see that
(e
i
e
j
) acts in the same way on 1
n
as the matrix 2B
ij
, thus we
may identify
(e
i
e
j
) = 2B
ij
.
We can rephrase the rst part of this proposition by saying that under the
symbol map : Cl
0,n
1
n
, the Lie algebra spin(n) gets mapped to
2
1
n
.
We end the section by a short description of some covering properties of
Spin
c
().
Proposition 8.22. The map
c
: Spin
c
() SO()U(1) given by [g, z]
((g), z
2
) is a double covering.
Proof. It is easy to see that it is well-dened (since (g) = (g) and (z)
2
=
z
2
). It is a covering map because it is the quotient map of the even Z
2
-action on
Spin
c
() given by (1, [g, z]) [g, z] = [g, z]. Thus it follows from Lemma
8.15.
Since SO(n) U(1) is compact, Spin
c
(n) is also compact. Furthermore, for
n 2 Spin
c
(n) is connected (cf. Proposition 8.9 and connectivity of Spin(n)).
Thus
1
(Spin
c
(n)) can be identied with a subgroup of
1
(SO(n) U(1)) =
Z
2
Z of index 2, i.e.
1
(Spin
c
(n))
= Z.
8.4 Spin Group Representations 135
8.4 Spin Group Representations
In this section we will treat the basics of the representation theory of spin
groups. In this section we will restrict our attention to a particular complex
representation of Spin(n), the spinor representation, dened as follows:
Denition 8.23 (Complex Spinor Representation). By the complex spinor
representation
n
of Spin(n) we understand the restriction
n
[
Spin(n)
Aut
C
(
n
)
of the complex spin representation
n
of Cl
C
n
to Spin(n).
Similarly, we dene the spin
c
-representation
c
n
of Spin
c
(n) by restricting the
spin representation to Spin
c
(n) Cl
C
n
.
We stress that we use the term spin representation for the irreducible Cliord
algebra representations and spinor representation for the associated spin group
representations.
Of course we can in a similar way dene the real spinor representation of
Spin(n) by restricting
n
to Spin(n), but we will not consider them here.
Theorem 8.24. For each n the complex spinor representation
n
is a faithful
representation of Spin(n).
Proof. If n = 2k is even, then
n
is a restriction of the isomorphism
n
:
Cl
C
n
End
C
(
2k
) and therefore injective.
So lets assume that n = 2k + 1. By denition we have
2k
=
2k+1
and
consequently Aut(
2k
) = Aut(
2k+1
). We can think of Spin(2k) as sitting
inside Spin(2k +1).
3
Denoting the injection the following diagram commutes:
Spin(2k)

2k
Aut(
2k
)
id
Spin(2k + 1)
2k+1
Aut(
2k+1
)
Now, put H := ker
2k+1
Spin(2k + 1). The goal is to verify H = 1, but
rst we show that H Spin(2k) = 1. The inclusion follows since 1 clearly
sits in ker
2k+1
. Now, assume that h H Spin(2k). In particular, h H and
h = (
h) for some

h Spin(2k). Since h sits in H,
2k+1
(h) = id
2k+1
. From
the commutativity of the diagram it follows that
2k
(
h) = id
2k+1
. But since
2k
is injective,
h must be 1, and so must h. This shows H Spin(2k) 1.

Identifying elements A SO(n) with elements in SO(2k + 1) of the form
diag(A, 1), we obtain SO(2k) SO(2k + 1) like the spin groups. Recall that
: Spin(2k +1) SO(2k +1) is a surjective homomorphism. Thus (H) is a
normal subgroup of SO(2k + 1), since H as a kernel is normal in Spin(2k + 1).
Now we claim
(H Spin(2k)) = (H) SO(2k).
The inclusion is obvious, and follows from the surjectivity of . Hence
we have (H) SO(2k) = I (here, I denotes the identity matrix). We want
to show that (H) = I, so let A (H) SO(2k + 1). Its characteristic
polynomial is of odd degree, and it thus has a real root. As A SO(2k + 1),
all eigenvalues have modulus 1. Moreover, A preserves orientation, so this root
must be 1. Denote the corresponding eigenvector by v
0
and choose an ordered,
positively oriented orthonormal basis for 1
2k+1
containing v
0
as the last vector.
3
If Cl
0,2k
is generated by {e
1
, . . . , e
2k
} and Cl
0,2k+1
is generated by {e
1
, . . . , e
2k+1
} then
we have a linear injection : Cl
0,2k
Cl
0,2k+1
by dening (e
j
) = e
j
. This restricts to an
injection : Spin(2k) Spin(2k + 1).
If B denotes the change-of-basis matrix, then we have the block diagonal matrix
BAB
1
= diag(
A, 1), where

A SO(2k) and 1 is the unit of 1. We can now
identify

A with BAB
1
. Hence, BAB
1
SO(2k), and since (H) was normal,
we also have BAB
1
(H). All together we have BAB
1
(H)SO(2k) =
I and so A = I.
Now we have (H) = I. We have two possibilities: H = 1 or H = 1.
But 1 cannot be in the kernel of the spinor representation (because its not in
the kernel of the spin representation, from which it came). Therefore, H = 1,
and
2k+1
is injective.
This theorem is not as innocent as it might look. It actually tells us that the
spinor representations do not arise as lifts of SO(n)-representations, since a lift
of an SO(n)-representation necessarily contains 1 in its kernel.
We now want to decompose the spinor representations into irreducible repre-
sentations. To this end we need:
Lemma 8.25. For any complex vector space V the endomorphism algebra
End(V ) is a simple algebra, i.e. the only ideals are the trivial ones. In par-
ticular if dimW < dimV , then any homomorphism : End(V ) End(W) is
trivial : 0.
Proof. Let n be the dimension of V and x a basis for V . Then we can think
of End(V ) as the algebra of complex n n-matrices. Now let J End(V )
be any non-zero ideal, and let 0 ,= a J. Then a has an eigenvalue ,= 0
(because C is algebraically closed). By a suitable basis transformation, given by
a change-of-basis-matrix b, and subsequently multiplying diag(1/, 0, . . . , 0) on
b
1
ab from the left we obtain the matrix diag(1, 0 . . . , 0). It has been constructed
from a just by multiplication, so its in J. By a similar argument we obtain
diag(0, . . . , 0, 1, 0 . . . , 0) J. The sum of all these is just the identity matrix,
which therefore is also in J. Thus, J = End(V ).
If : End(V ) End(W) is a homomorphism, ker is an ideal in End(V ),
thus ker = 0 or ker = End(V ). But since dimW < dimV injectivity of
is impossible. Therefore ker = End(V ) and 0.
Decomposing
2k+1
into irreducibles is easy:
Theorem 8.26. The spinor representation
2k+1
of Spin(2k+1) is irreducible.
Proof. Lets assume that 0 , = W _
2k+1
is a Spin(2k + 1)-invariant sub-
space, i.e. for each g Spin(2k + 1)
2k+1
(g)W W. Consider an element of the form e
i
e
j
, i < j (e
1
, . . . , e
2k+1
is an orthonormal basis for the vector space 1

2k+1
underlying Cl
0,2k+1
). It is an
element of Spin(2k +1) and therefore
2k+1
(e
i
e
j
)W W. I.e.
2k+1
is actually
dened on all elements e
i1
e
im
where i
1
< < i
m
and m is even. On the
other hand these elements constitute a complex basis for (Cl
C
2k+1
)
0
. Hence we
get an algebra representation : (Cl
C
2k+1
)
0
End(W) by extending
2k+1
linearly. But recall that (Cl
C
2k+1
)
0
= Cl
C
2k
= End(
2k
) (Proposition 7.16) so
that we get an algebra homomorphism
: End(
2k
) End(W).
W was a proper subspace of
2k+1
=
2k
, so dimW < dim
2k
. Lemma 8.25
now guarantees that 0. Since is an extension of
2k+1
, this should also be
zero. As
2k+1
is injective by Theorem 8.24 this is a contradiction, so W cannot
be invariant.
8.4 Spin Group Representations 137
Example 8.27. Again we consider our favorite spin group Spin(3). What is
the spinor representation of Spin(3)? Recall that Spin(3)

= SU(2) and that for
each n SU(2) has exactly one irreducible representation
n
of dimension n + 1
on the space of homogenous n-degree polynomials in two variables. The spinor
representation
3
is a 2-dimensional irreducible representation, thus it must be
equivalent to
1
.
2k
is not an irreducible representation, but it can quite easily be decom-
posed into such. To do this recall the volume element, the unique element in
Cl
0,2k
given by e
1
e
2k
. It commutes with everything in the even part of Cl
0,2k
and anti-commutes with the odd part. The map f = i
k
2k
(), which was an
involution, gave rise to a splitting
2k

=
+
2k

2k
.
Lemma 8.28.
+
2k
and
2k
are
2k
-invariant subspaces. Thus,
2k
induces
representations
2k
on
2k
such that
2k
=
+
2k

2k
.
Proof. We want to show
2k
(g)
2k

2k
for any g Spin(2k), so let be a
positive Weyl spinor. Then
f(
2k
(g)) =
2k
(g)f() =
2k
(g)
so
2k
(g)
+
2k
. Likewise if is a negative Weyl spinor.
Theorem 8.29.
2k
are irreducible representations of Spin(2k).
Proof. Like in the proof of Theorem 8.26 a
+
2k
-invariant subspace 0 , = W _
+
2k
gives rise to a representation
: (Cl
C
2k
)
0
End(W).
Again, by Proposition 7.16 and Corollary 7.20:
(Cl
C
2k
)
0

= Cl
C
2k1
= End(
2k1
) End(
2k1
).
We get homomorphisms
1
,
2
: End(
2k1
) End(W) simply by
1
(x) = (x, 1) and
2
(y) = (1, y).
By assumption dimW < dim
+
2k
= dim
2k1
and so by Lemma 8.25
1
,
2

0. This means (x, y) = ((x, 1)(1, y)) =
1
(x)
2
(y) = 0, hence 0 and thus
+
2k
0 which is a contradiction.
The covering space results in the previous section yields the following result
Corollary 8.30. Let : Spin(n) Aut(V ) be a nite-dimensional repre-
sentation of Spin(n) which is the restriction of an algebra representation :
Cl
0,n
End(V ) (e.g. the spinor representation
n
). Then for the induced
representation
: spin(n) End(V ) it holds that
(X)v = (X)v
for X spin(n) Cl
0,n
and v V .
Proof. Note that is a linear map, hence the induced representation of the
restriction [
Cl
0,n
is just itself (where we have identied cl
0,n
= Cl
0,n
). The
induced representation of = [
Spin(n)
is just the restriction of [
Cl
0,n
, hence
the formula follows.
A close examination of the proofs above will reveal that nothing is used which
does not hold for Spin
c
(n) as well. We may therefore summarize the results above
in the following statement about the spin
c
-representations:
Theorem 8.31. For each n the spin
c
-representation
c
n
is faithful. If n is odd,
the representation is irreducible and if n is even it splits in a direct sum of two
irreducible representations
c
n
= (
c
n
)
+
(
c
n
)
where (
c
n
)
are representations
on the space
n
.
Chapter 9
Topological K-Theory
9.1 The K-Functors
In this chapter we will describe the topological K-theory of topological spaces,
a theory developed in the 1950s, rst by Alexander Grothendieck in his attempt
to generalize the Riemann-Roch Theorem, later by Michael Atiyah, Friedrich
Hirzebruch and Raoul Bott, culminating with the proof of the Bott periodicity
in the late 1950s. The K stems from the German word Klassen, i.e. class.
Later it has, with great success, been vastly generalized to a non-commutative
version, a K-theory for C
-algebras.
In this chapter we shall deal with the K-theory for topological spaces. A
short outline of the present chapter is as follows: we begin by dening K-theory
for compact spaces and describing some elementary properties. In the section
to follow we discuss some cohomological aspects of K-theory such as relative
K-theory and the long-exact sequence, and then we state, without proof, Bott
periodicity. After this, we more or less start all over again by generalizing to
locally compact spaces carrying a G-action and nally we state the Thom Iso-
morphism Theorem and derive from this the Bott periodicity.
Since K-theory is dened in terms of vector bundles, we begin with a short
review, without proof, of some of the most important results in the theory of
vector bundles. Proofs of these results can be found in [Ha] Chapter 1.
Denition 9.1 (Vector Bundle). Let X be a topological space. A complex
vector bundle over X is a topological space E and a continuous projection map
: E X, such that for each point x X there exists a neighborhood U X
around x, a natural number k and a homeomorphism
:
1
(U) U C
k
such that
1
= . The pair (U, ) is called a trivialization of E and a cover
(U
i
,
i
) of trivializations is called a trivialization cover. The trivial bundles
X C
k
will be denoted by the short-hand notation I
k
.
Note, that we do not require all the bers to have the same dimension. On
components the dimension of the bers are the same, since the ber dimension,
by local triviality, is a continuous map X N
0
, but on dierent components
the dimensions need not be equal.
The rst proposition, which is really one of the founding pillars of K-theory,
explains why we here at the beginning restrict our attention to compact topo-
logical spaces.
139
140 Chapter 9 Topological K-Theory
Proposition 9.2. Let X be compact Hausdor and E a vector bundle over X.
Then there exists a vector bundle E
t
over X such that E E
t
is a trivial vector
bundle.
In the next chapter it will be of great importance that this result holds if X
is a smooth manifold, compact or not
1
.
If : E X is a vector bundle and f : Y X is a continuous map, we
can form a vector bundle f
E, called the pullback bundle, dened by

f
E = (x, v) Y E [ f(x) = (v)

with projection map : (x, v) x. This bundle is the unique bundle over Y
making the diagram
f
E
F
Y
f
X
commutative, where F is the map (x, v) v. The pullback construction sat-
ises the following properties
(f g)
E = g
(f
E)
f
(E F) = f
E f
F
f
(E F) = f
E f
F.
Proposition 9.3. Let E be a vector bundle over X [0, 1] where X is a para-
compact space, then the restrictions to X 0 and X 1 are isomorphic.
From this one can deduce the following homotopy-invariance of pullback bun-
dles:
Corollary 9.4. Let X be a paracompact space, and E a vector bundle over X.
If two maps f
0
, f
1
: Y X are homotopic, then f
0
E and f
1
E are isomorphic.
Let Vect
C
(X) denote the set of isomorphism classes of complex vector bundles
over X. The isomorphism class containing E will be denoted [E].
With the operation + dened as
[E] + [F] = [E F]
(one readily checks that this is well-dened) Vect
C
(X) becomes an abelian semi-
group (recall that E F

= F E). Furthermore this semigroup has the cancel-
lation property, i.e. if
[E] + [H] = [F] + [H]
then [E] = [F]. This can be seen in the following way: Let H
t
be a bundle
complementary to H, i.e. such that H H
t
= I
k
. Then by adding [H
t
] on both
sides we get that E I
k

= F I
k
, and this of course implies that E

= F.
Denition 9.5. Let X be compact, then we dene the (complex) K-group of
X to be the Grothendieck group
2
K(X) of the semigroup Vect
C
(X).
To be a little more specic, consider Vect
C
(X) Vect
C
(X) and dene an
equivalence relation on this product by ([E], [F]) ([G], [H]) if and only if
E H

= F G. Then dene K(X) = (Vect
C
(X) Vect
C
(X))/ .
1
Cf. [MT] Exercise 15.10.
2
A good account of the Grothendieck construction can be found in N. J. Laustsen, F.
Larsen and M. Rrdam: An Introduction to K-Theory for C*-Algebras.
9.1 The K-Functors 141
It is a standard result that any element in the Grothendieck group of a semi-
group with the cancellation property, that the semigroup sits inside the group,
and that any group element is just a formal dierence of two elements in the
semigroup. Thus any element of the K-group is of the form
= [E] [F],
although this representation of the element need not be unique. In fact we have
Lemma 9.6. Let [E] [F] be an element of K(X), then there exist H and I
k
such that [E] [F] = [H] [I
k
].
Proof. To see this use Proposition 9.2 to get a vector bundle F
t
such that
F F
t
= I
k
. Put H = E F
t
, then
[E] [F] = ([E] + [F
t
]) ([F] + [F
t
]) = [H] [I
k
].
Addition in this group is, of course given by
([E
1
] [F
1
]) + ([E
2
] [F
2
]) = [E
1
E
2
] [F
1
F
2
],
the neutral element is [I
0
], the isomorphism class of the trivial zero-dimensional
vector bundle over X, and this equals the formal dierence [E] [E] for any
vector bundle E. The inverse of [E] [F] is then seen to be [F] [E].
As a matter of fact, K(X) is not just a mere group, it is a commutative ring.
It should come as no surprise that the product in K(X) originates from tensor
products of vector bundles: we dene
([E] [F])([G] [H]) := ([E G] + [F H]) ([E H] + [F G]).
This is, as one can check, well-dened, and since E F

= F E this product
turns K(X) into a commutative ring with unit element [I
1
]. Although K(X) is
a ring, we will refer to it as the K-group.
If f : X Y is a continuous map, then we get a map f
: K(Y ) K(X)
simply by pulling back vector bundles
f
([E] [F]) = [f
E] [f
F].
This is well-dened and satises (f g)
= g
. Since f
(E F)
= f
(E)
f
(F) and f
(E F)

= f
(E) f
(F), it is a ring homomorphism, hence K

is a contravariant functor from the category of compact Hausdor spaces to
the category of commutative unital rings. Observe, that if f : X Y and
g : X Y are homotopic, then by Corollary 9.4 f
= g
. Thus we have shown

Proposition 9.7. Homotopic maps induce identical maps in K-theory. In par-
ticular, homotopy equivalences induce isomorphisms in K-theory.
Observe that all we have done so far could equally well have been done for real
vector bundles or, for that matter, quaternionic vector bundles. The correspond-
ing real K-group is denoted KO(X) and the quaternionic K-group KSp(X) (in
fact, complex K-theory is sometimes denoted KU(X) but as we will mostly
be interested in complex K-theory, this notation is unnecessarily cumbersome).
The dierence will become visible when discussing Bott-periodicity, where in
fact complex K-theory is substantially simpler than real K-theory.
Example 9.8. Lets calculate the K-group of a one-point space pt. As the
only vector bundles over pt are product bundles, and since these can only be
distinguished by their dimension we have a natural identication Vect
C
(pt)

=
N
0
, the identication being I
k
k. This is an isomorphism of semirings. If we
Grothendieck this semiring we get Z, hence K(pt)
= Z where the identication

is [I
m
] [I
n
] mn. If X is a contractible space, then all vector bundles over
X are trivial, thus Vect
C
(X)

= N
0
. Hence K(X)

= Z and the identication is
again [I
m
] [I
n
] mn. By exactly the same arguments we get KO(pt) =
KSp(pt) = Z.
With this example in mind we may proceed to dene the so-called reduced
K-theory. To this end let (X, x
0
) be a based compact Hausdor space, i.e. a
topological space with a distinct point x
0
, called the base point. The class of
such spaces forms a category with morphisms being continuous maps X Y
mapping base point to base point.
Consider the inclusion i : x
0
X. This gives rise to a ring homomorphism
i
: K(X) K(x
0
). If we compose i
with the canonical isomorphism

K(x
0
)

Z then it maps [E] [I
k
] to the integer dimE
x0
k, or, since ber
dimensions are constant on components, it maps [E] [I
k
] to the dimension of
E of the component containing x
0
minus k.
Denition 9.9. Let X be a compact based space and dene the reduced K-
group of X by
K(X) := ker i
.
Since

K(X) is an ideal in K(X), it is itself a commutative ring, albeit not
necessarily a unital one.
Similarly, the map i induces homomorphisms on real and quaternionic K-
theory, i
O
: KO(X) Z and i
Sp
: KSp(X) Z and we dene
KO(X) = ker i
O
and

KSp(X) = ker i
Sp
.
Obviously, these denitions are independent of the choice of one-point set. Later,
we will discuss how to get rid of the base point dependence. For now, lets verify
that

K is a functor on the category of based spaces: assume f : X Y to
be a continuous map between based spaces preserving base points. It induces a
map in unreduced K-theory: f
: K(Y ) K(X) and if

K(Y ), then
i
X
(f
()) = (f i
X
)
() = i
Y
() = 0
i.e. f
()

K(X). Thus by restriction we obtain a homomorphismf
:

K(Y )
K(X). This obviously satises the composition rule (f g)
= g
and thus
K is a functor.
Proposition 9.10. Let X be a compact based space. Then we have a natural
split exact sequence
0

K(X)
K(X)
i
K(x
0
)

c
0
where i : x
0
X is the inclusion and c : X x
0
is the trivial map.
Thus, we get a natural group isomorphism
K(X)

K(X) K(x
0
)
=

K(X) Z. (9.1)
Proof. Split exactness of the sequence follows since c i = id
x0]
. To see that
the split is natural, let Y be another compact based space and let f : X Y
be a based map. Then we get homomorphisms
f
: K(Y ) K(X) , f
:

K(Y )

K(X) , f
: K(y
0
) K(x
0
).
9.1 The K-Functors 143
We need to se that the diagram
0

K(Y )
K(Y )
i
K(y
0
)

c
0
0

K(X)
K(X)
i
K(x
0
)

c
0
is commutative. The rst square commutes since f
:

K(Y )

K(X) is just
restriction of f
: K(Y ) K(X) to

K(Y ). The second square commutes since
f
Y
= (c
Y
f)
= (f i
X
)
= i
X
f
and likewise f
Y
= i
X
f
. Thus the split is exact. It is well-known that a split

exact sequence of groups induces an isomorphism K(X)

K(X) K(x
0
)
and this isomorphism is natural since the split is natural.
Example 9.11. Again we consider a one-point set, or, generally, a compact,
contractible, based space X. Then K(X) = Z and by the isomorphism in the
Proposition above we get

K(X) = 0. Similarly

KO(X) =

KSp(X) = 0.
One can put a quite dierent view on the reduced K-group, one that will,
to a certain extend, allow us to disregard the base points. We consider the set
of complex vector bundles over X and dene an equivalence relation
s
, called
stable equivalence by E
s
F i there exist positive integers m and n such that
E I
m

= F I
n
. The stable equivalence class containing E is denoted [E]
s
.
Observe, that a bundle is stably equivalent to 0, if and only of it is trivial.
The following lemma is an immediate consequence of Proposition 9.2:
Lemma 9.12. Direct sum in Vect
C
(X) gives Vect
C
(X)/
s
the structure of an
abelian group.
Proposition 9.13. If X is a compact based space, then there is a group iso-
morphism

K(X) Vect
C
(X)/
s
.
Proof. Consider the group homomorphism K(X) Vect
C
(X)/
s
mapping
the element [E] [I
k
] to [E]
s
. It is easily checked that the map is well-dened
and surjective. Since [E]
s
= 0 if and only if E is trivial, we see that the kernel of
the map is [I
m
] [I
k
] [ m, k N
0
. But this is isomorphic to K(x
0
). From the
standard Isomorphism Theorem from group theory the homomorphism above
induces an isomorphism
K(X)/K(x
0
)

Vect
C
(X)/
s
.
But since K(X)/K(x
0
)
=

K(X) the result follows.
Thus we see that

K(X) is (as a group) independent of the choice of base
point: it is canonically isomorphic to Vect
C
(X)/
s
which is dened without
reference to base points. However if we want to view

K(X) as a subgroup of
K(X) the base point is needed. The way

K(X) sits inside K(X) does in fact
depend on the base point. Also the ring structure on

K(X) depends on the base
point!
Armed with this alternative description of

K(X) we are able to prove the
following nice result:
Proposition 9.14. Let X and Y be connected based spaces and assume fur-
thermore that X and Y are connected. Then there is a group isomorphism
K(X Y )

K(X)

K(Y ).
In fact if we dene
X
: XY X to be the map satisfying
X
[
X
= id
X
and
X
(Y ) = x
0
and similarly
Y
: XY Y , then the inverse

K(X)
K(Y )
K(X Y ) is given by
(
1
,
2
)
X
(
1
) +
Y
(
2
). (9.2)
Proof. The idea is to show that the groups t into a split short exact sequence.
Let i
X
: X X Y and i
Y
: Y X Y be the inclusions. The claim is
that the following sequence is split exact:
0

K(X)

K(X Y )
i

K(Y )

0 .
First we see that i
Y
= id
K(Y )
since
Y
i
Y
= id
Y
, thus the sequence is
split, and i
Y
is surjective. Moreover we see that i
X
= (
X
i
Y
)
and that
X
i
Y
factorizes as Y x
0
X. Since

K(x
0
) = 0, we get i
X
= 0,
thus im
X
ker i
Y
.
X
is injective: if p : E X is a bundle over X then
X
(E) is the bundle
over X Y which is E over X and the product Y p
1
(x
0
) over Y . Assume
this to be 0 over X Y , this means that the bundle is trivial, in particular E is
trivial over X, i.e. [E]
s
= 0. Thus
X
is injective. Finally let E be a bundle over
XY and assume [E]
s
to be in ker i
Y
. Thus when restricted to Y the bundle E
is trivial. But then it is easy to see that E

=
X
(i
X
E), i.e. [E]
s
im
X
. Thus
split exactness of the sequence has been shown and the isomorphism follows.
9.2 The Long Exact Sequence
In this section we will investigate K-theory of a pair. Let X be a compact
Hausdor space and A a closed subset. We call such a pair (X, A) a compact
pair. For such a pair, the quotient space X/A is a compact space with base point
A/A. Note that we do not require A to be non-empty, if A = then we interpret
X/ as X
+
, i.e. X with a disjoint base point added. Given two compact pairs
(X, A) and (Y, B), we consider continuous maps f : X Y mapping A into
B. As one can check, such a map gives rise to a map

f : X/A Y/B between
the quotient spaces.
Denition 9.15. Given a compact pair (X, A), we dene the relative K-group
by
K(X, A) :=

K(X/A).
Lemma 9.16. There are isomorphisms: K(X, ) =

K(X
+
)
= K(X).
Proof. The rst follows simply by denition. To verify the other one, let
: X X
+
= X x
0
denote the inclusion. This induces a group ho-
momorphism

K(X
+
) K(X
+
) K(X), which we denote by
. We will
show that this is the desired isomorphism. To see that it is injective, assume that
we have elements [E] [I
k
], [E
t
] [I
]

K(X
+
) K(X
+
) (which just means
that they are in the kernel of the map K(X) K(x
0
) i.e. that dimE
x0
= k
and dimE
t
x0
= ) such that:
([E] [I
k
]) =
([E
t
] [I
]).
9.2 The Long Exact Sequence 145
Since
is just restriction to X, this simply means that we have the identity

E[
X
I

= E
t
[
X
I
k
.
But this identity extends to all of X
+
by virtue of the dimension relations
dimE
x0
= k and dimE
t
x0
= above, since x
0
is an isolated point and this just
means that we have the identity [E] [I
k
] = [E
t
] [I
], i.e.
is injective.
To check that
is surjective, let [E] [I

k
] K(X). Dene

E over X
+
by
E[
X
= E and

E
x0
= x
0
C
k
for some k, and let

I
k
be the extension of I
k
to X
+
. Then [
E] [
I
k
]

K(X
+
) since their restrictions to x
0
have the same
dimension namely k. Thus
is surjective, hence an isomorphism.

A map f : (X, A) (Y, B) induces a map f : X/A Y/B and hence a
ring homomorphism f
: K(Y, B) K(X, A). The composition rule is easily

veried and thus we have extended the K-functor to a functor on the category
of compact pairs.
In homology and cohomology theories we have higher order groups, like H
n
(X)
or H
n
(X). We want to construct something similar in K-theory. For this recall
the notion of suspension: For a space X the (unreduced) suspension SX is the
quotient of X I where X 0 and X 1 are collapsed to two points. We
take the image of X 0 under the quotient map as the base point. If X is a
based space, we can dene the reduced suspension X as the quotient of SX
collapsing x
0
I to a point. It becomes a pointed space by taking this col-
lapsed point as the base point. It is easy to see that S is a functor for compact
spaces to compact based spaces and that is a functor from the category of
compact based spaces to itself.
We can apply the suspension to the suspension and thus obtaining iterated
suspensions S
n
X and
n
X (where, of course, we dene S
0
X =
0
X := X). For
the use in K-theory it is really of no concern which suspension we use, for as the
quotient map SX X collapses a contractible subspace, Proposition 9.20
below will guarantee that it induces ring isomorphisms K(X)

K(SX)
and

K(X)

K(SX).
Recall also the cone on X: it is the functor mapping a compact space X to
the quotient CX obtained from X I by collapsing X 0 to a point. We
take this collapsed set to be the base point. By identifying X with the subset
X 1 of CX we get a natural inclusion X CX. Collapsing X CX to
a point we obtain SX i.e. SX

= CX/X.
Denition 9.17. Let X be compact, A X a closed subset and Y a compact
based space. For n 0 dene
K
n
(X) :=

K(
n
(X
+
))
K
n
(Y ) :=

K(
n
Y )
K
n
(X, A) = K
n
(X, A) :=

K(
n
(X/A)).
Furthermore we dene the total K-groups
K
(X) :=
n0
K
n
(X) ,

K
(Y ) :=
n0
K
n
(Y )
K
(X, A) :=
n0
K
n
(X, A).
Since suspension and

K are functors, the above dened maps are functors as
well. Thus, a map f : X Y will induce maps f
: K
n
(Y ) K
n
(X) and
similarly in the reduced and relative cases.
As we shall see in the next section most of these groups coincide. In fact they
repeat themselves with a period of 2. This is the celebrated Bott periodicity.
Example 9.18. Lets play around with these new denitions. Assuming (X, A)
to be a compact pair we identify A as a subset of CA and thus we may form
the union X CA. It is intuitively clear that we have a homeomorphism (X
CA)/X CA/A
= SA. Thus we get a chain of isomorphisms

K(X CA, X) =

K((X CA)/X)
=

K(SA)
=

K(A) =

K
1
(A).
Our next goal is to obtain a long exact sequence for pairs. A great step is
taken by the following lemma:
Lemma 9.19. Let (X, A) be a compact pair with A a based space (the base point
of A will then also be the base point of X), let i : A X be the inclusion and
q : X X/A be the quotient map. Then the following sequence is exact:
K(X/A)
q

K(X)
i

K(A)
.
Proof. We see that i
= (q i)
and that q i factorizes as the composition

A A/A X/A. Since

K(A/A) = 0, i
factors through the trivial group,

i.e. is the zero map. Consequently, imq
ker i
.
Conversely, assume that [E]
s
ker i
i.e. that [i
E]
s
= 0. This means that E
is trivial when restricted to A. We need to show that E is the pullback along
q of some vector bundle over X/A. Let : E X denote the projection.
As E is trivial over A we have a trivialization map :
1
(A) A C
n
,
i.e. a homeomorphism which is linear on bers mapping bers isomorphically
to C
n
. Consider the space

E, the quotient of E under the equivalence relation
1
(x, v)
1
(y, v) for x, y A and v C
n
, i.e. we collapse all the bers over
A to one ber. Let q : E

E denote the quotient map. We get an obvious
projection map :

E X/A rendering the following diagram commutative:
E
q
X
q
X/A
To show that

E is a vector bundle we need a trivialization in a neighborhood
around the point A/A (around all other points in X/A we have trivializations
coming directly from the bundle E).
From the trivialization we get a local frame i.e. continuous maps s
1
, . . . , s
n
:
A C
n
such that
1
(x, s
i
(x)) are sections of E over A and such that at each
point x A the set s
1
(x), . . . , s
n
(x) is a basis for C
n
. Now, let U
j
be a covering
of A of open sets on which E is trivial and let
j
: p
1
(U
j
) U
j
C
n
be
the trivializations. Since A is compact, we can assume this covering to be nite.
As A U
j
is closed in U
j
and as X is normal we can (by Tietzes Extension
Theorem) extend s
i
from U
j
A to U
j
. This extended map we call s
ij
. This
gives rise to a section
1
j
(x, s
ij
(x)) of E over U
j
. Let
j
be a partition of
unity subordinate the the cover (U
j
), then we can patch these sections together
to n sections over

j
U
j
. These need not constitute a basis for
1
(x) for each
point in
j
U
j
but at least in a neighborhood around A they will. Thus we have
a local frame in a neighborhood of A and this is the same as having a local
trivialization i a neighborhood of A.
This gives the desired trivialization of

E around A/A and hence :

E
X/A is a vector bundle.
9.2 The Long Exact Sequence 147
Finally, consider the map : E q
E) given by v ((v), q(v)). This

is easily seen to be an isomorphism of vector bundles. Thus [E]
s
= q
E]
s
, i.e.
[E]
s
imq
.
From the proof of this lemma we can extract the following result which reects
some kind of excision property for the K-groups:
Proposition 9.20. Let A X be a closed contractible subspace, and let q :
X X/A denote the quotient map. Then the induced maps
q
: K(X/A)

K(X) and q
:

K(X/A)

K(X)
are ring isomorphisms.
Proof. We will show that the pullback map q
: Vect
C
(X/A) Vect
C
(X) is
an isomorphism of semirings, and this we do by constructing an explicit inverse.
Let E Vect
C
(X). Since A is contractible, E[
A
is trivial and as in the proof
above we can nd a trivialization on an open set U Aand from this construct
a vector bundle

E over X/A. We will show that E

E is the inverse to q
,
but rst we have to check that it is well-dened, i.e. that it is independent of the
choice of trivialization. So assume that
t
is another trivialization over U
t
A
and let

E
t
be the corresponding bundle. We see that
t
= (
t

1
) , where
g :=
t
1
: UU
t
GL(n, C) is the transition map. Since A is contractible,
g[
A
is null-homotopic, and since GL(n, C) is path-connected, g[
A
is homotopic
to the constant map A x I
n
. This implies that we get a homotopy H from
[
A
to
t
[
A
, where for each t [0, 1] the map H
t
is a trivialization of A. Thus
for each t we get a bundle

E
t
over X/A and these we clutch together to a bundle
E
I
over (X/A) I. Since

E
0
=

E and

E
1
=

E
t
it follows from Proposition 9.3
that E and E
t
are isomorphic, i.e. the map is well-dened.
We saw in the proof of the lemma that q
E) = E, and as

q
E = E, the map
is the desired inverse.
From the lemma we can furthermore derive the main result of this section,
namely the existence of long exact sequences in K-theory:
Theorem 9.21 (Long Exact Sequences). Let (X, A) be a compact pair then
there exist connecting homomorphisms : K
n
(A) K
n+1
(X, A) such that
the following sequence is exact
K
2
(A)

K
1
(X, A)
q
K
1
(X)
i
K
1
(A)
K
0
(X, A)
q
K
0
(X)
i
K
0
(A).
Likewise, if (X, Y ) is a compact pair and A is based, then the following sequence
is exact

K
2
(A)

K
1
(X, A)
q

K
1
(X)
i

K
1
(A)
K
0
(X, A)
q

K
0
(X)
i

K
0
(A).
Proof. The core of the proof is to establish exactness of the following sequence:
K
1
(X)
i

K
1
(A)

K
0
(X, A)
j

K
0
(X)
i

K
0
(A). (9.3)
From Lemma 9.19 we get exactness at

K
0
(X). To show exactness at the other
two spots we use Lemma 9.19 for the pairs (X CA, X) and ((X CA)
CX, X CA). Namely, consider the sequence of inclusion and quotient
X X CA

(X CA)/X.
From Lemma 9.19 we get an exact sequence
K((X CA)/X)

K(X CA)

K(X).
We have (X CA)/X = SA and by Proposition 9.20 a natural isomorphism
1
:

K(A)

K(SA). The map X CA X/A collapsing the cone
induces an isomorphism
2
:

K(X/A)

K(X CA) (since the cone is
contractible). Dene :=
1
2

1
:

K
1
(A) K(X, A). Since the
quotient q : X X/A factorizes as a composition X X CA

X/A
we see that and q
t into the following commutative diagram
K((X CA)/X)

K(X CA)

K(X)
K(A)
1

K(X/A)
2

K(X)
Since the top row is exact the bottom row will be exact as well, yielding exactness
at K
0
(X, A) in (9.3). Exactness at

K
1
(A) is proved in exactly the same way.
We can now extend the sequence to the left by replacing X and A by X
and A thus obtaining the following exact sequence
K
2
(X)

K
2
(A)

K
1
(X, A)

K
1
(X)

K
1
(A)
which coincides with (9.3) at the two last spots. Continuing this process will
give us an innite long exact sequence.
To get the corresponding sequence for unreduced K-theory, just replace X
and A by X
+
and A
+
and recall that K(X) =

K(X
+
) and that K(X, A) =
K(X
+
, A
+
).
The reader may wonder about the missing 0 at the end of the long
exact sequence, after all the long exact sequences for homology and cohomology
theories do end at 0. In general it is not true that the map K(X) K(A)
is surjective. However, it is true, if A is a retract of X i.e. if there is a map
r : X A such that r[
A
= id
A
.
Corollary 9.22. For compact based spaces X and Y there exists a group iso-
morphism
K(X Y )
=

K(X Y )

K(X)

K(Y ). (9.4)
To be more specic, if
1
: X Y X and
2
: X Y Y are the
projection maps and q : X Y X Y the quotient map, then the inverse
K(X Y )

K(X)

K(Y )

K(X Y ) is given by
(
1
,
2
,
3
)
1
(
1
) +
2
(
2
) +q
(
3
). (9.5)
Proof. Since we have (XY ) = XY we get

K
1
(XY )
=

K
1
(X)
K
1
(Y ) by Proposition 9.14 the last part of the reduced long exact sequence
for the pair (X Y, X Y ) takes the following form:
K((X Y ))

K(X)

K(Y )

K(X Y )

K(X Y )

K(X)

K(Y ).
The sequence splits for we have a map

K(X)

K(Y )

K(X Y ) by
(
1
,
2
)
1
(
1
) +
2
(
2
) where
1
: X Y X and
2
: X Y Y are
9.3 Exterior Products and Bott Periodicity 149
the projections. This means in particular that the map is surjective, and hence
we can extend the sequence above with a zero.
Similarly the map

K((XY ))

K(X)

K(Y ) splits, in particular it
is surjective. Consequently, the map

K(X)

K(Y )

K(XY ) is the zero
map. Thus the exact sequence above is equivalent to the following split short
exact sequence
0

K(X Y )

K(X Y )

K(X)

K(Y ) 0.
The statement of the proposition now follows from elementary results on split
short exact sequences.
The results of Lemma 9.19 as well as Theorem 9.36 and Corollary 9.22 hold
also in the real and quaternionic case.
9.3 Exterior Products and Bott Periodicity
In this section we want to introduce some extra algebraic structure on K
(X)
and

K
(X). The way this is accomplished is to introduce exterior products,

one for reduced K-theory and later an unreduced analogue.
Let X and Y be compact based spaces. The exterior product in reduced K-
theory is a map :

K(X)
Z

K(Y )

K(XY ) dened in the following way:
Let
1

K(X) and
2

K(Y ). Then retaining the notation of
1
and
2
for the
projections from X Y onto X and Y respectively, we can form the product
:=
1
(
1
)
2
(
2
)

K(X Y ).
By (9.5) we know that in a unique way can be written on the form
1
(
1
) +
2
(
2
) +q
(
3
) (9.6)
and we want to show that the rst two terms are 0. Let
X
: X XY denote
the map x (x, y
0
) and similarly let
Y
denote the map y (x
0
, y). We see
that
2

X
: X Y maps x y
0
, i.e.
2
:

K(Y )

K(X) factorizes
through

K(y
0
) = 0, i.e.
2
= 0. Similarly
X
q
= 0. Furthermore, since
1

X
= id
X
, we get by applying the map
X
to (9.6) that
X
(
1
(
1
))
X
(
2
) =
1
.
But the left-hand side is 0, since
2
= 0, and thus
1
= 0. In a similar fashion
we get
2
= 0, and therefore = q
(
3
). Now we simply dene (
1
2
) := ,
this is the unique element in

K(X Y ) to which
1
(
1
)
2
(
2
) pulls back.
Since
i
X = S
i
X this exterior product is easy to extend to the higher
K-groups, namely, dene
:

K
i
(X)
Z

K
j
(Y )

K
ij
(X Y )
to be the exterior product

K(S
i
X)
Z

K(S
j
Y )

K(S
i+j
X Y ). In
particular if X = Y = pt the exterior product turns

K
(pt) into a graded

ring and, since X pt = X, the exterior product
K
i
(pt)
Z

K
j
(X)

K
ij
(X)
turns

K
(X) into a graded module

3
over

K
(pt).
3
A graded module M over a graded ring R =
i
R
i
is a module over R with a grading
M =
j
M
j
such that R
i
M
j
M
i+j
.
This exterior product can easily be extended to the unreduced case: If we
replace X and Y with X
+
and Y
+
and recall Lemma 9.16, we get a map
: K
i
(X)
Z
K
j
(Y )

K
ij
(X
+
Y
+
) = K
ij
(X Y ).
In the same way as above this exterior product turns K
(pt) into a graded

ring and K
(X) into a graded module over K
(pt).
All this works equally well for real K-theory: we can dene exterior products
KO
i
(X)
Z

KO
j
(Y )

KO
ij
(X Y )
and
KO
i
(X)
Z
KO
j
(Y ) KO
ij
(X Y )
in exactly the same way as above. These exterior products give

KO
(pt) and
KO
(pt) the structure of graded rings and turn

KO
(X) and KO
(X) into
graded modules over the rings

KO
(pt) and KO
(pt) respectively.
One may have observed, that we have not yet been able to calculate the higher
K-groups, even for a 1-point space, not even mentioning more complicated
spaces. In fact, to do so requires a genius like Bott or Atiyah:
Theorem 9.23 (Bott Periodicity I). Let H be the canonical line bundle over
CP
1
= S
2
, and let K
2
(pt) =

K(S
2
) be the element [H] [I
1
]. Then there
is a ring isomorphism
K
(pt)
= Z[].
In particular K
i
(pt) = 0 if i is odd, K
i
(pt)

= Z if i is even and for any i
multiplication by gives an isomorphism K
i
(pt)

K
i2
(pt).
We will not prove it here. In a later section, we shall deduce it, partly, from
the so-called Thom isomorphism.
Out next aim is to calculate the K-groups of spheres. Since pt
+
= S
0
, exam-
ining the denitions gives
K
i
(pt) =

K(
i
S
0
) =

K(S
i
). (9.7)
Thus the K-groups of a sphere is closely related to those of a point. From (9.7)
we read o the reduced K-groups of the spheres immediately:
K(S
n
) =
_
Z if n is even
0 if n is odd.
Thanks to Proposition 9.10 we get the unreduced K-groups as well:
K(S
n
) =
_
Z Z if n is even
Z if n is odd.
We may even calculate the zeroth K-groups of a torus
2
= S
1
S
1
by virtue
of (9.4)
K(
2
) =

K(S
1
S
1
)

K(S
1
)

K(S
1
) = Z
and by Proposition 9.10 we have also K(
2
) = Z Z.
For the higher K-groups we get
K
i
(S
n
) =

K(
i
S
n
) =

K(S
i
S
n
) =

K(S
n+i
).
In particular, these K-groups repeat themselves with a period of 2, just as in
the case of a point. This is not a coincidence, in fact this remarkable property
holds for any space:
9.4 Equivariant K-theory 151
Theorem 9.24 (Bott Periodicity II). Let X be a compact space. Then the
module multiplication by as dened above, gives isomorphisms
K
i
(X)

K
i2
(X) and

K
i
(X)

K
i2
(X).
It is an immediate consequence of the denition, that also relative K-groups
repeat themselves with a period of 2:
K
i
(X, A) =

K(
i
(X/A)
+
) =

K
i
((X/A)
+
) =

K
i2
((X/A)
+
)
=

K(
i+2
(X/A)
+
) = K
n2
(X, A).
At this point it is very important to stress that this is where complex, real and
quaternionic K-theories go their separate ways: There are periodicity relations in
these theories as well, but in the real case the period is 8 and in the quaternionic
case the period is 6. The lower period in complex K-theory makes the complex
K-groups substantially simpler to compute.
Corollary 9.25. Composing the isomorphism K
0
(A)

K
2
(A) with the
connecting homomorphism : K
2
(A) K
1
(X, A) from the long exact
sequence we can reduce the long exact sequence of the pair (X, A) to the following
six-term exact sequence
K
0
(X, A)

K
0
(X)

K
0
(A)
K
1
(A)
K
1
(X)
K
1
(X, A)
A similar statement holds in the reduced case.

9.4 Equivariant K-theory
In this section we generalize the K-theory described thus far to topological
spaces X carrying a continuous action of a topological group, where by contin-
uous action we mean a continuous map : G X X, written (g, x)
g
(x) = g x satisfying g
1
(g
2
x) = (g
1
g
2
) x. For short we say that X is a
G-space. A G-map or equivariant map between to G-spaces X and Y is a map
f : X Y such that f(g x) = g f(x).
Denition 9.26. Let X be a G-space. A (complex) G-vector bundle or simply
a G-bundle (not to be confused with a principal G-bundle which is something
completely dierent) over X is a complex vector bundle : E X over X,
the total space of which is a G-space in such a way that is a G-map (i.e. if
v E
x
then g v E
gx
) and such that the action
g
: E
x
E
gx
is linear.
Note that all this reduces to the usual denition of a vector bundle if G is the
trivial group. Note, however, that a the total space of a G-bundle over a space
X with a trivial G-action need not be a trivial G-space
Example 9.27. 1) The product bundle: Let X be a G-space and V be a G-
module (i.e. a complex vector space with a representation : G Aut(V )).
Then V := X V is a G-bundle with the action g (x, v) = (g x, (g)v). This
obviously satises the requirements in the denition.
2) Complexied tangent bundle: Let X be a smooth manifold with a smooth
G-action and consider the complexied tangent bundle: TX
C
:= TX
R
C.
Then the dieomorphism
g
: X X induces a push-forward bundle map
(
g
)
: TX TX and hence also a self-map on the complexied bundle.

This map is easily seen to be a smooth action on the space TX
C
and since it
is linear on bers and maps (
g
)
: T
x
X
R
C T
gx
X
R
C the projection
map is equivariant with respect to these G-action, hence turning TX
C
into a
G-bundle.
Denition 9.28. Let E be a G-bundle. A section s of E which is also a G-map
is called equivariant. The vector space of equivariant sections is denoted
G
(E).
If E and F are two G-bundles over X then a bundle map : E F is called
equivariant or a G-bundle map if it is a G-map. The set of G-bundle maps we
denote by Hom
G
(E, F).
In particular a G-bundle isomorphism is a G-bundle map which is also a
bundle isomorphism. The set of G-isomorphism classes of G-bundles over X is
denoted Vect
G
C
(X). By [E] we denote the G-isomorphism class containing the
G-bundle E.
All the usual constructions of vector bundles carry over to G-bundles: We can
form the direct sum EF by endowing it with the action g (v, w) = (g v, g w).
We can form the tensor product E F, by giving it the action g (v w) =
g v g w (extended linearly). Finally, given a G-map f : Y X and a
G-bundle E over X we give the pullback bundle
f
E = (x, v) Y E [ f(x) = (v)

the action g (x, v) = (g x, g v). This is well-dened since f(g x) = g f(x) =
g (v) = (g v), i.e. g (x, v) f
E. It turns f
E into a G-bundle.
As for ordinary vector bundles (cf. Corollary 9.4) one can show
Proposition 9.29. Let f
0
, f
1
: Y X be two G-maps which are G-homotopic,
i.e. homotopic through G-maps, and let E be a G-bundle over X. Then the G-
bundles f
0
E and f
1
E are G-isomorphic.
We wont prove it here
4
. An important corollary is of course that any G-
bundle over a G-contractible base space is trivial.
Another important result that will become useful to us is the following gen-
eralization of Proposition 9.2
5
:
Proposition 9.30. Let X be a compact G-space and E a G-bundle over X.
Then there exists another G-bundle E
over X and a G-module V such that

E E

= V.
Thus, over compact spaces we always have complementary G-bundles.
The set of G-isomorphism classes of G-bundles over X is an abelian semigroup
w.r.t. direct sum, thus we may dene:
Denition 9.31. Let X be a compact G-space. The equivariant K-group
K
G
(X) of X is the Grothendieck group of the semigroup of G-bundles over
X.
By Proposition 9.30 the semigroup Vect
G
C
(X) has the cancellation property
and thus, as in ordinary K-theory the elements of K
G
(X) are formal dierences
[E
0
][E
1
] of isomorphism classes, two such dierences [E
0
][E
1
] and [F
0
][F
1
]
being equal if
E
0
F
1

= F
0
E
1
.
Direct sum of G-bundles gives K
G
(X) the structure of an abelian group, and
tensor product gives K
G
(X) the structure of a commutative ring.
4
For a proof see [Seg] Proposition 1.3.
5
A proof is given in [Seg] Proposition 2.4.
9.4 Equivariant K-theory 153
If f : Y X is a G-map the pullback map on G-bundles induces a map in
K-theory f
: K
G
(X) K
G
(Y ) and as in ordinary K-theory this is a ring-
homomorphism. Thus K
G
becomes a contravariant functor from the category of
G-spaces and G-maps to the category of commutative rings. From Proposition
9.29 we see that two G-homotopic maps induce the same maps in equivariant
K-theory.
Example 9.32. Let us calculate the equivariant K-group of a one-point G-
space. By Example 9.27 we see that Vect
G
C
(pt) equals the semigroup of nite-
dimensional G-modules. The Grothendieck group of this is the so-called repre-
sentation ring R(G) which is just the free abelian group generated by the set
of nite-dimensional G-modules. Thus K
G
(pt) = R(G). This is in accordance
with the result from usual K-theory, K(pt)
= Z, since the trivial group e has

only one irreducible representation so that its representation ring is the integer
ring.
Let X be a space. We say that X is a based or pointed G-space if it has a
base point x
0
satisfying g x
0
= x
0
for all g G. The inclusion of a base point
into the based space is readily seen to be a G-map.
Denition 9.33. Let X be a compact based G-space the base point of which
is denoted x
0
. Letting i : x
0
X denote the inclusion we denote by

K
G
(X)
the sub-ring ker i
of K
G
(X). This is called the reduced equivariant K-group of
X.
Thanks to Proposition 9.30 we have another description of this reduced group.
We say that two G-bundles E and F are G-stably equivalent if there exist G-
modules V and W such that
E V
= F W.
This equivalence relation on Vect
G
C
(X) we denote
s
. Now the proof of Propo-
sition 9.13 carries over more or less verbatim to yield:
Proposition 9.34. The quotient Vect
G
C
(X)/
s
is a group which is canonically
isomorphic to the group

K
G
(X).
Thus again, the reduced group itself does not depend on the base-point but
the ring structure and the way it sits inside K
G
(X) do.
Before we can dene the relative and higher order equivariant K-groups we
need to investigate how group actions behave with respect to quotients. Let X
be a compact space with G-action and A X a closed subspace which is
invariant under the group action, i.e. g A A for all g G. Such a pair is
called a compact G-pair. We want to give the quotient space X/A a G-action.
Let q : X X/A denote the quotient map. Then there is a unique continuous
map

: GX/A X/A making the diagram commute
GX
q
J
J
J
J
J
J
J
J
J
id
G
q
GX/A
X/A
and this is easily seen to be a G-action on X/A.
If X is a compact based G-space, then we can equip X I with the G-action
g (x, t) = (g x, t). With this G-action both X 0, X 1 and x
0
I
are G-invariant closed subspaces of X I and thus we can collapse them and
obtain a G-action on the quotient space which is of course just X.
In complete analogy with ordinary K-theory we can now dene dene:
Denition 9.35. Let (X, A) be a compact G-pair and Y a compact based
G-space. For n 0 dene
K
n
G
(X) :=

K
G
(
n
(X
+
))
K
n
G
(Y ) :=

K
G
(
n
Y )
K
n
G
(X, A) =

K
n
G
(X, A) :=

K
G
(
n
(X/A)).
Furthermore we dene
K
G
(X) :=
n0
K
n
G
(X) ,

K
G
(Y ) :=
n0
K
n
G
(Y )
K
G
(X, A) :=
n0
K
n
G
(X, A).
We even have long-exact sequences:
Theorem 9.36. Let (X, A) be a compact G-pair then there exist connecting
homomorphisms : K
n
G
(A) K
n+1
G
(X, A) such that the following sequence
is exact
K
2
G
(A)

K
1
G
(X, A)
q
K
1
G
(X)
i
K
1
G
(A)
K
0
G
(X, A)
q
K
0
G
(X)
i
K
0
G
(A).
Likewise, if (X, A) is a compact G-pair and A is furthermore a based G-space
then the following sequence in reduced K-theory is exact

K
2
G
(A)

K
1
G
(X, A)
q

K
1
G
(X)
i

K
1
G
(A)
K
0
G
(X, A)
q

K
0
G
(X)
i

K
0
G
(A).
The proof doesnt really involve the group G and so the earlier proof for usual
K-theory will hold in this case as well.
If X is a locally compact space, we can form the 1-point compactication
X which we denote by X
+
. This is in accordance with the previous
use of the same notation, for if X happens to be compact, then the 1-point
compactication is indeed just X with a disjoint point added. If X is a G-space,
we can extend the G-action to X
+
by dening g= . Since
g
: X X is a
homeomorphism, hence a proper map, the extension is continuous X
+
X
+
.
We thus extend equivariant K-theory to the category of locally compact spaces
in the following way:
Denition 9.37. Let X be a locally compact G-space. Then we dene
K
n
G
(X) :=

K
G
(
n
(X
+
)) and K
n
G
(X, A) := K
n
G
(X
+
, A
+
).
This is called equivariant K-theory with compact support. If X is already
compact, then one can show (as in Lemma 9.16) that

K
G
(X
+
)

= K
G
(X) and
hence that this new denition is a genuine extension of equivariant K-theory to
the category of locally compact spaces. However, the functorial properties are
a bit more complicated than before: In fact, K-theory with compact support is
a functor in two dierent ways: First, it is a contravariant functor with respect
to proper maps (i.e. maps whose pre-image of compact sets are compact sets).
For if f : X Y is a proper map, then this can be extended in an obvious
way to a continuous map f
+
: X
+
Y
+
and this induces a map (f
+
)
K
G
(Y
+
)

K
G
(X
+
) i.e. a map K
G
(Y ) K
G
(X).
9.5 The Thom Isomorphism 155
Second, it is a covariant functor with respect to inclusions: if X is a locally
compact space and U X is an open subset of X, then we get a homomorphism
K
G
(U) K
G
(X) induced by the map
X
+
X
+
/(X
+
U)

U
+
.
The following identities will become useful:
Lemma 9.38. Let X be a locally compact G-space and A closed a G-subset.
Then we have
K
n
G
(X) = K
G
(X 1
n
) and K
n
G
(X, A) = K
G
(X 1
n
, A1
n
).
Proof. This is an exercise in the denitions: Since
n
X = S
n
X
+
= (1
n
X)
+
we see that
K
n
G
(X) =

K
G
(
n
X
+
) =

K
G
((X 1
n
)
+
) = K
G
(X 1
n
).
To see that the second identity is true, we note that
n
(X
+
A
C(A
+
)) =
n
X
+
n
A
C(
n
(A
+
))
and that K
n
G
(X, A) =

K
G
(
n
(X
+
A
CA
+
)) and therefore
K
n
G
(X, A) =

K
G
(
n
(X
+
A
CA
+
)) =

K
G
(
n
X
+
n
A
C(
n
A
+
))
=

K
G
(X
+
S
n
C(A
+
S
n
)) =

K
G
(X
+
S
n
, A
+
S
n
)
= K
G
(X 1
n
, A1
n
).
9.5 The Thom Isomorphism
Now we will take a quite dierent approach to equivariant K-theory, a descrip-
tion which goes via complexes of G-bundles. This will help us dene the so-called
Thom isomorphism and ultimately lead us to Bott periodicity.
Out starting point will be the following: Let X be a G-space and consider a
nite complex E
of G-bundles over X

E
i1
di1
E
i
di
E
i+1
di+1

where the d
i
s are G-bundle maps such that d
i
d
i1
= 0 and nite means that
E
i
= 0 when [i[ N for some N. Such an object we call a G-complex and the
maps d
i
we call dierentials. We dene the support supp E
to be the set of
x X for which the sequence of vector spaces and linear maps

E
i1
x
di1
E
i
x
di
E
i+1
x
di+1

fails to be exact.
Lemma 9.39. The support of a complex is a closed set.
Proof. First note that the set
x [ dimimd
i
x
> k
is open, for if d
i
x0
has rank greater that k it has a matrix representation in which
there is a k k-submatrix which has non-zero determinant. Since this matrix
depends continuously on x there is at least a neighborhood around x
0
where the
same k k-submatrix has non-zero determinant. Thus the set above is open. In
the same way we have that
x [ dimker d
i
x
<
is open. Thus the sets
x [ dimimd
i
x
k and x [ dimker d
i
x
k
are closed.
We see that
supp E
=
_
i
x [ dimker d
i
x
> dimimd
i1
x

and that
x[ dimker d
i
x
> dimimd
i1
x
=
_
_
x[ dimker d
i
x
x[ dimimd
i1
x
> 1
_
is closed.
If the complex is exact at every point (i.e. if the support is empty) we say
that the complex is acyclic. An acyclic complex consisting of only two non-zero
bundles is called a simple acyclic complex. It is an elementary fact that any
acyclic complex can be written as a direct sum of simple acyclic complexes.
A morphism of complexes f : E
is a collection of G-bundle maps

f
i
: E
i
F
i
commuting with the dierentials, i.e. making the diagram com-
mutative:

E
i1
di1
f
i1
E
i
di
f
i
E
i+1
di+1
f
i+1

F
i1
di1
F
i
di
F
i+1
di+1

A morphism f of complexes for which each of the G-bundle maps f
i
is a bundle
isomorphism is called an isomorphism of complexes.
Example 9.40. An example which will become important shortly when we
construct the Thom isomorphism is the so-called Koszul complex. Let E be a
G-bundle over X and s and equivariant section of E. From this we construct
the following complex
0

C
d0
E
d1

2
E
d2

dn1

n
E

0
consisting of exterior powers of E, which are given the natural G-action:
g (
1

k
) = g
1
g
and with morphisms d
k
() = s(x) if
k
E
x
. The maps d
k
are easily
seen to be G-bundle maps: they obviously map bers to bers and if (U, ) is
a trivialization for E, then it also trivializes
k
E, i.e.
k
E[
U

= U
k
C
n
and
locally the map d
k
takes the form
U
k
C
n
(x, ) (x, (s(x)))
and this is continuous. Thus d
k
is a bundle map, and clearly d
k
d
k1
= 0. To see
that they respect the G-action, let be in the ber over x:
d
k
(g ) = (g ) s(g x) = (g ) g s(x) = g ( s(x)) = g d
k
().
Since wedging with a nonzero vector is known to yield an exact sequence, we
see that the support of the Koszul complex equals the set of zeros of s. If s is
everywhere non-zero, the complex is acyclic.
The usual operations on G-vector bundles can be generalized to G-complexes:
If E
and F
are two G-complexes we can form their direct sum E

E
i
F
i
did
E
i+1
F
i+1

where we simply form the direct sum of the corresponding vector bundles and
dierentials. We see that supp(E
) = supp E
supp F
, for if x is some
point where E
or F
fails to be exact, then the direct sum complex fails to be

exact at that point and conversely if both E
and F
are exact at x, then also

E
is exact at x.
In almost the same way we dene the tensor product of two G-complexes
E
. The kth bundle in this complex should be

(E
)
k
=
i+j=k
E
i
F
j
and the dierential D
k
: (E
)
k
(E
)
k+1
is given by
D
k
=
i+j=k
_
(d
i
id) + (id d
t
j
)
_
meaning that on the component E
i
F
j
the dierential maps it to E
i+1
F
j
by d
i
id and to E
i
F
j+1
by id d
t
j
. This time we have supp(E
) =
supp E
supp F
i.e. only at points x where both complexes fail to be exact,

the tensor product will fail to be exact.
The third and nal construction is the pullback. Let E
be a G-complex over
X and let f : Y X be a G-map. We dene a G-complex f
(E
) over Y by
f
(E
)
k
= f
E
k
with the dierential d
k
: f
E
k
f
E
k+1
given by
d
k
(y, ) = (y, d
k
()).
We see that supp(f
(E
)) = f
supp(E
).
Let (X, A) be a compact G-pair and let L
G
(X, A) denote the set of isomor-
phism classes of G-complexes having compact support inside X A, i.e. on
elements of A the complex is exact. In particular, all acyclic complexes live
in L
G
(X, A) for all A. If A happens to be the empty set, we will write L
G
(X)
instead of L
G
(X, ) for the set of isomorphism classes of G-complexes with com-
pact support. If the complexes E
and F
both have support inside X A then

as supp(E
) supp E
supp F
X A the space L
G
(X, A) is closed
under direct sum, in fact it is a semigroup w.r.t. direct sum. Furthermore, as
tensor product of complexes decreases support, L
G
(X, A) is closed under tensor
product as well, thus turning it into a semi-ring.
Inside L
G
(X, A) we dene the notion of homotopy. Two complexes E
and
F
in L
G
(X, A) are said to be homotopic if there exists a G-complex H

L
G
(X[0, 1], A[0, 1]) (the product space X[0, 1] is equipped with the usual
G-action g (x, t) = (g x, t)) such that H
[
X0]

= E
and H
[
X1]

= F
.
This equivalence relation is denoted by

=h
. It is easy to see that if E

=h

E
then
E

=h

E
. (9.8)
Lemma 9.41. Let X be compact, then any complex

E
i1
di1
E
i
di
E
i+1
di+1

is homotopic to the corresponding complex where the dierentials are all 0.
Proof. Let p denote the projection X [0, 1] X. Put

E
i
= p
E
i
then the
ber

E
i
(x,t)
over (x, t) is isomorphic to E
i
x
. Dene a dierential

d
i
:

E
i

E
i+1
by
d
i
(v) = td
i
(v)
if v

E
(x,t)

= E
x
. Then

E
becomes a complex over X [0, 1], and we see

that

E
[
X1]
is isomorphic to E
and that

E
[
X0]
is a complex with 0-
dierentials.
Introduce also in L
G
(X, A) the equivalence relation: E
i there exist
acyclic complexes H
0
and H
1
on X such that
E
=h
F
1
.
The equivalence class containing the complex E
is denoted E
. Observe, that
all acyclic complexes are equivalent.
Theorem 9.42. The quotient space L
G
(X, A)/ is naturally an abelian group
which is naturally isomorphic to K
G
(X, A).
Proof. We show it only in the very special case where X is compact and
A =
6
. The group G doesnt really matter, so without loss of generality we
may assume that G is the trivial group.
First we show that L(X)/ is a group. We dene addition by
E
+F
= E
.
From (9.8) it is easy to deduce that this operation is well-dened. If F
is an
acyclic complex, then by denition of , one can see that E
+F
= E
.
Thus, we have identied our neutral element.
The most complicated part is to construct inverses: Let a complex E
:
0

E
1
d1
E
2
d2

dn1
E
n
0
be given. To E
1
we pick a complementary bundle F
1
, i.e. E
1
F
1
= X C
k1
.
To E
2
we pick a bundle F
2
such that E
2
F
2
= X C
k2
where k
2
k
1
i.e. so that an injective linear map
1
: C
k1
C
k2
exists. Pick a bundle F
3
such that E
3
F
3
= X C
k3
, where k
3
is so big that an injective linear map
(im
1
)
C
k3
exists. Let
2
: C
k2
C
k3
be the linear extension which is 0
on im
1
. Continue in this way to nd bundles F
i
such that E
i
F
i
= X C
ki
and such that the sequence
0

C
k1
1
C
k2
2

n1
C
kn

0
is exact. Let F
denote the complex consisting of the bundles F

i
and zero dier-
entials. Let H
denote the complex with H

i
= X C
ki
and whose dierentials
are
d
i
(x, v) = (x,
i
(v)).
By construction H
is acyclic. Since E
and H
are homotopic (they are

by the lemma above both homotopic to the corresponding complex with zero
dierentials), we have E
+F
= H
= 0, i.e. F
is an inverse to E
.
Thus L(X)/ is a group.
Consider the map

: L(X) K(X) given by
E
k
(1)
k
[E
k
].
6
For a full proof, see [Seg] Proposition 3.1 and Appendix A.
This is easily seen to be additive. If E
and F
are homotopic complexes, then

by Proposition 9.3

(E
) =

(F
). Let H
be an acyclic complex. This we

write as a sum of simple acyclic complexes H
= H
1
H
n
, but since H
i
consists solely of two (isomorphic) non-zero bundles H
i
is mapped to zero by
. Thus if E
we have
(E
) =

(F
).
Thus we may dene : L(X)/K(X) by
(E
) =
k
(1)
k
E
k
and this is a group homomorphism. It is trivially surjective, for the element
[E] [F] K(X) is hit by the complex
0

E
0
F

0 .
To see that it is injective, assume that the complex E
0

E
1
d1
E
2
d2

dn1
E
n
0
is mapped to 0 by

. Let F
i
be a bundle, such that E
k
F
k
is trivial. Then
this implies that the bundle
H := F
1
E
2
F
3
E
4

is trivial.
By Lemma 9.41 the complex E
is homotopic to the complex

0

E
1
0
E
2
0
E
n
0
To this we add the complex 0

F
1
id
F
1
0 yielding the new
complex
0

I
i1
E
2
F
1

0
E
n
0
To this we add the complex
0

E
1
F
2
id
E
1
F
2
0
where the rst E
1
F
2
is added to E
2
E
1
(yielding a trivial complex I
i2
) and
the second is added to E
3
. Continuing in this way will ultimately give us the
following complex
0

I
i1
I
i2

I
in1
H

0
Thus, by adding acyclic complexes to E
we have obtained a complex consisting

solely of trivial complexes. In much the same way as we constructed inverses
in the beginning of this proof, one can to this trivial complex add acyclic
complexes such that it becomes homotopic to an acyclic trivial complex. Thus
[E
] represents 0 in L(X)/ and is injective.

Thus we have obtained a new picture of equivariant K-theory involving G-
complexes instead of G-bundles. Being able to switch back and forth between
the two pictures can often be useful, especially when we come to discuss index
theory.
The rst asset of this new picture is seen in the following when we construct
the Thom isomorphism. The problem is as follows: Given a G-vector bundle
: E X we would like somehow to obtain a homomorphism between K
G
(E)
and K
G
(X) and preferably (since E and X are homotopy equivalent topological
spaces) an isomorphism. However, homotopy equivalence alone cant save us, for
if X is compact, E is non-compact, so they belong to two dierent categories
of topological spaces, and thus the projection map does not induce a homo-
morphism in K-theory and if X is only locally compact, the projection map
is neither an inclusion nor a proper map, so neither in this case do we get an
induced homomorphism in K-theory.
So in order to obtain such a map, we have to proceed dierently, and this is
where the complexes enter the stage: Let : E X be some xed G-bundle
over X (which is still assumed to be at least locally compact). Pull E back
along to obtain the bundle
E over E. This bundle has a natural section s,

namely E (, )
E, and this is easily seen to be equivariant. From

the bundle
E and the section s we get a Koszul complex over E (cf. Example

9.40):
0

C
s(x)
E
s(x)

2
(
E)
s(x)

s(x)

n
(
E)

0
which in the sequel will be denoted
E
.
From a given G-complex F
L
G
(X) we pull back to E and tensor it with
the Koszul complex to obtain
(F
E
which is a complex over E. The support of this is
1
(supp F
) E
0
where E
0
denotes the zero-section of E which equals the set of zeros for s. This set is
homeomorphic to supp F
which is compact, i.e.
(F
E
L
G
(E). Thus
we get a map L
G
(X) L
G
(E) and this map is additive. One can check that
it induces a map on the quotient, thus giving a map
: K
G
(X) K
G
(E).
This is the Thom homomorphism.
We can extend this to the higher K-groups in the following way: Given a
G-bundle : E X we dene a G-bundle : E 1
n
X 1
n
by
(v, w) = ((v), w), i.e. the ber over the point (x, w) X 1
n
is E
x
w.
If we give X 1
n
the G-action g (x, w) = (g x, w) and E 1
n
the same
G-action g (v, w) = (g v, w), becomes an equivariant map, and E1
n
is a G-
bundle. Thus we get a Thom homomorphism K
G
(X1
n
) K
G
(E1
n
) and
composing with the isomorphisms K
n
G
(X)

K
G
(X1
n
) and K
n
G
(E)

K
n
G
(E 1
n
), we obtain a map
: K
n
G
(X) K
n
G
(E).
Putting all these together gives us a Thom homomorphism
: K
G
(X) K
G
(E).
Actually, the Thom homomorphism can be shown to be more than just a ho-
momorphism:
Theorem 9.43 (Thom Isomorphism). For a G-bundle E over a locally com-
pact space X, the Thom homomorphism K
G
(X) K
G
(E) is a natural group
isomorphism.
Of course since
, by construction, respects the grading,
: K
n
G
(X)
K
n
G
(E) is also an isomorphism for any i.
Example 9.44. Lets see what information we can extract from the Thom
isomorphism in the case of a trivial group G and a trivial bundle E = X C
n
.
Then we get an isomorphism
K(X)
= K(XC
n
) =

K((XC
n
)
+
) =

K(X
+
S
2n
) =

K(
2n
X
+
) = K
2n
(X).
In fact if we consider the trivial bundle (X 1
i
) C
n
over X 1
i
and recall
Lemma 9.38 we get
K
i
(X) = K(X 1
i
)
= K(X 1
i
C
n
) = K(X 1
2n+i
)
= K
i2n
(X),
that is we have retrieved the Bott periodicity from the Thom isomorphism.
That this is possible is a mark of how deep a result the existence of the Thom
isomorphism really is.
Assume we have two bundles : E X and : F X. Then :
E F X is a bundle over X, but it can also be viewed as a bundle over E,
by the obvious map p : E F E. How do the dierent Thom isomorphisms
interact?
Corollary 9.45 (Transitivity of the Thom isomorphism). Let the sit-
uation be as above, then the Thom isomorphisms for the bundles E X,
E F X and E F E make the following diagram commute
K
G
(X)

J
J
J
J
J
J
J
J
J
K
G
(E F)
K
G
(E)
q
q
q
q
q
q
q
q
q
q
Proof. First of all, note that E F E is isomorphic to the bundle
F,
thus we will reserve the notation E F to mean the bundle over X. Secondly,
observe that
EF
= p
E
q
F
.
This is a simple consequence of the fact that ( )
(EF) = p
Eq
F
and that
k
(E F) =
k
j=0
j
E
kj
F
(note how this ts with the denition of tensor products of complexes).
Now for the Thom isomorphisms: The isomorphism K
G
(X)

K
G
(E F)
is given by
H ( )
(H)
EF
= ( )
(H) p
E
q
F
.
The composition K
G
(X) K
G
(E) K
G
(E F) is given by
H
E
p
H p
F
.
We see that p
H equals ( )
H simply because = p, and that
F
= q
F
since p = q.
Finally, how does the Thom isomorphism for a pullback bundle relate to the
Thom isomorphism of the original bundle? Fortunately they are related in the
nicest possible way:
Proposition 9.46. Let : E X be a vector bundle and f : Y X a
continuous map. Let
: K(X)

K(E) denote the Thom isomorphism for
E and
: K(Y )

K(f
E) denote the Thom isomorphism for f
E. The
natural map F : f
E E is proper and hence induces a map F
: K(E)
K(f
E) and the following diagram commutes

K(X)

f
K(E)
F
K(Y )
K(f
E)
Unfortunately, Ive not been able to nd a proof of this claim.
Chapter 10
Characteristic Classes
10.1 Connections on Vector Bundles
In a smooth vector bundle there is a priori no way of taking the derivative of
sections. Well, for the tangent bundle and tensor products of such there is of
course the Lie derivative, but this has the serious drawback of not providing us
with a well-dened way of taking derivatives along curves. Therefore we need
the notion of a connection which, intuitively speaking, is a device which tells us
how the bers are packed together.
Before we get started, let us x the notation. Throughout this chapter, M and
N will denote smooth manifolds without boundary. By C
(M) = C
(M, 1) is
meant the algebra of smooth real-valued functions on M, and by C
(M, C) is
meant the algebra of smooth complex-valued functions on M. It is easy to see
that a smooth complex-valued function f can be written on the form f = f
1
+if
2
where f
1
, f
2
C
(M), and thus that

C
(M, C) = C
(M)
R
C.
If it is clear from the context that the scalar eld is K, we will use the short-hand
notation C
(M) to mean C
(M, K). Thus C
(M) need not just refer to the

algebra of real-valued functions.
TM and T
M denote the tangent bundle and cotangent bundle over M re-

spectively, and T
C
M denotes the complexied cotangent bundle. A section of
the complexied cotangent bundle is just a complex 1-form , i.e. if X is a
smooth vector eld, then (X) is a smooth complex-valued function on M.
The set of sections of T
C
M is denoted
1
(M, C), and it is easy to see that
1
(M, C) =
1
(M, 1)
R
C, i.e. we can split =
1
+i
2
where
1
and
2
are
real-valued 1-forms.
In general we can form the bundle
k
(T
K
M) and the sections of this, are
K-valued k-forms. The set of such is denoted
k
(M, K). It is well-known that
this is a C
(M)-module.
Now we can introduce connections on vector bundles. The setup is as fol-
lows: Let M be a smooth manifold and E a smooth vector bundle over K
(i.e. over 1 or C). By T
K
M we simply mean the cotangent bundle if K = 1
and Hom
R
(TM, C) = T
M
R
C (i.e. the complexied cotangent bundle) if
K = C. By
k
(M, E) we shall denote the module over C
(M) (recall that this

means C
(M, C) if K = C) of E-valued k-forms, i.e. sections of the bundle
k
(T
K
M)
K
E. Any such section is a linear combination of elements of the
form s where is an ordinary k-form and s is a section of E. It is easy to
see that if E = MK
n
then elements of
k
(M, E) are just n-tuples of k-forms.
163
164 Chapter 10 Characteristic Classes
Denition 10.1 (Connection). By a connection on E we understand a K-
linear map
: (E)
1
(M, E)
satisfying the Leibniz rule
(fs) = df s +fs (10.1)
for f C
(M) and s (E). s is called the covariant derivative of s.

The set of connections on E is denoted /
E
.
For each s (E), s is a smooth section of the bundle T
K
M E

=
Hom
K
(TM, E), i.e. s
p
is a K-linear map T
p
M
K
E
p
. But smooth sections of
the bundle Hom
K
(TM, E) are in 1-1 correspondence with smooth bundle maps
TM E which again are in 1-1 correspondence with C
(M)-linear maps
X(M) (E). Thus if X is a vector eld on M we get a smooth section of E
by dening
X
s := s(X) i.e.
(
X
s)(p) := s
p
(X
p
).
As mentioned this is C
(M)-linear in X, i.e.
gX
s = g
X
s (but this can also
be seen from a direct computation). In this description the Leibniz rule takes
the following form:
X
(fs) = (fs)(X) = (df s +fs)(X) = df(X)s +fs(X)
= (Xf)s +f
X
s.
Phrased in this way, a connection is a map : X(M) (E) (E) which is
C
(M)-linear in the rst variable, K-linear in the second and satises

X
(fs) =
(Xf)s +f
X
s. Sometimes this is the denition of a connection.
Of course, one might wonder: does such an object exist at all? The answer is
yes: Assume rst that E is a product bundle. This means that we have a canon-
ical global frame s
1
, . . . , s
n
. If v
1
, . . . , v
n

1
(M, E) are chosen arbitrarily
then dene a connection by
(f
1
s
1
+ +f
n
s
n
) =
n
i=1
(df
i
s
i
+f
i
v
i
). (10.2)
Clearly, this is a K-linear map (E)
1
(M, E) and we see
(fs) = (ff
1
s
1
+ +ff
n
s
n
) =
n
i=1
(d(ff
i
) s
i
+ff
i
s
i
)
=
n
i=1
_
f(df
i
s
i
) +ff
i
s
i
+f
i
df s
i
_
= f(f
1
s
1
+ +f
n
s
n
) +
n
i=1
df f
i
s
i
= f(s) +df s.
Thus, at least product bundles have connections and lots of them. If v
1
= =
v
n
= 0 then the corresponding connection is called the trivial connection.
If E is an arbitrary vector bundle and (U
) is a trivialization cover of M then

the results above apply to the trivial vector bundles E[
U
giving a connection
on E[
U
. Let (
) be a partition of unity subordinate to the cover, then

a calculation as the one above will show that

is a connection on E.
Therefore /
E
is non-empty. In fact, /
E
is quite plentiful: Let End(E) denote
10.1 Connections on Vector Bundles 165
the bundle E
E, then and let A

1
(M, End(E)) act on a section s (E)
by
A(s)(X) := A(X)s.
This is obviously K-linear in s, and we have
A(fs)(X) = A(X)(fs) = fA(X)s = fA(s)(X).
Thus we are in position to show:
Proposition 10.2. If is a connection on E and A
1
(M, End(E)) then
+A is a connection. Conversely, any connection is of the form +A where
A
1
(M, End(E)).
Proof. At rst we check that + A is a connection. The the identity above
we have
(+A)(fs) = (fs) +A(fs) = df s +f(s) +fA(s)
= df s +f(+A)s
so +A is a connection.
Now, let
t
be any connection. For f C
(M) and s (E) we see

(
t
)(fs) = (df s +f
t
s) (df s +fs) = f(
t
)s
i.e.
t
is a C
(M)-linear map (E)

1
(M, E). Thus it corresponds
to a bundle map E T
K
M E which again corresponds to a smooth section
of the bundle Hom(E, T
K
M E)

= E
K
M E

= T
K
M End(E). Thus
t

1
(M, End(E)).
Observe that connections are local operators i.e. if the section s vanishes on
an open set U, then s also vanishes on U. In other words decreases support.
To see this let s be a section which is 0 on U. Let p U be arbitrary and let
f be a smooth bump function with supp f U and which is identically 1 in a
small neighborhood of p. Then fs is the zero section and hence (fs) = 0 and
therefore
(s)
p
= f(p)(s)
p
= (fs)
p
(df s)
p
= 0.
Another way of formulating locality is that (s)
p
depends on s only in an
arbitrarily small neighborhood around p. Therefore if U is any open set in M
and E[
U
is the restriction of E to U it makes sense to restrict from E to E[
U
.
Denition 10.3 (Connection 1-form). Given a connection on the trivial
bundle, dene the connection 1-forms
ij
to be the unique dierential 1-forms
satisfying
s
j
=
n
i=1
ij
s
i
.
They are smooth for if X is a smooth vector eld then
X
s
j
=
n
i=1
ij
(X)s
i
is a smooth section of E, i.e.
ij
(X) are smooth functions and hence
ij
are
smooth 1-forms. Conversely, if
ij
is a matrix of 1-forms, then v
j
=
n
i=1
ij
s
i
are sections of T
K
ME and hence determine a connection by (10.2). Thus, given
a local frame, there is a 1-1 correspondence between connections on E[
U
and
matrices (
ij
) of 1-forms over U.
For the trivial connection on a product bundle, the connection forms relative
to the canonical global frame are indeed trivial: By denition s
j
= d(1) s
j
=
0, so
ij
= 0.
The next result describes how the connection 1-forms react to a change of
trivialization.
Lemma 10.4. Let (U
) and (U
) be two trivializations of E with non-

trivial overlap, and let g
: U
GL(n, K) be the transition function. If
and
denote the matrices of connection 1-forms on U
and U
respectively,
then for any p U
(p) = g
1
(p)
(p)g
(p) +g
1
(p)dg
(p).
where dg
is the entrywise exterior derivative.

Proof. First, we know that the trivializations
and
yield smooth local

frames (s
i
) and (t
i
) on U
and U
respectively by
s
i
(p) =
1
(p, e
i
), t
i
(p) =
1
(p, e
i
)
where (e
i
) denotes the standard basis for K
n
. If p U
then
t
j
(p) =
1
(p, e
j
) =
1
(p, e
j
) =
1
(p, g
(p)e
j
)
=
n
i=1
(g
)
ij
(p)
1
(p, e
i
),
in other words t
j
=
n
i=1
(g
)
ij
s
i
. We act on this by
X
and get
n
k=1
(
)
kj
(X)t
k
=
n
i=1
X
((g
)
ij
s
i
)
=
n
i=1
d(g
)
ij
(X)s
i
+
n
i,k=1
(g
)
kj
(
)
ik
(X)s
i
.
Now we plug in the expression for t
k
in the left hand side and get
n
i,k=1
(
)
kj
(X)(g
)
ik
s
i
=
n
i=1
d(g
)
ij
(X)s
i
+
n
i,k=1
(g
)
kj
(
)
ik
(X)s
i
.
Comparing the coecient to s
i
we get the equation
n
k=1
(g
)
ik
(
)
kj
(X) = d(g
)
ij
(X) +
n
k=1
(
)
ik
(X)(g
)
kj
which exactly states that g
= dg
. From this the desired formula

is an immediate consequence.
10.2 Connections on Associated Vector Bundles*
From principal G-bundles we can construct certain associated vector bundles
of great importance, particularly in spin geometry (spinor bundles). In this sec-
tion, which may be skipped at a rst reading, we will investigate how connections
on the principal bundle give rise to connections on the associated vector bundles.
First, however, we give a detailed outline of the construction of the associated
bundles:
10.2 Connections on Associated Vector Bundles* 167
Lemma 10.5. Consider a smooth G-principal bundle (P, , ) over a manifold
M, and assume that G is acting smoothly from the left on a manifold F. Con-
sider the following smooth right action of G on P F by (x, ) g = (x g, g
1
),
put P
G
F := (P F)/G the orbit space under this action, and dene a map
G
: P
G
F M by
G
([p, ]) = (p). Then (P
G
F,
G
) has the structure
of a smooth ber bundle over M with ber F.
Proof. It is easy to see, that the action mentioned in the theorem really is a
right action, and it is smooth since the coordinate maps ((p, ), g) p g and
((p, ), g) g are smooth, by assumption.
We then consider the orbit space P
G
F = (P F)/G where accordingly
two points (p
1
,
1
) and (p
2
,
2
) are identied if there is a g with (p
2
,
2
) =
(p
1
g, g
1

2
). The equivalence class containing (p, ) we denote by [p, ].
Dene a map q : P F P
G
F simply by
(p, )
q
[p, ],
and equip P
G
F with the quotient topology w.r.t. q. Now, P
G
F is a
topological space and we show that it has the structure of a topological ber
bundle over M.
To do this we need a projection
G
from P
G
F to M. We dene this map
on an equivalence class by
G
([p, ]) = (p).
This is well-dened, since any other element of [p, ] has the form (p g, g
1
)
and
G
([p g, g
1
]) = (p g) = (p).
To ensure continuity of
G
we remark that i renders the following diagram
commutative ( here, is the obvious extension of from the ber bundle, by
(p, ) = (p), it should not cause confusion to use the same notation for both
maps)
P F
P
P
P
P
P
P
P
P
P
P
P
P
P
P
q
P
G
F
M
q is a quotient map, is continuous, and thus by the characteristic property of
quotient maps,
G
is continuous.
The next thing we need to do is to construct trivializations. In order to do
so, we need to remark the following intermediate result
1
G
(x) = [p, ]
F .
for some xed p
1
(x). The inclusion is clear, since
G
([p, ]) = (p) = x.
Conversely, any element in P
G
F is of the form [p, ]. Assume, that [p, ] is
in
1
G
(x), then x =
G
([p, ]) = (p), so follows. From this it is easily seen
that
1
G
(V ) = [p, ]
(p) V, F . (10.3)
Back on track, we shall now construct trivializations. Fix a point x
0
M,
pick a trivializing neighborhood V for the principal bundle around x
0
, let

:
1
(V ) V G be the trivialization of the principal bundle, and let s :
V
1
(V ), s(x) =
1
(x, e) be the canonical local section associated with
the trivialization. Dene a map : V F
1
G
(V ) by
(x, ) = [s(x), ].
We will eventually see that this is the inverse of the trivialization we want.
To see that is injective, assume that (x, ) =

(x
t
,
t
), i.e. [s(x), ] =
[s(x
t
),
t
], which means that there exists a g such that s(x) = s(x
t
) g and
= g
1
. But then
x = (s(x)) = (s(x
t
) g) = (s(x
t
)) = x
t
,
and since the G-action on P is free, g = e, so (x, ) = (x
t
,
t
).
To prove surjectivity, let [p, ]
1
G
(V ) be arbitrary (every element of
1
G
(V )
is of this form by (10.3)) and put x = (p). As both p and s(x) are in the
same ber (namely
1
(x)), we have p = s(x) g for some g. Accordingly
[s(x) g, g
1

t
] = [s(x),
t
] for all
t
F. For the choice
t
= g , we get
(x, g ) = [s(x), g
t
] = [s(x) g, g
1
g ] = [p, ],
thus proving surjectivity and hence bijectivity.
is continuous, since its just the composition (x, ) (s(x), ) [s(x), ].
We denote the inverse by :
1
G
(V ) V F. To see that this is continuous,
consider this commutative diagram
P F q
1
(
1
G
(V ))
q
T
T
T
T
T
T
T
T
T
T
T
T
T
T
T
T
T
P
G
F
1
G
(V )
V F
Then q on q
1
(
1
G
(V )) is given by (p, )
q
[p, ]

((p), g ) which is
continuous. Again by the property of quotient maps, is continuous, and hence
the homeomorphism we want. In short, we have shown that a trivialization
(V,
) for the principal bundle renders a trivialization (V, ) for P

G
F.
Up till now we have veried that (P
G
F,
G
) is a topological ber bundle
over M. Our nal goal is to show that it is actually a smooth bundle, i.e. to equip
P
G
F with a smooth structure such that
G
is smooth and the trivializations
are dieomorphisms. To this end consider a trivialization cover V
for
the principal bundle. This gives us a trivialization cover V
for P
G
F.
Now, we simply declare P
G
F to have the unique smooth structure which
makes all
:
1
G
(V
) V
F dieomorphisms (V
i
F has a natural
smooth structure as an open subset op the manifold P F). In order for this
to be a well-dened smooth structure on P
G
F we only need to show that
: (V
)F (V
)F is smooth (it then follows automatically

that it is a dieomorphism).
(x, ) =
([s
(x), ]) =
([s
(x) g
(x), ])
=
([s
(x) g
(x), g
(x)
1
g
(x) ])
=
([s
(x), g
(x)])
= (x, g
(x) ).
This map is smooth. Hence P
G
F has a well-dened smooth structure.
The only thing left is to show, that
G
is smooth. We do that by showing that
G

1
: V F
1
G
(V ) is smooth for any trivialization . We observe
G

1
(x, ) =
G
([s(x), ]) = (s(x)) = x.
This it smooth, and since is a dieomorphism, also
G
is smooth.
The ber bundle whose existence is guaranteed by the preceding lemma is
called the associated ber bundle to the principal bundle (P, , ). This associ-
ated bundle depends, of course, on the choice of manifold F and on the G-action
on F.
If we instead of a mere group action on some manifold F have a representation
of a vector space, then we get a vector bundle
Lemma 10.6. Let (P, , ) be a smooth principal G-bundle, V a nite-dimensional
vector space and : G Aut(V ) a Lie group representation. Then the associ-
ated ber bundle (P
V,
) has the structure of a smooth vector bundle over

M.
Proof. As mentioned in the proof above, the bers are of the form
(x) = [p, v] [ v V
for some xed p
1
(x). These have natural vector space structures
a[p, v
1
] +b[p, v
2
] = [p, a
1
+bv
2
]
which turn the bers into dimV -dimensional vector spaces. Now, lets consider
restricted to the ber
1
(x) mapping injectively into x V , which has

the vector space structure a(x, v
1
) + b(x, v
2
) = (x, av
1
+ bv
2
) turning it into a
dimV -dimensional vector space as well. There is a g such that s(x) = p g
1
and so
[
(x)
(a[p, v
1
] +b[p, v
2
]) = ([p, av
1
+bv
2
]) = ([p g
1
, (g)(av
1
+bv
2
)])
= ([s(x), a(g)v
1
+b(g)v
2
])
= (x, a(g)v
1
+b(g)v
2
) = a(x, (g)v
1
) +b(x, (g)v
2
)
= a([s(x), (g)v
1
]) +b([s(x), (g)v
2
])
= a([p, v
1
]) +b([p, v
2
]).
restricted to the bers is accordingly a linear map and hence a vector space
isomorphism
1
(x)

x V . This proves the assertion.
As is apparent, the bers in the associated vector bundle can be a bit tricky
to handle. Thus it may cause some problems to work with sections of associated
vector bundles. Therefore it is a relieve, that we have alternative ways to deal
with such. If we still want to work globally, we can think of them as certain
functions on the total space of the principal bundle. Recall that a function f :
P V is called equivariant w.r.t. the representation if f(p g) = (g
1
)f(p).
Proposition 10.7. To each section (P
V ) corresponds an equivariant
function

: P V , given by (x) = [p,

(p)] when
(p) = x, and this

correspondence between smooth sections and equivariant functions is bijective.
Proof. To be a bit more precise on the denition of

: let p P, then we dene
(p) to be the unique element in V such that [p,

(p)] = (
(p)). Equivariance
of this function is seen as follows:
[p g,

(p g)] = [p, (g)
(p g)].
On the other hand
[p g,

(p g)] = ((p g)) = ((p)) = [p,

(p)].
Thus, by uniqueness, we get (g)
(p g) =

(p) and hence equivariance.
To see that it is smooth, let

be a trivialization of P over U, let s be the
associated local section, i.e. s(x) =

1
(x, e) and let be the corresponding
trivialization of the associated vector bundle. We note that
(x) = [s(x),

(s(x))]
and hence that ((x)) = (x,

(s(x))). Thus the map x

(s(x)) is smooth.
But then also

1
: U G V given by
(x, g)

(s(x) g) = (g
1
)
(s(x))
is smooth, and hence

is smooth.
Given a smooth equivariant function f we dene a corresponding section of
P
V by
f
(x) = [s(x), f(s(x))]. It is easy to see that these two operations
are inverses of each other.
Another description of the section is the local one. Observe the elementary
fact that for any vector bundle E with trivializations (U
)
I
and transition
functions g
: U
GL(n, K) there is a 1-1 correspondence between

smooth sections of E and collections (
)
I
of smooth functions
: U

1
n
satisfying
(x) = g
(x)
(x) for x U
: Given a section (E)

the obvious denition would be
(x) = (x,
(x)), i.e.
(x) = pr
2
((x))
which is obviously smooth. Since
(x) = pr
2
((x)) = pr
2

((x))
= pr
2
(
(x,
(x)))
= pr
2
(x, g
(x)
(x)) = g
(x)
(x).
Conversely, given such a collection (
) we dene a section by
(x) =
(x,
(x)). A calculation as the one above will reveal that this is well-dened.
Now, let E = P
V be an associated vector bundle and (E) a smooth

section. Furthermore, let (U
) be a set of trivializations of P and (U
)
the corresponding trivializations of E. Relative to these we have the local func-
tions
(x) = (pr
2
)((x))
Since
([s
(x), v]) = (x, v) we see that locally (x) = [s
(x),
(x)]. Moreover
we can express
in terms of the associated equivariant function

Proposition 10.8. For any (E) we have
(x) =

(s
(x))
Proof. Since (x) = [s
(x),

(s
(x))] on U
, the result follows immediately

by the calculations above.
Assume now that our principal G-bundle : P M has a connection, i.e.
a choice of horizontal subspace H
p
P T
p
P for each p P, or equivalently
a connection 1-form . Given a connection, the push-forward
: T
p
P
T
(p)
M restricts to an isomorphism H
p
P

T
(p)
M. Any vector eld X on
M can be lifted to a unique horizontal vector eld X on P. More specically,
X
p
=
1
(X
(p)
).
The rst step in constructing a connection on the associated bundle E =
P
V is the following lemma

Lemma 10.9. Let (E) be a section of the associated bundle, then for each
X X(M) the map P V given by p X
p
(
) is -equivariant.
Proof. This is an elementary calculation:
X
pg
(
) = (
g
)
X
p
(
) = X
p
(

g
)
= X
p
((g
1
)

) = (g
1
)X
p
(
)
the third identity being equivariance of

.
Theorem 10.10 (Connection on Associated Bundle). The map : X(M)
(E) (E) given by (X, )
X
where
X
is the section correspond-
ing to the equivariant function p X
p
(
) denes a connection on E. In short
X
(p) = X
p
(
). (10.4)
Since we spent some time discussing local expressions of sections, it is a nat-
ural question to pose: how does
X
look locally? The answer is given in the
following proposition
Proposition 10.11. Given a local section s
: U
P of the principal
bundle, then the local function of
X
relative to the corresponding trivialization
of E is given by
(
X
)
(x) = X
x
(A
(X
x
))
(x) (10.5)
where
: g End(V ) is the induced Lie algebra representation of and

A
= s
is the local gauge potential.

We can phrase this in a bit dierent way, when we view a connection as a
map (E) (T
M E). If (E) is a section of E and
are the local

functions, then is a section of (T
M E) whose local functions are given

by
()
(x) = d
(x)
(A
x
)
(x). (10.6)
This formula has to be interpreted correctly. To x U
corresponds an ele-
ment in T
p
M V

= 1
k
1
n
(if dimM = k and dimV = m), d should be
understood componentwise, i.e. pick a basis for V and apply d to the compo-
nents relative to this basis, and
(A
x
)
should act on a vector X

x
T
x
M by
(A
x
(X
x
))
(x). Letting 10.6 act on a vector eld X we thus get 10.5.

Lets end with a prominent example of an associated connection, namely
the Levi-Civita connection. Let M be an n-dimensional pseudo-Riemannian
manifold, i.e. a manifold with a metric g of signature (r, s). Then we can consider
the orthonormal frame bundle : F
O
(M) M. This is a principal O(r, s)-
bundle on M. It carries what we will call the canonical 1-form , an 1
n
-valued
1-form on P
O
(M), dened in the following way: let X
p
T
p
P
O
(M), where
p P
O
(M) is an isometric isomorphism p : 1
n
T
(p)M, then
(X
p
) = p
1
(
X
p
).
One can check that relative to the dening representation : O(r, s)
GL(n, 1) the 1-form satises
((
A
)
X
p
) = (A
1
)(X
p
)
and that it is zero on vertical vectors (this is obvious, since vertical vectors, by
denition, satisfy
X
p
= 0). Thus
1
(P
O
(M), 1
n
).
Given a connection on P
O
(M) we have a covariant exterior derivative D
1
(P
O
(M), 1
n
)
2
(P
O
(M), 1
n
) and the 2-form
:= D
is called the
torsion of the connection.
Theorem 10.12 (Fundamental Theorem of Riemannian Geometry I).
There exists a unique connection , called the Levi-Civita connection or the
Riemannian connection, on P
O
(M) with vanishing torsion 2-form.
We will not prove this theorem. For a proof consult [Bleecker] Theorem 6.2.5.
Exactly the same holds true in the case of an oriented pseudo-Riemannian man-
ifold where we replace the orthonormal frame bundle P
O
(M) with the oriented
orthonormal frame bundle P
SO
(M).
We can retrieve the tangent bundle TM from the frame bundle, it is simply the
vector bundle associated to the dening representation : O(r, s) GL(n, 1),
i.e.
TM = P
O
(M)
1
n
and the Levi-Civita connection induces on TM a connection which is also
called the Levi-Civita or Riemannian connection. If s
: U
P
O
(M) is some
local section and A
= s
is the gauge potential of the Levi-Civita connection,

formula (10.5) above gives us the following local formula for the connection on
TM:
(
Y
X)
(x) = X
x
Y
iI
A
i
(X
x
)B
i
_
Y
(x) (10.7)
where (B
i
)
iI
is some basis for the Lie algebra o(r, s).
10.3 Pullback Bundles and Pullback Connections
Assume that f : N M is a smooth map and : E M is a vector bundle.
Recall from Chapter 9 the pullback bundle:
f
E = (p, v) N E [f(p) = (v)

over N. f induces a pullback map on the sections f
: (E) (f
E) by
(f
s)(p) = (p, s(f(p))).

If we turn C
(N) into a C
(M)-module by dening scalar multiplication

:= f
() (where f
just means f) we can prove the following

Lemma 10.13. The map : C
(N)
C
(M)
(E) (f
E) given by the
C
(N)-linear extension of
s f
(s)
is an isomorphism of C
(N)-modules.
Proof. It is easy to see that it is well-dened: if C
(M), then
( s) = f
(s) = (f
)f
s = ( )f
s = ( s).
We prove lemma rst in the case where E is the trivial bundle, E

= MK
n
.
This being the case, f
E is also trivial: f
E

= N K
n
. Thus, sections of these
bundles are just smooth maps M K
n
or N K
n
. Let s
i
: M K
n
denote the section p e
i
of E, mapping p to the ith standard basis vector,
and t
i
: N K
n
denote the corresponding section of f
E. It is easy to see
that f
s
i
= t
i
.
Consider the section s =
1
t
1
+ +
n
t
n
(f
E), then
s =
1
f
s
1
+ +
n
f
s
n
= (
1
s
1
+ +
n
s
n
),
10.3 Pullback Bundles and Pullback Connections 173
i.e. is surjective.
To see that it is injective, assume ( s) = 0. As before, s is of the form
1
s
1
+ +
n
s
n
, so that
( (
1
s
1
+ +
n
s
n
)) = ( (
1
s
1
) + (
n
s
n
))
= ((f
1
) s
1
) + + ((f
n
) s
n
)
= ((f
1
), , (f
n
)).
Thus we must have (f
1
) = = (f
n
) = 0. Therefore also
s = (f
1
) s
1
+ + (f
n
) s
n
= 0,
i.e. is injective.
In the general case: Let F be a bundle such that E F is trivial
1
. Then the
result holds here, and by restriction, also on E E F.
From the pullback map of ordinary k-forms we get a pullback map f
k
(M, E)
k
(N, f
E) by a linear extension of
f
( s) = f
s.
One can check that this is well-dened.
Proposition 10.14. Given a connection on E there is a unique connection
t
on f
E rendering the following diagram commutative

(E)

f
1
(M, E)
f
(f
E)

1
(N, f
E)
This connection, denoted f
, is called the pullback connection on f
E.
Proof. Uniqueness rst. As shown above any section of f
E is a linear com-
bination of elements of the form f
(s) where s (E) and C
(N).
Assume we have two connections
t
and
tt
making the diagram commutative,
i.e.
t
(f
s) = f
(s) =
tt
(f
s). On elements of the form f
(s) we have
t
(f
s) = d f
s +
t
(f
s) = d f
s +
tt
(f
s)
=
tt
(f
s).
Thus, the two connections are equal.
To show existence we simply construct a connection tting into the diagram.
Lemma 10.13 gave us an isomorphism : C
(N)
C
(M)
(E) (f
E).
Thus we have
k
(N, f
E) =
k
(N)
C
(N)
(f
E)
=
k
(N)
C
(N)
C
(N)
C
(M)
(E)
=
k
(N)
C
(M)
(E).
This isomorphism we call
k
. It is not hard to see that the inverse is given by
1
k
( s) = f
s.
1
For smooth vector bundles such a complementary bundle exists over any manifold, com-
pact or not, cf. [MT] Exercise 15.10.
As in Lemma 10.13 the pullback map on forms f
:
k
(M)
k
(N) gives
a C
(M)-linear map
C
(N)
C
(M)

k
(M)
k
(N)
upon dening f
. This is no longer an isomorphism. Tensoring

(over the ring C
(M)) with (E) gives a map

C
(N)
C
(M)

k
(M)
C
(M)
(E)
k
(N)
C
(M)
(E)
(the map on the second factor is, of course, just the identity). Composing this
with the isomorphism
C
(N)
C
(M)

k
(M, E)

C
(N)
C
(M)

k
(M)
C
(M)
(E)
yields the map
: C
(N)
C
(M)

k
(M, E)
k
(N)
C
(M)
(E)
which is given explicitly by ( s) (f
) s.
Now we dene f
: (f
E)
1
(N, f
E) by
f
=
1
1
(d id +(id ))
1
. (10.8)
To check that it ts into the diagram, note that
1
(f
s) = 1 s, and
assuming s to be of the form s
t
we get
f
(f
s) =
1
1
(d id +(id ))(1 s)
=
1
1
(d(1) s) +
1
1
(1 s
t
)
=
1
1
(f
s
t
) = f
s
t
= f
( s
t
)
= f
(s).
Finally we need to check that it is indeed a connection. Obviously it is K-
linear. To see that it satises (10.1), let C
(N) and t (f
E). By Lemma
10.13 we may assume that t =
t
f
s, where
t
C
(N). Combining
t
into
a single function we only need to check (10.1) on sections of the form f
s
corresponding, under the isomorphism to s. Still assuming s to be of
the form s
t
we calculate
f
()(f
s) =
1
1
(d id +(id ))( s)
=
1
1
(d s +( s
t
)) = d f
s +
1
1
((f
) s
t
)
= d f
s +f
s
t
,
and since
f
s
t
= f
( s
t
) = f
(s) = f
(f
s)
(the last equality following from the commuting diagram) we see that f
satises (10.1).
Proposition 10.15. Let E be a vector bundle over M and a connection.
1) We have id
M
= .
2) For maps f : N M and g : K N we have, under the isomorphism
g
E

= (fg)
E, that g
= (fg)
.
3) If (
ij
) are the connection forms relative to some frame (s
i
) over a trivi-
alization U M, then (f
ij
) are the connection forms for f
relative
to the frame (f
s
i
) over f
1
(U).
10.4 Curvature 175
Proof. Point 1) follows trivially from the uniqueness part of Proposition 10.14.
For part 2) consider the following diagram
(E)

g
1
(M, E)
g
(g
E)
g
1
(K, g
E)
((fg)
E)
(fg)

1
(K, (fg)
E)
where and are the natural isomorphisms, induced by the bundle isomor-
phism g
E

= (fg)
E. The upper square is commutative and so is the outer

square. By uniqueness, this forces
g
=
1
((fg)
),
and this is exactly what it means for g
and (fg)
to be equal under the

isomorphism.
3) For this part we can assume E to be trivial and the frame (s
i
) to be global.
If E = M K
n
then also f
E = N K
n
. Lets abbreviate s
t
j
= f
s
j
. This is
easily seen to be a global frame for f
E. Then we see that

(f
)s
t
j
= (f
)(f
s
j
) = f
(s
j
)
= f
_
n
i=1
ij
s
i
_
=
n
i=1
(f
ij
) (f
s
i
)
=
n
i=1
(f
ij
) s
t
i
which is precisely what we wanted.
10.4 Curvature
We dene a wedge product :
k
(M)
l
(M, E)
k+l
(M, E) by
( s) := ( ) s.
In general we would need some kind of bilinear map on E in order to dene
a wedge product
k
(M, E)
l
(M, E)
k+l
(M, E). This is for instance
the case for the bundle End(E), where we have composition, or for the trivial
bundle MM(n, K) where we have matrix multiplication. Specically, if , are
dierential forms with values in M(n, K), they are matrices of usual dierential
forms, and the wedge product is dened to be the matrix of forms whose ijth
entry is given by
( )
ij
:=
n
k=1
ik

kj
.
The most obvious example at hand of a matrix-valued form is of course the
matrix of connection 1-forms.
For the spaces
k
(M) of K-valued forms on M we have the usual exterior
derivative d :
k
(M)
k+1
(M). We want to construct something similar for
k
(M, E):
Denition 10.16 (Covariant Derivative). Let be a connection on the
vector bundle E. For k 1 we dene a map

k
:
k
(M, E)
k+1
(M, E),
called the covariant exterior derivative, given on generators by
k
( s) = d s + (1)
k
s.
This is well-dened for on one hand
k
(f s) = d(f) s + (1)
k
f s
= (df +fd) s + (1)
k
fs (10.9)
and on the other
k
( fs) = d fs + (1)
k
(fs)
= d fs + (1)
k
(df s + (1)
k
fs)
= fd s + (1)
k
( df) s + (1)
k
fs
= (fd +df ) s + (1)
k
fs.
Thus the two sides are equal and from (10.9) we read o:
Lemma 10.17. The covariant exterior derivative

k
satises the Leibniz rule,
i.e. if f C
(M) and
k
(M, E) then:
k
(f) = df +f
k
(). (10.10)
The important map is

1
which we abbreviate to

.
Denition 10.18 (Curvature). Let E be a K-vector bundle and a connec-
tion. The K-linear map
R
:=

: (E)
2
(M, E)
is called the curvature of the connection. A connection with vanishing curvature
is called a at connection.
First we observe, that R
(or just R for brevity) is C
(M)-linear:
R(fs) =

(fs) =

(df s +fs)
= d(df) s df s +df s +f
s = fR(s).
Thus R induces a bundle map E
2
(T
K
M) E which again corresponds to
a section of the bundle E

2
(T
K
M) E

=
2
(T
K
M) End(E). Therefore
we can think of R either as a map (E)
2
(M, E) or we can think of it as
an End(E)-valued 2-form on M. If we adopt this last view of the curvature, we
write R
X,Y
(s) for the smooth section of E which we get by acting on s with the
vector elds X and Y .
For the trivial connection on a product bundle, the curvature is identically
0, indeed, since R is C
(M)-linear and Rs
i
=

(s
i
) =

(d(1) s
i
) = 0:
R(f
1
s
1
+ +f
n
s
n
) = f
1
R(s
1
) + +f
n
R(s
n
) = 0.
Thus the trivial connection is a at connection.
Next we investigate how the curvature looks locally
10.4 Curvature 177
Proposition 10.19. Let U be a trivialization neighborhood for E and
ij
the
connection 1-forms for relative to some local frame s
1
, . . . , s
n
. Put
ij
:=
d
ij
+
n
k=1
ik

kj
(or in compact form = d + ). Then
Rs
j
=
n
i=1
ij
s
i
.
The 2-forms
ij
are called the curvature 2-forms.
Proof. This hangs on the fact that s
j
=
ij
s
i
so that
Rs
j
=

(s
j
) =

_
n
i=1
ij
s
i
_
=
n
i=1
(
ij
s
i
)
=
n
i=1
(d
ij
s
i
ij
s
i
) =
n
i=1
_
d
ij
s
i
ij
(
n
k=1
ki
s
k
)
_
=
n
i=1
d
ij
s
i
i,k=1
ij
(
ki
s
k
)
=
n
i=1
d
ij
s
i
+
n
i=1
( )
ij
s
i
=
n
j=1
ij
s
i
.
And then globally:
Proposition 10.20. For vector elds X and Y and s (E) we have
R
X,Y
(s) = (
X
Y

Y
X

[X,Y ]
)s.
Proof. We will show it locally so let U be a trivialization for E and and
the matrices of connection 1-forms and curvature 2-forms relative to the
corresponding local frame s
1
, . . . , s
n
. Then
Y
s
j
=
X
_
n
i=1
ij
(Y )s
i
_
=
n
i=1
_
X(
ij
(Y ))s
i
+
ij
(Y )
X
s
i
_
=
n
i,k=1
ij
(Y )
ki
(X)s
k
+
n
i=1
X(
ij
(Y ))s
i
.
And therefore
(
X
Y

Y
X

[X,Y ]
)s
j
=
n
i=1
_
X(
ij
(Y )) Y (
ij
(X))
ij
([X, Y ])
_
s
i
+
n
i,k=1
_
kj
(Y )
ik
(X)
kj
(X)
ik
(Y )
_
s
i
=
n
i=1
d
ij
(X, Y )s
i
+
n
i,k=1
(
ik

kj
)(X, Y )s
i
=
n
i=1
ij
(X, Y )s
i
= R
X,Y
(s
j
),
and by C
(M)-linearity of R
X,Y
the result follows.
As for the connection forms, we also have a transition rule for the curvature
forms
Lemma 10.21. Let (U
) and (U
) be two trivializations of E with non-

trivial overlap, and let g
: U
GL(n, K) be the transition function.

If
and
denote the curvature 2-forms on U
and U
respectively then for

p U
(p) = g
1
(p)
(p)g
(p).
Proof. From Lemma 10.4 we know that the connection 1-forms are related by
(p) = g
1
(p)
(p)g
(p) + g
1
(p)dg
(p). For brevity we write g instead of

g
. By dg we understand the entry-wise exterior dierentiation of the matrix

of functions. Since g
1
g = I we get d(g
1
)g + g
1
dg = 0 and hence d(g
1
) =
g
1
(dg)g
1
. Inserting in the denition of
we get
= d
= d(g
1
g +g
1
dg) + (g
1
g +g
1
dg) (g
1
g +g
1
dg)
= (g
1
(dg)g
1
)
g +g
1
(d
)g g
1
dg (g
1
(dg)g
1
) dg
+ (g
1
g +g
1
dg) (g
1
g +g
1
dg)
= g
1
d
g +g
1
(
)g = g
1
g.
10.5 Metric Connections
Before we embark on the construction of characteristic classes we will in this
section briey discuss some of the interplay between metrics and connections on
vector bundles. Recall the denition of a metric or ber metric: It is a choice of
inner product (i.e. a conjugate-symmetric, positive denit sesquilinear form) g
p
on each of the bers varying smoothly in sense that if s, s
t
are smooth sections of
E, then p g
p
(s
p
, s
t
p
) is a smooth map. We will use the notation g or , for
such a metric. A vector bundle endowed with a metric is called a Riemannian
vector bundle.
Denition 10.22 (Metric Connection). Let E be a vector bundle equipped
with a smooth ber metric. A connection is said to be compatible with the metric
or a metric connection if for all X X(M) and s, s
t
(E) the following
equation is satised:
Xs, s
t
= s(X), s
t
+s, s
t
(X). (10.11)
Proposition 10.23. A metric connection always exists on a Riemannian vector
bundle.
Proof. We only need to show that a metric connection exists on a trivial
bundle for then a partition of unity subordinate to a trivialization cover of M
will provide the rest of the argument.
So let E = M K
n
be the trivial bundle with the metric coming from the
usual inner product on K
n
and let e
1
, . . . , e
n
be the standard basis for K
n
and consider the associated global orthonormal frame s
i
(p) = (p, e
i
). We show
that the trivial connection is metric: For sections
s =
n
i=1
a
i
s
i
and s
t
=
n
i=1
b
i
s
i
where a
i
, b
i
C
(M) we see that s, s

t
=
n
i=1
a
i
b
i
so that
Xs, s
t
=
n
i=1
X(a
i
b
i
) =
n
i=1
(Xa
i
)b
i
+
n
i=1
a
i
(Xb
i
)
whereas
s(X), s
t
=
_
n
i=1
(Xa
i
)s
i
,
n
j=1
b
j
s
j
_
=
n
i=1
(Xa
i
)b
i
and similarly s, s
t
(X) =
n
i=1
a
i
(Xb
i
). Thus the trivial connection is a met-
ric connection.
10.5 Metric Connections 179
If E is a Riemannian vector bundle and U is a trivialization neighborhood, we
can always assume the local sections over U to be orthonormal. If they are not,
we simply apply the Gram-Schmidt procedure to obtain the desired orthonormal
frame.
The impact on connection forms of a connection being metric is seen in the
following (where by
ij
we understand the complex 1-form, acting on vector
elds by
ij
(X) :=
ij
(X)).
Lemma 10.24. Let be a connection on the Riemannian bundle E. Then is
metric if and only if for any local orthonormal frame (s
i
) the associated matrix
of connection 1-forms (
ij
) is skew-adjoint, i.e.
ij
+
ji
= 0.
Furthermore the matrix of curvature forms of a metric connection associated
with an orthonormal frame is skew-adjoint.
In the case of a real bundle, the matrices will, of course, turn out to be anti-
symmetric.
Proof. Assume rst, that the connection is metric. We have s
i
, s
j
=
ij
and
thus
0 = Xs
i
, s
j
= s
i
(X), s
j
+s
i
, s
j
(X)
=
_
n
k=1
ki
(X)s
k
, s
j
_
+
_
s
i
,
n
k=1
kj
(X)s
k
_
=
ji
(X) +
ij
(X).
Thus they are skew-adjoint.
Conversely, assume that all matrices of connection 1-forms are skew-adjoint.
We want to show that (10.11) is satised. We do it locally, so let U be a triv-
ialization neighborhood and let s
1
, . . . , s
n
be a local orthonormal frame and
(
ij
) the connection 1-forms, which are by assumption skew-symmetric. It suf-
ces to show it for the local sections s
i
and s
j
. Since they are orthonormal we
have Xs
i
, s
j
= 0. On the right-hand side
s
i
(X), s
j
+s
i
, s
j
(X) =
_
n
k=1
ki
(X)s
k
, s
j
_
+
_
s
i
,
n
=1
j
(X)s
_
=
n
k=1
ki
(X)s
k
, s
j
+
n
=1
j
(X)s
i
, s
= (
ji
+
ij
)(X) = 0.
Thus the two sides are equal.
The last assertion follows from the simple calculation
ij
= d
ij
+
n
k=1
ik

kj
= d
ji
+
n
k=1
ki

jk
= d
ji
k=1
jk

ki
=
ji
.
One last bit of information on metric connections that will be needed: how do
metric connections react to pullbacks? First of all, if E is a vector bundle over
M with metric g, and f : N M is smooth, we can pull g back to a metric
f
g on f
E in the following way: Bearing in mind that sections t, t

t
of f
E are
of the form t(p) = (p, (p)) and t
t
(p) = (p,
t
(p)), where ,
t
: M E are
some functions satisfying f(p) = ((p)), we can dene the pullback by
(f
g)(t, t
t
)
p
= g
f(p)
((p),
t
(p)).
This is well-dened since (p) and
t
(p) are in the ber over f(p), and it is
easily seen to be sesquilinear, conjugate symmetric and positive denit, and
since the right-hand side depends smoothly on p, f
g is a ber metric on f
E.
Since (f
s)(p) = (p, s(f(p))) we have in particular that

(f
g)
p
(f
s, f
s
t
) = g
f(p)
(s(f(p)), s
t
(f(p))). (10.12)
Thus, for instance, if s
1
, . . . , s
n
is a local orthonormal frame w.r.t. g, then
f
s
1
, . . . , f
s
n
is orthonormal w.r.t. f
g.
Lemma 10.25. Let (E, g) be a Riemannian vector bundle and a metric
connection on E. Then f
is metric relative to the pullback metric f
g.
Proof. Let s
1
, . . . , s
n
be a local orthonormal frame for E, then by the
only-if-part of Lemma 10.24 the matrix of connection 1-forms is skew-adjoint.
f
s
1
, . . . , f
s
n
is a local orthonormal frame for f
E relative to f
g and by
Proposition 10.15 3) f
ij
are the connection 1-forms for f
relative to this
frame. But we easily see that f
ij
= f
ji
and thus by the if-part of Lemma
10.24 the connection f
is metric.
On the tangent bundle of a Riemannian manifold the situation is quite del-
icate. Not only do metric connections exists but we can in fact single out one
particular connection if we impose one further condition, namely symmetry. Let
be a connection on TM and dene the torsion tensor to be the covariant
2-tensor given by
(X, Y ) =
X
Y
Y
X [X, Y ].
The connection is said to be symmetric or torsion-free if the torsion tensor
vanishes identically, i.e. if
X
Y =
Y
X + [X, Y ].
Theorem 10.26 (Fundamental Theorem of Riemannian Geometry II).
On the tangent bundle of a Riemannian manifold there exists a unique con-
nection, called the Levi-Civita connection or Riemannian connection, which is
symmetric and compatible with the metric.
The proof of this important fact can be found in any textbook on Riemannian
geometry, for instance [Lee].
10.6 Characteristic Classes
Denition 10.27 (Invariant Polynomial). Let G be a matrix Lie group and
g its Lie algebra. A G-invariant polynomial is a map P : g K which is
polynomial in the entries of the elements of g and which satises P(AXA
1
)
for all X g and A G. The set of G-invariant polynomials is denoted I(G) and
the set of G-invariant polynomials of homogenous degree k is denoted I
k
(G).
Clearly we must have
I(G) =
n
k=1
I
k
(G).
In this section our interest we primarily be at GL(n, K)-invariant polynomials.
We have some obvious examples of such polynomials, the simplest being the
trace and the determinant. They are homogenous of degree 1 and n respectively.
Another example is
t
i
(X) := Tr X
i
as well as the following: dene for k = 1, . . . , n the map
k
: M
n
(K) K by
det(I +X) = 1 +
1
(X) +
2
2
(X) + +
n
n
(X).
10.6 Characteristic Classes 181
Its not hard to check that
k
is in fact a GL(n, K)-invariant polynomial of
homogenous degree k. Note that
n
(X) = det X.
Surprisingly enough, the t
k
s and
k
s generate the set of GL(n, K)-invariant
polynomials:
Theorem 10.28. The set t
1
, . . . , t
n
of invariant polynomials is independent,
in the sense that none of them can be written as a polynomial of the others. The
same is true for the set
1
, . . . ,
n
. Furthermore
I(GL(n, 1)) = 1[t
1
, . . . , t
n
] = 1[
1
, . . . ,
n
] (10.13)
i.e. any GL(n, 1)-invariant polynomial can be written as a polynomial of the
t
i
s or the
i
s.
Upon dening c
k
:= (
i
2
)
k
k
we get
I(GL(n, C)) = 1[c
1
, , c
n
]. (10.14)
An immediate consequence of (10.13) is that the
k
s can be written as poly-
nomials in the t
k
s. These relations are called Newton relations, and the rst of
them read
1
= t
1
,
2
=
1
2
(t
2
1
t
2
),
3
=
1
3
(
1
2
t
3
1
3t
1
t
2
+t
3
).
If R is the curvature of some connection on a vector bundle E and (U
) is
a trivialization cover, then we have the matrices of curvature 2-forms. If P is a
GL(n, K)-invariant polynomial we can form the K-valued form P(
) over U
. If
U
and U
are two trivialization neighborhoods that intersect nontrivially then,

by Lemma 10.21, we have
= g
1
on U
, where g
(p) GL(n, K)
and hence by invariance of P we have P(
) = P(
) on U
. Thus, we
can piece the forms P(
) together to a global dierential form P(R) satisfying

P(R)[
U
= P(
).
Lemma 10.29. P(R) is a closed dierential form.
Proof. We will show that this is true locally, i.e. that d(P()) = 0 whenever
is the curvature 2-form on some neighborhood U.
Recall the denition = d + . We then have
d = d d
= + (10.15)
=
(this is also known as the Bianchi identity).
Now we consider d(t
i
())
d(t
i
()) = d(Tr(
i
)) = Tr(d
i
)
= Tr(d
i1
+ d
i2
+ +
i1
d)
= Tr(( )
i1
+ ( )
i2
+
+
i1
( ))
= Tr(
i

i
).
During the calculation we exploited the Bianchi identity along with the fact that
the second last sum is telescoping. Now since Tr is an invariant polynomial we
have Tr(
i
) = Tr(
i
) and hence that d(t
i
()) = 0. As P is a polynomial
in the t
i
s we get d(P()) = 0.
Thus, if P is a homogenous invariant polynomial of degree k P(R) is a closed
2k-form on M. Therefore it represents a cohomology class [P(R)] in H
2k
dR
(M, K)
2
.
Lemma 10.30. The cohomology class [P(R
)] H
2k
dR
(M, K) is independent of
the choice of connection.
Proof. Let
0
and
1
be two connections on E and R
0
and R
1
the corre-
sponding curvature tensors. We will see that [P(R
0
)] = [P(R
1
)].
Consider the projection
1
: M 1 M and let E
t
be the bundle over
M 1 induced from E by this projection and let
t
0
and
t
1
be the pullback
connections on E
t
. We dene a new connection on E
t
by
(s)(p, t) = (1 t)
t
0
(s)(p, t) +t
t
1
(s)(p, t),
i.e. (s) is of the form f
0
t
0
(s) + f
1
t
1
(s) where in this case f
0
(p, t) = 1 t
and f
1
(p, t) = t. Let R
denote the corresponding curvature, then P(R
) is a
closed 2k-form over M 1.
Now let, for k = 0, 1,
k
: M M 1 denote the map p (p, k). As
1

k
= id
M
, we have
k
E
t
= E. By 2) of Proposition 10.15 we see that
0
((s)) =
0
(f
0
t
0
(s) +f
1
t
1
(s)) = (f
0

0
)
t
0
(s) + (f
1

0
)
0
(
t
1
(s))
=
t
0
(s) =
0
(s) =
0
(s)
and similarly that
1
=
1
. Therefore (since the pullback map is multiplicative
with respect to the wedge product)
P(R
0
) = P(R
) =
0
P(R
)
P(R
1
) = P(R
) =
1
P(R
).
But as
0
and
1
are obviously homotopic, they induce the same maps on coho-
mology and thus we have
[P(R
0
)] = [
0
P(R
)] =
0
[P(R
)] =
1
[P(R
)] = [P(R
1
)].
By this last result [P(R)] does not depend on the curvature but depends only
on the (isomorphism class of the) vector bundle E and therefore we may write
it P(E). A cohomology class obtained in this way is called a characteristic class
for the vector bundle E.
Proposition 10.31 (Naturality of Characteristic Classes). Let E be a
vector bundle over M and f : N M a smooth map. Then for any invariant
polynomial P we have
P(f
E) = f
P(E).
Proof. Let be a connection on E, then f
is a connection on f
E and by
point 3 of Proposition 10.15 we have P(R
f
) = f
P(R
) and therefore
P(f
E) = [P(R
f
)] = [f
P(R
)] = f
([P(R
)]) = f
P(E).
Proposition 10.32 (Isomorphism Invariance). Two isomorphic bundles
have the same characteristic classes.
2
H
n
dR
(M; 1) is just the usual de Rham cohomology i.e. the homology of the cocomplex
(
k
(M, 1), d). In the complex case we dene H
2k
dR
(M, C) to be the homology of the cocomplex
(
k
(M, C), d). Since
k
(M, C)

=
k
(M, 1) C, it is not hard to see that H
n
dR
(M, C)

=
H
n
dR
(M) C.
Proof. Let : E F be a bundle isomorphism. It induces an isomorphism
: (E)

(F) by
(s) = s and similarly an isomorphism
k
(M, E)

k
(M, F) by
( s) = ( s).
Given a connection on E there exists a connection
t
on F such that
(E)

1
(M, E)
(F)

1
(M, F)
commutes, simply dene
t
by
t
( s) :=
(s).
Since
t
is K-linear and since for any f C
(M) and s (E) we have
t
(f( s)) =
t
( (fs)) =
(fs)
=
(df s +fs) = df ( s) +f
(s)
= df ( s) +f
t
( s),
t
is a connection on F.
Now let s
1
, . . . , s
n
be a local frame for E and let
ij
be the corresponding
connection 1-forms for . Put t
i
=
s
i
, then
t
t
j
=
t
( t
j
) =
(s
j
) =
_
n
i=1
ij
s
i
_
=
n
i=1
ij
t
i
i.e.
t
has the same connection 1-forms relative to the frame t
1
, . . . , t
n
. Thus
they also have the same curvature 2-forms, and hence the same characteristic
classes.
Example 10.33. 1) For a bundle over a contractible manifold M all charac-
teristic classes in positive degree vanish since H
k
(M; K) = 0 for k 1.
2) Product bundles have vanishing characteristic classes as well: For on the
product bundle we have the trivial connection which is at, i.e. its curvature is
zero. Thus all curvature 2-forms are zero and so are the characteristic classes.
Another way to see this is the following: Consider the map f : M p
0
mapping all of M to some point p

0
. Then the product bundle MK
n
equals the
pullback f
(p
0
K
n
) and by naturality, the characteristic classes of M K
n
are pullbacks of the characteristic classes of p
0
K
n
which are zero by 1).
For the next proposition it is crucial for the vector bundles in question to be
real. The statement is not true for complex vector bundles!
Proposition 10.34. If E is a real vector bundle and P is an invariant poly-
nomial of odd degree, then P(E) = 0.
Proof. As mentioned above we have I(GL(n, 1)) = 1[t
1
, . . . , t
n
], so if P is an
invariant polynomial of odd degree then it must be a polynomial in t
1
, t
3
, t
5
, . . .
Thus it suces to show that t
k
(E) = 0 for k odd. The characteristic class
is independent of the connection, so equip E with a ber metric and choose
a metric connection . On each of the trivialization neighborhoods we pick
orthonormal frames, and thus, by Lemma 10.24 the matrix of curvature forms
is antisymmetric. If A is some antisymmetric matrix, then obviously A
k
is also
antisymmetric (k is still odd) and thus t
k
(A) = Tr A
k
= 0. Thus also t
k
(
ij
) = 0
and the result follows.
In view of Proposition 10.34 we dene:
Denition 10.35 (Pontrjagin Class). Let E be a real vector bundle of di-
mension n, and dene for k = 1, . . . , [n/2]
p
k
:=
1
(2)
2k
2k
.
The characteristic class p
k
(E) H
4k
dR
(M)

= H
4k
(M; 1) is called the kth Pon-
trjagin class of E.
The map M
n
(1) X det(I +
1
2
X) is an invariant polynomial as well,
and by denition of
k
we have
det
_
I +
1
2
X
_
= 1 +
n
k=1
1
(2)
k
k
(X).
Since
k
(E) = 0 for k odd (Proposition 10.34) we have
p(E) := det
_
I +
1
2
E
_
= 1 +p
1
(E) +p
2
(E) + +p
[n/2]
(E) H
(M; 1),
and this characteristic class is called the total Pontrjagin class.
Recalling the denition c
k
= (
i
2
)
k
k
, in the complex case we dene:
Denition 10.36 (Chern Class). Let E be a complex vector bundle of di-
mension n. The characteristic class c
k
(E) H
2k
dR
(M, C)

= H
2k
(M; C) is called
the kth Chern class of E.
The class
c(E) := det
_
I +
i
2
E
_
= 1 +c
1
(E) +c
2
(E) + +c
n
(E) H
(M; C)
is called the total Chern class.
Even though it is dened in terms of complex polynomials and complex con-
nection forms, it turns out that the Chern-classes are real cohomology classes:
Lemma 10.37. Let E be a complex vector bundle, then c
k
(E) H
2k
(M; 1)
and hence c(E) H
(M; 1).
Proof. The Chern class is independent of the choice of connection, so we can
pick a metric connection associated to some hermitian ber metric on E. Then,
relative to some orthonormal frame, the curvature matrix will be skew-adjoint,
and hence I +
i
2
will be self-adjoint. The determinant of a self-adjoint matrix
will be a real number, and thus the dierential form I +
i
2
does indeed take
values in 1: To be more specic: when action on a vector eld it produces a real
number, and thus it is a real cohomology class.
In Section 10.8 we go one step further to show that both Chern and Pontrjagin
classes are in fact integer cohomology classes.
The total Pontrjagin and Chern classes behave nicely with respect to direct
sums:
Theorem 10.38 (Whitney Sum Formulas). Let E and F be vector bundles
over K.
1) If K = 1: p(E F) = p(E) p(F) or explicitly
p
k
(E F) =
k
i=1
p
i
(E) p
ki
(F).
2) If K = C: c(E F) = c(E) c(F) or explicitly
c
k
(E F) =
k
i=1
c
i
(E) c
ki
(F).
Proof. We show it for Chern classes. The proof for Pontrjagin classes is com-
pletely analogous.
Let be a connection on E and
t
a connection on F. Then under the
isomorphisms (E F)

= (E) (F) and
1
(M, E F)

=
1
(M, E)
1
(M, F) it is a routine calculation to show that
t
dened by (
t
)(s, s
t
) = (s,
t
s
t
) mapping
(E) (F) (T
M E) (T
F)
= (T
M (E F))
is a connection on E F. This is called the direct sum connection. Consider
an open set U which is a trivialization for both E and F. Let (s
1
, . . . , s
m
) be a
local frame for E and (t
1
, . . . , t
n
) be a local frame for F, both over U, and let
E
and
F
be the corresponding connection matrices. Then the corresponding
connection form for E F relative to the frame (s
1
, . . . , s
m
, t
1
, . . . , t
n
) is
=
_
E
0
0
F
_
. (10.16)
Similarly the curvature matrix for E F is the block diagonal matrix of the
curvature forms of E and F respectively:
=
_
E
0
0
F
_
.
Thus over U we get
det
_
I +
i
2
_
= det
_
I +
i
2
E
_
det
_
I +
i
2
F
_
.
For any point p U there is a neighborhood where this identity holds, thus it
holds pointwise over all of M. Passing to the level of cohomology classes, the
wedge product is transformed to a cup product, and the formula is proved.
Example 10.39. Lets calculate the Pontrjagin classes of the tangent bundle
of a sphere S
n
. Upon adding to TS
n
the normal bundle NS
n
(which is a trivial
line bundle) we get the product bundle S
n
1
n
. But by Example 10.33 the total
Pontrjagin class of a product bundle is just 1, so by the Whitney sum formula:
1 = p(TS
n
NS
n
) = p(TS
n
) p(NS
n
) = p(TS
n
) 1 = p(TS
n
),
i.e. all the Pontrjagin classes for TS
n
are zero. This, however, does not imply
that the tangent bundles of the spheres are all trivial. Indeed, it is a highly
non-trivial fact, that only TS
1
, TS
3
and TS
7
are trivial.
From a vector space V , we can form the dual space V
. Similarly, for a
vector bundle E we can form the dual bundle E
, this is simply the bundle

whose bers are the dual spaces E
p
. This can be given a natural vector bundle
structure. A closely related object is the conjugate bundle E, whose bers E
x
are just the bers E
p
equipped with the scalar multiplication (z, v) zv.
Of course, if E is a real bundle, E = E. In any case, the conjugate bundle is
isomorphic to the dual bundle: Pick a ber metric on E, then the map E E
,
v
v
where
v
(w) = w, v is easily seen to be a bundle isomorphism. In
particular, if E is a real bundle, it is isomorphic to its dual, although the explicit
isomorphism depends on the choice of metric. In the complex case, E and E
are not isomorphic, as is seen by the following proposition

Proposition 10.40. The Chern classes of the conjugate and dual bundle of E
are given by
c
k
(E
) = c
k
(E) = (1)
k
c
k
(E). (10.17)
Proof. The rst identity is a consequence of the isomorphism E = E
. The
second one can be seen as follows: Let be a metric connection on E and
the matrix of curvature forms relative to some trivialization. Then is again a
connection on E and is the curvature form of this curvature. But as was
a metric connection, we have =
T
. As
k
is a polynomial of homogenous
degree k, the result follows.
We end this section with a few words on the relation between the Pontrjagin
and Chern classes. Let E be a real vector bundle. From this we can form the
a complex vector bundle E
C
:= E
R
C, the complexication of E. If is a
connection on E we can dene a connection on E
C
in the following way: Observe
that (E
R
C)

= (E)
C
(M,R)
C
(M, C) and thus that a section of E

C
is
of the form s z, where z : M C is a smooth function. A connection
C
is
then given by
C
(s z) = (s) z
(and extended linearly). Checking that this is well-dened and a connection
is trivial. Assume now, that (s
1
, . . . , s
n
) is a local frame for E and is the
corresponding connection form. Then (s
1
1, . . . , s
n
1) is a local frame for E
C
(now we allow the coecient functions to be complex-valued). We see that
C
(s
j
1) = (s
j
) 1 =
_
n
i=1
ij
s
i
_
1 =
n
i=1
ij
(s
i
1)
i.e. the connection form for
C
is just , now viewed as a complex form. In the
same manner: the curvature form of
C
is just the curvature form of , viewed
as a complex form. Therefore
p
k
(E) =
_
1
(2)
2k
2k
(R
)
_
=
_
(1)
k
_
i
2
_
2k
2k
(R
C)
_
= (1)
k
c
2k
(E
C
).
Thus we have showed the promised relation:
Proposition 10.41. Let E be a real vector bundle and E
C
its complexication.
Then
p
k
(E) = (1)
k
c
2k
(E
C
). (10.18)
10.7 Orientation and the Euler Class
In the previous section we were mostly interested in GL(n, K)-invariant poly-
nomials. In the following our group will be SO(2n) and the Lie algebra will
be so(2n), the set of skew-symmetric 2n 2n-matrices. We dene a SO(2n)-
invariant polynomial, called the Pfaan Pf : so(2n) 1 in the following
way: Let A so(2n) and dene
(A) :=
1i<jn
A
ij
e
i
e
j

2
(1
2n
).
Now the n times wedge product
1
n!
(A) (A) is a 2n-form, i.e. it is
proportional to the volume form := e
1
e
2n
on 1
2n
. We dene the
Pfaan of A to be exactly this proportionality factor, i.e.
1
n!
(A) (A) = Pf(A).
10.7 Orientation and the Euler Class 187
One can (but we will not) show the following properties of the Pfaan
3
:
Proposition 10.42. If A so(2n) and B M
2n
(1) then:
1) Pf(A) =
1
2
n
n!
S2n
sign A
(1)(2)
A
(2n1)(2n)
.
2) Pf(BAB
T
) = Pf(A) det B.
3) Pf(A)
2
= det A.
4) If A and B are anti-symmetric, then the Pfaan on the block-diagonal
matrix C = diag(A, B) satises
Pf(C) = Pf(A) Pf(B). (10.19)
Note, that the content of the rst two is, that the Pfaan is an SO(2n)-
invariant polynomial.
From this invariant polynomial we will construct a characteristic class. How-
ever, this cannot be done for all vector bundles, only for a privileged class: the
orientable bundles.
Denition 10.43 (Orientable Vector Bundle). Let E be a real vector bun-
dle over M. E is said to be orientable if there is a trivialization cover (U
i
,
i
)
such that the corresponding transition functions g
ij
map into GL
+
(n, 1), i.e.
have positive determinant. A choice of a specic set of trivializations is called an
orientation on E. If E is endowed with an orientation it is said to be oriented.
It is easy to see, that if M is connected, there are only two possible orienta-
tions. If M has n components there are two possibilities for each component,
i.e. a total of 2
n
possible orientations. A fancy way of phrasing this is by saying
that the set of orientations is in bijective correspondence with H
0
(M; Z
2
), the
zeroth singular cohomology group of M with coecients in Z
2
.
Let E be a complex vector bundle of rank n. Let E
R
denote the realication of
E, i.e. the same bundle but we forget the complex structure. This is a real vector
bundle of rank 2n and more interesting: it is orientable. This is for the following
reason: Let V be a complex vector space and V
R
its realication. If v
1
, . . . , v
n
is a complex basis for V , then v

1
, iv
1
, , v
n
, iv
n
is a real basis for V
R
. Let
w
1
, . . . , w
n
be another complex basis for V and let A = (b
kl
+ ic
kl
) be the
transition matrix. It takes but a trivial calculation to see, that the corresponding
transition matrix between the real bases is
A
R
=
_
_
_
_
_
_
_
b
11
c
11
b
n1
c
n1
c
11
b
11
c
n1
b
n1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
b
1n
c
1n
b
nn
c
nn
c
1n
b
1n
c
nn
b
nn
_
_
_
_
_
_
_
. (10.20)
The transition matrix A is an element of GL(n, C) which is path-connected, in
particular there exists a continuous curve joining A to the identity I GL(n, C).
Realifying this curve, will give a curve in GL(2n, 1) joining A
R
with I
2n
. Thus
A
R
will have positive determinant. The upshot of all this is therefore that two
dierent complex bases on a complex vector space induce the same orientation
on the realication, i.e. the realication has a canonical orientation.
If E is a complex vector bundle, we give each ber in the realication the
canonical orientation. This is easily seen to determine an orientation on E
R
,
thus we have proved:
3
For proofs consult [MT] Appendix B.
Lemma 10.44. The realication of a complex vector bundle is orientable and
has a canonical orientation.
Assume now that E is an oriented real vector bundle of rank 2n, and let
(U
) be an oriented trivialization set. We may assume that the associated

local sections constitute local orthonormal frames. Pick a ber metric on E
and let be a metric connection. With respect to the orthonormal frame in
question, the matrices of curvature 2-forms
are skew-symmetric. Thus, we

can apply the Pfaan to
to obtain a 2n-form Pf(

1
2
) over U
. On the
overlap U
we have
= g
1
and since the trivializations are oriented and the associated local frames are
orthonormal, g
takes values in SO(2n) we get

Pf(
1
2
) = det(g
) Pf(
1
2
) = Pf(
1
2
).
Thus we can piece the individual forms together to a global 2n-form, which we
will denote by Pf(R
).
Lemma 10.45. Pf(R
) is a closed 2n-form.
Proof. By using the Bianchi identity on the explicit formula for the Pfaan,
Proposition 10.42 1), a direct computation as the one in the proof of Lemma
10.29 will show that Pf(R
) is closed.
Thus, Pf(R
) determines a class in H
2n
(M; 1).
Lemma 10.46. The cohomology class [Pf(R
)] is independent of the choice of

metric and metric connection.
Proof. Assume rst that we have two metrics g
0
and g
1
on E and
0
and
1
two metric connections (relative of course to g
0
and g
1
respectively). Let
: M 1 M be the projection and i
k
: M M 1 be the inclusions
p (p, k) for k = 0, 1. We want to construct a metric on
E which pulls
back to g
0
and g
1
via i
0
and i
1
(note that i
k
(
E)
= E).
Consider the open cover M],
3
4
[ and M]
1
4
, [ of M1 and let (
0
,
1
)
be a partition of unity subordinate to the cover. Put g =
0
g
0
+
1
g
1
, this
is a metric on
E. Over M 0 it is just g
0
and over M 1 it is just g
1
.
Since the pullback i
k
is just restriction, we get i
k
g = g
k
.
Now, let

be a connection which is compatible with g, in particular it is
compatible with g on (
E)[
M]
1
8
,
7
8
[
(this is Lemma 10.25 applied to the inclu-
sion map M ]
1
8
,
7
8
[ M 1). Furthermore
0
and
1
are compatible
with
g
0
and
g
1
respectively and thus
0
will be compatible with g over
M ] ,
1
4
[ and
1
will be compatible with g over M ]
3
4
, [. Thus by
using a partition of unity subordinate to this new cover of M1 we get a con-
nection

on
E and this is compatible with g since (10.11) by construction

of

holds in a neighborhood of every point.
The pullback i
is the unique connection on i
k
(
E)

= E which satises
i
(i
k
s) = i
k
(
s). But this equation is, by construction of

, satised by
k
,
thus by uniqueness, i
=
k
.
Now, by Proposition 10.15 3) we have
i
k
(Pf(R
)) = Pf(R
i
) = Pf(R
k
).
Since i
0
and i
1
are homotopic, they induce the same map in cohomology, thus
Pf(R
0
) = Pf(R
1
).
10.7 Orientation and the Euler Class 189
Denition 10.47 (Euler Class). The cohomology class e(E) := [Pf(R
)]
H
2n
(M; 1) which depends only on E is called the Euler class of the oriented
bundle E.
The Euler class possesses some properties which are analogous to the proper-
ties of the Chern and Pontrjagin classes:
Proposition 10.48. The Euler class satises
1) e(E F) = e(E) e(F).
2) If E is a bundle over M and f : N M is smooth, then
e(f
E) = f
(e(E)).
where f
E is given the induced orientation.

Proof. 1) This a simple consequence of (10.16) in combination with (10.19).
2) Let be any metric connection on E. Pick f
on f
E, then by the same

sort of argument as in the proof of Proposition 10.31 we get that
e(f
E) = [Pf(R
f
)] = [f
Pf(R
)] = f
[Pf(R
)] = f
e(E).
If E is an complex bundle of rank n we saw above, that its realication was
an orientable bundle of rank 2n. Thus it makes sense to dene the Euler class
of an arbitrary complex vector bundle by e(E) := e(E
R
).
Proposition 10.49. The relations to other characteristic classes are as follows:
1) If E is an oriented rank 2n real vector bundle, then e(E)
2
= p
n
(E).
2) If E is a complex vector bundle of rank n, then e(E) = c
n
(E).
Proof. 1) Let g be a metric on E and a metric connection. Let be the
matrix of curvature 2-forms relative to some local orthonormal frame. Now we
complexify E to obtain E
C
, a rank 2n complex bundle. Let g
C
be the complex-
ied metric on E
C
, it is dened ber-wise by
g
C
(v z, v
t
z
t
) = g(v, v
t
)zz
t
.
The connection is , if extended by C-linearity to a map dened on (E
C
)

=
(E)
R
C, becomes a connection on E
C
, and it is compatible with g
C
. The local
frame mentioned above is still an orthonormal frame for E
C
, and will be the
curvature form for relative to this frame. Now we have locally that
c
2n
(E
C
) =
_
1
2i
_
2n
det().
Similarly we have
e(E)
2
=
_
1
2
_
2n
Pf()
2
=
_
1
2
_
2n
det().
Since c
2n
(E
C
) = (1)
k
p
n
(E) by Proposition 10.41 we reach the conclusion.
2) Choose a complex metric g on E and let be a corresponding metric
connection. It maps into
1
(M, E) viewed as
1
(M, C)
C
(M,C)
(E).
On E
R
we have a metric g
R
given by
g
R
(s, s
t
) := Re g(s, s
t
).
Under the isomorphism
1
(M, E)
=
1
(M, C)
C
(M,C)
(E)
=
1
(M, 1)
C
(M,R)
C
(M, C)
C
(M,C)
(E)
=
1
(M, 1)
C
(M,R)
(E
R
)
we can perceive as a connection on E
R
. Under this isomorphism all connection
and curvature forms will be real dierential forms. Obviously, is compatible
with g
R
simply take the real part of (10.11).
Let s
1
, . . . , s
n
be a local g-orthonormal frame for E. Then s
1
, is
1
, . . . , s
n
, is
n
is a local g
R
-orthonormal frame for E
R
. Let
ij
be the curvature 2-forms for
relative to the frame (s
i
). They are complex dierential forms, so we write them
as
ij
= a
ij
+ib
ij
. Just as in the beginning of this section one can show that the
matrix of curvature 2-forms for E
R
relative to the frame s
1
, is
1
, . . . , s
n
, is
n
is
A
R
=
_
_
_
_
_
_
_
a
11
b
11
a
n1
b
n1
b
11
a
11
b
n1
a
n1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
a
1n
b
1n
a
nn
b
nn
b
1n
a
1n
b
nn
a
nn
_
_
_
_
_
_
_
. (10.21)
This is exactly the way we turn a complex matrix into a real matrix. So in order
to show the desired identity we only need to verify the purely algebraic identity
Pf(A
R
) = (i)
n
det A for some skew-adjoint n n-matrix A.
Since A is skew-adjoint it can be diagonalized, i.e. there exists a unitary matrix
U such that UAU
1
= diag(i
1
, . . . , i
n
). Realifying we get U
R
SO(2n) and
U
R
A
R
U
1
R
= diag
_
_
0
1
1
0
_
, . . . ,
_
0
n
n
0
_
_
.
Thus by 2) and 4) of Proposition 10.42 we get
Pf(A
R
) = Pf(U
R
A
R
U
1
R
) =
1

n
.
On the other hand det(A) = i
n
1

n
and the identity is proved.
The importance of the Euler class (except for its use in the Atiyah-Singer
index formula) lies in the fact that it provides an obstruction to the existence
of an everywhere nonzero section of the bundle:
Theorem 10.50. Let E M be a real oriented vector bundle. If an every-
where nonzero section s : M E exists, then e(E) = 0.
We will not use this theorem and hence not prove it.
10.8 Splitting Principle, Multiplicative Sequences
In this section we dene the Chern character, a map which gives an important
link between cohomology of a space and its K-theory. There are several ways
to accomplish this, but the most elegant way, I think, involves cohomology with
integer coecients. Thus the rst task before us is to show that the Chern
classes as dened above are not just 1-cohomology classes but in fact integer
cohomology classes. An important ingredient in this is the following, which I do
not intend to prove
4
:
4
For a proof consult [MT] Chapter 20 or [Ha] Proposition 3.3.
10.8 Splitting Principle, Multiplicative Sequences 191
Theorem 10.51 (The Splitting Principle I). Let E be a real or complex vec-
tor bundle of rank n over the manifold M. Then there exists a smooth manifold
T (depending on E) and a smooth map f : T M such that
1) f
E = L
1
L
n
where L
1
, . . . , L
n
are real or complex line bundles
over T.
2) For each k and each coecient ring R the map f
: H
k
(M; R)
H
k
(T; R) is injective.
For real vector bundles there is an alternative version of the splitting principle
Theorem 10.52 (The Splitting Principle II). Let E be an oriented real
vector bundle of rank n over the manifold M. Then there exists a smooth
manifold T (depending on E) and a smooth map f : T M such that
f
: H
(M; R) H
(T; R) is injective and such that

1) f
(E
C
) = L
1
L
1
L
k
L
k
if n = 2k
2) f
(E
C
) = L
1
L
1
L
k
L
k
I
1
if n = 2k + 1
where L
i
are complex line bundles, L
i
is the conjugate bundle and I
1
is the
trivial complex line bundle.
Thanks to this splitting principle we can prove a uniqueness result for the
Chern-classes. In what follows, H
n
will denote the canonical complex line bundle
over CP
n
.
Theorem 10.53. There exists a unique set of maps c
k
: Vect
C
(M) H
2k
(M; 1),
k N satisfying
1)
_
M
(c
1
(H
1
)) = 1, c
k
(H
1
) = 0 for k > 1.
2) f
(c
k
(E)) = c
k
(f
E) for any smooth map f : N M.

3) c
k
(E F) =
k
j=1
c
j
(E) c
kj
(F).
Proof. We do know that they exist, after all that was, what all the eort in
the previous sections was for. The only thing missing is the rst point, but this
can be seen in [MT] Theorem 18.4.
The proof of uniqueness is divided into 3 steps: rst we show it for a line
bundle, then for a sum of line bundles and nally for a general bundle. First,
let : L M be a smooth line bundle over M and assume we have maps
c
1
, c
2
, . . . satisfying the 3 requirements above. There exists an integer n and a
rank n bundle L
over M such that L L

= M C
n+1
. Thus we may view
L as sitting inside M C
n+1
.
We dene a map p : M CP
n
in the following (slightly complicated) way:
For a point x M the ber L
x
in L is a complex line in C
n+1
and thus represents
an element in CP
n
. To see that this map is smooth we see how it looks locally:
Around any point in M there is a neighborhood U and a nowhere vanishing
section s : U L of L. On U the map p is given by x [
2
(s(x))] where
2
: M C
n+1
C
n+1
is the projection map and [ ] : C
n+1
CP
n
is the
quotient map. These are all smooth maps, and thus p is smooth over U.
In the same spirit we dene a map p : L H
n
in the following way: Let
U CP
n
be a trivialization neighborhood for H
n
such that p
1
(U) M is
a trivialization neighborhood for L. Then we dene p over
1
(U) L as the
composition
L[
1
(U)
p
1
(U) C
pid
U C

H
n
[
U
.
This gives a well-dened smooth bundle map which makes the following diagram
commutative:
L
p
H
n
M
p
CP
n
Since the pullback p
H
n
is the unique bundle over M making the diagram above
commutative, we have L
= p
H
n
. But then
c
1
(L) = c
1
(p
H
n
) = p
(c
1
(H
n
))
and for k > 1
c
k
(L) = p
(c
k
(H
n
)) = 0.
Thus the c
k
s are uniquely determined on a line bundle.
Consider now a sum L
1
L
n
of line bundles. Then inductive use of
requirement 3) gives c
k
in terms of c
1
(L
1
), . . . , c
1
(L
n
) which were uniquely de-
termined. Thus the c
k
s are uniquely determined on a sum of line bundles.
Finally, assume E to be a generic vector bundle over M. By the splitting
principle there exists a manifold T and a map f : T M such that f
E

=
L
1
L
n
is a sum of line bundles. Assume we have cohomology classes
c
k
(E), c
t
k
(E) H
2k
(M) satisfying the 3 conditions, then by point 2) we have
f
(c
k
(E)) = c
k
(f
E) = c
t
k
(f
E) = f
(c
t
k
(E))
and since f
: H
2k
(M; 1) H
2k
(T; 1) is injective, c
k
(E) = c
t
k
(E).
This uniqueness statement can be exploited to show that the Chern classes are
in fact integer cohomology classes, but we will not need that. For our purpose
it suces that they are real cohomology classes.
In the previous sections we have constructed characteristic classes from invari-
ant polynomials. However in order to construct the characteristic classes needed
for the Atiyah-Singer Index Theorem, we cannot conne ourselves to just poly-
nomials, we need to be able to construct characteristic classes out of innite
power series. The remainder of this section will create the proper framework
and in the next we will consider some very concrete and important examples.
We begin by introducing symmetric polynomials:
Denition 10.54 (Symmetric Polynomial). A symmetric polynomial p of
degree k is a polynomial of homogenous degree k satisfying
p(x
1
, . . . , x
n
) = p(x
(1)
, , x
(n)
)
for all S
n
.
An obvious example of such a polynomial is
t
k
(x
1
, , x
n
) := x
k
1
+ +x
k
n
.
Furthermore we have the so-called elementary symmetric polynomials given by
1
(x
1
, . . . , x
n
) = x
1
+ +x
n
2
(x
1
, . . . , x
n
) =
i1<i2
x
i1
x
i2
3
(x
1
, . . . , x
n
) =
i1<i2<i3
x
i1
x
i2
x
i3
.
.
.
n
(x
1
, . . . , x
n
) = x
1
x
n
.
The main result of symmetric polynomials is the following
Theorem 10.55. The set of elementary symmetric polynomials is algebraically
independent (i.e. no
i
can be written as a polynomial of the rest) and further-
more any symmetric polynomial can be expressed as a polynomial in the
i
s.
In particular the symmetric polynomial t
k
can be written as a polynomial
s
k
(
1
, . . . ,
k
). These particular polynomials s
k
are called the Newton polyno-
mials. The rst Newton polynomials are
s
1
=
1
s
2
=
2
1
2
2
s
3
=
3
1
3
1
2
+ 3
3
s
4
=
4
1
4
2
1
2
+ 4
1
3
+ 2
2
2
4
4
.
In general one has the following recursion formula
s
k
=
1
s
k1
2
s
k2
+ + (1)
k2
k1
s
1
+ (1)
k1
k
k
. (10.22)
We will return to these in the following section.
For now, consider
1[x] := 1 +a
1
x +a
2
x
2
+ [ a
i
1,
the set of formal power series in the indeterminate x with real coecients and
constant term 1 (the bar is to resemble the closure). Fixing f 1[x] we construct
a formal power series in n indeterminates by
f(x
1
) f(x
n
).
Group the terms according to their degree. The sum of the terms of degree k
will be a homogenous symmetric polynomial and can therefore be written as a
polynomial F
k
in the elementary symmetric polynomials, i.e.
f(x
1
) f(x
n
) = 1 +F
1
(
1
) +F
2
(
1
,
2
) +F
3
(
1
,
2
,
3
) +
The polynomials F
k
are independent of the number n of indeterminates because
k
(x
1
, . . . , x
n
, 0, . . . , 0) =
k
(x
1
, . . . , x
n
) if k n and
k
(x
1
, . . . , x
n
, 0, . . . , 0) =
0 if k > 0. In this way we obtain an innite sequence (F
k
)
k=1
of polynomials,
called the multiplicative sequence which is characteristic of the formal power
series f. A great source of power series is of course Taylor series of holomorphic
functions. The Taylor series of a product of holomorphic functions is, of course,
just the product of the Taylor series, as above. Therefore, if f = gh is a product
of holomorphic functions, and F, G and H are the corresponding multiplicative
sequences, we get
F(x
1
, . . . , x
n
) = G(x
1
, . . . , x
n
)H(x
1
, . . . , x
n
). (10.23)
Let B =
k
B
k
be a commutative N
0
-graded algebra over 1, examples in-
clude 1[x], the polynomial algebra, and H
2
(M; 1) the even cohomology with
real coecients. Let

B be the subalgebra consisting of elements of the form
1 +b
1
+b
2
+ +b
n
for b
i
B
i
and varying n. To a xed f 1[x] we associate a map F :

B

B
by dening
F(1 +b
1
+ +b
n
) = 1 +F
1
(b
1
) +F
2
(b
1
, b
2
) +
For the purpose of constructing characteristic classes of complex vector bun-
dles we will take as our algebra B the even cohomology H
2
(M; 1) of the
manifold. To a given f 1[x] we get a homomorphism F : H
2
(M; 1)
H
2
(M; 1). If E is a complex vector bundle, we dene its total F-class
F
C
(E) := F(c(E)) = 1 +F
1
(c
1
(E)) +F
2
(c
1
(E), c
2
(E)) +
If E is instead a real vector bundle, we take as our algebra H
4
(M; 1), identify
the total Pontrjagin class p(E) as an element of H
4
(M; 1) and dene its total
F-class by
F
R
(E) := F(p(E)).
A prime example of a real vector bundle is of course the tangent bundle TM.
In this case we write F(M) := F
R
(TM) H
4
(M; 1).
If we assume furthermore that M is compact, connected and oriented, then
H
n
(M; 1)

= 1 and the homology class corresponding to 1 is called the funda-
mental class of M and is denoted [M] (the isomorphism H
n
(M; 1) depends on
the orientation, thus changing the orientation changes the fundamental class).
We can evaluate the F-class F(M) on [M] to obtain a real number F(M) called
the F-genus of M. Its easy to see that
F(M) = F(M)[M] =
_
F
k
(p
1
(TM), . . . , p
k
(TM))[M] , n = 4k
0 , n ,= 0 mod 4
Proposition 10.56. Assume f 1[x] is xed. The total F-class has the fol-
lowing properties
1) It is isomorphism invariant: If E

= E
t
, then F
K
(E) = F
K
(E
t
).
2) It it natural : If g : N M is a map, then F
K
(g
E) = g
F
K
(E).
3) It is multiplicative: F
K
(E E
t
) = F
K
(E) F
K
(E
t
).
Proof. The rst assertion follows trivially from the fact, that the Chern and
Pontrjagin classes are isomorphism invariant. The second claim follows from a
small calculation (for which we assume E to be a complex bundle)
F
C
(g
E) = F(c(g
E)) = F(g
c(E))
= 1 +F
1
(g
c
1
(E)) +F
2
(g
c
1
(E), g
c
2
(E)) +
= g
(1 +F
1
(c
1
(E)) +F
2
(c
1
(E), c
2
(E)) + ).
To prove the third claim (as before, we assume the bundles to be complex,
the real case is analogous), assume rst that E = L
1
L
n
and E
t
=
L
n+1
L
n+m
are sums of line bundles. Then by the Whitney sum formula
c(E E
t
) = c(L
1
) c(L
n+m
) = (1 +x
1
) (1 +x
n+m
) = 1 +
1
+ +
n+m
where x
i
= c(L
i
) and the
k
s are the elementary symmetric polynomials in
n +m variables. But then
F
C
(E E
t
) = F(c(E E
t
)) = F(1 +
1
+ +
n+m
)
= f(x
1
) f(x
n+m
) = (f(x
1
) f(x
n
))(f(x
n+1
) f(x
n+m
))
= F(1 +
t
1
+ +
t
n
)F(1 +
tt
1
+ +
tt
m
)
= F(c(E))F(c(E
t
)) = F
C
(E)F
C
(E
t
)
where (
t
k
) and (
tt
k
) are the elementary symmetric polynomials in the variables
x
1
, . . . , x
n
resp. x
n+1
, . . . x
n+m
.
If E and E
t
are arbitrary bundles we apply the splitting principle. By the
splitting principle there exists a manifold T
1
and a smooth map g
1
: T
1
M
such that g
1
E is a direct sum of line bundles. The pullback g
1
E
t
need not be
a sum of line bundles, but applying the splitting principle once again we get a
manifold T
2
and a smooth map g
2
: T
2
T
1
such that g
2
(g
1
E
t
) is a sum of
line bundles. Obviously g
2
(g
1
E) is still a sum of line bundles, and thus putting
T := T
2
and g := g
1
g
2
: T M we have g
E and g
E
t
are sums of line
bundles over the same manifold. Hence by naturality
g
F
C
(E E
t
) = F
C
(g
E g
E
t
)
= F
C
(g
E)F
C
(g
E
t
) = g
(F
C
(E)F
C
(E
t
)).
Since g
: H
(M; 1) H
(T; 1) is injective, the assertion is proven.

By the Splitting principle we can in many cases reduce to the case of line
bundles. Therefore these deserve a special treatment. Let E be a real line bundle,
such that E
C
= L
1
L
1
L
n
L
n
, a sum of complex line bundles. By the
Whitney sum formula we get
c(E
C
) =
n
k=1
(1x
2
k
) = 1
1
(x
2
1
, . . . , x
2
n
) +
2
(x
2
1
, . . . , x
2
n
)
3
(x
2
1
, . . . , x
2
n
) +
and from this we get that c
2k
(E
C
) = (1)
k
k
(x
2
1
, . . . , x
2
n
). Thus
p(E) = 1 +p
1
(E) +p
2
(E) + = 1 c
2
(E
C
) +c
4
(E
C
) c
6
(E
C
) +
= 1 +
1
(x
1
, . . . , x
n
) +
2
(x
1
, . . . , x
n
) +
=
n
k=1
(1 +x
2
n
).
In particular, multiplicativity of the F-class above gives
F
R
(E) = f(x
2
1
) f(x
2
n
) (10.24)
still under the assumption that E
C
splits as a direct sum of line bundles as
above.
Example 10.57. 1) The Todd Class. Consider the formal power series
td(x) =
x
1 e
x
= 1 +
1
2
x +
1
12
x
2
1
720
x
4
+
The coecients here in this Taylor series are closely related to the so-called
Bernoulli numbers, in fact these are the coecients in the Taylor expansion of
the holomorphic function
z
e
z
1
.
The corresponding multiplicative sequence is called the Todd sequence (Td
m
).
From
td(x
1
) td(x
2
) = 1 +
1
2
(x
1
+x
2
) +
_
1
12
x
2
1
+
1
4
x
1
x
2
+
1
12
x
2
2
_
+
we see by comparing to the expressions of the elementary symmetric polynomials
that
Td
1
(
1
) =
1
2
1
Td
2
(
1
,
2
) =
1
12
(
2
+
2
1
)
Td
3
(
1
,
2
,
3
) =
1
24
c
2
c
1
and so on. If E is a complex vector bundle, its total Todd class is then dened
by
Td
C
(E) = 1 +
1
2
c
1
(E) +
1
12
(c
2
(E) +c
1
(E)
2
) + (10.25)
As a matter of fact, the Todd class has an inverse, namely consider the
inverse of td(x):
td
1
(x) =
1 e
x
x
= 1
x
2
+
x
2
6

x
3
24
+
Given a complex vector bundle E, we denote the corresponding character-
istic class by Td
1
(E). Since td
1
(x) td(x) = 1, we have by (10.23) that
Td
1
C
(E)Td
C
(E) = 1 in H
2
(M; 1). Sometimes we will use the slightly abusive
notation
1
Td
C
(E)
for the inverse Todd class.
2) The Total

A-Class. In this example we consider the formal power series given
by
a(x) =
x/2
sinh(
x/2)
= 1
1
24
x +
7
5760
x
2
+
The corresponding sequence, called the

A-sequence, has the rst terms
A
1
(
1
) =
1
24
A
2
(
1
,
2
) =
1
5760
(4
2
+ 7
2
1
)
A
3
(
1
,
2
,
3
) =
1
9676800
(16
3
44
2
1
+ 31
3
1
).
For a real bundle, the associated characteristic class

A
R
(E) is called the total
A-class.
3) Hirzebruch L-class. Finally consider the formal power series
l(x) =
x
tanh
x
= 1 +
1
3
x
1
45
x
2
+
The rst terms of the corresponding Hirzebruch L-sequence are
L
1
(
1
) =
1
3
1
L
2
(
1
,
2
) =
1
45
(7
2
2
1
)
L
3
(
1
,
2
,
3
) =
1
945
(62
3
13
1
2
+ 2
3
1
).
For a real bundle E the characteristic class
L
R
(E) = 1 +L
1
(p
1
(E)) +L
2
(p
1
(E), p
2
(E)) +
is called the Hirzebruch L-class or just the total L-class.
Proposition 10.58. Let E be an oriented real vector bundle. Then
Td
C
(E C) = (
A
R
(E))
2
. (10.26)
Proof. By the splitting principle we only need to show it in the case where
E C is a sum of line bundles. Assuming that E has dimension 2n then
E C
= L
1
L
1
L
n
L
n
.
10.9 The Chern Character 197
In this case we get (thanks to (10.17))
Td
C
(E C) =
n
j=1
x
j
1 e
xj
x
j
1 e
xj
where x
j
= c
1
(L
j
). Multiply the denominator by e
xj/2
e
xj/2
to obtain
Td
C
(E C) =
n
j=1
_
x
j
e
xj/2
e
xj/2
_
2
=
n
j=1
_
x
j
/2
sinh(x
j
/2)
_
2
= (
A
R
(E))
2
where the last equality follows from (10.24). The case where E is of odd dimen-
sion is analogous.
10.9 The Chern Character
We end this chapter by introducing the Chern character. This is again a charac-
teristic class dened from an formal power series. However its construction and
its properties are slightly dierent from the F-classes. Let L be a complex line
bundle over a manifold M and dene the Chern character ch(L) on L by
ch(L) = exp(c
1
(L)) = 1 +c
1
(L) +
1
2
c
1
(L)
2
+ +
1
n!
c
1
(L)
n
+
Since
1
k!
(L)
k
H
2k
(M; 1) and since H
n
(M; 1) = for n great enough, this series
terminates after nitely many terms. Thus the Chern character is a well-dened
real cohomology class.
For a direct sum L
1
L
n
of complex line bundles, we have
c(E) =
n
i=1
(1 +x
i
) = 1 +
1
(x
1
, . . . , x
n
) + +
n
(x
1
, . . . , x
n
)
(where x
i
:= c
1
(L
i
)) in other words that
c
k
(L
1
L
n
) =
k
(x
1
, . . . , x
n
). (10.27)
Upon dening the Chern character to be the sum of the Chern character of the
line bundles (this is where we depart from the F-classes)
ch(L
1
L
n
) =
n
i=1
ch(L
i
) =
n
i=1
exp(c
1
(L
i
))
= n +t
1
(x
1
, . . . , x
n
) + +t
k
(x
1
, . . . , x
n
)/k! + (10.28)
and substituting the Newton relations (10.22) into this and using (10.27) we get
ch(E) = n +
n
k=1
s
k
(c
1
(E), . . . , c
n
(E))
k!
. (10.29)
This formula is derived under the assumption that E is a direct sum of line
bundles. But the right hand side is expressed in terms of E only. Therefore we
may take (10.29) to be the denition of the Chern character in the general case:
Denition 10.59 (Chern Character). For a complex vector bundle E of
rank n, dene the Chern character of E by
ch(E) := n +
n
k=1
s
k
(c
1
(E), . . . , c
n
(E))
k!
. (10.30)
Compared to the Chern classes and F-classes, the Chern character has even
nicer properties
Proposition 10.60. Let E and F be complex vector bundles over a manifold
M and f : N M a smooth map, then the Chern character satises the
following
1) Naturality: ch(f
E) = f
ch E.
2) Additivity: ch(E F) = ch(E) + ch(F).
3) Multiplicativity:ch(E F) = ch(E) ch(F).
Proof. 1) This follows immediately from the denition (10.30) as well as from
naturality of the Chern classes and multiplicativity of f
.
2) By denition of the Chern character, this is true if E and F are direct sums
of line bundles. If E and F are generic vector bundles we will use the splitting
principle to reduce to the case of sums of line bundles:
By the splitting principle there exists a manifold T and a smooth map f :
T M such that f
E and f
F are direct sums of line bundles. By naturality

of the Chern character we have ch(f
E) = f
ch(E) and therefore

f
ch(E F) = ch(f
E f
F) = ch(f
E) + ch(f
F)
= f
(ch(E) + ch(F)).
Since f
: H
(M; 1) H
(T; 1) is injective, we have ch(E F) = ch(E) +

ch(F).
3) This can be proved in exactly the same way. All we need to know is that if
L
1
and L
2
are line bundles, then we have c
1
(L
1
L
2
) = c
1
(L
1
) +c
1
(L
2
)
5
, and
therefore
ch(L
1
L
2
) = e
c1(L1L2)
= e
c1(L1)+c1(L2)
= e
c1(L1)
e
c1(L2)
= ch(L
1
) ch(L
2
).
Iterating this formula to arbitrary sums of line bundles and using the splitting
principle as in the proof of 1), we get 2).
Example 10.61. Let E be a complex vector bundle. Lets try to derive a
formula for the alternating sum of the Chern character of the kth exterior
power of E, i.e. ch(
k
(1)
k
ch(
k
E) (articial as this may seem, it will however
turn out useful later on). First, assume that E = L
1
L
n
is a direct
sum of complex line bundles. By an iteration of the formula
k
(V W) =
k
i=0

i
V
ki
W we get
k
E =
1i1<<i
k
n
L
i1
L
i
k
.
Thus (writing x
i
:= c
1
(L
i
)):
ch
k
E =
1i1<...<i
k
n
ch(L
i1
) ch(L
i
k
)
=
1i1<<i
k
n
e
xi
1
e
xi
k
=
1i1<<i
k
n
e
xi
1
++xi
k
5
A proof of this statement may be found in [Ha], the proof of Proposition 3.10.
10.9 The Chern Character 199
and thereby
n
k=0
(1)
n
ch(
k
E) =
n
k=0
(1)
k
1i1<<i
k
n
e
xi
1
++xi
k
=
n
i=1
(1 e
xi
).
By the Whitney formula c(E) =
(1 +x
i
) we easily see that c
n
(E) = x
1
x
n
(still, of course, under the assumption, that E is a sum of line bundles). By
Proposition 10.40 we have c
n
(E
) = (1)
n
x
1
x
n
and hence
n
k=0
(1)
k
ch(
k
E) =
n
i=1
(1 e
xi
)
=
_
(1)
n
x
1
x
n
_
(1)
n
n
i=1
1 e
xi
x
i
= c
n
(E
)Td
1
C
(E
) (10.31)
(note that (10.23) guarantees that these manipulations are legal). By virtue of
the Splitting Principle, this formula holds for any complex vector bundle, being
a direct sum of line bundles or not.
We can immediately extend the denition of the Chern character to K-
theory. Let M be a compact manifold. Since the Chern character Vect
C
(M)
H
(M; 1) is additive it follows from standard properties of the Grothendieck

group, that it descends to a group homomorphism K(M) H
(M; 1) (or, as
the Chern classes are even dimensional, ch : K(M) H
2
(M; 1)), explicitly
it is given by
ch([E] [F]) = ch(E) ch(F).
The properties 2) and 3) of Proposition 10.60 ensure that this is a ring homo-
morphism and property 1) says that if f : N M is a smooth map, the
following diagram is commutative:
K(M)
ch
(M; 1)
f
K(N)
ch
(N; 1)
We will need also a Chern character over non-compact manifolds. The con-
struction is as follows: Let M be a manifold and M
+
its 1-point compacti-
cation. Let j : pt M
denote the inclusion of innity in M

+
, then by
denition K(M) =

K(M
+
) = ker j
. Furthermore we have the isomorphism

H
c
(M; 1)
= ker j
where j
: H
(M; 1) H
(pt; 1) is now the induced map

in cohomology. We have a Chern character ch : K(M
+
) H
(M
+
; 1) and by
naturality it commutes with j
. Thus it is easily seen that the Chern character

maps kernel to kernel, i.e. restricting it to K(M) =

K(M
+
) produces a ring
homomorphism
ch : K(M) H
c
(M; 1). (10.32)
This is our Chern character in the non-compact case.
Chapter 11
Dierential Operators
11.1 Dierential Operators on Manifolds
Let U be an open set in 1
n
. By a dierential operator of order k on U we will
understand an n
1
n
2
-matrix (A
ij
)
A
ij
=
]]k
a
ij
(x)
(11.1)
where a
ij
are smooth complex-valued functions on U.

Let M be a manifold, and let (U, x
1
, . . . , x
n
) be a chart on M. For a smooth
function f C
(U) it makes sense to write
f, it simply means consecutive

actions of the coordinate vector elds
i
= /x
i
on f. Thus we can talk about
a dierential operator over U namely a matrix of operators of the form (11.1).
A dierential operator on M is an operator which locally looks like the oper-
ators described above. More precisely, let
E
: E M and
F
: F M be
complex vector bundles over M (they could have been real as well). We say that
an open set U M is a proper neighborhood if it is the domain of a smooth chart
(U, x
1
, . . . , x
n
) for M and if there exist trivializations
E
:
1
E
(U) U C
n1
and
F
:
1
F
(U) U C
n2
for E and F over U. We formalize the notion of
a dierential operator in the following
Denition 11.1 (Dierential Operator). A C-linear map A : (E) (F)
is called a dierential operator provided:
1) The operator is local , i.e. if u (E) is zero on some open neighborhood
U then Au[
U
= 0 as well, i.e. A decreases support.
2) For each proper neighborhood U with trivializations
E
and
F
there
exists a matrix of dierential operators A
U
(called the local representative
for A) as in (11.1) such that for any u (E):
Au[
U
= (
1
F
(id
U
A
U
)
E
)u[
U
.
Note how point 1) is necessary for point 2) to make sense.
For later use we will investigate how the local representative A
U
depends on
the trivializations. Suppose we have two alternative trivializations
E
:
1
E
(U) U C
n1
and
F
:
1
F
(U) U C
n2
and let A
U
be the local representative for A relative to these trivializations.
We have transition functions g
E
: U GL(n
1
, C) and g
F
: U GL(n
2
, C)
201
202 Chapter 11 Dierential Operators
satisfying
E

1
E
(p, ) = (p, g
E
(p)),
F

1
F
(p, ) = (p, g
F
(p)).
From the equation
1
F
(id
U
A
U
)
E
=
1
F
(id
U
A
U
)
E
(they are both equal to Au[
U
) we get by isolating
id
U
A
U
= (
F

1
F
) (id
U
A
U
) (
E

1
E
) = id
U
(g
F
A
U
g
1
E
),
that is A
U
= g
F
A
U
g
1
E
.
In the following example we consider a number of dierential operators on
manifolds.
Example 11.2. 1) Vector Fields. Let X be a vector eld, and let L
X
:
C
(M) C
(M) be the Lie derivative on functions, i.e. L

X
f = Xf. It is an
operator on the trivial bundle E = M1. In a coordinate chart (U, x
1
, . . . , x
n
)
we have X =
n
i=1
X
i
i
and
L
X
f =
f
x
1
+ +
f
x
n
thus in this chart L
X
has local representative /x
1
+ + /x
n
(a 1 1-
matrix) and thus it is a rst order dierential operator on E.
2) The Exterior Derivative. Consider the vector bundles M1 and T
M (these
are now real vector bundles!), as well as the exterior derivative d : C
(M)
1
(M) between the sets of sections of these bundles. It is a local 1-linear oper-
ator, and in a coordinate patch (U, x
1
, . . . , x
n
) it is given by
df =
f
x
1
dx
1
+ +
f
x
n
dx
n
.
Consequently, relative to the trivialization for T
M given by the chart, this

operator is represented by the dierential operator
_

x
1
, . . . ,

x
n
_
.
Thus, the exterior dierential is a rst order dierential operator.
3) Connections. Let E be a vector bundle, a connection on E and X a
vector eld on M, then
X
: (E) (E) is a dierential operator. Triv-
ially, this is K-linear and local. To check that it is locally a dierential operator,
we might just a well assume E to be trivial. Furthermore in order to keep
the calculations as simple as possible, we we impose the additional assump-
tion, that E is 1-dimensional i.e. is the trivial line bundle M K. In this case
(E) = C
(M) (K-valued smooth functions) and the connection is given on a

section f C
(M) by (cf. (10.2))

f = df +fv
where v is a 1-form. Thus
X
f = Xf +v(X)f.
11.1 Dierential Operators on Manifolds 203
In a local coordinate patch (U, x
1
, . . . , x
n
) we write X =
n
i=1
X
i
i
, then
X
f =
n
i=1
X
i
(
i
f) +v(X)f
thus locally,
X
is represented by the dierential operator
X
1

x
1
+ +X
n

x
n
+v(X).
In fact, the connection itself : (E) (T
M E) is a dierential
operator. Again we assume that E = MK
m
is trivial, then the sections are of
course just m-tuples of smooth functions (the sections s
i
(p) = (p, e
i
) constitute
a global frame) and (T
M E) =
1
(M, E) consists of m-tuples of 1-forms
(the identication being s
1
(, 0, . . . , 0), s
2
(0, , 0, . . . , 0) and
so on). The connection acts on fs
1
= (f, 0, . . . , 0) in the following way
(fs
1
) = df s
1
+fv
where
v =
_
_
_
1
.
.
.
m
_
_
_ =
_
_
_
11
dx
1
+ +
1n
dx
n
.
.
.
m1
dx
1
+ +
mn
dx
n
_
_
_
1
(M, E)
is an m-tuple of 1-forms expressed in terms of the chart (U, x
1
, . . . , x
n
). Identify
df s
1
with the tuple (df, 0, . . . , 0) and we get
(fs
1
) =
_
_
_
df
.
.
.
0
_
_
_+f
_
_
_
11
dx
1
+ +
1n
dx
n
.
.
.
m1
dx
1
+ +
mn
dx
n
_
_
_
=
_
_
_
(f
11
+
f
x
1
)dx
1
+ + (f
1n
+
f
x
n
)dx
n
.
.
.
f
m1
dx
1
+ +f
mn
dx
n
_
_
_
The rst row of this vector of 1-forms corresponds to
_
f
11
+
f
x
1
_
(dx
1
s
1
) + +
_
f
1n
+
f
x
n
_
(dx
n
s
1
)
the mth row corresponds to
f
m1
(dx
1
s
m
) + +f
mn
(dx
n
s
m
)
and similarly with the other rows. Thus by choosing the trivialization of T
M
E corresponding to the local frame dx
i
s
j
we see that acts on the com-
ponents as a dierential operator.
For the remainder of this chapter suppose (unless otherwise specied) that
(M, g) is an oriented Riemannian manifold with volume form dV
g
. For a vector
bundle E over M we let
c
(E) denote the set of smooth sections of E with
compact support. If E is a Riemannian vector bundle we can dene on
c
(E)
an inner product ( [ )
E
by
(u[v)
E
:=
_
M
u(p), v(p)
p
dV
g
.
Denition 11.3 (Formal Adjoint). Let A : (E) (F) be a linear oper-
ator. If there exists an operator A
: (F) (E) such that

(Au[v)
F
= (u[A
v)
for all u
c
(E) and v
c
(F) we say that A
is the formal adjoint of A.

It is called a formal adjoint since it is not an adjoint in the Hilbert space
sense. We record some elementary properties:
Proposition 11.4. Let A : (E) (F) be an 1-linear operator, then the
following holds
1) A has at most one formal adjoint.
2) If A and B : (E) (F) have formal adjoints A
and B
, then A+B
has a formal adjoint and (A+B)
= A
+B
.
3) If A and B : (F) (H) have formal adjoints, then BA has a formal
adjoint and (BA)
= A
.
4) If A is a dierential operator a formal adjoint exists and A
is a dierential
operator.
Proof. The only non-trivial statement in this proposition is 4). We can prove
this by using a partition of unity to reduce to the trivial case.
First, if A is a usual dierential operator A =
]]k
a
(x)
over 1
n
or an
open subset hereof, integration by parts yields the existence of a formal adjoint
as well as the formula
A
u =
]]k
(1)
]]
(a
u).
If A = (A
ij
) is a dierential operator C
(U)
n1
C
(U)
n2
it is easy to see
that the formal adjoint exists and equals A
= (A
ji
) i.e. the transpose of the
matrix of formal adjoints.
Now, suppose A is a dierential operator on a manifold M, and suppose
(U
i
)
iI
is a cover of M of proper neighborhoods. Let A
Ui
denote the local rep-
resentatives of A (w.r.t. some trivializations which we dont bother to include
in the notation), then the formal adjoints A
Ui
exists as we have just seen. Fur-
thermore, if two neighborhoods U
i
and U
j
intersect non-trivially, then both A
Ui
and A
Uj
, when restricted to the intersection, are formal adjoints of A restricted
to U
i
U
j
. Thus, by uniqueness of formal adjoints, they agree on the overlap.
Consequently, there exists a dierential operator B having A
Ui
as local repre-
sentatives.
To see that this is the formal adjoint of A, let (
i
) be a partition of unity
subordinate to the cover (U
i
), then we have (by denition of the integral over
M, in fact) that
(Au[v) =
_
M
Au, v dV
g
=
iI
_
Ui
i
A
Ui
u, v dV
g
=
iI
_
Ui
i
u, A
Ui
v dV
g
=
_
M
u, Bv dV
g
= (u[Bv)
from which we conclude that B is the formal adjoint of A.
11.2 The Principal Symbol 205
Example 11.5. 1) Exterior Derivative. On a compact, oriented Riemannian
manifold it is well-known that we have a map d
:
k
(M)
k1
(M) given
by
d
= (1)
n(k+1)+1
d
where is the Hodge star operator and n is the dimension of M. This map
satises
(d
[) = ([d)
for all
k
(M) and
k1
(M) where ( [ ) is the inner product on
k
(M)
induced from the metric on M. Thus d
is the formal adjoint of d.

2) Vector elds. First we record the following formula for u, v C
(M) (we
use L
X
to denote the Lie derivative on both functions and dierential forms):
L
X
(uv dV
g
) = (L
X
u)v dV
g
+u(L
X
v)dV
g
+uv(L
X
dV
g
)
= (L
X
u)v dV
g
+u(L
X
+ div X)v dV
g
.
The Cartan formula states that L
X
= i
X
d + d i
X
where i
X
:
k
(M)
k1
(M) is contraction with X. Since uv dV
g
is already a top-dimensional form,
L
X
(uv dV
g
) = di
X
(uv dV
g
) and substituting this into the equation above yields
di
X
(uv dV
g
) = (L
X
u)v dV
g
+u(L
X
+ div X)v dV
g
. (11.2)
Integration on the left-hand side gives 0 by Stokes Theorem (since our manifold
by assumption has no boundary). Hence
_
M
(L
X
u)v dV
g
=
_
M
u(L
X
div X)v dV
g
,
i.e. L
X
= L
X
div(X), and this is again a dierential operator on the trivial
bundle M 1.
3) Connections. Let be a metric connection on a bundle E, i.e. for any
u, v (E) and X A(M) it satises
L
X
u, v =
X
u, v +u,
X
v.
Upon integrating we get
_
M
X
u, v+u,
X
vdV
g
=
_
M
L
X
(u, v)dV
g
=
_
M
(div X)u, vdV
g
=
_
M
u, (div X)vdV
g
,
where the second identity follows from (11.2) (simply replace u by u, v and v
by 1). Consequently
X
=
X
div X.
11.2 The Principal Symbol
Let U 1
n
be some open set and let A =
]]k
a
(x)
be a usual dierential
operator of order k, acting on functions f C
(U) (which can be either real- or

complex-valued but in this setting are assumed complex-valued). The principal
symbol , or just symbol
1
(A) of A is the map (A) : U 1
n
C obtained
by formally replacing
by i
]]
in the top degree terms of A:

(A)(x, ) := i
k
]]=k
a
(x)
.
More generally, if we have a dierential operator A : C
(U)
n1
C
(U)
n2
i.e. an n
2
n
1
matrix of dierential operators
(A)
ij
=
]]k
a
ij
(x)
as in (11.1), we have again a principal symbol (A) though this time a matrix-
valued one:
((A))
ij
(x, ) := i
k
]]=k
a
ij
(x)
.
We see that for a given x this is a homogenous polynomial in with coecients
in Mat(n
2
, n
1
), the set of n
2
n
1
-matrices.
In order to motivate the denition of the symbol for a dierential operator
on an arbitrary manifold below, lets give the symbol map a more formal guise:
let E = U C
n1
and F = U C
n2
denote trivial bundles, then A is an operator
(E) (F). Let (x,
0
) U 1
n
be xed, then the complex matrix (x,
0
)
may be viewed as a linear map E
x
E
x
. Identifying U1
n
with the cotangent
bundle : T
U U (shortly, it will become apparent why we have identied

it with the cotangent bundle and not the tangent bundle) the remarks above
translate into the following: For each point (x, ) T
U we have a linear map

E
x
F
x
or (since, by construction of the pullback bundle, the ber (
E)
(x,)
equals the ber E
x
) a linear map (
E)
(x,)
(
F)
(x,)
. Thus we see that
we may view the symbol as a smooth section of the bundle Hom(
E,
F), or
equivalently as a bundle map
F.
With this in mind, we generalize to a dierential operator A : (E)
(F) over an arbitrary manifold M. Let : T
M M denote the cotangent

bundle, and let
E and
F be the pullbacks over T
M. Let U be a proper
neighborhood and let
E
and
F
be trivializations of E and F over U.
Denition 11.6 (Principal Symbol). Dene the principal symbol (or just
symbol ) (A) of the dierential operator A to be the smooth bundle map
F corresponding to the smooth section (also denoted (A)) of the

bundle Hom(
E,
F) given locally by
(A)(
p
) = (
F
[
p
)
1
(A
U
)(p, )
E
[
p
: E
p
F
p
(11.3)
where
p
is an element in T
p
M, = (
1
, . . . ,
n
) are the coordinates relative to
the basis (dx
i
) and A
U
is the local representative of A.
There are several things we need to check in order to verify that this is well-
dened. First of all we note that (
E)
p
= E
p
and (
F)
p
= F
p
, so (A) is
indeed a section of Hom(
E,
F) as stated. Secondly we need to check that

it is independent of the choice of trivializations. To this end let
E
and
F
be
dierent trivializations over U. Then
(A)(
p
) = (
F
[
p
)
1
(A
U
)(p, )
E
[
p
= (
E
[
p
)
1
E
[
p
(
F
[
p
)
1
(A
U
)(p, )
E
[
p
(
E
[
p
)
1
E
[
p
= (
E
[
p
)
1
g
F
(p)(A
U
)(p, )g
E
(p)
1
E
[
p
= (
E
[
p
)
1
(A
U
)(p, )
E
[
p
.
1
In analysis on 1
n
one distinguishes between the symbol and the principal symbol. In this
manifold setting, however, only the principal part of the symbol makes coordinate independent
sense, and thus we will use the terms symbol and principal symbol interchangeably.
Finally we need to check that it is independent of the choice of smooth
coordinates (x
1
, . . . , x
n
) on U. Assume we have two charts = (x
1
, . . . , x
n
)
and = ( x
1
, . . . , x
n
). We let D(
1
) denote the GL(n, 1)-valued function
p ( x
j
/x
i
)(p) and A
U
and

A
U
denote the local representatives of A with
respect to the charts and . It is well-known that the corresponding principal
symbols are related by
2
(
U
)(p, ) = (A
U
)
_
p, D(
1
)(p)
T
_
(11.4)
for 1
n
. If
p
=
1
dx
1
p
+ +
n
dx
n
p
=

1
d x
1
p
+ +

n
d x
n
p
, we know that
the coecients are related by (
1
, . . . ,
n
) = D(
1
)(p)
T
(
1
, . . . ,
n
), i.e. by
exactly the same transformation rule as in (11.4). This shows that the expression
(11.3) is independent of the chart, and thus that it is well-dened.
Proposition 11.7. The symbol possesses the following properties:
1) If a, b C and A and B are dierential operators of the same order, then
(aA+bB) = a(A) +b(B).
2) If A : (E) (F) and B : (F) (H) are dierential operators
then
(BA)(
p
) = (B)(
p
) (A)(
p
).
3) If A is of order k then
(A
)(
p
) = (1)
k
(A)(
p
)
(11.5)
where (A)(
p
)
: F
p
E
p
is the adjoint of (A)(
p
).
Proof. 1) is obvious. To prove 2) consider rst two local representatives A
U
=
]]k
a
and B
U
=
]]
b
. Then, of course
A
U
B
U
v =
]]k,]]
a
(b
.v).
By the Leibniz rule, the term of highest order is just
]]=k,]]=
a
+
.
From this we deduce that (A
U
B
U
) = (A
U
)(B
U
). Globally we have
(AB)(
p
) = (
H
[
p
)
1
(A
U
B
U
)(p, )
E
[
p
= (
H
[
p
)
1
(A
U
)(p, )(B
U
)(p, )
E
[
p
= (
H
[
p
)
1
(A
U
)(p, )
F
[
p
(
F
[
p
)
1
(B
U
)(p, )
E
[
p
= (A)(
p
)(B)(
p
).
Lets calculate the symbol of some of the dierential operators discussed above
Example 11.8. 1) Vector Fields. We claim that (L
X
)(
p
) = i
p
(X). Let
(U, x
1
, . . . , x
n
) be a chart on M and write X =
j
X
j
j
, thus locally:
L
X
f =
n
j=1
X
j
f
x
j
= i
n
j=1
X
j
D
j
f.
2
See for instance [GG] (8.5).
Formally replacing
j
with i
j
gives the local symbol
(L
X
)[
U
(
1
, . . . ,
n
) = i
j=1
X
j
j
.
If
p
is a covector over a point i p U with
p
=
n
j=1
j
dx
j
p
, then
p
(X) =
j
X
j
j
, i.e. (L
X
)(
p
) = i
p
(X).
2) Exterior Derivative. Lets calculate the symbol of the exterior derivative
d :
1
(M)
2
(M). To keep the calculations to a manageable level, lets
assume M to have dimension 3. Let (U, x
1
, x
2
, x
3
) be a chart on M and write
=
3
i=1
i
dx
i
. It is well-known that
d =
i<j
_
j
x
i

i
x
j
_
dx
i
dx
j
and from this it is easily seen that the exterior derivative, relative to the local
frames (dx
1
, dx
2
, dx
3
) and (dx
1
dx
2
, dx
1
dx
3
, dx
2
dx
3
), is represented by
the matrix
_
_
/x
2
/x
1
0
/x
3
0 /x
1
0 /x
3
/x
2
_
_
. (11.6)
The claim is that the symbol of d is given by (d)(
p
) = ie(
p
), where e(
p
) is
the map
p
M
p
M given by
p

p
p
. To verify this, note that
if
p
=
3
i=1
i
dx
i
p
and T
p
M
p
=
3
j=1
j
dx
j
p
then
p

p
=
i<j
(
i
j

j
i
)dx
i
p
dx
j
p
i.e. the linear map e(
p
) is represented, in the basis (dx
i
p
), by the matrix
_
_
2

1
0
3
0
1
0
3

2
_
_
and comparing to (11.6) we see that the symbol of d is indeed equal to e(
p
) as
stated above. This can easily be generalized to manifolds of arbitrary dimension
and to d as an operator d :
(M)
(M).
3) Hodge-de Rham Operator. Consider the operator d+d
(M)
(M).
By Proposition 11.7 we need only calculate the adjoint of the linear map e(
p
)
in order to obtain (d
)(
p
). One can check that the adjoint of e(
p
) is c(
p
)
where c(
p
) :
p
M
1
T
p
M is contraction with the metric dual of
p
.
Thus by Proposition 11.7
(d +d
)(
p
) = (d)(
p
) (d)(
p
)
= ie(
p
) ic(
p
).
4) Hodge Laplacian. Since d d = d
= 0 we see that
(d +d
)
2
= dd
+d
d = :
(M)
(M),
i.e. the Hodge-de Rham operator is a square-root of the Hodge Laplacian. By
Proposition 11.7
()(
p
) = (d)(
p
)(d)(
p
)
(d)(
p
)
(d)(
p
)
=
_
ie(
p
)(ic(
p
)) + (ic(
p
))ie(
p
)
_
= e(
p
)c(
p
) +c(
p
)e(
p
).
Thus, to compute the symbol, we need only compute e(
p
)c(
p
) +c(
p
)e(
p
). To
do so, let
1
, . . . ,
n
be an orthonormal basis for T
p
M and a 1 arbitrary,
then
_
e(a
1
)c(a
1
) +c(a
1
)e(a
1
)
_
(
j1

j
k
)
= a
2
1
c(
1
)(
j1

j
k
) +a
2
c(
1
)(
1

j1

j
k
).
If j
1
= 1 then the second term vanishes and the rst term equals a
2
j1

j
k
. If
j
1
,= 1, then the rst term vanishes, while the second term equals a
2
j1

j
k
.
From this and linearity we get in general the following Cartan formula
e(
p
)c(
p
) +c(
p
)e(
p
) = |
p
|
2
. (11.7)
In conclusion we obtain
()(
p
) = e(
p
)c(
p
) +c(
p
)e(
p
) = |
p
|
2
.
5) Connections and the Connection Laplacian. Let s
1
, . . . , s
m
be a local frame
for a bundle E and let be a connection. A local expression for the order 1
terms of (f
1
s
1
+ +f
m
s
m
) is then
m
i=1
n
j=1
f
i
x
j
dx
j
s
i
.
As in 3) we see that the symbol of is then ()(
p
) = i
p
, i.e. tensoring with
i
p
from the left. The adjoint of
p
is contraction with
p
, therefore (
)(
p
) =
ic(
p
). Since
p
(
p
) = |
p
|
2
we get by Proposition 11.7 that (
)(
p
) =
|
p
|
2
. The operator
is known as the connection Laplacian.

Inspired by 4) a second order dierential operator A with symbol (A)(
p
) =
|
p
|
2
is called a generalized Laplacian. The Hodge Laplacian and the connection
Laplacian are examples of generalized Laplacians.
A rst order dierential operator A for which A
A and AA
are generalized
Laplacians is called a Dirac type operator (the reason for this will become ap-
parent in Section 11.3). Examples of Dirac type operators include the Hodge-de
Rham operator and the connection.
Denition 11.9 (Elliptic Operator). A dierential operator A is called el-
liptic if the symbol map (A) :
F is a pointwise isomorphism o the

zero section of T
M.
By the results of Example 11.8 we see that the Hodge Laplacian is an elliptic
operator. In fact any generalized Laplacian as well as formally self-adjoint Dirac
type operators are elliptic. Thus the Hodge-de Rham operator is elliptic whereas
a connection need not.
The composition of two elliptic operators is again elliptic. If A is elliptic of
order k and K is a dierential operator of order strictly less than k, then as
(A+K) = (A), the operator A+K is elliptic as well.
Elliptic operators have a lot of very beautiful analytic properties. In order
to shed some light on these we have to introduce Sobolev spaces. This is the
topic of Section 11.4. Before delving into that we want to devote a section to
the description of a type of operator which plays an immensely important part
in modern geometry: the Dirac operator.
11.3 Dirac Bundles and the Dirac Operator
In this section we let M denote an oriented Riemannian manifold. No compact-
ness condition is imposed unless specied.
Let E be a real oriented Riemannian vector bundle over M of rank n (often
we will take it to be the tangent bundle). Thus we can construct its oriented
frame bundle P
SO
(E), the principal SO(n)-bundle over M whose ber at p M
consists of oriented orthonormal bases for the vector space E
p
. We want to
construct the so-called Cliord bundle over E, i.e. an algebra bundle over M
whose ber at x is isomorphic to the Cliord algebra Cl(E
x
). The construction is
accomplished as an associated bundle in the following way: Consider the Cliord
algebra Cl
0,n
, the Cliord algebra over 1
n
with the usual negative denit inner
product. We have a representation of SO(n) on Cl
0,n
: any A SO(n) viewed
as a linear map 1
n
1
n
preserves the inner product, and hence induces an
algebra homomorphism

A : Cl
0,n
Cl
0,n
, so the representation is given as
(A) =

A.
Denition 11.10 (Cliord Bundle). The Cliord bundle of E is the associ-
ated bundle
Cl(E) := P
SO
(E)
Cl
0,n
.
Elements in Cl(E) are equivalence classes [p, ], where p P
SO
(E) and
Cl
0,n
and the equivalence relation on P
SO
Cl
0,n
is (p, ) (p A, (A)). The
projection map is : Cl(E) M, [p, ] (p) where : P
SO
(E) M is
the projection in the frame bundle. The vector space structure on the bers is
given by
a[p, ] +b[p,
t
] = [p, a +b
t
]
(note, by transitivity of the right SO(n)-action on each ber in P
SO
(E) we can
always assume the ps to be equal). Similarly, the algebra structure is given by
[p, ] [p,
t
] = [p,
t
],
and the identity element is [p, 1]. It is easy to check that these operations are
well-dened. Since
2
= ||
2
1 (by denition of the Cliord algebra), we get
[p, ][p, ] = [p, ] = [p, ||
2
1] = ||
2
[p, 1],
thus each ber is indeed a Cliord algebra of type (0, n). Thus Cl(E)
x
(the
ber in the Cliord bundle) is isomorphic to Cl(E
x
) (the Cliord algebra of the
vector space E
x
).
Observe that we have 1
n
Cl
0,n
and that 1
n
is an -invariant subspace and
that (A)[
R
n = A, i.e. restricted to this invariant subspace is just the dening
representation of SO(n) on 1
n
. We write it as id. But this means that we have
the subbundle P
SO
(E)
id
1
n
= E sitting inside Cl(E) (elements in E Cl(E)
are characterized by being of the form [p, v] where v 1
n
). In particular we
may view (E) as sitting inside (Cl(E)).
For the purpose of studying spinor bundles as we will do later in this section,
we need another description of the Cliord bundle. Assume that the oriented
Riemannian vector bundle E has a spin structure : P
Spin
(E) M with dou-
ble covering bundle map : P
Spin
(E) P
SO
(E). Consider the representation
Ad : Spin(n) Aut(Cl
0,n
) given by
Ad(g) = gg
1
(recall that Spin(n) sits inside Cl
0,n
, so the multiplication makes sense). Let
: Spin(n) SO(n) denote the double covering, then the following diagram
11.3 Dirac Bundles and the Dirac Operator 211
commutes (simply because Ad(g) is the unique extension of (g) to Cl
0,n
):
Spin(n)
Ad
Aut(Cl
0,n
)
SO(n)
l
l
l
l
l
l
l
l
l
l
l
l
l
l
From the principal Spin(n)-bundle P
Spin
(E) and the representation Ad we can
form the associated bundle P
Spin
(E)
Ad
Cl
0,n
.
Lemma 11.11. The map : P
Spin
(E)
Ad
Cl
0,n
Cl(E) = P
SO
(E)
Cl
0,n
given by [p, ] [(p), ] is a well-dened smooth algebra bundle isomorphism.
Proof. It is well-dened, since
([p g
1
, Ad(g)]) = [(p g
1
), Ad(g)] = [(p) (g)
1
, ((g))]
= [(p), ] = ([p, ]),
the third identity is a consequence of equivariance of and of the commut-
ing diagram just above. Restricted to the ber over x, the map is an algebra
isomorphism (we skip checking linearity):
x
([p, ][p,
t
]) =
x
([p,
t
]) = [(p),
t
] = [(p), ] [(p),
t
]
=
x
([p, ])
x
([p,
t
]).
It is injective, for if 0 =
x
([p, ]) = [(p), ], then must be 0, but then [p, ] =
0. Also
x
is surjective, since is. Thus is an algebra bundle isomorphism.
Denition 11.12 (Dirac Bundle). Let E be a real Riemannian vector bundle
over M and a metric connection. A complex vector bundle S over M, which
is a left Cl(E)-module, i.e. for each x M there is a representation of the
algebra Cl(E)
x
on S
x
is called a Dirac bundle, provided it is equipped with a
ber metric , and a compatible connection

satisfying the following two
additional conditions
1) Cliord multiplication is skew-adjoint, i.e. for each x M and each V
x

E
x
and
1
,
2
S
x
:
V
x

1
,
2
+
1
, V
x

2
= 0.
2) The connection on S is compatible with the connection on E in the fol-
lowing sense:
X
(V ) = (
X
V ) +V (
X
)
for X X(M), V (E) and (S).
Example 11.13. Exterior Bundle. Lets prove that the exterior bundle
M
is a Dirac bundle.
We extend the metric g on M to a ber metric on
M in the usual way.

On TM we have the Levi-Civita connection. This can be extended to T
M upon
dening
X
= (
X
, for
1
(M). This connection on T
M is metric:
X
1
,
2
= X
1
,
2
=
X
1
,
2
+
1
,
X
= (
X
1
)
2
+
1
, (
X
2
)
=
X
1
,
2
+
1
,
X
2
.
We can extend the connection to the exterior bundle by the requirement
X
( ) = (
X
) + (
X
).
One can check that this is indeed a metric connection on this bundle, and that
satises
X
(c(Y )) = c(
X
Y ) +c(Y )(
X
) (11.8)
where c is contraction with the vector eld Y .
Next, we need the exterior bundle to be a bundle of Cliord modules, that
is for each point x M we need an action of the Cliord algebra Cl(T
x
M) on
x
M. To this end it is fortunate that we have quantization map vector space
isomorphism Q :
x
M

Cl(T
x
M) given by
1

k

S
k
sign
(1)

(2)
.
We simply dene the Cliord action by
Cl(T
x
M)
x
M (v, ) v := Q
1
(vQ()).
If v happens to be an element of T
x
M Cl(T
x
M) we have an explicit formula
for this product
3
v = v
+c(v). (11.9)
As has been mentioned previously, the adjoint of v
is c(v) and so by (11.9)

condition 1 in Denition 11.12 is satised:
v , = v
, +c(v),
= , c(v) +, v

= , v
+c(v)
= , v .
Finally we need to check compatibility of the connection with the Cliord action.
Let X and Y be vector elds on M, and view Y as a section of Cl(M) (whose
values happen to be in the subbundle TM Cl(M)). Using (11.9) and (11.8)
and the dening property of the connection on
M we get
X
(Y ) =
X
(Y
+c(Y ))
= (
X
Y
) +Y

X
+
X
(c(Y ))
= (
X
Y )
+Y

X
+c(
X
Y ) +c(Y )(
X
)
= (
X
Y ) +Y (
X
).
Thus we have proved that the exterior bundle is a Dirac bundle.
Spinor Bundle. The next and single most example of a Dirac bundle is the
spinor bundle. The setup is as follows: let E be a real oriented Riemannian vector
bundle over a Riemannian manifold M and suppose it has a spin structure, i.e. its
oriented frame bundle lifts to a principal Spin(n)-bundle P
Spin
(E) M with
a double covering map : P
Spin
(E) P
SO
(E). Let
n
: Spin(n) Aut(
n
)
be the spinor representation of Spin(n), then the spinor bundle is the associated
bundle
S(E) := P
Spin
(E)
n

n
.
Smooth sections of this bundle are called spinor elds or Dirac spinor elds.
3
See [Lawson, Michelson] Proposition 3.9. Note the minus sign in their formula. This is
due to their convention v
2
= v
2
which is dierent from ours.
Our present task is to show that it is a Dirac bundle. First we equip it with
an action of the Cliord bundle Cl(E) = P
Spin
(E)
Ad
Cl
0,n
. Consider on the
threefold product P
Spin
(E) Cl
0,n
n
the equivalence relation
(p, , v) (p g
1
, Ad(g),
n
(g)v)
for any g Spin(n). Elements in the quotient space P
Spin
(E) Cl
0,n
n
/
are denoted [p, , v]. As in the proof of Lemma 11.11 one can show that the map
Cl(E) S(E) P
Spin
(E) Cl
0,n
n
/
given by ([p, ], [p, v]) [p, , v] is a well-dened bundle isomorphism. This
allows us to dene the Cliord action in the following way: Dene : P
Spin
(E)
Cl
0,n
n
P
Spin
(E)
n
by (q, , v) = (q,
n
()v) (where
n
: Cl
0,n

Aut(
n
) is the spin representation of the Cliord algebra) and note that for
any g Spin(n) it makes the following diagram commutative
P
Spin
(E) Cl
0,n
P
Spin
(E)
n
g
P
Spin
(E) Cl
0,n
P
Spin
(E)
n
where the rst vertical map is (p, , v) (p g
1
, Ad(g),
n
(g)v) and the
second is (p, v) (pg
1
,
n
(g)v). Thus it induces a map : Cl(E)S(E)
S(E), the desired Cliord action. It is given explicitly by the formula
[p, ] [p, v] := ([p, ], [p, v]) = [p,
n
()v].
Next we want to give S(E) a metric. Inside Cl
0,n
we have the nite group
G
n
:= e
i1
e
i
k
[ 1 k n, 1 i
1
< < i
k
n
(where e
1
, . . . , e
n
is some orthonormal basis for 1
n
). Restricting the spin
representation
n
of Cl
0,n
to G
n
gives a representation of G
n
on
n
, also denoted
n
. By a well-known result from representation theory, there exists an inner
product ,
n
on
n
relative to which
n
is a unitary representation, i.e.
n
(e
i1
e
i
k
)v,
n
(e
i1
e
i
k
)w
n
= v, w
n
.
(To make the notation in the following less cumbersome, we will simply write
the action of
n
() on v as v.) If =
n
i=1
a
i
e
i
is a unit vector in 1
n
Cl
0,n
then
n
(e) is a unitary operator as well: First we observe
e
i
v, e
j
w
n
= e
j
(e
i
v), e
2
j
w
n
= (e
j
e
i
) v, w
n
= (e
i
e
j
) v, w
n
= (e
2
i
e
j
) v, e
i
w
n
= e
j
v, e
i
w
n
,
and from this we get
v, w
n
=
_
n
i=1
a
i
e
i
v,
n
j=1
a
j
e
j
w
_
n
=
n
i=1
a
2
i
e
i
v, e
j
w
n
+
i,=j
a
i
a
j
e
i
v, e
j
w
n
=
n
i=1
a
2
i
v, w
n
+
i<j
a
i
a
j
_
e
i
v, e
j
w
n
+e
j
v, e
i
w
n
_
= v, w
n
.
Since Spin(n) is generated by unit vectors, we see immediately that
n
is a
unitary representation w.r.t. this inner product.
Furthermore, for any 1
n
, unit vector or not,
n
() is a skew-adjoint map:
v, w
n
=

||
( v),

||
w
n
=
1
||
2
2
v, w
n
= v, w
n
.
Note that this implies that
n
() is skew-adjoint, when Cl
1
0,n
(the odd part
of Cl
0,n
) and that
n
() is self-adjoint when Cl
0
0,n
(the even part of Cl
0,n
).
In particular,
n
=
n
[
Spin(n)
is self-adjoint.
We can readily extend this inner product to a ber metric on S(E), simply
by dening
[p, v], [p, w] := v, w
n
.
This is well-dened due to unitarity of
n
(g):
[p g
1
,
n
(g)v], [p g
1
,
n
(g)w] =
n
(g)v,
n
(g)w
n
= v, w
n
= [p, v], [p, w].
Checking condition 1 in the denition of a Dirac bundle is not hard: Let V
x

E
x
Cl(E)
x
and
1
,
2
S
x
(E). We have presentations V
x
= [p, v] and
i
=
[p, w
i
] where v 1
n
Cl
0,n
, p P
Spin
(E) and w
i

n
, and hence:
V
x

1
,
2
= [p, v] [p, w
1
], [p, w
2
] = [p, v w
1
], [p, w
2
]
= v w
1
, w
2
n
= w
1
, v w
2
n
= [p, w], [p, v] [p, w
2
]
=
1
, V
x

2
.
Next, we want to equip S(E) with a connection. Consider a connection on
P
SO
(E) given either as a horizontal tangent distribution on the total space
or as a so(n)-valued 1-form on P
SO
(E). This connection can be lifted to
a connection on the spin bundle P
Spin
(E) in the following way. We have the
double covering map : P
Spin
(E) P
SO
(E) and the induced map
:
T
q
P
Spin
(E) T
(q)
P
SO
(E) is an isomorphism. Therefore we can dene a
horizontal subspace H
q
P
Spin
(E) by
1
(H
(q)
P
SO
(E)). This is easily seen to
be a connection. Alternatively we can dene a connection 1-form on P
Spin
(E)
as follows: :=
1
. This is a connection 1-form: we only have to check

the usual two requirements. The rst one:

g
=
1
=
1
(g)
=
1
Ad((g
1
)) =
1
Ad((g
1
))(
)
= Ad(g
1
)
1
= Ad(g
1
) .
For the second one let p P
Spin
(E) and denote by
p
: Spin(n) P
Spin
(E)
the map g p g and note that
p
=
(p)
, where
(p)
is the similar
map on the bundle P
SO
(E). Then:

p
(
p
)
=
1
)
p
(
p
)
=
1

(p)

(
p
)
=
1

(p)
(
(p)
)
=
1
= id
spin(n)
.
This denes the same connection as before, for one can easily check that ker( )
p
=
H
p
P
Spin
(E).
Now we can transfer this connection to the spin bundle as done in Section
10.2, obtaining a connection

, on S(E). In the special case where E = TM is
the tangent bundle and on P
SO
(M) is the Levi-Civita connection, the lifted
connection, both on P
Spin
(E) and

on S(E), is called the spin connection.
Proposition 10.11 gives us local expressions for an induced connection on an
associated bundle. Lets see what it looks like in the case of the spin connection.
Let t
: U
P
Spin
(E) be a local section of the spin bundle and put s
:=
t
. This is a local section of the frame bundle P

SO
(E). Let
:= t
and A
:= s
denote the local gauge potentials of the connection on P

SO
(E), resp. on
P
Spin
(E). Then we have
= t
= t
(
1
) =
1
=
1
=
1
,
so the gauge potentials are related in the nicest possible way.
Putting this into (10.5) we obtain
(
X
)
(x) = X
x
(
n
)
(X
x
))
(x)
= X
x
n
(
1
(A
(X
x
)))
(x)
for x U
(recall that the action of (

n
)
is just
n
itself, cf. Corollary 8.30).
Now we pick the usual basis (B
ij
)
i<j
for so(n) and write the so(n)-valued 1-form
A
in terms of this basis: A
i<j
A
ij
B
ij
. Then (remembering Proposition
8.21) we get
(
X
)
(x) = X
x
n
_
i<j
A
ij
(X
x
)B
ij
_
_
(x)
= X
x
n
_
1
2
i<j
A
ij
(X
x
)e
i
e
j
_
(x)
= X
x
1
2
i<j
A
ij
(X
x
)e
i
e
j

(x).
Now, let us show compatibility of the spin connection with the ber met-
ric. We will use the local expression above, for note that since locally (x) =
[s
(x),
(x)] (for some section s
: U
P
Spin
(E)) we get by denition of
the ber metric
(x),
t
(x) =
(x),
t
(x)
n
for x U
. Thus we get
X
(x),
t
(x) =
_
X
x
1
2
i<j
A
ij
(X
x
)e
i
e
j

(x) ,
t
(x)
_
n
= X
x
,
t
(x)
n

1
2
i<j
A
ij
(X
x
)e
i
e
j

(x),
t
(x)
n
.
Note that
e
i
e
j

(x),
t
(x)
n
= e
j

(x), e
i

t
(x)
n
=
(x), e
j
e
i

t
(x)
n
=
(x), e
i
e
j

t
(x)
n
,
and therefore
X
(x),
t
(x) +(x),
X
t
(x) = X
x
,
t
(x)
n
+(x), X
x
n
.
As mentioned X
x
should be interpreted componentwise, i.e. pick a complex

basis v
1
, . . . , v
N
for
n
(of course N = 2
n
2
) and write
=
N
i=1
,i
v
i
where
,i
, the ith component of
, are complex valued functions on U
, then
X
x
=
N
i=1
(X
x
,i
)v
i

n
.
Since we have a metric in play, it would be wise of us to assume the basis
v
1
, . . . , v
N
to be orthonormal. Then
X
x
,
t
(x)
n
+
(x), X
x
n
=
N
i=1
(X
x
,i
)
t
,i
(x) +
N
i=1
,i
(x)X
x
t
,i
= X
x
_
N
i=1
,i
t
,i
_
= X
x
,
t
n
.
Thus we have proved that the spin connection is compatible with the metric.
Finally, we need to check condition 2 in the denition of a Dirac bundle, that
is
X
(Y )(x) = (
X
Y )(x) +Y
x
(
X
(x)) (11.10)
for each x M. Again we use the local expressions, i.e. we consider a cover
(U
) which are domains of trivializations of both TM and E. Let
denote
the trivializations of the tangent bundle (we may assume it to preserve the
metric on M, i.e.
x
: T
x
M

1
m
is an isometry). Since

X
is C-linear and
satises the Leibniz rule, it is sucient to verify the above condition for Y = E
k
where E
k
(x) =
1
(x, e
k
) are local orthonormal vector elds.
But rst, recall formula (8.3) and replace in that X by

i<j
A
ij
(X
x
)e
i
e
j

spin(n) to get
_
i<j
A
ij
(X
x
)e
i
e
j
_
e
k
= e
k
_
i<j
A
ij
(X
x
)e
i
e
j
_
+
i<j
A
ij
(X
x
)e
i
e
j
_
e
k
= e
k
_
i<j
A
ij
(X
x
)e
i
e
j
_
+ 2
i<j
A
ij
(X
x
)B
ij
e
k
. (11.11)
Note that concatenation here means multiplication inside the Cliord algebra,
and not the Cliord action. Recall also formula (10.7) for the local form of the
Levi-Civita connection. It will be used in the following calculations:
(
X
(E
k
))
(x) = X
x
(e
k

)
_
1
2
i<j
A
ij
(X
x
)e
i
e
j
_
e
k

(x)
= e
k
(X
x
) e
k
_
1
2
i<j
A
ij
(X
x
)e
i
e
j
_

(x)
i<j
(A
ij
(X
x
)B
ij
e
k
)
(x)
= e
k
(
X
)
(x) + (
X
E
k
)
(x)
(x)
and this is precisely the local form of the right-hand side of (11.10). For the
rst identity we used that (E
k
)
= e
k

and in the second we used (11.11)

as well as the fact, that e
k
is a linear map, and thus commutes with X
x
. This
veries condition 2 and hence we have shown that the spinor bundle S(E) is a
Dirac bundle.
Denition 11.14 (Dirac Operator). Let S be a Dirac bundle. The Dirac
operator / is then dened as the composition
(S)
(T
M S)

(TM S) (S) (11.12)
where the last map is Cliord multiplication X(M) (S) (S).
Lemma 11.15. The Dirac operator is a rst order dierential operator, and
given a local orthonormal frame E
1
, . . . , E
n
for TM over U M, the Dirac
operator takes the local form
/ [
U
=
n
j=1
E
j
(
Ej
). (11.13)
Proof. Let
i
= E
i

1
(U) be the metric dual of E
i
. Suppose

=
t
for some local section
t
of S, then / =

t
. Locally we have =
i
,
and
i
E
i
. Thus

Ei
=
i
t
and therefore
/ =

t
=
n
i=1
(
i
E
i
)
t
=
n
i=1
E
i
(
i
t
) =
n
i=1
E
i
(
Ei
).
From this formula, and the fact that

Ei
is a rst order dierential operator, it
is evident that / is also a rst order dierential operator.
As a dierential operator, we can compute the symbol of the Dirac operator:
Proposition 11.16. Let
x
T
x
M, and let
x
: S
x
S
x
be Cliord multi-
plication with the metric dual
x
of
x
. Then we have
(/ )(
x
) = i
x
and (/
2
)(
x
) = |
x
|
2
.
Thus /
2
(called the Dirac Laplacian) is a generalized Laplacian and hence both
/ and /
2
are elliptic.
Proof. Locally / =
j
E
j
Ej
. Choose normal coordinates (x
1
, . . . , x
n
) cen-
tered at p, such that
j
[
x
= E
j
(x) for all i. Pick a trivialization of S around x,
then by Example 11.2

Ej
acts componentwise as d
j
plus lower order terms.
Thus, at x, the rst order terms of / act as
j
E
j
(x)
j
[
x
. Let
1
, . . . ,
n
be
the metric duals of the E
i
s and decompose
x
=
j

j
j
. By formally replacing
j
by i
j
we get
(/ ) = i
n
j=1
E
j
j
= i
x
.
For /
2
, Proposition 11.7 says
(/
2
)(
x
) = (/ )(
x
)(/ )(
x
) =
x
= |
x
|
2
= |
x
|
2
.
Cliord multiplication by
x
has as inverse multiplication by

x
|
x
|
2
(as long as
x
,= 0), thus the Dirac operator is elliptic.
Proposition 11.17. The Dirac operator is formally self-adjoint, i.e. for
1
,
2

c
(S) (sections of S with compact support)
(/
1
[
2
) = (
1
[ /
2
). (11.14)
Proof. First we consider the berwise inner product /
1
,
2
x
at some point
x M. Let E
1
, . . . , E
n
be an orthonormal frame such that
Ej
E
j
(x) = 0 at
this particular x. Then
/
1
,
2
x
=
n
j=1
E
j

Ej
1
,
2
x
=
n
j=1
Ej
1
(x), E
j
(x)
2
(x)
=
n
j=1
E
j
(x)
1
(x), E
j
(x)
2
(x) +
n
j=1
1
(x), (
Ej
E
j
)(x)
2
(x)
+
n
j=1
1
(x), E
j
(x)

Ej
2
(x)
=
n
j=1
E
j
(x)
1
, E
j

2
+
n
j=1
1
(x), E
j
(x)

Ej
2
(x).
Secondly, we need a formula involving the divergence. Let X be the global
vector eld given uniquely by
X, Y =
1
, Y
2
for all Y X(M). Then at x (the rst identity following from [Lee] Problem
5-6)
(div X)
x
=
n
j=1
Ej
X, E
j
x
=
n
j=1
_
E
j
(x)X, E
j
X(x),
Ej
E
j
(x)
_
=
n
j=1
E
j
(x)X, E
j
=
n
j=1
E
j
(x)
1
, E
j

2
. (11.15)
Plugging this into the expression above we get
/
1
,
2
x
= (div X)
x
+
1
, /
2
x
,
and as x was chosen arbitrarily this holds for any x. Since M has no boundary,
it is a consequence of the divergence formula that integration of this identity
yields (11.14)
Now it is clear, that /
2
= /
/ = / /
, and this explains the term Dirac type

operator introduced in a previous section.
Example 11.18. Dirac Operator of the Exterior Bundle. In this example we
calculate the Dirac operator of the Dirac bundle
M and show that this

operator is none other than the Hodge-de Rham operator.
Let E
1
, . . . , E
n
denote a local orthonormal frame for TM. We want to show
the following local identities
d =
n
j=1
E
j

Ej
(11.16a)
d
=
n
j=1
c(E
j
)
Ej
. (11.16b)
First we note that the expressions are dened independently of the choice of
orthonormal frame. To be more precise, suppose

E
j
=
i
A
i
j
E
j
, is another or-
thonormal frame over the same neighborhood (i.e. A
i
j
is an O(n)-valued smooth
function), then
n
j=1
Ej
=
n
i,j,k=1
A
i
j
E
i
A
k
j
E
k
=
n
i,j,k=1
A
i
j
A
k
j
E
i

E
k
=
n
i,k=1
(AA
T
)
ik
E
i

E
k
=
n
i=1
E
i

Ei
.
Thus, we can choose whatever orthonormal frame that suit our needs, it makes
no dierence.
To show the actual identity (11.16a) we have to show that the right-hand side
satises the axioms dening d, i.e. that df(X) = Xf for any vector eld X, that
d( ) = d + (1)
]]
d and that d d = 0. The rst one is easy, one
readily checks that df =
j
(E
j
f)E
j
and hence
n
j=1
E
j

Ej
f =
n
j=1
E
j
(E
j
f) =
n
j=1
(E
j
f)E
j
= df,
thus the rst axiom is satised. Axiom 2 is also satised
n
j=1
E
j

Ej
( ) =
n
j=1
E
j

_
(
Ej
) + (
Ej
)
_
=
_
n
j=1
E
j

Ej
_
+ (1)
]]

_
n
j=1
E
j

Ej
_
(induction is tacit in these calculations).
Checking axiom 3 is the hardest part. Let p U and pick an orthonormal
frame (E
j
) such that
Ej
E
j
(x) = 0 and hence also
Ej
E
j
(x) = 0. To simplify
things we check this axiom on the form fE
1
E
k
for some smooth function
f.
n
i=1
E
i

Ei
_
n
j=1
E
j

Ej
(fE
1
E
k
)
_
=
n
i,j=1
E
i

Ei
E
j

Ej
(fE
1
E
k
)
+
n
i,j=1
E
i
E
j

Ei
_
Ej
(fE
1
E
k
)
_
.
Since
Ej
E
j
(x) = 0 most of these terms are zero when evaluated at p. The
surviving terms are
k<i<j
([E
j
, E
j
]f)E
i
E
j
E
1
E
k
.
By symmetry of the Levi-Civita connection we have [E
i
, E
j
] =
Ei
E
j
Ej
E
i
and this is 0 at x. Thus the entire expression is zero when evaluated at x and
this was chosen arbitrarily so the third axiom is satised and the expression
(11.16a) is veried.
Also the second expression (11.16b) is dened independently of the choice
of orthonormal frame. Thus we can apply our standard trick: let p M be
arbitrary and pick an orthonormal frame E
1
, . . . , E
n
such that
Ei
E
j
(x) = 0.
We can then derive (11.16b) from (11.16a) if we recall that (E
j
)
= c(E
j
)
and
Ej
=
Ej
div E
j
and that (div E
j
)
x
=
Ej
E
j
(x), E
j
(x) = 0:
d,
x
=
n
j=1
E
j
(x) (
Ej
)(x), (x) =
n
j=1
Ej
(x), c(E
j
)(x)
=
n
j=1
(x),
Ej
(c(E
j
))(x) =
n
j=1
(x), c(E
j
)(
Ej
(x))
where the last identity follows from (11.8). Finally, by (11.9) we get the Dirac
operator
D =
n
j=1
E
j
(
Ej
) =
n
j=1
E
j

Ej
+c(E
j
)(
Ej
)
= d +d
and realize that it equals the Hodge-de Rham operator.

Spin-Dirac operator. We consider the spinor bundle S(E) associated to some
oriented Riemannian vector bundle E carrying a spin structure. We saw earlier
that this vector bundle is a Dirac bundle. Thus it carries a Dirac operator / D
called the spin-Dirac operator or just the Dirac operator (in [Lawson] it is called
the Atiyah-Singer operator). Thanks to Lemma 11.15 and the calculations done
in the previous example, we arrive at the following local description
( / D )
(x) =
_
m
k=1
E
k

E
k
(x) =
m
k=1
e
k
(
E
k
)
(x)
=
m
k=1
e
k

_
(E
k
)
x
1
2
n
i<j
A
ij
((E
k
)
x
)e
i
e
j

(x)
_
=
m
k=1
e
k
(E
k
)
x
1
2
i<j
A
ij
((E
k
)
x
)e
k
e
i
e
j

(x) (11.17)
where E
k
(x) =
1
(x, e
k
) are local orthonormal vector elds (we pick the triv-
ialization
such that it preserves the metric) and m = dimM

4
.
11.4 Sobolev Spaces
First of all, we will introduce L
p
-spaces. The Sobolev spaces will then be dened
as certain subspaces of these. Most of the proofs in this section will be skipped.
The reader is referred to [Nicolaescu] or [Lawson-Michelson].
Consider a smooth Riemannian vector bundle E M over an oriented Rie-
mannian manifold (M, g). Both spaces E and M are, in particular, topological
spaces, so it makes sense to talk about Borel subsets of E and M. A (rough)
section u : M E is called measurable if u
1
(B) is a Borel set in M for
each Borel set B E. The set of measurable sections is denoted /(M, E). The
metric g gives rise to a Radon measure
g
on the Borel algebra of M
5
4
Note that often (E
k
)x is written as

e
k
(x).
5
This is for the following reason: on a manifold one can integrate compactly supported
continuous n-forms. Thus on a Riemannian manifold a linear form I : Cc(M) 1 is dened
by I(f) =
M
f dVg. By the Riesz Representation Theorem there exists a unique radon
measure g such that I(f) =
M
fg.
11.4 Sobolev Spaces 221
Denition 11.19. Let 1 p < . We say that a section u /(M, E) is
p-integrable if
_
M
|u(x)|
p
E
d
g
< .
The space of equivalence classes (equivalence meaning equality
g
-a.e.) of p-
integrable sections is denoted L
p
(M, E)
A measurable section u is called locally p-integrable if for each smooth com-
pactly supported function C
c
(M) the section u is p-integrable. The space
of equivalence classes of locally p-integrable sections is denoted L
p
loc
(M, E).
Proposition 11.20. Under the norm
|u|
L
p =
__
M
|u|
p
E
d
g
_
1/p
the space L
p
(M, E) is a Banach space. For p = 2 the space is a Hilbert space
with the inner product
u, v =
_
M
u(x), v(x)
E
d
g
.
Moreover
c
(E) is dense in L
p
(M, E).
Proof. Let (v
k
) be a Cauchy sequence in L
p
(M, E) and U M a proper
neighborhood. Then (v
k
[
U
) is again a Cauchy sequence and
v
k
[
U
=
_
_
_
v
1
k
.
.
.
v
m
k
_
_
_ L
p
(U) L
p
(U).
As U is a chart neighborhood it is isometrically dieomorphic to an open set
V 1
n
and thus L
p
(U)

= L
p
(V ) which is a Banach space. Hence v
k
[
U
will
converge. By a partition of unity argument we obtain convergence of v
k
, thus
L
p
(M, E) is a Banach space.
A partition of unity argument also applies to the second assertion. Just recall
the fact that C
c
() is dense in L
p
() for any open 1
n
.
If the volume of M is nite, for instance if M is compact, p q implies
L
q
(M, E) L
p
(M, E). Thus for any M (nite volume or not) L
q
loc
(M, E)
L
p
loc
(M, E). In particular, for any p:
L
p
loc
(M, E) L
1
loc
(M, E). (11.18)
Assume now that is a metric connection on E. The tangent bundle TM has
the Levi-Civita connection, and via the bundle isomorphism TM

T
M we
can turn it into a connection on T
M simply by dening
X
= (
X
. The
tensor product of these two connections is a connection on T
M E, and we
denote it by . It maps into T
MT
ME. Again, taking the tensor product

with the Levi-Civita connection yields a connection on T
MT
ME. Hence
we have maps
(E)

(T
M E)

(T
M T
M E)
We let
k
= denote the composition of the rst k of these maps.
Furthermore we have a ber metric on T
M E, simply the tensor product

of the ber metrics on T
M

= TM and E. Specically, if
p
s
p
,
p
s
p

T
p
M E
p
, then
p
s
p
,
p
s
p
:=
p
,
p
s
p
, s
p
,
which gives the norm
|
p
s
p
| = |
p
||s
p
|
on T
p
M E
p
. Similarly, we get metrics on the higher bundles (T
M)
k
E.
We use this to construct two new types of Banach spaces. First, recall that
a (rough) section u of E is said to be k times dierentiable if, relative to any
smooth local frame s
1
, . . . , s
m
, u can be written as u =
u
i
s
i
for some k
times dierentiable functions u
i
. The space of k times dierentiable sections of
E is denoted
k
(E). When M is compact this is also a Banach space when
equipped with the norm
|u|
k := sup
xM
k
j=0
|
j
u(x)|
E
.
Secondly dene for each 1 p < and each k N
0
the Sobolev norm | |
p,k
on
c
(E) by
|u|
p,k
:=
_
k
j=0
|
j
u|
p
L
pd
g
_
1/p
=
_
k
j=0
_
M
|
j
u|
p
E
d
g
_
1/p
.
The notation can be a bit misleading, for as a matter of fact, the Sobolev spaces
depend not only on M and E but also on the metric on M as well as the metric
and connection on E, thus priori there is no canonical choice of norm!
Example 11.21. Consider the manifold 1
n
and the trivial real line bundle
E := 1. Equip E with the trivial connection d and the metric inherited
from 1. Lets calculate the Sobolev norm | |
2,1
, on u
c
(E) = C
c
()
|u|
2
2,1
= |u|
2
L
2 +|du|
2
L
2.
Since
du =
u
x
1
dx
1
+ +
u
x
n
dx
n
and since (dx
i
) is an orthonormal frame for T
we get
|du|
E
=
u
x
1
+ +
u
x
n
.
Hence
|u|
2,1
=
_
[u[ +
u
x
1
+ +
u
x
n
dx
and this is just equal to the usual Sobolev norm on H
1
(). Likewise, one can
show that | |
2,k
equals the usual Sobolev norms on H
k
().
Had we chosen a dierent connection on the trivial line bundle in the previous
example, we would probably end up with a dierent norm than the usual one.
In the case of a compact base manifold, however, the situation is much brighter.
Lemma 11.22. Let M be a compact oriented Riemannian manifold and E a
vector bundle over M. Then the Sobolev norm is (up to equivalence) independent
of the choices of metrics and connections.
Proof. First we show that the Sobolev-norm | |
p,k
is independent of the
choices of connections. Assume we have connections and

on T
M (one of
them need not be the Levi-Civita connection) and connections
E
and

E
on
E. Let | |
p,k
be the Sobolev norm constructed from and
E
and let
p,k
be the norm constructed from

and

E
. We want to show that a constant
C
p,k
exists such that for each (E)
p,k
C
p,k
||
p,k
.
We do it by induction. Note, rst, that for k = 0 we have
p,0
= ||
p,0
since
this norm does not involve the connections. Thus C
p,0
= 1.
Assume now the conclusion holds for k 1 and note that we have
||
p
p,k
= ||
p
p,0
+|
E
|
p
p,k1
and
p
p,k
=
p
p,0
+

E

p
p,k1
.
Let A
1
(M, End(E)) denote the dierence

E

E
. We want to estimate
the term
p
p,k
:
p
p,k
= ||
p
p,0
+
p
p,k1
|p|
p
p,0
+C
p
p,k1
|
E
|
p
p,k1
= ||
p
p,0
+C
p
p,k1
|
E
+A|
p
p,k1
||
p
p,0
+C
p
p,k1
2
p
(|
E
|
p
p,k1
+|A|
p
p,k1
)
(1 + 2
p
C
p
p,k1
)(||
p
p,0
+|
E
|
p
p,k1
) +C
p
p,k1
|A|
p
p,k1
= (1 + 2
p
C
p
p,k1
)||
p
p,k
+C
p
p,k1
|A|
p
p,k1
.
We may assume that A = for some 1-form . By the Leibniz rule (which
can easily be extended to hold for
k
) we get
k1
( ) =
k1
i=0
_
k 1
i
_
(
k1i
)
i
,
and by this we can estimate | |
p
p,k1
:
| |
p
p,k1
=
k1
j=0
_
M
_
_
_
j
j=0
_
j
i
_
i

j1
_
_
_
p
E
d
g
k1
j=0
j
i=0
_
j
i
__
M
|
i
|
p
T
M
|
j1
|
p
E
d
g
2
k
||
p
k1
k1
j=0
j
i=0
_
M
|
j1
|
p
E
d
g
2
k
k||
p
k1
||
p
p,k1
2
k
k||
k1||
p
p,k
.
In the third estimate we used that
_
j
i
_
2
k
. Combining this we the inequality
above we can nd the desired constant C
p,k
such that
p,k
is dominated by
| |
p,k
. Thus the Sobolev norm is independent of the choice of connections.
We need to check that the norms are independent of the choices of metrics.
There are two impacts of a change of metric on M, rst i changes the Levi-Civita
connection, but this change yields an equivalent metric, secondly it changes the
measure by a positive function C
(M), i.e.
_
M
f d
g
=
_
M
f d
g
. But
as M is compact, is bounded and so the new Sobolev norm is equivalent to
the original one. The impact of a change of metric on E is also just a positive
function.
For this reason, on compact manifolds the following denition is well-posed.
Denition 11.23 (Sobolev Space). Let 1 p < and k N
0
and assume E
is a Riemannian vector bundle over a compact, orientable Riemannian manifold
M. Then, dene the Sobolev space W
p,k
(M, E) as the completion of (E) in
the Sobolev norm | |
p,k
.
With the Sobolev norm these spaces are Banach spaces. The case p = 2 is
particularly interesting, for in this case H
k
(M, E) := W
2,k
(M, E) (there should
be no risk of confusing this with the cohomology groups) are Hilbert spaces
when equipped with the inner product
u, v
2
=
k
j=0
j
u,
j
v
L
2 =
k
j=0
_
M
j
u,
j
v
E
d
g
.
The Sobolev norm in this case will be written | |
k
instead of | |
2,k
.
Note that the Sobolev norm | |
p,0
is just the L
p
-norm on (E). Thus
the zeroth Sobolev space W
p,0
(M, E) equals L
p
(M, E) for all p. In particular
H
0
(M, E) = L
2
(M, E).
There is another approach to dening Sobolev spaces, one that relates to the
classical Sobolev spaces and norms over 1
n
, and which makes it possible for
us to dene Sobolev spaces not just for k N
0
but for s 1. Sobolev spaces of
non-integer order is important in the study of pseudo-dierential operators.
Continue to let M denote a compact oriented Riemannian manifold and E
a smooth rank N vector bundle over M. Pick a nite atlas (U
i
,
i
)
J
i=1
for M
for which each open set U
i
is a trivialization neighborhood of E and let (
i
)
be a partition of unity with compact support subordinate to this cover. For
any (E) we have that (
i
)
1
i
is a smooth map 1
n
E with
compact support in
i
(U
i
). Since E is trivial over the support of
i
, we may
view (
i
)
1
i
as a smooth compactly supported map 1
n
C
N
. Thus for
any s 1 it is an element of H
s
(1
n
) H
s
(1
n
), and we dene the Sobolev
norm by
||
t
s
:=
J
i=1
|(
i
)
1
i
|
H
s
(R
n
)
N
where of course the norm on the right-hand side is the product norm. We may
therefore for any s 1 dene the Sobolev space H
s
(M, E) as the closure of
(E) in the norm | |
t
s
.
For s = k N
0
this new Sobolev norm is equivalent to the one dened in
terms of connections. Note rst, that since
i
is a dieomorphism we have by
Example ?? that
|(
i
)
1
i
|
H
k
(R
n
)
N C|
i
|
2,k
C
t
|(
i
)
1
i
|
H
k
(R
n
)
N.
Hence we get
||
2,k
=
_
_
_
J
i=1
_
_
_
J
i=1
|
i
|
2,k
C
t
J
i=1
|(
i
)
1
i
|
H
k
(R
n
)
N
= C
t
||
t
k
.
The other way round:
||
t
k
=
J
i=1
|(
i
)
1
i
|
H
s
(R
n
)
N C
J
i=1
|
i
|
2,k
JC||
2,k
.
Thus | |
t
k
and | |
2,k
are equivalent. Note how this implies that | |
t
k
is (at least
up to equivalence) independent of the choices of atlas and partition of unity.
Given this link to the classical Sobolev spaces, it should come as no surprise
that many of the results there carry over verbatim to the manifold case. One of
the most important such results is of course the Sobolev Theorem:
Theorem 11.24 (Sobolev Embedding Theorem). Let M be a compact
orientable Riemannian manifold of dimension n. For non-negative integers k
and m such that
1
2
n k m we have H
k
(M, E)
m
(E), more precisely each
class u H
k
(M, E) has a representative in
m
(E). Moreover the inclusion is a
continuous operator.
The relations among the Sobolev spaces are given by the Rellich Lemma:
Theorem 11.25 (Rellich Lemma). If s t then H
t
(M, E) H
s
(M, E) and
if the inequality is strict, the inclusion is a compact operator.
When having a dierential operator A : (E) (F) of order k (let N
1
and N
2
denote the rank of E and F respectively) we can extend it to Sobolev
spaces: The way to do it, is to show that it is continuous w.r.t. the Sobolev
norms. Picking a cover of proper charts (U
i
,
i
) and partitions of unity (
i
) we
get for any u (E):
|Au|
sk
=
|(
i
Au)
1
i
|
H
sk
(R
n
)
N
2
.
But over U
i
, the operator A acts as a matrix of usual dierential operators on
1
n
and we know that a dierential operator

a
on 1
n
with coecient
functions in o(1
n
) is continuous H
s
(1
n
) H
sk
(1
n
) for any s. Thus for
each i we can nd a constant C
i
(depending of course on s) such that
|(
i
Au)
1
i
|
H
sk
(R
n
)
N
2
C
i
|(
i
u)
1
i
|
H
s
(R
n
)
N
1
.
Letting C denote the maximum of the C
i
s we see that
|Au|
sk
C|u|
s
.
Consequently:
Proposition 11.26. A dierential operator A : (E) (F) over a compact,
oriented Riemannian manifold extends to a continuous operator A : H
s
(M, E)
H
sk
(M, F) for each s 1.
One can easily show that the Hilbert space adjoint A
: H
sk
(M, F)
H
s
(M, E) is exactly the extension of the formal adjoint A
to H
sk
(M, E),
thus the use of the star is unambiguous.
When working with an order k dierential equation like Au = v where v is
some xed smooth section or an element of a Sobolev space, one is interested in
how nice u behaves when compared to v. Unfortunately, one cannot conclude
that if v
m
(F) then u
k+m
(E), but if A is elliptic, we have something
similar:
Theorem 11.27 (Elliptic Regularity). Let A be an elliptic dierential op-
erator of order k. If Au = v for some v H
s
(M, F), then u H
s+k
(M, E).
Thus for elliptic operators Sobolev spaces are the natural spaces to work with.
Combining the elliptic regularity with the Sobolev Embedding Theorem, one can
relate dierentiability of u to dierentiability of v but not as schematically as
one could have hoped. However, we do have the following strong consequence
Corollary 11.28 (Weyl Lemma). If A is an elliptic order k dierential op-
erator A and Au = v for v (F), then u (E). In particular the kernel of
A : H
s
(M, E) H
sk
(M, F) is a subspace of (E) and is thus independent
of s.
Change of focus. In the rest of this section we will be concerned with co-
called Fredholm operators and their theory. We just present some basic stu on
Fredholm operators, for proofs we refer the reader to [Conway] for an excellent
presentation of the theory.
Denition 11.29 (Fredholm Operator). A linear operator A : V
1
V
2
between vector spaces is called a Fredholm operator if ker A and coker A :=
V/ imA are nite-dimensional. The Fredholm index of a Fredholm operator is
then dened as
ind(A) := dimker Adim coker A.
The main result which we present later in this section (without proof, though)
is that elliptic dierential operators are Fredholm operators. However, before
that we want to describe some key features of Fredholm operators and their
indices. First, the so-called Atkinson Theorem which gives some equivalent con-
ditions for a bounded operator to be Fredholm
Theorem 11.30 (Atkinson). Let A : H
1
H
2
be a bounded operator be-
tween Hilbert spaces. Then the following conditions are equivalent
1) A is Fredholm.
2) imA is closed and ker A and ker A
are nite-dimensional.
3) There exists a bounded operator B : H
2
H
1
(called a parametrix) such
that I AB and I BA are compact operators.
The set of bounded Fredholm operators H
1
H
2
is denoted F(H
1
, H
2
).
This is an open set inside B(H
1
, H
2
), when the latter is given the norm-topology.
The Fredholm index turns out to be an impressively robust and well-behaved
quantity.
Proposition 11.31. Let A F(H
1
, H
2
). Then
1) The adjoint A
is again a Fredholm operator and ind A
= ind A.
2) If B F(H
1
, H
2
) is a Fredholm operator, then BA is a Fredholm operator
and ind(BA) = ind(A) + ind(B).
3) If K : H
1
H
2
is a compact operator, then A+K is a bounded Fredholm
operator and ind(A+K) = ind(A).
4) The Fredholm index is a continuous map ind : F(H
1
, H
2
) Z. In
particular it is constant on path components of F(H
1
, H
2
).
With the basic properties of Fredholm operators in place we can return to
the study of dierential operators:
Theorem 11.32. An elliptic dierential operator A : H
s
(M, E) H
sk
(M, F)
is Fredholm.
As noted in the Weyl Lemma, ker A (E) is independent of the Sobolev
space on which we choose to represent A. Likewise with ker A
. Thus
ind A = dimker Adimker A
is independent of the choice of Sobolev space, and therefore we can talk unam-
biguously about the index of an elliptic dierential operator.
Let A
t
: (E) (F), t [0, 1] be a family of dierential operators. We call
this a continuous family of dierential operators if for any proper neighborhood
11.5 Elliptic Complexes 227
U, the coecients of
in the local representative (A

t
)
U
depend continuously
on t. Two elliptic operators A
0
and A
1
for which there exists a continuous family
A
t
of elliptic operators joining them are said to be homotopic.
Let J
m
[0, 1] be the set of t [0, 1] for which A
t
is of order at least m.
By continuity of the family J
m
is an open set. Since [0, 1] is compact, there is a
maximum
k = maxm[ J
k
[0, 1]
and thus we can assume all the operators A
t
to be of order k. Given such a contin-
uous family of dierential operators and a real number s, one can represent each
operator as a continuous map A
t
: H
s
(M, E) H
sk
(M, F) and one can show
that t A
t
is continuous in the norm topology on B(H
s
(M, E), H
sk
(M, F)).
Combining with property 4 of Proposition 11.31 we get
Corollary 11.33. The index of two homotopic elliptic dierential operators are
equal.
Corollary 11.34. For two elliptic dierential operators A
0
and A
1
having iden-
tical principal symbols, ind A
0
= ind A
1
.
Proof. Let A
t
= (1t)A
0
+tA
1
, then this is a continuous family of dierential
operators, and obviously (A
t
) = (A
0
), so A
t
is elliptic. Thus A
0
and A
1
are
homotopic, and they have the same index by the preceding corollary.
11.5 Elliptic Complexes
In the next chapter we will uncover a connection between dierential operators
and K-theory, a connection which is vital for the index theory. As a preparation
for this we introduce in this section the notion of an elliptic complex. As an
extra bonus we obtain a generalization of the classical Hodge Theorem.
Lets rst discuss the following simple case. Consider the nite complex V
0

V
0
d0
V
1
d1

dn1
V
n
0
of nite-dimensional inner product spaces and linear maps. Dene
H
k
(V
) := ker d
k
ker d
k1
and
H
(V
) :=
n
k=0
H
k
(V
),
these are the the harmonic elements of the complex. Furthermore, let H
k
(V
)
denote the cohomology of the complex, i.e.
H
k
(V
) := ker d
k
/ imd
k1
.
We put V :=
V
k
, and dene the map d :=
d
i
: V V as well as the
map := (d +d
)
2
.
Lemma 11.35. The quotient map
q
k
: H
k
(V
) H
k
(V
)
is a linear isomorphism. Furthermore
H
(V
) = ker = ker(d +d
).
In particular the complex will be exact if and only if d + d
: V V is an
isomorphism.
Proof. First, lets show that q
k
is an isomorphism. To show injectivity suppose
v ker d
k
ker d
k1
satises q
k
(v) = 0, i.e. v = d
k1
u for some u V
k1
. Since
v ker d
k1
we have 0 = d
k1
v = d
k1
d
k1
u and therefore
0 = d
k1
d
k1
u, u = |d
k1
u|
2
,
i.e. v = d
k1
u = 0, thus q
k
is injective. Showing that it is surjective is a bit
harder: Let [v] H
k
(V
) be a cohomology class represented by v. We can

identify this equivalence class with the ane subspace
C
v
:= v +d
k1
u [ u V
k1
V
k
.
Let v
0
be the unique point in C
v
with minimal norm
6
. We claim that d
k1
v
0
=
0. To see this, let u V
k1
be arbitrary and consider the smooth function
f
u
: 1 [0, [ given by
f
u
(t) = |v
0
+td
k1
u|
2
.
Since v
0
+ td
k1
u C
v
and v
0
was the element in C
v
with minimal norm, f
u
must have a minimum in 0. Thus f
t
u
(0) = 0, i.e.
0 =
d
dt
t=0
v
0
+td
k1
u, v
0
+td
k1
u
= d
k1
u, v
0
+v
0
, d
k1
u = v
0
, d
k1
u +v
0
, d
k1
u
= 2 Red
k1
v
0
, u.
Replacing u by d
k1
v
0
we see that d
k1
v
0
= 0. Moreover
d
k
(v
0
) = d
k
(v +d
k1
u) = d
k
v = 0.
Thus v
0
represents a class in H
k
(V
) and since v
0
C
v
we see that [v] = [v
0
] =
q
k
(v
0
). Since d
k1
v
0
= 0, v
0
is a homogenous element and thus q
k
: H
k
(V
)
H
k
(V
) is surjective.
Next we show the identity ker(d + d
) = ker . The inclusion is obvious.

Conversely, assume that v ker , then since d +d
is self-adjoint, we get
0 = (d +d
)
2
v, v = (d +d
)v, (d +d
)v = |(d +d
)v|
2
,
i.e. v ker(d+d
). The identity ker(d+d
) = H
(V
) follows from the following

decomposition
ker(d +d
) = (ker d
0
) (ker d
1
ker d
0
) (ker d
n1
).
Denition 11.36 (Elliptic Complex). Let E
0
, . . . , E
n
be complex Rieman-
nian vector bundles over M, and let D
k
: (E
k
) (E
k+1
) be dierential
operators such that D
k
D
k1
= 0. The complex
0

(E
0
)
D0
(E
1
)
D1

Dn1
(E
n
)

0 (11.19)
is called an elliptic complex if the complex
0

E
0
(D0)
E
1
(D1)

(Dn1)
E
n
0
over the cotangent bundle : T
M M is exact o the zero section of T
M.
6
It is a standard result in Hilbert spaces theory that if A is a convex subset of a Hilbert
space V , and q is a point in V , there exists a unique point p A such that pq = dist(p, A).
This applies to the situation here, the Hilbert space being of course V
k
and the convex set
being Cv.
Dene the bundles
E :=
n
k=0
E
k
, E
+
:=
n
k=0
E
2k
, E
:=
n
k=0
E
2k+1
,
let D denote the dierential operator

D
k
: (E) (E) and put D :=
D + D
as well as := D
2
. Furthermore we let D : (E
+
) (E
) denote
the restriction of D to (E
+
).
Proposition 11.37. The complex (11.19) is elliptic if and only if the operator
D : (E) (E) is an elliptic dierential operator.
Proof. Let 0 ,=
p
T
p
M. It is obvious from the properties of the symbol (cf.
Proposition 11.7) that
(D)(
p
) =
_
n
k=0
(D
k
)(
p
)
_
+
_
n
k=0
(D
k
)(
p
)
_
.
From Lemma 11.35 we see that the complex of vector spaces
0

E
0
p
(D0)(p)
E
1
p
(D1)(p)

(Dn1)(p)
E
n
p

0
is exact if and only if (D)(
p
) is an isomorphism, i.e. the complex (11.19) is
elliptic if and only if the operator D is elliptic.
We are now ready to state and prove the promised generalized Hodge Theo-
rem. As in the setup for Lemma 11.35 we let
H
k
(E
) := ker D
k
ker D
k1
and H
(E
) :=
n
k=0
H
k
(E
)
denote the harmonic sections of E
k
, resp. E. It is easy to see that
ker = ker D = H
(E
).
If the complex is elliptic, then so is the operator D, in particular it is Fredholm,
implying that ker D (and thus also H
(E
)) is nite-dimensional.
Furthermore, we let H
k
(E
) := ker D
k
/ imD
k1
denote the kth cohomology
group of the complex. If the vector spaces H
k
(E
) are nite-dimensional, the

integer dimH
k
(E
) (the dimension taken over 1 if the bundles E

k
are real,
and over C if the bundles are complex) is called the kth Betti number of the
complex and we dene the Euler characteristic (E
) of the complex to be the

alternating sum of the Betti numbers
(E
) :=
n
k=0
(1)
k
dimH
k
(E
).
One of the statements of the following theorem is that the cohomology groups
of an elliptic complex are always nite-dimensional.
Theorem 11.38 (Generalized Hodge). For an elliptic complex
0

(E
0
)
D0
(E
1
)
D1

Dn1
(E
n
)

0 .
the natural quotient map q
k
: H
k
(E
) H
k
(E
) is a linear isomorphism,
i.e. any cohomology class is represented by exactly one harmonic section. Con-
sequently, the cohomology groups of the complex are nite-dimensional.
Proof. Almost every argument from the proof of Lemma 11.35 holds in this
general case, except the one proving surjectivity of q
k
. Since the spaces (E
k
)
are no longer nite-dimensional or complete, we have to proceed dierently. The
problem is still the same: Let [v] H
k
(E
) where v ker D
k
. We need to nd
a representative for this class in H
k
(E
). Since D is an elliptic operator, the

kernel ker D = H
(E
) is nite-dimensional, in particular H
k
(E
) is nite-
dimensional. Thus, there exists a continuous projection p : (E
k
) H
k
(E
).
Letting v
0
= p(v), we assert that v
0
is the desired representative. Obviously,
v v
0
is orthogonal to ker D. Extend D to a continuous map H
1
(M, E
k1
)
L
2
(M, E
k
), then we know that L
2
(M, E
k1
) = imD ker D
= imD ker D
(since it is self-adjoint and Fredholm, so it has closed image) and so we must
have v v
0
= Du for some u H
1
(M, E
k1
). From the Weyl Lemma we obtain
smoothness of u, i.e. u (E
k1
).
Recall that v v
0
ker D
k
ker D and so by acting with D on the equation
v v
0
= (D +D
)u we get DD
u = 0. Thus
0 = DD
u, u = |D
u|
2
,
i.e. D
u = 0. Therefore v v
0
= Du = Du = D
k1
u, i.e. v and v
0
are cohomol-
ogous, and surjectivity is proven.
From the Hodge Theorem we derive our rst index result
Corollary 11.39. Let E
be an elliptic complex. Then the Euler characteristic

(E
) equals the index of the Fredholm operator

D : (E
+
) (E
).
Proof. First of all, the operator D is Fredholm since it is the restriction of the
elliptic operator D. Furthermore, it is easy to see that
ind D = dimker(D) dimker(D

)
=
n
k=0
dimH
2k
(E
)
n
k=0
dimH
2k+1
(E
)
=
n
k=0
dimH
2k
(E
)
n
k=0
dimH
2k+1
(E
)
= (E
).
Theorem 11.38 is indeed a generalization of the classical Hodge Theorem as
seen in the following example.
Example 11.40. The de Rham Complex. Let M be a compact, oriented Rie-
mannian manifold and consider the de Rham complex
0

C
(M)
d0

1
(M)
d1

dn1

n
(M)

0
where d
i
are the exterior derivatives.
Let
p
T
p
M be nonzero, then the corresponding symbol complex takes the
form
0

C
ip

1
(T
p
M
C
)
ip

ip

n
(T
p
M
C
)

0 ,
and it is well-known that this is exact, i.e. the de Rham complex is elliptic.
The cohomology of this complex is the de Rham cohomology, and from the
Hodge Theorem we derive that the de Rham cohomology groups are nite-
dimensional and that there is an isomorphism from the space of harmonic k-
forms to H
k
dR
(M). This is indeed the content of the classical Hodge Theorem.
The operator D is just the Hodge-de Rham operator, and Proposition 11.37
states that it is elliptic. Finally, the statement of Corollary 11.39
ind(D) = (M) (11.20)
is that the Fredholm index of the restricted Hodge-de Rham operator
D :
+
(M)
(M)
equals the Euler characteristic of the manifold. This is of course just a rephrase
of the de Rham Theorem.
Chapter 12
The Atiyah-Singer Index
Theorem
12.1 K-Theoretic Version
Let A be an elliptic dierential operator between vector bundles E and F over
a compact Riemannian manifold M. The assumption of compactness is impor-
tant and will be retained throughout the rest of this chapter unless otherwise
specied. The symbol map (A) is part of the complex
0

E
(A)
F

0
over T
M and since A was elliptic, this complex is exact o the zero section
of T
M. The zero section is dieomorphic to M, in particular it is compact,

and thus this complex represents a class in K(T
M). Composing with the iso-

morphism K(T
M)

K(TM) induced by the metric bundle isomorphism
TM

T
M, we obtain a class [(A)] K(TM) called the symbol class of

A. The statement of the Index Theorem (in its K-theoretic form) is that there
exists a homomorphism K(TM) Z, called the topological index which eval-
uated on the class [(A)] equals the Fredholm index of A. Our present job is to
dene this topological index.
To this end, assume i : Y M to be an inclusion of a proper submanifold
Y into M. Let N
Y
denote the normal bundle of Y in M and U
Y
the corre-
sponding tubular neighborhood
1
. To the inclusion Y M there corresponds
an inclusion TY TM, a tubular neighborhood U
TY
TM and a normal
bundle N
TY
(which is now a vector bundle over TY ) which happens to be (iso-
morphic to)
(N
Y
N
Y
)

=
N
Y

N
Y
, where : TY Y now denotes
the projection in the tangent bundle. N
TY
is a real bundle, but we can equip it
with a complex structure, namely the bundle map
_
0 id
id 0
_
.
Thus, we may view N
TY
as a complex bundle over TY . We can now dene a
group homomorphism i
!
: K(TY ) K(TM) as the composition
K(TY )

K(N
TY
)

K(U
TY
) K(TM) (12.1)
1
Given an inclusion Y M of a proper submanifold into a Riemannian manifold (compact
or not) one can dene the normal bundle N
Y
over M to be the smooth vector bundle over Y
whose ber over x Y is Nx := (TxY )
TxM. A tubular neighborhood is an open subset

U
Y
of M containing Y together with a dieomorphism N
Y
U
Y
which is an extension of
the dieomorphism N
0
Y of the zero section of N with Y . One can show that a tubular
neighborhood always exists, see for instance [Lang] Thm. IV.5.1.
233
234 Chapter 12 The Atiyah-Singer Index Theorem
where the rst map is the Thom isomorphism for the complex bundle N
TY

TY , the second is induced by the dieomorphism N
TY
U
TY
and the last
one is induced by the inclusion U
TY
TM.
Lemma 12.1. The map i
!
is independent of the choice of tubular neighborhood,
and for a smooth family i
t
: Y M of inclusions (meaning that the map
I Y M, (t, x) i
t
(x) is smooth) then (i
0
)
!
= (i
1
)
!
.
Finally, if i : X Y and j : Y M are inclusions, then (j i)
!
= j
!
i
!
.
Proof. Given two tubular neighborhoods U
TY
and

U
TY
, one can show that
they are homotopic, i.e. there exists a smooth map : I N
TY
TM
such that
0
maps N
TY
dieomorphically onto U
TY
and
1
maps N
TY
dif-
feomorphically onto

U
TY
. But then, by homotopy invariance, the two maps
K(N
TY
) K(TM) induced by
0
and
1
are identical.
The smooth family i
t
gives rise to a smooth family of inclusions
i
t
: TY
TM. Let TY
t
denote the image
i
t
(TY ) in TM. This new family again gives rise
to a smooth family of bundles isomorphisms N
TYt
U
TYt
between the corre-
sponding normal bundles and tubular neighborhoods and thus also to a smooth
family of inclusions U
TYt
TM. These smooth families also produces a dif-
feomorphism TY
0
TY
1
, simply take a point

i
0
(x) and move it to

i
1
(x).
Likewise, from the family of bundle isomorphisms, we obtain a bundle isomor-
phism N
TY0
N
TY1
which is compatible with the underlying dieomorphism
TY
0
TY
1
. Upon composing with the dieomorphisms N
TY
k
U
TY
k
,
k = 1, 2 we get a dieomorphism U
TY0
U
TY1
. By construction this map
makes the following diagram commute
N
TY0
N
TY1
U
TY0

U
TY1
Consider now the following commuting diagram
K(TY )

K
K
K
K
K
K
K
K
K
K(TY
0
)

K(N
TY0
)

K(U
TY0
)
L
L
L
L
L
L
L
L
L
L
K(TY
1
)

K(N
TY1
)

K(U
TY1
)

K(TM)
The vertical maps are induced by the dieomorphisms TY
1
TY
0
, N
TY1
N
TY0
and U
TY1
U
TY0
. The rst triangle commutes since the dieomor-
phisms between TY , TY
0
and TY
1
commute. The rst square commutes thanks
to naturality of the Thom isomorphism and the second square commutes since
the preceding diagram does. The last triangle commutes due to homotopy in-
variance. Composing the upper maps gives (i
0
)
!
while composing the lower
maps yields (i
1
)
!
. From this commutativity of the diagram, we see that (i
0
)
!
=
(i
1
)
!
.
As an example: Let i : M E be the zero-section of a vector bundle (real
or complex). But then the push-forward map i
: TM TE is just the zero-

section of the complex bundle TE TM. Then TE will be a normal bundle
for this inclusion as well as a tubular neighborhood. Thus, the two last maps
in the denition of i
!
are just identity maps, i.e. i
!
is nothing but the Thom
isomorphism for the bundle TE TM.
By the famous Nash Embedding Theorem we can embed any Riemannian
manifold M isometrically into 1
n
. Denoting the inclusion map by i we have by
12.1 K-Theoretic Version 235
the construction above a map i
!
: K(TM) K(T1
n
). Similarly, if j denotes
the inclusion of the 1-point set (zero dimensional manifold) P into 0 1
n
we
get a homomorphism
j
!
: Z
= K(TP) K(T1
n
).
But j is just the zero-section of the trivial bundle C
n
over a point, and therefore,
as noted above, j
!
: K(TP) K(T1
n
) is just the Thom isomorphism. In
particular it has an inverse (j
!
)
1
: K(T1
n
) K(TP)
= Z.
Denition 12.2 (Topological Index). Let M be a compact Riemannian man-
ifold and i : M 1
n
an isometric embedding. Dene the topological index
ind
T
: K(TM) Z as the homomorphism (j
!
)
1
i
!
.
Lemma 12.3. The topological index is independent of the choice of embedding.
Proof. Assume we have two embeddings i : M 1
n
and i
t
: M 1
m
and consider the diagonal embedding k(x) = (i(x), i
t
(x)) of M into 1
n+m
as
well as the smooth family of embeddings k
t
(x) = (i(x), ti
t
(x)). Lets compare i
and k
0
(these are not identical as i maps into 1
n
whereas k
0
maps into 1
n+m
).
Let N T1
n
denote the normal bundle for i(M), then obviously the normal
bundle for k
0
(M) will be N 1
n
. Viewing T1
n
and T1
n+m
as C
n
and C
n+m
respectively, and viewing C
n+m
as a (trivial) complex rank m vector bundle over
C
n
we get a Thom isomorphism
: K(T1
n
) K(T1
n+m
). Let P denote
the one-point space, and j : TP T1
n
and l : TP T1
n+m
denote the
inclusions, then, by transitivity of the Thom isomorphism each of the triangles
in the following diagram commutes
K(TM)
i
!
.s
s
s
s
s
s
s
s
s
s
(k0)
!
M
M
M
M
M
M
M
M
M
M
K(T1
n
)

K(T1
n+m
)
K(TP)
j
!
_K
K
K
K
K
K
K
K
K
K
l
!
q
q
q
q
q
q
q
q
q
q
Thus we see that (l
!
)
1
(k
0
)
!
= (j
!
)
1
i
!
. This shows that i and k
0
give the same
index maps. By Lemma 12.1 k
0
and k
1
= k give the same index map. Thus i
and k produce the same index map. By symmetry, the same is true for i
t
and
k, and hence i and i
t
produce the same index.
Without further Ado here is the theorem we have all been waiting for
Theorem 12.4 (Atiyah-Singer Index Theorem I). For an elliptic dier-
ential operator A : (E) (F) on a compact Riemannian manifold between
complex vector bundles, the Fredholm index of A equals the topological index of
its symbol class:
ind(A) = ind
T
([(A)]). (12.2)
Sketch of the proof. We give here a very rough sketch of the proof as it
is presented in the classical article by Atiyah and Singer(although they have
generalized slightly to equivariant K-theory). First of all, in Section 4 they
introduce the notion of an index function as well as two axioms, namely (1): if
M is a point, then the index is just the isomorphism K(TM) = K(pt)

Z
and (2) given an inclusion i : N M of N into M, then the index maps
ind
M
and ind
M
are compatible with the induced map i
: K(TN) K(TM),
i.e. ind
N
= ind
M
i
. The topological index is an index function and satises

these two axioms. Then they show that any index function satisfying these
axioms coincide with the topological index. Thus theses axioms characterize the
topological index uniquely. This is rather straightforward. The remainder of the
section is spend on breaking the second axiom down to a set of axioms which
are easier to handle.
In Section 6, after having reviewed some theory of pseudo-dierential opera-
tors they proceed to dene the so-called analytic index map. This is dened in
the following way: In section 2 they have showed that any element of K(TM)
can be obtained as the symbol class of some elliptic pseudo-dierential opera-
tor P over M. The index of P only depends on this symbol class, and hence
we obtain the analytic index ind
A
: K(TM) Z by sending an element of
K(TM), which is of the form [(P)], to ind(P). This is an index function and
it is relatively easy to realize that the rst axiom is satised. The real core of
the proof is to show that it satises the second axiom, or better, that it satises
the other axioms which implied the second one. This is carried out in Section
8 and 9. But then the analytic index and the topological index coincide! Since
ind
A
([(A)]) = ind A, by construction, (12.2) is true.
12.2 Cohomological Version
Our present task is to give a dierent (and some would say more applicable)
formula for the topological index, one involving cohomology classes instead of
K-classes. In this game the Chern character will play a prominent role.
First, however, we need to discuss a cohomological version of the Thom iso-
morphism. Let M be an oriented manifold of dimension m and N an oriented
manifold of dimension n. Then we have the Poincar duality isomorphisms
D
M
: H
p
c
(M; 1)

H
mp
(M; 1) and D
N
: H
p
c
(N; 1)

H
np
(N; 1)
where H
c
denotes cohomology with compact support. If f : M N is a
smooth map we obtain a homomorphism f
: H
p
c
(M; 1) H
p(mn)
c
(N; 1),
called the Gysin homomorphism or integration along the ber as the composition
H
p
c
(M; 1)
D
M
H
mp
(M; 1)
f

H
mp
(N; 1)
D
1
N
H
p(mn)
c
(N; 1)
where f
is the induced map in homology. In particular we can consider an

oriented vector bundle : E M of rank k (if the bundle is oriented, then
E as a manifold is also oriented). We also assume M to be compact, such that
H
p
c
(M; 1) = H
p
(M; 1). If i : M E denotes the inclusion as the zero-section,
then and i are homotopy inverses of each other, i.e.
and i
isomorphisms
in homology, thus
: H
p+k
c
(E; 1) H
p
c
(M; 1) and i
: H
p
c
(M; 1)
H
p+k
c
(E; 1) are isomorphisms. In fact they are inverses of each other, as is
easily seen.
Denition 12.5 (Thom Isomorphism). Let E be an oriented rank k-bundle
over an oriented manifold M. The map := i
: H
p
(M; 1)

H
p+k
c
(E; 1) is
called the Thom isomorphism in cohomology.
Note, that the bundle E need not be complex, contrary to the situation for
the K-theoretic Thom isomorphism.
Some basic and useful properties of the Thom isomorphism are listed in the
following Proposition
Proposition 12.6. Let : E M be an oriented vector bundle of rank k
over an oriented manifold M.
12.2 Cohomological Version 237
1) The class (1), called the Thom class, is the unique class in H
p+k
c
(E; 1)
with integral 1 over each ber such that
() =
() (1) (12.3)
and the Thom class is related to the Euler class of E by
e(E) = i
((1)). (12.4)
2) If H
p
c
(E; 1) is given the structure of a right H
p
(M; 1)-module by =

() ( H
p
c
(E; 1) and H
p
(M; 1)) then
1
=
is a module-
homomorphism, i.e.
1
(
()) =
1
().
3) If f : N M is a smooth map, then the natural map F : f
E E
is proper and if : H
p
(N; 1) H
p+n
(f
E; 1) denotes the the Thom

isomorphism for the pullback bundle, the following diagram commutes
H
p
(M; 1)
f
H
p
(N; 1)
H
p+n
c
(E; 1)
F
H
p+n
c
(f
E; 1)
The rst claim is proved in [Madsen, Tornehave] and the second one in [Bott,
Tu]. Applying i
to the formula (12.3) and using (12.4) we obtain

i
() = e(E) = e(E) . (12.5)

The last identity holds, since either e(E) H
2k
(M; 1) or e(E) = 0.
For the remainder of this section E M will denote a complex vector bun-
dle of rank k over a compact oriented Riemannian manifold (and thus considered
as a real vector bundle, it has rank 2k and is orientable). We have two Thom iso-
morphisms
: K(M) K(E) and : H

p
(M; 1) H
p+2k
c
(E; 1) and we
have Chern characters ch : K(M) H
(M; 1) and ch : K(E) H
c
(E; 1),
how are they related? Unfortunately, the Chern character does not intertwine
the two Thom isomorphisms, but almost: Introduce the Thom defect
I(E) :=
1
ch(
(1)).
One of the properties of the Thom defect below states that, up to the Thom
defect, the Chern character intertwines the Thom isomorphisms:
Lemma 12.7. Under the above conditions the following hold:
1) For all K(M) the Thom Defect Formula holds
1
ch(
) = I(E) ch . (12.6)
2) The Thom defect is natural, i.e. if f : N M is a smooth map then
I(f
E) = f
I(E).
3) The Thom defect is multiplicative, i.e. I(E F) = I(E)I(F).
4) If E is a trivial bundle, then I(E) = 1.
Proof. 1) First note, that by construction of the K-theoretic Thom isomor-
phism we have
(1)
(since M is compact, K(M) is unital ring).

Using that the Chern character is natural and multiplicative and that
1
is a
module homomorphism we get
1
ch(
) =
1
ch(
(1)
)
=
1
_
ch
(1) ch(
)
_
=
1
_
ch
(1)
(ch )
_
=
_
1
ch
(1)
_
ch = I(E) ch .
2) This follows from part 3) of Proposition 12.6 and the corresponding result
for the K-theory Thom isomorphism, Proposition 9.46. Concretely, let
:
K(M) K(E) be the Thom isomorphism for E and
: K(N) K(f
E)
the Thom isomorphism for f
E. If 1
M
K(M) and 1
N
K(N) denote the
identities, note that 1
N
= f
(1
M
) and so we have
(1
N
) =
(f
1
M
) = F
(1
M
).
Therefore
I(f
E) =
1
ch(F
(1
M
)) =
1
F
ch(
(1
M
))
= f
1
E
(ch
(1
M
)) = f
I(E).
3) and 4) are immediate consequences of (12.11) proved below.
Next, recall that if E and F are two vector bundles, then
k
(E F)
i+j=k
(
i
E) (
j
F). (12.7)
Dene a map
t
: Vect
C
(M) K(M)[ t ] by
t
(E) :=
k=0
[
k
E]t
k
.
Since
k
E = 0 if k is larger than the rank of E,
t
(E) is a polynomial in t with
coecients in K(M). If t = n is an integer, then
n
is an element of K(M).
Note that
t
(E F) =
k0
[
k
(E F)]t
k
=
k0
i+j=k
[
i
E][
j
F]t
i
t
j
=
i0
j0
[
i
E][
j
F]t
i
t
j
=
t
(E)
t
(F).
Example 12.8. Let L be a complex line bundle over M. Then
0
L = M C,
1
L = L and
k
L = 0 if k > 1. Since the trivial line bundle M C represents
the identity in K(M) we have
t
(L) = 1 + [L]t. As is multiplicative, we see
that
t
(L
1
L
n
) =
n
k=1
(1 + [L
k
]t). (12.8)
Thus we have determined
t
for any sum of line bundles.
In the following calculations our goal is to obtain a formula for the Thom
defect in terms of
1
(E) and the Euler class (since E is complex, the Euler
class is always dened). At rst, suppose that E = L
1
L
k
is a sum of
line bundles. Apply the Chern character to (12.8):
ch(
t
(E)) = ch
_
k
j=1
(1 + [L
j
]t)
_
=
k
j=1
ch(1 + [L
j
]t)
=
k
j=1
(1 +t ch L
j
) =
k
j=1
(1 +te
xj
)
where, as usual, x
j
= c
1
(L
j
). For t = 1 we get
ch(
1
(E)) =
k
j=1
(1 e
xj
). (12.9)
The importance of the map
1
is to produce a formula for the K-theoretic
Thom isomorphism which resembles (12.5):
Proposition 12.9. For a complex vector bundle E M over a compact
oriented manifold with i : M E denoting the zero section, the following
holds for any K(M):
i
() =
1
(E) . (12.10)
This is proved in [Lawson, Michelson]. Combining this with (12.5) we get
e(E)I(E) = i
(I(E)) = i
ch
(1) = ch(i
(1))
= ch(
1
(E)).
Now split the Euler class e(E) = c
k
(E) = x
1
x
k
and recall formula (12.9) to
obtain
x
1
x
k
I(E) =
k
j=1
(1 e
xj
).
Since x
1
x
(1 e
x
) is a perfectly holomorphic function (the singularity in
x = 0 is removable) we see that
I(E) =
k
j=1
1 e
xi
x
i
= (1)
k
Td
1
C
(E). (12.11)
But since both I and Td
1
C
are natural w.r.t. smooth maps, the Splitting Prin-
ciple guarantees that this formula holds for arbitrary vector bundles.
A nal notion, we need to introduce is that of the fundamental class or ori-
entation class. In the case of an oriented manifold M, one can show that there
exists a homology class [M] in the top degree homology group, which in a cer-
tain sense determines the orientation (a change of orientation will give produce
a dierent orientation class). Recall, that H
k
(M; 1)

= H
k
(M; 1)
(an identity
which holds for any eld, and not just 1), thus if M is compact (without bound-
ary) we can let a cohomology class in the top degree cohomology act on [M].
As a matter of fact, this is nothing but integration:
([M]) =
_
M
which is well-dened (i.e. independent of choice of representative for the coho-

mology class ) due to Stokes theorem, and the fact, that M has no boundary.
In the case of M being non-compact we can let elements of the top cohomol-
ogy with compact support act on the fundamental class, and the above formula
will still hold. In fact, we can let a compact cohomology class of arbitrary de-
gree act on the fundamental class, simply dene this action to be 0, unless the
cohomology class is of top degree.
Consider a tangent bundle : TM M and assume M is orientable. TM
is always orientable (whether M is orientable or not) and in local coordinates
(x
1
, . . . , x
n
, v
1
, . . . , v
n
) the orientation is given by dx
1
dv
1
dx
n
dv
n
. Thus
the fundamental class [TM] always exists. Given an element H
n
c
(TM; 1)
and assume M to have a global chart, so that appears as dv
1
dx
1
dv
n
dx
n
(we allow ourselves to identify the cohomology class with its representative)
then we see
[TM] =
_
TM
dv
1
dx
1
dv
n
dx
n
= (1)
n(n1)/2
_
TM
dx
1
dx
n
dv
1
dv
n
.
Integrating rst along the v
k
s is just integration along the ber, i.e. it equals
1
() where is the Thom isomorphism for the tangent bundle. Integrating
over the x
k
s is then just evaluation on the fundamental class [M]. Using a
partition of unity argument in the case where M is not covered by a single
chart, we get
[TM] = (1)
n(n1)/2
(
1
)[M]. (12.12)
Now we are in position to deduce the cohomological index formula:
Theorem 12.10 (Atiyah-Singer Index Theorem II). For an elliptic dier-
ential operator A : (E) (F) on a compact, oriented Riemannian manifold
of dimension n between complex vector bundles we have
ind(A) = (1)
n
_
ch[(A)]
A(M)
2
_
[TM] (12.13)
where : TM M is the projection in the tangent bundle.
Note, that : TM M is not a proper map. Thus
A(M)
2
is an element
of H
(TM; 1), and not necessarily of H
c
(TM; 1). However, ch[(A)] does have
compact support, and hence the product has compact support.
Proof. We begin by collecting some preliminary results which will become
useful later in the proof. Let T1
N
P be a scrunch map (P is just a one-
point set). Under the isomorphism T1
N
= C
N
we may view this as a complex
vector bundle over a point. Thus we have the Thom isomorphisms (
)
1
:
K(C
N
) K(P) = Z and : H
(P; 1) H
c
(C
N
; 1). By the Thom defect
formula with = (
)
1
u K(P) we get
1
ch u = I(C
N
) ch(
1
u) = ch(
1
u)
(as I is 1 on trivial bundles). Note that ch : K(P) H
(P; 1) is just the

identity Z Z, for the following reason: View C as the trivial line bundle
over the point, then c
k
(C) = 0 if for k 1 and therefore ch(C) = 1. Since C
represents the identity in K(P), the Chern character maps 1 to 1, but then it
has to be the identity. Thus we have
1
ch u =
1
u.
1
is integration along the ber, but here the ber is all of C
N
= T1
N
, thus
the left-hand side equals (ch u)[T1
N
] where [T1
N
] is the fundamental class of
T1
N
, (recall that TM is always an orientable manifold for any manifold M, in
fact, if (x
1
, . . . , x
n
, v
1
, . . . , v
n
) are local coordinates for TM, then an orientation
form is given locally by dx
1
dv
1
dx
n
dv
n
, thus the fundamental class
exists). Thus
u = (ch u)[T1
N
] (12.14)
where
= q
. This is our rst preliminary result.

Consider now a real vector bundle p : E M. Then also the push-forward
p
: TE TM is a vector bundle, in fact it can be shown to be isomorphic to
E =
E
R
C, where : TM M is the projection in the tangent
bundle. Thus it is a complex vector bundle over TM. The Thom isomorphism
in K-theory is denoted
whereas the Thom isomorphism in cohomology is

called . Since the inclusion i : TM TE of TM as the zero section is a
proper map, the Thom defect formula still holds and in this case yields
1
ch(
) = I(
E C) ch
for any K(TM). Evaluate this on the fundamental class [TM]:
1
ch(
)[TM] = (I(
E C) ch )[TM].
Since
1
is just integration along the bers we get
ch(
)[TE] = (I(
E C))[TM]. (12.15)
This is our second preliminary result.
Recall now, how we dened the topological index in the preceding section: We
took an embedding i : M 1
N
as well as an embedding j : P 1
N
where
P is a one-point space. Then we formed the maps i
!
: K(TM) K(T1
N
)
and j
!
: K(TP) K(T1
N
), and we noted that j
!
was nothing but the Thom
isomorphism
mentioned above. Let E M denote the normal bundle

to i(M) and let, as above,
denote the Thom isomorphism for the bundle

TE TM. For a given K(TM) the class
K(TE). This has compact

support. The inclusion E 1
N
extends to an inclusion TE T1
N
and
under the map in K-theory induced by this inclusion,
is mapped to a class
in K(T1
N
) with support inside TE. Since i
!
is exactly this extended class, we
obtain
(ch
)[TE] = ch(i
!
)[T1
N
]. (12.16)
But then we can combine the formulas obtained so far (replacing or u by
[(A)]) to get
ind A = (j
!
)
1
i
!
[(A)] =
1
i
!
[(A)]
= ch(i
!
[(A)])[T1
N
] = (ch
[(A)])[TE]
=
_
I(
E C) ch[(A)]
_
[TM]
the third identity being (12.14), the fourth being (12.16) and the nal being
(12.15).
Finally, we only need calculate I(
EC). First, note that it equals
I(E
C) since I is natural. Secondly we note that TM E = T1
N
, and thus, since
I(T1
N
) = 1 (as T1
N
is a trivial bundle) and since I is multiplicative, we
obtain
I(E C) = I(TM C)
1
= (1)
n
Td
C
(TM C)
the last identity being a consequence of the fact that TMC is its own conjugate
bundle. Applying Proposition 10.58 we get the desired formula.
Using (12.12) and the fact that
1
is a module homomorphism we obtain
the following alternative index-formula
Theorem 12.11 (Atiyah-Singer Index Theorem III). For an elliptic dif-
ferential operator A : (E) (F) between complex vector bundles on a
compact oriented Riemannian manifold of dimension n we have
ind(A) = (1)
n(n+1)/2
_
1
(ch[(A)])

A(M)
2
_
[M]. (12.17)
This powerful theorem has an impressively long list of corollaries. Here we
just mention two:
Corollary 12.12. An elliptic dierential operator on an odd-dimensional com-
pact manifold has index 0.
Proof. Consider the involution c : TM TM given by v v. This is
an orientation reversing bundle automorphism, for a basis E
1
, . . . , E
n
for one
of the bers T
p
M, when subjected to c is changed to E
1
, . . . , E
n
. The
transition matrix between these bases is just the diagonal diag(1, . . . , 1) and
this has determinant (1)
n
= 1. Thus c
[TM] = [TM].
How does c inuence the symbol? Well, rst of all, c induces a natural auto-
morphism c : T
M T
M and this is again given by . Let A be of

order m. Since the symbol (A) is homogenous of degree m in , we see that
c
(A) = (A) c = (1)

m
(A)
so (A) is either unchanged or changed to (A). But (A) and (A) can be
deformed continuously to each other through a path of elliptic symbols by t
e
it
(A). By construction of the K-classes of complexes, such a homotopy does
not change the corresponding K-theory class, thus c
[(A)] = [(A)]. Therefore

we get
ind(A) =
_
ch[(A)]
A(M)
2
_
[TM]
= c
_
ch[(A)]
A(M)
2
_
c
[TM]
=
_
ch(c
[(A)]) c
A(M)
2
_
([TM])
=
_
ch[(A)]
A(M)
2
_
[TM]
= ind(A)
(the second identity follows from involutivity of c and the fourth from the fact
that c = ). Thus ind(A) = 0.
The second corollary, we want to derive is the generalized Gauss-Bonnet The-
orem.
Corollary 12.13 (Gauss-Bonnet). Let M be a compact, oriented manifold of
even dimension n and let e(M) be the Euler class of the tangent bundle, then
(M) =
_
M
e(M). (12.18)
In the case of an odd-dimensional manifold (M) = 0.
Proof. Let i : M TM be the zero-section. By (12.5) we have i
() =
e(M). Replace by
1
to obtain
i
=
1
()e(M), for all H
c
(TM; 1).
Thus we also have
1
(ch u)e(M) = i
ch u = ch(i
u) (12.19)
for all u K(TM).
Now consider the complex Hodge-de Rham operator D = d+d
:
+
C
(M)
C
(M), i.e. the operator originating from the complexied de Rham complex
(in order for the Index Theorem to be applicable we need the bundles to be
complex). We have discussed this operator in the real case in Example 11.40
where we found that the index equals (M). This is still true in the complex
case, for then the Euler characteristic is simply the alternating sum of
dim
C
H
k
(M; C) = dim
C
(H
k
(M; 1)
R
C) = dim
R
H
k
(M; 1).
For now we work on the right-hand side of 12.17. First we note that the index
class of D equals
[(D)] =
n
i=0
(1)
i
[
(
i
TM
C
)]
where is the projection in the tangent bundle (in secret we have identied the
tangent and the cotangent bundle) and [
i
TM
C
] is the K-class represented by
the ith exterior bundle. Since i = id
M
we get
i
[(D)] =
n
i=0
(1)
i
i
(
i
TM
C
)] =
n
i=0
(1)
i
[
i
TM
C
].
The Chern character of this K-class has already in Example 10.61 been calcu-
lated to be c
n
(TM
C
)Td
1
C
(TM
C
). Plugging this K-class into (12.19) therefore
gives
1
(ch[(D)])e(M) = c
n
(TM
C
)Td
1
C
(TM
C
).
Noting that
c
n
(TM
C
) = (1)
n/2
p
n/2
(M) = (1)
n/2
e(M)
2
(cf. Proposition 10.41 and 10.49) we get
1
(ch[(D)])e(M) = (1)
n/2
e(M)
2
Td
1
C
(TM
C
).
From this we conclude, that
1
(ch[(D)]) = (1)
n/2
e(M)Td
1
C
(TM
C
).
Inserting this into (12.17) and recalling the identity

A
R
(TM))
2
= Td
C
(TM
C
)
(cf. Proposition 10.58) we get the desired expression e(M)[M].
Appendix A
Table of Cliord Algebras
The following table displays the rst of the real Cliord algebras Cl
p,q
.
q p 0 1 2 3 4
0 1 1 1 1(2) C(2) H(2)
1 C 1(2) 1(2) 1(2) 1(4) C(4)
2 H C(2) 1(4) 1(4) 1(4) 1(8)
3 HH H(2) C(4) 1(8) 1(8) 1(8)
4 H(2) H(2) H(2) H(4) C(8) 1(16)
5 C(4) H(4) H(4) H(4) H(8) C(16)
6 1(8) C(8) H(8) H(8) H(8) H(16)
7 1(8) 1(8) 1(16) C(16) H(16) H(16) H(16)
8 1(16) 1(16) 1(16) 1(32) C(32) H(32)
q p 5 6 7 8
0 H(2) H(2) H(4) C(8) 1(16)
1 H(4) H(4) H(4) H(8) C(16)
2 C(8) H(8) H(8) H(8) H(16)
3 1(16) C(16) H(16) H(16) H(16)
4 1(16) 1(16) 1(32) C(32) H(32)
5 1(32) 1(32) 1(32) 1(64) C(64)
6 C(32) 1(64) 1(64) 1(64) 1(128)
7 H(32) C(64) 1(128) 1(128) 1(128)
8 H(32) H(32) H(64) C(128) 1(256)
245
246 Chapter A Table of Cliord Algebras
Appendix B
Calculation of Fundamental
Groups
Just for the sake of completion Ive added this appendix on calculation of fun-
damental groups of some of the classical Lie groups.
The calculation will involve homotopy theory, so let us just recall the denition
of the higher homotopy groups: Let (X, x
0
) be a pointed topological space. As a
set, the nth homotopy group
n
(X, x
0
) is the set of homotopy classes of contin-
uous maps (I
n
, I
n
) (X, x
0
) relative to the boundary. One can equip this
set with a composition turning it into a group which is abelian if n 2. Further-
more one can show that the construction is a functor, i.e. given a continuous map
f : X Y there are induced homomorphisms f
:
n
(X, x
0
)
n
(Y, f(x
0
)).
The exact details will not concern us here.
Next, recall the notion of a ber bundle: A ber bundle (or a locally trivial
bundle) over some Hausdor topological space X with ber F (also a Hausdor
topological space) is a pair (P, ) of yet another topological space P and a
continuous surjective map : P X such that for each point x X we
can nd a neighborhood U around x and a homeomorphism (called a local
trivialization) :
1
(U) U F of the form (p) = ((p), (p)) where
:
1
(U) F is some continuous map. For such a bundle we will use the
handy notation F P X.
The rst result we will need in order to calculate fundamental groups is the
following standard result from homotopy theory:
Theorem B.1. Let F P X be a ber bundle. Choose a point x
0
X,
let F
0
=
1
(x
0
) P be the ber over x
0
and select p
0
F
0
. Denote by
the inclusion F
0
P. Then for each n 2 there exists a homomorphism
n
:
n
(X, x
0
)
n1
(F
0
, p
0
) such that the following long sequence is exact

n
(F
0
, p
0
)

n
(P, p
0
)

n
(X, x
0
)
n
n1
(F
0
, p
0
)
(B.1)
The long exact sequence (B.1) is known as the homotopy long exact sequence.
The connection with Lie groups is via homogenous manifolds. A homoge-
nous G-manifold is a smooth manifold M equipped with a smooth, transitive
left action of a Lie group G. The prototype of a homogenous manifold is the
following: let G be a Lie group and H a closed subgroup (which is then au-
tomatically a Lie subgroup) where H acts on G from the right by translation:
GH (g, h) gh. This action is smooth, free and proper and thus the orbit
space, denoted G/H, has a smooth structure. Now dene a left G-action on this
247
248 Chapter B Calculation of Fundamental Groups
space by (g, g
0
H) (gg
0
)H. This is a smooth, transitive action of G on G/H
making G/H a homogenous manifold. In fact, any homogenous G-manifold M
is of this form: for the closed subgroup H simply take the isotropy group G
p
of
any point p M, and M will be dieomorphic to G/G
p
. These are all classical
facts from smooth manifold theory.
Lets give some examples which we will need later. First the (n 1)-sphere
is a homogenous manifold: We let O(n) act on the sphere S
n1
1
n
in the
obvious way. The action is clearly smooth, and it is transitive, since we can
move the north pole N = (0, . . . , 0, 1) to any point on the sphere by an ap-
propriate rotation (this also shows that the action of SO(n) on S
n1
is smooth
and transitive, we return to that in an little while). Now we seek the isotropy
group at N that is the group of orthogonal matrices xing N. Such a matrix
must have the form R =
_
A 0
0 1
_
for some n1 n1 matrix A. The orthog-
onality condition R
T
R = I forces A to obey A
T
A, thus A O(n 1). We can
therefore identify the isotropy group at N with O(n 1) and hence we get a
dieomorphism O(n)/ O(n 1)
= S
n1
.
As we noted above, also SO(n) acts smoothly and transitively on S
n1
and as
before the SO(n)-matrices that x N have the form
_
A 0
0 1
_
for an orthogonal
matrix A O(n 1). But
1 = det
_
A 0
0 1
_
= det Adet 1 = det A,
hence A SO(n1) and so we get a dieomorphism SO(n)/ SO(n1)
= S
n1
.
The natural actions of U(n) and SU(n) on S
2n1
C
n
provide us, in the same
manner, with dieomorphisms U(n)/ U(n1)
= S
2n1
and SU(n)/ SU(n1)
=
S
2n1
. These dieomorphisms will come in quite handy later on.
We now provide the link between homogenous manifolds and Theorem B.1:
Proposition B.2. Let G/H be a homogenous manifold. Then G has the struc-
ture of a ber bundle over G/H with ber H.
If we use the notation from above, G would be a H G G/H ber bundle.
Proof. As projection map we simply use the natural map : G G/H
sending g to gH. It is clearly continuous and surjective.
Now we prove the trivialization part. Let x
0
G/H be arbitrary. According
to [14] Theorem 3.58 there exists a neighborhood U G/H around x
0
and
a smooth local section of G, that is a smooth map : U G such that
= id
U
. can be used to dene the trivialization :
1
(U) U H
(p) = ((p), ((p))
1
p).
This map is obviously continuous and has as inverse map
1
(x, h) = (x)h
which is also continuous. Thus, is the desired trivialization.
At this point we are really able to do some calculations. But before doing
so, we would like to get rid of the base point dependence by breaking the Lie
groups up into components (recall that for Lie groups, or generally for manifolds,
the components and path components are the same). Many of the classical Lie
groups are actually connected by virtue of the following result:
1
1
The proof of this result can be found either in [14] (Proposition 3.66) or in [9] (Proposition
9.34).
249
Lemma B.3. Suppose that G is a Lie group acting smoothly, freely and properly
on a manifold M. If G and M/G are connected then so is M.
We can now show:
Proposition B.4. For all n 1, the Lie groups SO(n), U(n) and SU(n) are
connected.
Proof. Lets verify the result for SO(n) by induction. Firstly, SO(1) is just
a point, hence connected. Assume then that SO(n 1) is connected. By the
dieomorphism SO(n)/ SO(n1)
= S
n1
and Lemma B.3 (S
n1
is connected)
we get that SO(n) is connected. The proof for U(n) and SU(n) is exactly the
same, just use that U(1) = S
1
is connected, and that SU(1), like SO(1), is just
a point.
The orthogonal group O(n) is not connected, albeit almost:
Proposition B.5. For every n 1 the group O(n) has two dieomorphic
components O(n)
+
= det
1
(1) and O(n)
= det
1
(1).
Proof. It is clear that O(n) is not connected since det(O(n)) = 1, 1, and
1, 1 is not connected.
Obviously, O(n) = O(n)
+
O(n)
and O(n)
+
O(n)
= . By denition
O(n)
+
is just SO(n) which is connected by Proposition B.4. Thus, O(n)
+
is a
component. Let A O(n) be an arbitrary matrix with determinant 1, then
the map O(n)
+
O(n)
given by X AX is a dieomorphism. Hence,

also O(n)
is connected.
Finally lets do what we set out to do: calculate some fundamental groups.
Theorem B.6. For all n 1 we have
1
(U(n)) = Z and
1
(SU(n)) = 0, i.e.
SU(n) is simply connected.
Proof. We only do U(n). Thanks to the dieomorphism U(n)/ U(n 1)

=
S
2n1
and Proposition B.2 which yields a ber bundle U(n 1) U(n)
S
2n1
, the long exact sequence of homotopy theory give us an exact sequence
2
(S
2n1
)
1
(U(n 1))
1
(U(n))
1
(S
2n1
).
For all n 2
2
(S
2n1
) =
1
(S
2n1
) = 0 and hence exactness of the se-
quence above gives an isomorphism
1
(U(n))
=
1
(U(n1)). Hence
1
(U(n))
1
(U(1)) = Z. Exactly the same for SU(n) except that
1
(SU(1)) = 0.
In the calculation of the fundamental group of SO(n) we will need the fol-
lowing very famous relationship between SO(3) and SU(2) which is important
enough to be stated as a result in its own right. It plays an interesting role in
the quantum mechanical theory of spin
2
.
Lemma B.7. There exists a Lie group isomorphism
SO(3)

SU(2)/I, I.
Theorem B.8. For SO(n) we have the following fundamental groups:
1
(SO(1)) = 0,
1
(SO(2)) = Z and
1
(SO(n)) = Z
2
for n 3.
2
For a proof see [1] Proposition 9.2.
250 Chapter B Calculation of Fundamental Groups
Proof. The rst two fundamental groups are obvious since SO(1) is just a
point, and SO(2) is the circle. For a general n 3 we proceed as before. We
have an exact sequence
2
(S
n1
)
1
(SO(n 1))
1
(SO(n))
1
(S
n1
).
which for n 4, says
1
(SO(n))
=
1
(SO(n1)) since
2
(S
n1
) =
1
(S
n1
) =
0. Inductively
1
(SO(n))
=
1
(SO(3)), so we only need to nd this fundamental
group. This is where we need Lemma B.7. Let I, I SU(2) act in the
natural way and consider the canonical map : SU(2) SU(2)/I, I onto
the orbit space which, according to Lemma B.7, is just SO(3). Since I, I
is a discrete subgroup, is a double covering map
3
and since SU(2) is simply
connected by Theorem B.6 is the universal covering map of SO(3). But then
the fundamental group of SO(3) has the same order as the number of sheets in
the covering, i.e.
1
(SO(3)) must be isomorphic to Z
2
.
Proposition B.5 says that the two components of O(n) are both dieomorphic
to SO(n) and so renders the following result:
Corollary B.9. For all n 1 we have
1
(O(n)
+
) =
1
(O(n)
= Z
2
.
3
See [9] Proposition 9.26.
Bibliography
[AA] M. F. Atiyah, D. W. Anderson: K-Theory. W. A. Benjamin Inc., 1967.
[BT] Raoul Bott, Loring Tu: Dierential Forms in Algebraic Topology. Gradu-
ate Texts in Mathematics nr. 82, Springer, 1982.
[LLR] N. J. Laustsen, F. Larsen and M. Rrdam: An Introduction to K-Theory
for C*-Algebras. London Mathematical Society, Student Texts, 2000.
[Lee] John M. Lee: Riemannian Manifolds. An Introduction to Curvature. Grad-
uate Texts in Mathematics, 1997.
[MT] Ib Madsen, Jrgen Tornehave: From Calculus to Cohomology. Cambridge
University Press, 1999.
[Mi] J. W. Milnor, J. Stashe: Characteristic Classes. Annals of Mathematics
Studies, 1974.
[Mo] Shigeyuki Morita: Geometry of Dierential Forms. Translation of Mathe-
matical Monographs, 2000.
251
Index
abstract root system, 68
reduced, 68
A-class, 196
action, 151
acyclic complex, 156
ad-nilpotent, 33
adjoint representation, 30
algebra
complexication of, 120
analytic index map, 236
angular momentum, 105, 107
angular momentum operator, 105
antisymmetric operator, 97
arc measure, 16
associated ber bundle, 169
associated measure, 15
associated vector bundle, 169
connection, 171
Atiyah-Singer Index Theorem, 235, 240,
242
Atiyah-Singer operator, 220
Atkinsons Theorem, 226
base point, 142
based G-space, 153
based space, 142
Bernoulli numbers, 195
Betti number, 229
Bianchi identity, 181
bilinear form
negative denite, 111
non-degenerate, 111
positive denite, 111
bosonic subalgebra, 116
Bott periodicity, 150, 151, 161
from Thom isomorphism, 161
bundle map
equivariant, 152
C
-vectors, 92
are dense in H, 95
cancellation property, 140, 152
canonical 1-form, 171
canonical anti-automorphism, 115
canonical automorphism, 116
canonical generator, 81
Cartan formula, 205, 209
Cartan subalgebra, 59, 87
existence, 59
real form, 67
Cartans Criterion, 36
Cartan-Bott Theorem, 119, 121
Cartan-Dieudonne Theorem, 129
Casimir element, 55
Cauchy-Schwartz inequality, 26
center, 31
centralizer, 30
character, 22, 24
irreducible, 22
characteristic class, 182
Chern character, 197
in K-theory, 199
on non-compact spaces, 199
Chern class, 184
sum formula, 184
total, 184
uniqueness of, 191
chiral spinors
even, 123
odd, 123
circle group, 14
irreduciblerepresentations, 14
class function, 20
Clebsch-Gordan theory, 12
Cl(), 125
Cliord algebra, 111
Cl
0,1
, 114
Cl
0,2
, 114
Cl
p,q
, 114
existence, 111
functoriality of, 113
uniqueness, 111
Z
2
-grading, 116
Cliord bundle, 210
Cliord group, 125
Cliord module, 211
compact G-pair, 153
compact pair, 144
completely reducible, 12, 26
complex
252
INDEX 253
elliptic, 228, 229
Euler characteristic, 229
harmonic section, 229
complex n-spinor, 123
complexication, 186
of a Lie algebra, 39
of a representation, 52
of a vector space, 39
cone, 145
conjugate bundle, 185
conjugation, 126
connection, 164
1-form, 165
complexication, 186
direct sum, 185
at, 176
formal adjoint, 205
is a dierential operator, 202
Levi-Civita, 180
metric, 178, 180
pullback, 173
Riemannian, 180
symmetric, 180
trivial, 164
connection Laplacian, 209
symbol, 209
co-root, 63
covariant derivative, 164
covariant exterior derivative, 176
covering action, 131
covering map, 131
covering space, 131
curvature, 176
2-form, 177
cyclic highest weight vector, 76
de Rham cohomology, 182
de Rham complex, 230
is elliptic, 230
dening representation, 10, 88
derived algebra, 31
derived series, 31
dierentiability, 91
dierential operator, 201
continuous family, 226
elliptic, 209
homotopy, 227
symbol, 206
Dirac bundle, 211
Dirac Laplacian, 217
Dirac operator, 220
formally self-adjoint, 217
is elliptic, 217
of the exterior bundle, 218
symbol, 217
Dirac spinor, 123
eld, 212
Dirac spinor eld, 212
Dirac type operator, 209
direct sum, 10, 51, 52
of Hilbert spaces, 10
directional derivative, 91
Dixmier-Malliavin Theorem, 104
dominant element, 73, 87
dominated convergence, 93
dual bundle, 185
elementary symmetric polynomials, 192
elliptic complex, 228
elliptic operator, 209
is Fredholm, 226
elliptic regularity, 225
endomorphism algebra, 30
Engels Theorem, 34
equivalence of representations, 11, 50
equivalent representations, 11, 50
equivariant K-group, 152
equivariant bundle map, 152
equivariant function, 169
equivariant map, 151
equivariant section, 152
essentially skew-adjoint, 97
Euler characteristic, 229
Euler class, 189
and nonzero sections, 190
and the Thom isomorphism, 237
even action, 131
evenly covered, 131
exterior bundle
Dirac operator, 218
is a Dirac bundle, 211
exterior derivative, 218
formal adjoint, 205, 218
symbol, 208
exterior product, 149, 150
fermionic subspace, 116
F-genus, 194
ber bundle, 247
ber metric, 178
ltration of the tensor algebra, 44
at connection, 176
formal adjoint, 204
formal power series, 193
Fourier coecient, 27
Fourier series, 27
Fourier theory, 27
Fredholm index, 226
Fredholm operator, 226
254 INDEX
fundamental class, 194, 239, 240
fundamental group
of O(n), 250
of SO(n), 249
of SU(n), 249
of U(n), 249
fundamental representation, 13
fundamental representations, 88
fundamental system, 70, 87
Fundamental Theorem of Riemannian
Geometry, 180
G-complex, 155
acyclic, 156
direct sum, 157
homotopy, 157
pullback, 157
support, 155
tensor product, 157
G-homotopic maps, 152
G-map, 151
G-module, 9
G-space, 151
G-vector bundle, 151
g-module, 49
Gauss-Bonnet Theorem, 242
Gelfands Theorem, 13
generalized Hodge Theorem, 229
generalized Laplacian, 209
graded module, 149
graded tensor product, 116
Gysin homomorphism, 236
Grding subspace, 94, 95
is dense in H, 95
Grding vector, 94
Grdings Theorem, 95
Haar integral, 15
left, 15
right, 15
Haar measure, 15
for , 16
for 1
n
, 15
left, 15
on compact group, 16
right, 15
half space, 70
harmonic analysis, 27
harmonic element, 227
harmonic section, 229
highest weight, 76, 80, 88
is a dominant integral element, 79
highest weight module, 80
Highest Weight Theorem, 86
highest weight vector, 76, 80
cyclic, 76
Hirzebruch L-class, 196
Hirzebruch L-sequence, 196
Hodge Laplacian, 208
is elliptic, 209
symbol, 208
Hodge star operator, 205
Hodge-de Rham operator, 208, 220, 231
and the Gauss-Bonnet Theorem,
243
Fredholm index, 231
is elliptic, 209, 231
symbol, 208
homogenous manifold, 247
is a ber bundle, 248
homotopy, 157
homotopy equivalence
induces isomorphism, 141
homotopy groups, 247
homotopy long exact sequence, 247
ideal, 29
index function, 235
index notation, 12
innitesimal representation, 50
integral element, 79
integration along ber, 236
intertwiner, 10, 50, 121
intertwining map, 10
intertwining number, 11
invariant polynomial, 180
invariant subspace, 11, 50, 121
irreducible character, 22
irreducible representation, 12, 24, 26,
50
isomorphism of complexes, 156
Jacobi identity, 29
K-group, 140
of a sphere, 150
of a torus, 150
quaternionic, 141
real, 141
relative, 144
ring structure, 141
total, 145
Killing form, 36, 37, 62, 64
radical of, 37
Koszul complex, 156, 160
K-theory
induced map, 141
is a contravariant functor, 141
with compact support, 154
INDEX 255
L-class, 196
left Haar integral, 15
left Haar measure, 15
left regular representation, 27
Leibniz rule, 176
level, 70
Levi-Civita connection, 172, 180
lexicographic ordering, 71
Lie algebra, 29
abelian, 31
indecomposable, 31
nilpotent, 32, 34
radical of, 32
reductive, 40
semisimple, 32, 37
simple, 31
solvable, 31
Lie algebra homomorphism, 30
Lie algebra isomorphism, 30
Lie algebra representation, 49, 137
dimension, 49
equivalence, 50
faithful, 49
induced, 50
irreducible, 50
Lie derivative, 202
Lie group representation, 49
Lie subalgebra, 29
Lipschitz group, 125
local operator, 165, 201
local trivialization, 247
locally trivial bundle, 247
long exact sequence in K-theory, 147,
154
matrix coecient, 18, 24
matrix Lie group, 10
maximal torus, 59
measurable section, 220
metric, 178
modular function, 17, 98
momentum, 102
momentum operator, 104
morphism of complexes, 156
multiplicative sequence, 193
Nash Embedding Theorem, 234
Newton polynomial, 193
recursion formula, 193
Newton relations, 181
norm, 126
normal bundle, 233
normalizer, 31
O(), 113
operator
antisymmetric, 97
essentially skew-adjoint, 97
orbital angular momentum, 105
orientation, 187
orientation class, 239, 240
orthogonal decomposition, 116
orthogonal group, 113
orthogonal isomorphism, 113
orthogonal linear map, 113
orthonormal basis, 115
parametrix, 226
Pauli spin matrices, 105
PBW-basis, 48
PBW-Theorem, 44
Peter-Weyl Theorem, 24, 26
Pfaan, 186
pin group, 128
is a Lie group, 131
Pin
c
(), 128
Pin(), 128
is a double covering of O(), 132
Pin(n), 128
is compact, 132
Poincar duality, 236
Poincar-Birkho-Witt Theorem, 44
polarization identity, 111
Pontrjagin class, 184
of tangent bundle of S
n
, 185
sum formula, 184
total, 184
positive integral, 15
positive root, 70, 75
level, 70
positive roots, 87
positive system, 70, 75
positivity, 70
principal symbol, 206
proper neighborhood, 201
pullback bundle, 140, 172
quadratic form, 111
negative denite, 111
non-degenerate, 111
positive denite, 111
quantization map, 113, 212
quantum mechanics, 102
quotient algebra, 30
radical, 32
of a bilinear form, 54
of the Killing form, 37
rank
256 INDEX
of a root system, 68
real form, 67
realication, 187
is orientable, 188
reduced K-group, 142, 153
reduced K-theory, 142, 153
reduced suspension, 145
reection through hyperplane, 125
relative K-group, 144
Rellich Lemma, 225
representation, 9, 121
completely reducible, 12
complexication, 121
dimension of, 9
direct sum, 22
equivalence, 11, 121
faithful, 9, 49
irreducible, 12, 24, 50, 121
of a Lie group, 49
of Cliord algebras, 122
of spin groups, 135
tensor product, 22
unitary, 10
retract, 148
Riemannian connection, 172, 180
Riemannian vector bundle, 178
Riesz Representation Theorem, 15
right Haar integral, 15
right Haar measure, 15
root, 61, 68
is an integral element, 79
positive, 70, 75
reduced, 68
simple, 70
root reection, 68
root space, 61
root space decomposition, 61, 75
root string, 65
root system
abstract, 68
basis, 70
irreducible, 69
isomorphism, 69
rank, 68
reduced, 68
reducible, 69
root vector, 61
root vectors, 87
rotation group, 104
rotation operator, 104
Schur Orthogonality, 19, 20
Schurs Lemma, 13, 51
section
p-integrable, 221
equivariant, 152
locally p-integrable, 221
measurable, 220
with compact support, 203
sheet of a covering, 131
simple acyclic complex, 156
simple system, 70
six-term exact sequence, 151
SN-decomposition, 35
SO(), 113
Sobolev Embedding Theorem, 103, 225
Sobolev norm, 222, 224
independence of connections, 222
Sobolev space, 223
special orthogonal group, 113
Spectral Theorem, 25
spin bundle
connection, 214
spin connection, 214
Spin group
Lie algebra, 134
spin group, 128
is a Lie group, 131
Spin(3), 130
spin representation
complex, 123
real, 122
spin-Dirac operator, 220
Spin
c
(), 128
spin
c
-representation, 135, 138
irreducibility, 138
is faithful, 138
Spin
c
(3), 130
Spin
c
(4), 131
Spin(), 128
is a double covering of SO(), 132
Spin(n), 128
is compact, 132
is simply connected, 133
is universal covering of SO(n), 133
spin(n), 134
spinor, 122
chiral, 123
Dirac, 123
eld, 212
Weyl, 123
spinor bundle, 212
connection, 214
Dirac operator, 220
is a Dirac bundle, 212
spinor eld, 212
spinor representation, 135
INDEX 257
irreducibility, 136, 137
is faithful, 135
Lie algebra representation, 137
spinorial representation, 12
splitting principle, 191, 198
stable equivalence, 143, 153
Stones Theorem, 102
strong operator topology, 9
super tensor product, 116
suspension, 145
reduced, 145
symbol
of a dierential operator, 206
symbol class, 233
symbol map, 113, 134
symmetric algebra, 42
symmetric polynomial, 192
elementary, 192
tensor algebra, 42
ltration, 44
tensor product, 10, 51, 52, 81
of Hilbert spaces, 10
Thom class, 237
Thom defect, 237
is multiplicative, 237
is natural, 237
of a trivial bundle, 237
Thom Defect Formula, 237
Thom homomorphism, 160
Thom isomorphism
K-theory, 160
and the Euler class, 237
in cohomology, 236
is a module homomorphism, 237
pullbacks, 162, 237
transitivity, 161
Todd class, 196
Todd sequence, 195
topological index, 233, 235
torsion, 171
torsion tensor, 180
torus, 59
maximal, 59
total F-class, 194
total Chern class, 184
sum formula, 184
total Pontrjagin class, 184
sum formula, 184
total

A-class, 196
translation operator, 102
trivial representation, 10
trivialization cover, 139
tubular neighborhood, 233
twisted adjoint representation, 125
kernel of, 126
unimodular group, 17
unitarization, 16
unitary representation, 10, 16
universal covering, 131
universal enveloping algebra, 42
Urysohns Lemma, 13
vector bundle, 139
conjugate, 185
dual, 185
orientable, 187
orientation, 187
oriented, 187
pullback, 140, 172
realication, 187
vector eld
formal adjoint, 205
symbol, 207
vector space
Verma module, 81
volume element, 123, 137
weak derivative, 103
weak integral, 93
wedge product, 175
weight, 60, 80, 87
highest, 76
is an integral element, 79
weight space, 60, 80
weight space decomposition, 60
weight vector, 60, 80
highest, 76
Weyl chamber, 75
Weyl group, 72
Weyl Lemma, 225
Weyl spinors
negative, 123
positive, 123
Weyls Theorem, 56
Whitney sum formula, 184

The Mathematics Any Physicist Should Know

Hochgeladen von

Dokumentinformationen

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

The Mathematics Any Physicist Should Know

Hochgeladen von

Copyright:

Verfügbare Formate

The Mathematics any Physicist

-algebras are involved. In the rst two sections

. Furthermore U is -invariant if and only if

, i.e. u U : (g)v, u = 0. But thats

is invariant by the above. We split

. Thus we have the desired formula.

, where = 1, 2 but the exact notation will

. On the other hand and are actually inequivalent as we will se in Sec-

is the hermitian adjoint of F. This yields

. Its a commutative unital C

-algebra, and all the elements are of the

, so lets nd them: It is well-known that the

are those of the form x e

is invariant as well cf. Lemma 1.8. If

the space of linear combinations of matrix coecient. Since

need not all be matrix coecients for .

(A)(g) = Tr((g) A). (1.5)

(A). Consider the map L

(A) is a linear combination of

A), it is easy to see

was surjective by Lemma 1.23. To show injectivity we dene

and show that this is unitary. Since the linear maps L

is unitary and in particular injective.

is bijective, so there is a unique A End(H) for which =

intertwines and leads to

which shows that

. In due course we will prove some powerful or-

= 1 and decompose into irreducibles:

,= 0. Now, suppose that M(G)

,= 0). Then we can pick an nite orthonormal basis (

be nonzero. As C(G) is dense in L

of T commutes with R(g):

T commutes with R(g). The Spectral

are nite-dimensional. They are also R-invariant,

T)f = R(g)(f) = (R(g)f) (1.8)

. Actually M(G) is R-invariant: all functions are of the form

we see that R(g)f M(G). But then also M(G)

denotes the orthogonal projection, then by Lemma 1.8, P commutes

are all R-invariant sub-

. These are very good candidates to the subspace we wanted:

are all 0. This means by denition of P that

T ker T, where the last inclu-

Tf, f = Tf, Tf, i.e. Tf = 0.

we picked at the beginning, we have

,= 0. If R restricted to this space is not irreducible, it

,= 0. We have a contradiction if we in W can nd

for all h. Using unitarity we get (h

U W is nite-dimensional, nontrivial (it contains T

denote the orthogonal projection onto

. It is not hard to see that P

maps class functions to class functions, hence

C(G, class), the last space being the 1-dimensional

by Lemma 1.27. Hence the space

f is just a scalar multiple of

which by Peter-Weyl I implies f = 0.

, consisting of -linear maps E , is 0, then

be arbitrary. The basis we chose for V readily gives us a

:= X g[ Y h : B(X, Y ) = 0. This is again an ideal in g for if X h

and Y g, then for all Z g we have [Y, Z] h and hence

. To show that we have a decomposition g = h h

is zero. We can do this by showing that it

would be a consequence of ad(h h

) is solvable we invoke the Cartan Criterion: For

we have Tr(ad(X) ad(Y )) = B(X, Y ) = 0